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"
59 #endif /* AFS_OSF_ENV */
61 #include "afs/sysincludes.h"
62 #include "afsincludes.h"
65 #include "rx_kmutex.h"
66 #include "rx_kernel.h"
70 #include "rx_globals.h"
72 #include "rx_atomic.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 "afs/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_atomic.h"
109 # include "rx_globals.h"
110 # include "rx_trace.h"
111 # include <afs/rxgen_consts.h>
115 #ifdef AFS_PTHREAD_ENV
117 int (*registerProgram) (pid_t, char *) = 0;
118 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
121 int (*registerProgram) (PROCESS, char *) = 0;
122 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
126 /* Local static routines */
127 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
128 #ifdef RX_ENABLE_LOCKS
129 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
132 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
134 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
135 afs_int32 rxi_start_in_error;
137 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
140 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
141 * currently allocated within rx. This number is used to allocate the
142 * memory required to return the statistics when queried.
143 * Protected by the rx_rpc_stats mutex.
146 static unsigned int rxi_rpc_peer_stat_cnt;
149 * rxi_rpc_process_stat_cnt counts the total number of local process stat
150 * structures currently allocated within rx. The number is used to allocate
151 * the memory required to return the statistics when queried.
152 * Protected by the rx_rpc_stats mutex.
155 static unsigned int rxi_rpc_process_stat_cnt;
157 #if !defined(offsetof)
158 #include <stddef.h> /* for definition of offsetof() */
161 #ifdef RX_ENABLE_LOCKS
162 afs_kmutex_t rx_atomic_mutex;
165 #ifdef AFS_PTHREAD_ENV
169 * Use procedural initialization of mutexes/condition variables
173 extern afs_kmutex_t rx_stats_mutex;
174 extern afs_kmutex_t rx_waiting_mutex;
175 extern afs_kmutex_t rx_quota_mutex;
176 extern afs_kmutex_t rx_pthread_mutex;
177 extern afs_kmutex_t rx_packets_mutex;
178 extern afs_kmutex_t des_init_mutex;
179 extern afs_kmutex_t des_random_mutex;
180 extern afs_kmutex_t rx_clock_mutex;
181 extern afs_kmutex_t rxi_connCacheMutex;
182 extern afs_kmutex_t rx_event_mutex;
183 extern afs_kmutex_t osi_malloc_mutex;
184 extern afs_kmutex_t event_handler_mutex;
185 extern afs_kmutex_t listener_mutex;
186 extern afs_kmutex_t rx_if_init_mutex;
187 extern afs_kmutex_t rx_if_mutex;
188 extern afs_kmutex_t rxkad_client_uid_mutex;
189 extern afs_kmutex_t rxkad_random_mutex;
191 extern afs_kcondvar_t rx_event_handler_cond;
192 extern afs_kcondvar_t rx_listener_cond;
194 static afs_kmutex_t epoch_mutex;
195 static afs_kmutex_t rx_init_mutex;
196 static afs_kmutex_t rx_debug_mutex;
197 static afs_kmutex_t rx_rpc_stats;
200 rxi_InitPthread(void)
202 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
203 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
204 MUTEX_INIT(&rx_waiting_mutex, "waiting", MUTEX_DEFAULT, 0);
205 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
206 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
207 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
208 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
209 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
210 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
211 MUTEX_INIT(&rx_event_mutex, "event", MUTEX_DEFAULT, 0);
212 MUTEX_INIT(&des_init_mutex, "des", MUTEX_DEFAULT, 0);
213 MUTEX_INIT(&des_random_mutex, "random", MUTEX_DEFAULT, 0);
214 MUTEX_INIT(&osi_malloc_mutex, "malloc", MUTEX_DEFAULT, 0);
215 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
216 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
217 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
218 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
219 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
220 MUTEX_INIT(&rxkad_client_uid_mutex, "uid", MUTEX_DEFAULT, 0);
221 MUTEX_INIT(&rxkad_random_mutex, "rxkad random", MUTEX_DEFAULT, 0);
222 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
224 assert(pthread_cond_init
225 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
226 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
228 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
229 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
231 rxkad_global_stats_init();
233 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
234 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
235 #ifdef RX_ENABLE_LOCKS
238 #endif /* RX_LOCKS_DB */
239 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
240 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
242 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
244 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
246 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
248 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
249 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
250 #endif /* RX_ENABLE_LOCKS */
253 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
254 #define INIT_PTHREAD_LOCKS \
255 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
257 * The rx_stats_mutex mutex protects the following global variables:
258 * rxi_lowConnRefCount
259 * rxi_lowPeerRefCount
268 * The rx_quota_mutex mutex protects the following global variables:
276 * The rx_freePktQ_lock protects the following global variables:
281 * The rx_packets_mutex mutex protects the following global variables:
289 * The rx_pthread_mutex mutex protects the following global variables:
293 #define INIT_PTHREAD_LOCKS
297 /* Variables for handling the minProcs implementation. availProcs gives the
298 * number of threads available in the pool at this moment (not counting dudes
299 * executing right now). totalMin gives the total number of procs required
300 * for handling all minProcs requests. minDeficit is a dynamic variable
301 * tracking the # of procs required to satisfy all of the remaining minProcs
303 * For fine grain locking to work, the quota check and the reservation of
304 * a server thread has to come while rxi_availProcs and rxi_minDeficit
305 * are locked. To this end, the code has been modified under #ifdef
306 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
307 * same time. A new function, ReturnToServerPool() returns the allocation.
309 * A call can be on several queue's (but only one at a time). When
310 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
311 * that no one else is touching the queue. To this end, we store the address
312 * of the queue lock in the call structure (under the call lock) when we
313 * put the call on a queue, and we clear the call_queue_lock when the
314 * call is removed from a queue (once the call lock has been obtained).
315 * This allows rxi_ResetCall to safely synchronize with others wishing
316 * to manipulate the queue.
319 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
320 static afs_kmutex_t rx_rpc_stats;
321 void rxi_StartUnlocked(struct rxevent *event, void *call,
322 void *arg1, int istack);
325 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
326 ** pretty good that the next packet coming in is from the same connection
327 ** as the last packet, since we're send multiple packets in a transmit window.
329 struct rx_connection *rxLastConn = 0;
331 #ifdef RX_ENABLE_LOCKS
332 /* The locking hierarchy for rx fine grain locking is composed of these
335 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
336 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
337 * call->lock - locks call data fields.
338 * These are independent of each other:
339 * rx_freeCallQueue_lock
344 * serverQueueEntry->lock
345 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
347 * peer->lock - locks peer data fields.
348 * conn_data_lock - that more than one thread is not updating a conn data
349 * field at the same time.
358 * Do we need a lock to protect the peer field in the conn structure?
359 * conn->peer was previously a constant for all intents and so has no
360 * lock protecting this field. The multihomed client delta introduced
361 * a RX code change : change the peer field in the connection structure
362 * to that remote interface from which the last packet for this
363 * connection was sent out. This may become an issue if further changes
366 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
367 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
369 /* rxdb_fileID is used to identify the lock location, along with line#. */
370 static int rxdb_fileID = RXDB_FILE_RX;
371 #endif /* RX_LOCKS_DB */
372 #else /* RX_ENABLE_LOCKS */
373 #define SET_CALL_QUEUE_LOCK(C, L)
374 #define CLEAR_CALL_QUEUE_LOCK(C)
375 #endif /* RX_ENABLE_LOCKS */
376 struct rx_serverQueueEntry *rx_waitForPacket = 0;
377 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
379 /* ------------Exported Interfaces------------- */
381 /* This function allows rxkad to set the epoch to a suitably random number
382 * which rx_NewConnection will use in the future. The principle purpose is to
383 * get rxnull connections to use the same epoch as the rxkad connections do, at
384 * least once the first rxkad connection is established. This is important now
385 * that the host/port addresses aren't used in FindConnection: the uniqueness
386 * of epoch/cid matters and the start time won't do. */
388 #ifdef AFS_PTHREAD_ENV
390 * This mutex protects the following global variables:
394 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
395 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
399 #endif /* AFS_PTHREAD_ENV */
402 rx_SetEpoch(afs_uint32 epoch)
409 /* Initialize rx. A port number may be mentioned, in which case this
410 * becomes the default port number for any service installed later.
411 * If 0 is provided for the port number, a random port will be chosen
412 * by the kernel. Whether this will ever overlap anything in
413 * /etc/services is anybody's guess... Returns 0 on success, -1 on
418 int rxinit_status = 1;
419 #ifdef AFS_PTHREAD_ENV
421 * This mutex protects the following global variables:
425 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
426 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
429 #define UNLOCK_RX_INIT
433 rx_InitHost(u_int host, u_int port)
440 char *htable, *ptable;
447 if (rxinit_status == 0) {
448 tmp_status = rxinit_status;
450 return tmp_status; /* Already started; return previous error code. */
456 if (afs_winsockInit() < 0)
462 * Initialize anything necessary to provide a non-premptive threading
465 rxi_InitializeThreadSupport();
468 /* Allocate and initialize a socket for client and perhaps server
471 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
472 if (rx_socket == OSI_NULLSOCKET) {
476 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
479 #endif /* RX_LOCKS_DB */
480 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
481 MUTEX_INIT(&rx_waiting_mutex, "rx_waiting_mutex", MUTEX_DEFAULT, 0);
482 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
483 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
484 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
485 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
486 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
487 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
488 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
490 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
492 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
494 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
496 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
497 #if defined(AFS_HPUX110_ENV)
499 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
500 #endif /* AFS_HPUX110_ENV */
501 #endif /* RX_ENABLE_LOCKS && KERNEL */
504 rx_connDeadTime = 12;
505 rx_tranquil = 0; /* reset flag */
506 memset(&rx_stats, 0, sizeof(struct rx_statistics));
508 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
509 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
510 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
511 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
512 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
513 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
515 /* Malloc up a bunch of packets & buffers */
517 queue_Init(&rx_freePacketQueue);
518 rxi_NeedMorePackets = FALSE;
519 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
521 /* enforce a minimum number of allocated packets */
522 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
523 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
525 /* allocate the initial free packet pool */
526 #ifdef RX_ENABLE_TSFPQ
527 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
528 #else /* RX_ENABLE_TSFPQ */
529 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
530 #endif /* RX_ENABLE_TSFPQ */
537 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
538 tv.tv_sec = clock_now.sec;
539 tv.tv_usec = clock_now.usec;
540 srand((unsigned int)tv.tv_usec);
547 #if defined(KERNEL) && !defined(UKERNEL)
548 /* Really, this should never happen in a real kernel */
551 struct sockaddr_in addr;
553 int addrlen = sizeof(addr);
555 socklen_t addrlen = sizeof(addr);
557 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
561 rx_port = addr.sin_port;
564 rx_stats.minRtt.sec = 9999999;
566 rx_SetEpoch(tv.tv_sec | 0x80000000);
568 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
569 * will provide a randomer value. */
571 MUTEX_ENTER(&rx_quota_mutex);
572 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
573 MUTEX_EXIT(&rx_quota_mutex);
574 /* *Slightly* random start time for the cid. This is just to help
575 * out with the hashing function at the peer */
576 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
577 rx_connHashTable = (struct rx_connection **)htable;
578 rx_peerHashTable = (struct rx_peer **)ptable;
580 rx_lastAckDelay.sec = 0;
581 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
582 rx_hardAckDelay.sec = 0;
583 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
584 rx_softAckDelay.sec = 0;
585 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
587 rxevent_Init(20, rxi_ReScheduleEvents);
589 /* Initialize various global queues */
590 queue_Init(&rx_idleServerQueue);
591 queue_Init(&rx_incomingCallQueue);
592 queue_Init(&rx_freeCallQueue);
594 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
595 /* Initialize our list of usable IP addresses. */
599 /* Start listener process (exact function is dependent on the
600 * implementation environment--kernel or user space) */
604 tmp_status = rxinit_status = 0;
612 return rx_InitHost(htonl(INADDR_ANY), port);
615 /* called with unincremented nRequestsRunning to see if it is OK to start
616 * a new thread in this service. Could be "no" for two reasons: over the
617 * max quota, or would prevent others from reaching their min quota.
619 #ifdef RX_ENABLE_LOCKS
620 /* This verion of QuotaOK reserves quota if it's ok while the
621 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
624 QuotaOK(struct rx_service *aservice)
626 /* check if over max quota */
627 if (aservice->nRequestsRunning >= aservice->maxProcs) {
631 /* under min quota, we're OK */
632 /* otherwise, can use only if there are enough to allow everyone
633 * to go to their min quota after this guy starts.
636 MUTEX_ENTER(&rx_quota_mutex);
637 if ((aservice->nRequestsRunning < aservice->minProcs)
638 || (rxi_availProcs > rxi_minDeficit)) {
639 aservice->nRequestsRunning++;
640 /* just started call in minProcs pool, need fewer to maintain
642 if (aservice->nRequestsRunning <= aservice->minProcs)
645 MUTEX_EXIT(&rx_quota_mutex);
648 MUTEX_EXIT(&rx_quota_mutex);
654 ReturnToServerPool(struct rx_service *aservice)
656 aservice->nRequestsRunning--;
657 MUTEX_ENTER(&rx_quota_mutex);
658 if (aservice->nRequestsRunning < aservice->minProcs)
661 MUTEX_EXIT(&rx_quota_mutex);
664 #else /* RX_ENABLE_LOCKS */
666 QuotaOK(struct rx_service *aservice)
669 /* under min quota, we're OK */
670 if (aservice->nRequestsRunning < aservice->minProcs)
673 /* check if over max quota */
674 if (aservice->nRequestsRunning >= aservice->maxProcs)
677 /* otherwise, can use only if there are enough to allow everyone
678 * to go to their min quota after this guy starts.
680 MUTEX_ENTER(&rx_quota_mutex);
681 if (rxi_availProcs > rxi_minDeficit)
683 MUTEX_EXIT(&rx_quota_mutex);
686 #endif /* RX_ENABLE_LOCKS */
689 /* Called by rx_StartServer to start up lwp's to service calls.
690 NExistingProcs gives the number of procs already existing, and which
691 therefore needn't be created. */
693 rxi_StartServerProcs(int nExistingProcs)
695 struct rx_service *service;
700 /* For each service, reserve N processes, where N is the "minimum"
701 * number of processes that MUST be able to execute a request in parallel,
702 * at any time, for that process. Also compute the maximum difference
703 * between any service's maximum number of processes that can run
704 * (i.e. the maximum number that ever will be run, and a guarantee
705 * that this number will run if other services aren't running), and its
706 * minimum number. The result is the extra number of processes that
707 * we need in order to provide the latter guarantee */
708 for (i = 0; i < RX_MAX_SERVICES; i++) {
710 service = rx_services[i];
711 if (service == (struct rx_service *)0)
713 nProcs += service->minProcs;
714 diff = service->maxProcs - service->minProcs;
718 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
719 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
720 for (i = 0; i < nProcs; i++) {
721 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
727 /* This routine is only required on Windows */
729 rx_StartClientThread(void)
731 #ifdef AFS_PTHREAD_ENV
733 pid = pthread_self();
734 #endif /* AFS_PTHREAD_ENV */
736 #endif /* AFS_NT40_ENV */
738 /* This routine must be called if any services are exported. If the
739 * donateMe flag is set, the calling process is donated to the server
742 rx_StartServer(int donateMe)
744 struct rx_service *service;
750 /* Start server processes, if necessary (exact function is dependent
751 * on the implementation environment--kernel or user space). DonateMe
752 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
753 * case, one less new proc will be created rx_StartServerProcs.
755 rxi_StartServerProcs(donateMe);
757 /* count up the # of threads in minProcs, and add set the min deficit to
758 * be that value, too.
760 for (i = 0; i < RX_MAX_SERVICES; i++) {
761 service = rx_services[i];
762 if (service == (struct rx_service *)0)
764 MUTEX_ENTER(&rx_quota_mutex);
765 rxi_totalMin += service->minProcs;
766 /* below works even if a thread is running, since minDeficit would
767 * still have been decremented and later re-incremented.
769 rxi_minDeficit += service->minProcs;
770 MUTEX_EXIT(&rx_quota_mutex);
773 /* Turn on reaping of idle server connections */
774 rxi_ReapConnections(NULL, NULL, NULL);
783 #ifdef AFS_PTHREAD_ENV
785 pid = afs_pointer_to_int(pthread_self());
786 #else /* AFS_PTHREAD_ENV */
788 LWP_CurrentProcess(&pid);
789 #endif /* AFS_PTHREAD_ENV */
791 sprintf(name, "srv_%d", ++nProcs);
793 (*registerProgram) (pid, name);
795 #endif /* AFS_NT40_ENV */
796 rx_ServerProc(NULL); /* Never returns */
798 #ifdef RX_ENABLE_TSFPQ
799 /* no use leaving packets around in this thread's local queue if
800 * it isn't getting donated to the server thread pool.
802 rxi_FlushLocalPacketsTSFPQ();
803 #endif /* RX_ENABLE_TSFPQ */
807 /* Create a new client connection to the specified service, using the
808 * specified security object to implement the security model for this
810 struct rx_connection *
811 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
812 struct rx_securityClass *securityObject,
813 int serviceSecurityIndex)
817 struct rx_connection *conn;
822 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
823 "serviceSecurityIndex %d)\n",
824 ntohl(shost), ntohs(sport), sservice, securityObject,
825 serviceSecurityIndex));
827 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
828 * the case of kmem_alloc? */
829 conn = rxi_AllocConnection();
830 #ifdef RX_ENABLE_LOCKS
831 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
832 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
833 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
836 MUTEX_ENTER(&rx_connHashTable_lock);
837 cid = (rx_nextCid += RX_MAXCALLS);
838 conn->type = RX_CLIENT_CONNECTION;
840 conn->epoch = rx_epoch;
841 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
842 conn->serviceId = sservice;
843 conn->securityObject = securityObject;
844 conn->securityData = (void *) 0;
845 conn->securityIndex = serviceSecurityIndex;
846 rx_SetConnDeadTime(conn, rx_connDeadTime);
847 rx_SetConnSecondsUntilNatPing(conn, 0);
848 conn->ackRate = RX_FAST_ACK_RATE;
850 conn->specific = NULL;
851 conn->challengeEvent = NULL;
852 conn->delayedAbortEvent = NULL;
853 conn->abortCount = 0;
855 for (i = 0; i < RX_MAXCALLS; i++) {
856 conn->twind[i] = rx_initSendWindow;
857 conn->rwind[i] = rx_initReceiveWindow;
860 RXS_NewConnection(securityObject, conn);
862 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
864 conn->refCount++; /* no lock required since only this thread knows... */
865 conn->next = rx_connHashTable[hashindex];
866 rx_connHashTable[hashindex] = conn;
868 rx_MutexIncrement(rx_stats.nClientConns, rx_stats_mutex);
869 MUTEX_EXIT(&rx_connHashTable_lock);
875 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
877 /* The idea is to set the dead time to a value that allows several
878 * keepalives to be dropped without timing out the connection. */
879 conn->secondsUntilDead = MAX(seconds, 6);
880 conn->secondsUntilPing = conn->secondsUntilDead / 6;
883 int rxi_lowPeerRefCount = 0;
884 int rxi_lowConnRefCount = 0;
887 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
888 * NOTE: must not be called with rx_connHashTable_lock held.
891 rxi_CleanupConnection(struct rx_connection *conn)
893 /* Notify the service exporter, if requested, that this connection
894 * is being destroyed */
895 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
896 (*conn->service->destroyConnProc) (conn);
898 /* Notify the security module that this connection is being destroyed */
899 RXS_DestroyConnection(conn->securityObject, conn);
901 /* If this is the last connection using the rx_peer struct, set its
902 * idle time to now. rxi_ReapConnections will reap it if it's still
903 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
905 MUTEX_ENTER(&rx_peerHashTable_lock);
906 if (conn->peer->refCount < 2) {
907 conn->peer->idleWhen = clock_Sec();
908 if (conn->peer->refCount < 1) {
909 conn->peer->refCount = 1;
910 if (rx_stats_active) {
911 MUTEX_ENTER(&rx_stats_mutex);
912 rxi_lowPeerRefCount++;
913 MUTEX_EXIT(&rx_stats_mutex);
917 conn->peer->refCount--;
918 MUTEX_EXIT(&rx_peerHashTable_lock);
922 if (conn->type == RX_SERVER_CONNECTION)
923 rx_MutexDecrement(rx_stats.nServerConns, rx_stats_mutex);
925 rx_MutexDecrement(rx_stats.nClientConns, rx_stats_mutex);
928 if (conn->specific) {
930 for (i = 0; i < conn->nSpecific; i++) {
931 if (conn->specific[i] && rxi_keyCreate_destructor[i])
932 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
933 conn->specific[i] = NULL;
935 free(conn->specific);
937 conn->specific = NULL;
941 MUTEX_DESTROY(&conn->conn_call_lock);
942 MUTEX_DESTROY(&conn->conn_data_lock);
943 CV_DESTROY(&conn->conn_call_cv);
945 rxi_FreeConnection(conn);
948 /* Destroy the specified connection */
950 rxi_DestroyConnection(struct rx_connection *conn)
952 MUTEX_ENTER(&rx_connHashTable_lock);
953 rxi_DestroyConnectionNoLock(conn);
954 /* conn should be at the head of the cleanup list */
955 if (conn == rx_connCleanup_list) {
956 rx_connCleanup_list = rx_connCleanup_list->next;
957 MUTEX_EXIT(&rx_connHashTable_lock);
958 rxi_CleanupConnection(conn);
960 #ifdef RX_ENABLE_LOCKS
962 MUTEX_EXIT(&rx_connHashTable_lock);
964 #endif /* RX_ENABLE_LOCKS */
968 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
970 struct rx_connection **conn_ptr;
972 struct rx_packet *packet;
979 MUTEX_ENTER(&conn->conn_data_lock);
980 if (conn->refCount > 0)
983 if (rx_stats_active) {
984 MUTEX_ENTER(&rx_stats_mutex);
985 rxi_lowConnRefCount++;
986 MUTEX_EXIT(&rx_stats_mutex);
990 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
991 /* Busy; wait till the last guy before proceeding */
992 MUTEX_EXIT(&conn->conn_data_lock);
997 /* If the client previously called rx_NewCall, but it is still
998 * waiting, treat this as a running call, and wait to destroy the
999 * connection later when the call completes. */
1000 if ((conn->type == RX_CLIENT_CONNECTION)
1001 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1002 conn->flags |= RX_CONN_DESTROY_ME;
1003 MUTEX_EXIT(&conn->conn_data_lock);
1007 MUTEX_EXIT(&conn->conn_data_lock);
1009 /* Check for extant references to this connection */
1010 for (i = 0; i < RX_MAXCALLS; i++) {
1011 struct rx_call *call = conn->call[i];
1014 if (conn->type == RX_CLIENT_CONNECTION) {
1015 MUTEX_ENTER(&call->lock);
1016 if (call->delayedAckEvent) {
1017 /* Push the final acknowledgment out now--there
1018 * won't be a subsequent call to acknowledge the
1019 * last reply packets */
1020 rxevent_Cancel(call->delayedAckEvent, call,
1021 RX_CALL_REFCOUNT_DELAY);
1022 if (call->state == RX_STATE_PRECALL
1023 || call->state == RX_STATE_ACTIVE) {
1024 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1026 rxi_AckAll(NULL, call, 0);
1029 MUTEX_EXIT(&call->lock);
1033 #ifdef RX_ENABLE_LOCKS
1035 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1036 MUTEX_EXIT(&conn->conn_data_lock);
1038 /* Someone is accessing a packet right now. */
1042 #endif /* RX_ENABLE_LOCKS */
1045 /* Don't destroy the connection if there are any call
1046 * structures still in use */
1047 MUTEX_ENTER(&conn->conn_data_lock);
1048 conn->flags |= RX_CONN_DESTROY_ME;
1049 MUTEX_EXIT(&conn->conn_data_lock);
1054 if (conn->natKeepAliveEvent) {
1055 rxi_NatKeepAliveOff(conn);
1058 if (conn->delayedAbortEvent) {
1059 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1060 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1062 MUTEX_ENTER(&conn->conn_data_lock);
1063 rxi_SendConnectionAbort(conn, packet, 0, 1);
1064 MUTEX_EXIT(&conn->conn_data_lock);
1065 rxi_FreePacket(packet);
1069 /* Remove from connection hash table before proceeding */
1071 &rx_connHashTable[CONN_HASH
1072 (peer->host, peer->port, conn->cid, conn->epoch,
1074 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1075 if (*conn_ptr == conn) {
1076 *conn_ptr = conn->next;
1080 /* if the conn that we are destroying was the last connection, then we
1081 * clear rxLastConn as well */
1082 if (rxLastConn == conn)
1085 /* Make sure the connection is completely reset before deleting it. */
1086 /* get rid of pending events that could zap us later */
1087 if (conn->challengeEvent)
1088 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1089 if (conn->checkReachEvent)
1090 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1091 if (conn->natKeepAliveEvent)
1092 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
1094 /* Add the connection to the list of destroyed connections that
1095 * need to be cleaned up. This is necessary to avoid deadlocks
1096 * in the routines we call to inform others that this connection is
1097 * being destroyed. */
1098 conn->next = rx_connCleanup_list;
1099 rx_connCleanup_list = conn;
1102 /* Externally available version */
1104 rx_DestroyConnection(struct rx_connection *conn)
1109 rxi_DestroyConnection(conn);
1114 rx_GetConnection(struct rx_connection *conn)
1119 MUTEX_ENTER(&conn->conn_data_lock);
1121 MUTEX_EXIT(&conn->conn_data_lock);
1125 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1126 /* Wait for the transmit queue to no longer be busy.
1127 * requires the call->lock to be held */
1128 static void rxi_WaitforTQBusy(struct rx_call *call) {
1129 while (call->flags & RX_CALL_TQ_BUSY) {
1130 call->flags |= RX_CALL_TQ_WAIT;
1132 #ifdef RX_ENABLE_LOCKS
1133 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1134 CV_WAIT(&call->cv_tq, &call->lock);
1135 #else /* RX_ENABLE_LOCKS */
1136 osi_rxSleep(&call->tq);
1137 #endif /* RX_ENABLE_LOCKS */
1139 if (call->tqWaiters == 0) {
1140 call->flags &= ~RX_CALL_TQ_WAIT;
1146 /* Start a new rx remote procedure call, on the specified connection.
1147 * If wait is set to 1, wait for a free call channel; otherwise return
1148 * 0. Maxtime gives the maximum number of seconds this call may take,
1149 * after rx_NewCall returns. After this time interval, a call to any
1150 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1151 * For fine grain locking, we hold the conn_call_lock in order to
1152 * to ensure that we don't get signalle after we found a call in an active
1153 * state and before we go to sleep.
1156 rx_NewCall(struct rx_connection *conn)
1159 struct rx_call *call;
1160 struct clock queueTime;
1164 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1167 clock_GetTime(&queueTime);
1169 * Check if there are others waiting for a new call.
1170 * If so, let them go first to avoid starving them.
1171 * This is a fairly simple scheme, and might not be
1172 * a complete solution for large numbers of waiters.
1174 * makeCallWaiters keeps track of the number of
1175 * threads waiting to make calls and the
1176 * RX_CONN_MAKECALL_WAITING flag bit is used to
1177 * indicate that there are indeed calls waiting.
1178 * The flag is set when the waiter is incremented.
1179 * It is only cleared when makeCallWaiters is 0.
1180 * This prevents us from accidently destroying the
1181 * connection while it is potentially about to be used.
1183 MUTEX_ENTER(&conn->conn_call_lock);
1184 MUTEX_ENTER(&conn->conn_data_lock);
1185 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1186 conn->flags |= RX_CONN_MAKECALL_WAITING;
1187 conn->makeCallWaiters++;
1188 MUTEX_EXIT(&conn->conn_data_lock);
1190 #ifdef RX_ENABLE_LOCKS
1191 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1195 MUTEX_ENTER(&conn->conn_data_lock);
1196 conn->makeCallWaiters--;
1197 if (conn->makeCallWaiters == 0)
1198 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1201 /* We are now the active thread in rx_NewCall */
1202 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1203 MUTEX_EXIT(&conn->conn_data_lock);
1208 for (i = 0; i < RX_MAXCALLS; i++) {
1209 call = conn->call[i];
1211 if (call->state == RX_STATE_DALLY) {
1212 MUTEX_ENTER(&call->lock);
1213 if (call->state == RX_STATE_DALLY) {
1215 * We are setting the state to RX_STATE_RESET to
1216 * ensure that no one else will attempt to use this
1217 * call once we drop the conn->conn_call_lock and
1218 * call->lock. We must drop the conn->conn_call_lock
1219 * before calling rxi_ResetCall because the process
1220 * of clearing the transmit queue can block for an
1221 * extended period of time. If we block while holding
1222 * the conn->conn_call_lock, then all rx_EndCall
1223 * processing will block as well. This has a detrimental
1224 * effect on overall system performance.
