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 osi_Assert(pthread_cond_init
233 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
234 osi_Assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
236 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
237 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
239 rxkad_global_stats_init();
241 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
242 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
243 #ifdef RX_ENABLE_LOCKS
246 #endif /* RX_LOCKS_DB */
247 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
248 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
250 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
252 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
254 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
256 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
257 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
258 #endif /* RX_ENABLE_LOCKS */
261 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
262 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
264 * The rx_stats_mutex mutex protects the following global variables:
265 * rxi_lowConnRefCount
266 * rxi_lowPeerRefCount
275 * The rx_quota_mutex mutex protects the following global variables:
283 * The rx_freePktQ_lock protects the following global variables:
288 * The rx_packets_mutex mutex protects the following global variables:
296 * The rx_pthread_mutex mutex protects the following global variables:
297 * rxi_fcfs_thread_num
300 #define INIT_PTHREAD_LOCKS
304 /* Variables for handling the minProcs implementation. availProcs gives the
305 * number of threads available in the pool at this moment (not counting dudes
306 * executing right now). totalMin gives the total number of procs required
307 * for handling all minProcs requests. minDeficit is a dynamic variable
308 * tracking the # of procs required to satisfy all of the remaining minProcs
310 * For fine grain locking to work, the quota check and the reservation of
311 * a server thread has to come while rxi_availProcs and rxi_minDeficit
312 * are locked. To this end, the code has been modified under #ifdef
313 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
314 * same time. A new function, ReturnToServerPool() returns the allocation.
316 * A call can be on several queue's (but only one at a time). When
317 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
318 * that no one else is touching the queue. To this end, we store the address
319 * of the queue lock in the call structure (under the call lock) when we
320 * put the call on a queue, and we clear the call_queue_lock when the
321 * call is removed from a queue (once the call lock has been obtained).
322 * This allows rxi_ResetCall to safely synchronize with others wishing
323 * to manipulate the queue.
326 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
327 static afs_kmutex_t rx_rpc_stats;
328 void rxi_StartUnlocked(struct rxevent *event, void *call,
329 void *arg1, int istack);
332 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
333 ** pretty good that the next packet coming in is from the same connection
334 ** as the last packet, since we're send multiple packets in a transmit window.
336 struct rx_connection *rxLastConn = 0;
338 #ifdef RX_ENABLE_LOCKS
339 /* The locking hierarchy for rx fine grain locking is composed of these
342 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
343 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
344 * call->lock - locks call data fields.
345 * These are independent of each other:
346 * rx_freeCallQueue_lock
351 * serverQueueEntry->lock
352 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
354 * peer->lock - locks peer data fields.
355 * conn_data_lock - that more than one thread is not updating a conn data
356 * field at the same time.
367 * Do we need a lock to protect the peer field in the conn structure?
368 * conn->peer was previously a constant for all intents and so has no
369 * lock protecting this field. The multihomed client delta introduced
370 * a RX code change : change the peer field in the connection structure
371 * to that remote interface from which the last packet for this
372 * connection was sent out. This may become an issue if further changes
375 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
376 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
378 /* rxdb_fileID is used to identify the lock location, along with line#. */
379 static int rxdb_fileID = RXDB_FILE_RX;
380 #endif /* RX_LOCKS_DB */
381 #else /* RX_ENABLE_LOCKS */
382 #define SET_CALL_QUEUE_LOCK(C, L)
383 #define CLEAR_CALL_QUEUE_LOCK(C)
384 #endif /* RX_ENABLE_LOCKS */
385 struct rx_serverQueueEntry *rx_waitForPacket = 0;
386 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
388 /* ------------Exported Interfaces------------- */
390 /* This function allows rxkad to set the epoch to a suitably random number
391 * which rx_NewConnection will use in the future. The principle purpose is to
392 * get rxnull connections to use the same epoch as the rxkad connections do, at
393 * least once the first rxkad connection is established. This is important now
394 * that the host/port addresses aren't used in FindConnection: the uniqueness
395 * of epoch/cid matters and the start time won't do. */
397 #ifdef AFS_PTHREAD_ENV
399 * This mutex protects the following global variables:
403 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
404 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
408 #endif /* AFS_PTHREAD_ENV */
411 rx_SetEpoch(afs_uint32 epoch)
418 /* Initialize rx. A port number may be mentioned, in which case this
419 * becomes the default port number for any service installed later.
420 * If 0 is provided for the port number, a random port will be chosen
421 * by the kernel. Whether this will ever overlap anything in
422 * /etc/services is anybody's guess... Returns 0 on success, -1 on
427 int rxinit_status = 1;
428 #ifdef AFS_PTHREAD_ENV
430 * This mutex protects the following global variables:
434 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
435 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
438 #define UNLOCK_RX_INIT
442 rx_InitHost(u_int host, u_int port)
449 char *htable, *ptable;
456 if (rxinit_status == 0) {
457 tmp_status = rxinit_status;
459 return tmp_status; /* Already started; return previous error code. */
465 if (afs_winsockInit() < 0)
471 * Initialize anything necessary to provide a non-premptive threading
474 rxi_InitializeThreadSupport();
477 /* Allocate and initialize a socket for client and perhaps server
480 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
481 if (rx_socket == OSI_NULLSOCKET) {
485 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
488 #endif /* RX_LOCKS_DB */
489 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
490 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
491 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
492 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
493 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
494 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
495 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
496 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
497 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
499 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
501 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
503 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
505 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
506 #if defined(AFS_HPUX110_ENV)
508 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
509 #endif /* AFS_HPUX110_ENV */
510 #endif /* RX_ENABLE_LOCKS && KERNEL */
513 rx_connDeadTime = 12;
514 rx_tranquil = 0; /* reset flag */
515 rxi_ResetStatistics();
517 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
518 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
519 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
520 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
521 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
522 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
524 /* Malloc up a bunch of packets & buffers */
526 queue_Init(&rx_freePacketQueue);
527 rxi_NeedMorePackets = FALSE;
528 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
530 /* enforce a minimum number of allocated packets */
531 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
532 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
534 /* allocate the initial free packet pool */
535 #ifdef RX_ENABLE_TSFPQ
536 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
537 #else /* RX_ENABLE_TSFPQ */
538 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
539 #endif /* RX_ENABLE_TSFPQ */
546 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
547 tv.tv_sec = clock_now.sec;
548 tv.tv_usec = clock_now.usec;
549 srand((unsigned int)tv.tv_usec);
556 #if defined(KERNEL) && !defined(UKERNEL)
557 /* Really, this should never happen in a real kernel */
560 struct sockaddr_in addr;
562 int addrlen = sizeof(addr);
564 socklen_t addrlen = sizeof(addr);
566 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
570 rx_port = addr.sin_port;
573 rx_stats.minRtt.sec = 9999999;
575 rx_SetEpoch(tv.tv_sec | 0x80000000);
577 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
578 * will provide a randomer value. */
580 MUTEX_ENTER(&rx_quota_mutex);
581 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
582 MUTEX_EXIT(&rx_quota_mutex);
583 /* *Slightly* random start time for the cid. This is just to help
584 * out with the hashing function at the peer */
585 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
586 rx_connHashTable = (struct rx_connection **)htable;
587 rx_peerHashTable = (struct rx_peer **)ptable;
589 rx_lastAckDelay.sec = 0;
590 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
591 rx_hardAckDelay.sec = 0;
592 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
593 rx_softAckDelay.sec = 0;
594 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
596 rxevent_Init(20, rxi_ReScheduleEvents);
598 /* Initialize various global queues */
599 queue_Init(&rx_idleServerQueue);
600 queue_Init(&rx_incomingCallQueue);
601 queue_Init(&rx_freeCallQueue);
603 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
604 /* Initialize our list of usable IP addresses. */
608 /* Start listener process (exact function is dependent on the
609 * implementation environment--kernel or user space) */
613 tmp_status = rxinit_status = 0;
621 return rx_InitHost(htonl(INADDR_ANY), port);
624 /* called with unincremented nRequestsRunning to see if it is OK to start
625 * a new thread in this service. Could be "no" for two reasons: over the
626 * max quota, or would prevent others from reaching their min quota.
628 #ifdef RX_ENABLE_LOCKS
629 /* This verion of QuotaOK reserves quota if it's ok while the
630 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
633 QuotaOK(struct rx_service *aservice)
635 /* check if over max quota */
636 if (aservice->nRequestsRunning >= aservice->maxProcs) {
640 /* under min quota, we're OK */
641 /* otherwise, can use only if there are enough to allow everyone
642 * to go to their min quota after this guy starts.
645 MUTEX_ENTER(&rx_quota_mutex);
646 if ((aservice->nRequestsRunning < aservice->minProcs)
647 || (rxi_availProcs > rxi_minDeficit)) {
648 aservice->nRequestsRunning++;
649 /* just started call in minProcs pool, need fewer to maintain
651 if (aservice->nRequestsRunning <= aservice->minProcs)
654 MUTEX_EXIT(&rx_quota_mutex);
657 MUTEX_EXIT(&rx_quota_mutex);
663 ReturnToServerPool(struct rx_service *aservice)
665 aservice->nRequestsRunning--;
666 MUTEX_ENTER(&rx_quota_mutex);
667 if (aservice->nRequestsRunning < aservice->minProcs)
670 MUTEX_EXIT(&rx_quota_mutex);
673 #else /* RX_ENABLE_LOCKS */
675 QuotaOK(struct rx_service *aservice)
678 /* under min quota, we're OK */
679 if (aservice->nRequestsRunning < aservice->minProcs)
682 /* check if over max quota */
683 if (aservice->nRequestsRunning >= aservice->maxProcs)
686 /* otherwise, can use only if there are enough to allow everyone
687 * to go to their min quota after this guy starts.
689 MUTEX_ENTER(&rx_quota_mutex);
690 if (rxi_availProcs > rxi_minDeficit)
692 MUTEX_EXIT(&rx_quota_mutex);
695 #endif /* RX_ENABLE_LOCKS */
698 /* Called by rx_StartServer to start up lwp's to service calls.
699 NExistingProcs gives the number of procs already existing, and which
700 therefore needn't be created. */
702 rxi_StartServerProcs(int nExistingProcs)
704 struct rx_service *service;
709 /* For each service, reserve N processes, where N is the "minimum"
710 * number of processes that MUST be able to execute a request in parallel,
711 * at any time, for that process. Also compute the maximum difference
712 * between any service's maximum number of processes that can run
713 * (i.e. the maximum number that ever will be run, and a guarantee
714 * that this number will run if other services aren't running), and its
715 * minimum number. The result is the extra number of processes that
716 * we need in order to provide the latter guarantee */
717 for (i = 0; i < RX_MAX_SERVICES; i++) {
719 service = rx_services[i];
720 if (service == (struct rx_service *)0)
722 nProcs += service->minProcs;
723 diff = service->maxProcs - service->minProcs;
727 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
728 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
729 for (i = 0; i < nProcs; i++) {
730 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
736 /* This routine is only required on Windows */
738 rx_StartClientThread(void)
740 #ifdef AFS_PTHREAD_ENV
742 pid = pthread_self();
743 #endif /* AFS_PTHREAD_ENV */
745 #endif /* AFS_NT40_ENV */
747 /* This routine must be called if any services are exported. If the
748 * donateMe flag is set, the calling process is donated to the server
751 rx_StartServer(int donateMe)
753 struct rx_service *service;
759 /* Start server processes, if necessary (exact function is dependent
760 * on the implementation environment--kernel or user space). DonateMe
761 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
762 * case, one less new proc will be created rx_StartServerProcs.
764 rxi_StartServerProcs(donateMe);
766 /* count up the # of threads in minProcs, and add set the min deficit to
767 * be that value, too.
769 for (i = 0; i < RX_MAX_SERVICES; i++) {
770 service = rx_services[i];
771 if (service == (struct rx_service *)0)
773 MUTEX_ENTER(&rx_quota_mutex);
774 rxi_totalMin += service->minProcs;
775 /* below works even if a thread is running, since minDeficit would
776 * still have been decremented and later re-incremented.
778 rxi_minDeficit += service->minProcs;
779 MUTEX_EXIT(&rx_quota_mutex);
782 /* Turn on reaping of idle server connections */
783 rxi_ReapConnections(NULL, NULL, NULL);
792 #ifdef AFS_PTHREAD_ENV
794 pid = afs_pointer_to_int(pthread_self());
795 #else /* AFS_PTHREAD_ENV */
797 LWP_CurrentProcess(&pid);
798 #endif /* AFS_PTHREAD_ENV */
800 sprintf(name, "srv_%d", ++nProcs);
802 (*registerProgram) (pid, name);
804 #endif /* AFS_NT40_ENV */
805 rx_ServerProc(NULL); /* Never returns */
807 #ifdef RX_ENABLE_TSFPQ
808 /* no use leaving packets around in this thread's local queue if
809 * it isn't getting donated to the server thread pool.
811 rxi_FlushLocalPacketsTSFPQ();
812 #endif /* RX_ENABLE_TSFPQ */
816 /* Create a new client connection to the specified service, using the
817 * specified security object to implement the security model for this
819 struct rx_connection *
820 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
821 struct rx_securityClass *securityObject,
822 int serviceSecurityIndex)
826 struct rx_connection *conn;
831 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
832 "serviceSecurityIndex %d)\n",
833 ntohl(shost), ntohs(sport), sservice, securityObject,
834 serviceSecurityIndex));
836 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
837 * the case of kmem_alloc? */
838 conn = rxi_AllocConnection();
839 #ifdef RX_ENABLE_LOCKS
840 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
841 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
842 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
845 MUTEX_ENTER(&rx_connHashTable_lock);
846 cid = (rx_nextCid += RX_MAXCALLS);
847 conn->type = RX_CLIENT_CONNECTION;
849 conn->epoch = rx_epoch;
850 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
851 conn->serviceId = sservice;
852 conn->securityObject = securityObject;
853 conn->securityData = (void *) 0;
854 conn->securityIndex = serviceSecurityIndex;
855 rx_SetConnDeadTime(conn, rx_connDeadTime);
856 rx_SetConnSecondsUntilNatPing(conn, 0);
857 conn->ackRate = RX_FAST_ACK_RATE;
859 conn->specific = NULL;
860 conn->challengeEvent = NULL;
861 conn->delayedAbortEvent = NULL;
862 conn->abortCount = 0;
864 for (i = 0; i < RX_MAXCALLS; i++) {
865 conn->twind[i] = rx_initSendWindow;
866 conn->rwind[i] = rx_initReceiveWindow;
869 RXS_NewConnection(securityObject, conn);
871 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
873 conn->refCount++; /* no lock required since only this thread knows... */
874 conn->next = rx_connHashTable[hashindex];
875 rx_connHashTable[hashindex] = conn;
877 rx_atomic_inc(&rx_stats.nClientConns);
878 MUTEX_EXIT(&rx_connHashTable_lock);
884 * Ensure a connection's timeout values are valid.
886 * @param[in] conn The connection to check
888 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
889 * unless idleDeadTime and/or hardDeadTime are not set
893 rxi_CheckConnTimeouts(struct rx_connection *conn)
895 /* a connection's timeouts must have the relationship
896 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
897 * total loss of network to a peer may cause an idle timeout instead of a
898 * dead timeout, simply because the idle timeout gets hit first. Also set
899 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
900 /* this logic is slightly complicated by the fact that
901 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
903 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
904 if (conn->idleDeadTime) {
905 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
907 if (conn->hardDeadTime) {
908 if (conn->idleDeadTime) {
909 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
911 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
917 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
919 /* The idea is to set the dead time to a value that allows several
920 * keepalives to be dropped without timing out the connection. */
921 conn->secondsUntilDead = seconds;
922 rxi_CheckConnTimeouts(conn);
923 conn->secondsUntilPing = conn->secondsUntilDead / 6;
927 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
929 conn->hardDeadTime = seconds;
930 rxi_CheckConnTimeouts(conn);
934 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
936 conn->idleDeadTime = seconds;
937 rxi_CheckConnTimeouts(conn);
940 int rxi_lowPeerRefCount = 0;
941 int rxi_lowConnRefCount = 0;
944 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
945 * NOTE: must not be called with rx_connHashTable_lock held.
948 rxi_CleanupConnection(struct rx_connection *conn)
950 /* Notify the service exporter, if requested, that this connection
951 * is being destroyed */
952 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
953 (*conn->service->destroyConnProc) (conn);
955 /* Notify the security module that this connection is being destroyed */
956 RXS_DestroyConnection(conn->securityObject, conn);
958 /* If this is the last connection using the rx_peer struct, set its
959 * idle time to now. rxi_ReapConnections will reap it if it's still
960 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
962 MUTEX_ENTER(&rx_peerHashTable_lock);
963 if (conn->peer->refCount < 2) {
964 conn->peer->idleWhen = clock_Sec();
965 if (conn->peer->refCount < 1) {
966 conn->peer->refCount = 1;
967 if (rx_stats_active) {
968 MUTEX_ENTER(&rx_stats_mutex);
969 rxi_lowPeerRefCount++;
970 MUTEX_EXIT(&rx_stats_mutex);
974 conn->peer->refCount--;
975 MUTEX_EXIT(&rx_peerHashTable_lock);
979 if (conn->type == RX_SERVER_CONNECTION)
980 rx_atomic_dec(&rx_stats.nServerConns);
982 rx_atomic_dec(&rx_stats.nClientConns);
985 if (conn->specific) {
987 for (i = 0; i < conn->nSpecific; i++) {
988 if (conn->specific[i] && rxi_keyCreate_destructor[i])
989 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
990 conn->specific[i] = NULL;
992 free(conn->specific);
994 conn->specific = NULL;
998 MUTEX_DESTROY(&conn->conn_call_lock);
999 MUTEX_DESTROY(&conn->conn_data_lock);
1000 CV_DESTROY(&conn->conn_call_cv);
1002 rxi_FreeConnection(conn);
1005 /* Destroy the specified connection */
1007 rxi_DestroyConnection(struct rx_connection *conn)
1009 MUTEX_ENTER(&rx_connHashTable_lock);
1010 rxi_DestroyConnectionNoLock(conn);
1011 /* conn should be at the head of the cleanup list */
1012 if (conn == rx_connCleanup_list) {
1013 rx_connCleanup_list = rx_connCleanup_list->next;
1014 MUTEX_EXIT(&rx_connHashTable_lock);
1015 rxi_CleanupConnection(conn);
1017 #ifdef RX_ENABLE_LOCKS
1019 MUTEX_EXIT(&rx_connHashTable_lock);
1021 #endif /* RX_ENABLE_LOCKS */
1025 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1027 struct rx_connection **conn_ptr;
1029 struct rx_packet *packet;
1036 MUTEX_ENTER(&conn->conn_data_lock);
1037 MUTEX_ENTER(&rx_refcnt_mutex);
1038 if (conn->refCount > 0)
1041 if (rx_stats_active) {
1042 MUTEX_ENTER(&rx_stats_mutex);
1043 rxi_lowConnRefCount++;
1044 MUTEX_EXIT(&rx_stats_mutex);
1048 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1049 /* Busy; wait till the last guy before proceeding */
1050 MUTEX_EXIT(&rx_refcnt_mutex);
1051 MUTEX_EXIT(&conn->conn_data_lock);
1056 /* If the client previously called rx_NewCall, but it is still
1057 * waiting, treat this as a running call, and wait to destroy the
1058 * connection later when the call completes. */
1059 if ((conn->type == RX_CLIENT_CONNECTION)
1060 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1061 conn->flags |= RX_CONN_DESTROY_ME;
1062 MUTEX_EXIT(&conn->conn_data_lock);
1066 MUTEX_EXIT(&rx_refcnt_mutex);
1067 MUTEX_EXIT(&conn->conn_data_lock);
1069 /* Check for extant references to this connection */
1070 for (i = 0; i < RX_MAXCALLS; i++) {
1071 struct rx_call *call = conn->call[i];
1074 if (conn->type == RX_CLIENT_CONNECTION) {
1075 MUTEX_ENTER(&call->lock);
1076 if (call->delayedAckEvent) {
1077 /* Push the final acknowledgment out now--there
1078 * won't be a subsequent call to acknowledge the
1079 * last reply packets */
1080 rxevent_Cancel(call->delayedAckEvent, call,
1081 RX_CALL_REFCOUNT_DELAY);
1082 if (call->state == RX_STATE_PRECALL
1083 || call->state == RX_STATE_ACTIVE) {
1084 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1086 rxi_AckAll(NULL, call, 0);
1089 MUTEX_EXIT(&call->lock);
1093 #ifdef RX_ENABLE_LOCKS
1095 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1096 MUTEX_EXIT(&conn->conn_data_lock);
1098 /* Someone is accessing a packet right now. */
1102 #endif /* RX_ENABLE_LOCKS */
1105 /* Don't destroy the connection if there are any call
1106 * structures still in use */
1107 MUTEX_ENTER(&conn->conn_data_lock);
1108 conn->flags |= RX_CONN_DESTROY_ME;
1109 MUTEX_EXIT(&conn->conn_data_lock);
1114 if (conn->natKeepAliveEvent) {
1115 rxi_NatKeepAliveOff(conn);
1118 if (conn->delayedAbortEvent) {
1119 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1120 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1122 MUTEX_ENTER(&conn->conn_data_lock);
1123 rxi_SendConnectionAbort(conn, packet, 0, 1);
1124 MUTEX_EXIT(&conn->conn_data_lock);
1125 rxi_FreePacket(packet);
1129 /* Remove from connection hash table before proceeding */
1131 &rx_connHashTable[CONN_HASH
1132 (peer->host, peer->port, conn->cid, conn->epoch,
1134 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1135 if (*conn_ptr == conn) {
1136 *conn_ptr = conn->next;
1140 /* if the conn that we are destroying was the last connection, then we
1141 * clear rxLastConn as well */
1142 if (rxLastConn == conn)
1145 /* Make sure the connection is completely reset before deleting it. */
1146 /* get rid of pending events that could zap us later */
1147 if (conn->challengeEvent)
1148 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1149 if (conn->checkReachEvent)
1150 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1151 if (conn->natKeepAliveEvent)
1152 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
1154 /* Add the connection to the list of destroyed connections that
1155 * need to be cleaned up. This is necessary to avoid deadlocks
1156 * in the routines we call to inform others that this connection is
1157 * being destroyed. */
1158 conn->next = rx_connCleanup_list;
1159 rx_connCleanup_list = conn;
1162 /* Externally available version */
1164 rx_DestroyConnection(struct rx_connection *conn)
1169 rxi_DestroyConnection(conn);
1174 rx_GetConnection(struct rx_connection *conn)
1179 MUTEX_ENTER(&rx_refcnt_mutex);
1181 MUTEX_EXIT(&rx_refcnt_mutex);
1185 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1186 /* Wait for the transmit queue to no longer be busy.
1187 * requires the call->lock to be held */
1189 rxi_WaitforTQBusy(struct rx_call *call) {
1190 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1191 call->flags |= RX_CALL_TQ_WAIT;
1193 #ifdef RX_ENABLE_LOCKS
1194 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1195 CV_WAIT(&call->cv_tq, &call->lock);
1196 #else /* RX_ENABLE_LOCKS */
1197 osi_rxSleep(&call->tq);
1198 #endif /* RX_ENABLE_LOCKS */
1200 if (call->tqWaiters == 0) {
1201 call->flags &= ~RX_CALL_TQ_WAIT;
1208 rxi_WakeUpTransmitQueue(struct rx_call *call)
1210 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1211 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1212 call, call->tqWaiters, call->flags));
1213 #ifdef RX_ENABLE_LOCKS
1214 osirx_AssertMine(&call->lock, "rxi_Start start");
1215 CV_BROADCAST(&call->cv_tq);
1216 #else /* RX_ENABLE_LOCKS */
1217 osi_rxWakeup(&call->tq);
1218 #endif /* RX_ENABLE_LOCKS */
1222 /* Start a new rx remote procedure call, on the specified connection.
1223 * If wait is set to 1, wait for a free call channel; otherwise return
1224 * 0. Maxtime gives the maximum number of seconds this call may take,
1225 * after rx_NewCall returns. After this time interval, a call to any
1226 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1227 * For fine grain locking, we hold the conn_call_lock in order to
1228 * to ensure that we don't get signalle after we found a call in an active
1229 * state and before we go to sleep.
1232 rx_NewCall(struct rx_connection *conn)
1235 struct rx_call *call;
1236 struct clock queueTime;
1240 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1243 clock_GetTime(&queueTime);
1245 * Check if there are others waiting for a new call.
1246 * If so, let them go first to avoid starving them.
1247 * This is a fairly simple scheme, and might not be
1248 * a complete solution for large numbers of waiters.
1250 * makeCallWaiters keeps track of the number of
1251 * threads waiting to make calls and the
1252 * RX_CONN_MAKECALL_WAITING flag bit is used to
1253 * indicate that there are indeed calls waiting.
1254 * The flag is set when the waiter is incremented.
1255 * It is only cleared when makeCallWaiters is 0.
1256 * This prevents us from accidently destroying the
1257 * connection while it is potentially about to be used.
1259 MUTEX_ENTER(&conn->conn_call_lock);
1260 MUTEX_ENTER(&conn->conn_data_lock);
1261 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1262 conn->flags |= RX_CONN_MAKECALL_WAITING;
1263 conn->makeCallWaiters++;
1264 MUTEX_EXIT(&conn->conn_data_lock);
1266 #ifdef RX_ENABLE_LOCKS
1267 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1271 MUTEX_ENTER(&conn->conn_data_lock);
1272 conn->makeCallWaiters--;
1273 if (conn->makeCallWaiters == 0)
1274 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1277 /* We are now the active thread in rx_NewCall */
1278 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1279 MUTEX_EXIT(&conn->conn_data_lock);
1284 for (i = 0; i < RX_MAXCALLS; i++) {
1285 call = conn->call[i];
1287 if (call->state == RX_STATE_DALLY) {
1288 MUTEX_ENTER(&call->lock);
1289 if (call->state == RX_STATE_DALLY) {
1291 * We are setting the state to RX_STATE_RESET to
1292 * ensure that no one else will attempt to use this
1293 * call once we drop the conn->conn_call_lock and
1294 * call->lock. We must drop the conn->conn_call_lock
1295 * before calling rxi_ResetCall because the process
1296 * of clearing the transmit queue can block for an
1297 * extended period of time. If we block while holding
1298 * the conn->conn_call_lock, then all rx_EndCall
1299 * processing will block as well. This has a detrimental
1300 * effect on overall system performance.
1302 call->state = RX_STATE_RESET;
1303 MUTEX_EXIT(&conn->conn_call_lock);
1304 MUTEX_ENTER(&rx_refcnt_mutex);
1305 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1306 MUTEX_EXIT(&rx_refcnt_mutex);
1307 rxi_ResetCall(call, 0);
1308 (*call->callNumber)++;
1309 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1313 * If we failed to be able to safely obtain the
1314 * conn->conn_call_lock we will have to drop the
1315 * call->lock to avoid a deadlock. When the call->lock
1316 * is released the state of the call can change. If it
1317 * is no longer RX_STATE_RESET then some other thread is
1320 MUTEX_EXIT(&call->lock);
1321 MUTEX_ENTER(&conn->conn_call_lock);
1322 MUTEX_ENTER(&call->lock);
1324 if (call->state == RX_STATE_RESET)
1328 * If we get here it means that after dropping
1329 * the conn->conn_call_lock and call->lock that
1330 * the call is no longer ours. If we can't find
1331 * a free call in the remaining slots we should
1332 * not go immediately to RX_CONN_MAKECALL_WAITING
1333 * because by dropping the conn->conn_call_lock
1334 * we have given up synchronization with rx_EndCall.
1335 * Instead, cycle through one more time to see if
1336 * we can find a call that can call our own.
1338 MUTEX_ENTER(&rx_refcnt_mutex);
1339 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1340 MUTEX_EXIT(&rx_refcnt_mutex);
1343 MUTEX_EXIT(&call->lock);
1346 /* rxi_NewCall returns with mutex locked */
1347 call = rxi_NewCall(conn, i);
1348 MUTEX_ENTER(&rx_refcnt_mutex);
1349 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1350 MUTEX_EXIT(&rx_refcnt_mutex);
1354 if (i < RX_MAXCALLS) {
1360 MUTEX_ENTER(&conn->conn_data_lock);
1361 conn->flags |= RX_CONN_MAKECALL_WAITING;
1362 conn->makeCallWaiters++;
1363 MUTEX_EXIT(&conn->conn_data_lock);
1365 #ifdef RX_ENABLE_LOCKS
1366 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1370 MUTEX_ENTER(&conn->conn_data_lock);
1371 conn->makeCallWaiters--;
1372 if (conn->makeCallWaiters == 0)
1373 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1374 MUTEX_EXIT(&conn->conn_data_lock);
1376 /* Client is initially in send mode */
1377 call->state = RX_STATE_ACTIVE;
1378 call->error = conn->error;
1380 call->mode = RX_MODE_ERROR;
1382 call->mode = RX_MODE_SENDING;
1384 /* remember start time for call in case we have hard dead time limit */
1385 call->queueTime = queueTime;
1386 clock_GetTime(&call->startTime);
1387 hzero(call->bytesSent);
1388 hzero(call->bytesRcvd);
1390 /* Turn on busy protocol. */
1391 rxi_KeepAliveOn(call);
1393 /* Attempt MTU discovery */
1394 rxi_GrowMTUOn(call);
1397 * We are no longer the active thread in rx_NewCall
1399 MUTEX_ENTER(&conn->conn_data_lock);
1400 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1401 MUTEX_EXIT(&conn->conn_data_lock);
1404 * Wake up anyone else who might be giving us a chance to
1405 * run (see code above that avoids resource starvation).
