2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
14 #include "afs/param.h"
16 #include <afs/param.h>
21 #include "afs/sysincludes.h"
22 #include "afsincludes.h"
28 #include <net/net_globals.h>
29 #endif /* AFS_OSF_ENV */
30 #ifdef AFS_LINUX20_ENV
33 #include "netinet/in.h"
35 #include "inet/common.h"
37 #include "inet/ip_ire.h"
39 #include "afs/afs_args.h"
40 #include "afs/afs_osi.h"
41 #ifdef RX_KERNEL_TRACE
42 #include "rx_kcommon.h"
44 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
48 #undef RXDEBUG /* turn off debugging */
50 #if defined(AFS_SGI_ENV)
51 #include "sys/debug.h"
60 #endif /* AFS_OSF_ENV */
62 #include "afs/sysincludes.h"
63 #include "afsincludes.h"
66 #include "rx_kmutex.h"
67 #include "rx_kernel.h"
71 #include "rx_globals.h"
73 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
74 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
75 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
77 extern afs_int32 afs_termState;
79 #include "sys/lockl.h"
80 #include "sys/lock_def.h"
81 #endif /* AFS_AIX41_ENV */
82 # include "rxgen_consts.h"
84 # include <sys/types.h>
94 # include <afs/afsutil.h>
95 # include <WINNT\afsreg.h>
97 # include <sys/socket.h>
98 # include <sys/file.h>
100 # include <sys/stat.h>
101 # include <netinet/in.h>
102 # include <sys/time.h>
105 # include "rx_user.h"
106 # include "rx_clock.h"
107 # include "rx_queue.h"
108 # include "rx_globals.h"
109 # include "rx_trace.h"
110 # include <afs/rxgen_consts.h>
114 #ifdef AFS_PTHREAD_ENV
116 int (*registerProgram) (pid_t, char *) = 0;
117 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
120 int (*registerProgram) (PROCESS, char *) = 0;
121 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
125 /* Local static routines */
126 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
127 #ifdef RX_ENABLE_LOCKS
128 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
131 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
133 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
134 afs_int32 rxi_start_in_error;
136 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
139 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
140 * currently allocated within rx. This number is used to allocate the
141 * memory required to return the statistics when queried.
142 * Protected by the rx_rpc_stats mutex.
145 static unsigned int rxi_rpc_peer_stat_cnt;
148 * rxi_rpc_process_stat_cnt counts the total number of local process stat
149 * structures currently allocated within rx. The number is used to allocate
150 * the memory required to return the statistics when queried.
151 * Protected by the rx_rpc_stats mutex.
154 static unsigned int rxi_rpc_process_stat_cnt;
156 #if !defined(offsetof)
157 #include <stddef.h> /* for definition of offsetof() */
160 #ifdef AFS_PTHREAD_ENV
164 * Use procedural initialization of mutexes/condition variables
168 extern afs_kmutex_t rx_stats_mutex;
169 extern afs_kmutex_t rx_waiting_mutex;
170 extern afs_kmutex_t rx_quota_mutex;
171 extern afs_kmutex_t rx_pthread_mutex;
172 extern afs_kmutex_t rx_packets_mutex;
173 extern afs_kmutex_t des_init_mutex;
174 extern afs_kmutex_t des_random_mutex;
175 extern afs_kmutex_t rx_clock_mutex;
176 extern afs_kmutex_t rxi_connCacheMutex;
177 extern afs_kmutex_t rx_event_mutex;
178 extern afs_kmutex_t osi_malloc_mutex;
179 extern afs_kmutex_t event_handler_mutex;
180 extern afs_kmutex_t listener_mutex;
181 extern afs_kmutex_t rx_if_init_mutex;
182 extern afs_kmutex_t rx_if_mutex;
183 extern afs_kmutex_t rxkad_client_uid_mutex;
184 extern afs_kmutex_t rxkad_random_mutex;
186 extern afs_kcondvar_t rx_event_handler_cond;
187 extern afs_kcondvar_t rx_listener_cond;
189 static afs_kmutex_t epoch_mutex;
190 static afs_kmutex_t rx_init_mutex;
191 static afs_kmutex_t rx_debug_mutex;
192 static afs_kmutex_t rx_rpc_stats;
195 rxi_InitPthread(void)
197 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
198 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
199 MUTEX_INIT(&rx_waiting_mutex, "waiting", MUTEX_DEFAULT, 0);
200 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
201 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
202 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
203 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
204 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
205 MUTEX_INIT(&rx_event_mutex, "event", MUTEX_DEFAULT, 0);
206 MUTEX_INIT(&des_init_mutex, "des", MUTEX_DEFAULT, 0);
207 MUTEX_INIT(&des_random_mutex, "random", MUTEX_DEFAULT, 0);
208 MUTEX_INIT(&osi_malloc_mutex, "malloc", MUTEX_DEFAULT, 0);
209 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
210 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
211 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
212 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
213 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
214 MUTEX_INIT(&rxkad_client_uid_mutex, "uid", MUTEX_DEFAULT, 0);
215 MUTEX_INIT(&rxkad_random_mutex, "rxkad random", MUTEX_DEFAULT, 0);
216 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
218 assert(pthread_cond_init
219 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
220 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
222 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
223 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
225 rxkad_global_stats_init();
227 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
228 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
229 #ifdef RX_ENABLE_LOCKS
232 #endif /* RX_LOCKS_DB */
233 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
234 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
236 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
238 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
240 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
242 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
243 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
244 #endif /* RX_ENABLE_LOCKS */
247 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
248 #define INIT_PTHREAD_LOCKS \
249 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
251 * The rx_stats_mutex mutex protects the following global variables:
252 * rxi_lowConnRefCount
253 * rxi_lowPeerRefCount
262 * The rx_quota_mutex mutex protects the following global variables:
270 * The rx_freePktQ_lock protects the following global variables:
275 * The rx_packets_mutex mutex protects the following global variables:
283 * The rx_pthread_mutex mutex protects the following global variables:
287 #define INIT_PTHREAD_LOCKS
291 /* Variables for handling the minProcs implementation. availProcs gives the
292 * number of threads available in the pool at this moment (not counting dudes
293 * executing right now). totalMin gives the total number of procs required
294 * for handling all minProcs requests. minDeficit is a dynamic variable
295 * tracking the # of procs required to satisfy all of the remaining minProcs
297 * For fine grain locking to work, the quota check and the reservation of
298 * a server thread has to come while rxi_availProcs and rxi_minDeficit
299 * are locked. To this end, the code has been modified under #ifdef
300 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
301 * same time. A new function, ReturnToServerPool() returns the allocation.
303 * A call can be on several queue's (but only one at a time). When
304 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
305 * that no one else is touching the queue. To this end, we store the address
306 * of the queue lock in the call structure (under the call lock) when we
307 * put the call on a queue, and we clear the call_queue_lock when the
308 * call is removed from a queue (once the call lock has been obtained).
309 * This allows rxi_ResetCall to safely synchronize with others wishing
310 * to manipulate the queue.
313 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
314 static afs_kmutex_t rx_rpc_stats;
315 void rxi_StartUnlocked(struct rxevent *event, void *call,
316 void *arg1, int istack);
319 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
320 ** pretty good that the next packet coming in is from the same connection
321 ** as the last packet, since we're send multiple packets in a transmit window.
323 struct rx_connection *rxLastConn = 0;
325 #ifdef RX_ENABLE_LOCKS
326 /* The locking hierarchy for rx fine grain locking is composed of these
329 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
330 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
331 * call->lock - locks call data fields.
332 * These are independent of each other:
333 * rx_freeCallQueue_lock
338 * serverQueueEntry->lock
339 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
341 * peer->lock - locks peer data fields.
342 * conn_data_lock - that more than one thread is not updating a conn data
343 * field at the same time.
351 * Do we need a lock to protect the peer field in the conn structure?
352 * conn->peer was previously a constant for all intents and so has no
353 * lock protecting this field. The multihomed client delta introduced
354 * a RX code change : change the peer field in the connection structure
355 * to that remote interface from which the last packet for this
356 * connection was sent out. This may become an issue if further changes
359 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
360 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
362 /* rxdb_fileID is used to identify the lock location, along with line#. */
363 static int rxdb_fileID = RXDB_FILE_RX;
364 #endif /* RX_LOCKS_DB */
365 #else /* RX_ENABLE_LOCKS */
366 #define SET_CALL_QUEUE_LOCK(C, L)
367 #define CLEAR_CALL_QUEUE_LOCK(C)
368 #endif /* RX_ENABLE_LOCKS */
369 struct rx_serverQueueEntry *rx_waitForPacket = 0;
370 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
372 /* ------------Exported Interfaces------------- */
374 /* This function allows rxkad to set the epoch to a suitably random number
375 * which rx_NewConnection will use in the future. The principle purpose is to
376 * get rxnull connections to use the same epoch as the rxkad connections do, at
377 * least once the first rxkad connection is established. This is important now
378 * that the host/port addresses aren't used in FindConnection: the uniqueness
379 * of epoch/cid matters and the start time won't do. */
381 #ifdef AFS_PTHREAD_ENV
383 * This mutex protects the following global variables:
387 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
388 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
392 #endif /* AFS_PTHREAD_ENV */
395 rx_SetEpoch(afs_uint32 epoch)
402 /* Initialize rx. A port number may be mentioned, in which case this
403 * becomes the default port number for any service installed later.
404 * If 0 is provided for the port number, a random port will be chosen
405 * by the kernel. Whether this will ever overlap anything in
406 * /etc/services is anybody's guess... Returns 0 on success, -1 on
411 int rxinit_status = 1;
412 #ifdef AFS_PTHREAD_ENV
414 * This mutex protects the following global variables:
418 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
419 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
422 #define UNLOCK_RX_INIT
426 rx_InitHost(u_int host, u_int port)
433 char *htable, *ptable;
440 if (rxinit_status == 0) {
441 tmp_status = rxinit_status;
443 return tmp_status; /* Already started; return previous error code. */
449 if (afs_winsockInit() < 0)
455 * Initialize anything necessary to provide a non-premptive threading
458 rxi_InitializeThreadSupport();
461 /* Allocate and initialize a socket for client and perhaps server
464 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
465 if (rx_socket == OSI_NULLSOCKET) {
469 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
472 #endif /* RX_LOCKS_DB */
473 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
474 MUTEX_INIT(&rx_waiting_mutex, "rx_waiting_mutex", MUTEX_DEFAULT, 0);
475 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
476 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
477 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
478 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
479 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
480 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
481 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
483 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
485 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
487 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
489 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
490 #if defined(AFS_HPUX110_ENV)
492 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
493 #endif /* AFS_HPUX110_ENV */
494 #endif /* RX_ENABLE_LOCKS && KERNEL */
497 rx_connDeadTime = 12;
498 rx_tranquil = 0; /* reset flag */
499 memset(&rx_stats, 0, sizeof(struct rx_statistics));
501 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
502 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
503 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
504 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
505 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
506 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
508 /* Malloc up a bunch of packets & buffers */
510 queue_Init(&rx_freePacketQueue);
511 rxi_NeedMorePackets = FALSE;
512 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
514 /* enforce a minimum number of allocated packets */
515 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
516 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
518 /* allocate the initial free packet pool */
519 #ifdef RX_ENABLE_TSFPQ
520 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
521 #else /* RX_ENABLE_TSFPQ */
522 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
523 #endif /* RX_ENABLE_TSFPQ */
530 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
531 tv.tv_sec = clock_now.sec;
532 tv.tv_usec = clock_now.usec;
533 srand((unsigned int)tv.tv_usec);
540 #if defined(KERNEL) && !defined(UKERNEL)
541 /* Really, this should never happen in a real kernel */
544 struct sockaddr_in addr;
546 int addrlen = sizeof(addr);
548 socklen_t addrlen = sizeof(addr);
550 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
554 rx_port = addr.sin_port;
557 rx_stats.minRtt.sec = 9999999;
559 rx_SetEpoch(tv.tv_sec | 0x80000000);
561 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
562 * will provide a randomer value. */
564 MUTEX_ENTER(&rx_quota_mutex);
565 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
566 MUTEX_EXIT(&rx_quota_mutex);
567 /* *Slightly* random start time for the cid. This is just to help
568 * out with the hashing function at the peer */
569 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
570 rx_connHashTable = (struct rx_connection **)htable;
571 rx_peerHashTable = (struct rx_peer **)ptable;
573 rx_lastAckDelay.sec = 0;
574 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
575 rx_hardAckDelay.sec = 0;
576 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
577 rx_softAckDelay.sec = 0;
578 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
580 rxevent_Init(20, rxi_ReScheduleEvents);
582 /* Initialize various global queues */
583 queue_Init(&rx_idleServerQueue);
584 queue_Init(&rx_incomingCallQueue);
585 queue_Init(&rx_freeCallQueue);
587 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
588 /* Initialize our list of usable IP addresses. */
592 /* Start listener process (exact function is dependent on the
593 * implementation environment--kernel or user space) */
597 tmp_status = rxinit_status = 0;
605 return rx_InitHost(htonl(INADDR_ANY), port);
608 /* called with unincremented nRequestsRunning to see if it is OK to start
609 * a new thread in this service. Could be "no" for two reasons: over the
610 * max quota, or would prevent others from reaching their min quota.
612 #ifdef RX_ENABLE_LOCKS
613 /* This verion of QuotaOK reserves quota if it's ok while the
614 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
617 QuotaOK(struct rx_service *aservice)
619 /* check if over max quota */
620 if (aservice->nRequestsRunning >= aservice->maxProcs) {
624 /* under min quota, we're OK */
625 /* otherwise, can use only if there are enough to allow everyone
626 * to go to their min quota after this guy starts.
629 MUTEX_ENTER(&rx_quota_mutex);
630 if ((aservice->nRequestsRunning < aservice->minProcs)
631 || (rxi_availProcs > rxi_minDeficit)) {
632 aservice->nRequestsRunning++;
633 /* just started call in minProcs pool, need fewer to maintain
635 if (aservice->nRequestsRunning <= aservice->minProcs)
638 MUTEX_EXIT(&rx_quota_mutex);
641 MUTEX_EXIT(&rx_quota_mutex);
647 ReturnToServerPool(struct rx_service *aservice)
649 aservice->nRequestsRunning--;
650 MUTEX_ENTER(&rx_quota_mutex);
651 if (aservice->nRequestsRunning < aservice->minProcs)
654 MUTEX_EXIT(&rx_quota_mutex);
657 #else /* RX_ENABLE_LOCKS */
659 QuotaOK(struct rx_service *aservice)
662 /* under min quota, we're OK */
663 if (aservice->nRequestsRunning < aservice->minProcs)
666 /* check if over max quota */
667 if (aservice->nRequestsRunning >= aservice->maxProcs)
670 /* otherwise, can use only if there are enough to allow everyone
671 * to go to their min quota after this guy starts.
673 MUTEX_ENTER(&rx_quota_mutex);
674 if (rxi_availProcs > rxi_minDeficit)
676 MUTEX_EXIT(&rx_quota_mutex);
679 #endif /* RX_ENABLE_LOCKS */
682 /* Called by rx_StartServer to start up lwp's to service calls.
683 NExistingProcs gives the number of procs already existing, and which
684 therefore needn't be created. */
686 rxi_StartServerProcs(int nExistingProcs)
688 struct rx_service *service;
693 /* For each service, reserve N processes, where N is the "minimum"
694 * number of processes that MUST be able to execute a request in parallel,
695 * at any time, for that process. Also compute the maximum difference
696 * between any service's maximum number of processes that can run
697 * (i.e. the maximum number that ever will be run, and a guarantee
698 * that this number will run if other services aren't running), and its
699 * minimum number. The result is the extra number of processes that
700 * we need in order to provide the latter guarantee */
701 for (i = 0; i < RX_MAX_SERVICES; i++) {
703 service = rx_services[i];
704 if (service == (struct rx_service *)0)
706 nProcs += service->minProcs;
707 diff = service->maxProcs - service->minProcs;
711 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
712 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
713 for (i = 0; i < nProcs; i++) {
714 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
720 /* This routine is only required on Windows */
722 rx_StartClientThread(void)
724 #ifdef AFS_PTHREAD_ENV
726 pid = pthread_self();
727 #endif /* AFS_PTHREAD_ENV */
729 #endif /* AFS_NT40_ENV */
731 /* This routine must be called if any services are exported. If the
732 * donateMe flag is set, the calling process is donated to the server
735 rx_StartServer(int donateMe)
737 struct rx_service *service;
743 /* Start server processes, if necessary (exact function is dependent
744 * on the implementation environment--kernel or user space). DonateMe
745 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
746 * case, one less new proc will be created rx_StartServerProcs.
748 rxi_StartServerProcs(donateMe);
750 /* count up the # of threads in minProcs, and add set the min deficit to
751 * be that value, too.
753 for (i = 0; i < RX_MAX_SERVICES; i++) {
754 service = rx_services[i];
755 if (service == (struct rx_service *)0)
757 MUTEX_ENTER(&rx_quota_mutex);
758 rxi_totalMin += service->minProcs;
759 /* below works even if a thread is running, since minDeficit would
760 * still have been decremented and later re-incremented.
762 rxi_minDeficit += service->minProcs;
763 MUTEX_EXIT(&rx_quota_mutex);
766 /* Turn on reaping of idle server connections */
767 rxi_ReapConnections(NULL, NULL, NULL);
776 #ifdef AFS_PTHREAD_ENV
778 pid = afs_pointer_to_int(pthread_self());
779 #else /* AFS_PTHREAD_ENV */
781 LWP_CurrentProcess(&pid);
782 #endif /* AFS_PTHREAD_ENV */
784 sprintf(name, "srv_%d", ++nProcs);
786 (*registerProgram) (pid, name);
788 #endif /* AFS_NT40_ENV */
789 rx_ServerProc(NULL); /* Never returns */
791 #ifdef RX_ENABLE_TSFPQ
792 /* no use leaving packets around in this thread's local queue if
793 * it isn't getting donated to the server thread pool.
795 rxi_FlushLocalPacketsTSFPQ();
796 #endif /* RX_ENABLE_TSFPQ */
800 /* Create a new client connection to the specified service, using the
801 * specified security object to implement the security model for this
803 struct rx_connection *
804 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
805 struct rx_securityClass *securityObject,
806 int serviceSecurityIndex)
810 struct rx_connection *conn;
815 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
816 "serviceSecurityIndex %d)\n",
817 ntohl(shost), ntohs(sport), sservice, securityObject,
818 serviceSecurityIndex));
820 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
821 * the case of kmem_alloc? */
822 conn = rxi_AllocConnection();
823 #ifdef RX_ENABLE_LOCKS
824 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
825 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
826 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
829 MUTEX_ENTER(&rx_connHashTable_lock);
830 cid = (rx_nextCid += RX_MAXCALLS);
831 conn->type = RX_CLIENT_CONNECTION;
833 conn->epoch = rx_epoch;
834 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
835 conn->serviceId = sservice;
836 conn->securityObject = securityObject;
837 conn->securityData = (void *) 0;
838 conn->securityIndex = serviceSecurityIndex;
839 rx_SetConnDeadTime(conn, rx_connDeadTime);
840 rx_SetConnSecondsUntilNatPing(conn, 0);
841 conn->ackRate = RX_FAST_ACK_RATE;
843 conn->specific = NULL;
844 conn->challengeEvent = NULL;
845 conn->delayedAbortEvent = NULL;
846 conn->abortCount = 0;
848 for (i = 0; i < RX_MAXCALLS; i++) {
849 conn->twind[i] = rx_initSendWindow;
850 conn->rwind[i] = rx_initReceiveWindow;
853 RXS_NewConnection(securityObject, conn);
855 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
857 conn->refCount++; /* no lock required since only this thread knows... */
858 conn->next = rx_connHashTable[hashindex];
859 rx_connHashTable[hashindex] = conn;
861 rx_MutexIncrement(rx_stats.nClientConns, rx_stats_mutex);
862 MUTEX_EXIT(&rx_connHashTable_lock);
868 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
870 /* The idea is to set the dead time to a value that allows several
871 * keepalives to be dropped without timing out the connection. */
872 conn->secondsUntilDead = MAX(seconds, 6);
873 conn->secondsUntilPing = conn->secondsUntilDead / 6;
876 int rxi_lowPeerRefCount = 0;
877 int rxi_lowConnRefCount = 0;
880 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
881 * NOTE: must not be called with rx_connHashTable_lock held.
884 rxi_CleanupConnection(struct rx_connection *conn)
886 /* Notify the service exporter, if requested, that this connection
887 * is being destroyed */
888 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
889 (*conn->service->destroyConnProc) (conn);
891 /* Notify the security module that this connection is being destroyed */
892 RXS_DestroyConnection(conn->securityObject, conn);
894 /* If this is the last connection using the rx_peer struct, set its
895 * idle time to now. rxi_ReapConnections will reap it if it's still
896 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
898 MUTEX_ENTER(&rx_peerHashTable_lock);
899 if (conn->peer->refCount < 2) {
900 conn->peer->idleWhen = clock_Sec();
901 if (conn->peer->refCount < 1) {
902 conn->peer->refCount = 1;
903 if (rx_stats_active) {
904 MUTEX_ENTER(&rx_stats_mutex);
905 rxi_lowPeerRefCount++;
906 MUTEX_EXIT(&rx_stats_mutex);
910 conn->peer->refCount--;
911 MUTEX_EXIT(&rx_peerHashTable_lock);
915 if (conn->type == RX_SERVER_CONNECTION)
916 rx_MutexDecrement(rx_stats.nServerConns, rx_stats_mutex);
918 rx_MutexDecrement(rx_stats.nClientConns, rx_stats_mutex);
921 if (conn->specific) {
923 for (i = 0; i < conn->nSpecific; i++) {
924 if (conn->specific[i] && rxi_keyCreate_destructor[i])
925 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
926 conn->specific[i] = NULL;
928 free(conn->specific);
930 conn->specific = NULL;
934 MUTEX_DESTROY(&conn->conn_call_lock);
935 MUTEX_DESTROY(&conn->conn_data_lock);
936 CV_DESTROY(&conn->conn_call_cv);
938 rxi_FreeConnection(conn);
941 /* Destroy the specified connection */
943 rxi_DestroyConnection(struct rx_connection *conn)
945 MUTEX_ENTER(&rx_connHashTable_lock);
946 rxi_DestroyConnectionNoLock(conn);
947 /* conn should be at the head of the cleanup list */
948 if (conn == rx_connCleanup_list) {
949 rx_connCleanup_list = rx_connCleanup_list->next;
950 MUTEX_EXIT(&rx_connHashTable_lock);
951 rxi_CleanupConnection(conn);
953 #ifdef RX_ENABLE_LOCKS
955 MUTEX_EXIT(&rx_connHashTable_lock);
957 #endif /* RX_ENABLE_LOCKS */
961 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
963 struct rx_connection **conn_ptr;
965 struct rx_packet *packet;
972 MUTEX_ENTER(&conn->conn_data_lock);
973 if (conn->refCount > 0)
976 if (rx_stats_active) {
977 MUTEX_ENTER(&rx_stats_mutex);
978 rxi_lowConnRefCount++;
979 MUTEX_EXIT(&rx_stats_mutex);
983 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
984 /* Busy; wait till the last guy before proceeding */
985 MUTEX_EXIT(&conn->conn_data_lock);
990 /* If the client previously called rx_NewCall, but it is still
991 * waiting, treat this as a running call, and wait to destroy the
992 * connection later when the call completes. */
993 if ((conn->type == RX_CLIENT_CONNECTION)
994 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
995 conn->flags |= RX_CONN_DESTROY_ME;
996 MUTEX_EXIT(&conn->conn_data_lock);
1000 MUTEX_EXIT(&conn->conn_data_lock);
1002 /* Check for extant references to this connection */
1003 for (i = 0; i < RX_MAXCALLS; i++) {
1004 struct rx_call *call = conn->call[i];
1007 if (conn->type == RX_CLIENT_CONNECTION) {
1008 MUTEX_ENTER(&call->lock);
1009 if (call->delayedAckEvent) {
1010 /* Push the final acknowledgment out now--there
1011 * won't be a subsequent call to acknowledge the
1012 * last reply packets */
1013 rxevent_Cancel(call->delayedAckEvent, call,
1014 RX_CALL_REFCOUNT_DELAY);
1015 if (call->state == RX_STATE_PRECALL
1016 || call->state == RX_STATE_ACTIVE) {
1017 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1019 rxi_AckAll(NULL, call, 0);
1022 MUTEX_EXIT(&call->lock);
1026 #ifdef RX_ENABLE_LOCKS
1028 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1029 MUTEX_EXIT(&conn->conn_data_lock);
1031 /* Someone is accessing a packet right now. */
1035 #endif /* RX_ENABLE_LOCKS */
1038 /* Don't destroy the connection if there are any call
1039 * structures still in use */
1040 MUTEX_ENTER(&conn->conn_data_lock);
1041 conn->flags |= RX_CONN_DESTROY_ME;
1042 MUTEX_EXIT(&conn->conn_data_lock);
1047 if (conn->natKeepAliveEvent) {
1048 rxi_NatKeepAliveOff(conn);
1051 if (conn->delayedAbortEvent) {
1052 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1053 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1055 MUTEX_ENTER(&conn->conn_data_lock);
1056 rxi_SendConnectionAbort(conn, packet, 0, 1);
1057 MUTEX_EXIT(&conn->conn_data_lock);
1058 rxi_FreePacket(packet);
1062 /* Remove from connection hash table before proceeding */
1064 &rx_connHashTable[CONN_HASH
1065 (peer->host, peer->port, conn->cid, conn->epoch,
1067 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1068 if (*conn_ptr == conn) {
1069 *conn_ptr = conn->next;
1073 /* if the conn that we are destroying was the last connection, then we
1074 * clear rxLastConn as well */
1075 if (rxLastConn == conn)
1078 /* Make sure the connection is completely reset before deleting it. */
1079 /* get rid of pending events that could zap us later */
1080 if (conn->challengeEvent)
1081 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1082 if (conn->checkReachEvent)
1083 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1084 if (conn->natKeepAliveEvent)
1085 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
1087 /* Add the connection to the list of destroyed connections that
1088 * need to be cleaned up. This is necessary to avoid deadlocks
1089 * in the routines we call to inform others that this connection is
1090 * being destroyed. */
1091 conn->next = rx_connCleanup_list;
1092 rx_connCleanup_list = conn;
1095 /* Externally available version */
1097 rx_DestroyConnection(struct rx_connection *conn)
1102 rxi_DestroyConnection(conn);
1107 rx_GetConnection(struct rx_connection *conn)
1112 MUTEX_ENTER(&conn->conn_data_lock);
1114 MUTEX_EXIT(&conn->conn_data_lock);
1118 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1119 /* Wait for the transmit queue to no longer be busy.
1120 * requires the call->lock to be held */
1121 static void rxi_WaitforTQBusy(struct rx_call *call) {
1122 while (call->flags & RX_CALL_TQ_BUSY) {
1123 call->flags |= RX_CALL_TQ_WAIT;
1125 #ifdef RX_ENABLE_LOCKS
1126 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1127 CV_WAIT(&call->cv_tq, &call->lock);
1128 #else /* RX_ENABLE_LOCKS */
1129 osi_rxSleep(&call->tq);
1130 #endif /* RX_ENABLE_LOCKS */
1132 if (call->tqWaiters == 0) {
1133 call->flags &= ~RX_CALL_TQ_WAIT;
1139 /* Start a new rx remote procedure call, on the specified connection.
1140 * If wait is set to 1, wait for a free call channel; otherwise return
1141 * 0. Maxtime gives the maximum number of seconds this call may take,
1142 * after rx_NewCall returns. After this time interval, a call to any
1143 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1144 * For fine grain locking, we hold the conn_call_lock in order to
1145 * to ensure that we don't get signalle after we found a call in an active
1146 * state and before we go to sleep.
1149 rx_NewCall(struct rx_connection *conn)
1152 struct rx_call *call;
1153 struct clock queueTime;
1157 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1160 clock_GetTime(&queueTime);
1162 * Check if there are others waiting for a new call.
1163 * If so, let them go first to avoid starving them.
1164 * This is a fairly simple scheme, and might not be
1165 * a complete solution for large numbers of waiters.
1167 * makeCallWaiters keeps track of the number of
1168 * threads waiting to make calls and the
1169 * RX_CONN_MAKECALL_WAITING flag bit is used to
1170 * indicate that there are indeed calls waiting.
1171 * The flag is set when the waiter is incremented.
1172 * It is only cleared when makeCallWaiters is 0.
1173 * This prevents us from accidently destroying the
1174 * connection while it is potentially about to be used.
