2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
14 #include "afs/param.h"
16 #include <afs/param.h>
23 #include "afs/sysincludes.h"
24 #include "afsincludes.h"
30 #include <net/net_globals.h>
31 #endif /* AFS_OSF_ENV */
32 #ifdef AFS_LINUX20_ENV
35 #include "netinet/in.h"
36 #include "afs/afs_args.h"
37 #include "afs/afs_osi.h"
38 #ifdef RX_KERNEL_TRACE
39 #include "rx_kcommon.h"
41 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
45 #undef RXDEBUG /* turn off debugging */
47 #if defined(AFS_SGI_ENV)
48 #include "sys/debug.h"
57 #endif /* AFS_OSF_ENV */
59 #include "afs/sysincludes.h"
60 #include "afsincludes.h"
63 #include "rx_kmutex.h"
64 #include "rx_kernel.h"
68 #include "rx_globals.h"
70 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
71 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
72 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
74 extern afs_int32 afs_termState;
76 #include "sys/lockl.h"
77 #include "sys/lock_def.h"
78 #endif /* AFS_AIX41_ENV */
79 # include "rxgen_consts.h"
81 # include <sys/types.h>
86 # include <afs/afsutil.h>
88 # include <sys/socket.h>
89 # include <sys/file.h>
91 # include <sys/stat.h>
92 # include <netinet/in.h>
93 # include <sys/time.h>
103 # include "rx_user.h"
104 # include "rx_clock.h"
105 # include "rx_queue.h"
106 # include "rx_globals.h"
107 # include "rx_trace.h"
108 # include <afs/rxgen_consts.h>
111 int (*registerProgram) () = 0;
112 int (*swapNameProgram) () = 0;
114 /* Local static routines */
115 static void rxi_DestroyConnectionNoLock(register struct rx_connection *conn);
116 #ifdef RX_ENABLE_LOCKS
117 static void rxi_SetAcksInTransmitQueue(register struct rx_call *call);
120 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
122 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
123 afs_int32 rxi_start_in_error;
125 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
128 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
129 * currently allocated within rx. This number is used to allocate the
130 * memory required to return the statistics when queried.
133 static unsigned int rxi_rpc_peer_stat_cnt;
136 * rxi_rpc_process_stat_cnt counts the total number of local process stat
137 * structures currently allocated within rx. The number is used to allocate
138 * the memory required to return the statistics when queried.
141 static unsigned int rxi_rpc_process_stat_cnt;
143 #if !defined(offsetof)
144 #include <stddef.h> /* for definition of offsetof() */
147 #ifdef AFS_PTHREAD_ENV
151 * Use procedural initialization of mutexes/condition variables
155 extern pthread_mutex_t rx_stats_mutex;
156 extern pthread_mutex_t des_init_mutex;
157 extern pthread_mutex_t des_random_mutex;
158 extern pthread_mutex_t rx_clock_mutex;
159 extern pthread_mutex_t rxi_connCacheMutex;
160 extern pthread_mutex_t rx_event_mutex;
161 extern pthread_mutex_t osi_malloc_mutex;
162 extern pthread_mutex_t event_handler_mutex;
163 extern pthread_mutex_t listener_mutex;
164 extern pthread_mutex_t rx_if_init_mutex;
165 extern pthread_mutex_t rx_if_mutex;
166 extern pthread_mutex_t rxkad_client_uid_mutex;
167 extern pthread_mutex_t rxkad_random_mutex;
169 extern pthread_cond_t rx_event_handler_cond;
170 extern pthread_cond_t rx_listener_cond;
172 static pthread_mutex_t epoch_mutex;
173 static pthread_mutex_t rx_init_mutex;
174 static pthread_mutex_t rx_debug_mutex;
177 rxi_InitPthread(void)
179 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
181 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
183 assert(pthread_mutex_init
184 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
185 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
187 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
189 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
191 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
193 assert(pthread_mutex_init
194 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
195 assert(pthread_mutex_init
196 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
197 assert(pthread_mutex_init
198 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
199 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
201 assert(pthread_mutex_init
202 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
203 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
205 assert(pthread_mutex_init
206 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
207 assert(pthread_mutex_init
208 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
209 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
212 assert(pthread_cond_init
213 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
214 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
216 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
217 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
219 rxkad_global_stats_init();
222 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
223 #define INIT_PTHREAD_LOCKS \
224 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
226 * The rx_stats_mutex mutex protects the following global variables:
231 * rxi_lowConnRefCount
232 * rxi_lowPeerRefCount
241 #define INIT_PTHREAD_LOCKS
245 /* Variables for handling the minProcs implementation. availProcs gives the
246 * number of threads available in the pool at this moment (not counting dudes
247 * executing right now). totalMin gives the total number of procs required
248 * for handling all minProcs requests. minDeficit is a dynamic variable
249 * tracking the # of procs required to satisfy all of the remaining minProcs
251 * For fine grain locking to work, the quota check and the reservation of
252 * a server thread has to come while rxi_availProcs and rxi_minDeficit
253 * are locked. To this end, the code has been modified under #ifdef
254 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
255 * same time. A new function, ReturnToServerPool() returns the allocation.
257 * A call can be on several queue's (but only one at a time). When
258 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
259 * that no one else is touching the queue. To this end, we store the address
260 * of the queue lock in the call structure (under the call lock) when we
261 * put the call on a queue, and we clear the call_queue_lock when the
262 * call is removed from a queue (once the call lock has been obtained).
263 * This allows rxi_ResetCall to safely synchronize with others wishing
264 * to manipulate the queue.
267 #ifdef RX_ENABLE_LOCKS
268 static afs_kmutex_t rx_rpc_stats;
269 void rxi_StartUnlocked();
272 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
273 ** pretty good that the next packet coming in is from the same connection
274 ** as the last packet, since we're send multiple packets in a transmit window.
276 struct rx_connection *rxLastConn = 0;
278 #ifdef RX_ENABLE_LOCKS
279 /* The locking hierarchy for rx fine grain locking is composed of these
282 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
283 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
284 * call->lock - locks call data fields.
285 * These are independent of each other:
286 * rx_freeCallQueue_lock
291 * serverQueueEntry->lock
293 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
294 * peer->lock - locks peer data fields.
295 * conn_data_lock - that more than one thread is not updating a conn data
296 * field at the same time.
304 * Do we need a lock to protect the peer field in the conn structure?
305 * conn->peer was previously a constant for all intents and so has no
306 * lock protecting this field. The multihomed client delta introduced
307 * a RX code change : change the peer field in the connection structure
308 * to that remote inetrface from which the last packet for this
309 * connection was sent out. This may become an issue if further changes
312 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
313 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
315 /* rxdb_fileID is used to identify the lock location, along with line#. */
316 static int rxdb_fileID = RXDB_FILE_RX;
317 #endif /* RX_LOCKS_DB */
318 #else /* RX_ENABLE_LOCKS */
319 #define SET_CALL_QUEUE_LOCK(C, L)
320 #define CLEAR_CALL_QUEUE_LOCK(C)
321 #endif /* RX_ENABLE_LOCKS */
322 struct rx_serverQueueEntry *rx_waitForPacket = 0;
323 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
325 /* ------------Exported Interfaces------------- */
327 /* This function allows rxkad to set the epoch to a suitably random number
328 * which rx_NewConnection will use in the future. The principle purpose is to
329 * get rxnull connections to use the same epoch as the rxkad connections do, at
330 * least once the first rxkad connection is established. This is important now
331 * that the host/port addresses aren't used in FindConnection: the uniqueness
332 * of epoch/cid matters and the start time won't do. */
334 #ifdef AFS_PTHREAD_ENV
336 * This mutex protects the following global variables:
340 #define LOCK_EPOCH assert(pthread_mutex_lock(&epoch_mutex)==0)
341 #define UNLOCK_EPOCH assert(pthread_mutex_unlock(&epoch_mutex)==0)
345 #endif /* AFS_PTHREAD_ENV */
348 rx_SetEpoch(afs_uint32 epoch)
355 /* Initialize rx. A port number may be mentioned, in which case this
356 * becomes the default port number for any service installed later.
357 * If 0 is provided for the port number, a random port will be chosen
358 * by the kernel. Whether this will ever overlap anything in
359 * /etc/services is anybody's guess... Returns 0 on success, -1 on
361 static int rxinit_status = 1;
362 #ifdef AFS_PTHREAD_ENV
364 * This mutex protects the following global variables:
368 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0)
369 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0)
372 #define UNLOCK_RX_INIT
376 rx_InitHost(u_int host, u_int port)
383 char *htable, *ptable;
386 #if defined(AFS_DJGPP_ENV) && !defined(DEBUG)
387 __djgpp_set_quiet_socket(1);
394 if (rxinit_status == 0) {
395 tmp_status = rxinit_status;
397 return tmp_status; /* Already started; return previous error code. */
400 if (afs_winsockInit() < 0)
406 * Initialize anything necessary to provide a non-premptive threading
409 rxi_InitializeThreadSupport();
412 /* Allocate and initialize a socket for client and perhaps server
415 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
416 if (rx_socket == OSI_NULLSOCKET) {
420 #ifdef RX_ENABLE_LOCKS
423 #endif /* RX_LOCKS_DB */
424 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
425 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
426 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
427 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
428 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
430 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
432 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
434 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
436 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
438 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
440 #if defined(KERNEL) && defined(AFS_HPUX110_ENV)
442 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
443 #endif /* KERNEL && AFS_HPUX110_ENV */
444 #else /* RX_ENABLE_LOCKS */
445 #if defined(KERNEL) && defined(AFS_GLOBAL_SUNLOCK) && !defined(AFS_HPUX_ENV) && !defined(AFS_OBSD_ENV)
446 mutex_init(&afs_rxglobal_lock, "afs_rxglobal_lock", MUTEX_DEFAULT, NULL);
447 #endif /* AFS_GLOBAL_SUNLOCK */
448 #endif /* RX_ENABLE_LOCKS */
451 rx_connDeadTime = 12;
452 rx_tranquil = 0; /* reset flag */
453 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
455 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
456 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
457 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
458 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
459 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
460 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
462 /* Malloc up a bunch of packets & buffers */
464 queue_Init(&rx_freePacketQueue);
465 rxi_NeedMorePackets = FALSE;
466 #ifdef RX_ENABLE_TSFPQ
467 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
468 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
469 #else /* RX_ENABLE_TSFPQ */
470 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
471 rxi_MorePackets(rx_nPackets);
472 #endif /* RX_ENABLE_TSFPQ */
479 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
480 tv.tv_sec = clock_now.sec;
481 tv.tv_usec = clock_now.usec;
482 srand((unsigned int)tv.tv_usec);
489 #if defined(KERNEL) && !defined(UKERNEL)
490 /* Really, this should never happen in a real kernel */
493 struct sockaddr_in addr;
494 int addrlen = sizeof(addr);
495 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
499 rx_port = addr.sin_port;
502 rx_stats.minRtt.sec = 9999999;
504 rx_SetEpoch(tv.tv_sec | 0x80000000);
506 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
507 * will provide a randomer value. */
509 MUTEX_ENTER(&rx_stats_mutex);
510 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
511 MUTEX_EXIT(&rx_stats_mutex);
512 /* *Slightly* random start time for the cid. This is just to help
513 * out with the hashing function at the peer */
514 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
515 rx_connHashTable = (struct rx_connection **)htable;
516 rx_peerHashTable = (struct rx_peer **)ptable;
518 rx_lastAckDelay.sec = 0;
519 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
520 rx_hardAckDelay.sec = 0;
521 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
522 rx_softAckDelay.sec = 0;
523 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
525 rxevent_Init(20, rxi_ReScheduleEvents);
527 /* Initialize various global queues */
528 queue_Init(&rx_idleServerQueue);
529 queue_Init(&rx_incomingCallQueue);
530 queue_Init(&rx_freeCallQueue);
532 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
533 /* Initialize our list of usable IP addresses. */
537 /* Start listener process (exact function is dependent on the
538 * implementation environment--kernel or user space) */
542 tmp_status = rxinit_status = 0;
550 return rx_InitHost(htonl(INADDR_ANY), port);
553 /* called with unincremented nRequestsRunning to see if it is OK to start
554 * a new thread in this service. Could be "no" for two reasons: over the
555 * max quota, or would prevent others from reaching their min quota.
557 #ifdef RX_ENABLE_LOCKS
558 /* This verion of QuotaOK reserves quota if it's ok while the
559 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
562 QuotaOK(register struct rx_service *aservice)
564 /* check if over max quota */
565 if (aservice->nRequestsRunning >= aservice->maxProcs) {
569 /* under min quota, we're OK */
570 /* otherwise, can use only if there are enough to allow everyone
571 * to go to their min quota after this guy starts.
573 MUTEX_ENTER(&rx_stats_mutex);
574 if ((aservice->nRequestsRunning < aservice->minProcs)
575 || (rxi_availProcs > rxi_minDeficit)) {
576 aservice->nRequestsRunning++;
577 /* just started call in minProcs pool, need fewer to maintain
579 if (aservice->nRequestsRunning <= aservice->minProcs)
582 MUTEX_EXIT(&rx_stats_mutex);
585 MUTEX_EXIT(&rx_stats_mutex);
591 ReturnToServerPool(register struct rx_service *aservice)
593 aservice->nRequestsRunning--;
594 MUTEX_ENTER(&rx_stats_mutex);
595 if (aservice->nRequestsRunning < aservice->minProcs)
598 MUTEX_EXIT(&rx_stats_mutex);
601 #else /* RX_ENABLE_LOCKS */
603 QuotaOK(register struct rx_service *aservice)
606 /* under min quota, we're OK */
607 if (aservice->nRequestsRunning < aservice->minProcs)
610 /* check if over max quota */
611 if (aservice->nRequestsRunning >= aservice->maxProcs)
614 /* otherwise, can use only if there are enough to allow everyone
615 * to go to their min quota after this guy starts.
617 if (rxi_availProcs > rxi_minDeficit)
621 #endif /* RX_ENABLE_LOCKS */
624 /* Called by rx_StartServer to start up lwp's to service calls.
625 NExistingProcs gives the number of procs already existing, and which
626 therefore needn't be created. */
628 rxi_StartServerProcs(int nExistingProcs)
630 register struct rx_service *service;
635 /* For each service, reserve N processes, where N is the "minimum"
636 * number of processes that MUST be able to execute a request in parallel,
637 * at any time, for that process. Also compute the maximum difference
638 * between any service's maximum number of processes that can run
639 * (i.e. the maximum number that ever will be run, and a guarantee
640 * that this number will run if other services aren't running), and its
641 * minimum number. The result is the extra number of processes that
642 * we need in order to provide the latter guarantee */
643 for (i = 0; i < RX_MAX_SERVICES; i++) {
645 service = rx_services[i];
646 if (service == (struct rx_service *)0)
648 nProcs += service->minProcs;
649 diff = service->maxProcs - service->minProcs;
653 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
654 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
655 for (i = 0; i < nProcs; i++) {
656 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
662 /* This routine is only required on Windows */
664 rx_StartClientThread(void)
666 #ifdef AFS_PTHREAD_ENV
668 pid = (int) pthread_self();
669 #endif /* AFS_PTHREAD_ENV */
671 #endif /* AFS_NT40_ENV */
673 /* This routine must be called if any services are exported. If the
674 * donateMe flag is set, the calling process is donated to the server
677 rx_StartServer(int donateMe)
679 register struct rx_service *service;
685 /* Start server processes, if necessary (exact function is dependent
686 * on the implementation environment--kernel or user space). DonateMe
687 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
688 * case, one less new proc will be created rx_StartServerProcs.
690 rxi_StartServerProcs(donateMe);
692 /* count up the # of threads in minProcs, and add set the min deficit to
693 * be that value, too.
695 for (i = 0; i < RX_MAX_SERVICES; i++) {
696 service = rx_services[i];
697 if (service == (struct rx_service *)0)
699 MUTEX_ENTER(&rx_stats_mutex);
700 rxi_totalMin += service->minProcs;
701 /* below works even if a thread is running, since minDeficit would
702 * still have been decremented and later re-incremented.
704 rxi_minDeficit += service->minProcs;
705 MUTEX_EXIT(&rx_stats_mutex);
708 /* Turn on reaping of idle server connections */
709 rxi_ReapConnections();
718 #ifdef AFS_PTHREAD_ENV
720 pid = (pid_t) pthread_self();
721 #else /* AFS_PTHREAD_ENV */
723 LWP_CurrentProcess(&pid);
724 #endif /* AFS_PTHREAD_ENV */
726 sprintf(name, "srv_%d", ++nProcs);
728 (*registerProgram) (pid, name);
730 #endif /* AFS_NT40_ENV */
731 rx_ServerProc(); /* Never returns */
733 #ifdef RX_ENABLE_TSFPQ
734 /* no use leaving packets around in this thread's local queue if
735 * it isn't getting donated to the server thread pool.
737 rxi_FlushLocalPacketsTSFPQ();
738 #endif /* RX_ENABLE_TSFPQ */
742 /* Create a new client connection to the specified service, using the
743 * specified security object to implement the security model for this
745 struct rx_connection *
746 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
747 register struct rx_securityClass *securityObject,
748 int serviceSecurityIndex)
752 register struct rx_connection *conn;
757 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", shost, sport, sservice, securityObject, serviceSecurityIndex));
759 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
760 * the case of kmem_alloc? */
761 conn = rxi_AllocConnection();
762 #ifdef RX_ENABLE_LOCKS
763 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
764 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
765 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
768 MUTEX_ENTER(&rx_connHashTable_lock);
769 cid = (rx_nextCid += RX_MAXCALLS);
770 conn->type = RX_CLIENT_CONNECTION;
772 conn->epoch = rx_epoch;
773 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
774 conn->serviceId = sservice;
775 conn->securityObject = securityObject;
776 /* This doesn't work in all compilers with void (they're buggy), so fake it
778 conn->securityData = (VOID *) 0;
779 conn->securityIndex = serviceSecurityIndex;
780 rx_SetConnDeadTime(conn, rx_connDeadTime);
781 conn->ackRate = RX_FAST_ACK_RATE;
783 conn->specific = NULL;
784 conn->challengeEvent = NULL;
785 conn->delayedAbortEvent = NULL;
786 conn->abortCount = 0;
789 RXS_NewConnection(securityObject, conn);
791 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
793 conn->refCount++; /* no lock required since only this thread knows... */
794 conn->next = rx_connHashTable[hashindex];
795 rx_connHashTable[hashindex] = conn;
796 MUTEX_ENTER(&rx_stats_mutex);
797 rx_stats.nClientConns++;
798 MUTEX_EXIT(&rx_stats_mutex);
800 MUTEX_EXIT(&rx_connHashTable_lock);
806 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
808 /* The idea is to set the dead time to a value that allows several
809 * keepalives to be dropped without timing out the connection. */
810 conn->secondsUntilDead = MAX(seconds, 6);
811 conn->secondsUntilPing = conn->secondsUntilDead / 6;
814 int rxi_lowPeerRefCount = 0;
815 int rxi_lowConnRefCount = 0;
818 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
819 * NOTE: must not be called with rx_connHashTable_lock held.
822 rxi_CleanupConnection(struct rx_connection *conn)
824 /* Notify the service exporter, if requested, that this connection
825 * is being destroyed */
826 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
827 (*conn->service->destroyConnProc) (conn);
829 /* Notify the security module that this connection is being destroyed */
830 RXS_DestroyConnection(conn->securityObject, conn);
832 /* If this is the last connection using the rx_peer struct, set its
833 * idle time to now. rxi_ReapConnections will reap it if it's still
834 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
836 MUTEX_ENTER(&rx_peerHashTable_lock);
837 if (conn->peer->refCount < 2) {
838 conn->peer->idleWhen = clock_Sec();
839 if (conn->peer->refCount < 1) {
840 conn->peer->refCount = 1;
841 MUTEX_ENTER(&rx_stats_mutex);
842 rxi_lowPeerRefCount++;
843 MUTEX_EXIT(&rx_stats_mutex);
846 conn->peer->refCount--;
847 MUTEX_EXIT(&rx_peerHashTable_lock);
849 MUTEX_ENTER(&rx_stats_mutex);
850 if (conn->type == RX_SERVER_CONNECTION)
851 rx_stats.nServerConns--;
853 rx_stats.nClientConns--;
854 MUTEX_EXIT(&rx_stats_mutex);
857 if (conn->specific) {
859 for (i = 0; i < conn->nSpecific; i++) {
860 if (conn->specific[i] && rxi_keyCreate_destructor[i])
861 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
862 conn->specific[i] = NULL;
864 free(conn->specific);
866 conn->specific = NULL;
870 MUTEX_DESTROY(&conn->conn_call_lock);
871 MUTEX_DESTROY(&conn->conn_data_lock);
872 CV_DESTROY(&conn->conn_call_cv);
874 rxi_FreeConnection(conn);
877 /* Destroy the specified connection */
879 rxi_DestroyConnection(register struct rx_connection *conn)
881 MUTEX_ENTER(&rx_connHashTable_lock);
882 rxi_DestroyConnectionNoLock(conn);
883 /* conn should be at the head of the cleanup list */
884 if (conn == rx_connCleanup_list) {
885 rx_connCleanup_list = rx_connCleanup_list->next;
886 MUTEX_EXIT(&rx_connHashTable_lock);
887 rxi_CleanupConnection(conn);
889 #ifdef RX_ENABLE_LOCKS
891 MUTEX_EXIT(&rx_connHashTable_lock);
893 #endif /* RX_ENABLE_LOCKS */
897 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
899 register struct rx_connection **conn_ptr;
900 register int havecalls = 0;
901 struct rx_packet *packet;
908 MUTEX_ENTER(&conn->conn_data_lock);
909 if (conn->refCount > 0)
912 MUTEX_ENTER(&rx_stats_mutex);
913 rxi_lowConnRefCount++;
914 MUTEX_EXIT(&rx_stats_mutex);
917 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
918 /* Busy; wait till the last guy before proceeding */
919 MUTEX_EXIT(&conn->conn_data_lock);
924 /* If the client previously called rx_NewCall, but it is still
925 * waiting, treat this as a running call, and wait to destroy the
926 * connection later when the call completes. */
927 if ((conn->type == RX_CLIENT_CONNECTION)
928 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
929 conn->flags |= RX_CONN_DESTROY_ME;
930 MUTEX_EXIT(&conn->conn_data_lock);
934 MUTEX_EXIT(&conn->conn_data_lock);
936 /* Check for extant references to this connection */
937 for (i = 0; i < RX_MAXCALLS; i++) {
938 register struct rx_call *call = conn->call[i];
941 if (conn->type == RX_CLIENT_CONNECTION) {
942 MUTEX_ENTER(&call->lock);
943 if (call->delayedAckEvent) {
944 /* Push the final acknowledgment out now--there
945 * won't be a subsequent call to acknowledge the
946 * last reply packets */
947 rxevent_Cancel(call->delayedAckEvent, call,
948 RX_CALL_REFCOUNT_DELAY);
949 if (call->state == RX_STATE_PRECALL
950 || call->state == RX_STATE_ACTIVE) {
951 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
953 rxi_AckAll(NULL, call, 0);
956 MUTEX_EXIT(&call->lock);
960 #ifdef RX_ENABLE_LOCKS
962 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
963 MUTEX_EXIT(&conn->conn_data_lock);
965 /* Someone is accessing a packet right now. */
969 #endif /* RX_ENABLE_LOCKS */
972 /* Don't destroy the connection if there are any call
973 * structures still in use */
974 MUTEX_ENTER(&conn->conn_data_lock);
975 conn->flags |= RX_CONN_DESTROY_ME;
976 MUTEX_EXIT(&conn->conn_data_lock);
981 if (conn->delayedAbortEvent) {
982 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
983 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
985 MUTEX_ENTER(&conn->conn_data_lock);
986 rxi_SendConnectionAbort(conn, packet, 0, 1);
987 MUTEX_EXIT(&conn->conn_data_lock);
988 rxi_FreePacket(packet);
992 /* Remove from connection hash table before proceeding */
994 &rx_connHashTable[CONN_HASH
995 (peer->host, peer->port, conn->cid, conn->epoch,
997 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
998 if (*conn_ptr == conn) {
999 *conn_ptr = conn->next;
1003 /* if the conn that we are destroying was the last connection, then we
1004 * clear rxLastConn as well */
1005 if (rxLastConn == conn)
1008 /* Make sure the connection is completely reset before deleting it. */
1009 /* get rid of pending events that could zap us later */
1010 if (conn->challengeEvent)
1011 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1012 if (conn->checkReachEvent)
1013 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1015 /* Add the connection to the list of destroyed connections that
1016 * need to be cleaned up. This is necessary to avoid deadlocks
1017 * in the routines we call to inform others that this connection is
1018 * being destroyed. */
1019 conn->next = rx_connCleanup_list;
1020 rx_connCleanup_list = conn;
1023 /* Externally available version */
1025 rx_DestroyConnection(register struct rx_connection *conn)
1030 rxi_DestroyConnection(conn);
1035 rx_GetConnection(register struct rx_connection *conn)
1040 MUTEX_ENTER(&conn->conn_data_lock);
1042 MUTEX_EXIT(&conn->conn_data_lock);
1046 /* Start a new rx remote procedure call, on the specified connection.
