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_ALPHA_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 rxkad_stats_mutex;
157 extern pthread_mutex_t des_init_mutex;
158 extern pthread_mutex_t des_random_mutex;
159 extern pthread_mutex_t rx_clock_mutex;
160 extern pthread_mutex_t rxi_connCacheMutex;
161 extern pthread_mutex_t rx_event_mutex;
162 extern pthread_mutex_t osi_malloc_mutex;
163 extern pthread_mutex_t event_handler_mutex;
164 extern pthread_mutex_t listener_mutex;
165 extern pthread_mutex_t rx_if_init_mutex;
166 extern pthread_mutex_t rx_if_mutex;
167 extern pthread_mutex_t rxkad_client_uid_mutex;
168 extern pthread_mutex_t rxkad_random_mutex;
170 extern pthread_cond_t rx_event_handler_cond;
171 extern pthread_cond_t rx_listener_cond;
173 static pthread_mutex_t epoch_mutex;
174 static pthread_mutex_t rx_init_mutex;
175 static pthread_mutex_t rx_debug_mutex;
178 rxi_InitPthread(void)
180 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
182 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
184 assert(pthread_mutex_init
185 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
186 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
188 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
190 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
192 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
194 assert(pthread_mutex_init
195 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
196 assert(pthread_mutex_init
197 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
198 assert(pthread_mutex_init
199 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
200 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
202 assert(pthread_mutex_init
203 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
204 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
206 assert(pthread_mutex_init
207 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
208 assert(pthread_mutex_init
209 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
210 assert(pthread_mutex_init
211 (&rxkad_stats_mutex, (const pthread_mutexattr_t *)0) == 0);
212 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
215 assert(pthread_cond_init
216 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
217 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
219 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
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 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
465 queue_Init(&rx_freePacketQueue);
466 rxi_NeedMorePackets = FALSE;
467 rxi_MorePackets(rx_nPackets);
475 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
476 tv.tv_sec = clock_now.sec;
477 tv.tv_usec = clock_now.usec;
478 srand((unsigned int)tv.tv_usec);
485 #if defined(KERNEL) && !defined(UKERNEL)
486 /* Really, this should never happen in a real kernel */
489 struct sockaddr_in addr;
490 int addrlen = sizeof(addr);
491 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
495 rx_port = addr.sin_port;
498 rx_stats.minRtt.sec = 9999999;
500 rx_SetEpoch(tv.tv_sec | 0x80000000);
502 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
503 * will provide a randomer value. */
505 MUTEX_ENTER(&rx_stats_mutex);
506 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
507 MUTEX_EXIT(&rx_stats_mutex);
508 /* *Slightly* random start time for the cid. This is just to help
509 * out with the hashing function at the peer */
510 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
511 rx_connHashTable = (struct rx_connection **)htable;
512 rx_peerHashTable = (struct rx_peer **)ptable;
514 rx_lastAckDelay.sec = 0;
515 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
516 rx_hardAckDelay.sec = 0;
517 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
518 rx_softAckDelay.sec = 0;
519 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
521 rxevent_Init(20, rxi_ReScheduleEvents);
523 /* Initialize various global queues */
524 queue_Init(&rx_idleServerQueue);
525 queue_Init(&rx_incomingCallQueue);
526 queue_Init(&rx_freeCallQueue);
528 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
529 /* Initialize our list of usable IP addresses. */
533 /* Start listener process (exact function is dependent on the
534 * implementation environment--kernel or user space) */
539 tmp_status = rxinit_status = 0;
547 return rx_InitHost(htonl(INADDR_ANY), port);
550 /* called with unincremented nRequestsRunning to see if it is OK to start
551 * a new thread in this service. Could be "no" for two reasons: over the
552 * max quota, or would prevent others from reaching their min quota.
554 #ifdef RX_ENABLE_LOCKS
555 /* This verion of QuotaOK reserves quota if it's ok while the
556 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
559 QuotaOK(register struct rx_service *aservice)
561 /* check if over max quota */
562 if (aservice->nRequestsRunning >= aservice->maxProcs) {
566 /* under min quota, we're OK */
567 /* otherwise, can use only if there are enough to allow everyone
568 * to go to their min quota after this guy starts.
570 MUTEX_ENTER(&rx_stats_mutex);
571 if ((aservice->nRequestsRunning < aservice->minProcs)
572 || (rxi_availProcs > rxi_minDeficit)) {
573 aservice->nRequestsRunning++;
574 /* just started call in minProcs pool, need fewer to maintain
576 if (aservice->nRequestsRunning <= aservice->minProcs)
579 MUTEX_EXIT(&rx_stats_mutex);
582 MUTEX_EXIT(&rx_stats_mutex);
588 ReturnToServerPool(register struct rx_service *aservice)
590 aservice->nRequestsRunning--;
591 MUTEX_ENTER(&rx_stats_mutex);
592 if (aservice->nRequestsRunning < aservice->minProcs)
595 MUTEX_EXIT(&rx_stats_mutex);
598 #else /* RX_ENABLE_LOCKS */
600 QuotaOK(register struct rx_service *aservice)
603 /* under min quota, we're OK */
604 if (aservice->nRequestsRunning < aservice->minProcs)
607 /* check if over max quota */
608 if (aservice->nRequestsRunning >= aservice->maxProcs)
611 /* otherwise, can use only if there are enough to allow everyone
612 * to go to their min quota after this guy starts.
614 if (rxi_availProcs > rxi_minDeficit)
618 #endif /* RX_ENABLE_LOCKS */
621 /* Called by rx_StartServer to start up lwp's to service calls.
622 NExistingProcs gives the number of procs already existing, and which
623 therefore needn't be created. */
625 rxi_StartServerProcs(int nExistingProcs)
627 register struct rx_service *service;
632 /* For each service, reserve N processes, where N is the "minimum"
633 * number of processes that MUST be able to execute a request in parallel,
634 * at any time, for that process. Also compute the maximum difference
635 * between any service's maximum number of processes that can run
636 * (i.e. the maximum number that ever will be run, and a guarantee
637 * that this number will run if other services aren't running), and its
638 * minimum number. The result is the extra number of processes that
639 * we need in order to provide the latter guarantee */
640 for (i = 0; i < RX_MAX_SERVICES; i++) {
642 service = rx_services[i];
643 if (service == (struct rx_service *)0)
645 nProcs += service->minProcs;
646 diff = service->maxProcs - service->minProcs;
650 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
651 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
652 for (i = 0; i < nProcs; i++) {
653 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
658 /* This routine must be called if any services are exported. If the
659 * donateMe flag is set, the calling process is donated to the server
662 rx_StartServer(int donateMe)
664 register struct rx_service *service;
671 /* Start server processes, if necessary (exact function is dependent
672 * on the implementation environment--kernel or user space). DonateMe
673 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
674 * case, one less new proc will be created rx_StartServerProcs.
676 rxi_StartServerProcs(donateMe);
678 /* count up the # of threads in minProcs, and add set the min deficit to
679 * be that value, too.
681 for (i = 0; i < RX_MAX_SERVICES; i++) {
682 service = rx_services[i];
683 if (service == (struct rx_service *)0)
685 MUTEX_ENTER(&rx_stats_mutex);
686 rxi_totalMin += service->minProcs;
687 /* below works even if a thread is running, since minDeficit would
688 * still have been decremented and later re-incremented.
690 rxi_minDeficit += service->minProcs;
691 MUTEX_EXIT(&rx_stats_mutex);
694 /* Turn on reaping of idle server connections */
695 rxi_ReapConnections();
705 #ifdef AFS_PTHREAD_ENV
707 pid = (pid_t) pthread_self();
708 #else /* AFS_PTHREAD_ENV */
710 LWP_CurrentProcess(&pid);
711 #endif /* AFS_PTHREAD_ENV */
713 sprintf(name, "srv_%d", ++nProcs);
715 (*registerProgram) (pid, name);
717 #endif /* AFS_NT40_ENV */
718 rx_ServerProc(); /* Never returns */
723 /* Create a new client connection to the specified service, using the
724 * specified security object to implement the security model for this
726 struct rx_connection *
727 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
728 register struct rx_securityClass *securityObject,
729 int serviceSecurityIndex)
733 register struct rx_connection *conn;
738 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", shost, sport, sservice, securityObject, serviceSecurityIndex));
740 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
741 * the case of kmem_alloc? */
742 conn = rxi_AllocConnection();
743 #ifdef RX_ENABLE_LOCKS
744 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
745 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
746 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
750 MUTEX_ENTER(&rx_connHashTable_lock);
751 cid = (rx_nextCid += RX_MAXCALLS);
752 conn->type = RX_CLIENT_CONNECTION;
754 conn->epoch = rx_epoch;
755 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
756 conn->serviceId = sservice;
757 conn->securityObject = securityObject;
758 /* This doesn't work in all compilers with void (they're buggy), so fake it
760 conn->securityData = (VOID *) 0;
761 conn->securityIndex = serviceSecurityIndex;
762 rx_SetConnDeadTime(conn, rx_connDeadTime);
763 conn->ackRate = RX_FAST_ACK_RATE;
765 conn->specific = NULL;
766 conn->challengeEvent = NULL;
767 conn->delayedAbortEvent = NULL;
768 conn->abortCount = 0;
771 RXS_NewConnection(securityObject, conn);
773 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
775 conn->refCount++; /* no lock required since only this thread knows... */
776 conn->next = rx_connHashTable[hashindex];
777 rx_connHashTable[hashindex] = conn;
778 MUTEX_ENTER(&rx_stats_mutex);
779 rx_stats.nClientConns++;
780 MUTEX_EXIT(&rx_stats_mutex);
782 MUTEX_EXIT(&rx_connHashTable_lock);
789 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
791 /* The idea is to set the dead time to a value that allows several
792 * keepalives to be dropped without timing out the connection. */
793 conn->secondsUntilDead = MAX(seconds, 6);
794 conn->secondsUntilPing = conn->secondsUntilDead / 6;
797 int rxi_lowPeerRefCount = 0;
798 int rxi_lowConnRefCount = 0;
801 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
802 * NOTE: must not be called with rx_connHashTable_lock held.
805 rxi_CleanupConnection(struct rx_connection *conn)
807 /* Notify the service exporter, if requested, that this connection
808 * is being destroyed */
809 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
810 (*conn->service->destroyConnProc) (conn);
812 /* Notify the security module that this connection is being destroyed */
813 RXS_DestroyConnection(conn->securityObject, conn);
815 /* If this is the last connection using the rx_peer struct, set its
816 * idle time to now. rxi_ReapConnections will reap it if it's still
817 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
819 MUTEX_ENTER(&rx_peerHashTable_lock);
820 if (conn->peer->refCount < 2) {
821 conn->peer->idleWhen = clock_Sec();
822 if (conn->peer->refCount < 1) {
823 conn->peer->refCount = 1;
824 MUTEX_ENTER(&rx_stats_mutex);
825 rxi_lowPeerRefCount++;
826 MUTEX_EXIT(&rx_stats_mutex);
829 conn->peer->refCount--;
830 MUTEX_EXIT(&rx_peerHashTable_lock);
832 MUTEX_ENTER(&rx_stats_mutex);
833 if (conn->type == RX_SERVER_CONNECTION)
834 rx_stats.nServerConns--;
836 rx_stats.nClientConns--;
837 MUTEX_EXIT(&rx_stats_mutex);
840 if (conn->specific) {
842 for (i = 0; i < conn->nSpecific; i++) {
843 if (conn->specific[i] && rxi_keyCreate_destructor[i])
844 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
845 conn->specific[i] = NULL;
847 free(conn->specific);
849 conn->specific = NULL;
853 MUTEX_DESTROY(&conn->conn_call_lock);
854 MUTEX_DESTROY(&conn->conn_data_lock);
855 CV_DESTROY(&conn->conn_call_cv);
857 rxi_FreeConnection(conn);
860 /* Destroy the specified connection */
862 rxi_DestroyConnection(register struct rx_connection *conn)
864 MUTEX_ENTER(&rx_connHashTable_lock);
865 rxi_DestroyConnectionNoLock(conn);
866 /* conn should be at the head of the cleanup list */
867 if (conn == rx_connCleanup_list) {
868 rx_connCleanup_list = rx_connCleanup_list->next;
869 MUTEX_EXIT(&rx_connHashTable_lock);
870 rxi_CleanupConnection(conn);
872 #ifdef RX_ENABLE_LOCKS
874 MUTEX_EXIT(&rx_connHashTable_lock);
876 #endif /* RX_ENABLE_LOCKS */
880 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
882 register struct rx_connection **conn_ptr;
883 register int havecalls = 0;
884 struct rx_packet *packet;
891 MUTEX_ENTER(&conn->conn_data_lock);
892 if (conn->refCount > 0)
895 MUTEX_ENTER(&rx_stats_mutex);
896 rxi_lowConnRefCount++;
897 MUTEX_EXIT(&rx_stats_mutex);
900 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
901 /* Busy; wait till the last guy before proceeding */
902 MUTEX_EXIT(&conn->conn_data_lock);
907 /* If the client previously called rx_NewCall, but it is still
908 * waiting, treat this as a running call, and wait to destroy the
909 * connection later when the call completes. */
910 if ((conn->type == RX_CLIENT_CONNECTION)
911 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
912 conn->flags |= RX_CONN_DESTROY_ME;
913 MUTEX_EXIT(&conn->conn_data_lock);
917 MUTEX_EXIT(&conn->conn_data_lock);
919 /* Check for extant references to this connection */
920 for (i = 0; i < RX_MAXCALLS; i++) {
921 register struct rx_call *call = conn->call[i];
924 if (conn->type == RX_CLIENT_CONNECTION) {
925 MUTEX_ENTER(&call->lock);
926 if (call->delayedAckEvent) {
927 /* Push the final acknowledgment out now--there
928 * won't be a subsequent call to acknowledge the
929 * last reply packets */
930 rxevent_Cancel(call->delayedAckEvent, call,
931 RX_CALL_REFCOUNT_DELAY);
932 if (call->state == RX_STATE_PRECALL
933 || call->state == RX_STATE_ACTIVE) {
934 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
936 rxi_AckAll(NULL, call, 0);
939 MUTEX_EXIT(&call->lock);
943 #ifdef RX_ENABLE_LOCKS
945 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
946 MUTEX_EXIT(&conn->conn_data_lock);
948 /* Someone is accessing a packet right now. */
952 #endif /* RX_ENABLE_LOCKS */
955 /* Don't destroy the connection if there are any call
956 * structures still in use */
957 MUTEX_ENTER(&conn->conn_data_lock);
958 conn->flags |= RX_CONN_DESTROY_ME;
959 MUTEX_EXIT(&conn->conn_data_lock);
964 if (conn->delayedAbortEvent) {
965 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
966 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
968 MUTEX_ENTER(&conn->conn_data_lock);
969 rxi_SendConnectionAbort(conn, packet, 0, 1);
970 MUTEX_EXIT(&conn->conn_data_lock);
971 rxi_FreePacket(packet);
975 /* Remove from connection hash table before proceeding */
977 &rx_connHashTable[CONN_HASH
978 (peer->host, peer->port, conn->cid, conn->epoch,
980 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
981 if (*conn_ptr == conn) {
982 *conn_ptr = conn->next;
986 /* if the conn that we are destroying was the last connection, then we
987 * clear rxLastConn as well */
988 if (rxLastConn == conn)
991 /* Make sure the connection is completely reset before deleting it. */
992 /* get rid of pending events that could zap us later */
993 if (conn->challengeEvent)
994 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
995 if (conn->checkReachEvent)
996 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
998 /* Add the connection to the list of destroyed connections that
999 * need to be cleaned up. This is necessary to avoid deadlocks
1000 * in the routines we call to inform others that this connection is
1001 * being destroyed. */
1002 conn->next = rx_connCleanup_list;
1003 rx_connCleanup_list = conn;
1006 /* Externally available version */
1008 rx_DestroyConnection(register struct rx_connection *conn)
1014 rxi_DestroyConnection(conn);
1020 rx_GetConnection(register struct rx_connection *conn)
1026 MUTEX_ENTER(&conn->conn_data_lock);
1028 MUTEX_EXIT(&conn->conn_data_lock);
1033 /* Start a new rx remote procedure call, on the specified connection.
1034 * If wait is set to 1, wait for a free call channel; otherwise return
1035 * 0. Maxtime gives the maximum number of seconds this call may take,
1036 * after rx_MakeCall returns. After this time interval, a call to any
1037 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1038 * For fine grain locking, we hold the conn_call_lock in order to
1039 * to ensure that we don't get signalle after we found a call in an active
1040 * state and before we go to sleep.
1043 rx_NewCall(register struct rx_connection *conn)
1046 register struct rx_call *call;
1047 struct clock queueTime;
1051 dpf(("rx_MakeCall(conn %x)\n", conn));
1054 clock_GetTime(&queueTime);
1056 MUTEX_ENTER(&conn->conn_call_lock);
1059 * Check if there are others waiting for a new call.
1060 * If so, let them go first to avoid starving them.
1061 * This is a fairly simple scheme, and might not be
1062 * a complete solution for large numbers of waiters.
1064 if (conn->makeCallWaiters) {
1065 #ifdef RX_ENABLE_LOCKS
1066 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1073 for (i = 0; i < RX_MAXCALLS; i++) {
1074 call = conn->call[i];
1076 MUTEX_ENTER(&call->lock);
1077 if (call->state == RX_STATE_DALLY) {
1078 rxi_ResetCall(call, 0);
1079 (*call->callNumber)++;
1082 MUTEX_EXIT(&call->lock);
1084 call = rxi_NewCall(conn, i);
1088 if (i < RX_MAXCALLS) {
1091 MUTEX_ENTER(&conn->conn_data_lock);
1092 conn->flags |= RX_CONN_MAKECALL_WAITING;
1093 MUTEX_EXIT(&conn->conn_data_lock);
1095 conn->makeCallWaiters++;
1096 #ifdef RX_ENABLE_LOCKS
1097 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1101 conn->makeCallWaiters--;
1104 * Wake up anyone else who might be giving us a chance to
1105 * run (see code above that avoids resource starvation).
1107 #ifdef RX_ENABLE_LOCKS
1108 CV_BROADCAST(&conn->conn_call_cv);
1113 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1115 /* Client is initially in send mode */
1116 call->state = RX_STATE_ACTIVE;
1117 call->mode = RX_MODE_SENDING;
1119 /* remember start time for call in case we have hard dead time limit */
1120 call->queueTime = queueTime;
1121 clock_GetTime(&call->startTime);
1122 hzero(call->bytesSent);
1123 hzero(call->bytesRcvd);
1125 /* Turn on busy protocol. */
1126 rxi_KeepAliveOn(call);
1128 MUTEX_EXIT(&call->lock);
1129 MUTEX_EXIT(&conn->conn_call_lock);
1133 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1134 /* Now, if TQ wasn't cleared earlier, do it now. */
1136 MUTEX_ENTER(&call->lock);
1137 while (call->flags & RX_CALL_TQ_BUSY) {
1138 call->flags |= RX_CALL_TQ_WAIT;
1139 #ifdef RX_ENABLE_LOCKS
1140 CV_WAIT(&call->cv_tq, &call->lock);
1141 #else /* RX_ENABLE_LOCKS */
1142 osi_rxSleep(&call->tq);
1143 #endif /* RX_ENABLE_LOCKS */
1145 if (call->flags & RX_CALL_TQ_CLEARME) {
1146 rxi_ClearTransmitQueue(call, 0);
1147 queue_Init(&call->tq);
1149 MUTEX_EXIT(&call->lock);
1151 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1157 rxi_HasActiveCalls(register struct rx_connection *aconn)
1160 register struct rx_call *tcall;
1164 for (i = 0; i < RX_MAXCALLS; i++) {
1165 if ((tcall = aconn->call[i])) {
1166 if ((tcall->state == RX_STATE_ACTIVE)
1167 || (tcall->state == RX_STATE_PRECALL)) {
1178 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1179 register afs_int32 * aint32s)
1182 register struct rx_call *tcall;
1186 for (i = 0; i < RX_MAXCALLS; i++) {
1187 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1188 aint32s[i] = aconn->callNumber[i] + 1;
1190 aint32s[i] = aconn->callNumber[i];
1197 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1198 register afs_int32 * aint32s)
1201 register struct rx_call *tcall;
1205 for (i = 0; i < RX_MAXCALLS; i++) {
1206 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1207 aconn->callNumber[i] = aint32s[i] - 1;
1209 aconn->callNumber[i] = aint32s[i];
1215 /* Advertise a new service. A service is named locally by a UDP port
1216 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1219 char *serviceName; Name for identification purposes (e.g. the
1220 service name might be used for probing for
1223 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1224 struct rx_securityClass **securityObjects, int nSecurityObjects,
1225 afs_int32(*serviceProc) (struct rx_call * acall))
1227 osi_socket socket = OSI_NULLSOCKET;
1228 register struct rx_service *tservice;
1234 if (serviceId == 0) {
1236 "rx_NewService: service id for service %s is not non-zero.\n",
1243 "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",
1251 tservice = rxi_AllocService();
1254 for (i = 0; i < RX_MAX_SERVICES; i++) {
1255 register struct rx_service *service = rx_services[i];
1257 if (port == service->servicePort) {
1258 if (service->serviceId == serviceId) {
1259 /* The identical service has already been
1260 * installed; if the caller was intending to
1261 * change the security classes used by this
1262 * service, he/she loses. */
1264 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1265 serviceName, serviceId, service->serviceName);
1268 rxi_FreeService(tservice);
1271 /* Different service, same port: re-use the socket
1272 * which is bound to the same port */
1273 socket = service->socket;
1276 if (socket == OSI_NULLSOCKET) {
1277 /* If we don't already have a socket (from another
1278 * service on same port) get a new one */
1279 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1280 if (socket == OSI_NULLSOCKET) {
1283 rxi_FreeService(tservice);
1288 service->socket = socket;
1289 service->servicePort = port;
1290 service->serviceId = serviceId;
1291 service->serviceName = serviceName;
1292 service->nSecurityObjects = nSecurityObjects;
1293 service->securityObjects = securityObjects;
1294 service->minProcs = 0;
1295 service->maxProcs = 1;
1296 service->idleDeadTime = 60;
1297 service->connDeadTime = rx_connDeadTime;
1298 service->executeRequestProc = serviceProc;
1299 service->checkReach = 0;
1300 rx_services[i] = service; /* not visible until now */
1308 rxi_FreeService(tservice);
1309 (osi_Msg "rx_NewService: cannot support > %d services\n",
1314 /* Generic request processing loop. This routine should be called
1315 * by the implementation dependent rx_ServerProc. If socketp is
1316 * non-null, it will be set to the file descriptor that this thread
1317 * is now listening on. If socketp is null, this routine will never
1320 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1322 register struct rx_call *call;
1323 register afs_int32 code;
1324 register struct rx_service *tservice = NULL;
1331 call = rx_GetCall(threadID, tservice, socketp);
1332 if (socketp && *socketp != OSI_NULLSOCKET) {
1333 /* We are now a listener thread */
1338 /* if server is restarting( typically smooth shutdown) then do not
1339 * allow any new calls.
