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();
1434 #ifdef RX_ENABLE_LOCKS
1435 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1438 for (i = 0; i < RX_MAX_SERVICES; i++) {
1439 struct rx_service *service = rx_services[i];
1441 if (port == service->servicePort && host == service->serviceHost) {
1442 if (service->serviceId == serviceId) {
1443 /* The identical service has already been
1444 * installed; if the caller was intending to
1445 * change the security classes used by this
1446 * service, he/she loses. */
1448 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1449 serviceName, serviceId, service->serviceName);
1451 rxi_FreeService(tservice);
1454 /* Different service, same port: re-use the socket
1455 * which is bound to the same port */
1456 socket = service->socket;
1459 if (socket == OSI_NULLSOCKET) {
1460 /* If we don't already have a socket (from another
1461 * service on same port) get a new one */
1462 socket = rxi_GetHostUDPSocket(host, port);
1463 if (socket == OSI_NULLSOCKET) {
1465 rxi_FreeService(tservice);
1470 service->socket = socket;
1471 service->serviceHost = host;
1472 service->servicePort = port;
1473 service->serviceId = serviceId;
1474 service->serviceName = serviceName;
1475 service->nSecurityObjects = nSecurityObjects;
1476 service->securityObjects = securityObjects;
1477 service->minProcs = 0;
1478 service->maxProcs = 1;
1479 service->idleDeadTime = 60;
1480 service->idleDeadErr = 0;
1481 service->connDeadTime = rx_connDeadTime;
1482 service->executeRequestProc = serviceProc;
1483 service->checkReach = 0;
1484 service->nSpecific = 0;
1485 service->specific = NULL;
1486 rx_services[i] = service; /* not visible until now */
1492 rxi_FreeService(tservice);
1493 (osi_Msg "rx_NewService: cannot support > %d services\n",
1498 /* Set configuration options for all of a service's security objects */
1501 rx_SetSecurityConfiguration(struct rx_service *service,
1502 rx_securityConfigVariables type,
1506 for (i = 0; i<service->nSecurityObjects; i++) {
1507 if (service->securityObjects[i]) {
1508 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1516 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1517 struct rx_securityClass **securityObjects, int nSecurityObjects,
1518 afs_int32(*serviceProc) (struct rx_call * acall))
1520 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1523 /* Generic request processing loop. This routine should be called
1524 * by the implementation dependent rx_ServerProc. If socketp is
1525 * non-null, it will be set to the file descriptor that this thread
1526 * is now listening on. If socketp is null, this routine will never
1529 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1531 struct rx_call *call;
1533 struct rx_service *tservice = NULL;
1540 call = rx_GetCall(threadID, tservice, socketp);
1541 if (socketp && *socketp != OSI_NULLSOCKET) {
1542 /* We are now a listener thread */
1547 /* if server is restarting( typically smooth shutdown) then do not
1548 * allow any new calls.
1551 if (rx_tranquil && (call != NULL)) {
1555 MUTEX_ENTER(&call->lock);
1557 rxi_CallError(call, RX_RESTARTING);
1558 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1560 MUTEX_EXIT(&call->lock);
1564 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1565 #ifdef RX_ENABLE_LOCKS
1567 #endif /* RX_ENABLE_LOCKS */
1568 afs_termState = AFSOP_STOP_AFS;
1569 afs_osi_Wakeup(&afs_termState);
1570 #ifdef RX_ENABLE_LOCKS
1572 #endif /* RX_ENABLE_LOCKS */
1577 tservice = call->conn->service;
1579 if (tservice->beforeProc)
1580 (*tservice->beforeProc) (call);
1582 code = call->conn->service->executeRequestProc(call);
1584 if (tservice->afterProc)
1585 (*tservice->afterProc) (call, code);
1587 rx_EndCall(call, code);
1588 if (rx_stats_active) {
1589 MUTEX_ENTER(&rx_stats_mutex);
1591 MUTEX_EXIT(&rx_stats_mutex);
1598 rx_WakeupServerProcs(void)
1600 struct rx_serverQueueEntry *np, *tqp;
1604 MUTEX_ENTER(&rx_serverPool_lock);
1606 #ifdef RX_ENABLE_LOCKS
1607 if (rx_waitForPacket)
1608 CV_BROADCAST(&rx_waitForPacket->cv);
1609 #else /* RX_ENABLE_LOCKS */
1610 if (rx_waitForPacket)
1611 osi_rxWakeup(rx_waitForPacket);
1612 #endif /* RX_ENABLE_LOCKS */
1613 MUTEX_ENTER(&freeSQEList_lock);
1614 for (np = rx_FreeSQEList; np; np = tqp) {
1615 tqp = *(struct rx_serverQueueEntry **)np;
1616 #ifdef RX_ENABLE_LOCKS
1617 CV_BROADCAST(&np->cv);
1618 #else /* RX_ENABLE_LOCKS */
1620 #endif /* RX_ENABLE_LOCKS */
1622 MUTEX_EXIT(&freeSQEList_lock);
1623 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1624 #ifdef RX_ENABLE_LOCKS
1625 CV_BROADCAST(&np->cv);
1626 #else /* RX_ENABLE_LOCKS */
1628 #endif /* RX_ENABLE_LOCKS */
1630 MUTEX_EXIT(&rx_serverPool_lock);
1635 * One thing that seems to happen is that all the server threads get
1636 * tied up on some empty or slow call, and then a whole bunch of calls
1637 * arrive at once, using up the packet pool, so now there are more
1638 * empty calls. The most critical resources here are server threads
1639 * and the free packet pool. The "doreclaim" code seems to help in
1640 * general. I think that eventually we arrive in this state: there
1641 * are lots of pending calls which do have all their packets present,
1642 * so they won't be reclaimed, are multi-packet calls, so they won't
1643 * be scheduled until later, and thus are tying up most of the free
1644 * packet pool for a very long time.
1646 * 1. schedule multi-packet calls if all the packets are present.
1647 * Probably CPU-bound operation, useful to return packets to pool.
1648 * Do what if there is a full window, but the last packet isn't here?
1649 * 3. preserve one thread which *only* runs "best" calls, otherwise
1650 * it sleeps and waits for that type of call.
1651 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1652 * the current dataquota business is badly broken. The quota isn't adjusted
1653 * to reflect how many packets are presently queued for a running call.
1654 * So, when we schedule a queued call with a full window of packets queued
1655 * up for it, that *should* free up a window full of packets for other 2d-class
1656 * calls to be able to use from the packet pool. But it doesn't.
1658 * NB. Most of the time, this code doesn't run -- since idle server threads
1659 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1660 * as a new call arrives.
1662 /* Sleep until a call arrives. Returns a pointer to the call, ready
1663 * for an rx_Read. */
1664 #ifdef RX_ENABLE_LOCKS
1666 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1668 struct rx_serverQueueEntry *sq;
1669 struct rx_call *call = (struct rx_call *)0;
1670 struct rx_service *service = NULL;
1673 MUTEX_ENTER(&freeSQEList_lock);
1675 if ((sq = rx_FreeSQEList)) {
1676 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1677 MUTEX_EXIT(&freeSQEList_lock);
1678 } else { /* otherwise allocate a new one and return that */
1679 MUTEX_EXIT(&freeSQEList_lock);
1680 sq = (struct rx_serverQueueEntry *)
1681 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1682 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1683 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1686 MUTEX_ENTER(&rx_serverPool_lock);
1687 if (cur_service != NULL) {
1688 ReturnToServerPool(cur_service);
1691 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1692 struct rx_call *tcall, *ncall, *choice2 = NULL;
1694 /* Scan for eligible incoming calls. A call is not eligible
1695 * if the maximum number of calls for its service type are
1696 * already executing */
1697 /* One thread will process calls FCFS (to prevent starvation),
1698 * while the other threads may run ahead looking for calls which
1699 * have all their input data available immediately. This helps
1700 * keep threads from blocking, waiting for data from the client. */
1701 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1702 service = tcall->conn->service;
1703 if (!QuotaOK(service)) {
1706 MUTEX_ENTER(&rx_pthread_mutex);
1707 if (tno == rxi_fcfs_thread_num
1708 || !tcall->queue_item_header.next) {
1709 MUTEX_EXIT(&rx_pthread_mutex);
1710 /* If we're the fcfs thread , then we'll just use
1711 * this call. If we haven't been able to find an optimal
1712 * choice, and we're at the end of the list, then use a
1713 * 2d choice if one has been identified. Otherwise... */
1714 call = (choice2 ? choice2 : tcall);
1715 service = call->conn->service;
1717 MUTEX_EXIT(&rx_pthread_mutex);
1718 if (!queue_IsEmpty(&tcall->rq)) {
1719 struct rx_packet *rp;
1720 rp = queue_First(&tcall->rq, rx_packet);
1721 if (rp->header.seq == 1) {
1723 || (rp->header.flags & RX_LAST_PACKET)) {
1725 } else if (rxi_2dchoice && !choice2
1726 && !(tcall->flags & RX_CALL_CLEARED)
1727 && (tcall->rprev > rxi_HardAckRate)) {
1737 ReturnToServerPool(service);
1744 MUTEX_EXIT(&rx_serverPool_lock);
1745 MUTEX_ENTER(&call->lock);
1747 if (call->flags & RX_CALL_WAIT_PROC) {
1748 call->flags &= ~RX_CALL_WAIT_PROC;
1749 MUTEX_ENTER(&rx_waiting_mutex);
1751 MUTEX_EXIT(&rx_waiting_mutex);
1754 if (call->state != RX_STATE_PRECALL || call->error) {
1755 MUTEX_EXIT(&call->lock);
1756 MUTEX_ENTER(&rx_serverPool_lock);
1757 ReturnToServerPool(service);
1762 if (queue_IsEmpty(&call->rq)
1763 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1764 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1766 CLEAR_CALL_QUEUE_LOCK(call);
1769 /* If there are no eligible incoming calls, add this process
1770 * to the idle server queue, to wait for one */
1774 *socketp = OSI_NULLSOCKET;
1776 sq->socketp = socketp;
1777 queue_Append(&rx_idleServerQueue, sq);
1778 #ifndef AFS_AIX41_ENV
1779 rx_waitForPacket = sq;
1781 rx_waitingForPacket = sq;
1782 #endif /* AFS_AIX41_ENV */
1784 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1786 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1787 MUTEX_EXIT(&rx_serverPool_lock);
1788 return (struct rx_call *)0;
1791 } while (!(call = sq->newcall)
1792 && !(socketp && *socketp != OSI_NULLSOCKET));
1793 MUTEX_EXIT(&rx_serverPool_lock);
1795 MUTEX_ENTER(&call->lock);
1801 MUTEX_ENTER(&freeSQEList_lock);
1802 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1803 rx_FreeSQEList = sq;
1804 MUTEX_EXIT(&freeSQEList_lock);
1807 clock_GetTime(&call->startTime);
1808 call->state = RX_STATE_ACTIVE;
1809 call->mode = RX_MODE_RECEIVING;
1810 #ifdef RX_KERNEL_TRACE
1811 if (ICL_SETACTIVE(afs_iclSetp)) {
1812 int glockOwner = ISAFS_GLOCK();
1815 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1816 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1823 rxi_calltrace(RX_CALL_START, call);
1824 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
1825 call->conn->service->servicePort, call->conn->service->serviceId,
1828 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1829 MUTEX_EXIT(&call->lock);
1831 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1836 #else /* RX_ENABLE_LOCKS */
1838 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1840 struct rx_serverQueueEntry *sq;
1841 struct rx_call *call = (struct rx_call *)0, *choice2;
1842 struct rx_service *service = NULL;
1846 MUTEX_ENTER(&freeSQEList_lock);
1848 if ((sq = rx_FreeSQEList)) {
1849 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1850 MUTEX_EXIT(&freeSQEList_lock);
1851 } else { /* otherwise allocate a new one and return that */
1852 MUTEX_EXIT(&freeSQEList_lock);
1853 sq = (struct rx_serverQueueEntry *)
1854 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1855 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1856 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1858 MUTEX_ENTER(&sq->lock);
1860 if (cur_service != NULL) {
1861 cur_service->nRequestsRunning--;
1862 MUTEX_ENTER(&rx_quota_mutex);
1863 if (cur_service->nRequestsRunning < cur_service->minProcs)
1866 MUTEX_EXIT(&rx_quota_mutex);
1868 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1869 struct rx_call *tcall, *ncall;
1870 /* Scan for eligible incoming calls. A call is not eligible
1871 * if the maximum number of calls for its service type are
1872 * already executing */
1873 /* One thread will process calls FCFS (to prevent starvation),
1874 * while the other threads may run ahead looking for calls which
1875 * have all their input data available immediately. This helps
1876 * keep threads from blocking, waiting for data from the client. */
1877 choice2 = (struct rx_call *)0;
1878 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1879 service = tcall->conn->service;
1880 if (QuotaOK(service)) {
1881 MUTEX_ENTER(&rx_pthread_mutex);
1882 if (tno == rxi_fcfs_thread_num
1883 || !tcall->queue_item_header.next) {
1884 MUTEX_EXIT(&rx_pthread_mutex);
1885 /* If we're the fcfs thread, then we'll just use
1886 * this call. If we haven't been able to find an optimal
1887 * choice, and we're at the end of the list, then use a
1888 * 2d choice if one has been identified. Otherwise... */
1889 call = (choice2 ? choice2 : tcall);
1890 service = call->conn->service;
1892 MUTEX_EXIT(&rx_pthread_mutex);
1893 if (!queue_IsEmpty(&tcall->rq)) {
1894 struct rx_packet *rp;
1895 rp = queue_First(&tcall->rq, rx_packet);
1896 if (rp->header.seq == 1
1898 || (rp->header.flags & RX_LAST_PACKET))) {
1900 } else if (rxi_2dchoice && !choice2
1901 && !(tcall->flags & RX_CALL_CLEARED)
1902 && (tcall->rprev > rxi_HardAckRate)) {
1916 /* we can't schedule a call if there's no data!!! */
1917 /* send an ack if there's no data, if we're missing the
1918 * first packet, or we're missing something between first
1919 * and last -- there's a "hole" in the incoming data. */
1920 if (queue_IsEmpty(&call->rq)
1921 || queue_First(&call->rq, rx_packet)->header.seq != 1
1922 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1923 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1925 call->flags &= (~RX_CALL_WAIT_PROC);
1926 service->nRequestsRunning++;
1927 /* just started call in minProcs pool, need fewer to maintain
1929 MUTEX_ENTER(&rx_quota_mutex);
1930 if (service->nRequestsRunning <= service->minProcs)
1933 MUTEX_EXIT(&rx_quota_mutex);
1935 /* MUTEX_EXIT(&call->lock); */
1937 /* If there are no eligible incoming calls, add this process
1938 * to the idle server queue, to wait for one */
1941 *socketp = OSI_NULLSOCKET;
1943 sq->socketp = socketp;
1944 queue_Append(&rx_idleServerQueue, sq);
1948 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1950 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1951 return (struct rx_call *)0;
1954 } while (!(call = sq->newcall)
1955 && !(socketp && *socketp != OSI_NULLSOCKET));
1957 MUTEX_EXIT(&sq->lock);
1959 MUTEX_ENTER(&freeSQEList_lock);
1960 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1961 rx_FreeSQEList = sq;
1962 MUTEX_EXIT(&freeSQEList_lock);
1965 clock_GetTime(&call->startTime);
1966 call->state = RX_STATE_ACTIVE;
1967 call->mode = RX_MODE_RECEIVING;
1968 #ifdef RX_KERNEL_TRACE
1969 if (ICL_SETACTIVE(afs_iclSetp)) {
1970 int glockOwner = ISAFS_GLOCK();
1973 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1974 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1981 rxi_calltrace(RX_CALL_START, call);
1982 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
1983 call->conn->service->servicePort, call->conn->service->serviceId,
1986 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1993 #endif /* RX_ENABLE_LOCKS */
1997 /* Establish a procedure to be called when a packet arrives for a
1998 * call. This routine will be called at most once after each call,
1999 * and will also be called if there is an error condition on the or
2000 * the call is complete. Used by multi rx to build a selection
2001 * function which determines which of several calls is likely to be a
2002 * good one to read from.
2003 * NOTE: the way this is currently implemented it is probably only a
2004 * good idea to (1) use it immediately after a newcall (clients only)
2005 * and (2) only use it once. Other uses currently void your warranty
2008 rx_SetArrivalProc(struct rx_call *call,
2009 void (*proc) (struct rx_call * call,
2012 void * handle, int arg)
2014 call->arrivalProc = proc;
2015 call->arrivalProcHandle = handle;
2016 call->arrivalProcArg = arg;
2019 /* Call is finished (possibly prematurely). Return rc to the peer, if
2020 * appropriate, and return the final error code from the conversation
2024 rx_EndCall(struct rx_call *call, afs_int32 rc)
2026 struct rx_connection *conn = call->conn;
2027 struct rx_service *service;
2031 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2032 call, rc, call->error, call->abortCode));
2035 MUTEX_ENTER(&call->lock);
2037 if (rc == 0 && call->error == 0) {
2038 call->abortCode = 0;
2039 call->abortCount = 0;
2042 call->arrivalProc = (void (*)())0;
2043 if (rc && call->error == 0) {
2044 rxi_CallError(call, rc);
2045 /* Send an abort message to the peer if this error code has
2046 * only just been set. If it was set previously, assume the
2047 * peer has already been sent the error code or will request it
2049 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2051 if (conn->type == RX_SERVER_CONNECTION) {
2052 /* Make sure reply or at least dummy reply is sent */
2053 if (call->mode == RX_MODE_RECEIVING) {
2054 rxi_WriteProc(call, 0, 0);
2056 if (call->mode == RX_MODE_SENDING) {
2057 rxi_FlushWrite(call);
2059 service = conn->service;
2060 rxi_calltrace(RX_CALL_END, call);
2061 /* Call goes to hold state until reply packets are acknowledged */
2062 if (call->tfirst + call->nSoftAcked < call->tnext) {
2063 call->state = RX_STATE_HOLD;
2065 call->state = RX_STATE_DALLY;
2066 rxi_ClearTransmitQueue(call, 0);
2067 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
2068 rxevent_Cancel(call->keepAliveEvent, call,
2069 RX_CALL_REFCOUNT_ALIVE);
2071 } else { /* Client connection */
2073 /* Make sure server receives input packets, in the case where
2074 * no reply arguments are expected */
2075 if ((call->mode == RX_MODE_SENDING)
2076 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2077 (void)rxi_ReadProc(call, &dummy, 1);
2080 /* If we had an outstanding delayed ack, be nice to the server
2081 * and force-send it now.
2083 if (call->delayedAckEvent) {
2084 rxevent_Cancel(call->delayedAckEvent, call,
2085 RX_CALL_REFCOUNT_DELAY);
2086 call->delayedAckEvent = NULL;
2087 rxi_SendDelayedAck(NULL, call, NULL);
2090 /* We need to release the call lock since it's lower than the
2091 * conn_call_lock and we don't want to hold the conn_call_lock
2092 * over the rx_ReadProc call. The conn_call_lock needs to be held
2093 * here for the case where rx_NewCall is perusing the calls on
2094 * the connection structure. We don't want to signal until
2095 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2096 * have checked this call, found it active and by the time it
2097 * goes to sleep, will have missed the signal.
2099 MUTEX_EXIT(&call->lock);
2100 MUTEX_ENTER(&conn->conn_call_lock);
2101 MUTEX_ENTER(&call->lock);
2102 MUTEX_ENTER(&conn->conn_data_lock);
2103 conn->flags |= RX_CONN_BUSY;
2104 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2105 MUTEX_EXIT(&conn->conn_data_lock);
2106 #ifdef RX_ENABLE_LOCKS
2107 CV_BROADCAST(&conn->conn_call_cv);
2112 #ifdef RX_ENABLE_LOCKS
2114 MUTEX_EXIT(&conn->conn_data_lock);
2116 #endif /* RX_ENABLE_LOCKS */
2117 call->state = RX_STATE_DALLY;
2119 error = call->error;
2121 /* currentPacket, nLeft, and NFree must be zeroed here, because
2122 * ResetCall cannot: ResetCall may be called at splnet(), in the
2123 * kernel version, and may interrupt the macros rx_Read or
2124 * rx_Write, which run at normal priority for efficiency. */
2125 if (call->currentPacket) {
2126 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2127 rxi_FreePacket(call->currentPacket);
2128 call->currentPacket = (struct rx_packet *)0;
2131 call->nLeft = call->nFree = call->curlen = 0;
2133 /* Free any packets from the last call to ReadvProc/WritevProc */
2134 #ifdef RXDEBUG_PACKET
2136 #endif /* RXDEBUG_PACKET */
2137 rxi_FreePackets(0, &call->iovq);
2139 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2140 MUTEX_EXIT(&call->lock);
2141 if (conn->type == RX_CLIENT_CONNECTION) {
2142 MUTEX_ENTER(&conn->conn_data_lock);
2143 conn->flags &= ~RX_CONN_BUSY;
2144 MUTEX_EXIT(&conn->conn_data_lock);
2145 MUTEX_EXIT(&conn->conn_call_lock);
2149 * Map errors to the local host's errno.h format.
2151 error = ntoh_syserr_conv(error);
2155 #if !defined(KERNEL)
2157 /* Call this routine when shutting down a server or client (especially
2158 * clients). This will allow Rx to gracefully garbage collect server
2159 * connections, and reduce the number of retries that a server might
2160 * make to a dead client.
2161 * This is not quite right, since some calls may still be ongoing and
2162 * we can't lock them to destroy them. */
2166 struct rx_connection **conn_ptr, **conn_end;
2170 if (rxinit_status == 1) {
2172 return; /* Already shutdown. */
2174 rxi_DeleteCachedConnections();
2175 if (rx_connHashTable) {
2176 MUTEX_ENTER(&rx_connHashTable_lock);
2177 for (conn_ptr = &rx_connHashTable[0], conn_end =
2178 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2180 struct rx_connection *conn, *next;
2181 for (conn = *conn_ptr; conn; conn = next) {
2183 if (conn->type == RX_CLIENT_CONNECTION) {
2184 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2186 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2187 #ifdef RX_ENABLE_LOCKS
2188 rxi_DestroyConnectionNoLock(conn);
2189 #else /* RX_ENABLE_LOCKS */
2190 rxi_DestroyConnection(conn);
2191 #endif /* RX_ENABLE_LOCKS */
2195 #ifdef RX_ENABLE_LOCKS
2196 while (rx_connCleanup_list) {
2197 struct rx_connection *conn;
2198 conn = rx_connCleanup_list;
2199 rx_connCleanup_list = rx_connCleanup_list->next;
2200 MUTEX_EXIT(&rx_connHashTable_lock);
2201 rxi_CleanupConnection(conn);
2202 MUTEX_ENTER(&rx_connHashTable_lock);
2204 MUTEX_EXIT(&rx_connHashTable_lock);
2205 #endif /* RX_ENABLE_LOCKS */
2210 afs_winsockCleanup();
2218 /* if we wakeup packet waiter too often, can get in loop with two
2219 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2221 rxi_PacketsUnWait(void)
2223 if (!rx_waitingForPackets) {
2227 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2228 return; /* still over quota */
2231 rx_waitingForPackets = 0;
2232 #ifdef RX_ENABLE_LOCKS
2233 CV_BROADCAST(&rx_waitingForPackets_cv);
2235 osi_rxWakeup(&rx_waitingForPackets);
2241 /* ------------------Internal interfaces------------------------- */
2243 /* Return this process's service structure for the
2244 * specified socket and service */
2246 rxi_FindService(osi_socket socket, u_short serviceId)
2248 struct rx_service **sp;
2249 for (sp = &rx_services[0]; *sp; sp++) {
2250 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2256 #ifdef RXDEBUG_PACKET
2257 #ifdef KDUMP_RX_LOCK
2258 static struct rx_call_rx_lock *rx_allCallsp = 0;
2260 static struct rx_call *rx_allCallsp = 0;
2262 #endif /* RXDEBUG_PACKET */
2264 /* Allocate a call structure, for the indicated channel of the
2265 * supplied connection. The mode and state of the call must be set by
2266 * the caller. Returns the call with mutex locked. */
2268 rxi_NewCall(struct rx_connection *conn, int channel)
2270 struct rx_call *call;
2271 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2272 struct rx_call *cp; /* Call pointer temp */
2273 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2274 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2276 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2278 /* Grab an existing call structure, or allocate a new one.