1047 * If wait is set to 1, wait for a free call channel; otherwise return
1048 * 0. Maxtime gives the maximum number of seconds this call may take,
1049 * after rx_MakeCall returns. After this time interval, a call to any
1050 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1051 * For fine grain locking, we hold the conn_call_lock in order to
1052 * to ensure that we don't get signalle after we found a call in an active
1053 * state and before we go to sleep.
1056 rx_NewCall(register struct rx_connection *conn)
1059 register struct rx_call *call;
1060 struct clock queueTime;
1064 dpf(("rx_MakeCall(conn %x)\n", conn));
1067 clock_GetTime(&queueTime);
1068 MUTEX_ENTER(&conn->conn_call_lock);
1071 * Check if there are others waiting for a new call.
1072 * If so, let them go first to avoid starving them.
1073 * This is a fairly simple scheme, and might not be
1074 * a complete solution for large numbers of waiters.
1076 * makeCallWaiters keeps track of the number of
1077 * threads waiting to make calls and the
1078 * RX_CONN_MAKECALL_WAITING flag bit is used to
1079 * indicate that there are indeed calls waiting.
1080 * The flag is set when the waiter is incremented.
1081 * It is only cleared in rx_EndCall when
1082 * makeCallWaiters is 0. This prevents us from
1083 * accidently destroying the connection while it
1084 * is potentially about to be used.
1086 MUTEX_ENTER(&conn->conn_data_lock);
1087 if (conn->makeCallWaiters) {
1088 conn->flags |= RX_CONN_MAKECALL_WAITING;
1089 conn->makeCallWaiters++;
1090 MUTEX_EXIT(&conn->conn_data_lock);
1092 #ifdef RX_ENABLE_LOCKS
1093 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1097 MUTEX_ENTER(&conn->conn_data_lock);
1098 conn->makeCallWaiters--;
1100 MUTEX_EXIT(&conn->conn_data_lock);
1103 for (i = 0; i < RX_MAXCALLS; i++) {
1104 call = conn->call[i];
1106 MUTEX_ENTER(&call->lock);
1107 if (call->state == RX_STATE_DALLY) {
1108 rxi_ResetCall(call, 0);
1109 (*call->callNumber)++;
1112 MUTEX_EXIT(&call->lock);
1114 call = rxi_NewCall(conn, i);
1118 if (i < RX_MAXCALLS) {
1121 MUTEX_ENTER(&conn->conn_data_lock);
1122 conn->flags |= RX_CONN_MAKECALL_WAITING;
1123 conn->makeCallWaiters++;
1124 MUTEX_EXIT(&conn->conn_data_lock);
1126 #ifdef RX_ENABLE_LOCKS
1127 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1131 MUTEX_ENTER(&conn->conn_data_lock);
1132 conn->makeCallWaiters--;
1133 MUTEX_EXIT(&conn->conn_data_lock);
1136 * Wake up anyone else who might be giving us a chance to
1137 * run (see code above that avoids resource starvation).
1139 #ifdef RX_ENABLE_LOCKS
1140 CV_BROADCAST(&conn->conn_call_cv);
1145 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1147 /* Client is initially in send mode */
1148 call->state = RX_STATE_ACTIVE;
1149 call->error = conn->error;
1151 call->mode = RX_MODE_ERROR;
1153 call->mode = RX_MODE_SENDING;
1155 /* remember start time for call in case we have hard dead time limit */
1156 call->queueTime = queueTime;
1157 clock_GetTime(&call->startTime);
1158 hzero(call->bytesSent);
1159 hzero(call->bytesRcvd);
1161 /* Turn on busy protocol. */
1162 rxi_KeepAliveOn(call);
1164 MUTEX_EXIT(&call->lock);
1165 MUTEX_EXIT(&conn->conn_call_lock);
1168 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1169 /* Now, if TQ wasn't cleared earlier, do it now. */
1170 MUTEX_ENTER(&call->lock);
1171 while (call->flags & RX_CALL_TQ_BUSY) {
1172 call->flags |= RX_CALL_TQ_WAIT;
1173 #ifdef RX_ENABLE_LOCKS
1174 CV_WAIT(&call->cv_tq, &call->lock);
1175 #else /* RX_ENABLE_LOCKS */
1176 osi_rxSleep(&call->tq);
1177 #endif /* RX_ENABLE_LOCKS */
1179 if (call->flags & RX_CALL_TQ_CLEARME) {
1180 rxi_ClearTransmitQueue(call, 0);
1181 queue_Init(&call->tq);
1183 MUTEX_EXIT(&call->lock);
1184 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1190 rxi_HasActiveCalls(register struct rx_connection *aconn)
1193 register struct rx_call *tcall;
1197 for (i = 0; i < RX_MAXCALLS; i++) {
1198 if ((tcall = aconn->call[i])) {
1199 if ((tcall->state == RX_STATE_ACTIVE)
1200 || (tcall->state == RX_STATE_PRECALL)) {
1211 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1212 register afs_int32 * aint32s)
1215 register struct rx_call *tcall;
1219 for (i = 0; i < RX_MAXCALLS; i++) {
1220 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1221 aint32s[i] = aconn->callNumber[i] + 1;
1223 aint32s[i] = aconn->callNumber[i];
1230 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1231 register afs_int32 * aint32s)
1234 register struct rx_call *tcall;
1238 for (i = 0; i < RX_MAXCALLS; i++) {
1239 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1240 aconn->callNumber[i] = aint32s[i] - 1;
1242 aconn->callNumber[i] = aint32s[i];
1248 /* Advertise a new service. A service is named locally by a UDP port
1249 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1252 char *serviceName; Name for identification purposes (e.g. the
1253 service name might be used for probing for
1256 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1257 struct rx_securityClass **securityObjects, int nSecurityObjects,
1258 afs_int32(*serviceProc) (struct rx_call * acall))
1260 osi_socket socket = OSI_NULLSOCKET;
1261 register struct rx_service *tservice;
1267 if (serviceId == 0) {
1269 "rx_NewService: service id for service %s is not non-zero.\n",
1276 "rx_NewService: A non-zero port must be specified on this call if a non-zero port was not provided at Rx initialization (service %s).\n",
1284 tservice = rxi_AllocService();
1286 for (i = 0; i < RX_MAX_SERVICES; i++) {
1287 register struct rx_service *service = rx_services[i];
1289 if (port == service->servicePort) {
1290 if (service->serviceId == serviceId) {
1291 /* The identical service has already been
1292 * installed; if the caller was intending to
1293 * change the security classes used by this
1294 * service, he/she loses. */
1296 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1297 serviceName, serviceId, service->serviceName);
1299 rxi_FreeService(tservice);
1302 /* Different service, same port: re-use the socket
1303 * which is bound to the same port */
1304 socket = service->socket;
1307 if (socket == OSI_NULLSOCKET) {
1308 /* If we don't already have a socket (from another
1309 * service on same port) get a new one */
1310 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1311 if (socket == OSI_NULLSOCKET) {
1313 rxi_FreeService(tservice);
1318 service->socket = socket;
1319 service->servicePort = port;
1320 service->serviceId = serviceId;
1321 service->serviceName = serviceName;
1322 service->nSecurityObjects = nSecurityObjects;
1323 service->securityObjects = securityObjects;
1324 service->minProcs = 0;
1325 service->maxProcs = 1;
1326 service->idleDeadTime = 60;
1327 service->connDeadTime = rx_connDeadTime;
1328 service->executeRequestProc = serviceProc;
1329 service->checkReach = 0;
1330 rx_services[i] = service; /* not visible until now */
1336 rxi_FreeService(tservice);
1337 (osi_Msg "rx_NewService: cannot support > %d services\n",
1342 /* Generic request processing loop. This routine should be called
1343 * by the implementation dependent rx_ServerProc. If socketp is
1344 * non-null, it will be set to the file descriptor that this thread
1345 * is now listening on. If socketp is null, this routine will never
1348 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1350 register struct rx_call *call;
1351 register afs_int32 code;
1352 register struct rx_service *tservice = NULL;
1359 call = rx_GetCall(threadID, tservice, socketp);
1360 if (socketp && *socketp != OSI_NULLSOCKET) {
1361 /* We are now a listener thread */
1366 /* if server is restarting( typically smooth shutdown) then do not
1367 * allow any new calls.
1370 if (rx_tranquil && (call != NULL)) {
1374 MUTEX_ENTER(&call->lock);
1376 rxi_CallError(call, RX_RESTARTING);
1377 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1379 MUTEX_EXIT(&call->lock);
1383 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1384 #ifdef RX_ENABLE_LOCKS
1386 #endif /* RX_ENABLE_LOCKS */
1387 afs_termState = AFSOP_STOP_AFS;
1388 afs_osi_Wakeup(&afs_termState);
1389 #ifdef RX_ENABLE_LOCKS
1391 #endif /* RX_ENABLE_LOCKS */
1396 tservice = call->conn->service;
1398 if (tservice->beforeProc)
1399 (*tservice->beforeProc) (call);
1401 code = call->conn->service->executeRequestProc(call);
1403 if (tservice->afterProc)
1404 (*tservice->afterProc) (call, code);
1406 rx_EndCall(call, code);
1407 MUTEX_ENTER(&rx_stats_mutex);
1409 MUTEX_EXIT(&rx_stats_mutex);
1415 rx_WakeupServerProcs(void)
1417 struct rx_serverQueueEntry *np, *tqp;
1421 MUTEX_ENTER(&rx_serverPool_lock);
1423 #ifdef RX_ENABLE_LOCKS
1424 if (rx_waitForPacket)
1425 CV_BROADCAST(&rx_waitForPacket->cv);
1426 #else /* RX_ENABLE_LOCKS */
1427 if (rx_waitForPacket)
1428 osi_rxWakeup(rx_waitForPacket);
1429 #endif /* RX_ENABLE_LOCKS */
1430 MUTEX_ENTER(&freeSQEList_lock);
1431 for (np = rx_FreeSQEList; np; np = tqp) {
1432 tqp = *(struct rx_serverQueueEntry **)np;
1433 #ifdef RX_ENABLE_LOCKS
1434 CV_BROADCAST(&np->cv);
1435 #else /* RX_ENABLE_LOCKS */
1437 #endif /* RX_ENABLE_LOCKS */
1439 MUTEX_EXIT(&freeSQEList_lock);
1440 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1441 #ifdef RX_ENABLE_LOCKS
1442 CV_BROADCAST(&np->cv);
1443 #else /* RX_ENABLE_LOCKS */
1445 #endif /* RX_ENABLE_LOCKS */
1447 MUTEX_EXIT(&rx_serverPool_lock);
1452 * One thing that seems to happen is that all the server threads get
1453 * tied up on some empty or slow call, and then a whole bunch of calls
1454 * arrive at once, using up the packet pool, so now there are more
1455 * empty calls. The most critical resources here are server threads
1456 * and the free packet pool. The "doreclaim" code seems to help in
1457 * general. I think that eventually we arrive in this state: there
1458 * are lots of pending calls which do have all their packets present,
1459 * so they won't be reclaimed, are multi-packet calls, so they won't
1460 * be scheduled until later, and thus are tying up most of the free
1461 * packet pool for a very long time.
1463 * 1. schedule multi-packet calls if all the packets are present.
1464 * Probably CPU-bound operation, useful to return packets to pool.
1465 * Do what if there is a full window, but the last packet isn't here?
1466 * 3. preserve one thread which *only* runs "best" calls, otherwise
1467 * it sleeps and waits for that type of call.
1468 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1469 * the current dataquota business is badly broken. The quota isn't adjusted
1470 * to reflect how many packets are presently queued for a running call.
1471 * So, when we schedule a queued call with a full window of packets queued
1472 * up for it, that *should* free up a window full of packets for other 2d-class
1473 * calls to be able to use from the packet pool. But it doesn't.
1475 * NB. Most of the time, this code doesn't run -- since idle server threads
1476 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1477 * as a new call arrives.
1479 /* Sleep until a call arrives. Returns a pointer to the call, ready
1480 * for an rx_Read. */
1481 #ifdef RX_ENABLE_LOCKS
1483 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1485 struct rx_serverQueueEntry *sq;
1486 register struct rx_call *call = (struct rx_call *)0;
1487 struct rx_service *service = NULL;
1490 MUTEX_ENTER(&freeSQEList_lock);
1492 if ((sq = rx_FreeSQEList)) {
1493 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1494 MUTEX_EXIT(&freeSQEList_lock);
1495 } else { /* otherwise allocate a new one and return that */
1496 MUTEX_EXIT(&freeSQEList_lock);
1497 sq = (struct rx_serverQueueEntry *)
1498 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1499 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1500 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1503 MUTEX_ENTER(&rx_serverPool_lock);
1504 if (cur_service != NULL) {
1505 ReturnToServerPool(cur_service);
1508 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1509 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1511 /* Scan for eligible incoming calls. A call is not eligible
1512 * if the maximum number of calls for its service type are
1513 * already executing */
1514 /* One thread will process calls FCFS (to prevent starvation),
1515 * while the other threads may run ahead looking for calls which
1516 * have all their input data available immediately. This helps
1517 * keep threads from blocking, waiting for data from the client. */
1518 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1519 service = tcall->conn->service;
1520 if (!QuotaOK(service)) {
1523 if (tno == rxi_fcfs_thread_num
1524 || !tcall->queue_item_header.next) {
1525 /* If we're the fcfs thread , then we'll just use
1526 * this call. If we haven't been able to find an optimal
1527 * choice, and we're at the end of the list, then use a
1528 * 2d choice if one has been identified. Otherwise... */
1529 call = (choice2 ? choice2 : tcall);
1530 service = call->conn->service;
1531 } else if (!queue_IsEmpty(&tcall->rq)) {
1532 struct rx_packet *rp;
1533 rp = queue_First(&tcall->rq, rx_packet);
1534 if (rp->header.seq == 1) {
1536 || (rp->header.flags & RX_LAST_PACKET)) {
1538 } else if (rxi_2dchoice && !choice2
1539 && !(tcall->flags & RX_CALL_CLEARED)
1540 && (tcall->rprev > rxi_HardAckRate)) {
1549 ReturnToServerPool(service);
1556 MUTEX_EXIT(&rx_serverPool_lock);
1557 MUTEX_ENTER(&call->lock);
1559 if (call->flags & RX_CALL_WAIT_PROC) {
1560 call->flags &= ~RX_CALL_WAIT_PROC;
1561 MUTEX_ENTER(&rx_stats_mutex);
1563 MUTEX_EXIT(&rx_stats_mutex);
1566 if (call->state != RX_STATE_PRECALL || call->error) {
1567 MUTEX_EXIT(&call->lock);
1568 MUTEX_ENTER(&rx_serverPool_lock);
1569 ReturnToServerPool(service);
1574 if (queue_IsEmpty(&call->rq)
1575 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1576 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1578 CLEAR_CALL_QUEUE_LOCK(call);
1581 /* If there are no eligible incoming calls, add this process
1582 * to the idle server queue, to wait for one */
1586 *socketp = OSI_NULLSOCKET;
1588 sq->socketp = socketp;
1589 queue_Append(&rx_idleServerQueue, sq);
1590 #ifndef AFS_AIX41_ENV
1591 rx_waitForPacket = sq;
1593 rx_waitingForPacket = sq;
1594 #endif /* AFS_AIX41_ENV */
1596 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1598 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1599 MUTEX_EXIT(&rx_serverPool_lock);
1600 return (struct rx_call *)0;
1603 } while (!(call = sq->newcall)
1604 && !(socketp && *socketp != OSI_NULLSOCKET));
1605 MUTEX_EXIT(&rx_serverPool_lock);
1607 MUTEX_ENTER(&call->lock);
1613 MUTEX_ENTER(&freeSQEList_lock);
1614 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1615 rx_FreeSQEList = sq;
1616 MUTEX_EXIT(&freeSQEList_lock);
1619 clock_GetTime(&call->startTime);
1620 call->state = RX_STATE_ACTIVE;
1621 call->mode = RX_MODE_RECEIVING;
1622 #ifdef RX_KERNEL_TRACE
1623 if (ICL_SETACTIVE(afs_iclSetp)) {
1624 int glockOwner = ISAFS_GLOCK();
1627 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1628 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1635 rxi_calltrace(RX_CALL_START, call);
1636 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1637 call->conn->service->servicePort, call->conn->service->serviceId,
1640 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1641 MUTEX_EXIT(&call->lock);
1643 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1648 #else /* RX_ENABLE_LOCKS */
1650 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1652 struct rx_serverQueueEntry *sq;
1653 register struct rx_call *call = (struct rx_call *)0, *choice2;
1654 struct rx_service *service = NULL;
1658 MUTEX_ENTER(&freeSQEList_lock);
1660 if ((sq = rx_FreeSQEList)) {
1661 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1662 MUTEX_EXIT(&freeSQEList_lock);
1663 } else { /* otherwise allocate a new one and return that */
1664 MUTEX_EXIT(&freeSQEList_lock);
1665 sq = (struct rx_serverQueueEntry *)
1666 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1667 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1668 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1670 MUTEX_ENTER(&sq->lock);
1672 if (cur_service != NULL) {
1673 cur_service->nRequestsRunning--;
1674 if (cur_service->nRequestsRunning < cur_service->minProcs)
1678 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1679 register struct rx_call *tcall, *ncall;
1680 /* Scan for eligible incoming calls. A call is not eligible
1681 * if the maximum number of calls for its service type are
1682 * already executing */
1683 /* One thread will process calls FCFS (to prevent starvation),
1684 * while the other threads may run ahead looking for calls which
1685 * have all their input data available immediately. This helps
1686 * keep threads from blocking, waiting for data from the client. */
1687 choice2 = (struct rx_call *)0;
1688 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1689 service = tcall->conn->service;
1690 if (QuotaOK(service)) {
1691 if (tno == rxi_fcfs_thread_num
1692 || !tcall->queue_item_header.next) {
1693 /* If we're the fcfs thread, then we'll just use
1694 * this call. If we haven't been able to find an optimal
1695 * choice, and we're at the end of the list, then use a
1696 * 2d choice if one has been identified. Otherwise... */
1697 call = (choice2 ? choice2 : tcall);
1698 service = call->conn->service;
1699 } else if (!queue_IsEmpty(&tcall->rq)) {
1700 struct rx_packet *rp;
1701 rp = queue_First(&tcall->rq, rx_packet);
1702 if (rp->header.seq == 1
1704 || (rp->header.flags & RX_LAST_PACKET))) {
1706 } else if (rxi_2dchoice && !choice2
1707 && !(tcall->flags & RX_CALL_CLEARED)
1708 && (tcall->rprev > rxi_HardAckRate)) {
1721 /* we can't schedule a call if there's no data!!! */
1722 /* send an ack if there's no data, if we're missing the
1723 * first packet, or we're missing something between first
1724 * and last -- there's a "hole" in the incoming data. */
1725 if (queue_IsEmpty(&call->rq)
1726 || queue_First(&call->rq, rx_packet)->header.seq != 1
1727 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1728 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1730 call->flags &= (~RX_CALL_WAIT_PROC);
1731 service->nRequestsRunning++;
1732 /* just started call in minProcs pool, need fewer to maintain
1734 if (service->nRequestsRunning <= service->minProcs)
1738 /* MUTEX_EXIT(&call->lock); */
1740 /* If there are no eligible incoming calls, add this process
1741 * to the idle server queue, to wait for one */
1744 *socketp = OSI_NULLSOCKET;
1746 sq->socketp = socketp;
1747 queue_Append(&rx_idleServerQueue, sq);
1751 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1753 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1754 return (struct rx_call *)0;
1757 } while (!(call = sq->newcall)
1758 && !(socketp && *socketp != OSI_NULLSOCKET));
1760 MUTEX_EXIT(&sq->lock);
1762 MUTEX_ENTER(&freeSQEList_lock);
1763 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1764 rx_FreeSQEList = sq;
1765 MUTEX_EXIT(&freeSQEList_lock);
1768 clock_GetTime(&call->startTime);
1769 call->state = RX_STATE_ACTIVE;
1770 call->mode = RX_MODE_RECEIVING;
1771 #ifdef RX_KERNEL_TRACE
1772 if (ICL_SETACTIVE(afs_iclSetp)) {
1773 int glockOwner = ISAFS_GLOCK();
1776 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1777 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1784 rxi_calltrace(RX_CALL_START, call);
1785 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1786 call->conn->service->servicePort, call->conn->service->serviceId,
1789 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1796 #endif /* RX_ENABLE_LOCKS */
1800 /* Establish a procedure to be called when a packet arrives for a
1801 * call. This routine will be called at most once after each call,
1802 * and will also be called if there is an error condition on the or
1803 * the call is complete. Used by multi rx to build a selection
1804 * function which determines which of several calls is likely to be a
1805 * good one to read from.
1806 * NOTE: the way this is currently implemented it is probably only a
1807 * good idea to (1) use it immediately after a newcall (clients only)
1808 * and (2) only use it once. Other uses currently void your warranty
1811 rx_SetArrivalProc(register struct rx_call *call,
1812 register void (*proc) (register struct rx_call * call,
1814 register int index),
1815 register VOID * handle, register int arg)
1817 call->arrivalProc = proc;
1818 call->arrivalProcHandle = handle;
1819 call->arrivalProcArg = arg;
1822 /* Call is finished (possibly prematurely). Return rc to the peer, if
1823 * appropriate, and return the final error code from the conversation
1827 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1829 register struct rx_connection *conn = call->conn;
1830 register struct rx_service *service;
1831 register struct rx_packet *tp; /* Temporary packet pointer */
1832 register struct rx_packet *nxp; /* Next packet pointer, for queue_Scan */
1836 dpf(("rx_EndCall(call %x)\n", call));
1839 MUTEX_ENTER(&call->lock);
1841 if (rc == 0 && call->error == 0) {
1842 call->abortCode = 0;
1843 call->abortCount = 0;
1846 call->arrivalProc = (void (*)())0;
1847 if (rc && call->error == 0) {
1848 rxi_CallError(call, rc);
1849 /* Send an abort message to the peer if this error code has
1850 * only just been set. If it was set previously, assume the
1851 * peer has already been sent the error code or will request it
1853 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1855 if (conn->type == RX_SERVER_CONNECTION) {
1856 /* Make sure reply or at least dummy reply is sent */
1857 if (call->mode == RX_MODE_RECEIVING) {
1858 rxi_WriteProc(call, 0, 0);
1860 if (call->mode == RX_MODE_SENDING) {
1861 rxi_FlushWrite(call);
1863 service = conn->service;
1864 rxi_calltrace(RX_CALL_END, call);
1865 /* Call goes to hold state until reply packets are acknowledged */
1866 if (call->tfirst + call->nSoftAcked < call->tnext) {
1867 call->state = RX_STATE_HOLD;
1869 call->state = RX_STATE_DALLY;
1870 rxi_ClearTransmitQueue(call, 0);
1871 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1872 rxevent_Cancel(call->keepAliveEvent, call,
1873 RX_CALL_REFCOUNT_ALIVE);
1875 } else { /* Client connection */
1877 /* Make sure server receives input packets, in the case where
1878 * no reply arguments are expected */
1879 if ((call->mode == RX_MODE_SENDING)
1880 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1881 (void)rxi_ReadProc(call, &dummy, 1);
1884 /* If we had an outstanding delayed ack, be nice to the server
1885 * and force-send it now.
1887 if (call->delayedAckEvent) {
1888 rxevent_Cancel(call->delayedAckEvent, call,
1889 RX_CALL_REFCOUNT_DELAY);
1890 call->delayedAckEvent = NULL;
1891 rxi_SendDelayedAck(NULL, call, NULL);
1894 /* We need to release the call lock since it's lower than the
1895 * conn_call_lock and we don't want to hold the conn_call_lock
1896 * over the rx_ReadProc call. The conn_call_lock needs to be held
1897 * here for the case where rx_NewCall is perusing the calls on
1898 * the connection structure. We don't want to signal until
1899 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1900 * have checked this call, found it active and by the time it
1901 * goes to sleep, will have missed the signal.
1903 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
1904 * there are threads waiting to use the conn object.
1906 MUTEX_EXIT(&call->lock);
1907 MUTEX_ENTER(&conn->conn_call_lock);
1908 MUTEX_ENTER(&call->lock);
1909 MUTEX_ENTER(&conn->conn_data_lock);
1910 conn->flags |= RX_CONN_BUSY;
1911 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1912 if (conn->makeCallWaiters == 0)
1913 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1914 MUTEX_EXIT(&conn->conn_data_lock);
1915 #ifdef RX_ENABLE_LOCKS
1916 CV_BROADCAST(&conn->conn_call_cv);
1921 #ifdef RX_ENABLE_LOCKS
1923 MUTEX_EXIT(&conn->conn_data_lock);
1925 #endif /* RX_ENABLE_LOCKS */
1926 call->state = RX_STATE_DALLY;
1928 error = call->error;
1930 /* currentPacket, nLeft, and NFree must be zeroed here, because
1931 * ResetCall cannot: ResetCall may be called at splnet(), in the
1932 * kernel version, and may interrupt the macros rx_Read or
1933 * rx_Write, which run at normal priority for efficiency. */
1934 if (call->currentPacket) {
1935 queue_Prepend(&call->iovq, call->currentPacket);
1936 call->currentPacket = (struct rx_packet *)0;
1939 call->nLeft = call->nFree = call->curlen = 0;
1941 /* Free any packets from the last call to ReadvProc/WritevProc */
1942 rxi_FreePackets(0, &call->iovq);
1944 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1945 MUTEX_EXIT(&call->lock);
1946 if (conn->type == RX_CLIENT_CONNECTION) {
1947 MUTEX_EXIT(&conn->conn_call_lock);
1948 conn->flags &= ~RX_CONN_BUSY;
1952 * Map errors to the local host's errno.h format.
