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>
13 #include "afs/param.h"
16 # include "afs/sysincludes.h"
17 # include "afsincludes.h"
22 # ifdef AFS_LINUX20_ENV
23 # include "h/socket.h"
25 # include "netinet/in.h"
27 # include "inet/common.h"
29 # include "inet/ip_ire.h"
31 # include "afs/afs_args.h"
32 # include "afs/afs_osi.h"
33 # ifdef RX_KERNEL_TRACE
34 # include "rx_kcommon.h"
36 # if defined(AFS_AIX_ENV)
40 # undef RXDEBUG /* turn off debugging */
42 # if defined(AFS_SGI_ENV)
43 # include "sys/debug.h"
46 # include "afs/sysincludes.h"
47 # include "afsincludes.h"
48 # endif /* !UKERNEL */
49 # include "afs/lock.h"
50 # include "rx_kmutex.h"
51 # include "rx_kernel.h"
52 # define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
53 # define AFSOP_STOP_AFS 211 /* Stop AFS process */
54 # define AFSOP_STOP_BKG 212 /* Stop BKG process */
55 extern afs_int32 afs_termState;
57 # include "sys/lockl.h"
58 # include "sys/lock_def.h"
59 # endif /* AFS_AIX41_ENV */
60 # include "afs/rxgen_consts.h"
63 # include <sys/types.h>
73 # include <afs/afsutil.h>
74 # include <WINNT\afsreg.h>
76 # include <sys/socket.h>
77 # include <sys/file.h>
79 # include <sys/stat.h>
80 # include <netinet/in.h>
81 # include <sys/time.h>
89 #include "rx_atomic.h"
90 #include "rx_globals.h"
92 #include "rx_internal.h"
95 #include <afs/rxgen_consts.h>
98 #ifdef AFS_PTHREAD_ENV
100 int (*registerProgram) (pid_t, char *) = 0;
101 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
104 int (*registerProgram) (PROCESS, char *) = 0;
105 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
109 /* Local static routines */
110 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
111 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
112 struct rx_peer *, struct clock *);
114 #ifdef RX_ENABLE_LOCKS
115 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
118 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
120 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
121 rx_atomic_t rxi_start_in_error;
123 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
126 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
127 * currently allocated within rx. This number is used to allocate the
128 * memory required to return the statistics when queried.
129 * Protected by the rx_rpc_stats mutex.
132 static unsigned int rxi_rpc_peer_stat_cnt;
135 * rxi_rpc_process_stat_cnt counts the total number of local process stat
136 * structures currently allocated within rx. The number is used to allocate
137 * the memory required to return the statistics when queried.
138 * Protected by the rx_rpc_stats mutex.
141 static unsigned int rxi_rpc_process_stat_cnt;
143 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
144 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
146 #if !defined(offsetof)
147 #include <stddef.h> /* for definition of offsetof() */
150 #ifdef RX_ENABLE_LOCKS
151 afs_kmutex_t rx_atomic_mutex;
154 #ifdef AFS_PTHREAD_ENV
157 * Use procedural initialization of mutexes/condition variables
161 extern afs_kmutex_t rx_quota_mutex;
162 extern afs_kmutex_t rx_pthread_mutex;
163 extern afs_kmutex_t rx_packets_mutex;
164 extern afs_kmutex_t rx_refcnt_mutex;
165 extern afs_kmutex_t des_init_mutex;
166 extern afs_kmutex_t des_random_mutex;
167 extern afs_kmutex_t rx_clock_mutex;
168 extern afs_kmutex_t rxi_connCacheMutex;
169 extern afs_kmutex_t rx_event_mutex;
170 extern afs_kmutex_t osi_malloc_mutex;
171 extern afs_kmutex_t event_handler_mutex;
172 extern afs_kmutex_t listener_mutex;
173 extern afs_kmutex_t rx_if_init_mutex;
174 extern afs_kmutex_t rx_if_mutex;
175 extern afs_kmutex_t rxkad_client_uid_mutex;
176 extern afs_kmutex_t rxkad_random_mutex;
178 extern afs_kcondvar_t rx_event_handler_cond;
179 extern afs_kcondvar_t rx_listener_cond;
181 static afs_kmutex_t epoch_mutex;
182 static afs_kmutex_t rx_init_mutex;
183 static afs_kmutex_t rx_debug_mutex;
184 static afs_kmutex_t rx_rpc_stats;
187 rxi_InitPthread(void)
189 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
190 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
191 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
192 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
193 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
194 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
195 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
196 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
197 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
198 MUTEX_INIT(&rx_event_mutex, "event", MUTEX_DEFAULT, 0);
199 MUTEX_INIT(&des_init_mutex, "des", MUTEX_DEFAULT, 0);
200 MUTEX_INIT(&des_random_mutex, "random", MUTEX_DEFAULT, 0);
201 MUTEX_INIT(&osi_malloc_mutex, "malloc", MUTEX_DEFAULT, 0);
202 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
203 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
204 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
205 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
206 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
207 MUTEX_INIT(&rxkad_client_uid_mutex, "uid", MUTEX_DEFAULT, 0);
208 MUTEX_INIT(&rxkad_random_mutex, "rxkad random", MUTEX_DEFAULT, 0);
209 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
211 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
212 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
214 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
215 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
217 rxkad_global_stats_init();
219 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
220 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
221 #ifdef RX_ENABLE_LOCKS
224 #endif /* RX_LOCKS_DB */
225 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
226 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
228 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
230 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
232 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
234 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
235 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
236 #endif /* RX_ENABLE_LOCKS */
239 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
240 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
242 * The rx_stats_mutex mutex protects the following global variables:
243 * rxi_lowConnRefCount
244 * rxi_lowPeerRefCount
253 * The rx_quota_mutex mutex protects the following global variables:
261 * The rx_freePktQ_lock protects the following global variables:
266 * The rx_packets_mutex mutex protects the following global variables:
274 * The rx_pthread_mutex mutex protects the following global variables:
275 * rxi_fcfs_thread_num
278 #define INIT_PTHREAD_LOCKS
282 /* Variables for handling the minProcs implementation. availProcs gives the
283 * number of threads available in the pool at this moment (not counting dudes
284 * executing right now). totalMin gives the total number of procs required
285 * for handling all minProcs requests. minDeficit is a dynamic variable
286 * tracking the # of procs required to satisfy all of the remaining minProcs
288 * For fine grain locking to work, the quota check and the reservation of
289 * a server thread has to come while rxi_availProcs and rxi_minDeficit
290 * are locked. To this end, the code has been modified under #ifdef
291 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
292 * same time. A new function, ReturnToServerPool() returns the allocation.
294 * A call can be on several queue's (but only one at a time). When
295 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
296 * that no one else is touching the queue. To this end, we store the address
297 * of the queue lock in the call structure (under the call lock) when we
298 * put the call on a queue, and we clear the call_queue_lock when the
299 * call is removed from a queue (once the call lock has been obtained).
300 * This allows rxi_ResetCall to safely synchronize with others wishing
301 * to manipulate the queue.
304 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
305 static afs_kmutex_t rx_rpc_stats;
306 void rxi_StartUnlocked(struct rxevent *event, void *call,
307 void *arg1, int istack);
310 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
311 ** pretty good that the next packet coming in is from the same connection
312 ** as the last packet, since we're send multiple packets in a transmit window.
314 struct rx_connection *rxLastConn = 0;
316 #ifdef RX_ENABLE_LOCKS
317 /* The locking hierarchy for rx fine grain locking is composed of these
320 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
321 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
322 * call->lock - locks call data fields.
323 * These are independent of each other:
324 * rx_freeCallQueue_lock
329 * serverQueueEntry->lock
330 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
332 * peer->lock - locks peer data fields.
333 * conn_data_lock - that more than one thread is not updating a conn data
334 * field at the same time.
345 * Do we need a lock to protect the peer field in the conn structure?
346 * conn->peer was previously a constant for all intents and so has no
347 * lock protecting this field. The multihomed client delta introduced
348 * a RX code change : change the peer field in the connection structure
349 * to that remote interface from which the last packet for this
350 * connection was sent out. This may become an issue if further changes
353 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
354 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
356 /* rxdb_fileID is used to identify the lock location, along with line#. */
357 static int rxdb_fileID = RXDB_FILE_RX;
358 #endif /* RX_LOCKS_DB */
359 #else /* RX_ENABLE_LOCKS */
360 #define SET_CALL_QUEUE_LOCK(C, L)
361 #define CLEAR_CALL_QUEUE_LOCK(C)
362 #endif /* RX_ENABLE_LOCKS */
363 struct rx_serverQueueEntry *rx_waitForPacket = 0;
364 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
366 /* ------------Exported Interfaces------------- */
368 /* This function allows rxkad to set the epoch to a suitably random number
369 * which rx_NewConnection will use in the future. The principle purpose is to
370 * get rxnull connections to use the same epoch as the rxkad connections do, at
371 * least once the first rxkad connection is established. This is important now
372 * that the host/port addresses aren't used in FindConnection: the uniqueness
373 * of epoch/cid matters and the start time won't do. */
375 #ifdef AFS_PTHREAD_ENV
377 * This mutex protects the following global variables:
381 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
382 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
386 #endif /* AFS_PTHREAD_ENV */
389 rx_SetEpoch(afs_uint32 epoch)
396 /* Initialize rx. A port number may be mentioned, in which case this
397 * becomes the default port number for any service installed later.
398 * If 0 is provided for the port number, a random port will be chosen
399 * by the kernel. Whether this will ever overlap anything in
400 * /etc/services is anybody's guess... Returns 0 on success, -1 on
405 int rxinit_status = 1;
406 #ifdef AFS_PTHREAD_ENV
408 * This mutex protects the following global variables:
412 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
413 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
416 #define UNLOCK_RX_INIT
420 rx_InitHost(u_int host, u_int port)
427 char *htable, *ptable;
434 if (rxinit_status == 0) {
435 tmp_status = rxinit_status;
437 return tmp_status; /* Already started; return previous error code. */
443 if (afs_winsockInit() < 0)
449 * Initialize anything necessary to provide a non-premptive threading
452 rxi_InitializeThreadSupport();
455 /* Allocate and initialize a socket for client and perhaps server
458 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
459 if (rx_socket == OSI_NULLSOCKET) {
463 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
466 #endif /* RX_LOCKS_DB */
467 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
468 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
469 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
470 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
471 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
472 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
473 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
474 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
475 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
477 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
479 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
481 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
483 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
484 #if defined(AFS_HPUX110_ENV)
486 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
487 #endif /* AFS_HPUX110_ENV */
488 #endif /* RX_ENABLE_LOCKS && KERNEL */
491 rx_connDeadTime = 12;
492 rx_tranquil = 0; /* reset flag */
493 rxi_ResetStatistics();
495 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
496 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
497 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
498 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
499 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
500 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
502 /* Malloc up a bunch of packets & buffers */
504 queue_Init(&rx_freePacketQueue);
505 rxi_NeedMorePackets = FALSE;
506 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
508 /* enforce a minimum number of allocated packets */
509 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
510 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
512 /* allocate the initial free packet pool */
513 #ifdef RX_ENABLE_TSFPQ
514 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
515 #else /* RX_ENABLE_TSFPQ */
516 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
517 #endif /* RX_ENABLE_TSFPQ */
524 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
525 tv.tv_sec = clock_now.sec;
526 tv.tv_usec = clock_now.usec;
527 srand((unsigned int)tv.tv_usec);
534 #if defined(KERNEL) && !defined(UKERNEL)
535 /* Really, this should never happen in a real kernel */
538 struct sockaddr_in addr;
540 int addrlen = sizeof(addr);
542 socklen_t addrlen = sizeof(addr);
544 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
548 rx_port = addr.sin_port;
551 rx_stats.minRtt.sec = 9999999;
553 rx_SetEpoch(tv.tv_sec | 0x80000000);
555 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
556 * will provide a randomer value. */
558 MUTEX_ENTER(&rx_quota_mutex);
559 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
560 MUTEX_EXIT(&rx_quota_mutex);
561 /* *Slightly* random start time for the cid. This is just to help
562 * out with the hashing function at the peer */
563 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
564 rx_connHashTable = (struct rx_connection **)htable;
565 rx_peerHashTable = (struct rx_peer **)ptable;
567 rx_lastAckDelay.sec = 0;
568 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
569 rx_hardAckDelay.sec = 0;
570 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
571 rx_softAckDelay.sec = 0;
572 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
574 rxevent_Init(20, rxi_ReScheduleEvents);
576 /* Initialize various global queues */
577 queue_Init(&rx_idleServerQueue);
578 queue_Init(&rx_incomingCallQueue);
579 queue_Init(&rx_freeCallQueue);
581 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
582 /* Initialize our list of usable IP addresses. */
586 /* Start listener process (exact function is dependent on the
587 * implementation environment--kernel or user space) */
591 tmp_status = rxinit_status = 0;
599 return rx_InitHost(htonl(INADDR_ANY), port);
602 /* called with unincremented nRequestsRunning to see if it is OK to start
603 * a new thread in this service. Could be "no" for two reasons: over the
604 * max quota, or would prevent others from reaching their min quota.
606 #ifdef RX_ENABLE_LOCKS
607 /* This verion of QuotaOK reserves quota if it's ok while the
608 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
611 QuotaOK(struct rx_service *aservice)
613 /* check if over max quota */
614 if (aservice->nRequestsRunning >= aservice->maxProcs) {
618 /* under min quota, we're OK */
619 /* otherwise, can use only if there are enough to allow everyone
620 * to go to their min quota after this guy starts.
623 MUTEX_ENTER(&rx_quota_mutex);
624 if ((aservice->nRequestsRunning < aservice->minProcs)
625 || (rxi_availProcs > rxi_minDeficit)) {
626 aservice->nRequestsRunning++;
627 /* just started call in minProcs pool, need fewer to maintain
629 if (aservice->nRequestsRunning <= aservice->minProcs)
632 MUTEX_EXIT(&rx_quota_mutex);
635 MUTEX_EXIT(&rx_quota_mutex);
641 ReturnToServerPool(struct rx_service *aservice)
643 aservice->nRequestsRunning--;
644 MUTEX_ENTER(&rx_quota_mutex);
645 if (aservice->nRequestsRunning < aservice->minProcs)
648 MUTEX_EXIT(&rx_quota_mutex);
651 #else /* RX_ENABLE_LOCKS */
653 QuotaOK(struct rx_service *aservice)
656 /* under min quota, we're OK */
657 if (aservice->nRequestsRunning < aservice->minProcs)
660 /* check if over max quota */
661 if (aservice->nRequestsRunning >= aservice->maxProcs)
664 /* otherwise, can use only if there are enough to allow everyone
665 * to go to their min quota after this guy starts.
667 MUTEX_ENTER(&rx_quota_mutex);
668 if (rxi_availProcs > rxi_minDeficit)
670 MUTEX_EXIT(&rx_quota_mutex);
673 #endif /* RX_ENABLE_LOCKS */
676 /* Called by rx_StartServer to start up lwp's to service calls.
677 NExistingProcs gives the number of procs already existing, and which
678 therefore needn't be created. */
680 rxi_StartServerProcs(int nExistingProcs)
682 struct rx_service *service;
687 /* For each service, reserve N processes, where N is the "minimum"
688 * number of processes that MUST be able to execute a request in parallel,
689 * at any time, for that process. Also compute the maximum difference
690 * between any service's maximum number of processes that can run
691 * (i.e. the maximum number that ever will be run, and a guarantee
692 * that this number will run if other services aren't running), and its
693 * minimum number. The result is the extra number of processes that
694 * we need in order to provide the latter guarantee */
695 for (i = 0; i < RX_MAX_SERVICES; i++) {
697 service = rx_services[i];
698 if (service == (struct rx_service *)0)
700 nProcs += service->minProcs;
701 diff = service->maxProcs - service->minProcs;
705 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
706 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
707 for (i = 0; i < nProcs; i++) {
708 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
714 /* This routine is only required on Windows */
716 rx_StartClientThread(void)
718 #ifdef AFS_PTHREAD_ENV
720 pid = pthread_self();
721 #endif /* AFS_PTHREAD_ENV */
723 #endif /* AFS_NT40_ENV */
725 /* This routine must be called if any services are exported. If the
726 * donateMe flag is set, the calling process is donated to the server
729 rx_StartServer(int donateMe)
731 struct rx_service *service;
737 /* Start server processes, if necessary (exact function is dependent
738 * on the implementation environment--kernel or user space). DonateMe
739 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
740 * case, one less new proc will be created rx_StartServerProcs.
742 rxi_StartServerProcs(donateMe);
744 /* count up the # of threads in minProcs, and add set the min deficit to
745 * be that value, too.
747 for (i = 0; i < RX_MAX_SERVICES; i++) {
748 service = rx_services[i];
749 if (service == (struct rx_service *)0)
751 MUTEX_ENTER(&rx_quota_mutex);
752 rxi_totalMin += service->minProcs;
753 /* below works even if a thread is running, since minDeficit would
754 * still have been decremented and later re-incremented.
756 rxi_minDeficit += service->minProcs;
757 MUTEX_EXIT(&rx_quota_mutex);
760 /* Turn on reaping of idle server connections */
761 rxi_ReapConnections(NULL, NULL, NULL);
770 #ifdef AFS_PTHREAD_ENV
772 pid = afs_pointer_to_int(pthread_self());
773 #else /* AFS_PTHREAD_ENV */
775 LWP_CurrentProcess(&pid);
776 #endif /* AFS_PTHREAD_ENV */
778 sprintf(name, "srv_%d", ++nProcs);
780 (*registerProgram) (pid, name);
782 #endif /* AFS_NT40_ENV */
783 rx_ServerProc(NULL); /* Never returns */
785 #ifdef RX_ENABLE_TSFPQ
786 /* no use leaving packets around in this thread's local queue if
787 * it isn't getting donated to the server thread pool.
789 rxi_FlushLocalPacketsTSFPQ();
790 #endif /* RX_ENABLE_TSFPQ */
794 /* Create a new client connection to the specified service, using the
795 * specified security object to implement the security model for this
797 struct rx_connection *
798 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
799 struct rx_securityClass *securityObject,
800 int serviceSecurityIndex)
804 struct rx_connection *conn;
809 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
810 "serviceSecurityIndex %d)\n",
811 ntohl(shost), ntohs(sport), sservice, securityObject,
812 serviceSecurityIndex));
814 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
815 * the case of kmem_alloc? */
816 conn = rxi_AllocConnection();
817 #ifdef RX_ENABLE_LOCKS
818 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
819 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
820 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
823 MUTEX_ENTER(&rx_connHashTable_lock);
824 cid = (rx_nextCid += RX_MAXCALLS);
825 conn->type = RX_CLIENT_CONNECTION;
827 conn->epoch = rx_epoch;
828 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
829 conn->serviceId = sservice;
830 conn->securityObject = securityObject;
831 conn->securityData = (void *) 0;
832 conn->securityIndex = serviceSecurityIndex;
833 rx_SetConnDeadTime(conn, rx_connDeadTime);
834 rx_SetConnSecondsUntilNatPing(conn, 0);
835 conn->ackRate = RX_FAST_ACK_RATE;
837 conn->specific = NULL;
838 conn->challengeEvent = NULL;
839 conn->delayedAbortEvent = NULL;
840 conn->abortCount = 0;
842 for (i = 0; i < RX_MAXCALLS; i++) {
843 conn->twind[i] = rx_initSendWindow;
844 conn->rwind[i] = rx_initReceiveWindow;
847 RXS_NewConnection(securityObject, conn);
849 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
851 conn->refCount++; /* no lock required since only this thread knows... */
852 conn->next = rx_connHashTable[hashindex];
853 rx_connHashTable[hashindex] = conn;
855 rx_atomic_inc(&rx_stats.nClientConns);
856 MUTEX_EXIT(&rx_connHashTable_lock);
862 * Ensure a connection's timeout values are valid.
864 * @param[in] conn The connection to check
866 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
867 * unless idleDeadTime and/or hardDeadTime are not set
871 rxi_CheckConnTimeouts(struct rx_connection *conn)
873 /* a connection's timeouts must have the relationship
874 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
875 * total loss of network to a peer may cause an idle timeout instead of a
876 * dead timeout, simply because the idle timeout gets hit first. Also set
877 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
878 /* this logic is slightly complicated by the fact that
879 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
881 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
882 if (conn->idleDeadTime) {
883 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
885 if (conn->hardDeadTime) {
886 if (conn->idleDeadTime) {
887 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
889 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
895 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
897 /* The idea is to set the dead time to a value that allows several
898 * keepalives to be dropped without timing out the connection. */
899 conn->secondsUntilDead = seconds;
900 rxi_CheckConnTimeouts(conn);
901 conn->secondsUntilPing = conn->secondsUntilDead / 6;
905 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
907 conn->hardDeadTime = seconds;
908 rxi_CheckConnTimeouts(conn);
912 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
914 conn->idleDeadTime = seconds;
915 rxi_CheckConnTimeouts(conn);
918 int rxi_lowPeerRefCount = 0;
919 int rxi_lowConnRefCount = 0;
922 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
923 * NOTE: must not be called with rx_connHashTable_lock held.
926 rxi_CleanupConnection(struct rx_connection *conn)
928 /* Notify the service exporter, if requested, that this connection
929 * is being destroyed */
930 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
931 (*conn->service->destroyConnProc) (conn);
933 /* Notify the security module that this connection is being destroyed */
934 RXS_DestroyConnection(conn->securityObject, conn);
936 /* If this is the last connection using the rx_peer struct, set its
937 * idle time to now. rxi_ReapConnections will reap it if it's still
938 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
940 MUTEX_ENTER(&rx_peerHashTable_lock);
941 if (conn->peer->refCount < 2) {
942 conn->peer->idleWhen = clock_Sec();
943 if (conn->peer->refCount < 1) {
944 conn->peer->refCount = 1;
945 if (rx_stats_active) {
946 MUTEX_ENTER(&rx_stats_mutex);
947 rxi_lowPeerRefCount++;
948 MUTEX_EXIT(&rx_stats_mutex);
952 conn->peer->refCount--;
953 MUTEX_EXIT(&rx_peerHashTable_lock);
957 if (conn->type == RX_SERVER_CONNECTION)
958 rx_atomic_dec(&rx_stats.nServerConns);
960 rx_atomic_dec(&rx_stats.nClientConns);
963 if (conn->specific) {
965 for (i = 0; i < conn->nSpecific; i++) {
966 if (conn->specific[i] && rxi_keyCreate_destructor[i])
967 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
968 conn->specific[i] = NULL;
970 free(conn->specific);
972 conn->specific = NULL;
976 MUTEX_DESTROY(&conn->conn_call_lock);
977 MUTEX_DESTROY(&conn->conn_data_lock);
978 CV_DESTROY(&conn->conn_call_cv);
980 rxi_FreeConnection(conn);
983 /* Destroy the specified connection */
985 rxi_DestroyConnection(struct rx_connection *conn)
987 MUTEX_ENTER(&rx_connHashTable_lock);
988 rxi_DestroyConnectionNoLock(conn);
989 /* conn should be at the head of the cleanup list */
990 if (conn == rx_connCleanup_list) {
991 rx_connCleanup_list = rx_connCleanup_list->next;
992 MUTEX_EXIT(&rx_connHashTable_lock);
993 rxi_CleanupConnection(conn);
995 #ifdef RX_ENABLE_LOCKS
997 MUTEX_EXIT(&rx_connHashTable_lock);
999 #endif /* RX_ENABLE_LOCKS */
1003 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1005 struct rx_connection **conn_ptr;
1007 struct rx_packet *packet;
1014 MUTEX_ENTER(&conn->conn_data_lock);
1015 MUTEX_ENTER(&rx_refcnt_mutex);
1016 if (conn->refCount > 0)
1019 if (rx_stats_active) {
1020 MUTEX_ENTER(&rx_stats_mutex);
1021 rxi_lowConnRefCount++;
1022 MUTEX_EXIT(&rx_stats_mutex);
1026 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1027 /* Busy; wait till the last guy before proceeding */
1028 MUTEX_EXIT(&rx_refcnt_mutex);
1029 MUTEX_EXIT(&conn->conn_data_lock);
1034 /* If the client previously called rx_NewCall, but it is still
1035 * waiting, treat this as a running call, and wait to destroy the
1036 * connection later when the call completes. */
1037 if ((conn->type == RX_CLIENT_CONNECTION)
1038 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1039 conn->flags |= RX_CONN_DESTROY_ME;
1040 MUTEX_EXIT(&conn->conn_data_lock);
1044 MUTEX_EXIT(&rx_refcnt_mutex);
1045 MUTEX_EXIT(&conn->conn_data_lock);
1047 /* Check for extant references to this connection */
1048 for (i = 0; i < RX_MAXCALLS; i++) {
1049 struct rx_call *call = conn->call[i];
1052 if (conn->type == RX_CLIENT_CONNECTION) {
1053 MUTEX_ENTER(&call->lock);
1054 if (call->delayedAckEvent) {
1055 /* Push the final acknowledgment out now--there
1056 * won't be a subsequent call to acknowledge the
1057 * last reply packets */
1058 rxevent_Cancel(call->delayedAckEvent, call,
1059 RX_CALL_REFCOUNT_DELAY);
1060 if (call->state == RX_STATE_PRECALL
1061 || call->state == RX_STATE_ACTIVE) {
1062 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1064 rxi_AckAll(NULL, call, 0);
1067 MUTEX_EXIT(&call->lock);
1071 #ifdef RX_ENABLE_LOCKS
1073 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1074 MUTEX_EXIT(&conn->conn_data_lock);
1076 /* Someone is accessing a packet right now. */
1080 #endif /* RX_ENABLE_LOCKS */
1083 /* Don't destroy the connection if there are any call
1084 * structures still in use */
1085 MUTEX_ENTER(&conn->conn_data_lock);
1086 conn->flags |= RX_CONN_DESTROY_ME;
1087 MUTEX_EXIT(&conn->conn_data_lock);
1092 if (conn->natKeepAliveEvent) {
1093 rxi_NatKeepAliveOff(conn);
1096 if (conn->delayedAbortEvent) {
1097 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1098 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1100 MUTEX_ENTER(&conn->conn_data_lock);
1101 rxi_SendConnectionAbort(conn, packet, 0, 1);
1102 MUTEX_EXIT(&conn->conn_data_lock);
1103 rxi_FreePacket(packet);
1107 /* Remove from connection hash table before proceeding */
1109 &rx_connHashTable[CONN_HASH
1110 (peer->host, peer->port, conn->cid, conn->epoch,
1112 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1113 if (*conn_ptr == conn) {
1114 *conn_ptr = conn->next;
1118 /* if the conn that we are destroying was the last connection, then we
1119 * clear rxLastConn as well */
1120 if (rxLastConn == conn)
1123 /* Make sure the connection is completely reset before deleting it. */
1124 /* get rid of pending events that could zap us later */
1125 if (conn->challengeEvent)
1126 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1127 if (conn->checkReachEvent)
1128 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1129 if (conn->natKeepAliveEvent)
1130 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
1132 /* Add the connection to the list of destroyed connections that
1133 * need to be cleaned up. This is necessary to avoid deadlocks
1134 * in the routines we call to inform others that this connection is
1135 * being destroyed. */
1136 conn->next = rx_connCleanup_list;
1137 rx_connCleanup_list = conn;
1140 /* Externally available version */
1142 rx_DestroyConnection(struct rx_connection *conn)
1147 rxi_DestroyConnection(conn);
1152 rx_GetConnection(struct rx_connection *conn)
1157 MUTEX_ENTER(&rx_refcnt_mutex);
1159 MUTEX_EXIT(&rx_refcnt_mutex);
1163 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1164 /* Wait for the transmit queue to no longer be busy.
1165 * requires the call->lock to be held */
1167 rxi_WaitforTQBusy(struct rx_call *call) {
1168 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1169 call->flags |= RX_CALL_TQ_WAIT;
1171 #ifdef RX_ENABLE_LOCKS
1172 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1173 CV_WAIT(&call->cv_tq, &call->lock);
1174 #else /* RX_ENABLE_LOCKS */
1175 osi_rxSleep(&call->tq);
1176 #endif /* RX_ENABLE_LOCKS */
1178 if (call->tqWaiters == 0) {
1179 call->flags &= ~RX_CALL_TQ_WAIT;
1186 rxi_WakeUpTransmitQueue(struct rx_call *call)
1188 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1189 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1190 call, call->tqWaiters, call->flags));
1191 #ifdef RX_ENABLE_LOCKS
1192 osirx_AssertMine(&call->lock, "rxi_Start start");
1193 CV_BROADCAST(&call->cv_tq);
1194 #else /* RX_ENABLE_LOCKS */
1195 osi_rxWakeup(&call->tq);
1196 #endif /* RX_ENABLE_LOCKS */
1200 /* Start a new rx remote procedure call, on the specified connection.
1201 * If wait is set to 1, wait for a free call channel; otherwise return
1202 * 0. Maxtime gives the maximum number of seconds this call may take,
1203 * after rx_NewCall returns. After this time interval, a call to any
1204 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1205 * For fine grain locking, we hold the conn_call_lock in order to
1206 * to ensure that we don't get signalle after we found a call in an active
1207 * state and before we go to sleep.
1210 rx_NewCall(struct rx_connection *conn)
1213 struct rx_call *call;
1214 struct clock queueTime;
1218 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1221 clock_GetTime(&queueTime);
1223 * Check if there are others waiting for a new call.