1226 call->state = RX_STATE_RESET;
1227 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1228 MUTEX_EXIT(&conn->conn_call_lock);
1229 rxi_ResetCall(call, 0);
1230 (*call->callNumber)++;
1231 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1235 * If we failed to be able to safely obtain the
1236 * conn->conn_call_lock we will have to drop the
1237 * call->lock to avoid a deadlock. When the call->lock
1238 * is released the state of the call can change. If it
1239 * is no longer RX_STATE_RESET then some other thread is
1242 MUTEX_EXIT(&call->lock);
1243 MUTEX_ENTER(&conn->conn_call_lock);
1244 MUTEX_ENTER(&call->lock);
1246 if (call->state == RX_STATE_RESET)
1250 * If we get here it means that after dropping
1251 * the conn->conn_call_lock and call->lock that
1252 * the call is no longer ours. If we can't find
1253 * a free call in the remaining slots we should
1254 * not go immediately to RX_CONN_MAKECALL_WAITING
1255 * because by dropping the conn->conn_call_lock
1256 * we have given up synchronization with rx_EndCall.
1257 * Instead, cycle through one more time to see if
1258 * we can find a call that can call our own.
1260 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1263 MUTEX_EXIT(&call->lock);
1266 /* rxi_NewCall returns with mutex locked */
1267 call = rxi_NewCall(conn, i);
1268 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1272 if (i < RX_MAXCALLS) {
1278 MUTEX_ENTER(&conn->conn_data_lock);
1279 conn->flags |= RX_CONN_MAKECALL_WAITING;
1280 conn->makeCallWaiters++;
1281 MUTEX_EXIT(&conn->conn_data_lock);
1283 #ifdef RX_ENABLE_LOCKS
1284 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1288 MUTEX_ENTER(&conn->conn_data_lock);
1289 conn->makeCallWaiters--;
1290 if (conn->makeCallWaiters == 0)
1291 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1292 MUTEX_EXIT(&conn->conn_data_lock);
1294 /* Client is initially in send mode */
1295 call->state = RX_STATE_ACTIVE;
1296 call->error = conn->error;
1298 call->mode = RX_MODE_ERROR;
1300 call->mode = RX_MODE_SENDING;
1302 /* remember start time for call in case we have hard dead time limit */
1303 call->queueTime = queueTime;
1304 clock_GetTime(&call->startTime);
1305 hzero(call->bytesSent);
1306 hzero(call->bytesRcvd);
1308 /* Turn on busy protocol. */
1309 rxi_KeepAliveOn(call);
1311 /* Attempt MTU discovery */
1312 rxi_GrowMTUOn(call);
1315 * We are no longer the active thread in rx_NewCall
1317 MUTEX_ENTER(&conn->conn_data_lock);
1318 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1319 MUTEX_EXIT(&conn->conn_data_lock);
1322 * Wake up anyone else who might be giving us a chance to
1323 * run (see code above that avoids resource starvation).
1325 #ifdef RX_ENABLE_LOCKS
1326 CV_BROADCAST(&conn->conn_call_cv);
1330 MUTEX_EXIT(&conn->conn_call_lock);
1332 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1333 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1334 osi_Panic("rx_NewCall call about to be used without an empty tq");
1336 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1338 MUTEX_EXIT(&call->lock);
1341 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1346 rxi_HasActiveCalls(struct rx_connection *aconn)
1349 struct rx_call *tcall;
1353 for (i = 0; i < RX_MAXCALLS; i++) {
1354 if ((tcall = aconn->call[i])) {
1355 if ((tcall->state == RX_STATE_ACTIVE)
1356 || (tcall->state == RX_STATE_PRECALL)) {
1367 rxi_GetCallNumberVector(struct rx_connection *aconn,
1368 afs_int32 * aint32s)
1371 struct rx_call *tcall;
1375 for (i = 0; i < RX_MAXCALLS; i++) {
1376 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1377 aint32s[i] = aconn->callNumber[i] + 1;
1379 aint32s[i] = aconn->callNumber[i];
1386 rxi_SetCallNumberVector(struct rx_connection *aconn,
1387 afs_int32 * aint32s)
1390 struct rx_call *tcall;
1394 for (i = 0; i < RX_MAXCALLS; i++) {
1395 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1396 aconn->callNumber[i] = aint32s[i] - 1;
1398 aconn->callNumber[i] = aint32s[i];
1404 /* Advertise a new service. A service is named locally by a UDP port
1405 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1408 char *serviceName; Name for identification purposes (e.g. the
1409 service name might be used for probing for
1412 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1413 char *serviceName, struct rx_securityClass **securityObjects,
1414 int nSecurityObjects,
1415 afs_int32(*serviceProc) (struct rx_call * acall))
1417 osi_socket socket = OSI_NULLSOCKET;
1418 struct rx_service *tservice;
1424 if (serviceId == 0) {
1426 "rx_NewService: service id for service %s is not non-zero.\n",
1433 "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",
1441 tservice = rxi_AllocService();
1444 #ifdef RX_ENABLE_LOCKS
1445 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1448 for (i = 0; i < RX_MAX_SERVICES; i++) {
1449 struct rx_service *service = rx_services[i];
1451 if (port == service->servicePort && host == service->serviceHost) {
1452 if (service->serviceId == serviceId) {
1453 /* The identical service has already been
1454 * installed; if the caller was intending to
1455 * change the security classes used by this
1456 * service, he/she loses. */
1458 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1459 serviceName, serviceId, service->serviceName);
1461 rxi_FreeService(tservice);
1464 /* Different service, same port: re-use the socket
1465 * which is bound to the same port */
1466 socket = service->socket;
1469 if (socket == OSI_NULLSOCKET) {
1470 /* If we don't already have a socket (from another
1471 * service on same port) get a new one */
1472 socket = rxi_GetHostUDPSocket(host, port);
1473 if (socket == OSI_NULLSOCKET) {
1475 rxi_FreeService(tservice);
1480 service->socket = socket;
1481 service->serviceHost = host;
1482 service->servicePort = port;
1483 service->serviceId = serviceId;
1484 service->serviceName = serviceName;
1485 service->nSecurityObjects = nSecurityObjects;
1486 service->securityObjects = securityObjects;
1487 service->minProcs = 0;
1488 service->maxProcs = 1;
1489 service->idleDeadTime = 60;
1490 service->idleDeadErr = 0;
1491 service->connDeadTime = rx_connDeadTime;
1492 service->executeRequestProc = serviceProc;
1493 service->checkReach = 0;
1494 service->nSpecific = 0;
1495 service->specific = NULL;
1496 rx_services[i] = service; /* not visible until now */
1502 rxi_FreeService(tservice);
1503 (osi_Msg "rx_NewService: cannot support > %d services\n",
1508 /* Set configuration options for all of a service's security objects */
1511 rx_SetSecurityConfiguration(struct rx_service *service,
1512 rx_securityConfigVariables type,
1516 for (i = 0; i<service->nSecurityObjects; i++) {
1517 if (service->securityObjects[i]) {
1518 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1526 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1527 struct rx_securityClass **securityObjects, int nSecurityObjects,
1528 afs_int32(*serviceProc) (struct rx_call * acall))
1530 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1533 /* Generic request processing loop. This routine should be called
1534 * by the implementation dependent rx_ServerProc. If socketp is
1535 * non-null, it will be set to the file descriptor that this thread
1536 * is now listening on. If socketp is null, this routine will never
1539 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1541 struct rx_call *call;
1543 struct rx_service *tservice = NULL;
1550 call = rx_GetCall(threadID, tservice, socketp);
1551 if (socketp && *socketp != OSI_NULLSOCKET) {
1552 /* We are now a listener thread */
1557 /* if server is restarting( typically smooth shutdown) then do not
1558 * allow any new calls.
1561 if (rx_tranquil && (call != NULL)) {
1565 MUTEX_ENTER(&call->lock);
1567 rxi_CallError(call, RX_RESTARTING);
1568 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1570 MUTEX_EXIT(&call->lock);
1574 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1575 #ifdef RX_ENABLE_LOCKS
1577 #endif /* RX_ENABLE_LOCKS */
1578 afs_termState = AFSOP_STOP_AFS;
1579 afs_osi_Wakeup(&afs_termState);
1580 #ifdef RX_ENABLE_LOCKS
1582 #endif /* RX_ENABLE_LOCKS */
1587 tservice = call->conn->service;
1589 if (tservice->beforeProc)
1590 (*tservice->beforeProc) (call);
1592 code = tservice->executeRequestProc(call);
1594 if (tservice->afterProc)
1595 (*tservice->afterProc) (call, code);
1597 rx_EndCall(call, code);
1598 if (rx_stats_active) {
1599 MUTEX_ENTER(&rx_stats_mutex);
1601 MUTEX_EXIT(&rx_stats_mutex);
1608 rx_WakeupServerProcs(void)
1610 struct rx_serverQueueEntry *np, *tqp;
1614 MUTEX_ENTER(&rx_serverPool_lock);
1616 #ifdef RX_ENABLE_LOCKS
1617 if (rx_waitForPacket)
1618 CV_BROADCAST(&rx_waitForPacket->cv);
1619 #else /* RX_ENABLE_LOCKS */
1620 if (rx_waitForPacket)
1621 osi_rxWakeup(rx_waitForPacket);
1622 #endif /* RX_ENABLE_LOCKS */
1623 MUTEX_ENTER(&freeSQEList_lock);
1624 for (np = rx_FreeSQEList; np; np = tqp) {
1625 tqp = *(struct rx_serverQueueEntry **)np;
1626 #ifdef RX_ENABLE_LOCKS
1627 CV_BROADCAST(&np->cv);
1628 #else /* RX_ENABLE_LOCKS */
1630 #endif /* RX_ENABLE_LOCKS */
1632 MUTEX_EXIT(&freeSQEList_lock);
1633 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1634 #ifdef RX_ENABLE_LOCKS
1635 CV_BROADCAST(&np->cv);
1636 #else /* RX_ENABLE_LOCKS */
1638 #endif /* RX_ENABLE_LOCKS */
1640 MUTEX_EXIT(&rx_serverPool_lock);
1645 * One thing that seems to happen is that all the server threads get
1646 * tied up on some empty or slow call, and then a whole bunch of calls
1647 * arrive at once, using up the packet pool, so now there are more
1648 * empty calls. The most critical resources here are server threads
1649 * and the free packet pool. The "doreclaim" code seems to help in
1650 * general. I think that eventually we arrive in this state: there
1651 * are lots of pending calls which do have all their packets present,
1652 * so they won't be reclaimed, are multi-packet calls, so they won't
1653 * be scheduled until later, and thus are tying up most of the free
1654 * packet pool for a very long time.
1656 * 1. schedule multi-packet calls if all the packets are present.
1657 * Probably CPU-bound operation, useful to return packets to pool.
1658 * Do what if there is a full window, but the last packet isn't here?
1659 * 3. preserve one thread which *only* runs "best" calls, otherwise
1660 * it sleeps and waits for that type of call.
1661 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1662 * the current dataquota business is badly broken. The quota isn't adjusted
1663 * to reflect how many packets are presently queued for a running call.
1664 * So, when we schedule a queued call with a full window of packets queued
1665 * up for it, that *should* free up a window full of packets for other 2d-class
1666 * calls to be able to use from the packet pool. But it doesn't.
1668 * NB. Most of the time, this code doesn't run -- since idle server threads
1669 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1670 * as a new call arrives.
1672 /* Sleep until a call arrives. Returns a pointer to the call, ready
1673 * for an rx_Read. */
1674 #ifdef RX_ENABLE_LOCKS
1676 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1678 struct rx_serverQueueEntry *sq;
1679 struct rx_call *call = (struct rx_call *)0;
1680 struct rx_service *service = NULL;
1683 MUTEX_ENTER(&freeSQEList_lock);
1685 if ((sq = rx_FreeSQEList)) {
1686 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1687 MUTEX_EXIT(&freeSQEList_lock);
1688 } else { /* otherwise allocate a new one and return that */
1689 MUTEX_EXIT(&freeSQEList_lock);
1690 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1691 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1692 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1695 MUTEX_ENTER(&rx_serverPool_lock);
1696 if (cur_service != NULL) {
1697 ReturnToServerPool(cur_service);
1700 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1701 struct rx_call *tcall, *ncall, *choice2 = NULL;
1703 /* Scan for eligible incoming calls. A call is not eligible
1704 * if the maximum number of calls for its service type are
1705 * already executing */
1706 /* One thread will process calls FCFS (to prevent starvation),
1707 * while the other threads may run ahead looking for calls which
1708 * have all their input data available immediately. This helps
1709 * keep threads from blocking, waiting for data from the client. */
1710 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1711 service = tcall->conn->service;
1712 if (!QuotaOK(service)) {
1715 MUTEX_ENTER(&rx_pthread_mutex);
1716 if (tno == rxi_fcfs_thread_num
1717 || !tcall->queue_item_header.next) {
1718 MUTEX_EXIT(&rx_pthread_mutex);
1719 /* If we're the fcfs thread , then we'll just use
1720 * this call. If we haven't been able to find an optimal
1721 * choice, and we're at the end of the list, then use a
1722 * 2d choice if one has been identified. Otherwise... */
1723 call = (choice2 ? choice2 : tcall);
1724 service = call->conn->service;
1726 MUTEX_EXIT(&rx_pthread_mutex);
1727 if (!queue_IsEmpty(&tcall->rq)) {
1728 struct rx_packet *rp;
1729 rp = queue_First(&tcall->rq, rx_packet);
1730 if (rp->header.seq == 1) {
1732 || (rp->header.flags & RX_LAST_PACKET)) {
1734 } else if (rxi_2dchoice && !choice2
1735 && !(tcall->flags & RX_CALL_CLEARED)
1736 && (tcall->rprev > rxi_HardAckRate)) {
1746 ReturnToServerPool(service);
1753 MUTEX_EXIT(&rx_serverPool_lock);
1754 MUTEX_ENTER(&call->lock);
1756 if (call->flags & RX_CALL_WAIT_PROC) {
1757 call->flags &= ~RX_CALL_WAIT_PROC;
1758 MUTEX_ENTER(&rx_waiting_mutex);
1760 MUTEX_EXIT(&rx_waiting_mutex);
1763 if (call->state != RX_STATE_PRECALL || call->error) {
1764 MUTEX_EXIT(&call->lock);
1765 MUTEX_ENTER(&rx_serverPool_lock);
1766 ReturnToServerPool(service);
1771 if (queue_IsEmpty(&call->rq)
1772 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1773 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1775 CLEAR_CALL_QUEUE_LOCK(call);
1778 /* If there are no eligible incoming calls, add this process
1779 * to the idle server queue, to wait for one */
1783 *socketp = OSI_NULLSOCKET;
1785 sq->socketp = socketp;
1786 queue_Append(&rx_idleServerQueue, sq);
1787 #ifndef AFS_AIX41_ENV
1788 rx_waitForPacket = sq;
1790 rx_waitingForPacket = sq;
1791 #endif /* AFS_AIX41_ENV */
1793 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1795 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1796 MUTEX_EXIT(&rx_serverPool_lock);
1797 return (struct rx_call *)0;
1800 } while (!(call = sq->newcall)
1801 && !(socketp && *socketp != OSI_NULLSOCKET));
1802 MUTEX_EXIT(&rx_serverPool_lock);
1804 MUTEX_ENTER(&call->lock);
1810 MUTEX_ENTER(&freeSQEList_lock);
1811 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1812 rx_FreeSQEList = sq;
1813 MUTEX_EXIT(&freeSQEList_lock);
1816 clock_GetTime(&call->startTime);
1817 call->state = RX_STATE_ACTIVE;
1818 call->mode = RX_MODE_RECEIVING;
1819 #ifdef RX_KERNEL_TRACE
1820 if (ICL_SETACTIVE(afs_iclSetp)) {
1821 int glockOwner = ISAFS_GLOCK();
1824 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1825 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1832 rxi_calltrace(RX_CALL_START, call);
1833 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
1834 call->conn->service->servicePort, call->conn->service->serviceId,
1837 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1838 MUTEX_EXIT(&call->lock);
1840 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1845 #else /* RX_ENABLE_LOCKS */
1847 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1849 struct rx_serverQueueEntry *sq;
1850 struct rx_call *call = (struct rx_call *)0, *choice2;
1851 struct rx_service *service = NULL;
1855 MUTEX_ENTER(&freeSQEList_lock);
1857 if ((sq = rx_FreeSQEList)) {
1858 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1859 MUTEX_EXIT(&freeSQEList_lock);
1860 } else { /* otherwise allocate a new one and return that */
1861 MUTEX_EXIT(&freeSQEList_lock);
1862 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1863 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1864 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1866 MUTEX_ENTER(&sq->lock);
1868 if (cur_service != NULL) {
1869 cur_service->nRequestsRunning--;
1870 MUTEX_ENTER(&rx_quota_mutex);
1871 if (cur_service->nRequestsRunning < cur_service->minProcs)
1874 MUTEX_EXIT(&rx_quota_mutex);
1876 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1877 struct rx_call *tcall, *ncall;
1878 /* Scan for eligible incoming calls. A call is not eligible
1879 * if the maximum number of calls for its service type are
1880 * already executing */
1881 /* One thread will process calls FCFS (to prevent starvation),
1882 * while the other threads may run ahead looking for calls which
1883 * have all their input data available immediately. This helps
1884 * keep threads from blocking, waiting for data from the client. */
1885 choice2 = (struct rx_call *)0;
1886 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1887 service = tcall->conn->service;
1888 if (QuotaOK(service)) {
1889 MUTEX_ENTER(&rx_pthread_mutex);
1890 if (tno == rxi_fcfs_thread_num
1891 || !tcall->queue_item_header.next) {
1892 MUTEX_EXIT(&rx_pthread_mutex);
1893 /* If we're the fcfs thread, then we'll just use
1894 * this call. If we haven't been able to find an optimal
1895 * choice, and we're at the end of the list, then use a
1896 * 2d choice if one has been identified. Otherwise... */
1897 call = (choice2 ? choice2 : tcall);
1898 service = call->conn->service;
1900 MUTEX_EXIT(&rx_pthread_mutex);
1901 if (!queue_IsEmpty(&tcall->rq)) {
1902 struct rx_packet *rp;
1903 rp = queue_First(&tcall->rq, rx_packet);
1904 if (rp->header.seq == 1
1906 || (rp->header.flags & RX_LAST_PACKET))) {
1908 } else if (rxi_2dchoice && !choice2
1909 && !(tcall->flags & RX_CALL_CLEARED)
1910 && (tcall->rprev > rxi_HardAckRate)) {
1924 /* we can't schedule a call if there's no data!!! */
1925 /* send an ack if there's no data, if we're missing the
1926 * first packet, or we're missing something between first
1927 * and last -- there's a "hole" in the incoming data. */
1928 if (queue_IsEmpty(&call->rq)
1929 || queue_First(&call->rq, rx_packet)->header.seq != 1
1930 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1931 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1933 call->flags &= (~RX_CALL_WAIT_PROC);
1934 service->nRequestsRunning++;
1935 /* just started call in minProcs pool, need fewer to maintain
1937 MUTEX_ENTER(&rx_quota_mutex);
1938 if (service->nRequestsRunning <= service->minProcs)
1941 MUTEX_EXIT(&rx_quota_mutex);
1943 /* MUTEX_EXIT(&call->lock); */
1945 /* If there are no eligible incoming calls, add this process
1946 * to the idle server queue, to wait for one */
1949 *socketp = OSI_NULLSOCKET;
1951 sq->socketp = socketp;
1952 queue_Append(&rx_idleServerQueue, sq);
1956 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1958 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1959 return (struct rx_call *)0;
1962 } while (!(call = sq->newcall)
1963 && !(socketp && *socketp != OSI_NULLSOCKET));
1965 MUTEX_EXIT(&sq->lock);
1967 MUTEX_ENTER(&freeSQEList_lock);
1968 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1969 rx_FreeSQEList = sq;
1970 MUTEX_EXIT(&freeSQEList_lock);
1973 clock_GetTime(&call->startTime);
1974 call->state = RX_STATE_ACTIVE;
1975 call->mode = RX_MODE_RECEIVING;
1976 #ifdef RX_KERNEL_TRACE
1977 if (ICL_SETACTIVE(afs_iclSetp)) {
1978 int glockOwner = ISAFS_GLOCK();
1981 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1982 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1989 rxi_calltrace(RX_CALL_START, call);
1990 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
1991 call->conn->service->servicePort, call->conn->service->serviceId,
1994 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2001 #endif /* RX_ENABLE_LOCKS */
2005 /* Establish a procedure to be called when a packet arrives for a
2006 * call. This routine will be called at most once after each call,
2007 * and will also be called if there is an error condition on the or
2008 * the call is complete. Used by multi rx to build a selection
2009 * function which determines which of several calls is likely to be a
2010 * good one to read from.
2011 * NOTE: the way this is currently implemented it is probably only a
2012 * good idea to (1) use it immediately after a newcall (clients only)
2013 * and (2) only use it once. Other uses currently void your warranty
2016 rx_SetArrivalProc(struct rx_call *call,
2017 void (*proc) (struct rx_call * call,
2020 void * handle, int arg)
2022 call->arrivalProc = proc;
2023 call->arrivalProcHandle = handle;
2024 call->arrivalProcArg = arg;
2027 /* Call is finished (possibly prematurely). Return rc to the peer, if
2028 * appropriate, and return the final error code from the conversation
2032 rx_EndCall(struct rx_call *call, afs_int32 rc)
2034 struct rx_connection *conn = call->conn;
2038 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2039 call, rc, call->error, call->abortCode));
2042 MUTEX_ENTER(&call->lock);
2044 if (rc == 0 && call->error == 0) {
2045 call->abortCode = 0;
2046 call->abortCount = 0;
2049 call->arrivalProc = (void (*)())0;
2050 if (rc && call->error == 0) {
2051 rxi_CallError(call, rc);
2052 /* Send an abort message to the peer if this error code has
2053 * only just been set. If it was set previously, assume the
2054 * peer has already been sent the error code or will request it
2056 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2058 if (conn->type == RX_SERVER_CONNECTION) {
2059 /* Make sure reply or at least dummy reply is sent */
2060 if (call->mode == RX_MODE_RECEIVING) {
2061 rxi_WriteProc(call, 0, 0);
2063 if (call->mode == RX_MODE_SENDING) {
2064 rxi_FlushWrite(call);
2066 rxi_calltrace(RX_CALL_END, call);
2067 /* Call goes to hold state until reply packets are acknowledged */
2068 if (call->tfirst + call->nSoftAcked < call->tnext) {
2069 call->state = RX_STATE_HOLD;
2071 call->state = RX_STATE_DALLY;
2072 rxi_ClearTransmitQueue(call, 0);
2073 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
2074 rxevent_Cancel(call->keepAliveEvent, call,
2075 RX_CALL_REFCOUNT_ALIVE);
2077 } else { /* Client connection */
2079 /* Make sure server receives input packets, in the case where
2080 * no reply arguments are expected */
2081 if ((call->mode == RX_MODE_SENDING)
2082 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2083 (void)rxi_ReadProc(call, &dummy, 1);
2086 /* If we had an outstanding delayed ack, be nice to the server
2087 * and force-send it now.
2089 if (call->delayedAckEvent) {
2090 rxevent_Cancel(call->delayedAckEvent, call,
2091 RX_CALL_REFCOUNT_DELAY);
2092 call->delayedAckEvent = NULL;
2093 rxi_SendDelayedAck(NULL, call, NULL);
2096 /* We need to release the call lock since it's lower than the
2097 * conn_call_lock and we don't want to hold the conn_call_lock
2098 * over the rx_ReadProc call. The conn_call_lock needs to be held
2099 * here for the case where rx_NewCall is perusing the calls on
2100 * the connection structure. We don't want to signal until
2101 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2102 * have checked this call, found it active and by the time it
2103 * goes to sleep, will have missed the signal.
2105 MUTEX_EXIT(&call->lock);
2106 MUTEX_ENTER(&conn->conn_call_lock);
2107 MUTEX_ENTER(&call->lock);
2108 MUTEX_ENTER(&conn->conn_data_lock);
2109 conn->flags |= RX_CONN_BUSY;
2110 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2111 MUTEX_EXIT(&conn->conn_data_lock);
2112 #ifdef RX_ENABLE_LOCKS
2113 CV_BROADCAST(&conn->conn_call_cv);
2118 #ifdef RX_ENABLE_LOCKS
2120 MUTEX_EXIT(&conn->conn_data_lock);
2122 #endif /* RX_ENABLE_LOCKS */
2123 call->state = RX_STATE_DALLY;
2125 error = call->error;
2127 /* currentPacket, nLeft, and NFree must be zeroed here, because
2128 * ResetCall cannot: ResetCall may be called at splnet(), in the
2129 * kernel version, and may interrupt the macros rx_Read or
2130 * rx_Write, which run at normal priority for efficiency. */
2131 if (call->currentPacket) {
2132 #ifdef RX_TRACK_PACKETS
2133 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2135 rxi_FreePacket(call->currentPacket);
2136 call->currentPacket = (struct rx_packet *)0;
2139 call->nLeft = call->nFree = call->curlen = 0;
2141 /* Free any packets from the last call to ReadvProc/WritevProc */
2142 #ifdef RXDEBUG_PACKET
2144 #endif /* RXDEBUG_PACKET */
2145 rxi_FreePackets(0, &call->iovq);
2147 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2148 MUTEX_EXIT(&call->lock);
2149 if (conn->type == RX_CLIENT_CONNECTION) {
2150 MUTEX_ENTER(&conn->conn_data_lock);
2151 conn->flags &= ~RX_CONN_BUSY;
2152 MUTEX_EXIT(&conn->conn_data_lock);
2153 MUTEX_EXIT(&conn->conn_call_lock);
2157 * Map errors to the local host's errno.h format.
2159 error = ntoh_syserr_conv(error);
2163 #if !defined(KERNEL)
2165 /* Call this routine when shutting down a server or client (especially
2166 * clients). This will allow Rx to gracefully garbage collect server
2167 * connections, and reduce the number of retries that a server might
2168 * make to a dead client.
2169 * This is not quite right, since some calls may still be ongoing and
2170 * we can't lock them to destroy them. */
2174 struct rx_connection **conn_ptr, **conn_end;
2178 if (rxinit_status == 1) {
2180 return; /* Already shutdown. */
2182 rxi_DeleteCachedConnections();
2183 if (rx_connHashTable) {
2184 MUTEX_ENTER(&rx_connHashTable_lock);
2185 for (conn_ptr = &rx_connHashTable[0], conn_end =
2186 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2188 struct rx_connection *conn, *next;
2189 for (conn = *conn_ptr; conn; conn = next) {
2191 if (conn->type == RX_CLIENT_CONNECTION) {
2192 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2194 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2195 #ifdef RX_ENABLE_LOCKS
2196 rxi_DestroyConnectionNoLock(conn);
2197 #else /* RX_ENABLE_LOCKS */
2198 rxi_DestroyConnection(conn);
2199 #endif /* RX_ENABLE_LOCKS */
2203 #ifdef RX_ENABLE_LOCKS
2204 while (rx_connCleanup_list) {
2205 struct rx_connection *conn;
2206 conn = rx_connCleanup_list;
2207 rx_connCleanup_list = rx_connCleanup_list->next;
2208 MUTEX_EXIT(&rx_connHashTable_lock);
2209 rxi_CleanupConnection(conn);
2210 MUTEX_ENTER(&rx_connHashTable_lock);
2212 MUTEX_EXIT(&rx_connHashTable_lock);
2213 #endif /* RX_ENABLE_LOCKS */
2218 afs_winsockCleanup();
2226 /* if we wakeup packet waiter too often, can get in loop with two
2227 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2229 rxi_PacketsUnWait(void)
2231 if (!rx_waitingForPackets) {
2235 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2236 return; /* still over quota */
2239 rx_waitingForPackets = 0;
2240 #ifdef RX_ENABLE_LOCKS
2241 CV_BROADCAST(&rx_waitingForPackets_cv);
2243 osi_rxWakeup(&rx_waitingForPackets);
2249 /* ------------------Internal interfaces------------------------- */
2251 /* Return this process's service structure for the
2252 * specified socket and service */
2254 rxi_FindService(osi_socket socket, u_short serviceId)
2256 struct rx_service **sp;
2257 for (sp = &rx_services[0]; *sp; sp++) {
2258 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2264 #ifdef RXDEBUG_PACKET
2265 #ifdef KDUMP_RX_LOCK
2266 static struct rx_call_rx_lock *rx_allCallsp = 0;
2268 static struct rx_call *rx_allCallsp = 0;
2270 #endif /* RXDEBUG_PACKET */
2272 /* Allocate a call structure, for the indicated channel of the
2273 * supplied connection. The mode and state of the call must be set by
2274 * the caller. Returns the call with mutex locked. */
2276 rxi_NewCall(struct rx_connection *conn, int channel)
2278 struct rx_call *call;
2279 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2280 struct rx_call *cp; /* Call pointer temp */
2281 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2282 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2284 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2286 /* Grab an existing call structure, or allocate a new one.
