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);
1019 rx_GetConnection(register struct rx_connection *conn)
1025 MUTEX_ENTER(&conn->conn_data_lock);
1027 MUTEX_EXIT(&conn->conn_data_lock);
1032 /* Start a new rx remote procedure call, on the specified connection.
1033 * If wait is set to 1, wait for a free call channel; otherwise return
1034 * 0. Maxtime gives the maximum number of seconds this call may take,
1035 * after rx_MakeCall returns. After this time interval, a call to any
1036 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1037 * For fine grain locking, we hold the conn_call_lock in order to
1038 * to ensure that we don't get signalle after we found a call in an active
1039 * state and before we go to sleep.
1042 rx_NewCall(register struct rx_connection *conn)
1045 register struct rx_call *call;
1046 struct clock queueTime;
1050 dpf(("rx_MakeCall(conn %x)\n", conn));
1053 clock_GetTime(&queueTime);
1055 MUTEX_ENTER(&conn->conn_call_lock);
1058 * Check if there are others waiting for a new call.
1059 * If so, let them go first to avoid starving them.
1060 * This is a fairly simple scheme, and might not be
1061 * a complete solution for large numbers of waiters.
1063 if (conn->makeCallWaiters) {
1064 #ifdef RX_ENABLE_LOCKS
1065 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1072 for (i = 0; i < RX_MAXCALLS; i++) {
1073 call = conn->call[i];
1075 MUTEX_ENTER(&call->lock);
1076 if (call->state == RX_STATE_DALLY) {
1077 rxi_ResetCall(call, 0);
1078 (*call->callNumber)++;
1081 MUTEX_EXIT(&call->lock);
1083 call = rxi_NewCall(conn, i);
1087 if (i < RX_MAXCALLS) {
1090 MUTEX_ENTER(&conn->conn_data_lock);
1091 conn->flags |= RX_CONN_MAKECALL_WAITING;
1092 MUTEX_EXIT(&conn->conn_data_lock);
1094 conn->makeCallWaiters++;
1095 #ifdef RX_ENABLE_LOCKS
1096 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1100 conn->makeCallWaiters--;
1103 * Wake up anyone else who might be giving us a chance to
1104 * run (see code above that avoids resource starvation).
1106 #ifdef RX_ENABLE_LOCKS
1107 CV_BROADCAST(&conn->conn_call_cv);
1112 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1114 /* Client is initially in send mode */
1115 call->state = RX_STATE_ACTIVE;
1116 call->mode = RX_MODE_SENDING;
1118 /* remember start time for call in case we have hard dead time limit */
1119 call->queueTime = queueTime;
1120 clock_GetTime(&call->startTime);
1121 hzero(call->bytesSent);
1122 hzero(call->bytesRcvd);
1124 /* Turn on busy protocol. */
1125 rxi_KeepAliveOn(call);
1127 MUTEX_EXIT(&call->lock);
1128 MUTEX_EXIT(&conn->conn_call_lock);
1132 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1133 /* Now, if TQ wasn't cleared earlier, do it now. */
1135 MUTEX_ENTER(&call->lock);
1136 while (call->flags & RX_CALL_TQ_BUSY) {
1137 call->flags |= RX_CALL_TQ_WAIT;
1138 #ifdef RX_ENABLE_LOCKS
1139 CV_WAIT(&call->cv_tq, &call->lock);
1140 #else /* RX_ENABLE_LOCKS */
1141 osi_rxSleep(&call->tq);
1142 #endif /* RX_ENABLE_LOCKS */
1144 if (call->flags & RX_CALL_TQ_CLEARME) {
1145 rxi_ClearTransmitQueue(call, 0);
1146 queue_Init(&call->tq);
1148 MUTEX_EXIT(&call->lock);
1150 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1156 rxi_HasActiveCalls(register struct rx_connection *aconn)
1159 register struct rx_call *tcall;
1163 for (i = 0; i < RX_MAXCALLS; i++) {
1164 if ((tcall = aconn->call[i])) {
1165 if ((tcall->state == RX_STATE_ACTIVE)
1166 || (tcall->state == RX_STATE_PRECALL)) {
1177 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1178 register afs_int32 * aint32s)
1181 register struct rx_call *tcall;
1185 for (i = 0; i < RX_MAXCALLS; i++) {
1186 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1187 aint32s[i] = aconn->callNumber[i] + 1;
1189 aint32s[i] = aconn->callNumber[i];
1196 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1197 register afs_int32 * aint32s)
1200 register struct rx_call *tcall;
1204 for (i = 0; i < RX_MAXCALLS; i++) {
1205 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1206 aconn->callNumber[i] = aint32s[i] - 1;
1208 aconn->callNumber[i] = aint32s[i];
1214 /* Advertise a new service. A service is named locally by a UDP port
1215 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1218 char *serviceName; Name for identification purposes (e.g. the
1219 service name might be used for probing for
1222 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1223 struct rx_securityClass **securityObjects, int nSecurityObjects,
1224 afs_int32(*serviceProc) (struct rx_call * acall))
1226 osi_socket socket = OSI_NULLSOCKET;
1227 register struct rx_service *tservice;
1233 if (serviceId == 0) {
1235 "rx_NewService: service id for service %s is not non-zero.\n",
1242 "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",
1250 tservice = rxi_AllocService();
1253 for (i = 0; i < RX_MAX_SERVICES; i++) {
1254 register struct rx_service *service = rx_services[i];
1256 if (port == service->servicePort) {
1257 if (service->serviceId == serviceId) {
1258 /* The identical service has already been
1259 * installed; if the caller was intending to
1260 * change the security classes used by this
1261 * service, he/she loses. */
1263 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1264 serviceName, serviceId, service->serviceName);
1267 rxi_FreeService(tservice);
1270 /* Different service, same port: re-use the socket
1271 * which is bound to the same port */
1272 socket = service->socket;
1275 if (socket == OSI_NULLSOCKET) {
1276 /* If we don't already have a socket (from another
1277 * service on same port) get a new one */
1278 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1279 if (socket == OSI_NULLSOCKET) {
1282 rxi_FreeService(tservice);
1287 service->socket = socket;
1288 service->servicePort = port;
1289 service->serviceId = serviceId;
1290 service->serviceName = serviceName;
1291 service->nSecurityObjects = nSecurityObjects;
1292 service->securityObjects = securityObjects;
1293 service->minProcs = 0;
1294 service->maxProcs = 1;
1295 service->idleDeadTime = 60;
1296 service->connDeadTime = rx_connDeadTime;
1297 service->executeRequestProc = serviceProc;
1298 service->checkReach = 0;
1299 rx_services[i] = service; /* not visible until now */
1307 rxi_FreeService(tservice);
1308 (osi_Msg "rx_NewService: cannot support > %d services\n",
1313 /* Generic request processing loop. This routine should be called
1314 * by the implementation dependent rx_ServerProc. If socketp is
1315 * non-null, it will be set to the file descriptor that this thread
1316 * is now listening on. If socketp is null, this routine will never
1319 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1321 register struct rx_call *call;
1322 register afs_int32 code;
1323 register struct rx_service *tservice = NULL;
1330 call = rx_GetCall(threadID, tservice, socketp);
1331 if (socketp && *socketp != OSI_NULLSOCKET) {
1332 /* We are now a listener thread */
1337 /* if server is restarting( typically smooth shutdown) then do not
1338 * allow any new calls.
1341 if (rx_tranquil && (call != NULL)) {
1346 MUTEX_ENTER(&call->lock);
1348 rxi_CallError(call, RX_RESTARTING);
1349 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1351 MUTEX_EXIT(&call->lock);
1356 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1357 #ifdef RX_ENABLE_LOCKS
1359 #endif /* RX_ENABLE_LOCKS */
1360 afs_termState = AFSOP_STOP_AFS;
1361 afs_osi_Wakeup(&afs_termState);
1362 #ifdef RX_ENABLE_LOCKS
1364 #endif /* RX_ENABLE_LOCKS */
1369 tservice = call->conn->service;
1371 if (tservice->beforeProc)
1372 (*tservice->beforeProc) (call);
1374 code = call->conn->service->executeRequestProc(call);
1376 if (tservice->afterProc)
1377 (*tservice->afterProc) (call, code);
1379 rx_EndCall(call, code);
1380 MUTEX_ENTER(&rx_stats_mutex);
1382 MUTEX_EXIT(&rx_stats_mutex);
1388 rx_WakeupServerProcs(void)
1390 struct rx_serverQueueEntry *np, *tqp;
1395 MUTEX_ENTER(&rx_serverPool_lock);
1397 #ifdef RX_ENABLE_LOCKS
1398 if (rx_waitForPacket)
1399 CV_BROADCAST(&rx_waitForPacket->cv);
1400 #else /* RX_ENABLE_LOCKS */
1401 if (rx_waitForPacket)
1402 osi_rxWakeup(rx_waitForPacket);
1403 #endif /* RX_ENABLE_LOCKS */
1404 MUTEX_ENTER(&freeSQEList_lock);
1405 for (np = rx_FreeSQEList; np; np = tqp) {
1406 tqp = *(struct rx_serverQueueEntry **)np;
1407 #ifdef RX_ENABLE_LOCKS
1408 CV_BROADCAST(&np->cv);
1409 #else /* RX_ENABLE_LOCKS */
1411 #endif /* RX_ENABLE_LOCKS */
1413 MUTEX_EXIT(&freeSQEList_lock);
1414 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1415 #ifdef RX_ENABLE_LOCKS
1416 CV_BROADCAST(&np->cv);
1417 #else /* RX_ENABLE_LOCKS */
1419 #endif /* RX_ENABLE_LOCKS */
1421 MUTEX_EXIT(&rx_serverPool_lock);
1427 * One thing that seems to happen is that all the server threads get
1428 * tied up on some empty or slow call, and then a whole bunch of calls
1429 * arrive at once, using up the packet pool, so now there are more
1430 * empty calls. The most critical resources here are server threads
1431 * and the free packet pool. The "doreclaim" code seems to help in
1432 * general. I think that eventually we arrive in this state: there
1433 * are lots of pending calls which do have all their packets present,
1434 * so they won't be reclaimed, are multi-packet calls, so they won't
1435 * be scheduled until later, and thus are tying up most of the free
1436 * packet pool for a very long time.
1438 * 1. schedule multi-packet calls if all the packets are present.
1439 * Probably CPU-bound operation, useful to return packets to pool.
1440 * Do what if there is a full window, but the last packet isn't here?
1441 * 3. preserve one thread which *only* runs "best" calls, otherwise
1442 * it sleeps and waits for that type of call.
1443 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1444 * the current dataquota business is badly broken. The quota isn't adjusted
1445 * to reflect how many packets are presently queued for a running call.
1446 * So, when we schedule a queued call with a full window of packets queued
1447 * up for it, that *should* free up a window full of packets for other 2d-class
1448 * calls to be able to use from the packet pool. But it doesn't.
1450 * NB. Most of the time, this code doesn't run -- since idle server threads
1451 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1452 * as a new call arrives.
1454 /* Sleep until a call arrives. Returns a pointer to the call, ready
1455 * for an rx_Read. */
1456 #ifdef RX_ENABLE_LOCKS
1458 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1460 struct rx_serverQueueEntry *sq;
1461 register struct rx_call *call = (struct rx_call *)0;
1462 struct rx_service *service = NULL;
1465 MUTEX_ENTER(&freeSQEList_lock);
1467 if ((sq = rx_FreeSQEList)) {
1468 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1469 MUTEX_EXIT(&freeSQEList_lock);
1470 } else { /* otherwise allocate a new one and return that */
1471 MUTEX_EXIT(&freeSQEList_lock);
1472 sq = (struct rx_serverQueueEntry *)
1473 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1474 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1475 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1478 MUTEX_ENTER(&rx_serverPool_lock);
1479 if (cur_service != NULL) {
1480 ReturnToServerPool(cur_service);
1483 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1484 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1486 /* Scan for eligible incoming calls. A call is not eligible
1487 * if the maximum number of calls for its service type are
1488 * already executing */
1489 /* One thread will process calls FCFS (to prevent starvation),
1490 * while the other threads may run ahead looking for calls which
1491 * have all their input data available immediately. This helps
1492 * keep threads from blocking, waiting for data from the client. */
1493 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1494 service = tcall->conn->service;
1495 if (!QuotaOK(service)) {
1498 if (tno == rxi_fcfs_thread_num
1499 || !tcall->queue_item_header.next) {
1500 /* If we're the fcfs thread , then we'll just use
1501 * this call. If we haven't been able to find an optimal
1502 * choice, and we're at the end of the list, then use a
1503 * 2d choice if one has been identified. Otherwise... */
1504 call = (choice2 ? choice2 : tcall);
1505 service = call->conn->service;
1506 } else if (!queue_IsEmpty(&tcall->rq)) {
1507 struct rx_packet *rp;
1508 rp = queue_First(&tcall->rq, rx_packet);
1509 if (rp->header.seq == 1) {
1511 || (rp->header.flags & RX_LAST_PACKET)) {
1513 } else if (rxi_2dchoice && !choice2
1514 && !(tcall->flags & RX_CALL_CLEARED)
1515 && (tcall->rprev > rxi_HardAckRate)) {
1524 ReturnToServerPool(service);
1531 MUTEX_EXIT(&rx_serverPool_lock);
1532 MUTEX_ENTER(&call->lock);
1534 if (call->flags & RX_CALL_WAIT_PROC) {
1535 call->flags &= ~RX_CALL_WAIT_PROC;
1536 MUTEX_ENTER(&rx_stats_mutex);
1538 MUTEX_EXIT(&rx_stats_mutex);
1541 if (call->state != RX_STATE_PRECALL || call->error) {
1542 MUTEX_EXIT(&call->lock);
1543 MUTEX_ENTER(&rx_serverPool_lock);
1544 ReturnToServerPool(service);
1549 if (queue_IsEmpty(&call->rq)
1550 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1551 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1553 CLEAR_CALL_QUEUE_LOCK(call);
1556 /* If there are no eligible incoming calls, add this process
1557 * to the idle server queue, to wait for one */
1561 *socketp = OSI_NULLSOCKET;
1563 sq->socketp = socketp;
1564 queue_Append(&rx_idleServerQueue, sq);
1565 #ifndef AFS_AIX41_ENV
1566 rx_waitForPacket = sq;
1568 rx_waitingForPacket = sq;
1569 #endif /* AFS_AIX41_ENV */
1571 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1573 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1574 MUTEX_EXIT(&rx_serverPool_lock);
1575 return (struct rx_call *)0;
1578 } while (!(call = sq->newcall)
1579 && !(socketp && *socketp != OSI_NULLSOCKET));
1580 MUTEX_EXIT(&rx_serverPool_lock);
1582 MUTEX_ENTER(&call->lock);
1588 MUTEX_ENTER(&freeSQEList_lock);
1589 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1590 rx_FreeSQEList = sq;
1591 MUTEX_EXIT(&freeSQEList_lock);
1594 clock_GetTime(&call->startTime);
1595 call->state = RX_STATE_ACTIVE;
1596 call->mode = RX_MODE_RECEIVING;
1597 #ifdef RX_KERNEL_TRACE
1598 if (ICL_SETACTIVE(afs_iclSetp)) {
1599 int glockOwner = ISAFS_GLOCK();
1602 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1603 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1610 rxi_calltrace(RX_CALL_START, call);
1611 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1612 call->conn->service->servicePort, call->conn->service->serviceId,
1615 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1616 MUTEX_EXIT(&call->lock);
1618 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1623 #else /* RX_ENABLE_LOCKS */
1625 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1627 struct rx_serverQueueEntry *sq;
1628 register struct rx_call *call = (struct rx_call *)0, *choice2;
1629 struct rx_service *service = NULL;
1634 MUTEX_ENTER(&freeSQEList_lock);
1636 if ((sq = rx_FreeSQEList)) {
1637 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1638 MUTEX_EXIT(&freeSQEList_lock);
1639 } else { /* otherwise allocate a new one and return that */
1640 MUTEX_EXIT(&freeSQEList_lock);
1641 sq = (struct rx_serverQueueEntry *)
1642 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1643 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1644 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1646 MUTEX_ENTER(&sq->lock);
1648 if (cur_service != NULL) {
1649 cur_service->nRequestsRunning--;
1650 if (cur_service->nRequestsRunning < cur_service->minProcs)
1654 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1655 register struct rx_call *tcall, *ncall;
1656 /* Scan for eligible incoming calls. A call is not eligible
1657 * if the maximum number of calls for its service type are
1658 * already executing */
1659 /* One thread will process calls FCFS (to prevent starvation),
1660 * while the other threads may run ahead looking for calls which
1661 * have all their input data available immediately. This helps
1662 * keep threads from blocking, waiting for data from the client. */
1663 choice2 = (struct rx_call *)0;
1664 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1665 service = tcall->conn->service;
1666 if (QuotaOK(service)) {
1667 if (tno == rxi_fcfs_thread_num
1668 || !tcall->queue_item_header.next) {
1669 /* If we're the fcfs thread, then we'll just use
1670 * this call. If we haven't been able to find an optimal
1671 * choice, and we're at the end of the list, then use a
1672 * 2d choice if one has been identified. Otherwise... */
1673 call = (choice2 ? choice2 : tcall);
1674 service = call->conn->service;
1675 } else if (!queue_IsEmpty(&tcall->rq)) {
1676 struct rx_packet *rp;
1677 rp = queue_First(&tcall->rq, rx_packet);
1678 if (rp->header.seq == 1
1680 || (rp->header.flags & RX_LAST_PACKET))) {
1682 } else if (rxi_2dchoice && !choice2
1683 && !(tcall->flags & RX_CALL_CLEARED)
1684 && (tcall->rprev > rxi_HardAckRate)) {
1697 /* we can't schedule a call if there's no data!!! */
1698 /* send an ack if there's no data, if we're missing the
1699 * first packet, or we're missing something between first
1700 * and last -- there's a "hole" in the incoming data. */
1701 if (queue_IsEmpty(&call->rq)
1702 || queue_First(&call->rq, rx_packet)->header.seq != 1
1703 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1704 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1706 call->flags &= (~RX_CALL_WAIT_PROC);
1707 service->nRequestsRunning++;
1708 /* just started call in minProcs pool, need fewer to maintain
1710 if (service->nRequestsRunning <= service->minProcs)
1714 /* MUTEX_EXIT(&call->lock); */
1716 /* If there are no eligible incoming calls, add this process
1717 * to the idle server queue, to wait for one */
1720 *socketp = OSI_NULLSOCKET;
1722 sq->socketp = socketp;
1723 queue_Append(&rx_idleServerQueue, sq);
1727 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1730 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1731 return (struct rx_call *)0;
1734 } while (!(call = sq->newcall)
1735 && !(socketp && *socketp != OSI_NULLSOCKET));
1737 MUTEX_EXIT(&sq->lock);
1739 MUTEX_ENTER(&freeSQEList_lock);
1740 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1741 rx_FreeSQEList = sq;
1742 MUTEX_EXIT(&freeSQEList_lock);
1745 clock_GetTime(&call->startTime);
1746 call->state = RX_STATE_ACTIVE;
1747 call->mode = RX_MODE_RECEIVING;
1748 #ifdef RX_KERNEL_TRACE
1749 if (ICL_SETACTIVE(afs_iclSetp)) {
1750 int glockOwner = ISAFS_GLOCK();
1753 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1754 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1761 rxi_calltrace(RX_CALL_START, call);
1762 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1763 call->conn->service->servicePort, call->conn->service->serviceId,
1766 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1774 #endif /* RX_ENABLE_LOCKS */
1778 /* Establish a procedure to be called when a packet arrives for a
1779 * call. This routine will be called at most once after each call,
1780 * and will also be called if there is an error condition on the or
1781 * the call is complete. Used by multi rx to build a selection
1782 * function which determines which of several calls is likely to be a
1783 * good one to read from.
1784 * NOTE: the way this is currently implemented it is probably only a
1785 * good idea to (1) use it immediately after a newcall (clients only)
1786 * and (2) only use it once. Other uses currently void your warranty
1789 rx_SetArrivalProc(register struct rx_call *call,
1790 register VOID(*proc) (register struct rx_call * call,
1791 register struct multi_handle * mh,
1792 register int index),
1793 register VOID * handle, register VOID * arg)
1795 call->arrivalProc = proc;
1796 call->arrivalProcHandle = handle;
1797 call->arrivalProcArg = arg;
1800 /* Call is finished (possibly prematurely). Return rc to the peer, if
1801 * appropriate, and return the final error code from the conversation
1805 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1807 register struct rx_connection *conn = call->conn;
1808 register struct rx_service *service;
1809 register struct rx_packet *tp; /* Temporary packet pointer */
1810 register struct rx_packet *nxp; /* Next packet pointer, for queue_Scan */
1814 dpf(("rx_EndCall(call %x)\n", call));
1818 MUTEX_ENTER(&call->lock);
1820 if (rc == 0 && call->error == 0) {
1821 call->abortCode = 0;
1822 call->abortCount = 0;
1825 call->arrivalProc = (VOID(*)())0;
1826 if (rc && call->error == 0) {
1827 rxi_CallError(call, rc);
1828 /* Send an abort message to the peer if this error code has
1829 * only just been set. If it was set previously, assume the
1830 * peer has already been sent the error code or will request it
1832 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1834 if (conn->type == RX_SERVER_CONNECTION) {
1835 /* Make sure reply or at least dummy reply is sent */
1836 if (call->mode == RX_MODE_RECEIVING) {
1837 rxi_WriteProc(call, 0, 0);
1839 if (call->mode == RX_MODE_SENDING) {
1840 rxi_FlushWrite(call);
1842 service = conn->service;
1843 rxi_calltrace(RX_CALL_END, call);
1844 /* Call goes to hold state until reply packets are acknowledged */
1845 if (call->tfirst + call->nSoftAcked < call->tnext) {
1846 call->state = RX_STATE_HOLD;
1848 call->state = RX_STATE_DALLY;
1849 rxi_ClearTransmitQueue(call, 0);
1850 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1851 rxevent_Cancel(call->keepAliveEvent, call,
1852 RX_CALL_REFCOUNT_ALIVE);
1854 } else { /* Client connection */
1856 /* Make sure server receives input packets, in the case where
1857 * no reply arguments are expected */
1858 if ((call->mode == RX_MODE_SENDING)
1859 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1860 (void)rxi_ReadProc(call, &dummy, 1);
1863 /* If we had an outstanding delayed ack, be nice to the server
1864 * and force-send it now.
