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>
21 #include "afs/sysincludes.h"
22 #include "afsincludes.h"
28 #include <net/net_globals.h>
29 #endif /* AFS_OSF_ENV */
30 #ifdef AFS_LINUX20_ENV
33 #include "netinet/in.h"
35 #include "inet/common.h"
37 #include "inet/ip_ire.h"
39 #include "afs/afs_args.h"
40 #include "afs/afs_osi.h"
41 #ifdef RX_KERNEL_TRACE
42 #include "rx_kcommon.h"
44 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
48 #undef RXDEBUG /* turn off debugging */
50 #if defined(AFS_SGI_ENV)
51 #include "sys/debug.h"
60 #endif /* AFS_OSF_ENV */
62 #include "afs/sysincludes.h"
63 #include "afsincludes.h"
66 #include "rx_kmutex.h"
67 #include "rx_kernel.h"
71 #include "rx_globals.h"
73 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
74 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
75 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
77 extern afs_int32 afs_termState;
79 #include "sys/lockl.h"
80 #include "sys/lock_def.h"
81 #endif /* AFS_AIX41_ENV */
82 # include "rxgen_consts.h"
84 # include <sys/types.h>
94 # include <afs/afsutil.h>
95 # include <WINNT\afsreg.h>
97 # include <sys/socket.h>
98 # include <sys/file.h>
100 # include <sys/stat.h>
101 # include <netinet/in.h>
102 # include <sys/time.h>
105 # include "rx_user.h"
106 # include "rx_clock.h"
107 # include "rx_queue.h"
108 # include "rx_globals.h"
109 # include "rx_trace.h"
110 # include <afs/rxgen_consts.h>
114 #ifdef AFS_PTHREAD_ENV
116 int (*registerProgram) (pid_t, char *) = 0;
117 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
120 int (*registerProgram) (PROCESS, char *) = 0;
121 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
125 /* Local static routines */
126 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
127 #ifdef RX_ENABLE_LOCKS
128 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
131 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
133 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
134 afs_int32 rxi_start_in_error;
136 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
139 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
140 * currently allocated within rx. This number is used to allocate the
141 * memory required to return the statistics when queried.
142 * Protected by the rx_rpc_stats mutex.
145 static unsigned int rxi_rpc_peer_stat_cnt;
148 * rxi_rpc_process_stat_cnt counts the total number of local process stat
149 * structures currently allocated within rx. The number is used to allocate
150 * the memory required to return the statistics when queried.
151 * Protected by the rx_rpc_stats mutex.
154 static unsigned int rxi_rpc_process_stat_cnt;
156 #if !defined(offsetof)
157 #include <stddef.h> /* for definition of offsetof() */
160 #ifdef AFS_PTHREAD_ENV
164 * Use procedural initialization of mutexes/condition variables
168 extern afs_kmutex_t rx_stats_mutex;
169 extern afs_kmutex_t rx_waiting_mutex;
170 extern afs_kmutex_t rx_quota_mutex;
171 extern afs_kmutex_t rx_pthread_mutex;
172 extern afs_kmutex_t rx_packets_mutex;
173 extern afs_kmutex_t des_init_mutex;
174 extern afs_kmutex_t des_random_mutex;
175 extern afs_kmutex_t rx_clock_mutex;
176 extern afs_kmutex_t rxi_connCacheMutex;
177 extern afs_kmutex_t rx_event_mutex;
178 extern afs_kmutex_t osi_malloc_mutex;
179 extern afs_kmutex_t event_handler_mutex;
180 extern afs_kmutex_t listener_mutex;
181 extern afs_kmutex_t rx_if_init_mutex;
182 extern afs_kmutex_t rx_if_mutex;
183 extern afs_kmutex_t rxkad_client_uid_mutex;
184 extern afs_kmutex_t rxkad_random_mutex;
186 extern afs_kcondvar_t rx_event_handler_cond;
187 extern afs_kcondvar_t rx_listener_cond;
189 static afs_kmutex_t epoch_mutex;
190 static afs_kmutex_t rx_init_mutex;
191 static afs_kmutex_t rx_debug_mutex;
192 static afs_kmutex_t rx_rpc_stats;
195 rxi_InitPthread(void)
197 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
198 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
199 MUTEX_INIT(&rx_waiting_mutex, "waiting", MUTEX_DEFAULT, 0);
200 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
201 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
202 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
203 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
204 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
205 MUTEX_INIT(&rx_event_mutex, "event", MUTEX_DEFAULT, 0);
206 MUTEX_INIT(&des_init_mutex, "des", MUTEX_DEFAULT, 0);
207 MUTEX_INIT(&des_random_mutex, "random", MUTEX_DEFAULT, 0);
208 MUTEX_INIT(&osi_malloc_mutex, "malloc", MUTEX_DEFAULT, 0);
209 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
210 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
211 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
212 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
213 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
214 MUTEX_INIT(&rxkad_client_uid_mutex, "uid", MUTEX_DEFAULT, 0);
215 MUTEX_INIT(&rxkad_random_mutex, "rxkad random", MUTEX_DEFAULT, 0);
216 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
218 assert(pthread_cond_init
219 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
220 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
222 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
223 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
225 rxkad_global_stats_init();
227 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
228 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
229 #ifdef RX_ENABLE_LOCKS
232 #endif /* RX_LOCKS_DB */
233 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
234 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
236 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
238 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
240 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
242 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
243 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
244 #endif /* RX_ENABLE_LOCKS */
247 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
248 #define INIT_PTHREAD_LOCKS \
249 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
251 * The rx_stats_mutex mutex protects the following global variables:
252 * rxi_lowConnRefCount
253 * rxi_lowPeerRefCount
262 * The rx_quota_mutex mutex protects the following global variables:
270 * The rx_freePktQ_lock protects the following global variables:
275 * The rx_packets_mutex mutex protects the following global variables:
283 * The rx_pthread_mutex mutex protects the following global variables:
287 #define INIT_PTHREAD_LOCKS
291 /* Variables for handling the minProcs implementation. availProcs gives the
292 * number of threads available in the pool at this moment (not counting dudes
293 * executing right now). totalMin gives the total number of procs required
294 * for handling all minProcs requests. minDeficit is a dynamic variable
295 * tracking the # of procs required to satisfy all of the remaining minProcs
297 * For fine grain locking to work, the quota check and the reservation of
298 * a server thread has to come while rxi_availProcs and rxi_minDeficit
299 * are locked. To this end, the code has been modified under #ifdef
300 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
301 * same time. A new function, ReturnToServerPool() returns the allocation.
303 * A call can be on several queue's (but only one at a time). When
304 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
305 * that no one else is touching the queue. To this end, we store the address
306 * of the queue lock in the call structure (under the call lock) when we
307 * put the call on a queue, and we clear the call_queue_lock when the
308 * call is removed from a queue (once the call lock has been obtained).
309 * This allows rxi_ResetCall to safely synchronize with others wishing
310 * to manipulate the queue.
313 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
314 static afs_kmutex_t rx_rpc_stats;
315 void rxi_StartUnlocked(struct rxevent *event, void *call,
316 void *arg1, int istack);
319 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
320 ** pretty good that the next packet coming in is from the same connection
321 ** as the last packet, since we're send multiple packets in a transmit window.
323 struct rx_connection *rxLastConn = 0;
325 #ifdef RX_ENABLE_LOCKS
326 /* The locking hierarchy for rx fine grain locking is composed of these
329 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
330 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
331 * call->lock - locks call data fields.
332 * These are independent of each other:
333 * rx_freeCallQueue_lock
338 * serverQueueEntry->lock
339 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
341 * peer->lock - locks peer data fields.
342 * conn_data_lock - that more than one thread is not updating a conn data
343 * field at the same time.
351 * Do we need a lock to protect the peer field in the conn structure?
352 * conn->peer was previously a constant for all intents and so has no
353 * lock protecting this field. The multihomed client delta introduced
354 * a RX code change : change the peer field in the connection structure
355 * to that remote interface from which the last packet for this
356 * connection was sent out. This may become an issue if further changes
359 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
360 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
362 /* rxdb_fileID is used to identify the lock location, along with line#. */
363 static int rxdb_fileID = RXDB_FILE_RX;
364 #endif /* RX_LOCKS_DB */
365 #else /* RX_ENABLE_LOCKS */
366 #define SET_CALL_QUEUE_LOCK(C, L)
367 #define CLEAR_CALL_QUEUE_LOCK(C)
368 #endif /* RX_ENABLE_LOCKS */
369 struct rx_serverQueueEntry *rx_waitForPacket = 0;
370 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
372 /* ------------Exported Interfaces------------- */
374 /* This function allows rxkad to set the epoch to a suitably random number
375 * which rx_NewConnection will use in the future. The principle purpose is to
376 * get rxnull connections to use the same epoch as the rxkad connections do, at
377 * least once the first rxkad connection is established. This is important now
378 * that the host/port addresses aren't used in FindConnection: the uniqueness
379 * of epoch/cid matters and the start time won't do. */
381 #ifdef AFS_PTHREAD_ENV
383 * This mutex protects the following global variables:
387 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
388 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
392 #endif /* AFS_PTHREAD_ENV */
395 rx_SetEpoch(afs_uint32 epoch)
402 /* Initialize rx. A port number may be mentioned, in which case this
403 * becomes the default port number for any service installed later.
404 * If 0 is provided for the port number, a random port will be chosen
405 * by the kernel. Whether this will ever overlap anything in
406 * /etc/services is anybody's guess... Returns 0 on success, -1 on
411 int rxinit_status = 1;
412 #ifdef AFS_PTHREAD_ENV
414 * This mutex protects the following global variables:
418 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
419 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
422 #define UNLOCK_RX_INIT
426 rx_InitHost(u_int host, u_int port)
433 char *htable, *ptable;
440 if (rxinit_status == 0) {
441 tmp_status = rxinit_status;
443 return tmp_status; /* Already started; return previous error code. */
449 if (afs_winsockInit() < 0)
455 * Initialize anything necessary to provide a non-premptive threading
458 rxi_InitializeThreadSupport();
461 /* Allocate and initialize a socket for client and perhaps server
464 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
465 if (rx_socket == OSI_NULLSOCKET) {
469 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
472 #endif /* RX_LOCKS_DB */
473 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
474 MUTEX_INIT(&rx_waiting_mutex, "rx_waiting_mutex", MUTEX_DEFAULT, 0);
475 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
476 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
477 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
478 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
479 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
480 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
481 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
483 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
485 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
487 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
489 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
490 #if defined(AFS_HPUX110_ENV)
492 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
493 #endif /* AFS_HPUX110_ENV */
494 #endif /* RX_ENABLE_LOCKS && KERNEL */
497 rx_connDeadTime = 12;
498 rx_tranquil = 0; /* reset flag */
499 memset(&rx_stats, 0, sizeof(struct rx_statistics));
501 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
502 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
503 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
504 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
505 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
506 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
508 /* Malloc up a bunch of packets & buffers */
510 queue_Init(&rx_freePacketQueue);
511 rxi_NeedMorePackets = FALSE;
512 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
514 /* enforce a minimum number of allocated packets */
515 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
516 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
518 /* allocate the initial free packet pool */
519 #ifdef RX_ENABLE_TSFPQ
520 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
521 #else /* RX_ENABLE_TSFPQ */
522 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
523 #endif /* RX_ENABLE_TSFPQ */
530 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
531 tv.tv_sec = clock_now.sec;
532 tv.tv_usec = clock_now.usec;
533 srand((unsigned int)tv.tv_usec);
540 #if defined(KERNEL) && !defined(UKERNEL)
541 /* Really, this should never happen in a real kernel */
544 struct sockaddr_in addr;
546 int addrlen = sizeof(addr);
548 socklen_t addrlen = sizeof(addr);
550 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
554 rx_port = addr.sin_port;
557 rx_stats.minRtt.sec = 9999999;
559 rx_SetEpoch(tv.tv_sec | 0x80000000);
561 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
562 * will provide a randomer value. */
564 MUTEX_ENTER(&rx_quota_mutex);
565 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
566 MUTEX_EXIT(&rx_quota_mutex);
567 /* *Slightly* random start time for the cid. This is just to help
568 * out with the hashing function at the peer */
569 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
570 rx_connHashTable = (struct rx_connection **)htable;
571 rx_peerHashTable = (struct rx_peer **)ptable;
573 rx_lastAckDelay.sec = 0;
574 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
575 rx_hardAckDelay.sec = 0;
576 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
577 rx_softAckDelay.sec = 0;
578 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
580 rxevent_Init(20, rxi_ReScheduleEvents);
582 /* Initialize various global queues */
583 queue_Init(&rx_idleServerQueue);
584 queue_Init(&rx_incomingCallQueue);
585 queue_Init(&rx_freeCallQueue);
587 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
588 /* Initialize our list of usable IP addresses. */
592 /* Start listener process (exact function is dependent on the
593 * implementation environment--kernel or user space) */
597 tmp_status = rxinit_status = 0;
605 return rx_InitHost(htonl(INADDR_ANY), port);
608 /* called with unincremented nRequestsRunning to see if it is OK to start
609 * a new thread in this service. Could be "no" for two reasons: over the
610 * max quota, or would prevent others from reaching their min quota.
612 #ifdef RX_ENABLE_LOCKS
613 /* This verion of QuotaOK reserves quota if it's ok while the
614 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
617 QuotaOK(struct rx_service *aservice)
619 /* check if over max quota */
620 if (aservice->nRequestsRunning >= aservice->maxProcs) {
624 /* under min quota, we're OK */
625 /* otherwise, can use only if there are enough to allow everyone
626 * to go to their min quota after this guy starts.
629 MUTEX_ENTER(&rx_quota_mutex);
630 if ((aservice->nRequestsRunning < aservice->minProcs)
631 || (rxi_availProcs > rxi_minDeficit)) {
632 aservice->nRequestsRunning++;
633 /* just started call in minProcs pool, need fewer to maintain
635 if (aservice->nRequestsRunning <= aservice->minProcs)
638 MUTEX_EXIT(&rx_quota_mutex);
641 MUTEX_EXIT(&rx_quota_mutex);
647 ReturnToServerPool(struct rx_service *aservice)
649 aservice->nRequestsRunning--;
650 MUTEX_ENTER(&rx_quota_mutex);
651 if (aservice->nRequestsRunning < aservice->minProcs)
654 MUTEX_EXIT(&rx_quota_mutex);
657 #else /* RX_ENABLE_LOCKS */
659 QuotaOK(struct rx_service *aservice)
662 /* under min quota, we're OK */
663 if (aservice->nRequestsRunning < aservice->minProcs)
666 /* check if over max quota */
667 if (aservice->nRequestsRunning >= aservice->maxProcs)
670 /* otherwise, can use only if there are enough to allow everyone
671 * to go to their min quota after this guy starts.
673 MUTEX_ENTER(&rx_quota_mutex);
674 if (rxi_availProcs > rxi_minDeficit)
676 MUTEX_EXIT(&rx_quota_mutex);
679 #endif /* RX_ENABLE_LOCKS */
682 /* Called by rx_StartServer to start up lwp's to service calls.
683 NExistingProcs gives the number of procs already existing, and which
684 therefore needn't be created. */
686 rxi_StartServerProcs(int nExistingProcs)
688 struct rx_service *service;
693 /* For each service, reserve N processes, where N is the "minimum"
694 * number of processes that MUST be able to execute a request in parallel,
695 * at any time, for that process. Also compute the maximum difference
696 * between any service's maximum number of processes that can run
697 * (i.e. the maximum number that ever will be run, and a guarantee
698 * that this number will run if other services aren't running), and its
699 * minimum number. The result is the extra number of processes that
700 * we need in order to provide the latter guarantee */
701 for (i = 0; i < RX_MAX_SERVICES; i++) {
703 service = rx_services[i];
704 if (service == (struct rx_service *)0)
706 nProcs += service->minProcs;
707 diff = service->maxProcs - service->minProcs;
711 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
712 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
713 for (i = 0; i < nProcs; i++) {
714 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
720 /* This routine is only required on Windows */
722 rx_StartClientThread(void)
724 #ifdef AFS_PTHREAD_ENV
726 pid = pthread_self();
727 #endif /* AFS_PTHREAD_ENV */
729 #endif /* AFS_NT40_ENV */
731 /* This routine must be called if any services are exported. If the
732 * donateMe flag is set, the calling process is donated to the server
735 rx_StartServer(int donateMe)
737 struct rx_service *service;
743 /* Start server processes, if necessary (exact function is dependent
744 * on the implementation environment--kernel or user space). DonateMe
745 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
746 * case, one less new proc will be created rx_StartServerProcs.
748 rxi_StartServerProcs(donateMe);
750 /* count up the # of threads in minProcs, and add set the min deficit to
751 * be that value, too.
753 for (i = 0; i < RX_MAX_SERVICES; i++) {
754 service = rx_services[i];
755 if (service == (struct rx_service *)0)
757 MUTEX_ENTER(&rx_quota_mutex);
758 rxi_totalMin += service->minProcs;
759 /* below works even if a thread is running, since minDeficit would
760 * still have been decremented and later re-incremented.
762 rxi_minDeficit += service->minProcs;
763 MUTEX_EXIT(&rx_quota_mutex);
766 /* Turn on reaping of idle server connections */
767 rxi_ReapConnections(NULL, NULL, NULL);
776 #ifdef AFS_PTHREAD_ENV
778 pid = afs_pointer_to_int(pthread_self());
779 #else /* AFS_PTHREAD_ENV */
781 LWP_CurrentProcess(&pid);
782 #endif /* AFS_PTHREAD_ENV */
784 sprintf(name, "srv_%d", ++nProcs);
786 (*registerProgram) (pid, name);
788 #endif /* AFS_NT40_ENV */
789 rx_ServerProc(NULL); /* Never returns */
791 #ifdef RX_ENABLE_TSFPQ
792 /* no use leaving packets around in this thread's local queue if
793 * it isn't getting donated to the server thread pool.
795 rxi_FlushLocalPacketsTSFPQ();
796 #endif /* RX_ENABLE_TSFPQ */
800 /* Create a new client connection to the specified service, using the
801 * specified security object to implement the security model for this
803 struct rx_connection *
804 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
805 struct rx_securityClass *securityObject,
806 int serviceSecurityIndex)
810 struct rx_connection *conn;
815 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
816 "serviceSecurityIndex %d)\n",
817 ntohl(shost), ntohs(sport), sservice, securityObject,
818 serviceSecurityIndex));
820 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
821 * the case of kmem_alloc? */
822 conn = rxi_AllocConnection();
823 #ifdef RX_ENABLE_LOCKS
824 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
825 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
826 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
829 MUTEX_ENTER(&rx_connHashTable_lock);
830 cid = (rx_nextCid += RX_MAXCALLS);
831 conn->type = RX_CLIENT_CONNECTION;
833 conn->epoch = rx_epoch;
834 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
835 conn->serviceId = sservice;
836 conn->securityObject = securityObject;
837 conn->securityData = (void *) 0;
838 conn->securityIndex = serviceSecurityIndex;
839 rx_SetConnDeadTime(conn, rx_connDeadTime);
840 rx_SetConnSecondsUntilNatPing(conn, 0);
841 conn->ackRate = RX_FAST_ACK_RATE;
843 conn->specific = NULL;
844 conn->challengeEvent = NULL;
845 conn->delayedAbortEvent = NULL;
846 conn->abortCount = 0;
848 for (i = 0; i < RX_MAXCALLS; i++) {
849 conn->twind[i] = rx_initSendWindow;
850 conn->rwind[i] = rx_initReceiveWindow;
853 RXS_NewConnection(securityObject, conn);
855 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
857 conn->refCount++; /* no lock required since only this thread knows... */
858 conn->next = rx_connHashTable[hashindex];
859 rx_connHashTable[hashindex] = conn;
861 rx_MutexIncrement(rx_stats.nClientConns, rx_stats_mutex);
862 MUTEX_EXIT(&rx_connHashTable_lock);
868 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
870 /* The idea is to set the dead time to a value that allows several
871 * keepalives to be dropped without timing out the connection. */
872 conn->secondsUntilDead = MAX(seconds, 6);
873 conn->secondsUntilPing = conn->secondsUntilDead / 6;
876 int rxi_lowPeerRefCount = 0;
877 int rxi_lowConnRefCount = 0;
880 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
881 * NOTE: must not be called with rx_connHashTable_lock held.
884 rxi_CleanupConnection(struct rx_connection *conn)
886 /* Notify the service exporter, if requested, that this connection
887 * is being destroyed */
888 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
889 (*conn->service->destroyConnProc) (conn);
891 /* Notify the security module that this connection is being destroyed */
892 RXS_DestroyConnection(conn->securityObject, conn);
894 /* If this is the last connection using the rx_peer struct, set its
895 * idle time to now. rxi_ReapConnections will reap it if it's still
896 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
898 MUTEX_ENTER(&rx_peerHashTable_lock);
899 if (conn->peer->refCount < 2) {
900 conn->peer->idleWhen = clock_Sec();
901 if (conn->peer->refCount < 1) {
902 conn->peer->refCount = 1;
903 if (rx_stats_active) {
904 MUTEX_ENTER(&rx_stats_mutex);
905 rxi_lowPeerRefCount++;
906 MUTEX_EXIT(&rx_stats_mutex);
910 conn->peer->refCount--;
911 MUTEX_EXIT(&rx_peerHashTable_lock);
915 if (conn->type == RX_SERVER_CONNECTION)
916 rx_MutexDecrement(rx_stats.nServerConns, rx_stats_mutex);
918 rx_MutexDecrement(rx_stats.nClientConns, rx_stats_mutex);
921 if (conn->specific) {
923 for (i = 0; i < conn->nSpecific; i++) {
924 if (conn->specific[i] && rxi_keyCreate_destructor[i])
925 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
926 conn->specific[i] = NULL;
928 free(conn->specific);
930 conn->specific = NULL;
934 MUTEX_DESTROY(&conn->conn_call_lock);
935 MUTEX_DESTROY(&conn->conn_data_lock);
936 CV_DESTROY(&conn->conn_call_cv);
938 rxi_FreeConnection(conn);
941 /* Destroy the specified connection */
943 rxi_DestroyConnection(struct rx_connection *conn)
945 MUTEX_ENTER(&rx_connHashTable_lock);
946 rxi_DestroyConnectionNoLock(conn);
947 /* conn should be at the head of the cleanup list */
948 if (conn == rx_connCleanup_list) {
949 rx_connCleanup_list = rx_connCleanup_list->next;
950 MUTEX_EXIT(&rx_connHashTable_lock);
951 rxi_CleanupConnection(conn);
953 #ifdef RX_ENABLE_LOCKS
955 MUTEX_EXIT(&rx_connHashTable_lock);
957 #endif /* RX_ENABLE_LOCKS */
961 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
963 struct rx_connection **conn_ptr;
965 struct rx_packet *packet;
972 MUTEX_ENTER(&conn->conn_data_lock);
973 if (conn->refCount > 0)
976 if (rx_stats_active) {
977 MUTEX_ENTER(&rx_stats_mutex);
978 rxi_lowConnRefCount++;
979 MUTEX_EXIT(&rx_stats_mutex);
983 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
984 /* Busy; wait till the last guy before proceeding */
985 MUTEX_EXIT(&conn->conn_data_lock);
990 /* If the client previously called rx_NewCall, but it is still
991 * waiting, treat this as a running call, and wait to destroy the
992 * connection later when the call completes. */
993 if ((conn->type == RX_CLIENT_CONNECTION)
994 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
995 conn->flags |= RX_CONN_DESTROY_ME;
996 MUTEX_EXIT(&conn->conn_data_lock);
1000 MUTEX_EXIT(&conn->conn_data_lock);
1002 /* Check for extant references to this connection */
1003 for (i = 0; i < RX_MAXCALLS; i++) {
1004 struct rx_call *call = conn->call[i];
1007 if (conn->type == RX_CLIENT_CONNECTION) {
1008 MUTEX_ENTER(&call->lock);
1009 if (call->delayedAckEvent) {
1010 /* Push the final acknowledgment out now--there
1011 * won't be a subsequent call to acknowledge the
1012 * last reply packets */
1013 rxevent_Cancel(call->delayedAckEvent, call,
1014 RX_CALL_REFCOUNT_DELAY);
1015 if (call->state == RX_STATE_PRECALL
1016 || call->state == RX_STATE_ACTIVE) {
1017 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1019 rxi_AckAll(NULL, call, 0);
1022 MUTEX_EXIT(&call->lock);
1026 #ifdef RX_ENABLE_LOCKS
1028 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1029 MUTEX_EXIT(&conn->conn_data_lock);
1031 /* Someone is accessing a packet right now. */
1035 #endif /* RX_ENABLE_LOCKS */
1038 /* Don't destroy the connection if there are any call
1039 * structures still in use */
1040 MUTEX_ENTER(&conn->conn_data_lock);
1041 conn->flags |= RX_CONN_DESTROY_ME;
1042 MUTEX_EXIT(&conn->conn_data_lock);
1047 if (conn->natKeepAliveEvent) {
1048 rxi_NatKeepAliveOff(conn);
1051 if (conn->delayedAbortEvent) {
1052 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1053 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1055 MUTEX_ENTER(&conn->conn_data_lock);
1056 rxi_SendConnectionAbort(conn, packet, 0, 1);
1057 MUTEX_EXIT(&conn->conn_data_lock);
1058 rxi_FreePacket(packet);
1062 /* Remove from connection hash table before proceeding */
1064 &rx_connHashTable[CONN_HASH
1065 (peer->host, peer->port, conn->cid, conn->epoch,
1067 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1068 if (*conn_ptr == conn) {
1069 *conn_ptr = conn->next;
1073 /* if the conn that we are destroying was the last connection, then we
1074 * clear rxLastConn as well */
1075 if (rxLastConn == conn)
1078 /* Make sure the connection is completely reset before deleting it. */
1079 /* get rid of pending events that could zap us later */
1080 if (conn->challengeEvent)
1081 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1082 if (conn->checkReachEvent)
1083 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1084 if (conn->natKeepAliveEvent)
1085 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
1087 /* Add the connection to the list of destroyed connections that
1088 * need to be cleaned up. This is necessary to avoid deadlocks
1089 * in the routines we call to inform others that this connection is
1090 * being destroyed. */
1091 conn->next = rx_connCleanup_list;
1092 rx_connCleanup_list = conn;
1095 /* Externally available version */
1097 rx_DestroyConnection(struct rx_connection *conn)
1102 rxi_DestroyConnection(conn);
1107 rx_GetConnection(struct rx_connection *conn)
1112 MUTEX_ENTER(&conn->conn_data_lock);
1114 MUTEX_EXIT(&conn->conn_data_lock);
1118 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1119 /* Wait for the transmit queue to no longer be busy.
1120 * requires the call->lock to be held */
1121 static void rxi_WaitforTQBusy(struct rx_call *call) {
1122 while (call->flags & RX_CALL_TQ_BUSY) {
1123 call->flags |= RX_CALL_TQ_WAIT;
1125 #ifdef RX_ENABLE_LOCKS
1126 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1127 CV_WAIT(&call->cv_tq, &call->lock);
1128 #else /* RX_ENABLE_LOCKS */
1129 osi_rxSleep(&call->tq);
1130 #endif /* RX_ENABLE_LOCKS */
1132 if (call->tqWaiters == 0) {
1133 call->flags &= ~RX_CALL_TQ_WAIT;
1139 /* Start a new rx remote procedure call, on the specified connection.
1140 * If wait is set to 1, wait for a free call channel; otherwise return
1141 * 0. Maxtime gives the maximum number of seconds this call may take,
1142 * after rx_NewCall returns. After this time interval, a call to any
1143 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1144 * For fine grain locking, we hold the conn_call_lock in order to
1145 * to ensure that we don't get signalle after we found a call in an active
1146 * state and before we go to sleep.
1149 rx_NewCall(struct rx_connection *conn)
1152 struct rx_call *call;
1153 struct clock queueTime;
1157 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1160 clock_GetTime(&queueTime);
1162 * Check if there are others waiting for a new call.