2279 * Existing call structures are assumed to have been left reset by
2281 MUTEX_ENTER(&rx_freeCallQueue_lock);
2283 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2285 * EXCEPT that the TQ might not yet be cleared out.
2286 * Skip over those with in-use TQs.
2289 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2290 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2296 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2297 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2298 call = queue_First(&rx_freeCallQueue, rx_call);
2299 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2301 if (rx_stats_active)
2302 rx_MutexDecrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2303 MUTEX_EXIT(&rx_freeCallQueue_lock);
2304 MUTEX_ENTER(&call->lock);
2305 CLEAR_CALL_QUEUE_LOCK(call);
2306 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2307 /* Now, if TQ wasn't cleared earlier, do it now. */
2308 rxi_WaitforTQBusy(call);
2309 if (call->flags & RX_CALL_TQ_CLEARME) {
2310 rxi_ClearTransmitQueue(call, 1);
2311 /*queue_Init(&call->tq);*/
2313 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2314 /* Bind the call to its connection structure */
2316 rxi_ResetCall(call, 1);
2319 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2320 #ifdef RXDEBUG_PACKET
2321 call->allNextp = rx_allCallsp;
2322 rx_allCallsp = call;
2324 #endif /* RXDEBUG_PACKET */
2325 rx_MutexIncrement(rx_stats.nCallStructs, rx_stats_mutex);
2327 MUTEX_EXIT(&rx_freeCallQueue_lock);
2328 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2329 MUTEX_ENTER(&call->lock);
2330 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2331 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2332 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2334 /* Initialize once-only items */
2335 queue_Init(&call->tq);
2336 queue_Init(&call->rq);
2337 queue_Init(&call->iovq);
2338 #ifdef RXDEBUG_PACKET
2339 call->rqc = call->tqc = call->iovqc = 0;
2340 #endif /* RXDEBUG_PACKET */
2341 /* Bind the call to its connection structure (prereq for reset) */
2343 rxi_ResetCall(call, 1);
2345 call->channel = channel;
2346 call->callNumber = &conn->callNumber[channel];
2347 call->rwind = conn->rwind[channel];
2348 call->twind = conn->twind[channel];
2349 /* Note that the next expected call number is retained (in
2350 * conn->callNumber[i]), even if we reallocate the call structure
2352 conn->call[channel] = call;
2353 /* if the channel's never been used (== 0), we should start at 1, otherwise
2354 * the call number is valid from the last time this channel was used */
2355 if (*call->callNumber == 0)
2356 *call->callNumber = 1;
2361 /* A call has been inactive long enough that so we can throw away
2362 * state, including the call structure, which is placed on the call
2364 * Call is locked upon entry.
2365 * haveCTLock set if called from rxi_ReapConnections
2367 #ifdef RX_ENABLE_LOCKS
2369 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2370 #else /* RX_ENABLE_LOCKS */
2372 rxi_FreeCall(struct rx_call *call)
2373 #endif /* RX_ENABLE_LOCKS */
2375 int channel = call->channel;
2376 struct rx_connection *conn = call->conn;
2379 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2380 (*call->callNumber)++;
2381 rxi_ResetCall(call, 0);
2382 call->conn->call[channel] = (struct rx_call *)0;
2384 MUTEX_ENTER(&rx_freeCallQueue_lock);
2385 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2386 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2387 /* A call may be free even though its transmit queue is still in use.
2388 * Since we search the call list from head to tail, put busy calls at
2389 * the head of the list, and idle calls at the tail.
2391 if (call->flags & RX_CALL_TQ_BUSY)
2392 queue_Prepend(&rx_freeCallQueue, call);
2394 queue_Append(&rx_freeCallQueue, call);
2395 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2396 queue_Append(&rx_freeCallQueue, call);
2397 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2398 if (rx_stats_active)
2399 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2400 MUTEX_EXIT(&rx_freeCallQueue_lock);
2402 /* Destroy the connection if it was previously slated for
2403 * destruction, i.e. the Rx client code previously called
2404 * rx_DestroyConnection (client connections), or
2405 * rxi_ReapConnections called the same routine (server
2406 * connections). Only do this, however, if there are no
2407 * outstanding calls. Note that for fine grain locking, there appears
2408 * to be a deadlock in that rxi_FreeCall has a call locked and
2409 * DestroyConnectionNoLock locks each call in the conn. But note a
2410 * few lines up where we have removed this call from the conn.
2411 * If someone else destroys a connection, they either have no
2412 * call lock held or are going through this section of code.
2414 MUTEX_ENTER(&conn->conn_data_lock);
2415 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2417 MUTEX_EXIT(&conn->conn_data_lock);
2418 #ifdef RX_ENABLE_LOCKS
2420 rxi_DestroyConnectionNoLock(conn);
2422 rxi_DestroyConnection(conn);
2423 #else /* RX_ENABLE_LOCKS */
2424 rxi_DestroyConnection(conn);
2425 #endif /* RX_ENABLE_LOCKS */
2427 MUTEX_EXIT(&conn->conn_data_lock);
2431 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2433 rxi_Alloc(size_t size)
2437 if (rx_stats_active)
2438 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2441 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2442 afs_osi_Alloc_NoSleep(size);
2447 osi_Panic("rxi_Alloc error");
2453 rxi_Free(void *addr, size_t size)
2455 if (rx_stats_active)
2456 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2457 osi_Free(addr, size);
2461 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2463 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2464 struct rx_peer *next = NULL;
2468 MUTEX_ENTER(&rx_peerHashTable_lock);
2470 peer_ptr = &rx_peerHashTable[0];
2471 peer_end = &rx_peerHashTable[rx_hashTableSize];
2474 for ( ; peer_ptr < peer_end; peer_ptr++) {
2477 for ( ; peer; peer = next) {
2479 if (host == peer->host)
2484 hashIndex = PEER_HASH(host, port);
2485 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2486 if ((peer->host == host) && (peer->port == port))
2491 MUTEX_ENTER(&rx_peerHashTable_lock);
2496 MUTEX_EXIT(&rx_peerHashTable_lock);
2498 MUTEX_ENTER(&peer->peer_lock);
2499 /* We don't handle dropping below min, so don't */
2500 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2501 peer->ifMTU=MIN(mtu, peer->ifMTU);
2502 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2503 /* if we tweaked this down, need to tune our peer MTU too */
2504 peer->MTU = MIN(peer->MTU, peer->natMTU);
2505 /* if we discovered a sub-1500 mtu, degrade */
2506 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2507 peer->maxDgramPackets = 1;
2508 /* We no longer have valid peer packet information */
2509 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2510 peer->maxPacketSize = 0;
2511 MUTEX_EXIT(&peer->peer_lock);
2513 MUTEX_ENTER(&rx_peerHashTable_lock);
2515 if (host && !port) {
2517 /* pick up where we left off */
2521 MUTEX_EXIT(&rx_peerHashTable_lock);
2524 /* Find the peer process represented by the supplied (host,port)
2525 * combination. If there is no appropriate active peer structure, a
2526 * new one will be allocated and initialized
2527 * The origPeer, if set, is a pointer to a peer structure on which the
2528 * refcount will be be decremented. This is used to replace the peer
2529 * structure hanging off a connection structure */
2531 rxi_FindPeer(afs_uint32 host, u_short port,
2532 struct rx_peer *origPeer, int create)
2536 hashIndex = PEER_HASH(host, port);
2537 MUTEX_ENTER(&rx_peerHashTable_lock);
2538 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2539 if ((pp->host == host) && (pp->port == port))
2544 pp = rxi_AllocPeer(); /* This bzero's *pp */
2545 pp->host = host; /* set here or in InitPeerParams is zero */
2547 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2548 queue_Init(&pp->congestionQueue);
2549 queue_Init(&pp->rpcStats);
2550 pp->next = rx_peerHashTable[hashIndex];
2551 rx_peerHashTable[hashIndex] = pp;
2552 rxi_InitPeerParams(pp);
2553 if (rx_stats_active)
2554 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2561 origPeer->refCount--;
2562 MUTEX_EXIT(&rx_peerHashTable_lock);
2567 /* Find the connection at (host, port) started at epoch, and with the
2568 * given connection id. Creates the server connection if necessary.
2569 * The type specifies whether a client connection or a server
2570 * connection is desired. In both cases, (host, port) specify the
2571 * peer's (host, pair) pair. Client connections are not made
2572 * automatically by this routine. The parameter socket gives the
2573 * socket descriptor on which the packet was received. This is used,
2574 * in the case of server connections, to check that *new* connections
2575 * come via a valid (port, serviceId). Finally, the securityIndex
2576 * parameter must match the existing index for the connection. If a
2577 * server connection is created, it will be created using the supplied
2578 * index, if the index is valid for this service */
2579 struct rx_connection *
2580 rxi_FindConnection(osi_socket socket, afs_int32 host,
2581 u_short port, u_short serviceId, afs_uint32 cid,
2582 afs_uint32 epoch, int type, u_int securityIndex)
2584 int hashindex, flag, i;
2585 struct rx_connection *conn;
2586 hashindex = CONN_HASH(host, port, cid, epoch, type);
2587 MUTEX_ENTER(&rx_connHashTable_lock);
2588 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2589 rx_connHashTable[hashindex],
2592 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2593 && (epoch == conn->epoch)) {
2594 struct rx_peer *pp = conn->peer;
2595 if (securityIndex != conn->securityIndex) {
2596 /* this isn't supposed to happen, but someone could forge a packet
2597 * like this, and there seems to be some CM bug that makes this
2598 * happen from time to time -- in which case, the fileserver
2600 MUTEX_EXIT(&rx_connHashTable_lock);
2601 return (struct rx_connection *)0;
2603 if (pp->host == host && pp->port == port)
2605 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2607 /* So what happens when it's a callback connection? */
2608 if ( /*type == RX_CLIENT_CONNECTION && */
2609 (conn->epoch & 0x80000000))
2613 /* the connection rxLastConn that was used the last time is not the
2614 ** one we are looking for now. Hence, start searching in the hash */
2616 conn = rx_connHashTable[hashindex];
2621 struct rx_service *service;
2622 if (type == RX_CLIENT_CONNECTION) {
2623 MUTEX_EXIT(&rx_connHashTable_lock);
2624 return (struct rx_connection *)0;
2626 service = rxi_FindService(socket, serviceId);
2627 if (!service || (securityIndex >= service->nSecurityObjects)
2628 || (service->securityObjects[securityIndex] == 0)) {
2629 MUTEX_EXIT(&rx_connHashTable_lock);
2630 return (struct rx_connection *)0;
2632 conn = rxi_AllocConnection(); /* This bzero's the connection */
2633 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2634 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2635 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2636 conn->next = rx_connHashTable[hashindex];
2637 rx_connHashTable[hashindex] = conn;
2638 conn->peer = rxi_FindPeer(host, port, 0, 1);
2639 conn->type = RX_SERVER_CONNECTION;
2640 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2641 conn->epoch = epoch;
2642 conn->cid = cid & RX_CIDMASK;
2643 /* conn->serial = conn->lastSerial = 0; */
2644 /* conn->timeout = 0; */
2645 conn->ackRate = RX_FAST_ACK_RATE;
2646 conn->service = service;
2647 conn->serviceId = serviceId;
2648 conn->securityIndex = securityIndex;
2649 conn->securityObject = service->securityObjects[securityIndex];
2650 conn->nSpecific = 0;
2651 conn->specific = NULL;
2652 rx_SetConnDeadTime(conn, service->connDeadTime);
2653 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2654 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2655 for (i = 0; i < RX_MAXCALLS; i++) {
2656 conn->twind[i] = rx_initSendWindow;
2657 conn->rwind[i] = rx_initReceiveWindow;
2659 /* Notify security object of the new connection */
2660 RXS_NewConnection(conn->securityObject, conn);
2661 /* XXXX Connection timeout? */
2662 if (service->newConnProc)
2663 (*service->newConnProc) (conn);
2664 if (rx_stats_active)
2665 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2668 MUTEX_ENTER(&conn->conn_data_lock);
2670 MUTEX_EXIT(&conn->conn_data_lock);
2672 rxLastConn = conn; /* store this connection as the last conn used */
2673 MUTEX_EXIT(&rx_connHashTable_lock);
2677 /* There are two packet tracing routines available for testing and monitoring
2678 * Rx. One is called just after every packet is received and the other is
2679 * called just before every packet is sent. Received packets, have had their
2680 * headers decoded, and packets to be sent have not yet had their headers
2681 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2682 * containing the network address. Both can be modified. The return value, if
2683 * non-zero, indicates that the packet should be dropped. */
2685 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2686 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2688 /* A packet has been received off the interface. Np is the packet, socket is
2689 * the socket number it was received from (useful in determining which service
2690 * this packet corresponds to), and (host, port) reflect the host,port of the
2691 * sender. This call returns the packet to the caller if it is finished with
2692 * it, rather than de-allocating it, just as a small performance hack */
2695 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
2696 afs_uint32 host, u_short port, int *tnop,
2697 struct rx_call **newcallp)
2699 struct rx_call *call;
2700 struct rx_connection *conn;
2702 afs_uint32 currentCallNumber;
2708 struct rx_packet *tnp;
2711 /* We don't print out the packet until now because (1) the time may not be
2712 * accurate enough until now in the lwp implementation (rx_Listener only gets
2713 * the time after the packet is read) and (2) from a protocol point of view,
2714 * this is the first time the packet has been seen */
2715 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2716 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2717 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT,
2718 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2719 np->header.epoch, np->header.cid, np->header.callNumber,
2720 np->header.seq, np->header.flags, np));
2723 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2724 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2727 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2728 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2731 /* If an input tracer function is defined, call it with the packet and
2732 * network address. Note this function may modify its arguments. */
2733 if (rx_justReceived) {
2734 struct sockaddr_in addr;
2736 addr.sin_family = AF_INET;
2737 addr.sin_port = port;
2738 addr.sin_addr.s_addr = host;
2739 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2740 addr.sin_len = sizeof(addr);
2741 #endif /* AFS_OSF_ENV */
2742 drop = (*rx_justReceived) (np, &addr);
2743 /* drop packet if return value is non-zero */
2746 port = addr.sin_port; /* in case fcn changed addr */
2747 host = addr.sin_addr.s_addr;
2751 /* If packet was not sent by the client, then *we* must be the client */
2752 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2753 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2755 /* Find the connection (or fabricate one, if we're the server & if
2756 * necessary) associated with this packet */
2758 rxi_FindConnection(socket, host, port, np->header.serviceId,
2759 np->header.cid, np->header.epoch, type,
2760 np->header.securityIndex);
2763 /* If no connection found or fabricated, just ignore the packet.
2764 * (An argument could be made for sending an abort packet for
2769 MUTEX_ENTER(&conn->conn_data_lock);
2770 if (conn->maxSerial < np->header.serial)
2771 conn->maxSerial = np->header.serial;
2772 MUTEX_EXIT(&conn->conn_data_lock);
2774 /* If the connection is in an error state, send an abort packet and ignore
2775 * the incoming packet */
2777 /* Don't respond to an abort packet--we don't want loops! */
2778 MUTEX_ENTER(&conn->conn_data_lock);
2779 if (np->header.type != RX_PACKET_TYPE_ABORT)
2780 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2782 MUTEX_EXIT(&conn->conn_data_lock);
2786 /* Check for connection-only requests (i.e. not call specific). */
2787 if (np->header.callNumber == 0) {
2788 switch (np->header.type) {
2789 case RX_PACKET_TYPE_ABORT: {
2790 /* What if the supplied error is zero? */
2791 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2792 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2793 rxi_ConnectionError(conn, errcode);
2794 MUTEX_ENTER(&conn->conn_data_lock);
2796 MUTEX_EXIT(&conn->conn_data_lock);
2799 case RX_PACKET_TYPE_CHALLENGE:
2800 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2801 MUTEX_ENTER(&conn->conn_data_lock);
2803 MUTEX_EXIT(&conn->conn_data_lock);
2805 case RX_PACKET_TYPE_RESPONSE:
2806 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2807 MUTEX_ENTER(&conn->conn_data_lock);
2809 MUTEX_EXIT(&conn->conn_data_lock);
2811 case RX_PACKET_TYPE_PARAMS:
2812 case RX_PACKET_TYPE_PARAMS + 1:
2813 case RX_PACKET_TYPE_PARAMS + 2:
2814 /* ignore these packet types for now */
2815 MUTEX_ENTER(&conn->conn_data_lock);
2817 MUTEX_EXIT(&conn->conn_data_lock);
2822 /* Should not reach here, unless the peer is broken: send an
2824 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2825 MUTEX_ENTER(&conn->conn_data_lock);
2826 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2828 MUTEX_EXIT(&conn->conn_data_lock);
2833 channel = np->header.cid & RX_CHANNELMASK;
2834 call = conn->call[channel];
2835 #ifdef RX_ENABLE_LOCKS
2837 MUTEX_ENTER(&call->lock);
2838 /* Test to see if call struct is still attached to conn. */
2839 if (call != conn->call[channel]) {
2841 MUTEX_EXIT(&call->lock);
2842 if (type == RX_SERVER_CONNECTION) {
2843 call = conn->call[channel];
2844 /* If we started with no call attached and there is one now,
2845 * another thread is also running this routine and has gotten
2846 * the connection channel. We should drop this packet in the tests
2847 * below. If there was a call on this connection and it's now
2848 * gone, then we'll be making a new call below.
2849 * If there was previously a call and it's now different then
2850 * the old call was freed and another thread running this routine
2851 * has created a call on this channel. One of these two threads
2852 * has a packet for the old call and the code below handles those
2856 MUTEX_ENTER(&call->lock);
2858 /* This packet can't be for this call. If the new call address is
2859 * 0 then no call is running on this channel. If there is a call
2860 * then, since this is a client connection we're getting data for
2861 * it must be for the previous call.
2863 if (rx_stats_active)
2864 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2865 MUTEX_ENTER(&conn->conn_data_lock);
2867 MUTEX_EXIT(&conn->conn_data_lock);
2872 currentCallNumber = conn->callNumber[channel];
2874 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2875 if (np->header.callNumber < currentCallNumber) {
2876 if (rx_stats_active)
2877 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2878 #ifdef RX_ENABLE_LOCKS
2880 MUTEX_EXIT(&call->lock);
2882 MUTEX_ENTER(&conn->conn_data_lock);
2884 MUTEX_EXIT(&conn->conn_data_lock);
2888 MUTEX_ENTER(&conn->conn_call_lock);
2889 call = rxi_NewCall(conn, channel);
2890 MUTEX_EXIT(&conn->conn_call_lock);
2891 *call->callNumber = np->header.callNumber;
2893 if (np->header.callNumber == 0)
2894 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%.06d len %d",
2895 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
2896 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
2897 np->header.flags, np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2899 call->state = RX_STATE_PRECALL;
2900 clock_GetTime(&call->queueTime);
2901 hzero(call->bytesSent);
2902 hzero(call->bytesRcvd);
2904 * If the number of queued calls exceeds the overload
2905 * threshold then abort this call.
2907 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2908 struct rx_packet *tp;
2910 rxi_CallError(call, rx_BusyError);
2911 tp = rxi_SendCallAbort(call, np, 1, 0);
2912 MUTEX_EXIT(&call->lock);
2913 MUTEX_ENTER(&conn->conn_data_lock);
2915 MUTEX_EXIT(&conn->conn_data_lock);
2916 if (rx_stats_active)
2917 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2920 rxi_KeepAliveOn(call);
2921 } else if (np->header.callNumber != currentCallNumber) {
2922 /* Wait until the transmit queue is idle before deciding
2923 * whether to reset the current call. Chances are that the
2924 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2927 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2928 while ((call->state == RX_STATE_ACTIVE)
2929 && (call->flags & RX_CALL_TQ_BUSY)) {
2930 call->flags |= RX_CALL_TQ_WAIT;
2932 #ifdef RX_ENABLE_LOCKS
2933 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2934 CV_WAIT(&call->cv_tq, &call->lock);
2935 #else /* RX_ENABLE_LOCKS */
2936 osi_rxSleep(&call->tq);
2937 #endif /* RX_ENABLE_LOCKS */
2939 if (call->tqWaiters == 0)
2940 call->flags &= ~RX_CALL_TQ_WAIT;
2942 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2943 /* If the new call cannot be taken right now send a busy and set
2944 * the error condition in this call, so that it terminates as
2945 * quickly as possible */
2946 if (call->state == RX_STATE_ACTIVE) {
2947 struct rx_packet *tp;
2949 rxi_CallError(call, RX_CALL_DEAD);
2950 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2952 MUTEX_EXIT(&call->lock);
2953 MUTEX_ENTER(&conn->conn_data_lock);
2955 MUTEX_EXIT(&conn->conn_data_lock);
2958 rxi_ResetCall(call, 0);
2959 *call->callNumber = np->header.callNumber;
2961 if (np->header.callNumber == 0)
2962 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%06d len %d",
2963 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
2964 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
2965 np->header.flags, np, np->retryTime.sec, np->retryTime.usec, np->length));
2967 call->state = RX_STATE_PRECALL;
2968 clock_GetTime(&call->queueTime);
2969 hzero(call->bytesSent);
2970 hzero(call->bytesRcvd);
2972 * If the number of queued calls exceeds the overload
2973 * threshold then abort this call.
2975 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2976 struct rx_packet *tp;
2978 rxi_CallError(call, rx_BusyError);
2979 tp = rxi_SendCallAbort(call, np, 1, 0);
2980 MUTEX_EXIT(&call->lock);
2981 MUTEX_ENTER(&conn->conn_data_lock);
2983 MUTEX_EXIT(&conn->conn_data_lock);
2984 if (rx_stats_active)
2985 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2988 rxi_KeepAliveOn(call);
2990 /* Continuing call; do nothing here. */
2992 } else { /* we're the client */
2993 /* Ignore all incoming acknowledgements for calls in DALLY state */
2994 if (call && (call->state == RX_STATE_DALLY)
2995 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2996 if (rx_stats_active)
2997 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
2998 #ifdef RX_ENABLE_LOCKS
3000 MUTEX_EXIT(&call->lock);
3003 MUTEX_ENTER(&conn->conn_data_lock);
3005 MUTEX_EXIT(&conn->conn_data_lock);
3009 /* Ignore anything that's not relevant to the current call. If there
3010 * isn't a current call, then no packet is relevant. */
3011 if (!call || (np->header.callNumber != currentCallNumber)) {
3012 if (rx_stats_active)
3013 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
3014 #ifdef RX_ENABLE_LOCKS
3016 MUTEX_EXIT(&call->lock);
3019 MUTEX_ENTER(&conn->conn_data_lock);
3021 MUTEX_EXIT(&conn->conn_data_lock);
3024 /* If the service security object index stamped in the packet does not
3025 * match the connection's security index, ignore the packet */
3026 if (np->header.securityIndex != conn->securityIndex) {
3027 #ifdef RX_ENABLE_LOCKS
3028 MUTEX_EXIT(&call->lock);
3030 MUTEX_ENTER(&conn->conn_data_lock);
3032 MUTEX_EXIT(&conn->conn_data_lock);
3036 /* If we're receiving the response, then all transmit packets are
3037 * implicitly acknowledged. Get rid of them. */
3038 if (np->header.type == RX_PACKET_TYPE_DATA) {
3039 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3040 /* XXX Hack. Because we must release the global rx lock when
3041 * sending packets (osi_NetSend) we drop all acks while we're
3042 * traversing the tq in rxi_Start sending packets out because
3043 * packets may move to the freePacketQueue as result of being here!
3044 * So we drop these packets until we're safely out of the
3045 * traversing. Really ugly!
