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;
1173 #ifdef RX_ENABLE_LOCKS
1174 CV_WAIT(&call->cv_tq, &call->lock);
1175 #else /* RX_ENABLE_LOCKS */
1176 osi_rxSleep(&call->tq);
1177 #endif /* RX_ENABLE_LOCKS */
1179 if (call->flags & RX_CALL_TQ_CLEARME) {
1180 rxi_ClearTransmitQueue(call, 0);
1181 queue_Init(&call->tq);
1183 MUTEX_EXIT(&call->lock);
1184 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1190 rxi_HasActiveCalls(register struct rx_connection *aconn)
1193 register struct rx_call *tcall;
1197 for (i = 0; i < RX_MAXCALLS; i++) {
1198 if ((tcall = aconn->call[i])) {
1199 if ((tcall->state == RX_STATE_ACTIVE)
1200 || (tcall->state == RX_STATE_PRECALL)) {
1211 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1212 register afs_int32 * aint32s)
1215 register struct rx_call *tcall;
1219 for (i = 0; i < RX_MAXCALLS; i++) {
1220 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1221 aint32s[i] = aconn->callNumber[i] + 1;
1223 aint32s[i] = aconn->callNumber[i];
1230 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1231 register afs_int32 * aint32s)
1234 register struct rx_call *tcall;
1238 for (i = 0; i < RX_MAXCALLS; i++) {
1239 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1240 aconn->callNumber[i] = aint32s[i] - 1;
1242 aconn->callNumber[i] = aint32s[i];
1248 /* Advertise a new service. A service is named locally by a UDP port
1249 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1252 char *serviceName; Name for identification purposes (e.g. the
1253 service name might be used for probing for
1256 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1257 struct rx_securityClass **securityObjects, int nSecurityObjects,
1258 afs_int32(*serviceProc) (struct rx_call * acall))
1260 osi_socket socket = OSI_NULLSOCKET;
1261 register struct rx_service *tservice;
1267 if (serviceId == 0) {
1269 "rx_NewService: service id for service %s is not non-zero.\n",
1276 "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",
1284 tservice = rxi_AllocService();
1286 for (i = 0; i < RX_MAX_SERVICES; i++) {
1287 register struct rx_service *service = rx_services[i];
1289 if (port == service->servicePort) {
1290 if (service->serviceId == serviceId) {
1291 /* The identical service has already been
1292 * installed; if the caller was intending to
1293 * change the security classes used by this
1294 * service, he/she loses. */
1296 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1297 serviceName, serviceId, service->serviceName);
1299 rxi_FreeService(tservice);
1302 /* Different service, same port: re-use the socket
1303 * which is bound to the same port */
1304 socket = service->socket;
1307 if (socket == OSI_NULLSOCKET) {
1308 /* If we don't already have a socket (from another
1309 * service on same port) get a new one */
1310 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1311 if (socket == OSI_NULLSOCKET) {
1313 rxi_FreeService(tservice);
1318 service->socket = socket;
1319 service->servicePort = port;
1320 service->serviceId = serviceId;
1321 service->serviceName = serviceName;
1322 service->nSecurityObjects = nSecurityObjects;
1323 service->securityObjects = securityObjects;
1324 service->minProcs = 0;
1325 service->maxProcs = 1;
1326 service->idleDeadTime = 60;
1327 service->connDeadTime = rx_connDeadTime;
1328 service->executeRequestProc = serviceProc;
1329 service->checkReach = 0;
1330 rx_services[i] = service; /* not visible until now */
1336 rxi_FreeService(tservice);
1337 (osi_Msg "rx_NewService: cannot support > %d services\n",
1342 /* Generic request processing loop. This routine should be called
1343 * by the implementation dependent rx_ServerProc. If socketp is
1344 * non-null, it will be set to the file descriptor that this thread
1345 * is now listening on. If socketp is null, this routine will never
1348 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1350 register struct rx_call *call;
1351 register afs_int32 code;
1352 register struct rx_service *tservice = NULL;
1359 call = rx_GetCall(threadID, tservice, socketp);
1360 if (socketp && *socketp != OSI_NULLSOCKET) {
1361 /* We are now a listener thread */
1366 /* if server is restarting( typically smooth shutdown) then do not
1367 * allow any new calls.
1370 if (rx_tranquil && (call != NULL)) {
1374 MUTEX_ENTER(&call->lock);
1376 rxi_CallError(call, RX_RESTARTING);
1377 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1379 MUTEX_EXIT(&call->lock);
1383 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1384 #ifdef RX_ENABLE_LOCKS
1386 #endif /* RX_ENABLE_LOCKS */
1387 afs_termState = AFSOP_STOP_AFS;
1388 afs_osi_Wakeup(&afs_termState);
1389 #ifdef RX_ENABLE_LOCKS
1391 #endif /* RX_ENABLE_LOCKS */
1396 tservice = call->conn->service;
1398 if (tservice->beforeProc)
1399 (*tservice->beforeProc) (call);
1401 code = call->conn->service->executeRequestProc(call);
1403 if (tservice->afterProc)
1404 (*tservice->afterProc) (call, code);
1406 rx_EndCall(call, code);
1407 MUTEX_ENTER(&rx_stats_mutex);
1409 MUTEX_EXIT(&rx_stats_mutex);
1415 rx_WakeupServerProcs(void)
1417 struct rx_serverQueueEntry *np, *tqp;
1421 MUTEX_ENTER(&rx_serverPool_lock);
1423 #ifdef RX_ENABLE_LOCKS
1424 if (rx_waitForPacket)
1425 CV_BROADCAST(&rx_waitForPacket->cv);
1426 #else /* RX_ENABLE_LOCKS */
1427 if (rx_waitForPacket)
1428 osi_rxWakeup(rx_waitForPacket);
1429 #endif /* RX_ENABLE_LOCKS */
1430 MUTEX_ENTER(&freeSQEList_lock);
1431 for (np = rx_FreeSQEList; np; np = tqp) {
1432 tqp = *(struct rx_serverQueueEntry **)np;
1433 #ifdef RX_ENABLE_LOCKS
1434 CV_BROADCAST(&np->cv);
1435 #else /* RX_ENABLE_LOCKS */
1437 #endif /* RX_ENABLE_LOCKS */
1439 MUTEX_EXIT(&freeSQEList_lock);
1440 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1441 #ifdef RX_ENABLE_LOCKS
1442 CV_BROADCAST(&np->cv);
1443 #else /* RX_ENABLE_LOCKS */
1445 #endif /* RX_ENABLE_LOCKS */
1447 MUTEX_EXIT(&rx_serverPool_lock);
1452 * One thing that seems to happen is that all the server threads get
1453 * tied up on some empty or slow call, and then a whole bunch of calls
1454 * arrive at once, using up the packet pool, so now there are more
1455 * empty calls. The most critical resources here are server threads
1456 * and the free packet pool. The "doreclaim" code seems to help in
1457 * general. I think that eventually we arrive in this state: there
1458 * are lots of pending calls which do have all their packets present,
1459 * so they won't be reclaimed, are multi-packet calls, so they won't
1460 * be scheduled until later, and thus are tying up most of the free
1461 * packet pool for a very long time.
1463 * 1. schedule multi-packet calls if all the packets are present.
1464 * Probably CPU-bound operation, useful to return packets to pool.
1465 * Do what if there is a full window, but the last packet isn't here?
1466 * 3. preserve one thread which *only* runs "best" calls, otherwise
1467 * it sleeps and waits for that type of call.
1468 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1469 * the current dataquota business is badly broken. The quota isn't adjusted
1470 * to reflect how many packets are presently queued for a running call.
1471 * So, when we schedule a queued call with a full window of packets queued
1472 * up for it, that *should* free up a window full of packets for other 2d-class
1473 * calls to be able to use from the packet pool. But it doesn't.
1475 * NB. Most of the time, this code doesn't run -- since idle server threads
1476 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1477 * as a new call arrives.
1479 /* Sleep until a call arrives. Returns a pointer to the call, ready
1480 * for an rx_Read. */
1481 #ifdef RX_ENABLE_LOCKS
1483 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1485 struct rx_serverQueueEntry *sq;
1486 register struct rx_call *call = (struct rx_call *)0;
1487 struct rx_service *service = NULL;
1490 MUTEX_ENTER(&freeSQEList_lock);
1492 if ((sq = rx_FreeSQEList)) {
1493 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1494 MUTEX_EXIT(&freeSQEList_lock);
1495 } else { /* otherwise allocate a new one and return that */
1496 MUTEX_EXIT(&freeSQEList_lock);
1497 sq = (struct rx_serverQueueEntry *)
1498 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1499 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1500 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1503 MUTEX_ENTER(&rx_serverPool_lock);
1504 if (cur_service != NULL) {
1505 ReturnToServerPool(cur_service);
1508 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1509 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1511 /* Scan for eligible incoming calls. A call is not eligible
1512 * if the maximum number of calls for its service type are
1513 * already executing */
1514 /* One thread will process calls FCFS (to prevent starvation),
1515 * while the other threads may run ahead looking for calls which
1516 * have all their input data available immediately. This helps
1517 * keep threads from blocking, waiting for data from the client. */
1518 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1519 service = tcall->conn->service;
1520 if (!QuotaOK(service)) {
1523 if (tno == rxi_fcfs_thread_num
1524 || !tcall->queue_item_header.next) {
1525 /* If we're the fcfs thread , then we'll just use
1526 * this call. If we haven't been able to find an optimal
1527 * choice, and we're at the end of the list, then use a
1528 * 2d choice if one has been identified. Otherwise... */
1529 call = (choice2 ? choice2 : tcall);
1530 service = call->conn->service;
1531 } else if (!queue_IsEmpty(&tcall->rq)) {
1532 struct rx_packet *rp;
1533 rp = queue_First(&tcall->rq, rx_packet);
1534 if (rp->header.seq == 1) {
1536 || (rp->header.flags & RX_LAST_PACKET)) {
1538 } else if (rxi_2dchoice && !choice2
1539 && !(tcall->flags & RX_CALL_CLEARED)
1540 && (tcall->rprev > rxi_HardAckRate)) {
1549 ReturnToServerPool(service);
1556 MUTEX_EXIT(&rx_serverPool_lock);
1557 MUTEX_ENTER(&call->lock);
1559 if (call->flags & RX_CALL_WAIT_PROC) {
1560 call->flags &= ~RX_CALL_WAIT_PROC;
1561 MUTEX_ENTER(&rx_stats_mutex);
1563 MUTEX_EXIT(&rx_stats_mutex);
1566 if (call->state != RX_STATE_PRECALL || call->error) {
1567 MUTEX_EXIT(&call->lock);
1568 MUTEX_ENTER(&rx_serverPool_lock);
1569 ReturnToServerPool(service);
1574 if (queue_IsEmpty(&call->rq)
1575 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1576 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1578 CLEAR_CALL_QUEUE_LOCK(call);
1581 /* If there are no eligible incoming calls, add this process
1582 * to the idle server queue, to wait for one */
1586 *socketp = OSI_NULLSOCKET;
1588 sq->socketp = socketp;
1589 queue_Append(&rx_idleServerQueue, sq);
1590 #ifndef AFS_AIX41_ENV
1591 rx_waitForPacket = sq;
1593 rx_waitingForPacket = sq;
1594 #endif /* AFS_AIX41_ENV */
1596 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1598 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1599 MUTEX_EXIT(&rx_serverPool_lock);
1600 return (struct rx_call *)0;
1603 } while (!(call = sq->newcall)
1604 && !(socketp && *socketp != OSI_NULLSOCKET));
1605 MUTEX_EXIT(&rx_serverPool_lock);
1607 MUTEX_ENTER(&call->lock);
1613 MUTEX_ENTER(&freeSQEList_lock);
1614 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1615 rx_FreeSQEList = sq;
1616 MUTEX_EXIT(&freeSQEList_lock);
1619 clock_GetTime(&call->startTime);
1620 call->state = RX_STATE_ACTIVE;
1621 call->mode = RX_MODE_RECEIVING;
1622 #ifdef RX_KERNEL_TRACE
1623 if (ICL_SETACTIVE(afs_iclSetp)) {
1624 int glockOwner = ISAFS_GLOCK();
1627 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1628 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1635 rxi_calltrace(RX_CALL_START, call);
1636 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1637 call->conn->service->servicePort, call->conn->service->serviceId,
1640 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1641 MUTEX_EXIT(&call->lock);
1643 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1648 #else /* RX_ENABLE_LOCKS */
1650 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1652 struct rx_serverQueueEntry *sq;
1653 register struct rx_call *call = (struct rx_call *)0, *choice2;
1654 struct rx_service *service = NULL;
1658 MUTEX_ENTER(&freeSQEList_lock);
1660 if ((sq = rx_FreeSQEList)) {
1661 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1662 MUTEX_EXIT(&freeSQEList_lock);
1663 } else { /* otherwise allocate a new one and return that */
1664 MUTEX_EXIT(&freeSQEList_lock);
1665 sq = (struct rx_serverQueueEntry *)
1666 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1667 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1668 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1670 MUTEX_ENTER(&sq->lock);
1672 if (cur_service != NULL) {
1673 cur_service->nRequestsRunning--;
1674 if (cur_service->nRequestsRunning < cur_service->minProcs)
1678 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1679 register struct rx_call *tcall, *ncall;
1680 /* Scan for eligible incoming calls. A call is not eligible
1681 * if the maximum number of calls for its service type are
1682 * already executing */
1683 /* One thread will process calls FCFS (to prevent starvation),
1684 * while the other threads may run ahead looking for calls which
1685 * have all their input data available immediately. This helps
1686 * keep threads from blocking, waiting for data from the client. */
1687 choice2 = (struct rx_call *)0;
1688 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1689 service = tcall->conn->service;
1690 if (QuotaOK(service)) {
1691 if (tno == rxi_fcfs_thread_num
1692 || !tcall->queue_item_header.next) {
1693 /* If we're the fcfs thread, then we'll just use
1694 * this call. If we haven't been able to find an optimal
1695 * choice, and we're at the end of the list, then use a
1696 * 2d choice if one has been identified. Otherwise... */
1697 call = (choice2 ? choice2 : tcall);
1698 service = call->conn->service;
1699 } else if (!queue_IsEmpty(&tcall->rq)) {
1700 struct rx_packet *rp;
1701 rp = queue_First(&tcall->rq, rx_packet);
1702 if (rp->header.seq == 1
1704 || (rp->header.flags & RX_LAST_PACKET))) {
1706 } else if (rxi_2dchoice && !choice2
1707 && !(tcall->flags & RX_CALL_CLEARED)
1708 && (tcall->rprev > rxi_HardAckRate)) {
1721 /* we can't schedule a call if there's no data!!! */
1722 /* send an ack if there's no data, if we're missing the
1723 * first packet, or we're missing something between first
1724 * and last -- there's a "hole" in the incoming data. */
1725 if (queue_IsEmpty(&call->rq)
1726 || queue_First(&call->rq, rx_packet)->header.seq != 1
1727 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1728 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1730 call->flags &= (~RX_CALL_WAIT_PROC);
1731 service->nRequestsRunning++;
1732 /* just started call in minProcs pool, need fewer to maintain
1734 if (service->nRequestsRunning <= service->minProcs)
1738 /* MUTEX_EXIT(&call->lock); */
1740 /* If there are no eligible incoming calls, add this process
1741 * to the idle server queue, to wait for one */
1744 *socketp = OSI_NULLSOCKET;
1746 sq->socketp = socketp;
1747 queue_Append(&rx_idleServerQueue, sq);
1751 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1753 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1754 return (struct rx_call *)0;
1757 } while (!(call = sq->newcall)
1758 && !(socketp && *socketp != OSI_NULLSOCKET));
1760 MUTEX_EXIT(&sq->lock);
1762 MUTEX_ENTER(&freeSQEList_lock);
1763 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1764 rx_FreeSQEList = sq;
1765 MUTEX_EXIT(&freeSQEList_lock);
1768 clock_GetTime(&call->startTime);
1769 call->state = RX_STATE_ACTIVE;
1770 call->mode = RX_MODE_RECEIVING;
1771 #ifdef RX_KERNEL_TRACE
1772 if (ICL_SETACTIVE(afs_iclSetp)) {
1773 int glockOwner = ISAFS_GLOCK();
1776 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1777 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1784 rxi_calltrace(RX_CALL_START, call);
1785 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1786 call->conn->service->servicePort, call->conn->service->serviceId,
1789 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1796 #endif /* RX_ENABLE_LOCKS */
1800 /* Establish a procedure to be called when a packet arrives for a
1801 * call. This routine will be called at most once after each call,
1802 * and will also be called if there is an error condition on the or
1803 * the call is complete. Used by multi rx to build a selection
1804 * function which determines which of several calls is likely to be a
1805 * good one to read from.
1806 * NOTE: the way this is currently implemented it is probably only a
1807 * good idea to (1) use it immediately after a newcall (clients only)
1808 * and (2) only use it once. Other uses currently void your warranty
1811 rx_SetArrivalProc(register struct rx_call *call,
1812 register void (*proc) (register struct rx_call * call,
1814 register int index),
1815 register VOID * handle, register int arg)
1817 call->arrivalProc = proc;
1818 call->arrivalProcHandle = handle;
1819 call->arrivalProcArg = arg;
1822 /* Call is finished (possibly prematurely). Return rc to the peer, if
1823 * appropriate, and return the final error code from the conversation
1827 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1829 register struct rx_connection *conn = call->conn;
1830 register struct rx_service *service;
1831 register struct rx_packet *tp; /* Temporary packet pointer */
1832 register struct rx_packet *nxp; /* Next packet pointer, for queue_Scan */
1836 dpf(("rx_EndCall(call %x)\n", call));
1839 MUTEX_ENTER(&call->lock);
1841 if (rc == 0 && call->error == 0) {
1842 call->abortCode = 0;
1843 call->abortCount = 0;
1846 call->arrivalProc = (void (*)())0;
1847 if (rc && call->error == 0) {
1848 rxi_CallError(call, rc);
1849 /* Send an abort message to the peer if this error code has
1850 * only just been set. If it was set previously, assume the
1851 * peer has already been sent the error code or will request it
1853 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1855 if (conn->type == RX_SERVER_CONNECTION) {
1856 /* Make sure reply or at least dummy reply is sent */
1857 if (call->mode == RX_MODE_RECEIVING) {
1858 rxi_WriteProc(call, 0, 0);
1860 if (call->mode == RX_MODE_SENDING) {
1861 rxi_FlushWrite(call);
1863 service = conn->service;
1864 rxi_calltrace(RX_CALL_END, call);
1865 /* Call goes to hold state until reply packets are acknowledged */
1866 if (call->tfirst + call->nSoftAcked < call->tnext) {
1867 call->state = RX_STATE_HOLD;
1869 call->state = RX_STATE_DALLY;
1870 rxi_ClearTransmitQueue(call, 0);
1871 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1872 rxevent_Cancel(call->keepAliveEvent, call,
1873 RX_CALL_REFCOUNT_ALIVE);
1875 } else { /* Client connection */
1877 /* Make sure server receives input packets, in the case where
1878 * no reply arguments are expected */
1879 if ((call->mode == RX_MODE_SENDING)
1880 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1881 (void)rxi_ReadProc(call, &dummy, 1);
1884 /* If we had an outstanding delayed ack, be nice to the server
1885 * and force-send it now.
1887 if (call->delayedAckEvent) {
1888 rxevent_Cancel(call->delayedAckEvent, call,
1889 RX_CALL_REFCOUNT_DELAY);
1890 call->delayedAckEvent = NULL;
1891 rxi_SendDelayedAck(NULL, call, NULL);
1894 /* We need to release the call lock since it's lower than the
1895 * conn_call_lock and we don't want to hold the conn_call_lock
1896 * over the rx_ReadProc call. The conn_call_lock needs to be held
1897 * here for the case where rx_NewCall is perusing the calls on
1898 * the connection structure. We don't want to signal until
1899 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1900 * have checked this call, found it active and by the time it
1901 * goes to sleep, will have missed the signal.
1903 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
1904 * there are threads waiting to use the conn object.
1906 MUTEX_EXIT(&call->lock);
1907 MUTEX_ENTER(&conn->conn_call_lock);
1908 MUTEX_ENTER(&call->lock);
1909 MUTEX_ENTER(&conn->conn_data_lock);
1910 conn->flags |= RX_CONN_BUSY;
1911 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1912 if (conn->makeCallWaiters == 0)
1913 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1914 MUTEX_EXIT(&conn->conn_data_lock);
1915 #ifdef RX_ENABLE_LOCKS
1916 CV_BROADCAST(&conn->conn_call_cv);
1921 #ifdef RX_ENABLE_LOCKS
1923 MUTEX_EXIT(&conn->conn_data_lock);
1925 #endif /* RX_ENABLE_LOCKS */
1926 call->state = RX_STATE_DALLY;
1928 error = call->error;
1930 /* currentPacket, nLeft, and NFree must be zeroed here, because
1931 * ResetCall cannot: ResetCall may be called at splnet(), in the
1932 * kernel version, and may interrupt the macros rx_Read or
1933 * rx_Write, which run at normal priority for efficiency. */
1934 if (call->currentPacket) {
1935 queue_Prepend(&call->iovq, call->currentPacket);
1936 call->currentPacket = (struct rx_packet *)0;
1939 call->nLeft = call->nFree = call->curlen = 0;
1941 /* Free any packets from the last call to ReadvProc/WritevProc */
1942 rxi_FreePackets(0, &call->iovq);
1944 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1945 MUTEX_EXIT(&call->lock);
1946 if (conn->type == RX_CLIENT_CONNECTION) {
1947 MUTEX_EXIT(&conn->conn_call_lock);
1948 conn->flags &= ~RX_CONN_BUSY;
1952 * Map errors to the local host's errno.h format.
