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>
91 # include <afs/afsutil.h>
92 # include <WINNT\afsreg.h>
94 # include <sys/socket.h>
95 # include <sys/file.h>
97 # include <sys/stat.h>
98 # include <netinet/in.h>
99 # include <sys/time.h>
102 # include "rx_user.h"
103 # include "rx_clock.h"
104 # include "rx_queue.h"
105 # include "rx_globals.h"
106 # include "rx_trace.h"
107 # include <afs/rxgen_consts.h>
111 #ifdef AFS_PTHREAD_ENV
113 int (*registerProgram) (pid_t, char *) = 0;
114 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
117 int (*registerProgram) (PROCESS, char *) = 0;
118 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
122 /* Local static routines */
123 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
124 #ifdef RX_ENABLE_LOCKS
125 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
128 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
130 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
131 afs_int32 rxi_start_in_error;
133 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
136 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
137 * currently allocated within rx. This number is used to allocate the
138 * memory required to return the statistics when queried.
141 static unsigned int rxi_rpc_peer_stat_cnt;
144 * rxi_rpc_process_stat_cnt counts the total number of local process stat
145 * structures currently allocated within rx. The number is used to allocate
146 * the memory required to return the statistics when queried.
149 static unsigned int rxi_rpc_process_stat_cnt;
151 #if !defined(offsetof)
152 #include <stddef.h> /* for definition of offsetof() */
155 #ifdef AFS_PTHREAD_ENV
159 * Use procedural initialization of mutexes/condition variables
163 extern afs_kmutex_t rx_stats_mutex;
164 extern afs_kmutex_t rx_waiting_mutex;
165 extern afs_kmutex_t rx_quota_mutex;
166 extern afs_kmutex_t rx_pthread_mutex;
167 extern afs_kmutex_t rx_packets_mutex;
168 extern afs_kmutex_t des_init_mutex;
169 extern afs_kmutex_t des_random_mutex;
170 extern afs_kmutex_t rx_clock_mutex;
171 extern afs_kmutex_t rxi_connCacheMutex;
172 extern afs_kmutex_t rx_event_mutex;
173 extern afs_kmutex_t osi_malloc_mutex;
174 extern afs_kmutex_t event_handler_mutex;
175 extern afs_kmutex_t listener_mutex;
176 extern afs_kmutex_t rx_if_init_mutex;
177 extern afs_kmutex_t rx_if_mutex;
178 extern afs_kmutex_t rxkad_client_uid_mutex;
179 extern afs_kmutex_t rxkad_random_mutex;
181 extern afs_kcondvar_t rx_event_handler_cond;
182 extern afs_kcondvar_t rx_listener_cond;
184 static afs_kmutex_t epoch_mutex;
185 static afs_kmutex_t rx_init_mutex;
186 static afs_kmutex_t rx_debug_mutex;
187 static afs_kmutex_t rx_rpc_stats;
190 rxi_InitPthread(void)
192 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
193 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
194 MUTEX_INIT(&rx_waiting_mutex, "waiting", MUTEX_DEFAULT, 0);
195 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
196 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
197 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
198 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
199 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
200 MUTEX_INIT(&rx_event_mutex, "event", MUTEX_DEFAULT, 0);
201 MUTEX_INIT(&des_init_mutex, "des", MUTEX_DEFAULT, 0);
202 MUTEX_INIT(&des_random_mutex, "random", MUTEX_DEFAULT, 0);
203 MUTEX_INIT(&osi_malloc_mutex, "malloc", MUTEX_DEFAULT, 0);
204 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
205 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
206 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
207 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
208 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
209 MUTEX_INIT(&rxkad_client_uid_mutex, "uid", MUTEX_DEFAULT, 0);
210 MUTEX_INIT(&rxkad_random_mutex, "rxkad random", MUTEX_DEFAULT, 0);
211 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
213 assert(pthread_cond_init
214 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
215 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
217 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
218 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
220 rxkad_global_stats_init();
222 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
223 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
224 #ifdef RX_ENABLE_LOCKS
227 #endif /* RX_LOCKS_DB */
228 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
229 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
231 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
233 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
235 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
237 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
238 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
239 #endif /* RX_ENABLE_LOCKS */
242 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
243 #define INIT_PTHREAD_LOCKS \
244 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
246 * The rx_stats_mutex mutex protects the following global variables:
247 * rxi_lowConnRefCount
248 * rxi_lowPeerRefCount
257 * The rx_quota_mutex mutex protects the following global variables:
265 * The rx_freePktQ_lock protects the following global variables:
270 * The rx_packets_mutex mutex protects the following global variables:
278 * The rx_pthread_mutex mutex protects the following global variables:
282 #define INIT_PTHREAD_LOCKS
286 /* Variables for handling the minProcs implementation. availProcs gives the
287 * number of threads available in the pool at this moment (not counting dudes
288 * executing right now). totalMin gives the total number of procs required
289 * for handling all minProcs requests. minDeficit is a dynamic variable
290 * tracking the # of procs required to satisfy all of the remaining minProcs
292 * For fine grain locking to work, the quota check and the reservation of
293 * a server thread has to come while rxi_availProcs and rxi_minDeficit
294 * are locked. To this end, the code has been modified under #ifdef
295 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
296 * same time. A new function, ReturnToServerPool() returns the allocation.
298 * A call can be on several queue's (but only one at a time). When
299 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
300 * that no one else is touching the queue. To this end, we store the address
301 * of the queue lock in the call structure (under the call lock) when we
302 * put the call on a queue, and we clear the call_queue_lock when the
303 * call is removed from a queue (once the call lock has been obtained).
304 * This allows rxi_ResetCall to safely synchronize with others wishing
305 * to manipulate the queue.
308 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
309 static afs_kmutex_t rx_rpc_stats;
310 void rxi_StartUnlocked(struct rxevent *event, void *call,
311 void *arg1, int istack);
314 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
315 ** pretty good that the next packet coming in is from the same connection
316 ** as the last packet, since we're send multiple packets in a transmit window.
318 struct rx_connection *rxLastConn = 0;
320 #ifdef RX_ENABLE_LOCKS
321 /* The locking hierarchy for rx fine grain locking is composed of these
324 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
325 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
326 * call->lock - locks call data fields.
327 * These are independent of each other:
328 * rx_freeCallQueue_lock
333 * serverQueueEntry->lock
335 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
336 * peer->lock - locks peer data fields.
337 * conn_data_lock - that more than one thread is not updating a conn data
338 * field at the same time.
346 * Do we need a lock to protect the peer field in the conn structure?
347 * conn->peer was previously a constant for all intents and so has no
348 * lock protecting this field. The multihomed client delta introduced
349 * a RX code change : change the peer field in the connection structure
350 * to that remote inetrface from which the last packet for this
351 * connection was sent out. This may become an issue if further changes
354 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
355 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
357 /* rxdb_fileID is used to identify the lock location, along with line#. */
358 static int rxdb_fileID = RXDB_FILE_RX;
359 #endif /* RX_LOCKS_DB */
360 #else /* RX_ENABLE_LOCKS */
361 #define SET_CALL_QUEUE_LOCK(C, L)
362 #define CLEAR_CALL_QUEUE_LOCK(C)
363 #endif /* RX_ENABLE_LOCKS */
364 struct rx_serverQueueEntry *rx_waitForPacket = 0;
365 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
367 /* ------------Exported Interfaces------------- */
369 /* This function allows rxkad to set the epoch to a suitably random number
370 * which rx_NewConnection will use in the future. The principle purpose is to
371 * get rxnull connections to use the same epoch as the rxkad connections do, at
372 * least once the first rxkad connection is established. This is important now
373 * that the host/port addresses aren't used in FindConnection: the uniqueness
374 * of epoch/cid matters and the start time won't do. */
376 #ifdef AFS_PTHREAD_ENV
378 * This mutex protects the following global variables:
382 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
383 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
387 #endif /* AFS_PTHREAD_ENV */
390 rx_SetEpoch(afs_uint32 epoch)
397 /* Initialize rx. A port number may be mentioned, in which case this
398 * becomes the default port number for any service installed later.
399 * If 0 is provided for the port number, a random port will be chosen
400 * by the kernel. Whether this will ever overlap anything in
401 * /etc/services is anybody's guess... Returns 0 on success, -1 on
406 int rxinit_status = 1;
407 #ifdef AFS_PTHREAD_ENV
409 * This mutex protects the following global variables:
413 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
414 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
417 #define UNLOCK_RX_INIT
421 rx_InitHost(u_int host, u_int port)
428 char *htable, *ptable;
435 if (rxinit_status == 0) {
436 tmp_status = rxinit_status;
438 return tmp_status; /* Already started; return previous error code. */
444 if (afs_winsockInit() < 0)
450 * Initialize anything necessary to provide a non-premptive threading
453 rxi_InitializeThreadSupport();
456 /* Allocate and initialize a socket for client and perhaps server
459 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
460 if (rx_socket == OSI_NULLSOCKET) {
464 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
467 #endif /* RX_LOCKS_DB */
468 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
469 MUTEX_INIT(&rx_waiting_mutex, "rx_waiting_mutex", MUTEX_DEFAULT, 0);
470 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
471 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
472 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
473 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
474 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
475 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
476 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
478 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
480 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
482 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
484 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
485 #if defined(AFS_HPUX110_ENV)
487 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
488 #endif /* AFS_HPUX110_ENV */
489 #endif /* RX_ENABLE_LOCKS && KERNEL */
492 rx_connDeadTime = 12;
493 rx_tranquil = 0; /* reset flag */
494 memset((char *)&rx_stats, 0, sizeof(struct rx_statistics));
496 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
497 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
498 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
499 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
500 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
501 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
503 /* Malloc up a bunch of packets & buffers */
505 queue_Init(&rx_freePacketQueue);
506 rxi_NeedMorePackets = FALSE;
507 #ifdef RX_ENABLE_TSFPQ
508 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
509 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
510 #else /* RX_ENABLE_TSFPQ */
511 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
512 rxi_MorePackets(rx_nPackets);
513 #endif /* RX_ENABLE_TSFPQ */
520 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
521 tv.tv_sec = clock_now.sec;
522 tv.tv_usec = clock_now.usec;
523 srand((unsigned int)tv.tv_usec);
530 #if defined(KERNEL) && !defined(UKERNEL)
531 /* Really, this should never happen in a real kernel */
534 struct sockaddr_in addr;
536 int addrlen = sizeof(addr);
538 socklen_t addrlen = sizeof(addr);
540 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
544 rx_port = addr.sin_port;
547 rx_stats.minRtt.sec = 9999999;
549 rx_SetEpoch(tv.tv_sec | 0x80000000);
551 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
552 * will provide a randomer value. */
554 MUTEX_ENTER(&rx_quota_mutex);
555 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
556 MUTEX_EXIT(&rx_quota_mutex);
557 /* *Slightly* random start time for the cid. This is just to help
558 * out with the hashing function at the peer */
559 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
560 rx_connHashTable = (struct rx_connection **)htable;
561 rx_peerHashTable = (struct rx_peer **)ptable;
563 rx_lastAckDelay.sec = 0;
564 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
565 rx_hardAckDelay.sec = 0;
566 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
567 rx_softAckDelay.sec = 0;
568 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
570 rxevent_Init(20, rxi_ReScheduleEvents);
572 /* Initialize various global queues */
573 queue_Init(&rx_idleServerQueue);
574 queue_Init(&rx_incomingCallQueue);
575 queue_Init(&rx_freeCallQueue);
577 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
578 /* Initialize our list of usable IP addresses. */
582 /* Start listener process (exact function is dependent on the
583 * implementation environment--kernel or user space) */
587 tmp_status = rxinit_status = 0;
595 return rx_InitHost(htonl(INADDR_ANY), port);
598 /* called with unincremented nRequestsRunning to see if it is OK to start
599 * a new thread in this service. Could be "no" for two reasons: over the
600 * max quota, or would prevent others from reaching their min quota.
602 #ifdef RX_ENABLE_LOCKS
603 /* This verion of QuotaOK reserves quota if it's ok while the
604 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
607 QuotaOK(struct rx_service *aservice)
609 /* check if over max quota */
610 if (aservice->nRequestsRunning >= aservice->maxProcs) {
614 /* under min quota, we're OK */
615 /* otherwise, can use only if there are enough to allow everyone
616 * to go to their min quota after this guy starts.
619 MUTEX_ENTER(&rx_quota_mutex);
620 if ((aservice->nRequestsRunning < aservice->minProcs)
621 || (rxi_availProcs > rxi_minDeficit)) {
622 aservice->nRequestsRunning++;
623 /* just started call in minProcs pool, need fewer to maintain
625 if (aservice->nRequestsRunning <= aservice->minProcs)
628 MUTEX_EXIT(&rx_quota_mutex);
631 MUTEX_EXIT(&rx_quota_mutex);
637 ReturnToServerPool(struct rx_service *aservice)
639 aservice->nRequestsRunning--;
640 MUTEX_ENTER(&rx_quota_mutex);
641 if (aservice->nRequestsRunning < aservice->minProcs)
644 MUTEX_EXIT(&rx_quota_mutex);
647 #else /* RX_ENABLE_LOCKS */
649 QuotaOK(struct rx_service *aservice)
652 /* under min quota, we're OK */
653 if (aservice->nRequestsRunning < aservice->minProcs)
656 /* check if over max quota */
657 if (aservice->nRequestsRunning >= aservice->maxProcs)
660 /* otherwise, can use only if there are enough to allow everyone
661 * to go to their min quota after this guy starts.
663 if (rxi_availProcs > rxi_minDeficit)
667 #endif /* RX_ENABLE_LOCKS */
670 /* Called by rx_StartServer to start up lwp's to service calls.
671 NExistingProcs gives the number of procs already existing, and which
672 therefore needn't be created. */
674 rxi_StartServerProcs(int nExistingProcs)
676 struct rx_service *service;
681 /* For each service, reserve N processes, where N is the "minimum"
682 * number of processes that MUST be able to execute a request in parallel,
683 * at any time, for that process. Also compute the maximum difference
684 * between any service's maximum number of processes that can run
685 * (i.e. the maximum number that ever will be run, and a guarantee
686 * that this number will run if other services aren't running), and its
687 * minimum number. The result is the extra number of processes that
688 * we need in order to provide the latter guarantee */
689 for (i = 0; i < RX_MAX_SERVICES; i++) {
691 service = rx_services[i];
692 if (service == (struct rx_service *)0)
694 nProcs += service->minProcs;
695 diff = service->maxProcs - service->minProcs;
699 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
700 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
701 for (i = 0; i < nProcs; i++) {
702 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
708 /* This routine is only required on Windows */
710 rx_StartClientThread(void)
712 #ifdef AFS_PTHREAD_ENV
714 pid = pthread_self();
715 #endif /* AFS_PTHREAD_ENV */
717 #endif /* AFS_NT40_ENV */
719 /* This routine must be called if any services are exported. If the
720 * donateMe flag is set, the calling process is donated to the server
723 rx_StartServer(int donateMe)
725 struct rx_service *service;
731 /* Start server processes, if necessary (exact function is dependent
732 * on the implementation environment--kernel or user space). DonateMe
733 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
734 * case, one less new proc will be created rx_StartServerProcs.
736 rxi_StartServerProcs(donateMe);
738 /* count up the # of threads in minProcs, and add set the min deficit to
739 * be that value, too.
741 for (i = 0; i < RX_MAX_SERVICES; i++) {
742 service = rx_services[i];
743 if (service == (struct rx_service *)0)
745 MUTEX_ENTER(&rx_quota_mutex);
746 rxi_totalMin += service->minProcs;
747 /* below works even if a thread is running, since minDeficit would
748 * still have been decremented and later re-incremented.
750 rxi_minDeficit += service->minProcs;
751 MUTEX_EXIT(&rx_quota_mutex);
754 /* Turn on reaping of idle server connections */
755 rxi_ReapConnections(NULL, NULL, NULL);
764 #ifdef AFS_PTHREAD_ENV
766 pid = (pid_t) pthread_self();
767 #else /* AFS_PTHREAD_ENV */
769 LWP_CurrentProcess(&pid);
770 #endif /* AFS_PTHREAD_ENV */
772 sprintf(name, "srv_%d", ++nProcs);
774 (*registerProgram) (pid, name);
776 #endif /* AFS_NT40_ENV */
777 rx_ServerProc(NULL); /* Never returns */
779 #ifdef RX_ENABLE_TSFPQ
780 /* no use leaving packets around in this thread's local queue if
781 * it isn't getting donated to the server thread pool.
783 rxi_FlushLocalPacketsTSFPQ();
784 #endif /* RX_ENABLE_TSFPQ */
788 /* Create a new client connection to the specified service, using the
789 * specified security object to implement the security model for this
791 struct rx_connection *
792 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
793 struct rx_securityClass *securityObject,
794 int serviceSecurityIndex)
798 struct rx_connection *conn;
803 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", ntohl(shost), ntohs(sport), sservice, securityObject, serviceSecurityIndex));
805 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
806 * the case of kmem_alloc? */
807 conn = rxi_AllocConnection();
808 #ifdef RX_ENABLE_LOCKS
809 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
810 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
811 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
814 MUTEX_ENTER(&rx_connHashTable_lock);
815 cid = (rx_nextCid += RX_MAXCALLS);
816 conn->type = RX_CLIENT_CONNECTION;
818 conn->epoch = rx_epoch;
819 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
820 conn->serviceId = sservice;
821 conn->securityObject = securityObject;
822 conn->securityData = (void *) 0;
823 conn->securityIndex = serviceSecurityIndex;
824 rx_SetConnDeadTime(conn, rx_connDeadTime);
825 conn->ackRate = RX_FAST_ACK_RATE;
827 conn->specific = NULL;
828 conn->challengeEvent = NULL;
829 conn->delayedAbortEvent = NULL;
830 conn->abortCount = 0;
832 for (i = 0; i < RX_MAXCALLS; i++) {
833 conn->twind[i] = rx_initSendWindow;
834 conn->rwind[i] = rx_initReceiveWindow;
837 RXS_NewConnection(securityObject, conn);
839 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
841 conn->refCount++; /* no lock required since only this thread knows... */
842 conn->next = rx_connHashTable[hashindex];
843 rx_connHashTable[hashindex] = conn;
845 rx_MutexIncrement(rx_stats.nClientConns, rx_stats_mutex);
846 MUTEX_EXIT(&rx_connHashTable_lock);
852 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
854 /* The idea is to set the dead time to a value that allows several
855 * keepalives to be dropped without timing out the connection. */
856 conn->secondsUntilDead = MAX(seconds, 6);
857 conn->secondsUntilPing = conn->secondsUntilDead / 6;
860 int rxi_lowPeerRefCount = 0;
861 int rxi_lowConnRefCount = 0;
864 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
865 * NOTE: must not be called with rx_connHashTable_lock held.