3046 * For fine grain RX locking, we set the acked field in the
3047 * packets and let rxi_Start remove them from the transmit queue.
3049 if (call->flags & RX_CALL_TQ_BUSY) {
3050 #ifdef RX_ENABLE_LOCKS
3051 rxi_SetAcksInTransmitQueue(call);
3054 return np; /* xmitting; drop packet */
3057 rxi_ClearTransmitQueue(call, 0);
3059 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3060 rxi_ClearTransmitQueue(call, 0);
3061 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3063 if (np->header.type == RX_PACKET_TYPE_ACK) {
3064 /* now check to see if this is an ack packet acknowledging that the
3065 * server actually *lost* some hard-acked data. If this happens we
3066 * ignore this packet, as it may indicate that the server restarted in
3067 * the middle of a call. It is also possible that this is an old ack
3068 * packet. We don't abort the connection in this case, because this
3069 * *might* just be an old ack packet. The right way to detect a server
3070 * restart in the midst of a call is to notice that the server epoch
3072 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3073 * XXX unacknowledged. I think that this is off-by-one, but
3074 * XXX I don't dare change it just yet, since it will
3075 * XXX interact badly with the server-restart detection
3076 * XXX code in receiveackpacket. */
3077 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3078 if (rx_stats_active)
3079 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
3080 MUTEX_EXIT(&call->lock);
3081 MUTEX_ENTER(&conn->conn_data_lock);
3083 MUTEX_EXIT(&conn->conn_data_lock);
3087 } /* else not a data packet */
3090 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3091 /* Set remote user defined status from packet */
3092 call->remoteStatus = np->header.userStatus;
3094 /* Note the gap between the expected next packet and the actual
3095 * packet that arrived, when the new packet has a smaller serial number
3096 * than expected. Rioses frequently reorder packets all by themselves,
3097 * so this will be quite important with very large window sizes.
3098 * Skew is checked against 0 here to avoid any dependence on the type of
3099 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3101 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3102 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3103 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3105 MUTEX_ENTER(&conn->conn_data_lock);
3106 skew = conn->lastSerial - np->header.serial;
3107 conn->lastSerial = np->header.serial;
3108 MUTEX_EXIT(&conn->conn_data_lock);
3110 struct rx_peer *peer;
3112 if (skew > peer->inPacketSkew) {
3113 dpf(("*** In skew changed from %d to %d\n",
3114 peer->inPacketSkew, skew));
3115 peer->inPacketSkew = skew;
3119 /* Now do packet type-specific processing */
3120 switch (np->header.type) {
3121 case RX_PACKET_TYPE_DATA:
3122 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3125 case RX_PACKET_TYPE_ACK:
3126 /* Respond immediately to ack packets requesting acknowledgement
3128 if (np->header.flags & RX_REQUEST_ACK) {
3130 (void)rxi_SendCallAbort(call, 0, 1, 0);
3132 (void)rxi_SendAck(call, 0, np->header.serial,
3133 RX_ACK_PING_RESPONSE, 1);
3135 np = rxi_ReceiveAckPacket(call, np, 1);
3137 case RX_PACKET_TYPE_ABORT: {
3138 /* An abort packet: reset the call, passing the error up to the user. */
3139 /* What if error is zero? */
3140 /* What if the error is -1? the application will treat it as a timeout. */
3141 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3142 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
3143 rxi_CallError(call, errdata);
3144 MUTEX_EXIT(&call->lock);
3145 MUTEX_ENTER(&conn->conn_data_lock);
3147 MUTEX_EXIT(&conn->conn_data_lock);
3148 return np; /* xmitting; drop packet */
3150 case RX_PACKET_TYPE_BUSY:
3153 case RX_PACKET_TYPE_ACKALL:
3154 /* All packets acknowledged, so we can drop all packets previously
3155 * readied for sending */
3156 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3157 /* XXX Hack. We because we can't release the global rx lock when
3158 * sending packets (osi_NetSend) we drop all ack pkts while we're
3159 * traversing the tq in rxi_Start sending packets out because
3160 * packets may move to the freePacketQueue as result of being
3161 * here! So we drop these packets until we're safely out of the
3162 * traversing. Really ugly!
3163 * For fine grain RX locking, we set the acked field in the packets
3164 * and let rxi_Start remove the packets from the transmit queue.
3166 if (call->flags & RX_CALL_TQ_BUSY) {
3167 #ifdef RX_ENABLE_LOCKS
3168 rxi_SetAcksInTransmitQueue(call);
3170 #else /* RX_ENABLE_LOCKS */
3171 MUTEX_EXIT(&call->lock);
3172 MUTEX_ENTER(&conn->conn_data_lock);
3174 MUTEX_EXIT(&conn->conn_data_lock);
3175 return np; /* xmitting; drop packet */
3176 #endif /* RX_ENABLE_LOCKS */
3178 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3179 rxi_ClearTransmitQueue(call, 0);
3180 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3183 /* Should not reach here, unless the peer is broken: send an abort
3185 rxi_CallError(call, RX_PROTOCOL_ERROR);
3186 np = rxi_SendCallAbort(call, np, 1, 0);
3189 /* Note when this last legitimate packet was received, for keep-alive
3190 * processing. Note, we delay getting the time until now in the hope that
3191 * the packet will be delivered to the user before any get time is required
3192 * (if not, then the time won't actually be re-evaluated here). */
3193 call->lastReceiveTime = clock_Sec();
3194 MUTEX_EXIT(&call->lock);
3195 MUTEX_ENTER(&conn->conn_data_lock);
3197 MUTEX_EXIT(&conn->conn_data_lock);
3201 /* return true if this is an "interesting" connection from the point of view
3202 of someone trying to debug the system */
3204 rxi_IsConnInteresting(struct rx_connection *aconn)
3207 struct rx_call *tcall;
3209 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3212 for (i = 0; i < RX_MAXCALLS; i++) {
3213 tcall = aconn->call[i];
3215 if ((tcall->state == RX_STATE_PRECALL)
3216 || (tcall->state == RX_STATE_ACTIVE))
3218 if ((tcall->mode == RX_MODE_SENDING)
3219 || (tcall->mode == RX_MODE_RECEIVING))
3227 /* if this is one of the last few packets AND it wouldn't be used by the
3228 receiving call to immediately satisfy a read request, then drop it on
3229 the floor, since accepting it might prevent a lock-holding thread from
3230 making progress in its reading. If a call has been cleared while in
3231 the precall state then ignore all subsequent packets until the call
3232 is assigned to a thread. */
3235 TooLow(struct rx_packet *ap, struct rx_call *acall)
3239 MUTEX_ENTER(&rx_quota_mutex);
3240 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3241 && (acall->state == RX_STATE_PRECALL))
3242 || ((rx_nFreePackets < rxi_dataQuota + 2)
3243 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3244 && (acall->flags & RX_CALL_READER_WAIT)))) {
3247 MUTEX_EXIT(&rx_quota_mutex);
3253 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3255 struct rx_connection *conn = arg1;
3256 struct rx_call *acall = arg2;
3257 struct rx_call *call = acall;
3258 struct clock when, now;
3261 MUTEX_ENTER(&conn->conn_data_lock);
3262 conn->checkReachEvent = NULL;
3263 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3266 MUTEX_EXIT(&conn->conn_data_lock);
3270 MUTEX_ENTER(&conn->conn_call_lock);
3271 MUTEX_ENTER(&conn->conn_data_lock);
3272 for (i = 0; i < RX_MAXCALLS; i++) {
3273 struct rx_call *tc = conn->call[i];
3274 if (tc && tc->state == RX_STATE_PRECALL) {
3280 /* Indicate that rxi_CheckReachEvent is no longer running by
3281 * clearing the flag. Must be atomic under conn_data_lock to
3282 * avoid a new call slipping by: rxi_CheckConnReach holds
3283 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3285 conn->flags &= ~RX_CONN_ATTACHWAIT;
3286 MUTEX_EXIT(&conn->conn_data_lock);
3287 MUTEX_EXIT(&conn->conn_call_lock);
3292 MUTEX_ENTER(&call->lock);
3293 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3295 MUTEX_EXIT(&call->lock);
3297 clock_GetTime(&now);
3299 when.sec += RX_CHECKREACH_TIMEOUT;
3300 MUTEX_ENTER(&conn->conn_data_lock);
3301 if (!conn->checkReachEvent) {
3303 conn->checkReachEvent =
3304 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3307 MUTEX_EXIT(&conn->conn_data_lock);
3313 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3315 struct rx_service *service = conn->service;
3316 struct rx_peer *peer = conn->peer;
3317 afs_uint32 now, lastReach;
3319 if (service->checkReach == 0)
3323 MUTEX_ENTER(&peer->peer_lock);
3324 lastReach = peer->lastReachTime;
3325 MUTEX_EXIT(&peer->peer_lock);
3326 if (now - lastReach < RX_CHECKREACH_TTL)
3329 MUTEX_ENTER(&conn->conn_data_lock);
3330 if (conn->flags & RX_CONN_ATTACHWAIT) {
3331 MUTEX_EXIT(&conn->conn_data_lock);
3334 conn->flags |= RX_CONN_ATTACHWAIT;
3335 MUTEX_EXIT(&conn->conn_data_lock);
3336 if (!conn->checkReachEvent)
3337 rxi_CheckReachEvent(NULL, conn, call);
3342 /* try to attach call, if authentication is complete */
3344 TryAttach(struct rx_call *acall, osi_socket socket,
3345 int *tnop, struct rx_call **newcallp,
3348 struct rx_connection *conn = acall->conn;
3350 if (conn->type == RX_SERVER_CONNECTION
3351 && acall->state == RX_STATE_PRECALL) {
3352 /* Don't attach until we have any req'd. authentication. */
3353 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3354 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3355 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3356 /* Note: this does not necessarily succeed; there
3357 * may not any proc available
3360 rxi_ChallengeOn(acall->conn);
3365 /* A data packet has been received off the interface. This packet is
3366 * appropriate to the call (the call is in the right state, etc.). This
3367 * routine can return a packet to the caller, for re-use */
3370 rxi_ReceiveDataPacket(struct rx_call *call,
3371 struct rx_packet *np, int istack,
3372 osi_socket socket, afs_uint32 host, u_short port,
3373 int *tnop, struct rx_call **newcallp)
3375 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3380 afs_uint32 serial=0, flags=0;
3382 struct rx_packet *tnp;
3383 struct clock when, now;
3384 if (rx_stats_active)
3385 rx_MutexIncrement(rx_stats.dataPacketsRead, rx_stats_mutex);
3388 /* If there are no packet buffers, drop this new packet, unless we can find
3389 * packet buffers from inactive calls */
3391 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3392 MUTEX_ENTER(&rx_freePktQ_lock);
3393 rxi_NeedMorePackets = TRUE;
3394 MUTEX_EXIT(&rx_freePktQ_lock);
3395 if (rx_stats_active)
3396 rx_MutexIncrement(rx_stats.noPacketBuffersOnRead, rx_stats_mutex);
3397 call->rprev = np->header.serial;
3398 rxi_calltrace(RX_TRACE_DROP, call);
3399 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems", np));
3401 rxi_ClearReceiveQueue(call);
3402 clock_GetTime(&now);
3404 clock_Add(&when, &rx_softAckDelay);
3405 if (!call->delayedAckEvent
3406 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3407 rxevent_Cancel(call->delayedAckEvent, call,
3408 RX_CALL_REFCOUNT_DELAY);
3409 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3410 call->delayedAckEvent =
3411 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3413 /* we've damaged this call already, might as well do it in. */
3419 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3420 * packet is one of several packets transmitted as a single
3421 * datagram. Do not send any soft or hard acks until all packets
3422 * in a jumbogram have been processed. Send negative acks right away.
3424 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3425 /* tnp is non-null when there are more packets in the
3426 * current jumbo gram */
3433 seq = np->header.seq;
3434 serial = np->header.serial;
3435 flags = np->header.flags;
3437 /* If the call is in an error state, send an abort message */
3439 return rxi_SendCallAbort(call, np, istack, 0);
3441 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3442 * AFS 3.5 jumbogram. */
3443 if (flags & RX_JUMBO_PACKET) {
3444 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3449 if (np->header.spare != 0) {
3450 MUTEX_ENTER(&call->conn->conn_data_lock);
3451 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3452 MUTEX_EXIT(&call->conn->conn_data_lock);
3455 /* The usual case is that this is the expected next packet */
3456 if (seq == call->rnext) {
3458 /* Check to make sure it is not a duplicate of one already queued */
3459 if (queue_IsNotEmpty(&call->rq)
3460 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3461 if (rx_stats_active)
3462 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3463 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate", np));
3464 rxevent_Cancel(call->delayedAckEvent, call,
3465 RX_CALL_REFCOUNT_DELAY);
3466 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3472 /* It's the next packet. Stick it on the receive queue
3473 * for this call. Set newPackets to make sure we wake
3474 * the reader once all packets have been processed */
3475 np->flags |= RX_PKTFLAG_RQ;
3476 queue_Prepend(&call->rq, np);
3477 #ifdef RXDEBUG_PACKET
3479 #endif /* RXDEBUG_PACKET */
3481 np = NULL; /* We can't use this anymore */
3484 /* If an ack is requested then set a flag to make sure we
3485 * send an acknowledgement for this packet */
3486 if (flags & RX_REQUEST_ACK) {
3487 ackNeeded = RX_ACK_REQUESTED;
3490 /* Keep track of whether we have received the last packet */
3491 if (flags & RX_LAST_PACKET) {
3492 call->flags |= RX_CALL_HAVE_LAST;
3496 /* Check whether we have all of the packets for this call */
3497 if (call->flags & RX_CALL_HAVE_LAST) {
3498 afs_uint32 tseq; /* temporary sequence number */
3499 struct rx_packet *tp; /* Temporary packet pointer */
3500 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3502 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3503 if (tseq != tp->header.seq)
3505 if (tp->header.flags & RX_LAST_PACKET) {
3506 call->flags |= RX_CALL_RECEIVE_DONE;
3513 /* Provide asynchronous notification for those who want it
3514 * (e.g. multi rx) */
3515 if (call->arrivalProc) {
3516 (*call->arrivalProc) (call, call->arrivalProcHandle,
3517 call->arrivalProcArg);
3518 call->arrivalProc = (void (*)())0;
3521 /* Update last packet received */
3524 /* If there is no server process serving this call, grab
3525 * one, if available. We only need to do this once. If a
3526 * server thread is available, this thread becomes a server
3527 * thread and the server thread becomes a listener thread. */
3529 TryAttach(call, socket, tnop, newcallp, 0);
3532 /* This is not the expected next packet. */
3534 /* Determine whether this is a new or old packet, and if it's
3535 * a new one, whether it fits into the current receive window.
3536 * Also figure out whether the packet was delivered in sequence.
3537 * We use the prev variable to determine whether the new packet
3538 * is the successor of its immediate predecessor in the
3539 * receive queue, and the missing flag to determine whether
3540 * any of this packets predecessors are missing. */
3542 afs_uint32 prev; /* "Previous packet" sequence number */
3543 struct rx_packet *tp; /* Temporary packet pointer */
3544 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3545 int missing; /* Are any predecessors missing? */
3547 /* If the new packet's sequence number has been sent to the
3548 * application already, then this is a duplicate */
3549 if (seq < call->rnext) {
3550 if (rx_stats_active)
3551 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3552 rxevent_Cancel(call->delayedAckEvent, call,
3553 RX_CALL_REFCOUNT_DELAY);
3554 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3560 /* If the sequence number is greater than what can be
3561 * accomodated by the current window, then send a negative
3562 * acknowledge and drop the packet */
3563 if ((call->rnext + call->rwind) <= seq) {
3564 rxevent_Cancel(call->delayedAckEvent, call,
3565 RX_CALL_REFCOUNT_DELAY);
3566 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3573 /* Look for the packet in the queue of old received packets */
3574 for (prev = call->rnext - 1, missing =
3575 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3576 /*Check for duplicate packet */
3577 if (seq == tp->header.seq) {
3578 if (rx_stats_active)
3579 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3580 rxevent_Cancel(call->delayedAckEvent, call,
3581 RX_CALL_REFCOUNT_DELAY);
3582 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3588 /* If we find a higher sequence packet, break out and
3589 * insert the new packet here. */
3590 if (seq < tp->header.seq)
3592 /* Check for missing packet */
3593 if (tp->header.seq != prev + 1) {
3597 prev = tp->header.seq;
3600 /* Keep track of whether we have received the last packet. */
3601 if (flags & RX_LAST_PACKET) {
3602 call->flags |= RX_CALL_HAVE_LAST;
3605 /* It's within the window: add it to the the receive queue.
3606 * tp is left by the previous loop either pointing at the
3607 * packet before which to insert the new packet, or at the
3608 * queue head if the queue is empty or the packet should be
3610 np->flags |= RX_PKTFLAG_RQ;
3611 #ifdef RXDEBUG_PACKET
3613 #endif /* RXDEBUG_PACKET */
3614 queue_InsertBefore(tp, np);
3618 /* Check whether we have all of the packets for this call */
3619 if ((call->flags & RX_CALL_HAVE_LAST)
3620 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3621 afs_uint32 tseq; /* temporary sequence number */
3624 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3625 if (tseq != tp->header.seq)
3627 if (tp->header.flags & RX_LAST_PACKET) {
3628 call->flags |= RX_CALL_RECEIVE_DONE;
3635 /* We need to send an ack of the packet is out of sequence,
3636 * or if an ack was requested by the peer. */
3637 if (seq != prev + 1 || missing) {
3638 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3639 } else if (flags & RX_REQUEST_ACK) {
3640 ackNeeded = RX_ACK_REQUESTED;
3643 /* Acknowledge the last packet for each call */
3644 if (flags & RX_LAST_PACKET) {
3655 * If the receiver is waiting for an iovec, fill the iovec
3656 * using the data from the receive queue */
3657 if (call->flags & RX_CALL_IOVEC_WAIT) {
3658 didHardAck = rxi_FillReadVec(call, serial);
3659 /* the call may have been aborted */
3668 /* Wakeup the reader if any */
3669 if ((call->flags & RX_CALL_READER_WAIT)
3670 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3671 || (call->iovNext >= call->iovMax)
3672 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3673 call->flags &= ~RX_CALL_READER_WAIT;
3674 #ifdef RX_ENABLE_LOCKS
3675 CV_BROADCAST(&call->cv_rq);
3677 osi_rxWakeup(&call->rq);
3683 * Send an ack when requested by the peer, or once every
3684 * rxi_SoftAckRate packets until the last packet has been
3685 * received. Always send a soft ack for the last packet in
3686 * the server's reply. */
3688 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3689 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3690 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3691 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3692 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3693 } else if (call->nSoftAcks) {
3694 clock_GetTime(&now);
3696 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3697 clock_Add(&when, &rx_lastAckDelay);
3699 clock_Add(&when, &rx_softAckDelay);
3701 if (!call->delayedAckEvent
3702 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3703 rxevent_Cancel(call->delayedAckEvent, call,
3704 RX_CALL_REFCOUNT_DELAY);
3705 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3706 call->delayedAckEvent =
3707 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3709 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3710 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3717 static void rxi_ComputeRate();
3721 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3723 struct rx_peer *peer = conn->peer;
3725 MUTEX_ENTER(&peer->peer_lock);
3726 peer->lastReachTime = clock_Sec();
3727 MUTEX_EXIT(&peer->peer_lock);
3729 MUTEX_ENTER(&conn->conn_data_lock);
3730 if (conn->flags & RX_CONN_ATTACHWAIT) {
3733 conn->flags &= ~RX_CONN_ATTACHWAIT;
3734 MUTEX_EXIT(&conn->conn_data_lock);
3736 for (i = 0; i < RX_MAXCALLS; i++) {
3737 struct rx_call *call = conn->call[i];
3740 MUTEX_ENTER(&call->lock);
3741 /* tnop can be null if newcallp is null */
3742 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3744 MUTEX_EXIT(&call->lock);
3748 MUTEX_EXIT(&conn->conn_data_lock);
3751 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
3753 rx_ack_reason(int reason)
3756 case RX_ACK_REQUESTED:
3758 case RX_ACK_DUPLICATE:
3760 case RX_ACK_OUT_OF_SEQUENCE:
3762 case RX_ACK_EXCEEDS_WINDOW:
3764 case RX_ACK_NOSPACE:
3768 case RX_ACK_PING_RESPONSE:
3781 /* rxi_ComputePeerNetStats
3783 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3784 * estimates (like RTT and throughput) based on ack packets. Caller
3785 * must ensure that the packet in question is the right one (i.e.
3786 * serial number matches).
3789 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3790 struct rx_ackPacket *ap, struct rx_packet *np)
3792 struct rx_peer *peer = call->conn->peer;
3794 /* Use RTT if not delayed by client and
3795 * ignore packets that were retransmitted. */
3796 if (!(p->flags & RX_PKTFLAG_ACKED) &&
3797 ap->reason != RX_ACK_DELAY &&
3798 clock_Eq(&p->timeSent, &p->firstSent))
3799 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3801 rxi_ComputeRate(peer, call, p, np, ap->reason);
3805 /* The real smarts of the whole thing. */
3807 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
3810 struct rx_ackPacket *ap;
3812 struct rx_packet *tp;
3813 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3814 struct rx_connection *conn = call->conn;
3815 struct rx_peer *peer = conn->peer;
3818 /* because there are CM's that are bogus, sending weird values for this. */
3819 afs_uint32 skew = 0;
3825 int newAckCount = 0;
3826 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3827 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3828 int pktsize = 0; /* Set if we need to update the peer mtu */
3830 if (rx_stats_active)
3831 rx_MutexIncrement(rx_stats.ackPacketsRead, rx_stats_mutex);
3832 ap = (struct rx_ackPacket *)rx_DataOf(np);
3833 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3835 return np; /* truncated ack packet */
3837 /* depends on ack packet struct */
3838 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3839 first = ntohl(ap->firstPacket);
3840 serial = ntohl(ap->serial);
3841 /* temporarily disabled -- needs to degrade over time
3842 * skew = ntohs(ap->maxSkew); */
3844 /* Ignore ack packets received out of order */
3845 if (first < call->tfirst) {
3849 if (np->header.flags & RX_SLOW_START_OK) {
3850 call->flags |= RX_CALL_SLOW_START_OK;
3853 if (ap->reason == RX_ACK_PING_RESPONSE)
3854 rxi_UpdatePeerReach(conn, call);
3856 if (conn->lastPacketSizeSeq) {
3857 MUTEX_ENTER(&conn->conn_data_lock);
3858 if (first >= conn->lastPacketSizeSeq) {
3859 pktsize = conn->lastPacketSize;
3860 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
3862 MUTEX_EXIT(&conn->conn_data_lock);
3863 MUTEX_ENTER(&peer->peer_lock);
3864 /* start somewhere */
3865 if (!peer->maxPacketSize)
3866 peer->maxPacketSize = np->length+RX_IPUDP_SIZE;
3868 if (pktsize > peer->maxPacketSize) {
3869 peer->maxPacketSize = pktsize;
3870 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
3871 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
3872 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
3875 MUTEX_EXIT(&peer->peer_lock);
3880 if (rxdebug_active) {
3884 len = _snprintf(msg, sizeof(msg),
3885 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3886 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3887 ntohl(ap->serial), ntohl(ap->previousPacket),
3888 (unsigned int)np->header.seq, (unsigned int)skew,
3889 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3893 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3894 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3898 OutputDebugString(msg);
3900 #else /* AFS_NT40_ENV */
3903 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3904 ap->reason, ntohl(ap->previousPacket),
3905 (unsigned int)np->header.seq, (unsigned int)serial,
3906 (unsigned int)skew, ntohl(ap->firstPacket));
3909 for (offset = 0; offset < nAcks; offset++)
3910 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3915 #endif /* AFS_NT40_ENV */
3918 /* Update the outgoing packet skew value to the latest value of
3919 * the peer's incoming packet skew value. The ack packet, of
3920 * course, could arrive out of order, but that won't affect things
3922 MUTEX_ENTER(&peer->peer_lock);
3923 peer->outPacketSkew = skew;
3925 /* Check for packets that no longer need to be transmitted, and
3926 * discard them. This only applies to packets positively
3927 * acknowledged as having been sent to the peer's upper level.