1342 if (rx_tranquil && (call != NULL)) {
1347 MUTEX_ENTER(&call->lock);
1349 rxi_CallError(call, RX_RESTARTING);
1350 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1352 MUTEX_EXIT(&call->lock);
1357 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1358 #ifdef RX_ENABLE_LOCKS
1360 #endif /* RX_ENABLE_LOCKS */
1361 afs_termState = AFSOP_STOP_AFS;
1362 afs_osi_Wakeup(&afs_termState);
1363 #ifdef RX_ENABLE_LOCKS
1365 #endif /* RX_ENABLE_LOCKS */
1370 tservice = call->conn->service;
1372 if (tservice->beforeProc)
1373 (*tservice->beforeProc) (call);
1375 code = call->conn->service->executeRequestProc(call);
1377 if (tservice->afterProc)
1378 (*tservice->afterProc) (call, code);
1380 rx_EndCall(call, code);
1381 MUTEX_ENTER(&rx_stats_mutex);
1383 MUTEX_EXIT(&rx_stats_mutex);
1389 rx_WakeupServerProcs(void)
1391 struct rx_serverQueueEntry *np, *tqp;
1396 MUTEX_ENTER(&rx_serverPool_lock);
1398 #ifdef RX_ENABLE_LOCKS
1399 if (rx_waitForPacket)
1400 CV_BROADCAST(&rx_waitForPacket->cv);
1401 #else /* RX_ENABLE_LOCKS */
1402 if (rx_waitForPacket)
1403 osi_rxWakeup(rx_waitForPacket);
1404 #endif /* RX_ENABLE_LOCKS */
1405 MUTEX_ENTER(&freeSQEList_lock);
1406 for (np = rx_FreeSQEList; np; np = tqp) {
1407 tqp = *(struct rx_serverQueueEntry **)np;
1408 #ifdef RX_ENABLE_LOCKS
1409 CV_BROADCAST(&np->cv);
1410 #else /* RX_ENABLE_LOCKS */
1412 #endif /* RX_ENABLE_LOCKS */
1414 MUTEX_EXIT(&freeSQEList_lock);
1415 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1416 #ifdef RX_ENABLE_LOCKS
1417 CV_BROADCAST(&np->cv);
1418 #else /* RX_ENABLE_LOCKS */
1420 #endif /* RX_ENABLE_LOCKS */
1422 MUTEX_EXIT(&rx_serverPool_lock);
1428 * One thing that seems to happen is that all the server threads get
1429 * tied up on some empty or slow call, and then a whole bunch of calls
1430 * arrive at once, using up the packet pool, so now there are more
1431 * empty calls. The most critical resources here are server threads
1432 * and the free packet pool. The "doreclaim" code seems to help in
1433 * general. I think that eventually we arrive in this state: there
1434 * are lots of pending calls which do have all their packets present,
1435 * so they won't be reclaimed, are multi-packet calls, so they won't
1436 * be scheduled until later, and thus are tying up most of the free
1437 * packet pool for a very long time.
1439 * 1. schedule multi-packet calls if all the packets are present.
1440 * Probably CPU-bound operation, useful to return packets to pool.
1441 * Do what if there is a full window, but the last packet isn't here?
1442 * 3. preserve one thread which *only* runs "best" calls, otherwise
1443 * it sleeps and waits for that type of call.
1444 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1445 * the current dataquota business is badly broken. The quota isn't adjusted
1446 * to reflect how many packets are presently queued for a running call.
1447 * So, when we schedule a queued call with a full window of packets queued
1448 * up for it, that *should* free up a window full of packets for other 2d-class
1449 * calls to be able to use from the packet pool. But it doesn't.
1451 * NB. Most of the time, this code doesn't run -- since idle server threads
1452 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1453 * as a new call arrives.
1455 /* Sleep until a call arrives. Returns a pointer to the call, ready
1456 * for an rx_Read. */
1457 #ifdef RX_ENABLE_LOCKS
1459 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1461 struct rx_serverQueueEntry *sq;
1462 register struct rx_call *call = (struct rx_call *)0;
1463 struct rx_service *service = NULL;
1466 MUTEX_ENTER(&freeSQEList_lock);
1468 if ((sq = rx_FreeSQEList)) {
1469 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1470 MUTEX_EXIT(&freeSQEList_lock);
1471 } else { /* otherwise allocate a new one and return that */
1472 MUTEX_EXIT(&freeSQEList_lock);
1473 sq = (struct rx_serverQueueEntry *)
1474 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1475 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1476 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1479 MUTEX_ENTER(&rx_serverPool_lock);
1480 if (cur_service != NULL) {
1481 ReturnToServerPool(cur_service);
1484 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1485 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1487 /* Scan for eligible incoming calls. A call is not eligible
1488 * if the maximum number of calls for its service type are
1489 * already executing */
1490 /* One thread will process calls FCFS (to prevent starvation),
1491 * while the other threads may run ahead looking for calls which
1492 * have all their input data available immediately. This helps
1493 * keep threads from blocking, waiting for data from the client. */
1494 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1495 service = tcall->conn->service;
1496 if (!QuotaOK(service)) {
1499 if (tno == rxi_fcfs_thread_num
1500 || !tcall->queue_item_header.next) {
1501 /* If we're the fcfs thread , then we'll just use
1502 * this call. If we haven't been able to find an optimal
1503 * choice, and we're at the end of the list, then use a
1504 * 2d choice if one has been identified. Otherwise... */
1505 call = (choice2 ? choice2 : tcall);
1506 service = call->conn->service;
1507 } else if (!queue_IsEmpty(&tcall->rq)) {
1508 struct rx_packet *rp;
1509 rp = queue_First(&tcall->rq, rx_packet);
1510 if (rp->header.seq == 1) {
1512 || (rp->header.flags & RX_LAST_PACKET)) {
1514 } else if (rxi_2dchoice && !choice2
1515 && !(tcall->flags & RX_CALL_CLEARED)
1516 && (tcall->rprev > rxi_HardAckRate)) {
1525 ReturnToServerPool(service);
1532 MUTEX_EXIT(&rx_serverPool_lock);
1533 MUTEX_ENTER(&call->lock);
1535 if (call->flags & RX_CALL_WAIT_PROC) {
1536 call->flags &= ~RX_CALL_WAIT_PROC;
1537 MUTEX_ENTER(&rx_stats_mutex);
1539 MUTEX_EXIT(&rx_stats_mutex);
1542 if (call->state != RX_STATE_PRECALL || call->error) {
1543 MUTEX_EXIT(&call->lock);
1544 MUTEX_ENTER(&rx_serverPool_lock);
1545 ReturnToServerPool(service);
1550 if (queue_IsEmpty(&call->rq)
1551 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1552 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1554 CLEAR_CALL_QUEUE_LOCK(call);
1557 /* If there are no eligible incoming calls, add this process
1558 * to the idle server queue, to wait for one */
1562 *socketp = OSI_NULLSOCKET;
1564 sq->socketp = socketp;
1565 queue_Append(&rx_idleServerQueue, sq);
1566 #ifndef AFS_AIX41_ENV
1567 rx_waitForPacket = sq;
1569 rx_waitingForPacket = sq;
1570 #endif /* AFS_AIX41_ENV */
1572 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1574 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1575 MUTEX_EXIT(&rx_serverPool_lock);
1576 return (struct rx_call *)0;
1579 } while (!(call = sq->newcall)
1580 && !(socketp && *socketp != OSI_NULLSOCKET));
1581 MUTEX_EXIT(&rx_serverPool_lock);
1583 MUTEX_ENTER(&call->lock);
1589 MUTEX_ENTER(&freeSQEList_lock);
1590 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1591 rx_FreeSQEList = sq;
1592 MUTEX_EXIT(&freeSQEList_lock);
1595 clock_GetTime(&call->startTime);
1596 call->state = RX_STATE_ACTIVE;
1597 call->mode = RX_MODE_RECEIVING;
1598 #ifdef RX_KERNEL_TRACE
1599 if (ICL_SETACTIVE(afs_iclSetp)) {
1600 int glockOwner = ISAFS_GLOCK();
1603 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1604 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1611 rxi_calltrace(RX_CALL_START, call);
1612 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1613 call->conn->service->servicePort, call->conn->service->serviceId,
1616 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1617 MUTEX_EXIT(&call->lock);
1619 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1624 #else /* RX_ENABLE_LOCKS */
1626 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1628 struct rx_serverQueueEntry *sq;
1629 register struct rx_call *call = (struct rx_call *)0, *choice2;
1630 struct rx_service *service = NULL;
1635 MUTEX_ENTER(&freeSQEList_lock);
1637 if ((sq = rx_FreeSQEList)) {
1638 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1639 MUTEX_EXIT(&freeSQEList_lock);
1640 } else { /* otherwise allocate a new one and return that */
1641 MUTEX_EXIT(&freeSQEList_lock);
1642 sq = (struct rx_serverQueueEntry *)
1643 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1644 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1645 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1647 MUTEX_ENTER(&sq->lock);
1649 if (cur_service != NULL) {
1650 cur_service->nRequestsRunning--;
1651 if (cur_service->nRequestsRunning < cur_service->minProcs)
1655 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1656 register struct rx_call *tcall, *ncall;
1657 /* Scan for eligible incoming calls. A call is not eligible
1658 * if the maximum number of calls for its service type are
1659 * already executing */
1660 /* One thread will process calls FCFS (to prevent starvation),
1661 * while the other threads may run ahead looking for calls which
1662 * have all their input data available immediately. This helps
1663 * keep threads from blocking, waiting for data from the client. */
1664 choice2 = (struct rx_call *)0;
1665 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1666 service = tcall->conn->service;
1667 if (QuotaOK(service)) {
1668 if (tno == rxi_fcfs_thread_num
1669 || !tcall->queue_item_header.next) {
1670 /* If we're the fcfs thread, then we'll just use
1671 * this call. If we haven't been able to find an optimal
1672 * choice, and we're at the end of the list, then use a
1673 * 2d choice if one has been identified. Otherwise... */
1674 call = (choice2 ? choice2 : tcall);
1675 service = call->conn->service;
1676 } else if (!queue_IsEmpty(&tcall->rq)) {
1677 struct rx_packet *rp;
1678 rp = queue_First(&tcall->rq, rx_packet);
1679 if (rp->header.seq == 1
1681 || (rp->header.flags & RX_LAST_PACKET))) {
1683 } else if (rxi_2dchoice && !choice2
1684 && !(tcall->flags & RX_CALL_CLEARED)
1685 && (tcall->rprev > rxi_HardAckRate)) {
1698 /* we can't schedule a call if there's no data!!! */
1699 /* send an ack if there's no data, if we're missing the
1700 * first packet, or we're missing something between first
1701 * and last -- there's a "hole" in the incoming data. */
1702 if (queue_IsEmpty(&call->rq)
1703 || queue_First(&call->rq, rx_packet)->header.seq != 1
1704 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1705 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1707 call->flags &= (~RX_CALL_WAIT_PROC);
1708 service->nRequestsRunning++;
1709 /* just started call in minProcs pool, need fewer to maintain
1711 if (service->nRequestsRunning <= service->minProcs)
1715 /* MUTEX_EXIT(&call->lock); */
1717 /* If there are no eligible incoming calls, add this process
1718 * to the idle server queue, to wait for one */
1721 *socketp = OSI_NULLSOCKET;
1723 sq->socketp = socketp;
1724 queue_Append(&rx_idleServerQueue, sq);
1728 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1731 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1732 return (struct rx_call *)0;
1735 } while (!(call = sq->newcall)
1736 && !(socketp && *socketp != OSI_NULLSOCKET));
1738 MUTEX_EXIT(&sq->lock);
1740 MUTEX_ENTER(&freeSQEList_lock);
1741 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1742 rx_FreeSQEList = sq;
1743 MUTEX_EXIT(&freeSQEList_lock);
1746 clock_GetTime(&call->startTime);
1747 call->state = RX_STATE_ACTIVE;
1748 call->mode = RX_MODE_RECEIVING;
1749 #ifdef RX_KERNEL_TRACE
1750 if (ICL_SETACTIVE(afs_iclSetp)) {
1751 int glockOwner = ISAFS_GLOCK();
1754 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1755 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1762 rxi_calltrace(RX_CALL_START, call);
1763 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1764 call->conn->service->servicePort, call->conn->service->serviceId,
1767 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1775 #endif /* RX_ENABLE_LOCKS */
1779 /* Establish a procedure to be called when a packet arrives for a
1780 * call. This routine will be called at most once after each call,
1781 * and will also be called if there is an error condition on the or
1782 * the call is complete. Used by multi rx to build a selection
1783 * function which determines which of several calls is likely to be a
1784 * good one to read from.
1785 * NOTE: the way this is currently implemented it is probably only a
1786 * good idea to (1) use it immediately after a newcall (clients only)
1787 * and (2) only use it once. Other uses currently void your warranty
1790 rx_SetArrivalProc(register struct rx_call *call,
1791 register VOID(*proc) (register struct rx_call * call,
1792 register struct multi_handle * mh,
1793 register int index),
1794 register VOID * handle, register VOID * arg)
1796 call->arrivalProc = proc;
1797 call->arrivalProcHandle = handle;
1798 call->arrivalProcArg = arg;
1801 /* Call is finished (possibly prematurely). Return rc to the peer, if
1802 * appropriate, and return the final error code from the conversation
1806 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1808 register struct rx_connection *conn = call->conn;
1809 register struct rx_service *service;
1810 register struct rx_packet *tp; /* Temporary packet pointer */
1811 register struct rx_packet *nxp; /* Next packet pointer, for queue_Scan */
1815 dpf(("rx_EndCall(call %x)\n", call));
1819 MUTEX_ENTER(&call->lock);
1821 if (rc == 0 && call->error == 0) {
1822 call->abortCode = 0;
1823 call->abortCount = 0;
1826 call->arrivalProc = (VOID(*)())0;
1827 if (rc && call->error == 0) {
1828 rxi_CallError(call, rc);
1829 /* Send an abort message to the peer if this error code has
1830 * only just been set. If it was set previously, assume the
1831 * peer has already been sent the error code or will request it
1833 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1835 if (conn->type == RX_SERVER_CONNECTION) {
1836 /* Make sure reply or at least dummy reply is sent */
1837 if (call->mode == RX_MODE_RECEIVING) {
1838 rxi_WriteProc(call, 0, 0);
1840 if (call->mode == RX_MODE_SENDING) {
1841 rxi_FlushWrite(call);
1843 service = conn->service;
1844 rxi_calltrace(RX_CALL_END, call);
1845 /* Call goes to hold state until reply packets are acknowledged */
1846 if (call->tfirst + call->nSoftAcked < call->tnext) {
1847 call->state = RX_STATE_HOLD;
1849 call->state = RX_STATE_DALLY;
1850 rxi_ClearTransmitQueue(call, 0);
1851 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1852 rxevent_Cancel(call->keepAliveEvent, call,
1853 RX_CALL_REFCOUNT_ALIVE);
1855 } else { /* Client connection */
1857 /* Make sure server receives input packets, in the case where
1858 * no reply arguments are expected */
1859 if ((call->mode == RX_MODE_SENDING)
1860 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1861 (void)rxi_ReadProc(call, &dummy, 1);
1864 /* If we had an outstanding delayed ack, be nice to the server
1865 * and force-send it now.
1867 if (call->delayedAckEvent) {
1868 rxevent_Cancel(call->delayedAckEvent, call,
1869 RX_CALL_REFCOUNT_DELAY);
1870 call->delayedAckEvent = NULL;
1871 rxi_SendDelayedAck(NULL, call, NULL);
1874 /* We need to release the call lock since it's lower than the
1875 * conn_call_lock and we don't want to hold the conn_call_lock
1876 * over the rx_ReadProc call. The conn_call_lock needs to be held
1877 * here for the case where rx_NewCall is perusing the calls on
1878 * the connection structure. We don't want to signal until
1879 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1880 * have checked this call, found it active and by the time it
1881 * goes to sleep, will have missed the signal.
1883 MUTEX_EXIT(&call->lock);
1884 MUTEX_ENTER(&conn->conn_call_lock);
1885 MUTEX_ENTER(&call->lock);
1886 MUTEX_ENTER(&conn->conn_data_lock);
1887 conn->flags |= RX_CONN_BUSY;
1888 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1889 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1890 MUTEX_EXIT(&conn->conn_data_lock);
1891 #ifdef RX_ENABLE_LOCKS
1892 CV_BROADCAST(&conn->conn_call_cv);
1897 #ifdef RX_ENABLE_LOCKS
1899 MUTEX_EXIT(&conn->conn_data_lock);
1901 #endif /* RX_ENABLE_LOCKS */
1902 call->state = RX_STATE_DALLY;
1904 error = call->error;
1906 /* currentPacket, nLeft, and NFree must be zeroed here, because
1907 * ResetCall cannot: ResetCall may be called at splnet(), in the
1908 * kernel version, and may interrupt the macros rx_Read or
1909 * rx_Write, which run at normal priority for efficiency. */
1910 if (call->currentPacket) {
1911 rxi_FreePacket(call->currentPacket);
1912 call->currentPacket = (struct rx_packet *)0;
1913 call->nLeft = call->nFree = call->curlen = 0;
1915 call->nLeft = call->nFree = call->curlen = 0;
1917 /* Free any packets from the last call to ReadvProc/WritevProc */
1918 for (queue_Scan(&call->iovq, tp, nxp, rx_packet)) {
1923 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1924 MUTEX_EXIT(&call->lock);
1925 if (conn->type == RX_CLIENT_CONNECTION) {
1926 MUTEX_EXIT(&conn->conn_call_lock);
1927 conn->flags &= ~RX_CONN_BUSY;
1932 * Map errors to the local host's errno.h format.
1934 error = ntoh_syserr_conv(error);
1938 #if !defined(KERNEL)
1940 /* Call this routine when shutting down a server or client (especially
1941 * clients). This will allow Rx to gracefully garbage collect server
1942 * connections, and reduce the number of retries that a server might
1943 * make to a dead client.
1944 * This is not quite right, since some calls may still be ongoing and
1945 * we can't lock them to destroy them. */
1949 register struct rx_connection **conn_ptr, **conn_end;
1953 if (rxinit_status == 1) {
1955 return; /* Already shutdown. */
1957 rxi_DeleteCachedConnections();
1958 if (rx_connHashTable) {
1959 MUTEX_ENTER(&rx_connHashTable_lock);
1960 for (conn_ptr = &rx_connHashTable[0], conn_end =
1961 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
1963 struct rx_connection *conn, *next;
1964 for (conn = *conn_ptr; conn; conn = next) {
1966 if (conn->type == RX_CLIENT_CONNECTION) {
1967 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
1969 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
1970 #ifdef RX_ENABLE_LOCKS
1971 rxi_DestroyConnectionNoLock(conn);
1972 #else /* RX_ENABLE_LOCKS */
1973 rxi_DestroyConnection(conn);
1974 #endif /* RX_ENABLE_LOCKS */
1978 #ifdef RX_ENABLE_LOCKS
1979 while (rx_connCleanup_list) {
1980 struct rx_connection *conn;
1981 conn = rx_connCleanup_list;
1982 rx_connCleanup_list = rx_connCleanup_list->next;
1983 MUTEX_EXIT(&rx_connHashTable_lock);
1984 rxi_CleanupConnection(conn);
1985 MUTEX_ENTER(&rx_connHashTable_lock);
1987 MUTEX_EXIT(&rx_connHashTable_lock);
1988 #endif /* RX_ENABLE_LOCKS */
1997 /* if we wakeup packet waiter too often, can get in loop with two
1998 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2000 rxi_PacketsUnWait(void)
2002 if (!rx_waitingForPackets) {
2006 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2007 return; /* still over quota */
2010 rx_waitingForPackets = 0;
2011 #ifdef RX_ENABLE_LOCKS
2012 CV_BROADCAST(&rx_waitingForPackets_cv);
2014 osi_rxWakeup(&rx_waitingForPackets);
2020 /* ------------------Internal interfaces------------------------- */
2022 /* Return this process's service structure for the
2023 * specified socket and service */
2025 rxi_FindService(register osi_socket socket, register u_short serviceId)
2027 register struct rx_service **sp;
2028 for (sp = &rx_services[0]; *sp; sp++) {
2029 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2035 /* Allocate a call structure, for the indicated channel of the
2036 * supplied connection. The mode and state of the call must be set by
2037 * the caller. Returns the call with mutex locked. */
2039 rxi_NewCall(register struct rx_connection *conn, register int channel)
2041 register struct rx_call *call;
2042 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2043 register struct rx_call *cp; /* Call pointer temp */
2044 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2045 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2047 /* Grab an existing call structure, or allocate a new one.
2048 * Existing call structures are assumed to have been left reset by
2050 MUTEX_ENTER(&rx_freeCallQueue_lock);
2052 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2054 * EXCEPT that the TQ might not yet be cleared out.
2055 * Skip over those with in-use TQs.
2058 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2059 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2065 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2066 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2067 call = queue_First(&rx_freeCallQueue, rx_call);
2068 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2070 MUTEX_ENTER(&rx_stats_mutex);
2071 rx_stats.nFreeCallStructs--;
2072 MUTEX_EXIT(&rx_stats_mutex);
2073 MUTEX_EXIT(&rx_freeCallQueue_lock);
2074 MUTEX_ENTER(&call->lock);
2075 CLEAR_CALL_QUEUE_LOCK(call);
2076 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2077 /* Now, if TQ wasn't cleared earlier, do it now. */
2078 if (call->flags & RX_CALL_TQ_CLEARME) {
2079 rxi_ClearTransmitQueue(call, 0);
2080 queue_Init(&call->tq);
2082 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2083 /* Bind the call to its connection structure */
2085 rxi_ResetCall(call, 1);
2087 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2089 MUTEX_EXIT(&rx_freeCallQueue_lock);
2090 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2091 MUTEX_ENTER(&call->lock);
2092 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2093 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2094 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2096 MUTEX_ENTER(&rx_stats_mutex);
2097 rx_stats.nCallStructs++;
2098 MUTEX_EXIT(&rx_stats_mutex);
2099 /* Initialize once-only items */
2100 queue_Init(&call->tq);
2101 queue_Init(&call->rq);
2102 queue_Init(&call->iovq);
2103 /* Bind the call to its connection structure (prereq for reset) */
2105 rxi_ResetCall(call, 1);
2107 call->channel = channel;
2108 call->callNumber = &conn->callNumber[channel];
2109 /* Note that the next expected call number is retained (in
2110 * conn->callNumber[i]), even if we reallocate the call structure
2112 conn->call[channel] = call;
2113 /* if the channel's never been used (== 0), we should start at 1, otherwise
2114 * the call number is valid from the last time this channel was used */
2115 if (*call->callNumber == 0)
2116 *call->callNumber = 1;
2121 /* A call has been inactive long enough that so we can throw away
2122 * state, including the call structure, which is placed on the call
2124 * Call is locked upon entry.
2125 * haveCTLock set if called from rxi_ReapConnections
2127 #ifdef RX_ENABLE_LOCKS
2129 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2130 #else /* RX_ENABLE_LOCKS */
2132 rxi_FreeCall(register struct rx_call *call)
2133 #endif /* RX_ENABLE_LOCKS */
2135 register int channel = call->channel;
2136 register struct rx_connection *conn = call->conn;
2139 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2140 (*call->callNumber)++;
2141 rxi_ResetCall(call, 0);
2142 call->conn->call[channel] = (struct rx_call *)0;
2144 MUTEX_ENTER(&rx_freeCallQueue_lock);
2145 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2146 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2147 /* A call may be free even though its transmit queue is still in use.
2148 * Since we search the call list from head to tail, put busy calls at
2149 * the head of the list, and idle calls at the tail.
2151 if (call->flags & RX_CALL_TQ_BUSY)
2152 queue_Prepend(&rx_freeCallQueue, call);
2154 queue_Append(&rx_freeCallQueue, call);
2155 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2156 queue_Append(&rx_freeCallQueue, call);
2157 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2158 MUTEX_ENTER(&rx_stats_mutex);
2159 rx_stats.nFreeCallStructs++;
2160 MUTEX_EXIT(&rx_stats_mutex);
2162 MUTEX_EXIT(&rx_freeCallQueue_lock);
2164 /* Destroy the connection if it was previously slated for
2165 * destruction, i.e. the Rx client code previously called
2166 * rx_DestroyConnection (client connections), or
2167 * rxi_ReapConnections called the same routine (server
2168 * connections). Only do this, however, if there are no
2169 * outstanding calls. Note that for fine grain locking, there appears
2170 * to be a deadlock in that rxi_FreeCall has a call locked and
2171 * DestroyConnectionNoLock locks each call in the conn. But note a
2172 * few lines up where we have removed this call from the conn.
2173 * If someone else destroys a connection, they either have no
2174 * call lock held or are going through this section of code.