868 rxi_CleanupConnection(struct rx_connection *conn)
870 /* Notify the service exporter, if requested, that this connection
871 * is being destroyed */
872 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
873 (*conn->service->destroyConnProc) (conn);
875 /* Notify the security module that this connection is being destroyed */
876 RXS_DestroyConnection(conn->securityObject, conn);
878 /* If this is the last connection using the rx_peer struct, set its
879 * idle time to now. rxi_ReapConnections will reap it if it's still
880 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
882 MUTEX_ENTER(&rx_peerHashTable_lock);
883 if (conn->peer->refCount < 2) {
884 conn->peer->idleWhen = clock_Sec();
885 if (conn->peer->refCount < 1) {
886 conn->peer->refCount = 1;
887 if (rx_stats_active) {
888 MUTEX_ENTER(&rx_stats_mutex);
889 rxi_lowPeerRefCount++;
890 MUTEX_EXIT(&rx_stats_mutex);
894 conn->peer->refCount--;
895 MUTEX_EXIT(&rx_peerHashTable_lock);
899 if (conn->type == RX_SERVER_CONNECTION)
900 rx_MutexDecrement(rx_stats.nServerConns, rx_stats_mutex);
902 rx_MutexDecrement(rx_stats.nClientConns, rx_stats_mutex);
905 if (conn->specific) {
907 for (i = 0; i < conn->nSpecific; i++) {
908 if (conn->specific[i] && rxi_keyCreate_destructor[i])
909 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
910 conn->specific[i] = NULL;
912 free(conn->specific);
914 conn->specific = NULL;
918 MUTEX_DESTROY(&conn->conn_call_lock);
919 MUTEX_DESTROY(&conn->conn_data_lock);
920 CV_DESTROY(&conn->conn_call_cv);
922 rxi_FreeConnection(conn);
925 /* Destroy the specified connection */
927 rxi_DestroyConnection(struct rx_connection *conn)
929 MUTEX_ENTER(&rx_connHashTable_lock);
930 rxi_DestroyConnectionNoLock(conn);
931 /* conn should be at the head of the cleanup list */
932 if (conn == rx_connCleanup_list) {
933 rx_connCleanup_list = rx_connCleanup_list->next;
934 MUTEX_EXIT(&rx_connHashTable_lock);
935 rxi_CleanupConnection(conn);
937 #ifdef RX_ENABLE_LOCKS
939 MUTEX_EXIT(&rx_connHashTable_lock);
941 #endif /* RX_ENABLE_LOCKS */
945 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
947 struct rx_connection **conn_ptr;
949 struct rx_packet *packet;
956 MUTEX_ENTER(&conn->conn_data_lock);
957 if (conn->refCount > 0)
960 if (rx_stats_active) {
961 MUTEX_ENTER(&rx_stats_mutex);
962 rxi_lowConnRefCount++;
963 MUTEX_EXIT(&rx_stats_mutex);
967 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
968 /* Busy; wait till the last guy before proceeding */
969 MUTEX_EXIT(&conn->conn_data_lock);
974 /* If the client previously called rx_NewCall, but it is still
975 * waiting, treat this as a running call, and wait to destroy the
976 * connection later when the call completes. */
977 if ((conn->type == RX_CLIENT_CONNECTION)
978 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
979 conn->flags |= RX_CONN_DESTROY_ME;
980 MUTEX_EXIT(&conn->conn_data_lock);
984 MUTEX_EXIT(&conn->conn_data_lock);
986 /* Check for extant references to this connection */
987 for (i = 0; i < RX_MAXCALLS; i++) {
988 struct rx_call *call = conn->call[i];
991 if (conn->type == RX_CLIENT_CONNECTION) {
992 MUTEX_ENTER(&call->lock);
993 if (call->delayedAckEvent) {
994 /* Push the final acknowledgment out now--there
995 * won't be a subsequent call to acknowledge the
996 * last reply packets */
997 rxevent_Cancel(call->delayedAckEvent, call,
998 RX_CALL_REFCOUNT_DELAY);
999 if (call->state == RX_STATE_PRECALL
1000 || call->state == RX_STATE_ACTIVE) {
1001 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1003 rxi_AckAll(NULL, call, 0);
1006 MUTEX_EXIT(&call->lock);
1010 #ifdef RX_ENABLE_LOCKS
1012 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1013 MUTEX_EXIT(&conn->conn_data_lock);
1015 /* Someone is accessing a packet right now. */
1019 #endif /* RX_ENABLE_LOCKS */
1022 /* Don't destroy the connection if there are any call
1023 * structures still in use */
1024 MUTEX_ENTER(&conn->conn_data_lock);
1025 conn->flags |= RX_CONN_DESTROY_ME;
1026 MUTEX_EXIT(&conn->conn_data_lock);
1031 if (conn->delayedAbortEvent) {
1032 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1033 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1035 MUTEX_ENTER(&conn->conn_data_lock);
1036 rxi_SendConnectionAbort(conn, packet, 0, 1);
1037 MUTEX_EXIT(&conn->conn_data_lock);
1038 rxi_FreePacket(packet);
1042 /* Remove from connection hash table before proceeding */
1044 &rx_connHashTable[CONN_HASH
1045 (peer->host, peer->port, conn->cid, conn->epoch,
1047 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1048 if (*conn_ptr == conn) {
1049 *conn_ptr = conn->next;
1053 /* if the conn that we are destroying was the last connection, then we
1054 * clear rxLastConn as well */
1055 if (rxLastConn == conn)
1058 /* Make sure the connection is completely reset before deleting it. */
1059 /* get rid of pending events that could zap us later */
1060 if (conn->challengeEvent)
1061 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1062 if (conn->checkReachEvent)
1063 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1065 /* Add the connection to the list of destroyed connections that
1066 * need to be cleaned up. This is necessary to avoid deadlocks
1067 * in the routines we call to inform others that this connection is
1068 * being destroyed. */
1069 conn->next = rx_connCleanup_list;
1070 rx_connCleanup_list = conn;
1073 /* Externally available version */
1075 rx_DestroyConnection(struct rx_connection *conn)
1080 rxi_DestroyConnection(conn);
1085 rx_GetConnection(struct rx_connection *conn)
1090 MUTEX_ENTER(&conn->conn_data_lock);
1092 MUTEX_EXIT(&conn->conn_data_lock);
1096 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1097 /* Wait for the transmit queue to no longer be busy.
1098 * requires the call->lock to be held */
1099 static void rxi_WaitforTQBusy(struct rx_call *call) {
1100 while (call->flags & RX_CALL_TQ_BUSY) {
1101 call->flags |= RX_CALL_TQ_WAIT;
1103 #ifdef RX_ENABLE_LOCKS
1104 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1105 CV_WAIT(&call->cv_tq, &call->lock);
1106 #else /* RX_ENABLE_LOCKS */
1107 osi_rxSleep(&call->tq);
1108 #endif /* RX_ENABLE_LOCKS */
1110 if (call->tqWaiters == 0) {
1111 call->flags &= ~RX_CALL_TQ_WAIT;
1117 /* Start a new rx remote procedure call, on the specified connection.
1118 * If wait is set to 1, wait for a free call channel; otherwise return
1119 * 0. Maxtime gives the maximum number of seconds this call may take,
1120 * after rx_NewCall returns. After this time interval, a call to any
1121 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1122 * For fine grain locking, we hold the conn_call_lock in order to
1123 * to ensure that we don't get signalle after we found a call in an active
1124 * state and before we go to sleep.
1127 rx_NewCall(struct rx_connection *conn)
1130 struct rx_call *call;
1131 struct clock queueTime;
1135 dpf(("rx_NewCall(conn %x)\n", conn));
1138 clock_GetTime(&queueTime);
1139 MUTEX_ENTER(&conn->conn_call_lock);
1142 * Check if there are others waiting for a new call.
1143 * If so, let them go first to avoid starving them.
1144 * This is a fairly simple scheme, and might not be
1145 * a complete solution for large numbers of waiters.
1147 * makeCallWaiters keeps track of the number of
1148 * threads waiting to make calls and the
1149 * RX_CONN_MAKECALL_WAITING flag bit is used to
1150 * indicate that there are indeed calls waiting.
1151 * The flag is set when the waiter is incremented.
1152 * It is only cleared in rx_EndCall when
1153 * makeCallWaiters is 0. This prevents us from
1154 * accidently destroying the connection while it
1155 * is potentially about to be used.
1157 MUTEX_ENTER(&conn->conn_data_lock);
1158 if (conn->makeCallWaiters) {
1159 conn->flags |= RX_CONN_MAKECALL_WAITING;
1160 conn->makeCallWaiters++;
1161 MUTEX_EXIT(&conn->conn_data_lock);
1163 #ifdef RX_ENABLE_LOCKS
1164 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1168 MUTEX_ENTER(&conn->conn_data_lock);
1169 conn->makeCallWaiters--;
1171 MUTEX_EXIT(&conn->conn_data_lock);
1174 for (i = 0; i < RX_MAXCALLS; i++) {
1175 call = conn->call[i];
1177 MUTEX_ENTER(&call->lock);
1178 if (call->state == RX_STATE_DALLY) {
1179 rxi_ResetCall(call, 0);
1180 (*call->callNumber)++;
1183 MUTEX_EXIT(&call->lock);
1185 call = rxi_NewCall(conn, i);
1189 if (i < RX_MAXCALLS) {
1192 MUTEX_ENTER(&conn->conn_data_lock);
1193 conn->flags |= RX_CONN_MAKECALL_WAITING;
1194 conn->makeCallWaiters++;
1195 MUTEX_EXIT(&conn->conn_data_lock);
1197 #ifdef RX_ENABLE_LOCKS
1198 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1202 MUTEX_ENTER(&conn->conn_data_lock);
1203 conn->makeCallWaiters--;
1204 MUTEX_EXIT(&conn->conn_data_lock);
1207 * Wake up anyone else who might be giving us a chance to
1208 * run (see code above that avoids resource starvation).
1210 #ifdef RX_ENABLE_LOCKS
1211 CV_BROADCAST(&conn->conn_call_cv);
1216 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1218 /* Client is initially in send mode */
1219 call->state = RX_STATE_ACTIVE;
1220 call->error = conn->error;
1222 call->mode = RX_MODE_ERROR;
1224 call->mode = RX_MODE_SENDING;
1226 /* remember start time for call in case we have hard dead time limit */
1227 call->queueTime = queueTime;
1228 clock_GetTime(&call->startTime);
1229 hzero(call->bytesSent);
1230 hzero(call->bytesRcvd);
1232 /* Turn on busy protocol. */
1233 rxi_KeepAliveOn(call);
1235 MUTEX_EXIT(&call->lock);
1236 MUTEX_EXIT(&conn->conn_call_lock);
1239 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1240 /* Now, if TQ wasn't cleared earlier, do it now. */
1241 MUTEX_ENTER(&call->lock);
1242 rxi_WaitforTQBusy(call);
1243 if (call->flags & RX_CALL_TQ_CLEARME) {
1244 rxi_ClearTransmitQueue(call, 1);
1245 /*queue_Init(&call->tq);*/
1247 MUTEX_EXIT(&call->lock);
1248 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1250 dpf(("rx_NewCall(call %x)\n", call));
1255 rxi_HasActiveCalls(struct rx_connection *aconn)
1258 struct rx_call *tcall;
1262 for (i = 0; i < RX_MAXCALLS; i++) {
1263 if ((tcall = aconn->call[i])) {
1264 if ((tcall->state == RX_STATE_ACTIVE)
1265 || (tcall->state == RX_STATE_PRECALL)) {
1276 rxi_GetCallNumberVector(struct rx_connection *aconn,
1277 afs_int32 * aint32s)
1280 struct rx_call *tcall;
1284 for (i = 0; i < RX_MAXCALLS; i++) {
1285 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1286 aint32s[i] = aconn->callNumber[i] + 1;
1288 aint32s[i] = aconn->callNumber[i];
1295 rxi_SetCallNumberVector(struct rx_connection *aconn,
1296 afs_int32 * aint32s)
1299 struct rx_call *tcall;
1303 for (i = 0; i < RX_MAXCALLS; i++) {
1304 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1305 aconn->callNumber[i] = aint32s[i] - 1;
1307 aconn->callNumber[i] = aint32s[i];
1313 /* Advertise a new service. A service is named locally by a UDP port
1314 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1317 char *serviceName; Name for identification purposes (e.g. the
1318 service name might be used for probing for
1321 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1322 char *serviceName, struct rx_securityClass **securityObjects,
1323 int nSecurityObjects,
1324 afs_int32(*serviceProc) (struct rx_call * acall))
1326 osi_socket socket = OSI_NULLSOCKET;
1327 struct rx_service *tservice;
1333 if (serviceId == 0) {
1335 "rx_NewService: service id for service %s is not non-zero.\n",
1342 "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",
1350 tservice = rxi_AllocService();
1352 for (i = 0; i < RX_MAX_SERVICES; i++) {
1353 struct rx_service *service = rx_services[i];
1355 if (port == service->servicePort && host == service->serviceHost) {
1356 if (service->serviceId == serviceId) {
1357 /* The identical service has already been
1358 * installed; if the caller was intending to
1359 * change the security classes used by this
1360 * service, he/she loses. */
1362 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1363 serviceName, serviceId, service->serviceName);
1365 rxi_FreeService(tservice);
1368 /* Different service, same port: re-use the socket
1369 * which is bound to the same port */
1370 socket = service->socket;
1373 if (socket == OSI_NULLSOCKET) {
1374 /* If we don't already have a socket (from another
1375 * service on same port) get a new one */
1376 socket = rxi_GetHostUDPSocket(host, port);
1377 if (socket == OSI_NULLSOCKET) {
1379 rxi_FreeService(tservice);
1384 service->socket = socket;
1385 service->serviceHost = host;
1386 service->servicePort = port;
1387 service->serviceId = serviceId;
1388 service->serviceName = serviceName;
1389 service->nSecurityObjects = nSecurityObjects;
1390 service->securityObjects = securityObjects;
1391 service->minProcs = 0;
1392 service->maxProcs = 1;
1393 service->idleDeadTime = 60;
1394 service->idleDeadErr = 0;
1395 service->connDeadTime = rx_connDeadTime;
1396 service->executeRequestProc = serviceProc;
1397 service->checkReach = 0;
1398 rx_services[i] = service; /* not visible until now */
1404 rxi_FreeService(tservice);
1405 (osi_Msg "rx_NewService: cannot support > %d services\n",
1410 /* Set configuration options for all of a service's security objects */
1413 rx_SetSecurityConfiguration(struct rx_service *service,
1414 rx_securityConfigVariables type,
1418 for (i = 0; i<service->nSecurityObjects; i++) {
1419 if (service->securityObjects[i]) {
1420 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1428 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1429 struct rx_securityClass **securityObjects, int nSecurityObjects,
1430 afs_int32(*serviceProc) (struct rx_call * acall))
1432 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1435 /* Generic request processing loop. This routine should be called
1436 * by the implementation dependent rx_ServerProc. If socketp is
1437 * non-null, it will be set to the file descriptor that this thread
1438 * is now listening on. If socketp is null, this routine will never
1441 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1443 struct rx_call *call;
1445 struct rx_service *tservice = NULL;
1452 call = rx_GetCall(threadID, tservice, socketp);
1453 if (socketp && *socketp != OSI_NULLSOCKET) {
1454 /* We are now a listener thread */
1459 /* if server is restarting( typically smooth shutdown) then do not
1460 * allow any new calls.
1463 if (rx_tranquil && (call != NULL)) {
1467 MUTEX_ENTER(&call->lock);
1469 rxi_CallError(call, RX_RESTARTING);
1470 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1472 MUTEX_EXIT(&call->lock);
1476 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1477 #ifdef RX_ENABLE_LOCKS
1479 #endif /* RX_ENABLE_LOCKS */
1480 afs_termState = AFSOP_STOP_AFS;
1481 afs_osi_Wakeup(&afs_termState);
1482 #ifdef RX_ENABLE_LOCKS
1484 #endif /* RX_ENABLE_LOCKS */
1489 tservice = call->conn->service;
1491 if (tservice->beforeProc)
1492 (*tservice->beforeProc) (call);
1494 code = call->conn->service->executeRequestProc(call);
1496 if (tservice->afterProc)
1497 (*tservice->afterProc) (call, code);
1499 rx_EndCall(call, code);
1500 if (rx_stats_active) {
1501 MUTEX_ENTER(&rx_stats_mutex);
1503 MUTEX_EXIT(&rx_stats_mutex);
1510 rx_WakeupServerProcs(void)
1512 struct rx_serverQueueEntry *np, *tqp;
1516 MUTEX_ENTER(&rx_serverPool_lock);
1518 #ifdef RX_ENABLE_LOCKS
1519 if (rx_waitForPacket)
1520 CV_BROADCAST(&rx_waitForPacket->cv);
1521 #else /* RX_ENABLE_LOCKS */
1522 if (rx_waitForPacket)
1523 osi_rxWakeup(rx_waitForPacket);
1524 #endif /* RX_ENABLE_LOCKS */
1525 MUTEX_ENTER(&freeSQEList_lock);
1526 for (np = rx_FreeSQEList; np; np = tqp) {
1527 tqp = *(struct rx_serverQueueEntry **)np;
1528 #ifdef RX_ENABLE_LOCKS
1529 CV_BROADCAST(&np->cv);
1530 #else /* RX_ENABLE_LOCKS */
1532 #endif /* RX_ENABLE_LOCKS */
1534 MUTEX_EXIT(&freeSQEList_lock);
1535 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1536 #ifdef RX_ENABLE_LOCKS
1537 CV_BROADCAST(&np->cv);
1538 #else /* RX_ENABLE_LOCKS */
1540 #endif /* RX_ENABLE_LOCKS */
1542 MUTEX_EXIT(&rx_serverPool_lock);
1547 * One thing that seems to happen is that all the server threads get
1548 * tied up on some empty or slow call, and then a whole bunch of calls
1549 * arrive at once, using up the packet pool, so now there are more
1550 * empty calls. The most critical resources here are server threads
1551 * and the free packet pool. The "doreclaim" code seems to help in
1552 * general. I think that eventually we arrive in this state: there
1553 * are lots of pending calls which do have all their packets present,
1554 * so they won't be reclaimed, are multi-packet calls, so they won't
1555 * be scheduled until later, and thus are tying up most of the free
1556 * packet pool for a very long time.