1866 if (call->delayedAckEvent) {
1867 rxevent_Cancel(call->delayedAckEvent, call,
1868 RX_CALL_REFCOUNT_DELAY);
1869 call->delayedAckEvent = NULL;
1870 rxi_SendDelayedAck(NULL, call, NULL);
1873 /* We need to release the call lock since it's lower than the
1874 * conn_call_lock and we don't want to hold the conn_call_lock
1875 * over the rx_ReadProc call. The conn_call_lock needs to be held
1876 * here for the case where rx_NewCall is perusing the calls on
1877 * the connection structure. We don't want to signal until
1878 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1879 * have checked this call, found it active and by the time it
1880 * goes to sleep, will have missed the signal.
1882 MUTEX_EXIT(&call->lock);
1883 MUTEX_ENTER(&conn->conn_call_lock);
1884 MUTEX_ENTER(&call->lock);
1885 MUTEX_ENTER(&conn->conn_data_lock);
1886 conn->flags |= RX_CONN_BUSY;
1887 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1888 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1889 MUTEX_EXIT(&conn->conn_data_lock);
1890 #ifdef RX_ENABLE_LOCKS
1891 CV_BROADCAST(&conn->conn_call_cv);
1896 #ifdef RX_ENABLE_LOCKS
1898 MUTEX_EXIT(&conn->conn_data_lock);
1900 #endif /* RX_ENABLE_LOCKS */
1901 call->state = RX_STATE_DALLY;
1903 error = call->error;
1905 /* currentPacket, nLeft, and NFree must be zeroed here, because
1906 * ResetCall cannot: ResetCall may be called at splnet(), in the
1907 * kernel version, and may interrupt the macros rx_Read or
1908 * rx_Write, which run at normal priority for efficiency. */
1909 if (call->currentPacket) {
1910 rxi_FreePacket(call->currentPacket);
1911 call->currentPacket = (struct rx_packet *)0;
1912 call->nLeft = call->nFree = call->curlen = 0;
1914 call->nLeft = call->nFree = call->curlen = 0;
1916 /* Free any packets from the last call to ReadvProc/WritevProc */
1917 for (queue_Scan(&call->iovq, tp, nxp, rx_packet)) {
1922 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1923 MUTEX_EXIT(&call->lock);
1924 if (conn->type == RX_CLIENT_CONNECTION) {
1925 MUTEX_EXIT(&conn->conn_call_lock);
1926 conn->flags &= ~RX_CONN_BUSY;
1931 * Map errors to the local host's errno.h format.
1933 error = ntoh_syserr_conv(error);
1937 #if !defined(KERNEL)
1939 /* Call this routine when shutting down a server or client (especially
1940 * clients). This will allow Rx to gracefully garbage collect server
1941 * connections, and reduce the number of retries that a server might
1942 * make to a dead client.
1943 * This is not quite right, since some calls may still be ongoing and
1944 * we can't lock them to destroy them. */
1948 register struct rx_connection **conn_ptr, **conn_end;
1952 if (rxinit_status == 1) {
1954 return; /* Already shutdown. */
1956 rxi_DeleteCachedConnections();
1957 if (rx_connHashTable) {
1958 MUTEX_ENTER(&rx_connHashTable_lock);
1959 for (conn_ptr = &rx_connHashTable[0], conn_end =
1960 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
1962 struct rx_connection *conn, *next;
1963 for (conn = *conn_ptr; conn; conn = next) {
1965 if (conn->type == RX_CLIENT_CONNECTION) {
1966 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
1968 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
1969 #ifdef RX_ENABLE_LOCKS
1970 rxi_DestroyConnectionNoLock(conn);
1971 #else /* RX_ENABLE_LOCKS */
1972 rxi_DestroyConnection(conn);
1973 #endif /* RX_ENABLE_LOCKS */
1977 #ifdef RX_ENABLE_LOCKS
1978 while (rx_connCleanup_list) {
1979 struct rx_connection *conn;
1980 conn = rx_connCleanup_list;
1981 rx_connCleanup_list = rx_connCleanup_list->next;
1982 MUTEX_EXIT(&rx_connHashTable_lock);
1983 rxi_CleanupConnection(conn);
1984 MUTEX_ENTER(&rx_connHashTable_lock);
1986 MUTEX_EXIT(&rx_connHashTable_lock);
1987 #endif /* RX_ENABLE_LOCKS */
1996 /* if we wakeup packet waiter too often, can get in loop with two
1997 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
1999 rxi_PacketsUnWait(void)
2001 if (!rx_waitingForPackets) {
2005 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2006 return; /* still over quota */
2009 rx_waitingForPackets = 0;
2010 #ifdef RX_ENABLE_LOCKS
2011 CV_BROADCAST(&rx_waitingForPackets_cv);
2013 osi_rxWakeup(&rx_waitingForPackets);
2019 /* ------------------Internal interfaces------------------------- */
2021 /* Return this process's service structure for the
2022 * specified socket and service */
2024 rxi_FindService(register osi_socket socket, register u_short serviceId)
2026 register struct rx_service **sp;
2027 for (sp = &rx_services[0]; *sp; sp++) {
2028 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2034 /* Allocate a call structure, for the indicated channel of the
2035 * supplied connection. The mode and state of the call must be set by
2036 * the caller. Returns the call with mutex locked. */
2038 rxi_NewCall(register struct rx_connection *conn, register int channel)
2040 register struct rx_call *call;
2041 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2042 register struct rx_call *cp; /* Call pointer temp */
2043 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2044 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2046 /* Grab an existing call structure, or allocate a new one.
2047 * Existing call structures are assumed to have been left reset by
2049 MUTEX_ENTER(&rx_freeCallQueue_lock);
2051 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2053 * EXCEPT that the TQ might not yet be cleared out.
2054 * Skip over those with in-use TQs.
2057 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2058 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2064 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2065 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2066 call = queue_First(&rx_freeCallQueue, rx_call);
2067 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2069 MUTEX_ENTER(&rx_stats_mutex);
2070 rx_stats.nFreeCallStructs--;
2071 MUTEX_EXIT(&rx_stats_mutex);
2072 MUTEX_EXIT(&rx_freeCallQueue_lock);
2073 MUTEX_ENTER(&call->lock);
2074 CLEAR_CALL_QUEUE_LOCK(call);
2075 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2076 /* Now, if TQ wasn't cleared earlier, do it now. */
2077 if (call->flags & RX_CALL_TQ_CLEARME) {
2078 rxi_ClearTransmitQueue(call, 0);
2079 queue_Init(&call->tq);
2081 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2082 /* Bind the call to its connection structure */
2084 rxi_ResetCall(call, 1);
2086 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2088 MUTEX_EXIT(&rx_freeCallQueue_lock);
2089 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2090 MUTEX_ENTER(&call->lock);
2091 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2092 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2093 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2095 MUTEX_ENTER(&rx_stats_mutex);
2096 rx_stats.nCallStructs++;
2097 MUTEX_EXIT(&rx_stats_mutex);
2098 /* Initialize once-only items */
2099 queue_Init(&call->tq);
2100 queue_Init(&call->rq);
2101 queue_Init(&call->iovq);
2102 /* Bind the call to its connection structure (prereq for reset) */
2104 rxi_ResetCall(call, 1);
2106 call->channel = channel;
2107 call->callNumber = &conn->callNumber[channel];
2108 /* Note that the next expected call number is retained (in
2109 * conn->callNumber[i]), even if we reallocate the call structure
2111 conn->call[channel] = call;
2112 /* if the channel's never been used (== 0), we should start at 1, otherwise
2113 * the call number is valid from the last time this channel was used */
2114 if (*call->callNumber == 0)
2115 *call->callNumber = 1;
2120 /* A call has been inactive long enough that so we can throw away
2121 * state, including the call structure, which is placed on the call
2123 * Call is locked upon entry.
2124 * haveCTLock set if called from rxi_ReapConnections
2126 #ifdef RX_ENABLE_LOCKS
2128 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2129 #else /* RX_ENABLE_LOCKS */
2131 rxi_FreeCall(register struct rx_call *call)
2132 #endif /* RX_ENABLE_LOCKS */
2134 register int channel = call->channel;
2135 register struct rx_connection *conn = call->conn;
2138 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2139 (*call->callNumber)++;
2140 rxi_ResetCall(call, 0);
2141 call->conn->call[channel] = (struct rx_call *)0;
2143 MUTEX_ENTER(&rx_freeCallQueue_lock);
2144 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2145 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2146 /* A call may be free even though its transmit queue is still in use.
2147 * Since we search the call list from head to tail, put busy calls at
2148 * the head of the list, and idle calls at the tail.
2150 if (call->flags & RX_CALL_TQ_BUSY)
2151 queue_Prepend(&rx_freeCallQueue, call);
2153 queue_Append(&rx_freeCallQueue, call);
2154 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2155 queue_Append(&rx_freeCallQueue, call);
2156 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2157 MUTEX_ENTER(&rx_stats_mutex);
2158 rx_stats.nFreeCallStructs++;
2159 MUTEX_EXIT(&rx_stats_mutex);
2161 MUTEX_EXIT(&rx_freeCallQueue_lock);
2163 /* Destroy the connection if it was previously slated for
2164 * destruction, i.e. the Rx client code previously called
2165 * rx_DestroyConnection (client connections), or
2166 * rxi_ReapConnections called the same routine (server
2167 * connections). Only do this, however, if there are no
2168 * outstanding calls. Note that for fine grain locking, there appears
2169 * to be a deadlock in that rxi_FreeCall has a call locked and
2170 * DestroyConnectionNoLock locks each call in the conn. But note a
2171 * few lines up where we have removed this call from the conn.
2172 * If someone else destroys a connection, they either have no
2173 * call lock held or are going through this section of code.
2175 if (conn->flags & RX_CONN_DESTROY_ME) {
2176 MUTEX_ENTER(&conn->conn_data_lock);
2178 MUTEX_EXIT(&conn->conn_data_lock);
2179 #ifdef RX_ENABLE_LOCKS
2181 rxi_DestroyConnectionNoLock(conn);
2183 rxi_DestroyConnection(conn);
2184 #else /* RX_ENABLE_LOCKS */
2185 rxi_DestroyConnection(conn);
2186 #endif /* RX_ENABLE_LOCKS */
2190 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2192 rxi_Alloc(register size_t size)
2196 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2197 /* Grab the AFS filesystem lock. See afs/osi.h for the lock
2200 int glockOwner = ISAFS_GLOCK();
2204 MUTEX_ENTER(&rx_stats_mutex);
2206 rxi_Allocsize += size;
2207 MUTEX_EXIT(&rx_stats_mutex);
2208 #if (defined(AFS_AIX32_ENV) || defined(AFS_HPUX_ENV)) && !defined(AFS_HPUX100_ENV) && defined(KERNEL)
2209 if (size > AFS_SMALLOCSIZ) {
2210 p = (char *)osi_AllocMediumSpace(size);
2212 p = (char *)osi_AllocSmall(size, 1);
2213 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2218 p = (char *)osi_Alloc(size);
2221 osi_Panic("rxi_Alloc error");
2227 rxi_Free(void *addr, register size_t size)
2229 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2230 /* Grab the AFS filesystem lock. See afs/osi.h for the lock
2233 int glockOwner = ISAFS_GLOCK();
2237 MUTEX_ENTER(&rx_stats_mutex);
2239 rxi_Allocsize -= size;
2240 MUTEX_EXIT(&rx_stats_mutex);
2241 #if (defined(AFS_AIX32_ENV) || defined(AFS_HPUX_ENV)) && !defined(AFS_HPUX100_ENV) && defined(KERNEL)
2242 if (size > AFS_SMALLOCSIZ)
2243 osi_FreeMediumSpace(addr);
2245 osi_FreeSmall(addr);
2246 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2251 osi_Free(addr, size);
2255 /* Find the peer process represented by the supplied (host,port)
2256 * combination. If there is no appropriate active peer structure, a
2257 * new one will be allocated and initialized
2258 * The origPeer, if set, is a pointer to a peer structure on which the
2259 * refcount will be be decremented. This is used to replace the peer
2260 * structure hanging off a connection structure */
2262 rxi_FindPeer(register afs_uint32 host, register u_short port,
2263 struct rx_peer *origPeer, int create)
2265 register struct rx_peer *pp;
2267 hashIndex = PEER_HASH(host, port);
2268 MUTEX_ENTER(&rx_peerHashTable_lock);
2269 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2270 if ((pp->host == host) && (pp->port == port))
2275 pp = rxi_AllocPeer(); /* This bzero's *pp */
2276 pp->host = host; /* set here or in InitPeerParams is zero */
2278 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2279 queue_Init(&pp->congestionQueue);
2280 queue_Init(&pp->rpcStats);
2281 pp->next = rx_peerHashTable[hashIndex];
2282 rx_peerHashTable[hashIndex] = pp;
2283 rxi_InitPeerParams(pp);
2284 MUTEX_ENTER(&rx_stats_mutex);
2285 rx_stats.nPeerStructs++;
2286 MUTEX_EXIT(&rx_stats_mutex);
2293 origPeer->refCount--;
2294 MUTEX_EXIT(&rx_peerHashTable_lock);
2299 /* Find the connection at (host, port) started at epoch, and with the
2300 * given connection id. Creates the server connection if necessary.
2301 * The type specifies whether a client connection or a server
2302 * connection is desired. In both cases, (host, port) specify the
2303 * peer's (host, pair) pair. Client connections are not made
2304 * automatically by this routine. The parameter socket gives the
2305 * socket descriptor on which the packet was received. This is used,
2306 * in the case of server connections, to check that *new* connections
2307 * come via a valid (port, serviceId). Finally, the securityIndex
2308 * parameter must match the existing index for the connection. If a
2309 * server connection is created, it will be created using the supplied
2310 * index, if the index is valid for this service */
2311 struct rx_connection *
2312 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2313 register u_short port, u_short serviceId, afs_uint32 cid,
2314 afs_uint32 epoch, int type, u_int securityIndex)
2316 int hashindex, flag;
2317 register struct rx_connection *conn;
2318 hashindex = CONN_HASH(host, port, cid, epoch, type);
2319 MUTEX_ENTER(&rx_connHashTable_lock);
2320 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2321 rx_connHashTable[hashindex],
2324 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2325 && (epoch == conn->epoch)) {
2326 register struct rx_peer *pp = conn->peer;
2327 if (securityIndex != conn->securityIndex) {
2328 /* this isn't supposed to happen, but someone could forge a packet
2329 * like this, and there seems to be some CM bug that makes this
2330 * happen from time to time -- in which case, the fileserver
2332 MUTEX_EXIT(&rx_connHashTable_lock);
2333 return (struct rx_connection *)0;
2335 if (pp->host == host && pp->port == port)
2337 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2339 /* So what happens when it's a callback connection? */
2340 if ( /*type == RX_CLIENT_CONNECTION && */
2341 (conn->epoch & 0x80000000))
2345 /* the connection rxLastConn that was used the last time is not the
2346 ** one we are looking for now. Hence, start searching in the hash */
2348 conn = rx_connHashTable[hashindex];
2353 struct rx_service *service;
2354 if (type == RX_CLIENT_CONNECTION) {
2355 MUTEX_EXIT(&rx_connHashTable_lock);
2356 return (struct rx_connection *)0;
2358 service = rxi_FindService(socket, serviceId);
2359 if (!service || (securityIndex >= service->nSecurityObjects)
2360 || (service->securityObjects[securityIndex] == 0)) {
2361 MUTEX_EXIT(&rx_connHashTable_lock);
2362 return (struct rx_connection *)0;
2364 conn = rxi_AllocConnection(); /* This bzero's the connection */
2365 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2366 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2367 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2368 conn->next = rx_connHashTable[hashindex];
2369 rx_connHashTable[hashindex] = conn;
2370 conn->peer = rxi_FindPeer(host, port, 0, 1);
2371 conn->type = RX_SERVER_CONNECTION;
2372 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2373 conn->epoch = epoch;
2374 conn->cid = cid & RX_CIDMASK;
2375 /* conn->serial = conn->lastSerial = 0; */
2376 /* conn->timeout = 0; */
2377 conn->ackRate = RX_FAST_ACK_RATE;
2378 conn->service = service;
2379 conn->serviceId = serviceId;
2380 conn->securityIndex = securityIndex;
2381 conn->securityObject = service->securityObjects[securityIndex];
2382 conn->nSpecific = 0;
2383 conn->specific = NULL;
2384 rx_SetConnDeadTime(conn, service->connDeadTime);
2385 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2386 /* Notify security object of the new connection */
2387 RXS_NewConnection(conn->securityObject, conn);
2388 /* XXXX Connection timeout? */
2389 if (service->newConnProc)
2390 (*service->newConnProc) (conn);
2391 MUTEX_ENTER(&rx_stats_mutex);
2392 rx_stats.nServerConns++;
2393 MUTEX_EXIT(&rx_stats_mutex);
2396 MUTEX_ENTER(&conn->conn_data_lock);
2398 MUTEX_EXIT(&conn->conn_data_lock);
2400 rxLastConn = conn; /* store this connection as the last conn used */
2401 MUTEX_EXIT(&rx_connHashTable_lock);
2405 /* There are two packet tracing routines available for testing and monitoring
2406 * Rx. One is called just after every packet is received and the other is
2407 * called just before every packet is sent. Received packets, have had their
2408 * headers decoded, and packets to be sent have not yet had their headers
2409 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2410 * containing the network address. Both can be modified. The return value, if
2411 * non-zero, indicates that the packet should be dropped. */
2413 int (*rx_justReceived) () = 0;
2414 int (*rx_almostSent) () = 0;
2416 /* A packet has been received off the interface. Np is the packet, socket is
2417 * the socket number it was received from (useful in determining which service
2418 * this packet corresponds to), and (host, port) reflect the host,port of the
2419 * sender. This call returns the packet to the caller if it is finished with
2420 * it, rather than de-allocating it, just as a small performance hack */
2423 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2424 afs_uint32 host, u_short port, int *tnop,
2425 struct rx_call **newcallp)
2427 register struct rx_call *call;
2428 register struct rx_connection *conn;
2430 afs_uint32 currentCallNumber;
2436 struct rx_packet *tnp;
2439 /* We don't print out the packet until now because (1) the time may not be
2440 * accurate enough until now in the lwp implementation (rx_Listener only gets
2441 * the time after the packet is read) and (2) from a protocol point of view,
2442 * this is the first time the packet has been seen */
2443 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2444 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2445 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2446 np->header.serial, packetType, host, port, np->header.serviceId,
2447 np->header.epoch, np->header.cid, np->header.callNumber,
2448 np->header.seq, np->header.flags, np));
2451 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2452 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2455 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2456 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2459 /* If an input tracer function is defined, call it with the packet and
2460 * network address. Note this function may modify its arguments. */
2461 if (rx_justReceived) {
2462 struct sockaddr_in addr;
2464 addr.sin_family = AF_INET;
2465 addr.sin_port = port;
2466 addr.sin_addr.s_addr = host;
2467 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2468 addr.sin_len = sizeof(addr);
2469 #endif /* AFS_OSF_ENV */
2470 drop = (*rx_justReceived) (np, &addr);
2471 /* drop packet if return value is non-zero */
2474 port = addr.sin_port; /* in case fcn changed addr */
2475 host = addr.sin_addr.s_addr;
2479 /* If packet was not sent by the client, then *we* must be the client */
2480 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2481 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2483 /* Find the connection (or fabricate one, if we're the server & if
2484 * necessary) associated with this packet */
2486 rxi_FindConnection(socket, host, port, np->header.serviceId,
2487 np->header.cid, np->header.epoch, type,
2488 np->header.securityIndex);
2491 /* If no connection found or fabricated, just ignore the packet.
2492 * (An argument could be made for sending an abort packet for
2497 MUTEX_ENTER(&conn->conn_data_lock);
2498 if (conn->maxSerial < np->header.serial)
2499 conn->maxSerial = np->header.serial;
2500 MUTEX_EXIT(&conn->conn_data_lock);
2502 /* If the connection is in an error state, send an abort packet and ignore
2503 * the incoming packet */
2505 /* Don't respond to an abort packet--we don't want loops! */
2506 MUTEX_ENTER(&conn->conn_data_lock);
2507 if (np->header.type != RX_PACKET_TYPE_ABORT)
2508 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2510 MUTEX_EXIT(&conn->conn_data_lock);
2514 /* Check for connection-only requests (i.e. not call specific). */
2515 if (np->header.callNumber == 0) {
2516 switch (np->header.type) {
2517 case RX_PACKET_TYPE_ABORT:
2518 /* What if the supplied error is zero? */
2519 rxi_ConnectionError(conn, ntohl(rx_GetInt32(np, 0)));
2520 MUTEX_ENTER(&conn->conn_data_lock);
2522 MUTEX_EXIT(&conn->conn_data_lock);
2524 case RX_PACKET_TYPE_CHALLENGE:
2525 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2526 MUTEX_ENTER(&conn->conn_data_lock);
2528 MUTEX_EXIT(&conn->conn_data_lock);
2530 case RX_PACKET_TYPE_RESPONSE:
2531 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2532 MUTEX_ENTER(&conn->conn_data_lock);
2534 MUTEX_EXIT(&conn->conn_data_lock);
2536 case RX_PACKET_TYPE_PARAMS:
2537 case RX_PACKET_TYPE_PARAMS + 1:
2538 case RX_PACKET_TYPE_PARAMS + 2:
2539 /* ignore these packet types for now */
2540 MUTEX_ENTER(&conn->conn_data_lock);
2542 MUTEX_EXIT(&conn->conn_data_lock);
2547 /* Should not reach here, unless the peer is broken: send an
2549 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2550 MUTEX_ENTER(&conn->conn_data_lock);
2551 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2553 MUTEX_EXIT(&conn->conn_data_lock);
2558 channel = np->header.cid & RX_CHANNELMASK;
2559 call = conn->call[channel];
2560 #ifdef RX_ENABLE_LOCKS
2562 MUTEX_ENTER(&call->lock);
2563 /* Test to see if call struct is still attached to conn. */
2564 if (call != conn->call[channel]) {
2566 MUTEX_EXIT(&call->lock);
2567 if (type == RX_SERVER_CONNECTION) {
2568 call = conn->call[channel];
2569 /* If we started with no call attached and there is one now,
2570 * another thread is also running this routine and has gotten
2571 * the connection channel. We should drop this packet in the tests
2572 * below. If there was a call on this connection and it's now
2573 * gone, then we'll be making a new call below.