1224 * If so, let them go first to avoid starving them.
1225 * This is a fairly simple scheme, and might not be
1226 * a complete solution for large numbers of waiters.
1228 * makeCallWaiters keeps track of the number of
1229 * threads waiting to make calls and the
1230 * RX_CONN_MAKECALL_WAITING flag bit is used to
1231 * indicate that there are indeed calls waiting.
1232 * The flag is set when the waiter is incremented.
1233 * It is only cleared when makeCallWaiters is 0.
1234 * This prevents us from accidently destroying the
1235 * connection while it is potentially about to be used.
1237 MUTEX_ENTER(&conn->conn_call_lock);
1238 MUTEX_ENTER(&conn->conn_data_lock);
1239 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1240 conn->flags |= RX_CONN_MAKECALL_WAITING;
1241 conn->makeCallWaiters++;
1242 MUTEX_EXIT(&conn->conn_data_lock);
1244 #ifdef RX_ENABLE_LOCKS
1245 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1249 MUTEX_ENTER(&conn->conn_data_lock);
1250 conn->makeCallWaiters--;
1251 if (conn->makeCallWaiters == 0)
1252 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1255 /* We are now the active thread in rx_NewCall */
1256 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1257 MUTEX_EXIT(&conn->conn_data_lock);
1262 for (i = 0; i < RX_MAXCALLS; i++) {
1263 call = conn->call[i];
1265 if (call->state == RX_STATE_DALLY) {
1266 MUTEX_ENTER(&call->lock);
1267 if (call->state == RX_STATE_DALLY) {
1269 * We are setting the state to RX_STATE_RESET to
1270 * ensure that no one else will attempt to use this
1271 * call once we drop the conn->conn_call_lock and
1272 * call->lock. We must drop the conn->conn_call_lock
1273 * before calling rxi_ResetCall because the process
1274 * of clearing the transmit queue can block for an
1275 * extended period of time. If we block while holding
1276 * the conn->conn_call_lock, then all rx_EndCall
1277 * processing will block as well. This has a detrimental
1278 * effect on overall system performance.
1280 call->state = RX_STATE_RESET;
1281 MUTEX_EXIT(&conn->conn_call_lock);
1282 MUTEX_ENTER(&rx_refcnt_mutex);
1283 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1284 MUTEX_EXIT(&rx_refcnt_mutex);
1285 rxi_ResetCall(call, 0);
1286 (*call->callNumber)++;
1287 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1291 * If we failed to be able to safely obtain the
1292 * conn->conn_call_lock we will have to drop the
1293 * call->lock to avoid a deadlock. When the call->lock
1294 * is released the state of the call can change. If it
1295 * is no longer RX_STATE_RESET then some other thread is
1298 MUTEX_EXIT(&call->lock);
1299 MUTEX_ENTER(&conn->conn_call_lock);
1300 MUTEX_ENTER(&call->lock);
1302 if (call->state == RX_STATE_RESET)
1306 * If we get here it means that after dropping
1307 * the conn->conn_call_lock and call->lock that
1308 * the call is no longer ours. If we can't find
1309 * a free call in the remaining slots we should
1310 * not go immediately to RX_CONN_MAKECALL_WAITING
1311 * because by dropping the conn->conn_call_lock
1312 * we have given up synchronization with rx_EndCall.
1313 * Instead, cycle through one more time to see if
1314 * we can find a call that can call our own.
1316 MUTEX_ENTER(&rx_refcnt_mutex);
1317 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1318 MUTEX_EXIT(&rx_refcnt_mutex);
1321 MUTEX_EXIT(&call->lock);
1324 /* rxi_NewCall returns with mutex locked */
1325 call = rxi_NewCall(conn, i);
1326 MUTEX_ENTER(&rx_refcnt_mutex);
1327 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1328 MUTEX_EXIT(&rx_refcnt_mutex);
1332 if (i < RX_MAXCALLS) {
1338 MUTEX_ENTER(&conn->conn_data_lock);
1339 conn->flags |= RX_CONN_MAKECALL_WAITING;
1340 conn->makeCallWaiters++;
1341 MUTEX_EXIT(&conn->conn_data_lock);
1343 #ifdef RX_ENABLE_LOCKS
1344 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1348 MUTEX_ENTER(&conn->conn_data_lock);
1349 conn->makeCallWaiters--;
1350 if (conn->makeCallWaiters == 0)
1351 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1352 MUTEX_EXIT(&conn->conn_data_lock);
1354 /* Client is initially in send mode */
1355 call->state = RX_STATE_ACTIVE;
1356 call->error = conn->error;
1358 call->mode = RX_MODE_ERROR;
1360 call->mode = RX_MODE_SENDING;
1362 /* remember start time for call in case we have hard dead time limit */
1363 call->queueTime = queueTime;
1364 clock_GetTime(&call->startTime);
1365 hzero(call->bytesSent);
1366 hzero(call->bytesRcvd);
1368 /* Turn on busy protocol. */
1369 rxi_KeepAliveOn(call);
1371 /* Attempt MTU discovery */
1372 rxi_GrowMTUOn(call);
1375 * We are no longer the active thread in rx_NewCall
1377 MUTEX_ENTER(&conn->conn_data_lock);
1378 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1379 MUTEX_EXIT(&conn->conn_data_lock);
1382 * Wake up anyone else who might be giving us a chance to
1383 * run (see code above that avoids resource starvation).
1385 #ifdef RX_ENABLE_LOCKS
1386 CV_BROADCAST(&conn->conn_call_cv);
1390 MUTEX_EXIT(&conn->conn_call_lock);
1392 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1393 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1394 osi_Panic("rx_NewCall call about to be used without an empty tq");
1396 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1398 MUTEX_EXIT(&call->lock);
1401 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1406 rxi_HasActiveCalls(struct rx_connection *aconn)
1409 struct rx_call *tcall;
1413 for (i = 0; i < RX_MAXCALLS; i++) {
1414 if ((tcall = aconn->call[i])) {
1415 if ((tcall->state == RX_STATE_ACTIVE)
1416 || (tcall->state == RX_STATE_PRECALL)) {
1427 rxi_GetCallNumberVector(struct rx_connection *aconn,
1428 afs_int32 * aint32s)
1431 struct rx_call *tcall;
1435 for (i = 0; i < RX_MAXCALLS; i++) {
1436 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1437 aint32s[i] = aconn->callNumber[i] + 1;
1439 aint32s[i] = aconn->callNumber[i];
1446 rxi_SetCallNumberVector(struct rx_connection *aconn,
1447 afs_int32 * aint32s)
1450 struct rx_call *tcall;
1454 for (i = 0; i < RX_MAXCALLS; i++) {
1455 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1456 aconn->callNumber[i] = aint32s[i] - 1;
1458 aconn->callNumber[i] = aint32s[i];
1464 /* Advertise a new service. A service is named locally by a UDP port
1465 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1468 char *serviceName; Name for identification purposes (e.g. the
1469 service name might be used for probing for
1472 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1473 char *serviceName, struct rx_securityClass **securityObjects,
1474 int nSecurityObjects,
1475 afs_int32(*serviceProc) (struct rx_call * acall))
1477 osi_socket socket = OSI_NULLSOCKET;
1478 struct rx_service *tservice;
1484 if (serviceId == 0) {
1486 "rx_NewService: service id for service %s is not non-zero.\n",
1493 "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",
1501 tservice = rxi_AllocService();
1504 #ifdef RX_ENABLE_LOCKS
1505 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1508 for (i = 0; i < RX_MAX_SERVICES; i++) {
1509 struct rx_service *service = rx_services[i];
1511 if (port == service->servicePort && host == service->serviceHost) {
1512 if (service->serviceId == serviceId) {
1513 /* The identical service has already been
1514 * installed; if the caller was intending to
1515 * change the security classes used by this
1516 * service, he/she loses. */
1518 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1519 serviceName, serviceId, service->serviceName);
1521 rxi_FreeService(tservice);
1524 /* Different service, same port: re-use the socket
1525 * which is bound to the same port */
1526 socket = service->socket;
1529 if (socket == OSI_NULLSOCKET) {
1530 /* If we don't already have a socket (from another
1531 * service on same port) get a new one */
1532 socket = rxi_GetHostUDPSocket(host, port);
1533 if (socket == OSI_NULLSOCKET) {
1535 rxi_FreeService(tservice);
1540 service->socket = socket;
1541 service->serviceHost = host;
1542 service->servicePort = port;
1543 service->serviceId = serviceId;
1544 service->serviceName = serviceName;
1545 service->nSecurityObjects = nSecurityObjects;
1546 service->securityObjects = securityObjects;
1547 service->minProcs = 0;
1548 service->maxProcs = 1;
1549 service->idleDeadTime = 60;
1550 service->idleDeadErr = 0;
1551 service->connDeadTime = rx_connDeadTime;
1552 service->executeRequestProc = serviceProc;
1553 service->checkReach = 0;
1554 service->nSpecific = 0;
1555 service->specific = NULL;
1556 rx_services[i] = service; /* not visible until now */
1562 rxi_FreeService(tservice);
1563 (osi_Msg "rx_NewService: cannot support > %d services\n",
1568 /* Set configuration options for all of a service's security objects */
1571 rx_SetSecurityConfiguration(struct rx_service *service,
1572 rx_securityConfigVariables type,
1576 for (i = 0; i<service->nSecurityObjects; i++) {
1577 if (service->securityObjects[i]) {
1578 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1586 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1587 struct rx_securityClass **securityObjects, int nSecurityObjects,
1588 afs_int32(*serviceProc) (struct rx_call * acall))
1590 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1593 /* Generic request processing loop. This routine should be called
1594 * by the implementation dependent rx_ServerProc. If socketp is
1595 * non-null, it will be set to the file descriptor that this thread
1596 * is now listening on. If socketp is null, this routine will never
1599 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1601 struct rx_call *call;
1603 struct rx_service *tservice = NULL;
1610 call = rx_GetCall(threadID, tservice, socketp);
1611 if (socketp && *socketp != OSI_NULLSOCKET) {
1612 /* We are now a listener thread */
1617 /* if server is restarting( typically smooth shutdown) then do not
1618 * allow any new calls.
1621 if (rx_tranquil && (call != NULL)) {
1625 MUTEX_ENTER(&call->lock);
1627 rxi_CallError(call, RX_RESTARTING);
1628 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1630 MUTEX_EXIT(&call->lock);
1634 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1635 #ifdef RX_ENABLE_LOCKS
1637 #endif /* RX_ENABLE_LOCKS */
1638 afs_termState = AFSOP_STOP_AFS;
1639 afs_osi_Wakeup(&afs_termState);
1640 #ifdef RX_ENABLE_LOCKS
1642 #endif /* RX_ENABLE_LOCKS */
1647 tservice = call->conn->service;
1649 if (tservice->beforeProc)
1650 (*tservice->beforeProc) (call);
1652 code = tservice->executeRequestProc(call);
1654 if (tservice->afterProc)
1655 (*tservice->afterProc) (call, code);
1657 rx_EndCall(call, code);
1658 if (rx_stats_active) {
1659 MUTEX_ENTER(&rx_stats_mutex);
1661 MUTEX_EXIT(&rx_stats_mutex);
1668 rx_WakeupServerProcs(void)
1670 struct rx_serverQueueEntry *np, *tqp;
1674 MUTEX_ENTER(&rx_serverPool_lock);
1676 #ifdef RX_ENABLE_LOCKS
1677 if (rx_waitForPacket)
1678 CV_BROADCAST(&rx_waitForPacket->cv);
1679 #else /* RX_ENABLE_LOCKS */
1680 if (rx_waitForPacket)
1681 osi_rxWakeup(rx_waitForPacket);
1682 #endif /* RX_ENABLE_LOCKS */
1683 MUTEX_ENTER(&freeSQEList_lock);
1684 for (np = rx_FreeSQEList; np; np = tqp) {
1685 tqp = *(struct rx_serverQueueEntry **)np;
1686 #ifdef RX_ENABLE_LOCKS
1687 CV_BROADCAST(&np->cv);
1688 #else /* RX_ENABLE_LOCKS */
1690 #endif /* RX_ENABLE_LOCKS */
1692 MUTEX_EXIT(&freeSQEList_lock);
1693 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1694 #ifdef RX_ENABLE_LOCKS
1695 CV_BROADCAST(&np->cv);
1696 #else /* RX_ENABLE_LOCKS */
1698 #endif /* RX_ENABLE_LOCKS */
1700 MUTEX_EXIT(&rx_serverPool_lock);
1705 * One thing that seems to happen is that all the server threads get
1706 * tied up on some empty or slow call, and then a whole bunch of calls
1707 * arrive at once, using up the packet pool, so now there are more
1708 * empty calls. The most critical resources here are server threads
1709 * and the free packet pool. The "doreclaim" code seems to help in
1710 * general. I think that eventually we arrive in this state: there
1711 * are lots of pending calls which do have all their packets present,
1712 * so they won't be reclaimed, are multi-packet calls, so they won't
1713 * be scheduled until later, and thus are tying up most of the free
1714 * packet pool for a very long time.
1716 * 1. schedule multi-packet calls if all the packets are present.
1717 * Probably CPU-bound operation, useful to return packets to pool.
1718 * Do what if there is a full window, but the last packet isn't here?
1719 * 3. preserve one thread which *only* runs "best" calls, otherwise
1720 * it sleeps and waits for that type of call.
1721 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1722 * the current dataquota business is badly broken. The quota isn't adjusted
1723 * to reflect how many packets are presently queued for a running call.
1724 * So, when we schedule a queued call with a full window of packets queued
1725 * up for it, that *should* free up a window full of packets for other 2d-class
1726 * calls to be able to use from the packet pool. But it doesn't.
1728 * NB. Most of the time, this code doesn't run -- since idle server threads
1729 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1730 * as a new call arrives.
1732 /* Sleep until a call arrives. Returns a pointer to the call, ready
1733 * for an rx_Read. */
1734 #ifdef RX_ENABLE_LOCKS
1736 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1738 struct rx_serverQueueEntry *sq;
1739 struct rx_call *call = (struct rx_call *)0;
1740 struct rx_service *service = NULL;
1743 MUTEX_ENTER(&freeSQEList_lock);
1745 if ((sq = rx_FreeSQEList)) {
1746 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1747 MUTEX_EXIT(&freeSQEList_lock);
1748 } else { /* otherwise allocate a new one and return that */
1749 MUTEX_EXIT(&freeSQEList_lock);
1750 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1751 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1752 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1755 MUTEX_ENTER(&rx_serverPool_lock);
1756 if (cur_service != NULL) {
1757 ReturnToServerPool(cur_service);
1760 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1761 struct rx_call *tcall, *ncall, *choice2 = NULL;
1763 /* Scan for eligible incoming calls. A call is not eligible
1764 * if the maximum number of calls for its service type are
1765 * already executing */
1766 /* One thread will process calls FCFS (to prevent starvation),
1767 * while the other threads may run ahead looking for calls which
1768 * have all their input data available immediately. This helps
1769 * keep threads from blocking, waiting for data from the client. */
1770 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1771 service = tcall->conn->service;
1772 if (!QuotaOK(service)) {
1775 MUTEX_ENTER(&rx_pthread_mutex);
1776 if (tno == rxi_fcfs_thread_num
1777 || !tcall->queue_item_header.next) {
1778 MUTEX_EXIT(&rx_pthread_mutex);
1779 /* If we're the fcfs thread , then we'll just use
1780 * this call. If we haven't been able to find an optimal
1781 * choice, and we're at the end of the list, then use a
1782 * 2d choice if one has been identified. Otherwise... */
1783 call = (choice2 ? choice2 : tcall);
1784 service = call->conn->service;
1786 MUTEX_EXIT(&rx_pthread_mutex);
1787 if (!queue_IsEmpty(&tcall->rq)) {
1788 struct rx_packet *rp;
1789 rp = queue_First(&tcall->rq, rx_packet);
1790 if (rp->header.seq == 1) {
1792 || (rp->header.flags & RX_LAST_PACKET)) {
1794 } else if (rxi_2dchoice && !choice2
1795 && !(tcall->flags & RX_CALL_CLEARED)
1796 && (tcall->rprev > rxi_HardAckRate)) {
1806 ReturnToServerPool(service);
1813 MUTEX_EXIT(&rx_serverPool_lock);
1814 MUTEX_ENTER(&call->lock);
1816 if (call->flags & RX_CALL_WAIT_PROC) {
1817 call->flags &= ~RX_CALL_WAIT_PROC;
1818 rx_atomic_dec(&rx_nWaiting);
1821 if (call->state != RX_STATE_PRECALL || call->error) {
1822 MUTEX_EXIT(&call->lock);
1823 MUTEX_ENTER(&rx_serverPool_lock);
1824 ReturnToServerPool(service);
1829 if (queue_IsEmpty(&call->rq)
1830 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1831 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1833 CLEAR_CALL_QUEUE_LOCK(call);
1836 /* If there are no eligible incoming calls, add this process
1837 * to the idle server queue, to wait for one */
1841 *socketp = OSI_NULLSOCKET;
1843 sq->socketp = socketp;
1844 queue_Append(&rx_idleServerQueue, sq);
1845 #ifndef AFS_AIX41_ENV
1846 rx_waitForPacket = sq;
1848 rx_waitingForPacket = sq;
1849 #endif /* AFS_AIX41_ENV */
1851 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1853 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1854 MUTEX_EXIT(&rx_serverPool_lock);
1855 return (struct rx_call *)0;
1858 } while (!(call = sq->newcall)
1859 && !(socketp && *socketp != OSI_NULLSOCKET));
1860 MUTEX_EXIT(&rx_serverPool_lock);
1862 MUTEX_ENTER(&call->lock);
1868 MUTEX_ENTER(&freeSQEList_lock);
1869 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1870 rx_FreeSQEList = sq;
1871 MUTEX_EXIT(&freeSQEList_lock);
1874 clock_GetTime(&call->startTime);
1875 call->state = RX_STATE_ACTIVE;
1876 call->mode = RX_MODE_RECEIVING;
1877 #ifdef RX_KERNEL_TRACE
1878 if (ICL_SETACTIVE(afs_iclSetp)) {
1879 int glockOwner = ISAFS_GLOCK();
1882 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1883 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1890 rxi_calltrace(RX_CALL_START, call);
1891 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
1892 call->conn->service->servicePort, call->conn->service->serviceId,
1895 MUTEX_EXIT(&call->lock);
1896 MUTEX_ENTER(&rx_refcnt_mutex);
1897 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1898 MUTEX_EXIT(&rx_refcnt_mutex);
1900 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1905 #else /* RX_ENABLE_LOCKS */
1907 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1909 struct rx_serverQueueEntry *sq;
1910 struct rx_call *call = (struct rx_call *)0, *choice2;
1911 struct rx_service *service = NULL;
1915 MUTEX_ENTER(&freeSQEList_lock);
1917 if ((sq = rx_FreeSQEList)) {
1918 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1919 MUTEX_EXIT(&freeSQEList_lock);
1920 } else { /* otherwise allocate a new one and return that */
1921 MUTEX_EXIT(&freeSQEList_lock);
1922 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1923 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1924 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1926 MUTEX_ENTER(&sq->lock);
1928 if (cur_service != NULL) {
1929 cur_service->nRequestsRunning--;
1930 MUTEX_ENTER(&rx_quota_mutex);
1931 if (cur_service->nRequestsRunning < cur_service->minProcs)
1934 MUTEX_EXIT(&rx_quota_mutex);
1936 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1937 struct rx_call *tcall, *ncall;
1938 /* Scan for eligible incoming calls. A call is not eligible
1939 * if the maximum number of calls for its service type are
1940 * already executing */
1941 /* One thread will process calls FCFS (to prevent starvation),
1942 * while the other threads may run ahead looking for calls which
1943 * have all their input data available immediately. This helps
1944 * keep threads from blocking, waiting for data from the client. */
1945 choice2 = (struct rx_call *)0;
1946 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1947 service = tcall->conn->service;
1948 if (QuotaOK(service)) {
1949 MUTEX_ENTER(&rx_pthread_mutex);
1950 if (tno == rxi_fcfs_thread_num
1951 || !tcall->queue_item_header.next) {
1952 MUTEX_EXIT(&rx_pthread_mutex);
1953 /* If we're the fcfs thread, then we'll just use
1954 * this call. If we haven't been able to find an optimal
1955 * choice, and we're at the end of the list, then use a
1956 * 2d choice if one has been identified. Otherwise... */
1957 call = (choice2 ? choice2 : tcall);
1958 service = call->conn->service;
1960 MUTEX_EXIT(&rx_pthread_mutex);
1961 if (!queue_IsEmpty(&tcall->rq)) {
1962 struct rx_packet *rp;
1963 rp = queue_First(&tcall->rq, rx_packet);
1964 if (rp->header.seq == 1
1966 || (rp->header.flags & RX_LAST_PACKET))) {
1968 } else if (rxi_2dchoice && !choice2
1969 && !(tcall->flags & RX_CALL_CLEARED)
1970 && (tcall->rprev > rxi_HardAckRate)) {
1984 /* we can't schedule a call if there's no data!!! */
1985 /* send an ack if there's no data, if we're missing the
1986 * first packet, or we're missing something between first
1987 * and last -- there's a "hole" in the incoming data. */
1988 if (queue_IsEmpty(&call->rq)
1989 || queue_First(&call->rq, rx_packet)->header.seq != 1
1990 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1991 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1993 call->flags &= (~RX_CALL_WAIT_PROC);
1994 service->nRequestsRunning++;
1995 /* just started call in minProcs pool, need fewer to maintain
1997 MUTEX_ENTER(&rx_quota_mutex);
1998 if (service->nRequestsRunning <= service->minProcs)
2001 MUTEX_EXIT(&rx_quota_mutex);
2002 rx_atomic_dec(&rx_nWaiting);
2003 /* MUTEX_EXIT(&call->lock); */
2005 /* If there are no eligible incoming calls, add this process
2006 * to the idle server queue, to wait for one */
2009 *socketp = OSI_NULLSOCKET;
2011 sq->socketp = socketp;
2012 queue_Append(&rx_idleServerQueue, sq);
2016 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2018 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2019 return (struct rx_call *)0;
2022 } while (!(call = sq->newcall)
2023 && !(socketp && *socketp != OSI_NULLSOCKET));
2025 MUTEX_EXIT(&sq->lock);
2027 MUTEX_ENTER(&freeSQEList_lock);
2028 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2029 rx_FreeSQEList = sq;
2030 MUTEX_EXIT(&freeSQEList_lock);
2033 clock_GetTime(&call->startTime);
2034 call->state = RX_STATE_ACTIVE;
2035 call->mode = RX_MODE_RECEIVING;
2036 #ifdef RX_KERNEL_TRACE
2037 if (ICL_SETACTIVE(afs_iclSetp)) {
2038 int glockOwner = ISAFS_GLOCK();
2041 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2042 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2049 rxi_calltrace(RX_CALL_START, call);
2050 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2051 call->conn->service->servicePort, call->conn->service->serviceId,
2054 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2061 #endif /* RX_ENABLE_LOCKS */
2065 /* Establish a procedure to be called when a packet arrives for a
2066 * call. This routine will be called at most once after each call,
2067 * and will also be called if there is an error condition on the or
2068 * the call is complete. Used by multi rx to build a selection
2069 * function which determines which of several calls is likely to be a
2070 * good one to read from.
2071 * NOTE: the way this is currently implemented it is probably only a
2072 * good idea to (1) use it immediately after a newcall (clients only)
2073 * and (2) only use it once. Other uses currently void your warranty
2076 rx_SetArrivalProc(struct rx_call *call,
2077 void (*proc) (struct rx_call * call,
2080 void * handle, int arg)
2082 call->arrivalProc = proc;
2083 call->arrivalProcHandle = handle;
2084 call->arrivalProcArg = arg;
2087 /* Call is finished (possibly prematurely). Return rc to the peer, if
2088 * appropriate, and return the final error code from the conversation
2092 rx_EndCall(struct rx_call *call, afs_int32 rc)
2094 struct rx_connection *conn = call->conn;
2098 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2099 call, rc, call->error, call->abortCode));
2102 MUTEX_ENTER(&call->lock);
2104 if (rc == 0 && call->error == 0) {
2105 call->abortCode = 0;
2106 call->abortCount = 0;
2109 call->arrivalProc = (void (*)())0;
2110 if (rc && call->error == 0) {
2111 rxi_CallError(call, rc);
2112 call->mode = RX_MODE_ERROR;
2113 /* Send an abort message to the peer if this error code has
2114 * only just been set. If it was set previously, assume the
2115 * peer has already been sent the error code or will request it
2117 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2119 if (conn->type == RX_SERVER_CONNECTION) {
2120 /* Make sure reply or at least dummy reply is sent */
2121 if (call->mode == RX_MODE_RECEIVING) {
2122 MUTEX_EXIT(&call->lock);
2123 rxi_WriteProc(call, 0, 0);
2124 MUTEX_ENTER(&call->lock);
2126 if (call->mode == RX_MODE_SENDING) {
2127 MUTEX_EXIT(&call->lock);
2128 rxi_FlushWrite(call);
2129 MUTEX_ENTER(&call->lock);
2131 rxi_calltrace(RX_CALL_END, call);
2132 /* Call goes to hold state until reply packets are acknowledged */
2133 if (call->tfirst + call->nSoftAcked < call->tnext) {
2134 call->state = RX_STATE_HOLD;
2136 call->state = RX_STATE_DALLY;
2137 rxi_ClearTransmitQueue(call, 0);
2138 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
2139 rxevent_Cancel(call->keepAliveEvent, call,
2140 RX_CALL_REFCOUNT_ALIVE);
2142 } else { /* Client connection */
2144 /* Make sure server receives input packets, in the case where
2145 * no reply arguments are expected */
2146 if ((call->mode == RX_MODE_SENDING)
2147 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2148 MUTEX_EXIT(&call->lock);
2149 (void)rxi_ReadProc(call, &dummy, 1);
2150 MUTEX_ENTER(&call->lock);
2153 /* If we had an outstanding delayed ack, be nice to the server
2154 * and force-send it now.
2156 if (call->delayedAckEvent) {
2157 rxevent_Cancel(call->delayedAckEvent, call,
2158 RX_CALL_REFCOUNT_DELAY);
2159 call->delayedAckEvent = NULL;
2160 rxi_SendDelayedAck(NULL, call, NULL);
2163 /* We need to release the call lock since it's lower than the
2164 * conn_call_lock and we don't want to hold the conn_call_lock
2165 * over the rx_ReadProc call. The conn_call_lock needs to be held
2166 * here for the case where rx_NewCall is perusing the calls on
2167 * the connection structure. We don't want to signal until
2168 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2169 * have checked this call, found it active and by the time it
2170 * goes to sleep, will have missed the signal.
2172 MUTEX_EXIT(&call->lock);
2173 MUTEX_ENTER(&conn->conn_call_lock);
2174 MUTEX_ENTER(&call->lock);
2175 MUTEX_ENTER(&conn->conn_data_lock);
2176 conn->flags |= RX_CONN_BUSY;
2177 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2178 MUTEX_EXIT(&conn->conn_data_lock);
2179 #ifdef RX_ENABLE_LOCKS
2180 CV_BROADCAST(&conn->conn_call_cv);
2185 #ifdef RX_ENABLE_LOCKS
2187 MUTEX_EXIT(&conn->conn_data_lock);
2189 #endif /* RX_ENABLE_LOCKS */
2190 call->state = RX_STATE_DALLY;
2192 error = call->error;
2194 /* currentPacket, nLeft, and NFree must be zeroed here, because
2195 * ResetCall cannot: ResetCall may be called at splnet(), in the
2196 * kernel version, and may interrupt the macros rx_Read or
2197 * rx_Write, which run at normal priority for efficiency. */
2198 if (call->currentPacket) {
2199 #ifdef RX_TRACK_PACKETS
2200 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2202 rxi_FreePacket(call->currentPacket);
2203 call->currentPacket = (struct rx_packet *)0;
2206 call->nLeft = call->nFree = call->curlen = 0;
2208 /* Free any packets from the last call to ReadvProc/WritevProc */
2209 #ifdef RXDEBUG_PACKET
2211 #endif /* RXDEBUG_PACKET */
2212 rxi_FreePackets(0, &call->iovq);
2213 MUTEX_EXIT(&call->lock);
2215 MUTEX_ENTER(&rx_refcnt_mutex);
2216 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2217 MUTEX_EXIT(&rx_refcnt_mutex);
2218 if (conn->type == RX_CLIENT_CONNECTION) {
2219 MUTEX_ENTER(&conn->conn_data_lock);
2220 conn->flags &= ~RX_CONN_BUSY;
2221 MUTEX_EXIT(&conn->conn_data_lock);
2222 MUTEX_EXIT(&conn->conn_call_lock);
2226 * Map errors to the local host's errno.h format.
2228 error = ntoh_syserr_conv(error);
2232 #if !defined(KERNEL)
2234 /* Call this routine when shutting down a server or client (especially
2235 * clients). This will allow Rx to gracefully garbage collect server
2236 * connections, and reduce the number of retries that a server might
2237 * make to a dead client.