1163 * If so, let them go first to avoid starving them.
1164 * This is a fairly simple scheme, and might not be
1165 * a complete solution for large numbers of waiters.
1167 * makeCallWaiters keeps track of the number of
1168 * threads waiting to make calls and the
1169 * RX_CONN_MAKECALL_WAITING flag bit is used to
1170 * indicate that there are indeed calls waiting.
1171 * The flag is set when the waiter is incremented.
1172 * It is only cleared when makeCallWaiters is 0.
1173 * This prevents us from accidently destroying the
1174 * connection while it is potentially about to be used.
1176 MUTEX_ENTER(&conn->conn_call_lock);
1177 MUTEX_ENTER(&conn->conn_data_lock);
1178 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1179 conn->flags |= RX_CONN_MAKECALL_WAITING;
1180 conn->makeCallWaiters++;
1181 MUTEX_EXIT(&conn->conn_data_lock);
1183 #ifdef RX_ENABLE_LOCKS
1184 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1188 MUTEX_ENTER(&conn->conn_data_lock);
1189 conn->makeCallWaiters--;
1190 if (conn->makeCallWaiters == 0)
1191 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1194 /* We are now the active thread in rx_NewCall */
1195 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1196 MUTEX_EXIT(&conn->conn_data_lock);
1201 for (i = 0; i < RX_MAXCALLS; i++) {
1202 call = conn->call[i];
1204 if (call->state == RX_STATE_DALLY) {
1205 MUTEX_ENTER(&call->lock);
1206 if (call->state == RX_STATE_DALLY) {
1208 * We are setting the state to RX_STATE_RESET to
1209 * ensure that no one else will attempt to use this
1210 * call once we drop the conn->conn_call_lock and
1211 * call->lock. We must drop the conn->conn_call_lock
1212 * before calling rxi_ResetCall because the process
1213 * of clearing the transmit queue can block for an
1214 * extended period of time. If we block while holding
1215 * the conn->conn_call_lock, then all rx_EndCall
1216 * processing will block as well. This has a detrimental
1217 * effect on overall system performance.
1219 call->state = RX_STATE_RESET;
1220 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1221 MUTEX_EXIT(&conn->conn_call_lock);
1222 rxi_ResetCall(call, 0);
1223 (*call->callNumber)++;
1224 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1228 * If we failed to be able to safely obtain the
1229 * conn->conn_call_lock we will have to drop the
1230 * call->lock to avoid a deadlock. When the call->lock
1231 * is released the state of the call can change. If it
1232 * is no longer RX_STATE_RESET then some other thread is
1235 MUTEX_EXIT(&call->lock);
1236 MUTEX_ENTER(&conn->conn_call_lock);
1237 MUTEX_ENTER(&call->lock);
1239 if (call->state == RX_STATE_RESET)
1243 * If we get here it means that after dropping
1244 * the conn->conn_call_lock and call->lock that
1245 * the call is no longer ours. If we can't find
1246 * a free call in the remaining slots we should
1247 * not go immediately to RX_CONN_MAKECALL_WAITING
1248 * because by dropping the conn->conn_call_lock
1249 * we have given up synchronization with rx_EndCall.
1250 * Instead, cycle through one more time to see if
1251 * we can find a call that can call our own.
1253 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1256 MUTEX_EXIT(&call->lock);
1259 /* rxi_NewCall returns with mutex locked */
1260 call = rxi_NewCall(conn, i);
1261 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1265 if (i < RX_MAXCALLS) {
1271 MUTEX_ENTER(&conn->conn_data_lock);
1272 conn->flags |= RX_CONN_MAKECALL_WAITING;
1273 conn->makeCallWaiters++;
1274 MUTEX_EXIT(&conn->conn_data_lock);
1276 #ifdef RX_ENABLE_LOCKS
1277 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1281 MUTEX_ENTER(&conn->conn_data_lock);
1282 conn->makeCallWaiters--;
1283 if (conn->makeCallWaiters == 0)
1284 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1285 MUTEX_EXIT(&conn->conn_data_lock);
1287 /* Client is initially in send mode */
1288 call->state = RX_STATE_ACTIVE;
1289 call->error = conn->error;
1291 call->mode = RX_MODE_ERROR;
1293 call->mode = RX_MODE_SENDING;
1295 /* remember start time for call in case we have hard dead time limit */
1296 call->queueTime = queueTime;
1297 clock_GetTime(&call->startTime);
1298 hzero(call->bytesSent);
1299 hzero(call->bytesRcvd);
1301 /* Turn on busy protocol. */
1302 rxi_KeepAliveOn(call);
1305 * We are no longer the active thread in rx_NewCall
1307 MUTEX_ENTER(&conn->conn_data_lock);
1308 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1309 MUTEX_EXIT(&conn->conn_data_lock);
1312 * Wake up anyone else who might be giving us a chance to
1313 * run (see code above that avoids resource starvation).
1315 #ifdef RX_ENABLE_LOCKS
1316 CV_BROADCAST(&conn->conn_call_cv);
1320 MUTEX_EXIT(&conn->conn_call_lock);
1322 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1323 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1324 osi_Panic("rx_NewCall call about to be used without an empty tq");
1326 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1328 MUTEX_EXIT(&call->lock);
1331 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1336 rxi_HasActiveCalls(struct rx_connection *aconn)
1339 struct rx_call *tcall;
1343 for (i = 0; i < RX_MAXCALLS; i++) {
1344 if ((tcall = aconn->call[i])) {
1345 if ((tcall->state == RX_STATE_ACTIVE)
1346 || (tcall->state == RX_STATE_PRECALL)) {
1357 rxi_GetCallNumberVector(struct rx_connection *aconn,
1358 afs_int32 * aint32s)
1361 struct rx_call *tcall;
1365 for (i = 0; i < RX_MAXCALLS; i++) {
1366 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1367 aint32s[i] = aconn->callNumber[i] + 1;
1369 aint32s[i] = aconn->callNumber[i];
1376 rxi_SetCallNumberVector(struct rx_connection *aconn,
1377 afs_int32 * aint32s)
1380 struct rx_call *tcall;
1384 for (i = 0; i < RX_MAXCALLS; i++) {
1385 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1386 aconn->callNumber[i] = aint32s[i] - 1;
1388 aconn->callNumber[i] = aint32s[i];
1394 /* Advertise a new service. A service is named locally by a UDP port
1395 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1398 char *serviceName; Name for identification purposes (e.g. the
1399 service name might be used for probing for
1402 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1403 char *serviceName, struct rx_securityClass **securityObjects,
1404 int nSecurityObjects,
1405 afs_int32(*serviceProc) (struct rx_call * acall))
1407 osi_socket socket = OSI_NULLSOCKET;
1408 struct rx_service *tservice;
1414 if (serviceId == 0) {
1416 "rx_NewService: service id for service %s is not non-zero.\n",
1423 "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",
1431 tservice = rxi_AllocService();
1433 for (i = 0; i < RX_MAX_SERVICES; i++) {
1434 struct rx_service *service = rx_services[i];
1436 if (port == service->servicePort && host == service->serviceHost) {
1437 if (service->serviceId == serviceId) {
1438 /* The identical service has already been
1439 * installed; if the caller was intending to
1440 * change the security classes used by this
1441 * service, he/she loses. */
1443 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1444 serviceName, serviceId, service->serviceName);
1446 rxi_FreeService(tservice);
1449 /* Different service, same port: re-use the socket
1450 * which is bound to the same port */
1451 socket = service->socket;
1454 if (socket == OSI_NULLSOCKET) {
1455 /* If we don't already have a socket (from another
1456 * service on same port) get a new one */
1457 socket = rxi_GetHostUDPSocket(host, port);
1458 if (socket == OSI_NULLSOCKET) {
1460 rxi_FreeService(tservice);
1465 service->socket = socket;
1466 service->serviceHost = host;
1467 service->servicePort = port;
1468 service->serviceId = serviceId;
1469 service->serviceName = serviceName;
1470 service->nSecurityObjects = nSecurityObjects;
1471 service->securityObjects = securityObjects;
1472 service->minProcs = 0;
1473 service->maxProcs = 1;
1474 service->idleDeadTime = 60;
1475 service->idleDeadErr = 0;
1476 service->connDeadTime = rx_connDeadTime;
1477 service->executeRequestProc = serviceProc;
1478 service->checkReach = 0;
1479 rx_services[i] = service; /* not visible until now */
1485 rxi_FreeService(tservice);
1486 (osi_Msg "rx_NewService: cannot support > %d services\n",
1491 /* Set configuration options for all of a service's security objects */
1494 rx_SetSecurityConfiguration(struct rx_service *service,
1495 rx_securityConfigVariables type,
1499 for (i = 0; i<service->nSecurityObjects; i++) {
1500 if (service->securityObjects[i]) {
1501 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1509 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1510 struct rx_securityClass **securityObjects, int nSecurityObjects,
1511 afs_int32(*serviceProc) (struct rx_call * acall))
1513 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1516 /* Generic request processing loop. This routine should be called
1517 * by the implementation dependent rx_ServerProc. If socketp is
1518 * non-null, it will be set to the file descriptor that this thread
1519 * is now listening on. If socketp is null, this routine will never
1522 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1524 struct rx_call *call;
1526 struct rx_service *tservice = NULL;
1533 call = rx_GetCall(threadID, tservice, socketp);
1534 if (socketp && *socketp != OSI_NULLSOCKET) {
1535 /* We are now a listener thread */
1540 /* if server is restarting( typically smooth shutdown) then do not
1541 * allow any new calls.
1544 if (rx_tranquil && (call != NULL)) {
1548 MUTEX_ENTER(&call->lock);
1550 rxi_CallError(call, RX_RESTARTING);
1551 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1553 MUTEX_EXIT(&call->lock);
1557 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1558 #ifdef RX_ENABLE_LOCKS
1560 #endif /* RX_ENABLE_LOCKS */
1561 afs_termState = AFSOP_STOP_AFS;
1562 afs_osi_Wakeup(&afs_termState);
1563 #ifdef RX_ENABLE_LOCKS
1565 #endif /* RX_ENABLE_LOCKS */
1570 tservice = call->conn->service;
1572 if (tservice->beforeProc)
1573 (*tservice->beforeProc) (call);
1575 code = call->conn->service->executeRequestProc(call);
1577 if (tservice->afterProc)
1578 (*tservice->afterProc) (call, code);
1580 rx_EndCall(call, code);
1581 if (rx_stats_active) {
1582 MUTEX_ENTER(&rx_stats_mutex);
1584 MUTEX_EXIT(&rx_stats_mutex);
1591 rx_WakeupServerProcs(void)
1593 struct rx_serverQueueEntry *np, *tqp;
1597 MUTEX_ENTER(&rx_serverPool_lock);
1599 #ifdef RX_ENABLE_LOCKS
1600 if (rx_waitForPacket)
1601 CV_BROADCAST(&rx_waitForPacket->cv);
1602 #else /* RX_ENABLE_LOCKS */
1603 if (rx_waitForPacket)
1604 osi_rxWakeup(rx_waitForPacket);
1605 #endif /* RX_ENABLE_LOCKS */
1606 MUTEX_ENTER(&freeSQEList_lock);
1607 for (np = rx_FreeSQEList; np; np = tqp) {
1608 tqp = *(struct rx_serverQueueEntry **)np;
1609 #ifdef RX_ENABLE_LOCKS
1610 CV_BROADCAST(&np->cv);
1611 #else /* RX_ENABLE_LOCKS */
1613 #endif /* RX_ENABLE_LOCKS */
1615 MUTEX_EXIT(&freeSQEList_lock);
1616 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1617 #ifdef RX_ENABLE_LOCKS
1618 CV_BROADCAST(&np->cv);
1619 #else /* RX_ENABLE_LOCKS */
1621 #endif /* RX_ENABLE_LOCKS */
1623 MUTEX_EXIT(&rx_serverPool_lock);
1628 * One thing that seems to happen is that all the server threads get
1629 * tied up on some empty or slow call, and then a whole bunch of calls
1630 * arrive at once, using up the packet pool, so now there are more
1631 * empty calls. The most critical resources here are server threads
1632 * and the free packet pool. The "doreclaim" code seems to help in
1633 * general. I think that eventually we arrive in this state: there
1634 * are lots of pending calls which do have all their packets present,
1635 * so they won't be reclaimed, are multi-packet calls, so they won't
1636 * be scheduled until later, and thus are tying up most of the free
1637 * packet pool for a very long time.
1639 * 1. schedule multi-packet calls if all the packets are present.
1640 * Probably CPU-bound operation, useful to return packets to pool.
1641 * Do what if there is a full window, but the last packet isn't here?
1642 * 3. preserve one thread which *only* runs "best" calls, otherwise
1643 * it sleeps and waits for that type of call.
1644 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1645 * the current dataquota business is badly broken. The quota isn't adjusted
1646 * to reflect how many packets are presently queued for a running call.
1647 * So, when we schedule a queued call with a full window of packets queued
1648 * up for it, that *should* free up a window full of packets for other 2d-class
1649 * calls to be able to use from the packet pool. But it doesn't.
1651 * NB. Most of the time, this code doesn't run -- since idle server threads
1652 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1653 * as a new call arrives.
1655 /* Sleep until a call arrives. Returns a pointer to the call, ready
1656 * for an rx_Read. */
1657 #ifdef RX_ENABLE_LOCKS
1659 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1661 struct rx_serverQueueEntry *sq;
1662 struct rx_call *call = (struct rx_call *)0;
1663 struct rx_service *service = NULL;
1666 MUTEX_ENTER(&freeSQEList_lock);
1668 if ((sq = rx_FreeSQEList)) {
1669 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1670 MUTEX_EXIT(&freeSQEList_lock);
1671 } else { /* otherwise allocate a new one and return that */
1672 MUTEX_EXIT(&freeSQEList_lock);
1673 sq = (struct rx_serverQueueEntry *)
1674 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1675 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1676 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1679 MUTEX_ENTER(&rx_serverPool_lock);
1680 if (cur_service != NULL) {
1681 ReturnToServerPool(cur_service);
1684 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1685 struct rx_call *tcall, *ncall, *choice2 = NULL;
1687 /* Scan for eligible incoming calls. A call is not eligible
1688 * if the maximum number of calls for its service type are
1689 * already executing */
1690 /* One thread will process calls FCFS (to prevent starvation),
1691 * while the other threads may run ahead looking for calls which
1692 * have all their input data available immediately. This helps
1693 * keep threads from blocking, waiting for data from the client. */
1694 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1695 service = tcall->conn->service;
1696 if (!QuotaOK(service)) {
1699 MUTEX_ENTER(&rx_pthread_mutex);
1700 if (tno == rxi_fcfs_thread_num
1701 || !tcall->queue_item_header.next) {
1702 MUTEX_EXIT(&rx_pthread_mutex);
1703 /* If we're the fcfs thread , then we'll just use
1704 * this call. If we haven't been able to find an optimal
1705 * choice, and we're at the end of the list, then use a
1706 * 2d choice if one has been identified. Otherwise... */
1707 call = (choice2 ? choice2 : tcall);
1708 service = call->conn->service;
1710 MUTEX_EXIT(&rx_pthread_mutex);
1711 if (!queue_IsEmpty(&tcall->rq)) {
1712 struct rx_packet *rp;
1713 rp = queue_First(&tcall->rq, rx_packet);
1714 if (rp->header.seq == 1) {
1716 || (rp->header.flags & RX_LAST_PACKET)) {
1718 } else if (rxi_2dchoice && !choice2
1719 && !(tcall->flags & RX_CALL_CLEARED)
1720 && (tcall->rprev > rxi_HardAckRate)) {
1730 ReturnToServerPool(service);
1737 MUTEX_EXIT(&rx_serverPool_lock);
1738 MUTEX_ENTER(&call->lock);
1740 if (call->flags & RX_CALL_WAIT_PROC) {
1741 call->flags &= ~RX_CALL_WAIT_PROC;
1742 MUTEX_ENTER(&rx_waiting_mutex);
1744 MUTEX_EXIT(&rx_waiting_mutex);
1747 if (call->state != RX_STATE_PRECALL || call->error) {
1748 MUTEX_EXIT(&call->lock);
1749 MUTEX_ENTER(&rx_serverPool_lock);
1750 ReturnToServerPool(service);
1755 if (queue_IsEmpty(&call->rq)
1756 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1757 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1759 CLEAR_CALL_QUEUE_LOCK(call);
1762 /* If there are no eligible incoming calls, add this process
1763 * to the idle server queue, to wait for one */
1767 *socketp = OSI_NULLSOCKET;
1769 sq->socketp = socketp;
1770 queue_Append(&rx_idleServerQueue, sq);
1771 #ifndef AFS_AIX41_ENV
1772 rx_waitForPacket = sq;
1774 rx_waitingForPacket = sq;
1775 #endif /* AFS_AIX41_ENV */
1777 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1779 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1780 MUTEX_EXIT(&rx_serverPool_lock);
1781 return (struct rx_call *)0;
1784 } while (!(call = sq->newcall)
1785 && !(socketp && *socketp != OSI_NULLSOCKET));
1786 MUTEX_EXIT(&rx_serverPool_lock);
1788 MUTEX_ENTER(&call->lock);
1794 MUTEX_ENTER(&freeSQEList_lock);
1795 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1796 rx_FreeSQEList = sq;
1797 MUTEX_EXIT(&freeSQEList_lock);
1800 clock_GetTime(&call->startTime);
1801 call->state = RX_STATE_ACTIVE;
1802 call->mode = RX_MODE_RECEIVING;
1803 #ifdef RX_KERNEL_TRACE
1804 if (ICL_SETACTIVE(afs_iclSetp)) {
1805 int glockOwner = ISAFS_GLOCK();
1808 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1809 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1816 rxi_calltrace(RX_CALL_START, call);
1817 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
1818 call->conn->service->servicePort, call->conn->service->serviceId,
1821 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1822 MUTEX_EXIT(&call->lock);
1824 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1829 #else /* RX_ENABLE_LOCKS */
1831 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1833 struct rx_serverQueueEntry *sq;
1834 struct rx_call *call = (struct rx_call *)0, *choice2;
1835 struct rx_service *service = NULL;
1839 MUTEX_ENTER(&freeSQEList_lock);
1841 if ((sq = rx_FreeSQEList)) {
1842 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1843 MUTEX_EXIT(&freeSQEList_lock);
1844 } else { /* otherwise allocate a new one and return that */
1845 MUTEX_EXIT(&freeSQEList_lock);
1846 sq = (struct rx_serverQueueEntry *)
1847 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1848 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1849 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1851 MUTEX_ENTER(&sq->lock);
1853 if (cur_service != NULL) {
1854 cur_service->nRequestsRunning--;
1855 MUTEX_ENTER(&rx_quota_mutex);
1856 if (cur_service->nRequestsRunning < cur_service->minProcs)
1859 MUTEX_EXIT(&rx_quota_mutex);
1861 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1862 struct rx_call *tcall, *ncall;
1863 /* Scan for eligible incoming calls. A call is not eligible
1864 * if the maximum number of calls for its service type are
1865 * already executing */
1866 /* One thread will process calls FCFS (to prevent starvation),
1867 * while the other threads may run ahead looking for calls which
1868 * have all their input data available immediately. This helps
1869 * keep threads from blocking, waiting for data from the client. */
1870 choice2 = (struct rx_call *)0;
1871 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1872 service = tcall->conn->service;
1873 if (QuotaOK(service)) {
1874 MUTEX_ENTER(&rx_pthread_mutex);
1875 if (tno == rxi_fcfs_thread_num
1876 || !tcall->queue_item_header.next) {
1877 MUTEX_EXIT(&rx_pthread_mutex);
1878 /* If we're the fcfs thread, then we'll just use
1879 * this call. If we haven't been able to find an optimal
1880 * choice, and we're at the end of the list, then use a
1881 * 2d choice if one has been identified. Otherwise... */
1882 call = (choice2 ? choice2 : tcall);
1883 service = call->conn->service;
1885 MUTEX_EXIT(&rx_pthread_mutex);
1886 if (!queue_IsEmpty(&tcall->rq)) {
1887 struct rx_packet *rp;
1888 rp = queue_First(&tcall->rq, rx_packet);
1889 if (rp->header.seq == 1
1891 || (rp->header.flags & RX_LAST_PACKET))) {
1893 } else if (rxi_2dchoice && !choice2
1894 && !(tcall->flags & RX_CALL_CLEARED)
1895 && (tcall->rprev > rxi_HardAckRate)) {
1909 /* we can't schedule a call if there's no data!!! */
1910 /* send an ack if there's no data, if we're missing the
1911 * first packet, or we're missing something between first
1912 * and last -- there's a "hole" in the incoming data. */
1913 if (queue_IsEmpty(&call->rq)
1914 || queue_First(&call->rq, rx_packet)->header.seq != 1
1915 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1916 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1918 call->flags &= (~RX_CALL_WAIT_PROC);
1919 service->nRequestsRunning++;
1920 /* just started call in minProcs pool, need fewer to maintain
1922 MUTEX_ENTER(&rx_quota_mutex);
1923 if (service->nRequestsRunning <= service->minProcs)
1926 MUTEX_EXIT(&rx_quota_mutex);
1928 /* MUTEX_EXIT(&call->lock); */
1930 /* If there are no eligible incoming calls, add this process
1931 * to the idle server queue, to wait for one */
1934 *socketp = OSI_NULLSOCKET;
1936 sq->socketp = socketp;
1937 queue_Append(&rx_idleServerQueue, sq);
1941 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1943 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1944 return (struct rx_call *)0;
1947 } while (!(call = sq->newcall)
1948 && !(socketp && *socketp != OSI_NULLSOCKET));
1950 MUTEX_EXIT(&sq->lock);
1952 MUTEX_ENTER(&freeSQEList_lock);
1953 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1954 rx_FreeSQEList = sq;
1955 MUTEX_EXIT(&freeSQEList_lock);
1958 clock_GetTime(&call->startTime);
1959 call->state = RX_STATE_ACTIVE;
1960 call->mode = RX_MODE_RECEIVING;
1961 #ifdef RX_KERNEL_TRACE
1962 if (ICL_SETACTIVE(afs_iclSetp)) {
1963 int glockOwner = ISAFS_GLOCK();
1966 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1967 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1974 rxi_calltrace(RX_CALL_START, call);
1975 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
1976 call->conn->service->servicePort, call->conn->service->serviceId,
1979 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1986 #endif /* RX_ENABLE_LOCKS */
1990 /* Establish a procedure to be called when a packet arrives for a
1991 * call. This routine will be called at most once after each call,
1992 * and will also be called if there is an error condition on the or
1993 * the call is complete. Used by multi rx to build a selection
1994 * function which determines which of several calls is likely to be a
1995 * good one to read from.
1996 * NOTE: the way this is currently implemented it is probably only a
1997 * good idea to (1) use it immediately after a newcall (clients only)
1998 * and (2) only use it once. Other uses currently void your warranty
2001 rx_SetArrivalProc(struct rx_call *call,
2002 void (*proc) (struct rx_call * call,
2005 void * handle, int arg)
2007 call->arrivalProc = proc;
2008 call->arrivalProcHandle = handle;
2009 call->arrivalProcArg = arg;
2012 /* Call is finished (possibly prematurely). Return rc to the peer, if
2013 * appropriate, and return the final error code from the conversation
2017 rx_EndCall(struct rx_call *call, afs_int32 rc)
2019 struct rx_connection *conn = call->conn;
2020 struct rx_service *service;
2024 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2025 call, rc, call->error, call->abortCode));
2028 MUTEX_ENTER(&call->lock);
2030 if (rc == 0 && call->error == 0) {
2031 call->abortCode = 0;
2032 call->abortCount = 0;
2035 call->arrivalProc = (void (*)())0;
2036 if (rc && call->error == 0) {
2037 rxi_CallError(call, rc);
2038 /* Send an abort message to the peer if this error code has
2039 * only just been set. If it was set previously, assume the
2040 * peer has already been sent the error code or will request it
2042 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2044 if (conn->type == RX_SERVER_CONNECTION) {
2045 /* Make sure reply or at least dummy reply is sent */
2046 if (call->mode == RX_MODE_RECEIVING) {
2047 rxi_WriteProc(call, 0, 0);
2049 if (call->mode == RX_MODE_SENDING) {
2050 rxi_FlushWrite(call);
2052 service = conn->service;
2053 rxi_calltrace(RX_CALL_END, call);
2054 /* Call goes to hold state until reply packets are acknowledged */
2055 if (call->tfirst + call->nSoftAcked < call->tnext) {
2056 call->state = RX_STATE_HOLD;
2058 call->state = RX_STATE_DALLY;
2059 rxi_ClearTransmitQueue(call, 0);
2060 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
2061 rxevent_Cancel(call->keepAliveEvent, call,
2062 RX_CALL_REFCOUNT_ALIVE);
2064 } else { /* Client connection */
2066 /* Make sure server receives input packets, in the case where
2067 * no reply arguments are expected */
2068 if ((call->mode == RX_MODE_SENDING)
2069 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2070 (void)rxi_ReadProc(call, &dummy, 1);
2073 /* If we had an outstanding delayed ack, be nice to the server
2074 * and force-send it now.
2076 if (call->delayedAckEvent) {
2077 rxevent_Cancel(call->delayedAckEvent, call,
2078 RX_CALL_REFCOUNT_DELAY);
2079 call->delayedAckEvent = NULL;
2080 rxi_SendDelayedAck(NULL, call, NULL);
2083 /* We need to release the call lock since it's lower than the
2084 * conn_call_lock and we don't want to hold the conn_call_lock
2085 * over the rx_ReadProc call. The conn_call_lock needs to be held
2086 * here for the case where rx_NewCall is perusing the calls on
2087 * the connection structure. We don't want to signal until
2088 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2089 * have checked this call, found it active and by the time it
2090 * goes to sleep, will have missed the signal.
2092 MUTEX_EXIT(&call->lock);
2093 MUTEX_ENTER(&conn->conn_call_lock);
2094 MUTEX_ENTER(&call->lock);
2095 MUTEX_ENTER(&conn->conn_data_lock);
2096 conn->flags |= RX_CONN_BUSY;
2097 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2098 MUTEX_EXIT(&conn->conn_data_lock);
2099 #ifdef RX_ENABLE_LOCKS
2100 CV_BROADCAST(&conn->conn_call_cv);
2105 #ifdef RX_ENABLE_LOCKS
2107 MUTEX_EXIT(&conn->conn_data_lock);
2109 #endif /* RX_ENABLE_LOCKS */
2110 call->state = RX_STATE_DALLY;
2112 error = call->error;
2114 /* currentPacket, nLeft, and NFree must be zeroed here, because
2115 * ResetCall cannot: ResetCall may be called at splnet(), in the
2116 * kernel version, and may interrupt the macros rx_Read or
2117 * rx_Write, which run at normal priority for efficiency. */
2118 if (call->currentPacket) {
2119 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2120 rxi_FreePacket(call->currentPacket);
2121 call->currentPacket = (struct rx_packet *)0;
2124 call->nLeft = call->nFree = call->curlen = 0;
2126 /* Free any packets from the last call to ReadvProc/WritevProc */
2127 #ifdef RXDEBUG_PACKET
2129 #endif /* RXDEBUG_PACKET */
2130 rxi_FreePackets(0, &call->iovq);
2132 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2133 MUTEX_EXIT(&call->lock);
2134 if (conn->type == RX_CLIENT_CONNECTION) {
2135 MUTEX_ENTER(&conn->conn_data_lock);
2136 conn->flags &= ~RX_CONN_BUSY;
2137 MUTEX_EXIT(&conn->conn_data_lock);
2138 MUTEX_EXIT(&conn->conn_call_lock);
2142 * Map errors to the local host's errno.h format.
2144 error = ntoh_syserr_conv(error);
2148 #if !defined(KERNEL)
2150 /* Call this routine when shutting down a server or client (especially
2151 * clients). This will allow Rx to gracefully garbage collect server
2152 * connections, and reduce the number of retries that a server might
2153 * make to a dead client.