2574 * If there was previously a call and it's now different then
2575 * the old call was freed and another thread running this routine
2576 * has created a call on this channel. One of these two threads
2577 * has a packet for the old call and the code below handles those
2581 MUTEX_ENTER(&call->lock);
2583 /* This packet can't be for this call. If the new call address is
2584 * 0 then no call is running on this channel. If there is a call
2585 * then, since this is a client connection we're getting data for
2586 * it must be for the previous call.
2588 MUTEX_ENTER(&rx_stats_mutex);
2589 rx_stats.spuriousPacketsRead++;
2590 MUTEX_EXIT(&rx_stats_mutex);
2591 MUTEX_ENTER(&conn->conn_data_lock);
2593 MUTEX_EXIT(&conn->conn_data_lock);
2598 currentCallNumber = conn->callNumber[channel];
2600 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2601 if (np->header.callNumber < currentCallNumber) {
2602 MUTEX_ENTER(&rx_stats_mutex);
2603 rx_stats.spuriousPacketsRead++;
2604 MUTEX_EXIT(&rx_stats_mutex);
2605 #ifdef RX_ENABLE_LOCKS
2607 MUTEX_EXIT(&call->lock);
2609 MUTEX_ENTER(&conn->conn_data_lock);
2611 MUTEX_EXIT(&conn->conn_data_lock);
2615 MUTEX_ENTER(&conn->conn_call_lock);
2616 call = rxi_NewCall(conn, channel);
2617 MUTEX_EXIT(&conn->conn_call_lock);
2618 *call->callNumber = np->header.callNumber;
2619 call->state = RX_STATE_PRECALL;
2620 clock_GetTime(&call->queueTime);
2621 hzero(call->bytesSent);
2622 hzero(call->bytesRcvd);
2623 rxi_KeepAliveOn(call);
2624 } else if (np->header.callNumber != currentCallNumber) {
2625 /* Wait until the transmit queue is idle before deciding
2626 * whether to reset the current call. Chances are that the
2627 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2630 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2631 while ((call->state == RX_STATE_ACTIVE)
2632 && (call->flags & RX_CALL_TQ_BUSY)) {
2633 call->flags |= RX_CALL_TQ_WAIT;
2634 #ifdef RX_ENABLE_LOCKS
2635 CV_WAIT(&call->cv_tq, &call->lock);
2636 #else /* RX_ENABLE_LOCKS */
2637 osi_rxSleep(&call->tq);
2638 #endif /* RX_ENABLE_LOCKS */
2640 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2641 /* If the new call cannot be taken right now send a busy and set
2642 * the error condition in this call, so that it terminates as
2643 * quickly as possible */
2644 if (call->state == RX_STATE_ACTIVE) {
2645 struct rx_packet *tp;
2647 rxi_CallError(call, RX_CALL_DEAD);
2648 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2650 MUTEX_EXIT(&call->lock);
2651 MUTEX_ENTER(&conn->conn_data_lock);
2653 MUTEX_EXIT(&conn->conn_data_lock);
2656 rxi_ResetCall(call, 0);
2657 *call->callNumber = np->header.callNumber;
2658 call->state = RX_STATE_PRECALL;
2659 clock_GetTime(&call->queueTime);
2660 hzero(call->bytesSent);
2661 hzero(call->bytesRcvd);
2663 * If the number of queued calls exceeds the overload
2664 * threshold then abort this call.
2666 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2667 struct rx_packet *tp;
2669 rxi_CallError(call, rx_BusyError);
2670 tp = rxi_SendCallAbort(call, np, 1, 0);
2671 MUTEX_EXIT(&call->lock);
2672 MUTEX_ENTER(&conn->conn_data_lock);
2674 MUTEX_EXIT(&conn->conn_data_lock);
2677 rxi_KeepAliveOn(call);
2679 /* Continuing call; do nothing here. */
2681 } else { /* we're the client */
2682 /* Ignore all incoming acknowledgements for calls in DALLY state */
2683 if (call && (call->state == RX_STATE_DALLY)
2684 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2685 MUTEX_ENTER(&rx_stats_mutex);
2686 rx_stats.ignorePacketDally++;
2687 MUTEX_EXIT(&rx_stats_mutex);
2688 #ifdef RX_ENABLE_LOCKS
2690 MUTEX_EXIT(&call->lock);
2693 MUTEX_ENTER(&conn->conn_data_lock);
2695 MUTEX_EXIT(&conn->conn_data_lock);
2699 /* Ignore anything that's not relevant to the current call. If there
2700 * isn't a current call, then no packet is relevant. */
2701 if (!call || (np->header.callNumber != currentCallNumber)) {
2702 MUTEX_ENTER(&rx_stats_mutex);
2703 rx_stats.spuriousPacketsRead++;
2704 MUTEX_EXIT(&rx_stats_mutex);
2705 #ifdef RX_ENABLE_LOCKS
2707 MUTEX_EXIT(&call->lock);
2710 MUTEX_ENTER(&conn->conn_data_lock);
2712 MUTEX_EXIT(&conn->conn_data_lock);
2715 /* If the service security object index stamped in the packet does not
2716 * match the connection's security index, ignore the packet */
2717 if (np->header.securityIndex != conn->securityIndex) {
2718 #ifdef RX_ENABLE_LOCKS
2719 MUTEX_EXIT(&call->lock);
2721 MUTEX_ENTER(&conn->conn_data_lock);
2723 MUTEX_EXIT(&conn->conn_data_lock);
2727 /* If we're receiving the response, then all transmit packets are
2728 * implicitly acknowledged. Get rid of them. */
2729 if (np->header.type == RX_PACKET_TYPE_DATA) {
2730 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2731 /* XXX Hack. Because we must release the global rx lock when
2732 * sending packets (osi_NetSend) we drop all acks while we're
2733 * traversing the tq in rxi_Start sending packets out because
2734 * packets may move to the freePacketQueue as result of being here!
2735 * So we drop these packets until we're safely out of the
2736 * traversing. Really ugly!
2737 * For fine grain RX locking, we set the acked field in the
2738 * packets and let rxi_Start remove them from the transmit queue.
2740 if (call->flags & RX_CALL_TQ_BUSY) {
2741 #ifdef RX_ENABLE_LOCKS
2742 rxi_SetAcksInTransmitQueue(call);
2745 return np; /* xmitting; drop packet */
2748 rxi_ClearTransmitQueue(call, 0);
2750 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2751 rxi_ClearTransmitQueue(call, 0);
2752 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2754 if (np->header.type == RX_PACKET_TYPE_ACK) {
2755 /* now check to see if this is an ack packet acknowledging that the
2756 * server actually *lost* some hard-acked data. If this happens we
2757 * ignore this packet, as it may indicate that the server restarted in
2758 * the middle of a call. It is also possible that this is an old ack
2759 * packet. We don't abort the connection in this case, because this
2760 * *might* just be an old ack packet. The right way to detect a server
2761 * restart in the midst of a call is to notice that the server epoch
2763 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2764 * XXX unacknowledged. I think that this is off-by-one, but
2765 * XXX I don't dare change it just yet, since it will
2766 * XXX interact badly with the server-restart detection
2767 * XXX code in receiveackpacket. */
2768 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2769 MUTEX_ENTER(&rx_stats_mutex);
2770 rx_stats.spuriousPacketsRead++;
2771 MUTEX_EXIT(&rx_stats_mutex);
2772 MUTEX_EXIT(&call->lock);
2773 MUTEX_ENTER(&conn->conn_data_lock);
2775 MUTEX_EXIT(&conn->conn_data_lock);
2779 } /* else not a data packet */
2782 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2783 /* Set remote user defined status from packet */
2784 call->remoteStatus = np->header.userStatus;
2786 /* Note the gap between the expected next packet and the actual
2787 * packet that arrived, when the new packet has a smaller serial number
2788 * than expected. Rioses frequently reorder packets all by themselves,
2789 * so this will be quite important with very large window sizes.
2790 * Skew is checked against 0 here to avoid any dependence on the type of
2791 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2793 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2794 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2795 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2797 MUTEX_ENTER(&conn->conn_data_lock);
2798 skew = conn->lastSerial - np->header.serial;
2799 conn->lastSerial = np->header.serial;
2800 MUTEX_EXIT(&conn->conn_data_lock);
2802 register struct rx_peer *peer;
2804 if (skew > peer->inPacketSkew) {
2805 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2807 peer->inPacketSkew = skew;
2811 /* Now do packet type-specific processing */
2812 switch (np->header.type) {
2813 case RX_PACKET_TYPE_DATA:
2814 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2817 case RX_PACKET_TYPE_ACK:
2818 /* Respond immediately to ack packets requesting acknowledgement
2820 if (np->header.flags & RX_REQUEST_ACK) {
2822 (void)rxi_SendCallAbort(call, 0, 1, 0);
2824 (void)rxi_SendAck(call, 0, np->header.serial,
2825 RX_ACK_PING_RESPONSE, 1);
2827 np = rxi_ReceiveAckPacket(call, np, 1);
2829 case RX_PACKET_TYPE_ABORT:
2830 /* An abort packet: reset the connection, passing the error up to
2832 /* What if error is zero? */
2833 rxi_CallError(call, ntohl(*(afs_int32 *) rx_DataOf(np)));
2835 case RX_PACKET_TYPE_BUSY:
2838 case RX_PACKET_TYPE_ACKALL:
2839 /* All packets acknowledged, so we can drop all packets previously
2840 * readied for sending */
2841 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2842 /* XXX Hack. We because we can't release the global rx lock when
2843 * sending packets (osi_NetSend) we drop all ack pkts while we're
2844 * traversing the tq in rxi_Start sending packets out because
2845 * packets may move to the freePacketQueue as result of being
2846 * here! So we drop these packets until we're safely out of the
2847 * traversing. Really ugly!
2848 * For fine grain RX locking, we set the acked field in the packets
2849 * and let rxi_Start remove the packets from the transmit queue.
2851 if (call->flags & RX_CALL_TQ_BUSY) {
2852 #ifdef RX_ENABLE_LOCKS
2853 rxi_SetAcksInTransmitQueue(call);
2855 #else /* RX_ENABLE_LOCKS */
2857 return np; /* xmitting; drop packet */
2858 #endif /* RX_ENABLE_LOCKS */
2860 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2861 rxi_ClearTransmitQueue(call, 0);
2864 /* Should not reach here, unless the peer is broken: send an abort
2866 rxi_CallError(call, RX_PROTOCOL_ERROR);
2867 np = rxi_SendCallAbort(call, np, 1, 0);
2870 /* Note when this last legitimate packet was received, for keep-alive
2871 * processing. Note, we delay getting the time until now in the hope that
2872 * the packet will be delivered to the user before any get time is required
2873 * (if not, then the time won't actually be re-evaluated here). */
2874 call->lastReceiveTime = clock_Sec();
2875 MUTEX_EXIT(&call->lock);
2876 MUTEX_ENTER(&conn->conn_data_lock);
2878 MUTEX_EXIT(&conn->conn_data_lock);
2882 /* return true if this is an "interesting" connection from the point of view
2883 of someone trying to debug the system */
2885 rxi_IsConnInteresting(struct rx_connection *aconn)
2888 register struct rx_call *tcall;
2890 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2892 for (i = 0; i < RX_MAXCALLS; i++) {
2893 tcall = aconn->call[i];
2895 if ((tcall->state == RX_STATE_PRECALL)
2896 || (tcall->state == RX_STATE_ACTIVE))
2898 if ((tcall->mode == RX_MODE_SENDING)
2899 || (tcall->mode == RX_MODE_RECEIVING))
2907 /* if this is one of the last few packets AND it wouldn't be used by the
2908 receiving call to immediately satisfy a read request, then drop it on
2909 the floor, since accepting it might prevent a lock-holding thread from
2910 making progress in its reading. If a call has been cleared while in
2911 the precall state then ignore all subsequent packets until the call
2912 is assigned to a thread. */
2915 TooLow(struct rx_packet *ap, struct rx_call *acall)
2918 MUTEX_ENTER(&rx_stats_mutex);
2919 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2920 && (acall->state == RX_STATE_PRECALL))
2921 || ((rx_nFreePackets < rxi_dataQuota + 2)
2922 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2923 && (acall->flags & RX_CALL_READER_WAIT)))) {
2926 MUTEX_EXIT(&rx_stats_mutex);
2932 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2933 struct rx_call *acall)
2935 struct rx_call *call = acall;
2939 MUTEX_ENTER(&conn->conn_data_lock);
2940 conn->checkReachEvent = NULL;
2941 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2944 MUTEX_EXIT(&conn->conn_data_lock);
2948 MUTEX_ENTER(&conn->conn_call_lock);
2949 MUTEX_ENTER(&conn->conn_data_lock);
2950 for (i = 0; i < RX_MAXCALLS; i++) {
2951 struct rx_call *tc = conn->call[i];
2952 if (tc && tc->state == RX_STATE_PRECALL) {
2958 /* Indicate that rxi_CheckReachEvent is no longer running by
2959 * clearing the flag. Must be atomic under conn_data_lock to
2960 * avoid a new call slipping by: rxi_CheckConnReach holds
2961 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
2963 conn->flags &= ~RX_CONN_ATTACHWAIT;
2964 MUTEX_EXIT(&conn->conn_data_lock);
2965 MUTEX_EXIT(&conn->conn_call_lock);
2970 MUTEX_ENTER(&call->lock);
2971 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
2973 MUTEX_EXIT(&call->lock);
2975 clock_GetTime(&when);
2976 when.sec += RX_CHECKREACH_TIMEOUT;
2977 MUTEX_ENTER(&conn->conn_data_lock);
2978 if (!conn->checkReachEvent) {
2980 conn->checkReachEvent =
2981 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
2983 MUTEX_EXIT(&conn->conn_data_lock);
2989 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
2991 struct rx_service *service = conn->service;
2992 struct rx_peer *peer = conn->peer;
2993 afs_uint32 now, lastReach;
2995 if (service->checkReach == 0)
2999 MUTEX_ENTER(&peer->peer_lock);
3000 lastReach = peer->lastReachTime;
3001 MUTEX_EXIT(&peer->peer_lock);
3002 if (now - lastReach < RX_CHECKREACH_TTL)
3005 MUTEX_ENTER(&conn->conn_data_lock);
3006 if (conn->flags & RX_CONN_ATTACHWAIT) {
3007 MUTEX_EXIT(&conn->conn_data_lock);
3010 conn->flags |= RX_CONN_ATTACHWAIT;
3011 MUTEX_EXIT(&conn->conn_data_lock);
3012 if (!conn->checkReachEvent)
3013 rxi_CheckReachEvent(NULL, conn, call);
3018 /* try to attach call, if authentication is complete */
3020 TryAttach(register struct rx_call *acall, register osi_socket socket,
3021 register int *tnop, register struct rx_call **newcallp,
3024 struct rx_connection *conn = acall->conn;
3026 if (conn->type == RX_SERVER_CONNECTION
3027 && acall->state == RX_STATE_PRECALL) {
3028 /* Don't attach until we have any req'd. authentication. */
3029 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3030 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3031 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3032 /* Note: this does not necessarily succeed; there
3033 * may not any proc available
3036 rxi_ChallengeOn(acall->conn);
3041 /* A data packet has been received off the interface. This packet is
3042 * appropriate to the call (the call is in the right state, etc.). This
3043 * routine can return a packet to the caller, for re-use */
3046 rxi_ReceiveDataPacket(register struct rx_call *call,
3047 register struct rx_packet *np, int istack,
3048 osi_socket socket, afs_uint32 host, u_short port,
3049 int *tnop, struct rx_call **newcallp)
3051 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3055 afs_uint32 seq, serial, flags;
3057 struct rx_packet *tnp;
3059 MUTEX_ENTER(&rx_stats_mutex);
3060 rx_stats.dataPacketsRead++;
3061 MUTEX_EXIT(&rx_stats_mutex);
3064 /* If there are no packet buffers, drop this new packet, unless we can find
3065 * packet buffers from inactive calls */
3067 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3068 MUTEX_ENTER(&rx_freePktQ_lock);
3069 rxi_NeedMorePackets = TRUE;
3070 MUTEX_EXIT(&rx_freePktQ_lock);
3071 MUTEX_ENTER(&rx_stats_mutex);
3072 rx_stats.noPacketBuffersOnRead++;
3073 MUTEX_EXIT(&rx_stats_mutex);
3074 call->rprev = np->header.serial;
3075 rxi_calltrace(RX_TRACE_DROP, call);
3076 dpf(("packet %x dropped on receipt - quota problems", np));
3078 rxi_ClearReceiveQueue(call);
3079 clock_GetTime(&when);
3080 clock_Add(&when, &rx_softAckDelay);
3081 if (!call->delayedAckEvent
3082 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3083 rxevent_Cancel(call->delayedAckEvent, call,
3084 RX_CALL_REFCOUNT_DELAY);
3085 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3086 call->delayedAckEvent =
3087 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3089 /* we've damaged this call already, might as well do it in. */
3095 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3096 * packet is one of several packets transmitted as a single
3097 * datagram. Do not send any soft or hard acks until all packets
3098 * in a jumbogram have been processed. Send negative acks right away.
3100 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3101 /* tnp is non-null when there are more packets in the
3102 * current jumbo gram */
3109 seq = np->header.seq;
3110 serial = np->header.serial;
3111 flags = np->header.flags;
3113 /* If the call is in an error state, send an abort message */
3115 return rxi_SendCallAbort(call, np, istack, 0);
3117 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3118 * AFS 3.5 jumbogram. */
3119 if (flags & RX_JUMBO_PACKET) {
3120 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3125 if (np->header.spare != 0) {
3126 MUTEX_ENTER(&call->conn->conn_data_lock);
3127 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3128 MUTEX_EXIT(&call->conn->conn_data_lock);
3131 /* The usual case is that this is the expected next packet */
3132 if (seq == call->rnext) {
3134 /* Check to make sure it is not a duplicate of one already queued */
3135 if (queue_IsNotEmpty(&call->rq)
3136 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3137 MUTEX_ENTER(&rx_stats_mutex);
3138 rx_stats.dupPacketsRead++;
3139 MUTEX_EXIT(&rx_stats_mutex);
3140 dpf(("packet %x dropped on receipt - duplicate", np));
3141 rxevent_Cancel(call->delayedAckEvent, call,
3142 RX_CALL_REFCOUNT_DELAY);
3143 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3149 /* It's the next packet. Stick it on the receive queue
3150 * for this call. Set newPackets to make sure we wake
3151 * the reader once all packets have been processed */
3152 queue_Prepend(&call->rq, np);
3154 np = NULL; /* We can't use this anymore */
3157 /* If an ack is requested then set a flag to make sure we
3158 * send an acknowledgement for this packet */
3159 if (flags & RX_REQUEST_ACK) {
3160 ackNeeded = RX_ACK_REQUESTED;
3163 /* Keep track of whether we have received the last packet */
3164 if (flags & RX_LAST_PACKET) {
3165 call->flags |= RX_CALL_HAVE_LAST;
3169 /* Check whether we have all of the packets for this call */
3170 if (call->flags & RX_CALL_HAVE_LAST) {
3171 afs_uint32 tseq; /* temporary sequence number */
3172 struct rx_packet *tp; /* Temporary packet pointer */
3173 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3175 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3176 if (tseq != tp->header.seq)
3178 if (tp->header.flags & RX_LAST_PACKET) {
3179 call->flags |= RX_CALL_RECEIVE_DONE;
3186 /* Provide asynchronous notification for those who want it
3187 * (e.g. multi rx) */
3188 if (call->arrivalProc) {
3189 (*call->arrivalProc) (call, call->arrivalProcHandle,
3190 (int)call->arrivalProcArg);
3191 call->arrivalProc = (VOID(*)())0;
3194 /* Update last packet received */
3197 /* If there is no server process serving this call, grab
3198 * one, if available. We only need to do this once. If a
3199 * server thread is available, this thread becomes a server
3200 * thread and the server thread becomes a listener thread. */
3202 TryAttach(call, socket, tnop, newcallp, 0);
3205 /* This is not the expected next packet. */
3207 /* Determine whether this is a new or old packet, and if it's
3208 * a new one, whether it fits into the current receive window.
3209 * Also figure out whether the packet was delivered in sequence.
3210 * We use the prev variable to determine whether the new packet
3211 * is the successor of its immediate predecessor in the
3212 * receive queue, and the missing flag to determine whether
3213 * any of this packets predecessors are missing. */
3215 afs_uint32 prev; /* "Previous packet" sequence number */
3216 struct rx_packet *tp; /* Temporary packet pointer */
3217 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3218 int missing; /* Are any predecessors missing? */
3220 /* If the new packet's sequence number has been sent to the
3221 * application already, then this is a duplicate */
3222 if (seq < call->rnext) {
3223 MUTEX_ENTER(&rx_stats_mutex);
3224 rx_stats.dupPacketsRead++;
3225 MUTEX_EXIT(&rx_stats_mutex);
3226 rxevent_Cancel(call->delayedAckEvent, call,
3227 RX_CALL_REFCOUNT_DELAY);
3228 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3234 /* If the sequence number is greater than what can be
3235 * accomodated by the current window, then send a negative
3236 * acknowledge and drop the packet */
3237 if ((call->rnext + call->rwind) <= seq) {
3238 rxevent_Cancel(call->delayedAckEvent, call,
3239 RX_CALL_REFCOUNT_DELAY);
3240 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3247 /* Look for the packet in the queue of old received packets */
3248 for (prev = call->rnext - 1, missing =
3249 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3250 /*Check for duplicate packet */
3251 if (seq == tp->header.seq) {
3252 MUTEX_ENTER(&rx_stats_mutex);
3253 rx_stats.dupPacketsRead++;
3254 MUTEX_EXIT(&rx_stats_mutex);
3255 rxevent_Cancel(call->delayedAckEvent, call,
3256 RX_CALL_REFCOUNT_DELAY);
3257 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3263 /* If we find a higher sequence packet, break out and
3264 * insert the new packet here. */
3265 if (seq < tp->header.seq)
3267 /* Check for missing packet */
3268 if (tp->header.seq != prev + 1) {
3272 prev = tp->header.seq;
3275 /* Keep track of whether we have received the last packet. */
3276 if (flags & RX_LAST_PACKET) {
3277 call->flags |= RX_CALL_HAVE_LAST;
3280 /* It's within the window: add it to the the receive queue.
