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 #include "rx_internal.h"
75 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
76 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
77 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
79 extern afs_int32 afs_termState;
81 #include "sys/lockl.h"
82 #include "sys/lock_def.h"
83 #endif /* AFS_AIX41_ENV */
84 # include "afs/rxgen_consts.h"
86 # include <sys/types.h>
96 # include <afs/afsutil.h>
97 # include <WINNT\afsreg.h>
99 # include <sys/socket.h>
100 # include <sys/file.h>
102 # include <sys/stat.h>
103 # include <netinet/in.h>
104 # include <sys/time.h>
107 # include "rx_user.h"
108 # include "rx_clock.h"
109 # include "rx_queue.h"
110 # include "rx_atomic.h"
111 # include "rx_globals.h"
112 # include "rx_trace.h"
113 # include "rx_internal.h"
114 # include "rx_stats.h"
115 # include <afs/rxgen_consts.h>
119 #ifdef AFS_PTHREAD_ENV
121 int (*registerProgram) (pid_t, char *) = 0;
122 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
125 int (*registerProgram) (PROCESS, char *) = 0;
126 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
130 /* Local static routines */
131 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
132 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
133 struct rx_peer *, struct clock *);
135 #ifdef RX_ENABLE_LOCKS
136 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
139 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
141 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
142 rx_atomic_t rxi_start_in_error;
144 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
147 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
148 * currently allocated within rx. This number is used to allocate the
149 * memory required to return the statistics when queried.
150 * Protected by the rx_rpc_stats mutex.
153 static unsigned int rxi_rpc_peer_stat_cnt;
156 * rxi_rpc_process_stat_cnt counts the total number of local process stat
157 * structures currently allocated within rx. The number is used to allocate
158 * the memory required to return the statistics when queried.
159 * Protected by the rx_rpc_stats mutex.
162 static unsigned int rxi_rpc_process_stat_cnt;
164 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
165 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
167 #if !defined(offsetof)
168 #include <stddef.h> /* for definition of offsetof() */
171 #ifdef RX_ENABLE_LOCKS
172 afs_kmutex_t rx_atomic_mutex;
175 #ifdef AFS_PTHREAD_ENV
178 * Use procedural initialization of mutexes/condition variables
182 extern afs_kmutex_t rx_quota_mutex;
183 extern afs_kmutex_t rx_pthread_mutex;
184 extern afs_kmutex_t rx_packets_mutex;
185 extern afs_kmutex_t rx_refcnt_mutex;
186 extern afs_kmutex_t des_init_mutex;
187 extern afs_kmutex_t des_random_mutex;
188 extern afs_kmutex_t rx_clock_mutex;
189 extern afs_kmutex_t rxi_connCacheMutex;
190 extern afs_kmutex_t rx_event_mutex;
191 extern afs_kmutex_t osi_malloc_mutex;
192 extern afs_kmutex_t event_handler_mutex;
193 extern afs_kmutex_t listener_mutex;
194 extern afs_kmutex_t rx_if_init_mutex;
195 extern afs_kmutex_t rx_if_mutex;
196 extern afs_kmutex_t rxkad_client_uid_mutex;
197 extern afs_kmutex_t rxkad_random_mutex;
199 extern afs_kcondvar_t rx_event_handler_cond;
200 extern afs_kcondvar_t rx_listener_cond;
202 static afs_kmutex_t epoch_mutex;
203 static afs_kmutex_t rx_init_mutex;
204 static afs_kmutex_t rx_debug_mutex;
205 static afs_kmutex_t rx_rpc_stats;
208 rxi_InitPthread(void)
210 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
211 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
212 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
213 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
214 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
215 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
216 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
217 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
218 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
219 MUTEX_INIT(&rx_event_mutex, "event", MUTEX_DEFAULT, 0);
220 MUTEX_INIT(&des_init_mutex, "des", MUTEX_DEFAULT, 0);
221 MUTEX_INIT(&des_random_mutex, "random", MUTEX_DEFAULT, 0);
222 MUTEX_INIT(&osi_malloc_mutex, "malloc", MUTEX_DEFAULT, 0);
223 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
224 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
225 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
226 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
227 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
228 MUTEX_INIT(&rxkad_client_uid_mutex, "uid", MUTEX_DEFAULT, 0);
229 MUTEX_INIT(&rxkad_random_mutex, "rxkad random", MUTEX_DEFAULT, 0);
230 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
232 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
233 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
235 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
236 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
238 rxkad_global_stats_init();
240 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
241 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
242 #ifdef RX_ENABLE_LOCKS
245 #endif /* RX_LOCKS_DB */
246 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
247 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
249 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
251 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
253 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
255 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
256 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
257 #endif /* RX_ENABLE_LOCKS */
260 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
261 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
263 * The rx_stats_mutex mutex protects the following global variables:
264 * rxi_lowConnRefCount
265 * rxi_lowPeerRefCount
274 * The rx_quota_mutex mutex protects the following global variables:
282 * The rx_freePktQ_lock protects the following global variables:
287 * The rx_packets_mutex mutex protects the following global variables:
295 * The rx_pthread_mutex mutex protects the following global variables:
296 * rxi_fcfs_thread_num
299 #define INIT_PTHREAD_LOCKS
303 /* Variables for handling the minProcs implementation. availProcs gives the
304 * number of threads available in the pool at this moment (not counting dudes
305 * executing right now). totalMin gives the total number of procs required
306 * for handling all minProcs requests. minDeficit is a dynamic variable
307 * tracking the # of procs required to satisfy all of the remaining minProcs
309 * For fine grain locking to work, the quota check and the reservation of
310 * a server thread has to come while rxi_availProcs and rxi_minDeficit
311 * are locked. To this end, the code has been modified under #ifdef
312 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
313 * same time. A new function, ReturnToServerPool() returns the allocation.
315 * A call can be on several queue's (but only one at a time). When
316 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
317 * that no one else is touching the queue. To this end, we store the address
318 * of the queue lock in the call structure (under the call lock) when we
319 * put the call on a queue, and we clear the call_queue_lock when the
320 * call is removed from a queue (once the call lock has been obtained).
321 * This allows rxi_ResetCall to safely synchronize with others wishing
322 * to manipulate the queue.
325 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
326 static afs_kmutex_t rx_rpc_stats;
327 void rxi_StartUnlocked(struct rxevent *event, void *call,
328 void *arg1, int istack);
331 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
332 ** pretty good that the next packet coming in is from the same connection
333 ** as the last packet, since we're send multiple packets in a transmit window.
335 struct rx_connection *rxLastConn = 0;
337 #ifdef RX_ENABLE_LOCKS
338 /* The locking hierarchy for rx fine grain locking is composed of these
341 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
342 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
343 * call->lock - locks call data fields.
344 * These are independent of each other:
345 * rx_freeCallQueue_lock
350 * serverQueueEntry->lock
351 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
353 * peer->lock - locks peer data fields.
354 * conn_data_lock - that more than one thread is not updating a conn data
355 * field at the same time.
366 * Do we need a lock to protect the peer field in the conn structure?
367 * conn->peer was previously a constant for all intents and so has no
368 * lock protecting this field. The multihomed client delta introduced
369 * a RX code change : change the peer field in the connection structure
370 * to that remote interface from which the last packet for this
371 * connection was sent out. This may become an issue if further changes
374 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
375 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
377 /* rxdb_fileID is used to identify the lock location, along with line#. */
378 static int rxdb_fileID = RXDB_FILE_RX;
379 #endif /* RX_LOCKS_DB */
380 #else /* RX_ENABLE_LOCKS */
381 #define SET_CALL_QUEUE_LOCK(C, L)
382 #define CLEAR_CALL_QUEUE_LOCK(C)
383 #endif /* RX_ENABLE_LOCKS */
384 struct rx_serverQueueEntry *rx_waitForPacket = 0;
385 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
387 /* ------------Exported Interfaces------------- */
389 /* This function allows rxkad to set the epoch to a suitably random number
390 * which rx_NewConnection will use in the future. The principle purpose is to
391 * get rxnull connections to use the same epoch as the rxkad connections do, at
392 * least once the first rxkad connection is established. This is important now
393 * that the host/port addresses aren't used in FindConnection: the uniqueness
394 * of epoch/cid matters and the start time won't do. */
396 #ifdef AFS_PTHREAD_ENV
398 * This mutex protects the following global variables:
402 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
403 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
407 #endif /* AFS_PTHREAD_ENV */
410 rx_SetEpoch(afs_uint32 epoch)
417 /* Initialize rx. A port number may be mentioned, in which case this
418 * becomes the default port number for any service installed later.
419 * If 0 is provided for the port number, a random port will be chosen
420 * by the kernel. Whether this will ever overlap anything in
421 * /etc/services is anybody's guess... Returns 0 on success, -1 on
426 int rxinit_status = 1;
427 #ifdef AFS_PTHREAD_ENV
429 * This mutex protects the following global variables:
433 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
434 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
437 #define UNLOCK_RX_INIT
441 rx_InitHost(u_int host, u_int port)
448 char *htable, *ptable;
455 if (rxinit_status == 0) {
456 tmp_status = rxinit_status;
458 return tmp_status; /* Already started; return previous error code. */
464 if (afs_winsockInit() < 0)
470 * Initialize anything necessary to provide a non-premptive threading
473 rxi_InitializeThreadSupport();
476 /* Allocate and initialize a socket for client and perhaps server
479 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
480 if (rx_socket == OSI_NULLSOCKET) {
484 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
487 #endif /* RX_LOCKS_DB */
488 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
489 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
490 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
491 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
492 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
493 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
494 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
495 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
496 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
498 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
500 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
502 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
504 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
505 #if defined(AFS_HPUX110_ENV)
507 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
508 #endif /* AFS_HPUX110_ENV */
509 #endif /* RX_ENABLE_LOCKS && KERNEL */
512 rx_connDeadTime = 12;
513 rx_tranquil = 0; /* reset flag */
514 rxi_ResetStatistics();
516 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
517 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
518 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
519 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
520 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
521 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
523 /* Malloc up a bunch of packets & buffers */
525 queue_Init(&rx_freePacketQueue);
526 rxi_NeedMorePackets = FALSE;
527 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
529 /* enforce a minimum number of allocated packets */
530 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
531 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
533 /* allocate the initial free packet pool */
534 #ifdef RX_ENABLE_TSFPQ
535 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
536 #else /* RX_ENABLE_TSFPQ */
537 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
538 #endif /* RX_ENABLE_TSFPQ */
545 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
546 tv.tv_sec = clock_now.sec;
547 tv.tv_usec = clock_now.usec;
548 srand((unsigned int)tv.tv_usec);
555 #if defined(KERNEL) && !defined(UKERNEL)
556 /* Really, this should never happen in a real kernel */
559 struct sockaddr_in addr;
561 int addrlen = sizeof(addr);
563 socklen_t addrlen = sizeof(addr);
565 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
569 rx_port = addr.sin_port;
572 rx_stats.minRtt.sec = 9999999;
574 rx_SetEpoch(tv.tv_sec | 0x80000000);
576 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
577 * will provide a randomer value. */
579 MUTEX_ENTER(&rx_quota_mutex);
580 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
581 MUTEX_EXIT(&rx_quota_mutex);
582 /* *Slightly* random start time for the cid. This is just to help
583 * out with the hashing function at the peer */
584 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
585 rx_connHashTable = (struct rx_connection **)htable;
586 rx_peerHashTable = (struct rx_peer **)ptable;
588 rx_lastAckDelay.sec = 0;
589 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
590 rx_hardAckDelay.sec = 0;
591 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
592 rx_softAckDelay.sec = 0;
593 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
595 rxevent_Init(20, rxi_ReScheduleEvents);
597 /* Initialize various global queues */
598 queue_Init(&rx_idleServerQueue);
599 queue_Init(&rx_incomingCallQueue);
600 queue_Init(&rx_freeCallQueue);
602 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
603 /* Initialize our list of usable IP addresses. */
607 /* Start listener process (exact function is dependent on the
608 * implementation environment--kernel or user space) */
612 tmp_status = rxinit_status = 0;
620 return rx_InitHost(htonl(INADDR_ANY), port);
623 /* called with unincremented nRequestsRunning to see if it is OK to start
624 * a new thread in this service. Could be "no" for two reasons: over the
625 * max quota, or would prevent others from reaching their min quota.
627 #ifdef RX_ENABLE_LOCKS
628 /* This verion of QuotaOK reserves quota if it's ok while the
629 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
632 QuotaOK(struct rx_service *aservice)
634 /* check if over max quota */
635 if (aservice->nRequestsRunning >= aservice->maxProcs) {
639 /* under min quota, we're OK */
640 /* otherwise, can use only if there are enough to allow everyone
641 * to go to their min quota after this guy starts.
644 MUTEX_ENTER(&rx_quota_mutex);
645 if ((aservice->nRequestsRunning < aservice->minProcs)
646 || (rxi_availProcs > rxi_minDeficit)) {
647 aservice->nRequestsRunning++;
648 /* just started call in minProcs pool, need fewer to maintain
650 if (aservice->nRequestsRunning <= aservice->minProcs)
653 MUTEX_EXIT(&rx_quota_mutex);
656 MUTEX_EXIT(&rx_quota_mutex);
662 ReturnToServerPool(struct rx_service *aservice)
664 aservice->nRequestsRunning--;
665 MUTEX_ENTER(&rx_quota_mutex);
666 if (aservice->nRequestsRunning < aservice->minProcs)
669 MUTEX_EXIT(&rx_quota_mutex);
672 #else /* RX_ENABLE_LOCKS */
674 QuotaOK(struct rx_service *aservice)
677 /* under min quota, we're OK */
678 if (aservice->nRequestsRunning < aservice->minProcs)
681 /* check if over max quota */
682 if (aservice->nRequestsRunning >= aservice->maxProcs)
685 /* otherwise, can use only if there are enough to allow everyone
686 * to go to their min quota after this guy starts.
688 MUTEX_ENTER(&rx_quota_mutex);
689 if (rxi_availProcs > rxi_minDeficit)
691 MUTEX_EXIT(&rx_quota_mutex);
694 #endif /* RX_ENABLE_LOCKS */
697 /* Called by rx_StartServer to start up lwp's to service calls.
698 NExistingProcs gives the number of procs already existing, and which
699 therefore needn't be created. */
701 rxi_StartServerProcs(int nExistingProcs)
703 struct rx_service *service;
708 /* For each service, reserve N processes, where N is the "minimum"
709 * number of processes that MUST be able to execute a request in parallel,
710 * at any time, for that process. Also compute the maximum difference
711 * between any service's maximum number of processes that can run
712 * (i.e. the maximum number that ever will be run, and a guarantee
713 * that this number will run if other services aren't running), and its
714 * minimum number. The result is the extra number of processes that
715 * we need in order to provide the latter guarantee */
716 for (i = 0; i < RX_MAX_SERVICES; i++) {
718 service = rx_services[i];
719 if (service == (struct rx_service *)0)
721 nProcs += service->minProcs;
722 diff = service->maxProcs - service->minProcs;
726 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
727 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
728 for (i = 0; i < nProcs; i++) {
729 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
735 /* This routine is only required on Windows */
737 rx_StartClientThread(void)
739 #ifdef AFS_PTHREAD_ENV
741 pid = pthread_self();
742 #endif /* AFS_PTHREAD_ENV */
744 #endif /* AFS_NT40_ENV */
746 /* This routine must be called if any services are exported. If the
747 * donateMe flag is set, the calling process is donated to the server
750 rx_StartServer(int donateMe)
752 struct rx_service *service;
758 /* Start server processes, if necessary (exact function is dependent
759 * on the implementation environment--kernel or user space). DonateMe
760 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
761 * case, one less new proc will be created rx_StartServerProcs.
763 rxi_StartServerProcs(donateMe);
765 /* count up the # of threads in minProcs, and add set the min deficit to
766 * be that value, too.
768 for (i = 0; i < RX_MAX_SERVICES; i++) {
769 service = rx_services[i];
770 if (service == (struct rx_service *)0)
772 MUTEX_ENTER(&rx_quota_mutex);
773 rxi_totalMin += service->minProcs;
774 /* below works even if a thread is running, since minDeficit would
775 * still have been decremented and later re-incremented.
777 rxi_minDeficit += service->minProcs;
778 MUTEX_EXIT(&rx_quota_mutex);
781 /* Turn on reaping of idle server connections */
782 rxi_ReapConnections(NULL, NULL, NULL);
791 #ifdef AFS_PTHREAD_ENV
793 pid = afs_pointer_to_int(pthread_self());
794 #else /* AFS_PTHREAD_ENV */
796 LWP_CurrentProcess(&pid);
797 #endif /* AFS_PTHREAD_ENV */
799 sprintf(name, "srv_%d", ++nProcs);
801 (*registerProgram) (pid, name);
803 #endif /* AFS_NT40_ENV */
804 rx_ServerProc(NULL); /* Never returns */
806 #ifdef RX_ENABLE_TSFPQ
807 /* no use leaving packets around in this thread's local queue if
808 * it isn't getting donated to the server thread pool.
810 rxi_FlushLocalPacketsTSFPQ();
811 #endif /* RX_ENABLE_TSFPQ */
815 /* Create a new client connection to the specified service, using the
816 * specified security object to implement the security model for this
818 struct rx_connection *
819 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
820 struct rx_securityClass *securityObject,
821 int serviceSecurityIndex)
825 struct rx_connection *conn;
830 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
831 "serviceSecurityIndex %d)\n",
832 ntohl(shost), ntohs(sport), sservice, securityObject,
833 serviceSecurityIndex));
835 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
836 * the case of kmem_alloc? */
837 conn = rxi_AllocConnection();
838 #ifdef RX_ENABLE_LOCKS
839 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
840 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
841 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
844 MUTEX_ENTER(&rx_connHashTable_lock);
845 cid = (rx_nextCid += RX_MAXCALLS);
846 conn->type = RX_CLIENT_CONNECTION;
848 conn->epoch = rx_epoch;
849 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
850 conn->serviceId = sservice;
851 conn->securityObject = securityObject;
852 conn->securityData = (void *) 0;
853 conn->securityIndex = serviceSecurityIndex;
854 rx_SetConnDeadTime(conn, rx_connDeadTime);
855 rx_SetConnSecondsUntilNatPing(conn, 0);
856 conn->ackRate = RX_FAST_ACK_RATE;
858 conn->specific = NULL;
859 conn->challengeEvent = NULL;
860 conn->delayedAbortEvent = NULL;
861 conn->abortCount = 0;
863 for (i = 0; i < RX_MAXCALLS; i++) {
864 conn->twind[i] = rx_initSendWindow;
865 conn->rwind[i] = rx_initReceiveWindow;
868 RXS_NewConnection(securityObject, conn);
870 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
872 conn->refCount++; /* no lock required since only this thread knows... */
873 conn->next = rx_connHashTable[hashindex];
874 rx_connHashTable[hashindex] = conn;
876 rx_atomic_inc(&rx_stats.nClientConns);
877 MUTEX_EXIT(&rx_connHashTable_lock);
883 * Ensure a connection's timeout values are valid.
885 * @param[in] conn The connection to check
887 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
888 * unless idleDeadTime and/or hardDeadTime are not set
892 rxi_CheckConnTimeouts(struct rx_connection *conn)
894 /* a connection's timeouts must have the relationship
895 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
896 * total loss of network to a peer may cause an idle timeout instead of a
897 * dead timeout, simply because the idle timeout gets hit first. Also set
898 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
899 /* this logic is slightly complicated by the fact that
900 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
902 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
903 if (conn->idleDeadTime) {
904 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
906 if (conn->hardDeadTime) {
907 if (conn->idleDeadTime) {
908 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
910 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
916 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
918 /* The idea is to set the dead time to a value that allows several
919 * keepalives to be dropped without timing out the connection. */
920 conn->secondsUntilDead = seconds;
921 rxi_CheckConnTimeouts(conn);
922 conn->secondsUntilPing = conn->secondsUntilDead / 6;
926 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
928 conn->hardDeadTime = seconds;
929 rxi_CheckConnTimeouts(conn);
933 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
935 conn->idleDeadTime = seconds;
936 rxi_CheckConnTimeouts(conn);
939 int rxi_lowPeerRefCount = 0;
940 int rxi_lowConnRefCount = 0;
943 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
944 * NOTE: must not be called with rx_connHashTable_lock held.
947 rxi_CleanupConnection(struct rx_connection *conn)
949 /* Notify the service exporter, if requested, that this connection
950 * is being destroyed */
951 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
952 (*conn->service->destroyConnProc) (conn);
954 /* Notify the security module that this connection is being destroyed */
955 RXS_DestroyConnection(conn->securityObject, conn);
957 /* If this is the last connection using the rx_peer struct, set its
958 * idle time to now. rxi_ReapConnections will reap it if it's still
959 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
961 MUTEX_ENTER(&rx_peerHashTable_lock);
962 if (conn->peer->refCount < 2) {
963 conn->peer->idleWhen = clock_Sec();
964 if (conn->peer->refCount < 1) {
965 conn->peer->refCount = 1;
966 if (rx_stats_active) {
967 MUTEX_ENTER(&rx_stats_mutex);
968 rxi_lowPeerRefCount++;
969 MUTEX_EXIT(&rx_stats_mutex);
973 conn->peer->refCount--;
974 MUTEX_EXIT(&rx_peerHashTable_lock);
978 if (conn->type == RX_SERVER_CONNECTION)
979 rx_atomic_dec(&rx_stats.nServerConns);
981 rx_atomic_dec(&rx_stats.nClientConns);
984 if (conn->specific) {
986 for (i = 0; i < conn->nSpecific; i++) {
987 if (conn->specific[i] && rxi_keyCreate_destructor[i])
988 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
989 conn->specific[i] = NULL;
991 free(conn->specific);
993 conn->specific = NULL;
997 MUTEX_DESTROY(&conn->conn_call_lock);
998 MUTEX_DESTROY(&conn->conn_data_lock);
999 CV_DESTROY(&conn->conn_call_cv);
1001 rxi_FreeConnection(conn);
1004 /* Destroy the specified connection */
1006 rxi_DestroyConnection(struct rx_connection *conn)
1008 MUTEX_ENTER(&rx_connHashTable_lock);
1009 rxi_DestroyConnectionNoLock(conn);
1010 /* conn should be at the head of the cleanup list */
1011 if (conn == rx_connCleanup_list) {
1012 rx_connCleanup_list = rx_connCleanup_list->next;
1013 MUTEX_EXIT(&rx_connHashTable_lock);
1014 rxi_CleanupConnection(conn);
1016 #ifdef RX_ENABLE_LOCKS
1018 MUTEX_EXIT(&rx_connHashTable_lock);
1020 #endif /* RX_ENABLE_LOCKS */
1024 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1026 struct rx_connection **conn_ptr;
1028 struct rx_packet *packet;
1035 MUTEX_ENTER(&conn->conn_data_lock);
1036 MUTEX_ENTER(&rx_refcnt_mutex);
1037 if (conn->refCount > 0)
1040 if (rx_stats_active) {
1041 MUTEX_ENTER(&rx_stats_mutex);
1042 rxi_lowConnRefCount++;
1043 MUTEX_EXIT(&rx_stats_mutex);
1047 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1048 /* Busy; wait till the last guy before proceeding */
1049 MUTEX_EXIT(&rx_refcnt_mutex);
1050 MUTEX_EXIT(&conn->conn_data_lock);
1055 /* If the client previously called rx_NewCall, but it is still
1056 * waiting, treat this as a running call, and wait to destroy the
1057 * connection later when the call completes. */
1058 if ((conn->type == RX_CLIENT_CONNECTION)
1059 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1060 conn->flags |= RX_CONN_DESTROY_ME;
1061 MUTEX_EXIT(&conn->conn_data_lock);
1065 MUTEX_EXIT(&rx_refcnt_mutex);
1066 MUTEX_EXIT(&conn->conn_data_lock);
1068 /* Check for extant references to this connection */
1069 for (i = 0; i < RX_MAXCALLS; i++) {
1070 struct rx_call *call = conn->call[i];
1073 if (conn->type == RX_CLIENT_CONNECTION) {
1074 MUTEX_ENTER(&call->lock);
1075 if (call->delayedAckEvent) {
1076 /* Push the final acknowledgment out now--there
1077 * won't be a subsequent call to acknowledge the
1078 * last reply packets */
1079 rxevent_Cancel(call->delayedAckEvent, call,
1080 RX_CALL_REFCOUNT_DELAY);
1081 if (call->state == RX_STATE_PRECALL
1082 || call->state == RX_STATE_ACTIVE) {
1083 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1085 rxi_AckAll(NULL, call, 0);
1088 MUTEX_EXIT(&call->lock);
1092 #ifdef RX_ENABLE_LOCKS
1094 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1095 MUTEX_EXIT(&conn->conn_data_lock);
1097 /* Someone is accessing a packet right now. */
1101 #endif /* RX_ENABLE_LOCKS */
1104 /* Don't destroy the connection if there are any call
1105 * structures still in use */
1106 MUTEX_ENTER(&conn->conn_data_lock);
1107 conn->flags |= RX_CONN_DESTROY_ME;
1108 MUTEX_EXIT(&conn->conn_data_lock);
1113 if (conn->natKeepAliveEvent) {
1114 rxi_NatKeepAliveOff(conn);
1117 if (conn->delayedAbortEvent) {
1118 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1119 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1121 MUTEX_ENTER(&conn->conn_data_lock);
1122 rxi_SendConnectionAbort(conn, packet, 0, 1);
1123 MUTEX_EXIT(&conn->conn_data_lock);
1124 rxi_FreePacket(packet);
1128 /* Remove from connection hash table before proceeding */
1130 &rx_connHashTable[CONN_HASH
1131 (peer->host, peer->port, conn->cid, conn->epoch,
1133 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1134 if (*conn_ptr == conn) {
1135 *conn_ptr = conn->next;
1139 /* if the conn that we are destroying was the last connection, then we
1140 * clear rxLastConn as well */
1141 if (rxLastConn == conn)
1144 /* Make sure the connection is completely reset before deleting it. */
1145 /* get rid of pending events that could zap us later */
1146 if (conn->challengeEvent)
1147 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1148 if (conn->checkReachEvent)
1149 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1150 if (conn->natKeepAliveEvent)
1151 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
1153 /* Add the connection to the list of destroyed connections that
1154 * need to be cleaned up. This is necessary to avoid deadlocks
1155 * in the routines we call to inform others that this connection is
1156 * being destroyed. */
1157 conn->next = rx_connCleanup_list;
1158 rx_connCleanup_list = conn;
1161 /* Externally available version */
1163 rx_DestroyConnection(struct rx_connection *conn)
1168 rxi_DestroyConnection(conn);
1173 rx_GetConnection(struct rx_connection *conn)
1178 MUTEX_ENTER(&rx_refcnt_mutex);
1180 MUTEX_EXIT(&rx_refcnt_mutex);
1184 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1185 /* Wait for the transmit queue to no longer be busy.
1186 * requires the call->lock to be held */
1188 rxi_WaitforTQBusy(struct rx_call *call) {
1189 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1190 call->flags |= RX_CALL_TQ_WAIT;
1192 #ifdef RX_ENABLE_LOCKS
1193 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1194 CV_WAIT(&call->cv_tq, &call->lock);
1195 #else /* RX_ENABLE_LOCKS */
1196 osi_rxSleep(&call->tq);
1197 #endif /* RX_ENABLE_LOCKS */
1199 if (call->tqWaiters == 0) {
1200 call->flags &= ~RX_CALL_TQ_WAIT;
1207 rxi_WakeUpTransmitQueue(struct rx_call *call)
1209 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1210 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1211 call, call->tqWaiters, call->flags));
1212 #ifdef RX_ENABLE_LOCKS
1213 osirx_AssertMine(&call->lock, "rxi_Start start");
1214 CV_BROADCAST(&call->cv_tq);
1215 #else /* RX_ENABLE_LOCKS */
1216 osi_rxWakeup(&call->tq);
1217 #endif /* RX_ENABLE_LOCKS */
1221 /* Start a new rx remote procedure call, on the specified connection.
1222 * If wait is set to 1, wait for a free call channel; otherwise return
1223 * 0. Maxtime gives the maximum number of seconds this call may take,
1224 * after rx_NewCall returns. After this time interval, a call to any
1225 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1226 * For fine grain locking, we hold the conn_call_lock in order to
1227 * to ensure that we don't get signalle after we found a call in an active
1228 * state and before we go to sleep.
1231 rx_NewCall(struct rx_connection *conn)
1234 struct rx_call *call;
1235 struct clock queueTime;
1239 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1242 clock_GetTime(&queueTime);
1244 * Check if there are others waiting for a new call.
1245 * If so, let them go first to avoid starving them.
1246 * This is a fairly simple scheme, and might not be
1247 * a complete solution for large numbers of waiters.
1249 * makeCallWaiters keeps track of the number of
1250 * threads waiting to make calls and the
1251 * RX_CONN_MAKECALL_WAITING flag bit is used to
1252 * indicate that there are indeed calls waiting.
1253 * The flag is set when the waiter is incremented.
1254 * It is only cleared when makeCallWaiters is 0.
1255 * This prevents us from accidently destroying the
1256 * connection while it is potentially about to be used.