1176 MUTEX_ENTER(&conn->conn_call_lock);
1177 MUTEX_ENTER(&conn->conn_data_lock);
1178 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1179 conn->flags |= RX_CONN_MAKECALL_WAITING;
1180 conn->makeCallWaiters++;
1181 MUTEX_EXIT(&conn->conn_data_lock);
1183 #ifdef RX_ENABLE_LOCKS
1184 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1188 MUTEX_ENTER(&conn->conn_data_lock);
1189 conn->makeCallWaiters--;
1190 if (conn->makeCallWaiters == 0)
1191 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1194 /* We are now the active thread in rx_NewCall */
1195 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1196 MUTEX_EXIT(&conn->conn_data_lock);
1201 for (i = 0; i < RX_MAXCALLS; i++) {
1202 call = conn->call[i];
1204 if (call->state == RX_STATE_DALLY) {
1205 MUTEX_ENTER(&call->lock);
1206 if (call->state == RX_STATE_DALLY) {
1208 * We are setting the state to RX_STATE_RESET to
1209 * ensure that no one else will attempt to use this
1210 * call once we drop the conn->conn_call_lock and
1211 * call->lock. We must drop the conn->conn_call_lock
1212 * before calling rxi_ResetCall because the process
1213 * of clearing the transmit queue can block for an
1214 * extended period of time. If we block while holding
1215 * the conn->conn_call_lock, then all rx_EndCall
1216 * processing will block as well. This has a detrimental
1217 * effect on overall system performance.
1219 call->state = RX_STATE_RESET;
1220 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1221 MUTEX_EXIT(&conn->conn_call_lock);
1222 rxi_ResetCall(call, 0);
1223 (*call->callNumber)++;
1224 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1228 * If we failed to be able to safely obtain the
1229 * conn->conn_call_lock we will have to drop the
1230 * call->lock to avoid a deadlock. When the call->lock
1231 * is released the state of the call can change. If it
1232 * is no longer RX_STATE_RESET then some other thread is
1235 MUTEX_EXIT(&call->lock);
1236 MUTEX_ENTER(&conn->conn_call_lock);
1237 MUTEX_ENTER(&call->lock);
1239 if (call->state == RX_STATE_RESET)
1243 * If we get here it means that after dropping
1244 * the conn->conn_call_lock and call->lock that
1245 * the call is no longer ours. If we can't find
1246 * a free call in the remaining slots we should
1247 * not go immediately to RX_CONN_MAKECALL_WAITING
1248 * because by dropping the conn->conn_call_lock
1249 * we have given up synchronization with rx_EndCall.
1250 * Instead, cycle through one more time to see if
1251 * we can find a call that can call our own.
1253 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1256 MUTEX_EXIT(&call->lock);
1259 /* rxi_NewCall returns with mutex locked */
1260 call = rxi_NewCall(conn, i);
1261 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1265 if (i < RX_MAXCALLS) {
1271 MUTEX_ENTER(&conn->conn_data_lock);
1272 conn->flags |= RX_CONN_MAKECALL_WAITING;
1273 conn->makeCallWaiters++;
1274 MUTEX_EXIT(&conn->conn_data_lock);
1276 #ifdef RX_ENABLE_LOCKS
1277 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1281 MUTEX_ENTER(&conn->conn_data_lock);
1282 conn->makeCallWaiters--;
1283 if (conn->makeCallWaiters == 0)
1284 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1285 MUTEX_EXIT(&conn->conn_data_lock);
1287 /* Client is initially in send mode */
1288 call->state = RX_STATE_ACTIVE;
1289 call->error = conn->error;
1291 call->mode = RX_MODE_ERROR;
1293 call->mode = RX_MODE_SENDING;
1295 /* remember start time for call in case we have hard dead time limit */
1296 call->queueTime = queueTime;
1297 clock_GetTime(&call->startTime);
1298 hzero(call->bytesSent);
1299 hzero(call->bytesRcvd);
1301 /* Turn on busy protocol. */
1302 rxi_KeepAliveOn(call);
1304 /* Attempt MTU discovery */
1305 rxi_GrowMTUOn(call);
1308 * We are no longer the active thread in rx_NewCall
1310 MUTEX_ENTER(&conn->conn_data_lock);
1311 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1312 MUTEX_EXIT(&conn->conn_data_lock);
1315 * Wake up anyone else who might be giving us a chance to
1316 * run (see code above that avoids resource starvation).
1318 #ifdef RX_ENABLE_LOCKS
1319 CV_BROADCAST(&conn->conn_call_cv);
1323 MUTEX_EXIT(&conn->conn_call_lock);
1325 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1326 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1327 osi_Panic("rx_NewCall call about to be used without an empty tq");
1329 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1331 MUTEX_EXIT(&call->lock);
1334 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1339 rxi_HasActiveCalls(struct rx_connection *aconn)
1342 struct rx_call *tcall;
1346 for (i = 0; i < RX_MAXCALLS; i++) {
1347 if ((tcall = aconn->call[i])) {
1348 if ((tcall->state == RX_STATE_ACTIVE)
1349 || (tcall->state == RX_STATE_PRECALL)) {
1360 rxi_GetCallNumberVector(struct rx_connection *aconn,
1361 afs_int32 * aint32s)
1364 struct rx_call *tcall;
1368 for (i = 0; i < RX_MAXCALLS; i++) {
1369 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1370 aint32s[i] = aconn->callNumber[i] + 1;
1372 aint32s[i] = aconn->callNumber[i];
1379 rxi_SetCallNumberVector(struct rx_connection *aconn,
1380 afs_int32 * aint32s)
1383 struct rx_call *tcall;
1387 for (i = 0; i < RX_MAXCALLS; i++) {
1388 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1389 aconn->callNumber[i] = aint32s[i] - 1;
1391 aconn->callNumber[i] = aint32s[i];
1397 /* Advertise a new service. A service is named locally by a UDP port
1398 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1401 char *serviceName; Name for identification purposes (e.g. the
1402 service name might be used for probing for
1405 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1406 char *serviceName, struct rx_securityClass **securityObjects,
1407 int nSecurityObjects,
1408 afs_int32(*serviceProc) (struct rx_call * acall))
1410 osi_socket socket = OSI_NULLSOCKET;
1411 struct rx_service *tservice;
1417 if (serviceId == 0) {
1419 "rx_NewService: service id for service %s is not non-zero.\n",
1426 "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",
1434 tservice = rxi_AllocService();
1437 #ifdef RX_ENABLE_LOCKS
1438 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1441 for (i = 0; i < RX_MAX_SERVICES; i++) {
1442 struct rx_service *service = rx_services[i];
1444 if (port == service->servicePort && host == service->serviceHost) {
1445 if (service->serviceId == serviceId) {
1446 /* The identical service has already been
1447 * installed; if the caller was intending to
1448 * change the security classes used by this
1449 * service, he/she loses. */
1451 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1452 serviceName, serviceId, service->serviceName);
1454 rxi_FreeService(tservice);
1457 /* Different service, same port: re-use the socket
1458 * which is bound to the same port */
1459 socket = service->socket;
1462 if (socket == OSI_NULLSOCKET) {
1463 /* If we don't already have a socket (from another
1464 * service on same port) get a new one */
1465 socket = rxi_GetHostUDPSocket(host, port);
1466 if (socket == OSI_NULLSOCKET) {
1468 rxi_FreeService(tservice);
1473 service->socket = socket;
1474 service->serviceHost = host;
1475 service->servicePort = port;
1476 service->serviceId = serviceId;
1477 service->serviceName = serviceName;
1478 service->nSecurityObjects = nSecurityObjects;
1479 service->securityObjects = securityObjects;
1480 service->minProcs = 0;
1481 service->maxProcs = 1;
1482 service->idleDeadTime = 60;
1483 service->idleDeadErr = 0;
1484 service->connDeadTime = rx_connDeadTime;
1485 service->executeRequestProc = serviceProc;
1486 service->checkReach = 0;
1487 service->nSpecific = 0;
1488 service->specific = NULL;
1489 rx_services[i] = service; /* not visible until now */
1495 rxi_FreeService(tservice);
1496 (osi_Msg "rx_NewService: cannot support > %d services\n",
1501 /* Set configuration options for all of a service's security objects */
1504 rx_SetSecurityConfiguration(struct rx_service *service,
1505 rx_securityConfigVariables type,
1509 for (i = 0; i<service->nSecurityObjects; i++) {
1510 if (service->securityObjects[i]) {
1511 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1519 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1520 struct rx_securityClass **securityObjects, int nSecurityObjects,
1521 afs_int32(*serviceProc) (struct rx_call * acall))
1523 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1526 /* Generic request processing loop. This routine should be called
1527 * by the implementation dependent rx_ServerProc. If socketp is
1528 * non-null, it will be set to the file descriptor that this thread
1529 * is now listening on. If socketp is null, this routine will never
1532 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1534 struct rx_call *call;
1536 struct rx_service *tservice = NULL;
1543 call = rx_GetCall(threadID, tservice, socketp);
1544 if (socketp && *socketp != OSI_NULLSOCKET) {
1545 /* We are now a listener thread */
1550 /* if server is restarting( typically smooth shutdown) then do not
1551 * allow any new calls.
1554 if (rx_tranquil && (call != NULL)) {
1558 MUTEX_ENTER(&call->lock);
1560 rxi_CallError(call, RX_RESTARTING);
1561 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1563 MUTEX_EXIT(&call->lock);
1567 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1568 #ifdef RX_ENABLE_LOCKS
1570 #endif /* RX_ENABLE_LOCKS */
1571 afs_termState = AFSOP_STOP_AFS;
1572 afs_osi_Wakeup(&afs_termState);
1573 #ifdef RX_ENABLE_LOCKS
1575 #endif /* RX_ENABLE_LOCKS */
1580 tservice = call->conn->service;
1582 if (tservice->beforeProc)
1583 (*tservice->beforeProc) (call);
1585 code = call->conn->service->executeRequestProc(call);
1587 if (tservice->afterProc)
1588 (*tservice->afterProc) (call, code);
1590 rx_EndCall(call, code);
1591 if (rx_stats_active) {
1592 MUTEX_ENTER(&rx_stats_mutex);
1594 MUTEX_EXIT(&rx_stats_mutex);
1601 rx_WakeupServerProcs(void)
1603 struct rx_serverQueueEntry *np, *tqp;
1607 MUTEX_ENTER(&rx_serverPool_lock);
1609 #ifdef RX_ENABLE_LOCKS
1610 if (rx_waitForPacket)
1611 CV_BROADCAST(&rx_waitForPacket->cv);
1612 #else /* RX_ENABLE_LOCKS */
1613 if (rx_waitForPacket)
1614 osi_rxWakeup(rx_waitForPacket);
1615 #endif /* RX_ENABLE_LOCKS */
1616 MUTEX_ENTER(&freeSQEList_lock);
1617 for (np = rx_FreeSQEList; np; np = tqp) {
1618 tqp = *(struct rx_serverQueueEntry **)np;
1619 #ifdef RX_ENABLE_LOCKS
1620 CV_BROADCAST(&np->cv);
1621 #else /* RX_ENABLE_LOCKS */
1623 #endif /* RX_ENABLE_LOCKS */
1625 MUTEX_EXIT(&freeSQEList_lock);
1626 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1627 #ifdef RX_ENABLE_LOCKS
1628 CV_BROADCAST(&np->cv);
1629 #else /* RX_ENABLE_LOCKS */
1631 #endif /* RX_ENABLE_LOCKS */
1633 MUTEX_EXIT(&rx_serverPool_lock);
1638 * One thing that seems to happen is that all the server threads get
1639 * tied up on some empty or slow call, and then a whole bunch of calls
1640 * arrive at once, using up the packet pool, so now there are more
1641 * empty calls. The most critical resources here are server threads
1642 * and the free packet pool. The "doreclaim" code seems to help in
1643 * general. I think that eventually we arrive in this state: there
1644 * are lots of pending calls which do have all their packets present,
1645 * so they won't be reclaimed, are multi-packet calls, so they won't
1646 * be scheduled until later, and thus are tying up most of the free
1647 * packet pool for a very long time.
1649 * 1. schedule multi-packet calls if all the packets are present.
1650 * Probably CPU-bound operation, useful to return packets to pool.
1651 * Do what if there is a full window, but the last packet isn't here?
1652 * 3. preserve one thread which *only* runs "best" calls, otherwise
1653 * it sleeps and waits for that type of call.
1654 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1655 * the current dataquota business is badly broken. The quota isn't adjusted
1656 * to reflect how many packets are presently queued for a running call.
1657 * So, when we schedule a queued call with a full window of packets queued
1658 * up for it, that *should* free up a window full of packets for other 2d-class
1659 * calls to be able to use from the packet pool. But it doesn't.
1661 * NB. Most of the time, this code doesn't run -- since idle server threads
1662 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1663 * as a new call arrives.
1665 /* Sleep until a call arrives. Returns a pointer to the call, ready
1666 * for an rx_Read. */
1667 #ifdef RX_ENABLE_LOCKS
1669 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1671 struct rx_serverQueueEntry *sq;
1672 struct rx_call *call = (struct rx_call *)0;
1673 struct rx_service *service = NULL;
1676 MUTEX_ENTER(&freeSQEList_lock);
1678 if ((sq = rx_FreeSQEList)) {
1679 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1680 MUTEX_EXIT(&freeSQEList_lock);
1681 } else { /* otherwise allocate a new one and return that */
1682 MUTEX_EXIT(&freeSQEList_lock);
1683 sq = (struct rx_serverQueueEntry *)
1684 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1685 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1686 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1689 MUTEX_ENTER(&rx_serverPool_lock);
1690 if (cur_service != NULL) {
1691 ReturnToServerPool(cur_service);
1694 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1695 struct rx_call *tcall, *ncall, *choice2 = NULL;
1697 /* Scan for eligible incoming calls. A call is not eligible
1698 * if the maximum number of calls for its service type are
1699 * already executing */
1700 /* One thread will process calls FCFS (to prevent starvation),
1701 * while the other threads may run ahead looking for calls which
1702 * have all their input data available immediately. This helps
1703 * keep threads from blocking, waiting for data from the client. */
1704 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1705 service = tcall->conn->service;
1706 if (!QuotaOK(service)) {
1709 MUTEX_ENTER(&rx_pthread_mutex);
1710 if (tno == rxi_fcfs_thread_num
1711 || !tcall->queue_item_header.next) {
1712 MUTEX_EXIT(&rx_pthread_mutex);
1713 /* If we're the fcfs thread , then we'll just use
1714 * this call. If we haven't been able to find an optimal
1715 * choice, and we're at the end of the list, then use a
1716 * 2d choice if one has been identified. Otherwise... */
1717 call = (choice2 ? choice2 : tcall);
1718 service = call->conn->service;
1720 MUTEX_EXIT(&rx_pthread_mutex);
1721 if (!queue_IsEmpty(&tcall->rq)) {
1722 struct rx_packet *rp;
1723 rp = queue_First(&tcall->rq, rx_packet);
1724 if (rp->header.seq == 1) {
1726 || (rp->header.flags & RX_LAST_PACKET)) {
1728 } else if (rxi_2dchoice && !choice2
1729 && !(tcall->flags & RX_CALL_CLEARED)
1730 && (tcall->rprev > rxi_HardAckRate)) {
1740 ReturnToServerPool(service);
1747 MUTEX_EXIT(&rx_serverPool_lock);
1748 MUTEX_ENTER(&call->lock);
1750 if (call->flags & RX_CALL_WAIT_PROC) {
1751 call->flags &= ~RX_CALL_WAIT_PROC;
1752 MUTEX_ENTER(&rx_waiting_mutex);
1754 MUTEX_EXIT(&rx_waiting_mutex);
1757 if (call->state != RX_STATE_PRECALL || call->error) {
1758 MUTEX_EXIT(&call->lock);
1759 MUTEX_ENTER(&rx_serverPool_lock);
1760 ReturnToServerPool(service);
1765 if (queue_IsEmpty(&call->rq)
1766 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1767 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1769 CLEAR_CALL_QUEUE_LOCK(call);
1772 /* If there are no eligible incoming calls, add this process
1773 * to the idle server queue, to wait for one */
1777 *socketp = OSI_NULLSOCKET;
1779 sq->socketp = socketp;
1780 queue_Append(&rx_idleServerQueue, sq);
1781 #ifndef AFS_AIX41_ENV
1782 rx_waitForPacket = sq;
1784 rx_waitingForPacket = sq;
1785 #endif /* AFS_AIX41_ENV */
1787 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1789 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1790 MUTEX_EXIT(&rx_serverPool_lock);
1791 return (struct rx_call *)0;
1794 } while (!(call = sq->newcall)
1795 && !(socketp && *socketp != OSI_NULLSOCKET));
1796 MUTEX_EXIT(&rx_serverPool_lock);
1798 MUTEX_ENTER(&call->lock);
1804 MUTEX_ENTER(&freeSQEList_lock);
1805 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1806 rx_FreeSQEList = sq;
1807 MUTEX_EXIT(&freeSQEList_lock);
1810 clock_GetTime(&call->startTime);
1811 call->state = RX_STATE_ACTIVE;
1812 call->mode = RX_MODE_RECEIVING;
1813 #ifdef RX_KERNEL_TRACE
1814 if (ICL_SETACTIVE(afs_iclSetp)) {
1815 int glockOwner = ISAFS_GLOCK();
1818 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1819 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1826 rxi_calltrace(RX_CALL_START, call);
1827 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
1828 call->conn->service->servicePort, call->conn->service->serviceId,
1831 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1832 MUTEX_EXIT(&call->lock);
1834 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1839 #else /* RX_ENABLE_LOCKS */
1841 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1843 struct rx_serverQueueEntry *sq;
1844 struct rx_call *call = (struct rx_call *)0, *choice2;
1845 struct rx_service *service = NULL;
1849 MUTEX_ENTER(&freeSQEList_lock);
1851 if ((sq = rx_FreeSQEList)) {
1852 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1853 MUTEX_EXIT(&freeSQEList_lock);
1854 } else { /* otherwise allocate a new one and return that */
1855 MUTEX_EXIT(&freeSQEList_lock);
1856 sq = (struct rx_serverQueueEntry *)
1857 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1858 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1859 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1861 MUTEX_ENTER(&sq->lock);
1863 if (cur_service != NULL) {
1864 cur_service->nRequestsRunning--;
1865 MUTEX_ENTER(&rx_quota_mutex);
1866 if (cur_service->nRequestsRunning < cur_service->minProcs)
1869 MUTEX_EXIT(&rx_quota_mutex);
1871 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1872 struct rx_call *tcall, *ncall;
1873 /* Scan for eligible incoming calls. A call is not eligible
1874 * if the maximum number of calls for its service type are
1875 * already executing */
1876 /* One thread will process calls FCFS (to prevent starvation),
1877 * while the other threads may run ahead looking for calls which
1878 * have all their input data available immediately. This helps
1879 * keep threads from blocking, waiting for data from the client. */
1880 choice2 = (struct rx_call *)0;
1881 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1882 service = tcall->conn->service;
1883 if (QuotaOK(service)) {
1884 MUTEX_ENTER(&rx_pthread_mutex);
1885 if (tno == rxi_fcfs_thread_num
1886 || !tcall->queue_item_header.next) {
1887 MUTEX_EXIT(&rx_pthread_mutex);
1888 /* If we're the fcfs thread, then we'll just use
1889 * this call. If we haven't been able to find an optimal
1890 * choice, and we're at the end of the list, then use a
1891 * 2d choice if one has been identified. Otherwise... */
1892 call = (choice2 ? choice2 : tcall);
1893 service = call->conn->service;
1895 MUTEX_EXIT(&rx_pthread_mutex);
1896 if (!queue_IsEmpty(&tcall->rq)) {
1897 struct rx_packet *rp;
1898 rp = queue_First(&tcall->rq, rx_packet);
1899 if (rp->header.seq == 1
1901 || (rp->header.flags & RX_LAST_PACKET))) {
1903 } else if (rxi_2dchoice && !choice2
1904 && !(tcall->flags & RX_CALL_CLEARED)
1905 && (tcall->rprev > rxi_HardAckRate)) {
1919 /* we can't schedule a call if there's no data!!! */
1920 /* send an ack if there's no data, if we're missing the
1921 * first packet, or we're missing something between first
1922 * and last -- there's a "hole" in the incoming data. */
1923 if (queue_IsEmpty(&call->rq)
1924 || queue_First(&call->rq, rx_packet)->header.seq != 1
1925 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1926 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1928 call->flags &= (~RX_CALL_WAIT_PROC);
1929 service->nRequestsRunning++;
1930 /* just started call in minProcs pool, need fewer to maintain
1932 MUTEX_ENTER(&rx_quota_mutex);
1933 if (service->nRequestsRunning <= service->minProcs)
1936 MUTEX_EXIT(&rx_quota_mutex);
1938 /* MUTEX_EXIT(&call->lock); */
1940 /* If there are no eligible incoming calls, add this process
1941 * to the idle server queue, to wait for one */
1944 *socketp = OSI_NULLSOCKET;
1946 sq->socketp = socketp;
1947 queue_Append(&rx_idleServerQueue, sq);
1951 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1953 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1954 return (struct rx_call *)0;
1957 } while (!(call = sq->newcall)
1958 && !(socketp && *socketp != OSI_NULLSOCKET));
1960 MUTEX_EXIT(&sq->lock);
1962 MUTEX_ENTER(&freeSQEList_lock);
1963 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1964 rx_FreeSQEList = sq;
1965 MUTEX_EXIT(&freeSQEList_lock);
1968 clock_GetTime(&call->startTime);
1969 call->state = RX_STATE_ACTIVE;
1970 call->mode = RX_MODE_RECEIVING;
1971 #ifdef RX_KERNEL_TRACE
1972 if (ICL_SETACTIVE(afs_iclSetp)) {
1973 int glockOwner = ISAFS_GLOCK();
1976 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1977 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1984 rxi_calltrace(RX_CALL_START, call);
1985 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
1986 call->conn->service->servicePort, call->conn->service->serviceId,
1989 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1996 #endif /* RX_ENABLE_LOCKS */
2000 /* Establish a procedure to be called when a packet arrives for a
2001 * call. This routine will be called at most once after each call,
2002 * and will also be called if there is an error condition on the or
2003 * the call is complete. Used by multi rx to build a selection
2004 * function which determines which of several calls is likely to be a
2005 * good one to read from.
2006 * NOTE: the way this is currently implemented it is probably only a
2007 * good idea to (1) use it immediately after a newcall (clients only)
2008 * and (2) only use it once. Other uses currently void your warranty
2011 rx_SetArrivalProc(struct rx_call *call,
2012 void (*proc) (struct rx_call * call,
2015 void * handle, int arg)
2017 call->arrivalProc = proc;
2018 call->arrivalProcHandle = handle;
2019 call->arrivalProcArg = arg;
2022 /* Call is finished (possibly prematurely). Return rc to the peer, if
2023 * appropriate, and return the final error code from the conversation
2027 rx_EndCall(struct rx_call *call, afs_int32 rc)
2029 struct rx_connection *conn = call->conn;
2030 struct rx_service *service;
2034 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2035 call, rc, call->error, call->abortCode));
2038 MUTEX_ENTER(&call->lock);
2040 if (rc == 0 && call->error == 0) {
2041 call->abortCode = 0;
2042 call->abortCount = 0;
2045 call->arrivalProc = (void (*)())0;
2046 if (rc && call->error == 0) {
2047 rxi_CallError(call, rc);
2048 /* Send an abort message to the peer if this error code has
2049 * only just been set. If it was set previously, assume the
2050 * peer has already been sent the error code or will request it
2052 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2054 if (conn->type == RX_SERVER_CONNECTION) {
2055 /* Make sure reply or at least dummy reply is sent */
2056 if (call->mode == RX_MODE_RECEIVING) {
2057 rxi_WriteProc(call, 0, 0);
2059 if (call->mode == RX_MODE_SENDING) {
2060 rxi_FlushWrite(call);
2062 service = conn->service;
2063 rxi_calltrace(RX_CALL_END, call);
2064 /* Call goes to hold state until reply packets are acknowledged */
2065 if (call->tfirst + call->nSoftAcked < call->tnext) {
2066 call->state = RX_STATE_HOLD;
2068 call->state = RX_STATE_DALLY;
2069 rxi_ClearTransmitQueue(call, 0);
2070 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
2071 rxevent_Cancel(call->keepAliveEvent, call,
2072 RX_CALL_REFCOUNT_ALIVE);
2074 } else { /* Client connection */
2076 /* Make sure server receives input packets, in the case where
2077 * no reply arguments are expected */
2078 if ((call->mode == RX_MODE_SENDING)
2079 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2080 (void)rxi_ReadProc(call, &dummy, 1);
2083 /* If we had an outstanding delayed ack, be nice to the server
2084 * and force-send it now.
2086 if (call->delayedAckEvent) {
2087 rxevent_Cancel(call->delayedAckEvent, call,
2088 RX_CALL_REFCOUNT_DELAY);
2089 call->delayedAckEvent = NULL;
2090 rxi_SendDelayedAck(NULL, call, NULL);
2093 /* We need to release the call lock since it's lower than the
2094 * conn_call_lock and we don't want to hold the conn_call_lock
2095 * over the rx_ReadProc call. The conn_call_lock needs to be held
2096 * here for the case where rx_NewCall is perusing the calls on
2097 * the connection structure. We don't want to signal until
2098 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2099 * have checked this call, found it active and by the time it
2100 * goes to sleep, will have missed the signal.
2102 MUTEX_EXIT(&call->lock);
2103 MUTEX_ENTER(&conn->conn_call_lock);
2104 MUTEX_ENTER(&call->lock);
2105 MUTEX_ENTER(&conn->conn_data_lock);
2106 conn->flags |= RX_CONN_BUSY;
2107 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2108 MUTEX_EXIT(&conn->conn_data_lock);
2109 #ifdef RX_ENABLE_LOCKS
2110 CV_BROADCAST(&conn->conn_call_cv);
2115 #ifdef RX_ENABLE_LOCKS
2117 MUTEX_EXIT(&conn->conn_data_lock);
2119 #endif /* RX_ENABLE_LOCKS */
2120 call->state = RX_STATE_DALLY;
2122 error = call->error;
2124 /* currentPacket, nLeft, and NFree must be zeroed here, because
2125 * ResetCall cannot: ResetCall may be called at splnet(), in the
2126 * kernel version, and may interrupt the macros rx_Read or
2127 * rx_Write, which run at normal priority for efficiency. */
2128 if (call->currentPacket) {
2129 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2130 rxi_FreePacket(call->currentPacket);
2131 call->currentPacket = (struct rx_packet *)0;
2134 call->nLeft = call->nFree = call->curlen = 0;
2136 /* Free any packets from the last call to ReadvProc/WritevProc */
2137 #ifdef RXDEBUG_PACKET
2139 #endif /* RXDEBUG_PACKET */
2140 rxi_FreePackets(0, &call->iovq);
2142 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2143 MUTEX_EXIT(&call->lock);
2144 if (conn->type == RX_CLIENT_CONNECTION) {
2145 MUTEX_ENTER(&conn->conn_data_lock);
2146 conn->flags &= ~RX_CONN_BUSY;
2147 MUTEX_EXIT(&conn->conn_data_lock);
2148 MUTEX_EXIT(&conn->conn_call_lock);
2152 * Map errors to the local host's errno.h format.
2154 error = ntoh_syserr_conv(error);
2158 #if !defined(KERNEL)
2160 /* Call this routine when shutting down a server or client (especially
2161 * clients). This will allow Rx to gracefully garbage collect server
2162 * connections, and reduce the number of retries that a server might
2163 * make to a dead client.