2287 * Existing call structures are assumed to have been left reset by
2289 MUTEX_ENTER(&rx_freeCallQueue_lock);
2291 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2293 * EXCEPT that the TQ might not yet be cleared out.
2294 * Skip over those with in-use TQs.
2297 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2298 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2304 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2305 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2306 call = queue_First(&rx_freeCallQueue, rx_call);
2307 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2309 if (rx_stats_active)
2310 rx_MutexDecrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2311 MUTEX_EXIT(&rx_freeCallQueue_lock);
2312 MUTEX_ENTER(&call->lock);
2313 CLEAR_CALL_QUEUE_LOCK(call);
2314 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2315 /* Now, if TQ wasn't cleared earlier, do it now. */
2316 rxi_WaitforTQBusy(call);
2317 if (call->flags & RX_CALL_TQ_CLEARME) {
2318 rxi_ClearTransmitQueue(call, 1);
2319 /*queue_Init(&call->tq);*/
2321 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2322 /* Bind the call to its connection structure */
2324 rxi_ResetCall(call, 1);
2327 call = rxi_Alloc(sizeof(struct rx_call));
2328 #ifdef RXDEBUG_PACKET
2329 call->allNextp = rx_allCallsp;
2330 rx_allCallsp = call;
2332 #endif /* RXDEBUG_PACKET */
2333 rx_MutexIncrement(rx_stats.nCallStructs, rx_stats_mutex);
2335 MUTEX_EXIT(&rx_freeCallQueue_lock);
2336 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2337 MUTEX_ENTER(&call->lock);
2338 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2339 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2340 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2342 /* Initialize once-only items */
2343 queue_Init(&call->tq);
2344 queue_Init(&call->rq);
2345 queue_Init(&call->iovq);
2346 #ifdef RXDEBUG_PACKET
2347 call->rqc = call->tqc = call->iovqc = 0;
2348 #endif /* RXDEBUG_PACKET */
2349 /* Bind the call to its connection structure (prereq for reset) */
2351 rxi_ResetCall(call, 1);
2353 call->channel = channel;
2354 call->callNumber = &conn->callNumber[channel];
2355 call->rwind = conn->rwind[channel];
2356 call->twind = conn->twind[channel];
2357 /* Note that the next expected call number is retained (in
2358 * conn->callNumber[i]), even if we reallocate the call structure
2360 conn->call[channel] = call;
2361 /* if the channel's never been used (== 0), we should start at 1, otherwise
2362 * the call number is valid from the last time this channel was used */
2363 if (*call->callNumber == 0)
2364 *call->callNumber = 1;
2369 /* A call has been inactive long enough that so we can throw away
2370 * state, including the call structure, which is placed on the call
2372 * Call is locked upon entry.
2373 * haveCTLock set if called from rxi_ReapConnections
2375 #ifdef RX_ENABLE_LOCKS
2377 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2378 #else /* RX_ENABLE_LOCKS */
2380 rxi_FreeCall(struct rx_call *call)
2381 #endif /* RX_ENABLE_LOCKS */
2383 int channel = call->channel;
2384 struct rx_connection *conn = call->conn;
2387 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2388 (*call->callNumber)++;
2389 rxi_ResetCall(call, 0);
2390 call->conn->call[channel] = (struct rx_call *)0;
2392 MUTEX_ENTER(&rx_freeCallQueue_lock);
2393 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2394 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2395 /* A call may be free even though its transmit queue is still in use.
2396 * Since we search the call list from head to tail, put busy calls at
2397 * the head of the list, and idle calls at the tail.
2399 if (call->flags & RX_CALL_TQ_BUSY)
2400 queue_Prepend(&rx_freeCallQueue, call);
2402 queue_Append(&rx_freeCallQueue, call);
2403 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2404 queue_Append(&rx_freeCallQueue, call);
2405 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2406 if (rx_stats_active)
2407 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2408 MUTEX_EXIT(&rx_freeCallQueue_lock);
2410 /* Destroy the connection if it was previously slated for
2411 * destruction, i.e. the Rx client code previously called
2412 * rx_DestroyConnection (client connections), or
2413 * rxi_ReapConnections called the same routine (server
2414 * connections). Only do this, however, if there are no
2415 * outstanding calls. Note that for fine grain locking, there appears
2416 * to be a deadlock in that rxi_FreeCall has a call locked and
2417 * DestroyConnectionNoLock locks each call in the conn. But note a
2418 * few lines up where we have removed this call from the conn.
2419 * If someone else destroys a connection, they either have no
2420 * call lock held or are going through this section of code.
2422 MUTEX_ENTER(&conn->conn_data_lock);
2423 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2425 MUTEX_EXIT(&conn->conn_data_lock);
2426 #ifdef RX_ENABLE_LOCKS
2428 rxi_DestroyConnectionNoLock(conn);
2430 rxi_DestroyConnection(conn);
2431 #else /* RX_ENABLE_LOCKS */
2432 rxi_DestroyConnection(conn);
2433 #endif /* RX_ENABLE_LOCKS */
2435 MUTEX_EXIT(&conn->conn_data_lock);
2439 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2441 rxi_Alloc(size_t size)
2445 if (rx_stats_active)
2446 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2449 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2450 afs_osi_Alloc_NoSleep(size);
2455 osi_Panic("rxi_Alloc error");
2461 rxi_Free(void *addr, size_t size)
2463 if (rx_stats_active)
2464 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2465 osi_Free(addr, size);
2469 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2471 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2472 struct rx_peer *next = NULL;
2476 MUTEX_ENTER(&rx_peerHashTable_lock);
2478 peer_ptr = &rx_peerHashTable[0];
2479 peer_end = &rx_peerHashTable[rx_hashTableSize];
2482 for ( ; peer_ptr < peer_end; peer_ptr++) {
2485 for ( ; peer; peer = next) {
2487 if (host == peer->host)
2492 hashIndex = PEER_HASH(host, port);
2493 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2494 if ((peer->host == host) && (peer->port == port))
2499 MUTEX_ENTER(&rx_peerHashTable_lock);
2504 MUTEX_EXIT(&rx_peerHashTable_lock);
2506 MUTEX_ENTER(&peer->peer_lock);
2507 /* We don't handle dropping below min, so don't */
2508 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2509 peer->ifMTU=MIN(mtu, peer->ifMTU);
2510 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2511 /* if we tweaked this down, need to tune our peer MTU too */
2512 peer->MTU = MIN(peer->MTU, peer->natMTU);
2513 /* if we discovered a sub-1500 mtu, degrade */
2514 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2515 peer->maxDgramPackets = 1;
2516 /* We no longer have valid peer packet information */
2517 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2518 peer->maxPacketSize = 0;
2519 MUTEX_EXIT(&peer->peer_lock);
2521 MUTEX_ENTER(&rx_peerHashTable_lock);
2523 if (host && !port) {
2525 /* pick up where we left off */
2529 MUTEX_EXIT(&rx_peerHashTable_lock);
2532 /* Find the peer process represented by the supplied (host,port)
2533 * combination. If there is no appropriate active peer structure, a
2534 * new one will be allocated and initialized
2535 * The origPeer, if set, is a pointer to a peer structure on which the
2536 * refcount will be be decremented. This is used to replace the peer
2537 * structure hanging off a connection structure */
2539 rxi_FindPeer(afs_uint32 host, u_short port,
2540 struct rx_peer *origPeer, int create)
2544 hashIndex = PEER_HASH(host, port);
2545 MUTEX_ENTER(&rx_peerHashTable_lock);
2546 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2547 if ((pp->host == host) && (pp->port == port))
2552 pp = rxi_AllocPeer(); /* This bzero's *pp */
2553 pp->host = host; /* set here or in InitPeerParams is zero */
2555 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2556 queue_Init(&pp->congestionQueue);
2557 queue_Init(&pp->rpcStats);
2558 pp->next = rx_peerHashTable[hashIndex];
2559 rx_peerHashTable[hashIndex] = pp;
2560 rxi_InitPeerParams(pp);
2561 if (rx_stats_active)
2562 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2569 origPeer->refCount--;
2570 MUTEX_EXIT(&rx_peerHashTable_lock);
2575 /* Find the connection at (host, port) started at epoch, and with the
2576 * given connection id. Creates the server connection if necessary.
2577 * The type specifies whether a client connection or a server
2578 * connection is desired. In both cases, (host, port) specify the
2579 * peer's (host, pair) pair. Client connections are not made
2580 * automatically by this routine. The parameter socket gives the
2581 * socket descriptor on which the packet was received. This is used,
2582 * in the case of server connections, to check that *new* connections
2583 * come via a valid (port, serviceId). Finally, the securityIndex
2584 * parameter must match the existing index for the connection. If a
2585 * server connection is created, it will be created using the supplied
2586 * index, if the index is valid for this service */
2587 struct rx_connection *
2588 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2589 u_short port, u_short serviceId, afs_uint32 cid,
2590 afs_uint32 epoch, int type, u_int securityIndex)
2592 int hashindex, flag, i;
2593 struct rx_connection *conn;
2594 hashindex = CONN_HASH(host, port, cid, epoch, type);
2595 MUTEX_ENTER(&rx_connHashTable_lock);
2596 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2597 rx_connHashTable[hashindex],
2600 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2601 && (epoch == conn->epoch)) {
2602 struct rx_peer *pp = conn->peer;
2603 if (securityIndex != conn->securityIndex) {
2604 /* this isn't supposed to happen, but someone could forge a packet
2605 * like this, and there seems to be some CM bug that makes this
2606 * happen from time to time -- in which case, the fileserver
2608 MUTEX_EXIT(&rx_connHashTable_lock);
2609 return (struct rx_connection *)0;
2611 if (pp->host == host && pp->port == port)
2613 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2615 /* So what happens when it's a callback connection? */
2616 if ( /*type == RX_CLIENT_CONNECTION && */
2617 (conn->epoch & 0x80000000))
2621 /* the connection rxLastConn that was used the last time is not the
2622 ** one we are looking for now. Hence, start searching in the hash */
2624 conn = rx_connHashTable[hashindex];
2629 struct rx_service *service;
2630 if (type == RX_CLIENT_CONNECTION) {
2631 MUTEX_EXIT(&rx_connHashTable_lock);
2632 return (struct rx_connection *)0;
2634 service = rxi_FindService(socket, serviceId);
2635 if (!service || (securityIndex >= service->nSecurityObjects)
2636 || (service->securityObjects[securityIndex] == 0)) {
2637 MUTEX_EXIT(&rx_connHashTable_lock);
2638 return (struct rx_connection *)0;
2640 conn = rxi_AllocConnection(); /* This bzero's the connection */
2641 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2642 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2643 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2644 conn->next = rx_connHashTable[hashindex];
2645 rx_connHashTable[hashindex] = conn;
2646 conn->peer = rxi_FindPeer(host, port, 0, 1);
2647 conn->type = RX_SERVER_CONNECTION;
2648 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2649 conn->epoch = epoch;
2650 conn->cid = cid & RX_CIDMASK;
2651 /* conn->serial = conn->lastSerial = 0; */
2652 /* conn->timeout = 0; */
2653 conn->ackRate = RX_FAST_ACK_RATE;
2654 conn->service = service;
2655 conn->serviceId = serviceId;
2656 conn->securityIndex = securityIndex;
2657 conn->securityObject = service->securityObjects[securityIndex];
2658 conn->nSpecific = 0;
2659 conn->specific = NULL;
2660 rx_SetConnDeadTime(conn, service->connDeadTime);
2661 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2662 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2663 for (i = 0; i < RX_MAXCALLS; i++) {
2664 conn->twind[i] = rx_initSendWindow;
2665 conn->rwind[i] = rx_initReceiveWindow;
2667 /* Notify security object of the new connection */
2668 RXS_NewConnection(conn->securityObject, conn);
2669 /* XXXX Connection timeout? */
2670 if (service->newConnProc)
2671 (*service->newConnProc) (conn);
2672 if (rx_stats_active)
2673 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2676 MUTEX_ENTER(&conn->conn_data_lock);
2678 MUTEX_EXIT(&conn->conn_data_lock);
2680 rxLastConn = conn; /* store this connection as the last conn used */
2681 MUTEX_EXIT(&rx_connHashTable_lock);
2685 /* There are two packet tracing routines available for testing and monitoring
2686 * Rx. One is called just after every packet is received and the other is
2687 * called just before every packet is sent. Received packets, have had their
2688 * headers decoded, and packets to be sent have not yet had their headers
2689 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2690 * containing the network address. Both can be modified. The return value, if
2691 * non-zero, indicates that the packet should be dropped. */
2693 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2694 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2696 /* A packet has been received off the interface. Np is the packet, socket is
2697 * the socket number it was received from (useful in determining which service
2698 * this packet corresponds to), and (host, port) reflect the host,port of the
2699 * sender. This call returns the packet to the caller if it is finished with
2700 * it, rather than de-allocating it, just as a small performance hack */
2703 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
2704 afs_uint32 host, u_short port, int *tnop,
2705 struct rx_call **newcallp)
2707 struct rx_call *call;
2708 struct rx_connection *conn;
2710 afs_uint32 currentCallNumber;
2716 struct rx_packet *tnp;
2719 /* We don't print out the packet until now because (1) the time may not be
2720 * accurate enough until now in the lwp implementation (rx_Listener only gets
2721 * the time after the packet is read) and (2) from a protocol point of view,
2722 * this is the first time the packet has been seen */
2723 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2724 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2725 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT,
2726 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2727 np->header.epoch, np->header.cid, np->header.callNumber,
2728 np->header.seq, np->header.flags, np));
2731 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2732 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2735 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2736 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2739 /* If an input tracer function is defined, call it with the packet and
2740 * network address. Note this function may modify its arguments. */
2741 if (rx_justReceived) {
2742 struct sockaddr_in addr;
2744 addr.sin_family = AF_INET;
2745 addr.sin_port = port;
2746 addr.sin_addr.s_addr = host;
2747 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2748 addr.sin_len = sizeof(addr);
2749 #endif /* AFS_OSF_ENV */
2750 drop = (*rx_justReceived) (np, &addr);
2751 /* drop packet if return value is non-zero */
2754 port = addr.sin_port; /* in case fcn changed addr */
2755 host = addr.sin_addr.s_addr;
2759 /* If packet was not sent by the client, then *we* must be the client */
2760 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2761 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2763 /* Find the connection (or fabricate one, if we're the server & if
2764 * necessary) associated with this packet */
2766 rxi_FindConnection(socket, host, port, np->header.serviceId,
2767 np->header.cid, np->header.epoch, type,
2768 np->header.securityIndex);
2771 /* If no connection found or fabricated, just ignore the packet.
2772 * (An argument could be made for sending an abort packet for
2777 MUTEX_ENTER(&conn->conn_data_lock);
2778 if (conn->maxSerial < np->header.serial)
2779 conn->maxSerial = np->header.serial;
2780 MUTEX_EXIT(&conn->conn_data_lock);
2782 /* If the connection is in an error state, send an abort packet and ignore
2783 * the incoming packet */
2785 /* Don't respond to an abort packet--we don't want loops! */
2786 MUTEX_ENTER(&conn->conn_data_lock);
2787 if (np->header.type != RX_PACKET_TYPE_ABORT)
2788 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2790 MUTEX_EXIT(&conn->conn_data_lock);
2794 /* Check for connection-only requests (i.e. not call specific). */
2795 if (np->header.callNumber == 0) {
2796 switch (np->header.type) {
2797 case RX_PACKET_TYPE_ABORT: {
2798 /* What if the supplied error is zero? */
2799 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2800 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2801 rxi_ConnectionError(conn, errcode);
2802 MUTEX_ENTER(&conn->conn_data_lock);
2804 MUTEX_EXIT(&conn->conn_data_lock);
2807 case RX_PACKET_TYPE_CHALLENGE:
2808 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2809 MUTEX_ENTER(&conn->conn_data_lock);
2811 MUTEX_EXIT(&conn->conn_data_lock);
2813 case RX_PACKET_TYPE_RESPONSE:
2814 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2815 MUTEX_ENTER(&conn->conn_data_lock);
2817 MUTEX_EXIT(&conn->conn_data_lock);
2819 case RX_PACKET_TYPE_PARAMS:
2820 case RX_PACKET_TYPE_PARAMS + 1:
2821 case RX_PACKET_TYPE_PARAMS + 2:
2822 /* ignore these packet types for now */
2823 MUTEX_ENTER(&conn->conn_data_lock);
2825 MUTEX_EXIT(&conn->conn_data_lock);
2830 /* Should not reach here, unless the peer is broken: send an
2832 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2833 MUTEX_ENTER(&conn->conn_data_lock);
2834 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2836 MUTEX_EXIT(&conn->conn_data_lock);
2841 channel = np->header.cid & RX_CHANNELMASK;
2842 call = conn->call[channel];
2843 #ifdef RX_ENABLE_LOCKS
2845 MUTEX_ENTER(&call->lock);
2846 /* Test to see if call struct is still attached to conn. */
2847 if (call != conn->call[channel]) {
2849 MUTEX_EXIT(&call->lock);
2850 if (type == RX_SERVER_CONNECTION) {
2851 call = conn->call[channel];
2852 /* If we started with no call attached and there is one now,
2853 * another thread is also running this routine and has gotten
2854 * the connection channel. We should drop this packet in the tests
2855 * below. If there was a call on this connection and it's now
2856 * gone, then we'll be making a new call below.
2857 * If there was previously a call and it's now different then
2858 * the old call was freed and another thread running this routine
2859 * has created a call on this channel. One of these two threads
2860 * has a packet for the old call and the code below handles those
2864 MUTEX_ENTER(&call->lock);
2866 /* This packet can't be for this call. If the new call address is
2867 * 0 then no call is running on this channel. If there is a call
2868 * then, since this is a client connection we're getting data for
2869 * it must be for the previous call.
2871 if (rx_stats_active)
2872 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2873 MUTEX_ENTER(&conn->conn_data_lock);
2875 MUTEX_EXIT(&conn->conn_data_lock);
2880 currentCallNumber = conn->callNumber[channel];
2882 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2883 if (np->header.callNumber < currentCallNumber) {
2884 if (rx_stats_active)
2885 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2886 #ifdef RX_ENABLE_LOCKS
2888 MUTEX_EXIT(&call->lock);
2890 MUTEX_ENTER(&conn->conn_data_lock);
2892 MUTEX_EXIT(&conn->conn_data_lock);
2896 MUTEX_ENTER(&conn->conn_call_lock);
2897 call = rxi_NewCall(conn, channel);
2898 MUTEX_EXIT(&conn->conn_call_lock);
2899 *call->callNumber = np->header.callNumber;
2901 if (np->header.callNumber == 0)
2902 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%.06d len %d",
2903 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
2904 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
2905 np->header.flags, np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2907 call->state = RX_STATE_PRECALL;
2908 clock_GetTime(&call->queueTime);
2909 hzero(call->bytesSent);
2910 hzero(call->bytesRcvd);
2912 * If the number of queued calls exceeds the overload
2913 * threshold then abort this call.
2915 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2916 struct rx_packet *tp;
2918 rxi_CallError(call, rx_BusyError);
2919 tp = rxi_SendCallAbort(call, np, 1, 0);
2920 MUTEX_EXIT(&call->lock);
2921 MUTEX_ENTER(&conn->conn_data_lock);
2923 MUTEX_EXIT(&conn->conn_data_lock);
2924 if (rx_stats_active)
2925 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2928 rxi_KeepAliveOn(call);
2929 } else if (np->header.callNumber != currentCallNumber) {
2930 /* Wait until the transmit queue is idle before deciding
2931 * whether to reset the current call. Chances are that the
2932 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2935 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2936 while ((call->state == RX_STATE_ACTIVE)
2937 && (call->flags & RX_CALL_TQ_BUSY)) {
2938 call->flags |= RX_CALL_TQ_WAIT;
2940 #ifdef RX_ENABLE_LOCKS
2941 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2942 CV_WAIT(&call->cv_tq, &call->lock);
2943 #else /* RX_ENABLE_LOCKS */
2944 osi_rxSleep(&call->tq);
2945 #endif /* RX_ENABLE_LOCKS */
2947 if (call->tqWaiters == 0)
2948 call->flags &= ~RX_CALL_TQ_WAIT;
2950 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2951 /* If the new call cannot be taken right now send a busy and set
2952 * the error condition in this call, so that it terminates as
2953 * quickly as possible */
2954 if (call->state == RX_STATE_ACTIVE) {
2955 struct rx_packet *tp;
2957 rxi_CallError(call, RX_CALL_DEAD);
2958 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2960 MUTEX_EXIT(&call->lock);
2961 MUTEX_ENTER(&conn->conn_data_lock);
2963 MUTEX_EXIT(&conn->conn_data_lock);
2966 rxi_ResetCall(call, 0);
2967 *call->callNumber = np->header.callNumber;
2969 if (np->header.callNumber == 0)
2970 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%06d len %d",
2971 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
2972 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
2973 np->header.flags, np, np->retryTime.sec, np->retryTime.usec, np->length));
2975 call->state = RX_STATE_PRECALL;
2976 clock_GetTime(&call->queueTime);
2977 hzero(call->bytesSent);
2978 hzero(call->bytesRcvd);
2980 * If the number of queued calls exceeds the overload
2981 * threshold then abort this call.
2983 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2984 struct rx_packet *tp;
2986 rxi_CallError(call, rx_BusyError);
2987 tp = rxi_SendCallAbort(call, np, 1, 0);
2988 MUTEX_EXIT(&call->lock);
2989 MUTEX_ENTER(&conn->conn_data_lock);
2991 MUTEX_EXIT(&conn->conn_data_lock);
2992 if (rx_stats_active)
2993 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2996 rxi_KeepAliveOn(call);
2998 /* Continuing call; do nothing here. */
3000 } else { /* we're the client */
3001 /* Ignore all incoming acknowledgements for calls in DALLY state */
3002 if (call && (call->state == RX_STATE_DALLY)
3003 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3004 if (rx_stats_active)
3005 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
3006 #ifdef RX_ENABLE_LOCKS
3008 MUTEX_EXIT(&call->lock);
3011 MUTEX_ENTER(&conn->conn_data_lock);
3013 MUTEX_EXIT(&conn->conn_data_lock);
3017 /* Ignore anything that's not relevant to the current call. If there
3018 * isn't a current call, then no packet is relevant. */
3019 if (!call || (np->header.callNumber != currentCallNumber)) {
3020 if (rx_stats_active)
3021 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
3022 #ifdef RX_ENABLE_LOCKS
3024 MUTEX_EXIT(&call->lock);
3027 MUTEX_ENTER(&conn->conn_data_lock);
3029 MUTEX_EXIT(&conn->conn_data_lock);
3032 /* If the service security object index stamped in the packet does not
3033 * match the connection's security index, ignore the packet */
3034 if (np->header.securityIndex != conn->securityIndex) {
3035 #ifdef RX_ENABLE_LOCKS
3036 MUTEX_EXIT(&call->lock);
3038 MUTEX_ENTER(&conn->conn_data_lock);
3040 MUTEX_EXIT(&conn->conn_data_lock);
3044 /* If we're receiving the response, then all transmit packets are
3045 * implicitly acknowledged. Get rid of them. */
3046 if (np->header.type == RX_PACKET_TYPE_DATA) {
3047 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3048 /* XXX Hack. Because we must release the global rx lock when
3049 * sending packets (osi_NetSend) we drop all acks while we're
3050 * traversing the tq in rxi_Start sending packets out because
3051 * packets may move to the freePacketQueue as result of being here!
3052 * So we drop these packets until we're safely out of the
3053 * traversing. Really ugly!
3054 * For fine grain RX locking, we set the acked field in the
3055 * packets and let rxi_Start remove them from the transmit queue.
3057 if (call->flags & RX_CALL_TQ_BUSY) {
3058 #ifdef RX_ENABLE_LOCKS
3059 rxi_SetAcksInTransmitQueue(call);
3062 return np; /* xmitting; drop packet */
3065 rxi_ClearTransmitQueue(call, 0);
3067 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3068 rxi_ClearTransmitQueue(call, 0);
3069 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3071 if (np->header.type == RX_PACKET_TYPE_ACK) {
3072 /* now check to see if this is an ack packet acknowledging that the
3073 * server actually *lost* some hard-acked data. If this happens we
3074 * ignore this packet, as it may indicate that the server restarted in
3075 * the middle of a call. It is also possible that this is an old ack
3076 * packet. We don't abort the connection in this case, because this
3077 * *might* just be an old ack packet. The right way to detect a server
3078 * restart in the midst of a call is to notice that the server epoch
3080 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3081 * XXX unacknowledged. I think that this is off-by-one, but
3082 * XXX I don't dare change it just yet, since it will
3083 * XXX interact badly with the server-restart detection
3084 * XXX code in receiveackpacket. */
3085 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3086 if (rx_stats_active)
3087 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
3088 MUTEX_EXIT(&call->lock);
3089 MUTEX_ENTER(&conn->conn_data_lock);
3091 MUTEX_EXIT(&conn->conn_data_lock);
3095 } /* else not a data packet */
3098 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3099 /* Set remote user defined status from packet */
3100 call->remoteStatus = np->header.userStatus;
3102 /* Note the gap between the expected next packet and the actual
3103 * packet that arrived, when the new packet has a smaller serial number
3104 * than expected. Rioses frequently reorder packets all by themselves,
3105 * so this will be quite important with very large window sizes.
3106 * Skew is checked against 0 here to avoid any dependence on the type of
3107 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3109 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3110 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3111 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3113 MUTEX_ENTER(&conn->conn_data_lock);
3114 skew = conn->lastSerial - np->header.serial;
3115 conn->lastSerial = np->header.serial;
3116 MUTEX_EXIT(&conn->conn_data_lock);
3118 struct rx_peer *peer;
3120 if (skew > peer->inPacketSkew) {
3121 dpf(("*** In skew changed from %d to %d\n",
3122 peer->inPacketSkew, skew));
3123 peer->inPacketSkew = skew;
3127 /* Now do packet type-specific processing */
3128 switch (np->header.type) {
3129 case RX_PACKET_TYPE_DATA:
3130 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3133 case RX_PACKET_TYPE_ACK:
3134 /* Respond immediately to ack packets requesting acknowledgement
3136 if (np->header.flags & RX_REQUEST_ACK) {
3138 (void)rxi_SendCallAbort(call, 0, 1, 0);
3140 (void)rxi_SendAck(call, 0, np->header.serial,
3141 RX_ACK_PING_RESPONSE, 1);
3143 np = rxi_ReceiveAckPacket(call, np, 1);
3145 case RX_PACKET_TYPE_ABORT: {
3146 /* An abort packet: reset the call, passing the error up to the user. */
3147 /* What if error is zero? */
3148 /* What if the error is -1? the application will treat it as a timeout. */
3149 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3150 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
3151 rxi_CallError(call, errdata);
3152 MUTEX_EXIT(&call->lock);
3153 MUTEX_ENTER(&conn->conn_data_lock);
3155 MUTEX_EXIT(&conn->conn_data_lock);
3156 return np; /* xmitting; drop packet */
3158 case RX_PACKET_TYPE_BUSY:
3161 case RX_PACKET_TYPE_ACKALL:
3162 /* All packets acknowledged, so we can drop all packets previously
3163 * readied for sending */
3164 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3165 /* XXX Hack. We because we can't release the global rx lock when
3166 * sending packets (osi_NetSend) we drop all ack pkts while we're
3167 * traversing the tq in rxi_Start sending packets out because
3168 * packets may move to the freePacketQueue as result of being
3169 * here! So we drop these packets until we're safely out of the
3170 * traversing. Really ugly!