2238 * This is not quite right, since some calls may still be ongoing and
2239 * we can't lock them to destroy them. */
2243 struct rx_connection **conn_ptr, **conn_end;
2247 if (rxinit_status == 1) {
2249 return; /* Already shutdown. */
2251 rxi_DeleteCachedConnections();
2252 if (rx_connHashTable) {
2253 MUTEX_ENTER(&rx_connHashTable_lock);
2254 for (conn_ptr = &rx_connHashTable[0], conn_end =
2255 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2257 struct rx_connection *conn, *next;
2258 for (conn = *conn_ptr; conn; conn = next) {
2260 if (conn->type == RX_CLIENT_CONNECTION) {
2261 MUTEX_ENTER(&rx_refcnt_mutex);
2263 MUTEX_EXIT(&rx_refcnt_mutex);
2264 #ifdef RX_ENABLE_LOCKS
2265 rxi_DestroyConnectionNoLock(conn);
2266 #else /* RX_ENABLE_LOCKS */
2267 rxi_DestroyConnection(conn);
2268 #endif /* RX_ENABLE_LOCKS */
2272 #ifdef RX_ENABLE_LOCKS
2273 while (rx_connCleanup_list) {
2274 struct rx_connection *conn;
2275 conn = rx_connCleanup_list;
2276 rx_connCleanup_list = rx_connCleanup_list->next;
2277 MUTEX_EXIT(&rx_connHashTable_lock);
2278 rxi_CleanupConnection(conn);
2279 MUTEX_ENTER(&rx_connHashTable_lock);
2281 MUTEX_EXIT(&rx_connHashTable_lock);
2282 #endif /* RX_ENABLE_LOCKS */
2287 afs_winsockCleanup();
2295 /* if we wakeup packet waiter too often, can get in loop with two
2296 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2298 rxi_PacketsUnWait(void)
2300 if (!rx_waitingForPackets) {
2304 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2305 return; /* still over quota */
2308 rx_waitingForPackets = 0;
2309 #ifdef RX_ENABLE_LOCKS
2310 CV_BROADCAST(&rx_waitingForPackets_cv);
2312 osi_rxWakeup(&rx_waitingForPackets);
2318 /* ------------------Internal interfaces------------------------- */
2320 /* Return this process's service structure for the
2321 * specified socket and service */
2323 rxi_FindService(osi_socket socket, u_short serviceId)
2325 struct rx_service **sp;
2326 for (sp = &rx_services[0]; *sp; sp++) {
2327 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2333 #ifdef RXDEBUG_PACKET
2334 #ifdef KDUMP_RX_LOCK
2335 static struct rx_call_rx_lock *rx_allCallsp = 0;
2337 static struct rx_call *rx_allCallsp = 0;
2339 #endif /* RXDEBUG_PACKET */
2341 /* Allocate a call structure, for the indicated channel of the
2342 * supplied connection. The mode and state of the call must be set by
2343 * the caller. Returns the call with mutex locked. */
2345 rxi_NewCall(struct rx_connection *conn, int channel)
2347 struct rx_call *call;
2348 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2349 struct rx_call *cp; /* Call pointer temp */
2350 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2351 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2353 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2355 /* Grab an existing call structure, or allocate a new one.
2356 * Existing call structures are assumed to have been left reset by
2358 MUTEX_ENTER(&rx_freeCallQueue_lock);
2360 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2362 * EXCEPT that the TQ might not yet be cleared out.
2363 * Skip over those with in-use TQs.
2366 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2367 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2373 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2374 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2375 call = queue_First(&rx_freeCallQueue, rx_call);
2376 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2378 if (rx_stats_active)
2379 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2380 MUTEX_EXIT(&rx_freeCallQueue_lock);
2381 MUTEX_ENTER(&call->lock);
2382 CLEAR_CALL_QUEUE_LOCK(call);
2383 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2384 /* Now, if TQ wasn't cleared earlier, do it now. */
2385 rxi_WaitforTQBusy(call);
2386 if (call->flags & RX_CALL_TQ_CLEARME) {
2387 rxi_ClearTransmitQueue(call, 1);
2388 /*queue_Init(&call->tq);*/
2390 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2391 /* Bind the call to its connection structure */
2393 rxi_ResetCall(call, 1);
2396 call = rxi_Alloc(sizeof(struct rx_call));
2397 #ifdef RXDEBUG_PACKET
2398 call->allNextp = rx_allCallsp;
2399 rx_allCallsp = call;
2401 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2402 #else /* RXDEBUG_PACKET */
2403 rx_atomic_inc(&rx_stats.nCallStructs);
2404 #endif /* RXDEBUG_PACKET */
2406 MUTEX_EXIT(&rx_freeCallQueue_lock);
2407 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2408 MUTEX_ENTER(&call->lock);
2409 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2410 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2411 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2413 /* Initialize once-only items */
2414 queue_Init(&call->tq);
2415 queue_Init(&call->rq);
2416 queue_Init(&call->iovq);
2417 #ifdef RXDEBUG_PACKET
2418 call->rqc = call->tqc = call->iovqc = 0;
2419 #endif /* RXDEBUG_PACKET */
2420 /* Bind the call to its connection structure (prereq for reset) */
2422 rxi_ResetCall(call, 1);
2424 call->channel = channel;
2425 call->callNumber = &conn->callNumber[channel];
2426 call->rwind = conn->rwind[channel];
2427 call->twind = conn->twind[channel];
2428 /* Note that the next expected call number is retained (in
2429 * conn->callNumber[i]), even if we reallocate the call structure
2431 conn->call[channel] = call;
2432 /* if the channel's never been used (== 0), we should start at 1, otherwise
2433 * the call number is valid from the last time this channel was used */
2434 if (*call->callNumber == 0)
2435 *call->callNumber = 1;
2440 /* A call has been inactive long enough that so we can throw away
2441 * state, including the call structure, which is placed on the call
2444 * call->lock amd rx_refcnt_mutex are held upon entry.
2445 * haveCTLock is set when called from rxi_ReapConnections.
2448 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2450 int channel = call->channel;
2451 struct rx_connection *conn = call->conn;
2454 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2455 (*call->callNumber)++;
2457 * We are setting the state to RX_STATE_RESET to
2458 * ensure that no one else will attempt to use this
2459 * call once we drop the refcnt lock. We must drop
2460 * the refcnt lock before calling rxi_ResetCall
2461 * because it cannot be held across acquiring the
2462 * freepktQ lock. NewCall does the same.
2464 call->state = RX_STATE_RESET;
2465 MUTEX_EXIT(&rx_refcnt_mutex);
2466 rxi_ResetCall(call, 0);
2467 call->conn->call[channel] = (struct rx_call *)0;
2469 MUTEX_ENTER(&rx_freeCallQueue_lock);
2470 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2471 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2472 /* A call may be free even though its transmit queue is still in use.
2473 * Since we search the call list from head to tail, put busy calls at
2474 * the head of the list, and idle calls at the tail.
2476 if (call->flags & RX_CALL_TQ_BUSY)
2477 queue_Prepend(&rx_freeCallQueue, call);
2479 queue_Append(&rx_freeCallQueue, call);
2480 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2481 queue_Append(&rx_freeCallQueue, call);
2482 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2483 if (rx_stats_active)
2484 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2485 MUTEX_EXIT(&rx_freeCallQueue_lock);
2487 /* Destroy the connection if it was previously slated for
2488 * destruction, i.e. the Rx client code previously called
2489 * rx_DestroyConnection (client connections), or
2490 * rxi_ReapConnections called the same routine (server
2491 * connections). Only do this, however, if there are no
2492 * outstanding calls. Note that for fine grain locking, there appears
2493 * to be a deadlock in that rxi_FreeCall has a call locked and
2494 * DestroyConnectionNoLock locks each call in the conn. But note a
2495 * few lines up where we have removed this call from the conn.
2496 * If someone else destroys a connection, they either have no
2497 * call lock held or are going through this section of code.
2499 MUTEX_ENTER(&conn->conn_data_lock);
2500 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2501 MUTEX_ENTER(&rx_refcnt_mutex);
2503 MUTEX_EXIT(&rx_refcnt_mutex);
2504 MUTEX_EXIT(&conn->conn_data_lock);
2505 #ifdef RX_ENABLE_LOCKS
2507 rxi_DestroyConnectionNoLock(conn);
2509 rxi_DestroyConnection(conn);
2510 #else /* RX_ENABLE_LOCKS */
2511 rxi_DestroyConnection(conn);
2512 #endif /* RX_ENABLE_LOCKS */
2514 MUTEX_EXIT(&conn->conn_data_lock);
2516 MUTEX_ENTER(&rx_refcnt_mutex);
2519 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2520 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2523 rxi_Alloc(size_t size)
2527 if (rx_stats_active) {
2528 rx_atomic_add(&rxi_Allocsize, (int) size);
2529 rx_atomic_inc(&rxi_Alloccnt);
2533 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2534 afs_osi_Alloc_NoSleep(size);
2539 osi_Panic("rxi_Alloc error");
2545 rxi_Free(void *addr, size_t size)
2547 if (rx_stats_active) {
2548 rx_atomic_sub(&rxi_Allocsize, (int) size);
2549 rx_atomic_dec(&rxi_Alloccnt);
2551 osi_Free(addr, size);
2555 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2557 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2558 struct rx_peer *next = NULL;
2562 MUTEX_ENTER(&rx_peerHashTable_lock);
2564 peer_ptr = &rx_peerHashTable[0];
2565 peer_end = &rx_peerHashTable[rx_hashTableSize];
2568 for ( ; peer_ptr < peer_end; peer_ptr++) {
2571 for ( ; peer; peer = next) {
2573 if (host == peer->host)
2578 hashIndex = PEER_HASH(host, port);
2579 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2580 if ((peer->host == host) && (peer->port == port))
2585 MUTEX_ENTER(&rx_peerHashTable_lock);
2590 MUTEX_EXIT(&rx_peerHashTable_lock);
2592 MUTEX_ENTER(&peer->peer_lock);
2593 /* We don't handle dropping below min, so don't */
2594 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2595 peer->ifMTU=MIN(mtu, peer->ifMTU);
2596 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2597 /* if we tweaked this down, need to tune our peer MTU too */
2598 peer->MTU = MIN(peer->MTU, peer->natMTU);
2599 /* if we discovered a sub-1500 mtu, degrade */
2600 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2601 peer->maxDgramPackets = 1;
2602 /* We no longer have valid peer packet information */
2603 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2604 peer->maxPacketSize = 0;
2605 MUTEX_EXIT(&peer->peer_lock);
2607 MUTEX_ENTER(&rx_peerHashTable_lock);
2609 if (host && !port) {
2611 /* pick up where we left off */
2615 MUTEX_EXIT(&rx_peerHashTable_lock);
2618 /* Find the peer process represented by the supplied (host,port)
2619 * combination. If there is no appropriate active peer structure, a
2620 * new one will be allocated and initialized
2621 * The origPeer, if set, is a pointer to a peer structure on which the
2622 * refcount will be be decremented. This is used to replace the peer
2623 * structure hanging off a connection structure */
2625 rxi_FindPeer(afs_uint32 host, u_short port,
2626 struct rx_peer *origPeer, int create)
2630 hashIndex = PEER_HASH(host, port);
2631 MUTEX_ENTER(&rx_peerHashTable_lock);
2632 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2633 if ((pp->host == host) && (pp->port == port))
2638 pp = rxi_AllocPeer(); /* This bzero's *pp */
2639 pp->host = host; /* set here or in InitPeerParams is zero */
2641 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2642 queue_Init(&pp->congestionQueue);
2643 queue_Init(&pp->rpcStats);
2644 pp->next = rx_peerHashTable[hashIndex];
2645 rx_peerHashTable[hashIndex] = pp;
2646 rxi_InitPeerParams(pp);
2647 if (rx_stats_active)
2648 rx_atomic_inc(&rx_stats.nPeerStructs);
2655 origPeer->refCount--;
2656 MUTEX_EXIT(&rx_peerHashTable_lock);
2661 /* Find the connection at (host, port) started at epoch, and with the
2662 * given connection id. Creates the server connection if necessary.
2663 * The type specifies whether a client connection or a server
2664 * connection is desired. In both cases, (host, port) specify the
2665 * peer's (host, pair) pair. Client connections are not made
2666 * automatically by this routine. The parameter socket gives the
2667 * socket descriptor on which the packet was received. This is used,
2668 * in the case of server connections, to check that *new* connections
2669 * come via a valid (port, serviceId). Finally, the securityIndex
2670 * parameter must match the existing index for the connection. If a
2671 * server connection is created, it will be created using the supplied
2672 * index, if the index is valid for this service */
2673 struct rx_connection *
2674 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2675 u_short port, u_short serviceId, afs_uint32 cid,
2676 afs_uint32 epoch, int type, u_int securityIndex)
2678 int hashindex, flag, i;
2679 struct rx_connection *conn;
2680 hashindex = CONN_HASH(host, port, cid, epoch, type);
2681 MUTEX_ENTER(&rx_connHashTable_lock);
2682 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2683 rx_connHashTable[hashindex],
2686 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2687 && (epoch == conn->epoch)) {
2688 struct rx_peer *pp = conn->peer;
2689 if (securityIndex != conn->securityIndex) {
2690 /* this isn't supposed to happen, but someone could forge a packet
2691 * like this, and there seems to be some CM bug that makes this
2692 * happen from time to time -- in which case, the fileserver
2694 MUTEX_EXIT(&rx_connHashTable_lock);
2695 return (struct rx_connection *)0;
2697 if (pp->host == host && pp->port == port)
2699 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2701 /* So what happens when it's a callback connection? */
2702 if ( /*type == RX_CLIENT_CONNECTION && */
2703 (conn->epoch & 0x80000000))
2707 /* the connection rxLastConn that was used the last time is not the
2708 ** one we are looking for now. Hence, start searching in the hash */
2710 conn = rx_connHashTable[hashindex];
2715 struct rx_service *service;
2716 if (type == RX_CLIENT_CONNECTION) {
2717 MUTEX_EXIT(&rx_connHashTable_lock);
2718 return (struct rx_connection *)0;
2720 service = rxi_FindService(socket, serviceId);
2721 if (!service || (securityIndex >= service->nSecurityObjects)
2722 || (service->securityObjects[securityIndex] == 0)) {
2723 MUTEX_EXIT(&rx_connHashTable_lock);
2724 return (struct rx_connection *)0;
2726 conn = rxi_AllocConnection(); /* This bzero's the connection */
2727 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2728 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2729 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2730 conn->next = rx_connHashTable[hashindex];
2731 rx_connHashTable[hashindex] = conn;
2732 conn->peer = rxi_FindPeer(host, port, 0, 1);
2733 conn->type = RX_SERVER_CONNECTION;
2734 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2735 conn->epoch = epoch;
2736 conn->cid = cid & RX_CIDMASK;
2737 /* conn->serial = conn->lastSerial = 0; */
2738 /* conn->timeout = 0; */
2739 conn->ackRate = RX_FAST_ACK_RATE;
2740 conn->service = service;
2741 conn->serviceId = serviceId;
2742 conn->securityIndex = securityIndex;
2743 conn->securityObject = service->securityObjects[securityIndex];
2744 conn->nSpecific = 0;
2745 conn->specific = NULL;
2746 rx_SetConnDeadTime(conn, service->connDeadTime);
2747 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2748 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2749 for (i = 0; i < RX_MAXCALLS; i++) {
2750 conn->twind[i] = rx_initSendWindow;
2751 conn->rwind[i] = rx_initReceiveWindow;
2753 /* Notify security object of the new connection */
2754 RXS_NewConnection(conn->securityObject, conn);
2755 /* XXXX Connection timeout? */
2756 if (service->newConnProc)
2757 (*service->newConnProc) (conn);
2758 if (rx_stats_active)
2759 rx_atomic_inc(&rx_stats.nServerConns);
2762 MUTEX_ENTER(&rx_refcnt_mutex);
2764 MUTEX_EXIT(&rx_refcnt_mutex);
2766 rxLastConn = conn; /* store this connection as the last conn used */
2767 MUTEX_EXIT(&rx_connHashTable_lock);
2771 /* There are two packet tracing routines available for testing and monitoring
2772 * Rx. One is called just after every packet is received and the other is
2773 * called just before every packet is sent. Received packets, have had their
2774 * headers decoded, and packets to be sent have not yet had their headers
2775 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2776 * containing the network address. Both can be modified. The return value, if
2777 * non-zero, indicates that the packet should be dropped. */
2779 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2780 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2782 /* A packet has been received off the interface. Np is the packet, socket is
2783 * the socket number it was received from (useful in determining which service
2784 * this packet corresponds to), and (host, port) reflect the host,port of the
2785 * sender. This call returns the packet to the caller if it is finished with
2786 * it, rather than de-allocating it, just as a small performance hack */
2789 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
2790 afs_uint32 host, u_short port, int *tnop,
2791 struct rx_call **newcallp)
2793 struct rx_call *call;
2794 struct rx_connection *conn;
2796 afs_uint32 currentCallNumber;
2802 struct rx_packet *tnp;
2805 /* We don't print out the packet until now because (1) the time may not be
2806 * accurate enough until now in the lwp implementation (rx_Listener only gets
2807 * the time after the packet is read) and (2) from a protocol point of view,
2808 * this is the first time the packet has been seen */
2809 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2810 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2811 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
2812 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2813 np->header.epoch, np->header.cid, np->header.callNumber,
2814 np->header.seq, np->header.flags, np));
2817 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2818 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2821 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2822 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2825 /* If an input tracer function is defined, call it with the packet and
2826 * network address. Note this function may modify its arguments. */
2827 if (rx_justReceived) {
2828 struct sockaddr_in addr;
2830 addr.sin_family = AF_INET;
2831 addr.sin_port = port;
2832 addr.sin_addr.s_addr = host;
2833 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2834 addr.sin_len = sizeof(addr);
2835 #endif /* AFS_OSF_ENV */
2836 drop = (*rx_justReceived) (np, &addr);
2837 /* drop packet if return value is non-zero */
2840 port = addr.sin_port; /* in case fcn changed addr */
2841 host = addr.sin_addr.s_addr;
2845 /* If packet was not sent by the client, then *we* must be the client */
2846 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2847 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2849 /* Find the connection (or fabricate one, if we're the server & if
2850 * necessary) associated with this packet */
2852 rxi_FindConnection(socket, host, port, np->header.serviceId,
2853 np->header.cid, np->header.epoch, type,
2854 np->header.securityIndex);
2857 /* If no connection found or fabricated, just ignore the packet.
2858 * (An argument could be made for sending an abort packet for
2863 MUTEX_ENTER(&conn->conn_data_lock);
2864 if (conn->maxSerial < np->header.serial)
2865 conn->maxSerial = np->header.serial;
2866 MUTEX_EXIT(&conn->conn_data_lock);
2868 /* If the connection is in an error state, send an abort packet and ignore
2869 * the incoming packet */
2871 /* Don't respond to an abort packet--we don't want loops! */
2872 MUTEX_ENTER(&conn->conn_data_lock);
2873 if (np->header.type != RX_PACKET_TYPE_ABORT)
2874 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2875 MUTEX_ENTER(&rx_refcnt_mutex);
2877 MUTEX_EXIT(&rx_refcnt_mutex);
2878 MUTEX_EXIT(&conn->conn_data_lock);
2882 /* Check for connection-only requests (i.e. not call specific). */
2883 if (np->header.callNumber == 0) {
2884 switch (np->header.type) {
2885 case RX_PACKET_TYPE_ABORT: {
2886 /* What if the supplied error is zero? */
2887 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2888 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
2889 rxi_ConnectionError(conn, errcode);
2890 MUTEX_ENTER(&rx_refcnt_mutex);
2892 MUTEX_EXIT(&rx_refcnt_mutex);
2895 case RX_PACKET_TYPE_CHALLENGE:
2896 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2897 MUTEX_ENTER(&rx_refcnt_mutex);
2899 MUTEX_EXIT(&rx_refcnt_mutex);
2901 case RX_PACKET_TYPE_RESPONSE:
2902 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2903 MUTEX_ENTER(&rx_refcnt_mutex);
2905 MUTEX_EXIT(&rx_refcnt_mutex);
2907 case RX_PACKET_TYPE_PARAMS:
2908 case RX_PACKET_TYPE_PARAMS + 1:
2909 case RX_PACKET_TYPE_PARAMS + 2:
2910 /* ignore these packet types for now */
2911 MUTEX_ENTER(&rx_refcnt_mutex);
2913 MUTEX_EXIT(&rx_refcnt_mutex);
2918 /* Should not reach here, unless the peer is broken: send an
2920 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2921 MUTEX_ENTER(&conn->conn_data_lock);
2922 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2923 MUTEX_ENTER(&rx_refcnt_mutex);
2925 MUTEX_EXIT(&rx_refcnt_mutex);
2926 MUTEX_EXIT(&conn->conn_data_lock);
2931 channel = np->header.cid & RX_CHANNELMASK;
2932 call = conn->call[channel];
2933 #ifdef RX_ENABLE_LOCKS
2935 MUTEX_ENTER(&call->lock);
2936 /* Test to see if call struct is still attached to conn. */
2937 if (call != conn->call[channel]) {
2939 MUTEX_EXIT(&call->lock);
2940 if (type == RX_SERVER_CONNECTION) {
2941 call = conn->call[channel];
2942 /* If we started with no call attached and there is one now,
2943 * another thread is also running this routine and has gotten
2944 * the connection channel. We should drop this packet in the tests
2945 * below. If there was a call on this connection and it's now
2946 * gone, then we'll be making a new call below.
2947 * If there was previously a call and it's now different then
2948 * the old call was freed and another thread running this routine
2949 * has created a call on this channel. One of these two threads
2950 * has a packet for the old call and the code below handles those
2954 MUTEX_ENTER(&call->lock);
2956 /* This packet can't be for this call. If the new call address is
2957 * 0 then no call is running on this channel. If there is a call
2958 * then, since this is a client connection we're getting data for
2959 * it must be for the previous call.
2961 if (rx_stats_active)
2962 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
2963 MUTEX_ENTER(&rx_refcnt_mutex);
2965 MUTEX_EXIT(&rx_refcnt_mutex);
2970 currentCallNumber = conn->callNumber[channel];
2972 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2973 if (np->header.callNumber < currentCallNumber) {
2974 if (rx_stats_active)
2975 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
2976 #ifdef RX_ENABLE_LOCKS
2978 MUTEX_EXIT(&call->lock);
2980 MUTEX_ENTER(&rx_refcnt_mutex);
2982 MUTEX_EXIT(&rx_refcnt_mutex);
2986 MUTEX_ENTER(&conn->conn_call_lock);
2987 call = rxi_NewCall(conn, channel);
2988 MUTEX_EXIT(&conn->conn_call_lock);
2989 *call->callNumber = np->header.callNumber;
2991 if (np->header.callNumber == 0)
2992 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",
2993 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
2994 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
2995 np->header.flags, np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2997 call->state = RX_STATE_PRECALL;
2998 clock_GetTime(&call->queueTime);
2999 hzero(call->bytesSent);
3000 hzero(call->bytesRcvd);
3002 * If the number of queued calls exceeds the overload
3003 * threshold then abort this call.
3005 if ((rx_BusyThreshold > 0) &&
3006 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3007 struct rx_packet *tp;
3009 rxi_CallError(call, rx_BusyError);
3010 tp = rxi_SendCallAbort(call, np, 1, 0);
3011 MUTEX_EXIT(&call->lock);
3012 MUTEX_ENTER(&rx_refcnt_mutex);
3014 MUTEX_EXIT(&rx_refcnt_mutex);
3015 if (rx_stats_active)
3016 rx_atomic_inc(&rx_stats.nBusies);
3019 rxi_KeepAliveOn(call);
3020 } else if (np->header.callNumber != currentCallNumber) {
3021 /* Wait until the transmit queue is idle before deciding
3022 * whether to reset the current call. Chances are that the
3023 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3026 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3027 if (call->state == RX_STATE_ACTIVE) {
3028 rxi_WaitforTQBusy(call);
3030 * If we entered error state while waiting,
3031 * must call rxi_CallError to permit rxi_ResetCall
3032 * to processed when the tqWaiter count hits zero.
3035 rxi_CallError(call, call->error);
3036 MUTEX_EXIT(&call->lock);
3037 MUTEX_ENTER(&rx_refcnt_mutex);
3039 MUTEX_EXIT(&rx_refcnt_mutex);
3043 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3044 /* If the new call cannot be taken right now send a busy and set
3045 * the error condition in this call, so that it terminates as
3046 * quickly as possible */
3047 if (call->state == RX_STATE_ACTIVE) {
3048 struct rx_packet *tp;
3050 rxi_CallError(call, RX_CALL_DEAD);
3051 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3053 MUTEX_EXIT(&call->lock);
3054 MUTEX_ENTER(&rx_refcnt_mutex);
3056 MUTEX_EXIT(&rx_refcnt_mutex);
3059 rxi_ResetCall(call, 0);
3060 *call->callNumber = np->header.callNumber;
3062 if (np->header.callNumber == 0)
3063 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",
3064 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3065 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3066 np->header.flags, np, np->retryTime.sec, np->retryTime.usec, np->length));
3068 call->state = RX_STATE_PRECALL;
3069 clock_GetTime(&call->queueTime);
3070 hzero(call->bytesSent);
3071 hzero(call->bytesRcvd);
3073 * If the number of queued calls exceeds the overload
3074 * threshold then abort this call.
3076 if ((rx_BusyThreshold > 0) &&
3077 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3078 struct rx_packet *tp;
3080 rxi_CallError(call, rx_BusyError);
3081 tp = rxi_SendCallAbort(call, np, 1, 0);
3082 MUTEX_EXIT(&call->lock);
3083 MUTEX_ENTER(&rx_refcnt_mutex);
3085 MUTEX_EXIT(&rx_refcnt_mutex);
3086 if (rx_stats_active)
3087 rx_atomic_inc(&rx_stats.nBusies);
3090 rxi_KeepAliveOn(call);
3092 /* Continuing call; do nothing here. */
3094 } else { /* we're the client */
3095 /* Ignore all incoming acknowledgements for calls in DALLY state */
3096 if (call && (call->state == RX_STATE_DALLY)
3097 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3098 if (rx_stats_active)
3099 rx_atomic_inc(&rx_stats.ignorePacketDally);
3100 #ifdef RX_ENABLE_LOCKS
3102 MUTEX_EXIT(&call->lock);
3105 MUTEX_ENTER(&rx_refcnt_mutex);
3107 MUTEX_EXIT(&rx_refcnt_mutex);
3111 /* Ignore anything that's not relevant to the current call. If there
3112 * isn't a current call, then no packet is relevant. */
3113 if (!call || (np->header.callNumber != currentCallNumber)) {
3114 if (rx_stats_active)
3115 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3116 #ifdef RX_ENABLE_LOCKS
3118 MUTEX_EXIT(&call->lock);
3121 MUTEX_ENTER(&rx_refcnt_mutex);
3123 MUTEX_EXIT(&rx_refcnt_mutex);
3126 /* If the service security object index stamped in the packet does not
3127 * match the connection's security index, ignore the packet */
3128 if (np->header.securityIndex != conn->securityIndex) {
3129 #ifdef RX_ENABLE_LOCKS
3130 MUTEX_EXIT(&call->lock);
3132 MUTEX_ENTER(&rx_refcnt_mutex);
3134 MUTEX_EXIT(&rx_refcnt_mutex);
3138 /* If we're receiving the response, then all transmit packets are
3139 * implicitly acknowledged. Get rid of them. */
3140 if (np->header.type == RX_PACKET_TYPE_DATA) {
3141 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3142 /* XXX Hack. Because we must release the global rx lock when
3143 * sending packets (osi_NetSend) we drop all acks while we're
3144 * traversing the tq in rxi_Start sending packets out because
3145 * packets may move to the freePacketQueue as result of being here!
3146 * So we drop these packets until we're safely out of the
3147 * traversing. Really ugly!
3148 * For fine grain RX locking, we set the acked field in the
3149 * packets and let rxi_Start remove them from the transmit queue.
3151 if (call->flags & RX_CALL_TQ_BUSY) {
3152 #ifdef RX_ENABLE_LOCKS
3153 rxi_SetAcksInTransmitQueue(call);
3155 MUTEX_ENTER(&rx_refcnt_mutex);
3157 MUTEX_EXIT(&rx_refcnt_mutex);
3158 return np; /* xmitting; drop packet */
3161 rxi_ClearTransmitQueue(call, 0);
3163 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3164 rxi_ClearTransmitQueue(call, 0);
3165 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3167 if (np->header.type == RX_PACKET_TYPE_ACK) {
3168 /* now check to see if this is an ack packet acknowledging that the
3169 * server actually *lost* some hard-acked data. If this happens we
3170 * ignore this packet, as it may indicate that the server restarted in
3171 * the middle of a call. It is also possible that this is an old ack
3172 * packet. We don't abort the connection in this case, because this
3173 * *might* just be an old ack packet. The right way to detect a server
3174 * restart in the midst of a call is to notice that the server epoch
3176 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3177 * XXX unacknowledged. I think that this is off-by-one, but
3178 * XXX I don't dare change it just yet, since it will
3179 * XXX interact badly with the server-restart detection
3180 * XXX code in receiveackpacket. */
3181 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3182 if (rx_stats_active)
3183 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3184 MUTEX_EXIT(&call->lock);
3185 MUTEX_ENTER(&rx_refcnt_mutex);
3187 MUTEX_EXIT(&rx_refcnt_mutex);
3191 } /* else not a data packet */
3194 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3195 /* Set remote user defined status from packet */
3196 call->remoteStatus = np->header.userStatus;
3198 /* Note the gap between the expected next packet and the actual
3199 * packet that arrived, when the new packet has a smaller serial number
3200 * than expected. Rioses frequently reorder packets all by themselves,
3201 * so this will be quite important with very large window sizes.