2164 * This is not quite right, since some calls may still be ongoing and
2165 * we can't lock them to destroy them. */
2169 struct rx_connection **conn_ptr, **conn_end;
2173 if (rxinit_status == 1) {
2175 return; /* Already shutdown. */
2177 rxi_DeleteCachedConnections();
2178 if (rx_connHashTable) {
2179 MUTEX_ENTER(&rx_connHashTable_lock);
2180 for (conn_ptr = &rx_connHashTable[0], conn_end =
2181 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2183 struct rx_connection *conn, *next;
2184 for (conn = *conn_ptr; conn; conn = next) {
2186 if (conn->type == RX_CLIENT_CONNECTION) {
2187 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2189 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2190 #ifdef RX_ENABLE_LOCKS
2191 rxi_DestroyConnectionNoLock(conn);
2192 #else /* RX_ENABLE_LOCKS */
2193 rxi_DestroyConnection(conn);
2194 #endif /* RX_ENABLE_LOCKS */
2198 #ifdef RX_ENABLE_LOCKS
2199 while (rx_connCleanup_list) {
2200 struct rx_connection *conn;
2201 conn = rx_connCleanup_list;
2202 rx_connCleanup_list = rx_connCleanup_list->next;
2203 MUTEX_EXIT(&rx_connHashTable_lock);
2204 rxi_CleanupConnection(conn);
2205 MUTEX_ENTER(&rx_connHashTable_lock);
2207 MUTEX_EXIT(&rx_connHashTable_lock);
2208 #endif /* RX_ENABLE_LOCKS */
2213 afs_winsockCleanup();
2221 /* if we wakeup packet waiter too often, can get in loop with two
2222 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2224 rxi_PacketsUnWait(void)
2226 if (!rx_waitingForPackets) {
2230 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2231 return; /* still over quota */
2234 rx_waitingForPackets = 0;
2235 #ifdef RX_ENABLE_LOCKS
2236 CV_BROADCAST(&rx_waitingForPackets_cv);
2238 osi_rxWakeup(&rx_waitingForPackets);
2244 /* ------------------Internal interfaces------------------------- */
2246 /* Return this process's service structure for the
2247 * specified socket and service */
2249 rxi_FindService(osi_socket socket, u_short serviceId)
2251 struct rx_service **sp;
2252 for (sp = &rx_services[0]; *sp; sp++) {
2253 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2259 #ifdef RXDEBUG_PACKET
2260 #ifdef KDUMP_RX_LOCK
2261 static struct rx_call_rx_lock *rx_allCallsp = 0;
2263 static struct rx_call *rx_allCallsp = 0;
2265 #endif /* RXDEBUG_PACKET */
2267 /* Allocate a call structure, for the indicated channel of the
2268 * supplied connection. The mode and state of the call must be set by
2269 * the caller. Returns the call with mutex locked. */
2271 rxi_NewCall(struct rx_connection *conn, int channel)
2273 struct rx_call *call;
2274 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2275 struct rx_call *cp; /* Call pointer temp */
2276 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2277 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2279 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2281 /* Grab an existing call structure, or allocate a new one.
2282 * Existing call structures are assumed to have been left reset by
2284 MUTEX_ENTER(&rx_freeCallQueue_lock);
2286 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2288 * EXCEPT that the TQ might not yet be cleared out.
2289 * Skip over those with in-use TQs.
2292 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2293 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2299 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2300 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2301 call = queue_First(&rx_freeCallQueue, rx_call);
2302 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2304 if (rx_stats_active)
2305 rx_MutexDecrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2306 MUTEX_EXIT(&rx_freeCallQueue_lock);
2307 MUTEX_ENTER(&call->lock);
2308 CLEAR_CALL_QUEUE_LOCK(call);
2309 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2310 /* Now, if TQ wasn't cleared earlier, do it now. */
2311 rxi_WaitforTQBusy(call);
2312 if (call->flags & RX_CALL_TQ_CLEARME) {
2313 rxi_ClearTransmitQueue(call, 1);
2314 /*queue_Init(&call->tq);*/
2316 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2317 /* Bind the call to its connection structure */
2319 rxi_ResetCall(call, 1);
2322 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2323 #ifdef RXDEBUG_PACKET
2324 call->allNextp = rx_allCallsp;
2325 rx_allCallsp = call;
2327 #endif /* RXDEBUG_PACKET */
2328 rx_MutexIncrement(rx_stats.nCallStructs, rx_stats_mutex);
2330 MUTEX_EXIT(&rx_freeCallQueue_lock);
2331 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2332 MUTEX_ENTER(&call->lock);
2333 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2334 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2335 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2337 /* Initialize once-only items */
2338 queue_Init(&call->tq);
2339 queue_Init(&call->rq);
2340 queue_Init(&call->iovq);
2341 #ifdef RXDEBUG_PACKET
2342 call->rqc = call->tqc = call->iovqc = 0;
2343 #endif /* RXDEBUG_PACKET */
2344 /* Bind the call to its connection structure (prereq for reset) */
2346 rxi_ResetCall(call, 1);
2348 call->channel = channel;
2349 call->callNumber = &conn->callNumber[channel];
2350 call->rwind = conn->rwind[channel];
2351 call->twind = conn->twind[channel];
2352 /* Note that the next expected call number is retained (in
2353 * conn->callNumber[i]), even if we reallocate the call structure
2355 conn->call[channel] = call;
2356 /* if the channel's never been used (== 0), we should start at 1, otherwise
2357 * the call number is valid from the last time this channel was used */
2358 if (*call->callNumber == 0)
2359 *call->callNumber = 1;
2364 /* A call has been inactive long enough that so we can throw away
2365 * state, including the call structure, which is placed on the call
2367 * Call is locked upon entry.
2368 * haveCTLock set if called from rxi_ReapConnections
2370 #ifdef RX_ENABLE_LOCKS
2372 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2373 #else /* RX_ENABLE_LOCKS */
2375 rxi_FreeCall(struct rx_call *call)
2376 #endif /* RX_ENABLE_LOCKS */
2378 int channel = call->channel;
2379 struct rx_connection *conn = call->conn;
2382 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2383 (*call->callNumber)++;
2384 rxi_ResetCall(call, 0);
2385 call->conn->call[channel] = (struct rx_call *)0;
2387 MUTEX_ENTER(&rx_freeCallQueue_lock);
2388 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2389 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2390 /* A call may be free even though its transmit queue is still in use.
2391 * Since we search the call list from head to tail, put busy calls at
2392 * the head of the list, and idle calls at the tail.
2394 if (call->flags & RX_CALL_TQ_BUSY)
2395 queue_Prepend(&rx_freeCallQueue, call);
2397 queue_Append(&rx_freeCallQueue, call);
2398 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2399 queue_Append(&rx_freeCallQueue, call);
2400 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2401 if (rx_stats_active)
2402 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2403 MUTEX_EXIT(&rx_freeCallQueue_lock);
2405 /* Destroy the connection if it was previously slated for
2406 * destruction, i.e. the Rx client code previously called
2407 * rx_DestroyConnection (client connections), or
2408 * rxi_ReapConnections called the same routine (server
2409 * connections). Only do this, however, if there are no
2410 * outstanding calls. Note that for fine grain locking, there appears
2411 * to be a deadlock in that rxi_FreeCall has a call locked and
2412 * DestroyConnectionNoLock locks each call in the conn. But note a
2413 * few lines up where we have removed this call from the conn.
2414 * If someone else destroys a connection, they either have no
2415 * call lock held or are going through this section of code.
2417 MUTEX_ENTER(&conn->conn_data_lock);
2418 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2420 MUTEX_EXIT(&conn->conn_data_lock);
2421 #ifdef RX_ENABLE_LOCKS
2423 rxi_DestroyConnectionNoLock(conn);
2425 rxi_DestroyConnection(conn);
2426 #else /* RX_ENABLE_LOCKS */
2427 rxi_DestroyConnection(conn);
2428 #endif /* RX_ENABLE_LOCKS */
2430 MUTEX_EXIT(&conn->conn_data_lock);
2434 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2436 rxi_Alloc(size_t size)
2440 if (rx_stats_active)
2441 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2444 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2445 afs_osi_Alloc_NoSleep(size);
2450 osi_Panic("rxi_Alloc error");
2456 rxi_Free(void *addr, size_t size)
2458 if (rx_stats_active)
2459 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2460 osi_Free(addr, size);
2464 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2466 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2467 struct rx_peer *next = NULL;
2471 MUTEX_ENTER(&rx_peerHashTable_lock);
2473 peer_ptr = &rx_peerHashTable[0];
2474 peer_end = &rx_peerHashTable[rx_hashTableSize];
2477 for ( ; peer_ptr < peer_end; peer_ptr++) {
2480 for ( ; peer; peer = next) {
2482 if (host == peer->host)
2487 hashIndex = PEER_HASH(host, port);
2488 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2489 if ((peer->host == host) && (peer->port == port))
2494 MUTEX_ENTER(&rx_peerHashTable_lock);
2499 MUTEX_EXIT(&rx_peerHashTable_lock);
2501 MUTEX_ENTER(&peer->peer_lock);
2502 /* We don't handle dropping below min, so don't */
2503 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2504 peer->ifMTU=MIN(mtu, peer->ifMTU);
2505 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2506 /* if we tweaked this down, need to tune our peer MTU too */
2507 peer->MTU = MIN(peer->MTU, peer->natMTU);
2508 /* if we discovered a sub-1500 mtu, degrade */
2509 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2510 peer->maxDgramPackets = 1;
2511 /* We no longer have valid peer packet information */
2512 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2513 peer->maxPacketSize = 0;
2514 MUTEX_EXIT(&peer->peer_lock);
2516 MUTEX_ENTER(&rx_peerHashTable_lock);
2518 if (host && !port) {
2520 /* pick up where we left off */
2524 MUTEX_EXIT(&rx_peerHashTable_lock);
2527 /* Find the peer process represented by the supplied (host,port)
2528 * combination. If there is no appropriate active peer structure, a
2529 * new one will be allocated and initialized
2530 * The origPeer, if set, is a pointer to a peer structure on which the
2531 * refcount will be be decremented. This is used to replace the peer
2532 * structure hanging off a connection structure */
2534 rxi_FindPeer(afs_uint32 host, u_short port,
2535 struct rx_peer *origPeer, int create)
2539 hashIndex = PEER_HASH(host, port);
2540 MUTEX_ENTER(&rx_peerHashTable_lock);
2541 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2542 if ((pp->host == host) && (pp->port == port))
2547 pp = rxi_AllocPeer(); /* This bzero's *pp */
2548 pp->host = host; /* set here or in InitPeerParams is zero */
2550 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2551 queue_Init(&pp->congestionQueue);
2552 queue_Init(&pp->rpcStats);
2553 pp->next = rx_peerHashTable[hashIndex];
2554 rx_peerHashTable[hashIndex] = pp;
2555 rxi_InitPeerParams(pp);
2556 if (rx_stats_active)
2557 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2564 origPeer->refCount--;
2565 MUTEX_EXIT(&rx_peerHashTable_lock);
2570 /* Find the connection at (host, port) started at epoch, and with the
2571 * given connection id. Creates the server connection if necessary.
2572 * The type specifies whether a client connection or a server
2573 * connection is desired. In both cases, (host, port) specify the
2574 * peer's (host, pair) pair. Client connections are not made
2575 * automatically by this routine. The parameter socket gives the
2576 * socket descriptor on which the packet was received. This is used,
2577 * in the case of server connections, to check that *new* connections
2578 * come via a valid (port, serviceId). Finally, the securityIndex
2579 * parameter must match the existing index for the connection. If a
2580 * server connection is created, it will be created using the supplied
2581 * index, if the index is valid for this service */
2582 struct rx_connection *
2583 rxi_FindConnection(osi_socket socket, afs_int32 host,
2584 u_short port, u_short serviceId, afs_uint32 cid,
2585 afs_uint32 epoch, int type, u_int securityIndex)
2587 int hashindex, flag, i;
2588 struct rx_connection *conn;
2589 hashindex = CONN_HASH(host, port, cid, epoch, type);
2590 MUTEX_ENTER(&rx_connHashTable_lock);
2591 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2592 rx_connHashTable[hashindex],
2595 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2596 && (epoch == conn->epoch)) {
2597 struct rx_peer *pp = conn->peer;
2598 if (securityIndex != conn->securityIndex) {
2599 /* this isn't supposed to happen, but someone could forge a packet
2600 * like this, and there seems to be some CM bug that makes this
2601 * happen from time to time -- in which case, the fileserver
2603 MUTEX_EXIT(&rx_connHashTable_lock);
2604 return (struct rx_connection *)0;
2606 if (pp->host == host && pp->port == port)
2608 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2610 /* So what happens when it's a callback connection? */
2611 if ( /*type == RX_CLIENT_CONNECTION && */
2612 (conn->epoch & 0x80000000))
2616 /* the connection rxLastConn that was used the last time is not the
2617 ** one we are looking for now. Hence, start searching in the hash */
2619 conn = rx_connHashTable[hashindex];
2624 struct rx_service *service;
2625 if (type == RX_CLIENT_CONNECTION) {
2626 MUTEX_EXIT(&rx_connHashTable_lock);
2627 return (struct rx_connection *)0;
2629 service = rxi_FindService(socket, serviceId);
2630 if (!service || (securityIndex >= service->nSecurityObjects)
2631 || (service->securityObjects[securityIndex] == 0)) {
2632 MUTEX_EXIT(&rx_connHashTable_lock);
2633 return (struct rx_connection *)0;
2635 conn = rxi_AllocConnection(); /* This bzero's the connection */
2636 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2637 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2638 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2639 conn->next = rx_connHashTable[hashindex];
2640 rx_connHashTable[hashindex] = conn;
2641 conn->peer = rxi_FindPeer(host, port, 0, 1);
2642 conn->type = RX_SERVER_CONNECTION;
2643 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2644 conn->epoch = epoch;
2645 conn->cid = cid & RX_CIDMASK;
2646 /* conn->serial = conn->lastSerial = 0; */
2647 /* conn->timeout = 0; */
2648 conn->ackRate = RX_FAST_ACK_RATE;
2649 conn->service = service;
2650 conn->serviceId = serviceId;
2651 conn->securityIndex = securityIndex;
2652 conn->securityObject = service->securityObjects[securityIndex];
2653 conn->nSpecific = 0;
2654 conn->specific = NULL;
2655 rx_SetConnDeadTime(conn, service->connDeadTime);
2656 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2657 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2658 for (i = 0; i < RX_MAXCALLS; i++) {
2659 conn->twind[i] = rx_initSendWindow;
2660 conn->rwind[i] = rx_initReceiveWindow;
2662 /* Notify security object of the new connection */
2663 RXS_NewConnection(conn->securityObject, conn);
2664 /* XXXX Connection timeout? */
2665 if (service->newConnProc)
2666 (*service->newConnProc) (conn);
2667 if (rx_stats_active)
2668 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2671 MUTEX_ENTER(&conn->conn_data_lock);
2673 MUTEX_EXIT(&conn->conn_data_lock);
2675 rxLastConn = conn; /* store this connection as the last conn used */
2676 MUTEX_EXIT(&rx_connHashTable_lock);
2680 /* There are two packet tracing routines available for testing and monitoring
2681 * Rx. One is called just after every packet is received and the other is
2682 * called just before every packet is sent. Received packets, have had their
2683 * headers decoded, and packets to be sent have not yet had their headers
2684 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2685 * containing the network address. Both can be modified. The return value, if
2686 * non-zero, indicates that the packet should be dropped. */
2688 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2689 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2691 /* A packet has been received off the interface. Np is the packet, socket is
2692 * the socket number it was received from (useful in determining which service
2693 * this packet corresponds to), and (host, port) reflect the host,port of the
2694 * sender. This call returns the packet to the caller if it is finished with
2695 * it, rather than de-allocating it, just as a small performance hack */
2698 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
2699 afs_uint32 host, u_short port, int *tnop,
2700 struct rx_call **newcallp)
2702 struct rx_call *call;
2703 struct rx_connection *conn;
2705 afs_uint32 currentCallNumber;
2711 struct rx_packet *tnp;
2714 /* We don't print out the packet until now because (1) the time may not be
2715 * accurate enough until now in the lwp implementation (rx_Listener only gets
2716 * the time after the packet is read) and (2) from a protocol point of view,
2717 * this is the first time the packet has been seen */
2718 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2719 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2720 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT,
2721 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2722 np->header.epoch, np->header.cid, np->header.callNumber,
2723 np->header.seq, np->header.flags, np));
2726 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2727 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2730 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2731 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2734 /* If an input tracer function is defined, call it with the packet and
2735 * network address. Note this function may modify its arguments. */
2736 if (rx_justReceived) {
2737 struct sockaddr_in addr;
2739 addr.sin_family = AF_INET;
2740 addr.sin_port = port;
2741 addr.sin_addr.s_addr = host;
2742 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2743 addr.sin_len = sizeof(addr);
2744 #endif /* AFS_OSF_ENV */
2745 drop = (*rx_justReceived) (np, &addr);
2746 /* drop packet if return value is non-zero */
2749 port = addr.sin_port; /* in case fcn changed addr */
2750 host = addr.sin_addr.s_addr;
2754 /* If packet was not sent by the client, then *we* must be the client */
2755 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2756 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2758 /* Find the connection (or fabricate one, if we're the server & if
2759 * necessary) associated with this packet */
2761 rxi_FindConnection(socket, host, port, np->header.serviceId,
2762 np->header.cid, np->header.epoch, type,
2763 np->header.securityIndex);
2766 /* If no connection found or fabricated, just ignore the packet.
2767 * (An argument could be made for sending an abort packet for
2772 MUTEX_ENTER(&conn->conn_data_lock);
2773 if (conn->maxSerial < np->header.serial)
2774 conn->maxSerial = np->header.serial;
2775 MUTEX_EXIT(&conn->conn_data_lock);
2777 /* If the connection is in an error state, send an abort packet and ignore
2778 * the incoming packet */
2780 /* Don't respond to an abort packet--we don't want loops! */
2781 MUTEX_ENTER(&conn->conn_data_lock);
2782 if (np->header.type != RX_PACKET_TYPE_ABORT)
2783 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2785 MUTEX_EXIT(&conn->conn_data_lock);
2789 /* Check for connection-only requests (i.e. not call specific). */
2790 if (np->header.callNumber == 0) {
2791 switch (np->header.type) {
2792 case RX_PACKET_TYPE_ABORT: {
2793 /* What if the supplied error is zero? */
2794 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2795 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2796 rxi_ConnectionError(conn, errcode);
2797 MUTEX_ENTER(&conn->conn_data_lock);
2799 MUTEX_EXIT(&conn->conn_data_lock);
2802 case RX_PACKET_TYPE_CHALLENGE:
2803 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2804 MUTEX_ENTER(&conn->conn_data_lock);
2806 MUTEX_EXIT(&conn->conn_data_lock);
2808 case RX_PACKET_TYPE_RESPONSE:
2809 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2810 MUTEX_ENTER(&conn->conn_data_lock);
2812 MUTEX_EXIT(&conn->conn_data_lock);
2814 case RX_PACKET_TYPE_PARAMS:
2815 case RX_PACKET_TYPE_PARAMS + 1:
2816 case RX_PACKET_TYPE_PARAMS + 2:
2817 /* ignore these packet types for now */
2818 MUTEX_ENTER(&conn->conn_data_lock);
2820 MUTEX_EXIT(&conn->conn_data_lock);
2825 /* Should not reach here, unless the peer is broken: send an
2827 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2828 MUTEX_ENTER(&conn->conn_data_lock);
2829 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2831 MUTEX_EXIT(&conn->conn_data_lock);
2836 channel = np->header.cid & RX_CHANNELMASK;
2837 call = conn->call[channel];
2838 #ifdef RX_ENABLE_LOCKS
2840 MUTEX_ENTER(&call->lock);
2841 /* Test to see if call struct is still attached to conn. */
2842 if (call != conn->call[channel]) {
2844 MUTEX_EXIT(&call->lock);
2845 if (type == RX_SERVER_CONNECTION) {
2846 call = conn->call[channel];
2847 /* If we started with no call attached and there is one now,
2848 * another thread is also running this routine and has gotten
2849 * the connection channel. We should drop this packet in the tests
2850 * below. If there was a call on this connection and it's now
2851 * gone, then we'll be making a new call below.
2852 * If there was previously a call and it's now different then
2853 * the old call was freed and another thread running this routine
2854 * has created a call on this channel. One of these two threads
2855 * has a packet for the old call and the code below handles those
2859 MUTEX_ENTER(&call->lock);
2861 /* This packet can't be for this call. If the new call address is
2862 * 0 then no call is running on this channel. If there is a call
2863 * then, since this is a client connection we're getting data for
2864 * it must be for the previous call.
2866 if (rx_stats_active)
2867 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2868 MUTEX_ENTER(&conn->conn_data_lock);
2870 MUTEX_EXIT(&conn->conn_data_lock);
2875 currentCallNumber = conn->callNumber[channel];
2877 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2878 if (np->header.callNumber < currentCallNumber) {
2879 if (rx_stats_active)
2880 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2881 #ifdef RX_ENABLE_LOCKS
2883 MUTEX_EXIT(&call->lock);
2885 MUTEX_ENTER(&conn->conn_data_lock);
2887 MUTEX_EXIT(&conn->conn_data_lock);
2891 MUTEX_ENTER(&conn->conn_call_lock);
2892 call = rxi_NewCall(conn, channel);
2893 MUTEX_EXIT(&conn->conn_call_lock);
2894 *call->callNumber = np->header.callNumber;
2896 if (np->header.callNumber == 0)
2897 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%.06d len %d",
2898 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
2899 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
2900 np->header.flags, np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2902 call->state = RX_STATE_PRECALL;
2903 clock_GetTime(&call->queueTime);
2904 hzero(call->bytesSent);
2905 hzero(call->bytesRcvd);
2907 * If the number of queued calls exceeds the overload
2908 * threshold then abort this call.
2910 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2911 struct rx_packet *tp;
2913 rxi_CallError(call, rx_BusyError);
2914 tp = rxi_SendCallAbort(call, np, 1, 0);
2915 MUTEX_EXIT(&call->lock);
2916 MUTEX_ENTER(&conn->conn_data_lock);
2918 MUTEX_EXIT(&conn->conn_data_lock);
2919 if (rx_stats_active)
2920 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2923 rxi_KeepAliveOn(call);
2924 } else if (np->header.callNumber != currentCallNumber) {
2925 /* Wait until the transmit queue is idle before deciding
2926 * whether to reset the current call. Chances are that the
2927 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2930 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2931 while ((call->state == RX_STATE_ACTIVE)
2932 && (call->flags & RX_CALL_TQ_BUSY)) {
2933 call->flags |= RX_CALL_TQ_WAIT;
2935 #ifdef RX_ENABLE_LOCKS
2936 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2937 CV_WAIT(&call->cv_tq, &call->lock);
2938 #else /* RX_ENABLE_LOCKS */
2939 osi_rxSleep(&call->tq);
2940 #endif /* RX_ENABLE_LOCKS */
2942 if (call->tqWaiters == 0)
2943 call->flags &= ~RX_CALL_TQ_WAIT;
2945 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2946 /* If the new call cannot be taken right now send a busy and set
2947 * the error condition in this call, so that it terminates as
2948 * quickly as possible */
2949 if (call->state == RX_STATE_ACTIVE) {
2950 struct rx_packet *tp;
2952 rxi_CallError(call, RX_CALL_DEAD);
2953 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2955 MUTEX_EXIT(&call->lock);
2956 MUTEX_ENTER(&conn->conn_data_lock);
2958 MUTEX_EXIT(&conn->conn_data_lock);
2961 rxi_ResetCall(call, 0);
2962 *call->callNumber = np->header.callNumber;
2964 if (np->header.callNumber == 0)
2965 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%06d len %d",
2966 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
2967 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
2968 np->header.flags, np, np->retryTime.sec, np->retryTime.usec, np->length));
2970 call->state = RX_STATE_PRECALL;
2971 clock_GetTime(&call->queueTime);
2972 hzero(call->bytesSent);
2973 hzero(call->bytesRcvd);
2975 * If the number of queued calls exceeds the overload
2976 * threshold then abort this call.
2978 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2979 struct rx_packet *tp;
2981 rxi_CallError(call, rx_BusyError);
2982 tp = rxi_SendCallAbort(call, np, 1, 0);
2983 MUTEX_EXIT(&call->lock);
2984 MUTEX_ENTER(&conn->conn_data_lock);
2986 MUTEX_EXIT(&conn->conn_data_lock);
2987 if (rx_stats_active)
2988 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2991 rxi_KeepAliveOn(call);
2993 /* Continuing call; do nothing here. */
2995 } else { /* we're the client */
2996 /* Ignore all incoming acknowledgements for calls in DALLY state */
2997 if (call && (call->state == RX_STATE_DALLY)
2998 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2999 if (rx_stats_active)
3000 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
3001 #ifdef RX_ENABLE_LOCKS
3003 MUTEX_EXIT(&call->lock);
3006 MUTEX_ENTER(&conn->conn_data_lock);
3008 MUTEX_EXIT(&conn->conn_data_lock);
3012 /* Ignore anything that's not relevant to the current call. If there
3013 * isn't a current call, then no packet is relevant. */
3014 if (!call || (np->header.callNumber != currentCallNumber)) {
3015 if (rx_stats_active)
3016 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
3017 #ifdef RX_ENABLE_LOCKS
3019 MUTEX_EXIT(&call->lock);
3022 MUTEX_ENTER(&conn->conn_data_lock);
3024 MUTEX_EXIT(&conn->conn_data_lock);
3027 /* If the service security object index stamped in the packet does not
3028 * match the connection's security index, ignore the packet */
3029 if (np->header.securityIndex != conn->securityIndex) {
3030 #ifdef RX_ENABLE_LOCKS
3031 MUTEX_EXIT(&call->lock);
3033 MUTEX_ENTER(&conn->conn_data_lock);
3035 MUTEX_EXIT(&conn->conn_data_lock);
3039 /* If we're receiving the response, then all transmit packets are
3040 * implicitly acknowledged. Get rid of them. */
3041 if (np->header.type == RX_PACKET_TYPE_DATA) {
3042 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3043 /* XXX Hack. Because we must release the global rx lock when
3044 * sending packets (osi_NetSend) we drop all acks while we're
3045 * traversing the tq in rxi_Start sending packets out because
3046 * packets may move to the freePacketQueue as result of being here!
3047 * So we drop these packets until we're safely out of the
3048 * traversing. Really ugly!
3049 * For fine grain RX locking, we set the acked field in the
3050 * packets and let rxi_Start remove them from the transmit queue.
3052 if (call->flags & RX_CALL_TQ_BUSY) {
3053 #ifdef RX_ENABLE_LOCKS
3054 rxi_SetAcksInTransmitQueue(call);
3057 return np; /* xmitting; drop packet */
3060 rxi_ClearTransmitQueue(call, 0);
3062 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3063 rxi_ClearTransmitQueue(call, 0);
3064 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3066 if (np->header.type == RX_PACKET_TYPE_ACK) {
3067 /* now check to see if this is an ack packet acknowledging that the
3068 * server actually *lost* some hard-acked data. If this happens we
3069 * ignore this packet, as it may indicate that the server restarted in
3070 * the middle of a call. It is also possible that this is an old ack
3071 * packet. We don't abort the connection in this case, because this
3072 * *might* just be an old ack packet. The right way to detect a server
3073 * restart in the midst of a call is to notice that the server epoch
3075 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3076 * XXX unacknowledged. I think that this is off-by-one, but
3077 * XXX I don't dare change it just yet, since it will
3078 * XXX interact badly with the server-restart detection
3079 * XXX code in receiveackpacket. */
3080 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3081 if (rx_stats_active)
3082 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
3083 MUTEX_EXIT(&call->lock);
3084 MUTEX_ENTER(&conn->conn_data_lock);
3086 MUTEX_EXIT(&conn->conn_data_lock);
3090 } /* else not a data packet */
3093 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3094 /* Set remote user defined status from packet */
3095 call->remoteStatus = np->header.userStatus;
3097 /* Note the gap between the expected next packet and the actual
3098 * packet that arrived, when the new packet has a smaller serial number
3099 * than expected. Rioses frequently reorder packets all by themselves,
3100 * so this will be quite important with very large window sizes.
3101 * Skew is checked against 0 here to avoid any dependence on the type of
3102 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3104 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3105 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3106 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3108 MUTEX_ENTER(&conn->conn_data_lock);
3109 skew = conn->lastSerial - np->header.serial;
3110 conn->lastSerial = np->header.serial;
3111 MUTEX_EXIT(&conn->conn_data_lock);
3113 struct rx_peer *peer;
3115 if (skew > peer->inPacketSkew) {
3116 dpf(("*** In skew changed from %d to %d\n",
3117 peer->inPacketSkew, skew));
3118 peer->inPacketSkew = skew;
3122 /* Now do packet type-specific processing */
3123 switch (np->header.type) {
3124 case RX_PACKET_TYPE_DATA:
3125 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3128 case RX_PACKET_TYPE_ACK:
3129 /* Respond immediately to ack packets requesting acknowledgement
3131 if (np->header.flags & RX_REQUEST_ACK) {
3133 (void)rxi_SendCallAbort(call, 0, 1, 0);
3135 (void)rxi_SendAck(call, 0, np->header.serial,
3136 RX_ACK_PING_RESPONSE, 1);
3138 np = rxi_ReceiveAckPacket(call, np, 1);
3140 case RX_PACKET_TYPE_ABORT: {
3141 /* An abort packet: reset the call, passing the error up to the user. */
3142 /* What if error is zero? */
3143 /* What if the error is -1? the application will treat it as a timeout. */
3144 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3145 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
3146 rxi_CallError(call, errdata);
3147 MUTEX_EXIT(&call->lock);
3148 MUTEX_ENTER(&conn->conn_data_lock);
3150 MUTEX_EXIT(&conn->conn_data_lock);
3151 return np; /* xmitting; drop packet */
3153 case RX_PACKET_TYPE_BUSY:
3156 case RX_PACKET_TYPE_ACKALL:
3157 /* All packets acknowledged, so we can drop all packets previously
3158 * readied for sending */
3159 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3160 /* XXX Hack. We because we can't release the global rx lock when
3161 * sending packets (osi_NetSend) we drop all ack pkts while we're
3162 * traversing the tq in rxi_Start sending packets out because
3163 * packets may move to the freePacketQueue as result of being
3164 * here! So we drop these packets until we're safely out of the
3165 * traversing. Really ugly!