1558 * 1. schedule multi-packet calls if all the packets are present.
1559 * Probably CPU-bound operation, useful to return packets to pool.
1560 * Do what if there is a full window, but the last packet isn't here?
1561 * 3. preserve one thread which *only* runs "best" calls, otherwise
1562 * it sleeps and waits for that type of call.
1563 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1564 * the current dataquota business is badly broken. The quota isn't adjusted
1565 * to reflect how many packets are presently queued for a running call.
1566 * So, when we schedule a queued call with a full window of packets queued
1567 * up for it, that *should* free up a window full of packets for other 2d-class
1568 * calls to be able to use from the packet pool. But it doesn't.
1570 * NB. Most of the time, this code doesn't run -- since idle server threads
1571 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1572 * as a new call arrives.
1574 /* Sleep until a call arrives. Returns a pointer to the call, ready
1575 * for an rx_Read. */
1576 #ifdef RX_ENABLE_LOCKS
1578 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1580 struct rx_serverQueueEntry *sq;
1581 struct rx_call *call = (struct rx_call *)0;
1582 struct rx_service *service = NULL;
1585 MUTEX_ENTER(&freeSQEList_lock);
1587 if ((sq = rx_FreeSQEList)) {
1588 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1589 MUTEX_EXIT(&freeSQEList_lock);
1590 } else { /* otherwise allocate a new one and return that */
1591 MUTEX_EXIT(&freeSQEList_lock);
1592 sq = (struct rx_serverQueueEntry *)
1593 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1594 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1595 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1598 MUTEX_ENTER(&rx_serverPool_lock);
1599 if (cur_service != NULL) {
1600 ReturnToServerPool(cur_service);
1603 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1604 struct rx_call *tcall, *ncall, *choice2 = NULL;
1606 /* Scan for eligible incoming calls. A call is not eligible
1607 * if the maximum number of calls for its service type are
1608 * already executing */
1609 /* One thread will process calls FCFS (to prevent starvation),
1610 * while the other threads may run ahead looking for calls which
1611 * have all their input data available immediately. This helps
1612 * keep threads from blocking, waiting for data from the client. */
1613 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1614 service = tcall->conn->service;
1615 if (!QuotaOK(service)) {
1618 MUTEX_ENTER(&rx_pthread_mutex);
1619 if (tno == rxi_fcfs_thread_num
1620 || !tcall->queue_item_header.next) {
1621 MUTEX_EXIT(&rx_pthread_mutex);
1622 /* If we're the fcfs thread , then we'll just use
1623 * this call. If we haven't been able to find an optimal
1624 * choice, and we're at the end of the list, then use a
1625 * 2d choice if one has been identified. Otherwise... */
1626 call = (choice2 ? choice2 : tcall);
1627 service = call->conn->service;
1629 MUTEX_EXIT(&rx_pthread_mutex);
1630 if (!queue_IsEmpty(&tcall->rq)) {
1631 struct rx_packet *rp;
1632 rp = queue_First(&tcall->rq, rx_packet);
1633 if (rp->header.seq == 1) {
1635 || (rp->header.flags & RX_LAST_PACKET)) {
1637 } else if (rxi_2dchoice && !choice2
1638 && !(tcall->flags & RX_CALL_CLEARED)
1639 && (tcall->rprev > rxi_HardAckRate)) {
1649 ReturnToServerPool(service);
1656 MUTEX_EXIT(&rx_serverPool_lock);
1657 MUTEX_ENTER(&call->lock);
1659 if (call->flags & RX_CALL_WAIT_PROC) {
1660 call->flags &= ~RX_CALL_WAIT_PROC;
1661 MUTEX_ENTER(&rx_waiting_mutex);
1663 MUTEX_EXIT(&rx_waiting_mutex);
1666 if (call->state != RX_STATE_PRECALL || call->error) {
1667 MUTEX_EXIT(&call->lock);
1668 MUTEX_ENTER(&rx_serverPool_lock);
1669 ReturnToServerPool(service);
1674 if (queue_IsEmpty(&call->rq)
1675 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1676 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1678 CLEAR_CALL_QUEUE_LOCK(call);
1681 /* If there are no eligible incoming calls, add this process
1682 * to the idle server queue, to wait for one */
1686 *socketp = OSI_NULLSOCKET;
1688 sq->socketp = socketp;
1689 queue_Append(&rx_idleServerQueue, sq);
1690 #ifndef AFS_AIX41_ENV
1691 rx_waitForPacket = sq;
1693 rx_waitingForPacket = sq;
1694 #endif /* AFS_AIX41_ENV */
1696 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1698 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1699 MUTEX_EXIT(&rx_serverPool_lock);
1700 return (struct rx_call *)0;
1703 } while (!(call = sq->newcall)
1704 && !(socketp && *socketp != OSI_NULLSOCKET));
1705 MUTEX_EXIT(&rx_serverPool_lock);
1707 MUTEX_ENTER(&call->lock);
1713 MUTEX_ENTER(&freeSQEList_lock);
1714 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1715 rx_FreeSQEList = sq;
1716 MUTEX_EXIT(&freeSQEList_lock);
1719 clock_GetTime(&call->startTime);
1720 call->state = RX_STATE_ACTIVE;
1721 call->mode = RX_MODE_RECEIVING;
1722 #ifdef RX_KERNEL_TRACE
1723 if (ICL_SETACTIVE(afs_iclSetp)) {
1724 int glockOwner = ISAFS_GLOCK();
1727 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1728 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1735 rxi_calltrace(RX_CALL_START, call);
1736 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1737 call->conn->service->servicePort, call->conn->service->serviceId,
1740 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1741 MUTEX_EXIT(&call->lock);
1743 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1748 #else /* RX_ENABLE_LOCKS */
1750 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1752 struct rx_serverQueueEntry *sq;
1753 struct rx_call *call = (struct rx_call *)0, *choice2;
1754 struct rx_service *service = NULL;
1758 MUTEX_ENTER(&freeSQEList_lock);
1760 if ((sq = rx_FreeSQEList)) {
1761 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1762 MUTEX_EXIT(&freeSQEList_lock);
1763 } else { /* otherwise allocate a new one and return that */
1764 MUTEX_EXIT(&freeSQEList_lock);
1765 sq = (struct rx_serverQueueEntry *)
1766 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1767 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1768 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1770 MUTEX_ENTER(&sq->lock);
1772 if (cur_service != NULL) {
1773 cur_service->nRequestsRunning--;
1774 if (cur_service->nRequestsRunning < cur_service->minProcs)
1778 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1779 struct rx_call *tcall, *ncall;
1780 /* Scan for eligible incoming calls. A call is not eligible
1781 * if the maximum number of calls for its service type are
1782 * already executing */
1783 /* One thread will process calls FCFS (to prevent starvation),
1784 * while the other threads may run ahead looking for calls which
1785 * have all their input data available immediately. This helps
1786 * keep threads from blocking, waiting for data from the client. */
1787 choice2 = (struct rx_call *)0;
1788 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1789 service = tcall->conn->service;
1790 if (QuotaOK(service)) {
1791 MUTEX_ENTER(&rx_pthread_mutex);
1792 if (tno == rxi_fcfs_thread_num
1793 || !tcall->queue_item_header.next) {
1794 MUTEX_EXIT(&rx_pthread_mutex);
1795 /* If we're the fcfs thread, then we'll just use
1796 * this call. If we haven't been able to find an optimal
1797 * choice, and we're at the end of the list, then use a
1798 * 2d choice if one has been identified. Otherwise... */
1799 call = (choice2 ? choice2 : tcall);
1800 service = call->conn->service;
1802 MUTEX_EXIT(&rx_pthread_mutex);
1803 if (!queue_IsEmpty(&tcall->rq)) {
1804 struct rx_packet *rp;
1805 rp = queue_First(&tcall->rq, rx_packet);
1806 if (rp->header.seq == 1
1808 || (rp->header.flags & RX_LAST_PACKET))) {
1810 } else if (rxi_2dchoice && !choice2
1811 && !(tcall->flags & RX_CALL_CLEARED)
1812 && (tcall->rprev > rxi_HardAckRate)) {
1826 /* we can't schedule a call if there's no data!!! */
1827 /* send an ack if there's no data, if we're missing the
1828 * first packet, or we're missing something between first
1829 * and last -- there's a "hole" in the incoming data. */
1830 if (queue_IsEmpty(&call->rq)
1831 || queue_First(&call->rq, rx_packet)->header.seq != 1
1832 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1833 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1835 call->flags &= (~RX_CALL_WAIT_PROC);
1836 service->nRequestsRunning++;
1837 /* just started call in minProcs pool, need fewer to maintain
1839 if (service->nRequestsRunning <= service->minProcs)
1843 /* MUTEX_EXIT(&call->lock); */
1845 /* If there are no eligible incoming calls, add this process
1846 * to the idle server queue, to wait for one */
1849 *socketp = OSI_NULLSOCKET;
1851 sq->socketp = socketp;
1852 queue_Append(&rx_idleServerQueue, sq);
1856 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1858 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1859 return (struct rx_call *)0;
1862 } while (!(call = sq->newcall)
1863 && !(socketp && *socketp != OSI_NULLSOCKET));
1865 MUTEX_EXIT(&sq->lock);
1867 MUTEX_ENTER(&freeSQEList_lock);
1868 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1869 rx_FreeSQEList = sq;
1870 MUTEX_EXIT(&freeSQEList_lock);
1873 clock_GetTime(&call->startTime);
1874 call->state = RX_STATE_ACTIVE;
1875 call->mode = RX_MODE_RECEIVING;
1876 #ifdef RX_KERNEL_TRACE
1877 if (ICL_SETACTIVE(afs_iclSetp)) {
1878 int glockOwner = ISAFS_GLOCK();
1881 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1882 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1889 rxi_calltrace(RX_CALL_START, call);
1890 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1891 call->conn->service->servicePort, call->conn->service->serviceId,
1894 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1901 #endif /* RX_ENABLE_LOCKS */
1905 /* Establish a procedure to be called when a packet arrives for a
1906 * call. This routine will be called at most once after each call,
1907 * and will also be called if there is an error condition on the or
1908 * the call is complete. Used by multi rx to build a selection
1909 * function which determines which of several calls is likely to be a
1910 * good one to read from.
1911 * NOTE: the way this is currently implemented it is probably only a
1912 * good idea to (1) use it immediately after a newcall (clients only)
1913 * and (2) only use it once. Other uses currently void your warranty
1916 rx_SetArrivalProc(struct rx_call *call,
1917 void (*proc) (struct rx_call * call,
1920 void * handle, int arg)
1922 call->arrivalProc = proc;
1923 call->arrivalProcHandle = handle;
1924 call->arrivalProcArg = arg;
1927 /* Call is finished (possibly prematurely). Return rc to the peer, if
1928 * appropriate, and return the final error code from the conversation
1932 rx_EndCall(struct rx_call *call, afs_int32 rc)
1934 struct rx_connection *conn = call->conn;
1935 struct rx_service *service;
1941 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1944 MUTEX_ENTER(&call->lock);
1946 if (rc == 0 && call->error == 0) {
1947 call->abortCode = 0;
1948 call->abortCount = 0;
1951 call->arrivalProc = (void (*)())0;
1952 if (rc && call->error == 0) {
1953 rxi_CallError(call, rc);
1954 /* Send an abort message to the peer if this error code has
1955 * only just been set. If it was set previously, assume the
1956 * peer has already been sent the error code or will request it
1958 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1960 if (conn->type == RX_SERVER_CONNECTION) {
1961 /* Make sure reply or at least dummy reply is sent */
1962 if (call->mode == RX_MODE_RECEIVING) {
1963 rxi_WriteProc(call, 0, 0);
1965 if (call->mode == RX_MODE_SENDING) {
1966 rxi_FlushWrite(call);
1968 service = conn->service;
1969 rxi_calltrace(RX_CALL_END, call);
1970 /* Call goes to hold state until reply packets are acknowledged */
1971 if (call->tfirst + call->nSoftAcked < call->tnext) {
1972 call->state = RX_STATE_HOLD;
1974 call->state = RX_STATE_DALLY;
1975 rxi_ClearTransmitQueue(call, 0);
1976 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1977 rxevent_Cancel(call->keepAliveEvent, call,
1978 RX_CALL_REFCOUNT_ALIVE);
1980 } else { /* Client connection */
1982 /* Make sure server receives input packets, in the case where
1983 * no reply arguments are expected */
1984 if ((call->mode == RX_MODE_SENDING)
1985 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1986 (void)rxi_ReadProc(call, &dummy, 1);
1989 /* If we had an outstanding delayed ack, be nice to the server
1990 * and force-send it now.
1992 if (call->delayedAckEvent) {
1993 rxevent_Cancel(call->delayedAckEvent, call,
1994 RX_CALL_REFCOUNT_DELAY);
1995 call->delayedAckEvent = NULL;
1996 rxi_SendDelayedAck(NULL, call, NULL);
1999 /* We need to release the call lock since it's lower than the
2000 * conn_call_lock and we don't want to hold the conn_call_lock
2001 * over the rx_ReadProc call. The conn_call_lock needs to be held
2002 * here for the case where rx_NewCall is perusing the calls on
2003 * the connection structure. We don't want to signal until
2004 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2005 * have checked this call, found it active and by the time it
2006 * goes to sleep, will have missed the signal.
2008 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
2009 * there are threads waiting to use the conn object.
2011 MUTEX_EXIT(&call->lock);
2012 MUTEX_ENTER(&conn->conn_call_lock);
2013 MUTEX_ENTER(&call->lock);
2014 MUTEX_ENTER(&conn->conn_data_lock);
2015 conn->flags |= RX_CONN_BUSY;
2016 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2017 if (conn->makeCallWaiters == 0)
2018 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
2019 MUTEX_EXIT(&conn->conn_data_lock);
2020 #ifdef RX_ENABLE_LOCKS
2021 CV_BROADCAST(&conn->conn_call_cv);
2026 #ifdef RX_ENABLE_LOCKS
2028 MUTEX_EXIT(&conn->conn_data_lock);
2030 #endif /* RX_ENABLE_LOCKS */
2031 call->state = RX_STATE_DALLY;
2033 error = call->error;
2035 /* currentPacket, nLeft, and NFree must be zeroed here, because
2036 * ResetCall cannot: ResetCall may be called at splnet(), in the
2037 * kernel version, and may interrupt the macros rx_Read or
2038 * rx_Write, which run at normal priority for efficiency. */
2039 if (call->currentPacket) {
2040 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2041 rxi_FreePacket(call->currentPacket);
2042 call->currentPacket = (struct rx_packet *)0;
2045 call->nLeft = call->nFree = call->curlen = 0;
2047 /* Free any packets from the last call to ReadvProc/WritevProc */
2048 #ifdef RXDEBUG_PACKET
2050 #endif /* RXDEBUG_PACKET */
2051 rxi_FreePackets(0, &call->iovq);
2053 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2054 MUTEX_EXIT(&call->lock);
2055 if (conn->type == RX_CLIENT_CONNECTION) {
2056 MUTEX_EXIT(&conn->conn_call_lock);
2057 conn->flags &= ~RX_CONN_BUSY;
2061 * Map errors to the local host's errno.h format.
2063 error = ntoh_syserr_conv(error);
2067 #if !defined(KERNEL)
2069 /* Call this routine when shutting down a server or client (especially
2070 * clients). This will allow Rx to gracefully garbage collect server
2071 * connections, and reduce the number of retries that a server might
2072 * make to a dead client.
2073 * This is not quite right, since some calls may still be ongoing and
2074 * we can't lock them to destroy them. */
2078 struct rx_connection **conn_ptr, **conn_end;
2082 if (rxinit_status == 1) {
2084 return; /* Already shutdown. */
2086 rxi_DeleteCachedConnections();
2087 if (rx_connHashTable) {
2088 MUTEX_ENTER(&rx_connHashTable_lock);
2089 for (conn_ptr = &rx_connHashTable[0], conn_end =
2090 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2092 struct rx_connection *conn, *next;
2093 for (conn = *conn_ptr; conn; conn = next) {
2095 if (conn->type == RX_CLIENT_CONNECTION) {
2096 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2098 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2099 #ifdef RX_ENABLE_LOCKS
2100 rxi_DestroyConnectionNoLock(conn);
2101 #else /* RX_ENABLE_LOCKS */
2102 rxi_DestroyConnection(conn);
2103 #endif /* RX_ENABLE_LOCKS */
2107 #ifdef RX_ENABLE_LOCKS
2108 while (rx_connCleanup_list) {
2109 struct rx_connection *conn;
2110 conn = rx_connCleanup_list;
2111 rx_connCleanup_list = rx_connCleanup_list->next;
2112 MUTEX_EXIT(&rx_connHashTable_lock);
2113 rxi_CleanupConnection(conn);
2114 MUTEX_ENTER(&rx_connHashTable_lock);
2116 MUTEX_EXIT(&rx_connHashTable_lock);
2117 #endif /* RX_ENABLE_LOCKS */
2122 afs_winsockCleanup();
2130 /* if we wakeup packet waiter too often, can get in loop with two
2131 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2133 rxi_PacketsUnWait(void)
2135 if (!rx_waitingForPackets) {
2139 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2140 return; /* still over quota */
2143 rx_waitingForPackets = 0;
2144 #ifdef RX_ENABLE_LOCKS
2145 CV_BROADCAST(&rx_waitingForPackets_cv);
2147 osi_rxWakeup(&rx_waitingForPackets);
2153 /* ------------------Internal interfaces------------------------- */
2155 /* Return this process's service structure for the
2156 * specified socket and service */
2158 rxi_FindService(osi_socket socket, u_short serviceId)
2160 struct rx_service **sp;
2161 for (sp = &rx_services[0]; *sp; sp++) {
2162 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2168 #ifdef RXDEBUG_PACKET
2169 #ifdef KDUMP_RX_LOCK
2170 static struct rx_call_rx_lock *rx_allCallsp = 0;
2172 static struct rx_call *rx_allCallsp = 0;
2174 #endif /* RXDEBUG_PACKET */
2176 /* Allocate a call structure, for the indicated channel of the
2177 * supplied connection. The mode and state of the call must be set by
2178 * the caller. Returns the call with mutex locked. */
2180 rxi_NewCall(struct rx_connection *conn, int channel)
2182 struct rx_call *call;
2183 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2184 struct rx_call *cp; /* Call pointer temp */
2185 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2186 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2188 dpf(("rxi_NewCall(conn %x, channel %d)\n", conn, channel));
2190 /* Grab an existing call structure, or allocate a new one.