1954 error = ntoh_syserr_conv(error);
1958 #if !defined(KERNEL)
1960 /* Call this routine when shutting down a server or client (especially
1961 * clients). This will allow Rx to gracefully garbage collect server
1962 * connections, and reduce the number of retries that a server might
1963 * make to a dead client.
1964 * This is not quite right, since some calls may still be ongoing and
1965 * we can't lock them to destroy them. */
1969 register struct rx_connection **conn_ptr, **conn_end;
1973 if (rxinit_status == 1) {
1975 return; /* Already shutdown. */
1977 rxi_DeleteCachedConnections();
1978 if (rx_connHashTable) {
1979 MUTEX_ENTER(&rx_connHashTable_lock);
1980 for (conn_ptr = &rx_connHashTable[0], conn_end =
1981 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
1983 struct rx_connection *conn, *next;
1984 for (conn = *conn_ptr; conn; conn = next) {
1986 if (conn->type == RX_CLIENT_CONNECTION) {
1987 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
1989 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
1990 #ifdef RX_ENABLE_LOCKS
1991 rxi_DestroyConnectionNoLock(conn);
1992 #else /* RX_ENABLE_LOCKS */
1993 rxi_DestroyConnection(conn);
1994 #endif /* RX_ENABLE_LOCKS */
1998 #ifdef RX_ENABLE_LOCKS
1999 while (rx_connCleanup_list) {
2000 struct rx_connection *conn;
2001 conn = rx_connCleanup_list;
2002 rx_connCleanup_list = rx_connCleanup_list->next;
2003 MUTEX_EXIT(&rx_connHashTable_lock);
2004 rxi_CleanupConnection(conn);
2005 MUTEX_ENTER(&rx_connHashTable_lock);
2007 MUTEX_EXIT(&rx_connHashTable_lock);
2008 #endif /* RX_ENABLE_LOCKS */
2017 /* if we wakeup packet waiter too often, can get in loop with two
2018 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2020 rxi_PacketsUnWait(void)
2022 if (!rx_waitingForPackets) {
2026 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2027 return; /* still over quota */
2030 rx_waitingForPackets = 0;
2031 #ifdef RX_ENABLE_LOCKS
2032 CV_BROADCAST(&rx_waitingForPackets_cv);
2034 osi_rxWakeup(&rx_waitingForPackets);
2040 /* ------------------Internal interfaces------------------------- */
2042 /* Return this process's service structure for the
2043 * specified socket and service */
2045 rxi_FindService(register osi_socket socket, register u_short serviceId)
2047 register struct rx_service **sp;
2048 for (sp = &rx_services[0]; *sp; sp++) {
2049 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2055 /* Allocate a call structure, for the indicated channel of the
2056 * supplied connection. The mode and state of the call must be set by
2057 * the caller. Returns the call with mutex locked. */
2059 rxi_NewCall(register struct rx_connection *conn, register int channel)
2061 register struct rx_call *call;
2062 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2063 register struct rx_call *cp; /* Call pointer temp */
2064 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2065 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2067 /* Grab an existing call structure, or allocate a new one.
2068 * Existing call structures are assumed to have been left reset by
2070 MUTEX_ENTER(&rx_freeCallQueue_lock);
2072 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2074 * EXCEPT that the TQ might not yet be cleared out.
2075 * Skip over those with in-use TQs.
2078 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2079 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2085 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2086 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2087 call = queue_First(&rx_freeCallQueue, rx_call);
2088 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2090 MUTEX_ENTER(&rx_stats_mutex);
2091 rx_stats.nFreeCallStructs--;
2092 MUTEX_EXIT(&rx_stats_mutex);
2093 MUTEX_EXIT(&rx_freeCallQueue_lock);
2094 MUTEX_ENTER(&call->lock);
2095 CLEAR_CALL_QUEUE_LOCK(call);
2096 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2097 /* Now, if TQ wasn't cleared earlier, do it now. */
2098 if (call->flags & RX_CALL_TQ_CLEARME) {
2099 rxi_ClearTransmitQueue(call, 0);
2100 queue_Init(&call->tq);
2102 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2103 /* Bind the call to its connection structure */
2105 rxi_ResetCall(call, 1);
2107 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2109 MUTEX_EXIT(&rx_freeCallQueue_lock);
2110 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2111 MUTEX_ENTER(&call->lock);
2112 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2113 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2114 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2116 MUTEX_ENTER(&rx_stats_mutex);
2117 rx_stats.nCallStructs++;
2118 MUTEX_EXIT(&rx_stats_mutex);
2119 /* Initialize once-only items */
2120 queue_Init(&call->tq);
2121 queue_Init(&call->rq);
2122 queue_Init(&call->iovq);
2123 /* Bind the call to its connection structure (prereq for reset) */
2125 rxi_ResetCall(call, 1);
2127 call->channel = channel;
2128 call->callNumber = &conn->callNumber[channel];
2129 /* Note that the next expected call number is retained (in
2130 * conn->callNumber[i]), even if we reallocate the call structure
2132 conn->call[channel] = call;
2133 /* if the channel's never been used (== 0), we should start at 1, otherwise
2134 * the call number is valid from the last time this channel was used */
2135 if (*call->callNumber == 0)
2136 *call->callNumber = 1;
2141 /* A call has been inactive long enough that so we can throw away
2142 * state, including the call structure, which is placed on the call
2144 * Call is locked upon entry.
2145 * haveCTLock set if called from rxi_ReapConnections
2147 #ifdef RX_ENABLE_LOCKS
2149 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2150 #else /* RX_ENABLE_LOCKS */
2152 rxi_FreeCall(register struct rx_call *call)
2153 #endif /* RX_ENABLE_LOCKS */
2155 register int channel = call->channel;
2156 register struct rx_connection *conn = call->conn;
2159 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2160 (*call->callNumber)++;
2161 rxi_ResetCall(call, 0);
2162 call->conn->call[channel] = (struct rx_call *)0;
2164 MUTEX_ENTER(&rx_freeCallQueue_lock);
2165 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2166 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2167 /* A call may be free even though its transmit queue is still in use.
2168 * Since we search the call list from head to tail, put busy calls at
2169 * the head of the list, and idle calls at the tail.
2171 if (call->flags & RX_CALL_TQ_BUSY)
2172 queue_Prepend(&rx_freeCallQueue, call);
2174 queue_Append(&rx_freeCallQueue, call);
2175 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2176 queue_Append(&rx_freeCallQueue, call);
2177 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2178 MUTEX_ENTER(&rx_stats_mutex);
2179 rx_stats.nFreeCallStructs++;
2180 MUTEX_EXIT(&rx_stats_mutex);
2182 MUTEX_EXIT(&rx_freeCallQueue_lock);
2184 /* Destroy the connection if it was previously slated for
2185 * destruction, i.e. the Rx client code previously called
2186 * rx_DestroyConnection (client connections), or
2187 * rxi_ReapConnections called the same routine (server
2188 * connections). Only do this, however, if there are no
2189 * outstanding calls. Note that for fine grain locking, there appears
2190 * to be a deadlock in that rxi_FreeCall has a call locked and
2191 * DestroyConnectionNoLock locks each call in the conn. But note a
2192 * few lines up where we have removed this call from the conn.
2193 * If someone else destroys a connection, they either have no
2194 * call lock held or are going through this section of code.
2196 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2197 MUTEX_ENTER(&conn->conn_data_lock);
2199 MUTEX_EXIT(&conn->conn_data_lock);
2200 #ifdef RX_ENABLE_LOCKS
2202 rxi_DestroyConnectionNoLock(conn);
2204 rxi_DestroyConnection(conn);
2205 #else /* RX_ENABLE_LOCKS */
2206 rxi_DestroyConnection(conn);
2207 #endif /* RX_ENABLE_LOCKS */
2211 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2213 rxi_Alloc(register size_t size)
2217 MUTEX_ENTER(&rx_stats_mutex);
2219 rxi_Allocsize += size;
2220 MUTEX_EXIT(&rx_stats_mutex);
2222 p = (char *)osi_Alloc(size);
2225 osi_Panic("rxi_Alloc error");
2231 rxi_Free(void *addr, register size_t size)
2233 MUTEX_ENTER(&rx_stats_mutex);
2235 rxi_Allocsize -= size;
2236 MUTEX_EXIT(&rx_stats_mutex);
2238 osi_Free(addr, size);
2241 /* Find the peer process represented by the supplied (host,port)
2242 * combination. If there is no appropriate active peer structure, a
2243 * new one will be allocated and initialized
2244 * The origPeer, if set, is a pointer to a peer structure on which the
2245 * refcount will be be decremented. This is used to replace the peer
2246 * structure hanging off a connection structure */
2248 rxi_FindPeer(register afs_uint32 host, register u_short port,
2249 struct rx_peer *origPeer, int create)
2251 register struct rx_peer *pp;
2253 hashIndex = PEER_HASH(host, port);
2254 MUTEX_ENTER(&rx_peerHashTable_lock);
2255 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2256 if ((pp->host == host) && (pp->port == port))
2261 pp = rxi_AllocPeer(); /* This bzero's *pp */
2262 pp->host = host; /* set here or in InitPeerParams is zero */
2264 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2265 queue_Init(&pp->congestionQueue);
2266 queue_Init(&pp->rpcStats);
2267 pp->next = rx_peerHashTable[hashIndex];
2268 rx_peerHashTable[hashIndex] = pp;
2269 rxi_InitPeerParams(pp);
2270 MUTEX_ENTER(&rx_stats_mutex);
2271 rx_stats.nPeerStructs++;
2272 MUTEX_EXIT(&rx_stats_mutex);
2279 origPeer->refCount--;
2280 MUTEX_EXIT(&rx_peerHashTable_lock);
2285 /* Find the connection at (host, port) started at epoch, and with the
2286 * given connection id. Creates the server connection if necessary.
2287 * The type specifies whether a client connection or a server
2288 * connection is desired. In both cases, (host, port) specify the
2289 * peer's (host, pair) pair. Client connections are not made
2290 * automatically by this routine. The parameter socket gives the
2291 * socket descriptor on which the packet was received. This is used,
2292 * in the case of server connections, to check that *new* connections
2293 * come via a valid (port, serviceId). Finally, the securityIndex
2294 * parameter must match the existing index for the connection. If a
2295 * server connection is created, it will be created using the supplied
2296 * index, if the index is valid for this service */
2297 struct rx_connection *
2298 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2299 register u_short port, u_short serviceId, afs_uint32 cid,
2300 afs_uint32 epoch, int type, u_int securityIndex)
2302 int hashindex, flag;
2303 register struct rx_connection *conn;
2304 hashindex = CONN_HASH(host, port, cid, epoch, type);
2305 MUTEX_ENTER(&rx_connHashTable_lock);
2306 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2307 rx_connHashTable[hashindex],
2310 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2311 && (epoch == conn->epoch)) {
2312 register struct rx_peer *pp = conn->peer;
2313 if (securityIndex != conn->securityIndex) {
2314 /* this isn't supposed to happen, but someone could forge a packet
2315 * like this, and there seems to be some CM bug that makes this
2316 * happen from time to time -- in which case, the fileserver
2318 MUTEX_EXIT(&rx_connHashTable_lock);
2319 return (struct rx_connection *)0;
2321 if (pp->host == host && pp->port == port)
2323 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2325 /* So what happens when it's a callback connection? */
2326 if ( /*type == RX_CLIENT_CONNECTION && */
2327 (conn->epoch & 0x80000000))
2331 /* the connection rxLastConn that was used the last time is not the
2332 ** one we are looking for now. Hence, start searching in the hash */
2334 conn = rx_connHashTable[hashindex];
2339 struct rx_service *service;
2340 if (type == RX_CLIENT_CONNECTION) {
2341 MUTEX_EXIT(&rx_connHashTable_lock);
2342 return (struct rx_connection *)0;
2344 service = rxi_FindService(socket, serviceId);
2345 if (!service || (securityIndex >= service->nSecurityObjects)
2346 || (service->securityObjects[securityIndex] == 0)) {
2347 MUTEX_EXIT(&rx_connHashTable_lock);
2348 return (struct rx_connection *)0;
2350 conn = rxi_AllocConnection(); /* This bzero's the connection */
2351 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2352 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2353 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2354 conn->next = rx_connHashTable[hashindex];
2355 rx_connHashTable[hashindex] = conn;
2356 conn->peer = rxi_FindPeer(host, port, 0, 1);
2357 conn->type = RX_SERVER_CONNECTION;
2358 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2359 conn->epoch = epoch;
2360 conn->cid = cid & RX_CIDMASK;
2361 /* conn->serial = conn->lastSerial = 0; */
2362 /* conn->timeout = 0; */
2363 conn->ackRate = RX_FAST_ACK_RATE;
2364 conn->service = service;
2365 conn->serviceId = serviceId;
2366 conn->securityIndex = securityIndex;
2367 conn->securityObject = service->securityObjects[securityIndex];
2368 conn->nSpecific = 0;
2369 conn->specific = NULL;
2370 rx_SetConnDeadTime(conn, service->connDeadTime);
2371 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2372 /* Notify security object of the new connection */
2373 RXS_NewConnection(conn->securityObject, conn);
2374 /* XXXX Connection timeout? */
2375 if (service->newConnProc)
2376 (*service->newConnProc) (conn);
2377 MUTEX_ENTER(&rx_stats_mutex);
2378 rx_stats.nServerConns++;
2379 MUTEX_EXIT(&rx_stats_mutex);
2382 MUTEX_ENTER(&conn->conn_data_lock);
2384 MUTEX_EXIT(&conn->conn_data_lock);
2386 rxLastConn = conn; /* store this connection as the last conn used */
2387 MUTEX_EXIT(&rx_connHashTable_lock);
2391 /* There are two packet tracing routines available for testing and monitoring
2392 * Rx. One is called just after every packet is received and the other is
2393 * called just before every packet is sent. Received packets, have had their
2394 * headers decoded, and packets to be sent have not yet had their headers
2395 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2396 * containing the network address. Both can be modified. The return value, if
2397 * non-zero, indicates that the packet should be dropped. */
2399 int (*rx_justReceived) () = 0;
2400 int (*rx_almostSent) () = 0;
2402 /* A packet has been received off the interface. Np is the packet, socket is
2403 * the socket number it was received from (useful in determining which service
2404 * this packet corresponds to), and (host, port) reflect the host,port of the
2405 * sender. This call returns the packet to the caller if it is finished with
2406 * it, rather than de-allocating it, just as a small performance hack */
2409 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2410 afs_uint32 host, u_short port, int *tnop,
2411 struct rx_call **newcallp)
2413 register struct rx_call *call;
2414 register struct rx_connection *conn;
2416 afs_uint32 currentCallNumber;
2422 struct rx_packet *tnp;
2425 /* We don't print out the packet until now because (1) the time may not be
2426 * accurate enough until now in the lwp implementation (rx_Listener only gets
2427 * the time after the packet is read) and (2) from a protocol point of view,
2428 * this is the first time the packet has been seen */
2429 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2430 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2431 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2432 np->header.serial, packetType, host, port, np->header.serviceId,
2433 np->header.epoch, np->header.cid, np->header.callNumber,
2434 np->header.seq, np->header.flags, np));
2437 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2438 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2441 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2442 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2445 /* If an input tracer function is defined, call it with the packet and
2446 * network address. Note this function may modify its arguments. */
2447 if (rx_justReceived) {
2448 struct sockaddr_in addr;
2450 addr.sin_family = AF_INET;
2451 addr.sin_port = port;
2452 addr.sin_addr.s_addr = host;
2453 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2454 addr.sin_len = sizeof(addr);
2455 #endif /* AFS_OSF_ENV */
2456 drop = (*rx_justReceived) (np, &addr);
2457 /* drop packet if return value is non-zero */
2460 port = addr.sin_port; /* in case fcn changed addr */
2461 host = addr.sin_addr.s_addr;
2465 /* If packet was not sent by the client, then *we* must be the client */
2466 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2467 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2469 /* Find the connection (or fabricate one, if we're the server & if
2470 * necessary) associated with this packet */
2472 rxi_FindConnection(socket, host, port, np->header.serviceId,
2473 np->header.cid, np->header.epoch, type,
2474 np->header.securityIndex);
2477 /* If no connection found or fabricated, just ignore the packet.
2478 * (An argument could be made for sending an abort packet for
2483 MUTEX_ENTER(&conn->conn_data_lock);
2484 if (conn->maxSerial < np->header.serial)
2485 conn->maxSerial = np->header.serial;
2486 MUTEX_EXIT(&conn->conn_data_lock);
2488 /* If the connection is in an error state, send an abort packet and ignore
2489 * the incoming packet */
2491 /* Don't respond to an abort packet--we don't want loops! */
2492 MUTEX_ENTER(&conn->conn_data_lock);
2493 if (np->header.type != RX_PACKET_TYPE_ABORT)
2494 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2496 MUTEX_EXIT(&conn->conn_data_lock);
2500 /* Check for connection-only requests (i.e. not call specific). */
2501 if (np->header.callNumber == 0) {
2502 switch (np->header.type) {
2503 case RX_PACKET_TYPE_ABORT:
2504 /* What if the supplied error is zero? */
2505 rxi_ConnectionError(conn, ntohl(rx_GetInt32(np, 0)));
2506 MUTEX_ENTER(&conn->conn_data_lock);
2508 MUTEX_EXIT(&conn->conn_data_lock);
2510 case RX_PACKET_TYPE_CHALLENGE:
2511 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2512 MUTEX_ENTER(&conn->conn_data_lock);
2514 MUTEX_EXIT(&conn->conn_data_lock);
2516 case RX_PACKET_TYPE_RESPONSE:
2517 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2518 MUTEX_ENTER(&conn->conn_data_lock);
2520 MUTEX_EXIT(&conn->conn_data_lock);
2522 case RX_PACKET_TYPE_PARAMS:
2523 case RX_PACKET_TYPE_PARAMS + 1:
2524 case RX_PACKET_TYPE_PARAMS + 2:
2525 /* ignore these packet types for now */
2526 MUTEX_ENTER(&conn->conn_data_lock);
2528 MUTEX_EXIT(&conn->conn_data_lock);
2533 /* Should not reach here, unless the peer is broken: send an
2535 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2536 MUTEX_ENTER(&conn->conn_data_lock);
2537 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2539 MUTEX_EXIT(&conn->conn_data_lock);
2544 channel = np->header.cid & RX_CHANNELMASK;
2545 call = conn->call[channel];
2546 #ifdef RX_ENABLE_LOCKS
2548 MUTEX_ENTER(&call->lock);
2549 /* Test to see if call struct is still attached to conn. */
2550 if (call != conn->call[channel]) {
2552 MUTEX_EXIT(&call->lock);
2553 if (type == RX_SERVER_CONNECTION) {
2554 call = conn->call[channel];
2555 /* If we started with no call attached and there is one now,
2556 * another thread is also running this routine and has gotten
2557 * the connection channel. We should drop this packet in the tests
2558 * below. If there was a call on this connection and it's now
2559 * gone, then we'll be making a new call below.
2560 * If there was previously a call and it's now different then
2561 * the old call was freed and another thread running this routine
2562 * has created a call on this channel. One of these two threads
2563 * has a packet for the old call and the code below handles those
2567 MUTEX_ENTER(&call->lock);
2569 /* This packet can't be for this call. If the new call address is
2570 * 0 then no call is running on this channel. If there is a call
2571 * then, since this is a client connection we're getting data for
2572 * it must be for the previous call.
2574 MUTEX_ENTER(&rx_stats_mutex);
2575 rx_stats.spuriousPacketsRead++;
2576 MUTEX_EXIT(&rx_stats_mutex);
2577 MUTEX_ENTER(&conn->conn_data_lock);
2579 MUTEX_EXIT(&conn->conn_data_lock);
2584 currentCallNumber = conn->callNumber[channel];
2586 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2587 if (np->header.callNumber < currentCallNumber) {
2588 MUTEX_ENTER(&rx_stats_mutex);
2589 rx_stats.spuriousPacketsRead++;
2590 MUTEX_EXIT(&rx_stats_mutex);
2591 #ifdef RX_ENABLE_LOCKS
2593 MUTEX_EXIT(&call->lock);
2595 MUTEX_ENTER(&conn->conn_data_lock);
2597 MUTEX_EXIT(&conn->conn_data_lock);
2601 MUTEX_ENTER(&conn->conn_call_lock);
2602 call = rxi_NewCall(conn, channel);
2603 MUTEX_EXIT(&conn->conn_call_lock);
2604 *call->callNumber = np->header.callNumber;
2605 if (np->header.callNumber == 0)
2606 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));
2608 call->state = RX_STATE_PRECALL;
2609 clock_GetTime(&call->queueTime);
2610 hzero(call->bytesSent);
2611 hzero(call->bytesRcvd);
2613 * If the number of queued calls exceeds the overload
2614 * threshold then abort this call.