3928 * All other packets must be retained. So only packets with
3929 * sequence numbers < ap->firstPacket are candidates. */
3930 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3931 if (tp->header.seq >= first)
3933 call->tfirst = tp->header.seq + 1;
3934 rxi_ComputePeerNetStats(call, tp, ap, np);
3935 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3938 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3939 /* XXX Hack. Because we have to release the global rx lock when sending
3940 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3941 * in rxi_Start sending packets out because packets may move to the
3942 * freePacketQueue as result of being here! So we drop these packets until
3943 * we're safely out of the traversing. Really ugly!
3944 * To make it even uglier, if we're using fine grain locking, we can
3945 * set the ack bits in the packets and have rxi_Start remove the packets
3946 * when it's done transmitting.
3948 if (call->flags & RX_CALL_TQ_BUSY) {
3949 #ifdef RX_ENABLE_LOCKS
3950 tp->flags |= RX_PKTFLAG_ACKED;
3951 call->flags |= RX_CALL_TQ_SOME_ACKED;
3952 #else /* RX_ENABLE_LOCKS */
3954 #endif /* RX_ENABLE_LOCKS */
3956 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3959 tp->flags &= ~RX_PKTFLAG_TQ;
3960 #ifdef RXDEBUG_PACKET
3962 #endif /* RXDEBUG_PACKET */
3963 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3968 /* Give rate detector a chance to respond to ping requests */
3969 if (ap->reason == RX_ACK_PING_RESPONSE) {
3970 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3974 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3976 /* Now go through explicit acks/nacks and record the results in
3977 * the waiting packets. These are packets that can't be released
3978 * yet, even with a positive acknowledge. This positive
3979 * acknowledge only means the packet has been received by the
3980 * peer, not that it will be retained long enough to be sent to
3981 * the peer's upper level. In addition, reset the transmit timers
3982 * of any missing packets (those packets that must be missing
3983 * because this packet was out of sequence) */
3985 call->nSoftAcked = 0;
3986 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3987 /* Update round trip time if the ack was stimulated on receipt
3989 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3990 #ifdef RX_ENABLE_LOCKS
3991 if (tp->header.seq >= first)
3992 #endif /* RX_ENABLE_LOCKS */
3993 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3994 rxi_ComputePeerNetStats(call, tp, ap, np);
3996 /* Set the acknowledge flag per packet based on the
3997 * information in the ack packet. An acknowlegded packet can
3998 * be downgraded when the server has discarded a packet it
3999 * soacked previously, or when an ack packet is received
4000 * out of sequence. */
4001 if (tp->header.seq < first) {
4002 /* Implicit ack information */
4003 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4006 tp->flags |= RX_PKTFLAG_ACKED;
4007 } else if (tp->header.seq < first + nAcks) {
4008 /* Explicit ack information: set it in the packet appropriately */
4009 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4010 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4012 tp->flags |= RX_PKTFLAG_ACKED;
4019 } else /* RX_ACK_TYPE_NACK */ {
4020 tp->flags &= ~RX_PKTFLAG_ACKED;
4024 tp->flags &= ~RX_PKTFLAG_ACKED;
4029 * Following the suggestion of Phil Kern, we back off the peer's
4030 * timeout value for future packets until a successful response
4031 * is received for an initial transmission.
4033 if (missing && !backedOff) {
4034 struct clock c = peer->timeout;
4035 struct clock max_to = {3, 0};
4037 clock_Add(&peer->timeout, &c);
4038 if (clock_Gt(&peer->timeout, &max_to))
4039 peer->timeout = max_to;
4043 /* If packet isn't yet acked, and it has been transmitted at least
4044 * once, reset retransmit time using latest timeout
4045 * ie, this should readjust the retransmit timer for all outstanding
4046 * packets... So we don't just retransmit when we should know better*/
4048 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
4049 tp->retryTime = tp->timeSent;
4050 clock_Add(&tp->retryTime, &peer->timeout);
4051 /* shift by eight because one quarter-sec ~ 256 milliseconds */
4052 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
4056 /* If the window has been extended by this acknowledge packet,
4057 * then wakeup a sender waiting in alloc for window space, or try
4058 * sending packets now, if he's been sitting on packets due to
4059 * lack of window space */
4060 if (call->tnext < (call->tfirst + call->twind)) {
4061 #ifdef RX_ENABLE_LOCKS
4062 CV_SIGNAL(&call->cv_twind);
4064 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4065 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4066 osi_rxWakeup(&call->twind);
4069 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4070 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4074 /* if the ack packet has a receivelen field hanging off it,
4075 * update our state */
4076 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4079 /* If the ack packet has a "recommended" size that is less than
4080 * what I am using now, reduce my size to match */
4081 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4082 (int)sizeof(afs_int32), &tSize);
4083 tSize = (afs_uint32) ntohl(tSize);
4084 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4086 /* Get the maximum packet size to send to this peer */
4087 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4089 tSize = (afs_uint32) ntohl(tSize);
4090 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4091 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4093 /* sanity check - peer might have restarted with different params.
4094 * If peer says "send less", dammit, send less... Peer should never
4095 * be unable to accept packets of the size that prior AFS versions would
4096 * send without asking. */
4097 if (peer->maxMTU != tSize) {
4098 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4100 peer->maxMTU = tSize;
4101 peer->MTU = MIN(tSize, peer->MTU);
4102 call->MTU = MIN(call->MTU, tSize);
4105 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4108 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4109 (int)sizeof(afs_int32), &tSize);
4110 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4111 if (tSize < call->twind) { /* smaller than our send */
4112 call->twind = tSize; /* window, we must send less... */
4113 call->ssthresh = MIN(call->twind, call->ssthresh);
4114 call->conn->twind[call->channel] = call->twind;
4117 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4118 * network MTU confused with the loopback MTU. Calculate the
4119 * maximum MTU here for use in the slow start code below.
4121 maxMTU = peer->maxMTU;
4122 /* Did peer restart with older RX version? */
4123 if (peer->maxDgramPackets > 1) {
4124 peer->maxDgramPackets = 1;
4126 } else if (np->length >=
4127 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4130 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4131 sizeof(afs_int32), &tSize);
4132 tSize = (afs_uint32) ntohl(tSize);
4134 * As of AFS 3.5 we set the send window to match the receive window.
4136 if (tSize < call->twind) {
4137 call->twind = tSize;
4138 call->conn->twind[call->channel] = call->twind;
4139 call->ssthresh = MIN(call->twind, call->ssthresh);
4140 } else if (tSize > call->twind) {
4141 call->twind = tSize;
4142 call->conn->twind[call->channel] = call->twind;
4146 * As of AFS 3.5, a jumbogram is more than one fixed size
4147 * packet transmitted in a single UDP datagram. If the remote
4148 * MTU is smaller than our local MTU then never send a datagram
4149 * larger than the natural MTU.
4152 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4153 (int)sizeof(afs_int32), &tSize);
4154 maxDgramPackets = (afs_uint32) ntohl(tSize);
4155 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4157 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4158 maxDgramPackets = MIN(maxDgramPackets, tSize);
4159 if (maxDgramPackets > 1) {
4160 peer->maxDgramPackets = maxDgramPackets;
4161 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4163 peer->maxDgramPackets = 1;
4164 call->MTU = peer->natMTU;
4166 } else if (peer->maxDgramPackets > 1) {
4167 /* Restarted with lower version of RX */
4168 peer->maxDgramPackets = 1;
4170 } else if (peer->maxDgramPackets > 1
4171 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4172 /* Restarted with lower version of RX */
4173 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4174 peer->natMTU = OLD_MAX_PACKET_SIZE;
4175 peer->MTU = OLD_MAX_PACKET_SIZE;
4176 peer->maxDgramPackets = 1;
4177 peer->nDgramPackets = 1;
4179 call->MTU = OLD_MAX_PACKET_SIZE;
4184 * Calculate how many datagrams were successfully received after
4185 * the first missing packet and adjust the negative ack counter
4190 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4191 if (call->nNacks < nNacked) {
4192 call->nNacks = nNacked;
4195 call->nAcks += newAckCount;
4199 if (call->flags & RX_CALL_FAST_RECOVER) {
4201 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4203 call->flags &= ~RX_CALL_FAST_RECOVER;
4204 call->cwind = call->nextCwind;
4205 call->nextCwind = 0;
4208 call->nCwindAcks = 0;
4209 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4210 /* Three negative acks in a row trigger congestion recovery */
4211 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4212 MUTEX_EXIT(&peer->peer_lock);
4213 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4214 /* someone else is waiting to start recovery */
4217 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4218 rxi_WaitforTQBusy(call);
4219 MUTEX_ENTER(&peer->peer_lock);
4220 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4221 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4222 call->flags |= RX_CALL_FAST_RECOVER;
4223 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4225 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4226 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4227 call->nextCwind = call->ssthresh;
4230 peer->MTU = call->MTU;
4231 peer->cwind = call->nextCwind;
4232 peer->nDgramPackets = call->nDgramPackets;
4234 call->congestSeq = peer->congestSeq;
4235 /* Reset the resend times on the packets that were nacked
4236 * so we will retransmit as soon as the window permits*/
4237 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4239 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4240 clock_Zero(&tp->retryTime);
4242 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4247 /* If cwind is smaller than ssthresh, then increase
4248 * the window one packet for each ack we receive (exponential
4250 * If cwind is greater than or equal to ssthresh then increase
4251 * the congestion window by one packet for each cwind acks we
4252 * receive (linear growth). */
4253 if (call->cwind < call->ssthresh) {
4255 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4256 call->nCwindAcks = 0;
4258 call->nCwindAcks += newAckCount;
4259 if (call->nCwindAcks >= call->cwind) {
4260 call->nCwindAcks = 0;
4261 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4265 * If we have received several acknowledgements in a row then
4266 * it is time to increase the size of our datagrams
4268 if ((int)call->nAcks > rx_nDgramThreshold) {
4269 if (peer->maxDgramPackets > 1) {
4270 if (call->nDgramPackets < peer->maxDgramPackets) {
4271 call->nDgramPackets++;
4273 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4274 } else if (call->MTU < peer->maxMTU) {
4275 /* don't upgrade if we can't handle it */
4276 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4277 call->MTU = peer->ifMTU;
4279 call->MTU += peer->natMTU;
4280 call->MTU = MIN(call->MTU, peer->maxMTU);
4287 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4289 /* Servers need to hold the call until all response packets have
4290 * been acknowledged. Soft acks are good enough since clients
4291 * are not allowed to clear their receive queues. */
4292 if (call->state == RX_STATE_HOLD
4293 && call->tfirst + call->nSoftAcked >= call->tnext) {
4294 call->state = RX_STATE_DALLY;
4295 rxi_ClearTransmitQueue(call, 0);
4296 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4297 } else if (!queue_IsEmpty(&call->tq)) {
4298 rxi_Start(0, call, 0, istack);
4303 /* Received a response to a challenge packet */
4305 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4306 struct rx_packet *np, int istack)
4310 /* Ignore the packet if we're the client */
4311 if (conn->type == RX_CLIENT_CONNECTION)
4314 /* If already authenticated, ignore the packet (it's probably a retry) */
4315 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4318 /* Otherwise, have the security object evaluate the response packet */
4319 error = RXS_CheckResponse(conn->securityObject, conn, np);
4321 /* If the response is invalid, reset the connection, sending
4322 * an abort to the peer */
4326 rxi_ConnectionError(conn, error);
4327 MUTEX_ENTER(&conn->conn_data_lock);
4328 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4329 MUTEX_EXIT(&conn->conn_data_lock);
4332 /* If the response is valid, any calls waiting to attach
4333 * servers can now do so */
4336 for (i = 0; i < RX_MAXCALLS; i++) {
4337 struct rx_call *call = conn->call[i];
4339 MUTEX_ENTER(&call->lock);
4340 if (call->state == RX_STATE_PRECALL)
4341 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4342 /* tnop can be null if newcallp is null */
4343 MUTEX_EXIT(&call->lock);
4347 /* Update the peer reachability information, just in case
4348 * some calls went into attach-wait while we were waiting
4349 * for authentication..
4351 rxi_UpdatePeerReach(conn, NULL);
4356 /* A client has received an authentication challenge: the security
4357 * object is asked to cough up a respectable response packet to send
4358 * back to the server. The server is responsible for retrying the
4359 * challenge if it fails to get a response. */
4362 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4363 struct rx_packet *np, int istack)
4367 /* Ignore the challenge if we're the server */
4368 if (conn->type == RX_SERVER_CONNECTION)
4371 /* Ignore the challenge if the connection is otherwise idle; someone's
4372 * trying to use us as an oracle. */
4373 if (!rxi_HasActiveCalls(conn))
4376 /* Send the security object the challenge packet. It is expected to fill
4377 * in the response. */
4378 error = RXS_GetResponse(conn->securityObject, conn, np);
4380 /* If the security object is unable to return a valid response, reset the
4381 * connection and send an abort to the peer. Otherwise send the response
4382 * packet to the peer connection. */
4384 rxi_ConnectionError(conn, error);
4385 MUTEX_ENTER(&conn->conn_data_lock);
4386 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4387 MUTEX_EXIT(&conn->conn_data_lock);
4389 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4390 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4396 /* Find an available server process to service the current request in
4397 * the given call structure. If one isn't available, queue up this
4398 * call so it eventually gets one */
4400 rxi_AttachServerProc(struct rx_call *call,
4401 osi_socket socket, int *tnop,
4402 struct rx_call **newcallp)
4404 struct rx_serverQueueEntry *sq;
4405 struct rx_service *service = call->conn->service;
4408 /* May already be attached */
4409 if (call->state == RX_STATE_ACTIVE)
4412 MUTEX_ENTER(&rx_serverPool_lock);
4414 haveQuota = QuotaOK(service);
4415 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4416 /* If there are no processes available to service this call,
4417 * put the call on the incoming call queue (unless it's
4418 * already on the queue).
4420 #ifdef RX_ENABLE_LOCKS
4422 ReturnToServerPool(service);
4423 #endif /* RX_ENABLE_LOCKS */
4425 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4426 call->flags |= RX_CALL_WAIT_PROC;
4427 MUTEX_ENTER(&rx_waiting_mutex);
4430 MUTEX_EXIT(&rx_waiting_mutex);
4431 rxi_calltrace(RX_CALL_ARRIVAL, call);
4432 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4433 queue_Append(&rx_incomingCallQueue, call);
4436 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4438 /* If hot threads are enabled, and both newcallp and sq->socketp
4439 * are non-null, then this thread will process the call, and the
4440 * idle server thread will start listening on this threads socket.
4443 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4446 *sq->socketp = socket;
4447 clock_GetTime(&call->startTime);
4448 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4452 if (call->flags & RX_CALL_WAIT_PROC) {
4453 /* Conservative: I don't think this should happen */
4454 call->flags &= ~RX_CALL_WAIT_PROC;
4455 if (queue_IsOnQueue(call)) {
4458 MUTEX_ENTER(&rx_waiting_mutex);
4460 MUTEX_EXIT(&rx_waiting_mutex);
4463 call->state = RX_STATE_ACTIVE;
4464 call->mode = RX_MODE_RECEIVING;
4465 #ifdef RX_KERNEL_TRACE
4467 int glockOwner = ISAFS_GLOCK();
4470 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4471 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4477 if (call->flags & RX_CALL_CLEARED) {
4478 /* send an ack now to start the packet flow up again */
4479 call->flags &= ~RX_CALL_CLEARED;
4480 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4482 #ifdef RX_ENABLE_LOCKS
4485 service->nRequestsRunning++;
4486 MUTEX_ENTER(&rx_quota_mutex);
4487 if (service->nRequestsRunning <= service->minProcs)
4490 MUTEX_EXIT(&rx_quota_mutex);
4494 MUTEX_EXIT(&rx_serverPool_lock);
4497 /* Delay the sending of an acknowledge event for a short while, while
4498 * a new call is being prepared (in the case of a client) or a reply
4499 * is being prepared (in the case of a server). Rather than sending
4500 * an ack packet, an ACKALL packet is sent. */
4502 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4504 #ifdef RX_ENABLE_LOCKS
4506 MUTEX_ENTER(&call->lock);
4507 call->delayedAckEvent = NULL;
4508 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4510 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4511 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4513 MUTEX_EXIT(&call->lock);
4514 #else /* RX_ENABLE_LOCKS */
4516 call->delayedAckEvent = NULL;
4517 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4518 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4519 #endif /* RX_ENABLE_LOCKS */
4523 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4525 struct rx_call *call = arg1;
4526 #ifdef RX_ENABLE_LOCKS
4528 MUTEX_ENTER(&call->lock);
4529 if (event == call->delayedAckEvent)
4530 call->delayedAckEvent = NULL;
4531 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4533 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4535 MUTEX_EXIT(&call->lock);
4536 #else /* RX_ENABLE_LOCKS */
4538 call->delayedAckEvent = NULL;
4539 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4540 #endif /* RX_ENABLE_LOCKS */
4544 #ifdef RX_ENABLE_LOCKS
4545 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4546 * clearing them out.
4549 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4551 struct rx_packet *p, *tp;
4554 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4555 p->flags |= RX_PKTFLAG_ACKED;
4559 call->flags |= RX_CALL_TQ_CLEARME;
4560 call->flags |= RX_CALL_TQ_SOME_ACKED;
4563 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4564 call->tfirst = call->tnext;
4565 call->nSoftAcked = 0;
4567 if (call->flags & RX_CALL_FAST_RECOVER) {
4568 call->flags &= ~RX_CALL_FAST_RECOVER;
4569 call->cwind = call->nextCwind;
4570 call->nextCwind = 0;
4573 CV_SIGNAL(&call->cv_twind);
4575 #endif /* RX_ENABLE_LOCKS */
4577 /* Clear out the transmit queue for the current call (all packets have
4578 * been received by peer) */
4580 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4582 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4583 struct rx_packet *p, *tp;
4585 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4587 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4588 p->flags |= RX_PKTFLAG_ACKED;
4592 call->flags |= RX_CALL_TQ_CLEARME;
4593 call->flags |= RX_CALL_TQ_SOME_ACKED;
4596 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4597 #ifdef RXDEBUG_PACKET
4599 #endif /* RXDEBUG_PACKET */
4600 rxi_FreePackets(0, &call->tq);
4601 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
4602 #ifdef RX_ENABLE_LOCKS
4603 CV_BROADCAST(&call->cv_tq);
4604 #else /* RX_ENABLE_LOCKS */
4605 osi_rxWakeup(&call->tq);
4606 #endif /* RX_ENABLE_LOCKS */
4608 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4609 call->flags &= ~RX_CALL_TQ_CLEARME;
4611 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4613 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4614 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4615 call->nSoftAcked = 0;
4617 if (call->flags & RX_CALL_FAST_RECOVER) {
4618 call->flags &= ~RX_CALL_FAST_RECOVER;
4619 call->cwind = call->nextCwind;
4621 #ifdef RX_ENABLE_LOCKS
4622 CV_SIGNAL(&call->cv_twind);
4624 osi_rxWakeup(&call->twind);
4629 rxi_ClearReceiveQueue(struct rx_call *call)
4631 if (queue_IsNotEmpty(&call->rq)) {
4634 count = rxi_FreePackets(0, &call->rq);
4635 rx_packetReclaims += count;
4636 #ifdef RXDEBUG_PACKET
4638 if ( call->rqc != 0 )
4639 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0", call, call->rqc));
4641 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4643 if (call->state == RX_STATE_PRECALL) {
4644 call->flags |= RX_CALL_CLEARED;
4648 /* Send an abort packet for the specified call */
4650 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
4651 int istack, int force)
4654 struct clock when, now;
4659 /* Clients should never delay abort messages */
4660 if (rx_IsClientConn(call->conn))
4663 if (call->abortCode != call->error) {
4664 call->abortCode = call->error;
4665 call->abortCount = 0;
4668 if (force || rxi_callAbortThreshhold == 0
4669 || call->abortCount < rxi_callAbortThreshhold) {
4670 if (call->delayedAbortEvent) {
4671 rxevent_Cancel(call->delayedAbortEvent, call,
4672 RX_CALL_REFCOUNT_ABORT);
4674 error = htonl(call->error);
4677 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4678 (char *)&error, sizeof(error), istack);
4679 } else if (!call->delayedAbortEvent) {
4680 clock_GetTime(&now);
4682 clock_Addmsec(&when, rxi_callAbortDelay);
4683 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4684 call->delayedAbortEvent =
4685 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4690 /* Send an abort packet for the specified connection. Packet is an
4691 * optional pointer to a packet that can be used to send the abort.
4692 * Once the number of abort messages reaches the threshhold, an
4693 * event is scheduled to send the abort. Setting the force flag
4694 * overrides sending delayed abort messages.
4696 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4697 * to send the abort packet.
4700 rxi_SendConnectionAbort(struct rx_connection *conn,
4701 struct rx_packet *packet, int istack, int force)
4704 struct clock when, now;
4709 /* Clients should never delay abort messages */
4710 if (rx_IsClientConn(conn))
4713 if (force || rxi_connAbortThreshhold == 0
4714 || conn->abortCount < rxi_connAbortThreshhold) {
4715 if (conn->delayedAbortEvent) {
4716 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4718 error = htonl(conn->error);
4720 MUTEX_EXIT(&conn->conn_data_lock);
4722 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4723 RX_PACKET_TYPE_ABORT, (char *)&error,
4724 sizeof(error), istack);
4725 MUTEX_ENTER(&conn->conn_data_lock);
4726 } else if (!conn->delayedAbortEvent) {
4727 clock_GetTime(&now);
4729 clock_Addmsec(&when, rxi_connAbortDelay);
4730 conn->delayedAbortEvent =
4731 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4736 /* Associate an error all of the calls owned by a connection. Called
4737 * with error non-zero. This is only for really fatal things, like
4738 * bad authentication responses. The connection itself is set in
4739 * error at this point, so that future packets received will be
4742 rxi_ConnectionError(struct rx_connection *conn,
4748 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d", conn, error));
4750 MUTEX_ENTER(&conn->conn_data_lock);
4751 if (conn->challengeEvent)
4752 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4753 if (conn->natKeepAliveEvent)
4754 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
4755 if (conn->checkReachEvent) {
4756 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4757 conn->checkReachEvent = 0;
4758 conn->flags &= ~RX_CONN_ATTACHWAIT;
4761 MUTEX_EXIT(&conn->conn_data_lock);
4762 for (i = 0; i < RX_MAXCALLS; i++) {
4763 struct rx_call *call = conn->call[i];
4765 MUTEX_ENTER(&call->lock);
4766 rxi_CallError(call, error);
4767 MUTEX_EXIT(&call->lock);
4770 conn->error = error;
4771 if (rx_stats_active)
4772 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4777 rxi_CallError(struct rx_call *call, afs_int32 error)
4780 osirx_AssertMine(&call->lock, "rxi_CallError");
4782 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d", call, error, call->error));
4784 error = call->error;
4786 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4787 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4788 rxi_ResetCall(call, 0);
4791 rxi_ResetCall(call, 0);
4793 call->error = error;
4794 call->mode = RX_MODE_ERROR;
4797 /* Reset various fields in a call structure, and wakeup waiting
4798 * processes. Some fields aren't changed: state & mode are not
4799 * touched (these must be set by the caller), and bufptr, nLeft, and
4800 * nFree are not reset, since these fields are manipulated by
4801 * unprotected macros, and may only be reset by non-interrupting code.