2154 * This is not quite right, since some calls may still be ongoing and
2155 * we can't lock them to destroy them. */
2159 struct rx_connection **conn_ptr, **conn_end;
2163 if (rxinit_status == 1) {
2165 return; /* Already shutdown. */
2167 rxi_DeleteCachedConnections();
2168 if (rx_connHashTable) {
2169 MUTEX_ENTER(&rx_connHashTable_lock);
2170 for (conn_ptr = &rx_connHashTable[0], conn_end =
2171 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2173 struct rx_connection *conn, *next;
2174 for (conn = *conn_ptr; conn; conn = next) {
2176 if (conn->type == RX_CLIENT_CONNECTION) {
2177 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2179 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2180 #ifdef RX_ENABLE_LOCKS
2181 rxi_DestroyConnectionNoLock(conn);
2182 #else /* RX_ENABLE_LOCKS */
2183 rxi_DestroyConnection(conn);
2184 #endif /* RX_ENABLE_LOCKS */
2188 #ifdef RX_ENABLE_LOCKS
2189 while (rx_connCleanup_list) {
2190 struct rx_connection *conn;
2191 conn = rx_connCleanup_list;
2192 rx_connCleanup_list = rx_connCleanup_list->next;
2193 MUTEX_EXIT(&rx_connHashTable_lock);
2194 rxi_CleanupConnection(conn);
2195 MUTEX_ENTER(&rx_connHashTable_lock);
2197 MUTEX_EXIT(&rx_connHashTable_lock);
2198 #endif /* RX_ENABLE_LOCKS */
2203 afs_winsockCleanup();
2211 /* if we wakeup packet waiter too often, can get in loop with two
2212 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2214 rxi_PacketsUnWait(void)
2216 if (!rx_waitingForPackets) {
2220 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2221 return; /* still over quota */
2224 rx_waitingForPackets = 0;
2225 #ifdef RX_ENABLE_LOCKS
2226 CV_BROADCAST(&rx_waitingForPackets_cv);
2228 osi_rxWakeup(&rx_waitingForPackets);
2234 /* ------------------Internal interfaces------------------------- */
2236 /* Return this process's service structure for the
2237 * specified socket and service */
2239 rxi_FindService(osi_socket socket, u_short serviceId)
2241 struct rx_service **sp;
2242 for (sp = &rx_services[0]; *sp; sp++) {
2243 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2249 #ifdef RXDEBUG_PACKET
2250 #ifdef KDUMP_RX_LOCK
2251 static struct rx_call_rx_lock *rx_allCallsp = 0;
2253 static struct rx_call *rx_allCallsp = 0;
2255 #endif /* RXDEBUG_PACKET */
2257 /* Allocate a call structure, for the indicated channel of the
2258 * supplied connection. The mode and state of the call must be set by
2259 * the caller. Returns the call with mutex locked. */
2261 rxi_NewCall(struct rx_connection *conn, int channel)
2263 struct rx_call *call;
2264 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2265 struct rx_call *cp; /* Call pointer temp */
2266 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2267 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2269 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2271 /* Grab an existing call structure, or allocate a new one.
2272 * Existing call structures are assumed to have been left reset by
2274 MUTEX_ENTER(&rx_freeCallQueue_lock);
2276 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2278 * EXCEPT that the TQ might not yet be cleared out.
2279 * Skip over those with in-use TQs.
2282 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2283 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2289 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2290 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2291 call = queue_First(&rx_freeCallQueue, rx_call);
2292 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2294 if (rx_stats_active)
2295 rx_MutexDecrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2296 MUTEX_EXIT(&rx_freeCallQueue_lock);
2297 MUTEX_ENTER(&call->lock);
2298 CLEAR_CALL_QUEUE_LOCK(call);
2299 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2300 /* Now, if TQ wasn't cleared earlier, do it now. */
2301 rxi_WaitforTQBusy(call);
2302 if (call->flags & RX_CALL_TQ_CLEARME) {
2303 rxi_ClearTransmitQueue(call, 1);
2304 /*queue_Init(&call->tq);*/
2306 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2307 /* Bind the call to its connection structure */
2309 rxi_ResetCall(call, 1);
2312 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2313 #ifdef RXDEBUG_PACKET
2314 call->allNextp = rx_allCallsp;
2315 rx_allCallsp = call;
2317 #endif /* RXDEBUG_PACKET */
2318 rx_MutexIncrement(rx_stats.nCallStructs, rx_stats_mutex);
2320 MUTEX_EXIT(&rx_freeCallQueue_lock);
2321 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2322 MUTEX_ENTER(&call->lock);
2323 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2324 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2325 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2327 /* Initialize once-only items */
2328 queue_Init(&call->tq);
2329 queue_Init(&call->rq);
2330 queue_Init(&call->iovq);
2331 #ifdef RXDEBUG_PACKET
2332 call->rqc = call->tqc = call->iovqc = 0;
2333 #endif /* RXDEBUG_PACKET */
2334 /* Bind the call to its connection structure (prereq for reset) */
2336 rxi_ResetCall(call, 1);
2338 call->channel = channel;
2339 call->callNumber = &conn->callNumber[channel];
2340 call->rwind = conn->rwind[channel];
2341 call->twind = conn->twind[channel];
2342 /* Note that the next expected call number is retained (in
2343 * conn->callNumber[i]), even if we reallocate the call structure
2345 conn->call[channel] = call;
2346 /* if the channel's never been used (== 0), we should start at 1, otherwise
2347 * the call number is valid from the last time this channel was used */
2348 if (*call->callNumber == 0)
2349 *call->callNumber = 1;
2354 /* A call has been inactive long enough that so we can throw away
2355 * state, including the call structure, which is placed on the call
2357 * Call is locked upon entry.
2358 * haveCTLock set if called from rxi_ReapConnections
2360 #ifdef RX_ENABLE_LOCKS
2362 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2363 #else /* RX_ENABLE_LOCKS */
2365 rxi_FreeCall(struct rx_call *call)
2366 #endif /* RX_ENABLE_LOCKS */
2368 int channel = call->channel;
2369 struct rx_connection *conn = call->conn;
2372 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2373 (*call->callNumber)++;
2374 rxi_ResetCall(call, 0);
2375 call->conn->call[channel] = (struct rx_call *)0;
2377 MUTEX_ENTER(&rx_freeCallQueue_lock);
2378 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2379 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2380 /* A call may be free even though its transmit queue is still in use.
2381 * Since we search the call list from head to tail, put busy calls at
2382 * the head of the list, and idle calls at the tail.
2384 if (call->flags & RX_CALL_TQ_BUSY)
2385 queue_Prepend(&rx_freeCallQueue, call);
2387 queue_Append(&rx_freeCallQueue, call);
2388 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2389 queue_Append(&rx_freeCallQueue, call);
2390 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2391 if (rx_stats_active)
2392 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2393 MUTEX_EXIT(&rx_freeCallQueue_lock);
2395 /* Destroy the connection if it was previously slated for
2396 * destruction, i.e. the Rx client code previously called
2397 * rx_DestroyConnection (client connections), or
2398 * rxi_ReapConnections called the same routine (server
2399 * connections). Only do this, however, if there are no
2400 * outstanding calls. Note that for fine grain locking, there appears
2401 * to be a deadlock in that rxi_FreeCall has a call locked and
2402 * DestroyConnectionNoLock locks each call in the conn. But note a
2403 * few lines up where we have removed this call from the conn.
2404 * If someone else destroys a connection, they either have no
2405 * call lock held or are going through this section of code.
2407 MUTEX_ENTER(&conn->conn_data_lock);
2408 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2410 MUTEX_EXIT(&conn->conn_data_lock);
2411 #ifdef RX_ENABLE_LOCKS
2413 rxi_DestroyConnectionNoLock(conn);
2415 rxi_DestroyConnection(conn);
2416 #else /* RX_ENABLE_LOCKS */
2417 rxi_DestroyConnection(conn);
2418 #endif /* RX_ENABLE_LOCKS */
2420 MUTEX_EXIT(&conn->conn_data_lock);
2424 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2426 rxi_Alloc(size_t size)
2430 if (rx_stats_active)
2431 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2434 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2435 afs_osi_Alloc_NoSleep(size);
2440 osi_Panic("rxi_Alloc error");
2446 rxi_Free(void *addr, size_t size)
2448 if (rx_stats_active)
2449 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2450 osi_Free(addr, size);
2454 rxi_SetPeerMtu(afs_uint32 host, afs_uint32 port, int mtu)
2456 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2457 struct rx_peer *peer = NULL, *next = NULL;
2460 MUTEX_ENTER(&rx_peerHashTable_lock);
2462 peer_ptr = &rx_peerHashTable[0];
2463 peer_end = &rx_peerHashTable[rx_hashTableSize];
2466 for ( ; peer_ptr < peer_end; peer_ptr++) {
2469 for ( ; peer; peer = next) {
2471 if (host == peer->host)
2476 hashIndex = PEER_HASH(host, port);
2477 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2478 if ((peer->host == host) && (peer->port == port))
2485 MUTEX_EXIT(&rx_peerHashTable_lock);
2487 MUTEX_ENTER(&peer->peer_lock);
2488 /* We don't handle dropping below min, so don't */
2489 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2490 peer->ifMTU=MIN(mtu, peer->ifMTU);
2491 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2492 MUTEX_EXIT(&peer->peer_lock);
2494 MUTEX_ENTER(&rx_peerHashTable_lock);
2498 /* pick up where we left off */
2502 MUTEX_EXIT(&rx_peerHashTable_lock);
2505 /* Find the peer process represented by the supplied (host,port)
2506 * combination. If there is no appropriate active peer structure, a
2507 * new one will be allocated and initialized
2508 * The origPeer, if set, is a pointer to a peer structure on which the
2509 * refcount will be be decremented. This is used to replace the peer
2510 * structure hanging off a connection structure */
2512 rxi_FindPeer(afs_uint32 host, u_short port,
2513 struct rx_peer *origPeer, int create)
2517 hashIndex = PEER_HASH(host, port);
2518 MUTEX_ENTER(&rx_peerHashTable_lock);
2519 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2520 if ((pp->host == host) && (pp->port == port))
2525 pp = rxi_AllocPeer(); /* This bzero's *pp */
2526 pp->host = host; /* set here or in InitPeerParams is zero */
2528 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2529 queue_Init(&pp->congestionQueue);
2530 queue_Init(&pp->rpcStats);
2531 pp->next = rx_peerHashTable[hashIndex];
2532 rx_peerHashTable[hashIndex] = pp;
2533 rxi_InitPeerParams(pp);
2534 if (rx_stats_active)
2535 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2542 origPeer->refCount--;
2543 MUTEX_EXIT(&rx_peerHashTable_lock);
2548 /* Find the connection at (host, port) started at epoch, and with the
2549 * given connection id. Creates the server connection if necessary.
2550 * The type specifies whether a client connection or a server
2551 * connection is desired. In both cases, (host, port) specify the
2552 * peer's (host, pair) pair. Client connections are not made
2553 * automatically by this routine. The parameter socket gives the
2554 * socket descriptor on which the packet was received. This is used,
2555 * in the case of server connections, to check that *new* connections
2556 * come via a valid (port, serviceId). Finally, the securityIndex
2557 * parameter must match the existing index for the connection. If a
2558 * server connection is created, it will be created using the supplied
2559 * index, if the index is valid for this service */
2560 struct rx_connection *
2561 rxi_FindConnection(osi_socket socket, afs_int32 host,
2562 u_short port, u_short serviceId, afs_uint32 cid,
2563 afs_uint32 epoch, int type, u_int securityIndex)
2565 int hashindex, flag, i;
2566 struct rx_connection *conn;
2567 hashindex = CONN_HASH(host, port, cid, epoch, type);
2568 MUTEX_ENTER(&rx_connHashTable_lock);
2569 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2570 rx_connHashTable[hashindex],
2573 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2574 && (epoch == conn->epoch)) {
2575 struct rx_peer *pp = conn->peer;
2576 if (securityIndex != conn->securityIndex) {
2577 /* this isn't supposed to happen, but someone could forge a packet
2578 * like this, and there seems to be some CM bug that makes this
2579 * happen from time to time -- in which case, the fileserver
2581 MUTEX_EXIT(&rx_connHashTable_lock);
2582 return (struct rx_connection *)0;
2584 if (pp->host == host && pp->port == port)
2586 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2588 /* So what happens when it's a callback connection? */
2589 if ( /*type == RX_CLIENT_CONNECTION && */
2590 (conn->epoch & 0x80000000))
2594 /* the connection rxLastConn that was used the last time is not the
2595 ** one we are looking for now. Hence, start searching in the hash */
2597 conn = rx_connHashTable[hashindex];
2602 struct rx_service *service;
2603 if (type == RX_CLIENT_CONNECTION) {
2604 MUTEX_EXIT(&rx_connHashTable_lock);
2605 return (struct rx_connection *)0;
2607 service = rxi_FindService(socket, serviceId);
2608 if (!service || (securityIndex >= service->nSecurityObjects)
2609 || (service->securityObjects[securityIndex] == 0)) {
2610 MUTEX_EXIT(&rx_connHashTable_lock);
2611 return (struct rx_connection *)0;
2613 conn = rxi_AllocConnection(); /* This bzero's the connection */
2614 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2615 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2616 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2617 conn->next = rx_connHashTable[hashindex];
2618 rx_connHashTable[hashindex] = conn;
2619 conn->peer = rxi_FindPeer(host, port, 0, 1);
2620 conn->type = RX_SERVER_CONNECTION;
2621 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2622 conn->epoch = epoch;
2623 conn->cid = cid & RX_CIDMASK;
2624 /* conn->serial = conn->lastSerial = 0; */
2625 /* conn->timeout = 0; */
2626 conn->ackRate = RX_FAST_ACK_RATE;
2627 conn->service = service;
2628 conn->serviceId = serviceId;
2629 conn->securityIndex = securityIndex;
2630 conn->securityObject = service->securityObjects[securityIndex];
2631 conn->nSpecific = 0;
2632 conn->specific = NULL;
2633 rx_SetConnDeadTime(conn, service->connDeadTime);
2634 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2635 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2636 for (i = 0; i < RX_MAXCALLS; i++) {
2637 conn->twind[i] = rx_initSendWindow;
2638 conn->rwind[i] = rx_initReceiveWindow;
2640 /* Notify security object of the new connection */
2641 RXS_NewConnection(conn->securityObject, conn);
2642 /* XXXX Connection timeout? */
2643 if (service->newConnProc)
2644 (*service->newConnProc) (conn);
2645 if (rx_stats_active)
2646 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2649 MUTEX_ENTER(&conn->conn_data_lock);
2651 MUTEX_EXIT(&conn->conn_data_lock);
2653 rxLastConn = conn; /* store this connection as the last conn used */
2654 MUTEX_EXIT(&rx_connHashTable_lock);
2658 /* There are two packet tracing routines available for testing and monitoring
2659 * Rx. One is called just after every packet is received and the other is
2660 * called just before every packet is sent. Received packets, have had their
2661 * headers decoded, and packets to be sent have not yet had their headers
2662 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2663 * containing the network address. Both can be modified. The return value, if
2664 * non-zero, indicates that the packet should be dropped. */
2666 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2667 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2669 /* A packet has been received off the interface. Np is the packet, socket is
2670 * the socket number it was received from (useful in determining which service
2671 * this packet corresponds to), and (host, port) reflect the host,port of the
2672 * sender. This call returns the packet to the caller if it is finished with
2673 * it, rather than de-allocating it, just as a small performance hack */
2676 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
2677 afs_uint32 host, u_short port, int *tnop,
2678 struct rx_call **newcallp)
2680 struct rx_call *call;
2681 struct rx_connection *conn;
2683 afs_uint32 currentCallNumber;
2689 struct rx_packet *tnp;
2692 /* We don't print out the packet until now because (1) the time may not be
2693 * accurate enough until now in the lwp implementation (rx_Listener only gets
2694 * the time after the packet is read) and (2) from a protocol point of view,
2695 * this is the first time the packet has been seen */
2696 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2697 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2698 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT,
2699 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2700 np->header.epoch, np->header.cid, np->header.callNumber,
2701 np->header.seq, np->header.flags, np));
2704 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2705 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2708 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2709 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2712 /* If an input tracer function is defined, call it with the packet and
2713 * network address. Note this function may modify its arguments. */
2714 if (rx_justReceived) {
2715 struct sockaddr_in addr;
2717 addr.sin_family = AF_INET;
2718 addr.sin_port = port;
2719 addr.sin_addr.s_addr = host;
2720 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2721 addr.sin_len = sizeof(addr);
2722 #endif /* AFS_OSF_ENV */
2723 drop = (*rx_justReceived) (np, &addr);
2724 /* drop packet if return value is non-zero */
2727 port = addr.sin_port; /* in case fcn changed addr */
2728 host = addr.sin_addr.s_addr;
2732 /* If packet was not sent by the client, then *we* must be the client */
2733 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2734 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2736 /* Find the connection (or fabricate one, if we're the server & if
2737 * necessary) associated with this packet */
2739 rxi_FindConnection(socket, host, port, np->header.serviceId,
2740 np->header.cid, np->header.epoch, type,
2741 np->header.securityIndex);
2744 /* If no connection found or fabricated, just ignore the packet.
2745 * (An argument could be made for sending an abort packet for
2750 MUTEX_ENTER(&conn->conn_data_lock);
2751 if (conn->maxSerial < np->header.serial)
2752 conn->maxSerial = np->header.serial;
2753 MUTEX_EXIT(&conn->conn_data_lock);
2755 /* If the connection is in an error state, send an abort packet and ignore
2756 * the incoming packet */
2758 /* Don't respond to an abort packet--we don't want loops! */
2759 MUTEX_ENTER(&conn->conn_data_lock);
2760 if (np->header.type != RX_PACKET_TYPE_ABORT)
2761 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2763 MUTEX_EXIT(&conn->conn_data_lock);
2767 /* Check for connection-only requests (i.e. not call specific). */
2768 if (np->header.callNumber == 0) {
2769 switch (np->header.type) {
2770 case RX_PACKET_TYPE_ABORT: {
2771 /* What if the supplied error is zero? */
2772 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2773 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2774 rxi_ConnectionError(conn, errcode);
2775 MUTEX_ENTER(&conn->conn_data_lock);
2777 MUTEX_EXIT(&conn->conn_data_lock);
2780 case RX_PACKET_TYPE_CHALLENGE:
2781 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2782 MUTEX_ENTER(&conn->conn_data_lock);
2784 MUTEX_EXIT(&conn->conn_data_lock);
2786 case RX_PACKET_TYPE_RESPONSE:
2787 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2788 MUTEX_ENTER(&conn->conn_data_lock);
2790 MUTEX_EXIT(&conn->conn_data_lock);
2792 case RX_PACKET_TYPE_PARAMS:
2793 case RX_PACKET_TYPE_PARAMS + 1:
2794 case RX_PACKET_TYPE_PARAMS + 2:
2795 /* ignore these packet types for now */
2796 MUTEX_ENTER(&conn->conn_data_lock);
2798 MUTEX_EXIT(&conn->conn_data_lock);
2803 /* Should not reach here, unless the peer is broken: send an
2805 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2806 MUTEX_ENTER(&conn->conn_data_lock);
2807 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2809 MUTEX_EXIT(&conn->conn_data_lock);
2814 channel = np->header.cid & RX_CHANNELMASK;
2815 call = conn->call[channel];
2816 #ifdef RX_ENABLE_LOCKS
2818 MUTEX_ENTER(&call->lock);
2819 /* Test to see if call struct is still attached to conn. */
2820 if (call != conn->call[channel]) {
2822 MUTEX_EXIT(&call->lock);
2823 if (type == RX_SERVER_CONNECTION) {
2824 call = conn->call[channel];
2825 /* If we started with no call attached and there is one now,
2826 * another thread is also running this routine and has gotten
2827 * the connection channel. We should drop this packet in the tests
2828 * below. If there was a call on this connection and it's now
2829 * gone, then we'll be making a new call below.
2830 * If there was previously a call and it's now different then
2831 * the old call was freed and another thread running this routine
2832 * has created a call on this channel. One of these two threads
2833 * has a packet for the old call and the code below handles those
2837 MUTEX_ENTER(&call->lock);
2839 /* This packet can't be for this call. If the new call address is
2840 * 0 then no call is running on this channel. If there is a call
2841 * then, since this is a client connection we're getting data for
2842 * it must be for the previous call.
2844 if (rx_stats_active)
2845 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2846 MUTEX_ENTER(&conn->conn_data_lock);
2848 MUTEX_EXIT(&conn->conn_data_lock);
2853 currentCallNumber = conn->callNumber[channel];
2855 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2856 if (np->header.callNumber < currentCallNumber) {
2857 if (rx_stats_active)
2858 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2859 #ifdef RX_ENABLE_LOCKS
2861 MUTEX_EXIT(&call->lock);
2863 MUTEX_ENTER(&conn->conn_data_lock);
2865 MUTEX_EXIT(&conn->conn_data_lock);
2869 MUTEX_ENTER(&conn->conn_call_lock);
2870 call = rxi_NewCall(conn, channel);
2871 MUTEX_EXIT(&conn->conn_call_lock);
2872 *call->callNumber = np->header.callNumber;
2874 if (np->header.callNumber == 0)
2875 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%.06d len %d",
2876 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
2877 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
2878 np->header.flags, np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2880 call->state = RX_STATE_PRECALL;
2881 clock_GetTime(&call->queueTime);
2882 hzero(call->bytesSent);
2883 hzero(call->bytesRcvd);
2885 * If the number of queued calls exceeds the overload
2886 * threshold then abort this call.
2888 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2889 struct rx_packet *tp;
2891 rxi_CallError(call, rx_BusyError);
2892 tp = rxi_SendCallAbort(call, np, 1, 0);
2893 MUTEX_EXIT(&call->lock);
2894 MUTEX_ENTER(&conn->conn_data_lock);
2896 MUTEX_EXIT(&conn->conn_data_lock);
2897 if (rx_stats_active)
2898 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2901 rxi_KeepAliveOn(call);
2902 } else if (np->header.callNumber != currentCallNumber) {
2903 /* Wait until the transmit queue is idle before deciding
2904 * whether to reset the current call. Chances are that the
2905 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2908 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2909 while ((call->state == RX_STATE_ACTIVE)
2910 && (call->flags & RX_CALL_TQ_BUSY)) {
2911 call->flags |= RX_CALL_TQ_WAIT;
2913 #ifdef RX_ENABLE_LOCKS
2914 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2915 CV_WAIT(&call->cv_tq, &call->lock);
2916 #else /* RX_ENABLE_LOCKS */
2917 osi_rxSleep(&call->tq);
2918 #endif /* RX_ENABLE_LOCKS */
2920 if (call->tqWaiters == 0)
2921 call->flags &= ~RX_CALL_TQ_WAIT;
2923 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2924 /* If the new call cannot be taken right now send a busy and set
2925 * the error condition in this call, so that it terminates as
2926 * quickly as possible */
2927 if (call->state == RX_STATE_ACTIVE) {
2928 struct rx_packet *tp;
2930 rxi_CallError(call, RX_CALL_DEAD);
2931 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2933 MUTEX_EXIT(&call->lock);
2934 MUTEX_ENTER(&conn->conn_data_lock);
2936 MUTEX_EXIT(&conn->conn_data_lock);
2939 rxi_ResetCall(call, 0);
2940 *call->callNumber = np->header.callNumber;
2942 if (np->header.callNumber == 0)
2943 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%06d len %d",
2944 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
2945 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
2946 np->header.flags, np, np->retryTime.sec, np->retryTime.usec, np->length));
2948 call->state = RX_STATE_PRECALL;
2949 clock_GetTime(&call->queueTime);
2950 hzero(call->bytesSent);
2951 hzero(call->bytesRcvd);
2953 * If the number of queued calls exceeds the overload
2954 * threshold then abort this call.
2956 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2957 struct rx_packet *tp;
2959 rxi_CallError(call, rx_BusyError);
2960 tp = rxi_SendCallAbort(call, np, 1, 0);
2961 MUTEX_EXIT(&call->lock);
2962 MUTEX_ENTER(&conn->conn_data_lock);
2964 MUTEX_EXIT(&conn->conn_data_lock);
2965 if (rx_stats_active)
2966 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2969 rxi_KeepAliveOn(call);
2971 /* Continuing call; do nothing here. */
2973 } else { /* we're the client */
2974 /* Ignore all incoming acknowledgements for calls in DALLY state */
2975 if (call && (call->state == RX_STATE_DALLY)
2976 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2977 if (rx_stats_active)
2978 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
2979 #ifdef RX_ENABLE_LOCKS
2981 MUTEX_EXIT(&call->lock);
2984 MUTEX_ENTER(&conn->conn_data_lock);
2986 MUTEX_EXIT(&conn->conn_data_lock);
2990 /* Ignore anything that's not relevant to the current call. If there
2991 * isn't a current call, then no packet is relevant. */
2992 if (!call || (np->header.callNumber != currentCallNumber)) {
2993 if (rx_stats_active)
2994 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2995 #ifdef RX_ENABLE_LOCKS
2997 MUTEX_EXIT(&call->lock);
3000 MUTEX_ENTER(&conn->conn_data_lock);
3002 MUTEX_EXIT(&conn->conn_data_lock);
3005 /* If the service security object index stamped in the packet does not
3006 * match the connection's security index, ignore the packet */
3007 if (np->header.securityIndex != conn->securityIndex) {
3008 #ifdef RX_ENABLE_LOCKS
3009 MUTEX_EXIT(&call->lock);
3011 MUTEX_ENTER(&conn->conn_data_lock);
3013 MUTEX_EXIT(&conn->conn_data_lock);
3017 /* If we're receiving the response, then all transmit packets are
3018 * implicitly acknowledged. Get rid of them. */
3019 if (np->header.type == RX_PACKET_TYPE_DATA) {
3020 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3021 /* XXX Hack. Because we must release the global rx lock when
3022 * sending packets (osi_NetSend) we drop all acks while we're
3023 * traversing the tq in rxi_Start sending packets out because
3024 * packets may move to the freePacketQueue as result of being here!
3025 * So we drop these packets until we're safely out of the
3026 * traversing. Really ugly!
3027 * For fine grain RX locking, we set the acked field in the
3028 * packets and let rxi_Start remove them from the transmit queue.
3030 if (call->flags & RX_CALL_TQ_BUSY) {
3031 #ifdef RX_ENABLE_LOCKS
3032 rxi_SetAcksInTransmitQueue(call);
3035 return np; /* xmitting; drop packet */
3038 rxi_ClearTransmitQueue(call, 0);
3040 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3041 rxi_ClearTransmitQueue(call, 0);
3042 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3044 if (np->header.type == RX_PACKET_TYPE_ACK) {
3045 /* now check to see if this is an ack packet acknowledging that the
3046 * server actually *lost* some hard-acked data. If this happens we
3047 * ignore this packet, as it may indicate that the server restarted in
3048 * the middle of a call. It is also possible that this is an old ack
3049 * packet. We don't abort the connection in this case, because this
3050 * *might* just be an old ack packet. The right way to detect a server
3051 * restart in the midst of a call is to notice that the server epoch
3053 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3054 * XXX unacknowledged. I think that this is off-by-one, but
3055 * XXX I don't dare change it just yet, since it will
3056 * XXX interact badly with the server-restart detection
3057 * XXX code in receiveackpacket. */
3058 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3059 if (rx_stats_active)
3060 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
3061 MUTEX_EXIT(&call->lock);
3062 MUTEX_ENTER(&conn->conn_data_lock);
3064 MUTEX_EXIT(&conn->conn_data_lock);
3068 } /* else not a data packet */
3071 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3072 /* Set remote user defined status from packet */
3073 call->remoteStatus = np->header.userStatus;
3075 /* Note the gap between the expected next packet and the actual
3076 * packet that arrived, when the new packet has a smaller serial number
3077 * than expected. Rioses frequently reorder packets all by themselves,
3078 * so this will be quite important with very large window sizes.