2616 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2617 struct rx_packet *tp;
2619 rxi_CallError(call, rx_BusyError);
2620 tp = rxi_SendCallAbort(call, np, 1, 0);
2621 MUTEX_EXIT(&call->lock);
2622 MUTEX_ENTER(&conn->conn_data_lock);
2624 MUTEX_EXIT(&conn->conn_data_lock);
2625 MUTEX_ENTER(&rx_stats_mutex);
2627 MUTEX_EXIT(&rx_stats_mutex);
2630 rxi_KeepAliveOn(call);
2631 } else if (np->header.callNumber != currentCallNumber) {
2632 /* Wait until the transmit queue is idle before deciding
2633 * whether to reset the current call. Chances are that the
2634 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2637 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2638 while ((call->state == RX_STATE_ACTIVE)
2639 && (call->flags & RX_CALL_TQ_BUSY)) {
2640 call->flags |= RX_CALL_TQ_WAIT;
2641 #ifdef RX_ENABLE_LOCKS
2642 CV_WAIT(&call->cv_tq, &call->lock);
2643 #else /* RX_ENABLE_LOCKS */
2644 osi_rxSleep(&call->tq);
2645 #endif /* RX_ENABLE_LOCKS */
2647 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2648 /* If the new call cannot be taken right now send a busy and set
2649 * the error condition in this call, so that it terminates as
2650 * quickly as possible */
2651 if (call->state == RX_STATE_ACTIVE) {
2652 struct rx_packet *tp;
2654 rxi_CallError(call, RX_CALL_DEAD);
2655 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2657 MUTEX_EXIT(&call->lock);
2658 MUTEX_ENTER(&conn->conn_data_lock);
2660 MUTEX_EXIT(&conn->conn_data_lock);
2663 rxi_ResetCall(call, 0);
2664 *call->callNumber = np->header.callNumber;
2665 if (np->header.callNumber == 0)
2666 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));
2668 call->state = RX_STATE_PRECALL;
2669 clock_GetTime(&call->queueTime);
2670 hzero(call->bytesSent);
2671 hzero(call->bytesRcvd);
2673 * If the number of queued calls exceeds the overload
2674 * threshold then abort this call.
2676 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2677 struct rx_packet *tp;
2679 rxi_CallError(call, rx_BusyError);
2680 tp = rxi_SendCallAbort(call, np, 1, 0);
2681 MUTEX_EXIT(&call->lock);
2682 MUTEX_ENTER(&conn->conn_data_lock);
2684 MUTEX_EXIT(&conn->conn_data_lock);
2685 MUTEX_ENTER(&rx_stats_mutex);
2687 MUTEX_EXIT(&rx_stats_mutex);
2690 rxi_KeepAliveOn(call);
2692 /* Continuing call; do nothing here. */
2694 } else { /* we're the client */
2695 /* Ignore all incoming acknowledgements for calls in DALLY state */
2696 if (call && (call->state == RX_STATE_DALLY)
2697 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2698 MUTEX_ENTER(&rx_stats_mutex);
2699 rx_stats.ignorePacketDally++;
2700 MUTEX_EXIT(&rx_stats_mutex);
2701 #ifdef RX_ENABLE_LOCKS
2703 MUTEX_EXIT(&call->lock);
2706 MUTEX_ENTER(&conn->conn_data_lock);
2708 MUTEX_EXIT(&conn->conn_data_lock);
2712 /* Ignore anything that's not relevant to the current call. If there
2713 * isn't a current call, then no packet is relevant. */
2714 if (!call || (np->header.callNumber != currentCallNumber)) {
2715 MUTEX_ENTER(&rx_stats_mutex);
2716 rx_stats.spuriousPacketsRead++;
2717 MUTEX_EXIT(&rx_stats_mutex);
2718 #ifdef RX_ENABLE_LOCKS
2720 MUTEX_EXIT(&call->lock);
2723 MUTEX_ENTER(&conn->conn_data_lock);
2725 MUTEX_EXIT(&conn->conn_data_lock);
2728 /* If the service security object index stamped in the packet does not
2729 * match the connection's security index, ignore the packet */
2730 if (np->header.securityIndex != conn->securityIndex) {
2731 #ifdef RX_ENABLE_LOCKS
2732 MUTEX_EXIT(&call->lock);
2734 MUTEX_ENTER(&conn->conn_data_lock);
2736 MUTEX_EXIT(&conn->conn_data_lock);
2740 /* If we're receiving the response, then all transmit packets are
2741 * implicitly acknowledged. Get rid of them. */
2742 if (np->header.type == RX_PACKET_TYPE_DATA) {
2743 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2744 /* XXX Hack. Because we must release the global rx lock when
2745 * sending packets (osi_NetSend) we drop all acks while we're
2746 * traversing the tq in rxi_Start sending packets out because
2747 * packets may move to the freePacketQueue as result of being here!
2748 * So we drop these packets until we're safely out of the
2749 * traversing. Really ugly!
2750 * For fine grain RX locking, we set the acked field in the
2751 * packets and let rxi_Start remove them from the transmit queue.
2753 if (call->flags & RX_CALL_TQ_BUSY) {
2754 #ifdef RX_ENABLE_LOCKS
2755 rxi_SetAcksInTransmitQueue(call);
2758 return np; /* xmitting; drop packet */
2761 rxi_ClearTransmitQueue(call, 0);
2763 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2764 rxi_ClearTransmitQueue(call, 0);
2765 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2767 if (np->header.type == RX_PACKET_TYPE_ACK) {
2768 /* now check to see if this is an ack packet acknowledging that the
2769 * server actually *lost* some hard-acked data. If this happens we
2770 * ignore this packet, as it may indicate that the server restarted in
2771 * the middle of a call. It is also possible that this is an old ack
2772 * packet. We don't abort the connection in this case, because this
2773 * *might* just be an old ack packet. The right way to detect a server
2774 * restart in the midst of a call is to notice that the server epoch
2776 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2777 * XXX unacknowledged. I think that this is off-by-one, but
2778 * XXX I don't dare change it just yet, since it will
2779 * XXX interact badly with the server-restart detection
2780 * XXX code in receiveackpacket. */
2781 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2782 MUTEX_ENTER(&rx_stats_mutex);
2783 rx_stats.spuriousPacketsRead++;
2784 MUTEX_EXIT(&rx_stats_mutex);
2785 MUTEX_EXIT(&call->lock);
2786 MUTEX_ENTER(&conn->conn_data_lock);
2788 MUTEX_EXIT(&conn->conn_data_lock);
2792 } /* else not a data packet */
2795 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2796 /* Set remote user defined status from packet */
2797 call->remoteStatus = np->header.userStatus;
2799 /* Note the gap between the expected next packet and the actual
2800 * packet that arrived, when the new packet has a smaller serial number
2801 * than expected. Rioses frequently reorder packets all by themselves,
2802 * so this will be quite important with very large window sizes.
2803 * Skew is checked against 0 here to avoid any dependence on the type of
2804 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2806 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2807 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2808 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2810 MUTEX_ENTER(&conn->conn_data_lock);
2811 skew = conn->lastSerial - np->header.serial;
2812 conn->lastSerial = np->header.serial;
2813 MUTEX_EXIT(&conn->conn_data_lock);
2815 register struct rx_peer *peer;
2817 if (skew > peer->inPacketSkew) {
2818 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2820 peer->inPacketSkew = skew;
2824 /* Now do packet type-specific processing */
2825 switch (np->header.type) {
2826 case RX_PACKET_TYPE_DATA:
2827 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2830 case RX_PACKET_TYPE_ACK:
2831 /* Respond immediately to ack packets requesting acknowledgement
2833 if (np->header.flags & RX_REQUEST_ACK) {
2835 (void)rxi_SendCallAbort(call, 0, 1, 0);
2837 (void)rxi_SendAck(call, 0, np->header.serial,
2838 RX_ACK_PING_RESPONSE, 1);
2840 np = rxi_ReceiveAckPacket(call, np, 1);
2842 case RX_PACKET_TYPE_ABORT:
2843 /* An abort packet: reset the connection, passing the error up to
2845 /* What if error is zero? */
2846 rxi_CallError(call, ntohl(*(afs_int32 *) rx_DataOf(np)));
2848 case RX_PACKET_TYPE_BUSY:
2851 case RX_PACKET_TYPE_ACKALL:
2852 /* All packets acknowledged, so we can drop all packets previously
2853 * readied for sending */
2854 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2855 /* XXX Hack. We because we can't release the global rx lock when
2856 * sending packets (osi_NetSend) we drop all ack pkts while we're
2857 * traversing the tq in rxi_Start sending packets out because
2858 * packets may move to the freePacketQueue as result of being
2859 * here! So we drop these packets until we're safely out of the
2860 * traversing. Really ugly!
2861 * For fine grain RX locking, we set the acked field in the packets
2862 * and let rxi_Start remove the packets from the transmit queue.
2864 if (call->flags & RX_CALL_TQ_BUSY) {
2865 #ifdef RX_ENABLE_LOCKS
2866 rxi_SetAcksInTransmitQueue(call);
2868 #else /* RX_ENABLE_LOCKS */
2870 return np; /* xmitting; drop packet */
2871 #endif /* RX_ENABLE_LOCKS */
2873 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2874 rxi_ClearTransmitQueue(call, 0);
2877 /* Should not reach here, unless the peer is broken: send an abort
2879 rxi_CallError(call, RX_PROTOCOL_ERROR);
2880 np = rxi_SendCallAbort(call, np, 1, 0);
2883 /* Note when this last legitimate packet was received, for keep-alive
2884 * processing. Note, we delay getting the time until now in the hope that
2885 * the packet will be delivered to the user before any get time is required
2886 * (if not, then the time won't actually be re-evaluated here). */
2887 call->lastReceiveTime = clock_Sec();
2888 MUTEX_EXIT(&call->lock);
2889 MUTEX_ENTER(&conn->conn_data_lock);
2891 MUTEX_EXIT(&conn->conn_data_lock);
2895 /* return true if this is an "interesting" connection from the point of view
2896 of someone trying to debug the system */
2898 rxi_IsConnInteresting(struct rx_connection *aconn)
2901 register struct rx_call *tcall;
2903 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2905 for (i = 0; i < RX_MAXCALLS; i++) {
2906 tcall = aconn->call[i];
2908 if ((tcall->state == RX_STATE_PRECALL)
2909 || (tcall->state == RX_STATE_ACTIVE))
2911 if ((tcall->mode == RX_MODE_SENDING)
2912 || (tcall->mode == RX_MODE_RECEIVING))
2920 /* if this is one of the last few packets AND it wouldn't be used by the
2921 receiving call to immediately satisfy a read request, then drop it on
2922 the floor, since accepting it might prevent a lock-holding thread from
2923 making progress in its reading. If a call has been cleared while in
2924 the precall state then ignore all subsequent packets until the call
2925 is assigned to a thread. */
2928 TooLow(struct rx_packet *ap, struct rx_call *acall)
2931 MUTEX_ENTER(&rx_stats_mutex);
2932 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2933 && (acall->state == RX_STATE_PRECALL))
2934 || ((rx_nFreePackets < rxi_dataQuota + 2)
2935 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2936 && (acall->flags & RX_CALL_READER_WAIT)))) {
2939 MUTEX_EXIT(&rx_stats_mutex);
2945 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2946 struct rx_call *acall)
2948 struct rx_call *call = acall;
2952 MUTEX_ENTER(&conn->conn_data_lock);
2953 conn->checkReachEvent = NULL;
2954 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2957 MUTEX_EXIT(&conn->conn_data_lock);
2961 MUTEX_ENTER(&conn->conn_call_lock);
2962 MUTEX_ENTER(&conn->conn_data_lock);
2963 for (i = 0; i < RX_MAXCALLS; i++) {
2964 struct rx_call *tc = conn->call[i];
2965 if (tc && tc->state == RX_STATE_PRECALL) {
2971 /* Indicate that rxi_CheckReachEvent is no longer running by
2972 * clearing the flag. Must be atomic under conn_data_lock to
2973 * avoid a new call slipping by: rxi_CheckConnReach holds
2974 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
2976 conn->flags &= ~RX_CONN_ATTACHWAIT;
2977 MUTEX_EXIT(&conn->conn_data_lock);
2978 MUTEX_EXIT(&conn->conn_call_lock);
2983 MUTEX_ENTER(&call->lock);
2984 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
2986 MUTEX_EXIT(&call->lock);
2988 clock_GetTime(&when);
2989 when.sec += RX_CHECKREACH_TIMEOUT;
2990 MUTEX_ENTER(&conn->conn_data_lock);
2991 if (!conn->checkReachEvent) {
2993 conn->checkReachEvent =
2994 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
2996 MUTEX_EXIT(&conn->conn_data_lock);
3002 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3004 struct rx_service *service = conn->service;
3005 struct rx_peer *peer = conn->peer;
3006 afs_uint32 now, lastReach;
3008 if (service->checkReach == 0)
3012 MUTEX_ENTER(&peer->peer_lock);
3013 lastReach = peer->lastReachTime;
3014 MUTEX_EXIT(&peer->peer_lock);
3015 if (now - lastReach < RX_CHECKREACH_TTL)
3018 MUTEX_ENTER(&conn->conn_data_lock);
3019 if (conn->flags & RX_CONN_ATTACHWAIT) {
3020 MUTEX_EXIT(&conn->conn_data_lock);
3023 conn->flags |= RX_CONN_ATTACHWAIT;
3024 MUTEX_EXIT(&conn->conn_data_lock);
3025 if (!conn->checkReachEvent)
3026 rxi_CheckReachEvent(NULL, conn, call);
3031 /* try to attach call, if authentication is complete */
3033 TryAttach(register struct rx_call *acall, register osi_socket socket,
3034 register int *tnop, register struct rx_call **newcallp,
3037 struct rx_connection *conn = acall->conn;
3039 if (conn->type == RX_SERVER_CONNECTION
3040 && acall->state == RX_STATE_PRECALL) {
3041 /* Don't attach until we have any req'd. authentication. */
3042 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3043 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3044 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3045 /* Note: this does not necessarily succeed; there
3046 * may not any proc available
3049 rxi_ChallengeOn(acall->conn);
3054 /* A data packet has been received off the interface. This packet is
3055 * appropriate to the call (the call is in the right state, etc.). This
3056 * routine can return a packet to the caller, for re-use */
3059 rxi_ReceiveDataPacket(register struct rx_call *call,
3060 register struct rx_packet *np, int istack,
3061 osi_socket socket, afs_uint32 host, u_short port,
3062 int *tnop, struct rx_call **newcallp)
3064 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3068 afs_uint32 seq, serial, flags;
3070 struct rx_packet *tnp;
3072 MUTEX_ENTER(&rx_stats_mutex);
3073 rx_stats.dataPacketsRead++;
3074 MUTEX_EXIT(&rx_stats_mutex);
3077 /* If there are no packet buffers, drop this new packet, unless we can find
3078 * packet buffers from inactive calls */
3080 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3081 MUTEX_ENTER(&rx_freePktQ_lock);
3082 rxi_NeedMorePackets = TRUE;
3083 MUTEX_EXIT(&rx_freePktQ_lock);
3084 MUTEX_ENTER(&rx_stats_mutex);
3085 rx_stats.noPacketBuffersOnRead++;
3086 MUTEX_EXIT(&rx_stats_mutex);
3087 call->rprev = np->header.serial;
3088 rxi_calltrace(RX_TRACE_DROP, call);
3089 dpf(("packet %x dropped on receipt - quota problems", np));
3091 rxi_ClearReceiveQueue(call);
3092 clock_GetTime(&when);
3093 clock_Add(&when, &rx_softAckDelay);
3094 if (!call->delayedAckEvent
3095 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3096 rxevent_Cancel(call->delayedAckEvent, call,
3097 RX_CALL_REFCOUNT_DELAY);
3098 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3099 call->delayedAckEvent =
3100 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3102 /* we've damaged this call already, might as well do it in. */
3108 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3109 * packet is one of several packets transmitted as a single
3110 * datagram. Do not send any soft or hard acks until all packets
3111 * in a jumbogram have been processed. Send negative acks right away.
3113 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3114 /* tnp is non-null when there are more packets in the
3115 * current jumbo gram */
3122 seq = np->header.seq;
3123 serial = np->header.serial;
3124 flags = np->header.flags;
3126 /* If the call is in an error state, send an abort message */
3128 return rxi_SendCallAbort(call, np, istack, 0);
3130 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3131 * AFS 3.5 jumbogram. */
3132 if (flags & RX_JUMBO_PACKET) {
3133 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3138 if (np->header.spare != 0) {
3139 MUTEX_ENTER(&call->conn->conn_data_lock);
3140 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3141 MUTEX_EXIT(&call->conn->conn_data_lock);
3144 /* The usual case is that this is the expected next packet */
3145 if (seq == call->rnext) {
3147 /* Check to make sure it is not a duplicate of one already queued */
3148 if (queue_IsNotEmpty(&call->rq)
3149 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3150 MUTEX_ENTER(&rx_stats_mutex);
3151 rx_stats.dupPacketsRead++;
3152 MUTEX_EXIT(&rx_stats_mutex);
3153 dpf(("packet %x dropped on receipt - duplicate", np));
3154 rxevent_Cancel(call->delayedAckEvent, call,
3155 RX_CALL_REFCOUNT_DELAY);
3156 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3162 /* It's the next packet. Stick it on the receive queue
3163 * for this call. Set newPackets to make sure we wake
3164 * the reader once all packets have been processed */
3165 queue_Prepend(&call->rq, np);
3167 np = NULL; /* We can't use this anymore */
3170 /* If an ack is requested then set a flag to make sure we
3171 * send an acknowledgement for this packet */
3172 if (flags & RX_REQUEST_ACK) {
3173 ackNeeded = RX_ACK_REQUESTED;
3176 /* Keep track of whether we have received the last packet */
3177 if (flags & RX_LAST_PACKET) {
3178 call->flags |= RX_CALL_HAVE_LAST;
3182 /* Check whether we have all of the packets for this call */
3183 if (call->flags & RX_CALL_HAVE_LAST) {
3184 afs_uint32 tseq; /* temporary sequence number */
3185 struct rx_packet *tp; /* Temporary packet pointer */
3186 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3188 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3189 if (tseq != tp->header.seq)
3191 if (tp->header.flags & RX_LAST_PACKET) {
3192 call->flags |= RX_CALL_RECEIVE_DONE;
3199 /* Provide asynchronous notification for those who want it
3200 * (e.g. multi rx) */
3201 if (call->arrivalProc) {
3202 (*call->arrivalProc) (call, call->arrivalProcHandle,
3203 call->arrivalProcArg);
3204 call->arrivalProc = (void (*)())0;
3207 /* Update last packet received */
3210 /* If there is no server process serving this call, grab
3211 * one, if available. We only need to do this once. If a
3212 * server thread is available, this thread becomes a server
3213 * thread and the server thread becomes a listener thread. */
3215 TryAttach(call, socket, tnop, newcallp, 0);
3218 /* This is not the expected next packet. */
3220 /* Determine whether this is a new or old packet, and if it's
3221 * a new one, whether it fits into the current receive window.
3222 * Also figure out whether the packet was delivered in sequence.
3223 * We use the prev variable to determine whether the new packet
3224 * is the successor of its immediate predecessor in the
3225 * receive queue, and the missing flag to determine whether
3226 * any of this packets predecessors are missing. */
3228 afs_uint32 prev; /* "Previous packet" sequence number */
3229 struct rx_packet *tp; /* Temporary packet pointer */
3230 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3231 int missing; /* Are any predecessors missing? */
3233 /* If the new packet's sequence number has been sent to the
3234 * application already, then this is a duplicate */
3235 if (seq < call->rnext) {
3236 MUTEX_ENTER(&rx_stats_mutex);
3237 rx_stats.dupPacketsRead++;
3238 MUTEX_EXIT(&rx_stats_mutex);
3239 rxevent_Cancel(call->delayedAckEvent, call,
3240 RX_CALL_REFCOUNT_DELAY);
3241 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3247 /* If the sequence number is greater than what can be
3248 * accomodated by the current window, then send a negative
3249 * acknowledge and drop the packet */
3250 if ((call->rnext + call->rwind) <= seq) {
3251 rxevent_Cancel(call->delayedAckEvent, call,
3252 RX_CALL_REFCOUNT_DELAY);
3253 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3260 /* Look for the packet in the queue of old received packets */
3261 for (prev = call->rnext - 1, missing =
3262 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3263 /*Check for duplicate packet */
3264 if (seq == tp->header.seq) {
3265 MUTEX_ENTER(&rx_stats_mutex);
3266 rx_stats.dupPacketsRead++;
3267 MUTEX_EXIT(&rx_stats_mutex);
3268 rxevent_Cancel(call->delayedAckEvent, call,
3269 RX_CALL_REFCOUNT_DELAY);
3270 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3276 /* If we find a higher sequence packet, break out and
3277 * insert the new packet here. */
3278 if (seq < tp->header.seq)
3280 /* Check for missing packet */
3281 if (tp->header.seq != prev + 1) {
3285 prev = tp->header.seq;
3288 /* Keep track of whether we have received the last packet. */
3289 if (flags & RX_LAST_PACKET) {
3290 call->flags |= RX_CALL_HAVE_LAST;
3293 /* It's within the window: add it to the the receive queue.