3202 * Skew is checked against 0 here to avoid any dependence on the type of
3203 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3205 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3206 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3207 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3209 MUTEX_ENTER(&conn->conn_data_lock);
3210 skew = conn->lastSerial - np->header.serial;
3211 conn->lastSerial = np->header.serial;
3212 MUTEX_EXIT(&conn->conn_data_lock);
3214 struct rx_peer *peer;
3216 if (skew > peer->inPacketSkew) {
3217 dpf(("*** In skew changed from %d to %d\n",
3218 peer->inPacketSkew, skew));
3219 peer->inPacketSkew = skew;
3223 /* Now do packet type-specific processing */
3224 switch (np->header.type) {
3225 case RX_PACKET_TYPE_DATA:
3226 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3229 case RX_PACKET_TYPE_ACK:
3230 /* Respond immediately to ack packets requesting acknowledgement
3232 if (np->header.flags & RX_REQUEST_ACK) {
3234 (void)rxi_SendCallAbort(call, 0, 1, 0);
3236 (void)rxi_SendAck(call, 0, np->header.serial,
3237 RX_ACK_PING_RESPONSE, 1);
3239 np = rxi_ReceiveAckPacket(call, np, 1);
3241 case RX_PACKET_TYPE_ABORT: {
3242 /* An abort packet: reset the call, passing the error up to the user. */
3243 /* What if error is zero? */
3244 /* What if the error is -1? the application will treat it as a timeout. */
3245 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3246 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3247 rxi_CallError(call, errdata);
3248 MUTEX_EXIT(&call->lock);
3249 MUTEX_ENTER(&rx_refcnt_mutex);
3251 MUTEX_EXIT(&rx_refcnt_mutex);
3252 return np; /* xmitting; drop packet */
3254 case RX_PACKET_TYPE_BUSY:
3257 case RX_PACKET_TYPE_ACKALL:
3258 /* All packets acknowledged, so we can drop all packets previously
3259 * readied for sending */
3260 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3261 /* XXX Hack. We because we can't release the global rx lock when
3262 * sending packets (osi_NetSend) we drop all ack pkts while we're
3263 * traversing the tq in rxi_Start sending packets out because
3264 * packets may move to the freePacketQueue as result of being
3265 * here! So we drop these packets until we're safely out of the
3266 * traversing. Really ugly!
3267 * For fine grain RX locking, we set the acked field in the packets
3268 * and let rxi_Start remove the packets from the transmit queue.
3270 if (call->flags & RX_CALL_TQ_BUSY) {
3271 #ifdef RX_ENABLE_LOCKS
3272 rxi_SetAcksInTransmitQueue(call);
3274 #else /* RX_ENABLE_LOCKS */
3275 MUTEX_EXIT(&call->lock);
3276 MUTEX_ENTER(&rx_refcnt_mutex);
3278 MUTEX_EXIT(&rx_refcnt_mutex);
3279 return np; /* xmitting; drop packet */
3280 #endif /* RX_ENABLE_LOCKS */
3282 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3283 rxi_ClearTransmitQueue(call, 0);
3284 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3287 /* Should not reach here, unless the peer is broken: send an abort
3289 rxi_CallError(call, RX_PROTOCOL_ERROR);
3290 np = rxi_SendCallAbort(call, np, 1, 0);
3293 /* Note when this last legitimate packet was received, for keep-alive
3294 * processing. Note, we delay getting the time until now in the hope that
3295 * the packet will be delivered to the user before any get time is required
3296 * (if not, then the time won't actually be re-evaluated here). */
3297 call->lastReceiveTime = clock_Sec();
3298 MUTEX_EXIT(&call->lock);
3299 MUTEX_ENTER(&rx_refcnt_mutex);
3301 MUTEX_EXIT(&rx_refcnt_mutex);
3305 /* return true if this is an "interesting" connection from the point of view
3306 of someone trying to debug the system */
3308 rxi_IsConnInteresting(struct rx_connection *aconn)
3311 struct rx_call *tcall;
3313 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3316 for (i = 0; i < RX_MAXCALLS; i++) {
3317 tcall = aconn->call[i];
3319 if ((tcall->state == RX_STATE_PRECALL)
3320 || (tcall->state == RX_STATE_ACTIVE))
3322 if ((tcall->mode == RX_MODE_SENDING)
3323 || (tcall->mode == RX_MODE_RECEIVING))
3331 /* if this is one of the last few packets AND it wouldn't be used by the
3332 receiving call to immediately satisfy a read request, then drop it on
3333 the floor, since accepting it might prevent a lock-holding thread from
3334 making progress in its reading. If a call has been cleared while in
3335 the precall state then ignore all subsequent packets until the call
3336 is assigned to a thread. */
3339 TooLow(struct rx_packet *ap, struct rx_call *acall)
3343 MUTEX_ENTER(&rx_quota_mutex);
3344 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3345 && (acall->state == RX_STATE_PRECALL))
3346 || ((rx_nFreePackets < rxi_dataQuota + 2)
3347 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3348 && (acall->flags & RX_CALL_READER_WAIT)))) {
3351 MUTEX_EXIT(&rx_quota_mutex);
3357 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3359 struct rx_connection *conn = arg1;
3360 struct rx_call *acall = arg2;
3361 struct rx_call *call = acall;
3362 struct clock when, now;
3365 MUTEX_ENTER(&conn->conn_data_lock);
3366 conn->checkReachEvent = NULL;
3367 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3369 MUTEX_ENTER(&rx_refcnt_mutex);
3371 MUTEX_EXIT(&rx_refcnt_mutex);
3373 MUTEX_EXIT(&conn->conn_data_lock);
3377 MUTEX_ENTER(&conn->conn_call_lock);
3378 MUTEX_ENTER(&conn->conn_data_lock);
3379 for (i = 0; i < RX_MAXCALLS; i++) {
3380 struct rx_call *tc = conn->call[i];
3381 if (tc && tc->state == RX_STATE_PRECALL) {
3387 /* Indicate that rxi_CheckReachEvent is no longer running by
3388 * clearing the flag. Must be atomic under conn_data_lock to
3389 * avoid a new call slipping by: rxi_CheckConnReach holds
3390 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3392 conn->flags &= ~RX_CONN_ATTACHWAIT;
3393 MUTEX_EXIT(&conn->conn_data_lock);
3394 MUTEX_EXIT(&conn->conn_call_lock);
3399 MUTEX_ENTER(&call->lock);
3400 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3402 MUTEX_EXIT(&call->lock);
3404 clock_GetTime(&now);
3406 when.sec += RX_CHECKREACH_TIMEOUT;
3407 MUTEX_ENTER(&conn->conn_data_lock);
3408 if (!conn->checkReachEvent) {
3409 MUTEX_ENTER(&rx_refcnt_mutex);
3411 MUTEX_EXIT(&rx_refcnt_mutex);
3412 conn->checkReachEvent =
3413 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3416 MUTEX_EXIT(&conn->conn_data_lock);
3422 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3424 struct rx_service *service = conn->service;
3425 struct rx_peer *peer = conn->peer;
3426 afs_uint32 now, lastReach;
3428 if (service->checkReach == 0)
3432 MUTEX_ENTER(&peer->peer_lock);
3433 lastReach = peer->lastReachTime;
3434 MUTEX_EXIT(&peer->peer_lock);
3435 if (now - lastReach < RX_CHECKREACH_TTL)
3438 MUTEX_ENTER(&conn->conn_data_lock);
3439 if (conn->flags & RX_CONN_ATTACHWAIT) {
3440 MUTEX_EXIT(&conn->conn_data_lock);
3443 conn->flags |= RX_CONN_ATTACHWAIT;
3444 MUTEX_EXIT(&conn->conn_data_lock);
3445 if (!conn->checkReachEvent)
3446 rxi_CheckReachEvent(NULL, conn, call);
3451 /* try to attach call, if authentication is complete */
3453 TryAttach(struct rx_call *acall, osi_socket socket,
3454 int *tnop, struct rx_call **newcallp,
3457 struct rx_connection *conn = acall->conn;
3459 if (conn->type == RX_SERVER_CONNECTION
3460 && acall->state == RX_STATE_PRECALL) {
3461 /* Don't attach until we have any req'd. authentication. */
3462 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3463 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3464 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3465 /* Note: this does not necessarily succeed; there
3466 * may not any proc available
3469 rxi_ChallengeOn(acall->conn);
3474 /* A data packet has been received off the interface. This packet is
3475 * appropriate to the call (the call is in the right state, etc.). This
3476 * routine can return a packet to the caller, for re-use */
3479 rxi_ReceiveDataPacket(struct rx_call *call,
3480 struct rx_packet *np, int istack,
3481 osi_socket socket, afs_uint32 host, u_short port,
3482 int *tnop, struct rx_call **newcallp)
3484 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3489 afs_uint32 serial=0, flags=0;
3491 struct rx_packet *tnp;
3492 struct clock when, now;
3493 if (rx_stats_active)
3494 rx_atomic_inc(&rx_stats.dataPacketsRead);
3497 /* If there are no packet buffers, drop this new packet, unless we can find
3498 * packet buffers from inactive calls */
3500 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3501 MUTEX_ENTER(&rx_freePktQ_lock);
3502 rxi_NeedMorePackets = TRUE;
3503 MUTEX_EXIT(&rx_freePktQ_lock);
3504 if (rx_stats_active)
3505 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3506 call->rprev = np->header.serial;
3507 rxi_calltrace(RX_TRACE_DROP, call);
3508 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3510 rxi_ClearReceiveQueue(call);
3511 clock_GetTime(&now);
3513 clock_Add(&when, &rx_softAckDelay);
3514 if (!call->delayedAckEvent
3515 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3516 rxevent_Cancel(call->delayedAckEvent, call,
3517 RX_CALL_REFCOUNT_DELAY);
3518 MUTEX_ENTER(&rx_refcnt_mutex);
3519 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3520 MUTEX_EXIT(&rx_refcnt_mutex);
3522 call->delayedAckEvent =
3523 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3525 /* we've damaged this call already, might as well do it in. */
3531 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3532 * packet is one of several packets transmitted as a single
3533 * datagram. Do not send any soft or hard acks until all packets
3534 * in a jumbogram have been processed. Send negative acks right away.
3536 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3537 /* tnp is non-null when there are more packets in the
3538 * current jumbo gram */
3545 seq = np->header.seq;
3546 serial = np->header.serial;
3547 flags = np->header.flags;
3549 /* If the call is in an error state, send an abort message */
3551 return rxi_SendCallAbort(call, np, istack, 0);
3553 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3554 * AFS 3.5 jumbogram. */
3555 if (flags & RX_JUMBO_PACKET) {
3556 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3561 if (np->header.spare != 0) {
3562 MUTEX_ENTER(&call->conn->conn_data_lock);
3563 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3564 MUTEX_EXIT(&call->conn->conn_data_lock);
3567 /* The usual case is that this is the expected next packet */
3568 if (seq == call->rnext) {
3570 /* Check to make sure it is not a duplicate of one already queued */
3571 if (queue_IsNotEmpty(&call->rq)
3572 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3573 if (rx_stats_active)
3574 rx_atomic_inc(&rx_stats.dupPacketsRead);
3575 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3576 rxevent_Cancel(call->delayedAckEvent, call,
3577 RX_CALL_REFCOUNT_DELAY);
3578 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3584 /* It's the next packet. Stick it on the receive queue
3585 * for this call. Set newPackets to make sure we wake
3586 * the reader once all packets have been processed */
3587 #ifdef RX_TRACK_PACKETS
3588 np->flags |= RX_PKTFLAG_RQ;
3590 queue_Prepend(&call->rq, np);
3591 #ifdef RXDEBUG_PACKET
3593 #endif /* RXDEBUG_PACKET */
3595 np = NULL; /* We can't use this anymore */
3598 /* If an ack is requested then set a flag to make sure we
3599 * send an acknowledgement for this packet */
3600 if (flags & RX_REQUEST_ACK) {
3601 ackNeeded = RX_ACK_REQUESTED;
3604 /* Keep track of whether we have received the last packet */
3605 if (flags & RX_LAST_PACKET) {
3606 call->flags |= RX_CALL_HAVE_LAST;
3610 /* Check whether we have all of the packets for this call */
3611 if (call->flags & RX_CALL_HAVE_LAST) {
3612 afs_uint32 tseq; /* temporary sequence number */
3613 struct rx_packet *tp; /* Temporary packet pointer */
3614 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3616 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3617 if (tseq != tp->header.seq)
3619 if (tp->header.flags & RX_LAST_PACKET) {
3620 call->flags |= RX_CALL_RECEIVE_DONE;
3627 /* Provide asynchronous notification for those who want it
3628 * (e.g. multi rx) */
3629 if (call->arrivalProc) {
3630 (*call->arrivalProc) (call, call->arrivalProcHandle,
3631 call->arrivalProcArg);
3632 call->arrivalProc = (void (*)())0;
3635 /* Update last packet received */
3638 /* If there is no server process serving this call, grab
3639 * one, if available. We only need to do this once. If a
3640 * server thread is available, this thread becomes a server
3641 * thread and the server thread becomes a listener thread. */
3643 TryAttach(call, socket, tnop, newcallp, 0);
3646 /* This is not the expected next packet. */
3648 /* Determine whether this is a new or old packet, and if it's
3649 * a new one, whether it fits into the current receive window.
3650 * Also figure out whether the packet was delivered in sequence.
3651 * We use the prev variable to determine whether the new packet
3652 * is the successor of its immediate predecessor in the
3653 * receive queue, and the missing flag to determine whether
3654 * any of this packets predecessors are missing. */
3656 afs_uint32 prev; /* "Previous packet" sequence number */
3657 struct rx_packet *tp; /* Temporary packet pointer */
3658 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3659 int missing; /* Are any predecessors missing? */
3661 /* If the new packet's sequence number has been sent to the
3662 * application already, then this is a duplicate */
3663 if (seq < call->rnext) {
3664 if (rx_stats_active)
3665 rx_atomic_inc(&rx_stats.dupPacketsRead);
3666 rxevent_Cancel(call->delayedAckEvent, call,
3667 RX_CALL_REFCOUNT_DELAY);
3668 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3674 /* If the sequence number is greater than what can be
3675 * accomodated by the current window, then send a negative
3676 * acknowledge and drop the packet */
3677 if ((call->rnext + call->rwind) <= seq) {
3678 rxevent_Cancel(call->delayedAckEvent, call,
3679 RX_CALL_REFCOUNT_DELAY);
3680 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3687 /* Look for the packet in the queue of old received packets */
3688 for (prev = call->rnext - 1, missing =
3689 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3690 /*Check for duplicate packet */
3691 if (seq == tp->header.seq) {
3692 if (rx_stats_active)
3693 rx_atomic_inc(&rx_stats.dupPacketsRead);
3694 rxevent_Cancel(call->delayedAckEvent, call,
3695 RX_CALL_REFCOUNT_DELAY);
3696 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3702 /* If we find a higher sequence packet, break out and
3703 * insert the new packet here. */
3704 if (seq < tp->header.seq)
3706 /* Check for missing packet */
3707 if (tp->header.seq != prev + 1) {
3711 prev = tp->header.seq;
3714 /* Keep track of whether we have received the last packet. */
3715 if (flags & RX_LAST_PACKET) {
3716 call->flags |= RX_CALL_HAVE_LAST;
3719 /* It's within the window: add it to the the receive queue.
3720 * tp is left by the previous loop either pointing at the
3721 * packet before which to insert the new packet, or at the
3722 * queue head if the queue is empty or the packet should be
3724 #ifdef RX_TRACK_PACKETS
3725 np->flags |= RX_PKTFLAG_RQ;
3727 #ifdef RXDEBUG_PACKET
3729 #endif /* RXDEBUG_PACKET */
3730 queue_InsertBefore(tp, np);
3734 /* Check whether we have all of the packets for this call */
3735 if ((call->flags & RX_CALL_HAVE_LAST)
3736 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3737 afs_uint32 tseq; /* temporary sequence number */
3740 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3741 if (tseq != tp->header.seq)
3743 if (tp->header.flags & RX_LAST_PACKET) {
3744 call->flags |= RX_CALL_RECEIVE_DONE;
3751 /* We need to send an ack of the packet is out of sequence,
3752 * or if an ack was requested by the peer. */
3753 if (seq != prev + 1 || missing) {
3754 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3755 } else if (flags & RX_REQUEST_ACK) {
3756 ackNeeded = RX_ACK_REQUESTED;
3759 /* Acknowledge the last packet for each call */
3760 if (flags & RX_LAST_PACKET) {
3771 * If the receiver is waiting for an iovec, fill the iovec
3772 * using the data from the receive queue */
3773 if (call->flags & RX_CALL_IOVEC_WAIT) {
3774 didHardAck = rxi_FillReadVec(call, serial);
3775 /* the call may have been aborted */
3784 /* Wakeup the reader if any */
3785 if ((call->flags & RX_CALL_READER_WAIT)
3786 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3787 || (call->iovNext >= call->iovMax)
3788 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3789 call->flags &= ~RX_CALL_READER_WAIT;
3790 #ifdef RX_ENABLE_LOCKS
3791 CV_BROADCAST(&call->cv_rq);
3793 osi_rxWakeup(&call->rq);
3799 * Send an ack when requested by the peer, or once every
3800 * rxi_SoftAckRate packets until the last packet has been
3801 * received. Always send a soft ack for the last packet in
3802 * the server's reply.
3804 * If we have received all of the packets for the call
3805 * immediately send an RX_PACKET_TYPE_ACKALL packet so that
3806 * the peer can empty its packet queue and cancel all resend
3809 if (call->flags & RX_CALL_RECEIVE_DONE) {
3810 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3811 rxi_AckAll(NULL, call, 0);
3812 } else if (ackNeeded) {
3813 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3814 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3815 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3816 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3817 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3818 } else if (call->nSoftAcks) {
3819 clock_GetTime(&now);
3821 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3822 clock_Add(&when, &rx_lastAckDelay);
3824 clock_Add(&when, &rx_softAckDelay);
3826 if (!call->delayedAckEvent
3827 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3828 rxevent_Cancel(call->delayedAckEvent, call,
3829 RX_CALL_REFCOUNT_DELAY);
3830 MUTEX_ENTER(&rx_refcnt_mutex);
3831 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3832 MUTEX_EXIT(&rx_refcnt_mutex);
3833 call->delayedAckEvent =
3834 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3842 static void rxi_ComputeRate();
3846 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3848 struct rx_peer *peer = conn->peer;
3850 MUTEX_ENTER(&peer->peer_lock);
3851 peer->lastReachTime = clock_Sec();
3852 MUTEX_EXIT(&peer->peer_lock);
3854 MUTEX_ENTER(&conn->conn_data_lock);
3855 if (conn->flags & RX_CONN_ATTACHWAIT) {
3858 conn->flags &= ~RX_CONN_ATTACHWAIT;
3859 MUTEX_EXIT(&conn->conn_data_lock);
3861 for (i = 0; i < RX_MAXCALLS; i++) {
3862 struct rx_call *call = conn->call[i];
3865 MUTEX_ENTER(&call->lock);
3866 /* tnop can be null if newcallp is null */
3867 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3869 MUTEX_EXIT(&call->lock);
3873 MUTEX_EXIT(&conn->conn_data_lock);
3876 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
3878 rx_ack_reason(int reason)
3881 case RX_ACK_REQUESTED:
3883 case RX_ACK_DUPLICATE:
3885 case RX_ACK_OUT_OF_SEQUENCE:
3887 case RX_ACK_EXCEEDS_WINDOW:
3889 case RX_ACK_NOSPACE:
3893 case RX_ACK_PING_RESPONSE:
3906 /* The real smarts of the whole thing. */
3908 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
3911 struct rx_ackPacket *ap;
3913 struct rx_packet *tp;
3914 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3915 struct rx_connection *conn = call->conn;
3916 struct rx_peer *peer = conn->peer;
3917 struct clock now; /* Current time, for RTT calculations */
3921 /* because there are CM's that are bogus, sending weird values for this. */
3922 afs_uint32 skew = 0;
3927 int newAckCount = 0;
3928 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3929 int pktsize = 0; /* Set if we need to update the peer mtu */
3930 int conn_data_locked = 0;
3932 if (rx_stats_active)
3933 rx_atomic_inc(&rx_stats.ackPacketsRead);
3934 ap = (struct rx_ackPacket *)rx_DataOf(np);
3935 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3937 return np; /* truncated ack packet */
3939 /* depends on ack packet struct */
3940 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3941 first = ntohl(ap->firstPacket);
3942 prev = ntohl(ap->previousPacket);
3943 serial = ntohl(ap->serial);
3944 /* temporarily disabled -- needs to degrade over time
3945 * skew = ntohs(ap->maxSkew); */
3947 /* Ignore ack packets received out of order */
3948 if (first < call->tfirst ||
3949 (first == call->tfirst && prev < call->tprev)) {
3955 if (np->header.flags & RX_SLOW_START_OK) {
3956 call->flags |= RX_CALL_SLOW_START_OK;
3959 if (ap->reason == RX_ACK_PING_RESPONSE)
3960 rxi_UpdatePeerReach(conn, call);
3962 if (conn->lastPacketSizeSeq) {
3963 MUTEX_ENTER(&conn->conn_data_lock);
3964 conn_data_locked = 1;
3965 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
3966 pktsize = conn->lastPacketSize;
3967 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
3970 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
3971 if (!conn_data_locked) {
3972 MUTEX_ENTER(&conn->conn_data_lock);
3973 conn_data_locked = 1;
3975 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
3976 /* process mtu ping ack */
3977 pktsize = conn->lastPingSize;
3978 conn->lastPingSizeSer = conn->lastPingSize = 0;
3982 if (conn_data_locked) {
3983 MUTEX_EXIT(&conn->conn_data_lock);
3984 conn_data_locked = 0;
3988 if (rxdebug_active) {
3992 len = _snprintf(msg, sizeof(msg),
3993 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3994 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3995 ntohl(ap->serial), ntohl(ap->previousPacket),
3996 (unsigned int)np->header.seq, (unsigned int)skew,
3997 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
4001 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4002 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4006 OutputDebugString(msg);
4008 #else /* AFS_NT40_ENV */
4011 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
4012 ap->reason, ntohl(ap->previousPacket),
4013 (unsigned int)np->header.seq, (unsigned int)serial,
4014 (unsigned int)skew, ntohl(ap->firstPacket));
4017 for (offset = 0; offset < nAcks; offset++)
4018 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4023 #endif /* AFS_NT40_ENV */
4026 MUTEX_ENTER(&peer->peer_lock);
4029 * Start somewhere. Can't assume we can send what we can receive,
4030 * but we are clearly receiving.
4032 if (!peer->maxPacketSize)
4033 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4035 if (pktsize > peer->maxPacketSize) {
4036 peer->maxPacketSize = pktsize;
4037 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4038 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4039 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4040 rxi_ScheduleGrowMTUEvent(call, 1);
4045 /* Update the outgoing packet skew value to the latest value of
4046 * the peer's incoming packet skew value. The ack packet, of
4047 * course, could arrive out of order, but that won't affect things
4049 peer->outPacketSkew = skew;
4051 /* Check for packets that no longer need to be transmitted, and
4052 * discard them. This only applies to packets positively
4053 * acknowledged as having been sent to the peer's upper level.
4054 * All other packets must be retained. So only packets with
4055 * sequence numbers < ap->firstPacket are candidates. */
4057 clock_GetTime(&now);
4059 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
4060 if (tp->header.seq >= first)
4062 call->tfirst = tp->header.seq + 1;
4064 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4067 rxi_ComputeRoundTripTime(tp, ap, call->conn->peer, &now);
4071 rxi_ComputeRate(call->conn->peer, call, p, np, ap->reason);
4074 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4075 /* XXX Hack. Because we have to release the global rx lock when sending
4076 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4077 * in rxi_Start sending packets out because packets may move to the
4078 * freePacketQueue as result of being here! So we drop these packets until
4079 * we're safely out of the traversing. Really ugly!
4080 * To make it even uglier, if we're using fine grain locking, we can
4081 * set the ack bits in the packets and have rxi_Start remove the packets
4082 * when it's done transmitting.
4084 if (call->flags & RX_CALL_TQ_BUSY) {
4085 #ifdef RX_ENABLE_LOCKS
4086 tp->flags |= RX_PKTFLAG_ACKED;
4087 call->flags |= RX_CALL_TQ_SOME_ACKED;
4088 #else /* RX_ENABLE_LOCKS */
4090 #endif /* RX_ENABLE_LOCKS */
4092 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4095 #ifdef RX_TRACK_PACKETS
4096 tp->flags &= ~RX_PKTFLAG_TQ;
4098 #ifdef RXDEBUG_PACKET
4100 #endif /* RXDEBUG_PACKET */
4101 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4106 /* Give rate detector a chance to respond to ping requests */
4107 if (ap->reason == RX_ACK_PING_RESPONSE) {
4108 rxi_ComputeRate(peer, call, 0, np, ap->reason);
4112 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4114 /* Now go through explicit acks/nacks and record the results in
4115 * the waiting packets. These are packets that can't be released
4116 * yet, even with a positive acknowledge. This positive
4117 * acknowledge only means the packet has been received by the
4118 * peer, not that it will be retained long enough to be sent to
4119 * the peer's upper level. In addition, reset the transmit timers
4120 * of any missing packets (those packets that must be missing
4121 * because this packet was out of sequence) */
4123 call->nSoftAcked = 0;
4124 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
4126 /* Set the acknowledge flag per packet based on the
4127 * information in the ack packet. An acknowlegded packet can
4128 * be downgraded when the server has discarded a packet it
4129 * soacked previously, or when an ack packet is received
4130 * out of sequence. */
4131 if (tp->header.seq < first) {
4132 /* Implicit ack information */
4133 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4136 tp->flags |= RX_PKTFLAG_ACKED;
4137 } else if (tp->header.seq < first + nAcks) {
4138 /* Explicit ack information: set it in the packet appropriately */
4139 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4140 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4142 tp->flags |= RX_PKTFLAG_ACKED;
4144 rxi_ComputeRoundTripTime(tp, ap, call->conn->peer, &now);
4146 rxi_ComputeRate(call->conn->peer, call, tp, np,
4155 } else /* RX_ACK_TYPE_NACK */ {
4156 tp->flags &= ~RX_PKTFLAG_ACKED;
4160 if (tp->flags & RX_PKTFLAG_ACKED) {
4161 tp->flags &= ~RX_PKTFLAG_ACKED;
4167 * Following the suggestion of Phil Kern, we back off the peer's
4168 * timeout value for future packets until a successful response
4169 * is received for an initial transmission.
4171 if (missing && !peer->backedOff) {
4172 struct clock c = peer->timeout;
4173 struct clock max_to = {3, 0};
4175 clock_Add(&peer->timeout, &c);
4176 if (clock_Gt(&peer->timeout, &max_to))
4177 peer->timeout = max_to;
4178 peer->backedOff = 1;
4181 /* If packet isn't yet acked, and it has been transmitted at least
4182 * once, reset retransmit time using latest timeout
4183 * ie, this should readjust the retransmit timer for all outstanding
4184 * packets... So we don't just retransmit when we should know better*/
4186 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
4187 tp->retryTime = tp->timeSent;
4188 clock_Add(&tp->retryTime, &peer->timeout);
4189 /* shift by eight because one quarter-sec ~ 256 milliseconds */
4190 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
4194 /* If the window has been extended by this acknowledge packet,
4195 * then wakeup a sender waiting in alloc for window space, or try
4196 * sending packets now, if he's been sitting on packets due to
4197 * lack of window space */
4198 if (call->tnext < (call->tfirst + call->twind)) {
4199 #ifdef RX_ENABLE_LOCKS
4200 CV_SIGNAL(&call->cv_twind);
4202 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4203 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4204 osi_rxWakeup(&call->twind);
4207 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4208 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4212 /* if the ack packet has a receivelen field hanging off it,
4213 * update our state */
4214 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4217 /* If the ack packet has a "recommended" size that is less than
4218 * what I am using now, reduce my size to match */
4219 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4220 (int)sizeof(afs_int32), &tSize);
4221 tSize = (afs_uint32) ntohl(tSize);
4222 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4224 /* Get the maximum packet size to send to this peer */
4225 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4227 tSize = (afs_uint32) ntohl(tSize);
4228 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4229 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4231 /* sanity check - peer might have restarted with different params.
4232 * If peer says "send less", dammit, send less... Peer should never
4233 * be unable to accept packets of the size that prior AFS versions would
4234 * send without asking. */
4235 if (peer->maxMTU != tSize) {
4236 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4238 peer->maxMTU = tSize;
4239 peer->MTU = MIN(tSize, peer->MTU);
4240 call->MTU = MIN(call->MTU, tSize);
4243 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4246 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4247 (int)sizeof(afs_int32), &tSize);
4248 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4249 if (tSize < call->twind) { /* smaller than our send */
4250 call->twind = tSize; /* window, we must send less... */
4251 call->ssthresh = MIN(call->twind, call->ssthresh);
4252 call->conn->twind[call->channel] = call->twind;
4255 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4256 * network MTU confused with the loopback MTU. Calculate the
4257 * maximum MTU here for use in the slow start code below.