3079 * Skew is checked against 0 here to avoid any dependence on the type of
3080 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3082 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3083 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3084 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3086 MUTEX_ENTER(&conn->conn_data_lock);
3087 skew = conn->lastSerial - np->header.serial;
3088 conn->lastSerial = np->header.serial;
3089 MUTEX_EXIT(&conn->conn_data_lock);
3091 struct rx_peer *peer;
3093 if (skew > peer->inPacketSkew) {
3094 dpf(("*** In skew changed from %d to %d\n",
3095 peer->inPacketSkew, skew));
3096 peer->inPacketSkew = skew;
3100 /* Now do packet type-specific processing */
3101 switch (np->header.type) {
3102 case RX_PACKET_TYPE_DATA:
3103 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3106 case RX_PACKET_TYPE_ACK:
3107 /* Respond immediately to ack packets requesting acknowledgement
3109 if (np->header.flags & RX_REQUEST_ACK) {
3111 (void)rxi_SendCallAbort(call, 0, 1, 0);
3113 (void)rxi_SendAck(call, 0, np->header.serial,
3114 RX_ACK_PING_RESPONSE, 1);
3116 np = rxi_ReceiveAckPacket(call, np, 1);
3118 case RX_PACKET_TYPE_ABORT: {
3119 /* An abort packet: reset the call, passing the error up to the user. */
3120 /* What if error is zero? */
3121 /* What if the error is -1? the application will treat it as a timeout. */
3122 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3123 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
3124 rxi_CallError(call, errdata);
3125 MUTEX_EXIT(&call->lock);
3126 MUTEX_ENTER(&conn->conn_data_lock);
3128 MUTEX_EXIT(&conn->conn_data_lock);
3129 return np; /* xmitting; drop packet */
3131 case RX_PACKET_TYPE_BUSY:
3134 case RX_PACKET_TYPE_ACKALL:
3135 /* All packets acknowledged, so we can drop all packets previously
3136 * readied for sending */
3137 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3138 /* XXX Hack. We because we can't release the global rx lock when
3139 * sending packets (osi_NetSend) we drop all ack pkts while we're
3140 * traversing the tq in rxi_Start sending packets out because
3141 * packets may move to the freePacketQueue as result of being
3142 * here! So we drop these packets until we're safely out of the
3143 * traversing. Really ugly!
3144 * For fine grain RX locking, we set the acked field in the packets
3145 * and let rxi_Start remove the packets from the transmit queue.
3147 if (call->flags & RX_CALL_TQ_BUSY) {
3148 #ifdef RX_ENABLE_LOCKS
3149 rxi_SetAcksInTransmitQueue(call);
3151 #else /* RX_ENABLE_LOCKS */
3152 MUTEX_EXIT(&call->lock);
3153 MUTEX_ENTER(&conn->conn_data_lock);
3155 MUTEX_EXIT(&conn->conn_data_lock);
3156 return np; /* xmitting; drop packet */
3157 #endif /* RX_ENABLE_LOCKS */
3159 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3160 rxi_ClearTransmitQueue(call, 0);
3161 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3164 /* Should not reach here, unless the peer is broken: send an abort
3166 rxi_CallError(call, RX_PROTOCOL_ERROR);
3167 np = rxi_SendCallAbort(call, np, 1, 0);
3170 /* Note when this last legitimate packet was received, for keep-alive
3171 * processing. Note, we delay getting the time until now in the hope that
3172 * the packet will be delivered to the user before any get time is required
3173 * (if not, then the time won't actually be re-evaluated here). */
3174 call->lastReceiveTime = clock_Sec();
3175 MUTEX_EXIT(&call->lock);
3176 MUTEX_ENTER(&conn->conn_data_lock);
3178 MUTEX_EXIT(&conn->conn_data_lock);
3182 /* return true if this is an "interesting" connection from the point of view
3183 of someone trying to debug the system */
3185 rxi_IsConnInteresting(struct rx_connection *aconn)
3188 struct rx_call *tcall;
3190 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3193 for (i = 0; i < RX_MAXCALLS; i++) {
3194 tcall = aconn->call[i];
3196 if ((tcall->state == RX_STATE_PRECALL)
3197 || (tcall->state == RX_STATE_ACTIVE))
3199 if ((tcall->mode == RX_MODE_SENDING)
3200 || (tcall->mode == RX_MODE_RECEIVING))
3208 /* if this is one of the last few packets AND it wouldn't be used by the
3209 receiving call to immediately satisfy a read request, then drop it on
3210 the floor, since accepting it might prevent a lock-holding thread from
3211 making progress in its reading. If a call has been cleared while in
3212 the precall state then ignore all subsequent packets until the call
3213 is assigned to a thread. */
3216 TooLow(struct rx_packet *ap, struct rx_call *acall)
3220 MUTEX_ENTER(&rx_quota_mutex);
3221 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3222 && (acall->state == RX_STATE_PRECALL))
3223 || ((rx_nFreePackets < rxi_dataQuota + 2)
3224 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3225 && (acall->flags & RX_CALL_READER_WAIT)))) {
3228 MUTEX_EXIT(&rx_quota_mutex);
3234 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3236 struct rx_connection *conn = arg1;
3237 struct rx_call *acall = arg2;
3238 struct rx_call *call = acall;
3239 struct clock when, now;
3242 MUTEX_ENTER(&conn->conn_data_lock);
3243 conn->checkReachEvent = NULL;
3244 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3247 MUTEX_EXIT(&conn->conn_data_lock);
3251 MUTEX_ENTER(&conn->conn_call_lock);
3252 MUTEX_ENTER(&conn->conn_data_lock);
3253 for (i = 0; i < RX_MAXCALLS; i++) {
3254 struct rx_call *tc = conn->call[i];
3255 if (tc && tc->state == RX_STATE_PRECALL) {
3261 /* Indicate that rxi_CheckReachEvent is no longer running by
3262 * clearing the flag. Must be atomic under conn_data_lock to
3263 * avoid a new call slipping by: rxi_CheckConnReach holds
3264 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3266 conn->flags &= ~RX_CONN_ATTACHWAIT;
3267 MUTEX_EXIT(&conn->conn_data_lock);
3268 MUTEX_EXIT(&conn->conn_call_lock);
3273 MUTEX_ENTER(&call->lock);
3274 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3276 MUTEX_EXIT(&call->lock);
3278 clock_GetTime(&now);
3280 when.sec += RX_CHECKREACH_TIMEOUT;
3281 MUTEX_ENTER(&conn->conn_data_lock);
3282 if (!conn->checkReachEvent) {
3284 conn->checkReachEvent =
3285 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3288 MUTEX_EXIT(&conn->conn_data_lock);
3294 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3296 struct rx_service *service = conn->service;
3297 struct rx_peer *peer = conn->peer;
3298 afs_uint32 now, lastReach;
3300 if (service->checkReach == 0)
3304 MUTEX_ENTER(&peer->peer_lock);
3305 lastReach = peer->lastReachTime;
3306 MUTEX_EXIT(&peer->peer_lock);
3307 if (now - lastReach < RX_CHECKREACH_TTL)
3310 MUTEX_ENTER(&conn->conn_data_lock);
3311 if (conn->flags & RX_CONN_ATTACHWAIT) {
3312 MUTEX_EXIT(&conn->conn_data_lock);
3315 conn->flags |= RX_CONN_ATTACHWAIT;
3316 MUTEX_EXIT(&conn->conn_data_lock);
3317 if (!conn->checkReachEvent)
3318 rxi_CheckReachEvent(NULL, conn, call);
3323 /* try to attach call, if authentication is complete */
3325 TryAttach(struct rx_call *acall, osi_socket socket,
3326 int *tnop, struct rx_call **newcallp,
3329 struct rx_connection *conn = acall->conn;
3331 if (conn->type == RX_SERVER_CONNECTION
3332 && acall->state == RX_STATE_PRECALL) {
3333 /* Don't attach until we have any req'd. authentication. */
3334 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3335 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3336 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3337 /* Note: this does not necessarily succeed; there
3338 * may not any proc available
3341 rxi_ChallengeOn(acall->conn);
3346 /* A data packet has been received off the interface. This packet is
3347 * appropriate to the call (the call is in the right state, etc.). This
3348 * routine can return a packet to the caller, for re-use */
3351 rxi_ReceiveDataPacket(struct rx_call *call,
3352 struct rx_packet *np, int istack,
3353 osi_socket socket, afs_uint32 host, u_short port,
3354 int *tnop, struct rx_call **newcallp)
3356 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3361 afs_uint32 serial=0, flags=0;
3363 struct rx_packet *tnp;
3364 struct clock when, now;
3365 if (rx_stats_active)
3366 rx_MutexIncrement(rx_stats.dataPacketsRead, rx_stats_mutex);
3369 /* If there are no packet buffers, drop this new packet, unless we can find
3370 * packet buffers from inactive calls */
3372 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3373 MUTEX_ENTER(&rx_freePktQ_lock);
3374 rxi_NeedMorePackets = TRUE;
3375 MUTEX_EXIT(&rx_freePktQ_lock);
3376 if (rx_stats_active)
3377 rx_MutexIncrement(rx_stats.noPacketBuffersOnRead, rx_stats_mutex);
3378 call->rprev = np->header.serial;
3379 rxi_calltrace(RX_TRACE_DROP, call);
3380 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems", np));
3382 rxi_ClearReceiveQueue(call);
3383 clock_GetTime(&now);
3385 clock_Add(&when, &rx_softAckDelay);
3386 if (!call->delayedAckEvent
3387 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3388 rxevent_Cancel(call->delayedAckEvent, call,
3389 RX_CALL_REFCOUNT_DELAY);
3390 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3391 call->delayedAckEvent =
3392 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3394 /* we've damaged this call already, might as well do it in. */
3400 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3401 * packet is one of several packets transmitted as a single
3402 * datagram. Do not send any soft or hard acks until all packets
3403 * in a jumbogram have been processed. Send negative acks right away.
3405 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3406 /* tnp is non-null when there are more packets in the
3407 * current jumbo gram */
3414 seq = np->header.seq;
3415 serial = np->header.serial;
3416 flags = np->header.flags;
3418 /* If the call is in an error state, send an abort message */
3420 return rxi_SendCallAbort(call, np, istack, 0);
3422 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3423 * AFS 3.5 jumbogram. */
3424 if (flags & RX_JUMBO_PACKET) {
3425 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3430 if (np->header.spare != 0) {
3431 MUTEX_ENTER(&call->conn->conn_data_lock);
3432 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3433 MUTEX_EXIT(&call->conn->conn_data_lock);
3436 /* The usual case is that this is the expected next packet */
3437 if (seq == call->rnext) {
3439 /* Check to make sure it is not a duplicate of one already queued */
3440 if (queue_IsNotEmpty(&call->rq)
3441 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3442 if (rx_stats_active)
3443 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3444 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate", np));
3445 rxevent_Cancel(call->delayedAckEvent, call,
3446 RX_CALL_REFCOUNT_DELAY);
3447 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3453 /* It's the next packet. Stick it on the receive queue
3454 * for this call. Set newPackets to make sure we wake
3455 * the reader once all packets have been processed */
3456 np->flags |= RX_PKTFLAG_RQ;
3457 queue_Prepend(&call->rq, np);
3458 #ifdef RXDEBUG_PACKET
3460 #endif /* RXDEBUG_PACKET */
3462 np = NULL; /* We can't use this anymore */
3465 /* If an ack is requested then set a flag to make sure we
3466 * send an acknowledgement for this packet */
3467 if (flags & RX_REQUEST_ACK) {
3468 ackNeeded = RX_ACK_REQUESTED;
3471 /* Keep track of whether we have received the last packet */
3472 if (flags & RX_LAST_PACKET) {
3473 call->flags |= RX_CALL_HAVE_LAST;
3477 /* Check whether we have all of the packets for this call */
3478 if (call->flags & RX_CALL_HAVE_LAST) {
3479 afs_uint32 tseq; /* temporary sequence number */
3480 struct rx_packet *tp; /* Temporary packet pointer */
3481 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3483 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3484 if (tseq != tp->header.seq)
3486 if (tp->header.flags & RX_LAST_PACKET) {
3487 call->flags |= RX_CALL_RECEIVE_DONE;
3494 /* Provide asynchronous notification for those who want it
3495 * (e.g. multi rx) */
3496 if (call->arrivalProc) {
3497 (*call->arrivalProc) (call, call->arrivalProcHandle,
3498 call->arrivalProcArg);
3499 call->arrivalProc = (void (*)())0;
3502 /* Update last packet received */
3505 /* If there is no server process serving this call, grab
3506 * one, if available. We only need to do this once. If a
3507 * server thread is available, this thread becomes a server
3508 * thread and the server thread becomes a listener thread. */
3510 TryAttach(call, socket, tnop, newcallp, 0);
3513 /* This is not the expected next packet. */
3515 /* Determine whether this is a new or old packet, and if it's
3516 * a new one, whether it fits into the current receive window.
3517 * Also figure out whether the packet was delivered in sequence.
3518 * We use the prev variable to determine whether the new packet
3519 * is the successor of its immediate predecessor in the
3520 * receive queue, and the missing flag to determine whether
3521 * any of this packets predecessors are missing. */
3523 afs_uint32 prev; /* "Previous packet" sequence number */
3524 struct rx_packet *tp; /* Temporary packet pointer */
3525 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3526 int missing; /* Are any predecessors missing? */
3528 /* If the new packet's sequence number has been sent to the
3529 * application already, then this is a duplicate */
3530 if (seq < call->rnext) {
3531 if (rx_stats_active)
3532 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3533 rxevent_Cancel(call->delayedAckEvent, call,
3534 RX_CALL_REFCOUNT_DELAY);
3535 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3541 /* If the sequence number is greater than what can be
3542 * accomodated by the current window, then send a negative
3543 * acknowledge and drop the packet */
3544 if ((call->rnext + call->rwind) <= seq) {
3545 rxevent_Cancel(call->delayedAckEvent, call,
3546 RX_CALL_REFCOUNT_DELAY);
3547 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3554 /* Look for the packet in the queue of old received packets */
3555 for (prev = call->rnext - 1, missing =
3556 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3557 /*Check for duplicate packet */
3558 if (seq == tp->header.seq) {
3559 if (rx_stats_active)
3560 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3561 rxevent_Cancel(call->delayedAckEvent, call,
3562 RX_CALL_REFCOUNT_DELAY);
3563 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3569 /* If we find a higher sequence packet, break out and
3570 * insert the new packet here. */
3571 if (seq < tp->header.seq)
3573 /* Check for missing packet */
3574 if (tp->header.seq != prev + 1) {
3578 prev = tp->header.seq;
3581 /* Keep track of whether we have received the last packet. */
3582 if (flags & RX_LAST_PACKET) {
3583 call->flags |= RX_CALL_HAVE_LAST;
3586 /* It's within the window: add it to the the receive queue.
3587 * tp is left by the previous loop either pointing at the
3588 * packet before which to insert the new packet, or at the
3589 * queue head if the queue is empty or the packet should be
3591 np->flags |= RX_PKTFLAG_RQ;
3592 #ifdef RXDEBUG_PACKET
3594 #endif /* RXDEBUG_PACKET */
3595 queue_InsertBefore(tp, np);
3599 /* Check whether we have all of the packets for this call */
3600 if ((call->flags & RX_CALL_HAVE_LAST)
3601 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3602 afs_uint32 tseq; /* temporary sequence number */
3605 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3606 if (tseq != tp->header.seq)
3608 if (tp->header.flags & RX_LAST_PACKET) {
3609 call->flags |= RX_CALL_RECEIVE_DONE;
3616 /* We need to send an ack of the packet is out of sequence,
3617 * or if an ack was requested by the peer. */
3618 if (seq != prev + 1 || missing) {
3619 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3620 } else if (flags & RX_REQUEST_ACK) {
3621 ackNeeded = RX_ACK_REQUESTED;
3624 /* Acknowledge the last packet for each call */
3625 if (flags & RX_LAST_PACKET) {
3636 * If the receiver is waiting for an iovec, fill the iovec
3637 * using the data from the receive queue */
3638 if (call->flags & RX_CALL_IOVEC_WAIT) {
3639 didHardAck = rxi_FillReadVec(call, serial);
3640 /* the call may have been aborted */
3649 /* Wakeup the reader if any */
3650 if ((call->flags & RX_CALL_READER_WAIT)
3651 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3652 || (call->iovNext >= call->iovMax)
3653 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3654 call->flags &= ~RX_CALL_READER_WAIT;
3655 #ifdef RX_ENABLE_LOCKS
3656 CV_BROADCAST(&call->cv_rq);
3658 osi_rxWakeup(&call->rq);
3664 * Send an ack when requested by the peer, or once every
3665 * rxi_SoftAckRate packets until the last packet has been
3666 * received. Always send a soft ack for the last packet in
3667 * the server's reply. */
3669 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3670 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3671 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3672 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3673 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3674 } else if (call->nSoftAcks) {
3675 clock_GetTime(&now);
3677 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3678 clock_Add(&when, &rx_lastAckDelay);
3680 clock_Add(&when, &rx_softAckDelay);
3682 if (!call->delayedAckEvent
3683 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3684 rxevent_Cancel(call->delayedAckEvent, call,
3685 RX_CALL_REFCOUNT_DELAY);
3686 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3687 call->delayedAckEvent =
3688 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3690 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3691 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3698 static void rxi_ComputeRate();
3702 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3704 struct rx_peer *peer = conn->peer;
3706 MUTEX_ENTER(&peer->peer_lock);
3707 peer->lastReachTime = clock_Sec();
3708 MUTEX_EXIT(&peer->peer_lock);
3710 MUTEX_ENTER(&conn->conn_data_lock);
3711 if (conn->flags & RX_CONN_ATTACHWAIT) {
3714 conn->flags &= ~RX_CONN_ATTACHWAIT;
3715 MUTEX_EXIT(&conn->conn_data_lock);
3717 for (i = 0; i < RX_MAXCALLS; i++) {
3718 struct rx_call *call = conn->call[i];
3721 MUTEX_ENTER(&call->lock);
3722 /* tnop can be null if newcallp is null */
3723 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3725 MUTEX_EXIT(&call->lock);
3729 MUTEX_EXIT(&conn->conn_data_lock);
3732 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
3734 rx_ack_reason(int reason)
3737 case RX_ACK_REQUESTED:
3739 case RX_ACK_DUPLICATE:
3741 case RX_ACK_OUT_OF_SEQUENCE:
3743 case RX_ACK_EXCEEDS_WINDOW:
3745 case RX_ACK_NOSPACE:
3749 case RX_ACK_PING_RESPONSE:
3762 /* rxi_ComputePeerNetStats
3764 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3765 * estimates (like RTT and throughput) based on ack packets. Caller
3766 * must ensure that the packet in question is the right one (i.e.
3767 * serial number matches).
3770 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3771 struct rx_ackPacket *ap, struct rx_packet *np)
3773 struct rx_peer *peer = call->conn->peer;
3775 /* Use RTT if not delayed by client and
3776 * ignore packets that were retransmitted. */
3777 if (!(p->flags & RX_PKTFLAG_ACKED) &&
3778 ap->reason != RX_ACK_DELAY &&
3779 clock_Eq(&p->timeSent, &p->firstSent))
3780 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3782 rxi_ComputeRate(peer, call, p, np, ap->reason);
3786 /* The real smarts of the whole thing. */
3788 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
3791 struct rx_ackPacket *ap;
3793 struct rx_packet *tp;
3794 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3795 struct rx_connection *conn = call->conn;
3796 struct rx_peer *peer = conn->peer;
3799 /* because there are CM's that are bogus, sending weird values for this. */
3800 afs_uint32 skew = 0;
3806 int newAckCount = 0;
3807 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3808 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3810 if (rx_stats_active)
3811 rx_MutexIncrement(rx_stats.ackPacketsRead, rx_stats_mutex);
3812 ap = (struct rx_ackPacket *)rx_DataOf(np);
3813 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3815 return np; /* truncated ack packet */
3817 /* depends on ack packet struct */
3818 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3819 first = ntohl(ap->firstPacket);
3820 serial = ntohl(ap->serial);
3821 /* temporarily disabled -- needs to degrade over time
3822 * skew = ntohs(ap->maxSkew); */
3824 /* Ignore ack packets received out of order */
3825 if (first < call->tfirst) {
3829 if (np->header.flags & RX_SLOW_START_OK) {
3830 call->flags |= RX_CALL_SLOW_START_OK;
3833 if (ap->reason == RX_ACK_PING_RESPONSE)
3834 rxi_UpdatePeerReach(conn, call);
3838 if (rxdebug_active) {
3842 len = _snprintf(msg, sizeof(msg),
3843 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3844 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3845 ntohl(ap->serial), ntohl(ap->previousPacket),
3846 (unsigned int)np->header.seq, (unsigned int)skew,
3847 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3851 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3852 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3856 OutputDebugString(msg);
3858 #else /* AFS_NT40_ENV */
3861 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3862 ap->reason, ntohl(ap->previousPacket),
3863 (unsigned int)np->header.seq, (unsigned int)serial,
3864 (unsigned int)skew, ntohl(ap->firstPacket));
3867 for (offset = 0; offset < nAcks; offset++)
3868 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3873 #endif /* AFS_NT40_ENV */
3876 /* Update the outgoing packet skew value to the latest value of
3877 * the peer's incoming packet skew value. The ack packet, of
3878 * course, could arrive out of order, but that won't affect things
3880 MUTEX_ENTER(&peer->peer_lock);
3881 peer->outPacketSkew = skew;
3883 /* Check for packets that no longer need to be transmitted, and
3884 * discard them. This only applies to packets positively
3885 * acknowledged as having been sent to the peer's upper level.
3886 * All other packets must be retained. So only packets with
3887 * sequence numbers < ap->firstPacket are candidates. */
3888 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3889 if (tp->header.seq >= first)
3891 call->tfirst = tp->header.seq + 1;
3892 rxi_ComputePeerNetStats(call, tp, ap, np);
3893 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3896 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3897 /* XXX Hack. Because we have to release the global rx lock when sending
3898 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3899 * in rxi_Start sending packets out because packets may move to the
3900 * freePacketQueue as result of being here! So we drop these packets until
3901 * we're safely out of the traversing. Really ugly!
3902 * To make it even uglier, if we're using fine grain locking, we can
3903 * set the ack bits in the packets and have rxi_Start remove the packets
3904 * when it's done transmitting.
3906 if (call->flags & RX_CALL_TQ_BUSY) {
3907 #ifdef RX_ENABLE_LOCKS
3908 tp->flags |= RX_PKTFLAG_ACKED;
3909 call->flags |= RX_CALL_TQ_SOME_ACKED;
3910 #else /* RX_ENABLE_LOCKS */
3912 #endif /* RX_ENABLE_LOCKS */
3914 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3917 tp->flags &= ~RX_PKTFLAG_TQ;
3918 #ifdef RXDEBUG_PACKET
3920 #endif /* RXDEBUG_PACKET */
3921 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3926 /* Give rate detector a chance to respond to ping requests */
3927 if (ap->reason == RX_ACK_PING_RESPONSE) {
3928 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3932 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3934 /* Now go through explicit acks/nacks and record the results in
3935 * the waiting packets. These are packets that can't be released
3936 * yet, even with a positive acknowledge. This positive
3937 * acknowledge only means the packet has been received by the
3938 * peer, not that it will be retained long enough to be sent to
3939 * the peer's upper level. In addition, reset the transmit timers
3940 * of any missing packets (those packets that must be missing
3941 * because this packet was out of sequence) */
3943 call->nSoftAcked = 0;
3944 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3945 /* Update round trip time if the ack was stimulated on receipt
3947 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3948 #ifdef RX_ENABLE_LOCKS
3949 if (tp->header.seq >= first)
3950 #endif /* RX_ENABLE_LOCKS */
3951 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3952 rxi_ComputePeerNetStats(call, tp, ap, np);
3954 /* Set the acknowledge flag per packet based on the
3955 * information in the ack packet. An acknowlegded packet can
3956 * be downgraded when the server has discarded a packet it
3957 * soacked previously, or when an ack packet is received
3958 * out of sequence. */
3959 if (tp->header.seq < first) {
3960 /* Implicit ack information */
3961 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3964 tp->flags |= RX_PKTFLAG_ACKED;
3965 } else if (tp->header.seq < first + nAcks) {
3966 /* Explicit ack information: set it in the packet appropriately */
3967 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3968 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3970 tp->flags |= RX_PKTFLAG_ACKED;
3977 } else /* RX_ACK_TYPE_NACK */ {
3978 tp->flags &= ~RX_PKTFLAG_ACKED;
3982 tp->flags &= ~RX_PKTFLAG_ACKED;
3987 * Following the suggestion of Phil Kern, we back off the peer's
3988 * timeout value for future packets until a successful response
3989 * is received for an initial transmission.
3991 if (missing && !backedOff) {
3992 struct clock c = peer->timeout;
3993 struct clock max_to = {3, 0};
3995 clock_Add(&peer->timeout, &c);
3996 if (clock_Gt(&peer->timeout, &max_to))
3997 peer->timeout = max_to;
4001 /* If packet isn't yet acked, and it has been transmitted at least
4002 * once, reset retransmit time using latest timeout
4003 * ie, this should readjust the retransmit timer for all outstanding
4004 * packets... So we don't just retransmit when we should know better*/
4006 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
4007 tp->retryTime = tp->timeSent;
4008 clock_Add(&tp->retryTime, &peer->timeout);
4009 /* shift by eight because one quarter-sec ~ 256 milliseconds */
4010 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
4014 /* If the window has been extended by this acknowledge packet,
4015 * then wakeup a sender waiting in alloc for window space, or try
4016 * sending packets now, if he's been sitting on packets due to
4017 * lack of window space */
4018 if (call->tnext < (call->tfirst + call->twind)) {
4019 #ifdef RX_ENABLE_LOCKS
4020 CV_SIGNAL(&call->cv_twind);
4022 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4023 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4024 osi_rxWakeup(&call->twind);
4027 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4028 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4032 /* if the ack packet has a receivelen field hanging off it,
4033 * update our state */
4034 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4037 /* If the ack packet has a "recommended" size that is less than
4038 * what I am using now, reduce my size to match */
4039 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4040 (int)sizeof(afs_int32), &tSize);
4041 tSize = (afs_uint32) ntohl(tSize);
4042 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4044 /* Get the maximum packet size to send to this peer */
4045 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4047 tSize = (afs_uint32) ntohl(tSize);
4048 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4049 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4051 /* sanity check - peer might have restarted with different params.
4052 * If peer says "send less", dammit, send less... Peer should never
4053 * be unable to accept packets of the size that prior AFS versions would
4054 * send without asking. */
4055 if (peer->maxMTU != tSize) {
4056 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4058 peer->maxMTU = tSize;
4059 peer->MTU = MIN(tSize, peer->MTU);
4060 call->MTU = MIN(call->MTU, tSize);
4063 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4066 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4067 (int)sizeof(afs_int32), &tSize);
4068 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4069 if (tSize < call->twind) { /* smaller than our send */
4070 call->twind = tSize; /* window, we must send less... */
4071 call->ssthresh = MIN(call->twind, call->ssthresh);
4072 call->conn->twind[call->channel] = call->twind;
4075 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4076 * network MTU confused with the loopback MTU. Calculate the
4077 * maximum MTU here for use in the slow start code below.
4079 maxMTU = peer->maxMTU;
4080 /* Did peer restart with older RX version? */
4081 if (peer->maxDgramPackets > 1) {
4082 peer->maxDgramPackets = 1;
4084 } else if (np->length >=
4085 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4088 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4089 sizeof(afs_int32), &tSize);
4090 tSize = (afs_uint32) ntohl(tSize);
4092 * As of AFS 3.5 we set the send window to match the receive window.