3281 * tp is left by the previous loop either pointing at the
3282 * packet before which to insert the new packet, or at the
3283 * queue head if the queue is empty or the packet should be
3285 queue_InsertBefore(tp, np);
3289 /* Check whether we have all of the packets for this call */
3290 if ((call->flags & RX_CALL_HAVE_LAST)
3291 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3292 afs_uint32 tseq; /* temporary sequence number */
3295 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3296 if (tseq != tp->header.seq)
3298 if (tp->header.flags & RX_LAST_PACKET) {
3299 call->flags |= RX_CALL_RECEIVE_DONE;
3306 /* We need to send an ack of the packet is out of sequence,
3307 * or if an ack was requested by the peer. */
3308 if (seq != prev + 1 || missing || (flags & RX_REQUEST_ACK)) {
3309 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3312 /* Acknowledge the last packet for each call */
3313 if (flags & RX_LAST_PACKET) {
3324 * If the receiver is waiting for an iovec, fill the iovec
3325 * using the data from the receive queue */
3326 if (call->flags & RX_CALL_IOVEC_WAIT) {
3327 didHardAck = rxi_FillReadVec(call, serial);
3328 /* the call may have been aborted */
3337 /* Wakeup the reader if any */
3338 if ((call->flags & RX_CALL_READER_WAIT)
3339 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3340 || (call->iovNext >= call->iovMax)
3341 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3342 call->flags &= ~RX_CALL_READER_WAIT;
3343 #ifdef RX_ENABLE_LOCKS
3344 CV_BROADCAST(&call->cv_rq);
3346 osi_rxWakeup(&call->rq);
3352 * Send an ack when requested by the peer, or once every
3353 * rxi_SoftAckRate packets until the last packet has been
3354 * received. Always send a soft ack for the last packet in
3355 * the server's reply. */
3357 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3358 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3359 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3360 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3361 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3362 } else if (call->nSoftAcks) {
3363 clock_GetTime(&when);
3364 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3365 clock_Add(&when, &rx_lastAckDelay);
3367 clock_Add(&when, &rx_softAckDelay);
3369 if (!call->delayedAckEvent
3370 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3371 rxevent_Cancel(call->delayedAckEvent, call,
3372 RX_CALL_REFCOUNT_DELAY);
3373 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3374 call->delayedAckEvent =
3375 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3377 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3378 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3385 static void rxi_ComputeRate();
3389 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3391 struct rx_peer *peer = conn->peer;
3393 MUTEX_ENTER(&peer->peer_lock);
3394 peer->lastReachTime = clock_Sec();
3395 MUTEX_EXIT(&peer->peer_lock);
3397 MUTEX_ENTER(&conn->conn_data_lock);
3398 if (conn->flags & RX_CONN_ATTACHWAIT) {
3401 conn->flags &= ~RX_CONN_ATTACHWAIT;
3402 MUTEX_EXIT(&conn->conn_data_lock);
3404 for (i = 0; i < RX_MAXCALLS; i++) {
3405 struct rx_call *call = conn->call[i];
3408 MUTEX_ENTER(&call->lock);
3409 /* tnop can be null if newcallp is null */
3410 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3412 MUTEX_EXIT(&call->lock);
3416 MUTEX_EXIT(&conn->conn_data_lock);
3419 /* rxi_ComputePeerNetStats
3421 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3422 * estimates (like RTT and throughput) based on ack packets. Caller
3423 * must ensure that the packet in question is the right one (i.e.
3424 * serial number matches).
3427 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3428 struct rx_ackPacket *ap, struct rx_packet *np)
3430 struct rx_peer *peer = call->conn->peer;
3432 /* Use RTT if not delayed by client. */
3433 if (ap->reason != RX_ACK_DELAY)
3434 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3436 rxi_ComputeRate(peer, call, p, np, ap->reason);
3440 /* The real smarts of the whole thing. */
3442 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3445 struct rx_ackPacket *ap;
3447 register struct rx_packet *tp;
3448 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3449 register struct rx_connection *conn = call->conn;
3450 struct rx_peer *peer = conn->peer;
3453 /* because there are CM's that are bogus, sending weird values for this. */
3454 afs_uint32 skew = 0;
3459 int newAckCount = 0;
3460 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3461 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3463 MUTEX_ENTER(&rx_stats_mutex);
3464 rx_stats.ackPacketsRead++;
3465 MUTEX_EXIT(&rx_stats_mutex);
3466 ap = (struct rx_ackPacket *)rx_DataOf(np);
3467 nbytes = rx_Contiguous(np) - ((ap->acks) - (u_char *) ap);
3469 return np; /* truncated ack packet */
3471 /* depends on ack packet struct */
3472 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3473 first = ntohl(ap->firstPacket);
3474 serial = ntohl(ap->serial);
3475 /* temporarily disabled -- needs to degrade over time
3476 * skew = ntohs(ap->maxSkew); */
3478 /* Ignore ack packets received out of order */
3479 if (first < call->tfirst) {
3483 if (np->header.flags & RX_SLOW_START_OK) {
3484 call->flags |= RX_CALL_SLOW_START_OK;
3487 if (ap->reason == RX_ACK_PING_RESPONSE)
3488 rxi_UpdatePeerReach(conn, call);
3493 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3494 ap->reason, ntohl(ap->previousPacket),
3495 (unsigned int)np->header.seq, (unsigned int)serial,
3496 (unsigned int)skew, ntohl(ap->firstPacket));
3499 for (offset = 0; offset < nAcks; offset++)
3500 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3507 /* Update the outgoing packet skew value to the latest value of
3508 * the peer's incoming packet skew value. The ack packet, of
3509 * course, could arrive out of order, but that won't affect things
3511 MUTEX_ENTER(&peer->peer_lock);
3512 peer->outPacketSkew = skew;
3514 /* Check for packets that no longer need to be transmitted, and
3515 * discard them. This only applies to packets positively
3516 * acknowledged as having been sent to the peer's upper level.
3517 * All other packets must be retained. So only packets with
3518 * sequence numbers < ap->firstPacket are candidates. */
3519 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3520 if (tp->header.seq >= first)
3522 call->tfirst = tp->header.seq + 1;
3524 && (tp->header.serial == serial || tp->firstSerial == serial))
3525 rxi_ComputePeerNetStats(call, tp, ap, np);
3526 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3527 /* XXX Hack. Because we have to release the global rx lock when sending
3528 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3529 * in rxi_Start sending packets out because packets may move to the
3530 * freePacketQueue as result of being here! So we drop these packets until
3531 * we're safely out of the traversing. Really ugly!
3532 * To make it even uglier, if we're using fine grain locking, we can
3533 * set the ack bits in the packets and have rxi_Start remove the packets
3534 * when it's done transmitting.
3536 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3539 if (call->flags & RX_CALL_TQ_BUSY) {
3540 #ifdef RX_ENABLE_LOCKS
3541 tp->flags |= RX_PKTFLAG_ACKED;
3542 call->flags |= RX_CALL_TQ_SOME_ACKED;
3543 #else /* RX_ENABLE_LOCKS */
3545 #endif /* RX_ENABLE_LOCKS */
3547 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3550 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3555 /* Give rate detector a chance to respond to ping requests */
3556 if (ap->reason == RX_ACK_PING_RESPONSE) {
3557 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3561 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3563 /* Now go through explicit acks/nacks and record the results in
3564 * the waiting packets. These are packets that can't be released
3565 * yet, even with a positive acknowledge. This positive
3566 * acknowledge only means the packet has been received by the
3567 * peer, not that it will be retained long enough to be sent to
3568 * the peer's upper level. In addition, reset the transmit timers
3569 * of any missing packets (those packets that must be missing
3570 * because this packet was out of sequence) */
3572 call->nSoftAcked = 0;
3573 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3574 /* Update round trip time if the ack was stimulated on receipt
3576 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3577 #ifdef RX_ENABLE_LOCKS
3578 if (tp->header.seq >= first)
3579 #endif /* RX_ENABLE_LOCKS */
3580 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3582 && (tp->header.serial == serial || tp->firstSerial == serial))
3583 rxi_ComputePeerNetStats(call, tp, ap, np);
3585 /* Set the acknowledge flag per packet based on the
3586 * information in the ack packet. An acknowlegded packet can
3587 * be downgraded when the server has discarded a packet it
3588 * soacked previously, or when an ack packet is received
3589 * out of sequence. */
3590 if (tp->header.seq < first) {
3591 /* Implicit ack information */
3592 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3595 tp->flags |= RX_PKTFLAG_ACKED;
3596 } else if (tp->header.seq < first + nAcks) {
3597 /* Explicit ack information: set it in the packet appropriately */
3598 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3599 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3601 tp->flags |= RX_PKTFLAG_ACKED;
3609 tp->flags &= ~RX_PKTFLAG_ACKED;
3613 tp->flags &= ~RX_PKTFLAG_ACKED;
3617 /* If packet isn't yet acked, and it has been transmitted at least
3618 * once, reset retransmit time using latest timeout
3619 * ie, this should readjust the retransmit timer for all outstanding
3620 * packets... So we don't just retransmit when we should know better*/
3622 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3623 tp->retryTime = tp->timeSent;
3624 clock_Add(&tp->retryTime, &peer->timeout);
3625 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3626 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3630 /* If the window has been extended by this acknowledge packet,
3631 * then wakeup a sender waiting in alloc for window space, or try
3632 * sending packets now, if he's been sitting on packets due to
3633 * lack of window space */
3634 if (call->tnext < (call->tfirst + call->twind)) {
3635 #ifdef RX_ENABLE_LOCKS
3636 CV_SIGNAL(&call->cv_twind);
3638 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3639 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3640 osi_rxWakeup(&call->twind);
3643 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3644 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3648 /* if the ack packet has a receivelen field hanging off it,
3649 * update our state */
3650 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3653 /* If the ack packet has a "recommended" size that is less than
3654 * what I am using now, reduce my size to match */
3655 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3656 sizeof(afs_int32), &tSize);
3657 tSize = (afs_uint32) ntohl(tSize);
3658 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3660 /* Get the maximum packet size to send to this peer */
3661 rx_packetread(np, rx_AckDataSize(ap->nAcks), sizeof(afs_int32),
3663 tSize = (afs_uint32) ntohl(tSize);
3664 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3665 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3667 /* sanity check - peer might have restarted with different params.
3668 * If peer says "send less", dammit, send less... Peer should never
3669 * be unable to accept packets of the size that prior AFS versions would
3670 * send without asking. */
3671 if (peer->maxMTU != tSize) {
3672 peer->maxMTU = tSize;
3673 peer->MTU = MIN(tSize, peer->MTU);
3674 call->MTU = MIN(call->MTU, tSize);
3678 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3681 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3682 sizeof(afs_int32), &tSize);
3683 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3684 if (tSize < call->twind) { /* smaller than our send */
3685 call->twind = tSize; /* window, we must send less... */
3686 call->ssthresh = MIN(call->twind, call->ssthresh);
3689 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3690 * network MTU confused with the loopback MTU. Calculate the
3691 * maximum MTU here for use in the slow start code below.
3693 maxMTU = peer->maxMTU;
3694 /* Did peer restart with older RX version? */
3695 if (peer->maxDgramPackets > 1) {
3696 peer->maxDgramPackets = 1;
3698 } else if (np->length >=
3699 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3702 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3703 sizeof(afs_int32), &tSize);
3704 tSize = (afs_uint32) ntohl(tSize);
3706 * As of AFS 3.5 we set the send window to match the receive window.
3708 if (tSize < call->twind) {
3709 call->twind = tSize;
3710 call->ssthresh = MIN(call->twind, call->ssthresh);
3711 } else if (tSize > call->twind) {
3712 call->twind = tSize;
3716 * As of AFS 3.5, a jumbogram is more than one fixed size
3717 * packet transmitted in a single UDP datagram. If the remote
3718 * MTU is smaller than our local MTU then never send a datagram
3719 * larger than the natural MTU.
3722 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3723 sizeof(afs_int32), &tSize);
3724 maxDgramPackets = (afs_uint32) ntohl(tSize);
3725 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3727 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3728 maxDgramPackets = MIN(maxDgramPackets, tSize);
3729 if (maxDgramPackets > 1) {
3730 peer->maxDgramPackets = maxDgramPackets;
3731 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3733 peer->maxDgramPackets = 1;
3734 call->MTU = peer->natMTU;
3736 } else if (peer->maxDgramPackets > 1) {
3737 /* Restarted with lower version of RX */
3738 peer->maxDgramPackets = 1;
3740 } else if (peer->maxDgramPackets > 1
3741 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3742 /* Restarted with lower version of RX */
3743 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3744 peer->natMTU = OLD_MAX_PACKET_SIZE;
3745 peer->MTU = OLD_MAX_PACKET_SIZE;
3746 peer->maxDgramPackets = 1;
3747 peer->nDgramPackets = 1;
3749 call->MTU = OLD_MAX_PACKET_SIZE;
3754 * Calculate how many datagrams were successfully received after
3755 * the first missing packet and adjust the negative ack counter
3760 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3761 if (call->nNacks < nNacked) {
3762 call->nNacks = nNacked;
3771 if (call->flags & RX_CALL_FAST_RECOVER) {
3773 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3775 call->flags &= ~RX_CALL_FAST_RECOVER;
3776 call->cwind = call->nextCwind;
3777 call->nextCwind = 0;
3780 call->nCwindAcks = 0;
3781 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3782 /* Three negative acks in a row trigger congestion recovery */
3783 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3784 MUTEX_EXIT(&peer->peer_lock);
3785 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3786 /* someone else is waiting to start recovery */
3789 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3790 while (call->flags & RX_CALL_TQ_BUSY) {
3791 call->flags |= RX_CALL_TQ_WAIT;
3792 #ifdef RX_ENABLE_LOCKS
3793 CV_WAIT(&call->cv_tq, &call->lock);
3794 #else /* RX_ENABLE_LOCKS */
3795 osi_rxSleep(&call->tq);
3796 #endif /* RX_ENABLE_LOCKS */
3798 MUTEX_ENTER(&peer->peer_lock);
3799 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3800 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3801 call->flags |= RX_CALL_FAST_RECOVER;
3802 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3804 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3805 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3806 call->nextCwind = call->ssthresh;
3809 peer->MTU = call->MTU;
3810 peer->cwind = call->nextCwind;
3811 peer->nDgramPackets = call->nDgramPackets;
3813 call->congestSeq = peer->congestSeq;
3814 /* Reset the resend times on the packets that were nacked
3815 * so we will retransmit as soon as the window permits*/
3816 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3818 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3819 clock_Zero(&tp->retryTime);
3821 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3826 /* If cwind is smaller than ssthresh, then increase
3827 * the window one packet for each ack we receive (exponential
3829 * If cwind is greater than or equal to ssthresh then increase
3830 * the congestion window by one packet for each cwind acks we
3831 * receive (linear growth). */
3832 if (call->cwind < call->ssthresh) {
3834 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3835 call->nCwindAcks = 0;
3837 call->nCwindAcks += newAckCount;
3838 if (call->nCwindAcks >= call->cwind) {
3839 call->nCwindAcks = 0;
3840 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3844 * If we have received several acknowledgements in a row then
3845 * it is time to increase the size of our datagrams
3847 if ((int)call->nAcks > rx_nDgramThreshold) {
3848 if (peer->maxDgramPackets > 1) {
3849 if (call->nDgramPackets < peer->maxDgramPackets) {
3850 call->nDgramPackets++;
3852 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3853 } else if (call->MTU < peer->maxMTU) {
3854 call->MTU += peer->natMTU;
3855 call->MTU = MIN(call->MTU, peer->maxMTU);
3861 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3863 /* Servers need to hold the call until all response packets have
3864 * been acknowledged. Soft acks are good enough since clients
3865 * are not allowed to clear their receive queues. */
3866 if (call->state == RX_STATE_HOLD
3867 && call->tfirst + call->nSoftAcked >= call->tnext) {
3868 call->state = RX_STATE_DALLY;
3869 rxi_ClearTransmitQueue(call, 0);
3870 } else if (!queue_IsEmpty(&call->tq)) {
3871 rxi_Start(0, call, istack);
3876 /* Received a response to a challenge packet */
3878 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3879 register struct rx_packet *np, int istack)
3883 /* Ignore the packet if we're the client */
3884 if (conn->type == RX_CLIENT_CONNECTION)
3887 /* If already authenticated, ignore the packet (it's probably a retry) */
3888 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3891 /* Otherwise, have the security object evaluate the response packet */
3892 error = RXS_CheckResponse(conn->securityObject, conn, np);
3894 /* If the response is invalid, reset the connection, sending
3895 * an abort to the peer */
3899 rxi_ConnectionError(conn, error);
3900 MUTEX_ENTER(&conn->conn_data_lock);
3901 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3902 MUTEX_EXIT(&conn->conn_data_lock);
3905 /* If the response is valid, any calls waiting to attach
3906 * servers can now do so */
3909 for (i = 0; i < RX_MAXCALLS; i++) {
3910 struct rx_call *call = conn->call[i];
3912 MUTEX_ENTER(&call->lock);
3913 if (call->state == RX_STATE_PRECALL)
3914 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
3915 /* tnop can be null if newcallp is null */
3916 MUTEX_EXIT(&call->lock);
3920 /* Update the peer reachability information, just in case
3921 * some calls went into attach-wait while we were waiting
3922 * for authentication..
3924 rxi_UpdatePeerReach(conn, NULL);
3929 /* A client has received an authentication challenge: the security
3930 * object is asked to cough up a respectable response packet to send
3931 * back to the server. The server is responsible for retrying the
3932 * challenge if it fails to get a response. */
3935 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
3936 register struct rx_packet *np, int istack)
3940 /* Ignore the challenge if we're the server */
3941 if (conn->type == RX_SERVER_CONNECTION)
3944 /* Ignore the challenge if the connection is otherwise idle; someone's
3945 * trying to use us as an oracle. */
3946 if (!rxi_HasActiveCalls(conn))
3949 /* Send the security object the challenge packet. It is expected to fill
3950 * in the response. */
3951 error = RXS_GetResponse(conn->securityObject, conn, np);
3953 /* If the security object is unable to return a valid response, reset the
3954 * connection and send an abort to the peer. Otherwise send the response
3955 * packet to the peer connection. */
3957 rxi_ConnectionError(conn, error);
3958 MUTEX_ENTER(&conn->conn_data_lock);
3959 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3960 MUTEX_EXIT(&conn->conn_data_lock);
3962 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
3963 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
3969 /* Find an available server process to service the current request in
3970 * the given call structure. If one isn't available, queue up this
3971 * call so it eventually gets one */
3973 rxi_AttachServerProc(register struct rx_call *call,
3974 register osi_socket socket, register int *tnop,
3975 register struct rx_call **newcallp)
3977 register struct rx_serverQueueEntry *sq;
3978 register struct rx_service *service = call->conn->service;
3979 register int haveQuota = 0;
3981 /* May already be attached */
3982 if (call->state == RX_STATE_ACTIVE)
3985 MUTEX_ENTER(&rx_serverPool_lock);
3987 haveQuota = QuotaOK(service);
3988 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
3989 /* If there are no processes available to service this call,
3990 * put the call on the incoming call queue (unless it's
3991 * already on the queue).
3993 #ifdef RX_ENABLE_LOCKS
3995 ReturnToServerPool(service);
3996 #endif /* RX_ENABLE_LOCKS */
3998 if (!(call->flags & RX_CALL_WAIT_PROC)) {
3999 call->flags |= RX_CALL_WAIT_PROC;
4000 MUTEX_ENTER(&rx_stats_mutex);
4003 MUTEX_EXIT(&rx_stats_mutex);
4004 rxi_calltrace(RX_CALL_ARRIVAL, call);
4005 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4006 queue_Append(&rx_incomingCallQueue, call);
4009 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4011 /* If hot threads are enabled, and both newcallp and sq->socketp
4012 * are non-null, then this thread will process the call, and the
4013 * idle server thread will start listening on this threads socket.
4016 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4019 *sq->socketp = socket;
4020 clock_GetTime(&call->startTime);
4021 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4025 if (call->flags & RX_CALL_WAIT_PROC) {
4026 /* Conservative: I don't think this should happen */
4027 call->flags &= ~RX_CALL_WAIT_PROC;
4028 if (queue_IsOnQueue(call)) {
4030 MUTEX_ENTER(&rx_stats_mutex);
4032 MUTEX_EXIT(&rx_stats_mutex);
4035 call->state = RX_STATE_ACTIVE;
4036 call->mode = RX_MODE_RECEIVING;
4037 #ifdef RX_KERNEL_TRACE
4039 int glockOwner = ISAFS_GLOCK();
4042 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4043 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4049 if (call->flags & RX_CALL_CLEARED) {
4050 /* send an ack now to start the packet flow up again */
4051 call->flags &= ~RX_CALL_CLEARED;
4052 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4054 #ifdef RX_ENABLE_LOCKS
4057 service->nRequestsRunning++;
4058 if (service->nRequestsRunning <= service->minProcs)
4064 MUTEX_EXIT(&rx_serverPool_lock);
4067 /* Delay the sending of an acknowledge event for a short while, while
4068 * a new call is being prepared (in the case of a client) or a reply
4069 * is being prepared (in the case of a server). Rather than sending
4070 * an ack packet, an ACKALL packet is sent. */
4072 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4074 #ifdef RX_ENABLE_LOCKS
4076 MUTEX_ENTER(&call->lock);
4077 call->delayedAckEvent = NULL;
4078 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4080 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4081 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4083 MUTEX_EXIT(&call->lock);
4084 #else /* RX_ENABLE_LOCKS */
4086 call->delayedAckEvent = NULL;
4087 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4088 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4089 #endif /* RX_ENABLE_LOCKS */
4093 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4096 #ifdef RX_ENABLE_LOCKS
4098 MUTEX_ENTER(&call->lock);
4099 if (event == call->delayedAckEvent)
4100 call->delayedAckEvent = NULL;
4101 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4103 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4105 MUTEX_EXIT(&call->lock);
4106 #else /* RX_ENABLE_LOCKS */
4108 call->delayedAckEvent = NULL;
4109 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4110 #endif /* RX_ENABLE_LOCKS */
4114 #ifdef RX_ENABLE_LOCKS
4115 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4116 * clearing them out.