2191 * Existing call structures are assumed to have been left reset by
2193 MUTEX_ENTER(&rx_freeCallQueue_lock);
2195 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2197 * EXCEPT that the TQ might not yet be cleared out.
2198 * Skip over those with in-use TQs.
2201 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2202 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2208 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2209 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2210 call = queue_First(&rx_freeCallQueue, rx_call);
2211 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2213 if (rx_stats_active)
2214 rx_MutexDecrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2215 MUTEX_EXIT(&rx_freeCallQueue_lock);
2216 MUTEX_ENTER(&call->lock);
2217 CLEAR_CALL_QUEUE_LOCK(call);
2218 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2219 /* Now, if TQ wasn't cleared earlier, do it now. */
2220 if (call->flags & RX_CALL_TQ_CLEARME) {
2221 rxi_ClearTransmitQueue(call, 1);
2222 /*queue_Init(&call->tq);*/
2224 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2225 /* Bind the call to its connection structure */
2227 rxi_ResetCall(call, 1);
2230 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2231 #ifdef RXDEBUG_PACKET
2232 call->allNextp = rx_allCallsp;
2233 rx_allCallsp = call;
2235 #endif /* RXDEBUG_PACKET */
2236 rx_MutexIncrement(rx_stats.nCallStructs, rx_stats_mutex);
2238 MUTEX_EXIT(&rx_freeCallQueue_lock);
2239 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2240 MUTEX_ENTER(&call->lock);
2241 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2242 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2243 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2245 /* Initialize once-only items */
2246 queue_Init(&call->tq);
2247 queue_Init(&call->rq);
2248 queue_Init(&call->iovq);
2249 #ifdef RXDEBUG_PACKET
2250 call->rqc = call->tqc = call->iovqc = 0;
2251 #endif /* RXDEBUG_PACKET */
2252 /* Bind the call to its connection structure (prereq for reset) */
2254 rxi_ResetCall(call, 1);
2256 call->channel = channel;
2257 call->callNumber = &conn->callNumber[channel];
2258 call->rwind = conn->rwind[channel];
2259 call->twind = conn->twind[channel];
2260 /* Note that the next expected call number is retained (in
2261 * conn->callNumber[i]), even if we reallocate the call structure
2263 conn->call[channel] = call;
2264 /* if the channel's never been used (== 0), we should start at 1, otherwise
2265 * the call number is valid from the last time this channel was used */
2266 if (*call->callNumber == 0)
2267 *call->callNumber = 1;
2272 /* A call has been inactive long enough that so we can throw away
2273 * state, including the call structure, which is placed on the call
2275 * Call is locked upon entry.
2276 * haveCTLock set if called from rxi_ReapConnections
2278 #ifdef RX_ENABLE_LOCKS
2280 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2281 #else /* RX_ENABLE_LOCKS */
2283 rxi_FreeCall(struct rx_call *call)
2284 #endif /* RX_ENABLE_LOCKS */
2286 int channel = call->channel;
2287 struct rx_connection *conn = call->conn;
2290 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2291 (*call->callNumber)++;
2292 rxi_ResetCall(call, 0);
2293 call->conn->call[channel] = (struct rx_call *)0;
2295 MUTEX_ENTER(&rx_freeCallQueue_lock);
2296 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2297 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2298 /* A call may be free even though its transmit queue is still in use.
2299 * Since we search the call list from head to tail, put busy calls at
2300 * the head of the list, and idle calls at the tail.
2302 if (call->flags & RX_CALL_TQ_BUSY)
2303 queue_Prepend(&rx_freeCallQueue, call);
2305 queue_Append(&rx_freeCallQueue, call);
2306 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2307 queue_Append(&rx_freeCallQueue, call);
2308 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2309 if (rx_stats_active)
2310 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2311 MUTEX_EXIT(&rx_freeCallQueue_lock);
2313 /* Destroy the connection if it was previously slated for
2314 * destruction, i.e. the Rx client code previously called
2315 * rx_DestroyConnection (client connections), or
2316 * rxi_ReapConnections called the same routine (server
2317 * connections). Only do this, however, if there are no
2318 * outstanding calls. Note that for fine grain locking, there appears
2319 * to be a deadlock in that rxi_FreeCall has a call locked and
2320 * DestroyConnectionNoLock locks each call in the conn. But note a
2321 * few lines up where we have removed this call from the conn.
2322 * If someone else destroys a connection, they either have no
2323 * call lock held or are going through this section of code.
2325 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2326 MUTEX_ENTER(&conn->conn_data_lock);
2328 MUTEX_EXIT(&conn->conn_data_lock);
2329 #ifdef RX_ENABLE_LOCKS
2331 rxi_DestroyConnectionNoLock(conn);
2333 rxi_DestroyConnection(conn);
2334 #else /* RX_ENABLE_LOCKS */
2335 rxi_DestroyConnection(conn);
2336 #endif /* RX_ENABLE_LOCKS */
2340 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2342 rxi_Alloc(size_t size)
2346 if (rx_stats_active)
2347 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2350 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2351 afs_osi_Alloc_NoSleep(size);
2356 osi_Panic("rxi_Alloc error");
2362 rxi_Free(void *addr, size_t size)
2364 if (rx_stats_active)
2365 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2366 osi_Free(addr, size);
2370 rxi_SetPeerMtu(afs_uint32 host, afs_uint32 port, int mtu)
2372 struct rx_peer **peer_ptr, **peer_end;
2375 MUTEX_ENTER(&rx_peerHashTable_lock);
2377 for (peer_ptr = &rx_peerHashTable[0], peer_end =
2378 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
2380 struct rx_peer *peer, *next;
2381 for (peer = *peer_ptr; peer; peer = next) {
2383 if (host == peer->host) {
2384 MUTEX_ENTER(&peer->peer_lock);
2385 peer->ifMTU=MIN(mtu, peer->ifMTU);
2386 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2387 MUTEX_EXIT(&peer->peer_lock);
2392 struct rx_peer *peer;
2393 hashIndex = PEER_HASH(host, port);
2394 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2395 if ((peer->host == host) && (peer->port == port)) {
2396 MUTEX_ENTER(&peer->peer_lock);
2397 peer->ifMTU=MIN(mtu, peer->ifMTU);
2398 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2399 MUTEX_EXIT(&peer->peer_lock);
2403 MUTEX_EXIT(&rx_peerHashTable_lock);
2406 /* Find the peer process represented by the supplied (host,port)
2407 * combination. If there is no appropriate active peer structure, a
2408 * new one will be allocated and initialized
2409 * The origPeer, if set, is a pointer to a peer structure on which the
2410 * refcount will be be decremented. This is used to replace the peer
2411 * structure hanging off a connection structure */
2413 rxi_FindPeer(afs_uint32 host, u_short port,
2414 struct rx_peer *origPeer, int create)
2418 hashIndex = PEER_HASH(host, port);
2419 MUTEX_ENTER(&rx_peerHashTable_lock);
2420 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2421 if ((pp->host == host) && (pp->port == port))
2426 pp = rxi_AllocPeer(); /* This bzero's *pp */
2427 pp->host = host; /* set here or in InitPeerParams is zero */
2429 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2430 queue_Init(&pp->congestionQueue);
2431 queue_Init(&pp->rpcStats);
2432 pp->next = rx_peerHashTable[hashIndex];
2433 rx_peerHashTable[hashIndex] = pp;
2434 rxi_InitPeerParams(pp);
2435 if (rx_stats_active)
2436 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2443 origPeer->refCount--;
2444 MUTEX_EXIT(&rx_peerHashTable_lock);
2449 /* Find the connection at (host, port) started at epoch, and with the
2450 * given connection id. Creates the server connection if necessary.
2451 * The type specifies whether a client connection or a server
2452 * connection is desired. In both cases, (host, port) specify the
2453 * peer's (host, pair) pair. Client connections are not made
2454 * automatically by this routine. The parameter socket gives the
2455 * socket descriptor on which the packet was received. This is used,
2456 * in the case of server connections, to check that *new* connections
2457 * come via a valid (port, serviceId). Finally, the securityIndex
2458 * parameter must match the existing index for the connection. If a
2459 * server connection is created, it will be created using the supplied
2460 * index, if the index is valid for this service */
2461 struct rx_connection *
2462 rxi_FindConnection(osi_socket socket, afs_int32 host,
2463 u_short port, u_short serviceId, afs_uint32 cid,
2464 afs_uint32 epoch, int type, u_int securityIndex)
2466 int hashindex, flag, i;
2467 struct rx_connection *conn;
2468 hashindex = CONN_HASH(host, port, cid, epoch, type);
2469 MUTEX_ENTER(&rx_connHashTable_lock);
2470 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2471 rx_connHashTable[hashindex],
2474 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2475 && (epoch == conn->epoch)) {
2476 struct rx_peer *pp = conn->peer;
2477 if (securityIndex != conn->securityIndex) {
2478 /* this isn't supposed to happen, but someone could forge a packet
2479 * like this, and there seems to be some CM bug that makes this
2480 * happen from time to time -- in which case, the fileserver
2482 MUTEX_EXIT(&rx_connHashTable_lock);
2483 return (struct rx_connection *)0;
2485 if (pp->host == host && pp->port == port)
2487 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2489 /* So what happens when it's a callback connection? */
2490 if ( /*type == RX_CLIENT_CONNECTION && */
2491 (conn->epoch & 0x80000000))
2495 /* the connection rxLastConn that was used the last time is not the
2496 ** one we are looking for now. Hence, start searching in the hash */
2498 conn = rx_connHashTable[hashindex];
2503 struct rx_service *service;
2504 if (type == RX_CLIENT_CONNECTION) {
2505 MUTEX_EXIT(&rx_connHashTable_lock);
2506 return (struct rx_connection *)0;
2508 service = rxi_FindService(socket, serviceId);
2509 if (!service || (securityIndex >= service->nSecurityObjects)
2510 || (service->securityObjects[securityIndex] == 0)) {
2511 MUTEX_EXIT(&rx_connHashTable_lock);
2512 return (struct rx_connection *)0;
2514 conn = rxi_AllocConnection(); /* This bzero's the connection */
2515 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2516 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2517 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2518 conn->next = rx_connHashTable[hashindex];
2519 rx_connHashTable[hashindex] = conn;
2520 conn->peer = rxi_FindPeer(host, port, 0, 1);
2521 conn->type = RX_SERVER_CONNECTION;
2522 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2523 conn->epoch = epoch;
2524 conn->cid = cid & RX_CIDMASK;
2525 /* conn->serial = conn->lastSerial = 0; */
2526 /* conn->timeout = 0; */
2527 conn->ackRate = RX_FAST_ACK_RATE;
2528 conn->service = service;
2529 conn->serviceId = serviceId;
2530 conn->securityIndex = securityIndex;
2531 conn->securityObject = service->securityObjects[securityIndex];
2532 conn->nSpecific = 0;
2533 conn->specific = NULL;
2534 rx_SetConnDeadTime(conn, service->connDeadTime);
2535 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2536 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2537 for (i = 0; i < RX_MAXCALLS; i++) {
2538 conn->twind[i] = rx_initSendWindow;
2539 conn->rwind[i] = rx_initReceiveWindow;
2541 /* Notify security object of the new connection */
2542 RXS_NewConnection(conn->securityObject, conn);
2543 /* XXXX Connection timeout? */
2544 if (service->newConnProc)
2545 (*service->newConnProc) (conn);
2546 if (rx_stats_active)
2547 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2550 MUTEX_ENTER(&conn->conn_data_lock);
2552 MUTEX_EXIT(&conn->conn_data_lock);
2554 rxLastConn = conn; /* store this connection as the last conn used */
2555 MUTEX_EXIT(&rx_connHashTable_lock);
2559 /* There are two packet tracing routines available for testing and monitoring
2560 * Rx. One is called just after every packet is received and the other is
2561 * called just before every packet is sent. Received packets, have had their
2562 * headers decoded, and packets to be sent have not yet had their headers
2563 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2564 * containing the network address. Both can be modified. The return value, if
2565 * non-zero, indicates that the packet should be dropped. */
2567 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2568 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2570 /* A packet has been received off the interface. Np is the packet, socket is
2571 * the socket number it was received from (useful in determining which service
2572 * this packet corresponds to), and (host, port) reflect the host,port of the
2573 * sender. This call returns the packet to the caller if it is finished with
2574 * it, rather than de-allocating it, just as a small performance hack */
2577 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
2578 afs_uint32 host, u_short port, int *tnop,
2579 struct rx_call **newcallp)
2581 struct rx_call *call;
2582 struct rx_connection *conn;
2584 afs_uint32 currentCallNumber;
2590 struct rx_packet *tnp;
2593 /* We don't print out the packet until now because (1) the time may not be
2594 * accurate enough until now in the lwp implementation (rx_Listener only gets
2595 * the time after the packet is read) and (2) from a protocol point of view,
2596 * this is the first time the packet has been seen */
2597 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2598 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2599 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2600 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2601 np->header.epoch, np->header.cid, np->header.callNumber,
2602 np->header.seq, np->header.flags, np));
2605 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2606 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2609 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2610 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2613 /* If an input tracer function is defined, call it with the packet and
2614 * network address. Note this function may modify its arguments. */
2615 if (rx_justReceived) {
2616 struct sockaddr_in addr;
2618 addr.sin_family = AF_INET;
2619 addr.sin_port = port;
2620 addr.sin_addr.s_addr = host;
2621 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2622 addr.sin_len = sizeof(addr);
2623 #endif /* AFS_OSF_ENV */
2624 drop = (*rx_justReceived) (np, &addr);
2625 /* drop packet if return value is non-zero */
2628 port = addr.sin_port; /* in case fcn changed addr */
2629 host = addr.sin_addr.s_addr;
2633 /* If packet was not sent by the client, then *we* must be the client */
2634 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2635 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2637 /* Find the connection (or fabricate one, if we're the server & if
2638 * necessary) associated with this packet */
2640 rxi_FindConnection(socket, host, port, np->header.serviceId,
2641 np->header.cid, np->header.epoch, type,
2642 np->header.securityIndex);
2645 /* If no connection found or fabricated, just ignore the packet.
2646 * (An argument could be made for sending an abort packet for
2651 MUTEX_ENTER(&conn->conn_data_lock);
2652 if (conn->maxSerial < np->header.serial)
2653 conn->maxSerial = np->header.serial;
2654 MUTEX_EXIT(&conn->conn_data_lock);
2656 /* If the connection is in an error state, send an abort packet and ignore
2657 * the incoming packet */
2659 /* Don't respond to an abort packet--we don't want loops! */
2660 MUTEX_ENTER(&conn->conn_data_lock);
2661 if (np->header.type != RX_PACKET_TYPE_ABORT)
2662 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2664 MUTEX_EXIT(&conn->conn_data_lock);
2668 /* Check for connection-only requests (i.e. not call specific). */
2669 if (np->header.callNumber == 0) {
2670 switch (np->header.type) {
2671 case RX_PACKET_TYPE_ABORT: {
2672 /* What if the supplied error is zero? */
2673 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2674 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2675 rxi_ConnectionError(conn, errcode);
2676 MUTEX_ENTER(&conn->conn_data_lock);
2678 MUTEX_EXIT(&conn->conn_data_lock);
2681 case RX_PACKET_TYPE_CHALLENGE:
2682 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2683 MUTEX_ENTER(&conn->conn_data_lock);
2685 MUTEX_EXIT(&conn->conn_data_lock);
2687 case RX_PACKET_TYPE_RESPONSE:
2688 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2689 MUTEX_ENTER(&conn->conn_data_lock);
2691 MUTEX_EXIT(&conn->conn_data_lock);
2693 case RX_PACKET_TYPE_PARAMS:
2694 case RX_PACKET_TYPE_PARAMS + 1:
2695 case RX_PACKET_TYPE_PARAMS + 2:
2696 /* ignore these packet types for now */
2697 MUTEX_ENTER(&conn->conn_data_lock);
2699 MUTEX_EXIT(&conn->conn_data_lock);
2704 /* Should not reach here, unless the peer is broken: send an
2706 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2707 MUTEX_ENTER(&conn->conn_data_lock);
2708 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2710 MUTEX_EXIT(&conn->conn_data_lock);
2715 channel = np->header.cid & RX_CHANNELMASK;
2716 call = conn->call[channel];
2717 #ifdef RX_ENABLE_LOCKS
2719 MUTEX_ENTER(&call->lock);
2720 /* Test to see if call struct is still attached to conn. */
2721 if (call != conn->call[channel]) {
2723 MUTEX_EXIT(&call->lock);
2724 if (type == RX_SERVER_CONNECTION) {
2725 call = conn->call[channel];
2726 /* If we started with no call attached and there is one now,
2727 * another thread is also running this routine and has gotten
2728 * the connection channel. We should drop this packet in the tests
2729 * below. If there was a call on this connection and it's now
2730 * gone, then we'll be making a new call below.