1258 MUTEX_ENTER(&conn->conn_call_lock);
1259 MUTEX_ENTER(&conn->conn_data_lock);
1260 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1261 conn->flags |= RX_CONN_MAKECALL_WAITING;
1262 conn->makeCallWaiters++;
1263 MUTEX_EXIT(&conn->conn_data_lock);
1265 #ifdef RX_ENABLE_LOCKS
1266 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1270 MUTEX_ENTER(&conn->conn_data_lock);
1271 conn->makeCallWaiters--;
1272 if (conn->makeCallWaiters == 0)
1273 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1276 /* We are now the active thread in rx_NewCall */
1277 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1278 MUTEX_EXIT(&conn->conn_data_lock);
1283 for (i = 0; i < RX_MAXCALLS; i++) {
1284 call = conn->call[i];
1286 if (call->state == RX_STATE_DALLY) {
1287 MUTEX_ENTER(&call->lock);
1288 if (call->state == RX_STATE_DALLY) {
1290 * We are setting the state to RX_STATE_RESET to
1291 * ensure that no one else will attempt to use this
1292 * call once we drop the conn->conn_call_lock and
1293 * call->lock. We must drop the conn->conn_call_lock
1294 * before calling rxi_ResetCall because the process
1295 * of clearing the transmit queue can block for an
1296 * extended period of time. If we block while holding
1297 * the conn->conn_call_lock, then all rx_EndCall
1298 * processing will block as well. This has a detrimental
1299 * effect on overall system performance.
1301 call->state = RX_STATE_RESET;
1302 MUTEX_EXIT(&conn->conn_call_lock);
1303 MUTEX_ENTER(&rx_refcnt_mutex);
1304 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1305 MUTEX_EXIT(&rx_refcnt_mutex);
1306 rxi_ResetCall(call, 0);
1307 (*call->callNumber)++;
1308 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1312 * If we failed to be able to safely obtain the
1313 * conn->conn_call_lock we will have to drop the
1314 * call->lock to avoid a deadlock. When the call->lock
1315 * is released the state of the call can change. If it
1316 * is no longer RX_STATE_RESET then some other thread is
1319 MUTEX_EXIT(&call->lock);
1320 MUTEX_ENTER(&conn->conn_call_lock);
1321 MUTEX_ENTER(&call->lock);
1323 if (call->state == RX_STATE_RESET)
1327 * If we get here it means that after dropping
1328 * the conn->conn_call_lock and call->lock that
1329 * the call is no longer ours. If we can't find
1330 * a free call in the remaining slots we should
1331 * not go immediately to RX_CONN_MAKECALL_WAITING
1332 * because by dropping the conn->conn_call_lock
1333 * we have given up synchronization with rx_EndCall.
1334 * Instead, cycle through one more time to see if
1335 * we can find a call that can call our own.
1337 MUTEX_ENTER(&rx_refcnt_mutex);
1338 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1339 MUTEX_EXIT(&rx_refcnt_mutex);
1342 MUTEX_EXIT(&call->lock);
1345 /* rxi_NewCall returns with mutex locked */
1346 call = rxi_NewCall(conn, i);
1347 MUTEX_ENTER(&rx_refcnt_mutex);
1348 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1349 MUTEX_EXIT(&rx_refcnt_mutex);
1353 if (i < RX_MAXCALLS) {
1359 MUTEX_ENTER(&conn->conn_data_lock);
1360 conn->flags |= RX_CONN_MAKECALL_WAITING;
1361 conn->makeCallWaiters++;
1362 MUTEX_EXIT(&conn->conn_data_lock);
1364 #ifdef RX_ENABLE_LOCKS
1365 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1369 MUTEX_ENTER(&conn->conn_data_lock);
1370 conn->makeCallWaiters--;
1371 if (conn->makeCallWaiters == 0)
1372 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1373 MUTEX_EXIT(&conn->conn_data_lock);
1375 /* Client is initially in send mode */
1376 call->state = RX_STATE_ACTIVE;
1377 call->error = conn->error;
1379 call->mode = RX_MODE_ERROR;
1381 call->mode = RX_MODE_SENDING;
1383 /* remember start time for call in case we have hard dead time limit */
1384 call->queueTime = queueTime;
1385 clock_GetTime(&call->startTime);
1386 hzero(call->bytesSent);
1387 hzero(call->bytesRcvd);
1389 /* Turn on busy protocol. */
1390 rxi_KeepAliveOn(call);
1392 /* Attempt MTU discovery */
1393 rxi_GrowMTUOn(call);
1396 * We are no longer the active thread in rx_NewCall
1398 MUTEX_ENTER(&conn->conn_data_lock);
1399 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1400 MUTEX_EXIT(&conn->conn_data_lock);
1403 * Wake up anyone else who might be giving us a chance to
1404 * run (see code above that avoids resource starvation).
1406 #ifdef RX_ENABLE_LOCKS
1407 CV_BROADCAST(&conn->conn_call_cv);
1411 MUTEX_EXIT(&conn->conn_call_lock);
1413 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1414 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1415 osi_Panic("rx_NewCall call about to be used without an empty tq");
1417 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1419 MUTEX_EXIT(&call->lock);
1422 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1427 rxi_HasActiveCalls(struct rx_connection *aconn)
1430 struct rx_call *tcall;
1434 for (i = 0; i < RX_MAXCALLS; i++) {
1435 if ((tcall = aconn->call[i])) {
1436 if ((tcall->state == RX_STATE_ACTIVE)
1437 || (tcall->state == RX_STATE_PRECALL)) {
1448 rxi_GetCallNumberVector(struct rx_connection *aconn,
1449 afs_int32 * aint32s)
1452 struct rx_call *tcall;
1456 for (i = 0; i < RX_MAXCALLS; i++) {
1457 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1458 aint32s[i] = aconn->callNumber[i] + 1;
1460 aint32s[i] = aconn->callNumber[i];
1467 rxi_SetCallNumberVector(struct rx_connection *aconn,
1468 afs_int32 * aint32s)
1471 struct rx_call *tcall;
1475 for (i = 0; i < RX_MAXCALLS; i++) {
1476 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1477 aconn->callNumber[i] = aint32s[i] - 1;
1479 aconn->callNumber[i] = aint32s[i];
1485 /* Advertise a new service. A service is named locally by a UDP port
1486 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1489 char *serviceName; Name for identification purposes (e.g. the
1490 service name might be used for probing for
1493 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1494 char *serviceName, struct rx_securityClass **securityObjects,
1495 int nSecurityObjects,
1496 afs_int32(*serviceProc) (struct rx_call * acall))
1498 osi_socket socket = OSI_NULLSOCKET;
1499 struct rx_service *tservice;
1505 if (serviceId == 0) {
1507 "rx_NewService: service id for service %s is not non-zero.\n",
1514 "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",
1522 tservice = rxi_AllocService();
1525 #ifdef RX_ENABLE_LOCKS
1526 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1529 for (i = 0; i < RX_MAX_SERVICES; i++) {
1530 struct rx_service *service = rx_services[i];
1532 if (port == service->servicePort && host == service->serviceHost) {
1533 if (service->serviceId == serviceId) {
1534 /* The identical service has already been
1535 * installed; if the caller was intending to
1536 * change the security classes used by this
1537 * service, he/she loses. */
1539 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1540 serviceName, serviceId, service->serviceName);
1542 rxi_FreeService(tservice);
1545 /* Different service, same port: re-use the socket
1546 * which is bound to the same port */
1547 socket = service->socket;
1550 if (socket == OSI_NULLSOCKET) {
1551 /* If we don't already have a socket (from another
1552 * service on same port) get a new one */
1553 socket = rxi_GetHostUDPSocket(host, port);
1554 if (socket == OSI_NULLSOCKET) {
1556 rxi_FreeService(tservice);
1561 service->socket = socket;
1562 service->serviceHost = host;
1563 service->servicePort = port;
1564 service->serviceId = serviceId;
1565 service->serviceName = serviceName;
1566 service->nSecurityObjects = nSecurityObjects;
1567 service->securityObjects = securityObjects;
1568 service->minProcs = 0;
1569 service->maxProcs = 1;
1570 service->idleDeadTime = 60;
1571 service->idleDeadErr = 0;
1572 service->connDeadTime = rx_connDeadTime;
1573 service->executeRequestProc = serviceProc;
1574 service->checkReach = 0;
1575 service->nSpecific = 0;
1576 service->specific = NULL;
1577 rx_services[i] = service; /* not visible until now */
1583 rxi_FreeService(tservice);
1584 (osi_Msg "rx_NewService: cannot support > %d services\n",
1589 /* Set configuration options for all of a service's security objects */
1592 rx_SetSecurityConfiguration(struct rx_service *service,
1593 rx_securityConfigVariables type,
1597 for (i = 0; i<service->nSecurityObjects; i++) {
1598 if (service->securityObjects[i]) {
1599 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1607 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1608 struct rx_securityClass **securityObjects, int nSecurityObjects,
1609 afs_int32(*serviceProc) (struct rx_call * acall))
1611 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1614 /* Generic request processing loop. This routine should be called
1615 * by the implementation dependent rx_ServerProc. If socketp is
1616 * non-null, it will be set to the file descriptor that this thread
1617 * is now listening on. If socketp is null, this routine will never
1620 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1622 struct rx_call *call;
1624 struct rx_service *tservice = NULL;
1631 call = rx_GetCall(threadID, tservice, socketp);
1632 if (socketp && *socketp != OSI_NULLSOCKET) {
1633 /* We are now a listener thread */
1638 /* if server is restarting( typically smooth shutdown) then do not
1639 * allow any new calls.
1642 if (rx_tranquil && (call != NULL)) {
1646 MUTEX_ENTER(&call->lock);
1648 rxi_CallError(call, RX_RESTARTING);
1649 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1651 MUTEX_EXIT(&call->lock);
1655 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1656 #ifdef RX_ENABLE_LOCKS
1658 #endif /* RX_ENABLE_LOCKS */
1659 afs_termState = AFSOP_STOP_AFS;
1660 afs_osi_Wakeup(&afs_termState);
1661 #ifdef RX_ENABLE_LOCKS
1663 #endif /* RX_ENABLE_LOCKS */
1668 tservice = call->conn->service;
1670 if (tservice->beforeProc)
1671 (*tservice->beforeProc) (call);
1673 code = tservice->executeRequestProc(call);
1675 if (tservice->afterProc)
1676 (*tservice->afterProc) (call, code);
1678 rx_EndCall(call, code);
1679 if (rx_stats_active) {
1680 MUTEX_ENTER(&rx_stats_mutex);
1682 MUTEX_EXIT(&rx_stats_mutex);
1689 rx_WakeupServerProcs(void)
1691 struct rx_serverQueueEntry *np, *tqp;
1695 MUTEX_ENTER(&rx_serverPool_lock);
1697 #ifdef RX_ENABLE_LOCKS
1698 if (rx_waitForPacket)
1699 CV_BROADCAST(&rx_waitForPacket->cv);
1700 #else /* RX_ENABLE_LOCKS */
1701 if (rx_waitForPacket)
1702 osi_rxWakeup(rx_waitForPacket);
1703 #endif /* RX_ENABLE_LOCKS */
1704 MUTEX_ENTER(&freeSQEList_lock);
1705 for (np = rx_FreeSQEList; np; np = tqp) {
1706 tqp = *(struct rx_serverQueueEntry **)np;
1707 #ifdef RX_ENABLE_LOCKS
1708 CV_BROADCAST(&np->cv);
1709 #else /* RX_ENABLE_LOCKS */
1711 #endif /* RX_ENABLE_LOCKS */
1713 MUTEX_EXIT(&freeSQEList_lock);
1714 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1715 #ifdef RX_ENABLE_LOCKS
1716 CV_BROADCAST(&np->cv);
1717 #else /* RX_ENABLE_LOCKS */
1719 #endif /* RX_ENABLE_LOCKS */
1721 MUTEX_EXIT(&rx_serverPool_lock);
1726 * One thing that seems to happen is that all the server threads get
1727 * tied up on some empty or slow call, and then a whole bunch of calls
1728 * arrive at once, using up the packet pool, so now there are more
1729 * empty calls. The most critical resources here are server threads
1730 * and the free packet pool. The "doreclaim" code seems to help in
1731 * general. I think that eventually we arrive in this state: there
1732 * are lots of pending calls which do have all their packets present,
1733 * so they won't be reclaimed, are multi-packet calls, so they won't
1734 * be scheduled until later, and thus are tying up most of the free
1735 * packet pool for a very long time.
1737 * 1. schedule multi-packet calls if all the packets are present.
1738 * Probably CPU-bound operation, useful to return packets to pool.
1739 * Do what if there is a full window, but the last packet isn't here?
1740 * 3. preserve one thread which *only* runs "best" calls, otherwise
1741 * it sleeps and waits for that type of call.
1742 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1743 * the current dataquota business is badly broken. The quota isn't adjusted
1744 * to reflect how many packets are presently queued for a running call.
1745 * So, when we schedule a queued call with a full window of packets queued
1746 * up for it, that *should* free up a window full of packets for other 2d-class
1747 * calls to be able to use from the packet pool. But it doesn't.
1749 * NB. Most of the time, this code doesn't run -- since idle server threads
1750 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1751 * as a new call arrives.
1753 /* Sleep until a call arrives. Returns a pointer to the call, ready
1754 * for an rx_Read. */
1755 #ifdef RX_ENABLE_LOCKS
1757 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1759 struct rx_serverQueueEntry *sq;
1760 struct rx_call *call = (struct rx_call *)0;
1761 struct rx_service *service = NULL;
1764 MUTEX_ENTER(&freeSQEList_lock);
1766 if ((sq = rx_FreeSQEList)) {
1767 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1768 MUTEX_EXIT(&freeSQEList_lock);
1769 } else { /* otherwise allocate a new one and return that */
1770 MUTEX_EXIT(&freeSQEList_lock);
1771 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1772 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1773 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1776 MUTEX_ENTER(&rx_serverPool_lock);
1777 if (cur_service != NULL) {
1778 ReturnToServerPool(cur_service);
1781 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1782 struct rx_call *tcall, *ncall, *choice2 = NULL;
1784 /* Scan for eligible incoming calls. A call is not eligible
1785 * if the maximum number of calls for its service type are
1786 * already executing */
1787 /* One thread will process calls FCFS (to prevent starvation),
1788 * while the other threads may run ahead looking for calls which
1789 * have all their input data available immediately. This helps
1790 * keep threads from blocking, waiting for data from the client. */
1791 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1792 service = tcall->conn->service;
1793 if (!QuotaOK(service)) {
1796 MUTEX_ENTER(&rx_pthread_mutex);
1797 if (tno == rxi_fcfs_thread_num
1798 || !tcall->queue_item_header.next) {
1799 MUTEX_EXIT(&rx_pthread_mutex);
1800 /* If we're the fcfs thread , then we'll just use
1801 * this call. If we haven't been able to find an optimal
1802 * choice, and we're at the end of the list, then use a
1803 * 2d choice if one has been identified. Otherwise... */
1804 call = (choice2 ? choice2 : tcall);
1805 service = call->conn->service;
1807 MUTEX_EXIT(&rx_pthread_mutex);
1808 if (!queue_IsEmpty(&tcall->rq)) {
1809 struct rx_packet *rp;
1810 rp = queue_First(&tcall->rq, rx_packet);
1811 if (rp->header.seq == 1) {
1813 || (rp->header.flags & RX_LAST_PACKET)) {
1815 } else if (rxi_2dchoice && !choice2
1816 && !(tcall->flags & RX_CALL_CLEARED)
1817 && (tcall->rprev > rxi_HardAckRate)) {
1827 ReturnToServerPool(service);
1834 MUTEX_EXIT(&rx_serverPool_lock);
1835 MUTEX_ENTER(&call->lock);
1837 if (call->flags & RX_CALL_WAIT_PROC) {
1838 call->flags &= ~RX_CALL_WAIT_PROC;
1839 rx_atomic_dec(&rx_nWaiting);
1842 if (call->state != RX_STATE_PRECALL || call->error) {
1843 MUTEX_EXIT(&call->lock);
1844 MUTEX_ENTER(&rx_serverPool_lock);
1845 ReturnToServerPool(service);
1850 if (queue_IsEmpty(&call->rq)
1851 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1852 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1854 CLEAR_CALL_QUEUE_LOCK(call);
1857 /* If there are no eligible incoming calls, add this process
1858 * to the idle server queue, to wait for one */
1862 *socketp = OSI_NULLSOCKET;
1864 sq->socketp = socketp;
1865 queue_Append(&rx_idleServerQueue, sq);
1866 #ifndef AFS_AIX41_ENV
1867 rx_waitForPacket = sq;
1869 rx_waitingForPacket = sq;
1870 #endif /* AFS_AIX41_ENV */
1872 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1874 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1875 MUTEX_EXIT(&rx_serverPool_lock);
1876 return (struct rx_call *)0;
1879 } while (!(call = sq->newcall)
1880 && !(socketp && *socketp != OSI_NULLSOCKET));
1881 MUTEX_EXIT(&rx_serverPool_lock);
1883 MUTEX_ENTER(&call->lock);
1889 MUTEX_ENTER(&freeSQEList_lock);
1890 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1891 rx_FreeSQEList = sq;
1892 MUTEX_EXIT(&freeSQEList_lock);
1895 clock_GetTime(&call->startTime);
1896 call->state = RX_STATE_ACTIVE;
1897 call->mode = RX_MODE_RECEIVING;
1898 #ifdef RX_KERNEL_TRACE
1899 if (ICL_SETACTIVE(afs_iclSetp)) {
1900 int glockOwner = ISAFS_GLOCK();
1903 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1904 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1911 rxi_calltrace(RX_CALL_START, call);
1912 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
1913 call->conn->service->servicePort, call->conn->service->serviceId,
1916 MUTEX_EXIT(&call->lock);
1917 MUTEX_ENTER(&rx_refcnt_mutex);
1918 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1919 MUTEX_EXIT(&rx_refcnt_mutex);
1921 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1926 #else /* RX_ENABLE_LOCKS */
1928 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1930 struct rx_serverQueueEntry *sq;
1931 struct rx_call *call = (struct rx_call *)0, *choice2;
1932 struct rx_service *service = NULL;
1936 MUTEX_ENTER(&freeSQEList_lock);
1938 if ((sq = rx_FreeSQEList)) {
1939 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1940 MUTEX_EXIT(&freeSQEList_lock);
1941 } else { /* otherwise allocate a new one and return that */
1942 MUTEX_EXIT(&freeSQEList_lock);
1943 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1944 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1945 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1947 MUTEX_ENTER(&sq->lock);
1949 if (cur_service != NULL) {
1950 cur_service->nRequestsRunning--;
1951 MUTEX_ENTER(&rx_quota_mutex);
1952 if (cur_service->nRequestsRunning < cur_service->minProcs)
1955 MUTEX_EXIT(&rx_quota_mutex);
1957 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1958 struct rx_call *tcall, *ncall;
1959 /* Scan for eligible incoming calls. A call is not eligible
1960 * if the maximum number of calls for its service type are
1961 * already executing */
1962 /* One thread will process calls FCFS (to prevent starvation),
1963 * while the other threads may run ahead looking for calls which
1964 * have all their input data available immediately. This helps
1965 * keep threads from blocking, waiting for data from the client. */
1966 choice2 = (struct rx_call *)0;
1967 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1968 service = tcall->conn->service;
1969 if (QuotaOK(service)) {
1970 MUTEX_ENTER(&rx_pthread_mutex);
1971 if (tno == rxi_fcfs_thread_num
1972 || !tcall->queue_item_header.next) {
1973 MUTEX_EXIT(&rx_pthread_mutex);
1974 /* If we're the fcfs thread, then we'll just use
1975 * this call. If we haven't been able to find an optimal
1976 * choice, and we're at the end of the list, then use a
1977 * 2d choice if one has been identified. Otherwise... */
1978 call = (choice2 ? choice2 : tcall);
1979 service = call->conn->service;
1981 MUTEX_EXIT(&rx_pthread_mutex);
1982 if (!queue_IsEmpty(&tcall->rq)) {
1983 struct rx_packet *rp;
1984 rp = queue_First(&tcall->rq, rx_packet);
1985 if (rp->header.seq == 1
1987 || (rp->header.flags & RX_LAST_PACKET))) {
1989 } else if (rxi_2dchoice && !choice2
1990 && !(tcall->flags & RX_CALL_CLEARED)
1991 && (tcall->rprev > rxi_HardAckRate)) {
2005 /* we can't schedule a call if there's no data!!! */
2006 /* send an ack if there's no data, if we're missing the
2007 * first packet, or we're missing something between first
2008 * and last -- there's a "hole" in the incoming data. */
2009 if (queue_IsEmpty(&call->rq)
2010 || queue_First(&call->rq, rx_packet)->header.seq != 1
2011 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
2012 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2014 call->flags &= (~RX_CALL_WAIT_PROC);
2015 service->nRequestsRunning++;
2016 /* just started call in minProcs pool, need fewer to maintain
2018 MUTEX_ENTER(&rx_quota_mutex);
2019 if (service->nRequestsRunning <= service->minProcs)
2022 MUTEX_EXIT(&rx_quota_mutex);
2023 rx_atomic_dec(&rx_nWaiting);
2024 /* MUTEX_EXIT(&call->lock); */
2026 /* If there are no eligible incoming calls, add this process
2027 * to the idle server queue, to wait for one */
2030 *socketp = OSI_NULLSOCKET;
2032 sq->socketp = socketp;
2033 queue_Append(&rx_idleServerQueue, sq);
2037 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2039 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2040 return (struct rx_call *)0;
2043 } while (!(call = sq->newcall)
2044 && !(socketp && *socketp != OSI_NULLSOCKET));
2046 MUTEX_EXIT(&sq->lock);
2048 MUTEX_ENTER(&freeSQEList_lock);
2049 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2050 rx_FreeSQEList = sq;
2051 MUTEX_EXIT(&freeSQEList_lock);
2054 clock_GetTime(&call->startTime);
2055 call->state = RX_STATE_ACTIVE;
2056 call->mode = RX_MODE_RECEIVING;
2057 #ifdef RX_KERNEL_TRACE
2058 if (ICL_SETACTIVE(afs_iclSetp)) {
2059 int glockOwner = ISAFS_GLOCK();
2062 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2063 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2070 rxi_calltrace(RX_CALL_START, call);
2071 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2072 call->conn->service->servicePort, call->conn->service->serviceId,
2075 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2082 #endif /* RX_ENABLE_LOCKS */
2086 /* Establish a procedure to be called when a packet arrives for a
2087 * call. This routine will be called at most once after each call,
2088 * and will also be called if there is an error condition on the or
2089 * the call is complete. Used by multi rx to build a selection
2090 * function which determines which of several calls is likely to be a
2091 * good one to read from.
2092 * NOTE: the way this is currently implemented it is probably only a
2093 * good idea to (1) use it immediately after a newcall (clients only)
2094 * and (2) only use it once. Other uses currently void your warranty
2097 rx_SetArrivalProc(struct rx_call *call,
2098 void (*proc) (struct rx_call * call,
2101 void * handle, int arg)
2103 call->arrivalProc = proc;
2104 call->arrivalProcHandle = handle;
2105 call->arrivalProcArg = arg;
2108 /* Call is finished (possibly prematurely). Return rc to the peer, if
2109 * appropriate, and return the final error code from the conversation
2113 rx_EndCall(struct rx_call *call, afs_int32 rc)
2115 struct rx_connection *conn = call->conn;
2119 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2120 call, rc, call->error, call->abortCode));
2123 MUTEX_ENTER(&call->lock);
2125 if (rc == 0 && call->error == 0) {
2126 call->abortCode = 0;
2127 call->abortCount = 0;
2130 call->arrivalProc = (void (*)())0;
2131 if (rc && call->error == 0) {
2132 rxi_CallError(call, rc);
2133 call->mode = RX_MODE_ERROR;
2134 /* Send an abort message to the peer if this error code has
2135 * only just been set. If it was set previously, assume the
2136 * peer has already been sent the error code or will request it
2138 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2140 if (conn->type == RX_SERVER_CONNECTION) {
2141 /* Make sure reply or at least dummy reply is sent */
2142 if (call->mode == RX_MODE_RECEIVING) {
2143 MUTEX_EXIT(&call->lock);
2144 rxi_WriteProc(call, 0, 0);
2145 MUTEX_ENTER(&call->lock);
2147 if (call->mode == RX_MODE_SENDING) {
2148 MUTEX_EXIT(&call->lock);
2149 rxi_FlushWrite(call);
2150 MUTEX_ENTER(&call->lock);
2152 rxi_calltrace(RX_CALL_END, call);
2153 /* Call goes to hold state until reply packets are acknowledged */
2154 if (call->tfirst + call->nSoftAcked < call->tnext) {
2155 call->state = RX_STATE_HOLD;
2157 call->state = RX_STATE_DALLY;
2158 rxi_ClearTransmitQueue(call, 0);
2159 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
2160 rxevent_Cancel(call->keepAliveEvent, call,
2161 RX_CALL_REFCOUNT_ALIVE);
2163 } else { /* Client connection */
2165 /* Make sure server receives input packets, in the case where
2166 * no reply arguments are expected */
2167 if ((call->mode == RX_MODE_SENDING)
2168 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2169 MUTEX_EXIT(&call->lock);
2170 (void)rxi_ReadProc(call, &dummy, 1);
2171 MUTEX_ENTER(&call->lock);
2174 /* If we had an outstanding delayed ack, be nice to the server
2175 * and force-send it now.
2177 if (call->delayedAckEvent) {
2178 rxevent_Cancel(call->delayedAckEvent, call,
2179 RX_CALL_REFCOUNT_DELAY);
2180 call->delayedAckEvent = NULL;
2181 rxi_SendDelayedAck(NULL, call, NULL);
2184 /* We need to release the call lock since it's lower than the
2185 * conn_call_lock and we don't want to hold the conn_call_lock
2186 * over the rx_ReadProc call. The conn_call_lock needs to be held
2187 * here for the case where rx_NewCall is perusing the calls on
2188 * the connection structure. We don't want to signal until
2189 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2190 * have checked this call, found it active and by the time it
2191 * goes to sleep, will have missed the signal.
2193 MUTEX_EXIT(&call->lock);
2194 MUTEX_ENTER(&conn->conn_call_lock);
2195 MUTEX_ENTER(&call->lock);
2196 MUTEX_ENTER(&conn->conn_data_lock);
2197 conn->flags |= RX_CONN_BUSY;
2198 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2199 MUTEX_EXIT(&conn->conn_data_lock);
2200 #ifdef RX_ENABLE_LOCKS
2201 CV_BROADCAST(&conn->conn_call_cv);
2206 #ifdef RX_ENABLE_LOCKS
2208 MUTEX_EXIT(&conn->conn_data_lock);
2210 #endif /* RX_ENABLE_LOCKS */
2211 call->state = RX_STATE_DALLY;
2213 error = call->error;
2215 /* currentPacket, nLeft, and NFree must be zeroed here, because
2216 * ResetCall cannot: ResetCall may be called at splnet(), in the
2217 * kernel version, and may interrupt the macros rx_Read or
2218 * rx_Write, which run at normal priority for efficiency. */
2219 if (call->currentPacket) {
2220 #ifdef RX_TRACK_PACKETS
2221 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2223 rxi_FreePacket(call->currentPacket);
2224 call->currentPacket = (struct rx_packet *)0;
2227 call->nLeft = call->nFree = call->curlen = 0;
2229 /* Free any packets from the last call to ReadvProc/WritevProc */
2230 #ifdef RXDEBUG_PACKET
2232 #endif /* RXDEBUG_PACKET */
2233 rxi_FreePackets(0, &call->iovq);
2234 MUTEX_EXIT(&call->lock);
2236 MUTEX_ENTER(&rx_refcnt_mutex);
2237 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2238 MUTEX_EXIT(&rx_refcnt_mutex);
2239 if (conn->type == RX_CLIENT_CONNECTION) {
2240 MUTEX_ENTER(&conn->conn_data_lock);
2241 conn->flags &= ~RX_CONN_BUSY;
2242 MUTEX_EXIT(&conn->conn_data_lock);
2243 MUTEX_EXIT(&conn->conn_call_lock);
2247 * Map errors to the local host's errno.h format.
2249 error = ntoh_syserr_conv(error);
2253 #if !defined(KERNEL)
2255 /* Call this routine when shutting down a server or client (especially
2256 * clients). This will allow Rx to gracefully garbage collect server
2257 * connections, and reduce the number of retries that a server might
2258 * make to a dead client.