4259 /* Did peer restart with older RX version? */
4260 if (peer->maxDgramPackets > 1) {
4261 peer->maxDgramPackets = 1;
4263 } else if (np->length >=
4264 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4267 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4268 sizeof(afs_int32), &tSize);
4269 tSize = (afs_uint32) ntohl(tSize);
4271 * As of AFS 3.5 we set the send window to match the receive window.
4273 if (tSize < call->twind) {
4274 call->twind = tSize;
4275 call->conn->twind[call->channel] = call->twind;
4276 call->ssthresh = MIN(call->twind, call->ssthresh);
4277 } else if (tSize > call->twind) {
4278 call->twind = tSize;
4279 call->conn->twind[call->channel] = call->twind;
4283 * As of AFS 3.5, a jumbogram is more than one fixed size
4284 * packet transmitted in a single UDP datagram. If the remote
4285 * MTU is smaller than our local MTU then never send a datagram
4286 * larger than the natural MTU.
4289 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4290 (int)sizeof(afs_int32), &tSize);
4291 maxDgramPackets = (afs_uint32) ntohl(tSize);
4292 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4294 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4295 if (maxDgramPackets > 1) {
4296 peer->maxDgramPackets = maxDgramPackets;
4297 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4299 peer->maxDgramPackets = 1;
4300 call->MTU = peer->natMTU;
4302 } else if (peer->maxDgramPackets > 1) {
4303 /* Restarted with lower version of RX */
4304 peer->maxDgramPackets = 1;
4306 } else if (peer->maxDgramPackets > 1
4307 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4308 /* Restarted with lower version of RX */
4309 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4310 peer->natMTU = OLD_MAX_PACKET_SIZE;
4311 peer->MTU = OLD_MAX_PACKET_SIZE;
4312 peer->maxDgramPackets = 1;
4313 peer->nDgramPackets = 1;
4315 call->MTU = OLD_MAX_PACKET_SIZE;
4320 * Calculate how many datagrams were successfully received after
4321 * the first missing packet and adjust the negative ack counter
4326 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4327 if (call->nNacks < nNacked) {
4328 call->nNacks = nNacked;
4331 call->nAcks += newAckCount;
4335 if (call->flags & RX_CALL_FAST_RECOVER) {
4337 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4339 call->flags &= ~RX_CALL_FAST_RECOVER;
4340 call->cwind = call->nextCwind;
4341 call->nextCwind = 0;
4344 call->nCwindAcks = 0;
4345 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4346 /* Three negative acks in a row trigger congestion recovery */
4347 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4348 MUTEX_EXIT(&peer->peer_lock);
4349 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4350 /* someone else is waiting to start recovery */
4353 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4354 rxi_WaitforTQBusy(call);
4355 MUTEX_ENTER(&peer->peer_lock);
4356 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4357 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4358 call->flags |= RX_CALL_FAST_RECOVER;
4359 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4361 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4362 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4363 call->nextCwind = call->ssthresh;
4366 peer->MTU = call->MTU;
4367 peer->cwind = call->nextCwind;
4368 peer->nDgramPackets = call->nDgramPackets;
4370 call->congestSeq = peer->congestSeq;
4371 /* Reset the resend times on the packets that were nacked
4372 * so we will retransmit as soon as the window permits*/
4373 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4375 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4376 clock_Zero(&tp->retryTime);
4378 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4383 /* If cwind is smaller than ssthresh, then increase
4384 * the window one packet for each ack we receive (exponential
4386 * If cwind is greater than or equal to ssthresh then increase
4387 * the congestion window by one packet for each cwind acks we
4388 * receive (linear growth). */
4389 if (call->cwind < call->ssthresh) {
4391 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4392 call->nCwindAcks = 0;
4394 call->nCwindAcks += newAckCount;
4395 if (call->nCwindAcks >= call->cwind) {
4396 call->nCwindAcks = 0;
4397 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4401 * If we have received several acknowledgements in a row then
4402 * it is time to increase the size of our datagrams
4404 if ((int)call->nAcks > rx_nDgramThreshold) {
4405 if (peer->maxDgramPackets > 1) {
4406 if (call->nDgramPackets < peer->maxDgramPackets) {
4407 call->nDgramPackets++;
4409 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4410 } else if (call->MTU < peer->maxMTU) {
4411 /* don't upgrade if we can't handle it */
4412 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4413 call->MTU = peer->ifMTU;
4415 call->MTU += peer->natMTU;
4416 call->MTU = MIN(call->MTU, peer->maxMTU);
4423 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4425 /* Servers need to hold the call until all response packets have
4426 * been acknowledged. Soft acks are good enough since clients
4427 * are not allowed to clear their receive queues. */
4428 if (call->state == RX_STATE_HOLD
4429 && call->tfirst + call->nSoftAcked >= call->tnext) {
4430 call->state = RX_STATE_DALLY;
4431 rxi_ClearTransmitQueue(call, 0);
4432 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4433 } else if (!queue_IsEmpty(&call->tq)) {
4434 rxi_Start(0, call, 0, istack);
4439 /* Received a response to a challenge packet */
4441 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4442 struct rx_packet *np, int istack)
4446 /* Ignore the packet if we're the client */
4447 if (conn->type == RX_CLIENT_CONNECTION)
4450 /* If already authenticated, ignore the packet (it's probably a retry) */
4451 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4454 /* Otherwise, have the security object evaluate the response packet */
4455 error = RXS_CheckResponse(conn->securityObject, conn, np);
4457 /* If the response is invalid, reset the connection, sending
4458 * an abort to the peer */
4462 rxi_ConnectionError(conn, error);
4463 MUTEX_ENTER(&conn->conn_data_lock);
4464 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4465 MUTEX_EXIT(&conn->conn_data_lock);
4468 /* If the response is valid, any calls waiting to attach
4469 * servers can now do so */
4472 for (i = 0; i < RX_MAXCALLS; i++) {
4473 struct rx_call *call = conn->call[i];
4475 MUTEX_ENTER(&call->lock);
4476 if (call->state == RX_STATE_PRECALL)
4477 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4478 /* tnop can be null if newcallp is null */
4479 MUTEX_EXIT(&call->lock);
4483 /* Update the peer reachability information, just in case
4484 * some calls went into attach-wait while we were waiting
4485 * for authentication..
4487 rxi_UpdatePeerReach(conn, NULL);
4492 /* A client has received an authentication challenge: the security
4493 * object is asked to cough up a respectable response packet to send
4494 * back to the server. The server is responsible for retrying the
4495 * challenge if it fails to get a response. */
4498 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4499 struct rx_packet *np, int istack)
4503 /* Ignore the challenge if we're the server */
4504 if (conn->type == RX_SERVER_CONNECTION)
4507 /* Ignore the challenge if the connection is otherwise idle; someone's
4508 * trying to use us as an oracle. */
4509 if (!rxi_HasActiveCalls(conn))
4512 /* Send the security object the challenge packet. It is expected to fill
4513 * in the response. */
4514 error = RXS_GetResponse(conn->securityObject, conn, np);
4516 /* If the security object is unable to return a valid response, reset the
4517 * connection and send an abort to the peer. Otherwise send the response
4518 * packet to the peer connection. */
4520 rxi_ConnectionError(conn, error);
4521 MUTEX_ENTER(&conn->conn_data_lock);
4522 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4523 MUTEX_EXIT(&conn->conn_data_lock);
4525 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4526 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4532 /* Find an available server process to service the current request in
4533 * the given call structure. If one isn't available, queue up this
4534 * call so it eventually gets one */
4536 rxi_AttachServerProc(struct rx_call *call,
4537 osi_socket socket, int *tnop,
4538 struct rx_call **newcallp)
4540 struct rx_serverQueueEntry *sq;
4541 struct rx_service *service = call->conn->service;
4544 /* May already be attached */
4545 if (call->state == RX_STATE_ACTIVE)
4548 MUTEX_ENTER(&rx_serverPool_lock);
4550 haveQuota = QuotaOK(service);
4551 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4552 /* If there are no processes available to service this call,
4553 * put the call on the incoming call queue (unless it's
4554 * already on the queue).
4556 #ifdef RX_ENABLE_LOCKS
4558 ReturnToServerPool(service);
4559 #endif /* RX_ENABLE_LOCKS */
4561 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4562 call->flags |= RX_CALL_WAIT_PROC;
4563 rx_atomic_inc(&rx_nWaiting);
4564 rx_atomic_inc(&rx_nWaited);
4565 rxi_calltrace(RX_CALL_ARRIVAL, call);
4566 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4567 queue_Append(&rx_incomingCallQueue, call);
4570 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4572 /* If hot threads are enabled, and both newcallp and sq->socketp
4573 * are non-null, then this thread will process the call, and the
4574 * idle server thread will start listening on this threads socket.
4577 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4580 *sq->socketp = socket;
4581 clock_GetTime(&call->startTime);
4582 MUTEX_ENTER(&rx_refcnt_mutex);
4583 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4584 MUTEX_EXIT(&rx_refcnt_mutex);
4588 if (call->flags & RX_CALL_WAIT_PROC) {
4589 /* Conservative: I don't think this should happen */
4590 call->flags &= ~RX_CALL_WAIT_PROC;
4591 if (queue_IsOnQueue(call)) {
4594 rx_atomic_dec(&rx_nWaiting);
4597 call->state = RX_STATE_ACTIVE;
4598 call->mode = RX_MODE_RECEIVING;
4599 #ifdef RX_KERNEL_TRACE
4601 int glockOwner = ISAFS_GLOCK();
4604 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4605 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4611 if (call->flags & RX_CALL_CLEARED) {
4612 /* send an ack now to start the packet flow up again */
4613 call->flags &= ~RX_CALL_CLEARED;
4614 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4616 #ifdef RX_ENABLE_LOCKS
4619 service->nRequestsRunning++;
4620 MUTEX_ENTER(&rx_quota_mutex);
4621 if (service->nRequestsRunning <= service->minProcs)
4624 MUTEX_EXIT(&rx_quota_mutex);
4628 MUTEX_EXIT(&rx_serverPool_lock);
4631 /* Delay the sending of an acknowledge event for a short while, while
4632 * a new call is being prepared (in the case of a client) or a reply
4633 * is being prepared (in the case of a server). Rather than sending
4634 * an ack packet, an ACKALL packet is sent. */
4636 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4638 #ifdef RX_ENABLE_LOCKS
4640 MUTEX_ENTER(&call->lock);
4641 call->delayedAckEvent = NULL;
4642 MUTEX_ENTER(&rx_refcnt_mutex);
4643 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4644 MUTEX_EXIT(&rx_refcnt_mutex);
4646 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4647 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4649 MUTEX_EXIT(&call->lock);
4650 #else /* RX_ENABLE_LOCKS */
4652 call->delayedAckEvent = NULL;
4653 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4654 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4655 #endif /* RX_ENABLE_LOCKS */
4659 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4661 struct rx_call *call = arg1;
4662 #ifdef RX_ENABLE_LOCKS
4664 MUTEX_ENTER(&call->lock);
4665 if (event == call->delayedAckEvent)
4666 call->delayedAckEvent = NULL;
4667 MUTEX_ENTER(&rx_refcnt_mutex);
4668 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4669 MUTEX_EXIT(&rx_refcnt_mutex);
4671 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4673 MUTEX_EXIT(&call->lock);
4674 #else /* RX_ENABLE_LOCKS */
4676 call->delayedAckEvent = NULL;
4677 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4678 #endif /* RX_ENABLE_LOCKS */
4682 #ifdef RX_ENABLE_LOCKS
4683 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4684 * clearing them out.
4687 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4689 struct rx_packet *p, *tp;
4692 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4693 p->flags |= RX_PKTFLAG_ACKED;
4697 call->flags |= RX_CALL_TQ_CLEARME;
4698 call->flags |= RX_CALL_TQ_SOME_ACKED;
4701 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4702 call->tfirst = call->tnext;
4703 call->nSoftAcked = 0;
4705 if (call->flags & RX_CALL_FAST_RECOVER) {
4706 call->flags &= ~RX_CALL_FAST_RECOVER;
4707 call->cwind = call->nextCwind;
4708 call->nextCwind = 0;
4711 CV_SIGNAL(&call->cv_twind);
4713 #endif /* RX_ENABLE_LOCKS */
4715 /* Clear out the transmit queue for the current call (all packets have
4716 * been received by peer) */
4718 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4720 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4721 struct rx_packet *p, *tp;
4723 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4725 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4726 p->flags |= RX_PKTFLAG_ACKED;
4730 call->flags |= RX_CALL_TQ_CLEARME;
4731 call->flags |= RX_CALL_TQ_SOME_ACKED;
4734 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4735 #ifdef RXDEBUG_PACKET
4737 #endif /* RXDEBUG_PACKET */
4738 rxi_FreePackets(0, &call->tq);
4739 rxi_WakeUpTransmitQueue(call);
4740 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4741 call->flags &= ~RX_CALL_TQ_CLEARME;
4743 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4745 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4746 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4747 call->nSoftAcked = 0;
4749 if (call->flags & RX_CALL_FAST_RECOVER) {
4750 call->flags &= ~RX_CALL_FAST_RECOVER;
4751 call->cwind = call->nextCwind;
4753 #ifdef RX_ENABLE_LOCKS
4754 CV_SIGNAL(&call->cv_twind);
4756 osi_rxWakeup(&call->twind);
4761 rxi_ClearReceiveQueue(struct rx_call *call)
4763 if (queue_IsNotEmpty(&call->rq)) {
4766 count = rxi_FreePackets(0, &call->rq);
4767 rx_packetReclaims += count;
4768 #ifdef RXDEBUG_PACKET
4770 if ( call->rqc != 0 )
4771 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
4773 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4775 if (call->state == RX_STATE_PRECALL) {
4776 call->flags |= RX_CALL_CLEARED;
4780 /* Send an abort packet for the specified call */
4782 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
4783 int istack, int force)
4786 struct clock when, now;
4791 /* Clients should never delay abort messages */
4792 if (rx_IsClientConn(call->conn))
4795 if (call->abortCode != call->error) {
4796 call->abortCode = call->error;
4797 call->abortCount = 0;
4800 if (force || rxi_callAbortThreshhold == 0
4801 || call->abortCount < rxi_callAbortThreshhold) {
4802 if (call->delayedAbortEvent) {
4803 rxevent_Cancel(call->delayedAbortEvent, call,
4804 RX_CALL_REFCOUNT_ABORT);
4806 error = htonl(call->error);
4809 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4810 (char *)&error, sizeof(error), istack);
4811 } else if (!call->delayedAbortEvent) {
4812 clock_GetTime(&now);
4814 clock_Addmsec(&when, rxi_callAbortDelay);
4815 MUTEX_ENTER(&rx_refcnt_mutex);
4816 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4817 MUTEX_EXIT(&rx_refcnt_mutex);
4818 call->delayedAbortEvent =
4819 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4824 /* Send an abort packet for the specified connection. Packet is an
4825 * optional pointer to a packet that can be used to send the abort.
4826 * Once the number of abort messages reaches the threshhold, an
4827 * event is scheduled to send the abort. Setting the force flag
4828 * overrides sending delayed abort messages.
4830 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4831 * to send the abort packet.
4834 rxi_SendConnectionAbort(struct rx_connection *conn,
4835 struct rx_packet *packet, int istack, int force)
4838 struct clock when, now;
4843 /* Clients should never delay abort messages */
4844 if (rx_IsClientConn(conn))
4847 if (force || rxi_connAbortThreshhold == 0
4848 || conn->abortCount < rxi_connAbortThreshhold) {
4849 if (conn->delayedAbortEvent) {
4850 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4852 error = htonl(conn->error);
4854 MUTEX_EXIT(&conn->conn_data_lock);
4856 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4857 RX_PACKET_TYPE_ABORT, (char *)&error,
4858 sizeof(error), istack);
4859 MUTEX_ENTER(&conn->conn_data_lock);
4860 } else if (!conn->delayedAbortEvent) {
4861 clock_GetTime(&now);
4863 clock_Addmsec(&when, rxi_connAbortDelay);
4864 conn->delayedAbortEvent =
4865 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4870 /* Associate an error all of the calls owned by a connection. Called
4871 * with error non-zero. This is only for really fatal things, like
4872 * bad authentication responses. The connection itself is set in
4873 * error at this point, so that future packets received will be
4876 rxi_ConnectionError(struct rx_connection *conn,
4882 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
4884 MUTEX_ENTER(&conn->conn_data_lock);
4885 if (conn->challengeEvent)
4886 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4887 if (conn->natKeepAliveEvent)
4888 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
4889 if (conn->checkReachEvent) {
4890 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4891 conn->checkReachEvent = 0;
4892 conn->flags &= ~RX_CONN_ATTACHWAIT;
4893 MUTEX_ENTER(&rx_refcnt_mutex);
4895 MUTEX_EXIT(&rx_refcnt_mutex);
4897 MUTEX_EXIT(&conn->conn_data_lock);
4898 for (i = 0; i < RX_MAXCALLS; i++) {
4899 struct rx_call *call = conn->call[i];
4901 MUTEX_ENTER(&call->lock);
4902 rxi_CallError(call, error);
4903 MUTEX_EXIT(&call->lock);
4906 conn->error = error;
4907 if (rx_stats_active)
4908 rx_atomic_inc(&rx_stats.fatalErrors);
4913 * Interrupt an in-progress call with the specified error and wakeup waiters.
4915 * @param[in] call The call to interrupt
4916 * @param[in] error The error code to send to the peer
4919 rx_InterruptCall(struct rx_call *call, afs_int32 error)
4921 MUTEX_ENTER(&call->lock);
4922 rxi_CallError(call, error);
4923 rxi_SendCallAbort(call, NULL, 0, 1);
4924 MUTEX_EXIT(&call->lock);
4928 rxi_CallError(struct rx_call *call, afs_int32 error)
4931 osirx_AssertMine(&call->lock, "rxi_CallError");
4933 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
4935 error = call->error;
4937 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4938 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4939 rxi_ResetCall(call, 0);
4942 rxi_ResetCall(call, 0);
4944 call->error = error;
4947 /* Reset various fields in a call structure, and wakeup waiting
4948 * processes. Some fields aren't changed: state & mode are not
4949 * touched (these must be set by the caller), and bufptr, nLeft, and
4950 * nFree are not reset, since these fields are manipulated by
4951 * unprotected macros, and may only be reset by non-interrupting code.
4954 /* this code requires that call->conn be set properly as a pre-condition. */
4955 #endif /* ADAPT_WINDOW */
4958 rxi_ResetCall(struct rx_call *call, int newcall)
4961 struct rx_peer *peer;
4962 struct rx_packet *packet;
4964 osirx_AssertMine(&call->lock, "rxi_ResetCall");
4966 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
4968 /* Notify anyone who is waiting for asynchronous packet arrival */
4969 if (call->arrivalProc) {
4970 (*call->arrivalProc) (call, call->arrivalProcHandle,
4971 call->arrivalProcArg);
4972 call->arrivalProc = (void (*)())0;
4975 if (call->delayedAbortEvent) {
4976 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4977 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4979 rxi_SendCallAbort(call, packet, 0, 1);
4980 rxi_FreePacket(packet);
4985 * Update the peer with the congestion information in this call
4986 * so other calls on this connection can pick up where this call
4987 * left off. If the congestion sequence numbers don't match then
4988 * another call experienced a retransmission.
4990 peer = call->conn->peer;
4991 MUTEX_ENTER(&peer->peer_lock);
4993 if (call->congestSeq == peer->congestSeq) {
4994 peer->cwind = MAX(peer->cwind, call->cwind);
4995 peer->MTU = MAX(peer->MTU, call->MTU);
4996 peer->nDgramPackets =
4997 MAX(peer->nDgramPackets, call->nDgramPackets);
5000 call->abortCode = 0;
5001 call->abortCount = 0;
5003 if (peer->maxDgramPackets > 1) {
5004 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5006 call->MTU = peer->MTU;
5008 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5009 call->ssthresh = rx_maxSendWindow;
5010 call->nDgramPackets = peer->nDgramPackets;
5011 call->congestSeq = peer->congestSeq;
5012 MUTEX_EXIT(&peer->peer_lock);
5014 flags = call->flags;
5015 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5016 rxi_WaitforTQBusy(call);
5017 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5019 rxi_ClearTransmitQueue(call, 1);
5020 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5021 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5025 rxi_ClearReceiveQueue(call);
5026 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5030 call->twind = call->conn->twind[call->channel];
5031 call->rwind = call->conn->rwind[call->channel];
5032 call->nSoftAcked = 0;
5033 call->nextCwind = 0;
5036 call->nCwindAcks = 0;
5037 call->nSoftAcks = 0;
5038 call->nHardAcks = 0;
5040 call->tfirst = call->rnext = call->tnext = 1;
5043 call->lastAcked = 0;
5044 call->localStatus = call->remoteStatus = 0;
5046 if (flags & RX_CALL_READER_WAIT) {
5047 #ifdef RX_ENABLE_LOCKS
5048 CV_BROADCAST(&call->cv_rq);
5050 osi_rxWakeup(&call->rq);
5053 if (flags & RX_CALL_WAIT_PACKETS) {
5054 MUTEX_ENTER(&rx_freePktQ_lock);
5055 rxi_PacketsUnWait(); /* XXX */
5056 MUTEX_EXIT(&rx_freePktQ_lock);
5058 #ifdef RX_ENABLE_LOCKS
5059 CV_SIGNAL(&call->cv_twind);
5061 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5062 osi_rxWakeup(&call->twind);
5065 #ifdef RX_ENABLE_LOCKS
5066 /* The following ensures that we don't mess with any queue while some
5067 * other thread might also be doing so. The call_queue_lock field is
5068 * is only modified under the call lock. If the call is in the process
5069 * of being removed from a queue, the call is not locked until the
5070 * the queue lock is dropped and only then is the call_queue_lock field
5071 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5072 * Note that any other routine which removes a call from a queue has to
5073 * obtain the queue lock before examing the queue and removing the call.
5075 if (call->call_queue_lock) {
5076 MUTEX_ENTER(call->call_queue_lock);
5077 if (queue_IsOnQueue(call)) {
5079 if (flags & RX_CALL_WAIT_PROC) {
5080 rx_atomic_dec(&rx_nWaiting);
5083 MUTEX_EXIT(call->call_queue_lock);
5084 CLEAR_CALL_QUEUE_LOCK(call);
5086 #else /* RX_ENABLE_LOCKS */
5087 if (queue_IsOnQueue(call)) {
5089 if (flags & RX_CALL_WAIT_PROC)
5090 rx_atomic_dec(&rx_nWaiting);
5092 #endif /* RX_ENABLE_LOCKS */
5094 rxi_KeepAliveOff(call);
5095 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5098 /* Send an acknowledge for the indicated packet (seq,serial) of the
5099 * indicated call, for the indicated reason (reason). This
5100 * acknowledge will specifically acknowledge receiving the packet, and
5101 * will also specify which other packets for this call have been
5102 * received. This routine returns the packet that was used to the
5103 * caller. The caller is responsible for freeing it or re-using it.
5104 * This acknowledgement also returns the highest sequence number
5105 * actually read out by the higher level to the sender; the sender
5106 * promises to keep around packets that have not been read by the
5107 * higher level yet (unless, of course, the sender decides to abort
5108 * the call altogether). Any of p, seq, serial, pflags, or reason may
5109 * be set to zero without ill effect. That is, if they are zero, they
5110 * will not convey any information.
5111 * NOW there is a trailer field, after the ack where it will safely be
5112 * ignored by mundanes, which indicates the maximum size packet this
5113 * host can swallow. */
5115 struct rx_packet *optionalPacket; use to send ack (or null)
5116 int seq; Sequence number of the packet we are acking
5117 int serial; Serial number of the packet
5118 int pflags; Flags field from packet header
5119 int reason; Reason an acknowledge was prompted
5123 rxi_SendAck(struct rx_call *call,
5124 struct rx_packet *optionalPacket, int serial, int reason,
5127 struct rx_ackPacket *ap;
5128 struct rx_packet *rqp;
5129 struct rx_packet *nxp; /* For queue_Scan */
5130 struct rx_packet *p;
5133 afs_uint32 padbytes = 0;
5134 #ifdef RX_ENABLE_TSFPQ
5135 struct rx_ts_info_t * rx_ts_info;
5139 * Open the receive window once a thread starts reading packets
5141 if (call->rnext > 1) {
5142 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5145 /* Don't attempt to grow MTU if this is a critical ping */
5146 if (reason == RX_ACK_MTU) {
5147 /* keep track of per-call attempts, if we're over max, do in small
5148 * otherwise in larger? set a size to increment by, decrease
5151 if (call->conn->peer->maxPacketSize &&
5152 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5154 padbytes = call->conn->peer->maxPacketSize+16;
5156 padbytes = call->conn->peer->maxMTU + 128;
5158 /* do always try a minimum size ping */
5159 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5161 /* subtract the ack payload */
5162 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5163 reason = RX_ACK_PING;
5166 call->nHardAcks = 0;
5167 call->nSoftAcks = 0;
5168 if (call->rnext > call->lastAcked)
5169 call->lastAcked = call->rnext;
5173 rx_computelen(p, p->length); /* reset length, you never know */
5174 } /* where that's been... */
5175 #ifdef RX_ENABLE_TSFPQ
5177 RX_TS_INFO_GET(rx_ts_info);
5178 if ((p = rx_ts_info->local_special_packet)) {
5179 rx_computelen(p, p->length);
5180 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5181 rx_ts_info->local_special_packet = p;
5182 } else { /* We won't send the ack, but don't panic. */
5183 return optionalPacket;
5187 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5188 /* We won't send the ack, but don't panic. */
5189 return optionalPacket;
5194 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5197 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5198 #ifndef RX_ENABLE_TSFPQ
5199 if (!optionalPacket)
5202 return optionalPacket;
5204 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5205 if (rx_Contiguous(p) < templ) {
5206 #ifndef RX_ENABLE_TSFPQ
5207 if (!optionalPacket)
5210 return optionalPacket;
5215 /* MTUXXX failing to send an ack is very serious. We should */
5216 /* try as hard as possible to send even a partial ack; it's */
5217 /* better than nothing. */
5218 ap = (struct rx_ackPacket *)rx_DataOf(p);
5219 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5220 ap->reason = reason;
5222 /* The skew computation used to be bogus, I think it's better now. */
5223 /* We should start paying attention to skew. XXX */
5224 ap->serial = htonl(serial);
5225 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5227 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
5228 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5230 /* No fear of running out of ack packet here because there can only be at most
5231 * one window full of unacknowledged packets. The window size must be constrained
5232 * to be less than the maximum ack size, of course. Also, an ack should always
5233 * fit into a single packet -- it should not ever be fragmented. */
5234 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5235 if (!rqp || !call->rq.next
5236 || (rqp->header.seq > (call->rnext + call->rwind))) {
5237 #ifndef RX_ENABLE_TSFPQ
5238 if (!optionalPacket)
5241 rxi_CallError(call, RX_CALL_DEAD);
5242 return optionalPacket;
5245 while (rqp->header.seq > call->rnext + offset)
5246 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5247 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5249 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5250 #ifndef RX_ENABLE_TSFPQ
5251 if (!optionalPacket)
5254 rxi_CallError(call, RX_CALL_DEAD);
5255 return optionalPacket;
5260 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5262 /* these are new for AFS 3.3 */
5263 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5264 templ = htonl(templ);
5265 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5266 templ = htonl(call->conn->peer->ifMTU);
5267 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5268 sizeof(afs_int32), &templ);
5270 /* new for AFS 3.4 */
5271 templ = htonl(call->rwind);
5272 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5273 sizeof(afs_int32), &templ);
5275 /* new for AFS 3.5 */
5276 templ = htonl(call->conn->peer->ifDgramPackets);
5277 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5278 sizeof(afs_int32), &templ);
5280 p->header.serviceId = call->conn->serviceId;
5281 p->header.cid = (call->conn->cid | call->channel);
5282 p->header.callNumber = *call->callNumber;
5284 p->header.securityIndex = call->conn->securityIndex;
5285 p->header.epoch = call->conn->epoch;
5286 p->header.type = RX_PACKET_TYPE_ACK;
5287 p->header.flags = RX_SLOW_START_OK;
5288 if (reason == RX_ACK_PING) {
5289 p->header.flags |= RX_REQUEST_ACK;
5291 clock_GetTime(&call->pingRequestTime);
5294 p->length = padbytes +
5295 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5298 /* not fast but we can potentially use this if truncated
5299 * fragments are delivered to figure out the mtu.