3294 * tp is left by the previous loop either pointing at the
3295 * packet before which to insert the new packet, or at the
3296 * queue head if the queue is empty or the packet should be
3298 queue_InsertBefore(tp, np);
3302 /* Check whether we have all of the packets for this call */
3303 if ((call->flags & RX_CALL_HAVE_LAST)
3304 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3305 afs_uint32 tseq; /* temporary sequence number */
3308 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3309 if (tseq != tp->header.seq)
3311 if (tp->header.flags & RX_LAST_PACKET) {
3312 call->flags |= RX_CALL_RECEIVE_DONE;
3319 /* We need to send an ack of the packet is out of sequence,
3320 * or if an ack was requested by the peer. */
3321 if (seq != prev + 1 || missing || (flags & RX_REQUEST_ACK)) {
3322 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3325 /* Acknowledge the last packet for each call */
3326 if (flags & RX_LAST_PACKET) {
3337 * If the receiver is waiting for an iovec, fill the iovec
3338 * using the data from the receive queue */
3339 if (call->flags & RX_CALL_IOVEC_WAIT) {
3340 didHardAck = rxi_FillReadVec(call, serial);
3341 /* the call may have been aborted */
3350 /* Wakeup the reader if any */
3351 if ((call->flags & RX_CALL_READER_WAIT)
3352 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3353 || (call->iovNext >= call->iovMax)
3354 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3355 call->flags &= ~RX_CALL_READER_WAIT;
3356 #ifdef RX_ENABLE_LOCKS
3357 CV_BROADCAST(&call->cv_rq);
3359 osi_rxWakeup(&call->rq);
3365 * Send an ack when requested by the peer, or once every
3366 * rxi_SoftAckRate packets until the last packet has been
3367 * received. Always send a soft ack for the last packet in
3368 * the server's reply. */
3370 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3371 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3372 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3373 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3374 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3375 } else if (call->nSoftAcks) {
3376 clock_GetTime(&when);
3377 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3378 clock_Add(&when, &rx_lastAckDelay);
3380 clock_Add(&when, &rx_softAckDelay);
3382 if (!call->delayedAckEvent
3383 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3384 rxevent_Cancel(call->delayedAckEvent, call,
3385 RX_CALL_REFCOUNT_DELAY);
3386 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3387 call->delayedAckEvent =
3388 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3390 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3391 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3398 static void rxi_ComputeRate();
3402 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3404 struct rx_peer *peer = conn->peer;
3406 MUTEX_ENTER(&peer->peer_lock);
3407 peer->lastReachTime = clock_Sec();
3408 MUTEX_EXIT(&peer->peer_lock);
3410 MUTEX_ENTER(&conn->conn_data_lock);
3411 if (conn->flags & RX_CONN_ATTACHWAIT) {
3414 conn->flags &= ~RX_CONN_ATTACHWAIT;
3415 MUTEX_EXIT(&conn->conn_data_lock);
3417 for (i = 0; i < RX_MAXCALLS; i++) {
3418 struct rx_call *call = conn->call[i];
3421 MUTEX_ENTER(&call->lock);
3422 /* tnop can be null if newcallp is null */
3423 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3425 MUTEX_EXIT(&call->lock);
3429 MUTEX_EXIT(&conn->conn_data_lock);
3432 /* rxi_ComputePeerNetStats
3434 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3435 * estimates (like RTT and throughput) based on ack packets. Caller
3436 * must ensure that the packet in question is the right one (i.e.
3437 * serial number matches).
3440 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3441 struct rx_ackPacket *ap, struct rx_packet *np)
3443 struct rx_peer *peer = call->conn->peer;
3445 /* Use RTT if not delayed by client. */
3446 if (ap->reason != RX_ACK_DELAY)
3447 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3449 rxi_ComputeRate(peer, call, p, np, ap->reason);
3453 /* The real smarts of the whole thing. */
3455 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3458 struct rx_ackPacket *ap;
3460 register struct rx_packet *tp;
3461 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3462 register struct rx_connection *conn = call->conn;
3463 struct rx_peer *peer = conn->peer;
3466 /* because there are CM's that are bogus, sending weird values for this. */
3467 afs_uint32 skew = 0;
3472 int newAckCount = 0;
3473 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3474 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3476 MUTEX_ENTER(&rx_stats_mutex);
3477 rx_stats.ackPacketsRead++;
3478 MUTEX_EXIT(&rx_stats_mutex);
3479 ap = (struct rx_ackPacket *)rx_DataOf(np);
3480 nbytes = rx_Contiguous(np) - ((ap->acks) - (u_char *) ap);
3482 return np; /* truncated ack packet */
3484 /* depends on ack packet struct */
3485 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3486 first = ntohl(ap->firstPacket);
3487 serial = ntohl(ap->serial);
3488 /* temporarily disabled -- needs to degrade over time
3489 * skew = ntohs(ap->maxSkew); */
3491 /* Ignore ack packets received out of order */
3492 if (first < call->tfirst) {
3496 if (np->header.flags & RX_SLOW_START_OK) {
3497 call->flags |= RX_CALL_SLOW_START_OK;
3500 if (ap->reason == RX_ACK_PING_RESPONSE)
3501 rxi_UpdatePeerReach(conn, call);
3506 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3507 ap->reason, ntohl(ap->previousPacket),
3508 (unsigned int)np->header.seq, (unsigned int)serial,
3509 (unsigned int)skew, ntohl(ap->firstPacket));
3512 for (offset = 0; offset < nAcks; offset++)
3513 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3520 /* Update the outgoing packet skew value to the latest value of
3521 * the peer's incoming packet skew value. The ack packet, of
3522 * course, could arrive out of order, but that won't affect things
3524 MUTEX_ENTER(&peer->peer_lock);
3525 peer->outPacketSkew = skew;
3527 /* Check for packets that no longer need to be transmitted, and
3528 * discard them. This only applies to packets positively
3529 * acknowledged as having been sent to the peer's upper level.
3530 * All other packets must be retained. So only packets with
3531 * sequence numbers < ap->firstPacket are candidates. */
3532 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3533 if (tp->header.seq >= first)
3535 call->tfirst = tp->header.seq + 1;
3537 && (tp->header.serial == serial || tp->firstSerial == serial))
3538 rxi_ComputePeerNetStats(call, tp, ap, np);
3539 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3540 /* XXX Hack. Because we have to release the global rx lock when sending
3541 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3542 * in rxi_Start sending packets out because packets may move to the
3543 * freePacketQueue as result of being here! So we drop these packets until
3544 * we're safely out of the traversing. Really ugly!
3545 * To make it even uglier, if we're using fine grain locking, we can
3546 * set the ack bits in the packets and have rxi_Start remove the packets
3547 * when it's done transmitting.
3549 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3552 if (call->flags & RX_CALL_TQ_BUSY) {
3553 #ifdef RX_ENABLE_LOCKS
3554 tp->flags |= RX_PKTFLAG_ACKED;
3555 call->flags |= RX_CALL_TQ_SOME_ACKED;
3556 #else /* RX_ENABLE_LOCKS */
3558 #endif /* RX_ENABLE_LOCKS */
3560 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3563 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3568 /* Give rate detector a chance to respond to ping requests */
3569 if (ap->reason == RX_ACK_PING_RESPONSE) {
3570 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3574 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3576 /* Now go through explicit acks/nacks and record the results in
3577 * the waiting packets. These are packets that can't be released
3578 * yet, even with a positive acknowledge. This positive
3579 * acknowledge only means the packet has been received by the
3580 * peer, not that it will be retained long enough to be sent to
3581 * the peer's upper level. In addition, reset the transmit timers
3582 * of any missing packets (those packets that must be missing
3583 * because this packet was out of sequence) */
3585 call->nSoftAcked = 0;
3586 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3587 /* Update round trip time if the ack was stimulated on receipt
3589 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3590 #ifdef RX_ENABLE_LOCKS
3591 if (tp->header.seq >= first)
3592 #endif /* RX_ENABLE_LOCKS */
3593 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3595 && (tp->header.serial == serial || tp->firstSerial == serial))
3596 rxi_ComputePeerNetStats(call, tp, ap, np);
3598 /* Set the acknowledge flag per packet based on the
3599 * information in the ack packet. An acknowlegded packet can
3600 * be downgraded when the server has discarded a packet it
3601 * soacked previously, or when an ack packet is received
3602 * out of sequence. */
3603 if (tp->header.seq < first) {
3604 /* Implicit ack information */
3605 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3608 tp->flags |= RX_PKTFLAG_ACKED;
3609 } else if (tp->header.seq < first + nAcks) {
3610 /* Explicit ack information: set it in the packet appropriately */
3611 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3612 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3614 tp->flags |= RX_PKTFLAG_ACKED;
3622 tp->flags &= ~RX_PKTFLAG_ACKED;
3626 tp->flags &= ~RX_PKTFLAG_ACKED;
3630 /* If packet isn't yet acked, and it has been transmitted at least
3631 * once, reset retransmit time using latest timeout
3632 * ie, this should readjust the retransmit timer for all outstanding
3633 * packets... So we don't just retransmit when we should know better*/
3635 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3636 tp->retryTime = tp->timeSent;
3637 clock_Add(&tp->retryTime, &peer->timeout);
3638 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3639 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3643 /* If the window has been extended by this acknowledge packet,
3644 * then wakeup a sender waiting in alloc for window space, or try
3645 * sending packets now, if he's been sitting on packets due to
3646 * lack of window space */
3647 if (call->tnext < (call->tfirst + call->twind)) {
3648 #ifdef RX_ENABLE_LOCKS
3649 CV_SIGNAL(&call->cv_twind);
3651 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3652 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3653 osi_rxWakeup(&call->twind);
3656 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3657 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3661 /* if the ack packet has a receivelen field hanging off it,
3662 * update our state */
3663 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3666 /* If the ack packet has a "recommended" size that is less than
3667 * what I am using now, reduce my size to match */
3668 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3669 sizeof(afs_int32), &tSize);
3670 tSize = (afs_uint32) ntohl(tSize);
3671 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3673 /* Get the maximum packet size to send to this peer */
3674 rx_packetread(np, rx_AckDataSize(ap->nAcks), sizeof(afs_int32),
3676 tSize = (afs_uint32) ntohl(tSize);
3677 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3678 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3680 /* sanity check - peer might have restarted with different params.
3681 * If peer says "send less", dammit, send less... Peer should never
3682 * be unable to accept packets of the size that prior AFS versions would
3683 * send without asking. */
3684 if (peer->maxMTU != tSize) {
3685 peer->maxMTU = tSize;
3686 peer->MTU = MIN(tSize, peer->MTU);
3687 call->MTU = MIN(call->MTU, tSize);
3691 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3694 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3695 sizeof(afs_int32), &tSize);
3696 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3697 if (tSize < call->twind) { /* smaller than our send */
3698 call->twind = tSize; /* window, we must send less... */
3699 call->ssthresh = MIN(call->twind, call->ssthresh);
3702 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3703 * network MTU confused with the loopback MTU. Calculate the
3704 * maximum MTU here for use in the slow start code below.
3706 maxMTU = peer->maxMTU;
3707 /* Did peer restart with older RX version? */
3708 if (peer->maxDgramPackets > 1) {
3709 peer->maxDgramPackets = 1;
3711 } else if (np->length >=
3712 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3715 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3716 sizeof(afs_int32), &tSize);
3717 tSize = (afs_uint32) ntohl(tSize);
3719 * As of AFS 3.5 we set the send window to match the receive window.
3721 if (tSize < call->twind) {
3722 call->twind = tSize;
3723 call->ssthresh = MIN(call->twind, call->ssthresh);
3724 } else if (tSize > call->twind) {
3725 call->twind = tSize;
3729 * As of AFS 3.5, a jumbogram is more than one fixed size
3730 * packet transmitted in a single UDP datagram. If the remote
3731 * MTU is smaller than our local MTU then never send a datagram
3732 * larger than the natural MTU.
3735 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3736 sizeof(afs_int32), &tSize);
3737 maxDgramPackets = (afs_uint32) ntohl(tSize);
3738 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3740 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3741 maxDgramPackets = MIN(maxDgramPackets, tSize);
3742 if (maxDgramPackets > 1) {
3743 peer->maxDgramPackets = maxDgramPackets;
3744 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3746 peer->maxDgramPackets = 1;
3747 call->MTU = peer->natMTU;
3749 } else if (peer->maxDgramPackets > 1) {
3750 /* Restarted with lower version of RX */
3751 peer->maxDgramPackets = 1;
3753 } else if (peer->maxDgramPackets > 1
3754 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3755 /* Restarted with lower version of RX */
3756 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3757 peer->natMTU = OLD_MAX_PACKET_SIZE;
3758 peer->MTU = OLD_MAX_PACKET_SIZE;
3759 peer->maxDgramPackets = 1;
3760 peer->nDgramPackets = 1;
3762 call->MTU = OLD_MAX_PACKET_SIZE;
3767 * Calculate how many datagrams were successfully received after
3768 * the first missing packet and adjust the negative ack counter
3773 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3774 if (call->nNacks < nNacked) {
3775 call->nNacks = nNacked;
3784 if (call->flags & RX_CALL_FAST_RECOVER) {
3786 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3788 call->flags &= ~RX_CALL_FAST_RECOVER;
3789 call->cwind = call->nextCwind;
3790 call->nextCwind = 0;
3793 call->nCwindAcks = 0;
3794 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3795 /* Three negative acks in a row trigger congestion recovery */
3796 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3797 MUTEX_EXIT(&peer->peer_lock);
3798 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3799 /* someone else is waiting to start recovery */
3802 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3803 while (call->flags & RX_CALL_TQ_BUSY) {
3804 call->flags |= RX_CALL_TQ_WAIT;
3805 #ifdef RX_ENABLE_LOCKS
3806 CV_WAIT(&call->cv_tq, &call->lock);
3807 #else /* RX_ENABLE_LOCKS */
3808 osi_rxSleep(&call->tq);
3809 #endif /* RX_ENABLE_LOCKS */
3811 MUTEX_ENTER(&peer->peer_lock);
3812 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3813 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3814 call->flags |= RX_CALL_FAST_RECOVER;
3815 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3817 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3818 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3819 call->nextCwind = call->ssthresh;
3822 peer->MTU = call->MTU;
3823 peer->cwind = call->nextCwind;
3824 peer->nDgramPackets = call->nDgramPackets;
3826 call->congestSeq = peer->congestSeq;
3827 /* Reset the resend times on the packets that were nacked
3828 * so we will retransmit as soon as the window permits*/
3829 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3831 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3832 clock_Zero(&tp->retryTime);
3834 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3839 /* If cwind is smaller than ssthresh, then increase
3840 * the window one packet for each ack we receive (exponential
3842 * If cwind is greater than or equal to ssthresh then increase
3843 * the congestion window by one packet for each cwind acks we
3844 * receive (linear growth). */
3845 if (call->cwind < call->ssthresh) {
3847 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3848 call->nCwindAcks = 0;
3850 call->nCwindAcks += newAckCount;
3851 if (call->nCwindAcks >= call->cwind) {
3852 call->nCwindAcks = 0;
3853 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3857 * If we have received several acknowledgements in a row then
3858 * it is time to increase the size of our datagrams
3860 if ((int)call->nAcks > rx_nDgramThreshold) {
3861 if (peer->maxDgramPackets > 1) {
3862 if (call->nDgramPackets < peer->maxDgramPackets) {
3863 call->nDgramPackets++;
3865 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3866 } else if (call->MTU < peer->maxMTU) {
3867 call->MTU += peer->natMTU;
3868 call->MTU = MIN(call->MTU, peer->maxMTU);
3874 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3876 /* Servers need to hold the call until all response packets have
3877 * been acknowledged. Soft acks are good enough since clients
3878 * are not allowed to clear their receive queues. */
3879 if (call->state == RX_STATE_HOLD
3880 && call->tfirst + call->nSoftAcked >= call->tnext) {
3881 call->state = RX_STATE_DALLY;
3882 rxi_ClearTransmitQueue(call, 0);
3883 } else if (!queue_IsEmpty(&call->tq)) {
3884 rxi_Start(0, call, 0, istack);
3889 /* Received a response to a challenge packet */
3891 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3892 register struct rx_packet *np, int istack)
3896 /* Ignore the packet if we're the client */
3897 if (conn->type == RX_CLIENT_CONNECTION)
3900 /* If already authenticated, ignore the packet (it's probably a retry) */
3901 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3904 /* Otherwise, have the security object evaluate the response packet */
3905 error = RXS_CheckResponse(conn->securityObject, conn, np);
3907 /* If the response is invalid, reset the connection, sending
3908 * an abort to the peer */
3912 rxi_ConnectionError(conn, error);
3913 MUTEX_ENTER(&conn->conn_data_lock);
3914 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3915 MUTEX_EXIT(&conn->conn_data_lock);
3918 /* If the response is valid, any calls waiting to attach
3919 * servers can now do so */
3922 for (i = 0; i < RX_MAXCALLS; i++) {
3923 struct rx_call *call = conn->call[i];
3925 MUTEX_ENTER(&call->lock);
3926 if (call->state == RX_STATE_PRECALL)
3927 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
3928 /* tnop can be null if newcallp is null */
3929 MUTEX_EXIT(&call->lock);
3933 /* Update the peer reachability information, just in case
3934 * some calls went into attach-wait while we were waiting
3935 * for authentication..
3937 rxi_UpdatePeerReach(conn, NULL);
3942 /* A client has received an authentication challenge: the security
3943 * object is asked to cough up a respectable response packet to send
3944 * back to the server. The server is responsible for retrying the
3945 * challenge if it fails to get a response. */
3948 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
3949 register struct rx_packet *np, int istack)
3953 /* Ignore the challenge if we're the server */
3954 if (conn->type == RX_SERVER_CONNECTION)
3957 /* Ignore the challenge if the connection is otherwise idle; someone's
3958 * trying to use us as an oracle. */
3959 if (!rxi_HasActiveCalls(conn))
3962 /* Send the security object the challenge packet. It is expected to fill
3963 * in the response. */
3964 error = RXS_GetResponse(conn->securityObject, conn, np);
3966 /* If the security object is unable to return a valid response, reset the
3967 * connection and send an abort to the peer. Otherwise send the response
3968 * packet to the peer connection. */
3970 rxi_ConnectionError(conn, error);
3971 MUTEX_ENTER(&conn->conn_data_lock);
3972 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3973 MUTEX_EXIT(&conn->conn_data_lock);
3975 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
3976 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
3982 /* Find an available server process to service the current request in
3983 * the given call structure. If one isn't available, queue up this
3984 * call so it eventually gets one */
3986 rxi_AttachServerProc(register struct rx_call *call,
3987 register osi_socket socket, register int *tnop,
3988 register struct rx_call **newcallp)
3990 register struct rx_serverQueueEntry *sq;
3991 register struct rx_service *service = call->conn->service;
3992 register int haveQuota = 0;
3994 /* May already be attached */
3995 if (call->state == RX_STATE_ACTIVE)
3998 MUTEX_ENTER(&rx_serverPool_lock);
4000 haveQuota = QuotaOK(service);
4001 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4002 /* If there are no processes available to service this call,
4003 * put the call on the incoming call queue (unless it's
4004 * already on the queue).
4006 #ifdef RX_ENABLE_LOCKS
4008 ReturnToServerPool(service);
4009 #endif /* RX_ENABLE_LOCKS */
4011 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4012 call->flags |= RX_CALL_WAIT_PROC;
4013 MUTEX_ENTER(&rx_stats_mutex);
4016 MUTEX_EXIT(&rx_stats_mutex);
4017 rxi_calltrace(RX_CALL_ARRIVAL, call);
4018 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4019 queue_Append(&rx_incomingCallQueue, call);
4022 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4024 /* If hot threads are enabled, and both newcallp and sq->socketp
4025 * are non-null, then this thread will process the call, and the
4026 * idle server thread will start listening on this threads socket.
4029 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4032 *sq->socketp = socket;
4033 clock_GetTime(&call->startTime);
4034 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4038 if (call->flags & RX_CALL_WAIT_PROC) {
4039 /* Conservative: I don't think this should happen */
4040 call->flags &= ~RX_CALL_WAIT_PROC;
4041 if (queue_IsOnQueue(call)) {
4043 MUTEX_ENTER(&rx_stats_mutex);
4045 MUTEX_EXIT(&rx_stats_mutex);
4048 call->state = RX_STATE_ACTIVE;
4049 call->mode = RX_MODE_RECEIVING;
4050 #ifdef RX_KERNEL_TRACE
4052 int glockOwner = ISAFS_GLOCK();
4055 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4056 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4062 if (call->flags & RX_CALL_CLEARED) {
4063 /* send an ack now to start the packet flow up again */
4064 call->flags &= ~RX_CALL_CLEARED;
4065 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4067 #ifdef RX_ENABLE_LOCKS
4070 service->nRequestsRunning++;
4071 if (service->nRequestsRunning <= service->minProcs)
4077 MUTEX_EXIT(&rx_serverPool_lock);
4080 /* Delay the sending of an acknowledge event for a short while, while
4081 * a new call is being prepared (in the case of a client) or a reply
4082 * is being prepared (in the case of a server). Rather than sending
4083 * an ack packet, an ACKALL packet is sent. */
4085 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4087 #ifdef RX_ENABLE_LOCKS
4089 MUTEX_ENTER(&call->lock);
4090 call->delayedAckEvent = NULL;
4091 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4093 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4094 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4096 MUTEX_EXIT(&call->lock);
4097 #else /* RX_ENABLE_LOCKS */
4099 call->delayedAckEvent = NULL;
4100 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4101 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4102 #endif /* RX_ENABLE_LOCKS */
4106 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4109 #ifdef RX_ENABLE_LOCKS
4111 MUTEX_ENTER(&call->lock);
4112 if (event == call->delayedAckEvent)
4113 call->delayedAckEvent = NULL;
4114 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4116 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4118 MUTEX_EXIT(&call->lock);
4119 #else /* RX_ENABLE_LOCKS */
4121 call->delayedAckEvent = NULL;
4122 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4123 #endif /* RX_ENABLE_LOCKS */
4127 #ifdef RX_ENABLE_LOCKS
4128 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4129 * clearing them out.