2259 * This is not quite right, since some calls may still be ongoing and
2260 * we can't lock them to destroy them. */
2264 struct rx_connection **conn_ptr, **conn_end;
2268 if (rxinit_status == 1) {
2270 return; /* Already shutdown. */
2272 rxi_DeleteCachedConnections();
2273 if (rx_connHashTable) {
2274 MUTEX_ENTER(&rx_connHashTable_lock);
2275 for (conn_ptr = &rx_connHashTable[0], conn_end =
2276 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2278 struct rx_connection *conn, *next;
2279 for (conn = *conn_ptr; conn; conn = next) {
2281 if (conn->type == RX_CLIENT_CONNECTION) {
2282 MUTEX_ENTER(&rx_refcnt_mutex);
2284 MUTEX_EXIT(&rx_refcnt_mutex);
2285 #ifdef RX_ENABLE_LOCKS
2286 rxi_DestroyConnectionNoLock(conn);
2287 #else /* RX_ENABLE_LOCKS */
2288 rxi_DestroyConnection(conn);
2289 #endif /* RX_ENABLE_LOCKS */
2293 #ifdef RX_ENABLE_LOCKS
2294 while (rx_connCleanup_list) {
2295 struct rx_connection *conn;
2296 conn = rx_connCleanup_list;
2297 rx_connCleanup_list = rx_connCleanup_list->next;
2298 MUTEX_EXIT(&rx_connHashTable_lock);
2299 rxi_CleanupConnection(conn);
2300 MUTEX_ENTER(&rx_connHashTable_lock);
2302 MUTEX_EXIT(&rx_connHashTable_lock);
2303 #endif /* RX_ENABLE_LOCKS */
2308 afs_winsockCleanup();
2316 /* if we wakeup packet waiter too often, can get in loop with two
2317 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2319 rxi_PacketsUnWait(void)
2321 if (!rx_waitingForPackets) {
2325 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2326 return; /* still over quota */
2329 rx_waitingForPackets = 0;
2330 #ifdef RX_ENABLE_LOCKS
2331 CV_BROADCAST(&rx_waitingForPackets_cv);
2333 osi_rxWakeup(&rx_waitingForPackets);
2339 /* ------------------Internal interfaces------------------------- */
2341 /* Return this process's service structure for the
2342 * specified socket and service */
2344 rxi_FindService(osi_socket socket, u_short serviceId)
2346 struct rx_service **sp;
2347 for (sp = &rx_services[0]; *sp; sp++) {
2348 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2354 #ifdef RXDEBUG_PACKET
2355 #ifdef KDUMP_RX_LOCK
2356 static struct rx_call_rx_lock *rx_allCallsp = 0;
2358 static struct rx_call *rx_allCallsp = 0;
2360 #endif /* RXDEBUG_PACKET */
2362 /* Allocate a call structure, for the indicated channel of the
2363 * supplied connection. The mode and state of the call must be set by
2364 * the caller. Returns the call with mutex locked. */
2366 rxi_NewCall(struct rx_connection *conn, int channel)
2368 struct rx_call *call;
2369 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2370 struct rx_call *cp; /* Call pointer temp */
2371 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2372 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2374 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2376 /* Grab an existing call structure, or allocate a new one.
2377 * Existing call structures are assumed to have been left reset by
2379 MUTEX_ENTER(&rx_freeCallQueue_lock);
2381 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2383 * EXCEPT that the TQ might not yet be cleared out.
2384 * Skip over those with in-use TQs.
2387 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2388 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2394 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2395 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2396 call = queue_First(&rx_freeCallQueue, rx_call);
2397 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2399 if (rx_stats_active)
2400 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2401 MUTEX_EXIT(&rx_freeCallQueue_lock);
2402 MUTEX_ENTER(&call->lock);
2403 CLEAR_CALL_QUEUE_LOCK(call);
2404 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2405 /* Now, if TQ wasn't cleared earlier, do it now. */
2406 rxi_WaitforTQBusy(call);
2407 if (call->flags & RX_CALL_TQ_CLEARME) {
2408 rxi_ClearTransmitQueue(call, 1);
2409 /*queue_Init(&call->tq);*/
2411 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2412 /* Bind the call to its connection structure */
2414 rxi_ResetCall(call, 1);
2417 call = rxi_Alloc(sizeof(struct rx_call));
2418 #ifdef RXDEBUG_PACKET
2419 call->allNextp = rx_allCallsp;
2420 rx_allCallsp = call;
2422 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2423 #else /* RXDEBUG_PACKET */
2424 rx_atomic_inc(&rx_stats.nCallStructs);
2425 #endif /* RXDEBUG_PACKET */
2427 MUTEX_EXIT(&rx_freeCallQueue_lock);
2428 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2429 MUTEX_ENTER(&call->lock);
2430 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2431 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2432 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2434 /* Initialize once-only items */
2435 queue_Init(&call->tq);
2436 queue_Init(&call->rq);
2437 queue_Init(&call->iovq);
2438 #ifdef RXDEBUG_PACKET
2439 call->rqc = call->tqc = call->iovqc = 0;
2440 #endif /* RXDEBUG_PACKET */
2441 /* Bind the call to its connection structure (prereq for reset) */
2443 rxi_ResetCall(call, 1);
2445 call->channel = channel;
2446 call->callNumber = &conn->callNumber[channel];
2447 call->rwind = conn->rwind[channel];
2448 call->twind = conn->twind[channel];
2449 /* Note that the next expected call number is retained (in
2450 * conn->callNumber[i]), even if we reallocate the call structure
2452 conn->call[channel] = call;
2453 /* if the channel's never been used (== 0), we should start at 1, otherwise
2454 * the call number is valid from the last time this channel was used */
2455 if (*call->callNumber == 0)
2456 *call->callNumber = 1;
2461 /* A call has been inactive long enough that so we can throw away
2462 * state, including the call structure, which is placed on the call
2465 * call->lock amd rx_refcnt_mutex are held upon entry.
2466 * haveCTLock is set when called from rxi_ReapConnections.
2469 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2471 int channel = call->channel;
2472 struct rx_connection *conn = call->conn;
2475 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2476 (*call->callNumber)++;
2478 * We are setting the state to RX_STATE_RESET to
2479 * ensure that no one else will attempt to use this
2480 * call once we drop the refcnt lock. We must drop
2481 * the refcnt lock before calling rxi_ResetCall
2482 * because it cannot be held across acquiring the
2483 * freepktQ lock. NewCall does the same.
2485 call->state = RX_STATE_RESET;
2486 MUTEX_EXIT(&rx_refcnt_mutex);
2487 rxi_ResetCall(call, 0);
2488 call->conn->call[channel] = (struct rx_call *)0;
2490 MUTEX_ENTER(&rx_freeCallQueue_lock);
2491 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2492 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2493 /* A call may be free even though its transmit queue is still in use.
2494 * Since we search the call list from head to tail, put busy calls at
2495 * the head of the list, and idle calls at the tail.
2497 if (call->flags & RX_CALL_TQ_BUSY)
2498 queue_Prepend(&rx_freeCallQueue, call);
2500 queue_Append(&rx_freeCallQueue, call);
2501 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2502 queue_Append(&rx_freeCallQueue, call);
2503 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2504 if (rx_stats_active)
2505 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2506 MUTEX_EXIT(&rx_freeCallQueue_lock);
2508 /* Destroy the connection if it was previously slated for
2509 * destruction, i.e. the Rx client code previously called
2510 * rx_DestroyConnection (client connections), or
2511 * rxi_ReapConnections called the same routine (server
2512 * connections). Only do this, however, if there are no
2513 * outstanding calls. Note that for fine grain locking, there appears
2514 * to be a deadlock in that rxi_FreeCall has a call locked and
2515 * DestroyConnectionNoLock locks each call in the conn. But note a
2516 * few lines up where we have removed this call from the conn.
2517 * If someone else destroys a connection, they either have no
2518 * call lock held or are going through this section of code.
2520 MUTEX_ENTER(&conn->conn_data_lock);
2521 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2522 MUTEX_ENTER(&rx_refcnt_mutex);
2524 MUTEX_EXIT(&rx_refcnt_mutex);
2525 MUTEX_EXIT(&conn->conn_data_lock);
2526 #ifdef RX_ENABLE_LOCKS
2528 rxi_DestroyConnectionNoLock(conn);
2530 rxi_DestroyConnection(conn);
2531 #else /* RX_ENABLE_LOCKS */
2532 rxi_DestroyConnection(conn);
2533 #endif /* RX_ENABLE_LOCKS */
2535 MUTEX_EXIT(&conn->conn_data_lock);
2537 MUTEX_ENTER(&rx_refcnt_mutex);
2540 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2541 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2544 rxi_Alloc(size_t size)
2548 if (rx_stats_active) {
2549 rx_atomic_add(&rxi_Allocsize, (int) size);
2550 rx_atomic_inc(&rxi_Alloccnt);
2554 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2555 afs_osi_Alloc_NoSleep(size);
2560 osi_Panic("rxi_Alloc error");
2566 rxi_Free(void *addr, size_t size)
2568 if (rx_stats_active) {
2569 rx_atomic_sub(&rxi_Allocsize, (int) size);
2570 rx_atomic_dec(&rxi_Alloccnt);
2572 osi_Free(addr, size);
2576 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2578 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2579 struct rx_peer *next = NULL;
2583 MUTEX_ENTER(&rx_peerHashTable_lock);
2585 peer_ptr = &rx_peerHashTable[0];
2586 peer_end = &rx_peerHashTable[rx_hashTableSize];
2589 for ( ; peer_ptr < peer_end; peer_ptr++) {
2592 for ( ; peer; peer = next) {
2594 if (host == peer->host)
2599 hashIndex = PEER_HASH(host, port);
2600 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2601 if ((peer->host == host) && (peer->port == port))
2606 MUTEX_ENTER(&rx_peerHashTable_lock);
2611 MUTEX_EXIT(&rx_peerHashTable_lock);
2613 MUTEX_ENTER(&peer->peer_lock);
2614 /* We don't handle dropping below min, so don't */
2615 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2616 peer->ifMTU=MIN(mtu, peer->ifMTU);
2617 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2618 /* if we tweaked this down, need to tune our peer MTU too */
2619 peer->MTU = MIN(peer->MTU, peer->natMTU);
2620 /* if we discovered a sub-1500 mtu, degrade */
2621 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2622 peer->maxDgramPackets = 1;
2623 /* We no longer have valid peer packet information */
2624 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2625 peer->maxPacketSize = 0;
2626 MUTEX_EXIT(&peer->peer_lock);
2628 MUTEX_ENTER(&rx_peerHashTable_lock);
2630 if (host && !port) {
2632 /* pick up where we left off */
2636 MUTEX_EXIT(&rx_peerHashTable_lock);
2639 /* Find the peer process represented by the supplied (host,port)
2640 * combination. If there is no appropriate active peer structure, a
2641 * new one will be allocated and initialized
2642 * The origPeer, if set, is a pointer to a peer structure on which the
2643 * refcount will be be decremented. This is used to replace the peer
2644 * structure hanging off a connection structure */
2646 rxi_FindPeer(afs_uint32 host, u_short port,
2647 struct rx_peer *origPeer, int create)
2651 hashIndex = PEER_HASH(host, port);
2652 MUTEX_ENTER(&rx_peerHashTable_lock);
2653 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2654 if ((pp->host == host) && (pp->port == port))
2659 pp = rxi_AllocPeer(); /* This bzero's *pp */
2660 pp->host = host; /* set here or in InitPeerParams is zero */
2662 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2663 queue_Init(&pp->congestionQueue);
2664 queue_Init(&pp->rpcStats);
2665 pp->next = rx_peerHashTable[hashIndex];
2666 rx_peerHashTable[hashIndex] = pp;
2667 rxi_InitPeerParams(pp);
2668 if (rx_stats_active)
2669 rx_atomic_inc(&rx_stats.nPeerStructs);
2676 origPeer->refCount--;
2677 MUTEX_EXIT(&rx_peerHashTable_lock);
2682 /* Find the connection at (host, port) started at epoch, and with the
2683 * given connection id. Creates the server connection if necessary.
2684 * The type specifies whether a client connection or a server
2685 * connection is desired. In both cases, (host, port) specify the
2686 * peer's (host, pair) pair. Client connections are not made
2687 * automatically by this routine. The parameter socket gives the
2688 * socket descriptor on which the packet was received. This is used,
2689 * in the case of server connections, to check that *new* connections
2690 * come via a valid (port, serviceId). Finally, the securityIndex
2691 * parameter must match the existing index for the connection. If a
2692 * server connection is created, it will be created using the supplied
2693 * index, if the index is valid for this service */
2694 struct rx_connection *
2695 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2696 u_short port, u_short serviceId, afs_uint32 cid,
2697 afs_uint32 epoch, int type, u_int securityIndex)
2699 int hashindex, flag, i;
2700 struct rx_connection *conn;
2701 hashindex = CONN_HASH(host, port, cid, epoch, type);
2702 MUTEX_ENTER(&rx_connHashTable_lock);
2703 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2704 rx_connHashTable[hashindex],
2707 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2708 && (epoch == conn->epoch)) {
2709 struct rx_peer *pp = conn->peer;
2710 if (securityIndex != conn->securityIndex) {
2711 /* this isn't supposed to happen, but someone could forge a packet
2712 * like this, and there seems to be some CM bug that makes this
2713 * happen from time to time -- in which case, the fileserver
2715 MUTEX_EXIT(&rx_connHashTable_lock);
2716 return (struct rx_connection *)0;
2718 if (pp->host == host && pp->port == port)
2720 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2722 /* So what happens when it's a callback connection? */
2723 if ( /*type == RX_CLIENT_CONNECTION && */
2724 (conn->epoch & 0x80000000))
2728 /* the connection rxLastConn that was used the last time is not the
2729 ** one we are looking for now. Hence, start searching in the hash */
2731 conn = rx_connHashTable[hashindex];
2736 struct rx_service *service;
2737 if (type == RX_CLIENT_CONNECTION) {
2738 MUTEX_EXIT(&rx_connHashTable_lock);
2739 return (struct rx_connection *)0;
2741 service = rxi_FindService(socket, serviceId);
2742 if (!service || (securityIndex >= service->nSecurityObjects)
2743 || (service->securityObjects[securityIndex] == 0)) {
2744 MUTEX_EXIT(&rx_connHashTable_lock);
2745 return (struct rx_connection *)0;
2747 conn = rxi_AllocConnection(); /* This bzero's the connection */
2748 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2749 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2750 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2751 conn->next = rx_connHashTable[hashindex];
2752 rx_connHashTable[hashindex] = conn;
2753 conn->peer = rxi_FindPeer(host, port, 0, 1);
2754 conn->type = RX_SERVER_CONNECTION;
2755 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2756 conn->epoch = epoch;
2757 conn->cid = cid & RX_CIDMASK;
2758 /* conn->serial = conn->lastSerial = 0; */
2759 /* conn->timeout = 0; */
2760 conn->ackRate = RX_FAST_ACK_RATE;
2761 conn->service = service;
2762 conn->serviceId = serviceId;
2763 conn->securityIndex = securityIndex;
2764 conn->securityObject = service->securityObjects[securityIndex];
2765 conn->nSpecific = 0;
2766 conn->specific = NULL;
2767 rx_SetConnDeadTime(conn, service->connDeadTime);
2768 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2769 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2770 for (i = 0; i < RX_MAXCALLS; i++) {
2771 conn->twind[i] = rx_initSendWindow;
2772 conn->rwind[i] = rx_initReceiveWindow;
2774 /* Notify security object of the new connection */
2775 RXS_NewConnection(conn->securityObject, conn);
2776 /* XXXX Connection timeout? */
2777 if (service->newConnProc)
2778 (*service->newConnProc) (conn);
2779 if (rx_stats_active)
2780 rx_atomic_inc(&rx_stats.nServerConns);
2783 MUTEX_ENTER(&rx_refcnt_mutex);
2785 MUTEX_EXIT(&rx_refcnt_mutex);
2787 rxLastConn = conn; /* store this connection as the last conn used */
2788 MUTEX_EXIT(&rx_connHashTable_lock);
2792 /* There are two packet tracing routines available for testing and monitoring
2793 * Rx. One is called just after every packet is received and the other is
2794 * called just before every packet is sent. Received packets, have had their
2795 * headers decoded, and packets to be sent have not yet had their headers
2796 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2797 * containing the network address. Both can be modified. The return value, if
2798 * non-zero, indicates that the packet should be dropped. */
2800 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2801 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2803 /* A packet has been received off the interface. Np is the packet, socket is
2804 * the socket number it was received from (useful in determining which service
2805 * this packet corresponds to), and (host, port) reflect the host,port of the
2806 * sender. This call returns the packet to the caller if it is finished with
2807 * it, rather than de-allocating it, just as a small performance hack */
2810 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
2811 afs_uint32 host, u_short port, int *tnop,
2812 struct rx_call **newcallp)
2814 struct rx_call *call;
2815 struct rx_connection *conn;
2817 afs_uint32 currentCallNumber;
2823 struct rx_packet *tnp;
2826 /* We don't print out the packet until now because (1) the time may not be
2827 * accurate enough until now in the lwp implementation (rx_Listener only gets
2828 * the time after the packet is read) and (2) from a protocol point of view,
2829 * this is the first time the packet has been seen */
2830 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2831 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2832 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
2833 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2834 np->header.epoch, np->header.cid, np->header.callNumber,
2835 np->header.seq, np->header.flags, np));
2838 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2839 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2842 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2843 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2846 /* If an input tracer function is defined, call it with the packet and
2847 * network address. Note this function may modify its arguments. */
2848 if (rx_justReceived) {
2849 struct sockaddr_in addr;
2851 addr.sin_family = AF_INET;
2852 addr.sin_port = port;
2853 addr.sin_addr.s_addr = host;
2854 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2855 addr.sin_len = sizeof(addr);
2856 #endif /* AFS_OSF_ENV */
2857 drop = (*rx_justReceived) (np, &addr);
2858 /* drop packet if return value is non-zero */
2861 port = addr.sin_port; /* in case fcn changed addr */
2862 host = addr.sin_addr.s_addr;
2866 /* If packet was not sent by the client, then *we* must be the client */
2867 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2868 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2870 /* Find the connection (or fabricate one, if we're the server & if
2871 * necessary) associated with this packet */
2873 rxi_FindConnection(socket, host, port, np->header.serviceId,
2874 np->header.cid, np->header.epoch, type,
2875 np->header.securityIndex);
2878 /* If no connection found or fabricated, just ignore the packet.
2879 * (An argument could be made for sending an abort packet for
2884 MUTEX_ENTER(&conn->conn_data_lock);
2885 if (conn->maxSerial < np->header.serial)
2886 conn->maxSerial = np->header.serial;
2887 MUTEX_EXIT(&conn->conn_data_lock);
2889 /* If the connection is in an error state, send an abort packet and ignore
2890 * the incoming packet */
2892 /* Don't respond to an abort packet--we don't want loops! */
2893 MUTEX_ENTER(&conn->conn_data_lock);
2894 if (np->header.type != RX_PACKET_TYPE_ABORT)
2895 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2896 MUTEX_ENTER(&rx_refcnt_mutex);
2898 MUTEX_EXIT(&rx_refcnt_mutex);
2899 MUTEX_EXIT(&conn->conn_data_lock);
2903 /* Check for connection-only requests (i.e. not call specific). */
2904 if (np->header.callNumber == 0) {
2905 switch (np->header.type) {
2906 case RX_PACKET_TYPE_ABORT: {
2907 /* What if the supplied error is zero? */
2908 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2909 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
2910 rxi_ConnectionError(conn, errcode);
2911 MUTEX_ENTER(&rx_refcnt_mutex);
2913 MUTEX_EXIT(&rx_refcnt_mutex);
2916 case RX_PACKET_TYPE_CHALLENGE:
2917 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2918 MUTEX_ENTER(&rx_refcnt_mutex);
2920 MUTEX_EXIT(&rx_refcnt_mutex);
2922 case RX_PACKET_TYPE_RESPONSE:
2923 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2924 MUTEX_ENTER(&rx_refcnt_mutex);
2926 MUTEX_EXIT(&rx_refcnt_mutex);
2928 case RX_PACKET_TYPE_PARAMS:
2929 case RX_PACKET_TYPE_PARAMS + 1:
2930 case RX_PACKET_TYPE_PARAMS + 2:
2931 /* ignore these packet types for now */
2932 MUTEX_ENTER(&rx_refcnt_mutex);
2934 MUTEX_EXIT(&rx_refcnt_mutex);
2939 /* Should not reach here, unless the peer is broken: send an
2941 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2942 MUTEX_ENTER(&conn->conn_data_lock);
2943 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2944 MUTEX_ENTER(&rx_refcnt_mutex);
2946 MUTEX_EXIT(&rx_refcnt_mutex);
2947 MUTEX_EXIT(&conn->conn_data_lock);
2952 channel = np->header.cid & RX_CHANNELMASK;
2953 call = conn->call[channel];
2954 #ifdef RX_ENABLE_LOCKS
2956 MUTEX_ENTER(&call->lock);
2957 /* Test to see if call struct is still attached to conn. */
2958 if (call != conn->call[channel]) {
2960 MUTEX_EXIT(&call->lock);
2961 if (type == RX_SERVER_CONNECTION) {
2962 call = conn->call[channel];
2963 /* If we started with no call attached and there is one now,
2964 * another thread is also running this routine and has gotten
2965 * the connection channel. We should drop this packet in the tests
2966 * below. If there was a call on this connection and it's now
2967 * gone, then we'll be making a new call below.
2968 * If there was previously a call and it's now different then
2969 * the old call was freed and another thread running this routine
2970 * has created a call on this channel. One of these two threads
2971 * has a packet for the old call and the code below handles those
2975 MUTEX_ENTER(&call->lock);
2977 /* This packet can't be for this call. If the new call address is
2978 * 0 then no call is running on this channel. If there is a call
2979 * then, since this is a client connection we're getting data for
2980 * it must be for the previous call.
2982 if (rx_stats_active)
2983 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
2984 MUTEX_ENTER(&rx_refcnt_mutex);
2986 MUTEX_EXIT(&rx_refcnt_mutex);
2991 currentCallNumber = conn->callNumber[channel];
2993 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2994 if (np->header.callNumber < currentCallNumber) {
2995 if (rx_stats_active)
2996 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
2997 #ifdef RX_ENABLE_LOCKS
2999 MUTEX_EXIT(&call->lock);
3001 MUTEX_ENTER(&rx_refcnt_mutex);
3003 MUTEX_EXIT(&rx_refcnt_mutex);
3007 MUTEX_ENTER(&conn->conn_call_lock);
3008 call = rxi_NewCall(conn, channel);
3009 MUTEX_EXIT(&conn->conn_call_lock);
3010 *call->callNumber = np->header.callNumber;
3012 if (np->header.callNumber == 0)
3013 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%.06d len %d\n",
3014 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3015 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3016 np->header.flags, np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
3018 call->state = RX_STATE_PRECALL;
3019 clock_GetTime(&call->queueTime);
3020 hzero(call->bytesSent);
3021 hzero(call->bytesRcvd);
3023 * If the number of queued calls exceeds the overload
3024 * threshold then abort this call.
3026 if ((rx_BusyThreshold > 0) &&
3027 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3028 struct rx_packet *tp;
3030 rxi_CallError(call, rx_BusyError);
3031 tp = rxi_SendCallAbort(call, np, 1, 0);
3032 MUTEX_EXIT(&call->lock);
3033 MUTEX_ENTER(&rx_refcnt_mutex);
3035 MUTEX_EXIT(&rx_refcnt_mutex);
3036 if (rx_stats_active)
3037 rx_atomic_inc(&rx_stats.nBusies);
3040 rxi_KeepAliveOn(call);
3041 } else if (np->header.callNumber != currentCallNumber) {
3042 /* Wait until the transmit queue is idle before deciding
3043 * whether to reset the current call. Chances are that the
3044 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3047 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3048 if (call->state == RX_STATE_ACTIVE) {
3049 rxi_WaitforTQBusy(call);
3051 * If we entered error state while waiting,
3052 * must call rxi_CallError to permit rxi_ResetCall
3053 * to processed when the tqWaiter count hits zero.
3056 rxi_CallError(call, call->error);
3057 MUTEX_EXIT(&call->lock);
3058 MUTEX_ENTER(&rx_refcnt_mutex);
3060 MUTEX_EXIT(&rx_refcnt_mutex);
3064 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3065 /* If the new call cannot be taken right now send a busy and set
3066 * the error condition in this call, so that it terminates as
3067 * quickly as possible */
3068 if (call->state == RX_STATE_ACTIVE) {
3069 struct rx_packet *tp;
3071 rxi_CallError(call, RX_CALL_DEAD);
3072 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3074 MUTEX_EXIT(&call->lock);
3075 MUTEX_ENTER(&rx_refcnt_mutex);
3077 MUTEX_EXIT(&rx_refcnt_mutex);
3080 rxi_ResetCall(call, 0);
3081 *call->callNumber = np->header.callNumber;
3083 if (np->header.callNumber == 0)
3084 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%06d len %d\n",
3085 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3086 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3087 np->header.flags, np, np->retryTime.sec, np->retryTime.usec, np->length));
3089 call->state = RX_STATE_PRECALL;
3090 clock_GetTime(&call->queueTime);
3091 hzero(call->bytesSent);
3092 hzero(call->bytesRcvd);
3094 * If the number of queued calls exceeds the overload
3095 * threshold then abort this call.
3097 if ((rx_BusyThreshold > 0) &&
3098 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3099 struct rx_packet *tp;
3101 rxi_CallError(call, rx_BusyError);
3102 tp = rxi_SendCallAbort(call, np, 1, 0);
3103 MUTEX_EXIT(&call->lock);
3104 MUTEX_ENTER(&rx_refcnt_mutex);
3106 MUTEX_EXIT(&rx_refcnt_mutex);
3107 if (rx_stats_active)
3108 rx_atomic_inc(&rx_stats.nBusies);
3111 rxi_KeepAliveOn(call);
3113 /* Continuing call; do nothing here. */
3115 } else { /* we're the client */
3116 /* Ignore all incoming acknowledgements for calls in DALLY state */
3117 if (call && (call->state == RX_STATE_DALLY)
3118 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3119 if (rx_stats_active)
3120 rx_atomic_inc(&rx_stats.ignorePacketDally);
3121 #ifdef RX_ENABLE_LOCKS
3123 MUTEX_EXIT(&call->lock);
3126 MUTEX_ENTER(&rx_refcnt_mutex);
3128 MUTEX_EXIT(&rx_refcnt_mutex);
3132 /* Ignore anything that's not relevant to the current call. If there
3133 * isn't a current call, then no packet is relevant. */
3134 if (!call || (np->header.callNumber != currentCallNumber)) {
3135 if (rx_stats_active)
3136 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3137 #ifdef RX_ENABLE_LOCKS
3139 MUTEX_EXIT(&call->lock);
3142 MUTEX_ENTER(&rx_refcnt_mutex);
3144 MUTEX_EXIT(&rx_refcnt_mutex);
3147 /* If the service security object index stamped in the packet does not
3148 * match the connection's security index, ignore the packet */
3149 if (np->header.securityIndex != conn->securityIndex) {
3150 #ifdef RX_ENABLE_LOCKS
3151 MUTEX_EXIT(&call->lock);
3153 MUTEX_ENTER(&rx_refcnt_mutex);
3155 MUTEX_EXIT(&rx_refcnt_mutex);
3159 /* If we're receiving the response, then all transmit packets are
3160 * implicitly acknowledged. Get rid of them. */
3161 if (np->header.type == RX_PACKET_TYPE_DATA) {
3162 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3163 /* XXX Hack. Because we must release the global rx lock when
3164 * sending packets (osi_NetSend) we drop all acks while we're
3165 * traversing the tq in rxi_Start sending packets out because
3166 * packets may move to the freePacketQueue as result of being here!
3167 * So we drop these packets until we're safely out of the
3168 * traversing. Really ugly!
3169 * For fine grain RX locking, we set the acked field in the
3170 * packets and let rxi_Start remove them from the transmit queue.
3172 if (call->flags & RX_CALL_TQ_BUSY) {
3173 #ifdef RX_ENABLE_LOCKS
3174 rxi_SetAcksInTransmitQueue(call);
3176 MUTEX_ENTER(&rx_refcnt_mutex);
3178 MUTEX_EXIT(&rx_refcnt_mutex);
3179 return np; /* xmitting; drop packet */
3182 rxi_ClearTransmitQueue(call, 0);
3184 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3185 rxi_ClearTransmitQueue(call, 0);
3186 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3188 if (np->header.type == RX_PACKET_TYPE_ACK) {
3189 /* now check to see if this is an ack packet acknowledging that the
3190 * server actually *lost* some hard-acked data. If this happens we
3191 * ignore this packet, as it may indicate that the server restarted in
3192 * the middle of a call. It is also possible that this is an old ack
3193 * packet. We don't abort the connection in this case, because this
3194 * *might* just be an old ack packet. The right way to detect a server
3195 * restart in the midst of a call is to notice that the server epoch
3197 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3198 * XXX unacknowledged. I think that this is off-by-one, but
3199 * XXX I don't dare change it just yet, since it will
3200 * XXX interact badly with the server-restart detection
3201 * XXX code in receiveackpacket. */
3202 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3203 if (rx_stats_active)
3204 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3205 MUTEX_EXIT(&call->lock);
3206 MUTEX_ENTER(&rx_refcnt_mutex);
3208 MUTEX_EXIT(&rx_refcnt_mutex);
3212 } /* else not a data packet */
3215 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3216 /* Set remote user defined status from packet */
3217 call->remoteStatus = np->header.userStatus;
3219 /* Note the gap between the expected next packet and the actual
3220 * packet that arrived, when the new packet has a smaller serial number
3221 * than expected. Rioses frequently reorder packets all by themselves,
3222 * so this will be quite important with very large window sizes.