3171 * For fine grain RX locking, we set the acked field in the packets
3172 * and let rxi_Start remove the packets from the transmit queue.
3174 if (call->flags & RX_CALL_TQ_BUSY) {
3175 #ifdef RX_ENABLE_LOCKS
3176 rxi_SetAcksInTransmitQueue(call);
3178 #else /* RX_ENABLE_LOCKS */
3179 MUTEX_EXIT(&call->lock);
3180 MUTEX_ENTER(&conn->conn_data_lock);
3182 MUTEX_EXIT(&conn->conn_data_lock);
3183 return np; /* xmitting; drop packet */
3184 #endif /* RX_ENABLE_LOCKS */
3186 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3187 rxi_ClearTransmitQueue(call, 0);
3188 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3191 /* Should not reach here, unless the peer is broken: send an abort
3193 rxi_CallError(call, RX_PROTOCOL_ERROR);
3194 np = rxi_SendCallAbort(call, np, 1, 0);
3197 /* Note when this last legitimate packet was received, for keep-alive
3198 * processing. Note, we delay getting the time until now in the hope that
3199 * the packet will be delivered to the user before any get time is required
3200 * (if not, then the time won't actually be re-evaluated here). */
3201 call->lastReceiveTime = clock_Sec();
3202 MUTEX_EXIT(&call->lock);
3203 MUTEX_ENTER(&conn->conn_data_lock);
3205 MUTEX_EXIT(&conn->conn_data_lock);
3209 /* return true if this is an "interesting" connection from the point of view
3210 of someone trying to debug the system */
3212 rxi_IsConnInteresting(struct rx_connection *aconn)
3215 struct rx_call *tcall;
3217 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3220 for (i = 0; i < RX_MAXCALLS; i++) {
3221 tcall = aconn->call[i];
3223 if ((tcall->state == RX_STATE_PRECALL)
3224 || (tcall->state == RX_STATE_ACTIVE))
3226 if ((tcall->mode == RX_MODE_SENDING)
3227 || (tcall->mode == RX_MODE_RECEIVING))
3235 /* if this is one of the last few packets AND it wouldn't be used by the
3236 receiving call to immediately satisfy a read request, then drop it on
3237 the floor, since accepting it might prevent a lock-holding thread from
3238 making progress in its reading. If a call has been cleared while in
3239 the precall state then ignore all subsequent packets until the call
3240 is assigned to a thread. */
3243 TooLow(struct rx_packet *ap, struct rx_call *acall)
3247 MUTEX_ENTER(&rx_quota_mutex);
3248 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3249 && (acall->state == RX_STATE_PRECALL))
3250 || ((rx_nFreePackets < rxi_dataQuota + 2)
3251 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3252 && (acall->flags & RX_CALL_READER_WAIT)))) {
3255 MUTEX_EXIT(&rx_quota_mutex);
3261 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3263 struct rx_connection *conn = arg1;
3264 struct rx_call *acall = arg2;
3265 struct rx_call *call = acall;
3266 struct clock when, now;
3269 MUTEX_ENTER(&conn->conn_data_lock);
3270 conn->checkReachEvent = NULL;
3271 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3274 MUTEX_EXIT(&conn->conn_data_lock);
3278 MUTEX_ENTER(&conn->conn_call_lock);
3279 MUTEX_ENTER(&conn->conn_data_lock);
3280 for (i = 0; i < RX_MAXCALLS; i++) {
3281 struct rx_call *tc = conn->call[i];
3282 if (tc && tc->state == RX_STATE_PRECALL) {
3288 /* Indicate that rxi_CheckReachEvent is no longer running by
3289 * clearing the flag. Must be atomic under conn_data_lock to
3290 * avoid a new call slipping by: rxi_CheckConnReach holds
3291 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3293 conn->flags &= ~RX_CONN_ATTACHWAIT;
3294 MUTEX_EXIT(&conn->conn_data_lock);
3295 MUTEX_EXIT(&conn->conn_call_lock);
3300 MUTEX_ENTER(&call->lock);
3301 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3303 MUTEX_EXIT(&call->lock);
3305 clock_GetTime(&now);
3307 when.sec += RX_CHECKREACH_TIMEOUT;
3308 MUTEX_ENTER(&conn->conn_data_lock);
3309 if (!conn->checkReachEvent) {
3311 conn->checkReachEvent =
3312 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3315 MUTEX_EXIT(&conn->conn_data_lock);
3321 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3323 struct rx_service *service = conn->service;
3324 struct rx_peer *peer = conn->peer;
3325 afs_uint32 now, lastReach;
3327 if (service->checkReach == 0)
3331 MUTEX_ENTER(&peer->peer_lock);
3332 lastReach = peer->lastReachTime;
3333 MUTEX_EXIT(&peer->peer_lock);
3334 if (now - lastReach < RX_CHECKREACH_TTL)
3337 MUTEX_ENTER(&conn->conn_data_lock);
3338 if (conn->flags & RX_CONN_ATTACHWAIT) {
3339 MUTEX_EXIT(&conn->conn_data_lock);
3342 conn->flags |= RX_CONN_ATTACHWAIT;
3343 MUTEX_EXIT(&conn->conn_data_lock);
3344 if (!conn->checkReachEvent)
3345 rxi_CheckReachEvent(NULL, conn, call);
3350 /* try to attach call, if authentication is complete */
3352 TryAttach(struct rx_call *acall, osi_socket socket,
3353 int *tnop, struct rx_call **newcallp,
3356 struct rx_connection *conn = acall->conn;
3358 if (conn->type == RX_SERVER_CONNECTION
3359 && acall->state == RX_STATE_PRECALL) {
3360 /* Don't attach until we have any req'd. authentication. */
3361 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3362 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3363 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3364 /* Note: this does not necessarily succeed; there
3365 * may not any proc available
3368 rxi_ChallengeOn(acall->conn);
3373 /* A data packet has been received off the interface. This packet is
3374 * appropriate to the call (the call is in the right state, etc.). This
3375 * routine can return a packet to the caller, for re-use */
3378 rxi_ReceiveDataPacket(struct rx_call *call,
3379 struct rx_packet *np, int istack,
3380 osi_socket socket, afs_uint32 host, u_short port,
3381 int *tnop, struct rx_call **newcallp)
3383 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3388 afs_uint32 serial=0, flags=0;
3390 struct rx_packet *tnp;
3391 struct clock when, now;
3392 if (rx_stats_active)
3393 rx_MutexIncrement(rx_stats.dataPacketsRead, rx_stats_mutex);
3396 /* If there are no packet buffers, drop this new packet, unless we can find
3397 * packet buffers from inactive calls */
3399 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3400 MUTEX_ENTER(&rx_freePktQ_lock);
3401 rxi_NeedMorePackets = TRUE;
3402 MUTEX_EXIT(&rx_freePktQ_lock);
3403 if (rx_stats_active)
3404 rx_MutexIncrement(rx_stats.noPacketBuffersOnRead, rx_stats_mutex);
3405 call->rprev = np->header.serial;
3406 rxi_calltrace(RX_TRACE_DROP, call);
3407 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems", np));
3409 rxi_ClearReceiveQueue(call);
3410 clock_GetTime(&now);
3412 clock_Add(&when, &rx_softAckDelay);
3413 if (!call->delayedAckEvent
3414 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3415 rxevent_Cancel(call->delayedAckEvent, call,
3416 RX_CALL_REFCOUNT_DELAY);
3417 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3418 call->delayedAckEvent =
3419 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3421 /* we've damaged this call already, might as well do it in. */
3427 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3428 * packet is one of several packets transmitted as a single
3429 * datagram. Do not send any soft or hard acks until all packets
3430 * in a jumbogram have been processed. Send negative acks right away.
3432 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3433 /* tnp is non-null when there are more packets in the
3434 * current jumbo gram */
3441 seq = np->header.seq;
3442 serial = np->header.serial;
3443 flags = np->header.flags;
3445 /* If the call is in an error state, send an abort message */
3447 return rxi_SendCallAbort(call, np, istack, 0);
3449 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3450 * AFS 3.5 jumbogram. */
3451 if (flags & RX_JUMBO_PACKET) {
3452 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3457 if (np->header.spare != 0) {
3458 MUTEX_ENTER(&call->conn->conn_data_lock);
3459 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3460 MUTEX_EXIT(&call->conn->conn_data_lock);
3463 /* The usual case is that this is the expected next packet */
3464 if (seq == call->rnext) {
3466 /* Check to make sure it is not a duplicate of one already queued */
3467 if (queue_IsNotEmpty(&call->rq)
3468 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3469 if (rx_stats_active)
3470 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3471 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate", np));
3472 rxevent_Cancel(call->delayedAckEvent, call,
3473 RX_CALL_REFCOUNT_DELAY);
3474 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3480 /* It's the next packet. Stick it on the receive queue
3481 * for this call. Set newPackets to make sure we wake
3482 * the reader once all packets have been processed */
3483 #ifdef RX_TRACK_PACKETS
3484 np->flags |= RX_PKTFLAG_RQ;
3486 queue_Prepend(&call->rq, np);
3487 #ifdef RXDEBUG_PACKET
3489 #endif /* RXDEBUG_PACKET */
3491 np = NULL; /* We can't use this anymore */
3494 /* If an ack is requested then set a flag to make sure we
3495 * send an acknowledgement for this packet */
3496 if (flags & RX_REQUEST_ACK) {
3497 ackNeeded = RX_ACK_REQUESTED;
3500 /* Keep track of whether we have received the last packet */
3501 if (flags & RX_LAST_PACKET) {
3502 call->flags |= RX_CALL_HAVE_LAST;
3506 /* Check whether we have all of the packets for this call */
3507 if (call->flags & RX_CALL_HAVE_LAST) {
3508 afs_uint32 tseq; /* temporary sequence number */
3509 struct rx_packet *tp; /* Temporary packet pointer */
3510 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3512 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3513 if (tseq != tp->header.seq)
3515 if (tp->header.flags & RX_LAST_PACKET) {
3516 call->flags |= RX_CALL_RECEIVE_DONE;
3523 /* Provide asynchronous notification for those who want it
3524 * (e.g. multi rx) */
3525 if (call->arrivalProc) {
3526 (*call->arrivalProc) (call, call->arrivalProcHandle,
3527 call->arrivalProcArg);
3528 call->arrivalProc = (void (*)())0;
3531 /* Update last packet received */
3534 /* If there is no server process serving this call, grab
3535 * one, if available. We only need to do this once. If a
3536 * server thread is available, this thread becomes a server
3537 * thread and the server thread becomes a listener thread. */
3539 TryAttach(call, socket, tnop, newcallp, 0);
3542 /* This is not the expected next packet. */
3544 /* Determine whether this is a new or old packet, and if it's
3545 * a new one, whether it fits into the current receive window.
3546 * Also figure out whether the packet was delivered in sequence.
3547 * We use the prev variable to determine whether the new packet
3548 * is the successor of its immediate predecessor in the
3549 * receive queue, and the missing flag to determine whether
3550 * any of this packets predecessors are missing. */
3552 afs_uint32 prev; /* "Previous packet" sequence number */
3553 struct rx_packet *tp; /* Temporary packet pointer */
3554 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3555 int missing; /* Are any predecessors missing? */
3557 /* If the new packet's sequence number has been sent to the
3558 * application already, then this is a duplicate */
3559 if (seq < call->rnext) {
3560 if (rx_stats_active)
3561 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3562 rxevent_Cancel(call->delayedAckEvent, call,
3563 RX_CALL_REFCOUNT_DELAY);
3564 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3570 /* If the sequence number is greater than what can be
3571 * accomodated by the current window, then send a negative
3572 * acknowledge and drop the packet */
3573 if ((call->rnext + call->rwind) <= seq) {
3574 rxevent_Cancel(call->delayedAckEvent, call,
3575 RX_CALL_REFCOUNT_DELAY);
3576 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3583 /* Look for the packet in the queue of old received packets */
3584 for (prev = call->rnext - 1, missing =
3585 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3586 /*Check for duplicate packet */
3587 if (seq == tp->header.seq) {
3588 if (rx_stats_active)
3589 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3590 rxevent_Cancel(call->delayedAckEvent, call,
3591 RX_CALL_REFCOUNT_DELAY);
3592 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3598 /* If we find a higher sequence packet, break out and
3599 * insert the new packet here. */
3600 if (seq < tp->header.seq)
3602 /* Check for missing packet */
3603 if (tp->header.seq != prev + 1) {
3607 prev = tp->header.seq;
3610 /* Keep track of whether we have received the last packet. */
3611 if (flags & RX_LAST_PACKET) {
3612 call->flags |= RX_CALL_HAVE_LAST;
3615 /* It's within the window: add it to the the receive queue.
3616 * tp is left by the previous loop either pointing at the
3617 * packet before which to insert the new packet, or at the
3618 * queue head if the queue is empty or the packet should be
3620 #ifdef RX_TRACK_PACKETS
3621 np->flags |= RX_PKTFLAG_RQ;
3623 #ifdef RXDEBUG_PACKET
3625 #endif /* RXDEBUG_PACKET */
3626 queue_InsertBefore(tp, np);
3630 /* Check whether we have all of the packets for this call */
3631 if ((call->flags & RX_CALL_HAVE_LAST)
3632 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3633 afs_uint32 tseq; /* temporary sequence number */
3636 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3637 if (tseq != tp->header.seq)
3639 if (tp->header.flags & RX_LAST_PACKET) {
3640 call->flags |= RX_CALL_RECEIVE_DONE;
3647 /* We need to send an ack of the packet is out of sequence,
3648 * or if an ack was requested by the peer. */
3649 if (seq != prev + 1 || missing) {
3650 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3651 } else if (flags & RX_REQUEST_ACK) {
3652 ackNeeded = RX_ACK_REQUESTED;
3655 /* Acknowledge the last packet for each call */
3656 if (flags & RX_LAST_PACKET) {
3667 * If the receiver is waiting for an iovec, fill the iovec
3668 * using the data from the receive queue */
3669 if (call->flags & RX_CALL_IOVEC_WAIT) {
3670 didHardAck = rxi_FillReadVec(call, serial);
3671 /* the call may have been aborted */
3680 /* Wakeup the reader if any */
3681 if ((call->flags & RX_CALL_READER_WAIT)
3682 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3683 || (call->iovNext >= call->iovMax)
3684 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3685 call->flags &= ~RX_CALL_READER_WAIT;
3686 #ifdef RX_ENABLE_LOCKS
3687 CV_BROADCAST(&call->cv_rq);
3689 osi_rxWakeup(&call->rq);
3695 * Send an ack when requested by the peer, or once every
3696 * rxi_SoftAckRate packets until the last packet has been
3697 * received. Always send a soft ack for the last packet in
3698 * the server's reply. */
3700 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3701 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3702 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3703 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3704 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3705 } else if (call->nSoftAcks) {
3706 clock_GetTime(&now);
3708 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3709 clock_Add(&when, &rx_lastAckDelay);
3711 clock_Add(&when, &rx_softAckDelay);
3713 if (!call->delayedAckEvent
3714 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3715 rxevent_Cancel(call->delayedAckEvent, call,
3716 RX_CALL_REFCOUNT_DELAY);
3717 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3718 call->delayedAckEvent =
3719 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3721 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3722 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3729 static void rxi_ComputeRate();
3733 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3735 struct rx_peer *peer = conn->peer;
3737 MUTEX_ENTER(&peer->peer_lock);
3738 peer->lastReachTime = clock_Sec();
3739 MUTEX_EXIT(&peer->peer_lock);
3741 MUTEX_ENTER(&conn->conn_data_lock);
3742 if (conn->flags & RX_CONN_ATTACHWAIT) {
3745 conn->flags &= ~RX_CONN_ATTACHWAIT;
3746 MUTEX_EXIT(&conn->conn_data_lock);
3748 for (i = 0; i < RX_MAXCALLS; i++) {
3749 struct rx_call *call = conn->call[i];
3752 MUTEX_ENTER(&call->lock);
3753 /* tnop can be null if newcallp is null */
3754 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3756 MUTEX_EXIT(&call->lock);
3760 MUTEX_EXIT(&conn->conn_data_lock);
3763 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
3765 rx_ack_reason(int reason)
3768 case RX_ACK_REQUESTED:
3770 case RX_ACK_DUPLICATE:
3772 case RX_ACK_OUT_OF_SEQUENCE:
3774 case RX_ACK_EXCEEDS_WINDOW:
3776 case RX_ACK_NOSPACE:
3780 case RX_ACK_PING_RESPONSE:
3793 /* rxi_ComputePeerNetStats
3795 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3796 * estimates (like RTT and throughput) based on ack packets. Caller
3797 * must ensure that the packet in question is the right one (i.e.
3798 * serial number matches).
3801 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3802 struct rx_ackPacket *ap, struct rx_packet *np)
3804 struct rx_peer *peer = call->conn->peer;
3806 /* Use RTT if not delayed by client and
3807 * ignore packets that were retransmitted. */
3808 if (!(p->flags & RX_PKTFLAG_ACKED) &&
3809 ap->reason != RX_ACK_DELAY &&
3810 clock_Eq(&p->timeSent, &p->firstSent))
3811 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3813 rxi_ComputeRate(peer, call, p, np, ap->reason);
3817 /* The real smarts of the whole thing. */
3819 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
3822 struct rx_ackPacket *ap;
3824 struct rx_packet *tp;
3825 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3826 struct rx_connection *conn = call->conn;
3827 struct rx_peer *peer = conn->peer;
3830 /* because there are CM's that are bogus, sending weird values for this. */
3831 afs_uint32 skew = 0;
3836 int newAckCount = 0;
3837 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3838 int pktsize = 0; /* Set if we need to update the peer mtu */
3839 int conn_data_locked = 0;
3841 if (rx_stats_active)
3842 rx_MutexIncrement(rx_stats.ackPacketsRead, rx_stats_mutex);
3843 ap = (struct rx_ackPacket *)rx_DataOf(np);
3844 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3846 return np; /* truncated ack packet */
3848 /* depends on ack packet struct */
3849 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3850 first = ntohl(ap->firstPacket);
3851 serial = ntohl(ap->serial);
3852 /* temporarily disabled -- needs to degrade over time
3853 * skew = ntohs(ap->maxSkew); */
3855 /* Ignore ack packets received out of order */
3856 if (first < call->tfirst) {
3860 if (np->header.flags & RX_SLOW_START_OK) {
3861 call->flags |= RX_CALL_SLOW_START_OK;
3864 if (ap->reason == RX_ACK_PING_RESPONSE)
3865 rxi_UpdatePeerReach(conn, call);
3867 if (conn->lastPacketSizeSeq) {
3868 MUTEX_ENTER(&conn->conn_data_lock);
3869 conn_data_locked = 1;
3870 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
3871 pktsize = conn->lastPacketSize;
3872 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
3875 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
3876 if (!conn_data_locked) {
3877 MUTEX_ENTER(&conn->conn_data_lock);
3878 conn_data_locked = 1;
3880 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
3881 /* process mtu ping ack */
3882 pktsize = conn->lastPingSize;
3883 conn->lastPingSizeSer = conn->lastPingSize = 0;
3887 if (conn_data_locked) {
3888 MUTEX_EXIT(&conn->conn_data_lock);
3889 conn_data_locked = 0;
3893 if (rxdebug_active) {
3897 len = _snprintf(msg, sizeof(msg),
3898 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3899 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3900 ntohl(ap->serial), ntohl(ap->previousPacket),
3901 (unsigned int)np->header.seq, (unsigned int)skew,
3902 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3906 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3907 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3911 OutputDebugString(msg);
3913 #else /* AFS_NT40_ENV */
3916 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3917 ap->reason, ntohl(ap->previousPacket),
3918 (unsigned int)np->header.seq, (unsigned int)serial,
3919 (unsigned int)skew, ntohl(ap->firstPacket));
3922 for (offset = 0; offset < nAcks; offset++)
3923 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3928 #endif /* AFS_NT40_ENV */
3931 MUTEX_ENTER(&peer->peer_lock);
3934 * Start somewhere. Can't assume we can send what we can receive,
3935 * but we are clearly receiving.
3937 if (!peer->maxPacketSize)
3938 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
3940 if (pktsize > peer->maxPacketSize) {
3941 peer->maxPacketSize = pktsize;
3942 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
3943 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
3944 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
3945 rxi_ScheduleGrowMTUEvent(call, 1);
3950 /* Update the outgoing packet skew value to the latest value of
3951 * the peer's incoming packet skew value. The ack packet, of
3952 * course, could arrive out of order, but that won't affect things
3954 peer->outPacketSkew = skew;
3956 /* Check for packets that no longer need to be transmitted, and
3957 * discard them. This only applies to packets positively
3958 * acknowledged as having been sent to the peer's upper level.
3959 * All other packets must be retained. So only packets with
3960 * sequence numbers < ap->firstPacket are candidates. */
3961 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3962 if (tp->header.seq >= first)
3964 call->tfirst = tp->header.seq + 1;
3965 rxi_ComputePeerNetStats(call, tp, ap, np);
3966 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3969 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3970 /* XXX Hack. Because we have to release the global rx lock when sending
3971 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3972 * in rxi_Start sending packets out because packets may move to the
3973 * freePacketQueue as result of being here! So we drop these packets until
3974 * we're safely out of the traversing. Really ugly!
3975 * To make it even uglier, if we're using fine grain locking, we can
3976 * set the ack bits in the packets and have rxi_Start remove the packets
3977 * when it's done transmitting.
3979 if (call->flags & RX_CALL_TQ_BUSY) {
3980 #ifdef RX_ENABLE_LOCKS
3981 tp->flags |= RX_PKTFLAG_ACKED;
3982 call->flags |= RX_CALL_TQ_SOME_ACKED;
3983 #else /* RX_ENABLE_LOCKS */
3985 #endif /* RX_ENABLE_LOCKS */
3987 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3990 #ifdef RX_TRACK_PACKETS
3991 tp->flags &= ~RX_PKTFLAG_TQ;
3993 #ifdef RXDEBUG_PACKET
3995 #endif /* RXDEBUG_PACKET */
3996 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4001 /* Give rate detector a chance to respond to ping requests */
4002 if (ap->reason == RX_ACK_PING_RESPONSE) {
4003 rxi_ComputeRate(peer, call, 0, np, ap->reason);
4007 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4009 /* Now go through explicit acks/nacks and record the results in
4010 * the waiting packets. These are packets that can't be released
4011 * yet, even with a positive acknowledge. This positive
4012 * acknowledge only means the packet has been received by the
4013 * peer, not that it will be retained long enough to be sent to
4014 * the peer's upper level. In addition, reset the transmit timers
4015 * of any missing packets (those packets that must be missing
4016 * because this packet was out of sequence) */
4018 call->nSoftAcked = 0;
4019 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
4020 /* Update round trip time if the ack was stimulated on receipt
4022 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4023 #ifdef RX_ENABLE_LOCKS
4024 if (tp->header.seq >= first)
4025 #endif /* RX_ENABLE_LOCKS */
4026 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4027 rxi_ComputePeerNetStats(call, tp, ap, np);
4029 /* Set the acknowledge flag per packet based on the
4030 * information in the ack packet. An acknowlegded packet can
4031 * be downgraded when the server has discarded a packet it
4032 * soacked previously, or when an ack packet is received
4033 * out of sequence. */
4034 if (tp->header.seq < first) {
4035 /* Implicit ack information */
4036 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4039 tp->flags |= RX_PKTFLAG_ACKED;
4040 } else if (tp->header.seq < first + nAcks) {
4041 /* Explicit ack information: set it in the packet appropriately */
4042 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4043 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4045 tp->flags |= RX_PKTFLAG_ACKED;
4052 } else /* RX_ACK_TYPE_NACK */ {
4053 tp->flags &= ~RX_PKTFLAG_ACKED;
4057 tp->flags &= ~RX_PKTFLAG_ACKED;
4062 * Following the suggestion of Phil Kern, we back off the peer's
4063 * timeout value for future packets until a successful response
4064 * is received for an initial transmission.
4066 if (missing && !peer->backedOff) {
4067 struct clock c = peer->timeout;
4068 struct clock max_to = {3, 0};
4070 clock_Add(&peer->timeout, &c);
4071 if (clock_Gt(&peer->timeout, &max_to))
4072 peer->timeout = max_to;
4073 peer->backedOff = 1;
4076 /* If packet isn't yet acked, and it has been transmitted at least
4077 * once, reset retransmit time using latest timeout
4078 * ie, this should readjust the retransmit timer for all outstanding
4079 * packets... So we don't just retransmit when we should know better*/
4081 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
4082 tp->retryTime = tp->timeSent;
4083 clock_Add(&tp->retryTime, &peer->timeout);
4084 /* shift by eight because one quarter-sec ~ 256 milliseconds */
4085 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
4089 /* If the window has been extended by this acknowledge packet,
4090 * then wakeup a sender waiting in alloc for window space, or try
4091 * sending packets now, if he's been sitting on packets due to
4092 * lack of window space */
4093 if (call->tnext < (call->tfirst + call->twind)) {
4094 #ifdef RX_ENABLE_LOCKS
4095 CV_SIGNAL(&call->cv_twind);
4097 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4098 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4099 osi_rxWakeup(&call->twind);
4102 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4103 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4107 /* if the ack packet has a receivelen field hanging off it,
4108 * update our state */
4109 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4112 /* If the ack packet has a "recommended" size that is less than
4113 * what I am using now, reduce my size to match */
4114 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4115 (int)sizeof(afs_int32), &tSize);
4116 tSize = (afs_uint32) ntohl(tSize);
4117 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4119 /* Get the maximum packet size to send to this peer */
4120 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4122 tSize = (afs_uint32) ntohl(tSize);
4123 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4124 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4126 /* sanity check - peer might have restarted with different params.
4127 * If peer says "send less", dammit, send less... Peer should never
4128 * be unable to accept packets of the size that prior AFS versions would
4129 * send without asking. */
4130 if (peer->maxMTU != tSize) {
4131 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4133 peer->maxMTU = tSize;
4134 peer->MTU = MIN(tSize, peer->MTU);
4135 call->MTU = MIN(call->MTU, tSize);
4138 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4141 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4142 (int)sizeof(afs_int32), &tSize);
4143 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4144 if (tSize < call->twind) { /* smaller than our send */
4145 call->twind = tSize; /* window, we must send less... */
4146 call->ssthresh = MIN(call->twind, call->ssthresh);
4147 call->conn->twind[call->channel] = call->twind;
4150 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4151 * network MTU confused with the loopback MTU. Calculate the
4152 * maximum MTU here for use in the slow start code below.
4154 /* Did peer restart with older RX version? */
4155 if (peer->maxDgramPackets > 1) {
4156 peer->maxDgramPackets = 1;
4158 } else if (np->length >=
4159 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4162 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4163 sizeof(afs_int32), &tSize);
4164 tSize = (afs_uint32) ntohl(tSize);
4166 * As of AFS 3.5 we set the send window to match the receive window.
4168 if (tSize < call->twind) {
4169 call->twind = tSize;
4170 call->conn->twind[call->channel] = call->twind;
4171 call->ssthresh = MIN(call->twind, call->ssthresh);
4172 } else if (tSize > call->twind) {
4173 call->twind = tSize;
4174 call->conn->twind[call->channel] = call->twind;
4178 * As of AFS 3.5, a jumbogram is more than one fixed size
4179 * packet transmitted in a single UDP datagram. If the remote
4180 * MTU is smaller than our local MTU then never send a datagram
4181 * larger than the natural MTU.