4132 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4134 register struct rx_packet *p, *tp;
4137 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4138 p->flags |= RX_PKTFLAG_ACKED;
4142 call->flags |= RX_CALL_TQ_CLEARME;
4143 call->flags |= RX_CALL_TQ_SOME_ACKED;
4146 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4147 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4148 call->tfirst = call->tnext;
4149 call->nSoftAcked = 0;
4151 if (call->flags & RX_CALL_FAST_RECOVER) {
4152 call->flags &= ~RX_CALL_FAST_RECOVER;
4153 call->cwind = call->nextCwind;
4154 call->nextCwind = 0;
4157 CV_SIGNAL(&call->cv_twind);
4159 #endif /* RX_ENABLE_LOCKS */
4161 /* Clear out the transmit queue for the current call (all packets have
4162 * been received by peer) */
4164 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4166 register struct rx_packet *p, *tp;
4168 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4169 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4171 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4172 p->flags |= RX_PKTFLAG_ACKED;
4176 call->flags |= RX_CALL_TQ_CLEARME;
4177 call->flags |= RX_CALL_TQ_SOME_ACKED;
4180 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4181 rxi_FreePackets(0, &call->tq);
4182 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4183 call->flags &= ~RX_CALL_TQ_CLEARME;
4185 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4187 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4188 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4189 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4190 call->nSoftAcked = 0;
4192 if (call->flags & RX_CALL_FAST_RECOVER) {
4193 call->flags &= ~RX_CALL_FAST_RECOVER;
4194 call->cwind = call->nextCwind;
4196 #ifdef RX_ENABLE_LOCKS
4197 CV_SIGNAL(&call->cv_twind);
4199 osi_rxWakeup(&call->twind);
4204 rxi_ClearReceiveQueue(register struct rx_call *call)
4206 if (queue_IsNotEmpty(&call->rq)) {
4207 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4208 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4210 if (call->state == RX_STATE_PRECALL) {
4211 call->flags |= RX_CALL_CLEARED;
4215 /* Send an abort packet for the specified call */
4217 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4218 int istack, int force)
4226 /* Clients should never delay abort messages */
4227 if (rx_IsClientConn(call->conn))
4230 if (call->abortCode != call->error) {
4231 call->abortCode = call->error;
4232 call->abortCount = 0;
4235 if (force || rxi_callAbortThreshhold == 0
4236 || call->abortCount < rxi_callAbortThreshhold) {
4237 if (call->delayedAbortEvent) {
4238 rxevent_Cancel(call->delayedAbortEvent, call,
4239 RX_CALL_REFCOUNT_ABORT);
4241 error = htonl(call->error);
4244 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4245 (char *)&error, sizeof(error), istack);
4246 } else if (!call->delayedAbortEvent) {
4247 clock_GetTime(&when);
4248 clock_Addmsec(&when, rxi_callAbortDelay);
4249 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4250 call->delayedAbortEvent =
4251 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4256 /* Send an abort packet for the specified connection. Packet is an
4257 * optional pointer to a packet that can be used to send the abort.
4258 * Once the number of abort messages reaches the threshhold, an
4259 * event is scheduled to send the abort. Setting the force flag
4260 * overrides sending delayed abort messages.
4262 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4263 * to send the abort packet.
4266 rxi_SendConnectionAbort(register struct rx_connection *conn,
4267 struct rx_packet *packet, int istack, int force)
4275 /* Clients should never delay abort messages */
4276 if (rx_IsClientConn(conn))
4279 if (force || rxi_connAbortThreshhold == 0
4280 || conn->abortCount < rxi_connAbortThreshhold) {
4281 if (conn->delayedAbortEvent) {
4282 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4284 error = htonl(conn->error);
4286 MUTEX_EXIT(&conn->conn_data_lock);
4288 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4289 RX_PACKET_TYPE_ABORT, (char *)&error,
4290 sizeof(error), istack);
4291 MUTEX_ENTER(&conn->conn_data_lock);
4292 } else if (!conn->delayedAbortEvent) {
4293 clock_GetTime(&when);
4294 clock_Addmsec(&when, rxi_connAbortDelay);
4295 conn->delayedAbortEvent =
4296 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4301 /* Associate an error all of the calls owned by a connection. Called
4302 * with error non-zero. This is only for really fatal things, like
4303 * bad authentication responses. The connection itself is set in
4304 * error at this point, so that future packets received will be
4307 rxi_ConnectionError(register struct rx_connection *conn,
4308 register afs_int32 error)
4312 MUTEX_ENTER(&conn->conn_data_lock);
4313 if (conn->challengeEvent)
4314 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4315 if (conn->checkReachEvent) {
4316 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4317 conn->checkReachEvent = 0;
4318 conn->flags &= ~RX_CONN_ATTACHWAIT;
4321 MUTEX_EXIT(&conn->conn_data_lock);
4322 for (i = 0; i < RX_MAXCALLS; i++) {
4323 struct rx_call *call = conn->call[i];
4325 MUTEX_ENTER(&call->lock);
4326 rxi_CallError(call, error);
4327 MUTEX_EXIT(&call->lock);
4330 conn->error = error;
4331 MUTEX_ENTER(&rx_stats_mutex);
4332 rx_stats.fatalErrors++;
4333 MUTEX_EXIT(&rx_stats_mutex);
4338 rxi_CallError(register struct rx_call *call, afs_int32 error)
4341 error = call->error;
4342 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4343 if (!(call->flags & RX_CALL_TQ_BUSY)) {
4344 rxi_ResetCall(call, 0);
4347 rxi_ResetCall(call, 0);
4349 call->error = error;
4350 call->mode = RX_MODE_ERROR;
4353 /* Reset various fields in a call structure, and wakeup waiting
4354 * processes. Some fields aren't changed: state & mode are not
4355 * touched (these must be set by the caller), and bufptr, nLeft, and
4356 * nFree are not reset, since these fields are manipulated by
4357 * unprotected macros, and may only be reset by non-interrupting code.
4360 /* this code requires that call->conn be set properly as a pre-condition. */
4361 #endif /* ADAPT_WINDOW */
4364 rxi_ResetCall(register struct rx_call *call, register int newcall)
4367 register struct rx_peer *peer;
4368 struct rx_packet *packet;
4370 /* Notify anyone who is waiting for asynchronous packet arrival */
4371 if (call->arrivalProc) {
4372 (*call->arrivalProc) (call, call->arrivalProcHandle,
4373 call->arrivalProcArg);
4374 call->arrivalProc = (void (*)())0;
4377 if (call->delayedAbortEvent) {
4378 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4379 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4381 rxi_SendCallAbort(call, packet, 0, 1);
4382 rxi_FreePacket(packet);
4387 * Update the peer with the congestion information in this call
4388 * so other calls on this connection can pick up where this call
4389 * left off. If the congestion sequence numbers don't match then
4390 * another call experienced a retransmission.
4392 peer = call->conn->peer;
4393 MUTEX_ENTER(&peer->peer_lock);
4395 if (call->congestSeq == peer->congestSeq) {
4396 peer->cwind = MAX(peer->cwind, call->cwind);
4397 peer->MTU = MAX(peer->MTU, call->MTU);
4398 peer->nDgramPackets =
4399 MAX(peer->nDgramPackets, call->nDgramPackets);
4402 call->abortCode = 0;
4403 call->abortCount = 0;
4405 if (peer->maxDgramPackets > 1) {
4406 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4408 call->MTU = peer->MTU;
4410 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4411 call->ssthresh = rx_maxSendWindow;
4412 call->nDgramPackets = peer->nDgramPackets;
4413 call->congestSeq = peer->congestSeq;
4414 MUTEX_EXIT(&peer->peer_lock);
4416 flags = call->flags;
4417 rxi_ClearReceiveQueue(call);
4418 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4419 if (call->flags & RX_CALL_TQ_BUSY) {
4420 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4421 call->flags |= (flags & RX_CALL_TQ_WAIT);
4423 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4425 rxi_ClearTransmitQueue(call, 0);
4426 queue_Init(&call->tq);
4429 queue_Init(&call->rq);
4431 call->rwind = rx_initReceiveWindow;
4432 call->twind = rx_initSendWindow;
4433 call->nSoftAcked = 0;
4434 call->nextCwind = 0;
4437 call->nCwindAcks = 0;
4438 call->nSoftAcks = 0;
4439 call->nHardAcks = 0;
4441 call->tfirst = call->rnext = call->tnext = 1;
4443 call->lastAcked = 0;
4444 call->localStatus = call->remoteStatus = 0;
4446 if (flags & RX_CALL_READER_WAIT) {
4447 #ifdef RX_ENABLE_LOCKS
4448 CV_BROADCAST(&call->cv_rq);
4450 osi_rxWakeup(&call->rq);
4453 if (flags & RX_CALL_WAIT_PACKETS) {
4454 MUTEX_ENTER(&rx_freePktQ_lock);
4455 rxi_PacketsUnWait(); /* XXX */
4456 MUTEX_EXIT(&rx_freePktQ_lock);
4458 #ifdef RX_ENABLE_LOCKS
4459 CV_SIGNAL(&call->cv_twind);
4461 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4462 osi_rxWakeup(&call->twind);
4465 #ifdef RX_ENABLE_LOCKS
4466 /* The following ensures that we don't mess with any queue while some
4467 * other thread might also be doing so. The call_queue_lock field is
4468 * is only modified under the call lock. If the call is in the process
4469 * of being removed from a queue, the call is not locked until the
4470 * the queue lock is dropped and only then is the call_queue_lock field
4471 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4472 * Note that any other routine which removes a call from a queue has to
4473 * obtain the queue lock before examing the queue and removing the call.
4475 if (call->call_queue_lock) {
4476 MUTEX_ENTER(call->call_queue_lock);
4477 if (queue_IsOnQueue(call)) {
4479 if (flags & RX_CALL_WAIT_PROC) {
4480 MUTEX_ENTER(&rx_stats_mutex);
4482 MUTEX_EXIT(&rx_stats_mutex);
4485 MUTEX_EXIT(call->call_queue_lock);
4486 CLEAR_CALL_QUEUE_LOCK(call);
4488 #else /* RX_ENABLE_LOCKS */
4489 if (queue_IsOnQueue(call)) {
4491 if (flags & RX_CALL_WAIT_PROC)
4494 #endif /* RX_ENABLE_LOCKS */
4496 rxi_KeepAliveOff(call);
4497 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4500 /* Send an acknowledge for the indicated packet (seq,serial) of the
4501 * indicated call, for the indicated reason (reason). This
4502 * acknowledge will specifically acknowledge receiving the packet, and
4503 * will also specify which other packets for this call have been
4504 * received. This routine returns the packet that was used to the
4505 * caller. The caller is responsible for freeing it or re-using it.
4506 * This acknowledgement also returns the highest sequence number
4507 * actually read out by the higher level to the sender; the sender
4508 * promises to keep around packets that have not been read by the
4509 * higher level yet (unless, of course, the sender decides to abort
4510 * the call altogether). Any of p, seq, serial, pflags, or reason may
4511 * be set to zero without ill effect. That is, if they are zero, they
4512 * will not convey any information.
4513 * NOW there is a trailer field, after the ack where it will safely be
4514 * ignored by mundanes, which indicates the maximum size packet this
4515 * host can swallow. */
4517 register struct rx_packet *optionalPacket; use to send ack (or null)
4518 int seq; Sequence number of the packet we are acking
4519 int serial; Serial number of the packet
4520 int pflags; Flags field from packet header
4521 int reason; Reason an acknowledge was prompted
4525 rxi_SendAck(register struct rx_call *call,
4526 register struct rx_packet *optionalPacket, int serial, int reason,
4529 struct rx_ackPacket *ap;
4530 register struct rx_packet *rqp;
4531 register struct rx_packet *nxp; /* For queue_Scan */
4532 register struct rx_packet *p;
4535 #ifdef RX_ENABLE_TSFPQ
4536 struct rx_ts_info_t * rx_ts_info;
4540 * Open the receive window once a thread starts reading packets
4542 if (call->rnext > 1) {
4543 call->rwind = rx_maxReceiveWindow;
4546 call->nHardAcks = 0;
4547 call->nSoftAcks = 0;
4548 if (call->rnext > call->lastAcked)
4549 call->lastAcked = call->rnext;
4553 rx_computelen(p, p->length); /* reset length, you never know */
4554 } /* where that's been... */
4555 #ifdef RX_ENABLE_TSFPQ
4557 RX_TS_INFO_GET(rx_ts_info);
4558 if ((p = rx_ts_info->local_special_packet)) {
4559 rx_computelen(p, p->length);
4560 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4561 rx_ts_info->local_special_packet = p;
4562 } else { /* We won't send the ack, but don't panic. */
4563 return optionalPacket;
4567 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4568 /* We won't send the ack, but don't panic. */
4569 return optionalPacket;
4574 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4577 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4578 #ifndef RX_ENABLE_TSFPQ
4579 if (!optionalPacket)
4582 return optionalPacket;
4584 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4585 if (rx_Contiguous(p) < templ) {
4586 #ifndef RX_ENABLE_TSFPQ
4587 if (!optionalPacket)
4590 return optionalPacket;
4595 /* MTUXXX failing to send an ack is very serious. We should */
4596 /* try as hard as possible to send even a partial ack; it's */
4597 /* better than nothing. */
4598 ap = (struct rx_ackPacket *)rx_DataOf(p);
4599 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4600 ap->reason = reason;
4602 /* The skew computation used to be bogus, I think it's better now. */
4603 /* We should start paying attention to skew. XXX */
4604 ap->serial = htonl(serial);
4605 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4607 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4608 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4610 /* No fear of running out of ack packet here because there can only be at most
4611 * one window full of unacknowledged packets. The window size must be constrained
4612 * to be less than the maximum ack size, of course. Also, an ack should always
4613 * fit into a single packet -- it should not ever be fragmented. */
4614 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4615 if (!rqp || !call->rq.next
4616 || (rqp->header.seq > (call->rnext + call->rwind))) {
4617 #ifndef RX_ENABLE_TSFPQ
4618 if (!optionalPacket)
4621 rxi_CallError(call, RX_CALL_DEAD);
4622 return optionalPacket;
4625 while (rqp->header.seq > call->rnext + offset)
4626 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4627 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4629 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4630 #ifndef RX_ENABLE_TSFPQ
4631 if (!optionalPacket)
4634 rxi_CallError(call, RX_CALL_DEAD);
4635 return optionalPacket;
4640 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4642 /* these are new for AFS 3.3 */
4643 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4644 templ = htonl(templ);
4645 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4646 templ = htonl(call->conn->peer->ifMTU);
4647 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4648 sizeof(afs_int32), &templ);
4650 /* new for AFS 3.4 */
4651 templ = htonl(call->rwind);
4652 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4653 sizeof(afs_int32), &templ);
4655 /* new for AFS 3.5 */
4656 templ = htonl(call->conn->peer->ifDgramPackets);
4657 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4658 sizeof(afs_int32), &templ);
4660 p->header.serviceId = call->conn->serviceId;
4661 p->header.cid = (call->conn->cid | call->channel);
4662 p->header.callNumber = *call->callNumber;
4664 p->header.securityIndex = call->conn->securityIndex;
4665 p->header.epoch = call->conn->epoch;
4666 p->header.type = RX_PACKET_TYPE_ACK;
4667 p->header.flags = RX_SLOW_START_OK;
4668 if (reason == RX_ACK_PING) {
4669 p->header.flags |= RX_REQUEST_ACK;
4671 clock_GetTime(&call->pingRequestTime);
4674 if (call->conn->type == RX_CLIENT_CONNECTION)
4675 p->header.flags |= RX_CLIENT_INITIATED;
4679 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u",
4680 ap->reason, ntohl(ap->previousPacket),
4681 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4683 for (offset = 0; offset < ap->nAcks; offset++)
4684 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4692 register int i, nbytes = p->length;
4694 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4695 if (nbytes <= p->wirevec[i].iov_len) {
4696 register int savelen, saven;
4698 savelen = p->wirevec[i].iov_len;
4700 p->wirevec[i].iov_len = nbytes;
4702 rxi_Send(call, p, istack);
4703 p->wirevec[i].iov_len = savelen;
4707 nbytes -= p->wirevec[i].iov_len;
4710 MUTEX_ENTER(&rx_stats_mutex);
4711 rx_stats.ackPacketsSent++;
4712 MUTEX_EXIT(&rx_stats_mutex);
4713 #ifndef RX_ENABLE_TSFPQ
4714 if (!optionalPacket)
4717 return optionalPacket; /* Return packet for re-use by caller */
4720 /* Send all of the packets in the list in single datagram */
4722 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4723 int istack, int moreFlag, struct clock *now,
4724 struct clock *retryTime, int resending)
4729 struct rx_connection *conn = call->conn;
4730 struct rx_peer *peer = conn->peer;
4732 MUTEX_ENTER(&peer->peer_lock);
4735 peer->reSends += len;
4736 MUTEX_ENTER(&rx_stats_mutex);
4737 rx_stats.dataPacketsSent += len;
4738 MUTEX_EXIT(&rx_stats_mutex);
4739 MUTEX_EXIT(&peer->peer_lock);
4741 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4745 /* Set the packet flags and schedule the resend events */
4746 /* Only request an ack for the last packet in the list */
4747 for (i = 0; i < len; i++) {
4748 list[i]->retryTime = *retryTime;
4749 if (list[i]->header.serial) {
4750 /* Exponentially backoff retry times */
4751 if (list[i]->backoff < MAXBACKOFF) {
4752 /* so it can't stay == 0 */
4753 list[i]->backoff = (list[i]->backoff << 1) + 1;
4756 clock_Addmsec(&(list[i]->retryTime),
4757 ((afs_uint32) list[i]->backoff) << 8);
4760 /* Wait a little extra for the ack on the last packet */
4761 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4762 clock_Addmsec(&(list[i]->retryTime), 400);
4765 /* Record the time sent */
4766 list[i]->timeSent = *now;
4768 /* Ask for an ack on retransmitted packets, on every other packet
4769 * if the peer doesn't support slow start. Ask for an ack on every
4770 * packet until the congestion window reaches the ack rate. */
4771 if (list[i]->header.serial) {
4773 MUTEX_ENTER(&rx_stats_mutex);
4774 rx_stats.dataPacketsReSent++;
4775 MUTEX_EXIT(&rx_stats_mutex);
4777 /* improved RTO calculation- not Karn */
4778 list[i]->firstSent = *now;
4779 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4780 || (!(call->flags & RX_CALL_SLOW_START_OK)
4781 && (list[i]->header.seq & 1)))) {
4786 MUTEX_ENTER(&peer->peer_lock);
4790 MUTEX_ENTER(&rx_stats_mutex);
4791 rx_stats.dataPacketsSent++;
4792 MUTEX_EXIT(&rx_stats_mutex);
4793 MUTEX_EXIT(&peer->peer_lock);
4795 /* Tag this packet as not being the last in this group,
4796 * for the receiver's benefit */
4797 if (i < len - 1 || moreFlag) {
4798 list[i]->header.flags |= RX_MORE_PACKETS;
4801 /* Install the new retransmit time for the packet, and
4802 * record the time sent */
4803 list[i]->timeSent = *now;
4807 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4810 /* Since we're about to send a data packet to the peer, it's
4811 * safe to nuke any scheduled end-of-packets ack */
4812 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4814 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4815 MUTEX_EXIT(&call->lock);
4817 rxi_SendPacketList(call, conn, list, len, istack);
4819 rxi_SendPacket(call, conn, list[0], istack);
4821 MUTEX_ENTER(&call->lock);
4822 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4824 /* Update last send time for this call (for keep-alive
4825 * processing), and for the connection (so that we can discover
4826 * idle connections) */
4827 conn->lastSendTime = call->lastSendTime = clock_Sec();
4830 /* When sending packets we need to follow these rules:
4831 * 1. Never send more than maxDgramPackets in a jumbogram.
4832 * 2. Never send a packet with more than two iovecs in a jumbogram.
4833 * 3. Never send a retransmitted packet in a jumbogram.
4834 * 4. Never send more than cwind/4 packets in a jumbogram
4835 * We always keep the last list we should have sent so we
4836 * can set the RX_MORE_PACKETS flags correctly.
4839 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4840 int istack, struct clock *now, struct clock *retryTime,
4843 int i, cnt, lastCnt = 0;
4844 struct rx_packet **listP, **lastP = 0;
4845 struct rx_peer *peer = call->conn->peer;
4846 int morePackets = 0;
4848 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4849 /* Does the current packet force us to flush the current list? */
4851 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4852 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4854 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4856 /* If the call enters an error state stop sending, or if
4857 * we entered congestion recovery mode, stop sending */
4858 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4866 /* Add the current packet to the list if it hasn't been acked.