4094 if (tSize < call->twind) {
4095 call->twind = tSize;
4096 call->conn->twind[call->channel] = call->twind;
4097 call->ssthresh = MIN(call->twind, call->ssthresh);
4098 } else if (tSize > call->twind) {
4099 call->twind = tSize;
4100 call->conn->twind[call->channel] = call->twind;
4104 * As of AFS 3.5, a jumbogram is more than one fixed size
4105 * packet transmitted in a single UDP datagram. If the remote
4106 * MTU is smaller than our local MTU then never send a datagram
4107 * larger than the natural MTU.
4110 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4111 (int)sizeof(afs_int32), &tSize);
4112 maxDgramPackets = (afs_uint32) ntohl(tSize);
4113 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4115 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4116 maxDgramPackets = MIN(maxDgramPackets, tSize);
4117 if (maxDgramPackets > 1) {
4118 peer->maxDgramPackets = maxDgramPackets;
4119 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4121 peer->maxDgramPackets = 1;
4122 call->MTU = peer->natMTU;
4124 } else if (peer->maxDgramPackets > 1) {
4125 /* Restarted with lower version of RX */
4126 peer->maxDgramPackets = 1;
4128 } else if (peer->maxDgramPackets > 1
4129 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4130 /* Restarted with lower version of RX */
4131 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4132 peer->natMTU = OLD_MAX_PACKET_SIZE;
4133 peer->MTU = OLD_MAX_PACKET_SIZE;
4134 peer->maxDgramPackets = 1;
4135 peer->nDgramPackets = 1;
4137 call->MTU = OLD_MAX_PACKET_SIZE;
4142 * Calculate how many datagrams were successfully received after
4143 * the first missing packet and adjust the negative ack counter
4148 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4149 if (call->nNacks < nNacked) {
4150 call->nNacks = nNacked;
4153 call->nAcks += newAckCount;
4157 if (call->flags & RX_CALL_FAST_RECOVER) {
4159 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4161 call->flags &= ~RX_CALL_FAST_RECOVER;
4162 call->cwind = call->nextCwind;
4163 call->nextCwind = 0;
4166 call->nCwindAcks = 0;
4167 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4168 /* Three negative acks in a row trigger congestion recovery */
4169 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4170 MUTEX_EXIT(&peer->peer_lock);
4171 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4172 /* someone else is waiting to start recovery */
4175 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4176 rxi_WaitforTQBusy(call);
4177 MUTEX_ENTER(&peer->peer_lock);
4178 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4179 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4180 call->flags |= RX_CALL_FAST_RECOVER;
4181 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4183 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4184 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4185 call->nextCwind = call->ssthresh;
4188 peer->MTU = call->MTU;
4189 peer->cwind = call->nextCwind;
4190 peer->nDgramPackets = call->nDgramPackets;
4192 call->congestSeq = peer->congestSeq;
4193 /* Reset the resend times on the packets that were nacked
4194 * so we will retransmit as soon as the window permits*/
4195 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4197 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4198 clock_Zero(&tp->retryTime);
4200 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4205 /* If cwind is smaller than ssthresh, then increase
4206 * the window one packet for each ack we receive (exponential
4208 * If cwind is greater than or equal to ssthresh then increase
4209 * the congestion window by one packet for each cwind acks we
4210 * receive (linear growth). */
4211 if (call->cwind < call->ssthresh) {
4213 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4214 call->nCwindAcks = 0;
4216 call->nCwindAcks += newAckCount;
4217 if (call->nCwindAcks >= call->cwind) {
4218 call->nCwindAcks = 0;
4219 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4223 * If we have received several acknowledgements in a row then
4224 * it is time to increase the size of our datagrams
4226 if ((int)call->nAcks > rx_nDgramThreshold) {
4227 if (peer->maxDgramPackets > 1) {
4228 if (call->nDgramPackets < peer->maxDgramPackets) {
4229 call->nDgramPackets++;
4231 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4232 } else if (call->MTU < peer->maxMTU) {
4233 call->MTU += peer->natMTU;
4234 call->MTU = MIN(call->MTU, peer->maxMTU);
4240 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4242 /* Servers need to hold the call until all response packets have
4243 * been acknowledged. Soft acks are good enough since clients
4244 * are not allowed to clear their receive queues. */
4245 if (call->state == RX_STATE_HOLD
4246 && call->tfirst + call->nSoftAcked >= call->tnext) {
4247 call->state = RX_STATE_DALLY;
4248 rxi_ClearTransmitQueue(call, 0);
4249 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4250 } else if (!queue_IsEmpty(&call->tq)) {
4251 rxi_Start(0, call, 0, istack);
4256 /* Received a response to a challenge packet */
4258 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4259 struct rx_packet *np, int istack)
4263 /* Ignore the packet if we're the client */
4264 if (conn->type == RX_CLIENT_CONNECTION)
4267 /* If already authenticated, ignore the packet (it's probably a retry) */
4268 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4271 /* Otherwise, have the security object evaluate the response packet */
4272 error = RXS_CheckResponse(conn->securityObject, conn, np);
4274 /* If the response is invalid, reset the connection, sending
4275 * an abort to the peer */
4279 rxi_ConnectionError(conn, error);
4280 MUTEX_ENTER(&conn->conn_data_lock);
4281 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4282 MUTEX_EXIT(&conn->conn_data_lock);
4285 /* If the response is valid, any calls waiting to attach
4286 * servers can now do so */
4289 for (i = 0; i < RX_MAXCALLS; i++) {
4290 struct rx_call *call = conn->call[i];
4292 MUTEX_ENTER(&call->lock);
4293 if (call->state == RX_STATE_PRECALL)
4294 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4295 /* tnop can be null if newcallp is null */
4296 MUTEX_EXIT(&call->lock);
4300 /* Update the peer reachability information, just in case
4301 * some calls went into attach-wait while we were waiting
4302 * for authentication..
4304 rxi_UpdatePeerReach(conn, NULL);
4309 /* A client has received an authentication challenge: the security
4310 * object is asked to cough up a respectable response packet to send
4311 * back to the server. The server is responsible for retrying the
4312 * challenge if it fails to get a response. */
4315 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4316 struct rx_packet *np, int istack)
4320 /* Ignore the challenge if we're the server */
4321 if (conn->type == RX_SERVER_CONNECTION)
4324 /* Ignore the challenge if the connection is otherwise idle; someone's
4325 * trying to use us as an oracle. */
4326 if (!rxi_HasActiveCalls(conn))
4329 /* Send the security object the challenge packet. It is expected to fill
4330 * in the response. */
4331 error = RXS_GetResponse(conn->securityObject, conn, np);
4333 /* If the security object is unable to return a valid response, reset the
4334 * connection and send an abort to the peer. Otherwise send the response
4335 * packet to the peer connection. */
4337 rxi_ConnectionError(conn, error);
4338 MUTEX_ENTER(&conn->conn_data_lock);
4339 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4340 MUTEX_EXIT(&conn->conn_data_lock);
4342 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4343 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4349 /* Find an available server process to service the current request in
4350 * the given call structure. If one isn't available, queue up this
4351 * call so it eventually gets one */
4353 rxi_AttachServerProc(struct rx_call *call,
4354 osi_socket socket, int *tnop,
4355 struct rx_call **newcallp)
4357 struct rx_serverQueueEntry *sq;
4358 struct rx_service *service = call->conn->service;
4361 /* May already be attached */
4362 if (call->state == RX_STATE_ACTIVE)
4365 MUTEX_ENTER(&rx_serverPool_lock);
4367 haveQuota = QuotaOK(service);
4368 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4369 /* If there are no processes available to service this call,
4370 * put the call on the incoming call queue (unless it's
4371 * already on the queue).
4373 #ifdef RX_ENABLE_LOCKS
4375 ReturnToServerPool(service);
4376 #endif /* RX_ENABLE_LOCKS */
4378 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4379 call->flags |= RX_CALL_WAIT_PROC;
4380 MUTEX_ENTER(&rx_waiting_mutex);
4383 MUTEX_EXIT(&rx_waiting_mutex);
4384 rxi_calltrace(RX_CALL_ARRIVAL, call);
4385 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4386 queue_Append(&rx_incomingCallQueue, call);
4389 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4391 /* If hot threads are enabled, and both newcallp and sq->socketp
4392 * are non-null, then this thread will process the call, and the
4393 * idle server thread will start listening on this threads socket.
4396 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4399 *sq->socketp = socket;
4400 clock_GetTime(&call->startTime);
4401 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4405 if (call->flags & RX_CALL_WAIT_PROC) {
4406 /* Conservative: I don't think this should happen */
4407 call->flags &= ~RX_CALL_WAIT_PROC;
4408 if (queue_IsOnQueue(call)) {
4411 MUTEX_ENTER(&rx_waiting_mutex);
4413 MUTEX_EXIT(&rx_waiting_mutex);
4416 call->state = RX_STATE_ACTIVE;
4417 call->mode = RX_MODE_RECEIVING;
4418 #ifdef RX_KERNEL_TRACE
4420 int glockOwner = ISAFS_GLOCK();
4423 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4424 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4430 if (call->flags & RX_CALL_CLEARED) {
4431 /* send an ack now to start the packet flow up again */
4432 call->flags &= ~RX_CALL_CLEARED;
4433 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4435 #ifdef RX_ENABLE_LOCKS
4438 service->nRequestsRunning++;
4439 MUTEX_ENTER(&rx_quota_mutex);
4440 if (service->nRequestsRunning <= service->minProcs)
4443 MUTEX_EXIT(&rx_quota_mutex);
4447 MUTEX_EXIT(&rx_serverPool_lock);
4450 /* Delay the sending of an acknowledge event for a short while, while
4451 * a new call is being prepared (in the case of a client) or a reply
4452 * is being prepared (in the case of a server). Rather than sending
4453 * an ack packet, an ACKALL packet is sent. */
4455 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4457 #ifdef RX_ENABLE_LOCKS
4459 MUTEX_ENTER(&call->lock);
4460 call->delayedAckEvent = NULL;
4461 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4463 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4464 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4466 MUTEX_EXIT(&call->lock);
4467 #else /* RX_ENABLE_LOCKS */
4469 call->delayedAckEvent = NULL;
4470 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4471 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4472 #endif /* RX_ENABLE_LOCKS */
4476 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4478 struct rx_call *call = arg1;
4479 #ifdef RX_ENABLE_LOCKS
4481 MUTEX_ENTER(&call->lock);
4482 if (event == call->delayedAckEvent)
4483 call->delayedAckEvent = NULL;
4484 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4486 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4488 MUTEX_EXIT(&call->lock);
4489 #else /* RX_ENABLE_LOCKS */
4491 call->delayedAckEvent = NULL;
4492 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4493 #endif /* RX_ENABLE_LOCKS */
4497 #ifdef RX_ENABLE_LOCKS
4498 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4499 * clearing them out.
4502 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4504 struct rx_packet *p, *tp;
4507 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4508 p->flags |= RX_PKTFLAG_ACKED;
4512 call->flags |= RX_CALL_TQ_CLEARME;
4513 call->flags |= RX_CALL_TQ_SOME_ACKED;
4516 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4517 call->tfirst = call->tnext;
4518 call->nSoftAcked = 0;
4520 if (call->flags & RX_CALL_FAST_RECOVER) {
4521 call->flags &= ~RX_CALL_FAST_RECOVER;
4522 call->cwind = call->nextCwind;
4523 call->nextCwind = 0;
4526 CV_SIGNAL(&call->cv_twind);
4528 #endif /* RX_ENABLE_LOCKS */
4530 /* Clear out the transmit queue for the current call (all packets have
4531 * been received by peer) */
4533 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4535 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4536 struct rx_packet *p, *tp;
4538 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4540 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4541 p->flags |= RX_PKTFLAG_ACKED;
4545 call->flags |= RX_CALL_TQ_CLEARME;
4546 call->flags |= RX_CALL_TQ_SOME_ACKED;
4549 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4550 #ifdef RXDEBUG_PACKET
4552 #endif /* RXDEBUG_PACKET */
4553 rxi_FreePackets(0, &call->tq);
4554 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
4555 #ifdef RX_ENABLE_LOCKS
4556 CV_BROADCAST(&call->cv_tq);
4557 #else /* RX_ENABLE_LOCKS */
4558 osi_rxWakeup(&call->tq);
4559 #endif /* RX_ENABLE_LOCKS */
4561 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4562 call->flags &= ~RX_CALL_TQ_CLEARME;
4564 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4566 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4567 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4568 call->nSoftAcked = 0;
4570 if (call->flags & RX_CALL_FAST_RECOVER) {
4571 call->flags &= ~RX_CALL_FAST_RECOVER;
4572 call->cwind = call->nextCwind;
4574 #ifdef RX_ENABLE_LOCKS
4575 CV_SIGNAL(&call->cv_twind);
4577 osi_rxWakeup(&call->twind);
4582 rxi_ClearReceiveQueue(struct rx_call *call)
4584 if (queue_IsNotEmpty(&call->rq)) {
4587 count = rxi_FreePackets(0, &call->rq);
4588 rx_packetReclaims += count;
4589 #ifdef RXDEBUG_PACKET
4591 if ( call->rqc != 0 )
4592 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0", call, call->rqc));
4594 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4596 if (call->state == RX_STATE_PRECALL) {
4597 call->flags |= RX_CALL_CLEARED;
4601 /* Send an abort packet for the specified call */
4603 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
4604 int istack, int force)
4607 struct clock when, now;
4612 /* Clients should never delay abort messages */
4613 if (rx_IsClientConn(call->conn))
4616 if (call->abortCode != call->error) {
4617 call->abortCode = call->error;
4618 call->abortCount = 0;
4621 if (force || rxi_callAbortThreshhold == 0
4622 || call->abortCount < rxi_callAbortThreshhold) {
4623 if (call->delayedAbortEvent) {
4624 rxevent_Cancel(call->delayedAbortEvent, call,
4625 RX_CALL_REFCOUNT_ABORT);
4627 error = htonl(call->error);
4630 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4631 (char *)&error, sizeof(error), istack);
4632 } else if (!call->delayedAbortEvent) {
4633 clock_GetTime(&now);
4635 clock_Addmsec(&when, rxi_callAbortDelay);
4636 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4637 call->delayedAbortEvent =
4638 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4643 /* Send an abort packet for the specified connection. Packet is an
4644 * optional pointer to a packet that can be used to send the abort.
4645 * Once the number of abort messages reaches the threshhold, an
4646 * event is scheduled to send the abort. Setting the force flag
4647 * overrides sending delayed abort messages.
4649 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4650 * to send the abort packet.
4653 rxi_SendConnectionAbort(struct rx_connection *conn,
4654 struct rx_packet *packet, int istack, int force)
4657 struct clock when, now;
4662 /* Clients should never delay abort messages */
4663 if (rx_IsClientConn(conn))
4666 if (force || rxi_connAbortThreshhold == 0
4667 || conn->abortCount < rxi_connAbortThreshhold) {
4668 if (conn->delayedAbortEvent) {
4669 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4671 error = htonl(conn->error);
4673 MUTEX_EXIT(&conn->conn_data_lock);
4675 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4676 RX_PACKET_TYPE_ABORT, (char *)&error,
4677 sizeof(error), istack);
4678 MUTEX_ENTER(&conn->conn_data_lock);
4679 } else if (!conn->delayedAbortEvent) {
4680 clock_GetTime(&now);
4682 clock_Addmsec(&when, rxi_connAbortDelay);
4683 conn->delayedAbortEvent =
4684 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4689 /* Associate an error all of the calls owned by a connection. Called
4690 * with error non-zero. This is only for really fatal things, like
4691 * bad authentication responses. The connection itself is set in
4692 * error at this point, so that future packets received will be
4695 rxi_ConnectionError(struct rx_connection *conn,
4701 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d", conn, error));
4703 MUTEX_ENTER(&conn->conn_data_lock);
4704 if (conn->challengeEvent)
4705 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4706 if (conn->natKeepAliveEvent)
4707 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
4708 if (conn->checkReachEvent) {
4709 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4710 conn->checkReachEvent = 0;
4711 conn->flags &= ~RX_CONN_ATTACHWAIT;
4714 MUTEX_EXIT(&conn->conn_data_lock);
4715 for (i = 0; i < RX_MAXCALLS; i++) {
4716 struct rx_call *call = conn->call[i];
4718 MUTEX_ENTER(&call->lock);
4719 rxi_CallError(call, error);
4720 MUTEX_EXIT(&call->lock);
4723 conn->error = error;
4724 if (rx_stats_active)
4725 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4730 rxi_CallError(struct rx_call *call, afs_int32 error)
4733 osirx_AssertMine(&call->lock, "rxi_CallError");
4735 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d", call, error, call->error));
4737 error = call->error;
4739 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4740 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4741 rxi_ResetCall(call, 0);
4744 rxi_ResetCall(call, 0);
4746 call->error = error;
4747 call->mode = RX_MODE_ERROR;
4750 /* Reset various fields in a call structure, and wakeup waiting
4751 * processes. Some fields aren't changed: state & mode are not
4752 * touched (these must be set by the caller), and bufptr, nLeft, and
4753 * nFree are not reset, since these fields are manipulated by
4754 * unprotected macros, and may only be reset by non-interrupting code.
4757 /* this code requires that call->conn be set properly as a pre-condition. */
4758 #endif /* ADAPT_WINDOW */
4761 rxi_ResetCall(struct rx_call *call, int newcall)
4764 struct rx_peer *peer;
4765 struct rx_packet *packet;
4767 osirx_AssertMine(&call->lock, "rxi_ResetCall");
4769 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
4771 /* Notify anyone who is waiting for asynchronous packet arrival */
4772 if (call->arrivalProc) {
4773 (*call->arrivalProc) (call, call->arrivalProcHandle,
4774 call->arrivalProcArg);
4775 call->arrivalProc = (void (*)())0;
4778 if (call->delayedAbortEvent) {
4779 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4780 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4782 rxi_SendCallAbort(call, packet, 0, 1);
4783 rxi_FreePacket(packet);
4788 * Update the peer with the congestion information in this call
4789 * so other calls on this connection can pick up where this call
4790 * left off. If the congestion sequence numbers don't match then
4791 * another call experienced a retransmission.
4793 peer = call->conn->peer;
4794 MUTEX_ENTER(&peer->peer_lock);
4796 if (call->congestSeq == peer->congestSeq) {
4797 peer->cwind = MAX(peer->cwind, call->cwind);
4798 peer->MTU = MAX(peer->MTU, call->MTU);
4799 peer->nDgramPackets =
4800 MAX(peer->nDgramPackets, call->nDgramPackets);
4803 call->abortCode = 0;
4804 call->abortCount = 0;
4806 if (peer->maxDgramPackets > 1) {
4807 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4809 call->MTU = peer->MTU;
4811 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4812 call->ssthresh = rx_maxSendWindow;
4813 call->nDgramPackets = peer->nDgramPackets;
4814 call->congestSeq = peer->congestSeq;
4815 MUTEX_EXIT(&peer->peer_lock);
4817 flags = call->flags;
4818 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4819 rxi_WaitforTQBusy(call);
4820 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4822 rxi_ClearTransmitQueue(call, 1);
4823 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4824 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4828 rxi_ClearReceiveQueue(call);
4829 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
4831 if (call->currentPacket) {
4832 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
4833 call->currentPacket->flags |= RX_PKTFLAG_IOVQ;
4834 queue_Prepend(&call->iovq, call->currentPacket);
4835 #ifdef RXDEBUG_PACKET
4837 #endif /* RXDEBUG_PACKET */
4838 call->currentPacket = (struct rx_packet *)0;
4840 call->curlen = call->nLeft = call->nFree = 0;
4842 #ifdef RXDEBUG_PACKET
4845 rxi_FreePackets(0, &call->iovq);
4848 call->twind = call->conn->twind[call->channel];
4849 call->rwind = call->conn->rwind[call->channel];
4850 call->nSoftAcked = 0;
4851 call->nextCwind = 0;
4854 call->nCwindAcks = 0;
4855 call->nSoftAcks = 0;
4856 call->nHardAcks = 0;
4858 call->tfirst = call->rnext = call->tnext = 1;
4860 call->lastAcked = 0;
4861 call->localStatus = call->remoteStatus = 0;
4863 if (flags & RX_CALL_READER_WAIT) {
4864 #ifdef RX_ENABLE_LOCKS
4865 CV_BROADCAST(&call->cv_rq);
4867 osi_rxWakeup(&call->rq);
4870 if (flags & RX_CALL_WAIT_PACKETS) {
4871 MUTEX_ENTER(&rx_freePktQ_lock);
4872 rxi_PacketsUnWait(); /* XXX */
4873 MUTEX_EXIT(&rx_freePktQ_lock);
4875 #ifdef RX_ENABLE_LOCKS
4876 CV_SIGNAL(&call->cv_twind);
4878 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4879 osi_rxWakeup(&call->twind);
4882 #ifdef RX_ENABLE_LOCKS
4883 /* The following ensures that we don't mess with any queue while some
4884 * other thread might also be doing so. The call_queue_lock field is
4885 * is only modified under the call lock. If the call is in the process
4886 * of being removed from a queue, the call is not locked until the
4887 * the queue lock is dropped and only then is the call_queue_lock field
4888 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4889 * Note that any other routine which removes a call from a queue has to
4890 * obtain the queue lock before examing the queue and removing the call.
4892 if (call->call_queue_lock) {
4893 MUTEX_ENTER(call->call_queue_lock);
4894 if (queue_IsOnQueue(call)) {
4896 if (flags & RX_CALL_WAIT_PROC) {
4898 MUTEX_ENTER(&rx_waiting_mutex);
4900 MUTEX_EXIT(&rx_waiting_mutex);
4903 MUTEX_EXIT(call->call_queue_lock);
4904 CLEAR_CALL_QUEUE_LOCK(call);
4906 #else /* RX_ENABLE_LOCKS */
4907 if (queue_IsOnQueue(call)) {
4909 if (flags & RX_CALL_WAIT_PROC)
4912 #endif /* RX_ENABLE_LOCKS */
4914 rxi_KeepAliveOff(call);
4915 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4918 /* Send an acknowledge for the indicated packet (seq,serial) of the
4919 * indicated call, for the indicated reason (reason). This
4920 * acknowledge will specifically acknowledge receiving the packet, and
4921 * will also specify which other packets for this call have been
4922 * received. This routine returns the packet that was used to the
4923 * caller. The caller is responsible for freeing it or re-using it.
4924 * This acknowledgement also returns the highest sequence number
4925 * actually read out by the higher level to the sender; the sender
4926 * promises to keep around packets that have not been read by the
4927 * higher level yet (unless, of course, the sender decides to abort
4928 * the call altogether). Any of p, seq, serial, pflags, or reason may
4929 * be set to zero without ill effect. That is, if they are zero, they
4930 * will not convey any information.
4931 * NOW there is a trailer field, after the ack where it will safely be
4932 * ignored by mundanes, which indicates the maximum size packet this
4933 * host can swallow. */
4935 struct rx_packet *optionalPacket; use to send ack (or null)
4936 int seq; Sequence number of the packet we are acking
4937 int serial; Serial number of the packet
4938 int pflags; Flags field from packet header
4939 int reason; Reason an acknowledge was prompted
4943 rxi_SendAck(struct rx_call *call,
4944 struct rx_packet *optionalPacket, int serial, int reason,
4947 struct rx_ackPacket *ap;
4948 struct rx_packet *rqp;
4949 struct rx_packet *nxp; /* For queue_Scan */
4950 struct rx_packet *p;
4953 #ifdef RX_ENABLE_TSFPQ
4954 struct rx_ts_info_t * rx_ts_info;
4958 * Open the receive window once a thread starts reading packets
4960 if (call->rnext > 1) {
4961 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
4964 call->nHardAcks = 0;
4965 call->nSoftAcks = 0;
4966 if (call->rnext > call->lastAcked)
4967 call->lastAcked = call->rnext;
4971 rx_computelen(p, p->length); /* reset length, you never know */
4972 } /* where that's been... */
4973 #ifdef RX_ENABLE_TSFPQ
4975 RX_TS_INFO_GET(rx_ts_info);
4976 if ((p = rx_ts_info->local_special_packet)) {
4977 rx_computelen(p, p->length);
4978 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4979 rx_ts_info->local_special_packet = p;
4980 } else { /* We won't send the ack, but don't panic. */
4981 return optionalPacket;
4985 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4986 /* We won't send the ack, but don't panic. */
4987 return optionalPacket;
4992 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4995 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4996 #ifndef RX_ENABLE_TSFPQ
4997 if (!optionalPacket)
5000 return optionalPacket;
5002 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5003 if (rx_Contiguous(p) < templ) {
5004 #ifndef RX_ENABLE_TSFPQ
5005 if (!optionalPacket)
5008 return optionalPacket;
5013 /* MTUXXX failing to send an ack is very serious. We should */
5014 /* try as hard as possible to send even a partial ack; it's */
5015 /* better than nothing. */
5016 ap = (struct rx_ackPacket *)rx_DataOf(p);
5017 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5018 ap->reason = reason;
5020 /* The skew computation used to be bogus, I think it's better now. */
5021 /* We should start paying attention to skew. XXX */
5022 ap->serial = htonl(serial);
5023 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5025 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
5026 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5028 /* No fear of running out of ack packet here because there can only be at most
5029 * one window full of unacknowledged packets. The window size must be constrained
5030 * to be less than the maximum ack size, of course. Also, an ack should always
5031 * fit into a single packet -- it should not ever be fragmented. */
5032 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5033 if (!rqp || !call->rq.next
5034 || (rqp->header.seq > (call->rnext + call->rwind))) {
5035 #ifndef RX_ENABLE_TSFPQ
5036 if (!optionalPacket)
5039 rxi_CallError(call, RX_CALL_DEAD);
5040 return optionalPacket;
5043 while (rqp->header.seq > call->rnext + offset)
5044 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5045 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5047 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5048 #ifndef RX_ENABLE_TSFPQ
5049 if (!optionalPacket)
5052 rxi_CallError(call, RX_CALL_DEAD);
5053 return optionalPacket;
5058 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5060 /* these are new for AFS 3.3 */
5061 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5062 templ = htonl(templ);
5063 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5064 templ = htonl(call->conn->peer->ifMTU);
5065 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5066 sizeof(afs_int32), &templ);
5068 /* new for AFS 3.4 */
5069 templ = htonl(call->rwind);
5070 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5071 sizeof(afs_int32), &templ);
5073 /* new for AFS 3.5 */
5074 templ = htonl(call->conn->peer->ifDgramPackets);
5075 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5076 sizeof(afs_int32), &templ);
5078 p->header.serviceId = call->conn->serviceId;
5079 p->header.cid = (call->conn->cid | call->channel);
5080 p->header.callNumber = *call->callNumber;
5082 p->header.securityIndex = call->conn->securityIndex;
5083 p->header.epoch = call->conn->epoch;
5084 p->header.type = RX_PACKET_TYPE_ACK;
5085 p->header.flags = RX_SLOW_START_OK;
5086 if (reason == RX_ACK_PING) {
5087 p->header.flags |= RX_REQUEST_ACK;
5089 clock_GetTime(&call->pingRequestTime);
5092 if (call->conn->type == RX_CLIENT_CONNECTION)
5093 p->header.flags |= RX_CLIENT_INITIATED;
5097 if (rxdebug_active) {
5101 len = _snprintf(msg, sizeof(msg),
5102 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5103 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5104 ntohl(ap->serial), ntohl(ap->previousPacket),
5105 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5106 ap->nAcks, ntohs(ap->bufferSpace) );
5110 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5111 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5115 OutputDebugString(msg);
5117 #else /* AFS_NT40_ENV */
5119 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5120 ap->reason, ntohl(ap->previousPacket),
5121 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5123 for (offset = 0; offset < ap->nAcks; offset++)
5124 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5129 #endif /* AFS_NT40_ENV */
5132 int i, nbytes = p->length;
5134 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5135 if (nbytes <= p->wirevec[i].iov_len) {
5138 savelen = p->wirevec[i].iov_len;
5140 p->wirevec[i].iov_len = nbytes;
5142 rxi_Send(call, p, istack);
5143 p->wirevec[i].iov_len = savelen;
5147 nbytes -= p->wirevec[i].iov_len;
5150 if (rx_stats_active)
5151 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
5152 #ifndef RX_ENABLE_TSFPQ
5153 if (!optionalPacket)
5156 return optionalPacket; /* Return packet for re-use by caller */
5159 /* Send all of the packets in the list in single datagram */
5161 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
5162 int istack, int moreFlag, struct clock *now,
5163 struct clock *retryTime, int resending)
5168 struct rx_connection *conn = call->conn;
5169 struct rx_peer *peer = conn->peer;
5171 MUTEX_ENTER(&peer->peer_lock);
5174 peer->reSends += len;
5175 if (rx_stats_active)
5176 rx_MutexAdd(rx_stats.dataPacketsSent, len, rx_stats_mutex);
5177 MUTEX_EXIT(&peer->peer_lock);
5179 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
5183 /* Set the packet flags and schedule the resend events */
5184 /* Only request an ack for the last packet in the list */
5185 for (i = 0; i < len; i++) {
5186 list[i]->retryTime = *retryTime;
5187 if (list[i]->header.serial) {
5188 /* Exponentially backoff retry times */
5189 if (list[i]->backoff < MAXBACKOFF) {
5190 /* so it can't stay == 0 */
5191 list[i]->backoff = (list[i]->backoff << 1) + 1;
5194 clock_Addmsec(&(list[i]->retryTime),
5195 ((afs_uint32) list[i]->backoff) << 8);
5198 /* Wait a little extra for the ack on the last packet */
5199 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
5200 clock_Addmsec(&(list[i]->retryTime), 400);
5203 /* Record the time sent */
5204 list[i]->timeSent = *now;
5206 /* Ask for an ack on retransmitted packets, on every other packet
5207 * if the peer doesn't support slow start. Ask for an ack on every
5208 * packet until the congestion window reaches the ack rate. */
5209 if (list[i]->header.serial) {
5211 if (rx_stats_active)
5212 rx_MutexIncrement(rx_stats.dataPacketsReSent, rx_stats_mutex);
5214 /* improved RTO calculation- not Karn */
5215 list[i]->firstSent = *now;
5216 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5217 || (!(call->flags & RX_CALL_SLOW_START_OK)
5218 && (list[i]->header.seq & 1)))) {
5223 /* Tag this packet as not being the last in this group,
5224 * for the receiver's benefit */
5225 if (i < len - 1 || moreFlag) {
5226 list[i]->header.flags |= RX_MORE_PACKETS;
5229 /* Install the new retransmit time for the packet, and
5230 * record the time sent */
5231 list[i]->timeSent = *now;
5235 list[len - 1]->header.flags |= RX_REQUEST_ACK;
5238 /* Since we're about to send a data packet to the peer, it's
5239 * safe to nuke any scheduled end-of-packets ack */
5240 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5242 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5243 MUTEX_EXIT(&call->lock);
5245 rxi_SendPacketList(call, conn, list, len, istack);
5247 rxi_SendPacket(call, conn, list[0], istack);
5249 MUTEX_ENTER(&call->lock);
5250 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5252 /* Update last send time for this call (for keep-alive
5253 * processing), and for the connection (so that we can discover
5254 * idle connections) */
5255 call->lastSendData = conn->lastSendTime = call->lastSendTime = clock_Sec();
5258 /* When sending packets we need to follow these rules:
5259 * 1. Never send more than maxDgramPackets in a jumbogram.