4184 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4185 (int)sizeof(afs_int32), &tSize);
4186 maxDgramPackets = (afs_uint32) ntohl(tSize);
4187 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4189 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4190 maxDgramPackets = MIN(maxDgramPackets, tSize);
4191 if (maxDgramPackets > 1) {
4192 peer->maxDgramPackets = maxDgramPackets;
4193 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4195 peer->maxDgramPackets = 1;
4196 call->MTU = peer->natMTU;
4198 } else if (peer->maxDgramPackets > 1) {
4199 /* Restarted with lower version of RX */
4200 peer->maxDgramPackets = 1;
4202 } else if (peer->maxDgramPackets > 1
4203 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4204 /* Restarted with lower version of RX */
4205 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4206 peer->natMTU = OLD_MAX_PACKET_SIZE;
4207 peer->MTU = OLD_MAX_PACKET_SIZE;
4208 peer->maxDgramPackets = 1;
4209 peer->nDgramPackets = 1;
4211 call->MTU = OLD_MAX_PACKET_SIZE;
4216 * Calculate how many datagrams were successfully received after
4217 * the first missing packet and adjust the negative ack counter
4222 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4223 if (call->nNacks < nNacked) {
4224 call->nNacks = nNacked;
4227 call->nAcks += newAckCount;
4231 if (call->flags & RX_CALL_FAST_RECOVER) {
4233 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4235 call->flags &= ~RX_CALL_FAST_RECOVER;
4236 call->cwind = call->nextCwind;
4237 call->nextCwind = 0;
4240 call->nCwindAcks = 0;
4241 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4242 /* Three negative acks in a row trigger congestion recovery */
4243 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4244 MUTEX_EXIT(&peer->peer_lock);
4245 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4246 /* someone else is waiting to start recovery */
4249 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4250 rxi_WaitforTQBusy(call);
4251 MUTEX_ENTER(&peer->peer_lock);
4252 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4253 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4254 call->flags |= RX_CALL_FAST_RECOVER;
4255 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4257 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4258 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4259 call->nextCwind = call->ssthresh;
4262 peer->MTU = call->MTU;
4263 peer->cwind = call->nextCwind;
4264 peer->nDgramPackets = call->nDgramPackets;
4266 call->congestSeq = peer->congestSeq;
4267 /* Reset the resend times on the packets that were nacked
4268 * so we will retransmit as soon as the window permits*/
4269 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4271 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4272 clock_Zero(&tp->retryTime);
4274 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4279 /* If cwind is smaller than ssthresh, then increase
4280 * the window one packet for each ack we receive (exponential
4282 * If cwind is greater than or equal to ssthresh then increase
4283 * the congestion window by one packet for each cwind acks we
4284 * receive (linear growth). */
4285 if (call->cwind < call->ssthresh) {
4287 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4288 call->nCwindAcks = 0;
4290 call->nCwindAcks += newAckCount;
4291 if (call->nCwindAcks >= call->cwind) {
4292 call->nCwindAcks = 0;
4293 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4297 * If we have received several acknowledgements in a row then
4298 * it is time to increase the size of our datagrams
4300 if ((int)call->nAcks > rx_nDgramThreshold) {
4301 if (peer->maxDgramPackets > 1) {
4302 if (call->nDgramPackets < peer->maxDgramPackets) {
4303 call->nDgramPackets++;
4305 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4306 } else if (call->MTU < peer->maxMTU) {
4307 /* don't upgrade if we can't handle it */
4308 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4309 call->MTU = peer->ifMTU;
4311 call->MTU += peer->natMTU;
4312 call->MTU = MIN(call->MTU, peer->maxMTU);
4319 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4321 /* Servers need to hold the call until all response packets have
4322 * been acknowledged. Soft acks are good enough since clients
4323 * are not allowed to clear their receive queues. */
4324 if (call->state == RX_STATE_HOLD
4325 && call->tfirst + call->nSoftAcked >= call->tnext) {
4326 call->state = RX_STATE_DALLY;
4327 rxi_ClearTransmitQueue(call, 0);
4328 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4329 } else if (!queue_IsEmpty(&call->tq)) {
4330 rxi_Start(0, call, 0, istack);
4335 /* Received a response to a challenge packet */
4337 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4338 struct rx_packet *np, int istack)
4342 /* Ignore the packet if we're the client */
4343 if (conn->type == RX_CLIENT_CONNECTION)
4346 /* If already authenticated, ignore the packet (it's probably a retry) */
4347 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4350 /* Otherwise, have the security object evaluate the response packet */
4351 error = RXS_CheckResponse(conn->securityObject, conn, np);
4353 /* If the response is invalid, reset the connection, sending
4354 * an abort to the peer */
4358 rxi_ConnectionError(conn, error);
4359 MUTEX_ENTER(&conn->conn_data_lock);
4360 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4361 MUTEX_EXIT(&conn->conn_data_lock);
4364 /* If the response is valid, any calls waiting to attach
4365 * servers can now do so */
4368 for (i = 0; i < RX_MAXCALLS; i++) {
4369 struct rx_call *call = conn->call[i];
4371 MUTEX_ENTER(&call->lock);
4372 if (call->state == RX_STATE_PRECALL)
4373 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4374 /* tnop can be null if newcallp is null */
4375 MUTEX_EXIT(&call->lock);
4379 /* Update the peer reachability information, just in case
4380 * some calls went into attach-wait while we were waiting
4381 * for authentication..
4383 rxi_UpdatePeerReach(conn, NULL);
4388 /* A client has received an authentication challenge: the security
4389 * object is asked to cough up a respectable response packet to send
4390 * back to the server. The server is responsible for retrying the
4391 * challenge if it fails to get a response. */
4394 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4395 struct rx_packet *np, int istack)
4399 /* Ignore the challenge if we're the server */
4400 if (conn->type == RX_SERVER_CONNECTION)
4403 /* Ignore the challenge if the connection is otherwise idle; someone's
4404 * trying to use us as an oracle. */
4405 if (!rxi_HasActiveCalls(conn))
4408 /* Send the security object the challenge packet. It is expected to fill
4409 * in the response. */
4410 error = RXS_GetResponse(conn->securityObject, conn, np);
4412 /* If the security object is unable to return a valid response, reset the
4413 * connection and send an abort to the peer. Otherwise send the response
4414 * packet to the peer connection. */
4416 rxi_ConnectionError(conn, error);
4417 MUTEX_ENTER(&conn->conn_data_lock);
4418 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4419 MUTEX_EXIT(&conn->conn_data_lock);
4421 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4422 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4428 /* Find an available server process to service the current request in
4429 * the given call structure. If one isn't available, queue up this
4430 * call so it eventually gets one */
4432 rxi_AttachServerProc(struct rx_call *call,
4433 osi_socket socket, int *tnop,
4434 struct rx_call **newcallp)
4436 struct rx_serverQueueEntry *sq;
4437 struct rx_service *service = call->conn->service;
4440 /* May already be attached */
4441 if (call->state == RX_STATE_ACTIVE)
4444 MUTEX_ENTER(&rx_serverPool_lock);
4446 haveQuota = QuotaOK(service);
4447 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4448 /* If there are no processes available to service this call,
4449 * put the call on the incoming call queue (unless it's
4450 * already on the queue).
4452 #ifdef RX_ENABLE_LOCKS
4454 ReturnToServerPool(service);
4455 #endif /* RX_ENABLE_LOCKS */
4457 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4458 call->flags |= RX_CALL_WAIT_PROC;
4459 MUTEX_ENTER(&rx_waiting_mutex);
4462 MUTEX_EXIT(&rx_waiting_mutex);
4463 rxi_calltrace(RX_CALL_ARRIVAL, call);
4464 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4465 queue_Append(&rx_incomingCallQueue, call);
4468 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4470 /* If hot threads are enabled, and both newcallp and sq->socketp
4471 * are non-null, then this thread will process the call, and the
4472 * idle server thread will start listening on this threads socket.
4475 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4478 *sq->socketp = socket;
4479 clock_GetTime(&call->startTime);
4480 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4484 if (call->flags & RX_CALL_WAIT_PROC) {
4485 /* Conservative: I don't think this should happen */
4486 call->flags &= ~RX_CALL_WAIT_PROC;
4487 if (queue_IsOnQueue(call)) {
4490 MUTEX_ENTER(&rx_waiting_mutex);
4492 MUTEX_EXIT(&rx_waiting_mutex);
4495 call->state = RX_STATE_ACTIVE;
4496 call->mode = RX_MODE_RECEIVING;
4497 #ifdef RX_KERNEL_TRACE
4499 int glockOwner = ISAFS_GLOCK();
4502 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4503 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4509 if (call->flags & RX_CALL_CLEARED) {
4510 /* send an ack now to start the packet flow up again */
4511 call->flags &= ~RX_CALL_CLEARED;
4512 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4514 #ifdef RX_ENABLE_LOCKS
4517 service->nRequestsRunning++;
4518 MUTEX_ENTER(&rx_quota_mutex);
4519 if (service->nRequestsRunning <= service->minProcs)
4522 MUTEX_EXIT(&rx_quota_mutex);
4526 MUTEX_EXIT(&rx_serverPool_lock);
4529 /* Delay the sending of an acknowledge event for a short while, while
4530 * a new call is being prepared (in the case of a client) or a reply
4531 * is being prepared (in the case of a server). Rather than sending
4532 * an ack packet, an ACKALL packet is sent. */
4534 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4536 #ifdef RX_ENABLE_LOCKS
4538 MUTEX_ENTER(&call->lock);
4539 call->delayedAckEvent = NULL;
4540 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4542 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4543 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4545 MUTEX_EXIT(&call->lock);
4546 #else /* RX_ENABLE_LOCKS */
4548 call->delayedAckEvent = NULL;
4549 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4550 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4551 #endif /* RX_ENABLE_LOCKS */
4555 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4557 struct rx_call *call = arg1;
4558 #ifdef RX_ENABLE_LOCKS
4560 MUTEX_ENTER(&call->lock);
4561 if (event == call->delayedAckEvent)
4562 call->delayedAckEvent = NULL;
4563 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4565 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4567 MUTEX_EXIT(&call->lock);
4568 #else /* RX_ENABLE_LOCKS */
4570 call->delayedAckEvent = NULL;
4571 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4572 #endif /* RX_ENABLE_LOCKS */
4576 #ifdef RX_ENABLE_LOCKS
4577 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4578 * clearing them out.
4581 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4583 struct rx_packet *p, *tp;
4586 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4587 p->flags |= RX_PKTFLAG_ACKED;
4591 call->flags |= RX_CALL_TQ_CLEARME;
4592 call->flags |= RX_CALL_TQ_SOME_ACKED;
4595 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4596 call->tfirst = call->tnext;
4597 call->nSoftAcked = 0;
4599 if (call->flags & RX_CALL_FAST_RECOVER) {
4600 call->flags &= ~RX_CALL_FAST_RECOVER;
4601 call->cwind = call->nextCwind;
4602 call->nextCwind = 0;
4605 CV_SIGNAL(&call->cv_twind);
4607 #endif /* RX_ENABLE_LOCKS */
4609 /* Clear out the transmit queue for the current call (all packets have
4610 * been received by peer) */
4612 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4614 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4615 struct rx_packet *p, *tp;
4617 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4619 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4620 p->flags |= RX_PKTFLAG_ACKED;
4624 call->flags |= RX_CALL_TQ_CLEARME;
4625 call->flags |= RX_CALL_TQ_SOME_ACKED;
4628 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4629 #ifdef RXDEBUG_PACKET
4631 #endif /* RXDEBUG_PACKET */
4632 rxi_FreePackets(0, &call->tq);
4633 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
4634 #ifdef RX_ENABLE_LOCKS
4635 CV_BROADCAST(&call->cv_tq);
4636 #else /* RX_ENABLE_LOCKS */
4637 osi_rxWakeup(&call->tq);
4638 #endif /* RX_ENABLE_LOCKS */
4640 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4641 call->flags &= ~RX_CALL_TQ_CLEARME;
4643 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4645 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4646 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4647 call->nSoftAcked = 0;
4649 if (call->flags & RX_CALL_FAST_RECOVER) {
4650 call->flags &= ~RX_CALL_FAST_RECOVER;
4651 call->cwind = call->nextCwind;
4653 #ifdef RX_ENABLE_LOCKS
4654 CV_SIGNAL(&call->cv_twind);
4656 osi_rxWakeup(&call->twind);
4661 rxi_ClearReceiveQueue(struct rx_call *call)
4663 if (queue_IsNotEmpty(&call->rq)) {
4666 count = rxi_FreePackets(0, &call->rq);
4667 rx_packetReclaims += count;
4668 #ifdef RXDEBUG_PACKET
4670 if ( call->rqc != 0 )
4671 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0", call, call->rqc));
4673 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4675 if (call->state == RX_STATE_PRECALL) {
4676 call->flags |= RX_CALL_CLEARED;
4680 /* Send an abort packet for the specified call */
4682 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
4683 int istack, int force)
4686 struct clock when, now;
4691 /* Clients should never delay abort messages */
4692 if (rx_IsClientConn(call->conn))
4695 if (call->abortCode != call->error) {
4696 call->abortCode = call->error;
4697 call->abortCount = 0;
4700 if (force || rxi_callAbortThreshhold == 0
4701 || call->abortCount < rxi_callAbortThreshhold) {
4702 if (call->delayedAbortEvent) {
4703 rxevent_Cancel(call->delayedAbortEvent, call,
4704 RX_CALL_REFCOUNT_ABORT);
4706 error = htonl(call->error);
4709 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4710 (char *)&error, sizeof(error), istack);
4711 } else if (!call->delayedAbortEvent) {
4712 clock_GetTime(&now);
4714 clock_Addmsec(&when, rxi_callAbortDelay);
4715 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4716 call->delayedAbortEvent =
4717 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4722 /* Send an abort packet for the specified connection. Packet is an
4723 * optional pointer to a packet that can be used to send the abort.
4724 * Once the number of abort messages reaches the threshhold, an
4725 * event is scheduled to send the abort. Setting the force flag
4726 * overrides sending delayed abort messages.
4728 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4729 * to send the abort packet.
4732 rxi_SendConnectionAbort(struct rx_connection *conn,
4733 struct rx_packet *packet, int istack, int force)
4736 struct clock when, now;
4741 /* Clients should never delay abort messages */
4742 if (rx_IsClientConn(conn))
4745 if (force || rxi_connAbortThreshhold == 0
4746 || conn->abortCount < rxi_connAbortThreshhold) {
4747 if (conn->delayedAbortEvent) {
4748 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4750 error = htonl(conn->error);
4752 MUTEX_EXIT(&conn->conn_data_lock);
4754 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4755 RX_PACKET_TYPE_ABORT, (char *)&error,
4756 sizeof(error), istack);
4757 MUTEX_ENTER(&conn->conn_data_lock);
4758 } else if (!conn->delayedAbortEvent) {
4759 clock_GetTime(&now);
4761 clock_Addmsec(&when, rxi_connAbortDelay);
4762 conn->delayedAbortEvent =
4763 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4768 /* Associate an error all of the calls owned by a connection. Called
4769 * with error non-zero. This is only for really fatal things, like
4770 * bad authentication responses. The connection itself is set in
4771 * error at this point, so that future packets received will be
4774 rxi_ConnectionError(struct rx_connection *conn,
4780 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d", conn, error));
4782 MUTEX_ENTER(&conn->conn_data_lock);
4783 if (conn->challengeEvent)
4784 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4785 if (conn->natKeepAliveEvent)
4786 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
4787 if (conn->checkReachEvent) {
4788 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4789 conn->checkReachEvent = 0;
4790 conn->flags &= ~RX_CONN_ATTACHWAIT;
4793 MUTEX_EXIT(&conn->conn_data_lock);
4794 for (i = 0; i < RX_MAXCALLS; i++) {
4795 struct rx_call *call = conn->call[i];
4797 MUTEX_ENTER(&call->lock);
4798 rxi_CallError(call, error);
4799 MUTEX_EXIT(&call->lock);
4802 conn->error = error;
4803 if (rx_stats_active)
4804 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4809 rxi_CallError(struct rx_call *call, afs_int32 error)
4812 osirx_AssertMine(&call->lock, "rxi_CallError");
4814 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d", call, error, call->error));
4816 error = call->error;
4818 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4819 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4820 rxi_ResetCall(call, 0);
4823 rxi_ResetCall(call, 0);
4825 call->error = error;
4826 call->mode = RX_MODE_ERROR;
4829 /* Reset various fields in a call structure, and wakeup waiting
4830 * processes. Some fields aren't changed: state & mode are not
4831 * touched (these must be set by the caller), and bufptr, nLeft, and
4832 * nFree are not reset, since these fields are manipulated by
4833 * unprotected macros, and may only be reset by non-interrupting code.
4836 /* this code requires that call->conn be set properly as a pre-condition. */
4837 #endif /* ADAPT_WINDOW */
4840 rxi_ResetCall(struct rx_call *call, int newcall)
4843 struct rx_peer *peer;
4844 struct rx_packet *packet;
4846 osirx_AssertMine(&call->lock, "rxi_ResetCall");
4848 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
4850 /* Notify anyone who is waiting for asynchronous packet arrival */
4851 if (call->arrivalProc) {
4852 (*call->arrivalProc) (call, call->arrivalProcHandle,
4853 call->arrivalProcArg);
4854 call->arrivalProc = (void (*)())0;
4857 if (call->delayedAbortEvent) {
4858 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4859 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4861 rxi_SendCallAbort(call, packet, 0, 1);
4862 rxi_FreePacket(packet);
4867 * Update the peer with the congestion information in this call
4868 * so other calls on this connection can pick up where this call
4869 * left off. If the congestion sequence numbers don't match then
4870 * another call experienced a retransmission.
4872 peer = call->conn->peer;
4873 MUTEX_ENTER(&peer->peer_lock);
4875 if (call->congestSeq == peer->congestSeq) {
4876 peer->cwind = MAX(peer->cwind, call->cwind);
4877 peer->MTU = MAX(peer->MTU, call->MTU);
4878 peer->nDgramPackets =
4879 MAX(peer->nDgramPackets, call->nDgramPackets);
4882 call->abortCode = 0;
4883 call->abortCount = 0;
4885 if (peer->maxDgramPackets > 1) {
4886 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4888 call->MTU = peer->MTU;
4890 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4891 call->ssthresh = rx_maxSendWindow;
4892 call->nDgramPackets = peer->nDgramPackets;
4893 call->congestSeq = peer->congestSeq;
4894 MUTEX_EXIT(&peer->peer_lock);
4896 flags = call->flags;
4897 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4898 rxi_WaitforTQBusy(call);
4899 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4901 rxi_ClearTransmitQueue(call, 1);
4902 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4903 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4907 rxi_ClearReceiveQueue(call);
4908 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
4912 call->twind = call->conn->twind[call->channel];
4913 call->rwind = call->conn->rwind[call->channel];
4914 call->nSoftAcked = 0;
4915 call->nextCwind = 0;
4918 call->nCwindAcks = 0;
4919 call->nSoftAcks = 0;
4920 call->nHardAcks = 0;
4922 call->tfirst = call->rnext = call->tnext = 1;
4924 call->lastAcked = 0;
4925 call->localStatus = call->remoteStatus = 0;
4927 if (flags & RX_CALL_READER_WAIT) {
4928 #ifdef RX_ENABLE_LOCKS
4929 CV_BROADCAST(&call->cv_rq);
4931 osi_rxWakeup(&call->rq);
4934 if (flags & RX_CALL_WAIT_PACKETS) {
4935 MUTEX_ENTER(&rx_freePktQ_lock);
4936 rxi_PacketsUnWait(); /* XXX */
4937 MUTEX_EXIT(&rx_freePktQ_lock);
4939 #ifdef RX_ENABLE_LOCKS
4940 CV_SIGNAL(&call->cv_twind);
4942 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4943 osi_rxWakeup(&call->twind);
4946 #ifdef RX_ENABLE_LOCKS
4947 /* The following ensures that we don't mess with any queue while some
4948 * other thread might also be doing so. The call_queue_lock field is
4949 * is only modified under the call lock. If the call is in the process
4950 * of being removed from a queue, the call is not locked until the
4951 * the queue lock is dropped and only then is the call_queue_lock field
4952 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4953 * Note that any other routine which removes a call from a queue has to
4954 * obtain the queue lock before examing the queue and removing the call.
4956 if (call->call_queue_lock) {
4957 MUTEX_ENTER(call->call_queue_lock);
4958 if (queue_IsOnQueue(call)) {
4960 if (flags & RX_CALL_WAIT_PROC) {
4962 MUTEX_ENTER(&rx_waiting_mutex);
4964 MUTEX_EXIT(&rx_waiting_mutex);
4967 MUTEX_EXIT(call->call_queue_lock);
4968 CLEAR_CALL_QUEUE_LOCK(call);
4970 #else /* RX_ENABLE_LOCKS */
4971 if (queue_IsOnQueue(call)) {
4973 if (flags & RX_CALL_WAIT_PROC)
4976 #endif /* RX_ENABLE_LOCKS */
4978 rxi_KeepAliveOff(call);
4979 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4982 /* Send an acknowledge for the indicated packet (seq,serial) of the
4983 * indicated call, for the indicated reason (reason). This
4984 * acknowledge will specifically acknowledge receiving the packet, and
4985 * will also specify which other packets for this call have been
4986 * received. This routine returns the packet that was used to the
4987 * caller. The caller is responsible for freeing it or re-using it.
4988 * This acknowledgement also returns the highest sequence number
4989 * actually read out by the higher level to the sender; the sender
4990 * promises to keep around packets that have not been read by the
4991 * higher level yet (unless, of course, the sender decides to abort
4992 * the call altogether). Any of p, seq, serial, pflags, or reason may
4993 * be set to zero without ill effect. That is, if they are zero, they
4994 * will not convey any information.
4995 * NOW there is a trailer field, after the ack where it will safely be
4996 * ignored by mundanes, which indicates the maximum size packet this
4997 * host can swallow. */
4999 struct rx_packet *optionalPacket; use to send ack (or null)
5000 int seq; Sequence number of the packet we are acking
5001 int serial; Serial number of the packet
5002 int pflags; Flags field from packet header
5003 int reason; Reason an acknowledge was prompted
5007 rxi_SendAck(struct rx_call *call,
5008 struct rx_packet *optionalPacket, int serial, int reason,
5011 struct rx_ackPacket *ap;
5012 struct rx_packet *rqp;
5013 struct rx_packet *nxp; /* For queue_Scan */
5014 struct rx_packet *p;
5017 afs_uint32 padbytes = 0;
5018 #ifdef RX_ENABLE_TSFPQ
5019 struct rx_ts_info_t * rx_ts_info;
5023 * Open the receive window once a thread starts reading packets
5025 if (call->rnext > 1) {
5026 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5029 /* Don't attempt to grow MTU if this is a critical ping */
5030 if (reason == RX_ACK_MTU) {
5031 /* keep track of per-call attempts, if we're over max, do in small
5032 * otherwise in larger? set a size to increment by, decrease
5035 if (call->conn->peer->maxPacketSize &&
5036 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5038 padbytes = call->conn->peer->maxPacketSize+16;
5040 padbytes = call->conn->peer->maxMTU + 128;
5042 /* do always try a minimum size ping */
5043 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5045 /* subtract the ack payload */
5046 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5047 reason = RX_ACK_PING;
5050 call->nHardAcks = 0;
5051 call->nSoftAcks = 0;
5052 if (call->rnext > call->lastAcked)
5053 call->lastAcked = call->rnext;
5057 rx_computelen(p, p->length); /* reset length, you never know */
5058 } /* where that's been... */
5059 #ifdef RX_ENABLE_TSFPQ
5061 RX_TS_INFO_GET(rx_ts_info);
5062 if ((p = rx_ts_info->local_special_packet)) {
5063 rx_computelen(p, p->length);
5064 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5065 rx_ts_info->local_special_packet = p;
5066 } else { /* We won't send the ack, but don't panic. */
5067 return optionalPacket;
5071 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5072 /* We won't send the ack, but don't panic. */
5073 return optionalPacket;
5078 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5081 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5082 #ifndef RX_ENABLE_TSFPQ
5083 if (!optionalPacket)
5086 return optionalPacket;
5088 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5089 if (rx_Contiguous(p) < templ) {
5090 #ifndef RX_ENABLE_TSFPQ
5091 if (!optionalPacket)
5094 return optionalPacket;
5099 /* MTUXXX failing to send an ack is very serious. We should */
5100 /* try as hard as possible to send even a partial ack; it's */
5101 /* better than nothing. */
5102 ap = (struct rx_ackPacket *)rx_DataOf(p);
5103 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5104 ap->reason = reason;
5106 /* The skew computation used to be bogus, I think it's better now. */
5107 /* We should start paying attention to skew. XXX */
5108 ap->serial = htonl(serial);
5109 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5111 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
5112 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5114 /* No fear of running out of ack packet here because there can only be at most
5115 * one window full of unacknowledged packets. The window size must be constrained
5116 * to be less than the maximum ack size, of course. Also, an ack should always
5117 * fit into a single packet -- it should not ever be fragmented. */
5118 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5119 if (!rqp || !call->rq.next
5120 || (rqp->header.seq > (call->rnext + call->rwind))) {
5121 #ifndef RX_ENABLE_TSFPQ
5122 if (!optionalPacket)
5125 rxi_CallError(call, RX_CALL_DEAD);
5126 return optionalPacket;
5129 while (rqp->header.seq > call->rnext + offset)
5130 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5131 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5133 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5134 #ifndef RX_ENABLE_TSFPQ
5135 if (!optionalPacket)
5138 rxi_CallError(call, RX_CALL_DEAD);
5139 return optionalPacket;
5144 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5146 /* these are new for AFS 3.3 */
5147 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5148 templ = htonl(templ);
5149 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5150 templ = htonl(call->conn->peer->ifMTU);
5151 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5152 sizeof(afs_int32), &templ);
5154 /* new for AFS 3.4 */
5155 templ = htonl(call->rwind);
5156 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5157 sizeof(afs_int32), &templ);
5159 /* new for AFS 3.5 */
5160 templ = htonl(call->conn->peer->ifDgramPackets);
5161 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5162 sizeof(afs_int32), &templ);
5164 p->header.serviceId = call->conn->serviceId;
5165 p->header.cid = (call->conn->cid | call->channel);
5166 p->header.callNumber = *call->callNumber;
5168 p->header.securityIndex = call->conn->securityIndex;
5169 p->header.epoch = call->conn->epoch;
5170 p->header.type = RX_PACKET_TYPE_ACK;
5171 p->header.flags = RX_SLOW_START_OK;
5172 if (reason == RX_ACK_PING) {
5173 p->header.flags |= RX_REQUEST_ACK;
5175 clock_GetTime(&call->pingRequestTime);
5178 p->length = padbytes +
5179 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5182 /* not fast but we can potentially use this if truncated
5183 * fragments are delivered to figure out the mtu.
5185 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5186 sizeof(afs_int32), sizeof(afs_int32),
5190 if (call->conn->type == RX_CLIENT_CONNECTION)
5191 p->header.flags |= RX_CLIENT_INITIATED;
5195 if (rxdebug_active) {
5199 len = _snprintf(msg, sizeof(msg),
5200 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5201 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5202 ntohl(ap->serial), ntohl(ap->previousPacket),
5203 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5204 ap->nAcks, ntohs(ap->bufferSpace) );
5208 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5209 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5213 OutputDebugString(msg);
5215 #else /* AFS_NT40_ENV */
5217 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5218 ap->reason, ntohl(ap->previousPacket),
5219 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5221 for (offset = 0; offset < ap->nAcks; offset++)
5222 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5227 #endif /* AFS_NT40_ENV */
5230 int i, nbytes = p->length;
5232 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5233 if (nbytes <= p->wirevec[i].iov_len) {
5236 savelen = p->wirevec[i].iov_len;
5238 p->wirevec[i].iov_len = nbytes;
5240 rxi_Send(call, p, istack);
5241 p->wirevec[i].iov_len = savelen;
5245 nbytes -= p->wirevec[i].iov_len;
5248 if (rx_stats_active)
5249 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
5250 #ifndef RX_ENABLE_TSFPQ
5251 if (!optionalPacket)
5254 return optionalPacket; /* Return packet for re-use by caller */
5257 /* Send all of the packets in the list in single datagram */
5259 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
5260 int istack, int moreFlag, struct clock *now,
5261 struct clock *retryTime, int resending)
5266 struct rx_connection *conn = call->conn;
5267 struct rx_peer *peer = conn->peer;
5269 MUTEX_ENTER(&peer->peer_lock);
5272 peer->reSends += len;
5273 MUTEX_EXIT(&peer->peer_lock);
5275 if (rx_stats_active) {
5277 rx_MutexAdd(rx_stats.dataPacketsReSent, len, rx_stats_mutex);
5279 rx_MutexAdd(rx_stats.dataPacketsSent, len, rx_stats_mutex);
5282 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
5286 /* Set the packet flags and schedule the resend events */
5287 /* Only request an ack for the last packet in the list */
5288 for (i = 0; i < len; i++) {
5289 list[i]->retryTime = *retryTime;
5290 if (list[i]->header.serial) {
5291 /* Exponentially backoff retry times */
5292 if (list[i]->backoff < MAXBACKOFF) {
5293 /* so it can't stay == 0 */
5294 list[i]->backoff = (list[i]->backoff << 1) + 1;
5297 clock_Addmsec(&(list[i]->retryTime),
5298 ((afs_uint32) list[i]->backoff) << 8);
5301 /* Wait a little extra for the ack on the last packet */
5302 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
5303 clock_Addmsec(&(list[i]->retryTime), 400);
5306 /* Record the time sent */
5307 list[i]->timeSent = *now;
5309 /* Ask for an ack on retransmitted packets, on every other packet
5310 * if the peer doesn't support slow start. Ask for an ack on every
5311 * packet until the congestion window reaches the ack rate. */
5312 if (list[i]->header.serial) {
5315 /* improved RTO calculation- not Karn */
5316 list[i]->firstSent = *now;
5317 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5318 || (!(call->flags & RX_CALL_SLOW_START_OK)
5319 && (list[i]->header.seq & 1)))) {
5324 /* Tag this packet as not being the last in this group,
5325 * for the receiver's benefit */
5326 if (i < len - 1 || moreFlag) {
5327 list[i]->header.flags |= RX_MORE_PACKETS;
5330 /* Install the new retransmit time for the packet, and
5331 * record the time sent */
5332 list[i]->timeSent = *now;
5336 list[len - 1]->header.flags |= RX_REQUEST_ACK;
5339 /* Since we're about to send a data packet to the peer, it's
5340 * safe to nuke any scheduled end-of-packets ack */
5341 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5343 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5344 MUTEX_EXIT(&call->lock);
5346 rxi_SendPacketList(call, conn, list, len, istack);
5348 rxi_SendPacket(call, conn, list[0], istack);
5350 MUTEX_ENTER(&call->lock);
5351 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5353 /* Update last send time for this call (for keep-alive
5354 * processing), and for the connection (so that we can discover
5355 * idle connections) */
5356 conn->lastSendTime = call->lastSendTime = clock_Sec();
5357 /* Let a set of retransmits trigger an idle timeout */
5359 call->lastSendData = call->lastSendTime;
5362 /* When sending packets we need to follow these rules:
5363 * 1. Never send more than maxDgramPackets in a jumbogram.