4867 * Otherwise adjust the list pointer to skip the current packet. */
4868 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
4870 /* Do we need to flush the list? */
4871 if (cnt >= (int)peer->maxDgramPackets
4872 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
4873 || list[i]->header.serial
4874 || list[i]->length != RX_JUMBOBUFFERSIZE) {
4876 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
4877 retryTime, resending);
4878 /* If the call enters an error state stop sending, or if
4879 * we entered congestion recovery mode, stop sending */
4881 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4886 listP = &list[i + 1];
4891 osi_Panic("rxi_SendList error");
4893 listP = &list[i + 1];
4897 /* Send the whole list when the call is in receive mode, when
4898 * the call is in eof mode, when we are in fast recovery mode,
4899 * and when we have the last packet */
4900 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
4901 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
4902 || (call->flags & RX_CALL_FAST_RECOVER)) {
4903 /* Check for the case where the current list contains
4904 * an acked packet. Since we always send retransmissions
4905 * in a separate packet, we only need to check the first
4906 * packet in the list */
4907 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
4911 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
4912 retryTime, resending);
4913 /* If the call enters an error state stop sending, or if
4914 * we entered congestion recovery mode, stop sending */
4915 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4919 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
4922 } else if (lastCnt > 0) {
4923 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
4928 #ifdef RX_ENABLE_LOCKS
4929 /* Call rxi_Start, below, but with the call lock held. */
4931 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
4932 void *arg1, int istack)
4934 MUTEX_ENTER(&call->lock);
4935 rxi_Start(event, call, arg1, istack);
4936 MUTEX_EXIT(&call->lock);
4938 #endif /* RX_ENABLE_LOCKS */
4940 /* This routine is called when new packets are readied for
4941 * transmission and when retransmission may be necessary, or when the
4942 * transmission window or burst count are favourable. This should be
4943 * better optimized for new packets, the usual case, now that we've
4944 * got rid of queues of send packets. XXXXXXXXXXX */
4946 rxi_Start(struct rxevent *event, register struct rx_call *call,
4947 void *arg1, int istack)
4949 struct rx_packet *p;
4950 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
4951 struct rx_peer *peer = call->conn->peer;
4952 struct clock now, retryTime;
4956 struct rx_packet **xmitList;
4959 /* If rxi_Start is being called as a result of a resend event,
4960 * then make sure that the event pointer is removed from the call
4961 * structure, since there is no longer a per-call retransmission
4963 if (event && event == call->resendEvent) {
4964 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
4965 call->resendEvent = NULL;
4967 if (queue_IsEmpty(&call->tq)) {
4971 /* Timeouts trigger congestion recovery */
4972 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4973 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4974 /* someone else is waiting to start recovery */
4977 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4978 while (call->flags & RX_CALL_TQ_BUSY) {
4979 call->flags |= RX_CALL_TQ_WAIT;
4980 #ifdef RX_ENABLE_LOCKS
4981 CV_WAIT(&call->cv_tq, &call->lock);
4982 #else /* RX_ENABLE_LOCKS */
4983 osi_rxSleep(&call->tq);
4984 #endif /* RX_ENABLE_LOCKS */
4986 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4987 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4988 call->flags |= RX_CALL_FAST_RECOVER;
4989 if (peer->maxDgramPackets > 1) {
4990 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4992 call->MTU = MIN(peer->natMTU, peer->maxMTU);
4994 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4995 call->nDgramPackets = 1;
4997 call->nextCwind = 1;
5000 MUTEX_ENTER(&peer->peer_lock);
5001 peer->MTU = call->MTU;
5002 peer->cwind = call->cwind;
5003 peer->nDgramPackets = 1;
5005 call->congestSeq = peer->congestSeq;
5006 MUTEX_EXIT(&peer->peer_lock);
5007 /* Clear retry times on packets. Otherwise, it's possible for
5008 * some packets in the queue to force resends at rates faster
5009 * than recovery rates.
5011 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5012 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5013 clock_Zero(&p->retryTime);
5018 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5019 MUTEX_ENTER(&rx_stats_mutex);
5020 rx_tq_debug.rxi_start_in_error++;
5021 MUTEX_EXIT(&rx_stats_mutex);
5026 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5027 /* Get clock to compute the re-transmit time for any packets
5028 * in this burst. Note, if we back off, it's reasonable to
5029 * back off all of the packets in the same manner, even if
5030 * some of them have been retransmitted more times than more
5031 * recent additions */
5032 clock_GetTime(&now);
5033 retryTime = now; /* initialize before use */
5034 MUTEX_ENTER(&peer->peer_lock);
5035 clock_Add(&retryTime, &peer->timeout);
5036 MUTEX_EXIT(&peer->peer_lock);
5038 /* Send (or resend) any packets that need it, subject to
5039 * window restrictions and congestion burst control
5040 * restrictions. Ask for an ack on the last packet sent in
5041 * this burst. For now, we're relying upon the window being
5042 * considerably bigger than the largest number of packets that
5043 * are typically sent at once by one initial call to
5044 * rxi_Start. This is probably bogus (perhaps we should ask
5045 * for an ack when we're half way through the current
5046 * window?). Also, for non file transfer applications, this
5047 * may end up asking for an ack for every packet. Bogus. XXXX
5050 * But check whether we're here recursively, and let the other guy
5053 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5054 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5055 call->flags |= RX_CALL_TQ_BUSY;
5057 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5059 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5060 call->flags &= ~RX_CALL_NEED_START;
5061 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5063 maxXmitPackets = MIN(call->twind, call->cwind);
5064 xmitList = (struct rx_packet **)
5065 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5066 if (xmitList == NULL)
5067 osi_Panic("rxi_Start, failed to allocate xmit list");
5068 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5069 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5070 /* We shouldn't be sending packets if a thread is waiting
5071 * to initiate congestion recovery */
5075 && (call->flags & RX_CALL_FAST_RECOVER)) {
5076 /* Only send one packet during fast recovery */
5079 if ((p->flags & RX_PKTFLAG_FREE)
5080 || (!queue_IsEnd(&call->tq, nxp)
5081 && (nxp->flags & RX_PKTFLAG_FREE))
5082 || (p == (struct rx_packet *)&rx_freePacketQueue)
5083 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5084 osi_Panic("rxi_Start: xmit queue clobbered");
5086 if (p->flags & RX_PKTFLAG_ACKED) {
5087 MUTEX_ENTER(&rx_stats_mutex);
5088 rx_stats.ignoreAckedPacket++;
5089 MUTEX_EXIT(&rx_stats_mutex);
5090 continue; /* Ignore this packet if it has been acknowledged */
5093 /* Turn off all flags except these ones, which are the same
5094 * on each transmission */
5095 p->header.flags &= RX_PRESET_FLAGS;
5097 if (p->header.seq >=
5098 call->tfirst + MIN((int)call->twind,
5099 (int)(call->nSoftAcked +
5101 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5102 /* Note: if we're waiting for more window space, we can
5103 * still send retransmits; hence we don't return here, but
5104 * break out to schedule a retransmit event */
5105 dpf(("call %d waiting for window",
5106 *(call->callNumber)));
5110 /* Transmit the packet if it needs to be sent. */
5111 if (!clock_Lt(&now, &p->retryTime)) {
5112 if (nXmitPackets == maxXmitPackets) {
5113 rxi_SendXmitList(call, xmitList, nXmitPackets,
5114 istack, &now, &retryTime,
5116 osi_Free(xmitList, maxXmitPackets *
5117 sizeof(struct rx_packet *));
5120 xmitList[nXmitPackets++] = p;
5124 /* xmitList now hold pointers to all of the packets that are
5125 * ready to send. Now we loop to send the packets */
5126 if (nXmitPackets > 0) {
5127 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5128 &now, &retryTime, resending);
5131 maxXmitPackets * sizeof(struct rx_packet *));
5133 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5135 * TQ references no longer protected by this flag; they must remain
5136 * protected by the global lock.
5138 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5139 call->flags &= ~RX_CALL_TQ_BUSY;
5140 if (call->flags & RX_CALL_TQ_WAIT) {
5141 call->flags &= ~RX_CALL_TQ_WAIT;
5142 #ifdef RX_ENABLE_LOCKS
5143 CV_BROADCAST(&call->cv_tq);
5144 #else /* RX_ENABLE_LOCKS */
5145 osi_rxWakeup(&call->tq);
5146 #endif /* RX_ENABLE_LOCKS */
5151 /* We went into the error state while sending packets. Now is
5152 * the time to reset the call. This will also inform the using
5153 * process that the call is in an error state.
5155 MUTEX_ENTER(&rx_stats_mutex);
5156 rx_tq_debug.rxi_start_aborted++;
5157 MUTEX_EXIT(&rx_stats_mutex);
5158 call->flags &= ~RX_CALL_TQ_BUSY;
5159 if (call->flags & RX_CALL_TQ_WAIT) {
5160 call->flags &= ~RX_CALL_TQ_WAIT;
5161 #ifdef RX_ENABLE_LOCKS
5162 CV_BROADCAST(&call->cv_tq);
5163 #else /* RX_ENABLE_LOCKS */
5164 osi_rxWakeup(&call->tq);
5165 #endif /* RX_ENABLE_LOCKS */
5167 rxi_CallError(call, call->error);
5170 #ifdef RX_ENABLE_LOCKS
5171 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5172 register int missing;
5173 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5174 /* Some packets have received acks. If they all have, we can clear
5175 * the transmit queue.
5178 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5179 if (p->header.seq < call->tfirst
5180 && (p->flags & RX_PKTFLAG_ACKED)) {
5187 call->flags |= RX_CALL_TQ_CLEARME;
5189 #endif /* RX_ENABLE_LOCKS */
5190 /* Don't bother doing retransmits if the TQ is cleared. */
5191 if (call->flags & RX_CALL_TQ_CLEARME) {
5192 rxi_ClearTransmitQueue(call, 1);
5194 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5197 /* Always post a resend event, if there is anything in the
5198 * queue, and resend is possible. There should be at least
5199 * one unacknowledged packet in the queue ... otherwise none
5200 * of these packets should be on the queue in the first place.
5202 if (call->resendEvent) {
5203 /* Cancel the existing event and post a new one */
5204 rxevent_Cancel(call->resendEvent, call,
5205 RX_CALL_REFCOUNT_RESEND);
5208 /* The retry time is the retry time on the first unacknowledged
5209 * packet inside the current window */
5211 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5212 /* Don't set timers for packets outside the window */
5213 if (p->header.seq >= call->tfirst + call->twind) {
5217 if (!(p->flags & RX_PKTFLAG_ACKED)
5218 && !clock_IsZero(&p->retryTime)) {
5220 retryTime = p->retryTime;
5225 /* Post a new event to re-run rxi_Start when retries may be needed */
5226 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5227 #ifdef RX_ENABLE_LOCKS
5228 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5230 rxevent_Post2(&retryTime, rxi_StartUnlocked,
5231 (void *)call, 0, istack);
5232 #else /* RX_ENABLE_LOCKS */
5234 rxevent_Post2(&retryTime, rxi_Start, (void *)call,
5236 #endif /* RX_ENABLE_LOCKS */
5239 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5240 } while (call->flags & RX_CALL_NEED_START);
5242 * TQ references no longer protected by this flag; they must remain
5243 * protected by the global lock.
5245 call->flags &= ~RX_CALL_TQ_BUSY;
5246 if (call->flags & RX_CALL_TQ_WAIT) {
5247 call->flags &= ~RX_CALL_TQ_WAIT;
5248 #ifdef RX_ENABLE_LOCKS
5249 CV_BROADCAST(&call->cv_tq);
5250 #else /* RX_ENABLE_LOCKS */
5251 osi_rxWakeup(&call->tq);
5252 #endif /* RX_ENABLE_LOCKS */
5255 call->flags |= RX_CALL_NEED_START;
5257 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5259 if (call->resendEvent) {
5260 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5265 /* Also adjusts the keep alive parameters for the call, to reflect
5266 * that we have just sent a packet (so keep alives aren't sent
5269 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5272 register struct rx_connection *conn = call->conn;
5274 /* Stamp each packet with the user supplied status */
5275 p->header.userStatus = call->localStatus;
5277 /* Allow the security object controlling this call's security to
5278 * make any last-minute changes to the packet */
5279 RXS_SendPacket(conn->securityObject, call, p);
5281 /* Since we're about to send SOME sort of packet to the peer, it's
5282 * safe to nuke any scheduled end-of-packets ack */
5283 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5285 /* Actually send the packet, filling in more connection-specific fields */
5286 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5287 MUTEX_EXIT(&call->lock);
5288 rxi_SendPacket(call, conn, p, istack);
5289 MUTEX_ENTER(&call->lock);
5290 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5292 /* Update last send time for this call (for keep-alive
5293 * processing), and for the connection (so that we can discover
5294 * idle connections) */
5295 conn->lastSendTime = call->lastSendTime = clock_Sec();
5299 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5300 * that things are fine. Also called periodically to guarantee that nothing
5301 * falls through the cracks (e.g. (error + dally) connections have keepalive
5302 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5304 * haveCTLock Set if calling from rxi_ReapConnections
5306 #ifdef RX_ENABLE_LOCKS
5308 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5309 #else /* RX_ENABLE_LOCKS */
5311 rxi_CheckCall(register struct rx_call *call)
5312 #endif /* RX_ENABLE_LOCKS */
5314 register struct rx_connection *conn = call->conn;
5316 afs_uint32 deadTime;
5318 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5319 if (call->flags & RX_CALL_TQ_BUSY) {
5320 /* Call is active and will be reset by rxi_Start if it's
5321 * in an error state.
5326 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5328 (((afs_uint32) conn->secondsUntilDead << 10) +
5329 ((afs_uint32) conn->peer->rtt >> 3) +
5330 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5332 /* These are computed to the second (+- 1 second). But that's
5333 * good enough for these values, which should be a significant
5334 * number of seconds. */
5335 if (now > (call->lastReceiveTime + deadTime)) {
5336 if (call->state == RX_STATE_ACTIVE) {
5337 rxi_CallError(call, RX_CALL_DEAD);
5340 #ifdef RX_ENABLE_LOCKS
5341 /* Cancel pending events */
5342 rxevent_Cancel(call->delayedAckEvent, call,
5343 RX_CALL_REFCOUNT_DELAY);
5344 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5345 rxevent_Cancel(call->keepAliveEvent, call,
5346 RX_CALL_REFCOUNT_ALIVE);
5347 if (call->refCount == 0) {
5348 rxi_FreeCall(call, haveCTLock);
5352 #else /* RX_ENABLE_LOCKS */
5355 #endif /* RX_ENABLE_LOCKS */
5357 /* Non-active calls are destroyed if they are not responding
5358 * to pings; active calls are simply flagged in error, so the
5359 * attached process can die reasonably gracefully. */
5361 /* see if we have a non-activity timeout */
5362 if (call->startWait && conn->idleDeadTime
5363 && ((call->startWait + conn->idleDeadTime) < now)) {
5364 if (call->state == RX_STATE_ACTIVE) {
5365 rxi_CallError(call, RX_CALL_TIMEOUT);
5369 /* see if we have a hard timeout */
5370 if (conn->hardDeadTime
5371 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5372 if (call->state == RX_STATE_ACTIVE)
5373 rxi_CallError(call, RX_CALL_TIMEOUT);
5380 /* When a call is in progress, this routine is called occasionally to
5381 * make sure that some traffic has arrived (or been sent to) the peer.
5382 * If nothing has arrived in a reasonable amount of time, the call is
5383 * declared dead; if nothing has been sent for a while, we send a
5384 * keep-alive packet (if we're actually trying to keep the call alive)
5387 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5390 struct rx_connection *conn;
5393 MUTEX_ENTER(&call->lock);
5394 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5395 if (event == call->keepAliveEvent)
5396 call->keepAliveEvent = NULL;
5399 #ifdef RX_ENABLE_LOCKS
5400 if (rxi_CheckCall(call, 0)) {
5401 MUTEX_EXIT(&call->lock);
5404 #else /* RX_ENABLE_LOCKS */
5405 if (rxi_CheckCall(call))
5407 #endif /* RX_ENABLE_LOCKS */
5409 /* Don't try to keep alive dallying calls */
5410 if (call->state == RX_STATE_DALLY) {
5411 MUTEX_EXIT(&call->lock);
5416 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5417 /* Don't try to send keepalives if there is unacknowledged data */
5418 /* the rexmit code should be good enough, this little hack
5419 * doesn't quite work XXX */
5420 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5422 rxi_ScheduleKeepAliveEvent(call);
5423 MUTEX_EXIT(&call->lock);
5428 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5430 if (!call->keepAliveEvent) {
5432 clock_GetTime(&when);
5433 when.sec += call->conn->secondsUntilPing;
5434 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5435 call->keepAliveEvent =
5436 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5440 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5442 rxi_KeepAliveOn(register struct rx_call *call)
5444 /* Pretend last packet received was received now--i.e. if another
5445 * packet isn't received within the keep alive time, then the call
5446 * will die; Initialize last send time to the current time--even
5447 * if a packet hasn't been sent yet. This will guarantee that a
5448 * keep-alive is sent within the ping time */
5449 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5450 rxi_ScheduleKeepAliveEvent(call);
5453 /* This routine is called to send connection abort messages
5454 * that have been delayed to throttle looping clients. */
5456 rxi_SendDelayedConnAbort(struct rxevent *event,
5457 register struct rx_connection *conn, char *dummy)
5460 struct rx_packet *packet;
5462 MUTEX_ENTER(&conn->conn_data_lock);
5463 conn->delayedAbortEvent = NULL;
5464 error = htonl(conn->error);
5466 MUTEX_EXIT(&conn->conn_data_lock);
5467 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5470 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5471 RX_PACKET_TYPE_ABORT, (char *)&error,
5473 rxi_FreePacket(packet);
5477 /* This routine is called to send call abort messages
5478 * that have been delayed to throttle looping clients. */
5480 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5484 struct rx_packet *packet;
5486 MUTEX_ENTER(&call->lock);
5487 call->delayedAbortEvent = NULL;
5488 error = htonl(call->error);
5490 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5493 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5494 (char *)&error, sizeof(error), 0);
5495 rxi_FreePacket(packet);
5497 MUTEX_EXIT(&call->lock);
5500 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5501 * seconds) to ask the client to authenticate itself. The routine
5502 * issues a challenge to the client, which is obtained from the
5503 * security object associated with the connection */
5505 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5506 void *arg1, int tries)
5508 conn->challengeEvent = NULL;
5509 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5510 register struct rx_packet *packet;
5514 /* We've failed to authenticate for too long.
5515 * Reset any calls waiting for authentication;
5516 * they are all in RX_STATE_PRECALL.
5520 MUTEX_ENTER(&conn->conn_call_lock);
5521 for (i = 0; i < RX_MAXCALLS; i++) {
5522 struct rx_call *call = conn->call[i];
5524 MUTEX_ENTER(&call->lock);
5525 if (call->state == RX_STATE_PRECALL) {
5526 rxi_CallError(call, RX_CALL_DEAD);
5527 rxi_SendCallAbort(call, NULL, 0, 0);
5529 MUTEX_EXIT(&call->lock);
5532 MUTEX_EXIT(&conn->conn_call_lock);
5536 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5538 /* If there's no packet available, do this later. */
5539 RXS_GetChallenge(conn->securityObject, conn, packet);
5540 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5541 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5542 rxi_FreePacket(packet);
5544 clock_GetTime(&when);
5545 when.sec += RX_CHALLENGE_TIMEOUT;
5546 conn->challengeEvent =
5547 rxevent_Post2(&when, rxi_ChallengeEvent, conn, 0,
5552 /* Call this routine to start requesting the client to authenticate
5553 * itself. This will continue until authentication is established,
5554 * the call times out, or an invalid response is returned. The
5555 * security object associated with the connection is asked to create
5556 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5557 * defined earlier. */
5559 rxi_ChallengeOn(register struct rx_connection *conn)
5561 if (!conn->challengeEvent) {
5562 RXS_CreateChallenge(conn->securityObject, conn);
5563 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5568 /* Compute round trip time of the packet provided, in *rttp.
5571 /* rxi_ComputeRoundTripTime is called with peer locked. */
5572 /* sentp and/or peer may be null */
5574 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5575 register struct clock *sentp,
5576 register struct rx_peer *peer)
5578 struct clock thisRtt, *rttp = &thisRtt;
5580 register int rtt_timeout;
5582 clock_GetTime(rttp);
5584 if (clock_Lt(rttp, sentp)) {
5586 return; /* somebody set the clock back, don't count this time. */
5588 clock_Sub(rttp, sentp);
5589 MUTEX_ENTER(&rx_stats_mutex);
5590 if (clock_Lt(rttp, &rx_stats.minRtt))
5591 rx_stats.minRtt = *rttp;
5592 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5593 if (rttp->sec > 60) {
5594 MUTEX_EXIT(&rx_stats_mutex);
5595 return; /* somebody set the clock ahead */
5597 rx_stats.maxRtt = *rttp;
5599 clock_Add(&rx_stats.totalRtt, rttp);
5600 rx_stats.nRttSamples++;
5601 MUTEX_EXIT(&rx_stats_mutex);
5603 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5605 /* Apply VanJacobson round-trip estimations */
5610 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5611 * srtt is stored as fixed point with 3 bits after the binary
5612 * point (i.e., scaled by 8). The following magic is
5613 * equivalent to the smoothing algorithm in rfc793 with an
5614 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5615 * srtt*8 = srtt*8 + rtt - srtt
5616 * srtt = srtt + rtt/8 - srtt/8
5619 delta = MSEC(rttp) - (peer->rtt >> 3);
5623 * We accumulate a smoothed rtt variance (actually, a smoothed
5624 * mean difference), then set the retransmit timer to smoothed
5625 * rtt + 4 times the smoothed variance (was 2x in van's original
5626 * paper, but 4x works better for me, and apparently for him as
5628 * rttvar is stored as
5629 * fixed point with 2 bits after the binary point (scaled by
5630 * 4). The following is equivalent to rfc793 smoothing with
5631 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5632 * replaces rfc793's wired-in beta.
5633 * dev*4 = dev*4 + (|actual - expected| - dev)
5639 delta -= (peer->rtt_dev >> 2);
5640 peer->rtt_dev += delta;
5642 /* I don't have a stored RTT so I start with this value. Since I'm
5643 * probably just starting a call, and will be pushing more data down
5644 * this, I expect congestion to increase rapidly. So I fudge a
5645 * little, and I set deviance to half the rtt. In practice,
5646 * deviance tends to approach something a little less than
5647 * half the smoothed rtt. */
5648 peer->rtt = (MSEC(rttp) << 3) + 8;
5649 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5651 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5652 * the other of these connections is usually in a user process, and can
5653 * be switched and/or swapped out. So on fast, reliable networks, the
5654 * timeout would otherwise be too short.