1954 error = ntoh_syserr_conv(error);
1958 #if !defined(KERNEL)
1960 /* Call this routine when shutting down a server or client (especially
1961 * clients). This will allow Rx to gracefully garbage collect server
1962 * connections, and reduce the number of retries that a server might
1963 * make to a dead client.
1964 * This is not quite right, since some calls may still be ongoing and
1965 * we can't lock them to destroy them. */
1969 register struct rx_connection **conn_ptr, **conn_end;
1973 if (rxinit_status == 1) {
1975 return; /* Already shutdown. */
1977 rxi_DeleteCachedConnections();
1978 if (rx_connHashTable) {
1979 MUTEX_ENTER(&rx_connHashTable_lock);
1980 for (conn_ptr = &rx_connHashTable[0], conn_end =
1981 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
1983 struct rx_connection *conn, *next;
1984 for (conn = *conn_ptr; conn; conn = next) {
1986 if (conn->type == RX_CLIENT_CONNECTION) {
1987 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
1989 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
1990 #ifdef RX_ENABLE_LOCKS
1991 rxi_DestroyConnectionNoLock(conn);
1992 #else /* RX_ENABLE_LOCKS */
1993 rxi_DestroyConnection(conn);
1994 #endif /* RX_ENABLE_LOCKS */
1998 #ifdef RX_ENABLE_LOCKS
1999 while (rx_connCleanup_list) {
2000 struct rx_connection *conn;
2001 conn = rx_connCleanup_list;
2002 rx_connCleanup_list = rx_connCleanup_list->next;
2003 MUTEX_EXIT(&rx_connHashTable_lock);
2004 rxi_CleanupConnection(conn);
2005 MUTEX_ENTER(&rx_connHashTable_lock);
2007 MUTEX_EXIT(&rx_connHashTable_lock);
2008 #endif /* RX_ENABLE_LOCKS */
2017 /* if we wakeup packet waiter too often, can get in loop with two
2018 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2020 rxi_PacketsUnWait(void)
2022 if (!rx_waitingForPackets) {
2026 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2027 return; /* still over quota */
2030 rx_waitingForPackets = 0;
2031 #ifdef RX_ENABLE_LOCKS
2032 CV_BROADCAST(&rx_waitingForPackets_cv);
2034 osi_rxWakeup(&rx_waitingForPackets);
2040 /* ------------------Internal interfaces------------------------- */
2042 /* Return this process's service structure for the
2043 * specified socket and service */
2045 rxi_FindService(register osi_socket socket, register u_short serviceId)
2047 register struct rx_service **sp;
2048 for (sp = &rx_services[0]; *sp; sp++) {
2049 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2055 /* Allocate a call structure, for the indicated channel of the
2056 * supplied connection. The mode and state of the call must be set by
2057 * the caller. Returns the call with mutex locked. */
2059 rxi_NewCall(register struct rx_connection *conn, register int channel)
2061 register struct rx_call *call;
2062 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2063 register struct rx_call *cp; /* Call pointer temp */
2064 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2065 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2067 /* Grab an existing call structure, or allocate a new one.
2068 * Existing call structures are assumed to have been left reset by
2070 MUTEX_ENTER(&rx_freeCallQueue_lock);
2072 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2074 * EXCEPT that the TQ might not yet be cleared out.
2075 * Skip over those with in-use TQs.
2078 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2079 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2085 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2086 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2087 call = queue_First(&rx_freeCallQueue, rx_call);
2088 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2090 MUTEX_ENTER(&rx_stats_mutex);
2091 rx_stats.nFreeCallStructs--;
2092 MUTEX_EXIT(&rx_stats_mutex);
2093 MUTEX_EXIT(&rx_freeCallQueue_lock);
2094 MUTEX_ENTER(&call->lock);
2095 CLEAR_CALL_QUEUE_LOCK(call);
2096 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2097 /* Now, if TQ wasn't cleared earlier, do it now. */
2098 if (call->flags & RX_CALL_TQ_CLEARME) {
2099 rxi_ClearTransmitQueue(call, 0);
2100 queue_Init(&call->tq);
2102 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2103 /* Bind the call to its connection structure */
2105 rxi_ResetCall(call, 1);
2107 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2109 MUTEX_EXIT(&rx_freeCallQueue_lock);
2110 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2111 MUTEX_ENTER(&call->lock);
2112 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2113 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2114 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2116 MUTEX_ENTER(&rx_stats_mutex);
2117 rx_stats.nCallStructs++;
2118 MUTEX_EXIT(&rx_stats_mutex);
2119 /* Initialize once-only items */
2120 queue_Init(&call->tq);
2121 queue_Init(&call->rq);
2122 queue_Init(&call->iovq);
2123 /* Bind the call to its connection structure (prereq for reset) */
2125 rxi_ResetCall(call, 1);
2127 call->channel = channel;
2128 call->callNumber = &conn->callNumber[channel];
2129 /* Note that the next expected call number is retained (in
2130 * conn->callNumber[i]), even if we reallocate the call structure
2132 conn->call[channel] = call;
2133 /* if the channel's never been used (== 0), we should start at 1, otherwise
2134 * the call number is valid from the last time this channel was used */
2135 if (*call->callNumber == 0)
2136 *call->callNumber = 1;
2141 /* A call has been inactive long enough that so we can throw away
2142 * state, including the call structure, which is placed on the call
2144 * Call is locked upon entry.
2145 * haveCTLock set if called from rxi_ReapConnections
2147 #ifdef RX_ENABLE_LOCKS
2149 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2150 #else /* RX_ENABLE_LOCKS */
2152 rxi_FreeCall(register struct rx_call *call)
2153 #endif /* RX_ENABLE_LOCKS */
2155 register int channel = call->channel;
2156 register struct rx_connection *conn = call->conn;
2159 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2160 (*call->callNumber)++;
2161 rxi_ResetCall(call, 0);
2162 call->conn->call[channel] = (struct rx_call *)0;
2164 MUTEX_ENTER(&rx_freeCallQueue_lock);
2165 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2166 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2167 /* A call may be free even though its transmit queue is still in use.
2168 * Since we search the call list from head to tail, put busy calls at
2169 * the head of the list, and idle calls at the tail.
2171 if (call->flags & RX_CALL_TQ_BUSY)
2172 queue_Prepend(&rx_freeCallQueue, call);
2174 queue_Append(&rx_freeCallQueue, call);
2175 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2176 queue_Append(&rx_freeCallQueue, call);
2177 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2178 MUTEX_ENTER(&rx_stats_mutex);
2179 rx_stats.nFreeCallStructs++;
2180 MUTEX_EXIT(&rx_stats_mutex);
2182 MUTEX_EXIT(&rx_freeCallQueue_lock);
2184 /* Destroy the connection if it was previously slated for
2185 * destruction, i.e. the Rx client code previously called
2186 * rx_DestroyConnection (client connections), or
2187 * rxi_ReapConnections called the same routine (server
2188 * connections). Only do this, however, if there are no
2189 * outstanding calls. Note that for fine grain locking, there appears
2190 * to be a deadlock in that rxi_FreeCall has a call locked and
2191 * DestroyConnectionNoLock locks each call in the conn. But note a
2192 * few lines up where we have removed this call from the conn.
2193 * If someone else destroys a connection, they either have no
2194 * call lock held or are going through this section of code.
2196 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2197 MUTEX_ENTER(&conn->conn_data_lock);
2199 MUTEX_EXIT(&conn->conn_data_lock);
2200 #ifdef RX_ENABLE_LOCKS
2202 rxi_DestroyConnectionNoLock(conn);
2204 rxi_DestroyConnection(conn);
2205 #else /* RX_ENABLE_LOCKS */
2206 rxi_DestroyConnection(conn);
2207 #endif /* RX_ENABLE_LOCKS */
2211 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2213 rxi_Alloc(register size_t size)
2217 MUTEX_ENTER(&rx_stats_mutex);
2219 rxi_Allocsize += size;
2220 MUTEX_EXIT(&rx_stats_mutex);
2222 p = (char *)osi_Alloc(size);
2225 osi_Panic("rxi_Alloc error");
2231 rxi_Free(void *addr, register size_t size)
2233 MUTEX_ENTER(&rx_stats_mutex);
2235 rxi_Allocsize -= size;
2236 MUTEX_EXIT(&rx_stats_mutex);
2238 osi_Free(addr, size);
2241 /* Find the peer process represented by the supplied (host,port)
2242 * combination. If there is no appropriate active peer structure, a
2243 * new one will be allocated and initialized
2244 * The origPeer, if set, is a pointer to a peer structure on which the
2245 * refcount will be be decremented. This is used to replace the peer
2246 * structure hanging off a connection structure */
2248 rxi_FindPeer(register afs_uint32 host, register u_short port,
2249 struct rx_peer *origPeer, int create)
2251 register struct rx_peer *pp;
2253 hashIndex = PEER_HASH(host, port);
2254 MUTEX_ENTER(&rx_peerHashTable_lock);
2255 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2256 if ((pp->host == host) && (pp->port == port))
2261 pp = rxi_AllocPeer(); /* This bzero's *pp */
2262 pp->host = host; /* set here or in InitPeerParams is zero */
2264 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2265 queue_Init(&pp->congestionQueue);
2266 queue_Init(&pp->rpcStats);
2267 pp->next = rx_peerHashTable[hashIndex];
2268 rx_peerHashTable[hashIndex] = pp;
2269 rxi_InitPeerParams(pp);
2270 MUTEX_ENTER(&rx_stats_mutex);
2271 rx_stats.nPeerStructs++;
2272 MUTEX_EXIT(&rx_stats_mutex);
2279 origPeer->refCount--;
2280 MUTEX_EXIT(&rx_peerHashTable_lock);
2285 /* Find the connection at (host, port) started at epoch, and with the
2286 * given connection id. Creates the server connection if necessary.
2287 * The type specifies whether a client connection or a server
2288 * connection is desired. In both cases, (host, port) specify the
2289 * peer's (host, pair) pair. Client connections are not made
2290 * automatically by this routine. The parameter socket gives the
2291 * socket descriptor on which the packet was received. This is used,
2292 * in the case of server connections, to check that *new* connections
2293 * come via a valid (port, serviceId). Finally, the securityIndex
2294 * parameter must match the existing index for the connection. If a
2295 * server connection is created, it will be created using the supplied
2296 * index, if the index is valid for this service */
2297 struct rx_connection *
2298 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2299 register u_short port, u_short serviceId, afs_uint32 cid,
2300 afs_uint32 epoch, int type, u_int securityIndex)
2302 int hashindex, flag;
2303 register struct rx_connection *conn;
2304 hashindex = CONN_HASH(host, port, cid, epoch, type);
2305 MUTEX_ENTER(&rx_connHashTable_lock);
2306 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2307 rx_connHashTable[hashindex],
2310 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2311 && (epoch == conn->epoch)) {
2312 register struct rx_peer *pp = conn->peer;
2313 if (securityIndex != conn->securityIndex) {
2314 /* this isn't supposed to happen, but someone could forge a packet
2315 * like this, and there seems to be some CM bug that makes this
2316 * happen from time to time -- in which case, the fileserver
2318 MUTEX_EXIT(&rx_connHashTable_lock);
2319 return (struct rx_connection *)0;
2321 if (pp->host == host && pp->port == port)
2323 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2325 /* So what happens when it's a callback connection? */
2326 if ( /*type == RX_CLIENT_CONNECTION && */
2327 (conn->epoch & 0x80000000))
2331 /* the connection rxLastConn that was used the last time is not the
2332 ** one we are looking for now. Hence, start searching in the hash */
2334 conn = rx_connHashTable[hashindex];
2339 struct rx_service *service;
2340 if (type == RX_CLIENT_CONNECTION) {
2341 MUTEX_EXIT(&rx_connHashTable_lock);
2342 return (struct rx_connection *)0;
2344 service = rxi_FindService(socket, serviceId);
2345 if (!service || (securityIndex >= service->nSecurityObjects)
2346 || (service->securityObjects[securityIndex] == 0)) {
2347 MUTEX_EXIT(&rx_connHashTable_lock);
2348 return (struct rx_connection *)0;
2350 conn = rxi_AllocConnection(); /* This bzero's the connection */
2351 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2352 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2353 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2354 conn->next = rx_connHashTable[hashindex];
2355 rx_connHashTable[hashindex] = conn;
2356 conn->peer = rxi_FindPeer(host, port, 0, 1);
2357 conn->type = RX_SERVER_CONNECTION;
2358 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2359 conn->epoch = epoch;
2360 conn->cid = cid & RX_CIDMASK;
2361 /* conn->serial = conn->lastSerial = 0; */
2362 /* conn->timeout = 0; */
2363 conn->ackRate = RX_FAST_ACK_RATE;
2364 conn->service = service;
2365 conn->serviceId = serviceId;
2366 conn->securityIndex = securityIndex;
2367 conn->securityObject = service->securityObjects[securityIndex];
2368 conn->nSpecific = 0;
2369 conn->specific = NULL;
2370 rx_SetConnDeadTime(conn, service->connDeadTime);
2371 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2372 /* Notify security object of the new connection */
2373 RXS_NewConnection(conn->securityObject, conn);
2374 /* XXXX Connection timeout? */
2375 if (service->newConnProc)
2376 (*service->newConnProc) (conn);
2377 MUTEX_ENTER(&rx_stats_mutex);
2378 rx_stats.nServerConns++;
2379 MUTEX_EXIT(&rx_stats_mutex);
2382 MUTEX_ENTER(&conn->conn_data_lock);
2384 MUTEX_EXIT(&conn->conn_data_lock);
2386 rxLastConn = conn; /* store this connection as the last conn used */
2387 MUTEX_EXIT(&rx_connHashTable_lock);
2391 /* There are two packet tracing routines available for testing and monitoring
2392 * Rx. One is called just after every packet is received and the other is
2393 * called just before every packet is sent. Received packets, have had their
2394 * headers decoded, and packets to be sent have not yet had their headers
2395 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2396 * containing the network address. Both can be modified. The return value, if
2397 * non-zero, indicates that the packet should be dropped. */
2399 int (*rx_justReceived) () = 0;
2400 int (*rx_almostSent) () = 0;
2402 /* A packet has been received off the interface. Np is the packet, socket is
2403 * the socket number it was received from (useful in determining which service
2404 * this packet corresponds to), and (host, port) reflect the host,port of the
2405 * sender. This call returns the packet to the caller if it is finished with
2406 * it, rather than de-allocating it, just as a small performance hack */
2409 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2410 afs_uint32 host, u_short port, int *tnop,
2411 struct rx_call **newcallp)
2413 register struct rx_call *call;
2414 register struct rx_connection *conn;
2416 afs_uint32 currentCallNumber;
2422 struct rx_packet *tnp;
2425 /* We don't print out the packet until now because (1) the time may not be
2426 * accurate enough until now in the lwp implementation (rx_Listener only gets
2427 * the time after the packet is read) and (2) from a protocol point of view,
2428 * this is the first time the packet has been seen */
2429 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2430 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2431 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2432 np->header.serial, packetType, host, port, np->header.serviceId,
2433 np->header.epoch, np->header.cid, np->header.callNumber,
2434 np->header.seq, np->header.flags, np));
2437 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2438 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2441 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2442 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2445 /* If an input tracer function is defined, call it with the packet and
2446 * network address. Note this function may modify its arguments. */
2447 if (rx_justReceived) {
2448 struct sockaddr_in addr;
2450 addr.sin_family = AF_INET;
2451 addr.sin_port = port;
2452 addr.sin_addr.s_addr = host;
2453 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2454 addr.sin_len = sizeof(addr);
2455 #endif /* AFS_OSF_ENV */
2456 drop = (*rx_justReceived) (np, &addr);
2457 /* drop packet if return value is non-zero */
2460 port = addr.sin_port; /* in case fcn changed addr */
2461 host = addr.sin_addr.s_addr;
2465 /* If packet was not sent by the client, then *we* must be the client */
2466 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2467 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2469 /* Find the connection (or fabricate one, if we're the server & if
2470 * necessary) associated with this packet */
2472 rxi_FindConnection(socket, host, port, np->header.serviceId,
2473 np->header.cid, np->header.epoch, type,
2474 np->header.securityIndex);
2477 /* If no connection found or fabricated, just ignore the packet.
2478 * (An argument could be made for sending an abort packet for
2483 MUTEX_ENTER(&conn->conn_data_lock);
2484 if (conn->maxSerial < np->header.serial)
2485 conn->maxSerial = np->header.serial;
2486 MUTEX_EXIT(&conn->conn_data_lock);
2488 /* If the connection is in an error state, send an abort packet and ignore
2489 * the incoming packet */
2491 /* Don't respond to an abort packet--we don't want loops! */
2492 MUTEX_ENTER(&conn->conn_data_lock);
2493 if (np->header.type != RX_PACKET_TYPE_ABORT)
2494 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2496 MUTEX_EXIT(&conn->conn_data_lock);
2500 /* Check for connection-only requests (i.e. not call specific). */
2501 if (np->header.callNumber == 0) {
2502 switch (np->header.type) {
2503 case RX_PACKET_TYPE_ABORT:
2504 /* What if the supplied error is zero? */
2505 rxi_ConnectionError(conn, ntohl(rx_GetInt32(np, 0)));
2506 MUTEX_ENTER(&conn->conn_data_lock);
2508 MUTEX_EXIT(&conn->conn_data_lock);
2510 case RX_PACKET_TYPE_CHALLENGE:
2511 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2512 MUTEX_ENTER(&conn->conn_data_lock);
2514 MUTEX_EXIT(&conn->conn_data_lock);
2516 case RX_PACKET_TYPE_RESPONSE:
2517 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2518 MUTEX_ENTER(&conn->conn_data_lock);
2520 MUTEX_EXIT(&conn->conn_data_lock);
2522 case RX_PACKET_TYPE_PARAMS:
2523 case RX_PACKET_TYPE_PARAMS + 1:
2524 case RX_PACKET_TYPE_PARAMS + 2:
2525 /* ignore these packet types for now */
2526 MUTEX_ENTER(&conn->conn_data_lock);
2528 MUTEX_EXIT(&conn->conn_data_lock);
2533 /* Should not reach here, unless the peer is broken: send an
2535 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2536 MUTEX_ENTER(&conn->conn_data_lock);
2537 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2539 MUTEX_EXIT(&conn->conn_data_lock);
2544 channel = np->header.cid & RX_CHANNELMASK;
2545 call = conn->call[channel];
2546 #ifdef RX_ENABLE_LOCKS
2548 MUTEX_ENTER(&call->lock);
2549 /* Test to see if call struct is still attached to conn. */
2550 if (call != conn->call[channel]) {
2552 MUTEX_EXIT(&call->lock);
2553 if (type == RX_SERVER_CONNECTION) {
2554 call = conn->call[channel];
2555 /* If we started with no call attached and there is one now,
2556 * another thread is also running this routine and has gotten
2557 * the connection channel. We should drop this packet in the tests
2558 * below. If there was a call on this connection and it's now
2559 * gone, then we'll be making a new call below.
2560 * If there was previously a call and it's now different then
2561 * the old call was freed and another thread running this routine
2562 * has created a call on this channel. One of these two threads
2563 * has a packet for the old call and the code below handles those
2567 MUTEX_ENTER(&call->lock);
2569 /* This packet can't be for this call. If the new call address is
2570 * 0 then no call is running on this channel. If there is a call
2571 * then, since this is a client connection we're getting data for
2572 * it must be for the previous call.
2574 MUTEX_ENTER(&rx_stats_mutex);
2575 rx_stats.spuriousPacketsRead++;
2576 MUTEX_EXIT(&rx_stats_mutex);
2577 MUTEX_ENTER(&conn->conn_data_lock);
2579 MUTEX_EXIT(&conn->conn_data_lock);
2584 currentCallNumber = conn->callNumber[channel];
2586 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2587 if (np->header.callNumber < currentCallNumber) {
2588 MUTEX_ENTER(&rx_stats_mutex);
2589 rx_stats.spuriousPacketsRead++;
2590 MUTEX_EXIT(&rx_stats_mutex);
2591 #ifdef RX_ENABLE_LOCKS
2593 MUTEX_EXIT(&call->lock);
2595 MUTEX_ENTER(&conn->conn_data_lock);
2597 MUTEX_EXIT(&conn->conn_data_lock);
2601 MUTEX_ENTER(&conn->conn_call_lock);
2602 call = rxi_NewCall(conn, channel);
2603 MUTEX_EXIT(&conn->conn_call_lock);
2604 *call->callNumber = np->header.callNumber;
2605 call->state = RX_STATE_PRECALL;
2606 clock_GetTime(&call->queueTime);
2607 hzero(call->bytesSent);
2608 hzero(call->bytesRcvd);
2610 * If the number of queued calls exceeds the overload
2611 * threshold then abort this call.
2613 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2614 struct rx_packet *tp;
2616 rxi_CallError(call, rx_BusyError);
2617 tp = rxi_SendCallAbort(call, np, 1, 0);
2618 MUTEX_EXIT(&call->lock);
2619 MUTEX_ENTER(&conn->conn_data_lock);
2621 MUTEX_EXIT(&conn->conn_data_lock);
2622 MUTEX_ENTER(&rx_stats_mutex);
2624 MUTEX_EXIT(&rx_stats_mutex);
2627 rxi_KeepAliveOn(call);
2628 } else if (np->header.callNumber != currentCallNumber) {
2629 /* Wait until the transmit queue is idle before deciding
2630 * whether to reset the current call. Chances are that the
2631 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2634 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2635 while ((call->state == RX_STATE_ACTIVE)
2636 && (call->flags & RX_CALL_TQ_BUSY)) {
2637 call->flags |= RX_CALL_TQ_WAIT;
2638 #ifdef RX_ENABLE_LOCKS
2639 CV_WAIT(&call->cv_tq, &call->lock);
2640 #else /* RX_ENABLE_LOCKS */
2641 osi_rxSleep(&call->tq);
2642 #endif /* RX_ENABLE_LOCKS */
2644 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2645 /* If the new call cannot be taken right now send a busy and set
2646 * the error condition in this call, so that it terminates as
2647 * quickly as possible */
2648 if (call->state == RX_STATE_ACTIVE) {
2649 struct rx_packet *tp;
2651 rxi_CallError(call, RX_CALL_DEAD);
2652 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2654 MUTEX_EXIT(&call->lock);
2655 MUTEX_ENTER(&conn->conn_data_lock);
2657 MUTEX_EXIT(&conn->conn_data_lock);
2660 rxi_ResetCall(call, 0);
2661 *call->callNumber = np->header.callNumber;
2662 call->state = RX_STATE_PRECALL;
2663 clock_GetTime(&call->queueTime);
2664 hzero(call->bytesSent);
2665 hzero(call->bytesRcvd);
2667 * If the number of queued calls exceeds the overload
2668 * threshold then abort this call.
2670 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2671 struct rx_packet *tp;
2673 rxi_CallError(call, rx_BusyError);
2674 tp = rxi_SendCallAbort(call, np, 1, 0);
2675 MUTEX_EXIT(&call->lock);
2676 MUTEX_ENTER(&conn->conn_data_lock);
2678 MUTEX_EXIT(&conn->conn_data_lock);
2679 MUTEX_ENTER(&rx_stats_mutex);
2681 MUTEX_EXIT(&rx_stats_mutex);
2684 rxi_KeepAliveOn(call);
2686 /* Continuing call; do nothing here. */
2688 } else { /* we're the client */
2689 /* Ignore all incoming acknowledgements for calls in DALLY state */
2690 if (call && (call->state == RX_STATE_DALLY)
2691 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2692 MUTEX_ENTER(&rx_stats_mutex);
2693 rx_stats.ignorePacketDally++;
2694 MUTEX_EXIT(&rx_stats_mutex);
2695 #ifdef RX_ENABLE_LOCKS
2697 MUTEX_EXIT(&call->lock);
2700 MUTEX_ENTER(&conn->conn_data_lock);
2702 MUTEX_EXIT(&conn->conn_data_lock);
2706 /* Ignore anything that's not relevant to the current call. If there
2707 * isn't a current call, then no packet is relevant. */
2708 if (!call || (np->header.callNumber != currentCallNumber)) {
2709 MUTEX_ENTER(&rx_stats_mutex);
2710 rx_stats.spuriousPacketsRead++;
2711 MUTEX_EXIT(&rx_stats_mutex);
2712 #ifdef RX_ENABLE_LOCKS
2714 MUTEX_EXIT(&call->lock);
2717 MUTEX_ENTER(&conn->conn_data_lock);
2719 MUTEX_EXIT(&conn->conn_data_lock);
2722 /* If the service security object index stamped in the packet does not
2723 * match the connection's security index, ignore the packet */
2724 if (np->header.securityIndex != conn->securityIndex) {
2725 #ifdef RX_ENABLE_LOCKS
2726 MUTEX_EXIT(&call->lock);
2728 MUTEX_ENTER(&conn->conn_data_lock);
2730 MUTEX_EXIT(&conn->conn_data_lock);
2734 /* If we're receiving the response, then all transmit packets are
2735 * implicitly acknowledged. Get rid of them. */
2736 if (np->header.type == RX_PACKET_TYPE_DATA) {
2737 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2738 /* XXX Hack. Because we must release the global rx lock when
2739 * sending packets (osi_NetSend) we drop all acks while we're
2740 * traversing the tq in rxi_Start sending packets out because
2741 * packets may move to the freePacketQueue as result of being here!