5364 * 2. Never send a packet with more than two iovecs in a jumbogram.
5365 * 3. Never send a retransmitted packet in a jumbogram.
5366 * 4. Never send more than cwind/4 packets in a jumbogram
5367 * We always keep the last list we should have sent so we
5368 * can set the RX_MORE_PACKETS flags correctly.
5371 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5372 int istack, struct clock *now, struct clock *retryTime,
5375 int i, cnt, lastCnt = 0;
5376 struct rx_packet **listP, **lastP = 0;
5377 struct rx_peer *peer = call->conn->peer;
5378 int morePackets = 0;
5380 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
5381 /* Does the current packet force us to flush the current list? */
5383 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5384 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5386 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
5388 /* If the call enters an error state stop sending, or if
5389 * we entered congestion recovery mode, stop sending */
5390 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5398 /* Add the current packet to the list if it hasn't been acked.
5399 * Otherwise adjust the list pointer to skip the current packet. */
5400 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5402 /* Do we need to flush the list? */
5403 if (cnt >= (int)peer->maxDgramPackets
5404 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5405 || list[i]->header.serial
5406 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5408 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5409 retryTime, resending);
5410 /* If the call enters an error state stop sending, or if
5411 * we entered congestion recovery mode, stop sending */
5413 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5418 listP = &list[i + 1];
5423 osi_Panic("rxi_SendList error");
5425 listP = &list[i + 1];
5429 /* Send the whole list when the call is in receive mode, when
5430 * the call is in eof mode, when we are in fast recovery mode,
5431 * and when we have the last packet */
5432 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5433 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5434 || (call->flags & RX_CALL_FAST_RECOVER)) {
5435 /* Check for the case where the current list contains
5436 * an acked packet. Since we always send retransmissions
5437 * in a separate packet, we only need to check the first
5438 * packet in the list */
5439 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5443 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5444 retryTime, resending);
5445 /* If the call enters an error state stop sending, or if
5446 * we entered congestion recovery mode, stop sending */
5447 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5451 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5454 } else if (lastCnt > 0) {
5455 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5460 #ifdef RX_ENABLE_LOCKS
5461 /* Call rxi_Start, below, but with the call lock held. */
5463 rxi_StartUnlocked(struct rxevent *event,
5464 void *arg0, void *arg1, int istack)
5466 struct rx_call *call = arg0;
5468 MUTEX_ENTER(&call->lock);
5469 rxi_Start(event, call, arg1, istack);
5470 MUTEX_EXIT(&call->lock);
5472 #endif /* RX_ENABLE_LOCKS */
5474 /* This routine is called when new packets are readied for
5475 * transmission and when retransmission may be necessary, or when the
5476 * transmission window or burst count are favourable. This should be
5477 * better optimized for new packets, the usual case, now that we've
5478 * got rid of queues of send packets. XXXXXXXXXXX */
5480 rxi_Start(struct rxevent *event,
5481 void *arg0, void *arg1, int istack)
5483 struct rx_call *call = arg0;
5485 struct rx_packet *p;
5486 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5487 struct rx_peer *peer = call->conn->peer;
5488 struct clock now, usenow, retryTime;
5492 struct rx_packet **xmitList;
5495 /* If rxi_Start is being called as a result of a resend event,
5496 * then make sure that the event pointer is removed from the call
5497 * structure, since there is no longer a per-call retransmission
5499 if (event && event == call->resendEvent) {
5500 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5501 call->resendEvent = NULL;
5503 if (queue_IsEmpty(&call->tq)) {
5507 /* Timeouts trigger congestion recovery */
5508 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5509 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5510 /* someone else is waiting to start recovery */
5513 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5514 rxi_WaitforTQBusy(call);
5515 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5516 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5517 call->flags |= RX_CALL_FAST_RECOVER;
5518 if (peer->maxDgramPackets > 1) {
5519 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5521 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5523 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5524 call->nDgramPackets = 1;
5526 call->nextCwind = 1;
5529 MUTEX_ENTER(&peer->peer_lock);
5530 peer->MTU = call->MTU;
5531 peer->cwind = call->cwind;
5532 peer->nDgramPackets = 1;
5534 call->congestSeq = peer->congestSeq;
5535 MUTEX_EXIT(&peer->peer_lock);
5536 /* Clear retry times on packets. Otherwise, it's possible for
5537 * some packets in the queue to force resends at rates faster
5538 * than recovery rates.
5540 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5541 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5542 clock_Zero(&p->retryTime);
5547 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5548 if (rx_stats_active)
5549 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5554 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5555 /* Get clock to compute the re-transmit time for any packets
5556 * in this burst. Note, if we back off, it's reasonable to
5557 * back off all of the packets in the same manner, even if
5558 * some of them have been retransmitted more times than more
5560 * Do a dance to avoid blocking after setting now. */
5561 MUTEX_ENTER(&peer->peer_lock);
5562 retryTime = peer->timeout;
5563 MUTEX_EXIT(&peer->peer_lock);
5565 clock_GetTime(&now);
5566 clock_Add(&retryTime, &now);
5568 /* Send (or resend) any packets that need it, subject to
5569 * window restrictions and congestion burst control
5570 * restrictions. Ask for an ack on the last packet sent in
5571 * this burst. For now, we're relying upon the window being
5572 * considerably bigger than the largest number of packets that
5573 * are typically sent at once by one initial call to
5574 * rxi_Start. This is probably bogus (perhaps we should ask
5575 * for an ack when we're half way through the current
5576 * window?). Also, for non file transfer applications, this
5577 * may end up asking for an ack for every packet. Bogus. XXXX
5580 * But check whether we're here recursively, and let the other guy
5583 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5584 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5585 call->flags |= RX_CALL_TQ_BUSY;
5587 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5589 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5590 call->flags &= ~RX_CALL_NEED_START;
5591 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5593 maxXmitPackets = MIN(call->twind, call->cwind);
5594 xmitList = (struct rx_packet **)
5595 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
5596 /* XXXX else we must drop any mtx we hold */
5597 afs_osi_Alloc_NoSleep(maxXmitPackets * sizeof(struct rx_packet *));
5599 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5601 if (xmitList == NULL)
5602 osi_Panic("rxi_Start, failed to allocate xmit list");
5603 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5604 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5605 /* We shouldn't be sending packets if a thread is waiting
5606 * to initiate congestion recovery */
5607 dpf(("call %d waiting to initiate fast recovery\n",
5608 *(call->callNumber)));
5612 && (call->flags & RX_CALL_FAST_RECOVER)) {
5613 /* Only send one packet during fast recovery */
5614 dpf(("call %d restricted to one packet per send during fast recovery\n",
5615 *(call->callNumber)));
5618 #ifdef RX_TRACK_PACKETS
5619 if ((p->flags & RX_PKTFLAG_FREE)
5620 || (!queue_IsEnd(&call->tq, nxp)
5621 && (nxp->flags & RX_PKTFLAG_FREE))
5622 || (p == (struct rx_packet *)&rx_freePacketQueue)
5623 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5624 osi_Panic("rxi_Start: xmit queue clobbered");
5627 if (p->flags & RX_PKTFLAG_ACKED) {
5628 /* Since we may block, don't trust this */
5629 usenow.sec = usenow.usec = 0;
5630 if (rx_stats_active)
5631 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5632 continue; /* Ignore this packet if it has been acknowledged */
5635 /* Turn off all flags except these ones, which are the same
5636 * on each transmission */
5637 p->header.flags &= RX_PRESET_FLAGS;
5639 if (p->header.seq >=
5640 call->tfirst + MIN((int)call->twind,
5641 (int)(call->nSoftAcked +
5643 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5644 /* Note: if we're waiting for more window space, we can
5645 * still send retransmits; hence we don't return here, but
5646 * break out to schedule a retransmit event */
5647 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
5648 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
5653 /* Transmit the packet if it needs to be sent. */
5654 if (!clock_Lt(&now, &p->retryTime)) {
5655 if (nXmitPackets == maxXmitPackets) {
5656 rxi_SendXmitList(call, xmitList, nXmitPackets,
5657 istack, &now, &retryTime,
5659 osi_Free(xmitList, maxXmitPackets *
5660 sizeof(struct rx_packet *));
5663 dpf(("call %d xmit packet %"AFS_PTR_FMT" now %u.%06u retryTime %u.%06u nextRetry %u.%06u\n",
5664 *(call->callNumber), p,
5666 p->retryTime.sec, p->retryTime.usec,
5667 retryTime.sec, retryTime.usec));
5668 xmitList[nXmitPackets++] = p;
5672 /* xmitList now hold pointers to all of the packets that are
5673 * ready to send. Now we loop to send the packets */
5674 if (nXmitPackets > 0) {
5675 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5676 &now, &retryTime, resending);
5679 maxXmitPackets * sizeof(struct rx_packet *));
5681 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5683 * TQ references no longer protected by this flag; they must remain
5684 * protected by the global lock.
5686 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5687 call->flags &= ~RX_CALL_TQ_BUSY;
5688 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5689 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
5690 call, call->tqWaiters, call->flags));
5691 #ifdef RX_ENABLE_LOCKS
5692 osirx_AssertMine(&call->lock, "rxi_Start start");
5693 CV_BROADCAST(&call->cv_tq);
5694 #else /* RX_ENABLE_LOCKS */
5695 osi_rxWakeup(&call->tq);
5696 #endif /* RX_ENABLE_LOCKS */
5701 /* We went into the error state while sending packets. Now is
5702 * the time to reset the call. This will also inform the using
5703 * process that the call is in an error state.
5705 if (rx_stats_active)
5706 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5707 call->flags &= ~RX_CALL_TQ_BUSY;
5708 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5709 dpf(("call error %d while xmit %p has %d waiters and flags %d\n",
5710 call->error, call, call->tqWaiters, call->flags));
5711 #ifdef RX_ENABLE_LOCKS
5712 osirx_AssertMine(&call->lock, "rxi_Start middle");
5713 CV_BROADCAST(&call->cv_tq);
5714 #else /* RX_ENABLE_LOCKS */
5715 osi_rxWakeup(&call->tq);
5716 #endif /* RX_ENABLE_LOCKS */
5718 rxi_CallError(call, call->error);
5721 #ifdef RX_ENABLE_LOCKS
5722 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5724 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5725 /* Some packets have received acks. If they all have, we can clear
5726 * the transmit queue.
5729 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5730 if (p->header.seq < call->tfirst
5731 && (p->flags & RX_PKTFLAG_ACKED)) {
5733 #ifdef RX_TRACK_PACKETS
5734 p->flags &= ~RX_PKTFLAG_TQ;
5736 #ifdef RXDEBUG_PACKET
5744 call->flags |= RX_CALL_TQ_CLEARME;
5746 #endif /* RX_ENABLE_LOCKS */
5747 /* Don't bother doing retransmits if the TQ is cleared. */
5748 if (call->flags & RX_CALL_TQ_CLEARME) {
5749 rxi_ClearTransmitQueue(call, 1);
5751 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5754 /* Always post a resend event, if there is anything in the
5755 * queue, and resend is possible. There should be at least
5756 * one unacknowledged packet in the queue ... otherwise none
5757 * of these packets should be on the queue in the first place.
5759 if (call->resendEvent) {
5760 /* Cancel the existing event and post a new one */
5761 rxevent_Cancel(call->resendEvent, call,
5762 RX_CALL_REFCOUNT_RESEND);
5765 /* The retry time is the retry time on the first unacknowledged
5766 * packet inside the current window */
5768 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5769 /* Don't set timers for packets outside the window */
5770 if (p->header.seq >= call->tfirst + call->twind) {
5774 if (!(p->flags & RX_PKTFLAG_ACKED)
5775 && !clock_IsZero(&p->retryTime)) {
5777 retryTime = p->retryTime;
5782 /* Post a new event to re-run rxi_Start when retries may be needed */
5783 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5784 #ifdef RX_ENABLE_LOCKS
5785 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5787 rxevent_PostNow2(&retryTime, &usenow,
5789 (void *)call, 0, istack);
5790 #else /* RX_ENABLE_LOCKS */
5792 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5793 (void *)call, 0, istack);
5794 #endif /* RX_ENABLE_LOCKS */
5797 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5798 } while (call->flags & RX_CALL_NEED_START);
5800 * TQ references no longer protected by this flag; they must remain
5801 * protected by the global lock.
5803 call->flags &= ~RX_CALL_TQ_BUSY;
5804 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5805 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
5806 call, call->tqWaiters, call->flags));
5807 #ifdef RX_ENABLE_LOCKS
5808 osirx_AssertMine(&call->lock, "rxi_Start end");
5809 CV_BROADCAST(&call->cv_tq);
5810 #else /* RX_ENABLE_LOCKS */
5811 osi_rxWakeup(&call->tq);
5812 #endif /* RX_ENABLE_LOCKS */
5815 call->flags |= RX_CALL_NEED_START;
5817 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5819 if (call->resendEvent) {
5820 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5825 /* Also adjusts the keep alive parameters for the call, to reflect
5826 * that we have just sent a packet (so keep alives aren't sent
5829 rxi_Send(struct rx_call *call, struct rx_packet *p,
5832 struct rx_connection *conn = call->conn;
5834 /* Stamp each packet with the user supplied status */
5835 p->header.userStatus = call->localStatus;
5837 /* Allow the security object controlling this call's security to
5838 * make any last-minute changes to the packet */
5839 RXS_SendPacket(conn->securityObject, call, p);
5841 /* Since we're about to send SOME sort of packet to the peer, it's
5842 * safe to nuke any scheduled end-of-packets ack */
5843 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5845 /* Actually send the packet, filling in more connection-specific fields */
5846 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5847 MUTEX_EXIT(&call->lock);
5848 rxi_SendPacket(call, conn, p, istack);
5849 MUTEX_ENTER(&call->lock);
5850 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5852 /* Update last send time for this call (for keep-alive
5853 * processing), and for the connection (so that we can discover
5854 * idle connections) */
5855 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5856 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
5857 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
5859 conn->lastSendTime = call->lastSendTime = clock_Sec();
5860 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
5861 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5862 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
5863 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
5864 RX_ACK_PING_RESPONSE)))
5865 call->lastSendData = call->lastSendTime;
5869 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5870 * that things are fine. Also called periodically to guarantee that nothing
5871 * falls through the cracks (e.g. (error + dally) connections have keepalive
5872 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5874 * haveCTLock Set if calling from rxi_ReapConnections
5876 #ifdef RX_ENABLE_LOCKS
5878 rxi_CheckCall(struct rx_call *call, int haveCTLock)
5879 #else /* RX_ENABLE_LOCKS */
5881 rxi_CheckCall(struct rx_call *call)
5882 #endif /* RX_ENABLE_LOCKS */
5884 struct rx_connection *conn = call->conn;
5886 afs_uint32 deadTime;
5890 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5891 if (call->flags & RX_CALL_TQ_BUSY) {
5892 /* Call is active and will be reset by rxi_Start if it's
5893 * in an error state.
5898 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5900 (((afs_uint32) conn->secondsUntilDead << 10) +
5901 ((afs_uint32) conn->peer->rtt >> 3) +
5902 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5904 /* These are computed to the second (+- 1 second). But that's
5905 * good enough for these values, which should be a significant
5906 * number of seconds. */
5907 if (now > (call->lastReceiveTime + deadTime)) {
5908 if (call->state == RX_STATE_ACTIVE) {
5910 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5912 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5913 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5914 ip_stack_t *ipst = ns->netstack_ip;
5916 ire = ire_cache_lookup(conn->peer->host
5917 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5919 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5921 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5928 if (ire && ire->ire_max_frag > 0)
5929 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
5931 #if defined(GLOBAL_NETSTACKID)
5935 #endif /* ADAPT_PMTU */
5936 cerror = RX_CALL_DEAD;
5939 #ifdef RX_ENABLE_LOCKS
5940 /* Cancel pending events */
5941 rxevent_Cancel(call->delayedAckEvent, call,
5942 RX_CALL_REFCOUNT_DELAY);
5943 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5944 rxevent_Cancel(call->keepAliveEvent, call,
5945 RX_CALL_REFCOUNT_ALIVE);
5946 if (call->refCount == 0) {
5947 rxi_FreeCall(call, haveCTLock);
5951 #else /* RX_ENABLE_LOCKS */
5954 #endif /* RX_ENABLE_LOCKS */
5956 /* Non-active calls are destroyed if they are not responding
5957 * to pings; active calls are simply flagged in error, so the
5958 * attached process can die reasonably gracefully. */
5960 /* see if we have a non-activity timeout */
5961 if (call->startWait && conn->idleDeadTime
5962 && ((call->startWait + conn->idleDeadTime) < now) &&
5963 (call->flags & RX_CALL_READER_WAIT)) {
5964 if (call->state == RX_STATE_ACTIVE) {
5965 cerror = RX_CALL_TIMEOUT;
5969 if (call->lastSendData && conn->idleDeadTime && (conn->idleDeadErr != 0)
5970 && ((call->lastSendData + conn->idleDeadTime) < now)) {
5971 if (call->state == RX_STATE_ACTIVE) {
5972 cerror = conn->idleDeadErr;
5976 /* see if we have a hard timeout */
5977 if (conn->hardDeadTime
5978 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5979 if (call->state == RX_STATE_ACTIVE)
5980 rxi_CallError(call, RX_CALL_TIMEOUT);
5985 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT) {
5986 int oldMTU = conn->peer->ifMTU;
5988 /* if we thought we could send more, perhaps things got worse */
5989 if (call->conn->peer->maxPacketSize > conn->lastPacketSize)
5990 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
5991 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
5992 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
5994 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
5996 /* minimum capped in SetPeerMtu */
5997 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6000 conn->lastPacketSize = 0;
6002 /* needed so ResetCall doesn't clobber us. */
6003 call->MTU = conn->peer->ifMTU;
6005 /* if we never succeeded, let the error pass out as-is */
6006 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6007 cerror = conn->msgsizeRetryErr;
6010 rxi_CallError(call, cerror);
6015 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6017 struct rx_connection *conn = arg1;
6018 struct rx_header theader;
6020 struct sockaddr_in taddr;
6023 struct iovec tmpiov[2];
6026 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6029 tp = &tbuffer[sizeof(struct rx_header)];
6030 taddr.sin_family = AF_INET;
6031 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6032 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6033 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6034 taddr.sin_len = sizeof(struct sockaddr_in);
6036 memset(&theader, 0, sizeof(theader));
6037 theader.epoch = htonl(999);
6039 theader.callNumber = 0;
6042 theader.type = RX_PACKET_TYPE_VERSION;
6043 theader.flags = RX_LAST_PACKET;
6044 theader.serviceId = 0;
6046 memcpy(tbuffer, &theader, sizeof(theader));
6047 memcpy(tp, &a, sizeof(a));
6048 tmpiov[0].iov_base = tbuffer;
6049 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6051 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6053 MUTEX_ENTER(&conn->conn_data_lock);
6054 /* Only reschedule ourselves if the connection would not be destroyed */
6055 if (conn->refCount <= 1) {
6056 conn->natKeepAliveEvent = NULL;
6057 MUTEX_EXIT(&conn->conn_data_lock);
6058 rx_DestroyConnection(conn); /* drop the reference for this */
6060 conn->natKeepAliveEvent = NULL;
6061 conn->refCount--; /* drop the reference for this */
6062 rxi_ScheduleNatKeepAliveEvent(conn);
6063 MUTEX_EXIT(&conn->conn_data_lock);
6068 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6070 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6071 struct clock when, now;
6072 clock_GetTime(&now);
6074 when.sec += conn->secondsUntilNatPing;
6075 conn->refCount++; /* hold a reference for this */
6076 conn->natKeepAliveEvent =
6077 rxevent_PostNow(&when, &now, rxi_NatKeepAliveEvent, conn, 0);
6082 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6084 MUTEX_ENTER(&conn->conn_data_lock);
6085 conn->secondsUntilNatPing = seconds;
6087 rxi_ScheduleNatKeepAliveEvent(conn);
6088 MUTEX_EXIT(&conn->conn_data_lock);
6092 rxi_NatKeepAliveOn(struct rx_connection *conn)
6094 MUTEX_ENTER(&conn->conn_data_lock);
6095 rxi_ScheduleNatKeepAliveEvent(conn);
6096 MUTEX_EXIT(&conn->conn_data_lock);
6099 /* When a call is in progress, this routine is called occasionally to
6100 * make sure that some traffic has arrived (or been sent to) the peer.
6101 * If nothing has arrived in a reasonable amount of time, the call is
6102 * declared dead; if nothing has been sent for a while, we send a
6103 * keep-alive packet (if we're actually trying to keep the call alive)
6106 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6108 struct rx_call *call = arg1;
6109 struct rx_connection *conn;
6112 MUTEX_ENTER(&call->lock);
6113 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6114 if (event == call->keepAliveEvent)
6115 call->keepAliveEvent = NULL;
6118 #ifdef RX_ENABLE_LOCKS
6119 if (rxi_CheckCall(call, 0)) {
6120 MUTEX_EXIT(&call->lock);
6123 #else /* RX_ENABLE_LOCKS */
6124 if (rxi_CheckCall(call))
6126 #endif /* RX_ENABLE_LOCKS */
6128 /* Don't try to keep alive dallying calls */
6129 if (call->state == RX_STATE_DALLY) {
6130 MUTEX_EXIT(&call->lock);
6135 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6136 /* Don't try to send keepalives if there is unacknowledged data */
6137 /* the rexmit code should be good enough, this little hack
6138 * doesn't quite work XXX */
6139 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6141 rxi_ScheduleKeepAliveEvent(call);
6142 MUTEX_EXIT(&call->lock);
6145 /* Does what's on the nameplate. */
6147 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy)
6149 struct rx_call *call = arg1;
6150 struct rx_connection *conn;
6152 MUTEX_ENTER(&call->lock);
6153 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6154 if (event == call->growMTUEvent)
6155 call->growMTUEvent = NULL;
6157 #ifdef RX_ENABLE_LOCKS
6158 if (rxi_CheckCall(call, 0)) {
6159 MUTEX_EXIT(&call->lock);
6162 #else /* RX_ENABLE_LOCKS */
6163 if (rxi_CheckCall(call))
6165 #endif /* RX_ENABLE_LOCKS */
6167 /* Don't bother with dallying calls */
6168 if (call->state == RX_STATE_DALLY) {
6169 MUTEX_EXIT(&call->lock);
6176 * keep being scheduled, just don't do anything if we're at peak,
6177 * or we're not set up to be properly handled (idle timeout required)
6179 if ((conn->peer->maxPacketSize != 0) &&
6180 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6181 (conn->idleDeadErr))
6182 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6183 rxi_ScheduleGrowMTUEvent(call, 0);
6184 MUTEX_EXIT(&call->lock);
6188 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6190 if (!call->keepAliveEvent) {
6191 struct clock when, now;
6192 clock_GetTime(&now);
6194 when.sec += call->conn->secondsUntilPing;
6195 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6196 call->keepAliveEvent =
6197 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
6202 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6204 if (!call->growMTUEvent) {
6205 struct clock when, now;
6207 clock_GetTime(&now);
6210 if (call->conn->secondsUntilPing)
6211 secs = (6*call->conn->secondsUntilPing)-1;
6213 if (call->conn->secondsUntilDead)
6214 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6218 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6219 call->growMTUEvent =
6220 rxevent_PostNow(&when, &now, rxi_GrowMTUEvent, call, 0);
6224 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6226 rxi_KeepAliveOn(struct rx_call *call)
6228 /* Pretend last packet received was received now--i.e. if another
6229 * packet isn't received within the keep alive time, then the call
6230 * will die; Initialize last send time to the current time--even
6231 * if a packet hasn't been sent yet. This will guarantee that a
6232 * keep-alive is sent within the ping time */
6233 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6234 rxi_ScheduleKeepAliveEvent(call);
6238 rxi_GrowMTUOn(struct rx_call *call)
6240 struct rx_connection *conn = call->conn;
6241 MUTEX_ENTER(&conn->conn_data_lock);
6242 conn->lastPingSizeSer = conn->lastPingSize = 0;
6243 MUTEX_EXIT(&conn->conn_data_lock);
6244 rxi_ScheduleGrowMTUEvent(call, 1);
6247 /* This routine is called to send connection abort messages
6248 * that have been delayed to throttle looping clients. */
6250 rxi_SendDelayedConnAbort(struct rxevent *event,
6251 void *arg1, void *unused)
6253 struct rx_connection *conn = arg1;
6256 struct rx_packet *packet;
6258 MUTEX_ENTER(&conn->conn_data_lock);
6259 conn->delayedAbortEvent = NULL;
6260 error = htonl(conn->error);
6262 MUTEX_EXIT(&conn->conn_data_lock);
6263 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6266 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6267 RX_PACKET_TYPE_ABORT, (char *)&error,
6269 rxi_FreePacket(packet);
6273 /* This routine is called to send call abort messages
6274 * that have been delayed to throttle looping clients. */
6276 rxi_SendDelayedCallAbort(struct rxevent *event,
6277 void *arg1, void *dummy)
6279 struct rx_call *call = arg1;
6282 struct rx_packet *packet;
6284 MUTEX_ENTER(&call->lock);
6285 call->delayedAbortEvent = NULL;
6286 error = htonl(call->error);
6288 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6291 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6292 (char *)&error, sizeof(error), 0);
6293 rxi_FreePacket(packet);
6295 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6296 MUTEX_EXIT(&call->lock);
6299 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6300 * seconds) to ask the client to authenticate itself. The routine
6301 * issues a challenge to the client, which is obtained from the
6302 * security object associated with the connection */
6304 rxi_ChallengeEvent(struct rxevent *event,
6305 void *arg0, void *arg1, int tries)
6307 struct rx_connection *conn = arg0;
6309 conn->challengeEvent = NULL;
6310 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6311 struct rx_packet *packet;
6312 struct clock when, now;
6315 /* We've failed to authenticate for too long.
6316 * Reset any calls waiting for authentication;
6317 * they are all in RX_STATE_PRECALL.
6321 MUTEX_ENTER(&conn->conn_call_lock);
6322 for (i = 0; i < RX_MAXCALLS; i++) {
6323 struct rx_call *call = conn->call[i];
6325 MUTEX_ENTER(&call->lock);
6326 if (call->state == RX_STATE_PRECALL) {
6327 rxi_CallError(call, RX_CALL_DEAD);
6328 rxi_SendCallAbort(call, NULL, 0, 0);
6330 MUTEX_EXIT(&call->lock);
6333 MUTEX_EXIT(&conn->conn_call_lock);
6337 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6339 /* If there's no packet available, do this later. */
6340 RXS_GetChallenge(conn->securityObject, conn, packet);
6341 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6342 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6343 rxi_FreePacket(packet);
6345 clock_GetTime(&now);
6347 when.sec += RX_CHALLENGE_TIMEOUT;
6348 conn->challengeEvent =
6349 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
6354 /* Call this routine to start requesting the client to authenticate
6355 * itself. This will continue until authentication is established,
6356 * the call times out, or an invalid response is returned. The
6357 * security object associated with the connection is asked to create
6358 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6359 * defined earlier. */
6361 rxi_ChallengeOn(struct rx_connection *conn)
6363 if (!conn->challengeEvent) {
6364 RXS_CreateChallenge(conn->securityObject, conn);
6365 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6370 /* Compute round trip time of the packet provided, in *rttp.
