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 <= 0) {
821 conn->peer->idleWhen = clock_Sec();
822 if (conn->peer->refCount < 0) {
823 conn->peer->refCount = 0;
824 MUTEX_ENTER(&rx_stats_mutex);
825 rxi_lowPeerRefCount++;
826 MUTEX_EXIT(&rx_stats_mutex);
829 MUTEX_EXIT(&rx_peerHashTable_lock);
831 MUTEX_ENTER(&rx_stats_mutex);
832 if (conn->type == RX_SERVER_CONNECTION)
833 rx_stats.nServerConns--;
835 rx_stats.nClientConns--;
836 MUTEX_EXIT(&rx_stats_mutex);
839 if (conn->specific) {
841 for (i = 0; i < conn->nSpecific; i++) {
842 if (conn->specific[i] && rxi_keyCreate_destructor[i])
843 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
844 conn->specific[i] = NULL;
846 free(conn->specific);
848 conn->specific = NULL;
852 MUTEX_DESTROY(&conn->conn_call_lock);
853 MUTEX_DESTROY(&conn->conn_data_lock);
854 CV_DESTROY(&conn->conn_call_cv);
856 rxi_FreeConnection(conn);
859 /* Destroy the specified connection */
861 rxi_DestroyConnection(register struct rx_connection *conn)
863 MUTEX_ENTER(&rx_connHashTable_lock);
864 rxi_DestroyConnectionNoLock(conn);
865 /* conn should be at the head of the cleanup list */
866 if (conn == rx_connCleanup_list) {
867 rx_connCleanup_list = rx_connCleanup_list->next;
868 MUTEX_EXIT(&rx_connHashTable_lock);
869 rxi_CleanupConnection(conn);
871 #ifdef RX_ENABLE_LOCKS
873 MUTEX_EXIT(&rx_connHashTable_lock);
875 #endif /* RX_ENABLE_LOCKS */
879 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
881 register struct rx_connection **conn_ptr;
882 register int havecalls = 0;
883 struct rx_packet *packet;
890 MUTEX_ENTER(&conn->conn_data_lock);
891 if (conn->refCount > 0)
894 MUTEX_ENTER(&rx_stats_mutex);
895 rxi_lowConnRefCount++;
896 MUTEX_EXIT(&rx_stats_mutex);
899 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
900 /* Busy; wait till the last guy before proceeding */
901 MUTEX_EXIT(&conn->conn_data_lock);
906 /* If the client previously called rx_NewCall, but it is still
907 * waiting, treat this as a running call, and wait to destroy the
908 * connection later when the call completes. */
909 if ((conn->type == RX_CLIENT_CONNECTION)
910 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
911 conn->flags |= RX_CONN_DESTROY_ME;
912 MUTEX_EXIT(&conn->conn_data_lock);
916 MUTEX_EXIT(&conn->conn_data_lock);
918 /* Check for extant references to this connection */
919 for (i = 0; i < RX_MAXCALLS; i++) {
920 register struct rx_call *call = conn->call[i];
923 if (conn->type == RX_CLIENT_CONNECTION) {
924 MUTEX_ENTER(&call->lock);
925 if (call->delayedAckEvent) {
926 /* Push the final acknowledgment out now--there
927 * won't be a subsequent call to acknowledge the
928 * last reply packets */
929 rxevent_Cancel(call->delayedAckEvent, call,
930 RX_CALL_REFCOUNT_DELAY);
931 if (call->state == RX_STATE_PRECALL
932 || call->state == RX_STATE_ACTIVE) {
933 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
935 rxi_AckAll(NULL, call, 0);
938 MUTEX_EXIT(&call->lock);
942 #ifdef RX_ENABLE_LOCKS
944 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
945 MUTEX_EXIT(&conn->conn_data_lock);
947 /* Someone is accessing a packet right now. */
951 #endif /* RX_ENABLE_LOCKS */
954 /* Don't destroy the connection if there are any call
955 * structures still in use */
956 MUTEX_ENTER(&conn->conn_data_lock);
957 conn->flags |= RX_CONN_DESTROY_ME;
958 MUTEX_EXIT(&conn->conn_data_lock);
963 if (conn->delayedAbortEvent) {
964 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
965 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
967 MUTEX_ENTER(&conn->conn_data_lock);
968 rxi_SendConnectionAbort(conn, packet, 0, 1);
969 MUTEX_EXIT(&conn->conn_data_lock);
970 rxi_FreePacket(packet);
974 /* Remove from connection hash table before proceeding */
976 &rx_connHashTable[CONN_HASH
977 (peer->host, peer->port, conn->cid, conn->epoch,
979 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
980 if (*conn_ptr == conn) {
981 *conn_ptr = conn->next;
985 /* if the conn that we are destroying was the last connection, then we
986 * clear rxLastConn as well */
987 if (rxLastConn == conn)
990 /* Make sure the connection is completely reset before deleting it. */
991 /* get rid of pending events that could zap us later */
992 if (conn->challengeEvent)
993 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
994 if (conn->checkReachEvent)
995 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
997 /* Add the connection to the list of destroyed connections that
998 * need to be cleaned up. This is necessary to avoid deadlocks
999 * in the routines we call to inform others that this connection is
1000 * being destroyed. */
1001 conn->next = rx_connCleanup_list;
1002 rx_connCleanup_list = conn;
1005 /* Externally available version */
1007 rx_DestroyConnection(register struct rx_connection *conn)
1013 rxi_DestroyConnection(conn);
1018 /* Start a new rx remote procedure call, on the specified connection.
1019 * If wait is set to 1, wait for a free call channel; otherwise return
1020 * 0. Maxtime gives the maximum number of seconds this call may take,
1021 * after rx_MakeCall returns. After this time interval, a call to any
1022 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1023 * For fine grain locking, we hold the conn_call_lock in order to
1024 * to ensure that we don't get signalle after we found a call in an active
1025 * state and before we go to sleep.
1028 rx_NewCall(register struct rx_connection *conn)
1031 register struct rx_call *call;
1032 struct clock queueTime;
1036 dpf(("rx_MakeCall(conn %x)\n", conn));
1039 clock_GetTime(&queueTime);
1041 MUTEX_ENTER(&conn->conn_call_lock);
1044 * Check if there are others waiting for a new call.
1045 * If so, let them go first to avoid starving them.
1046 * This is a fairly simple scheme, and might not be
1047 * a complete solution for large numbers of waiters.
1049 if (conn->makeCallWaiters) {
1050 #ifdef RX_ENABLE_LOCKS
1051 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1058 for (i = 0; i < RX_MAXCALLS; i++) {
1059 call = conn->call[i];
1061 MUTEX_ENTER(&call->lock);
1062 if (call->state == RX_STATE_DALLY) {
1063 rxi_ResetCall(call, 0);
1064 (*call->callNumber)++;
1067 MUTEX_EXIT(&call->lock);
1069 call = rxi_NewCall(conn, i);
1073 if (i < RX_MAXCALLS) {
1076 MUTEX_ENTER(&conn->conn_data_lock);
1077 conn->flags |= RX_CONN_MAKECALL_WAITING;
1078 MUTEX_EXIT(&conn->conn_data_lock);
1080 conn->makeCallWaiters++;
1081 #ifdef RX_ENABLE_LOCKS
1082 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1086 conn->makeCallWaiters--;
1089 * Wake up anyone else who might be giving us a chance to
1090 * run (see code above that avoids resource starvation).
1092 #ifdef RX_ENABLE_LOCKS
1093 CV_BROADCAST(&conn->conn_call_cv);
1098 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1100 /* Client is initially in send mode */
1101 call->state = RX_STATE_ACTIVE;
1102 call->mode = RX_MODE_SENDING;
1104 /* remember start time for call in case we have hard dead time limit */
1105 call->queueTime = queueTime;
1106 clock_GetTime(&call->startTime);
1107 hzero(call->bytesSent);
1108 hzero(call->bytesRcvd);
1110 /* Turn on busy protocol. */
1111 rxi_KeepAliveOn(call);
1113 MUTEX_EXIT(&call->lock);
1114 MUTEX_EXIT(&conn->conn_call_lock);
1118 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1119 /* Now, if TQ wasn't cleared earlier, do it now. */
1121 MUTEX_ENTER(&call->lock);
1122 while (call->flags & RX_CALL_TQ_BUSY) {
1123 call->flags |= RX_CALL_TQ_WAIT;
1124 #ifdef RX_ENABLE_LOCKS
1125 CV_WAIT(&call->cv_tq, &call->lock);
1126 #else /* RX_ENABLE_LOCKS */
1127 osi_rxSleep(&call->tq);
1128 #endif /* RX_ENABLE_LOCKS */
1130 if (call->flags & RX_CALL_TQ_CLEARME) {
1131 rxi_ClearTransmitQueue(call, 0);
1132 queue_Init(&call->tq);
1134 MUTEX_EXIT(&call->lock);
1136 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1142 rxi_HasActiveCalls(register struct rx_connection *aconn)
1145 register struct rx_call *tcall;
1149 for (i = 0; i < RX_MAXCALLS; i++) {
1150 if ((tcall = aconn->call[i])) {
1151 if ((tcall->state == RX_STATE_ACTIVE)
1152 || (tcall->state == RX_STATE_PRECALL)) {
1163 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1164 register afs_int32 * aint32s)
1167 register struct rx_call *tcall;
1171 for (i = 0; i < RX_MAXCALLS; i++) {
1172 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1173 aint32s[i] = aconn->callNumber[i] + 1;
1175 aint32s[i] = aconn->callNumber[i];
1182 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1183 register afs_int32 * aint32s)
1186 register struct rx_call *tcall;
1190 for (i = 0; i < RX_MAXCALLS; i++) {
1191 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1192 aconn->callNumber[i] = aint32s[i] - 1;
1194 aconn->callNumber[i] = aint32s[i];
1200 /* Advertise a new service. A service is named locally by a UDP port
1201 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1204 char *serviceName; Name for identification purposes (e.g. the
1205 service name might be used for probing for
1208 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1209 struct rx_securityClass **securityObjects, int nSecurityObjects,
1210 afs_int32(*serviceProc) (struct rx_call * acall))
1212 osi_socket socket = OSI_NULLSOCKET;
1213 register struct rx_service *tservice;
1219 if (serviceId == 0) {
1221 "rx_NewService: service id for service %s is not non-zero.\n",
1228 "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",
1236 tservice = rxi_AllocService();
1239 for (i = 0; i < RX_MAX_SERVICES; i++) {
1240 register struct rx_service *service = rx_services[i];
1242 if (port == service->servicePort) {
1243 if (service->serviceId == serviceId) {
1244 /* The identical service has already been
1245 * installed; if the caller was intending to
1246 * change the security classes used by this
1247 * service, he/she loses. */
1249 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1250 serviceName, serviceId, service->serviceName);
1253 rxi_FreeService(tservice);
1256 /* Different service, same port: re-use the socket
1257 * which is bound to the same port */
1258 socket = service->socket;
1261 if (socket == OSI_NULLSOCKET) {
1262 /* If we don't already have a socket (from another
1263 * service on same port) get a new one */
1264 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1265 if (socket == OSI_NULLSOCKET) {
1268 rxi_FreeService(tservice);
1273 service->socket = socket;
1274 service->servicePort = port;
1275 service->serviceId = serviceId;
1276 service->serviceName = serviceName;
1277 service->nSecurityObjects = nSecurityObjects;
1278 service->securityObjects = securityObjects;
1279 service->minProcs = 0;
1280 service->maxProcs = 1;
1281 service->idleDeadTime = 60;
1282 service->connDeadTime = rx_connDeadTime;
1283 service->executeRequestProc = serviceProc;
1284 service->checkReach = 0;
1285 rx_services[i] = service; /* not visible until now */
1293 rxi_FreeService(tservice);
1294 (osi_Msg "rx_NewService: cannot support > %d services\n",
1299 /* Generic request processing loop. This routine should be called
1300 * by the implementation dependent rx_ServerProc. If socketp is
1301 * non-null, it will be set to the file descriptor that this thread
1302 * is now listening on. If socketp is null, this routine will never
1305 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1307 register struct rx_call *call;
1308 register afs_int32 code;
1309 register struct rx_service *tservice = NULL;
1316 call = rx_GetCall(threadID, tservice, socketp);
1317 if (socketp && *socketp != OSI_NULLSOCKET) {
1318 /* We are now a listener thread */
1323 /* if server is restarting( typically smooth shutdown) then do not
1324 * allow any new calls.
1327 if (rx_tranquil && (call != NULL)) {
1332 MUTEX_ENTER(&call->lock);
1334 rxi_CallError(call, RX_RESTARTING);
1335 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1337 MUTEX_EXIT(&call->lock);
1342 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1343 #ifdef RX_ENABLE_LOCKS
1345 #endif /* RX_ENABLE_LOCKS */
1346 afs_termState = AFSOP_STOP_AFS;
1347 afs_osi_Wakeup(&afs_termState);
1348 #ifdef RX_ENABLE_LOCKS
1350 #endif /* RX_ENABLE_LOCKS */
1355 tservice = call->conn->service;
1357 if (tservice->beforeProc)
1358 (*tservice->beforeProc) (call);
1360 code = call->conn->service->executeRequestProc(call);
1362 if (tservice->afterProc)
1363 (*tservice->afterProc) (call, code);
1365 rx_EndCall(call, code);
1366 MUTEX_ENTER(&rx_stats_mutex);
1368 MUTEX_EXIT(&rx_stats_mutex);
1374 rx_WakeupServerProcs(void)
1376 struct rx_serverQueueEntry *np, *tqp;
1381 MUTEX_ENTER(&rx_serverPool_lock);
1383 #ifdef RX_ENABLE_LOCKS
1384 if (rx_waitForPacket)
1385 CV_BROADCAST(&rx_waitForPacket->cv);
1386 #else /* RX_ENABLE_LOCKS */
1387 if (rx_waitForPacket)
1388 osi_rxWakeup(rx_waitForPacket);
1389 #endif /* RX_ENABLE_LOCKS */
1390 MUTEX_ENTER(&freeSQEList_lock);
1391 for (np = rx_FreeSQEList; np; np = tqp) {
1392 tqp = *(struct rx_serverQueueEntry **)np;
1393 #ifdef RX_ENABLE_LOCKS
1394 CV_BROADCAST(&np->cv);
1395 #else /* RX_ENABLE_LOCKS */
1397 #endif /* RX_ENABLE_LOCKS */
1399 MUTEX_EXIT(&freeSQEList_lock);
1400 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1401 #ifdef RX_ENABLE_LOCKS
1402 CV_BROADCAST(&np->cv);
1403 #else /* RX_ENABLE_LOCKS */
1405 #endif /* RX_ENABLE_LOCKS */
1407 MUTEX_EXIT(&rx_serverPool_lock);
1413 * One thing that seems to happen is that all the server threads get
1414 * tied up on some empty or slow call, and then a whole bunch of calls
1415 * arrive at once, using up the packet pool, so now there are more
1416 * empty calls. The most critical resources here are server threads
1417 * and the free packet pool. The "doreclaim" code seems to help in
1418 * general. I think that eventually we arrive in this state: there
1419 * are lots of pending calls which do have all their packets present,
1420 * so they won't be reclaimed, are multi-packet calls, so they won't
1421 * be scheduled until later, and thus are tying up most of the free
1422 * packet pool for a very long time.
1424 * 1. schedule multi-packet calls if all the packets are present.
1425 * Probably CPU-bound operation, useful to return packets to pool.
1426 * Do what if there is a full window, but the last packet isn't here?
1427 * 3. preserve one thread which *only* runs "best" calls, otherwise
1428 * it sleeps and waits for that type of call.
1429 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1430 * the current dataquota business is badly broken. The quota isn't adjusted
1431 * to reflect how many packets are presently queued for a running call.
1432 * So, when we schedule a queued call with a full window of packets queued
1433 * up for it, that *should* free up a window full of packets for other 2d-class
1434 * calls to be able to use from the packet pool. But it doesn't.
1436 * NB. Most of the time, this code doesn't run -- since idle server threads
1437 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1438 * as a new call arrives.
1440 /* Sleep until a call arrives. Returns a pointer to the call, ready
1441 * for an rx_Read. */
1442 #ifdef RX_ENABLE_LOCKS
1444 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1446 struct rx_serverQueueEntry *sq;
1447 register struct rx_call *call = (struct rx_call *)0;
1448 struct rx_service *service = NULL;
1451 MUTEX_ENTER(&freeSQEList_lock);
1453 if ((sq = rx_FreeSQEList)) {
1454 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1455 MUTEX_EXIT(&freeSQEList_lock);
1456 } else { /* otherwise allocate a new one and return that */
1457 MUTEX_EXIT(&freeSQEList_lock);
1458 sq = (struct rx_serverQueueEntry *)
1459 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1460 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1461 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1464 MUTEX_ENTER(&rx_serverPool_lock);
1465 if (cur_service != NULL) {
1466 ReturnToServerPool(cur_service);
1469 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1470 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1472 /* Scan for eligible incoming calls. A call is not eligible
1473 * if the maximum number of calls for its service type are
1474 * already executing */
1475 /* One thread will process calls FCFS (to prevent starvation),
1476 * while the other threads may run ahead looking for calls which
1477 * have all their input data available immediately. This helps
1478 * keep threads from blocking, waiting for data from the client. */
1479 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1480 service = tcall->conn->service;
1481 if (!QuotaOK(service)) {
1484 if (tno == rxi_fcfs_thread_num
1485 || !tcall->queue_item_header.next) {
1486 /* If we're the fcfs thread , then we'll just use
1487 * this call. If we haven't been able to find an optimal
1488 * choice, and we're at the end of the list, then use a
1489 * 2d choice if one has been identified. Otherwise... */
1490 call = (choice2 ? choice2 : tcall);
1491 service = call->conn->service;
1492 } else if (!queue_IsEmpty(&tcall->rq)) {
1493 struct rx_packet *rp;
1494 rp = queue_First(&tcall->rq, rx_packet);
1495 if (rp->header.seq == 1) {
1497 || (rp->header.flags & RX_LAST_PACKET)) {
1499 } else if (rxi_2dchoice && !choice2
1500 && !(tcall->flags & RX_CALL_CLEARED)
1501 && (tcall->rprev > rxi_HardAckRate)) {
1510 ReturnToServerPool(service);
1517 MUTEX_EXIT(&rx_serverPool_lock);
1518 MUTEX_ENTER(&call->lock);
1520 if (call->flags & RX_CALL_WAIT_PROC) {
1521 call->flags &= ~RX_CALL_WAIT_PROC;
1522 MUTEX_ENTER(&rx_stats_mutex);
1524 MUTEX_EXIT(&rx_stats_mutex);
1527 if (call->state != RX_STATE_PRECALL || call->error) {
1528 MUTEX_EXIT(&call->lock);
1529 MUTEX_ENTER(&rx_serverPool_lock);
1530 ReturnToServerPool(service);
1535 if (queue_IsEmpty(&call->rq)
1536 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1537 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1539 CLEAR_CALL_QUEUE_LOCK(call);
1542 /* If there are no eligible incoming calls, add this process
1543 * to the idle server queue, to wait for one */
1547 *socketp = OSI_NULLSOCKET;
1549 sq->socketp = socketp;
1550 queue_Append(&rx_idleServerQueue, sq);
1551 #ifndef AFS_AIX41_ENV
1552 rx_waitForPacket = sq;
1554 rx_waitingForPacket = sq;
1555 #endif /* AFS_AIX41_ENV */
1557 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1559 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1560 MUTEX_EXIT(&rx_serverPool_lock);
1561 return (struct rx_call *)0;
1564 } while (!(call = sq->newcall)
1565 && !(socketp && *socketp != OSI_NULLSOCKET));
1566 MUTEX_EXIT(&rx_serverPool_lock);
1568 MUTEX_ENTER(&call->lock);
1574 MUTEX_ENTER(&freeSQEList_lock);
1575 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1576 rx_FreeSQEList = sq;
1577 MUTEX_EXIT(&freeSQEList_lock);
1580 clock_GetTime(&call->startTime);
1581 call->state = RX_STATE_ACTIVE;
1582 call->mode = RX_MODE_RECEIVING;
1583 #ifdef RX_KERNEL_TRACE
1584 if (ICL_SETACTIVE(afs_iclSetp)) {
1585 int glockOwner = ISAFS_GLOCK();
1588 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1589 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1596 rxi_calltrace(RX_CALL_START, call);
1597 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1598 call->conn->service->servicePort, call->conn->service->serviceId,
1601 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1602 MUTEX_EXIT(&call->lock);
1604 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1609 #else /* RX_ENABLE_LOCKS */
1611 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1613 struct rx_serverQueueEntry *sq;
1614 register struct rx_call *call = (struct rx_call *)0, *choice2;
1615 struct rx_service *service = NULL;
1620 MUTEX_ENTER(&freeSQEList_lock);
1622 if ((sq = rx_FreeSQEList)) {
1623 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1624 MUTEX_EXIT(&freeSQEList_lock);
1625 } else { /* otherwise allocate a new one and return that */
1626 MUTEX_EXIT(&freeSQEList_lock);
1627 sq = (struct rx_serverQueueEntry *)
1628 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1629 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1630 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1632 MUTEX_ENTER(&sq->lock);
1634 if (cur_service != NULL) {
1635 cur_service->nRequestsRunning--;
1636 if (cur_service->nRequestsRunning < cur_service->minProcs)
1640 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1641 register struct rx_call *tcall, *ncall;
1642 /* Scan for eligible incoming calls. A call is not eligible
1643 * if the maximum number of calls for its service type are
1644 * already executing */
1645 /* One thread will process calls FCFS (to prevent starvation),
1646 * while the other threads may run ahead looking for calls which
1647 * have all their input data available immediately. This helps
1648 * keep threads from blocking, waiting for data from the client. */
1649 choice2 = (struct rx_call *)0;
1650 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1651 service = tcall->conn->service;
1652 if (QuotaOK(service)) {
1653 if (tno == rxi_fcfs_thread_num
1654 || !tcall->queue_item_header.next) {
1655 /* If we're the fcfs thread, then we'll just use
1656 * this call. If we haven't been able to find an optimal
1657 * choice, and we're at the end of the list, then use a
1658 * 2d choice if one has been identified. Otherwise... */
1659 call = (choice2 ? choice2 : tcall);
1660 service = call->conn->service;
1661 } else if (!queue_IsEmpty(&tcall->rq)) {
1662 struct rx_packet *rp;
1663 rp = queue_First(&tcall->rq, rx_packet);
1664 if (rp->header.seq == 1
1666 || (rp->header.flags & RX_LAST_PACKET))) {
1668 } else if (rxi_2dchoice && !choice2
1669 && !(tcall->flags & RX_CALL_CLEARED)
1670 && (tcall->rprev > rxi_HardAckRate)) {
1683 /* we can't schedule a call if there's no data!!! */
1684 /* send an ack if there's no data, if we're missing the
1685 * first packet, or we're missing something between first
1686 * and last -- there's a "hole" in the incoming data. */
1687 if (queue_IsEmpty(&call->rq)
1688 || queue_First(&call->rq, rx_packet)->header.seq != 1
1689 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1690 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1692 call->flags &= (~RX_CALL_WAIT_PROC);
1693 service->nRequestsRunning++;
1694 /* just started call in minProcs pool, need fewer to maintain
1696 if (service->nRequestsRunning <= service->minProcs)
1700 /* MUTEX_EXIT(&call->lock); */
1702 /* If there are no eligible incoming calls, add this process
1703 * to the idle server queue, to wait for one */
1706 *socketp = OSI_NULLSOCKET;
1708 sq->socketp = socketp;
1709 queue_Append(&rx_idleServerQueue, sq);
1713 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1716 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1717 return (struct rx_call *)0;
1720 } while (!(call = sq->newcall)
1721 && !(socketp && *socketp != OSI_NULLSOCKET));
1723 MUTEX_EXIT(&sq->lock);
1725 MUTEX_ENTER(&freeSQEList_lock);
1726 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1727 rx_FreeSQEList = sq;
1728 MUTEX_EXIT(&freeSQEList_lock);
1731 clock_GetTime(&call->startTime);
1732 call->state = RX_STATE_ACTIVE;
1733 call->mode = RX_MODE_RECEIVING;
1734 #ifdef RX_KERNEL_TRACE
1735 if (ICL_SETACTIVE(afs_iclSetp)) {
1736 int glockOwner = ISAFS_GLOCK();
1739 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1740 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1747 rxi_calltrace(RX_CALL_START, call);
1748 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1749 call->conn->service->servicePort, call->conn->service->serviceId,
1752 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1760 #endif /* RX_ENABLE_LOCKS */
1764 /* Establish a procedure to be called when a packet arrives for a
1765 * call. This routine will be called at most once after each call,
1766 * and will also be called if there is an error condition on the or
1767 * the call is complete. Used by multi rx to build a selection
1768 * function which determines which of several calls is likely to be a
1769 * good one to read from.