3223 * Skew is checked against 0 here to avoid any dependence on the type of
3224 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3226 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3227 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3228 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3230 MUTEX_ENTER(&conn->conn_data_lock);
3231 skew = conn->lastSerial - np->header.serial;
3232 conn->lastSerial = np->header.serial;
3233 MUTEX_EXIT(&conn->conn_data_lock);
3235 struct rx_peer *peer;
3237 if (skew > peer->inPacketSkew) {
3238 dpf(("*** In skew changed from %d to %d\n",
3239 peer->inPacketSkew, skew));
3240 peer->inPacketSkew = skew;
3244 /* Now do packet type-specific processing */
3245 switch (np->header.type) {
3246 case RX_PACKET_TYPE_DATA:
3247 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3250 case RX_PACKET_TYPE_ACK:
3251 /* Respond immediately to ack packets requesting acknowledgement
3253 if (np->header.flags & RX_REQUEST_ACK) {
3255 (void)rxi_SendCallAbort(call, 0, 1, 0);
3257 (void)rxi_SendAck(call, 0, np->header.serial,
3258 RX_ACK_PING_RESPONSE, 1);
3260 np = rxi_ReceiveAckPacket(call, np, 1);
3262 case RX_PACKET_TYPE_ABORT: {
3263 /* An abort packet: reset the call, passing the error up to the user. */
3264 /* What if error is zero? */
3265 /* What if the error is -1? the application will treat it as a timeout. */
3266 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3267 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3268 rxi_CallError(call, errdata);
3269 MUTEX_EXIT(&call->lock);
3270 MUTEX_ENTER(&rx_refcnt_mutex);
3272 MUTEX_EXIT(&rx_refcnt_mutex);
3273 return np; /* xmitting; drop packet */
3275 case RX_PACKET_TYPE_BUSY:
3278 case RX_PACKET_TYPE_ACKALL:
3279 /* All packets acknowledged, so we can drop all packets previously
3280 * readied for sending */
3281 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3282 /* XXX Hack. We because we can't release the global rx lock when
3283 * sending packets (osi_NetSend) we drop all ack pkts while we're
3284 * traversing the tq in rxi_Start sending packets out because
3285 * packets may move to the freePacketQueue as result of being
3286 * here! So we drop these packets until we're safely out of the
3287 * traversing. Really ugly!
3288 * For fine grain RX locking, we set the acked field in the packets
3289 * and let rxi_Start remove the packets from the transmit queue.
3291 if (call->flags & RX_CALL_TQ_BUSY) {
3292 #ifdef RX_ENABLE_LOCKS
3293 rxi_SetAcksInTransmitQueue(call);
3295 #else /* RX_ENABLE_LOCKS */
3296 MUTEX_EXIT(&call->lock);
3297 MUTEX_ENTER(&rx_refcnt_mutex);
3299 MUTEX_EXIT(&rx_refcnt_mutex);
3300 return np; /* xmitting; drop packet */
3301 #endif /* RX_ENABLE_LOCKS */
3303 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3304 rxi_ClearTransmitQueue(call, 0);
3305 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3308 /* Should not reach here, unless the peer is broken: send an abort
3310 rxi_CallError(call, RX_PROTOCOL_ERROR);
3311 np = rxi_SendCallAbort(call, np, 1, 0);
3314 /* Note when this last legitimate packet was received, for keep-alive
3315 * processing. Note, we delay getting the time until now in the hope that
3316 * the packet will be delivered to the user before any get time is required
3317 * (if not, then the time won't actually be re-evaluated here). */
3318 call->lastReceiveTime = clock_Sec();
3319 MUTEX_EXIT(&call->lock);
3320 MUTEX_ENTER(&rx_refcnt_mutex);
3322 MUTEX_EXIT(&rx_refcnt_mutex);
3326 /* return true if this is an "interesting" connection from the point of view
3327 of someone trying to debug the system */
3329 rxi_IsConnInteresting(struct rx_connection *aconn)
3332 struct rx_call *tcall;
3334 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3337 for (i = 0; i < RX_MAXCALLS; i++) {
3338 tcall = aconn->call[i];
3340 if ((tcall->state == RX_STATE_PRECALL)
3341 || (tcall->state == RX_STATE_ACTIVE))
3343 if ((tcall->mode == RX_MODE_SENDING)
3344 || (tcall->mode == RX_MODE_RECEIVING))
3352 /* if this is one of the last few packets AND it wouldn't be used by the
3353 receiving call to immediately satisfy a read request, then drop it on
3354 the floor, since accepting it might prevent a lock-holding thread from
3355 making progress in its reading. If a call has been cleared while in
3356 the precall state then ignore all subsequent packets until the call
3357 is assigned to a thread. */
3360 TooLow(struct rx_packet *ap, struct rx_call *acall)
3364 MUTEX_ENTER(&rx_quota_mutex);
3365 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3366 && (acall->state == RX_STATE_PRECALL))
3367 || ((rx_nFreePackets < rxi_dataQuota + 2)
3368 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3369 && (acall->flags & RX_CALL_READER_WAIT)))) {
3372 MUTEX_EXIT(&rx_quota_mutex);
3378 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3380 struct rx_connection *conn = arg1;
3381 struct rx_call *acall = arg2;
3382 struct rx_call *call = acall;
3383 struct clock when, now;
3386 MUTEX_ENTER(&conn->conn_data_lock);
3387 conn->checkReachEvent = NULL;
3388 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3390 MUTEX_ENTER(&rx_refcnt_mutex);
3392 MUTEX_EXIT(&rx_refcnt_mutex);
3394 MUTEX_EXIT(&conn->conn_data_lock);
3398 MUTEX_ENTER(&conn->conn_call_lock);
3399 MUTEX_ENTER(&conn->conn_data_lock);
3400 for (i = 0; i < RX_MAXCALLS; i++) {
3401 struct rx_call *tc = conn->call[i];
3402 if (tc && tc->state == RX_STATE_PRECALL) {
3408 /* Indicate that rxi_CheckReachEvent is no longer running by
3409 * clearing the flag. Must be atomic under conn_data_lock to
3410 * avoid a new call slipping by: rxi_CheckConnReach holds
3411 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3413 conn->flags &= ~RX_CONN_ATTACHWAIT;
3414 MUTEX_EXIT(&conn->conn_data_lock);
3415 MUTEX_EXIT(&conn->conn_call_lock);
3420 MUTEX_ENTER(&call->lock);
3421 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3423 MUTEX_EXIT(&call->lock);
3425 clock_GetTime(&now);
3427 when.sec += RX_CHECKREACH_TIMEOUT;
3428 MUTEX_ENTER(&conn->conn_data_lock);
3429 if (!conn->checkReachEvent) {
3430 MUTEX_ENTER(&rx_refcnt_mutex);
3432 MUTEX_EXIT(&rx_refcnt_mutex);
3433 conn->checkReachEvent =
3434 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3437 MUTEX_EXIT(&conn->conn_data_lock);
3443 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3445 struct rx_service *service = conn->service;
3446 struct rx_peer *peer = conn->peer;
3447 afs_uint32 now, lastReach;
3449 if (service->checkReach == 0)
3453 MUTEX_ENTER(&peer->peer_lock);
3454 lastReach = peer->lastReachTime;
3455 MUTEX_EXIT(&peer->peer_lock);
3456 if (now - lastReach < RX_CHECKREACH_TTL)
3459 MUTEX_ENTER(&conn->conn_data_lock);
3460 if (conn->flags & RX_CONN_ATTACHWAIT) {
3461 MUTEX_EXIT(&conn->conn_data_lock);
3464 conn->flags |= RX_CONN_ATTACHWAIT;
3465 MUTEX_EXIT(&conn->conn_data_lock);
3466 if (!conn->checkReachEvent)
3467 rxi_CheckReachEvent(NULL, conn, call);
3472 /* try to attach call, if authentication is complete */
3474 TryAttach(struct rx_call *acall, osi_socket socket,
3475 int *tnop, struct rx_call **newcallp,
3478 struct rx_connection *conn = acall->conn;
3480 if (conn->type == RX_SERVER_CONNECTION
3481 && acall->state == RX_STATE_PRECALL) {
3482 /* Don't attach until we have any req'd. authentication. */
3483 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3484 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3485 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3486 /* Note: this does not necessarily succeed; there
3487 * may not any proc available
3490 rxi_ChallengeOn(acall->conn);
3495 /* A data packet has been received off the interface. This packet is
3496 * appropriate to the call (the call is in the right state, etc.). This
3497 * routine can return a packet to the caller, for re-use */
3500 rxi_ReceiveDataPacket(struct rx_call *call,
3501 struct rx_packet *np, int istack,
3502 osi_socket socket, afs_uint32 host, u_short port,
3503 int *tnop, struct rx_call **newcallp)
3505 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3510 afs_uint32 serial=0, flags=0;
3512 struct rx_packet *tnp;
3513 struct clock when, now;
3514 if (rx_stats_active)
3515 rx_atomic_inc(&rx_stats.dataPacketsRead);
3518 /* If there are no packet buffers, drop this new packet, unless we can find
3519 * packet buffers from inactive calls */
3521 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3522 MUTEX_ENTER(&rx_freePktQ_lock);
3523 rxi_NeedMorePackets = TRUE;
3524 MUTEX_EXIT(&rx_freePktQ_lock);
3525 if (rx_stats_active)
3526 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3527 call->rprev = np->header.serial;
3528 rxi_calltrace(RX_TRACE_DROP, call);
3529 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3531 rxi_ClearReceiveQueue(call);
3532 clock_GetTime(&now);
3534 clock_Add(&when, &rx_softAckDelay);
3535 if (!call->delayedAckEvent
3536 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3537 rxevent_Cancel(call->delayedAckEvent, call,
3538 RX_CALL_REFCOUNT_DELAY);
3539 MUTEX_ENTER(&rx_refcnt_mutex);
3540 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3541 MUTEX_EXIT(&rx_refcnt_mutex);
3543 call->delayedAckEvent =
3544 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3546 /* we've damaged this call already, might as well do it in. */
3552 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3553 * packet is one of several packets transmitted as a single
3554 * datagram. Do not send any soft or hard acks until all packets
3555 * in a jumbogram have been processed. Send negative acks right away.
3557 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3558 /* tnp is non-null when there are more packets in the
3559 * current jumbo gram */
3566 seq = np->header.seq;
3567 serial = np->header.serial;
3568 flags = np->header.flags;
3570 /* If the call is in an error state, send an abort message */
3572 return rxi_SendCallAbort(call, np, istack, 0);
3574 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3575 * AFS 3.5 jumbogram. */
3576 if (flags & RX_JUMBO_PACKET) {
3577 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3582 if (np->header.spare != 0) {
3583 MUTEX_ENTER(&call->conn->conn_data_lock);
3584 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3585 MUTEX_EXIT(&call->conn->conn_data_lock);
3588 /* The usual case is that this is the expected next packet */
3589 if (seq == call->rnext) {
3591 /* Check to make sure it is not a duplicate of one already queued */
3592 if (queue_IsNotEmpty(&call->rq)
3593 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3594 if (rx_stats_active)
3595 rx_atomic_inc(&rx_stats.dupPacketsRead);
3596 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3597 rxevent_Cancel(call->delayedAckEvent, call,
3598 RX_CALL_REFCOUNT_DELAY);
3599 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3605 /* It's the next packet. Stick it on the receive queue
3606 * for this call. Set newPackets to make sure we wake
3607 * the reader once all packets have been processed */
3608 #ifdef RX_TRACK_PACKETS
3609 np->flags |= RX_PKTFLAG_RQ;
3611 queue_Prepend(&call->rq, np);
3612 #ifdef RXDEBUG_PACKET
3614 #endif /* RXDEBUG_PACKET */
3616 np = NULL; /* We can't use this anymore */
3619 /* If an ack is requested then set a flag to make sure we
3620 * send an acknowledgement for this packet */
3621 if (flags & RX_REQUEST_ACK) {
3622 ackNeeded = RX_ACK_REQUESTED;
3625 /* Keep track of whether we have received the last packet */
3626 if (flags & RX_LAST_PACKET) {
3627 call->flags |= RX_CALL_HAVE_LAST;
3631 /* Check whether we have all of the packets for this call */
3632 if (call->flags & RX_CALL_HAVE_LAST) {
3633 afs_uint32 tseq; /* temporary sequence number */
3634 struct rx_packet *tp; /* Temporary packet pointer */
3635 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3637 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3638 if (tseq != tp->header.seq)
3640 if (tp->header.flags & RX_LAST_PACKET) {
3641 call->flags |= RX_CALL_RECEIVE_DONE;
3648 /* Provide asynchronous notification for those who want it
3649 * (e.g. multi rx) */
3650 if (call->arrivalProc) {
3651 (*call->arrivalProc) (call, call->arrivalProcHandle,
3652 call->arrivalProcArg);
3653 call->arrivalProc = (void (*)())0;
3656 /* Update last packet received */
3659 /* If there is no server process serving this call, grab
3660 * one, if available. We only need to do this once. If a
3661 * server thread is available, this thread becomes a server
3662 * thread and the server thread becomes a listener thread. */
3664 TryAttach(call, socket, tnop, newcallp, 0);
3667 /* This is not the expected next packet. */
3669 /* Determine whether this is a new or old packet, and if it's
3670 * a new one, whether it fits into the current receive window.
3671 * Also figure out whether the packet was delivered in sequence.
3672 * We use the prev variable to determine whether the new packet
3673 * is the successor of its immediate predecessor in the
3674 * receive queue, and the missing flag to determine whether
3675 * any of this packets predecessors are missing. */
3677 afs_uint32 prev; /* "Previous packet" sequence number */
3678 struct rx_packet *tp; /* Temporary packet pointer */
3679 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3680 int missing; /* Are any predecessors missing? */
3682 /* If the new packet's sequence number has been sent to the
3683 * application already, then this is a duplicate */
3684 if (seq < call->rnext) {
3685 if (rx_stats_active)
3686 rx_atomic_inc(&rx_stats.dupPacketsRead);
3687 rxevent_Cancel(call->delayedAckEvent, call,
3688 RX_CALL_REFCOUNT_DELAY);
3689 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3695 /* If the sequence number is greater than what can be
3696 * accomodated by the current window, then send a negative
3697 * acknowledge and drop the packet */
3698 if ((call->rnext + call->rwind) <= seq) {
3699 rxevent_Cancel(call->delayedAckEvent, call,
3700 RX_CALL_REFCOUNT_DELAY);
3701 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3708 /* Look for the packet in the queue of old received packets */
3709 for (prev = call->rnext - 1, missing =
3710 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3711 /*Check for duplicate packet */
3712 if (seq == tp->header.seq) {
3713 if (rx_stats_active)
3714 rx_atomic_inc(&rx_stats.dupPacketsRead);
3715 rxevent_Cancel(call->delayedAckEvent, call,
3716 RX_CALL_REFCOUNT_DELAY);
3717 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3723 /* If we find a higher sequence packet, break out and
3724 * insert the new packet here. */
3725 if (seq < tp->header.seq)
3727 /* Check for missing packet */
3728 if (tp->header.seq != prev + 1) {
3732 prev = tp->header.seq;
3735 /* Keep track of whether we have received the last packet. */
3736 if (flags & RX_LAST_PACKET) {
3737 call->flags |= RX_CALL_HAVE_LAST;
3740 /* It's within the window: add it to the the receive queue.
3741 * tp is left by the previous loop either pointing at the
3742 * packet before which to insert the new packet, or at the
3743 * queue head if the queue is empty or the packet should be
3745 #ifdef RX_TRACK_PACKETS
3746 np->flags |= RX_PKTFLAG_RQ;
3748 #ifdef RXDEBUG_PACKET
3750 #endif /* RXDEBUG_PACKET */
3751 queue_InsertBefore(tp, np);
3755 /* Check whether we have all of the packets for this call */
3756 if ((call->flags & RX_CALL_HAVE_LAST)
3757 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3758 afs_uint32 tseq; /* temporary sequence number */
3761 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3762 if (tseq != tp->header.seq)
3764 if (tp->header.flags & RX_LAST_PACKET) {
3765 call->flags |= RX_CALL_RECEIVE_DONE;
3772 /* We need to send an ack of the packet is out of sequence,
3773 * or if an ack was requested by the peer. */
3774 if (seq != prev + 1 || missing) {
3775 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3776 } else if (flags & RX_REQUEST_ACK) {
3777 ackNeeded = RX_ACK_REQUESTED;
3780 /* Acknowledge the last packet for each call */
3781 if (flags & RX_LAST_PACKET) {
3792 * If the receiver is waiting for an iovec, fill the iovec
3793 * using the data from the receive queue */
3794 if (call->flags & RX_CALL_IOVEC_WAIT) {
3795 didHardAck = rxi_FillReadVec(call, serial);
3796 /* the call may have been aborted */
3805 /* Wakeup the reader if any */
3806 if ((call->flags & RX_CALL_READER_WAIT)
3807 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3808 || (call->iovNext >= call->iovMax)
3809 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3810 call->flags &= ~RX_CALL_READER_WAIT;
3811 #ifdef RX_ENABLE_LOCKS
3812 CV_BROADCAST(&call->cv_rq);
3814 osi_rxWakeup(&call->rq);
3820 * Send an ack when requested by the peer, or once every
3821 * rxi_SoftAckRate packets until the last packet has been
3822 * received. Always send a soft ack for the last packet in
3823 * the server's reply.
3825 * If we have received all of the packets for the call
3826 * immediately send an RX_PACKET_TYPE_ACKALL packet so that
3827 * the peer can empty its packet queue and cancel all resend
3830 if (call->flags & RX_CALL_RECEIVE_DONE) {
3831 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3832 rxi_AckAll(NULL, call, 0);
3833 } else if (ackNeeded) {
3834 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3835 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3836 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3837 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3838 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3839 } else if (call->nSoftAcks) {
3840 clock_GetTime(&now);
3842 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3843 clock_Add(&when, &rx_lastAckDelay);
3845 clock_Add(&when, &rx_softAckDelay);
3847 if (!call->delayedAckEvent
3848 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3849 rxevent_Cancel(call->delayedAckEvent, call,
3850 RX_CALL_REFCOUNT_DELAY);
3851 MUTEX_ENTER(&rx_refcnt_mutex);
3852 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3853 MUTEX_EXIT(&rx_refcnt_mutex);
3854 call->delayedAckEvent =
3855 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3863 static void rxi_ComputeRate();
3867 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3869 struct rx_peer *peer = conn->peer;
3871 MUTEX_ENTER(&peer->peer_lock);
3872 peer->lastReachTime = clock_Sec();
3873 MUTEX_EXIT(&peer->peer_lock);
3875 MUTEX_ENTER(&conn->conn_data_lock);
3876 if (conn->flags & RX_CONN_ATTACHWAIT) {
3879 conn->flags &= ~RX_CONN_ATTACHWAIT;
3880 MUTEX_EXIT(&conn->conn_data_lock);
3882 for (i = 0; i < RX_MAXCALLS; i++) {
3883 struct rx_call *call = conn->call[i];
3886 MUTEX_ENTER(&call->lock);
3887 /* tnop can be null if newcallp is null */
3888 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3890 MUTEX_EXIT(&call->lock);
3894 MUTEX_EXIT(&conn->conn_data_lock);
3897 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
3899 rx_ack_reason(int reason)
3902 case RX_ACK_REQUESTED:
3904 case RX_ACK_DUPLICATE:
3906 case RX_ACK_OUT_OF_SEQUENCE:
3908 case RX_ACK_EXCEEDS_WINDOW:
3910 case RX_ACK_NOSPACE:
3914 case RX_ACK_PING_RESPONSE:
3927 /* The real smarts of the whole thing. */
3929 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
3932 struct rx_ackPacket *ap;
3934 struct rx_packet *tp;
3935 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3936 struct rx_connection *conn = call->conn;
3937 struct rx_peer *peer = conn->peer;
3938 struct clock now; /* Current time, for RTT calculations */
3942 /* because there are CM's that are bogus, sending weird values for this. */
3943 afs_uint32 skew = 0;
3948 int newAckCount = 0;
3949 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3950 int pktsize = 0; /* Set if we need to update the peer mtu */
3951 int conn_data_locked = 0;
3953 if (rx_stats_active)
3954 rx_atomic_inc(&rx_stats.ackPacketsRead);
3955 ap = (struct rx_ackPacket *)rx_DataOf(np);
3956 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3958 return np; /* truncated ack packet */
3960 /* depends on ack packet struct */
3961 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3962 first = ntohl(ap->firstPacket);
3963 prev = ntohl(ap->previousPacket);
3964 serial = ntohl(ap->serial);
3965 /* temporarily disabled -- needs to degrade over time
3966 * skew = ntohs(ap->maxSkew); */
3968 /* Ignore ack packets received out of order */
3969 if (first < call->tfirst ||
3970 (first == call->tfirst && prev < call->tprev)) {
3976 if (np->header.flags & RX_SLOW_START_OK) {
3977 call->flags |= RX_CALL_SLOW_START_OK;
3980 if (ap->reason == RX_ACK_PING_RESPONSE)
3981 rxi_UpdatePeerReach(conn, call);
3983 if (conn->lastPacketSizeSeq) {
3984 MUTEX_ENTER(&conn->conn_data_lock);
3985 conn_data_locked = 1;
3986 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
3987 pktsize = conn->lastPacketSize;
3988 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
3991 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
3992 if (!conn_data_locked) {
3993 MUTEX_ENTER(&conn->conn_data_lock);
3994 conn_data_locked = 1;
3996 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
3997 /* process mtu ping ack */
3998 pktsize = conn->lastPingSize;
3999 conn->lastPingSizeSer = conn->lastPingSize = 0;
4003 if (conn_data_locked) {
4004 MUTEX_EXIT(&conn->conn_data_lock);
4005 conn_data_locked = 0;
4009 if (rxdebug_active) {
4013 len = _snprintf(msg, sizeof(msg),
4014 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
4015 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4016 ntohl(ap->serial), ntohl(ap->previousPacket),
4017 (unsigned int)np->header.seq, (unsigned int)skew,
4018 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
4022 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4023 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4027 OutputDebugString(msg);
4029 #else /* AFS_NT40_ENV */
4032 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
4033 ap->reason, ntohl(ap->previousPacket),
4034 (unsigned int)np->header.seq, (unsigned int)serial,
4035 (unsigned int)skew, ntohl(ap->firstPacket));
4038 for (offset = 0; offset < nAcks; offset++)
4039 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4044 #endif /* AFS_NT40_ENV */
4047 MUTEX_ENTER(&peer->peer_lock);
4050 * Start somewhere. Can't assume we can send what we can receive,
4051 * but we are clearly receiving.
4053 if (!peer->maxPacketSize)
4054 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4056 if (pktsize > peer->maxPacketSize) {
4057 peer->maxPacketSize = pktsize;
4058 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4059 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4060 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4061 rxi_ScheduleGrowMTUEvent(call, 1);
4066 /* Update the outgoing packet skew value to the latest value of
4067 * the peer's incoming packet skew value. The ack packet, of
4068 * course, could arrive out of order, but that won't affect things
4070 peer->outPacketSkew = skew;
4072 /* Check for packets that no longer need to be transmitted, and
4073 * discard them. This only applies to packets positively
4074 * acknowledged as having been sent to the peer's upper level.
4075 * All other packets must be retained. So only packets with
4076 * sequence numbers < ap->firstPacket are candidates. */
4078 clock_GetTime(&now);
4080 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
4081 if (tp->header.seq >= first)
4083 call->tfirst = tp->header.seq + 1;
4085 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4088 rxi_ComputeRoundTripTime(tp, ap, call->conn->peer, &now);
4092 rxi_ComputeRate(call->conn->peer, call, p, np, ap->reason);
4095 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4096 /* XXX Hack. Because we have to release the global rx lock when sending
4097 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4098 * in rxi_Start sending packets out because packets may move to the
4099 * freePacketQueue as result of being here! So we drop these packets until
4100 * we're safely out of the traversing. Really ugly!
4101 * To make it even uglier, if we're using fine grain locking, we can
4102 * set the ack bits in the packets and have rxi_Start remove the packets
4103 * when it's done transmitting.
4105 if (call->flags & RX_CALL_TQ_BUSY) {
4106 #ifdef RX_ENABLE_LOCKS
4107 tp->flags |= RX_PKTFLAG_ACKED;
4108 call->flags |= RX_CALL_TQ_SOME_ACKED;
4109 #else /* RX_ENABLE_LOCKS */
4111 #endif /* RX_ENABLE_LOCKS */
4113 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4116 #ifdef RX_TRACK_PACKETS
4117 tp->flags &= ~RX_PKTFLAG_TQ;
4119 #ifdef RXDEBUG_PACKET
4121 #endif /* RXDEBUG_PACKET */
4122 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4127 /* Give rate detector a chance to respond to ping requests */
4128 if (ap->reason == RX_ACK_PING_RESPONSE) {
4129 rxi_ComputeRate(peer, call, 0, np, ap->reason);
4133 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4135 /* Now go through explicit acks/nacks and record the results in
4136 * the waiting packets. These are packets that can't be released
4137 * yet, even with a positive acknowledge. This positive
4138 * acknowledge only means the packet has been received by the
4139 * peer, not that it will be retained long enough to be sent to
4140 * the peer's upper level. In addition, reset the transmit timers
4141 * of any missing packets (those packets that must be missing
4142 * because this packet was out of sequence) */
4144 call->nSoftAcked = 0;
4145 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
4147 /* Set the acknowledge flag per packet based on the
4148 * information in the ack packet. An acknowlegded packet can
4149 * be downgraded when the server has discarded a packet it
4150 * soacked previously, or when an ack packet is received
4151 * out of sequence. */
4152 if (tp->header.seq < first) {
4153 /* Implicit ack information */
4154 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4157 tp->flags |= RX_PKTFLAG_ACKED;
4158 } else if (tp->header.seq < first + nAcks) {
4159 /* Explicit ack information: set it in the packet appropriately */
4160 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4161 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4163 tp->flags |= RX_PKTFLAG_ACKED;
4165 rxi_ComputeRoundTripTime(tp, ap, call->conn->peer, &now);
4167 rxi_ComputeRate(call->conn->peer, call, tp, np,
4176 } else /* RX_ACK_TYPE_NACK */ {
4177 tp->flags &= ~RX_PKTFLAG_ACKED;
4181 if (tp->flags & RX_PKTFLAG_ACKED) {
4182 tp->flags &= ~RX_PKTFLAG_ACKED;
4188 * Following the suggestion of Phil Kern, we back off the peer's
4189 * timeout value for future packets until a successful response
4190 * is received for an initial transmission.
4192 if (missing && !peer->backedOff) {
4193 struct clock c = peer->timeout;
4194 struct clock max_to = {3, 0};
4196 clock_Add(&peer->timeout, &c);
4197 if (clock_Gt(&peer->timeout, &max_to))
4198 peer->timeout = max_to;
4199 peer->backedOff = 1;
4202 /* If packet isn't yet acked, and it has been transmitted at least
4203 * once, reset retransmit time using latest timeout
4204 * ie, this should readjust the retransmit timer for all outstanding
4205 * packets... So we don't just retransmit when we should know better*/
4207 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
4208 tp->retryTime = tp->timeSent;
4209 clock_Add(&tp->retryTime, &peer->timeout);
4210 /* shift by eight because one quarter-sec ~ 256 milliseconds */
4211 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
4215 /* If the window has been extended by this acknowledge packet,
4216 * then wakeup a sender waiting in alloc for window space, or try
4217 * sending packets now, if he's been sitting on packets due to
4218 * lack of window space */
4219 if (call->tnext < (call->tfirst + call->twind)) {
4220 #ifdef RX_ENABLE_LOCKS
4221 CV_SIGNAL(&call->cv_twind);
4223 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4224 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4225 osi_rxWakeup(&call->twind);
4228 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4229 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4233 /* if the ack packet has a receivelen field hanging off it,
4234 * update our state */
4235 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4238 /* If the ack packet has a "recommended" size that is less than
4239 * what I am using now, reduce my size to match */
4240 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4241 (int)sizeof(afs_int32), &tSize);
4242 tSize = (afs_uint32) ntohl(tSize);
4243 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4245 /* Get the maximum packet size to send to this peer */
4246 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4248 tSize = (afs_uint32) ntohl(tSize);
4249 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4250 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4252 /* sanity check - peer might have restarted with different params.
4253 * If peer says "send less", dammit, send less... Peer should never
4254 * be unable to accept packets of the size that prior AFS versions would
4255 * send without asking. */
4256 if (peer->maxMTU != tSize) {
4257 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4259 peer->maxMTU = tSize;
4260 peer->MTU = MIN(tSize, peer->MTU);
4261 call->MTU = MIN(call->MTU, tSize);
4264 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4267 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4268 (int)sizeof(afs_int32), &tSize);
4269 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4270 if (tSize < call->twind) { /* smaller than our send */
4271 call->twind = tSize; /* window, we must send less... */
4272 call->ssthresh = MIN(call->twind, call->ssthresh);
4273 call->conn->twind[call->channel] = call->twind;
4276 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4277 * network MTU confused with the loopback MTU. Calculate the
4278 * maximum MTU here for use in the slow start code below.
4280 /* Did peer restart with older RX version? */
4281 if (peer->maxDgramPackets > 1) {
4282 peer->maxDgramPackets = 1;
4284 } else if (np->length >=
4285 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4288 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4289 sizeof(afs_int32), &tSize);
4290 tSize = (afs_uint32) ntohl(tSize);
4292 * As of AFS 3.5 we set the send window to match the receive window.