4119 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4121 register struct rx_packet *p, *tp;
4124 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4127 p->flags |= RX_PKTFLAG_ACKED;
4131 call->flags |= RX_CALL_TQ_CLEARME;
4132 call->flags |= RX_CALL_TQ_SOME_ACKED;
4135 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4136 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4137 call->tfirst = call->tnext;
4138 call->nSoftAcked = 0;
4140 if (call->flags & RX_CALL_FAST_RECOVER) {
4141 call->flags &= ~RX_CALL_FAST_RECOVER;
4142 call->cwind = call->nextCwind;
4143 call->nextCwind = 0;
4146 CV_SIGNAL(&call->cv_twind);
4148 #endif /* RX_ENABLE_LOCKS */
4150 /* Clear out the transmit queue for the current call (all packets have
4151 * been received by peer) */
4153 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4155 register struct rx_packet *p, *tp;
4157 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4158 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4160 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4163 p->flags |= RX_PKTFLAG_ACKED;
4167 call->flags |= RX_CALL_TQ_CLEARME;
4168 call->flags |= RX_CALL_TQ_SOME_ACKED;
4171 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4172 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4178 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4179 call->flags &= ~RX_CALL_TQ_CLEARME;
4181 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4183 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4184 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4185 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4186 call->nSoftAcked = 0;
4188 if (call->flags & RX_CALL_FAST_RECOVER) {
4189 call->flags &= ~RX_CALL_FAST_RECOVER;
4190 call->cwind = call->nextCwind;
4192 #ifdef RX_ENABLE_LOCKS
4193 CV_SIGNAL(&call->cv_twind);
4195 osi_rxWakeup(&call->twind);
4200 rxi_ClearReceiveQueue(register struct rx_call *call)
4202 register struct rx_packet *p, *tp;
4203 if (queue_IsNotEmpty(&call->rq)) {
4204 for (queue_Scan(&call->rq, p, tp, rx_packet)) {
4209 rx_packetReclaims++;
4211 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4213 if (call->state == RX_STATE_PRECALL) {
4214 call->flags |= RX_CALL_CLEARED;
4218 /* Send an abort packet for the specified call */
4220 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4221 int istack, int force)
4229 /* Clients should never delay abort messages */
4230 if (rx_IsClientConn(call->conn))
4233 if (call->abortCode != call->error) {
4234 call->abortCode = call->error;
4235 call->abortCount = 0;
4238 if (force || rxi_callAbortThreshhold == 0
4239 || call->abortCount < rxi_callAbortThreshhold) {
4240 if (call->delayedAbortEvent) {
4241 rxevent_Cancel(call->delayedAbortEvent, call,
4242 RX_CALL_REFCOUNT_ABORT);
4244 error = htonl(call->error);
4247 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4248 (char *)&error, sizeof(error), istack);
4249 } else if (!call->delayedAbortEvent) {
4250 clock_GetTime(&when);
4251 clock_Addmsec(&when, rxi_callAbortDelay);
4252 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4253 call->delayedAbortEvent =
4254 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4259 /* Send an abort packet for the specified connection. Packet is an
4260 * optional pointer to a packet that can be used to send the abort.
4261 * Once the number of abort messages reaches the threshhold, an
4262 * event is scheduled to send the abort. Setting the force flag
4263 * overrides sending delayed abort messages.
4265 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4266 * to send the abort packet.
4269 rxi_SendConnectionAbort(register struct rx_connection *conn,
4270 struct rx_packet *packet, int istack, int force)
4278 /* Clients should never delay abort messages */
4279 if (rx_IsClientConn(conn))
4282 if (force || rxi_connAbortThreshhold == 0
4283 || conn->abortCount < rxi_connAbortThreshhold) {
4284 if (conn->delayedAbortEvent) {
4285 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4287 error = htonl(conn->error);
4289 MUTEX_EXIT(&conn->conn_data_lock);
4291 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4292 RX_PACKET_TYPE_ABORT, (char *)&error,
4293 sizeof(error), istack);
4294 MUTEX_ENTER(&conn->conn_data_lock);
4295 } else if (!conn->delayedAbortEvent) {
4296 clock_GetTime(&when);
4297 clock_Addmsec(&when, rxi_connAbortDelay);
4298 conn->delayedAbortEvent =
4299 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4304 /* Associate an error all of the calls owned by a connection. Called
4305 * with error non-zero. This is only for really fatal things, like
4306 * bad authentication responses. The connection itself is set in
4307 * error at this point, so that future packets received will be
4310 rxi_ConnectionError(register struct rx_connection *conn,
4311 register afs_int32 error)
4315 MUTEX_ENTER(&conn->conn_data_lock);
4316 if (conn->challengeEvent)
4317 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4318 if (conn->checkReachEvent) {
4319 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4320 conn->checkReachEvent = 0;
4321 conn->flags &= ~RX_CONN_ATTACHWAIT;
4324 MUTEX_EXIT(&conn->conn_data_lock);
4325 for (i = 0; i < RX_MAXCALLS; i++) {
4326 struct rx_call *call = conn->call[i];
4328 MUTEX_ENTER(&call->lock);
4329 rxi_CallError(call, error);
4330 MUTEX_EXIT(&call->lock);
4333 conn->error = error;
4334 MUTEX_ENTER(&rx_stats_mutex);
4335 rx_stats.fatalErrors++;
4336 MUTEX_EXIT(&rx_stats_mutex);
4341 rxi_CallError(register struct rx_call *call, afs_int32 error)
4344 error = call->error;
4345 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4346 if (!(call->flags & RX_CALL_TQ_BUSY)) {
4347 rxi_ResetCall(call, 0);
4350 rxi_ResetCall(call, 0);
4352 call->error = error;
4353 call->mode = RX_MODE_ERROR;
4356 /* Reset various fields in a call structure, and wakeup waiting
4357 * processes. Some fields aren't changed: state & mode are not
4358 * touched (these must be set by the caller), and bufptr, nLeft, and
4359 * nFree are not reset, since these fields are manipulated by
4360 * unprotected macros, and may only be reset by non-interrupting code.
4363 /* this code requires that call->conn be set properly as a pre-condition. */
4364 #endif /* ADAPT_WINDOW */
4367 rxi_ResetCall(register struct rx_call *call, register int newcall)
4370 register struct rx_peer *peer;
4371 struct rx_packet *packet;
4373 /* Notify anyone who is waiting for asynchronous packet arrival */
4374 if (call->arrivalProc) {
4375 (*call->arrivalProc) (call, call->arrivalProcHandle,
4376 (int)call->arrivalProcArg);
4377 call->arrivalProc = (VOID(*)())0;
4380 if (call->delayedAbortEvent) {
4381 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4382 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4384 rxi_SendCallAbort(call, packet, 0, 1);
4385 rxi_FreePacket(packet);
4390 * Update the peer with the congestion information in this call
4391 * so other calls on this connection can pick up where this call
4392 * left off. If the congestion sequence numbers don't match then
4393 * another call experienced a retransmission.
4395 peer = call->conn->peer;
4396 MUTEX_ENTER(&peer->peer_lock);
4398 if (call->congestSeq == peer->congestSeq) {
4399 peer->cwind = MAX(peer->cwind, call->cwind);
4400 peer->MTU = MAX(peer->MTU, call->MTU);
4401 peer->nDgramPackets =
4402 MAX(peer->nDgramPackets, call->nDgramPackets);
4405 call->abortCode = 0;
4406 call->abortCount = 0;
4408 if (peer->maxDgramPackets > 1) {
4409 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4411 call->MTU = peer->MTU;
4413 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4414 call->ssthresh = rx_maxSendWindow;
4415 call->nDgramPackets = peer->nDgramPackets;
4416 call->congestSeq = peer->congestSeq;
4417 MUTEX_EXIT(&peer->peer_lock);
4419 flags = call->flags;
4420 rxi_ClearReceiveQueue(call);
4421 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4422 if (call->flags & RX_CALL_TQ_BUSY) {
4423 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4424 call->flags |= (flags & RX_CALL_TQ_WAIT);
4426 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4428 rxi_ClearTransmitQueue(call, 0);
4429 queue_Init(&call->tq);
4432 queue_Init(&call->rq);
4434 call->rwind = rx_initReceiveWindow;
4435 call->twind = rx_initSendWindow;
4436 call->nSoftAcked = 0;
4437 call->nextCwind = 0;
4440 call->nCwindAcks = 0;
4441 call->nSoftAcks = 0;
4442 call->nHardAcks = 0;
4444 call->tfirst = call->rnext = call->tnext = 1;
4446 call->lastAcked = 0;
4447 call->localStatus = call->remoteStatus = 0;
4449 if (flags & RX_CALL_READER_WAIT) {
4450 #ifdef RX_ENABLE_LOCKS
4451 CV_BROADCAST(&call->cv_rq);
4453 osi_rxWakeup(&call->rq);
4456 if (flags & RX_CALL_WAIT_PACKETS) {
4457 MUTEX_ENTER(&rx_freePktQ_lock);
4458 rxi_PacketsUnWait(); /* XXX */
4459 MUTEX_EXIT(&rx_freePktQ_lock);
4461 #ifdef RX_ENABLE_LOCKS
4462 CV_SIGNAL(&call->cv_twind);
4464 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4465 osi_rxWakeup(&call->twind);
4468 #ifdef RX_ENABLE_LOCKS
4469 /* The following ensures that we don't mess with any queue while some
4470 * other thread might also be doing so. The call_queue_lock field is
4471 * is only modified under the call lock. If the call is in the process
4472 * of being removed from a queue, the call is not locked until the
4473 * the queue lock is dropped and only then is the call_queue_lock field
4474 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4475 * Note that any other routine which removes a call from a queue has to
4476 * obtain the queue lock before examing the queue and removing the call.
4478 if (call->call_queue_lock) {
4479 MUTEX_ENTER(call->call_queue_lock);
4480 if (queue_IsOnQueue(call)) {
4482 if (flags & RX_CALL_WAIT_PROC) {
4483 MUTEX_ENTER(&rx_stats_mutex);
4485 MUTEX_EXIT(&rx_stats_mutex);
4488 MUTEX_EXIT(call->call_queue_lock);
4489 CLEAR_CALL_QUEUE_LOCK(call);
4491 #else /* RX_ENABLE_LOCKS */
4492 if (queue_IsOnQueue(call)) {
4494 if (flags & RX_CALL_WAIT_PROC)
4497 #endif /* RX_ENABLE_LOCKS */
4499 rxi_KeepAliveOff(call);
4500 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4503 /* Send an acknowledge for the indicated packet (seq,serial) of the
4504 * indicated call, for the indicated reason (reason). This
4505 * acknowledge will specifically acknowledge receiving the packet, and
4506 * will also specify which other packets for this call have been
4507 * received. This routine returns the packet that was used to the
4508 * caller. The caller is responsible for freeing it or re-using it.
4509 * This acknowledgement also returns the highest sequence number
4510 * actually read out by the higher level to the sender; the sender
4511 * promises to keep around packets that have not been read by the
4512 * higher level yet (unless, of course, the sender decides to abort
4513 * the call altogether). Any of p, seq, serial, pflags, or reason may
4514 * be set to zero without ill effect. That is, if they are zero, they
4515 * will not convey any information.
4516 * NOW there is a trailer field, after the ack where it will safely be
4517 * ignored by mundanes, which indicates the maximum size packet this
4518 * host can swallow. */
4520 register struct rx_packet *optionalPacket; use to send ack (or null)
4521 int seq; Sequence number of the packet we are acking
4522 int serial; Serial number of the packet
4523 int pflags; Flags field from packet header
4524 int reason; Reason an acknowledge was prompted
4528 rxi_SendAck(register struct rx_call *call,
4529 register struct rx_packet *optionalPacket, int serial, int reason,
4532 struct rx_ackPacket *ap;
4533 register struct rx_packet *rqp;
4534 register struct rx_packet *nxp; /* For queue_Scan */
4535 register struct rx_packet *p;
4540 * Open the receive window once a thread starts reading packets
4542 if (call->rnext > 1) {
4543 call->rwind = rx_maxReceiveWindow;
4546 call->nHardAcks = 0;
4547 call->nSoftAcks = 0;
4548 if (call->rnext > call->lastAcked)
4549 call->lastAcked = call->rnext;
4553 rx_computelen(p, p->length); /* reset length, you never know */
4554 } /* where that's been... */
4555 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4556 /* We won't send the ack, but don't panic. */
4557 return optionalPacket;
4561 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4564 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL)) {
4565 if (!optionalPacket)
4567 return optionalPacket;
4569 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4570 if (rx_Contiguous(p) < templ) {
4571 if (!optionalPacket)
4573 return optionalPacket;
4578 /* MTUXXX failing to send an ack is very serious. We should */
4579 /* try as hard as possible to send even a partial ack; it's */
4580 /* better than nothing. */
4581 ap = (struct rx_ackPacket *)rx_DataOf(p);
4582 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4583 ap->reason = reason;
4585 /* The skew computation used to be bogus, I think it's better now. */
4586 /* We should start paying attention to skew. XXX */
4587 ap->serial = htonl(serial);
4588 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4590 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4591 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4593 /* No fear of running out of ack packet here because there can only be at most
4594 * one window full of unacknowledged packets. The window size must be constrained
4595 * to be less than the maximum ack size, of course. Also, an ack should always
4596 * fit into a single packet -- it should not ever be fragmented. */
4597 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4598 if (!rqp || !call->rq.next
4599 || (rqp->header.seq > (call->rnext + call->rwind))) {
4600 if (!optionalPacket)
4602 rxi_CallError(call, RX_CALL_DEAD);
4603 return optionalPacket;
4606 while (rqp->header.seq > call->rnext + offset)
4607 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4608 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4610 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4611 if (!optionalPacket)
4613 rxi_CallError(call, RX_CALL_DEAD);
4614 return optionalPacket;
4619 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4621 /* these are new for AFS 3.3 */
4622 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4623 templ = htonl(templ);
4624 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4625 templ = htonl(call->conn->peer->ifMTU);
4626 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4627 sizeof(afs_int32), &templ);
4629 /* new for AFS 3.4 */
4630 templ = htonl(call->rwind);
4631 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4632 sizeof(afs_int32), &templ);
4634 /* new for AFS 3.5 */
4635 templ = htonl(call->conn->peer->ifDgramPackets);
4636 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4637 sizeof(afs_int32), &templ);
4639 p->header.serviceId = call->conn->serviceId;
4640 p->header.cid = (call->conn->cid | call->channel);
4641 p->header.callNumber = *call->callNumber;
4643 p->header.securityIndex = call->conn->securityIndex;
4644 p->header.epoch = call->conn->epoch;
4645 p->header.type = RX_PACKET_TYPE_ACK;
4646 p->header.flags = RX_SLOW_START_OK;
4647 if (reason == RX_ACK_PING) {
4648 p->header.flags |= RX_REQUEST_ACK;
4650 clock_GetTime(&call->pingRequestTime);
4653 if (call->conn->type == RX_CLIENT_CONNECTION)
4654 p->header.flags |= RX_CLIENT_INITIATED;
4658 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u",
4659 ap->reason, ntohl(ap->previousPacket),
4660 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4662 for (offset = 0; offset < ap->nAcks; offset++)
4663 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4671 register int i, nbytes = p->length;
4673 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4674 if (nbytes <= p->wirevec[i].iov_len) {
4675 register int savelen, saven;
4677 savelen = p->wirevec[i].iov_len;
4679 p->wirevec[i].iov_len = nbytes;
4681 rxi_Send(call, p, istack);
4682 p->wirevec[i].iov_len = savelen;
4686 nbytes -= p->wirevec[i].iov_len;
4689 MUTEX_ENTER(&rx_stats_mutex);
4690 rx_stats.ackPacketsSent++;
4691 MUTEX_EXIT(&rx_stats_mutex);
4692 if (!optionalPacket)
4694 return optionalPacket; /* Return packet for re-use by caller */
4697 /* Send all of the packets in the list in single datagram */
4699 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4700 int istack, int moreFlag, struct clock *now,
4701 struct clock *retryTime, int resending)
4706 struct rx_connection *conn = call->conn;
4707 struct rx_peer *peer = conn->peer;
4709 MUTEX_ENTER(&peer->peer_lock);
4712 peer->reSends += len;
4713 MUTEX_ENTER(&rx_stats_mutex);
4714 rx_stats.dataPacketsSent += len;
4715 MUTEX_EXIT(&rx_stats_mutex);
4716 MUTEX_EXIT(&peer->peer_lock);
4718 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4722 /* Set the packet flags and schedule the resend events */
4723 /* Only request an ack for the last packet in the list */
4724 for (i = 0; i < len; i++) {
4725 list[i]->retryTime = *retryTime;
4726 if (list[i]->header.serial) {
4727 /* Exponentially backoff retry times */
4728 if (list[i]->backoff < MAXBACKOFF) {
4729 /* so it can't stay == 0 */
4730 list[i]->backoff = (list[i]->backoff << 1) + 1;
4733 clock_Addmsec(&(list[i]->retryTime),
4734 ((afs_uint32) list[i]->backoff) << 8);
4737 /* Wait a little extra for the ack on the last packet */
4738 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4739 clock_Addmsec(&(list[i]->retryTime), 400);
4742 /* Record the time sent */
4743 list[i]->timeSent = *now;
4745 /* Ask for an ack on retransmitted packets, on every other packet
4746 * if the peer doesn't support slow start. Ask for an ack on every
4747 * packet until the congestion window reaches the ack rate. */
4748 if (list[i]->header.serial) {
4750 MUTEX_ENTER(&rx_stats_mutex);
4751 rx_stats.dataPacketsReSent++;
4752 MUTEX_EXIT(&rx_stats_mutex);
4754 /* improved RTO calculation- not Karn */
4755 list[i]->firstSent = *now;
4756 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4757 || (!(call->flags & RX_CALL_SLOW_START_OK)
4758 && (list[i]->header.seq & 1)))) {
4763 MUTEX_ENTER(&peer->peer_lock);
4767 MUTEX_ENTER(&rx_stats_mutex);
4768 rx_stats.dataPacketsSent++;
4769 MUTEX_EXIT(&rx_stats_mutex);
4770 MUTEX_EXIT(&peer->peer_lock);
4772 /* Tag this packet as not being the last in this group,
4773 * for the receiver's benefit */
4774 if (i < len - 1 || moreFlag) {
4775 list[i]->header.flags |= RX_MORE_PACKETS;
4778 /* Install the new retransmit time for the packet, and
4779 * record the time sent */
4780 list[i]->timeSent = *now;
4784 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4787 /* Since we're about to send a data packet to the peer, it's
4788 * safe to nuke any scheduled end-of-packets ack */
4789 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4791 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4792 MUTEX_EXIT(&call->lock);
4794 rxi_SendPacketList(call, conn, list, len, istack);
4796 rxi_SendPacket(call, conn, list[0], istack);
4798 MUTEX_ENTER(&call->lock);
4799 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4801 /* Update last send time for this call (for keep-alive
4802 * processing), and for the connection (so that we can discover
4803 * idle connections) */
4804 conn->lastSendTime = call->lastSendTime = clock_Sec();
4807 /* When sending packets we need to follow these rules:
4808 * 1. Never send more than maxDgramPackets in a jumbogram.
4809 * 2. Never send a packet with more than two iovecs in a jumbogram.
4810 * 3. Never send a retransmitted packet in a jumbogram.
4811 * 4. Never send more than cwind/4 packets in a jumbogram
4812 * We always keep the last list we should have sent so we
4813 * can set the RX_MORE_PACKETS flags correctly.
4816 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4817 int istack, struct clock *now, struct clock *retryTime,
4820 int i, cnt, lastCnt = 0;
4821 struct rx_packet **listP, **lastP = 0;
4822 struct rx_peer *peer = call->conn->peer;
4823 int morePackets = 0;
4825 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4826 /* Does the current packet force us to flush the current list? */
4828 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4829 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4831 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4833 /* If the call enters an error state stop sending, or if
4834 * we entered congestion recovery mode, stop sending */
4835 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4843 /* Add the current packet to the list if it hasn't been acked.