1770 * NOTE: the way this is currently implemented it is probably only a
1771 * good idea to (1) use it immediately after a newcall (clients only)
1772 * and (2) only use it once. Other uses currently void your warranty
1775 rx_SetArrivalProc(register struct rx_call *call,
1776 register VOID(*proc) (register struct rx_call * call,
1777 register struct multi_handle * mh,
1778 register int index),
1779 register VOID * handle, register VOID * arg)
1781 call->arrivalProc = proc;
1782 call->arrivalProcHandle = handle;
1783 call->arrivalProcArg = arg;
1786 /* Call is finished (possibly prematurely). Return rc to the peer, if
1787 * appropriate, and return the final error code from the conversation
1791 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1793 register struct rx_connection *conn = call->conn;
1794 register struct rx_service *service;
1795 register struct rx_packet *tp; /* Temporary packet pointer */
1796 register struct rx_packet *nxp; /* Next packet pointer, for queue_Scan */
1800 dpf(("rx_EndCall(call %x)\n", call));
1804 MUTEX_ENTER(&call->lock);
1806 if (rc == 0 && call->error == 0) {
1807 call->abortCode = 0;
1808 call->abortCount = 0;
1811 call->arrivalProc = (VOID(*)())0;
1812 if (rc && call->error == 0) {
1813 rxi_CallError(call, rc);
1814 /* Send an abort message to the peer if this error code has
1815 * only just been set. If it was set previously, assume the
1816 * peer has already been sent the error code or will request it
1818 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1820 if (conn->type == RX_SERVER_CONNECTION) {
1821 /* Make sure reply or at least dummy reply is sent */
1822 if (call->mode == RX_MODE_RECEIVING) {
1823 rxi_WriteProc(call, 0, 0);
1825 if (call->mode == RX_MODE_SENDING) {
1826 rxi_FlushWrite(call);
1828 service = conn->service;
1829 rxi_calltrace(RX_CALL_END, call);
1830 /* Call goes to hold state until reply packets are acknowledged */
1831 if (call->tfirst + call->nSoftAcked < call->tnext) {
1832 call->state = RX_STATE_HOLD;
1834 call->state = RX_STATE_DALLY;
1835 rxi_ClearTransmitQueue(call, 0);
1836 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1837 rxevent_Cancel(call->keepAliveEvent, call,
1838 RX_CALL_REFCOUNT_ALIVE);
1840 } else { /* Client connection */
1842 /* Make sure server receives input packets, in the case where
1843 * no reply arguments are expected */
1844 if ((call->mode == RX_MODE_SENDING)
1845 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1846 (void)rxi_ReadProc(call, &dummy, 1);
1849 /* If we had an outstanding delayed ack, be nice to the server
1850 * and force-send it now.
1852 if (call->delayedAckEvent) {
1853 rxevent_Cancel(call->delayedAckEvent, call,
1854 RX_CALL_REFCOUNT_DELAY);
1855 call->delayedAckEvent = NULL;
1856 rxi_SendDelayedAck(NULL, call, NULL);
1859 /* We need to release the call lock since it's lower than the
1860 * conn_call_lock and we don't want to hold the conn_call_lock
1861 * over the rx_ReadProc call. The conn_call_lock needs to be held
1862 * here for the case where rx_NewCall is perusing the calls on
1863 * the connection structure. We don't want to signal until
1864 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1865 * have checked this call, found it active and by the time it
1866 * goes to sleep, will have missed the signal.
1868 MUTEX_EXIT(&call->lock);
1869 MUTEX_ENTER(&conn->conn_call_lock);
1870 MUTEX_ENTER(&call->lock);
1871 MUTEX_ENTER(&conn->conn_data_lock);
1872 conn->flags |= RX_CONN_BUSY;
1873 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1874 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1875 MUTEX_EXIT(&conn->conn_data_lock);
1876 #ifdef RX_ENABLE_LOCKS
1877 CV_BROADCAST(&conn->conn_call_cv);
1882 #ifdef RX_ENABLE_LOCKS
1884 MUTEX_EXIT(&conn->conn_data_lock);
1886 #endif /* RX_ENABLE_LOCKS */
1887 call->state = RX_STATE_DALLY;
1889 error = call->error;
1891 /* currentPacket, nLeft, and NFree must be zeroed here, because
1892 * ResetCall cannot: ResetCall may be called at splnet(), in the
1893 * kernel version, and may interrupt the macros rx_Read or
1894 * rx_Write, which run at normal priority for efficiency. */
1895 if (call->currentPacket) {
1896 rxi_FreePacket(call->currentPacket);
1897 call->currentPacket = (struct rx_packet *)0;
1898 call->nLeft = call->nFree = call->curlen = 0;
1900 call->nLeft = call->nFree = call->curlen = 0;
1902 /* Free any packets from the last call to ReadvProc/WritevProc */
1903 for (queue_Scan(&call->iovq, tp, nxp, rx_packet)) {
1908 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1909 MUTEX_EXIT(&call->lock);
1910 if (conn->type == RX_CLIENT_CONNECTION) {
1911 MUTEX_EXIT(&conn->conn_call_lock);
1912 conn->flags &= ~RX_CONN_BUSY;
1917 * Map errors to the local host's errno.h format.
1919 error = ntoh_syserr_conv(error);
1923 #if !defined(KERNEL)
1925 /* Call this routine when shutting down a server or client (especially
1926 * clients). This will allow Rx to gracefully garbage collect server
1927 * connections, and reduce the number of retries that a server might
1928 * make to a dead client.
1929 * This is not quite right, since some calls may still be ongoing and
1930 * we can't lock them to destroy them. */
1934 register struct rx_connection **conn_ptr, **conn_end;
1938 if (rxinit_status == 1) {
1940 return; /* Already shutdown. */
1942 rxi_DeleteCachedConnections();
1943 if (rx_connHashTable) {
1944 MUTEX_ENTER(&rx_connHashTable_lock);
1945 for (conn_ptr = &rx_connHashTable[0], conn_end =
1946 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
1948 struct rx_connection *conn, *next;
1949 for (conn = *conn_ptr; conn; conn = next) {
1951 if (conn->type == RX_CLIENT_CONNECTION) {
1952 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
1954 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
1955 #ifdef RX_ENABLE_LOCKS
1956 rxi_DestroyConnectionNoLock(conn);
1957 #else /* RX_ENABLE_LOCKS */
1958 rxi_DestroyConnection(conn);
1959 #endif /* RX_ENABLE_LOCKS */
1963 #ifdef RX_ENABLE_LOCKS
1964 while (rx_connCleanup_list) {
1965 struct rx_connection *conn;
1966 conn = rx_connCleanup_list;
1967 rx_connCleanup_list = rx_connCleanup_list->next;
1968 MUTEX_EXIT(&rx_connHashTable_lock);
1969 rxi_CleanupConnection(conn);
1970 MUTEX_ENTER(&rx_connHashTable_lock);
1972 MUTEX_EXIT(&rx_connHashTable_lock);
1973 #endif /* RX_ENABLE_LOCKS */
1982 /* if we wakeup packet waiter too often, can get in loop with two
1983 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
1985 rxi_PacketsUnWait(void)
1987 if (!rx_waitingForPackets) {
1991 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
1992 return; /* still over quota */
1995 rx_waitingForPackets = 0;
1996 #ifdef RX_ENABLE_LOCKS
1997 CV_BROADCAST(&rx_waitingForPackets_cv);
1999 osi_rxWakeup(&rx_waitingForPackets);
2005 /* ------------------Internal interfaces------------------------- */
2007 /* Return this process's service structure for the
2008 * specified socket and service */
2010 rxi_FindService(register osi_socket socket, register u_short serviceId)
2012 register struct rx_service **sp;
2013 for (sp = &rx_services[0]; *sp; sp++) {
2014 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2020 /* Allocate a call structure, for the indicated channel of the
2021 * supplied connection. The mode and state of the call must be set by
2022 * the caller. Returns the call with mutex locked. */
2024 rxi_NewCall(register struct rx_connection *conn, register int channel)
2026 register struct rx_call *call;
2027 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2028 register struct rx_call *cp; /* Call pointer temp */
2029 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2030 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2032 /* Grab an existing call structure, or allocate a new one.
2033 * Existing call structures are assumed to have been left reset by
2035 MUTEX_ENTER(&rx_freeCallQueue_lock);
2037 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2039 * EXCEPT that the TQ might not yet be cleared out.
2040 * Skip over those with in-use TQs.
2043 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2044 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2050 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2051 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2052 call = queue_First(&rx_freeCallQueue, rx_call);
2053 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2055 MUTEX_ENTER(&rx_stats_mutex);
2056 rx_stats.nFreeCallStructs--;
2057 MUTEX_EXIT(&rx_stats_mutex);
2058 MUTEX_EXIT(&rx_freeCallQueue_lock);
2059 MUTEX_ENTER(&call->lock);
2060 CLEAR_CALL_QUEUE_LOCK(call);
2061 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2062 /* Now, if TQ wasn't cleared earlier, do it now. */
2063 if (call->flags & RX_CALL_TQ_CLEARME) {
2064 rxi_ClearTransmitQueue(call, 0);
2065 queue_Init(&call->tq);
2067 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2068 /* Bind the call to its connection structure */
2070 rxi_ResetCall(call, 1);
2072 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2074 MUTEX_EXIT(&rx_freeCallQueue_lock);
2075 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2076 MUTEX_ENTER(&call->lock);
2077 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2078 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2079 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2081 MUTEX_ENTER(&rx_stats_mutex);
2082 rx_stats.nCallStructs++;
2083 MUTEX_EXIT(&rx_stats_mutex);
2084 /* Initialize once-only items */
2085 queue_Init(&call->tq);
2086 queue_Init(&call->rq);
2087 queue_Init(&call->iovq);
2088 /* Bind the call to its connection structure (prereq for reset) */
2090 rxi_ResetCall(call, 1);
2092 call->channel = channel;
2093 call->callNumber = &conn->callNumber[channel];
2094 /* Note that the next expected call number is retained (in
2095 * conn->callNumber[i]), even if we reallocate the call structure
2097 conn->call[channel] = call;
2098 /* if the channel's never been used (== 0), we should start at 1, otherwise
2099 * the call number is valid from the last time this channel was used */
2100 if (*call->callNumber == 0)
2101 *call->callNumber = 1;
2106 /* A call has been inactive long enough that so we can throw away
2107 * state, including the call structure, which is placed on the call
2109 * Call is locked upon entry.
2110 * haveCTLock set if called from rxi_ReapConnections
2112 #ifdef RX_ENABLE_LOCKS
2114 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2115 #else /* RX_ENABLE_LOCKS */
2117 rxi_FreeCall(register struct rx_call *call)
2118 #endif /* RX_ENABLE_LOCKS */
2120 register int channel = call->channel;
2121 register struct rx_connection *conn = call->conn;
2124 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2125 (*call->callNumber)++;
2126 rxi_ResetCall(call, 0);
2127 call->conn->call[channel] = (struct rx_call *)0;
2129 MUTEX_ENTER(&rx_freeCallQueue_lock);
2130 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2131 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2132 /* A call may be free even though its transmit queue is still in use.
2133 * Since we search the call list from head to tail, put busy calls at
2134 * the head of the list, and idle calls at the tail.
2136 if (call->flags & RX_CALL_TQ_BUSY)
2137 queue_Prepend(&rx_freeCallQueue, call);
2139 queue_Append(&rx_freeCallQueue, call);
2140 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2141 queue_Append(&rx_freeCallQueue, call);
2142 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2143 MUTEX_ENTER(&rx_stats_mutex);
2144 rx_stats.nFreeCallStructs++;
2145 MUTEX_EXIT(&rx_stats_mutex);
2147 MUTEX_EXIT(&rx_freeCallQueue_lock);
2149 /* Destroy the connection if it was previously slated for
2150 * destruction, i.e. the Rx client code previously called
2151 * rx_DestroyConnection (client connections), or
2152 * rxi_ReapConnections called the same routine (server
2153 * connections). Only do this, however, if there are no
2154 * outstanding calls. Note that for fine grain locking, there appears
2155 * to be a deadlock in that rxi_FreeCall has a call locked and
2156 * DestroyConnectionNoLock locks each call in the conn. But note a
2157 * few lines up where we have removed this call from the conn.
2158 * If someone else destroys a connection, they either have no
2159 * call lock held or are going through this section of code.
2161 if (conn->flags & RX_CONN_DESTROY_ME) {
2162 MUTEX_ENTER(&conn->conn_data_lock);
2164 MUTEX_EXIT(&conn->conn_data_lock);
2165 #ifdef RX_ENABLE_LOCKS
2167 rxi_DestroyConnectionNoLock(conn);
2169 rxi_DestroyConnection(conn);
2170 #else /* RX_ENABLE_LOCKS */
2171 rxi_DestroyConnection(conn);
2172 #endif /* RX_ENABLE_LOCKS */
2176 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2178 rxi_Alloc(register size_t size)
2182 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2183 /* Grab the AFS filesystem lock. See afs/osi.h for the lock
2186 int glockOwner = ISAFS_GLOCK();
2190 MUTEX_ENTER(&rx_stats_mutex);
2192 rxi_Allocsize += size;
2193 MUTEX_EXIT(&rx_stats_mutex);
2194 #if (defined(AFS_AIX32_ENV) || defined(AFS_HPUX_ENV)) && !defined(AFS_HPUX100_ENV) && defined(KERNEL)
2195 if (size > AFS_SMALLOCSIZ) {
2196 p = (char *)osi_AllocMediumSpace(size);
2198 p = (char *)osi_AllocSmall(size, 1);
2199 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2204 p = (char *)osi_Alloc(size);
2207 osi_Panic("rxi_Alloc error");
2213 rxi_Free(void *addr, register size_t size)
2215 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2216 /* Grab the AFS filesystem lock. See afs/osi.h for the lock
2219 int glockOwner = ISAFS_GLOCK();
2223 MUTEX_ENTER(&rx_stats_mutex);
2225 rxi_Allocsize -= size;
2226 MUTEX_EXIT(&rx_stats_mutex);
2227 #if (defined(AFS_AIX32_ENV) || defined(AFS_HPUX_ENV)) && !defined(AFS_HPUX100_ENV) && defined(KERNEL)
2228 if (size > AFS_SMALLOCSIZ)
2229 osi_FreeMediumSpace(addr);
2231 osi_FreeSmall(addr);
2232 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2237 osi_Free(addr, size);
2241 /* Find the peer process represented by the supplied (host,port)
2242 * combination. If there is no appropriate active peer structure, a
2243 * new one will be allocated and initialized
2244 * The origPeer, if set, is a pointer to a peer structure on which the
2245 * refcount will be be decremented. This is used to replace the peer
2246 * structure hanging off a connection structure */
2248 rxi_FindPeer(register afs_uint32 host, register u_short port,
2249 struct rx_peer *origPeer, int create)
2251 register struct rx_peer *pp;
2253 hashIndex = PEER_HASH(host, port);
2254 MUTEX_ENTER(&rx_peerHashTable_lock);
2255 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2256 if ((pp->host == host) && (pp->port == port))
2261 pp = rxi_AllocPeer(); /* This bzero's *pp */
2262 pp->host = host; /* set here or in InitPeerParams is zero */
2264 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2265 queue_Init(&pp->congestionQueue);
2266 queue_Init(&pp->rpcStats);
2267 pp->next = rx_peerHashTable[hashIndex];
2268 rx_peerHashTable[hashIndex] = pp;
2269 rxi_InitPeerParams(pp);
2270 MUTEX_ENTER(&rx_stats_mutex);
2271 rx_stats.nPeerStructs++;
2272 MUTEX_EXIT(&rx_stats_mutex);
2279 origPeer->refCount--;
2280 MUTEX_EXIT(&rx_peerHashTable_lock);
2285 /* Find the connection at (host, port) started at epoch, and with the
2286 * given connection id. Creates the server connection if necessary.
2287 * The type specifies whether a client connection or a server
2288 * connection is desired. In both cases, (host, port) specify the
2289 * peer's (host, pair) pair. Client connections are not made
2290 * automatically by this routine. The parameter socket gives the
2291 * socket descriptor on which the packet was received. This is used,
2292 * in the case of server connections, to check that *new* connections
2293 * come via a valid (port, serviceId). Finally, the securityIndex
2294 * parameter must match the existing index for the connection. If a
2295 * server connection is created, it will be created using the supplied
2296 * index, if the index is valid for this service */
2297 struct rx_connection *
2298 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2299 register u_short port, u_short serviceId, afs_uint32 cid,
2300 afs_uint32 epoch, int type, u_int securityIndex)
2302 int hashindex, flag;
2303 register struct rx_connection *conn;
2304 hashindex = CONN_HASH(host, port, cid, epoch, type);
2305 MUTEX_ENTER(&rx_connHashTable_lock);
2306 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2307 rx_connHashTable[hashindex],
2310 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2311 && (epoch == conn->epoch)) {
2312 register struct rx_peer *pp = conn->peer;
2313 if (securityIndex != conn->securityIndex) {
2314 /* this isn't supposed to happen, but someone could forge a packet
2315 * like this, and there seems to be some CM bug that makes this
2316 * happen from time to time -- in which case, the fileserver
2318 MUTEX_EXIT(&rx_connHashTable_lock);
2319 return (struct rx_connection *)0;
2321 if (pp->host == host && pp->port == port)
2323 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2325 /* So what happens when it's a callback connection? */
2326 if ( /*type == RX_CLIENT_CONNECTION && */
2327 (conn->epoch & 0x80000000))
2331 /* the connection rxLastConn that was used the last time is not the
2332 ** one we are looking for now. Hence, start searching in the hash */
2334 conn = rx_connHashTable[hashindex];
2339 struct rx_service *service;
2340 if (type == RX_CLIENT_CONNECTION) {
2341 MUTEX_EXIT(&rx_connHashTable_lock);
2342 return (struct rx_connection *)0;
2344 service = rxi_FindService(socket, serviceId);
2345 if (!service || (securityIndex >= service->nSecurityObjects)
2346 || (service->securityObjects[securityIndex] == 0)) {
2347 MUTEX_EXIT(&rx_connHashTable_lock);
2348 return (struct rx_connection *)0;
2350 conn = rxi_AllocConnection(); /* This bzero's the connection */
2351 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2352 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2353 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2354 conn->next = rx_connHashTable[hashindex];
2355 rx_connHashTable[hashindex] = conn;
2356 conn->peer = rxi_FindPeer(host, port, 0, 1);
2357 conn->type = RX_SERVER_CONNECTION;
2358 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2359 conn->epoch = epoch;
2360 conn->cid = cid & RX_CIDMASK;
2361 /* conn->serial = conn->lastSerial = 0; */
2362 /* conn->timeout = 0; */
2363 conn->ackRate = RX_FAST_ACK_RATE;
2364 conn->service = service;
2365 conn->serviceId = serviceId;
2366 conn->securityIndex = securityIndex;
2367 conn->securityObject = service->securityObjects[securityIndex];
2368 conn->nSpecific = 0;
2369 conn->specific = NULL;
2370 rx_SetConnDeadTime(conn, service->connDeadTime);
2371 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2372 /* Notify security object of the new connection */
2373 RXS_NewConnection(conn->securityObject, conn);
2374 /* XXXX Connection timeout? */
2375 if (service->newConnProc)
2376 (*service->newConnProc) (conn);
2377 MUTEX_ENTER(&rx_stats_mutex);
2378 rx_stats.nServerConns++;
2379 MUTEX_EXIT(&rx_stats_mutex);
2382 MUTEX_ENTER(&conn->conn_data_lock);
2384 MUTEX_EXIT(&conn->conn_data_lock);
2386 rxLastConn = conn; /* store this connection as the last conn used */
2387 MUTEX_EXIT(&rx_connHashTable_lock);
2391 /* There are two packet tracing routines available for testing and monitoring
2392 * Rx. One is called just after every packet is received and the other is
2393 * called just before every packet is sent. Received packets, have had their
2394 * headers decoded, and packets to be sent have not yet had their headers
2395 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2396 * containing the network address. Both can be modified. The return value, if
2397 * non-zero, indicates that the packet should be dropped. */
2399 int (*rx_justReceived) () = 0;
2400 int (*rx_almostSent) () = 0;
2402 /* A packet has been received off the interface. Np is the packet, socket is
2403 * the socket number it was received from (useful in determining which service
2404 * this packet corresponds to), and (host, port) reflect the host,port of the
2405 * sender. This call returns the packet to the caller if it is finished with
2406 * it, rather than de-allocating it, just as a small performance hack */
2409 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2410 afs_uint32 host, u_short port, int *tnop,
2411 struct rx_call **newcallp)
2413 register struct rx_call *call;
2414 register struct rx_connection *conn;
2416 afs_uint32 currentCallNumber;
2422 struct rx_packet *tnp;
2425 /* We don't print out the packet until now because (1) the time may not be
2426 * accurate enough until now in the lwp implementation (rx_Listener only gets
2427 * the time after the packet is read) and (2) from a protocol point of view,
2428 * this is the first time the packet has been seen */
2429 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2430 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2431 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2432 np->header.serial, packetType, host, port, np->header.serviceId,
2433 np->header.epoch, np->header.cid, np->header.callNumber,
2434 np->header.seq, np->header.flags, np));
2437 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2438 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2441 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2442 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2445 /* If an input tracer function is defined, call it with the packet and
2446 * network address. Note this function may modify its arguments. */
2447 if (rx_justReceived) {
2448 struct sockaddr_in addr;
2450 addr.sin_family = AF_INET;
2451 addr.sin_port = port;
2452 addr.sin_addr.s_addr = host;
2453 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2454 addr.sin_len = sizeof(addr);
2455 #endif /* AFS_OSF_ENV */
2456 drop = (*rx_justReceived) (np, &addr);
2457 /* drop packet if return value is non-zero */
2460 port = addr.sin_port; /* in case fcn changed addr */
2461 host = addr.sin_addr.s_addr;
2465 /* If packet was not sent by the client, then *we* must be the client */
2466 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2467 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2469 /* Find the connection (or fabricate one, if we're the server & if
2470 * necessary) associated with this packet */
2472 rxi_FindConnection(socket, host, port, np->header.serviceId,
2473 np->header.cid, np->header.epoch, type,
2474 np->header.securityIndex);
2477 /* If no connection found or fabricated, just ignore the packet.