4294 if (tSize < call->twind) {
4295 call->twind = tSize;
4296 call->conn->twind[call->channel] = call->twind;
4297 call->ssthresh = MIN(call->twind, call->ssthresh);
4298 } else if (tSize > call->twind) {
4299 call->twind = tSize;
4300 call->conn->twind[call->channel] = call->twind;
4304 * As of AFS 3.5, a jumbogram is more than one fixed size
4305 * packet transmitted in a single UDP datagram. If the remote
4306 * MTU is smaller than our local MTU then never send a datagram
4307 * larger than the natural MTU.
4310 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4311 (int)sizeof(afs_int32), &tSize);
4312 maxDgramPackets = (afs_uint32) ntohl(tSize);
4313 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4315 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4316 if (maxDgramPackets > 1) {
4317 peer->maxDgramPackets = maxDgramPackets;
4318 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4320 peer->maxDgramPackets = 1;
4321 call->MTU = peer->natMTU;
4323 } else if (peer->maxDgramPackets > 1) {
4324 /* Restarted with lower version of RX */
4325 peer->maxDgramPackets = 1;
4327 } else if (peer->maxDgramPackets > 1
4328 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4329 /* Restarted with lower version of RX */
4330 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4331 peer->natMTU = OLD_MAX_PACKET_SIZE;
4332 peer->MTU = OLD_MAX_PACKET_SIZE;
4333 peer->maxDgramPackets = 1;
4334 peer->nDgramPackets = 1;
4336 call->MTU = OLD_MAX_PACKET_SIZE;
4341 * Calculate how many datagrams were successfully received after
4342 * the first missing packet and adjust the negative ack counter
4347 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4348 if (call->nNacks < nNacked) {
4349 call->nNacks = nNacked;
4352 call->nAcks += newAckCount;
4356 if (call->flags & RX_CALL_FAST_RECOVER) {
4358 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4360 call->flags &= ~RX_CALL_FAST_RECOVER;
4361 call->cwind = call->nextCwind;
4362 call->nextCwind = 0;
4365 call->nCwindAcks = 0;
4366 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4367 /* Three negative acks in a row trigger congestion recovery */
4368 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4369 MUTEX_EXIT(&peer->peer_lock);
4370 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4371 /* someone else is waiting to start recovery */
4374 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4375 rxi_WaitforTQBusy(call);
4376 MUTEX_ENTER(&peer->peer_lock);
4377 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4378 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4379 call->flags |= RX_CALL_FAST_RECOVER;
4380 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4382 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4383 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4384 call->nextCwind = call->ssthresh;
4387 peer->MTU = call->MTU;
4388 peer->cwind = call->nextCwind;
4389 peer->nDgramPackets = call->nDgramPackets;
4391 call->congestSeq = peer->congestSeq;
4392 /* Reset the resend times on the packets that were nacked
4393 * so we will retransmit as soon as the window permits*/
4394 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4396 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4397 clock_Zero(&tp->retryTime);
4399 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4404 /* If cwind is smaller than ssthresh, then increase
4405 * the window one packet for each ack we receive (exponential
4407 * If cwind is greater than or equal to ssthresh then increase
4408 * the congestion window by one packet for each cwind acks we
4409 * receive (linear growth). */
4410 if (call->cwind < call->ssthresh) {
4412 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4413 call->nCwindAcks = 0;
4415 call->nCwindAcks += newAckCount;
4416 if (call->nCwindAcks >= call->cwind) {
4417 call->nCwindAcks = 0;
4418 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4422 * If we have received several acknowledgements in a row then
4423 * it is time to increase the size of our datagrams
4425 if ((int)call->nAcks > rx_nDgramThreshold) {
4426 if (peer->maxDgramPackets > 1) {
4427 if (call->nDgramPackets < peer->maxDgramPackets) {
4428 call->nDgramPackets++;
4430 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4431 } else if (call->MTU < peer->maxMTU) {
4432 /* don't upgrade if we can't handle it */
4433 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4434 call->MTU = peer->ifMTU;
4436 call->MTU += peer->natMTU;
4437 call->MTU = MIN(call->MTU, peer->maxMTU);
4444 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4446 /* Servers need to hold the call until all response packets have
4447 * been acknowledged. Soft acks are good enough since clients
4448 * are not allowed to clear their receive queues. */
4449 if (call->state == RX_STATE_HOLD
4450 && call->tfirst + call->nSoftAcked >= call->tnext) {
4451 call->state = RX_STATE_DALLY;
4452 rxi_ClearTransmitQueue(call, 0);
4453 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4454 } else if (!queue_IsEmpty(&call->tq)) {
4455 rxi_Start(0, call, 0, istack);
4460 /* Received a response to a challenge packet */
4462 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4463 struct rx_packet *np, int istack)
4467 /* Ignore the packet if we're the client */
4468 if (conn->type == RX_CLIENT_CONNECTION)
4471 /* If already authenticated, ignore the packet (it's probably a retry) */
4472 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4475 /* Otherwise, have the security object evaluate the response packet */
4476 error = RXS_CheckResponse(conn->securityObject, conn, np);
4478 /* If the response is invalid, reset the connection, sending
4479 * an abort to the peer */
4483 rxi_ConnectionError(conn, error);
4484 MUTEX_ENTER(&conn->conn_data_lock);
4485 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4486 MUTEX_EXIT(&conn->conn_data_lock);
4489 /* If the response is valid, any calls waiting to attach
4490 * servers can now do so */
4493 for (i = 0; i < RX_MAXCALLS; i++) {
4494 struct rx_call *call = conn->call[i];
4496 MUTEX_ENTER(&call->lock);
4497 if (call->state == RX_STATE_PRECALL)
4498 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4499 /* tnop can be null if newcallp is null */
4500 MUTEX_EXIT(&call->lock);
4504 /* Update the peer reachability information, just in case
4505 * some calls went into attach-wait while we were waiting
4506 * for authentication..
4508 rxi_UpdatePeerReach(conn, NULL);
4513 /* A client has received an authentication challenge: the security
4514 * object is asked to cough up a respectable response packet to send
4515 * back to the server. The server is responsible for retrying the
4516 * challenge if it fails to get a response. */
4519 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4520 struct rx_packet *np, int istack)
4524 /* Ignore the challenge if we're the server */
4525 if (conn->type == RX_SERVER_CONNECTION)
4528 /* Ignore the challenge if the connection is otherwise idle; someone's
4529 * trying to use us as an oracle. */
4530 if (!rxi_HasActiveCalls(conn))
4533 /* Send the security object the challenge packet. It is expected to fill
4534 * in the response. */
4535 error = RXS_GetResponse(conn->securityObject, conn, np);
4537 /* If the security object is unable to return a valid response, reset the
4538 * connection and send an abort to the peer. Otherwise send the response
4539 * packet to the peer connection. */
4541 rxi_ConnectionError(conn, error);
4542 MUTEX_ENTER(&conn->conn_data_lock);
4543 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4544 MUTEX_EXIT(&conn->conn_data_lock);
4546 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4547 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4553 /* Find an available server process to service the current request in
4554 * the given call structure. If one isn't available, queue up this
4555 * call so it eventually gets one */
4557 rxi_AttachServerProc(struct rx_call *call,
4558 osi_socket socket, int *tnop,
4559 struct rx_call **newcallp)
4561 struct rx_serverQueueEntry *sq;
4562 struct rx_service *service = call->conn->service;
4565 /* May already be attached */
4566 if (call->state == RX_STATE_ACTIVE)
4569 MUTEX_ENTER(&rx_serverPool_lock);
4571 haveQuota = QuotaOK(service);
4572 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4573 /* If there are no processes available to service this call,
4574 * put the call on the incoming call queue (unless it's
4575 * already on the queue).
4577 #ifdef RX_ENABLE_LOCKS
4579 ReturnToServerPool(service);
4580 #endif /* RX_ENABLE_LOCKS */
4582 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4583 call->flags |= RX_CALL_WAIT_PROC;
4584 rx_atomic_inc(&rx_nWaiting);
4585 rx_atomic_inc(&rx_nWaited);
4586 rxi_calltrace(RX_CALL_ARRIVAL, call);
4587 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4588 queue_Append(&rx_incomingCallQueue, call);
4591 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4593 /* If hot threads are enabled, and both newcallp and sq->socketp
4594 * are non-null, then this thread will process the call, and the
4595 * idle server thread will start listening on this threads socket.
4598 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4601 *sq->socketp = socket;
4602 clock_GetTime(&call->startTime);
4603 MUTEX_ENTER(&rx_refcnt_mutex);
4604 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4605 MUTEX_EXIT(&rx_refcnt_mutex);
4609 if (call->flags & RX_CALL_WAIT_PROC) {
4610 /* Conservative: I don't think this should happen */
4611 call->flags &= ~RX_CALL_WAIT_PROC;
4612 if (queue_IsOnQueue(call)) {
4615 rx_atomic_dec(&rx_nWaiting);
4618 call->state = RX_STATE_ACTIVE;
4619 call->mode = RX_MODE_RECEIVING;
4620 #ifdef RX_KERNEL_TRACE
4622 int glockOwner = ISAFS_GLOCK();
4625 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4626 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4632 if (call->flags & RX_CALL_CLEARED) {
4633 /* send an ack now to start the packet flow up again */
4634 call->flags &= ~RX_CALL_CLEARED;
4635 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4637 #ifdef RX_ENABLE_LOCKS
4640 service->nRequestsRunning++;
4641 MUTEX_ENTER(&rx_quota_mutex);
4642 if (service->nRequestsRunning <= service->minProcs)
4645 MUTEX_EXIT(&rx_quota_mutex);
4649 MUTEX_EXIT(&rx_serverPool_lock);
4652 /* Delay the sending of an acknowledge event for a short while, while
4653 * a new call is being prepared (in the case of a client) or a reply
4654 * is being prepared (in the case of a server). Rather than sending
4655 * an ack packet, an ACKALL packet is sent. */
4657 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4659 #ifdef RX_ENABLE_LOCKS
4661 MUTEX_ENTER(&call->lock);
4662 call->delayedAckEvent = NULL;
4663 MUTEX_ENTER(&rx_refcnt_mutex);
4664 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4665 MUTEX_EXIT(&rx_refcnt_mutex);
4667 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4668 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4670 MUTEX_EXIT(&call->lock);
4671 #else /* RX_ENABLE_LOCKS */
4673 call->delayedAckEvent = NULL;
4674 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4675 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4676 #endif /* RX_ENABLE_LOCKS */
4680 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4682 struct rx_call *call = arg1;
4683 #ifdef RX_ENABLE_LOCKS
4685 MUTEX_ENTER(&call->lock);
4686 if (event == call->delayedAckEvent)
4687 call->delayedAckEvent = NULL;
4688 MUTEX_ENTER(&rx_refcnt_mutex);
4689 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4690 MUTEX_EXIT(&rx_refcnt_mutex);
4692 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4694 MUTEX_EXIT(&call->lock);
4695 #else /* RX_ENABLE_LOCKS */
4697 call->delayedAckEvent = NULL;
4698 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4699 #endif /* RX_ENABLE_LOCKS */
4703 #ifdef RX_ENABLE_LOCKS
4704 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4705 * clearing them out.
4708 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4710 struct rx_packet *p, *tp;
4713 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4714 p->flags |= RX_PKTFLAG_ACKED;
4718 call->flags |= RX_CALL_TQ_CLEARME;
4719 call->flags |= RX_CALL_TQ_SOME_ACKED;
4722 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4723 call->tfirst = call->tnext;
4724 call->nSoftAcked = 0;
4726 if (call->flags & RX_CALL_FAST_RECOVER) {
4727 call->flags &= ~RX_CALL_FAST_RECOVER;
4728 call->cwind = call->nextCwind;
4729 call->nextCwind = 0;
4732 CV_SIGNAL(&call->cv_twind);
4734 #endif /* RX_ENABLE_LOCKS */
4736 /* Clear out the transmit queue for the current call (all packets have
4737 * been received by peer) */
4739 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4741 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4742 struct rx_packet *p, *tp;
4744 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4746 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4747 p->flags |= RX_PKTFLAG_ACKED;
4751 call->flags |= RX_CALL_TQ_CLEARME;
4752 call->flags |= RX_CALL_TQ_SOME_ACKED;
4755 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4756 #ifdef RXDEBUG_PACKET
4758 #endif /* RXDEBUG_PACKET */
4759 rxi_FreePackets(0, &call->tq);
4760 rxi_WakeUpTransmitQueue(call);
4761 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4762 call->flags &= ~RX_CALL_TQ_CLEARME;
4764 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4766 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4767 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4768 call->nSoftAcked = 0;
4770 if (call->flags & RX_CALL_FAST_RECOVER) {
4771 call->flags &= ~RX_CALL_FAST_RECOVER;
4772 call->cwind = call->nextCwind;
4774 #ifdef RX_ENABLE_LOCKS
4775 CV_SIGNAL(&call->cv_twind);
4777 osi_rxWakeup(&call->twind);
4782 rxi_ClearReceiveQueue(struct rx_call *call)
4784 if (queue_IsNotEmpty(&call->rq)) {
4787 count = rxi_FreePackets(0, &call->rq);
4788 rx_packetReclaims += count;
4789 #ifdef RXDEBUG_PACKET
4791 if ( call->rqc != 0 )
4792 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
4794 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4796 if (call->state == RX_STATE_PRECALL) {
4797 call->flags |= RX_CALL_CLEARED;
4801 /* Send an abort packet for the specified call */
4803 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
4804 int istack, int force)
4807 struct clock when, now;
4812 /* Clients should never delay abort messages */
4813 if (rx_IsClientConn(call->conn))
4816 if (call->abortCode != call->error) {
4817 call->abortCode = call->error;
4818 call->abortCount = 0;
4821 if (force || rxi_callAbortThreshhold == 0
4822 || call->abortCount < rxi_callAbortThreshhold) {
4823 if (call->delayedAbortEvent) {
4824 rxevent_Cancel(call->delayedAbortEvent, call,
4825 RX_CALL_REFCOUNT_ABORT);
4827 error = htonl(call->error);
4830 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4831 (char *)&error, sizeof(error), istack);
4832 } else if (!call->delayedAbortEvent) {
4833 clock_GetTime(&now);
4835 clock_Addmsec(&when, rxi_callAbortDelay);
4836 MUTEX_ENTER(&rx_refcnt_mutex);
4837 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4838 MUTEX_EXIT(&rx_refcnt_mutex);
4839 call->delayedAbortEvent =
4840 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4845 /* Send an abort packet for the specified connection. Packet is an
4846 * optional pointer to a packet that can be used to send the abort.
4847 * Once the number of abort messages reaches the threshhold, an
4848 * event is scheduled to send the abort. Setting the force flag
4849 * overrides sending delayed abort messages.
4851 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4852 * to send the abort packet.
4855 rxi_SendConnectionAbort(struct rx_connection *conn,
4856 struct rx_packet *packet, int istack, int force)
4859 struct clock when, now;
4864 /* Clients should never delay abort messages */
4865 if (rx_IsClientConn(conn))
4868 if (force || rxi_connAbortThreshhold == 0
4869 || conn->abortCount < rxi_connAbortThreshhold) {
4870 if (conn->delayedAbortEvent) {
4871 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4873 error = htonl(conn->error);
4875 MUTEX_EXIT(&conn->conn_data_lock);
4877 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4878 RX_PACKET_TYPE_ABORT, (char *)&error,
4879 sizeof(error), istack);
4880 MUTEX_ENTER(&conn->conn_data_lock);
4881 } else if (!conn->delayedAbortEvent) {
4882 clock_GetTime(&now);
4884 clock_Addmsec(&when, rxi_connAbortDelay);
4885 conn->delayedAbortEvent =
4886 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4891 /* Associate an error all of the calls owned by a connection. Called
4892 * with error non-zero. This is only for really fatal things, like
4893 * bad authentication responses. The connection itself is set in
4894 * error at this point, so that future packets received will be
4897 rxi_ConnectionError(struct rx_connection *conn,
4903 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
4905 MUTEX_ENTER(&conn->conn_data_lock);
4906 if (conn->challengeEvent)
4907 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4908 if (conn->natKeepAliveEvent)
4909 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
4910 if (conn->checkReachEvent) {
4911 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4912 conn->checkReachEvent = 0;
4913 conn->flags &= ~RX_CONN_ATTACHWAIT;
4914 MUTEX_ENTER(&rx_refcnt_mutex);
4916 MUTEX_EXIT(&rx_refcnt_mutex);
4918 MUTEX_EXIT(&conn->conn_data_lock);
4919 for (i = 0; i < RX_MAXCALLS; i++) {
4920 struct rx_call *call = conn->call[i];
4922 MUTEX_ENTER(&call->lock);
4923 rxi_CallError(call, error);
4924 MUTEX_EXIT(&call->lock);
4927 conn->error = error;
4928 if (rx_stats_active)
4929 rx_atomic_inc(&rx_stats.fatalErrors);
4934 * Interrupt an in-progress call with the specified error and wakeup waiters.
4936 * @param[in] call The call to interrupt
4937 * @param[in] error The error code to send to the peer
4940 rx_InterruptCall(struct rx_call *call, afs_int32 error)
4942 MUTEX_ENTER(&call->lock);
4943 rxi_CallError(call, error);
4944 rxi_SendCallAbort(call, NULL, 0, 1);
4945 MUTEX_EXIT(&call->lock);
4949 rxi_CallError(struct rx_call *call, afs_int32 error)
4952 osirx_AssertMine(&call->lock, "rxi_CallError");
4954 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
4956 error = call->error;
4958 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4959 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4960 rxi_ResetCall(call, 0);
4963 rxi_ResetCall(call, 0);
4965 call->error = error;
4968 /* Reset various fields in a call structure, and wakeup waiting
4969 * processes. Some fields aren't changed: state & mode are not
4970 * touched (these must be set by the caller), and bufptr, nLeft, and
4971 * nFree are not reset, since these fields are manipulated by
4972 * unprotected macros, and may only be reset by non-interrupting code.
4975 /* this code requires that call->conn be set properly as a pre-condition. */
4976 #endif /* ADAPT_WINDOW */
4979 rxi_ResetCall(struct rx_call *call, int newcall)
4982 struct rx_peer *peer;
4983 struct rx_packet *packet;
4985 osirx_AssertMine(&call->lock, "rxi_ResetCall");
4987 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
4989 /* Notify anyone who is waiting for asynchronous packet arrival */
4990 if (call->arrivalProc) {
4991 (*call->arrivalProc) (call, call->arrivalProcHandle,
4992 call->arrivalProcArg);
4993 call->arrivalProc = (void (*)())0;
4996 if (call->delayedAbortEvent) {
4997 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4998 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5000 rxi_SendCallAbort(call, packet, 0, 1);
5001 rxi_FreePacket(packet);
5006 * Update the peer with the congestion information in this call
5007 * so other calls on this connection can pick up where this call
5008 * left off. If the congestion sequence numbers don't match then
5009 * another call experienced a retransmission.
5011 peer = call->conn->peer;
5012 MUTEX_ENTER(&peer->peer_lock);
5014 if (call->congestSeq == peer->congestSeq) {
5015 peer->cwind = MAX(peer->cwind, call->cwind);
5016 peer->MTU = MAX(peer->MTU, call->MTU);
5017 peer->nDgramPackets =
5018 MAX(peer->nDgramPackets, call->nDgramPackets);
5021 call->abortCode = 0;
5022 call->abortCount = 0;
5024 if (peer->maxDgramPackets > 1) {
5025 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5027 call->MTU = peer->MTU;
5029 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5030 call->ssthresh = rx_maxSendWindow;
5031 call->nDgramPackets = peer->nDgramPackets;
5032 call->congestSeq = peer->congestSeq;
5033 MUTEX_EXIT(&peer->peer_lock);
5035 flags = call->flags;
5036 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5037 rxi_WaitforTQBusy(call);
5038 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5040 rxi_ClearTransmitQueue(call, 1);
5041 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5042 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5046 rxi_ClearReceiveQueue(call);
5047 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5051 call->twind = call->conn->twind[call->channel];
5052 call->rwind = call->conn->rwind[call->channel];
5053 call->nSoftAcked = 0;
5054 call->nextCwind = 0;
5057 call->nCwindAcks = 0;
5058 call->nSoftAcks = 0;
5059 call->nHardAcks = 0;
5061 call->tfirst = call->rnext = call->tnext = 1;
5064 call->lastAcked = 0;
5065 call->localStatus = call->remoteStatus = 0;
5067 if (flags & RX_CALL_READER_WAIT) {
5068 #ifdef RX_ENABLE_LOCKS
5069 CV_BROADCAST(&call->cv_rq);
5071 osi_rxWakeup(&call->rq);
5074 if (flags & RX_CALL_WAIT_PACKETS) {
5075 MUTEX_ENTER(&rx_freePktQ_lock);
5076 rxi_PacketsUnWait(); /* XXX */
5077 MUTEX_EXIT(&rx_freePktQ_lock);
5079 #ifdef RX_ENABLE_LOCKS
5080 CV_SIGNAL(&call->cv_twind);
5082 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5083 osi_rxWakeup(&call->twind);
5086 #ifdef RX_ENABLE_LOCKS
5087 /* The following ensures that we don't mess with any queue while some
5088 * other thread might also be doing so. The call_queue_lock field is
5089 * is only modified under the call lock. If the call is in the process
5090 * of being removed from a queue, the call is not locked until the
5091 * the queue lock is dropped and only then is the call_queue_lock field
5092 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5093 * Note that any other routine which removes a call from a queue has to
5094 * obtain the queue lock before examing the queue and removing the call.
5096 if (call->call_queue_lock) {
5097 MUTEX_ENTER(call->call_queue_lock);
5098 if (queue_IsOnQueue(call)) {
5100 if (flags & RX_CALL_WAIT_PROC) {
5101 rx_atomic_dec(&rx_nWaiting);
5104 MUTEX_EXIT(call->call_queue_lock);
5105 CLEAR_CALL_QUEUE_LOCK(call);
5107 #else /* RX_ENABLE_LOCKS */
5108 if (queue_IsOnQueue(call)) {
5110 if (flags & RX_CALL_WAIT_PROC)
5111 rx_atomic_dec(&rx_nWaiting);
5113 #endif /* RX_ENABLE_LOCKS */
5115 rxi_KeepAliveOff(call);
5116 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5119 /* Send an acknowledge for the indicated packet (seq,serial) of the
5120 * indicated call, for the indicated reason (reason). This
5121 * acknowledge will specifically acknowledge receiving the packet, and
5122 * will also specify which other packets for this call have been
5123 * received. This routine returns the packet that was used to the
5124 * caller. The caller is responsible for freeing it or re-using it.
5125 * This acknowledgement also returns the highest sequence number
5126 * actually read out by the higher level to the sender; the sender
5127 * promises to keep around packets that have not been read by the
5128 * higher level yet (unless, of course, the sender decides to abort
5129 * the call altogether). Any of p, seq, serial, pflags, or reason may
5130 * be set to zero without ill effect. That is, if they are zero, they
5131 * will not convey any information.
5132 * NOW there is a trailer field, after the ack where it will safely be
5133 * ignored by mundanes, which indicates the maximum size packet this
5134 * host can swallow. */
5136 struct rx_packet *optionalPacket; use to send ack (or null)
5137 int seq; Sequence number of the packet we are acking
5138 int serial; Serial number of the packet
5139 int pflags; Flags field from packet header
5140 int reason; Reason an acknowledge was prompted
5144 rxi_SendAck(struct rx_call *call,
5145 struct rx_packet *optionalPacket, int serial, int reason,
5148 struct rx_ackPacket *ap;
5149 struct rx_packet *rqp;
5150 struct rx_packet *nxp; /* For queue_Scan */
5151 struct rx_packet *p;
5154 afs_uint32 padbytes = 0;
5155 #ifdef RX_ENABLE_TSFPQ
5156 struct rx_ts_info_t * rx_ts_info;
5160 * Open the receive window once a thread starts reading packets
5162 if (call->rnext > 1) {
5163 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5166 /* Don't attempt to grow MTU if this is a critical ping */
5167 if (reason == RX_ACK_MTU) {
5168 /* keep track of per-call attempts, if we're over max, do in small
5169 * otherwise in larger? set a size to increment by, decrease
5172 if (call->conn->peer->maxPacketSize &&
5173 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5175 padbytes = call->conn->peer->maxPacketSize+16;
5177 padbytes = call->conn->peer->maxMTU + 128;
5179 /* do always try a minimum size ping */
5180 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5182 /* subtract the ack payload */
5183 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5184 reason = RX_ACK_PING;
5187 call->nHardAcks = 0;
5188 call->nSoftAcks = 0;
5189 if (call->rnext > call->lastAcked)
5190 call->lastAcked = call->rnext;
5194 rx_computelen(p, p->length); /* reset length, you never know */
5195 } /* where that's been... */
5196 #ifdef RX_ENABLE_TSFPQ
5198 RX_TS_INFO_GET(rx_ts_info);
5199 if ((p = rx_ts_info->local_special_packet)) {
5200 rx_computelen(p, p->length);
5201 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5202 rx_ts_info->local_special_packet = p;
5203 } else { /* We won't send the ack, but don't panic. */
5204 return optionalPacket;
5208 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5209 /* We won't send the ack, but don't panic. */
5210 return optionalPacket;
5215 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5218 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5219 #ifndef RX_ENABLE_TSFPQ
5220 if (!optionalPacket)
5223 return optionalPacket;
5225 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5226 if (rx_Contiguous(p) < templ) {
5227 #ifndef RX_ENABLE_TSFPQ
5228 if (!optionalPacket)
5231 return optionalPacket;
5236 /* MTUXXX failing to send an ack is very serious. We should */
5237 /* try as hard as possible to send even a partial ack; it's */
5238 /* better than nothing. */
5239 ap = (struct rx_ackPacket *)rx_DataOf(p);
5240 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5241 ap->reason = reason;
5243 /* The skew computation used to be bogus, I think it's better now. */
5244 /* We should start paying attention to skew. XXX */
5245 ap->serial = htonl(serial);
5246 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5248 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
5249 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5251 /* No fear of running out of ack packet here because there can only be at most
5252 * one window full of unacknowledged packets. The window size must be constrained
5253 * to be less than the maximum ack size, of course. Also, an ack should always
5254 * fit into a single packet -- it should not ever be fragmented. */
5255 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5256 if (!rqp || !call->rq.next
5257 || (rqp->header.seq > (call->rnext + call->rwind))) {
5258 #ifndef RX_ENABLE_TSFPQ
5259 if (!optionalPacket)
5262 rxi_CallError(call, RX_CALL_DEAD);
5263 return optionalPacket;
5266 while (rqp->header.seq > call->rnext + offset)
5267 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5268 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5270 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5271 #ifndef RX_ENABLE_TSFPQ
5272 if (!optionalPacket)
5275 rxi_CallError(call, RX_CALL_DEAD);
5276 return optionalPacket;
5281 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5283 /* these are new for AFS 3.3 */
5284 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5285 templ = htonl(templ);
5286 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5287 templ = htonl(call->conn->peer->ifMTU);
5288 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5289 sizeof(afs_int32), &templ);
5291 /* new for AFS 3.4 */
5292 templ = htonl(call->rwind);
5293 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5294 sizeof(afs_int32), &templ);
5296 /* new for AFS 3.5 */
5297 templ = htonl(call->conn->peer->ifDgramPackets);
5298 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5299 sizeof(afs_int32), &templ);
5301 p->header.serviceId = call->conn->serviceId;
5302 p->header.cid = (call->conn->cid | call->channel);
5303 p->header.callNumber = *call->callNumber;
5305 p->header.securityIndex = call->conn->securityIndex;
5306 p->header.epoch = call->conn->epoch;
5307 p->header.type = RX_PACKET_TYPE_ACK;
5308 p->header.flags = RX_SLOW_START_OK;
5309 if (reason == RX_ACK_PING) {
5310 p->header.flags |= RX_REQUEST_ACK;
5312 clock_GetTime(&call->pingRequestTime);
5315 p->length = padbytes +
5316 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5319 /* not fast but we can potentially use this if truncated
5320 * fragments are delivered to figure out the mtu.