2176 if (conn->flags & RX_CONN_DESTROY_ME) {
2177 MUTEX_ENTER(&conn->conn_data_lock);
2179 MUTEX_EXIT(&conn->conn_data_lock);
2180 #ifdef RX_ENABLE_LOCKS
2182 rxi_DestroyConnectionNoLock(conn);
2184 rxi_DestroyConnection(conn);
2185 #else /* RX_ENABLE_LOCKS */
2186 rxi_DestroyConnection(conn);
2187 #endif /* RX_ENABLE_LOCKS */
2191 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2193 rxi_Alloc(register size_t size)
2197 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2198 /* Grab the AFS filesystem lock. See afs/osi.h for the lock
2201 int glockOwner = ISAFS_GLOCK();
2205 MUTEX_ENTER(&rx_stats_mutex);
2207 rxi_Allocsize += size;
2208 MUTEX_EXIT(&rx_stats_mutex);
2209 #if (defined(AFS_AIX32_ENV) || defined(AFS_HPUX_ENV)) && !defined(AFS_HPUX100_ENV) && defined(KERNEL)
2210 if (size > AFS_SMALLOCSIZ) {
2211 p = (char *)osi_AllocMediumSpace(size);
2213 p = (char *)osi_AllocSmall(size, 1);
2214 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2219 p = (char *)osi_Alloc(size);
2222 osi_Panic("rxi_Alloc error");
2228 rxi_Free(void *addr, register size_t size)
2230 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2231 /* Grab the AFS filesystem lock. See afs/osi.h for the lock
2234 int glockOwner = ISAFS_GLOCK();
2238 MUTEX_ENTER(&rx_stats_mutex);
2240 rxi_Allocsize -= size;
2241 MUTEX_EXIT(&rx_stats_mutex);
2242 #if (defined(AFS_AIX32_ENV) || defined(AFS_HPUX_ENV)) && !defined(AFS_HPUX100_ENV) && defined(KERNEL)
2243 if (size > AFS_SMALLOCSIZ)
2244 osi_FreeMediumSpace(addr);
2246 osi_FreeSmall(addr);
2247 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2252 osi_Free(addr, size);
2256 /* Find the peer process represented by the supplied (host,port)
2257 * combination. If there is no appropriate active peer structure, a
2258 * new one will be allocated and initialized
2259 * The origPeer, if set, is a pointer to a peer structure on which the
2260 * refcount will be be decremented. This is used to replace the peer
2261 * structure hanging off a connection structure */
2263 rxi_FindPeer(register afs_uint32 host, register u_short port,
2264 struct rx_peer *origPeer, int create)
2266 register struct rx_peer *pp;
2268 hashIndex = PEER_HASH(host, port);
2269 MUTEX_ENTER(&rx_peerHashTable_lock);
2270 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2271 if ((pp->host == host) && (pp->port == port))
2276 pp = rxi_AllocPeer(); /* This bzero's *pp */
2277 pp->host = host; /* set here or in InitPeerParams is zero */
2279 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2280 queue_Init(&pp->congestionQueue);
2281 queue_Init(&pp->rpcStats);
2282 pp->next = rx_peerHashTable[hashIndex];
2283 rx_peerHashTable[hashIndex] = pp;
2284 rxi_InitPeerParams(pp);
2285 MUTEX_ENTER(&rx_stats_mutex);
2286 rx_stats.nPeerStructs++;
2287 MUTEX_EXIT(&rx_stats_mutex);
2294 origPeer->refCount--;
2295 MUTEX_EXIT(&rx_peerHashTable_lock);
2300 /* Find the connection at (host, port) started at epoch, and with the
2301 * given connection id. Creates the server connection if necessary.
2302 * The type specifies whether a client connection or a server
2303 * connection is desired. In both cases, (host, port) specify the
2304 * peer's (host, pair) pair. Client connections are not made
2305 * automatically by this routine. The parameter socket gives the
2306 * socket descriptor on which the packet was received. This is used,
2307 * in the case of server connections, to check that *new* connections
2308 * come via a valid (port, serviceId). Finally, the securityIndex
2309 * parameter must match the existing index for the connection. If a
2310 * server connection is created, it will be created using the supplied
2311 * index, if the index is valid for this service */
2312 struct rx_connection *
2313 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2314 register u_short port, u_short serviceId, afs_uint32 cid,
2315 afs_uint32 epoch, int type, u_int securityIndex)
2317 int hashindex, flag;
2318 register struct rx_connection *conn;
2319 hashindex = CONN_HASH(host, port, cid, epoch, type);
2320 MUTEX_ENTER(&rx_connHashTable_lock);
2321 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2322 rx_connHashTable[hashindex],
2325 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2326 && (epoch == conn->epoch)) {
2327 register struct rx_peer *pp = conn->peer;
2328 if (securityIndex != conn->securityIndex) {
2329 /* this isn't supposed to happen, but someone could forge a packet
2330 * like this, and there seems to be some CM bug that makes this
2331 * happen from time to time -- in which case, the fileserver
2333 MUTEX_EXIT(&rx_connHashTable_lock);
2334 return (struct rx_connection *)0;
2336 if (pp->host == host && pp->port == port)
2338 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2340 /* So what happens when it's a callback connection? */
2341 if ( /*type == RX_CLIENT_CONNECTION && */
2342 (conn->epoch & 0x80000000))
2346 /* the connection rxLastConn that was used the last time is not the
2347 ** one we are looking for now. Hence, start searching in the hash */
2349 conn = rx_connHashTable[hashindex];
2354 struct rx_service *service;
2355 if (type == RX_CLIENT_CONNECTION) {
2356 MUTEX_EXIT(&rx_connHashTable_lock);
2357 return (struct rx_connection *)0;
2359 service = rxi_FindService(socket, serviceId);
2360 if (!service || (securityIndex >= service->nSecurityObjects)
2361 || (service->securityObjects[securityIndex] == 0)) {
2362 MUTEX_EXIT(&rx_connHashTable_lock);
2363 return (struct rx_connection *)0;
2365 conn = rxi_AllocConnection(); /* This bzero's the connection */
2366 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2367 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2368 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2369 conn->next = rx_connHashTable[hashindex];
2370 rx_connHashTable[hashindex] = conn;
2371 conn->peer = rxi_FindPeer(host, port, 0, 1);
2372 conn->type = RX_SERVER_CONNECTION;
2373 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2374 conn->epoch = epoch;
2375 conn->cid = cid & RX_CIDMASK;
2376 /* conn->serial = conn->lastSerial = 0; */
2377 /* conn->timeout = 0; */
2378 conn->ackRate = RX_FAST_ACK_RATE;
2379 conn->service = service;
2380 conn->serviceId = serviceId;
2381 conn->securityIndex = securityIndex;
2382 conn->securityObject = service->securityObjects[securityIndex];
2383 conn->nSpecific = 0;
2384 conn->specific = NULL;
2385 rx_SetConnDeadTime(conn, service->connDeadTime);
2386 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2387 /* Notify security object of the new connection */
2388 RXS_NewConnection(conn->securityObject, conn);
2389 /* XXXX Connection timeout? */
2390 if (service->newConnProc)
2391 (*service->newConnProc) (conn);
2392 MUTEX_ENTER(&rx_stats_mutex);
2393 rx_stats.nServerConns++;
2394 MUTEX_EXIT(&rx_stats_mutex);
2397 MUTEX_ENTER(&conn->conn_data_lock);
2399 MUTEX_EXIT(&conn->conn_data_lock);
2401 rxLastConn = conn; /* store this connection as the last conn used */
2402 MUTEX_EXIT(&rx_connHashTable_lock);
2406 /* There are two packet tracing routines available for testing and monitoring
2407 * Rx. One is called just after every packet is received and the other is
2408 * called just before every packet is sent. Received packets, have had their
2409 * headers decoded, and packets to be sent have not yet had their headers
2410 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2411 * containing the network address. Both can be modified. The return value, if
2412 * non-zero, indicates that the packet should be dropped. */
2414 int (*rx_justReceived) () = 0;
2415 int (*rx_almostSent) () = 0;
2417 /* A packet has been received off the interface. Np is the packet, socket is
2418 * the socket number it was received from (useful in determining which service
2419 * this packet corresponds to), and (host, port) reflect the host,port of the
2420 * sender. This call returns the packet to the caller if it is finished with
2421 * it, rather than de-allocating it, just as a small performance hack */
2424 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2425 afs_uint32 host, u_short port, int *tnop,
2426 struct rx_call **newcallp)
2428 register struct rx_call *call;
2429 register struct rx_connection *conn;
2431 afs_uint32 currentCallNumber;
2437 struct rx_packet *tnp;
2440 /* We don't print out the packet until now because (1) the time may not be
2441 * accurate enough until now in the lwp implementation (rx_Listener only gets
2442 * the time after the packet is read) and (2) from a protocol point of view,
2443 * this is the first time the packet has been seen */
2444 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2445 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2446 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2447 np->header.serial, packetType, host, port, np->header.serviceId,
2448 np->header.epoch, np->header.cid, np->header.callNumber,
2449 np->header.seq, np->header.flags, np));
2452 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2453 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2456 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2457 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2460 /* If an input tracer function is defined, call it with the packet and
2461 * network address. Note this function may modify its arguments. */
2462 if (rx_justReceived) {
2463 struct sockaddr_in addr;
2465 addr.sin_family = AF_INET;
2466 addr.sin_port = port;
2467 addr.sin_addr.s_addr = host;
2468 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2469 addr.sin_len = sizeof(addr);
2470 #endif /* AFS_OSF_ENV */
2471 drop = (*rx_justReceived) (np, &addr);
2472 /* drop packet if return value is non-zero */
2475 port = addr.sin_port; /* in case fcn changed addr */
2476 host = addr.sin_addr.s_addr;
2480 /* If packet was not sent by the client, then *we* must be the client */
2481 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2482 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2484 /* Find the connection (or fabricate one, if we're the server & if
2485 * necessary) associated with this packet */
2487 rxi_FindConnection(socket, host, port, np->header.serviceId,
2488 np->header.cid, np->header.epoch, type,
2489 np->header.securityIndex);
2492 /* If no connection found or fabricated, just ignore the packet.
2493 * (An argument could be made for sending an abort packet for
2498 MUTEX_ENTER(&conn->conn_data_lock);
2499 if (conn->maxSerial < np->header.serial)
2500 conn->maxSerial = np->header.serial;
2501 MUTEX_EXIT(&conn->conn_data_lock);
2503 /* If the connection is in an error state, send an abort packet and ignore
2504 * the incoming packet */
2506 /* Don't respond to an abort packet--we don't want loops! */
2507 MUTEX_ENTER(&conn->conn_data_lock);
2508 if (np->header.type != RX_PACKET_TYPE_ABORT)
2509 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2511 MUTEX_EXIT(&conn->conn_data_lock);
2515 /* Check for connection-only requests (i.e. not call specific). */
2516 if (np->header.callNumber == 0) {
2517 switch (np->header.type) {
2518 case RX_PACKET_TYPE_ABORT:
2519 /* What if the supplied error is zero? */
2520 rxi_ConnectionError(conn, ntohl(rx_GetInt32(np, 0)));
2521 MUTEX_ENTER(&conn->conn_data_lock);
2523 MUTEX_EXIT(&conn->conn_data_lock);
2525 case RX_PACKET_TYPE_CHALLENGE:
2526 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2527 MUTEX_ENTER(&conn->conn_data_lock);
2529 MUTEX_EXIT(&conn->conn_data_lock);
2531 case RX_PACKET_TYPE_RESPONSE:
2532 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2533 MUTEX_ENTER(&conn->conn_data_lock);
2535 MUTEX_EXIT(&conn->conn_data_lock);
2537 case RX_PACKET_TYPE_PARAMS:
2538 case RX_PACKET_TYPE_PARAMS + 1:
2539 case RX_PACKET_TYPE_PARAMS + 2:
2540 /* ignore these packet types for now */
2541 MUTEX_ENTER(&conn->conn_data_lock);
2543 MUTEX_EXIT(&conn->conn_data_lock);
2548 /* Should not reach here, unless the peer is broken: send an
2550 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2551 MUTEX_ENTER(&conn->conn_data_lock);
2552 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2554 MUTEX_EXIT(&conn->conn_data_lock);
2559 channel = np->header.cid & RX_CHANNELMASK;
2560 call = conn->call[channel];
2561 #ifdef RX_ENABLE_LOCKS
2563 MUTEX_ENTER(&call->lock);
2564 /* Test to see if call struct is still attached to conn. */
2565 if (call != conn->call[channel]) {
2567 MUTEX_EXIT(&call->lock);
2568 if (type == RX_SERVER_CONNECTION) {
2569 call = conn->call[channel];
2570 /* If we started with no call attached and there is one now,
2571 * another thread is also running this routine and has gotten
2572 * the connection channel. We should drop this packet in the tests
2573 * below. If there was a call on this connection and it's now
2574 * gone, then we'll be making a new call below.
2575 * If there was previously a call and it's now different then
2576 * the old call was freed and another thread running this routine
2577 * has created a call on this channel. One of these two threads
2578 * has a packet for the old call and the code below handles those
2582 MUTEX_ENTER(&call->lock);
2584 /* This packet can't be for this call. If the new call address is
2585 * 0 then no call is running on this channel. If there is a call
2586 * then, since this is a client connection we're getting data for
2587 * it must be for the previous call.
2589 MUTEX_ENTER(&rx_stats_mutex);
2590 rx_stats.spuriousPacketsRead++;
2591 MUTEX_EXIT(&rx_stats_mutex);
2592 MUTEX_ENTER(&conn->conn_data_lock);
2594 MUTEX_EXIT(&conn->conn_data_lock);
2599 currentCallNumber = conn->callNumber[channel];
2601 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2602 if (np->header.callNumber < currentCallNumber) {
2603 MUTEX_ENTER(&rx_stats_mutex);
2604 rx_stats.spuriousPacketsRead++;
2605 MUTEX_EXIT(&rx_stats_mutex);
2606 #ifdef RX_ENABLE_LOCKS
2608 MUTEX_EXIT(&call->lock);
2610 MUTEX_ENTER(&conn->conn_data_lock);
2612 MUTEX_EXIT(&conn->conn_data_lock);
2616 MUTEX_ENTER(&conn->conn_call_lock);
2617 call = rxi_NewCall(conn, channel);
2618 MUTEX_EXIT(&conn->conn_call_lock);
2619 *call->callNumber = np->header.callNumber;
2620 call->state = RX_STATE_PRECALL;
2621 clock_GetTime(&call->queueTime);
2622 hzero(call->bytesSent);
2623 hzero(call->bytesRcvd);
2624 rxi_KeepAliveOn(call);
2625 } else if (np->header.callNumber != currentCallNumber) {
2626 /* Wait until the transmit queue is idle before deciding
2627 * whether to reset the current call. Chances are that the
2628 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2631 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2632 while ((call->state == RX_STATE_ACTIVE)
2633 && (call->flags & RX_CALL_TQ_BUSY)) {
2634 call->flags |= RX_CALL_TQ_WAIT;
2635 #ifdef RX_ENABLE_LOCKS
2636 CV_WAIT(&call->cv_tq, &call->lock);
2637 #else /* RX_ENABLE_LOCKS */
2638 osi_rxSleep(&call->tq);
2639 #endif /* RX_ENABLE_LOCKS */
2641 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2642 /* If the new call cannot be taken right now send a busy and set
2643 * the error condition in this call, so that it terminates as
2644 * quickly as possible */
2645 if (call->state == RX_STATE_ACTIVE) {
2646 struct rx_packet *tp;
2648 rxi_CallError(call, RX_CALL_DEAD);
2649 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2651 MUTEX_EXIT(&call->lock);
2652 MUTEX_ENTER(&conn->conn_data_lock);
2654 MUTEX_EXIT(&conn->conn_data_lock);
2657 rxi_ResetCall(call, 0);
2658 *call->callNumber = np->header.callNumber;
2659 call->state = RX_STATE_PRECALL;
2660 clock_GetTime(&call->queueTime);
2661 hzero(call->bytesSent);
2662 hzero(call->bytesRcvd);
2664 * If the number of queued calls exceeds the overload
2665 * threshold then abort this call.
2667 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2668 struct rx_packet *tp;
2670 rxi_CallError(call, rx_BusyError);
2671 tp = rxi_SendCallAbort(call, np, 1, 0);
2672 MUTEX_EXIT(&call->lock);
2673 MUTEX_ENTER(&conn->conn_data_lock);
2675 MUTEX_EXIT(&conn->conn_data_lock);
2678 rxi_KeepAliveOn(call);
2680 /* Continuing call; do nothing here. */
2682 } else { /* we're the client */
2683 /* Ignore all incoming acknowledgements for calls in DALLY state */
2684 if (call && (call->state == RX_STATE_DALLY)
2685 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2686 MUTEX_ENTER(&rx_stats_mutex);
2687 rx_stats.ignorePacketDally++;
2688 MUTEX_EXIT(&rx_stats_mutex);
2689 #ifdef RX_ENABLE_LOCKS
2691 MUTEX_EXIT(&call->lock);
2694 MUTEX_ENTER(&conn->conn_data_lock);
2696 MUTEX_EXIT(&conn->conn_data_lock);
2700 /* Ignore anything that's not relevant to the current call. If there
2701 * isn't a current call, then no packet is relevant. */
2702 if (!call || (np->header.callNumber != currentCallNumber)) {
2703 MUTEX_ENTER(&rx_stats_mutex);
2704 rx_stats.spuriousPacketsRead++;
2705 MUTEX_EXIT(&rx_stats_mutex);
2706 #ifdef RX_ENABLE_LOCKS
2708 MUTEX_EXIT(&call->lock);
2711 MUTEX_ENTER(&conn->conn_data_lock);
2713 MUTEX_EXIT(&conn->conn_data_lock);
2716 /* If the service security object index stamped in the packet does not
2717 * match the connection's security index, ignore the packet */
2718 if (np->header.securityIndex != conn->securityIndex) {
2719 #ifdef RX_ENABLE_LOCKS
2720 MUTEX_EXIT(&call->lock);
2722 MUTEX_ENTER(&conn->conn_data_lock);
2724 MUTEX_EXIT(&conn->conn_data_lock);
2728 /* If we're receiving the response, then all transmit packets are
2729 * implicitly acknowledged. Get rid of them. */
2730 if (np->header.type == RX_PACKET_TYPE_DATA) {
2731 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2732 /* XXX Hack. Because we must release the global rx lock when
2733 * sending packets (osi_NetSend) we drop all acks while we're
2734 * traversing the tq in rxi_Start sending packets out because
2735 * packets may move to the freePacketQueue as result of being here!
2736 * So we drop these packets until we're safely out of the
2737 * traversing. Really ugly!
2738 * For fine grain RX locking, we set the acked field in the
2739 * packets and let rxi_Start remove them from the transmit queue.
2741 if (call->flags & RX_CALL_TQ_BUSY) {
2742 #ifdef RX_ENABLE_LOCKS
2743 rxi_SetAcksInTransmitQueue(call);
2746 return np; /* xmitting; drop packet */
2749 rxi_ClearTransmitQueue(call, 0);
2751 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2752 rxi_ClearTransmitQueue(call, 0);
2753 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2755 if (np->header.type == RX_PACKET_TYPE_ACK) {
2756 /* now check to see if this is an ack packet acknowledging that the
2757 * server actually *lost* some hard-acked data. If this happens we
2758 * ignore this packet, as it may indicate that the server restarted in
2759 * the middle of a call. It is also possible that this is an old ack
2760 * packet. We don't abort the connection in this case, because this
2761 * *might* just be an old ack packet. The right way to detect a server
2762 * restart in the midst of a call is to notice that the server epoch
2764 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2765 * XXX unacknowledged. I think that this is off-by-one, but
2766 * XXX I don't dare change it just yet, since it will
2767 * XXX interact badly with the server-restart detection
2768 * XXX code in receiveackpacket. */
2769 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2770 MUTEX_ENTER(&rx_stats_mutex);
2771 rx_stats.spuriousPacketsRead++;
2772 MUTEX_EXIT(&rx_stats_mutex);
2773 MUTEX_EXIT(&call->lock);
2774 MUTEX_ENTER(&conn->conn_data_lock);
2776 MUTEX_EXIT(&conn->conn_data_lock);
2780 } /* else not a data packet */
2783 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2784 /* Set remote user defined status from packet */
2785 call->remoteStatus = np->header.userStatus;
2787 /* Note the gap between the expected next packet and the actual
2788 * packet that arrived, when the new packet has a smaller serial number
2789 * than expected. Rioses frequently reorder packets all by themselves,
2790 * so this will be quite important with very large window sizes.
2791 * Skew is checked against 0 here to avoid any dependence on the type of
2792 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2794 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2795 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2796 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2798 MUTEX_ENTER(&conn->conn_data_lock);
2799 skew = conn->lastSerial - np->header.serial;
2800 conn->lastSerial = np->header.serial;
2801 MUTEX_EXIT(&conn->conn_data_lock);
2803 register struct rx_peer *peer;
2805 if (skew > peer->inPacketSkew) {
2806 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2808 peer->inPacketSkew = skew;
2812 /* Now do packet type-specific processing */
2813 switch (np->header.type) {
2814 case RX_PACKET_TYPE_DATA:
2815 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2818 case RX_PACKET_TYPE_ACK:
2819 /* Respond immediately to ack packets requesting acknowledgement
2821 if (np->header.flags & RX_REQUEST_ACK) {
2823 (void)rxi_SendCallAbort(call, 0, 1, 0);
2825 (void)rxi_SendAck(call, 0, np->header.serial,
2826 RX_ACK_PING_RESPONSE, 1);
2828 np = rxi_ReceiveAckPacket(call, np, 1);
2830 case RX_PACKET_TYPE_ABORT:
2831 /* An abort packet: reset the connection, passing the error up to
2833 /* What if error is zero? */
2834 rxi_CallError(call, ntohl(*(afs_int32 *) rx_DataOf(np)));
2836 case RX_PACKET_TYPE_BUSY:
2839 case RX_PACKET_TYPE_ACKALL:
2840 /* All packets acknowledged, so we can drop all packets previously
2841 * readied for sending */
2842 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2843 /* XXX Hack. We because we can't release the global rx lock when
2844 * sending packets (osi_NetSend) we drop all ack pkts while we're
2845 * traversing the tq in rxi_Start sending packets out because
2846 * packets may move to the freePacketQueue as result of being
2847 * here! So we drop these packets until we're safely out of the
2848 * traversing. Really ugly!
2849 * For fine grain RX locking, we set the acked field in the packets
2850 * and let rxi_Start remove the packets from the transmit queue.