2478 * (An argument could be made for sending an abort packet for
2483 MUTEX_ENTER(&conn->conn_data_lock);
2484 if (conn->maxSerial < np->header.serial)
2485 conn->maxSerial = np->header.serial;
2486 MUTEX_EXIT(&conn->conn_data_lock);
2488 /* If the connection is in an error state, send an abort packet and ignore
2489 * the incoming packet */
2491 /* Don't respond to an abort packet--we don't want loops! */
2492 MUTEX_ENTER(&conn->conn_data_lock);
2493 if (np->header.type != RX_PACKET_TYPE_ABORT)
2494 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2496 MUTEX_EXIT(&conn->conn_data_lock);
2500 /* Check for connection-only requests (i.e. not call specific). */
2501 if (np->header.callNumber == 0) {
2502 switch (np->header.type) {
2503 case RX_PACKET_TYPE_ABORT:
2504 /* What if the supplied error is zero? */
2505 rxi_ConnectionError(conn, ntohl(rx_GetInt32(np, 0)));
2506 MUTEX_ENTER(&conn->conn_data_lock);
2508 MUTEX_EXIT(&conn->conn_data_lock);
2510 case RX_PACKET_TYPE_CHALLENGE:
2511 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2512 MUTEX_ENTER(&conn->conn_data_lock);
2514 MUTEX_EXIT(&conn->conn_data_lock);
2516 case RX_PACKET_TYPE_RESPONSE:
2517 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2518 MUTEX_ENTER(&conn->conn_data_lock);
2520 MUTEX_EXIT(&conn->conn_data_lock);
2522 case RX_PACKET_TYPE_PARAMS:
2523 case RX_PACKET_TYPE_PARAMS + 1:
2524 case RX_PACKET_TYPE_PARAMS + 2:
2525 /* ignore these packet types for now */
2526 MUTEX_ENTER(&conn->conn_data_lock);
2528 MUTEX_EXIT(&conn->conn_data_lock);
2533 /* Should not reach here, unless the peer is broken: send an
2535 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2536 MUTEX_ENTER(&conn->conn_data_lock);
2537 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2539 MUTEX_EXIT(&conn->conn_data_lock);
2544 channel = np->header.cid & RX_CHANNELMASK;
2545 call = conn->call[channel];
2546 #ifdef RX_ENABLE_LOCKS
2548 MUTEX_ENTER(&call->lock);
2549 /* Test to see if call struct is still attached to conn. */
2550 if (call != conn->call[channel]) {
2552 MUTEX_EXIT(&call->lock);
2553 if (type == RX_SERVER_CONNECTION) {
2554 call = conn->call[channel];
2555 /* If we started with no call attached and there is one now,
2556 * another thread is also running this routine and has gotten
2557 * the connection channel. We should drop this packet in the tests
2558 * below. If there was a call on this connection and it's now
2559 * gone, then we'll be making a new call below.
2560 * If there was previously a call and it's now different then
2561 * the old call was freed and another thread running this routine
2562 * has created a call on this channel. One of these two threads
2563 * has a packet for the old call and the code below handles those
2567 MUTEX_ENTER(&call->lock);
2569 /* This packet can't be for this call. If the new call address is
2570 * 0 then no call is running on this channel. If there is a call
2571 * then, since this is a client connection we're getting data for
2572 * it must be for the previous call.
2574 MUTEX_ENTER(&rx_stats_mutex);
2575 rx_stats.spuriousPacketsRead++;
2576 MUTEX_EXIT(&rx_stats_mutex);
2577 MUTEX_ENTER(&conn->conn_data_lock);
2579 MUTEX_EXIT(&conn->conn_data_lock);
2584 currentCallNumber = conn->callNumber[channel];
2586 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2587 if (np->header.callNumber < currentCallNumber) {
2588 MUTEX_ENTER(&rx_stats_mutex);
2589 rx_stats.spuriousPacketsRead++;
2590 MUTEX_EXIT(&rx_stats_mutex);
2591 #ifdef RX_ENABLE_LOCKS
2593 MUTEX_EXIT(&call->lock);
2595 MUTEX_ENTER(&conn->conn_data_lock);
2597 MUTEX_EXIT(&conn->conn_data_lock);
2601 MUTEX_ENTER(&conn->conn_call_lock);
2602 call = rxi_NewCall(conn, channel);
2603 MUTEX_EXIT(&conn->conn_call_lock);
2604 *call->callNumber = np->header.callNumber;
2605 call->state = RX_STATE_PRECALL;
2606 clock_GetTime(&call->queueTime);
2607 hzero(call->bytesSent);
2608 hzero(call->bytesRcvd);
2609 rxi_KeepAliveOn(call);
2610 } else if (np->header.callNumber != currentCallNumber) {
2611 /* Wait until the transmit queue is idle before deciding
2612 * whether to reset the current call. Chances are that the
2613 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2616 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2617 while ((call->state == RX_STATE_ACTIVE)
2618 && (call->flags & RX_CALL_TQ_BUSY)) {
2619 call->flags |= RX_CALL_TQ_WAIT;
2620 #ifdef RX_ENABLE_LOCKS
2621 CV_WAIT(&call->cv_tq, &call->lock);
2622 #else /* RX_ENABLE_LOCKS */
2623 osi_rxSleep(&call->tq);
2624 #endif /* RX_ENABLE_LOCKS */
2626 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2627 /* If the new call cannot be taken right now send a busy and set
2628 * the error condition in this call, so that it terminates as
2629 * quickly as possible */
2630 if (call->state == RX_STATE_ACTIVE) {
2631 struct rx_packet *tp;
2633 rxi_CallError(call, RX_CALL_DEAD);
2634 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2636 MUTEX_EXIT(&call->lock);
2637 MUTEX_ENTER(&conn->conn_data_lock);
2639 MUTEX_EXIT(&conn->conn_data_lock);
2642 rxi_ResetCall(call, 0);
2643 *call->callNumber = np->header.callNumber;
2644 call->state = RX_STATE_PRECALL;
2645 clock_GetTime(&call->queueTime);
2646 hzero(call->bytesSent);
2647 hzero(call->bytesRcvd);
2649 * If the number of queued calls exceeds the overload
2650 * threshold then abort this call.
2652 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2653 struct rx_packet *tp;
2655 rxi_CallError(call, rx_BusyError);
2656 tp = rxi_SendCallAbort(call, np, 1, 0);
2657 MUTEX_EXIT(&call->lock);
2658 MUTEX_ENTER(&conn->conn_data_lock);
2660 MUTEX_EXIT(&conn->conn_data_lock);
2663 rxi_KeepAliveOn(call);
2665 /* Continuing call; do nothing here. */
2667 } else { /* we're the client */
2668 /* Ignore all incoming acknowledgements for calls in DALLY state */
2669 if (call && (call->state == RX_STATE_DALLY)
2670 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2671 MUTEX_ENTER(&rx_stats_mutex);
2672 rx_stats.ignorePacketDally++;
2673 MUTEX_EXIT(&rx_stats_mutex);
2674 #ifdef RX_ENABLE_LOCKS
2676 MUTEX_EXIT(&call->lock);
2679 MUTEX_ENTER(&conn->conn_data_lock);
2681 MUTEX_EXIT(&conn->conn_data_lock);
2685 /* Ignore anything that's not relevant to the current call. If there
2686 * isn't a current call, then no packet is relevant. */
2687 if (!call || (np->header.callNumber != currentCallNumber)) {
2688 MUTEX_ENTER(&rx_stats_mutex);
2689 rx_stats.spuriousPacketsRead++;
2690 MUTEX_EXIT(&rx_stats_mutex);
2691 #ifdef RX_ENABLE_LOCKS
2693 MUTEX_EXIT(&call->lock);
2696 MUTEX_ENTER(&conn->conn_data_lock);
2698 MUTEX_EXIT(&conn->conn_data_lock);
2701 /* If the service security object index stamped in the packet does not
2702 * match the connection's security index, ignore the packet */
2703 if (np->header.securityIndex != conn->securityIndex) {
2704 #ifdef RX_ENABLE_LOCKS
2705 MUTEX_EXIT(&call->lock);
2707 MUTEX_ENTER(&conn->conn_data_lock);
2709 MUTEX_EXIT(&conn->conn_data_lock);
2713 /* If we're receiving the response, then all transmit packets are
2714 * implicitly acknowledged. Get rid of them. */
2715 if (np->header.type == RX_PACKET_TYPE_DATA) {
2716 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2717 /* XXX Hack. Because we must release the global rx lock when
2718 * sending packets (osi_NetSend) we drop all acks while we're
2719 * traversing the tq in rxi_Start sending packets out because
2720 * packets may move to the freePacketQueue as result of being here!
2721 * So we drop these packets until we're safely out of the
2722 * traversing. Really ugly!
2723 * For fine grain RX locking, we set the acked field in the
2724 * packets and let rxi_Start remove them from the transmit queue.
2726 if (call->flags & RX_CALL_TQ_BUSY) {
2727 #ifdef RX_ENABLE_LOCKS
2728 rxi_SetAcksInTransmitQueue(call);
2731 return np; /* xmitting; drop packet */
2734 rxi_ClearTransmitQueue(call, 0);
2736 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2737 rxi_ClearTransmitQueue(call, 0);
2738 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2740 if (np->header.type == RX_PACKET_TYPE_ACK) {
2741 /* now check to see if this is an ack packet acknowledging that the
2742 * server actually *lost* some hard-acked data. If this happens we
2743 * ignore this packet, as it may indicate that the server restarted in
2744 * the middle of a call. It is also possible that this is an old ack
2745 * packet. We don't abort the connection in this case, because this
2746 * *might* just be an old ack packet. The right way to detect a server
2747 * restart in the midst of a call is to notice that the server epoch
2749 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2750 * XXX unacknowledged. I think that this is off-by-one, but
2751 * XXX I don't dare change it just yet, since it will
2752 * XXX interact badly with the server-restart detection
2753 * XXX code in receiveackpacket. */
2754 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2755 MUTEX_ENTER(&rx_stats_mutex);
2756 rx_stats.spuriousPacketsRead++;
2757 MUTEX_EXIT(&rx_stats_mutex);
2758 MUTEX_EXIT(&call->lock);
2759 MUTEX_ENTER(&conn->conn_data_lock);
2761 MUTEX_EXIT(&conn->conn_data_lock);
2765 } /* else not a data packet */
2768 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2769 /* Set remote user defined status from packet */
2770 call->remoteStatus = np->header.userStatus;
2772 /* Note the gap between the expected next packet and the actual
2773 * packet that arrived, when the new packet has a smaller serial number
2774 * than expected. Rioses frequently reorder packets all by themselves,
2775 * so this will be quite important with very large window sizes.
2776 * Skew is checked against 0 here to avoid any dependence on the type of
2777 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2779 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2780 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2781 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2783 MUTEX_ENTER(&conn->conn_data_lock);
2784 skew = conn->lastSerial - np->header.serial;
2785 conn->lastSerial = np->header.serial;
2786 MUTEX_EXIT(&conn->conn_data_lock);
2788 register struct rx_peer *peer;
2790 if (skew > peer->inPacketSkew) {
2791 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2793 peer->inPacketSkew = skew;
2797 /* Now do packet type-specific processing */
2798 switch (np->header.type) {
2799 case RX_PACKET_TYPE_DATA:
2800 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2803 case RX_PACKET_TYPE_ACK:
2804 /* Respond immediately to ack packets requesting acknowledgement
2806 if (np->header.flags & RX_REQUEST_ACK) {
2808 (void)rxi_SendCallAbort(call, 0, 1, 0);
2810 (void)rxi_SendAck(call, 0, np->header.serial,
2811 RX_ACK_PING_RESPONSE, 1);
2813 np = rxi_ReceiveAckPacket(call, np, 1);
2815 case RX_PACKET_TYPE_ABORT:
2816 /* An abort packet: reset the connection, passing the error up to
2818 /* What if error is zero? */
2819 rxi_CallError(call, ntohl(*(afs_int32 *) rx_DataOf(np)));
2821 case RX_PACKET_TYPE_BUSY:
2824 case RX_PACKET_TYPE_ACKALL:
2825 /* All packets acknowledged, so we can drop all packets previously
2826 * readied for sending */
2827 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2828 /* XXX Hack. We because we can't release the global rx lock when
2829 * sending packets (osi_NetSend) we drop all ack pkts while we're
2830 * traversing the tq in rxi_Start sending packets out because
2831 * packets may move to the freePacketQueue as result of being
2832 * here! So we drop these packets until we're safely out of the
2833 * traversing. Really ugly!
2834 * For fine grain RX locking, we set the acked field in the packets
2835 * and let rxi_Start remove the packets from the transmit queue.
2837 if (call->flags & RX_CALL_TQ_BUSY) {
2838 #ifdef RX_ENABLE_LOCKS
2839 rxi_SetAcksInTransmitQueue(call);
2841 #else /* RX_ENABLE_LOCKS */
2843 return np; /* xmitting; drop packet */
2844 #endif /* RX_ENABLE_LOCKS */
2846 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2847 rxi_ClearTransmitQueue(call, 0);
2850 /* Should not reach here, unless the peer is broken: send an abort
2852 rxi_CallError(call, RX_PROTOCOL_ERROR);
2853 np = rxi_SendCallAbort(call, np, 1, 0);
2856 /* Note when this last legitimate packet was received, for keep-alive
2857 * processing. Note, we delay getting the time until now in the hope that
2858 * the packet will be delivered to the user before any get time is required
2859 * (if not, then the time won't actually be re-evaluated here). */
2860 call->lastReceiveTime = clock_Sec();
2861 MUTEX_EXIT(&call->lock);
2862 MUTEX_ENTER(&conn->conn_data_lock);
2864 MUTEX_EXIT(&conn->conn_data_lock);
2868 /* return true if this is an "interesting" connection from the point of view
2869 of someone trying to debug the system */
2871 rxi_IsConnInteresting(struct rx_connection *aconn)
2874 register struct rx_call *tcall;
2876 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2878 for (i = 0; i < RX_MAXCALLS; i++) {
2879 tcall = aconn->call[i];
2881 if ((tcall->state == RX_STATE_PRECALL)
2882 || (tcall->state == RX_STATE_ACTIVE))
2884 if ((tcall->mode == RX_MODE_SENDING)
2885 || (tcall->mode == RX_MODE_RECEIVING))
2893 /* if this is one of the last few packets AND it wouldn't be used by the
2894 receiving call to immediately satisfy a read request, then drop it on
2895 the floor, since accepting it might prevent a lock-holding thread from
2896 making progress in its reading. If a call has been cleared while in
2897 the precall state then ignore all subsequent packets until the call
2898 is assigned to a thread. */
2901 TooLow(struct rx_packet *ap, struct rx_call *acall)
2904 MUTEX_ENTER(&rx_stats_mutex);
2905 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2906 && (acall->state == RX_STATE_PRECALL))
2907 || ((rx_nFreePackets < rxi_dataQuota + 2)
2908 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2909 && (acall->flags & RX_CALL_READER_WAIT)))) {
2912 MUTEX_EXIT(&rx_stats_mutex);
2918 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2919 struct rx_call *acall)
2921 struct rx_call *call = acall;
2925 MUTEX_ENTER(&conn->conn_data_lock);
2926 conn->checkReachEvent = NULL;
2927 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2930 MUTEX_EXIT(&conn->conn_data_lock);
2934 MUTEX_ENTER(&conn->conn_call_lock);
2935 MUTEX_ENTER(&conn->conn_data_lock);
2936 for (i = 0; i < RX_MAXCALLS; i++) {
2937 struct rx_call *tc = conn->call[i];
2938 if (tc && tc->state == RX_STATE_PRECALL) {
2944 /* Indicate that rxi_CheckReachEvent is no longer running by
2945 * clearing the flag. Must be atomic under conn_data_lock to
2946 * avoid a new call slipping by: rxi_CheckConnReach holds
2947 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
2949 conn->flags &= ~RX_CONN_ATTACHWAIT;
2950 MUTEX_EXIT(&conn->conn_data_lock);
2951 MUTEX_EXIT(&conn->conn_call_lock);
2956 MUTEX_ENTER(&call->lock);
2957 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
2959 MUTEX_EXIT(&call->lock);
2961 clock_GetTime(&when);
2962 when.sec += RX_CHECKREACH_TIMEOUT;
2963 MUTEX_ENTER(&conn->conn_data_lock);
2964 if (!conn->checkReachEvent) {
2966 conn->checkReachEvent =
2967 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
2969 MUTEX_EXIT(&conn->conn_data_lock);
2975 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
2977 struct rx_service *service = conn->service;
2978 struct rx_peer *peer = conn->peer;
2979 afs_uint32 now, lastReach;
2981 if (service->checkReach == 0)
2985 MUTEX_ENTER(&peer->peer_lock);
2986 lastReach = peer->lastReachTime;
2987 MUTEX_EXIT(&peer->peer_lock);
2988 if (now - lastReach < RX_CHECKREACH_TTL)
2991 MUTEX_ENTER(&conn->conn_data_lock);
2992 if (conn->flags & RX_CONN_ATTACHWAIT) {
2993 MUTEX_EXIT(&conn->conn_data_lock);
2996 conn->flags |= RX_CONN_ATTACHWAIT;
2997 MUTEX_EXIT(&conn->conn_data_lock);
2998 if (!conn->checkReachEvent)
2999 rxi_CheckReachEvent(NULL, conn, call);
3004 /* try to attach call, if authentication is complete */
3006 TryAttach(register struct rx_call *acall, register osi_socket socket,
3007 register int *tnop, register struct rx_call **newcallp,
3010 struct rx_connection *conn = acall->conn;
3012 if (conn->type == RX_SERVER_CONNECTION
3013 && acall->state == RX_STATE_PRECALL) {
3014 /* Don't attach until we have any req'd. authentication. */
3015 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3016 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3017 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3018 /* Note: this does not necessarily succeed; there
3019 * may not any proc available
3022 rxi_ChallengeOn(acall->conn);
3027 /* A data packet has been received off the interface. This packet is
3028 * appropriate to the call (the call is in the right state, etc.). This
3029 * routine can return a packet to the caller, for re-use */
3032 rxi_ReceiveDataPacket(register struct rx_call *call,
3033 register struct rx_packet *np, int istack,
3034 osi_socket socket, afs_uint32 host, u_short port,
3035 int *tnop, struct rx_call **newcallp)
3037 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3041 afs_uint32 seq, serial, flags;
3043 struct rx_packet *tnp;
3045 MUTEX_ENTER(&rx_stats_mutex);
3046 rx_stats.dataPacketsRead++;
3047 MUTEX_EXIT(&rx_stats_mutex);
3050 /* If there are no packet buffers, drop this new packet, unless we can find
3051 * packet buffers from inactive calls */
3053 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3054 MUTEX_ENTER(&rx_freePktQ_lock);
3055 rxi_NeedMorePackets = TRUE;
3056 MUTEX_EXIT(&rx_freePktQ_lock);
3057 MUTEX_ENTER(&rx_stats_mutex);
3058 rx_stats.noPacketBuffersOnRead++;
3059 MUTEX_EXIT(&rx_stats_mutex);
3060 call->rprev = np->header.serial;
3061 rxi_calltrace(RX_TRACE_DROP, call);
3062 dpf(("packet %x dropped on receipt - quota problems", np));
3064 rxi_ClearReceiveQueue(call);
3065 clock_GetTime(&when);
3066 clock_Add(&when, &rx_softAckDelay);
3067 if (!call->delayedAckEvent
3068 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3069 rxevent_Cancel(call->delayedAckEvent, call,
3070 RX_CALL_REFCOUNT_DELAY);
3071 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3072 call->delayedAckEvent =
3073 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3075 /* we've damaged this call already, might as well do it in. */
3081 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3082 * packet is one of several packets transmitted as a single
3083 * datagram. Do not send any soft or hard acks until all packets
3084 * in a jumbogram have been processed. Send negative acks right away.