2731 * If there was previously a call and it's now different then
2732 * the old call was freed and another thread running this routine
2733 * has created a call on this channel. One of these two threads
2734 * has a packet for the old call and the code below handles those
2738 MUTEX_ENTER(&call->lock);
2740 /* This packet can't be for this call. If the new call address is
2741 * 0 then no call is running on this channel. If there is a call
2742 * then, since this is a client connection we're getting data for
2743 * it must be for the previous call.
2745 if (rx_stats_active)
2746 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2747 MUTEX_ENTER(&conn->conn_data_lock);
2749 MUTEX_EXIT(&conn->conn_data_lock);
2754 currentCallNumber = conn->callNumber[channel];
2756 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2757 if (np->header.callNumber < currentCallNumber) {
2758 if (rx_stats_active)
2759 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2760 #ifdef RX_ENABLE_LOCKS
2762 MUTEX_EXIT(&call->lock);
2764 MUTEX_ENTER(&conn->conn_data_lock);
2766 MUTEX_EXIT(&conn->conn_data_lock);
2770 MUTEX_ENTER(&conn->conn_call_lock);
2771 call = rxi_NewCall(conn, channel);
2772 MUTEX_EXIT(&conn->conn_call_lock);
2773 *call->callNumber = np->header.callNumber;
2775 if (np->header.callNumber == 0)
2776 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %lx resend %d.%0.3d len %d", np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port), np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq, np->header.flags, (unsigned long)np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2778 call->state = RX_STATE_PRECALL;
2779 clock_GetTime(&call->queueTime);
2780 hzero(call->bytesSent);
2781 hzero(call->bytesRcvd);
2783 * If the number of queued calls exceeds the overload
2784 * threshold then abort this call.
2786 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2787 struct rx_packet *tp;
2789 rxi_CallError(call, rx_BusyError);
2790 tp = rxi_SendCallAbort(call, np, 1, 0);
2791 MUTEX_EXIT(&call->lock);
2792 MUTEX_ENTER(&conn->conn_data_lock);
2794 MUTEX_EXIT(&conn->conn_data_lock);
2795 if (rx_stats_active)
2796 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2799 rxi_KeepAliveOn(call);
2800 } else if (np->header.callNumber != currentCallNumber) {
2801 /* Wait until the transmit queue is idle before deciding
2802 * whether to reset the current call. Chances are that the
2803 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2806 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2807 while ((call->state == RX_STATE_ACTIVE)
2808 && (call->flags & RX_CALL_TQ_BUSY)) {
2809 call->flags |= RX_CALL_TQ_WAIT;
2811 #ifdef RX_ENABLE_LOCKS
2812 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2813 CV_WAIT(&call->cv_tq, &call->lock);
2814 #else /* RX_ENABLE_LOCKS */
2815 osi_rxSleep(&call->tq);
2816 #endif /* RX_ENABLE_LOCKS */
2818 if (call->tqWaiters == 0)
2819 call->flags &= ~RX_CALL_TQ_WAIT;
2821 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2822 /* If the new call cannot be taken right now send a busy and set
2823 * the error condition in this call, so that it terminates as
2824 * quickly as possible */
2825 if (call->state == RX_STATE_ACTIVE) {
2826 struct rx_packet *tp;
2828 rxi_CallError(call, RX_CALL_DEAD);
2829 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2831 MUTEX_EXIT(&call->lock);
2832 MUTEX_ENTER(&conn->conn_data_lock);
2834 MUTEX_EXIT(&conn->conn_data_lock);
2837 rxi_ResetCall(call, 0);
2838 *call->callNumber = np->header.callNumber;
2840 if (np->header.callNumber == 0)
2841 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %lx resend %d.%06d len %d", np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port), np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq, np->header.flags, (unsigned long)np, np->retryTime.sec, np->retryTime.usec, np->length));
2843 call->state = RX_STATE_PRECALL;
2844 clock_GetTime(&call->queueTime);
2845 hzero(call->bytesSent);
2846 hzero(call->bytesRcvd);
2848 * If the number of queued calls exceeds the overload
2849 * threshold then abort this call.
2851 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2852 struct rx_packet *tp;
2854 rxi_CallError(call, rx_BusyError);
2855 tp = rxi_SendCallAbort(call, np, 1, 0);
2856 MUTEX_EXIT(&call->lock);
2857 MUTEX_ENTER(&conn->conn_data_lock);
2859 MUTEX_EXIT(&conn->conn_data_lock);
2860 if (rx_stats_active)
2861 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2864 rxi_KeepAliveOn(call);
2866 /* Continuing call; do nothing here. */
2868 } else { /* we're the client */
2869 /* Ignore all incoming acknowledgements for calls in DALLY state */
2870 if (call && (call->state == RX_STATE_DALLY)
2871 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2872 if (rx_stats_active)
2873 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
2874 #ifdef RX_ENABLE_LOCKS
2876 MUTEX_EXIT(&call->lock);
2879 MUTEX_ENTER(&conn->conn_data_lock);
2881 MUTEX_EXIT(&conn->conn_data_lock);
2885 /* Ignore anything that's not relevant to the current call. If there
2886 * isn't a current call, then no packet is relevant. */
2887 if (!call || (np->header.callNumber != currentCallNumber)) {
2888 if (rx_stats_active)
2889 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2890 #ifdef RX_ENABLE_LOCKS
2892 MUTEX_EXIT(&call->lock);
2895 MUTEX_ENTER(&conn->conn_data_lock);
2897 MUTEX_EXIT(&conn->conn_data_lock);
2900 /* If the service security object index stamped in the packet does not
2901 * match the connection's security index, ignore the packet */
2902 if (np->header.securityIndex != conn->securityIndex) {
2903 #ifdef RX_ENABLE_LOCKS
2904 MUTEX_EXIT(&call->lock);
2906 MUTEX_ENTER(&conn->conn_data_lock);
2908 MUTEX_EXIT(&conn->conn_data_lock);
2912 /* If we're receiving the response, then all transmit packets are
2913 * implicitly acknowledged. Get rid of them. */
2914 if (np->header.type == RX_PACKET_TYPE_DATA) {
2915 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2916 /* XXX Hack. Because we must release the global rx lock when
2917 * sending packets (osi_NetSend) we drop all acks while we're
2918 * traversing the tq in rxi_Start sending packets out because
2919 * packets may move to the freePacketQueue as result of being here!
2920 * So we drop these packets until we're safely out of the
2921 * traversing. Really ugly!
2922 * For fine grain RX locking, we set the acked field in the
2923 * packets and let rxi_Start remove them from the transmit queue.
2925 if (call->flags & RX_CALL_TQ_BUSY) {
2926 #ifdef RX_ENABLE_LOCKS
2927 rxi_SetAcksInTransmitQueue(call);
2930 return np; /* xmitting; drop packet */
2933 rxi_ClearTransmitQueue(call, 0);
2935 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2936 rxi_ClearTransmitQueue(call, 0);
2937 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2939 if (np->header.type == RX_PACKET_TYPE_ACK) {
2940 /* now check to see if this is an ack packet acknowledging that the
2941 * server actually *lost* some hard-acked data. If this happens we
2942 * ignore this packet, as it may indicate that the server restarted in
2943 * the middle of a call. It is also possible that this is an old ack
2944 * packet. We don't abort the connection in this case, because this
2945 * *might* just be an old ack packet. The right way to detect a server
2946 * restart in the midst of a call is to notice that the server epoch
2948 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2949 * XXX unacknowledged. I think that this is off-by-one, but
2950 * XXX I don't dare change it just yet, since it will
2951 * XXX interact badly with the server-restart detection
2952 * XXX code in receiveackpacket. */
2953 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2954 if (rx_stats_active)
2955 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2956 MUTEX_EXIT(&call->lock);
2957 MUTEX_ENTER(&conn->conn_data_lock);
2959 MUTEX_EXIT(&conn->conn_data_lock);
2963 } /* else not a data packet */
2966 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2967 /* Set remote user defined status from packet */
2968 call->remoteStatus = np->header.userStatus;
2970 /* Note the gap between the expected next packet and the actual
2971 * packet that arrived, when the new packet has a smaller serial number
2972 * than expected. Rioses frequently reorder packets all by themselves,
2973 * so this will be quite important with very large window sizes.
2974 * Skew is checked against 0 here to avoid any dependence on the type of
2975 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2977 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2978 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2979 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2981 MUTEX_ENTER(&conn->conn_data_lock);
2982 skew = conn->lastSerial - np->header.serial;
2983 conn->lastSerial = np->header.serial;
2984 MUTEX_EXIT(&conn->conn_data_lock);
2986 struct rx_peer *peer;
2988 if (skew > peer->inPacketSkew) {
2989 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2991 peer->inPacketSkew = skew;
2995 /* Now do packet type-specific processing */
2996 switch (np->header.type) {
2997 case RX_PACKET_TYPE_DATA:
2998 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3001 case RX_PACKET_TYPE_ACK:
3002 /* Respond immediately to ack packets requesting acknowledgement
3004 if (np->header.flags & RX_REQUEST_ACK) {
3006 (void)rxi_SendCallAbort(call, 0, 1, 0);
3008 (void)rxi_SendAck(call, 0, np->header.serial,
3009 RX_ACK_PING_RESPONSE, 1);
3011 np = rxi_ReceiveAckPacket(call, np, 1);
3013 case RX_PACKET_TYPE_ABORT: {
3014 /* An abort packet: reset the call, passing the error up to the user. */
3015 /* What if error is zero? */
3016 /* What if the error is -1? the application will treat it as a timeout. */
3017 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3018 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
3019 rxi_CallError(call, errdata);
3020 MUTEX_EXIT(&call->lock);
3021 MUTEX_ENTER(&conn->conn_data_lock);
3023 MUTEX_EXIT(&conn->conn_data_lock);
3024 return np; /* xmitting; drop packet */
3026 case RX_PACKET_TYPE_BUSY:
3029 case RX_PACKET_TYPE_ACKALL:
3030 /* All packets acknowledged, so we can drop all packets previously
3031 * readied for sending */
3032 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3033 /* XXX Hack. We because we can't release the global rx lock when
3034 * sending packets (osi_NetSend) we drop all ack pkts while we're
3035 * traversing the tq in rxi_Start sending packets out because
3036 * packets may move to the freePacketQueue as result of being
3037 * here! So we drop these packets until we're safely out of the
3038 * traversing. Really ugly!
3039 * For fine grain RX locking, we set the acked field in the packets
3040 * and let rxi_Start remove the packets from the transmit queue.
3042 if (call->flags & RX_CALL_TQ_BUSY) {
3043 #ifdef RX_ENABLE_LOCKS
3044 rxi_SetAcksInTransmitQueue(call);
3046 #else /* RX_ENABLE_LOCKS */
3047 MUTEX_EXIT(&call->lock);
3048 MUTEX_ENTER(&conn->conn_data_lock);
3050 MUTEX_EXIT(&conn->conn_data_lock);
3051 return np; /* xmitting; drop packet */
3052 #endif /* RX_ENABLE_LOCKS */
3054 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3055 rxi_ClearTransmitQueue(call, 0);
3056 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3059 /* Should not reach here, unless the peer is broken: send an abort
3061 rxi_CallError(call, RX_PROTOCOL_ERROR);
3062 np = rxi_SendCallAbort(call, np, 1, 0);
3065 /* Note when this last legitimate packet was received, for keep-alive
3066 * processing. Note, we delay getting the time until now in the hope that
3067 * the packet will be delivered to the user before any get time is required
3068 * (if not, then the time won't actually be re-evaluated here). */
3069 call->lastReceiveTime = clock_Sec();
3070 MUTEX_EXIT(&call->lock);
3071 MUTEX_ENTER(&conn->conn_data_lock);
3073 MUTEX_EXIT(&conn->conn_data_lock);
3077 /* return true if this is an "interesting" connection from the point of view
3078 of someone trying to debug the system */
3080 rxi_IsConnInteresting(struct rx_connection *aconn)
3083 struct rx_call *tcall;
3085 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3087 for (i = 0; i < RX_MAXCALLS; i++) {
3088 tcall = aconn->call[i];
3090 if ((tcall->state == RX_STATE_PRECALL)
3091 || (tcall->state == RX_STATE_ACTIVE))
3093 if ((tcall->mode == RX_MODE_SENDING)
3094 || (tcall->mode == RX_MODE_RECEIVING))
3102 /* if this is one of the last few packets AND it wouldn't be used by the
3103 receiving call to immediately satisfy a read request, then drop it on
3104 the floor, since accepting it might prevent a lock-holding thread from
3105 making progress in its reading. If a call has been cleared while in
3106 the precall state then ignore all subsequent packets until the call
3107 is assigned to a thread. */
3110 TooLow(struct rx_packet *ap, struct rx_call *acall)
3114 MUTEX_ENTER(&rx_quota_mutex);
3115 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3116 && (acall->state == RX_STATE_PRECALL))
3117 || ((rx_nFreePackets < rxi_dataQuota + 2)
3118 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3119 && (acall->flags & RX_CALL_READER_WAIT)))) {
3122 MUTEX_EXIT(&rx_quota_mutex);
3128 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3130 struct rx_connection *conn = arg1;
3131 struct rx_call *acall = arg2;
3132 struct rx_call *call = acall;
3133 struct clock when, now;
3136 MUTEX_ENTER(&conn->conn_data_lock);
3137 conn->checkReachEvent = NULL;
3138 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3141 MUTEX_EXIT(&conn->conn_data_lock);
3145 MUTEX_ENTER(&conn->conn_call_lock);
3146 MUTEX_ENTER(&conn->conn_data_lock);
3147 for (i = 0; i < RX_MAXCALLS; i++) {
3148 struct rx_call *tc = conn->call[i];
3149 if (tc && tc->state == RX_STATE_PRECALL) {
3155 /* Indicate that rxi_CheckReachEvent is no longer running by
3156 * clearing the flag. Must be atomic under conn_data_lock to
3157 * avoid a new call slipping by: rxi_CheckConnReach holds
3158 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3160 conn->flags &= ~RX_CONN_ATTACHWAIT;
3161 MUTEX_EXIT(&conn->conn_data_lock);
3162 MUTEX_EXIT(&conn->conn_call_lock);
3167 MUTEX_ENTER(&call->lock);
3168 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3170 MUTEX_EXIT(&call->lock);
3172 clock_GetTime(&now);
3174 when.sec += RX_CHECKREACH_TIMEOUT;
3175 MUTEX_ENTER(&conn->conn_data_lock);
3176 if (!conn->checkReachEvent) {
3178 conn->checkReachEvent =
3179 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3182 MUTEX_EXIT(&conn->conn_data_lock);
3188 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3190 struct rx_service *service = conn->service;
3191 struct rx_peer *peer = conn->peer;
3192 afs_uint32 now, lastReach;
3194 if (service->checkReach == 0)
3198 MUTEX_ENTER(&peer->peer_lock);
3199 lastReach = peer->lastReachTime;
3200 MUTEX_EXIT(&peer->peer_lock);
3201 if (now - lastReach < RX_CHECKREACH_TTL)
3204 MUTEX_ENTER(&conn->conn_data_lock);
3205 if (conn->flags & RX_CONN_ATTACHWAIT) {
3206 MUTEX_EXIT(&conn->conn_data_lock);
3209 conn->flags |= RX_CONN_ATTACHWAIT;
3210 MUTEX_EXIT(&conn->conn_data_lock);
3211 if (!conn->checkReachEvent)
3212 rxi_CheckReachEvent(NULL, conn, call);
3217 /* try to attach call, if authentication is complete */
3219 TryAttach(struct rx_call *acall, osi_socket socket,
3220 int *tnop, struct rx_call **newcallp,
3223 struct rx_connection *conn = acall->conn;
3225 if (conn->type == RX_SERVER_CONNECTION
3226 && acall->state == RX_STATE_PRECALL) {
3227 /* Don't attach until we have any req'd. authentication. */
3228 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3229 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3230 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3231 /* Note: this does not necessarily succeed; there
3232 * may not any proc available
3235 rxi_ChallengeOn(acall->conn);
3240 /* A data packet has been received off the interface. This packet is
3241 * appropriate to the call (the call is in the right state, etc.). This
3242 * routine can return a packet to the caller, for re-use */
3245 rxi_ReceiveDataPacket(struct rx_call *call,
3246 struct rx_packet *np, int istack,
3247 osi_socket socket, afs_uint32 host, u_short port,
3248 int *tnop, struct rx_call **newcallp)
3250 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3255 afs_uint32 serial=0, flags=0;
3257 struct rx_packet *tnp;
3258 struct clock when, now;
3259 if (rx_stats_active)
3260 rx_MutexIncrement(rx_stats.dataPacketsRead, rx_stats_mutex);
3263 /* If there are no packet buffers, drop this new packet, unless we can find
3264 * packet buffers from inactive calls */
3266 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3267 MUTEX_ENTER(&rx_freePktQ_lock);
3268 rxi_NeedMorePackets = TRUE;
3269 MUTEX_EXIT(&rx_freePktQ_lock);
3270 if (rx_stats_active)
3271 rx_MutexIncrement(rx_stats.noPacketBuffersOnRead, rx_stats_mutex);
3272 call->rprev = np->header.serial;
3273 rxi_calltrace(RX_TRACE_DROP, call);
3274 dpf(("packet %x dropped on receipt - quota problems", np));
3276 rxi_ClearReceiveQueue(call);
3277 clock_GetTime(&now);
3279 clock_Add(&when, &rx_softAckDelay);
3280 if (!call->delayedAckEvent
3281 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3282 rxevent_Cancel(call->delayedAckEvent, call,
3283 RX_CALL_REFCOUNT_DELAY);
3284 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3285 call->delayedAckEvent =
3286 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3288 /* we've damaged this call already, might as well do it in. */
3294 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3295 * packet is one of several packets transmitted as a single
3296 * datagram. Do not send any soft or hard acks until all packets
3297 * in a jumbogram have been processed. Send negative acks right away.