5656 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5657 clock_Zero(&(peer->timeout));
5658 clock_Addmsec(&(peer->timeout), rtt_timeout);
5660 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)));
5664 /* Find all server connections that have not been active for a long time, and
5667 rxi_ReapConnections(void)
5670 clock_GetTime(&now);
5672 /* Find server connection structures that haven't been used for
5673 * greater than rx_idleConnectionTime */
5675 struct rx_connection **conn_ptr, **conn_end;
5676 int i, havecalls = 0;
5677 MUTEX_ENTER(&rx_connHashTable_lock);
5678 for (conn_ptr = &rx_connHashTable[0], conn_end =
5679 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5681 struct rx_connection *conn, *next;
5682 struct rx_call *call;
5686 for (conn = *conn_ptr; conn; conn = next) {
5687 /* XXX -- Shouldn't the connection be locked? */
5690 for (i = 0; i < RX_MAXCALLS; i++) {
5691 call = conn->call[i];
5694 MUTEX_ENTER(&call->lock);
5695 #ifdef RX_ENABLE_LOCKS
5696 result = rxi_CheckCall(call, 1);
5697 #else /* RX_ENABLE_LOCKS */
5698 result = rxi_CheckCall(call);
5699 #endif /* RX_ENABLE_LOCKS */
5700 MUTEX_EXIT(&call->lock);
5702 /* If CheckCall freed the call, it might
5703 * have destroyed the connection as well,
5704 * which screws up the linked lists.
5710 if (conn->type == RX_SERVER_CONNECTION) {
5711 /* This only actually destroys the connection if
5712 * there are no outstanding calls */
5713 MUTEX_ENTER(&conn->conn_data_lock);
5714 if (!havecalls && !conn->refCount
5715 && ((conn->lastSendTime + rx_idleConnectionTime) <
5717 conn->refCount++; /* it will be decr in rx_DestroyConn */
5718 MUTEX_EXIT(&conn->conn_data_lock);
5719 #ifdef RX_ENABLE_LOCKS
5720 rxi_DestroyConnectionNoLock(conn);
5721 #else /* RX_ENABLE_LOCKS */
5722 rxi_DestroyConnection(conn);
5723 #endif /* RX_ENABLE_LOCKS */
5725 #ifdef RX_ENABLE_LOCKS
5727 MUTEX_EXIT(&conn->conn_data_lock);
5729 #endif /* RX_ENABLE_LOCKS */
5733 #ifdef RX_ENABLE_LOCKS
5734 while (rx_connCleanup_list) {
5735 struct rx_connection *conn;
5736 conn = rx_connCleanup_list;
5737 rx_connCleanup_list = rx_connCleanup_list->next;
5738 MUTEX_EXIT(&rx_connHashTable_lock);
5739 rxi_CleanupConnection(conn);
5740 MUTEX_ENTER(&rx_connHashTable_lock);
5742 MUTEX_EXIT(&rx_connHashTable_lock);
5743 #endif /* RX_ENABLE_LOCKS */
5746 /* Find any peer structures that haven't been used (haven't had an
5747 * associated connection) for greater than rx_idlePeerTime */
5749 struct rx_peer **peer_ptr, **peer_end;
5751 MUTEX_ENTER(&rx_rpc_stats);
5752 MUTEX_ENTER(&rx_peerHashTable_lock);
5753 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5754 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5756 struct rx_peer *peer, *next, *prev;
5757 for (prev = peer = *peer_ptr; peer; peer = next) {
5759 code = MUTEX_TRYENTER(&peer->peer_lock);
5760 if ((code) && (peer->refCount == 0)
5761 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5762 rx_interface_stat_p rpc_stat, nrpc_stat;
5764 MUTEX_EXIT(&peer->peer_lock);
5765 MUTEX_DESTROY(&peer->peer_lock);
5767 (&peer->rpcStats, rpc_stat, nrpc_stat,
5768 rx_interface_stat)) {
5769 unsigned int num_funcs;
5772 queue_Remove(&rpc_stat->queue_header);
5773 queue_Remove(&rpc_stat->all_peers);
5774 num_funcs = rpc_stat->stats[0].func_total;
5776 sizeof(rx_interface_stat_t) +
5777 rpc_stat->stats[0].func_total *
5778 sizeof(rx_function_entry_v1_t);
5780 rxi_Free(rpc_stat, space);
5781 rxi_rpc_peer_stat_cnt -= num_funcs;
5784 MUTEX_ENTER(&rx_stats_mutex);
5785 rx_stats.nPeerStructs--;
5786 MUTEX_EXIT(&rx_stats_mutex);
5787 if (peer == *peer_ptr) {
5794 MUTEX_EXIT(&peer->peer_lock);
5800 MUTEX_EXIT(&rx_peerHashTable_lock);
5801 MUTEX_EXIT(&rx_rpc_stats);
5804 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5805 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5806 * GC, just below. Really, we shouldn't have to keep moving packets from
5807 * one place to another, but instead ought to always know if we can
5808 * afford to hold onto a packet in its particular use. */
5809 MUTEX_ENTER(&rx_freePktQ_lock);
5810 if (rx_waitingForPackets) {
5811 rx_waitingForPackets = 0;
5812 #ifdef RX_ENABLE_LOCKS
5813 CV_BROADCAST(&rx_waitingForPackets_cv);
5815 osi_rxWakeup(&rx_waitingForPackets);
5818 MUTEX_EXIT(&rx_freePktQ_lock);
5820 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5821 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5825 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5826 * rx.h is sort of strange this is better. This is called with a security
5827 * object before it is discarded. Each connection using a security object has
5828 * its own refcount to the object so it won't actually be freed until the last
5829 * connection is destroyed.
5831 * This is the only rxs module call. A hold could also be written but no one
5835 rxs_Release(struct rx_securityClass *aobj)
5837 return RXS_Close(aobj);
5841 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5842 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5843 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5844 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5846 /* Adjust our estimate of the transmission rate to this peer, given
5847 * that the packet p was just acked. We can adjust peer->timeout and
5848 * call->twind. Pragmatically, this is called
5849 * only with packets of maximal length.
5850 * Called with peer and call locked.
5854 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5855 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5857 afs_int32 xferSize, xferMs;
5858 register afs_int32 minTime;
5861 /* Count down packets */
5862 if (peer->rateFlag > 0)
5864 /* Do nothing until we're enabled */
5865 if (peer->rateFlag != 0)
5870 /* Count only when the ack seems legitimate */
5871 switch (ackReason) {
5872 case RX_ACK_REQUESTED:
5874 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
5878 case RX_ACK_PING_RESPONSE:
5879 if (p) /* want the response to ping-request, not data send */
5881 clock_GetTime(&newTO);
5882 if (clock_Gt(&newTO, &call->pingRequestTime)) {
5883 clock_Sub(&newTO, &call->pingRequestTime);
5884 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
5888 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
5895 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));
5897 /* Track only packets that are big enough. */
5898 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
5902 /* absorb RTT data (in milliseconds) for these big packets */
5903 if (peer->smRtt == 0) {
5904 peer->smRtt = xferMs;
5906 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
5911 if (peer->countDown) {
5915 peer->countDown = 10; /* recalculate only every so often */
5917 /* In practice, we can measure only the RTT for full packets,
5918 * because of the way Rx acks the data that it receives. (If it's
5919 * smaller than a full packet, it often gets implicitly acked
5920 * either by the call response (from a server) or by the next call
5921 * (from a client), and either case confuses transmission times
5922 * with processing times.) Therefore, replace the above
5923 * more-sophisticated processing with a simpler version, where the
5924 * smoothed RTT is kept for full-size packets, and the time to
5925 * transmit a windowful of full-size packets is simply RTT *
5926 * windowSize. Again, we take two steps:
5927 - ensure the timeout is large enough for a single packet's RTT;
5928 - ensure that the window is small enough to fit in the desired timeout.*/
5930 /* First, the timeout check. */
5931 minTime = peer->smRtt;
5932 /* Get a reasonable estimate for a timeout period */
5934 newTO.sec = minTime / 1000;
5935 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
5937 /* Increase the timeout period so that we can always do at least
5938 * one packet exchange */
5939 if (clock_Gt(&newTO, &peer->timeout)) {
5941 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));
5943 peer->timeout = newTO;
5946 /* Now, get an estimate for the transmit window size. */
5947 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
5948 /* Now, convert to the number of full packets that could fit in a
5949 * reasonable fraction of that interval */
5950 minTime /= (peer->smRtt << 1);
5951 xferSize = minTime; /* (make a copy) */
5953 /* Now clamp the size to reasonable bounds. */
5956 else if (minTime > rx_Window)
5957 minTime = rx_Window;
5958 /* if (minTime != peer->maxWindow) {
5959 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
5960 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
5961 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
5963 peer->maxWindow = minTime;
5964 elide... call->twind = minTime;
5968 /* Cut back on the peer timeout if it had earlier grown unreasonably.
5969 * Discern this by calculating the timeout necessary for rx_Window
5971 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
5972 /* calculate estimate for transmission interval in milliseconds */
5973 minTime = rx_Window * peer->smRtt;
5974 if (minTime < 1000) {
5975 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
5976 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
5977 peer->timeout.usec, peer->smRtt, peer->packetSize));
5979 newTO.sec = 0; /* cut back on timeout by half a second */
5980 newTO.usec = 500000;
5981 clock_Sub(&peer->timeout, &newTO);
5986 } /* end of rxi_ComputeRate */
5987 #endif /* ADAPT_WINDOW */
5995 /* Don't call this debugging routine directly; use dpf */
5997 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
5998 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
6002 clock_GetTime(&now);
6003 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6004 (unsigned int)now.usec / 1000);
6005 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6013 * This function is used to process the rx_stats structure that is local
6014 * to a process as well as an rx_stats structure received from a remote
6015 * process (via rxdebug). Therefore, it needs to do minimal version
6019 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6020 afs_int32 freePackets, char version)
6024 if (size != sizeof(struct rx_stats)) {
6026 "Unexpected size of stats structure: was %d, expected %d\n",
6027 size, sizeof(struct rx_stats));
6030 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6033 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6034 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6035 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6036 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6037 s->specialPktAllocFailures);
6039 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6040 s->receivePktAllocFailures, s->sendPktAllocFailures,
6041 s->specialPktAllocFailures);
6045 " greedy %d, " "bogusReads %d (last from host %x), "
6046 "noPackets %d, " "noBuffers %d, " "selects %d, "
6047 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6048 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6049 s->selects, s->sendSelects);
6051 fprintf(file, " packets read: ");
6052 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6053 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6055 fprintf(file, "\n");
6058 " other read counters: data %d, " "ack %d, " "dup %d "
6059 "spurious %d " "dally %d\n", s->dataPacketsRead,
6060 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6061 s->ignorePacketDally);
6063 fprintf(file, " packets sent: ");
6064 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6065 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6067 fprintf(file, "\n");
6070 " other send counters: ack %d, " "data %d (not resends), "
6071 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6072 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6073 s->dataPacketsPushed, s->ignoreAckedPacket);
6076 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6077 s->netSendFailures, (int)s->fatalErrors);
6079 if (s->nRttSamples) {
6080 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6081 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6083 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6084 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6088 " %d server connections, " "%d client connections, "
6089 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6090 s->nServerConns, s->nClientConns, s->nPeerStructs,
6091 s->nCallStructs, s->nFreeCallStructs);
6093 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6094 fprintf(file, " %d clock updates\n", clock_nUpdates);
6099 /* for backward compatibility */
6101 rx_PrintStats(FILE * file)
6103 MUTEX_ENTER(&rx_stats_mutex);
6104 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6106 MUTEX_EXIT(&rx_stats_mutex);
6110 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6112 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6113 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6114 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6117 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6118 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6119 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6122 " Packet size %d, " "max in packet skew %d, "
6123 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6124 (int)peer->outPacketSkew);
6127 #ifdef AFS_PTHREAD_ENV
6129 * This mutex protects the following static variables:
6133 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6134 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6136 #define LOCK_RX_DEBUG
6137 #define UNLOCK_RX_DEBUG
6138 #endif /* AFS_PTHREAD_ENV */
6141 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6142 u_char type, void *inputData, size_t inputLength,
6143 void *outputData, size_t outputLength)
6145 static afs_int32 counter = 100;
6147 struct rx_header theader;
6149 register afs_int32 code;
6151 struct sockaddr_in taddr, faddr;
6156 endTime = time(0) + 20; /* try for 20 seconds */
6160 tp = &tbuffer[sizeof(struct rx_header)];
6161 taddr.sin_family = AF_INET;
6162 taddr.sin_port = remotePort;
6163 taddr.sin_addr.s_addr = remoteAddr;
6164 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6165 taddr.sin_len = sizeof(struct sockaddr_in);
6168 memset(&theader, 0, sizeof(theader));
6169 theader.epoch = htonl(999);
6171 theader.callNumber = htonl(counter);
6174 theader.type = type;
6175 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6176 theader.serviceId = 0;
6178 memcpy(tbuffer, &theader, sizeof(theader));
6179 memcpy(tp, inputData, inputLength);
6181 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6182 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6184 /* see if there's a packet available */
6186 FD_SET(socket, &imask);
6189 code = select(socket + 1, &imask, 0, 0, &tv);
6190 if (code == 1 && FD_ISSET(socket, &imask)) {
6191 /* now receive a packet */
6192 faddrLen = sizeof(struct sockaddr_in);
6194 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6195 (struct sockaddr *)&faddr, &faddrLen);
6198 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6199 if (counter == ntohl(theader.callNumber))
6204 /* see if we've timed out */
6205 if (endTime < time(0))
6208 code -= sizeof(struct rx_header);
6209 if (code > outputLength)
6210 code = outputLength;
6211 memcpy(outputData, tp, code);
6216 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6217 afs_uint16 remotePort, struct rx_debugStats * stat,
6218 afs_uint32 * supportedValues)
6220 struct rx_debugIn in;
6223 *supportedValues = 0;
6224 in.type = htonl(RX_DEBUGI_GETSTATS);
6227 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6228 &in, sizeof(in), stat, sizeof(*stat));
6231 * If the call was successful, fixup the version and indicate
6232 * what contents of the stat structure are valid.
6233 * Also do net to host conversion of fields here.
6237 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6238 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6240 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6241 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6243 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6244 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6246 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6247 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6249 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6250 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6252 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6253 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6255 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6256 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6258 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6259 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6262 stat->nFreePackets = ntohl(stat->nFreePackets);
6263 stat->packetReclaims = ntohl(stat->packetReclaims);
6264 stat->callsExecuted = ntohl(stat->callsExecuted);
6265 stat->nWaiting = ntohl(stat->nWaiting);
6266 stat->idleThreads = ntohl(stat->idleThreads);
6273 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6274 afs_uint16 remotePort, struct rx_stats * stat,
6275 afs_uint32 * supportedValues)
6277 struct rx_debugIn in;
6278 afs_int32 *lp = (afs_int32 *) stat;
6283 * supportedValues is currently unused, but added to allow future
6284 * versioning of this function.
6287 *supportedValues = 0;
6288 in.type = htonl(RX_DEBUGI_RXSTATS);
6290 memset(stat, 0, sizeof(*stat));
6292 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6293 &in, sizeof(in), stat, sizeof(*stat));
6298 * Do net to host conversion here
6301 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6310 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6311 afs_uint16 remotePort, size_t version_length,
6315 return MakeDebugCall(socket, remoteAddr, remotePort,
6316 RX_PACKET_TYPE_VERSION, a, 1, version,
6321 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6322 afs_uint16 remotePort, afs_int32 * nextConnection,
6323 int allConnections, afs_uint32 debugSupportedValues,
6324 struct rx_debugConn * conn,
6325 afs_uint32 * supportedValues)
6327 struct rx_debugIn in;
6332 * supportedValues is currently unused, but added to allow future
6333 * versioning of this function.
6336 *supportedValues = 0;
6337 if (allConnections) {
6338 in.type = htonl(RX_DEBUGI_GETALLCONN);
6340 in.type = htonl(RX_DEBUGI_GETCONN);
6342 in.index = htonl(*nextConnection);
6343 memset(conn, 0, sizeof(*conn));
6345 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6346 &in, sizeof(in), conn, sizeof(*conn));
6349 *nextConnection += 1;
6352 * Convert old connection format to new structure.
6355 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6356 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6357 #define MOVEvL(a) (conn->a = vL->a)
6359 /* any old or unrecognized version... */
6360 for (i = 0; i < RX_MAXCALLS; i++) {
6361 MOVEvL(callState[i]);
6362 MOVEvL(callMode[i]);
6363 MOVEvL(callFlags[i]);
6364 MOVEvL(callOther[i]);
6366 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6367 MOVEvL(secStats.type);
6368 MOVEvL(secStats.level);
6369 MOVEvL(secStats.flags);
6370 MOVEvL(secStats.expires);
6371 MOVEvL(secStats.packetsReceived);
6372 MOVEvL(secStats.packetsSent);
6373 MOVEvL(secStats.bytesReceived);
6374 MOVEvL(secStats.bytesSent);
6379 * Do net to host conversion here
6381 * I don't convert host or port since we are most likely
6382 * going to want these in NBO.
6384 conn->cid = ntohl(conn->cid);
6385 conn->serial = ntohl(conn->serial);
6386 for (i = 0; i < RX_MAXCALLS; i++) {
6387 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6389 conn->error = ntohl(conn->error);
6390 conn->secStats.flags = ntohl(conn->secStats.flags);
6391 conn->secStats.expires = ntohl(conn->secStats.expires);
6392 conn->secStats.packetsReceived =
6393 ntohl(conn->secStats.packetsReceived);
6394 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6395 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6396 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6397 conn->epoch = ntohl(conn->epoch);
6398 conn->natMTU = ntohl(conn->natMTU);
6405 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6406 afs_uint16 remotePort, afs_int32 * nextPeer,
6407 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6408 afs_uint32 * supportedValues)
6410 struct rx_debugIn in;
6414 * supportedValues is currently unused, but added to allow future
6415 * versioning of this function.
6418 *supportedValues = 0;
6419 in.type = htonl(RX_DEBUGI_GETPEER);
6420 in.index = htonl(*nextPeer);
6421 memset(peer, 0, sizeof(*peer));
6423 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6424 &in, sizeof(in), peer, sizeof(*peer));
6430 * Do net to host conversion here
6432 * I don't convert host or port since we are most likely
6433 * going to want these in NBO.
6435 peer->ifMTU = ntohs(peer->ifMTU);
6436 peer->idleWhen = ntohl(peer->idleWhen);
6437 peer->refCount = ntohs(peer->refCount);
6438 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6439 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6440 peer->rtt = ntohl(peer->rtt);
6441 peer->rtt_dev = ntohl(peer->rtt_dev);
6442 peer->timeout.sec = ntohl(peer->timeout.sec);
6443 peer->timeout.usec = ntohl(peer->timeout.usec);
6444 peer->nSent = ntohl(peer->nSent);
6445 peer->reSends = ntohl(peer->reSends);
6446 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6447 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6448 peer->rateFlag = ntohl(peer->rateFlag);
6449 peer->natMTU = ntohs(peer->natMTU);
6450 peer->maxMTU = ntohs(peer->maxMTU);
6451 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6452 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6453 peer->MTU = ntohs(peer->MTU);
6454 peer->cwind = ntohs(peer->cwind);
6455 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6456 peer->congestSeq = ntohs(peer->congestSeq);
6457 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6458 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6459 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6460 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6465 #endif /* RXDEBUG */
6470 struct rx_serverQueueEntry *np;
6473 register struct rx_call *call;
6474 register struct rx_serverQueueEntry *sq;
6478 if (rxinit_status == 1) {
6480 return; /* Already shutdown. */
6484 #ifndef AFS_PTHREAD_ENV
6485 FD_ZERO(&rx_selectMask);
6486 #endif /* AFS_PTHREAD_ENV */
6487 rxi_dataQuota = RX_MAX_QUOTA;
6488 #ifndef AFS_PTHREAD_ENV
6490 #endif /* AFS_PTHREAD_ENV */
6493 #ifndef AFS_PTHREAD_ENV
6494 #ifndef AFS_USE_GETTIMEOFDAY
6496 #endif /* AFS_USE_GETTIMEOFDAY */
6497 #endif /* AFS_PTHREAD_ENV */
6499 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6500 call = queue_First(&rx_freeCallQueue, rx_call);
6502 rxi_Free(call, sizeof(struct rx_call));
6505 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6506 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6512 struct rx_peer **peer_ptr, **peer_end;
6513 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6514 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6516 struct rx_peer *peer, *next;
6517 for (peer = *peer_ptr; peer; peer = next) {
6518 rx_interface_stat_p rpc_stat, nrpc_stat;
6521 (&peer->rpcStats, rpc_stat, nrpc_stat,
6522 rx_interface_stat)) {
6523 unsigned int num_funcs;
6526 queue_Remove(&rpc_stat->queue_header);
6527 queue_Remove(&rpc_stat->all_peers);
6528 num_funcs = rpc_stat->stats[0].func_total;
6530 sizeof(rx_interface_stat_t) +
6531 rpc_stat->stats[0].func_total *
6532 sizeof(rx_function_entry_v1_t);
6534 rxi_Free(rpc_stat, space);
6535 MUTEX_ENTER(&rx_rpc_stats);
6536 rxi_rpc_peer_stat_cnt -= num_funcs;
6537 MUTEX_EXIT(&rx_rpc_stats);
6541 MUTEX_ENTER(&rx_stats_mutex);
6542 rx_stats.nPeerStructs--;
6543 MUTEX_EXIT(&rx_stats_mutex);
6547 for (i = 0; i < RX_MAX_SERVICES; i++) {
6549 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6551 for (i = 0; i < rx_hashTableSize; i++) {
6552 register struct rx_connection *tc, *ntc;
6553 MUTEX_ENTER(&rx_connHashTable_lock);
6554 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6556 for (j = 0; j < RX_MAXCALLS; j++) {
6558 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6561 rxi_Free(tc, sizeof(*tc));
6563 MUTEX_EXIT(&rx_connHashTable_lock);
6566 MUTEX_ENTER(&freeSQEList_lock);
6568 while ((np = rx_FreeSQEList)) {
6569 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6570 MUTEX_DESTROY(&np->lock);
6571 rxi_Free(np, sizeof(*np));
6574 MUTEX_EXIT(&freeSQEList_lock);
6575 MUTEX_DESTROY(&freeSQEList_lock);
6576 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6577 MUTEX_DESTROY(&rx_connHashTable_lock);
6578 MUTEX_DESTROY(&rx_peerHashTable_lock);
6579 MUTEX_DESTROY(&rx_serverPool_lock);
6581 osi_Free(rx_connHashTable,
6582 rx_hashTableSize * sizeof(struct rx_connection *));
6583 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6585 UNPIN(rx_connHashTable,
6586 rx_hashTableSize * sizeof(struct rx_connection *));
6587 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6589 rxi_FreeAllPackets();
6591 MUTEX_ENTER(&rx_stats_mutex);
6592 rxi_dataQuota = RX_MAX_QUOTA;
6593 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6594 MUTEX_EXIT(&rx_stats_mutex);
6600 #ifdef RX_ENABLE_LOCKS
6602 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6604 if (!MUTEX_ISMINE(lockaddr))
6605 osi_Panic("Lock not held: %s", msg);
6607 #endif /* RX_ENABLE_LOCKS */
6612 * Routines to implement connection specific data.