1407 #ifdef RX_ENABLE_LOCKS
1408 CV_BROADCAST(&conn->conn_call_cv);
1412 MUTEX_EXIT(&conn->conn_call_lock);
1414 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1415 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1416 osi_Panic("rx_NewCall call about to be used without an empty tq");
1418 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1420 MUTEX_EXIT(&call->lock);
1423 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1428 rxi_HasActiveCalls(struct rx_connection *aconn)
1431 struct rx_call *tcall;
1435 for (i = 0; i < RX_MAXCALLS; i++) {
1436 if ((tcall = aconn->call[i])) {
1437 if ((tcall->state == RX_STATE_ACTIVE)
1438 || (tcall->state == RX_STATE_PRECALL)) {
1449 rxi_GetCallNumberVector(struct rx_connection *aconn,
1450 afs_int32 * aint32s)
1453 struct rx_call *tcall;
1457 for (i = 0; i < RX_MAXCALLS; i++) {
1458 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1459 aint32s[i] = aconn->callNumber[i] + 1;
1461 aint32s[i] = aconn->callNumber[i];
1468 rxi_SetCallNumberVector(struct rx_connection *aconn,
1469 afs_int32 * aint32s)
1472 struct rx_call *tcall;
1476 for (i = 0; i < RX_MAXCALLS; i++) {
1477 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1478 aconn->callNumber[i] = aint32s[i] - 1;
1480 aconn->callNumber[i] = aint32s[i];
1486 /* Advertise a new service. A service is named locally by a UDP port
1487 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1490 char *serviceName; Name for identification purposes (e.g. the
1491 service name might be used for probing for
1494 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1495 char *serviceName, struct rx_securityClass **securityObjects,
1496 int nSecurityObjects,
1497 afs_int32(*serviceProc) (struct rx_call * acall))
1499 osi_socket socket = OSI_NULLSOCKET;
1500 struct rx_service *tservice;
1506 if (serviceId == 0) {
1508 "rx_NewService: service id for service %s is not non-zero.\n",
1515 "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",
1523 tservice = rxi_AllocService();
1526 #ifdef RX_ENABLE_LOCKS
1527 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1530 for (i = 0; i < RX_MAX_SERVICES; i++) {
1531 struct rx_service *service = rx_services[i];
1533 if (port == service->servicePort && host == service->serviceHost) {
1534 if (service->serviceId == serviceId) {
1535 /* The identical service has already been
1536 * installed; if the caller was intending to
1537 * change the security classes used by this
1538 * service, he/she loses. */
1540 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1541 serviceName, serviceId, service->serviceName);
1543 rxi_FreeService(tservice);
1546 /* Different service, same port: re-use the socket
1547 * which is bound to the same port */
1548 socket = service->socket;
1551 if (socket == OSI_NULLSOCKET) {
1552 /* If we don't already have a socket (from another
1553 * service on same port) get a new one */
1554 socket = rxi_GetHostUDPSocket(host, port);
1555 if (socket == OSI_NULLSOCKET) {
1557 rxi_FreeService(tservice);
1562 service->socket = socket;
1563 service->serviceHost = host;
1564 service->servicePort = port;
1565 service->serviceId = serviceId;
1566 service->serviceName = serviceName;
1567 service->nSecurityObjects = nSecurityObjects;
1568 service->securityObjects = securityObjects;
1569 service->minProcs = 0;
1570 service->maxProcs = 1;
1571 service->idleDeadTime = 60;
1572 service->idleDeadErr = 0;
1573 service->connDeadTime = rx_connDeadTime;
1574 service->executeRequestProc = serviceProc;
1575 service->checkReach = 0;
1576 service->nSpecific = 0;
1577 service->specific = NULL;
1578 rx_services[i] = service; /* not visible until now */
1584 rxi_FreeService(tservice);
1585 (osi_Msg "rx_NewService: cannot support > %d services\n",
1590 /* Set configuration options for all of a service's security objects */
1593 rx_SetSecurityConfiguration(struct rx_service *service,
1594 rx_securityConfigVariables type,
1598 for (i = 0; i<service->nSecurityObjects; i++) {
1599 if (service->securityObjects[i]) {
1600 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1608 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1609 struct rx_securityClass **securityObjects, int nSecurityObjects,
1610 afs_int32(*serviceProc) (struct rx_call * acall))
1612 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1615 /* Generic request processing loop. This routine should be called
1616 * by the implementation dependent rx_ServerProc. If socketp is
1617 * non-null, it will be set to the file descriptor that this thread
1618 * is now listening on. If socketp is null, this routine will never
1621 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1623 struct rx_call *call;
1625 struct rx_service *tservice = NULL;
1632 call = rx_GetCall(threadID, tservice, socketp);
1633 if (socketp && *socketp != OSI_NULLSOCKET) {
1634 /* We are now a listener thread */
1639 /* if server is restarting( typically smooth shutdown) then do not
1640 * allow any new calls.
1643 if (rx_tranquil && (call != NULL)) {
1647 MUTEX_ENTER(&call->lock);
1649 rxi_CallError(call, RX_RESTARTING);
1650 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1652 MUTEX_EXIT(&call->lock);
1656 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1657 #ifdef RX_ENABLE_LOCKS
1659 #endif /* RX_ENABLE_LOCKS */
1660 afs_termState = AFSOP_STOP_AFS;
1661 afs_osi_Wakeup(&afs_termState);
1662 #ifdef RX_ENABLE_LOCKS
1664 #endif /* RX_ENABLE_LOCKS */
1669 tservice = call->conn->service;
1671 if (tservice->beforeProc)
1672 (*tservice->beforeProc) (call);
1674 code = tservice->executeRequestProc(call);
1676 if (tservice->afterProc)
1677 (*tservice->afterProc) (call, code);
1679 rx_EndCall(call, code);
1680 if (rx_stats_active) {
1681 MUTEX_ENTER(&rx_stats_mutex);
1683 MUTEX_EXIT(&rx_stats_mutex);
1690 rx_WakeupServerProcs(void)
1692 struct rx_serverQueueEntry *np, *tqp;
1696 MUTEX_ENTER(&rx_serverPool_lock);
1698 #ifdef RX_ENABLE_LOCKS
1699 if (rx_waitForPacket)
1700 CV_BROADCAST(&rx_waitForPacket->cv);
1701 #else /* RX_ENABLE_LOCKS */
1702 if (rx_waitForPacket)
1703 osi_rxWakeup(rx_waitForPacket);
1704 #endif /* RX_ENABLE_LOCKS */
1705 MUTEX_ENTER(&freeSQEList_lock);
1706 for (np = rx_FreeSQEList; np; np = tqp) {
1707 tqp = *(struct rx_serverQueueEntry **)np;
1708 #ifdef RX_ENABLE_LOCKS
1709 CV_BROADCAST(&np->cv);
1710 #else /* RX_ENABLE_LOCKS */
1712 #endif /* RX_ENABLE_LOCKS */
1714 MUTEX_EXIT(&freeSQEList_lock);
1715 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1716 #ifdef RX_ENABLE_LOCKS
1717 CV_BROADCAST(&np->cv);
1718 #else /* RX_ENABLE_LOCKS */
1720 #endif /* RX_ENABLE_LOCKS */
1722 MUTEX_EXIT(&rx_serverPool_lock);
1727 * One thing that seems to happen is that all the server threads get
1728 * tied up on some empty or slow call, and then a whole bunch of calls
1729 * arrive at once, using up the packet pool, so now there are more
1730 * empty calls. The most critical resources here are server threads
1731 * and the free packet pool. The "doreclaim" code seems to help in
1732 * general. I think that eventually we arrive in this state: there
1733 * are lots of pending calls which do have all their packets present,
1734 * so they won't be reclaimed, are multi-packet calls, so they won't
1735 * be scheduled until later, and thus are tying up most of the free
1736 * packet pool for a very long time.
1738 * 1. schedule multi-packet calls if all the packets are present.
1739 * Probably CPU-bound operation, useful to return packets to pool.
1740 * Do what if there is a full window, but the last packet isn't here?
1741 * 3. preserve one thread which *only* runs "best" calls, otherwise
1742 * it sleeps and waits for that type of call.
1743 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1744 * the current dataquota business is badly broken. The quota isn't adjusted
1745 * to reflect how many packets are presently queued for a running call.
1746 * So, when we schedule a queued call with a full window of packets queued
1747 * up for it, that *should* free up a window full of packets for other 2d-class
1748 * calls to be able to use from the packet pool. But it doesn't.
1750 * NB. Most of the time, this code doesn't run -- since idle server threads
1751 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1752 * as a new call arrives.
1754 /* Sleep until a call arrives. Returns a pointer to the call, ready
1755 * for an rx_Read. */
1756 #ifdef RX_ENABLE_LOCKS
1758 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1760 struct rx_serverQueueEntry *sq;
1761 struct rx_call *call = (struct rx_call *)0;
1762 struct rx_service *service = NULL;
1765 MUTEX_ENTER(&freeSQEList_lock);
1767 if ((sq = rx_FreeSQEList)) {
1768 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1769 MUTEX_EXIT(&freeSQEList_lock);
1770 } else { /* otherwise allocate a new one and return that */
1771 MUTEX_EXIT(&freeSQEList_lock);
1772 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1773 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1774 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1777 MUTEX_ENTER(&rx_serverPool_lock);
1778 if (cur_service != NULL) {
1779 ReturnToServerPool(cur_service);
1782 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1783 struct rx_call *tcall, *ncall, *choice2 = NULL;
1785 /* Scan for eligible incoming calls. A call is not eligible
1786 * if the maximum number of calls for its service type are
1787 * already executing */
1788 /* One thread will process calls FCFS (to prevent starvation),
1789 * while the other threads may run ahead looking for calls which
1790 * have all their input data available immediately. This helps
1791 * keep threads from blocking, waiting for data from the client. */
1792 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1793 service = tcall->conn->service;
1794 if (!QuotaOK(service)) {
1797 MUTEX_ENTER(&rx_pthread_mutex);
1798 if (tno == rxi_fcfs_thread_num
1799 || !tcall->queue_item_header.next) {
1800 MUTEX_EXIT(&rx_pthread_mutex);
1801 /* If we're the fcfs thread , then we'll just use
1802 * this call. If we haven't been able to find an optimal
1803 * choice, and we're at the end of the list, then use a
1804 * 2d choice if one has been identified. Otherwise... */
1805 call = (choice2 ? choice2 : tcall);
1806 service = call->conn->service;
1808 MUTEX_EXIT(&rx_pthread_mutex);
1809 if (!queue_IsEmpty(&tcall->rq)) {
1810 struct rx_packet *rp;
1811 rp = queue_First(&tcall->rq, rx_packet);
1812 if (rp->header.seq == 1) {
1814 || (rp->header.flags & RX_LAST_PACKET)) {
1816 } else if (rxi_2dchoice && !choice2
1817 && !(tcall->flags & RX_CALL_CLEARED)
1818 && (tcall->rprev > rxi_HardAckRate)) {
1828 ReturnToServerPool(service);
1835 MUTEX_EXIT(&rx_serverPool_lock);
1836 MUTEX_ENTER(&call->lock);
1838 if (call->flags & RX_CALL_WAIT_PROC) {
1839 call->flags &= ~RX_CALL_WAIT_PROC;
1840 rx_atomic_dec(&rx_nWaiting);
1843 if (call->state != RX_STATE_PRECALL || call->error) {
1844 MUTEX_EXIT(&call->lock);
1845 MUTEX_ENTER(&rx_serverPool_lock);
1846 ReturnToServerPool(service);
1851 if (queue_IsEmpty(&call->rq)
1852 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1853 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1855 CLEAR_CALL_QUEUE_LOCK(call);
1858 /* If there are no eligible incoming calls, add this process
1859 * to the idle server queue, to wait for one */
1863 *socketp = OSI_NULLSOCKET;
1865 sq->socketp = socketp;
1866 queue_Append(&rx_idleServerQueue, sq);
1867 #ifndef AFS_AIX41_ENV
1868 rx_waitForPacket = sq;
1870 rx_waitingForPacket = sq;
1871 #endif /* AFS_AIX41_ENV */
1873 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1875 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1876 MUTEX_EXIT(&rx_serverPool_lock);
1877 return (struct rx_call *)0;
1880 } while (!(call = sq->newcall)
1881 && !(socketp && *socketp != OSI_NULLSOCKET));
1882 MUTEX_EXIT(&rx_serverPool_lock);
1884 MUTEX_ENTER(&call->lock);
1890 MUTEX_ENTER(&freeSQEList_lock);
1891 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1892 rx_FreeSQEList = sq;
1893 MUTEX_EXIT(&freeSQEList_lock);
1896 clock_GetTime(&call->startTime);
1897 call->state = RX_STATE_ACTIVE;
1898 call->mode = RX_MODE_RECEIVING;
1899 #ifdef RX_KERNEL_TRACE
1900 if (ICL_SETACTIVE(afs_iclSetp)) {
1901 int glockOwner = ISAFS_GLOCK();
1904 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1905 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1912 rxi_calltrace(RX_CALL_START, call);
1913 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
1914 call->conn->service->servicePort, call->conn->service->serviceId,
1917 MUTEX_EXIT(&call->lock);
1918 MUTEX_ENTER(&rx_refcnt_mutex);
1919 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1920 MUTEX_EXIT(&rx_refcnt_mutex);
1922 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1927 #else /* RX_ENABLE_LOCKS */
1929 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1931 struct rx_serverQueueEntry *sq;
1932 struct rx_call *call = (struct rx_call *)0, *choice2;
1933 struct rx_service *service = NULL;
1937 MUTEX_ENTER(&freeSQEList_lock);
1939 if ((sq = rx_FreeSQEList)) {
1940 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1941 MUTEX_EXIT(&freeSQEList_lock);
1942 } else { /* otherwise allocate a new one and return that */
1943 MUTEX_EXIT(&freeSQEList_lock);
1944 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1945 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1946 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1948 MUTEX_ENTER(&sq->lock);
1950 if (cur_service != NULL) {
1951 cur_service->nRequestsRunning--;
1952 MUTEX_ENTER(&rx_quota_mutex);
1953 if (cur_service->nRequestsRunning < cur_service->minProcs)
1956 MUTEX_EXIT(&rx_quota_mutex);
1958 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1959 struct rx_call *tcall, *ncall;
1960 /* Scan for eligible incoming calls. A call is not eligible
1961 * if the maximum number of calls for its service type are
1962 * already executing */
1963 /* One thread will process calls FCFS (to prevent starvation),
1964 * while the other threads may run ahead looking for calls which
1965 * have all their input data available immediately. This helps
1966 * keep threads from blocking, waiting for data from the client. */
1967 choice2 = (struct rx_call *)0;
1968 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1969 service = tcall->conn->service;
1970 if (QuotaOK(service)) {
1971 MUTEX_ENTER(&rx_pthread_mutex);
1972 if (tno == rxi_fcfs_thread_num
1973 || !tcall->queue_item_header.next) {
1974 MUTEX_EXIT(&rx_pthread_mutex);
1975 /* If we're the fcfs thread, then we'll just use
1976 * this call. If we haven't been able to find an optimal
1977 * choice, and we're at the end of the list, then use a
1978 * 2d choice if one has been identified. Otherwise... */
1979 call = (choice2 ? choice2 : tcall);
1980 service = call->conn->service;
1982 MUTEX_EXIT(&rx_pthread_mutex);
1983 if (!queue_IsEmpty(&tcall->rq)) {
1984 struct rx_packet *rp;
1985 rp = queue_First(&tcall->rq, rx_packet);
1986 if (rp->header.seq == 1
1988 || (rp->header.flags & RX_LAST_PACKET))) {
1990 } else if (rxi_2dchoice && !choice2
1991 && !(tcall->flags & RX_CALL_CLEARED)
1992 && (tcall->rprev > rxi_HardAckRate)) {
2006 /* we can't schedule a call if there's no data!!! */
2007 /* send an ack if there's no data, if we're missing the
2008 * first packet, or we're missing something between first
2009 * and last -- there's a "hole" in the incoming data. */
2010 if (queue_IsEmpty(&call->rq)
2011 || queue_First(&call->rq, rx_packet)->header.seq != 1
2012 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
2013 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2015 call->flags &= (~RX_CALL_WAIT_PROC);
2016 service->nRequestsRunning++;
2017 /* just started call in minProcs pool, need fewer to maintain
2019 MUTEX_ENTER(&rx_quota_mutex);
2020 if (service->nRequestsRunning <= service->minProcs)
2023 MUTEX_EXIT(&rx_quota_mutex);
2024 rx_atomic_dec(&rx_nWaiting);
2025 /* MUTEX_EXIT(&call->lock); */
2027 /* If there are no eligible incoming calls, add this process
2028 * to the idle server queue, to wait for one */
2031 *socketp = OSI_NULLSOCKET;
2033 sq->socketp = socketp;
2034 queue_Append(&rx_idleServerQueue, sq);
2038 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2040 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2041 return (struct rx_call *)0;
2044 } while (!(call = sq->newcall)
2045 && !(socketp && *socketp != OSI_NULLSOCKET));
2047 MUTEX_EXIT(&sq->lock);
2049 MUTEX_ENTER(&freeSQEList_lock);
2050 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2051 rx_FreeSQEList = sq;
2052 MUTEX_EXIT(&freeSQEList_lock);
2055 clock_GetTime(&call->startTime);
2056 call->state = RX_STATE_ACTIVE;
2057 call->mode = RX_MODE_RECEIVING;
2058 #ifdef RX_KERNEL_TRACE
2059 if (ICL_SETACTIVE(afs_iclSetp)) {
2060 int glockOwner = ISAFS_GLOCK();
2063 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2064 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2071 rxi_calltrace(RX_CALL_START, call);
2072 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2073 call->conn->service->servicePort, call->conn->service->serviceId,
2076 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2083 #endif /* RX_ENABLE_LOCKS */
2087 /* Establish a procedure to be called when a packet arrives for a
2088 * call. This routine will be called at most once after each call,
2089 * and will also be called if there is an error condition on the or
2090 * the call is complete. Used by multi rx to build a selection
2091 * function which determines which of several calls is likely to be a
2092 * good one to read from.
2093 * NOTE: the way this is currently implemented it is probably only a
2094 * good idea to (1) use it immediately after a newcall (clients only)
2095 * and (2) only use it once. Other uses currently void your warranty
2098 rx_SetArrivalProc(struct rx_call *call,
2099 void (*proc) (struct rx_call * call,
2102 void * handle, int arg)
2104 call->arrivalProc = proc;
2105 call->arrivalProcHandle = handle;
2106 call->arrivalProcArg = arg;
2109 /* Call is finished (possibly prematurely). Return rc to the peer, if
2110 * appropriate, and return the final error code from the conversation
2114 rx_EndCall(struct rx_call *call, afs_int32 rc)
2116 struct rx_connection *conn = call->conn;
2120 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2121 call, rc, call->error, call->abortCode));
2124 MUTEX_ENTER(&call->lock);
2126 if (rc == 0 && call->error == 0) {
2127 call->abortCode = 0;
2128 call->abortCount = 0;
2131 call->arrivalProc = (void (*)())0;
2132 if (rc && call->error == 0) {
2133 rxi_CallError(call, rc);
2134 call->mode = RX_MODE_ERROR;
2135 /* Send an abort message to the peer if this error code has
2136 * only just been set. If it was set previously, assume the
2137 * peer has already been sent the error code or will request it
2139 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2141 if (conn->type == RX_SERVER_CONNECTION) {
2142 /* Make sure reply or at least dummy reply is sent */
2143 if (call->mode == RX_MODE_RECEIVING) {
2144 MUTEX_EXIT(&call->lock);
2145 rxi_WriteProc(call, 0, 0);
2146 MUTEX_ENTER(&call->lock);
2148 if (call->mode == RX_MODE_SENDING) {
2149 MUTEX_EXIT(&call->lock);
2150 rxi_FlushWrite(call);
2151 MUTEX_ENTER(&call->lock);
2153 rxi_calltrace(RX_CALL_END, call);
2154 /* Call goes to hold state until reply packets are acknowledged */
2155 if (call->tfirst + call->nSoftAcked < call->tnext) {
2156 call->state = RX_STATE_HOLD;
2158 call->state = RX_STATE_DALLY;
2159 rxi_ClearTransmitQueue(call, 0);
2160 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
2161 rxevent_Cancel(call->keepAliveEvent, call,
2162 RX_CALL_REFCOUNT_ALIVE);
2164 } else { /* Client connection */
2166 /* Make sure server receives input packets, in the case where
2167 * no reply arguments are expected */
2168 if ((call->mode == RX_MODE_SENDING)
2169 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2170 MUTEX_EXIT(&call->lock);
2171 (void)rxi_ReadProc(call, &dummy, 1);
2172 MUTEX_ENTER(&call->lock);
2175 /* If we had an outstanding delayed ack, be nice to the server
2176 * and force-send it now.
2178 if (call->delayedAckEvent) {
2179 rxevent_Cancel(call->delayedAckEvent, call,
2180 RX_CALL_REFCOUNT_DELAY);
2181 call->delayedAckEvent = NULL;
2182 rxi_SendDelayedAck(NULL, call, NULL);
2185 /* We need to release the call lock since it's lower than the
2186 * conn_call_lock and we don't want to hold the conn_call_lock
2187 * over the rx_ReadProc call. The conn_call_lock needs to be held
2188 * here for the case where rx_NewCall is perusing the calls on
2189 * the connection structure. We don't want to signal until
2190 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2191 * have checked this call, found it active and by the time it
2192 * goes to sleep, will have missed the signal.
2194 MUTEX_EXIT(&call->lock);
2195 MUTEX_ENTER(&conn->conn_call_lock);
2196 MUTEX_ENTER(&call->lock);
2197 MUTEX_ENTER(&conn->conn_data_lock);
2198 conn->flags |= RX_CONN_BUSY;
2199 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2200 MUTEX_EXIT(&conn->conn_data_lock);
2201 #ifdef RX_ENABLE_LOCKS
2202 CV_BROADCAST(&conn->conn_call_cv);
2207 #ifdef RX_ENABLE_LOCKS
2209 MUTEX_EXIT(&conn->conn_data_lock);
2211 #endif /* RX_ENABLE_LOCKS */
2212 call->state = RX_STATE_DALLY;
2214 error = call->error;
2216 /* currentPacket, nLeft, and NFree must be zeroed here, because
2217 * ResetCall cannot: ResetCall may be called at splnet(), in the
2218 * kernel version, and may interrupt the macros rx_Read or
2219 * rx_Write, which run at normal priority for efficiency. */
2220 if (call->currentPacket) {
2221 #ifdef RX_TRACK_PACKETS
2222 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2224 rxi_FreePacket(call->currentPacket);
2225 call->currentPacket = (struct rx_packet *)0;
2228 call->nLeft = call->nFree = call->curlen = 0;
2230 /* Free any packets from the last call to ReadvProc/WritevProc */
2231 #ifdef RXDEBUG_PACKET
2233 #endif /* RXDEBUG_PACKET */
2234 rxi_FreePackets(0, &call->iovq);
2235 MUTEX_EXIT(&call->lock);
2237 MUTEX_ENTER(&rx_refcnt_mutex);
2238 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2239 MUTEX_EXIT(&rx_refcnt_mutex);
2240 if (conn->type == RX_CLIENT_CONNECTION) {
2241 MUTEX_ENTER(&conn->conn_data_lock);
2242 conn->flags &= ~RX_CONN_BUSY;
2243 MUTEX_EXIT(&conn->conn_data_lock);
2244 MUTEX_EXIT(&conn->conn_call_lock);
2248 * Map errors to the local host's errno.h format.
2250 error = ntoh_syserr_conv(error);
2254 #if !defined(KERNEL)
2256 /* Call this routine when shutting down a server or client (especially
2257 * clients). This will allow Rx to gracefully garbage collect server
2258 * connections, and reduce the number of retries that a server might
2259 * make to a dead client.
2260 * This is not quite right, since some calls may still be ongoing and
2261 * we can't lock them to destroy them. */
2265 struct rx_connection **conn_ptr, **conn_end;
2269 if (rxinit_status == 1) {
2271 return; /* Already shutdown. */
2273 rxi_DeleteCachedConnections();
2274 if (rx_connHashTable) {
2275 MUTEX_ENTER(&rx_connHashTable_lock);
2276 for (conn_ptr = &rx_connHashTable[0], conn_end =
2277 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2279 struct rx_connection *conn, *next;
2280 for (conn = *conn_ptr; conn; conn = next) {
2282 if (conn->type == RX_CLIENT_CONNECTION) {
2283 MUTEX_ENTER(&rx_refcnt_mutex);
2285 MUTEX_EXIT(&rx_refcnt_mutex);
2286 #ifdef RX_ENABLE_LOCKS
2287 rxi_DestroyConnectionNoLock(conn);
2288 #else /* RX_ENABLE_LOCKS */
2289 rxi_DestroyConnection(conn);
2290 #endif /* RX_ENABLE_LOCKS */
2294 #ifdef RX_ENABLE_LOCKS
2295 while (rx_connCleanup_list) {
2296 struct rx_connection *conn;
2297 conn = rx_connCleanup_list;
2298 rx_connCleanup_list = rx_connCleanup_list->next;
2299 MUTEX_EXIT(&rx_connHashTable_lock);
2300 rxi_CleanupConnection(conn);
2301 MUTEX_ENTER(&rx_connHashTable_lock);
2303 MUTEX_EXIT(&rx_connHashTable_lock);
2304 #endif /* RX_ENABLE_LOCKS */
2309 afs_winsockCleanup();
2317 /* if we wakeup packet waiter too often, can get in loop with two
2318 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2320 rxi_PacketsUnWait(void)
2322 if (!rx_waitingForPackets) {
2326 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2327 return; /* still over quota */
2330 rx_waitingForPackets = 0;
2331 #ifdef RX_ENABLE_LOCKS
2332 CV_BROADCAST(&rx_waitingForPackets_cv);
2334 osi_rxWakeup(&rx_waitingForPackets);
2340 /* ------------------Internal interfaces------------------------- */
2342 /* Return this process's service structure for the
2343 * specified socket and service */
2345 rxi_FindService(osi_socket socket, u_short serviceId)
2347 struct rx_service **sp;
2348 for (sp = &rx_services[0]; *sp; sp++) {
2349 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2355 #ifdef RXDEBUG_PACKET
2356 #ifdef KDUMP_RX_LOCK
2357 static struct rx_call_rx_lock *rx_allCallsp = 0;
2359 static struct rx_call *rx_allCallsp = 0;
2361 #endif /* RXDEBUG_PACKET */
2363 /* Allocate a call structure, for the indicated channel of the
2364 * supplied connection. The mode and state of the call must be set by
2365 * the caller. Returns the call with mutex locked. */
2367 rxi_NewCall(struct rx_connection *conn, int channel)
2369 struct rx_call *call;
2370 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2371 struct rx_call *cp; /* Call pointer temp */
2372 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2373 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2375 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2377 /* Grab an existing call structure, or allocate a new one.
2378 * Existing call structures are assumed to have been left reset by
2380 MUTEX_ENTER(&rx_freeCallQueue_lock);
2382 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2384 * EXCEPT that the TQ might not yet be cleared out.
2385 * Skip over those with in-use TQs.