5260 * 2. Never send a packet with more than two iovecs in a jumbogram.
5261 * 3. Never send a retransmitted packet in a jumbogram.
5262 * 4. Never send more than cwind/4 packets in a jumbogram
5263 * We always keep the last list we should have sent so we
5264 * can set the RX_MORE_PACKETS flags correctly.
5267 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5268 int istack, struct clock *now, struct clock *retryTime,
5271 int i, cnt, lastCnt = 0;
5272 struct rx_packet **listP, **lastP = 0;
5273 struct rx_peer *peer = call->conn->peer;
5274 int morePackets = 0;
5276 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
5277 /* Does the current packet force us to flush the current list? */
5279 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5280 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5282 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
5284 /* If the call enters an error state stop sending, or if
5285 * we entered congestion recovery mode, stop sending */
5286 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5294 /* Add the current packet to the list if it hasn't been acked.
5295 * Otherwise adjust the list pointer to skip the current packet. */
5296 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5298 /* Do we need to flush the list? */
5299 if (cnt >= (int)peer->maxDgramPackets
5300 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5301 || list[i]->header.serial
5302 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5304 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5305 retryTime, resending);
5306 /* If the call enters an error state stop sending, or if
5307 * we entered congestion recovery mode, stop sending */
5309 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5314 listP = &list[i + 1];
5319 osi_Panic("rxi_SendList error");
5321 listP = &list[i + 1];
5325 /* Send the whole list when the call is in receive mode, when
5326 * the call is in eof mode, when we are in fast recovery mode,
5327 * and when we have the last packet */
5328 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5329 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5330 || (call->flags & RX_CALL_FAST_RECOVER)) {
5331 /* Check for the case where the current list contains
5332 * an acked packet. Since we always send retransmissions
5333 * in a separate packet, we only need to check the first
5334 * packet in the list */
5335 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5339 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5340 retryTime, resending);
5341 /* If the call enters an error state stop sending, or if
5342 * we entered congestion recovery mode, stop sending */
5343 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5347 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5350 } else if (lastCnt > 0) {
5351 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5356 #ifdef RX_ENABLE_LOCKS
5357 /* Call rxi_Start, below, but with the call lock held. */
5359 rxi_StartUnlocked(struct rxevent *event,
5360 void *arg0, void *arg1, int istack)
5362 struct rx_call *call = arg0;
5364 MUTEX_ENTER(&call->lock);
5365 rxi_Start(event, call, arg1, istack);
5366 MUTEX_EXIT(&call->lock);
5368 #endif /* RX_ENABLE_LOCKS */
5370 /* This routine is called when new packets are readied for
5371 * transmission and when retransmission may be necessary, or when the
5372 * transmission window or burst count are favourable. This should be
5373 * better optimized for new packets, the usual case, now that we've
5374 * got rid of queues of send packets. XXXXXXXXXXX */
5376 rxi_Start(struct rxevent *event,
5377 void *arg0, void *arg1, int istack)
5379 struct rx_call *call = arg0;
5381 struct rx_packet *p;
5382 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5383 struct rx_peer *peer = call->conn->peer;
5384 struct clock now, usenow, retryTime;
5388 struct rx_packet **xmitList;
5391 /* If rxi_Start is being called as a result of a resend event,
5392 * then make sure that the event pointer is removed from the call
5393 * structure, since there is no longer a per-call retransmission
5395 if (event && event == call->resendEvent) {
5396 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5397 call->resendEvent = NULL;
5399 if (queue_IsEmpty(&call->tq)) {
5403 /* Timeouts trigger congestion recovery */
5404 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5405 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5406 /* someone else is waiting to start recovery */
5409 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5410 rxi_WaitforTQBusy(call);
5411 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5412 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5413 call->flags |= RX_CALL_FAST_RECOVER;
5414 if (peer->maxDgramPackets > 1) {
5415 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5417 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5419 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5420 call->nDgramPackets = 1;
5422 call->nextCwind = 1;
5425 MUTEX_ENTER(&peer->peer_lock);
5426 peer->MTU = call->MTU;
5427 peer->cwind = call->cwind;
5428 peer->nDgramPackets = 1;
5430 call->congestSeq = peer->congestSeq;
5431 MUTEX_EXIT(&peer->peer_lock);
5432 /* Clear retry times on packets. Otherwise, it's possible for
5433 * some packets in the queue to force resends at rates faster
5434 * than recovery rates.
5436 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5437 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5438 clock_Zero(&p->retryTime);
5443 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5444 if (rx_stats_active)
5445 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5450 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5451 /* Get clock to compute the re-transmit time for any packets
5452 * in this burst. Note, if we back off, it's reasonable to
5453 * back off all of the packets in the same manner, even if
5454 * some of them have been retransmitted more times than more
5456 * Do a dance to avoid blocking after setting now. */
5457 MUTEX_ENTER(&peer->peer_lock);
5458 retryTime = peer->timeout;
5459 MUTEX_EXIT(&peer->peer_lock);
5460 clock_GetTime(&now);
5461 clock_Add(&retryTime, &now);
5463 /* Send (or resend) any packets that need it, subject to
5464 * window restrictions and congestion burst control
5465 * restrictions. Ask for an ack on the last packet sent in
5466 * this burst. For now, we're relying upon the window being
5467 * considerably bigger than the largest number of packets that
5468 * are typically sent at once by one initial call to
5469 * rxi_Start. This is probably bogus (perhaps we should ask
5470 * for an ack when we're half way through the current
5471 * window?). Also, for non file transfer applications, this
5472 * may end up asking for an ack for every packet. Bogus. XXXX
5475 * But check whether we're here recursively, and let the other guy
5478 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5479 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5480 call->flags |= RX_CALL_TQ_BUSY;
5482 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5484 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5485 call->flags &= ~RX_CALL_NEED_START;
5486 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5488 maxXmitPackets = MIN(call->twind, call->cwind);
5489 xmitList = (struct rx_packet **)
5490 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
5491 /* XXXX else we must drop any mtx we hold */
5492 afs_osi_Alloc_NoSleep(maxXmitPackets * sizeof(struct rx_packet *));
5494 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5496 if (xmitList == NULL)
5497 osi_Panic("rxi_Start, failed to allocate xmit list");
5498 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5499 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5500 /* We shouldn't be sending packets if a thread is waiting
5501 * to initiate congestion recovery */
5502 dpf(("call %d waiting to initiate fast recovery\n",
5503 *(call->callNumber)));
5507 && (call->flags & RX_CALL_FAST_RECOVER)) {
5508 /* Only send one packet during fast recovery */
5509 dpf(("call %d restricted to one packet per send during fast recovery\n",
5510 *(call->callNumber)));
5513 if ((p->flags & RX_PKTFLAG_FREE)
5514 || (!queue_IsEnd(&call->tq, nxp)
5515 && (nxp->flags & RX_PKTFLAG_FREE))
5516 || (p == (struct rx_packet *)&rx_freePacketQueue)
5517 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5518 osi_Panic("rxi_Start: xmit queue clobbered");
5520 if (p->flags & RX_PKTFLAG_ACKED) {
5521 /* Since we may block, don't trust this */
5522 usenow.sec = usenow.usec = 0;
5523 if (rx_stats_active)
5524 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5525 continue; /* Ignore this packet if it has been acknowledged */
5528 /* Turn off all flags except these ones, which are the same
5529 * on each transmission */
5530 p->header.flags &= RX_PRESET_FLAGS;
5532 if (p->header.seq >=
5533 call->tfirst + MIN((int)call->twind,
5534 (int)(call->nSoftAcked +
5536 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5537 /* Note: if we're waiting for more window space, we can
5538 * still send retransmits; hence we don't return here, but
5539 * break out to schedule a retransmit event */
5540 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
5541 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
5546 /* Transmit the packet if it needs to be sent. */
5547 if (!clock_Lt(&now, &p->retryTime)) {
5548 if (nXmitPackets == maxXmitPackets) {
5549 rxi_SendXmitList(call, xmitList, nXmitPackets,
5550 istack, &now, &retryTime,
5552 osi_Free(xmitList, maxXmitPackets *
5553 sizeof(struct rx_packet *));
5556 dpf(("call %d xmit packet %"AFS_PTR_FMT" now %u.%06u retryTime %u.%06u nextRetry %u.%06u\n",
5557 *(call->callNumber), p,
5559 p->retryTime.sec, p->retryTime.usec,
5560 retryTime.sec, retryTime.usec));
5561 xmitList[nXmitPackets++] = p;
5565 /* xmitList now hold pointers to all of the packets that are
5566 * ready to send. Now we loop to send the packets */
5567 if (nXmitPackets > 0) {
5568 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5569 &now, &retryTime, resending);
5572 maxXmitPackets * sizeof(struct rx_packet *));
5574 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5576 * TQ references no longer protected by this flag; they must remain
5577 * protected by the global lock.
5579 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5580 call->flags &= ~RX_CALL_TQ_BUSY;
5581 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5582 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
5583 call, call->tqWaiters, call->flags));
5584 #ifdef RX_ENABLE_LOCKS
5585 osirx_AssertMine(&call->lock, "rxi_Start start");
5586 CV_BROADCAST(&call->cv_tq);
5587 #else /* RX_ENABLE_LOCKS */
5588 osi_rxWakeup(&call->tq);
5589 #endif /* RX_ENABLE_LOCKS */
5594 /* We went into the error state while sending packets. Now is
5595 * the time to reset the call. This will also inform the using
5596 * process that the call is in an error state.
5598 if (rx_stats_active)
5599 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5600 call->flags &= ~RX_CALL_TQ_BUSY;
5601 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5602 dpf(("call error %d while xmit %p has %d waiters and flags %d\n",
5603 call->error, call, call->tqWaiters, call->flags));
5604 #ifdef RX_ENABLE_LOCKS
5605 osirx_AssertMine(&call->lock, "rxi_Start middle");
5606 CV_BROADCAST(&call->cv_tq);
5607 #else /* RX_ENABLE_LOCKS */
5608 osi_rxWakeup(&call->tq);
5609 #endif /* RX_ENABLE_LOCKS */
5611 rxi_CallError(call, call->error);
5614 #ifdef RX_ENABLE_LOCKS
5615 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5617 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5618 /* Some packets have received acks. If they all have, we can clear
5619 * the transmit queue.
5622 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5623 if (p->header.seq < call->tfirst
5624 && (p->flags & RX_PKTFLAG_ACKED)) {
5626 p->flags &= ~RX_PKTFLAG_TQ;
5627 #ifdef RXDEBUG_PACKET
5635 call->flags |= RX_CALL_TQ_CLEARME;
5637 #endif /* RX_ENABLE_LOCKS */
5638 /* Don't bother doing retransmits if the TQ is cleared. */
5639 if (call->flags & RX_CALL_TQ_CLEARME) {
5640 rxi_ClearTransmitQueue(call, 1);
5642 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5645 /* Always post a resend event, if there is anything in the
5646 * queue, and resend is possible. There should be at least
5647 * one unacknowledged packet in the queue ... otherwise none
5648 * of these packets should be on the queue in the first place.
5650 if (call->resendEvent) {
5651 /* Cancel the existing event and post a new one */
5652 rxevent_Cancel(call->resendEvent, call,
5653 RX_CALL_REFCOUNT_RESEND);
5656 /* The retry time is the retry time on the first unacknowledged
5657 * packet inside the current window */
5659 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5660 /* Don't set timers for packets outside the window */
5661 if (p->header.seq >= call->tfirst + call->twind) {
5665 if (!(p->flags & RX_PKTFLAG_ACKED)
5666 && !clock_IsZero(&p->retryTime)) {
5668 retryTime = p->retryTime;
5673 /* Post a new event to re-run rxi_Start when retries may be needed */
5674 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5675 #ifdef RX_ENABLE_LOCKS
5676 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5678 rxevent_PostNow2(&retryTime, &usenow,
5680 (void *)call, 0, istack);
5681 #else /* RX_ENABLE_LOCKS */
5683 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5684 (void *)call, 0, istack);
5685 #endif /* RX_ENABLE_LOCKS */
5688 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5689 } while (call->flags & RX_CALL_NEED_START);
5691 * TQ references no longer protected by this flag; they must remain
5692 * protected by the global lock.
5694 call->flags &= ~RX_CALL_TQ_BUSY;
5695 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5696 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
5697 call, call->tqWaiters, call->flags));
5698 #ifdef RX_ENABLE_LOCKS
5699 osirx_AssertMine(&call->lock, "rxi_Start end");
5700 CV_BROADCAST(&call->cv_tq);
5701 #else /* RX_ENABLE_LOCKS */
5702 osi_rxWakeup(&call->tq);
5703 #endif /* RX_ENABLE_LOCKS */
5706 call->flags |= RX_CALL_NEED_START;
5708 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5710 if (call->resendEvent) {
5711 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5716 /* Also adjusts the keep alive parameters for the call, to reflect
5717 * that we have just sent a packet (so keep alives aren't sent
5720 rxi_Send(struct rx_call *call, struct rx_packet *p,
5723 struct rx_connection *conn = call->conn;
5725 /* Stamp each packet with the user supplied status */
5726 p->header.userStatus = call->localStatus;
5728 /* Allow the security object controlling this call's security to
5729 * make any last-minute changes to the packet */
5730 RXS_SendPacket(conn->securityObject, call, p);
5732 /* Since we're about to send SOME sort of packet to the peer, it's
5733 * safe to nuke any scheduled end-of-packets ack */
5734 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5736 /* Actually send the packet, filling in more connection-specific fields */
5737 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5738 MUTEX_EXIT(&call->lock);
5739 rxi_SendPacket(call, conn, p, istack);
5740 MUTEX_ENTER(&call->lock);
5741 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5743 /* Update last send time for this call (for keep-alive
5744 * processing), and for the connection (so that we can discover
5745 * idle connections) */
5746 conn->lastSendTime = call->lastSendTime = clock_Sec();
5747 /* Don't count keepalives here, so idleness can be tracked. */
5748 if ((p->header.type != RX_PACKET_TYPE_ACK) || (((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING))
5749 call->lastSendData = call->lastSendTime;
5753 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5754 * that things are fine. Also called periodically to guarantee that nothing
5755 * falls through the cracks (e.g. (error + dally) connections have keepalive
5756 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5758 * haveCTLock Set if calling from rxi_ReapConnections
5760 #ifdef RX_ENABLE_LOCKS
5762 rxi_CheckCall(struct rx_call *call, int haveCTLock)
5763 #else /* RX_ENABLE_LOCKS */
5765 rxi_CheckCall(struct rx_call *call)
5766 #endif /* RX_ENABLE_LOCKS */
5768 struct rx_connection *conn = call->conn;
5770 afs_uint32 deadTime;
5772 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5773 if (call->flags & RX_CALL_TQ_BUSY) {
5774 /* Call is active and will be reset by rxi_Start if it's
5775 * in an error state.
5780 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5782 (((afs_uint32) conn->secondsUntilDead << 10) +
5783 ((afs_uint32) conn->peer->rtt >> 3) +
5784 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5786 /* These are computed to the second (+- 1 second). But that's
5787 * good enough for these values, which should be a significant
5788 * number of seconds. */
5789 if (now > (call->lastReceiveTime + deadTime)) {
5790 if (call->state == RX_STATE_ACTIVE) {
5792 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5794 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5795 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5796 ip_stack_t *ipst = ns->netstack_ip;
5798 ire = ire_cache_lookup(call->conn->peer->host
5799 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5801 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5803 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5810 if (ire && ire->ire_max_frag > 0)
5811 rxi_SetPeerMtu(call->conn->peer->host, 0, ire->ire_max_frag);
5812 #if defined(GLOBAL_NETSTACKID)
5816 #endif /* ADAPT_PMTU */
5817 rxi_CallError(call, RX_CALL_DEAD);
5820 #ifdef RX_ENABLE_LOCKS
5821 /* Cancel pending events */
5822 rxevent_Cancel(call->delayedAckEvent, call,
5823 RX_CALL_REFCOUNT_DELAY);
5824 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5825 rxevent_Cancel(call->keepAliveEvent, call,
5826 RX_CALL_REFCOUNT_ALIVE);
5827 if (call->refCount == 0) {
5828 rxi_FreeCall(call, haveCTLock);
5832 #else /* RX_ENABLE_LOCKS */
5835 #endif /* RX_ENABLE_LOCKS */
5837 /* Non-active calls are destroyed if they are not responding
5838 * to pings; active calls are simply flagged in error, so the
5839 * attached process can die reasonably gracefully. */
5841 /* see if we have a non-activity timeout */
5842 if (call->startWait && conn->idleDeadTime
5843 && ((call->startWait + conn->idleDeadTime) < now) &&
5844 (call->flags & RX_CALL_READER_WAIT)) {
5845 if (call->state == RX_STATE_ACTIVE) {
5846 rxi_CallError(call, RX_CALL_TIMEOUT);
5850 if (call->lastSendData && conn->idleDeadTime && (conn->idleDeadErr != 0)
5851 && ((call->lastSendData + conn->idleDeadTime) < now)) {
5852 if (call->state == RX_STATE_ACTIVE) {
5853 rxi_CallError(call, conn->idleDeadErr);
5857 /* see if we have a hard timeout */
5858 if (conn->hardDeadTime
5859 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5860 if (call->state == RX_STATE_ACTIVE)
5861 rxi_CallError(call, RX_CALL_TIMEOUT);
5868 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
5870 struct rx_connection *conn = arg1;
5871 struct rx_header theader;
5873 struct sockaddr_in taddr;
5876 struct iovec tmpiov[2];
5879 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
5882 tp = &tbuffer[sizeof(struct rx_header)];
5883 taddr.sin_family = AF_INET;
5884 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
5885 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
5886 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
5887 taddr.sin_len = sizeof(struct sockaddr_in);
5889 memset(&theader, 0, sizeof(theader));
5890 theader.epoch = htonl(999);
5892 theader.callNumber = 0;
5895 theader.type = RX_PACKET_TYPE_VERSION;
5896 theader.flags = RX_LAST_PACKET;
5897 theader.serviceId = 0;
5899 memcpy(tbuffer, &theader, sizeof(theader));
5900 memcpy(tp, &a, sizeof(a));
5901 tmpiov[0].iov_base = tbuffer;
5902 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
5904 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
5906 MUTEX_ENTER(&conn->conn_data_lock);
5907 /* Only reschedule ourselves if the connection would not be destroyed */
5908 if (conn->refCount <= 1) {
5909 conn->natKeepAliveEvent = NULL;
5910 MUTEX_EXIT(&conn->conn_data_lock);
5911 rx_DestroyConnection(conn); /* drop the reference for this */
5913 conn->natKeepAliveEvent = NULL;
5914 conn->refCount--; /* drop the reference for this */
5915 rxi_ScheduleNatKeepAliveEvent(conn);
5916 MUTEX_EXIT(&conn->conn_data_lock);
5921 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
5923 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
5924 struct clock when, now;
5925 clock_GetTime(&now);
5927 when.sec += conn->secondsUntilNatPing;
5928 conn->refCount++; /* hold a reference for this */
5929 conn->natKeepAliveEvent =
5930 rxevent_PostNow(&when, &now, rxi_NatKeepAliveEvent, conn, 0);
5935 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
5937 MUTEX_ENTER(&conn->conn_data_lock);
5938 conn->secondsUntilNatPing = seconds;
5940 rxi_ScheduleNatKeepAliveEvent(conn);
5941 MUTEX_EXIT(&conn->conn_data_lock);
5945 rxi_NatKeepAliveOn(struct rx_connection *conn)
5947 MUTEX_ENTER(&conn->conn_data_lock);
5948 rxi_ScheduleNatKeepAliveEvent(conn);
5949 MUTEX_EXIT(&conn->conn_data_lock);
5952 /* When a call is in progress, this routine is called occasionally to
5953 * make sure that some traffic has arrived (or been sent to) the peer.
5954 * If nothing has arrived in a reasonable amount of time, the call is
5955 * declared dead; if nothing has been sent for a while, we send a
5956 * keep-alive packet (if we're actually trying to keep the call alive)
5959 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
5961 struct rx_call *call = arg1;
5962 struct rx_connection *conn;
5965 MUTEX_ENTER(&call->lock);
5966 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5967 if (event == call->keepAliveEvent)
5968 call->keepAliveEvent = NULL;
5971 #ifdef RX_ENABLE_LOCKS
5972 if (rxi_CheckCall(call, 0)) {
5973 MUTEX_EXIT(&call->lock);
5976 #else /* RX_ENABLE_LOCKS */
5977 if (rxi_CheckCall(call))
5979 #endif /* RX_ENABLE_LOCKS */
5981 /* Don't try to keep alive dallying calls */
5982 if (call->state == RX_STATE_DALLY) {
5983 MUTEX_EXIT(&call->lock);
5988 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5989 /* Don't try to send keepalives if there is unacknowledged data */
5990 /* the rexmit code should be good enough, this little hack
5991 * doesn't quite work XXX */
5992 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5994 rxi_ScheduleKeepAliveEvent(call);
5995 MUTEX_EXIT(&call->lock);
6000 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6002 if (!call->keepAliveEvent) {
6003 struct clock when, now;
6004 clock_GetTime(&now);
6006 when.sec += call->conn->secondsUntilPing;
6007 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6008 call->keepAliveEvent =
6009 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
6013 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6015 rxi_KeepAliveOn(struct rx_call *call)
6017 /* Pretend last packet received was received now--i.e. if another
6018 * packet isn't received within the keep alive time, then the call
6019 * will die; Initialize last send time to the current time--even
6020 * if a packet hasn't been sent yet. This will guarantee that a
6021 * keep-alive is sent within the ping time */
6022 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6023 rxi_ScheduleKeepAliveEvent(call);
6026 /* This routine is called to send connection abort messages
6027 * that have been delayed to throttle looping clients. */
6029 rxi_SendDelayedConnAbort(struct rxevent *event,
6030 void *arg1, void *unused)
6032 struct rx_connection *conn = arg1;
6035 struct rx_packet *packet;
6037 MUTEX_ENTER(&conn->conn_data_lock);
6038 conn->delayedAbortEvent = NULL;
6039 error = htonl(conn->error);
6041 MUTEX_EXIT(&conn->conn_data_lock);
6042 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6045 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6046 RX_PACKET_TYPE_ABORT, (char *)&error,
6048 rxi_FreePacket(packet);
6052 /* This routine is called to send call abort messages
6053 * that have been delayed to throttle looping clients. */
6055 rxi_SendDelayedCallAbort(struct rxevent *event,
6056 void *arg1, void *dummy)
6058 struct rx_call *call = arg1;
6061 struct rx_packet *packet;
6063 MUTEX_ENTER(&call->lock);
6064 call->delayedAbortEvent = NULL;
6065 error = htonl(call->error);
6067 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6070 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6071 (char *)&error, sizeof(error), 0);
6072 rxi_FreePacket(packet);
6074 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6075 MUTEX_EXIT(&call->lock);
6078 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6079 * seconds) to ask the client to authenticate itself. The routine
6080 * issues a challenge to the client, which is obtained from the
6081 * security object associated with the connection */
6083 rxi_ChallengeEvent(struct rxevent *event,
6084 void *arg0, void *arg1, int tries)
6086 struct rx_connection *conn = arg0;
6088 conn->challengeEvent = NULL;
6089 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6090 struct rx_packet *packet;
6091 struct clock when, now;
6094 /* We've failed to authenticate for too long.
6095 * Reset any calls waiting for authentication;
6096 * they are all in RX_STATE_PRECALL.
6100 MUTEX_ENTER(&conn->conn_call_lock);
6101 for (i = 0; i < RX_MAXCALLS; i++) {
6102 struct rx_call *call = conn->call[i];
6104 MUTEX_ENTER(&call->lock);
6105 if (call->state == RX_STATE_PRECALL) {
6106 rxi_CallError(call, RX_CALL_DEAD);
6107 rxi_SendCallAbort(call, NULL, 0, 0);
6109 MUTEX_EXIT(&call->lock);
6112 MUTEX_EXIT(&conn->conn_call_lock);
6116 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6118 /* If there's no packet available, do this later. */
6119 RXS_GetChallenge(conn->securityObject, conn, packet);
6120 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6121 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6122 rxi_FreePacket(packet);
6124 clock_GetTime(&now);
6126 when.sec += RX_CHALLENGE_TIMEOUT;
6127 conn->challengeEvent =
6128 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
6133 /* Call this routine to start requesting the client to authenticate
6134 * itself. This will continue until authentication is established,
6135 * the call times out, or an invalid response is returned. The
6136 * security object associated with the connection is asked to create
6137 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6138 * defined earlier. */
6140 rxi_ChallengeOn(struct rx_connection *conn)
6142 if (!conn->challengeEvent) {
6143 RXS_CreateChallenge(conn->securityObject, conn);
6144 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6149 /* Compute round trip time of the packet provided, in *rttp.