5322 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5323 sizeof(afs_int32), sizeof(afs_int32),
5327 if (call->conn->type == RX_CLIENT_CONNECTION)
5328 p->header.flags |= RX_CLIENT_INITIATED;
5332 if (rxdebug_active) {
5336 len = _snprintf(msg, sizeof(msg),
5337 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5338 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5339 ntohl(ap->serial), ntohl(ap->previousPacket),
5340 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5341 ap->nAcks, ntohs(ap->bufferSpace) );
5345 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5346 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5350 OutputDebugString(msg);
5352 #else /* AFS_NT40_ENV */
5354 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5355 ap->reason, ntohl(ap->previousPacket),
5356 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5358 for (offset = 0; offset < ap->nAcks; offset++)
5359 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5364 #endif /* AFS_NT40_ENV */
5367 int i, nbytes = p->length;
5369 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5370 if (nbytes <= p->wirevec[i].iov_len) {
5373 savelen = p->wirevec[i].iov_len;
5375 p->wirevec[i].iov_len = nbytes;
5377 rxi_Send(call, p, istack);
5378 p->wirevec[i].iov_len = savelen;
5382 nbytes -= p->wirevec[i].iov_len;
5385 if (rx_stats_active)
5386 rx_atomic_inc(&rx_stats.ackPacketsSent);
5387 #ifndef RX_ENABLE_TSFPQ
5388 if (!optionalPacket)
5391 return optionalPacket; /* Return packet for re-use by caller */
5395 struct rx_packet **list;
5400 /* Send all of the packets in the list in single datagram */
5402 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5403 int istack, int moreFlag)
5408 struct clock now, retryTime;
5409 struct rx_connection *conn = call->conn;
5410 struct rx_peer *peer = conn->peer;
5412 MUTEX_ENTER(&peer->peer_lock);
5413 peer->nSent += xmit->len;
5414 if (xmit->resending)
5415 peer->reSends += xmit->len;
5416 retryTime = peer->timeout;
5417 MUTEX_EXIT(&peer->peer_lock);
5419 if (rx_stats_active) {
5420 if (xmit->resending)
5421 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5423 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5426 clock_GetTime(&now);
5427 clock_Add(&retryTime, &now);
5429 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5433 /* Set the packet flags and schedule the resend events */
5434 /* Only request an ack for the last packet in the list */
5435 for (i = 0; i < xmit->len; i++) {
5436 struct rx_packet *packet = xmit->list[i];
5438 packet->retryTime = retryTime;
5439 if (packet->header.serial) {
5440 /* Exponentially backoff retry times */
5441 if (packet->backoff < MAXBACKOFF) {
5442 /* so it can't stay == 0 */
5443 packet->backoff = (packet->backoff << 1) + 1;
5446 clock_Addmsec(&(packet->retryTime),
5447 ((afs_uint32) packet->backoff) << 8);
5450 /* Wait a little extra for the ack on the last packet */
5452 && !(packet->header.flags & RX_CLIENT_INITIATED)) {
5453 clock_Addmsec(&(packet->retryTime), 400);
5456 /* Record the time sent */
5457 packet->timeSent = now;
5459 /* Ask for an ack on retransmitted packets, on every other packet
5460 * if the peer doesn't support slow start. Ask for an ack on every
5461 * packet until the congestion window reaches the ack rate. */
5462 if (packet->header.serial) {
5465 /* improved RTO calculation- not Karn */
5466 packet->firstSent = now;
5467 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5468 || (!(call->flags & RX_CALL_SLOW_START_OK)
5469 && (packet->header.seq & 1)))) {
5474 /* Tag this packet as not being the last in this group,
5475 * for the receiver's benefit */
5476 if (i < xmit->len - 1 || moreFlag) {
5477 packet->header.flags |= RX_MORE_PACKETS;
5482 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5485 /* Since we're about to send a data packet to the peer, it's
5486 * safe to nuke any scheduled end-of-packets ack */
5487 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5489 MUTEX_EXIT(&call->lock);
5490 MUTEX_ENTER(&rx_refcnt_mutex);
5491 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5492 MUTEX_EXIT(&rx_refcnt_mutex);
5493 if (xmit->len > 1) {
5494 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5496 rxi_SendPacket(call, conn, xmit->list[0], istack);
5498 MUTEX_ENTER(&call->lock);
5499 MUTEX_ENTER(&rx_refcnt_mutex);
5500 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5501 MUTEX_EXIT(&rx_refcnt_mutex);
5503 /* Update last send time for this call (for keep-alive
5504 * processing), and for the connection (so that we can discover
5505 * idle connections) */
5506 conn->lastSendTime = call->lastSendTime = clock_Sec();
5507 /* Let a set of retransmits trigger an idle timeout */
5508 if (!xmit->resending)
5509 call->lastSendData = call->lastSendTime;
5512 /* When sending packets we need to follow these rules:
5513 * 1. Never send more than maxDgramPackets in a jumbogram.
5514 * 2. Never send a packet with more than two iovecs in a jumbogram.
5515 * 3. Never send a retransmitted packet in a jumbogram.
5516 * 4. Never send more than cwind/4 packets in a jumbogram
5517 * We always keep the last list we should have sent so we
5518 * can set the RX_MORE_PACKETS flags correctly.
5522 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5526 struct xmitlist working;
5527 struct xmitlist last;
5529 struct rx_peer *peer = call->conn->peer;
5530 int morePackets = 0;
5532 memset(&last, 0, sizeof(struct xmitlist));
5533 working.list = &list[0];
5535 working.resending = 0;
5537 for (i = 0; i < len; i++) {
5538 /* Does the current packet force us to flush the current list? */
5540 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5541 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5543 /* This sends the 'last' list and then rolls the current working
5544 * set into the 'last' one, and resets the working set */
5547 rxi_SendList(call, &last, istack, 1);
5548 /* If the call enters an error state stop sending, or if
5549 * we entered congestion recovery mode, stop sending */
5550 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5555 working.resending = 0;
5556 working.list = &list[i];
5558 /* Add the current packet to the list if it hasn't been acked.
5559 * Otherwise adjust the list pointer to skip the current packet. */
5560 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5563 if (list[i]->header.serial)
5564 working.resending = 1;
5566 /* Do we need to flush the list? */
5567 if (working.len >= (int)peer->maxDgramPackets
5568 || working.len >= (int)call->nDgramPackets
5569 || working.len >= (int)call->cwind
5570 || list[i]->header.serial
5571 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5573 rxi_SendList(call, &last, istack, 1);
5574 /* If the call enters an error state stop sending, or if
5575 * we entered congestion recovery mode, stop sending */
5577 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5582 working.resending = 0;
5583 working.list = &list[i + 1];
5586 if (working.len != 0) {
5587 osi_Panic("rxi_SendList error");
5589 working.list = &list[i + 1];
5593 /* Send the whole list when the call is in receive mode, when
5594 * the call is in eof mode, when we are in fast recovery mode,
5595 * and when we have the last packet */
5596 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5597 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5598 || (call->flags & RX_CALL_FAST_RECOVER)) {
5599 /* Check for the case where the current list contains
5600 * an acked packet. Since we always send retransmissions
5601 * in a separate packet, we only need to check the first
5602 * packet in the list */
5603 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5607 rxi_SendList(call, &last, istack, morePackets);
5608 /* If the call enters an error state stop sending, or if
5609 * we entered congestion recovery mode, stop sending */
5610 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5614 rxi_SendList(call, &working, istack, 0);
5616 } else if (last.len > 0) {
5617 rxi_SendList(call, &last, istack, 0);
5618 /* Packets which are in 'working' are not sent by this call */
5622 #ifdef RX_ENABLE_LOCKS
5623 /* Call rxi_Start, below, but with the call lock held. */
5625 rxi_StartUnlocked(struct rxevent *event,
5626 void *arg0, void *arg1, int istack)
5628 struct rx_call *call = arg0;
5630 MUTEX_ENTER(&call->lock);
5631 rxi_Start(event, call, arg1, istack);
5632 MUTEX_EXIT(&call->lock);
5634 #endif /* RX_ENABLE_LOCKS */
5636 /* This routine is called when new packets are readied for
5637 * transmission and when retransmission may be necessary, or when the
5638 * transmission window or burst count are favourable. This should be
5639 * better optimized for new packets, the usual case, now that we've
5640 * got rid of queues of send packets. XXXXXXXXXXX */
5642 rxi_Start(struct rxevent *event,
5643 void *arg0, void *arg1, int istack)
5645 struct rx_call *call = arg0;
5647 struct rx_packet *p;
5648 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5649 struct clock now, usenow, retryTime;
5654 /* If rxi_Start is being called as a result of a resend event,
5655 * then make sure that the event pointer is removed from the call
5656 * structure, since there is no longer a per-call retransmission
5658 if (event && event == call->resendEvent) {
5659 MUTEX_ENTER(&rx_refcnt_mutex);
5660 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5661 MUTEX_EXIT(&rx_refcnt_mutex);
5662 call->resendEvent = NULL;
5663 if (queue_IsEmpty(&call->tq)) {
5670 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5671 if (rx_stats_active)
5672 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5677 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5679 clock_GetTime(&now);
5682 /* Send (or resend) any packets that need it, subject to
5683 * window restrictions and congestion burst control
5684 * restrictions. Ask for an ack on the last packet sent in
5685 * this burst. For now, we're relying upon the window being
5686 * considerably bigger than the largest number of packets that
5687 * are typically sent at once by one initial call to
5688 * rxi_Start. This is probably bogus (perhaps we should ask
5689 * for an ack when we're half way through the current
5690 * window?). Also, for non file transfer applications, this
5691 * may end up asking for an ack for every packet. Bogus. XXXX
5694 * But check whether we're here recursively, and let the other guy
5697 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5698 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5699 call->flags |= RX_CALL_TQ_BUSY;
5701 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5703 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5704 call->flags &= ~RX_CALL_NEED_START;
5705 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5707 maxXmitPackets = MIN(call->twind, call->cwind);
5708 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5709 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5710 /* We shouldn't be sending packets if a thread is waiting
5711 * to initiate congestion recovery */
5712 dpf(("call %d waiting to initiate fast recovery\n",
5713 *(call->callNumber)));
5717 && (call->flags & RX_CALL_FAST_RECOVER)) {
5718 /* Only send one packet during fast recovery */
5719 dpf(("call %d restricted to one packet per send during fast recovery\n",
5720 *(call->callNumber)));
5723 #ifdef RX_TRACK_PACKETS
5724 if ((p->flags & RX_PKTFLAG_FREE)
5725 || (!queue_IsEnd(&call->tq, nxp)
5726 && (nxp->flags & RX_PKTFLAG_FREE))
5727 || (p == (struct rx_packet *)&rx_freePacketQueue)
5728 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5729 osi_Panic("rxi_Start: xmit queue clobbered");
5732 if (p->flags & RX_PKTFLAG_ACKED) {
5733 /* Since we may block, don't trust this */
5734 usenow.sec = usenow.usec = 0;
5735 if (rx_stats_active)
5736 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
5737 continue; /* Ignore this packet if it has been acknowledged */
5740 /* Turn off all flags except these ones, which are the same
5741 * on each transmission */
5742 p->header.flags &= RX_PRESET_FLAGS;
5744 if (p->header.seq >=
5745 call->tfirst + MIN((int)call->twind,
5746 (int)(call->nSoftAcked +
5748 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5749 /* Note: if we're waiting for more window space, we can
5750 * still send retransmits; hence we don't return here, but
5751 * break out to schedule a retransmit event */
5752 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
5753 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
5758 /* Transmit the packet if it needs to be sent. */
5759 if (!clock_Lt(&now, &p->retryTime)) {
5760 if (nXmitPackets == maxXmitPackets) {
5761 rxi_SendXmitList(call, call->xmitList,
5762 nXmitPackets, istack);
5765 dpf(("call %d xmit packet %"AFS_PTR_FMT" now %u.%06u retryTime %u.%06u\n",
5766 *(call->callNumber), p,
5768 p->retryTime.sec, p->retryTime.usec));
5769 call->xmitList[nXmitPackets++] = p;
5773 /* xmitList now hold pointers to all of the packets that are
5774 * ready to send. Now we loop to send the packets */
5775 if (nXmitPackets > 0) {
5776 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
5780 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5782 * TQ references no longer protected by this flag; they must remain
5783 * protected by the global lock.
5785 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5786 call->flags &= ~RX_CALL_TQ_BUSY;
5787 rxi_WakeUpTransmitQueue(call);
5791 /* We went into the error state while sending packets. Now is
5792 * the time to reset the call. This will also inform the using
5793 * process that the call is in an error state.
5795 if (rx_stats_active)
5796 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
5797 call->flags &= ~RX_CALL_TQ_BUSY;
5798 rxi_WakeUpTransmitQueue(call);
5799 rxi_CallError(call, call->error);
5802 #ifdef RX_ENABLE_LOCKS
5803 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5805 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5806 /* Some packets have received acks. If they all have, we can clear
5807 * the transmit queue.
5810 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5811 if (p->header.seq < call->tfirst
5812 && (p->flags & RX_PKTFLAG_ACKED)) {
5814 #ifdef RX_TRACK_PACKETS
5815 p->flags &= ~RX_PKTFLAG_TQ;
5817 #ifdef RXDEBUG_PACKET
5825 call->flags |= RX_CALL_TQ_CLEARME;
5827 #endif /* RX_ENABLE_LOCKS */
5828 /* Don't bother doing retransmits if the TQ is cleared. */
5829 if (call->flags & RX_CALL_TQ_CLEARME) {
5830 rxi_ClearTransmitQueue(call, 1);
5832 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5835 /* Always post a resend event, if there is anything in the
5836 * queue, and resend is possible. There should be at least
5837 * one unacknowledged packet in the queue ... otherwise none
5838 * of these packets should be on the queue in the first place.
5840 if (call->resendEvent) {
5841 /* Cancel the existing event and post a new one */
5842 rxevent_Cancel(call->resendEvent, call,
5843 RX_CALL_REFCOUNT_RESEND);
5846 /* The retry time is the retry time on the first unacknowledged
5847 * packet inside the current window */
5849 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5850 /* Don't set timers for packets outside the window */
5851 if (p->header.seq >= call->tfirst + call->twind) {
5855 if (!(p->flags & RX_PKTFLAG_ACKED)
5856 && !clock_IsZero(&p->retryTime)) {
5858 retryTime = p->retryTime;
5863 /* Post a new event to re-run rxi_Start when retries may be needed */
5864 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5865 #ifdef RX_ENABLE_LOCKS
5866 MUTEX_ENTER(&rx_refcnt_mutex);
5867 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5868 MUTEX_EXIT(&rx_refcnt_mutex);
5870 rxevent_PostNow2(&retryTime, &usenow,
5872 (void *)call, 0, istack);
5873 #else /* RX_ENABLE_LOCKS */
5875 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5876 (void *)call, 0, istack);
5877 #endif /* RX_ENABLE_LOCKS */
5880 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5881 } while (call->flags & RX_CALL_NEED_START);
5883 * TQ references no longer protected by this flag; they must remain
5884 * protected by the global lock.
5886 call->flags &= ~RX_CALL_TQ_BUSY;
5887 rxi_WakeUpTransmitQueue(call);
5889 call->flags |= RX_CALL_NEED_START;
5891 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5893 if (call->resendEvent) {
5894 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5899 /* Also adjusts the keep alive parameters for the call, to reflect
5900 * that we have just sent a packet (so keep alives aren't sent
5903 rxi_Send(struct rx_call *call, struct rx_packet *p,
5906 struct rx_connection *conn = call->conn;
5908 /* Stamp each packet with the user supplied status */
5909 p->header.userStatus = call->localStatus;
5911 /* Allow the security object controlling this call's security to
5912 * make any last-minute changes to the packet */
5913 RXS_SendPacket(conn->securityObject, call, p);
5915 /* Since we're about to send SOME sort of packet to the peer, it's
5916 * safe to nuke any scheduled end-of-packets ack */
5917 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5919 /* Actually send the packet, filling in more connection-specific fields */
5920 MUTEX_EXIT(&call->lock);
5921 MUTEX_ENTER(&rx_refcnt_mutex);
5922 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5923 MUTEX_EXIT(&rx_refcnt_mutex);
5924 rxi_SendPacket(call, conn, p, istack);
5925 MUTEX_ENTER(&rx_refcnt_mutex);
5926 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5927 MUTEX_EXIT(&rx_refcnt_mutex);
5928 MUTEX_ENTER(&call->lock);
5930 /* Update last send time for this call (for keep-alive
5931 * processing), and for the connection (so that we can discover
5932 * idle connections) */
5933 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5934 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
5935 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
5937 conn->lastSendTime = call->lastSendTime = clock_Sec();
5938 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
5939 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5940 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
5941 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
5942 RX_ACK_PING_RESPONSE)))
5943 call->lastSendData = call->lastSendTime;
5947 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5948 * that things are fine. Also called periodically to guarantee that nothing
5949 * falls through the cracks (e.g. (error + dally) connections have keepalive
5950 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5952 * haveCTLock Set if calling from rxi_ReapConnections
5954 #ifdef RX_ENABLE_LOCKS
5956 rxi_CheckCall(struct rx_call *call, int haveCTLock)
5957 #else /* RX_ENABLE_LOCKS */
5959 rxi_CheckCall(struct rx_call *call)
5960 #endif /* RX_ENABLE_LOCKS */
5962 struct rx_connection *conn = call->conn;
5964 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
5965 afs_uint32 fudgeFactor;
5969 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5970 if (call->flags & RX_CALL_TQ_BUSY) {
5971 /* Call is active and will be reset by rxi_Start if it's
5972 * in an error state.
5977 /* RTT + 8*MDEV, rounded up to the next second. */
5978 fudgeFactor = (((afs_uint32) conn->peer->rtt >> 3) +
5979 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5981 deadTime = conn->secondsUntilDead + fudgeFactor;
5983 /* These are computed to the second (+- 1 second). But that's
5984 * good enough for these values, which should be a significant
5985 * number of seconds. */
5986 if (now > (call->lastReceiveTime + deadTime)) {
5987 if (call->state == RX_STATE_ACTIVE) {
5989 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5991 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5992 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5993 ip_stack_t *ipst = ns->netstack_ip;
5995 ire = ire_cache_lookup(conn->peer->host
5996 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5998 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
6000 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6007 if (ire && ire->ire_max_frag > 0)
6008 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6010 #if defined(GLOBAL_NETSTACKID)
6014 #endif /* ADAPT_PMTU */
6015 cerror = RX_CALL_DEAD;
6018 #ifdef RX_ENABLE_LOCKS
6019 /* Cancel pending events */
6020 rxevent_Cancel(call->delayedAckEvent, call,
6021 RX_CALL_REFCOUNT_DELAY);
6022 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
6023 rxevent_Cancel(call->keepAliveEvent, call,
6024 RX_CALL_REFCOUNT_ALIVE);
6025 MUTEX_ENTER(&rx_refcnt_mutex);
6026 if (call->refCount == 0) {
6027 rxi_FreeCall(call, haveCTLock);
6028 MUTEX_EXIT(&rx_refcnt_mutex);
6031 MUTEX_EXIT(&rx_refcnt_mutex);
6033 #else /* RX_ENABLE_LOCKS */
6034 rxi_FreeCall(call, 0);
6036 #endif /* RX_ENABLE_LOCKS */
6038 /* Non-active calls are destroyed if they are not responding
6039 * to pings; active calls are simply flagged in error, so the
6040 * attached process can die reasonably gracefully. */
6043 if (conn->idleDeadTime) {
6044 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6047 /* see if we have a non-activity timeout */
6048 if (call->startWait && idleDeadTime
6049 && ((call->startWait + idleDeadTime) < now) &&
6050 (call->flags & RX_CALL_READER_WAIT)) {
6051 if (call->state == RX_STATE_ACTIVE) {
6052 cerror = RX_CALL_TIMEOUT;
6056 if (call->lastSendData && idleDeadTime && (conn->idleDeadErr != 0)
6057 && ((call->lastSendData + idleDeadTime) < now)) {
6058 if (call->state == RX_STATE_ACTIVE) {
6059 cerror = conn->idleDeadErr;
6065 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6068 /* see if we have a hard timeout */
6070 && (now > (hardDeadTime + call->startTime.sec))) {
6071 if (call->state == RX_STATE_ACTIVE)
6072 rxi_CallError(call, RX_CALL_TIMEOUT);
6077 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT
6078 && call->lastReceiveTime) {
6079 int oldMTU = conn->peer->ifMTU;
6081 /* if we thought we could send more, perhaps things got worse */
6082 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6083 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6084 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6085 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6087 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6089 /* minimum capped in SetPeerMtu */
6090 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6093 conn->lastPacketSize = 0;
6095 /* needed so ResetCall doesn't clobber us. */
6096 call->MTU = conn->peer->ifMTU;
6098 /* if we never succeeded, let the error pass out as-is */
6099 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6100 cerror = conn->msgsizeRetryErr;
6103 rxi_CallError(call, cerror);
6108 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6110 struct rx_connection *conn = arg1;
6111 struct rx_header theader;
6113 struct sockaddr_in taddr;
6116 struct iovec tmpiov[2];
6119 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6122 tp = &tbuffer[sizeof(struct rx_header)];
6123 taddr.sin_family = AF_INET;
6124 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6125 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6126 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6127 taddr.sin_len = sizeof(struct sockaddr_in);
6129 memset(&theader, 0, sizeof(theader));
6130 theader.epoch = htonl(999);
6132 theader.callNumber = 0;
6135 theader.type = RX_PACKET_TYPE_VERSION;
6136 theader.flags = RX_LAST_PACKET;
6137 theader.serviceId = 0;
6139 memcpy(tbuffer, &theader, sizeof(theader));
6140 memcpy(tp, &a, sizeof(a));
6141 tmpiov[0].iov_base = tbuffer;
6142 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6144 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6146 MUTEX_ENTER(&conn->conn_data_lock);
6147 MUTEX_ENTER(&rx_refcnt_mutex);
6148 /* Only reschedule ourselves if the connection would not be destroyed */
6149 if (conn->refCount <= 1) {
6150 conn->natKeepAliveEvent = NULL;
6151 MUTEX_EXIT(&rx_refcnt_mutex);
6152 MUTEX_EXIT(&conn->conn_data_lock);
6153 rx_DestroyConnection(conn); /* drop the reference for this */
6155 conn->refCount--; /* drop the reference for this */
6156 MUTEX_EXIT(&rx_refcnt_mutex);
6157 conn->natKeepAliveEvent = NULL;
6158 rxi_ScheduleNatKeepAliveEvent(conn);
6159 MUTEX_EXIT(&conn->conn_data_lock);
6164 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6166 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6167 struct clock when, now;
6168 clock_GetTime(&now);
6170 when.sec += conn->secondsUntilNatPing;
6171 MUTEX_ENTER(&rx_refcnt_mutex);
6172 conn->refCount++; /* hold a reference for this */
6173 MUTEX_EXIT(&rx_refcnt_mutex);
6174 conn->natKeepAliveEvent =
6175 rxevent_PostNow(&when, &now, rxi_NatKeepAliveEvent, conn, 0);
6180 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6182 MUTEX_ENTER(&conn->conn_data_lock);
6183 conn->secondsUntilNatPing = seconds;
6185 rxi_ScheduleNatKeepAliveEvent(conn);
6186 MUTEX_EXIT(&conn->conn_data_lock);
6190 rxi_NatKeepAliveOn(struct rx_connection *conn)
6192 MUTEX_ENTER(&conn->conn_data_lock);
6193 rxi_ScheduleNatKeepAliveEvent(conn);
6194 MUTEX_EXIT(&conn->conn_data_lock);
6197 /* When a call is in progress, this routine is called occasionally to
6198 * make sure that some traffic has arrived (or been sent to) the peer.
6199 * If nothing has arrived in a reasonable amount of time, the call is
6200 * declared dead; if nothing has been sent for a while, we send a
6201 * keep-alive packet (if we're actually trying to keep the call alive)
6204 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6206 struct rx_call *call = arg1;
6207 struct rx_connection *conn;
6210 MUTEX_ENTER(&rx_refcnt_mutex);
6211 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6212 MUTEX_EXIT(&rx_refcnt_mutex);
6213 MUTEX_ENTER(&call->lock);
6214 if (event == call->keepAliveEvent)
6215 call->keepAliveEvent = NULL;
6218 #ifdef RX_ENABLE_LOCKS
6219 if (rxi_CheckCall(call, 0)) {
6220 MUTEX_EXIT(&call->lock);
6223 #else /* RX_ENABLE_LOCKS */
6224 if (rxi_CheckCall(call))
6226 #endif /* RX_ENABLE_LOCKS */
6228 /* Don't try to keep alive dallying calls */
6229 if (call->state == RX_STATE_DALLY) {
6230 MUTEX_EXIT(&call->lock);
6235 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6236 /* Don't try to send keepalives if there is unacknowledged data */
6237 /* the rexmit code should be good enough, this little hack
6238 * doesn't quite work XXX */
6239 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6241 rxi_ScheduleKeepAliveEvent(call);
6242 MUTEX_EXIT(&call->lock);
6245 /* Does what's on the nameplate. */
6247 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy)
6249 struct rx_call *call = arg1;
6250 struct rx_connection *conn;
6252 MUTEX_ENTER(&rx_refcnt_mutex);
6253 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6254 MUTEX_EXIT(&rx_refcnt_mutex);
6255 MUTEX_ENTER(&call->lock);
6257 if (event == call->growMTUEvent)
6258 call->growMTUEvent = NULL;
6260 #ifdef RX_ENABLE_LOCKS
6261 if (rxi_CheckCall(call, 0)) {
6262 MUTEX_EXIT(&call->lock);
6265 #else /* RX_ENABLE_LOCKS */
6266 if (rxi_CheckCall(call))
6268 #endif /* RX_ENABLE_LOCKS */
6270 /* Don't bother with dallying calls */
6271 if (call->state == RX_STATE_DALLY) {
6272 MUTEX_EXIT(&call->lock);
6279 * keep being scheduled, just don't do anything if we're at peak,
6280 * or we're not set up to be properly handled (idle timeout required)
6282 if ((conn->peer->maxPacketSize != 0) &&
6283 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6284 (conn->idleDeadErr))
6285 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6286 rxi_ScheduleGrowMTUEvent(call, 0);
6287 MUTEX_EXIT(&call->lock);
6291 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6293 if (!call->keepAliveEvent) {
6294 struct clock when, now;
6295 clock_GetTime(&now);
6297 when.sec += call->conn->secondsUntilPing;
6298 MUTEX_ENTER(&rx_refcnt_mutex);
6299 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6300 MUTEX_EXIT(&rx_refcnt_mutex);
6301 call->keepAliveEvent =
6302 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
6307 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6309 if (!call->growMTUEvent) {
6310 struct clock when, now;
6312 clock_GetTime(&now);
6315 if (call->conn->secondsUntilPing)
6316 secs = (6*call->conn->secondsUntilPing)-1;
6318 if (call->conn->secondsUntilDead)
6319 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6323 MUTEX_ENTER(&rx_refcnt_mutex);
6324 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6325 MUTEX_EXIT(&rx_refcnt_mutex);
6326 call->growMTUEvent =
6327 rxevent_PostNow(&when, &now, rxi_GrowMTUEvent, call, 0);
6331 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6333 rxi_KeepAliveOn(struct rx_call *call)
6335 /* Pretend last packet received was received now--i.e. if another
6336 * packet isn't received within the keep alive time, then the call
6337 * will die; Initialize last send time to the current time--even
6338 * if a packet hasn't been sent yet. This will guarantee that a
6339 * keep-alive is sent within the ping time */
6340 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6341 rxi_ScheduleKeepAliveEvent(call);
6345 rxi_GrowMTUOn(struct rx_call *call)
6347 struct rx_connection *conn = call->conn;
6348 MUTEX_ENTER(&conn->conn_data_lock);
6349 conn->lastPingSizeSer = conn->lastPingSize = 0;
6350 MUTEX_EXIT(&conn->conn_data_lock);
6351 rxi_ScheduleGrowMTUEvent(call, 1);
6354 /* This routine is called to send connection abort messages
6355 * that have been delayed to throttle looping clients. */
6357 rxi_SendDelayedConnAbort(struct rxevent *event,
6358 void *arg1, void *unused)
6360 struct rx_connection *conn = arg1;
6363 struct rx_packet *packet;
6365 MUTEX_ENTER(&conn->conn_data_lock);
6366 conn->delayedAbortEvent = NULL;
6367 error = htonl(conn->error);
6369 MUTEX_EXIT(&conn->conn_data_lock);
6370 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6373 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6374 RX_PACKET_TYPE_ABORT, (char *)&error,
6376 rxi_FreePacket(packet);
6380 /* This routine is called to send call abort messages
6381 * that have been delayed to throttle looping clients. */
6383 rxi_SendDelayedCallAbort(struct rxevent *event,
6384 void *arg1, void *dummy)
6386 struct rx_call *call = arg1;
6389 struct rx_packet *packet;
6391 MUTEX_ENTER(&call->lock);
6392 call->delayedAbortEvent = NULL;
6393 error = htonl(call->error);
6395 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6398 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6399 (char *)&error, sizeof(error), 0);
6400 rxi_FreePacket(packet);
6402 MUTEX_EXIT(&call->lock);
6403 MUTEX_ENTER(&rx_refcnt_mutex);
6404 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6405 MUTEX_EXIT(&rx_refcnt_mutex);
6408 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6409 * seconds) to ask the client to authenticate itself. The routine
6410 * issues a challenge to the client, which is obtained from the
6411 * security object associated with the connection */
6413 rxi_ChallengeEvent(struct rxevent *event,
6414 void *arg0, void *arg1, int tries)
6416 struct rx_connection *conn = arg0;
6418 conn->challengeEvent = NULL;
6419 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6420 struct rx_packet *packet;
6421 struct clock when, now;
6424 /* We've failed to authenticate for too long.
6425 * Reset any calls waiting for authentication;
6426 * they are all in RX_STATE_PRECALL.