2388 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2389 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2395 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2396 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2397 call = queue_First(&rx_freeCallQueue, rx_call);
2398 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2400 if (rx_stats_active)
2401 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2402 MUTEX_EXIT(&rx_freeCallQueue_lock);
2403 MUTEX_ENTER(&call->lock);
2404 CLEAR_CALL_QUEUE_LOCK(call);
2405 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2406 /* Now, if TQ wasn't cleared earlier, do it now. */
2407 rxi_WaitforTQBusy(call);
2408 if (call->flags & RX_CALL_TQ_CLEARME) {
2409 rxi_ClearTransmitQueue(call, 1);
2410 /*queue_Init(&call->tq);*/
2412 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2413 /* Bind the call to its connection structure */
2415 rxi_ResetCall(call, 1);
2418 call = rxi_Alloc(sizeof(struct rx_call));
2419 #ifdef RXDEBUG_PACKET
2420 call->allNextp = rx_allCallsp;
2421 rx_allCallsp = call;
2423 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2424 #else /* RXDEBUG_PACKET */
2425 rx_atomic_inc(&rx_stats.nCallStructs);
2426 #endif /* RXDEBUG_PACKET */
2428 MUTEX_EXIT(&rx_freeCallQueue_lock);
2429 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2430 MUTEX_ENTER(&call->lock);
2431 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2432 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2433 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2435 /* Initialize once-only items */
2436 queue_Init(&call->tq);
2437 queue_Init(&call->rq);
2438 queue_Init(&call->iovq);
2439 #ifdef RXDEBUG_PACKET
2440 call->rqc = call->tqc = call->iovqc = 0;
2441 #endif /* RXDEBUG_PACKET */
2442 /* Bind the call to its connection structure (prereq for reset) */
2444 rxi_ResetCall(call, 1);
2446 call->channel = channel;
2447 call->callNumber = &conn->callNumber[channel];
2448 call->rwind = conn->rwind[channel];
2449 call->twind = conn->twind[channel];
2450 /* Note that the next expected call number is retained (in
2451 * conn->callNumber[i]), even if we reallocate the call structure
2453 conn->call[channel] = call;
2454 /* if the channel's never been used (== 0), we should start at 1, otherwise
2455 * the call number is valid from the last time this channel was used */
2456 if (*call->callNumber == 0)
2457 *call->callNumber = 1;
2462 /* A call has been inactive long enough that so we can throw away
2463 * state, including the call structure, which is placed on the call
2466 * call->lock amd rx_refcnt_mutex are held upon entry.
2467 * haveCTLock is set when called from rxi_ReapConnections.
2470 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2472 int channel = call->channel;
2473 struct rx_connection *conn = call->conn;
2476 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2477 (*call->callNumber)++;
2478 rxi_ResetCall(call, 0);
2479 call->conn->call[channel] = (struct rx_call *)0;
2480 MUTEX_EXIT(&rx_refcnt_mutex);
2482 MUTEX_ENTER(&rx_freeCallQueue_lock);
2483 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2484 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2485 /* A call may be free even though its transmit queue is still in use.
2486 * Since we search the call list from head to tail, put busy calls at
2487 * the head of the list, and idle calls at the tail.
2489 if (call->flags & RX_CALL_TQ_BUSY)
2490 queue_Prepend(&rx_freeCallQueue, call);
2492 queue_Append(&rx_freeCallQueue, call);
2493 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2494 queue_Append(&rx_freeCallQueue, call);
2495 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2496 if (rx_stats_active)
2497 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2498 MUTEX_EXIT(&rx_freeCallQueue_lock);
2500 /* Destroy the connection if it was previously slated for
2501 * destruction, i.e. the Rx client code previously called
2502 * rx_DestroyConnection (client connections), or
2503 * rxi_ReapConnections called the same routine (server
2504 * connections). Only do this, however, if there are no
2505 * outstanding calls. Note that for fine grain locking, there appears
2506 * to be a deadlock in that rxi_FreeCall has a call locked and
2507 * DestroyConnectionNoLock locks each call in the conn. But note a
2508 * few lines up where we have removed this call from the conn.
2509 * If someone else destroys a connection, they either have no
2510 * call lock held or are going through this section of code.
2512 MUTEX_ENTER(&conn->conn_data_lock);
2513 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2514 MUTEX_ENTER(&rx_refcnt_mutex);
2516 MUTEX_EXIT(&rx_refcnt_mutex);
2517 MUTEX_EXIT(&conn->conn_data_lock);
2518 #ifdef RX_ENABLE_LOCKS
2520 rxi_DestroyConnectionNoLock(conn);
2522 rxi_DestroyConnection(conn);
2523 #else /* RX_ENABLE_LOCKS */
2524 rxi_DestroyConnection(conn);
2525 #endif /* RX_ENABLE_LOCKS */
2527 MUTEX_EXIT(&conn->conn_data_lock);
2529 MUTEX_ENTER(&rx_refcnt_mutex);
2532 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2533 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2536 rxi_Alloc(size_t size)
2540 if (rx_stats_active) {
2541 rx_atomic_add(&rxi_Allocsize, (int) size);
2542 rx_atomic_inc(&rxi_Alloccnt);
2546 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2547 afs_osi_Alloc_NoSleep(size);
2552 osi_Panic("rxi_Alloc error");
2558 rxi_Free(void *addr, size_t size)
2560 if (rx_stats_active) {
2561 rx_atomic_sub(&rxi_Allocsize, (int) size);
2562 rx_atomic_dec(&rxi_Alloccnt);
2564 osi_Free(addr, size);
2568 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2570 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2571 struct rx_peer *next = NULL;
2575 MUTEX_ENTER(&rx_peerHashTable_lock);
2577 peer_ptr = &rx_peerHashTable[0];
2578 peer_end = &rx_peerHashTable[rx_hashTableSize];
2581 for ( ; peer_ptr < peer_end; peer_ptr++) {
2584 for ( ; peer; peer = next) {
2586 if (host == peer->host)
2591 hashIndex = PEER_HASH(host, port);
2592 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2593 if ((peer->host == host) && (peer->port == port))
2598 MUTEX_ENTER(&rx_peerHashTable_lock);
2603 MUTEX_EXIT(&rx_peerHashTable_lock);
2605 MUTEX_ENTER(&peer->peer_lock);
2606 /* We don't handle dropping below min, so don't */
2607 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2608 peer->ifMTU=MIN(mtu, peer->ifMTU);
2609 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2610 /* if we tweaked this down, need to tune our peer MTU too */
2611 peer->MTU = MIN(peer->MTU, peer->natMTU);
2612 /* if we discovered a sub-1500 mtu, degrade */
2613 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2614 peer->maxDgramPackets = 1;
2615 /* We no longer have valid peer packet information */
2616 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2617 peer->maxPacketSize = 0;
2618 MUTEX_EXIT(&peer->peer_lock);
2620 MUTEX_ENTER(&rx_peerHashTable_lock);
2622 if (host && !port) {
2624 /* pick up where we left off */
2628 MUTEX_EXIT(&rx_peerHashTable_lock);
2631 /* Find the peer process represented by the supplied (host,port)
2632 * combination. If there is no appropriate active peer structure, a
2633 * new one will be allocated and initialized
2634 * The origPeer, if set, is a pointer to a peer structure on which the
2635 * refcount will be be decremented. This is used to replace the peer
2636 * structure hanging off a connection structure */
2638 rxi_FindPeer(afs_uint32 host, u_short port,
2639 struct rx_peer *origPeer, int create)
2643 hashIndex = PEER_HASH(host, port);
2644 MUTEX_ENTER(&rx_peerHashTable_lock);
2645 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2646 if ((pp->host == host) && (pp->port == port))
2651 pp = rxi_AllocPeer(); /* This bzero's *pp */
2652 pp->host = host; /* set here or in InitPeerParams is zero */
2654 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2655 queue_Init(&pp->congestionQueue);
2656 queue_Init(&pp->rpcStats);
2657 pp->next = rx_peerHashTable[hashIndex];
2658 rx_peerHashTable[hashIndex] = pp;
2659 rxi_InitPeerParams(pp);
2660 if (rx_stats_active)
2661 rx_atomic_inc(&rx_stats.nPeerStructs);
2668 origPeer->refCount--;
2669 MUTEX_EXIT(&rx_peerHashTable_lock);
2674 /* Find the connection at (host, port) started at epoch, and with the
2675 * given connection id. Creates the server connection if necessary.
2676 * The type specifies whether a client connection or a server
2677 * connection is desired. In both cases, (host, port) specify the
2678 * peer's (host, pair) pair. Client connections are not made
2679 * automatically by this routine. The parameter socket gives the
2680 * socket descriptor on which the packet was received. This is used,
2681 * in the case of server connections, to check that *new* connections
2682 * come via a valid (port, serviceId). Finally, the securityIndex
2683 * parameter must match the existing index for the connection. If a
2684 * server connection is created, it will be created using the supplied
2685 * index, if the index is valid for this service */
2686 struct rx_connection *
2687 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2688 u_short port, u_short serviceId, afs_uint32 cid,
2689 afs_uint32 epoch, int type, u_int securityIndex)
2691 int hashindex, flag, i;
2692 struct rx_connection *conn;
2693 hashindex = CONN_HASH(host, port, cid, epoch, type);
2694 MUTEX_ENTER(&rx_connHashTable_lock);
2695 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2696 rx_connHashTable[hashindex],
2699 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2700 && (epoch == conn->epoch)) {
2701 struct rx_peer *pp = conn->peer;
2702 if (securityIndex != conn->securityIndex) {
2703 /* this isn't supposed to happen, but someone could forge a packet
2704 * like this, and there seems to be some CM bug that makes this
2705 * happen from time to time -- in which case, the fileserver
2707 MUTEX_EXIT(&rx_connHashTable_lock);
2708 return (struct rx_connection *)0;
2710 if (pp->host == host && pp->port == port)
2712 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2714 /* So what happens when it's a callback connection? */
2715 if ( /*type == RX_CLIENT_CONNECTION && */
2716 (conn->epoch & 0x80000000))
2720 /* the connection rxLastConn that was used the last time is not the
2721 ** one we are looking for now. Hence, start searching in the hash */
2723 conn = rx_connHashTable[hashindex];
2728 struct rx_service *service;
2729 if (type == RX_CLIENT_CONNECTION) {
2730 MUTEX_EXIT(&rx_connHashTable_lock);
2731 return (struct rx_connection *)0;
2733 service = rxi_FindService(socket, serviceId);
2734 if (!service || (securityIndex >= service->nSecurityObjects)
2735 || (service->securityObjects[securityIndex] == 0)) {
2736 MUTEX_EXIT(&rx_connHashTable_lock);
2737 return (struct rx_connection *)0;
2739 conn = rxi_AllocConnection(); /* This bzero's the connection */
2740 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2741 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2742 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2743 conn->next = rx_connHashTable[hashindex];
2744 rx_connHashTable[hashindex] = conn;
2745 conn->peer = rxi_FindPeer(host, port, 0, 1);
2746 conn->type = RX_SERVER_CONNECTION;
2747 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2748 conn->epoch = epoch;
2749 conn->cid = cid & RX_CIDMASK;
2750 /* conn->serial = conn->lastSerial = 0; */
2751 /* conn->timeout = 0; */
2752 conn->ackRate = RX_FAST_ACK_RATE;
2753 conn->service = service;
2754 conn->serviceId = serviceId;
2755 conn->securityIndex = securityIndex;
2756 conn->securityObject = service->securityObjects[securityIndex];
2757 conn->nSpecific = 0;
2758 conn->specific = NULL;
2759 rx_SetConnDeadTime(conn, service->connDeadTime);
2760 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2761 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2762 for (i = 0; i < RX_MAXCALLS; i++) {
2763 conn->twind[i] = rx_initSendWindow;
2764 conn->rwind[i] = rx_initReceiveWindow;
2766 /* Notify security object of the new connection */
2767 RXS_NewConnection(conn->securityObject, conn);
2768 /* XXXX Connection timeout? */
2769 if (service->newConnProc)
2770 (*service->newConnProc) (conn);
2771 if (rx_stats_active)
2772 rx_atomic_inc(&rx_stats.nServerConns);
2775 MUTEX_ENTER(&rx_refcnt_mutex);
2777 MUTEX_EXIT(&rx_refcnt_mutex);
2779 rxLastConn = conn; /* store this connection as the last conn used */
2780 MUTEX_EXIT(&rx_connHashTable_lock);
2784 /* There are two packet tracing routines available for testing and monitoring
2785 * Rx. One is called just after every packet is received and the other is
2786 * called just before every packet is sent. Received packets, have had their
2787 * headers decoded, and packets to be sent have not yet had their headers
2788 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2789 * containing the network address. Both can be modified. The return value, if
2790 * non-zero, indicates that the packet should be dropped. */
2792 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2793 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2795 /* A packet has been received off the interface. Np is the packet, socket is
2796 * the socket number it was received from (useful in determining which service
2797 * this packet corresponds to), and (host, port) reflect the host,port of the
2798 * sender. This call returns the packet to the caller if it is finished with
2799 * it, rather than de-allocating it, just as a small performance hack */
2802 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
2803 afs_uint32 host, u_short port, int *tnop,
2804 struct rx_call **newcallp)
2806 struct rx_call *call;
2807 struct rx_connection *conn;
2809 afs_uint32 currentCallNumber;
2815 struct rx_packet *tnp;
2818 /* We don't print out the packet until now because (1) the time may not be
2819 * accurate enough until now in the lwp implementation (rx_Listener only gets
2820 * the time after the packet is read) and (2) from a protocol point of view,
2821 * this is the first time the packet has been seen */
2822 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2823 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2824 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
2825 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2826 np->header.epoch, np->header.cid, np->header.callNumber,
2827 np->header.seq, np->header.flags, np));
2830 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2831 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2834 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2835 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2838 /* If an input tracer function is defined, call it with the packet and
2839 * network address. Note this function may modify its arguments. */
2840 if (rx_justReceived) {
2841 struct sockaddr_in addr;
2843 addr.sin_family = AF_INET;
2844 addr.sin_port = port;
2845 addr.sin_addr.s_addr = host;
2846 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2847 addr.sin_len = sizeof(addr);
2848 #endif /* AFS_OSF_ENV */
2849 drop = (*rx_justReceived) (np, &addr);
2850 /* drop packet if return value is non-zero */
2853 port = addr.sin_port; /* in case fcn changed addr */
2854 host = addr.sin_addr.s_addr;
2858 /* If packet was not sent by the client, then *we* must be the client */
2859 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2860 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2862 /* Find the connection (or fabricate one, if we're the server & if
2863 * necessary) associated with this packet */
2865 rxi_FindConnection(socket, host, port, np->header.serviceId,
2866 np->header.cid, np->header.epoch, type,
2867 np->header.securityIndex);
2870 /* If no connection found or fabricated, just ignore the packet.
2871 * (An argument could be made for sending an abort packet for
2876 MUTEX_ENTER(&conn->conn_data_lock);
2877 if (conn->maxSerial < np->header.serial)
2878 conn->maxSerial = np->header.serial;
2879 MUTEX_EXIT(&conn->conn_data_lock);
2881 /* If the connection is in an error state, send an abort packet and ignore
2882 * the incoming packet */
2884 /* Don't respond to an abort packet--we don't want loops! */
2885 MUTEX_ENTER(&conn->conn_data_lock);
2886 if (np->header.type != RX_PACKET_TYPE_ABORT)
2887 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2888 MUTEX_ENTER(&rx_refcnt_mutex);
2890 MUTEX_EXIT(&rx_refcnt_mutex);
2891 MUTEX_EXIT(&conn->conn_data_lock);
2895 /* Check for connection-only requests (i.e. not call specific). */
2896 if (np->header.callNumber == 0) {
2897 switch (np->header.type) {
2898 case RX_PACKET_TYPE_ABORT: {
2899 /* What if the supplied error is zero? */
2900 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2901 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
2902 rxi_ConnectionError(conn, errcode);
2903 MUTEX_ENTER(&rx_refcnt_mutex);
2905 MUTEX_EXIT(&rx_refcnt_mutex);
2908 case RX_PACKET_TYPE_CHALLENGE:
2909 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2910 MUTEX_ENTER(&rx_refcnt_mutex);
2912 MUTEX_EXIT(&rx_refcnt_mutex);
2914 case RX_PACKET_TYPE_RESPONSE:
2915 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2916 MUTEX_ENTER(&rx_refcnt_mutex);
2918 MUTEX_EXIT(&rx_refcnt_mutex);
2920 case RX_PACKET_TYPE_PARAMS:
2921 case RX_PACKET_TYPE_PARAMS + 1:
2922 case RX_PACKET_TYPE_PARAMS + 2:
2923 /* ignore these packet types for now */
2924 MUTEX_ENTER(&rx_refcnt_mutex);
2926 MUTEX_EXIT(&rx_refcnt_mutex);
2931 /* Should not reach here, unless the peer is broken: send an
2933 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2934 MUTEX_ENTER(&conn->conn_data_lock);
2935 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2936 MUTEX_ENTER(&rx_refcnt_mutex);
2938 MUTEX_EXIT(&rx_refcnt_mutex);
2939 MUTEX_EXIT(&conn->conn_data_lock);
2944 channel = np->header.cid & RX_CHANNELMASK;
2945 call = conn->call[channel];
2946 #ifdef RX_ENABLE_LOCKS
2948 MUTEX_ENTER(&call->lock);
2949 /* Test to see if call struct is still attached to conn. */
2950 if (call != conn->call[channel]) {
2952 MUTEX_EXIT(&call->lock);
2953 if (type == RX_SERVER_CONNECTION) {
2954 call = conn->call[channel];
2955 /* If we started with no call attached and there is one now,
2956 * another thread is also running this routine and has gotten
2957 * the connection channel. We should drop this packet in the tests
2958 * below. If there was a call on this connection and it's now
2959 * gone, then we'll be making a new call below.
2960 * If there was previously a call and it's now different then
2961 * the old call was freed and another thread running this routine
2962 * has created a call on this channel. One of these two threads
2963 * has a packet for the old call and the code below handles those
2967 MUTEX_ENTER(&call->lock);
2969 /* This packet can't be for this call. If the new call address is
2970 * 0 then no call is running on this channel. If there is a call
2971 * then, since this is a client connection we're getting data for
2972 * it must be for the previous call.
2974 if (rx_stats_active)
2975 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
2976 MUTEX_ENTER(&rx_refcnt_mutex);
2978 MUTEX_EXIT(&rx_refcnt_mutex);
2983 currentCallNumber = conn->callNumber[channel];
2985 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2986 if (np->header.callNumber < currentCallNumber) {
2987 if (rx_stats_active)
2988 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
2989 #ifdef RX_ENABLE_LOCKS
2991 MUTEX_EXIT(&call->lock);
2993 MUTEX_ENTER(&rx_refcnt_mutex);
2995 MUTEX_EXIT(&rx_refcnt_mutex);
2999 MUTEX_ENTER(&conn->conn_call_lock);
3000 call = rxi_NewCall(conn, channel);
3001 MUTEX_EXIT(&conn->conn_call_lock);
3002 *call->callNumber = np->header.callNumber;
3004 if (np->header.callNumber == 0)
3005 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",
3006 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3007 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3008 np->header.flags, np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
3010 call->state = RX_STATE_PRECALL;
3011 clock_GetTime(&call->queueTime);
3012 hzero(call->bytesSent);
3013 hzero(call->bytesRcvd);
3015 * If the number of queued calls exceeds the overload
3016 * threshold then abort this call.
3018 if ((rx_BusyThreshold > 0) &&
3019 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3020 struct rx_packet *tp;
3022 rxi_CallError(call, rx_BusyError);
3023 tp = rxi_SendCallAbort(call, np, 1, 0);
3024 MUTEX_EXIT(&call->lock);
3025 MUTEX_ENTER(&rx_refcnt_mutex);
3027 MUTEX_EXIT(&rx_refcnt_mutex);
3028 if (rx_stats_active)
3029 rx_atomic_inc(&rx_stats.nBusies);
3032 rxi_KeepAliveOn(call);
3033 } else if (np->header.callNumber != currentCallNumber) {
3034 /* Wait until the transmit queue is idle before deciding
3035 * whether to reset the current call. Chances are that the
3036 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3039 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3040 if (call->state == RX_STATE_ACTIVE) {
3041 rxi_WaitforTQBusy(call);
3043 * If we entered error state while waiting,
3044 * must call rxi_CallError to permit rxi_ResetCall
3045 * to processed when the tqWaiter count hits zero.
3048 rxi_CallError(call, call->error);
3049 MUTEX_EXIT(&call->lock);
3050 MUTEX_ENTER(&rx_refcnt_mutex);
3052 MUTEX_EXIT(&rx_refcnt_mutex);
3056 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3057 /* If the new call cannot be taken right now send a busy and set
3058 * the error condition in this call, so that it terminates as
3059 * quickly as possible */
3060 if (call->state == RX_STATE_ACTIVE) {
3061 struct rx_packet *tp;
3063 rxi_CallError(call, RX_CALL_DEAD);
3064 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3066 MUTEX_EXIT(&call->lock);
3067 MUTEX_ENTER(&rx_refcnt_mutex);
3069 MUTEX_EXIT(&rx_refcnt_mutex);
3072 rxi_ResetCall(call, 0);
3073 *call->callNumber = np->header.callNumber;
3075 if (np->header.callNumber == 0)
3076 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",
3077 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3078 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3079 np->header.flags, np, np->retryTime.sec, np->retryTime.usec, np->length));
3081 call->state = RX_STATE_PRECALL;
3082 clock_GetTime(&call->queueTime);
3083 hzero(call->bytesSent);
3084 hzero(call->bytesRcvd);
3086 * If the number of queued calls exceeds the overload
3087 * threshold then abort this call.
3089 if ((rx_BusyThreshold > 0) &&
3090 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3091 struct rx_packet *tp;
3093 rxi_CallError(call, rx_BusyError);
3094 tp = rxi_SendCallAbort(call, np, 1, 0);
3095 MUTEX_EXIT(&call->lock);
3096 MUTEX_ENTER(&rx_refcnt_mutex);
3098 MUTEX_EXIT(&rx_refcnt_mutex);
3099 if (rx_stats_active)
3100 rx_atomic_inc(&rx_stats.nBusies);
3103 rxi_KeepAliveOn(call);
3105 /* Continuing call; do nothing here. */
3107 } else { /* we're the client */
3108 /* Ignore all incoming acknowledgements for calls in DALLY state */
3109 if (call && (call->state == RX_STATE_DALLY)
3110 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3111 if (rx_stats_active)
3112 rx_atomic_inc(&rx_stats.ignorePacketDally);
3113 #ifdef RX_ENABLE_LOCKS
3115 MUTEX_EXIT(&call->lock);
3118 MUTEX_ENTER(&rx_refcnt_mutex);
3120 MUTEX_EXIT(&rx_refcnt_mutex);
3124 /* Ignore anything that's not relevant to the current call. If there
3125 * isn't a current call, then no packet is relevant. */
3126 if (!call || (np->header.callNumber != currentCallNumber)) {
3127 if (rx_stats_active)
3128 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3129 #ifdef RX_ENABLE_LOCKS
3131 MUTEX_EXIT(&call->lock);
3134 MUTEX_ENTER(&rx_refcnt_mutex);
3136 MUTEX_EXIT(&rx_refcnt_mutex);
3139 /* If the service security object index stamped in the packet does not
3140 * match the connection's security index, ignore the packet */
3141 if (np->header.securityIndex != conn->securityIndex) {
3142 #ifdef RX_ENABLE_LOCKS
3143 MUTEX_EXIT(&call->lock);
3145 MUTEX_ENTER(&rx_refcnt_mutex);
3147 MUTEX_EXIT(&rx_refcnt_mutex);
3151 /* If we're receiving the response, then all transmit packets are
3152 * implicitly acknowledged. Get rid of them. */
3153 if (np->header.type == RX_PACKET_TYPE_DATA) {
3154 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3155 /* XXX Hack. Because we must release the global rx lock when
3156 * sending packets (osi_NetSend) we drop all acks while we're
3157 * traversing the tq in rxi_Start sending packets out because
3158 * packets may move to the freePacketQueue as result of being here!
3159 * So we drop these packets until we're safely out of the
3160 * traversing. Really ugly!
3161 * For fine grain RX locking, we set the acked field in the
3162 * packets and let rxi_Start remove them from the transmit queue.
3164 if (call->flags & RX_CALL_TQ_BUSY) {
3165 #ifdef RX_ENABLE_LOCKS
3166 rxi_SetAcksInTransmitQueue(call);
3168 MUTEX_ENTER(&rx_refcnt_mutex);
3170 MUTEX_EXIT(&rx_refcnt_mutex);
3171 return np; /* xmitting; drop packet */
3174 rxi_ClearTransmitQueue(call, 0);
3176 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3177 rxi_ClearTransmitQueue(call, 0);
3178 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3180 if (np->header.type == RX_PACKET_TYPE_ACK) {
3181 /* now check to see if this is an ack packet acknowledging that the
3182 * server actually *lost* some hard-acked data. If this happens we
3183 * ignore this packet, as it may indicate that the server restarted in
3184 * the middle of a call. It is also possible that this is an old ack
3185 * packet. We don't abort the connection in this case, because this
3186 * *might* just be an old ack packet. The right way to detect a server
3187 * restart in the midst of a call is to notice that the server epoch
3189 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3190 * XXX unacknowledged. I think that this is off-by-one, but
3191 * XXX I don't dare change it just yet, since it will
3192 * XXX interact badly with the server-restart detection
3193 * XXX code in receiveackpacket. */
3194 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3195 if (rx_stats_active)
3196 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3197 MUTEX_EXIT(&call->lock);
3198 MUTEX_ENTER(&rx_refcnt_mutex);
3200 MUTEX_EXIT(&rx_refcnt_mutex);
3204 } /* else not a data packet */
3207 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3208 /* Set remote user defined status from packet */
3209 call->remoteStatus = np->header.userStatus;
3211 /* Note the gap between the expected next packet and the actual
3212 * packet that arrived, when the new packet has a smaller serial number
3213 * than expected. Rioses frequently reorder packets all by themselves,
3214 * so this will be quite important with very large window sizes.
3215 * Skew is checked against 0 here to avoid any dependence on the type of
3216 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3218 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3219 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3220 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3222 MUTEX_ENTER(&conn->conn_data_lock);
3223 skew = conn->lastSerial - np->header.serial;
3224 conn->lastSerial = np->header.serial;
3225 MUTEX_EXIT(&conn->conn_data_lock);
3227 struct rx_peer *peer;
3229 if (skew > peer->inPacketSkew) {
3230 dpf(("*** In skew changed from %d to %d\n",
3231 peer->inPacketSkew, skew));
3232 peer->inPacketSkew = skew;
3236 /* Now do packet type-specific processing */
3237 switch (np->header.type) {
3238 case RX_PACKET_TYPE_DATA:
3239 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3242 case RX_PACKET_TYPE_ACK:
3243 /* Respond immediately to ack packets requesting acknowledgement
3245 if (np->header.flags & RX_REQUEST_ACK) {
3247 (void)rxi_SendCallAbort(call, 0, 1, 0);
3249 (void)rxi_SendAck(call, 0, np->header.serial,
3250 RX_ACK_PING_RESPONSE, 1);
3252 np = rxi_ReceiveAckPacket(call, np, 1);
3254 case RX_PACKET_TYPE_ABORT: {
3255 /* An abort packet: reset the call, passing the error up to the user. */
3256 /* What if error is zero? */
3257 /* What if the error is -1? the application will treat it as a timeout. */
3258 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3259 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3260 rxi_CallError(call, errdata);
3261 MUTEX_EXIT(&call->lock);
3262 MUTEX_ENTER(&rx_refcnt_mutex);
3264 MUTEX_EXIT(&rx_refcnt_mutex);
3265 return np; /* xmitting; drop packet */
3267 case RX_PACKET_TYPE_BUSY:
3270 case RX_PACKET_TYPE_ACKALL:
3271 /* All packets acknowledged, so we can drop all packets previously
3272 * readied for sending */
3273 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3274 /* XXX Hack. We because we can't release the global rx lock when
3275 * sending packets (osi_NetSend) we drop all ack pkts while we're
3276 * traversing the tq in rxi_Start sending packets out because
3277 * packets may move to the freePacketQueue as result of being
3278 * here! So we drop these packets until we're safely out of the
3279 * traversing. Really ugly!
3280 * For fine grain RX locking, we set the acked field in the packets
3281 * and let rxi_Start remove the packets from the transmit queue.