6616 rx_KeyCreate(rx_destructor_t rtn)
6619 MUTEX_ENTER(&rxi_keyCreate_lock);
6620 key = rxi_keyCreate_counter++;
6621 rxi_keyCreate_destructor = (rx_destructor_t *)
6622 realloc((void *)rxi_keyCreate_destructor,
6623 (key + 1) * sizeof(rx_destructor_t));
6624 rxi_keyCreate_destructor[key] = rtn;
6625 MUTEX_EXIT(&rxi_keyCreate_lock);
6630 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6633 MUTEX_ENTER(&conn->conn_data_lock);
6634 if (!conn->specific) {
6635 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6636 for (i = 0; i < key; i++)
6637 conn->specific[i] = NULL;
6638 conn->nSpecific = key + 1;
6639 conn->specific[key] = ptr;
6640 } else if (key >= conn->nSpecific) {
6641 conn->specific = (void **)
6642 realloc(conn->specific, (key + 1) * sizeof(void *));
6643 for (i = conn->nSpecific; i < key; i++)
6644 conn->specific[i] = NULL;
6645 conn->nSpecific = key + 1;
6646 conn->specific[key] = ptr;
6648 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6649 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6650 conn->specific[key] = ptr;
6652 MUTEX_EXIT(&conn->conn_data_lock);
6656 rx_GetSpecific(struct rx_connection *conn, int key)
6659 MUTEX_ENTER(&conn->conn_data_lock);
6660 if (key >= conn->nSpecific)
6663 ptr = conn->specific[key];
6664 MUTEX_EXIT(&conn->conn_data_lock);
6668 #endif /* !KERNEL */
6671 * processStats is a queue used to store the statistics for the local
6672 * process. Its contents are similar to the contents of the rpcStats
6673 * queue on a rx_peer structure, but the actual data stored within
6674 * this queue contains totals across the lifetime of the process (assuming
6675 * the stats have not been reset) - unlike the per peer structures
6676 * which can come and go based upon the peer lifetime.
6679 static struct rx_queue processStats = { &processStats, &processStats };
6682 * peerStats is a queue used to store the statistics for all peer structs.
6683 * Its contents are the union of all the peer rpcStats queues.
6686 static struct rx_queue peerStats = { &peerStats, &peerStats };
6689 * rxi_monitor_processStats is used to turn process wide stat collection
6693 static int rxi_monitor_processStats = 0;
6696 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6699 static int rxi_monitor_peerStats = 0;
6702 * rxi_AddRpcStat - given all of the information for a particular rpc
6703 * call, create (if needed) and update the stat totals for the rpc.
6707 * IN stats - the queue of stats that will be updated with the new value
6709 * IN rxInterface - a unique number that identifies the rpc interface
6711 * IN currentFunc - the index of the function being invoked
6713 * IN totalFunc - the total number of functions in this interface
6715 * IN queueTime - the amount of time this function waited for a thread
6717 * IN execTime - the amount of time this function invocation took to execute
6719 * IN bytesSent - the number bytes sent by this invocation
6721 * IN bytesRcvd - the number bytes received by this invocation
6723 * IN isServer - if true, this invocation was made to a server
6725 * IN remoteHost - the ip address of the remote host
6727 * IN remotePort - the port of the remote host
6729 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6731 * INOUT counter - if a new stats structure is allocated, the counter will
6732 * be updated with the new number of allocated stat structures
6740 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6741 afs_uint32 currentFunc, afs_uint32 totalFunc,
6742 struct clock *queueTime, struct clock *execTime,
6743 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6744 afs_uint32 remoteHost, afs_uint32 remotePort,
6745 int addToPeerList, unsigned int *counter)
6748 rx_interface_stat_p rpc_stat, nrpc_stat;
6751 * See if there's already a structure for this interface
6754 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6755 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6756 && (rpc_stat->stats[0].remote_is_server == isServer))
6761 * Didn't find a match so allocate a new structure and add it to the
6765 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6766 || (rpc_stat->stats[0].interfaceId != rxInterface)
6767 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6772 sizeof(rx_interface_stat_t) +
6773 totalFunc * sizeof(rx_function_entry_v1_t);
6775 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6776 if (rpc_stat == NULL) {
6780 *counter += totalFunc;
6781 for (i = 0; i < totalFunc; i++) {
6782 rpc_stat->stats[i].remote_peer = remoteHost;
6783 rpc_stat->stats[i].remote_port = remotePort;
6784 rpc_stat->stats[i].remote_is_server = isServer;
6785 rpc_stat->stats[i].interfaceId = rxInterface;
6786 rpc_stat->stats[i].func_total = totalFunc;
6787 rpc_stat->stats[i].func_index = i;
6788 hzero(rpc_stat->stats[i].invocations);
6789 hzero(rpc_stat->stats[i].bytes_sent);
6790 hzero(rpc_stat->stats[i].bytes_rcvd);
6791 rpc_stat->stats[i].queue_time_sum.sec = 0;
6792 rpc_stat->stats[i].queue_time_sum.usec = 0;
6793 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6794 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6795 rpc_stat->stats[i].queue_time_min.sec = 9999999;
6796 rpc_stat->stats[i].queue_time_min.usec = 9999999;
6797 rpc_stat->stats[i].queue_time_max.sec = 0;
6798 rpc_stat->stats[i].queue_time_max.usec = 0;
6799 rpc_stat->stats[i].execution_time_sum.sec = 0;
6800 rpc_stat->stats[i].execution_time_sum.usec = 0;
6801 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
6802 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
6803 rpc_stat->stats[i].execution_time_min.sec = 9999999;
6804 rpc_stat->stats[i].execution_time_min.usec = 9999999;
6805 rpc_stat->stats[i].execution_time_max.sec = 0;
6806 rpc_stat->stats[i].execution_time_max.usec = 0;
6808 queue_Prepend(stats, rpc_stat);
6809 if (addToPeerList) {
6810 queue_Prepend(&peerStats, &rpc_stat->all_peers);
6815 * Increment the stats for this function
6818 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
6819 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
6820 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
6821 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
6822 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
6823 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
6824 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
6826 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
6827 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
6829 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
6830 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
6832 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
6833 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
6835 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
6836 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
6844 * rx_IncrementTimeAndCount - increment the times and count for a particular
6849 * IN peer - the peer who invoked the rpc
6851 * IN rxInterface - a unique number that identifies the rpc interface
6853 * IN currentFunc - the index of the function being invoked
6855 * IN totalFunc - the total number of functions in this interface
6857 * IN queueTime - the amount of time this function waited for a thread
6859 * IN execTime - the amount of time this function invocation took to execute
6861 * IN bytesSent - the number bytes sent by this invocation
6863 * IN bytesRcvd - the number bytes received by this invocation
6865 * IN isServer - if true, this invocation was made to a server
6873 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
6874 afs_uint32 currentFunc, afs_uint32 totalFunc,
6875 struct clock *queueTime, struct clock *execTime,
6876 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
6880 MUTEX_ENTER(&rx_rpc_stats);
6881 MUTEX_ENTER(&peer->peer_lock);
6883 if (rxi_monitor_peerStats) {
6884 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
6885 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6886 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
6889 if (rxi_monitor_processStats) {
6890 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
6891 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6892 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
6895 MUTEX_EXIT(&peer->peer_lock);
6896 MUTEX_EXIT(&rx_rpc_stats);
6901 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
6905 * IN callerVersion - the rpc stat version of the caller.
6907 * IN count - the number of entries to marshall.
6909 * IN stats - pointer to stats to be marshalled.
6911 * OUT ptr - Where to store the marshalled data.
6918 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
6919 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
6925 * We only support the first version
6927 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
6928 *(ptr++) = stats->remote_peer;
6929 *(ptr++) = stats->remote_port;
6930 *(ptr++) = stats->remote_is_server;
6931 *(ptr++) = stats->interfaceId;
6932 *(ptr++) = stats->func_total;
6933 *(ptr++) = stats->func_index;
6934 *(ptr++) = hgethi(stats->invocations);
6935 *(ptr++) = hgetlo(stats->invocations);
6936 *(ptr++) = hgethi(stats->bytes_sent);
6937 *(ptr++) = hgetlo(stats->bytes_sent);
6938 *(ptr++) = hgethi(stats->bytes_rcvd);
6939 *(ptr++) = hgetlo(stats->bytes_rcvd);
6940 *(ptr++) = stats->queue_time_sum.sec;
6941 *(ptr++) = stats->queue_time_sum.usec;
6942 *(ptr++) = stats->queue_time_sum_sqr.sec;
6943 *(ptr++) = stats->queue_time_sum_sqr.usec;
6944 *(ptr++) = stats->queue_time_min.sec;
6945 *(ptr++) = stats->queue_time_min.usec;
6946 *(ptr++) = stats->queue_time_max.sec;
6947 *(ptr++) = stats->queue_time_max.usec;
6948 *(ptr++) = stats->execution_time_sum.sec;
6949 *(ptr++) = stats->execution_time_sum.usec;
6950 *(ptr++) = stats->execution_time_sum_sqr.sec;
6951 *(ptr++) = stats->execution_time_sum_sqr.usec;
6952 *(ptr++) = stats->execution_time_min.sec;
6953 *(ptr++) = stats->execution_time_min.usec;
6954 *(ptr++) = stats->execution_time_max.sec;
6955 *(ptr++) = stats->execution_time_max.usec;
6961 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
6966 * IN callerVersion - the rpc stat version of the caller
6968 * OUT myVersion - the rpc stat version of this function
6970 * OUT clock_sec - local time seconds
6972 * OUT clock_usec - local time microseconds
6974 * OUT allocSize - the number of bytes allocated to contain stats
6976 * OUT statCount - the number stats retrieved from this process.
6978 * OUT stats - the actual stats retrieved from this process.
6982 * Returns void. If successful, stats will != NULL.
6986 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
6987 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
6988 size_t * allocSize, afs_uint32 * statCount,
6989 afs_uint32 ** stats)
6999 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7002 * Check to see if stats are enabled
7005 MUTEX_ENTER(&rx_rpc_stats);
7006 if (!rxi_monitor_processStats) {
7007 MUTEX_EXIT(&rx_rpc_stats);
7011 clock_GetTime(&now);
7012 *clock_sec = now.sec;
7013 *clock_usec = now.usec;
7016 * Allocate the space based upon the caller version
7018 * If the client is at an older version than we are,
7019 * we return the statistic data in the older data format, but
7020 * we still return our version number so the client knows we
7021 * are maintaining more data than it can retrieve.
7024 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7025 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7026 *statCount = rxi_rpc_process_stat_cnt;
7029 * This can't happen yet, but in the future version changes
7030 * can be handled by adding additional code here
7034 if (space > (size_t) 0) {
7036 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7039 rx_interface_stat_p rpc_stat, nrpc_stat;
7043 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7045 * Copy the data based upon the caller version
7047 rx_MarshallProcessRPCStats(callerVersion,
7048 rpc_stat->stats[0].func_total,
7049 rpc_stat->stats, &ptr);
7055 MUTEX_EXIT(&rx_rpc_stats);
7060 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7064 * IN callerVersion - the rpc stat version of the caller
7066 * OUT myVersion - the rpc stat version of this function
7068 * OUT clock_sec - local time seconds
7070 * OUT clock_usec - local time microseconds
7072 * OUT allocSize - the number of bytes allocated to contain stats
7074 * OUT statCount - the number of stats retrieved from the individual
7077 * OUT stats - the actual stats retrieved from the individual peer structures.
7081 * Returns void. If successful, stats will != NULL.
7085 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7086 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7087 size_t * allocSize, afs_uint32 * statCount,
7088 afs_uint32 ** stats)
7098 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7101 * Check to see if stats are enabled
7104 MUTEX_ENTER(&rx_rpc_stats);
7105 if (!rxi_monitor_peerStats) {
7106 MUTEX_EXIT(&rx_rpc_stats);
7110 clock_GetTime(&now);
7111 *clock_sec = now.sec;
7112 *clock_usec = now.usec;
7115 * Allocate the space based upon the caller version
7117 * If the client is at an older version than we are,
7118 * we return the statistic data in the older data format, but
7119 * we still return our version number so the client knows we
7120 * are maintaining more data than it can retrieve.
7123 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7124 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7125 *statCount = rxi_rpc_peer_stat_cnt;
7128 * This can't happen yet, but in the future version changes
7129 * can be handled by adding additional code here
7133 if (space > (size_t) 0) {
7135 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7138 rx_interface_stat_p rpc_stat, nrpc_stat;
7142 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7144 * We have to fix the offset of rpc_stat since we are
7145 * keeping this structure on two rx_queues. The rx_queue
7146 * package assumes that the rx_queue member is the first
7147 * member of the structure. That is, rx_queue assumes that
7148 * any one item is only on one queue at a time. We are
7149 * breaking that assumption and so we have to do a little
7150 * math to fix our pointers.
7153 fix_offset = (char *)rpc_stat;
7154 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7155 rpc_stat = (rx_interface_stat_p) fix_offset;
7158 * Copy the data based upon the caller version
7160 rx_MarshallProcessRPCStats(callerVersion,
7161 rpc_stat->stats[0].func_total,
7162 rpc_stat->stats, &ptr);
7168 MUTEX_EXIT(&rx_rpc_stats);
7173 * rx_FreeRPCStats - free memory allocated by
7174 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7178 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7179 * rx_RetrievePeerRPCStats
7181 * IN allocSize - the number of bytes in stats.
7189 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7191 rxi_Free(stats, allocSize);
7195 * rx_queryProcessRPCStats - see if process rpc stat collection is
7196 * currently enabled.
7202 * Returns 0 if stats are not enabled != 0 otherwise
7206 rx_queryProcessRPCStats(void)
7209 MUTEX_ENTER(&rx_rpc_stats);
7210 rc = rxi_monitor_processStats;
7211 MUTEX_EXIT(&rx_rpc_stats);
7216 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7222 * Returns 0 if stats are not enabled != 0 otherwise
7226 rx_queryPeerRPCStats(void)
7229 MUTEX_ENTER(&rx_rpc_stats);
7230 rc = rxi_monitor_peerStats;
7231 MUTEX_EXIT(&rx_rpc_stats);
7236 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7246 rx_enableProcessRPCStats(void)
7248 MUTEX_ENTER(&rx_rpc_stats);
7249 rx_enable_stats = 1;
7250 rxi_monitor_processStats = 1;
7251 MUTEX_EXIT(&rx_rpc_stats);
7255 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7265 rx_enablePeerRPCStats(void)
7267 MUTEX_ENTER(&rx_rpc_stats);
7268 rx_enable_stats = 1;
7269 rxi_monitor_peerStats = 1;
7270 MUTEX_EXIT(&rx_rpc_stats);
7274 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7284 rx_disableProcessRPCStats(void)
7286 rx_interface_stat_p rpc_stat, nrpc_stat;
7289 MUTEX_ENTER(&rx_rpc_stats);
7292 * Turn off process statistics and if peer stats is also off, turn
7296 rxi_monitor_processStats = 0;
7297 if (rxi_monitor_peerStats == 0) {
7298 rx_enable_stats = 0;
7301 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7302 unsigned int num_funcs = 0;
7305 queue_Remove(rpc_stat);
7306 num_funcs = rpc_stat->stats[0].func_total;
7308 sizeof(rx_interface_stat_t) +
7309 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7311 rxi_Free(rpc_stat, space);
7312 rxi_rpc_process_stat_cnt -= num_funcs;
7314 MUTEX_EXIT(&rx_rpc_stats);
7318 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7328 rx_disablePeerRPCStats(void)
7330 struct rx_peer **peer_ptr, **peer_end;
7333 MUTEX_ENTER(&rx_rpc_stats);
7336 * Turn off peer statistics and if process stats is also off, turn
7340 rxi_monitor_peerStats = 0;
7341 if (rxi_monitor_processStats == 0) {
7342 rx_enable_stats = 0;
7345 MUTEX_ENTER(&rx_peerHashTable_lock);
7346 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7347 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7349 struct rx_peer *peer, *next, *prev;
7350 for (prev = peer = *peer_ptr; peer; peer = next) {
7352 code = MUTEX_TRYENTER(&peer->peer_lock);
7354 rx_interface_stat_p rpc_stat, nrpc_stat;
7357 (&peer->rpcStats, rpc_stat, nrpc_stat,
7358 rx_interface_stat)) {
7359 unsigned int num_funcs = 0;
7362 queue_Remove(&rpc_stat->queue_header);
7363 queue_Remove(&rpc_stat->all_peers);
7364 num_funcs = rpc_stat->stats[0].func_total;
7366 sizeof(rx_interface_stat_t) +
7367 rpc_stat->stats[0].func_total *
7368 sizeof(rx_function_entry_v1_t);
7370 rxi_Free(rpc_stat, space);
7371 rxi_rpc_peer_stat_cnt -= num_funcs;
7373 MUTEX_EXIT(&peer->peer_lock);
7374 if (prev == *peer_ptr) {
7384 MUTEX_EXIT(&rx_peerHashTable_lock);
7385 MUTEX_EXIT(&rx_rpc_stats);
7389 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7394 * IN clearFlag - flag indicating which stats to clear
7402 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7404 rx_interface_stat_p rpc_stat, nrpc_stat;
7406 MUTEX_ENTER(&rx_rpc_stats);
7408 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7409 unsigned int num_funcs = 0, i;
7410 num_funcs = rpc_stat->stats[0].func_total;
7411 for (i = 0; i < num_funcs; i++) {
7412 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7413 hzero(rpc_stat->stats[i].invocations);
7415 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7416 hzero(rpc_stat->stats[i].bytes_sent);
7418 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7419 hzero(rpc_stat->stats[i].bytes_rcvd);
7421 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7422 rpc_stat->stats[i].queue_time_sum.sec = 0;
7423 rpc_stat->stats[i].queue_time_sum.usec = 0;
7425 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7426 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7427 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7429 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7430 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7431 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7433 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7434 rpc_stat->stats[i].queue_time_max.sec = 0;
7435 rpc_stat->stats[i].queue_time_max.usec = 0;
7437 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7438 rpc_stat->stats[i].execution_time_sum.sec = 0;
7439 rpc_stat->stats[i].execution_time_sum.usec = 0;
7441 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7442 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7443 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7445 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7446 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7447 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7449 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7450 rpc_stat->stats[i].execution_time_max.sec = 0;
7451 rpc_stat->stats[i].execution_time_max.usec = 0;
7456 MUTEX_EXIT(&rx_rpc_stats);
7460 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7465 * IN clearFlag - flag indicating which stats to clear
7473 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7475 rx_interface_stat_p rpc_stat, nrpc_stat;
7477 MUTEX_ENTER(&rx_rpc_stats);
7479 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7480 unsigned int num_funcs = 0, i;
7483 * We have to fix the offset of rpc_stat since we are
7484 * keeping this structure on two rx_queues. The rx_queue
7485 * package assumes that the rx_queue member is the first
7486 * member of the structure. That is, rx_queue assumes that
7487 * any one item is only on one queue at a time. We are
7488 * breaking that assumption and so we have to do a little
7489 * math to fix our pointers.
7492 fix_offset = (char *)rpc_stat;
7493 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7494 rpc_stat = (rx_interface_stat_p) fix_offset;
7496 num_funcs = rpc_stat->stats[0].func_total;
7497 for (i = 0; i < num_funcs; i++) {
7498 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7499 hzero(rpc_stat->stats[i].invocations);
7501 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7502 hzero(rpc_stat->stats[i].bytes_sent);
7504 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7505 hzero(rpc_stat->stats[i].bytes_rcvd);
7507 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7508 rpc_stat->stats[i].queue_time_sum.sec = 0;
7509 rpc_stat->stats[i].queue_time_sum.usec = 0;
7511 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7512 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7513 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7515 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7516 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7517 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7519 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7520 rpc_stat->stats[i].queue_time_max.sec = 0;
7521 rpc_stat->stats[i].queue_time_max.usec = 0;
7523 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7524 rpc_stat->stats[i].execution_time_sum.sec = 0;
7525 rpc_stat->stats[i].execution_time_sum.usec = 0;
7527 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7528 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7529 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7531 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7532 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7533 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7535 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7536 rpc_stat->stats[i].execution_time_max.sec = 0;
7537 rpc_stat->stats[i].execution_time_max.usec = 0;
7542 MUTEX_EXIT(&rx_rpc_stats);
7546 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7547 * is authorized to enable/disable/clear RX statistics.
7549 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7552 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7554 rxi_rxstat_userok = proc;
7558 rx_RxStatUserOk(struct rx_call *call)
7560 if (!rxi_rxstat_userok)
7562 return rxi_rxstat_userok(call);