3086 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3087 /* tnp is non-null when there are more packets in the
3088 * current jumbo gram */
3095 seq = np->header.seq;
3096 serial = np->header.serial;
3097 flags = np->header.flags;
3099 /* If the call is in an error state, send an abort message */
3101 return rxi_SendCallAbort(call, np, istack, 0);
3103 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3104 * AFS 3.5 jumbogram. */
3105 if (flags & RX_JUMBO_PACKET) {
3106 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3111 if (np->header.spare != 0) {
3112 MUTEX_ENTER(&call->conn->conn_data_lock);
3113 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3114 MUTEX_EXIT(&call->conn->conn_data_lock);
3117 /* The usual case is that this is the expected next packet */
3118 if (seq == call->rnext) {
3120 /* Check to make sure it is not a duplicate of one already queued */
3121 if (queue_IsNotEmpty(&call->rq)
3122 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3123 MUTEX_ENTER(&rx_stats_mutex);
3124 rx_stats.dupPacketsRead++;
3125 MUTEX_EXIT(&rx_stats_mutex);
3126 dpf(("packet %x dropped on receipt - duplicate", np));
3127 rxevent_Cancel(call->delayedAckEvent, call,
3128 RX_CALL_REFCOUNT_DELAY);
3129 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3135 /* It's the next packet. Stick it on the receive queue
3136 * for this call. Set newPackets to make sure we wake
3137 * the reader once all packets have been processed */
3138 queue_Prepend(&call->rq, np);
3140 np = NULL; /* We can't use this anymore */
3143 /* If an ack is requested then set a flag to make sure we
3144 * send an acknowledgement for this packet */
3145 if (flags & RX_REQUEST_ACK) {
3146 ackNeeded = RX_ACK_REQUESTED;
3149 /* Keep track of whether we have received the last packet */
3150 if (flags & RX_LAST_PACKET) {
3151 call->flags |= RX_CALL_HAVE_LAST;
3155 /* Check whether we have all of the packets for this call */
3156 if (call->flags & RX_CALL_HAVE_LAST) {
3157 afs_uint32 tseq; /* temporary sequence number */
3158 struct rx_packet *tp; /* Temporary packet pointer */
3159 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3161 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3162 if (tseq != tp->header.seq)
3164 if (tp->header.flags & RX_LAST_PACKET) {
3165 call->flags |= RX_CALL_RECEIVE_DONE;
3172 /* Provide asynchronous notification for those who want it
3173 * (e.g. multi rx) */
3174 if (call->arrivalProc) {
3175 (*call->arrivalProc) (call, call->arrivalProcHandle,
3176 (int)call->arrivalProcArg);
3177 call->arrivalProc = (VOID(*)())0;
3180 /* Update last packet received */
3183 /* If there is no server process serving this call, grab
3184 * one, if available. We only need to do this once. If a
3185 * server thread is available, this thread becomes a server
3186 * thread and the server thread becomes a listener thread. */
3188 TryAttach(call, socket, tnop, newcallp, 0);
3191 /* This is not the expected next packet. */
3193 /* Determine whether this is a new or old packet, and if it's
3194 * a new one, whether it fits into the current receive window.
3195 * Also figure out whether the packet was delivered in sequence.
3196 * We use the prev variable to determine whether the new packet
3197 * is the successor of its immediate predecessor in the
3198 * receive queue, and the missing flag to determine whether
3199 * any of this packets predecessors are missing. */
3201 afs_uint32 prev; /* "Previous packet" sequence number */
3202 struct rx_packet *tp; /* Temporary packet pointer */
3203 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3204 int missing; /* Are any predecessors missing? */
3206 /* If the new packet's sequence number has been sent to the
3207 * application already, then this is a duplicate */
3208 if (seq < call->rnext) {
3209 MUTEX_ENTER(&rx_stats_mutex);
3210 rx_stats.dupPacketsRead++;
3211 MUTEX_EXIT(&rx_stats_mutex);
3212 rxevent_Cancel(call->delayedAckEvent, call,
3213 RX_CALL_REFCOUNT_DELAY);
3214 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3220 /* If the sequence number is greater than what can be
3221 * accomodated by the current window, then send a negative
3222 * acknowledge and drop the packet */
3223 if ((call->rnext + call->rwind) <= seq) {
3224 rxevent_Cancel(call->delayedAckEvent, call,
3225 RX_CALL_REFCOUNT_DELAY);
3226 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3233 /* Look for the packet in the queue of old received packets */
3234 for (prev = call->rnext - 1, missing =
3235 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3236 /*Check for duplicate packet */
3237 if (seq == tp->header.seq) {
3238 MUTEX_ENTER(&rx_stats_mutex);
3239 rx_stats.dupPacketsRead++;
3240 MUTEX_EXIT(&rx_stats_mutex);
3241 rxevent_Cancel(call->delayedAckEvent, call,
3242 RX_CALL_REFCOUNT_DELAY);
3243 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3249 /* If we find a higher sequence packet, break out and
3250 * insert the new packet here. */
3251 if (seq < tp->header.seq)
3253 /* Check for missing packet */
3254 if (tp->header.seq != prev + 1) {
3258 prev = tp->header.seq;
3261 /* Keep track of whether we have received the last packet. */
3262 if (flags & RX_LAST_PACKET) {
3263 call->flags |= RX_CALL_HAVE_LAST;
3266 /* It's within the window: add it to the the receive queue.
3267 * tp is left by the previous loop either pointing at the
3268 * packet before which to insert the new packet, or at the
3269 * queue head if the queue is empty or the packet should be
3271 queue_InsertBefore(tp, np);
3275 /* Check whether we have all of the packets for this call */
3276 if ((call->flags & RX_CALL_HAVE_LAST)
3277 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3278 afs_uint32 tseq; /* temporary sequence number */
3281 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3282 if (tseq != tp->header.seq)
3284 if (tp->header.flags & RX_LAST_PACKET) {
3285 call->flags |= RX_CALL_RECEIVE_DONE;
3292 /* We need to send an ack of the packet is out of sequence,
3293 * or if an ack was requested by the peer. */
3294 if (seq != prev + 1 || missing || (flags & RX_REQUEST_ACK)) {
3295 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3298 /* Acknowledge the last packet for each call */
3299 if (flags & RX_LAST_PACKET) {
3310 * If the receiver is waiting for an iovec, fill the iovec
3311 * using the data from the receive queue */
3312 if (call->flags & RX_CALL_IOVEC_WAIT) {
3313 didHardAck = rxi_FillReadVec(call, serial);
3314 /* the call may have been aborted */
3323 /* Wakeup the reader if any */
3324 if ((call->flags & RX_CALL_READER_WAIT)
3325 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3326 || (call->iovNext >= call->iovMax)
3327 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3328 call->flags &= ~RX_CALL_READER_WAIT;
3329 #ifdef RX_ENABLE_LOCKS
3330 CV_BROADCAST(&call->cv_rq);
3332 osi_rxWakeup(&call->rq);
3338 * Send an ack when requested by the peer, or once every
3339 * rxi_SoftAckRate packets until the last packet has been
3340 * received. Always send a soft ack for the last packet in
3341 * the server's reply. */
3343 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3344 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3345 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3346 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3347 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3348 } else if (call->nSoftAcks) {
3349 clock_GetTime(&when);
3350 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3351 clock_Add(&when, &rx_lastAckDelay);
3353 clock_Add(&when, &rx_softAckDelay);
3355 if (!call->delayedAckEvent
3356 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3357 rxevent_Cancel(call->delayedAckEvent, call,
3358 RX_CALL_REFCOUNT_DELAY);
3359 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3360 call->delayedAckEvent =
3361 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3363 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3364 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3371 static void rxi_ComputeRate();
3375 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3377 struct rx_peer *peer = conn->peer;
3379 MUTEX_ENTER(&peer->peer_lock);
3380 peer->lastReachTime = clock_Sec();
3381 MUTEX_EXIT(&peer->peer_lock);
3383 MUTEX_ENTER(&conn->conn_data_lock);
3384 if (conn->flags & RX_CONN_ATTACHWAIT) {
3387 conn->flags &= ~RX_CONN_ATTACHWAIT;
3388 MUTEX_EXIT(&conn->conn_data_lock);
3390 for (i = 0; i < RX_MAXCALLS; i++) {
3391 struct rx_call *call = conn->call[i];
3394 MUTEX_ENTER(&call->lock);
3395 /* tnop can be null if newcallp is null */
3396 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3398 MUTEX_EXIT(&call->lock);
3402 MUTEX_EXIT(&conn->conn_data_lock);
3405 /* rxi_ComputePeerNetStats
3407 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3408 * estimates (like RTT and throughput) based on ack packets. Caller
3409 * must ensure that the packet in question is the right one (i.e.
3410 * serial number matches).
3413 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3414 struct rx_ackPacket *ap, struct rx_packet *np)
3416 struct rx_peer *peer = call->conn->peer;
3418 /* Use RTT if not delayed by client. */
3419 if (ap->reason != RX_ACK_DELAY)
3420 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3422 rxi_ComputeRate(peer, call, p, np, ap->reason);
3426 /* The real smarts of the whole thing. */
3428 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3431 struct rx_ackPacket *ap;
3433 register struct rx_packet *tp;
3434 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3435 register struct rx_connection *conn = call->conn;
3436 struct rx_peer *peer = conn->peer;
3439 /* because there are CM's that are bogus, sending weird values for this. */
3440 afs_uint32 skew = 0;
3445 int newAckCount = 0;
3446 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3447 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3449 MUTEX_ENTER(&rx_stats_mutex);
3450 rx_stats.ackPacketsRead++;
3451 MUTEX_EXIT(&rx_stats_mutex);
3452 ap = (struct rx_ackPacket *)rx_DataOf(np);
3453 nbytes = rx_Contiguous(np) - ((ap->acks) - (u_char *) ap);
3455 return np; /* truncated ack packet */
3457 /* depends on ack packet struct */
3458 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3459 first = ntohl(ap->firstPacket);
3460 serial = ntohl(ap->serial);
3461 /* temporarily disabled -- needs to degrade over time
3462 * skew = ntohs(ap->maxSkew); */
3464 /* Ignore ack packets received out of order */
3465 if (first < call->tfirst) {
3469 if (np->header.flags & RX_SLOW_START_OK) {
3470 call->flags |= RX_CALL_SLOW_START_OK;
3473 if (ap->reason == RX_ACK_PING_RESPONSE)
3474 rxi_UpdatePeerReach(conn, call);
3479 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3480 ap->reason, ntohl(ap->previousPacket),
3481 (unsigned int)np->header.seq, (unsigned int)serial,
3482 (unsigned int)skew, ntohl(ap->firstPacket));
3485 for (offset = 0; offset < nAcks; offset++)
3486 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3493 /* Update the outgoing packet skew value to the latest value of
3494 * the peer's incoming packet skew value. The ack packet, of
3495 * course, could arrive out of order, but that won't affect things
3497 MUTEX_ENTER(&peer->peer_lock);
3498 peer->outPacketSkew = skew;
3500 /* Check for packets that no longer need to be transmitted, and
3501 * discard them. This only applies to packets positively
3502 * acknowledged as having been sent to the peer's upper level.
3503 * All other packets must be retained. So only packets with
3504 * sequence numbers < ap->firstPacket are candidates. */
3505 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3506 if (tp->header.seq >= first)
3508 call->tfirst = tp->header.seq + 1;
3510 && (tp->header.serial == serial || tp->firstSerial == serial))
3511 rxi_ComputePeerNetStats(call, tp, ap, np);
3512 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3513 /* XXX Hack. Because we have to release the global rx lock when sending
3514 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3515 * in rxi_Start sending packets out because packets may move to the
3516 * freePacketQueue as result of being here! So we drop these packets until
3517 * we're safely out of the traversing. Really ugly!
3518 * To make it even uglier, if we're using fine grain locking, we can
3519 * set the ack bits in the packets and have rxi_Start remove the packets
3520 * when it's done transmitting.
3522 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3525 if (call->flags & RX_CALL_TQ_BUSY) {
3526 #ifdef RX_ENABLE_LOCKS
3527 tp->flags |= RX_PKTFLAG_ACKED;
3528 call->flags |= RX_CALL_TQ_SOME_ACKED;
3529 #else /* RX_ENABLE_LOCKS */
3531 #endif /* RX_ENABLE_LOCKS */
3533 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3536 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3541 /* Give rate detector a chance to respond to ping requests */
3542 if (ap->reason == RX_ACK_PING_RESPONSE) {
3543 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3547 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3549 /* Now go through explicit acks/nacks and record the results in
3550 * the waiting packets. These are packets that can't be released
3551 * yet, even with a positive acknowledge. This positive
3552 * acknowledge only means the packet has been received by the
3553 * peer, not that it will be retained long enough to be sent to
3554 * the peer's upper level. In addition, reset the transmit timers
3555 * of any missing packets (those packets that must be missing
3556 * because this packet was out of sequence) */
3558 call->nSoftAcked = 0;
3559 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3560 /* Update round trip time if the ack was stimulated on receipt
3562 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3563 #ifdef RX_ENABLE_LOCKS
3564 if (tp->header.seq >= first)
3565 #endif /* RX_ENABLE_LOCKS */
3566 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3568 && (tp->header.serial == serial || tp->firstSerial == serial))
3569 rxi_ComputePeerNetStats(call, tp, ap, np);
3571 /* Set the acknowledge flag per packet based on the
3572 * information in the ack packet. An acknowlegded packet can
3573 * be downgraded when the server has discarded a packet it
3574 * soacked previously, or when an ack packet is received
3575 * out of sequence. */
3576 if (tp->header.seq < first) {
3577 /* Implicit ack information */
3578 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3581 tp->flags |= RX_PKTFLAG_ACKED;
3582 } else if (tp->header.seq < first + nAcks) {
3583 /* Explicit ack information: set it in the packet appropriately */
3584 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3585 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3587 tp->flags |= RX_PKTFLAG_ACKED;
3595 tp->flags &= ~RX_PKTFLAG_ACKED;
3599 tp->flags &= ~RX_PKTFLAG_ACKED;
3603 /* If packet isn't yet acked, and it has been transmitted at least
3604 * once, reset retransmit time using latest timeout
3605 * ie, this should readjust the retransmit timer for all outstanding
3606 * packets... So we don't just retransmit when we should know better*/
3608 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3609 tp->retryTime = tp->timeSent;
3610 clock_Add(&tp->retryTime, &peer->timeout);
3611 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3612 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3616 /* If the window has been extended by this acknowledge packet,
3617 * then wakeup a sender waiting in alloc for window space, or try
3618 * sending packets now, if he's been sitting on packets due to
3619 * lack of window space */
3620 if (call->tnext < (call->tfirst + call->twind)) {
3621 #ifdef RX_ENABLE_LOCKS
3622 CV_SIGNAL(&call->cv_twind);
3624 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3625 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3626 osi_rxWakeup(&call->twind);
3629 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3630 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3634 /* if the ack packet has a receivelen field hanging off it,
3635 * update our state */
3636 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3639 /* If the ack packet has a "recommended" size that is less than
3640 * what I am using now, reduce my size to match */
3641 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3642 sizeof(afs_int32), &tSize);
3643 tSize = (afs_uint32) ntohl(tSize);
3644 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3646 /* Get the maximum packet size to send to this peer */
3647 rx_packetread(np, rx_AckDataSize(ap->nAcks), sizeof(afs_int32),
3649 tSize = (afs_uint32) ntohl(tSize);
3650 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3651 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3653 /* sanity check - peer might have restarted with different params.
3654 * If peer says "send less", dammit, send less... Peer should never
3655 * be unable to accept packets of the size that prior AFS versions would
3656 * send without asking. */
3657 if (peer->maxMTU != tSize) {
3658 peer->maxMTU = tSize;
3659 peer->MTU = MIN(tSize, peer->MTU);
3660 call->MTU = MIN(call->MTU, tSize);
3664 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3667 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3668 sizeof(afs_int32), &tSize);
3669 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3670 if (tSize < call->twind) { /* smaller than our send */
3671 call->twind = tSize; /* window, we must send less... */
3672 call->ssthresh = MIN(call->twind, call->ssthresh);
3675 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3676 * network MTU confused with the loopback MTU. Calculate the
3677 * maximum MTU here for use in the slow start code below.
3679 maxMTU = peer->maxMTU;
3680 /* Did peer restart with older RX version? */
3681 if (peer->maxDgramPackets > 1) {
3682 peer->maxDgramPackets = 1;
3684 } else if (np->length >=
3685 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3688 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3689 sizeof(afs_int32), &tSize);
3690 tSize = (afs_uint32) ntohl(tSize);
3692 * As of AFS 3.5 we set the send window to match the receive window.
3694 if (tSize < call->twind) {
3695 call->twind = tSize;
3696 call->ssthresh = MIN(call->twind, call->ssthresh);
3697 } else if (tSize > call->twind) {
3698 call->twind = tSize;
3702 * As of AFS 3.5, a jumbogram is more than one fixed size
3703 * packet transmitted in a single UDP datagram. If the remote
3704 * MTU is smaller than our local MTU then never send a datagram
3705 * larger than the natural MTU.
3708 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3709 sizeof(afs_int32), &tSize);
3710 maxDgramPackets = (afs_uint32) ntohl(tSize);
3711 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3713 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3714 maxDgramPackets = MIN(maxDgramPackets, tSize);
3715 if (maxDgramPackets > 1) {
3716 peer->maxDgramPackets = maxDgramPackets;
3717 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3719 peer->maxDgramPackets = 1;
3720 call->MTU = peer->natMTU;
3722 } else if (peer->maxDgramPackets > 1) {
3723 /* Restarted with lower version of RX */
3724 peer->maxDgramPackets = 1;
3726 } else if (peer->maxDgramPackets > 1
3727 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3728 /* Restarted with lower version of RX */
3729 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3730 peer->natMTU = OLD_MAX_PACKET_SIZE;
3731 peer->MTU = OLD_MAX_PACKET_SIZE;
3732 peer->maxDgramPackets = 1;
3733 peer->nDgramPackets = 1;
3735 call->MTU = OLD_MAX_PACKET_SIZE;
3740 * Calculate how many datagrams were successfully received after
3741 * the first missing packet and adjust the negative ack counter
3746 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3747 if (call->nNacks < nNacked) {
3748 call->nNacks = nNacked;
3757 if (call->flags & RX_CALL_FAST_RECOVER) {
3759 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3761 call->flags &= ~RX_CALL_FAST_RECOVER;
3762 call->cwind = call->nextCwind;
3763 call->nextCwind = 0;
3766 call->nCwindAcks = 0;
3767 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3768 /* Three negative acks in a row trigger congestion recovery */
3769 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3770 MUTEX_EXIT(&peer->peer_lock);
3771 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3772 /* someone else is waiting to start recovery */
3775 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3776 while (call->flags & RX_CALL_TQ_BUSY) {
3777 call->flags |= RX_CALL_TQ_WAIT;
3778 #ifdef RX_ENABLE_LOCKS
3779 CV_WAIT(&call->cv_tq, &call->lock);
3780 #else /* RX_ENABLE_LOCKS */
3781 osi_rxSleep(&call->tq);
3782 #endif /* RX_ENABLE_LOCKS */
3784 MUTEX_ENTER(&peer->peer_lock);
3785 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3786 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3787 call->flags |= RX_CALL_FAST_RECOVER;
3788 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3790 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3791 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3792 call->nextCwind = call->ssthresh;
3795 peer->MTU = call->MTU;
3796 peer->cwind = call->nextCwind;
3797 peer->nDgramPackets = call->nDgramPackets;
3799 call->congestSeq = peer->congestSeq;
3800 /* Reset the resend times on the packets that were nacked
3801 * so we will retransmit as soon as the window permits*/
3802 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3804 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3805 clock_Zero(&tp->retryTime);
3807 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3812 /* If cwind is smaller than ssthresh, then increase
3813 * the window one packet for each ack we receive (exponential
3815 * If cwind is greater than or equal to ssthresh then increase
3816 * the congestion window by one packet for each cwind acks we
3817 * receive (linear growth). */
3818 if (call->cwind < call->ssthresh) {
3820 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3821 call->nCwindAcks = 0;
3823 call->nCwindAcks += newAckCount;
3824 if (call->nCwindAcks >= call->cwind) {
3825 call->nCwindAcks = 0;
3826 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3830 * If we have received several acknowledgements in a row then
3831 * it is time to increase the size of our datagrams
3833 if ((int)call->nAcks > rx_nDgramThreshold) {
3834 if (peer->maxDgramPackets > 1) {
3835 if (call->nDgramPackets < peer->maxDgramPackets) {
3836 call->nDgramPackets++;
3838 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3839 } else if (call->MTU < peer->maxMTU) {
3840 call->MTU += peer->natMTU;
3841 call->MTU = MIN(call->MTU, peer->maxMTU);
3847 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3849 /* Servers need to hold the call until all response packets have
3850 * been acknowledged. Soft acks are good enough since clients
3851 * are not allowed to clear their receive queues. */
3852 if (call->state == RX_STATE_HOLD
3853 && call->tfirst + call->nSoftAcked >= call->tnext) {
3854 call->state = RX_STATE_DALLY;
3855 rxi_ClearTransmitQueue(call, 0);
3856 } else if (!queue_IsEmpty(&call->tq)) {
3857 rxi_Start(0, call, istack);
3862 /* Received a response to a challenge packet */
3864 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3865 register struct rx_packet *np, int istack)
3869 /* Ignore the packet if we're the client */
3870 if (conn->type == RX_CLIENT_CONNECTION)
3873 /* If already authenticated, ignore the packet (it's probably a retry) */
3874 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3877 /* Otherwise, have the security object evaluate the response packet */
3878 error = RXS_CheckResponse(conn->securityObject, conn, np);
3880 /* If the response is invalid, reset the connection, sending
3881 * an abort to the peer */
3885 rxi_ConnectionError(conn, error);
3886 MUTEX_ENTER(&conn->conn_data_lock);
3887 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3888 MUTEX_EXIT(&conn->conn_data_lock);
3891 /* If the response is valid, any calls waiting to attach
3892 * servers can now do so */
3895 for (i = 0; i < RX_MAXCALLS; i++) {
3896 struct rx_call *call = conn->call[i];
3898 MUTEX_ENTER(&call->lock);
3899 if (call->state == RX_STATE_PRECALL)
3900 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
3901 /* tnop can be null if newcallp is null */
3902 MUTEX_EXIT(&call->lock);
3906 /* Update the peer reachability information, just in case
3907 * some calls went into attach-wait while we were waiting
3908 * for authentication..
3910 rxi_UpdatePeerReach(conn, NULL);
3915 /* A client has received an authentication challenge: the security
3916 * object is asked to cough up a respectable response packet to send
3917 * back to the server. The server is responsible for retrying the
3918 * challenge if it fails to get a response. */
3921 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
3922 register struct rx_packet *np, int istack)
3926 /* Ignore the challenge if we're the server */
3927 if (conn->type == RX_SERVER_CONNECTION)
3930 /* Ignore the challenge if the connection is otherwise idle; someone's
3931 * trying to use us as an oracle. */
3932 if (!rxi_HasActiveCalls(conn))
3935 /* Send the security object the challenge packet. It is expected to fill
3936 * in the response. */
3937 error = RXS_GetResponse(conn->securityObject, conn, np);
3939 /* If the security object is unable to return a valid response, reset the
3940 * connection and send an abort to the peer. Otherwise send the response
3941 * packet to the peer connection. */
3943 rxi_ConnectionError(conn, error);
3944 MUTEX_ENTER(&conn->conn_data_lock);
3945 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3946 MUTEX_EXIT(&conn->conn_data_lock);
3948 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
3949 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
3955 /* Find an available server process to service the current request in
3956 * the given call structure. If one isn't available, queue up this
3957 * call so it eventually gets one */
3959 rxi_AttachServerProc(register struct rx_call *call,
3960 register osi_socket socket, register int *tnop,
3961 register struct rx_call **newcallp)
3963 register struct rx_serverQueueEntry *sq;
3964 register struct rx_service *service = call->conn->service;
3965 register int haveQuota = 0;
3967 /* May already be attached */
3968 if (call->state == RX_STATE_ACTIVE)
3971 MUTEX_ENTER(&rx_serverPool_lock);
3973 haveQuota = QuotaOK(service);
3974 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
3975 /* If there are no processes available to service this call,
3976 * put the call on the incoming call queue (unless it's
3977 * already on the queue).