5301 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5302 sizeof(afs_int32), sizeof(afs_int32),
5306 if (call->conn->type == RX_CLIENT_CONNECTION)
5307 p->header.flags |= RX_CLIENT_INITIATED;
5311 if (rxdebug_active) {
5315 len = _snprintf(msg, sizeof(msg),
5316 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5317 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5318 ntohl(ap->serial), ntohl(ap->previousPacket),
5319 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5320 ap->nAcks, ntohs(ap->bufferSpace) );
5324 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5325 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5329 OutputDebugString(msg);
5331 #else /* AFS_NT40_ENV */
5333 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5334 ap->reason, ntohl(ap->previousPacket),
5335 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5337 for (offset = 0; offset < ap->nAcks; offset++)
5338 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5343 #endif /* AFS_NT40_ENV */
5346 int i, nbytes = p->length;
5348 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5349 if (nbytes <= p->wirevec[i].iov_len) {
5352 savelen = p->wirevec[i].iov_len;
5354 p->wirevec[i].iov_len = nbytes;
5356 rxi_Send(call, p, istack);
5357 p->wirevec[i].iov_len = savelen;
5361 nbytes -= p->wirevec[i].iov_len;
5364 if (rx_stats_active)
5365 rx_atomic_inc(&rx_stats.ackPacketsSent);
5366 #ifndef RX_ENABLE_TSFPQ
5367 if (!optionalPacket)
5370 return optionalPacket; /* Return packet for re-use by caller */
5374 struct rx_packet **list;
5379 /* Send all of the packets in the list in single datagram */
5381 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5382 int istack, int moreFlag)
5387 struct clock now, retryTime;
5388 struct rx_connection *conn = call->conn;
5389 struct rx_peer *peer = conn->peer;
5391 MUTEX_ENTER(&peer->peer_lock);
5392 peer->nSent += xmit->len;
5393 if (xmit->resending)
5394 peer->reSends += xmit->len;
5395 retryTime = peer->timeout;
5396 MUTEX_EXIT(&peer->peer_lock);
5398 if (rx_stats_active) {
5399 if (xmit->resending)
5400 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5402 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5405 clock_GetTime(&now);
5406 clock_Add(&retryTime, &now);
5408 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5412 /* Set the packet flags and schedule the resend events */
5413 /* Only request an ack for the last packet in the list */
5414 for (i = 0; i < xmit->len; i++) {
5415 struct rx_packet *packet = xmit->list[i];
5417 packet->retryTime = retryTime;
5418 if (packet->header.serial) {
5419 /* Exponentially backoff retry times */
5420 if (packet->backoff < MAXBACKOFF) {
5421 /* so it can't stay == 0 */
5422 packet->backoff = (packet->backoff << 1) + 1;
5425 clock_Addmsec(&(packet->retryTime),
5426 ((afs_uint32) packet->backoff) << 8);
5429 /* Wait a little extra for the ack on the last packet */
5431 && !(packet->header.flags & RX_CLIENT_INITIATED)) {
5432 clock_Addmsec(&(packet->retryTime), 400);
5435 /* Record the time sent */
5436 packet->timeSent = now;
5438 /* Ask for an ack on retransmitted packets, on every other packet
5439 * if the peer doesn't support slow start. Ask for an ack on every
5440 * packet until the congestion window reaches the ack rate. */
5441 if (packet->header.serial) {
5444 /* improved RTO calculation- not Karn */
5445 packet->firstSent = now;
5446 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5447 || (!(call->flags & RX_CALL_SLOW_START_OK)
5448 && (packet->header.seq & 1)))) {
5453 /* Tag this packet as not being the last in this group,
5454 * for the receiver's benefit */
5455 if (i < xmit->len - 1 || moreFlag) {
5456 packet->header.flags |= RX_MORE_PACKETS;
5461 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5464 /* Since we're about to send a data packet to the peer, it's
5465 * safe to nuke any scheduled end-of-packets ack */
5466 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5468 MUTEX_EXIT(&call->lock);
5469 MUTEX_ENTER(&rx_refcnt_mutex);
5470 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5471 MUTEX_EXIT(&rx_refcnt_mutex);
5472 if (xmit->len > 1) {
5473 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5475 rxi_SendPacket(call, conn, xmit->list[0], istack);
5477 MUTEX_ENTER(&call->lock);
5478 MUTEX_ENTER(&rx_refcnt_mutex);
5479 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5480 MUTEX_EXIT(&rx_refcnt_mutex);
5482 /* Update last send time for this call (for keep-alive
5483 * processing), and for the connection (so that we can discover
5484 * idle connections) */
5485 conn->lastSendTime = call->lastSendTime = clock_Sec();
5486 /* Let a set of retransmits trigger an idle timeout */
5487 if (!xmit->resending)
5488 call->lastSendData = call->lastSendTime;
5491 /* When sending packets we need to follow these rules:
5492 * 1. Never send more than maxDgramPackets in a jumbogram.
5493 * 2. Never send a packet with more than two iovecs in a jumbogram.
5494 * 3. Never send a retransmitted packet in a jumbogram.
5495 * 4. Never send more than cwind/4 packets in a jumbogram
5496 * We always keep the last list we should have sent so we
5497 * can set the RX_MORE_PACKETS flags correctly.
5501 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5505 struct xmitlist working;
5506 struct xmitlist last;
5508 struct rx_peer *peer = call->conn->peer;
5509 int morePackets = 0;
5511 memset(&last, 0, sizeof(struct xmitlist));
5512 working.list = &list[0];
5514 working.resending = 0;
5516 for (i = 0; i < len; i++) {
5517 /* Does the current packet force us to flush the current list? */
5519 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5520 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5522 /* This sends the 'last' list and then rolls the current working
5523 * set into the 'last' one, and resets the working set */
5526 rxi_SendList(call, &last, istack, 1);
5527 /* If the call enters an error state stop sending, or if
5528 * we entered congestion recovery mode, stop sending */
5529 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5534 working.resending = 0;
5535 working.list = &list[i];
5537 /* Add the current packet to the list if it hasn't been acked.
5538 * Otherwise adjust the list pointer to skip the current packet. */
5539 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5542 if (list[i]->header.serial)
5543 working.resending = 1;
5545 /* Do we need to flush the list? */
5546 if (working.len >= (int)peer->maxDgramPackets
5547 || working.len >= (int)call->nDgramPackets
5548 || working.len >= (int)call->cwind
5549 || list[i]->header.serial
5550 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5552 rxi_SendList(call, &last, istack, 1);
5553 /* If the call enters an error state stop sending, or if
5554 * we entered congestion recovery mode, stop sending */
5556 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5561 working.resending = 0;
5562 working.list = &list[i + 1];
5565 if (working.len != 0) {
5566 osi_Panic("rxi_SendList error");
5568 working.list = &list[i + 1];
5572 /* Send the whole list when the call is in receive mode, when
5573 * the call is in eof mode, when we are in fast recovery mode,
5574 * and when we have the last packet */
5575 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5576 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5577 || (call->flags & RX_CALL_FAST_RECOVER)) {
5578 /* Check for the case where the current list contains
5579 * an acked packet. Since we always send retransmissions
5580 * in a separate packet, we only need to check the first
5581 * packet in the list */
5582 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5586 rxi_SendList(call, &last, istack, morePackets);
5587 /* If the call enters an error state stop sending, or if
5588 * we entered congestion recovery mode, stop sending */
5589 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5593 rxi_SendList(call, &working, istack, 0);
5595 } else if (last.len > 0) {
5596 rxi_SendList(call, &last, istack, 0);
5597 /* Packets which are in 'working' are not sent by this call */
5601 #ifdef RX_ENABLE_LOCKS
5602 /* Call rxi_Start, below, but with the call lock held. */
5604 rxi_StartUnlocked(struct rxevent *event,
5605 void *arg0, void *arg1, int istack)
5607 struct rx_call *call = arg0;
5609 MUTEX_ENTER(&call->lock);
5610 rxi_Start(event, call, arg1, istack);
5611 MUTEX_EXIT(&call->lock);
5613 #endif /* RX_ENABLE_LOCKS */
5615 /* This routine is called when new packets are readied for
5616 * transmission and when retransmission may be necessary, or when the
5617 * transmission window or burst count are favourable. This should be
5618 * better optimized for new packets, the usual case, now that we've
5619 * got rid of queues of send packets. XXXXXXXXXXX */
5621 rxi_Start(struct rxevent *event,
5622 void *arg0, void *arg1, int istack)
5624 struct rx_call *call = arg0;
5626 struct rx_packet *p;
5627 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5628 struct clock now, usenow, retryTime;
5633 /* If rxi_Start is being called as a result of a resend event,
5634 * then make sure that the event pointer is removed from the call
5635 * structure, since there is no longer a per-call retransmission
5637 if (event && event == call->resendEvent) {
5638 MUTEX_ENTER(&rx_refcnt_mutex);
5639 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5640 MUTEX_EXIT(&rx_refcnt_mutex);
5641 call->resendEvent = NULL;
5642 if (queue_IsEmpty(&call->tq)) {
5649 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5650 if (rx_stats_active)
5651 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5656 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5658 clock_GetTime(&now);
5661 /* Send (or resend) any packets that need it, subject to
5662 * window restrictions and congestion burst control
5663 * restrictions. Ask for an ack on the last packet sent in
5664 * this burst. For now, we're relying upon the window being
5665 * considerably bigger than the largest number of packets that
5666 * are typically sent at once by one initial call to
5667 * rxi_Start. This is probably bogus (perhaps we should ask
5668 * for an ack when we're half way through the current
5669 * window?). Also, for non file transfer applications, this
5670 * may end up asking for an ack for every packet. Bogus. XXXX
5673 * But check whether we're here recursively, and let the other guy
5676 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5677 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5678 call->flags |= RX_CALL_TQ_BUSY;
5680 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5682 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5683 call->flags &= ~RX_CALL_NEED_START;
5684 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5686 maxXmitPackets = MIN(call->twind, call->cwind);
5687 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5688 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5689 /* We shouldn't be sending packets if a thread is waiting
5690 * to initiate congestion recovery */
5691 dpf(("call %d waiting to initiate fast recovery\n",
5692 *(call->callNumber)));
5696 && (call->flags & RX_CALL_FAST_RECOVER)) {
5697 /* Only send one packet during fast recovery */
5698 dpf(("call %d restricted to one packet per send during fast recovery\n",
5699 *(call->callNumber)));
5702 #ifdef RX_TRACK_PACKETS
5703 if ((p->flags & RX_PKTFLAG_FREE)
5704 || (!queue_IsEnd(&call->tq, nxp)
5705 && (nxp->flags & RX_PKTFLAG_FREE))
5706 || (p == (struct rx_packet *)&rx_freePacketQueue)
5707 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5708 osi_Panic("rxi_Start: xmit queue clobbered");
5711 if (p->flags & RX_PKTFLAG_ACKED) {
5712 /* Since we may block, don't trust this */
5713 usenow.sec = usenow.usec = 0;
5714 if (rx_stats_active)
5715 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
5716 continue; /* Ignore this packet if it has been acknowledged */
5719 /* Turn off all flags except these ones, which are the same
5720 * on each transmission */
5721 p->header.flags &= RX_PRESET_FLAGS;
5723 if (p->header.seq >=
5724 call->tfirst + MIN((int)call->twind,
5725 (int)(call->nSoftAcked +
5727 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5728 /* Note: if we're waiting for more window space, we can
5729 * still send retransmits; hence we don't return here, but
5730 * break out to schedule a retransmit event */
5731 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
5732 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
5737 /* Transmit the packet if it needs to be sent. */
5738 if (!clock_Lt(&now, &p->retryTime)) {
5739 if (nXmitPackets == maxXmitPackets) {
5740 rxi_SendXmitList(call, call->xmitList,
5741 nXmitPackets, istack);
5744 dpf(("call %d xmit packet %"AFS_PTR_FMT" now %u.%06u retryTime %u.%06u\n",
5745 *(call->callNumber), p,
5747 p->retryTime.sec, p->retryTime.usec));
5748 call->xmitList[nXmitPackets++] = p;
5752 /* xmitList now hold pointers to all of the packets that are
5753 * ready to send. Now we loop to send the packets */
5754 if (nXmitPackets > 0) {
5755 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
5759 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5761 * TQ references no longer protected by this flag; they must remain
5762 * protected by the global lock.
5764 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5765 call->flags &= ~RX_CALL_TQ_BUSY;
5766 rxi_WakeUpTransmitQueue(call);
5770 /* We went into the error state while sending packets. Now is
5771 * the time to reset the call. This will also inform the using
5772 * process that the call is in an error state.
5774 if (rx_stats_active)
5775 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
5776 call->flags &= ~RX_CALL_TQ_BUSY;
5777 rxi_WakeUpTransmitQueue(call);
5778 rxi_CallError(call, call->error);
5781 #ifdef RX_ENABLE_LOCKS
5782 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5784 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5785 /* Some packets have received acks. If they all have, we can clear
5786 * the transmit queue.
5789 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5790 if (p->header.seq < call->tfirst
5791 && (p->flags & RX_PKTFLAG_ACKED)) {
5793 #ifdef RX_TRACK_PACKETS
5794 p->flags &= ~RX_PKTFLAG_TQ;
5796 #ifdef RXDEBUG_PACKET
5804 call->flags |= RX_CALL_TQ_CLEARME;
5806 #endif /* RX_ENABLE_LOCKS */
5807 /* Don't bother doing retransmits if the TQ is cleared. */
5808 if (call->flags & RX_CALL_TQ_CLEARME) {
5809 rxi_ClearTransmitQueue(call, 1);
5811 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5814 /* Always post a resend event, if there is anything in the
5815 * queue, and resend is possible. There should be at least
5816 * one unacknowledged packet in the queue ... otherwise none
5817 * of these packets should be on the queue in the first place.
5819 if (call->resendEvent) {
5820 /* Cancel the existing event and post a new one */
5821 rxevent_Cancel(call->resendEvent, call,
5822 RX_CALL_REFCOUNT_RESEND);
5825 /* The retry time is the retry time on the first unacknowledged
5826 * packet inside the current window */
5828 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5829 /* Don't set timers for packets outside the window */
5830 if (p->header.seq >= call->tfirst + call->twind) {
5834 if (!(p->flags & RX_PKTFLAG_ACKED)
5835 && !clock_IsZero(&p->retryTime)) {
5837 retryTime = p->retryTime;
5842 /* Post a new event to re-run rxi_Start when retries may be needed */
5843 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5844 #ifdef RX_ENABLE_LOCKS
5845 MUTEX_ENTER(&rx_refcnt_mutex);
5846 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5847 MUTEX_EXIT(&rx_refcnt_mutex);
5849 rxevent_PostNow2(&retryTime, &usenow,
5851 (void *)call, 0, istack);
5852 #else /* RX_ENABLE_LOCKS */
5854 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5855 (void *)call, 0, istack);
5856 #endif /* RX_ENABLE_LOCKS */
5859 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5860 } while (call->flags & RX_CALL_NEED_START);
5862 * TQ references no longer protected by this flag; they must remain
5863 * protected by the global lock.
5865 call->flags &= ~RX_CALL_TQ_BUSY;
5866 rxi_WakeUpTransmitQueue(call);
5868 call->flags |= RX_CALL_NEED_START;
5870 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5872 if (call->resendEvent) {
5873 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5878 /* Also adjusts the keep alive parameters for the call, to reflect
5879 * that we have just sent a packet (so keep alives aren't sent
5882 rxi_Send(struct rx_call *call, struct rx_packet *p,
5885 struct rx_connection *conn = call->conn;
5887 /* Stamp each packet with the user supplied status */
5888 p->header.userStatus = call->localStatus;
5890 /* Allow the security object controlling this call's security to
5891 * make any last-minute changes to the packet */
5892 RXS_SendPacket(conn->securityObject, call, p);
5894 /* Since we're about to send SOME sort of packet to the peer, it's
5895 * safe to nuke any scheduled end-of-packets ack */
5896 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5898 /* Actually send the packet, filling in more connection-specific fields */
5899 MUTEX_EXIT(&call->lock);
5900 MUTEX_ENTER(&rx_refcnt_mutex);
5901 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5902 MUTEX_EXIT(&rx_refcnt_mutex);
5903 rxi_SendPacket(call, conn, p, istack);
5904 MUTEX_ENTER(&rx_refcnt_mutex);
5905 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5906 MUTEX_EXIT(&rx_refcnt_mutex);
5907 MUTEX_ENTER(&call->lock);
5909 /* Update last send time for this call (for keep-alive
5910 * processing), and for the connection (so that we can discover
5911 * idle connections) */
5912 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5913 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
5914 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
5916 conn->lastSendTime = call->lastSendTime = clock_Sec();
5917 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
5918 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5919 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
5920 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
5921 RX_ACK_PING_RESPONSE)))
5922 call->lastSendData = call->lastSendTime;
5926 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5927 * that things are fine. Also called periodically to guarantee that nothing
5928 * falls through the cracks (e.g. (error + dally) connections have keepalive
5929 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5931 * haveCTLock Set if calling from rxi_ReapConnections
5933 #ifdef RX_ENABLE_LOCKS
5935 rxi_CheckCall(struct rx_call *call, int haveCTLock)
5936 #else /* RX_ENABLE_LOCKS */
5938 rxi_CheckCall(struct rx_call *call)
5939 #endif /* RX_ENABLE_LOCKS */
5941 struct rx_connection *conn = call->conn;
5943 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
5944 afs_uint32 fudgeFactor;
5948 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5949 if (call->flags & RX_CALL_TQ_BUSY) {
5950 /* Call is active and will be reset by rxi_Start if it's
5951 * in an error state.
5956 /* RTT + 8*MDEV, rounded up to the next second. */
5957 fudgeFactor = (((afs_uint32) conn->peer->rtt >> 3) +
5958 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5960 deadTime = conn->secondsUntilDead + fudgeFactor;
5962 /* These are computed to the second (+- 1 second). But that's
5963 * good enough for these values, which should be a significant
5964 * number of seconds. */
5965 if (now > (call->lastReceiveTime + deadTime)) {
5966 if (call->state == RX_STATE_ACTIVE) {
5968 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5970 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5971 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5972 ip_stack_t *ipst = ns->netstack_ip;
5974 ire = ire_cache_lookup(conn->peer->host
5975 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5977 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5979 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5986 if (ire && ire->ire_max_frag > 0)
5987 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
5989 #if defined(GLOBAL_NETSTACKID)
5993 #endif /* ADAPT_PMTU */
5994 cerror = RX_CALL_DEAD;
5997 #ifdef RX_ENABLE_LOCKS
5998 /* Cancel pending events */
5999 rxevent_Cancel(call->delayedAckEvent, call,
6000 RX_CALL_REFCOUNT_DELAY);
6001 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
6002 rxevent_Cancel(call->keepAliveEvent, call,
6003 RX_CALL_REFCOUNT_ALIVE);
6004 MUTEX_ENTER(&rx_refcnt_mutex);
6005 if (call->refCount == 0) {
6006 rxi_FreeCall(call, haveCTLock);
6007 MUTEX_EXIT(&rx_refcnt_mutex);
6010 MUTEX_EXIT(&rx_refcnt_mutex);
6012 #else /* RX_ENABLE_LOCKS */
6013 rxi_FreeCall(call, 0);
6015 #endif /* RX_ENABLE_LOCKS */
6017 /* Non-active calls are destroyed if they are not responding
6018 * to pings; active calls are simply flagged in error, so the
6019 * attached process can die reasonably gracefully. */
6022 if (conn->idleDeadTime) {
6023 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6026 /* see if we have a non-activity timeout */
6027 if (call->startWait && idleDeadTime
6028 && ((call->startWait + idleDeadTime) < now) &&
6029 (call->flags & RX_CALL_READER_WAIT)) {
6030 if (call->state == RX_STATE_ACTIVE) {
6031 cerror = RX_CALL_TIMEOUT;
6035 if (call->lastSendData && idleDeadTime && (conn->idleDeadErr != 0)
6036 && ((call->lastSendData + idleDeadTime) < now)) {
6037 if (call->state == RX_STATE_ACTIVE) {
6038 cerror = conn->idleDeadErr;
6044 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6047 /* see if we have a hard timeout */
6049 && (now > (hardDeadTime + call->startTime.sec))) {
6050 if (call->state == RX_STATE_ACTIVE)
6051 rxi_CallError(call, RX_CALL_TIMEOUT);
6056 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT
6057 && call->lastReceiveTime) {
6058 int oldMTU = conn->peer->ifMTU;
6060 /* if we thought we could send more, perhaps things got worse */
6061 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6062 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6063 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6064 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6066 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6068 /* minimum capped in SetPeerMtu */
6069 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6072 conn->lastPacketSize = 0;
6074 /* needed so ResetCall doesn't clobber us. */
6075 call->MTU = conn->peer->ifMTU;
6077 /* if we never succeeded, let the error pass out as-is */
6078 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6079 cerror = conn->msgsizeRetryErr;
6082 rxi_CallError(call, cerror);
6087 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6089 struct rx_connection *conn = arg1;
6090 struct rx_header theader;
6092 struct sockaddr_in taddr;
6095 struct iovec tmpiov[2];
6098 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6101 tp = &tbuffer[sizeof(struct rx_header)];
6102 taddr.sin_family = AF_INET;
6103 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6104 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6105 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6106 taddr.sin_len = sizeof(struct sockaddr_in);
6108 memset(&theader, 0, sizeof(theader));
6109 theader.epoch = htonl(999);
6111 theader.callNumber = 0;
6114 theader.type = RX_PACKET_TYPE_VERSION;
6115 theader.flags = RX_LAST_PACKET;
6116 theader.serviceId = 0;
6118 memcpy(tbuffer, &theader, sizeof(theader));
6119 memcpy(tp, &a, sizeof(a));
6120 tmpiov[0].iov_base = tbuffer;
6121 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6123 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6125 MUTEX_ENTER(&conn->conn_data_lock);
6126 MUTEX_ENTER(&rx_refcnt_mutex);
6127 /* Only reschedule ourselves if the connection would not be destroyed */
6128 if (conn->refCount <= 1) {
6129 conn->natKeepAliveEvent = NULL;
6130 MUTEX_EXIT(&rx_refcnt_mutex);
6131 MUTEX_EXIT(&conn->conn_data_lock);
6132 rx_DestroyConnection(conn); /* drop the reference for this */
6134 conn->refCount--; /* drop the reference for this */
6135 MUTEX_EXIT(&rx_refcnt_mutex);
6136 conn->natKeepAliveEvent = NULL;
6137 rxi_ScheduleNatKeepAliveEvent(conn);
6138 MUTEX_EXIT(&conn->conn_data_lock);
6143 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6145 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6146 struct clock when, now;
6147 clock_GetTime(&now);
6149 when.sec += conn->secondsUntilNatPing;
6150 MUTEX_ENTER(&rx_refcnt_mutex);
6151 conn->refCount++; /* hold a reference for this */
6152 MUTEX_EXIT(&rx_refcnt_mutex);
6153 conn->natKeepAliveEvent =
6154 rxevent_PostNow(&when, &now, rxi_NatKeepAliveEvent, conn, 0);
6159 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6161 MUTEX_ENTER(&conn->conn_data_lock);
6162 conn->secondsUntilNatPing = seconds;
6164 rxi_ScheduleNatKeepAliveEvent(conn);
6165 MUTEX_EXIT(&conn->conn_data_lock);
6169 rxi_NatKeepAliveOn(struct rx_connection *conn)
6171 MUTEX_ENTER(&conn->conn_data_lock);
6172 rxi_ScheduleNatKeepAliveEvent(conn);
6173 MUTEX_EXIT(&conn->conn_data_lock);
6176 /* When a call is in progress, this routine is called occasionally to
6177 * make sure that some traffic has arrived (or been sent to) the peer.
6178 * If nothing has arrived in a reasonable amount of time, the call is
6179 * declared dead; if nothing has been sent for a while, we send a
6180 * keep-alive packet (if we're actually trying to keep the call alive)
6183 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6185 struct rx_call *call = arg1;
6186 struct rx_connection *conn;
6189 MUTEX_ENTER(&rx_refcnt_mutex);
6190 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6191 MUTEX_EXIT(&rx_refcnt_mutex);
6192 MUTEX_ENTER(&call->lock);
6193 if (event == call->keepAliveEvent)
6194 call->keepAliveEvent = NULL;
6197 #ifdef RX_ENABLE_LOCKS
6198 if (rxi_CheckCall(call, 0)) {
6199 MUTEX_EXIT(&call->lock);
6202 #else /* RX_ENABLE_LOCKS */
6203 if (rxi_CheckCall(call))
6205 #endif /* RX_ENABLE_LOCKS */
6207 /* Don't try to keep alive dallying calls */
6208 if (call->state == RX_STATE_DALLY) {
6209 MUTEX_EXIT(&call->lock);
6214 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6215 /* Don't try to send keepalives if there is unacknowledged data */
6216 /* the rexmit code should be good enough, this little hack
6217 * doesn't quite work XXX */
6218 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6220 rxi_ScheduleKeepAliveEvent(call);
6221 MUTEX_EXIT(&call->lock);
6224 /* Does what's on the nameplate. */
6226 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy)
6228 struct rx_call *call = arg1;
6229 struct rx_connection *conn;
6231 MUTEX_ENTER(&rx_refcnt_mutex);
6232 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6233 MUTEX_EXIT(&rx_refcnt_mutex);
6234 MUTEX_ENTER(&call->lock);
6236 if (event == call->growMTUEvent)
6237 call->growMTUEvent = NULL;
6239 #ifdef RX_ENABLE_LOCKS
6240 if (rxi_CheckCall(call, 0)) {
6241 MUTEX_EXIT(&call->lock);
6244 #else /* RX_ENABLE_LOCKS */
6245 if (rxi_CheckCall(call))
6247 #endif /* RX_ENABLE_LOCKS */
6249 /* Don't bother with dallying calls */
6250 if (call->state == RX_STATE_DALLY) {
6251 MUTEX_EXIT(&call->lock);
6258 * keep being scheduled, just don't do anything if we're at peak,
6259 * or we're not set up to be properly handled (idle timeout required)
6261 if ((conn->peer->maxPacketSize != 0) &&
6262 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6263 (conn->idleDeadErr))
6264 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6265 rxi_ScheduleGrowMTUEvent(call, 0);
6266 MUTEX_EXIT(&call->lock);
6270 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6272 if (!call->keepAliveEvent) {
6273 struct clock when, now;
6274 clock_GetTime(&now);
6276 when.sec += call->conn->secondsUntilPing;
6277 MUTEX_ENTER(&rx_refcnt_mutex);
6278 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6279 MUTEX_EXIT(&rx_refcnt_mutex);
6280 call->keepAliveEvent =
6281 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
6286 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6288 if (!call->growMTUEvent) {
6289 struct clock when, now;
6291 clock_GetTime(&now);
6294 if (call->conn->secondsUntilPing)
6295 secs = (6*call->conn->secondsUntilPing)-1;
6297 if (call->conn->secondsUntilDead)
6298 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6302 MUTEX_ENTER(&rx_refcnt_mutex);
6303 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6304 MUTEX_EXIT(&rx_refcnt_mutex);
6305 call->growMTUEvent =
6306 rxevent_PostNow(&when, &now, rxi_GrowMTUEvent, call, 0);
6310 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6312 rxi_KeepAliveOn(struct rx_call *call)
6314 /* Pretend last packet received was received now--i.e. if another
6315 * packet isn't received within the keep alive time, then the call
6316 * will die; Initialize last send time to the current time--even
6317 * if a packet hasn't been sent yet. This will guarantee that a
6318 * keep-alive is sent within the ping time */
6319 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6320 rxi_ScheduleKeepAliveEvent(call);
6324 rxi_GrowMTUOn(struct rx_call *call)
6326 struct rx_connection *conn = call->conn;
6327 MUTEX_ENTER(&conn->conn_data_lock);
6328 conn->lastPingSizeSer = conn->lastPingSize = 0;
6329 MUTEX_EXIT(&conn->conn_data_lock);
6330 rxi_ScheduleGrowMTUEvent(call, 1);
6333 /* This routine is called to send connection abort messages
6334 * that have been delayed to throttle looping clients. */
6336 rxi_SendDelayedConnAbort(struct rxevent *event,
6337 void *arg1, void *unused)
6339 struct rx_connection *conn = arg1;
6342 struct rx_packet *packet;
6344 MUTEX_ENTER(&conn->conn_data_lock);
6345 conn->delayedAbortEvent = NULL;
6346 error = htonl(conn->error);
6348 MUTEX_EXIT(&conn->conn_data_lock);
6349 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6352 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6353 RX_PACKET_TYPE_ABORT, (char *)&error,
6355 rxi_FreePacket(packet);
6359 /* This routine is called to send call abort messages
6360 * that have been delayed to throttle looping clients. */
6362 rxi_SendDelayedCallAbort(struct rxevent *event,
6363 void *arg1, void *dummy)
6365 struct rx_call *call = arg1;
6368 struct rx_packet *packet;
6370 MUTEX_ENTER(&call->lock);
6371 call->delayedAbortEvent = NULL;
6372 error = htonl(call->error);
6374 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6377 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6378 (char *)&error, sizeof(error), 0);
6379 rxi_FreePacket(packet);
6381 MUTEX_EXIT(&call->lock);
6382 MUTEX_ENTER(&rx_refcnt_mutex);
6383 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6384 MUTEX_EXIT(&rx_refcnt_mutex);
6387 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6388 * seconds) to ask the client to authenticate itself. The routine
6389 * issues a challenge to the client, which is obtained from the
6390 * security object associated with the connection */
6392 rxi_ChallengeEvent(struct rxevent *event,
6393 void *arg0, void *arg1, int tries)
6395 struct rx_connection *conn = arg0;
6397 conn->challengeEvent = NULL;
6398 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6399 struct rx_packet *packet;
6400 struct clock when, now;
6403 /* We've failed to authenticate for too long.
6404 * Reset any calls waiting for authentication;
6405 * they are all in RX_STATE_PRECALL.
