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"
36 #include "afs/afs_args.h"
37 #include "afs/afs_osi.h"
38 #ifdef RX_KERNEL_TRACE
39 #include "rx_kcommon.h"
41 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
45 #undef RXDEBUG /* turn off debugging */
47 #if defined(AFS_SGI_ENV)
48 #include "sys/debug.h"
57 #endif /* AFS_OSF_ENV */
59 #include "afs/sysincludes.h"
60 #include "afsincludes.h"
63 #include "rx_kmutex.h"
64 #include "rx_kernel.h"
68 #include "rx_globals.h"
70 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
71 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
72 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
74 extern afs_int32 afs_termState;
76 #include "sys/lockl.h"
77 #include "sys/lock_def.h"
78 #endif /* AFS_AIX41_ENV */
79 # include "rxgen_consts.h"
81 # include <sys/types.h>
86 # include <afs/afsutil.h>
88 # include <sys/socket.h>
89 # include <sys/file.h>
91 # include <sys/stat.h>
92 # include <netinet/in.h>
93 # include <sys/time.h>
103 # include "rx_user.h"
104 # include "rx_clock.h"
105 # include "rx_queue.h"
106 # include "rx_globals.h"
107 # include "rx_trace.h"
108 # include <afs/rxgen_consts.h>
111 int (*registerProgram) () = 0;
112 int (*swapNameProgram) () = 0;
114 /* Local static routines */
115 static void rxi_DestroyConnectionNoLock(register struct rx_connection *conn);
116 #ifdef RX_ENABLE_LOCKS
117 static void rxi_SetAcksInTransmitQueue(register struct rx_call *call);
120 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
122 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
123 afs_int32 rxi_start_in_error;
125 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
128 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
129 * currently allocated within rx. This number is used to allocate the
130 * memory required to return the statistics when queried.
133 static unsigned int rxi_rpc_peer_stat_cnt;
136 * rxi_rpc_process_stat_cnt counts the total number of local process stat
137 * structures currently allocated within rx. The number is used to allocate
138 * the memory required to return the statistics when queried.
141 static unsigned int rxi_rpc_process_stat_cnt;
143 #if !defined(offsetof)
144 #include <stddef.h> /* for definition of offsetof() */
147 #ifdef AFS_PTHREAD_ENV
151 * Use procedural initialization of mutexes/condition variables
155 extern pthread_mutex_t rx_stats_mutex;
156 extern pthread_mutex_t des_init_mutex;
157 extern pthread_mutex_t des_random_mutex;
158 extern pthread_mutex_t rx_clock_mutex;
159 extern pthread_mutex_t rxi_connCacheMutex;
160 extern pthread_mutex_t rx_event_mutex;
161 extern pthread_mutex_t osi_malloc_mutex;
162 extern pthread_mutex_t event_handler_mutex;
163 extern pthread_mutex_t listener_mutex;
164 extern pthread_mutex_t rx_if_init_mutex;
165 extern pthread_mutex_t rx_if_mutex;
166 extern pthread_mutex_t rxkad_client_uid_mutex;
167 extern pthread_mutex_t rxkad_random_mutex;
169 extern pthread_cond_t rx_event_handler_cond;
170 extern pthread_cond_t rx_listener_cond;
172 static pthread_mutex_t epoch_mutex;
173 static pthread_mutex_t rx_init_mutex;
174 static pthread_mutex_t rx_debug_mutex;
177 rxi_InitPthread(void)
179 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
181 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
183 assert(pthread_mutex_init
184 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
185 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
187 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
189 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
191 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
193 assert(pthread_mutex_init
194 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
195 assert(pthread_mutex_init
196 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
197 assert(pthread_mutex_init
198 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
199 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
201 assert(pthread_mutex_init
202 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
203 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
205 assert(pthread_mutex_init
206 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
207 assert(pthread_mutex_init
208 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
209 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
212 assert(pthread_cond_init
213 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
214 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
216 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
217 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
219 rxkad_global_stats_init();
222 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
223 #define INIT_PTHREAD_LOCKS \
224 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
226 * The rx_stats_mutex mutex protects the following global variables:
231 * rxi_lowConnRefCount
232 * rxi_lowPeerRefCount
241 #define INIT_PTHREAD_LOCKS
245 /* Variables for handling the minProcs implementation. availProcs gives the
246 * number of threads available in the pool at this moment (not counting dudes
247 * executing right now). totalMin gives the total number of procs required
248 * for handling all minProcs requests. minDeficit is a dynamic variable
249 * tracking the # of procs required to satisfy all of the remaining minProcs
251 * For fine grain locking to work, the quota check and the reservation of
252 * a server thread has to come while rxi_availProcs and rxi_minDeficit
253 * are locked. To this end, the code has been modified under #ifdef
254 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
255 * same time. A new function, ReturnToServerPool() returns the allocation.
257 * A call can be on several queue's (but only one at a time). When
258 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
259 * that no one else is touching the queue. To this end, we store the address
260 * of the queue lock in the call structure (under the call lock) when we
261 * put the call on a queue, and we clear the call_queue_lock when the
262 * call is removed from a queue (once the call lock has been obtained).
263 * This allows rxi_ResetCall to safely synchronize with others wishing
264 * to manipulate the queue.
267 #ifdef RX_ENABLE_LOCKS
268 static afs_kmutex_t rx_rpc_stats;
269 void rxi_StartUnlocked();
272 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
273 ** pretty good that the next packet coming in is from the same connection
274 ** as the last packet, since we're send multiple packets in a transmit window.
276 struct rx_connection *rxLastConn = 0;
278 #ifdef RX_ENABLE_LOCKS
279 /* The locking hierarchy for rx fine grain locking is composed of these
282 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
283 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
284 * call->lock - locks call data fields.
285 * These are independent of each other:
286 * rx_freeCallQueue_lock
291 * serverQueueEntry->lock
293 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
294 * peer->lock - locks peer data fields.
295 * conn_data_lock - that more than one thread is not updating a conn data
296 * field at the same time.
304 * Do we need a lock to protect the peer field in the conn structure?
305 * conn->peer was previously a constant for all intents and so has no
306 * lock protecting this field. The multihomed client delta introduced
307 * a RX code change : change the peer field in the connection structure
308 * to that remote inetrface from which the last packet for this
309 * connection was sent out. This may become an issue if further changes
312 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
313 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
315 /* rxdb_fileID is used to identify the lock location, along with line#. */
316 static int rxdb_fileID = RXDB_FILE_RX;
317 #endif /* RX_LOCKS_DB */
318 #else /* RX_ENABLE_LOCKS */
319 #define SET_CALL_QUEUE_LOCK(C, L)
320 #define CLEAR_CALL_QUEUE_LOCK(C)
321 #endif /* RX_ENABLE_LOCKS */
322 struct rx_serverQueueEntry *rx_waitForPacket = 0;
323 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
325 /* ------------Exported Interfaces------------- */
327 /* This function allows rxkad to set the epoch to a suitably random number
328 * which rx_NewConnection will use in the future. The principle purpose is to
329 * get rxnull connections to use the same epoch as the rxkad connections do, at
330 * least once the first rxkad connection is established. This is important now
331 * that the host/port addresses aren't used in FindConnection: the uniqueness
332 * of epoch/cid matters and the start time won't do. */
334 #ifdef AFS_PTHREAD_ENV
336 * This mutex protects the following global variables:
340 #define LOCK_EPOCH assert(pthread_mutex_lock(&epoch_mutex)==0)
341 #define UNLOCK_EPOCH assert(pthread_mutex_unlock(&epoch_mutex)==0)
345 #endif /* AFS_PTHREAD_ENV */
348 rx_SetEpoch(afs_uint32 epoch)
355 /* Initialize rx. A port number may be mentioned, in which case this
356 * becomes the default port number for any service installed later.
357 * If 0 is provided for the port number, a random port will be chosen
358 * by the kernel. Whether this will ever overlap anything in
359 * /etc/services is anybody's guess... Returns 0 on success, -1 on
361 static int rxinit_status = 1;
362 #ifdef AFS_PTHREAD_ENV
364 * This mutex protects the following global variables:
368 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0)
369 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0)
372 #define UNLOCK_RX_INIT
376 rx_InitHost(u_int host, u_int port)
383 char *htable, *ptable;
386 #if defined(AFS_DJGPP_ENV) && !defined(DEBUG)
387 __djgpp_set_quiet_socket(1);
394 if (rxinit_status == 0) {
395 tmp_status = rxinit_status;
397 return tmp_status; /* Already started; return previous error code. */
400 if (afs_winsockInit() < 0)
406 * Initialize anything necessary to provide a non-premptive threading
409 rxi_InitializeThreadSupport();
412 /* Allocate and initialize a socket for client and perhaps server
415 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
416 if (rx_socket == OSI_NULLSOCKET) {
420 #ifdef RX_ENABLE_LOCKS
423 #endif /* RX_LOCKS_DB */
424 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
425 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
426 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
427 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
428 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
430 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
432 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
434 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
436 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
438 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
440 #if defined(KERNEL) && defined(AFS_HPUX110_ENV)
442 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
443 #endif /* KERNEL && AFS_HPUX110_ENV */
444 #else /* RX_ENABLE_LOCKS */
445 #if defined(KERNEL) && defined(AFS_GLOBAL_SUNLOCK) && !defined(AFS_HPUX_ENV) && !defined(AFS_OBSD_ENV)
446 mutex_init(&afs_rxglobal_lock, "afs_rxglobal_lock", MUTEX_DEFAULT, NULL);
447 #endif /* AFS_GLOBAL_SUNLOCK */
448 #endif /* RX_ENABLE_LOCKS */
451 rx_connDeadTime = 12;
452 rx_tranquil = 0; /* reset flag */
453 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
455 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
456 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
457 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
458 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
459 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
460 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
462 /* Malloc up a bunch of packets & buffers */
464 queue_Init(&rx_freePacketQueue);
465 rxi_NeedMorePackets = FALSE;
466 #ifdef RX_ENABLE_TSFPQ
467 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
468 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
469 #else /* RX_ENABLE_TSFPQ */
470 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
471 rxi_MorePackets(rx_nPackets);
472 #endif /* RX_ENABLE_TSFPQ */
479 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
480 tv.tv_sec = clock_now.sec;
481 tv.tv_usec = clock_now.usec;
482 srand((unsigned int)tv.tv_usec);
489 #if defined(KERNEL) && !defined(UKERNEL)
490 /* Really, this should never happen in a real kernel */
493 struct sockaddr_in addr;
494 int addrlen = sizeof(addr);
495 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
499 rx_port = addr.sin_port;
502 rx_stats.minRtt.sec = 9999999;
504 rx_SetEpoch(tv.tv_sec | 0x80000000);
506 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
507 * will provide a randomer value. */
509 MUTEX_ENTER(&rx_stats_mutex);
510 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
511 MUTEX_EXIT(&rx_stats_mutex);
512 /* *Slightly* random start time for the cid. This is just to help
513 * out with the hashing function at the peer */
514 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
515 rx_connHashTable = (struct rx_connection **)htable;
516 rx_peerHashTable = (struct rx_peer **)ptable;
518 rx_lastAckDelay.sec = 0;
519 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
520 rx_hardAckDelay.sec = 0;
521 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
522 rx_softAckDelay.sec = 0;
523 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
525 rxevent_Init(20, rxi_ReScheduleEvents);
527 /* Initialize various global queues */
528 queue_Init(&rx_idleServerQueue);
529 queue_Init(&rx_incomingCallQueue);
530 queue_Init(&rx_freeCallQueue);
532 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
533 /* Initialize our list of usable IP addresses. */
537 /* Start listener process (exact function is dependent on the
538 * implementation environment--kernel or user space) */
542 tmp_status = rxinit_status = 0;
550 return rx_InitHost(htonl(INADDR_ANY), port);
553 /* called with unincremented nRequestsRunning to see if it is OK to start
554 * a new thread in this service. Could be "no" for two reasons: over the
555 * max quota, or would prevent others from reaching their min quota.
557 #ifdef RX_ENABLE_LOCKS
558 /* This verion of QuotaOK reserves quota if it's ok while the
559 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
562 QuotaOK(register struct rx_service *aservice)
564 /* check if over max quota */
565 if (aservice->nRequestsRunning >= aservice->maxProcs) {
569 /* under min quota, we're OK */
570 /* otherwise, can use only if there are enough to allow everyone
571 * to go to their min quota after this guy starts.
573 MUTEX_ENTER(&rx_stats_mutex);
574 if ((aservice->nRequestsRunning < aservice->minProcs)
575 || (rxi_availProcs > rxi_minDeficit)) {
576 aservice->nRequestsRunning++;
577 /* just started call in minProcs pool, need fewer to maintain
579 if (aservice->nRequestsRunning <= aservice->minProcs)
582 MUTEX_EXIT(&rx_stats_mutex);
585 MUTEX_EXIT(&rx_stats_mutex);
591 ReturnToServerPool(register struct rx_service *aservice)
593 aservice->nRequestsRunning--;
594 MUTEX_ENTER(&rx_stats_mutex);
595 if (aservice->nRequestsRunning < aservice->minProcs)
598 MUTEX_EXIT(&rx_stats_mutex);
601 #else /* RX_ENABLE_LOCKS */
603 QuotaOK(register struct rx_service *aservice)
606 /* under min quota, we're OK */
607 if (aservice->nRequestsRunning < aservice->minProcs)
610 /* check if over max quota */
611 if (aservice->nRequestsRunning >= aservice->maxProcs)
614 /* otherwise, can use only if there are enough to allow everyone
615 * to go to their min quota after this guy starts.
617 if (rxi_availProcs > rxi_minDeficit)
621 #endif /* RX_ENABLE_LOCKS */
624 /* Called by rx_StartServer to start up lwp's to service calls.
625 NExistingProcs gives the number of procs already existing, and which
626 therefore needn't be created. */
628 rxi_StartServerProcs(int nExistingProcs)
630 register struct rx_service *service;
635 /* For each service, reserve N processes, where N is the "minimum"
636 * number of processes that MUST be able to execute a request in parallel,
637 * at any time, for that process. Also compute the maximum difference
638 * between any service's maximum number of processes that can run
639 * (i.e. the maximum number that ever will be run, and a guarantee
640 * that this number will run if other services aren't running), and its
641 * minimum number. The result is the extra number of processes that
642 * we need in order to provide the latter guarantee */
643 for (i = 0; i < RX_MAX_SERVICES; i++) {
645 service = rx_services[i];
646 if (service == (struct rx_service *)0)
648 nProcs += service->minProcs;
649 diff = service->maxProcs - service->minProcs;
653 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
654 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
655 for (i = 0; i < nProcs; i++) {
656 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
662 /* This routine is only required on Windows */
664 rx_StartClientThread(void)
666 #ifdef AFS_PTHREAD_ENV
668 pid = (int) pthread_self();
669 #endif /* AFS_PTHREAD_ENV */
671 #endif /* AFS_NT40_ENV */
673 /* This routine must be called if any services are exported. If the
674 * donateMe flag is set, the calling process is donated to the server
677 rx_StartServer(int donateMe)
679 register struct rx_service *service;
685 /* Start server processes, if necessary (exact function is dependent
686 * on the implementation environment--kernel or user space). DonateMe
687 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
688 * case, one less new proc will be created rx_StartServerProcs.
690 rxi_StartServerProcs(donateMe);
692 /* count up the # of threads in minProcs, and add set the min deficit to
693 * be that value, too.
695 for (i = 0; i < RX_MAX_SERVICES; i++) {
696 service = rx_services[i];
697 if (service == (struct rx_service *)0)
699 MUTEX_ENTER(&rx_stats_mutex);
700 rxi_totalMin += service->minProcs;
701 /* below works even if a thread is running, since minDeficit would
702 * still have been decremented and later re-incremented.
704 rxi_minDeficit += service->minProcs;
705 MUTEX_EXIT(&rx_stats_mutex);
708 /* Turn on reaping of idle server connections */
709 rxi_ReapConnections();
718 #ifdef AFS_PTHREAD_ENV
720 pid = (pid_t) pthread_self();
721 #else /* AFS_PTHREAD_ENV */
723 LWP_CurrentProcess(&pid);
724 #endif /* AFS_PTHREAD_ENV */
726 sprintf(name, "srv_%d", ++nProcs);
728 (*registerProgram) (pid, name);
730 #endif /* AFS_NT40_ENV */
731 rx_ServerProc(); /* Never returns */
733 #ifdef RX_ENABLE_TSFPQ
734 /* no use leaving packets around in this thread's local queue if
735 * it isn't getting donated to the server thread pool.
737 rxi_FlushLocalPacketsTSFPQ();
738 #endif /* RX_ENABLE_TSFPQ */
742 /* Create a new client connection to the specified service, using the
743 * specified security object to implement the security model for this
745 struct rx_connection *
746 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
747 register struct rx_securityClass *securityObject,
748 int serviceSecurityIndex)
752 register struct rx_connection *conn;
757 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", shost, sport, sservice, securityObject, serviceSecurityIndex));
759 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
760 * the case of kmem_alloc? */
761 conn = rxi_AllocConnection();
762 #ifdef RX_ENABLE_LOCKS
763 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
764 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
765 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
768 MUTEX_ENTER(&rx_connHashTable_lock);
769 cid = (rx_nextCid += RX_MAXCALLS);
770 conn->type = RX_CLIENT_CONNECTION;
772 conn->epoch = rx_epoch;
773 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
774 conn->serviceId = sservice;
775 conn->securityObject = securityObject;
776 /* This doesn't work in all compilers with void (they're buggy), so fake it
778 conn->securityData = (VOID *) 0;
779 conn->securityIndex = serviceSecurityIndex;
780 rx_SetConnDeadTime(conn, rx_connDeadTime);
781 conn->ackRate = RX_FAST_ACK_RATE;
783 conn->specific = NULL;
784 conn->challengeEvent = NULL;
785 conn->delayedAbortEvent = NULL;
786 conn->abortCount = 0;
789 RXS_NewConnection(securityObject, conn);
791 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
793 conn->refCount++; /* no lock required since only this thread knows... */
794 conn->next = rx_connHashTable[hashindex];
795 rx_connHashTable[hashindex] = conn;
796 MUTEX_ENTER(&rx_stats_mutex);
797 rx_stats.nClientConns++;
798 MUTEX_EXIT(&rx_stats_mutex);
800 MUTEX_EXIT(&rx_connHashTable_lock);
806 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
808 /* The idea is to set the dead time to a value that allows several
809 * keepalives to be dropped without timing out the connection. */
810 conn->secondsUntilDead = MAX(seconds, 6);
811 conn->secondsUntilPing = conn->secondsUntilDead / 6;
814 int rxi_lowPeerRefCount = 0;
815 int rxi_lowConnRefCount = 0;
818 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
819 * NOTE: must not be called with rx_connHashTable_lock held.
822 rxi_CleanupConnection(struct rx_connection *conn)
824 /* Notify the service exporter, if requested, that this connection
825 * is being destroyed */
826 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
827 (*conn->service->destroyConnProc) (conn);
829 /* Notify the security module that this connection is being destroyed */
830 RXS_DestroyConnection(conn->securityObject, conn);
832 /* If this is the last connection using the rx_peer struct, set its
833 * idle time to now. rxi_ReapConnections will reap it if it's still
834 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
836 MUTEX_ENTER(&rx_peerHashTable_lock);
837 if (conn->peer->refCount < 2) {
838 conn->peer->idleWhen = clock_Sec();
839 if (conn->peer->refCount < 1) {
840 conn->peer->refCount = 1;
841 MUTEX_ENTER(&rx_stats_mutex);
842 rxi_lowPeerRefCount++;
843 MUTEX_EXIT(&rx_stats_mutex);
846 conn->peer->refCount--;
847 MUTEX_EXIT(&rx_peerHashTable_lock);
849 MUTEX_ENTER(&rx_stats_mutex);
850 if (conn->type == RX_SERVER_CONNECTION)
851 rx_stats.nServerConns--;
853 rx_stats.nClientConns--;
854 MUTEX_EXIT(&rx_stats_mutex);
857 if (conn->specific) {
859 for (i = 0; i < conn->nSpecific; i++) {
860 if (conn->specific[i] && rxi_keyCreate_destructor[i])
861 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
862 conn->specific[i] = NULL;
864 free(conn->specific);
866 conn->specific = NULL;
870 MUTEX_DESTROY(&conn->conn_call_lock);
871 MUTEX_DESTROY(&conn->conn_data_lock);
872 CV_DESTROY(&conn->conn_call_cv);
874 rxi_FreeConnection(conn);
877 /* Destroy the specified connection */
879 rxi_DestroyConnection(register struct rx_connection *conn)
881 MUTEX_ENTER(&rx_connHashTable_lock);
882 rxi_DestroyConnectionNoLock(conn);
883 /* conn should be at the head of the cleanup list */
884 if (conn == rx_connCleanup_list) {
885 rx_connCleanup_list = rx_connCleanup_list->next;
886 MUTEX_EXIT(&rx_connHashTable_lock);
887 rxi_CleanupConnection(conn);
889 #ifdef RX_ENABLE_LOCKS
891 MUTEX_EXIT(&rx_connHashTable_lock);
893 #endif /* RX_ENABLE_LOCKS */
897 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
899 register struct rx_connection **conn_ptr;
900 register int havecalls = 0;
901 struct rx_packet *packet;
908 MUTEX_ENTER(&conn->conn_data_lock);
909 if (conn->refCount > 0)
912 MUTEX_ENTER(&rx_stats_mutex);
913 rxi_lowConnRefCount++;
914 MUTEX_EXIT(&rx_stats_mutex);
917 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
918 /* Busy; wait till the last guy before proceeding */
919 MUTEX_EXIT(&conn->conn_data_lock);
924 /* If the client previously called rx_NewCall, but it is still
925 * waiting, treat this as a running call, and wait to destroy the
926 * connection later when the call completes. */
927 if ((conn->type == RX_CLIENT_CONNECTION)
928 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
929 conn->flags |= RX_CONN_DESTROY_ME;
930 MUTEX_EXIT(&conn->conn_data_lock);
934 MUTEX_EXIT(&conn->conn_data_lock);
936 /* Check for extant references to this connection */
937 for (i = 0; i < RX_MAXCALLS; i++) {
938 register struct rx_call *call = conn->call[i];
941 if (conn->type == RX_CLIENT_CONNECTION) {
942 MUTEX_ENTER(&call->lock);
943 if (call->delayedAckEvent) {
944 /* Push the final acknowledgment out now--there
945 * won't be a subsequent call to acknowledge the
946 * last reply packets */
947 rxevent_Cancel(call->delayedAckEvent, call,
948 RX_CALL_REFCOUNT_DELAY);
949 if (call->state == RX_STATE_PRECALL
950 || call->state == RX_STATE_ACTIVE) {
951 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
953 rxi_AckAll(NULL, call, 0);
956 MUTEX_EXIT(&call->lock);
960 #ifdef RX_ENABLE_LOCKS
962 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
963 MUTEX_EXIT(&conn->conn_data_lock);
965 /* Someone is accessing a packet right now. */
969 #endif /* RX_ENABLE_LOCKS */
972 /* Don't destroy the connection if there are any call
973 * structures still in use */
974 MUTEX_ENTER(&conn->conn_data_lock);
975 conn->flags |= RX_CONN_DESTROY_ME;
976 MUTEX_EXIT(&conn->conn_data_lock);
981 if (conn->delayedAbortEvent) {
982 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
983 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
985 MUTEX_ENTER(&conn->conn_data_lock);
986 rxi_SendConnectionAbort(conn, packet, 0, 1);
987 MUTEX_EXIT(&conn->conn_data_lock);
988 rxi_FreePacket(packet);
992 /* Remove from connection hash table before proceeding */
994 &rx_connHashTable[CONN_HASH
995 (peer->host, peer->port, conn->cid, conn->epoch,
997 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
998 if (*conn_ptr == conn) {
999 *conn_ptr = conn->next;
1003 /* if the conn that we are destroying was the last connection, then we
1004 * clear rxLastConn as well */
1005 if (rxLastConn == conn)
1008 /* Make sure the connection is completely reset before deleting it. */
1009 /* get rid of pending events that could zap us later */
1010 if (conn->challengeEvent)
1011 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1012 if (conn->checkReachEvent)
1013 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1015 /* Add the connection to the list of destroyed connections that
1016 * need to be cleaned up. This is necessary to avoid deadlocks
1017 * in the routines we call to inform others that this connection is
1018 * being destroyed. */
1019 conn->next = rx_connCleanup_list;
1020 rx_connCleanup_list = conn;
1023 /* Externally available version */
1025 rx_DestroyConnection(register struct rx_connection *conn)
1030 rxi_DestroyConnection(conn);
1035 rx_GetConnection(register struct rx_connection *conn)
1040 MUTEX_ENTER(&conn->conn_data_lock);
1042 MUTEX_EXIT(&conn->conn_data_lock);
1046 /* Start a new rx remote procedure call, on the specified connection.
1047 * If wait is set to 1, wait for a free call channel; otherwise return
1048 * 0. Maxtime gives the maximum number of seconds this call may take,
1049 * after rx_MakeCall returns. After this time interval, a call to any
1050 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1051 * For fine grain locking, we hold the conn_call_lock in order to
1052 * to ensure that we don't get signalle after we found a call in an active
1053 * state and before we go to sleep.