2742 * So we drop these packets until we're safely out of the
2743 * traversing. Really ugly!
2744 * For fine grain RX locking, we set the acked field in the
2745 * packets and let rxi_Start remove them from the transmit queue.
2747 if (call->flags & RX_CALL_TQ_BUSY) {
2748 #ifdef RX_ENABLE_LOCKS
2749 rxi_SetAcksInTransmitQueue(call);
2752 return np; /* xmitting; drop packet */
2755 rxi_ClearTransmitQueue(call, 0);
2757 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2758 rxi_ClearTransmitQueue(call, 0);
2759 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2761 if (np->header.type == RX_PACKET_TYPE_ACK) {
2762 /* now check to see if this is an ack packet acknowledging that the
2763 * server actually *lost* some hard-acked data. If this happens we
2764 * ignore this packet, as it may indicate that the server restarted in
2765 * the middle of a call. It is also possible that this is an old ack
2766 * packet. We don't abort the connection in this case, because this
2767 * *might* just be an old ack packet. The right way to detect a server
2768 * restart in the midst of a call is to notice that the server epoch
2770 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2771 * XXX unacknowledged. I think that this is off-by-one, but
2772 * XXX I don't dare change it just yet, since it will
2773 * XXX interact badly with the server-restart detection
2774 * XXX code in receiveackpacket. */
2775 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2776 MUTEX_ENTER(&rx_stats_mutex);
2777 rx_stats.spuriousPacketsRead++;
2778 MUTEX_EXIT(&rx_stats_mutex);
2779 MUTEX_EXIT(&call->lock);
2780 MUTEX_ENTER(&conn->conn_data_lock);
2782 MUTEX_EXIT(&conn->conn_data_lock);
2786 } /* else not a data packet */
2789 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2790 /* Set remote user defined status from packet */
2791 call->remoteStatus = np->header.userStatus;
2793 /* Note the gap between the expected next packet and the actual
2794 * packet that arrived, when the new packet has a smaller serial number
2795 * than expected. Rioses frequently reorder packets all by themselves,
2796 * so this will be quite important with very large window sizes.
2797 * Skew is checked against 0 here to avoid any dependence on the type of
2798 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2800 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2801 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2802 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2804 MUTEX_ENTER(&conn->conn_data_lock);
2805 skew = conn->lastSerial - np->header.serial;
2806 conn->lastSerial = np->header.serial;
2807 MUTEX_EXIT(&conn->conn_data_lock);
2809 register struct rx_peer *peer;
2811 if (skew > peer->inPacketSkew) {
2812 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2814 peer->inPacketSkew = skew;
2818 /* Now do packet type-specific processing */
2819 switch (np->header.type) {
2820 case RX_PACKET_TYPE_DATA:
2821 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2824 case RX_PACKET_TYPE_ACK:
2825 /* Respond immediately to ack packets requesting acknowledgement
2827 if (np->header.flags & RX_REQUEST_ACK) {
2829 (void)rxi_SendCallAbort(call, 0, 1, 0);
2831 (void)rxi_SendAck(call, 0, np->header.serial,
2832 RX_ACK_PING_RESPONSE, 1);
2834 np = rxi_ReceiveAckPacket(call, np, 1);
2836 case RX_PACKET_TYPE_ABORT:
2837 /* An abort packet: reset the connection, passing the error up to
2839 /* What if error is zero? */
2840 rxi_CallError(call, ntohl(*(afs_int32 *) rx_DataOf(np)));
2842 case RX_PACKET_TYPE_BUSY:
2845 case RX_PACKET_TYPE_ACKALL:
2846 /* All packets acknowledged, so we can drop all packets previously
2847 * readied for sending */
2848 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2849 /* XXX Hack. We because we can't release the global rx lock when
2850 * sending packets (osi_NetSend) we drop all ack pkts while we're
2851 * traversing the tq in rxi_Start sending packets out because
2852 * packets may move to the freePacketQueue as result of being
2853 * here! So we drop these packets until we're safely out of the
2854 * traversing. Really ugly!
2855 * For fine grain RX locking, we set the acked field in the packets
2856 * and let rxi_Start remove the packets from the transmit queue.
2858 if (call->flags & RX_CALL_TQ_BUSY) {
2859 #ifdef RX_ENABLE_LOCKS
2860 rxi_SetAcksInTransmitQueue(call);
2862 #else /* RX_ENABLE_LOCKS */
2864 return np; /* xmitting; drop packet */
2865 #endif /* RX_ENABLE_LOCKS */
2867 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2868 rxi_ClearTransmitQueue(call, 0);
2871 /* Should not reach here, unless the peer is broken: send an abort
2873 rxi_CallError(call, RX_PROTOCOL_ERROR);
2874 np = rxi_SendCallAbort(call, np, 1, 0);
2877 /* Note when this last legitimate packet was received, for keep-alive
2878 * processing. Note, we delay getting the time until now in the hope that
2879 * the packet will be delivered to the user before any get time is required
2880 * (if not, then the time won't actually be re-evaluated here). */
2881 call->lastReceiveTime = clock_Sec();
2882 MUTEX_EXIT(&call->lock);
2883 MUTEX_ENTER(&conn->conn_data_lock);
2885 MUTEX_EXIT(&conn->conn_data_lock);
2889 /* return true if this is an "interesting" connection from the point of view
2890 of someone trying to debug the system */
2892 rxi_IsConnInteresting(struct rx_connection *aconn)
2895 register struct rx_call *tcall;
2897 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2899 for (i = 0; i < RX_MAXCALLS; i++) {
2900 tcall = aconn->call[i];
2902 if ((tcall->state == RX_STATE_PRECALL)
2903 || (tcall->state == RX_STATE_ACTIVE))
2905 if ((tcall->mode == RX_MODE_SENDING)
2906 || (tcall->mode == RX_MODE_RECEIVING))
2914 /* if this is one of the last few packets AND it wouldn't be used by the
2915 receiving call to immediately satisfy a read request, then drop it on
2916 the floor, since accepting it might prevent a lock-holding thread from
2917 making progress in its reading. If a call has been cleared while in
2918 the precall state then ignore all subsequent packets until the call
2919 is assigned to a thread. */
2922 TooLow(struct rx_packet *ap, struct rx_call *acall)
2925 MUTEX_ENTER(&rx_stats_mutex);
2926 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2927 && (acall->state == RX_STATE_PRECALL))
2928 || ((rx_nFreePackets < rxi_dataQuota + 2)
2929 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2930 && (acall->flags & RX_CALL_READER_WAIT)))) {
2933 MUTEX_EXIT(&rx_stats_mutex);
2939 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2940 struct rx_call *acall)
2942 struct rx_call *call = acall;
2946 MUTEX_ENTER(&conn->conn_data_lock);
2947 conn->checkReachEvent = NULL;
2948 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2951 MUTEX_EXIT(&conn->conn_data_lock);
2955 MUTEX_ENTER(&conn->conn_call_lock);
2956 MUTEX_ENTER(&conn->conn_data_lock);
2957 for (i = 0; i < RX_MAXCALLS; i++) {
2958 struct rx_call *tc = conn->call[i];
2959 if (tc && tc->state == RX_STATE_PRECALL) {
2965 /* Indicate that rxi_CheckReachEvent is no longer running by
2966 * clearing the flag. Must be atomic under conn_data_lock to
2967 * avoid a new call slipping by: rxi_CheckConnReach holds
2968 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
2970 conn->flags &= ~RX_CONN_ATTACHWAIT;
2971 MUTEX_EXIT(&conn->conn_data_lock);
2972 MUTEX_EXIT(&conn->conn_call_lock);
2977 MUTEX_ENTER(&call->lock);
2978 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
2980 MUTEX_EXIT(&call->lock);
2982 clock_GetTime(&when);
2983 when.sec += RX_CHECKREACH_TIMEOUT;
2984 MUTEX_ENTER(&conn->conn_data_lock);
2985 if (!conn->checkReachEvent) {
2987 conn->checkReachEvent =
2988 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
2990 MUTEX_EXIT(&conn->conn_data_lock);
2996 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
2998 struct rx_service *service = conn->service;
2999 struct rx_peer *peer = conn->peer;
3000 afs_uint32 now, lastReach;
3002 if (service->checkReach == 0)
3006 MUTEX_ENTER(&peer->peer_lock);
3007 lastReach = peer->lastReachTime;
3008 MUTEX_EXIT(&peer->peer_lock);
3009 if (now - lastReach < RX_CHECKREACH_TTL)
3012 MUTEX_ENTER(&conn->conn_data_lock);
3013 if (conn->flags & RX_CONN_ATTACHWAIT) {
3014 MUTEX_EXIT(&conn->conn_data_lock);
3017 conn->flags |= RX_CONN_ATTACHWAIT;
3018 MUTEX_EXIT(&conn->conn_data_lock);
3019 if (!conn->checkReachEvent)
3020 rxi_CheckReachEvent(NULL, conn, call);
3025 /* try to attach call, if authentication is complete */
3027 TryAttach(register struct rx_call *acall, register osi_socket socket,
3028 register int *tnop, register struct rx_call **newcallp,
3031 struct rx_connection *conn = acall->conn;
3033 if (conn->type == RX_SERVER_CONNECTION
3034 && acall->state == RX_STATE_PRECALL) {
3035 /* Don't attach until we have any req'd. authentication. */
3036 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3037 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3038 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3039 /* Note: this does not necessarily succeed; there
3040 * may not any proc available
3043 rxi_ChallengeOn(acall->conn);
3048 /* A data packet has been received off the interface. This packet is
3049 * appropriate to the call (the call is in the right state, etc.). This
3050 * routine can return a packet to the caller, for re-use */
3053 rxi_ReceiveDataPacket(register struct rx_call *call,
3054 register struct rx_packet *np, int istack,
3055 osi_socket socket, afs_uint32 host, u_short port,
3056 int *tnop, struct rx_call **newcallp)
3058 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3062 afs_uint32 seq, serial, flags;
3064 struct rx_packet *tnp;
3066 MUTEX_ENTER(&rx_stats_mutex);
3067 rx_stats.dataPacketsRead++;
3068 MUTEX_EXIT(&rx_stats_mutex);
3071 /* If there are no packet buffers, drop this new packet, unless we can find
3072 * packet buffers from inactive calls */
3074 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3075 MUTEX_ENTER(&rx_freePktQ_lock);
3076 rxi_NeedMorePackets = TRUE;
3077 MUTEX_EXIT(&rx_freePktQ_lock);
3078 MUTEX_ENTER(&rx_stats_mutex);
3079 rx_stats.noPacketBuffersOnRead++;
3080 MUTEX_EXIT(&rx_stats_mutex);
3081 call->rprev = np->header.serial;
3082 rxi_calltrace(RX_TRACE_DROP, call);
3083 dpf(("packet %x dropped on receipt - quota problems", np));
3085 rxi_ClearReceiveQueue(call);
3086 clock_GetTime(&when);
3087 clock_Add(&when, &rx_softAckDelay);
3088 if (!call->delayedAckEvent
3089 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3090 rxevent_Cancel(call->delayedAckEvent, call,
3091 RX_CALL_REFCOUNT_DELAY);
3092 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3093 call->delayedAckEvent =
3094 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3096 /* we've damaged this call already, might as well do it in. */
3102 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3103 * packet is one of several packets transmitted as a single
3104 * datagram. Do not send any soft or hard acks until all packets
3105 * in a jumbogram have been processed. Send negative acks right away.
3107 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3108 /* tnp is non-null when there are more packets in the
3109 * current jumbo gram */
3116 seq = np->header.seq;
3117 serial = np->header.serial;
3118 flags = np->header.flags;
3120 /* If the call is in an error state, send an abort message */
3122 return rxi_SendCallAbort(call, np, istack, 0);
3124 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3125 * AFS 3.5 jumbogram. */
3126 if (flags & RX_JUMBO_PACKET) {
3127 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3132 if (np->header.spare != 0) {
3133 MUTEX_ENTER(&call->conn->conn_data_lock);
3134 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3135 MUTEX_EXIT(&call->conn->conn_data_lock);
3138 /* The usual case is that this is the expected next packet */
3139 if (seq == call->rnext) {
3141 /* Check to make sure it is not a duplicate of one already queued */
3142 if (queue_IsNotEmpty(&call->rq)
3143 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3144 MUTEX_ENTER(&rx_stats_mutex);
3145 rx_stats.dupPacketsRead++;
3146 MUTEX_EXIT(&rx_stats_mutex);
3147 dpf(("packet %x dropped on receipt - duplicate", np));
3148 rxevent_Cancel(call->delayedAckEvent, call,
3149 RX_CALL_REFCOUNT_DELAY);
3150 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3156 /* It's the next packet. Stick it on the receive queue
3157 * for this call. Set newPackets to make sure we wake
3158 * the reader once all packets have been processed */
3159 queue_Prepend(&call->rq, np);
3161 np = NULL; /* We can't use this anymore */
3164 /* If an ack is requested then set a flag to make sure we
3165 * send an acknowledgement for this packet */
3166 if (flags & RX_REQUEST_ACK) {
3167 ackNeeded = RX_ACK_REQUESTED;
3170 /* Keep track of whether we have received the last packet */
3171 if (flags & RX_LAST_PACKET) {
3172 call->flags |= RX_CALL_HAVE_LAST;
3176 /* Check whether we have all of the packets for this call */
3177 if (call->flags & RX_CALL_HAVE_LAST) {
3178 afs_uint32 tseq; /* temporary sequence number */
3179 struct rx_packet *tp; /* Temporary packet pointer */
3180 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3182 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3183 if (tseq != tp->header.seq)
3185 if (tp->header.flags & RX_LAST_PACKET) {
3186 call->flags |= RX_CALL_RECEIVE_DONE;
3193 /* Provide asynchronous notification for those who want it
3194 * (e.g. multi rx) */
3195 if (call->arrivalProc) {
3196 (*call->arrivalProc) (call, call->arrivalProcHandle,
3197 call->arrivalProcArg);
3198 call->arrivalProc = (void (*)())0;
3201 /* Update last packet received */
3204 /* If there is no server process serving this call, grab
3205 * one, if available. We only need to do this once. If a
3206 * server thread is available, this thread becomes a server
3207 * thread and the server thread becomes a listener thread. */
3209 TryAttach(call, socket, tnop, newcallp, 0);
3212 /* This is not the expected next packet. */
3214 /* Determine whether this is a new or old packet, and if it's
3215 * a new one, whether it fits into the current receive window.
3216 * Also figure out whether the packet was delivered in sequence.
3217 * We use the prev variable to determine whether the new packet
3218 * is the successor of its immediate predecessor in the
3219 * receive queue, and the missing flag to determine whether
3220 * any of this packets predecessors are missing. */
3222 afs_uint32 prev; /* "Previous packet" sequence number */
3223 struct rx_packet *tp; /* Temporary packet pointer */
3224 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3225 int missing; /* Are any predecessors missing? */
3227 /* If the new packet's sequence number has been sent to the
3228 * application already, then this is a duplicate */
3229 if (seq < call->rnext) {
3230 MUTEX_ENTER(&rx_stats_mutex);
3231 rx_stats.dupPacketsRead++;
3232 MUTEX_EXIT(&rx_stats_mutex);
3233 rxevent_Cancel(call->delayedAckEvent, call,
3234 RX_CALL_REFCOUNT_DELAY);
3235 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3241 /* If the sequence number is greater than what can be
3242 * accomodated by the current window, then send a negative
3243 * acknowledge and drop the packet */
3244 if ((call->rnext + call->rwind) <= seq) {
3245 rxevent_Cancel(call->delayedAckEvent, call,
3246 RX_CALL_REFCOUNT_DELAY);
3247 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3254 /* Look for the packet in the queue of old received packets */
3255 for (prev = call->rnext - 1, missing =
3256 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3257 /*Check for duplicate packet */
3258 if (seq == tp->header.seq) {
3259 MUTEX_ENTER(&rx_stats_mutex);
3260 rx_stats.dupPacketsRead++;
3261 MUTEX_EXIT(&rx_stats_mutex);
3262 rxevent_Cancel(call->delayedAckEvent, call,
3263 RX_CALL_REFCOUNT_DELAY);
3264 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3270 /* If we find a higher sequence packet, break out and
3271 * insert the new packet here. */
3272 if (seq < tp->header.seq)
3274 /* Check for missing packet */
3275 if (tp->header.seq != prev + 1) {
3279 prev = tp->header.seq;
3282 /* Keep track of whether we have received the last packet. */
3283 if (flags & RX_LAST_PACKET) {
3284 call->flags |= RX_CALL_HAVE_LAST;
3287 /* It's within the window: add it to the the receive queue.
3288 * tp is left by the previous loop either pointing at the
3289 * packet before which to insert the new packet, or at the
3290 * queue head if the queue is empty or the packet should be
3292 queue_InsertBefore(tp, np);
3296 /* Check whether we have all of the packets for this call */
3297 if ((call->flags & RX_CALL_HAVE_LAST)
3298 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3299 afs_uint32 tseq; /* temporary sequence number */
3302 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3303 if (tseq != tp->header.seq)
3305 if (tp->header.flags & RX_LAST_PACKET) {
3306 call->flags |= RX_CALL_RECEIVE_DONE;
3313 /* We need to send an ack of the packet is out of sequence,
3314 * or if an ack was requested by the peer. */
3315 if (seq != prev + 1 || missing || (flags & RX_REQUEST_ACK)) {
3316 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3319 /* Acknowledge the last packet for each call */
3320 if (flags & RX_LAST_PACKET) {
3331 * If the receiver is waiting for an iovec, fill the iovec
3332 * using the data from the receive queue */
3333 if (call->flags & RX_CALL_IOVEC_WAIT) {
3334 didHardAck = rxi_FillReadVec(call, serial);
3335 /* the call may have been aborted */
3344 /* Wakeup the reader if any */
3345 if ((call->flags & RX_CALL_READER_WAIT)
3346 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3347 || (call->iovNext >= call->iovMax)
3348 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3349 call->flags &= ~RX_CALL_READER_WAIT;
3350 #ifdef RX_ENABLE_LOCKS
3351 CV_BROADCAST(&call->cv_rq);
3353 osi_rxWakeup(&call->rq);
3359 * Send an ack when requested by the peer, or once every
3360 * rxi_SoftAckRate packets until the last packet has been
3361 * received. Always send a soft ack for the last packet in
3362 * the server's reply. */
3364 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3365 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3366 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3367 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3368 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3369 } else if (call->nSoftAcks) {
3370 clock_GetTime(&when);
3371 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3372 clock_Add(&when, &rx_lastAckDelay);
3374 clock_Add(&when, &rx_softAckDelay);
3376 if (!call->delayedAckEvent
3377 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3378 rxevent_Cancel(call->delayedAckEvent, call,
3379 RX_CALL_REFCOUNT_DELAY);
3380 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3381 call->delayedAckEvent =
3382 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3384 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3385 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3392 static void rxi_ComputeRate();
3396 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3398 struct rx_peer *peer = conn->peer;
3400 MUTEX_ENTER(&peer->peer_lock);
3401 peer->lastReachTime = clock_Sec();
3402 MUTEX_EXIT(&peer->peer_lock);
3404 MUTEX_ENTER(&conn->conn_data_lock);
3405 if (conn->flags & RX_CONN_ATTACHWAIT) {
3408 conn->flags &= ~RX_CONN_ATTACHWAIT;
3409 MUTEX_EXIT(&conn->conn_data_lock);
3411 for (i = 0; i < RX_MAXCALLS; i++) {
3412 struct rx_call *call = conn->call[i];
3415 MUTEX_ENTER(&call->lock);
3416 /* tnop can be null if newcallp is null */
3417 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3419 MUTEX_EXIT(&call->lock);
3423 MUTEX_EXIT(&conn->conn_data_lock);
3426 /* rxi_ComputePeerNetStats
3428 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3429 * estimates (like RTT and throughput) based on ack packets. Caller
3430 * must ensure that the packet in question is the right one (i.e.
3431 * serial number matches).
3434 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3435 struct rx_ackPacket *ap, struct rx_packet *np)
3437 struct rx_peer *peer = call->conn->peer;
3439 /* Use RTT if not delayed by client. */
3440 if (ap->reason != RX_ACK_DELAY)
3441 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3443 rxi_ComputeRate(peer, call, p, np, ap->reason);
3447 /* The real smarts of the whole thing. */
3449 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3452 struct rx_ackPacket *ap;
3454 register struct rx_packet *tp;
3455 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3456 register struct rx_connection *conn = call->conn;
3457 struct rx_peer *peer = conn->peer;
3460 /* because there are CM's that are bogus, sending weird values for this. */
3461 afs_uint32 skew = 0;
3466 int newAckCount = 0;
3467 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3468 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3470 MUTEX_ENTER(&rx_stats_mutex);
3471 rx_stats.ackPacketsRead++;
3472 MUTEX_EXIT(&rx_stats_mutex);
3473 ap = (struct rx_ackPacket *)rx_DataOf(np);
3474 nbytes = rx_Contiguous(np) - ((ap->acks) - (u_char *) ap);
3476 return np; /* truncated ack packet */
3478 /* depends on ack packet struct */
3479 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3480 first = ntohl(ap->firstPacket);
3481 serial = ntohl(ap->serial);
3482 /* temporarily disabled -- needs to degrade over time
3483 * skew = ntohs(ap->maxSkew); */
3485 /* Ignore ack packets received out of order */
3486 if (first < call->tfirst) {
3490 if (np->header.flags & RX_SLOW_START_OK) {
3491 call->flags |= RX_CALL_SLOW_START_OK;
3494 if (ap->reason == RX_ACK_PING_RESPONSE)
3495 rxi_UpdatePeerReach(conn, call);
3500 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3501 ap->reason, ntohl(ap->previousPacket),
3502 (unsigned int)np->header.seq, (unsigned int)serial,
3503 (unsigned int)skew, ntohl(ap->firstPacket));
3506 for (offset = 0; offset < nAcks; offset++)
3507 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3514 /* Update the outgoing packet skew value to the latest value of
3515 * the peer's incoming packet skew value. The ack packet, of
3516 * course, could arrive out of order, but that won't affect things
3518 MUTEX_ENTER(&peer->peer_lock);
3519 peer->outPacketSkew = skew;
3521 /* Check for packets that no longer need to be transmitted, and
3522 * discard them. This only applies to packets positively
3523 * acknowledged as having been sent to the peer's upper level.
3524 * All other packets must be retained. So only packets with
3525 * sequence numbers < ap->firstPacket are candidates. */
3526 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3527 if (tp->header.seq >= first)
3529 call->tfirst = tp->header.seq + 1;
3531 && (tp->header.serial == serial || tp->firstSerial == serial))
3532 rxi_ComputePeerNetStats(call, tp, ap, np);
3533 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3534 /* XXX Hack. Because we have to release the global rx lock when sending
3535 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3536 * in rxi_Start sending packets out because packets may move to the
3537 * freePacketQueue as result of being here! So we drop these packets until
3538 * we're safely out of the traversing. Really ugly!
3539 * To make it even uglier, if we're using fine grain locking, we can
3540 * set the ack bits in the packets and have rxi_Start remove the packets
3541 * when it's done transmitting.