2852 if (call->flags & RX_CALL_TQ_BUSY) {
2853 #ifdef RX_ENABLE_LOCKS
2854 rxi_SetAcksInTransmitQueue(call);
2856 #else /* RX_ENABLE_LOCKS */
2858 return np; /* xmitting; drop packet */
2859 #endif /* RX_ENABLE_LOCKS */
2861 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2862 rxi_ClearTransmitQueue(call, 0);
2865 /* Should not reach here, unless the peer is broken: send an abort
2867 rxi_CallError(call, RX_PROTOCOL_ERROR);
2868 np = rxi_SendCallAbort(call, np, 1, 0);
2871 /* Note when this last legitimate packet was received, for keep-alive
2872 * processing. Note, we delay getting the time until now in the hope that
2873 * the packet will be delivered to the user before any get time is required
2874 * (if not, then the time won't actually be re-evaluated here). */
2875 call->lastReceiveTime = clock_Sec();
2876 MUTEX_EXIT(&call->lock);
2877 MUTEX_ENTER(&conn->conn_data_lock);
2879 MUTEX_EXIT(&conn->conn_data_lock);
2883 /* return true if this is an "interesting" connection from the point of view
2884 of someone trying to debug the system */
2886 rxi_IsConnInteresting(struct rx_connection *aconn)
2889 register struct rx_call *tcall;
2891 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2893 for (i = 0; i < RX_MAXCALLS; i++) {
2894 tcall = aconn->call[i];
2896 if ((tcall->state == RX_STATE_PRECALL)
2897 || (tcall->state == RX_STATE_ACTIVE))
2899 if ((tcall->mode == RX_MODE_SENDING)
2900 || (tcall->mode == RX_MODE_RECEIVING))
2908 /* if this is one of the last few packets AND it wouldn't be used by the
2909 receiving call to immediately satisfy a read request, then drop it on
2910 the floor, since accepting it might prevent a lock-holding thread from
2911 making progress in its reading. If a call has been cleared while in
2912 the precall state then ignore all subsequent packets until the call
2913 is assigned to a thread. */
2916 TooLow(struct rx_packet *ap, struct rx_call *acall)
2919 MUTEX_ENTER(&rx_stats_mutex);
2920 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2921 && (acall->state == RX_STATE_PRECALL))
2922 || ((rx_nFreePackets < rxi_dataQuota + 2)
2923 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2924 && (acall->flags & RX_CALL_READER_WAIT)))) {
2927 MUTEX_EXIT(&rx_stats_mutex);
2933 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2934 struct rx_call *acall)
2936 struct rx_call *call = acall;
2940 MUTEX_ENTER(&conn->conn_data_lock);
2941 conn->checkReachEvent = NULL;
2942 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2945 MUTEX_EXIT(&conn->conn_data_lock);
2949 MUTEX_ENTER(&conn->conn_call_lock);
2950 MUTEX_ENTER(&conn->conn_data_lock);
2951 for (i = 0; i < RX_MAXCALLS; i++) {
2952 struct rx_call *tc = conn->call[i];
2953 if (tc && tc->state == RX_STATE_PRECALL) {
2959 /* Indicate that rxi_CheckReachEvent is no longer running by
2960 * clearing the flag. Must be atomic under conn_data_lock to
2961 * avoid a new call slipping by: rxi_CheckConnReach holds
2962 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
2964 conn->flags &= ~RX_CONN_ATTACHWAIT;
2965 MUTEX_EXIT(&conn->conn_data_lock);
2966 MUTEX_EXIT(&conn->conn_call_lock);
2971 MUTEX_ENTER(&call->lock);
2972 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
2974 MUTEX_EXIT(&call->lock);
2976 clock_GetTime(&when);
2977 when.sec += RX_CHECKREACH_TIMEOUT;
2978 MUTEX_ENTER(&conn->conn_data_lock);
2979 if (!conn->checkReachEvent) {
2981 conn->checkReachEvent =
2982 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
2984 MUTEX_EXIT(&conn->conn_data_lock);
2990 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
2992 struct rx_service *service = conn->service;
2993 struct rx_peer *peer = conn->peer;
2994 afs_uint32 now, lastReach;
2996 if (service->checkReach == 0)
3000 MUTEX_ENTER(&peer->peer_lock);
3001 lastReach = peer->lastReachTime;
3002 MUTEX_EXIT(&peer->peer_lock);
3003 if (now - lastReach < RX_CHECKREACH_TTL)
3006 MUTEX_ENTER(&conn->conn_data_lock);
3007 if (conn->flags & RX_CONN_ATTACHWAIT) {
3008 MUTEX_EXIT(&conn->conn_data_lock);
3011 conn->flags |= RX_CONN_ATTACHWAIT;
3012 MUTEX_EXIT(&conn->conn_data_lock);
3013 if (!conn->checkReachEvent)
3014 rxi_CheckReachEvent(NULL, conn, call);
3019 /* try to attach call, if authentication is complete */
3021 TryAttach(register struct rx_call *acall, register osi_socket socket,
3022 register int *tnop, register struct rx_call **newcallp,
3025 struct rx_connection *conn = acall->conn;
3027 if (conn->type == RX_SERVER_CONNECTION
3028 && acall->state == RX_STATE_PRECALL) {
3029 /* Don't attach until we have any req'd. authentication. */
3030 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3031 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3032 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3033 /* Note: this does not necessarily succeed; there
3034 * may not any proc available
3037 rxi_ChallengeOn(acall->conn);
3042 /* A data packet has been received off the interface. This packet is
3043 * appropriate to the call (the call is in the right state, etc.). This
3044 * routine can return a packet to the caller, for re-use */
3047 rxi_ReceiveDataPacket(register struct rx_call *call,
3048 register struct rx_packet *np, int istack,
3049 osi_socket socket, afs_uint32 host, u_short port,
3050 int *tnop, struct rx_call **newcallp)
3052 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3056 afs_uint32 seq, serial, flags;
3058 struct rx_packet *tnp;
3060 MUTEX_ENTER(&rx_stats_mutex);
3061 rx_stats.dataPacketsRead++;
3062 MUTEX_EXIT(&rx_stats_mutex);
3065 /* If there are no packet buffers, drop this new packet, unless we can find
3066 * packet buffers from inactive calls */
3068 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3069 MUTEX_ENTER(&rx_freePktQ_lock);
3070 rxi_NeedMorePackets = TRUE;
3071 MUTEX_EXIT(&rx_freePktQ_lock);
3072 MUTEX_ENTER(&rx_stats_mutex);
3073 rx_stats.noPacketBuffersOnRead++;
3074 MUTEX_EXIT(&rx_stats_mutex);
3075 call->rprev = np->header.serial;
3076 rxi_calltrace(RX_TRACE_DROP, call);
3077 dpf(("packet %x dropped on receipt - quota problems", np));
3079 rxi_ClearReceiveQueue(call);
3080 clock_GetTime(&when);
3081 clock_Add(&when, &rx_softAckDelay);
3082 if (!call->delayedAckEvent
3083 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3084 rxevent_Cancel(call->delayedAckEvent, call,
3085 RX_CALL_REFCOUNT_DELAY);
3086 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3087 call->delayedAckEvent =
3088 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3090 /* we've damaged this call already, might as well do it in. */
3096 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3097 * packet is one of several packets transmitted as a single
3098 * datagram. Do not send any soft or hard acks until all packets
3099 * in a jumbogram have been processed. Send negative acks right away.
3101 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3102 /* tnp is non-null when there are more packets in the
3103 * current jumbo gram */
3110 seq = np->header.seq;
3111 serial = np->header.serial;
3112 flags = np->header.flags;
3114 /* If the call is in an error state, send an abort message */
3116 return rxi_SendCallAbort(call, np, istack, 0);
3118 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3119 * AFS 3.5 jumbogram. */
3120 if (flags & RX_JUMBO_PACKET) {
3121 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3126 if (np->header.spare != 0) {
3127 MUTEX_ENTER(&call->conn->conn_data_lock);
3128 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3129 MUTEX_EXIT(&call->conn->conn_data_lock);
3132 /* The usual case is that this is the expected next packet */
3133 if (seq == call->rnext) {
3135 /* Check to make sure it is not a duplicate of one already queued */
3136 if (queue_IsNotEmpty(&call->rq)
3137 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3138 MUTEX_ENTER(&rx_stats_mutex);
3139 rx_stats.dupPacketsRead++;
3140 MUTEX_EXIT(&rx_stats_mutex);
3141 dpf(("packet %x dropped on receipt - duplicate", np));
3142 rxevent_Cancel(call->delayedAckEvent, call,
3143 RX_CALL_REFCOUNT_DELAY);
3144 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3150 /* It's the next packet. Stick it on the receive queue
3151 * for this call. Set newPackets to make sure we wake
3152 * the reader once all packets have been processed */
3153 queue_Prepend(&call->rq, np);
3155 np = NULL; /* We can't use this anymore */
3158 /* If an ack is requested then set a flag to make sure we
3159 * send an acknowledgement for this packet */
3160 if (flags & RX_REQUEST_ACK) {
3161 ackNeeded = RX_ACK_REQUESTED;
3164 /* Keep track of whether we have received the last packet */
3165 if (flags & RX_LAST_PACKET) {
3166 call->flags |= RX_CALL_HAVE_LAST;
3170 /* Check whether we have all of the packets for this call */
3171 if (call->flags & RX_CALL_HAVE_LAST) {
3172 afs_uint32 tseq; /* temporary sequence number */
3173 struct rx_packet *tp; /* Temporary packet pointer */
3174 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3176 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3177 if (tseq != tp->header.seq)
3179 if (tp->header.flags & RX_LAST_PACKET) {
3180 call->flags |= RX_CALL_RECEIVE_DONE;
3187 /* Provide asynchronous notification for those who want it
3188 * (e.g. multi rx) */
3189 if (call->arrivalProc) {
3190 (*call->arrivalProc) (call, call->arrivalProcHandle,
3191 (int)call->arrivalProcArg);
3192 call->arrivalProc = (VOID(*)())0;
3195 /* Update last packet received */
3198 /* If there is no server process serving this call, grab
3199 * one, if available. We only need to do this once. If a
3200 * server thread is available, this thread becomes a server
3201 * thread and the server thread becomes a listener thread. */
3203 TryAttach(call, socket, tnop, newcallp, 0);
3206 /* This is not the expected next packet. */
3208 /* Determine whether this is a new or old packet, and if it's
3209 * a new one, whether it fits into the current receive window.
3210 * Also figure out whether the packet was delivered in sequence.
3211 * We use the prev variable to determine whether the new packet
3212 * is the successor of its immediate predecessor in the
3213 * receive queue, and the missing flag to determine whether
3214 * any of this packets predecessors are missing. */
3216 afs_uint32 prev; /* "Previous packet" sequence number */
3217 struct rx_packet *tp; /* Temporary packet pointer */
3218 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3219 int missing; /* Are any predecessors missing? */
3221 /* If the new packet's sequence number has been sent to the
3222 * application already, then this is a duplicate */
3223 if (seq < call->rnext) {
3224 MUTEX_ENTER(&rx_stats_mutex);
3225 rx_stats.dupPacketsRead++;
3226 MUTEX_EXIT(&rx_stats_mutex);
3227 rxevent_Cancel(call->delayedAckEvent, call,
3228 RX_CALL_REFCOUNT_DELAY);
3229 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3235 /* If the sequence number is greater than what can be
3236 * accomodated by the current window, then send a negative
3237 * acknowledge and drop the packet */
3238 if ((call->rnext + call->rwind) <= seq) {
3239 rxevent_Cancel(call->delayedAckEvent, call,
3240 RX_CALL_REFCOUNT_DELAY);
3241 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3248 /* Look for the packet in the queue of old received packets */
3249 for (prev = call->rnext - 1, missing =
3250 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3251 /*Check for duplicate packet */
3252 if (seq == tp->header.seq) {
3253 MUTEX_ENTER(&rx_stats_mutex);
3254 rx_stats.dupPacketsRead++;
3255 MUTEX_EXIT(&rx_stats_mutex);
3256 rxevent_Cancel(call->delayedAckEvent, call,
3257 RX_CALL_REFCOUNT_DELAY);
3258 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3264 /* If we find a higher sequence packet, break out and
3265 * insert the new packet here. */
3266 if (seq < tp->header.seq)
3268 /* Check for missing packet */
3269 if (tp->header.seq != prev + 1) {
3273 prev = tp->header.seq;
3276 /* Keep track of whether we have received the last packet. */
3277 if (flags & RX_LAST_PACKET) {
3278 call->flags |= RX_CALL_HAVE_LAST;
3281 /* It's within the window: add it to the the receive queue.
3282 * tp is left by the previous loop either pointing at the
3283 * packet before which to insert the new packet, or at the
3284 * queue head if the queue is empty or the packet should be
3286 queue_InsertBefore(tp, np);
3290 /* Check whether we have all of the packets for this call */
3291 if ((call->flags & RX_CALL_HAVE_LAST)
3292 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3293 afs_uint32 tseq; /* temporary sequence number */
3296 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3297 if (tseq != tp->header.seq)
3299 if (tp->header.flags & RX_LAST_PACKET) {
3300 call->flags |= RX_CALL_RECEIVE_DONE;
3307 /* We need to send an ack of the packet is out of sequence,
3308 * or if an ack was requested by the peer. */
3309 if (seq != prev + 1 || missing || (flags & RX_REQUEST_ACK)) {
3310 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3313 /* Acknowledge the last packet for each call */
3314 if (flags & RX_LAST_PACKET) {
3325 * If the receiver is waiting for an iovec, fill the iovec
3326 * using the data from the receive queue */
3327 if (call->flags & RX_CALL_IOVEC_WAIT) {
3328 didHardAck = rxi_FillReadVec(call, serial);
3329 /* the call may have been aborted */
3338 /* Wakeup the reader if any */
3339 if ((call->flags & RX_CALL_READER_WAIT)
3340 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3341 || (call->iovNext >= call->iovMax)
3342 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3343 call->flags &= ~RX_CALL_READER_WAIT;
3344 #ifdef RX_ENABLE_LOCKS
3345 CV_BROADCAST(&call->cv_rq);
3347 osi_rxWakeup(&call->rq);
3353 * Send an ack when requested by the peer, or once every
3354 * rxi_SoftAckRate packets until the last packet has been
3355 * received. Always send a soft ack for the last packet in
3356 * the server's reply. */
3358 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3359 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3360 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3361 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3362 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3363 } else if (call->nSoftAcks) {
3364 clock_GetTime(&when);
3365 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3366 clock_Add(&when, &rx_lastAckDelay);
3368 clock_Add(&when, &rx_softAckDelay);
3370 if (!call->delayedAckEvent
3371 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3372 rxevent_Cancel(call->delayedAckEvent, call,
3373 RX_CALL_REFCOUNT_DELAY);
3374 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3375 call->delayedAckEvent =
3376 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3378 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3379 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3386 static void rxi_ComputeRate();
3390 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3392 struct rx_peer *peer = conn->peer;
3394 MUTEX_ENTER(&peer->peer_lock);
3395 peer->lastReachTime = clock_Sec();
3396 MUTEX_EXIT(&peer->peer_lock);
3398 MUTEX_ENTER(&conn->conn_data_lock);
3399 if (conn->flags & RX_CONN_ATTACHWAIT) {
3402 conn->flags &= ~RX_CONN_ATTACHWAIT;
3403 MUTEX_EXIT(&conn->conn_data_lock);
3405 for (i = 0; i < RX_MAXCALLS; i++) {
3406 struct rx_call *call = conn->call[i];
3409 MUTEX_ENTER(&call->lock);
3410 /* tnop can be null if newcallp is null */
3411 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3413 MUTEX_EXIT(&call->lock);
3417 MUTEX_EXIT(&conn->conn_data_lock);
3420 /* rxi_ComputePeerNetStats
3422 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3423 * estimates (like RTT and throughput) based on ack packets. Caller
3424 * must ensure that the packet in question is the right one (i.e.
3425 * serial number matches).
3428 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3429 struct rx_ackPacket *ap, struct rx_packet *np)
3431 struct rx_peer *peer = call->conn->peer;
3433 /* Use RTT if not delayed by client. */
3434 if (ap->reason != RX_ACK_DELAY)
3435 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3437 rxi_ComputeRate(peer, call, p, np, ap->reason);
3441 /* The real smarts of the whole thing. */
3443 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3446 struct rx_ackPacket *ap;
3448 register struct rx_packet *tp;
3449 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3450 register struct rx_connection *conn = call->conn;
3451 struct rx_peer *peer = conn->peer;
3454 /* because there are CM's that are bogus, sending weird values for this. */
3455 afs_uint32 skew = 0;
3460 int newAckCount = 0;
3461 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3462 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3464 MUTEX_ENTER(&rx_stats_mutex);
3465 rx_stats.ackPacketsRead++;
3466 MUTEX_EXIT(&rx_stats_mutex);
3467 ap = (struct rx_ackPacket *)rx_DataOf(np);
3468 nbytes = rx_Contiguous(np) - ((ap->acks) - (u_char *) ap);
3470 return np; /* truncated ack packet */
3472 /* depends on ack packet struct */
3473 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3474 first = ntohl(ap->firstPacket);
3475 serial = ntohl(ap->serial);
3476 /* temporarily disabled -- needs to degrade over time
3477 * skew = ntohs(ap->maxSkew); */
3479 /* Ignore ack packets received out of order */
3480 if (first < call->tfirst) {
3484 if (np->header.flags & RX_SLOW_START_OK) {
3485 call->flags |= RX_CALL_SLOW_START_OK;
3488 if (ap->reason == RX_ACK_PING_RESPONSE)
3489 rxi_UpdatePeerReach(conn, call);
3494 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3495 ap->reason, ntohl(ap->previousPacket),
3496 (unsigned int)np->header.seq, (unsigned int)serial,
3497 (unsigned int)skew, ntohl(ap->firstPacket));
3500 for (offset = 0; offset < nAcks; offset++)
3501 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3508 /* Update the outgoing packet skew value to the latest value of
3509 * the peer's incoming packet skew value. The ack packet, of
3510 * course, could arrive out of order, but that won't affect things
3512 MUTEX_ENTER(&peer->peer_lock);
3513 peer->outPacketSkew = skew;
3515 /* Check for packets that no longer need to be transmitted, and
3516 * discard them. This only applies to packets positively
3517 * acknowledged as having been sent to the peer's upper level.
3518 * All other packets must be retained. So only packets with
3519 * sequence numbers < ap->firstPacket are candidates. */
3520 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3521 if (tp->header.seq >= first)
3523 call->tfirst = tp->header.seq + 1;
3525 && (tp->header.serial == serial || tp->firstSerial == serial))
3526 rxi_ComputePeerNetStats(call, tp, ap, np);
3527 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3528 /* XXX Hack. Because we have to release the global rx lock when sending
3529 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3530 * in rxi_Start sending packets out because packets may move to the
3531 * freePacketQueue as result of being here! So we drop these packets until
3532 * we're safely out of the traversing. Really ugly!
3533 * To make it even uglier, if we're using fine grain locking, we can
3534 * set the ack bits in the packets and have rxi_Start remove the packets
3535 * when it's done transmitting.
3537 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3540 if (call->flags & RX_CALL_TQ_BUSY) {
3541 #ifdef RX_ENABLE_LOCKS
3542 tp->flags |= RX_PKTFLAG_ACKED;
3543 call->flags |= RX_CALL_TQ_SOME_ACKED;
3544 #else /* RX_ENABLE_LOCKS */
3546 #endif /* RX_ENABLE_LOCKS */
3548 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3551 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3556 /* Give rate detector a chance to respond to ping requests */
3557 if (ap->reason == RX_ACK_PING_RESPONSE) {
3558 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3562 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3564 /* Now go through explicit acks/nacks and record the results in
3565 * the waiting packets. These are packets that can't be released
3566 * yet, even with a positive acknowledge. This positive
3567 * acknowledge only means the packet has been received by the
3568 * peer, not that it will be retained long enough to be sent to
3569 * the peer's upper level. In addition, reset the transmit timers
3570 * of any missing packets (those packets that must be missing
3571 * because this packet was out of sequence) */
3573 call->nSoftAcked = 0;
3574 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3575 /* Update round trip time if the ack was stimulated on receipt
3577 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3578 #ifdef RX_ENABLE_LOCKS
3579 if (tp->header.seq >= first)
3580 #endif /* RX_ENABLE_LOCKS */
3581 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3583 && (tp->header.serial == serial || tp->firstSerial == serial))
3584 rxi_ComputePeerNetStats(call, tp, ap, np);
3586 /* Set the acknowledge flag per packet based on the
3587 * information in the ack packet. An acknowlegded packet can
3588 * be downgraded when the server has discarded a packet it
3589 * soacked previously, or when an ack packet is received
3590 * out of sequence. */
3591 if (tp->header.seq < first) {
3592 /* Implicit ack information */
3593 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3596 tp->flags |= RX_PKTFLAG_ACKED;
3597 } else if (tp->header.seq < first + nAcks) {
3598 /* Explicit ack information: set it in the packet appropriately */
3599 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3600 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3602 tp->flags |= RX_PKTFLAG_ACKED;
3610 tp->flags &= ~RX_PKTFLAG_ACKED;
3614 tp->flags &= ~RX_PKTFLAG_ACKED;
3618 /* If packet isn't yet acked, and it has been transmitted at least
3619 * once, reset retransmit time using latest timeout
3620 * ie, this should readjust the retransmit timer for all outstanding
3621 * packets... So we don't just retransmit when we should know better*/
3623 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3624 tp->retryTime = tp->timeSent;
3625 clock_Add(&tp->retryTime, &peer->timeout);
3626 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3627 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3631 /* If the window has been extended by this acknowledge packet,
3632 * then wakeup a sender waiting in alloc for window space, or try
3633 * sending packets now, if he's been sitting on packets due to
3634 * lack of window space */
3635 if (call->tnext < (call->tfirst + call->twind)) {
3636 #ifdef RX_ENABLE_LOCKS
3637 CV_SIGNAL(&call->cv_twind);
3639 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3640 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3641 osi_rxWakeup(&call->twind);
3644 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3645 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3649 /* if the ack packet has a receivelen field hanging off it,
3650 * update our state */
3651 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3654 /* If the ack packet has a "recommended" size that is less than
3655 * what I am using now, reduce my size to match */
3656 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3657 sizeof(afs_int32), &tSize);
3658 tSize = (afs_uint32) ntohl(tSize);
3659 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3661 /* Get the maximum packet size to send to this peer */
3662 rx_packetread(np, rx_AckDataSize(ap->nAcks), sizeof(afs_int32),
3664 tSize = (afs_uint32) ntohl(tSize);
3665 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3666 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3668 /* sanity check - peer might have restarted with different params.
3669 * If peer says "send less", dammit, send less... Peer should never
3670 * be unable to accept packets of the size that prior AFS versions would
3671 * send without asking. */
3672 if (peer->maxMTU != tSize) {
3673 peer->maxMTU = tSize;
3674 peer->MTU = MIN(tSize, peer->MTU);
3675 call->MTU = MIN(call->MTU, tSize);
3679 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3682 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3683 sizeof(afs_int32), &tSize);
3684 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3685 if (tSize < call->twind) { /* smaller than our send */
3686 call->twind = tSize; /* window, we must send less... */
3687 call->ssthresh = MIN(call->twind, call->ssthresh);
3690 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3691 * network MTU confused with the loopback MTU. Calculate the
3692 * maximum MTU here for use in the slow start code below.
3694 maxMTU = peer->maxMTU;
3695 /* Did peer restart with older RX version? */
3696 if (peer->maxDgramPackets > 1) {
3697 peer->maxDgramPackets = 1;
3699 } else if (np->length >=
3700 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3703 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3704 sizeof(afs_int32), &tSize);
3705 tSize = (afs_uint32) ntohl(tSize);
3707 * As of AFS 3.5 we set the send window to match the receive window.
3709 if (tSize < call->twind) {
3710 call->twind = tSize;
3711 call->ssthresh = MIN(call->twind, call->ssthresh);
3712 } else if (tSize > call->twind) {
3713 call->twind = tSize;
3717 * As of AFS 3.5, a jumbogram is more than one fixed size
3718 * packet transmitted in a single UDP datagram. If the remote
3719 * MTU is smaller than our local MTU then never send a datagram
3720 * larger than the natural MTU.