4844 * Otherwise adjust the list pointer to skip the current packet. */
4845 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
4847 /* Do we need to flush the list? */
4848 if (cnt >= (int)peer->maxDgramPackets
4849 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
4850 || list[i]->header.serial
4851 || list[i]->length != RX_JUMBOBUFFERSIZE) {
4853 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
4854 retryTime, resending);
4855 /* If the call enters an error state stop sending, or if
4856 * we entered congestion recovery mode, stop sending */
4858 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4863 listP = &list[i + 1];
4868 osi_Panic("rxi_SendList error");
4870 listP = &list[i + 1];
4874 /* Send the whole list when the call is in receive mode, when
4875 * the call is in eof mode, when we are in fast recovery mode,
4876 * and when we have the last packet */
4877 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
4878 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
4879 || (call->flags & RX_CALL_FAST_RECOVER)) {
4880 /* Check for the case where the current list contains
4881 * an acked packet. Since we always send retransmissions
4882 * in a separate packet, we only need to check the first
4883 * packet in the list */
4884 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
4888 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
4889 retryTime, resending);
4890 /* If the call enters an error state stop sending, or if
4891 * we entered congestion recovery mode, stop sending */
4892 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4896 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
4899 } else if (lastCnt > 0) {
4900 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
4905 #ifdef RX_ENABLE_LOCKS
4906 /* Call rxi_Start, below, but with the call lock held. */
4908 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
4911 MUTEX_ENTER(&call->lock);
4912 rxi_Start(event, call, istack);
4913 MUTEX_EXIT(&call->lock);
4915 #endif /* RX_ENABLE_LOCKS */
4917 /* This routine is called when new packets are readied for
4918 * transmission and when retransmission may be necessary, or when the
4919 * transmission window or burst count are favourable. This should be
4920 * better optimized for new packets, the usual case, now that we've
4921 * got rid of queues of send packets. XXXXXXXXXXX */
4923 rxi_Start(struct rxevent *event, register struct rx_call *call, int istack)
4925 struct rx_packet *p;
4926 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
4927 struct rx_peer *peer = call->conn->peer;
4928 struct clock now, retryTime;
4932 struct rx_packet **xmitList;
4935 /* If rxi_Start is being called as a result of a resend event,
4936 * then make sure that the event pointer is removed from the call
4937 * structure, since there is no longer a per-call retransmission
4939 if (event && event == call->resendEvent) {
4940 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
4941 call->resendEvent = NULL;
4943 if (queue_IsEmpty(&call->tq)) {
4947 /* Timeouts trigger congestion recovery */
4948 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4949 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4950 /* someone else is waiting to start recovery */
4953 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4954 while (call->flags & RX_CALL_TQ_BUSY) {
4955 call->flags |= RX_CALL_TQ_WAIT;
4956 #ifdef RX_ENABLE_LOCKS
4957 CV_WAIT(&call->cv_tq, &call->lock);
4958 #else /* RX_ENABLE_LOCKS */
4959 osi_rxSleep(&call->tq);
4960 #endif /* RX_ENABLE_LOCKS */
4962 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4963 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4964 call->flags |= RX_CALL_FAST_RECOVER;
4965 if (peer->maxDgramPackets > 1) {
4966 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4968 call->MTU = MIN(peer->natMTU, peer->maxMTU);
4970 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4971 call->nDgramPackets = 1;
4973 call->nextCwind = 1;
4976 MUTEX_ENTER(&peer->peer_lock);
4977 peer->MTU = call->MTU;
4978 peer->cwind = call->cwind;
4979 peer->nDgramPackets = 1;
4981 call->congestSeq = peer->congestSeq;
4982 MUTEX_EXIT(&peer->peer_lock);
4983 /* Clear retry times on packets. Otherwise, it's possible for
4984 * some packets in the queue to force resends at rates faster
4985 * than recovery rates.
4987 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
4988 if (!(p->flags & RX_PKTFLAG_ACKED)) {
4989 clock_Zero(&p->retryTime);
4994 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4995 MUTEX_ENTER(&rx_stats_mutex);
4996 rx_tq_debug.rxi_start_in_error++;
4997 MUTEX_EXIT(&rx_stats_mutex);
5002 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5003 /* Get clock to compute the re-transmit time for any packets
5004 * in this burst. Note, if we back off, it's reasonable to
5005 * back off all of the packets in the same manner, even if
5006 * some of them have been retransmitted more times than more
5007 * recent additions */
5008 clock_GetTime(&now);
5009 retryTime = now; /* initialize before use */
5010 MUTEX_ENTER(&peer->peer_lock);
5011 clock_Add(&retryTime, &peer->timeout);
5012 MUTEX_EXIT(&peer->peer_lock);
5014 /* Send (or resend) any packets that need it, subject to
5015 * window restrictions and congestion burst control
5016 * restrictions. Ask for an ack on the last packet sent in
5017 * this burst. For now, we're relying upon the window being
5018 * considerably bigger than the largest number of packets that
5019 * are typically sent at once by one initial call to
5020 * rxi_Start. This is probably bogus (perhaps we should ask
5021 * for an ack when we're half way through the current
5022 * window?). Also, for non file transfer applications, this
5023 * may end up asking for an ack for every packet. Bogus. XXXX
5026 * But check whether we're here recursively, and let the other guy
5029 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5030 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5031 call->flags |= RX_CALL_TQ_BUSY;
5033 call->flags &= ~RX_CALL_NEED_START;
5034 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5036 maxXmitPackets = MIN(call->twind, call->cwind);
5037 xmitList = (struct rx_packet **)
5038 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5039 if (xmitList == NULL)
5040 osi_Panic("rxi_Start, failed to allocate xmit list");
5041 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5042 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5043 /* We shouldn't be sending packets if a thread is waiting
5044 * to initiate congestion recovery */
5048 && (call->flags & RX_CALL_FAST_RECOVER)) {
5049 /* Only send one packet during fast recovery */
5052 if ((p->flags & RX_PKTFLAG_FREE)
5053 || (!queue_IsEnd(&call->tq, nxp)
5054 && (nxp->flags & RX_PKTFLAG_FREE))
5055 || (p == (struct rx_packet *)&rx_freePacketQueue)
5056 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5057 osi_Panic("rxi_Start: xmit queue clobbered");
5059 if (p->flags & RX_PKTFLAG_ACKED) {
5060 MUTEX_ENTER(&rx_stats_mutex);
5061 rx_stats.ignoreAckedPacket++;
5062 MUTEX_EXIT(&rx_stats_mutex);
5063 continue; /* Ignore this packet if it has been acknowledged */
5066 /* Turn off all flags except these ones, which are the same
5067 * on each transmission */
5068 p->header.flags &= RX_PRESET_FLAGS;
5070 if (p->header.seq >=
5071 call->tfirst + MIN((int)call->twind,
5072 (int)(call->nSoftAcked +
5074 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5075 /* Note: if we're waiting for more window space, we can
5076 * still send retransmits; hence we don't return here, but
5077 * break out to schedule a retransmit event */
5078 dpf(("call %d waiting for window",
5079 *(call->callNumber)));
5083 /* Transmit the packet if it needs to be sent. */
5084 if (!clock_Lt(&now, &p->retryTime)) {
5085 if (nXmitPackets == maxXmitPackets) {
5086 osi_Panic("rxi_Start: xmit list overflowed");
5088 xmitList[nXmitPackets++] = p;
5092 /* xmitList now hold pointers to all of the packets that are
5093 * ready to send. Now we loop to send the packets */
5094 if (nXmitPackets > 0) {
5095 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5096 &now, &retryTime, resending);
5099 maxXmitPackets * sizeof(struct rx_packet *));
5101 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5103 * TQ references no longer protected by this flag; they must remain
5104 * protected by the global lock.
5106 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5107 call->flags &= ~RX_CALL_TQ_BUSY;
5108 if (call->flags & RX_CALL_TQ_WAIT) {
5109 call->flags &= ~RX_CALL_TQ_WAIT;
5110 #ifdef RX_ENABLE_LOCKS
5111 CV_BROADCAST(&call->cv_tq);
5112 #else /* RX_ENABLE_LOCKS */
5113 osi_rxWakeup(&call->tq);
5114 #endif /* RX_ENABLE_LOCKS */
5119 /* We went into the error state while sending packets. Now is
5120 * the time to reset the call. This will also inform the using
5121 * process that the call is in an error state.
5123 MUTEX_ENTER(&rx_stats_mutex);
5124 rx_tq_debug.rxi_start_aborted++;
5125 MUTEX_EXIT(&rx_stats_mutex);
5126 call->flags &= ~RX_CALL_TQ_BUSY;
5127 if (call->flags & RX_CALL_TQ_WAIT) {
5128 call->flags &= ~RX_CALL_TQ_WAIT;
5129 #ifdef RX_ENABLE_LOCKS
5130 CV_BROADCAST(&call->cv_tq);
5131 #else /* RX_ENABLE_LOCKS */
5132 osi_rxWakeup(&call->tq);
5133 #endif /* RX_ENABLE_LOCKS */
5135 rxi_CallError(call, call->error);
5138 #ifdef RX_ENABLE_LOCKS
5139 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5140 register int missing;
5141 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5142 /* Some packets have received acks. If they all have, we can clear
5143 * the transmit queue.
5146 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5147 if (p->header.seq < call->tfirst
5148 && (p->flags & RX_PKTFLAG_ACKED)) {
5155 call->flags |= RX_CALL_TQ_CLEARME;
5157 #endif /* RX_ENABLE_LOCKS */
5158 /* Don't bother doing retransmits if the TQ is cleared. */
5159 if (call->flags & RX_CALL_TQ_CLEARME) {
5160 rxi_ClearTransmitQueue(call, 1);
5162 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5165 /* Always post a resend event, if there is anything in the
5166 * queue, and resend is possible. There should be at least
5167 * one unacknowledged packet in the queue ... otherwise none
5168 * of these packets should be on the queue in the first place.
5170 if (call->resendEvent) {
5171 /* Cancel the existing event and post a new one */
5172 rxevent_Cancel(call->resendEvent, call,
5173 RX_CALL_REFCOUNT_RESEND);
5176 /* The retry time is the retry time on the first unacknowledged
5177 * packet inside the current window */
5179 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5180 /* Don't set timers for packets outside the window */
5181 if (p->header.seq >= call->tfirst + call->twind) {
5185 if (!(p->flags & RX_PKTFLAG_ACKED)
5186 && !clock_IsZero(&p->retryTime)) {
5188 retryTime = p->retryTime;
5193 /* Post a new event to re-run rxi_Start when retries may be needed */
5194 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5195 #ifdef RX_ENABLE_LOCKS
5196 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5198 rxevent_Post(&retryTime, rxi_StartUnlocked,
5199 (void *)call, (void *)istack);
5200 #else /* RX_ENABLE_LOCKS */
5202 rxevent_Post(&retryTime, rxi_Start, (void *)call,
5204 #endif /* RX_ENABLE_LOCKS */
5207 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5208 } while (call->flags & RX_CALL_NEED_START);
5210 * TQ references no longer protected by this flag; they must remain
5211 * protected by the global lock.
5213 call->flags &= ~RX_CALL_TQ_BUSY;
5214 if (call->flags & RX_CALL_TQ_WAIT) {
5215 call->flags &= ~RX_CALL_TQ_WAIT;
5216 #ifdef RX_ENABLE_LOCKS
5217 CV_BROADCAST(&call->cv_tq);
5218 #else /* RX_ENABLE_LOCKS */
5219 osi_rxWakeup(&call->tq);
5220 #endif /* RX_ENABLE_LOCKS */
5223 call->flags |= RX_CALL_NEED_START;
5225 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5227 if (call->resendEvent) {
5228 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5233 /* Also adjusts the keep alive parameters for the call, to reflect
5234 * that we have just sent a packet (so keep alives aren't sent
5237 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5240 register struct rx_connection *conn = call->conn;
5242 /* Stamp each packet with the user supplied status */
5243 p->header.userStatus = call->localStatus;
5245 /* Allow the security object controlling this call's security to
5246 * make any last-minute changes to the packet */
5247 RXS_SendPacket(conn->securityObject, call, p);
5249 /* Since we're about to send SOME sort of packet to the peer, it's
5250 * safe to nuke any scheduled end-of-packets ack */
5251 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5253 /* Actually send the packet, filling in more connection-specific fields */
5254 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5255 MUTEX_EXIT(&call->lock);
5256 rxi_SendPacket(call, conn, p, istack);
5257 MUTEX_ENTER(&call->lock);
5258 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5260 /* Update last send time for this call (for keep-alive
5261 * processing), and for the connection (so that we can discover
5262 * idle connections) */
5263 conn->lastSendTime = call->lastSendTime = clock_Sec();
5267 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5268 * that things are fine. Also called periodically to guarantee that nothing
5269 * falls through the cracks (e.g. (error + dally) connections have keepalive
5270 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5272 * haveCTLock Set if calling from rxi_ReapConnections
5274 #ifdef RX_ENABLE_LOCKS
5276 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5277 #else /* RX_ENABLE_LOCKS */
5279 rxi_CheckCall(register struct rx_call *call)
5280 #endif /* RX_ENABLE_LOCKS */
5282 register struct rx_connection *conn = call->conn;
5284 afs_uint32 deadTime;
5286 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5287 if (call->flags & RX_CALL_TQ_BUSY) {
5288 /* Call is active and will be reset by rxi_Start if it's
5289 * in an error state.
5294 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5296 (((afs_uint32) conn->secondsUntilDead << 10) +
5297 ((afs_uint32) conn->peer->rtt >> 3) +
5298 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5300 /* These are computed to the second (+- 1 second). But that's
5301 * good enough for these values, which should be a significant
5302 * number of seconds. */
5303 if (now > (call->lastReceiveTime + deadTime)) {
5304 if (call->state == RX_STATE_ACTIVE) {
5305 rxi_CallError(call, RX_CALL_DEAD);
5308 #ifdef RX_ENABLE_LOCKS
5309 /* Cancel pending events */
5310 rxevent_Cancel(call->delayedAckEvent, call,
5311 RX_CALL_REFCOUNT_DELAY);
5312 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5313 rxevent_Cancel(call->keepAliveEvent, call,
5314 RX_CALL_REFCOUNT_ALIVE);
5315 if (call->refCount == 0) {
5316 rxi_FreeCall(call, haveCTLock);
5320 #else /* RX_ENABLE_LOCKS */
5323 #endif /* RX_ENABLE_LOCKS */
5325 /* Non-active calls are destroyed if they are not responding
5326 * to pings; active calls are simply flagged in error, so the
5327 * attached process can die reasonably gracefully. */
5329 /* see if we have a non-activity timeout */
5330 if (call->startWait && conn->idleDeadTime
5331 && ((call->startWait + conn->idleDeadTime) < now)) {
5332 if (call->state == RX_STATE_ACTIVE) {
5333 rxi_CallError(call, RX_CALL_TIMEOUT);
5337 /* see if we have a hard timeout */
5338 if (conn->hardDeadTime
5339 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5340 if (call->state == RX_STATE_ACTIVE)
5341 rxi_CallError(call, RX_CALL_TIMEOUT);
5348 /* When a call is in progress, this routine is called occasionally to
5349 * make sure that some traffic has arrived (or been sent to) the peer.
5350 * If nothing has arrived in a reasonable amount of time, the call is
5351 * declared dead; if nothing has been sent for a while, we send a
5352 * keep-alive packet (if we're actually trying to keep the call alive)
5355 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5358 struct rx_connection *conn;
5361 MUTEX_ENTER(&call->lock);
5362 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5363 if (event == call->keepAliveEvent)
5364 call->keepAliveEvent = NULL;
5367 #ifdef RX_ENABLE_LOCKS
5368 if (rxi_CheckCall(call, 0)) {
5369 MUTEX_EXIT(&call->lock);
5372 #else /* RX_ENABLE_LOCKS */
5373 if (rxi_CheckCall(call))
5375 #endif /* RX_ENABLE_LOCKS */
5377 /* Don't try to keep alive dallying calls */
5378 if (call->state == RX_STATE_DALLY) {
5379 MUTEX_EXIT(&call->lock);
5384 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5385 /* Don't try to send keepalives if there is unacknowledged data */
5386 /* the rexmit code should be good enough, this little hack
5387 * doesn't quite work XXX */
5388 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5390 rxi_ScheduleKeepAliveEvent(call);
5391 MUTEX_EXIT(&call->lock);
5396 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5398 if (!call->keepAliveEvent) {
5400 clock_GetTime(&when);
5401 when.sec += call->conn->secondsUntilPing;
5402 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5403 call->keepAliveEvent =
5404 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5408 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5410 rxi_KeepAliveOn(register struct rx_call *call)
5412 /* Pretend last packet received was received now--i.e. if another
5413 * packet isn't received within the keep alive time, then the call
5414 * will die; Initialize last send time to the current time--even
5415 * if a packet hasn't been sent yet. This will guarantee that a
5416 * keep-alive is sent within the ping time */
5417 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5418 rxi_ScheduleKeepAliveEvent(call);
5421 /* This routine is called to send connection abort messages
5422 * that have been delayed to throttle looping clients. */
5424 rxi_SendDelayedConnAbort(struct rxevent *event,
5425 register struct rx_connection *conn, char *dummy)
5428 struct rx_packet *packet;
5430 MUTEX_ENTER(&conn->conn_data_lock);
5431 conn->delayedAbortEvent = NULL;
5432 error = htonl(conn->error);
5434 MUTEX_EXIT(&conn->conn_data_lock);
5435 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5438 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5439 RX_PACKET_TYPE_ABORT, (char *)&error,
5441 rxi_FreePacket(packet);
5445 /* This routine is called to send call abort messages
5446 * that have been delayed to throttle looping clients. */
5448 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5452 struct rx_packet *packet;
5454 MUTEX_ENTER(&call->lock);
5455 call->delayedAbortEvent = NULL;
5456 error = htonl(call->error);
5458 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5461 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5462 (char *)&error, sizeof(error), 0);
5463 rxi_FreePacket(packet);
5465 MUTEX_EXIT(&call->lock);
5468 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5469 * seconds) to ask the client to authenticate itself. The routine
5470 * issues a challenge to the client, which is obtained from the
5471 * security object associated with the connection */
5473 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5476 int tries = (int)atries;
5477 conn->challengeEvent = NULL;
5478 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5479 register struct rx_packet *packet;
5483 /* We've failed to authenticate for too long.
5484 * Reset any calls waiting for authentication;
5485 * they are all in RX_STATE_PRECALL.
5489 MUTEX_ENTER(&conn->conn_call_lock);
5490 for (i = 0; i < RX_MAXCALLS; i++) {
5491 struct rx_call *call = conn->call[i];
5493 MUTEX_ENTER(&call->lock);
5494 if (call->state == RX_STATE_PRECALL) {
5495 rxi_CallError(call, RX_CALL_DEAD);
5496 rxi_SendCallAbort(call, NULL, 0, 0);
5498 MUTEX_EXIT(&call->lock);
5501 MUTEX_EXIT(&conn->conn_call_lock);
5505 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5507 /* If there's no packet available, do this later. */
5508 RXS_GetChallenge(conn->securityObject, conn, packet);
5509 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5510 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5511 rxi_FreePacket(packet);
5513 clock_GetTime(&when);
5514 when.sec += RX_CHALLENGE_TIMEOUT;
5515 conn->challengeEvent =
5516 rxevent_Post(&when, rxi_ChallengeEvent, conn,
5517 (void *)(tries - 1));
5521 /* Call this routine to start requesting the client to authenticate
5522 * itself. This will continue until authentication is established,
5523 * the call times out, or an invalid response is returned. The
5524 * security object associated with the connection is asked to create
5525 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5526 * defined earlier. */
5528 rxi_ChallengeOn(register struct rx_connection *conn)
5530 if (!conn->challengeEvent) {
5531 RXS_CreateChallenge(conn->securityObject, conn);
5532 rxi_ChallengeEvent(NULL, conn, (void *)RX_CHALLENGE_MAXTRIES);
5537 /* Compute round trip time of the packet provided, in *rttp.
5540 /* rxi_ComputeRoundTripTime is called with peer locked. */
5541 /* sentp and/or peer may be null */
5543 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5544 register struct clock *sentp,
5545 register struct rx_peer *peer)
5547 struct clock thisRtt, *rttp = &thisRtt;
5549 #if defined(AFS_ALPHA_LINUX22_ENV) && defined(AFS_PTHREAD_ENV) && !defined(KERNEL)
5550 /* making year 2038 bugs to get this running now - stroucki */
5551 struct timeval temptime;
5553 register int rtt_timeout;
5555 #if defined(AFS_ALPHA_LINUX20_ENV) && defined(AFS_PTHREAD_ENV) && !defined(KERNEL)
5556 /* yet again. This was the worst Heisenbug of the port - stroucki */
5557 clock_GetTime(&temptime);
5558 rttp->sec = (afs_int32) temptime.tv_sec;
5559 rttp->usec = (afs_int32) temptime.tv_usec;
5561 clock_GetTime(rttp);
5563 if (clock_Lt(rttp, sentp)) {
5565 return; /* somebody set the clock back, don't count this time. */
5567 clock_Sub(rttp, sentp);
5568 MUTEX_ENTER(&rx_stats_mutex);
5569 if (clock_Lt(rttp, &rx_stats.minRtt))
5570 rx_stats.minRtt = *rttp;
5571 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5572 if (rttp->sec > 60) {
5573 MUTEX_EXIT(&rx_stats_mutex);
5574 return; /* somebody set the clock ahead */
5576 rx_stats.maxRtt = *rttp;
5578 clock_Add(&rx_stats.totalRtt, rttp);
5579 rx_stats.nRttSamples++;
5580 MUTEX_EXIT(&rx_stats_mutex);
5582 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5584 /* Apply VanJacobson round-trip estimations */
5589 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5590 * srtt is stored as fixed point with 3 bits after the binary
5591 * point (i.e., scaled by 8). The following magic is
5592 * equivalent to the smoothing algorithm in rfc793 with an
5593 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5594 * srtt*8 = srtt*8 + rtt - srtt
5595 * srtt = srtt + rtt/8 - srtt/8
5598 delta = MSEC(rttp) - (peer->rtt >> 3);
5602 * We accumulate a smoothed rtt variance (actually, a smoothed
5603 * mean difference), then set the retransmit timer to smoothed
5604 * rtt + 4 times the smoothed variance (was 2x in van's original
5605 * paper, but 4x works better for me, and apparently for him as
5607 * rttvar is stored as
5608 * fixed point with 2 bits after the binary point (scaled by
5609 * 4). The following is equivalent to rfc793 smoothing with
5610 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5611 * replaces rfc793's wired-in beta.
5612 * dev*4 = dev*4 + (|actual - expected| - dev)
5618 delta -= (peer->rtt_dev >> 2);
5619 peer->rtt_dev += delta;
5621 /* I don't have a stored RTT so I start with this value. Since I'm
5622 * probably just starting a call, and will be pushing more data down
5623 * this, I expect congestion to increase rapidly. So I fudge a
5624 * little, and I set deviance to half the rtt. In practice,
5625 * deviance tends to approach something a little less than
5626 * half the smoothed rtt. */
5627 peer->rtt = (MSEC(rttp) << 3) + 8;
5628 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5630 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5631 * the other of these connections is usually in a user process, and can
5632 * be switched and/or swapped out. So on fast, reliable networks, the
5633 * timeout would otherwise be too short.