3283 if (call->flags & RX_CALL_TQ_BUSY) {
3284 #ifdef RX_ENABLE_LOCKS
3285 rxi_SetAcksInTransmitQueue(call);
3287 #else /* RX_ENABLE_LOCKS */
3288 MUTEX_EXIT(&call->lock);
3289 MUTEX_ENTER(&rx_refcnt_mutex);
3291 MUTEX_EXIT(&rx_refcnt_mutex);
3292 return np; /* xmitting; drop packet */
3293 #endif /* RX_ENABLE_LOCKS */
3295 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3296 rxi_ClearTransmitQueue(call, 0);
3297 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3300 /* Should not reach here, unless the peer is broken: send an abort
3302 rxi_CallError(call, RX_PROTOCOL_ERROR);
3303 np = rxi_SendCallAbort(call, np, 1, 0);
3306 /* Note when this last legitimate packet was received, for keep-alive
3307 * processing. Note, we delay getting the time until now in the hope that
3308 * the packet will be delivered to the user before any get time is required
3309 * (if not, then the time won't actually be re-evaluated here). */
3310 call->lastReceiveTime = clock_Sec();
3311 MUTEX_EXIT(&call->lock);
3312 MUTEX_ENTER(&rx_refcnt_mutex);
3314 MUTEX_EXIT(&rx_refcnt_mutex);
3318 /* return true if this is an "interesting" connection from the point of view
3319 of someone trying to debug the system */
3321 rxi_IsConnInteresting(struct rx_connection *aconn)
3324 struct rx_call *tcall;
3326 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3329 for (i = 0; i < RX_MAXCALLS; i++) {
3330 tcall = aconn->call[i];
3332 if ((tcall->state == RX_STATE_PRECALL)
3333 || (tcall->state == RX_STATE_ACTIVE))
3335 if ((tcall->mode == RX_MODE_SENDING)
3336 || (tcall->mode == RX_MODE_RECEIVING))
3344 /* if this is one of the last few packets AND it wouldn't be used by the
3345 receiving call to immediately satisfy a read request, then drop it on
3346 the floor, since accepting it might prevent a lock-holding thread from
3347 making progress in its reading. If a call has been cleared while in
3348 the precall state then ignore all subsequent packets until the call
3349 is assigned to a thread. */
3352 TooLow(struct rx_packet *ap, struct rx_call *acall)
3356 MUTEX_ENTER(&rx_quota_mutex);
3357 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3358 && (acall->state == RX_STATE_PRECALL))
3359 || ((rx_nFreePackets < rxi_dataQuota + 2)
3360 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3361 && (acall->flags & RX_CALL_READER_WAIT)))) {
3364 MUTEX_EXIT(&rx_quota_mutex);
3370 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3372 struct rx_connection *conn = arg1;
3373 struct rx_call *acall = arg2;
3374 struct rx_call *call = acall;
3375 struct clock when, now;
3378 MUTEX_ENTER(&conn->conn_data_lock);
3379 conn->checkReachEvent = NULL;
3380 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3382 MUTEX_ENTER(&rx_refcnt_mutex);
3384 MUTEX_EXIT(&rx_refcnt_mutex);
3386 MUTEX_EXIT(&conn->conn_data_lock);
3390 MUTEX_ENTER(&conn->conn_call_lock);
3391 MUTEX_ENTER(&conn->conn_data_lock);
3392 for (i = 0; i < RX_MAXCALLS; i++) {
3393 struct rx_call *tc = conn->call[i];
3394 if (tc && tc->state == RX_STATE_PRECALL) {
3400 /* Indicate that rxi_CheckReachEvent is no longer running by
3401 * clearing the flag. Must be atomic under conn_data_lock to
3402 * avoid a new call slipping by: rxi_CheckConnReach holds
3403 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3405 conn->flags &= ~RX_CONN_ATTACHWAIT;
3406 MUTEX_EXIT(&conn->conn_data_lock);
3407 MUTEX_EXIT(&conn->conn_call_lock);
3412 MUTEX_ENTER(&call->lock);
3413 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3415 MUTEX_EXIT(&call->lock);
3417 clock_GetTime(&now);
3419 when.sec += RX_CHECKREACH_TIMEOUT;
3420 MUTEX_ENTER(&conn->conn_data_lock);
3421 if (!conn->checkReachEvent) {
3422 MUTEX_ENTER(&rx_refcnt_mutex);
3424 MUTEX_EXIT(&rx_refcnt_mutex);
3425 conn->checkReachEvent =
3426 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3429 MUTEX_EXIT(&conn->conn_data_lock);
3435 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3437 struct rx_service *service = conn->service;
3438 struct rx_peer *peer = conn->peer;
3439 afs_uint32 now, lastReach;
3441 if (service->checkReach == 0)
3445 MUTEX_ENTER(&peer->peer_lock);
3446 lastReach = peer->lastReachTime;
3447 MUTEX_EXIT(&peer->peer_lock);
3448 if (now - lastReach < RX_CHECKREACH_TTL)
3451 MUTEX_ENTER(&conn->conn_data_lock);
3452 if (conn->flags & RX_CONN_ATTACHWAIT) {
3453 MUTEX_EXIT(&conn->conn_data_lock);
3456 conn->flags |= RX_CONN_ATTACHWAIT;
3457 MUTEX_EXIT(&conn->conn_data_lock);
3458 if (!conn->checkReachEvent)
3459 rxi_CheckReachEvent(NULL, conn, call);
3464 /* try to attach call, if authentication is complete */
3466 TryAttach(struct rx_call *acall, osi_socket socket,
3467 int *tnop, struct rx_call **newcallp,
3470 struct rx_connection *conn = acall->conn;
3472 if (conn->type == RX_SERVER_CONNECTION
3473 && acall->state == RX_STATE_PRECALL) {
3474 /* Don't attach until we have any req'd. authentication. */
3475 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3476 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3477 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3478 /* Note: this does not necessarily succeed; there
3479 * may not any proc available
3482 rxi_ChallengeOn(acall->conn);
3487 /* A data packet has been received off the interface. This packet is
3488 * appropriate to the call (the call is in the right state, etc.). This
3489 * routine can return a packet to the caller, for re-use */
3492 rxi_ReceiveDataPacket(struct rx_call *call,
3493 struct rx_packet *np, int istack,
3494 osi_socket socket, afs_uint32 host, u_short port,
3495 int *tnop, struct rx_call **newcallp)
3497 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3502 afs_uint32 serial=0, flags=0;
3504 struct rx_packet *tnp;
3505 struct clock when, now;
3506 if (rx_stats_active)
3507 rx_atomic_inc(&rx_stats.dataPacketsRead);
3510 /* If there are no packet buffers, drop this new packet, unless we can find
3511 * packet buffers from inactive calls */
3513 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3514 MUTEX_ENTER(&rx_freePktQ_lock);
3515 rxi_NeedMorePackets = TRUE;
3516 MUTEX_EXIT(&rx_freePktQ_lock);
3517 if (rx_stats_active)
3518 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3519 call->rprev = np->header.serial;
3520 rxi_calltrace(RX_TRACE_DROP, call);
3521 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3523 rxi_ClearReceiveQueue(call);
3524 clock_GetTime(&now);
3526 clock_Add(&when, &rx_softAckDelay);
3527 if (!call->delayedAckEvent
3528 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3529 rxevent_Cancel(call->delayedAckEvent, call,
3530 RX_CALL_REFCOUNT_DELAY);
3531 MUTEX_ENTER(&rx_refcnt_mutex);
3532 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3533 MUTEX_EXIT(&rx_refcnt_mutex);
3535 call->delayedAckEvent =
3536 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3538 /* we've damaged this call already, might as well do it in. */
3544 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3545 * packet is one of several packets transmitted as a single
3546 * datagram. Do not send any soft or hard acks until all packets
3547 * in a jumbogram have been processed. Send negative acks right away.
3549 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3550 /* tnp is non-null when there are more packets in the
3551 * current jumbo gram */
3558 seq = np->header.seq;
3559 serial = np->header.serial;
3560 flags = np->header.flags;
3562 /* If the call is in an error state, send an abort message */
3564 return rxi_SendCallAbort(call, np, istack, 0);
3566 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3567 * AFS 3.5 jumbogram. */
3568 if (flags & RX_JUMBO_PACKET) {
3569 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3574 if (np->header.spare != 0) {
3575 MUTEX_ENTER(&call->conn->conn_data_lock);
3576 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3577 MUTEX_EXIT(&call->conn->conn_data_lock);
3580 /* The usual case is that this is the expected next packet */
3581 if (seq == call->rnext) {
3583 /* Check to make sure it is not a duplicate of one already queued */
3584 if (queue_IsNotEmpty(&call->rq)
3585 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3586 if (rx_stats_active)
3587 rx_atomic_inc(&rx_stats.dupPacketsRead);
3588 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3589 rxevent_Cancel(call->delayedAckEvent, call,
3590 RX_CALL_REFCOUNT_DELAY);
3591 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3597 /* It's the next packet. Stick it on the receive queue
3598 * for this call. Set newPackets to make sure we wake
3599 * the reader once all packets have been processed */
3600 #ifdef RX_TRACK_PACKETS
3601 np->flags |= RX_PKTFLAG_RQ;
3603 queue_Prepend(&call->rq, np);
3604 #ifdef RXDEBUG_PACKET
3606 #endif /* RXDEBUG_PACKET */
3608 np = NULL; /* We can't use this anymore */
3611 /* If an ack is requested then set a flag to make sure we
3612 * send an acknowledgement for this packet */
3613 if (flags & RX_REQUEST_ACK) {
3614 ackNeeded = RX_ACK_REQUESTED;
3617 /* Keep track of whether we have received the last packet */
3618 if (flags & RX_LAST_PACKET) {
3619 call->flags |= RX_CALL_HAVE_LAST;
3623 /* Check whether we have all of the packets for this call */
3624 if (call->flags & RX_CALL_HAVE_LAST) {
3625 afs_uint32 tseq; /* temporary sequence number */
3626 struct rx_packet *tp; /* Temporary packet pointer */
3627 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3629 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3630 if (tseq != tp->header.seq)
3632 if (tp->header.flags & RX_LAST_PACKET) {
3633 call->flags |= RX_CALL_RECEIVE_DONE;
3640 /* Provide asynchronous notification for those who want it
3641 * (e.g. multi rx) */
3642 if (call->arrivalProc) {
3643 (*call->arrivalProc) (call, call->arrivalProcHandle,
3644 call->arrivalProcArg);
3645 call->arrivalProc = (void (*)())0;
3648 /* Update last packet received */
3651 /* If there is no server process serving this call, grab
3652 * one, if available. We only need to do this once. If a
3653 * server thread is available, this thread becomes a server
3654 * thread and the server thread becomes a listener thread. */
3656 TryAttach(call, socket, tnop, newcallp, 0);
3659 /* This is not the expected next packet. */
3661 /* Determine whether this is a new or old packet, and if it's
3662 * a new one, whether it fits into the current receive window.
3663 * Also figure out whether the packet was delivered in sequence.
3664 * We use the prev variable to determine whether the new packet
3665 * is the successor of its immediate predecessor in the
3666 * receive queue, and the missing flag to determine whether
3667 * any of this packets predecessors are missing. */
3669 afs_uint32 prev; /* "Previous packet" sequence number */
3670 struct rx_packet *tp; /* Temporary packet pointer */
3671 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3672 int missing; /* Are any predecessors missing? */
3674 /* If the new packet's sequence number has been sent to the
3675 * application already, then this is a duplicate */
3676 if (seq < call->rnext) {
3677 if (rx_stats_active)
3678 rx_atomic_inc(&rx_stats.dupPacketsRead);
3679 rxevent_Cancel(call->delayedAckEvent, call,
3680 RX_CALL_REFCOUNT_DELAY);
3681 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3687 /* If the sequence number is greater than what can be
3688 * accomodated by the current window, then send a negative
3689 * acknowledge and drop the packet */
3690 if ((call->rnext + call->rwind) <= seq) {
3691 rxevent_Cancel(call->delayedAckEvent, call,
3692 RX_CALL_REFCOUNT_DELAY);
3693 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3700 /* Look for the packet in the queue of old received packets */
3701 for (prev = call->rnext - 1, missing =
3702 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3703 /*Check for duplicate packet */
3704 if (seq == tp->header.seq) {
3705 if (rx_stats_active)
3706 rx_atomic_inc(&rx_stats.dupPacketsRead);
3707 rxevent_Cancel(call->delayedAckEvent, call,
3708 RX_CALL_REFCOUNT_DELAY);
3709 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3715 /* If we find a higher sequence packet, break out and
3716 * insert the new packet here. */
3717 if (seq < tp->header.seq)
3719 /* Check for missing packet */
3720 if (tp->header.seq != prev + 1) {
3724 prev = tp->header.seq;
3727 /* Keep track of whether we have received the last packet. */
3728 if (flags & RX_LAST_PACKET) {
3729 call->flags |= RX_CALL_HAVE_LAST;
3732 /* It's within the window: add it to the the receive queue.
3733 * tp is left by the previous loop either pointing at the
3734 * packet before which to insert the new packet, or at the
3735 * queue head if the queue is empty or the packet should be
3737 #ifdef RX_TRACK_PACKETS
3738 np->flags |= RX_PKTFLAG_RQ;
3740 #ifdef RXDEBUG_PACKET
3742 #endif /* RXDEBUG_PACKET */
3743 queue_InsertBefore(tp, np);
3747 /* Check whether we have all of the packets for this call */
3748 if ((call->flags & RX_CALL_HAVE_LAST)
3749 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3750 afs_uint32 tseq; /* temporary sequence number */
3753 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3754 if (tseq != tp->header.seq)
3756 if (tp->header.flags & RX_LAST_PACKET) {
3757 call->flags |= RX_CALL_RECEIVE_DONE;
3764 /* We need to send an ack of the packet is out of sequence,
3765 * or if an ack was requested by the peer. */
3766 if (seq != prev + 1 || missing) {
3767 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3768 } else if (flags & RX_REQUEST_ACK) {
3769 ackNeeded = RX_ACK_REQUESTED;
3772 /* Acknowledge the last packet for each call */
3773 if (flags & RX_LAST_PACKET) {
3784 * If the receiver is waiting for an iovec, fill the iovec
3785 * using the data from the receive queue */
3786 if (call->flags & RX_CALL_IOVEC_WAIT) {
3787 didHardAck = rxi_FillReadVec(call, serial);
3788 /* the call may have been aborted */
3797 /* Wakeup the reader if any */
3798 if ((call->flags & RX_CALL_READER_WAIT)
3799 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3800 || (call->iovNext >= call->iovMax)
3801 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3802 call->flags &= ~RX_CALL_READER_WAIT;
3803 #ifdef RX_ENABLE_LOCKS
3804 CV_BROADCAST(&call->cv_rq);
3806 osi_rxWakeup(&call->rq);
3812 * Send an ack when requested by the peer, or once every
3813 * rxi_SoftAckRate packets until the last packet has been
3814 * received. Always send a soft ack for the last packet in
3815 * the server's reply.
3817 * If we have received all of the packets for the call
3818 * immediately send an RX_PACKET_TYPE_ACKALL packet so that
3819 * the peer can empty its packet queue and cancel all resend
3822 if (call->flags & RX_CALL_RECEIVE_DONE) {
3823 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3824 rxi_AckAll(NULL, call, 0);
3825 } else if (ackNeeded) {
3826 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3827 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3828 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3829 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3830 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3831 } else if (call->nSoftAcks) {
3832 clock_GetTime(&now);
3834 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3835 clock_Add(&when, &rx_lastAckDelay);
3837 clock_Add(&when, &rx_softAckDelay);
3839 if (!call->delayedAckEvent
3840 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3841 rxevent_Cancel(call->delayedAckEvent, call,
3842 RX_CALL_REFCOUNT_DELAY);
3843 MUTEX_ENTER(&rx_refcnt_mutex);
3844 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3845 MUTEX_EXIT(&rx_refcnt_mutex);
3846 call->delayedAckEvent =
3847 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3855 static void rxi_ComputeRate();
3859 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3861 struct rx_peer *peer = conn->peer;
3863 MUTEX_ENTER(&peer->peer_lock);
3864 peer->lastReachTime = clock_Sec();
3865 MUTEX_EXIT(&peer->peer_lock);
3867 MUTEX_ENTER(&conn->conn_data_lock);
3868 if (conn->flags & RX_CONN_ATTACHWAIT) {
3871 conn->flags &= ~RX_CONN_ATTACHWAIT;
3872 MUTEX_EXIT(&conn->conn_data_lock);
3874 for (i = 0; i < RX_MAXCALLS; i++) {
3875 struct rx_call *call = conn->call[i];
3878 MUTEX_ENTER(&call->lock);
3879 /* tnop can be null if newcallp is null */
3880 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3882 MUTEX_EXIT(&call->lock);
3886 MUTEX_EXIT(&conn->conn_data_lock);
3889 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
3891 rx_ack_reason(int reason)
3894 case RX_ACK_REQUESTED:
3896 case RX_ACK_DUPLICATE:
3898 case RX_ACK_OUT_OF_SEQUENCE:
3900 case RX_ACK_EXCEEDS_WINDOW:
3902 case RX_ACK_NOSPACE:
3906 case RX_ACK_PING_RESPONSE:
3919 /* The real smarts of the whole thing. */
3921 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
3924 struct rx_ackPacket *ap;
3926 struct rx_packet *tp;
3927 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3928 struct rx_connection *conn = call->conn;
3929 struct rx_peer *peer = conn->peer;
3930 struct clock now; /* Current time, for RTT calculations */
3934 /* because there are CM's that are bogus, sending weird values for this. */
3935 afs_uint32 skew = 0;
3940 int newAckCount = 0;
3941 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3942 int pktsize = 0; /* Set if we need to update the peer mtu */
3943 int conn_data_locked = 0;
3945 if (rx_stats_active)
3946 rx_atomic_inc(&rx_stats.ackPacketsRead);
3947 ap = (struct rx_ackPacket *)rx_DataOf(np);
3948 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3950 return np; /* truncated ack packet */
3952 /* depends on ack packet struct */
3953 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3954 first = ntohl(ap->firstPacket);
3955 prev = ntohl(ap->previousPacket);
3956 serial = ntohl(ap->serial);
3957 /* temporarily disabled -- needs to degrade over time
3958 * skew = ntohs(ap->maxSkew); */
3960 /* Ignore ack packets received out of order */
3961 if (first < call->tfirst ||
3962 (first == call->tfirst && prev < call->tprev)) {
3968 if (np->header.flags & RX_SLOW_START_OK) {
3969 call->flags |= RX_CALL_SLOW_START_OK;
3972 if (ap->reason == RX_ACK_PING_RESPONSE)
3973 rxi_UpdatePeerReach(conn, call);
3975 if (conn->lastPacketSizeSeq) {
3976 MUTEX_ENTER(&conn->conn_data_lock);
3977 conn_data_locked = 1;
3978 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
3979 pktsize = conn->lastPacketSize;
3980 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
3983 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
3984 if (!conn_data_locked) {
3985 MUTEX_ENTER(&conn->conn_data_lock);
3986 conn_data_locked = 1;
3988 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
3989 /* process mtu ping ack */
3990 pktsize = conn->lastPingSize;
3991 conn->lastPingSizeSer = conn->lastPingSize = 0;
3995 if (conn_data_locked) {
3996 MUTEX_EXIT(&conn->conn_data_lock);
3997 conn_data_locked = 0;
4001 if (rxdebug_active) {
4005 len = _snprintf(msg, sizeof(msg),
4006 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
4007 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4008 ntohl(ap->serial), ntohl(ap->previousPacket),
4009 (unsigned int)np->header.seq, (unsigned int)skew,
4010 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
4014 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4015 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4019 OutputDebugString(msg);
4021 #else /* AFS_NT40_ENV */
4024 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
4025 ap->reason, ntohl(ap->previousPacket),
4026 (unsigned int)np->header.seq, (unsigned int)serial,
4027 (unsigned int)skew, ntohl(ap->firstPacket));
4030 for (offset = 0; offset < nAcks; offset++)
4031 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4036 #endif /* AFS_NT40_ENV */
4039 MUTEX_ENTER(&peer->peer_lock);
4042 * Start somewhere. Can't assume we can send what we can receive,
4043 * but we are clearly receiving.
4045 if (!peer->maxPacketSize)
4046 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4048 if (pktsize > peer->maxPacketSize) {
4049 peer->maxPacketSize = pktsize;
4050 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4051 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4052 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4053 rxi_ScheduleGrowMTUEvent(call, 1);
4058 /* Update the outgoing packet skew value to the latest value of
4059 * the peer's incoming packet skew value. The ack packet, of
4060 * course, could arrive out of order, but that won't affect things
4062 peer->outPacketSkew = skew;
4064 /* Check for packets that no longer need to be transmitted, and
4065 * discard them. This only applies to packets positively
4066 * acknowledged as having been sent to the peer's upper level.
4067 * All other packets must be retained. So only packets with
4068 * sequence numbers < ap->firstPacket are candidates. */
4070 clock_GetTime(&now);
4072 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
4073 if (tp->header.seq >= first)
4075 call->tfirst = tp->header.seq + 1;
4077 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4080 rxi_ComputeRoundTripTime(tp, ap, call->conn->peer, &now);
4084 rxi_ComputeRate(call->conn->peer, call, p, np, ap->reason);
4087 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4088 /* XXX Hack. Because we have to release the global rx lock when sending
4089 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4090 * in rxi_Start sending packets out because packets may move to the
4091 * freePacketQueue as result of being here! So we drop these packets until
4092 * we're safely out of the traversing. Really ugly!
4093 * To make it even uglier, if we're using fine grain locking, we can
4094 * set the ack bits in the packets and have rxi_Start remove the packets
4095 * when it's done transmitting.
4097 if (call->flags & RX_CALL_TQ_BUSY) {
4098 #ifdef RX_ENABLE_LOCKS
4099 tp->flags |= RX_PKTFLAG_ACKED;
4100 call->flags |= RX_CALL_TQ_SOME_ACKED;
4101 #else /* RX_ENABLE_LOCKS */
4103 #endif /* RX_ENABLE_LOCKS */
4105 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4108 #ifdef RX_TRACK_PACKETS
4109 tp->flags &= ~RX_PKTFLAG_TQ;
4111 #ifdef RXDEBUG_PACKET
4113 #endif /* RXDEBUG_PACKET */
4114 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4119 /* Give rate detector a chance to respond to ping requests */
4120 if (ap->reason == RX_ACK_PING_RESPONSE) {
4121 rxi_ComputeRate(peer, call, 0, np, ap->reason);
4125 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4127 /* Now go through explicit acks/nacks and record the results in
4128 * the waiting packets. These are packets that can't be released
4129 * yet, even with a positive acknowledge. This positive
4130 * acknowledge only means the packet has been received by the
4131 * peer, not that it will be retained long enough to be sent to
4132 * the peer's upper level. In addition, reset the transmit timers
4133 * of any missing packets (those packets that must be missing
4134 * because this packet was out of sequence) */
4136 call->nSoftAcked = 0;
4137 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
4139 /* Set the acknowledge flag per packet based on the
4140 * information in the ack packet. An acknowlegded packet can
4141 * be downgraded when the server has discarded a packet it
4142 * soacked previously, or when an ack packet is received
4143 * out of sequence. */
4144 if (tp->header.seq < first) {
4145 /* Implicit ack information */
4146 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4149 tp->flags |= RX_PKTFLAG_ACKED;
4150 } else if (tp->header.seq < first + nAcks) {
4151 /* Explicit ack information: set it in the packet appropriately */
4152 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4153 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4155 tp->flags |= RX_PKTFLAG_ACKED;
4157 rxi_ComputeRoundTripTime(tp, ap, call->conn->peer, &now);
4159 rxi_ComputeRate(call->conn->peer, call, tp, np,
4168 } else /* RX_ACK_TYPE_NACK */ {
4169 tp->flags &= ~RX_PKTFLAG_ACKED;
4173 if (tp->flags & RX_PKTFLAG_ACKED) {
4174 tp->flags &= ~RX_PKTFLAG_ACKED;
4180 * Following the suggestion of Phil Kern, we back off the peer's
4181 * timeout value for future packets until a successful response
4182 * is received for an initial transmission.
4184 if (missing && !peer->backedOff) {
4185 struct clock c = peer->timeout;
4186 struct clock max_to = {3, 0};
4188 clock_Add(&peer->timeout, &c);
4189 if (clock_Gt(&peer->timeout, &max_to))
4190 peer->timeout = max_to;
4191 peer->backedOff = 1;
4194 /* If packet isn't yet acked, and it has been transmitted at least
4195 * once, reset retransmit time using latest timeout
4196 * ie, this should readjust the retransmit timer for all outstanding
4197 * packets... So we don't just retransmit when we should know better*/
4199 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
4200 tp->retryTime = tp->timeSent;
4201 clock_Add(&tp->retryTime, &peer->timeout);
4202 /* shift by eight because one quarter-sec ~ 256 milliseconds */
4203 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
4207 /* If the window has been extended by this acknowledge packet,
4208 * then wakeup a sender waiting in alloc for window space, or try
4209 * sending packets now, if he's been sitting on packets due to
4210 * lack of window space */
4211 if (call->tnext < (call->tfirst + call->twind)) {
4212 #ifdef RX_ENABLE_LOCKS
4213 CV_SIGNAL(&call->cv_twind);
4215 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4216 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4217 osi_rxWakeup(&call->twind);
4220 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4221 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4225 /* if the ack packet has a receivelen field hanging off it,
4226 * update our state */
4227 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4230 /* If the ack packet has a "recommended" size that is less than
4231 * what I am using now, reduce my size to match */
4232 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4233 (int)sizeof(afs_int32), &tSize);
4234 tSize = (afs_uint32) ntohl(tSize);
4235 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4237 /* Get the maximum packet size to send to this peer */
4238 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4240 tSize = (afs_uint32) ntohl(tSize);
4241 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4242 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4244 /* sanity check - peer might have restarted with different params.
4245 * If peer says "send less", dammit, send less... Peer should never
4246 * be unable to accept packets of the size that prior AFS versions would
4247 * send without asking. */
4248 if (peer->maxMTU != tSize) {
4249 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4251 peer->maxMTU = tSize;
4252 peer->MTU = MIN(tSize, peer->MTU);
4253 call->MTU = MIN(call->MTU, tSize);
4256 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4259 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4260 (int)sizeof(afs_int32), &tSize);
4261 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4262 if (tSize < call->twind) { /* smaller than our send */
4263 call->twind = tSize; /* window, we must send less... */
4264 call->ssthresh = MIN(call->twind, call->ssthresh);
4265 call->conn->twind[call->channel] = call->twind;
4268 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4269 * network MTU confused with the loopback MTU. Calculate the
4270 * maximum MTU here for use in the slow start code below.
4272 /* Did peer restart with older RX version? */
4273 if (peer->maxDgramPackets > 1) {
4274 peer->maxDgramPackets = 1;
4276 } else if (np->length >=
4277 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4280 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4281 sizeof(afs_int32), &tSize);
4282 tSize = (afs_uint32) ntohl(tSize);
4284 * As of AFS 3.5 we set the send window to match the receive window.
4286 if (tSize < call->twind) {
4287 call->twind = tSize;
4288 call->conn->twind[call->channel] = call->twind;
4289 call->ssthresh = MIN(call->twind, call->ssthresh);
4290 } else if (tSize > call->twind) {
4291 call->twind = tSize;
4292 call->conn->twind[call->channel] = call->twind;
4296 * As of AFS 3.5, a jumbogram is more than one fixed size
4297 * packet transmitted in a single UDP datagram. If the remote
4298 * MTU is smaller than our local MTU then never send a datagram
4299 * larger than the natural MTU.