3543 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3546 if (call->flags & RX_CALL_TQ_BUSY) {
3547 #ifdef RX_ENABLE_LOCKS
3548 tp->flags |= RX_PKTFLAG_ACKED;
3549 call->flags |= RX_CALL_TQ_SOME_ACKED;
3550 #else /* RX_ENABLE_LOCKS */
3552 #endif /* RX_ENABLE_LOCKS */
3554 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3557 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3562 /* Give rate detector a chance to respond to ping requests */
3563 if (ap->reason == RX_ACK_PING_RESPONSE) {
3564 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3568 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3570 /* Now go through explicit acks/nacks and record the results in
3571 * the waiting packets. These are packets that can't be released
3572 * yet, even with a positive acknowledge. This positive
3573 * acknowledge only means the packet has been received by the
3574 * peer, not that it will be retained long enough to be sent to
3575 * the peer's upper level. In addition, reset the transmit timers
3576 * of any missing packets (those packets that must be missing
3577 * because this packet was out of sequence) */
3579 call->nSoftAcked = 0;
3580 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3581 /* Update round trip time if the ack was stimulated on receipt
3583 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3584 #ifdef RX_ENABLE_LOCKS
3585 if (tp->header.seq >= first)
3586 #endif /* RX_ENABLE_LOCKS */
3587 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3589 && (tp->header.serial == serial || tp->firstSerial == serial))
3590 rxi_ComputePeerNetStats(call, tp, ap, np);
3592 /* Set the acknowledge flag per packet based on the
3593 * information in the ack packet. An acknowlegded packet can
3594 * be downgraded when the server has discarded a packet it
3595 * soacked previously, or when an ack packet is received
3596 * out of sequence. */
3597 if (tp->header.seq < first) {
3598 /* Implicit ack information */
3599 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3602 tp->flags |= RX_PKTFLAG_ACKED;
3603 } else if (tp->header.seq < first + nAcks) {
3604 /* Explicit ack information: set it in the packet appropriately */
3605 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3606 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3608 tp->flags |= RX_PKTFLAG_ACKED;
3616 tp->flags &= ~RX_PKTFLAG_ACKED;
3620 tp->flags &= ~RX_PKTFLAG_ACKED;
3624 /* If packet isn't yet acked, and it has been transmitted at least
3625 * once, reset retransmit time using latest timeout
3626 * ie, this should readjust the retransmit timer for all outstanding
3627 * packets... So we don't just retransmit when we should know better*/
3629 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3630 tp->retryTime = tp->timeSent;
3631 clock_Add(&tp->retryTime, &peer->timeout);
3632 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3633 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3637 /* If the window has been extended by this acknowledge packet,
3638 * then wakeup a sender waiting in alloc for window space, or try
3639 * sending packets now, if he's been sitting on packets due to
3640 * lack of window space */
3641 if (call->tnext < (call->tfirst + call->twind)) {
3642 #ifdef RX_ENABLE_LOCKS
3643 CV_SIGNAL(&call->cv_twind);
3645 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3646 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3647 osi_rxWakeup(&call->twind);
3650 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3651 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3655 /* if the ack packet has a receivelen field hanging off it,
3656 * update our state */
3657 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3660 /* If the ack packet has a "recommended" size that is less than
3661 * what I am using now, reduce my size to match */
3662 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3663 sizeof(afs_int32), &tSize);
3664 tSize = (afs_uint32) ntohl(tSize);
3665 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3667 /* Get the maximum packet size to send to this peer */
3668 rx_packetread(np, rx_AckDataSize(ap->nAcks), sizeof(afs_int32),
3670 tSize = (afs_uint32) ntohl(tSize);
3671 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3672 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3674 /* sanity check - peer might have restarted with different params.
3675 * If peer says "send less", dammit, send less... Peer should never
3676 * be unable to accept packets of the size that prior AFS versions would
3677 * send without asking. */
3678 if (peer->maxMTU != tSize) {
3679 peer->maxMTU = tSize;
3680 peer->MTU = MIN(tSize, peer->MTU);
3681 call->MTU = MIN(call->MTU, tSize);
3685 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3688 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3689 sizeof(afs_int32), &tSize);
3690 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3691 if (tSize < call->twind) { /* smaller than our send */
3692 call->twind = tSize; /* window, we must send less... */
3693 call->ssthresh = MIN(call->twind, call->ssthresh);
3696 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3697 * network MTU confused with the loopback MTU. Calculate the
3698 * maximum MTU here for use in the slow start code below.
3700 maxMTU = peer->maxMTU;
3701 /* Did peer restart with older RX version? */
3702 if (peer->maxDgramPackets > 1) {
3703 peer->maxDgramPackets = 1;
3705 } else if (np->length >=
3706 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3709 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3710 sizeof(afs_int32), &tSize);
3711 tSize = (afs_uint32) ntohl(tSize);
3713 * As of AFS 3.5 we set the send window to match the receive window.
3715 if (tSize < call->twind) {
3716 call->twind = tSize;
3717 call->ssthresh = MIN(call->twind, call->ssthresh);
3718 } else if (tSize > call->twind) {
3719 call->twind = tSize;
3723 * As of AFS 3.5, a jumbogram is more than one fixed size
3724 * packet transmitted in a single UDP datagram. If the remote
3725 * MTU is smaller than our local MTU then never send a datagram
3726 * larger than the natural MTU.
3729 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3730 sizeof(afs_int32), &tSize);
3731 maxDgramPackets = (afs_uint32) ntohl(tSize);
3732 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3734 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3735 maxDgramPackets = MIN(maxDgramPackets, tSize);
3736 if (maxDgramPackets > 1) {
3737 peer->maxDgramPackets = maxDgramPackets;
3738 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3740 peer->maxDgramPackets = 1;
3741 call->MTU = peer->natMTU;
3743 } else if (peer->maxDgramPackets > 1) {
3744 /* Restarted with lower version of RX */
3745 peer->maxDgramPackets = 1;
3747 } else if (peer->maxDgramPackets > 1
3748 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3749 /* Restarted with lower version of RX */
3750 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3751 peer->natMTU = OLD_MAX_PACKET_SIZE;
3752 peer->MTU = OLD_MAX_PACKET_SIZE;
3753 peer->maxDgramPackets = 1;
3754 peer->nDgramPackets = 1;
3756 call->MTU = OLD_MAX_PACKET_SIZE;
3761 * Calculate how many datagrams were successfully received after
3762 * the first missing packet and adjust the negative ack counter
3767 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3768 if (call->nNacks < nNacked) {
3769 call->nNacks = nNacked;
3778 if (call->flags & RX_CALL_FAST_RECOVER) {
3780 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3782 call->flags &= ~RX_CALL_FAST_RECOVER;
3783 call->cwind = call->nextCwind;
3784 call->nextCwind = 0;
3787 call->nCwindAcks = 0;
3788 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3789 /* Three negative acks in a row trigger congestion recovery */
3790 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3791 MUTEX_EXIT(&peer->peer_lock);
3792 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3793 /* someone else is waiting to start recovery */
3796 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3797 while (call->flags & RX_CALL_TQ_BUSY) {
3798 call->flags |= RX_CALL_TQ_WAIT;
3799 #ifdef RX_ENABLE_LOCKS
3800 CV_WAIT(&call->cv_tq, &call->lock);
3801 #else /* RX_ENABLE_LOCKS */
3802 osi_rxSleep(&call->tq);
3803 #endif /* RX_ENABLE_LOCKS */
3805 MUTEX_ENTER(&peer->peer_lock);
3806 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3807 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3808 call->flags |= RX_CALL_FAST_RECOVER;
3809 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3811 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3812 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3813 call->nextCwind = call->ssthresh;
3816 peer->MTU = call->MTU;
3817 peer->cwind = call->nextCwind;
3818 peer->nDgramPackets = call->nDgramPackets;
3820 call->congestSeq = peer->congestSeq;
3821 /* Reset the resend times on the packets that were nacked
3822 * so we will retransmit as soon as the window permits*/
3823 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3825 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3826 clock_Zero(&tp->retryTime);
3828 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3833 /* If cwind is smaller than ssthresh, then increase
3834 * the window one packet for each ack we receive (exponential
3836 * If cwind is greater than or equal to ssthresh then increase
3837 * the congestion window by one packet for each cwind acks we
3838 * receive (linear growth). */
3839 if (call->cwind < call->ssthresh) {
3841 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3842 call->nCwindAcks = 0;
3844 call->nCwindAcks += newAckCount;
3845 if (call->nCwindAcks >= call->cwind) {
3846 call->nCwindAcks = 0;
3847 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3851 * If we have received several acknowledgements in a row then
3852 * it is time to increase the size of our datagrams
3854 if ((int)call->nAcks > rx_nDgramThreshold) {
3855 if (peer->maxDgramPackets > 1) {
3856 if (call->nDgramPackets < peer->maxDgramPackets) {
3857 call->nDgramPackets++;
3859 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3860 } else if (call->MTU < peer->maxMTU) {
3861 call->MTU += peer->natMTU;
3862 call->MTU = MIN(call->MTU, peer->maxMTU);
3868 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3870 /* Servers need to hold the call until all response packets have
3871 * been acknowledged. Soft acks are good enough since clients
3872 * are not allowed to clear their receive queues. */
3873 if (call->state == RX_STATE_HOLD
3874 && call->tfirst + call->nSoftAcked >= call->tnext) {
3875 call->state = RX_STATE_DALLY;
3876 rxi_ClearTransmitQueue(call, 0);
3877 } else if (!queue_IsEmpty(&call->tq)) {
3878 rxi_Start(0, call, 0, istack);
3883 /* Received a response to a challenge packet */
3885 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3886 register struct rx_packet *np, int istack)
3890 /* Ignore the packet if we're the client */
3891 if (conn->type == RX_CLIENT_CONNECTION)
3894 /* If already authenticated, ignore the packet (it's probably a retry) */
3895 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3898 /* Otherwise, have the security object evaluate the response packet */
3899 error = RXS_CheckResponse(conn->securityObject, conn, np);
3901 /* If the response is invalid, reset the connection, sending
3902 * an abort to the peer */
3906 rxi_ConnectionError(conn, error);
3907 MUTEX_ENTER(&conn->conn_data_lock);
3908 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3909 MUTEX_EXIT(&conn->conn_data_lock);
3912 /* If the response is valid, any calls waiting to attach
3913 * servers can now do so */
3916 for (i = 0; i < RX_MAXCALLS; i++) {
3917 struct rx_call *call = conn->call[i];
3919 MUTEX_ENTER(&call->lock);
3920 if (call->state == RX_STATE_PRECALL)
3921 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
3922 /* tnop can be null if newcallp is null */
3923 MUTEX_EXIT(&call->lock);
3927 /* Update the peer reachability information, just in case
3928 * some calls went into attach-wait while we were waiting
3929 * for authentication..
3931 rxi_UpdatePeerReach(conn, NULL);
3936 /* A client has received an authentication challenge: the security
3937 * object is asked to cough up a respectable response packet to send
3938 * back to the server. The server is responsible for retrying the
3939 * challenge if it fails to get a response. */
3942 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
3943 register struct rx_packet *np, int istack)
3947 /* Ignore the challenge if we're the server */
3948 if (conn->type == RX_SERVER_CONNECTION)
3951 /* Ignore the challenge if the connection is otherwise idle; someone's
3952 * trying to use us as an oracle. */
3953 if (!rxi_HasActiveCalls(conn))
3956 /* Send the security object the challenge packet. It is expected to fill
3957 * in the response. */
3958 error = RXS_GetResponse(conn->securityObject, conn, np);
3960 /* If the security object is unable to return a valid response, reset the
3961 * connection and send an abort to the peer. Otherwise send the response
3962 * packet to the peer connection. */
3964 rxi_ConnectionError(conn, error);
3965 MUTEX_ENTER(&conn->conn_data_lock);
3966 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3967 MUTEX_EXIT(&conn->conn_data_lock);
3969 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
3970 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
3976 /* Find an available server process to service the current request in
3977 * the given call structure. If one isn't available, queue up this
3978 * call so it eventually gets one */
3980 rxi_AttachServerProc(register struct rx_call *call,
3981 register osi_socket socket, register int *tnop,
3982 register struct rx_call **newcallp)
3984 register struct rx_serverQueueEntry *sq;
3985 register struct rx_service *service = call->conn->service;
3986 register int haveQuota = 0;
3988 /* May already be attached */
3989 if (call->state == RX_STATE_ACTIVE)
3992 MUTEX_ENTER(&rx_serverPool_lock);
3994 haveQuota = QuotaOK(service);
3995 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
3996 /* If there are no processes available to service this call,
3997 * put the call on the incoming call queue (unless it's
3998 * already on the queue).
4000 #ifdef RX_ENABLE_LOCKS
4002 ReturnToServerPool(service);
4003 #endif /* RX_ENABLE_LOCKS */
4005 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4006 call->flags |= RX_CALL_WAIT_PROC;
4007 MUTEX_ENTER(&rx_stats_mutex);
4010 MUTEX_EXIT(&rx_stats_mutex);
4011 rxi_calltrace(RX_CALL_ARRIVAL, call);
4012 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4013 queue_Append(&rx_incomingCallQueue, call);
4016 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4018 /* If hot threads are enabled, and both newcallp and sq->socketp
4019 * are non-null, then this thread will process the call, and the
4020 * idle server thread will start listening on this threads socket.
4023 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4026 *sq->socketp = socket;
4027 clock_GetTime(&call->startTime);
4028 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4032 if (call->flags & RX_CALL_WAIT_PROC) {
4033 /* Conservative: I don't think this should happen */
4034 call->flags &= ~RX_CALL_WAIT_PROC;
4035 if (queue_IsOnQueue(call)) {
4037 MUTEX_ENTER(&rx_stats_mutex);
4039 MUTEX_EXIT(&rx_stats_mutex);
4042 call->state = RX_STATE_ACTIVE;
4043 call->mode = RX_MODE_RECEIVING;
4044 #ifdef RX_KERNEL_TRACE
4046 int glockOwner = ISAFS_GLOCK();
4049 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4050 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4056 if (call->flags & RX_CALL_CLEARED) {
4057 /* send an ack now to start the packet flow up again */
4058 call->flags &= ~RX_CALL_CLEARED;
4059 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4061 #ifdef RX_ENABLE_LOCKS
4064 service->nRequestsRunning++;
4065 if (service->nRequestsRunning <= service->minProcs)
4071 MUTEX_EXIT(&rx_serverPool_lock);
4074 /* Delay the sending of an acknowledge event for a short while, while
4075 * a new call is being prepared (in the case of a client) or a reply
4076 * is being prepared (in the case of a server). Rather than sending
4077 * an ack packet, an ACKALL packet is sent. */
4079 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4081 #ifdef RX_ENABLE_LOCKS
4083 MUTEX_ENTER(&call->lock);
4084 call->delayedAckEvent = NULL;
4085 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4087 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4088 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4090 MUTEX_EXIT(&call->lock);
4091 #else /* RX_ENABLE_LOCKS */
4093 call->delayedAckEvent = NULL;
4094 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4095 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4096 #endif /* RX_ENABLE_LOCKS */
4100 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4103 #ifdef RX_ENABLE_LOCKS
4105 MUTEX_ENTER(&call->lock);
4106 if (event == call->delayedAckEvent)
4107 call->delayedAckEvent = NULL;
4108 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4110 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4112 MUTEX_EXIT(&call->lock);
4113 #else /* RX_ENABLE_LOCKS */
4115 call->delayedAckEvent = NULL;
4116 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4117 #endif /* RX_ENABLE_LOCKS */
4121 #ifdef RX_ENABLE_LOCKS
4122 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4123 * clearing them out.
4126 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4128 register struct rx_packet *p, *tp;
4131 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4132 p->flags |= RX_PKTFLAG_ACKED;
4136 call->flags |= RX_CALL_TQ_CLEARME;
4137 call->flags |= RX_CALL_TQ_SOME_ACKED;
4140 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4141 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4142 call->tfirst = call->tnext;
4143 call->nSoftAcked = 0;
4145 if (call->flags & RX_CALL_FAST_RECOVER) {
4146 call->flags &= ~RX_CALL_FAST_RECOVER;
4147 call->cwind = call->nextCwind;
4148 call->nextCwind = 0;
4151 CV_SIGNAL(&call->cv_twind);
4153 #endif /* RX_ENABLE_LOCKS */
4155 /* Clear out the transmit queue for the current call (all packets have
4156 * been received by peer) */
4158 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4160 register struct rx_packet *p, *tp;
4162 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4163 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4165 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4166 p->flags |= RX_PKTFLAG_ACKED;
4170 call->flags |= RX_CALL_TQ_CLEARME;
4171 call->flags |= RX_CALL_TQ_SOME_ACKED;
4174 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4175 rxi_FreePackets(0, &call->tq);
4176 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4177 call->flags &= ~RX_CALL_TQ_CLEARME;
4179 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4181 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4182 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4183 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4184 call->nSoftAcked = 0;
4186 if (call->flags & RX_CALL_FAST_RECOVER) {
4187 call->flags &= ~RX_CALL_FAST_RECOVER;
4188 call->cwind = call->nextCwind;
4190 #ifdef RX_ENABLE_LOCKS
4191 CV_SIGNAL(&call->cv_twind);
4193 osi_rxWakeup(&call->twind);
4198 rxi_ClearReceiveQueue(register struct rx_call *call)
4200 if (queue_IsNotEmpty(&call->rq)) {
4201 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4202 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4204 if (call->state == RX_STATE_PRECALL) {
4205 call->flags |= RX_CALL_CLEARED;
4209 /* Send an abort packet for the specified call */
4211 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4212 int istack, int force)
4220 /* Clients should never delay abort messages */
4221 if (rx_IsClientConn(call->conn))
4224 if (call->abortCode != call->error) {
4225 call->abortCode = call->error;
4226 call->abortCount = 0;
4229 if (force || rxi_callAbortThreshhold == 0
4230 || call->abortCount < rxi_callAbortThreshhold) {
4231 if (call->delayedAbortEvent) {
4232 rxevent_Cancel(call->delayedAbortEvent, call,
4233 RX_CALL_REFCOUNT_ABORT);
4235 error = htonl(call->error);
4238 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4239 (char *)&error, sizeof(error), istack);
4240 } else if (!call->delayedAbortEvent) {
4241 clock_GetTime(&when);
4242 clock_Addmsec(&when, rxi_callAbortDelay);
4243 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4244 call->delayedAbortEvent =
4245 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4250 /* Send an abort packet for the specified connection. Packet is an
4251 * optional pointer to a packet that can be used to send the abort.
4252 * Once the number of abort messages reaches the threshhold, an
4253 * event is scheduled to send the abort. Setting the force flag
4254 * overrides sending delayed abort messages.
4256 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4257 * to send the abort packet.
4260 rxi_SendConnectionAbort(register struct rx_connection *conn,
4261 struct rx_packet *packet, int istack, int force)
4269 /* Clients should never delay abort messages */
4270 if (rx_IsClientConn(conn))
4273 if (force || rxi_connAbortThreshhold == 0
4274 || conn->abortCount < rxi_connAbortThreshhold) {
4275 if (conn->delayedAbortEvent) {
4276 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4278 error = htonl(conn->error);
4280 MUTEX_EXIT(&conn->conn_data_lock);
4282 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4283 RX_PACKET_TYPE_ABORT, (char *)&error,
4284 sizeof(error), istack);
4285 MUTEX_ENTER(&conn->conn_data_lock);
4286 } else if (!conn->delayedAbortEvent) {
4287 clock_GetTime(&when);
4288 clock_Addmsec(&when, rxi_connAbortDelay);
4289 conn->delayedAbortEvent =
4290 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4295 /* Associate an error all of the calls owned by a connection. Called
4296 * with error non-zero. This is only for really fatal things, like
4297 * bad authentication responses. The connection itself is set in
4298 * error at this point, so that future packets received will be
4301 rxi_ConnectionError(register struct rx_connection *conn,
4302 register afs_int32 error)
4306 MUTEX_ENTER(&conn->conn_data_lock);
4307 if (conn->challengeEvent)
4308 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4309 if (conn->checkReachEvent) {
4310 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4311 conn->checkReachEvent = 0;
4312 conn->flags &= ~RX_CONN_ATTACHWAIT;
4315 MUTEX_EXIT(&conn->conn_data_lock);
4316 for (i = 0; i < RX_MAXCALLS; i++) {
4317 struct rx_call *call = conn->call[i];
4319 MUTEX_ENTER(&call->lock);
4320 rxi_CallError(call, error);
4321 MUTEX_EXIT(&call->lock);
4324 conn->error = error;
4325 MUTEX_ENTER(&rx_stats_mutex);
4326 rx_stats.fatalErrors++;
4327 MUTEX_EXIT(&rx_stats_mutex);
4332 rxi_CallError(register struct rx_call *call, afs_int32 error)
4335 error = call->error;
4336 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4337 if (!(call->flags & RX_CALL_TQ_BUSY)) {
4338 rxi_ResetCall(call, 0);
4341 rxi_ResetCall(call, 0);
4343 call->error = error;
4344 call->mode = RX_MODE_ERROR;
4347 /* Reset various fields in a call structure, and wakeup waiting
4348 * processes. Some fields aren't changed: state & mode are not
4349 * touched (these must be set by the caller), and bufptr, nLeft, and
4350 * nFree are not reset, since these fields are manipulated by
4351 * unprotected macros, and may only be reset by non-interrupting code.
4354 /* this code requires that call->conn be set properly as a pre-condition. */
4355 #endif /* ADAPT_WINDOW */
4358 rxi_ResetCall(register struct rx_call *call, register int newcall)
4361 register struct rx_peer *peer;
4362 struct rx_packet *packet;
4364 /* Notify anyone who is waiting for asynchronous packet arrival */
4365 if (call->arrivalProc) {
4366 (*call->arrivalProc) (call, call->arrivalProcHandle,
4367 call->arrivalProcArg);
4368 call->arrivalProc = (void (*)())0;
4371 if (call->delayedAbortEvent) {
4372 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4373 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4375 rxi_SendCallAbort(call, packet, 0, 1);
4376 rxi_FreePacket(packet);
4381 * Update the peer with the congestion information in this call
4382 * so other calls on this connection can pick up where this call
4383 * left off. If the congestion sequence numbers don't match then
4384 * another call experienced a retransmission.
4386 peer = call->conn->peer;
4387 MUTEX_ENTER(&peer->peer_lock);
4389 if (call->congestSeq == peer->congestSeq) {
4390 peer->cwind = MAX(peer->cwind, call->cwind);
4391 peer->MTU = MAX(peer->MTU, call->MTU);
4392 peer->nDgramPackets =
4393 MAX(peer->nDgramPackets, call->nDgramPackets);
4396 call->abortCode = 0;
4397 call->abortCount = 0;
4399 if (peer->maxDgramPackets > 1) {
4400 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4402 call->MTU = peer->MTU;
4404 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4405 call->ssthresh = rx_maxSendWindow;
4406 call->nDgramPackets = peer->nDgramPackets;
4407 call->congestSeq = peer->congestSeq;
4408 MUTEX_EXIT(&peer->peer_lock);
4410 flags = call->flags;
4411 rxi_ClearReceiveQueue(call);
4412 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4413 if (call->flags & RX_CALL_TQ_BUSY) {
4414 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4415 call->flags |= (flags & RX_CALL_TQ_WAIT);
4417 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4419 rxi_ClearTransmitQueue(call, 0);
4420 queue_Init(&call->tq);
4423 queue_Init(&call->rq);
4425 call->rwind = rx_initReceiveWindow;
4426 call->twind = rx_initSendWindow;
4427 call->nSoftAcked = 0;
4428 call->nextCwind = 0;
4431 call->nCwindAcks = 0;
4432 call->nSoftAcks = 0;
4433 call->nHardAcks = 0;
4435 call->tfirst = call->rnext = call->tnext = 1;
4437 call->lastAcked = 0;
4438 call->localStatus = call->remoteStatus = 0;
4440 if (flags & RX_CALL_READER_WAIT) {
4441 #ifdef RX_ENABLE_LOCKS
4442 CV_BROADCAST(&call->cv_rq);
4444 osi_rxWakeup(&call->rq);
4447 if (flags & RX_CALL_WAIT_PACKETS) {
4448 MUTEX_ENTER(&rx_freePktQ_lock);
4449 rxi_PacketsUnWait(); /* XXX */
4450 MUTEX_EXIT(&rx_freePktQ_lock);
4452 #ifdef RX_ENABLE_LOCKS
4453 CV_SIGNAL(&call->cv_twind);
4455 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4456 osi_rxWakeup(&call->twind);
4459 #ifdef RX_ENABLE_LOCKS
4460 /* The following ensures that we don't mess with any queue while some
4461 * other thread might also be doing so. The call_queue_lock field is
4462 * is only modified under the call lock. If the call is in the process
4463 * of being removed from a queue, the call is not locked until the
4464 * the queue lock is dropped and only then is the call_queue_lock field
4465 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4466 * Note that any other routine which removes a call from a queue has to
4467 * obtain the queue lock before examing the queue and removing the call.
4469 if (call->call_queue_lock) {
4470 MUTEX_ENTER(call->call_queue_lock);
4471 if (queue_IsOnQueue(call)) {
4473 if (flags & RX_CALL_WAIT_PROC) {
4474 MUTEX_ENTER(&rx_stats_mutex);
4476 MUTEX_EXIT(&rx_stats_mutex);
4479 MUTEX_EXIT(call->call_queue_lock);
4480 CLEAR_CALL_QUEUE_LOCK(call);
4482 #else /* RX_ENABLE_LOCKS */
4483 if (queue_IsOnQueue(call)) {
4485 if (flags & RX_CALL_WAIT_PROC)
4488 #endif /* RX_ENABLE_LOCKS */
4490 rxi_KeepAliveOff(call);
4491 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4494 /* Send an acknowledge for the indicated packet (seq,serial) of the
4495 * indicated call, for the indicated reason (reason). This
4496 * acknowledge will specifically acknowledge receiving the packet, and
4497 * will also specify which other packets for this call have been
4498 * received. This routine returns the packet that was used to the
4499 * caller. The caller is responsible for freeing it or re-using it.
4500 * This acknowledgement also returns the highest sequence number
4501 * actually read out by the higher level to the sender; the sender
4502 * promises to keep around packets that have not been read by the
4503 * higher level yet (unless, of course, the sender decides to abort
4504 * the call altogether). Any of p, seq, serial, pflags, or reason may
4505 * be set to zero without ill effect. That is, if they are zero, they
4506 * will not convey any information.