3166 * For fine grain RX locking, we set the acked field in the packets
3167 * and let rxi_Start remove the packets from the transmit queue.
3169 if (call->flags & RX_CALL_TQ_BUSY) {
3170 #ifdef RX_ENABLE_LOCKS
3171 rxi_SetAcksInTransmitQueue(call);
3173 #else /* RX_ENABLE_LOCKS */
3174 MUTEX_EXIT(&call->lock);
3175 MUTEX_ENTER(&conn->conn_data_lock);
3177 MUTEX_EXIT(&conn->conn_data_lock);
3178 return np; /* xmitting; drop packet */
3179 #endif /* RX_ENABLE_LOCKS */
3181 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3182 rxi_ClearTransmitQueue(call, 0);
3183 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3186 /* Should not reach here, unless the peer is broken: send an abort
3188 rxi_CallError(call, RX_PROTOCOL_ERROR);
3189 np = rxi_SendCallAbort(call, np, 1, 0);
3192 /* Note when this last legitimate packet was received, for keep-alive
3193 * processing. Note, we delay getting the time until now in the hope that
3194 * the packet will be delivered to the user before any get time is required
3195 * (if not, then the time won't actually be re-evaluated here). */
3196 call->lastReceiveTime = clock_Sec();
3197 MUTEX_EXIT(&call->lock);
3198 MUTEX_ENTER(&conn->conn_data_lock);
3200 MUTEX_EXIT(&conn->conn_data_lock);
3204 /* return true if this is an "interesting" connection from the point of view
3205 of someone trying to debug the system */
3207 rxi_IsConnInteresting(struct rx_connection *aconn)
3210 struct rx_call *tcall;
3212 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3215 for (i = 0; i < RX_MAXCALLS; i++) {
3216 tcall = aconn->call[i];
3218 if ((tcall->state == RX_STATE_PRECALL)
3219 || (tcall->state == RX_STATE_ACTIVE))
3221 if ((tcall->mode == RX_MODE_SENDING)
3222 || (tcall->mode == RX_MODE_RECEIVING))
3230 /* if this is one of the last few packets AND it wouldn't be used by the
3231 receiving call to immediately satisfy a read request, then drop it on
3232 the floor, since accepting it might prevent a lock-holding thread from
3233 making progress in its reading. If a call has been cleared while in
3234 the precall state then ignore all subsequent packets until the call
3235 is assigned to a thread. */
3238 TooLow(struct rx_packet *ap, struct rx_call *acall)
3242 MUTEX_ENTER(&rx_quota_mutex);
3243 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3244 && (acall->state == RX_STATE_PRECALL))
3245 || ((rx_nFreePackets < rxi_dataQuota + 2)
3246 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3247 && (acall->flags & RX_CALL_READER_WAIT)))) {
3250 MUTEX_EXIT(&rx_quota_mutex);
3256 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3258 struct rx_connection *conn = arg1;
3259 struct rx_call *acall = arg2;
3260 struct rx_call *call = acall;
3261 struct clock when, now;
3264 MUTEX_ENTER(&conn->conn_data_lock);
3265 conn->checkReachEvent = NULL;
3266 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3269 MUTEX_EXIT(&conn->conn_data_lock);
3273 MUTEX_ENTER(&conn->conn_call_lock);
3274 MUTEX_ENTER(&conn->conn_data_lock);
3275 for (i = 0; i < RX_MAXCALLS; i++) {
3276 struct rx_call *tc = conn->call[i];
3277 if (tc && tc->state == RX_STATE_PRECALL) {
3283 /* Indicate that rxi_CheckReachEvent is no longer running by
3284 * clearing the flag. Must be atomic under conn_data_lock to
3285 * avoid a new call slipping by: rxi_CheckConnReach holds
3286 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3288 conn->flags &= ~RX_CONN_ATTACHWAIT;
3289 MUTEX_EXIT(&conn->conn_data_lock);
3290 MUTEX_EXIT(&conn->conn_call_lock);
3295 MUTEX_ENTER(&call->lock);
3296 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3298 MUTEX_EXIT(&call->lock);
3300 clock_GetTime(&now);
3302 when.sec += RX_CHECKREACH_TIMEOUT;
3303 MUTEX_ENTER(&conn->conn_data_lock);
3304 if (!conn->checkReachEvent) {
3306 conn->checkReachEvent =
3307 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3310 MUTEX_EXIT(&conn->conn_data_lock);
3316 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3318 struct rx_service *service = conn->service;
3319 struct rx_peer *peer = conn->peer;
3320 afs_uint32 now, lastReach;
3322 if (service->checkReach == 0)
3326 MUTEX_ENTER(&peer->peer_lock);
3327 lastReach = peer->lastReachTime;
3328 MUTEX_EXIT(&peer->peer_lock);
3329 if (now - lastReach < RX_CHECKREACH_TTL)
3332 MUTEX_ENTER(&conn->conn_data_lock);
3333 if (conn->flags & RX_CONN_ATTACHWAIT) {
3334 MUTEX_EXIT(&conn->conn_data_lock);
3337 conn->flags |= RX_CONN_ATTACHWAIT;
3338 MUTEX_EXIT(&conn->conn_data_lock);
3339 if (!conn->checkReachEvent)
3340 rxi_CheckReachEvent(NULL, conn, call);
3345 /* try to attach call, if authentication is complete */
3347 TryAttach(struct rx_call *acall, osi_socket socket,
3348 int *tnop, struct rx_call **newcallp,
3351 struct rx_connection *conn = acall->conn;
3353 if (conn->type == RX_SERVER_CONNECTION
3354 && acall->state == RX_STATE_PRECALL) {
3355 /* Don't attach until we have any req'd. authentication. */
3356 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3357 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3358 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3359 /* Note: this does not necessarily succeed; there
3360 * may not any proc available
3363 rxi_ChallengeOn(acall->conn);
3368 /* A data packet has been received off the interface. This packet is
3369 * appropriate to the call (the call is in the right state, etc.). This
3370 * routine can return a packet to the caller, for re-use */
3373 rxi_ReceiveDataPacket(struct rx_call *call,
3374 struct rx_packet *np, int istack,
3375 osi_socket socket, afs_uint32 host, u_short port,
3376 int *tnop, struct rx_call **newcallp)
3378 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3383 afs_uint32 serial=0, flags=0;
3385 struct rx_packet *tnp;
3386 struct clock when, now;
3387 if (rx_stats_active)
3388 rx_MutexIncrement(rx_stats.dataPacketsRead, rx_stats_mutex);
3391 /* If there are no packet buffers, drop this new packet, unless we can find
3392 * packet buffers from inactive calls */
3394 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3395 MUTEX_ENTER(&rx_freePktQ_lock);
3396 rxi_NeedMorePackets = TRUE;
3397 MUTEX_EXIT(&rx_freePktQ_lock);
3398 if (rx_stats_active)
3399 rx_MutexIncrement(rx_stats.noPacketBuffersOnRead, rx_stats_mutex);
3400 call->rprev = np->header.serial;
3401 rxi_calltrace(RX_TRACE_DROP, call);
3402 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems", np));
3404 rxi_ClearReceiveQueue(call);
3405 clock_GetTime(&now);
3407 clock_Add(&when, &rx_softAckDelay);
3408 if (!call->delayedAckEvent
3409 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3410 rxevent_Cancel(call->delayedAckEvent, call,
3411 RX_CALL_REFCOUNT_DELAY);
3412 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3413 call->delayedAckEvent =
3414 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3416 /* we've damaged this call already, might as well do it in. */
3422 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3423 * packet is one of several packets transmitted as a single
3424 * datagram. Do not send any soft or hard acks until all packets
3425 * in a jumbogram have been processed. Send negative acks right away.
3427 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3428 /* tnp is non-null when there are more packets in the
3429 * current jumbo gram */
3436 seq = np->header.seq;
3437 serial = np->header.serial;
3438 flags = np->header.flags;
3440 /* If the call is in an error state, send an abort message */
3442 return rxi_SendCallAbort(call, np, istack, 0);
3444 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3445 * AFS 3.5 jumbogram. */
3446 if (flags & RX_JUMBO_PACKET) {
3447 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3452 if (np->header.spare != 0) {
3453 MUTEX_ENTER(&call->conn->conn_data_lock);
3454 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3455 MUTEX_EXIT(&call->conn->conn_data_lock);
3458 /* The usual case is that this is the expected next packet */
3459 if (seq == call->rnext) {
3461 /* Check to make sure it is not a duplicate of one already queued */
3462 if (queue_IsNotEmpty(&call->rq)
3463 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3464 if (rx_stats_active)
3465 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3466 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate", np));
3467 rxevent_Cancel(call->delayedAckEvent, call,
3468 RX_CALL_REFCOUNT_DELAY);
3469 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3475 /* It's the next packet. Stick it on the receive queue
3476 * for this call. Set newPackets to make sure we wake
3477 * the reader once all packets have been processed */
3478 np->flags |= RX_PKTFLAG_RQ;
3479 queue_Prepend(&call->rq, np);
3480 #ifdef RXDEBUG_PACKET
3482 #endif /* RXDEBUG_PACKET */
3484 np = NULL; /* We can't use this anymore */
3487 /* If an ack is requested then set a flag to make sure we
3488 * send an acknowledgement for this packet */
3489 if (flags & RX_REQUEST_ACK) {
3490 ackNeeded = RX_ACK_REQUESTED;
3493 /* Keep track of whether we have received the last packet */
3494 if (flags & RX_LAST_PACKET) {
3495 call->flags |= RX_CALL_HAVE_LAST;
3499 /* Check whether we have all of the packets for this call */
3500 if (call->flags & RX_CALL_HAVE_LAST) {
3501 afs_uint32 tseq; /* temporary sequence number */
3502 struct rx_packet *tp; /* Temporary packet pointer */
3503 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3505 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3506 if (tseq != tp->header.seq)
3508 if (tp->header.flags & RX_LAST_PACKET) {
3509 call->flags |= RX_CALL_RECEIVE_DONE;
3516 /* Provide asynchronous notification for those who want it
3517 * (e.g. multi rx) */
3518 if (call->arrivalProc) {
3519 (*call->arrivalProc) (call, call->arrivalProcHandle,
3520 call->arrivalProcArg);
3521 call->arrivalProc = (void (*)())0;
3524 /* Update last packet received */
3527 /* If there is no server process serving this call, grab
3528 * one, if available. We only need to do this once. If a
3529 * server thread is available, this thread becomes a server
3530 * thread and the server thread becomes a listener thread. */
3532 TryAttach(call, socket, tnop, newcallp, 0);
3535 /* This is not the expected next packet. */
3537 /* Determine whether this is a new or old packet, and if it's
3538 * a new one, whether it fits into the current receive window.
3539 * Also figure out whether the packet was delivered in sequence.
3540 * We use the prev variable to determine whether the new packet
3541 * is the successor of its immediate predecessor in the
3542 * receive queue, and the missing flag to determine whether
3543 * any of this packets predecessors are missing. */
3545 afs_uint32 prev; /* "Previous packet" sequence number */
3546 struct rx_packet *tp; /* Temporary packet pointer */
3547 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3548 int missing; /* Are any predecessors missing? */
3550 /* If the new packet's sequence number has been sent to the
3551 * application already, then this is a duplicate */
3552 if (seq < call->rnext) {
3553 if (rx_stats_active)
3554 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3555 rxevent_Cancel(call->delayedAckEvent, call,
3556 RX_CALL_REFCOUNT_DELAY);
3557 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3563 /* If the sequence number is greater than what can be
3564 * accomodated by the current window, then send a negative
3565 * acknowledge and drop the packet */
3566 if ((call->rnext + call->rwind) <= seq) {
3567 rxevent_Cancel(call->delayedAckEvent, call,
3568 RX_CALL_REFCOUNT_DELAY);
3569 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3576 /* Look for the packet in the queue of old received packets */
3577 for (prev = call->rnext - 1, missing =
3578 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3579 /*Check for duplicate packet */
3580 if (seq == tp->header.seq) {
3581 if (rx_stats_active)
3582 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3583 rxevent_Cancel(call->delayedAckEvent, call,
3584 RX_CALL_REFCOUNT_DELAY);
3585 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3591 /* If we find a higher sequence packet, break out and
3592 * insert the new packet here. */
3593 if (seq < tp->header.seq)
3595 /* Check for missing packet */
3596 if (tp->header.seq != prev + 1) {
3600 prev = tp->header.seq;
3603 /* Keep track of whether we have received the last packet. */
3604 if (flags & RX_LAST_PACKET) {
3605 call->flags |= RX_CALL_HAVE_LAST;
3608 /* It's within the window: add it to the the receive queue.
3609 * tp is left by the previous loop either pointing at the
3610 * packet before which to insert the new packet, or at the
3611 * queue head if the queue is empty or the packet should be
3613 np->flags |= RX_PKTFLAG_RQ;
3614 #ifdef RXDEBUG_PACKET
3616 #endif /* RXDEBUG_PACKET */
3617 queue_InsertBefore(tp, np);
3621 /* Check whether we have all of the packets for this call */
3622 if ((call->flags & RX_CALL_HAVE_LAST)
3623 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3624 afs_uint32 tseq; /* temporary sequence number */
3627 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3628 if (tseq != tp->header.seq)
3630 if (tp->header.flags & RX_LAST_PACKET) {
3631 call->flags |= RX_CALL_RECEIVE_DONE;
3638 /* We need to send an ack of the packet is out of sequence,
3639 * or if an ack was requested by the peer. */
3640 if (seq != prev + 1 || missing) {
3641 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3642 } else if (flags & RX_REQUEST_ACK) {
3643 ackNeeded = RX_ACK_REQUESTED;
3646 /* Acknowledge the last packet for each call */
3647 if (flags & RX_LAST_PACKET) {
3658 * If the receiver is waiting for an iovec, fill the iovec
3659 * using the data from the receive queue */
3660 if (call->flags & RX_CALL_IOVEC_WAIT) {
3661 didHardAck = rxi_FillReadVec(call, serial);
3662 /* the call may have been aborted */
3671 /* Wakeup the reader if any */
3672 if ((call->flags & RX_CALL_READER_WAIT)
3673 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3674 || (call->iovNext >= call->iovMax)
3675 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3676 call->flags &= ~RX_CALL_READER_WAIT;
3677 #ifdef RX_ENABLE_LOCKS
3678 CV_BROADCAST(&call->cv_rq);
3680 osi_rxWakeup(&call->rq);
3686 * Send an ack when requested by the peer, or once every
3687 * rxi_SoftAckRate packets until the last packet has been
3688 * received. Always send a soft ack for the last packet in
3689 * the server's reply. */
3691 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3692 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3693 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3694 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3695 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3696 } else if (call->nSoftAcks) {
3697 clock_GetTime(&now);
3699 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3700 clock_Add(&when, &rx_lastAckDelay);
3702 clock_Add(&when, &rx_softAckDelay);
3704 if (!call->delayedAckEvent
3705 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3706 rxevent_Cancel(call->delayedAckEvent, call,
3707 RX_CALL_REFCOUNT_DELAY);
3708 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3709 call->delayedAckEvent =
3710 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3712 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3713 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3720 static void rxi_ComputeRate();
3724 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3726 struct rx_peer *peer = conn->peer;
3728 MUTEX_ENTER(&peer->peer_lock);
3729 peer->lastReachTime = clock_Sec();
3730 MUTEX_EXIT(&peer->peer_lock);
3732 MUTEX_ENTER(&conn->conn_data_lock);
3733 if (conn->flags & RX_CONN_ATTACHWAIT) {
3736 conn->flags &= ~RX_CONN_ATTACHWAIT;
3737 MUTEX_EXIT(&conn->conn_data_lock);
3739 for (i = 0; i < RX_MAXCALLS; i++) {
3740 struct rx_call *call = conn->call[i];
3743 MUTEX_ENTER(&call->lock);
3744 /* tnop can be null if newcallp is null */
3745 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3747 MUTEX_EXIT(&call->lock);
3751 MUTEX_EXIT(&conn->conn_data_lock);
3754 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
3756 rx_ack_reason(int reason)
3759 case RX_ACK_REQUESTED:
3761 case RX_ACK_DUPLICATE:
3763 case RX_ACK_OUT_OF_SEQUENCE:
3765 case RX_ACK_EXCEEDS_WINDOW:
3767 case RX_ACK_NOSPACE:
3771 case RX_ACK_PING_RESPONSE:
3784 /* rxi_ComputePeerNetStats
3786 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3787 * estimates (like RTT and throughput) based on ack packets. Caller
3788 * must ensure that the packet in question is the right one (i.e.
3789 * serial number matches).
3792 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3793 struct rx_ackPacket *ap, struct rx_packet *np)
3795 struct rx_peer *peer = call->conn->peer;
3797 /* Use RTT if not delayed by client and
3798 * ignore packets that were retransmitted. */
3799 if (!(p->flags & RX_PKTFLAG_ACKED) &&
3800 ap->reason != RX_ACK_DELAY &&
3801 clock_Eq(&p->timeSent, &p->firstSent))
3802 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3804 rxi_ComputeRate(peer, call, p, np, ap->reason);
3808 /* The real smarts of the whole thing. */
3810 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
3813 struct rx_ackPacket *ap;
3815 struct rx_packet *tp;
3816 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3817 struct rx_connection *conn = call->conn;
3818 struct rx_peer *peer = conn->peer;
3821 /* because there are CM's that are bogus, sending weird values for this. */
3822 afs_uint32 skew = 0;
3828 int newAckCount = 0;
3829 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3830 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3831 int pktsize = 0; /* Set if we need to update the peer mtu */
3833 if (rx_stats_active)
3834 rx_MutexIncrement(rx_stats.ackPacketsRead, rx_stats_mutex);
3835 ap = (struct rx_ackPacket *)rx_DataOf(np);
3836 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3838 return np; /* truncated ack packet */
3840 /* depends on ack packet struct */
3841 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3842 first = ntohl(ap->firstPacket);
3843 serial = ntohl(ap->serial);
3844 /* temporarily disabled -- needs to degrade over time
3845 * skew = ntohs(ap->maxSkew); */
3847 /* Ignore ack packets received out of order */
3848 if (first < call->tfirst) {
3852 if (np->header.flags & RX_SLOW_START_OK) {
3853 call->flags |= RX_CALL_SLOW_START_OK;
3856 if (ap->reason == RX_ACK_PING_RESPONSE)
3857 rxi_UpdatePeerReach(conn, call);
3859 if (conn->lastPacketSizeSeq) {
3860 MUTEX_ENTER(&conn->conn_data_lock);
3861 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
3862 pktsize = conn->lastPacketSize;
3863 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
3865 MUTEX_EXIT(&conn->conn_data_lock);
3867 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
3868 MUTEX_ENTER(&conn->conn_data_lock);
3869 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
3870 /* process mtu ping ack */
3871 pktsize = conn->lastPingSize;
3872 conn->lastPingSizeSer = conn->lastPingSize = 0;
3874 MUTEX_EXIT(&conn->conn_data_lock);
3878 MUTEX_ENTER(&peer->peer_lock);
3880 * Start somewhere. Can't assume we can send what we can receive,
3881 * but we are clearly receiving.
3883 if (!peer->maxPacketSize)
3884 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
3886 if (pktsize > peer->maxPacketSize) {
3887 peer->maxPacketSize = pktsize;
3888 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
3889 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
3890 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
3891 rxi_ScheduleGrowMTUEvent(call, 1);
3894 MUTEX_EXIT(&peer->peer_lock);
3899 if (rxdebug_active) {
3903 len = _snprintf(msg, sizeof(msg),
3904 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3905 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3906 ntohl(ap->serial), ntohl(ap->previousPacket),
3907 (unsigned int)np->header.seq, (unsigned int)skew,
3908 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3912 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3913 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3917 OutputDebugString(msg);
3919 #else /* AFS_NT40_ENV */
3922 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3923 ap->reason, ntohl(ap->previousPacket),
3924 (unsigned int)np->header.seq, (unsigned int)serial,
3925 (unsigned int)skew, ntohl(ap->firstPacket));
3928 for (offset = 0; offset < nAcks; offset++)
3929 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3934 #endif /* AFS_NT40_ENV */
3937 /* Update the outgoing packet skew value to the latest value of
3938 * the peer's incoming packet skew value. The ack packet, of
3939 * course, could arrive out of order, but that won't affect things
3941 MUTEX_ENTER(&peer->peer_lock);
3942 peer->outPacketSkew = skew;
3944 /* Check for packets that no longer need to be transmitted, and
3945 * discard them. This only applies to packets positively
3946 * acknowledged as having been sent to the peer's upper level.
3947 * All other packets must be retained. So only packets with
3948 * sequence numbers < ap->firstPacket are candidates. */
3949 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3950 if (tp->header.seq >= first)
3952 call->tfirst = tp->header.seq + 1;
3953 rxi_ComputePeerNetStats(call, tp, ap, np);
3954 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3957 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3958 /* XXX Hack. Because we have to release the global rx lock when sending
3959 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3960 * in rxi_Start sending packets out because packets may move to the
3961 * freePacketQueue as result of being here! So we drop these packets until
3962 * we're safely out of the traversing. Really ugly!
3963 * To make it even uglier, if we're using fine grain locking, we can
3964 * set the ack bits in the packets and have rxi_Start remove the packets
3965 * when it's done transmitting.
3967 if (call->flags & RX_CALL_TQ_BUSY) {
3968 #ifdef RX_ENABLE_LOCKS
3969 tp->flags |= RX_PKTFLAG_ACKED;
3970 call->flags |= RX_CALL_TQ_SOME_ACKED;
3971 #else /* RX_ENABLE_LOCKS */
3973 #endif /* RX_ENABLE_LOCKS */
3975 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3978 tp->flags &= ~RX_PKTFLAG_TQ;
3979 #ifdef RXDEBUG_PACKET
3981 #endif /* RXDEBUG_PACKET */
3982 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3987 /* Give rate detector a chance to respond to ping requests */
3988 if (ap->reason == RX_ACK_PING_RESPONSE) {
3989 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3993 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3995 /* Now go through explicit acks/nacks and record the results in
3996 * the waiting packets. These are packets that can't be released
3997 * yet, even with a positive acknowledge. This positive
3998 * acknowledge only means the packet has been received by the
3999 * peer, not that it will be retained long enough to be sent to
4000 * the peer's upper level. In addition, reset the transmit timers
4001 * of any missing packets (those packets that must be missing
4002 * because this packet was out of sequence) */
4004 call->nSoftAcked = 0;
4005 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
4006 /* Update round trip time if the ack was stimulated on receipt
4008 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4009 #ifdef RX_ENABLE_LOCKS
4010 if (tp->header.seq >= first)
4011 #endif /* RX_ENABLE_LOCKS */
4012 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4013 rxi_ComputePeerNetStats(call, tp, ap, np);
4015 /* Set the acknowledge flag per packet based on the
4016 * information in the ack packet. An acknowlegded packet can
4017 * be downgraded when the server has discarded a packet it
4018 * soacked previously, or when an ack packet is received
4019 * out of sequence. */
4020 if (tp->header.seq < first) {
4021 /* Implicit ack information */
4022 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4025 tp->flags |= RX_PKTFLAG_ACKED;
4026 } else if (tp->header.seq < first + nAcks) {
4027 /* Explicit ack information: set it in the packet appropriately */
4028 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4029 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4031 tp->flags |= RX_PKTFLAG_ACKED;
4038 } else /* RX_ACK_TYPE_NACK */ {
4039 tp->flags &= ~RX_PKTFLAG_ACKED;
4043 tp->flags &= ~RX_PKTFLAG_ACKED;
4048 * Following the suggestion of Phil Kern, we back off the peer's
4049 * timeout value for future packets until a successful response
4050 * is received for an initial transmission.
4052 if (missing && !backedOff) {
4053 struct clock c = peer->timeout;
4054 struct clock max_to = {3, 0};
4056 clock_Add(&peer->timeout, &c);
4057 if (clock_Gt(&peer->timeout, &max_to))
4058 peer->timeout = max_to;
4062 /* If packet isn't yet acked, and it has been transmitted at least
4063 * once, reset retransmit time using latest timeout
4064 * ie, this should readjust the retransmit timer for all outstanding
4065 * packets... So we don't just retransmit when we should know better*/
4067 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
4068 tp->retryTime = tp->timeSent;
4069 clock_Add(&tp->retryTime, &peer->timeout);
4070 /* shift by eight because one quarter-sec ~ 256 milliseconds */
4071 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
4075 /* If the window has been extended by this acknowledge packet,
4076 * then wakeup a sender waiting in alloc for window space, or try
4077 * sending packets now, if he's been sitting on packets due to
4078 * lack of window space */
4079 if (call->tnext < (call->tfirst + call->twind)) {
4080 #ifdef RX_ENABLE_LOCKS
4081 CV_SIGNAL(&call->cv_twind);
4083 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4084 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4085 osi_rxWakeup(&call->twind);
4088 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4089 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4093 /* if the ack packet has a receivelen field hanging off it,
4094 * update our state */
4095 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4098 /* If the ack packet has a "recommended" size that is less than
4099 * what I am using now, reduce my size to match */
4100 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4101 (int)sizeof(afs_int32), &tSize);
4102 tSize = (afs_uint32) ntohl(tSize);
4103 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4105 /* Get the maximum packet size to send to this peer */
4106 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4108 tSize = (afs_uint32) ntohl(tSize);
4109 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4110 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4112 /* sanity check - peer might have restarted with different params.
4113 * If peer says "send less", dammit, send less... Peer should never
4114 * be unable to accept packets of the size that prior AFS versions would
4115 * send without asking. */
4116 if (peer->maxMTU != tSize) {
4117 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4119 peer->maxMTU = tSize;
4120 peer->MTU = MIN(tSize, peer->MTU);
4121 call->MTU = MIN(call->MTU, tSize);
4124 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4127 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4128 (int)sizeof(afs_int32), &tSize);
4129 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4130 if (tSize < call->twind) { /* smaller than our send */
4131 call->twind = tSize; /* window, we must send less... */
4132 call->ssthresh = MIN(call->twind, call->ssthresh);
4133 call->conn->twind[call->channel] = call->twind;
4136 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4137 * network MTU confused with the loopback MTU. Calculate the
4138 * maximum MTU here for use in the slow start code below.
4140 maxMTU = peer->maxMTU;
4141 /* Did peer restart with older RX version? */
4142 if (peer->maxDgramPackets > 1) {
4143 peer->maxDgramPackets = 1;
4145 } else if (np->length >=
4146 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4149 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4150 sizeof(afs_int32), &tSize);
4151 tSize = (afs_uint32) ntohl(tSize);
4153 * As of AFS 3.5 we set the send window to match the receive window.
4155 if (tSize < call->twind) {
4156 call->twind = tSize;
4157 call->conn->twind[call->channel] = call->twind;
4158 call->ssthresh = MIN(call->twind, call->ssthresh);
4159 } else if (tSize > call->twind) {
4160 call->twind = tSize;
4161 call->conn->twind[call->channel] = call->twind;
4165 * As of AFS 3.5, a jumbogram is more than one fixed size
4166 * packet transmitted in a single UDP datagram. If the remote
4167 * MTU is smaller than our local MTU then never send a datagram
4168 * larger than the natural MTU.