6409 MUTEX_ENTER(&conn->conn_call_lock);
6410 for (i = 0; i < RX_MAXCALLS; i++) {
6411 struct rx_call *call = conn->call[i];
6413 MUTEX_ENTER(&call->lock);
6414 if (call->state == RX_STATE_PRECALL) {
6415 rxi_CallError(call, RX_CALL_DEAD);
6416 rxi_SendCallAbort(call, NULL, 0, 0);
6418 MUTEX_EXIT(&call->lock);
6421 MUTEX_EXIT(&conn->conn_call_lock);
6425 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6427 /* If there's no packet available, do this later. */
6428 RXS_GetChallenge(conn->securityObject, conn, packet);
6429 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6430 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6431 rxi_FreePacket(packet);
6433 clock_GetTime(&now);
6435 when.sec += RX_CHALLENGE_TIMEOUT;
6436 conn->challengeEvent =
6437 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
6442 /* Call this routine to start requesting the client to authenticate
6443 * itself. This will continue until authentication is established,
6444 * the call times out, or an invalid response is returned. The
6445 * security object associated with the connection is asked to create
6446 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6447 * defined earlier. */
6449 rxi_ChallengeOn(struct rx_connection *conn)
6451 if (!conn->challengeEvent) {
6452 RXS_CreateChallenge(conn->securityObject, conn);
6453 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6458 /* rxi_ComputeRoundTripTime is called with peer locked. */
6459 /* peer may be null */
6461 rxi_ComputeRoundTripTime(struct rx_packet *p,
6462 struct rx_ackPacket *ack,
6463 struct rx_peer *peer,
6466 struct clock thisRtt, *sentp;
6470 /* If the ACK is delayed, then do nothing */
6471 if (ack->reason == RX_ACK_DELAY)
6474 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6475 * their RTT multiple times, so only include the RTT of the last packet
6477 if (p->flags & RX_JUMBO_PACKET)
6480 /* Use the serial number to determine which transmission the ACK is for,
6481 * and set the sent time to match this. If we have no serial number, then
6482 * only use the ACK for RTT calculations if the packet has not been
6486 serial = ntohl(ack->serial);
6488 if (serial == p->header.serial) {
6489 sentp = &p->timeSent;
6490 } else if (serial == p->firstSerial) {
6491 sentp = &p->firstSent;
6492 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6493 sentp = &p->firstSent;
6497 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6498 sentp = &p->firstSent;
6505 if (clock_Lt(&thisRtt, sentp))
6506 return; /* somebody set the clock back, don't count this time. */
6508 clock_Sub(&thisRtt, sentp);
6509 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6510 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6512 if (clock_IsZero(&thisRtt)) {
6514 * The actual round trip time is shorter than the
6515 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6516 * Since we can't tell which at the moment we will assume 1ms.
6518 thisRtt.usec = 1000;
6521 if (rx_stats_active) {
6522 MUTEX_ENTER(&rx_stats_mutex);
6523 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6524 rx_stats.minRtt = thisRtt;
6525 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6526 if (thisRtt.sec > 60) {
6527 MUTEX_EXIT(&rx_stats_mutex);
6528 return; /* somebody set the clock ahead */
6530 rx_stats.maxRtt = thisRtt;
6532 clock_Add(&rx_stats.totalRtt, &thisRtt);
6533 rx_atomic_inc(&rx_stats.nRttSamples);
6534 MUTEX_EXIT(&rx_stats_mutex);
6537 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6539 /* Apply VanJacobson round-trip estimations */
6544 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
6545 * srtt is stored as fixed point with 3 bits after the binary
6546 * point (i.e., scaled by 8). The following magic is
6547 * equivalent to the smoothing algorithm in rfc793 with an
6548 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6549 * srtt'*8 = rtt + srtt*7
6550 * srtt'*8 = srtt*8 + rtt - srtt
6551 * srtt' = srtt + rtt/8 - srtt/8
6552 * srtt' = srtt + (rtt - srtt)/8
6555 delta = _8THMSEC(&thisRtt) - peer->rtt;
6556 peer->rtt += (delta >> 3);
6559 * We accumulate a smoothed rtt variance (actually, a smoothed
6560 * mean difference), then set the retransmit timer to smoothed
6561 * rtt + 4 times the smoothed variance (was 2x in van's original
6562 * paper, but 4x works better for me, and apparently for him as
6564 * rttvar is stored as
6565 * fixed point with 2 bits after the binary point (scaled by
6566 * 4). The following is equivalent to rfc793 smoothing with
6567 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6568 * rttvar'*4 = rttvar*3 + |delta|
6569 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6570 * rttvar' = rttvar + |delta|/4 - rttvar/4
6571 * rttvar' = rttvar + (|delta| - rttvar)/4
6572 * This replaces rfc793's wired-in beta.
6573 * dev*4 = dev*4 + (|actual - expected| - dev)
6579 delta -= (peer->rtt_dev << 1);
6580 peer->rtt_dev += (delta >> 3);
6582 /* I don't have a stored RTT so I start with this value. Since I'm
6583 * probably just starting a call, and will be pushing more data down
6584 * this, I expect congestion to increase rapidly. So I fudge a
6585 * little, and I set deviance to half the rtt. In practice,
6586 * deviance tends to approach something a little less than
6587 * half the smoothed rtt. */
6588 peer->rtt = _8THMSEC(&thisRtt) + 8;
6589 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
6591 /* the timeout is RTT + 4*MDEV + rx_minPeerTimeout msec.
6592 * This is because one end or the other of these connections is usually
6593 * in a user process, and can be switched and/or swapped out. So on fast,
6594 * reliable networks, the timeout would otherwise be too short. */
6595 rtt_timeout = ((peer->rtt >> 3) + peer->rtt_dev) + rx_minPeerTimeout;
6596 clock_Zero(&(peer->timeout));
6597 clock_Addmsec(&(peer->timeout), rtt_timeout);
6599 /* Reset the backedOff flag since we just computed a new timeout value */
6600 peer->backedOff = 0;
6602 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6603 p->header.callNumber, p, MSEC(&thisRtt), peer->rtt >> 3, peer->rtt_dev >> 2, (peer->timeout.sec), (peer->timeout.usec)));
6607 /* Find all server connections that have not been active for a long time, and
6610 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6612 struct clock now, when;
6613 clock_GetTime(&now);
6615 /* Find server connection structures that haven't been used for
6616 * greater than rx_idleConnectionTime */
6618 struct rx_connection **conn_ptr, **conn_end;
6619 int i, havecalls = 0;
6620 MUTEX_ENTER(&rx_connHashTable_lock);
6621 for (conn_ptr = &rx_connHashTable[0], conn_end =
6622 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6624 struct rx_connection *conn, *next;
6625 struct rx_call *call;
6629 for (conn = *conn_ptr; conn; conn = next) {
6630 /* XXX -- Shouldn't the connection be locked? */
6633 for (i = 0; i < RX_MAXCALLS; i++) {
6634 call = conn->call[i];
6638 code = MUTEX_TRYENTER(&call->lock);
6641 #ifdef RX_ENABLE_LOCKS
6642 result = rxi_CheckCall(call, 1);
6643 #else /* RX_ENABLE_LOCKS */
6644 result = rxi_CheckCall(call);
6645 #endif /* RX_ENABLE_LOCKS */
6646 MUTEX_EXIT(&call->lock);
6648 /* If CheckCall freed the call, it might
6649 * have destroyed the connection as well,
6650 * which screws up the linked lists.
6656 if (conn->type == RX_SERVER_CONNECTION) {
6657 /* This only actually destroys the connection if
6658 * there are no outstanding calls */
6659 MUTEX_ENTER(&conn->conn_data_lock);
6660 MUTEX_ENTER(&rx_refcnt_mutex);
6661 if (!havecalls && !conn->refCount
6662 && ((conn->lastSendTime + rx_idleConnectionTime) <
6664 conn->refCount++; /* it will be decr in rx_DestroyConn */
6665 MUTEX_EXIT(&rx_refcnt_mutex);
6666 MUTEX_EXIT(&conn->conn_data_lock);
6667 #ifdef RX_ENABLE_LOCKS
6668 rxi_DestroyConnectionNoLock(conn);
6669 #else /* RX_ENABLE_LOCKS */
6670 rxi_DestroyConnection(conn);
6671 #endif /* RX_ENABLE_LOCKS */
6673 #ifdef RX_ENABLE_LOCKS
6675 MUTEX_EXIT(&rx_refcnt_mutex);
6676 MUTEX_EXIT(&conn->conn_data_lock);
6678 #endif /* RX_ENABLE_LOCKS */
6682 #ifdef RX_ENABLE_LOCKS
6683 while (rx_connCleanup_list) {
6684 struct rx_connection *conn;
6685 conn = rx_connCleanup_list;
6686 rx_connCleanup_list = rx_connCleanup_list->next;
6687 MUTEX_EXIT(&rx_connHashTable_lock);
6688 rxi_CleanupConnection(conn);
6689 MUTEX_ENTER(&rx_connHashTable_lock);
6691 MUTEX_EXIT(&rx_connHashTable_lock);
6692 #endif /* RX_ENABLE_LOCKS */
6695 /* Find any peer structures that haven't been used (haven't had an
6696 * associated connection) for greater than rx_idlePeerTime */
6698 struct rx_peer **peer_ptr, **peer_end;
6702 * Why do we need to hold the rx_peerHashTable_lock across
6703 * the incrementing of peer_ptr since the rx_peerHashTable
6704 * array is not changing? We don't.
6706 * By dropping the lock periodically we can permit other
6707 * activities to be performed while a rxi_ReapConnections
6708 * call is in progress. The goal of reap connections
6709 * is to clean up quickly without causing large amounts
6710 * of contention. Therefore, it is important that global
6711 * mutexes not be held for extended periods of time.
6713 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6714 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6716 struct rx_peer *peer, *next, *prev;
6718 MUTEX_ENTER(&rx_peerHashTable_lock);
6719 for (prev = peer = *peer_ptr; peer; peer = next) {
6721 code = MUTEX_TRYENTER(&peer->peer_lock);
6722 if ((code) && (peer->refCount == 0)
6723 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6724 rx_interface_stat_p rpc_stat, nrpc_stat;
6728 * now know that this peer object is one to be
6729 * removed from the hash table. Once it is removed
6730 * it can't be referenced by other threads.
6731 * Lets remove it first and decrement the struct
6732 * nPeerStructs count.
6734 if (peer == *peer_ptr) {
6740 if (rx_stats_active)
6741 rx_atomic_dec(&rx_stats.nPeerStructs);
6744 * Now if we hold references on 'prev' and 'next'
6745 * we can safely drop the rx_peerHashTable_lock
6746 * while we destroy this 'peer' object.
6752 MUTEX_EXIT(&rx_peerHashTable_lock);
6754 MUTEX_EXIT(&peer->peer_lock);
6755 MUTEX_DESTROY(&peer->peer_lock);
6757 (&peer->rpcStats, rpc_stat, nrpc_stat,
6758 rx_interface_stat)) {
6759 unsigned int num_funcs;
6762 queue_Remove(&rpc_stat->queue_header);
6763 queue_Remove(&rpc_stat->all_peers);
6764 num_funcs = rpc_stat->stats[0].func_total;
6766 sizeof(rx_interface_stat_t) +
6767 rpc_stat->stats[0].func_total *
6768 sizeof(rx_function_entry_v1_t);
6770 rxi_Free(rpc_stat, space);
6772 MUTEX_ENTER(&rx_rpc_stats);
6773 rxi_rpc_peer_stat_cnt -= num_funcs;
6774 MUTEX_EXIT(&rx_rpc_stats);
6779 * Regain the rx_peerHashTable_lock and
6780 * decrement the reference count on 'prev'
6783 MUTEX_ENTER(&rx_peerHashTable_lock);
6790 MUTEX_EXIT(&peer->peer_lock);
6795 MUTEX_EXIT(&rx_peerHashTable_lock);
6799 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6800 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6801 * GC, just below. Really, we shouldn't have to keep moving packets from
6802 * one place to another, but instead ought to always know if we can
6803 * afford to hold onto a packet in its particular use. */
6804 MUTEX_ENTER(&rx_freePktQ_lock);
6805 if (rx_waitingForPackets) {
6806 rx_waitingForPackets = 0;
6807 #ifdef RX_ENABLE_LOCKS
6808 CV_BROADCAST(&rx_waitingForPackets_cv);
6810 osi_rxWakeup(&rx_waitingForPackets);
6813 MUTEX_EXIT(&rx_freePktQ_lock);
6816 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6817 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6821 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6822 * rx.h is sort of strange this is better. This is called with a security
6823 * object before it is discarded. Each connection using a security object has
6824 * its own refcount to the object so it won't actually be freed until the last
6825 * connection is destroyed.
6827 * This is the only rxs module call. A hold could also be written but no one
6831 rxs_Release(struct rx_securityClass *aobj)
6833 return RXS_Close(aobj);
6837 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6838 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6839 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6840 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6842 /* Adjust our estimate of the transmission rate to this peer, given
6843 * that the packet p was just acked. We can adjust peer->timeout and
6844 * call->twind. Pragmatically, this is called
6845 * only with packets of maximal length.
6846 * Called with peer and call locked.
6850 rxi_ComputeRate(struct rx_peer *peer, struct rx_call *call,
6851 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6853 afs_int32 xferSize, xferMs;
6857 /* Count down packets */
6858 if (peer->rateFlag > 0)
6860 /* Do nothing until we're enabled */
6861 if (peer->rateFlag != 0)
6866 /* Count only when the ack seems legitimate */
6867 switch (ackReason) {
6868 case RX_ACK_REQUESTED:
6870 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6874 case RX_ACK_PING_RESPONSE:
6875 if (p) /* want the response to ping-request, not data send */
6877 clock_GetTime(&newTO);
6878 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6879 clock_Sub(&newTO, &call->pingRequestTime);
6880 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6884 xferSize = rx_AckDataSize(rx_maxSendWindow) + RX_HEADER_SIZE;
6891 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %d.%06d, rtt %u, ps %u)\n",
6892 ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"),
6893 xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
6895 /* Track only packets that are big enough. */
6896 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6900 /* absorb RTT data (in milliseconds) for these big packets */
6901 if (peer->smRtt == 0) {
6902 peer->smRtt = xferMs;
6904 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6909 if (peer->countDown) {
6913 peer->countDown = 10; /* recalculate only every so often */
6915 /* In practice, we can measure only the RTT for full packets,
6916 * because of the way Rx acks the data that it receives. (If it's
6917 * smaller than a full packet, it often gets implicitly acked
6918 * either by the call response (from a server) or by the next call
6919 * (from a client), and either case confuses transmission times
6920 * with processing times.) Therefore, replace the above
6921 * more-sophisticated processing with a simpler version, where the
6922 * smoothed RTT is kept for full-size packets, and the time to
6923 * transmit a windowful of full-size packets is simply RTT *
6924 * windowSize. Again, we take two steps:
6925 - ensure the timeout is large enough for a single packet's RTT;
6926 - ensure that the window is small enough to fit in the desired timeout.*/
6928 /* First, the timeout check. */
6929 minTime = peer->smRtt;
6930 /* Get a reasonable estimate for a timeout period */
6932 newTO.sec = minTime / 1000;
6933 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6935 /* Increase the timeout period so that we can always do at least
6936 * one packet exchange */
6937 if (clock_Gt(&newTO, &peer->timeout)) {
6939 dpf(("CONG peer %lx/%u: timeout %d.%06d ==> %ld.%06d (rtt %u)\n",
6940 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec,
6941 newTO.sec, newTO.usec, peer->smRtt));
6943 peer->timeout = newTO;
6946 /* Now, get an estimate for the transmit window size. */
6947 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6948 /* Now, convert to the number of full packets that could fit in a
6949 * reasonable fraction of that interval */
6950 minTime /= (peer->smRtt << 1);
6951 minTime = MAX(minTime, rx_minPeerTimeout);
6952 xferSize = minTime; /* (make a copy) */
6954 /* Now clamp the size to reasonable bounds. */
6957 else if (minTime > rx_maxSendWindow)
6958 minTime = rx_maxSendWindow;
6959 /* if (minTime != peer->maxWindow) {
6960 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u)\n",
6961 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6962 peer->timeout.sec, peer->timeout.usec, peer->smRtt));
6963 peer->maxWindow = minTime;
6964 elide... call->twind = minTime;
6968 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6969 * Discern this by calculating the timeout necessary for rx_Window
6971 if ((xferSize > rx_maxSendWindow) && (peer->timeout.sec >= 3)) {
6972 /* calculate estimate for transmission interval in milliseconds */
6973 minTime = rx_maxSendWindow * peer->smRtt;
6974 if (minTime < 1000) {
6975 dpf(("CONG peer %lx/%u: cut TO %d.%06d by 0.5 (rtt %u)\n",
6976 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6977 peer->timeout.usec, peer->smRtt));
6979 newTO.sec = 0; /* cut back on timeout by half a second */
6980 newTO.usec = 500000;
6981 clock_Sub(&peer->timeout, &newTO);
6986 } /* end of rxi_ComputeRate */
6987 #endif /* ADAPT_WINDOW */
6995 #define TRACE_OPTION_RX_DEBUG 16
7003 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7004 0, KEY_QUERY_VALUE, &parmKey);
7005 if (code != ERROR_SUCCESS)
7008 dummyLen = sizeof(TraceOption);
7009 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7010 (BYTE *) &TraceOption, &dummyLen);
7011 if (code == ERROR_SUCCESS) {
7012 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7014 RegCloseKey (parmKey);
7015 #endif /* AFS_NT40_ENV */
7020 rx_DebugOnOff(int on)
7024 rxdebug_active = on;
7030 rx_StatsOnOff(int on)
7032 rx_stats_active = on;
7036 /* Don't call this debugging routine directly; use dpf */
7038 rxi_DebugPrint(char *format, ...)
7047 va_start(ap, format);
7049 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7052 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7054 OutputDebugString(msg);
7060 va_start(ap, format);
7062 clock_GetTime(&now);
7063 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7064 (unsigned int)now.usec);
7065 vfprintf(rx_Log, format, ap);
7073 * This function is used to process the rx_stats structure that is local
7074 * to a process as well as an rx_stats structure received from a remote
7075 * process (via rxdebug). Therefore, it needs to do minimal version
7079 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7080 afs_int32 freePackets, char version)
7084 if (size != sizeof(struct rx_statistics)) {
7086 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7087 size, sizeof(struct rx_statistics));
7090 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7093 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7094 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7095 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7096 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7097 s->specialPktAllocFailures);
7099 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7100 s->receivePktAllocFailures, s->sendPktAllocFailures,
7101 s->specialPktAllocFailures);
7105 " greedy %u, " "bogusReads %u (last from host %x), "
7106 "noPackets %u, " "noBuffers %u, " "selects %u, "
7107 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7108 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7109 s->selects, s->sendSelects);
7111 fprintf(file, " packets read: ");
7112 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7113 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7115 fprintf(file, "\n");
7118 " other read counters: data %u, " "ack %u, " "dup %u "
7119 "spurious %u " "dally %u\n", s->dataPacketsRead,
7120 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7121 s->ignorePacketDally);
7123 fprintf(file, " packets sent: ");
7124 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7125 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7127 fprintf(file, "\n");
7130 " other send counters: ack %u, " "data %u (not resends), "
7131 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7132 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7133 s->dataPacketsPushed, s->ignoreAckedPacket);
7136 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7137 s->netSendFailures, (int)s->fatalErrors);
7139 if (s->nRttSamples) {
7140 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7141 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7143 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7144 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7148 " %d server connections, " "%d client connections, "
7149 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7150 s->nServerConns, s->nClientConns, s->nPeerStructs,
7151 s->nCallStructs, s->nFreeCallStructs);
7153 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7154 fprintf(file, " %d clock updates\n", clock_nUpdates);
7158 /* for backward compatibility */
7160 rx_PrintStats(FILE * file)
7162 MUTEX_ENTER(&rx_stats_mutex);
7163 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7164 sizeof(rx_stats), rx_nFreePackets,
7166 MUTEX_EXIT(&rx_stats_mutex);
7170 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7172 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
7173 ntohl(peer->host), (int)ntohs(peer->port), (int)peer->burstSize,
7174 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
7177 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
7178 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
7179 (int)peer->timeout.usec, peer->nSent, peer->reSends);
7182 " Packet size %d, " "max in packet skew %d, "
7183 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
7184 (int)peer->outPacketSkew);
7188 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7190 * This mutex protects the following static variables:
7194 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7195 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7197 #define LOCK_RX_DEBUG
7198 #define UNLOCK_RX_DEBUG
7199 #endif /* AFS_PTHREAD_ENV */
7201 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7203 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7204 u_char type, void *inputData, size_t inputLength,
7205 void *outputData, size_t outputLength)
7207 static afs_int32 counter = 100;
7208 time_t waitTime, waitCount;
7209 struct rx_header theader;
7212 struct timeval tv_now, tv_wake, tv_delta;
7213 struct sockaddr_in taddr, faddr;
7227 tp = &tbuffer[sizeof(struct rx_header)];
7228 taddr.sin_family = AF_INET;
7229 taddr.sin_port = remotePort;
7230 taddr.sin_addr.s_addr = remoteAddr;
7231 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7232 taddr.sin_len = sizeof(struct sockaddr_in);
7235 memset(&theader, 0, sizeof(theader));
7236 theader.epoch = htonl(999);
7238 theader.callNumber = htonl(counter);
7241 theader.type = type;
7242 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7243 theader.serviceId = 0;
7245 memcpy(tbuffer, &theader, sizeof(theader));
7246 memcpy(tp, inputData, inputLength);
7248 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7249 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7251 /* see if there's a packet available */
7252 gettimeofday(&tv_wake,0);
7253 tv_wake.tv_sec += waitTime;
7256 FD_SET(socket, &imask);
7257 tv_delta.tv_sec = tv_wake.tv_sec;
7258 tv_delta.tv_usec = tv_wake.tv_usec;
7259 gettimeofday(&tv_now, 0);
7261 if (tv_delta.tv_usec < tv_now.tv_usec) {
7263 tv_delta.tv_usec += 1000000;
7266 tv_delta.tv_usec -= tv_now.tv_usec;
7268 if (tv_delta.tv_sec < tv_now.tv_sec) {
7272 tv_delta.tv_sec -= tv_now.tv_sec;
7275 code = select(0, &imask, 0, 0, &tv_delta);
7276 #else /* AFS_NT40_ENV */
7277 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7278 #endif /* AFS_NT40_ENV */
7279 if (code == 1 && FD_ISSET(socket, &imask)) {
7280 /* now receive a packet */
7281 faddrLen = sizeof(struct sockaddr_in);
7283 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7284 (struct sockaddr *)&faddr, &faddrLen);
7287 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7288 if (counter == ntohl(theader.callNumber))
7296 /* see if we've timed out */
7304 code -= sizeof(struct rx_header);
7305 if (code > outputLength)
7306 code = outputLength;
7307 memcpy(outputData, tp, code);
7310 #endif /* RXDEBUG */
7313 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7314 afs_uint16 remotePort, struct rx_debugStats * stat,
7315 afs_uint32 * supportedValues)
7317 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7319 struct rx_debugIn in;
7321 *supportedValues = 0;
7322 in.type = htonl(RX_DEBUGI_GETSTATS);
7325 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7326 &in, sizeof(in), stat, sizeof(*stat));
7329 * If the call was successful, fixup the version and indicate
7330 * what contents of the stat structure are valid.
7331 * Also do net to host conversion of fields here.
7335 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7336 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7338 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7339 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7341 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7342 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7344 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7345 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7347 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7348 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7350 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7351 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7353 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7354 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7356 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7357 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7359 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7360 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7362 stat->nFreePackets = ntohl(stat->nFreePackets);
7363 stat->packetReclaims = ntohl(stat->packetReclaims);
7364 stat->callsExecuted = ntohl(stat->callsExecuted);
7365 stat->nWaiting = ntohl(stat->nWaiting);
7366 stat->idleThreads = ntohl(stat->idleThreads);
7367 stat->nWaited = ntohl(stat->nWaited);
7368 stat->nPackets = ntohl(stat->nPackets);
7377 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7378 afs_uint16 remotePort, struct rx_statistics * stat,
7379 afs_uint32 * supportedValues)
7381 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7383 struct rx_debugIn in;
7384 afs_int32 *lp = (afs_int32 *) stat;
7388 * supportedValues is currently unused, but added to allow future
7389 * versioning of this function.
7392 *supportedValues = 0;
7393 in.type = htonl(RX_DEBUGI_RXSTATS);
7395 memset(stat, 0, sizeof(*stat));
7397 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7398 &in, sizeof(in), stat, sizeof(*stat));
7403 * Do net to host conversion here
7406 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7417 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7418 afs_uint16 remotePort, size_t version_length,
7421 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7423 return MakeDebugCall(socket, remoteAddr, remotePort,
7424 RX_PACKET_TYPE_VERSION, a, 1, version,
7432 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7433 afs_uint16 remotePort, afs_int32 * nextConnection,
7434 int allConnections, afs_uint32 debugSupportedValues,
7435 struct rx_debugConn * conn,
7436 afs_uint32 * supportedValues)
7438 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7440 struct rx_debugIn in;
7444 * supportedValues is currently unused, but added to allow future
7445 * versioning of this function.
7448 *supportedValues = 0;
7449 if (allConnections) {
7450 in.type = htonl(RX_DEBUGI_GETALLCONN);
7452 in.type = htonl(RX_DEBUGI_GETCONN);
7454 in.index = htonl(*nextConnection);
7455 memset(conn, 0, sizeof(*conn));
7457 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7458 &in, sizeof(in), conn, sizeof(*conn));
7461 *nextConnection += 1;
7464 * Convert old connection format to new structure.
7467 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7468 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7469 #define MOVEvL(a) (conn->a = vL->a)
7471 /* any old or unrecognized version... */
7472 for (i = 0; i < RX_MAXCALLS; i++) {
7473 MOVEvL(callState[i]);
7474 MOVEvL(callMode[i]);
7475 MOVEvL(callFlags[i]);
7476 MOVEvL(callOther[i]);
7478 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7479 MOVEvL(secStats.type);
7480 MOVEvL(secStats.level);
7481 MOVEvL(secStats.flags);
7482 MOVEvL(secStats.expires);
7483 MOVEvL(secStats.packetsReceived);
7484 MOVEvL(secStats.packetsSent);
7485 MOVEvL(secStats.bytesReceived);
7486 MOVEvL(secStats.bytesSent);
7491 * Do net to host conversion here
7493 * I don't convert host or port since we are most likely
7494 * going to want these in NBO.
7496 conn->cid = ntohl(conn->cid);
7497 conn->serial = ntohl(conn->serial);
7498 for (i = 0; i < RX_MAXCALLS; i++) {
7499 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7501 conn->error = ntohl(conn->error);
7502 conn->secStats.flags = ntohl(conn->secStats.flags);
7503 conn->secStats.expires = ntohl(conn->secStats.expires);
7504 conn->secStats.packetsReceived =
7505 ntohl(conn->secStats.packetsReceived);
7506 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7507 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7508 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7509 conn->epoch = ntohl(conn->epoch);
7510 conn->natMTU = ntohl(conn->natMTU);
7519 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7520 afs_uint16 remotePort, afs_int32 * nextPeer,
7521 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7522 afs_uint32 * supportedValues)
7524 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7526 struct rx_debugIn in;
7529 * supportedValues is currently unused, but added to allow future
7530 * versioning of this function.
7533 *supportedValues = 0;
7534 in.type = htonl(RX_DEBUGI_GETPEER);
7535 in.index = htonl(*nextPeer);
7536 memset(peer, 0, sizeof(*peer));
7538 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7539 &in, sizeof(in), peer, sizeof(*peer));
7545 * Do net to host conversion here
7547 * I don't convert host or port since we are most likely
7548 * going to want these in NBO.