4507 * NOW there is a trailer field, after the ack where it will safely be
4508 * ignored by mundanes, which indicates the maximum size packet this
4509 * host can swallow. */
4511 register struct rx_packet *optionalPacket; use to send ack (or null)
4512 int seq; Sequence number of the packet we are acking
4513 int serial; Serial number of the packet
4514 int pflags; Flags field from packet header
4515 int reason; Reason an acknowledge was prompted
4519 rxi_SendAck(register struct rx_call *call,
4520 register struct rx_packet *optionalPacket, int serial, int reason,
4523 struct rx_ackPacket *ap;
4524 register struct rx_packet *rqp;
4525 register struct rx_packet *nxp; /* For queue_Scan */
4526 register struct rx_packet *p;
4529 #ifdef RX_ENABLE_TSFPQ
4530 struct rx_ts_info_t * rx_ts_info;
4534 * Open the receive window once a thread starts reading packets
4536 if (call->rnext > 1) {
4537 call->rwind = rx_maxReceiveWindow;
4540 call->nHardAcks = 0;
4541 call->nSoftAcks = 0;
4542 if (call->rnext > call->lastAcked)
4543 call->lastAcked = call->rnext;
4547 rx_computelen(p, p->length); /* reset length, you never know */
4548 } /* where that's been... */
4549 #ifdef RX_ENABLE_TSFPQ
4551 RX_TS_INFO_GET(rx_ts_info);
4552 if ((p = rx_ts_info->local_special_packet)) {
4553 rx_computelen(p, p->length);
4554 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4555 rx_ts_info->local_special_packet = p;
4556 } else { /* We won't send the ack, but don't panic. */
4557 return optionalPacket;
4561 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4562 /* We won't send the ack, but don't panic. */
4563 return optionalPacket;
4568 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4571 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4572 #ifndef RX_ENABLE_TSFPQ
4573 if (!optionalPacket)
4576 return optionalPacket;
4578 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4579 if (rx_Contiguous(p) < templ) {
4580 #ifndef RX_ENABLE_TSFPQ
4581 if (!optionalPacket)
4584 return optionalPacket;
4589 /* MTUXXX failing to send an ack is very serious. We should */
4590 /* try as hard as possible to send even a partial ack; it's */
4591 /* better than nothing. */
4592 ap = (struct rx_ackPacket *)rx_DataOf(p);
4593 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4594 ap->reason = reason;
4596 /* The skew computation used to be bogus, I think it's better now. */
4597 /* We should start paying attention to skew. XXX */
4598 ap->serial = htonl(serial);
4599 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4601 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4602 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4604 /* No fear of running out of ack packet here because there can only be at most
4605 * one window full of unacknowledged packets. The window size must be constrained
4606 * to be less than the maximum ack size, of course. Also, an ack should always
4607 * fit into a single packet -- it should not ever be fragmented. */
4608 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4609 if (!rqp || !call->rq.next
4610 || (rqp->header.seq > (call->rnext + call->rwind))) {
4611 #ifndef RX_ENABLE_TSFPQ
4612 if (!optionalPacket)
4615 rxi_CallError(call, RX_CALL_DEAD);
4616 return optionalPacket;
4619 while (rqp->header.seq > call->rnext + offset)
4620 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4621 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4623 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4624 #ifndef RX_ENABLE_TSFPQ
4625 if (!optionalPacket)
4628 rxi_CallError(call, RX_CALL_DEAD);
4629 return optionalPacket;
4634 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4636 /* these are new for AFS 3.3 */
4637 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4638 templ = htonl(templ);
4639 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4640 templ = htonl(call->conn->peer->ifMTU);
4641 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4642 sizeof(afs_int32), &templ);
4644 /* new for AFS 3.4 */
4645 templ = htonl(call->rwind);
4646 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4647 sizeof(afs_int32), &templ);
4649 /* new for AFS 3.5 */
4650 templ = htonl(call->conn->peer->ifDgramPackets);
4651 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4652 sizeof(afs_int32), &templ);
4654 p->header.serviceId = call->conn->serviceId;
4655 p->header.cid = (call->conn->cid | call->channel);
4656 p->header.callNumber = *call->callNumber;
4658 p->header.securityIndex = call->conn->securityIndex;
4659 p->header.epoch = call->conn->epoch;
4660 p->header.type = RX_PACKET_TYPE_ACK;
4661 p->header.flags = RX_SLOW_START_OK;
4662 if (reason == RX_ACK_PING) {
4663 p->header.flags |= RX_REQUEST_ACK;
4665 clock_GetTime(&call->pingRequestTime);
4668 if (call->conn->type == RX_CLIENT_CONNECTION)
4669 p->header.flags |= RX_CLIENT_INITIATED;
4673 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u",
4674 ap->reason, ntohl(ap->previousPacket),
4675 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4677 for (offset = 0; offset < ap->nAcks; offset++)
4678 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4686 register int i, nbytes = p->length;
4688 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4689 if (nbytes <= p->wirevec[i].iov_len) {
4690 register int savelen, saven;
4692 savelen = p->wirevec[i].iov_len;
4694 p->wirevec[i].iov_len = nbytes;
4696 rxi_Send(call, p, istack);
4697 p->wirevec[i].iov_len = savelen;
4701 nbytes -= p->wirevec[i].iov_len;
4704 MUTEX_ENTER(&rx_stats_mutex);
4705 rx_stats.ackPacketsSent++;
4706 MUTEX_EXIT(&rx_stats_mutex);
4707 #ifndef RX_ENABLE_TSFPQ
4708 if (!optionalPacket)
4711 return optionalPacket; /* Return packet for re-use by caller */
4714 /* Send all of the packets in the list in single datagram */
4716 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4717 int istack, int moreFlag, struct clock *now,
4718 struct clock *retryTime, int resending)
4723 struct rx_connection *conn = call->conn;
4724 struct rx_peer *peer = conn->peer;
4726 MUTEX_ENTER(&peer->peer_lock);
4729 peer->reSends += len;
4730 MUTEX_ENTER(&rx_stats_mutex);
4731 rx_stats.dataPacketsSent += len;
4732 MUTEX_EXIT(&rx_stats_mutex);
4733 MUTEX_EXIT(&peer->peer_lock);
4735 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4739 /* Set the packet flags and schedule the resend events */
4740 /* Only request an ack for the last packet in the list */
4741 for (i = 0; i < len; i++) {
4742 list[i]->retryTime = *retryTime;
4743 if (list[i]->header.serial) {
4744 /* Exponentially backoff retry times */
4745 if (list[i]->backoff < MAXBACKOFF) {
4746 /* so it can't stay == 0 */
4747 list[i]->backoff = (list[i]->backoff << 1) + 1;
4750 clock_Addmsec(&(list[i]->retryTime),
4751 ((afs_uint32) list[i]->backoff) << 8);
4754 /* Wait a little extra for the ack on the last packet */
4755 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4756 clock_Addmsec(&(list[i]->retryTime), 400);
4759 /* Record the time sent */
4760 list[i]->timeSent = *now;
4762 /* Ask for an ack on retransmitted packets, on every other packet
4763 * if the peer doesn't support slow start. Ask for an ack on every
4764 * packet until the congestion window reaches the ack rate. */
4765 if (list[i]->header.serial) {
4767 MUTEX_ENTER(&rx_stats_mutex);
4768 rx_stats.dataPacketsReSent++;
4769 MUTEX_EXIT(&rx_stats_mutex);
4771 /* improved RTO calculation- not Karn */
4772 list[i]->firstSent = *now;
4773 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4774 || (!(call->flags & RX_CALL_SLOW_START_OK)
4775 && (list[i]->header.seq & 1)))) {
4780 MUTEX_ENTER(&peer->peer_lock);
4784 MUTEX_ENTER(&rx_stats_mutex);
4785 rx_stats.dataPacketsSent++;
4786 MUTEX_EXIT(&rx_stats_mutex);
4787 MUTEX_EXIT(&peer->peer_lock);
4789 /* Tag this packet as not being the last in this group,
4790 * for the receiver's benefit */
4791 if (i < len - 1 || moreFlag) {
4792 list[i]->header.flags |= RX_MORE_PACKETS;
4795 /* Install the new retransmit time for the packet, and
4796 * record the time sent */
4797 list[i]->timeSent = *now;
4801 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4804 /* Since we're about to send a data packet to the peer, it's
4805 * safe to nuke any scheduled end-of-packets ack */
4806 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4808 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4809 MUTEX_EXIT(&call->lock);
4811 rxi_SendPacketList(call, conn, list, len, istack);
4813 rxi_SendPacket(call, conn, list[0], istack);
4815 MUTEX_ENTER(&call->lock);
4816 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4818 /* Update last send time for this call (for keep-alive
4819 * processing), and for the connection (so that we can discover
4820 * idle connections) */
4821 conn->lastSendTime = call->lastSendTime = clock_Sec();
4824 /* When sending packets we need to follow these rules:
4825 * 1. Never send more than maxDgramPackets in a jumbogram.
4826 * 2. Never send a packet with more than two iovecs in a jumbogram.
4827 * 3. Never send a retransmitted packet in a jumbogram.
4828 * 4. Never send more than cwind/4 packets in a jumbogram
4829 * We always keep the last list we should have sent so we
4830 * can set the RX_MORE_PACKETS flags correctly.
4833 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4834 int istack, struct clock *now, struct clock *retryTime,
4837 int i, cnt, lastCnt = 0;
4838 struct rx_packet **listP, **lastP = 0;
4839 struct rx_peer *peer = call->conn->peer;
4840 int morePackets = 0;
4842 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4843 /* Does the current packet force us to flush the current list? */
4845 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4846 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4848 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4850 /* If the call enters an error state stop sending, or if
4851 * we entered congestion recovery mode, stop sending */
4852 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4860 /* Add the current packet to the list if it hasn't been acked.
4861 * Otherwise adjust the list pointer to skip the current packet. */
4862 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
4864 /* Do we need to flush the list? */
4865 if (cnt >= (int)peer->maxDgramPackets
4866 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
4867 || list[i]->header.serial
4868 || list[i]->length != RX_JUMBOBUFFERSIZE) {
4870 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
4871 retryTime, resending);
4872 /* If the call enters an error state stop sending, or if
4873 * we entered congestion recovery mode, stop sending */
4875 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4880 listP = &list[i + 1];
4885 osi_Panic("rxi_SendList error");
4887 listP = &list[i + 1];
4891 /* Send the whole list when the call is in receive mode, when
4892 * the call is in eof mode, when we are in fast recovery mode,
4893 * and when we have the last packet */
4894 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
4895 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
4896 || (call->flags & RX_CALL_FAST_RECOVER)) {
4897 /* Check for the case where the current list contains
4898 * an acked packet. Since we always send retransmissions
4899 * in a separate packet, we only need to check the first
4900 * packet in the list */
4901 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
4905 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
4906 retryTime, resending);
4907 /* If the call enters an error state stop sending, or if
4908 * we entered congestion recovery mode, stop sending */
4909 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4913 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
4916 } else if (lastCnt > 0) {
4917 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
4922 #ifdef RX_ENABLE_LOCKS
4923 /* Call rxi_Start, below, but with the call lock held. */
4925 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
4926 void *arg1, int istack)
4928 MUTEX_ENTER(&call->lock);
4929 rxi_Start(event, call, arg1, istack);
4930 MUTEX_EXIT(&call->lock);
4932 #endif /* RX_ENABLE_LOCKS */
4934 /* This routine is called when new packets are readied for
4935 * transmission and when retransmission may be necessary, or when the
4936 * transmission window or burst count are favourable. This should be
4937 * better optimized for new packets, the usual case, now that we've
4938 * got rid of queues of send packets. XXXXXXXXXXX */
4940 rxi_Start(struct rxevent *event, register struct rx_call *call,
4941 void *arg1, int istack)
4943 struct rx_packet *p;
4944 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
4945 struct rx_peer *peer = call->conn->peer;
4946 struct clock now, retryTime;
4950 struct rx_packet **xmitList;
4953 /* If rxi_Start is being called as a result of a resend event,
4954 * then make sure that the event pointer is removed from the call
4955 * structure, since there is no longer a per-call retransmission
4957 if (event && event == call->resendEvent) {
4958 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
4959 call->resendEvent = NULL;
4961 if (queue_IsEmpty(&call->tq)) {
4965 /* Timeouts trigger congestion recovery */
4966 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4967 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4968 /* someone else is waiting to start recovery */
4971 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4972 while (call->flags & RX_CALL_TQ_BUSY) {
4973 call->flags |= RX_CALL_TQ_WAIT;
4974 #ifdef RX_ENABLE_LOCKS
4975 CV_WAIT(&call->cv_tq, &call->lock);
4976 #else /* RX_ENABLE_LOCKS */
4977 osi_rxSleep(&call->tq);
4978 #endif /* RX_ENABLE_LOCKS */
4980 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4981 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4982 call->flags |= RX_CALL_FAST_RECOVER;
4983 if (peer->maxDgramPackets > 1) {
4984 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4986 call->MTU = MIN(peer->natMTU, peer->maxMTU);
4988 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4989 call->nDgramPackets = 1;
4991 call->nextCwind = 1;
4994 MUTEX_ENTER(&peer->peer_lock);
4995 peer->MTU = call->MTU;
4996 peer->cwind = call->cwind;
4997 peer->nDgramPackets = 1;
4999 call->congestSeq = peer->congestSeq;
5000 MUTEX_EXIT(&peer->peer_lock);
5001 /* Clear retry times on packets. Otherwise, it's possible for
5002 * some packets in the queue to force resends at rates faster
5003 * than recovery rates.
5005 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5006 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5007 clock_Zero(&p->retryTime);
5012 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5013 MUTEX_ENTER(&rx_stats_mutex);
5014 rx_tq_debug.rxi_start_in_error++;
5015 MUTEX_EXIT(&rx_stats_mutex);
5020 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5021 /* Get clock to compute the re-transmit time for any packets
5022 * in this burst. Note, if we back off, it's reasonable to
5023 * back off all of the packets in the same manner, even if
5024 * some of them have been retransmitted more times than more
5025 * recent additions */
5026 clock_GetTime(&now);
5027 retryTime = now; /* initialize before use */
5028 MUTEX_ENTER(&peer->peer_lock);
5029 clock_Add(&retryTime, &peer->timeout);
5030 MUTEX_EXIT(&peer->peer_lock);
5032 /* Send (or resend) any packets that need it, subject to
5033 * window restrictions and congestion burst control
5034 * restrictions. Ask for an ack on the last packet sent in
5035 * this burst. For now, we're relying upon the window being
5036 * considerably bigger than the largest number of packets that
5037 * are typically sent at once by one initial call to
5038 * rxi_Start. This is probably bogus (perhaps we should ask
5039 * for an ack when we're half way through the current
5040 * window?). Also, for non file transfer applications, this
5041 * may end up asking for an ack for every packet. Bogus. XXXX
5044 * But check whether we're here recursively, and let the other guy
5047 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5048 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5049 call->flags |= RX_CALL_TQ_BUSY;
5051 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5053 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5054 call->flags &= ~RX_CALL_NEED_START;
5055 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5057 maxXmitPackets = MIN(call->twind, call->cwind);
5058 xmitList = (struct rx_packet **)
5059 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5060 if (xmitList == NULL)
5061 osi_Panic("rxi_Start, failed to allocate xmit list");
5062 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5063 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5064 /* We shouldn't be sending packets if a thread is waiting
5065 * to initiate congestion recovery */
5069 && (call->flags & RX_CALL_FAST_RECOVER)) {
5070 /* Only send one packet during fast recovery */
5073 if ((p->flags & RX_PKTFLAG_FREE)
5074 || (!queue_IsEnd(&call->tq, nxp)
5075 && (nxp->flags & RX_PKTFLAG_FREE))
5076 || (p == (struct rx_packet *)&rx_freePacketQueue)
5077 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5078 osi_Panic("rxi_Start: xmit queue clobbered");
5080 if (p->flags & RX_PKTFLAG_ACKED) {
5081 MUTEX_ENTER(&rx_stats_mutex);
5082 rx_stats.ignoreAckedPacket++;
5083 MUTEX_EXIT(&rx_stats_mutex);
5084 continue; /* Ignore this packet if it has been acknowledged */
5087 /* Turn off all flags except these ones, which are the same
5088 * on each transmission */
5089 p->header.flags &= RX_PRESET_FLAGS;
5091 if (p->header.seq >=
5092 call->tfirst + MIN((int)call->twind,
5093 (int)(call->nSoftAcked +
5095 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5096 /* Note: if we're waiting for more window space, we can
5097 * still send retransmits; hence we don't return here, but
5098 * break out to schedule a retransmit event */
5099 dpf(("call %d waiting for window",
5100 *(call->callNumber)));
5104 /* Transmit the packet if it needs to be sent. */
5105 if (!clock_Lt(&now, &p->retryTime)) {
5106 if (nXmitPackets == maxXmitPackets) {
5107 rxi_SendXmitList(call, xmitList, nXmitPackets,
5108 istack, &now, &retryTime,
5110 osi_Free(xmitList, maxXmitPackets *
5111 sizeof(struct rx_packet *));
5114 xmitList[nXmitPackets++] = p;
5118 /* xmitList now hold pointers to all of the packets that are
5119 * ready to send. Now we loop to send the packets */
5120 if (nXmitPackets > 0) {
5121 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5122 &now, &retryTime, resending);
5125 maxXmitPackets * sizeof(struct rx_packet *));
5127 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5129 * TQ references no longer protected by this flag; they must remain
5130 * protected by the global lock.
5132 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5133 call->flags &= ~RX_CALL_TQ_BUSY;
5134 if (call->flags & RX_CALL_TQ_WAIT) {
5135 call->flags &= ~RX_CALL_TQ_WAIT;
5136 #ifdef RX_ENABLE_LOCKS
5137 CV_BROADCAST(&call->cv_tq);
5138 #else /* RX_ENABLE_LOCKS */
5139 osi_rxWakeup(&call->tq);
5140 #endif /* RX_ENABLE_LOCKS */
5145 /* We went into the error state while sending packets. Now is
5146 * the time to reset the call. This will also inform the using
5147 * process that the call is in an error state.
5149 MUTEX_ENTER(&rx_stats_mutex);
5150 rx_tq_debug.rxi_start_aborted++;
5151 MUTEX_EXIT(&rx_stats_mutex);
5152 call->flags &= ~RX_CALL_TQ_BUSY;
5153 if (call->flags & RX_CALL_TQ_WAIT) {
5154 call->flags &= ~RX_CALL_TQ_WAIT;
5155 #ifdef RX_ENABLE_LOCKS
5156 CV_BROADCAST(&call->cv_tq);
5157 #else /* RX_ENABLE_LOCKS */
5158 osi_rxWakeup(&call->tq);
5159 #endif /* RX_ENABLE_LOCKS */
5161 rxi_CallError(call, call->error);
5164 #ifdef RX_ENABLE_LOCKS
5165 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5166 register int missing;
5167 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5168 /* Some packets have received acks. If they all have, we can clear
5169 * the transmit queue.
5172 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5173 if (p->header.seq < call->tfirst
5174 && (p->flags & RX_PKTFLAG_ACKED)) {
5181 call->flags |= RX_CALL_TQ_CLEARME;
5183 #endif /* RX_ENABLE_LOCKS */
5184 /* Don't bother doing retransmits if the TQ is cleared. */
5185 if (call->flags & RX_CALL_TQ_CLEARME) {
5186 rxi_ClearTransmitQueue(call, 1);
5188 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5191 /* Always post a resend event, if there is anything in the
5192 * queue, and resend is possible. There should be at least
5193 * one unacknowledged packet in the queue ... otherwise none
5194 * of these packets should be on the queue in the first place.
5196 if (call->resendEvent) {
5197 /* Cancel the existing event and post a new one */
5198 rxevent_Cancel(call->resendEvent, call,
5199 RX_CALL_REFCOUNT_RESEND);
5202 /* The retry time is the retry time on the first unacknowledged
5203 * packet inside the current window */
5205 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5206 /* Don't set timers for packets outside the window */
5207 if (p->header.seq >= call->tfirst + call->twind) {
5211 if (!(p->flags & RX_PKTFLAG_ACKED)
5212 && !clock_IsZero(&p->retryTime)) {
5214 retryTime = p->retryTime;
5219 /* Post a new event to re-run rxi_Start when retries may be needed */
5220 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5221 #ifdef RX_ENABLE_LOCKS
5222 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5224 rxevent_Post2(&retryTime, rxi_StartUnlocked,
5225 (void *)call, 0, istack);
5226 #else /* RX_ENABLE_LOCKS */
5228 rxevent_Post2(&retryTime, rxi_Start, (void *)call,
5230 #endif /* RX_ENABLE_LOCKS */
5233 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5234 } while (call->flags & RX_CALL_NEED_START);
5236 * TQ references no longer protected by this flag; they must remain
5237 * protected by the global lock.
5239 call->flags &= ~RX_CALL_TQ_BUSY;
5240 if (call->flags & RX_CALL_TQ_WAIT) {
5241 call->flags &= ~RX_CALL_TQ_WAIT;
5242 #ifdef RX_ENABLE_LOCKS
5243 CV_BROADCAST(&call->cv_tq);
5244 #else /* RX_ENABLE_LOCKS */
5245 osi_rxWakeup(&call->tq);
5246 #endif /* RX_ENABLE_LOCKS */
5249 call->flags |= RX_CALL_NEED_START;
5251 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5253 if (call->resendEvent) {
5254 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5259 /* Also adjusts the keep alive parameters for the call, to reflect
5260 * that we have just sent a packet (so keep alives aren't sent
5263 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5266 register struct rx_connection *conn = call->conn;
5268 /* Stamp each packet with the user supplied status */
5269 p->header.userStatus = call->localStatus;
5271 /* Allow the security object controlling this call's security to
5272 * make any last-minute changes to the packet */
5273 RXS_SendPacket(conn->securityObject, call, p);
5275 /* Since we're about to send SOME sort of packet to the peer, it's
5276 * safe to nuke any scheduled end-of-packets ack */
5277 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5279 /* Actually send the packet, filling in more connection-specific fields */
5280 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5281 MUTEX_EXIT(&call->lock);
5282 rxi_SendPacket(call, conn, p, istack);
5283 MUTEX_ENTER(&call->lock);
5284 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5286 /* Update last send time for this call (for keep-alive
5287 * processing), and for the connection (so that we can discover
5288 * idle connections) */
5289 conn->lastSendTime = call->lastSendTime = clock_Sec();
5293 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5294 * that things are fine. Also called periodically to guarantee that nothing
5295 * falls through the cracks (e.g. (error + dally) connections have keepalive
5296 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5298 * haveCTLock Set if calling from rxi_ReapConnections
5300 #ifdef RX_ENABLE_LOCKS
5302 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5303 #else /* RX_ENABLE_LOCKS */
5305 rxi_CheckCall(register struct rx_call *call)
5306 #endif /* RX_ENABLE_LOCKS */
5308 register struct rx_connection *conn = call->conn;
5310 afs_uint32 deadTime;
5312 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5313 if (call->flags & RX_CALL_TQ_BUSY) {
5314 /* Call is active and will be reset by rxi_Start if it's
5315 * in an error state.
5320 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5322 (((afs_uint32) conn->secondsUntilDead << 10) +
5323 ((afs_uint32) conn->peer->rtt >> 3) +
5324 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5326 /* These are computed to the second (+- 1 second). But that's
5327 * good enough for these values, which should be a significant
5328 * number of seconds. */
5329 if (now > (call->lastReceiveTime + deadTime)) {
5330 if (call->state == RX_STATE_ACTIVE) {
5331 rxi_CallError(call, RX_CALL_DEAD);
5334 #ifdef RX_ENABLE_LOCKS
5335 /* Cancel pending events */
5336 rxevent_Cancel(call->delayedAckEvent, call,
5337 RX_CALL_REFCOUNT_DELAY);
5338 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5339 rxevent_Cancel(call->keepAliveEvent, call,
5340 RX_CALL_REFCOUNT_ALIVE);
5341 if (call->refCount == 0) {
5342 rxi_FreeCall(call, haveCTLock);
5346 #else /* RX_ENABLE_LOCKS */
5349 #endif /* RX_ENABLE_LOCKS */
5351 /* Non-active calls are destroyed if they are not responding
5352 * to pings; active calls are simply flagged in error, so the
5353 * attached process can die reasonably gracefully. */
5355 /* see if we have a non-activity timeout */
5356 if (call->startWait && conn->idleDeadTime
5357 && ((call->startWait + conn->idleDeadTime) < now)) {
5358 if (call->state == RX_STATE_ACTIVE) {
5359 rxi_CallError(call, RX_CALL_TIMEOUT);
5363 /* see if we have a hard timeout */
5364 if (conn->hardDeadTime
5365 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5366 if (call->state == RX_STATE_ACTIVE)
5367 rxi_CallError(call, RX_CALL_TIMEOUT);
5374 /* When a call is in progress, this routine is called occasionally to
5375 * make sure that some traffic has arrived (or been sent to) the peer.