3723 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3724 sizeof(afs_int32), &tSize);
3725 maxDgramPackets = (afs_uint32) ntohl(tSize);
3726 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3728 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3729 maxDgramPackets = MIN(maxDgramPackets, tSize);
3730 if (maxDgramPackets > 1) {
3731 peer->maxDgramPackets = maxDgramPackets;
3732 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3734 peer->maxDgramPackets = 1;
3735 call->MTU = peer->natMTU;
3737 } else if (peer->maxDgramPackets > 1) {
3738 /* Restarted with lower version of RX */
3739 peer->maxDgramPackets = 1;
3741 } else if (peer->maxDgramPackets > 1
3742 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3743 /* Restarted with lower version of RX */
3744 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3745 peer->natMTU = OLD_MAX_PACKET_SIZE;
3746 peer->MTU = OLD_MAX_PACKET_SIZE;
3747 peer->maxDgramPackets = 1;
3748 peer->nDgramPackets = 1;
3750 call->MTU = OLD_MAX_PACKET_SIZE;
3755 * Calculate how many datagrams were successfully received after
3756 * the first missing packet and adjust the negative ack counter
3761 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3762 if (call->nNacks < nNacked) {
3763 call->nNacks = nNacked;
3772 if (call->flags & RX_CALL_FAST_RECOVER) {
3774 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3776 call->flags &= ~RX_CALL_FAST_RECOVER;
3777 call->cwind = call->nextCwind;
3778 call->nextCwind = 0;
3781 call->nCwindAcks = 0;
3782 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3783 /* Three negative acks in a row trigger congestion recovery */
3784 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3785 MUTEX_EXIT(&peer->peer_lock);
3786 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3787 /* someone else is waiting to start recovery */
3790 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3791 while (call->flags & RX_CALL_TQ_BUSY) {
3792 call->flags |= RX_CALL_TQ_WAIT;
3793 #ifdef RX_ENABLE_LOCKS
3794 CV_WAIT(&call->cv_tq, &call->lock);
3795 #else /* RX_ENABLE_LOCKS */
3796 osi_rxSleep(&call->tq);
3797 #endif /* RX_ENABLE_LOCKS */
3799 MUTEX_ENTER(&peer->peer_lock);
3800 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3801 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3802 call->flags |= RX_CALL_FAST_RECOVER;
3803 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3805 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3806 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3807 call->nextCwind = call->ssthresh;
3810 peer->MTU = call->MTU;
3811 peer->cwind = call->nextCwind;
3812 peer->nDgramPackets = call->nDgramPackets;
3814 call->congestSeq = peer->congestSeq;
3815 /* Reset the resend times on the packets that were nacked
3816 * so we will retransmit as soon as the window permits*/
3817 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3819 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3820 clock_Zero(&tp->retryTime);
3822 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3827 /* If cwind is smaller than ssthresh, then increase
3828 * the window one packet for each ack we receive (exponential
3830 * If cwind is greater than or equal to ssthresh then increase
3831 * the congestion window by one packet for each cwind acks we
3832 * receive (linear growth). */
3833 if (call->cwind < call->ssthresh) {
3835 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3836 call->nCwindAcks = 0;
3838 call->nCwindAcks += newAckCount;
3839 if (call->nCwindAcks >= call->cwind) {
3840 call->nCwindAcks = 0;
3841 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3845 * If we have received several acknowledgements in a row then
3846 * it is time to increase the size of our datagrams
3848 if ((int)call->nAcks > rx_nDgramThreshold) {
3849 if (peer->maxDgramPackets > 1) {
3850 if (call->nDgramPackets < peer->maxDgramPackets) {
3851 call->nDgramPackets++;
3853 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3854 } else if (call->MTU < peer->maxMTU) {
3855 call->MTU += peer->natMTU;
3856 call->MTU = MIN(call->MTU, peer->maxMTU);
3862 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3864 /* Servers need to hold the call until all response packets have
3865 * been acknowledged. Soft acks are good enough since clients
3866 * are not allowed to clear their receive queues. */
3867 if (call->state == RX_STATE_HOLD
3868 && call->tfirst + call->nSoftAcked >= call->tnext) {
3869 call->state = RX_STATE_DALLY;
3870 rxi_ClearTransmitQueue(call, 0);
3871 } else if (!queue_IsEmpty(&call->tq)) {
3872 rxi_Start(0, call, istack);
3877 /* Received a response to a challenge packet */
3879 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3880 register struct rx_packet *np, int istack)
3884 /* Ignore the packet if we're the client */
3885 if (conn->type == RX_CLIENT_CONNECTION)
3888 /* If already authenticated, ignore the packet (it's probably a retry) */
3889 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3892 /* Otherwise, have the security object evaluate the response packet */
3893 error = RXS_CheckResponse(conn->securityObject, conn, np);
3895 /* If the response is invalid, reset the connection, sending
3896 * an abort to the peer */
3900 rxi_ConnectionError(conn, error);
3901 MUTEX_ENTER(&conn->conn_data_lock);
3902 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3903 MUTEX_EXIT(&conn->conn_data_lock);
3906 /* If the response is valid, any calls waiting to attach
3907 * servers can now do so */
3910 for (i = 0; i < RX_MAXCALLS; i++) {
3911 struct rx_call *call = conn->call[i];
3913 MUTEX_ENTER(&call->lock);
3914 if (call->state == RX_STATE_PRECALL)
3915 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
3916 /* tnop can be null if newcallp is null */
3917 MUTEX_EXIT(&call->lock);
3921 /* Update the peer reachability information, just in case
3922 * some calls went into attach-wait while we were waiting
3923 * for authentication..
3925 rxi_UpdatePeerReach(conn, NULL);
3930 /* A client has received an authentication challenge: the security
3931 * object is asked to cough up a respectable response packet to send
3932 * back to the server. The server is responsible for retrying the
3933 * challenge if it fails to get a response. */
3936 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
3937 register struct rx_packet *np, int istack)
3941 /* Ignore the challenge if we're the server */
3942 if (conn->type == RX_SERVER_CONNECTION)
3945 /* Ignore the challenge if the connection is otherwise idle; someone's
3946 * trying to use us as an oracle. */
3947 if (!rxi_HasActiveCalls(conn))
3950 /* Send the security object the challenge packet. It is expected to fill
3951 * in the response. */
3952 error = RXS_GetResponse(conn->securityObject, conn, np);
3954 /* If the security object is unable to return a valid response, reset the
3955 * connection and send an abort to the peer. Otherwise send the response
3956 * packet to the peer connection. */
3958 rxi_ConnectionError(conn, error);
3959 MUTEX_ENTER(&conn->conn_data_lock);
3960 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3961 MUTEX_EXIT(&conn->conn_data_lock);
3963 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
3964 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
3970 /* Find an available server process to service the current request in
3971 * the given call structure. If one isn't available, queue up this
3972 * call so it eventually gets one */
3974 rxi_AttachServerProc(register struct rx_call *call,
3975 register osi_socket socket, register int *tnop,
3976 register struct rx_call **newcallp)
3978 register struct rx_serverQueueEntry *sq;
3979 register struct rx_service *service = call->conn->service;
3980 register int haveQuota = 0;
3982 /* May already be attached */
3983 if (call->state == RX_STATE_ACTIVE)
3986 MUTEX_ENTER(&rx_serverPool_lock);
3988 haveQuota = QuotaOK(service);
3989 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
3990 /* If there are no processes available to service this call,
3991 * put the call on the incoming call queue (unless it's
3992 * already on the queue).
3994 #ifdef RX_ENABLE_LOCKS
3996 ReturnToServerPool(service);
3997 #endif /* RX_ENABLE_LOCKS */
3999 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4000 call->flags |= RX_CALL_WAIT_PROC;
4001 MUTEX_ENTER(&rx_stats_mutex);
4004 MUTEX_EXIT(&rx_stats_mutex);
4005 rxi_calltrace(RX_CALL_ARRIVAL, call);
4006 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4007 queue_Append(&rx_incomingCallQueue, call);
4010 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4012 /* If hot threads are enabled, and both newcallp and sq->socketp
4013 * are non-null, then this thread will process the call, and the
4014 * idle server thread will start listening on this threads socket.
4017 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4020 *sq->socketp = socket;
4021 clock_GetTime(&call->startTime);
4022 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4026 if (call->flags & RX_CALL_WAIT_PROC) {
4027 /* Conservative: I don't think this should happen */
4028 call->flags &= ~RX_CALL_WAIT_PROC;
4029 if (queue_IsOnQueue(call)) {
4031 MUTEX_ENTER(&rx_stats_mutex);
4033 MUTEX_EXIT(&rx_stats_mutex);
4036 call->state = RX_STATE_ACTIVE;
4037 call->mode = RX_MODE_RECEIVING;
4038 #ifdef RX_KERNEL_TRACE
4040 int glockOwner = ISAFS_GLOCK();
4043 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4044 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4050 if (call->flags & RX_CALL_CLEARED) {
4051 /* send an ack now to start the packet flow up again */
4052 call->flags &= ~RX_CALL_CLEARED;
4053 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4055 #ifdef RX_ENABLE_LOCKS
4058 service->nRequestsRunning++;
4059 if (service->nRequestsRunning <= service->minProcs)
4065 MUTEX_EXIT(&rx_serverPool_lock);
4068 /* Delay the sending of an acknowledge event for a short while, while
4069 * a new call is being prepared (in the case of a client) or a reply
4070 * is being prepared (in the case of a server). Rather than sending
4071 * an ack packet, an ACKALL packet is sent. */
4073 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4075 #ifdef RX_ENABLE_LOCKS
4077 MUTEX_ENTER(&call->lock);
4078 call->delayedAckEvent = NULL;
4079 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4081 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4082 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4084 MUTEX_EXIT(&call->lock);
4085 #else /* RX_ENABLE_LOCKS */
4087 call->delayedAckEvent = NULL;
4088 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4089 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4090 #endif /* RX_ENABLE_LOCKS */
4094 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4097 #ifdef RX_ENABLE_LOCKS
4099 MUTEX_ENTER(&call->lock);
4100 if (event == call->delayedAckEvent)
4101 call->delayedAckEvent = NULL;
4102 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4104 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4106 MUTEX_EXIT(&call->lock);
4107 #else /* RX_ENABLE_LOCKS */
4109 call->delayedAckEvent = NULL;
4110 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4111 #endif /* RX_ENABLE_LOCKS */
4115 #ifdef RX_ENABLE_LOCKS
4116 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4117 * clearing them out.
4120 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4122 register struct rx_packet *p, *tp;
4125 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4128 p->flags |= RX_PKTFLAG_ACKED;
4132 call->flags |= RX_CALL_TQ_CLEARME;
4133 call->flags |= RX_CALL_TQ_SOME_ACKED;
4136 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4137 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4138 call->tfirst = call->tnext;
4139 call->nSoftAcked = 0;
4141 if (call->flags & RX_CALL_FAST_RECOVER) {
4142 call->flags &= ~RX_CALL_FAST_RECOVER;
4143 call->cwind = call->nextCwind;
4144 call->nextCwind = 0;
4147 CV_SIGNAL(&call->cv_twind);
4149 #endif /* RX_ENABLE_LOCKS */
4151 /* Clear out the transmit queue for the current call (all packets have
4152 * been received by peer) */
4154 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4156 register struct rx_packet *p, *tp;
4158 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4159 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4161 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4164 p->flags |= RX_PKTFLAG_ACKED;
4168 call->flags |= RX_CALL_TQ_CLEARME;
4169 call->flags |= RX_CALL_TQ_SOME_ACKED;
4172 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4173 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4179 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4180 call->flags &= ~RX_CALL_TQ_CLEARME;
4182 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4184 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4185 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4186 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4187 call->nSoftAcked = 0;
4189 if (call->flags & RX_CALL_FAST_RECOVER) {
4190 call->flags &= ~RX_CALL_FAST_RECOVER;
4191 call->cwind = call->nextCwind;
4193 #ifdef RX_ENABLE_LOCKS
4194 CV_SIGNAL(&call->cv_twind);
4196 osi_rxWakeup(&call->twind);
4201 rxi_ClearReceiveQueue(register struct rx_call *call)
4203 register struct rx_packet *p, *tp;
4204 if (queue_IsNotEmpty(&call->rq)) {
4205 for (queue_Scan(&call->rq, p, tp, rx_packet)) {
4210 rx_packetReclaims++;
4212 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4214 if (call->state == RX_STATE_PRECALL) {
4215 call->flags |= RX_CALL_CLEARED;
4219 /* Send an abort packet for the specified call */
4221 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4222 int istack, int force)
4230 /* Clients should never delay abort messages */
4231 if (rx_IsClientConn(call->conn))
4234 if (call->abortCode != call->error) {
4235 call->abortCode = call->error;
4236 call->abortCount = 0;
4239 if (force || rxi_callAbortThreshhold == 0
4240 || call->abortCount < rxi_callAbortThreshhold) {
4241 if (call->delayedAbortEvent) {
4242 rxevent_Cancel(call->delayedAbortEvent, call,
4243 RX_CALL_REFCOUNT_ABORT);
4245 error = htonl(call->error);
4248 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4249 (char *)&error, sizeof(error), istack);
4250 } else if (!call->delayedAbortEvent) {
4251 clock_GetTime(&when);
4252 clock_Addmsec(&when, rxi_callAbortDelay);
4253 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4254 call->delayedAbortEvent =
4255 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4260 /* Send an abort packet for the specified connection. Packet is an
4261 * optional pointer to a packet that can be used to send the abort.
4262 * Once the number of abort messages reaches the threshhold, an
4263 * event is scheduled to send the abort. Setting the force flag
4264 * overrides sending delayed abort messages.
4266 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4267 * to send the abort packet.
4270 rxi_SendConnectionAbort(register struct rx_connection *conn,
4271 struct rx_packet *packet, int istack, int force)
4279 /* Clients should never delay abort messages */
4280 if (rx_IsClientConn(conn))
4283 if (force || rxi_connAbortThreshhold == 0
4284 || conn->abortCount < rxi_connAbortThreshhold) {
4285 if (conn->delayedAbortEvent) {
4286 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4288 error = htonl(conn->error);
4290 MUTEX_EXIT(&conn->conn_data_lock);
4292 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4293 RX_PACKET_TYPE_ABORT, (char *)&error,
4294 sizeof(error), istack);
4295 MUTEX_ENTER(&conn->conn_data_lock);
4296 } else if (!conn->delayedAbortEvent) {
4297 clock_GetTime(&when);
4298 clock_Addmsec(&when, rxi_connAbortDelay);
4299 conn->delayedAbortEvent =
4300 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4305 /* Associate an error all of the calls owned by a connection. Called
4306 * with error non-zero. This is only for really fatal things, like
4307 * bad authentication responses. The connection itself is set in
4308 * error at this point, so that future packets received will be
4311 rxi_ConnectionError(register struct rx_connection *conn,
4312 register afs_int32 error)
4316 MUTEX_ENTER(&conn->conn_data_lock);
4317 if (conn->challengeEvent)
4318 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4319 if (conn->checkReachEvent) {
4320 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4321 conn->checkReachEvent = 0;
4322 conn->flags &= ~RX_CONN_ATTACHWAIT;
4325 MUTEX_EXIT(&conn->conn_data_lock);
4326 for (i = 0; i < RX_MAXCALLS; i++) {
4327 struct rx_call *call = conn->call[i];
4329 MUTEX_ENTER(&call->lock);
4330 rxi_CallError(call, error);
4331 MUTEX_EXIT(&call->lock);
4334 conn->error = error;
4335 MUTEX_ENTER(&rx_stats_mutex);
4336 rx_stats.fatalErrors++;
4337 MUTEX_EXIT(&rx_stats_mutex);
4342 rxi_CallError(register struct rx_call *call, afs_int32 error)
4345 error = call->error;
4346 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4347 if (!(call->flags & RX_CALL_TQ_BUSY)) {
4348 rxi_ResetCall(call, 0);
4351 rxi_ResetCall(call, 0);
4353 call->error = error;
4354 call->mode = RX_MODE_ERROR;
4357 /* Reset various fields in a call structure, and wakeup waiting
4358 * processes. Some fields aren't changed: state & mode are not
4359 * touched (these must be set by the caller), and bufptr, nLeft, and
4360 * nFree are not reset, since these fields are manipulated by
4361 * unprotected macros, and may only be reset by non-interrupting code.
4364 /* this code requires that call->conn be set properly as a pre-condition. */
4365 #endif /* ADAPT_WINDOW */
4368 rxi_ResetCall(register struct rx_call *call, register int newcall)
4371 register struct rx_peer *peer;
4372 struct rx_packet *packet;
4374 /* Notify anyone who is waiting for asynchronous packet arrival */
4375 if (call->arrivalProc) {
4376 (*call->arrivalProc) (call, call->arrivalProcHandle,
4377 (int)call->arrivalProcArg);
4378 call->arrivalProc = (VOID(*)())0;
4381 if (call->delayedAbortEvent) {
4382 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4383 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4385 rxi_SendCallAbort(call, packet, 0, 1);
4386 rxi_FreePacket(packet);
4391 * Update the peer with the congestion information in this call
4392 * so other calls on this connection can pick up where this call
4393 * left off. If the congestion sequence numbers don't match then
4394 * another call experienced a retransmission.
4396 peer = call->conn->peer;
4397 MUTEX_ENTER(&peer->peer_lock);
4399 if (call->congestSeq == peer->congestSeq) {
4400 peer->cwind = MAX(peer->cwind, call->cwind);
4401 peer->MTU = MAX(peer->MTU, call->MTU);
4402 peer->nDgramPackets =
4403 MAX(peer->nDgramPackets, call->nDgramPackets);
4406 call->abortCode = 0;
4407 call->abortCount = 0;
4409 if (peer->maxDgramPackets > 1) {
4410 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4412 call->MTU = peer->MTU;
4414 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4415 call->ssthresh = rx_maxSendWindow;
4416 call->nDgramPackets = peer->nDgramPackets;
4417 call->congestSeq = peer->congestSeq;
4418 MUTEX_EXIT(&peer->peer_lock);
4420 flags = call->flags;
4421 rxi_ClearReceiveQueue(call);
4422 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4423 if (call->flags & RX_CALL_TQ_BUSY) {
4424 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4425 call->flags |= (flags & RX_CALL_TQ_WAIT);
4427 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4429 rxi_ClearTransmitQueue(call, 0);
4430 queue_Init(&call->tq);
4433 queue_Init(&call->rq);
4435 call->rwind = rx_initReceiveWindow;
4436 call->twind = rx_initSendWindow;
4437 call->nSoftAcked = 0;
4438 call->nextCwind = 0;
4441 call->nCwindAcks = 0;
4442 call->nSoftAcks = 0;
4443 call->nHardAcks = 0;
4445 call->tfirst = call->rnext = call->tnext = 1;
4447 call->lastAcked = 0;
4448 call->localStatus = call->remoteStatus = 0;
4450 if (flags & RX_CALL_READER_WAIT) {
4451 #ifdef RX_ENABLE_LOCKS
4452 CV_BROADCAST(&call->cv_rq);
4454 osi_rxWakeup(&call->rq);
4457 if (flags & RX_CALL_WAIT_PACKETS) {
4458 MUTEX_ENTER(&rx_freePktQ_lock);
4459 rxi_PacketsUnWait(); /* XXX */
4460 MUTEX_EXIT(&rx_freePktQ_lock);
4462 #ifdef RX_ENABLE_LOCKS
4463 CV_SIGNAL(&call->cv_twind);
4465 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4466 osi_rxWakeup(&call->twind);
4469 #ifdef RX_ENABLE_LOCKS
4470 /* The following ensures that we don't mess with any queue while some
4471 * other thread might also be doing so. The call_queue_lock field is
4472 * is only modified under the call lock. If the call is in the process
4473 * of being removed from a queue, the call is not locked until the
4474 * the queue lock is dropped and only then is the call_queue_lock field
4475 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4476 * Note that any other routine which removes a call from a queue has to
4477 * obtain the queue lock before examing the queue and removing the call.
4479 if (call->call_queue_lock) {
4480 MUTEX_ENTER(call->call_queue_lock);
4481 if (queue_IsOnQueue(call)) {
4483 if (flags & RX_CALL_WAIT_PROC) {
4484 MUTEX_ENTER(&rx_stats_mutex);
4486 MUTEX_EXIT(&rx_stats_mutex);
4489 MUTEX_EXIT(call->call_queue_lock);
4490 CLEAR_CALL_QUEUE_LOCK(call);
4492 #else /* RX_ENABLE_LOCKS */
4493 if (queue_IsOnQueue(call)) {
4495 if (flags & RX_CALL_WAIT_PROC)
4498 #endif /* RX_ENABLE_LOCKS */
4500 rxi_KeepAliveOff(call);
4501 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4504 /* Send an acknowledge for the indicated packet (seq,serial) of the
4505 * indicated call, for the indicated reason (reason). This
4506 * acknowledge will specifically acknowledge receiving the packet, and
4507 * will also specify which other packets for this call have been
4508 * received. This routine returns the packet that was used to the
4509 * caller. The caller is responsible for freeing it or re-using it.
4510 * This acknowledgement also returns the highest sequence number
4511 * actually read out by the higher level to the sender; the sender
4512 * promises to keep around packets that have not been read by the
4513 * higher level yet (unless, of course, the sender decides to abort
4514 * the call altogether). Any of p, seq, serial, pflags, or reason may
4515 * be set to zero without ill effect. That is, if they are zero, they
4516 * will not convey any information.
4517 * NOW there is a trailer field, after the ack where it will safely be
4518 * ignored by mundanes, which indicates the maximum size packet this
4519 * host can swallow. */
4521 register struct rx_packet *optionalPacket; use to send ack (or null)
4522 int seq; Sequence number of the packet we are acking
4523 int serial; Serial number of the packet
4524 int pflags; Flags field from packet header
4525 int reason; Reason an acknowledge was prompted
4529 rxi_SendAck(register struct rx_call *call,
4530 register struct rx_packet *optionalPacket, int serial, int reason,
4533 struct rx_ackPacket *ap;
4534 register struct rx_packet *rqp;
4535 register struct rx_packet *nxp; /* For queue_Scan */
4536 register struct rx_packet *p;
4541 * Open the receive window once a thread starts reading packets
4543 if (call->rnext > 1) {
4544 call->rwind = rx_maxReceiveWindow;
4547 call->nHardAcks = 0;
4548 call->nSoftAcks = 0;
4549 if (call->rnext > call->lastAcked)
4550 call->lastAcked = call->rnext;
4554 rx_computelen(p, p->length); /* reset length, you never know */
4555 } /* where that's been... */
4556 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4557 /* We won't send the ack, but don't panic. */
4558 return optionalPacket;
4562 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4565 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL)) {
4566 if (!optionalPacket)
4568 return optionalPacket;
4570 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4571 if (rx_Contiguous(p) < templ) {
4572 if (!optionalPacket)
4574 return optionalPacket;
4579 /* MTUXXX failing to send an ack is very serious. We should */
4580 /* try as hard as possible to send even a partial ack; it's */
4581 /* better than nothing. */
4582 ap = (struct rx_ackPacket *)rx_DataOf(p);
4583 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4584 ap->reason = reason;
4586 /* The skew computation used to be bogus, I think it's better now. */
4587 /* We should start paying attention to skew. XXX */
4588 ap->serial = htonl(serial);
4589 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4591 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4592 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4594 /* No fear of running out of ack packet here because there can only be at most
4595 * one window full of unacknowledged packets. The window size must be constrained
4596 * to be less than the maximum ack size, of course. Also, an ack should always
4597 * fit into a single packet -- it should not ever be fragmented. */
4598 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4599 if (!rqp || !call->rq.next
4600 || (rqp->header.seq > (call->rnext + call->rwind))) {
4601 if (!optionalPacket)
4603 rxi_CallError(call, RX_CALL_DEAD);
4604 return optionalPacket;
4607 while (rqp->header.seq > call->rnext + offset)
4608 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4609 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4611 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4612 if (!optionalPacket)
4614 rxi_CallError(call, RX_CALL_DEAD);
4615 return optionalPacket;
4620 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4622 /* these are new for AFS 3.3 */
4623 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4624 templ = htonl(templ);
4625 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4626 templ = htonl(call->conn->peer->ifMTU);
4627 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4628 sizeof(afs_int32), &templ);
4630 /* new for AFS 3.4 */
4631 templ = htonl(call->rwind);
4632 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4633 sizeof(afs_int32), &templ);
4635 /* new for AFS 3.5 */
4636 templ = htonl(call->conn->peer->ifDgramPackets);
4637 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4638 sizeof(afs_int32), &templ);
4640 p->header.serviceId = call->conn->serviceId;
4641 p->header.cid = (call->conn->cid | call->channel);
4642 p->header.callNumber = *call->callNumber;
4644 p->header.securityIndex = call->conn->securityIndex;
4645 p->header.epoch = call->conn->epoch;
4646 p->header.type = RX_PACKET_TYPE_ACK;
4647 p->header.flags = RX_SLOW_START_OK;
4648 if (reason == RX_ACK_PING) {
4649 p->header.flags |= RX_REQUEST_ACK;
4651 clock_GetTime(&call->pingRequestTime);
4654 if (call->conn->type == RX_CLIENT_CONNECTION)
4655 p->header.flags |= RX_CLIENT_INITIATED;
4659 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u",
4660 ap->reason, ntohl(ap->previousPacket),
4661 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4663 for (offset = 0; offset < ap->nAcks; offset++)
4664 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4672 register int i, nbytes = p->length;
4674 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4675 if (nbytes <= p->wirevec[i].iov_len) {
4676 register int savelen, saven;
4678 savelen = p->wirevec[i].iov_len;
4680 p->wirevec[i].iov_len = nbytes;
4682 rxi_Send(call, p, istack);
4683 p->wirevec[i].iov_len = savelen;
4687 nbytes -= p->wirevec[i].iov_len;
4690 MUTEX_ENTER(&rx_stats_mutex);
4691 rx_stats.ackPacketsSent++;
4692 MUTEX_EXIT(&rx_stats_mutex);
4693 if (!optionalPacket)
4695 return optionalPacket; /* Return packet for re-use by caller */
4698 /* Send all of the packets in the list in single datagram */
4700 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4701 int istack, int moreFlag, struct clock *now,
4702 struct clock *retryTime, int resending)
4707 struct rx_connection *conn = call->conn;
4708 struct rx_peer *peer = conn->peer;
4710 MUTEX_ENTER(&peer->peer_lock);
4713 peer->reSends += len;
4714 MUTEX_ENTER(&rx_stats_mutex);
4715 rx_stats.dataPacketsSent += len;
4716 MUTEX_EXIT(&rx_stats_mutex);
4717 MUTEX_EXIT(&peer->peer_lock);
4719 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4723 /* Set the packet flags and schedule the resend events */
4724 /* Only request an ack for the last packet in the list */
4725 for (i = 0; i < len; i++) {
4726 list[i]->retryTime = *retryTime;
4727 if (list[i]->header.serial) {
4728 /* Exponentially backoff retry times */
4729 if (list[i]->backoff < MAXBACKOFF) {
4730 /* so it can't stay == 0 */
4731 list[i]->backoff = (list[i]->backoff << 1) + 1;
4734 clock_Addmsec(&(list[i]->retryTime),
4735 ((afs_uint32) list[i]->backoff) << 8);
4738 /* Wait a little extra for the ack on the last packet */
4739 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4740 clock_Addmsec(&(list[i]->retryTime), 400);
4743 /* Record the time sent */
4744 list[i]->timeSent = *now;
4746 /* Ask for an ack on retransmitted packets, on every other packet
4747 * if the peer doesn't support slow start. Ask for an ack on every
4748 * packet until the congestion window reaches the ack rate. */
4749 if (list[i]->header.serial) {
4751 MUTEX_ENTER(&rx_stats_mutex);
4752 rx_stats.dataPacketsReSent++;
4753 MUTEX_EXIT(&rx_stats_mutex);
4755 /* improved RTO calculation- not Karn */
4756 list[i]->firstSent = *now;
4757 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4758 || (!(call->flags & RX_CALL_SLOW_START_OK)
4759 && (list[i]->header.seq & 1)))) {
4764 MUTEX_ENTER(&peer->peer_lock);
4768 MUTEX_ENTER(&rx_stats_mutex);
4769 rx_stats.dataPacketsSent++;
4770 MUTEX_EXIT(&rx_stats_mutex);
4771 MUTEX_EXIT(&peer->peer_lock);
4773 /* Tag this packet as not being the last in this group,
4774 * for the receiver's benefit */
4775 if (i < len - 1 || moreFlag) {
4776 list[i]->header.flags |= RX_MORE_PACKETS;
4779 /* Install the new retransmit time for the packet, and
4780 * record the time sent */
4781 list[i]->timeSent = *now;
4785 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4788 /* Since we're about to send a data packet to the peer, it's
4789 * safe to nuke any scheduled end-of-packets ack */
4790 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4792 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4793 MUTEX_EXIT(&call->lock);
4795 rxi_SendPacketList(call, conn, list, len, istack);
4797 rxi_SendPacket(call, conn, list[0], istack);
4799 MUTEX_ENTER(&call->lock);
4800 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4802 /* Update last send time for this call (for keep-alive
4803 * processing), and for the connection (so that we can discover
4804 * idle connections) */
4805 conn->lastSendTime = call->lastSendTime = clock_Sec();
4808 /* When sending packets we need to follow these rules:
4809 * 1. Never send more than maxDgramPackets in a jumbogram.