1056 rx_NewCall(register struct rx_connection *conn)
1059 register struct rx_call *call;
1060 struct clock queueTime;
1064 dpf(("rx_MakeCall(conn %x)\n", conn));
1067 clock_GetTime(&queueTime);
1068 MUTEX_ENTER(&conn->conn_call_lock);
1071 * Check if there are others waiting for a new call.
1072 * If so, let them go first to avoid starving them.
1073 * This is a fairly simple scheme, and might not be
1074 * a complete solution for large numbers of waiters.
1076 * makeCallWaiters keeps track of the number of
1077 * threads waiting to make calls and the
1078 * RX_CONN_MAKECALL_WAITING flag bit is used to
1079 * indicate that there are indeed calls waiting.
1080 * The flag is set when the waiter is incremented.
1081 * It is only cleared in rx_EndCall when
1082 * makeCallWaiters is 0. This prevents us from
1083 * accidently destroying the connection while it
1084 * is potentially about to be used.
1086 MUTEX_ENTER(&conn->conn_data_lock);
1087 if (conn->makeCallWaiters) {
1088 conn->flags |= RX_CONN_MAKECALL_WAITING;
1089 conn->makeCallWaiters++;
1090 MUTEX_EXIT(&conn->conn_data_lock);
1092 #ifdef RX_ENABLE_LOCKS
1093 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1097 MUTEX_ENTER(&conn->conn_data_lock);
1098 conn->makeCallWaiters--;
1100 MUTEX_EXIT(&conn->conn_data_lock);
1103 for (i = 0; i < RX_MAXCALLS; i++) {
1104 call = conn->call[i];
1106 MUTEX_ENTER(&call->lock);
1107 if (call->state == RX_STATE_DALLY) {
1108 rxi_ResetCall(call, 0);
1109 (*call->callNumber)++;
1112 MUTEX_EXIT(&call->lock);
1114 call = rxi_NewCall(conn, i);
1118 if (i < RX_MAXCALLS) {
1121 MUTEX_ENTER(&conn->conn_data_lock);
1122 conn->flags |= RX_CONN_MAKECALL_WAITING;
1123 conn->makeCallWaiters++;
1124 MUTEX_EXIT(&conn->conn_data_lock);
1126 #ifdef RX_ENABLE_LOCKS
1127 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1131 MUTEX_ENTER(&conn->conn_data_lock);
1132 conn->makeCallWaiters--;
1133 MUTEX_EXIT(&conn->conn_data_lock);
1136 * Wake up anyone else who might be giving us a chance to
1137 * run (see code above that avoids resource starvation).
1139 #ifdef RX_ENABLE_LOCKS
1140 CV_BROADCAST(&conn->conn_call_cv);
1145 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1147 /* Client is initially in send mode */
1148 call->state = RX_STATE_ACTIVE;
1149 call->error = conn->error;
1151 call->mode = RX_MODE_ERROR;
1153 call->mode = RX_MODE_SENDING;
1155 /* remember start time for call in case we have hard dead time limit */
1156 call->queueTime = queueTime;
1157 clock_GetTime(&call->startTime);
1158 hzero(call->bytesSent);
1159 hzero(call->bytesRcvd);
1161 /* Turn on busy protocol. */
1162 rxi_KeepAliveOn(call);
1164 MUTEX_EXIT(&call->lock);
1165 MUTEX_EXIT(&conn->conn_call_lock);
1168 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1169 /* Now, if TQ wasn't cleared earlier, do it now. */
1170 MUTEX_ENTER(&call->lock);
1171 while (call->flags & RX_CALL_TQ_BUSY) {
1172 call->flags |= RX_CALL_TQ_WAIT;
1174 #ifdef RX_ENABLE_LOCKS
1175 osirx_AssertMine(&call->lock, "rxi_Start lock4");
1176 CV_WAIT(&call->cv_tq, &call->lock);
1177 #else /* RX_ENABLE_LOCKS */
1178 osi_rxSleep(&call->tq);
1179 #endif /* RX_ENABLE_LOCKS */
1181 if (call->tqWaiters == 0) {
1182 call->flags &= ~RX_CALL_TQ_WAIT;
1185 if (call->flags & RX_CALL_TQ_CLEARME) {
1186 rxi_ClearTransmitQueue(call, 0);
1187 queue_Init(&call->tq);
1189 MUTEX_EXIT(&call->lock);
1190 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1196 rxi_HasActiveCalls(register struct rx_connection *aconn)
1199 register struct rx_call *tcall;
1203 for (i = 0; i < RX_MAXCALLS; i++) {
1204 if ((tcall = aconn->call[i])) {
1205 if ((tcall->state == RX_STATE_ACTIVE)
1206 || (tcall->state == RX_STATE_PRECALL)) {
1217 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1218 register afs_int32 * aint32s)
1221 register struct rx_call *tcall;
1225 for (i = 0; i < RX_MAXCALLS; i++) {
1226 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1227 aint32s[i] = aconn->callNumber[i] + 1;
1229 aint32s[i] = aconn->callNumber[i];
1236 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1237 register afs_int32 * aint32s)
1240 register struct rx_call *tcall;
1244 for (i = 0; i < RX_MAXCALLS; i++) {
1245 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1246 aconn->callNumber[i] = aint32s[i] - 1;
1248 aconn->callNumber[i] = aint32s[i];
1254 /* Advertise a new service. A service is named locally by a UDP port
1255 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1258 char *serviceName; Name for identification purposes (e.g. the
1259 service name might be used for probing for
1262 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1263 struct rx_securityClass **securityObjects, int nSecurityObjects,
1264 afs_int32(*serviceProc) (struct rx_call * acall))
1266 osi_socket socket = OSI_NULLSOCKET;
1267 register struct rx_service *tservice;
1273 if (serviceId == 0) {
1275 "rx_NewService: service id for service %s is not non-zero.\n",
1282 "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",
1290 tservice = rxi_AllocService();
1292 for (i = 0; i < RX_MAX_SERVICES; i++) {
1293 register struct rx_service *service = rx_services[i];
1295 if (port == service->servicePort) {
1296 if (service->serviceId == serviceId) {
1297 /* The identical service has already been
1298 * installed; if the caller was intending to
1299 * change the security classes used by this
1300 * service, he/she loses. */
1302 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1303 serviceName, serviceId, service->serviceName);
1305 rxi_FreeService(tservice);
1308 /* Different service, same port: re-use the socket
1309 * which is bound to the same port */
1310 socket = service->socket;
1313 if (socket == OSI_NULLSOCKET) {
1314 /* If we don't already have a socket (from another
1315 * service on same port) get a new one */
1316 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1317 if (socket == OSI_NULLSOCKET) {
1319 rxi_FreeService(tservice);
1324 service->socket = socket;
1325 service->servicePort = port;
1326 service->serviceId = serviceId;
1327 service->serviceName = serviceName;
1328 service->nSecurityObjects = nSecurityObjects;
1329 service->securityObjects = securityObjects;
1330 service->minProcs = 0;
1331 service->maxProcs = 1;
1332 service->idleDeadTime = 60;
1333 service->connDeadTime = rx_connDeadTime;
1334 service->executeRequestProc = serviceProc;
1335 service->checkReach = 0;
1336 rx_services[i] = service; /* not visible until now */
1342 rxi_FreeService(tservice);
1343 (osi_Msg "rx_NewService: cannot support > %d services\n",
1348 /* Generic request processing loop. This routine should be called
1349 * by the implementation dependent rx_ServerProc. If socketp is
1350 * non-null, it will be set to the file descriptor that this thread
1351 * is now listening on. If socketp is null, this routine will never
1354 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1356 register struct rx_call *call;
1357 register afs_int32 code;
1358 register struct rx_service *tservice = NULL;
1365 call = rx_GetCall(threadID, tservice, socketp);
1366 if (socketp && *socketp != OSI_NULLSOCKET) {
1367 /* We are now a listener thread */
1372 /* if server is restarting( typically smooth shutdown) then do not
1373 * allow any new calls.
1376 if (rx_tranquil && (call != NULL)) {
1380 MUTEX_ENTER(&call->lock);
1382 rxi_CallError(call, RX_RESTARTING);
1383 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1385 MUTEX_EXIT(&call->lock);
1389 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1390 #ifdef RX_ENABLE_LOCKS
1392 #endif /* RX_ENABLE_LOCKS */
1393 afs_termState = AFSOP_STOP_AFS;
1394 afs_osi_Wakeup(&afs_termState);
1395 #ifdef RX_ENABLE_LOCKS
1397 #endif /* RX_ENABLE_LOCKS */
1402 tservice = call->conn->service;
1404 if (tservice->beforeProc)
1405 (*tservice->beforeProc) (call);
1407 code = call->conn->service->executeRequestProc(call);
1409 if (tservice->afterProc)
1410 (*tservice->afterProc) (call, code);
1412 rx_EndCall(call, code);
1413 MUTEX_ENTER(&rx_stats_mutex);
1415 MUTEX_EXIT(&rx_stats_mutex);
1421 rx_WakeupServerProcs(void)
1423 struct rx_serverQueueEntry *np, *tqp;
1427 MUTEX_ENTER(&rx_serverPool_lock);
1429 #ifdef RX_ENABLE_LOCKS
1430 if (rx_waitForPacket)
1431 CV_BROADCAST(&rx_waitForPacket->cv);
1432 #else /* RX_ENABLE_LOCKS */
1433 if (rx_waitForPacket)
1434 osi_rxWakeup(rx_waitForPacket);
1435 #endif /* RX_ENABLE_LOCKS */
1436 MUTEX_ENTER(&freeSQEList_lock);
1437 for (np = rx_FreeSQEList; np; np = tqp) {
1438 tqp = *(struct rx_serverQueueEntry **)np;
1439 #ifdef RX_ENABLE_LOCKS
1440 CV_BROADCAST(&np->cv);
1441 #else /* RX_ENABLE_LOCKS */
1443 #endif /* RX_ENABLE_LOCKS */
1445 MUTEX_EXIT(&freeSQEList_lock);
1446 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1447 #ifdef RX_ENABLE_LOCKS
1448 CV_BROADCAST(&np->cv);
1449 #else /* RX_ENABLE_LOCKS */
1451 #endif /* RX_ENABLE_LOCKS */
1453 MUTEX_EXIT(&rx_serverPool_lock);
1458 * One thing that seems to happen is that all the server threads get
1459 * tied up on some empty or slow call, and then a whole bunch of calls
1460 * arrive at once, using up the packet pool, so now there are more
1461 * empty calls. The most critical resources here are server threads
1462 * and the free packet pool. The "doreclaim" code seems to help in
1463 * general. I think that eventually we arrive in this state: there
1464 * are lots of pending calls which do have all their packets present,
1465 * so they won't be reclaimed, are multi-packet calls, so they won't
1466 * be scheduled until later, and thus are tying up most of the free
1467 * packet pool for a very long time.
1469 * 1. schedule multi-packet calls if all the packets are present.
1470 * Probably CPU-bound operation, useful to return packets to pool.
1471 * Do what if there is a full window, but the last packet isn't here?
1472 * 3. preserve one thread which *only* runs "best" calls, otherwise
1473 * it sleeps and waits for that type of call.
1474 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1475 * the current dataquota business is badly broken. The quota isn't adjusted
1476 * to reflect how many packets are presently queued for a running call.
1477 * So, when we schedule a queued call with a full window of packets queued
1478 * up for it, that *should* free up a window full of packets for other 2d-class
1479 * calls to be able to use from the packet pool. But it doesn't.
1481 * NB. Most of the time, this code doesn't run -- since idle server threads
1482 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1483 * as a new call arrives.
1485 /* Sleep until a call arrives. Returns a pointer to the call, ready
1486 * for an rx_Read. */
1487 #ifdef RX_ENABLE_LOCKS
1489 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1491 struct rx_serverQueueEntry *sq;
1492 register struct rx_call *call = (struct rx_call *)0;
1493 struct rx_service *service = NULL;
1496 MUTEX_ENTER(&freeSQEList_lock);
1498 if ((sq = rx_FreeSQEList)) {
1499 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1500 MUTEX_EXIT(&freeSQEList_lock);
1501 } else { /* otherwise allocate a new one and return that */
1502 MUTEX_EXIT(&freeSQEList_lock);
1503 sq = (struct rx_serverQueueEntry *)
1504 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1505 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1506 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1509 MUTEX_ENTER(&rx_serverPool_lock);
1510 if (cur_service != NULL) {
1511 ReturnToServerPool(cur_service);
1514 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1515 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1517 /* Scan for eligible incoming calls. A call is not eligible
1518 * if the maximum number of calls for its service type are
1519 * already executing */
1520 /* One thread will process calls FCFS (to prevent starvation),
1521 * while the other threads may run ahead looking for calls which
1522 * have all their input data available immediately. This helps
1523 * keep threads from blocking, waiting for data from the client. */
1524 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1525 service = tcall->conn->service;
1526 if (!QuotaOK(service)) {
1529 if (tno == rxi_fcfs_thread_num
1530 || !tcall->queue_item_header.next) {
1531 /* If we're the fcfs thread , then we'll just use
1532 * this call. If we haven't been able to find an optimal
1533 * choice, and we're at the end of the list, then use a
1534 * 2d choice if one has been identified. Otherwise... */
1535 call = (choice2 ? choice2 : tcall);
1536 service = call->conn->service;
1537 } else if (!queue_IsEmpty(&tcall->rq)) {
1538 struct rx_packet *rp;
1539 rp = queue_First(&tcall->rq, rx_packet);
1540 if (rp->header.seq == 1) {
1542 || (rp->header.flags & RX_LAST_PACKET)) {
1544 } else if (rxi_2dchoice && !choice2
1545 && !(tcall->flags & RX_CALL_CLEARED)
1546 && (tcall->rprev > rxi_HardAckRate)) {
1555 ReturnToServerPool(service);
1562 MUTEX_EXIT(&rx_serverPool_lock);
1563 MUTEX_ENTER(&call->lock);
1565 if (call->flags & RX_CALL_WAIT_PROC) {
1566 call->flags &= ~RX_CALL_WAIT_PROC;
1567 MUTEX_ENTER(&rx_stats_mutex);
1569 MUTEX_EXIT(&rx_stats_mutex);
1572 if (call->state != RX_STATE_PRECALL || call->error) {
1573 MUTEX_EXIT(&call->lock);
1574 MUTEX_ENTER(&rx_serverPool_lock);
1575 ReturnToServerPool(service);
1580 if (queue_IsEmpty(&call->rq)
1581 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1582 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1584 CLEAR_CALL_QUEUE_LOCK(call);
1587 /* If there are no eligible incoming calls, add this process
1588 * to the idle server queue, to wait for one */
1592 *socketp = OSI_NULLSOCKET;
1594 sq->socketp = socketp;
1595 queue_Append(&rx_idleServerQueue, sq);
1596 #ifndef AFS_AIX41_ENV
1597 rx_waitForPacket = sq;
1599 rx_waitingForPacket = sq;
1600 #endif /* AFS_AIX41_ENV */
1602 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1604 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1605 MUTEX_EXIT(&rx_serverPool_lock);
1606 return (struct rx_call *)0;
1609 } while (!(call = sq->newcall)
1610 && !(socketp && *socketp != OSI_NULLSOCKET));
1611 MUTEX_EXIT(&rx_serverPool_lock);
1613 MUTEX_ENTER(&call->lock);
1619 MUTEX_ENTER(&freeSQEList_lock);
1620 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1621 rx_FreeSQEList = sq;
1622 MUTEX_EXIT(&freeSQEList_lock);
1625 clock_GetTime(&call->startTime);
1626 call->state = RX_STATE_ACTIVE;
1627 call->mode = RX_MODE_RECEIVING;
1628 #ifdef RX_KERNEL_TRACE
1629 if (ICL_SETACTIVE(afs_iclSetp)) {
1630 int glockOwner = ISAFS_GLOCK();
1633 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1634 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1641 rxi_calltrace(RX_CALL_START, call);
1642 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1643 call->conn->service->servicePort, call->conn->service->serviceId,
1646 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1647 MUTEX_EXIT(&call->lock);
1649 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1654 #else /* RX_ENABLE_LOCKS */
1656 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1658 struct rx_serverQueueEntry *sq;
1659 register struct rx_call *call = (struct rx_call *)0, *choice2;
1660 struct rx_service *service = NULL;
1664 MUTEX_ENTER(&freeSQEList_lock);
1666 if ((sq = rx_FreeSQEList)) {
1667 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1668 MUTEX_EXIT(&freeSQEList_lock);
1669 } else { /* otherwise allocate a new one and return that */
1670 MUTEX_EXIT(&freeSQEList_lock);
1671 sq = (struct rx_serverQueueEntry *)
1672 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1673 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1674 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1676 MUTEX_ENTER(&sq->lock);
1678 if (cur_service != NULL) {
1679 cur_service->nRequestsRunning--;
1680 if (cur_service->nRequestsRunning < cur_service->minProcs)
1684 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1685 register struct rx_call *tcall, *ncall;
1686 /* Scan for eligible incoming calls. A call is not eligible
1687 * if the maximum number of calls for its service type are
1688 * already executing */
1689 /* One thread will process calls FCFS (to prevent starvation),
1690 * while the other threads may run ahead looking for calls which
1691 * have all their input data available immediately. This helps
1692 * keep threads from blocking, waiting for data from the client. */
1693 choice2 = (struct rx_call *)0;
1694 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1695 service = tcall->conn->service;
1696 if (QuotaOK(service)) {
1697 if (tno == rxi_fcfs_thread_num
1698 || !tcall->queue_item_header.next) {
1699 /* If we're the fcfs thread, then we'll just use
1700 * this call. If we haven't been able to find an optimal
1701 * choice, and we're at the end of the list, then use a
1702 * 2d choice if one has been identified. Otherwise... */
1703 call = (choice2 ? choice2 : tcall);
1704 service = call->conn->service;
1705 } else if (!queue_IsEmpty(&tcall->rq)) {
1706 struct rx_packet *rp;
1707 rp = queue_First(&tcall->rq, rx_packet);
1708 if (rp->header.seq == 1
1710 || (rp->header.flags & RX_LAST_PACKET))) {
1712 } else if (rxi_2dchoice && !choice2
1713 && !(tcall->flags & RX_CALL_CLEARED)
1714 && (tcall->rprev > rxi_HardAckRate)) {
1727 /* we can't schedule a call if there's no data!!! */
1728 /* send an ack if there's no data, if we're missing the
1729 * first packet, or we're missing something between first
1730 * and last -- there's a "hole" in the incoming data. */
1731 if (queue_IsEmpty(&call->rq)
1732 || queue_First(&call->rq, rx_packet)->header.seq != 1
1733 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1734 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1736 call->flags &= (~RX_CALL_WAIT_PROC);
1737 service->nRequestsRunning++;
1738 /* just started call in minProcs pool, need fewer to maintain
1740 if (service->nRequestsRunning <= service->minProcs)
1744 /* MUTEX_EXIT(&call->lock); */
1746 /* If there are no eligible incoming calls, add this process
1747 * to the idle server queue, to wait for one */
1750 *socketp = OSI_NULLSOCKET;
1752 sq->socketp = socketp;
1753 queue_Append(&rx_idleServerQueue, sq);
1757 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1759 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1760 return (struct rx_call *)0;
1763 } while (!(call = sq->newcall)
1764 && !(socketp && *socketp != OSI_NULLSOCKET));
1766 MUTEX_EXIT(&sq->lock);
1768 MUTEX_ENTER(&freeSQEList_lock);
1769 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1770 rx_FreeSQEList = sq;
1771 MUTEX_EXIT(&freeSQEList_lock);
1774 clock_GetTime(&call->startTime);
1775 call->state = RX_STATE_ACTIVE;
1776 call->mode = RX_MODE_RECEIVING;
1777 #ifdef RX_KERNEL_TRACE
1778 if (ICL_SETACTIVE(afs_iclSetp)) {
1779 int glockOwner = ISAFS_GLOCK();
1782 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1783 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1790 rxi_calltrace(RX_CALL_START, call);
1791 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1792 call->conn->service->servicePort, call->conn->service->serviceId,
1795 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1802 #endif /* RX_ENABLE_LOCKS */
1806 /* Establish a procedure to be called when a packet arrives for a
1807 * call. This routine will be called at most once after each call,
1808 * and will also be called if there is an error condition on the or
1809 * the call is complete. Used by multi rx to build a selection
1810 * function which determines which of several calls is likely to be a
1811 * good one to read from.
1812 * NOTE: the way this is currently implemented it is probably only a
1813 * good idea to (1) use it immediately after a newcall (clients only)
1814 * and (2) only use it once. Other uses currently void your warranty
1817 rx_SetArrivalProc(register struct rx_call *call,
1818 register void (*proc) (register struct rx_call * call,
1820 register int index),
1821 register VOID * handle, register int arg)
1823 call->arrivalProc = proc;
1824 call->arrivalProcHandle = handle;
1825 call->arrivalProcArg = arg;
1828 /* Call is finished (possibly prematurely). Return rc to the peer, if
1829 * appropriate, and return the final error code from the conversation
1833 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1835 register struct rx_connection *conn = call->conn;
1836 register struct rx_service *service;
1837 register struct rx_packet *tp; /* Temporary packet pointer */
1838 register struct rx_packet *nxp; /* Next packet pointer, for queue_Scan */
1842 dpf(("rx_EndCall(call %x)\n", call));
1845 MUTEX_ENTER(&call->lock);
1847 if (rc == 0 && call->error == 0) {
1848 call->abortCode = 0;
1849 call->abortCount = 0;
1852 call->arrivalProc = (void (*)())0;
1853 if (rc && call->error == 0) {
1854 rxi_CallError(call, rc);
1855 /* Send an abort message to the peer if this error code has
1856 * only just been set. If it was set previously, assume the
1857 * peer has already been sent the error code or will request it
1859 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1861 if (conn->type == RX_SERVER_CONNECTION) {
1862 /* Make sure reply or at least dummy reply is sent */
1863 if (call->mode == RX_MODE_RECEIVING) {
1864 rxi_WriteProc(call, 0, 0);
1866 if (call->mode == RX_MODE_SENDING) {
1867 rxi_FlushWrite(call);
1869 service = conn->service;
1870 rxi_calltrace(RX_CALL_END, call);
1871 /* Call goes to hold state until reply packets are acknowledged */
1872 if (call->tfirst + call->nSoftAcked < call->tnext) {
1873 call->state = RX_STATE_HOLD;
1875 call->state = RX_STATE_DALLY;
1876 rxi_ClearTransmitQueue(call, 0);
1877 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1878 rxevent_Cancel(call->keepAliveEvent, call,
1879 RX_CALL_REFCOUNT_ALIVE);
1881 } else { /* Client connection */
1883 /* Make sure server receives input packets, in the case where
1884 * no reply arguments are expected */
1885 if ((call->mode == RX_MODE_SENDING)
1886 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1887 (void)rxi_ReadProc(call, &dummy, 1);
1890 /* If we had an outstanding delayed ack, be nice to the server
1891 * and force-send it now.
1893 if (call->delayedAckEvent) {
1894 rxevent_Cancel(call->delayedAckEvent, call,
1895 RX_CALL_REFCOUNT_DELAY);
1896 call->delayedAckEvent = NULL;
1897 rxi_SendDelayedAck(NULL, call, NULL);
1900 /* We need to release the call lock since it's lower than the
1901 * conn_call_lock and we don't want to hold the conn_call_lock
1902 * over the rx_ReadProc call. The conn_call_lock needs to be held
1903 * here for the case where rx_NewCall is perusing the calls on
1904 * the connection structure. We don't want to signal until
1905 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1906 * have checked this call, found it active and by the time it
1907 * goes to sleep, will have missed the signal.
1909 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
1910 * there are threads waiting to use the conn object.
1912 MUTEX_EXIT(&call->lock);
1913 MUTEX_ENTER(&conn->conn_call_lock);
1914 MUTEX_ENTER(&call->lock);
1915 MUTEX_ENTER(&conn->conn_data_lock);
1916 conn->flags |= RX_CONN_BUSY;
1917 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1918 if (conn->makeCallWaiters == 0)
1919 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1920 MUTEX_EXIT(&conn->conn_data_lock);
1921 #ifdef RX_ENABLE_LOCKS
1922 CV_BROADCAST(&conn->conn_call_cv);
1927 #ifdef RX_ENABLE_LOCKS
1929 MUTEX_EXIT(&conn->conn_data_lock);
1931 #endif /* RX_ENABLE_LOCKS */
1932 call->state = RX_STATE_DALLY;
1934 error = call->error;
1936 /* currentPacket, nLeft, and NFree must be zeroed here, because
1937 * ResetCall cannot: ResetCall may be called at splnet(), in the
1938 * kernel version, and may interrupt the macros rx_Read or
1939 * rx_Write, which run at normal priority for efficiency. */
1940 if (call->currentPacket) {
1941 queue_Prepend(&call->iovq, call->currentPacket);
1942 call->currentPacket = (struct rx_packet *)0;
1945 call->nLeft = call->nFree = call->curlen = 0;
1947 /* Free any packets from the last call to ReadvProc/WritevProc */
1948 rxi_FreePackets(0, &call->iovq);
1950 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1951 MUTEX_EXIT(&call->lock);
1952 if (conn->type == RX_CLIENT_CONNECTION) {
1953 MUTEX_EXIT(&conn->conn_call_lock);
1954 conn->flags &= ~RX_CONN_BUSY;
1958 * Map errors to the local host's errno.h format.