5635 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5636 clock_Zero(&(peer->timeout));
5637 clock_Addmsec(&(peer->timeout), rtt_timeout);
5639 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)));
5643 /* Find all server connections that have not been active for a long time, and
5646 rxi_ReapConnections(void)
5649 clock_GetTime(&now);
5651 /* Find server connection structures that haven't been used for
5652 * greater than rx_idleConnectionTime */
5654 struct rx_connection **conn_ptr, **conn_end;
5655 int i, havecalls = 0;
5656 MUTEX_ENTER(&rx_connHashTable_lock);
5657 for (conn_ptr = &rx_connHashTable[0], conn_end =
5658 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5660 struct rx_connection *conn, *next;
5661 struct rx_call *call;
5665 for (conn = *conn_ptr; conn; conn = next) {
5666 /* XXX -- Shouldn't the connection be locked? */
5669 for (i = 0; i < RX_MAXCALLS; i++) {
5670 call = conn->call[i];
5673 MUTEX_ENTER(&call->lock);
5674 #ifdef RX_ENABLE_LOCKS
5675 result = rxi_CheckCall(call, 1);
5676 #else /* RX_ENABLE_LOCKS */
5677 result = rxi_CheckCall(call);
5678 #endif /* RX_ENABLE_LOCKS */
5679 MUTEX_EXIT(&call->lock);
5681 /* If CheckCall freed the call, it might
5682 * have destroyed the connection as well,
5683 * which screws up the linked lists.
5689 if (conn->type == RX_SERVER_CONNECTION) {
5690 /* This only actually destroys the connection if
5691 * there are no outstanding calls */
5692 MUTEX_ENTER(&conn->conn_data_lock);
5693 if (!havecalls && !conn->refCount
5694 && ((conn->lastSendTime + rx_idleConnectionTime) <
5696 conn->refCount++; /* it will be decr in rx_DestroyConn */
5697 MUTEX_EXIT(&conn->conn_data_lock);
5698 #ifdef RX_ENABLE_LOCKS
5699 rxi_DestroyConnectionNoLock(conn);
5700 #else /* RX_ENABLE_LOCKS */
5701 rxi_DestroyConnection(conn);
5702 #endif /* RX_ENABLE_LOCKS */
5704 #ifdef RX_ENABLE_LOCKS
5706 MUTEX_EXIT(&conn->conn_data_lock);
5708 #endif /* RX_ENABLE_LOCKS */
5712 #ifdef RX_ENABLE_LOCKS
5713 while (rx_connCleanup_list) {
5714 struct rx_connection *conn;
5715 conn = rx_connCleanup_list;
5716 rx_connCleanup_list = rx_connCleanup_list->next;
5717 MUTEX_EXIT(&rx_connHashTable_lock);
5718 rxi_CleanupConnection(conn);
5719 MUTEX_ENTER(&rx_connHashTable_lock);
5721 MUTEX_EXIT(&rx_connHashTable_lock);
5722 #endif /* RX_ENABLE_LOCKS */
5725 /* Find any peer structures that haven't been used (haven't had an
5726 * associated connection) for greater than rx_idlePeerTime */
5728 struct rx_peer **peer_ptr, **peer_end;
5730 MUTEX_ENTER(&rx_rpc_stats);
5731 MUTEX_ENTER(&rx_peerHashTable_lock);
5732 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5733 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5735 struct rx_peer *peer, *next, *prev;
5736 for (prev = peer = *peer_ptr; peer; peer = next) {
5738 code = MUTEX_TRYENTER(&peer->peer_lock);
5739 if ((code) && (peer->refCount == 0)
5740 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5741 rx_interface_stat_p rpc_stat, nrpc_stat;
5743 MUTEX_EXIT(&peer->peer_lock);
5744 MUTEX_DESTROY(&peer->peer_lock);
5746 (&peer->rpcStats, rpc_stat, nrpc_stat,
5747 rx_interface_stat)) {
5748 unsigned int num_funcs;
5751 queue_Remove(&rpc_stat->queue_header);
5752 queue_Remove(&rpc_stat->all_peers);
5753 num_funcs = rpc_stat->stats[0].func_total;
5755 sizeof(rx_interface_stat_t) +
5756 rpc_stat->stats[0].func_total *
5757 sizeof(rx_function_entry_v1_t);
5759 rxi_Free(rpc_stat, space);
5760 rxi_rpc_peer_stat_cnt -= num_funcs;
5763 MUTEX_ENTER(&rx_stats_mutex);
5764 rx_stats.nPeerStructs--;
5765 MUTEX_EXIT(&rx_stats_mutex);
5766 if (peer == *peer_ptr) {
5773 MUTEX_EXIT(&peer->peer_lock);
5779 MUTEX_EXIT(&rx_peerHashTable_lock);
5780 MUTEX_EXIT(&rx_rpc_stats);
5783 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5784 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5785 * GC, just below. Really, we shouldn't have to keep moving packets from
5786 * one place to another, but instead ought to always know if we can
5787 * afford to hold onto a packet in its particular use. */
5788 MUTEX_ENTER(&rx_freePktQ_lock);
5789 if (rx_waitingForPackets) {
5790 rx_waitingForPackets = 0;
5791 #ifdef RX_ENABLE_LOCKS
5792 CV_BROADCAST(&rx_waitingForPackets_cv);
5794 osi_rxWakeup(&rx_waitingForPackets);
5797 MUTEX_EXIT(&rx_freePktQ_lock);
5799 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5800 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5804 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5805 * rx.h is sort of strange this is better. This is called with a security
5806 * object before it is discarded. Each connection using a security object has
5807 * its own refcount to the object so it won't actually be freed until the last
5808 * connection is destroyed.
5810 * This is the only rxs module call. A hold could also be written but no one
5814 rxs_Release(struct rx_securityClass *aobj)
5816 return RXS_Close(aobj);
5820 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5821 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5822 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5823 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5825 /* Adjust our estimate of the transmission rate to this peer, given
5826 * that the packet p was just acked. We can adjust peer->timeout and
5827 * call->twind. Pragmatically, this is called
5828 * only with packets of maximal length.
5829 * Called with peer and call locked.
5833 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5834 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5836 afs_int32 xferSize, xferMs;
5837 register afs_int32 minTime;
5840 /* Count down packets */
5841 if (peer->rateFlag > 0)
5843 /* Do nothing until we're enabled */
5844 if (peer->rateFlag != 0)
5849 /* Count only when the ack seems legitimate */
5850 switch (ackReason) {
5851 case RX_ACK_REQUESTED:
5853 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
5857 case RX_ACK_PING_RESPONSE:
5858 if (p) /* want the response to ping-request, not data send */
5860 clock_GetTime(&newTO);
5861 if (clock_Gt(&newTO, &call->pingRequestTime)) {
5862 clock_Sub(&newTO, &call->pingRequestTime);
5863 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
5867 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
5874 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));
5876 /* Track only packets that are big enough. */
5877 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
5881 /* absorb RTT data (in milliseconds) for these big packets */
5882 if (peer->smRtt == 0) {
5883 peer->smRtt = xferMs;
5885 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
5890 if (peer->countDown) {
5894 peer->countDown = 10; /* recalculate only every so often */
5896 /* In practice, we can measure only the RTT for full packets,
5897 * because of the way Rx acks the data that it receives. (If it's
5898 * smaller than a full packet, it often gets implicitly acked
5899 * either by the call response (from a server) or by the next call
5900 * (from a client), and either case confuses transmission times
5901 * with processing times.) Therefore, replace the above
5902 * more-sophisticated processing with a simpler version, where the
5903 * smoothed RTT is kept for full-size packets, and the time to
5904 * transmit a windowful of full-size packets is simply RTT *
5905 * windowSize. Again, we take two steps:
5906 - ensure the timeout is large enough for a single packet's RTT;
5907 - ensure that the window is small enough to fit in the desired timeout.*/
5909 /* First, the timeout check. */
5910 minTime = peer->smRtt;
5911 /* Get a reasonable estimate for a timeout period */
5913 newTO.sec = minTime / 1000;
5914 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
5916 /* Increase the timeout period so that we can always do at least
5917 * one packet exchange */
5918 if (clock_Gt(&newTO, &peer->timeout)) {
5920 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));
5922 peer->timeout = newTO;
5925 /* Now, get an estimate for the transmit window size. */
5926 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
5927 /* Now, convert to the number of full packets that could fit in a
5928 * reasonable fraction of that interval */
5929 minTime /= (peer->smRtt << 1);
5930 xferSize = minTime; /* (make a copy) */
5932 /* Now clamp the size to reasonable bounds. */
5935 else if (minTime > rx_Window)
5936 minTime = rx_Window;
5937 /* if (minTime != peer->maxWindow) {
5938 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
5939 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
5940 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
5942 peer->maxWindow = minTime;
5943 elide... call->twind = minTime;
5947 /* Cut back on the peer timeout if it had earlier grown unreasonably.
5948 * Discern this by calculating the timeout necessary for rx_Window
5950 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
5951 /* calculate estimate for transmission interval in milliseconds */
5952 minTime = rx_Window * peer->smRtt;
5953 if (minTime < 1000) {
5954 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
5955 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
5956 peer->timeout.usec, peer->smRtt, peer->packetSize));
5958 newTO.sec = 0; /* cut back on timeout by half a second */
5959 newTO.usec = 500000;
5960 clock_Sub(&peer->timeout, &newTO);
5965 } /* end of rxi_ComputeRate */
5966 #endif /* ADAPT_WINDOW */
5974 /* Don't call this debugging routine directly; use dpf */
5976 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
5977 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
5981 clock_GetTime(&now);
5982 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
5983 (unsigned int)now.usec / 1000);
5984 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
5992 * This function is used to process the rx_stats structure that is local
5993 * to a process as well as an rx_stats structure received from a remote
5994 * process (via rxdebug). Therefore, it needs to do minimal version
5998 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
5999 afs_int32 freePackets, char version)
6003 if (size != sizeof(struct rx_stats)) {
6005 "Unexpected size of stats structure: was %d, expected %d\n",
6006 size, sizeof(struct rx_stats));
6009 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6012 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6013 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6014 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6015 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6016 s->specialPktAllocFailures);
6018 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6019 s->receivePktAllocFailures, s->sendPktAllocFailures,
6020 s->specialPktAllocFailures);
6024 " greedy %d, " "bogusReads %d (last from host %x), "
6025 "noPackets %d, " "noBuffers %d, " "selects %d, "
6026 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6027 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6028 s->selects, s->sendSelects);
6030 fprintf(file, " packets read: ");
6031 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6032 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6034 fprintf(file, "\n");
6037 " other read counters: data %d, " "ack %d, " "dup %d "
6038 "spurious %d " "dally %d\n", s->dataPacketsRead,
6039 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6040 s->ignorePacketDally);
6042 fprintf(file, " packets sent: ");
6043 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6044 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6046 fprintf(file, "\n");
6049 " other send counters: ack %d, " "data %d (not resends), "
6050 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6051 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6052 s->dataPacketsPushed, s->ignoreAckedPacket);
6055 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6056 s->netSendFailures, (int)s->fatalErrors);
6058 if (s->nRttSamples) {
6059 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6060 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6062 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6063 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6067 " %d server connections, " "%d client connections, "
6068 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6069 s->nServerConns, s->nClientConns, s->nPeerStructs,
6070 s->nCallStructs, s->nFreeCallStructs);
6072 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6073 fprintf(file, " %d clock updates\n", clock_nUpdates);
6078 /* for backward compatibility */
6080 rx_PrintStats(FILE * file)
6082 MUTEX_ENTER(&rx_stats_mutex);
6083 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6085 MUTEX_EXIT(&rx_stats_mutex);
6089 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6091 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6092 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6093 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6096 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6097 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6098 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6101 " Packet size %d, " "max in packet skew %d, "
6102 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6103 (int)peer->outPacketSkew);
6106 #ifdef AFS_PTHREAD_ENV
6108 * This mutex protects the following static variables:
6112 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6113 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6115 #define LOCK_RX_DEBUG
6116 #define UNLOCK_RX_DEBUG
6117 #endif /* AFS_PTHREAD_ENV */
6120 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6121 u_char type, void *inputData, size_t inputLength,
6122 void *outputData, size_t outputLength)
6124 static afs_int32 counter = 100;
6126 struct rx_header theader;
6128 register afs_int32 code;
6130 struct sockaddr_in taddr, faddr;
6135 endTime = time(0) + 20; /* try for 20 seconds */
6139 tp = &tbuffer[sizeof(struct rx_header)];
6140 taddr.sin_family = AF_INET;
6141 taddr.sin_port = remotePort;
6142 taddr.sin_addr.s_addr = remoteAddr;
6143 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6144 taddr.sin_len = sizeof(struct sockaddr_in);
6147 memset(&theader, 0, sizeof(theader));
6148 theader.epoch = htonl(999);
6150 theader.callNumber = htonl(counter);
6153 theader.type = type;
6154 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6155 theader.serviceId = 0;
6157 memcpy(tbuffer, &theader, sizeof(theader));
6158 memcpy(tp, inputData, inputLength);
6160 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6161 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6163 /* see if there's a packet available */
6165 FD_SET(socket, &imask);
6168 code = select(socket + 1, &imask, 0, 0, &tv);
6169 if (code == 1 && FD_ISSET(socket, &imask)) {
6170 /* now receive a packet */
6171 faddrLen = sizeof(struct sockaddr_in);
6173 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6174 (struct sockaddr *)&faddr, &faddrLen);
6177 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6178 if (counter == ntohl(theader.callNumber))
6183 /* see if we've timed out */
6184 if (endTime < time(0))
6187 code -= sizeof(struct rx_header);
6188 if (code > outputLength)
6189 code = outputLength;
6190 memcpy(outputData, tp, code);
6195 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6196 afs_uint16 remotePort, struct rx_debugStats * stat,
6197 afs_uint32 * supportedValues)
6199 struct rx_debugIn in;
6202 *supportedValues = 0;
6203 in.type = htonl(RX_DEBUGI_GETSTATS);
6206 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6207 &in, sizeof(in), stat, sizeof(*stat));
6210 * If the call was successful, fixup the version and indicate
6211 * what contents of the stat structure are valid.
6212 * Also do net to host conversion of fields here.
6216 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6217 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6219 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6220 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6222 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6223 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6225 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6226 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6228 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6229 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6231 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6232 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6234 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6235 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6237 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6238 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6241 stat->nFreePackets = ntohl(stat->nFreePackets);
6242 stat->packetReclaims = ntohl(stat->packetReclaims);
6243 stat->callsExecuted = ntohl(stat->callsExecuted);
6244 stat->nWaiting = ntohl(stat->nWaiting);
6245 stat->idleThreads = ntohl(stat->idleThreads);
6252 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6253 afs_uint16 remotePort, struct rx_stats * stat,
6254 afs_uint32 * supportedValues)
6256 struct rx_debugIn in;
6257 afs_int32 *lp = (afs_int32 *) stat;
6262 * supportedValues is currently unused, but added to allow future
6263 * versioning of this function.
6266 *supportedValues = 0;
6267 in.type = htonl(RX_DEBUGI_RXSTATS);
6269 memset(stat, 0, sizeof(*stat));
6271 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6272 &in, sizeof(in), stat, sizeof(*stat));
6277 * Do net to host conversion here
6280 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6289 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6290 afs_uint16 remotePort, size_t version_length,
6294 return MakeDebugCall(socket, remoteAddr, remotePort,
6295 RX_PACKET_TYPE_VERSION, a, 1, version,
6300 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6301 afs_uint16 remotePort, afs_int32 * nextConnection,
6302 int allConnections, afs_uint32 debugSupportedValues,
6303 struct rx_debugConn * conn,
6304 afs_uint32 * supportedValues)
6306 struct rx_debugIn in;
6311 * supportedValues is currently unused, but added to allow future
6312 * versioning of this function.
6315 *supportedValues = 0;
6316 if (allConnections) {
6317 in.type = htonl(RX_DEBUGI_GETALLCONN);
6319 in.type = htonl(RX_DEBUGI_GETCONN);
6321 in.index = htonl(*nextConnection);
6322 memset(conn, 0, sizeof(*conn));
6324 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6325 &in, sizeof(in), conn, sizeof(*conn));
6328 *nextConnection += 1;
6331 * Convert old connection format to new structure.
6334 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6335 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6336 #define MOVEvL(a) (conn->a = vL->a)
6338 /* any old or unrecognized version... */
6339 for (i = 0; i < RX_MAXCALLS; i++) {
6340 MOVEvL(callState[i]);
6341 MOVEvL(callMode[i]);
6342 MOVEvL(callFlags[i]);
6343 MOVEvL(callOther[i]);
6345 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6346 MOVEvL(secStats.type);
6347 MOVEvL(secStats.level);
6348 MOVEvL(secStats.flags);
6349 MOVEvL(secStats.expires);
6350 MOVEvL(secStats.packetsReceived);
6351 MOVEvL(secStats.packetsSent);
6352 MOVEvL(secStats.bytesReceived);
6353 MOVEvL(secStats.bytesSent);
6358 * Do net to host conversion here
6360 * I don't convert host or port since we are most likely
6361 * going to want these in NBO.
6363 conn->cid = ntohl(conn->cid);
6364 conn->serial = ntohl(conn->serial);
6365 for (i = 0; i < RX_MAXCALLS; i++) {
6366 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6368 conn->error = ntohl(conn->error);
6369 conn->secStats.flags = ntohl(conn->secStats.flags);
6370 conn->secStats.expires = ntohl(conn->secStats.expires);
6371 conn->secStats.packetsReceived =
6372 ntohl(conn->secStats.packetsReceived);
6373 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6374 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6375 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6376 conn->epoch = ntohl(conn->epoch);
6377 conn->natMTU = ntohl(conn->natMTU);
6384 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6385 afs_uint16 remotePort, afs_int32 * nextPeer,
6386 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6387 afs_uint32 * supportedValues)
6389 struct rx_debugIn in;
6393 * supportedValues is currently unused, but added to allow future
6394 * versioning of this function.
6397 *supportedValues = 0;
6398 in.type = htonl(RX_DEBUGI_GETPEER);
6399 in.index = htonl(*nextPeer);
6400 memset(peer, 0, sizeof(*peer));
6402 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6403 &in, sizeof(in), peer, sizeof(*peer));
6409 * Do net to host conversion here
6411 * I don't convert host or port since we are most likely
6412 * going to want these in NBO.
6414 peer->ifMTU = ntohs(peer->ifMTU);
6415 peer->idleWhen = ntohl(peer->idleWhen);
6416 peer->refCount = ntohs(peer->refCount);
6417 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6418 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6419 peer->rtt = ntohl(peer->rtt);
6420 peer->rtt_dev = ntohl(peer->rtt_dev);
6421 peer->timeout.sec = ntohl(peer->timeout.sec);
6422 peer->timeout.usec = ntohl(peer->timeout.usec);
6423 peer->nSent = ntohl(peer->nSent);
6424 peer->reSends = ntohl(peer->reSends);
6425 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6426 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6427 peer->rateFlag = ntohl(peer->rateFlag);
6428 peer->natMTU = ntohs(peer->natMTU);
6429 peer->maxMTU = ntohs(peer->maxMTU);
6430 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6431 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6432 peer->MTU = ntohs(peer->MTU);
6433 peer->cwind = ntohs(peer->cwind);
6434 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6435 peer->congestSeq = ntohs(peer->congestSeq);
6436 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6437 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6438 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6439 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6444 #endif /* RXDEBUG */
6449 struct rx_serverQueueEntry *np;
6452 register struct rx_call *call;
6453 register struct rx_serverQueueEntry *sq;
6457 if (rxinit_status == 1) {
6459 return; /* Already shutdown. */
6463 #ifndef AFS_PTHREAD_ENV
6464 FD_ZERO(&rx_selectMask);
6465 #endif /* AFS_PTHREAD_ENV */
6466 rxi_dataQuota = RX_MAX_QUOTA;
6467 #ifndef AFS_PTHREAD_ENV
6469 #endif /* AFS_PTHREAD_ENV */
6472 #ifndef AFS_PTHREAD_ENV
6473 #ifndef AFS_USE_GETTIMEOFDAY
6475 #endif /* AFS_USE_GETTIMEOFDAY */
6476 #endif /* AFS_PTHREAD_ENV */
6478 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6479 call = queue_First(&rx_freeCallQueue, rx_call);
6481 rxi_Free(call, sizeof(struct rx_call));
6484 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6485 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6491 struct rx_peer **peer_ptr, **peer_end;
6492 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6493 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6495 struct rx_peer *peer, *next;
6496 for (peer = *peer_ptr; peer; peer = next) {
6497 rx_interface_stat_p rpc_stat, nrpc_stat;
6500 (&peer->rpcStats, rpc_stat, nrpc_stat,
6501 rx_interface_stat)) {
6502 unsigned int num_funcs;
6505 queue_Remove(&rpc_stat->queue_header);
6506 queue_Remove(&rpc_stat->all_peers);
6507 num_funcs = rpc_stat->stats[0].func_total;
6509 sizeof(rx_interface_stat_t) +
6510 rpc_stat->stats[0].func_total *
6511 sizeof(rx_function_entry_v1_t);
6513 rxi_Free(rpc_stat, space);
6514 MUTEX_ENTER(&rx_rpc_stats);
6515 rxi_rpc_peer_stat_cnt -= num_funcs;
6516 MUTEX_EXIT(&rx_rpc_stats);
6520 MUTEX_ENTER(&rx_stats_mutex);
6521 rx_stats.nPeerStructs--;
6522 MUTEX_EXIT(&rx_stats_mutex);
6526 for (i = 0; i < RX_MAX_SERVICES; i++) {
6528 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6530 for (i = 0; i < rx_hashTableSize; i++) {
6531 register struct rx_connection *tc, *ntc;
6532 MUTEX_ENTER(&rx_connHashTable_lock);
6533 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6535 for (j = 0; j < RX_MAXCALLS; j++) {
6537 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6540 rxi_Free(tc, sizeof(*tc));
6542 MUTEX_EXIT(&rx_connHashTable_lock);
6545 MUTEX_ENTER(&freeSQEList_lock);
6547 while ((np = rx_FreeSQEList)) {
6548 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6549 MUTEX_DESTROY(&np->lock);
6550 rxi_Free(np, sizeof(*np));
6553 MUTEX_EXIT(&freeSQEList_lock);
6554 MUTEX_DESTROY(&freeSQEList_lock);
6555 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6556 MUTEX_DESTROY(&rx_connHashTable_lock);
6557 MUTEX_DESTROY(&rx_peerHashTable_lock);
6558 MUTEX_DESTROY(&rx_serverPool_lock);
6560 osi_Free(rx_connHashTable,
6561 rx_hashTableSize * sizeof(struct rx_connection *));
6562 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6564 UNPIN(rx_connHashTable,
6565 rx_hashTableSize * sizeof(struct rx_connection *));
6566 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6568 rxi_FreeAllPackets();
6570 MUTEX_ENTER(&rx_stats_mutex);
6571 rxi_dataQuota = RX_MAX_QUOTA;
6572 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6573 MUTEX_EXIT(&rx_stats_mutex);
6579 #ifdef RX_ENABLE_LOCKS
6581 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6583 if (!MUTEX_ISMINE(lockaddr))
6584 osi_Panic("Lock not held: %s", msg);
6586 #endif /* RX_ENABLE_LOCKS */
6591 * Routines to implement connection specific data.