6430 MUTEX_ENTER(&conn->conn_call_lock);
6431 for (i = 0; i < RX_MAXCALLS; i++) {
6432 struct rx_call *call = conn->call[i];
6434 MUTEX_ENTER(&call->lock);
6435 if (call->state == RX_STATE_PRECALL) {
6436 rxi_CallError(call, RX_CALL_DEAD);
6437 rxi_SendCallAbort(call, NULL, 0, 0);
6439 MUTEX_EXIT(&call->lock);
6442 MUTEX_EXIT(&conn->conn_call_lock);
6446 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6448 /* If there's no packet available, do this later. */
6449 RXS_GetChallenge(conn->securityObject, conn, packet);
6450 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6451 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6452 rxi_FreePacket(packet);
6454 clock_GetTime(&now);
6456 when.sec += RX_CHALLENGE_TIMEOUT;
6457 conn->challengeEvent =
6458 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
6463 /* Call this routine to start requesting the client to authenticate
6464 * itself. This will continue until authentication is established,
6465 * the call times out, or an invalid response is returned. The
6466 * security object associated with the connection is asked to create
6467 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6468 * defined earlier. */
6470 rxi_ChallengeOn(struct rx_connection *conn)
6472 if (!conn->challengeEvent) {
6473 RXS_CreateChallenge(conn->securityObject, conn);
6474 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6479 /* rxi_ComputeRoundTripTime is called with peer locked. */
6480 /* peer may be null */
6482 rxi_ComputeRoundTripTime(struct rx_packet *p,
6483 struct rx_ackPacket *ack,
6484 struct rx_peer *peer,
6487 struct clock thisRtt, *sentp;
6491 /* If the ACK is delayed, then do nothing */
6492 if (ack->reason == RX_ACK_DELAY)
6495 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6496 * their RTT multiple times, so only include the RTT of the last packet
6498 if (p->flags & RX_JUMBO_PACKET)
6501 /* Use the serial number to determine which transmission the ACK is for,
6502 * and set the sent time to match this. If we have no serial number, then
6503 * only use the ACK for RTT calculations if the packet has not been
6507 serial = ntohl(ack->serial);
6509 if (serial == p->header.serial) {
6510 sentp = &p->timeSent;
6511 } else if (serial == p->firstSerial) {
6512 sentp = &p->firstSent;
6513 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6514 sentp = &p->firstSent;
6518 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6519 sentp = &p->firstSent;
6526 if (clock_Lt(&thisRtt, sentp))
6527 return; /* somebody set the clock back, don't count this time. */
6529 clock_Sub(&thisRtt, sentp);
6530 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6531 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6533 if (clock_IsZero(&thisRtt)) {
6535 * The actual round trip time is shorter than the
6536 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6537 * Since we can't tell which at the moment we will assume 1ms.
6539 thisRtt.usec = 1000;
6542 if (rx_stats_active) {
6543 MUTEX_ENTER(&rx_stats_mutex);
6544 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6545 rx_stats.minRtt = thisRtt;
6546 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6547 if (thisRtt.sec > 60) {
6548 MUTEX_EXIT(&rx_stats_mutex);
6549 return; /* somebody set the clock ahead */
6551 rx_stats.maxRtt = thisRtt;
6553 clock_Add(&rx_stats.totalRtt, &thisRtt);
6554 rx_atomic_inc(&rx_stats.nRttSamples);
6555 MUTEX_EXIT(&rx_stats_mutex);
6558 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6560 /* Apply VanJacobson round-trip estimations */
6565 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
6566 * srtt is stored as fixed point with 3 bits after the binary
6567 * point (i.e., scaled by 8). The following magic is
6568 * equivalent to the smoothing algorithm in rfc793 with an
6569 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6570 * srtt'*8 = rtt + srtt*7
6571 * srtt'*8 = srtt*8 + rtt - srtt
6572 * srtt' = srtt + rtt/8 - srtt/8
6573 * srtt' = srtt + (rtt - srtt)/8
6576 delta = _8THMSEC(&thisRtt) - peer->rtt;
6577 peer->rtt += (delta >> 3);
6580 * We accumulate a smoothed rtt variance (actually, a smoothed
6581 * mean difference), then set the retransmit timer to smoothed
6582 * rtt + 4 times the smoothed variance (was 2x in van's original
6583 * paper, but 4x works better for me, and apparently for him as
6585 * rttvar is stored as
6586 * fixed point with 2 bits after the binary point (scaled by
6587 * 4). The following is equivalent to rfc793 smoothing with
6588 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6589 * rttvar'*4 = rttvar*3 + |delta|
6590 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6591 * rttvar' = rttvar + |delta|/4 - rttvar/4
6592 * rttvar' = rttvar + (|delta| - rttvar)/4
6593 * This replaces rfc793's wired-in beta.
6594 * dev*4 = dev*4 + (|actual - expected| - dev)
6600 delta -= (peer->rtt_dev << 1);
6601 peer->rtt_dev += (delta >> 3);
6603 /* I don't have a stored RTT so I start with this value. Since I'm
6604 * probably just starting a call, and will be pushing more data down
6605 * this, I expect congestion to increase rapidly. So I fudge a
6606 * little, and I set deviance to half the rtt. In practice,
6607 * deviance tends to approach something a little less than
6608 * half the smoothed rtt. */
6609 peer->rtt = _8THMSEC(&thisRtt) + 8;
6610 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
6612 /* the timeout is RTT + 4*MDEV + rx_minPeerTimeout msec.
6613 * This is because one end or the other of these connections is usually
6614 * in a user process, and can be switched and/or swapped out. So on fast,
6615 * reliable networks, the timeout would otherwise be too short. */
6616 rtt_timeout = ((peer->rtt >> 3) + peer->rtt_dev) + rx_minPeerTimeout;
6617 clock_Zero(&(peer->timeout));
6618 clock_Addmsec(&(peer->timeout), rtt_timeout);
6620 /* Reset the backedOff flag since we just computed a new timeout value */
6621 peer->backedOff = 0;
6623 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6624 p->header.callNumber, p, MSEC(&thisRtt), peer->rtt >> 3, peer->rtt_dev >> 2, (peer->timeout.sec), (peer->timeout.usec)));
6628 /* Find all server connections that have not been active for a long time, and
6631 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6633 struct clock now, when;
6634 clock_GetTime(&now);
6636 /* Find server connection structures that haven't been used for
6637 * greater than rx_idleConnectionTime */
6639 struct rx_connection **conn_ptr, **conn_end;
6640 int i, havecalls = 0;
6641 MUTEX_ENTER(&rx_connHashTable_lock);
6642 for (conn_ptr = &rx_connHashTable[0], conn_end =
6643 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6645 struct rx_connection *conn, *next;
6646 struct rx_call *call;
6650 for (conn = *conn_ptr; conn; conn = next) {
6651 /* XXX -- Shouldn't the connection be locked? */
6654 for (i = 0; i < RX_MAXCALLS; i++) {
6655 call = conn->call[i];
6659 code = MUTEX_TRYENTER(&call->lock);
6662 #ifdef RX_ENABLE_LOCKS
6663 result = rxi_CheckCall(call, 1);
6664 #else /* RX_ENABLE_LOCKS */
6665 result = rxi_CheckCall(call);
6666 #endif /* RX_ENABLE_LOCKS */
6667 MUTEX_EXIT(&call->lock);
6669 /* If CheckCall freed the call, it might
6670 * have destroyed the connection as well,
6671 * which screws up the linked lists.
6677 if (conn->type == RX_SERVER_CONNECTION) {
6678 /* This only actually destroys the connection if
6679 * there are no outstanding calls */
6680 MUTEX_ENTER(&conn->conn_data_lock);
6681 MUTEX_ENTER(&rx_refcnt_mutex);
6682 if (!havecalls && !conn->refCount
6683 && ((conn->lastSendTime + rx_idleConnectionTime) <
6685 conn->refCount++; /* it will be decr in rx_DestroyConn */
6686 MUTEX_EXIT(&rx_refcnt_mutex);
6687 MUTEX_EXIT(&conn->conn_data_lock);
6688 #ifdef RX_ENABLE_LOCKS
6689 rxi_DestroyConnectionNoLock(conn);
6690 #else /* RX_ENABLE_LOCKS */
6691 rxi_DestroyConnection(conn);
6692 #endif /* RX_ENABLE_LOCKS */
6694 #ifdef RX_ENABLE_LOCKS
6696 MUTEX_EXIT(&rx_refcnt_mutex);
6697 MUTEX_EXIT(&conn->conn_data_lock);
6699 #endif /* RX_ENABLE_LOCKS */
6703 #ifdef RX_ENABLE_LOCKS
6704 while (rx_connCleanup_list) {
6705 struct rx_connection *conn;
6706 conn = rx_connCleanup_list;
6707 rx_connCleanup_list = rx_connCleanup_list->next;
6708 MUTEX_EXIT(&rx_connHashTable_lock);
6709 rxi_CleanupConnection(conn);
6710 MUTEX_ENTER(&rx_connHashTable_lock);
6712 MUTEX_EXIT(&rx_connHashTable_lock);
6713 #endif /* RX_ENABLE_LOCKS */
6716 /* Find any peer structures that haven't been used (haven't had an
6717 * associated connection) for greater than rx_idlePeerTime */
6719 struct rx_peer **peer_ptr, **peer_end;
6723 * Why do we need to hold the rx_peerHashTable_lock across
6724 * the incrementing of peer_ptr since the rx_peerHashTable
6725 * array is not changing? We don't.
6727 * By dropping the lock periodically we can permit other
6728 * activities to be performed while a rxi_ReapConnections
6729 * call is in progress. The goal of reap connections
6730 * is to clean up quickly without causing large amounts
6731 * of contention. Therefore, it is important that global
6732 * mutexes not be held for extended periods of time.
6734 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6735 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6737 struct rx_peer *peer, *next, *prev;
6739 MUTEX_ENTER(&rx_peerHashTable_lock);
6740 for (prev = peer = *peer_ptr; peer; peer = next) {
6742 code = MUTEX_TRYENTER(&peer->peer_lock);
6743 if ((code) && (peer->refCount == 0)
6744 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6745 rx_interface_stat_p rpc_stat, nrpc_stat;
6749 * now know that this peer object is one to be
6750 * removed from the hash table. Once it is removed
6751 * it can't be referenced by other threads.
6752 * Lets remove it first and decrement the struct
6753 * nPeerStructs count.
6755 if (peer == *peer_ptr) {
6761 if (rx_stats_active)
6762 rx_atomic_dec(&rx_stats.nPeerStructs);
6765 * Now if we hold references on 'prev' and 'next'
6766 * we can safely drop the rx_peerHashTable_lock
6767 * while we destroy this 'peer' object.
6773 MUTEX_EXIT(&rx_peerHashTable_lock);
6775 MUTEX_EXIT(&peer->peer_lock);
6776 MUTEX_DESTROY(&peer->peer_lock);
6778 (&peer->rpcStats, rpc_stat, nrpc_stat,
6779 rx_interface_stat)) {
6780 unsigned int num_funcs;
6783 queue_Remove(&rpc_stat->queue_header);
6784 queue_Remove(&rpc_stat->all_peers);
6785 num_funcs = rpc_stat->stats[0].func_total;
6787 sizeof(rx_interface_stat_t) +
6788 rpc_stat->stats[0].func_total *
6789 sizeof(rx_function_entry_v1_t);
6791 rxi_Free(rpc_stat, space);
6793 MUTEX_ENTER(&rx_rpc_stats);
6794 rxi_rpc_peer_stat_cnt -= num_funcs;
6795 MUTEX_EXIT(&rx_rpc_stats);
6800 * Regain the rx_peerHashTable_lock and
6801 * decrement the reference count on 'prev'
6804 MUTEX_ENTER(&rx_peerHashTable_lock);
6811 MUTEX_EXIT(&peer->peer_lock);
6816 MUTEX_EXIT(&rx_peerHashTable_lock);
6820 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6821 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6822 * GC, just below. Really, we shouldn't have to keep moving packets from
6823 * one place to another, but instead ought to always know if we can
6824 * afford to hold onto a packet in its particular use. */
6825 MUTEX_ENTER(&rx_freePktQ_lock);
6826 if (rx_waitingForPackets) {
6827 rx_waitingForPackets = 0;
6828 #ifdef RX_ENABLE_LOCKS
6829 CV_BROADCAST(&rx_waitingForPackets_cv);
6831 osi_rxWakeup(&rx_waitingForPackets);
6834 MUTEX_EXIT(&rx_freePktQ_lock);
6837 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6838 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6842 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6843 * rx.h is sort of strange this is better. This is called with a security
6844 * object before it is discarded. Each connection using a security object has
6845 * its own refcount to the object so it won't actually be freed until the last
6846 * connection is destroyed.
6848 * This is the only rxs module call. A hold could also be written but no one
6852 rxs_Release(struct rx_securityClass *aobj)
6854 return RXS_Close(aobj);
6858 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6859 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6860 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6861 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6863 /* Adjust our estimate of the transmission rate to this peer, given
6864 * that the packet p was just acked. We can adjust peer->timeout and
6865 * call->twind. Pragmatically, this is called
6866 * only with packets of maximal length.
6867 * Called with peer and call locked.
6871 rxi_ComputeRate(struct rx_peer *peer, struct rx_call *call,
6872 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6874 afs_int32 xferSize, xferMs;
6878 /* Count down packets */
6879 if (peer->rateFlag > 0)
6881 /* Do nothing until we're enabled */
6882 if (peer->rateFlag != 0)
6887 /* Count only when the ack seems legitimate */
6888 switch (ackReason) {
6889 case RX_ACK_REQUESTED:
6891 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6895 case RX_ACK_PING_RESPONSE:
6896 if (p) /* want the response to ping-request, not data send */
6898 clock_GetTime(&newTO);
6899 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6900 clock_Sub(&newTO, &call->pingRequestTime);
6901 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6905 xferSize = rx_AckDataSize(rx_maxSendWindow) + RX_HEADER_SIZE;
6912 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %d.%06d, rtt %u, ps %u)\n",
6913 ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"),
6914 xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
6916 /* Track only packets that are big enough. */
6917 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6921 /* absorb RTT data (in milliseconds) for these big packets */
6922 if (peer->smRtt == 0) {
6923 peer->smRtt = xferMs;
6925 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6930 if (peer->countDown) {
6934 peer->countDown = 10; /* recalculate only every so often */
6936 /* In practice, we can measure only the RTT for full packets,
6937 * because of the way Rx acks the data that it receives. (If it's
6938 * smaller than a full packet, it often gets implicitly acked
6939 * either by the call response (from a server) or by the next call
6940 * (from a client), and either case confuses transmission times
6941 * with processing times.) Therefore, replace the above
6942 * more-sophisticated processing with a simpler version, where the
6943 * smoothed RTT is kept for full-size packets, and the time to
6944 * transmit a windowful of full-size packets is simply RTT *
6945 * windowSize. Again, we take two steps:
6946 - ensure the timeout is large enough for a single packet's RTT;
6947 - ensure that the window is small enough to fit in the desired timeout.*/
6949 /* First, the timeout check. */
6950 minTime = peer->smRtt;
6951 /* Get a reasonable estimate for a timeout period */
6953 newTO.sec = minTime / 1000;
6954 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6956 /* Increase the timeout period so that we can always do at least
6957 * one packet exchange */
6958 if (clock_Gt(&newTO, &peer->timeout)) {
6960 dpf(("CONG peer %lx/%u: timeout %d.%06d ==> %ld.%06d (rtt %u)\n",
6961 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec,
6962 newTO.sec, newTO.usec, peer->smRtt));
6964 peer->timeout = newTO;
6967 /* Now, get an estimate for the transmit window size. */
6968 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6969 /* Now, convert to the number of full packets that could fit in a
6970 * reasonable fraction of that interval */
6971 minTime /= (peer->smRtt << 1);
6972 minTime = MAX(minTime, rx_minPeerTimeout);
6973 xferSize = minTime; /* (make a copy) */
6975 /* Now clamp the size to reasonable bounds. */
6978 else if (minTime > rx_maxSendWindow)
6979 minTime = rx_maxSendWindow;
6980 /* if (minTime != peer->maxWindow) {
6981 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u)\n",
6982 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6983 peer->timeout.sec, peer->timeout.usec, peer->smRtt));
6984 peer->maxWindow = minTime;
6985 elide... call->twind = minTime;
6989 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6990 * Discern this by calculating the timeout necessary for rx_Window
6992 if ((xferSize > rx_maxSendWindow) && (peer->timeout.sec >= 3)) {
6993 /* calculate estimate for transmission interval in milliseconds */
6994 minTime = rx_maxSendWindow * peer->smRtt;
6995 if (minTime < 1000) {
6996 dpf(("CONG peer %lx/%u: cut TO %d.%06d by 0.5 (rtt %u)\n",
6997 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6998 peer->timeout.usec, peer->smRtt));
7000 newTO.sec = 0; /* cut back on timeout by half a second */
7001 newTO.usec = 500000;
7002 clock_Sub(&peer->timeout, &newTO);
7007 } /* end of rxi_ComputeRate */
7008 #endif /* ADAPT_WINDOW */
7016 #define TRACE_OPTION_RX_DEBUG 16
7024 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7025 0, KEY_QUERY_VALUE, &parmKey);
7026 if (code != ERROR_SUCCESS)
7029 dummyLen = sizeof(TraceOption);
7030 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7031 (BYTE *) &TraceOption, &dummyLen);
7032 if (code == ERROR_SUCCESS) {
7033 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7035 RegCloseKey (parmKey);
7036 #endif /* AFS_NT40_ENV */
7041 rx_DebugOnOff(int on)
7045 rxdebug_active = on;
7051 rx_StatsOnOff(int on)
7053 rx_stats_active = on;
7057 /* Don't call this debugging routine directly; use dpf */
7059 rxi_DebugPrint(char *format, ...)
7068 va_start(ap, format);
7070 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7073 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7075 OutputDebugString(msg);
7081 va_start(ap, format);
7083 clock_GetTime(&now);
7084 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7085 (unsigned int)now.usec);
7086 vfprintf(rx_Log, format, ap);
7094 * This function is used to process the rx_stats structure that is local
7095 * to a process as well as an rx_stats structure received from a remote
7096 * process (via rxdebug). Therefore, it needs to do minimal version
7100 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7101 afs_int32 freePackets, char version)
7105 if (size != sizeof(struct rx_statistics)) {
7107 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7108 size, sizeof(struct rx_statistics));
7111 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7114 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7115 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7116 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7117 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7118 s->specialPktAllocFailures);
7120 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7121 s->receivePktAllocFailures, s->sendPktAllocFailures,
7122 s->specialPktAllocFailures);
7126 " greedy %u, " "bogusReads %u (last from host %x), "
7127 "noPackets %u, " "noBuffers %u, " "selects %u, "
7128 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7129 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7130 s->selects, s->sendSelects);
7132 fprintf(file, " packets read: ");
7133 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7134 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7136 fprintf(file, "\n");
7139 " other read counters: data %u, " "ack %u, " "dup %u "
7140 "spurious %u " "dally %u\n", s->dataPacketsRead,
7141 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7142 s->ignorePacketDally);
7144 fprintf(file, " packets sent: ");
7145 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7146 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7148 fprintf(file, "\n");
7151 " other send counters: ack %u, " "data %u (not resends), "
7152 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7153 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7154 s->dataPacketsPushed, s->ignoreAckedPacket);
7157 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7158 s->netSendFailures, (int)s->fatalErrors);
7160 if (s->nRttSamples) {
7161 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7162 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7164 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7165 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7169 " %d server connections, " "%d client connections, "
7170 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7171 s->nServerConns, s->nClientConns, s->nPeerStructs,
7172 s->nCallStructs, s->nFreeCallStructs);
7174 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7175 fprintf(file, " %d clock updates\n", clock_nUpdates);
7179 /* for backward compatibility */
7181 rx_PrintStats(FILE * file)
7183 MUTEX_ENTER(&rx_stats_mutex);
7184 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7185 sizeof(rx_stats), rx_nFreePackets,
7187 MUTEX_EXIT(&rx_stats_mutex);
7191 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7193 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
7194 ntohl(peer->host), (int)ntohs(peer->port), (int)peer->burstSize,
7195 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
7198 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
7199 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
7200 (int)peer->timeout.usec, peer->nSent, peer->reSends);
7203 " Packet size %d, " "max in packet skew %d, "
7204 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
7205 (int)peer->outPacketSkew);
7209 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7211 * This mutex protects the following static variables:
7215 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7216 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7218 #define LOCK_RX_DEBUG
7219 #define UNLOCK_RX_DEBUG
7220 #endif /* AFS_PTHREAD_ENV */
7222 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7224 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7225 u_char type, void *inputData, size_t inputLength,
7226 void *outputData, size_t outputLength)
7228 static afs_int32 counter = 100;
7229 time_t waitTime, waitCount;
7230 struct rx_header theader;
7233 struct timeval tv_now, tv_wake, tv_delta;
7234 struct sockaddr_in taddr, faddr;
7248 tp = &tbuffer[sizeof(struct rx_header)];
7249 taddr.sin_family = AF_INET;
7250 taddr.sin_port = remotePort;
7251 taddr.sin_addr.s_addr = remoteAddr;
7252 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7253 taddr.sin_len = sizeof(struct sockaddr_in);
7256 memset(&theader, 0, sizeof(theader));
7257 theader.epoch = htonl(999);
7259 theader.callNumber = htonl(counter);
7262 theader.type = type;
7263 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7264 theader.serviceId = 0;
7266 memcpy(tbuffer, &theader, sizeof(theader));
7267 memcpy(tp, inputData, inputLength);
7269 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7270 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7272 /* see if there's a packet available */
7273 gettimeofday(&tv_wake,0);
7274 tv_wake.tv_sec += waitTime;
7277 FD_SET(socket, &imask);
7278 tv_delta.tv_sec = tv_wake.tv_sec;
7279 tv_delta.tv_usec = tv_wake.tv_usec;
7280 gettimeofday(&tv_now, 0);
7282 if (tv_delta.tv_usec < tv_now.tv_usec) {
7284 tv_delta.tv_usec += 1000000;
7287 tv_delta.tv_usec -= tv_now.tv_usec;
7289 if (tv_delta.tv_sec < tv_now.tv_sec) {
7293 tv_delta.tv_sec -= tv_now.tv_sec;
7296 code = select(0, &imask, 0, 0, &tv_delta);
7297 #else /* AFS_NT40_ENV */
7298 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7299 #endif /* AFS_NT40_ENV */
7300 if (code == 1 && FD_ISSET(socket, &imask)) {
7301 /* now receive a packet */
7302 faddrLen = sizeof(struct sockaddr_in);
7304 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7305 (struct sockaddr *)&faddr, &faddrLen);
7308 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7309 if (counter == ntohl(theader.callNumber))
7317 /* see if we've timed out */
7325 code -= sizeof(struct rx_header);
7326 if (code > outputLength)
7327 code = outputLength;
7328 memcpy(outputData, tp, code);
7331 #endif /* RXDEBUG */
7334 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7335 afs_uint16 remotePort, struct rx_debugStats * stat,
7336 afs_uint32 * supportedValues)
7338 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7340 struct rx_debugIn in;
7342 *supportedValues = 0;
7343 in.type = htonl(RX_DEBUGI_GETSTATS);
7346 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7347 &in, sizeof(in), stat, sizeof(*stat));
7350 * If the call was successful, fixup the version and indicate
7351 * what contents of the stat structure are valid.
7352 * Also do net to host conversion of fields here.
7356 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7357 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7359 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7360 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7362 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7363 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7365 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7366 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7368 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7369 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7371 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7372 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7374 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7375 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7377 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7378 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7380 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7381 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7383 stat->nFreePackets = ntohl(stat->nFreePackets);
7384 stat->packetReclaims = ntohl(stat->packetReclaims);
7385 stat->callsExecuted = ntohl(stat->callsExecuted);
7386 stat->nWaiting = ntohl(stat->nWaiting);
7387 stat->idleThreads = ntohl(stat->idleThreads);
7388 stat->nWaited = ntohl(stat->nWaited);
7389 stat->nPackets = ntohl(stat->nPackets);
7398 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7399 afs_uint16 remotePort, struct rx_statistics * stat,
7400 afs_uint32 * supportedValues)
7402 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7404 struct rx_debugIn in;
7405 afs_int32 *lp = (afs_int32 *) stat;
7409 * supportedValues is currently unused, but added to allow future
7410 * versioning of this function.
7413 *supportedValues = 0;
7414 in.type = htonl(RX_DEBUGI_RXSTATS);
7416 memset(stat, 0, sizeof(*stat));
7418 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7419 &in, sizeof(in), stat, sizeof(*stat));
7424 * Do net to host conversion here
7427 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7438 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7439 afs_uint16 remotePort, size_t version_length,
7442 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7444 return MakeDebugCall(socket, remoteAddr, remotePort,
7445 RX_PACKET_TYPE_VERSION, a, 1, version,
7453 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7454 afs_uint16 remotePort, afs_int32 * nextConnection,
7455 int allConnections, afs_uint32 debugSupportedValues,
7456 struct rx_debugConn * conn,
7457 afs_uint32 * supportedValues)
7459 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7461 struct rx_debugIn in;
7465 * supportedValues is currently unused, but added to allow future
7466 * versioning of this function.
7469 *supportedValues = 0;
7470 if (allConnections) {
7471 in.type = htonl(RX_DEBUGI_GETALLCONN);
7473 in.type = htonl(RX_DEBUGI_GETCONN);
7475 in.index = htonl(*nextConnection);
7476 memset(conn, 0, sizeof(*conn));
7478 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7479 &in, sizeof(in), conn, sizeof(*conn));
7482 *nextConnection += 1;
7485 * Convert old connection format to new structure.
7488 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7489 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7490 #define MOVEvL(a) (conn->a = vL->a)
7492 /* any old or unrecognized version... */
7493 for (i = 0; i < RX_MAXCALLS; i++) {
7494 MOVEvL(callState[i]);
7495 MOVEvL(callMode[i]);
7496 MOVEvL(callFlags[i]);
7497 MOVEvL(callOther[i]);
7499 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7500 MOVEvL(secStats.type);
7501 MOVEvL(secStats.level);
7502 MOVEvL(secStats.flags);
7503 MOVEvL(secStats.expires);
7504 MOVEvL(secStats.packetsReceived);
7505 MOVEvL(secStats.packetsSent);
7506 MOVEvL(secStats.bytesReceived);
7507 MOVEvL(secStats.bytesSent);
7512 * Do net to host conversion here
7514 * I don't convert host or port since we are most likely
7515 * going to want these in NBO.
7517 conn->cid = ntohl(conn->cid);
7518 conn->serial = ntohl(conn->serial);
7519 for (i = 0; i < RX_MAXCALLS; i++) {
7520 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7522 conn->error = ntohl(conn->error);
7523 conn->secStats.flags = ntohl(conn->secStats.flags);
7524 conn->secStats.expires = ntohl(conn->secStats.expires);
7525 conn->secStats.packetsReceived =
7526 ntohl(conn->secStats.packetsReceived);
7527 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7528 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7529 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7530 conn->epoch = ntohl(conn->epoch);
7531 conn->natMTU = ntohl(conn->natMTU);
7540 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7541 afs_uint16 remotePort, afs_int32 * nextPeer,
7542 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7543 afs_uint32 * supportedValues)
7545 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7547 struct rx_debugIn in;
7550 * supportedValues is currently unused, but added to allow future
7551 * versioning of this function.
7554 *supportedValues = 0;
7555 in.type = htonl(RX_DEBUGI_GETPEER);
7556 in.index = htonl(*nextPeer);
7557 memset(peer, 0, sizeof(*peer));
7559 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7560 &in, sizeof(in), peer, sizeof(*peer));
7566 * Do net to host conversion here
7568 * I don't convert host or port since we are most likely
7569 * going to want these in NBO.