6152 /* rxi_ComputeRoundTripTime is called with peer locked. */
6153 /* sentp and/or peer may be null */
6155 rxi_ComputeRoundTripTime(struct rx_packet *p,
6156 struct clock *sentp,
6157 struct rx_peer *peer)
6159 struct clock thisRtt, *rttp = &thisRtt;
6163 clock_GetTime(rttp);
6165 if (clock_Lt(rttp, sentp)) {
6167 return; /* somebody set the clock back, don't count this time. */
6169 clock_Sub(rttp, sentp);
6170 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6171 p->header.callNumber, p, rttp->sec, rttp->usec));
6173 if (rttp->sec == 0 && rttp->usec == 0) {
6175 * The actual round trip time is shorter than the
6176 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6177 * Since we can't tell which at the moment we will assume 1ms.
6182 if (rx_stats_active) {
6183 MUTEX_ENTER(&rx_stats_mutex);
6184 if (clock_Lt(rttp, &rx_stats.minRtt))
6185 rx_stats.minRtt = *rttp;
6186 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
6187 if (rttp->sec > 60) {
6188 MUTEX_EXIT(&rx_stats_mutex);
6189 return; /* somebody set the clock ahead */
6191 rx_stats.maxRtt = *rttp;
6193 clock_Add(&rx_stats.totalRtt, rttp);
6194 rx_stats.nRttSamples++;
6195 MUTEX_EXIT(&rx_stats_mutex);
6198 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6200 /* Apply VanJacobson round-trip estimations */
6205 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
6206 * srtt is stored as fixed point with 3 bits after the binary
6207 * point (i.e., scaled by 8). The following magic is
6208 * equivalent to the smoothing algorithm in rfc793 with an
6209 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6210 * srtt'*8 = rtt + srtt*7
6211 * srtt'*8 = srtt*8 + rtt - srtt
6212 * srtt' = srtt + rtt/8 - srtt/8
6213 * srtt' = srtt + (rtt - srtt)/8
6216 delta = _8THMSEC(rttp) - peer->rtt;
6217 peer->rtt += (delta >> 3);
6220 * We accumulate a smoothed rtt variance (actually, a smoothed
6221 * mean difference), then set the retransmit timer to smoothed
6222 * rtt + 4 times the smoothed variance (was 2x in van's original
6223 * paper, but 4x works better for me, and apparently for him as
6225 * rttvar is stored as
6226 * fixed point with 2 bits after the binary point (scaled by
6227 * 4). The following is equivalent to rfc793 smoothing with
6228 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6229 * rttvar'*4 = rttvar*3 + |delta|
6230 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6231 * rttvar' = rttvar + |delta|/4 - rttvar/4
6232 * rttvar' = rttvar + (|delta| - rttvar)/4
6233 * This replaces rfc793's wired-in beta.
6234 * dev*4 = dev*4 + (|actual - expected| - dev)
6240 delta -= (peer->rtt_dev << 1);
6241 peer->rtt_dev += (delta >> 3);
6243 /* I don't have a stored RTT so I start with this value. Since I'm
6244 * probably just starting a call, and will be pushing more data down
6245 * this, I expect congestion to increase rapidly. So I fudge a
6246 * little, and I set deviance to half the rtt. In practice,
6247 * deviance tends to approach something a little less than
6248 * half the smoothed rtt. */
6249 peer->rtt = _8THMSEC(rttp) + 8;
6250 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
6252 /* the timeout is RTT + 4*MDEV but no less than rx_minPeerTimeout msec.
6253 * This is because one end or the other of these connections is usually
6254 * in a user process, and can be switched and/or swapped out. So on fast,
6255 * reliable networks, the timeout would otherwise be too short. */
6256 rtt_timeout = MAX(((peer->rtt >> 3) + peer->rtt_dev), rx_minPeerTimeout);
6257 clock_Zero(&(peer->timeout));
6258 clock_Addmsec(&(peer->timeout), rtt_timeout);
6260 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6261 p->header.callNumber, p, MSEC(rttp), peer->rtt >> 3, peer->rtt_dev >> 2, (peer->timeout.sec), (peer->timeout.usec)));
6265 /* Find all server connections that have not been active for a long time, and
6268 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6270 struct clock now, when;
6271 clock_GetTime(&now);
6273 /* Find server connection structures that haven't been used for
6274 * greater than rx_idleConnectionTime */
6276 struct rx_connection **conn_ptr, **conn_end;
6277 int i, havecalls = 0;
6278 MUTEX_ENTER(&rx_connHashTable_lock);
6279 for (conn_ptr = &rx_connHashTable[0], conn_end =
6280 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6282 struct rx_connection *conn, *next;
6283 struct rx_call *call;
6287 for (conn = *conn_ptr; conn; conn = next) {
6288 /* XXX -- Shouldn't the connection be locked? */
6291 for (i = 0; i < RX_MAXCALLS; i++) {
6292 call = conn->call[i];
6296 code = MUTEX_TRYENTER(&call->lock);
6299 #ifdef RX_ENABLE_LOCKS
6300 result = rxi_CheckCall(call, 1);
6301 #else /* RX_ENABLE_LOCKS */
6302 result = rxi_CheckCall(call);
6303 #endif /* RX_ENABLE_LOCKS */
6304 MUTEX_EXIT(&call->lock);
6306 /* If CheckCall freed the call, it might
6307 * have destroyed the connection as well,
6308 * which screws up the linked lists.
6314 if (conn->type == RX_SERVER_CONNECTION) {
6315 /* This only actually destroys the connection if
6316 * there are no outstanding calls */
6317 MUTEX_ENTER(&conn->conn_data_lock);
6318 if (!havecalls && !conn->refCount
6319 && ((conn->lastSendTime + rx_idleConnectionTime) <
6321 conn->refCount++; /* it will be decr in rx_DestroyConn */
6322 MUTEX_EXIT(&conn->conn_data_lock);
6323 #ifdef RX_ENABLE_LOCKS
6324 rxi_DestroyConnectionNoLock(conn);
6325 #else /* RX_ENABLE_LOCKS */
6326 rxi_DestroyConnection(conn);
6327 #endif /* RX_ENABLE_LOCKS */
6329 #ifdef RX_ENABLE_LOCKS
6331 MUTEX_EXIT(&conn->conn_data_lock);
6333 #endif /* RX_ENABLE_LOCKS */
6337 #ifdef RX_ENABLE_LOCKS
6338 while (rx_connCleanup_list) {
6339 struct rx_connection *conn;
6340 conn = rx_connCleanup_list;
6341 rx_connCleanup_list = rx_connCleanup_list->next;
6342 MUTEX_EXIT(&rx_connHashTable_lock);
6343 rxi_CleanupConnection(conn);
6344 MUTEX_ENTER(&rx_connHashTable_lock);
6346 MUTEX_EXIT(&rx_connHashTable_lock);
6347 #endif /* RX_ENABLE_LOCKS */
6350 /* Find any peer structures that haven't been used (haven't had an
6351 * associated connection) for greater than rx_idlePeerTime */
6353 struct rx_peer **peer_ptr, **peer_end;
6357 * Why do we need to hold the rx_peerHashTable_lock across
6358 * the incrementing of peer_ptr since the rx_peerHashTable
6359 * array is not changing? We don't.
6361 * By dropping the lock periodically we can permit other
6362 * activities to be performed while a rxi_ReapConnections
6363 * call is in progress. The goal of reap connections
6364 * is to clean up quickly without causing large amounts
6365 * of contention. Therefore, it is important that global
6366 * mutexes not be held for extended periods of time.
6368 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6369 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6371 struct rx_peer *peer, *next, *prev;
6373 MUTEX_ENTER(&rx_peerHashTable_lock);
6374 for (prev = peer = *peer_ptr; peer; peer = next) {
6376 code = MUTEX_TRYENTER(&peer->peer_lock);
6377 if ((code) && (peer->refCount == 0)
6378 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6379 rx_interface_stat_p rpc_stat, nrpc_stat;
6383 * now know that this peer object is one to be
6384 * removed from the hash table. Once it is removed
6385 * it can't be referenced by other threads.
6386 * Lets remove it first and decrement the struct
6387 * nPeerStructs count.
6389 if (peer == *peer_ptr) {
6395 if (rx_stats_active)
6396 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6399 * Now if we hold references on 'prev' and 'next'
6400 * we can safely drop the rx_peerHashTable_lock
6401 * while we destroy this 'peer' object.
6407 MUTEX_EXIT(&rx_peerHashTable_lock);
6409 MUTEX_EXIT(&peer->peer_lock);
6410 MUTEX_DESTROY(&peer->peer_lock);
6412 (&peer->rpcStats, rpc_stat, nrpc_stat,
6413 rx_interface_stat)) {
6414 unsigned int num_funcs;
6417 queue_Remove(&rpc_stat->queue_header);
6418 queue_Remove(&rpc_stat->all_peers);
6419 num_funcs = rpc_stat->stats[0].func_total;
6421 sizeof(rx_interface_stat_t) +
6422 rpc_stat->stats[0].func_total *
6423 sizeof(rx_function_entry_v1_t);
6425 rxi_Free(rpc_stat, space);
6427 MUTEX_ENTER(&rx_rpc_stats);
6428 rxi_rpc_peer_stat_cnt -= num_funcs;
6429 MUTEX_EXIT(&rx_rpc_stats);
6434 * Regain the rx_peerHashTable_lock and
6435 * decrement the reference count on 'prev'
6438 MUTEX_ENTER(&rx_peerHashTable_lock);
6445 MUTEX_EXIT(&peer->peer_lock);
6450 MUTEX_EXIT(&rx_peerHashTable_lock);
6454 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6455 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6456 * GC, just below. Really, we shouldn't have to keep moving packets from
6457 * one place to another, but instead ought to always know if we can
6458 * afford to hold onto a packet in its particular use. */
6459 MUTEX_ENTER(&rx_freePktQ_lock);
6460 if (rx_waitingForPackets) {
6461 rx_waitingForPackets = 0;
6462 #ifdef RX_ENABLE_LOCKS
6463 CV_BROADCAST(&rx_waitingForPackets_cv);
6465 osi_rxWakeup(&rx_waitingForPackets);
6468 MUTEX_EXIT(&rx_freePktQ_lock);
6471 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6472 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6476 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6477 * rx.h is sort of strange this is better. This is called with a security
6478 * object before it is discarded. Each connection using a security object has
6479 * its own refcount to the object so it won't actually be freed until the last
6480 * connection is destroyed.
6482 * This is the only rxs module call. A hold could also be written but no one
6486 rxs_Release(struct rx_securityClass *aobj)
6488 return RXS_Close(aobj);
6492 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6493 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6494 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6495 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6497 /* Adjust our estimate of the transmission rate to this peer, given
6498 * that the packet p was just acked. We can adjust peer->timeout and
6499 * call->twind. Pragmatically, this is called
6500 * only with packets of maximal length.
6501 * Called with peer and call locked.
6505 rxi_ComputeRate(struct rx_peer *peer, struct rx_call *call,
6506 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6508 afs_int32 xferSize, xferMs;
6512 /* Count down packets */
6513 if (peer->rateFlag > 0)
6515 /* Do nothing until we're enabled */
6516 if (peer->rateFlag != 0)
6521 /* Count only when the ack seems legitimate */
6522 switch (ackReason) {
6523 case RX_ACK_REQUESTED:
6525 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6529 case RX_ACK_PING_RESPONSE:
6530 if (p) /* want the response to ping-request, not data send */
6532 clock_GetTime(&newTO);
6533 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6534 clock_Sub(&newTO, &call->pingRequestTime);
6535 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6539 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6546 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %d.%06d, rtt %u, ps %u)",
6547 ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"),
6548 xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
6550 /* Track only packets that are big enough. */
6551 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6555 /* absorb RTT data (in milliseconds) for these big packets */
6556 if (peer->smRtt == 0) {
6557 peer->smRtt = xferMs;
6559 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6564 if (peer->countDown) {
6568 peer->countDown = 10; /* recalculate only every so often */
6570 /* In practice, we can measure only the RTT for full packets,
6571 * because of the way Rx acks the data that it receives. (If it's
6572 * smaller than a full packet, it often gets implicitly acked
6573 * either by the call response (from a server) or by the next call
6574 * (from a client), and either case confuses transmission times
6575 * with processing times.) Therefore, replace the above
6576 * more-sophisticated processing with a simpler version, where the
6577 * smoothed RTT is kept for full-size packets, and the time to
6578 * transmit a windowful of full-size packets is simply RTT *
6579 * windowSize. Again, we take two steps:
6580 - ensure the timeout is large enough for a single packet's RTT;
6581 - ensure that the window is small enough to fit in the desired timeout.*/
6583 /* First, the timeout check. */
6584 minTime = peer->smRtt;
6585 /* Get a reasonable estimate for a timeout period */
6587 newTO.sec = minTime / 1000;
6588 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6590 /* Increase the timeout period so that we can always do at least
6591 * one packet exchange */
6592 if (clock_Gt(&newTO, &peer->timeout)) {
6594 dpf(("CONG peer %lx/%u: timeout %d.%06d ==> %ld.%06d (rtt %u, ps %u)",
6595 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec,
6596 newTO.sec, newTO.usec, peer->smRtt, peer->packetSize));
6598 peer->timeout = newTO;
6601 /* Now, get an estimate for the transmit window size. */
6602 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6603 /* Now, convert to the number of full packets that could fit in a
6604 * reasonable fraction of that interval */
6605 minTime /= (peer->smRtt << 1);
6606 xferSize = minTime; /* (make a copy) */
6608 /* Now clamp the size to reasonable bounds. */
6611 else if (minTime > rx_Window)
6612 minTime = rx_Window;
6613 /* if (minTime != peer->maxWindow) {
6614 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6615 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6616 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6618 peer->maxWindow = minTime;
6619 elide... call->twind = minTime;
6623 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6624 * Discern this by calculating the timeout necessary for rx_Window
6626 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6627 /* calculate estimate for transmission interval in milliseconds */
6628 minTime = rx_Window * peer->smRtt;
6629 if (minTime < 1000) {
6630 dpf(("CONG peer %lx/%u: cut TO %d.%06d by 0.5 (rtt %u, ps %u)",
6631 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6632 peer->timeout.usec, peer->smRtt, peer->packetSize));
6634 newTO.sec = 0; /* cut back on timeout by half a second */
6635 newTO.usec = 500000;
6636 clock_Sub(&peer->timeout, &newTO);
6641 } /* end of rxi_ComputeRate */
6642 #endif /* ADAPT_WINDOW */
6650 #define TRACE_OPTION_RX_DEBUG 16
6658 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6659 0, KEY_QUERY_VALUE, &parmKey);
6660 if (code != ERROR_SUCCESS)
6663 dummyLen = sizeof(TraceOption);
6664 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6665 (BYTE *) &TraceOption, &dummyLen);
6666 if (code == ERROR_SUCCESS) {
6667 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
6669 RegCloseKey (parmKey);
6670 #endif /* AFS_NT40_ENV */
6675 rx_DebugOnOff(int on)
6679 rxdebug_active = on;
6685 rx_StatsOnOff(int on)
6688 rx_stats_active = on;
6693 /* Don't call this debugging routine directly; use dpf */
6695 rxi_DebugPrint(char *format, ...)
6704 va_start(ap, format);
6706 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6709 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
6711 if (msg[len-1] != '\n') {
6715 OutputDebugString(msg);
6722 va_start(ap, format);
6724 clock_GetTime(&now);
6725 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
6726 (unsigned int)now.usec);
6727 vfprintf(rx_Log, format, ap);
6736 * This function is used to process the rx_stats structure that is local
6737 * to a process as well as an rx_stats structure received from a remote
6738 * process (via rxdebug). Therefore, it needs to do minimal version
6742 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
6743 afs_int32 freePackets, char version)
6748 if (size != sizeof(struct rx_statistics)) {
6750 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
6751 size, sizeof(struct rx_statistics));
6754 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6757 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6758 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
6759 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6760 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6761 s->specialPktAllocFailures);
6763 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
6764 s->receivePktAllocFailures, s->sendPktAllocFailures,
6765 s->specialPktAllocFailures);
6769 " greedy %u, " "bogusReads %u (last from host %x), "
6770 "noPackets %u, " "noBuffers %u, " "selects %u, "
6771 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
6772 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6773 s->selects, s->sendSelects);
6775 fprintf(file, " packets read: ");
6776 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6777 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
6779 fprintf(file, "\n");
6782 " other read counters: data %u, " "ack %u, " "dup %u "
6783 "spurious %u " "dally %u\n", s->dataPacketsRead,
6784 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6785 s->ignorePacketDally);
6787 fprintf(file, " packets sent: ");
6788 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6789 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
6791 fprintf(file, "\n");
6794 " other send counters: ack %u, " "data %u (not resends), "
6795 "resends %u, " "pushed %u, " "acked&ignored %u\n",
6796 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6797 s->dataPacketsPushed, s->ignoreAckedPacket);
6800 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
6801 s->netSendFailures, (int)s->fatalErrors);
6803 if (s->nRttSamples) {
6804 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6805 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6807 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6808 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6812 " %d server connections, " "%d client connections, "
6813 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6814 s->nServerConns, s->nClientConns, s->nPeerStructs,
6815 s->nCallStructs, s->nFreeCallStructs);
6817 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6818 fprintf(file, " %d clock updates\n", clock_nUpdates);
6821 fprintf(file, "ERROR: compiled without RXDEBUG\n");
6825 /* for backward compatibility */
6827 rx_PrintStats(FILE * file)
6829 MUTEX_ENTER(&rx_stats_mutex);
6830 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6832 MUTEX_EXIT(&rx_stats_mutex);
6836 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6838 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
6839 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6840 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6843 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6844 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6845 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6848 " Packet size %d, " "max in packet skew %d, "
6849 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6850 (int)peer->outPacketSkew);
6854 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
6856 * This mutex protects the following static variables:
6860 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
6861 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
6863 #define LOCK_RX_DEBUG
6864 #define UNLOCK_RX_DEBUG
6865 #endif /* AFS_PTHREAD_ENV */
6869 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6870 u_char type, void *inputData, size_t inputLength,
6871 void *outputData, size_t outputLength)
6873 static afs_int32 counter = 100;
6874 time_t waitTime, waitCount, startTime;
6875 struct rx_header theader;
6878 struct timeval tv_now, tv_wake, tv_delta;
6879 struct sockaddr_in taddr, faddr;
6888 startTime = time(0);
6894 tp = &tbuffer[sizeof(struct rx_header)];
6895 taddr.sin_family = AF_INET;
6896 taddr.sin_port = remotePort;
6897 taddr.sin_addr.s_addr = remoteAddr;
6898 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6899 taddr.sin_len = sizeof(struct sockaddr_in);
6902 memset(&theader, 0, sizeof(theader));
6903 theader.epoch = htonl(999);
6905 theader.callNumber = htonl(counter);
6908 theader.type = type;
6909 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6910 theader.serviceId = 0;
6912 memcpy(tbuffer, &theader, sizeof(theader));
6913 memcpy(tp, inputData, inputLength);
6915 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6916 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6918 /* see if there's a packet available */
6919 gettimeofday(&tv_wake,0);
6920 tv_wake.tv_sec += waitTime;
6923 FD_SET(socket, &imask);
6924 tv_delta.tv_sec = tv_wake.tv_sec;
6925 tv_delta.tv_usec = tv_wake.tv_usec;
6926 gettimeofday(&tv_now, 0);
6928 if (tv_delta.tv_usec < tv_now.tv_usec) {
6930 tv_delta.tv_usec += 1000000;
6933 tv_delta.tv_usec -= tv_now.tv_usec;
6935 if (tv_delta.tv_sec < tv_now.tv_sec) {
6939 tv_delta.tv_sec -= tv_now.tv_sec;
6942 code = select(0, &imask, 0, 0, &tv_delta);
6943 #else /* AFS_NT40_ENV */
6944 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6945 #endif /* AFS_NT40_ENV */
6946 if (code == 1 && FD_ISSET(socket, &imask)) {
6947 /* now receive a packet */
6948 faddrLen = sizeof(struct sockaddr_in);
6950 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6951 (struct sockaddr *)&faddr, &faddrLen);
6954 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6955 if (counter == ntohl(theader.callNumber))
6963 /* see if we've timed out */
6971 code -= sizeof(struct rx_header);
6972 if (code > outputLength)
6973 code = outputLength;
6974 memcpy(outputData, tp, code);
6977 #endif /* RXDEBUG */
6980 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6981 afs_uint16 remotePort, struct rx_debugStats * stat,
6982 afs_uint32 * supportedValues)
6988 struct rx_debugIn in;
6990 *supportedValues = 0;
6991 in.type = htonl(RX_DEBUGI_GETSTATS);
6994 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6995 &in, sizeof(in), stat, sizeof(*stat));
6998 * If the call was successful, fixup the version and indicate
6999 * what contents of the stat structure are valid.
7000 * Also do net to host conversion of fields here.
7004 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7005 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7007 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7008 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7010 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7011 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7013 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7014 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7016 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7017 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7019 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7020 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7022 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7023 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7025 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7026 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7028 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7029 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7031 stat->nFreePackets = ntohl(stat->nFreePackets);
7032 stat->packetReclaims = ntohl(stat->packetReclaims);
7033 stat->callsExecuted = ntohl(stat->callsExecuted);
7034 stat->nWaiting = ntohl(stat->nWaiting);
7035 stat->idleThreads = ntohl(stat->idleThreads);
7036 stat->nWaited = ntohl(stat->nWaited);
7037 stat->nPackets = ntohl(stat->nPackets);
7044 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7045 afs_uint16 remotePort, struct rx_statistics * stat,
7046 afs_uint32 * supportedValues)
7052 struct rx_debugIn in;
7053 afs_int32 *lp = (afs_int32 *) stat;
7057 * supportedValues is currently unused, but added to allow future
7058 * versioning of this function.
7061 *supportedValues = 0;
7062 in.type = htonl(RX_DEBUGI_RXSTATS);
7064 memset(stat, 0, sizeof(*stat));
7066 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7067 &in, sizeof(in), stat, sizeof(*stat));
7072 * Do net to host conversion here
7075 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7084 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7085 afs_uint16 remotePort, size_t version_length,
7090 return MakeDebugCall(socket, remoteAddr, remotePort,
7091 RX_PACKET_TYPE_VERSION, a, 1, version,
7099 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7100 afs_uint16 remotePort, afs_int32 * nextConnection,
7101 int allConnections, afs_uint32 debugSupportedValues,
7102 struct rx_debugConn * conn,
7103 afs_uint32 * supportedValues)
7109 struct rx_debugIn in;
7113 * supportedValues is currently unused, but added to allow future
7114 * versioning of this function.
7117 *supportedValues = 0;
7118 if (allConnections) {
7119 in.type = htonl(RX_DEBUGI_GETALLCONN);
7121 in.type = htonl(RX_DEBUGI_GETCONN);
7123 in.index = htonl(*nextConnection);
7124 memset(conn, 0, sizeof(*conn));
7126 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7127 &in, sizeof(in), conn, sizeof(*conn));
7130 *nextConnection += 1;
7133 * Convert old connection format to new structure.
7136 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7137 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7138 #define MOVEvL(a) (conn->a = vL->a)
7140 /* any old or unrecognized version... */
7141 for (i = 0; i < RX_MAXCALLS; i++) {
7142 MOVEvL(callState[i]);
7143 MOVEvL(callMode[i]);
7144 MOVEvL(callFlags[i]);
7145 MOVEvL(callOther[i]);
7147 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7148 MOVEvL(secStats.type);
7149 MOVEvL(secStats.level);
7150 MOVEvL(secStats.flags);
7151 MOVEvL(secStats.expires);
7152 MOVEvL(secStats.packetsReceived);
7153 MOVEvL(secStats.packetsSent);
7154 MOVEvL(secStats.bytesReceived);
7155 MOVEvL(secStats.bytesSent);
7160 * Do net to host conversion here
7162 * I don't convert host or port since we are most likely
7163 * going to want these in NBO.
7165 conn->cid = ntohl(conn->cid);
7166 conn->serial = ntohl(conn->serial);
7167 for (i = 0; i < RX_MAXCALLS; i++) {
7168 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7170 conn->error = ntohl(conn->error);
7171 conn->secStats.flags = ntohl(conn->secStats.flags);
7172 conn->secStats.expires = ntohl(conn->secStats.expires);
7173 conn->secStats.packetsReceived =
7174 ntohl(conn->secStats.packetsReceived);
7175 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7176 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7177 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7178 conn->epoch = ntohl(conn->epoch);
7179 conn->natMTU = ntohl(conn->natMTU);
7186 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7187 afs_uint16 remotePort, afs_int32 * nextPeer,
7188 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7189 afs_uint32 * supportedValues)
7195 struct rx_debugIn in;
7198 * supportedValues is currently unused, but added to allow future
7199 * versioning of this function.
7202 *supportedValues = 0;
7203 in.type = htonl(RX_DEBUGI_GETPEER);
7204 in.index = htonl(*nextPeer);
7205 memset(peer, 0, sizeof(*peer));
7207 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7208 &in, sizeof(in), peer, sizeof(*peer));
7214 * Do net to host conversion here
7216 * I don't convert host or port since we are most likely
7217 * going to want these in NBO.