3299 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3300 /* tnp is non-null when there are more packets in the
3301 * current jumbo gram */
3308 seq = np->header.seq;
3309 serial = np->header.serial;
3310 flags = np->header.flags;
3312 /* If the call is in an error state, send an abort message */
3314 return rxi_SendCallAbort(call, np, istack, 0);
3316 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3317 * AFS 3.5 jumbogram. */
3318 if (flags & RX_JUMBO_PACKET) {
3319 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3324 if (np->header.spare != 0) {
3325 MUTEX_ENTER(&call->conn->conn_data_lock);
3326 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3327 MUTEX_EXIT(&call->conn->conn_data_lock);
3330 /* The usual case is that this is the expected next packet */
3331 if (seq == call->rnext) {
3333 /* Check to make sure it is not a duplicate of one already queued */
3334 if (queue_IsNotEmpty(&call->rq)
3335 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3336 if (rx_stats_active)
3337 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3338 dpf(("packet %x dropped on receipt - duplicate", np));
3339 rxevent_Cancel(call->delayedAckEvent, call,
3340 RX_CALL_REFCOUNT_DELAY);
3341 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3347 /* It's the next packet. Stick it on the receive queue
3348 * for this call. Set newPackets to make sure we wake
3349 * the reader once all packets have been processed */
3350 np->flags |= RX_PKTFLAG_RQ;
3351 queue_Prepend(&call->rq, np);
3352 #ifdef RXDEBUG_PACKET
3354 #endif /* RXDEBUG_PACKET */
3356 np = NULL; /* We can't use this anymore */
3359 /* If an ack is requested then set a flag to make sure we
3360 * send an acknowledgement for this packet */
3361 if (flags & RX_REQUEST_ACK) {
3362 ackNeeded = RX_ACK_REQUESTED;
3365 /* Keep track of whether we have received the last packet */
3366 if (flags & RX_LAST_PACKET) {
3367 call->flags |= RX_CALL_HAVE_LAST;
3371 /* Check whether we have all of the packets for this call */
3372 if (call->flags & RX_CALL_HAVE_LAST) {
3373 afs_uint32 tseq; /* temporary sequence number */
3374 struct rx_packet *tp; /* Temporary packet pointer */
3375 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3377 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3378 if (tseq != tp->header.seq)
3380 if (tp->header.flags & RX_LAST_PACKET) {
3381 call->flags |= RX_CALL_RECEIVE_DONE;
3388 /* Provide asynchronous notification for those who want it
3389 * (e.g. multi rx) */
3390 if (call->arrivalProc) {
3391 (*call->arrivalProc) (call, call->arrivalProcHandle,
3392 call->arrivalProcArg);
3393 call->arrivalProc = (void (*)())0;
3396 /* Update last packet received */
3399 /* If there is no server process serving this call, grab
3400 * one, if available. We only need to do this once. If a
3401 * server thread is available, this thread becomes a server
3402 * thread and the server thread becomes a listener thread. */
3404 TryAttach(call, socket, tnop, newcallp, 0);
3407 /* This is not the expected next packet. */
3409 /* Determine whether this is a new or old packet, and if it's
3410 * a new one, whether it fits into the current receive window.
3411 * Also figure out whether the packet was delivered in sequence.
3412 * We use the prev variable to determine whether the new packet
3413 * is the successor of its immediate predecessor in the
3414 * receive queue, and the missing flag to determine whether
3415 * any of this packets predecessors are missing. */
3417 afs_uint32 prev; /* "Previous packet" sequence number */
3418 struct rx_packet *tp; /* Temporary packet pointer */
3419 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3420 int missing; /* Are any predecessors missing? */
3422 /* If the new packet's sequence number has been sent to the
3423 * application already, then this is a duplicate */
3424 if (seq < call->rnext) {
3425 if (rx_stats_active)
3426 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3427 rxevent_Cancel(call->delayedAckEvent, call,
3428 RX_CALL_REFCOUNT_DELAY);
3429 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3435 /* If the sequence number is greater than what can be
3436 * accomodated by the current window, then send a negative
3437 * acknowledge and drop the packet */
3438 if ((call->rnext + call->rwind) <= seq) {
3439 rxevent_Cancel(call->delayedAckEvent, call,
3440 RX_CALL_REFCOUNT_DELAY);
3441 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3448 /* Look for the packet in the queue of old received packets */
3449 for (prev = call->rnext - 1, missing =
3450 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3451 /*Check for duplicate packet */
3452 if (seq == tp->header.seq) {
3453 if (rx_stats_active)
3454 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3455 rxevent_Cancel(call->delayedAckEvent, call,
3456 RX_CALL_REFCOUNT_DELAY);
3457 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3463 /* If we find a higher sequence packet, break out and
3464 * insert the new packet here. */
3465 if (seq < tp->header.seq)
3467 /* Check for missing packet */
3468 if (tp->header.seq != prev + 1) {
3472 prev = tp->header.seq;
3475 /* Keep track of whether we have received the last packet. */
3476 if (flags & RX_LAST_PACKET) {
3477 call->flags |= RX_CALL_HAVE_LAST;
3480 /* It's within the window: add it to the the receive queue.
3481 * tp is left by the previous loop either pointing at the
3482 * packet before which to insert the new packet, or at the
3483 * queue head if the queue is empty or the packet should be
3485 np->flags |= RX_PKTFLAG_RQ;
3486 #ifdef RXDEBUG_PACKET
3488 #endif /* RXDEBUG_PACKET */
3489 queue_InsertBefore(tp, np);
3493 /* Check whether we have all of the packets for this call */
3494 if ((call->flags & RX_CALL_HAVE_LAST)
3495 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3496 afs_uint32 tseq; /* temporary sequence number */
3499 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3500 if (tseq != tp->header.seq)
3502 if (tp->header.flags & RX_LAST_PACKET) {
3503 call->flags |= RX_CALL_RECEIVE_DONE;
3510 /* We need to send an ack of the packet is out of sequence,
3511 * or if an ack was requested by the peer. */
3512 if (seq != prev + 1 || missing) {
3513 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3514 } else if (flags & RX_REQUEST_ACK) {
3515 ackNeeded = RX_ACK_REQUESTED;
3518 /* Acknowledge the last packet for each call */
3519 if (flags & RX_LAST_PACKET) {
3530 * If the receiver is waiting for an iovec, fill the iovec
3531 * using the data from the receive queue */
3532 if (call->flags & RX_CALL_IOVEC_WAIT) {
3533 didHardAck = rxi_FillReadVec(call, serial);
3534 /* the call may have been aborted */
3543 /* Wakeup the reader if any */
3544 if ((call->flags & RX_CALL_READER_WAIT)
3545 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3546 || (call->iovNext >= call->iovMax)
3547 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3548 call->flags &= ~RX_CALL_READER_WAIT;
3549 #ifdef RX_ENABLE_LOCKS
3550 CV_BROADCAST(&call->cv_rq);
3552 osi_rxWakeup(&call->rq);
3558 * Send an ack when requested by the peer, or once every
3559 * rxi_SoftAckRate packets until the last packet has been
3560 * received. Always send a soft ack for the last packet in
3561 * the server's reply. */
3563 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3564 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3565 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3566 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3567 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3568 } else if (call->nSoftAcks) {
3569 clock_GetTime(&now);
3571 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3572 clock_Add(&when, &rx_lastAckDelay);
3574 clock_Add(&when, &rx_softAckDelay);
3576 if (!call->delayedAckEvent
3577 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3578 rxevent_Cancel(call->delayedAckEvent, call,
3579 RX_CALL_REFCOUNT_DELAY);
3580 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3581 call->delayedAckEvent =
3582 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3584 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3585 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3592 static void rxi_ComputeRate();
3596 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3598 struct rx_peer *peer = conn->peer;
3600 MUTEX_ENTER(&peer->peer_lock);
3601 peer->lastReachTime = clock_Sec();
3602 MUTEX_EXIT(&peer->peer_lock);
3604 MUTEX_ENTER(&conn->conn_data_lock);
3605 if (conn->flags & RX_CONN_ATTACHWAIT) {
3608 conn->flags &= ~RX_CONN_ATTACHWAIT;
3609 MUTEX_EXIT(&conn->conn_data_lock);
3611 for (i = 0; i < RX_MAXCALLS; i++) {
3612 struct rx_call *call = conn->call[i];
3615 MUTEX_ENTER(&call->lock);
3616 /* tnop can be null if newcallp is null */
3617 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3619 MUTEX_EXIT(&call->lock);
3623 MUTEX_EXIT(&conn->conn_data_lock);
3626 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
3628 rx_ack_reason(int reason)
3631 case RX_ACK_REQUESTED:
3633 case RX_ACK_DUPLICATE:
3635 case RX_ACK_OUT_OF_SEQUENCE:
3637 case RX_ACK_EXCEEDS_WINDOW:
3639 case RX_ACK_NOSPACE:
3643 case RX_ACK_PING_RESPONSE:
3656 /* rxi_ComputePeerNetStats
3658 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3659 * estimates (like RTT and throughput) based on ack packets. Caller
3660 * must ensure that the packet in question is the right one (i.e.
3661 * serial number matches).
3664 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3665 struct rx_ackPacket *ap, struct rx_packet *np)
3667 struct rx_peer *peer = call->conn->peer;
3669 /* Use RTT if not delayed by client. */
3670 if (ap->reason != RX_ACK_DELAY)
3671 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3673 rxi_ComputeRate(peer, call, p, np, ap->reason);
3677 /* The real smarts of the whole thing. */
3679 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
3682 struct rx_ackPacket *ap;
3684 struct rx_packet *tp;
3685 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3686 struct rx_connection *conn = call->conn;
3687 struct rx_peer *peer = conn->peer;
3690 /* because there are CM's that are bogus, sending weird values for this. */
3691 afs_uint32 skew = 0;
3696 int newAckCount = 0;
3697 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3698 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3700 if (rx_stats_active)
3701 rx_MutexIncrement(rx_stats.ackPacketsRead, rx_stats_mutex);
3702 ap = (struct rx_ackPacket *)rx_DataOf(np);
3703 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3705 return np; /* truncated ack packet */
3707 /* depends on ack packet struct */
3708 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3709 first = ntohl(ap->firstPacket);
3710 serial = ntohl(ap->serial);
3711 /* temporarily disabled -- needs to degrade over time
3712 * skew = ntohs(ap->maxSkew); */
3714 /* Ignore ack packets received out of order */
3715 if (first < call->tfirst) {
3719 if (np->header.flags & RX_SLOW_START_OK) {
3720 call->flags |= RX_CALL_SLOW_START_OK;
3723 if (ap->reason == RX_ACK_PING_RESPONSE)
3724 rxi_UpdatePeerReach(conn, call);
3728 if (rxdebug_active) {
3732 len = _snprintf(msg, sizeof(msg),
3733 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3734 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3735 ntohl(ap->serial), ntohl(ap->previousPacket),
3736 (unsigned int)np->header.seq, (unsigned int)skew,
3737 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3741 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3742 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3746 OutputDebugString(msg);
3748 #else /* AFS_NT40_ENV */
3751 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3752 ap->reason, ntohl(ap->previousPacket),
3753 (unsigned int)np->header.seq, (unsigned int)serial,
3754 (unsigned int)skew, ntohl(ap->firstPacket));
3757 for (offset = 0; offset < nAcks; offset++)
3758 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3763 #endif /* AFS_NT40_ENV */
3766 /* Update the outgoing packet skew value to the latest value of
3767 * the peer's incoming packet skew value. The ack packet, of
3768 * course, could arrive out of order, but that won't affect things
3770 MUTEX_ENTER(&peer->peer_lock);
3771 peer->outPacketSkew = skew;
3773 /* Check for packets that no longer need to be transmitted, and
3774 * discard them. This only applies to packets positively
3775 * acknowledged as having been sent to the peer's upper level.
3776 * All other packets must be retained. So only packets with
3777 * sequence numbers < ap->firstPacket are candidates. */
3778 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3779 if (tp->header.seq >= first)
3781 call->tfirst = tp->header.seq + 1;
3783 && (tp->header.serial == serial || tp->firstSerial == serial))
3784 rxi_ComputePeerNetStats(call, tp, ap, np);
3785 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3788 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3789 /* XXX Hack. Because we have to release the global rx lock when sending
3790 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3791 * in rxi_Start sending packets out because packets may move to the
3792 * freePacketQueue as result of being here! So we drop these packets until
3793 * we're safely out of the traversing. Really ugly!
3794 * To make it even uglier, if we're using fine grain locking, we can
3795 * set the ack bits in the packets and have rxi_Start remove the packets
3796 * when it's done transmitting.
3798 if (call->flags & RX_CALL_TQ_BUSY) {
3799 #ifdef RX_ENABLE_LOCKS
3800 tp->flags |= RX_PKTFLAG_ACKED;
3801 call->flags |= RX_CALL_TQ_SOME_ACKED;
3802 #else /* RX_ENABLE_LOCKS */
3804 #endif /* RX_ENABLE_LOCKS */
3806 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3809 tp->flags &= ~RX_PKTFLAG_TQ;
3810 #ifdef RXDEBUG_PACKET
3812 #endif /* RXDEBUG_PACKET */
3813 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3818 /* Give rate detector a chance to respond to ping requests */
3819 if (ap->reason == RX_ACK_PING_RESPONSE) {
3820 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3824 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3826 /* Now go through explicit acks/nacks and record the results in
3827 * the waiting packets. These are packets that can't be released
3828 * yet, even with a positive acknowledge. This positive
3829 * acknowledge only means the packet has been received by the
3830 * peer, not that it will be retained long enough to be sent to
3831 * the peer's upper level. In addition, reset the transmit timers
3832 * of any missing packets (those packets that must be missing
3833 * because this packet was out of sequence) */
3835 call->nSoftAcked = 0;
3836 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3837 /* Update round trip time if the ack was stimulated on receipt
3839 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3840 #ifdef RX_ENABLE_LOCKS
3841 if (tp->header.seq >= first)
3842 #endif /* RX_ENABLE_LOCKS */
3843 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3845 && (tp->header.serial == serial || tp->firstSerial == serial))
3846 rxi_ComputePeerNetStats(call, tp, ap, np);
3848 /* Set the acknowledge flag per packet based on the
3849 * information in the ack packet. An acknowlegded packet can
3850 * be downgraded when the server has discarded a packet it
3851 * soacked previously, or when an ack packet is received
3852 * out of sequence. */
3853 if (tp->header.seq < first) {
3854 /* Implicit ack information */
3855 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3858 tp->flags |= RX_PKTFLAG_ACKED;
3859 } else if (tp->header.seq < first + nAcks) {
3860 /* Explicit ack information: set it in the packet appropriately */
3861 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3862 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3864 tp->flags |= RX_PKTFLAG_ACKED;
3871 } else /* RX_ACK_TYPE_NACK */ {
3872 tp->flags &= ~RX_PKTFLAG_ACKED;
3876 tp->flags &= ~RX_PKTFLAG_ACKED;
3880 /* If packet isn't yet acked, and it has been transmitted at least
3881 * once, reset retransmit time using latest timeout
3882 * ie, this should readjust the retransmit timer for all outstanding
3883 * packets... So we don't just retransmit when we should know better*/
3885 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3886 tp->retryTime = tp->timeSent;
3887 clock_Add(&tp->retryTime, &peer->timeout);
3888 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3889 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3893 /* If the window has been extended by this acknowledge packet,
3894 * then wakeup a sender waiting in alloc for window space, or try
3895 * sending packets now, if he's been sitting on packets due to
3896 * lack of window space */
3897 if (call->tnext < (call->tfirst + call->twind)) {
3898 #ifdef RX_ENABLE_LOCKS
3899 CV_SIGNAL(&call->cv_twind);
3901 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3902 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3903 osi_rxWakeup(&call->twind);
3906 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3907 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3911 /* if the ack packet has a receivelen field hanging off it,
3912 * update our state */
3913 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3916 /* If the ack packet has a "recommended" size that is less than
3917 * what I am using now, reduce my size to match */
3918 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3919 (int)sizeof(afs_int32), &tSize);
3920 tSize = (afs_uint32) ntohl(tSize);
3921 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3923 /* Get the maximum packet size to send to this peer */
3924 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3926 tSize = (afs_uint32) ntohl(tSize);
3927 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3928 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3930 /* sanity check - peer might have restarted with different params.
3931 * If peer says "send less", dammit, send less... Peer should never
3932 * be unable to accept packets of the size that prior AFS versions would
3933 * send without asking. */
3934 if (peer->maxMTU != tSize) {
3935 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3937 peer->maxMTU = tSize;
3938 peer->MTU = MIN(tSize, peer->MTU);
3939 call->MTU = MIN(call->MTU, tSize);
3942 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3945 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3946 (int)sizeof(afs_int32), &tSize);
3947 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3948 if (tSize < call->twind) { /* smaller than our send */
3949 call->twind = tSize; /* window, we must send less... */
3950 call->ssthresh = MIN(call->twind, call->ssthresh);
3951 call->conn->twind[call->channel] = call->twind;
3954 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3955 * network MTU confused with the loopback MTU. Calculate the
3956 * maximum MTU here for use in the slow start code below.
3958 maxMTU = peer->maxMTU;
3959 /* Did peer restart with older RX version? */
3960 if (peer->maxDgramPackets > 1) {
3961 peer->maxDgramPackets = 1;
3963 } else if (np->length >=
3964 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3967 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3968 sizeof(afs_int32), &tSize);
3969 tSize = (afs_uint32) ntohl(tSize);
3971 * As of AFS 3.5 we set the send window to match the receive window.
3973 if (tSize < call->twind) {
3974 call->twind = tSize;
3975 call->conn->twind[call->channel] = call->twind;
3976 call->ssthresh = MIN(call->twind, call->ssthresh);
3977 } else if (tSize > call->twind) {
3978 call->twind = tSize;
3979 call->conn->twind[call->channel] = call->twind;
3983 * As of AFS 3.5, a jumbogram is more than one fixed size
3984 * packet transmitted in a single UDP datagram. If the remote
3985 * MTU is smaller than our local MTU then never send a datagram
3986 * larger than the natural MTU.