4810 * 2. Never send a packet with more than two iovecs in a jumbogram.
4811 * 3. Never send a retransmitted packet in a jumbogram.
4812 * 4. Never send more than cwind/4 packets in a jumbogram
4813 * We always keep the last list we should have sent so we
4814 * can set the RX_MORE_PACKETS flags correctly.
4817 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4818 int istack, struct clock *now, struct clock *retryTime,
4821 int i, cnt, lastCnt = 0;
4822 struct rx_packet **listP, **lastP = 0;
4823 struct rx_peer *peer = call->conn->peer;
4824 int morePackets = 0;
4826 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4827 /* Does the current packet force us to flush the current list? */
4829 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4830 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4832 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4834 /* If the call enters an error state stop sending, or if
4835 * we entered congestion recovery mode, stop sending */
4836 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4844 /* Add the current packet to the list if it hasn't been acked.
4845 * Otherwise adjust the list pointer to skip the current packet. */
4846 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
4848 /* Do we need to flush the list? */
4849 if (cnt >= (int)peer->maxDgramPackets
4850 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
4851 || list[i]->header.serial
4852 || list[i]->length != RX_JUMBOBUFFERSIZE) {
4854 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
4855 retryTime, resending);
4856 /* If the call enters an error state stop sending, or if
4857 * we entered congestion recovery mode, stop sending */
4859 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4864 listP = &list[i + 1];
4869 osi_Panic("rxi_SendList error");
4871 listP = &list[i + 1];
4875 /* Send the whole list when the call is in receive mode, when
4876 * the call is in eof mode, when we are in fast recovery mode,
4877 * and when we have the last packet */
4878 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
4879 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
4880 || (call->flags & RX_CALL_FAST_RECOVER)) {
4881 /* Check for the case where the current list contains
4882 * an acked packet. Since we always send retransmissions
4883 * in a separate packet, we only need to check the first
4884 * packet in the list */
4885 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
4889 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
4890 retryTime, resending);
4891 /* If the call enters an error state stop sending, or if
4892 * we entered congestion recovery mode, stop sending */
4893 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4897 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
4900 } else if (lastCnt > 0) {
4901 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
4906 #ifdef RX_ENABLE_LOCKS
4907 /* Call rxi_Start, below, but with the call lock held. */
4909 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
4912 MUTEX_ENTER(&call->lock);
4913 rxi_Start(event, call, istack);
4914 MUTEX_EXIT(&call->lock);
4916 #endif /* RX_ENABLE_LOCKS */
4918 /* This routine is called when new packets are readied for
4919 * transmission and when retransmission may be necessary, or when the
4920 * transmission window or burst count are favourable. This should be
4921 * better optimized for new packets, the usual case, now that we've
4922 * got rid of queues of send packets. XXXXXXXXXXX */
4924 rxi_Start(struct rxevent *event, register struct rx_call *call, int istack)
4926 struct rx_packet *p;
4927 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
4928 struct rx_peer *peer = call->conn->peer;
4929 struct clock now, retryTime;
4933 struct rx_packet **xmitList;
4936 /* If rxi_Start is being called as a result of a resend event,
4937 * then make sure that the event pointer is removed from the call
4938 * structure, since there is no longer a per-call retransmission
4940 if (event && event == call->resendEvent) {
4941 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
4942 call->resendEvent = NULL;
4944 if (queue_IsEmpty(&call->tq)) {
4948 /* Timeouts trigger congestion recovery */
4949 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4950 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4951 /* someone else is waiting to start recovery */
4954 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4955 while (call->flags & RX_CALL_TQ_BUSY) {
4956 call->flags |= RX_CALL_TQ_WAIT;
4957 #ifdef RX_ENABLE_LOCKS
4958 CV_WAIT(&call->cv_tq, &call->lock);
4959 #else /* RX_ENABLE_LOCKS */
4960 osi_rxSleep(&call->tq);
4961 #endif /* RX_ENABLE_LOCKS */
4963 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4964 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4965 call->flags |= RX_CALL_FAST_RECOVER;
4966 if (peer->maxDgramPackets > 1) {
4967 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4969 call->MTU = MIN(peer->natMTU, peer->maxMTU);
4971 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4972 call->nDgramPackets = 1;
4974 call->nextCwind = 1;
4977 MUTEX_ENTER(&peer->peer_lock);
4978 peer->MTU = call->MTU;
4979 peer->cwind = call->cwind;
4980 peer->nDgramPackets = 1;
4982 call->congestSeq = peer->congestSeq;
4983 MUTEX_EXIT(&peer->peer_lock);
4984 /* Clear retry times on packets. Otherwise, it's possible for
4985 * some packets in the queue to force resends at rates faster
4986 * than recovery rates.
4988 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
4989 if (!(p->flags & RX_PKTFLAG_ACKED)) {
4990 clock_Zero(&p->retryTime);
4995 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4996 MUTEX_ENTER(&rx_stats_mutex);
4997 rx_tq_debug.rxi_start_in_error++;
4998 MUTEX_EXIT(&rx_stats_mutex);
5003 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5004 /* Get clock to compute the re-transmit time for any packets
5005 * in this burst. Note, if we back off, it's reasonable to
5006 * back off all of the packets in the same manner, even if
5007 * some of them have been retransmitted more times than more
5008 * recent additions */
5009 clock_GetTime(&now);
5010 retryTime = now; /* initialize before use */
5011 MUTEX_ENTER(&peer->peer_lock);
5012 clock_Add(&retryTime, &peer->timeout);
5013 MUTEX_EXIT(&peer->peer_lock);
5015 /* Send (or resend) any packets that need it, subject to
5016 * window restrictions and congestion burst control
5017 * restrictions. Ask for an ack on the last packet sent in
5018 * this burst. For now, we're relying upon the window being
5019 * considerably bigger than the largest number of packets that
5020 * are typically sent at once by one initial call to
5021 * rxi_Start. This is probably bogus (perhaps we should ask
5022 * for an ack when we're half way through the current
5023 * window?). Also, for non file transfer applications, this
5024 * may end up asking for an ack for every packet. Bogus. XXXX
5027 * But check whether we're here recursively, and let the other guy
5030 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5031 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5032 call->flags |= RX_CALL_TQ_BUSY;
5034 call->flags &= ~RX_CALL_NEED_START;
5035 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5037 maxXmitPackets = MIN(call->twind, call->cwind);
5038 xmitList = (struct rx_packet **)
5039 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5040 if (xmitList == NULL)
5041 osi_Panic("rxi_Start, failed to allocate xmit list");
5042 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5043 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5044 /* We shouldn't be sending packets if a thread is waiting
5045 * to initiate congestion recovery */
5049 && (call->flags & RX_CALL_FAST_RECOVER)) {
5050 /* Only send one packet during fast recovery */
5053 if ((p->flags & RX_PKTFLAG_FREE)
5054 || (!queue_IsEnd(&call->tq, nxp)
5055 && (nxp->flags & RX_PKTFLAG_FREE))
5056 || (p == (struct rx_packet *)&rx_freePacketQueue)
5057 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5058 osi_Panic("rxi_Start: xmit queue clobbered");
5060 if (p->flags & RX_PKTFLAG_ACKED) {
5061 MUTEX_ENTER(&rx_stats_mutex);
5062 rx_stats.ignoreAckedPacket++;
5063 MUTEX_EXIT(&rx_stats_mutex);
5064 continue; /* Ignore this packet if it has been acknowledged */
5067 /* Turn off all flags except these ones, which are the same
5068 * on each transmission */
5069 p->header.flags &= RX_PRESET_FLAGS;
5071 if (p->header.seq >=
5072 call->tfirst + MIN((int)call->twind,
5073 (int)(call->nSoftAcked +
5075 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5076 /* Note: if we're waiting for more window space, we can
5077 * still send retransmits; hence we don't return here, but
5078 * break out to schedule a retransmit event */
5079 dpf(("call %d waiting for window",
5080 *(call->callNumber)));
5084 /* Transmit the packet if it needs to be sent. */
5085 if (!clock_Lt(&now, &p->retryTime)) {
5086 if (nXmitPackets == maxXmitPackets) {
5087 osi_Panic("rxi_Start: xmit list overflowed");
5089 xmitList[nXmitPackets++] = p;
5093 /* xmitList now hold pointers to all of the packets that are
5094 * ready to send. Now we loop to send the packets */
5095 if (nXmitPackets > 0) {
5096 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5097 &now, &retryTime, resending);
5100 maxXmitPackets * sizeof(struct rx_packet *));
5102 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5104 * TQ references no longer protected by this flag; they must remain
5105 * protected by the global lock.
5107 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5108 call->flags &= ~RX_CALL_TQ_BUSY;
5109 if (call->flags & RX_CALL_TQ_WAIT) {
5110 call->flags &= ~RX_CALL_TQ_WAIT;
5111 #ifdef RX_ENABLE_LOCKS
5112 CV_BROADCAST(&call->cv_tq);
5113 #else /* RX_ENABLE_LOCKS */
5114 osi_rxWakeup(&call->tq);
5115 #endif /* RX_ENABLE_LOCKS */
5120 /* We went into the error state while sending packets. Now is
5121 * the time to reset the call. This will also inform the using
5122 * process that the call is in an error state.
5124 MUTEX_ENTER(&rx_stats_mutex);
5125 rx_tq_debug.rxi_start_aborted++;
5126 MUTEX_EXIT(&rx_stats_mutex);
5127 call->flags &= ~RX_CALL_TQ_BUSY;
5128 if (call->flags & RX_CALL_TQ_WAIT) {
5129 call->flags &= ~RX_CALL_TQ_WAIT;
5130 #ifdef RX_ENABLE_LOCKS
5131 CV_BROADCAST(&call->cv_tq);
5132 #else /* RX_ENABLE_LOCKS */
5133 osi_rxWakeup(&call->tq);
5134 #endif /* RX_ENABLE_LOCKS */
5136 rxi_CallError(call, call->error);
5139 #ifdef RX_ENABLE_LOCKS
5140 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5141 register int missing;
5142 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5143 /* Some packets have received acks. If they all have, we can clear
5144 * the transmit queue.
5147 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5148 if (p->header.seq < call->tfirst
5149 && (p->flags & RX_PKTFLAG_ACKED)) {
5156 call->flags |= RX_CALL_TQ_CLEARME;
5158 #endif /* RX_ENABLE_LOCKS */
5159 /* Don't bother doing retransmits if the TQ is cleared. */
5160 if (call->flags & RX_CALL_TQ_CLEARME) {
5161 rxi_ClearTransmitQueue(call, 1);
5163 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5166 /* Always post a resend event, if there is anything in the
5167 * queue, and resend is possible. There should be at least
5168 * one unacknowledged packet in the queue ... otherwise none
5169 * of these packets should be on the queue in the first place.
5171 if (call->resendEvent) {
5172 /* Cancel the existing event and post a new one */
5173 rxevent_Cancel(call->resendEvent, call,
5174 RX_CALL_REFCOUNT_RESEND);
5177 /* The retry time is the retry time on the first unacknowledged
5178 * packet inside the current window */
5180 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5181 /* Don't set timers for packets outside the window */
5182 if (p->header.seq >= call->tfirst + call->twind) {
5186 if (!(p->flags & RX_PKTFLAG_ACKED)
5187 && !clock_IsZero(&p->retryTime)) {
5189 retryTime = p->retryTime;
5194 /* Post a new event to re-run rxi_Start when retries may be needed */
5195 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5196 #ifdef RX_ENABLE_LOCKS
5197 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5199 rxevent_Post(&retryTime, rxi_StartUnlocked,
5200 (void *)call, (void *)istack);
5201 #else /* RX_ENABLE_LOCKS */
5203 rxevent_Post(&retryTime, rxi_Start, (void *)call,
5205 #endif /* RX_ENABLE_LOCKS */
5208 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5209 } while (call->flags & RX_CALL_NEED_START);
5211 * TQ references no longer protected by this flag; they must remain
5212 * protected by the global lock.
5214 call->flags &= ~RX_CALL_TQ_BUSY;
5215 if (call->flags & RX_CALL_TQ_WAIT) {
5216 call->flags &= ~RX_CALL_TQ_WAIT;
5217 #ifdef RX_ENABLE_LOCKS
5218 CV_BROADCAST(&call->cv_tq);
5219 #else /* RX_ENABLE_LOCKS */
5220 osi_rxWakeup(&call->tq);
5221 #endif /* RX_ENABLE_LOCKS */
5224 call->flags |= RX_CALL_NEED_START;
5226 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5228 if (call->resendEvent) {
5229 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5234 /* Also adjusts the keep alive parameters for the call, to reflect
5235 * that we have just sent a packet (so keep alives aren't sent
5238 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5241 register struct rx_connection *conn = call->conn;
5243 /* Stamp each packet with the user supplied status */
5244 p->header.userStatus = call->localStatus;
5246 /* Allow the security object controlling this call's security to
5247 * make any last-minute changes to the packet */
5248 RXS_SendPacket(conn->securityObject, call, p);
5250 /* Since we're about to send SOME sort of packet to the peer, it's
5251 * safe to nuke any scheduled end-of-packets ack */
5252 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5254 /* Actually send the packet, filling in more connection-specific fields */
5255 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5256 MUTEX_EXIT(&call->lock);
5257 rxi_SendPacket(call, conn, p, istack);
5258 MUTEX_ENTER(&call->lock);
5259 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5261 /* Update last send time for this call (for keep-alive
5262 * processing), and for the connection (so that we can discover
5263 * idle connections) */
5264 conn->lastSendTime = call->lastSendTime = clock_Sec();
5268 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5269 * that things are fine. Also called periodically to guarantee that nothing
5270 * falls through the cracks (e.g. (error + dally) connections have keepalive
5271 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5273 * haveCTLock Set if calling from rxi_ReapConnections
5275 #ifdef RX_ENABLE_LOCKS
5277 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5278 #else /* RX_ENABLE_LOCKS */
5280 rxi_CheckCall(register struct rx_call *call)
5281 #endif /* RX_ENABLE_LOCKS */
5283 register struct rx_connection *conn = call->conn;
5285 afs_uint32 deadTime;
5287 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5288 if (call->flags & RX_CALL_TQ_BUSY) {
5289 /* Call is active and will be reset by rxi_Start if it's
5290 * in an error state.
5295 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5297 (((afs_uint32) conn->secondsUntilDead << 10) +
5298 ((afs_uint32) conn->peer->rtt >> 3) +
5299 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5301 /* These are computed to the second (+- 1 second). But that's
5302 * good enough for these values, which should be a significant
5303 * number of seconds. */
5304 if (now > (call->lastReceiveTime + deadTime)) {
5305 if (call->state == RX_STATE_ACTIVE) {
5306 rxi_CallError(call, RX_CALL_DEAD);
5309 #ifdef RX_ENABLE_LOCKS
5310 /* Cancel pending events */
5311 rxevent_Cancel(call->delayedAckEvent, call,
5312 RX_CALL_REFCOUNT_DELAY);
5313 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5314 rxevent_Cancel(call->keepAliveEvent, call,
5315 RX_CALL_REFCOUNT_ALIVE);
5316 if (call->refCount == 0) {
5317 rxi_FreeCall(call, haveCTLock);
5321 #else /* RX_ENABLE_LOCKS */
5324 #endif /* RX_ENABLE_LOCKS */
5326 /* Non-active calls are destroyed if they are not responding
5327 * to pings; active calls are simply flagged in error, so the
5328 * attached process can die reasonably gracefully. */
5330 /* see if we have a non-activity timeout */
5331 if (call->startWait && conn->idleDeadTime
5332 && ((call->startWait + conn->idleDeadTime) < now)) {
5333 if (call->state == RX_STATE_ACTIVE) {
5334 rxi_CallError(call, RX_CALL_TIMEOUT);
5338 /* see if we have a hard timeout */
5339 if (conn->hardDeadTime
5340 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5341 if (call->state == RX_STATE_ACTIVE)
5342 rxi_CallError(call, RX_CALL_TIMEOUT);
5349 /* When a call is in progress, this routine is called occasionally to
5350 * make sure that some traffic has arrived (or been sent to) the peer.
5351 * If nothing has arrived in a reasonable amount of time, the call is
5352 * declared dead; if nothing has been sent for a while, we send a
5353 * keep-alive packet (if we're actually trying to keep the call alive)
5356 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5359 struct rx_connection *conn;
5362 MUTEX_ENTER(&call->lock);
5363 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5364 if (event == call->keepAliveEvent)
5365 call->keepAliveEvent = NULL;
5368 #ifdef RX_ENABLE_LOCKS
5369 if (rxi_CheckCall(call, 0)) {
5370 MUTEX_EXIT(&call->lock);
5373 #else /* RX_ENABLE_LOCKS */
5374 if (rxi_CheckCall(call))
5376 #endif /* RX_ENABLE_LOCKS */
5378 /* Don't try to keep alive dallying calls */
5379 if (call->state == RX_STATE_DALLY) {
5380 MUTEX_EXIT(&call->lock);
5385 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5386 /* Don't try to send keepalives if there is unacknowledged data */
5387 /* the rexmit code should be good enough, this little hack
5388 * doesn't quite work XXX */
5389 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5391 rxi_ScheduleKeepAliveEvent(call);
5392 MUTEX_EXIT(&call->lock);
5397 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5399 if (!call->keepAliveEvent) {
5401 clock_GetTime(&when);
5402 when.sec += call->conn->secondsUntilPing;
5403 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5404 call->keepAliveEvent =
5405 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5409 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5411 rxi_KeepAliveOn(register struct rx_call *call)
5413 /* Pretend last packet received was received now--i.e. if another
5414 * packet isn't received within the keep alive time, then the call
5415 * will die; Initialize last send time to the current time--even
5416 * if a packet hasn't been sent yet. This will guarantee that a
5417 * keep-alive is sent within the ping time */
5418 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5419 rxi_ScheduleKeepAliveEvent(call);
5422 /* This routine is called to send connection abort messages
5423 * that have been delayed to throttle looping clients. */
5425 rxi_SendDelayedConnAbort(struct rxevent *event,
5426 register struct rx_connection *conn, char *dummy)
5429 struct rx_packet *packet;
5431 MUTEX_ENTER(&conn->conn_data_lock);
5432 conn->delayedAbortEvent = NULL;
5433 error = htonl(conn->error);
5435 MUTEX_EXIT(&conn->conn_data_lock);
5436 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5439 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5440 RX_PACKET_TYPE_ABORT, (char *)&error,
5442 rxi_FreePacket(packet);
5446 /* This routine is called to send call abort messages
5447 * that have been delayed to throttle looping clients. */
5449 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5453 struct rx_packet *packet;
5455 MUTEX_ENTER(&call->lock);
5456 call->delayedAbortEvent = NULL;
5457 error = htonl(call->error);
5459 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5462 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5463 (char *)&error, sizeof(error), 0);
5464 rxi_FreePacket(packet);
5466 MUTEX_EXIT(&call->lock);
5469 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5470 * seconds) to ask the client to authenticate itself. The routine
5471 * issues a challenge to the client, which is obtained from the
5472 * security object associated with the connection */
5474 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5477 int tries = (int)atries;
5478 conn->challengeEvent = NULL;
5479 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5480 register struct rx_packet *packet;
5484 /* We've failed to authenticate for too long.
5485 * Reset any calls waiting for authentication;
5486 * they are all in RX_STATE_PRECALL.
5490 MUTEX_ENTER(&conn->conn_call_lock);
5491 for (i = 0; i < RX_MAXCALLS; i++) {
5492 struct rx_call *call = conn->call[i];
5494 MUTEX_ENTER(&call->lock);
5495 if (call->state == RX_STATE_PRECALL) {
5496 rxi_CallError(call, RX_CALL_DEAD);
5497 rxi_SendCallAbort(call, NULL, 0, 0);
5499 MUTEX_EXIT(&call->lock);
5502 MUTEX_EXIT(&conn->conn_call_lock);
5506 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5508 /* If there's no packet available, do this later. */
5509 RXS_GetChallenge(conn->securityObject, conn, packet);
5510 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5511 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5512 rxi_FreePacket(packet);
5514 clock_GetTime(&when);
5515 when.sec += RX_CHALLENGE_TIMEOUT;
5516 conn->challengeEvent =
5517 rxevent_Post(&when, rxi_ChallengeEvent, conn,
5518 (void *)(tries - 1));
5522 /* Call this routine to start requesting the client to authenticate
5523 * itself. This will continue until authentication is established,
5524 * the call times out, or an invalid response is returned. The
5525 * security object associated with the connection is asked to create
5526 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5527 * defined earlier. */
5529 rxi_ChallengeOn(register struct rx_connection *conn)
5531 if (!conn->challengeEvent) {
5532 RXS_CreateChallenge(conn->securityObject, conn);
5533 rxi_ChallengeEvent(NULL, conn, (void *)RX_CHALLENGE_MAXTRIES);
5538 /* Compute round trip time of the packet provided, in *rttp.