1960 error = ntoh_syserr_conv(error);
1964 #if !defined(KERNEL)
1966 /* Call this routine when shutting down a server or client (especially
1967 * clients). This will allow Rx to gracefully garbage collect server
1968 * connections, and reduce the number of retries that a server might
1969 * make to a dead client.
1970 * This is not quite right, since some calls may still be ongoing and
1971 * we can't lock them to destroy them. */
1975 register struct rx_connection **conn_ptr, **conn_end;
1979 if (rxinit_status == 1) {
1981 return; /* Already shutdown. */
1983 rxi_DeleteCachedConnections();
1984 if (rx_connHashTable) {
1985 MUTEX_ENTER(&rx_connHashTable_lock);
1986 for (conn_ptr = &rx_connHashTable[0], conn_end =
1987 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
1989 struct rx_connection *conn, *next;
1990 for (conn = *conn_ptr; conn; conn = next) {
1992 if (conn->type == RX_CLIENT_CONNECTION) {
1993 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
1995 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
1996 #ifdef RX_ENABLE_LOCKS
1997 rxi_DestroyConnectionNoLock(conn);
1998 #else /* RX_ENABLE_LOCKS */
1999 rxi_DestroyConnection(conn);
2000 #endif /* RX_ENABLE_LOCKS */
2004 #ifdef RX_ENABLE_LOCKS
2005 while (rx_connCleanup_list) {
2006 struct rx_connection *conn;
2007 conn = rx_connCleanup_list;
2008 rx_connCleanup_list = rx_connCleanup_list->next;
2009 MUTEX_EXIT(&rx_connHashTable_lock);
2010 rxi_CleanupConnection(conn);
2011 MUTEX_ENTER(&rx_connHashTable_lock);
2013 MUTEX_EXIT(&rx_connHashTable_lock);
2014 #endif /* RX_ENABLE_LOCKS */
2023 /* if we wakeup packet waiter too often, can get in loop with two
2024 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2026 rxi_PacketsUnWait(void)
2028 if (!rx_waitingForPackets) {
2032 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2033 return; /* still over quota */
2036 rx_waitingForPackets = 0;
2037 #ifdef RX_ENABLE_LOCKS
2038 CV_BROADCAST(&rx_waitingForPackets_cv);
2040 osi_rxWakeup(&rx_waitingForPackets);
2046 /* ------------------Internal interfaces------------------------- */
2048 /* Return this process's service structure for the
2049 * specified socket and service */
2051 rxi_FindService(register osi_socket socket, register u_short serviceId)
2053 register struct rx_service **sp;
2054 for (sp = &rx_services[0]; *sp; sp++) {
2055 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2061 /* Allocate a call structure, for the indicated channel of the
2062 * supplied connection. The mode and state of the call must be set by
2063 * the caller. Returns the call with mutex locked. */
2065 rxi_NewCall(register struct rx_connection *conn, register int channel)
2067 register struct rx_call *call;
2068 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2069 register struct rx_call *cp; /* Call pointer temp */
2070 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2071 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2073 /* Grab an existing call structure, or allocate a new one.
2074 * Existing call structures are assumed to have been left reset by
2076 MUTEX_ENTER(&rx_freeCallQueue_lock);
2078 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2080 * EXCEPT that the TQ might not yet be cleared out.
2081 * Skip over those with in-use TQs.
2084 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2085 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2091 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2092 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2093 call = queue_First(&rx_freeCallQueue, rx_call);
2094 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2096 MUTEX_ENTER(&rx_stats_mutex);
2097 rx_stats.nFreeCallStructs--;
2098 MUTEX_EXIT(&rx_stats_mutex);
2099 MUTEX_EXIT(&rx_freeCallQueue_lock);
2100 MUTEX_ENTER(&call->lock);
2101 CLEAR_CALL_QUEUE_LOCK(call);
2102 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2103 /* Now, if TQ wasn't cleared earlier, do it now. */
2104 if (call->flags & RX_CALL_TQ_CLEARME) {
2105 rxi_ClearTransmitQueue(call, 0);
2106 queue_Init(&call->tq);
2108 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2109 /* Bind the call to its connection structure */
2111 rxi_ResetCall(call, 1);
2113 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2115 MUTEX_EXIT(&rx_freeCallQueue_lock);
2116 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2117 MUTEX_ENTER(&call->lock);
2118 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2119 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2120 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2122 MUTEX_ENTER(&rx_stats_mutex);
2123 rx_stats.nCallStructs++;
2124 MUTEX_EXIT(&rx_stats_mutex);
2125 /* Initialize once-only items */
2126 queue_Init(&call->tq);
2127 queue_Init(&call->rq);
2128 queue_Init(&call->iovq);
2129 /* Bind the call to its connection structure (prereq for reset) */
2131 rxi_ResetCall(call, 1);
2133 call->channel = channel;
2134 call->callNumber = &conn->callNumber[channel];
2135 /* Note that the next expected call number is retained (in
2136 * conn->callNumber[i]), even if we reallocate the call structure
2138 conn->call[channel] = call;
2139 /* if the channel's never been used (== 0), we should start at 1, otherwise
2140 * the call number is valid from the last time this channel was used */
2141 if (*call->callNumber == 0)
2142 *call->callNumber = 1;
2147 /* A call has been inactive long enough that so we can throw away
2148 * state, including the call structure, which is placed on the call
2150 * Call is locked upon entry.
2151 * haveCTLock set if called from rxi_ReapConnections
2153 #ifdef RX_ENABLE_LOCKS
2155 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2156 #else /* RX_ENABLE_LOCKS */
2158 rxi_FreeCall(register struct rx_call *call)
2159 #endif /* RX_ENABLE_LOCKS */
2161 register int channel = call->channel;
2162 register struct rx_connection *conn = call->conn;
2165 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2166 (*call->callNumber)++;
2167 rxi_ResetCall(call, 0);
2168 call->conn->call[channel] = (struct rx_call *)0;
2170 MUTEX_ENTER(&rx_freeCallQueue_lock);
2171 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2172 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2173 /* A call may be free even though its transmit queue is still in use.
2174 * Since we search the call list from head to tail, put busy calls at
2175 * the head of the list, and idle calls at the tail.
2177 if (call->flags & RX_CALL_TQ_BUSY)
2178 queue_Prepend(&rx_freeCallQueue, call);
2180 queue_Append(&rx_freeCallQueue, call);
2181 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2182 queue_Append(&rx_freeCallQueue, call);
2183 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2184 MUTEX_ENTER(&rx_stats_mutex);
2185 rx_stats.nFreeCallStructs++;
2186 MUTEX_EXIT(&rx_stats_mutex);
2188 MUTEX_EXIT(&rx_freeCallQueue_lock);
2190 /* Destroy the connection if it was previously slated for
2191 * destruction, i.e. the Rx client code previously called
2192 * rx_DestroyConnection (client connections), or
2193 * rxi_ReapConnections called the same routine (server
2194 * connections). Only do this, however, if there are no
2195 * outstanding calls. Note that for fine grain locking, there appears
2196 * to be a deadlock in that rxi_FreeCall has a call locked and
2197 * DestroyConnectionNoLock locks each call in the conn. But note a
2198 * few lines up where we have removed this call from the conn.
2199 * If someone else destroys a connection, they either have no
2200 * call lock held or are going through this section of code.
2202 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2203 MUTEX_ENTER(&conn->conn_data_lock);
2205 MUTEX_EXIT(&conn->conn_data_lock);
2206 #ifdef RX_ENABLE_LOCKS
2208 rxi_DestroyConnectionNoLock(conn);
2210 rxi_DestroyConnection(conn);
2211 #else /* RX_ENABLE_LOCKS */
2212 rxi_DestroyConnection(conn);
2213 #endif /* RX_ENABLE_LOCKS */
2217 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2219 rxi_Alloc(register size_t size)
2223 MUTEX_ENTER(&rx_stats_mutex);
2225 rxi_Allocsize += size;
2226 MUTEX_EXIT(&rx_stats_mutex);
2228 p = (char *)osi_Alloc(size);
2231 osi_Panic("rxi_Alloc error");
2237 rxi_Free(void *addr, register size_t size)
2239 MUTEX_ENTER(&rx_stats_mutex);
2241 rxi_Allocsize -= size;
2242 MUTEX_EXIT(&rx_stats_mutex);
2244 osi_Free(addr, size);
2247 /* Find the peer process represented by the supplied (host,port)
2248 * combination. If there is no appropriate active peer structure, a
2249 * new one will be allocated and initialized
2250 * The origPeer, if set, is a pointer to a peer structure on which the
2251 * refcount will be be decremented. This is used to replace the peer
2252 * structure hanging off a connection structure */
2254 rxi_FindPeer(register afs_uint32 host, register u_short port,
2255 struct rx_peer *origPeer, int create)
2257 register struct rx_peer *pp;
2259 hashIndex = PEER_HASH(host, port);
2260 MUTEX_ENTER(&rx_peerHashTable_lock);
2261 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2262 if ((pp->host == host) && (pp->port == port))
2267 pp = rxi_AllocPeer(); /* This bzero's *pp */
2268 pp->host = host; /* set here or in InitPeerParams is zero */
2270 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2271 queue_Init(&pp->congestionQueue);
2272 queue_Init(&pp->rpcStats);
2273 pp->next = rx_peerHashTable[hashIndex];
2274 rx_peerHashTable[hashIndex] = pp;
2275 rxi_InitPeerParams(pp);
2276 MUTEX_ENTER(&rx_stats_mutex);
2277 rx_stats.nPeerStructs++;
2278 MUTEX_EXIT(&rx_stats_mutex);
2285 origPeer->refCount--;
2286 MUTEX_EXIT(&rx_peerHashTable_lock);
2291 /* Find the connection at (host, port) started at epoch, and with the
2292 * given connection id. Creates the server connection if necessary.
2293 * The type specifies whether a client connection or a server
2294 * connection is desired. In both cases, (host, port) specify the
2295 * peer's (host, pair) pair. Client connections are not made
2296 * automatically by this routine. The parameter socket gives the
2297 * socket descriptor on which the packet was received. This is used,
2298 * in the case of server connections, to check that *new* connections
2299 * come via a valid (port, serviceId). Finally, the securityIndex
2300 * parameter must match the existing index for the connection. If a
2301 * server connection is created, it will be created using the supplied
2302 * index, if the index is valid for this service */
2303 struct rx_connection *
2304 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2305 register u_short port, u_short serviceId, afs_uint32 cid,
2306 afs_uint32 epoch, int type, u_int securityIndex)
2308 int hashindex, flag;
2309 register struct rx_connection *conn;
2310 hashindex = CONN_HASH(host, port, cid, epoch, type);
2311 MUTEX_ENTER(&rx_connHashTable_lock);
2312 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2313 rx_connHashTable[hashindex],
2316 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2317 && (epoch == conn->epoch)) {
2318 register struct rx_peer *pp = conn->peer;
2319 if (securityIndex != conn->securityIndex) {
2320 /* this isn't supposed to happen, but someone could forge a packet
2321 * like this, and there seems to be some CM bug that makes this
2322 * happen from time to time -- in which case, the fileserver
2324 MUTEX_EXIT(&rx_connHashTable_lock);
2325 return (struct rx_connection *)0;
2327 if (pp->host == host && pp->port == port)
2329 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2331 /* So what happens when it's a callback connection? */
2332 if ( /*type == RX_CLIENT_CONNECTION && */
2333 (conn->epoch & 0x80000000))
2337 /* the connection rxLastConn that was used the last time is not the
2338 ** one we are looking for now. Hence, start searching in the hash */
2340 conn = rx_connHashTable[hashindex];
2345 struct rx_service *service;
2346 if (type == RX_CLIENT_CONNECTION) {
2347 MUTEX_EXIT(&rx_connHashTable_lock);
2348 return (struct rx_connection *)0;
2350 service = rxi_FindService(socket, serviceId);
2351 if (!service || (securityIndex >= service->nSecurityObjects)
2352 || (service->securityObjects[securityIndex] == 0)) {
2353 MUTEX_EXIT(&rx_connHashTable_lock);
2354 return (struct rx_connection *)0;
2356 conn = rxi_AllocConnection(); /* This bzero's the connection */
2357 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2358 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2359 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2360 conn->next = rx_connHashTable[hashindex];
2361 rx_connHashTable[hashindex] = conn;
2362 conn->peer = rxi_FindPeer(host, port, 0, 1);
2363 conn->type = RX_SERVER_CONNECTION;
2364 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2365 conn->epoch = epoch;
2366 conn->cid = cid & RX_CIDMASK;
2367 /* conn->serial = conn->lastSerial = 0; */
2368 /* conn->timeout = 0; */
2369 conn->ackRate = RX_FAST_ACK_RATE;
2370 conn->service = service;
2371 conn->serviceId = serviceId;
2372 conn->securityIndex = securityIndex;
2373 conn->securityObject = service->securityObjects[securityIndex];
2374 conn->nSpecific = 0;
2375 conn->specific = NULL;
2376 rx_SetConnDeadTime(conn, service->connDeadTime);
2377 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2378 /* Notify security object of the new connection */
2379 RXS_NewConnection(conn->securityObject, conn);
2380 /* XXXX Connection timeout? */
2381 if (service->newConnProc)
2382 (*service->newConnProc) (conn);
2383 MUTEX_ENTER(&rx_stats_mutex);
2384 rx_stats.nServerConns++;
2385 MUTEX_EXIT(&rx_stats_mutex);
2388 MUTEX_ENTER(&conn->conn_data_lock);
2390 MUTEX_EXIT(&conn->conn_data_lock);
2392 rxLastConn = conn; /* store this connection as the last conn used */
2393 MUTEX_EXIT(&rx_connHashTable_lock);
2397 /* There are two packet tracing routines available for testing and monitoring
2398 * Rx. One is called just after every packet is received and the other is
2399 * called just before every packet is sent. Received packets, have had their
2400 * headers decoded, and packets to be sent have not yet had their headers
2401 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2402 * containing the network address. Both can be modified. The return value, if
2403 * non-zero, indicates that the packet should be dropped. */
2405 int (*rx_justReceived) () = 0;
2406 int (*rx_almostSent) () = 0;
2408 /* A packet has been received off the interface. Np is the packet, socket is
2409 * the socket number it was received from (useful in determining which service
2410 * this packet corresponds to), and (host, port) reflect the host,port of the
2411 * sender. This call returns the packet to the caller if it is finished with
2412 * it, rather than de-allocating it, just as a small performance hack */
2415 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2416 afs_uint32 host, u_short port, int *tnop,
2417 struct rx_call **newcallp)
2419 register struct rx_call *call;
2420 register struct rx_connection *conn;
2422 afs_uint32 currentCallNumber;
2428 struct rx_packet *tnp;
2431 /* We don't print out the packet until now because (1) the time may not be
2432 * accurate enough until now in the lwp implementation (rx_Listener only gets
2433 * the time after the packet is read) and (2) from a protocol point of view,
2434 * this is the first time the packet has been seen */
2435 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2436 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2437 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2438 np->header.serial, packetType, host, port, np->header.serviceId,
2439 np->header.epoch, np->header.cid, np->header.callNumber,
2440 np->header.seq, np->header.flags, np));
2443 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2444 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2447 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2448 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2451 /* If an input tracer function is defined, call it with the packet and
2452 * network address. Note this function may modify its arguments. */
2453 if (rx_justReceived) {
2454 struct sockaddr_in addr;
2456 addr.sin_family = AF_INET;
2457 addr.sin_port = port;
2458 addr.sin_addr.s_addr = host;
2459 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2460 addr.sin_len = sizeof(addr);
2461 #endif /* AFS_OSF_ENV */
2462 drop = (*rx_justReceived) (np, &addr);
2463 /* drop packet if return value is non-zero */
2466 port = addr.sin_port; /* in case fcn changed addr */
2467 host = addr.sin_addr.s_addr;
2471 /* If packet was not sent by the client, then *we* must be the client */
2472 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2473 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2475 /* Find the connection (or fabricate one, if we're the server & if
2476 * necessary) associated with this packet */
2478 rxi_FindConnection(socket, host, port, np->header.serviceId,
2479 np->header.cid, np->header.epoch, type,
2480 np->header.securityIndex);
2483 /* If no connection found or fabricated, just ignore the packet.
2484 * (An argument could be made for sending an abort packet for
2489 MUTEX_ENTER(&conn->conn_data_lock);
2490 if (conn->maxSerial < np->header.serial)
2491 conn->maxSerial = np->header.serial;
2492 MUTEX_EXIT(&conn->conn_data_lock);
2494 /* If the connection is in an error state, send an abort packet and ignore
2495 * the incoming packet */
2497 /* Don't respond to an abort packet--we don't want loops! */
2498 MUTEX_ENTER(&conn->conn_data_lock);
2499 if (np->header.type != RX_PACKET_TYPE_ABORT)
2500 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2502 MUTEX_EXIT(&conn->conn_data_lock);
2506 /* Check for connection-only requests (i.e. not call specific). */
2507 if (np->header.callNumber == 0) {
2508 switch (np->header.type) {
2509 case RX_PACKET_TYPE_ABORT:
2510 /* What if the supplied error is zero? */
2511 rxi_ConnectionError(conn, ntohl(rx_GetInt32(np, 0)));
2512 MUTEX_ENTER(&conn->conn_data_lock);
2514 MUTEX_EXIT(&conn->conn_data_lock);
2516 case RX_PACKET_TYPE_CHALLENGE:
2517 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2518 MUTEX_ENTER(&conn->conn_data_lock);
2520 MUTEX_EXIT(&conn->conn_data_lock);
2522 case RX_PACKET_TYPE_RESPONSE:
2523 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2524 MUTEX_ENTER(&conn->conn_data_lock);
2526 MUTEX_EXIT(&conn->conn_data_lock);
2528 case RX_PACKET_TYPE_PARAMS:
2529 case RX_PACKET_TYPE_PARAMS + 1:
2530 case RX_PACKET_TYPE_PARAMS + 2:
2531 /* ignore these packet types for now */
2532 MUTEX_ENTER(&conn->conn_data_lock);
2534 MUTEX_EXIT(&conn->conn_data_lock);
2539 /* Should not reach here, unless the peer is broken: send an
2541 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2542 MUTEX_ENTER(&conn->conn_data_lock);
2543 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2545 MUTEX_EXIT(&conn->conn_data_lock);
2550 channel = np->header.cid & RX_CHANNELMASK;
2551 call = conn->call[channel];
2552 #ifdef RX_ENABLE_LOCKS
2554 MUTEX_ENTER(&call->lock);
2555 /* Test to see if call struct is still attached to conn. */
2556 if (call != conn->call[channel]) {
2558 MUTEX_EXIT(&call->lock);
2559 if (type == RX_SERVER_CONNECTION) {
2560 call = conn->call[channel];
2561 /* If we started with no call attached and there is one now,
2562 * another thread is also running this routine and has gotten
2563 * the connection channel. We should drop this packet in the tests
2564 * below. If there was a call on this connection and it's now
2565 * gone, then we'll be making a new call below.
2566 * If there was previously a call and it's now different then
2567 * the old call was freed and another thread running this routine
2568 * has created a call on this channel. One of these two threads
2569 * has a packet for the old call and the code below handles those
2573 MUTEX_ENTER(&call->lock);
2575 /* This packet can't be for this call. If the new call address is
2576 * 0 then no call is running on this channel. If there is a call
2577 * then, since this is a client connection we're getting data for
2578 * it must be for the previous call.
2580 MUTEX_ENTER(&rx_stats_mutex);
2581 rx_stats.spuriousPacketsRead++;
2582 MUTEX_EXIT(&rx_stats_mutex);
2583 MUTEX_ENTER(&conn->conn_data_lock);
2585 MUTEX_EXIT(&conn->conn_data_lock);
2590 currentCallNumber = conn->callNumber[channel];
2592 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2593 if (np->header.callNumber < currentCallNumber) {
2594 MUTEX_ENTER(&rx_stats_mutex);
2595 rx_stats.spuriousPacketsRead++;
2596 MUTEX_EXIT(&rx_stats_mutex);
2597 #ifdef RX_ENABLE_LOCKS
2599 MUTEX_EXIT(&call->lock);
2601 MUTEX_ENTER(&conn->conn_data_lock);
2603 MUTEX_EXIT(&conn->conn_data_lock);
2607 MUTEX_ENTER(&conn->conn_call_lock);
2608 call = rxi_NewCall(conn, channel);
2609 MUTEX_EXIT(&conn->conn_call_lock);
2610 *call->callNumber = np->header.callNumber;
2611 if (np->header.callNumber == 0)
2612 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], conn->peer->host, 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));
2614 call->state = RX_STATE_PRECALL;
2615 clock_GetTime(&call->queueTime);
2616 hzero(call->bytesSent);
2617 hzero(call->bytesRcvd);
2619 * If the number of queued calls exceeds the overload
2620 * threshold then abort this call.
2622 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2623 struct rx_packet *tp;
2625 rxi_CallError(call, rx_BusyError);
2626 tp = rxi_SendCallAbort(call, np, 1, 0);
2627 MUTEX_EXIT(&call->lock);
2628 MUTEX_ENTER(&conn->conn_data_lock);
2630 MUTEX_EXIT(&conn->conn_data_lock);
2631 MUTEX_ENTER(&rx_stats_mutex);
2633 MUTEX_EXIT(&rx_stats_mutex);
2636 rxi_KeepAliveOn(call);
2637 } else if (np->header.callNumber != currentCallNumber) {
2638 /* Wait until the transmit queue is idle before deciding
2639 * whether to reset the current call. Chances are that the
2640 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2643 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2644 while ((call->state == RX_STATE_ACTIVE)
2645 && (call->flags & RX_CALL_TQ_BUSY)) {
2646 call->flags |= RX_CALL_TQ_WAIT;
2648 #ifdef RX_ENABLE_LOCKS
2649 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2650 CV_WAIT(&call->cv_tq, &call->lock);
2651 #else /* RX_ENABLE_LOCKS */
2652 osi_rxSleep(&call->tq);
2653 #endif /* RX_ENABLE_LOCKS */
2655 if (call->tqWaiters == 0)
2656 call->flags &= ~RX_CALL_TQ_WAIT;
2658 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2659 /* If the new call cannot be taken right now send a busy and set
2660 * the error condition in this call, so that it terminates as
2661 * quickly as possible */
2662 if (call->state == RX_STATE_ACTIVE) {
2663 struct rx_packet *tp;
2665 rxi_CallError(call, RX_CALL_DEAD);
2666 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2668 MUTEX_EXIT(&call->lock);
2669 MUTEX_ENTER(&conn->conn_data_lock);
2671 MUTEX_EXIT(&conn->conn_data_lock);
2674 rxi_ResetCall(call, 0);
2675 *call->callNumber = np->header.callNumber;
2676 if (np->header.callNumber == 0)
2677 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], conn->peer->host, 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));
2679 call->state = RX_STATE_PRECALL;
2680 clock_GetTime(&call->queueTime);
2681 hzero(call->bytesSent);
2682 hzero(call->bytesRcvd);
2684 * If the number of queued calls exceeds the overload
2685 * threshold then abort this call.
2687 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2688 struct rx_packet *tp;
2690 rxi_CallError(call, rx_BusyError);
2691 tp = rxi_SendCallAbort(call, np, 1, 0);
2692 MUTEX_EXIT(&call->lock);
2693 MUTEX_ENTER(&conn->conn_data_lock);
2695 MUTEX_EXIT(&conn->conn_data_lock);
2696 MUTEX_ENTER(&rx_stats_mutex);
2698 MUTEX_EXIT(&rx_stats_mutex);
2701 rxi_KeepAliveOn(call);
2703 /* Continuing call; do nothing here. */
2705 } else { /* we're the client */
2706 /* Ignore all incoming acknowledgements for calls in DALLY state */
2707 if (call && (call->state == RX_STATE_DALLY)
2708 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2709 MUTEX_ENTER(&rx_stats_mutex);
2710 rx_stats.ignorePacketDally++;
2711 MUTEX_EXIT(&rx_stats_mutex);
2712 #ifdef RX_ENABLE_LOCKS
2714 MUTEX_EXIT(&call->lock);
2717 MUTEX_ENTER(&conn->conn_data_lock);
2719 MUTEX_EXIT(&conn->conn_data_lock);
2723 /* Ignore anything that's not relevant to the current call. If there
2724 * isn't a current call, then no packet is relevant. */
2725 if (!call || (np->header.callNumber != currentCallNumber)) {
2726 MUTEX_ENTER(&rx_stats_mutex);
2727 rx_stats.spuriousPacketsRead++;
2728 MUTEX_EXIT(&rx_stats_mutex);
2729 #ifdef RX_ENABLE_LOCKS
2731 MUTEX_EXIT(&call->lock);
2734 MUTEX_ENTER(&conn->conn_data_lock);
2736 MUTEX_EXIT(&conn->conn_data_lock);
2739 /* If the service security object index stamped in the packet does not
2740 * match the connection's security index, ignore the packet */
2741 if (np->header.securityIndex != conn->securityIndex) {
2742 #ifdef RX_ENABLE_LOCKS
2743 MUTEX_EXIT(&call->lock);
2745 MUTEX_ENTER(&conn->conn_data_lock);
2747 MUTEX_EXIT(&conn->conn_data_lock);
2751 /* If we're receiving the response, then all transmit packets are
2752 * implicitly acknowledged. Get rid of them. */
2753 if (np->header.type == RX_PACKET_TYPE_DATA) {
2754 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2755 /* XXX Hack. Because we must release the global rx lock when
2756 * sending packets (osi_NetSend) we drop all acks while we're
2757 * traversing the tq in rxi_Start sending packets out because
2758 * packets may move to the freePacketQueue as result of being here!