4171 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4172 (int)sizeof(afs_int32), &tSize);
4173 maxDgramPackets = (afs_uint32) ntohl(tSize);
4174 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4176 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4177 maxDgramPackets = MIN(maxDgramPackets, tSize);
4178 if (maxDgramPackets > 1) {
4179 peer->maxDgramPackets = maxDgramPackets;
4180 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4182 peer->maxDgramPackets = 1;
4183 call->MTU = peer->natMTU;
4185 } else if (peer->maxDgramPackets > 1) {
4186 /* Restarted with lower version of RX */
4187 peer->maxDgramPackets = 1;
4189 } else if (peer->maxDgramPackets > 1
4190 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4191 /* Restarted with lower version of RX */
4192 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4193 peer->natMTU = OLD_MAX_PACKET_SIZE;
4194 peer->MTU = OLD_MAX_PACKET_SIZE;
4195 peer->maxDgramPackets = 1;
4196 peer->nDgramPackets = 1;
4198 call->MTU = OLD_MAX_PACKET_SIZE;
4203 * Calculate how many datagrams were successfully received after
4204 * the first missing packet and adjust the negative ack counter
4209 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4210 if (call->nNacks < nNacked) {
4211 call->nNacks = nNacked;
4214 call->nAcks += newAckCount;
4218 if (call->flags & RX_CALL_FAST_RECOVER) {
4220 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4222 call->flags &= ~RX_CALL_FAST_RECOVER;
4223 call->cwind = call->nextCwind;
4224 call->nextCwind = 0;
4227 call->nCwindAcks = 0;
4228 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4229 /* Three negative acks in a row trigger congestion recovery */
4230 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4231 MUTEX_EXIT(&peer->peer_lock);
4232 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4233 /* someone else is waiting to start recovery */
4236 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4237 rxi_WaitforTQBusy(call);
4238 MUTEX_ENTER(&peer->peer_lock);
4239 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4240 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4241 call->flags |= RX_CALL_FAST_RECOVER;
4242 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4244 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4245 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4246 call->nextCwind = call->ssthresh;
4249 peer->MTU = call->MTU;
4250 peer->cwind = call->nextCwind;
4251 peer->nDgramPackets = call->nDgramPackets;
4253 call->congestSeq = peer->congestSeq;
4254 /* Reset the resend times on the packets that were nacked
4255 * so we will retransmit as soon as the window permits*/
4256 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4258 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4259 clock_Zero(&tp->retryTime);
4261 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4266 /* If cwind is smaller than ssthresh, then increase
4267 * the window one packet for each ack we receive (exponential
4269 * If cwind is greater than or equal to ssthresh then increase
4270 * the congestion window by one packet for each cwind acks we
4271 * receive (linear growth). */
4272 if (call->cwind < call->ssthresh) {
4274 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4275 call->nCwindAcks = 0;
4277 call->nCwindAcks += newAckCount;
4278 if (call->nCwindAcks >= call->cwind) {
4279 call->nCwindAcks = 0;
4280 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4284 * If we have received several acknowledgements in a row then
4285 * it is time to increase the size of our datagrams
4287 if ((int)call->nAcks > rx_nDgramThreshold) {
4288 if (peer->maxDgramPackets > 1) {
4289 if (call->nDgramPackets < peer->maxDgramPackets) {
4290 call->nDgramPackets++;
4292 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4293 } else if (call->MTU < peer->maxMTU) {
4294 /* don't upgrade if we can't handle it */
4295 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4296 call->MTU = peer->ifMTU;
4298 call->MTU += peer->natMTU;
4299 call->MTU = MIN(call->MTU, peer->maxMTU);
4306 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4308 /* Servers need to hold the call until all response packets have
4309 * been acknowledged. Soft acks are good enough since clients
4310 * are not allowed to clear their receive queues. */
4311 if (call->state == RX_STATE_HOLD
4312 && call->tfirst + call->nSoftAcked >= call->tnext) {
4313 call->state = RX_STATE_DALLY;
4314 rxi_ClearTransmitQueue(call, 0);
4315 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4316 } else if (!queue_IsEmpty(&call->tq)) {
4317 rxi_Start(0, call, 0, istack);
4322 /* Received a response to a challenge packet */
4324 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4325 struct rx_packet *np, int istack)
4329 /* Ignore the packet if we're the client */
4330 if (conn->type == RX_CLIENT_CONNECTION)
4333 /* If already authenticated, ignore the packet (it's probably a retry) */
4334 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4337 /* Otherwise, have the security object evaluate the response packet */
4338 error = RXS_CheckResponse(conn->securityObject, conn, np);
4340 /* If the response is invalid, reset the connection, sending
4341 * an abort to the peer */
4345 rxi_ConnectionError(conn, error);
4346 MUTEX_ENTER(&conn->conn_data_lock);
4347 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4348 MUTEX_EXIT(&conn->conn_data_lock);
4351 /* If the response is valid, any calls waiting to attach
4352 * servers can now do so */
4355 for (i = 0; i < RX_MAXCALLS; i++) {
4356 struct rx_call *call = conn->call[i];
4358 MUTEX_ENTER(&call->lock);
4359 if (call->state == RX_STATE_PRECALL)
4360 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4361 /* tnop can be null if newcallp is null */
4362 MUTEX_EXIT(&call->lock);
4366 /* Update the peer reachability information, just in case
4367 * some calls went into attach-wait while we were waiting
4368 * for authentication..
4370 rxi_UpdatePeerReach(conn, NULL);
4375 /* A client has received an authentication challenge: the security
4376 * object is asked to cough up a respectable response packet to send
4377 * back to the server. The server is responsible for retrying the
4378 * challenge if it fails to get a response. */
4381 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4382 struct rx_packet *np, int istack)
4386 /* Ignore the challenge if we're the server */
4387 if (conn->type == RX_SERVER_CONNECTION)
4390 /* Ignore the challenge if the connection is otherwise idle; someone's
4391 * trying to use us as an oracle. */
4392 if (!rxi_HasActiveCalls(conn))
4395 /* Send the security object the challenge packet. It is expected to fill
4396 * in the response. */
4397 error = RXS_GetResponse(conn->securityObject, conn, np);
4399 /* If the security object is unable to return a valid response, reset the
4400 * connection and send an abort to the peer. Otherwise send the response
4401 * packet to the peer connection. */
4403 rxi_ConnectionError(conn, error);
4404 MUTEX_ENTER(&conn->conn_data_lock);
4405 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4406 MUTEX_EXIT(&conn->conn_data_lock);
4408 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4409 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4415 /* Find an available server process to service the current request in
4416 * the given call structure. If one isn't available, queue up this
4417 * call so it eventually gets one */
4419 rxi_AttachServerProc(struct rx_call *call,
4420 osi_socket socket, int *tnop,
4421 struct rx_call **newcallp)
4423 struct rx_serverQueueEntry *sq;
4424 struct rx_service *service = call->conn->service;
4427 /* May already be attached */
4428 if (call->state == RX_STATE_ACTIVE)
4431 MUTEX_ENTER(&rx_serverPool_lock);
4433 haveQuota = QuotaOK(service);
4434 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4435 /* If there are no processes available to service this call,
4436 * put the call on the incoming call queue (unless it's
4437 * already on the queue).
4439 #ifdef RX_ENABLE_LOCKS
4441 ReturnToServerPool(service);
4442 #endif /* RX_ENABLE_LOCKS */
4444 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4445 call->flags |= RX_CALL_WAIT_PROC;
4446 MUTEX_ENTER(&rx_waiting_mutex);
4449 MUTEX_EXIT(&rx_waiting_mutex);
4450 rxi_calltrace(RX_CALL_ARRIVAL, call);
4451 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4452 queue_Append(&rx_incomingCallQueue, call);
4455 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4457 /* If hot threads are enabled, and both newcallp and sq->socketp
4458 * are non-null, then this thread will process the call, and the
4459 * idle server thread will start listening on this threads socket.
4462 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4465 *sq->socketp = socket;
4466 clock_GetTime(&call->startTime);
4467 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4471 if (call->flags & RX_CALL_WAIT_PROC) {
4472 /* Conservative: I don't think this should happen */
4473 call->flags &= ~RX_CALL_WAIT_PROC;
4474 if (queue_IsOnQueue(call)) {
4477 MUTEX_ENTER(&rx_waiting_mutex);
4479 MUTEX_EXIT(&rx_waiting_mutex);
4482 call->state = RX_STATE_ACTIVE;
4483 call->mode = RX_MODE_RECEIVING;
4484 #ifdef RX_KERNEL_TRACE
4486 int glockOwner = ISAFS_GLOCK();
4489 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4490 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4496 if (call->flags & RX_CALL_CLEARED) {
4497 /* send an ack now to start the packet flow up again */
4498 call->flags &= ~RX_CALL_CLEARED;
4499 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4501 #ifdef RX_ENABLE_LOCKS
4504 service->nRequestsRunning++;
4505 MUTEX_ENTER(&rx_quota_mutex);
4506 if (service->nRequestsRunning <= service->minProcs)
4509 MUTEX_EXIT(&rx_quota_mutex);
4513 MUTEX_EXIT(&rx_serverPool_lock);
4516 /* Delay the sending of an acknowledge event for a short while, while
4517 * a new call is being prepared (in the case of a client) or a reply
4518 * is being prepared (in the case of a server). Rather than sending
4519 * an ack packet, an ACKALL packet is sent. */
4521 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4523 #ifdef RX_ENABLE_LOCKS
4525 MUTEX_ENTER(&call->lock);
4526 call->delayedAckEvent = NULL;
4527 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4529 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4530 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4532 MUTEX_EXIT(&call->lock);
4533 #else /* RX_ENABLE_LOCKS */
4535 call->delayedAckEvent = NULL;
4536 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4537 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4538 #endif /* RX_ENABLE_LOCKS */
4542 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4544 struct rx_call *call = arg1;
4545 #ifdef RX_ENABLE_LOCKS
4547 MUTEX_ENTER(&call->lock);
4548 if (event == call->delayedAckEvent)
4549 call->delayedAckEvent = NULL;
4550 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4552 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4554 MUTEX_EXIT(&call->lock);
4555 #else /* RX_ENABLE_LOCKS */
4557 call->delayedAckEvent = NULL;
4558 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4559 #endif /* RX_ENABLE_LOCKS */
4563 #ifdef RX_ENABLE_LOCKS
4564 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4565 * clearing them out.
4568 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4570 struct rx_packet *p, *tp;
4573 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4574 p->flags |= RX_PKTFLAG_ACKED;
4578 call->flags |= RX_CALL_TQ_CLEARME;
4579 call->flags |= RX_CALL_TQ_SOME_ACKED;
4582 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4583 call->tfirst = call->tnext;
4584 call->nSoftAcked = 0;
4586 if (call->flags & RX_CALL_FAST_RECOVER) {
4587 call->flags &= ~RX_CALL_FAST_RECOVER;
4588 call->cwind = call->nextCwind;
4589 call->nextCwind = 0;
4592 CV_SIGNAL(&call->cv_twind);
4594 #endif /* RX_ENABLE_LOCKS */
4596 /* Clear out the transmit queue for the current call (all packets have
4597 * been received by peer) */
4599 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4601 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4602 struct rx_packet *p, *tp;
4604 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4606 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4607 p->flags |= RX_PKTFLAG_ACKED;
4611 call->flags |= RX_CALL_TQ_CLEARME;
4612 call->flags |= RX_CALL_TQ_SOME_ACKED;
4615 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4616 #ifdef RXDEBUG_PACKET
4618 #endif /* RXDEBUG_PACKET */
4619 rxi_FreePackets(0, &call->tq);
4620 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
4621 #ifdef RX_ENABLE_LOCKS
4622 CV_BROADCAST(&call->cv_tq);
4623 #else /* RX_ENABLE_LOCKS */
4624 osi_rxWakeup(&call->tq);
4625 #endif /* RX_ENABLE_LOCKS */
4627 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4628 call->flags &= ~RX_CALL_TQ_CLEARME;
4630 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4632 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4633 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4634 call->nSoftAcked = 0;
4636 if (call->flags & RX_CALL_FAST_RECOVER) {
4637 call->flags &= ~RX_CALL_FAST_RECOVER;
4638 call->cwind = call->nextCwind;
4640 #ifdef RX_ENABLE_LOCKS
4641 CV_SIGNAL(&call->cv_twind);
4643 osi_rxWakeup(&call->twind);
4648 rxi_ClearReceiveQueue(struct rx_call *call)
4650 if (queue_IsNotEmpty(&call->rq)) {
4653 count = rxi_FreePackets(0, &call->rq);
4654 rx_packetReclaims += count;
4655 #ifdef RXDEBUG_PACKET
4657 if ( call->rqc != 0 )
4658 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0", call, call->rqc));
4660 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4662 if (call->state == RX_STATE_PRECALL) {
4663 call->flags |= RX_CALL_CLEARED;
4667 /* Send an abort packet for the specified call */
4669 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
4670 int istack, int force)
4673 struct clock when, now;
4678 /* Clients should never delay abort messages */
4679 if (rx_IsClientConn(call->conn))
4682 if (call->abortCode != call->error) {
4683 call->abortCode = call->error;
4684 call->abortCount = 0;
4687 if (force || rxi_callAbortThreshhold == 0
4688 || call->abortCount < rxi_callAbortThreshhold) {
4689 if (call->delayedAbortEvent) {
4690 rxevent_Cancel(call->delayedAbortEvent, call,
4691 RX_CALL_REFCOUNT_ABORT);
4693 error = htonl(call->error);
4696 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4697 (char *)&error, sizeof(error), istack);
4698 } else if (!call->delayedAbortEvent) {
4699 clock_GetTime(&now);
4701 clock_Addmsec(&when, rxi_callAbortDelay);
4702 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4703 call->delayedAbortEvent =
4704 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4709 /* Send an abort packet for the specified connection. Packet is an
4710 * optional pointer to a packet that can be used to send the abort.
4711 * Once the number of abort messages reaches the threshhold, an
4712 * event is scheduled to send the abort. Setting the force flag
4713 * overrides sending delayed abort messages.
4715 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4716 * to send the abort packet.
4719 rxi_SendConnectionAbort(struct rx_connection *conn,
4720 struct rx_packet *packet, int istack, int force)
4723 struct clock when, now;
4728 /* Clients should never delay abort messages */
4729 if (rx_IsClientConn(conn))
4732 if (force || rxi_connAbortThreshhold == 0
4733 || conn->abortCount < rxi_connAbortThreshhold) {
4734 if (conn->delayedAbortEvent) {
4735 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4737 error = htonl(conn->error);
4739 MUTEX_EXIT(&conn->conn_data_lock);
4741 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4742 RX_PACKET_TYPE_ABORT, (char *)&error,
4743 sizeof(error), istack);
4744 MUTEX_ENTER(&conn->conn_data_lock);
4745 } else if (!conn->delayedAbortEvent) {
4746 clock_GetTime(&now);
4748 clock_Addmsec(&when, rxi_connAbortDelay);
4749 conn->delayedAbortEvent =
4750 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4755 /* Associate an error all of the calls owned by a connection. Called
4756 * with error non-zero. This is only for really fatal things, like
4757 * bad authentication responses. The connection itself is set in
4758 * error at this point, so that future packets received will be
4761 rxi_ConnectionError(struct rx_connection *conn,
4767 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d", conn, error));
4769 MUTEX_ENTER(&conn->conn_data_lock);
4770 if (conn->challengeEvent)
4771 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4772 if (conn->natKeepAliveEvent)
4773 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
4774 if (conn->checkReachEvent) {
4775 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4776 conn->checkReachEvent = 0;
4777 conn->flags &= ~RX_CONN_ATTACHWAIT;
4780 MUTEX_EXIT(&conn->conn_data_lock);
4781 for (i = 0; i < RX_MAXCALLS; i++) {
4782 struct rx_call *call = conn->call[i];
4784 MUTEX_ENTER(&call->lock);
4785 rxi_CallError(call, error);
4786 MUTEX_EXIT(&call->lock);
4789 conn->error = error;
4790 if (rx_stats_active)
4791 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4796 rxi_CallError(struct rx_call *call, afs_int32 error)
4799 osirx_AssertMine(&call->lock, "rxi_CallError");
4801 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d", call, error, call->error));
4803 error = call->error;
4805 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4806 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4807 rxi_ResetCall(call, 0);
4810 rxi_ResetCall(call, 0);
4812 call->error = error;
4813 call->mode = RX_MODE_ERROR;
4816 /* Reset various fields in a call structure, and wakeup waiting
4817 * processes. Some fields aren't changed: state & mode are not
4818 * touched (these must be set by the caller), and bufptr, nLeft, and
4819 * nFree are not reset, since these fields are manipulated by
4820 * unprotected macros, and may only be reset by non-interrupting code.
4823 /* this code requires that call->conn be set properly as a pre-condition. */
4824 #endif /* ADAPT_WINDOW */
4827 rxi_ResetCall(struct rx_call *call, int newcall)
4830 struct rx_peer *peer;
4831 struct rx_packet *packet;
4833 osirx_AssertMine(&call->lock, "rxi_ResetCall");
4835 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
4837 /* Notify anyone who is waiting for asynchronous packet arrival */
4838 if (call->arrivalProc) {
4839 (*call->arrivalProc) (call, call->arrivalProcHandle,
4840 call->arrivalProcArg);
4841 call->arrivalProc = (void (*)())0;
4844 if (call->delayedAbortEvent) {
4845 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4846 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4848 rxi_SendCallAbort(call, packet, 0, 1);
4849 rxi_FreePacket(packet);
4854 * Update the peer with the congestion information in this call
4855 * so other calls on this connection can pick up where this call
4856 * left off. If the congestion sequence numbers don't match then
4857 * another call experienced a retransmission.
4859 peer = call->conn->peer;
4860 MUTEX_ENTER(&peer->peer_lock);
4862 if (call->congestSeq == peer->congestSeq) {
4863 peer->cwind = MAX(peer->cwind, call->cwind);
4864 peer->MTU = MAX(peer->MTU, call->MTU);
4865 peer->nDgramPackets =
4866 MAX(peer->nDgramPackets, call->nDgramPackets);
4869 call->abortCode = 0;
4870 call->abortCount = 0;
4872 if (peer->maxDgramPackets > 1) {
4873 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4875 call->MTU = peer->MTU;
4877 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4878 call->ssthresh = rx_maxSendWindow;
4879 call->nDgramPackets = peer->nDgramPackets;
4880 call->congestSeq = peer->congestSeq;
4881 MUTEX_EXIT(&peer->peer_lock);
4883 flags = call->flags;
4884 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4885 rxi_WaitforTQBusy(call);
4886 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4888 rxi_ClearTransmitQueue(call, 1);
4889 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4890 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4894 rxi_ClearReceiveQueue(call);
4895 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
4897 if (call->currentPacket) {
4898 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
4899 call->currentPacket->flags |= RX_PKTFLAG_IOVQ;
4900 queue_Prepend(&call->iovq, call->currentPacket);
4901 #ifdef RXDEBUG_PACKET
4903 #endif /* RXDEBUG_PACKET */
4904 call->currentPacket = (struct rx_packet *)0;
4906 call->curlen = call->nLeft = call->nFree = 0;
4908 #ifdef RXDEBUG_PACKET
4911 rxi_FreePackets(0, &call->iovq);
4914 call->twind = call->conn->twind[call->channel];
4915 call->rwind = call->conn->rwind[call->channel];
4916 call->nSoftAcked = 0;
4917 call->nextCwind = 0;
4920 call->nCwindAcks = 0;
4921 call->nSoftAcks = 0;
4922 call->nHardAcks = 0;
4924 call->tfirst = call->rnext = call->tnext = 1;
4926 call->lastAcked = 0;
4927 call->localStatus = call->remoteStatus = 0;
4929 if (flags & RX_CALL_READER_WAIT) {
4930 #ifdef RX_ENABLE_LOCKS
4931 CV_BROADCAST(&call->cv_rq);
4933 osi_rxWakeup(&call->rq);
4936 if (flags & RX_CALL_WAIT_PACKETS) {
4937 MUTEX_ENTER(&rx_freePktQ_lock);
4938 rxi_PacketsUnWait(); /* XXX */
4939 MUTEX_EXIT(&rx_freePktQ_lock);
4941 #ifdef RX_ENABLE_LOCKS
4942 CV_SIGNAL(&call->cv_twind);
4944 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4945 osi_rxWakeup(&call->twind);
4948 #ifdef RX_ENABLE_LOCKS
4949 /* The following ensures that we don't mess with any queue while some
4950 * other thread might also be doing so. The call_queue_lock field is
4951 * is only modified under the call lock. If the call is in the process
4952 * of being removed from a queue, the call is not locked until the
4953 * the queue lock is dropped and only then is the call_queue_lock field
4954 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4955 * Note that any other routine which removes a call from a queue has to
4956 * obtain the queue lock before examing the queue and removing the call.
4958 if (call->call_queue_lock) {
4959 MUTEX_ENTER(call->call_queue_lock);
4960 if (queue_IsOnQueue(call)) {
4962 if (flags & RX_CALL_WAIT_PROC) {
4964 MUTEX_ENTER(&rx_waiting_mutex);
4966 MUTEX_EXIT(&rx_waiting_mutex);
4969 MUTEX_EXIT(call->call_queue_lock);
4970 CLEAR_CALL_QUEUE_LOCK(call);
4972 #else /* RX_ENABLE_LOCKS */
4973 if (queue_IsOnQueue(call)) {
4975 if (flags & RX_CALL_WAIT_PROC)
4978 #endif /* RX_ENABLE_LOCKS */
4980 rxi_KeepAliveOff(call);
4981 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4984 /* Send an acknowledge for the indicated packet (seq,serial) of the
4985 * indicated call, for the indicated reason (reason). This
4986 * acknowledge will specifically acknowledge receiving the packet, and
4987 * will also specify which other packets for this call have been
4988 * received. This routine returns the packet that was used to the
4989 * caller. The caller is responsible for freeing it or re-using it.
4990 * This acknowledgement also returns the highest sequence number
4991 * actually read out by the higher level to the sender; the sender
4992 * promises to keep around packets that have not been read by the
4993 * higher level yet (unless, of course, the sender decides to abort
4994 * the call altogether). Any of p, seq, serial, pflags, or reason may
4995 * be set to zero without ill effect. That is, if they are zero, they
4996 * will not convey any information.
4997 * NOW there is a trailer field, after the ack where it will safely be
4998 * ignored by mundanes, which indicates the maximum size packet this
4999 * host can swallow. */
5001 struct rx_packet *optionalPacket; use to send ack (or null)
5002 int seq; Sequence number of the packet we are acking
5003 int serial; Serial number of the packet
5004 int pflags; Flags field from packet header
5005 int reason; Reason an acknowledge was prompted
5009 rxi_SendAck(struct rx_call *call,
5010 struct rx_packet *optionalPacket, int serial, int reason,
5013 struct rx_ackPacket *ap;
5014 struct rx_packet *rqp;
5015 struct rx_packet *nxp; /* For queue_Scan */
5016 struct rx_packet *p;
5019 afs_uint32 padbytes = 0;
5020 #ifdef RX_ENABLE_TSFPQ
5021 struct rx_ts_info_t * rx_ts_info;
5025 * Open the receive window once a thread starts reading packets
5027 if (call->rnext > 1) {
5028 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5031 /* Don't attempt to grow MTU if this is a critical ping */
5032 if (reason == RX_ACK_MTU) {
5033 /* keep track of per-call attempts, if we're over max, do in small
5034 * otherwise in larger? set a size to increment by, decrease
5037 if (call->conn->peer->maxPacketSize &&
5038 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5040 padbytes = call->conn->peer->maxPacketSize+16;
5042 padbytes = call->conn->peer->maxMTU + 128;
5044 /* do always try a minimum size ping */
5045 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5047 /* subtract the ack payload */
5048 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5049 reason = RX_ACK_PING;
5052 call->nHardAcks = 0;
5053 call->nSoftAcks = 0;
5054 if (call->rnext > call->lastAcked)
5055 call->lastAcked = call->rnext;
5059 rx_computelen(p, p->length); /* reset length, you never know */
5060 } /* where that's been... */
5061 #ifdef RX_ENABLE_TSFPQ
5063 RX_TS_INFO_GET(rx_ts_info);
5064 if ((p = rx_ts_info->local_special_packet)) {
5065 rx_computelen(p, p->length);
5066 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5067 rx_ts_info->local_special_packet = p;
5068 } else { /* We won't send the ack, but don't panic. */
5069 return optionalPacket;
5073 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5074 /* We won't send the ack, but don't panic. */
5075 return optionalPacket;
5080 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5083 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5084 #ifndef RX_ENABLE_TSFPQ
5085 if (!optionalPacket)
5088 return optionalPacket;
5090 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5091 if (rx_Contiguous(p) < templ) {
5092 #ifndef RX_ENABLE_TSFPQ
5093 if (!optionalPacket)
5096 return optionalPacket;
5101 /* MTUXXX failing to send an ack is very serious. We should */
5102 /* try as hard as possible to send even a partial ack; it's */
5103 /* better than nothing. */
5104 ap = (struct rx_ackPacket *)rx_DataOf(p);
5105 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5106 ap->reason = reason;
5108 /* The skew computation used to be bogus, I think it's better now. */
5109 /* We should start paying attention to skew. XXX */
5110 ap->serial = htonl(serial);
5111 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5113 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
5114 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5116 /* No fear of running out of ack packet here because there can only be at most
5117 * one window full of unacknowledged packets. The window size must be constrained
5118 * to be less than the maximum ack size, of course. Also, an ack should always
5119 * fit into a single packet -- it should not ever be fragmented. */
5120 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5121 if (!rqp || !call->rq.next
5122 || (rqp->header.seq > (call->rnext + call->rwind))) {
5123 #ifndef RX_ENABLE_TSFPQ
5124 if (!optionalPacket)
5127 rxi_CallError(call, RX_CALL_DEAD);
5128 return optionalPacket;
5131 while (rqp->header.seq > call->rnext + offset)
5132 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5133 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5135 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5136 #ifndef RX_ENABLE_TSFPQ
5137 if (!optionalPacket)
5140 rxi_CallError(call, RX_CALL_DEAD);
5141 return optionalPacket;
5146 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5148 /* these are new for AFS 3.3 */
5149 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5150 templ = htonl(templ);
5151 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5152 templ = htonl(call->conn->peer->ifMTU);
5153 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5154 sizeof(afs_int32), &templ);
5156 /* new for AFS 3.4 */
5157 templ = htonl(call->rwind);
5158 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5159 sizeof(afs_int32), &templ);
5161 /* new for AFS 3.5 */
5162 templ = htonl(call->conn->peer->ifDgramPackets);
5163 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5164 sizeof(afs_int32), &templ);
5166 p->header.serviceId = call->conn->serviceId;
5167 p->header.cid = (call->conn->cid | call->channel);
5168 p->header.callNumber = *call->callNumber;
5170 p->header.securityIndex = call->conn->securityIndex;
5171 p->header.epoch = call->conn->epoch;
5172 p->header.type = RX_PACKET_TYPE_ACK;
5173 p->header.flags = RX_SLOW_START_OK;
5174 if (reason == RX_ACK_PING) {
5175 p->header.flags |= RX_REQUEST_ACK;
5177 clock_GetTime(&call->pingRequestTime);
5180 p->length = padbytes +
5181 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5184 /* not fast but we can potentially use this if truncated
5185 * fragments are delivered to figure out the mtu.