7219 peer->ifMTU = ntohs(peer->ifMTU);
7220 peer->idleWhen = ntohl(peer->idleWhen);
7221 peer->refCount = ntohs(peer->refCount);
7222 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7223 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7224 peer->rtt = ntohl(peer->rtt);
7225 peer->rtt_dev = ntohl(peer->rtt_dev);
7226 peer->timeout.sec = ntohl(peer->timeout.sec);
7227 peer->timeout.usec = ntohl(peer->timeout.usec);
7228 peer->nSent = ntohl(peer->nSent);
7229 peer->reSends = ntohl(peer->reSends);
7230 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7231 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7232 peer->rateFlag = ntohl(peer->rateFlag);
7233 peer->natMTU = ntohs(peer->natMTU);
7234 peer->maxMTU = ntohs(peer->maxMTU);
7235 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7236 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7237 peer->MTU = ntohs(peer->MTU);
7238 peer->cwind = ntohs(peer->cwind);
7239 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7240 peer->congestSeq = ntohs(peer->congestSeq);
7241 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7242 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7243 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7244 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7251 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7252 struct rx_debugPeer * peerStats)
7255 afs_int32 error = 1; /* default to "did not succeed" */
7256 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7258 MUTEX_ENTER(&rx_peerHashTable_lock);
7259 for(tp = rx_peerHashTable[hashValue];
7260 tp != NULL; tp = tp->next) {
7261 if (tp->host == peerHost)
7267 MUTEX_EXIT(&rx_peerHashTable_lock);
7271 MUTEX_ENTER(&tp->peer_lock);
7272 peerStats->host = tp->host;
7273 peerStats->port = tp->port;
7274 peerStats->ifMTU = tp->ifMTU;
7275 peerStats->idleWhen = tp->idleWhen;
7276 peerStats->refCount = tp->refCount;
7277 peerStats->burstSize = tp->burstSize;
7278 peerStats->burst = tp->burst;
7279 peerStats->burstWait.sec = tp->burstWait.sec;
7280 peerStats->burstWait.usec = tp->burstWait.usec;
7281 peerStats->rtt = tp->rtt;
7282 peerStats->rtt_dev = tp->rtt_dev;
7283 peerStats->timeout.sec = tp->timeout.sec;
7284 peerStats->timeout.usec = tp->timeout.usec;
7285 peerStats->nSent = tp->nSent;
7286 peerStats->reSends = tp->reSends;
7287 peerStats->inPacketSkew = tp->inPacketSkew;
7288 peerStats->outPacketSkew = tp->outPacketSkew;
7289 peerStats->rateFlag = tp->rateFlag;
7290 peerStats->natMTU = tp->natMTU;
7291 peerStats->maxMTU = tp->maxMTU;
7292 peerStats->maxDgramPackets = tp->maxDgramPackets;
7293 peerStats->ifDgramPackets = tp->ifDgramPackets;
7294 peerStats->MTU = tp->MTU;
7295 peerStats->cwind = tp->cwind;
7296 peerStats->nDgramPackets = tp->nDgramPackets;
7297 peerStats->congestSeq = tp->congestSeq;
7298 peerStats->bytesSent.high = tp->bytesSent.high;
7299 peerStats->bytesSent.low = tp->bytesSent.low;
7300 peerStats->bytesReceived.high = tp->bytesReceived.high;
7301 peerStats->bytesReceived.low = tp->bytesReceived.low;
7302 MUTEX_EXIT(&tp->peer_lock);
7304 MUTEX_ENTER(&rx_peerHashTable_lock);
7307 MUTEX_EXIT(&rx_peerHashTable_lock);
7315 struct rx_serverQueueEntry *np;
7318 struct rx_call *call;
7319 struct rx_serverQueueEntry *sq;
7323 if (rxinit_status == 1) {
7325 return; /* Already shutdown. */
7329 #ifndef AFS_PTHREAD_ENV
7330 FD_ZERO(&rx_selectMask);
7331 #endif /* AFS_PTHREAD_ENV */
7332 rxi_dataQuota = RX_MAX_QUOTA;
7333 #ifndef AFS_PTHREAD_ENV
7335 #endif /* AFS_PTHREAD_ENV */
7338 #ifndef AFS_PTHREAD_ENV
7339 #ifndef AFS_USE_GETTIMEOFDAY
7341 #endif /* AFS_USE_GETTIMEOFDAY */
7342 #endif /* AFS_PTHREAD_ENV */
7344 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7345 call = queue_First(&rx_freeCallQueue, rx_call);
7347 rxi_Free(call, sizeof(struct rx_call));
7350 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7351 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7357 struct rx_peer **peer_ptr, **peer_end;
7358 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7359 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7361 struct rx_peer *peer, *next;
7363 MUTEX_ENTER(&rx_peerHashTable_lock);
7364 for (peer = *peer_ptr; peer; peer = next) {
7365 rx_interface_stat_p rpc_stat, nrpc_stat;
7368 MUTEX_ENTER(&rx_rpc_stats);
7369 MUTEX_ENTER(&peer->peer_lock);
7371 (&peer->rpcStats, rpc_stat, nrpc_stat,
7372 rx_interface_stat)) {
7373 unsigned int num_funcs;
7376 queue_Remove(&rpc_stat->queue_header);
7377 queue_Remove(&rpc_stat->all_peers);
7378 num_funcs = rpc_stat->stats[0].func_total;
7380 sizeof(rx_interface_stat_t) +
7381 rpc_stat->stats[0].func_total *
7382 sizeof(rx_function_entry_v1_t);
7384 rxi_Free(rpc_stat, space);
7386 /* rx_rpc_stats must be held */
7387 rxi_rpc_peer_stat_cnt -= num_funcs;
7389 MUTEX_EXIT(&peer->peer_lock);
7390 MUTEX_EXIT(&rx_rpc_stats);
7394 if (rx_stats_active)
7395 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
7397 MUTEX_EXIT(&rx_peerHashTable_lock);
7400 for (i = 0; i < RX_MAX_SERVICES; i++) {
7402 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7404 for (i = 0; i < rx_hashTableSize; i++) {
7405 struct rx_connection *tc, *ntc;
7406 MUTEX_ENTER(&rx_connHashTable_lock);
7407 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7409 for (j = 0; j < RX_MAXCALLS; j++) {
7411 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7414 rxi_Free(tc, sizeof(*tc));
7416 MUTEX_EXIT(&rx_connHashTable_lock);
7419 MUTEX_ENTER(&freeSQEList_lock);
7421 while ((np = rx_FreeSQEList)) {
7422 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7423 MUTEX_DESTROY(&np->lock);
7424 rxi_Free(np, sizeof(*np));
7427 MUTEX_EXIT(&freeSQEList_lock);
7428 MUTEX_DESTROY(&freeSQEList_lock);
7429 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7430 MUTEX_DESTROY(&rx_connHashTable_lock);
7431 MUTEX_DESTROY(&rx_peerHashTable_lock);
7432 MUTEX_DESTROY(&rx_serverPool_lock);
7434 osi_Free(rx_connHashTable,
7435 rx_hashTableSize * sizeof(struct rx_connection *));
7436 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7438 UNPIN(rx_connHashTable,
7439 rx_hashTableSize * sizeof(struct rx_connection *));
7440 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7442 rxi_FreeAllPackets();
7444 MUTEX_ENTER(&rx_quota_mutex);
7445 rxi_dataQuota = RX_MAX_QUOTA;
7446 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7447 MUTEX_EXIT(&rx_quota_mutex);
7452 #ifdef RX_ENABLE_LOCKS
7454 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7456 if (!MUTEX_ISMINE(lockaddr))
7457 osi_Panic("Lock not held: %s", msg);
7459 #endif /* RX_ENABLE_LOCKS */
7464 * Routines to implement connection specific data.
7468 rx_KeyCreate(rx_destructor_t rtn)
7471 MUTEX_ENTER(&rxi_keyCreate_lock);
7472 key = rxi_keyCreate_counter++;
7473 rxi_keyCreate_destructor = (rx_destructor_t *)
7474 realloc((void *)rxi_keyCreate_destructor,
7475 (key + 1) * sizeof(rx_destructor_t));
7476 rxi_keyCreate_destructor[key] = rtn;
7477 MUTEX_EXIT(&rxi_keyCreate_lock);
7482 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7485 MUTEX_ENTER(&conn->conn_data_lock);
7486 if (!conn->specific) {
7487 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7488 for (i = 0; i < key; i++)
7489 conn->specific[i] = NULL;
7490 conn->nSpecific = key + 1;
7491 conn->specific[key] = ptr;
7492 } else if (key >= conn->nSpecific) {
7493 conn->specific = (void **)
7494 realloc(conn->specific, (key + 1) * sizeof(void *));
7495 for (i = conn->nSpecific; i < key; i++)
7496 conn->specific[i] = NULL;
7497 conn->nSpecific = key + 1;
7498 conn->specific[key] = ptr;
7500 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7501 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7502 conn->specific[key] = ptr;
7504 MUTEX_EXIT(&conn->conn_data_lock);
7508 rx_GetSpecific(struct rx_connection *conn, int key)
7511 MUTEX_ENTER(&conn->conn_data_lock);
7512 if (key >= conn->nSpecific)
7515 ptr = conn->specific[key];
7516 MUTEX_EXIT(&conn->conn_data_lock);
7520 #endif /* !KERNEL */
7523 * processStats is a queue used to store the statistics for the local
7524 * process. Its contents are similar to the contents of the rpcStats
7525 * queue on a rx_peer structure, but the actual data stored within
7526 * this queue contains totals across the lifetime of the process (assuming
7527 * the stats have not been reset) - unlike the per peer structures
7528 * which can come and go based upon the peer lifetime.
7531 static struct rx_queue processStats = { &processStats, &processStats };
7534 * peerStats is a queue used to store the statistics for all peer structs.
7535 * Its contents are the union of all the peer rpcStats queues.
7538 static struct rx_queue peerStats = { &peerStats, &peerStats };
7541 * rxi_monitor_processStats is used to turn process wide stat collection
7545 static int rxi_monitor_processStats = 0;
7548 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7551 static int rxi_monitor_peerStats = 0;
7554 * rxi_AddRpcStat - given all of the information for a particular rpc
7555 * call, create (if needed) and update the stat totals for the rpc.
7559 * IN stats - the queue of stats that will be updated with the new value
7561 * IN rxInterface - a unique number that identifies the rpc interface
7563 * IN currentFunc - the index of the function being invoked
7565 * IN totalFunc - the total number of functions in this interface
7567 * IN queueTime - the amount of time this function waited for a thread
7569 * IN execTime - the amount of time this function invocation took to execute
7571 * IN bytesSent - the number bytes sent by this invocation
7573 * IN bytesRcvd - the number bytes received by this invocation
7575 * IN isServer - if true, this invocation was made to a server
7577 * IN remoteHost - the ip address of the remote host
7579 * IN remotePort - the port of the remote host
7581 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7583 * INOUT counter - if a new stats structure is allocated, the counter will
7584 * be updated with the new number of allocated stat structures
7592 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
7593 afs_uint32 currentFunc, afs_uint32 totalFunc,
7594 struct clock *queueTime, struct clock *execTime,
7595 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
7596 afs_uint32 remoteHost, afs_uint32 remotePort,
7597 int addToPeerList, unsigned int *counter)
7600 rx_interface_stat_p rpc_stat, nrpc_stat;
7603 * See if there's already a structure for this interface
7606 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7607 if ((rpc_stat->stats[0].interfaceId == rxInterface)
7608 && (rpc_stat->stats[0].remote_is_server == isServer))
7613 * Didn't find a match so allocate a new structure and add it to the
7617 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
7618 || (rpc_stat->stats[0].interfaceId != rxInterface)
7619 || (rpc_stat->stats[0].remote_is_server != isServer)) {
7624 sizeof(rx_interface_stat_t) +
7625 totalFunc * sizeof(rx_function_entry_v1_t);
7627 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
7628 if (rpc_stat == NULL) {
7632 *counter += totalFunc;
7633 for (i = 0; i < totalFunc; i++) {
7634 rpc_stat->stats[i].remote_peer = remoteHost;
7635 rpc_stat->stats[i].remote_port = remotePort;
7636 rpc_stat->stats[i].remote_is_server = isServer;
7637 rpc_stat->stats[i].interfaceId = rxInterface;
7638 rpc_stat->stats[i].func_total = totalFunc;
7639 rpc_stat->stats[i].func_index = i;
7640 hzero(rpc_stat->stats[i].invocations);
7641 hzero(rpc_stat->stats[i].bytes_sent);
7642 hzero(rpc_stat->stats[i].bytes_rcvd);
7643 rpc_stat->stats[i].queue_time_sum.sec = 0;
7644 rpc_stat->stats[i].queue_time_sum.usec = 0;
7645 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7646 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7647 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7648 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7649 rpc_stat->stats[i].queue_time_max.sec = 0;
7650 rpc_stat->stats[i].queue_time_max.usec = 0;
7651 rpc_stat->stats[i].execution_time_sum.sec = 0;
7652 rpc_stat->stats[i].execution_time_sum.usec = 0;
7653 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7654 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7655 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7656 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7657 rpc_stat->stats[i].execution_time_max.sec = 0;
7658 rpc_stat->stats[i].execution_time_max.usec = 0;
7660 queue_Prepend(stats, rpc_stat);
7661 if (addToPeerList) {
7662 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7667 * Increment the stats for this function
7670 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7671 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7672 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7673 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7674 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7675 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7676 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7678 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7679 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7681 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7682 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7684 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7685 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7687 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7688 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7696 * rx_IncrementTimeAndCount - increment the times and count for a particular
7701 * IN peer - the peer who invoked the rpc
7703 * IN rxInterface - a unique number that identifies the rpc interface
7705 * IN currentFunc - the index of the function being invoked
7707 * IN totalFunc - the total number of functions in this interface
7709 * IN queueTime - the amount of time this function waited for a thread
7711 * IN execTime - the amount of time this function invocation took to execute
7713 * IN bytesSent - the number bytes sent by this invocation
7715 * IN bytesRcvd - the number bytes received by this invocation
7717 * IN isServer - if true, this invocation was made to a server
7725 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7726 afs_uint32 currentFunc, afs_uint32 totalFunc,
7727 struct clock *queueTime, struct clock *execTime,
7728 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7732 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7735 MUTEX_ENTER(&rx_rpc_stats);
7737 if (rxi_monitor_peerStats) {
7738 MUTEX_ENTER(&peer->peer_lock);
7739 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7740 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7741 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7742 MUTEX_EXIT(&peer->peer_lock);
7745 if (rxi_monitor_processStats) {
7746 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7747 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7748 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7751 MUTEX_EXIT(&rx_rpc_stats);
7756 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7760 * IN callerVersion - the rpc stat version of the caller.
7762 * IN count - the number of entries to marshall.
7764 * IN stats - pointer to stats to be marshalled.
7766 * OUT ptr - Where to store the marshalled data.
7773 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7774 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7780 * We only support the first version
7782 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7783 *(ptr++) = stats->remote_peer;
7784 *(ptr++) = stats->remote_port;
7785 *(ptr++) = stats->remote_is_server;
7786 *(ptr++) = stats->interfaceId;
7787 *(ptr++) = stats->func_total;
7788 *(ptr++) = stats->func_index;
7789 *(ptr++) = hgethi(stats->invocations);
7790 *(ptr++) = hgetlo(stats->invocations);
7791 *(ptr++) = hgethi(stats->bytes_sent);
7792 *(ptr++) = hgetlo(stats->bytes_sent);
7793 *(ptr++) = hgethi(stats->bytes_rcvd);
7794 *(ptr++) = hgetlo(stats->bytes_rcvd);
7795 *(ptr++) = stats->queue_time_sum.sec;
7796 *(ptr++) = stats->queue_time_sum.usec;
7797 *(ptr++) = stats->queue_time_sum_sqr.sec;
7798 *(ptr++) = stats->queue_time_sum_sqr.usec;
7799 *(ptr++) = stats->queue_time_min.sec;
7800 *(ptr++) = stats->queue_time_min.usec;
7801 *(ptr++) = stats->queue_time_max.sec;
7802 *(ptr++) = stats->queue_time_max.usec;
7803 *(ptr++) = stats->execution_time_sum.sec;
7804 *(ptr++) = stats->execution_time_sum.usec;
7805 *(ptr++) = stats->execution_time_sum_sqr.sec;
7806 *(ptr++) = stats->execution_time_sum_sqr.usec;
7807 *(ptr++) = stats->execution_time_min.sec;
7808 *(ptr++) = stats->execution_time_min.usec;
7809 *(ptr++) = stats->execution_time_max.sec;
7810 *(ptr++) = stats->execution_time_max.usec;
7816 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7821 * IN callerVersion - the rpc stat version of the caller
7823 * OUT myVersion - the rpc stat version of this function
7825 * OUT clock_sec - local time seconds
7827 * OUT clock_usec - local time microseconds
7829 * OUT allocSize - the number of bytes allocated to contain stats
7831 * OUT statCount - the number stats retrieved from this process.
7833 * OUT stats - the actual stats retrieved from this process.
7837 * Returns void. If successful, stats will != NULL.
7841 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7842 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7843 size_t * allocSize, afs_uint32 * statCount,
7844 afs_uint32 ** stats)
7854 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7857 * Check to see if stats are enabled
7860 MUTEX_ENTER(&rx_rpc_stats);
7861 if (!rxi_monitor_processStats) {
7862 MUTEX_EXIT(&rx_rpc_stats);
7866 clock_GetTime(&now);
7867 *clock_sec = now.sec;
7868 *clock_usec = now.usec;
7871 * Allocate the space based upon the caller version
7873 * If the client is at an older version than we are,
7874 * we return the statistic data in the older data format, but
7875 * we still return our version number so the client knows we
7876 * are maintaining more data than it can retrieve.
7879 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7880 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7881 *statCount = rxi_rpc_process_stat_cnt;
7884 * This can't happen yet, but in the future version changes
7885 * can be handled by adding additional code here
7889 if (space > (size_t) 0) {
7891 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7894 rx_interface_stat_p rpc_stat, nrpc_stat;
7898 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7900 * Copy the data based upon the caller version
7902 rx_MarshallProcessRPCStats(callerVersion,
7903 rpc_stat->stats[0].func_total,
7904 rpc_stat->stats, &ptr);
7910 MUTEX_EXIT(&rx_rpc_stats);
7915 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7919 * IN callerVersion - the rpc stat version of the caller
7921 * OUT myVersion - the rpc stat version of this function
7923 * OUT clock_sec - local time seconds
7925 * OUT clock_usec - local time microseconds
7927 * OUT allocSize - the number of bytes allocated to contain stats
7929 * OUT statCount - the number of stats retrieved from the individual
7932 * OUT stats - the actual stats retrieved from the individual peer structures.
7936 * Returns void. If successful, stats will != NULL.
7940 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7941 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7942 size_t * allocSize, afs_uint32 * statCount,
7943 afs_uint32 ** stats)
7953 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7956 * Check to see if stats are enabled
7959 MUTEX_ENTER(&rx_rpc_stats);
7960 if (!rxi_monitor_peerStats) {
7961 MUTEX_EXIT(&rx_rpc_stats);
7965 clock_GetTime(&now);
7966 *clock_sec = now.sec;
7967 *clock_usec = now.usec;
7970 * Allocate the space based upon the caller version
7972 * If the client is at an older version than we are,
7973 * we return the statistic data in the older data format, but
7974 * we still return our version number so the client knows we
7975 * are maintaining more data than it can retrieve.
7978 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7979 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7980 *statCount = rxi_rpc_peer_stat_cnt;
7983 * This can't happen yet, but in the future version changes
7984 * can be handled by adding additional code here
7988 if (space > (size_t) 0) {
7990 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7993 rx_interface_stat_p rpc_stat, nrpc_stat;
7997 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7999 * We have to fix the offset of rpc_stat since we are
8000 * keeping this structure on two rx_queues. The rx_queue
8001 * package assumes that the rx_queue member is the first
8002 * member of the structure. That is, rx_queue assumes that
8003 * any one item is only on one queue at a time. We are
8004 * breaking that assumption and so we have to do a little
8005 * math to fix our pointers.
8008 fix_offset = (char *)rpc_stat;
8009 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8010 rpc_stat = (rx_interface_stat_p) fix_offset;
8013 * Copy the data based upon the caller version
8015 rx_MarshallProcessRPCStats(callerVersion,
8016 rpc_stat->stats[0].func_total,
8017 rpc_stat->stats, &ptr);
8023 MUTEX_EXIT(&rx_rpc_stats);
8028 * rx_FreeRPCStats - free memory allocated by
8029 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8033 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8034 * rx_RetrievePeerRPCStats
8036 * IN allocSize - the number of bytes in stats.
8044 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8046 rxi_Free(stats, allocSize);
8050 * rx_queryProcessRPCStats - see if process rpc stat collection is
8051 * currently enabled.
8057 * Returns 0 if stats are not enabled != 0 otherwise
8061 rx_queryProcessRPCStats(void)
8064 MUTEX_ENTER(&rx_rpc_stats);
8065 rc = rxi_monitor_processStats;
8066 MUTEX_EXIT(&rx_rpc_stats);
8071 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8077 * Returns 0 if stats are not enabled != 0 otherwise
8081 rx_queryPeerRPCStats(void)
8084 MUTEX_ENTER(&rx_rpc_stats);
8085 rc = rxi_monitor_peerStats;
8086 MUTEX_EXIT(&rx_rpc_stats);
8091 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8101 rx_enableProcessRPCStats(void)
8103 MUTEX_ENTER(&rx_rpc_stats);
8104 rx_enable_stats = 1;
8105 rxi_monitor_processStats = 1;
8106 MUTEX_EXIT(&rx_rpc_stats);
8110 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8120 rx_enablePeerRPCStats(void)
8122 MUTEX_ENTER(&rx_rpc_stats);
8123 rx_enable_stats = 1;
8124 rxi_monitor_peerStats = 1;
8125 MUTEX_EXIT(&rx_rpc_stats);
8129 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8139 rx_disableProcessRPCStats(void)
8141 rx_interface_stat_p rpc_stat, nrpc_stat;
8144 MUTEX_ENTER(&rx_rpc_stats);
8147 * Turn off process statistics and if peer stats is also off, turn
8151 rxi_monitor_processStats = 0;
8152 if (rxi_monitor_peerStats == 0) {
8153 rx_enable_stats = 0;
8156 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8157 unsigned int num_funcs = 0;
8160 queue_Remove(rpc_stat);
8161 num_funcs = rpc_stat->stats[0].func_total;
8163 sizeof(rx_interface_stat_t) +
8164 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8166 rxi_Free(rpc_stat, space);
8167 rxi_rpc_process_stat_cnt -= num_funcs;
8169 MUTEX_EXIT(&rx_rpc_stats);
8173 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8183 rx_disablePeerRPCStats(void)
8185 struct rx_peer **peer_ptr, **peer_end;
8189 * Turn off peer statistics and if process stats is also off, turn
8193 rxi_monitor_peerStats = 0;
8194 if (rxi_monitor_processStats == 0) {
8195 rx_enable_stats = 0;
8198 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8199 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8201 struct rx_peer *peer, *next, *prev;
8203 MUTEX_ENTER(&rx_peerHashTable_lock);
8204 MUTEX_ENTER(&rx_rpc_stats);
8205 for (prev = peer = *peer_ptr; peer; peer = next) {
8207 code = MUTEX_TRYENTER(&peer->peer_lock);
8209 rx_interface_stat_p rpc_stat, nrpc_stat;
8212 if (prev == *peer_ptr) {
8223 MUTEX_EXIT(&rx_peerHashTable_lock);
8226 (&peer->rpcStats, rpc_stat, nrpc_stat,
8227 rx_interface_stat)) {
8228 unsigned int num_funcs = 0;
8231 queue_Remove(&rpc_stat->queue_header);
8232 queue_Remove(&rpc_stat->all_peers);
8233 num_funcs = rpc_stat->stats[0].func_total;
8235 sizeof(rx_interface_stat_t) +
8236 rpc_stat->stats[0].func_total *
8237 sizeof(rx_function_entry_v1_t);
8239 rxi_Free(rpc_stat, space);
8240 rxi_rpc_peer_stat_cnt -= num_funcs;
8242 MUTEX_EXIT(&peer->peer_lock);
8244 MUTEX_ENTER(&rx_peerHashTable_lock);
8254 MUTEX_EXIT(&rx_rpc_stats);
8255 MUTEX_EXIT(&rx_peerHashTable_lock);
8260 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8265 * IN clearFlag - flag indicating which stats to clear
8273 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8275 rx_interface_stat_p rpc_stat, nrpc_stat;
8277 MUTEX_ENTER(&rx_rpc_stats);
8279 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8280 unsigned int num_funcs = 0, i;
8281 num_funcs = rpc_stat->stats[0].func_total;
8282 for (i = 0; i < num_funcs; i++) {
8283 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8284 hzero(rpc_stat->stats[i].invocations);
8286 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8287 hzero(rpc_stat->stats[i].bytes_sent);
8289 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8290 hzero(rpc_stat->stats[i].bytes_rcvd);
8292 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8293 rpc_stat->stats[i].queue_time_sum.sec = 0;
8294 rpc_stat->stats[i].queue_time_sum.usec = 0;
8296 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8297 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8298 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8300 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8301 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8302 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8304 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8305 rpc_stat->stats[i].queue_time_max.sec = 0;
8306 rpc_stat->stats[i].queue_time_max.usec = 0;
8308 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8309 rpc_stat->stats[i].execution_time_sum.sec = 0;
8310 rpc_stat->stats[i].execution_time_sum.usec = 0;
8312 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8313 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8314 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8316 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8317 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8318 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8320 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8321 rpc_stat->stats[i].execution_time_max.sec = 0;
8322 rpc_stat->stats[i].execution_time_max.usec = 0;
8327 MUTEX_EXIT(&rx_rpc_stats);
8331 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8336 * IN clearFlag - flag indicating which stats to clear
8344 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8346 rx_interface_stat_p rpc_stat, nrpc_stat;
8348 MUTEX_ENTER(&rx_rpc_stats);
8350 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8351 unsigned int num_funcs = 0, i;
8354 * We have to fix the offset of rpc_stat since we are
8355 * keeping this structure on two rx_queues. The rx_queue
8356 * package assumes that the rx_queue member is the first
8357 * member of the structure. That is, rx_queue assumes that
8358 * any one item is only on one queue at a time. We are
8359 * breaking that assumption and so we have to do a little
8360 * math to fix our pointers.
8363 fix_offset = (char *)rpc_stat;
8364 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8365 rpc_stat = (rx_interface_stat_p) fix_offset;
8367 num_funcs = rpc_stat->stats[0].func_total;
8368 for (i = 0; i < num_funcs; i++) {
8369 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8370 hzero(rpc_stat->stats[i].invocations);
8372 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8373 hzero(rpc_stat->stats[i].bytes_sent);
8375 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8376 hzero(rpc_stat->stats[i].bytes_rcvd);
8378 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8379 rpc_stat->stats[i].queue_time_sum.sec = 0;
8380 rpc_stat->stats[i].queue_time_sum.usec = 0;
8382 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8383 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8384 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8386 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8387 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8388 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8390 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8391 rpc_stat->stats[i].queue_time_max.sec = 0;
8392 rpc_stat->stats[i].queue_time_max.usec = 0;
8394 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8395 rpc_stat->stats[i].execution_time_sum.sec = 0;
8396 rpc_stat->stats[i].execution_time_sum.usec = 0;
8398 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8399 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8400 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8402 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8403 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8404 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8406 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8407 rpc_stat->stats[i].execution_time_max.sec = 0;
8408 rpc_stat->stats[i].execution_time_max.usec = 0;
8413 MUTEX_EXIT(&rx_rpc_stats);
8417 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8418 * is authorized to enable/disable/clear RX statistics.
8420 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8423 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8425 rxi_rxstat_userok = proc;
8429 rx_RxStatUserOk(struct rx_call *call)
8431 if (!rxi_rxstat_userok)
8433 return rxi_rxstat_userok(call);
8438 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8439 * function in the MSVC runtime DLL (msvcrt.dll).
8441 * Note: the system serializes calls to this function.
8444 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8445 DWORD reason, /* reason function is being called */
8446 LPVOID reserved) /* reserved for future use */
8449 case DLL_PROCESS_ATTACH:
8450 /* library is being attached to a process */
8454 case DLL_PROCESS_DETACH:
8461 #endif /* AFS_NT40_ENV */
8464 int rx_DumpCalls(FILE *outputFile, char *cookie)
8466 #ifdef RXDEBUG_PACKET
8467 #ifdef KDUMP_RX_LOCK
8468 struct rx_call_rx_lock *c;
8475 #define RXDPRINTF sprintf
8476 #define RXDPRINTOUT output
8478 #define RXDPRINTF fprintf
8479 #define RXDPRINTOUT outputFile
8482 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8484 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8487 for (c = rx_allCallsp; c; c = c->allNextp) {
8488 u_short rqc, tqc, iovqc;
8489 struct rx_packet *p, *np;
8491 MUTEX_ENTER(&c->lock);
8492 queue_Count(&c->rq, p, np, rx_packet, rqc);
8493 queue_Count(&c->tq, p, np, rx_packet, tqc);
8494 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8496 RXDPRINTF(RXDPRINTOUT, "%s - call=0x%p, id=%u, state=%u, mode=%u, conn=%p, epoch=%u, cid=%u, callNum=%u, connFlags=0x%x, flags=0x%x, "
8497 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8498 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8499 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8500 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8501 #ifdef RX_ENABLE_LOCKS
8504 #ifdef RX_REFCOUNT_CHECK
8505 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
8506 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
8509 cookie, c, c->call_id, (afs_uint32)c->state, (afs_uint32)c->mode, c->conn, c->conn?c->conn->epoch:0, c->conn?c->conn->cid:0,
8510 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
8511 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
8512 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
8513 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
8514 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
8515 #ifdef RX_ENABLE_LOCKS
8516 , (afs_uint32)c->refCount
8518 #ifdef RX_REFCOUNT_CHECK
8519 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
8522 MUTEX_EXIT(&c->lock);
8525 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8528 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
8530 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8532 #endif /* RXDEBUG_PACKET */