2759 * So we drop these packets until we're safely out of the
2760 * traversing. Really ugly!
2761 * For fine grain RX locking, we set the acked field in the
2762 * packets and let rxi_Start remove them from the transmit queue.
2764 if (call->flags & RX_CALL_TQ_BUSY) {
2765 #ifdef RX_ENABLE_LOCKS
2766 rxi_SetAcksInTransmitQueue(call);
2769 return np; /* xmitting; drop packet */
2772 rxi_ClearTransmitQueue(call, 0);
2774 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2775 rxi_ClearTransmitQueue(call, 0);
2776 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2778 if (np->header.type == RX_PACKET_TYPE_ACK) {
2779 /* now check to see if this is an ack packet acknowledging that the
2780 * server actually *lost* some hard-acked data. If this happens we
2781 * ignore this packet, as it may indicate that the server restarted in
2782 * the middle of a call. It is also possible that this is an old ack
2783 * packet. We don't abort the connection in this case, because this
2784 * *might* just be an old ack packet. The right way to detect a server
2785 * restart in the midst of a call is to notice that the server epoch
2787 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2788 * XXX unacknowledged. I think that this is off-by-one, but
2789 * XXX I don't dare change it just yet, since it will
2790 * XXX interact badly with the server-restart detection
2791 * XXX code in receiveackpacket. */
2792 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2793 MUTEX_ENTER(&rx_stats_mutex);
2794 rx_stats.spuriousPacketsRead++;
2795 MUTEX_EXIT(&rx_stats_mutex);
2796 MUTEX_EXIT(&call->lock);
2797 MUTEX_ENTER(&conn->conn_data_lock);
2799 MUTEX_EXIT(&conn->conn_data_lock);
2803 } /* else not a data packet */
2806 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2807 /* Set remote user defined status from packet */
2808 call->remoteStatus = np->header.userStatus;
2810 /* Note the gap between the expected next packet and the actual
2811 * packet that arrived, when the new packet has a smaller serial number
2812 * than expected. Rioses frequently reorder packets all by themselves,
2813 * so this will be quite important with very large window sizes.
2814 * Skew is checked against 0 here to avoid any dependence on the type of
2815 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2817 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2818 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2819 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2821 MUTEX_ENTER(&conn->conn_data_lock);
2822 skew = conn->lastSerial - np->header.serial;
2823 conn->lastSerial = np->header.serial;
2824 MUTEX_EXIT(&conn->conn_data_lock);
2826 register struct rx_peer *peer;
2828 if (skew > peer->inPacketSkew) {
2829 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2831 peer->inPacketSkew = skew;
2835 /* Now do packet type-specific processing */
2836 switch (np->header.type) {
2837 case RX_PACKET_TYPE_DATA:
2838 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2841 case RX_PACKET_TYPE_ACK:
2842 /* Respond immediately to ack packets requesting acknowledgement
2844 if (np->header.flags & RX_REQUEST_ACK) {
2846 (void)rxi_SendCallAbort(call, 0, 1, 0);
2848 (void)rxi_SendAck(call, 0, np->header.serial,
2849 RX_ACK_PING_RESPONSE, 1);
2851 np = rxi_ReceiveAckPacket(call, np, 1);
2853 case RX_PACKET_TYPE_ABORT:
2854 /* An abort packet: reset the connection, passing the error up to
2856 /* What if error is zero? */
2857 rxi_CallError(call, ntohl(*(afs_int32 *) rx_DataOf(np)));
2859 case RX_PACKET_TYPE_BUSY:
2862 case RX_PACKET_TYPE_ACKALL:
2863 /* All packets acknowledged, so we can drop all packets previously
2864 * readied for sending */
2865 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2866 /* XXX Hack. We because we can't release the global rx lock when
2867 * sending packets (osi_NetSend) we drop all ack pkts while we're
2868 * traversing the tq in rxi_Start sending packets out because
2869 * packets may move to the freePacketQueue as result of being
2870 * here! So we drop these packets until we're safely out of the
2871 * traversing. Really ugly!
2872 * For fine grain RX locking, we set the acked field in the packets
2873 * and let rxi_Start remove the packets from the transmit queue.
2875 if (call->flags & RX_CALL_TQ_BUSY) {
2876 #ifdef RX_ENABLE_LOCKS
2877 rxi_SetAcksInTransmitQueue(call);
2879 #else /* RX_ENABLE_LOCKS */
2881 return np; /* xmitting; drop packet */
2882 #endif /* RX_ENABLE_LOCKS */
2884 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2885 rxi_ClearTransmitQueue(call, 0);
2888 /* Should not reach here, unless the peer is broken: send an abort
2890 rxi_CallError(call, RX_PROTOCOL_ERROR);
2891 np = rxi_SendCallAbort(call, np, 1, 0);
2894 /* Note when this last legitimate packet was received, for keep-alive
2895 * processing. Note, we delay getting the time until now in the hope that
2896 * the packet will be delivered to the user before any get time is required
2897 * (if not, then the time won't actually be re-evaluated here). */
2898 call->lastReceiveTime = clock_Sec();
2899 MUTEX_EXIT(&call->lock);
2900 MUTEX_ENTER(&conn->conn_data_lock);
2902 MUTEX_EXIT(&conn->conn_data_lock);
2906 /* return true if this is an "interesting" connection from the point of view
2907 of someone trying to debug the system */
2909 rxi_IsConnInteresting(struct rx_connection *aconn)
2912 register struct rx_call *tcall;
2914 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2916 for (i = 0; i < RX_MAXCALLS; i++) {
2917 tcall = aconn->call[i];
2919 if ((tcall->state == RX_STATE_PRECALL)
2920 || (tcall->state == RX_STATE_ACTIVE))
2922 if ((tcall->mode == RX_MODE_SENDING)
2923 || (tcall->mode == RX_MODE_RECEIVING))
2931 /* if this is one of the last few packets AND it wouldn't be used by the
2932 receiving call to immediately satisfy a read request, then drop it on
2933 the floor, since accepting it might prevent a lock-holding thread from
2934 making progress in its reading. If a call has been cleared while in
2935 the precall state then ignore all subsequent packets until the call
2936 is assigned to a thread. */
2939 TooLow(struct rx_packet *ap, struct rx_call *acall)
2942 MUTEX_ENTER(&rx_stats_mutex);
2943 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2944 && (acall->state == RX_STATE_PRECALL))
2945 || ((rx_nFreePackets < rxi_dataQuota + 2)
2946 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2947 && (acall->flags & RX_CALL_READER_WAIT)))) {
2950 MUTEX_EXIT(&rx_stats_mutex);
2956 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2957 struct rx_call *acall)
2959 struct rx_call *call = acall;
2963 MUTEX_ENTER(&conn->conn_data_lock);
2964 conn->checkReachEvent = NULL;
2965 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2968 MUTEX_EXIT(&conn->conn_data_lock);
2972 MUTEX_ENTER(&conn->conn_call_lock);
2973 MUTEX_ENTER(&conn->conn_data_lock);
2974 for (i = 0; i < RX_MAXCALLS; i++) {
2975 struct rx_call *tc = conn->call[i];
2976 if (tc && tc->state == RX_STATE_PRECALL) {
2982 /* Indicate that rxi_CheckReachEvent is no longer running by
2983 * clearing the flag. Must be atomic under conn_data_lock to
2984 * avoid a new call slipping by: rxi_CheckConnReach holds
2985 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
2987 conn->flags &= ~RX_CONN_ATTACHWAIT;
2988 MUTEX_EXIT(&conn->conn_data_lock);
2989 MUTEX_EXIT(&conn->conn_call_lock);
2994 MUTEX_ENTER(&call->lock);
2995 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
2997 MUTEX_EXIT(&call->lock);
2999 clock_GetTime(&when);
3000 when.sec += RX_CHECKREACH_TIMEOUT;
3001 MUTEX_ENTER(&conn->conn_data_lock);
3002 if (!conn->checkReachEvent) {
3004 conn->checkReachEvent =
3005 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
3007 MUTEX_EXIT(&conn->conn_data_lock);
3013 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3015 struct rx_service *service = conn->service;
3016 struct rx_peer *peer = conn->peer;
3017 afs_uint32 now, lastReach;
3019 if (service->checkReach == 0)
3023 MUTEX_ENTER(&peer->peer_lock);
3024 lastReach = peer->lastReachTime;
3025 MUTEX_EXIT(&peer->peer_lock);
3026 if (now - lastReach < RX_CHECKREACH_TTL)
3029 MUTEX_ENTER(&conn->conn_data_lock);
3030 if (conn->flags & RX_CONN_ATTACHWAIT) {
3031 MUTEX_EXIT(&conn->conn_data_lock);
3034 conn->flags |= RX_CONN_ATTACHWAIT;
3035 MUTEX_EXIT(&conn->conn_data_lock);
3036 if (!conn->checkReachEvent)
3037 rxi_CheckReachEvent(NULL, conn, call);
3042 /* try to attach call, if authentication is complete */
3044 TryAttach(register struct rx_call *acall, register osi_socket socket,
3045 register int *tnop, register struct rx_call **newcallp,
3048 struct rx_connection *conn = acall->conn;
3050 if (conn->type == RX_SERVER_CONNECTION
3051 && acall->state == RX_STATE_PRECALL) {
3052 /* Don't attach until we have any req'd. authentication. */
3053 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3054 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3055 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3056 /* Note: this does not necessarily succeed; there
3057 * may not any proc available
3060 rxi_ChallengeOn(acall->conn);
3065 /* A data packet has been received off the interface. This packet is
3066 * appropriate to the call (the call is in the right state, etc.). This
3067 * routine can return a packet to the caller, for re-use */
3070 rxi_ReceiveDataPacket(register struct rx_call *call,
3071 register struct rx_packet *np, int istack,
3072 osi_socket socket, afs_uint32 host, u_short port,
3073 int *tnop, struct rx_call **newcallp)
3075 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3079 afs_uint32 seq, serial, flags;
3081 struct rx_packet *tnp;
3083 MUTEX_ENTER(&rx_stats_mutex);
3084 rx_stats.dataPacketsRead++;
3085 MUTEX_EXIT(&rx_stats_mutex);
3088 /* If there are no packet buffers, drop this new packet, unless we can find
3089 * packet buffers from inactive calls */
3091 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3092 MUTEX_ENTER(&rx_freePktQ_lock);
3093 rxi_NeedMorePackets = TRUE;
3094 MUTEX_EXIT(&rx_freePktQ_lock);
3095 MUTEX_ENTER(&rx_stats_mutex);
3096 rx_stats.noPacketBuffersOnRead++;
3097 MUTEX_EXIT(&rx_stats_mutex);
3098 call->rprev = np->header.serial;
3099 rxi_calltrace(RX_TRACE_DROP, call);
3100 dpf(("packet %x dropped on receipt - quota problems", np));
3102 rxi_ClearReceiveQueue(call);
3103 clock_GetTime(&when);
3104 clock_Add(&when, &rx_softAckDelay);
3105 if (!call->delayedAckEvent
3106 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3107 rxevent_Cancel(call->delayedAckEvent, call,
3108 RX_CALL_REFCOUNT_DELAY);
3109 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3110 call->delayedAckEvent =
3111 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3113 /* we've damaged this call already, might as well do it in. */
3119 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3120 * packet is one of several packets transmitted as a single
3121 * datagram. Do not send any soft or hard acks until all packets
3122 * in a jumbogram have been processed. Send negative acks right away.
3124 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3125 /* tnp is non-null when there are more packets in the
3126 * current jumbo gram */
3133 seq = np->header.seq;
3134 serial = np->header.serial;
3135 flags = np->header.flags;
3137 /* If the call is in an error state, send an abort message */
3139 return rxi_SendCallAbort(call, np, istack, 0);
3141 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3142 * AFS 3.5 jumbogram. */
3143 if (flags & RX_JUMBO_PACKET) {
3144 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3149 if (np->header.spare != 0) {
3150 MUTEX_ENTER(&call->conn->conn_data_lock);
3151 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3152 MUTEX_EXIT(&call->conn->conn_data_lock);
3155 /* The usual case is that this is the expected next packet */
3156 if (seq == call->rnext) {
3158 /* Check to make sure it is not a duplicate of one already queued */
3159 if (queue_IsNotEmpty(&call->rq)
3160 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3161 MUTEX_ENTER(&rx_stats_mutex);
3162 rx_stats.dupPacketsRead++;
3163 MUTEX_EXIT(&rx_stats_mutex);
3164 dpf(("packet %x dropped on receipt - duplicate", np));
3165 rxevent_Cancel(call->delayedAckEvent, call,
3166 RX_CALL_REFCOUNT_DELAY);
3167 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3173 /* It's the next packet. Stick it on the receive queue
3174 * for this call. Set newPackets to make sure we wake
3175 * the reader once all packets have been processed */
3176 queue_Prepend(&call->rq, np);
3178 np = NULL; /* We can't use this anymore */
3181 /* If an ack is requested then set a flag to make sure we
3182 * send an acknowledgement for this packet */
3183 if (flags & RX_REQUEST_ACK) {
3184 ackNeeded = RX_ACK_REQUESTED;
3187 /* Keep track of whether we have received the last packet */
3188 if (flags & RX_LAST_PACKET) {
3189 call->flags |= RX_CALL_HAVE_LAST;
3193 /* Check whether we have all of the packets for this call */
3194 if (call->flags & RX_CALL_HAVE_LAST) {
3195 afs_uint32 tseq; /* temporary sequence number */
3196 struct rx_packet *tp; /* Temporary packet pointer */
3197 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3199 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3200 if (tseq != tp->header.seq)
3202 if (tp->header.flags & RX_LAST_PACKET) {
3203 call->flags |= RX_CALL_RECEIVE_DONE;
3210 /* Provide asynchronous notification for those who want it
3211 * (e.g. multi rx) */
3212 if (call->arrivalProc) {
3213 (*call->arrivalProc) (call, call->arrivalProcHandle,
3214 call->arrivalProcArg);
3215 call->arrivalProc = (void (*)())0;
3218 /* Update last packet received */
3221 /* If there is no server process serving this call, grab
3222 * one, if available. We only need to do this once. If a
3223 * server thread is available, this thread becomes a server
3224 * thread and the server thread becomes a listener thread. */
3226 TryAttach(call, socket, tnop, newcallp, 0);
3229 /* This is not the expected next packet. */
3231 /* Determine whether this is a new or old packet, and if it's
3232 * a new one, whether it fits into the current receive window.
3233 * Also figure out whether the packet was delivered in sequence.
3234 * We use the prev variable to determine whether the new packet
3235 * is the successor of its immediate predecessor in the
3236 * receive queue, and the missing flag to determine whether
3237 * any of this packets predecessors are missing. */
3239 afs_uint32 prev; /* "Previous packet" sequence number */
3240 struct rx_packet *tp; /* Temporary packet pointer */
3241 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3242 int missing; /* Are any predecessors missing? */
3244 /* If the new packet's sequence number has been sent to the
3245 * application already, then this is a duplicate */
3246 if (seq < call->rnext) {
3247 MUTEX_ENTER(&rx_stats_mutex);
3248 rx_stats.dupPacketsRead++;
3249 MUTEX_EXIT(&rx_stats_mutex);
3250 rxevent_Cancel(call->delayedAckEvent, call,
3251 RX_CALL_REFCOUNT_DELAY);
3252 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3258 /* If the sequence number is greater than what can be
3259 * accomodated by the current window, then send a negative
3260 * acknowledge and drop the packet */
3261 if ((call->rnext + call->rwind) <= seq) {
3262 rxevent_Cancel(call->delayedAckEvent, call,
3263 RX_CALL_REFCOUNT_DELAY);
3264 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3271 /* Look for the packet in the queue of old received packets */
3272 for (prev = call->rnext - 1, missing =
3273 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3274 /*Check for duplicate packet */
3275 if (seq == tp->header.seq) {
3276 MUTEX_ENTER(&rx_stats_mutex);
3277 rx_stats.dupPacketsRead++;
3278 MUTEX_EXIT(&rx_stats_mutex);
3279 rxevent_Cancel(call->delayedAckEvent, call,
3280 RX_CALL_REFCOUNT_DELAY);
3281 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3287 /* If we find a higher sequence packet, break out and
3288 * insert the new packet here. */
3289 if (seq < tp->header.seq)
3291 /* Check for missing packet */
3292 if (tp->header.seq != prev + 1) {
3296 prev = tp->header.seq;
3299 /* Keep track of whether we have received the last packet. */
3300 if (flags & RX_LAST_PACKET) {
3301 call->flags |= RX_CALL_HAVE_LAST;
3304 /* It's within the window: add it to the the receive queue.
3305 * tp is left by the previous loop either pointing at the
3306 * packet before which to insert the new packet, or at the
3307 * queue head if the queue is empty or the packet should be
3309 queue_InsertBefore(tp, np);
3313 /* Check whether we have all of the packets for this call */
3314 if ((call->flags & RX_CALL_HAVE_LAST)
3315 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3316 afs_uint32 tseq; /* temporary sequence number */
3319 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3320 if (tseq != tp->header.seq)
3322 if (tp->header.flags & RX_LAST_PACKET) {
3323 call->flags |= RX_CALL_RECEIVE_DONE;
3330 /* We need to send an ack of the packet is out of sequence,
3331 * or if an ack was requested by the peer. */
3332 if (seq != prev + 1 || missing || (flags & RX_REQUEST_ACK)) {
3333 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3336 /* Acknowledge the last packet for each call */
3337 if (flags & RX_LAST_PACKET) {
3348 * If the receiver is waiting for an iovec, fill the iovec
3349 * using the data from the receive queue */
3350 if (call->flags & RX_CALL_IOVEC_WAIT) {
3351 didHardAck = rxi_FillReadVec(call, serial);
3352 /* the call may have been aborted */
3361 /* Wakeup the reader if any */
3362 if ((call->flags & RX_CALL_READER_WAIT)
3363 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3364 || (call->iovNext >= call->iovMax)
3365 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3366 call->flags &= ~RX_CALL_READER_WAIT;
3367 #ifdef RX_ENABLE_LOCKS
3368 CV_BROADCAST(&call->cv_rq);
3370 osi_rxWakeup(&call->rq);
3376 * Send an ack when requested by the peer, or once every
3377 * rxi_SoftAckRate packets until the last packet has been
3378 * received. Always send a soft ack for the last packet in
3379 * the server's reply. */
3381 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3382 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3383 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3384 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3385 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3386 } else if (call->nSoftAcks) {
3387 clock_GetTime(&when);
3388 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3389 clock_Add(&when, &rx_lastAckDelay);
3391 clock_Add(&when, &rx_softAckDelay);
3393 if (!call->delayedAckEvent
3394 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3395 rxevent_Cancel(call->delayedAckEvent, call,
3396 RX_CALL_REFCOUNT_DELAY);
3397 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3398 call->delayedAckEvent =
3399 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3401 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3402 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3409 static void rxi_ComputeRate();
3413 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3415 struct rx_peer *peer = conn->peer;
3417 MUTEX_ENTER(&peer->peer_lock);
3418 peer->lastReachTime = clock_Sec();
3419 MUTEX_EXIT(&peer->peer_lock);
3421 MUTEX_ENTER(&conn->conn_data_lock);
3422 if (conn->flags & RX_CONN_ATTACHWAIT) {
3425 conn->flags &= ~RX_CONN_ATTACHWAIT;
3426 MUTEX_EXIT(&conn->conn_data_lock);
3428 for (i = 0; i < RX_MAXCALLS; i++) {
3429 struct rx_call *call = conn->call[i];
3432 MUTEX_ENTER(&call->lock);
3433 /* tnop can be null if newcallp is null */
3434 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3436 MUTEX_EXIT(&call->lock);
3440 MUTEX_EXIT(&conn->conn_data_lock);
3443 /* rxi_ComputePeerNetStats
3445 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3446 * estimates (like RTT and throughput) based on ack packets. Caller
3447 * must ensure that the packet in question is the right one (i.e.
3448 * serial number matches).
3451 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3452 struct rx_ackPacket *ap, struct rx_packet *np)
3454 struct rx_peer *peer = call->conn->peer;
3456 /* Use RTT if not delayed by client. */
3457 if (ap->reason != RX_ACK_DELAY)
3458 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3460 rxi_ComputeRate(peer, call, p, np, ap->reason);
3464 /* The real smarts of the whole thing. */
3466 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3469 struct rx_ackPacket *ap;
3471 register struct rx_packet *tp;
3472 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3473 register struct rx_connection *conn = call->conn;
3474 struct rx_peer *peer = conn->peer;
3477 /* because there are CM's that are bogus, sending weird values for this. */
3478 afs_uint32 skew = 0;
3483 int newAckCount = 0;
3484 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3485 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3487 MUTEX_ENTER(&rx_stats_mutex);
3488 rx_stats.ackPacketsRead++;
3489 MUTEX_EXIT(&rx_stats_mutex);
3490 ap = (struct rx_ackPacket *)rx_DataOf(np);
3491 nbytes = rx_Contiguous(np) - ((ap->acks) - (u_char *) ap);
3493 return np; /* truncated ack packet */
3495 /* depends on ack packet struct */
3496 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3497 first = ntohl(ap->firstPacket);
3498 serial = ntohl(ap->serial);
3499 /* temporarily disabled -- needs to degrade over time
3500 * skew = ntohs(ap->maxSkew); */
3502 /* Ignore ack packets received out of order */
3503 if (first < call->tfirst) {
3507 if (np->header.flags & RX_SLOW_START_OK) {
3508 call->flags |= RX_CALL_SLOW_START_OK;
3511 if (ap->reason == RX_ACK_PING_RESPONSE)
3512 rxi_UpdatePeerReach(conn, call);
3517 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3518 ap->reason, ntohl(ap->previousPacket),
3519 (unsigned int)np->header.seq, (unsigned int)serial,
3520 (unsigned int)skew, ntohl(ap->firstPacket));
3523 for (offset = 0; offset < nAcks; offset++)
3524 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3531 /* Update the outgoing packet skew value to the latest value of
3532 * the peer's incoming packet skew value. The ack packet, of
3533 * course, could arrive out of order, but that won't affect things
3535 MUTEX_ENTER(&peer->peer_lock);
3536 peer->outPacketSkew = skew;
3538 /* Check for packets that no longer need to be transmitted, and
3539 * discard them. This only applies to packets positively
3540 * acknowledged as having been sent to the peer's upper level.
3541 * All other packets must be retained. So only packets with
3542 * sequence numbers < ap->firstPacket are candidates. */
3543 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3544 if (tp->header.seq >= first)
3546 call->tfirst = tp->header.seq + 1;
3548 && (tp->header.serial == serial || tp->firstSerial == serial))
3549 rxi_ComputePeerNetStats(call, tp, ap, np);
3550 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3551 /* XXX Hack. Because we have to release the global rx lock when sending
3552 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3553 * in rxi_Start sending packets out because packets may move to the
3554 * freePacketQueue as result of being here! So we drop these packets until
3555 * we're safely out of the traversing. Really ugly!
3556 * To make it even uglier, if we're using fine grain locking, we can
3557 * set the ack bits in the packets and have rxi_Start remove the packets
3558 * when it's done transmitting.
3560 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3563 if (call->flags & RX_CALL_TQ_BUSY) {
3564 #ifdef RX_ENABLE_LOCKS
3565 tp->flags |= RX_PKTFLAG_ACKED;
3566 call->flags |= RX_CALL_TQ_SOME_ACKED;
3567 #else /* RX_ENABLE_LOCKS */
3569 #endif /* RX_ENABLE_LOCKS */
3571 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3574 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3579 /* Give rate detector a chance to respond to ping requests */
3580 if (ap->reason == RX_ACK_PING_RESPONSE) {
3581 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3585 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3587 /* Now go through explicit acks/nacks and record the results in
3588 * the waiting packets. These are packets that can't be released
3589 * yet, even with a positive acknowledge. This positive
3590 * acknowledge only means the packet has been received by the
3591 * peer, not that it will be retained long enough to be sent to
3592 * the peer's upper level. In addition, reset the transmit timers
3593 * of any missing packets (those packets that must be missing
3594 * because this packet was out of sequence) */
3596 call->nSoftAcked = 0;
3597 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3598 /* Update round trip time if the ack was stimulated on receipt
3600 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3601 #ifdef RX_ENABLE_LOCKS
3602 if (tp->header.seq >= first)
3603 #endif /* RX_ENABLE_LOCKS */
3604 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3606 && (tp->header.serial == serial || tp->firstSerial == serial))
3607 rxi_ComputePeerNetStats(call, tp, ap, np);
3609 /* Set the acknowledge flag per packet based on the
3610 * information in the ack packet. An acknowlegded packet can
3611 * be downgraded when the server has discarded a packet it
3612 * soacked previously, or when an ack packet is received
3613 * out of sequence. */
3614 if (tp->header.seq < first) {
3615 /* Implicit ack information */
3616 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3619 tp->flags |= RX_PKTFLAG_ACKED;
3620 } else if (tp->header.seq < first + nAcks) {
3621 /* Explicit ack information: set it in the packet appropriately */
3622 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3623 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3625 tp->flags |= RX_PKTFLAG_ACKED;
3633 tp->flags &= ~RX_PKTFLAG_ACKED;
3637 tp->flags &= ~RX_PKTFLAG_ACKED;
3641 /* If packet isn't yet acked, and it has been transmitted at least
3642 * once, reset retransmit time using latest timeout
3643 * ie, this should readjust the retransmit timer for all outstanding
3644 * packets... So we don't just retransmit when we should know better*/
3646 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3647 tp->retryTime = tp->timeSent;
3648 clock_Add(&tp->retryTime, &peer->timeout);
3649 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3650 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3654 /* If the window has been extended by this acknowledge packet,
3655 * then wakeup a sender waiting in alloc for window space, or try
3656 * sending packets now, if he's been sitting on packets due to
3657 * lack of window space */
3658 if (call->tnext < (call->tfirst + call->twind)) {
3659 #ifdef RX_ENABLE_LOCKS
3660 CV_SIGNAL(&call->cv_twind);
3662 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3663 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3664 osi_rxWakeup(&call->twind);
3667 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3668 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3672 /* if the ack packet has a receivelen field hanging off it,
3673 * update our state */
3674 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3677 /* If the ack packet has a "recommended" size that is less than
3678 * what I am using now, reduce my size to match */
3679 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3680 sizeof(afs_int32), &tSize);
3681 tSize = (afs_uint32) ntohl(tSize);
3682 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3684 /* Get the maximum packet size to send to this peer */
3685 rx_packetread(np, rx_AckDataSize(ap->nAcks), sizeof(afs_int32),
3687 tSize = (afs_uint32) ntohl(tSize);
3688 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3689 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3691 /* sanity check - peer might have restarted with different params.