5376 * If nothing has arrived in a reasonable amount of time, the call is
5377 * declared dead; if nothing has been sent for a while, we send a
5378 * keep-alive packet (if we're actually trying to keep the call alive)
5381 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5384 struct rx_connection *conn;
5387 MUTEX_ENTER(&call->lock);
5388 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5389 if (event == call->keepAliveEvent)
5390 call->keepAliveEvent = NULL;
5393 #ifdef RX_ENABLE_LOCKS
5394 if (rxi_CheckCall(call, 0)) {
5395 MUTEX_EXIT(&call->lock);
5398 #else /* RX_ENABLE_LOCKS */
5399 if (rxi_CheckCall(call))
5401 #endif /* RX_ENABLE_LOCKS */
5403 /* Don't try to keep alive dallying calls */
5404 if (call->state == RX_STATE_DALLY) {
5405 MUTEX_EXIT(&call->lock);
5410 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5411 /* Don't try to send keepalives if there is unacknowledged data */
5412 /* the rexmit code should be good enough, this little hack
5413 * doesn't quite work XXX */
5414 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5416 rxi_ScheduleKeepAliveEvent(call);
5417 MUTEX_EXIT(&call->lock);
5422 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5424 if (!call->keepAliveEvent) {
5426 clock_GetTime(&when);
5427 when.sec += call->conn->secondsUntilPing;
5428 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5429 call->keepAliveEvent =
5430 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5434 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5436 rxi_KeepAliveOn(register struct rx_call *call)
5438 /* Pretend last packet received was received now--i.e. if another
5439 * packet isn't received within the keep alive time, then the call
5440 * will die; Initialize last send time to the current time--even
5441 * if a packet hasn't been sent yet. This will guarantee that a
5442 * keep-alive is sent within the ping time */
5443 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5444 rxi_ScheduleKeepAliveEvent(call);
5447 /* This routine is called to send connection abort messages
5448 * that have been delayed to throttle looping clients. */
5450 rxi_SendDelayedConnAbort(struct rxevent *event,
5451 register struct rx_connection *conn, char *dummy)
5454 struct rx_packet *packet;
5456 MUTEX_ENTER(&conn->conn_data_lock);
5457 conn->delayedAbortEvent = NULL;
5458 error = htonl(conn->error);
5460 MUTEX_EXIT(&conn->conn_data_lock);
5461 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5464 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5465 RX_PACKET_TYPE_ABORT, (char *)&error,
5467 rxi_FreePacket(packet);
5471 /* This routine is called to send call abort messages
5472 * that have been delayed to throttle looping clients. */
5474 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5478 struct rx_packet *packet;
5480 MUTEX_ENTER(&call->lock);
5481 call->delayedAbortEvent = NULL;
5482 error = htonl(call->error);
5484 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5487 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5488 (char *)&error, sizeof(error), 0);
5489 rxi_FreePacket(packet);
5491 MUTEX_EXIT(&call->lock);
5494 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5495 * seconds) to ask the client to authenticate itself. The routine
5496 * issues a challenge to the client, which is obtained from the
5497 * security object associated with the connection */
5499 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5500 void *arg1, int tries)
5502 conn->challengeEvent = NULL;
5503 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5504 register struct rx_packet *packet;
5508 /* We've failed to authenticate for too long.
5509 * Reset any calls waiting for authentication;
5510 * they are all in RX_STATE_PRECALL.
5514 MUTEX_ENTER(&conn->conn_call_lock);
5515 for (i = 0; i < RX_MAXCALLS; i++) {
5516 struct rx_call *call = conn->call[i];
5518 MUTEX_ENTER(&call->lock);
5519 if (call->state == RX_STATE_PRECALL) {
5520 rxi_CallError(call, RX_CALL_DEAD);
5521 rxi_SendCallAbort(call, NULL, 0, 0);
5523 MUTEX_EXIT(&call->lock);
5526 MUTEX_EXIT(&conn->conn_call_lock);
5530 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5532 /* If there's no packet available, do this later. */
5533 RXS_GetChallenge(conn->securityObject, conn, packet);
5534 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5535 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5536 rxi_FreePacket(packet);
5538 clock_GetTime(&when);
5539 when.sec += RX_CHALLENGE_TIMEOUT;
5540 conn->challengeEvent =
5541 rxevent_Post2(&when, rxi_ChallengeEvent, conn, 0,
5546 /* Call this routine to start requesting the client to authenticate
5547 * itself. This will continue until authentication is established,
5548 * the call times out, or an invalid response is returned. The
5549 * security object associated with the connection is asked to create
5550 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5551 * defined earlier. */
5553 rxi_ChallengeOn(register struct rx_connection *conn)
5555 if (!conn->challengeEvent) {
5556 RXS_CreateChallenge(conn->securityObject, conn);
5557 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5562 /* Compute round trip time of the packet provided, in *rttp.
5565 /* rxi_ComputeRoundTripTime is called with peer locked. */
5566 /* sentp and/or peer may be null */
5568 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5569 register struct clock *sentp,
5570 register struct rx_peer *peer)
5572 struct clock thisRtt, *rttp = &thisRtt;
5574 register int rtt_timeout;
5576 clock_GetTime(rttp);
5578 if (clock_Lt(rttp, sentp)) {
5580 return; /* somebody set the clock back, don't count this time. */
5582 clock_Sub(rttp, sentp);
5583 MUTEX_ENTER(&rx_stats_mutex);
5584 if (clock_Lt(rttp, &rx_stats.minRtt))
5585 rx_stats.minRtt = *rttp;
5586 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5587 if (rttp->sec > 60) {
5588 MUTEX_EXIT(&rx_stats_mutex);
5589 return; /* somebody set the clock ahead */
5591 rx_stats.maxRtt = *rttp;
5593 clock_Add(&rx_stats.totalRtt, rttp);
5594 rx_stats.nRttSamples++;
5595 MUTEX_EXIT(&rx_stats_mutex);
5597 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5599 /* Apply VanJacobson round-trip estimations */
5604 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5605 * srtt is stored as fixed point with 3 bits after the binary
5606 * point (i.e., scaled by 8). The following magic is
5607 * equivalent to the smoothing algorithm in rfc793 with an
5608 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5609 * srtt*8 = srtt*8 + rtt - srtt
5610 * srtt = srtt + rtt/8 - srtt/8
5613 delta = MSEC(rttp) - (peer->rtt >> 3);
5617 * We accumulate a smoothed rtt variance (actually, a smoothed
5618 * mean difference), then set the retransmit timer to smoothed
5619 * rtt + 4 times the smoothed variance (was 2x in van's original
5620 * paper, but 4x works better for me, and apparently for him as
5622 * rttvar is stored as
5623 * fixed point with 2 bits after the binary point (scaled by
5624 * 4). The following is equivalent to rfc793 smoothing with
5625 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5626 * replaces rfc793's wired-in beta.
5627 * dev*4 = dev*4 + (|actual - expected| - dev)
5633 delta -= (peer->rtt_dev >> 2);
5634 peer->rtt_dev += delta;
5636 /* I don't have a stored RTT so I start with this value. Since I'm
5637 * probably just starting a call, and will be pushing more data down
5638 * this, I expect congestion to increase rapidly. So I fudge a
5639 * little, and I set deviance to half the rtt. In practice,
5640 * deviance tends to approach something a little less than
5641 * half the smoothed rtt. */
5642 peer->rtt = (MSEC(rttp) << 3) + 8;
5643 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5645 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5646 * the other of these connections is usually in a user process, and can
5647 * be switched and/or swapped out. So on fast, reliable networks, the
5648 * timeout would otherwise be too short.
5650 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5651 clock_Zero(&(peer->timeout));
5652 clock_Addmsec(&(peer->timeout), rtt_timeout);
5654 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)));
5658 /* Find all server connections that have not been active for a long time, and
5661 rxi_ReapConnections(void)
5664 clock_GetTime(&now);
5666 /* Find server connection structures that haven't been used for
5667 * greater than rx_idleConnectionTime */
5669 struct rx_connection **conn_ptr, **conn_end;
5670 int i, havecalls = 0;
5671 MUTEX_ENTER(&rx_connHashTable_lock);
5672 for (conn_ptr = &rx_connHashTable[0], conn_end =
5673 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5675 struct rx_connection *conn, *next;
5676 struct rx_call *call;
5680 for (conn = *conn_ptr; conn; conn = next) {
5681 /* XXX -- Shouldn't the connection be locked? */
5684 for (i = 0; i < RX_MAXCALLS; i++) {
5685 call = conn->call[i];
5688 MUTEX_ENTER(&call->lock);
5689 #ifdef RX_ENABLE_LOCKS
5690 result = rxi_CheckCall(call, 1);
5691 #else /* RX_ENABLE_LOCKS */
5692 result = rxi_CheckCall(call);
5693 #endif /* RX_ENABLE_LOCKS */
5694 MUTEX_EXIT(&call->lock);
5696 /* If CheckCall freed the call, it might
5697 * have destroyed the connection as well,
5698 * which screws up the linked lists.
5704 if (conn->type == RX_SERVER_CONNECTION) {
5705 /* This only actually destroys the connection if
5706 * there are no outstanding calls */
5707 MUTEX_ENTER(&conn->conn_data_lock);
5708 if (!havecalls && !conn->refCount
5709 && ((conn->lastSendTime + rx_idleConnectionTime) <
5711 conn->refCount++; /* it will be decr in rx_DestroyConn */
5712 MUTEX_EXIT(&conn->conn_data_lock);
5713 #ifdef RX_ENABLE_LOCKS
5714 rxi_DestroyConnectionNoLock(conn);
5715 #else /* RX_ENABLE_LOCKS */
5716 rxi_DestroyConnection(conn);
5717 #endif /* RX_ENABLE_LOCKS */
5719 #ifdef RX_ENABLE_LOCKS
5721 MUTEX_EXIT(&conn->conn_data_lock);
5723 #endif /* RX_ENABLE_LOCKS */
5727 #ifdef RX_ENABLE_LOCKS
5728 while (rx_connCleanup_list) {
5729 struct rx_connection *conn;
5730 conn = rx_connCleanup_list;
5731 rx_connCleanup_list = rx_connCleanup_list->next;
5732 MUTEX_EXIT(&rx_connHashTable_lock);
5733 rxi_CleanupConnection(conn);
5734 MUTEX_ENTER(&rx_connHashTable_lock);
5736 MUTEX_EXIT(&rx_connHashTable_lock);
5737 #endif /* RX_ENABLE_LOCKS */
5740 /* Find any peer structures that haven't been used (haven't had an
5741 * associated connection) for greater than rx_idlePeerTime */
5743 struct rx_peer **peer_ptr, **peer_end;
5745 MUTEX_ENTER(&rx_rpc_stats);
5746 MUTEX_ENTER(&rx_peerHashTable_lock);
5747 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5748 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5750 struct rx_peer *peer, *next, *prev;
5751 for (prev = peer = *peer_ptr; peer; peer = next) {
5753 code = MUTEX_TRYENTER(&peer->peer_lock);
5754 if ((code) && (peer->refCount == 0)
5755 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5756 rx_interface_stat_p rpc_stat, nrpc_stat;
5758 MUTEX_EXIT(&peer->peer_lock);
5759 MUTEX_DESTROY(&peer->peer_lock);
5761 (&peer->rpcStats, rpc_stat, nrpc_stat,
5762 rx_interface_stat)) {
5763 unsigned int num_funcs;
5766 queue_Remove(&rpc_stat->queue_header);
5767 queue_Remove(&rpc_stat->all_peers);
5768 num_funcs = rpc_stat->stats[0].func_total;
5770 sizeof(rx_interface_stat_t) +
5771 rpc_stat->stats[0].func_total *
5772 sizeof(rx_function_entry_v1_t);
5774 rxi_Free(rpc_stat, space);
5775 rxi_rpc_peer_stat_cnt -= num_funcs;
5778 MUTEX_ENTER(&rx_stats_mutex);
5779 rx_stats.nPeerStructs--;
5780 MUTEX_EXIT(&rx_stats_mutex);
5781 if (peer == *peer_ptr) {
5788 MUTEX_EXIT(&peer->peer_lock);
5794 MUTEX_EXIT(&rx_peerHashTable_lock);
5795 MUTEX_EXIT(&rx_rpc_stats);
5798 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5799 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5800 * GC, just below. Really, we shouldn't have to keep moving packets from
5801 * one place to another, but instead ought to always know if we can
5802 * afford to hold onto a packet in its particular use. */
5803 MUTEX_ENTER(&rx_freePktQ_lock);
5804 if (rx_waitingForPackets) {
5805 rx_waitingForPackets = 0;
5806 #ifdef RX_ENABLE_LOCKS
5807 CV_BROADCAST(&rx_waitingForPackets_cv);
5809 osi_rxWakeup(&rx_waitingForPackets);
5812 MUTEX_EXIT(&rx_freePktQ_lock);
5814 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5815 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5819 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5820 * rx.h is sort of strange this is better. This is called with a security
5821 * object before it is discarded. Each connection using a security object has
5822 * its own refcount to the object so it won't actually be freed until the last
5823 * connection is destroyed.
5825 * This is the only rxs module call. A hold could also be written but no one
5829 rxs_Release(struct rx_securityClass *aobj)
5831 return RXS_Close(aobj);
5835 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5836 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5837 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5838 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5840 /* Adjust our estimate of the transmission rate to this peer, given
5841 * that the packet p was just acked. We can adjust peer->timeout and
5842 * call->twind. Pragmatically, this is called
5843 * only with packets of maximal length.
5844 * Called with peer and call locked.
5848 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5849 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5851 afs_int32 xferSize, xferMs;
5852 register afs_int32 minTime;
5855 /* Count down packets */
5856 if (peer->rateFlag > 0)
5858 /* Do nothing until we're enabled */
5859 if (peer->rateFlag != 0)
5864 /* Count only when the ack seems legitimate */
5865 switch (ackReason) {
5866 case RX_ACK_REQUESTED:
5868 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
5872 case RX_ACK_PING_RESPONSE:
5873 if (p) /* want the response to ping-request, not data send */
5875 clock_GetTime(&newTO);
5876 if (clock_Gt(&newTO, &call->pingRequestTime)) {
5877 clock_Sub(&newTO, &call->pingRequestTime);
5878 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
5882 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
5889 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));
5891 /* Track only packets that are big enough. */
5892 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
5896 /* absorb RTT data (in milliseconds) for these big packets */
5897 if (peer->smRtt == 0) {
5898 peer->smRtt = xferMs;
5900 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
5905 if (peer->countDown) {
5909 peer->countDown = 10; /* recalculate only every so often */
5911 /* In practice, we can measure only the RTT for full packets,
5912 * because of the way Rx acks the data that it receives. (If it's
5913 * smaller than a full packet, it often gets implicitly acked
5914 * either by the call response (from a server) or by the next call
5915 * (from a client), and either case confuses transmission times
5916 * with processing times.) Therefore, replace the above
5917 * more-sophisticated processing with a simpler version, where the
5918 * smoothed RTT is kept for full-size packets, and the time to
5919 * transmit a windowful of full-size packets is simply RTT *
5920 * windowSize. Again, we take two steps:
5921 - ensure the timeout is large enough for a single packet's RTT;
5922 - ensure that the window is small enough to fit in the desired timeout.*/
5924 /* First, the timeout check. */
5925 minTime = peer->smRtt;
5926 /* Get a reasonable estimate for a timeout period */
5928 newTO.sec = minTime / 1000;
5929 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
5931 /* Increase the timeout period so that we can always do at least
5932 * one packet exchange */
5933 if (clock_Gt(&newTO, &peer->timeout)) {
5935 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));
5937 peer->timeout = newTO;
5940 /* Now, get an estimate for the transmit window size. */
5941 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
5942 /* Now, convert to the number of full packets that could fit in a
5943 * reasonable fraction of that interval */
5944 minTime /= (peer->smRtt << 1);
5945 xferSize = minTime; /* (make a copy) */
5947 /* Now clamp the size to reasonable bounds. */
5950 else if (minTime > rx_Window)
5951 minTime = rx_Window;
5952 /* if (minTime != peer->maxWindow) {
5953 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
5954 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
5955 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
5957 peer->maxWindow = minTime;
5958 elide... call->twind = minTime;
5962 /* Cut back on the peer timeout if it had earlier grown unreasonably.
5963 * Discern this by calculating the timeout necessary for rx_Window
5965 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
5966 /* calculate estimate for transmission interval in milliseconds */
5967 minTime = rx_Window * peer->smRtt;
5968 if (minTime < 1000) {
5969 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
5970 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
5971 peer->timeout.usec, peer->smRtt, peer->packetSize));
5973 newTO.sec = 0; /* cut back on timeout by half a second */
5974 newTO.usec = 500000;
5975 clock_Sub(&peer->timeout, &newTO);
5980 } /* end of rxi_ComputeRate */
5981 #endif /* ADAPT_WINDOW */
5989 /* Don't call this debugging routine directly; use dpf */
5991 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
5992 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
5996 clock_GetTime(&now);
5997 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
5998 (unsigned int)now.usec / 1000);
5999 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6007 * This function is used to process the rx_stats structure that is local
6008 * to a process as well as an rx_stats structure received from a remote
6009 * process (via rxdebug). Therefore, it needs to do minimal version
6013 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6014 afs_int32 freePackets, char version)
6018 if (size != sizeof(struct rx_stats)) {
6020 "Unexpected size of stats structure: was %d, expected %d\n",
6021 size, sizeof(struct rx_stats));
6024 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6027 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6028 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6029 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6030 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6031 s->specialPktAllocFailures);
6033 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6034 s->receivePktAllocFailures, s->sendPktAllocFailures,
6035 s->specialPktAllocFailures);
6039 " greedy %d, " "bogusReads %d (last from host %x), "
6040 "noPackets %d, " "noBuffers %d, " "selects %d, "
6041 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6042 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6043 s->selects, s->sendSelects);
6045 fprintf(file, " packets read: ");
6046 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6047 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6049 fprintf(file, "\n");
6052 " other read counters: data %d, " "ack %d, " "dup %d "
6053 "spurious %d " "dally %d\n", s->dataPacketsRead,
6054 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6055 s->ignorePacketDally);
6057 fprintf(file, " packets sent: ");
6058 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6059 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6061 fprintf(file, "\n");
6064 " other send counters: ack %d, " "data %d (not resends), "
6065 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6066 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6067 s->dataPacketsPushed, s->ignoreAckedPacket);
6070 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6071 s->netSendFailures, (int)s->fatalErrors);
6073 if (s->nRttSamples) {
6074 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6075 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6077 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6078 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6082 " %d server connections, " "%d client connections, "
6083 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6084 s->nServerConns, s->nClientConns, s->nPeerStructs,
6085 s->nCallStructs, s->nFreeCallStructs);
6087 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6088 fprintf(file, " %d clock updates\n", clock_nUpdates);
6093 /* for backward compatibility */
6095 rx_PrintStats(FILE * file)
6097 MUTEX_ENTER(&rx_stats_mutex);
6098 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6100 MUTEX_EXIT(&rx_stats_mutex);
6104 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6106 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6107 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6108 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6111 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6112 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6113 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6116 " Packet size %d, " "max in packet skew %d, "
6117 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6118 (int)peer->outPacketSkew);
6121 #ifdef AFS_PTHREAD_ENV
6123 * This mutex protects the following static variables:
6127 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6128 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6130 #define LOCK_RX_DEBUG
6131 #define UNLOCK_RX_DEBUG
6132 #endif /* AFS_PTHREAD_ENV */
6135 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6136 u_char type, void *inputData, size_t inputLength,
6137 void *outputData, size_t outputLength)
6139 static afs_int32 counter = 100;
6141 struct rx_header theader;
6143 register afs_int32 code;
6145 struct sockaddr_in taddr, faddr;
6150 endTime = time(0) + 20; /* try for 20 seconds */
6154 tp = &tbuffer[sizeof(struct rx_header)];
6155 taddr.sin_family = AF_INET;
6156 taddr.sin_port = remotePort;
6157 taddr.sin_addr.s_addr = remoteAddr;
6158 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6159 taddr.sin_len = sizeof(struct sockaddr_in);
6162 memset(&theader, 0, sizeof(theader));
6163 theader.epoch = htonl(999);
6165 theader.callNumber = htonl(counter);
6168 theader.type = type;
6169 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6170 theader.serviceId = 0;
6172 memcpy(tbuffer, &theader, sizeof(theader));
6173 memcpy(tp, inputData, inputLength);
6175 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6176 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6178 /* see if there's a packet available */
6180 FD_SET(socket, &imask);
6183 code = select(socket + 1, &imask, 0, 0, &tv);
6184 if (code == 1 && FD_ISSET(socket, &imask)) {
6185 /* now receive a packet */
6186 faddrLen = sizeof(struct sockaddr_in);
6188 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6189 (struct sockaddr *)&faddr, &faddrLen);
6192 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6193 if (counter == ntohl(theader.callNumber))
6198 /* see if we've timed out */
6199 if (endTime < time(0))
6202 code -= sizeof(struct rx_header);
6203 if (code > outputLength)
6204 code = outputLength;
6205 memcpy(outputData, tp, code);
6210 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6211 afs_uint16 remotePort, struct rx_debugStats * stat,
6212 afs_uint32 * supportedValues)
6214 struct rx_debugIn in;
6217 *supportedValues = 0;
6218 in.type = htonl(RX_DEBUGI_GETSTATS);
6221 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6222 &in, sizeof(in), stat, sizeof(*stat));
6225 * If the call was successful, fixup the version and indicate
6226 * what contents of the stat structure are valid.
6227 * Also do net to host conversion of fields here.
6231 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6232 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6234 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6235 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6237 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6238 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6240 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6241 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6243 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6244 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6246 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6247 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6249 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6250 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6252 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6253 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6256 stat->nFreePackets = ntohl(stat->nFreePackets);
6257 stat->packetReclaims = ntohl(stat->packetReclaims);
6258 stat->callsExecuted = ntohl(stat->callsExecuted);
6259 stat->nWaiting = ntohl(stat->nWaiting);
6260 stat->idleThreads = ntohl(stat->idleThreads);
6267 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6268 afs_uint16 remotePort, struct rx_stats * stat,
6269 afs_uint32 * supportedValues)
6271 struct rx_debugIn in;
6272 afs_int32 *lp = (afs_int32 *) stat;
6277 * supportedValues is currently unused, but added to allow future
6278 * versioning of this function.
6281 *supportedValues = 0;
6282 in.type = htonl(RX_DEBUGI_RXSTATS);
6284 memset(stat, 0, sizeof(*stat));
6286 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6287 &in, sizeof(in), stat, sizeof(*stat));
6292 * Do net to host conversion here
6295 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6304 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6305 afs_uint16 remotePort, size_t version_length,
6309 return MakeDebugCall(socket, remoteAddr, remotePort,
6310 RX_PACKET_TYPE_VERSION, a, 1, version,
6315 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6316 afs_uint16 remotePort, afs_int32 * nextConnection,
6317 int allConnections, afs_uint32 debugSupportedValues,
6318 struct rx_debugConn * conn,
6319 afs_uint32 * supportedValues)
6321 struct rx_debugIn in;
6326 * supportedValues is currently unused, but added to allow future
6327 * versioning of this function.
6330 *supportedValues = 0;
6331 if (allConnections) {
6332 in.type = htonl(RX_DEBUGI_GETALLCONN);
6334 in.type = htonl(RX_DEBUGI_GETCONN);
6336 in.index = htonl(*nextConnection);
6337 memset(conn, 0, sizeof(*conn));
6339 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6340 &in, sizeof(in), conn, sizeof(*conn));
6343 *nextConnection += 1;
6346 * Convert old connection format to new structure.
6349 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6350 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6351 #define MOVEvL(a) (conn->a = vL->a)
6353 /* any old or unrecognized version... */
6354 for (i = 0; i < RX_MAXCALLS; i++) {
6355 MOVEvL(callState[i]);
6356 MOVEvL(callMode[i]);
6357 MOVEvL(callFlags[i]);
6358 MOVEvL(callOther[i]);
6360 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6361 MOVEvL(secStats.type);
6362 MOVEvL(secStats.level);
6363 MOVEvL(secStats.flags);
6364 MOVEvL(secStats.expires);
6365 MOVEvL(secStats.packetsReceived);
6366 MOVEvL(secStats.packetsSent);
6367 MOVEvL(secStats.bytesReceived);
6368 MOVEvL(secStats.bytesSent);
6373 * Do net to host conversion here
6375 * I don't convert host or port since we are most likely
6376 * going to want these in NBO.
6378 conn->cid = ntohl(conn->cid);
6379 conn->serial = ntohl(conn->serial);
6380 for (i = 0; i < RX_MAXCALLS; i++) {
6381 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6383 conn->error = ntohl(conn->error);
6384 conn->secStats.flags = ntohl(conn->secStats.flags);
6385 conn->secStats.expires = ntohl(conn->secStats.expires);
6386 conn->secStats.packetsReceived =
6387 ntohl(conn->secStats.packetsReceived);
6388 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6389 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6390 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6391 conn->epoch = ntohl(conn->epoch);
6392 conn->natMTU = ntohl(conn->natMTU);
6399 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6400 afs_uint16 remotePort, afs_int32 * nextPeer,
6401 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6402 afs_uint32 * supportedValues)
6404 struct rx_debugIn in;
6408 * supportedValues is currently unused, but added to allow future
6409 * versioning of this function.
6412 *supportedValues = 0;
6413 in.type = htonl(RX_DEBUGI_GETPEER);
6414 in.index = htonl(*nextPeer);
6415 memset(peer, 0, sizeof(*peer));
6417 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6418 &in, sizeof(in), peer, sizeof(*peer));
6424 * Do net to host conversion here
6426 * I don't convert host or port since we are most likely
6427 * going to want these in NBO.