4804 /* this code requires that call->conn be set properly as a pre-condition. */
4805 #endif /* ADAPT_WINDOW */
4808 rxi_ResetCall(struct rx_call *call, int newcall)
4811 struct rx_peer *peer;
4812 struct rx_packet *packet;
4814 osirx_AssertMine(&call->lock, "rxi_ResetCall");
4816 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
4818 /* Notify anyone who is waiting for asynchronous packet arrival */
4819 if (call->arrivalProc) {
4820 (*call->arrivalProc) (call, call->arrivalProcHandle,
4821 call->arrivalProcArg);
4822 call->arrivalProc = (void (*)())0;
4825 if (call->delayedAbortEvent) {
4826 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4827 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4829 rxi_SendCallAbort(call, packet, 0, 1);
4830 rxi_FreePacket(packet);
4835 * Update the peer with the congestion information in this call
4836 * so other calls on this connection can pick up where this call
4837 * left off. If the congestion sequence numbers don't match then
4838 * another call experienced a retransmission.
4840 peer = call->conn->peer;
4841 MUTEX_ENTER(&peer->peer_lock);
4843 if (call->congestSeq == peer->congestSeq) {
4844 peer->cwind = MAX(peer->cwind, call->cwind);
4845 peer->MTU = MAX(peer->MTU, call->MTU);
4846 peer->nDgramPackets =
4847 MAX(peer->nDgramPackets, call->nDgramPackets);
4850 call->abortCode = 0;
4851 call->abortCount = 0;
4853 if (peer->maxDgramPackets > 1) {
4854 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4856 call->MTU = peer->MTU;
4858 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4859 call->ssthresh = rx_maxSendWindow;
4860 call->nDgramPackets = peer->nDgramPackets;
4861 call->congestSeq = peer->congestSeq;
4862 MUTEX_EXIT(&peer->peer_lock);
4864 flags = call->flags;
4865 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4866 rxi_WaitforTQBusy(call);
4867 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4869 rxi_ClearTransmitQueue(call, 1);
4870 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4871 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4875 rxi_ClearReceiveQueue(call);
4876 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
4878 if (call->currentPacket) {
4879 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
4880 call->currentPacket->flags |= RX_PKTFLAG_IOVQ;
4881 queue_Prepend(&call->iovq, call->currentPacket);
4882 #ifdef RXDEBUG_PACKET
4884 #endif /* RXDEBUG_PACKET */
4885 call->currentPacket = (struct rx_packet *)0;
4887 call->curlen = call->nLeft = call->nFree = 0;
4889 #ifdef RXDEBUG_PACKET
4892 rxi_FreePackets(0, &call->iovq);
4895 call->twind = call->conn->twind[call->channel];
4896 call->rwind = call->conn->rwind[call->channel];
4897 call->nSoftAcked = 0;
4898 call->nextCwind = 0;
4901 call->nCwindAcks = 0;
4902 call->nSoftAcks = 0;
4903 call->nHardAcks = 0;
4905 call->tfirst = call->rnext = call->tnext = 1;
4907 call->lastAcked = 0;
4908 call->localStatus = call->remoteStatus = 0;
4910 if (flags & RX_CALL_READER_WAIT) {
4911 #ifdef RX_ENABLE_LOCKS
4912 CV_BROADCAST(&call->cv_rq);
4914 osi_rxWakeup(&call->rq);
4917 if (flags & RX_CALL_WAIT_PACKETS) {
4918 MUTEX_ENTER(&rx_freePktQ_lock);
4919 rxi_PacketsUnWait(); /* XXX */
4920 MUTEX_EXIT(&rx_freePktQ_lock);
4922 #ifdef RX_ENABLE_LOCKS
4923 CV_SIGNAL(&call->cv_twind);
4925 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4926 osi_rxWakeup(&call->twind);
4929 #ifdef RX_ENABLE_LOCKS
4930 /* The following ensures that we don't mess with any queue while some
4931 * other thread might also be doing so. The call_queue_lock field is
4932 * is only modified under the call lock. If the call is in the process
4933 * of being removed from a queue, the call is not locked until the
4934 * the queue lock is dropped and only then is the call_queue_lock field
4935 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4936 * Note that any other routine which removes a call from a queue has to
4937 * obtain the queue lock before examing the queue and removing the call.
4939 if (call->call_queue_lock) {
4940 MUTEX_ENTER(call->call_queue_lock);
4941 if (queue_IsOnQueue(call)) {
4943 if (flags & RX_CALL_WAIT_PROC) {
4945 MUTEX_ENTER(&rx_waiting_mutex);
4947 MUTEX_EXIT(&rx_waiting_mutex);
4950 MUTEX_EXIT(call->call_queue_lock);
4951 CLEAR_CALL_QUEUE_LOCK(call);
4953 #else /* RX_ENABLE_LOCKS */
4954 if (queue_IsOnQueue(call)) {
4956 if (flags & RX_CALL_WAIT_PROC)
4959 #endif /* RX_ENABLE_LOCKS */
4961 rxi_KeepAliveOff(call);
4962 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4965 /* Send an acknowledge for the indicated packet (seq,serial) of the
4966 * indicated call, for the indicated reason (reason). This
4967 * acknowledge will specifically acknowledge receiving the packet, and
4968 * will also specify which other packets for this call have been
4969 * received. This routine returns the packet that was used to the
4970 * caller. The caller is responsible for freeing it or re-using it.
4971 * This acknowledgement also returns the highest sequence number
4972 * actually read out by the higher level to the sender; the sender
4973 * promises to keep around packets that have not been read by the
4974 * higher level yet (unless, of course, the sender decides to abort
4975 * the call altogether). Any of p, seq, serial, pflags, or reason may
4976 * be set to zero without ill effect. That is, if they are zero, they
4977 * will not convey any information.
4978 * NOW there is a trailer field, after the ack where it will safely be
4979 * ignored by mundanes, which indicates the maximum size packet this
4980 * host can swallow. */
4982 struct rx_packet *optionalPacket; use to send ack (or null)
4983 int seq; Sequence number of the packet we are acking
4984 int serial; Serial number of the packet
4985 int pflags; Flags field from packet header
4986 int reason; Reason an acknowledge was prompted
4990 rxi_SendAck(struct rx_call *call,
4991 struct rx_packet *optionalPacket, int serial, int reason,
4994 struct rx_ackPacket *ap;
4995 struct rx_packet *rqp;
4996 struct rx_packet *nxp; /* For queue_Scan */
4997 struct rx_packet *p;
5000 #ifdef RX_ENABLE_TSFPQ
5001 struct rx_ts_info_t * rx_ts_info;
5005 * Open the receive window once a thread starts reading packets
5007 if (call->rnext > 1) {
5008 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5011 call->nHardAcks = 0;
5012 call->nSoftAcks = 0;
5013 if (call->rnext > call->lastAcked)
5014 call->lastAcked = call->rnext;
5018 rx_computelen(p, p->length); /* reset length, you never know */
5019 } /* where that's been... */
5020 #ifdef RX_ENABLE_TSFPQ
5022 RX_TS_INFO_GET(rx_ts_info);
5023 if ((p = rx_ts_info->local_special_packet)) {
5024 rx_computelen(p, p->length);
5025 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5026 rx_ts_info->local_special_packet = p;
5027 } else { /* We won't send the ack, but don't panic. */
5028 return optionalPacket;
5032 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5033 /* We won't send the ack, but don't panic. */
5034 return optionalPacket;
5039 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5042 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5043 #ifndef RX_ENABLE_TSFPQ
5044 if (!optionalPacket)
5047 return optionalPacket;
5049 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5050 if (rx_Contiguous(p) < templ) {
5051 #ifndef RX_ENABLE_TSFPQ
5052 if (!optionalPacket)
5055 return optionalPacket;
5060 /* MTUXXX failing to send an ack is very serious. We should */
5061 /* try as hard as possible to send even a partial ack; it's */
5062 /* better than nothing. */
5063 ap = (struct rx_ackPacket *)rx_DataOf(p);
5064 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5065 ap->reason = reason;
5067 /* The skew computation used to be bogus, I think it's better now. */
5068 /* We should start paying attention to skew. XXX */
5069 ap->serial = htonl(serial);
5070 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5072 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
5073 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5075 /* No fear of running out of ack packet here because there can only be at most
5076 * one window full of unacknowledged packets. The window size must be constrained
5077 * to be less than the maximum ack size, of course. Also, an ack should always
5078 * fit into a single packet -- it should not ever be fragmented. */
5079 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5080 if (!rqp || !call->rq.next
5081 || (rqp->header.seq > (call->rnext + call->rwind))) {
5082 #ifndef RX_ENABLE_TSFPQ
5083 if (!optionalPacket)
5086 rxi_CallError(call, RX_CALL_DEAD);
5087 return optionalPacket;
5090 while (rqp->header.seq > call->rnext + offset)
5091 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5092 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5094 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5095 #ifndef RX_ENABLE_TSFPQ
5096 if (!optionalPacket)
5099 rxi_CallError(call, RX_CALL_DEAD);
5100 return optionalPacket;
5105 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5107 /* these are new for AFS 3.3 */
5108 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5109 templ = htonl(templ);
5110 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5111 templ = htonl(call->conn->peer->ifMTU);
5112 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5113 sizeof(afs_int32), &templ);
5115 /* new for AFS 3.4 */
5116 templ = htonl(call->rwind);
5117 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5118 sizeof(afs_int32), &templ);
5120 /* new for AFS 3.5 */
5121 templ = htonl(call->conn->peer->ifDgramPackets);
5122 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5123 sizeof(afs_int32), &templ);
5125 p->header.serviceId = call->conn->serviceId;
5126 p->header.cid = (call->conn->cid | call->channel);
5127 p->header.callNumber = *call->callNumber;
5129 p->header.securityIndex = call->conn->securityIndex;
5130 p->header.epoch = call->conn->epoch;
5131 p->header.type = RX_PACKET_TYPE_ACK;
5132 p->header.flags = RX_SLOW_START_OK;
5133 if (reason == RX_ACK_PING) {
5134 p->header.flags |= RX_REQUEST_ACK;
5136 clock_GetTime(&call->pingRequestTime);
5139 if (call->conn->type == RX_CLIENT_CONNECTION)
5140 p->header.flags |= RX_CLIENT_INITIATED;
5144 if (rxdebug_active) {
5148 len = _snprintf(msg, sizeof(msg),
5149 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5150 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5151 ntohl(ap->serial), ntohl(ap->previousPacket),
5152 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5153 ap->nAcks, ntohs(ap->bufferSpace) );
5157 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5158 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5162 OutputDebugString(msg);
5164 #else /* AFS_NT40_ENV */
5166 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5167 ap->reason, ntohl(ap->previousPacket),
5168 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5170 for (offset = 0; offset < ap->nAcks; offset++)
5171 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5176 #endif /* AFS_NT40_ENV */
5179 int i, nbytes = p->length;
5181 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5182 if (nbytes <= p->wirevec[i].iov_len) {
5185 savelen = p->wirevec[i].iov_len;
5187 p->wirevec[i].iov_len = nbytes;
5189 rxi_Send(call, p, istack);
5190 p->wirevec[i].iov_len = savelen;
5194 nbytes -= p->wirevec[i].iov_len;
5197 if (rx_stats_active)
5198 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
5199 #ifndef RX_ENABLE_TSFPQ
5200 if (!optionalPacket)
5203 return optionalPacket; /* Return packet for re-use by caller */
5206 /* Send all of the packets in the list in single datagram */
5208 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
5209 int istack, int moreFlag, struct clock *now,
5210 struct clock *retryTime, int resending)
5215 struct rx_connection *conn = call->conn;
5216 struct rx_peer *peer = conn->peer;
5218 MUTEX_ENTER(&peer->peer_lock);
5221 peer->reSends += len;
5222 if (rx_stats_active)
5223 rx_MutexAdd(rx_stats.dataPacketsSent, len, rx_stats_mutex);
5224 MUTEX_EXIT(&peer->peer_lock);
5226 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
5230 /* Set the packet flags and schedule the resend events */
5231 /* Only request an ack for the last packet in the list */
5232 for (i = 0; i < len; i++) {
5233 list[i]->retryTime = *retryTime;
5234 if (list[i]->header.serial) {
5235 /* Exponentially backoff retry times */
5236 if (list[i]->backoff < MAXBACKOFF) {
5237 /* so it can't stay == 0 */
5238 list[i]->backoff = (list[i]->backoff << 1) + 1;
5241 clock_Addmsec(&(list[i]->retryTime),
5242 ((afs_uint32) list[i]->backoff) << 8);
5245 /* Wait a little extra for the ack on the last packet */
5246 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
5247 clock_Addmsec(&(list[i]->retryTime), 400);
5250 /* Record the time sent */
5251 list[i]->timeSent = *now;
5253 /* Ask for an ack on retransmitted packets, on every other packet
5254 * if the peer doesn't support slow start. Ask for an ack on every
5255 * packet until the congestion window reaches the ack rate. */
5256 if (list[i]->header.serial) {
5258 if (rx_stats_active)
5259 rx_MutexIncrement(rx_stats.dataPacketsReSent, rx_stats_mutex);
5261 /* improved RTO calculation- not Karn */
5262 list[i]->firstSent = *now;
5263 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5264 || (!(call->flags & RX_CALL_SLOW_START_OK)
5265 && (list[i]->header.seq & 1)))) {
5270 /* Tag this packet as not being the last in this group,
5271 * for the receiver's benefit */
5272 if (i < len - 1 || moreFlag) {
5273 list[i]->header.flags |= RX_MORE_PACKETS;
5276 /* Install the new retransmit time for the packet, and
5277 * record the time sent */
5278 list[i]->timeSent = *now;
5282 list[len - 1]->header.flags |= RX_REQUEST_ACK;
5285 /* Since we're about to send a data packet to the peer, it's
5286 * safe to nuke any scheduled end-of-packets ack */
5287 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5289 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5290 MUTEX_EXIT(&call->lock);
5292 rxi_SendPacketList(call, conn, list, len, istack);
5294 rxi_SendPacket(call, conn, list[0], istack);
5296 MUTEX_ENTER(&call->lock);
5297 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5299 /* Update last send time for this call (for keep-alive
5300 * processing), and for the connection (so that we can discover
5301 * idle connections) */
5302 conn->lastSendTime = call->lastSendTime = clock_Sec();
5303 /* Let a set of retransmits trigger an idle timeout */
5305 call->lastSendData = call->lastSendTime;
5308 /* When sending packets we need to follow these rules:
5309 * 1. Never send more than maxDgramPackets in a jumbogram.
5310 * 2. Never send a packet with more than two iovecs in a jumbogram.
5311 * 3. Never send a retransmitted packet in a jumbogram.
5312 * 4. Never send more than cwind/4 packets in a jumbogram
5313 * We always keep the last list we should have sent so we
5314 * can set the RX_MORE_PACKETS flags correctly.
5317 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5318 int istack, struct clock *now, struct clock *retryTime,
5321 int i, cnt, lastCnt = 0;
5322 struct rx_packet **listP, **lastP = 0;
5323 struct rx_peer *peer = call->conn->peer;
5324 int morePackets = 0;
5326 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
5327 /* Does the current packet force us to flush the current list? */
5329 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5330 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5332 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
5334 /* If the call enters an error state stop sending, or if
5335 * we entered congestion recovery mode, stop sending */
5336 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5344 /* Add the current packet to the list if it hasn't been acked.
5345 * Otherwise adjust the list pointer to skip the current packet. */
5346 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5348 /* Do we need to flush the list? */
5349 if (cnt >= (int)peer->maxDgramPackets
5350 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5351 || list[i]->header.serial
5352 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5354 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5355 retryTime, resending);
5356 /* If the call enters an error state stop sending, or if
5357 * we entered congestion recovery mode, stop sending */
5359 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5364 listP = &list[i + 1];
5369 osi_Panic("rxi_SendList error");
5371 listP = &list[i + 1];
5375 /* Send the whole list when the call is in receive mode, when
5376 * the call is in eof mode, when we are in fast recovery mode,
5377 * and when we have the last packet */
5378 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5379 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5380 || (call->flags & RX_CALL_FAST_RECOVER)) {
5381 /* Check for the case where the current list contains
5382 * an acked packet. Since we always send retransmissions
5383 * in a separate packet, we only need to check the first
5384 * packet in the list */
5385 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5389 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5390 retryTime, resending);
5391 /* If the call enters an error state stop sending, or if
5392 * we entered congestion recovery mode, stop sending */
5393 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5397 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5400 } else if (lastCnt > 0) {
5401 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5406 #ifdef RX_ENABLE_LOCKS
5407 /* Call rxi_Start, below, but with the call lock held. */
5409 rxi_StartUnlocked(struct rxevent *event,
5410 void *arg0, void *arg1, int istack)
5412 struct rx_call *call = arg0;
5414 MUTEX_ENTER(&call->lock);
5415 rxi_Start(event, call, arg1, istack);
5416 MUTEX_EXIT(&call->lock);
5418 #endif /* RX_ENABLE_LOCKS */
5420 /* This routine is called when new packets are readied for
5421 * transmission and when retransmission may be necessary, or when the
5422 * transmission window or burst count are favourable. This should be
5423 * better optimized for new packets, the usual case, now that we've
5424 * got rid of queues of send packets. XXXXXXXXXXX */
5426 rxi_Start(struct rxevent *event,
5427 void *arg0, void *arg1, int istack)
5429 struct rx_call *call = arg0;
5431 struct rx_packet *p;
5432 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5433 struct rx_peer *peer = call->conn->peer;
5434 struct clock now, usenow, retryTime;
5438 struct rx_packet **xmitList;
5441 /* If rxi_Start is being called as a result of a resend event,
5442 * then make sure that the event pointer is removed from the call
5443 * structure, since there is no longer a per-call retransmission
5445 if (event && event == call->resendEvent) {
5446 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5447 call->resendEvent = NULL;
5449 if (queue_IsEmpty(&call->tq)) {
5453 /* Timeouts trigger congestion recovery */
5454 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5455 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5456 /* someone else is waiting to start recovery */
5459 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5460 rxi_WaitforTQBusy(call);
5461 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5462 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5463 call->flags |= RX_CALL_FAST_RECOVER;
5464 if (peer->maxDgramPackets > 1) {
5465 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5467 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5469 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5470 call->nDgramPackets = 1;
5472 call->nextCwind = 1;
5475 MUTEX_ENTER(&peer->peer_lock);
5476 peer->MTU = call->MTU;
5477 peer->cwind = call->cwind;
5478 peer->nDgramPackets = 1;
5480 call->congestSeq = peer->congestSeq;
5481 MUTEX_EXIT(&peer->peer_lock);
5482 /* Clear retry times on packets. Otherwise, it's possible for
5483 * some packets in the queue to force resends at rates faster
5484 * than recovery rates.
5486 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5487 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5488 clock_Zero(&p->retryTime);
5493 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5494 if (rx_stats_active)
5495 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5500 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5501 /* Get clock to compute the re-transmit time for any packets
5502 * in this burst. Note, if we back off, it's reasonable to
5503 * back off all of the packets in the same manner, even if
5504 * some of them have been retransmitted more times than more
5506 * Do a dance to avoid blocking after setting now. */
5507 MUTEX_ENTER(&peer->peer_lock);
5508 retryTime = peer->timeout;
5509 MUTEX_EXIT(&peer->peer_lock);
5510 clock_GetTime(&now);
5511 clock_Add(&retryTime, &now);
5513 /* Send (or resend) any packets that need it, subject to
5514 * window restrictions and congestion burst control
5515 * restrictions. Ask for an ack on the last packet sent in
5516 * this burst. For now, we're relying upon the window being
5517 * considerably bigger than the largest number of packets that
5518 * are typically sent at once by one initial call to
5519 * rxi_Start. This is probably bogus (perhaps we should ask
5520 * for an ack when we're half way through the current
5521 * window?). Also, for non file transfer applications, this
5522 * may end up asking for an ack for every packet. Bogus. XXXX
5525 * But check whether we're here recursively, and let the other guy
5528 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5529 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5530 call->flags |= RX_CALL_TQ_BUSY;
5532 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5534 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5535 call->flags &= ~RX_CALL_NEED_START;
5536 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5538 maxXmitPackets = MIN(call->twind, call->cwind);
5539 xmitList = (struct rx_packet **)
5540 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
5541 /* XXXX else we must drop any mtx we hold */
5542 afs_osi_Alloc_NoSleep(maxXmitPackets * sizeof(struct rx_packet *));
5544 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5546 if (xmitList == NULL)
5547 osi_Panic("rxi_Start, failed to allocate xmit list");
5548 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5549 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5550 /* We shouldn't be sending packets if a thread is waiting
5551 * to initiate congestion recovery */
5552 dpf(("call %d waiting to initiate fast recovery\n",
5553 *(call->callNumber)));
5557 && (call->flags & RX_CALL_FAST_RECOVER)) {
5558 /* Only send one packet during fast recovery */
5559 dpf(("call %d restricted to one packet per send during fast recovery\n",
5560 *(call->callNumber)));
5563 if ((p->flags & RX_PKTFLAG_FREE)
5564 || (!queue_IsEnd(&call->tq, nxp)
5565 && (nxp->flags & RX_PKTFLAG_FREE))
5566 || (p == (struct rx_packet *)&rx_freePacketQueue)
5567 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5568 osi_Panic("rxi_Start: xmit queue clobbered");
5570 if (p->flags & RX_PKTFLAG_ACKED) {
5571 /* Since we may block, don't trust this */
5572 usenow.sec = usenow.usec = 0;
5573 if (rx_stats_active)
5574 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5575 continue; /* Ignore this packet if it has been acknowledged */
5578 /* Turn off all flags except these ones, which are the same
5579 * on each transmission */
5580 p->header.flags &= RX_PRESET_FLAGS;
5582 if (p->header.seq >=
5583 call->tfirst + MIN((int)call->twind,
5584 (int)(call->nSoftAcked +
5586 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5587 /* Note: if we're waiting for more window space, we can
5588 * still send retransmits; hence we don't return here, but
5589 * break out to schedule a retransmit event */
5590 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
5591 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
5596 /* Transmit the packet if it needs to be sent. */
5597 if (!clock_Lt(&now, &p->retryTime)) {
5598 if (nXmitPackets == maxXmitPackets) {
5599 rxi_SendXmitList(call, xmitList, nXmitPackets,
5600 istack, &now, &retryTime,
5602 osi_Free(xmitList, maxXmitPackets *
5603 sizeof(struct rx_packet *));
5606 dpf(("call %d xmit packet %"AFS_PTR_FMT" now %u.%06u retryTime %u.%06u nextRetry %u.%06u\n",
5607 *(call->callNumber), p,
5609 p->retryTime.sec, p->retryTime.usec,
5610 retryTime.sec, retryTime.usec));
5611 xmitList[nXmitPackets++] = p;
5615 /* xmitList now hold pointers to all of the packets that are
5616 * ready to send. Now we loop to send the packets */
5617 if (nXmitPackets > 0) {
5618 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5619 &now, &retryTime, resending);
5622 maxXmitPackets * sizeof(struct rx_packet *));
5624 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5626 * TQ references no longer protected by this flag; they must remain
5627 * protected by the global lock.
5629 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5630 call->flags &= ~RX_CALL_TQ_BUSY;
5631 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5632 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
5633 call, call->tqWaiters, call->flags));
5634 #ifdef RX_ENABLE_LOCKS
5635 osirx_AssertMine(&call->lock, "rxi_Start start");
5636 CV_BROADCAST(&call->cv_tq);
5637 #else /* RX_ENABLE_LOCKS */
5638 osi_rxWakeup(&call->tq);
5639 #endif /* RX_ENABLE_LOCKS */
5644 /* We went into the error state while sending packets. Now is
5645 * the time to reset the call. This will also inform the using
5646 * process that the call is in an error state.
5648 if (rx_stats_active)
5649 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5650 call->flags &= ~RX_CALL_TQ_BUSY;
5651 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5652 dpf(("call error %d while xmit %p has %d waiters and flags %d\n",
5653 call->error, call, call->tqWaiters, call->flags));
5654 #ifdef RX_ENABLE_LOCKS
5655 osirx_AssertMine(&call->lock, "rxi_Start middle");
5656 CV_BROADCAST(&call->cv_tq);
5657 #else /* RX_ENABLE_LOCKS */
5658 osi_rxWakeup(&call->tq);
5659 #endif /* RX_ENABLE_LOCKS */
5661 rxi_CallError(call, call->error);
5664 #ifdef RX_ENABLE_LOCKS
5665 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5667 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5668 /* Some packets have received acks. If they all have, we can clear
5669 * the transmit queue.