3692 * If peer says "send less", dammit, send less... Peer should never
3693 * be unable to accept packets of the size that prior AFS versions would
3694 * send without asking. */
3695 if (peer->maxMTU != tSize) {
3696 peer->maxMTU = tSize;
3697 peer->MTU = MIN(tSize, peer->MTU);
3698 call->MTU = MIN(call->MTU, tSize);
3702 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3705 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3706 sizeof(afs_int32), &tSize);
3707 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3708 if (tSize < call->twind) { /* smaller than our send */
3709 call->twind = tSize; /* window, we must send less... */
3710 call->ssthresh = MIN(call->twind, call->ssthresh);
3713 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3714 * network MTU confused with the loopback MTU. Calculate the
3715 * maximum MTU here for use in the slow start code below.
3717 maxMTU = peer->maxMTU;
3718 /* Did peer restart with older RX version? */
3719 if (peer->maxDgramPackets > 1) {
3720 peer->maxDgramPackets = 1;
3722 } else if (np->length >=
3723 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3726 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3727 sizeof(afs_int32), &tSize);
3728 tSize = (afs_uint32) ntohl(tSize);
3730 * As of AFS 3.5 we set the send window to match the receive window.
3732 if (tSize < call->twind) {
3733 call->twind = tSize;
3734 call->ssthresh = MIN(call->twind, call->ssthresh);
3735 } else if (tSize > call->twind) {
3736 call->twind = tSize;
3740 * As of AFS 3.5, a jumbogram is more than one fixed size
3741 * packet transmitted in a single UDP datagram. If the remote
3742 * MTU is smaller than our local MTU then never send a datagram
3743 * larger than the natural MTU.
3746 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3747 sizeof(afs_int32), &tSize);
3748 maxDgramPackets = (afs_uint32) ntohl(tSize);
3749 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3751 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3752 maxDgramPackets = MIN(maxDgramPackets, tSize);
3753 if (maxDgramPackets > 1) {
3754 peer->maxDgramPackets = maxDgramPackets;
3755 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3757 peer->maxDgramPackets = 1;
3758 call->MTU = peer->natMTU;
3760 } else if (peer->maxDgramPackets > 1) {
3761 /* Restarted with lower version of RX */
3762 peer->maxDgramPackets = 1;
3764 } else if (peer->maxDgramPackets > 1
3765 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3766 /* Restarted with lower version of RX */
3767 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3768 peer->natMTU = OLD_MAX_PACKET_SIZE;
3769 peer->MTU = OLD_MAX_PACKET_SIZE;
3770 peer->maxDgramPackets = 1;
3771 peer->nDgramPackets = 1;
3773 call->MTU = OLD_MAX_PACKET_SIZE;
3778 * Calculate how many datagrams were successfully received after
3779 * the first missing packet and adjust the negative ack counter
3784 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3785 if (call->nNacks < nNacked) {
3786 call->nNacks = nNacked;
3795 if (call->flags & RX_CALL_FAST_RECOVER) {
3797 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3799 call->flags &= ~RX_CALL_FAST_RECOVER;
3800 call->cwind = call->nextCwind;
3801 call->nextCwind = 0;
3804 call->nCwindAcks = 0;
3805 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3806 /* Three negative acks in a row trigger congestion recovery */
3807 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3808 MUTEX_EXIT(&peer->peer_lock);
3809 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3810 /* someone else is waiting to start recovery */
3813 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3814 while (call->flags & RX_CALL_TQ_BUSY) {
3815 call->flags |= RX_CALL_TQ_WAIT;
3817 #ifdef RX_ENABLE_LOCKS
3818 osirx_AssertMine(&call->lock, "rxi_Start lock2");
3819 CV_WAIT(&call->cv_tq, &call->lock);
3820 #else /* RX_ENABLE_LOCKS */
3821 osi_rxSleep(&call->tq);
3822 #endif /* RX_ENABLE_LOCKS */
3824 if (call->tqWaiters == 0)
3825 call->flags &= ~RX_CALL_TQ_WAIT;
3827 MUTEX_ENTER(&peer->peer_lock);
3828 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3829 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3830 call->flags |= RX_CALL_FAST_RECOVER;
3831 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3833 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3834 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3835 call->nextCwind = call->ssthresh;
3838 peer->MTU = call->MTU;
3839 peer->cwind = call->nextCwind;
3840 peer->nDgramPackets = call->nDgramPackets;
3842 call->congestSeq = peer->congestSeq;
3843 /* Reset the resend times on the packets that were nacked
3844 * so we will retransmit as soon as the window permits*/
3845 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3847 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3848 clock_Zero(&tp->retryTime);
3850 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3855 /* If cwind is smaller than ssthresh, then increase
3856 * the window one packet for each ack we receive (exponential
3858 * If cwind is greater than or equal to ssthresh then increase
3859 * the congestion window by one packet for each cwind acks we
3860 * receive (linear growth). */
3861 if (call->cwind < call->ssthresh) {
3863 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3864 call->nCwindAcks = 0;
3866 call->nCwindAcks += newAckCount;
3867 if (call->nCwindAcks >= call->cwind) {
3868 call->nCwindAcks = 0;
3869 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3873 * If we have received several acknowledgements in a row then
3874 * it is time to increase the size of our datagrams
3876 if ((int)call->nAcks > rx_nDgramThreshold) {
3877 if (peer->maxDgramPackets > 1) {
3878 if (call->nDgramPackets < peer->maxDgramPackets) {
3879 call->nDgramPackets++;
3881 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3882 } else if (call->MTU < peer->maxMTU) {
3883 call->MTU += peer->natMTU;
3884 call->MTU = MIN(call->MTU, peer->maxMTU);
3890 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3892 /* Servers need to hold the call until all response packets have
3893 * been acknowledged. Soft acks are good enough since clients
3894 * are not allowed to clear their receive queues. */
3895 if (call->state == RX_STATE_HOLD
3896 && call->tfirst + call->nSoftAcked >= call->tnext) {
3897 call->state = RX_STATE_DALLY;
3898 rxi_ClearTransmitQueue(call, 0);
3899 } else if (!queue_IsEmpty(&call->tq)) {
3900 rxi_Start(0, call, 0, istack);
3905 /* Received a response to a challenge packet */
3907 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3908 register struct rx_packet *np, int istack)
3912 /* Ignore the packet if we're the client */
3913 if (conn->type == RX_CLIENT_CONNECTION)
3916 /* If already authenticated, ignore the packet (it's probably a retry) */
3917 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3920 /* Otherwise, have the security object evaluate the response packet */
3921 error = RXS_CheckResponse(conn->securityObject, conn, np);
3923 /* If the response is invalid, reset the connection, sending
3924 * an abort to the peer */
3928 rxi_ConnectionError(conn, error);
3929 MUTEX_ENTER(&conn->conn_data_lock);
3930 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3931 MUTEX_EXIT(&conn->conn_data_lock);
3934 /* If the response is valid, any calls waiting to attach
3935 * servers can now do so */
3938 for (i = 0; i < RX_MAXCALLS; i++) {
3939 struct rx_call *call = conn->call[i];
3941 MUTEX_ENTER(&call->lock);
3942 if (call->state == RX_STATE_PRECALL)
3943 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
3944 /* tnop can be null if newcallp is null */
3945 MUTEX_EXIT(&call->lock);
3949 /* Update the peer reachability information, just in case
3950 * some calls went into attach-wait while we were waiting
3951 * for authentication..
3953 rxi_UpdatePeerReach(conn, NULL);
3958 /* A client has received an authentication challenge: the security
3959 * object is asked to cough up a respectable response packet to send
3960 * back to the server. The server is responsible for retrying the
3961 * challenge if it fails to get a response. */
3964 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
3965 register struct rx_packet *np, int istack)
3969 /* Ignore the challenge if we're the server */
3970 if (conn->type == RX_SERVER_CONNECTION)
3973 /* Ignore the challenge if the connection is otherwise idle; someone's
3974 * trying to use us as an oracle. */
3975 if (!rxi_HasActiveCalls(conn))
3978 /* Send the security object the challenge packet. It is expected to fill
3979 * in the response. */
3980 error = RXS_GetResponse(conn->securityObject, conn, np);
3982 /* If the security object is unable to return a valid response, reset the
3983 * connection and send an abort to the peer. Otherwise send the response
3984 * packet to the peer connection. */
3986 rxi_ConnectionError(conn, error);
3987 MUTEX_ENTER(&conn->conn_data_lock);
3988 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3989 MUTEX_EXIT(&conn->conn_data_lock);
3991 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
3992 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
3998 /* Find an available server process to service the current request in
3999 * the given call structure. If one isn't available, queue up this
4000 * call so it eventually gets one */
4002 rxi_AttachServerProc(register struct rx_call *call,
4003 register osi_socket socket, register int *tnop,
4004 register struct rx_call **newcallp)
4006 register struct rx_serverQueueEntry *sq;
4007 register struct rx_service *service = call->conn->service;
4008 register int haveQuota = 0;
4010 /* May already be attached */
4011 if (call->state == RX_STATE_ACTIVE)
4014 MUTEX_ENTER(&rx_serverPool_lock);
4016 haveQuota = QuotaOK(service);
4017 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4018 /* If there are no processes available to service this call,
4019 * put the call on the incoming call queue (unless it's
4020 * already on the queue).
4022 #ifdef RX_ENABLE_LOCKS
4024 ReturnToServerPool(service);
4025 #endif /* RX_ENABLE_LOCKS */
4027 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4028 call->flags |= RX_CALL_WAIT_PROC;
4029 MUTEX_ENTER(&rx_stats_mutex);
4032 MUTEX_EXIT(&rx_stats_mutex);
4033 rxi_calltrace(RX_CALL_ARRIVAL, call);
4034 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4035 queue_Append(&rx_incomingCallQueue, call);
4038 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4040 /* If hot threads are enabled, and both newcallp and sq->socketp
4041 * are non-null, then this thread will process the call, and the
4042 * idle server thread will start listening on this threads socket.
4045 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4048 *sq->socketp = socket;
4049 clock_GetTime(&call->startTime);
4050 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4054 if (call->flags & RX_CALL_WAIT_PROC) {
4055 /* Conservative: I don't think this should happen */
4056 call->flags &= ~RX_CALL_WAIT_PROC;
4057 if (queue_IsOnQueue(call)) {
4059 MUTEX_ENTER(&rx_stats_mutex);
4061 MUTEX_EXIT(&rx_stats_mutex);
4064 call->state = RX_STATE_ACTIVE;
4065 call->mode = RX_MODE_RECEIVING;
4066 #ifdef RX_KERNEL_TRACE
4068 int glockOwner = ISAFS_GLOCK();
4071 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4072 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4078 if (call->flags & RX_CALL_CLEARED) {
4079 /* send an ack now to start the packet flow up again */
4080 call->flags &= ~RX_CALL_CLEARED;
4081 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4083 #ifdef RX_ENABLE_LOCKS
4086 service->nRequestsRunning++;
4087 if (service->nRequestsRunning <= service->minProcs)
4093 MUTEX_EXIT(&rx_serverPool_lock);
4096 /* Delay the sending of an acknowledge event for a short while, while
4097 * a new call is being prepared (in the case of a client) or a reply
4098 * is being prepared (in the case of a server). Rather than sending
4099 * an ack packet, an ACKALL packet is sent. */
4101 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4103 #ifdef RX_ENABLE_LOCKS
4105 MUTEX_ENTER(&call->lock);
4106 call->delayedAckEvent = NULL;
4107 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4109 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4110 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4112 MUTEX_EXIT(&call->lock);
4113 #else /* RX_ENABLE_LOCKS */
4115 call->delayedAckEvent = NULL;
4116 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4117 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4118 #endif /* RX_ENABLE_LOCKS */
4122 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4125 #ifdef RX_ENABLE_LOCKS
4127 MUTEX_ENTER(&call->lock);
4128 if (event == call->delayedAckEvent)
4129 call->delayedAckEvent = NULL;
4130 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4132 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4134 MUTEX_EXIT(&call->lock);
4135 #else /* RX_ENABLE_LOCKS */
4137 call->delayedAckEvent = NULL;
4138 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4139 #endif /* RX_ENABLE_LOCKS */
4143 #ifdef RX_ENABLE_LOCKS
4144 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4145 * clearing them out.
4148 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4150 register struct rx_packet *p, *tp;
4153 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4154 p->flags |= RX_PKTFLAG_ACKED;
4158 call->flags |= RX_CALL_TQ_CLEARME;
4159 call->flags |= RX_CALL_TQ_SOME_ACKED;
4162 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4163 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4164 call->tfirst = call->tnext;
4165 call->nSoftAcked = 0;
4167 if (call->flags & RX_CALL_FAST_RECOVER) {
4168 call->flags &= ~RX_CALL_FAST_RECOVER;
4169 call->cwind = call->nextCwind;
4170 call->nextCwind = 0;
4173 CV_SIGNAL(&call->cv_twind);
4175 #endif /* RX_ENABLE_LOCKS */
4177 /* Clear out the transmit queue for the current call (all packets have
4178 * been received by peer) */
4180 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4182 register struct rx_packet *p, *tp;
4184 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4185 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4187 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4188 p->flags |= RX_PKTFLAG_ACKED;
4192 call->flags |= RX_CALL_TQ_CLEARME;
4193 call->flags |= RX_CALL_TQ_SOME_ACKED;
4196 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4197 rxi_FreePackets(0, &call->tq);
4198 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4199 call->flags &= ~RX_CALL_TQ_CLEARME;
4201 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4203 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4204 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4205 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4206 call->nSoftAcked = 0;
4208 if (call->flags & RX_CALL_FAST_RECOVER) {
4209 call->flags &= ~RX_CALL_FAST_RECOVER;
4210 call->cwind = call->nextCwind;
4212 #ifdef RX_ENABLE_LOCKS
4213 CV_SIGNAL(&call->cv_twind);
4215 osi_rxWakeup(&call->twind);
4220 rxi_ClearReceiveQueue(register struct rx_call *call)
4222 if (queue_IsNotEmpty(&call->rq)) {
4223 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4224 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4226 if (call->state == RX_STATE_PRECALL) {
4227 call->flags |= RX_CALL_CLEARED;
4231 /* Send an abort packet for the specified call */
4233 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4234 int istack, int force)
4242 /* Clients should never delay abort messages */
4243 if (rx_IsClientConn(call->conn))
4246 if (call->abortCode != call->error) {
4247 call->abortCode = call->error;
4248 call->abortCount = 0;
4251 if (force || rxi_callAbortThreshhold == 0
4252 || call->abortCount < rxi_callAbortThreshhold) {
4253 if (call->delayedAbortEvent) {
4254 rxevent_Cancel(call->delayedAbortEvent, call,
4255 RX_CALL_REFCOUNT_ABORT);
4257 error = htonl(call->error);
4260 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4261 (char *)&error, sizeof(error), istack);
4262 } else if (!call->delayedAbortEvent) {
4263 clock_GetTime(&when);
4264 clock_Addmsec(&when, rxi_callAbortDelay);
4265 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4266 call->delayedAbortEvent =
4267 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4272 /* Send an abort packet for the specified connection. Packet is an
4273 * optional pointer to a packet that can be used to send the abort.
4274 * Once the number of abort messages reaches the threshhold, an
4275 * event is scheduled to send the abort. Setting the force flag
4276 * overrides sending delayed abort messages.
4278 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4279 * to send the abort packet.
4282 rxi_SendConnectionAbort(register struct rx_connection *conn,
4283 struct rx_packet *packet, int istack, int force)
4291 /* Clients should never delay abort messages */
4292 if (rx_IsClientConn(conn))
4295 if (force || rxi_connAbortThreshhold == 0
4296 || conn->abortCount < rxi_connAbortThreshhold) {
4297 if (conn->delayedAbortEvent) {
4298 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4300 error = htonl(conn->error);
4302 MUTEX_EXIT(&conn->conn_data_lock);
4304 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4305 RX_PACKET_TYPE_ABORT, (char *)&error,
4306 sizeof(error), istack);
4307 MUTEX_ENTER(&conn->conn_data_lock);
4308 } else if (!conn->delayedAbortEvent) {
4309 clock_GetTime(&when);
4310 clock_Addmsec(&when, rxi_connAbortDelay);
4311 conn->delayedAbortEvent =
4312 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4317 /* Associate an error all of the calls owned by a connection. Called
4318 * with error non-zero. This is only for really fatal things, like
4319 * bad authentication responses. The connection itself is set in
4320 * error at this point, so that future packets received will be
4323 rxi_ConnectionError(register struct rx_connection *conn,
4324 register afs_int32 error)
4328 MUTEX_ENTER(&conn->conn_data_lock);
4329 if (conn->challengeEvent)
4330 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4331 if (conn->checkReachEvent) {
4332 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4333 conn->checkReachEvent = 0;
4334 conn->flags &= ~RX_CONN_ATTACHWAIT;
4337 MUTEX_EXIT(&conn->conn_data_lock);
4338 for (i = 0; i < RX_MAXCALLS; i++) {
4339 struct rx_call *call = conn->call[i];
4341 MUTEX_ENTER(&call->lock);
4342 rxi_CallError(call, error);
4343 MUTEX_EXIT(&call->lock);
4346 conn->error = error;
4347 MUTEX_ENTER(&rx_stats_mutex);
4348 rx_stats.fatalErrors++;
4349 MUTEX_EXIT(&rx_stats_mutex);
4354 rxi_CallError(register struct rx_call *call, afs_int32 error)
4357 error = call->error;
4358 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4359 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4360 rxi_ResetCall(call, 0);
4363 rxi_ResetCall(call, 0);
4365 call->error = error;
4366 call->mode = RX_MODE_ERROR;
4369 /* Reset various fields in a call structure, and wakeup waiting
4370 * processes. Some fields aren't changed: state & mode are not
4371 * touched (these must be set by the caller), and bufptr, nLeft, and
4372 * nFree are not reset, since these fields are manipulated by
4373 * unprotected macros, and may only be reset by non-interrupting code.
4376 /* this code requires that call->conn be set properly as a pre-condition. */
4377 #endif /* ADAPT_WINDOW */
4380 rxi_ResetCall(register struct rx_call *call, register int newcall)
4383 register struct rx_peer *peer;
4384 struct rx_packet *packet;
4386 /* Notify anyone who is waiting for asynchronous packet arrival */
4387 if (call->arrivalProc) {
4388 (*call->arrivalProc) (call, call->arrivalProcHandle,
4389 call->arrivalProcArg);
4390 call->arrivalProc = (void (*)())0;
4393 if (call->delayedAbortEvent) {
4394 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4395 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4397 rxi_SendCallAbort(call, packet, 0, 1);
4398 rxi_FreePacket(packet);
4403 * Update the peer with the congestion information in this call
4404 * so other calls on this connection can pick up where this call
4405 * left off. If the congestion sequence numbers don't match then
4406 * another call experienced a retransmission.
4408 peer = call->conn->peer;
4409 MUTEX_ENTER(&peer->peer_lock);
4411 if (call->congestSeq == peer->congestSeq) {
4412 peer->cwind = MAX(peer->cwind, call->cwind);
4413 peer->MTU = MAX(peer->MTU, call->MTU);
4414 peer->nDgramPackets =
4415 MAX(peer->nDgramPackets, call->nDgramPackets);
4418 call->abortCode = 0;
4419 call->abortCount = 0;
4421 if (peer->maxDgramPackets > 1) {
4422 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4424 call->MTU = peer->MTU;
4426 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4427 call->ssthresh = rx_maxSendWindow;
4428 call->nDgramPackets = peer->nDgramPackets;
4429 call->congestSeq = peer->congestSeq;
4430 MUTEX_EXIT(&peer->peer_lock);
4432 flags = call->flags;
4433 rxi_ClearReceiveQueue(call);
4434 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4435 if (flags & RX_CALL_TQ_BUSY) {
4436 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4437 call->flags |= (flags & RX_CALL_TQ_WAIT);
4439 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4441 rxi_ClearTransmitQueue(call, 0);
4442 queue_Init(&call->tq);
4443 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4444 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4447 while (call->tqWaiters) {
4448 #ifdef RX_ENABLE_LOCKS
4449 CV_BROADCAST(&call->cv_tq);
4450 #else /* RX_ENABLE_LOCKS */
4451 osi_rxWakeup(&call->tq);
4452 #endif /* RX_ENABLE_LOCKS */
4456 queue_Init(&call->rq);
4458 call->rwind = rx_initReceiveWindow;
4459 call->twind = rx_initSendWindow;
4460 call->nSoftAcked = 0;
4461 call->nextCwind = 0;
4464 call->nCwindAcks = 0;
4465 call->nSoftAcks = 0;
4466 call->nHardAcks = 0;
4468 call->tfirst = call->rnext = call->tnext = 1;
4470 call->lastAcked = 0;
4471 call->localStatus = call->remoteStatus = 0;
4473 if (flags & RX_CALL_READER_WAIT) {
4474 #ifdef RX_ENABLE_LOCKS
4475 CV_BROADCAST(&call->cv_rq);
4477 osi_rxWakeup(&call->rq);
4480 if (flags & RX_CALL_WAIT_PACKETS) {
4481 MUTEX_ENTER(&rx_freePktQ_lock);
4482 rxi_PacketsUnWait(); /* XXX */
4483 MUTEX_EXIT(&rx_freePktQ_lock);
4485 #ifdef RX_ENABLE_LOCKS
4486 CV_SIGNAL(&call->cv_twind);
4488 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4489 osi_rxWakeup(&call->twind);
4492 #ifdef RX_ENABLE_LOCKS
4493 /* The following ensures that we don't mess with any queue while some
4494 * other thread might also be doing so. The call_queue_lock field is
4495 * is only modified under the call lock. If the call is in the process
4496 * of being removed from a queue, the call is not locked until the
4497 * the queue lock is dropped and only then is the call_queue_lock field
4498 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4499 * Note that any other routine which removes a call from a queue has to
4500 * obtain the queue lock before examing the queue and removing the call.
4502 if (call->call_queue_lock) {
4503 MUTEX_ENTER(call->call_queue_lock);
4504 if (queue_IsOnQueue(call)) {
4506 if (flags & RX_CALL_WAIT_PROC) {
4507 MUTEX_ENTER(&rx_stats_mutex);
4509 MUTEX_EXIT(&rx_stats_mutex);
4512 MUTEX_EXIT(call->call_queue_lock);
4513 CLEAR_CALL_QUEUE_LOCK(call);
4515 #else /* RX_ENABLE_LOCKS */
4516 if (queue_IsOnQueue(call)) {
4518 if (flags & RX_CALL_WAIT_PROC)
4521 #endif /* RX_ENABLE_LOCKS */
4523 rxi_KeepAliveOff(call);
4524 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4527 /* Send an acknowledge for the indicated packet (seq,serial) of the
4528 * indicated call, for the indicated reason (reason). This
4529 * acknowledge will specifically acknowledge receiving the packet, and
4530 * will also specify which other packets for this call have been
4531 * received. This routine returns the packet that was used to the
4532 * caller. The caller is responsible for freeing it or re-using it.
4533 * This acknowledgement also returns the highest sequence number
4534 * actually read out by the higher level to the sender; the sender
4535 * promises to keep around packets that have not been read by the
4536 * higher level yet (unless, of course, the sender decides to abort
4537 * the call altogether). Any of p, seq, serial, pflags, or reason may
4538 * be set to zero without ill effect. That is, if they are zero, they
4539 * will not convey any information.