5541 /* rxi_ComputeRoundTripTime is called with peer locked. */
5542 /* sentp and/or peer may be null */
5544 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5545 register struct clock *sentp,
5546 register struct rx_peer *peer)
5548 struct clock thisRtt, *rttp = &thisRtt;
5550 #if defined(AFS_ALPHA_LINUX22_ENV) && defined(AFS_PTHREAD_ENV) && !defined(KERNEL)
5551 /* making year 2038 bugs to get this running now - stroucki */
5552 struct timeval temptime;
5554 register int rtt_timeout;
5556 #if defined(AFS_ALPHA_LINUX20_ENV) && defined(AFS_PTHREAD_ENV) && !defined(KERNEL)
5557 /* yet again. This was the worst Heisenbug of the port - stroucki */
5558 clock_GetTime(&temptime);
5559 rttp->sec = (afs_int32) temptime.tv_sec;
5560 rttp->usec = (afs_int32) temptime.tv_usec;
5562 clock_GetTime(rttp);
5564 if (clock_Lt(rttp, sentp)) {
5566 return; /* somebody set the clock back, don't count this time. */
5568 clock_Sub(rttp, sentp);
5569 MUTEX_ENTER(&rx_stats_mutex);
5570 if (clock_Lt(rttp, &rx_stats.minRtt))
5571 rx_stats.minRtt = *rttp;
5572 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5573 if (rttp->sec > 60) {
5574 MUTEX_EXIT(&rx_stats_mutex);
5575 return; /* somebody set the clock ahead */
5577 rx_stats.maxRtt = *rttp;
5579 clock_Add(&rx_stats.totalRtt, rttp);
5580 rx_stats.nRttSamples++;
5581 MUTEX_EXIT(&rx_stats_mutex);
5583 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5585 /* Apply VanJacobson round-trip estimations */
5590 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5591 * srtt is stored as fixed point with 3 bits after the binary
5592 * point (i.e., scaled by 8). The following magic is
5593 * equivalent to the smoothing algorithm in rfc793 with an
5594 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5595 * srtt*8 = srtt*8 + rtt - srtt
5596 * srtt = srtt + rtt/8 - srtt/8
5599 delta = MSEC(rttp) - (peer->rtt >> 3);
5603 * We accumulate a smoothed rtt variance (actually, a smoothed
5604 * mean difference), then set the retransmit timer to smoothed
5605 * rtt + 4 times the smoothed variance (was 2x in van's original
5606 * paper, but 4x works better for me, and apparently for him as
5608 * rttvar is stored as
5609 * fixed point with 2 bits after the binary point (scaled by
5610 * 4). The following is equivalent to rfc793 smoothing with
5611 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5612 * replaces rfc793's wired-in beta.
5613 * dev*4 = dev*4 + (|actual - expected| - dev)
5619 delta -= (peer->rtt_dev >> 2);
5620 peer->rtt_dev += delta;
5622 /* I don't have a stored RTT so I start with this value. Since I'm
5623 * probably just starting a call, and will be pushing more data down
5624 * this, I expect congestion to increase rapidly. So I fudge a
5625 * little, and I set deviance to half the rtt. In practice,
5626 * deviance tends to approach something a little less than
5627 * half the smoothed rtt. */
5628 peer->rtt = (MSEC(rttp) << 3) + 8;
5629 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5631 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5632 * the other of these connections is usually in a user process, and can
5633 * be switched and/or swapped out. So on fast, reliable networks, the
5634 * timeout would otherwise be too short.
5636 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5637 clock_Zero(&(peer->timeout));
5638 clock_Addmsec(&(peer->timeout), rtt_timeout);
5640 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)));
5644 /* Find all server connections that have not been active for a long time, and
5647 rxi_ReapConnections(void)
5650 clock_GetTime(&now);
5652 /* Find server connection structures that haven't been used for
5653 * greater than rx_idleConnectionTime */
5655 struct rx_connection **conn_ptr, **conn_end;
5656 int i, havecalls = 0;
5657 MUTEX_ENTER(&rx_connHashTable_lock);
5658 for (conn_ptr = &rx_connHashTable[0], conn_end =
5659 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5661 struct rx_connection *conn, *next;
5662 struct rx_call *call;
5666 for (conn = *conn_ptr; conn; conn = next) {
5667 /* XXX -- Shouldn't the connection be locked? */
5670 for (i = 0; i < RX_MAXCALLS; i++) {
5671 call = conn->call[i];
5674 MUTEX_ENTER(&call->lock);
5675 #ifdef RX_ENABLE_LOCKS
5676 result = rxi_CheckCall(call, 1);
5677 #else /* RX_ENABLE_LOCKS */
5678 result = rxi_CheckCall(call);
5679 #endif /* RX_ENABLE_LOCKS */
5680 MUTEX_EXIT(&call->lock);
5682 /* If CheckCall freed the call, it might
5683 * have destroyed the connection as well,
5684 * which screws up the linked lists.
5690 if (conn->type == RX_SERVER_CONNECTION) {
5691 /* This only actually destroys the connection if
5692 * there are no outstanding calls */
5693 MUTEX_ENTER(&conn->conn_data_lock);
5694 if (!havecalls && !conn->refCount
5695 && ((conn->lastSendTime + rx_idleConnectionTime) <
5697 conn->refCount++; /* it will be decr in rx_DestroyConn */
5698 MUTEX_EXIT(&conn->conn_data_lock);
5699 #ifdef RX_ENABLE_LOCKS
5700 rxi_DestroyConnectionNoLock(conn);
5701 #else /* RX_ENABLE_LOCKS */
5702 rxi_DestroyConnection(conn);
5703 #endif /* RX_ENABLE_LOCKS */
5705 #ifdef RX_ENABLE_LOCKS
5707 MUTEX_EXIT(&conn->conn_data_lock);
5709 #endif /* RX_ENABLE_LOCKS */
5713 #ifdef RX_ENABLE_LOCKS
5714 while (rx_connCleanup_list) {
5715 struct rx_connection *conn;
5716 conn = rx_connCleanup_list;
5717 rx_connCleanup_list = rx_connCleanup_list->next;
5718 MUTEX_EXIT(&rx_connHashTable_lock);
5719 rxi_CleanupConnection(conn);
5720 MUTEX_ENTER(&rx_connHashTable_lock);
5722 MUTEX_EXIT(&rx_connHashTable_lock);
5723 #endif /* RX_ENABLE_LOCKS */
5726 /* Find any peer structures that haven't been used (haven't had an
5727 * associated connection) for greater than rx_idlePeerTime */
5729 struct rx_peer **peer_ptr, **peer_end;
5731 MUTEX_ENTER(&rx_rpc_stats);
5732 MUTEX_ENTER(&rx_peerHashTable_lock);
5733 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5734 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5736 struct rx_peer *peer, *next, *prev;
5737 for (prev = peer = *peer_ptr; peer; peer = next) {
5739 code = MUTEX_TRYENTER(&peer->peer_lock);
5740 if ((code) && (peer->refCount == 0)
5741 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5742 rx_interface_stat_p rpc_stat, nrpc_stat;
5744 MUTEX_EXIT(&peer->peer_lock);
5745 MUTEX_DESTROY(&peer->peer_lock);
5747 (&peer->rpcStats, rpc_stat, nrpc_stat,
5748 rx_interface_stat)) {
5749 unsigned int num_funcs;
5752 queue_Remove(&rpc_stat->queue_header);
5753 queue_Remove(&rpc_stat->all_peers);
5754 num_funcs = rpc_stat->stats[0].func_total;
5756 sizeof(rx_interface_stat_t) +
5757 rpc_stat->stats[0].func_total *
5758 sizeof(rx_function_entry_v1_t);
5760 rxi_Free(rpc_stat, space);
5761 rxi_rpc_peer_stat_cnt -= num_funcs;
5764 MUTEX_ENTER(&rx_stats_mutex);
5765 rx_stats.nPeerStructs--;
5766 MUTEX_EXIT(&rx_stats_mutex);
5767 if (peer == *peer_ptr) {
5774 MUTEX_EXIT(&peer->peer_lock);
5780 MUTEX_EXIT(&rx_peerHashTable_lock);
5781 MUTEX_EXIT(&rx_rpc_stats);
5784 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5785 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5786 * GC, just below. Really, we shouldn't have to keep moving packets from
5787 * one place to another, but instead ought to always know if we can
5788 * afford to hold onto a packet in its particular use. */
5789 MUTEX_ENTER(&rx_freePktQ_lock);
5790 if (rx_waitingForPackets) {
5791 rx_waitingForPackets = 0;
5792 #ifdef RX_ENABLE_LOCKS
5793 CV_BROADCAST(&rx_waitingForPackets_cv);
5795 osi_rxWakeup(&rx_waitingForPackets);
5798 MUTEX_EXIT(&rx_freePktQ_lock);
5800 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5801 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5805 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5806 * rx.h is sort of strange this is better. This is called with a security
5807 * object before it is discarded. Each connection using a security object has
5808 * its own refcount to the object so it won't actually be freed until the last
5809 * connection is destroyed.
5811 * This is the only rxs module call. A hold could also be written but no one
5815 rxs_Release(struct rx_securityClass *aobj)
5817 return RXS_Close(aobj);
5821 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5822 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5823 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5824 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5826 /* Adjust our estimate of the transmission rate to this peer, given
5827 * that the packet p was just acked. We can adjust peer->timeout and
5828 * call->twind. Pragmatically, this is called
5829 * only with packets of maximal length.
5830 * Called with peer and call locked.
5834 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5835 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5837 afs_int32 xferSize, xferMs;
5838 register afs_int32 minTime;
5841 /* Count down packets */
5842 if (peer->rateFlag > 0)
5844 /* Do nothing until we're enabled */
5845 if (peer->rateFlag != 0)
5850 /* Count only when the ack seems legitimate */
5851 switch (ackReason) {
5852 case RX_ACK_REQUESTED:
5854 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
5858 case RX_ACK_PING_RESPONSE:
5859 if (p) /* want the response to ping-request, not data send */
5861 clock_GetTime(&newTO);
5862 if (clock_Gt(&newTO, &call->pingRequestTime)) {
5863 clock_Sub(&newTO, &call->pingRequestTime);
5864 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
5868 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
5875 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));
5877 /* Track only packets that are big enough. */
5878 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
5882 /* absorb RTT data (in milliseconds) for these big packets */
5883 if (peer->smRtt == 0) {
5884 peer->smRtt = xferMs;
5886 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
5891 if (peer->countDown) {
5895 peer->countDown = 10; /* recalculate only every so often */
5897 /* In practice, we can measure only the RTT for full packets,
5898 * because of the way Rx acks the data that it receives. (If it's
5899 * smaller than a full packet, it often gets implicitly acked
5900 * either by the call response (from a server) or by the next call
5901 * (from a client), and either case confuses transmission times
5902 * with processing times.) Therefore, replace the above
5903 * more-sophisticated processing with a simpler version, where the
5904 * smoothed RTT is kept for full-size packets, and the time to
5905 * transmit a windowful of full-size packets is simply RTT *
5906 * windowSize. Again, we take two steps:
5907 - ensure the timeout is large enough for a single packet's RTT;
5908 - ensure that the window is small enough to fit in the desired timeout.*/
5910 /* First, the timeout check. */
5911 minTime = peer->smRtt;
5912 /* Get a reasonable estimate for a timeout period */
5914 newTO.sec = minTime / 1000;
5915 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
5917 /* Increase the timeout period so that we can always do at least
5918 * one packet exchange */
5919 if (clock_Gt(&newTO, &peer->timeout)) {
5921 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));
5923 peer->timeout = newTO;
5926 /* Now, get an estimate for the transmit window size. */
5927 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
5928 /* Now, convert to the number of full packets that could fit in a
5929 * reasonable fraction of that interval */
5930 minTime /= (peer->smRtt << 1);
5931 xferSize = minTime; /* (make a copy) */
5933 /* Now clamp the size to reasonable bounds. */
5936 else if (minTime > rx_Window)
5937 minTime = rx_Window;
5938 /* if (minTime != peer->maxWindow) {
5939 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
5940 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
5941 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
5943 peer->maxWindow = minTime;
5944 elide... call->twind = minTime;
5948 /* Cut back on the peer timeout if it had earlier grown unreasonably.
5949 * Discern this by calculating the timeout necessary for rx_Window
5951 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
5952 /* calculate estimate for transmission interval in milliseconds */
5953 minTime = rx_Window * peer->smRtt;
5954 if (minTime < 1000) {
5955 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
5956 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
5957 peer->timeout.usec, peer->smRtt, peer->packetSize));
5959 newTO.sec = 0; /* cut back on timeout by half a second */
5960 newTO.usec = 500000;
5961 clock_Sub(&peer->timeout, &newTO);
5966 } /* end of rxi_ComputeRate */
5967 #endif /* ADAPT_WINDOW */
5975 /* Don't call this debugging routine directly; use dpf */
5977 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
5978 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
5982 clock_GetTime(&now);
5983 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
5984 (unsigned int)now.usec / 1000);
5985 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
5993 * This function is used to process the rx_stats structure that is local
5994 * to a process as well as an rx_stats structure received from a remote
5995 * process (via rxdebug). Therefore, it needs to do minimal version
5999 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6000 afs_int32 freePackets, char version)
6004 if (size != sizeof(struct rx_stats)) {
6006 "Unexpected size of stats structure: was %d, expected %d\n",
6007 size, sizeof(struct rx_stats));
6010 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6013 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6014 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6015 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6016 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6017 s->specialPktAllocFailures);
6019 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6020 s->receivePktAllocFailures, s->sendPktAllocFailures,
6021 s->specialPktAllocFailures);
6025 " greedy %d, " "bogusReads %d (last from host %x), "
6026 "noPackets %d, " "noBuffers %d, " "selects %d, "
6027 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6028 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6029 s->selects, s->sendSelects);
6031 fprintf(file, " packets read: ");
6032 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6033 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6035 fprintf(file, "\n");
6038 " other read counters: data %d, " "ack %d, " "dup %d "
6039 "spurious %d " "dally %d\n", s->dataPacketsRead,
6040 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6041 s->ignorePacketDally);
6043 fprintf(file, " packets sent: ");
6044 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6045 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6047 fprintf(file, "\n");
6050 " other send counters: ack %d, " "data %d (not resends), "
6051 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6052 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6053 s->dataPacketsPushed, s->ignoreAckedPacket);
6056 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6057 s->netSendFailures, (int)s->fatalErrors);
6059 if (s->nRttSamples) {
6060 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6061 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6063 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6064 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6068 " %d server connections, " "%d client connections, "
6069 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6070 s->nServerConns, s->nClientConns, s->nPeerStructs,
6071 s->nCallStructs, s->nFreeCallStructs);
6073 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6074 fprintf(file, " %d clock updates\n", clock_nUpdates);
6079 /* for backward compatibility */
6081 rx_PrintStats(FILE * file)
6083 MUTEX_ENTER(&rx_stats_mutex);
6084 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6086 MUTEX_EXIT(&rx_stats_mutex);
6090 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6092 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6093 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6094 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6097 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6098 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6099 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6102 " Packet size %d, " "max in packet skew %d, "
6103 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6104 (int)peer->outPacketSkew);
6107 #ifdef AFS_PTHREAD_ENV
6109 * This mutex protects the following static variables:
6113 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6114 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6116 #define LOCK_RX_DEBUG
6117 #define UNLOCK_RX_DEBUG
6118 #endif /* AFS_PTHREAD_ENV */
6121 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6122 u_char type, void *inputData, size_t inputLength,
6123 void *outputData, size_t outputLength)
6125 static afs_int32 counter = 100;
6127 struct rx_header theader;
6129 register afs_int32 code;
6131 struct sockaddr_in taddr, faddr;
6136 endTime = time(0) + 20; /* try for 20 seconds */
6140 tp = &tbuffer[sizeof(struct rx_header)];
6141 taddr.sin_family = AF_INET;
6142 taddr.sin_port = remotePort;
6143 taddr.sin_addr.s_addr = remoteAddr;
6144 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6145 taddr.sin_len = sizeof(struct sockaddr_in);
6148 memset(&theader, 0, sizeof(theader));
6149 theader.epoch = htonl(999);
6151 theader.callNumber = htonl(counter);
6154 theader.type = type;
6155 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6156 theader.serviceId = 0;
6158 memcpy(tbuffer, &theader, sizeof(theader));
6159 memcpy(tp, inputData, inputLength);
6161 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6162 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6164 /* see if there's a packet available */
6166 FD_SET(socket, &imask);
6169 code = select(socket + 1, &imask, 0, 0, &tv);
6170 if (code == 1 && FD_ISSET(socket, &imask)) {
6171 /* now receive a packet */
6172 faddrLen = sizeof(struct sockaddr_in);
6174 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6175 (struct sockaddr *)&faddr, &faddrLen);
6178 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6179 if (counter == ntohl(theader.callNumber))
6184 /* see if we've timed out */
6185 if (endTime < time(0))
6188 code -= sizeof(struct rx_header);
6189 if (code > outputLength)
6190 code = outputLength;
6191 memcpy(outputData, tp, code);
6196 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6197 afs_uint16 remotePort, struct rx_debugStats * stat,
6198 afs_uint32 * supportedValues)
6200 struct rx_debugIn in;
6203 *supportedValues = 0;
6204 in.type = htonl(RX_DEBUGI_GETSTATS);
6207 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6208 &in, sizeof(in), stat, sizeof(*stat));
6211 * If the call was successful, fixup the version and indicate
6212 * what contents of the stat structure are valid.
6213 * Also do net to host conversion of fields here.
6217 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6218 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6220 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6221 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6223 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6224 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6226 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6227 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6229 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6230 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6232 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6233 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6235 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6236 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6238 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6239 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6242 stat->nFreePackets = ntohl(stat->nFreePackets);
6243 stat->packetReclaims = ntohl(stat->packetReclaims);
6244 stat->callsExecuted = ntohl(stat->callsExecuted);
6245 stat->nWaiting = ntohl(stat->nWaiting);
6246 stat->idleThreads = ntohl(stat->idleThreads);
6253 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6254 afs_uint16 remotePort, struct rx_stats * stat,
6255 afs_uint32 * supportedValues)
6257 struct rx_debugIn in;
6258 afs_int32 *lp = (afs_int32 *) stat;
6263 * supportedValues is currently unused, but added to allow future
6264 * versioning of this function.
6267 *supportedValues = 0;
6268 in.type = htonl(RX_DEBUGI_RXSTATS);
6270 memset(stat, 0, sizeof(*stat));
6272 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6273 &in, sizeof(in), stat, sizeof(*stat));
6278 * Do net to host conversion here
6281 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6290 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6291 afs_uint16 remotePort, size_t version_length,
6295 return MakeDebugCall(socket, remoteAddr, remotePort,
6296 RX_PACKET_TYPE_VERSION, a, 1, version,
6301 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6302 afs_uint16 remotePort, afs_int32 * nextConnection,
6303 int allConnections, afs_uint32 debugSupportedValues,
6304 struct rx_debugConn * conn,
6305 afs_uint32 * supportedValues)
6307 struct rx_debugIn in;
6312 * supportedValues is currently unused, but added to allow future
6313 * versioning of this function.
6316 *supportedValues = 0;
6317 if (allConnections) {
6318 in.type = htonl(RX_DEBUGI_GETALLCONN);
6320 in.type = htonl(RX_DEBUGI_GETCONN);
6322 in.index = htonl(*nextConnection);
6323 memset(conn, 0, sizeof(*conn));
6325 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6326 &in, sizeof(in), conn, sizeof(*conn));
6329 *nextConnection += 1;
6332 * Convert old connection format to new structure.
6335 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6336 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6337 #define MOVEvL(a) (conn->a = vL->a)
6339 /* any old or unrecognized version... */
6340 for (i = 0; i < RX_MAXCALLS; i++) {
6341 MOVEvL(callState[i]);
6342 MOVEvL(callMode[i]);
6343 MOVEvL(callFlags[i]);
6344 MOVEvL(callOther[i]);
6346 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6347 MOVEvL(secStats.type);
6348 MOVEvL(secStats.level);
6349 MOVEvL(secStats.flags);
6350 MOVEvL(secStats.expires);
6351 MOVEvL(secStats.packetsReceived);
6352 MOVEvL(secStats.packetsSent);
6353 MOVEvL(secStats.bytesReceived);
6354 MOVEvL(secStats.bytesSent);
6359 * Do net to host conversion here
6361 * I don't convert host or port since we are most likely
6362 * going to want these in NBO.
6364 conn->cid = ntohl(conn->cid);
6365 conn->serial = ntohl(conn->serial);
6366 for (i = 0; i < RX_MAXCALLS; i++) {
6367 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6369 conn->error = ntohl(conn->error);
6370 conn->secStats.flags = ntohl(conn->secStats.flags);
6371 conn->secStats.expires = ntohl(conn->secStats.expires);
6372 conn->secStats.packetsReceived =
6373 ntohl(conn->secStats.packetsReceived);
6374 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6375 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6376 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6377 conn->epoch = ntohl(conn->epoch);
6378 conn->natMTU = ntohl(conn->natMTU);
6385 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6386 afs_uint16 remotePort, afs_int32 * nextPeer,
6387 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6388 afs_uint32 * supportedValues)
6390 struct rx_debugIn in;
6394 * supportedValues is currently unused, but added to allow future
6395 * versioning of this function.
6398 *supportedValues = 0;
6399 in.type = htonl(RX_DEBUGI_GETPEER);
6400 in.index = htonl(*nextPeer);
6401 memset(peer, 0, sizeof(*peer));
6403 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6404 &in, sizeof(in), peer, sizeof(*peer));
6410 * Do net to host conversion here
6412 * I don't convert host or port since we are most likely
6413 * going to want these in NBO.