3989 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3990 sizeof(afs_int32), &tSize);
3991 maxDgramPackets = (afs_uint32) ntohl(tSize);
3992 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3994 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3995 maxDgramPackets = MIN(maxDgramPackets, tSize);
3996 if (maxDgramPackets > 1) {
3997 peer->maxDgramPackets = maxDgramPackets;
3998 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4000 peer->maxDgramPackets = 1;
4001 call->MTU = peer->natMTU;
4003 } else if (peer->maxDgramPackets > 1) {
4004 /* Restarted with lower version of RX */
4005 peer->maxDgramPackets = 1;
4007 } else if (peer->maxDgramPackets > 1
4008 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4009 /* Restarted with lower version of RX */
4010 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4011 peer->natMTU = OLD_MAX_PACKET_SIZE;
4012 peer->MTU = OLD_MAX_PACKET_SIZE;
4013 peer->maxDgramPackets = 1;
4014 peer->nDgramPackets = 1;
4016 call->MTU = OLD_MAX_PACKET_SIZE;
4021 * Calculate how many datagrams were successfully received after
4022 * the first missing packet and adjust the negative ack counter
4027 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4028 if (call->nNacks < nNacked) {
4029 call->nNacks = nNacked;
4032 call->nAcks += newAckCount;
4036 if (call->flags & RX_CALL_FAST_RECOVER) {
4038 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4040 call->flags &= ~RX_CALL_FAST_RECOVER;
4041 call->cwind = call->nextCwind;
4042 call->nextCwind = 0;
4045 call->nCwindAcks = 0;
4046 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4047 /* Three negative acks in a row trigger congestion recovery */
4048 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4049 MUTEX_EXIT(&peer->peer_lock);
4050 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4051 /* someone else is waiting to start recovery */
4054 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4055 rxi_WaitforTQBusy(call);
4056 MUTEX_ENTER(&peer->peer_lock);
4057 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4058 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4059 call->flags |= RX_CALL_FAST_RECOVER;
4060 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4062 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4063 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4064 call->nextCwind = call->ssthresh;
4067 peer->MTU = call->MTU;
4068 peer->cwind = call->nextCwind;
4069 peer->nDgramPackets = call->nDgramPackets;
4071 call->congestSeq = peer->congestSeq;
4072 /* Reset the resend times on the packets that were nacked
4073 * so we will retransmit as soon as the window permits*/
4074 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4076 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4077 clock_Zero(&tp->retryTime);
4079 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4084 /* If cwind is smaller than ssthresh, then increase
4085 * the window one packet for each ack we receive (exponential
4087 * If cwind is greater than or equal to ssthresh then increase
4088 * the congestion window by one packet for each cwind acks we
4089 * receive (linear growth). */
4090 if (call->cwind < call->ssthresh) {
4092 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4093 call->nCwindAcks = 0;
4095 call->nCwindAcks += newAckCount;
4096 if (call->nCwindAcks >= call->cwind) {
4097 call->nCwindAcks = 0;
4098 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4102 * If we have received several acknowledgements in a row then
4103 * it is time to increase the size of our datagrams
4105 if ((int)call->nAcks > rx_nDgramThreshold) {
4106 if (peer->maxDgramPackets > 1) {
4107 if (call->nDgramPackets < peer->maxDgramPackets) {
4108 call->nDgramPackets++;
4110 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4111 } else if (call->MTU < peer->maxMTU) {
4112 call->MTU += peer->natMTU;
4113 call->MTU = MIN(call->MTU, peer->maxMTU);
4119 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4121 /* Servers need to hold the call until all response packets have
4122 * been acknowledged. Soft acks are good enough since clients
4123 * are not allowed to clear their receive queues. */
4124 if (call->state == RX_STATE_HOLD
4125 && call->tfirst + call->nSoftAcked >= call->tnext) {
4126 call->state = RX_STATE_DALLY;
4127 rxi_ClearTransmitQueue(call, 0);
4128 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4129 } else if (!queue_IsEmpty(&call->tq)) {
4130 rxi_Start(0, call, 0, istack);
4135 /* Received a response to a challenge packet */
4137 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4138 struct rx_packet *np, int istack)
4142 /* Ignore the packet if we're the client */
4143 if (conn->type == RX_CLIENT_CONNECTION)
4146 /* If already authenticated, ignore the packet (it's probably a retry) */
4147 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4150 /* Otherwise, have the security object evaluate the response packet */
4151 error = RXS_CheckResponse(conn->securityObject, conn, np);
4153 /* If the response is invalid, reset the connection, sending
4154 * an abort to the peer */
4158 rxi_ConnectionError(conn, error);
4159 MUTEX_ENTER(&conn->conn_data_lock);
4160 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4161 MUTEX_EXIT(&conn->conn_data_lock);
4164 /* If the response is valid, any calls waiting to attach
4165 * servers can now do so */
4168 for (i = 0; i < RX_MAXCALLS; i++) {
4169 struct rx_call *call = conn->call[i];
4171 MUTEX_ENTER(&call->lock);
4172 if (call->state == RX_STATE_PRECALL)
4173 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4174 /* tnop can be null if newcallp is null */
4175 MUTEX_EXIT(&call->lock);
4179 /* Update the peer reachability information, just in case
4180 * some calls went into attach-wait while we were waiting
4181 * for authentication..
4183 rxi_UpdatePeerReach(conn, NULL);
4188 /* A client has received an authentication challenge: the security
4189 * object is asked to cough up a respectable response packet to send
4190 * back to the server. The server is responsible for retrying the
4191 * challenge if it fails to get a response. */
4194 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4195 struct rx_packet *np, int istack)
4199 /* Ignore the challenge if we're the server */
4200 if (conn->type == RX_SERVER_CONNECTION)
4203 /* Ignore the challenge if the connection is otherwise idle; someone's
4204 * trying to use us as an oracle. */
4205 if (!rxi_HasActiveCalls(conn))
4208 /* Send the security object the challenge packet. It is expected to fill
4209 * in the response. */
4210 error = RXS_GetResponse(conn->securityObject, conn, np);
4212 /* If the security object is unable to return a valid response, reset the
4213 * connection and send an abort to the peer. Otherwise send the response
4214 * packet to the peer connection. */
4216 rxi_ConnectionError(conn, error);
4217 MUTEX_ENTER(&conn->conn_data_lock);
4218 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4219 MUTEX_EXIT(&conn->conn_data_lock);
4221 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4222 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4228 /* Find an available server process to service the current request in
4229 * the given call structure. If one isn't available, queue up this
4230 * call so it eventually gets one */
4232 rxi_AttachServerProc(struct rx_call *call,
4233 osi_socket socket, int *tnop,
4234 struct rx_call **newcallp)
4236 struct rx_serverQueueEntry *sq;
4237 struct rx_service *service = call->conn->service;
4240 /* May already be attached */
4241 if (call->state == RX_STATE_ACTIVE)
4244 MUTEX_ENTER(&rx_serverPool_lock);
4246 haveQuota = QuotaOK(service);
4247 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4248 /* If there are no processes available to service this call,
4249 * put the call on the incoming call queue (unless it's
4250 * already on the queue).
4252 #ifdef RX_ENABLE_LOCKS
4254 ReturnToServerPool(service);
4255 #endif /* RX_ENABLE_LOCKS */
4257 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4258 call->flags |= RX_CALL_WAIT_PROC;
4259 MUTEX_ENTER(&rx_waiting_mutex);
4262 MUTEX_EXIT(&rx_waiting_mutex);
4263 rxi_calltrace(RX_CALL_ARRIVAL, call);
4264 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4265 queue_Append(&rx_incomingCallQueue, call);
4268 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4270 /* If hot threads are enabled, and both newcallp and sq->socketp
4271 * are non-null, then this thread will process the call, and the
4272 * idle server thread will start listening on this threads socket.
4275 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4278 *sq->socketp = socket;
4279 clock_GetTime(&call->startTime);
4280 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4284 if (call->flags & RX_CALL_WAIT_PROC) {
4285 /* Conservative: I don't think this should happen */
4286 call->flags &= ~RX_CALL_WAIT_PROC;
4287 if (queue_IsOnQueue(call)) {
4290 MUTEX_ENTER(&rx_waiting_mutex);
4292 MUTEX_EXIT(&rx_waiting_mutex);
4295 call->state = RX_STATE_ACTIVE;
4296 call->mode = RX_MODE_RECEIVING;
4297 #ifdef RX_KERNEL_TRACE
4299 int glockOwner = ISAFS_GLOCK();
4302 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4303 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4309 if (call->flags & RX_CALL_CLEARED) {
4310 /* send an ack now to start the packet flow up again */
4311 call->flags &= ~RX_CALL_CLEARED;
4312 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4314 #ifdef RX_ENABLE_LOCKS
4317 service->nRequestsRunning++;
4318 if (service->nRequestsRunning <= service->minProcs)
4324 MUTEX_EXIT(&rx_serverPool_lock);
4327 /* Delay the sending of an acknowledge event for a short while, while
4328 * a new call is being prepared (in the case of a client) or a reply
4329 * is being prepared (in the case of a server). Rather than sending
4330 * an ack packet, an ACKALL packet is sent. */
4332 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4334 #ifdef RX_ENABLE_LOCKS
4336 MUTEX_ENTER(&call->lock);
4337 call->delayedAckEvent = NULL;
4338 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4340 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4341 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4343 MUTEX_EXIT(&call->lock);
4344 #else /* RX_ENABLE_LOCKS */
4346 call->delayedAckEvent = NULL;
4347 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4348 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4349 #endif /* RX_ENABLE_LOCKS */
4353 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4355 struct rx_call *call = arg1;
4356 #ifdef RX_ENABLE_LOCKS
4358 MUTEX_ENTER(&call->lock);
4359 if (event == call->delayedAckEvent)
4360 call->delayedAckEvent = NULL;
4361 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4363 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4365 MUTEX_EXIT(&call->lock);
4366 #else /* RX_ENABLE_LOCKS */
4368 call->delayedAckEvent = NULL;
4369 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4370 #endif /* RX_ENABLE_LOCKS */
4374 #ifdef RX_ENABLE_LOCKS
4375 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4376 * clearing them out.
4379 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4381 struct rx_packet *p, *tp;
4384 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4385 p->flags |= RX_PKTFLAG_ACKED;
4389 call->flags |= RX_CALL_TQ_CLEARME;
4390 call->flags |= RX_CALL_TQ_SOME_ACKED;
4393 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4394 call->tfirst = call->tnext;
4395 call->nSoftAcked = 0;
4397 if (call->flags & RX_CALL_FAST_RECOVER) {
4398 call->flags &= ~RX_CALL_FAST_RECOVER;
4399 call->cwind = call->nextCwind;
4400 call->nextCwind = 0;
4403 CV_SIGNAL(&call->cv_twind);
4405 #endif /* RX_ENABLE_LOCKS */
4407 /* Clear out the transmit queue for the current call (all packets have
4408 * been received by peer) */
4410 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4412 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4413 struct rx_packet *p, *tp;
4415 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4417 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4418 p->flags |= RX_PKTFLAG_ACKED;
4422 call->flags |= RX_CALL_TQ_CLEARME;
4423 call->flags |= RX_CALL_TQ_SOME_ACKED;
4426 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4427 #ifdef RXDEBUG_PACKET
4429 #endif /* RXDEBUG_PACKET */
4430 rxi_FreePackets(0, &call->tq);
4431 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4432 call->flags &= ~RX_CALL_TQ_CLEARME;
4434 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4436 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4437 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4438 call->nSoftAcked = 0;
4440 if (call->flags & RX_CALL_FAST_RECOVER) {
4441 call->flags &= ~RX_CALL_FAST_RECOVER;
4442 call->cwind = call->nextCwind;
4444 #ifdef RX_ENABLE_LOCKS
4445 CV_SIGNAL(&call->cv_twind);
4447 osi_rxWakeup(&call->twind);
4452 rxi_ClearReceiveQueue(struct rx_call *call)
4454 if (queue_IsNotEmpty(&call->rq)) {
4457 count = rxi_FreePackets(0, &call->rq);
4458 rx_packetReclaims += count;
4459 #ifdef RXDEBUG_PACKET
4461 if ( call->rqc != 0 )
4462 dpf(("rxi_ClearReceiveQueue call %x rqc %u != 0", call, call->rqc));
4464 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4466 if (call->state == RX_STATE_PRECALL) {
4467 call->flags |= RX_CALL_CLEARED;
4471 /* Send an abort packet for the specified call */
4473 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
4474 int istack, int force)
4477 struct clock when, now;
4482 /* Clients should never delay abort messages */
4483 if (rx_IsClientConn(call->conn))
4486 if (call->abortCode != call->error) {
4487 call->abortCode = call->error;
4488 call->abortCount = 0;
4491 if (force || rxi_callAbortThreshhold == 0
4492 || call->abortCount < rxi_callAbortThreshhold) {
4493 if (call->delayedAbortEvent) {
4494 rxevent_Cancel(call->delayedAbortEvent, call,
4495 RX_CALL_REFCOUNT_ABORT);
4497 error = htonl(call->error);
4500 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4501 (char *)&error, sizeof(error), istack);
4502 } else if (!call->delayedAbortEvent) {
4503 clock_GetTime(&now);
4505 clock_Addmsec(&when, rxi_callAbortDelay);
4506 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4507 call->delayedAbortEvent =
4508 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4513 /* Send an abort packet for the specified connection. Packet is an
4514 * optional pointer to a packet that can be used to send the abort.
4515 * Once the number of abort messages reaches the threshhold, an
4516 * event is scheduled to send the abort. Setting the force flag
4517 * overrides sending delayed abort messages.
4519 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4520 * to send the abort packet.
4523 rxi_SendConnectionAbort(struct rx_connection *conn,
4524 struct rx_packet *packet, int istack, int force)
4527 struct clock when, now;
4532 /* Clients should never delay abort messages */
4533 if (rx_IsClientConn(conn))
4536 if (force || rxi_connAbortThreshhold == 0
4537 || conn->abortCount < rxi_connAbortThreshhold) {
4538 if (conn->delayedAbortEvent) {
4539 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4541 error = htonl(conn->error);
4543 MUTEX_EXIT(&conn->conn_data_lock);
4545 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4546 RX_PACKET_TYPE_ABORT, (char *)&error,
4547 sizeof(error), istack);
4548 MUTEX_ENTER(&conn->conn_data_lock);
4549 } else if (!conn->delayedAbortEvent) {
4550 clock_GetTime(&now);
4552 clock_Addmsec(&when, rxi_connAbortDelay);
4553 conn->delayedAbortEvent =
4554 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4559 /* Associate an error all of the calls owned by a connection. Called
4560 * with error non-zero. This is only for really fatal things, like
4561 * bad authentication responses. The connection itself is set in
4562 * error at this point, so that future packets received will be
4565 rxi_ConnectionError(struct rx_connection *conn,
4571 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4573 MUTEX_ENTER(&conn->conn_data_lock);
4574 if (conn->challengeEvent)
4575 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4576 if (conn->checkReachEvent) {
4577 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4578 conn->checkReachEvent = 0;
4579 conn->flags &= ~RX_CONN_ATTACHWAIT;
4582 MUTEX_EXIT(&conn->conn_data_lock);
4583 for (i = 0; i < RX_MAXCALLS; i++) {
4584 struct rx_call *call = conn->call[i];
4586 MUTEX_ENTER(&call->lock);
4587 rxi_CallError(call, error);
4588 MUTEX_EXIT(&call->lock);
4591 conn->error = error;
4592 if (rx_stats_active)
4593 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4598 rxi_CallError(struct rx_call *call, afs_int32 error)
4601 osirx_AssertMine(&call->lock, "rxi_CallError");
4603 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4605 error = call->error;
4607 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4608 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4609 rxi_ResetCall(call, 0);
4612 rxi_ResetCall(call, 0);
4614 call->error = error;
4615 call->mode = RX_MODE_ERROR;
4618 /* Reset various fields in a call structure, and wakeup waiting
4619 * processes. Some fields aren't changed: state & mode are not
4620 * touched (these must be set by the caller), and bufptr, nLeft, and
4621 * nFree are not reset, since these fields are manipulated by
4622 * unprotected macros, and may only be reset by non-interrupting code.
4625 /* this code requires that call->conn be set properly as a pre-condition. */
4626 #endif /* ADAPT_WINDOW */
4629 rxi_ResetCall(struct rx_call *call, int newcall)
4632 struct rx_peer *peer;
4633 struct rx_packet *packet;
4635 osirx_AssertMine(&call->lock, "rxi_ResetCall");
4637 dpf(("rxi_ResetCall(call %x, newcall %d)\n", call, newcall));
4639 /* Notify anyone who is waiting for asynchronous packet arrival */
4640 if (call->arrivalProc) {
4641 (*call->arrivalProc) (call, call->arrivalProcHandle,
4642 call->arrivalProcArg);
4643 call->arrivalProc = (void (*)())0;
4646 if (call->delayedAbortEvent) {
4647 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4648 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4650 rxi_SendCallAbort(call, packet, 0, 1);
4651 rxi_FreePacket(packet);
4656 * Update the peer with the congestion information in this call
4657 * so other calls on this connection can pick up where this call
4658 * left off. If the congestion sequence numbers don't match then
4659 * another call experienced a retransmission.