4540 * NOW there is a trailer field, after the ack where it will safely be
4541 * ignored by mundanes, which indicates the maximum size packet this
4542 * host can swallow. */
4544 register struct rx_packet *optionalPacket; use to send ack (or null)
4545 int seq; Sequence number of the packet we are acking
4546 int serial; Serial number of the packet
4547 int pflags; Flags field from packet header
4548 int reason; Reason an acknowledge was prompted
4552 rxi_SendAck(register struct rx_call *call,
4553 register struct rx_packet *optionalPacket, int serial, int reason,
4556 struct rx_ackPacket *ap;
4557 register struct rx_packet *rqp;
4558 register struct rx_packet *nxp; /* For queue_Scan */
4559 register struct rx_packet *p;
4562 #ifdef RX_ENABLE_TSFPQ
4563 struct rx_ts_info_t * rx_ts_info;
4567 * Open the receive window once a thread starts reading packets
4569 if (call->rnext > 1) {
4570 call->rwind = rx_maxReceiveWindow;
4573 call->nHardAcks = 0;
4574 call->nSoftAcks = 0;
4575 if (call->rnext > call->lastAcked)
4576 call->lastAcked = call->rnext;
4580 rx_computelen(p, p->length); /* reset length, you never know */
4581 } /* where that's been... */
4582 #ifdef RX_ENABLE_TSFPQ
4584 RX_TS_INFO_GET(rx_ts_info);
4585 if ((p = rx_ts_info->local_special_packet)) {
4586 rx_computelen(p, p->length);
4587 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4588 rx_ts_info->local_special_packet = p;
4589 } else { /* We won't send the ack, but don't panic. */
4590 return optionalPacket;
4594 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4595 /* We won't send the ack, but don't panic. */
4596 return optionalPacket;
4601 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4604 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4605 #ifndef RX_ENABLE_TSFPQ
4606 if (!optionalPacket)
4609 return optionalPacket;
4611 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4612 if (rx_Contiguous(p) < templ) {
4613 #ifndef RX_ENABLE_TSFPQ
4614 if (!optionalPacket)
4617 return optionalPacket;
4622 /* MTUXXX failing to send an ack is very serious. We should */
4623 /* try as hard as possible to send even a partial ack; it's */
4624 /* better than nothing. */
4625 ap = (struct rx_ackPacket *)rx_DataOf(p);
4626 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4627 ap->reason = reason;
4629 /* The skew computation used to be bogus, I think it's better now. */
4630 /* We should start paying attention to skew. XXX */
4631 ap->serial = htonl(serial);
4632 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4634 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4635 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4637 /* No fear of running out of ack packet here because there can only be at most
4638 * one window full of unacknowledged packets. The window size must be constrained
4639 * to be less than the maximum ack size, of course. Also, an ack should always
4640 * fit into a single packet -- it should not ever be fragmented. */
4641 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4642 if (!rqp || !call->rq.next
4643 || (rqp->header.seq > (call->rnext + call->rwind))) {
4644 #ifndef RX_ENABLE_TSFPQ
4645 if (!optionalPacket)
4648 rxi_CallError(call, RX_CALL_DEAD);
4649 return optionalPacket;
4652 while (rqp->header.seq > call->rnext + offset)
4653 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4654 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4656 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4657 #ifndef RX_ENABLE_TSFPQ
4658 if (!optionalPacket)
4661 rxi_CallError(call, RX_CALL_DEAD);
4662 return optionalPacket;
4667 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4669 /* these are new for AFS 3.3 */
4670 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4671 templ = htonl(templ);
4672 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4673 templ = htonl(call->conn->peer->ifMTU);
4674 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4675 sizeof(afs_int32), &templ);
4677 /* new for AFS 3.4 */
4678 templ = htonl(call->rwind);
4679 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4680 sizeof(afs_int32), &templ);
4682 /* new for AFS 3.5 */
4683 templ = htonl(call->conn->peer->ifDgramPackets);
4684 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4685 sizeof(afs_int32), &templ);
4687 p->header.serviceId = call->conn->serviceId;
4688 p->header.cid = (call->conn->cid | call->channel);
4689 p->header.callNumber = *call->callNumber;
4691 p->header.securityIndex = call->conn->securityIndex;
4692 p->header.epoch = call->conn->epoch;
4693 p->header.type = RX_PACKET_TYPE_ACK;
4694 p->header.flags = RX_SLOW_START_OK;
4695 if (reason == RX_ACK_PING) {
4696 p->header.flags |= RX_REQUEST_ACK;
4698 clock_GetTime(&call->pingRequestTime);
4701 if (call->conn->type == RX_CLIENT_CONNECTION)
4702 p->header.flags |= RX_CLIENT_INITIATED;
4706 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u",
4707 ap->reason, ntohl(ap->previousPacket),
4708 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4710 for (offset = 0; offset < ap->nAcks; offset++)
4711 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4719 register int i, nbytes = p->length;
4721 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4722 if (nbytes <= p->wirevec[i].iov_len) {
4723 register int savelen, saven;
4725 savelen = p->wirevec[i].iov_len;
4727 p->wirevec[i].iov_len = nbytes;
4729 rxi_Send(call, p, istack);
4730 p->wirevec[i].iov_len = savelen;
4734 nbytes -= p->wirevec[i].iov_len;
4737 MUTEX_ENTER(&rx_stats_mutex);
4738 rx_stats.ackPacketsSent++;
4739 MUTEX_EXIT(&rx_stats_mutex);
4740 #ifndef RX_ENABLE_TSFPQ
4741 if (!optionalPacket)
4744 return optionalPacket; /* Return packet for re-use by caller */
4747 /* Send all of the packets in the list in single datagram */
4749 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4750 int istack, int moreFlag, struct clock *now,
4751 struct clock *retryTime, int resending)
4756 struct rx_connection *conn = call->conn;
4757 struct rx_peer *peer = conn->peer;
4759 MUTEX_ENTER(&peer->peer_lock);
4762 peer->reSends += len;
4763 MUTEX_ENTER(&rx_stats_mutex);
4764 rx_stats.dataPacketsSent += len;
4765 MUTEX_EXIT(&rx_stats_mutex);
4766 MUTEX_EXIT(&peer->peer_lock);
4768 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4772 /* Set the packet flags and schedule the resend events */
4773 /* Only request an ack for the last packet in the list */
4774 for (i = 0; i < len; i++) {
4775 list[i]->retryTime = *retryTime;
4776 if (list[i]->header.serial) {
4777 /* Exponentially backoff retry times */
4778 if (list[i]->backoff < MAXBACKOFF) {
4779 /* so it can't stay == 0 */
4780 list[i]->backoff = (list[i]->backoff << 1) + 1;
4783 clock_Addmsec(&(list[i]->retryTime),
4784 ((afs_uint32) list[i]->backoff) << 8);
4787 /* Wait a little extra for the ack on the last packet */
4788 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4789 clock_Addmsec(&(list[i]->retryTime), 400);
4792 /* Record the time sent */
4793 list[i]->timeSent = *now;
4795 /* Ask for an ack on retransmitted packets, on every other packet
4796 * if the peer doesn't support slow start. Ask for an ack on every
4797 * packet until the congestion window reaches the ack rate. */
4798 if (list[i]->header.serial) {
4800 MUTEX_ENTER(&rx_stats_mutex);
4801 rx_stats.dataPacketsReSent++;
4802 MUTEX_EXIT(&rx_stats_mutex);
4804 /* improved RTO calculation- not Karn */
4805 list[i]->firstSent = *now;
4806 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4807 || (!(call->flags & RX_CALL_SLOW_START_OK)
4808 && (list[i]->header.seq & 1)))) {
4813 MUTEX_ENTER(&peer->peer_lock);
4817 MUTEX_ENTER(&rx_stats_mutex);
4818 rx_stats.dataPacketsSent++;
4819 MUTEX_EXIT(&rx_stats_mutex);
4820 MUTEX_EXIT(&peer->peer_lock);
4822 /* Tag this packet as not being the last in this group,
4823 * for the receiver's benefit */
4824 if (i < len - 1 || moreFlag) {
4825 list[i]->header.flags |= RX_MORE_PACKETS;
4828 /* Install the new retransmit time for the packet, and
4829 * record the time sent */
4830 list[i]->timeSent = *now;
4834 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4837 /* Since we're about to send a data packet to the peer, it's
4838 * safe to nuke any scheduled end-of-packets ack */
4839 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4841 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4842 MUTEX_EXIT(&call->lock);
4844 rxi_SendPacketList(call, conn, list, len, istack);
4846 rxi_SendPacket(call, conn, list[0], istack);
4848 MUTEX_ENTER(&call->lock);
4849 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4851 /* Update last send time for this call (for keep-alive
4852 * processing), and for the connection (so that we can discover
4853 * idle connections) */
4854 conn->lastSendTime = call->lastSendTime = clock_Sec();
4857 /* When sending packets we need to follow these rules:
4858 * 1. Never send more than maxDgramPackets in a jumbogram.
4859 * 2. Never send a packet with more than two iovecs in a jumbogram.
4860 * 3. Never send a retransmitted packet in a jumbogram.
4861 * 4. Never send more than cwind/4 packets in a jumbogram
4862 * We always keep the last list we should have sent so we
4863 * can set the RX_MORE_PACKETS flags correctly.
4866 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4867 int istack, struct clock *now, struct clock *retryTime,
4870 int i, cnt, lastCnt = 0;
4871 struct rx_packet **listP, **lastP = 0;
4872 struct rx_peer *peer = call->conn->peer;
4873 int morePackets = 0;
4875 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4876 /* Does the current packet force us to flush the current list? */
4878 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4879 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4881 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4883 /* If the call enters an error state stop sending, or if
4884 * we entered congestion recovery mode, stop sending */
4885 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4893 /* Add the current packet to the list if it hasn't been acked.
4894 * Otherwise adjust the list pointer to skip the current packet. */
4895 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
4897 /* Do we need to flush the list? */
4898 if (cnt >= (int)peer->maxDgramPackets
4899 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
4900 || list[i]->header.serial
4901 || list[i]->length != RX_JUMBOBUFFERSIZE) {
4903 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
4904 retryTime, resending);
4905 /* If the call enters an error state stop sending, or if
4906 * we entered congestion recovery mode, stop sending */
4908 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4913 listP = &list[i + 1];
4918 osi_Panic("rxi_SendList error");
4920 listP = &list[i + 1];
4924 /* Send the whole list when the call is in receive mode, when
4925 * the call is in eof mode, when we are in fast recovery mode,
4926 * and when we have the last packet */
4927 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
4928 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
4929 || (call->flags & RX_CALL_FAST_RECOVER)) {
4930 /* Check for the case where the current list contains
4931 * an acked packet. Since we always send retransmissions
4932 * in a separate packet, we only need to check the first
4933 * packet in the list */
4934 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
4938 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
4939 retryTime, resending);
4940 /* If the call enters an error state stop sending, or if
4941 * we entered congestion recovery mode, stop sending */
4942 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4946 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
4949 } else if (lastCnt > 0) {
4950 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
4955 #ifdef RX_ENABLE_LOCKS
4956 /* Call rxi_Start, below, but with the call lock held. */
4958 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
4959 void *arg1, int istack)
4961 MUTEX_ENTER(&call->lock);
4962 rxi_Start(event, call, arg1, istack);
4963 MUTEX_EXIT(&call->lock);
4965 #endif /* RX_ENABLE_LOCKS */
4967 /* This routine is called when new packets are readied for
4968 * transmission and when retransmission may be necessary, or when the
4969 * transmission window or burst count are favourable. This should be
4970 * better optimized for new packets, the usual case, now that we've
4971 * got rid of queues of send packets. XXXXXXXXXXX */
4973 rxi_Start(struct rxevent *event, register struct rx_call *call,
4974 void *arg1, int istack)
4976 struct rx_packet *p;
4977 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
4978 struct rx_peer *peer = call->conn->peer;
4979 struct clock now, retryTime;
4983 struct rx_packet **xmitList;
4986 /* If rxi_Start is being called as a result of a resend event,
4987 * then make sure that the event pointer is removed from the call
4988 * structure, since there is no longer a per-call retransmission
4990 if (event && event == call->resendEvent) {
4991 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
4992 call->resendEvent = NULL;
4994 if (queue_IsEmpty(&call->tq)) {
4998 /* Timeouts trigger congestion recovery */
4999 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5000 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5001 /* someone else is waiting to start recovery */
5004 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5005 while (call->flags & RX_CALL_TQ_BUSY) {
5006 call->flags |= RX_CALL_TQ_WAIT;
5008 #ifdef RX_ENABLE_LOCKS
5009 osirx_AssertMine(&call->lock, "rxi_Start lock1");
5010 CV_WAIT(&call->cv_tq, &call->lock);
5011 #else /* RX_ENABLE_LOCKS */
5012 osi_rxSleep(&call->tq);
5013 #endif /* RX_ENABLE_LOCKS */
5015 if (call->tqWaiters == 0)
5016 call->flags &= ~RX_CALL_TQ_WAIT;
5018 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5019 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5020 call->flags |= RX_CALL_FAST_RECOVER;
5021 if (peer->maxDgramPackets > 1) {
5022 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5024 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5026 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5027 call->nDgramPackets = 1;
5029 call->nextCwind = 1;
5032 MUTEX_ENTER(&peer->peer_lock);
5033 peer->MTU = call->MTU;
5034 peer->cwind = call->cwind;
5035 peer->nDgramPackets = 1;
5037 call->congestSeq = peer->congestSeq;
5038 MUTEX_EXIT(&peer->peer_lock);
5039 /* Clear retry times on packets. Otherwise, it's possible for
5040 * some packets in the queue to force resends at rates faster
5041 * than recovery rates.
5043 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5044 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5045 clock_Zero(&p->retryTime);
5050 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5051 MUTEX_ENTER(&rx_stats_mutex);
5052 rx_tq_debug.rxi_start_in_error++;
5053 MUTEX_EXIT(&rx_stats_mutex);
5058 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5059 /* Get clock to compute the re-transmit time for any packets
5060 * in this burst. Note, if we back off, it's reasonable to
5061 * back off all of the packets in the same manner, even if
5062 * some of them have been retransmitted more times than more
5063 * recent additions */
5064 clock_GetTime(&now);
5065 retryTime = now; /* initialize before use */
5066 MUTEX_ENTER(&peer->peer_lock);
5067 clock_Add(&retryTime, &peer->timeout);
5068 MUTEX_EXIT(&peer->peer_lock);
5070 /* Send (or resend) any packets that need it, subject to
5071 * window restrictions and congestion burst control
5072 * restrictions. Ask for an ack on the last packet sent in
5073 * this burst. For now, we're relying upon the window being
5074 * considerably bigger than the largest number of packets that
5075 * are typically sent at once by one initial call to
5076 * rxi_Start. This is probably bogus (perhaps we should ask
5077 * for an ack when we're half way through the current
5078 * window?). Also, for non file transfer applications, this
5079 * may end up asking for an ack for every packet. Bogus. XXXX
5082 * But check whether we're here recursively, and let the other guy
5085 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5086 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5087 call->flags |= RX_CALL_TQ_BUSY;
5089 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5091 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5092 call->flags &= ~RX_CALL_NEED_START;
5093 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5095 maxXmitPackets = MIN(call->twind, call->cwind);
5096 xmitList = (struct rx_packet **)
5097 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5098 if (xmitList == NULL)
5099 osi_Panic("rxi_Start, failed to allocate xmit list");
5100 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5101 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5102 /* We shouldn't be sending packets if a thread is waiting
5103 * to initiate congestion recovery */
5107 && (call->flags & RX_CALL_FAST_RECOVER)) {
5108 /* Only send one packet during fast recovery */
5111 if ((p->flags & RX_PKTFLAG_FREE)
5112 || (!queue_IsEnd(&call->tq, nxp)
5113 && (nxp->flags & RX_PKTFLAG_FREE))
5114 || (p == (struct rx_packet *)&rx_freePacketQueue)
5115 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5116 osi_Panic("rxi_Start: xmit queue clobbered");
5118 if (p->flags & RX_PKTFLAG_ACKED) {
5119 MUTEX_ENTER(&rx_stats_mutex);
5120 rx_stats.ignoreAckedPacket++;
5121 MUTEX_EXIT(&rx_stats_mutex);
5122 continue; /* Ignore this packet if it has been acknowledged */
5125 /* Turn off all flags except these ones, which are the same
5126 * on each transmission */
5127 p->header.flags &= RX_PRESET_FLAGS;
5129 if (p->header.seq >=
5130 call->tfirst + MIN((int)call->twind,
5131 (int)(call->nSoftAcked +
5133 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5134 /* Note: if we're waiting for more window space, we can
5135 * still send retransmits; hence we don't return here, but
5136 * break out to schedule a retransmit event */
5137 dpf(("call %d waiting for window",
5138 *(call->callNumber)));
5142 /* Transmit the packet if it needs to be sent. */
5143 if (!clock_Lt(&now, &p->retryTime)) {
5144 if (nXmitPackets == maxXmitPackets) {
5145 rxi_SendXmitList(call, xmitList, nXmitPackets,
5146 istack, &now, &retryTime,
5148 osi_Free(xmitList, maxXmitPackets *
5149 sizeof(struct rx_packet *));
5152 xmitList[nXmitPackets++] = p;
5156 /* xmitList now hold pointers to all of the packets that are
5157 * ready to send. Now we loop to send the packets */
5158 if (nXmitPackets > 0) {
5159 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5160 &now, &retryTime, resending);
5163 maxXmitPackets * sizeof(struct rx_packet *));
5165 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5167 * TQ references no longer protected by this flag; they must remain
5168 * protected by the global lock.
5170 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5171 call->flags &= ~RX_CALL_TQ_BUSY;
5172 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5173 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5175 #ifdef RX_ENABLE_LOCKS
5176 osirx_AssertMine(&call->lock, "rxi_Start start");
5177 CV_BROADCAST(&call->cv_tq);
5178 #else /* RX_ENABLE_LOCKS */
5179 osi_rxWakeup(&call->tq);
5180 #endif /* RX_ENABLE_LOCKS */
5184 /* We went into the error state while sending packets. Now is
5185 * the time to reset the call. This will also inform the using
5186 * process that the call is in an error state.
5188 MUTEX_ENTER(&rx_stats_mutex);
5189 rx_tq_debug.rxi_start_aborted++;
5190 MUTEX_EXIT(&rx_stats_mutex);
5191 call->flags &= ~RX_CALL_TQ_BUSY;
5192 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5193 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5195 #ifdef RX_ENABLE_LOCKS
5196 osirx_AssertMine(&call->lock, "rxi_Start middle");
5197 CV_BROADCAST(&call->cv_tq);
5198 #else /* RX_ENABLE_LOCKS */
5199 osi_rxWakeup(&call->tq);
5200 #endif /* RX_ENABLE_LOCKS */
5201 rxi_CallError(call, call->error);
5204 #ifdef RX_ENABLE_LOCKS
5205 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5206 register int missing;
5207 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5208 /* Some packets have received acks. If they all have, we can clear
5209 * the transmit queue.
5212 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5213 if (p->header.seq < call->tfirst
5214 && (p->flags & RX_PKTFLAG_ACKED)) {
5221 call->flags |= RX_CALL_TQ_CLEARME;
5223 #endif /* RX_ENABLE_LOCKS */
5224 /* Don't bother doing retransmits if the TQ is cleared. */
5225 if (call->flags & RX_CALL_TQ_CLEARME) {
5226 rxi_ClearTransmitQueue(call, 1);
5228 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5231 /* Always post a resend event, if there is anything in the
5232 * queue, and resend is possible. There should be at least
5233 * one unacknowledged packet in the queue ... otherwise none
5234 * of these packets should be on the queue in the first place.
5236 if (call->resendEvent) {
5237 /* Cancel the existing event and post a new one */
5238 rxevent_Cancel(call->resendEvent, call,
5239 RX_CALL_REFCOUNT_RESEND);
5242 /* The retry time is the retry time on the first unacknowledged
5243 * packet inside the current window */
5245 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5246 /* Don't set timers for packets outside the window */
5247 if (p->header.seq >= call->tfirst + call->twind) {
5251 if (!(p->flags & RX_PKTFLAG_ACKED)
5252 && !clock_IsZero(&p->retryTime)) {
5254 retryTime = p->retryTime;
5259 /* Post a new event to re-run rxi_Start when retries may be needed */
5260 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5261 #ifdef RX_ENABLE_LOCKS
5262 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5264 rxevent_Post2(&retryTime, rxi_StartUnlocked,
5265 (void *)call, 0, istack);
5266 #else /* RX_ENABLE_LOCKS */
5268 rxevent_Post2(&retryTime, rxi_Start, (void *)call,
5270 #endif /* RX_ENABLE_LOCKS */
5273 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5274 } while (call->flags & RX_CALL_NEED_START);
5276 * TQ references no longer protected by this flag; they must remain
5277 * protected by the global lock.
5279 call->flags &= ~RX_CALL_TQ_BUSY;
5280 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5281 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5283 #ifdef RX_ENABLE_LOCKS
5284 osirx_AssertMine(&call->lock, "rxi_Start end");
5285 CV_BROADCAST(&call->cv_tq);
5286 #else /* RX_ENABLE_LOCKS */
5287 osi_rxWakeup(&call->tq);
5288 #endif /* RX_ENABLE_LOCKS */
5290 call->flags |= RX_CALL_NEED_START;
5292 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5294 if (call->resendEvent) {
5295 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5300 /* Also adjusts the keep alive parameters for the call, to reflect
5301 * that we have just sent a packet (so keep alives aren't sent
5304 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5307 register struct rx_connection *conn = call->conn;
5309 /* Stamp each packet with the user supplied status */
5310 p->header.userStatus = call->localStatus;
5312 /* Allow the security object controlling this call's security to
5313 * make any last-minute changes to the packet */
5314 RXS_SendPacket(conn->securityObject, call, p);
5316 /* Since we're about to send SOME sort of packet to the peer, it's
5317 * safe to nuke any scheduled end-of-packets ack */
5318 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5320 /* Actually send the packet, filling in more connection-specific fields */
5321 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5322 MUTEX_EXIT(&call->lock);
5323 rxi_SendPacket(call, conn, p, istack);
5324 MUTEX_ENTER(&call->lock);
5325 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5327 /* Update last send time for this call (for keep-alive
5328 * processing), and for the connection (so that we can discover
5329 * idle connections) */
5330 conn->lastSendTime = call->lastSendTime = clock_Sec();
5334 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5335 * that things are fine. Also called periodically to guarantee that nothing
5336 * falls through the cracks (e.g. (error + dally) connections have keepalive
5337 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5339 * haveCTLock Set if calling from rxi_ReapConnections
5341 #ifdef RX_ENABLE_LOCKS
5343 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5344 #else /* RX_ENABLE_LOCKS */
5346 rxi_CheckCall(register struct rx_call *call)
5347 #endif /* RX_ENABLE_LOCKS */
5349 register struct rx_connection *conn = call->conn;
5351 afs_uint32 deadTime;
5353 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5354 if (call->flags & RX_CALL_TQ_BUSY) {
5355 /* Call is active and will be reset by rxi_Start if it's
5356 * in an error state.
5361 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5363 (((afs_uint32) conn->secondsUntilDead << 10) +
5364 ((afs_uint32) conn->peer->rtt >> 3) +
5365 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5367 /* These are computed to the second (+- 1 second). But that's
5368 * good enough for these values, which should be a significant
5369 * number of seconds. */
5370 if (now > (call->lastReceiveTime + deadTime)) {
5371 if (call->state == RX_STATE_ACTIVE) {
5372 rxi_CallError(call, RX_CALL_DEAD);
5375 #ifdef RX_ENABLE_LOCKS
5376 /* Cancel pending events */
5377 rxevent_Cancel(call->delayedAckEvent, call,
5378 RX_CALL_REFCOUNT_DELAY);
5379 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5380 rxevent_Cancel(call->keepAliveEvent, call,
5381 RX_CALL_REFCOUNT_ALIVE);
5382 if (call->refCount == 0) {
5383 rxi_FreeCall(call, haveCTLock);
5387 #else /* RX_ENABLE_LOCKS */
5390 #endif /* RX_ENABLE_LOCKS */
5392 /* Non-active calls are destroyed if they are not responding
5393 * to pings; active calls are simply flagged in error, so the
5394 * attached process can die reasonably gracefully. */
5396 /* see if we have a non-activity timeout */
5397 if (call->startWait && conn->idleDeadTime
5398 && ((call->startWait + conn->idleDeadTime) < now)) {
5399 if (call->state == RX_STATE_ACTIVE) {
5400 rxi_CallError(call, RX_CALL_TIMEOUT);
5404 /* see if we have a hard timeout */
5405 if (conn->hardDeadTime
5406 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5407 if (call->state == RX_STATE_ACTIVE)
5408 rxi_CallError(call, RX_CALL_TIMEOUT);
5415 /* When a call is in progress, this routine is called occasionally to
5416 * make sure that some traffic has arrived (or been sent to) the peer.