7571 peer->ifMTU = ntohs(peer->ifMTU);
7572 peer->idleWhen = ntohl(peer->idleWhen);
7573 peer->refCount = ntohs(peer->refCount);
7574 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7575 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7576 peer->rtt = ntohl(peer->rtt);
7577 peer->rtt_dev = ntohl(peer->rtt_dev);
7578 peer->timeout.sec = ntohl(peer->timeout.sec);
7579 peer->timeout.usec = ntohl(peer->timeout.usec);
7580 peer->nSent = ntohl(peer->nSent);
7581 peer->reSends = ntohl(peer->reSends);
7582 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7583 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7584 peer->rateFlag = ntohl(peer->rateFlag);
7585 peer->natMTU = ntohs(peer->natMTU);
7586 peer->maxMTU = ntohs(peer->maxMTU);
7587 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7588 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7589 peer->MTU = ntohs(peer->MTU);
7590 peer->cwind = ntohs(peer->cwind);
7591 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7592 peer->congestSeq = ntohs(peer->congestSeq);
7593 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7594 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7595 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7596 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7605 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7606 struct rx_debugPeer * peerStats)
7609 afs_int32 error = 1; /* default to "did not succeed" */
7610 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7612 MUTEX_ENTER(&rx_peerHashTable_lock);
7613 for(tp = rx_peerHashTable[hashValue];
7614 tp != NULL; tp = tp->next) {
7615 if (tp->host == peerHost)
7621 MUTEX_EXIT(&rx_peerHashTable_lock);
7625 MUTEX_ENTER(&tp->peer_lock);
7626 peerStats->host = tp->host;
7627 peerStats->port = tp->port;
7628 peerStats->ifMTU = tp->ifMTU;
7629 peerStats->idleWhen = tp->idleWhen;
7630 peerStats->refCount = tp->refCount;
7631 peerStats->burstSize = tp->burstSize;
7632 peerStats->burst = tp->burst;
7633 peerStats->burstWait.sec = tp->burstWait.sec;
7634 peerStats->burstWait.usec = tp->burstWait.usec;
7635 peerStats->rtt = tp->rtt;
7636 peerStats->rtt_dev = tp->rtt_dev;
7637 peerStats->timeout.sec = tp->timeout.sec;
7638 peerStats->timeout.usec = tp->timeout.usec;
7639 peerStats->nSent = tp->nSent;
7640 peerStats->reSends = tp->reSends;
7641 peerStats->inPacketSkew = tp->inPacketSkew;
7642 peerStats->outPacketSkew = tp->outPacketSkew;
7643 peerStats->rateFlag = tp->rateFlag;
7644 peerStats->natMTU = tp->natMTU;
7645 peerStats->maxMTU = tp->maxMTU;
7646 peerStats->maxDgramPackets = tp->maxDgramPackets;
7647 peerStats->ifDgramPackets = tp->ifDgramPackets;
7648 peerStats->MTU = tp->MTU;
7649 peerStats->cwind = tp->cwind;
7650 peerStats->nDgramPackets = tp->nDgramPackets;
7651 peerStats->congestSeq = tp->congestSeq;
7652 peerStats->bytesSent.high = tp->bytesSent.high;
7653 peerStats->bytesSent.low = tp->bytesSent.low;
7654 peerStats->bytesReceived.high = tp->bytesReceived.high;
7655 peerStats->bytesReceived.low = tp->bytesReceived.low;
7656 MUTEX_EXIT(&tp->peer_lock);
7658 MUTEX_ENTER(&rx_peerHashTable_lock);
7661 MUTEX_EXIT(&rx_peerHashTable_lock);
7669 struct rx_serverQueueEntry *np;
7672 struct rx_call *call;
7673 struct rx_serverQueueEntry *sq;
7677 if (rxinit_status == 1) {
7679 return; /* Already shutdown. */
7683 #ifndef AFS_PTHREAD_ENV
7684 FD_ZERO(&rx_selectMask);
7685 #endif /* AFS_PTHREAD_ENV */
7686 rxi_dataQuota = RX_MAX_QUOTA;
7687 #ifndef AFS_PTHREAD_ENV
7689 #endif /* AFS_PTHREAD_ENV */
7692 #ifndef AFS_PTHREAD_ENV
7693 #ifndef AFS_USE_GETTIMEOFDAY
7695 #endif /* AFS_USE_GETTIMEOFDAY */
7696 #endif /* AFS_PTHREAD_ENV */
7698 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7699 call = queue_First(&rx_freeCallQueue, rx_call);
7701 rxi_Free(call, sizeof(struct rx_call));
7704 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7705 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7711 struct rx_peer **peer_ptr, **peer_end;
7712 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7713 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7715 struct rx_peer *peer, *next;
7717 MUTEX_ENTER(&rx_peerHashTable_lock);
7718 for (peer = *peer_ptr; peer; peer = next) {
7719 rx_interface_stat_p rpc_stat, nrpc_stat;
7722 MUTEX_ENTER(&rx_rpc_stats);
7723 MUTEX_ENTER(&peer->peer_lock);
7725 (&peer->rpcStats, rpc_stat, nrpc_stat,
7726 rx_interface_stat)) {
7727 unsigned int num_funcs;
7730 queue_Remove(&rpc_stat->queue_header);
7731 queue_Remove(&rpc_stat->all_peers);
7732 num_funcs = rpc_stat->stats[0].func_total;
7734 sizeof(rx_interface_stat_t) +
7735 rpc_stat->stats[0].func_total *
7736 sizeof(rx_function_entry_v1_t);
7738 rxi_Free(rpc_stat, space);
7740 /* rx_rpc_stats must be held */
7741 rxi_rpc_peer_stat_cnt -= num_funcs;
7743 MUTEX_EXIT(&peer->peer_lock);
7744 MUTEX_EXIT(&rx_rpc_stats);
7748 if (rx_stats_active)
7749 rx_atomic_dec(&rx_stats.nPeerStructs);
7751 MUTEX_EXIT(&rx_peerHashTable_lock);
7754 for (i = 0; i < RX_MAX_SERVICES; i++) {
7756 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7758 for (i = 0; i < rx_hashTableSize; i++) {
7759 struct rx_connection *tc, *ntc;
7760 MUTEX_ENTER(&rx_connHashTable_lock);
7761 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7763 for (j = 0; j < RX_MAXCALLS; j++) {
7765 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7768 rxi_Free(tc, sizeof(*tc));
7770 MUTEX_EXIT(&rx_connHashTable_lock);
7773 MUTEX_ENTER(&freeSQEList_lock);
7775 while ((np = rx_FreeSQEList)) {
7776 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7777 MUTEX_DESTROY(&np->lock);
7778 rxi_Free(np, sizeof(*np));
7781 MUTEX_EXIT(&freeSQEList_lock);
7782 MUTEX_DESTROY(&freeSQEList_lock);
7783 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7784 MUTEX_DESTROY(&rx_connHashTable_lock);
7785 MUTEX_DESTROY(&rx_peerHashTable_lock);
7786 MUTEX_DESTROY(&rx_serverPool_lock);
7788 osi_Free(rx_connHashTable,
7789 rx_hashTableSize * sizeof(struct rx_connection *));
7790 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7792 UNPIN(rx_connHashTable,
7793 rx_hashTableSize * sizeof(struct rx_connection *));
7794 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7796 rxi_FreeAllPackets();
7798 MUTEX_ENTER(&rx_quota_mutex);
7799 rxi_dataQuota = RX_MAX_QUOTA;
7800 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7801 MUTEX_EXIT(&rx_quota_mutex);
7806 #ifdef RX_ENABLE_LOCKS
7808 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7810 if (!MUTEX_ISMINE(lockaddr))
7811 osi_Panic("Lock not held: %s", msg);
7813 #endif /* RX_ENABLE_LOCKS */
7818 * Routines to implement connection specific data.
7822 rx_KeyCreate(rx_destructor_t rtn)
7825 MUTEX_ENTER(&rxi_keyCreate_lock);
7826 key = rxi_keyCreate_counter++;
7827 rxi_keyCreate_destructor = (rx_destructor_t *)
7828 realloc((void *)rxi_keyCreate_destructor,
7829 (key + 1) * sizeof(rx_destructor_t));
7830 rxi_keyCreate_destructor[key] = rtn;
7831 MUTEX_EXIT(&rxi_keyCreate_lock);
7836 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7839 MUTEX_ENTER(&conn->conn_data_lock);
7840 if (!conn->specific) {
7841 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7842 for (i = 0; i < key; i++)
7843 conn->specific[i] = NULL;
7844 conn->nSpecific = key + 1;
7845 conn->specific[key] = ptr;
7846 } else if (key >= conn->nSpecific) {
7847 conn->specific = (void **)
7848 realloc(conn->specific, (key + 1) * sizeof(void *));
7849 for (i = conn->nSpecific; i < key; i++)
7850 conn->specific[i] = NULL;
7851 conn->nSpecific = key + 1;
7852 conn->specific[key] = ptr;
7854 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7855 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7856 conn->specific[key] = ptr;
7858 MUTEX_EXIT(&conn->conn_data_lock);
7862 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7865 MUTEX_ENTER(&svc->svc_data_lock);
7866 if (!svc->specific) {
7867 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
7868 for (i = 0; i < key; i++)
7869 svc->specific[i] = NULL;
7870 svc->nSpecific = key + 1;
7871 svc->specific[key] = ptr;
7872 } else if (key >= svc->nSpecific) {
7873 svc->specific = (void **)
7874 realloc(svc->specific, (key + 1) * sizeof(void *));
7875 for (i = svc->nSpecific; i < key; i++)
7876 svc->specific[i] = NULL;
7877 svc->nSpecific = key + 1;
7878 svc->specific[key] = ptr;
7880 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7881 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7882 svc->specific[key] = ptr;
7884 MUTEX_EXIT(&svc->svc_data_lock);
7888 rx_GetSpecific(struct rx_connection *conn, int key)
7891 MUTEX_ENTER(&conn->conn_data_lock);
7892 if (key >= conn->nSpecific)
7895 ptr = conn->specific[key];
7896 MUTEX_EXIT(&conn->conn_data_lock);
7901 rx_GetServiceSpecific(struct rx_service *svc, int key)
7904 MUTEX_ENTER(&svc->svc_data_lock);
7905 if (key >= svc->nSpecific)
7908 ptr = svc->specific[key];
7909 MUTEX_EXIT(&svc->svc_data_lock);
7914 #endif /* !KERNEL */
7917 * processStats is a queue used to store the statistics for the local
7918 * process. Its contents are similar to the contents of the rpcStats
7919 * queue on a rx_peer structure, but the actual data stored within
7920 * this queue contains totals across the lifetime of the process (assuming
7921 * the stats have not been reset) - unlike the per peer structures
7922 * which can come and go based upon the peer lifetime.
7925 static struct rx_queue processStats = { &processStats, &processStats };
7928 * peerStats is a queue used to store the statistics for all peer structs.
7929 * Its contents are the union of all the peer rpcStats queues.
7932 static struct rx_queue peerStats = { &peerStats, &peerStats };
7935 * rxi_monitor_processStats is used to turn process wide stat collection
7939 static int rxi_monitor_processStats = 0;
7942 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7945 static int rxi_monitor_peerStats = 0;
7948 * rxi_AddRpcStat - given all of the information for a particular rpc
7949 * call, create (if needed) and update the stat totals for the rpc.
7953 * IN stats - the queue of stats that will be updated with the new value
7955 * IN rxInterface - a unique number that identifies the rpc interface
7957 * IN currentFunc - the index of the function being invoked
7959 * IN totalFunc - the total number of functions in this interface
7961 * IN queueTime - the amount of time this function waited for a thread
7963 * IN execTime - the amount of time this function invocation took to execute
7965 * IN bytesSent - the number bytes sent by this invocation
7967 * IN bytesRcvd - the number bytes received by this invocation
7969 * IN isServer - if true, this invocation was made to a server
7971 * IN remoteHost - the ip address of the remote host
7973 * IN remotePort - the port of the remote host
7975 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7977 * INOUT counter - if a new stats structure is allocated, the counter will
7978 * be updated with the new number of allocated stat structures
7986 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
7987 afs_uint32 currentFunc, afs_uint32 totalFunc,
7988 struct clock *queueTime, struct clock *execTime,
7989 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
7990 afs_uint32 remoteHost, afs_uint32 remotePort,
7991 int addToPeerList, unsigned int *counter)
7994 rx_interface_stat_p rpc_stat, nrpc_stat;
7997 * See if there's already a structure for this interface
8000 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8001 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8002 && (rpc_stat->stats[0].remote_is_server == isServer))
8007 * Didn't find a match so allocate a new structure and add it to the
8011 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
8012 || (rpc_stat->stats[0].interfaceId != rxInterface)
8013 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8018 sizeof(rx_interface_stat_t) +
8019 totalFunc * sizeof(rx_function_entry_v1_t);
8021 rpc_stat = rxi_Alloc(space);
8022 if (rpc_stat == NULL) {
8026 *counter += totalFunc;
8027 for (i = 0; i < totalFunc; i++) {
8028 rpc_stat->stats[i].remote_peer = remoteHost;
8029 rpc_stat->stats[i].remote_port = remotePort;
8030 rpc_stat->stats[i].remote_is_server = isServer;
8031 rpc_stat->stats[i].interfaceId = rxInterface;
8032 rpc_stat->stats[i].func_total = totalFunc;
8033 rpc_stat->stats[i].func_index = i;
8034 hzero(rpc_stat->stats[i].invocations);
8035 hzero(rpc_stat->stats[i].bytes_sent);
8036 hzero(rpc_stat->stats[i].bytes_rcvd);
8037 rpc_stat->stats[i].queue_time_sum.sec = 0;
8038 rpc_stat->stats[i].queue_time_sum.usec = 0;
8039 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8040 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8041 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8042 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8043 rpc_stat->stats[i].queue_time_max.sec = 0;
8044 rpc_stat->stats[i].queue_time_max.usec = 0;
8045 rpc_stat->stats[i].execution_time_sum.sec = 0;
8046 rpc_stat->stats[i].execution_time_sum.usec = 0;
8047 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8048 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8049 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8050 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8051 rpc_stat->stats[i].execution_time_max.sec = 0;
8052 rpc_stat->stats[i].execution_time_max.usec = 0;
8054 queue_Prepend(stats, rpc_stat);
8055 if (addToPeerList) {
8056 queue_Prepend(&peerStats, &rpc_stat->all_peers);
8061 * Increment the stats for this function
8064 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
8065 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
8066 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
8067 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8068 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8069 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8070 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8072 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8073 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8075 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8076 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8078 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8079 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8081 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8082 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8090 * rx_IncrementTimeAndCount - increment the times and count for a particular
8095 * IN peer - the peer who invoked the rpc
8097 * IN rxInterface - a unique number that identifies the rpc interface
8099 * IN currentFunc - the index of the function being invoked
8101 * IN totalFunc - the total number of functions in this interface
8103 * IN queueTime - the amount of time this function waited for a thread
8105 * IN execTime - the amount of time this function invocation took to execute
8107 * IN bytesSent - the number bytes sent by this invocation
8109 * IN bytesRcvd - the number bytes received by this invocation
8111 * IN isServer - if true, this invocation was made to a server
8119 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8120 afs_uint32 currentFunc, afs_uint32 totalFunc,
8121 struct clock *queueTime, struct clock *execTime,
8122 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8126 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8129 MUTEX_ENTER(&rx_rpc_stats);
8131 if (rxi_monitor_peerStats) {
8132 MUTEX_ENTER(&peer->peer_lock);
8133 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8134 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8135 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8136 MUTEX_EXIT(&peer->peer_lock);
8139 if (rxi_monitor_processStats) {
8140 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8141 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8142 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8145 MUTEX_EXIT(&rx_rpc_stats);
8150 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8154 * IN callerVersion - the rpc stat version of the caller.
8156 * IN count - the number of entries to marshall.
8158 * IN stats - pointer to stats to be marshalled.
8160 * OUT ptr - Where to store the marshalled data.
8167 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8168 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8174 * We only support the first version
8176 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8177 *(ptr++) = stats->remote_peer;
8178 *(ptr++) = stats->remote_port;
8179 *(ptr++) = stats->remote_is_server;
8180 *(ptr++) = stats->interfaceId;
8181 *(ptr++) = stats->func_total;
8182 *(ptr++) = stats->func_index;
8183 *(ptr++) = hgethi(stats->invocations);
8184 *(ptr++) = hgetlo(stats->invocations);
8185 *(ptr++) = hgethi(stats->bytes_sent);
8186 *(ptr++) = hgetlo(stats->bytes_sent);
8187 *(ptr++) = hgethi(stats->bytes_rcvd);
8188 *(ptr++) = hgetlo(stats->bytes_rcvd);
8189 *(ptr++) = stats->queue_time_sum.sec;
8190 *(ptr++) = stats->queue_time_sum.usec;
8191 *(ptr++) = stats->queue_time_sum_sqr.sec;
8192 *(ptr++) = stats->queue_time_sum_sqr.usec;
8193 *(ptr++) = stats->queue_time_min.sec;
8194 *(ptr++) = stats->queue_time_min.usec;
8195 *(ptr++) = stats->queue_time_max.sec;
8196 *(ptr++) = stats->queue_time_max.usec;
8197 *(ptr++) = stats->execution_time_sum.sec;
8198 *(ptr++) = stats->execution_time_sum.usec;
8199 *(ptr++) = stats->execution_time_sum_sqr.sec;
8200 *(ptr++) = stats->execution_time_sum_sqr.usec;
8201 *(ptr++) = stats->execution_time_min.sec;
8202 *(ptr++) = stats->execution_time_min.usec;
8203 *(ptr++) = stats->execution_time_max.sec;
8204 *(ptr++) = stats->execution_time_max.usec;
8210 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8215 * IN callerVersion - the rpc stat version of the caller
8217 * OUT myVersion - the rpc stat version of this function
8219 * OUT clock_sec - local time seconds
8221 * OUT clock_usec - local time microseconds
8223 * OUT allocSize - the number of bytes allocated to contain stats
8225 * OUT statCount - the number stats retrieved from this process.
8227 * OUT stats - the actual stats retrieved from this process.
8231 * Returns void. If successful, stats will != NULL.
8235 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8236 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8237 size_t * allocSize, afs_uint32 * statCount,
8238 afs_uint32 ** stats)
8248 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8251 * Check to see if stats are enabled
8254 MUTEX_ENTER(&rx_rpc_stats);
8255 if (!rxi_monitor_processStats) {
8256 MUTEX_EXIT(&rx_rpc_stats);
8260 clock_GetTime(&now);
8261 *clock_sec = now.sec;
8262 *clock_usec = now.usec;
8265 * Allocate the space based upon the caller version
8267 * If the client is at an older version than we are,
8268 * we return the statistic data in the older data format, but
8269 * we still return our version number so the client knows we
8270 * are maintaining more data than it can retrieve.
8273 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8274 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8275 *statCount = rxi_rpc_process_stat_cnt;
8278 * This can't happen yet, but in the future version changes
8279 * can be handled by adding additional code here
8283 if (space > (size_t) 0) {
8285 ptr = *stats = rxi_Alloc(space);
8288 rx_interface_stat_p rpc_stat, nrpc_stat;
8292 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8294 * Copy the data based upon the caller version
8296 rx_MarshallProcessRPCStats(callerVersion,
8297 rpc_stat->stats[0].func_total,
8298 rpc_stat->stats, &ptr);
8304 MUTEX_EXIT(&rx_rpc_stats);
8309 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8313 * IN callerVersion - the rpc stat version of the caller
8315 * OUT myVersion - the rpc stat version of this function
8317 * OUT clock_sec - local time seconds
8319 * OUT clock_usec - local time microseconds
8321 * OUT allocSize - the number of bytes allocated to contain stats
8323 * OUT statCount - the number of stats retrieved from the individual
8326 * OUT stats - the actual stats retrieved from the individual peer structures.
8330 * Returns void. If successful, stats will != NULL.
8334 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8335 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8336 size_t * allocSize, afs_uint32 * statCount,
8337 afs_uint32 ** stats)
8347 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8350 * Check to see if stats are enabled
8353 MUTEX_ENTER(&rx_rpc_stats);
8354 if (!rxi_monitor_peerStats) {
8355 MUTEX_EXIT(&rx_rpc_stats);
8359 clock_GetTime(&now);
8360 *clock_sec = now.sec;
8361 *clock_usec = now.usec;
8364 * Allocate the space based upon the caller version
8366 * If the client is at an older version than we are,
8367 * we return the statistic data in the older data format, but
8368 * we still return our version number so the client knows we
8369 * are maintaining more data than it can retrieve.
8372 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8373 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8374 *statCount = rxi_rpc_peer_stat_cnt;
8377 * This can't happen yet, but in the future version changes
8378 * can be handled by adding additional code here
8382 if (space > (size_t) 0) {
8384 ptr = *stats = rxi_Alloc(space);
8387 rx_interface_stat_p rpc_stat, nrpc_stat;
8391 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8393 * We have to fix the offset of rpc_stat since we are
8394 * keeping this structure on two rx_queues. The rx_queue
8395 * package assumes that the rx_queue member is the first
8396 * member of the structure. That is, rx_queue assumes that
8397 * any one item is only on one queue at a time. We are
8398 * breaking that assumption and so we have to do a little
8399 * math to fix our pointers.
8402 fix_offset = (char *)rpc_stat;
8403 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8404 rpc_stat = (rx_interface_stat_p) fix_offset;
8407 * Copy the data based upon the caller version
8409 rx_MarshallProcessRPCStats(callerVersion,
8410 rpc_stat->stats[0].func_total,
8411 rpc_stat->stats, &ptr);
8417 MUTEX_EXIT(&rx_rpc_stats);
8422 * rx_FreeRPCStats - free memory allocated by
8423 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8427 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8428 * rx_RetrievePeerRPCStats
8430 * IN allocSize - the number of bytes in stats.
8438 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8440 rxi_Free(stats, allocSize);
8444 * rx_queryProcessRPCStats - see if process rpc stat collection is
8445 * currently enabled.
8451 * Returns 0 if stats are not enabled != 0 otherwise
8455 rx_queryProcessRPCStats(void)
8458 MUTEX_ENTER(&rx_rpc_stats);
8459 rc = rxi_monitor_processStats;
8460 MUTEX_EXIT(&rx_rpc_stats);
8465 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8471 * Returns 0 if stats are not enabled != 0 otherwise
8475 rx_queryPeerRPCStats(void)
8478 MUTEX_ENTER(&rx_rpc_stats);
8479 rc = rxi_monitor_peerStats;
8480 MUTEX_EXIT(&rx_rpc_stats);
8485 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8495 rx_enableProcessRPCStats(void)
8497 MUTEX_ENTER(&rx_rpc_stats);
8498 rx_enable_stats = 1;
8499 rxi_monitor_processStats = 1;
8500 MUTEX_EXIT(&rx_rpc_stats);
8504 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8514 rx_enablePeerRPCStats(void)
8516 MUTEX_ENTER(&rx_rpc_stats);
8517 rx_enable_stats = 1;
8518 rxi_monitor_peerStats = 1;
8519 MUTEX_EXIT(&rx_rpc_stats);
8523 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8533 rx_disableProcessRPCStats(void)
8535 rx_interface_stat_p rpc_stat, nrpc_stat;
8538 MUTEX_ENTER(&rx_rpc_stats);
8541 * Turn off process statistics and if peer stats is also off, turn
8545 rxi_monitor_processStats = 0;
8546 if (rxi_monitor_peerStats == 0) {
8547 rx_enable_stats = 0;
8550 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8551 unsigned int num_funcs = 0;
8554 queue_Remove(rpc_stat);
8555 num_funcs = rpc_stat->stats[0].func_total;
8557 sizeof(rx_interface_stat_t) +
8558 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8560 rxi_Free(rpc_stat, space);
8561 rxi_rpc_process_stat_cnt -= num_funcs;
8563 MUTEX_EXIT(&rx_rpc_stats);
8567 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8577 rx_disablePeerRPCStats(void)
8579 struct rx_peer **peer_ptr, **peer_end;
8583 * Turn off peer statistics and if process stats is also off, turn
8587 rxi_monitor_peerStats = 0;
8588 if (rxi_monitor_processStats == 0) {
8589 rx_enable_stats = 0;
8592 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8593 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8595 struct rx_peer *peer, *next, *prev;
8597 MUTEX_ENTER(&rx_peerHashTable_lock);
8598 MUTEX_ENTER(&rx_rpc_stats);
8599 for (prev = peer = *peer_ptr; peer; peer = next) {
8601 code = MUTEX_TRYENTER(&peer->peer_lock);
8603 rx_interface_stat_p rpc_stat, nrpc_stat;
8606 if (prev == *peer_ptr) {
8617 MUTEX_EXIT(&rx_peerHashTable_lock);
8620 (&peer->rpcStats, rpc_stat, nrpc_stat,
8621 rx_interface_stat)) {
8622 unsigned int num_funcs = 0;
8625 queue_Remove(&rpc_stat->queue_header);
8626 queue_Remove(&rpc_stat->all_peers);
8627 num_funcs = rpc_stat->stats[0].func_total;
8629 sizeof(rx_interface_stat_t) +
8630 rpc_stat->stats[0].func_total *
8631 sizeof(rx_function_entry_v1_t);
8633 rxi_Free(rpc_stat, space);
8634 rxi_rpc_peer_stat_cnt -= num_funcs;
8636 MUTEX_EXIT(&peer->peer_lock);
8638 MUTEX_ENTER(&rx_peerHashTable_lock);
8648 MUTEX_EXIT(&rx_rpc_stats);
8649 MUTEX_EXIT(&rx_peerHashTable_lock);
8654 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8659 * IN clearFlag - flag indicating which stats to clear
8667 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8669 rx_interface_stat_p rpc_stat, nrpc_stat;
8671 MUTEX_ENTER(&rx_rpc_stats);
8673 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8674 unsigned int num_funcs = 0, i;
8675 num_funcs = rpc_stat->stats[0].func_total;
8676 for (i = 0; i < num_funcs; i++) {
8677 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8678 hzero(rpc_stat->stats[i].invocations);
8680 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8681 hzero(rpc_stat->stats[i].bytes_sent);
8683 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8684 hzero(rpc_stat->stats[i].bytes_rcvd);
8686 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8687 rpc_stat->stats[i].queue_time_sum.sec = 0;
8688 rpc_stat->stats[i].queue_time_sum.usec = 0;
8690 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8691 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8692 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8694 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8695 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8696 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8698 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8699 rpc_stat->stats[i].queue_time_max.sec = 0;
8700 rpc_stat->stats[i].queue_time_max.usec = 0;
8702 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8703 rpc_stat->stats[i].execution_time_sum.sec = 0;
8704 rpc_stat->stats[i].execution_time_sum.usec = 0;
8706 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8707 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8708 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8710 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8711 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8712 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8714 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8715 rpc_stat->stats[i].execution_time_max.sec = 0;
8716 rpc_stat->stats[i].execution_time_max.usec = 0;
8721 MUTEX_EXIT(&rx_rpc_stats);
8725 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8730 * IN clearFlag - flag indicating which stats to clear
8738 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8740 rx_interface_stat_p rpc_stat, nrpc_stat;
8742 MUTEX_ENTER(&rx_rpc_stats);
8744 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8745 unsigned int num_funcs = 0, i;
8748 * We have to fix the offset of rpc_stat since we are
8749 * keeping this structure on two rx_queues. The rx_queue
8750 * package assumes that the rx_queue member is the first
8751 * member of the structure. That is, rx_queue assumes that
8752 * any one item is only on one queue at a time. We are
8753 * breaking that assumption and so we have to do a little
8754 * math to fix our pointers.
8757 fix_offset = (char *)rpc_stat;
8758 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8759 rpc_stat = (rx_interface_stat_p) fix_offset;
8761 num_funcs = rpc_stat->stats[0].func_total;
8762 for (i = 0; i < num_funcs; i++) {
8763 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8764 hzero(rpc_stat->stats[i].invocations);
8766 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8767 hzero(rpc_stat->stats[i].bytes_sent);
8769 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8770 hzero(rpc_stat->stats[i].bytes_rcvd);
8772 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8773 rpc_stat->stats[i].queue_time_sum.sec = 0;
8774 rpc_stat->stats[i].queue_time_sum.usec = 0;
8776 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8777 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8778 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8780 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8781 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8782 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8784 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8785 rpc_stat->stats[i].queue_time_max.sec = 0;
8786 rpc_stat->stats[i].queue_time_max.usec = 0;
8788 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8789 rpc_stat->stats[i].execution_time_sum.sec = 0;
8790 rpc_stat->stats[i].execution_time_sum.usec = 0;
8792 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8793 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8794 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8796 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8797 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8798 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8800 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8801 rpc_stat->stats[i].execution_time_max.sec = 0;
8802 rpc_stat->stats[i].execution_time_max.usec = 0;
8807 MUTEX_EXIT(&rx_rpc_stats);
8811 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8812 * is authorized to enable/disable/clear RX statistics.
8814 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8817 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8819 rxi_rxstat_userok = proc;
8823 rx_RxStatUserOk(struct rx_call *call)
8825 if (!rxi_rxstat_userok)
8827 return rxi_rxstat_userok(call);
8832 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8833 * function in the MSVC runtime DLL (msvcrt.dll).
8835 * Note: the system serializes calls to this function.
8838 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8839 DWORD reason, /* reason function is being called */
8840 LPVOID reserved) /* reserved for future use */
8843 case DLL_PROCESS_ATTACH:
8844 /* library is being attached to a process */
8848 case DLL_PROCESS_DETACH:
8855 #endif /* AFS_NT40_ENV */
8858 int rx_DumpCalls(FILE *outputFile, char *cookie)
8860 #ifdef RXDEBUG_PACKET
8861 #ifdef KDUMP_RX_LOCK
8862 struct rx_call_rx_lock *c;
8869 #define RXDPRINTF sprintf
8870 #define RXDPRINTOUT output
8872 #define RXDPRINTF fprintf
8873 #define RXDPRINTOUT outputFile
8876 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8878 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8881 for (c = rx_allCallsp; c; c = c->allNextp) {
8882 u_short rqc, tqc, iovqc;
8883 struct rx_packet *p, *np;
8885 MUTEX_ENTER(&c->lock);
8886 queue_Count(&c->rq, p, np, rx_packet, rqc);
8887 queue_Count(&c->tq, p, np, rx_packet, tqc);
8888 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8890 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, "
8891 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8892 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8893 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8894 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8895 #ifdef RX_ENABLE_LOCKS
8898 #ifdef RX_REFCOUNT_CHECK
8899 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
8900 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
8903 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,
8904 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
8905 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
8906 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
8907 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
8908 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
8909 #ifdef RX_ENABLE_LOCKS
8910 , (afs_uint32)c->refCount
8912 #ifdef RX_REFCOUNT_CHECK
8913 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
8916 MUTEX_EXIT(&c->lock);
8919 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8922 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
8924 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8926 #endif /* RXDEBUG_PACKET */