5187 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5188 sizeof(afs_int32), sizeof(afs_int32),
5192 if (call->conn->type == RX_CLIENT_CONNECTION)
5193 p->header.flags |= RX_CLIENT_INITIATED;
5197 if (rxdebug_active) {
5201 len = _snprintf(msg, sizeof(msg),
5202 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5203 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5204 ntohl(ap->serial), ntohl(ap->previousPacket),
5205 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5206 ap->nAcks, ntohs(ap->bufferSpace) );
5210 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5211 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5215 OutputDebugString(msg);
5217 #else /* AFS_NT40_ENV */
5219 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5220 ap->reason, ntohl(ap->previousPacket),
5221 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5223 for (offset = 0; offset < ap->nAcks; offset++)
5224 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5229 #endif /* AFS_NT40_ENV */
5232 int i, nbytes = p->length;
5234 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5235 if (nbytes <= p->wirevec[i].iov_len) {
5238 savelen = p->wirevec[i].iov_len;
5240 p->wirevec[i].iov_len = nbytes;
5242 rxi_Send(call, p, istack);
5243 p->wirevec[i].iov_len = savelen;
5247 nbytes -= p->wirevec[i].iov_len;
5250 if (rx_stats_active)
5251 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
5252 #ifndef RX_ENABLE_TSFPQ
5253 if (!optionalPacket)
5256 return optionalPacket; /* Return packet for re-use by caller */
5259 /* Send all of the packets in the list in single datagram */
5261 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
5262 int istack, int moreFlag, struct clock *now,
5263 struct clock *retryTime, int resending)
5268 struct rx_connection *conn = call->conn;
5269 struct rx_peer *peer = conn->peer;
5271 MUTEX_ENTER(&peer->peer_lock);
5274 peer->reSends += len;
5275 if (rx_stats_active)
5276 rx_MutexAdd(rx_stats.dataPacketsSent, len, rx_stats_mutex);
5277 MUTEX_EXIT(&peer->peer_lock);
5279 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
5283 /* Set the packet flags and schedule the resend events */
5284 /* Only request an ack for the last packet in the list */
5285 for (i = 0; i < len; i++) {
5286 list[i]->retryTime = *retryTime;
5287 if (list[i]->header.serial) {
5288 /* Exponentially backoff retry times */
5289 if (list[i]->backoff < MAXBACKOFF) {
5290 /* so it can't stay == 0 */
5291 list[i]->backoff = (list[i]->backoff << 1) + 1;
5294 clock_Addmsec(&(list[i]->retryTime),
5295 ((afs_uint32) list[i]->backoff) << 8);
5298 /* Wait a little extra for the ack on the last packet */
5299 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
5300 clock_Addmsec(&(list[i]->retryTime), 400);
5303 /* Record the time sent */
5304 list[i]->timeSent = *now;
5306 /* Ask for an ack on retransmitted packets, on every other packet
5307 * if the peer doesn't support slow start. Ask for an ack on every
5308 * packet until the congestion window reaches the ack rate. */
5309 if (list[i]->header.serial) {
5311 if (rx_stats_active)
5312 rx_MutexIncrement(rx_stats.dataPacketsReSent, rx_stats_mutex);
5314 /* improved RTO calculation- not Karn */
5315 list[i]->firstSent = *now;
5316 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5317 || (!(call->flags & RX_CALL_SLOW_START_OK)
5318 && (list[i]->header.seq & 1)))) {
5323 /* Tag this packet as not being the last in this group,
5324 * for the receiver's benefit */
5325 if (i < len - 1 || moreFlag) {
5326 list[i]->header.flags |= RX_MORE_PACKETS;
5329 /* Install the new retransmit time for the packet, and
5330 * record the time sent */
5331 list[i]->timeSent = *now;
5335 list[len - 1]->header.flags |= RX_REQUEST_ACK;
5338 /* Since we're about to send a data packet to the peer, it's
5339 * safe to nuke any scheduled end-of-packets ack */
5340 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5342 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5343 MUTEX_EXIT(&call->lock);
5345 rxi_SendPacketList(call, conn, list, len, istack);
5347 rxi_SendPacket(call, conn, list[0], istack);
5349 MUTEX_ENTER(&call->lock);
5350 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5352 /* Update last send time for this call (for keep-alive
5353 * processing), and for the connection (so that we can discover
5354 * idle connections) */
5355 conn->lastSendTime = call->lastSendTime = clock_Sec();
5356 /* Let a set of retransmits trigger an idle timeout */
5358 call->lastSendData = call->lastSendTime;
5361 /* When sending packets we need to follow these rules:
5362 * 1. Never send more than maxDgramPackets in a jumbogram.
5363 * 2. Never send a packet with more than two iovecs in a jumbogram.
5364 * 3. Never send a retransmitted packet in a jumbogram.
5365 * 4. Never send more than cwind/4 packets in a jumbogram
5366 * We always keep the last list we should have sent so we
5367 * can set the RX_MORE_PACKETS flags correctly.
5370 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5371 int istack, struct clock *now, struct clock *retryTime,
5374 int i, cnt, lastCnt = 0;
5375 struct rx_packet **listP, **lastP = 0;
5376 struct rx_peer *peer = call->conn->peer;
5377 int morePackets = 0;
5379 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
5380 /* Does the current packet force us to flush the current list? */
5382 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5383 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5385 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
5387 /* If the call enters an error state stop sending, or if
5388 * we entered congestion recovery mode, stop sending */
5389 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5397 /* Add the current packet to the list if it hasn't been acked.
5398 * Otherwise adjust the list pointer to skip the current packet. */
5399 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5401 /* Do we need to flush the list? */
5402 if (cnt >= (int)peer->maxDgramPackets
5403 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5404 || list[i]->header.serial
5405 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5407 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5408 retryTime, resending);
5409 /* If the call enters an error state stop sending, or if
5410 * we entered congestion recovery mode, stop sending */
5412 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5417 listP = &list[i + 1];
5422 osi_Panic("rxi_SendList error");
5424 listP = &list[i + 1];
5428 /* Send the whole list when the call is in receive mode, when
5429 * the call is in eof mode, when we are in fast recovery mode,
5430 * and when we have the last packet */
5431 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5432 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5433 || (call->flags & RX_CALL_FAST_RECOVER)) {
5434 /* Check for the case where the current list contains
5435 * an acked packet. Since we always send retransmissions
5436 * in a separate packet, we only need to check the first
5437 * packet in the list */
5438 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5442 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5443 retryTime, resending);
5444 /* If the call enters an error state stop sending, or if
5445 * we entered congestion recovery mode, stop sending */
5446 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5450 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5453 } else if (lastCnt > 0) {
5454 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5459 #ifdef RX_ENABLE_LOCKS
5460 /* Call rxi_Start, below, but with the call lock held. */
5462 rxi_StartUnlocked(struct rxevent *event,
5463 void *arg0, void *arg1, int istack)
5465 struct rx_call *call = arg0;
5467 MUTEX_ENTER(&call->lock);
5468 rxi_Start(event, call, arg1, istack);
5469 MUTEX_EXIT(&call->lock);
5471 #endif /* RX_ENABLE_LOCKS */
5473 /* This routine is called when new packets are readied for
5474 * transmission and when retransmission may be necessary, or when the
5475 * transmission window or burst count are favourable. This should be
5476 * better optimized for new packets, the usual case, now that we've
5477 * got rid of queues of send packets. XXXXXXXXXXX */
5479 rxi_Start(struct rxevent *event,
5480 void *arg0, void *arg1, int istack)
5482 struct rx_call *call = arg0;
5484 struct rx_packet *p;
5485 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5486 struct rx_peer *peer = call->conn->peer;
5487 struct clock now, usenow, retryTime;
5491 struct rx_packet **xmitList;
5494 /* If rxi_Start is being called as a result of a resend event,
5495 * then make sure that the event pointer is removed from the call
5496 * structure, since there is no longer a per-call retransmission
5498 if (event && event == call->resendEvent) {
5499 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5500 call->resendEvent = NULL;
5502 if (queue_IsEmpty(&call->tq)) {
5506 /* Timeouts trigger congestion recovery */
5507 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5508 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5509 /* someone else is waiting to start recovery */
5512 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5513 rxi_WaitforTQBusy(call);
5514 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5515 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5516 call->flags |= RX_CALL_FAST_RECOVER;
5517 if (peer->maxDgramPackets > 1) {
5518 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5520 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5522 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5523 call->nDgramPackets = 1;
5525 call->nextCwind = 1;
5528 MUTEX_ENTER(&peer->peer_lock);
5529 peer->MTU = call->MTU;
5530 peer->cwind = call->cwind;
5531 peer->nDgramPackets = 1;
5533 call->congestSeq = peer->congestSeq;
5534 MUTEX_EXIT(&peer->peer_lock);
5535 /* Clear retry times on packets. Otherwise, it's possible for
5536 * some packets in the queue to force resends at rates faster
5537 * than recovery rates.
5539 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5540 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5541 clock_Zero(&p->retryTime);
5546 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5547 if (rx_stats_active)
5548 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5553 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5554 /* Get clock to compute the re-transmit time for any packets
5555 * in this burst. Note, if we back off, it's reasonable to
5556 * back off all of the packets in the same manner, even if
5557 * some of them have been retransmitted more times than more
5559 * Do a dance to avoid blocking after setting now. */
5560 MUTEX_ENTER(&peer->peer_lock);
5561 retryTime = peer->timeout;
5562 MUTEX_EXIT(&peer->peer_lock);
5563 clock_GetTime(&now);
5564 clock_Add(&retryTime, &now);
5566 /* Send (or resend) any packets that need it, subject to
5567 * window restrictions and congestion burst control
5568 * restrictions. Ask for an ack on the last packet sent in
5569 * this burst. For now, we're relying upon the window being
5570 * considerably bigger than the largest number of packets that
5571 * are typically sent at once by one initial call to
5572 * rxi_Start. This is probably bogus (perhaps we should ask
5573 * for an ack when we're half way through the current
5574 * window?). Also, for non file transfer applications, this
5575 * may end up asking for an ack for every packet. Bogus. XXXX
5578 * But check whether we're here recursively, and let the other guy
5581 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5582 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5583 call->flags |= RX_CALL_TQ_BUSY;
5585 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5587 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5588 call->flags &= ~RX_CALL_NEED_START;
5589 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5591 maxXmitPackets = MIN(call->twind, call->cwind);
5592 xmitList = (struct rx_packet **)
5593 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
5594 /* XXXX else we must drop any mtx we hold */
5595 afs_osi_Alloc_NoSleep(maxXmitPackets * sizeof(struct rx_packet *));
5597 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5599 if (xmitList == NULL)
5600 osi_Panic("rxi_Start, failed to allocate xmit list");
5601 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5602 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5603 /* We shouldn't be sending packets if a thread is waiting
5604 * to initiate congestion recovery */
5605 dpf(("call %d waiting to initiate fast recovery\n",
5606 *(call->callNumber)));
5610 && (call->flags & RX_CALL_FAST_RECOVER)) {
5611 /* Only send one packet during fast recovery */
5612 dpf(("call %d restricted to one packet per send during fast recovery\n",
5613 *(call->callNumber)));
5616 if ((p->flags & RX_PKTFLAG_FREE)
5617 || (!queue_IsEnd(&call->tq, nxp)
5618 && (nxp->flags & RX_PKTFLAG_FREE))
5619 || (p == (struct rx_packet *)&rx_freePacketQueue)
5620 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5621 osi_Panic("rxi_Start: xmit queue clobbered");
5623 if (p->flags & RX_PKTFLAG_ACKED) {
5624 /* Since we may block, don't trust this */
5625 usenow.sec = usenow.usec = 0;
5626 if (rx_stats_active)
5627 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5628 continue; /* Ignore this packet if it has been acknowledged */
5631 /* Turn off all flags except these ones, which are the same
5632 * on each transmission */
5633 p->header.flags &= RX_PRESET_FLAGS;
5635 if (p->header.seq >=
5636 call->tfirst + MIN((int)call->twind,
5637 (int)(call->nSoftAcked +
5639 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5640 /* Note: if we're waiting for more window space, we can
5641 * still send retransmits; hence we don't return here, but
5642 * break out to schedule a retransmit event */
5643 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
5644 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
5649 /* Transmit the packet if it needs to be sent. */
5650 if (!clock_Lt(&now, &p->retryTime)) {
5651 if (nXmitPackets == maxXmitPackets) {
5652 rxi_SendXmitList(call, xmitList, nXmitPackets,
5653 istack, &now, &retryTime,
5655 osi_Free(xmitList, maxXmitPackets *
5656 sizeof(struct rx_packet *));
5659 dpf(("call %d xmit packet %"AFS_PTR_FMT" now %u.%06u retryTime %u.%06u nextRetry %u.%06u\n",
5660 *(call->callNumber), p,
5662 p->retryTime.sec, p->retryTime.usec,
5663 retryTime.sec, retryTime.usec));
5664 xmitList[nXmitPackets++] = p;
5668 /* xmitList now hold pointers to all of the packets that are
5669 * ready to send. Now we loop to send the packets */
5670 if (nXmitPackets > 0) {
5671 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5672 &now, &retryTime, resending);
5675 maxXmitPackets * sizeof(struct rx_packet *));
5677 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5679 * TQ references no longer protected by this flag; they must remain
5680 * protected by the global lock.
5682 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5683 call->flags &= ~RX_CALL_TQ_BUSY;
5684 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5685 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
5686 call, call->tqWaiters, call->flags));
5687 #ifdef RX_ENABLE_LOCKS
5688 osirx_AssertMine(&call->lock, "rxi_Start start");
5689 CV_BROADCAST(&call->cv_tq);
5690 #else /* RX_ENABLE_LOCKS */
5691 osi_rxWakeup(&call->tq);
5692 #endif /* RX_ENABLE_LOCKS */
5697 /* We went into the error state while sending packets. Now is
5698 * the time to reset the call. This will also inform the using
5699 * process that the call is in an error state.
5701 if (rx_stats_active)
5702 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5703 call->flags &= ~RX_CALL_TQ_BUSY;
5704 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5705 dpf(("call error %d while xmit %p has %d waiters and flags %d\n",
5706 call->error, call, call->tqWaiters, call->flags));
5707 #ifdef RX_ENABLE_LOCKS
5708 osirx_AssertMine(&call->lock, "rxi_Start middle");
5709 CV_BROADCAST(&call->cv_tq);
5710 #else /* RX_ENABLE_LOCKS */
5711 osi_rxWakeup(&call->tq);
5712 #endif /* RX_ENABLE_LOCKS */
5714 rxi_CallError(call, call->error);
5717 #ifdef RX_ENABLE_LOCKS
5718 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5720 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5721 /* Some packets have received acks. If they all have, we can clear
5722 * the transmit queue.
5725 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5726 if (p->header.seq < call->tfirst
5727 && (p->flags & RX_PKTFLAG_ACKED)) {
5729 p->flags &= ~RX_PKTFLAG_TQ;
5730 #ifdef RXDEBUG_PACKET
5738 call->flags |= RX_CALL_TQ_CLEARME;
5740 #endif /* RX_ENABLE_LOCKS */
5741 /* Don't bother doing retransmits if the TQ is cleared. */
5742 if (call->flags & RX_CALL_TQ_CLEARME) {
5743 rxi_ClearTransmitQueue(call, 1);
5745 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5748 /* Always post a resend event, if there is anything in the
5749 * queue, and resend is possible. There should be at least
5750 * one unacknowledged packet in the queue ... otherwise none
5751 * of these packets should be on the queue in the first place.
5753 if (call->resendEvent) {
5754 /* Cancel the existing event and post a new one */
5755 rxevent_Cancel(call->resendEvent, call,
5756 RX_CALL_REFCOUNT_RESEND);
5759 /* The retry time is the retry time on the first unacknowledged
5760 * packet inside the current window */
5762 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5763 /* Don't set timers for packets outside the window */
5764 if (p->header.seq >= call->tfirst + call->twind) {
5768 if (!(p->flags & RX_PKTFLAG_ACKED)
5769 && !clock_IsZero(&p->retryTime)) {
5771 retryTime = p->retryTime;
5776 /* Post a new event to re-run rxi_Start when retries may be needed */
5777 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5778 #ifdef RX_ENABLE_LOCKS
5779 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5781 rxevent_PostNow2(&retryTime, &usenow,
5783 (void *)call, 0, istack);
5784 #else /* RX_ENABLE_LOCKS */
5786 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5787 (void *)call, 0, istack);
5788 #endif /* RX_ENABLE_LOCKS */
5791 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5792 } while (call->flags & RX_CALL_NEED_START);
5794 * TQ references no longer protected by this flag; they must remain
5795 * protected by the global lock.
5797 call->flags &= ~RX_CALL_TQ_BUSY;
5798 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5799 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
5800 call, call->tqWaiters, call->flags));
5801 #ifdef RX_ENABLE_LOCKS
5802 osirx_AssertMine(&call->lock, "rxi_Start end");
5803 CV_BROADCAST(&call->cv_tq);
5804 #else /* RX_ENABLE_LOCKS */
5805 osi_rxWakeup(&call->tq);
5806 #endif /* RX_ENABLE_LOCKS */
5809 call->flags |= RX_CALL_NEED_START;
5811 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5813 if (call->resendEvent) {
5814 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5819 /* Also adjusts the keep alive parameters for the call, to reflect
5820 * that we have just sent a packet (so keep alives aren't sent
5823 rxi_Send(struct rx_call *call, struct rx_packet *p,
5826 struct rx_connection *conn = call->conn;
5828 /* Stamp each packet with the user supplied status */
5829 p->header.userStatus = call->localStatus;
5831 /* Allow the security object controlling this call's security to
5832 * make any last-minute changes to the packet */
5833 RXS_SendPacket(conn->securityObject, call, p);
5835 /* Since we're about to send SOME sort of packet to the peer, it's
5836 * safe to nuke any scheduled end-of-packets ack */
5837 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5839 /* Actually send the packet, filling in more connection-specific fields */
5840 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5841 MUTEX_EXIT(&call->lock);
5842 rxi_SendPacket(call, conn, p, istack);
5843 MUTEX_ENTER(&call->lock);
5844 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5846 /* Update last send time for this call (for keep-alive
5847 * processing), and for the connection (so that we can discover
5848 * idle connections) */
5849 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5850 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
5851 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
5853 conn->lastSendTime = call->lastSendTime = clock_Sec();
5854 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
5855 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5856 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
5857 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
5858 RX_ACK_PING_RESPONSE)))
5859 call->lastSendData = call->lastSendTime;
5863 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5864 * that things are fine. Also called periodically to guarantee that nothing
5865 * falls through the cracks (e.g. (error + dally) connections have keepalive
5866 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5868 * haveCTLock Set if calling from rxi_ReapConnections
5870 #ifdef RX_ENABLE_LOCKS
5872 rxi_CheckCall(struct rx_call *call, int haveCTLock)
5873 #else /* RX_ENABLE_LOCKS */
5875 rxi_CheckCall(struct rx_call *call)
5876 #endif /* RX_ENABLE_LOCKS */
5878 struct rx_connection *conn = call->conn;
5880 afs_uint32 deadTime;
5884 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5885 if (call->flags & RX_CALL_TQ_BUSY) {
5886 /* Call is active and will be reset by rxi_Start if it's
5887 * in an error state.
5892 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5894 (((afs_uint32) conn->secondsUntilDead << 10) +
5895 ((afs_uint32) conn->peer->rtt >> 3) +
5896 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5898 /* These are computed to the second (+- 1 second). But that's
5899 * good enough for these values, which should be a significant
5900 * number of seconds. */
5901 if (now > (call->lastReceiveTime + deadTime)) {
5902 if (call->state == RX_STATE_ACTIVE) {
5904 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5906 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5907 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5908 ip_stack_t *ipst = ns->netstack_ip;
5910 ire = ire_cache_lookup(conn->peer->host
5911 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5913 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5915 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5922 if (ire && ire->ire_max_frag > 0)
5923 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
5925 #if defined(GLOBAL_NETSTACKID)
5929 #endif /* ADAPT_PMTU */
5930 cerror = RX_CALL_DEAD;
5933 #ifdef RX_ENABLE_LOCKS
5934 /* Cancel pending events */
5935 rxevent_Cancel(call->delayedAckEvent, call,
5936 RX_CALL_REFCOUNT_DELAY);
5937 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5938 rxevent_Cancel(call->keepAliveEvent, call,
5939 RX_CALL_REFCOUNT_ALIVE);
5940 if (call->refCount == 0) {
5941 rxi_FreeCall(call, haveCTLock);
5945 #else /* RX_ENABLE_LOCKS */
5948 #endif /* RX_ENABLE_LOCKS */
5950 /* Non-active calls are destroyed if they are not responding
5951 * to pings; active calls are simply flagged in error, so the
5952 * attached process can die reasonably gracefully. */
5954 /* see if we have a non-activity timeout */
5955 if (call->startWait && conn->idleDeadTime
5956 && ((call->startWait + conn->idleDeadTime) < now) &&
5957 (call->flags & RX_CALL_READER_WAIT)) {
5958 if (call->state == RX_STATE_ACTIVE) {
5959 cerror = RX_CALL_TIMEOUT;
5963 if (call->lastSendData && conn->idleDeadTime && (conn->idleDeadErr != 0)
5964 && ((call->lastSendData + conn->idleDeadTime) < now)) {
5965 if (call->state == RX_STATE_ACTIVE) {
5966 cerror = conn->idleDeadErr;
5970 /* see if we have a hard timeout */
5971 if (conn->hardDeadTime
5972 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5973 if (call->state == RX_STATE_ACTIVE)
5974 rxi_CallError(call, RX_CALL_TIMEOUT);
5979 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT) {
5980 /* if we never succeeded, let the error pass out as-is */
5981 if (conn->peer->maxPacketSize)
5982 cerror = conn->msgsizeRetryErr;
5984 /* if we thought we could send more, perhaps things got worse */
5985 if (call->conn->peer->maxPacketSize > conn->lastPacketSize)
5986 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
5987 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
5988 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
5990 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
5992 /* minimum capped in SetPeerMtu */
5993 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
5996 conn->lastPacketSize = 0;
5998 /* needed so ResetCall doesn't clobber us. */
5999 call->MTU = conn->peer->ifMTU;
6001 rxi_CallError(call, cerror);
6006 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6008 struct rx_connection *conn = arg1;
6009 struct rx_header theader;
6011 struct sockaddr_in taddr;
6014 struct iovec tmpiov[2];
6017 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6020 tp = &tbuffer[sizeof(struct rx_header)];
6021 taddr.sin_family = AF_INET;
6022 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6023 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6024 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6025 taddr.sin_len = sizeof(struct sockaddr_in);
6027 memset(&theader, 0, sizeof(theader));
6028 theader.epoch = htonl(999);
6030 theader.callNumber = 0;
6033 theader.type = RX_PACKET_TYPE_VERSION;
6034 theader.flags = RX_LAST_PACKET;
6035 theader.serviceId = 0;
6037 memcpy(tbuffer, &theader, sizeof(theader));
6038 memcpy(tp, &a, sizeof(a));
6039 tmpiov[0].iov_base = tbuffer;
6040 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6042 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6044 MUTEX_ENTER(&conn->conn_data_lock);
6045 /* Only reschedule ourselves if the connection would not be destroyed */
6046 if (conn->refCount <= 1) {
6047 conn->natKeepAliveEvent = NULL;
6048 MUTEX_EXIT(&conn->conn_data_lock);
6049 rx_DestroyConnection(conn); /* drop the reference for this */
6051 conn->natKeepAliveEvent = NULL;
6052 conn->refCount--; /* drop the reference for this */
6053 rxi_ScheduleNatKeepAliveEvent(conn);
6054 MUTEX_EXIT(&conn->conn_data_lock);
6059 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6061 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6062 struct clock when, now;
6063 clock_GetTime(&now);
6065 when.sec += conn->secondsUntilNatPing;
6066 conn->refCount++; /* hold a reference for this */
6067 conn->natKeepAliveEvent =
6068 rxevent_PostNow(&when, &now, rxi_NatKeepAliveEvent, conn, 0);
6073 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6075 MUTEX_ENTER(&conn->conn_data_lock);
6076 conn->secondsUntilNatPing = seconds;
6078 rxi_ScheduleNatKeepAliveEvent(conn);
6079 MUTEX_EXIT(&conn->conn_data_lock);
6083 rxi_NatKeepAliveOn(struct rx_connection *conn)
6085 MUTEX_ENTER(&conn->conn_data_lock);
6086 rxi_ScheduleNatKeepAliveEvent(conn);
6087 MUTEX_EXIT(&conn->conn_data_lock);
6090 /* When a call is in progress, this routine is called occasionally to
6091 * make sure that some traffic has arrived (or been sent to) the peer.
6092 * If nothing has arrived in a reasonable amount of time, the call is
6093 * declared dead; if nothing has been sent for a while, we send a
6094 * keep-alive packet (if we're actually trying to keep the call alive)
6097 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6099 struct rx_call *call = arg1;
6100 struct rx_connection *conn;
6103 MUTEX_ENTER(&call->lock);
6104 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6105 if (event == call->keepAliveEvent)
6106 call->keepAliveEvent = NULL;
6109 #ifdef RX_ENABLE_LOCKS
6110 if (rxi_CheckCall(call, 0)) {
6111 MUTEX_EXIT(&call->lock);
6114 #else /* RX_ENABLE_LOCKS */
6115 if (rxi_CheckCall(call))
6117 #endif /* RX_ENABLE_LOCKS */
6119 /* Don't try to keep alive dallying calls */
6120 if (call->state == RX_STATE_DALLY) {
6121 MUTEX_EXIT(&call->lock);
6126 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6127 /* Don't try to send keepalives if there is unacknowledged data */
6128 /* the rexmit code should be good enough, this little hack
6129 * doesn't quite work XXX */
6130 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6132 rxi_ScheduleKeepAliveEvent(call);
6133 MUTEX_EXIT(&call->lock);
6136 /* Does what's on the nameplate. */
6138 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy)
6140 struct rx_call *call = arg1;
6141 struct rx_connection *conn;
6143 MUTEX_ENTER(&call->lock);
6144 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6145 if (event == call->growMTUEvent)
6146 call->growMTUEvent = NULL;
6148 #ifdef RX_ENABLE_LOCKS
6149 if (rxi_CheckCall(call, 0)) {
6150 MUTEX_EXIT(&call->lock);
6153 #else /* RX_ENABLE_LOCKS */
6154 if (rxi_CheckCall(call))
6156 #endif /* RX_ENABLE_LOCKS */
6158 /* Don't bother with dallying calls */
6159 if (call->state == RX_STATE_DALLY) {
6160 MUTEX_EXIT(&call->lock);
6167 * keep being scheduled, just don't do anything if we're at peak,
6168 * or we're not set up to be properly handled (idle timeout required)
6170 if ((conn->peer->maxPacketSize != 0) &&
6171 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6172 (conn->idleDeadErr))
6173 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6174 rxi_ScheduleGrowMTUEvent(call, 0);
6175 MUTEX_EXIT(&call->lock);
6179 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6181 if (!call->keepAliveEvent) {
6182 struct clock when, now;
6183 clock_GetTime(&now);
6185 when.sec += call->conn->secondsUntilPing;
6186 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6187 call->keepAliveEvent =
6188 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
6193 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6195 if (!call->growMTUEvent) {
6196 struct clock when, now;
6198 clock_GetTime(&now);
6201 if (call->conn->secondsUntilPing)
6202 secs = (6*call->conn->secondsUntilPing)-1;
6204 if (call->conn->secondsUntilDead)
6205 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6209 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6210 call->growMTUEvent =
6211 rxevent_PostNow(&when, &now, rxi_GrowMTUEvent, call, 0);
6215 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6217 rxi_KeepAliveOn(struct rx_call *call)
6219 /* Pretend last packet received was received now--i.e. if another
6220 * packet isn't received within the keep alive time, then the call
6221 * will die; Initialize last send time to the current time--even
6222 * if a packet hasn't been sent yet. This will guarantee that a
6223 * keep-alive is sent within the ping time */
6224 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6225 rxi_ScheduleKeepAliveEvent(call);
6229 rxi_GrowMTUOn(struct rx_call *call)
6231 struct rx_connection *conn = call->conn;
6232 MUTEX_ENTER(&conn->conn_data_lock);
6233 conn->lastPingSizeSer = conn->lastPingSize = 0;
6234 MUTEX_EXIT(&conn->conn_data_lock);
6235 rxi_ScheduleGrowMTUEvent(call, 1);
6238 /* This routine is called to send connection abort messages
6239 * that have been delayed to throttle looping clients. */
6241 rxi_SendDelayedConnAbort(struct rxevent *event,
6242 void *arg1, void *unused)
6244 struct rx_connection *conn = arg1;
6247 struct rx_packet *packet;
6249 MUTEX_ENTER(&conn->conn_data_lock);
6250 conn->delayedAbortEvent = NULL;
6251 error = htonl(conn->error);
6253 MUTEX_EXIT(&conn->conn_data_lock);
6254 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6257 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6258 RX_PACKET_TYPE_ABORT, (char *)&error,
6260 rxi_FreePacket(packet);
6264 /* This routine is called to send call abort messages
6265 * that have been delayed to throttle looping clients. */
6267 rxi_SendDelayedCallAbort(struct rxevent *event,
6268 void *arg1, void *dummy)
6270 struct rx_call *call = arg1;
6273 struct rx_packet *packet;
6275 MUTEX_ENTER(&call->lock);
6276 call->delayedAbortEvent = NULL;
6277 error = htonl(call->error);
6279 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6282 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6283 (char *)&error, sizeof(error), 0);
6284 rxi_FreePacket(packet);
6286 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6287 MUTEX_EXIT(&call->lock);
6290 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6291 * seconds) to ask the client to authenticate itself. The routine
6292 * issues a challenge to the client, which is obtained from the
6293 * security object associated with the connection */
6295 rxi_ChallengeEvent(struct rxevent *event,
6296 void *arg0, void *arg1, int tries)
6298 struct rx_connection *conn = arg0;
6300 conn->challengeEvent = NULL;
6301 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6302 struct rx_packet *packet;
6303 struct clock when, now;
6306 /* We've failed to authenticate for too long.
6307 * Reset any calls waiting for authentication;
6308 * they are all in RX_STATE_PRECALL.