4302 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4303 (int)sizeof(afs_int32), &tSize);
4304 maxDgramPackets = (afs_uint32) ntohl(tSize);
4305 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4307 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4308 if (maxDgramPackets > 1) {
4309 peer->maxDgramPackets = maxDgramPackets;
4310 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4312 peer->maxDgramPackets = 1;
4313 call->MTU = peer->natMTU;
4315 } else if (peer->maxDgramPackets > 1) {
4316 /* Restarted with lower version of RX */
4317 peer->maxDgramPackets = 1;
4319 } else if (peer->maxDgramPackets > 1
4320 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4321 /* Restarted with lower version of RX */
4322 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4323 peer->natMTU = OLD_MAX_PACKET_SIZE;
4324 peer->MTU = OLD_MAX_PACKET_SIZE;
4325 peer->maxDgramPackets = 1;
4326 peer->nDgramPackets = 1;
4328 call->MTU = OLD_MAX_PACKET_SIZE;
4333 * Calculate how many datagrams were successfully received after
4334 * the first missing packet and adjust the negative ack counter
4339 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4340 if (call->nNacks < nNacked) {
4341 call->nNacks = nNacked;
4344 call->nAcks += newAckCount;
4348 if (call->flags & RX_CALL_FAST_RECOVER) {
4350 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4352 call->flags &= ~RX_CALL_FAST_RECOVER;
4353 call->cwind = call->nextCwind;
4354 call->nextCwind = 0;
4357 call->nCwindAcks = 0;
4358 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4359 /* Three negative acks in a row trigger congestion recovery */
4360 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4361 MUTEX_EXIT(&peer->peer_lock);
4362 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4363 /* someone else is waiting to start recovery */
4366 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4367 rxi_WaitforTQBusy(call);
4368 MUTEX_ENTER(&peer->peer_lock);
4369 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4370 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4371 call->flags |= RX_CALL_FAST_RECOVER;
4372 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4374 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4375 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4376 call->nextCwind = call->ssthresh;
4379 peer->MTU = call->MTU;
4380 peer->cwind = call->nextCwind;
4381 peer->nDgramPackets = call->nDgramPackets;
4383 call->congestSeq = peer->congestSeq;
4384 /* Reset the resend times on the packets that were nacked
4385 * so we will retransmit as soon as the window permits*/
4386 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4388 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4389 clock_Zero(&tp->retryTime);
4391 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4396 /* If cwind is smaller than ssthresh, then increase
4397 * the window one packet for each ack we receive (exponential
4399 * If cwind is greater than or equal to ssthresh then increase
4400 * the congestion window by one packet for each cwind acks we
4401 * receive (linear growth). */
4402 if (call->cwind < call->ssthresh) {
4404 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4405 call->nCwindAcks = 0;
4407 call->nCwindAcks += newAckCount;
4408 if (call->nCwindAcks >= call->cwind) {
4409 call->nCwindAcks = 0;
4410 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4414 * If we have received several acknowledgements in a row then
4415 * it is time to increase the size of our datagrams
4417 if ((int)call->nAcks > rx_nDgramThreshold) {
4418 if (peer->maxDgramPackets > 1) {
4419 if (call->nDgramPackets < peer->maxDgramPackets) {
4420 call->nDgramPackets++;
4422 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4423 } else if (call->MTU < peer->maxMTU) {
4424 /* don't upgrade if we can't handle it */
4425 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4426 call->MTU = peer->ifMTU;
4428 call->MTU += peer->natMTU;
4429 call->MTU = MIN(call->MTU, peer->maxMTU);
4436 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4438 /* Servers need to hold the call until all response packets have
4439 * been acknowledged. Soft acks are good enough since clients
4440 * are not allowed to clear their receive queues. */
4441 if (call->state == RX_STATE_HOLD
4442 && call->tfirst + call->nSoftAcked >= call->tnext) {
4443 call->state = RX_STATE_DALLY;
4444 rxi_ClearTransmitQueue(call, 0);
4445 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4446 } else if (!queue_IsEmpty(&call->tq)) {
4447 rxi_Start(0, call, 0, istack);
4452 /* Received a response to a challenge packet */
4454 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4455 struct rx_packet *np, int istack)
4459 /* Ignore the packet if we're the client */
4460 if (conn->type == RX_CLIENT_CONNECTION)
4463 /* If already authenticated, ignore the packet (it's probably a retry) */
4464 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4467 /* Otherwise, have the security object evaluate the response packet */
4468 error = RXS_CheckResponse(conn->securityObject, conn, np);
4470 /* If the response is invalid, reset the connection, sending
4471 * an abort to the peer */
4475 rxi_ConnectionError(conn, error);
4476 MUTEX_ENTER(&conn->conn_data_lock);
4477 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4478 MUTEX_EXIT(&conn->conn_data_lock);
4481 /* If the response is valid, any calls waiting to attach
4482 * servers can now do so */
4485 for (i = 0; i < RX_MAXCALLS; i++) {
4486 struct rx_call *call = conn->call[i];
4488 MUTEX_ENTER(&call->lock);
4489 if (call->state == RX_STATE_PRECALL)
4490 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4491 /* tnop can be null if newcallp is null */
4492 MUTEX_EXIT(&call->lock);
4496 /* Update the peer reachability information, just in case
4497 * some calls went into attach-wait while we were waiting
4498 * for authentication..
4500 rxi_UpdatePeerReach(conn, NULL);
4505 /* A client has received an authentication challenge: the security
4506 * object is asked to cough up a respectable response packet to send
4507 * back to the server. The server is responsible for retrying the
4508 * challenge if it fails to get a response. */
4511 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4512 struct rx_packet *np, int istack)
4516 /* Ignore the challenge if we're the server */
4517 if (conn->type == RX_SERVER_CONNECTION)
4520 /* Ignore the challenge if the connection is otherwise idle; someone's
4521 * trying to use us as an oracle. */
4522 if (!rxi_HasActiveCalls(conn))
4525 /* Send the security object the challenge packet. It is expected to fill
4526 * in the response. */
4527 error = RXS_GetResponse(conn->securityObject, conn, np);
4529 /* If the security object is unable to return a valid response, reset the
4530 * connection and send an abort to the peer. Otherwise send the response
4531 * packet to the peer connection. */
4533 rxi_ConnectionError(conn, error);
4534 MUTEX_ENTER(&conn->conn_data_lock);
4535 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4536 MUTEX_EXIT(&conn->conn_data_lock);
4538 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4539 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4545 /* Find an available server process to service the current request in
4546 * the given call structure. If one isn't available, queue up this
4547 * call so it eventually gets one */
4549 rxi_AttachServerProc(struct rx_call *call,
4550 osi_socket socket, int *tnop,
4551 struct rx_call **newcallp)
4553 struct rx_serverQueueEntry *sq;
4554 struct rx_service *service = call->conn->service;
4557 /* May already be attached */
4558 if (call->state == RX_STATE_ACTIVE)
4561 MUTEX_ENTER(&rx_serverPool_lock);
4563 haveQuota = QuotaOK(service);
4564 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4565 /* If there are no processes available to service this call,
4566 * put the call on the incoming call queue (unless it's
4567 * already on the queue).
4569 #ifdef RX_ENABLE_LOCKS
4571 ReturnToServerPool(service);
4572 #endif /* RX_ENABLE_LOCKS */
4574 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4575 call->flags |= RX_CALL_WAIT_PROC;
4576 rx_atomic_inc(&rx_nWaiting);
4577 rx_atomic_inc(&rx_nWaited);
4578 rxi_calltrace(RX_CALL_ARRIVAL, call);
4579 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4580 queue_Append(&rx_incomingCallQueue, call);
4583 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4585 /* If hot threads are enabled, and both newcallp and sq->socketp
4586 * are non-null, then this thread will process the call, and the
4587 * idle server thread will start listening on this threads socket.
4590 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4593 *sq->socketp = socket;
4594 clock_GetTime(&call->startTime);
4595 MUTEX_ENTER(&rx_refcnt_mutex);
4596 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4597 MUTEX_EXIT(&rx_refcnt_mutex);
4601 if (call->flags & RX_CALL_WAIT_PROC) {
4602 /* Conservative: I don't think this should happen */
4603 call->flags &= ~RX_CALL_WAIT_PROC;
4604 if (queue_IsOnQueue(call)) {
4607 rx_atomic_dec(&rx_nWaiting);
4610 call->state = RX_STATE_ACTIVE;
4611 call->mode = RX_MODE_RECEIVING;
4612 #ifdef RX_KERNEL_TRACE
4614 int glockOwner = ISAFS_GLOCK();
4617 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4618 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4624 if (call->flags & RX_CALL_CLEARED) {
4625 /* send an ack now to start the packet flow up again */
4626 call->flags &= ~RX_CALL_CLEARED;
4627 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4629 #ifdef RX_ENABLE_LOCKS
4632 service->nRequestsRunning++;
4633 MUTEX_ENTER(&rx_quota_mutex);
4634 if (service->nRequestsRunning <= service->minProcs)
4637 MUTEX_EXIT(&rx_quota_mutex);
4641 MUTEX_EXIT(&rx_serverPool_lock);
4644 /* Delay the sending of an acknowledge event for a short while, while
4645 * a new call is being prepared (in the case of a client) or a reply
4646 * is being prepared (in the case of a server). Rather than sending
4647 * an ack packet, an ACKALL packet is sent. */
4649 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4651 #ifdef RX_ENABLE_LOCKS
4653 MUTEX_ENTER(&call->lock);
4654 call->delayedAckEvent = NULL;
4655 MUTEX_ENTER(&rx_refcnt_mutex);
4656 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4657 MUTEX_EXIT(&rx_refcnt_mutex);
4659 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4660 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4662 MUTEX_EXIT(&call->lock);
4663 #else /* RX_ENABLE_LOCKS */
4665 call->delayedAckEvent = NULL;
4666 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4667 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4668 #endif /* RX_ENABLE_LOCKS */
4672 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4674 struct rx_call *call = arg1;
4675 #ifdef RX_ENABLE_LOCKS
4677 MUTEX_ENTER(&call->lock);
4678 if (event == call->delayedAckEvent)
4679 call->delayedAckEvent = NULL;
4680 MUTEX_ENTER(&rx_refcnt_mutex);
4681 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4682 MUTEX_EXIT(&rx_refcnt_mutex);
4684 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4686 MUTEX_EXIT(&call->lock);
4687 #else /* RX_ENABLE_LOCKS */
4689 call->delayedAckEvent = NULL;
4690 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4691 #endif /* RX_ENABLE_LOCKS */
4695 #ifdef RX_ENABLE_LOCKS
4696 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4697 * clearing them out.
4700 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4702 struct rx_packet *p, *tp;
4705 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4706 p->flags |= RX_PKTFLAG_ACKED;
4710 call->flags |= RX_CALL_TQ_CLEARME;
4711 call->flags |= RX_CALL_TQ_SOME_ACKED;
4714 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4715 call->tfirst = call->tnext;
4716 call->nSoftAcked = 0;
4718 if (call->flags & RX_CALL_FAST_RECOVER) {
4719 call->flags &= ~RX_CALL_FAST_RECOVER;
4720 call->cwind = call->nextCwind;
4721 call->nextCwind = 0;
4724 CV_SIGNAL(&call->cv_twind);
4726 #endif /* RX_ENABLE_LOCKS */
4728 /* Clear out the transmit queue for the current call (all packets have
4729 * been received by peer) */
4731 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4733 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4734 struct rx_packet *p, *tp;
4736 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4738 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4739 p->flags |= RX_PKTFLAG_ACKED;
4743 call->flags |= RX_CALL_TQ_CLEARME;
4744 call->flags |= RX_CALL_TQ_SOME_ACKED;
4747 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4748 #ifdef RXDEBUG_PACKET
4750 #endif /* RXDEBUG_PACKET */
4751 rxi_FreePackets(0, &call->tq);
4752 rxi_WakeUpTransmitQueue(call);
4753 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4754 call->flags &= ~RX_CALL_TQ_CLEARME;
4756 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4758 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4759 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4760 call->nSoftAcked = 0;
4762 if (call->flags & RX_CALL_FAST_RECOVER) {
4763 call->flags &= ~RX_CALL_FAST_RECOVER;
4764 call->cwind = call->nextCwind;
4766 #ifdef RX_ENABLE_LOCKS
4767 CV_SIGNAL(&call->cv_twind);
4769 osi_rxWakeup(&call->twind);
4774 rxi_ClearReceiveQueue(struct rx_call *call)
4776 if (queue_IsNotEmpty(&call->rq)) {
4779 count = rxi_FreePackets(0, &call->rq);
4780 rx_packetReclaims += count;
4781 #ifdef RXDEBUG_PACKET
4783 if ( call->rqc != 0 )
4784 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
4786 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4788 if (call->state == RX_STATE_PRECALL) {
4789 call->flags |= RX_CALL_CLEARED;
4793 /* Send an abort packet for the specified call */
4795 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
4796 int istack, int force)
4799 struct clock when, now;
4804 /* Clients should never delay abort messages */
4805 if (rx_IsClientConn(call->conn))
4808 if (call->abortCode != call->error) {
4809 call->abortCode = call->error;
4810 call->abortCount = 0;
4813 if (force || rxi_callAbortThreshhold == 0
4814 || call->abortCount < rxi_callAbortThreshhold) {
4815 if (call->delayedAbortEvent) {
4816 rxevent_Cancel(call->delayedAbortEvent, call,
4817 RX_CALL_REFCOUNT_ABORT);
4819 error = htonl(call->error);
4822 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4823 (char *)&error, sizeof(error), istack);
4824 } else if (!call->delayedAbortEvent) {
4825 clock_GetTime(&now);
4827 clock_Addmsec(&when, rxi_callAbortDelay);
4828 MUTEX_ENTER(&rx_refcnt_mutex);
4829 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4830 MUTEX_EXIT(&rx_refcnt_mutex);
4831 call->delayedAbortEvent =
4832 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4837 /* Send an abort packet for the specified connection. Packet is an
4838 * optional pointer to a packet that can be used to send the abort.
4839 * Once the number of abort messages reaches the threshhold, an
4840 * event is scheduled to send the abort. Setting the force flag
4841 * overrides sending delayed abort messages.
4843 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4844 * to send the abort packet.
4847 rxi_SendConnectionAbort(struct rx_connection *conn,
4848 struct rx_packet *packet, int istack, int force)
4851 struct clock when, now;
4856 /* Clients should never delay abort messages */
4857 if (rx_IsClientConn(conn))
4860 if (force || rxi_connAbortThreshhold == 0
4861 || conn->abortCount < rxi_connAbortThreshhold) {
4862 if (conn->delayedAbortEvent) {
4863 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4865 error = htonl(conn->error);
4867 MUTEX_EXIT(&conn->conn_data_lock);
4869 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4870 RX_PACKET_TYPE_ABORT, (char *)&error,
4871 sizeof(error), istack);
4872 MUTEX_ENTER(&conn->conn_data_lock);
4873 } else if (!conn->delayedAbortEvent) {
4874 clock_GetTime(&now);
4876 clock_Addmsec(&when, rxi_connAbortDelay);
4877 conn->delayedAbortEvent =
4878 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4883 /* Associate an error all of the calls owned by a connection. Called
4884 * with error non-zero. This is only for really fatal things, like
4885 * bad authentication responses. The connection itself is set in
4886 * error at this point, so that future packets received will be
4889 rxi_ConnectionError(struct rx_connection *conn,
4895 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
4897 MUTEX_ENTER(&conn->conn_data_lock);
4898 if (conn->challengeEvent)
4899 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4900 if (conn->natKeepAliveEvent)
4901 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
4902 if (conn->checkReachEvent) {
4903 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4904 conn->checkReachEvent = 0;
4905 conn->flags &= ~RX_CONN_ATTACHWAIT;
4906 MUTEX_ENTER(&rx_refcnt_mutex);
4908 MUTEX_EXIT(&rx_refcnt_mutex);
4910 MUTEX_EXIT(&conn->conn_data_lock);
4911 for (i = 0; i < RX_MAXCALLS; i++) {
4912 struct rx_call *call = conn->call[i];
4914 MUTEX_ENTER(&call->lock);
4915 rxi_CallError(call, error);
4916 MUTEX_EXIT(&call->lock);
4919 conn->error = error;
4920 if (rx_stats_active)
4921 rx_atomic_inc(&rx_stats.fatalErrors);
4926 * Interrupt an in-progress call with the specified error and wakeup waiters.
4928 * @param[in] call The call to interrupt
4929 * @param[in] error The error code to send to the peer
4932 rx_InterruptCall(struct rx_call *call, afs_int32 error)
4934 MUTEX_ENTER(&call->lock);
4935 rxi_CallError(call, error);
4936 rxi_SendCallAbort(call, NULL, 0, 1);
4937 MUTEX_EXIT(&call->lock);
4941 rxi_CallError(struct rx_call *call, afs_int32 error)
4944 osirx_AssertMine(&call->lock, "rxi_CallError");
4946 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
4948 error = call->error;
4950 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4951 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4952 rxi_ResetCall(call, 0);
4955 rxi_ResetCall(call, 0);
4957 call->error = error;
4960 /* Reset various fields in a call structure, and wakeup waiting
4961 * processes. Some fields aren't changed: state & mode are not
4962 * touched (these must be set by the caller), and bufptr, nLeft, and
4963 * nFree are not reset, since these fields are manipulated by
4964 * unprotected macros, and may only be reset by non-interrupting code.
4967 /* this code requires that call->conn be set properly as a pre-condition. */
4968 #endif /* ADAPT_WINDOW */
4971 rxi_ResetCall(struct rx_call *call, int newcall)
4974 struct rx_peer *peer;
4975 struct rx_packet *packet;
4977 osirx_AssertMine(&call->lock, "rxi_ResetCall");
4979 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
4981 /* Notify anyone who is waiting for asynchronous packet arrival */
4982 if (call->arrivalProc) {
4983 (*call->arrivalProc) (call, call->arrivalProcHandle,
4984 call->arrivalProcArg);
4985 call->arrivalProc = (void (*)())0;
4988 if (call->delayedAbortEvent) {
4989 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4990 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4992 rxi_SendCallAbort(call, packet, 0, 1);
4993 rxi_FreePacket(packet);
4998 * Update the peer with the congestion information in this call
4999 * so other calls on this connection can pick up where this call
5000 * left off. If the congestion sequence numbers don't match then
5001 * another call experienced a retransmission.
5003 peer = call->conn->peer;
5004 MUTEX_ENTER(&peer->peer_lock);
5006 if (call->congestSeq == peer->congestSeq) {
5007 peer->cwind = MAX(peer->cwind, call->cwind);
5008 peer->MTU = MAX(peer->MTU, call->MTU);
5009 peer->nDgramPackets =
5010 MAX(peer->nDgramPackets, call->nDgramPackets);
5013 call->abortCode = 0;
5014 call->abortCount = 0;
5016 if (peer->maxDgramPackets > 1) {
5017 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5019 call->MTU = peer->MTU;
5021 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5022 call->ssthresh = rx_maxSendWindow;
5023 call->nDgramPackets = peer->nDgramPackets;
5024 call->congestSeq = peer->congestSeq;
5025 MUTEX_EXIT(&peer->peer_lock);
5027 flags = call->flags;
5028 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5029 rxi_WaitforTQBusy(call);
5030 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5032 rxi_ClearTransmitQueue(call, 1);
5033 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5034 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5038 rxi_ClearReceiveQueue(call);
5039 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5043 call->twind = call->conn->twind[call->channel];
5044 call->rwind = call->conn->rwind[call->channel];
5045 call->nSoftAcked = 0;
5046 call->nextCwind = 0;
5049 call->nCwindAcks = 0;
5050 call->nSoftAcks = 0;
5051 call->nHardAcks = 0;
5053 call->tfirst = call->rnext = call->tnext = 1;
5056 call->lastAcked = 0;
5057 call->localStatus = call->remoteStatus = 0;
5059 if (flags & RX_CALL_READER_WAIT) {
5060 #ifdef RX_ENABLE_LOCKS
5061 CV_BROADCAST(&call->cv_rq);
5063 osi_rxWakeup(&call->rq);
5066 if (flags & RX_CALL_WAIT_PACKETS) {
5067 MUTEX_ENTER(&rx_freePktQ_lock);
5068 rxi_PacketsUnWait(); /* XXX */
5069 MUTEX_EXIT(&rx_freePktQ_lock);
5071 #ifdef RX_ENABLE_LOCKS
5072 CV_SIGNAL(&call->cv_twind);
5074 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5075 osi_rxWakeup(&call->twind);
5078 #ifdef RX_ENABLE_LOCKS
5079 /* The following ensures that we don't mess with any queue while some
5080 * other thread might also be doing so. The call_queue_lock field is
5081 * is only modified under the call lock. If the call is in the process
5082 * of being removed from a queue, the call is not locked until the
5083 * the queue lock is dropped and only then is the call_queue_lock field
5084 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5085 * Note that any other routine which removes a call from a queue has to
5086 * obtain the queue lock before examing the queue and removing the call.
5088 if (call->call_queue_lock) {
5089 MUTEX_ENTER(call->call_queue_lock);
5090 if (queue_IsOnQueue(call)) {
5092 if (flags & RX_CALL_WAIT_PROC) {
5093 rx_atomic_dec(&rx_nWaiting);
5096 MUTEX_EXIT(call->call_queue_lock);
5097 CLEAR_CALL_QUEUE_LOCK(call);
5099 #else /* RX_ENABLE_LOCKS */
5100 if (queue_IsOnQueue(call)) {
5102 if (flags & RX_CALL_WAIT_PROC)
5103 rx_atomic_dec(&rx_nWaiting);
5105 #endif /* RX_ENABLE_LOCKS */
5107 rxi_KeepAliveOff(call);
5108 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5111 /* Send an acknowledge for the indicated packet (seq,serial) of the
5112 * indicated call, for the indicated reason (reason). This
5113 * acknowledge will specifically acknowledge receiving the packet, and
5114 * will also specify which other packets for this call have been
5115 * received. This routine returns the packet that was used to the
5116 * caller. The caller is responsible for freeing it or re-using it.
5117 * This acknowledgement also returns the highest sequence number
5118 * actually read out by the higher level to the sender; the sender
5119 * promises to keep around packets that have not been read by the
5120 * higher level yet (unless, of course, the sender decides to abort
5121 * the call altogether). Any of p, seq, serial, pflags, or reason may
5122 * be set to zero without ill effect. That is, if they are zero, they
5123 * will not convey any information.
5124 * NOW there is a trailer field, after the ack where it will safely be
5125 * ignored by mundanes, which indicates the maximum size packet this
5126 * host can swallow. */
5128 struct rx_packet *optionalPacket; use to send ack (or null)
5129 int seq; Sequence number of the packet we are acking
5130 int serial; Serial number of the packet
5131 int pflags; Flags field from packet header
5132 int reason; Reason an acknowledge was prompted
5136 rxi_SendAck(struct rx_call *call,
5137 struct rx_packet *optionalPacket, int serial, int reason,
5140 struct rx_ackPacket *ap;
5141 struct rx_packet *rqp;
5142 struct rx_packet *nxp; /* For queue_Scan */
5143 struct rx_packet *p;
5146 afs_uint32 padbytes = 0;
5147 #ifdef RX_ENABLE_TSFPQ
5148 struct rx_ts_info_t * rx_ts_info;
5152 * Open the receive window once a thread starts reading packets
5154 if (call->rnext > 1) {
5155 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5158 /* Don't attempt to grow MTU if this is a critical ping */
5159 if (reason == RX_ACK_MTU) {
5160 /* keep track of per-call attempts, if we're over max, do in small
5161 * otherwise in larger? set a size to increment by, decrease
5164 if (call->conn->peer->maxPacketSize &&
5165 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5167 padbytes = call->conn->peer->maxPacketSize+16;
5169 padbytes = call->conn->peer->maxMTU + 128;
5171 /* do always try a minimum size ping */
5172 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5174 /* subtract the ack payload */
5175 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5176 reason = RX_ACK_PING;
5179 call->nHardAcks = 0;
5180 call->nSoftAcks = 0;
5181 if (call->rnext > call->lastAcked)
5182 call->lastAcked = call->rnext;
5186 rx_computelen(p, p->length); /* reset length, you never know */
5187 } /* where that's been... */
5188 #ifdef RX_ENABLE_TSFPQ
5190 RX_TS_INFO_GET(rx_ts_info);
5191 if ((p = rx_ts_info->local_special_packet)) {
5192 rx_computelen(p, p->length);
5193 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5194 rx_ts_info->local_special_packet = p;
5195 } else { /* We won't send the ack, but don't panic. */
5196 return optionalPacket;
5200 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5201 /* We won't send the ack, but don't panic. */
5202 return optionalPacket;
5207 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5210 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5211 #ifndef RX_ENABLE_TSFPQ
5212 if (!optionalPacket)
5215 return optionalPacket;
5217 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5218 if (rx_Contiguous(p) < templ) {
5219 #ifndef RX_ENABLE_TSFPQ
5220 if (!optionalPacket)
5223 return optionalPacket;
5228 /* MTUXXX failing to send an ack is very serious. We should */
5229 /* try as hard as possible to send even a partial ack; it's */
5230 /* better than nothing. */
5231 ap = (struct rx_ackPacket *)rx_DataOf(p);
5232 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5233 ap->reason = reason;
5235 /* The skew computation used to be bogus, I think it's better now. */
5236 /* We should start paying attention to skew. XXX */
5237 ap->serial = htonl(serial);
5238 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5240 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
5241 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5243 /* No fear of running out of ack packet here because there can only be at most
5244 * one window full of unacknowledged packets. The window size must be constrained
5245 * to be less than the maximum ack size, of course. Also, an ack should always
5246 * fit into a single packet -- it should not ever be fragmented. */
5247 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5248 if (!rqp || !call->rq.next
5249 || (rqp->header.seq > (call->rnext + call->rwind))) {
5250 #ifndef RX_ENABLE_TSFPQ
5251 if (!optionalPacket)
5254 rxi_CallError(call, RX_CALL_DEAD);
5255 return optionalPacket;
5258 while (rqp->header.seq > call->rnext + offset)
5259 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5260 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5262 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5263 #ifndef RX_ENABLE_TSFPQ
5264 if (!optionalPacket)
5267 rxi_CallError(call, RX_CALL_DEAD);
5268 return optionalPacket;
5273 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5275 /* these are new for AFS 3.3 */
5276 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5277 templ = htonl(templ);
5278 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5279 templ = htonl(call->conn->peer->ifMTU);
5280 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5281 sizeof(afs_int32), &templ);
5283 /* new for AFS 3.4 */
5284 templ = htonl(call->rwind);
5285 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5286 sizeof(afs_int32), &templ);
5288 /* new for AFS 3.5 */
5289 templ = htonl(call->conn->peer->ifDgramPackets);
5290 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5291 sizeof(afs_int32), &templ);
5293 p->header.serviceId = call->conn->serviceId;
5294 p->header.cid = (call->conn->cid | call->channel);
5295 p->header.callNumber = *call->callNumber;
5297 p->header.securityIndex = call->conn->securityIndex;
5298 p->header.epoch = call->conn->epoch;
5299 p->header.type = RX_PACKET_TYPE_ACK;
5300 p->header.flags = RX_SLOW_START_OK;
5301 if (reason == RX_ACK_PING) {
5302 p->header.flags |= RX_REQUEST_ACK;
5304 clock_GetTime(&call->pingRequestTime);
5307 p->length = padbytes +
5308 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5311 /* not fast but we can potentially use this if truncated
5312 * fragments are delivered to figure out the mtu.
5314 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5315 sizeof(afs_int32), sizeof(afs_int32),
5319 if (call->conn->type == RX_CLIENT_CONNECTION)
5320 p->header.flags |= RX_CLIENT_INITIATED;
5324 if (rxdebug_active) {
5328 len = _snprintf(msg, sizeof(msg),
5329 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5330 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5331 ntohl(ap->serial), ntohl(ap->previousPacket),
5332 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5333 ap->nAcks, ntohs(ap->bufferSpace) );
5337 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5338 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5342 OutputDebugString(msg);
5344 #else /* AFS_NT40_ENV */
5346 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5347 ap->reason, ntohl(ap->previousPacket),
5348 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5350 for (offset = 0; offset < ap->nAcks; offset++)
5351 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5356 #endif /* AFS_NT40_ENV */
5359 int i, nbytes = p->length;
5361 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5362 if (nbytes <= p->wirevec[i].iov_len) {
5365 savelen = p->wirevec[i].iov_len;
5367 p->wirevec[i].iov_len = nbytes;
5369 rxi_Send(call, p, istack);
5370 p->wirevec[i].iov_len = savelen;
5374 nbytes -= p->wirevec[i].iov_len;
5377 if (rx_stats_active)
5378 rx_atomic_inc(&rx_stats.ackPacketsSent);
5379 #ifndef RX_ENABLE_TSFPQ
5380 if (!optionalPacket)
5383 return optionalPacket; /* Return packet for re-use by caller */
5387 struct rx_packet **list;
5392 /* Send all of the packets in the list in single datagram */
5394 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5395 int istack, int moreFlag, struct clock *now,
5396 struct clock *retryTime)
5401 struct rx_connection *conn = call->conn;
5402 struct rx_peer *peer = conn->peer;
5404 MUTEX_ENTER(&peer->peer_lock);
5405 peer->nSent += xmit->len;
5406 if (xmit->resending)
5407 peer->reSends += xmit->len;
5408 MUTEX_EXIT(&peer->peer_lock);
5410 if (rx_stats_active) {
5411 if (xmit->resending)
5412 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5414 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5417 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5421 /* Set the packet flags and schedule the resend events */
5422 /* Only request an ack for the last packet in the list */
5423 for (i = 0; i < xmit->len; i++) {
5424 struct rx_packet *packet = xmit->list[i];
5426 packet->retryTime = *retryTime;
5427 if (packet->header.serial) {
5428 /* Exponentially backoff retry times */
5429 if (packet->backoff < MAXBACKOFF) {
5430 /* so it can't stay == 0 */
5431 packet->backoff = (packet->backoff << 1) + 1;
5434 clock_Addmsec(&(packet->retryTime),
5435 ((afs_uint32) packet->backoff) << 8);
5438 /* Wait a little extra for the ack on the last packet */
5440 && !(packet->header.flags & RX_CLIENT_INITIATED)) {
5441 clock_Addmsec(&(packet->retryTime), 400);
5444 /* Record the time sent */
5445 packet->timeSent = *now;
5447 /* Ask for an ack on retransmitted packets, on every other packet
5448 * if the peer doesn't support slow start. Ask for an ack on every
5449 * packet until the congestion window reaches the ack rate. */
5450 if (packet->header.serial) {
5453 /* improved RTO calculation- not Karn */
5454 packet->firstSent = *now;
5455 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5456 || (!(call->flags & RX_CALL_SLOW_START_OK)
5457 && (packet->header.seq & 1)))) {
5462 /* Tag this packet as not being the last in this group,
5463 * for the receiver's benefit */
5464 if (i < xmit->len - 1 || moreFlag) {
5465 packet->header.flags |= RX_MORE_PACKETS;
5468 /* Install the new retransmit time for the packet, and
5469 * record the time sent */
5470 packet->timeSent = *now;
5474 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5477 /* Since we're about to send a data packet to the peer, it's
5478 * safe to nuke any scheduled end-of-packets ack */
5479 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5481 MUTEX_EXIT(&call->lock);
5482 MUTEX_ENTER(&rx_refcnt_mutex);
5483 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5484 MUTEX_EXIT(&rx_refcnt_mutex);
5485 if (xmit->len > 1) {
5486 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5488 rxi_SendPacket(call, conn, xmit->list[0], istack);
5490 MUTEX_ENTER(&call->lock);
5491 MUTEX_ENTER(&rx_refcnt_mutex);
5492 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5493 MUTEX_EXIT(&rx_refcnt_mutex);
5495 /* Update last send time for this call (for keep-alive
5496 * processing), and for the connection (so that we can discover
5497 * idle connections) */
5498 conn->lastSendTime = call->lastSendTime = clock_Sec();
5499 /* Let a set of retransmits trigger an idle timeout */
5500 if (!xmit->resending)
5501 call->lastSendData = call->lastSendTime;
5504 /* When sending packets we need to follow these rules:
5505 * 1. Never send more than maxDgramPackets in a jumbogram.