3979 #ifdef RX_ENABLE_LOCKS
3981 ReturnToServerPool(service);
3982 #endif /* RX_ENABLE_LOCKS */
3984 if (!(call->flags & RX_CALL_WAIT_PROC)) {
3985 call->flags |= RX_CALL_WAIT_PROC;
3986 MUTEX_ENTER(&rx_stats_mutex);
3989 MUTEX_EXIT(&rx_stats_mutex);
3990 rxi_calltrace(RX_CALL_ARRIVAL, call);
3991 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
3992 queue_Append(&rx_incomingCallQueue, call);
3995 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
3997 /* If hot threads are enabled, and both newcallp and sq->socketp
3998 * are non-null, then this thread will process the call, and the
3999 * idle server thread will start listening on this threads socket.
4002 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4005 *sq->socketp = socket;
4006 clock_GetTime(&call->startTime);
4007 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4011 if (call->flags & RX_CALL_WAIT_PROC) {
4012 /* Conservative: I don't think this should happen */
4013 call->flags &= ~RX_CALL_WAIT_PROC;
4014 if (queue_IsOnQueue(call)) {
4016 MUTEX_ENTER(&rx_stats_mutex);
4018 MUTEX_EXIT(&rx_stats_mutex);
4021 call->state = RX_STATE_ACTIVE;
4022 call->mode = RX_MODE_RECEIVING;
4023 #ifdef RX_KERNEL_TRACE
4025 int glockOwner = ISAFS_GLOCK();
4028 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4029 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4035 if (call->flags & RX_CALL_CLEARED) {
4036 /* send an ack now to start the packet flow up again */
4037 call->flags &= ~RX_CALL_CLEARED;
4038 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4040 #ifdef RX_ENABLE_LOCKS
4043 service->nRequestsRunning++;
4044 if (service->nRequestsRunning <= service->minProcs)
4050 MUTEX_EXIT(&rx_serverPool_lock);
4053 /* Delay the sending of an acknowledge event for a short while, while
4054 * a new call is being prepared (in the case of a client) or a reply
4055 * is being prepared (in the case of a server). Rather than sending
4056 * an ack packet, an ACKALL packet is sent. */
4058 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4060 #ifdef RX_ENABLE_LOCKS
4062 MUTEX_ENTER(&call->lock);
4063 call->delayedAckEvent = NULL;
4064 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4066 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4067 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4069 MUTEX_EXIT(&call->lock);
4070 #else /* RX_ENABLE_LOCKS */
4072 call->delayedAckEvent = NULL;
4073 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4074 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4075 #endif /* RX_ENABLE_LOCKS */
4079 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4082 #ifdef RX_ENABLE_LOCKS
4084 MUTEX_ENTER(&call->lock);
4085 if (event == call->delayedAckEvent)
4086 call->delayedAckEvent = NULL;
4087 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4089 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4091 MUTEX_EXIT(&call->lock);
4092 #else /* RX_ENABLE_LOCKS */
4094 call->delayedAckEvent = NULL;
4095 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4096 #endif /* RX_ENABLE_LOCKS */
4100 #ifdef RX_ENABLE_LOCKS
4101 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4102 * clearing them out.
4105 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4107 register struct rx_packet *p, *tp;
4110 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4113 p->flags |= RX_PKTFLAG_ACKED;
4117 call->flags |= RX_CALL_TQ_CLEARME;
4118 call->flags |= RX_CALL_TQ_SOME_ACKED;
4121 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4122 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4123 call->tfirst = call->tnext;
4124 call->nSoftAcked = 0;
4126 if (call->flags & RX_CALL_FAST_RECOVER) {
4127 call->flags &= ~RX_CALL_FAST_RECOVER;
4128 call->cwind = call->nextCwind;
4129 call->nextCwind = 0;
4132 CV_SIGNAL(&call->cv_twind);
4134 #endif /* RX_ENABLE_LOCKS */
4136 /* Clear out the transmit queue for the current call (all packets have
4137 * been received by peer) */
4139 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4141 register struct rx_packet *p, *tp;
4143 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4144 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4146 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4149 p->flags |= RX_PKTFLAG_ACKED;
4153 call->flags |= RX_CALL_TQ_CLEARME;
4154 call->flags |= RX_CALL_TQ_SOME_ACKED;
4157 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4158 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4164 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4165 call->flags &= ~RX_CALL_TQ_CLEARME;
4167 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4169 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4170 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4171 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4172 call->nSoftAcked = 0;
4174 if (call->flags & RX_CALL_FAST_RECOVER) {
4175 call->flags &= ~RX_CALL_FAST_RECOVER;
4176 call->cwind = call->nextCwind;
4178 #ifdef RX_ENABLE_LOCKS
4179 CV_SIGNAL(&call->cv_twind);
4181 osi_rxWakeup(&call->twind);
4186 rxi_ClearReceiveQueue(register struct rx_call *call)
4188 register struct rx_packet *p, *tp;
4189 if (queue_IsNotEmpty(&call->rq)) {
4190 for (queue_Scan(&call->rq, p, tp, rx_packet)) {
4195 rx_packetReclaims++;
4197 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4199 if (call->state == RX_STATE_PRECALL) {
4200 call->flags |= RX_CALL_CLEARED;
4204 /* Send an abort packet for the specified call */
4206 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4207 int istack, int force)
4215 /* Clients should never delay abort messages */
4216 if (rx_IsClientConn(call->conn))
4219 if (call->abortCode != call->error) {
4220 call->abortCode = call->error;
4221 call->abortCount = 0;
4224 if (force || rxi_callAbortThreshhold == 0
4225 || call->abortCount < rxi_callAbortThreshhold) {
4226 if (call->delayedAbortEvent) {
4227 rxevent_Cancel(call->delayedAbortEvent, call,
4228 RX_CALL_REFCOUNT_ABORT);
4230 error = htonl(call->error);
4233 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4234 (char *)&error, sizeof(error), istack);
4235 } else if (!call->delayedAbortEvent) {
4236 clock_GetTime(&when);
4237 clock_Addmsec(&when, rxi_callAbortDelay);
4238 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4239 call->delayedAbortEvent =
4240 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4245 /* Send an abort packet for the specified connection. Packet is an
4246 * optional pointer to a packet that can be used to send the abort.
4247 * Once the number of abort messages reaches the threshhold, an
4248 * event is scheduled to send the abort. Setting the force flag
4249 * overrides sending delayed abort messages.
4251 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4252 * to send the abort packet.
4255 rxi_SendConnectionAbort(register struct rx_connection *conn,
4256 struct rx_packet *packet, int istack, int force)
4264 /* Clients should never delay abort messages */
4265 if (rx_IsClientConn(conn))
4268 if (force || rxi_connAbortThreshhold == 0
4269 || conn->abortCount < rxi_connAbortThreshhold) {
4270 if (conn->delayedAbortEvent) {
4271 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4273 error = htonl(conn->error);
4275 MUTEX_EXIT(&conn->conn_data_lock);
4277 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4278 RX_PACKET_TYPE_ABORT, (char *)&error,
4279 sizeof(error), istack);
4280 MUTEX_ENTER(&conn->conn_data_lock);
4281 } else if (!conn->delayedAbortEvent) {
4282 clock_GetTime(&when);
4283 clock_Addmsec(&when, rxi_connAbortDelay);
4284 conn->delayedAbortEvent =
4285 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4290 /* Associate an error all of the calls owned by a connection. Called
4291 * with error non-zero. This is only for really fatal things, like
4292 * bad authentication responses. The connection itself is set in
4293 * error at this point, so that future packets received will be
4296 rxi_ConnectionError(register struct rx_connection *conn,
4297 register afs_int32 error)
4301 MUTEX_ENTER(&conn->conn_data_lock);
4302 if (conn->challengeEvent)
4303 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4304 if (conn->checkReachEvent) {
4305 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4306 conn->checkReachEvent = 0;
4307 conn->flags &= ~RX_CONN_ATTACHWAIT;
4310 MUTEX_EXIT(&conn->conn_data_lock);
4311 for (i = 0; i < RX_MAXCALLS; i++) {
4312 struct rx_call *call = conn->call[i];
4314 MUTEX_ENTER(&call->lock);
4315 rxi_CallError(call, error);
4316 MUTEX_EXIT(&call->lock);
4319 conn->error = error;
4320 MUTEX_ENTER(&rx_stats_mutex);
4321 rx_stats.fatalErrors++;
4322 MUTEX_EXIT(&rx_stats_mutex);
4327 rxi_CallError(register struct rx_call *call, afs_int32 error)
4330 error = call->error;
4331 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4332 if (!(call->flags & RX_CALL_TQ_BUSY)) {
4333 rxi_ResetCall(call, 0);
4336 rxi_ResetCall(call, 0);
4338 call->error = error;
4339 call->mode = RX_MODE_ERROR;
4342 /* Reset various fields in a call structure, and wakeup waiting
4343 * processes. Some fields aren't changed: state & mode are not
4344 * touched (these must be set by the caller), and bufptr, nLeft, and
4345 * nFree are not reset, since these fields are manipulated by
4346 * unprotected macros, and may only be reset by non-interrupting code.
4349 /* this code requires that call->conn be set properly as a pre-condition. */
4350 #endif /* ADAPT_WINDOW */
4353 rxi_ResetCall(register struct rx_call *call, register int newcall)
4356 register struct rx_peer *peer;
4357 struct rx_packet *packet;
4359 /* Notify anyone who is waiting for asynchronous packet arrival */
4360 if (call->arrivalProc) {
4361 (*call->arrivalProc) (call, call->arrivalProcHandle,
4362 (int)call->arrivalProcArg);
4363 call->arrivalProc = (VOID(*)())0;
4366 if (call->delayedAbortEvent) {
4367 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4368 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4370 rxi_SendCallAbort(call, packet, 0, 1);
4371 rxi_FreePacket(packet);
4376 * Update the peer with the congestion information in this call
4377 * so other calls on this connection can pick up where this call
4378 * left off. If the congestion sequence numbers don't match then
4379 * another call experienced a retransmission.
4381 peer = call->conn->peer;
4382 MUTEX_ENTER(&peer->peer_lock);
4384 if (call->congestSeq == peer->congestSeq) {
4385 peer->cwind = MAX(peer->cwind, call->cwind);
4386 peer->MTU = MAX(peer->MTU, call->MTU);
4387 peer->nDgramPackets =
4388 MAX(peer->nDgramPackets, call->nDgramPackets);
4391 call->abortCode = 0;
4392 call->abortCount = 0;
4394 if (peer->maxDgramPackets > 1) {
4395 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4397 call->MTU = peer->MTU;
4399 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4400 call->ssthresh = rx_maxSendWindow;
4401 call->nDgramPackets = peer->nDgramPackets;
4402 call->congestSeq = peer->congestSeq;
4403 MUTEX_EXIT(&peer->peer_lock);
4405 flags = call->flags;
4406 rxi_ClearReceiveQueue(call);
4407 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4408 if (call->flags & RX_CALL_TQ_BUSY) {
4409 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4410 call->flags |= (flags & RX_CALL_TQ_WAIT);
4412 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4414 rxi_ClearTransmitQueue(call, 0);
4415 queue_Init(&call->tq);
4418 queue_Init(&call->rq);
4420 call->rwind = rx_initReceiveWindow;
4421 call->twind = rx_initSendWindow;
4422 call->nSoftAcked = 0;
4423 call->nextCwind = 0;
4426 call->nCwindAcks = 0;
4427 call->nSoftAcks = 0;
4428 call->nHardAcks = 0;
4430 call->tfirst = call->rnext = call->tnext = 1;
4432 call->lastAcked = 0;
4433 call->localStatus = call->remoteStatus = 0;
4435 if (flags & RX_CALL_READER_WAIT) {
4436 #ifdef RX_ENABLE_LOCKS
4437 CV_BROADCAST(&call->cv_rq);
4439 osi_rxWakeup(&call->rq);
4442 if (flags & RX_CALL_WAIT_PACKETS) {
4443 MUTEX_ENTER(&rx_freePktQ_lock);
4444 rxi_PacketsUnWait(); /* XXX */
4445 MUTEX_EXIT(&rx_freePktQ_lock);
4447 #ifdef RX_ENABLE_LOCKS
4448 CV_SIGNAL(&call->cv_twind);
4450 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4451 osi_rxWakeup(&call->twind);
4454 #ifdef RX_ENABLE_LOCKS
4455 /* The following ensures that we don't mess with any queue while some
4456 * other thread might also be doing so. The call_queue_lock field is
4457 * is only modified under the call lock. If the call is in the process
4458 * of being removed from a queue, the call is not locked until the
4459 * the queue lock is dropped and only then is the call_queue_lock field
4460 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4461 * Note that any other routine which removes a call from a queue has to
4462 * obtain the queue lock before examing the queue and removing the call.
4464 if (call->call_queue_lock) {
4465 MUTEX_ENTER(call->call_queue_lock);
4466 if (queue_IsOnQueue(call)) {
4468 if (flags & RX_CALL_WAIT_PROC) {
4469 MUTEX_ENTER(&rx_stats_mutex);
4471 MUTEX_EXIT(&rx_stats_mutex);
4474 MUTEX_EXIT(call->call_queue_lock);
4475 CLEAR_CALL_QUEUE_LOCK(call);
4477 #else /* RX_ENABLE_LOCKS */
4478 if (queue_IsOnQueue(call)) {
4480 if (flags & RX_CALL_WAIT_PROC)
4483 #endif /* RX_ENABLE_LOCKS */
4485 rxi_KeepAliveOff(call);
4486 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4489 /* Send an acknowledge for the indicated packet (seq,serial) of the
4490 * indicated call, for the indicated reason (reason). This
4491 * acknowledge will specifically acknowledge receiving the packet, and
4492 * will also specify which other packets for this call have been
4493 * received. This routine returns the packet that was used to the
4494 * caller. The caller is responsible for freeing it or re-using it.
4495 * This acknowledgement also returns the highest sequence number
4496 * actually read out by the higher level to the sender; the sender
4497 * promises to keep around packets that have not been read by the
4498 * higher level yet (unless, of course, the sender decides to abort
4499 * the call altogether). Any of p, seq, serial, pflags, or reason may
4500 * be set to zero without ill effect. That is, if they are zero, they
4501 * will not convey any information.
4502 * NOW there is a trailer field, after the ack where it will safely be
4503 * ignored by mundanes, which indicates the maximum size packet this
4504 * host can swallow. */
4506 register struct rx_packet *optionalPacket; use to send ack (or null)
4507 int seq; Sequence number of the packet we are acking
4508 int serial; Serial number of the packet
4509 int pflags; Flags field from packet header
4510 int reason; Reason an acknowledge was prompted
4514 rxi_SendAck(register struct rx_call *call,
4515 register struct rx_packet *optionalPacket, int serial, int reason,
4518 struct rx_ackPacket *ap;
4519 register struct rx_packet *rqp;
4520 register struct rx_packet *nxp; /* For queue_Scan */
4521 register struct rx_packet *p;
4526 * Open the receive window once a thread starts reading packets
4528 if (call->rnext > 1) {
4529 call->rwind = rx_maxReceiveWindow;
4532 call->nHardAcks = 0;
4533 call->nSoftAcks = 0;
4534 if (call->rnext > call->lastAcked)
4535 call->lastAcked = call->rnext;
4539 rx_computelen(p, p->length); /* reset length, you never know */
4540 } /* where that's been... */
4541 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4542 /* We won't send the ack, but don't panic. */
4543 return optionalPacket;
4547 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4550 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL)) {
4551 if (!optionalPacket)
4553 return optionalPacket;
4555 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4556 if (rx_Contiguous(p) < templ) {
4557 if (!optionalPacket)
4559 return optionalPacket;
4564 /* MTUXXX failing to send an ack is very serious. We should */
4565 /* try as hard as possible to send even a partial ack; it's */
4566 /* better than nothing. */
4567 ap = (struct rx_ackPacket *)rx_DataOf(p);
4568 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4569 ap->reason = reason;
4571 /* The skew computation used to be bogus, I think it's better now. */
4572 /* We should start paying attention to skew. XXX */
4573 ap->serial = htonl(serial);
4574 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4576 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4577 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4579 /* No fear of running out of ack packet here because there can only be at most
4580 * one window full of unacknowledged packets. The window size must be constrained
4581 * to be less than the maximum ack size, of course. Also, an ack should always
4582 * fit into a single packet -- it should not ever be fragmented. */
4583 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4584 if (!rqp || !call->rq.next
4585 || (rqp->header.seq > (call->rnext + call->rwind))) {
4586 if (!optionalPacket)
4588 rxi_CallError(call, RX_CALL_DEAD);
4589 return optionalPacket;
4592 while (rqp->header.seq > call->rnext + offset)
4593 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4594 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4596 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4597 if (!optionalPacket)
4599 rxi_CallError(call, RX_CALL_DEAD);
4600 return optionalPacket;
4605 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4607 /* these are new for AFS 3.3 */
4608 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4609 templ = htonl(templ);
4610 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4611 templ = htonl(call->conn->peer->ifMTU);
4612 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4613 sizeof(afs_int32), &templ);
4615 /* new for AFS 3.4 */
4616 templ = htonl(call->rwind);
4617 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4618 sizeof(afs_int32), &templ);
4620 /* new for AFS 3.5 */
4621 templ = htonl(call->conn->peer->ifDgramPackets);
4622 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4623 sizeof(afs_int32), &templ);
4625 p->header.serviceId = call->conn->serviceId;
4626 p->header.cid = (call->conn->cid | call->channel);
4627 p->header.callNumber = *call->callNumber;
4629 p->header.securityIndex = call->conn->securityIndex;
4630 p->header.epoch = call->conn->epoch;
4631 p->header.type = RX_PACKET_TYPE_ACK;
4632 p->header.flags = RX_SLOW_START_OK;
4633 if (reason == RX_ACK_PING) {
4634 p->header.flags |= RX_REQUEST_ACK;
4636 clock_GetTime(&call->pingRequestTime);
4639 if (call->conn->type == RX_CLIENT_CONNECTION)
4640 p->header.flags |= RX_CLIENT_INITIATED;
4644 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u",
4645 ap->reason, ntohl(ap->previousPacket),
4646 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4648 for (offset = 0; offset < ap->nAcks; offset++)
4649 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4657 register int i, nbytes = p->length;
4659 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4660 if (nbytes <= p->wirevec[i].iov_len) {
4661 register int savelen, saven;
4663 savelen = p->wirevec[i].iov_len;
4665 p->wirevec[i].iov_len = nbytes;
4667 rxi_Send(call, p, istack);
4668 p->wirevec[i].iov_len = savelen;
4672 nbytes -= p->wirevec[i].iov_len;
4675 MUTEX_ENTER(&rx_stats_mutex);
4676 rx_stats.ackPacketsSent++;
4677 MUTEX_EXIT(&rx_stats_mutex);
4678 if (!optionalPacket)
4680 return optionalPacket; /* Return packet for re-use by caller */
4683 /* Send all of the packets in the list in single datagram */
4685 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4686 int istack, int moreFlag, struct clock *now,
4687 struct clock *retryTime, int resending)
4692 struct rx_connection *conn = call->conn;
4693 struct rx_peer *peer = conn->peer;
4695 MUTEX_ENTER(&peer->peer_lock);
4698 peer->reSends += len;
4699 MUTEX_ENTER(&rx_stats_mutex);
4700 rx_stats.dataPacketsSent += len;
4701 MUTEX_EXIT(&rx_stats_mutex);
4702 MUTEX_EXIT(&peer->peer_lock);
4704 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4708 /* Set the packet flags and schedule the resend events */
4709 /* Only request an ack for the last packet in the list */
4710 for (i = 0; i < len; i++) {
4711 list[i]->retryTime = *retryTime;
4712 if (list[i]->header.serial) {
4713 /* Exponentially backoff retry times */
4714 if (list[i]->backoff < MAXBACKOFF) {
4715 /* so it can't stay == 0 */
4716 list[i]->backoff = (list[i]->backoff << 1) + 1;
4719 clock_Addmsec(&(list[i]->retryTime),
4720 ((afs_uint32) list[i]->backoff) << 8);
4723 /* Wait a little extra for the ack on the last packet */
4724 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4725 clock_Addmsec(&(list[i]->retryTime), 400);
4728 /* Record the time sent */
4729 list[i]->timeSent = *now;
4731 /* Ask for an ack on retransmitted packets, on every other packet
4732 * if the peer doesn't support slow start. Ask for an ack on every
4733 * packet until the congestion window reaches the ack rate. */
4734 if (list[i]->header.serial) {
4736 MUTEX_ENTER(&rx_stats_mutex);
4737 rx_stats.dataPacketsReSent++;
4738 MUTEX_EXIT(&rx_stats_mutex);
4740 /* improved RTO calculation- not Karn */
4741 list[i]->firstSent = *now;
4742 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4743 || (!(call->flags & RX_CALL_SLOW_START_OK)
4744 && (list[i]->header.seq & 1)))) {
4749 MUTEX_ENTER(&peer->peer_lock);
4753 MUTEX_ENTER(&rx_stats_mutex);
4754 rx_stats.dataPacketsSent++;
4755 MUTEX_EXIT(&rx_stats_mutex);
4756 MUTEX_EXIT(&peer->peer_lock);
4758 /* Tag this packet as not being the last in this group,
4759 * for the receiver's benefit */
4760 if (i < len - 1 || moreFlag) {
4761 list[i]->header.flags |= RX_MORE_PACKETS;
4764 /* Install the new retransmit time for the packet, and
4765 * record the time sent */
4766 list[i]->timeSent = *now;
4770 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4773 /* Since we're about to send a data packet to the peer, it's
4774 * safe to nuke any scheduled end-of-packets ack */
4775 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4777 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4778 MUTEX_EXIT(&call->lock);
4780 rxi_SendPacketList(call, conn, list, len, istack);
4782 rxi_SendPacket(call, conn, list[0], istack);
4784 MUTEX_ENTER(&call->lock);
4785 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4787 /* Update last send time for this call (for keep-alive
4788 * processing), and for the connection (so that we can discover
4789 * idle connections) */
4790 conn->lastSendTime = call->lastSendTime = clock_Sec();
4793 /* When sending packets we need to follow these rules:
4794 * 1. Never send more than maxDgramPackets in a jumbogram.