6312 MUTEX_ENTER(&conn->conn_call_lock);
6313 for (i = 0; i < RX_MAXCALLS; i++) {
6314 struct rx_call *call = conn->call[i];
6316 MUTEX_ENTER(&call->lock);
6317 if (call->state == RX_STATE_PRECALL) {
6318 rxi_CallError(call, RX_CALL_DEAD);
6319 rxi_SendCallAbort(call, NULL, 0, 0);
6321 MUTEX_EXIT(&call->lock);
6324 MUTEX_EXIT(&conn->conn_call_lock);
6328 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6330 /* If there's no packet available, do this later. */
6331 RXS_GetChallenge(conn->securityObject, conn, packet);
6332 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6333 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6334 rxi_FreePacket(packet);
6336 clock_GetTime(&now);
6338 when.sec += RX_CHALLENGE_TIMEOUT;
6339 conn->challengeEvent =
6340 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
6345 /* Call this routine to start requesting the client to authenticate
6346 * itself. This will continue until authentication is established,
6347 * the call times out, or an invalid response is returned. The
6348 * security object associated with the connection is asked to create
6349 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6350 * defined earlier. */
6352 rxi_ChallengeOn(struct rx_connection *conn)
6354 if (!conn->challengeEvent) {
6355 RXS_CreateChallenge(conn->securityObject, conn);
6356 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6361 /* Compute round trip time of the packet provided, in *rttp.
6364 /* rxi_ComputeRoundTripTime is called with peer locked. */
6365 /* sentp and/or peer may be null */
6367 rxi_ComputeRoundTripTime(struct rx_packet *p,
6368 struct clock *sentp,
6369 struct rx_peer *peer)
6371 struct clock thisRtt, *rttp = &thisRtt;
6375 clock_GetTime(rttp);
6377 if (clock_Lt(rttp, sentp)) {
6379 return; /* somebody set the clock back, don't count this time. */
6381 clock_Sub(rttp, sentp);
6382 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6383 p->header.callNumber, p, rttp->sec, rttp->usec));
6385 if (rttp->sec == 0 && rttp->usec == 0) {
6387 * The actual round trip time is shorter than the
6388 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6389 * Since we can't tell which at the moment we will assume 1ms.
6394 if (rx_stats_active) {
6395 MUTEX_ENTER(&rx_stats_mutex);
6396 if (clock_Lt(rttp, &rx_stats.minRtt))
6397 rx_stats.minRtt = *rttp;
6398 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
6399 if (rttp->sec > 60) {
6400 MUTEX_EXIT(&rx_stats_mutex);
6401 return; /* somebody set the clock ahead */
6403 rx_stats.maxRtt = *rttp;
6405 clock_Add(&rx_stats.totalRtt, rttp);
6406 rx_stats.nRttSamples++;
6407 MUTEX_EXIT(&rx_stats_mutex);
6410 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6412 /* Apply VanJacobson round-trip estimations */
6417 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
6418 * srtt is stored as fixed point with 3 bits after the binary
6419 * point (i.e., scaled by 8). The following magic is
6420 * equivalent to the smoothing algorithm in rfc793 with an
6421 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6422 * srtt'*8 = rtt + srtt*7
6423 * srtt'*8 = srtt*8 + rtt - srtt
6424 * srtt' = srtt + rtt/8 - srtt/8
6425 * srtt' = srtt + (rtt - srtt)/8
6428 delta = _8THMSEC(rttp) - peer->rtt;
6429 peer->rtt += (delta >> 3);
6432 * We accumulate a smoothed rtt variance (actually, a smoothed
6433 * mean difference), then set the retransmit timer to smoothed
6434 * rtt + 4 times the smoothed variance (was 2x in van's original
6435 * paper, but 4x works better for me, and apparently for him as
6437 * rttvar is stored as
6438 * fixed point with 2 bits after the binary point (scaled by
6439 * 4). The following is equivalent to rfc793 smoothing with
6440 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6441 * rttvar'*4 = rttvar*3 + |delta|
6442 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6443 * rttvar' = rttvar + |delta|/4 - rttvar/4
6444 * rttvar' = rttvar + (|delta| - rttvar)/4
6445 * This replaces rfc793's wired-in beta.
6446 * dev*4 = dev*4 + (|actual - expected| - dev)
6452 delta -= (peer->rtt_dev << 1);
6453 peer->rtt_dev += (delta >> 3);
6455 /* I don't have a stored RTT so I start with this value. Since I'm
6456 * probably just starting a call, and will be pushing more data down
6457 * this, I expect congestion to increase rapidly. So I fudge a
6458 * little, and I set deviance to half the rtt. In practice,
6459 * deviance tends to approach something a little less than
6460 * half the smoothed rtt. */
6461 peer->rtt = _8THMSEC(rttp) + 8;
6462 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
6464 /* the timeout is RTT + 4*MDEV but no less than rx_minPeerTimeout msec.
6465 * This is because one end or the other of these connections is usually
6466 * in a user process, and can be switched and/or swapped out. So on fast,
6467 * reliable networks, the timeout would otherwise be too short. */
6468 rtt_timeout = MAX(((peer->rtt >> 3) + peer->rtt_dev), rx_minPeerTimeout);
6469 clock_Zero(&(peer->timeout));
6470 clock_Addmsec(&(peer->timeout), rtt_timeout);
6472 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6473 p->header.callNumber, p, MSEC(rttp), peer->rtt >> 3, peer->rtt_dev >> 2, (peer->timeout.sec), (peer->timeout.usec)));
6477 /* Find all server connections that have not been active for a long time, and
6480 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6482 struct clock now, when;
6483 clock_GetTime(&now);
6485 /* Find server connection structures that haven't been used for
6486 * greater than rx_idleConnectionTime */
6488 struct rx_connection **conn_ptr, **conn_end;
6489 int i, havecalls = 0;
6490 MUTEX_ENTER(&rx_connHashTable_lock);
6491 for (conn_ptr = &rx_connHashTable[0], conn_end =
6492 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6494 struct rx_connection *conn, *next;
6495 struct rx_call *call;
6499 for (conn = *conn_ptr; conn; conn = next) {
6500 /* XXX -- Shouldn't the connection be locked? */
6503 for (i = 0; i < RX_MAXCALLS; i++) {
6504 call = conn->call[i];
6508 code = MUTEX_TRYENTER(&call->lock);
6511 #ifdef RX_ENABLE_LOCKS
6512 result = rxi_CheckCall(call, 1);
6513 #else /* RX_ENABLE_LOCKS */
6514 result = rxi_CheckCall(call);
6515 #endif /* RX_ENABLE_LOCKS */
6516 MUTEX_EXIT(&call->lock);
6518 /* If CheckCall freed the call, it might
6519 * have destroyed the connection as well,
6520 * which screws up the linked lists.
6526 if (conn->type == RX_SERVER_CONNECTION) {
6527 /* This only actually destroys the connection if
6528 * there are no outstanding calls */
6529 MUTEX_ENTER(&conn->conn_data_lock);
6530 if (!havecalls && !conn->refCount
6531 && ((conn->lastSendTime + rx_idleConnectionTime) <
6533 conn->refCount++; /* it will be decr in rx_DestroyConn */
6534 MUTEX_EXIT(&conn->conn_data_lock);
6535 #ifdef RX_ENABLE_LOCKS
6536 rxi_DestroyConnectionNoLock(conn);
6537 #else /* RX_ENABLE_LOCKS */
6538 rxi_DestroyConnection(conn);
6539 #endif /* RX_ENABLE_LOCKS */
6541 #ifdef RX_ENABLE_LOCKS
6543 MUTEX_EXIT(&conn->conn_data_lock);
6545 #endif /* RX_ENABLE_LOCKS */
6549 #ifdef RX_ENABLE_LOCKS
6550 while (rx_connCleanup_list) {
6551 struct rx_connection *conn;
6552 conn = rx_connCleanup_list;
6553 rx_connCleanup_list = rx_connCleanup_list->next;
6554 MUTEX_EXIT(&rx_connHashTable_lock);
6555 rxi_CleanupConnection(conn);
6556 MUTEX_ENTER(&rx_connHashTable_lock);
6558 MUTEX_EXIT(&rx_connHashTable_lock);
6559 #endif /* RX_ENABLE_LOCKS */
6562 /* Find any peer structures that haven't been used (haven't had an
6563 * associated connection) for greater than rx_idlePeerTime */
6565 struct rx_peer **peer_ptr, **peer_end;
6569 * Why do we need to hold the rx_peerHashTable_lock across
6570 * the incrementing of peer_ptr since the rx_peerHashTable
6571 * array is not changing? We don't.
6573 * By dropping the lock periodically we can permit other
6574 * activities to be performed while a rxi_ReapConnections
6575 * call is in progress. The goal of reap connections
6576 * is to clean up quickly without causing large amounts
6577 * of contention. Therefore, it is important that global
6578 * mutexes not be held for extended periods of time.
6580 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6581 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6583 struct rx_peer *peer, *next, *prev;
6585 MUTEX_ENTER(&rx_peerHashTable_lock);
6586 for (prev = peer = *peer_ptr; peer; peer = next) {
6588 code = MUTEX_TRYENTER(&peer->peer_lock);
6589 if ((code) && (peer->refCount == 0)
6590 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6591 rx_interface_stat_p rpc_stat, nrpc_stat;
6595 * now know that this peer object is one to be
6596 * removed from the hash table. Once it is removed
6597 * it can't be referenced by other threads.
6598 * Lets remove it first and decrement the struct
6599 * nPeerStructs count.
6601 if (peer == *peer_ptr) {
6607 if (rx_stats_active)
6608 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6611 * Now if we hold references on 'prev' and 'next'
6612 * we can safely drop the rx_peerHashTable_lock
6613 * while we destroy this 'peer' object.
6619 MUTEX_EXIT(&rx_peerHashTable_lock);
6621 MUTEX_EXIT(&peer->peer_lock);
6622 MUTEX_DESTROY(&peer->peer_lock);
6624 (&peer->rpcStats, rpc_stat, nrpc_stat,
6625 rx_interface_stat)) {
6626 unsigned int num_funcs;
6629 queue_Remove(&rpc_stat->queue_header);
6630 queue_Remove(&rpc_stat->all_peers);
6631 num_funcs = rpc_stat->stats[0].func_total;
6633 sizeof(rx_interface_stat_t) +
6634 rpc_stat->stats[0].func_total *
6635 sizeof(rx_function_entry_v1_t);
6637 rxi_Free(rpc_stat, space);
6639 MUTEX_ENTER(&rx_rpc_stats);
6640 rxi_rpc_peer_stat_cnt -= num_funcs;
6641 MUTEX_EXIT(&rx_rpc_stats);
6646 * Regain the rx_peerHashTable_lock and
6647 * decrement the reference count on 'prev'
6650 MUTEX_ENTER(&rx_peerHashTable_lock);
6657 MUTEX_EXIT(&peer->peer_lock);
6662 MUTEX_EXIT(&rx_peerHashTable_lock);
6666 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6667 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6668 * GC, just below. Really, we shouldn't have to keep moving packets from
6669 * one place to another, but instead ought to always know if we can
6670 * afford to hold onto a packet in its particular use. */
6671 MUTEX_ENTER(&rx_freePktQ_lock);
6672 if (rx_waitingForPackets) {
6673 rx_waitingForPackets = 0;
6674 #ifdef RX_ENABLE_LOCKS
6675 CV_BROADCAST(&rx_waitingForPackets_cv);
6677 osi_rxWakeup(&rx_waitingForPackets);
6680 MUTEX_EXIT(&rx_freePktQ_lock);
6683 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6684 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6688 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6689 * rx.h is sort of strange this is better. This is called with a security
6690 * object before it is discarded. Each connection using a security object has
6691 * its own refcount to the object so it won't actually be freed until the last
6692 * connection is destroyed.
6694 * This is the only rxs module call. A hold could also be written but no one
6698 rxs_Release(struct rx_securityClass *aobj)
6700 return RXS_Close(aobj);
6704 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6705 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6706 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6707 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6709 /* Adjust our estimate of the transmission rate to this peer, given
6710 * that the packet p was just acked. We can adjust peer->timeout and
6711 * call->twind. Pragmatically, this is called
6712 * only with packets of maximal length.
6713 * Called with peer and call locked.
6717 rxi_ComputeRate(struct rx_peer *peer, struct rx_call *call,
6718 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6720 afs_int32 xferSize, xferMs;
6724 /* Count down packets */
6725 if (peer->rateFlag > 0)
6727 /* Do nothing until we're enabled */
6728 if (peer->rateFlag != 0)
6733 /* Count only when the ack seems legitimate */
6734 switch (ackReason) {
6735 case RX_ACK_REQUESTED:
6737 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6741 case RX_ACK_PING_RESPONSE:
6742 if (p) /* want the response to ping-request, not data send */
6744 clock_GetTime(&newTO);
6745 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6746 clock_Sub(&newTO, &call->pingRequestTime);
6747 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6751 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6758 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %d.%06d, rtt %u, ps %u)",
6759 ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"),
6760 xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
6762 /* Track only packets that are big enough. */
6763 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6767 /* absorb RTT data (in milliseconds) for these big packets */
6768 if (peer->smRtt == 0) {
6769 peer->smRtt = xferMs;
6771 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6776 if (peer->countDown) {
6780 peer->countDown = 10; /* recalculate only every so often */
6782 /* In practice, we can measure only the RTT for full packets,
6783 * because of the way Rx acks the data that it receives. (If it's
6784 * smaller than a full packet, it often gets implicitly acked
6785 * either by the call response (from a server) or by the next call
6786 * (from a client), and either case confuses transmission times
6787 * with processing times.) Therefore, replace the above
6788 * more-sophisticated processing with a simpler version, where the
6789 * smoothed RTT is kept for full-size packets, and the time to
6790 * transmit a windowful of full-size packets is simply RTT *
6791 * windowSize. Again, we take two steps:
6792 - ensure the timeout is large enough for a single packet's RTT;
6793 - ensure that the window is small enough to fit in the desired timeout.*/
6795 /* First, the timeout check. */
6796 minTime = peer->smRtt;
6797 /* Get a reasonable estimate for a timeout period */
6799 newTO.sec = minTime / 1000;
6800 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6802 /* Increase the timeout period so that we can always do at least
6803 * one packet exchange */
6804 if (clock_Gt(&newTO, &peer->timeout)) {
6806 dpf(("CONG peer %lx/%u: timeout %d.%06d ==> %ld.%06d (rtt %u, ps %u)",
6807 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec,
6808 newTO.sec, newTO.usec, peer->smRtt, peer->packetSize));
6810 peer->timeout = newTO;
6813 /* Now, get an estimate for the transmit window size. */
6814 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6815 /* Now, convert to the number of full packets that could fit in a
6816 * reasonable fraction of that interval */
6817 minTime /= (peer->smRtt << 1);
6818 xferSize = minTime; /* (make a copy) */
6820 /* Now clamp the size to reasonable bounds. */
6823 else if (minTime > rx_Window)
6824 minTime = rx_Window;
6825 /* if (minTime != peer->maxWindow) {
6826 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6827 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6828 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6830 peer->maxWindow = minTime;
6831 elide... call->twind = minTime;
6835 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6836 * Discern this by calculating the timeout necessary for rx_Window
6838 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6839 /* calculate estimate for transmission interval in milliseconds */
6840 minTime = rx_Window * peer->smRtt;
6841 if (minTime < 1000) {
6842 dpf(("CONG peer %lx/%u: cut TO %d.%06d by 0.5 (rtt %u, ps %u)",
6843 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6844 peer->timeout.usec, peer->smRtt, peer->packetSize));
6846 newTO.sec = 0; /* cut back on timeout by half a second */
6847 newTO.usec = 500000;
6848 clock_Sub(&peer->timeout, &newTO);
6853 } /* end of rxi_ComputeRate */
6854 #endif /* ADAPT_WINDOW */
6862 #define TRACE_OPTION_RX_DEBUG 16
6870 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6871 0, KEY_QUERY_VALUE, &parmKey);
6872 if (code != ERROR_SUCCESS)
6875 dummyLen = sizeof(TraceOption);
6876 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6877 (BYTE *) &TraceOption, &dummyLen);
6878 if (code == ERROR_SUCCESS) {
6879 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
6881 RegCloseKey (parmKey);
6882 #endif /* AFS_NT40_ENV */
6887 rx_DebugOnOff(int on)
6891 rxdebug_active = on;
6897 rx_StatsOnOff(int on)
6900 rx_stats_active = on;
6905 /* Don't call this debugging routine directly; use dpf */
6907 rxi_DebugPrint(char *format, ...)
6916 va_start(ap, format);
6918 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6921 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
6923 if (msg[len-1] != '\n') {
6927 OutputDebugString(msg);
6934 va_start(ap, format);
6936 clock_GetTime(&now);
6937 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
6938 (unsigned int)now.usec);
6939 vfprintf(rx_Log, format, ap);
6948 * This function is used to process the rx_stats structure that is local
6949 * to a process as well as an rx_stats structure received from a remote
6950 * process (via rxdebug). Therefore, it needs to do minimal version
6954 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
6955 afs_int32 freePackets, char version)
6960 if (size != sizeof(struct rx_statistics)) {
6962 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
6963 size, sizeof(struct rx_statistics));
6966 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6969 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6970 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
6971 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6972 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6973 s->specialPktAllocFailures);
6975 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
6976 s->receivePktAllocFailures, s->sendPktAllocFailures,
6977 s->specialPktAllocFailures);
6981 " greedy %u, " "bogusReads %u (last from host %x), "
6982 "noPackets %u, " "noBuffers %u, " "selects %u, "
6983 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
6984 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6985 s->selects, s->sendSelects);
6987 fprintf(file, " packets read: ");
6988 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6989 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
6991 fprintf(file, "\n");
6994 " other read counters: data %u, " "ack %u, " "dup %u "
6995 "spurious %u " "dally %u\n", s->dataPacketsRead,
6996 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6997 s->ignorePacketDally);
6999 fprintf(file, " packets sent: ");
7000 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7001 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7003 fprintf(file, "\n");
7006 " other send counters: ack %u, " "data %u (not resends), "
7007 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7008 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7009 s->dataPacketsPushed, s->ignoreAckedPacket);
7012 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7013 s->netSendFailures, (int)s->fatalErrors);
7015 if (s->nRttSamples) {
7016 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7017 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7019 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7020 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7024 " %d server connections, " "%d client connections, "
7025 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7026 s->nServerConns, s->nClientConns, s->nPeerStructs,
7027 s->nCallStructs, s->nFreeCallStructs);
7029 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7030 fprintf(file, " %d clock updates\n", clock_nUpdates);
7033 fprintf(file, "ERROR: compiled without RXDEBUG\n");
7037 /* for backward compatibility */
7039 rx_PrintStats(FILE * file)
7041 MUTEX_ENTER(&rx_stats_mutex);
7042 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
7044 MUTEX_EXIT(&rx_stats_mutex);
7048 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7050 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
7051 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
7052 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
7055 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
7056 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
7057 (int)peer->timeout.usec, peer->nSent, peer->reSends);
7060 " Packet size %d, " "max in packet skew %d, "
7061 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
7062 (int)peer->outPacketSkew);
7066 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7068 * This mutex protects the following static variables:
7072 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7073 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7075 #define LOCK_RX_DEBUG
7076 #define UNLOCK_RX_DEBUG
7077 #endif /* AFS_PTHREAD_ENV */
7081 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7082 u_char type, void *inputData, size_t inputLength,
7083 void *outputData, size_t outputLength)
7085 static afs_int32 counter = 100;
7086 time_t waitTime, waitCount, startTime;
7087 struct rx_header theader;
7090 struct timeval tv_now, tv_wake, tv_delta;
7091 struct sockaddr_in taddr, faddr;
7100 startTime = time(0);
7106 tp = &tbuffer[sizeof(struct rx_header)];
7107 taddr.sin_family = AF_INET;
7108 taddr.sin_port = remotePort;
7109 taddr.sin_addr.s_addr = remoteAddr;
7110 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7111 taddr.sin_len = sizeof(struct sockaddr_in);
7114 memset(&theader, 0, sizeof(theader));
7115 theader.epoch = htonl(999);
7117 theader.callNumber = htonl(counter);
7120 theader.type = type;
7121 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7122 theader.serviceId = 0;
7124 memcpy(tbuffer, &theader, sizeof(theader));
7125 memcpy(tp, inputData, inputLength);
7127 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7128 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7130 /* see if there's a packet available */
7131 gettimeofday(&tv_wake,0);
7132 tv_wake.tv_sec += waitTime;
7135 FD_SET(socket, &imask);
7136 tv_delta.tv_sec = tv_wake.tv_sec;
7137 tv_delta.tv_usec = tv_wake.tv_usec;
7138 gettimeofday(&tv_now, 0);
7140 if (tv_delta.tv_usec < tv_now.tv_usec) {
7142 tv_delta.tv_usec += 1000000;
7145 tv_delta.tv_usec -= tv_now.tv_usec;
7147 if (tv_delta.tv_sec < tv_now.tv_sec) {
7151 tv_delta.tv_sec -= tv_now.tv_sec;
7154 code = select(0, &imask, 0, 0, &tv_delta);
7155 #else /* AFS_NT40_ENV */
7156 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7157 #endif /* AFS_NT40_ENV */
7158 if (code == 1 && FD_ISSET(socket, &imask)) {
7159 /* now receive a packet */
7160 faddrLen = sizeof(struct sockaddr_in);
7162 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7163 (struct sockaddr *)&faddr, &faddrLen);
7166 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7167 if (counter == ntohl(theader.callNumber))
7175 /* see if we've timed out */
7183 code -= sizeof(struct rx_header);
7184 if (code > outputLength)
7185 code = outputLength;
7186 memcpy(outputData, tp, code);
7189 #endif /* RXDEBUG */
7192 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7193 afs_uint16 remotePort, struct rx_debugStats * stat,
7194 afs_uint32 * supportedValues)
7200 struct rx_debugIn in;
7202 *supportedValues = 0;
7203 in.type = htonl(RX_DEBUGI_GETSTATS);
7206 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7207 &in, sizeof(in), stat, sizeof(*stat));
7210 * If the call was successful, fixup the version and indicate
7211 * what contents of the stat structure are valid.
7212 * Also do net to host conversion of fields here.
7216 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7217 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7219 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7220 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7222 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7223 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7225 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7226 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7228 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7229 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7231 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7232 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7234 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7235 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7237 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7238 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7240 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7241 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7243 stat->nFreePackets = ntohl(stat->nFreePackets);
7244 stat->packetReclaims = ntohl(stat->packetReclaims);
7245 stat->callsExecuted = ntohl(stat->callsExecuted);
7246 stat->nWaiting = ntohl(stat->nWaiting);
7247 stat->idleThreads = ntohl(stat->idleThreads);
7248 stat->nWaited = ntohl(stat->nWaited);
7249 stat->nPackets = ntohl(stat->nPackets);
7256 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7257 afs_uint16 remotePort, struct rx_statistics * stat,
7258 afs_uint32 * supportedValues)
7264 struct rx_debugIn in;
7265 afs_int32 *lp = (afs_int32 *) stat;
7269 * supportedValues is currently unused, but added to allow future
7270 * versioning of this function.
7273 *supportedValues = 0;
7274 in.type = htonl(RX_DEBUGI_RXSTATS);
7276 memset(stat, 0, sizeof(*stat));
7278 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7279 &in, sizeof(in), stat, sizeof(*stat));
7284 * Do net to host conversion here
7287 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7296 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7297 afs_uint16 remotePort, size_t version_length,
7302 return MakeDebugCall(socket, remoteAddr, remotePort,
7303 RX_PACKET_TYPE_VERSION, a, 1, version,
7311 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7312 afs_uint16 remotePort, afs_int32 * nextConnection,
7313 int allConnections, afs_uint32 debugSupportedValues,
7314 struct rx_debugConn * conn,
7315 afs_uint32 * supportedValues)
7321 struct rx_debugIn in;
7325 * supportedValues is currently unused, but added to allow future
7326 * versioning of this function.
7329 *supportedValues = 0;
7330 if (allConnections) {
7331 in.type = htonl(RX_DEBUGI_GETALLCONN);
7333 in.type = htonl(RX_DEBUGI_GETCONN);
7335 in.index = htonl(*nextConnection);
7336 memset(conn, 0, sizeof(*conn));
7338 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7339 &in, sizeof(in), conn, sizeof(*conn));
7342 *nextConnection += 1;
7345 * Convert old connection format to new structure.
7348 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7349 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7350 #define MOVEvL(a) (conn->a = vL->a)
7352 /* any old or unrecognized version... */
7353 for (i = 0; i < RX_MAXCALLS; i++) {
7354 MOVEvL(callState[i]);
7355 MOVEvL(callMode[i]);
7356 MOVEvL(callFlags[i]);
7357 MOVEvL(callOther[i]);
7359 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7360 MOVEvL(secStats.type);
7361 MOVEvL(secStats.level);
7362 MOVEvL(secStats.flags);
7363 MOVEvL(secStats.expires);
7364 MOVEvL(secStats.packetsReceived);
7365 MOVEvL(secStats.packetsSent);
7366 MOVEvL(secStats.bytesReceived);
7367 MOVEvL(secStats.bytesSent);
7372 * Do net to host conversion here
7374 * I don't convert host or port since we are most likely
7375 * going to want these in NBO.
7377 conn->cid = ntohl(conn->cid);
7378 conn->serial = ntohl(conn->serial);
7379 for (i = 0; i < RX_MAXCALLS; i++) {
7380 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7382 conn->error = ntohl(conn->error);
7383 conn->secStats.flags = ntohl(conn->secStats.flags);
7384 conn->secStats.expires = ntohl(conn->secStats.expires);
7385 conn->secStats.packetsReceived =
7386 ntohl(conn->secStats.packetsReceived);
7387 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7388 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7389 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7390 conn->epoch = ntohl(conn->epoch);
7391 conn->natMTU = ntohl(conn->natMTU);
7398 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7399 afs_uint16 remotePort, afs_int32 * nextPeer,
7400 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7401 afs_uint32 * supportedValues)
7407 struct rx_debugIn in;
7410 * supportedValues is currently unused, but added to allow future
7411 * versioning of this function.
7414 *supportedValues = 0;
7415 in.type = htonl(RX_DEBUGI_GETPEER);
7416 in.index = htonl(*nextPeer);
7417 memset(peer, 0, sizeof(*peer));
7419 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7420 &in, sizeof(in), peer, sizeof(*peer));
7426 * Do net to host conversion here
7428 * I don't convert host or port since we are most likely
7429 * going to want these in NBO.