5506 * 2. Never send a packet with more than two iovecs in a jumbogram.
5507 * 3. Never send a retransmitted packet in a jumbogram.
5508 * 4. Never send more than cwind/4 packets in a jumbogram
5509 * We always keep the last list we should have sent so we
5510 * can set the RX_MORE_PACKETS flags correctly.
5514 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5515 int istack, struct clock *now, struct clock *retryTime)
5518 struct xmitlist working;
5519 struct xmitlist last;
5521 struct rx_peer *peer = call->conn->peer;
5522 int morePackets = 0;
5524 memset(&last, 0, sizeof(struct xmitlist));
5525 working.list = &list[0];
5527 working.resending = 0;
5529 for (i = 0; i < len; i++) {
5530 /* Does the current packet force us to flush the current list? */
5532 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5533 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5535 /* This sends the 'last' list and then rolls the current working
5536 * set into the 'last' one, and resets the working set */
5539 rxi_SendList(call, &last, istack, 1, now, retryTime);
5540 /* If the call enters an error state stop sending, or if
5541 * we entered congestion recovery mode, stop sending */
5542 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5547 working.resending = 0;
5548 working.list = &list[i];
5550 /* Add the current packet to the list if it hasn't been acked.
5551 * Otherwise adjust the list pointer to skip the current packet. */
5552 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5555 if (list[i]->header.serial)
5556 working.resending = 1;
5558 /* Do we need to flush the list? */
5559 if (working.len >= (int)peer->maxDgramPackets
5560 || working.len >= (int)call->nDgramPackets
5561 || working.len >= (int)call->cwind
5562 || list[i]->header.serial
5563 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5565 rxi_SendList(call, &last, istack, 1, now, retryTime);
5566 /* If the call enters an error state stop sending, or if
5567 * we entered congestion recovery mode, stop sending */
5569 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5574 working.resending = 0;
5575 working.list = &list[i + 1];
5578 if (working.len != 0) {
5579 osi_Panic("rxi_SendList error");
5581 working.list = &list[i + 1];
5585 /* Send the whole list when the call is in receive mode, when
5586 * the call is in eof mode, when we are in fast recovery mode,
5587 * and when we have the last packet */
5588 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5589 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5590 || (call->flags & RX_CALL_FAST_RECOVER)) {
5591 /* Check for the case where the current list contains
5592 * an acked packet. Since we always send retransmissions
5593 * in a separate packet, we only need to check the first
5594 * packet in the list */
5595 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5599 rxi_SendList(call, &last, istack, morePackets, now, retryTime);
5600 /* If the call enters an error state stop sending, or if
5601 * we entered congestion recovery mode, stop sending */
5602 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5606 rxi_SendList(call, &working, istack, 0, now, retryTime);
5608 } else if (last.len > 0) {
5609 rxi_SendList(call, &last, istack, 0, now, retryTime);
5610 /* Packets which are in 'working' are not sent by this call */
5614 #ifdef RX_ENABLE_LOCKS
5615 /* Call rxi_Start, below, but with the call lock held. */
5617 rxi_StartUnlocked(struct rxevent *event,
5618 void *arg0, void *arg1, int istack)
5620 struct rx_call *call = arg0;
5622 MUTEX_ENTER(&call->lock);
5623 rxi_Start(event, call, arg1, istack);
5624 MUTEX_EXIT(&call->lock);
5626 #endif /* RX_ENABLE_LOCKS */
5628 /* This routine is called when new packets are readied for
5629 * transmission and when retransmission may be necessary, or when the
5630 * transmission window or burst count are favourable. This should be
5631 * better optimized for new packets, the usual case, now that we've
5632 * got rid of queues of send packets. XXXXXXXXXXX */
5634 rxi_Start(struct rxevent *event,
5635 void *arg0, void *arg1, int istack)
5637 struct rx_call *call = arg0;
5639 struct rx_packet *p;
5640 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5641 struct rx_peer *peer = call->conn->peer;
5642 struct clock now, usenow, retryTime;
5647 /* If rxi_Start is being called as a result of a resend event,
5648 * then make sure that the event pointer is removed from the call
5649 * structure, since there is no longer a per-call retransmission
5651 if (event && event == call->resendEvent) {
5652 MUTEX_ENTER(&rx_refcnt_mutex);
5653 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5654 MUTEX_EXIT(&rx_refcnt_mutex);
5655 call->resendEvent = NULL;
5656 if (queue_IsEmpty(&call->tq)) {
5660 /* Timeouts trigger congestion recovery */
5661 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5662 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5663 /* someone else is waiting to start recovery */
5666 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5667 rxi_WaitforTQBusy(call);
5668 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5669 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5670 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5672 if (rx_stats_active)
5673 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5677 call->flags |= RX_CALL_FAST_RECOVER;
5679 if (peer->maxDgramPackets > 1) {
5680 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5682 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5684 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5685 call->nDgramPackets = 1;
5687 call->nextCwind = 1;
5690 MUTEX_ENTER(&peer->peer_lock);
5691 peer->MTU = call->MTU;
5692 peer->cwind = call->cwind;
5693 peer->nDgramPackets = 1;
5695 call->congestSeq = peer->congestSeq;
5696 MUTEX_EXIT(&peer->peer_lock);
5697 /* Clear retry times on packets. Otherwise, it's possible for
5698 * some packets in the queue to force resends at rates faster
5699 * than recovery rates.
5701 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5702 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5703 clock_Zero(&p->retryTime);
5708 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5709 if (rx_stats_active)
5710 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5715 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5716 /* Get clock to compute the re-transmit time for any packets
5717 * in this burst. Note, if we back off, it's reasonable to
5718 * back off all of the packets in the same manner, even if
5719 * some of them have been retransmitted more times than more
5721 * Do a dance to avoid blocking after setting now. */
5722 MUTEX_ENTER(&peer->peer_lock);
5723 retryTime = peer->timeout;
5724 MUTEX_EXIT(&peer->peer_lock);
5726 clock_GetTime(&now);
5727 clock_Add(&retryTime, &now);
5729 /* Send (or resend) any packets that need it, subject to
5730 * window restrictions and congestion burst control
5731 * restrictions. Ask for an ack on the last packet sent in
5732 * this burst. For now, we're relying upon the window being
5733 * considerably bigger than the largest number of packets that
5734 * are typically sent at once by one initial call to
5735 * rxi_Start. This is probably bogus (perhaps we should ask
5736 * for an ack when we're half way through the current
5737 * window?). Also, for non file transfer applications, this
5738 * may end up asking for an ack for every packet. Bogus. XXXX
5741 * But check whether we're here recursively, and let the other guy
5744 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5745 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5746 call->flags |= RX_CALL_TQ_BUSY;
5748 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5750 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5751 call->flags &= ~RX_CALL_NEED_START;
5752 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5754 maxXmitPackets = MIN(call->twind, call->cwind);
5755 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5756 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5757 /* We shouldn't be sending packets if a thread is waiting
5758 * to initiate congestion recovery */
5759 dpf(("call %d waiting to initiate fast recovery\n",
5760 *(call->callNumber)));
5764 && (call->flags & RX_CALL_FAST_RECOVER)) {
5765 /* Only send one packet during fast recovery */
5766 dpf(("call %d restricted to one packet per send during fast recovery\n",
5767 *(call->callNumber)));
5770 #ifdef RX_TRACK_PACKETS
5771 if ((p->flags & RX_PKTFLAG_FREE)
5772 || (!queue_IsEnd(&call->tq, nxp)
5773 && (nxp->flags & RX_PKTFLAG_FREE))
5774 || (p == (struct rx_packet *)&rx_freePacketQueue)
5775 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5776 osi_Panic("rxi_Start: xmit queue clobbered");
5779 if (p->flags & RX_PKTFLAG_ACKED) {
5780 /* Since we may block, don't trust this */
5781 usenow.sec = usenow.usec = 0;
5782 if (rx_stats_active)
5783 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
5784 continue; /* Ignore this packet if it has been acknowledged */
5787 /* Turn off all flags except these ones, which are the same
5788 * on each transmission */
5789 p->header.flags &= RX_PRESET_FLAGS;
5791 if (p->header.seq >=
5792 call->tfirst + MIN((int)call->twind,
5793 (int)(call->nSoftAcked +
5795 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5796 /* Note: if we're waiting for more window space, we can
5797 * still send retransmits; hence we don't return here, but
5798 * break out to schedule a retransmit event */
5799 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
5800 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
5805 /* Transmit the packet if it needs to be sent. */
5806 if (!clock_Lt(&now, &p->retryTime)) {
5807 if (nXmitPackets == maxXmitPackets) {
5808 rxi_SendXmitList(call, call->xmitList,
5809 nXmitPackets, istack, &now,
5813 dpf(("call %d xmit packet %"AFS_PTR_FMT" now %u.%06u retryTime %u.%06u nextRetry %u.%06u\n",
5814 *(call->callNumber), p,
5816 p->retryTime.sec, p->retryTime.usec,
5817 retryTime.sec, retryTime.usec));
5818 call->xmitList[nXmitPackets++] = p;
5822 /* xmitList now hold pointers to all of the packets that are
5823 * ready to send. Now we loop to send the packets */
5824 if (nXmitPackets > 0) {
5825 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
5826 istack, &now, &retryTime);
5829 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5831 * TQ references no longer protected by this flag; they must remain
5832 * protected by the global lock.
5834 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5835 call->flags &= ~RX_CALL_TQ_BUSY;
5836 rxi_WakeUpTransmitQueue(call);
5840 /* We went into the error state while sending packets. Now is
5841 * the time to reset the call. This will also inform the using
5842 * process that the call is in an error state.
5844 if (rx_stats_active)
5845 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
5846 call->flags &= ~RX_CALL_TQ_BUSY;
5847 rxi_WakeUpTransmitQueue(call);
5848 rxi_CallError(call, call->error);
5851 #ifdef RX_ENABLE_LOCKS
5852 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5854 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5855 /* Some packets have received acks. If they all have, we can clear
5856 * the transmit queue.
5859 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5860 if (p->header.seq < call->tfirst
5861 && (p->flags & RX_PKTFLAG_ACKED)) {
5863 #ifdef RX_TRACK_PACKETS
5864 p->flags &= ~RX_PKTFLAG_TQ;
5866 #ifdef RXDEBUG_PACKET
5874 call->flags |= RX_CALL_TQ_CLEARME;
5876 #endif /* RX_ENABLE_LOCKS */
5877 /* Don't bother doing retransmits if the TQ is cleared. */
5878 if (call->flags & RX_CALL_TQ_CLEARME) {
5879 rxi_ClearTransmitQueue(call, 1);
5881 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5884 /* Always post a resend event, if there is anything in the
5885 * queue, and resend is possible. There should be at least
5886 * one unacknowledged packet in the queue ... otherwise none
5887 * of these packets should be on the queue in the first place.
5889 if (call->resendEvent) {
5890 /* Cancel the existing event and post a new one */
5891 rxevent_Cancel(call->resendEvent, call,
5892 RX_CALL_REFCOUNT_RESEND);
5895 /* The retry time is the retry time on the first unacknowledged
5896 * packet inside the current window */
5898 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5899 /* Don't set timers for packets outside the window */
5900 if (p->header.seq >= call->tfirst + call->twind) {
5904 if (!(p->flags & RX_PKTFLAG_ACKED)
5905 && !clock_IsZero(&p->retryTime)) {
5907 retryTime = p->retryTime;
5912 /* Post a new event to re-run rxi_Start when retries may be needed */
5913 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5914 #ifdef RX_ENABLE_LOCKS
5915 MUTEX_ENTER(&rx_refcnt_mutex);
5916 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5917 MUTEX_EXIT(&rx_refcnt_mutex);
5919 rxevent_PostNow2(&retryTime, &usenow,
5921 (void *)call, 0, istack);
5922 #else /* RX_ENABLE_LOCKS */
5924 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5925 (void *)call, 0, istack);
5926 #endif /* RX_ENABLE_LOCKS */
5929 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5930 } while (call->flags & RX_CALL_NEED_START);
5932 * TQ references no longer protected by this flag; they must remain
5933 * protected by the global lock.
5935 call->flags &= ~RX_CALL_TQ_BUSY;
5936 rxi_WakeUpTransmitQueue(call);
5938 call->flags |= RX_CALL_NEED_START;
5940 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5942 if (call->resendEvent) {
5943 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5948 /* Also adjusts the keep alive parameters for the call, to reflect
5949 * that we have just sent a packet (so keep alives aren't sent
5952 rxi_Send(struct rx_call *call, struct rx_packet *p,
5955 struct rx_connection *conn = call->conn;
5957 /* Stamp each packet with the user supplied status */
5958 p->header.userStatus = call->localStatus;
5960 /* Allow the security object controlling this call's security to
5961 * make any last-minute changes to the packet */
5962 RXS_SendPacket(conn->securityObject, call, p);
5964 /* Since we're about to send SOME sort of packet to the peer, it's
5965 * safe to nuke any scheduled end-of-packets ack */
5966 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5968 /* Actually send the packet, filling in more connection-specific fields */
5969 MUTEX_EXIT(&call->lock);
5970 MUTEX_ENTER(&rx_refcnt_mutex);
5971 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5972 MUTEX_EXIT(&rx_refcnt_mutex);
5973 rxi_SendPacket(call, conn, p, istack);
5974 MUTEX_ENTER(&rx_refcnt_mutex);
5975 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5976 MUTEX_EXIT(&rx_refcnt_mutex);
5977 MUTEX_ENTER(&call->lock);
5979 /* Update last send time for this call (for keep-alive
5980 * processing), and for the connection (so that we can discover
5981 * idle connections) */
5982 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5983 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
5984 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
5986 conn->lastSendTime = call->lastSendTime = clock_Sec();
5987 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
5988 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5989 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
5990 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
5991 RX_ACK_PING_RESPONSE)))
5992 call->lastSendData = call->lastSendTime;
5996 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5997 * that things are fine. Also called periodically to guarantee that nothing
5998 * falls through the cracks (e.g. (error + dally) connections have keepalive
5999 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6001 * haveCTLock Set if calling from rxi_ReapConnections
6003 #ifdef RX_ENABLE_LOCKS
6005 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6006 #else /* RX_ENABLE_LOCKS */
6008 rxi_CheckCall(struct rx_call *call)
6009 #endif /* RX_ENABLE_LOCKS */
6011 struct rx_connection *conn = call->conn;
6013 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6014 afs_uint32 fudgeFactor;
6018 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6019 if (call->flags & RX_CALL_TQ_BUSY) {
6020 /* Call is active and will be reset by rxi_Start if it's
6021 * in an error state.
6026 /* RTT + 8*MDEV, rounded up to the next second. */
6027 fudgeFactor = (((afs_uint32) conn->peer->rtt >> 3) +
6028 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
6030 deadTime = conn->secondsUntilDead + fudgeFactor;
6032 /* These are computed to the second (+- 1 second). But that's
6033 * good enough for these values, which should be a significant
6034 * number of seconds. */
6035 if (now > (call->lastReceiveTime + deadTime)) {
6036 if (call->state == RX_STATE_ACTIVE) {
6038 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
6040 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6041 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6042 ip_stack_t *ipst = ns->netstack_ip;
6044 ire = ire_cache_lookup(conn->peer->host
6045 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6047 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
6049 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6056 if (ire && ire->ire_max_frag > 0)
6057 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6059 #if defined(GLOBAL_NETSTACKID)
6063 #endif /* ADAPT_PMTU */
6064 cerror = RX_CALL_DEAD;
6067 #ifdef RX_ENABLE_LOCKS
6068 /* Cancel pending events */
6069 rxevent_Cancel(call->delayedAckEvent, call,
6070 RX_CALL_REFCOUNT_DELAY);
6071 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
6072 rxevent_Cancel(call->keepAliveEvent, call,
6073 RX_CALL_REFCOUNT_ALIVE);
6074 MUTEX_ENTER(&rx_refcnt_mutex);
6075 if (call->refCount == 0) {
6076 rxi_FreeCall(call, haveCTLock);
6077 MUTEX_EXIT(&rx_refcnt_mutex);
6080 MUTEX_EXIT(&rx_refcnt_mutex);
6082 #else /* RX_ENABLE_LOCKS */
6083 rxi_FreeCall(call, 0);
6085 #endif /* RX_ENABLE_LOCKS */
6087 /* Non-active calls are destroyed if they are not responding
6088 * to pings; active calls are simply flagged in error, so the
6089 * attached process can die reasonably gracefully. */
6092 if (conn->idleDeadTime) {
6093 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6096 /* see if we have a non-activity timeout */
6097 if (call->startWait && idleDeadTime
6098 && ((call->startWait + idleDeadTime) < now) &&
6099 (call->flags & RX_CALL_READER_WAIT)) {
6100 if (call->state == RX_STATE_ACTIVE) {
6101 cerror = RX_CALL_TIMEOUT;
6105 if (call->lastSendData && idleDeadTime && (conn->idleDeadErr != 0)
6106 && ((call->lastSendData + idleDeadTime) < now)) {
6107 if (call->state == RX_STATE_ACTIVE) {
6108 cerror = conn->idleDeadErr;
6114 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6117 /* see if we have a hard timeout */
6119 && (now > (hardDeadTime + call->startTime.sec))) {
6120 if (call->state == RX_STATE_ACTIVE)
6121 rxi_CallError(call, RX_CALL_TIMEOUT);
6126 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT
6127 && call->lastReceiveTime) {
6128 int oldMTU = conn->peer->ifMTU;
6130 /* if we thought we could send more, perhaps things got worse */
6131 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6132 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6133 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6134 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6136 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6138 /* minimum capped in SetPeerMtu */
6139 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6142 conn->lastPacketSize = 0;
6144 /* needed so ResetCall doesn't clobber us. */
6145 call->MTU = conn->peer->ifMTU;
6147 /* if we never succeeded, let the error pass out as-is */
6148 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6149 cerror = conn->msgsizeRetryErr;
6152 rxi_CallError(call, cerror);
6157 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6159 struct rx_connection *conn = arg1;
6160 struct rx_header theader;
6162 struct sockaddr_in taddr;
6165 struct iovec tmpiov[2];
6168 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6171 tp = &tbuffer[sizeof(struct rx_header)];
6172 taddr.sin_family = AF_INET;
6173 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6174 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6175 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6176 taddr.sin_len = sizeof(struct sockaddr_in);
6178 memset(&theader, 0, sizeof(theader));
6179 theader.epoch = htonl(999);
6181 theader.callNumber = 0;
6184 theader.type = RX_PACKET_TYPE_VERSION;
6185 theader.flags = RX_LAST_PACKET;
6186 theader.serviceId = 0;
6188 memcpy(tbuffer, &theader, sizeof(theader));
6189 memcpy(tp, &a, sizeof(a));
6190 tmpiov[0].iov_base = tbuffer;
6191 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6193 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6195 MUTEX_ENTER(&conn->conn_data_lock);
6196 MUTEX_ENTER(&rx_refcnt_mutex);
6197 /* Only reschedule ourselves if the connection would not be destroyed */
6198 if (conn->refCount <= 1) {
6199 conn->natKeepAliveEvent = NULL;
6200 MUTEX_EXIT(&rx_refcnt_mutex);
6201 MUTEX_EXIT(&conn->conn_data_lock);
6202 rx_DestroyConnection(conn); /* drop the reference for this */
6204 conn->refCount--; /* drop the reference for this */
6205 MUTEX_EXIT(&rx_refcnt_mutex);
6206 conn->natKeepAliveEvent = NULL;
6207 rxi_ScheduleNatKeepAliveEvent(conn);
6208 MUTEX_EXIT(&conn->conn_data_lock);
6213 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6215 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6216 struct clock when, now;
6217 clock_GetTime(&now);
6219 when.sec += conn->secondsUntilNatPing;
6220 MUTEX_ENTER(&rx_refcnt_mutex);
6221 conn->refCount++; /* hold a reference for this */
6222 MUTEX_EXIT(&rx_refcnt_mutex);
6223 conn->natKeepAliveEvent =
6224 rxevent_PostNow(&when, &now, rxi_NatKeepAliveEvent, conn, 0);
6229 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6231 MUTEX_ENTER(&conn->conn_data_lock);
6232 conn->secondsUntilNatPing = seconds;
6234 rxi_ScheduleNatKeepAliveEvent(conn);
6235 MUTEX_EXIT(&conn->conn_data_lock);
6239 rxi_NatKeepAliveOn(struct rx_connection *conn)
6241 MUTEX_ENTER(&conn->conn_data_lock);
6242 rxi_ScheduleNatKeepAliveEvent(conn);
6243 MUTEX_EXIT(&conn->conn_data_lock);
6246 /* When a call is in progress, this routine is called occasionally to
6247 * make sure that some traffic has arrived (or been sent to) the peer.
6248 * If nothing has arrived in a reasonable amount of time, the call is
6249 * declared dead; if nothing has been sent for a while, we send a
6250 * keep-alive packet (if we're actually trying to keep the call alive)
6253 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6255 struct rx_call *call = arg1;
6256 struct rx_connection *conn;
6259 MUTEX_ENTER(&rx_refcnt_mutex);
6260 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6261 MUTEX_EXIT(&rx_refcnt_mutex);
6262 MUTEX_ENTER(&call->lock);
6263 if (event == call->keepAliveEvent)
6264 call->keepAliveEvent = NULL;
6267 #ifdef RX_ENABLE_LOCKS
6268 if (rxi_CheckCall(call, 0)) {
6269 MUTEX_EXIT(&call->lock);
6272 #else /* RX_ENABLE_LOCKS */
6273 if (rxi_CheckCall(call))
6275 #endif /* RX_ENABLE_LOCKS */
6277 /* Don't try to keep alive dallying calls */
6278 if (call->state == RX_STATE_DALLY) {
6279 MUTEX_EXIT(&call->lock);
6284 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6285 /* Don't try to send keepalives if there is unacknowledged data */
6286 /* the rexmit code should be good enough, this little hack
6287 * doesn't quite work XXX */
6288 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6290 rxi_ScheduleKeepAliveEvent(call);
6291 MUTEX_EXIT(&call->lock);
6294 /* Does what's on the nameplate. */
6296 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy)
6298 struct rx_call *call = arg1;
6299 struct rx_connection *conn;
6301 MUTEX_ENTER(&rx_refcnt_mutex);
6302 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6303 MUTEX_EXIT(&rx_refcnt_mutex);
6304 MUTEX_ENTER(&call->lock);
6306 if (event == call->growMTUEvent)
6307 call->growMTUEvent = NULL;
6309 #ifdef RX_ENABLE_LOCKS
6310 if (rxi_CheckCall(call, 0)) {
6311 MUTEX_EXIT(&call->lock);
6314 #else /* RX_ENABLE_LOCKS */
6315 if (rxi_CheckCall(call))
6317 #endif /* RX_ENABLE_LOCKS */
6319 /* Don't bother with dallying calls */
6320 if (call->state == RX_STATE_DALLY) {
6321 MUTEX_EXIT(&call->lock);
6328 * keep being scheduled, just don't do anything if we're at peak,
6329 * or we're not set up to be properly handled (idle timeout required)
6331 if ((conn->peer->maxPacketSize != 0) &&
6332 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6333 (conn->idleDeadErr))
6334 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6335 rxi_ScheduleGrowMTUEvent(call, 0);
6336 MUTEX_EXIT(&call->lock);
6340 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6342 if (!call->keepAliveEvent) {
6343 struct clock when, now;
6344 clock_GetTime(&now);
6346 when.sec += call->conn->secondsUntilPing;
6347 MUTEX_ENTER(&rx_refcnt_mutex);
6348 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6349 MUTEX_EXIT(&rx_refcnt_mutex);
6350 call->keepAliveEvent =
6351 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
6356 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6358 if (!call->growMTUEvent) {
6359 struct clock when, now;
6361 clock_GetTime(&now);
6364 if (call->conn->secondsUntilPing)
6365 secs = (6*call->conn->secondsUntilPing)-1;
6367 if (call->conn->secondsUntilDead)
6368 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6372 MUTEX_ENTER(&rx_refcnt_mutex);
6373 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6374 MUTEX_EXIT(&rx_refcnt_mutex);
6375 call->growMTUEvent =
6376 rxevent_PostNow(&when, &now, rxi_GrowMTUEvent, call, 0);
6380 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6382 rxi_KeepAliveOn(struct rx_call *call)
6384 /* Pretend last packet received was received now--i.e. if another
6385 * packet isn't received within the keep alive time, then the call
6386 * will die; Initialize last send time to the current time--even
6387 * if a packet hasn't been sent yet. This will guarantee that a
6388 * keep-alive is sent within the ping time */
6389 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6390 rxi_ScheduleKeepAliveEvent(call);
6394 rxi_GrowMTUOn(struct rx_call *call)
6396 struct rx_connection *conn = call->conn;
6397 MUTEX_ENTER(&conn->conn_data_lock);
6398 conn->lastPingSizeSer = conn->lastPingSize = 0;
6399 MUTEX_EXIT(&conn->conn_data_lock);
6400 rxi_ScheduleGrowMTUEvent(call, 1);
6403 /* This routine is called to send connection abort messages
6404 * that have been delayed to throttle looping clients. */
6406 rxi_SendDelayedConnAbort(struct rxevent *event,
6407 void *arg1, void *unused)
6409 struct rx_connection *conn = arg1;
6412 struct rx_packet *packet;
6414 MUTEX_ENTER(&conn->conn_data_lock);
6415 conn->delayedAbortEvent = NULL;
6416 error = htonl(conn->error);
6418 MUTEX_EXIT(&conn->conn_data_lock);
6419 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6422 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6423 RX_PACKET_TYPE_ABORT, (char *)&error,
6425 rxi_FreePacket(packet);
6429 /* This routine is called to send call abort messages
6430 * that have been delayed to throttle looping clients. */
6432 rxi_SendDelayedCallAbort(struct rxevent *event,
6433 void *arg1, void *dummy)
6435 struct rx_call *call = arg1;
6438 struct rx_packet *packet;
6440 MUTEX_ENTER(&call->lock);
6441 call->delayedAbortEvent = NULL;
6442 error = htonl(call->error);
6444 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6447 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6448 (char *)&error, sizeof(error), 0);
6449 rxi_FreePacket(packet);
6451 MUTEX_EXIT(&call->lock);
6452 MUTEX_ENTER(&rx_refcnt_mutex);
6453 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6454 MUTEX_EXIT(&rx_refcnt_mutex);
6457 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6458 * seconds) to ask the client to authenticate itself. The routine
6459 * issues a challenge to the client, which is obtained from the
6460 * security object associated with the connection */
6462 rxi_ChallengeEvent(struct rxevent *event,
6463 void *arg0, void *arg1, int tries)
6465 struct rx_connection *conn = arg0;
6467 conn->challengeEvent = NULL;
6468 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6469 struct rx_packet *packet;
6470 struct clock when, now;
6473 /* We've failed to authenticate for too long.
6474 * Reset any calls waiting for authentication;
6475 * they are all in RX_STATE_PRECALL.