5672 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5673 if (p->header.seq < call->tfirst
5674 && (p->flags & RX_PKTFLAG_ACKED)) {
5676 p->flags &= ~RX_PKTFLAG_TQ;
5677 #ifdef RXDEBUG_PACKET
5685 call->flags |= RX_CALL_TQ_CLEARME;
5687 #endif /* RX_ENABLE_LOCKS */
5688 /* Don't bother doing retransmits if the TQ is cleared. */
5689 if (call->flags & RX_CALL_TQ_CLEARME) {
5690 rxi_ClearTransmitQueue(call, 1);
5692 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5695 /* Always post a resend event, if there is anything in the
5696 * queue, and resend is possible. There should be at least
5697 * one unacknowledged packet in the queue ... otherwise none
5698 * of these packets should be on the queue in the first place.
5700 if (call->resendEvent) {
5701 /* Cancel the existing event and post a new one */
5702 rxevent_Cancel(call->resendEvent, call,
5703 RX_CALL_REFCOUNT_RESEND);
5706 /* The retry time is the retry time on the first unacknowledged
5707 * packet inside the current window */
5709 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5710 /* Don't set timers for packets outside the window */
5711 if (p->header.seq >= call->tfirst + call->twind) {
5715 if (!(p->flags & RX_PKTFLAG_ACKED)
5716 && !clock_IsZero(&p->retryTime)) {
5718 retryTime = p->retryTime;
5723 /* Post a new event to re-run rxi_Start when retries may be needed */
5724 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5725 #ifdef RX_ENABLE_LOCKS
5726 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5728 rxevent_PostNow2(&retryTime, &usenow,
5730 (void *)call, 0, istack);
5731 #else /* RX_ENABLE_LOCKS */
5733 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5734 (void *)call, 0, istack);
5735 #endif /* RX_ENABLE_LOCKS */
5738 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5739 } while (call->flags & RX_CALL_NEED_START);
5741 * TQ references no longer protected by this flag; they must remain
5742 * protected by the global lock.
5744 call->flags &= ~RX_CALL_TQ_BUSY;
5745 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5746 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
5747 call, call->tqWaiters, call->flags));
5748 #ifdef RX_ENABLE_LOCKS
5749 osirx_AssertMine(&call->lock, "rxi_Start end");
5750 CV_BROADCAST(&call->cv_tq);
5751 #else /* RX_ENABLE_LOCKS */
5752 osi_rxWakeup(&call->tq);
5753 #endif /* RX_ENABLE_LOCKS */
5756 call->flags |= RX_CALL_NEED_START;
5758 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5760 if (call->resendEvent) {
5761 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5766 /* Also adjusts the keep alive parameters for the call, to reflect
5767 * that we have just sent a packet (so keep alives aren't sent
5770 rxi_Send(struct rx_call *call, struct rx_packet *p,
5773 struct rx_connection *conn = call->conn;
5775 /* Stamp each packet with the user supplied status */
5776 p->header.userStatus = call->localStatus;
5778 /* Allow the security object controlling this call's security to
5779 * make any last-minute changes to the packet */
5780 RXS_SendPacket(conn->securityObject, call, p);
5782 /* Since we're about to send SOME sort of packet to the peer, it's
5783 * safe to nuke any scheduled end-of-packets ack */
5784 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5786 /* Actually send the packet, filling in more connection-specific fields */
5787 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5788 MUTEX_EXIT(&call->lock);
5789 rxi_SendPacket(call, conn, p, istack);
5790 MUTEX_ENTER(&call->lock);
5791 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5793 /* Update last send time for this call (for keep-alive
5794 * processing), and for the connection (so that we can discover
5795 * idle connections) */
5796 conn->lastSendTime = call->lastSendTime = clock_Sec();
5797 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
5798 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5799 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
5800 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
5801 RX_ACK_PING_RESPONSE)))
5802 call->lastSendData = call->lastSendTime;
5806 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5807 * that things are fine. Also called periodically to guarantee that nothing
5808 * falls through the cracks (e.g. (error + dally) connections have keepalive
5809 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5811 * haveCTLock Set if calling from rxi_ReapConnections
5813 #ifdef RX_ENABLE_LOCKS
5815 rxi_CheckCall(struct rx_call *call, int haveCTLock)
5816 #else /* RX_ENABLE_LOCKS */
5818 rxi_CheckCall(struct rx_call *call)
5819 #endif /* RX_ENABLE_LOCKS */
5821 struct rx_connection *conn = call->conn;
5823 afs_uint32 deadTime;
5827 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5828 if (call->flags & RX_CALL_TQ_BUSY) {
5829 /* Call is active and will be reset by rxi_Start if it's
5830 * in an error state.
5835 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5837 (((afs_uint32) conn->secondsUntilDead << 10) +
5838 ((afs_uint32) conn->peer->rtt >> 3) +
5839 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5841 /* These are computed to the second (+- 1 second). But that's
5842 * good enough for these values, which should be a significant
5843 * number of seconds. */
5844 if (now > (call->lastReceiveTime + deadTime)) {
5845 if (call->state == RX_STATE_ACTIVE) {
5847 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5849 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5850 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5851 ip_stack_t *ipst = ns->netstack_ip;
5853 ire = ire_cache_lookup(call->conn->peer->host
5854 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5856 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5858 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5865 if (ire && ire->ire_max_frag > 0)
5866 rxi_SetPeerMtu(NULL, call->conn->peer->host, 0,
5868 #if defined(GLOBAL_NETSTACKID)
5872 #endif /* ADAPT_PMTU */
5873 cerror = RX_CALL_DEAD;
5876 #ifdef RX_ENABLE_LOCKS
5877 /* Cancel pending events */
5878 rxevent_Cancel(call->delayedAckEvent, call,
5879 RX_CALL_REFCOUNT_DELAY);
5880 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5881 rxevent_Cancel(call->keepAliveEvent, call,
5882 RX_CALL_REFCOUNT_ALIVE);
5883 if (call->refCount == 0) {
5884 rxi_FreeCall(call, haveCTLock);
5888 #else /* RX_ENABLE_LOCKS */
5891 #endif /* RX_ENABLE_LOCKS */
5893 /* Non-active calls are destroyed if they are not responding
5894 * to pings; active calls are simply flagged in error, so the
5895 * attached process can die reasonably gracefully. */
5897 /* see if we have a non-activity timeout */
5898 if (call->startWait && conn->idleDeadTime
5899 && ((call->startWait + conn->idleDeadTime) < now) &&
5900 (call->flags & RX_CALL_READER_WAIT)) {
5901 if (call->state == RX_STATE_ACTIVE) {
5902 rxi_CallError(call, RX_CALL_TIMEOUT);
5906 if (call->lastSendData && conn->idleDeadTime && (conn->idleDeadErr != 0)
5907 && ((call->lastSendData + conn->idleDeadTime) < now)) {
5908 if (call->state == RX_STATE_ACTIVE) {
5909 rxi_CallError(call, conn->idleDeadErr);
5913 /* see if we have a hard timeout */
5914 if (conn->hardDeadTime
5915 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5916 if (call->state == RX_STATE_ACTIVE)
5917 rxi_CallError(call, RX_CALL_TIMEOUT);
5922 if (call->conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT) {
5923 /* if we never succeeded, let the error pass out as-is */
5924 if (call->conn->peer->maxPacketSize)
5925 cerror = call->conn->msgsizeRetryErr;
5927 /* if we thought we could send more, perhaps things got worse */
5928 if (call->conn->peer->maxPacketSize > conn->lastPacketSize)
5929 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
5930 newmtu = MAX(call->conn->peer->maxPacketSize-RX_IPUDP_SIZE,
5931 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
5933 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
5935 /* minimum capped in SetPeerMtu */
5936 rxi_SetPeerMtu(call->conn->peer, 0, 0, newmtu);
5939 conn->lastPacketSize = 0;
5941 /* needed so ResetCall doesn't clobber us. */
5942 call->MTU = call->conn->peer->ifMTU;
5944 rxi_CallError(call, cerror);
5949 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
5951 struct rx_connection *conn = arg1;
5952 struct rx_header theader;
5954 struct sockaddr_in taddr;
5957 struct iovec tmpiov[2];
5960 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
5963 tp = &tbuffer[sizeof(struct rx_header)];
5964 taddr.sin_family = AF_INET;
5965 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
5966 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
5967 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
5968 taddr.sin_len = sizeof(struct sockaddr_in);
5970 memset(&theader, 0, sizeof(theader));
5971 theader.epoch = htonl(999);
5973 theader.callNumber = 0;
5976 theader.type = RX_PACKET_TYPE_VERSION;
5977 theader.flags = RX_LAST_PACKET;
5978 theader.serviceId = 0;
5980 memcpy(tbuffer, &theader, sizeof(theader));
5981 memcpy(tp, &a, sizeof(a));
5982 tmpiov[0].iov_base = tbuffer;
5983 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
5985 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
5987 MUTEX_ENTER(&conn->conn_data_lock);
5988 /* Only reschedule ourselves if the connection would not be destroyed */
5989 if (conn->refCount <= 1) {
5990 conn->natKeepAliveEvent = NULL;
5991 MUTEX_EXIT(&conn->conn_data_lock);
5992 rx_DestroyConnection(conn); /* drop the reference for this */
5994 conn->natKeepAliveEvent = NULL;
5995 conn->refCount--; /* drop the reference for this */
5996 rxi_ScheduleNatKeepAliveEvent(conn);
5997 MUTEX_EXIT(&conn->conn_data_lock);
6002 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6004 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6005 struct clock when, now;
6006 clock_GetTime(&now);
6008 when.sec += conn->secondsUntilNatPing;
6009 conn->refCount++; /* hold a reference for this */
6010 conn->natKeepAliveEvent =
6011 rxevent_PostNow(&when, &now, rxi_NatKeepAliveEvent, conn, 0);
6016 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6018 MUTEX_ENTER(&conn->conn_data_lock);
6019 conn->secondsUntilNatPing = seconds;
6021 rxi_ScheduleNatKeepAliveEvent(conn);
6022 MUTEX_EXIT(&conn->conn_data_lock);
6026 rxi_NatKeepAliveOn(struct rx_connection *conn)
6028 MUTEX_ENTER(&conn->conn_data_lock);
6029 rxi_ScheduleNatKeepAliveEvent(conn);
6030 MUTEX_EXIT(&conn->conn_data_lock);
6033 /* When a call is in progress, this routine is called occasionally to
6034 * make sure that some traffic has arrived (or been sent to) the peer.
6035 * If nothing has arrived in a reasonable amount of time, the call is
6036 * declared dead; if nothing has been sent for a while, we send a
6037 * keep-alive packet (if we're actually trying to keep the call alive)
6040 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6042 struct rx_call *call = arg1;
6043 struct rx_connection *conn;
6046 MUTEX_ENTER(&call->lock);
6047 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6048 if (event == call->keepAliveEvent)
6049 call->keepAliveEvent = NULL;
6052 #ifdef RX_ENABLE_LOCKS
6053 if (rxi_CheckCall(call, 0)) {
6054 MUTEX_EXIT(&call->lock);
6057 #else /* RX_ENABLE_LOCKS */
6058 if (rxi_CheckCall(call))
6060 #endif /* RX_ENABLE_LOCKS */
6062 /* Don't try to keep alive dallying calls */
6063 if (call->state == RX_STATE_DALLY) {
6064 MUTEX_EXIT(&call->lock);
6069 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6070 /* Don't try to send keepalives if there is unacknowledged data */
6071 /* the rexmit code should be good enough, this little hack
6072 * doesn't quite work XXX */
6073 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6075 rxi_ScheduleKeepAliveEvent(call);
6076 MUTEX_EXIT(&call->lock);
6081 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6083 if (!call->keepAliveEvent) {
6084 struct clock when, now;
6085 clock_GetTime(&now);
6087 when.sec += call->conn->secondsUntilPing;
6088 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6089 call->keepAliveEvent =
6090 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
6094 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6096 rxi_KeepAliveOn(struct rx_call *call)
6098 /* Pretend last packet received was received now--i.e. if another
6099 * packet isn't received within the keep alive time, then the call
6100 * will die; Initialize last send time to the current time--even
6101 * if a packet hasn't been sent yet. This will guarantee that a
6102 * keep-alive is sent within the ping time */
6103 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6104 rxi_ScheduleKeepAliveEvent(call);
6107 /* This routine is called to send connection abort messages
6108 * that have been delayed to throttle looping clients. */
6110 rxi_SendDelayedConnAbort(struct rxevent *event,
6111 void *arg1, void *unused)
6113 struct rx_connection *conn = arg1;
6116 struct rx_packet *packet;
6118 MUTEX_ENTER(&conn->conn_data_lock);
6119 conn->delayedAbortEvent = NULL;
6120 error = htonl(conn->error);
6122 MUTEX_EXIT(&conn->conn_data_lock);
6123 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6126 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6127 RX_PACKET_TYPE_ABORT, (char *)&error,
6129 rxi_FreePacket(packet);
6133 /* This routine is called to send call abort messages
6134 * that have been delayed to throttle looping clients. */
6136 rxi_SendDelayedCallAbort(struct rxevent *event,
6137 void *arg1, void *dummy)
6139 struct rx_call *call = arg1;
6142 struct rx_packet *packet;
6144 MUTEX_ENTER(&call->lock);
6145 call->delayedAbortEvent = NULL;
6146 error = htonl(call->error);
6148 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6151 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6152 (char *)&error, sizeof(error), 0);
6153 rxi_FreePacket(packet);
6155 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6156 MUTEX_EXIT(&call->lock);
6159 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6160 * seconds) to ask the client to authenticate itself. The routine
6161 * issues a challenge to the client, which is obtained from the
6162 * security object associated with the connection */
6164 rxi_ChallengeEvent(struct rxevent *event,
6165 void *arg0, void *arg1, int tries)
6167 struct rx_connection *conn = arg0;
6169 conn->challengeEvent = NULL;
6170 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6171 struct rx_packet *packet;
6172 struct clock when, now;
6175 /* We've failed to authenticate for too long.
6176 * Reset any calls waiting for authentication;
6177 * they are all in RX_STATE_PRECALL.
6181 MUTEX_ENTER(&conn->conn_call_lock);
6182 for (i = 0; i < RX_MAXCALLS; i++) {
6183 struct rx_call *call = conn->call[i];
6185 MUTEX_ENTER(&call->lock);
6186 if (call->state == RX_STATE_PRECALL) {
6187 rxi_CallError(call, RX_CALL_DEAD);
6188 rxi_SendCallAbort(call, NULL, 0, 0);
6190 MUTEX_EXIT(&call->lock);
6193 MUTEX_EXIT(&conn->conn_call_lock);
6197 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6199 /* If there's no packet available, do this later. */
6200 RXS_GetChallenge(conn->securityObject, conn, packet);
6201 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6202 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6203 rxi_FreePacket(packet);
6205 clock_GetTime(&now);
6207 when.sec += RX_CHALLENGE_TIMEOUT;
6208 conn->challengeEvent =
6209 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
6214 /* Call this routine to start requesting the client to authenticate
6215 * itself. This will continue until authentication is established,
6216 * the call times out, or an invalid response is returned. The
6217 * security object associated with the connection is asked to create
6218 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6219 * defined earlier. */
6221 rxi_ChallengeOn(struct rx_connection *conn)
6223 if (!conn->challengeEvent) {
6224 RXS_CreateChallenge(conn->securityObject, conn);
6225 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6230 /* Compute round trip time of the packet provided, in *rttp.
6233 /* rxi_ComputeRoundTripTime is called with peer locked. */
6234 /* sentp and/or peer may be null */
6236 rxi_ComputeRoundTripTime(struct rx_packet *p,
6237 struct clock *sentp,
6238 struct rx_peer *peer)
6240 struct clock thisRtt, *rttp = &thisRtt;
6244 clock_GetTime(rttp);
6246 if (clock_Lt(rttp, sentp)) {
6248 return; /* somebody set the clock back, don't count this time. */
6250 clock_Sub(rttp, sentp);
6251 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6252 p->header.callNumber, p, rttp->sec, rttp->usec));
6254 if (rttp->sec == 0 && rttp->usec == 0) {
6256 * The actual round trip time is shorter than the
6257 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6258 * Since we can't tell which at the moment we will assume 1ms.
6263 if (rx_stats_active) {
6264 MUTEX_ENTER(&rx_stats_mutex);
6265 if (clock_Lt(rttp, &rx_stats.minRtt))
6266 rx_stats.minRtt = *rttp;
6267 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
6268 if (rttp->sec > 60) {
6269 MUTEX_EXIT(&rx_stats_mutex);
6270 return; /* somebody set the clock ahead */
6272 rx_stats.maxRtt = *rttp;
6274 clock_Add(&rx_stats.totalRtt, rttp);
6275 rx_stats.nRttSamples++;
6276 MUTEX_EXIT(&rx_stats_mutex);
6279 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6281 /* Apply VanJacobson round-trip estimations */
6286 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
6287 * srtt is stored as fixed point with 3 bits after the binary
6288 * point (i.e., scaled by 8). The following magic is
6289 * equivalent to the smoothing algorithm in rfc793 with an
6290 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6291 * srtt'*8 = rtt + srtt*7
6292 * srtt'*8 = srtt*8 + rtt - srtt
6293 * srtt' = srtt + rtt/8 - srtt/8
6294 * srtt' = srtt + (rtt - srtt)/8
6297 delta = _8THMSEC(rttp) - peer->rtt;
6298 peer->rtt += (delta >> 3);
6301 * We accumulate a smoothed rtt variance (actually, a smoothed
6302 * mean difference), then set the retransmit timer to smoothed
6303 * rtt + 4 times the smoothed variance (was 2x in van's original
6304 * paper, but 4x works better for me, and apparently for him as
6306 * rttvar is stored as
6307 * fixed point with 2 bits after the binary point (scaled by
6308 * 4). The following is equivalent to rfc793 smoothing with
6309 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6310 * rttvar'*4 = rttvar*3 + |delta|
6311 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6312 * rttvar' = rttvar + |delta|/4 - rttvar/4
6313 * rttvar' = rttvar + (|delta| - rttvar)/4
6314 * This replaces rfc793's wired-in beta.
6315 * dev*4 = dev*4 + (|actual - expected| - dev)
6321 delta -= (peer->rtt_dev << 1);
6322 peer->rtt_dev += (delta >> 3);
6324 /* I don't have a stored RTT so I start with this value. Since I'm
6325 * probably just starting a call, and will be pushing more data down
6326 * this, I expect congestion to increase rapidly. So I fudge a
6327 * little, and I set deviance to half the rtt. In practice,
6328 * deviance tends to approach something a little less than
6329 * half the smoothed rtt. */
6330 peer->rtt = _8THMSEC(rttp) + 8;
6331 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
6333 /* the timeout is RTT + 4*MDEV but no less than rx_minPeerTimeout msec.
6334 * This is because one end or the other of these connections is usually
6335 * in a user process, and can be switched and/or swapped out. So on fast,
6336 * reliable networks, the timeout would otherwise be too short. */
6337 rtt_timeout = MAX(((peer->rtt >> 3) + peer->rtt_dev), rx_minPeerTimeout);
6338 clock_Zero(&(peer->timeout));
6339 clock_Addmsec(&(peer->timeout), rtt_timeout);
6341 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6342 p->header.callNumber, p, MSEC(rttp), peer->rtt >> 3, peer->rtt_dev >> 2, (peer->timeout.sec), (peer->timeout.usec)));
6346 /* Find all server connections that have not been active for a long time, and
6349 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6351 struct clock now, when;
6352 clock_GetTime(&now);
6354 /* Find server connection structures that haven't been used for
6355 * greater than rx_idleConnectionTime */
6357 struct rx_connection **conn_ptr, **conn_end;
6358 int i, havecalls = 0;
6359 MUTEX_ENTER(&rx_connHashTable_lock);
6360 for (conn_ptr = &rx_connHashTable[0], conn_end =
6361 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6363 struct rx_connection *conn, *next;
6364 struct rx_call *call;
6368 for (conn = *conn_ptr; conn; conn = next) {
6369 /* XXX -- Shouldn't the connection be locked? */
6372 for (i = 0; i < RX_MAXCALLS; i++) {
6373 call = conn->call[i];
6377 code = MUTEX_TRYENTER(&call->lock);
6380 #ifdef RX_ENABLE_LOCKS
6381 result = rxi_CheckCall(call, 1);
6382 #else /* RX_ENABLE_LOCKS */
6383 result = rxi_CheckCall(call);
6384 #endif /* RX_ENABLE_LOCKS */
6385 MUTEX_EXIT(&call->lock);
6387 /* If CheckCall freed the call, it might
6388 * have destroyed the connection as well,
6389 * which screws up the linked lists.
6395 if (conn->type == RX_SERVER_CONNECTION) {
6396 /* This only actually destroys the connection if
6397 * there are no outstanding calls */
6398 MUTEX_ENTER(&conn->conn_data_lock);
6399 if (!havecalls && !conn->refCount
6400 && ((conn->lastSendTime + rx_idleConnectionTime) <
6402 conn->refCount++; /* it will be decr in rx_DestroyConn */
6403 MUTEX_EXIT(&conn->conn_data_lock);
6404 #ifdef RX_ENABLE_LOCKS
6405 rxi_DestroyConnectionNoLock(conn);
6406 #else /* RX_ENABLE_LOCKS */
6407 rxi_DestroyConnection(conn);
6408 #endif /* RX_ENABLE_LOCKS */
6410 #ifdef RX_ENABLE_LOCKS
6412 MUTEX_EXIT(&conn->conn_data_lock);
6414 #endif /* RX_ENABLE_LOCKS */
6418 #ifdef RX_ENABLE_LOCKS
6419 while (rx_connCleanup_list) {
6420 struct rx_connection *conn;
6421 conn = rx_connCleanup_list;
6422 rx_connCleanup_list = rx_connCleanup_list->next;
6423 MUTEX_EXIT(&rx_connHashTable_lock);
6424 rxi_CleanupConnection(conn);
6425 MUTEX_ENTER(&rx_connHashTable_lock);
6427 MUTEX_EXIT(&rx_connHashTable_lock);
6428 #endif /* RX_ENABLE_LOCKS */
6431 /* Find any peer structures that haven't been used (haven't had an
6432 * associated connection) for greater than rx_idlePeerTime */
6434 struct rx_peer **peer_ptr, **peer_end;
6438 * Why do we need to hold the rx_peerHashTable_lock across
6439 * the incrementing of peer_ptr since the rx_peerHashTable
6440 * array is not changing? We don't.
6442 * By dropping the lock periodically we can permit other
6443 * activities to be performed while a rxi_ReapConnections
6444 * call is in progress. The goal of reap connections
6445 * is to clean up quickly without causing large amounts
6446 * of contention. Therefore, it is important that global
6447 * mutexes not be held for extended periods of time.
6449 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6450 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6452 struct rx_peer *peer, *next, *prev;
6454 MUTEX_ENTER(&rx_peerHashTable_lock);
6455 for (prev = peer = *peer_ptr; peer; peer = next) {
6457 code = MUTEX_TRYENTER(&peer->peer_lock);
6458 if ((code) && (peer->refCount == 0)
6459 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6460 rx_interface_stat_p rpc_stat, nrpc_stat;
6464 * now know that this peer object is one to be
6465 * removed from the hash table. Once it is removed
6466 * it can't be referenced by other threads.
6467 * Lets remove it first and decrement the struct
6468 * nPeerStructs count.
6470 if (peer == *peer_ptr) {
6476 if (rx_stats_active)
6477 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6480 * Now if we hold references on 'prev' and 'next'
6481 * we can safely drop the rx_peerHashTable_lock
6482 * while we destroy this 'peer' object.