7431 peer->ifMTU = ntohs(peer->ifMTU);
7432 peer->idleWhen = ntohl(peer->idleWhen);
7433 peer->refCount = ntohs(peer->refCount);
7434 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7435 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7436 peer->rtt = ntohl(peer->rtt);
7437 peer->rtt_dev = ntohl(peer->rtt_dev);
7438 peer->timeout.sec = ntohl(peer->timeout.sec);
7439 peer->timeout.usec = ntohl(peer->timeout.usec);
7440 peer->nSent = ntohl(peer->nSent);
7441 peer->reSends = ntohl(peer->reSends);
7442 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7443 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7444 peer->rateFlag = ntohl(peer->rateFlag);
7445 peer->natMTU = ntohs(peer->natMTU);
7446 peer->maxMTU = ntohs(peer->maxMTU);
7447 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7448 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7449 peer->MTU = ntohs(peer->MTU);
7450 peer->cwind = ntohs(peer->cwind);
7451 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7452 peer->congestSeq = ntohs(peer->congestSeq);
7453 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7454 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7455 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7456 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7463 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7464 struct rx_debugPeer * peerStats)
7467 afs_int32 error = 1; /* default to "did not succeed" */
7468 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7470 MUTEX_ENTER(&rx_peerHashTable_lock);
7471 for(tp = rx_peerHashTable[hashValue];
7472 tp != NULL; tp = tp->next) {
7473 if (tp->host == peerHost)
7479 MUTEX_EXIT(&rx_peerHashTable_lock);
7483 MUTEX_ENTER(&tp->peer_lock);
7484 peerStats->host = tp->host;
7485 peerStats->port = tp->port;
7486 peerStats->ifMTU = tp->ifMTU;
7487 peerStats->idleWhen = tp->idleWhen;
7488 peerStats->refCount = tp->refCount;
7489 peerStats->burstSize = tp->burstSize;
7490 peerStats->burst = tp->burst;
7491 peerStats->burstWait.sec = tp->burstWait.sec;
7492 peerStats->burstWait.usec = tp->burstWait.usec;
7493 peerStats->rtt = tp->rtt;
7494 peerStats->rtt_dev = tp->rtt_dev;
7495 peerStats->timeout.sec = tp->timeout.sec;
7496 peerStats->timeout.usec = tp->timeout.usec;
7497 peerStats->nSent = tp->nSent;
7498 peerStats->reSends = tp->reSends;
7499 peerStats->inPacketSkew = tp->inPacketSkew;
7500 peerStats->outPacketSkew = tp->outPacketSkew;
7501 peerStats->rateFlag = tp->rateFlag;
7502 peerStats->natMTU = tp->natMTU;
7503 peerStats->maxMTU = tp->maxMTU;
7504 peerStats->maxDgramPackets = tp->maxDgramPackets;
7505 peerStats->ifDgramPackets = tp->ifDgramPackets;
7506 peerStats->MTU = tp->MTU;
7507 peerStats->cwind = tp->cwind;
7508 peerStats->nDgramPackets = tp->nDgramPackets;
7509 peerStats->congestSeq = tp->congestSeq;
7510 peerStats->bytesSent.high = tp->bytesSent.high;
7511 peerStats->bytesSent.low = tp->bytesSent.low;
7512 peerStats->bytesReceived.high = tp->bytesReceived.high;
7513 peerStats->bytesReceived.low = tp->bytesReceived.low;
7514 MUTEX_EXIT(&tp->peer_lock);
7516 MUTEX_ENTER(&rx_peerHashTable_lock);
7519 MUTEX_EXIT(&rx_peerHashTable_lock);
7527 struct rx_serverQueueEntry *np;
7530 struct rx_call *call;
7531 struct rx_serverQueueEntry *sq;
7535 if (rxinit_status == 1) {
7537 return; /* Already shutdown. */
7541 #ifndef AFS_PTHREAD_ENV
7542 FD_ZERO(&rx_selectMask);
7543 #endif /* AFS_PTHREAD_ENV */
7544 rxi_dataQuota = RX_MAX_QUOTA;
7545 #ifndef AFS_PTHREAD_ENV
7547 #endif /* AFS_PTHREAD_ENV */
7550 #ifndef AFS_PTHREAD_ENV
7551 #ifndef AFS_USE_GETTIMEOFDAY
7553 #endif /* AFS_USE_GETTIMEOFDAY */
7554 #endif /* AFS_PTHREAD_ENV */
7556 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7557 call = queue_First(&rx_freeCallQueue, rx_call);
7559 rxi_Free(call, sizeof(struct rx_call));
7562 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7563 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7569 struct rx_peer **peer_ptr, **peer_end;
7570 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7571 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7573 struct rx_peer *peer, *next;
7575 MUTEX_ENTER(&rx_peerHashTable_lock);
7576 for (peer = *peer_ptr; peer; peer = next) {
7577 rx_interface_stat_p rpc_stat, nrpc_stat;
7580 MUTEX_ENTER(&rx_rpc_stats);
7581 MUTEX_ENTER(&peer->peer_lock);
7583 (&peer->rpcStats, rpc_stat, nrpc_stat,
7584 rx_interface_stat)) {
7585 unsigned int num_funcs;
7588 queue_Remove(&rpc_stat->queue_header);
7589 queue_Remove(&rpc_stat->all_peers);
7590 num_funcs = rpc_stat->stats[0].func_total;
7592 sizeof(rx_interface_stat_t) +
7593 rpc_stat->stats[0].func_total *
7594 sizeof(rx_function_entry_v1_t);
7596 rxi_Free(rpc_stat, space);
7598 /* rx_rpc_stats must be held */
7599 rxi_rpc_peer_stat_cnt -= num_funcs;
7601 MUTEX_EXIT(&peer->peer_lock);
7602 MUTEX_EXIT(&rx_rpc_stats);
7606 if (rx_stats_active)
7607 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
7609 MUTEX_EXIT(&rx_peerHashTable_lock);
7612 for (i = 0; i < RX_MAX_SERVICES; i++) {
7614 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7616 for (i = 0; i < rx_hashTableSize; i++) {
7617 struct rx_connection *tc, *ntc;
7618 MUTEX_ENTER(&rx_connHashTable_lock);
7619 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7621 for (j = 0; j < RX_MAXCALLS; j++) {
7623 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7626 rxi_Free(tc, sizeof(*tc));
7628 MUTEX_EXIT(&rx_connHashTable_lock);
7631 MUTEX_ENTER(&freeSQEList_lock);
7633 while ((np = rx_FreeSQEList)) {
7634 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7635 MUTEX_DESTROY(&np->lock);
7636 rxi_Free(np, sizeof(*np));
7639 MUTEX_EXIT(&freeSQEList_lock);
7640 MUTEX_DESTROY(&freeSQEList_lock);
7641 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7642 MUTEX_DESTROY(&rx_connHashTable_lock);
7643 MUTEX_DESTROY(&rx_peerHashTable_lock);
7644 MUTEX_DESTROY(&rx_serverPool_lock);
7646 osi_Free(rx_connHashTable,
7647 rx_hashTableSize * sizeof(struct rx_connection *));
7648 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7650 UNPIN(rx_connHashTable,
7651 rx_hashTableSize * sizeof(struct rx_connection *));
7652 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7654 rxi_FreeAllPackets();
7656 MUTEX_ENTER(&rx_quota_mutex);
7657 rxi_dataQuota = RX_MAX_QUOTA;
7658 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7659 MUTEX_EXIT(&rx_quota_mutex);
7664 #ifdef RX_ENABLE_LOCKS
7666 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7668 if (!MUTEX_ISMINE(lockaddr))
7669 osi_Panic("Lock not held: %s", msg);
7671 #endif /* RX_ENABLE_LOCKS */
7676 * Routines to implement connection specific data.
7680 rx_KeyCreate(rx_destructor_t rtn)
7683 MUTEX_ENTER(&rxi_keyCreate_lock);
7684 key = rxi_keyCreate_counter++;
7685 rxi_keyCreate_destructor = (rx_destructor_t *)
7686 realloc((void *)rxi_keyCreate_destructor,
7687 (key + 1) * sizeof(rx_destructor_t));
7688 rxi_keyCreate_destructor[key] = rtn;
7689 MUTEX_EXIT(&rxi_keyCreate_lock);
7694 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7697 MUTEX_ENTER(&conn->conn_data_lock);
7698 if (!conn->specific) {
7699 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7700 for (i = 0; i < key; i++)
7701 conn->specific[i] = NULL;
7702 conn->nSpecific = key + 1;
7703 conn->specific[key] = ptr;
7704 } else if (key >= conn->nSpecific) {
7705 conn->specific = (void **)
7706 realloc(conn->specific, (key + 1) * sizeof(void *));
7707 for (i = conn->nSpecific; i < key; i++)
7708 conn->specific[i] = NULL;
7709 conn->nSpecific = key + 1;
7710 conn->specific[key] = ptr;
7712 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7713 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7714 conn->specific[key] = ptr;
7716 MUTEX_EXIT(&conn->conn_data_lock);
7720 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7723 MUTEX_ENTER(&svc->svc_data_lock);
7724 if (!svc->specific) {
7725 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
7726 for (i = 0; i < key; i++)
7727 svc->specific[i] = NULL;
7728 svc->nSpecific = key + 1;
7729 svc->specific[key] = ptr;
7730 } else if (key >= svc->nSpecific) {
7731 svc->specific = (void **)
7732 realloc(svc->specific, (key + 1) * sizeof(void *));
7733 for (i = svc->nSpecific; i < key; i++)
7734 svc->specific[i] = NULL;
7735 svc->nSpecific = key + 1;
7736 svc->specific[key] = ptr;
7738 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7739 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7740 svc->specific[key] = ptr;
7742 MUTEX_EXIT(&svc->svc_data_lock);
7746 rx_GetSpecific(struct rx_connection *conn, int key)
7749 MUTEX_ENTER(&conn->conn_data_lock);
7750 if (key >= conn->nSpecific)
7753 ptr = conn->specific[key];
7754 MUTEX_EXIT(&conn->conn_data_lock);
7759 rx_GetServiceSpecific(struct rx_service *svc, int key)
7762 MUTEX_ENTER(&svc->svc_data_lock);
7763 if (key >= svc->nSpecific)
7766 ptr = svc->specific[key];
7767 MUTEX_EXIT(&svc->svc_data_lock);
7772 #endif /* !KERNEL */
7775 * processStats is a queue used to store the statistics for the local
7776 * process. Its contents are similar to the contents of the rpcStats
7777 * queue on a rx_peer structure, but the actual data stored within
7778 * this queue contains totals across the lifetime of the process (assuming
7779 * the stats have not been reset) - unlike the per peer structures
7780 * which can come and go based upon the peer lifetime.
7783 static struct rx_queue processStats = { &processStats, &processStats };
7786 * peerStats is a queue used to store the statistics for all peer structs.
7787 * Its contents are the union of all the peer rpcStats queues.
7790 static struct rx_queue peerStats = { &peerStats, &peerStats };
7793 * rxi_monitor_processStats is used to turn process wide stat collection
7797 static int rxi_monitor_processStats = 0;
7800 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7803 static int rxi_monitor_peerStats = 0;
7806 * rxi_AddRpcStat - given all of the information for a particular rpc
7807 * call, create (if needed) and update the stat totals for the rpc.
7811 * IN stats - the queue of stats that will be updated with the new value
7813 * IN rxInterface - a unique number that identifies the rpc interface
7815 * IN currentFunc - the index of the function being invoked
7817 * IN totalFunc - the total number of functions in this interface
7819 * IN queueTime - the amount of time this function waited for a thread
7821 * IN execTime - the amount of time this function invocation took to execute
7823 * IN bytesSent - the number bytes sent by this invocation
7825 * IN bytesRcvd - the number bytes received by this invocation
7827 * IN isServer - if true, this invocation was made to a server
7829 * IN remoteHost - the ip address of the remote host
7831 * IN remotePort - the port of the remote host
7833 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7835 * INOUT counter - if a new stats structure is allocated, the counter will
7836 * be updated with the new number of allocated stat structures
7844 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
7845 afs_uint32 currentFunc, afs_uint32 totalFunc,
7846 struct clock *queueTime, struct clock *execTime,
7847 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
7848 afs_uint32 remoteHost, afs_uint32 remotePort,
7849 int addToPeerList, unsigned int *counter)
7852 rx_interface_stat_p rpc_stat, nrpc_stat;
7855 * See if there's already a structure for this interface
7858 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7859 if ((rpc_stat->stats[0].interfaceId == rxInterface)
7860 && (rpc_stat->stats[0].remote_is_server == isServer))
7865 * Didn't find a match so allocate a new structure and add it to the
7869 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
7870 || (rpc_stat->stats[0].interfaceId != rxInterface)
7871 || (rpc_stat->stats[0].remote_is_server != isServer)) {
7876 sizeof(rx_interface_stat_t) +
7877 totalFunc * sizeof(rx_function_entry_v1_t);
7879 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
7880 if (rpc_stat == NULL) {
7884 *counter += totalFunc;
7885 for (i = 0; i < totalFunc; i++) {
7886 rpc_stat->stats[i].remote_peer = remoteHost;
7887 rpc_stat->stats[i].remote_port = remotePort;
7888 rpc_stat->stats[i].remote_is_server = isServer;
7889 rpc_stat->stats[i].interfaceId = rxInterface;
7890 rpc_stat->stats[i].func_total = totalFunc;
7891 rpc_stat->stats[i].func_index = i;
7892 hzero(rpc_stat->stats[i].invocations);
7893 hzero(rpc_stat->stats[i].bytes_sent);
7894 hzero(rpc_stat->stats[i].bytes_rcvd);
7895 rpc_stat->stats[i].queue_time_sum.sec = 0;
7896 rpc_stat->stats[i].queue_time_sum.usec = 0;
7897 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7898 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7899 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7900 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7901 rpc_stat->stats[i].queue_time_max.sec = 0;
7902 rpc_stat->stats[i].queue_time_max.usec = 0;
7903 rpc_stat->stats[i].execution_time_sum.sec = 0;
7904 rpc_stat->stats[i].execution_time_sum.usec = 0;
7905 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7906 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7907 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7908 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7909 rpc_stat->stats[i].execution_time_max.sec = 0;
7910 rpc_stat->stats[i].execution_time_max.usec = 0;
7912 queue_Prepend(stats, rpc_stat);
7913 if (addToPeerList) {
7914 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7919 * Increment the stats for this function
7922 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7923 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7924 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7925 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7926 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7927 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7928 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7930 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7931 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7933 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7934 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7936 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7937 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7939 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7940 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7948 * rx_IncrementTimeAndCount - increment the times and count for a particular
7953 * IN peer - the peer who invoked the rpc
7955 * IN rxInterface - a unique number that identifies the rpc interface
7957 * IN currentFunc - the index of the function being invoked
7959 * IN totalFunc - the total number of functions in this interface
7961 * IN queueTime - the amount of time this function waited for a thread
7963 * IN execTime - the amount of time this function invocation took to execute
7965 * IN bytesSent - the number bytes sent by this invocation
7967 * IN bytesRcvd - the number bytes received by this invocation
7969 * IN isServer - if true, this invocation was made to a server
7977 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7978 afs_uint32 currentFunc, afs_uint32 totalFunc,
7979 struct clock *queueTime, struct clock *execTime,
7980 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7984 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7987 MUTEX_ENTER(&rx_rpc_stats);
7989 if (rxi_monitor_peerStats) {
7990 MUTEX_ENTER(&peer->peer_lock);
7991 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7992 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7993 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7994 MUTEX_EXIT(&peer->peer_lock);
7997 if (rxi_monitor_processStats) {
7998 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7999 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8000 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8003 MUTEX_EXIT(&rx_rpc_stats);
8008 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8012 * IN callerVersion - the rpc stat version of the caller.
8014 * IN count - the number of entries to marshall.
8016 * IN stats - pointer to stats to be marshalled.
8018 * OUT ptr - Where to store the marshalled data.
8025 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8026 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8032 * We only support the first version
8034 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8035 *(ptr++) = stats->remote_peer;
8036 *(ptr++) = stats->remote_port;
8037 *(ptr++) = stats->remote_is_server;
8038 *(ptr++) = stats->interfaceId;
8039 *(ptr++) = stats->func_total;
8040 *(ptr++) = stats->func_index;
8041 *(ptr++) = hgethi(stats->invocations);
8042 *(ptr++) = hgetlo(stats->invocations);
8043 *(ptr++) = hgethi(stats->bytes_sent);
8044 *(ptr++) = hgetlo(stats->bytes_sent);
8045 *(ptr++) = hgethi(stats->bytes_rcvd);
8046 *(ptr++) = hgetlo(stats->bytes_rcvd);
8047 *(ptr++) = stats->queue_time_sum.sec;
8048 *(ptr++) = stats->queue_time_sum.usec;
8049 *(ptr++) = stats->queue_time_sum_sqr.sec;
8050 *(ptr++) = stats->queue_time_sum_sqr.usec;
8051 *(ptr++) = stats->queue_time_min.sec;
8052 *(ptr++) = stats->queue_time_min.usec;
8053 *(ptr++) = stats->queue_time_max.sec;
8054 *(ptr++) = stats->queue_time_max.usec;
8055 *(ptr++) = stats->execution_time_sum.sec;
8056 *(ptr++) = stats->execution_time_sum.usec;
8057 *(ptr++) = stats->execution_time_sum_sqr.sec;
8058 *(ptr++) = stats->execution_time_sum_sqr.usec;
8059 *(ptr++) = stats->execution_time_min.sec;
8060 *(ptr++) = stats->execution_time_min.usec;
8061 *(ptr++) = stats->execution_time_max.sec;
8062 *(ptr++) = stats->execution_time_max.usec;
8068 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8073 * IN callerVersion - the rpc stat version of the caller
8075 * OUT myVersion - the rpc stat version of this function
8077 * OUT clock_sec - local time seconds
8079 * OUT clock_usec - local time microseconds
8081 * OUT allocSize - the number of bytes allocated to contain stats
8083 * OUT statCount - the number stats retrieved from this process.
8085 * OUT stats - the actual stats retrieved from this process.
8089 * Returns void. If successful, stats will != NULL.
8093 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8094 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8095 size_t * allocSize, afs_uint32 * statCount,
8096 afs_uint32 ** stats)
8106 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8109 * Check to see if stats are enabled
8112 MUTEX_ENTER(&rx_rpc_stats);
8113 if (!rxi_monitor_processStats) {
8114 MUTEX_EXIT(&rx_rpc_stats);
8118 clock_GetTime(&now);
8119 *clock_sec = now.sec;
8120 *clock_usec = now.usec;
8123 * Allocate the space based upon the caller version
8125 * If the client is at an older version than we are,
8126 * we return the statistic data in the older data format, but
8127 * we still return our version number so the client knows we
8128 * are maintaining more data than it can retrieve.
8131 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8132 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8133 *statCount = rxi_rpc_process_stat_cnt;
8136 * This can't happen yet, but in the future version changes
8137 * can be handled by adding additional code here
8141 if (space > (size_t) 0) {
8143 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
8146 rx_interface_stat_p rpc_stat, nrpc_stat;
8150 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8152 * Copy the data based upon the caller version
8154 rx_MarshallProcessRPCStats(callerVersion,
8155 rpc_stat->stats[0].func_total,
8156 rpc_stat->stats, &ptr);
8162 MUTEX_EXIT(&rx_rpc_stats);
8167 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8171 * IN callerVersion - the rpc stat version of the caller
8173 * OUT myVersion - the rpc stat version of this function
8175 * OUT clock_sec - local time seconds
8177 * OUT clock_usec - local time microseconds
8179 * OUT allocSize - the number of bytes allocated to contain stats
8181 * OUT statCount - the number of stats retrieved from the individual
8184 * OUT stats - the actual stats retrieved from the individual peer structures.
8188 * Returns void. If successful, stats will != NULL.
8192 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8193 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8194 size_t * allocSize, afs_uint32 * statCount,
8195 afs_uint32 ** stats)
8205 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8208 * Check to see if stats are enabled
8211 MUTEX_ENTER(&rx_rpc_stats);
8212 if (!rxi_monitor_peerStats) {
8213 MUTEX_EXIT(&rx_rpc_stats);
8217 clock_GetTime(&now);
8218 *clock_sec = now.sec;
8219 *clock_usec = now.usec;
8222 * Allocate the space based upon the caller version
8224 * If the client is at an older version than we are,
8225 * we return the statistic data in the older data format, but
8226 * we still return our version number so the client knows we
8227 * are maintaining more data than it can retrieve.
8230 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8231 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8232 *statCount = rxi_rpc_peer_stat_cnt;
8235 * This can't happen yet, but in the future version changes
8236 * can be handled by adding additional code here
8240 if (space > (size_t) 0) {
8242 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
8245 rx_interface_stat_p rpc_stat, nrpc_stat;
8249 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8251 * We have to fix the offset of rpc_stat since we are
8252 * keeping this structure on two rx_queues. The rx_queue
8253 * package assumes that the rx_queue member is the first
8254 * member of the structure. That is, rx_queue assumes that
8255 * any one item is only on one queue at a time. We are
8256 * breaking that assumption and so we have to do a little
8257 * math to fix our pointers.
8260 fix_offset = (char *)rpc_stat;
8261 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8262 rpc_stat = (rx_interface_stat_p) fix_offset;
8265 * Copy the data based upon the caller version
8267 rx_MarshallProcessRPCStats(callerVersion,
8268 rpc_stat->stats[0].func_total,
8269 rpc_stat->stats, &ptr);
8275 MUTEX_EXIT(&rx_rpc_stats);
8280 * rx_FreeRPCStats - free memory allocated by
8281 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8285 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8286 * rx_RetrievePeerRPCStats
8288 * IN allocSize - the number of bytes in stats.
8296 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8298 rxi_Free(stats, allocSize);
8302 * rx_queryProcessRPCStats - see if process rpc stat collection is
8303 * currently enabled.
8309 * Returns 0 if stats are not enabled != 0 otherwise
8313 rx_queryProcessRPCStats(void)
8316 MUTEX_ENTER(&rx_rpc_stats);
8317 rc = rxi_monitor_processStats;
8318 MUTEX_EXIT(&rx_rpc_stats);
8323 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8329 * Returns 0 if stats are not enabled != 0 otherwise
8333 rx_queryPeerRPCStats(void)
8336 MUTEX_ENTER(&rx_rpc_stats);
8337 rc = rxi_monitor_peerStats;
8338 MUTEX_EXIT(&rx_rpc_stats);
8343 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8353 rx_enableProcessRPCStats(void)
8355 MUTEX_ENTER(&rx_rpc_stats);
8356 rx_enable_stats = 1;
8357 rxi_monitor_processStats = 1;
8358 MUTEX_EXIT(&rx_rpc_stats);
8362 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8372 rx_enablePeerRPCStats(void)
8374 MUTEX_ENTER(&rx_rpc_stats);
8375 rx_enable_stats = 1;
8376 rxi_monitor_peerStats = 1;
8377 MUTEX_EXIT(&rx_rpc_stats);
8381 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8391 rx_disableProcessRPCStats(void)
8393 rx_interface_stat_p rpc_stat, nrpc_stat;
8396 MUTEX_ENTER(&rx_rpc_stats);
8399 * Turn off process statistics and if peer stats is also off, turn
8403 rxi_monitor_processStats = 0;
8404 if (rxi_monitor_peerStats == 0) {
8405 rx_enable_stats = 0;
8408 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8409 unsigned int num_funcs = 0;
8412 queue_Remove(rpc_stat);
8413 num_funcs = rpc_stat->stats[0].func_total;
8415 sizeof(rx_interface_stat_t) +
8416 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8418 rxi_Free(rpc_stat, space);
8419 rxi_rpc_process_stat_cnt -= num_funcs;
8421 MUTEX_EXIT(&rx_rpc_stats);
8425 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8435 rx_disablePeerRPCStats(void)
8437 struct rx_peer **peer_ptr, **peer_end;
8441 * Turn off peer statistics and if process stats is also off, turn
8445 rxi_monitor_peerStats = 0;
8446 if (rxi_monitor_processStats == 0) {
8447 rx_enable_stats = 0;
8450 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8451 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8453 struct rx_peer *peer, *next, *prev;
8455 MUTEX_ENTER(&rx_peerHashTable_lock);
8456 MUTEX_ENTER(&rx_rpc_stats);
8457 for (prev = peer = *peer_ptr; peer; peer = next) {
8459 code = MUTEX_TRYENTER(&peer->peer_lock);
8461 rx_interface_stat_p rpc_stat, nrpc_stat;
8464 if (prev == *peer_ptr) {
8475 MUTEX_EXIT(&rx_peerHashTable_lock);
8478 (&peer->rpcStats, rpc_stat, nrpc_stat,
8479 rx_interface_stat)) {
8480 unsigned int num_funcs = 0;
8483 queue_Remove(&rpc_stat->queue_header);
8484 queue_Remove(&rpc_stat->all_peers);
8485 num_funcs = rpc_stat->stats[0].func_total;
8487 sizeof(rx_interface_stat_t) +
8488 rpc_stat->stats[0].func_total *
8489 sizeof(rx_function_entry_v1_t);
8491 rxi_Free(rpc_stat, space);
8492 rxi_rpc_peer_stat_cnt -= num_funcs;
8494 MUTEX_EXIT(&peer->peer_lock);
8496 MUTEX_ENTER(&rx_peerHashTable_lock);
8506 MUTEX_EXIT(&rx_rpc_stats);
8507 MUTEX_EXIT(&rx_peerHashTable_lock);
8512 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8517 * IN clearFlag - flag indicating which stats to clear
8525 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8527 rx_interface_stat_p rpc_stat, nrpc_stat;
8529 MUTEX_ENTER(&rx_rpc_stats);
8531 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8532 unsigned int num_funcs = 0, i;
8533 num_funcs = rpc_stat->stats[0].func_total;
8534 for (i = 0; i < num_funcs; i++) {
8535 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8536 hzero(rpc_stat->stats[i].invocations);
8538 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8539 hzero(rpc_stat->stats[i].bytes_sent);
8541 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8542 hzero(rpc_stat->stats[i].bytes_rcvd);
8544 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8545 rpc_stat->stats[i].queue_time_sum.sec = 0;
8546 rpc_stat->stats[i].queue_time_sum.usec = 0;
8548 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8549 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8550 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8552 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8553 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8554 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8556 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8557 rpc_stat->stats[i].queue_time_max.sec = 0;
8558 rpc_stat->stats[i].queue_time_max.usec = 0;
8560 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8561 rpc_stat->stats[i].execution_time_sum.sec = 0;
8562 rpc_stat->stats[i].execution_time_sum.usec = 0;
8564 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8565 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8566 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8568 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8569 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8570 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8572 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8573 rpc_stat->stats[i].execution_time_max.sec = 0;
8574 rpc_stat->stats[i].execution_time_max.usec = 0;
8579 MUTEX_EXIT(&rx_rpc_stats);
8583 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8588 * IN clearFlag - flag indicating which stats to clear
8596 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8598 rx_interface_stat_p rpc_stat, nrpc_stat;
8600 MUTEX_ENTER(&rx_rpc_stats);
8602 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8603 unsigned int num_funcs = 0, i;
8606 * We have to fix the offset of rpc_stat since we are
8607 * keeping this structure on two rx_queues. The rx_queue
8608 * package assumes that the rx_queue member is the first
8609 * member of the structure. That is, rx_queue assumes that
8610 * any one item is only on one queue at a time. We are
8611 * breaking that assumption and so we have to do a little
8612 * math to fix our pointers.
8615 fix_offset = (char *)rpc_stat;
8616 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8617 rpc_stat = (rx_interface_stat_p) fix_offset;
8619 num_funcs = rpc_stat->stats[0].func_total;
8620 for (i = 0; i < num_funcs; i++) {
8621 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8622 hzero(rpc_stat->stats[i].invocations);
8624 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8625 hzero(rpc_stat->stats[i].bytes_sent);
8627 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8628 hzero(rpc_stat->stats[i].bytes_rcvd);
8630 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8631 rpc_stat->stats[i].queue_time_sum.sec = 0;
8632 rpc_stat->stats[i].queue_time_sum.usec = 0;
8634 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8635 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8636 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8638 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8639 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8640 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8642 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8643 rpc_stat->stats[i].queue_time_max.sec = 0;
8644 rpc_stat->stats[i].queue_time_max.usec = 0;
8646 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8647 rpc_stat->stats[i].execution_time_sum.sec = 0;
8648 rpc_stat->stats[i].execution_time_sum.usec = 0;
8650 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8651 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8652 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8654 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8655 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8656 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8658 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8659 rpc_stat->stats[i].execution_time_max.sec = 0;
8660 rpc_stat->stats[i].execution_time_max.usec = 0;
8665 MUTEX_EXIT(&rx_rpc_stats);
8669 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8670 * is authorized to enable/disable/clear RX statistics.
8672 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8675 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8677 rxi_rxstat_userok = proc;
8681 rx_RxStatUserOk(struct rx_call *call)
8683 if (!rxi_rxstat_userok)
8685 return rxi_rxstat_userok(call);
8690 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8691 * function in the MSVC runtime DLL (msvcrt.dll).
8693 * Note: the system serializes calls to this function.
8696 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8697 DWORD reason, /* reason function is being called */
8698 LPVOID reserved) /* reserved for future use */
8701 case DLL_PROCESS_ATTACH:
8702 /* library is being attached to a process */
8706 case DLL_PROCESS_DETACH:
8713 #endif /* AFS_NT40_ENV */
8716 int rx_DumpCalls(FILE *outputFile, char *cookie)
8718 #ifdef RXDEBUG_PACKET
8719 #ifdef KDUMP_RX_LOCK
8720 struct rx_call_rx_lock *c;
8727 #define RXDPRINTF sprintf
8728 #define RXDPRINTOUT output
8730 #define RXDPRINTF fprintf
8731 #define RXDPRINTOUT outputFile
8734 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8736 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8739 for (c = rx_allCallsp; c; c = c->allNextp) {
8740 u_short rqc, tqc, iovqc;
8741 struct rx_packet *p, *np;
8743 MUTEX_ENTER(&c->lock);
8744 queue_Count(&c->rq, p, np, rx_packet, rqc);
8745 queue_Count(&c->tq, p, np, rx_packet, tqc);
8746 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8748 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, "
8749 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8750 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8751 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8752 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8753 #ifdef RX_ENABLE_LOCKS
8756 #ifdef RX_REFCOUNT_CHECK
8757 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
8758 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
8761 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,
8762 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
8763 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
8764 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
8765 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
8766 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
8767 #ifdef RX_ENABLE_LOCKS
8768 , (afs_uint32)c->refCount
8770 #ifdef RX_REFCOUNT_CHECK
8771 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
8774 MUTEX_EXIT(&c->lock);
8777 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8780 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
8782 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8784 #endif /* RXDEBUG_PACKET */