6373 /* rxi_ComputeRoundTripTime is called with peer locked. */
6374 /* sentp and/or peer may be null */
6376 rxi_ComputeRoundTripTime(struct rx_packet *p,
6377 struct clock *sentp,
6378 struct rx_peer *peer)
6380 struct clock thisRtt, *rttp = &thisRtt;
6384 clock_GetTime(rttp);
6386 if (clock_Lt(rttp, sentp)) {
6388 return; /* somebody set the clock back, don't count this time. */
6390 clock_Sub(rttp, sentp);
6391 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6392 p->header.callNumber, p, rttp->sec, rttp->usec));
6394 if (rttp->sec == 0 && rttp->usec == 0) {
6396 * The actual round trip time is shorter than the
6397 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6398 * Since we can't tell which at the moment we will assume 1ms.
6403 if (rx_stats_active) {
6404 MUTEX_ENTER(&rx_stats_mutex);
6405 if (clock_Lt(rttp, &rx_stats.minRtt))
6406 rx_stats.minRtt = *rttp;
6407 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
6408 if (rttp->sec > 60) {
6409 MUTEX_EXIT(&rx_stats_mutex);
6410 return; /* somebody set the clock ahead */
6412 rx_stats.maxRtt = *rttp;
6414 clock_Add(&rx_stats.totalRtt, rttp);
6415 rx_stats.nRttSamples++;
6416 MUTEX_EXIT(&rx_stats_mutex);
6419 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6421 /* Apply VanJacobson round-trip estimations */
6426 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
6427 * srtt is stored as fixed point with 3 bits after the binary
6428 * point (i.e., scaled by 8). The following magic is
6429 * equivalent to the smoothing algorithm in rfc793 with an
6430 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6431 * srtt'*8 = rtt + srtt*7
6432 * srtt'*8 = srtt*8 + rtt - srtt
6433 * srtt' = srtt + rtt/8 - srtt/8
6434 * srtt' = srtt + (rtt - srtt)/8
6437 delta = _8THMSEC(rttp) - peer->rtt;
6438 peer->rtt += (delta >> 3);
6441 * We accumulate a smoothed rtt variance (actually, a smoothed
6442 * mean difference), then set the retransmit timer to smoothed
6443 * rtt + 4 times the smoothed variance (was 2x in van's original
6444 * paper, but 4x works better for me, and apparently for him as
6446 * rttvar is stored as
6447 * fixed point with 2 bits after the binary point (scaled by
6448 * 4). The following is equivalent to rfc793 smoothing with
6449 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6450 * rttvar'*4 = rttvar*3 + |delta|
6451 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6452 * rttvar' = rttvar + |delta|/4 - rttvar/4
6453 * rttvar' = rttvar + (|delta| - rttvar)/4
6454 * This replaces rfc793's wired-in beta.
6455 * dev*4 = dev*4 + (|actual - expected| - dev)
6461 delta -= (peer->rtt_dev << 1);
6462 peer->rtt_dev += (delta >> 3);
6464 /* I don't have a stored RTT so I start with this value. Since I'm
6465 * probably just starting a call, and will be pushing more data down
6466 * this, I expect congestion to increase rapidly. So I fudge a
6467 * little, and I set deviance to half the rtt. In practice,
6468 * deviance tends to approach something a little less than
6469 * half the smoothed rtt. */
6470 peer->rtt = _8THMSEC(rttp) + 8;
6471 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
6473 /* the timeout is RTT + 4*MDEV but no less than rx_minPeerTimeout msec.
6474 * This is because one end or the other of these connections is usually
6475 * in a user process, and can be switched and/or swapped out. So on fast,
6476 * reliable networks, the timeout would otherwise be too short. */
6477 rtt_timeout = MAX(((peer->rtt >> 3) + peer->rtt_dev), rx_minPeerTimeout);
6478 clock_Zero(&(peer->timeout));
6479 clock_Addmsec(&(peer->timeout), rtt_timeout);
6481 /* Reset the backedOff flag since we just computed a new timeout value */
6482 peer->backedOff = 0;
6484 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6485 p->header.callNumber, p, MSEC(rttp), peer->rtt >> 3, peer->rtt_dev >> 2, (peer->timeout.sec), (peer->timeout.usec)));
6489 /* Find all server connections that have not been active for a long time, and
6492 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6494 struct clock now, when;
6495 clock_GetTime(&now);
6497 /* Find server connection structures that haven't been used for
6498 * greater than rx_idleConnectionTime */
6500 struct rx_connection **conn_ptr, **conn_end;
6501 int i, havecalls = 0;
6502 MUTEX_ENTER(&rx_connHashTable_lock);
6503 for (conn_ptr = &rx_connHashTable[0], conn_end =
6504 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6506 struct rx_connection *conn, *next;
6507 struct rx_call *call;
6511 for (conn = *conn_ptr; conn; conn = next) {
6512 /* XXX -- Shouldn't the connection be locked? */
6515 for (i = 0; i < RX_MAXCALLS; i++) {
6516 call = conn->call[i];
6520 code = MUTEX_TRYENTER(&call->lock);
6523 #ifdef RX_ENABLE_LOCKS
6524 result = rxi_CheckCall(call, 1);
6525 #else /* RX_ENABLE_LOCKS */
6526 result = rxi_CheckCall(call);
6527 #endif /* RX_ENABLE_LOCKS */
6528 MUTEX_EXIT(&call->lock);
6530 /* If CheckCall freed the call, it might
6531 * have destroyed the connection as well,
6532 * which screws up the linked lists.
6538 if (conn->type == RX_SERVER_CONNECTION) {
6539 /* This only actually destroys the connection if
6540 * there are no outstanding calls */
6541 MUTEX_ENTER(&conn->conn_data_lock);
6542 if (!havecalls && !conn->refCount
6543 && ((conn->lastSendTime + rx_idleConnectionTime) <
6545 conn->refCount++; /* it will be decr in rx_DestroyConn */
6546 MUTEX_EXIT(&conn->conn_data_lock);
6547 #ifdef RX_ENABLE_LOCKS
6548 rxi_DestroyConnectionNoLock(conn);
6549 #else /* RX_ENABLE_LOCKS */
6550 rxi_DestroyConnection(conn);
6551 #endif /* RX_ENABLE_LOCKS */
6553 #ifdef RX_ENABLE_LOCKS
6555 MUTEX_EXIT(&conn->conn_data_lock);
6557 #endif /* RX_ENABLE_LOCKS */
6561 #ifdef RX_ENABLE_LOCKS
6562 while (rx_connCleanup_list) {
6563 struct rx_connection *conn;
6564 conn = rx_connCleanup_list;
6565 rx_connCleanup_list = rx_connCleanup_list->next;
6566 MUTEX_EXIT(&rx_connHashTable_lock);
6567 rxi_CleanupConnection(conn);
6568 MUTEX_ENTER(&rx_connHashTable_lock);
6570 MUTEX_EXIT(&rx_connHashTable_lock);
6571 #endif /* RX_ENABLE_LOCKS */
6574 /* Find any peer structures that haven't been used (haven't had an
6575 * associated connection) for greater than rx_idlePeerTime */
6577 struct rx_peer **peer_ptr, **peer_end;
6581 * Why do we need to hold the rx_peerHashTable_lock across
6582 * the incrementing of peer_ptr since the rx_peerHashTable
6583 * array is not changing? We don't.
6585 * By dropping the lock periodically we can permit other
6586 * activities to be performed while a rxi_ReapConnections
6587 * call is in progress. The goal of reap connections
6588 * is to clean up quickly without causing large amounts
6589 * of contention. Therefore, it is important that global
6590 * mutexes not be held for extended periods of time.
6592 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6593 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6595 struct rx_peer *peer, *next, *prev;
6597 MUTEX_ENTER(&rx_peerHashTable_lock);
6598 for (prev = peer = *peer_ptr; peer; peer = next) {
6600 code = MUTEX_TRYENTER(&peer->peer_lock);
6601 if ((code) && (peer->refCount == 0)
6602 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6603 rx_interface_stat_p rpc_stat, nrpc_stat;
6607 * now know that this peer object is one to be
6608 * removed from the hash table. Once it is removed
6609 * it can't be referenced by other threads.
6610 * Lets remove it first and decrement the struct
6611 * nPeerStructs count.
6613 if (peer == *peer_ptr) {
6619 if (rx_stats_active)
6620 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6623 * Now if we hold references on 'prev' and 'next'
6624 * we can safely drop the rx_peerHashTable_lock
6625 * while we destroy this 'peer' object.
6631 MUTEX_EXIT(&rx_peerHashTable_lock);
6633 MUTEX_EXIT(&peer->peer_lock);
6634 MUTEX_DESTROY(&peer->peer_lock);
6636 (&peer->rpcStats, rpc_stat, nrpc_stat,
6637 rx_interface_stat)) {
6638 unsigned int num_funcs;
6641 queue_Remove(&rpc_stat->queue_header);
6642 queue_Remove(&rpc_stat->all_peers);
6643 num_funcs = rpc_stat->stats[0].func_total;
6645 sizeof(rx_interface_stat_t) +
6646 rpc_stat->stats[0].func_total *
6647 sizeof(rx_function_entry_v1_t);
6649 rxi_Free(rpc_stat, space);
6651 MUTEX_ENTER(&rx_rpc_stats);
6652 rxi_rpc_peer_stat_cnt -= num_funcs;
6653 MUTEX_EXIT(&rx_rpc_stats);
6658 * Regain the rx_peerHashTable_lock and
6659 * decrement the reference count on 'prev'
6662 MUTEX_ENTER(&rx_peerHashTable_lock);
6669 MUTEX_EXIT(&peer->peer_lock);
6674 MUTEX_EXIT(&rx_peerHashTable_lock);
6678 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6679 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6680 * GC, just below. Really, we shouldn't have to keep moving packets from
6681 * one place to another, but instead ought to always know if we can
6682 * afford to hold onto a packet in its particular use. */
6683 MUTEX_ENTER(&rx_freePktQ_lock);
6684 if (rx_waitingForPackets) {
6685 rx_waitingForPackets = 0;
6686 #ifdef RX_ENABLE_LOCKS
6687 CV_BROADCAST(&rx_waitingForPackets_cv);
6689 osi_rxWakeup(&rx_waitingForPackets);
6692 MUTEX_EXIT(&rx_freePktQ_lock);
6695 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6696 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6700 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6701 * rx.h is sort of strange this is better. This is called with a security
6702 * object before it is discarded. Each connection using a security object has
6703 * its own refcount to the object so it won't actually be freed until the last
6704 * connection is destroyed.
6706 * This is the only rxs module call. A hold could also be written but no one
6710 rxs_Release(struct rx_securityClass *aobj)
6712 return RXS_Close(aobj);
6716 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6717 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6718 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6719 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6721 /* Adjust our estimate of the transmission rate to this peer, given
6722 * that the packet p was just acked. We can adjust peer->timeout and
6723 * call->twind. Pragmatically, this is called
6724 * only with packets of maximal length.
6725 * Called with peer and call locked.
6729 rxi_ComputeRate(struct rx_peer *peer, struct rx_call *call,
6730 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6732 afs_int32 xferSize, xferMs;
6736 /* Count down packets */
6737 if (peer->rateFlag > 0)
6739 /* Do nothing until we're enabled */
6740 if (peer->rateFlag != 0)
6745 /* Count only when the ack seems legitimate */
6746 switch (ackReason) {
6747 case RX_ACK_REQUESTED:
6749 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6753 case RX_ACK_PING_RESPONSE:
6754 if (p) /* want the response to ping-request, not data send */
6756 clock_GetTime(&newTO);
6757 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6758 clock_Sub(&newTO, &call->pingRequestTime);
6759 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6763 xferSize = rx_AckDataSize(rx_maxSendWindow) + RX_HEADER_SIZE;
6770 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %d.%06d, rtt %u, ps %u)",
6771 ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"),
6772 xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
6774 /* Track only packets that are big enough. */
6775 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6779 /* absorb RTT data (in milliseconds) for these big packets */
6780 if (peer->smRtt == 0) {
6781 peer->smRtt = xferMs;
6783 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6788 if (peer->countDown) {
6792 peer->countDown = 10; /* recalculate only every so often */
6794 /* In practice, we can measure only the RTT for full packets,
6795 * because of the way Rx acks the data that it receives. (If it's
6796 * smaller than a full packet, it often gets implicitly acked
6797 * either by the call response (from a server) or by the next call
6798 * (from a client), and either case confuses transmission times
6799 * with processing times.) Therefore, replace the above
6800 * more-sophisticated processing with a simpler version, where the
6801 * smoothed RTT is kept for full-size packets, and the time to
6802 * transmit a windowful of full-size packets is simply RTT *
6803 * windowSize. Again, we take two steps:
6804 - ensure the timeout is large enough for a single packet's RTT;
6805 - ensure that the window is small enough to fit in the desired timeout.*/
6807 /* First, the timeout check. */
6808 minTime = peer->smRtt;
6809 /* Get a reasonable estimate for a timeout period */
6811 newTO.sec = minTime / 1000;
6812 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6814 /* Increase the timeout period so that we can always do at least
6815 * one packet exchange */
6816 if (clock_Gt(&newTO, &peer->timeout)) {
6818 dpf(("CONG peer %lx/%u: timeout %d.%06d ==> %ld.%06d (rtt %u)",
6819 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec,
6820 newTO.sec, newTO.usec, peer->smRtt));
6822 peer->timeout = newTO;
6825 /* Now, get an estimate for the transmit window size. */
6826 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6827 /* Now, convert to the number of full packets that could fit in a
6828 * reasonable fraction of that interval */
6829 minTime /= (peer->smRtt << 1);
6830 minTime = MAX(minTime, rx_minPeerTimeout);
6831 xferSize = minTime; /* (make a copy) */
6833 /* Now clamp the size to reasonable bounds. */
6836 else if (minTime > rx_maxSendWindow)
6837 minTime = rx_maxSendWindow;
6838 /* if (minTime != peer->maxWindow) {
6839 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u)",
6840 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6841 peer->timeout.sec, peer->timeout.usec, peer->smRtt));
6842 peer->maxWindow = minTime;
6843 elide... call->twind = minTime;
6847 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6848 * Discern this by calculating the timeout necessary for rx_Window
6850 if ((xferSize > rx_maxSendWindow) && (peer->timeout.sec >= 3)) {
6851 /* calculate estimate for transmission interval in milliseconds */
6852 minTime = rx_maxSendWindow * peer->smRtt;
6853 if (minTime < 1000) {
6854 dpf(("CONG peer %lx/%u: cut TO %d.%06d by 0.5 (rtt %u)",
6855 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6856 peer->timeout.usec, peer->smRtt));
6858 newTO.sec = 0; /* cut back on timeout by half a second */
6859 newTO.usec = 500000;
6860 clock_Sub(&peer->timeout, &newTO);
6865 } /* end of rxi_ComputeRate */
6866 #endif /* ADAPT_WINDOW */
6874 #define TRACE_OPTION_RX_DEBUG 16
6882 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6883 0, KEY_QUERY_VALUE, &parmKey);
6884 if (code != ERROR_SUCCESS)
6887 dummyLen = sizeof(TraceOption);
6888 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6889 (BYTE *) &TraceOption, &dummyLen);
6890 if (code == ERROR_SUCCESS) {
6891 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
6893 RegCloseKey (parmKey);
6894 #endif /* AFS_NT40_ENV */
6899 rx_DebugOnOff(int on)
6903 rxdebug_active = on;
6909 rx_StatsOnOff(int on)
6912 rx_stats_active = on;
6917 /* Don't call this debugging routine directly; use dpf */
6919 rxi_DebugPrint(char *format, ...)
6928 va_start(ap, format);
6930 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6933 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
6935 if (msg[len-1] != '\n') {
6939 OutputDebugString(msg);
6946 va_start(ap, format);
6948 clock_GetTime(&now);
6949 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
6950 (unsigned int)now.usec);
6951 vfprintf(rx_Log, format, ap);
6960 * This function is used to process the rx_stats structure that is local
6961 * to a process as well as an rx_stats structure received from a remote
6962 * process (via rxdebug). Therefore, it needs to do minimal version
6966 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
6967 afs_int32 freePackets, char version)
6971 if (size != sizeof(struct rx_statistics)) {
6973 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
6974 size, sizeof(struct rx_statistics));
6977 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6980 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6981 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
6982 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6983 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6984 s->specialPktAllocFailures);
6986 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
6987 s->receivePktAllocFailures, s->sendPktAllocFailures,
6988 s->specialPktAllocFailures);
6992 " greedy %u, " "bogusReads %u (last from host %x), "
6993 "noPackets %u, " "noBuffers %u, " "selects %u, "
6994 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
6995 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6996 s->selects, s->sendSelects);
6998 fprintf(file, " packets read: ");
6999 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7000 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7002 fprintf(file, "\n");
7005 " other read counters: data %u, " "ack %u, " "dup %u "
7006 "spurious %u " "dally %u\n", s->dataPacketsRead,
7007 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7008 s->ignorePacketDally);
7010 fprintf(file, " packets sent: ");
7011 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7012 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7014 fprintf(file, "\n");
7017 " other send counters: ack %u, " "data %u (not resends), "
7018 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7019 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7020 s->dataPacketsPushed, s->ignoreAckedPacket);
7023 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7024 s->netSendFailures, (int)s->fatalErrors);
7026 if (s->nRttSamples) {
7027 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7028 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7030 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7031 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7035 " %d server connections, " "%d client connections, "
7036 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7037 s->nServerConns, s->nClientConns, s->nPeerStructs,
7038 s->nCallStructs, s->nFreeCallStructs);
7040 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7041 fprintf(file, " %d clock updates\n", clock_nUpdates);
7045 /* for backward compatibility */
7047 rx_PrintStats(FILE * file)
7049 MUTEX_ENTER(&rx_stats_mutex);
7050 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
7052 MUTEX_EXIT(&rx_stats_mutex);
7056 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7058 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
7059 ntohl(peer->host), (int)ntohs(peer->port), (int)peer->burstSize,
7060 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
7063 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
7064 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
7065 (int)peer->timeout.usec, peer->nSent, peer->reSends);
7068 " Packet size %d, " "max in packet skew %d, "
7069 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
7070 (int)peer->outPacketSkew);
7074 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7076 * This mutex protects the following static variables:
7080 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7081 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7083 #define LOCK_RX_DEBUG
7084 #define UNLOCK_RX_DEBUG
7085 #endif /* AFS_PTHREAD_ENV */
7087 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7089 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7090 u_char type, void *inputData, size_t inputLength,
7091 void *outputData, size_t outputLength)
7093 static afs_int32 counter = 100;
7094 time_t waitTime, waitCount;
7095 struct rx_header theader;
7098 struct timeval tv_now, tv_wake, tv_delta;
7099 struct sockaddr_in taddr, faddr;
7113 tp = &tbuffer[sizeof(struct rx_header)];
7114 taddr.sin_family = AF_INET;
7115 taddr.sin_port = remotePort;
7116 taddr.sin_addr.s_addr = remoteAddr;
7117 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7118 taddr.sin_len = sizeof(struct sockaddr_in);
7121 memset(&theader, 0, sizeof(theader));
7122 theader.epoch = htonl(999);
7124 theader.callNumber = htonl(counter);
7127 theader.type = type;
7128 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7129 theader.serviceId = 0;
7131 memcpy(tbuffer, &theader, sizeof(theader));
7132 memcpy(tp, inputData, inputLength);
7134 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7135 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7137 /* see if there's a packet available */
7138 gettimeofday(&tv_wake,0);
7139 tv_wake.tv_sec += waitTime;
7142 FD_SET(socket, &imask);
7143 tv_delta.tv_sec = tv_wake.tv_sec;
7144 tv_delta.tv_usec = tv_wake.tv_usec;
7145 gettimeofday(&tv_now, 0);
7147 if (tv_delta.tv_usec < tv_now.tv_usec) {
7149 tv_delta.tv_usec += 1000000;
7152 tv_delta.tv_usec -= tv_now.tv_usec;
7154 if (tv_delta.tv_sec < tv_now.tv_sec) {
7158 tv_delta.tv_sec -= tv_now.tv_sec;
7161 code = select(0, &imask, 0, 0, &tv_delta);
7162 #else /* AFS_NT40_ENV */
7163 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7164 #endif /* AFS_NT40_ENV */
7165 if (code == 1 && FD_ISSET(socket, &imask)) {
7166 /* now receive a packet */
7167 faddrLen = sizeof(struct sockaddr_in);
7169 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7170 (struct sockaddr *)&faddr, &faddrLen);
7173 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7174 if (counter == ntohl(theader.callNumber))
7182 /* see if we've timed out */
7190 code -= sizeof(struct rx_header);
7191 if (code > outputLength)
7192 code = outputLength;
7193 memcpy(outputData, tp, code);
7196 #endif /* RXDEBUG */
7199 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7200 afs_uint16 remotePort, struct rx_debugStats * stat,
7201 afs_uint32 * supportedValues)
7203 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7205 struct rx_debugIn in;
7207 *supportedValues = 0;
7208 in.type = htonl(RX_DEBUGI_GETSTATS);
7211 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7212 &in, sizeof(in), stat, sizeof(*stat));
7215 * If the call was successful, fixup the version and indicate
7216 * what contents of the stat structure are valid.
7217 * Also do net to host conversion of fields here.
7221 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7222 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7224 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7225 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7227 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7228 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7230 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7231 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7233 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7234 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7236 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7237 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7239 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7240 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7242 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7243 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7245 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7246 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7248 stat->nFreePackets = ntohl(stat->nFreePackets);
7249 stat->packetReclaims = ntohl(stat->packetReclaims);
7250 stat->callsExecuted = ntohl(stat->callsExecuted);
7251 stat->nWaiting = ntohl(stat->nWaiting);
7252 stat->idleThreads = ntohl(stat->idleThreads);
7253 stat->nWaited = ntohl(stat->nWaited);
7254 stat->nPackets = ntohl(stat->nPackets);
7263 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7264 afs_uint16 remotePort, struct rx_statistics * stat,
7265 afs_uint32 * supportedValues)
7267 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7269 struct rx_debugIn in;
7270 afs_int32 *lp = (afs_int32 *) stat;
7274 * supportedValues is currently unused, but added to allow future
7275 * versioning of this function.
7278 *supportedValues = 0;
7279 in.type = htonl(RX_DEBUGI_RXSTATS);
7281 memset(stat, 0, sizeof(*stat));
7283 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7284 &in, sizeof(in), stat, sizeof(*stat));
7289 * Do net to host conversion here
7292 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7303 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7304 afs_uint16 remotePort, size_t version_length,
7307 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7309 return MakeDebugCall(socket, remoteAddr, remotePort,
7310 RX_PACKET_TYPE_VERSION, a, 1, version,
7318 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7319 afs_uint16 remotePort, afs_int32 * nextConnection,
7320 int allConnections, afs_uint32 debugSupportedValues,
7321 struct rx_debugConn * conn,
7322 afs_uint32 * supportedValues)
7324 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7326 struct rx_debugIn in;
7330 * supportedValues is currently unused, but added to allow future
7331 * versioning of this function.
7334 *supportedValues = 0;
7335 if (allConnections) {
7336 in.type = htonl(RX_DEBUGI_GETALLCONN);
7338 in.type = htonl(RX_DEBUGI_GETCONN);
7340 in.index = htonl(*nextConnection);
7341 memset(conn, 0, sizeof(*conn));
7343 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7344 &in, sizeof(in), conn, sizeof(*conn));
7347 *nextConnection += 1;
7350 * Convert old connection format to new structure.
7353 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7354 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7355 #define MOVEvL(a) (conn->a = vL->a)
7357 /* any old or unrecognized version... */
7358 for (i = 0; i < RX_MAXCALLS; i++) {
7359 MOVEvL(callState[i]);
7360 MOVEvL(callMode[i]);
7361 MOVEvL(callFlags[i]);
7362 MOVEvL(callOther[i]);
7364 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7365 MOVEvL(secStats.type);
7366 MOVEvL(secStats.level);
7367 MOVEvL(secStats.flags);
7368 MOVEvL(secStats.expires);
7369 MOVEvL(secStats.packetsReceived);
7370 MOVEvL(secStats.packetsSent);
7371 MOVEvL(secStats.bytesReceived);
7372 MOVEvL(secStats.bytesSent);
7377 * Do net to host conversion here
7379 * I don't convert host or port since we are most likely
7380 * going to want these in NBO.
7382 conn->cid = ntohl(conn->cid);
7383 conn->serial = ntohl(conn->serial);
7384 for (i = 0; i < RX_MAXCALLS; i++) {
7385 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7387 conn->error = ntohl(conn->error);
7388 conn->secStats.flags = ntohl(conn->secStats.flags);
7389 conn->secStats.expires = ntohl(conn->secStats.expires);
7390 conn->secStats.packetsReceived =
7391 ntohl(conn->secStats.packetsReceived);
7392 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7393 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7394 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7395 conn->epoch = ntohl(conn->epoch);
7396 conn->natMTU = ntohl(conn->natMTU);
7405 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7406 afs_uint16 remotePort, afs_int32 * nextPeer,
7407 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7408 afs_uint32 * supportedValues)
7410 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7412 struct rx_debugIn in;
7415 * supportedValues is currently unused, but added to allow future
7416 * versioning of this function.
7419 *supportedValues = 0;
7420 in.type = htonl(RX_DEBUGI_GETPEER);
7421 in.index = htonl(*nextPeer);
7422 memset(peer, 0, sizeof(*peer));
7424 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7425 &in, sizeof(in), peer, sizeof(*peer));
7431 * Do net to host conversion here
7433 * I don't convert host or port since we are most likely
7434 * going to want these in NBO.