6488 MUTEX_EXIT(&rx_peerHashTable_lock);
6490 MUTEX_EXIT(&peer->peer_lock);
6491 MUTEX_DESTROY(&peer->peer_lock);
6493 (&peer->rpcStats, rpc_stat, nrpc_stat,
6494 rx_interface_stat)) {
6495 unsigned int num_funcs;
6498 queue_Remove(&rpc_stat->queue_header);
6499 queue_Remove(&rpc_stat->all_peers);
6500 num_funcs = rpc_stat->stats[0].func_total;
6502 sizeof(rx_interface_stat_t) +
6503 rpc_stat->stats[0].func_total *
6504 sizeof(rx_function_entry_v1_t);
6506 rxi_Free(rpc_stat, space);
6508 MUTEX_ENTER(&rx_rpc_stats);
6509 rxi_rpc_peer_stat_cnt -= num_funcs;
6510 MUTEX_EXIT(&rx_rpc_stats);
6515 * Regain the rx_peerHashTable_lock and
6516 * decrement the reference count on 'prev'
6519 MUTEX_ENTER(&rx_peerHashTable_lock);
6526 MUTEX_EXIT(&peer->peer_lock);
6531 MUTEX_EXIT(&rx_peerHashTable_lock);
6535 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6536 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6537 * GC, just below. Really, we shouldn't have to keep moving packets from
6538 * one place to another, but instead ought to always know if we can
6539 * afford to hold onto a packet in its particular use. */
6540 MUTEX_ENTER(&rx_freePktQ_lock);
6541 if (rx_waitingForPackets) {
6542 rx_waitingForPackets = 0;
6543 #ifdef RX_ENABLE_LOCKS
6544 CV_BROADCAST(&rx_waitingForPackets_cv);
6546 osi_rxWakeup(&rx_waitingForPackets);
6549 MUTEX_EXIT(&rx_freePktQ_lock);
6552 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6553 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6557 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6558 * rx.h is sort of strange this is better. This is called with a security
6559 * object before it is discarded. Each connection using a security object has
6560 * its own refcount to the object so it won't actually be freed until the last
6561 * connection is destroyed.
6563 * This is the only rxs module call. A hold could also be written but no one
6567 rxs_Release(struct rx_securityClass *aobj)
6569 return RXS_Close(aobj);
6573 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6574 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6575 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6576 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6578 /* Adjust our estimate of the transmission rate to this peer, given
6579 * that the packet p was just acked. We can adjust peer->timeout and
6580 * call->twind. Pragmatically, this is called
6581 * only with packets of maximal length.
6582 * Called with peer and call locked.
6586 rxi_ComputeRate(struct rx_peer *peer, struct rx_call *call,
6587 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6589 afs_int32 xferSize, xferMs;
6593 /* Count down packets */
6594 if (peer->rateFlag > 0)
6596 /* Do nothing until we're enabled */
6597 if (peer->rateFlag != 0)
6602 /* Count only when the ack seems legitimate */
6603 switch (ackReason) {
6604 case RX_ACK_REQUESTED:
6606 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6610 case RX_ACK_PING_RESPONSE:
6611 if (p) /* want the response to ping-request, not data send */
6613 clock_GetTime(&newTO);
6614 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6615 clock_Sub(&newTO, &call->pingRequestTime);
6616 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6620 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6627 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %d.%06d, rtt %u, ps %u)",
6628 ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"),
6629 xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
6631 /* Track only packets that are big enough. */
6632 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6636 /* absorb RTT data (in milliseconds) for these big packets */
6637 if (peer->smRtt == 0) {
6638 peer->smRtt = xferMs;
6640 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6645 if (peer->countDown) {
6649 peer->countDown = 10; /* recalculate only every so often */
6651 /* In practice, we can measure only the RTT for full packets,
6652 * because of the way Rx acks the data that it receives. (If it's
6653 * smaller than a full packet, it often gets implicitly acked
6654 * either by the call response (from a server) or by the next call
6655 * (from a client), and either case confuses transmission times
6656 * with processing times.) Therefore, replace the above
6657 * more-sophisticated processing with a simpler version, where the
6658 * smoothed RTT is kept for full-size packets, and the time to
6659 * transmit a windowful of full-size packets is simply RTT *
6660 * windowSize. Again, we take two steps:
6661 - ensure the timeout is large enough for a single packet's RTT;
6662 - ensure that the window is small enough to fit in the desired timeout.*/
6664 /* First, the timeout check. */
6665 minTime = peer->smRtt;
6666 /* Get a reasonable estimate for a timeout period */
6668 newTO.sec = minTime / 1000;
6669 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6671 /* Increase the timeout period so that we can always do at least
6672 * one packet exchange */
6673 if (clock_Gt(&newTO, &peer->timeout)) {
6675 dpf(("CONG peer %lx/%u: timeout %d.%06d ==> %ld.%06d (rtt %u, ps %u)",
6676 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec,
6677 newTO.sec, newTO.usec, peer->smRtt, peer->packetSize));
6679 peer->timeout = newTO;
6682 /* Now, get an estimate for the transmit window size. */
6683 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6684 /* Now, convert to the number of full packets that could fit in a
6685 * reasonable fraction of that interval */
6686 minTime /= (peer->smRtt << 1);
6687 xferSize = minTime; /* (make a copy) */
6689 /* Now clamp the size to reasonable bounds. */
6692 else if (minTime > rx_Window)
6693 minTime = rx_Window;
6694 /* if (minTime != peer->maxWindow) {
6695 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6696 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6697 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6699 peer->maxWindow = minTime;
6700 elide... call->twind = minTime;
6704 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6705 * Discern this by calculating the timeout necessary for rx_Window
6707 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6708 /* calculate estimate for transmission interval in milliseconds */
6709 minTime = rx_Window * peer->smRtt;
6710 if (minTime < 1000) {
6711 dpf(("CONG peer %lx/%u: cut TO %d.%06d by 0.5 (rtt %u, ps %u)",
6712 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6713 peer->timeout.usec, peer->smRtt, peer->packetSize));
6715 newTO.sec = 0; /* cut back on timeout by half a second */
6716 newTO.usec = 500000;
6717 clock_Sub(&peer->timeout, &newTO);
6722 } /* end of rxi_ComputeRate */
6723 #endif /* ADAPT_WINDOW */
6731 #define TRACE_OPTION_RX_DEBUG 16
6739 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6740 0, KEY_QUERY_VALUE, &parmKey);
6741 if (code != ERROR_SUCCESS)
6744 dummyLen = sizeof(TraceOption);
6745 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6746 (BYTE *) &TraceOption, &dummyLen);
6747 if (code == ERROR_SUCCESS) {
6748 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
6750 RegCloseKey (parmKey);
6751 #endif /* AFS_NT40_ENV */
6756 rx_DebugOnOff(int on)
6760 rxdebug_active = on;
6766 rx_StatsOnOff(int on)
6769 rx_stats_active = on;
6774 /* Don't call this debugging routine directly; use dpf */
6776 rxi_DebugPrint(char *format, ...)
6785 va_start(ap, format);
6787 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6790 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
6792 if (msg[len-1] != '\n') {
6796 OutputDebugString(msg);
6803 va_start(ap, format);
6805 clock_GetTime(&now);
6806 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
6807 (unsigned int)now.usec);
6808 vfprintf(rx_Log, format, ap);
6817 * This function is used to process the rx_stats structure that is local
6818 * to a process as well as an rx_stats structure received from a remote
6819 * process (via rxdebug). Therefore, it needs to do minimal version
6823 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
6824 afs_int32 freePackets, char version)
6829 if (size != sizeof(struct rx_statistics)) {
6831 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
6832 size, sizeof(struct rx_statistics));
6835 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6838 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6839 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
6840 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6841 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6842 s->specialPktAllocFailures);
6844 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
6845 s->receivePktAllocFailures, s->sendPktAllocFailures,
6846 s->specialPktAllocFailures);
6850 " greedy %u, " "bogusReads %u (last from host %x), "
6851 "noPackets %u, " "noBuffers %u, " "selects %u, "
6852 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
6853 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6854 s->selects, s->sendSelects);
6856 fprintf(file, " packets read: ");
6857 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6858 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
6860 fprintf(file, "\n");
6863 " other read counters: data %u, " "ack %u, " "dup %u "
6864 "spurious %u " "dally %u\n", s->dataPacketsRead,
6865 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6866 s->ignorePacketDally);
6868 fprintf(file, " packets sent: ");
6869 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6870 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
6872 fprintf(file, "\n");
6875 " other send counters: ack %u, " "data %u (not resends), "
6876 "resends %u, " "pushed %u, " "acked&ignored %u\n",
6877 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6878 s->dataPacketsPushed, s->ignoreAckedPacket);
6881 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
6882 s->netSendFailures, (int)s->fatalErrors);
6884 if (s->nRttSamples) {
6885 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6886 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6888 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6889 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6893 " %d server connections, " "%d client connections, "
6894 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6895 s->nServerConns, s->nClientConns, s->nPeerStructs,
6896 s->nCallStructs, s->nFreeCallStructs);
6898 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6899 fprintf(file, " %d clock updates\n", clock_nUpdates);
6902 fprintf(file, "ERROR: compiled without RXDEBUG\n");
6906 /* for backward compatibility */
6908 rx_PrintStats(FILE * file)
6910 MUTEX_ENTER(&rx_stats_mutex);
6911 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6913 MUTEX_EXIT(&rx_stats_mutex);
6917 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6919 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
6920 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6921 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6924 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6925 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6926 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6929 " Packet size %d, " "max in packet skew %d, "
6930 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6931 (int)peer->outPacketSkew);
6935 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
6937 * This mutex protects the following static variables:
6941 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
6942 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
6944 #define LOCK_RX_DEBUG
6945 #define UNLOCK_RX_DEBUG
6946 #endif /* AFS_PTHREAD_ENV */
6950 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6951 u_char type, void *inputData, size_t inputLength,
6952 void *outputData, size_t outputLength)
6954 static afs_int32 counter = 100;
6955 time_t waitTime, waitCount, startTime;
6956 struct rx_header theader;
6959 struct timeval tv_now, tv_wake, tv_delta;
6960 struct sockaddr_in taddr, faddr;
6969 startTime = time(0);
6975 tp = &tbuffer[sizeof(struct rx_header)];
6976 taddr.sin_family = AF_INET;
6977 taddr.sin_port = remotePort;
6978 taddr.sin_addr.s_addr = remoteAddr;
6979 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6980 taddr.sin_len = sizeof(struct sockaddr_in);
6983 memset(&theader, 0, sizeof(theader));
6984 theader.epoch = htonl(999);
6986 theader.callNumber = htonl(counter);
6989 theader.type = type;
6990 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6991 theader.serviceId = 0;
6993 memcpy(tbuffer, &theader, sizeof(theader));
6994 memcpy(tp, inputData, inputLength);
6996 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6997 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6999 /* see if there's a packet available */
7000 gettimeofday(&tv_wake,0);
7001 tv_wake.tv_sec += waitTime;
7004 FD_SET(socket, &imask);
7005 tv_delta.tv_sec = tv_wake.tv_sec;
7006 tv_delta.tv_usec = tv_wake.tv_usec;
7007 gettimeofday(&tv_now, 0);
7009 if (tv_delta.tv_usec < tv_now.tv_usec) {
7011 tv_delta.tv_usec += 1000000;
7014 tv_delta.tv_usec -= tv_now.tv_usec;
7016 if (tv_delta.tv_sec < tv_now.tv_sec) {
7020 tv_delta.tv_sec -= tv_now.tv_sec;
7023 code = select(0, &imask, 0, 0, &tv_delta);
7024 #else /* AFS_NT40_ENV */
7025 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7026 #endif /* AFS_NT40_ENV */
7027 if (code == 1 && FD_ISSET(socket, &imask)) {
7028 /* now receive a packet */
7029 faddrLen = sizeof(struct sockaddr_in);
7031 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7032 (struct sockaddr *)&faddr, &faddrLen);
7035 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7036 if (counter == ntohl(theader.callNumber))
7044 /* see if we've timed out */
7052 code -= sizeof(struct rx_header);
7053 if (code > outputLength)
7054 code = outputLength;
7055 memcpy(outputData, tp, code);
7058 #endif /* RXDEBUG */
7061 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7062 afs_uint16 remotePort, struct rx_debugStats * stat,
7063 afs_uint32 * supportedValues)
7069 struct rx_debugIn in;
7071 *supportedValues = 0;
7072 in.type = htonl(RX_DEBUGI_GETSTATS);
7075 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7076 &in, sizeof(in), stat, sizeof(*stat));
7079 * If the call was successful, fixup the version and indicate
7080 * what contents of the stat structure are valid.
7081 * Also do net to host conversion of fields here.
7085 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7086 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7088 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7089 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7091 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7092 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7094 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7095 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7097 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7098 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7100 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7101 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7103 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7104 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7106 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7107 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7109 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7110 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7112 stat->nFreePackets = ntohl(stat->nFreePackets);
7113 stat->packetReclaims = ntohl(stat->packetReclaims);
7114 stat->callsExecuted = ntohl(stat->callsExecuted);
7115 stat->nWaiting = ntohl(stat->nWaiting);
7116 stat->idleThreads = ntohl(stat->idleThreads);
7117 stat->nWaited = ntohl(stat->nWaited);
7118 stat->nPackets = ntohl(stat->nPackets);
7125 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7126 afs_uint16 remotePort, struct rx_statistics * stat,
7127 afs_uint32 * supportedValues)
7133 struct rx_debugIn in;
7134 afs_int32 *lp = (afs_int32 *) stat;
7138 * supportedValues is currently unused, but added to allow future
7139 * versioning of this function.
7142 *supportedValues = 0;
7143 in.type = htonl(RX_DEBUGI_RXSTATS);
7145 memset(stat, 0, sizeof(*stat));
7147 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7148 &in, sizeof(in), stat, sizeof(*stat));
7153 * Do net to host conversion here
7156 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7165 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7166 afs_uint16 remotePort, size_t version_length,
7171 return MakeDebugCall(socket, remoteAddr, remotePort,
7172 RX_PACKET_TYPE_VERSION, a, 1, version,
7180 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7181 afs_uint16 remotePort, afs_int32 * nextConnection,
7182 int allConnections, afs_uint32 debugSupportedValues,
7183 struct rx_debugConn * conn,
7184 afs_uint32 * supportedValues)
7190 struct rx_debugIn in;
7194 * supportedValues is currently unused, but added to allow future
7195 * versioning of this function.
7198 *supportedValues = 0;
7199 if (allConnections) {
7200 in.type = htonl(RX_DEBUGI_GETALLCONN);
7202 in.type = htonl(RX_DEBUGI_GETCONN);
7204 in.index = htonl(*nextConnection);
7205 memset(conn, 0, sizeof(*conn));
7207 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7208 &in, sizeof(in), conn, sizeof(*conn));
7211 *nextConnection += 1;
7214 * Convert old connection format to new structure.
7217 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7218 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7219 #define MOVEvL(a) (conn->a = vL->a)
7221 /* any old or unrecognized version... */
7222 for (i = 0; i < RX_MAXCALLS; i++) {
7223 MOVEvL(callState[i]);
7224 MOVEvL(callMode[i]);
7225 MOVEvL(callFlags[i]);
7226 MOVEvL(callOther[i]);
7228 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7229 MOVEvL(secStats.type);
7230 MOVEvL(secStats.level);
7231 MOVEvL(secStats.flags);
7232 MOVEvL(secStats.expires);
7233 MOVEvL(secStats.packetsReceived);
7234 MOVEvL(secStats.packetsSent);
7235 MOVEvL(secStats.bytesReceived);
7236 MOVEvL(secStats.bytesSent);
7241 * Do net to host conversion here
7243 * I don't convert host or port since we are most likely
7244 * going to want these in NBO.
7246 conn->cid = ntohl(conn->cid);
7247 conn->serial = ntohl(conn->serial);
7248 for (i = 0; i < RX_MAXCALLS; i++) {
7249 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7251 conn->error = ntohl(conn->error);
7252 conn->secStats.flags = ntohl(conn->secStats.flags);
7253 conn->secStats.expires = ntohl(conn->secStats.expires);
7254 conn->secStats.packetsReceived =
7255 ntohl(conn->secStats.packetsReceived);
7256 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7257 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7258 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7259 conn->epoch = ntohl(conn->epoch);
7260 conn->natMTU = ntohl(conn->natMTU);
7267 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7268 afs_uint16 remotePort, afs_int32 * nextPeer,
7269 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7270 afs_uint32 * supportedValues)
7276 struct rx_debugIn in;
7279 * supportedValues is currently unused, but added to allow future
7280 * versioning of this function.
7283 *supportedValues = 0;
7284 in.type = htonl(RX_DEBUGI_GETPEER);
7285 in.index = htonl(*nextPeer);
7286 memset(peer, 0, sizeof(*peer));
7288 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7289 &in, sizeof(in), peer, sizeof(*peer));
7295 * Do net to host conversion here
7297 * I don't convert host or port since we are most likely
7298 * going to want these in NBO.
7300 peer->ifMTU = ntohs(peer->ifMTU);
7301 peer->idleWhen = ntohl(peer->idleWhen);
7302 peer->refCount = ntohs(peer->refCount);
7303 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7304 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7305 peer->rtt = ntohl(peer->rtt);
7306 peer->rtt_dev = ntohl(peer->rtt_dev);
7307 peer->timeout.sec = ntohl(peer->timeout.sec);
7308 peer->timeout.usec = ntohl(peer->timeout.usec);
7309 peer->nSent = ntohl(peer->nSent);
7310 peer->reSends = ntohl(peer->reSends);
7311 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7312 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7313 peer->rateFlag = ntohl(peer->rateFlag);
7314 peer->natMTU = ntohs(peer->natMTU);
7315 peer->maxMTU = ntohs(peer->maxMTU);
7316 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7317 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7318 peer->MTU = ntohs(peer->MTU);
7319 peer->cwind = ntohs(peer->cwind);
7320 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7321 peer->congestSeq = ntohs(peer->congestSeq);
7322 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7323 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7324 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7325 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7332 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7333 struct rx_debugPeer * peerStats)
7336 afs_int32 error = 1; /* default to "did not succeed" */
7337 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7339 MUTEX_ENTER(&rx_peerHashTable_lock);
7340 for(tp = rx_peerHashTable[hashValue];
7341 tp != NULL; tp = tp->next) {
7342 if (tp->host == peerHost)
7348 MUTEX_EXIT(&rx_peerHashTable_lock);
7352 MUTEX_ENTER(&tp->peer_lock);
7353 peerStats->host = tp->host;
7354 peerStats->port = tp->port;
7355 peerStats->ifMTU = tp->ifMTU;
7356 peerStats->idleWhen = tp->idleWhen;
7357 peerStats->refCount = tp->refCount;
7358 peerStats->burstSize = tp->burstSize;
7359 peerStats->burst = tp->burst;
7360 peerStats->burstWait.sec = tp->burstWait.sec;
7361 peerStats->burstWait.usec = tp->burstWait.usec;
7362 peerStats->rtt = tp->rtt;
7363 peerStats->rtt_dev = tp->rtt_dev;
7364 peerStats->timeout.sec = tp->timeout.sec;
7365 peerStats->timeout.usec = tp->timeout.usec;
7366 peerStats->nSent = tp->nSent;
7367 peerStats->reSends = tp->reSends;
7368 peerStats->inPacketSkew = tp->inPacketSkew;
7369 peerStats->outPacketSkew = tp->outPacketSkew;
7370 peerStats->rateFlag = tp->rateFlag;
7371 peerStats->natMTU = tp->natMTU;
7372 peerStats->maxMTU = tp->maxMTU;
7373 peerStats->maxDgramPackets = tp->maxDgramPackets;
7374 peerStats->ifDgramPackets = tp->ifDgramPackets;
7375 peerStats->MTU = tp->MTU;
7376 peerStats->cwind = tp->cwind;
7377 peerStats->nDgramPackets = tp->nDgramPackets;
7378 peerStats->congestSeq = tp->congestSeq;
7379 peerStats->bytesSent.high = tp->bytesSent.high;
7380 peerStats->bytesSent.low = tp->bytesSent.low;
7381 peerStats->bytesReceived.high = tp->bytesReceived.high;
7382 peerStats->bytesReceived.low = tp->bytesReceived.low;
7383 MUTEX_EXIT(&tp->peer_lock);
7385 MUTEX_ENTER(&rx_peerHashTable_lock);
7388 MUTEX_EXIT(&rx_peerHashTable_lock);
7396 struct rx_serverQueueEntry *np;
7399 struct rx_call *call;
7400 struct rx_serverQueueEntry *sq;
7404 if (rxinit_status == 1) {
7406 return; /* Already shutdown. */
7410 #ifndef AFS_PTHREAD_ENV
7411 FD_ZERO(&rx_selectMask);
7412 #endif /* AFS_PTHREAD_ENV */
7413 rxi_dataQuota = RX_MAX_QUOTA;
7414 #ifndef AFS_PTHREAD_ENV
7416 #endif /* AFS_PTHREAD_ENV */
7419 #ifndef AFS_PTHREAD_ENV
7420 #ifndef AFS_USE_GETTIMEOFDAY
7422 #endif /* AFS_USE_GETTIMEOFDAY */
7423 #endif /* AFS_PTHREAD_ENV */
7425 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7426 call = queue_First(&rx_freeCallQueue, rx_call);
7428 rxi_Free(call, sizeof(struct rx_call));
7431 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7432 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7438 struct rx_peer **peer_ptr, **peer_end;
7439 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7440 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7442 struct rx_peer *peer, *next;
7444 MUTEX_ENTER(&rx_peerHashTable_lock);
7445 for (peer = *peer_ptr; peer; peer = next) {
7446 rx_interface_stat_p rpc_stat, nrpc_stat;
7449 MUTEX_ENTER(&rx_rpc_stats);
7450 MUTEX_ENTER(&peer->peer_lock);
7452 (&peer->rpcStats, rpc_stat, nrpc_stat,
7453 rx_interface_stat)) {
7454 unsigned int num_funcs;
7457 queue_Remove(&rpc_stat->queue_header);
7458 queue_Remove(&rpc_stat->all_peers);
7459 num_funcs = rpc_stat->stats[0].func_total;
7461 sizeof(rx_interface_stat_t) +
7462 rpc_stat->stats[0].func_total *
7463 sizeof(rx_function_entry_v1_t);
7465 rxi_Free(rpc_stat, space);
7467 /* rx_rpc_stats must be held */
7468 rxi_rpc_peer_stat_cnt -= num_funcs;
7470 MUTEX_EXIT(&peer->peer_lock);
7471 MUTEX_EXIT(&rx_rpc_stats);
7475 if (rx_stats_active)
7476 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
7478 MUTEX_EXIT(&rx_peerHashTable_lock);
7481 for (i = 0; i < RX_MAX_SERVICES; i++) {
7483 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7485 for (i = 0; i < rx_hashTableSize; i++) {
7486 struct rx_connection *tc, *ntc;
7487 MUTEX_ENTER(&rx_connHashTable_lock);
7488 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7490 for (j = 0; j < RX_MAXCALLS; j++) {
7492 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7495 rxi_Free(tc, sizeof(*tc));
7497 MUTEX_EXIT(&rx_connHashTable_lock);
7500 MUTEX_ENTER(&freeSQEList_lock);
7502 while ((np = rx_FreeSQEList)) {
7503 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7504 MUTEX_DESTROY(&np->lock);
7505 rxi_Free(np, sizeof(*np));
7508 MUTEX_EXIT(&freeSQEList_lock);
7509 MUTEX_DESTROY(&freeSQEList_lock);
7510 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7511 MUTEX_DESTROY(&rx_connHashTable_lock);
7512 MUTEX_DESTROY(&rx_peerHashTable_lock);
7513 MUTEX_DESTROY(&rx_serverPool_lock);
7515 osi_Free(rx_connHashTable,
7516 rx_hashTableSize * sizeof(struct rx_connection *));
7517 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7519 UNPIN(rx_connHashTable,
7520 rx_hashTableSize * sizeof(struct rx_connection *));
7521 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7523 rxi_FreeAllPackets();
7525 MUTEX_ENTER(&rx_quota_mutex);
7526 rxi_dataQuota = RX_MAX_QUOTA;
7527 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7528 MUTEX_EXIT(&rx_quota_mutex);
7533 #ifdef RX_ENABLE_LOCKS
7535 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7537 if (!MUTEX_ISMINE(lockaddr))
7538 osi_Panic("Lock not held: %s", msg);
7540 #endif /* RX_ENABLE_LOCKS */
7545 * Routines to implement connection specific data.