4795 * 2. Never send a packet with more than two iovecs in a jumbogram.
4796 * 3. Never send a retransmitted packet in a jumbogram.
4797 * 4. Never send more than cwind/4 packets in a jumbogram
4798 * We always keep the last list we should have sent so we
4799 * can set the RX_MORE_PACKETS flags correctly.
4802 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4803 int istack, struct clock *now, struct clock *retryTime,
4806 int i, cnt, lastCnt = 0;
4807 struct rx_packet **listP, **lastP = 0;
4808 struct rx_peer *peer = call->conn->peer;
4809 int morePackets = 0;
4811 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4812 /* Does the current packet force us to flush the current list? */
4814 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4815 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4817 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4819 /* If the call enters an error state stop sending, or if
4820 * we entered congestion recovery mode, stop sending */
4821 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4829 /* Add the current packet to the list if it hasn't been acked.
4830 * Otherwise adjust the list pointer to skip the current packet. */
4831 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
4833 /* Do we need to flush the list? */
4834 if (cnt >= (int)peer->maxDgramPackets
4835 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
4836 || list[i]->header.serial
4837 || list[i]->length != RX_JUMBOBUFFERSIZE) {
4839 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
4840 retryTime, resending);
4841 /* If the call enters an error state stop sending, or if
4842 * we entered congestion recovery mode, stop sending */
4844 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4849 listP = &list[i + 1];
4854 osi_Panic("rxi_SendList error");
4856 listP = &list[i + 1];
4860 /* Send the whole list when the call is in receive mode, when
4861 * the call is in eof mode, when we are in fast recovery mode,
4862 * and when we have the last packet */
4863 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
4864 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
4865 || (call->flags & RX_CALL_FAST_RECOVER)) {
4866 /* Check for the case where the current list contains
4867 * an acked packet. Since we always send retransmissions
4868 * in a separate packet, we only need to check the first
4869 * packet in the list */
4870 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
4874 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
4875 retryTime, resending);
4876 /* If the call enters an error state stop sending, or if
4877 * we entered congestion recovery mode, stop sending */
4878 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4882 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
4885 } else if (lastCnt > 0) {
4886 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
4891 #ifdef RX_ENABLE_LOCKS
4892 /* Call rxi_Start, below, but with the call lock held. */
4894 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
4897 MUTEX_ENTER(&call->lock);
4898 rxi_Start(event, call, istack);
4899 MUTEX_EXIT(&call->lock);
4901 #endif /* RX_ENABLE_LOCKS */
4903 /* This routine is called when new packets are readied for
4904 * transmission and when retransmission may be necessary, or when the
4905 * transmission window or burst count are favourable. This should be
4906 * better optimized for new packets, the usual case, now that we've
4907 * got rid of queues of send packets. XXXXXXXXXXX */
4909 rxi_Start(struct rxevent *event, register struct rx_call *call, int istack)
4911 struct rx_packet *p;
4912 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
4913 struct rx_peer *peer = call->conn->peer;
4914 struct clock now, retryTime;
4918 struct rx_packet **xmitList;
4921 /* If rxi_Start is being called as a result of a resend event,
4922 * then make sure that the event pointer is removed from the call
4923 * structure, since there is no longer a per-call retransmission
4925 if (event && event == call->resendEvent) {
4926 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
4927 call->resendEvent = NULL;
4929 if (queue_IsEmpty(&call->tq)) {
4933 /* Timeouts trigger congestion recovery */
4934 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4935 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4936 /* someone else is waiting to start recovery */
4939 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4940 while (call->flags & RX_CALL_TQ_BUSY) {
4941 call->flags |= RX_CALL_TQ_WAIT;
4942 #ifdef RX_ENABLE_LOCKS
4943 CV_WAIT(&call->cv_tq, &call->lock);
4944 #else /* RX_ENABLE_LOCKS */
4945 osi_rxSleep(&call->tq);
4946 #endif /* RX_ENABLE_LOCKS */
4948 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4949 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4950 call->flags |= RX_CALL_FAST_RECOVER;
4951 if (peer->maxDgramPackets > 1) {
4952 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4954 call->MTU = MIN(peer->natMTU, peer->maxMTU);
4956 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4957 call->nDgramPackets = 1;
4959 call->nextCwind = 1;
4962 MUTEX_ENTER(&peer->peer_lock);
4963 peer->MTU = call->MTU;
4964 peer->cwind = call->cwind;
4965 peer->nDgramPackets = 1;
4967 call->congestSeq = peer->congestSeq;
4968 MUTEX_EXIT(&peer->peer_lock);
4969 /* Clear retry times on packets. Otherwise, it's possible for
4970 * some packets in the queue to force resends at rates faster
4971 * than recovery rates.
4973 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
4974 if (!(p->flags & RX_PKTFLAG_ACKED)) {
4975 clock_Zero(&p->retryTime);
4980 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4981 MUTEX_ENTER(&rx_stats_mutex);
4982 rx_tq_debug.rxi_start_in_error++;
4983 MUTEX_EXIT(&rx_stats_mutex);
4988 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
4989 /* Get clock to compute the re-transmit time for any packets
4990 * in this burst. Note, if we back off, it's reasonable to
4991 * back off all of the packets in the same manner, even if
4992 * some of them have been retransmitted more times than more
4993 * recent additions */
4994 clock_GetTime(&now);
4995 retryTime = now; /* initialize before use */
4996 MUTEX_ENTER(&peer->peer_lock);
4997 clock_Add(&retryTime, &peer->timeout);
4998 MUTEX_EXIT(&peer->peer_lock);
5000 /* Send (or resend) any packets that need it, subject to
5001 * window restrictions and congestion burst control
5002 * restrictions. Ask for an ack on the last packet sent in
5003 * this burst. For now, we're relying upon the window being
5004 * considerably bigger than the largest number of packets that
5005 * are typically sent at once by one initial call to
5006 * rxi_Start. This is probably bogus (perhaps we should ask
5007 * for an ack when we're half way through the current
5008 * window?). Also, for non file transfer applications, this
5009 * may end up asking for an ack for every packet. Bogus. XXXX
5012 * But check whether we're here recursively, and let the other guy
5015 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5016 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5017 call->flags |= RX_CALL_TQ_BUSY;
5019 call->flags &= ~RX_CALL_NEED_START;
5020 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5022 maxXmitPackets = MIN(call->twind, call->cwind);
5023 xmitList = (struct rx_packet **)
5024 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5025 if (xmitList == NULL)
5026 osi_Panic("rxi_Start, failed to allocate xmit list");
5027 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5028 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5029 /* We shouldn't be sending packets if a thread is waiting
5030 * to initiate congestion recovery */
5034 && (call->flags & RX_CALL_FAST_RECOVER)) {
5035 /* Only send one packet during fast recovery */
5038 if ((p->flags & RX_PKTFLAG_FREE)
5039 || (!queue_IsEnd(&call->tq, nxp)
5040 && (nxp->flags & RX_PKTFLAG_FREE))
5041 || (p == (struct rx_packet *)&rx_freePacketQueue)
5042 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5043 osi_Panic("rxi_Start: xmit queue clobbered");
5045 if (p->flags & RX_PKTFLAG_ACKED) {
5046 MUTEX_ENTER(&rx_stats_mutex);
5047 rx_stats.ignoreAckedPacket++;
5048 MUTEX_EXIT(&rx_stats_mutex);
5049 continue; /* Ignore this packet if it has been acknowledged */
5052 /* Turn off all flags except these ones, which are the same
5053 * on each transmission */
5054 p->header.flags &= RX_PRESET_FLAGS;
5056 if (p->header.seq >=
5057 call->tfirst + MIN((int)call->twind,
5058 (int)(call->nSoftAcked +
5060 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5061 /* Note: if we're waiting for more window space, we can
5062 * still send retransmits; hence we don't return here, but
5063 * break out to schedule a retransmit event */
5064 dpf(("call %d waiting for window",
5065 *(call->callNumber)));
5069 /* Transmit the packet if it needs to be sent. */
5070 if (!clock_Lt(&now, &p->retryTime)) {
5071 if (nXmitPackets == maxXmitPackets) {
5072 osi_Panic("rxi_Start: xmit list overflowed");
5074 xmitList[nXmitPackets++] = p;
5078 /* xmitList now hold pointers to all of the packets that are
5079 * ready to send. Now we loop to send the packets */
5080 if (nXmitPackets > 0) {
5081 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5082 &now, &retryTime, resending);
5085 maxXmitPackets * sizeof(struct rx_packet *));
5087 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5089 * TQ references no longer protected by this flag; they must remain
5090 * protected by the global lock.
5092 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5093 call->flags &= ~RX_CALL_TQ_BUSY;
5094 if (call->flags & RX_CALL_TQ_WAIT) {
5095 call->flags &= ~RX_CALL_TQ_WAIT;
5096 #ifdef RX_ENABLE_LOCKS
5097 CV_BROADCAST(&call->cv_tq);
5098 #else /* RX_ENABLE_LOCKS */
5099 osi_rxWakeup(&call->tq);
5100 #endif /* RX_ENABLE_LOCKS */
5105 /* We went into the error state while sending packets. Now is
5106 * the time to reset the call. This will also inform the using
5107 * process that the call is in an error state.
5109 MUTEX_ENTER(&rx_stats_mutex);
5110 rx_tq_debug.rxi_start_aborted++;
5111 MUTEX_EXIT(&rx_stats_mutex);
5112 call->flags &= ~RX_CALL_TQ_BUSY;
5113 if (call->flags & RX_CALL_TQ_WAIT) {
5114 call->flags &= ~RX_CALL_TQ_WAIT;
5115 #ifdef RX_ENABLE_LOCKS
5116 CV_BROADCAST(&call->cv_tq);
5117 #else /* RX_ENABLE_LOCKS */
5118 osi_rxWakeup(&call->tq);
5119 #endif /* RX_ENABLE_LOCKS */
5121 rxi_CallError(call, call->error);
5124 #ifdef RX_ENABLE_LOCKS
5125 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5126 register int missing;
5127 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5128 /* Some packets have received acks. If they all have, we can clear
5129 * the transmit queue.
5132 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5133 if (p->header.seq < call->tfirst
5134 && (p->flags & RX_PKTFLAG_ACKED)) {
5141 call->flags |= RX_CALL_TQ_CLEARME;
5143 #endif /* RX_ENABLE_LOCKS */
5144 /* Don't bother doing retransmits if the TQ is cleared. */
5145 if (call->flags & RX_CALL_TQ_CLEARME) {
5146 rxi_ClearTransmitQueue(call, 1);
5148 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5151 /* Always post a resend event, if there is anything in the
5152 * queue, and resend is possible. There should be at least
5153 * one unacknowledged packet in the queue ... otherwise none
5154 * of these packets should be on the queue in the first place.
5156 if (call->resendEvent) {
5157 /* Cancel the existing event and post a new one */
5158 rxevent_Cancel(call->resendEvent, call,
5159 RX_CALL_REFCOUNT_RESEND);
5162 /* The retry time is the retry time on the first unacknowledged
5163 * packet inside the current window */
5165 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5166 /* Don't set timers for packets outside the window */
5167 if (p->header.seq >= call->tfirst + call->twind) {
5171 if (!(p->flags & RX_PKTFLAG_ACKED)
5172 && !clock_IsZero(&p->retryTime)) {
5174 retryTime = p->retryTime;
5179 /* Post a new event to re-run rxi_Start when retries may be needed */
5180 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5181 #ifdef RX_ENABLE_LOCKS
5182 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5184 rxevent_Post(&retryTime, rxi_StartUnlocked,
5185 (void *)call, (void *)istack);
5186 #else /* RX_ENABLE_LOCKS */
5188 rxevent_Post(&retryTime, rxi_Start, (void *)call,
5190 #endif /* RX_ENABLE_LOCKS */
5193 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5194 } while (call->flags & RX_CALL_NEED_START);
5196 * TQ references no longer protected by this flag; they must remain
5197 * protected by the global lock.
5199 call->flags &= ~RX_CALL_TQ_BUSY;
5200 if (call->flags & RX_CALL_TQ_WAIT) {
5201 call->flags &= ~RX_CALL_TQ_WAIT;
5202 #ifdef RX_ENABLE_LOCKS
5203 CV_BROADCAST(&call->cv_tq);
5204 #else /* RX_ENABLE_LOCKS */
5205 osi_rxWakeup(&call->tq);
5206 #endif /* RX_ENABLE_LOCKS */
5209 call->flags |= RX_CALL_NEED_START;
5211 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5213 if (call->resendEvent) {
5214 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5219 /* Also adjusts the keep alive parameters for the call, to reflect
5220 * that we have just sent a packet (so keep alives aren't sent
5223 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5226 register struct rx_connection *conn = call->conn;
5228 /* Stamp each packet with the user supplied status */
5229 p->header.userStatus = call->localStatus;
5231 /* Allow the security object controlling this call's security to
5232 * make any last-minute changes to the packet */
5233 RXS_SendPacket(conn->securityObject, call, p);
5235 /* Since we're about to send SOME sort of packet to the peer, it's
5236 * safe to nuke any scheduled end-of-packets ack */
5237 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5239 /* Actually send the packet, filling in more connection-specific fields */
5240 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5241 MUTEX_EXIT(&call->lock);
5242 rxi_SendPacket(call, conn, p, istack);
5243 MUTEX_ENTER(&call->lock);
5244 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5246 /* Update last send time for this call (for keep-alive
5247 * processing), and for the connection (so that we can discover
5248 * idle connections) */
5249 conn->lastSendTime = call->lastSendTime = clock_Sec();
5253 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5254 * that things are fine. Also called periodically to guarantee that nothing
5255 * falls through the cracks (e.g. (error + dally) connections have keepalive
5256 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5258 * haveCTLock Set if calling from rxi_ReapConnections
5260 #ifdef RX_ENABLE_LOCKS
5262 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5263 #else /* RX_ENABLE_LOCKS */
5265 rxi_CheckCall(register struct rx_call *call)
5266 #endif /* RX_ENABLE_LOCKS */
5268 register struct rx_connection *conn = call->conn;
5270 afs_uint32 deadTime;
5272 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5273 if (call->flags & RX_CALL_TQ_BUSY) {
5274 /* Call is active and will be reset by rxi_Start if it's
5275 * in an error state.
5280 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5282 (((afs_uint32) conn->secondsUntilDead << 10) +
5283 ((afs_uint32) conn->peer->rtt >> 3) +
5284 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5286 /* These are computed to the second (+- 1 second). But that's
5287 * good enough for these values, which should be a significant
5288 * number of seconds. */
5289 if (now > (call->lastReceiveTime + deadTime)) {
5290 if (call->state == RX_STATE_ACTIVE) {
5291 rxi_CallError(call, RX_CALL_DEAD);
5294 #ifdef RX_ENABLE_LOCKS
5295 /* Cancel pending events */
5296 rxevent_Cancel(call->delayedAckEvent, call,
5297 RX_CALL_REFCOUNT_DELAY);
5298 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5299 rxevent_Cancel(call->keepAliveEvent, call,
5300 RX_CALL_REFCOUNT_ALIVE);
5301 if (call->refCount == 0) {
5302 rxi_FreeCall(call, haveCTLock);
5306 #else /* RX_ENABLE_LOCKS */
5309 #endif /* RX_ENABLE_LOCKS */
5311 /* Non-active calls are destroyed if they are not responding
5312 * to pings; active calls are simply flagged in error, so the
5313 * attached process can die reasonably gracefully. */
5315 /* see if we have a non-activity timeout */
5316 if (call->startWait && conn->idleDeadTime
5317 && ((call->startWait + conn->idleDeadTime) < now)) {
5318 if (call->state == RX_STATE_ACTIVE) {
5319 rxi_CallError(call, RX_CALL_TIMEOUT);
5323 /* see if we have a hard timeout */
5324 if (conn->hardDeadTime
5325 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5326 if (call->state == RX_STATE_ACTIVE)
5327 rxi_CallError(call, RX_CALL_TIMEOUT);
5334 /* When a call is in progress, this routine is called occasionally to
5335 * make sure that some traffic has arrived (or been sent to) the peer.
5336 * If nothing has arrived in a reasonable amount of time, the call is
5337 * declared dead; if nothing has been sent for a while, we send a
5338 * keep-alive packet (if we're actually trying to keep the call alive)
5341 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5344 struct rx_connection *conn;
5347 MUTEX_ENTER(&call->lock);
5348 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5349 if (event == call->keepAliveEvent)
5350 call->keepAliveEvent = NULL;
5353 #ifdef RX_ENABLE_LOCKS
5354 if (rxi_CheckCall(call, 0)) {
5355 MUTEX_EXIT(&call->lock);
5358 #else /* RX_ENABLE_LOCKS */
5359 if (rxi_CheckCall(call))
5361 #endif /* RX_ENABLE_LOCKS */
5363 /* Don't try to keep alive dallying calls */
5364 if (call->state == RX_STATE_DALLY) {
5365 MUTEX_EXIT(&call->lock);
5370 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5371 /* Don't try to send keepalives if there is unacknowledged data */
5372 /* the rexmit code should be good enough, this little hack
5373 * doesn't quite work XXX */
5374 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5376 rxi_ScheduleKeepAliveEvent(call);
5377 MUTEX_EXIT(&call->lock);
5382 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5384 if (!call->keepAliveEvent) {
5386 clock_GetTime(&when);
5387 when.sec += call->conn->secondsUntilPing;
5388 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5389 call->keepAliveEvent =
5390 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5394 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5396 rxi_KeepAliveOn(register struct rx_call *call)
5398 /* Pretend last packet received was received now--i.e. if another
5399 * packet isn't received within the keep alive time, then the call
5400 * will die; Initialize last send time to the current time--even
5401 * if a packet hasn't been sent yet. This will guarantee that a
5402 * keep-alive is sent within the ping time */
5403 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5404 rxi_ScheduleKeepAliveEvent(call);
5407 /* This routine is called to send connection abort messages
5408 * that have been delayed to throttle looping clients. */
5410 rxi_SendDelayedConnAbort(struct rxevent *event,
5411 register struct rx_connection *conn, char *dummy)
5414 struct rx_packet *packet;
5416 MUTEX_ENTER(&conn->conn_data_lock);
5417 conn->delayedAbortEvent = NULL;
5418 error = htonl(conn->error);
5420 MUTEX_EXIT(&conn->conn_data_lock);
5421 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5424 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5425 RX_PACKET_TYPE_ABORT, (char *)&error,
5427 rxi_FreePacket(packet);
5431 /* This routine is called to send call abort messages
5432 * that have been delayed to throttle looping clients. */
5434 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5438 struct rx_packet *packet;
5440 MUTEX_ENTER(&call->lock);
5441 call->delayedAbortEvent = NULL;
5442 error = htonl(call->error);
5444 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5447 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5448 (char *)&error, sizeof(error), 0);
5449 rxi_FreePacket(packet);
5451 MUTEX_EXIT(&call->lock);
5454 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5455 * seconds) to ask the client to authenticate itself. The routine
5456 * issues a challenge to the client, which is obtained from the
5457 * security object associated with the connection */
5459 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5462 int tries = (int)atries;
5463 conn->challengeEvent = NULL;
5464 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5465 register struct rx_packet *packet;
5469 /* We've failed to authenticate for too long.
5470 * Reset any calls waiting for authentication;
5471 * they are all in RX_STATE_PRECALL.
5475 MUTEX_ENTER(&conn->conn_call_lock);
5476 for (i = 0; i < RX_MAXCALLS; i++) {
5477 struct rx_call *call = conn->call[i];
5479 MUTEX_ENTER(&call->lock);
5480 if (call->state == RX_STATE_PRECALL) {
5481 rxi_CallError(call, RX_CALL_DEAD);
5482 rxi_SendCallAbort(call, NULL, 0, 0);
5484 MUTEX_EXIT(&call->lock);
5487 MUTEX_EXIT(&conn->conn_call_lock);
5491 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5493 /* If there's no packet available, do this later. */
5494 RXS_GetChallenge(conn->securityObject, conn, packet);
5495 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5496 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5497 rxi_FreePacket(packet);
5499 clock_GetTime(&when);
5500 when.sec += RX_CHALLENGE_TIMEOUT;
5501 conn->challengeEvent =
5502 rxevent_Post(&when, rxi_ChallengeEvent, conn,
5503 (void *)(tries - 1));
5507 /* Call this routine to start requesting the client to authenticate
5508 * itself. This will continue until authentication is established,
5509 * the call times out, or an invalid response is returned. The
5510 * security object associated with the connection is asked to create
5511 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5512 * defined earlier. */
5514 rxi_ChallengeOn(register struct rx_connection *conn)
5516 if (!conn->challengeEvent) {
5517 RXS_CreateChallenge(conn->securityObject, conn);
5518 rxi_ChallengeEvent(NULL, conn, (void *)RX_CHALLENGE_MAXTRIES);
5523 /* Compute round trip time of the packet provided, in *rttp.
5526 /* rxi_ComputeRoundTripTime is called with peer locked. */
5527 /* sentp and/or peer may be null */
5529 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5530 register struct clock *sentp,
5531 register struct rx_peer *peer)
5533 struct clock thisRtt, *rttp = &thisRtt;
5535 #if defined(AFS_ALPHA_LINUX22_ENV) && defined(AFS_PTHREAD_ENV) && !defined(KERNEL)
5536 /* making year 2038 bugs to get this running now - stroucki */
5537 struct timeval temptime;
5539 register int rtt_timeout;
5541 #if defined(AFS_ALPHA_LINUX20_ENV) && defined(AFS_PTHREAD_ENV) && !defined(KERNEL)
5542 /* yet again. This was the worst Heisenbug of the port - stroucki */
5543 clock_GetTime(&temptime);
5544 rttp->sec = (afs_int32) temptime.tv_sec;
5545 rttp->usec = (afs_int32) temptime.tv_usec;
5547 clock_GetTime(rttp);
5549 if (clock_Lt(rttp, sentp)) {
5551 return; /* somebody set the clock back, don't count this time. */
5553 clock_Sub(rttp, sentp);
5554 MUTEX_ENTER(&rx_stats_mutex);
5555 if (clock_Lt(rttp, &rx_stats.minRtt))
5556 rx_stats.minRtt = *rttp;
5557 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5558 if (rttp->sec > 60) {
5559 MUTEX_EXIT(&rx_stats_mutex);
5560 return; /* somebody set the clock ahead */
5562 rx_stats.maxRtt = *rttp;
5564 clock_Add(&rx_stats.totalRtt, rttp);
5565 rx_stats.nRttSamples++;
5566 MUTEX_EXIT(&rx_stats_mutex);
5568 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5570 /* Apply VanJacobson round-trip estimations */
5575 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5576 * srtt is stored as fixed point with 3 bits after the binary
5577 * point (i.e., scaled by 8). The following magic is
5578 * equivalent to the smoothing algorithm in rfc793 with an
5579 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5580 * srtt*8 = srtt*8 + rtt - srtt
5581 * srtt = srtt + rtt/8 - srtt/8
5584 delta = MSEC(rttp) - (peer->rtt >> 3);
5588 * We accumulate a smoothed rtt variance (actually, a smoothed
5589 * mean difference), then set the retransmit timer to smoothed
5590 * rtt + 4 times the smoothed variance (was 2x in van's original
5591 * paper, but 4x works better for me, and apparently for him as
5593 * rttvar is stored as
5594 * fixed point with 2 bits after the binary point (scaled by
5595 * 4). The following is equivalent to rfc793 smoothing with
5596 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5597 * replaces rfc793's wired-in beta.