7436 peer->ifMTU = ntohs(peer->ifMTU);
7437 peer->idleWhen = ntohl(peer->idleWhen);
7438 peer->refCount = ntohs(peer->refCount);
7439 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7440 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7441 peer->rtt = ntohl(peer->rtt);
7442 peer->rtt_dev = ntohl(peer->rtt_dev);
7443 peer->timeout.sec = ntohl(peer->timeout.sec);
7444 peer->timeout.usec = ntohl(peer->timeout.usec);
7445 peer->nSent = ntohl(peer->nSent);
7446 peer->reSends = ntohl(peer->reSends);
7447 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7448 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7449 peer->rateFlag = ntohl(peer->rateFlag);
7450 peer->natMTU = ntohs(peer->natMTU);
7451 peer->maxMTU = ntohs(peer->maxMTU);
7452 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7453 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7454 peer->MTU = ntohs(peer->MTU);
7455 peer->cwind = ntohs(peer->cwind);
7456 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7457 peer->congestSeq = ntohs(peer->congestSeq);
7458 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7459 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7460 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7461 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7470 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7471 struct rx_debugPeer * peerStats)
7474 afs_int32 error = 1; /* default to "did not succeed" */
7475 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7477 MUTEX_ENTER(&rx_peerHashTable_lock);
7478 for(tp = rx_peerHashTable[hashValue];
7479 tp != NULL; tp = tp->next) {
7480 if (tp->host == peerHost)
7486 MUTEX_EXIT(&rx_peerHashTable_lock);
7490 MUTEX_ENTER(&tp->peer_lock);
7491 peerStats->host = tp->host;
7492 peerStats->port = tp->port;
7493 peerStats->ifMTU = tp->ifMTU;
7494 peerStats->idleWhen = tp->idleWhen;
7495 peerStats->refCount = tp->refCount;
7496 peerStats->burstSize = tp->burstSize;
7497 peerStats->burst = tp->burst;
7498 peerStats->burstWait.sec = tp->burstWait.sec;
7499 peerStats->burstWait.usec = tp->burstWait.usec;
7500 peerStats->rtt = tp->rtt;
7501 peerStats->rtt_dev = tp->rtt_dev;
7502 peerStats->timeout.sec = tp->timeout.sec;
7503 peerStats->timeout.usec = tp->timeout.usec;
7504 peerStats->nSent = tp->nSent;
7505 peerStats->reSends = tp->reSends;
7506 peerStats->inPacketSkew = tp->inPacketSkew;
7507 peerStats->outPacketSkew = tp->outPacketSkew;
7508 peerStats->rateFlag = tp->rateFlag;
7509 peerStats->natMTU = tp->natMTU;
7510 peerStats->maxMTU = tp->maxMTU;
7511 peerStats->maxDgramPackets = tp->maxDgramPackets;
7512 peerStats->ifDgramPackets = tp->ifDgramPackets;
7513 peerStats->MTU = tp->MTU;
7514 peerStats->cwind = tp->cwind;
7515 peerStats->nDgramPackets = tp->nDgramPackets;
7516 peerStats->congestSeq = tp->congestSeq;
7517 peerStats->bytesSent.high = tp->bytesSent.high;
7518 peerStats->bytesSent.low = tp->bytesSent.low;
7519 peerStats->bytesReceived.high = tp->bytesReceived.high;
7520 peerStats->bytesReceived.low = tp->bytesReceived.low;
7521 MUTEX_EXIT(&tp->peer_lock);
7523 MUTEX_ENTER(&rx_peerHashTable_lock);
7526 MUTEX_EXIT(&rx_peerHashTable_lock);
7534 struct rx_serverQueueEntry *np;
7537 struct rx_call *call;
7538 struct rx_serverQueueEntry *sq;
7542 if (rxinit_status == 1) {
7544 return; /* Already shutdown. */
7548 #ifndef AFS_PTHREAD_ENV
7549 FD_ZERO(&rx_selectMask);
7550 #endif /* AFS_PTHREAD_ENV */
7551 rxi_dataQuota = RX_MAX_QUOTA;
7552 #ifndef AFS_PTHREAD_ENV
7554 #endif /* AFS_PTHREAD_ENV */
7557 #ifndef AFS_PTHREAD_ENV
7558 #ifndef AFS_USE_GETTIMEOFDAY
7560 #endif /* AFS_USE_GETTIMEOFDAY */
7561 #endif /* AFS_PTHREAD_ENV */
7563 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7564 call = queue_First(&rx_freeCallQueue, rx_call);
7566 rxi_Free(call, sizeof(struct rx_call));
7569 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7570 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7576 struct rx_peer **peer_ptr, **peer_end;
7577 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7578 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7580 struct rx_peer *peer, *next;
7582 MUTEX_ENTER(&rx_peerHashTable_lock);
7583 for (peer = *peer_ptr; peer; peer = next) {
7584 rx_interface_stat_p rpc_stat, nrpc_stat;
7587 MUTEX_ENTER(&rx_rpc_stats);
7588 MUTEX_ENTER(&peer->peer_lock);
7590 (&peer->rpcStats, rpc_stat, nrpc_stat,
7591 rx_interface_stat)) {
7592 unsigned int num_funcs;
7595 queue_Remove(&rpc_stat->queue_header);
7596 queue_Remove(&rpc_stat->all_peers);
7597 num_funcs = rpc_stat->stats[0].func_total;
7599 sizeof(rx_interface_stat_t) +
7600 rpc_stat->stats[0].func_total *
7601 sizeof(rx_function_entry_v1_t);
7603 rxi_Free(rpc_stat, space);
7605 /* rx_rpc_stats must be held */
7606 rxi_rpc_peer_stat_cnt -= num_funcs;
7608 MUTEX_EXIT(&peer->peer_lock);
7609 MUTEX_EXIT(&rx_rpc_stats);
7613 if (rx_stats_active)
7614 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
7616 MUTEX_EXIT(&rx_peerHashTable_lock);
7619 for (i = 0; i < RX_MAX_SERVICES; i++) {
7621 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7623 for (i = 0; i < rx_hashTableSize; i++) {
7624 struct rx_connection *tc, *ntc;
7625 MUTEX_ENTER(&rx_connHashTable_lock);
7626 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7628 for (j = 0; j < RX_MAXCALLS; j++) {
7630 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7633 rxi_Free(tc, sizeof(*tc));
7635 MUTEX_EXIT(&rx_connHashTable_lock);
7638 MUTEX_ENTER(&freeSQEList_lock);
7640 while ((np = rx_FreeSQEList)) {
7641 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7642 MUTEX_DESTROY(&np->lock);
7643 rxi_Free(np, sizeof(*np));
7646 MUTEX_EXIT(&freeSQEList_lock);
7647 MUTEX_DESTROY(&freeSQEList_lock);
7648 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7649 MUTEX_DESTROY(&rx_connHashTable_lock);
7650 MUTEX_DESTROY(&rx_peerHashTable_lock);
7651 MUTEX_DESTROY(&rx_serverPool_lock);
7653 osi_Free(rx_connHashTable,
7654 rx_hashTableSize * sizeof(struct rx_connection *));
7655 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7657 UNPIN(rx_connHashTable,
7658 rx_hashTableSize * sizeof(struct rx_connection *));
7659 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7661 rxi_FreeAllPackets();
7663 MUTEX_ENTER(&rx_quota_mutex);
7664 rxi_dataQuota = RX_MAX_QUOTA;
7665 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7666 MUTEX_EXIT(&rx_quota_mutex);
7671 #ifdef RX_ENABLE_LOCKS
7673 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7675 if (!MUTEX_ISMINE(lockaddr))
7676 osi_Panic("Lock not held: %s", msg);
7678 #endif /* RX_ENABLE_LOCKS */
7683 * Routines to implement connection specific data.
7687 rx_KeyCreate(rx_destructor_t rtn)
7690 MUTEX_ENTER(&rxi_keyCreate_lock);
7691 key = rxi_keyCreate_counter++;
7692 rxi_keyCreate_destructor = (rx_destructor_t *)
7693 realloc((void *)rxi_keyCreate_destructor,
7694 (key + 1) * sizeof(rx_destructor_t));
7695 rxi_keyCreate_destructor[key] = rtn;
7696 MUTEX_EXIT(&rxi_keyCreate_lock);
7701 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7704 MUTEX_ENTER(&conn->conn_data_lock);
7705 if (!conn->specific) {
7706 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7707 for (i = 0; i < key; i++)
7708 conn->specific[i] = NULL;
7709 conn->nSpecific = key + 1;
7710 conn->specific[key] = ptr;
7711 } else if (key >= conn->nSpecific) {
7712 conn->specific = (void **)
7713 realloc(conn->specific, (key + 1) * sizeof(void *));
7714 for (i = conn->nSpecific; i < key; i++)
7715 conn->specific[i] = NULL;
7716 conn->nSpecific = key + 1;
7717 conn->specific[key] = ptr;
7719 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7720 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7721 conn->specific[key] = ptr;
7723 MUTEX_EXIT(&conn->conn_data_lock);
7727 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7730 MUTEX_ENTER(&svc->svc_data_lock);
7731 if (!svc->specific) {
7732 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
7733 for (i = 0; i < key; i++)
7734 svc->specific[i] = NULL;
7735 svc->nSpecific = key + 1;
7736 svc->specific[key] = ptr;
7737 } else if (key >= svc->nSpecific) {
7738 svc->specific = (void **)
7739 realloc(svc->specific, (key + 1) * sizeof(void *));
7740 for (i = svc->nSpecific; i < key; i++)
7741 svc->specific[i] = NULL;
7742 svc->nSpecific = key + 1;
7743 svc->specific[key] = ptr;
7745 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7746 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7747 svc->specific[key] = ptr;
7749 MUTEX_EXIT(&svc->svc_data_lock);
7753 rx_GetSpecific(struct rx_connection *conn, int key)
7756 MUTEX_ENTER(&conn->conn_data_lock);
7757 if (key >= conn->nSpecific)
7760 ptr = conn->specific[key];
7761 MUTEX_EXIT(&conn->conn_data_lock);
7766 rx_GetServiceSpecific(struct rx_service *svc, int key)
7769 MUTEX_ENTER(&svc->svc_data_lock);
7770 if (key >= svc->nSpecific)
7773 ptr = svc->specific[key];
7774 MUTEX_EXIT(&svc->svc_data_lock);
7779 #endif /* !KERNEL */
7782 * processStats is a queue used to store the statistics for the local
7783 * process. Its contents are similar to the contents of the rpcStats
7784 * queue on a rx_peer structure, but the actual data stored within
7785 * this queue contains totals across the lifetime of the process (assuming
7786 * the stats have not been reset) - unlike the per peer structures
7787 * which can come and go based upon the peer lifetime.
7790 static struct rx_queue processStats = { &processStats, &processStats };
7793 * peerStats is a queue used to store the statistics for all peer structs.
7794 * Its contents are the union of all the peer rpcStats queues.
7797 static struct rx_queue peerStats = { &peerStats, &peerStats };
7800 * rxi_monitor_processStats is used to turn process wide stat collection
7804 static int rxi_monitor_processStats = 0;
7807 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7810 static int rxi_monitor_peerStats = 0;
7813 * rxi_AddRpcStat - given all of the information for a particular rpc
7814 * call, create (if needed) and update the stat totals for the rpc.
7818 * IN stats - the queue of stats that will be updated with the new value
7820 * IN rxInterface - a unique number that identifies the rpc interface
7822 * IN currentFunc - the index of the function being invoked
7824 * IN totalFunc - the total number of functions in this interface
7826 * IN queueTime - the amount of time this function waited for a thread
7828 * IN execTime - the amount of time this function invocation took to execute
7830 * IN bytesSent - the number bytes sent by this invocation
7832 * IN bytesRcvd - the number bytes received by this invocation
7834 * IN isServer - if true, this invocation was made to a server
7836 * IN remoteHost - the ip address of the remote host
7838 * IN remotePort - the port of the remote host
7840 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7842 * INOUT counter - if a new stats structure is allocated, the counter will
7843 * be updated with the new number of allocated stat structures
7851 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
7852 afs_uint32 currentFunc, afs_uint32 totalFunc,
7853 struct clock *queueTime, struct clock *execTime,
7854 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
7855 afs_uint32 remoteHost, afs_uint32 remotePort,
7856 int addToPeerList, unsigned int *counter)
7859 rx_interface_stat_p rpc_stat, nrpc_stat;
7862 * See if there's already a structure for this interface
7865 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7866 if ((rpc_stat->stats[0].interfaceId == rxInterface)
7867 && (rpc_stat->stats[0].remote_is_server == isServer))
7872 * Didn't find a match so allocate a new structure and add it to the
7876 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
7877 || (rpc_stat->stats[0].interfaceId != rxInterface)
7878 || (rpc_stat->stats[0].remote_is_server != isServer)) {
7883 sizeof(rx_interface_stat_t) +
7884 totalFunc * sizeof(rx_function_entry_v1_t);
7886 rpc_stat = rxi_Alloc(space);
7887 if (rpc_stat == NULL) {
7891 *counter += totalFunc;
7892 for (i = 0; i < totalFunc; i++) {
7893 rpc_stat->stats[i].remote_peer = remoteHost;
7894 rpc_stat->stats[i].remote_port = remotePort;
7895 rpc_stat->stats[i].remote_is_server = isServer;
7896 rpc_stat->stats[i].interfaceId = rxInterface;
7897 rpc_stat->stats[i].func_total = totalFunc;
7898 rpc_stat->stats[i].func_index = i;
7899 hzero(rpc_stat->stats[i].invocations);
7900 hzero(rpc_stat->stats[i].bytes_sent);
7901 hzero(rpc_stat->stats[i].bytes_rcvd);
7902 rpc_stat->stats[i].queue_time_sum.sec = 0;
7903 rpc_stat->stats[i].queue_time_sum.usec = 0;
7904 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7905 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7906 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7907 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7908 rpc_stat->stats[i].queue_time_max.sec = 0;
7909 rpc_stat->stats[i].queue_time_max.usec = 0;
7910 rpc_stat->stats[i].execution_time_sum.sec = 0;
7911 rpc_stat->stats[i].execution_time_sum.usec = 0;
7912 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7913 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7914 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7915 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7916 rpc_stat->stats[i].execution_time_max.sec = 0;
7917 rpc_stat->stats[i].execution_time_max.usec = 0;
7919 queue_Prepend(stats, rpc_stat);
7920 if (addToPeerList) {
7921 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7926 * Increment the stats for this function
7929 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7930 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7931 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7932 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7933 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7934 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7935 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7937 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7938 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7940 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7941 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7943 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7944 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7946 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7947 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7955 * rx_IncrementTimeAndCount - increment the times and count for a particular
7960 * IN peer - the peer who invoked the rpc
7962 * IN rxInterface - a unique number that identifies the rpc interface
7964 * IN currentFunc - the index of the function being invoked
7966 * IN totalFunc - the total number of functions in this interface
7968 * IN queueTime - the amount of time this function waited for a thread
7970 * IN execTime - the amount of time this function invocation took to execute
7972 * IN bytesSent - the number bytes sent by this invocation
7974 * IN bytesRcvd - the number bytes received by this invocation
7976 * IN isServer - if true, this invocation was made to a server
7984 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7985 afs_uint32 currentFunc, afs_uint32 totalFunc,
7986 struct clock *queueTime, struct clock *execTime,
7987 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7991 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7994 MUTEX_ENTER(&rx_rpc_stats);
7996 if (rxi_monitor_peerStats) {
7997 MUTEX_ENTER(&peer->peer_lock);
7998 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7999 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8000 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8001 MUTEX_EXIT(&peer->peer_lock);
8004 if (rxi_monitor_processStats) {
8005 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8006 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8007 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8010 MUTEX_EXIT(&rx_rpc_stats);
8015 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8019 * IN callerVersion - the rpc stat version of the caller.
8021 * IN count - the number of entries to marshall.
8023 * IN stats - pointer to stats to be marshalled.
8025 * OUT ptr - Where to store the marshalled data.
8032 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8033 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8039 * We only support the first version
8041 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8042 *(ptr++) = stats->remote_peer;
8043 *(ptr++) = stats->remote_port;
8044 *(ptr++) = stats->remote_is_server;
8045 *(ptr++) = stats->interfaceId;
8046 *(ptr++) = stats->func_total;
8047 *(ptr++) = stats->func_index;
8048 *(ptr++) = hgethi(stats->invocations);
8049 *(ptr++) = hgetlo(stats->invocations);
8050 *(ptr++) = hgethi(stats->bytes_sent);
8051 *(ptr++) = hgetlo(stats->bytes_sent);
8052 *(ptr++) = hgethi(stats->bytes_rcvd);
8053 *(ptr++) = hgetlo(stats->bytes_rcvd);
8054 *(ptr++) = stats->queue_time_sum.sec;
8055 *(ptr++) = stats->queue_time_sum.usec;
8056 *(ptr++) = stats->queue_time_sum_sqr.sec;
8057 *(ptr++) = stats->queue_time_sum_sqr.usec;
8058 *(ptr++) = stats->queue_time_min.sec;
8059 *(ptr++) = stats->queue_time_min.usec;
8060 *(ptr++) = stats->queue_time_max.sec;
8061 *(ptr++) = stats->queue_time_max.usec;
8062 *(ptr++) = stats->execution_time_sum.sec;
8063 *(ptr++) = stats->execution_time_sum.usec;
8064 *(ptr++) = stats->execution_time_sum_sqr.sec;
8065 *(ptr++) = stats->execution_time_sum_sqr.usec;
8066 *(ptr++) = stats->execution_time_min.sec;
8067 *(ptr++) = stats->execution_time_min.usec;
8068 *(ptr++) = stats->execution_time_max.sec;
8069 *(ptr++) = stats->execution_time_max.usec;
8075 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8080 * IN callerVersion - the rpc stat version of the caller
8082 * OUT myVersion - the rpc stat version of this function
8084 * OUT clock_sec - local time seconds
8086 * OUT clock_usec - local time microseconds
8088 * OUT allocSize - the number of bytes allocated to contain stats
8090 * OUT statCount - the number stats retrieved from this process.
8092 * OUT stats - the actual stats retrieved from this process.
8096 * Returns void. If successful, stats will != NULL.
8100 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8101 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8102 size_t * allocSize, afs_uint32 * statCount,
8103 afs_uint32 ** stats)
8113 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8116 * Check to see if stats are enabled
8119 MUTEX_ENTER(&rx_rpc_stats);
8120 if (!rxi_monitor_processStats) {
8121 MUTEX_EXIT(&rx_rpc_stats);
8125 clock_GetTime(&now);
8126 *clock_sec = now.sec;
8127 *clock_usec = now.usec;
8130 * Allocate the space based upon the caller version
8132 * If the client is at an older version than we are,
8133 * we return the statistic data in the older data format, but
8134 * we still return our version number so the client knows we
8135 * are maintaining more data than it can retrieve.
8138 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8139 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8140 *statCount = rxi_rpc_process_stat_cnt;
8143 * This can't happen yet, but in the future version changes
8144 * can be handled by adding additional code here
8148 if (space > (size_t) 0) {
8150 ptr = *stats = rxi_Alloc(space);
8153 rx_interface_stat_p rpc_stat, nrpc_stat;
8157 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8159 * Copy the data based upon the caller version
8161 rx_MarshallProcessRPCStats(callerVersion,
8162 rpc_stat->stats[0].func_total,
8163 rpc_stat->stats, &ptr);
8169 MUTEX_EXIT(&rx_rpc_stats);
8174 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8178 * IN callerVersion - the rpc stat version of the caller
8180 * OUT myVersion - the rpc stat version of this function
8182 * OUT clock_sec - local time seconds
8184 * OUT clock_usec - local time microseconds
8186 * OUT allocSize - the number of bytes allocated to contain stats
8188 * OUT statCount - the number of stats retrieved from the individual
8191 * OUT stats - the actual stats retrieved from the individual peer structures.
8195 * Returns void. If successful, stats will != NULL.
8199 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8200 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8201 size_t * allocSize, afs_uint32 * statCount,
8202 afs_uint32 ** stats)
8212 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8215 * Check to see if stats are enabled
8218 MUTEX_ENTER(&rx_rpc_stats);
8219 if (!rxi_monitor_peerStats) {
8220 MUTEX_EXIT(&rx_rpc_stats);
8224 clock_GetTime(&now);
8225 *clock_sec = now.sec;
8226 *clock_usec = now.usec;
8229 * Allocate the space based upon the caller version
8231 * If the client is at an older version than we are,
8232 * we return the statistic data in the older data format, but
8233 * we still return our version number so the client knows we
8234 * are maintaining more data than it can retrieve.
8237 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8238 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8239 *statCount = rxi_rpc_peer_stat_cnt;
8242 * This can't happen yet, but in the future version changes
8243 * can be handled by adding additional code here
8247 if (space > (size_t) 0) {
8249 ptr = *stats = rxi_Alloc(space);
8252 rx_interface_stat_p rpc_stat, nrpc_stat;
8256 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8258 * We have to fix the offset of rpc_stat since we are
8259 * keeping this structure on two rx_queues. The rx_queue
8260 * package assumes that the rx_queue member is the first
8261 * member of the structure. That is, rx_queue assumes that
8262 * any one item is only on one queue at a time. We are
8263 * breaking that assumption and so we have to do a little
8264 * math to fix our pointers.
8267 fix_offset = (char *)rpc_stat;
8268 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8269 rpc_stat = (rx_interface_stat_p) fix_offset;
8272 * Copy the data based upon the caller version
8274 rx_MarshallProcessRPCStats(callerVersion,
8275 rpc_stat->stats[0].func_total,
8276 rpc_stat->stats, &ptr);
8282 MUTEX_EXIT(&rx_rpc_stats);
8287 * rx_FreeRPCStats - free memory allocated by
8288 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8292 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8293 * rx_RetrievePeerRPCStats
8295 * IN allocSize - the number of bytes in stats.
8303 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8305 rxi_Free(stats, allocSize);
8309 * rx_queryProcessRPCStats - see if process rpc stat collection is
8310 * currently enabled.
8316 * Returns 0 if stats are not enabled != 0 otherwise
8320 rx_queryProcessRPCStats(void)
8323 MUTEX_ENTER(&rx_rpc_stats);
8324 rc = rxi_monitor_processStats;
8325 MUTEX_EXIT(&rx_rpc_stats);
8330 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8336 * Returns 0 if stats are not enabled != 0 otherwise
8340 rx_queryPeerRPCStats(void)
8343 MUTEX_ENTER(&rx_rpc_stats);
8344 rc = rxi_monitor_peerStats;
8345 MUTEX_EXIT(&rx_rpc_stats);
8350 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8360 rx_enableProcessRPCStats(void)
8362 MUTEX_ENTER(&rx_rpc_stats);
8363 rx_enable_stats = 1;
8364 rxi_monitor_processStats = 1;
8365 MUTEX_EXIT(&rx_rpc_stats);
8369 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8379 rx_enablePeerRPCStats(void)
8381 MUTEX_ENTER(&rx_rpc_stats);
8382 rx_enable_stats = 1;
8383 rxi_monitor_peerStats = 1;
8384 MUTEX_EXIT(&rx_rpc_stats);
8388 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8398 rx_disableProcessRPCStats(void)
8400 rx_interface_stat_p rpc_stat, nrpc_stat;
8403 MUTEX_ENTER(&rx_rpc_stats);
8406 * Turn off process statistics and if peer stats is also off, turn
8410 rxi_monitor_processStats = 0;
8411 if (rxi_monitor_peerStats == 0) {
8412 rx_enable_stats = 0;
8415 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8416 unsigned int num_funcs = 0;
8419 queue_Remove(rpc_stat);
8420 num_funcs = rpc_stat->stats[0].func_total;
8422 sizeof(rx_interface_stat_t) +
8423 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8425 rxi_Free(rpc_stat, space);
8426 rxi_rpc_process_stat_cnt -= num_funcs;
8428 MUTEX_EXIT(&rx_rpc_stats);
8432 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8442 rx_disablePeerRPCStats(void)
8444 struct rx_peer **peer_ptr, **peer_end;
8448 * Turn off peer statistics and if process stats is also off, turn
8452 rxi_monitor_peerStats = 0;
8453 if (rxi_monitor_processStats == 0) {
8454 rx_enable_stats = 0;
8457 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8458 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8460 struct rx_peer *peer, *next, *prev;
8462 MUTEX_ENTER(&rx_peerHashTable_lock);
8463 MUTEX_ENTER(&rx_rpc_stats);
8464 for (prev = peer = *peer_ptr; peer; peer = next) {
8466 code = MUTEX_TRYENTER(&peer->peer_lock);
8468 rx_interface_stat_p rpc_stat, nrpc_stat;
8471 if (prev == *peer_ptr) {
8482 MUTEX_EXIT(&rx_peerHashTable_lock);
8485 (&peer->rpcStats, rpc_stat, nrpc_stat,
8486 rx_interface_stat)) {
8487 unsigned int num_funcs = 0;
8490 queue_Remove(&rpc_stat->queue_header);
8491 queue_Remove(&rpc_stat->all_peers);
8492 num_funcs = rpc_stat->stats[0].func_total;
8494 sizeof(rx_interface_stat_t) +
8495 rpc_stat->stats[0].func_total *
8496 sizeof(rx_function_entry_v1_t);
8498 rxi_Free(rpc_stat, space);
8499 rxi_rpc_peer_stat_cnt -= num_funcs;
8501 MUTEX_EXIT(&peer->peer_lock);
8503 MUTEX_ENTER(&rx_peerHashTable_lock);
8513 MUTEX_EXIT(&rx_rpc_stats);
8514 MUTEX_EXIT(&rx_peerHashTable_lock);
8519 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8524 * IN clearFlag - flag indicating which stats to clear
8532 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8534 rx_interface_stat_p rpc_stat, nrpc_stat;
8536 MUTEX_ENTER(&rx_rpc_stats);
8538 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8539 unsigned int num_funcs = 0, i;
8540 num_funcs = rpc_stat->stats[0].func_total;
8541 for (i = 0; i < num_funcs; i++) {
8542 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8543 hzero(rpc_stat->stats[i].invocations);
8545 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8546 hzero(rpc_stat->stats[i].bytes_sent);
8548 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8549 hzero(rpc_stat->stats[i].bytes_rcvd);
8551 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8552 rpc_stat->stats[i].queue_time_sum.sec = 0;
8553 rpc_stat->stats[i].queue_time_sum.usec = 0;
8555 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8556 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8557 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8559 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8560 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8561 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8563 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8564 rpc_stat->stats[i].queue_time_max.sec = 0;
8565 rpc_stat->stats[i].queue_time_max.usec = 0;
8567 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8568 rpc_stat->stats[i].execution_time_sum.sec = 0;
8569 rpc_stat->stats[i].execution_time_sum.usec = 0;
8571 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8572 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8573 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8575 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8576 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8577 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8579 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8580 rpc_stat->stats[i].execution_time_max.sec = 0;
8581 rpc_stat->stats[i].execution_time_max.usec = 0;
8586 MUTEX_EXIT(&rx_rpc_stats);
8590 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8595 * IN clearFlag - flag indicating which stats to clear
8603 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8605 rx_interface_stat_p rpc_stat, nrpc_stat;
8607 MUTEX_ENTER(&rx_rpc_stats);
8609 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8610 unsigned int num_funcs = 0, i;
8613 * We have to fix the offset of rpc_stat since we are
8614 * keeping this structure on two rx_queues. The rx_queue
8615 * package assumes that the rx_queue member is the first
8616 * member of the structure. That is, rx_queue assumes that
8617 * any one item is only on one queue at a time. We are
8618 * breaking that assumption and so we have to do a little
8619 * math to fix our pointers.
8622 fix_offset = (char *)rpc_stat;
8623 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8624 rpc_stat = (rx_interface_stat_p) fix_offset;
8626 num_funcs = rpc_stat->stats[0].func_total;
8627 for (i = 0; i < num_funcs; i++) {
8628 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8629 hzero(rpc_stat->stats[i].invocations);
8631 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8632 hzero(rpc_stat->stats[i].bytes_sent);
8634 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8635 hzero(rpc_stat->stats[i].bytes_rcvd);
8637 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8638 rpc_stat->stats[i].queue_time_sum.sec = 0;
8639 rpc_stat->stats[i].queue_time_sum.usec = 0;
8641 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8642 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8643 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8645 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8646 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8647 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8649 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8650 rpc_stat->stats[i].queue_time_max.sec = 0;
8651 rpc_stat->stats[i].queue_time_max.usec = 0;
8653 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8654 rpc_stat->stats[i].execution_time_sum.sec = 0;
8655 rpc_stat->stats[i].execution_time_sum.usec = 0;
8657 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8658 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8659 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8661 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8662 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8663 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8665 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8666 rpc_stat->stats[i].execution_time_max.sec = 0;
8667 rpc_stat->stats[i].execution_time_max.usec = 0;
8672 MUTEX_EXIT(&rx_rpc_stats);
8676 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8677 * is authorized to enable/disable/clear RX statistics.
8679 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8682 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8684 rxi_rxstat_userok = proc;
8688 rx_RxStatUserOk(struct rx_call *call)
8690 if (!rxi_rxstat_userok)
8692 return rxi_rxstat_userok(call);
8697 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8698 * function in the MSVC runtime DLL (msvcrt.dll).
8700 * Note: the system serializes calls to this function.
8703 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8704 DWORD reason, /* reason function is being called */
8705 LPVOID reserved) /* reserved for future use */
8708 case DLL_PROCESS_ATTACH:
8709 /* library is being attached to a process */
8713 case DLL_PROCESS_DETACH:
8720 #endif /* AFS_NT40_ENV */
8723 int rx_DumpCalls(FILE *outputFile, char *cookie)
8725 #ifdef RXDEBUG_PACKET
8726 #ifdef KDUMP_RX_LOCK
8727 struct rx_call_rx_lock *c;
8734 #define RXDPRINTF sprintf
8735 #define RXDPRINTOUT output
8737 #define RXDPRINTF fprintf
8738 #define RXDPRINTOUT outputFile
8741 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8743 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8746 for (c = rx_allCallsp; c; c = c->allNextp) {
8747 u_short rqc, tqc, iovqc;
8748 struct rx_packet *p, *np;
8750 MUTEX_ENTER(&c->lock);
8751 queue_Count(&c->rq, p, np, rx_packet, rqc);
8752 queue_Count(&c->tq, p, np, rx_packet, tqc);
8753 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8755 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, "
8756 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8757 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8758 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8759 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8760 #ifdef RX_ENABLE_LOCKS
8763 #ifdef RX_REFCOUNT_CHECK
8764 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
8765 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
8768 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,
8769 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
8770 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
8771 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
8772 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
8773 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
8774 #ifdef RX_ENABLE_LOCKS
8775 , (afs_uint32)c->refCount
8777 #ifdef RX_REFCOUNT_CHECK
8778 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
8781 MUTEX_EXIT(&c->lock);
8784 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8787 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
8789 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8791 #endif /* RXDEBUG_PACKET */