7549 rx_KeyCreate(rx_destructor_t rtn)
7552 MUTEX_ENTER(&rxi_keyCreate_lock);
7553 key = rxi_keyCreate_counter++;
7554 rxi_keyCreate_destructor = (rx_destructor_t *)
7555 realloc((void *)rxi_keyCreate_destructor,
7556 (key + 1) * sizeof(rx_destructor_t));
7557 rxi_keyCreate_destructor[key] = rtn;
7558 MUTEX_EXIT(&rxi_keyCreate_lock);
7563 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7566 MUTEX_ENTER(&conn->conn_data_lock);
7567 if (!conn->specific) {
7568 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7569 for (i = 0; i < key; i++)
7570 conn->specific[i] = NULL;
7571 conn->nSpecific = key + 1;
7572 conn->specific[key] = ptr;
7573 } else if (key >= conn->nSpecific) {
7574 conn->specific = (void **)
7575 realloc(conn->specific, (key + 1) * sizeof(void *));
7576 for (i = conn->nSpecific; i < key; i++)
7577 conn->specific[i] = NULL;
7578 conn->nSpecific = key + 1;
7579 conn->specific[key] = ptr;
7581 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7582 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7583 conn->specific[key] = ptr;
7585 MUTEX_EXIT(&conn->conn_data_lock);
7589 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7592 MUTEX_ENTER(&svc->svc_data_lock);
7593 if (!svc->specific) {
7594 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
7595 for (i = 0; i < key; i++)
7596 svc->specific[i] = NULL;
7597 svc->nSpecific = key + 1;
7598 svc->specific[key] = ptr;
7599 } else if (key >= svc->nSpecific) {
7600 svc->specific = (void **)
7601 realloc(svc->specific, (key + 1) * sizeof(void *));
7602 for (i = svc->nSpecific; i < key; i++)
7603 svc->specific[i] = NULL;
7604 svc->nSpecific = key + 1;
7605 svc->specific[key] = ptr;
7607 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7608 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7609 svc->specific[key] = ptr;
7611 MUTEX_EXIT(&svc->svc_data_lock);
7615 rx_GetSpecific(struct rx_connection *conn, int key)
7618 MUTEX_ENTER(&conn->conn_data_lock);
7619 if (key >= conn->nSpecific)
7622 ptr = conn->specific[key];
7623 MUTEX_EXIT(&conn->conn_data_lock);
7628 rx_GetServiceSpecific(struct rx_service *svc, int key)
7631 MUTEX_ENTER(&svc->svc_data_lock);
7632 if (key >= svc->nSpecific)
7635 ptr = svc->specific[key];
7636 MUTEX_EXIT(&svc->svc_data_lock);
7641 #endif /* !KERNEL */
7644 * processStats is a queue used to store the statistics for the local
7645 * process. Its contents are similar to the contents of the rpcStats
7646 * queue on a rx_peer structure, but the actual data stored within
7647 * this queue contains totals across the lifetime of the process (assuming
7648 * the stats have not been reset) - unlike the per peer structures
7649 * which can come and go based upon the peer lifetime.
7652 static struct rx_queue processStats = { &processStats, &processStats };
7655 * peerStats is a queue used to store the statistics for all peer structs.
7656 * Its contents are the union of all the peer rpcStats queues.
7659 static struct rx_queue peerStats = { &peerStats, &peerStats };
7662 * rxi_monitor_processStats is used to turn process wide stat collection
7666 static int rxi_monitor_processStats = 0;
7669 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7672 static int rxi_monitor_peerStats = 0;
7675 * rxi_AddRpcStat - given all of the information for a particular rpc
7676 * call, create (if needed) and update the stat totals for the rpc.
7680 * IN stats - the queue of stats that will be updated with the new value
7682 * IN rxInterface - a unique number that identifies the rpc interface
7684 * IN currentFunc - the index of the function being invoked
7686 * IN totalFunc - the total number of functions in this interface
7688 * IN queueTime - the amount of time this function waited for a thread
7690 * IN execTime - the amount of time this function invocation took to execute
7692 * IN bytesSent - the number bytes sent by this invocation
7694 * IN bytesRcvd - the number bytes received by this invocation
7696 * IN isServer - if true, this invocation was made to a server
7698 * IN remoteHost - the ip address of the remote host
7700 * IN remotePort - the port of the remote host
7702 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7704 * INOUT counter - if a new stats structure is allocated, the counter will
7705 * be updated with the new number of allocated stat structures
7713 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
7714 afs_uint32 currentFunc, afs_uint32 totalFunc,
7715 struct clock *queueTime, struct clock *execTime,
7716 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
7717 afs_uint32 remoteHost, afs_uint32 remotePort,
7718 int addToPeerList, unsigned int *counter)
7721 rx_interface_stat_p rpc_stat, nrpc_stat;
7724 * See if there's already a structure for this interface
7727 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7728 if ((rpc_stat->stats[0].interfaceId == rxInterface)
7729 && (rpc_stat->stats[0].remote_is_server == isServer))
7734 * Didn't find a match so allocate a new structure and add it to the
7738 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
7739 || (rpc_stat->stats[0].interfaceId != rxInterface)
7740 || (rpc_stat->stats[0].remote_is_server != isServer)) {
7745 sizeof(rx_interface_stat_t) +
7746 totalFunc * sizeof(rx_function_entry_v1_t);
7748 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
7749 if (rpc_stat == NULL) {
7753 *counter += totalFunc;
7754 for (i = 0; i < totalFunc; i++) {
7755 rpc_stat->stats[i].remote_peer = remoteHost;
7756 rpc_stat->stats[i].remote_port = remotePort;
7757 rpc_stat->stats[i].remote_is_server = isServer;
7758 rpc_stat->stats[i].interfaceId = rxInterface;
7759 rpc_stat->stats[i].func_total = totalFunc;
7760 rpc_stat->stats[i].func_index = i;
7761 hzero(rpc_stat->stats[i].invocations);
7762 hzero(rpc_stat->stats[i].bytes_sent);
7763 hzero(rpc_stat->stats[i].bytes_rcvd);
7764 rpc_stat->stats[i].queue_time_sum.sec = 0;
7765 rpc_stat->stats[i].queue_time_sum.usec = 0;
7766 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7767 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7768 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7769 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7770 rpc_stat->stats[i].queue_time_max.sec = 0;
7771 rpc_stat->stats[i].queue_time_max.usec = 0;
7772 rpc_stat->stats[i].execution_time_sum.sec = 0;
7773 rpc_stat->stats[i].execution_time_sum.usec = 0;
7774 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7775 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7776 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7777 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7778 rpc_stat->stats[i].execution_time_max.sec = 0;
7779 rpc_stat->stats[i].execution_time_max.usec = 0;
7781 queue_Prepend(stats, rpc_stat);
7782 if (addToPeerList) {
7783 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7788 * Increment the stats for this function
7791 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7792 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7793 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7794 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7795 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7796 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7797 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7799 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7800 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7802 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7803 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7805 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7806 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7808 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7809 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7817 * rx_IncrementTimeAndCount - increment the times and count for a particular
7822 * IN peer - the peer who invoked the rpc
7824 * IN rxInterface - a unique number that identifies the rpc interface
7826 * IN currentFunc - the index of the function being invoked
7828 * IN totalFunc - the total number of functions in this interface
7830 * IN queueTime - the amount of time this function waited for a thread
7832 * IN execTime - the amount of time this function invocation took to execute
7834 * IN bytesSent - the number bytes sent by this invocation
7836 * IN bytesRcvd - the number bytes received by this invocation
7838 * IN isServer - if true, this invocation was made to a server
7846 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7847 afs_uint32 currentFunc, afs_uint32 totalFunc,
7848 struct clock *queueTime, struct clock *execTime,
7849 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7853 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7856 MUTEX_ENTER(&rx_rpc_stats);
7858 if (rxi_monitor_peerStats) {
7859 MUTEX_ENTER(&peer->peer_lock);
7860 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7861 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7862 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7863 MUTEX_EXIT(&peer->peer_lock);
7866 if (rxi_monitor_processStats) {
7867 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7868 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7869 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7872 MUTEX_EXIT(&rx_rpc_stats);
7877 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7881 * IN callerVersion - the rpc stat version of the caller.
7883 * IN count - the number of entries to marshall.
7885 * IN stats - pointer to stats to be marshalled.
7887 * OUT ptr - Where to store the marshalled data.
7894 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7895 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7901 * We only support the first version
7903 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7904 *(ptr++) = stats->remote_peer;
7905 *(ptr++) = stats->remote_port;
7906 *(ptr++) = stats->remote_is_server;
7907 *(ptr++) = stats->interfaceId;
7908 *(ptr++) = stats->func_total;
7909 *(ptr++) = stats->func_index;
7910 *(ptr++) = hgethi(stats->invocations);
7911 *(ptr++) = hgetlo(stats->invocations);
7912 *(ptr++) = hgethi(stats->bytes_sent);
7913 *(ptr++) = hgetlo(stats->bytes_sent);
7914 *(ptr++) = hgethi(stats->bytes_rcvd);
7915 *(ptr++) = hgetlo(stats->bytes_rcvd);
7916 *(ptr++) = stats->queue_time_sum.sec;
7917 *(ptr++) = stats->queue_time_sum.usec;
7918 *(ptr++) = stats->queue_time_sum_sqr.sec;
7919 *(ptr++) = stats->queue_time_sum_sqr.usec;
7920 *(ptr++) = stats->queue_time_min.sec;
7921 *(ptr++) = stats->queue_time_min.usec;
7922 *(ptr++) = stats->queue_time_max.sec;
7923 *(ptr++) = stats->queue_time_max.usec;
7924 *(ptr++) = stats->execution_time_sum.sec;
7925 *(ptr++) = stats->execution_time_sum.usec;
7926 *(ptr++) = stats->execution_time_sum_sqr.sec;
7927 *(ptr++) = stats->execution_time_sum_sqr.usec;
7928 *(ptr++) = stats->execution_time_min.sec;
7929 *(ptr++) = stats->execution_time_min.usec;
7930 *(ptr++) = stats->execution_time_max.sec;
7931 *(ptr++) = stats->execution_time_max.usec;
7937 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7942 * IN callerVersion - the rpc stat version of the caller
7944 * OUT myVersion - the rpc stat version of this function
7946 * OUT clock_sec - local time seconds
7948 * OUT clock_usec - local time microseconds
7950 * OUT allocSize - the number of bytes allocated to contain stats
7952 * OUT statCount - the number stats retrieved from this process.
7954 * OUT stats - the actual stats retrieved from this process.
7958 * Returns void. If successful, stats will != NULL.
7962 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7963 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7964 size_t * allocSize, afs_uint32 * statCount,
7965 afs_uint32 ** stats)
7975 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7978 * Check to see if stats are enabled
7981 MUTEX_ENTER(&rx_rpc_stats);
7982 if (!rxi_monitor_processStats) {
7983 MUTEX_EXIT(&rx_rpc_stats);
7987 clock_GetTime(&now);
7988 *clock_sec = now.sec;
7989 *clock_usec = now.usec;
7992 * Allocate the space based upon the caller version
7994 * If the client is at an older version than we are,
7995 * we return the statistic data in the older data format, but
7996 * we still return our version number so the client knows we
7997 * are maintaining more data than it can retrieve.
8000 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8001 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8002 *statCount = rxi_rpc_process_stat_cnt;
8005 * This can't happen yet, but in the future version changes
8006 * can be handled by adding additional code here
8010 if (space > (size_t) 0) {
8012 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
8015 rx_interface_stat_p rpc_stat, nrpc_stat;
8019 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8021 * Copy the data based upon the caller version
8023 rx_MarshallProcessRPCStats(callerVersion,
8024 rpc_stat->stats[0].func_total,
8025 rpc_stat->stats, &ptr);
8031 MUTEX_EXIT(&rx_rpc_stats);
8036 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8040 * IN callerVersion - the rpc stat version of the caller
8042 * OUT myVersion - the rpc stat version of this function
8044 * OUT clock_sec - local time seconds
8046 * OUT clock_usec - local time microseconds
8048 * OUT allocSize - the number of bytes allocated to contain stats
8050 * OUT statCount - the number of stats retrieved from the individual
8053 * OUT stats - the actual stats retrieved from the individual peer structures.
8057 * Returns void. If successful, stats will != NULL.
8061 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8062 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8063 size_t * allocSize, afs_uint32 * statCount,
8064 afs_uint32 ** stats)
8074 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8077 * Check to see if stats are enabled
8080 MUTEX_ENTER(&rx_rpc_stats);
8081 if (!rxi_monitor_peerStats) {
8082 MUTEX_EXIT(&rx_rpc_stats);
8086 clock_GetTime(&now);
8087 *clock_sec = now.sec;
8088 *clock_usec = now.usec;
8091 * Allocate the space based upon the caller version
8093 * If the client is at an older version than we are,
8094 * we return the statistic data in the older data format, but
8095 * we still return our version number so the client knows we
8096 * are maintaining more data than it can retrieve.
8099 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8100 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8101 *statCount = rxi_rpc_peer_stat_cnt;
8104 * This can't happen yet, but in the future version changes
8105 * can be handled by adding additional code here
8109 if (space > (size_t) 0) {
8111 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
8114 rx_interface_stat_p rpc_stat, nrpc_stat;
8118 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8120 * We have to fix the offset of rpc_stat since we are
8121 * keeping this structure on two rx_queues. The rx_queue
8122 * package assumes that the rx_queue member is the first
8123 * member of the structure. That is, rx_queue assumes that
8124 * any one item is only on one queue at a time. We are
8125 * breaking that assumption and so we have to do a little
8126 * math to fix our pointers.
8129 fix_offset = (char *)rpc_stat;
8130 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8131 rpc_stat = (rx_interface_stat_p) fix_offset;
8134 * Copy the data based upon the caller version
8136 rx_MarshallProcessRPCStats(callerVersion,
8137 rpc_stat->stats[0].func_total,
8138 rpc_stat->stats, &ptr);
8144 MUTEX_EXIT(&rx_rpc_stats);
8149 * rx_FreeRPCStats - free memory allocated by
8150 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8154 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8155 * rx_RetrievePeerRPCStats
8157 * IN allocSize - the number of bytes in stats.
8165 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8167 rxi_Free(stats, allocSize);
8171 * rx_queryProcessRPCStats - see if process rpc stat collection is
8172 * currently enabled.
8178 * Returns 0 if stats are not enabled != 0 otherwise
8182 rx_queryProcessRPCStats(void)
8185 MUTEX_ENTER(&rx_rpc_stats);
8186 rc = rxi_monitor_processStats;
8187 MUTEX_EXIT(&rx_rpc_stats);
8192 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8198 * Returns 0 if stats are not enabled != 0 otherwise
8202 rx_queryPeerRPCStats(void)
8205 MUTEX_ENTER(&rx_rpc_stats);
8206 rc = rxi_monitor_peerStats;
8207 MUTEX_EXIT(&rx_rpc_stats);
8212 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8222 rx_enableProcessRPCStats(void)
8224 MUTEX_ENTER(&rx_rpc_stats);
8225 rx_enable_stats = 1;
8226 rxi_monitor_processStats = 1;
8227 MUTEX_EXIT(&rx_rpc_stats);
8231 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8241 rx_enablePeerRPCStats(void)
8243 MUTEX_ENTER(&rx_rpc_stats);
8244 rx_enable_stats = 1;
8245 rxi_monitor_peerStats = 1;
8246 MUTEX_EXIT(&rx_rpc_stats);
8250 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8260 rx_disableProcessRPCStats(void)
8262 rx_interface_stat_p rpc_stat, nrpc_stat;
8265 MUTEX_ENTER(&rx_rpc_stats);
8268 * Turn off process statistics and if peer stats is also off, turn
8272 rxi_monitor_processStats = 0;
8273 if (rxi_monitor_peerStats == 0) {
8274 rx_enable_stats = 0;
8277 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8278 unsigned int num_funcs = 0;
8281 queue_Remove(rpc_stat);
8282 num_funcs = rpc_stat->stats[0].func_total;
8284 sizeof(rx_interface_stat_t) +
8285 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8287 rxi_Free(rpc_stat, space);
8288 rxi_rpc_process_stat_cnt -= num_funcs;
8290 MUTEX_EXIT(&rx_rpc_stats);
8294 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8304 rx_disablePeerRPCStats(void)
8306 struct rx_peer **peer_ptr, **peer_end;
8310 * Turn off peer statistics and if process stats is also off, turn
8314 rxi_monitor_peerStats = 0;
8315 if (rxi_monitor_processStats == 0) {
8316 rx_enable_stats = 0;
8319 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8320 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8322 struct rx_peer *peer, *next, *prev;
8324 MUTEX_ENTER(&rx_peerHashTable_lock);
8325 MUTEX_ENTER(&rx_rpc_stats);
8326 for (prev = peer = *peer_ptr; peer; peer = next) {
8328 code = MUTEX_TRYENTER(&peer->peer_lock);
8330 rx_interface_stat_p rpc_stat, nrpc_stat;
8333 if (prev == *peer_ptr) {
8344 MUTEX_EXIT(&rx_peerHashTable_lock);
8347 (&peer->rpcStats, rpc_stat, nrpc_stat,
8348 rx_interface_stat)) {
8349 unsigned int num_funcs = 0;
8352 queue_Remove(&rpc_stat->queue_header);
8353 queue_Remove(&rpc_stat->all_peers);
8354 num_funcs = rpc_stat->stats[0].func_total;
8356 sizeof(rx_interface_stat_t) +
8357 rpc_stat->stats[0].func_total *
8358 sizeof(rx_function_entry_v1_t);
8360 rxi_Free(rpc_stat, space);
8361 rxi_rpc_peer_stat_cnt -= num_funcs;
8363 MUTEX_EXIT(&peer->peer_lock);
8365 MUTEX_ENTER(&rx_peerHashTable_lock);
8375 MUTEX_EXIT(&rx_rpc_stats);
8376 MUTEX_EXIT(&rx_peerHashTable_lock);
8381 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8386 * IN clearFlag - flag indicating which stats to clear
8394 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8396 rx_interface_stat_p rpc_stat, nrpc_stat;
8398 MUTEX_ENTER(&rx_rpc_stats);
8400 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8401 unsigned int num_funcs = 0, i;
8402 num_funcs = rpc_stat->stats[0].func_total;
8403 for (i = 0; i < num_funcs; i++) {
8404 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8405 hzero(rpc_stat->stats[i].invocations);
8407 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8408 hzero(rpc_stat->stats[i].bytes_sent);
8410 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8411 hzero(rpc_stat->stats[i].bytes_rcvd);
8413 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8414 rpc_stat->stats[i].queue_time_sum.sec = 0;
8415 rpc_stat->stats[i].queue_time_sum.usec = 0;
8417 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8418 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8419 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8421 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8422 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8423 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8425 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8426 rpc_stat->stats[i].queue_time_max.sec = 0;
8427 rpc_stat->stats[i].queue_time_max.usec = 0;
8429 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8430 rpc_stat->stats[i].execution_time_sum.sec = 0;
8431 rpc_stat->stats[i].execution_time_sum.usec = 0;
8433 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8434 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8435 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8437 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8438 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8439 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8441 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8442 rpc_stat->stats[i].execution_time_max.sec = 0;
8443 rpc_stat->stats[i].execution_time_max.usec = 0;
8448 MUTEX_EXIT(&rx_rpc_stats);
8452 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8457 * IN clearFlag - flag indicating which stats to clear
8465 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8467 rx_interface_stat_p rpc_stat, nrpc_stat;
8469 MUTEX_ENTER(&rx_rpc_stats);
8471 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8472 unsigned int num_funcs = 0, i;
8475 * We have to fix the offset of rpc_stat since we are
8476 * keeping this structure on two rx_queues. The rx_queue
8477 * package assumes that the rx_queue member is the first
8478 * member of the structure. That is, rx_queue assumes that
8479 * any one item is only on one queue at a time. We are
8480 * breaking that assumption and so we have to do a little
8481 * math to fix our pointers.
8484 fix_offset = (char *)rpc_stat;
8485 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8486 rpc_stat = (rx_interface_stat_p) fix_offset;
8488 num_funcs = rpc_stat->stats[0].func_total;
8489 for (i = 0; i < num_funcs; i++) {
8490 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8491 hzero(rpc_stat->stats[i].invocations);
8493 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8494 hzero(rpc_stat->stats[i].bytes_sent);
8496 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8497 hzero(rpc_stat->stats[i].bytes_rcvd);
8499 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8500 rpc_stat->stats[i].queue_time_sum.sec = 0;
8501 rpc_stat->stats[i].queue_time_sum.usec = 0;
8503 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8504 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8505 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8507 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8508 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8509 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8511 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8512 rpc_stat->stats[i].queue_time_max.sec = 0;
8513 rpc_stat->stats[i].queue_time_max.usec = 0;
8515 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8516 rpc_stat->stats[i].execution_time_sum.sec = 0;
8517 rpc_stat->stats[i].execution_time_sum.usec = 0;
8519 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8520 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8521 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8523 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8524 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8525 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8527 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8528 rpc_stat->stats[i].execution_time_max.sec = 0;
8529 rpc_stat->stats[i].execution_time_max.usec = 0;
8534 MUTEX_EXIT(&rx_rpc_stats);
8538 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8539 * is authorized to enable/disable/clear RX statistics.
8541 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8544 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8546 rxi_rxstat_userok = proc;
8550 rx_RxStatUserOk(struct rx_call *call)
8552 if (!rxi_rxstat_userok)
8554 return rxi_rxstat_userok(call);
8559 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8560 * function in the MSVC runtime DLL (msvcrt.dll).
8562 * Note: the system serializes calls to this function.
8565 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8566 DWORD reason, /* reason function is being called */
8567 LPVOID reserved) /* reserved for future use */
8570 case DLL_PROCESS_ATTACH:
8571 /* library is being attached to a process */
8575 case DLL_PROCESS_DETACH:
8582 #endif /* AFS_NT40_ENV */
8585 int rx_DumpCalls(FILE *outputFile, char *cookie)
8587 #ifdef RXDEBUG_PACKET
8588 #ifdef KDUMP_RX_LOCK
8589 struct rx_call_rx_lock *c;
8596 #define RXDPRINTF sprintf
8597 #define RXDPRINTOUT output
8599 #define RXDPRINTF fprintf
8600 #define RXDPRINTOUT outputFile
8603 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8605 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8608 for (c = rx_allCallsp; c; c = c->allNextp) {
8609 u_short rqc, tqc, iovqc;
8610 struct rx_packet *p, *np;
8612 MUTEX_ENTER(&c->lock);
8613 queue_Count(&c->rq, p, np, rx_packet, rqc);
8614 queue_Count(&c->tq, p, np, rx_packet, tqc);
8615 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8617 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, "
8618 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8619 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8620 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8621 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8622 #ifdef RX_ENABLE_LOCKS
8625 #ifdef RX_REFCOUNT_CHECK
8626 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
8627 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
8630 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,
8631 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
8632 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
8633 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
8634 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
8635 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
8636 #ifdef RX_ENABLE_LOCKS
8637 , (afs_uint32)c->refCount
8639 #ifdef RX_REFCOUNT_CHECK
8640 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
8643 MUTEX_EXIT(&c->lock);
8646 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8649 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
8651 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8653 #endif /* RXDEBUG_PACKET */