5598 * dev*4 = dev*4 + (|actual - expected| - dev)
5604 delta -= (peer->rtt_dev >> 2);
5605 peer->rtt_dev += delta;
5607 /* I don't have a stored RTT so I start with this value. Since I'm
5608 * probably just starting a call, and will be pushing more data down
5609 * this, I expect congestion to increase rapidly. So I fudge a
5610 * little, and I set deviance to half the rtt. In practice,
5611 * deviance tends to approach something a little less than
5612 * half the smoothed rtt. */
5613 peer->rtt = (MSEC(rttp) << 3) + 8;
5614 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5616 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5617 * the other of these connections is usually in a user process, and can
5618 * be switched and/or swapped out. So on fast, reliable networks, the
5619 * timeout would otherwise be too short.
5621 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5622 clock_Zero(&(peer->timeout));
5623 clock_Addmsec(&(peer->timeout), rtt_timeout);
5625 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)));
5629 /* Find all server connections that have not been active for a long time, and
5632 rxi_ReapConnections(void)
5635 clock_GetTime(&now);
5637 /* Find server connection structures that haven't been used for
5638 * greater than rx_idleConnectionTime */
5640 struct rx_connection **conn_ptr, **conn_end;
5641 int i, havecalls = 0;
5642 MUTEX_ENTER(&rx_connHashTable_lock);
5643 for (conn_ptr = &rx_connHashTable[0], conn_end =
5644 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5646 struct rx_connection *conn, *next;
5647 struct rx_call *call;
5651 for (conn = *conn_ptr; conn; conn = next) {
5652 /* XXX -- Shouldn't the connection be locked? */
5655 for (i = 0; i < RX_MAXCALLS; i++) {
5656 call = conn->call[i];
5659 MUTEX_ENTER(&call->lock);
5660 #ifdef RX_ENABLE_LOCKS
5661 result = rxi_CheckCall(call, 1);
5662 #else /* RX_ENABLE_LOCKS */
5663 result = rxi_CheckCall(call);
5664 #endif /* RX_ENABLE_LOCKS */
5665 MUTEX_EXIT(&call->lock);
5667 /* If CheckCall freed the call, it might
5668 * have destroyed the connection as well,
5669 * which screws up the linked lists.
5675 if (conn->type == RX_SERVER_CONNECTION) {
5676 /* This only actually destroys the connection if
5677 * there are no outstanding calls */
5678 MUTEX_ENTER(&conn->conn_data_lock);
5679 if (!havecalls && !conn->refCount
5680 && ((conn->lastSendTime + rx_idleConnectionTime) <
5682 conn->refCount++; /* it will be decr in rx_DestroyConn */
5683 MUTEX_EXIT(&conn->conn_data_lock);
5684 #ifdef RX_ENABLE_LOCKS
5685 rxi_DestroyConnectionNoLock(conn);
5686 #else /* RX_ENABLE_LOCKS */
5687 rxi_DestroyConnection(conn);
5688 #endif /* RX_ENABLE_LOCKS */
5690 #ifdef RX_ENABLE_LOCKS
5692 MUTEX_EXIT(&conn->conn_data_lock);
5694 #endif /* RX_ENABLE_LOCKS */
5698 #ifdef RX_ENABLE_LOCKS
5699 while (rx_connCleanup_list) {
5700 struct rx_connection *conn;
5701 conn = rx_connCleanup_list;
5702 rx_connCleanup_list = rx_connCleanup_list->next;
5703 MUTEX_EXIT(&rx_connHashTable_lock);
5704 rxi_CleanupConnection(conn);
5705 MUTEX_ENTER(&rx_connHashTable_lock);
5707 MUTEX_EXIT(&rx_connHashTable_lock);
5708 #endif /* RX_ENABLE_LOCKS */
5711 /* Find any peer structures that haven't been used (haven't had an
5712 * associated connection) for greater than rx_idlePeerTime */
5714 struct rx_peer **peer_ptr, **peer_end;
5716 MUTEX_ENTER(&rx_rpc_stats);
5717 MUTEX_ENTER(&rx_peerHashTable_lock);
5718 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5719 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5721 struct rx_peer *peer, *next, *prev;
5722 for (prev = peer = *peer_ptr; peer; peer = next) {
5724 code = MUTEX_TRYENTER(&peer->peer_lock);
5725 if ((code) && (peer->refCount == 0)
5726 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5727 rx_interface_stat_p rpc_stat, nrpc_stat;
5729 MUTEX_EXIT(&peer->peer_lock);
5730 MUTEX_DESTROY(&peer->peer_lock);
5732 (&peer->rpcStats, rpc_stat, nrpc_stat,
5733 rx_interface_stat)) {
5734 unsigned int num_funcs;
5737 queue_Remove(&rpc_stat->queue_header);
5738 queue_Remove(&rpc_stat->all_peers);
5739 num_funcs = rpc_stat->stats[0].func_total;
5741 sizeof(rx_interface_stat_t) +
5742 rpc_stat->stats[0].func_total *
5743 sizeof(rx_function_entry_v1_t);
5745 rxi_Free(rpc_stat, space);
5746 rxi_rpc_peer_stat_cnt -= num_funcs;
5749 MUTEX_ENTER(&rx_stats_mutex);
5750 rx_stats.nPeerStructs--;
5751 MUTEX_EXIT(&rx_stats_mutex);
5752 if (prev == *peer_ptr) {
5759 MUTEX_EXIT(&peer->peer_lock);
5765 MUTEX_EXIT(&rx_peerHashTable_lock);
5766 MUTEX_EXIT(&rx_rpc_stats);
5769 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5770 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5771 * GC, just below. Really, we shouldn't have to keep moving packets from
5772 * one place to another, but instead ought to always know if we can
5773 * afford to hold onto a packet in its particular use. */
5774 MUTEX_ENTER(&rx_freePktQ_lock);
5775 if (rx_waitingForPackets) {
5776 rx_waitingForPackets = 0;
5777 #ifdef RX_ENABLE_LOCKS
5778 CV_BROADCAST(&rx_waitingForPackets_cv);
5780 osi_rxWakeup(&rx_waitingForPackets);
5783 MUTEX_EXIT(&rx_freePktQ_lock);
5785 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5786 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5790 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5791 * rx.h is sort of strange this is better. This is called with a security
5792 * object before it is discarded. Each connection using a security object has
5793 * its own refcount to the object so it won't actually be freed until the last
5794 * connection is destroyed.
5796 * This is the only rxs module call. A hold could also be written but no one
5800 rxs_Release(struct rx_securityClass *aobj)
5802 return RXS_Close(aobj);
5806 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5807 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5808 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5809 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5811 /* Adjust our estimate of the transmission rate to this peer, given
5812 * that the packet p was just acked. We can adjust peer->timeout and
5813 * call->twind. Pragmatically, this is called
5814 * only with packets of maximal length.
5815 * Called with peer and call locked.
5819 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5820 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5822 afs_int32 xferSize, xferMs;
5823 register afs_int32 minTime;
5826 /* Count down packets */
5827 if (peer->rateFlag > 0)
5829 /* Do nothing until we're enabled */
5830 if (peer->rateFlag != 0)
5835 /* Count only when the ack seems legitimate */
5836 switch (ackReason) {
5837 case RX_ACK_REQUESTED:
5839 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
5843 case RX_ACK_PING_RESPONSE:
5844 if (p) /* want the response to ping-request, not data send */
5846 clock_GetTime(&newTO);
5847 if (clock_Gt(&newTO, &call->pingRequestTime)) {
5848 clock_Sub(&newTO, &call->pingRequestTime);
5849 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
5853 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
5860 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));
5862 /* Track only packets that are big enough. */
5863 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
5867 /* absorb RTT data (in milliseconds) for these big packets */
5868 if (peer->smRtt == 0) {
5869 peer->smRtt = xferMs;
5871 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
5876 if (peer->countDown) {
5880 peer->countDown = 10; /* recalculate only every so often */
5882 /* In practice, we can measure only the RTT for full packets,
5883 * because of the way Rx acks the data that it receives. (If it's
5884 * smaller than a full packet, it often gets implicitly acked
5885 * either by the call response (from a server) or by the next call
5886 * (from a client), and either case confuses transmission times
5887 * with processing times.) Therefore, replace the above
5888 * more-sophisticated processing with a simpler version, where the
5889 * smoothed RTT is kept for full-size packets, and the time to
5890 * transmit a windowful of full-size packets is simply RTT *
5891 * windowSize. Again, we take two steps:
5892 - ensure the timeout is large enough for a single packet's RTT;
5893 - ensure that the window is small enough to fit in the desired timeout.*/
5895 /* First, the timeout check. */
5896 minTime = peer->smRtt;
5897 /* Get a reasonable estimate for a timeout period */
5899 newTO.sec = minTime / 1000;
5900 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
5902 /* Increase the timeout period so that we can always do at least
5903 * one packet exchange */
5904 if (clock_Gt(&newTO, &peer->timeout)) {
5906 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));
5908 peer->timeout = newTO;
5911 /* Now, get an estimate for the transmit window size. */
5912 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
5913 /* Now, convert to the number of full packets that could fit in a
5914 * reasonable fraction of that interval */
5915 minTime /= (peer->smRtt << 1);
5916 xferSize = minTime; /* (make a copy) */
5918 /* Now clamp the size to reasonable bounds. */
5921 else if (minTime > rx_Window)
5922 minTime = rx_Window;
5923 /* if (minTime != peer->maxWindow) {
5924 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
5925 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
5926 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
5928 peer->maxWindow = minTime;
5929 elide... call->twind = minTime;
5933 /* Cut back on the peer timeout if it had earlier grown unreasonably.
5934 * Discern this by calculating the timeout necessary for rx_Window
5936 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
5937 /* calculate estimate for transmission interval in milliseconds */
5938 minTime = rx_Window * peer->smRtt;
5939 if (minTime < 1000) {
5940 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
5941 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
5942 peer->timeout.usec, peer->smRtt, peer->packetSize));
5944 newTO.sec = 0; /* cut back on timeout by half a second */
5945 newTO.usec = 500000;
5946 clock_Sub(&peer->timeout, &newTO);
5951 } /* end of rxi_ComputeRate */
5952 #endif /* ADAPT_WINDOW */
5960 /* Don't call this debugging routine directly; use dpf */
5962 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
5963 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
5967 clock_GetTime(&now);
5968 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
5969 (unsigned int)now.usec / 1000);
5970 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
5978 * This function is used to process the rx_stats structure that is local
5979 * to a process as well as an rx_stats structure received from a remote
5980 * process (via rxdebug). Therefore, it needs to do minimal version
5984 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
5985 afs_int32 freePackets, char version)
5989 if (size != sizeof(struct rx_stats)) {
5991 "Unexpected size of stats structure: was %d, expected %d\n",
5992 size, sizeof(struct rx_stats));
5995 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
5998 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
5999 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6000 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6001 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6002 s->specialPktAllocFailures);
6004 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6005 s->receivePktAllocFailures, s->sendPktAllocFailures,
6006 s->specialPktAllocFailures);
6010 " greedy %d, " "bogusReads %d (last from host %x), "
6011 "noPackets %d, " "noBuffers %d, " "selects %d, "
6012 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6013 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6014 s->selects, s->sendSelects);
6016 fprintf(file, " packets read: ");
6017 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6018 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6020 fprintf(file, "\n");
6023 " other read counters: data %d, " "ack %d, " "dup %d "
6024 "spurious %d " "dally %d\n", s->dataPacketsRead,
6025 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6026 s->ignorePacketDally);
6028 fprintf(file, " packets sent: ");
6029 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6030 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6032 fprintf(file, "\n");
6035 " other send counters: ack %d, " "data %d (not resends), "
6036 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6037 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6038 s->dataPacketsPushed, s->ignoreAckedPacket);
6041 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6042 s->netSendFailures, (int)s->fatalErrors);
6044 if (s->nRttSamples) {
6045 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6046 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6048 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6049 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6053 " %d server connections, " "%d client connections, "
6054 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6055 s->nServerConns, s->nClientConns, s->nPeerStructs,
6056 s->nCallStructs, s->nFreeCallStructs);
6058 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6059 fprintf(file, " %d clock updates\n", clock_nUpdates);
6064 /* for backward compatibility */
6066 rx_PrintStats(FILE * file)
6068 MUTEX_ENTER(&rx_stats_mutex);
6069 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6071 MUTEX_EXIT(&rx_stats_mutex);
6075 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6077 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6078 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6079 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6082 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6083 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6084 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6087 " Packet size %d, " "max in packet skew %d, "
6088 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6089 (int)peer->outPacketSkew);
6092 #ifdef AFS_PTHREAD_ENV
6094 * This mutex protects the following static variables:
6098 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6099 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6101 #define LOCK_RX_DEBUG
6102 #define UNLOCK_RX_DEBUG
6103 #endif /* AFS_PTHREAD_ENV */
6106 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6107 u_char type, void *inputData, size_t inputLength,
6108 void *outputData, size_t outputLength)
6110 static afs_int32 counter = 100;
6112 struct rx_header theader;
6114 register afs_int32 code;
6116 struct sockaddr_in taddr, faddr;
6121 endTime = time(0) + 20; /* try for 20 seconds */
6125 tp = &tbuffer[sizeof(struct rx_header)];
6126 taddr.sin_family = AF_INET;
6127 taddr.sin_port = remotePort;
6128 taddr.sin_addr.s_addr = remoteAddr;
6129 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6130 taddr.sin_len = sizeof(struct sockaddr_in);
6133 memset(&theader, 0, sizeof(theader));
6134 theader.epoch = htonl(999);
6136 theader.callNumber = htonl(counter);
6139 theader.type = type;
6140 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6141 theader.serviceId = 0;
6143 memcpy(tbuffer, &theader, sizeof(theader));
6144 memcpy(tp, inputData, inputLength);
6146 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6147 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6149 /* see if there's a packet available */
6151 FD_SET(socket, &imask);
6154 code = select(socket + 1, &imask, 0, 0, &tv);
6155 if (code == 1 && FD_ISSET(socket, &imask)) {
6156 /* now receive a packet */
6157 faddrLen = sizeof(struct sockaddr_in);
6159 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6160 (struct sockaddr *)&faddr, &faddrLen);
6163 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6164 if (counter == ntohl(theader.callNumber))
6169 /* see if we've timed out */
6170 if (endTime < time(0))
6173 code -= sizeof(struct rx_header);
6174 if (code > outputLength)
6175 code = outputLength;
6176 memcpy(outputData, tp, code);
6181 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6182 afs_uint16 remotePort, struct rx_debugStats * stat,
6183 afs_uint32 * supportedValues)
6185 struct rx_debugIn in;
6188 *supportedValues = 0;
6189 in.type = htonl(RX_DEBUGI_GETSTATS);
6192 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6193 &in, sizeof(in), stat, sizeof(*stat));
6196 * If the call was successful, fixup the version and indicate
6197 * what contents of the stat structure are valid.
6198 * Also do net to host conversion of fields here.
6202 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6203 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6205 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6206 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6208 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6209 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6211 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6212 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6214 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6215 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6217 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6218 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6220 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6221 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6223 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6224 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6227 stat->nFreePackets = ntohl(stat->nFreePackets);
6228 stat->packetReclaims = ntohl(stat->packetReclaims);
6229 stat->callsExecuted = ntohl(stat->callsExecuted);
6230 stat->nWaiting = ntohl(stat->nWaiting);
6231 stat->idleThreads = ntohl(stat->idleThreads);
6238 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6239 afs_uint16 remotePort, struct rx_stats * stat,
6240 afs_uint32 * supportedValues)
6242 struct rx_debugIn in;
6243 afs_int32 *lp = (afs_int32 *) stat;
6248 * supportedValues is currently unused, but added to allow future
6249 * versioning of this function.
6252 *supportedValues = 0;
6253 in.type = htonl(RX_DEBUGI_RXSTATS);
6255 memset(stat, 0, sizeof(*stat));
6257 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6258 &in, sizeof(in), stat, sizeof(*stat));
6263 * Do net to host conversion here
6266 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6275 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6276 afs_uint16 remotePort, size_t version_length,
6280 return MakeDebugCall(socket, remoteAddr, remotePort,
6281 RX_PACKET_TYPE_VERSION, a, 1, version,
6286 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6287 afs_uint16 remotePort, afs_int32 * nextConnection,
6288 int allConnections, afs_uint32 debugSupportedValues,
6289 struct rx_debugConn * conn,
6290 afs_uint32 * supportedValues)
6292 struct rx_debugIn in;
6297 * supportedValues is currently unused, but added to allow future
6298 * versioning of this function.
6301 *supportedValues = 0;
6302 if (allConnections) {
6303 in.type = htonl(RX_DEBUGI_GETALLCONN);
6305 in.type = htonl(RX_DEBUGI_GETCONN);
6307 in.index = htonl(*nextConnection);
6308 memset(conn, 0, sizeof(*conn));
6310 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6311 &in, sizeof(in), conn, sizeof(*conn));
6314 *nextConnection += 1;
6317 * Convert old connection format to new structure.
6320 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6321 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6322 #define MOVEvL(a) (conn->a = vL->a)
6324 /* any old or unrecognized version... */
6325 for (i = 0; i < RX_MAXCALLS; i++) {
6326 MOVEvL(callState[i]);
6327 MOVEvL(callMode[i]);
6328 MOVEvL(callFlags[i]);
6329 MOVEvL(callOther[i]);
6331 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6332 MOVEvL(secStats.type);
6333 MOVEvL(secStats.level);
6334 MOVEvL(secStats.flags);
6335 MOVEvL(secStats.expires);
6336 MOVEvL(secStats.packetsReceived);
6337 MOVEvL(secStats.packetsSent);
6338 MOVEvL(secStats.bytesReceived);
6339 MOVEvL(secStats.bytesSent);
6344 * Do net to host conversion here
6346 * I don't convert host or port since we are most likely
6347 * going to want these in NBO.
6349 conn->cid = ntohl(conn->cid);
6350 conn->serial = ntohl(conn->serial);
6351 for (i = 0; i < RX_MAXCALLS; i++) {
6352 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6354 conn->error = ntohl(conn->error);
6355 conn->secStats.flags = ntohl(conn->secStats.flags);
6356 conn->secStats.expires = ntohl(conn->secStats.expires);
6357 conn->secStats.packetsReceived =
6358 ntohl(conn->secStats.packetsReceived);
6359 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6360 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6361 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6362 conn->epoch = ntohl(conn->epoch);
6363 conn->natMTU = ntohl(conn->natMTU);
6370 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6371 afs_uint16 remotePort, afs_int32 * nextPeer,
6372 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6373 afs_uint32 * supportedValues)
6375 struct rx_debugIn in;
6379 * supportedValues is currently unused, but added to allow future
6380 * versioning of this function.
6383 *supportedValues = 0;
6384 in.type = htonl(RX_DEBUGI_GETPEER);
6385 in.index = htonl(*nextPeer);
6386 memset(peer, 0, sizeof(*peer));
6388 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6389 &in, sizeof(in), peer, sizeof(*peer));
6395 * Do net to host conversion here
6397 * I don't convert host or port since we are most likely
6398 * going to want these in NBO.
6400 peer->ifMTU = ntohs(peer->ifMTU);
6401 peer->idleWhen = ntohl(peer->idleWhen);
6402 peer->refCount = ntohs(peer->refCount);
6403 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6404 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6405 peer->rtt = ntohl(peer->rtt);
6406 peer->rtt_dev = ntohl(peer->rtt_dev);
6407 peer->timeout.sec = ntohl(peer->timeout.sec);
6408 peer->timeout.usec = ntohl(peer->timeout.usec);
6409 peer->nSent = ntohl(peer->nSent);
6410 peer->reSends = ntohl(peer->reSends);
6411 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6412 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6413 peer->rateFlag = ntohl(peer->rateFlag);
6414 peer->natMTU = ntohs(peer->natMTU);
6415 peer->maxMTU = ntohs(peer->maxMTU);
6416 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6417 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6418 peer->MTU = ntohs(peer->MTU);
6419 peer->cwind = ntohs(peer->cwind);
6420 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6421 peer->congestSeq = ntohs(peer->congestSeq);
6422 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6423 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6424 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6425 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6430 #endif /* RXDEBUG */
6435 struct rx_serverQueueEntry *np;
6438 register struct rx_call *call;
6439 register struct rx_serverQueueEntry *sq;
6443 if (rxinit_status == 1) {
6445 return; /* Already shutdown. */
6449 #ifndef AFS_PTHREAD_ENV
6450 FD_ZERO(&rx_selectMask);
6451 #endif /* AFS_PTHREAD_ENV */
6452 rxi_dataQuota = RX_MAX_QUOTA;
6453 #ifndef AFS_PTHREAD_ENV
6455 #endif /* AFS_PTHREAD_ENV */
6458 #ifndef AFS_PTHREAD_ENV
6459 #ifndef AFS_USE_GETTIMEOFDAY
6461 #endif /* AFS_USE_GETTIMEOFDAY */
6462 #endif /* AFS_PTHREAD_ENV */
6464 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6465 call = queue_First(&rx_freeCallQueue, rx_call);
6467 rxi_Free(call, sizeof(struct rx_call));
6470 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6471 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6477 struct rx_peer **peer_ptr, **peer_end;
6478 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6479 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6481 struct rx_peer *peer, *next;
6482 for (peer = *peer_ptr; peer; peer = next) {
6483 rx_interface_stat_p rpc_stat, nrpc_stat;
6486 (&peer->rpcStats, rpc_stat, nrpc_stat,
6487 rx_interface_stat)) {
6488 unsigned int num_funcs;
6491 queue_Remove(&rpc_stat->queue_header);
6492 queue_Remove(&rpc_stat->all_peers);
6493 num_funcs = rpc_stat->stats[0].func_total;
6495 sizeof(rx_interface_stat_t) +
6496 rpc_stat->stats[0].func_total *
6497 sizeof(rx_function_entry_v1_t);
6499 rxi_Free(rpc_stat, space);
6500 MUTEX_ENTER(&rx_rpc_stats);
6501 rxi_rpc_peer_stat_cnt -= num_funcs;
6502 MUTEX_EXIT(&rx_rpc_stats);
6506 MUTEX_ENTER(&rx_stats_mutex);
6507 rx_stats.nPeerStructs--;
6508 MUTEX_EXIT(&rx_stats_mutex);
6512 for (i = 0; i < RX_MAX_SERVICES; i++) {
6514 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6516 for (i = 0; i < rx_hashTableSize; i++) {
6517 register struct rx_connection *tc, *ntc;
6518 MUTEX_ENTER(&rx_connHashTable_lock);
6519 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6521 for (j = 0; j < RX_MAXCALLS; j++) {
6523 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6526 rxi_Free(tc, sizeof(*tc));
6528 MUTEX_EXIT(&rx_connHashTable_lock);
6531 MUTEX_ENTER(&freeSQEList_lock);
6533 while ((np = rx_FreeSQEList)) {
6534 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6535 MUTEX_DESTROY(&np->lock);
6536 rxi_Free(np, sizeof(*np));
6539 MUTEX_EXIT(&freeSQEList_lock);
6540 MUTEX_DESTROY(&freeSQEList_lock);
6541 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6542 MUTEX_DESTROY(&rx_connHashTable_lock);
6543 MUTEX_DESTROY(&rx_peerHashTable_lock);
6544 MUTEX_DESTROY(&rx_serverPool_lock);
6546 osi_Free(rx_connHashTable,
6547 rx_hashTableSize * sizeof(struct rx_connection *));
6548 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6550 UNPIN(rx_connHashTable,
6551 rx_hashTableSize * sizeof(struct rx_connection *));
6552 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6554 rxi_FreeAllPackets();
6556 MUTEX_ENTER(&rx_stats_mutex);
6557 rxi_dataQuota = RX_MAX_QUOTA;
6558 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6559 MUTEX_EXIT(&rx_stats_mutex);
6565 #ifdef RX_ENABLE_LOCKS
6567 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6569 if (!MUTEX_ISMINE(lockaddr))
6570 osi_Panic("Lock not held: %s", msg);
6572 #endif /* RX_ENABLE_LOCKS */
6577 * Routines to implement connection specific data.