6595 rx_KeyCreate(rx_destructor_t rtn)
6598 MUTEX_ENTER(&rxi_keyCreate_lock);
6599 key = rxi_keyCreate_counter++;
6600 rxi_keyCreate_destructor = (rx_destructor_t *)
6601 realloc((void *)rxi_keyCreate_destructor,
6602 (key + 1) * sizeof(rx_destructor_t));
6603 rxi_keyCreate_destructor[key] = rtn;
6604 MUTEX_EXIT(&rxi_keyCreate_lock);
6609 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6612 MUTEX_ENTER(&conn->conn_data_lock);
6613 if (!conn->specific) {
6614 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6615 for (i = 0; i < key; i++)
6616 conn->specific[i] = NULL;
6617 conn->nSpecific = key + 1;
6618 conn->specific[key] = ptr;
6619 } else if (key >= conn->nSpecific) {
6620 conn->specific = (void **)
6621 realloc(conn->specific, (key + 1) * sizeof(void *));
6622 for (i = conn->nSpecific; i < key; i++)
6623 conn->specific[i] = NULL;
6624 conn->nSpecific = key + 1;
6625 conn->specific[key] = ptr;
6627 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6628 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6629 conn->specific[key] = ptr;
6631 MUTEX_EXIT(&conn->conn_data_lock);
6635 rx_GetSpecific(struct rx_connection *conn, int key)
6638 MUTEX_ENTER(&conn->conn_data_lock);
6639 if (key >= conn->nSpecific)
6642 ptr = conn->specific[key];
6643 MUTEX_EXIT(&conn->conn_data_lock);
6647 #endif /* !KERNEL */
6650 * processStats is a queue used to store the statistics for the local
6651 * process. Its contents are similar to the contents of the rpcStats
6652 * queue on a rx_peer structure, but the actual data stored within
6653 * this queue contains totals across the lifetime of the process (assuming
6654 * the stats have not been reset) - unlike the per peer structures
6655 * which can come and go based upon the peer lifetime.
6658 static struct rx_queue processStats = { &processStats, &processStats };
6661 * peerStats is a queue used to store the statistics for all peer structs.
6662 * Its contents are the union of all the peer rpcStats queues.
6665 static struct rx_queue peerStats = { &peerStats, &peerStats };
6668 * rxi_monitor_processStats is used to turn process wide stat collection
6672 static int rxi_monitor_processStats = 0;
6675 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6678 static int rxi_monitor_peerStats = 0;
6681 * rxi_AddRpcStat - given all of the information for a particular rpc
6682 * call, create (if needed) and update the stat totals for the rpc.
6686 * IN stats - the queue of stats that will be updated with the new value
6688 * IN rxInterface - a unique number that identifies the rpc interface
6690 * IN currentFunc - the index of the function being invoked
6692 * IN totalFunc - the total number of functions in this interface
6694 * IN queueTime - the amount of time this function waited for a thread
6696 * IN execTime - the amount of time this function invocation took to execute
6698 * IN bytesSent - the number bytes sent by this invocation
6700 * IN bytesRcvd - the number bytes received by this invocation
6702 * IN isServer - if true, this invocation was made to a server
6704 * IN remoteHost - the ip address of the remote host
6706 * IN remotePort - the port of the remote host
6708 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6710 * INOUT counter - if a new stats structure is allocated, the counter will
6711 * be updated with the new number of allocated stat structures
6719 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6720 afs_uint32 currentFunc, afs_uint32 totalFunc,
6721 struct clock *queueTime, struct clock *execTime,
6722 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6723 afs_uint32 remoteHost, afs_uint32 remotePort,
6724 int addToPeerList, unsigned int *counter)
6727 rx_interface_stat_p rpc_stat, nrpc_stat;
6730 * See if there's already a structure for this interface
6733 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6734 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6735 && (rpc_stat->stats[0].remote_is_server == isServer))
6740 * Didn't find a match so allocate a new structure and add it to the
6744 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6745 || (rpc_stat->stats[0].interfaceId != rxInterface)
6746 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6751 sizeof(rx_interface_stat_t) +
6752 totalFunc * sizeof(rx_function_entry_v1_t);
6754 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6755 if (rpc_stat == NULL) {
6759 *counter += totalFunc;
6760 for (i = 0; i < totalFunc; i++) {
6761 rpc_stat->stats[i].remote_peer = remoteHost;
6762 rpc_stat->stats[i].remote_port = remotePort;
6763 rpc_stat->stats[i].remote_is_server = isServer;
6764 rpc_stat->stats[i].interfaceId = rxInterface;
6765 rpc_stat->stats[i].func_total = totalFunc;
6766 rpc_stat->stats[i].func_index = i;
6767 hzero(rpc_stat->stats[i].invocations);
6768 hzero(rpc_stat->stats[i].bytes_sent);
6769 hzero(rpc_stat->stats[i].bytes_rcvd);
6770 rpc_stat->stats[i].queue_time_sum.sec = 0;
6771 rpc_stat->stats[i].queue_time_sum.usec = 0;
6772 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6773 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6774 rpc_stat->stats[i].queue_time_min.sec = 9999999;
6775 rpc_stat->stats[i].queue_time_min.usec = 9999999;
6776 rpc_stat->stats[i].queue_time_max.sec = 0;
6777 rpc_stat->stats[i].queue_time_max.usec = 0;
6778 rpc_stat->stats[i].execution_time_sum.sec = 0;
6779 rpc_stat->stats[i].execution_time_sum.usec = 0;
6780 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
6781 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
6782 rpc_stat->stats[i].execution_time_min.sec = 9999999;
6783 rpc_stat->stats[i].execution_time_min.usec = 9999999;
6784 rpc_stat->stats[i].execution_time_max.sec = 0;
6785 rpc_stat->stats[i].execution_time_max.usec = 0;
6787 queue_Prepend(stats, rpc_stat);
6788 if (addToPeerList) {
6789 queue_Prepend(&peerStats, &rpc_stat->all_peers);
6794 * Increment the stats for this function
6797 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
6798 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
6799 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
6800 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
6801 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
6802 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
6803 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
6805 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
6806 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
6808 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
6809 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
6811 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
6812 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
6814 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
6815 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
6823 * rx_IncrementTimeAndCount - increment the times and count for a particular
6828 * IN peer - the peer who invoked the rpc
6830 * IN rxInterface - a unique number that identifies the rpc interface
6832 * IN currentFunc - the index of the function being invoked
6834 * IN totalFunc - the total number of functions in this interface
6836 * IN queueTime - the amount of time this function waited for a thread
6838 * IN execTime - the amount of time this function invocation took to execute
6840 * IN bytesSent - the number bytes sent by this invocation
6842 * IN bytesRcvd - the number bytes received by this invocation
6844 * IN isServer - if true, this invocation was made to a server
6852 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
6853 afs_uint32 currentFunc, afs_uint32 totalFunc,
6854 struct clock *queueTime, struct clock *execTime,
6855 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
6859 MUTEX_ENTER(&rx_rpc_stats);
6860 MUTEX_ENTER(&peer->peer_lock);
6862 if (rxi_monitor_peerStats) {
6863 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
6864 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6865 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
6868 if (rxi_monitor_processStats) {
6869 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
6870 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6871 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
6874 MUTEX_EXIT(&peer->peer_lock);
6875 MUTEX_EXIT(&rx_rpc_stats);
6880 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
6884 * IN callerVersion - the rpc stat version of the caller.
6886 * IN count - the number of entries to marshall.
6888 * IN stats - pointer to stats to be marshalled.
6890 * OUT ptr - Where to store the marshalled data.
6897 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
6898 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
6904 * We only support the first version
6906 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
6907 *(ptr++) = stats->remote_peer;
6908 *(ptr++) = stats->remote_port;
6909 *(ptr++) = stats->remote_is_server;
6910 *(ptr++) = stats->interfaceId;
6911 *(ptr++) = stats->func_total;
6912 *(ptr++) = stats->func_index;
6913 *(ptr++) = hgethi(stats->invocations);
6914 *(ptr++) = hgetlo(stats->invocations);
6915 *(ptr++) = hgethi(stats->bytes_sent);
6916 *(ptr++) = hgetlo(stats->bytes_sent);
6917 *(ptr++) = hgethi(stats->bytes_rcvd);
6918 *(ptr++) = hgetlo(stats->bytes_rcvd);
6919 *(ptr++) = stats->queue_time_sum.sec;
6920 *(ptr++) = stats->queue_time_sum.usec;
6921 *(ptr++) = stats->queue_time_sum_sqr.sec;
6922 *(ptr++) = stats->queue_time_sum_sqr.usec;
6923 *(ptr++) = stats->queue_time_min.sec;
6924 *(ptr++) = stats->queue_time_min.usec;
6925 *(ptr++) = stats->queue_time_max.sec;
6926 *(ptr++) = stats->queue_time_max.usec;
6927 *(ptr++) = stats->execution_time_sum.sec;
6928 *(ptr++) = stats->execution_time_sum.usec;
6929 *(ptr++) = stats->execution_time_sum_sqr.sec;
6930 *(ptr++) = stats->execution_time_sum_sqr.usec;
6931 *(ptr++) = stats->execution_time_min.sec;
6932 *(ptr++) = stats->execution_time_min.usec;
6933 *(ptr++) = stats->execution_time_max.sec;
6934 *(ptr++) = stats->execution_time_max.usec;
6940 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
6945 * IN callerVersion - the rpc stat version of the caller
6947 * OUT myVersion - the rpc stat version of this function
6949 * OUT clock_sec - local time seconds
6951 * OUT clock_usec - local time microseconds
6953 * OUT allocSize - the number of bytes allocated to contain stats
6955 * OUT statCount - the number stats retrieved from this process.
6957 * OUT stats - the actual stats retrieved from this process.
6961 * Returns void. If successful, stats will != NULL.
6965 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
6966 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
6967 size_t * allocSize, afs_uint32 * statCount,
6968 afs_uint32 ** stats)
6978 *myVersion = RX_STATS_RETRIEVAL_VERSION;
6981 * Check to see if stats are enabled
6984 MUTEX_ENTER(&rx_rpc_stats);
6985 if (!rxi_monitor_processStats) {
6986 MUTEX_EXIT(&rx_rpc_stats);
6990 clock_GetTime(&now);
6991 *clock_sec = now.sec;
6992 *clock_usec = now.usec;
6995 * Allocate the space based upon the caller version
6997 * If the client is at an older version than we are,
6998 * we return the statistic data in the older data format, but
6999 * we still return our version number so the client knows we
7000 * are maintaining more data than it can retrieve.
7003 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7004 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7005 *statCount = rxi_rpc_process_stat_cnt;
7008 * This can't happen yet, but in the future version changes
7009 * can be handled by adding additional code here
7013 if (space > (size_t) 0) {
7015 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7018 rx_interface_stat_p rpc_stat, nrpc_stat;
7022 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7024 * Copy the data based upon the caller version
7026 rx_MarshallProcessRPCStats(callerVersion,
7027 rpc_stat->stats[0].func_total,
7028 rpc_stat->stats, &ptr);
7034 MUTEX_EXIT(&rx_rpc_stats);
7039 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7043 * IN callerVersion - the rpc stat version of the caller
7045 * OUT myVersion - the rpc stat version of this function
7047 * OUT clock_sec - local time seconds
7049 * OUT clock_usec - local time microseconds
7051 * OUT allocSize - the number of bytes allocated to contain stats
7053 * OUT statCount - the number of stats retrieved from the individual
7056 * OUT stats - the actual stats retrieved from the individual peer structures.
7060 * Returns void. If successful, stats will != NULL.
7064 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7065 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7066 size_t * allocSize, afs_uint32 * statCount,
7067 afs_uint32 ** stats)
7077 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7080 * Check to see if stats are enabled
7083 MUTEX_ENTER(&rx_rpc_stats);
7084 if (!rxi_monitor_peerStats) {
7085 MUTEX_EXIT(&rx_rpc_stats);
7089 clock_GetTime(&now);
7090 *clock_sec = now.sec;
7091 *clock_usec = now.usec;
7094 * Allocate the space based upon the caller version
7096 * If the client is at an older version than we are,
7097 * we return the statistic data in the older data format, but
7098 * we still return our version number so the client knows we
7099 * are maintaining more data than it can retrieve.
7102 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7103 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7104 *statCount = rxi_rpc_peer_stat_cnt;
7107 * This can't happen yet, but in the future version changes
7108 * can be handled by adding additional code here
7112 if (space > (size_t) 0) {
7114 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7117 rx_interface_stat_p rpc_stat, nrpc_stat;
7121 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7123 * We have to fix the offset of rpc_stat since we are
7124 * keeping this structure on two rx_queues. The rx_queue
7125 * package assumes that the rx_queue member is the first
7126 * member of the structure. That is, rx_queue assumes that
7127 * any one item is only on one queue at a time. We are
7128 * breaking that assumption and so we have to do a little
7129 * math to fix our pointers.
7132 fix_offset = (char *)rpc_stat;
7133 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7134 rpc_stat = (rx_interface_stat_p) fix_offset;
7137 * Copy the data based upon the caller version
7139 rx_MarshallProcessRPCStats(callerVersion,
7140 rpc_stat->stats[0].func_total,
7141 rpc_stat->stats, &ptr);
7147 MUTEX_EXIT(&rx_rpc_stats);
7152 * rx_FreeRPCStats - free memory allocated by
7153 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7157 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7158 * rx_RetrievePeerRPCStats
7160 * IN allocSize - the number of bytes in stats.
7168 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7170 rxi_Free(stats, allocSize);
7174 * rx_queryProcessRPCStats - see if process rpc stat collection is
7175 * currently enabled.
7181 * Returns 0 if stats are not enabled != 0 otherwise
7185 rx_queryProcessRPCStats(void)
7188 MUTEX_ENTER(&rx_rpc_stats);
7189 rc = rxi_monitor_processStats;
7190 MUTEX_EXIT(&rx_rpc_stats);
7195 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7201 * Returns 0 if stats are not enabled != 0 otherwise
7205 rx_queryPeerRPCStats(void)
7208 MUTEX_ENTER(&rx_rpc_stats);
7209 rc = rxi_monitor_peerStats;
7210 MUTEX_EXIT(&rx_rpc_stats);
7215 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7225 rx_enableProcessRPCStats(void)
7227 MUTEX_ENTER(&rx_rpc_stats);
7228 rx_enable_stats = 1;
7229 rxi_monitor_processStats = 1;
7230 MUTEX_EXIT(&rx_rpc_stats);
7234 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7244 rx_enablePeerRPCStats(void)
7246 MUTEX_ENTER(&rx_rpc_stats);
7247 rx_enable_stats = 1;
7248 rxi_monitor_peerStats = 1;
7249 MUTEX_EXIT(&rx_rpc_stats);
7253 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7263 rx_disableProcessRPCStats(void)
7265 rx_interface_stat_p rpc_stat, nrpc_stat;
7268 MUTEX_ENTER(&rx_rpc_stats);
7271 * Turn off process statistics and if peer stats is also off, turn
7275 rxi_monitor_processStats = 0;
7276 if (rxi_monitor_peerStats == 0) {
7277 rx_enable_stats = 0;
7280 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7281 unsigned int num_funcs = 0;
7284 queue_Remove(rpc_stat);
7285 num_funcs = rpc_stat->stats[0].func_total;
7287 sizeof(rx_interface_stat_t) +
7288 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7290 rxi_Free(rpc_stat, space);
7291 rxi_rpc_process_stat_cnt -= num_funcs;
7293 MUTEX_EXIT(&rx_rpc_stats);
7297 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7307 rx_disablePeerRPCStats(void)
7309 struct rx_peer **peer_ptr, **peer_end;
7312 MUTEX_ENTER(&rx_rpc_stats);
7315 * Turn off peer statistics and if process stats is also off, turn
7319 rxi_monitor_peerStats = 0;
7320 if (rxi_monitor_processStats == 0) {
7321 rx_enable_stats = 0;
7324 MUTEX_ENTER(&rx_peerHashTable_lock);
7325 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7326 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7328 struct rx_peer *peer, *next, *prev;
7329 for (prev = peer = *peer_ptr; peer; peer = next) {
7331 code = MUTEX_TRYENTER(&peer->peer_lock);
7333 rx_interface_stat_p rpc_stat, nrpc_stat;
7336 (&peer->rpcStats, rpc_stat, nrpc_stat,
7337 rx_interface_stat)) {
7338 unsigned int num_funcs = 0;
7341 queue_Remove(&rpc_stat->queue_header);
7342 queue_Remove(&rpc_stat->all_peers);
7343 num_funcs = rpc_stat->stats[0].func_total;
7345 sizeof(rx_interface_stat_t) +
7346 rpc_stat->stats[0].func_total *
7347 sizeof(rx_function_entry_v1_t);
7349 rxi_Free(rpc_stat, space);
7350 rxi_rpc_peer_stat_cnt -= num_funcs;
7352 MUTEX_EXIT(&peer->peer_lock);
7353 if (prev == *peer_ptr) {
7363 MUTEX_EXIT(&rx_peerHashTable_lock);
7364 MUTEX_EXIT(&rx_rpc_stats);
7368 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7373 * IN clearFlag - flag indicating which stats to clear
7381 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7383 rx_interface_stat_p rpc_stat, nrpc_stat;
7385 MUTEX_ENTER(&rx_rpc_stats);
7387 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7388 unsigned int num_funcs = 0, i;
7389 num_funcs = rpc_stat->stats[0].func_total;
7390 for (i = 0; i < num_funcs; i++) {
7391 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7392 hzero(rpc_stat->stats[i].invocations);
7394 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7395 hzero(rpc_stat->stats[i].bytes_sent);
7397 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7398 hzero(rpc_stat->stats[i].bytes_rcvd);
7400 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7401 rpc_stat->stats[i].queue_time_sum.sec = 0;
7402 rpc_stat->stats[i].queue_time_sum.usec = 0;
7404 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7405 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7406 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7408 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7409 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7410 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7412 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7413 rpc_stat->stats[i].queue_time_max.sec = 0;
7414 rpc_stat->stats[i].queue_time_max.usec = 0;
7416 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7417 rpc_stat->stats[i].execution_time_sum.sec = 0;
7418 rpc_stat->stats[i].execution_time_sum.usec = 0;
7420 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7421 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7422 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7424 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7425 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7426 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7428 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7429 rpc_stat->stats[i].execution_time_max.sec = 0;
7430 rpc_stat->stats[i].execution_time_max.usec = 0;
7435 MUTEX_EXIT(&rx_rpc_stats);
7439 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7444 * IN clearFlag - flag indicating which stats to clear
7452 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7454 rx_interface_stat_p rpc_stat, nrpc_stat;
7456 MUTEX_ENTER(&rx_rpc_stats);
7458 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7459 unsigned int num_funcs = 0, i;
7462 * We have to fix the offset of rpc_stat since we are
7463 * keeping this structure on two rx_queues. The rx_queue
7464 * package assumes that the rx_queue member is the first
7465 * member of the structure. That is, rx_queue assumes that
7466 * any one item is only on one queue at a time. We are
7467 * breaking that assumption and so we have to do a little
7468 * math to fix our pointers.
7471 fix_offset = (char *)rpc_stat;
7472 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7473 rpc_stat = (rx_interface_stat_p) fix_offset;
7475 num_funcs = rpc_stat->stats[0].func_total;
7476 for (i = 0; i < num_funcs; i++) {
7477 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7478 hzero(rpc_stat->stats[i].invocations);
7480 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7481 hzero(rpc_stat->stats[i].bytes_sent);
7483 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7484 hzero(rpc_stat->stats[i].bytes_rcvd);
7486 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7487 rpc_stat->stats[i].queue_time_sum.sec = 0;
7488 rpc_stat->stats[i].queue_time_sum.usec = 0;
7490 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7491 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7492 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7494 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7495 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7496 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7498 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7499 rpc_stat->stats[i].queue_time_max.sec = 0;
7500 rpc_stat->stats[i].queue_time_max.usec = 0;
7502 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7503 rpc_stat->stats[i].execution_time_sum.sec = 0;
7504 rpc_stat->stats[i].execution_time_sum.usec = 0;
7506 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7507 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7508 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7510 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7511 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7512 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7514 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7515 rpc_stat->stats[i].execution_time_max.sec = 0;
7516 rpc_stat->stats[i].execution_time_max.usec = 0;
7521 MUTEX_EXIT(&rx_rpc_stats);
7525 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7526 * is authorized to enable/disable/clear RX statistics.
7528 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7531 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7533 rxi_rxstat_userok = proc;
7537 rx_RxStatUserOk(struct rx_call *call)
7539 if (!rxi_rxstat_userok)
7541 return rxi_rxstat_userok(call);