5417 * If nothing has arrived in a reasonable amount of time, the call is
5418 * declared dead; if nothing has been sent for a while, we send a
5419 * keep-alive packet (if we're actually trying to keep the call alive)
5422 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5425 struct rx_connection *conn;
5428 MUTEX_ENTER(&call->lock);
5429 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5430 if (event == call->keepAliveEvent)
5431 call->keepAliveEvent = NULL;
5434 #ifdef RX_ENABLE_LOCKS
5435 if (rxi_CheckCall(call, 0)) {
5436 MUTEX_EXIT(&call->lock);
5439 #else /* RX_ENABLE_LOCKS */
5440 if (rxi_CheckCall(call))
5442 #endif /* RX_ENABLE_LOCKS */
5444 /* Don't try to keep alive dallying calls */
5445 if (call->state == RX_STATE_DALLY) {
5446 MUTEX_EXIT(&call->lock);
5451 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5452 /* Don't try to send keepalives if there is unacknowledged data */
5453 /* the rexmit code should be good enough, this little hack
5454 * doesn't quite work XXX */
5455 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5457 rxi_ScheduleKeepAliveEvent(call);
5458 MUTEX_EXIT(&call->lock);
5463 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5465 if (!call->keepAliveEvent) {
5467 clock_GetTime(&when);
5468 when.sec += call->conn->secondsUntilPing;
5469 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5470 call->keepAliveEvent =
5471 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5475 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5477 rxi_KeepAliveOn(register struct rx_call *call)
5479 /* Pretend last packet received was received now--i.e. if another
5480 * packet isn't received within the keep alive time, then the call
5481 * will die; Initialize last send time to the current time--even
5482 * if a packet hasn't been sent yet. This will guarantee that a
5483 * keep-alive is sent within the ping time */
5484 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5485 rxi_ScheduleKeepAliveEvent(call);
5488 /* This routine is called to send connection abort messages
5489 * that have been delayed to throttle looping clients. */
5491 rxi_SendDelayedConnAbort(struct rxevent *event,
5492 register struct rx_connection *conn, char *dummy)
5495 struct rx_packet *packet;
5497 MUTEX_ENTER(&conn->conn_data_lock);
5498 conn->delayedAbortEvent = NULL;
5499 error = htonl(conn->error);
5501 MUTEX_EXIT(&conn->conn_data_lock);
5502 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5505 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5506 RX_PACKET_TYPE_ABORT, (char *)&error,
5508 rxi_FreePacket(packet);
5512 /* This routine is called to send call abort messages
5513 * that have been delayed to throttle looping clients. */
5515 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5519 struct rx_packet *packet;
5521 MUTEX_ENTER(&call->lock);
5522 call->delayedAbortEvent = NULL;
5523 error = htonl(call->error);
5525 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5528 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5529 (char *)&error, sizeof(error), 0);
5530 rxi_FreePacket(packet);
5532 MUTEX_EXIT(&call->lock);
5535 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5536 * seconds) to ask the client to authenticate itself. The routine
5537 * issues a challenge to the client, which is obtained from the
5538 * security object associated with the connection */
5540 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5541 void *arg1, int tries)
5543 conn->challengeEvent = NULL;
5544 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5545 register struct rx_packet *packet;
5549 /* We've failed to authenticate for too long.
5550 * Reset any calls waiting for authentication;
5551 * they are all in RX_STATE_PRECALL.
5555 MUTEX_ENTER(&conn->conn_call_lock);
5556 for (i = 0; i < RX_MAXCALLS; i++) {
5557 struct rx_call *call = conn->call[i];
5559 MUTEX_ENTER(&call->lock);
5560 if (call->state == RX_STATE_PRECALL) {
5561 rxi_CallError(call, RX_CALL_DEAD);
5562 rxi_SendCallAbort(call, NULL, 0, 0);
5564 MUTEX_EXIT(&call->lock);
5567 MUTEX_EXIT(&conn->conn_call_lock);
5571 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5573 /* If there's no packet available, do this later. */
5574 RXS_GetChallenge(conn->securityObject, conn, packet);
5575 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5576 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5577 rxi_FreePacket(packet);
5579 clock_GetTime(&when);
5580 when.sec += RX_CHALLENGE_TIMEOUT;
5581 conn->challengeEvent =
5582 rxevent_Post2(&when, rxi_ChallengeEvent, conn, 0,
5587 /* Call this routine to start requesting the client to authenticate
5588 * itself. This will continue until authentication is established,
5589 * the call times out, or an invalid response is returned. The
5590 * security object associated with the connection is asked to create
5591 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5592 * defined earlier. */
5594 rxi_ChallengeOn(register struct rx_connection *conn)
5596 if (!conn->challengeEvent) {
5597 RXS_CreateChallenge(conn->securityObject, conn);
5598 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5603 /* Compute round trip time of the packet provided, in *rttp.
5606 /* rxi_ComputeRoundTripTime is called with peer locked. */
5607 /* sentp and/or peer may be null */
5609 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5610 register struct clock *sentp,
5611 register struct rx_peer *peer)
5613 struct clock thisRtt, *rttp = &thisRtt;
5615 register int rtt_timeout;
5617 clock_GetTime(rttp);
5619 if (clock_Lt(rttp, sentp)) {
5621 return; /* somebody set the clock back, don't count this time. */
5623 clock_Sub(rttp, sentp);
5624 MUTEX_ENTER(&rx_stats_mutex);
5625 if (clock_Lt(rttp, &rx_stats.minRtt))
5626 rx_stats.minRtt = *rttp;
5627 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5628 if (rttp->sec > 60) {
5629 MUTEX_EXIT(&rx_stats_mutex);
5630 return; /* somebody set the clock ahead */
5632 rx_stats.maxRtt = *rttp;
5634 clock_Add(&rx_stats.totalRtt, rttp);
5635 rx_stats.nRttSamples++;
5636 MUTEX_EXIT(&rx_stats_mutex);
5638 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5640 /* Apply VanJacobson round-trip estimations */
5645 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5646 * srtt is stored as fixed point with 3 bits after the binary
5647 * point (i.e., scaled by 8). The following magic is
5648 * equivalent to the smoothing algorithm in rfc793 with an
5649 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5650 * srtt*8 = srtt*8 + rtt - srtt
5651 * srtt = srtt + rtt/8 - srtt/8
5654 delta = MSEC(rttp) - (peer->rtt >> 3);
5658 * We accumulate a smoothed rtt variance (actually, a smoothed
5659 * mean difference), then set the retransmit timer to smoothed
5660 * rtt + 4 times the smoothed variance (was 2x in van's original
5661 * paper, but 4x works better for me, and apparently for him as
5663 * rttvar is stored as
5664 * fixed point with 2 bits after the binary point (scaled by
5665 * 4). The following is equivalent to rfc793 smoothing with
5666 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5667 * replaces rfc793's wired-in beta.
5668 * dev*4 = dev*4 + (|actual - expected| - dev)
5674 delta -= (peer->rtt_dev >> 2);
5675 peer->rtt_dev += delta;
5677 /* I don't have a stored RTT so I start with this value. Since I'm
5678 * probably just starting a call, and will be pushing more data down
5679 * this, I expect congestion to increase rapidly. So I fudge a
5680 * little, and I set deviance to half the rtt. In practice,
5681 * deviance tends to approach something a little less than
5682 * half the smoothed rtt. */
5683 peer->rtt = (MSEC(rttp) << 3) + 8;
5684 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5686 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5687 * the other of these connections is usually in a user process, and can
5688 * be switched and/or swapped out. So on fast, reliable networks, the
5689 * timeout would otherwise be too short.
5691 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5692 clock_Zero(&(peer->timeout));
5693 clock_Addmsec(&(peer->timeout), rtt_timeout);
5695 dpf(("rxi_ComputeRoundTripTime(rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%0.3d sec)\n", MSEC(rttp), peer->rtt >> 3, peer->rtt_dev >> 2, (peer->timeout.sec), (peer->timeout.usec)));
5699 /* Find all server connections that have not been active for a long time, and
5702 rxi_ReapConnections(void)
5705 clock_GetTime(&now);
5707 /* Find server connection structures that haven't been used for
5708 * greater than rx_idleConnectionTime */
5710 struct rx_connection **conn_ptr, **conn_end;
5711 int i, havecalls = 0;
5712 MUTEX_ENTER(&rx_connHashTable_lock);
5713 for (conn_ptr = &rx_connHashTable[0], conn_end =
5714 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5716 struct rx_connection *conn, *next;
5717 struct rx_call *call;
5721 for (conn = *conn_ptr; conn; conn = next) {
5722 /* XXX -- Shouldn't the connection be locked? */
5725 for (i = 0; i < RX_MAXCALLS; i++) {
5726 call = conn->call[i];
5729 MUTEX_ENTER(&call->lock);
5730 #ifdef RX_ENABLE_LOCKS
5731 result = rxi_CheckCall(call, 1);
5732 #else /* RX_ENABLE_LOCKS */
5733 result = rxi_CheckCall(call);
5734 #endif /* RX_ENABLE_LOCKS */
5735 MUTEX_EXIT(&call->lock);
5737 /* If CheckCall freed the call, it might
5738 * have destroyed the connection as well,
5739 * which screws up the linked lists.
5745 if (conn->type == RX_SERVER_CONNECTION) {
5746 /* This only actually destroys the connection if
5747 * there are no outstanding calls */
5748 MUTEX_ENTER(&conn->conn_data_lock);
5749 if (!havecalls && !conn->refCount
5750 && ((conn->lastSendTime + rx_idleConnectionTime) <
5752 conn->refCount++; /* it will be decr in rx_DestroyConn */
5753 MUTEX_EXIT(&conn->conn_data_lock);
5754 #ifdef RX_ENABLE_LOCKS
5755 rxi_DestroyConnectionNoLock(conn);
5756 #else /* RX_ENABLE_LOCKS */
5757 rxi_DestroyConnection(conn);
5758 #endif /* RX_ENABLE_LOCKS */
5760 #ifdef RX_ENABLE_LOCKS
5762 MUTEX_EXIT(&conn->conn_data_lock);
5764 #endif /* RX_ENABLE_LOCKS */
5768 #ifdef RX_ENABLE_LOCKS
5769 while (rx_connCleanup_list) {
5770 struct rx_connection *conn;
5771 conn = rx_connCleanup_list;
5772 rx_connCleanup_list = rx_connCleanup_list->next;
5773 MUTEX_EXIT(&rx_connHashTable_lock);
5774 rxi_CleanupConnection(conn);
5775 MUTEX_ENTER(&rx_connHashTable_lock);
5777 MUTEX_EXIT(&rx_connHashTable_lock);
5778 #endif /* RX_ENABLE_LOCKS */
5781 /* Find any peer structures that haven't been used (haven't had an
5782 * associated connection) for greater than rx_idlePeerTime */
5784 struct rx_peer **peer_ptr, **peer_end;
5786 MUTEX_ENTER(&rx_rpc_stats);
5787 MUTEX_ENTER(&rx_peerHashTable_lock);
5788 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5789 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5791 struct rx_peer *peer, *next, *prev;
5792 for (prev = peer = *peer_ptr; peer; peer = next) {
5794 code = MUTEX_TRYENTER(&peer->peer_lock);
5795 if ((code) && (peer->refCount == 0)
5796 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5797 rx_interface_stat_p rpc_stat, nrpc_stat;
5799 MUTEX_EXIT(&peer->peer_lock);
5800 MUTEX_DESTROY(&peer->peer_lock);
5802 (&peer->rpcStats, rpc_stat, nrpc_stat,
5803 rx_interface_stat)) {
5804 unsigned int num_funcs;
5807 queue_Remove(&rpc_stat->queue_header);
5808 queue_Remove(&rpc_stat->all_peers);
5809 num_funcs = rpc_stat->stats[0].func_total;
5811 sizeof(rx_interface_stat_t) +
5812 rpc_stat->stats[0].func_total *
5813 sizeof(rx_function_entry_v1_t);
5815 rxi_Free(rpc_stat, space);
5816 rxi_rpc_peer_stat_cnt -= num_funcs;
5819 MUTEX_ENTER(&rx_stats_mutex);
5820 rx_stats.nPeerStructs--;
5821 MUTEX_EXIT(&rx_stats_mutex);
5822 if (peer == *peer_ptr) {
5829 MUTEX_EXIT(&peer->peer_lock);
5835 MUTEX_EXIT(&rx_peerHashTable_lock);
5836 MUTEX_EXIT(&rx_rpc_stats);
5839 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5840 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5841 * GC, just below. Really, we shouldn't have to keep moving packets from
5842 * one place to another, but instead ought to always know if we can
5843 * afford to hold onto a packet in its particular use. */
5844 MUTEX_ENTER(&rx_freePktQ_lock);
5845 if (rx_waitingForPackets) {
5846 rx_waitingForPackets = 0;
5847 #ifdef RX_ENABLE_LOCKS
5848 CV_BROADCAST(&rx_waitingForPackets_cv);
5850 osi_rxWakeup(&rx_waitingForPackets);
5853 MUTEX_EXIT(&rx_freePktQ_lock);
5855 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5856 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5860 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5861 * rx.h is sort of strange this is better. This is called with a security
5862 * object before it is discarded. Each connection using a security object has
5863 * its own refcount to the object so it won't actually be freed until the last
5864 * connection is destroyed.
5866 * This is the only rxs module call. A hold could also be written but no one
5870 rxs_Release(struct rx_securityClass *aobj)
5872 return RXS_Close(aobj);
5876 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5877 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5878 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5879 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5881 /* Adjust our estimate of the transmission rate to this peer, given
5882 * that the packet p was just acked. We can adjust peer->timeout and
5883 * call->twind. Pragmatically, this is called
5884 * only with packets of maximal length.
5885 * Called with peer and call locked.
5889 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5890 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5892 afs_int32 xferSize, xferMs;
5893 register afs_int32 minTime;
5896 /* Count down packets */
5897 if (peer->rateFlag > 0)
5899 /* Do nothing until we're enabled */
5900 if (peer->rateFlag != 0)
5905 /* Count only when the ack seems legitimate */
5906 switch (ackReason) {
5907 case RX_ACK_REQUESTED:
5909 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
5913 case RX_ACK_PING_RESPONSE:
5914 if (p) /* want the response to ping-request, not data send */
5916 clock_GetTime(&newTO);
5917 if (clock_Gt(&newTO, &call->pingRequestTime)) {
5918 clock_Sub(&newTO, &call->pingRequestTime);
5919 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
5923 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
5930 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %lu.%06lu, rtt %u, ps %u)", ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"), xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
5932 /* Track only packets that are big enough. */
5933 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
5937 /* absorb RTT data (in milliseconds) for these big packets */
5938 if (peer->smRtt == 0) {
5939 peer->smRtt = xferMs;
5941 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
5946 if (peer->countDown) {
5950 peer->countDown = 10; /* recalculate only every so often */
5952 /* In practice, we can measure only the RTT for full packets,
5953 * because of the way Rx acks the data that it receives. (If it's
5954 * smaller than a full packet, it often gets implicitly acked
5955 * either by the call response (from a server) or by the next call
5956 * (from a client), and either case confuses transmission times
5957 * with processing times.) Therefore, replace the above
5958 * more-sophisticated processing with a simpler version, where the
5959 * smoothed RTT is kept for full-size packets, and the time to
5960 * transmit a windowful of full-size packets is simply RTT *
5961 * windowSize. Again, we take two steps:
5962 - ensure the timeout is large enough for a single packet's RTT;
5963 - ensure that the window is small enough to fit in the desired timeout.*/
5965 /* First, the timeout check. */
5966 minTime = peer->smRtt;
5967 /* Get a reasonable estimate for a timeout period */
5969 newTO.sec = minTime / 1000;
5970 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
5972 /* Increase the timeout period so that we can always do at least
5973 * one packet exchange */
5974 if (clock_Gt(&newTO, &peer->timeout)) {
5976 dpf(("CONG peer %lx/%u: timeout %lu.%06lu ==> %lu.%06lu (rtt %u, ps %u)", ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec, newTO.sec, newTO.usec, peer->smRtt, peer->packetSize));
5978 peer->timeout = newTO;
5981 /* Now, get an estimate for the transmit window size. */
5982 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
5983 /* Now, convert to the number of full packets that could fit in a
5984 * reasonable fraction of that interval */
5985 minTime /= (peer->smRtt << 1);
5986 xferSize = minTime; /* (make a copy) */
5988 /* Now clamp the size to reasonable bounds. */
5991 else if (minTime > rx_Window)
5992 minTime = rx_Window;
5993 /* if (minTime != peer->maxWindow) {
5994 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
5995 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
5996 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
5998 peer->maxWindow = minTime;
5999 elide... call->twind = minTime;
6003 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6004 * Discern this by calculating the timeout necessary for rx_Window
6006 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6007 /* calculate estimate for transmission interval in milliseconds */
6008 minTime = rx_Window * peer->smRtt;
6009 if (minTime < 1000) {
6010 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6011 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6012 peer->timeout.usec, peer->smRtt, peer->packetSize));
6014 newTO.sec = 0; /* cut back on timeout by half a second */
6015 newTO.usec = 500000;
6016 clock_Sub(&peer->timeout, &newTO);
6021 } /* end of rxi_ComputeRate */
6022 #endif /* ADAPT_WINDOW */
6030 /* Don't call this debugging routine directly; use dpf */
6032 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
6033 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
6040 len = _snprintf(msg, sizeof(msg)-2,
6041 format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
6042 a11, a12, a13, a14, a15);
6044 if (msg[len-1] != '\n') {
6048 OutputDebugString(msg);
6052 clock_GetTime(&now);
6053 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6054 (unsigned int)now.usec / 1000);
6055 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6062 * This function is used to process the rx_stats structure that is local
6063 * to a process as well as an rx_stats structure received from a remote
6064 * process (via rxdebug). Therefore, it needs to do minimal version
6068 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6069 afs_int32 freePackets, char version)
6073 if (size != sizeof(struct rx_stats)) {
6075 "Unexpected size of stats structure: was %d, expected %d\n",
6076 size, sizeof(struct rx_stats));
6079 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6082 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6083 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6084 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6085 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6086 s->specialPktAllocFailures);
6088 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6089 s->receivePktAllocFailures, s->sendPktAllocFailures,
6090 s->specialPktAllocFailures);
6094 " greedy %d, " "bogusReads %d (last from host %x), "
6095 "noPackets %d, " "noBuffers %d, " "selects %d, "
6096 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6097 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6098 s->selects, s->sendSelects);
6100 fprintf(file, " packets read: ");
6101 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6102 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6104 fprintf(file, "\n");
6107 " other read counters: data %d, " "ack %d, " "dup %d "
6108 "spurious %d " "dally %d\n", s->dataPacketsRead,
6109 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6110 s->ignorePacketDally);
6112 fprintf(file, " packets sent: ");
6113 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6114 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6116 fprintf(file, "\n");
6119 " other send counters: ack %d, " "data %d (not resends), "
6120 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6121 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6122 s->dataPacketsPushed, s->ignoreAckedPacket);
6125 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6126 s->netSendFailures, (int)s->fatalErrors);
6128 if (s->nRttSamples) {
6129 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6130 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6132 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6133 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6137 " %d server connections, " "%d client connections, "
6138 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6139 s->nServerConns, s->nClientConns, s->nPeerStructs,
6140 s->nCallStructs, s->nFreeCallStructs);
6142 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6143 fprintf(file, " %d clock updates\n", clock_nUpdates);
6148 /* for backward compatibility */
6150 rx_PrintStats(FILE * file)
6152 MUTEX_ENTER(&rx_stats_mutex);
6153 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6155 MUTEX_EXIT(&rx_stats_mutex);
6159 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6161 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6162 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6163 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6166 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6167 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6168 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6171 " Packet size %d, " "max in packet skew %d, "
6172 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6173 (int)peer->outPacketSkew);
6176 #ifdef AFS_PTHREAD_ENV
6178 * This mutex protects the following static variables:
6182 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6183 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6185 #define LOCK_RX_DEBUG
6186 #define UNLOCK_RX_DEBUG
6187 #endif /* AFS_PTHREAD_ENV */
6190 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6191 u_char type, void *inputData, size_t inputLength,
6192 void *outputData, size_t outputLength)
6194 static afs_int32 counter = 100;
6196 struct rx_header theader;
6198 register afs_int32 code;
6200 struct sockaddr_in taddr, faddr;
6205 endTime = time(0) + 20; /* try for 20 seconds */
6209 tp = &tbuffer[sizeof(struct rx_header)];
6210 taddr.sin_family = AF_INET;
6211 taddr.sin_port = remotePort;
6212 taddr.sin_addr.s_addr = remoteAddr;
6213 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6214 taddr.sin_len = sizeof(struct sockaddr_in);
6217 memset(&theader, 0, sizeof(theader));
6218 theader.epoch = htonl(999);
6220 theader.callNumber = htonl(counter);
6223 theader.type = type;
6224 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6225 theader.serviceId = 0;
6227 memcpy(tbuffer, &theader, sizeof(theader));
6228 memcpy(tp, inputData, inputLength);
6230 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6231 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6233 /* see if there's a packet available */
6235 FD_SET(socket, &imask);
6238 code = select(socket + 1, &imask, 0, 0, &tv);
6239 if (code == 1 && FD_ISSET(socket, &imask)) {
6240 /* now receive a packet */
6241 faddrLen = sizeof(struct sockaddr_in);
6243 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6244 (struct sockaddr *)&faddr, &faddrLen);
6247 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6248 if (counter == ntohl(theader.callNumber))
6253 /* see if we've timed out */
6254 if (endTime < time(0))
6257 code -= sizeof(struct rx_header);
6258 if (code > outputLength)
6259 code = outputLength;
6260 memcpy(outputData, tp, code);
6265 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6266 afs_uint16 remotePort, struct rx_debugStats * stat,
6267 afs_uint32 * supportedValues)
6269 struct rx_debugIn in;
6272 *supportedValues = 0;
6273 in.type = htonl(RX_DEBUGI_GETSTATS);
6276 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6277 &in, sizeof(in), stat, sizeof(*stat));
6280 * If the call was successful, fixup the version and indicate
6281 * what contents of the stat structure are valid.
6282 * Also do net to host conversion of fields here.
6286 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6287 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6289 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6290 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6292 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6293 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6295 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6296 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6298 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6299 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6301 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6302 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6304 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6305 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6307 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6308 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6311 stat->nFreePackets = ntohl(stat->nFreePackets);
6312 stat->packetReclaims = ntohl(stat->packetReclaims);
6313 stat->callsExecuted = ntohl(stat->callsExecuted);
6314 stat->nWaiting = ntohl(stat->nWaiting);
6315 stat->idleThreads = ntohl(stat->idleThreads);
6322 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6323 afs_uint16 remotePort, struct rx_stats * stat,
6324 afs_uint32 * supportedValues)
6326 struct rx_debugIn in;
6327 afs_int32 *lp = (afs_int32 *) stat;
6332 * supportedValues is currently unused, but added to allow future
6333 * versioning of this function.
6336 *supportedValues = 0;
6337 in.type = htonl(RX_DEBUGI_RXSTATS);
6339 memset(stat, 0, sizeof(*stat));
6341 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6342 &in, sizeof(in), stat, sizeof(*stat));
6347 * Do net to host conversion here
6350 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6359 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6360 afs_uint16 remotePort, size_t version_length,
6364 return MakeDebugCall(socket, remoteAddr, remotePort,
6365 RX_PACKET_TYPE_VERSION, a, 1, version,
6370 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6371 afs_uint16 remotePort, afs_int32 * nextConnection,
6372 int allConnections, afs_uint32 debugSupportedValues,
6373 struct rx_debugConn * conn,
6374 afs_uint32 * supportedValues)
6376 struct rx_debugIn in;
6381 * supportedValues is currently unused, but added to allow future
6382 * versioning of this function.
6385 *supportedValues = 0;
6386 if (allConnections) {
6387 in.type = htonl(RX_DEBUGI_GETALLCONN);
6389 in.type = htonl(RX_DEBUGI_GETCONN);
6391 in.index = htonl(*nextConnection);
6392 memset(conn, 0, sizeof(*conn));
6394 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6395 &in, sizeof(in), conn, sizeof(*conn));
6398 *nextConnection += 1;
6401 * Convert old connection format to new structure.
6404 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6405 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6406 #define MOVEvL(a) (conn->a = vL->a)
6408 /* any old or unrecognized version... */
6409 for (i = 0; i < RX_MAXCALLS; i++) {
6410 MOVEvL(callState[i]);
6411 MOVEvL(callMode[i]);
6412 MOVEvL(callFlags[i]);
6413 MOVEvL(callOther[i]);
6415 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6416 MOVEvL(secStats.type);
6417 MOVEvL(secStats.level);
6418 MOVEvL(secStats.flags);
6419 MOVEvL(secStats.expires);
6420 MOVEvL(secStats.packetsReceived);
6421 MOVEvL(secStats.packetsSent);
6422 MOVEvL(secStats.bytesReceived);
6423 MOVEvL(secStats.bytesSent);
6428 * Do net to host conversion here
6430 * I don't convert host or port since we are most likely
6431 * going to want these in NBO.
6433 conn->cid = ntohl(conn->cid);
6434 conn->serial = ntohl(conn->serial);
6435 for (i = 0; i < RX_MAXCALLS; i++) {
6436 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6438 conn->error = ntohl(conn->error);
6439 conn->secStats.flags = ntohl(conn->secStats.flags);
6440 conn->secStats.expires = ntohl(conn->secStats.expires);
6441 conn->secStats.packetsReceived =
6442 ntohl(conn->secStats.packetsReceived);
6443 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6444 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6445 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6446 conn->epoch = ntohl(conn->epoch);
6447 conn->natMTU = ntohl(conn->natMTU);
6454 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6455 afs_uint16 remotePort, afs_int32 * nextPeer,
6456 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6457 afs_uint32 * supportedValues)
6459 struct rx_debugIn in;
6463 * supportedValues is currently unused, but added to allow future
6464 * versioning of this function.
6467 *supportedValues = 0;
6468 in.type = htonl(RX_DEBUGI_GETPEER);
6469 in.index = htonl(*nextPeer);
6470 memset(peer, 0, sizeof(*peer));
6472 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6473 &in, sizeof(in), peer, sizeof(*peer));
6479 * Do net to host conversion here
6481 * I don't convert host or port since we are most likely
6482 * going to want these in NBO.