6429 peer->ifMTU = ntohs(peer->ifMTU);
6430 peer->idleWhen = ntohl(peer->idleWhen);
6431 peer->refCount = ntohs(peer->refCount);
6432 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6433 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6434 peer->rtt = ntohl(peer->rtt);
6435 peer->rtt_dev = ntohl(peer->rtt_dev);
6436 peer->timeout.sec = ntohl(peer->timeout.sec);
6437 peer->timeout.usec = ntohl(peer->timeout.usec);
6438 peer->nSent = ntohl(peer->nSent);
6439 peer->reSends = ntohl(peer->reSends);
6440 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6441 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6442 peer->rateFlag = ntohl(peer->rateFlag);
6443 peer->natMTU = ntohs(peer->natMTU);
6444 peer->maxMTU = ntohs(peer->maxMTU);
6445 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6446 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6447 peer->MTU = ntohs(peer->MTU);
6448 peer->cwind = ntohs(peer->cwind);
6449 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6450 peer->congestSeq = ntohs(peer->congestSeq);
6451 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6452 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6453 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6454 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6459 #endif /* RXDEBUG */
6464 struct rx_serverQueueEntry *np;
6467 register struct rx_call *call;
6468 register struct rx_serverQueueEntry *sq;
6472 if (rxinit_status == 1) {
6474 return; /* Already shutdown. */
6478 #ifndef AFS_PTHREAD_ENV
6479 FD_ZERO(&rx_selectMask);
6480 #endif /* AFS_PTHREAD_ENV */
6481 rxi_dataQuota = RX_MAX_QUOTA;
6482 #ifndef AFS_PTHREAD_ENV
6484 #endif /* AFS_PTHREAD_ENV */
6487 #ifndef AFS_PTHREAD_ENV
6488 #ifndef AFS_USE_GETTIMEOFDAY
6490 #endif /* AFS_USE_GETTIMEOFDAY */
6491 #endif /* AFS_PTHREAD_ENV */
6493 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6494 call = queue_First(&rx_freeCallQueue, rx_call);
6496 rxi_Free(call, sizeof(struct rx_call));
6499 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6500 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6506 struct rx_peer **peer_ptr, **peer_end;
6507 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6508 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6510 struct rx_peer *peer, *next;
6511 for (peer = *peer_ptr; peer; peer = next) {
6512 rx_interface_stat_p rpc_stat, nrpc_stat;
6515 (&peer->rpcStats, rpc_stat, nrpc_stat,
6516 rx_interface_stat)) {
6517 unsigned int num_funcs;
6520 queue_Remove(&rpc_stat->queue_header);
6521 queue_Remove(&rpc_stat->all_peers);
6522 num_funcs = rpc_stat->stats[0].func_total;
6524 sizeof(rx_interface_stat_t) +
6525 rpc_stat->stats[0].func_total *
6526 sizeof(rx_function_entry_v1_t);
6528 rxi_Free(rpc_stat, space);
6529 MUTEX_ENTER(&rx_rpc_stats);
6530 rxi_rpc_peer_stat_cnt -= num_funcs;
6531 MUTEX_EXIT(&rx_rpc_stats);
6535 MUTEX_ENTER(&rx_stats_mutex);
6536 rx_stats.nPeerStructs--;
6537 MUTEX_EXIT(&rx_stats_mutex);
6541 for (i = 0; i < RX_MAX_SERVICES; i++) {
6543 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6545 for (i = 0; i < rx_hashTableSize; i++) {
6546 register struct rx_connection *tc, *ntc;
6547 MUTEX_ENTER(&rx_connHashTable_lock);
6548 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6550 for (j = 0; j < RX_MAXCALLS; j++) {
6552 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6555 rxi_Free(tc, sizeof(*tc));
6557 MUTEX_EXIT(&rx_connHashTable_lock);
6560 MUTEX_ENTER(&freeSQEList_lock);
6562 while ((np = rx_FreeSQEList)) {
6563 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6564 MUTEX_DESTROY(&np->lock);
6565 rxi_Free(np, sizeof(*np));
6568 MUTEX_EXIT(&freeSQEList_lock);
6569 MUTEX_DESTROY(&freeSQEList_lock);
6570 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6571 MUTEX_DESTROY(&rx_connHashTable_lock);
6572 MUTEX_DESTROY(&rx_peerHashTable_lock);
6573 MUTEX_DESTROY(&rx_serverPool_lock);
6575 osi_Free(rx_connHashTable,
6576 rx_hashTableSize * sizeof(struct rx_connection *));
6577 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6579 UNPIN(rx_connHashTable,
6580 rx_hashTableSize * sizeof(struct rx_connection *));
6581 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6583 rxi_FreeAllPackets();
6585 MUTEX_ENTER(&rx_stats_mutex);
6586 rxi_dataQuota = RX_MAX_QUOTA;
6587 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6588 MUTEX_EXIT(&rx_stats_mutex);
6594 #ifdef RX_ENABLE_LOCKS
6596 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6598 if (!MUTEX_ISMINE(lockaddr))
6599 osi_Panic("Lock not held: %s", msg);
6601 #endif /* RX_ENABLE_LOCKS */
6606 * Routines to implement connection specific data.
6610 rx_KeyCreate(rx_destructor_t rtn)
6613 MUTEX_ENTER(&rxi_keyCreate_lock);
6614 key = rxi_keyCreate_counter++;
6615 rxi_keyCreate_destructor = (rx_destructor_t *)
6616 realloc((void *)rxi_keyCreate_destructor,
6617 (key + 1) * sizeof(rx_destructor_t));
6618 rxi_keyCreate_destructor[key] = rtn;
6619 MUTEX_EXIT(&rxi_keyCreate_lock);
6624 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6627 MUTEX_ENTER(&conn->conn_data_lock);
6628 if (!conn->specific) {
6629 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6630 for (i = 0; i < key; i++)
6631 conn->specific[i] = NULL;
6632 conn->nSpecific = key + 1;
6633 conn->specific[key] = ptr;
6634 } else if (key >= conn->nSpecific) {
6635 conn->specific = (void **)
6636 realloc(conn->specific, (key + 1) * sizeof(void *));
6637 for (i = conn->nSpecific; i < key; i++)
6638 conn->specific[i] = NULL;
6639 conn->nSpecific = key + 1;
6640 conn->specific[key] = ptr;
6642 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6643 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6644 conn->specific[key] = ptr;
6646 MUTEX_EXIT(&conn->conn_data_lock);
6650 rx_GetSpecific(struct rx_connection *conn, int key)
6653 MUTEX_ENTER(&conn->conn_data_lock);
6654 if (key >= conn->nSpecific)
6657 ptr = conn->specific[key];
6658 MUTEX_EXIT(&conn->conn_data_lock);
6662 #endif /* !KERNEL */
6665 * processStats is a queue used to store the statistics for the local
6666 * process. Its contents are similar to the contents of the rpcStats
6667 * queue on a rx_peer structure, but the actual data stored within
6668 * this queue contains totals across the lifetime of the process (assuming
6669 * the stats have not been reset) - unlike the per peer structures
6670 * which can come and go based upon the peer lifetime.
6673 static struct rx_queue processStats = { &processStats, &processStats };
6676 * peerStats is a queue used to store the statistics for all peer structs.
6677 * Its contents are the union of all the peer rpcStats queues.
6680 static struct rx_queue peerStats = { &peerStats, &peerStats };
6683 * rxi_monitor_processStats is used to turn process wide stat collection
6687 static int rxi_monitor_processStats = 0;
6690 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6693 static int rxi_monitor_peerStats = 0;
6696 * rxi_AddRpcStat - given all of the information for a particular rpc
6697 * call, create (if needed) and update the stat totals for the rpc.
6701 * IN stats - the queue of stats that will be updated with the new value
6703 * IN rxInterface - a unique number that identifies the rpc interface
6705 * IN currentFunc - the index of the function being invoked
6707 * IN totalFunc - the total number of functions in this interface
6709 * IN queueTime - the amount of time this function waited for a thread
6711 * IN execTime - the amount of time this function invocation took to execute
6713 * IN bytesSent - the number bytes sent by this invocation
6715 * IN bytesRcvd - the number bytes received by this invocation
6717 * IN isServer - if true, this invocation was made to a server
6719 * IN remoteHost - the ip address of the remote host
6721 * IN remotePort - the port of the remote host
6723 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6725 * INOUT counter - if a new stats structure is allocated, the counter will
6726 * be updated with the new number of allocated stat structures
6734 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6735 afs_uint32 currentFunc, afs_uint32 totalFunc,
6736 struct clock *queueTime, struct clock *execTime,
6737 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6738 afs_uint32 remoteHost, afs_uint32 remotePort,
6739 int addToPeerList, unsigned int *counter)
6742 rx_interface_stat_p rpc_stat, nrpc_stat;
6745 * See if there's already a structure for this interface
6748 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6749 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6750 && (rpc_stat->stats[0].remote_is_server == isServer))
6755 * Didn't find a match so allocate a new structure and add it to the
6759 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6760 || (rpc_stat->stats[0].interfaceId != rxInterface)
6761 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6766 sizeof(rx_interface_stat_t) +
6767 totalFunc * sizeof(rx_function_entry_v1_t);
6769 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6770 if (rpc_stat == NULL) {
6774 *counter += totalFunc;
6775 for (i = 0; i < totalFunc; i++) {
6776 rpc_stat->stats[i].remote_peer = remoteHost;
6777 rpc_stat->stats[i].remote_port = remotePort;
6778 rpc_stat->stats[i].remote_is_server = isServer;
6779 rpc_stat->stats[i].interfaceId = rxInterface;
6780 rpc_stat->stats[i].func_total = totalFunc;
6781 rpc_stat->stats[i].func_index = i;
6782 hzero(rpc_stat->stats[i].invocations);
6783 hzero(rpc_stat->stats[i].bytes_sent);
6784 hzero(rpc_stat->stats[i].bytes_rcvd);
6785 rpc_stat->stats[i].queue_time_sum.sec = 0;
6786 rpc_stat->stats[i].queue_time_sum.usec = 0;
6787 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6788 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6789 rpc_stat->stats[i].queue_time_min.sec = 9999999;
6790 rpc_stat->stats[i].queue_time_min.usec = 9999999;
6791 rpc_stat->stats[i].queue_time_max.sec = 0;
6792 rpc_stat->stats[i].queue_time_max.usec = 0;
6793 rpc_stat->stats[i].execution_time_sum.sec = 0;
6794 rpc_stat->stats[i].execution_time_sum.usec = 0;
6795 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
6796 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
6797 rpc_stat->stats[i].execution_time_min.sec = 9999999;
6798 rpc_stat->stats[i].execution_time_min.usec = 9999999;
6799 rpc_stat->stats[i].execution_time_max.sec = 0;
6800 rpc_stat->stats[i].execution_time_max.usec = 0;
6802 queue_Prepend(stats, rpc_stat);
6803 if (addToPeerList) {
6804 queue_Prepend(&peerStats, &rpc_stat->all_peers);
6809 * Increment the stats for this function
6812 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
6813 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
6814 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
6815 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
6816 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
6817 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
6818 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
6820 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
6821 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
6823 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
6824 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
6826 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
6827 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
6829 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
6830 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
6838 * rx_IncrementTimeAndCount - increment the times and count for a particular
6843 * IN peer - the peer who invoked the rpc
6845 * IN rxInterface - a unique number that identifies the rpc interface
6847 * IN currentFunc - the index of the function being invoked
6849 * IN totalFunc - the total number of functions in this interface
6851 * IN queueTime - the amount of time this function waited for a thread
6853 * IN execTime - the amount of time this function invocation took to execute
6855 * IN bytesSent - the number bytes sent by this invocation
6857 * IN bytesRcvd - the number bytes received by this invocation
6859 * IN isServer - if true, this invocation was made to a server
6867 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
6868 afs_uint32 currentFunc, afs_uint32 totalFunc,
6869 struct clock *queueTime, struct clock *execTime,
6870 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
6874 MUTEX_ENTER(&rx_rpc_stats);
6875 MUTEX_ENTER(&peer->peer_lock);
6877 if (rxi_monitor_peerStats) {
6878 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
6879 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6880 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
6883 if (rxi_monitor_processStats) {
6884 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
6885 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6886 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
6889 MUTEX_EXIT(&peer->peer_lock);
6890 MUTEX_EXIT(&rx_rpc_stats);
6895 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
6899 * IN callerVersion - the rpc stat version of the caller.
6901 * IN count - the number of entries to marshall.
6903 * IN stats - pointer to stats to be marshalled.
6905 * OUT ptr - Where to store the marshalled data.
6912 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
6913 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
6919 * We only support the first version
6921 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
6922 *(ptr++) = stats->remote_peer;
6923 *(ptr++) = stats->remote_port;
6924 *(ptr++) = stats->remote_is_server;
6925 *(ptr++) = stats->interfaceId;
6926 *(ptr++) = stats->func_total;
6927 *(ptr++) = stats->func_index;
6928 *(ptr++) = hgethi(stats->invocations);
6929 *(ptr++) = hgetlo(stats->invocations);
6930 *(ptr++) = hgethi(stats->bytes_sent);
6931 *(ptr++) = hgetlo(stats->bytes_sent);
6932 *(ptr++) = hgethi(stats->bytes_rcvd);
6933 *(ptr++) = hgetlo(stats->bytes_rcvd);
6934 *(ptr++) = stats->queue_time_sum.sec;
6935 *(ptr++) = stats->queue_time_sum.usec;
6936 *(ptr++) = stats->queue_time_sum_sqr.sec;
6937 *(ptr++) = stats->queue_time_sum_sqr.usec;
6938 *(ptr++) = stats->queue_time_min.sec;
6939 *(ptr++) = stats->queue_time_min.usec;
6940 *(ptr++) = stats->queue_time_max.sec;
6941 *(ptr++) = stats->queue_time_max.usec;
6942 *(ptr++) = stats->execution_time_sum.sec;
6943 *(ptr++) = stats->execution_time_sum.usec;
6944 *(ptr++) = stats->execution_time_sum_sqr.sec;
6945 *(ptr++) = stats->execution_time_sum_sqr.usec;
6946 *(ptr++) = stats->execution_time_min.sec;
6947 *(ptr++) = stats->execution_time_min.usec;
6948 *(ptr++) = stats->execution_time_max.sec;
6949 *(ptr++) = stats->execution_time_max.usec;
6955 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
6960 * IN callerVersion - the rpc stat version of the caller
6962 * OUT myVersion - the rpc stat version of this function
6964 * OUT clock_sec - local time seconds
6966 * OUT clock_usec - local time microseconds
6968 * OUT allocSize - the number of bytes allocated to contain stats
6970 * OUT statCount - the number stats retrieved from this process.
6972 * OUT stats - the actual stats retrieved from this process.
6976 * Returns void. If successful, stats will != NULL.
6980 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
6981 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
6982 size_t * allocSize, afs_uint32 * statCount,
6983 afs_uint32 ** stats)
6993 *myVersion = RX_STATS_RETRIEVAL_VERSION;
6996 * Check to see if stats are enabled
6999 MUTEX_ENTER(&rx_rpc_stats);
7000 if (!rxi_monitor_processStats) {
7001 MUTEX_EXIT(&rx_rpc_stats);
7005 clock_GetTime(&now);
7006 *clock_sec = now.sec;
7007 *clock_usec = now.usec;
7010 * Allocate the space based upon the caller version
7012 * If the client is at an older version than we are,
7013 * we return the statistic data in the older data format, but
7014 * we still return our version number so the client knows we
7015 * are maintaining more data than it can retrieve.
7018 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7019 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7020 *statCount = rxi_rpc_process_stat_cnt;
7023 * This can't happen yet, but in the future version changes
7024 * can be handled by adding additional code here
7028 if (space > (size_t) 0) {
7030 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7033 rx_interface_stat_p rpc_stat, nrpc_stat;
7037 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7039 * Copy the data based upon the caller version
7041 rx_MarshallProcessRPCStats(callerVersion,
7042 rpc_stat->stats[0].func_total,
7043 rpc_stat->stats, &ptr);
7049 MUTEX_EXIT(&rx_rpc_stats);
7054 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7058 * IN callerVersion - the rpc stat version of the caller
7060 * OUT myVersion - the rpc stat version of this function
7062 * OUT clock_sec - local time seconds
7064 * OUT clock_usec - local time microseconds
7066 * OUT allocSize - the number of bytes allocated to contain stats
7068 * OUT statCount - the number of stats retrieved from the individual
7071 * OUT stats - the actual stats retrieved from the individual peer structures.
7075 * Returns void. If successful, stats will != NULL.
7079 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7080 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7081 size_t * allocSize, afs_uint32 * statCount,
7082 afs_uint32 ** stats)
7092 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7095 * Check to see if stats are enabled
7098 MUTEX_ENTER(&rx_rpc_stats);
7099 if (!rxi_monitor_peerStats) {
7100 MUTEX_EXIT(&rx_rpc_stats);
7104 clock_GetTime(&now);
7105 *clock_sec = now.sec;
7106 *clock_usec = now.usec;
7109 * Allocate the space based upon the caller version
7111 * If the client is at an older version than we are,
7112 * we return the statistic data in the older data format, but
7113 * we still return our version number so the client knows we
7114 * are maintaining more data than it can retrieve.
7117 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7118 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7119 *statCount = rxi_rpc_peer_stat_cnt;
7122 * This can't happen yet, but in the future version changes
7123 * can be handled by adding additional code here
7127 if (space > (size_t) 0) {
7129 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7132 rx_interface_stat_p rpc_stat, nrpc_stat;
7136 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7138 * We have to fix the offset of rpc_stat since we are
7139 * keeping this structure on two rx_queues. The rx_queue
7140 * package assumes that the rx_queue member is the first
7141 * member of the structure. That is, rx_queue assumes that
7142 * any one item is only on one queue at a time. We are
7143 * breaking that assumption and so we have to do a little
7144 * math to fix our pointers.
7147 fix_offset = (char *)rpc_stat;
7148 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7149 rpc_stat = (rx_interface_stat_p) fix_offset;
7152 * Copy the data based upon the caller version
7154 rx_MarshallProcessRPCStats(callerVersion,
7155 rpc_stat->stats[0].func_total,
7156 rpc_stat->stats, &ptr);
7162 MUTEX_EXIT(&rx_rpc_stats);
7167 * rx_FreeRPCStats - free memory allocated by
7168 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7172 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7173 * rx_RetrievePeerRPCStats
7175 * IN allocSize - the number of bytes in stats.
7183 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7185 rxi_Free(stats, allocSize);
7189 * rx_queryProcessRPCStats - see if process rpc stat collection is
7190 * currently enabled.
7196 * Returns 0 if stats are not enabled != 0 otherwise
7200 rx_queryProcessRPCStats(void)
7203 MUTEX_ENTER(&rx_rpc_stats);
7204 rc = rxi_monitor_processStats;
7205 MUTEX_EXIT(&rx_rpc_stats);
7210 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7216 * Returns 0 if stats are not enabled != 0 otherwise
7220 rx_queryPeerRPCStats(void)
7223 MUTEX_ENTER(&rx_rpc_stats);
7224 rc = rxi_monitor_peerStats;
7225 MUTEX_EXIT(&rx_rpc_stats);
7230 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7240 rx_enableProcessRPCStats(void)
7242 MUTEX_ENTER(&rx_rpc_stats);
7243 rx_enable_stats = 1;
7244 rxi_monitor_processStats = 1;
7245 MUTEX_EXIT(&rx_rpc_stats);
7249 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7259 rx_enablePeerRPCStats(void)
7261 MUTEX_ENTER(&rx_rpc_stats);
7262 rx_enable_stats = 1;
7263 rxi_monitor_peerStats = 1;
7264 MUTEX_EXIT(&rx_rpc_stats);
7268 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7278 rx_disableProcessRPCStats(void)
7280 rx_interface_stat_p rpc_stat, nrpc_stat;
7283 MUTEX_ENTER(&rx_rpc_stats);
7286 * Turn off process statistics and if peer stats is also off, turn
7290 rxi_monitor_processStats = 0;
7291 if (rxi_monitor_peerStats == 0) {
7292 rx_enable_stats = 0;
7295 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7296 unsigned int num_funcs = 0;
7299 queue_Remove(rpc_stat);
7300 num_funcs = rpc_stat->stats[0].func_total;
7302 sizeof(rx_interface_stat_t) +
7303 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7305 rxi_Free(rpc_stat, space);
7306 rxi_rpc_process_stat_cnt -= num_funcs;
7308 MUTEX_EXIT(&rx_rpc_stats);
7312 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7322 rx_disablePeerRPCStats(void)
7324 struct rx_peer **peer_ptr, **peer_end;
7327 MUTEX_ENTER(&rx_rpc_stats);
7330 * Turn off peer statistics and if process stats is also off, turn
7334 rxi_monitor_peerStats = 0;
7335 if (rxi_monitor_processStats == 0) {
7336 rx_enable_stats = 0;
7339 MUTEX_ENTER(&rx_peerHashTable_lock);
7340 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7341 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7343 struct rx_peer *peer, *next, *prev;
7344 for (prev = peer = *peer_ptr; peer; peer = next) {
7346 code = MUTEX_TRYENTER(&peer->peer_lock);
7348 rx_interface_stat_p rpc_stat, nrpc_stat;
7351 (&peer->rpcStats, rpc_stat, nrpc_stat,
7352 rx_interface_stat)) {
7353 unsigned int num_funcs = 0;
7356 queue_Remove(&rpc_stat->queue_header);
7357 queue_Remove(&rpc_stat->all_peers);
7358 num_funcs = rpc_stat->stats[0].func_total;
7360 sizeof(rx_interface_stat_t) +
7361 rpc_stat->stats[0].func_total *
7362 sizeof(rx_function_entry_v1_t);
7364 rxi_Free(rpc_stat, space);
7365 rxi_rpc_peer_stat_cnt -= num_funcs;
7367 MUTEX_EXIT(&peer->peer_lock);
7368 if (prev == *peer_ptr) {
7378 MUTEX_EXIT(&rx_peerHashTable_lock);
7379 MUTEX_EXIT(&rx_rpc_stats);
7383 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7388 * IN clearFlag - flag indicating which stats to clear
7396 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7398 rx_interface_stat_p rpc_stat, nrpc_stat;
7400 MUTEX_ENTER(&rx_rpc_stats);
7402 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7403 unsigned int num_funcs = 0, i;
7404 num_funcs = rpc_stat->stats[0].func_total;
7405 for (i = 0; i < num_funcs; i++) {
7406 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7407 hzero(rpc_stat->stats[i].invocations);
7409 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7410 hzero(rpc_stat->stats[i].bytes_sent);
7412 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7413 hzero(rpc_stat->stats[i].bytes_rcvd);
7415 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7416 rpc_stat->stats[i].queue_time_sum.sec = 0;
7417 rpc_stat->stats[i].queue_time_sum.usec = 0;
7419 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7420 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7421 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7423 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7424 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7425 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7427 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7428 rpc_stat->stats[i].queue_time_max.sec = 0;
7429 rpc_stat->stats[i].queue_time_max.usec = 0;
7431 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7432 rpc_stat->stats[i].execution_time_sum.sec = 0;
7433 rpc_stat->stats[i].execution_time_sum.usec = 0;
7435 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7436 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7437 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7439 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7440 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7441 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7443 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7444 rpc_stat->stats[i].execution_time_max.sec = 0;
7445 rpc_stat->stats[i].execution_time_max.usec = 0;
7450 MUTEX_EXIT(&rx_rpc_stats);
7454 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7459 * IN clearFlag - flag indicating which stats to clear
7467 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7469 rx_interface_stat_p rpc_stat, nrpc_stat;
7471 MUTEX_ENTER(&rx_rpc_stats);
7473 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7474 unsigned int num_funcs = 0, i;
7477 * We have to fix the offset of rpc_stat since we are
7478 * keeping this structure on two rx_queues. The rx_queue
7479 * package assumes that the rx_queue member is the first
7480 * member of the structure. That is, rx_queue assumes that
7481 * any one item is only on one queue at a time. We are
7482 * breaking that assumption and so we have to do a little
7483 * math to fix our pointers.
7486 fix_offset = (char *)rpc_stat;
7487 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7488 rpc_stat = (rx_interface_stat_p) fix_offset;
7490 num_funcs = rpc_stat->stats[0].func_total;
7491 for (i = 0; i < num_funcs; i++) {
7492 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7493 hzero(rpc_stat->stats[i].invocations);
7495 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7496 hzero(rpc_stat->stats[i].bytes_sent);
7498 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7499 hzero(rpc_stat->stats[i].bytes_rcvd);
7501 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7502 rpc_stat->stats[i].queue_time_sum.sec = 0;
7503 rpc_stat->stats[i].queue_time_sum.usec = 0;
7505 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7506 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7507 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7509 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7510 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7511 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7513 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7514 rpc_stat->stats[i].queue_time_max.sec = 0;
7515 rpc_stat->stats[i].queue_time_max.usec = 0;
7517 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7518 rpc_stat->stats[i].execution_time_sum.sec = 0;
7519 rpc_stat->stats[i].execution_time_sum.usec = 0;
7521 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7522 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7523 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7525 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7526 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7527 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7529 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7530 rpc_stat->stats[i].execution_time_max.sec = 0;
7531 rpc_stat->stats[i].execution_time_max.usec = 0;
7536 MUTEX_EXIT(&rx_rpc_stats);
7540 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7541 * is authorized to enable/disable/clear RX statistics.
7543 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7546 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7548 rxi_rxstat_userok = proc;
7552 rx_RxStatUserOk(struct rx_call *call)
7554 if (!rxi_rxstat_userok)
7556 return rxi_rxstat_userok(call);