6479 MUTEX_ENTER(&conn->conn_call_lock);
6480 for (i = 0; i < RX_MAXCALLS; i++) {
6481 struct rx_call *call = conn->call[i];
6483 MUTEX_ENTER(&call->lock);
6484 if (call->state == RX_STATE_PRECALL) {
6485 rxi_CallError(call, RX_CALL_DEAD);
6486 rxi_SendCallAbort(call, NULL, 0, 0);
6488 MUTEX_EXIT(&call->lock);
6491 MUTEX_EXIT(&conn->conn_call_lock);
6495 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6497 /* If there's no packet available, do this later. */
6498 RXS_GetChallenge(conn->securityObject, conn, packet);
6499 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6500 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6501 rxi_FreePacket(packet);
6503 clock_GetTime(&now);
6505 when.sec += RX_CHALLENGE_TIMEOUT;
6506 conn->challengeEvent =
6507 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
6512 /* Call this routine to start requesting the client to authenticate
6513 * itself. This will continue until authentication is established,
6514 * the call times out, or an invalid response is returned. The
6515 * security object associated with the connection is asked to create
6516 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6517 * defined earlier. */
6519 rxi_ChallengeOn(struct rx_connection *conn)
6521 if (!conn->challengeEvent) {
6522 RXS_CreateChallenge(conn->securityObject, conn);
6523 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6528 /* rxi_ComputeRoundTripTime is called with peer locked. */
6529 /* peer may be null */
6531 rxi_ComputeRoundTripTime(struct rx_packet *p,
6532 struct rx_ackPacket *ack,
6533 struct rx_peer *peer,
6536 struct clock thisRtt, *sentp;
6540 /* If the ACK is delayed, then do nothing */
6541 if (ack->reason == RX_ACK_DELAY)
6544 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6545 * their RTT multiple times, so only include the RTT of the last packet
6547 if (p->flags & RX_JUMBO_PACKET)
6550 /* Use the serial number to determine which transmission the ACK is for,
6551 * and set the sent time to match this. If we have no serial number, then
6552 * only use the ACK for RTT calculations if the packet has not been
6556 serial = ntohl(ack->serial);
6558 if (serial == p->header.serial) {
6559 sentp = &p->timeSent;
6560 } else if (serial == p->firstSerial) {
6561 sentp = &p->firstSent;
6562 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6563 sentp = &p->firstSent;
6567 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6568 sentp = &p->firstSent;
6575 if (clock_Lt(&thisRtt, sentp))
6576 return; /* somebody set the clock back, don't count this time. */
6578 clock_Sub(&thisRtt, sentp);
6579 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6580 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6582 if (clock_IsZero(&thisRtt)) {
6584 * The actual round trip time is shorter than the
6585 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6586 * Since we can't tell which at the moment we will assume 1ms.
6588 thisRtt.usec = 1000;
6591 if (rx_stats_active) {
6592 MUTEX_ENTER(&rx_stats_mutex);
6593 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6594 rx_stats.minRtt = thisRtt;
6595 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6596 if (thisRtt.sec > 60) {
6597 MUTEX_EXIT(&rx_stats_mutex);
6598 return; /* somebody set the clock ahead */
6600 rx_stats.maxRtt = thisRtt;
6602 clock_Add(&rx_stats.totalRtt, &thisRtt);
6603 rx_atomic_inc(&rx_stats.nRttSamples);
6604 MUTEX_EXIT(&rx_stats_mutex);
6607 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6609 /* Apply VanJacobson round-trip estimations */
6614 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
6615 * srtt is stored as fixed point with 3 bits after the binary
6616 * point (i.e., scaled by 8). The following magic is
6617 * equivalent to the smoothing algorithm in rfc793 with an
6618 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6619 * srtt'*8 = rtt + srtt*7
6620 * srtt'*8 = srtt*8 + rtt - srtt
6621 * srtt' = srtt + rtt/8 - srtt/8
6622 * srtt' = srtt + (rtt - srtt)/8
6625 delta = _8THMSEC(&thisRtt) - peer->rtt;
6626 peer->rtt += (delta >> 3);
6629 * We accumulate a smoothed rtt variance (actually, a smoothed
6630 * mean difference), then set the retransmit timer to smoothed
6631 * rtt + 4 times the smoothed variance (was 2x in van's original
6632 * paper, but 4x works better for me, and apparently for him as
6634 * rttvar is stored as
6635 * fixed point with 2 bits after the binary point (scaled by
6636 * 4). The following is equivalent to rfc793 smoothing with
6637 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6638 * rttvar'*4 = rttvar*3 + |delta|
6639 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6640 * rttvar' = rttvar + |delta|/4 - rttvar/4
6641 * rttvar' = rttvar + (|delta| - rttvar)/4
6642 * This replaces rfc793's wired-in beta.
6643 * dev*4 = dev*4 + (|actual - expected| - dev)
6649 delta -= (peer->rtt_dev << 1);
6650 peer->rtt_dev += (delta >> 3);
6652 /* I don't have a stored RTT so I start with this value. Since I'm
6653 * probably just starting a call, and will be pushing more data down
6654 * this, I expect congestion to increase rapidly. So I fudge a
6655 * little, and I set deviance to half the rtt. In practice,
6656 * deviance tends to approach something a little less than
6657 * half the smoothed rtt. */
6658 peer->rtt = _8THMSEC(&thisRtt) + 8;
6659 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
6661 /* the timeout is RTT + 4*MDEV + rx_minPeerTimeout msec.
6662 * This is because one end or the other of these connections is usually
6663 * in a user process, and can be switched and/or swapped out. So on fast,
6664 * reliable networks, the timeout would otherwise be too short. */
6665 rtt_timeout = ((peer->rtt >> 3) + peer->rtt_dev) + rx_minPeerTimeout;
6666 clock_Zero(&(peer->timeout));
6667 clock_Addmsec(&(peer->timeout), rtt_timeout);
6669 /* Reset the backedOff flag since we just computed a new timeout value */
6670 peer->backedOff = 0;
6672 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6673 p->header.callNumber, p, MSEC(&thisRtt), peer->rtt >> 3, peer->rtt_dev >> 2, (peer->timeout.sec), (peer->timeout.usec)));
6677 /* Find all server connections that have not been active for a long time, and
6680 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6682 struct clock now, when;
6683 clock_GetTime(&now);
6685 /* Find server connection structures that haven't been used for
6686 * greater than rx_idleConnectionTime */
6688 struct rx_connection **conn_ptr, **conn_end;
6689 int i, havecalls = 0;
6690 MUTEX_ENTER(&rx_connHashTable_lock);
6691 for (conn_ptr = &rx_connHashTable[0], conn_end =
6692 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6694 struct rx_connection *conn, *next;
6695 struct rx_call *call;
6699 for (conn = *conn_ptr; conn; conn = next) {
6700 /* XXX -- Shouldn't the connection be locked? */
6703 for (i = 0; i < RX_MAXCALLS; i++) {
6704 call = conn->call[i];
6708 code = MUTEX_TRYENTER(&call->lock);
6711 #ifdef RX_ENABLE_LOCKS
6712 result = rxi_CheckCall(call, 1);
6713 #else /* RX_ENABLE_LOCKS */
6714 result = rxi_CheckCall(call);
6715 #endif /* RX_ENABLE_LOCKS */
6716 MUTEX_EXIT(&call->lock);
6718 /* If CheckCall freed the call, it might
6719 * have destroyed the connection as well,
6720 * which screws up the linked lists.
6726 if (conn->type == RX_SERVER_CONNECTION) {
6727 /* This only actually destroys the connection if
6728 * there are no outstanding calls */
6729 MUTEX_ENTER(&conn->conn_data_lock);
6730 MUTEX_ENTER(&rx_refcnt_mutex);
6731 if (!havecalls && !conn->refCount
6732 && ((conn->lastSendTime + rx_idleConnectionTime) <
6734 conn->refCount++; /* it will be decr in rx_DestroyConn */
6735 MUTEX_EXIT(&rx_refcnt_mutex);
6736 MUTEX_EXIT(&conn->conn_data_lock);
6737 #ifdef RX_ENABLE_LOCKS
6738 rxi_DestroyConnectionNoLock(conn);
6739 #else /* RX_ENABLE_LOCKS */
6740 rxi_DestroyConnection(conn);
6741 #endif /* RX_ENABLE_LOCKS */
6743 #ifdef RX_ENABLE_LOCKS
6745 MUTEX_EXIT(&rx_refcnt_mutex);
6746 MUTEX_EXIT(&conn->conn_data_lock);
6748 #endif /* RX_ENABLE_LOCKS */
6752 #ifdef RX_ENABLE_LOCKS
6753 while (rx_connCleanup_list) {
6754 struct rx_connection *conn;
6755 conn = rx_connCleanup_list;
6756 rx_connCleanup_list = rx_connCleanup_list->next;
6757 MUTEX_EXIT(&rx_connHashTable_lock);
6758 rxi_CleanupConnection(conn);
6759 MUTEX_ENTER(&rx_connHashTable_lock);
6761 MUTEX_EXIT(&rx_connHashTable_lock);
6762 #endif /* RX_ENABLE_LOCKS */
6765 /* Find any peer structures that haven't been used (haven't had an
6766 * associated connection) for greater than rx_idlePeerTime */
6768 struct rx_peer **peer_ptr, **peer_end;
6772 * Why do we need to hold the rx_peerHashTable_lock across
6773 * the incrementing of peer_ptr since the rx_peerHashTable
6774 * array is not changing? We don't.
6776 * By dropping the lock periodically we can permit other
6777 * activities to be performed while a rxi_ReapConnections
6778 * call is in progress. The goal of reap connections
6779 * is to clean up quickly without causing large amounts
6780 * of contention. Therefore, it is important that global
6781 * mutexes not be held for extended periods of time.
6783 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6784 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6786 struct rx_peer *peer, *next, *prev;
6788 MUTEX_ENTER(&rx_peerHashTable_lock);
6789 for (prev = peer = *peer_ptr; peer; peer = next) {
6791 code = MUTEX_TRYENTER(&peer->peer_lock);
6792 if ((code) && (peer->refCount == 0)
6793 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6794 rx_interface_stat_p rpc_stat, nrpc_stat;
6798 * now know that this peer object is one to be
6799 * removed from the hash table. Once it is removed
6800 * it can't be referenced by other threads.
6801 * Lets remove it first and decrement the struct
6802 * nPeerStructs count.
6804 if (peer == *peer_ptr) {
6810 if (rx_stats_active)
6811 rx_atomic_dec(&rx_stats.nPeerStructs);
6814 * Now if we hold references on 'prev' and 'next'
6815 * we can safely drop the rx_peerHashTable_lock
6816 * while we destroy this 'peer' object.
6822 MUTEX_EXIT(&rx_peerHashTable_lock);
6824 MUTEX_EXIT(&peer->peer_lock);
6825 MUTEX_DESTROY(&peer->peer_lock);
6827 (&peer->rpcStats, rpc_stat, nrpc_stat,
6828 rx_interface_stat)) {
6829 unsigned int num_funcs;
6832 queue_Remove(&rpc_stat->queue_header);
6833 queue_Remove(&rpc_stat->all_peers);
6834 num_funcs = rpc_stat->stats[0].func_total;
6836 sizeof(rx_interface_stat_t) +
6837 rpc_stat->stats[0].func_total *
6838 sizeof(rx_function_entry_v1_t);
6840 rxi_Free(rpc_stat, space);
6842 MUTEX_ENTER(&rx_rpc_stats);
6843 rxi_rpc_peer_stat_cnt -= num_funcs;
6844 MUTEX_EXIT(&rx_rpc_stats);
6849 * Regain the rx_peerHashTable_lock and
6850 * decrement the reference count on 'prev'
6853 MUTEX_ENTER(&rx_peerHashTable_lock);
6860 MUTEX_EXIT(&peer->peer_lock);
6865 MUTEX_EXIT(&rx_peerHashTable_lock);
6869 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6870 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6871 * GC, just below. Really, we shouldn't have to keep moving packets from
6872 * one place to another, but instead ought to always know if we can
6873 * afford to hold onto a packet in its particular use. */
6874 MUTEX_ENTER(&rx_freePktQ_lock);
6875 if (rx_waitingForPackets) {
6876 rx_waitingForPackets = 0;
6877 #ifdef RX_ENABLE_LOCKS
6878 CV_BROADCAST(&rx_waitingForPackets_cv);
6880 osi_rxWakeup(&rx_waitingForPackets);
6883 MUTEX_EXIT(&rx_freePktQ_lock);
6886 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6887 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6891 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6892 * rx.h is sort of strange this is better. This is called with a security
6893 * object before it is discarded. Each connection using a security object has
6894 * its own refcount to the object so it won't actually be freed until the last
6895 * connection is destroyed.
6897 * This is the only rxs module call. A hold could also be written but no one
6901 rxs_Release(struct rx_securityClass *aobj)
6903 return RXS_Close(aobj);
6907 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6908 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6909 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6910 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6912 /* Adjust our estimate of the transmission rate to this peer, given
6913 * that the packet p was just acked. We can adjust peer->timeout and
6914 * call->twind. Pragmatically, this is called
6915 * only with packets of maximal length.
6916 * Called with peer and call locked.
6920 rxi_ComputeRate(struct rx_peer *peer, struct rx_call *call,
6921 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6923 afs_int32 xferSize, xferMs;
6927 /* Count down packets */
6928 if (peer->rateFlag > 0)
6930 /* Do nothing until we're enabled */
6931 if (peer->rateFlag != 0)
6936 /* Count only when the ack seems legitimate */
6937 switch (ackReason) {
6938 case RX_ACK_REQUESTED:
6940 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6944 case RX_ACK_PING_RESPONSE:
6945 if (p) /* want the response to ping-request, not data send */
6947 clock_GetTime(&newTO);
6948 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6949 clock_Sub(&newTO, &call->pingRequestTime);
6950 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6954 xferSize = rx_AckDataSize(rx_maxSendWindow) + RX_HEADER_SIZE;
6961 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %d.%06d, rtt %u, ps %u)\n",
6962 ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"),
6963 xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
6965 /* Track only packets that are big enough. */
6966 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6970 /* absorb RTT data (in milliseconds) for these big packets */
6971 if (peer->smRtt == 0) {
6972 peer->smRtt = xferMs;
6974 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6979 if (peer->countDown) {
6983 peer->countDown = 10; /* recalculate only every so often */
6985 /* In practice, we can measure only the RTT for full packets,
6986 * because of the way Rx acks the data that it receives. (If it's
6987 * smaller than a full packet, it often gets implicitly acked
6988 * either by the call response (from a server) or by the next call
6989 * (from a client), and either case confuses transmission times
6990 * with processing times.) Therefore, replace the above
6991 * more-sophisticated processing with a simpler version, where the
6992 * smoothed RTT is kept for full-size packets, and the time to
6993 * transmit a windowful of full-size packets is simply RTT *
6994 * windowSize. Again, we take two steps:
6995 - ensure the timeout is large enough for a single packet's RTT;
6996 - ensure that the window is small enough to fit in the desired timeout.*/
6998 /* First, the timeout check. */
6999 minTime = peer->smRtt;
7000 /* Get a reasonable estimate for a timeout period */
7002 newTO.sec = minTime / 1000;
7003 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
7005 /* Increase the timeout period so that we can always do at least
7006 * one packet exchange */
7007 if (clock_Gt(&newTO, &peer->timeout)) {
7009 dpf(("CONG peer %lx/%u: timeout %d.%06d ==> %ld.%06d (rtt %u)\n",
7010 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec,
7011 newTO.sec, newTO.usec, peer->smRtt));
7013 peer->timeout = newTO;
7016 /* Now, get an estimate for the transmit window size. */
7017 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
7018 /* Now, convert to the number of full packets that could fit in a
7019 * reasonable fraction of that interval */
7020 minTime /= (peer->smRtt << 1);
7021 minTime = MAX(minTime, rx_minPeerTimeout);
7022 xferSize = minTime; /* (make a copy) */
7024 /* Now clamp the size to reasonable bounds. */
7027 else if (minTime > rx_maxSendWindow)
7028 minTime = rx_maxSendWindow;
7029 /* if (minTime != peer->maxWindow) {
7030 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u)\n",
7031 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
7032 peer->timeout.sec, peer->timeout.usec, peer->smRtt));
7033 peer->maxWindow = minTime;
7034 elide... call->twind = minTime;
7038 /* Cut back on the peer timeout if it had earlier grown unreasonably.
7039 * Discern this by calculating the timeout necessary for rx_Window
7041 if ((xferSize > rx_maxSendWindow) && (peer->timeout.sec >= 3)) {
7042 /* calculate estimate for transmission interval in milliseconds */
7043 minTime = rx_maxSendWindow * peer->smRtt;
7044 if (minTime < 1000) {
7045 dpf(("CONG peer %lx/%u: cut TO %d.%06d by 0.5 (rtt %u)\n",
7046 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
7047 peer->timeout.usec, peer->smRtt));
7049 newTO.sec = 0; /* cut back on timeout by half a second */
7050 newTO.usec = 500000;
7051 clock_Sub(&peer->timeout, &newTO);
7056 } /* end of rxi_ComputeRate */
7057 #endif /* ADAPT_WINDOW */
7065 #define TRACE_OPTION_RX_DEBUG 16
7073 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7074 0, KEY_QUERY_VALUE, &parmKey);
7075 if (code != ERROR_SUCCESS)
7078 dummyLen = sizeof(TraceOption);
7079 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7080 (BYTE *) &TraceOption, &dummyLen);
7081 if (code == ERROR_SUCCESS) {
7082 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7084 RegCloseKey (parmKey);
7085 #endif /* AFS_NT40_ENV */
7090 rx_DebugOnOff(int on)
7094 rxdebug_active = on;
7100 rx_StatsOnOff(int on)
7102 rx_stats_active = on;
7106 /* Don't call this debugging routine directly; use dpf */
7108 rxi_DebugPrint(char *format, ...)
7117 va_start(ap, format);
7119 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7122 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7124 OutputDebugString(msg);
7130 va_start(ap, format);
7132 clock_GetTime(&now);
7133 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7134 (unsigned int)now.usec);
7135 vfprintf(rx_Log, format, ap);
7143 * This function is used to process the rx_stats structure that is local
7144 * to a process as well as an rx_stats structure received from a remote
7145 * process (via rxdebug). Therefore, it needs to do minimal version
7149 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7150 afs_int32 freePackets, char version)
7154 if (size != sizeof(struct rx_statistics)) {
7156 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7157 size, sizeof(struct rx_statistics));
7160 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7163 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7164 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7165 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7166 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7167 s->specialPktAllocFailures);
7169 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7170 s->receivePktAllocFailures, s->sendPktAllocFailures,
7171 s->specialPktAllocFailures);
7175 " greedy %u, " "bogusReads %u (last from host %x), "
7176 "noPackets %u, " "noBuffers %u, " "selects %u, "
7177 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7178 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7179 s->selects, s->sendSelects);
7181 fprintf(file, " packets read: ");
7182 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7183 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7185 fprintf(file, "\n");
7188 " other read counters: data %u, " "ack %u, " "dup %u "
7189 "spurious %u " "dally %u\n", s->dataPacketsRead,
7190 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7191 s->ignorePacketDally);
7193 fprintf(file, " packets sent: ");
7194 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7195 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7197 fprintf(file, "\n");
7200 " other send counters: ack %u, " "data %u (not resends), "
7201 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7202 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7203 s->dataPacketsPushed, s->ignoreAckedPacket);
7206 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7207 s->netSendFailures, (int)s->fatalErrors);
7209 if (s->nRttSamples) {
7210 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7211 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7213 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7214 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7218 " %d server connections, " "%d client connections, "
7219 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7220 s->nServerConns, s->nClientConns, s->nPeerStructs,
7221 s->nCallStructs, s->nFreeCallStructs);
7223 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7224 fprintf(file, " %d clock updates\n", clock_nUpdates);
7228 /* for backward compatibility */
7230 rx_PrintStats(FILE * file)
7232 MUTEX_ENTER(&rx_stats_mutex);
7233 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7234 sizeof(rx_stats), rx_nFreePackets,
7236 MUTEX_EXIT(&rx_stats_mutex);
7240 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7242 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
7243 ntohl(peer->host), (int)ntohs(peer->port), (int)peer->burstSize,
7244 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
7247 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
7248 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
7249 (int)peer->timeout.usec, peer->nSent, peer->reSends);
7252 " Packet size %d, " "max in packet skew %d, "
7253 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
7254 (int)peer->outPacketSkew);
7258 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7260 * This mutex protects the following static variables:
7264 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7265 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7267 #define LOCK_RX_DEBUG
7268 #define UNLOCK_RX_DEBUG
7269 #endif /* AFS_PTHREAD_ENV */
7271 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7273 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7274 u_char type, void *inputData, size_t inputLength,
7275 void *outputData, size_t outputLength)
7277 static afs_int32 counter = 100;
7278 time_t waitTime, waitCount;
7279 struct rx_header theader;
7282 struct timeval tv_now, tv_wake, tv_delta;
7283 struct sockaddr_in taddr, faddr;
7297 tp = &tbuffer[sizeof(struct rx_header)];
7298 taddr.sin_family = AF_INET;
7299 taddr.sin_port = remotePort;
7300 taddr.sin_addr.s_addr = remoteAddr;
7301 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7302 taddr.sin_len = sizeof(struct sockaddr_in);
7305 memset(&theader, 0, sizeof(theader));
7306 theader.epoch = htonl(999);
7308 theader.callNumber = htonl(counter);
7311 theader.type = type;
7312 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7313 theader.serviceId = 0;
7315 memcpy(tbuffer, &theader, sizeof(theader));
7316 memcpy(tp, inputData, inputLength);
7318 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7319 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7321 /* see if there's a packet available */
7322 gettimeofday(&tv_wake,0);
7323 tv_wake.tv_sec += waitTime;
7326 FD_SET(socket, &imask);
7327 tv_delta.tv_sec = tv_wake.tv_sec;
7328 tv_delta.tv_usec = tv_wake.tv_usec;
7329 gettimeofday(&tv_now, 0);
7331 if (tv_delta.tv_usec < tv_now.tv_usec) {
7333 tv_delta.tv_usec += 1000000;
7336 tv_delta.tv_usec -= tv_now.tv_usec;
7338 if (tv_delta.tv_sec < tv_now.tv_sec) {
7342 tv_delta.tv_sec -= tv_now.tv_sec;
7345 code = select(0, &imask, 0, 0, &tv_delta);
7346 #else /* AFS_NT40_ENV */
7347 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7348 #endif /* AFS_NT40_ENV */
7349 if (code == 1 && FD_ISSET(socket, &imask)) {
7350 /* now receive a packet */
7351 faddrLen = sizeof(struct sockaddr_in);
7353 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7354 (struct sockaddr *)&faddr, &faddrLen);
7357 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7358 if (counter == ntohl(theader.callNumber))
7366 /* see if we've timed out */
7374 code -= sizeof(struct rx_header);
7375 if (code > outputLength)
7376 code = outputLength;
7377 memcpy(outputData, tp, code);
7380 #endif /* RXDEBUG */
7383 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7384 afs_uint16 remotePort, struct rx_debugStats * stat,
7385 afs_uint32 * supportedValues)
7387 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7389 struct rx_debugIn in;
7391 *supportedValues = 0;
7392 in.type = htonl(RX_DEBUGI_GETSTATS);
7395 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7396 &in, sizeof(in), stat, sizeof(*stat));
7399 * If the call was successful, fixup the version and indicate
7400 * what contents of the stat structure are valid.
7401 * Also do net to host conversion of fields here.
7405 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7406 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7408 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7409 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7411 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7412 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7414 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7415 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7417 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7418 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7420 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7421 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7423 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7424 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7426 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7427 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7429 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7430 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7432 stat->nFreePackets = ntohl(stat->nFreePackets);
7433 stat->packetReclaims = ntohl(stat->packetReclaims);
7434 stat->callsExecuted = ntohl(stat->callsExecuted);
7435 stat->nWaiting = ntohl(stat->nWaiting);
7436 stat->idleThreads = ntohl(stat->idleThreads);
7437 stat->nWaited = ntohl(stat->nWaited);
7438 stat->nPackets = ntohl(stat->nPackets);
7447 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7448 afs_uint16 remotePort, struct rx_statistics * stat,
7449 afs_uint32 * supportedValues)
7451 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7453 struct rx_debugIn in;
7454 afs_int32 *lp = (afs_int32 *) stat;
7458 * supportedValues is currently unused, but added to allow future
7459 * versioning of this function.
7462 *supportedValues = 0;
7463 in.type = htonl(RX_DEBUGI_RXSTATS);
7465 memset(stat, 0, sizeof(*stat));
7467 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7468 &in, sizeof(in), stat, sizeof(*stat));
7473 * Do net to host conversion here
7476 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7487 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7488 afs_uint16 remotePort, size_t version_length,
7491 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7493 return MakeDebugCall(socket, remoteAddr, remotePort,
7494 RX_PACKET_TYPE_VERSION, a, 1, version,
7502 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7503 afs_uint16 remotePort, afs_int32 * nextConnection,
7504 int allConnections, afs_uint32 debugSupportedValues,
7505 struct rx_debugConn * conn,
7506 afs_uint32 * supportedValues)
7508 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7510 struct rx_debugIn in;
7514 * supportedValues is currently unused, but added to allow future
7515 * versioning of this function.
7518 *supportedValues = 0;
7519 if (allConnections) {
7520 in.type = htonl(RX_DEBUGI_GETALLCONN);
7522 in.type = htonl(RX_DEBUGI_GETCONN);
7524 in.index = htonl(*nextConnection);
7525 memset(conn, 0, sizeof(*conn));
7527 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7528 &in, sizeof(in), conn, sizeof(*conn));
7531 *nextConnection += 1;
7534 * Convert old connection format to new structure.
7537 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7538 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7539 #define MOVEvL(a) (conn->a = vL->a)
7541 /* any old or unrecognized version... */
7542 for (i = 0; i < RX_MAXCALLS; i++) {
7543 MOVEvL(callState[i]);
7544 MOVEvL(callMode[i]);
7545 MOVEvL(callFlags[i]);
7546 MOVEvL(callOther[i]);
7548 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7549 MOVEvL(secStats.type);
7550 MOVEvL(secStats.level);
7551 MOVEvL(secStats.flags);
7552 MOVEvL(secStats.expires);
7553 MOVEvL(secStats.packetsReceived);
7554 MOVEvL(secStats.packetsSent);
7555 MOVEvL(secStats.bytesReceived);
7556 MOVEvL(secStats.bytesSent);
7561 * Do net to host conversion here
7563 * I don't convert host or port since we are most likely
7564 * going to want these in NBO.
7566 conn->cid = ntohl(conn->cid);
7567 conn->serial = ntohl(conn->serial);
7568 for (i = 0; i < RX_MAXCALLS; i++) {
7569 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7571 conn->error = ntohl(conn->error);
7572 conn->secStats.flags = ntohl(conn->secStats.flags);
7573 conn->secStats.expires = ntohl(conn->secStats.expires);
7574 conn->secStats.packetsReceived =
7575 ntohl(conn->secStats.packetsReceived);
7576 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7577 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7578 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7579 conn->epoch = ntohl(conn->epoch);
7580 conn->natMTU = ntohl(conn->natMTU);
7589 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7590 afs_uint16 remotePort, afs_int32 * nextPeer,
7591 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7592 afs_uint32 * supportedValues)
7594 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7596 struct rx_debugIn in;
7599 * supportedValues is currently unused, but added to allow future
7600 * versioning of this function.
7603 *supportedValues = 0;
7604 in.type = htonl(RX_DEBUGI_GETPEER);
7605 in.index = htonl(*nextPeer);
7606 memset(peer, 0, sizeof(*peer));
7608 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7609 &in, sizeof(in), peer, sizeof(*peer));
7615 * Do net to host conversion here
7617 * I don't convert host or port since we are most likely
7618 * going to want these in NBO.