6581 rx_KeyCreate(rx_destructor_t rtn)
6584 MUTEX_ENTER(&rxi_keyCreate_lock);
6585 key = rxi_keyCreate_counter++;
6586 rxi_keyCreate_destructor = (rx_destructor_t *)
6587 realloc((void *)rxi_keyCreate_destructor,
6588 (key + 1) * sizeof(rx_destructor_t));
6589 rxi_keyCreate_destructor[key] = rtn;
6590 MUTEX_EXIT(&rxi_keyCreate_lock);
6595 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6598 MUTEX_ENTER(&conn->conn_data_lock);
6599 if (!conn->specific) {
6600 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6601 for (i = 0; i < key; i++)
6602 conn->specific[i] = NULL;
6603 conn->nSpecific = key + 1;
6604 conn->specific[key] = ptr;
6605 } else if (key >= conn->nSpecific) {
6606 conn->specific = (void **)
6607 realloc(conn->specific, (key + 1) * sizeof(void *));
6608 for (i = conn->nSpecific; i < key; i++)
6609 conn->specific[i] = NULL;
6610 conn->nSpecific = key + 1;
6611 conn->specific[key] = ptr;
6613 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6614 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6615 conn->specific[key] = ptr;
6617 MUTEX_EXIT(&conn->conn_data_lock);
6621 rx_GetSpecific(struct rx_connection *conn, int key)
6624 MUTEX_ENTER(&conn->conn_data_lock);
6625 if (key >= conn->nSpecific)
6628 ptr = conn->specific[key];
6629 MUTEX_EXIT(&conn->conn_data_lock);
6633 #endif /* !KERNEL */
6636 * processStats is a queue used to store the statistics for the local
6637 * process. Its contents are similar to the contents of the rpcStats
6638 * queue on a rx_peer structure, but the actual data stored within
6639 * this queue contains totals across the lifetime of the process (assuming
6640 * the stats have not been reset) - unlike the per peer structures
6641 * which can come and go based upon the peer lifetime.
6644 static struct rx_queue processStats = { &processStats, &processStats };
6647 * peerStats is a queue used to store the statistics for all peer structs.
6648 * Its contents are the union of all the peer rpcStats queues.
6651 static struct rx_queue peerStats = { &peerStats, &peerStats };
6654 * rxi_monitor_processStats is used to turn process wide stat collection
6658 static int rxi_monitor_processStats = 0;
6661 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6664 static int rxi_monitor_peerStats = 0;
6667 * rxi_AddRpcStat - given all of the information for a particular rpc
6668 * call, create (if needed) and update the stat totals for the rpc.
6672 * IN stats - the queue of stats that will be updated with the new value
6674 * IN rxInterface - a unique number that identifies the rpc interface
6676 * IN currentFunc - the index of the function being invoked
6678 * IN totalFunc - the total number of functions in this interface
6680 * IN queueTime - the amount of time this function waited for a thread
6682 * IN execTime - the amount of time this function invocation took to execute
6684 * IN bytesSent - the number bytes sent by this invocation
6686 * IN bytesRcvd - the number bytes received by this invocation
6688 * IN isServer - if true, this invocation was made to a server
6690 * IN remoteHost - the ip address of the remote host
6692 * IN remotePort - the port of the remote host
6694 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6696 * INOUT counter - if a new stats structure is allocated, the counter will
6697 * be updated with the new number of allocated stat structures
6705 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6706 afs_uint32 currentFunc, afs_uint32 totalFunc,
6707 struct clock *queueTime, struct clock *execTime,
6708 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6709 afs_uint32 remoteHost, afs_uint32 remotePort,
6710 int addToPeerList, unsigned int *counter)
6713 rx_interface_stat_p rpc_stat, nrpc_stat;
6716 * See if there's already a structure for this interface
6719 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6720 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6721 && (rpc_stat->stats[0].remote_is_server == isServer))
6726 * Didn't find a match so allocate a new structure and add it to the
6730 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6731 || (rpc_stat->stats[0].interfaceId != rxInterface)
6732 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6737 sizeof(rx_interface_stat_t) +
6738 totalFunc * sizeof(rx_function_entry_v1_t);
6740 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6741 if (rpc_stat == NULL) {
6745 *counter += totalFunc;
6746 for (i = 0; i < totalFunc; i++) {
6747 rpc_stat->stats[i].remote_peer = remoteHost;
6748 rpc_stat->stats[i].remote_port = remotePort;
6749 rpc_stat->stats[i].remote_is_server = isServer;
6750 rpc_stat->stats[i].interfaceId = rxInterface;
6751 rpc_stat->stats[i].func_total = totalFunc;
6752 rpc_stat->stats[i].func_index = i;
6753 hzero(rpc_stat->stats[i].invocations);
6754 hzero(rpc_stat->stats[i].bytes_sent);
6755 hzero(rpc_stat->stats[i].bytes_rcvd);
6756 rpc_stat->stats[i].queue_time_sum.sec = 0;
6757 rpc_stat->stats[i].queue_time_sum.usec = 0;
6758 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6759 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6760 rpc_stat->stats[i].queue_time_min.sec = 9999999;
6761 rpc_stat->stats[i].queue_time_min.usec = 9999999;
6762 rpc_stat->stats[i].queue_time_max.sec = 0;
6763 rpc_stat->stats[i].queue_time_max.usec = 0;
6764 rpc_stat->stats[i].execution_time_sum.sec = 0;
6765 rpc_stat->stats[i].execution_time_sum.usec = 0;
6766 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
6767 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
6768 rpc_stat->stats[i].execution_time_min.sec = 9999999;
6769 rpc_stat->stats[i].execution_time_min.usec = 9999999;
6770 rpc_stat->stats[i].execution_time_max.sec = 0;
6771 rpc_stat->stats[i].execution_time_max.usec = 0;
6773 queue_Prepend(stats, rpc_stat);
6774 if (addToPeerList) {
6775 queue_Prepend(&peerStats, &rpc_stat->all_peers);
6780 * Increment the stats for this function
6783 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
6784 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
6785 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
6786 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
6787 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
6788 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
6789 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
6791 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
6792 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
6794 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
6795 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
6797 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
6798 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
6800 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
6801 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
6809 * rx_IncrementTimeAndCount - increment the times and count for a particular
6814 * IN peer - the peer who invoked the rpc
6816 * IN rxInterface - a unique number that identifies the rpc interface
6818 * IN currentFunc - the index of the function being invoked
6820 * IN totalFunc - the total number of functions in this interface
6822 * IN queueTime - the amount of time this function waited for a thread
6824 * IN execTime - the amount of time this function invocation took to execute
6826 * IN bytesSent - the number bytes sent by this invocation
6828 * IN bytesRcvd - the number bytes received by this invocation
6830 * IN isServer - if true, this invocation was made to a server
6838 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
6839 afs_uint32 currentFunc, afs_uint32 totalFunc,
6840 struct clock *queueTime, struct clock *execTime,
6841 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
6845 MUTEX_ENTER(&rx_rpc_stats);
6846 MUTEX_ENTER(&peer->peer_lock);
6848 if (rxi_monitor_peerStats) {
6849 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
6850 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6851 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
6854 if (rxi_monitor_processStats) {
6855 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
6856 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6857 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
6860 MUTEX_EXIT(&peer->peer_lock);
6861 MUTEX_EXIT(&rx_rpc_stats);
6866 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
6870 * IN callerVersion - the rpc stat version of the caller.
6872 * IN count - the number of entries to marshall.
6874 * IN stats - pointer to stats to be marshalled.
6876 * OUT ptr - Where to store the marshalled data.
6883 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
6884 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
6890 * We only support the first version
6892 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
6893 *(ptr++) = stats->remote_peer;
6894 *(ptr++) = stats->remote_port;
6895 *(ptr++) = stats->remote_is_server;
6896 *(ptr++) = stats->interfaceId;
6897 *(ptr++) = stats->func_total;
6898 *(ptr++) = stats->func_index;
6899 *(ptr++) = hgethi(stats->invocations);
6900 *(ptr++) = hgetlo(stats->invocations);
6901 *(ptr++) = hgethi(stats->bytes_sent);
6902 *(ptr++) = hgetlo(stats->bytes_sent);
6903 *(ptr++) = hgethi(stats->bytes_rcvd);
6904 *(ptr++) = hgetlo(stats->bytes_rcvd);
6905 *(ptr++) = stats->queue_time_sum.sec;
6906 *(ptr++) = stats->queue_time_sum.usec;
6907 *(ptr++) = stats->queue_time_sum_sqr.sec;
6908 *(ptr++) = stats->queue_time_sum_sqr.usec;
6909 *(ptr++) = stats->queue_time_min.sec;
6910 *(ptr++) = stats->queue_time_min.usec;
6911 *(ptr++) = stats->queue_time_max.sec;
6912 *(ptr++) = stats->queue_time_max.usec;
6913 *(ptr++) = stats->execution_time_sum.sec;
6914 *(ptr++) = stats->execution_time_sum.usec;
6915 *(ptr++) = stats->execution_time_sum_sqr.sec;
6916 *(ptr++) = stats->execution_time_sum_sqr.usec;
6917 *(ptr++) = stats->execution_time_min.sec;
6918 *(ptr++) = stats->execution_time_min.usec;
6919 *(ptr++) = stats->execution_time_max.sec;
6920 *(ptr++) = stats->execution_time_max.usec;
6926 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
6931 * IN callerVersion - the rpc stat version of the caller
6933 * OUT myVersion - the rpc stat version of this function
6935 * OUT clock_sec - local time seconds
6937 * OUT clock_usec - local time microseconds
6939 * OUT allocSize - the number of bytes allocated to contain stats
6941 * OUT statCount - the number stats retrieved from this process.
6943 * OUT stats - the actual stats retrieved from this process.
6947 * Returns void. If successful, stats will != NULL.
6951 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
6952 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
6953 size_t * allocSize, afs_uint32 * statCount,
6954 afs_uint32 ** stats)
6964 *myVersion = RX_STATS_RETRIEVAL_VERSION;
6967 * Check to see if stats are enabled
6970 MUTEX_ENTER(&rx_rpc_stats);
6971 if (!rxi_monitor_processStats) {
6972 MUTEX_EXIT(&rx_rpc_stats);
6976 clock_GetTime(&now);
6977 *clock_sec = now.sec;
6978 *clock_usec = now.usec;
6981 * Allocate the space based upon the caller version
6983 * If the client is at an older version than we are,
6984 * we return the statistic data in the older data format, but
6985 * we still return our version number so the client knows we
6986 * are maintaining more data than it can retrieve.
6989 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
6990 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
6991 *statCount = rxi_rpc_process_stat_cnt;
6994 * This can't happen yet, but in the future version changes
6995 * can be handled by adding additional code here
6999 if (space > (size_t) 0) {
7001 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7004 rx_interface_stat_p rpc_stat, nrpc_stat;
7008 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7010 * Copy the data based upon the caller version
7012 rx_MarshallProcessRPCStats(callerVersion,
7013 rpc_stat->stats[0].func_total,
7014 rpc_stat->stats, &ptr);
7020 MUTEX_EXIT(&rx_rpc_stats);
7025 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7029 * IN callerVersion - the rpc stat version of the caller
7031 * OUT myVersion - the rpc stat version of this function
7033 * OUT clock_sec - local time seconds
7035 * OUT clock_usec - local time microseconds
7037 * OUT allocSize - the number of bytes allocated to contain stats
7039 * OUT statCount - the number of stats retrieved from the individual
7042 * OUT stats - the actual stats retrieved from the individual peer structures.
7046 * Returns void. If successful, stats will != NULL.
7050 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7051 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7052 size_t * allocSize, afs_uint32 * statCount,
7053 afs_uint32 ** stats)
7063 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7066 * Check to see if stats are enabled
7069 MUTEX_ENTER(&rx_rpc_stats);
7070 if (!rxi_monitor_peerStats) {
7071 MUTEX_EXIT(&rx_rpc_stats);
7075 clock_GetTime(&now);
7076 *clock_sec = now.sec;
7077 *clock_usec = now.usec;
7080 * Allocate the space based upon the caller version
7082 * If the client is at an older version than we are,
7083 * we return the statistic data in the older data format, but
7084 * we still return our version number so the client knows we
7085 * are maintaining more data than it can retrieve.
7088 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7089 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7090 *statCount = rxi_rpc_peer_stat_cnt;
7093 * This can't happen yet, but in the future version changes
7094 * can be handled by adding additional code here
7098 if (space > (size_t) 0) {
7100 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7103 rx_interface_stat_p rpc_stat, nrpc_stat;
7107 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7109 * We have to fix the offset of rpc_stat since we are
7110 * keeping this structure on two rx_queues. The rx_queue
7111 * package assumes that the rx_queue member is the first
7112 * member of the structure. That is, rx_queue assumes that
7113 * any one item is only on one queue at a time. We are
7114 * breaking that assumption and so we have to do a little
7115 * math to fix our pointers.
7118 fix_offset = (char *)rpc_stat;
7119 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7120 rpc_stat = (rx_interface_stat_p) fix_offset;
7123 * Copy the data based upon the caller version
7125 rx_MarshallProcessRPCStats(callerVersion,
7126 rpc_stat->stats[0].func_total,
7127 rpc_stat->stats, &ptr);
7133 MUTEX_EXIT(&rx_rpc_stats);
7138 * rx_FreeRPCStats - free memory allocated by
7139 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7143 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7144 * rx_RetrievePeerRPCStats
7146 * IN allocSize - the number of bytes in stats.
7154 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7156 rxi_Free(stats, allocSize);
7160 * rx_queryProcessRPCStats - see if process rpc stat collection is
7161 * currently enabled.
7167 * Returns 0 if stats are not enabled != 0 otherwise
7171 rx_queryProcessRPCStats(void)
7174 MUTEX_ENTER(&rx_rpc_stats);
7175 rc = rxi_monitor_processStats;
7176 MUTEX_EXIT(&rx_rpc_stats);
7181 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7187 * Returns 0 if stats are not enabled != 0 otherwise
7191 rx_queryPeerRPCStats(void)
7194 MUTEX_ENTER(&rx_rpc_stats);
7195 rc = rxi_monitor_peerStats;
7196 MUTEX_EXIT(&rx_rpc_stats);
7201 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7211 rx_enableProcessRPCStats(void)
7213 MUTEX_ENTER(&rx_rpc_stats);
7214 rx_enable_stats = 1;
7215 rxi_monitor_processStats = 1;
7216 MUTEX_EXIT(&rx_rpc_stats);
7220 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7230 rx_enablePeerRPCStats(void)
7232 MUTEX_ENTER(&rx_rpc_stats);
7233 rx_enable_stats = 1;
7234 rxi_monitor_peerStats = 1;
7235 MUTEX_EXIT(&rx_rpc_stats);
7239 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7249 rx_disableProcessRPCStats(void)
7251 rx_interface_stat_p rpc_stat, nrpc_stat;
7254 MUTEX_ENTER(&rx_rpc_stats);
7257 * Turn off process statistics and if peer stats is also off, turn
7261 rxi_monitor_processStats = 0;
7262 if (rxi_monitor_peerStats == 0) {
7263 rx_enable_stats = 0;
7266 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7267 unsigned int num_funcs = 0;
7270 queue_Remove(rpc_stat);
7271 num_funcs = rpc_stat->stats[0].func_total;
7273 sizeof(rx_interface_stat_t) +
7274 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7276 rxi_Free(rpc_stat, space);
7277 rxi_rpc_process_stat_cnt -= num_funcs;
7279 MUTEX_EXIT(&rx_rpc_stats);
7283 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7293 rx_disablePeerRPCStats(void)
7295 struct rx_peer **peer_ptr, **peer_end;
7298 MUTEX_ENTER(&rx_rpc_stats);
7301 * Turn off peer statistics and if process stats is also off, turn
7305 rxi_monitor_peerStats = 0;
7306 if (rxi_monitor_processStats == 0) {
7307 rx_enable_stats = 0;
7310 MUTEX_ENTER(&rx_peerHashTable_lock);
7311 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7312 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7314 struct rx_peer *peer, *next, *prev;
7315 for (prev = peer = *peer_ptr; peer; peer = next) {
7317 code = MUTEX_TRYENTER(&peer->peer_lock);
7319 rx_interface_stat_p rpc_stat, nrpc_stat;
7322 (&peer->rpcStats, rpc_stat, nrpc_stat,
7323 rx_interface_stat)) {
7324 unsigned int num_funcs = 0;
7327 queue_Remove(&rpc_stat->queue_header);
7328 queue_Remove(&rpc_stat->all_peers);
7329 num_funcs = rpc_stat->stats[0].func_total;
7331 sizeof(rx_interface_stat_t) +
7332 rpc_stat->stats[0].func_total *
7333 sizeof(rx_function_entry_v1_t);
7335 rxi_Free(rpc_stat, space);
7336 rxi_rpc_peer_stat_cnt -= num_funcs;
7338 MUTEX_EXIT(&peer->peer_lock);
7339 if (prev == *peer_ptr) {
7349 MUTEX_EXIT(&rx_peerHashTable_lock);
7350 MUTEX_EXIT(&rx_rpc_stats);
7354 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7359 * IN clearFlag - flag indicating which stats to clear
7367 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7369 rx_interface_stat_p rpc_stat, nrpc_stat;
7371 MUTEX_ENTER(&rx_rpc_stats);
7373 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7374 unsigned int num_funcs = 0, i;
7375 num_funcs = rpc_stat->stats[0].func_total;
7376 for (i = 0; i < num_funcs; i++) {
7377 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7378 hzero(rpc_stat->stats[i].invocations);
7380 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7381 hzero(rpc_stat->stats[i].bytes_sent);
7383 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7384 hzero(rpc_stat->stats[i].bytes_rcvd);
7386 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7387 rpc_stat->stats[i].queue_time_sum.sec = 0;
7388 rpc_stat->stats[i].queue_time_sum.usec = 0;
7390 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7391 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7392 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7394 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7395 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7396 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7398 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7399 rpc_stat->stats[i].queue_time_max.sec = 0;
7400 rpc_stat->stats[i].queue_time_max.usec = 0;
7402 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7403 rpc_stat->stats[i].execution_time_sum.sec = 0;
7404 rpc_stat->stats[i].execution_time_sum.usec = 0;
7406 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7407 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7408 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7410 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7411 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7412 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7414 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7415 rpc_stat->stats[i].execution_time_max.sec = 0;
7416 rpc_stat->stats[i].execution_time_max.usec = 0;
7421 MUTEX_EXIT(&rx_rpc_stats);
7425 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7430 * IN clearFlag - flag indicating which stats to clear
7438 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7440 rx_interface_stat_p rpc_stat, nrpc_stat;
7442 MUTEX_ENTER(&rx_rpc_stats);
7444 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7445 unsigned int num_funcs = 0, i;
7448 * We have to fix the offset of rpc_stat since we are
7449 * keeping this structure on two rx_queues. The rx_queue
7450 * package assumes that the rx_queue member is the first
7451 * member of the structure. That is, rx_queue assumes that
7452 * any one item is only on one queue at a time. We are
7453 * breaking that assumption and so we have to do a little
7454 * math to fix our pointers.
7457 fix_offset = (char *)rpc_stat;
7458 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7459 rpc_stat = (rx_interface_stat_p) fix_offset;
7461 num_funcs = rpc_stat->stats[0].func_total;
7462 for (i = 0; i < num_funcs; i++) {
7463 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7464 hzero(rpc_stat->stats[i].invocations);
7466 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7467 hzero(rpc_stat->stats[i].bytes_sent);
7469 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7470 hzero(rpc_stat->stats[i].bytes_rcvd);
7472 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7473 rpc_stat->stats[i].queue_time_sum.sec = 0;
7474 rpc_stat->stats[i].queue_time_sum.usec = 0;
7476 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7477 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7478 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7480 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7481 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7482 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7484 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7485 rpc_stat->stats[i].queue_time_max.sec = 0;
7486 rpc_stat->stats[i].queue_time_max.usec = 0;
7488 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7489 rpc_stat->stats[i].execution_time_sum.sec = 0;
7490 rpc_stat->stats[i].execution_time_sum.usec = 0;
7492 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7493 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7494 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7496 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7497 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7498 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7500 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7501 rpc_stat->stats[i].execution_time_max.sec = 0;
7502 rpc_stat->stats[i].execution_time_max.usec = 0;
7507 MUTEX_EXIT(&rx_rpc_stats);
7511 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7512 * is authorized to enable/disable/clear RX statistics.
7514 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7517 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7519 rxi_rxstat_userok = proc;
7523 rx_RxStatUserOk(struct rx_call *call)
7525 if (!rxi_rxstat_userok)
7527 return rxi_rxstat_userok(call);