7550 peer->ifMTU = ntohs(peer->ifMTU);
7551 peer->idleWhen = ntohl(peer->idleWhen);
7552 peer->refCount = ntohs(peer->refCount);
7553 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7554 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7555 peer->rtt = ntohl(peer->rtt);
7556 peer->rtt_dev = ntohl(peer->rtt_dev);
7557 peer->timeout.sec = ntohl(peer->timeout.sec);
7558 peer->timeout.usec = ntohl(peer->timeout.usec);
7559 peer->nSent = ntohl(peer->nSent);
7560 peer->reSends = ntohl(peer->reSends);
7561 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7562 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7563 peer->rateFlag = ntohl(peer->rateFlag);
7564 peer->natMTU = ntohs(peer->natMTU);
7565 peer->maxMTU = ntohs(peer->maxMTU);
7566 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7567 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7568 peer->MTU = ntohs(peer->MTU);
7569 peer->cwind = ntohs(peer->cwind);
7570 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7571 peer->congestSeq = ntohs(peer->congestSeq);
7572 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7573 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7574 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7575 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7584 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7585 struct rx_debugPeer * peerStats)
7588 afs_int32 error = 1; /* default to "did not succeed" */
7589 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7591 MUTEX_ENTER(&rx_peerHashTable_lock);
7592 for(tp = rx_peerHashTable[hashValue];
7593 tp != NULL; tp = tp->next) {
7594 if (tp->host == peerHost)
7600 MUTEX_EXIT(&rx_peerHashTable_lock);
7604 MUTEX_ENTER(&tp->peer_lock);
7605 peerStats->host = tp->host;
7606 peerStats->port = tp->port;
7607 peerStats->ifMTU = tp->ifMTU;
7608 peerStats->idleWhen = tp->idleWhen;
7609 peerStats->refCount = tp->refCount;
7610 peerStats->burstSize = tp->burstSize;
7611 peerStats->burst = tp->burst;
7612 peerStats->burstWait.sec = tp->burstWait.sec;
7613 peerStats->burstWait.usec = tp->burstWait.usec;
7614 peerStats->rtt = tp->rtt;
7615 peerStats->rtt_dev = tp->rtt_dev;
7616 peerStats->timeout.sec = tp->timeout.sec;
7617 peerStats->timeout.usec = tp->timeout.usec;
7618 peerStats->nSent = tp->nSent;
7619 peerStats->reSends = tp->reSends;
7620 peerStats->inPacketSkew = tp->inPacketSkew;
7621 peerStats->outPacketSkew = tp->outPacketSkew;
7622 peerStats->rateFlag = tp->rateFlag;
7623 peerStats->natMTU = tp->natMTU;
7624 peerStats->maxMTU = tp->maxMTU;
7625 peerStats->maxDgramPackets = tp->maxDgramPackets;
7626 peerStats->ifDgramPackets = tp->ifDgramPackets;
7627 peerStats->MTU = tp->MTU;
7628 peerStats->cwind = tp->cwind;
7629 peerStats->nDgramPackets = tp->nDgramPackets;
7630 peerStats->congestSeq = tp->congestSeq;
7631 peerStats->bytesSent.high = tp->bytesSent.high;
7632 peerStats->bytesSent.low = tp->bytesSent.low;
7633 peerStats->bytesReceived.high = tp->bytesReceived.high;
7634 peerStats->bytesReceived.low = tp->bytesReceived.low;
7635 MUTEX_EXIT(&tp->peer_lock);
7637 MUTEX_ENTER(&rx_peerHashTable_lock);
7640 MUTEX_EXIT(&rx_peerHashTable_lock);
7648 struct rx_serverQueueEntry *np;
7651 struct rx_call *call;
7652 struct rx_serverQueueEntry *sq;
7656 if (rxinit_status == 1) {
7658 return; /* Already shutdown. */
7662 #ifndef AFS_PTHREAD_ENV
7663 FD_ZERO(&rx_selectMask);
7664 #endif /* AFS_PTHREAD_ENV */
7665 rxi_dataQuota = RX_MAX_QUOTA;
7666 #ifndef AFS_PTHREAD_ENV
7668 #endif /* AFS_PTHREAD_ENV */
7671 #ifndef AFS_PTHREAD_ENV
7672 #ifndef AFS_USE_GETTIMEOFDAY
7674 #endif /* AFS_USE_GETTIMEOFDAY */
7675 #endif /* AFS_PTHREAD_ENV */
7677 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7678 call = queue_First(&rx_freeCallQueue, rx_call);
7680 rxi_Free(call, sizeof(struct rx_call));
7683 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7684 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7690 struct rx_peer **peer_ptr, **peer_end;
7691 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7692 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7694 struct rx_peer *peer, *next;
7696 MUTEX_ENTER(&rx_peerHashTable_lock);
7697 for (peer = *peer_ptr; peer; peer = next) {
7698 rx_interface_stat_p rpc_stat, nrpc_stat;
7701 MUTEX_ENTER(&rx_rpc_stats);
7702 MUTEX_ENTER(&peer->peer_lock);
7704 (&peer->rpcStats, rpc_stat, nrpc_stat,
7705 rx_interface_stat)) {
7706 unsigned int num_funcs;
7709 queue_Remove(&rpc_stat->queue_header);
7710 queue_Remove(&rpc_stat->all_peers);
7711 num_funcs = rpc_stat->stats[0].func_total;
7713 sizeof(rx_interface_stat_t) +
7714 rpc_stat->stats[0].func_total *
7715 sizeof(rx_function_entry_v1_t);
7717 rxi_Free(rpc_stat, space);
7719 /* rx_rpc_stats must be held */
7720 rxi_rpc_peer_stat_cnt -= num_funcs;
7722 MUTEX_EXIT(&peer->peer_lock);
7723 MUTEX_EXIT(&rx_rpc_stats);
7727 if (rx_stats_active)
7728 rx_atomic_dec(&rx_stats.nPeerStructs);
7730 MUTEX_EXIT(&rx_peerHashTable_lock);
7733 for (i = 0; i < RX_MAX_SERVICES; i++) {
7735 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7737 for (i = 0; i < rx_hashTableSize; i++) {
7738 struct rx_connection *tc, *ntc;
7739 MUTEX_ENTER(&rx_connHashTable_lock);
7740 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7742 for (j = 0; j < RX_MAXCALLS; j++) {
7744 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7747 rxi_Free(tc, sizeof(*tc));
7749 MUTEX_EXIT(&rx_connHashTable_lock);
7752 MUTEX_ENTER(&freeSQEList_lock);
7754 while ((np = rx_FreeSQEList)) {
7755 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7756 MUTEX_DESTROY(&np->lock);
7757 rxi_Free(np, sizeof(*np));
7760 MUTEX_EXIT(&freeSQEList_lock);
7761 MUTEX_DESTROY(&freeSQEList_lock);
7762 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7763 MUTEX_DESTROY(&rx_connHashTable_lock);
7764 MUTEX_DESTROY(&rx_peerHashTable_lock);
7765 MUTEX_DESTROY(&rx_serverPool_lock);
7767 osi_Free(rx_connHashTable,
7768 rx_hashTableSize * sizeof(struct rx_connection *));
7769 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7771 UNPIN(rx_connHashTable,
7772 rx_hashTableSize * sizeof(struct rx_connection *));
7773 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7775 rxi_FreeAllPackets();
7777 MUTEX_ENTER(&rx_quota_mutex);
7778 rxi_dataQuota = RX_MAX_QUOTA;
7779 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7780 MUTEX_EXIT(&rx_quota_mutex);
7785 #ifdef RX_ENABLE_LOCKS
7787 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7789 if (!MUTEX_ISMINE(lockaddr))
7790 osi_Panic("Lock not held: %s", msg);
7792 #endif /* RX_ENABLE_LOCKS */
7797 * Routines to implement connection specific data.
7801 rx_KeyCreate(rx_destructor_t rtn)
7804 MUTEX_ENTER(&rxi_keyCreate_lock);
7805 key = rxi_keyCreate_counter++;
7806 rxi_keyCreate_destructor = (rx_destructor_t *)
7807 realloc((void *)rxi_keyCreate_destructor,
7808 (key + 1) * sizeof(rx_destructor_t));
7809 rxi_keyCreate_destructor[key] = rtn;
7810 MUTEX_EXIT(&rxi_keyCreate_lock);
7815 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7818 MUTEX_ENTER(&conn->conn_data_lock);
7819 if (!conn->specific) {
7820 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7821 for (i = 0; i < key; i++)
7822 conn->specific[i] = NULL;
7823 conn->nSpecific = key + 1;
7824 conn->specific[key] = ptr;
7825 } else if (key >= conn->nSpecific) {
7826 conn->specific = (void **)
7827 realloc(conn->specific, (key + 1) * sizeof(void *));
7828 for (i = conn->nSpecific; i < key; i++)
7829 conn->specific[i] = NULL;
7830 conn->nSpecific = key + 1;
7831 conn->specific[key] = ptr;
7833 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7834 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7835 conn->specific[key] = ptr;
7837 MUTEX_EXIT(&conn->conn_data_lock);
7841 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7844 MUTEX_ENTER(&svc->svc_data_lock);
7845 if (!svc->specific) {
7846 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
7847 for (i = 0; i < key; i++)
7848 svc->specific[i] = NULL;
7849 svc->nSpecific = key + 1;
7850 svc->specific[key] = ptr;
7851 } else if (key >= svc->nSpecific) {
7852 svc->specific = (void **)
7853 realloc(svc->specific, (key + 1) * sizeof(void *));
7854 for (i = svc->nSpecific; i < key; i++)
7855 svc->specific[i] = NULL;
7856 svc->nSpecific = key + 1;
7857 svc->specific[key] = ptr;
7859 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7860 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7861 svc->specific[key] = ptr;
7863 MUTEX_EXIT(&svc->svc_data_lock);
7867 rx_GetSpecific(struct rx_connection *conn, int key)
7870 MUTEX_ENTER(&conn->conn_data_lock);
7871 if (key >= conn->nSpecific)
7874 ptr = conn->specific[key];
7875 MUTEX_EXIT(&conn->conn_data_lock);
7880 rx_GetServiceSpecific(struct rx_service *svc, int key)
7883 MUTEX_ENTER(&svc->svc_data_lock);
7884 if (key >= svc->nSpecific)
7887 ptr = svc->specific[key];
7888 MUTEX_EXIT(&svc->svc_data_lock);
7893 #endif /* !KERNEL */
7896 * processStats is a queue used to store the statistics for the local
7897 * process. Its contents are similar to the contents of the rpcStats
7898 * queue on a rx_peer structure, but the actual data stored within
7899 * this queue contains totals across the lifetime of the process (assuming
7900 * the stats have not been reset) - unlike the per peer structures
7901 * which can come and go based upon the peer lifetime.
7904 static struct rx_queue processStats = { &processStats, &processStats };
7907 * peerStats is a queue used to store the statistics for all peer structs.
7908 * Its contents are the union of all the peer rpcStats queues.
7911 static struct rx_queue peerStats = { &peerStats, &peerStats };
7914 * rxi_monitor_processStats is used to turn process wide stat collection
7918 static int rxi_monitor_processStats = 0;
7921 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7924 static int rxi_monitor_peerStats = 0;
7927 * rxi_AddRpcStat - given all of the information for a particular rpc
7928 * call, create (if needed) and update the stat totals for the rpc.
7932 * IN stats - the queue of stats that will be updated with the new value
7934 * IN rxInterface - a unique number that identifies the rpc interface
7936 * IN currentFunc - the index of the function being invoked
7938 * IN totalFunc - the total number of functions in this interface
7940 * IN queueTime - the amount of time this function waited for a thread
7942 * IN execTime - the amount of time this function invocation took to execute
7944 * IN bytesSent - the number bytes sent by this invocation
7946 * IN bytesRcvd - the number bytes received by this invocation
7948 * IN isServer - if true, this invocation was made to a server
7950 * IN remoteHost - the ip address of the remote host
7952 * IN remotePort - the port of the remote host
7954 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7956 * INOUT counter - if a new stats structure is allocated, the counter will
7957 * be updated with the new number of allocated stat structures
7965 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
7966 afs_uint32 currentFunc, afs_uint32 totalFunc,
7967 struct clock *queueTime, struct clock *execTime,
7968 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
7969 afs_uint32 remoteHost, afs_uint32 remotePort,
7970 int addToPeerList, unsigned int *counter)
7973 rx_interface_stat_p rpc_stat, nrpc_stat;
7976 * See if there's already a structure for this interface
7979 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7980 if ((rpc_stat->stats[0].interfaceId == rxInterface)
7981 && (rpc_stat->stats[0].remote_is_server == isServer))
7986 * Didn't find a match so allocate a new structure and add it to the
7990 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
7991 || (rpc_stat->stats[0].interfaceId != rxInterface)
7992 || (rpc_stat->stats[0].remote_is_server != isServer)) {
7997 sizeof(rx_interface_stat_t) +
7998 totalFunc * sizeof(rx_function_entry_v1_t);
8000 rpc_stat = rxi_Alloc(space);
8001 if (rpc_stat == NULL) {
8005 *counter += totalFunc;
8006 for (i = 0; i < totalFunc; i++) {
8007 rpc_stat->stats[i].remote_peer = remoteHost;
8008 rpc_stat->stats[i].remote_port = remotePort;
8009 rpc_stat->stats[i].remote_is_server = isServer;
8010 rpc_stat->stats[i].interfaceId = rxInterface;
8011 rpc_stat->stats[i].func_total = totalFunc;
8012 rpc_stat->stats[i].func_index = i;
8013 hzero(rpc_stat->stats[i].invocations);
8014 hzero(rpc_stat->stats[i].bytes_sent);
8015 hzero(rpc_stat->stats[i].bytes_rcvd);
8016 rpc_stat->stats[i].queue_time_sum.sec = 0;
8017 rpc_stat->stats[i].queue_time_sum.usec = 0;
8018 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8019 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8020 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8021 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8022 rpc_stat->stats[i].queue_time_max.sec = 0;
8023 rpc_stat->stats[i].queue_time_max.usec = 0;
8024 rpc_stat->stats[i].execution_time_sum.sec = 0;
8025 rpc_stat->stats[i].execution_time_sum.usec = 0;
8026 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8027 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8028 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8029 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8030 rpc_stat->stats[i].execution_time_max.sec = 0;
8031 rpc_stat->stats[i].execution_time_max.usec = 0;
8033 queue_Prepend(stats, rpc_stat);
8034 if (addToPeerList) {
8035 queue_Prepend(&peerStats, &rpc_stat->all_peers);
8040 * Increment the stats for this function
8043 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
8044 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
8045 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
8046 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8047 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8048 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8049 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8051 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8052 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8054 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8055 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8057 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8058 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8060 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8061 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8069 * rx_IncrementTimeAndCount - increment the times and count for a particular
8074 * IN peer - the peer who invoked the rpc
8076 * IN rxInterface - a unique number that identifies the rpc interface
8078 * IN currentFunc - the index of the function being invoked
8080 * IN totalFunc - the total number of functions in this interface
8082 * IN queueTime - the amount of time this function waited for a thread
8084 * IN execTime - the amount of time this function invocation took to execute
8086 * IN bytesSent - the number bytes sent by this invocation
8088 * IN bytesRcvd - the number bytes received by this invocation
8090 * IN isServer - if true, this invocation was made to a server
8098 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8099 afs_uint32 currentFunc, afs_uint32 totalFunc,
8100 struct clock *queueTime, struct clock *execTime,
8101 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8105 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8108 MUTEX_ENTER(&rx_rpc_stats);
8110 if (rxi_monitor_peerStats) {
8111 MUTEX_ENTER(&peer->peer_lock);
8112 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8113 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8114 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8115 MUTEX_EXIT(&peer->peer_lock);
8118 if (rxi_monitor_processStats) {
8119 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8120 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8121 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8124 MUTEX_EXIT(&rx_rpc_stats);
8129 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8133 * IN callerVersion - the rpc stat version of the caller.
8135 * IN count - the number of entries to marshall.
8137 * IN stats - pointer to stats to be marshalled.
8139 * OUT ptr - Where to store the marshalled data.
8146 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8147 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8153 * We only support the first version
8155 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8156 *(ptr++) = stats->remote_peer;
8157 *(ptr++) = stats->remote_port;
8158 *(ptr++) = stats->remote_is_server;
8159 *(ptr++) = stats->interfaceId;
8160 *(ptr++) = stats->func_total;
8161 *(ptr++) = stats->func_index;
8162 *(ptr++) = hgethi(stats->invocations);
8163 *(ptr++) = hgetlo(stats->invocations);
8164 *(ptr++) = hgethi(stats->bytes_sent);
8165 *(ptr++) = hgetlo(stats->bytes_sent);
8166 *(ptr++) = hgethi(stats->bytes_rcvd);
8167 *(ptr++) = hgetlo(stats->bytes_rcvd);
8168 *(ptr++) = stats->queue_time_sum.sec;
8169 *(ptr++) = stats->queue_time_sum.usec;
8170 *(ptr++) = stats->queue_time_sum_sqr.sec;
8171 *(ptr++) = stats->queue_time_sum_sqr.usec;
8172 *(ptr++) = stats->queue_time_min.sec;
8173 *(ptr++) = stats->queue_time_min.usec;
8174 *(ptr++) = stats->queue_time_max.sec;
8175 *(ptr++) = stats->queue_time_max.usec;
8176 *(ptr++) = stats->execution_time_sum.sec;
8177 *(ptr++) = stats->execution_time_sum.usec;
8178 *(ptr++) = stats->execution_time_sum_sqr.sec;
8179 *(ptr++) = stats->execution_time_sum_sqr.usec;
8180 *(ptr++) = stats->execution_time_min.sec;
8181 *(ptr++) = stats->execution_time_min.usec;
8182 *(ptr++) = stats->execution_time_max.sec;
8183 *(ptr++) = stats->execution_time_max.usec;
8189 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8194 * IN callerVersion - the rpc stat version of the caller
8196 * OUT myVersion - the rpc stat version of this function
8198 * OUT clock_sec - local time seconds
8200 * OUT clock_usec - local time microseconds
8202 * OUT allocSize - the number of bytes allocated to contain stats
8204 * OUT statCount - the number stats retrieved from this process.
8206 * OUT stats - the actual stats retrieved from this process.
8210 * Returns void. If successful, stats will != NULL.
8214 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8215 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8216 size_t * allocSize, afs_uint32 * statCount,
8217 afs_uint32 ** stats)
8227 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8230 * Check to see if stats are enabled
8233 MUTEX_ENTER(&rx_rpc_stats);
8234 if (!rxi_monitor_processStats) {
8235 MUTEX_EXIT(&rx_rpc_stats);
8239 clock_GetTime(&now);
8240 *clock_sec = now.sec;
8241 *clock_usec = now.usec;
8244 * Allocate the space based upon the caller version
8246 * If the client is at an older version than we are,
8247 * we return the statistic data in the older data format, but
8248 * we still return our version number so the client knows we
8249 * are maintaining more data than it can retrieve.
8252 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8253 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8254 *statCount = rxi_rpc_process_stat_cnt;
8257 * This can't happen yet, but in the future version changes
8258 * can be handled by adding additional code here
8262 if (space > (size_t) 0) {
8264 ptr = *stats = rxi_Alloc(space);
8267 rx_interface_stat_p rpc_stat, nrpc_stat;
8271 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8273 * Copy the data based upon the caller version
8275 rx_MarshallProcessRPCStats(callerVersion,
8276 rpc_stat->stats[0].func_total,
8277 rpc_stat->stats, &ptr);
8283 MUTEX_EXIT(&rx_rpc_stats);
8288 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8292 * IN callerVersion - the rpc stat version of the caller
8294 * OUT myVersion - the rpc stat version of this function
8296 * OUT clock_sec - local time seconds
8298 * OUT clock_usec - local time microseconds
8300 * OUT allocSize - the number of bytes allocated to contain stats
8302 * OUT statCount - the number of stats retrieved from the individual
8305 * OUT stats - the actual stats retrieved from the individual peer structures.
8309 * Returns void. If successful, stats will != NULL.
8313 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8314 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8315 size_t * allocSize, afs_uint32 * statCount,
8316 afs_uint32 ** stats)
8326 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8329 * Check to see if stats are enabled
8332 MUTEX_ENTER(&rx_rpc_stats);
8333 if (!rxi_monitor_peerStats) {
8334 MUTEX_EXIT(&rx_rpc_stats);
8338 clock_GetTime(&now);
8339 *clock_sec = now.sec;
8340 *clock_usec = now.usec;
8343 * Allocate the space based upon the caller version
8345 * If the client is at an older version than we are,
8346 * we return the statistic data in the older data format, but
8347 * we still return our version number so the client knows we
8348 * are maintaining more data than it can retrieve.
8351 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8352 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8353 *statCount = rxi_rpc_peer_stat_cnt;
8356 * This can't happen yet, but in the future version changes
8357 * can be handled by adding additional code here
8361 if (space > (size_t) 0) {
8363 ptr = *stats = rxi_Alloc(space);
8366 rx_interface_stat_p rpc_stat, nrpc_stat;
8370 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8372 * We have to fix the offset of rpc_stat since we are
8373 * keeping this structure on two rx_queues. The rx_queue
8374 * package assumes that the rx_queue member is the first
8375 * member of the structure. That is, rx_queue assumes that
8376 * any one item is only on one queue at a time. We are
8377 * breaking that assumption and so we have to do a little
8378 * math to fix our pointers.
8381 fix_offset = (char *)rpc_stat;
8382 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8383 rpc_stat = (rx_interface_stat_p) fix_offset;
8386 * Copy the data based upon the caller version
8388 rx_MarshallProcessRPCStats(callerVersion,
8389 rpc_stat->stats[0].func_total,
8390 rpc_stat->stats, &ptr);
8396 MUTEX_EXIT(&rx_rpc_stats);
8401 * rx_FreeRPCStats - free memory allocated by
8402 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8406 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8407 * rx_RetrievePeerRPCStats
8409 * IN allocSize - the number of bytes in stats.
8417 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8419 rxi_Free(stats, allocSize);
8423 * rx_queryProcessRPCStats - see if process rpc stat collection is
8424 * currently enabled.
8430 * Returns 0 if stats are not enabled != 0 otherwise
8434 rx_queryProcessRPCStats(void)
8437 MUTEX_ENTER(&rx_rpc_stats);
8438 rc = rxi_monitor_processStats;
8439 MUTEX_EXIT(&rx_rpc_stats);
8444 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8450 * Returns 0 if stats are not enabled != 0 otherwise
8454 rx_queryPeerRPCStats(void)
8457 MUTEX_ENTER(&rx_rpc_stats);
8458 rc = rxi_monitor_peerStats;
8459 MUTEX_EXIT(&rx_rpc_stats);
8464 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8474 rx_enableProcessRPCStats(void)
8476 MUTEX_ENTER(&rx_rpc_stats);
8477 rx_enable_stats = 1;
8478 rxi_monitor_processStats = 1;
8479 MUTEX_EXIT(&rx_rpc_stats);
8483 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8493 rx_enablePeerRPCStats(void)
8495 MUTEX_ENTER(&rx_rpc_stats);
8496 rx_enable_stats = 1;
8497 rxi_monitor_peerStats = 1;
8498 MUTEX_EXIT(&rx_rpc_stats);
8502 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8512 rx_disableProcessRPCStats(void)
8514 rx_interface_stat_p rpc_stat, nrpc_stat;
8517 MUTEX_ENTER(&rx_rpc_stats);
8520 * Turn off process statistics and if peer stats is also off, turn
8524 rxi_monitor_processStats = 0;
8525 if (rxi_monitor_peerStats == 0) {
8526 rx_enable_stats = 0;
8529 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8530 unsigned int num_funcs = 0;
8533 queue_Remove(rpc_stat);
8534 num_funcs = rpc_stat->stats[0].func_total;
8536 sizeof(rx_interface_stat_t) +
8537 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8539 rxi_Free(rpc_stat, space);
8540 rxi_rpc_process_stat_cnt -= num_funcs;
8542 MUTEX_EXIT(&rx_rpc_stats);
8546 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8556 rx_disablePeerRPCStats(void)
8558 struct rx_peer **peer_ptr, **peer_end;
8562 * Turn off peer statistics and if process stats is also off, turn
8566 rxi_monitor_peerStats = 0;
8567 if (rxi_monitor_processStats == 0) {
8568 rx_enable_stats = 0;
8571 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8572 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8574 struct rx_peer *peer, *next, *prev;
8576 MUTEX_ENTER(&rx_peerHashTable_lock);
8577 MUTEX_ENTER(&rx_rpc_stats);
8578 for (prev = peer = *peer_ptr; peer; peer = next) {
8580 code = MUTEX_TRYENTER(&peer->peer_lock);
8582 rx_interface_stat_p rpc_stat, nrpc_stat;
8585 if (prev == *peer_ptr) {
8596 MUTEX_EXIT(&rx_peerHashTable_lock);
8599 (&peer->rpcStats, rpc_stat, nrpc_stat,
8600 rx_interface_stat)) {
8601 unsigned int num_funcs = 0;
8604 queue_Remove(&rpc_stat->queue_header);
8605 queue_Remove(&rpc_stat->all_peers);
8606 num_funcs = rpc_stat->stats[0].func_total;
8608 sizeof(rx_interface_stat_t) +
8609 rpc_stat->stats[0].func_total *
8610 sizeof(rx_function_entry_v1_t);
8612 rxi_Free(rpc_stat, space);
8613 rxi_rpc_peer_stat_cnt -= num_funcs;
8615 MUTEX_EXIT(&peer->peer_lock);
8617 MUTEX_ENTER(&rx_peerHashTable_lock);
8627 MUTEX_EXIT(&rx_rpc_stats);
8628 MUTEX_EXIT(&rx_peerHashTable_lock);
8633 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8638 * IN clearFlag - flag indicating which stats to clear
8646 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8648 rx_interface_stat_p rpc_stat, nrpc_stat;
8650 MUTEX_ENTER(&rx_rpc_stats);
8652 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8653 unsigned int num_funcs = 0, i;
8654 num_funcs = rpc_stat->stats[0].func_total;
8655 for (i = 0; i < num_funcs; i++) {
8656 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8657 hzero(rpc_stat->stats[i].invocations);
8659 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8660 hzero(rpc_stat->stats[i].bytes_sent);
8662 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8663 hzero(rpc_stat->stats[i].bytes_rcvd);
8665 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8666 rpc_stat->stats[i].queue_time_sum.sec = 0;
8667 rpc_stat->stats[i].queue_time_sum.usec = 0;
8669 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8670 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8671 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8673 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8674 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8675 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8677 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8678 rpc_stat->stats[i].queue_time_max.sec = 0;
8679 rpc_stat->stats[i].queue_time_max.usec = 0;
8681 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8682 rpc_stat->stats[i].execution_time_sum.sec = 0;
8683 rpc_stat->stats[i].execution_time_sum.usec = 0;
8685 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8686 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8687 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8689 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8690 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8691 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8693 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8694 rpc_stat->stats[i].execution_time_max.sec = 0;
8695 rpc_stat->stats[i].execution_time_max.usec = 0;
8700 MUTEX_EXIT(&rx_rpc_stats);
8704 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8709 * IN clearFlag - flag indicating which stats to clear
8717 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8719 rx_interface_stat_p rpc_stat, nrpc_stat;
8721 MUTEX_ENTER(&rx_rpc_stats);
8723 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8724 unsigned int num_funcs = 0, i;
8727 * We have to fix the offset of rpc_stat since we are
8728 * keeping this structure on two rx_queues. The rx_queue
8729 * package assumes that the rx_queue member is the first
8730 * member of the structure. That is, rx_queue assumes that
8731 * any one item is only on one queue at a time. We are
8732 * breaking that assumption and so we have to do a little
8733 * math to fix our pointers.
8736 fix_offset = (char *)rpc_stat;
8737 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8738 rpc_stat = (rx_interface_stat_p) fix_offset;
8740 num_funcs = rpc_stat->stats[0].func_total;
8741 for (i = 0; i < num_funcs; i++) {
8742 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8743 hzero(rpc_stat->stats[i].invocations);
8745 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8746 hzero(rpc_stat->stats[i].bytes_sent);
8748 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8749 hzero(rpc_stat->stats[i].bytes_rcvd);
8751 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8752 rpc_stat->stats[i].queue_time_sum.sec = 0;
8753 rpc_stat->stats[i].queue_time_sum.usec = 0;
8755 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8756 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8757 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8759 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8760 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8761 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8763 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8764 rpc_stat->stats[i].queue_time_max.sec = 0;
8765 rpc_stat->stats[i].queue_time_max.usec = 0;
8767 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8768 rpc_stat->stats[i].execution_time_sum.sec = 0;
8769 rpc_stat->stats[i].execution_time_sum.usec = 0;
8771 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8772 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8773 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8775 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8776 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8777 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8779 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8780 rpc_stat->stats[i].execution_time_max.sec = 0;
8781 rpc_stat->stats[i].execution_time_max.usec = 0;
8786 MUTEX_EXIT(&rx_rpc_stats);
8790 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8791 * is authorized to enable/disable/clear RX statistics.
8793 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8796 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8798 rxi_rxstat_userok = proc;
8802 rx_RxStatUserOk(struct rx_call *call)
8804 if (!rxi_rxstat_userok)
8806 return rxi_rxstat_userok(call);
8811 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8812 * function in the MSVC runtime DLL (msvcrt.dll).
8814 * Note: the system serializes calls to this function.
8817 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8818 DWORD reason, /* reason function is being called */
8819 LPVOID reserved) /* reserved for future use */
8822 case DLL_PROCESS_ATTACH:
8823 /* library is being attached to a process */
8827 case DLL_PROCESS_DETACH:
8834 #endif /* AFS_NT40_ENV */
8837 int rx_DumpCalls(FILE *outputFile, char *cookie)
8839 #ifdef RXDEBUG_PACKET
8840 #ifdef KDUMP_RX_LOCK
8841 struct rx_call_rx_lock *c;
8848 #define RXDPRINTF sprintf
8849 #define RXDPRINTOUT output
8851 #define RXDPRINTF fprintf
8852 #define RXDPRINTOUT outputFile
8855 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8857 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8860 for (c = rx_allCallsp; c; c = c->allNextp) {
8861 u_short rqc, tqc, iovqc;
8862 struct rx_packet *p, *np;
8864 MUTEX_ENTER(&c->lock);
8865 queue_Count(&c->rq, p, np, rx_packet, rqc);
8866 queue_Count(&c->tq, p, np, rx_packet, tqc);
8867 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8869 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, "
8870 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8871 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8872 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8873 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8874 #ifdef RX_ENABLE_LOCKS
8877 #ifdef RX_REFCOUNT_CHECK
8878 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
8879 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
8882 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,
8883 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
8884 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
8885 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
8886 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
8887 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
8888 #ifdef RX_ENABLE_LOCKS
8889 , (afs_uint32)c->refCount
8891 #ifdef RX_REFCOUNT_CHECK
8892 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
8895 MUTEX_EXIT(&c->lock);
8898 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8901 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
8903 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8905 #endif /* RXDEBUG_PACKET */