7620 peer->ifMTU = ntohs(peer->ifMTU);
7621 peer->idleWhen = ntohl(peer->idleWhen);
7622 peer->refCount = ntohs(peer->refCount);
7623 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7624 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7625 peer->rtt = ntohl(peer->rtt);
7626 peer->rtt_dev = ntohl(peer->rtt_dev);
7627 peer->timeout.sec = ntohl(peer->timeout.sec);
7628 peer->timeout.usec = ntohl(peer->timeout.usec);
7629 peer->nSent = ntohl(peer->nSent);
7630 peer->reSends = ntohl(peer->reSends);
7631 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7632 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7633 peer->rateFlag = ntohl(peer->rateFlag);
7634 peer->natMTU = ntohs(peer->natMTU);
7635 peer->maxMTU = ntohs(peer->maxMTU);
7636 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7637 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7638 peer->MTU = ntohs(peer->MTU);
7639 peer->cwind = ntohs(peer->cwind);
7640 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7641 peer->congestSeq = ntohs(peer->congestSeq);
7642 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7643 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7644 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7645 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7654 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7655 struct rx_debugPeer * peerStats)
7658 afs_int32 error = 1; /* default to "did not succeed" */
7659 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7661 MUTEX_ENTER(&rx_peerHashTable_lock);
7662 for(tp = rx_peerHashTable[hashValue];
7663 tp != NULL; tp = tp->next) {
7664 if (tp->host == peerHost)
7670 MUTEX_EXIT(&rx_peerHashTable_lock);
7674 MUTEX_ENTER(&tp->peer_lock);
7675 peerStats->host = tp->host;
7676 peerStats->port = tp->port;
7677 peerStats->ifMTU = tp->ifMTU;
7678 peerStats->idleWhen = tp->idleWhen;
7679 peerStats->refCount = tp->refCount;
7680 peerStats->burstSize = tp->burstSize;
7681 peerStats->burst = tp->burst;
7682 peerStats->burstWait.sec = tp->burstWait.sec;
7683 peerStats->burstWait.usec = tp->burstWait.usec;
7684 peerStats->rtt = tp->rtt;
7685 peerStats->rtt_dev = tp->rtt_dev;
7686 peerStats->timeout.sec = tp->timeout.sec;
7687 peerStats->timeout.usec = tp->timeout.usec;
7688 peerStats->nSent = tp->nSent;
7689 peerStats->reSends = tp->reSends;
7690 peerStats->inPacketSkew = tp->inPacketSkew;
7691 peerStats->outPacketSkew = tp->outPacketSkew;
7692 peerStats->rateFlag = tp->rateFlag;
7693 peerStats->natMTU = tp->natMTU;
7694 peerStats->maxMTU = tp->maxMTU;
7695 peerStats->maxDgramPackets = tp->maxDgramPackets;
7696 peerStats->ifDgramPackets = tp->ifDgramPackets;
7697 peerStats->MTU = tp->MTU;
7698 peerStats->cwind = tp->cwind;
7699 peerStats->nDgramPackets = tp->nDgramPackets;
7700 peerStats->congestSeq = tp->congestSeq;
7701 peerStats->bytesSent.high = tp->bytesSent.high;
7702 peerStats->bytesSent.low = tp->bytesSent.low;
7703 peerStats->bytesReceived.high = tp->bytesReceived.high;
7704 peerStats->bytesReceived.low = tp->bytesReceived.low;
7705 MUTEX_EXIT(&tp->peer_lock);
7707 MUTEX_ENTER(&rx_peerHashTable_lock);
7710 MUTEX_EXIT(&rx_peerHashTable_lock);
7718 struct rx_serverQueueEntry *np;
7721 struct rx_call *call;
7722 struct rx_serverQueueEntry *sq;
7726 if (rxinit_status == 1) {
7728 return; /* Already shutdown. */
7732 #ifndef AFS_PTHREAD_ENV
7733 FD_ZERO(&rx_selectMask);
7734 #endif /* AFS_PTHREAD_ENV */
7735 rxi_dataQuota = RX_MAX_QUOTA;
7736 #ifndef AFS_PTHREAD_ENV
7738 #endif /* AFS_PTHREAD_ENV */
7741 #ifndef AFS_PTHREAD_ENV
7742 #ifndef AFS_USE_GETTIMEOFDAY
7744 #endif /* AFS_USE_GETTIMEOFDAY */
7745 #endif /* AFS_PTHREAD_ENV */
7747 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7748 call = queue_First(&rx_freeCallQueue, rx_call);
7750 rxi_Free(call, sizeof(struct rx_call));
7753 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7754 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7760 struct rx_peer **peer_ptr, **peer_end;
7761 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7762 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7764 struct rx_peer *peer, *next;
7766 MUTEX_ENTER(&rx_peerHashTable_lock);
7767 for (peer = *peer_ptr; peer; peer = next) {
7768 rx_interface_stat_p rpc_stat, nrpc_stat;
7771 MUTEX_ENTER(&rx_rpc_stats);
7772 MUTEX_ENTER(&peer->peer_lock);
7774 (&peer->rpcStats, rpc_stat, nrpc_stat,
7775 rx_interface_stat)) {
7776 unsigned int num_funcs;
7779 queue_Remove(&rpc_stat->queue_header);
7780 queue_Remove(&rpc_stat->all_peers);
7781 num_funcs = rpc_stat->stats[0].func_total;
7783 sizeof(rx_interface_stat_t) +
7784 rpc_stat->stats[0].func_total *
7785 sizeof(rx_function_entry_v1_t);
7787 rxi_Free(rpc_stat, space);
7789 /* rx_rpc_stats must be held */
7790 rxi_rpc_peer_stat_cnt -= num_funcs;
7792 MUTEX_EXIT(&peer->peer_lock);
7793 MUTEX_EXIT(&rx_rpc_stats);
7797 if (rx_stats_active)
7798 rx_atomic_dec(&rx_stats.nPeerStructs);
7800 MUTEX_EXIT(&rx_peerHashTable_lock);
7803 for (i = 0; i < RX_MAX_SERVICES; i++) {
7805 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7807 for (i = 0; i < rx_hashTableSize; i++) {
7808 struct rx_connection *tc, *ntc;
7809 MUTEX_ENTER(&rx_connHashTable_lock);
7810 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7812 for (j = 0; j < RX_MAXCALLS; j++) {
7814 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7817 rxi_Free(tc, sizeof(*tc));
7819 MUTEX_EXIT(&rx_connHashTable_lock);
7822 MUTEX_ENTER(&freeSQEList_lock);
7824 while ((np = rx_FreeSQEList)) {
7825 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7826 MUTEX_DESTROY(&np->lock);
7827 rxi_Free(np, sizeof(*np));
7830 MUTEX_EXIT(&freeSQEList_lock);
7831 MUTEX_DESTROY(&freeSQEList_lock);
7832 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7833 MUTEX_DESTROY(&rx_connHashTable_lock);
7834 MUTEX_DESTROY(&rx_peerHashTable_lock);
7835 MUTEX_DESTROY(&rx_serverPool_lock);
7837 osi_Free(rx_connHashTable,
7838 rx_hashTableSize * sizeof(struct rx_connection *));
7839 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7841 UNPIN(rx_connHashTable,
7842 rx_hashTableSize * sizeof(struct rx_connection *));
7843 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7845 rxi_FreeAllPackets();
7847 MUTEX_ENTER(&rx_quota_mutex);
7848 rxi_dataQuota = RX_MAX_QUOTA;
7849 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7850 MUTEX_EXIT(&rx_quota_mutex);
7855 #ifdef RX_ENABLE_LOCKS
7857 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7859 if (!MUTEX_ISMINE(lockaddr))
7860 osi_Panic("Lock not held: %s", msg);
7862 #endif /* RX_ENABLE_LOCKS */
7867 * Routines to implement connection specific data.
7871 rx_KeyCreate(rx_destructor_t rtn)
7874 MUTEX_ENTER(&rxi_keyCreate_lock);
7875 key = rxi_keyCreate_counter++;
7876 rxi_keyCreate_destructor = (rx_destructor_t *)
7877 realloc((void *)rxi_keyCreate_destructor,
7878 (key + 1) * sizeof(rx_destructor_t));
7879 rxi_keyCreate_destructor[key] = rtn;
7880 MUTEX_EXIT(&rxi_keyCreate_lock);
7885 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7888 MUTEX_ENTER(&conn->conn_data_lock);
7889 if (!conn->specific) {
7890 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7891 for (i = 0; i < key; i++)
7892 conn->specific[i] = NULL;
7893 conn->nSpecific = key + 1;
7894 conn->specific[key] = ptr;
7895 } else if (key >= conn->nSpecific) {
7896 conn->specific = (void **)
7897 realloc(conn->specific, (key + 1) * sizeof(void *));
7898 for (i = conn->nSpecific; i < key; i++)
7899 conn->specific[i] = NULL;
7900 conn->nSpecific = key + 1;
7901 conn->specific[key] = ptr;
7903 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7904 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7905 conn->specific[key] = ptr;
7907 MUTEX_EXIT(&conn->conn_data_lock);
7911 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7914 MUTEX_ENTER(&svc->svc_data_lock);
7915 if (!svc->specific) {
7916 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
7917 for (i = 0; i < key; i++)
7918 svc->specific[i] = NULL;
7919 svc->nSpecific = key + 1;
7920 svc->specific[key] = ptr;
7921 } else if (key >= svc->nSpecific) {
7922 svc->specific = (void **)
7923 realloc(svc->specific, (key + 1) * sizeof(void *));
7924 for (i = svc->nSpecific; i < key; i++)
7925 svc->specific[i] = NULL;
7926 svc->nSpecific = key + 1;
7927 svc->specific[key] = ptr;
7929 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7930 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7931 svc->specific[key] = ptr;
7933 MUTEX_EXIT(&svc->svc_data_lock);
7937 rx_GetSpecific(struct rx_connection *conn, int key)
7940 MUTEX_ENTER(&conn->conn_data_lock);
7941 if (key >= conn->nSpecific)
7944 ptr = conn->specific[key];
7945 MUTEX_EXIT(&conn->conn_data_lock);
7950 rx_GetServiceSpecific(struct rx_service *svc, int key)
7953 MUTEX_ENTER(&svc->svc_data_lock);
7954 if (key >= svc->nSpecific)
7957 ptr = svc->specific[key];
7958 MUTEX_EXIT(&svc->svc_data_lock);
7963 #endif /* !KERNEL */
7966 * processStats is a queue used to store the statistics for the local
7967 * process. Its contents are similar to the contents of the rpcStats
7968 * queue on a rx_peer structure, but the actual data stored within
7969 * this queue contains totals across the lifetime of the process (assuming
7970 * the stats have not been reset) - unlike the per peer structures
7971 * which can come and go based upon the peer lifetime.
7974 static struct rx_queue processStats = { &processStats, &processStats };
7977 * peerStats is a queue used to store the statistics for all peer structs.
7978 * Its contents are the union of all the peer rpcStats queues.
7981 static struct rx_queue peerStats = { &peerStats, &peerStats };
7984 * rxi_monitor_processStats is used to turn process wide stat collection
7988 static int rxi_monitor_processStats = 0;
7991 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7994 static int rxi_monitor_peerStats = 0;
7997 * rxi_AddRpcStat - given all of the information for a particular rpc
7998 * call, create (if needed) and update the stat totals for the rpc.
8002 * IN stats - the queue of stats that will be updated with the new value
8004 * IN rxInterface - a unique number that identifies the rpc interface
8006 * IN currentFunc - the index of the function being invoked
8008 * IN totalFunc - the total number of functions in this interface
8010 * IN queueTime - the amount of time this function waited for a thread
8012 * IN execTime - the amount of time this function invocation took to execute
8014 * IN bytesSent - the number bytes sent by this invocation
8016 * IN bytesRcvd - the number bytes received by this invocation
8018 * IN isServer - if true, this invocation was made to a server
8020 * IN remoteHost - the ip address of the remote host
8022 * IN remotePort - the port of the remote host
8024 * IN addToPeerList - if != 0, add newly created stat to the global peer list
8026 * INOUT counter - if a new stats structure is allocated, the counter will
8027 * be updated with the new number of allocated stat structures
8035 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
8036 afs_uint32 currentFunc, afs_uint32 totalFunc,
8037 struct clock *queueTime, struct clock *execTime,
8038 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
8039 afs_uint32 remoteHost, afs_uint32 remotePort,
8040 int addToPeerList, unsigned int *counter)
8043 rx_interface_stat_p rpc_stat, nrpc_stat;
8046 * See if there's already a structure for this interface
8049 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8050 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8051 && (rpc_stat->stats[0].remote_is_server == isServer))
8056 * Didn't find a match so allocate a new structure and add it to the
8060 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
8061 || (rpc_stat->stats[0].interfaceId != rxInterface)
8062 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8067 sizeof(rx_interface_stat_t) +
8068 totalFunc * sizeof(rx_function_entry_v1_t);
8070 rpc_stat = rxi_Alloc(space);
8071 if (rpc_stat == NULL) {
8075 *counter += totalFunc;
8076 for (i = 0; i < totalFunc; i++) {
8077 rpc_stat->stats[i].remote_peer = remoteHost;
8078 rpc_stat->stats[i].remote_port = remotePort;
8079 rpc_stat->stats[i].remote_is_server = isServer;
8080 rpc_stat->stats[i].interfaceId = rxInterface;
8081 rpc_stat->stats[i].func_total = totalFunc;
8082 rpc_stat->stats[i].func_index = i;
8083 hzero(rpc_stat->stats[i].invocations);
8084 hzero(rpc_stat->stats[i].bytes_sent);
8085 hzero(rpc_stat->stats[i].bytes_rcvd);
8086 rpc_stat->stats[i].queue_time_sum.sec = 0;
8087 rpc_stat->stats[i].queue_time_sum.usec = 0;
8088 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8089 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8090 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8091 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8092 rpc_stat->stats[i].queue_time_max.sec = 0;
8093 rpc_stat->stats[i].queue_time_max.usec = 0;
8094 rpc_stat->stats[i].execution_time_sum.sec = 0;
8095 rpc_stat->stats[i].execution_time_sum.usec = 0;
8096 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8097 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8098 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8099 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8100 rpc_stat->stats[i].execution_time_max.sec = 0;
8101 rpc_stat->stats[i].execution_time_max.usec = 0;
8103 queue_Prepend(stats, rpc_stat);
8104 if (addToPeerList) {
8105 queue_Prepend(&peerStats, &rpc_stat->all_peers);
8110 * Increment the stats for this function
8113 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
8114 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
8115 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
8116 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8117 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8118 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8119 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8121 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8122 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8124 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8125 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8127 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8128 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8130 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8131 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8139 * rx_IncrementTimeAndCount - increment the times and count for a particular
8144 * IN peer - the peer who invoked the rpc
8146 * IN rxInterface - a unique number that identifies the rpc interface
8148 * IN currentFunc - the index of the function being invoked
8150 * IN totalFunc - the total number of functions in this interface
8152 * IN queueTime - the amount of time this function waited for a thread
8154 * IN execTime - the amount of time this function invocation took to execute
8156 * IN bytesSent - the number bytes sent by this invocation
8158 * IN bytesRcvd - the number bytes received by this invocation
8160 * IN isServer - if true, this invocation was made to a server
8168 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8169 afs_uint32 currentFunc, afs_uint32 totalFunc,
8170 struct clock *queueTime, struct clock *execTime,
8171 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8175 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8178 MUTEX_ENTER(&rx_rpc_stats);
8180 if (rxi_monitor_peerStats) {
8181 MUTEX_ENTER(&peer->peer_lock);
8182 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8183 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8184 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8185 MUTEX_EXIT(&peer->peer_lock);
8188 if (rxi_monitor_processStats) {
8189 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8190 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8191 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8194 MUTEX_EXIT(&rx_rpc_stats);
8199 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8203 * IN callerVersion - the rpc stat version of the caller.
8205 * IN count - the number of entries to marshall.
8207 * IN stats - pointer to stats to be marshalled.
8209 * OUT ptr - Where to store the marshalled data.
8216 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8217 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8223 * We only support the first version
8225 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8226 *(ptr++) = stats->remote_peer;
8227 *(ptr++) = stats->remote_port;
8228 *(ptr++) = stats->remote_is_server;
8229 *(ptr++) = stats->interfaceId;
8230 *(ptr++) = stats->func_total;
8231 *(ptr++) = stats->func_index;
8232 *(ptr++) = hgethi(stats->invocations);
8233 *(ptr++) = hgetlo(stats->invocations);
8234 *(ptr++) = hgethi(stats->bytes_sent);
8235 *(ptr++) = hgetlo(stats->bytes_sent);
8236 *(ptr++) = hgethi(stats->bytes_rcvd);
8237 *(ptr++) = hgetlo(stats->bytes_rcvd);
8238 *(ptr++) = stats->queue_time_sum.sec;
8239 *(ptr++) = stats->queue_time_sum.usec;
8240 *(ptr++) = stats->queue_time_sum_sqr.sec;
8241 *(ptr++) = stats->queue_time_sum_sqr.usec;
8242 *(ptr++) = stats->queue_time_min.sec;
8243 *(ptr++) = stats->queue_time_min.usec;
8244 *(ptr++) = stats->queue_time_max.sec;
8245 *(ptr++) = stats->queue_time_max.usec;
8246 *(ptr++) = stats->execution_time_sum.sec;
8247 *(ptr++) = stats->execution_time_sum.usec;
8248 *(ptr++) = stats->execution_time_sum_sqr.sec;
8249 *(ptr++) = stats->execution_time_sum_sqr.usec;
8250 *(ptr++) = stats->execution_time_min.sec;
8251 *(ptr++) = stats->execution_time_min.usec;
8252 *(ptr++) = stats->execution_time_max.sec;
8253 *(ptr++) = stats->execution_time_max.usec;
8259 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8264 * IN callerVersion - the rpc stat version of the caller
8266 * OUT myVersion - the rpc stat version of this function
8268 * OUT clock_sec - local time seconds
8270 * OUT clock_usec - local time microseconds
8272 * OUT allocSize - the number of bytes allocated to contain stats
8274 * OUT statCount - the number stats retrieved from this process.
8276 * OUT stats - the actual stats retrieved from this process.
8280 * Returns void. If successful, stats will != NULL.
8284 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8285 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8286 size_t * allocSize, afs_uint32 * statCount,
8287 afs_uint32 ** stats)
8297 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8300 * Check to see if stats are enabled
8303 MUTEX_ENTER(&rx_rpc_stats);
8304 if (!rxi_monitor_processStats) {
8305 MUTEX_EXIT(&rx_rpc_stats);
8309 clock_GetTime(&now);
8310 *clock_sec = now.sec;
8311 *clock_usec = now.usec;
8314 * Allocate the space based upon the caller version
8316 * If the client is at an older version than we are,
8317 * we return the statistic data in the older data format, but
8318 * we still return our version number so the client knows we
8319 * are maintaining more data than it can retrieve.
8322 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8323 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8324 *statCount = rxi_rpc_process_stat_cnt;
8327 * This can't happen yet, but in the future version changes
8328 * can be handled by adding additional code here
8332 if (space > (size_t) 0) {
8334 ptr = *stats = rxi_Alloc(space);
8337 rx_interface_stat_p rpc_stat, nrpc_stat;
8341 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8343 * Copy the data based upon the caller version
8345 rx_MarshallProcessRPCStats(callerVersion,
8346 rpc_stat->stats[0].func_total,
8347 rpc_stat->stats, &ptr);
8353 MUTEX_EXIT(&rx_rpc_stats);
8358 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8362 * IN callerVersion - the rpc stat version of the caller
8364 * OUT myVersion - the rpc stat version of this function
8366 * OUT clock_sec - local time seconds
8368 * OUT clock_usec - local time microseconds
8370 * OUT allocSize - the number of bytes allocated to contain stats
8372 * OUT statCount - the number of stats retrieved from the individual
8375 * OUT stats - the actual stats retrieved from the individual peer structures.
8379 * Returns void. If successful, stats will != NULL.
8383 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8384 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8385 size_t * allocSize, afs_uint32 * statCount,
8386 afs_uint32 ** stats)
8396 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8399 * Check to see if stats are enabled
8402 MUTEX_ENTER(&rx_rpc_stats);
8403 if (!rxi_monitor_peerStats) {
8404 MUTEX_EXIT(&rx_rpc_stats);
8408 clock_GetTime(&now);
8409 *clock_sec = now.sec;
8410 *clock_usec = now.usec;
8413 * Allocate the space based upon the caller version
8415 * If the client is at an older version than we are,
8416 * we return the statistic data in the older data format, but
8417 * we still return our version number so the client knows we
8418 * are maintaining more data than it can retrieve.
8421 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8422 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8423 *statCount = rxi_rpc_peer_stat_cnt;
8426 * This can't happen yet, but in the future version changes
8427 * can be handled by adding additional code here
8431 if (space > (size_t) 0) {
8433 ptr = *stats = rxi_Alloc(space);
8436 rx_interface_stat_p rpc_stat, nrpc_stat;
8440 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8442 * We have to fix the offset of rpc_stat since we are
8443 * keeping this structure on two rx_queues. The rx_queue
8444 * package assumes that the rx_queue member is the first
8445 * member of the structure. That is, rx_queue assumes that
8446 * any one item is only on one queue at a time. We are
8447 * breaking that assumption and so we have to do a little
8448 * math to fix our pointers.
8451 fix_offset = (char *)rpc_stat;
8452 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8453 rpc_stat = (rx_interface_stat_p) fix_offset;
8456 * Copy the data based upon the caller version
8458 rx_MarshallProcessRPCStats(callerVersion,
8459 rpc_stat->stats[0].func_total,
8460 rpc_stat->stats, &ptr);
8466 MUTEX_EXIT(&rx_rpc_stats);
8471 * rx_FreeRPCStats - free memory allocated by
8472 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8476 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8477 * rx_RetrievePeerRPCStats
8479 * IN allocSize - the number of bytes in stats.
8487 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8489 rxi_Free(stats, allocSize);
8493 * rx_queryProcessRPCStats - see if process rpc stat collection is
8494 * currently enabled.
8500 * Returns 0 if stats are not enabled != 0 otherwise
8504 rx_queryProcessRPCStats(void)
8507 MUTEX_ENTER(&rx_rpc_stats);
8508 rc = rxi_monitor_processStats;
8509 MUTEX_EXIT(&rx_rpc_stats);
8514 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8520 * Returns 0 if stats are not enabled != 0 otherwise
8524 rx_queryPeerRPCStats(void)
8527 MUTEX_ENTER(&rx_rpc_stats);
8528 rc = rxi_monitor_peerStats;
8529 MUTEX_EXIT(&rx_rpc_stats);
8534 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8544 rx_enableProcessRPCStats(void)
8546 MUTEX_ENTER(&rx_rpc_stats);
8547 rx_enable_stats = 1;
8548 rxi_monitor_processStats = 1;
8549 MUTEX_EXIT(&rx_rpc_stats);
8553 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8563 rx_enablePeerRPCStats(void)
8565 MUTEX_ENTER(&rx_rpc_stats);
8566 rx_enable_stats = 1;
8567 rxi_monitor_peerStats = 1;
8568 MUTEX_EXIT(&rx_rpc_stats);
8572 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8582 rx_disableProcessRPCStats(void)
8584 rx_interface_stat_p rpc_stat, nrpc_stat;
8587 MUTEX_ENTER(&rx_rpc_stats);
8590 * Turn off process statistics and if peer stats is also off, turn
8594 rxi_monitor_processStats = 0;
8595 if (rxi_monitor_peerStats == 0) {
8596 rx_enable_stats = 0;
8599 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8600 unsigned int num_funcs = 0;
8603 queue_Remove(rpc_stat);
8604 num_funcs = rpc_stat->stats[0].func_total;
8606 sizeof(rx_interface_stat_t) +
8607 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8609 rxi_Free(rpc_stat, space);
8610 rxi_rpc_process_stat_cnt -= num_funcs;
8612 MUTEX_EXIT(&rx_rpc_stats);
8616 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8626 rx_disablePeerRPCStats(void)
8628 struct rx_peer **peer_ptr, **peer_end;
8632 * Turn off peer statistics and if process stats is also off, turn
8636 rxi_monitor_peerStats = 0;
8637 if (rxi_monitor_processStats == 0) {
8638 rx_enable_stats = 0;
8641 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8642 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8644 struct rx_peer *peer, *next, *prev;
8646 MUTEX_ENTER(&rx_peerHashTable_lock);
8647 MUTEX_ENTER(&rx_rpc_stats);
8648 for (prev = peer = *peer_ptr; peer; peer = next) {
8650 code = MUTEX_TRYENTER(&peer->peer_lock);
8652 rx_interface_stat_p rpc_stat, nrpc_stat;
8655 if (prev == *peer_ptr) {
8666 MUTEX_EXIT(&rx_peerHashTable_lock);
8669 (&peer->rpcStats, rpc_stat, nrpc_stat,
8670 rx_interface_stat)) {
8671 unsigned int num_funcs = 0;
8674 queue_Remove(&rpc_stat->queue_header);
8675 queue_Remove(&rpc_stat->all_peers);
8676 num_funcs = rpc_stat->stats[0].func_total;
8678 sizeof(rx_interface_stat_t) +
8679 rpc_stat->stats[0].func_total *
8680 sizeof(rx_function_entry_v1_t);
8682 rxi_Free(rpc_stat, space);
8683 rxi_rpc_peer_stat_cnt -= num_funcs;
8685 MUTEX_EXIT(&peer->peer_lock);
8687 MUTEX_ENTER(&rx_peerHashTable_lock);
8697 MUTEX_EXIT(&rx_rpc_stats);
8698 MUTEX_EXIT(&rx_peerHashTable_lock);
8703 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8708 * IN clearFlag - flag indicating which stats to clear
8716 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8718 rx_interface_stat_p rpc_stat, nrpc_stat;
8720 MUTEX_ENTER(&rx_rpc_stats);
8722 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8723 unsigned int num_funcs = 0, i;
8724 num_funcs = rpc_stat->stats[0].func_total;
8725 for (i = 0; i < num_funcs; i++) {
8726 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8727 hzero(rpc_stat->stats[i].invocations);
8729 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8730 hzero(rpc_stat->stats[i].bytes_sent);
8732 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8733 hzero(rpc_stat->stats[i].bytes_rcvd);
8735 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8736 rpc_stat->stats[i].queue_time_sum.sec = 0;
8737 rpc_stat->stats[i].queue_time_sum.usec = 0;
8739 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8740 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8741 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8743 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8744 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8745 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8747 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8748 rpc_stat->stats[i].queue_time_max.sec = 0;
8749 rpc_stat->stats[i].queue_time_max.usec = 0;
8751 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8752 rpc_stat->stats[i].execution_time_sum.sec = 0;
8753 rpc_stat->stats[i].execution_time_sum.usec = 0;
8755 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8756 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8757 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8759 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8760 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8761 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8763 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8764 rpc_stat->stats[i].execution_time_max.sec = 0;
8765 rpc_stat->stats[i].execution_time_max.usec = 0;
8770 MUTEX_EXIT(&rx_rpc_stats);
8774 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8779 * IN clearFlag - flag indicating which stats to clear
8787 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8789 rx_interface_stat_p rpc_stat, nrpc_stat;
8791 MUTEX_ENTER(&rx_rpc_stats);
8793 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8794 unsigned int num_funcs = 0, i;
8797 * We have to fix the offset of rpc_stat since we are
8798 * keeping this structure on two rx_queues. The rx_queue
8799 * package assumes that the rx_queue member is the first
8800 * member of the structure. That is, rx_queue assumes that
8801 * any one item is only on one queue at a time. We are
8802 * breaking that assumption and so we have to do a little
8803 * math to fix our pointers.
8806 fix_offset = (char *)rpc_stat;
8807 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8808 rpc_stat = (rx_interface_stat_p) fix_offset;
8810 num_funcs = rpc_stat->stats[0].func_total;
8811 for (i = 0; i < num_funcs; i++) {
8812 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8813 hzero(rpc_stat->stats[i].invocations);
8815 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8816 hzero(rpc_stat->stats[i].bytes_sent);
8818 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8819 hzero(rpc_stat->stats[i].bytes_rcvd);
8821 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8822 rpc_stat->stats[i].queue_time_sum.sec = 0;
8823 rpc_stat->stats[i].queue_time_sum.usec = 0;
8825 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8826 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8827 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8829 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8830 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8831 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8833 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8834 rpc_stat->stats[i].queue_time_max.sec = 0;
8835 rpc_stat->stats[i].queue_time_max.usec = 0;
8837 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8838 rpc_stat->stats[i].execution_time_sum.sec = 0;
8839 rpc_stat->stats[i].execution_time_sum.usec = 0;
8841 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8842 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8843 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8845 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8846 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8847 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8849 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8850 rpc_stat->stats[i].execution_time_max.sec = 0;
8851 rpc_stat->stats[i].execution_time_max.usec = 0;
8856 MUTEX_EXIT(&rx_rpc_stats);
8860 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8861 * is authorized to enable/disable/clear RX statistics.
8863 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8866 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8868 rxi_rxstat_userok = proc;
8872 rx_RxStatUserOk(struct rx_call *call)
8874 if (!rxi_rxstat_userok)
8876 return rxi_rxstat_userok(call);
8881 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8882 * function in the MSVC runtime DLL (msvcrt.dll).
8884 * Note: the system serializes calls to this function.
8887 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8888 DWORD reason, /* reason function is being called */
8889 LPVOID reserved) /* reserved for future use */
8892 case DLL_PROCESS_ATTACH:
8893 /* library is being attached to a process */
8897 case DLL_PROCESS_DETACH:
8904 #endif /* AFS_NT40_ENV */
8907 int rx_DumpCalls(FILE *outputFile, char *cookie)
8909 #ifdef RXDEBUG_PACKET
8910 #ifdef KDUMP_RX_LOCK
8911 struct rx_call_rx_lock *c;
8918 #define RXDPRINTF sprintf
8919 #define RXDPRINTOUT output
8921 #define RXDPRINTF fprintf
8922 #define RXDPRINTOUT outputFile
8925 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8927 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8930 for (c = rx_allCallsp; c; c = c->allNextp) {
8931 u_short rqc, tqc, iovqc;
8932 struct rx_packet *p, *np;
8934 MUTEX_ENTER(&c->lock);
8935 queue_Count(&c->rq, p, np, rx_packet, rqc);
8936 queue_Count(&c->tq, p, np, rx_packet, tqc);
8937 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8939 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, "
8940 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8941 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8942 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8943 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8944 #ifdef RX_ENABLE_LOCKS
8947 #ifdef RX_REFCOUNT_CHECK
8948 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
8949 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
8952 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,
8953 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
8954 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
8955 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
8956 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
8957 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
8958 #ifdef RX_ENABLE_LOCKS
8959 , (afs_uint32)c->refCount
8961 #ifdef RX_REFCOUNT_CHECK
8962 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
8965 MUTEX_EXIT(&c->lock);
8968 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8971 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
8973 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8975 #endif /* RXDEBUG_PACKET */