6484 peer->ifMTU = ntohs(peer->ifMTU);
6485 peer->idleWhen = ntohl(peer->idleWhen);
6486 peer->refCount = ntohs(peer->refCount);
6487 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6488 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6489 peer->rtt = ntohl(peer->rtt);
6490 peer->rtt_dev = ntohl(peer->rtt_dev);
6491 peer->timeout.sec = ntohl(peer->timeout.sec);
6492 peer->timeout.usec = ntohl(peer->timeout.usec);
6493 peer->nSent = ntohl(peer->nSent);
6494 peer->reSends = ntohl(peer->reSends);
6495 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6496 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6497 peer->rateFlag = ntohl(peer->rateFlag);
6498 peer->natMTU = ntohs(peer->natMTU);
6499 peer->maxMTU = ntohs(peer->maxMTU);
6500 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6501 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6502 peer->MTU = ntohs(peer->MTU);
6503 peer->cwind = ntohs(peer->cwind);
6504 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6505 peer->congestSeq = ntohs(peer->congestSeq);
6506 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6507 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6508 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6509 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6514 #endif /* RXDEBUG */
6519 struct rx_serverQueueEntry *np;
6522 register struct rx_call *call;
6523 register struct rx_serverQueueEntry *sq;
6527 if (rxinit_status == 1) {
6529 return; /* Already shutdown. */
6533 #ifndef AFS_PTHREAD_ENV
6534 FD_ZERO(&rx_selectMask);
6535 #endif /* AFS_PTHREAD_ENV */
6536 rxi_dataQuota = RX_MAX_QUOTA;
6537 #ifndef AFS_PTHREAD_ENV
6539 #endif /* AFS_PTHREAD_ENV */
6542 #ifndef AFS_PTHREAD_ENV
6543 #ifndef AFS_USE_GETTIMEOFDAY
6545 #endif /* AFS_USE_GETTIMEOFDAY */
6546 #endif /* AFS_PTHREAD_ENV */
6548 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6549 call = queue_First(&rx_freeCallQueue, rx_call);
6551 rxi_Free(call, sizeof(struct rx_call));
6554 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6555 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6561 struct rx_peer **peer_ptr, **peer_end;
6562 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6563 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6565 struct rx_peer *peer, *next;
6566 for (peer = *peer_ptr; peer; peer = next) {
6567 rx_interface_stat_p rpc_stat, nrpc_stat;
6570 (&peer->rpcStats, rpc_stat, nrpc_stat,
6571 rx_interface_stat)) {
6572 unsigned int num_funcs;
6575 queue_Remove(&rpc_stat->queue_header);
6576 queue_Remove(&rpc_stat->all_peers);
6577 num_funcs = rpc_stat->stats[0].func_total;
6579 sizeof(rx_interface_stat_t) +
6580 rpc_stat->stats[0].func_total *
6581 sizeof(rx_function_entry_v1_t);
6583 rxi_Free(rpc_stat, space);
6584 MUTEX_ENTER(&rx_rpc_stats);
6585 rxi_rpc_peer_stat_cnt -= num_funcs;
6586 MUTEX_EXIT(&rx_rpc_stats);
6590 MUTEX_ENTER(&rx_stats_mutex);
6591 rx_stats.nPeerStructs--;
6592 MUTEX_EXIT(&rx_stats_mutex);
6596 for (i = 0; i < RX_MAX_SERVICES; i++) {
6598 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6600 for (i = 0; i < rx_hashTableSize; i++) {
6601 register struct rx_connection *tc, *ntc;
6602 MUTEX_ENTER(&rx_connHashTable_lock);
6603 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6605 for (j = 0; j < RX_MAXCALLS; j++) {
6607 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6610 rxi_Free(tc, sizeof(*tc));
6612 MUTEX_EXIT(&rx_connHashTable_lock);
6615 MUTEX_ENTER(&freeSQEList_lock);
6617 while ((np = rx_FreeSQEList)) {
6618 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6619 MUTEX_DESTROY(&np->lock);
6620 rxi_Free(np, sizeof(*np));
6623 MUTEX_EXIT(&freeSQEList_lock);
6624 MUTEX_DESTROY(&freeSQEList_lock);
6625 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6626 MUTEX_DESTROY(&rx_connHashTable_lock);
6627 MUTEX_DESTROY(&rx_peerHashTable_lock);
6628 MUTEX_DESTROY(&rx_serverPool_lock);
6630 osi_Free(rx_connHashTable,
6631 rx_hashTableSize * sizeof(struct rx_connection *));
6632 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6634 UNPIN(rx_connHashTable,
6635 rx_hashTableSize * sizeof(struct rx_connection *));
6636 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6638 rxi_FreeAllPackets();
6640 MUTEX_ENTER(&rx_stats_mutex);
6641 rxi_dataQuota = RX_MAX_QUOTA;
6642 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6643 MUTEX_EXIT(&rx_stats_mutex);
6649 #ifdef RX_ENABLE_LOCKS
6651 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6653 if (!MUTEX_ISMINE(lockaddr))
6654 osi_Panic("Lock not held: %s", msg);
6656 #endif /* RX_ENABLE_LOCKS */
6661 * Routines to implement connection specific data.
6665 rx_KeyCreate(rx_destructor_t rtn)
6668 MUTEX_ENTER(&rxi_keyCreate_lock);
6669 key = rxi_keyCreate_counter++;
6670 rxi_keyCreate_destructor = (rx_destructor_t *)
6671 realloc((void *)rxi_keyCreate_destructor,
6672 (key + 1) * sizeof(rx_destructor_t));
6673 rxi_keyCreate_destructor[key] = rtn;
6674 MUTEX_EXIT(&rxi_keyCreate_lock);
6679 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6682 MUTEX_ENTER(&conn->conn_data_lock);
6683 if (!conn->specific) {
6684 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6685 for (i = 0; i < key; i++)
6686 conn->specific[i] = NULL;
6687 conn->nSpecific = key + 1;
6688 conn->specific[key] = ptr;
6689 } else if (key >= conn->nSpecific) {
6690 conn->specific = (void **)
6691 realloc(conn->specific, (key + 1) * sizeof(void *));
6692 for (i = conn->nSpecific; i < key; i++)
6693 conn->specific[i] = NULL;
6694 conn->nSpecific = key + 1;
6695 conn->specific[key] = ptr;
6697 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6698 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6699 conn->specific[key] = ptr;
6701 MUTEX_EXIT(&conn->conn_data_lock);
6705 rx_GetSpecific(struct rx_connection *conn, int key)
6708 MUTEX_ENTER(&conn->conn_data_lock);
6709 if (key >= conn->nSpecific)
6712 ptr = conn->specific[key];
6713 MUTEX_EXIT(&conn->conn_data_lock);
6717 #endif /* !KERNEL */
6720 * processStats is a queue used to store the statistics for the local
6721 * process. Its contents are similar to the contents of the rpcStats
6722 * queue on a rx_peer structure, but the actual data stored within
6723 * this queue contains totals across the lifetime of the process (assuming
6724 * the stats have not been reset) - unlike the per peer structures
6725 * which can come and go based upon the peer lifetime.
6728 static struct rx_queue processStats = { &processStats, &processStats };
6731 * peerStats is a queue used to store the statistics for all peer structs.
6732 * Its contents are the union of all the peer rpcStats queues.
6735 static struct rx_queue peerStats = { &peerStats, &peerStats };
6738 * rxi_monitor_processStats is used to turn process wide stat collection
6742 static int rxi_monitor_processStats = 0;
6745 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6748 static int rxi_monitor_peerStats = 0;
6751 * rxi_AddRpcStat - given all of the information for a particular rpc
6752 * call, create (if needed) and update the stat totals for the rpc.
6756 * IN stats - the queue of stats that will be updated with the new value
6758 * IN rxInterface - a unique number that identifies the rpc interface
6760 * IN currentFunc - the index of the function being invoked
6762 * IN totalFunc - the total number of functions in this interface
6764 * IN queueTime - the amount of time this function waited for a thread
6766 * IN execTime - the amount of time this function invocation took to execute
6768 * IN bytesSent - the number bytes sent by this invocation
6770 * IN bytesRcvd - the number bytes received by this invocation
6772 * IN isServer - if true, this invocation was made to a server
6774 * IN remoteHost - the ip address of the remote host
6776 * IN remotePort - the port of the remote host
6778 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6780 * INOUT counter - if a new stats structure is allocated, the counter will
6781 * be updated with the new number of allocated stat structures
6789 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6790 afs_uint32 currentFunc, afs_uint32 totalFunc,
6791 struct clock *queueTime, struct clock *execTime,
6792 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6793 afs_uint32 remoteHost, afs_uint32 remotePort,
6794 int addToPeerList, unsigned int *counter)
6797 rx_interface_stat_p rpc_stat, nrpc_stat;
6800 * See if there's already a structure for this interface
6803 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6804 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6805 && (rpc_stat->stats[0].remote_is_server == isServer))
6810 * Didn't find a match so allocate a new structure and add it to the
6814 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6815 || (rpc_stat->stats[0].interfaceId != rxInterface)
6816 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6821 sizeof(rx_interface_stat_t) +
6822 totalFunc * sizeof(rx_function_entry_v1_t);
6824 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6825 if (rpc_stat == NULL) {
6829 *counter += totalFunc;
6830 for (i = 0; i < totalFunc; i++) {
6831 rpc_stat->stats[i].remote_peer = remoteHost;
6832 rpc_stat->stats[i].remote_port = remotePort;
6833 rpc_stat->stats[i].remote_is_server = isServer;
6834 rpc_stat->stats[i].interfaceId = rxInterface;
6835 rpc_stat->stats[i].func_total = totalFunc;
6836 rpc_stat->stats[i].func_index = i;
6837 hzero(rpc_stat->stats[i].invocations);
6838 hzero(rpc_stat->stats[i].bytes_sent);
6839 hzero(rpc_stat->stats[i].bytes_rcvd);
6840 rpc_stat->stats[i].queue_time_sum.sec = 0;
6841 rpc_stat->stats[i].queue_time_sum.usec = 0;
6842 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6843 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6844 rpc_stat->stats[i].queue_time_min.sec = 9999999;
6845 rpc_stat->stats[i].queue_time_min.usec = 9999999;
6846 rpc_stat->stats[i].queue_time_max.sec = 0;
6847 rpc_stat->stats[i].queue_time_max.usec = 0;
6848 rpc_stat->stats[i].execution_time_sum.sec = 0;
6849 rpc_stat->stats[i].execution_time_sum.usec = 0;
6850 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
6851 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
6852 rpc_stat->stats[i].execution_time_min.sec = 9999999;
6853 rpc_stat->stats[i].execution_time_min.usec = 9999999;
6854 rpc_stat->stats[i].execution_time_max.sec = 0;
6855 rpc_stat->stats[i].execution_time_max.usec = 0;
6857 queue_Prepend(stats, rpc_stat);
6858 if (addToPeerList) {
6859 queue_Prepend(&peerStats, &rpc_stat->all_peers);
6864 * Increment the stats for this function
6867 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
6868 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
6869 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
6870 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
6871 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
6872 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
6873 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
6875 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
6876 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
6878 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
6879 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
6881 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
6882 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
6884 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
6885 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
6893 * rx_IncrementTimeAndCount - increment the times and count for a particular
6898 * IN peer - the peer who invoked the rpc
6900 * IN rxInterface - a unique number that identifies the rpc interface
6902 * IN currentFunc - the index of the function being invoked
6904 * IN totalFunc - the total number of functions in this interface
6906 * IN queueTime - the amount of time this function waited for a thread
6908 * IN execTime - the amount of time this function invocation took to execute
6910 * IN bytesSent - the number bytes sent by this invocation
6912 * IN bytesRcvd - the number bytes received by this invocation
6914 * IN isServer - if true, this invocation was made to a server
6922 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
6923 afs_uint32 currentFunc, afs_uint32 totalFunc,
6924 struct clock *queueTime, struct clock *execTime,
6925 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
6929 MUTEX_ENTER(&rx_rpc_stats);
6930 MUTEX_ENTER(&peer->peer_lock);
6932 if (rxi_monitor_peerStats) {
6933 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
6934 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6935 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
6938 if (rxi_monitor_processStats) {
6939 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
6940 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6941 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
6944 MUTEX_EXIT(&peer->peer_lock);
6945 MUTEX_EXIT(&rx_rpc_stats);
6950 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
6954 * IN callerVersion - the rpc stat version of the caller.
6956 * IN count - the number of entries to marshall.
6958 * IN stats - pointer to stats to be marshalled.
6960 * OUT ptr - Where to store the marshalled data.
6967 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
6968 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
6974 * We only support the first version
6976 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
6977 *(ptr++) = stats->remote_peer;
6978 *(ptr++) = stats->remote_port;
6979 *(ptr++) = stats->remote_is_server;
6980 *(ptr++) = stats->interfaceId;
6981 *(ptr++) = stats->func_total;
6982 *(ptr++) = stats->func_index;
6983 *(ptr++) = hgethi(stats->invocations);
6984 *(ptr++) = hgetlo(stats->invocations);
6985 *(ptr++) = hgethi(stats->bytes_sent);
6986 *(ptr++) = hgetlo(stats->bytes_sent);
6987 *(ptr++) = hgethi(stats->bytes_rcvd);
6988 *(ptr++) = hgetlo(stats->bytes_rcvd);
6989 *(ptr++) = stats->queue_time_sum.sec;
6990 *(ptr++) = stats->queue_time_sum.usec;
6991 *(ptr++) = stats->queue_time_sum_sqr.sec;
6992 *(ptr++) = stats->queue_time_sum_sqr.usec;
6993 *(ptr++) = stats->queue_time_min.sec;
6994 *(ptr++) = stats->queue_time_min.usec;
6995 *(ptr++) = stats->queue_time_max.sec;
6996 *(ptr++) = stats->queue_time_max.usec;
6997 *(ptr++) = stats->execution_time_sum.sec;
6998 *(ptr++) = stats->execution_time_sum.usec;
6999 *(ptr++) = stats->execution_time_sum_sqr.sec;
7000 *(ptr++) = stats->execution_time_sum_sqr.usec;
7001 *(ptr++) = stats->execution_time_min.sec;
7002 *(ptr++) = stats->execution_time_min.usec;
7003 *(ptr++) = stats->execution_time_max.sec;
7004 *(ptr++) = stats->execution_time_max.usec;
7010 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7015 * IN callerVersion - the rpc stat version of the caller
7017 * OUT myVersion - the rpc stat version of this function
7019 * OUT clock_sec - local time seconds
7021 * OUT clock_usec - local time microseconds
7023 * OUT allocSize - the number of bytes allocated to contain stats
7025 * OUT statCount - the number stats retrieved from this process.
7027 * OUT stats - the actual stats retrieved from this process.
7031 * Returns void. If successful, stats will != NULL.
7035 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7036 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7037 size_t * allocSize, afs_uint32 * statCount,
7038 afs_uint32 ** stats)
7048 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7051 * Check to see if stats are enabled
7054 MUTEX_ENTER(&rx_rpc_stats);
7055 if (!rxi_monitor_processStats) {
7056 MUTEX_EXIT(&rx_rpc_stats);
7060 clock_GetTime(&now);
7061 *clock_sec = now.sec;
7062 *clock_usec = now.usec;
7065 * Allocate the space based upon the caller version
7067 * If the client is at an older version than we are,
7068 * we return the statistic data in the older data format, but
7069 * we still return our version number so the client knows we
7070 * are maintaining more data than it can retrieve.
7073 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7074 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7075 *statCount = rxi_rpc_process_stat_cnt;
7078 * This can't happen yet, but in the future version changes
7079 * can be handled by adding additional code here
7083 if (space > (size_t) 0) {
7085 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7088 rx_interface_stat_p rpc_stat, nrpc_stat;
7092 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7094 * Copy the data based upon the caller version
7096 rx_MarshallProcessRPCStats(callerVersion,
7097 rpc_stat->stats[0].func_total,
7098 rpc_stat->stats, &ptr);
7104 MUTEX_EXIT(&rx_rpc_stats);
7109 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7113 * IN callerVersion - the rpc stat version of the caller
7115 * OUT myVersion - the rpc stat version of this function
7117 * OUT clock_sec - local time seconds
7119 * OUT clock_usec - local time microseconds
7121 * OUT allocSize - the number of bytes allocated to contain stats
7123 * OUT statCount - the number of stats retrieved from the individual
7126 * OUT stats - the actual stats retrieved from the individual peer structures.
7130 * Returns void. If successful, stats will != NULL.
7134 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7135 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7136 size_t * allocSize, afs_uint32 * statCount,
7137 afs_uint32 ** stats)
7147 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7150 * Check to see if stats are enabled
7153 MUTEX_ENTER(&rx_rpc_stats);
7154 if (!rxi_monitor_peerStats) {
7155 MUTEX_EXIT(&rx_rpc_stats);
7159 clock_GetTime(&now);
7160 *clock_sec = now.sec;
7161 *clock_usec = now.usec;
7164 * Allocate the space based upon the caller version
7166 * If the client is at an older version than we are,
7167 * we return the statistic data in the older data format, but
7168 * we still return our version number so the client knows we
7169 * are maintaining more data than it can retrieve.
7172 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7173 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7174 *statCount = rxi_rpc_peer_stat_cnt;
7177 * This can't happen yet, but in the future version changes
7178 * can be handled by adding additional code here
7182 if (space > (size_t) 0) {
7184 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7187 rx_interface_stat_p rpc_stat, nrpc_stat;
7191 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7193 * We have to fix the offset of rpc_stat since we are
7194 * keeping this structure on two rx_queues. The rx_queue
7195 * package assumes that the rx_queue member is the first
7196 * member of the structure. That is, rx_queue assumes that
7197 * any one item is only on one queue at a time. We are
7198 * breaking that assumption and so we have to do a little
7199 * math to fix our pointers.
7202 fix_offset = (char *)rpc_stat;
7203 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7204 rpc_stat = (rx_interface_stat_p) fix_offset;
7207 * Copy the data based upon the caller version
7209 rx_MarshallProcessRPCStats(callerVersion,
7210 rpc_stat->stats[0].func_total,
7211 rpc_stat->stats, &ptr);
7217 MUTEX_EXIT(&rx_rpc_stats);
7222 * rx_FreeRPCStats - free memory allocated by
7223 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7227 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7228 * rx_RetrievePeerRPCStats
7230 * IN allocSize - the number of bytes in stats.
7238 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7240 rxi_Free(stats, allocSize);
7244 * rx_queryProcessRPCStats - see if process rpc stat collection is
7245 * currently enabled.
7251 * Returns 0 if stats are not enabled != 0 otherwise
7255 rx_queryProcessRPCStats(void)
7258 MUTEX_ENTER(&rx_rpc_stats);
7259 rc = rxi_monitor_processStats;
7260 MUTEX_EXIT(&rx_rpc_stats);
7265 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7271 * Returns 0 if stats are not enabled != 0 otherwise
7275 rx_queryPeerRPCStats(void)
7278 MUTEX_ENTER(&rx_rpc_stats);
7279 rc = rxi_monitor_peerStats;
7280 MUTEX_EXIT(&rx_rpc_stats);
7285 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7295 rx_enableProcessRPCStats(void)
7297 MUTEX_ENTER(&rx_rpc_stats);
7298 rx_enable_stats = 1;
7299 rxi_monitor_processStats = 1;
7300 MUTEX_EXIT(&rx_rpc_stats);
7304 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7314 rx_enablePeerRPCStats(void)
7316 MUTEX_ENTER(&rx_rpc_stats);
7317 rx_enable_stats = 1;
7318 rxi_monitor_peerStats = 1;
7319 MUTEX_EXIT(&rx_rpc_stats);
7323 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7333 rx_disableProcessRPCStats(void)
7335 rx_interface_stat_p rpc_stat, nrpc_stat;
7338 MUTEX_ENTER(&rx_rpc_stats);
7341 * Turn off process statistics and if peer stats is also off, turn
7345 rxi_monitor_processStats = 0;
7346 if (rxi_monitor_peerStats == 0) {
7347 rx_enable_stats = 0;
7350 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7351 unsigned int num_funcs = 0;
7354 queue_Remove(rpc_stat);
7355 num_funcs = rpc_stat->stats[0].func_total;
7357 sizeof(rx_interface_stat_t) +
7358 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7360 rxi_Free(rpc_stat, space);
7361 rxi_rpc_process_stat_cnt -= num_funcs;
7363 MUTEX_EXIT(&rx_rpc_stats);
7367 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7377 rx_disablePeerRPCStats(void)
7379 struct rx_peer **peer_ptr, **peer_end;
7382 MUTEX_ENTER(&rx_rpc_stats);
7385 * Turn off peer statistics and if process stats is also off, turn
7389 rxi_monitor_peerStats = 0;
7390 if (rxi_monitor_processStats == 0) {
7391 rx_enable_stats = 0;
7394 MUTEX_ENTER(&rx_peerHashTable_lock);
7395 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7396 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7398 struct rx_peer *peer, *next, *prev;
7399 for (prev = peer = *peer_ptr; peer; peer = next) {
7401 code = MUTEX_TRYENTER(&peer->peer_lock);
7403 rx_interface_stat_p rpc_stat, nrpc_stat;
7406 (&peer->rpcStats, rpc_stat, nrpc_stat,
7407 rx_interface_stat)) {
7408 unsigned int num_funcs = 0;
7411 queue_Remove(&rpc_stat->queue_header);
7412 queue_Remove(&rpc_stat->all_peers);
7413 num_funcs = rpc_stat->stats[0].func_total;
7415 sizeof(rx_interface_stat_t) +
7416 rpc_stat->stats[0].func_total *
7417 sizeof(rx_function_entry_v1_t);
7419 rxi_Free(rpc_stat, space);
7420 rxi_rpc_peer_stat_cnt -= num_funcs;
7422 MUTEX_EXIT(&peer->peer_lock);
7423 if (prev == *peer_ptr) {
7433 MUTEX_EXIT(&rx_peerHashTable_lock);
7434 MUTEX_EXIT(&rx_rpc_stats);
7438 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7443 * IN clearFlag - flag indicating which stats to clear
7451 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7453 rx_interface_stat_p rpc_stat, nrpc_stat;
7455 MUTEX_ENTER(&rx_rpc_stats);
7457 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7458 unsigned int num_funcs = 0, i;
7459 num_funcs = rpc_stat->stats[0].func_total;
7460 for (i = 0; i < num_funcs; i++) {
7461 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7462 hzero(rpc_stat->stats[i].invocations);
7464 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7465 hzero(rpc_stat->stats[i].bytes_sent);
7467 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7468 hzero(rpc_stat->stats[i].bytes_rcvd);
7470 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7471 rpc_stat->stats[i].queue_time_sum.sec = 0;
7472 rpc_stat->stats[i].queue_time_sum.usec = 0;
7474 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7475 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7476 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7478 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7479 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7480 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7482 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7483 rpc_stat->stats[i].queue_time_max.sec = 0;
7484 rpc_stat->stats[i].queue_time_max.usec = 0;
7486 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7487 rpc_stat->stats[i].execution_time_sum.sec = 0;
7488 rpc_stat->stats[i].execution_time_sum.usec = 0;
7490 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7491 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7492 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7494 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7495 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7496 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7498 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7499 rpc_stat->stats[i].execution_time_max.sec = 0;
7500 rpc_stat->stats[i].execution_time_max.usec = 0;
7505 MUTEX_EXIT(&rx_rpc_stats);
7509 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7514 * IN clearFlag - flag indicating which stats to clear
7522 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7524 rx_interface_stat_p rpc_stat, nrpc_stat;
7526 MUTEX_ENTER(&rx_rpc_stats);
7528 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7529 unsigned int num_funcs = 0, i;
7532 * We have to fix the offset of rpc_stat since we are
7533 * keeping this structure on two rx_queues. The rx_queue
7534 * package assumes that the rx_queue member is the first
7535 * member of the structure. That is, rx_queue assumes that
7536 * any one item is only on one queue at a time. We are
7537 * breaking that assumption and so we have to do a little
7538 * math to fix our pointers.
7541 fix_offset = (char *)rpc_stat;
7542 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7543 rpc_stat = (rx_interface_stat_p) fix_offset;
7545 num_funcs = rpc_stat->stats[0].func_total;
7546 for (i = 0; i < num_funcs; i++) {
7547 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7548 hzero(rpc_stat->stats[i].invocations);
7550 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7551 hzero(rpc_stat->stats[i].bytes_sent);
7553 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7554 hzero(rpc_stat->stats[i].bytes_rcvd);
7556 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7557 rpc_stat->stats[i].queue_time_sum.sec = 0;
7558 rpc_stat->stats[i].queue_time_sum.usec = 0;
7560 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7561 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7562 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7564 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7565 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7566 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7568 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7569 rpc_stat->stats[i].queue_time_max.sec = 0;
7570 rpc_stat->stats[i].queue_time_max.usec = 0;
7572 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7573 rpc_stat->stats[i].execution_time_sum.sec = 0;
7574 rpc_stat->stats[i].execution_time_sum.usec = 0;
7576 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7577 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7578 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7580 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7581 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7582 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7584 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7585 rpc_stat->stats[i].execution_time_max.sec = 0;
7586 rpc_stat->stats[i].execution_time_max.usec = 0;
7591 MUTEX_EXIT(&rx_rpc_stats);
7595 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7596 * is authorized to enable/disable/clear RX statistics.
7598 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7601 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7603 rxi_rxstat_userok = proc;
7607 rx_RxStatUserOk(struct rx_call *call)
7609 if (!rxi_rxstat_userok)
7611 return rxi_rxstat_userok(call);