6415 peer->ifMTU = ntohs(peer->ifMTU);
6416 peer->idleWhen = ntohl(peer->idleWhen);
6417 peer->refCount = ntohs(peer->refCount);
6418 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6419 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6420 peer->rtt = ntohl(peer->rtt);
6421 peer->rtt_dev = ntohl(peer->rtt_dev);
6422 peer->timeout.sec = ntohl(peer->timeout.sec);
6423 peer->timeout.usec = ntohl(peer->timeout.usec);
6424 peer->nSent = ntohl(peer->nSent);
6425 peer->reSends = ntohl(peer->reSends);
6426 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6427 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6428 peer->rateFlag = ntohl(peer->rateFlag);
6429 peer->natMTU = ntohs(peer->natMTU);
6430 peer->maxMTU = ntohs(peer->maxMTU);
6431 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6432 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6433 peer->MTU = ntohs(peer->MTU);
6434 peer->cwind = ntohs(peer->cwind);
6435 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6436 peer->congestSeq = ntohs(peer->congestSeq);
6437 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6438 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6439 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6440 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6445 #endif /* RXDEBUG */
6450 struct rx_serverQueueEntry *np;
6453 register struct rx_call *call;
6454 register struct rx_serverQueueEntry *sq;
6458 if (rxinit_status == 1) {
6460 return; /* Already shutdown. */
6464 #ifndef AFS_PTHREAD_ENV
6465 FD_ZERO(&rx_selectMask);
6466 #endif /* AFS_PTHREAD_ENV */
6467 rxi_dataQuota = RX_MAX_QUOTA;
6468 #ifndef AFS_PTHREAD_ENV
6470 #endif /* AFS_PTHREAD_ENV */
6473 #ifndef AFS_PTHREAD_ENV
6474 #ifndef AFS_USE_GETTIMEOFDAY
6476 #endif /* AFS_USE_GETTIMEOFDAY */
6477 #endif /* AFS_PTHREAD_ENV */
6479 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6480 call = queue_First(&rx_freeCallQueue, rx_call);
6482 rxi_Free(call, sizeof(struct rx_call));
6485 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6486 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6492 struct rx_peer **peer_ptr, **peer_end;
6493 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6494 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6496 struct rx_peer *peer, *next;
6497 for (peer = *peer_ptr; peer; peer = next) {
6498 rx_interface_stat_p rpc_stat, nrpc_stat;
6501 (&peer->rpcStats, rpc_stat, nrpc_stat,
6502 rx_interface_stat)) {
6503 unsigned int num_funcs;
6506 queue_Remove(&rpc_stat->queue_header);
6507 queue_Remove(&rpc_stat->all_peers);
6508 num_funcs = rpc_stat->stats[0].func_total;
6510 sizeof(rx_interface_stat_t) +
6511 rpc_stat->stats[0].func_total *
6512 sizeof(rx_function_entry_v1_t);
6514 rxi_Free(rpc_stat, space);
6515 MUTEX_ENTER(&rx_rpc_stats);
6516 rxi_rpc_peer_stat_cnt -= num_funcs;
6517 MUTEX_EXIT(&rx_rpc_stats);
6521 MUTEX_ENTER(&rx_stats_mutex);
6522 rx_stats.nPeerStructs--;
6523 MUTEX_EXIT(&rx_stats_mutex);
6527 for (i = 0; i < RX_MAX_SERVICES; i++) {
6529 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6531 for (i = 0; i < rx_hashTableSize; i++) {
6532 register struct rx_connection *tc, *ntc;
6533 MUTEX_ENTER(&rx_connHashTable_lock);
6534 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6536 for (j = 0; j < RX_MAXCALLS; j++) {
6538 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6541 rxi_Free(tc, sizeof(*tc));
6543 MUTEX_EXIT(&rx_connHashTable_lock);
6546 MUTEX_ENTER(&freeSQEList_lock);
6548 while ((np = rx_FreeSQEList)) {
6549 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6550 MUTEX_DESTROY(&np->lock);
6551 rxi_Free(np, sizeof(*np));
6554 MUTEX_EXIT(&freeSQEList_lock);
6555 MUTEX_DESTROY(&freeSQEList_lock);
6556 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6557 MUTEX_DESTROY(&rx_connHashTable_lock);
6558 MUTEX_DESTROY(&rx_peerHashTable_lock);
6559 MUTEX_DESTROY(&rx_serverPool_lock);
6561 osi_Free(rx_connHashTable,
6562 rx_hashTableSize * sizeof(struct rx_connection *));
6563 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6565 UNPIN(rx_connHashTable,
6566 rx_hashTableSize * sizeof(struct rx_connection *));
6567 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6569 rxi_FreeAllPackets();
6571 MUTEX_ENTER(&rx_stats_mutex);
6572 rxi_dataQuota = RX_MAX_QUOTA;
6573 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6574 MUTEX_EXIT(&rx_stats_mutex);
6580 #ifdef RX_ENABLE_LOCKS
6582 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6584 if (!MUTEX_ISMINE(lockaddr))
6585 osi_Panic("Lock not held: %s", msg);
6587 #endif /* RX_ENABLE_LOCKS */
6592 * Routines to implement connection specific data.
6596 rx_KeyCreate(rx_destructor_t rtn)
6599 MUTEX_ENTER(&rxi_keyCreate_lock);
6600 key = rxi_keyCreate_counter++;
6601 rxi_keyCreate_destructor = (rx_destructor_t *)
6602 realloc((void *)rxi_keyCreate_destructor,
6603 (key + 1) * sizeof(rx_destructor_t));
6604 rxi_keyCreate_destructor[key] = rtn;
6605 MUTEX_EXIT(&rxi_keyCreate_lock);
6610 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6613 MUTEX_ENTER(&conn->conn_data_lock);
6614 if (!conn->specific) {
6615 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6616 for (i = 0; i < key; i++)
6617 conn->specific[i] = NULL;
6618 conn->nSpecific = key + 1;
6619 conn->specific[key] = ptr;
6620 } else if (key >= conn->nSpecific) {
6621 conn->specific = (void **)
6622 realloc(conn->specific, (key + 1) * sizeof(void *));
6623 for (i = conn->nSpecific; i < key; i++)
6624 conn->specific[i] = NULL;
6625 conn->nSpecific = key + 1;
6626 conn->specific[key] = ptr;
6628 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6629 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6630 conn->specific[key] = ptr;
6632 MUTEX_EXIT(&conn->conn_data_lock);
6636 rx_GetSpecific(struct rx_connection *conn, int key)
6639 MUTEX_ENTER(&conn->conn_data_lock);
6640 if (key >= conn->nSpecific)
6643 ptr = conn->specific[key];
6644 MUTEX_EXIT(&conn->conn_data_lock);
6648 #endif /* !KERNEL */
6651 * processStats is a queue used to store the statistics for the local
6652 * process. Its contents are similar to the contents of the rpcStats
6653 * queue on a rx_peer structure, but the actual data stored within
6654 * this queue contains totals across the lifetime of the process (assuming
6655 * the stats have not been reset) - unlike the per peer structures
6656 * which can come and go based upon the peer lifetime.
6659 static struct rx_queue processStats = { &processStats, &processStats };
6662 * peerStats is a queue used to store the statistics for all peer structs.
6663 * Its contents are the union of all the peer rpcStats queues.
6666 static struct rx_queue peerStats = { &peerStats, &peerStats };
6669 * rxi_monitor_processStats is used to turn process wide stat collection
6673 static int rxi_monitor_processStats = 0;
6676 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6679 static int rxi_monitor_peerStats = 0;
6682 * rxi_AddRpcStat - given all of the information for a particular rpc
6683 * call, create (if needed) and update the stat totals for the rpc.
6687 * IN stats - the queue of stats that will be updated with the new value
6689 * IN rxInterface - a unique number that identifies the rpc interface
6691 * IN currentFunc - the index of the function being invoked
6693 * IN totalFunc - the total number of functions in this interface
6695 * IN queueTime - the amount of time this function waited for a thread
6697 * IN execTime - the amount of time this function invocation took to execute
6699 * IN bytesSent - the number bytes sent by this invocation
6701 * IN bytesRcvd - the number bytes received by this invocation
6703 * IN isServer - if true, this invocation was made to a server
6705 * IN remoteHost - the ip address of the remote host
6707 * IN remotePort - the port of the remote host
6709 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6711 * INOUT counter - if a new stats structure is allocated, the counter will
6712 * be updated with the new number of allocated stat structures
6720 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6721 afs_uint32 currentFunc, afs_uint32 totalFunc,
6722 struct clock *queueTime, struct clock *execTime,
6723 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6724 afs_uint32 remoteHost, afs_uint32 remotePort,
6725 int addToPeerList, unsigned int *counter)
6728 rx_interface_stat_p rpc_stat, nrpc_stat;
6731 * See if there's already a structure for this interface
6734 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6735 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6736 && (rpc_stat->stats[0].remote_is_server == isServer))
6741 * Didn't find a match so allocate a new structure and add it to the
6745 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6746 || (rpc_stat->stats[0].interfaceId != rxInterface)
6747 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6752 sizeof(rx_interface_stat_t) +
6753 totalFunc * sizeof(rx_function_entry_v1_t);
6755 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6756 if (rpc_stat == NULL) {
6760 *counter += totalFunc;
6761 for (i = 0; i < totalFunc; i++) {
6762 rpc_stat->stats[i].remote_peer = remoteHost;
6763 rpc_stat->stats[i].remote_port = remotePort;
6764 rpc_stat->stats[i].remote_is_server = isServer;
6765 rpc_stat->stats[i].interfaceId = rxInterface;
6766 rpc_stat->stats[i].func_total = totalFunc;
6767 rpc_stat->stats[i].func_index = i;
6768 hzero(rpc_stat->stats[i].invocations);
6769 hzero(rpc_stat->stats[i].bytes_sent);
6770 hzero(rpc_stat->stats[i].bytes_rcvd);
6771 rpc_stat->stats[i].queue_time_sum.sec = 0;
6772 rpc_stat->stats[i].queue_time_sum.usec = 0;
6773 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6774 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6775 rpc_stat->stats[i].queue_time_min.sec = 9999999;
6776 rpc_stat->stats[i].queue_time_min.usec = 9999999;
6777 rpc_stat->stats[i].queue_time_max.sec = 0;
6778 rpc_stat->stats[i].queue_time_max.usec = 0;
6779 rpc_stat->stats[i].execution_time_sum.sec = 0;
6780 rpc_stat->stats[i].execution_time_sum.usec = 0;
6781 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
6782 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
6783 rpc_stat->stats[i].execution_time_min.sec = 9999999;
6784 rpc_stat->stats[i].execution_time_min.usec = 9999999;
6785 rpc_stat->stats[i].execution_time_max.sec = 0;
6786 rpc_stat->stats[i].execution_time_max.usec = 0;
6788 queue_Prepend(stats, rpc_stat);
6789 if (addToPeerList) {
6790 queue_Prepend(&peerStats, &rpc_stat->all_peers);
6795 * Increment the stats for this function
6798 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
6799 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
6800 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
6801 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
6802 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
6803 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
6804 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
6806 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
6807 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
6809 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
6810 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
6812 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
6813 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
6815 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
6816 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
6824 * rx_IncrementTimeAndCount - increment the times and count for a particular
6829 * IN peer - the peer who invoked the rpc
6831 * IN rxInterface - a unique number that identifies the rpc interface
6833 * IN currentFunc - the index of the function being invoked
6835 * IN totalFunc - the total number of functions in this interface
6837 * IN queueTime - the amount of time this function waited for a thread
6839 * IN execTime - the amount of time this function invocation took to execute
6841 * IN bytesSent - the number bytes sent by this invocation
6843 * IN bytesRcvd - the number bytes received by this invocation
6845 * IN isServer - if true, this invocation was made to a server
6853 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
6854 afs_uint32 currentFunc, afs_uint32 totalFunc,
6855 struct clock *queueTime, struct clock *execTime,
6856 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
6860 MUTEX_ENTER(&rx_rpc_stats);
6861 MUTEX_ENTER(&peer->peer_lock);
6863 if (rxi_monitor_peerStats) {
6864 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
6865 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6866 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
6869 if (rxi_monitor_processStats) {
6870 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
6871 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6872 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
6875 MUTEX_EXIT(&peer->peer_lock);
6876 MUTEX_EXIT(&rx_rpc_stats);
6881 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
6885 * IN callerVersion - the rpc stat version of the caller.
6887 * IN count - the number of entries to marshall.
6889 * IN stats - pointer to stats to be marshalled.
6891 * OUT ptr - Where to store the marshalled data.
6898 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
6899 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
6905 * We only support the first version
6907 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
6908 *(ptr++) = stats->remote_peer;
6909 *(ptr++) = stats->remote_port;
6910 *(ptr++) = stats->remote_is_server;
6911 *(ptr++) = stats->interfaceId;
6912 *(ptr++) = stats->func_total;
6913 *(ptr++) = stats->func_index;
6914 *(ptr++) = hgethi(stats->invocations);
6915 *(ptr++) = hgetlo(stats->invocations);
6916 *(ptr++) = hgethi(stats->bytes_sent);
6917 *(ptr++) = hgetlo(stats->bytes_sent);
6918 *(ptr++) = hgethi(stats->bytes_rcvd);
6919 *(ptr++) = hgetlo(stats->bytes_rcvd);
6920 *(ptr++) = stats->queue_time_sum.sec;
6921 *(ptr++) = stats->queue_time_sum.usec;
6922 *(ptr++) = stats->queue_time_sum_sqr.sec;
6923 *(ptr++) = stats->queue_time_sum_sqr.usec;
6924 *(ptr++) = stats->queue_time_min.sec;
6925 *(ptr++) = stats->queue_time_min.usec;
6926 *(ptr++) = stats->queue_time_max.sec;
6927 *(ptr++) = stats->queue_time_max.usec;
6928 *(ptr++) = stats->execution_time_sum.sec;
6929 *(ptr++) = stats->execution_time_sum.usec;
6930 *(ptr++) = stats->execution_time_sum_sqr.sec;
6931 *(ptr++) = stats->execution_time_sum_sqr.usec;
6932 *(ptr++) = stats->execution_time_min.sec;
6933 *(ptr++) = stats->execution_time_min.usec;
6934 *(ptr++) = stats->execution_time_max.sec;
6935 *(ptr++) = stats->execution_time_max.usec;
6941 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
6946 * IN callerVersion - the rpc stat version of the caller
6948 * OUT myVersion - the rpc stat version of this function
6950 * OUT clock_sec - local time seconds
6952 * OUT clock_usec - local time microseconds
6954 * OUT allocSize - the number of bytes allocated to contain stats
6956 * OUT statCount - the number stats retrieved from this process.
6958 * OUT stats - the actual stats retrieved from this process.
6962 * Returns void. If successful, stats will != NULL.
6966 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
6967 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
6968 size_t * allocSize, afs_uint32 * statCount,
6969 afs_uint32 ** stats)
6979 *myVersion = RX_STATS_RETRIEVAL_VERSION;
6982 * Check to see if stats are enabled
6985 MUTEX_ENTER(&rx_rpc_stats);
6986 if (!rxi_monitor_processStats) {
6987 MUTEX_EXIT(&rx_rpc_stats);
6991 clock_GetTime(&now);
6992 *clock_sec = now.sec;
6993 *clock_usec = now.usec;
6996 * Allocate the space based upon the caller version
6998 * If the client is at an older version than we are,
6999 * we return the statistic data in the older data format, but
7000 * we still return our version number so the client knows we
7001 * are maintaining more data than it can retrieve.
7004 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7005 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7006 *statCount = rxi_rpc_process_stat_cnt;
7009 * This can't happen yet, but in the future version changes
7010 * can be handled by adding additional code here
7014 if (space > (size_t) 0) {
7016 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7019 rx_interface_stat_p rpc_stat, nrpc_stat;
7023 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7025 * Copy the data based upon the caller version
7027 rx_MarshallProcessRPCStats(callerVersion,
7028 rpc_stat->stats[0].func_total,
7029 rpc_stat->stats, &ptr);
7035 MUTEX_EXIT(&rx_rpc_stats);
7040 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7044 * IN callerVersion - the rpc stat version of the caller
7046 * OUT myVersion - the rpc stat version of this function
7048 * OUT clock_sec - local time seconds
7050 * OUT clock_usec - local time microseconds
7052 * OUT allocSize - the number of bytes allocated to contain stats
7054 * OUT statCount - the number of stats retrieved from the individual
7057 * OUT stats - the actual stats retrieved from the individual peer structures.
7061 * Returns void. If successful, stats will != NULL.
7065 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7066 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7067 size_t * allocSize, afs_uint32 * statCount,
7068 afs_uint32 ** stats)
7078 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7081 * Check to see if stats are enabled
7084 MUTEX_ENTER(&rx_rpc_stats);
7085 if (!rxi_monitor_peerStats) {
7086 MUTEX_EXIT(&rx_rpc_stats);
7090 clock_GetTime(&now);
7091 *clock_sec = now.sec;
7092 *clock_usec = now.usec;
7095 * Allocate the space based upon the caller version
7097 * If the client is at an older version than we are,
7098 * we return the statistic data in the older data format, but
7099 * we still return our version number so the client knows we
7100 * are maintaining more data than it can retrieve.
7103 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7104 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7105 *statCount = rxi_rpc_peer_stat_cnt;
7108 * This can't happen yet, but in the future version changes
7109 * can be handled by adding additional code here
7113 if (space > (size_t) 0) {
7115 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7118 rx_interface_stat_p rpc_stat, nrpc_stat;
7122 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7124 * We have to fix the offset of rpc_stat since we are
7125 * keeping this structure on two rx_queues. The rx_queue
7126 * package assumes that the rx_queue member is the first
7127 * member of the structure. That is, rx_queue assumes that
7128 * any one item is only on one queue at a time. We are
7129 * breaking that assumption and so we have to do a little
7130 * math to fix our pointers.
7133 fix_offset = (char *)rpc_stat;
7134 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7135 rpc_stat = (rx_interface_stat_p) fix_offset;
7138 * Copy the data based upon the caller version
7140 rx_MarshallProcessRPCStats(callerVersion,
7141 rpc_stat->stats[0].func_total,
7142 rpc_stat->stats, &ptr);
7148 MUTEX_EXIT(&rx_rpc_stats);
7153 * rx_FreeRPCStats - free memory allocated by
7154 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7158 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7159 * rx_RetrievePeerRPCStats
7161 * IN allocSize - the number of bytes in stats.
7169 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7171 rxi_Free(stats, allocSize);
7175 * rx_queryProcessRPCStats - see if process rpc stat collection is
7176 * currently enabled.
7182 * Returns 0 if stats are not enabled != 0 otherwise
7186 rx_queryProcessRPCStats(void)
7189 MUTEX_ENTER(&rx_rpc_stats);
7190 rc = rxi_monitor_processStats;
7191 MUTEX_EXIT(&rx_rpc_stats);
7196 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7202 * Returns 0 if stats are not enabled != 0 otherwise
7206 rx_queryPeerRPCStats(void)
7209 MUTEX_ENTER(&rx_rpc_stats);
7210 rc = rxi_monitor_peerStats;
7211 MUTEX_EXIT(&rx_rpc_stats);
7216 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7226 rx_enableProcessRPCStats(void)
7228 MUTEX_ENTER(&rx_rpc_stats);
7229 rx_enable_stats = 1;
7230 rxi_monitor_processStats = 1;
7231 MUTEX_EXIT(&rx_rpc_stats);
7235 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7245 rx_enablePeerRPCStats(void)
7247 MUTEX_ENTER(&rx_rpc_stats);
7248 rx_enable_stats = 1;
7249 rxi_monitor_peerStats = 1;
7250 MUTEX_EXIT(&rx_rpc_stats);
7254 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7264 rx_disableProcessRPCStats(void)
7266 rx_interface_stat_p rpc_stat, nrpc_stat;
7269 MUTEX_ENTER(&rx_rpc_stats);
7272 * Turn off process statistics and if peer stats is also off, turn
7276 rxi_monitor_processStats = 0;
7277 if (rxi_monitor_peerStats == 0) {
7278 rx_enable_stats = 0;
7281 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7282 unsigned int num_funcs = 0;
7285 queue_Remove(rpc_stat);
7286 num_funcs = rpc_stat->stats[0].func_total;
7288 sizeof(rx_interface_stat_t) +
7289 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7291 rxi_Free(rpc_stat, space);
7292 rxi_rpc_process_stat_cnt -= num_funcs;
7294 MUTEX_EXIT(&rx_rpc_stats);
7298 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7308 rx_disablePeerRPCStats(void)
7310 struct rx_peer **peer_ptr, **peer_end;
7313 MUTEX_ENTER(&rx_rpc_stats);
7316 * Turn off peer statistics and if process stats is also off, turn
7320 rxi_monitor_peerStats = 0;
7321 if (rxi_monitor_processStats == 0) {
7322 rx_enable_stats = 0;
7325 MUTEX_ENTER(&rx_peerHashTable_lock);
7326 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7327 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7329 struct rx_peer *peer, *next, *prev;
7330 for (prev = peer = *peer_ptr; peer; peer = next) {
7332 code = MUTEX_TRYENTER(&peer->peer_lock);
7334 rx_interface_stat_p rpc_stat, nrpc_stat;
7337 (&peer->rpcStats, rpc_stat, nrpc_stat,
7338 rx_interface_stat)) {
7339 unsigned int num_funcs = 0;
7342 queue_Remove(&rpc_stat->queue_header);
7343 queue_Remove(&rpc_stat->all_peers);
7344 num_funcs = rpc_stat->stats[0].func_total;
7346 sizeof(rx_interface_stat_t) +
7347 rpc_stat->stats[0].func_total *
7348 sizeof(rx_function_entry_v1_t);
7350 rxi_Free(rpc_stat, space);
7351 rxi_rpc_peer_stat_cnt -= num_funcs;
7353 MUTEX_EXIT(&peer->peer_lock);
7354 if (prev == *peer_ptr) {
7364 MUTEX_EXIT(&rx_peerHashTable_lock);
7365 MUTEX_EXIT(&rx_rpc_stats);
7369 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7374 * IN clearFlag - flag indicating which stats to clear
7382 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7384 rx_interface_stat_p rpc_stat, nrpc_stat;
7386 MUTEX_ENTER(&rx_rpc_stats);
7388 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7389 unsigned int num_funcs = 0, i;
7390 num_funcs = rpc_stat->stats[0].func_total;
7391 for (i = 0; i < num_funcs; i++) {
7392 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7393 hzero(rpc_stat->stats[i].invocations);
7395 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7396 hzero(rpc_stat->stats[i].bytes_sent);
7398 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7399 hzero(rpc_stat->stats[i].bytes_rcvd);
7401 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7402 rpc_stat->stats[i].queue_time_sum.sec = 0;
7403 rpc_stat->stats[i].queue_time_sum.usec = 0;
7405 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7406 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7407 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7409 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7410 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7411 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7413 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7414 rpc_stat->stats[i].queue_time_max.sec = 0;
7415 rpc_stat->stats[i].queue_time_max.usec = 0;
7417 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7418 rpc_stat->stats[i].execution_time_sum.sec = 0;
7419 rpc_stat->stats[i].execution_time_sum.usec = 0;
7421 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7422 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7423 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7425 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7426 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7427 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7429 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7430 rpc_stat->stats[i].execution_time_max.sec = 0;
7431 rpc_stat->stats[i].execution_time_max.usec = 0;
7436 MUTEX_EXIT(&rx_rpc_stats);
7440 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7445 * IN clearFlag - flag indicating which stats to clear
7453 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7455 rx_interface_stat_p rpc_stat, nrpc_stat;
7457 MUTEX_ENTER(&rx_rpc_stats);
7459 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7460 unsigned int num_funcs = 0, i;
7463 * We have to fix the offset of rpc_stat since we are
7464 * keeping this structure on two rx_queues. The rx_queue
7465 * package assumes that the rx_queue member is the first
7466 * member of the structure. That is, rx_queue assumes that
7467 * any one item is only on one queue at a time. We are
7468 * breaking that assumption and so we have to do a little
7469 * math to fix our pointers.
7472 fix_offset = (char *)rpc_stat;
7473 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7474 rpc_stat = (rx_interface_stat_p) fix_offset;
7476 num_funcs = rpc_stat->stats[0].func_total;
7477 for (i = 0; i < num_funcs; i++) {
7478 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7479 hzero(rpc_stat->stats[i].invocations);
7481 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7482 hzero(rpc_stat->stats[i].bytes_sent);
7484 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7485 hzero(rpc_stat->stats[i].bytes_rcvd);
7487 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7488 rpc_stat->stats[i].queue_time_sum.sec = 0;
7489 rpc_stat->stats[i].queue_time_sum.usec = 0;
7491 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7492 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7493 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7495 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7496 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7497 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7499 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7500 rpc_stat->stats[i].queue_time_max.sec = 0;
7501 rpc_stat->stats[i].queue_time_max.usec = 0;
7503 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7504 rpc_stat->stats[i].execution_time_sum.sec = 0;
7505 rpc_stat->stats[i].execution_time_sum.usec = 0;
7507 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7508 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7509 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7511 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7512 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7513 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7515 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7516 rpc_stat->stats[i].execution_time_max.sec = 0;
7517 rpc_stat->stats[i].execution_time_max.usec = 0;
7522 MUTEX_EXIT(&rx_rpc_stats);
7526 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7527 * is authorized to enable/disable/clear RX statistics.
7529 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7532 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7534 rxi_rxstat_userok = proc;
7538 rx_RxStatUserOk(struct rx_call *call)
7540 if (!rxi_rxstat_userok)
7542 return rxi_rxstat_userok(call);