4661 peer = call->conn->peer;
4662 MUTEX_ENTER(&peer->peer_lock);
4664 if (call->congestSeq == peer->congestSeq) {
4665 peer->cwind = MAX(peer->cwind, call->cwind);
4666 peer->MTU = MAX(peer->MTU, call->MTU);
4667 peer->nDgramPackets =
4668 MAX(peer->nDgramPackets, call->nDgramPackets);
4671 call->abortCode = 0;
4672 call->abortCount = 0;
4674 if (peer->maxDgramPackets > 1) {
4675 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4677 call->MTU = peer->MTU;
4679 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4680 call->ssthresh = rx_maxSendWindow;
4681 call->nDgramPackets = peer->nDgramPackets;
4682 call->congestSeq = peer->congestSeq;
4683 MUTEX_EXIT(&peer->peer_lock);
4685 flags = call->flags;
4686 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4687 if (flags & RX_CALL_TQ_BUSY) {
4688 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4689 call->flags |= (flags & RX_CALL_TQ_WAIT);
4691 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4693 rxi_ClearTransmitQueue(call, 1);
4694 /* why init the queue if you just emptied it? queue_Init(&call->tq); */
4695 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4696 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4699 while (call->tqWaiters) {
4700 #ifdef RX_ENABLE_LOCKS
4701 CV_BROADCAST(&call->cv_tq);
4702 #else /* RX_ENABLE_LOCKS */
4703 osi_rxWakeup(&call->tq);
4704 #endif /* RX_ENABLE_LOCKS */
4709 rxi_ClearReceiveQueue(call);
4710 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
4712 if (call->currentPacket) {
4713 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
4714 call->currentPacket->flags |= RX_PKTFLAG_IOVQ;
4715 queue_Prepend(&call->iovq, call->currentPacket);
4716 #ifdef RXDEBUG_PACKET
4718 #endif /* RXDEBUG_PACKET */
4719 call->currentPacket = (struct rx_packet *)0;
4721 call->curlen = call->nLeft = call->nFree = 0;
4723 #ifdef RXDEBUG_PACKET
4726 rxi_FreePackets(0, &call->iovq);
4729 call->twind = call->conn->twind[call->channel];
4730 call->rwind = call->conn->rwind[call->channel];
4731 call->nSoftAcked = 0;
4732 call->nextCwind = 0;
4735 call->nCwindAcks = 0;
4736 call->nSoftAcks = 0;
4737 call->nHardAcks = 0;
4739 call->tfirst = call->rnext = call->tnext = 1;
4741 call->lastAcked = 0;
4742 call->localStatus = call->remoteStatus = 0;
4744 if (flags & RX_CALL_READER_WAIT) {
4745 #ifdef RX_ENABLE_LOCKS
4746 CV_BROADCAST(&call->cv_rq);
4748 osi_rxWakeup(&call->rq);
4751 if (flags & RX_CALL_WAIT_PACKETS) {
4752 MUTEX_ENTER(&rx_freePktQ_lock);
4753 rxi_PacketsUnWait(); /* XXX */
4754 MUTEX_EXIT(&rx_freePktQ_lock);
4756 #ifdef RX_ENABLE_LOCKS
4757 CV_SIGNAL(&call->cv_twind);
4759 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4760 osi_rxWakeup(&call->twind);
4763 #ifdef RX_ENABLE_LOCKS
4764 /* The following ensures that we don't mess with any queue while some
4765 * other thread might also be doing so. The call_queue_lock field is
4766 * is only modified under the call lock. If the call is in the process
4767 * of being removed from a queue, the call is not locked until the
4768 * the queue lock is dropped and only then is the call_queue_lock field
4769 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4770 * Note that any other routine which removes a call from a queue has to
4771 * obtain the queue lock before examing the queue and removing the call.
4773 if (call->call_queue_lock) {
4774 MUTEX_ENTER(call->call_queue_lock);
4775 if (queue_IsOnQueue(call)) {
4777 if (flags & RX_CALL_WAIT_PROC) {
4779 MUTEX_ENTER(&rx_waiting_mutex);
4781 MUTEX_EXIT(&rx_waiting_mutex);
4784 MUTEX_EXIT(call->call_queue_lock);
4785 CLEAR_CALL_QUEUE_LOCK(call);
4787 #else /* RX_ENABLE_LOCKS */
4788 if (queue_IsOnQueue(call)) {
4790 if (flags & RX_CALL_WAIT_PROC)
4793 #endif /* RX_ENABLE_LOCKS */
4795 rxi_KeepAliveOff(call);
4796 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4799 /* Send an acknowledge for the indicated packet (seq,serial) of the
4800 * indicated call, for the indicated reason (reason). This
4801 * acknowledge will specifically acknowledge receiving the packet, and
4802 * will also specify which other packets for this call have been
4803 * received. This routine returns the packet that was used to the
4804 * caller. The caller is responsible for freeing it or re-using it.
4805 * This acknowledgement also returns the highest sequence number
4806 * actually read out by the higher level to the sender; the sender
4807 * promises to keep around packets that have not been read by the
4808 * higher level yet (unless, of course, the sender decides to abort
4809 * the call altogether). Any of p, seq, serial, pflags, or reason may
4810 * be set to zero without ill effect. That is, if they are zero, they
4811 * will not convey any information.
4812 * NOW there is a trailer field, after the ack where it will safely be
4813 * ignored by mundanes, which indicates the maximum size packet this
4814 * host can swallow. */
4816 struct rx_packet *optionalPacket; use to send ack (or null)
4817 int seq; Sequence number of the packet we are acking
4818 int serial; Serial number of the packet
4819 int pflags; Flags field from packet header
4820 int reason; Reason an acknowledge was prompted
4824 rxi_SendAck(struct rx_call *call,
4825 struct rx_packet *optionalPacket, int serial, int reason,
4828 struct rx_ackPacket *ap;
4829 struct rx_packet *rqp;
4830 struct rx_packet *nxp; /* For queue_Scan */
4831 struct rx_packet *p;
4834 #ifdef RX_ENABLE_TSFPQ
4835 struct rx_ts_info_t * rx_ts_info;
4839 * Open the receive window once a thread starts reading packets
4841 if (call->rnext > 1) {
4842 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
4845 call->nHardAcks = 0;
4846 call->nSoftAcks = 0;
4847 if (call->rnext > call->lastAcked)
4848 call->lastAcked = call->rnext;
4852 rx_computelen(p, p->length); /* reset length, you never know */
4853 } /* where that's been... */
4854 #ifdef RX_ENABLE_TSFPQ
4856 RX_TS_INFO_GET(rx_ts_info);
4857 if ((p = rx_ts_info->local_special_packet)) {
4858 rx_computelen(p, p->length);
4859 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4860 rx_ts_info->local_special_packet = p;
4861 } else { /* We won't send the ack, but don't panic. */
4862 return optionalPacket;
4866 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4867 /* We won't send the ack, but don't panic. */
4868 return optionalPacket;
4873 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4876 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4877 #ifndef RX_ENABLE_TSFPQ
4878 if (!optionalPacket)
4881 return optionalPacket;
4883 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4884 if (rx_Contiguous(p) < templ) {
4885 #ifndef RX_ENABLE_TSFPQ
4886 if (!optionalPacket)
4889 return optionalPacket;
4894 /* MTUXXX failing to send an ack is very serious. We should */
4895 /* try as hard as possible to send even a partial ack; it's */
4896 /* better than nothing. */
4897 ap = (struct rx_ackPacket *)rx_DataOf(p);
4898 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4899 ap->reason = reason;
4901 /* The skew computation used to be bogus, I think it's better now. */
4902 /* We should start paying attention to skew. XXX */
4903 ap->serial = htonl(serial);
4904 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4906 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4907 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4909 /* No fear of running out of ack packet here because there can only be at most
4910 * one window full of unacknowledged packets. The window size must be constrained
4911 * to be less than the maximum ack size, of course. Also, an ack should always
4912 * fit into a single packet -- it should not ever be fragmented. */
4913 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4914 if (!rqp || !call->rq.next
4915 || (rqp->header.seq > (call->rnext + call->rwind))) {
4916 #ifndef RX_ENABLE_TSFPQ
4917 if (!optionalPacket)
4920 rxi_CallError(call, RX_CALL_DEAD);
4921 return optionalPacket;
4924 while (rqp->header.seq > call->rnext + offset)
4925 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4926 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4928 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4929 #ifndef RX_ENABLE_TSFPQ
4930 if (!optionalPacket)
4933 rxi_CallError(call, RX_CALL_DEAD);
4934 return optionalPacket;
4939 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4941 /* these are new for AFS 3.3 */
4942 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4943 templ = htonl(templ);
4944 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4945 templ = htonl(call->conn->peer->ifMTU);
4946 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4947 sizeof(afs_int32), &templ);
4949 /* new for AFS 3.4 */
4950 templ = htonl(call->rwind);
4951 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4952 sizeof(afs_int32), &templ);
4954 /* new for AFS 3.5 */
4955 templ = htonl(call->conn->peer->ifDgramPackets);
4956 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4957 sizeof(afs_int32), &templ);
4959 p->header.serviceId = call->conn->serviceId;
4960 p->header.cid = (call->conn->cid | call->channel);
4961 p->header.callNumber = *call->callNumber;
4963 p->header.securityIndex = call->conn->securityIndex;
4964 p->header.epoch = call->conn->epoch;
4965 p->header.type = RX_PACKET_TYPE_ACK;
4966 p->header.flags = RX_SLOW_START_OK;
4967 if (reason == RX_ACK_PING) {
4968 p->header.flags |= RX_REQUEST_ACK;
4970 clock_GetTime(&call->pingRequestTime);
4973 if (call->conn->type == RX_CLIENT_CONNECTION)
4974 p->header.flags |= RX_CLIENT_INITIATED;
4978 if (rxdebug_active) {
4982 len = _snprintf(msg, sizeof(msg),
4983 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4984 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4985 ntohl(ap->serial), ntohl(ap->previousPacket),
4986 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4987 ap->nAcks, ntohs(ap->bufferSpace) );
4991 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4992 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4996 OutputDebugString(msg);
4998 #else /* AFS_NT40_ENV */
5000 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5001 ap->reason, ntohl(ap->previousPacket),
5002 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5004 for (offset = 0; offset < ap->nAcks; offset++)
5005 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5010 #endif /* AFS_NT40_ENV */
5013 int i, nbytes = p->length;
5015 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5016 if (nbytes <= p->wirevec[i].iov_len) {
5019 savelen = p->wirevec[i].iov_len;
5021 p->wirevec[i].iov_len = nbytes;
5023 rxi_Send(call, p, istack);
5024 p->wirevec[i].iov_len = savelen;
5028 nbytes -= p->wirevec[i].iov_len;
5031 if (rx_stats_active)
5032 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
5033 #ifndef RX_ENABLE_TSFPQ
5034 if (!optionalPacket)
5037 return optionalPacket; /* Return packet for re-use by caller */
5040 /* Send all of the packets in the list in single datagram */
5042 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
5043 int istack, int moreFlag, struct clock *now,
5044 struct clock *retryTime, int resending)
5049 struct rx_connection *conn = call->conn;
5050 struct rx_peer *peer = conn->peer;
5052 MUTEX_ENTER(&peer->peer_lock);
5055 peer->reSends += len;
5056 if (rx_stats_active)
5057 rx_MutexAdd(rx_stats.dataPacketsSent, len, rx_stats_mutex);
5058 MUTEX_EXIT(&peer->peer_lock);
5060 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
5064 /* Set the packet flags and schedule the resend events */
5065 /* Only request an ack for the last packet in the list */
5066 for (i = 0; i < len; i++) {
5067 list[i]->retryTime = *retryTime;
5068 if (list[i]->header.serial) {
5069 /* Exponentially backoff retry times */
5070 if (list[i]->backoff < MAXBACKOFF) {
5071 /* so it can't stay == 0 */
5072 list[i]->backoff = (list[i]->backoff << 1) + 1;
5075 clock_Addmsec(&(list[i]->retryTime),
5076 ((afs_uint32) list[i]->backoff) << 8);
5079 /* Wait a little extra for the ack on the last packet */
5080 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
5081 clock_Addmsec(&(list[i]->retryTime), 400);
5084 /* Record the time sent */
5085 list[i]->timeSent = *now;
5087 /* Ask for an ack on retransmitted packets, on every other packet
5088 * if the peer doesn't support slow start. Ask for an ack on every
5089 * packet until the congestion window reaches the ack rate. */
5090 if (list[i]->header.serial) {
5092 if (rx_stats_active)
5093 rx_MutexIncrement(rx_stats.dataPacketsReSent, rx_stats_mutex);
5095 /* improved RTO calculation- not Karn */
5096 list[i]->firstSent = *now;
5097 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5098 || (!(call->flags & RX_CALL_SLOW_START_OK)
5099 && (list[i]->header.seq & 1)))) {
5104 /* Tag this packet as not being the last in this group,
5105 * for the receiver's benefit */
5106 if (i < len - 1 || moreFlag) {
5107 list[i]->header.flags |= RX_MORE_PACKETS;
5110 /* Install the new retransmit time for the packet, and
5111 * record the time sent */
5112 list[i]->timeSent = *now;
5116 list[len - 1]->header.flags |= RX_REQUEST_ACK;
5119 /* Since we're about to send a data packet to the peer, it's
5120 * safe to nuke any scheduled end-of-packets ack */
5121 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5123 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5124 MUTEX_EXIT(&call->lock);
5126 rxi_SendPacketList(call, conn, list, len, istack);
5128 rxi_SendPacket(call, conn, list[0], istack);
5130 MUTEX_ENTER(&call->lock);
5131 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5133 /* Update last send time for this call (for keep-alive
5134 * processing), and for the connection (so that we can discover
5135 * idle connections) */
5136 call->lastSendData = conn->lastSendTime = call->lastSendTime = clock_Sec();
5139 /* When sending packets we need to follow these rules:
5140 * 1. Never send more than maxDgramPackets in a jumbogram.
5141 * 2. Never send a packet with more than two iovecs in a jumbogram.
5142 * 3. Never send a retransmitted packet in a jumbogram.
5143 * 4. Never send more than cwind/4 packets in a jumbogram
5144 * We always keep the last list we should have sent so we
5145 * can set the RX_MORE_PACKETS flags correctly.
5148 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5149 int istack, struct clock *now, struct clock *retryTime,
5152 int i, cnt, lastCnt = 0;
5153 struct rx_packet **listP, **lastP = 0;
5154 struct rx_peer *peer = call->conn->peer;
5155 int morePackets = 0;
5157 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
5158 /* Does the current packet force us to flush the current list? */
5160 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5161 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5163 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
5165 /* If the call enters an error state stop sending, or if
5166 * we entered congestion recovery mode, stop sending */
5167 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5175 /* Add the current packet to the list if it hasn't been acked.
5176 * Otherwise adjust the list pointer to skip the current packet. */
5177 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5179 /* Do we need to flush the list? */
5180 if (cnt >= (int)peer->maxDgramPackets
5181 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5182 || list[i]->header.serial
5183 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5185 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5186 retryTime, resending);
5187 /* If the call enters an error state stop sending, or if
5188 * we entered congestion recovery mode, stop sending */
5190 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5195 listP = &list[i + 1];
5200 osi_Panic("rxi_SendList error");
5202 listP = &list[i + 1];
5206 /* Send the whole list when the call is in receive mode, when
5207 * the call is in eof mode, when we are in fast recovery mode,
5208 * and when we have the last packet */
5209 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5210 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5211 || (call->flags & RX_CALL_FAST_RECOVER)) {
5212 /* Check for the case where the current list contains
5213 * an acked packet. Since we always send retransmissions
5214 * in a separate packet, we only need to check the first
5215 * packet in the list */
5216 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5220 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5221 retryTime, resending);
5222 /* If the call enters an error state stop sending, or if
5223 * we entered congestion recovery mode, stop sending */
5224 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5228 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5231 } else if (lastCnt > 0) {
5232 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5237 #ifdef RX_ENABLE_LOCKS
5238 /* Call rxi_Start, below, but with the call lock held. */
5240 rxi_StartUnlocked(struct rxevent *event,
5241 void *arg0, void *arg1, int istack)
5243 struct rx_call *call = arg0;
5245 MUTEX_ENTER(&call->lock);
5246 rxi_Start(event, call, arg1, istack);
5247 MUTEX_EXIT(&call->lock);
5249 #endif /* RX_ENABLE_LOCKS */
5251 /* This routine is called when new packets are readied for
5252 * transmission and when retransmission may be necessary, or when the
5253 * transmission window or burst count are favourable. This should be
5254 * better optimized for new packets, the usual case, now that we've
5255 * got rid of queues of send packets. XXXXXXXXXXX */
5257 rxi_Start(struct rxevent *event,
5258 void *arg0, void *arg1, int istack)
5260 struct rx_call *call = arg0;
5262 struct rx_packet *p;
5263 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5264 struct rx_peer *peer = call->conn->peer;
5265 struct clock now, usenow, retryTime;
5269 struct rx_packet **xmitList;
5272 /* If rxi_Start is being called as a result of a resend event,
5273 * then make sure that the event pointer is removed from the call
5274 * structure, since there is no longer a per-call retransmission
5276 if (event && event == call->resendEvent) {
5277 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5278 call->resendEvent = NULL;
5280 if (queue_IsEmpty(&call->tq)) {
5284 /* Timeouts trigger congestion recovery */
5285 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5286 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5287 /* someone else is waiting to start recovery */
5290 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5291 rxi_WaitforTQBusy(call);
5292 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5293 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5294 call->flags |= RX_CALL_FAST_RECOVER;
5295 if (peer->maxDgramPackets > 1) {
5296 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5298 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5300 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5301 call->nDgramPackets = 1;
5303 call->nextCwind = 1;
5306 MUTEX_ENTER(&peer->peer_lock);
5307 peer->MTU = call->MTU;
5308 peer->cwind = call->cwind;
5309 peer->nDgramPackets = 1;
5311 call->congestSeq = peer->congestSeq;
5312 MUTEX_EXIT(&peer->peer_lock);
5313 /* Clear retry times on packets. Otherwise, it's possible for
5314 * some packets in the queue to force resends at rates faster
5315 * than recovery rates.
5317 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5318 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5319 clock_Zero(&p->retryTime);
5324 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5325 if (rx_stats_active)
5326 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5331 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5332 /* Get clock to compute the re-transmit time for any packets
5333 * in this burst. Note, if we back off, it's reasonable to
5334 * back off all of the packets in the same manner, even if
5335 * some of them have been retransmitted more times than more
5337 * Do a dance to avoid blocking after setting now. */
5338 clock_Zero(&retryTime);
5339 MUTEX_ENTER(&peer->peer_lock);
5340 clock_Add(&retryTime, &peer->timeout);
5341 MUTEX_EXIT(&peer->peer_lock);
5342 clock_GetTime(&now);
5343 clock_Add(&retryTime, &now);
5345 /* Send (or resend) any packets that need it, subject to
5346 * window restrictions and congestion burst control
5347 * restrictions. Ask for an ack on the last packet sent in
5348 * this burst. For now, we're relying upon the window being
5349 * considerably bigger than the largest number of packets that
5350 * are typically sent at once by one initial call to
5351 * rxi_Start. This is probably bogus (perhaps we should ask
5352 * for an ack when we're half way through the current
5353 * window?). Also, for non file transfer applications, this
5354 * may end up asking for an ack for every packet. Bogus. XXXX
5357 * But check whether we're here recursively, and let the other guy
5360 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5361 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5362 call->flags |= RX_CALL_TQ_BUSY;
5364 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5366 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5367 call->flags &= ~RX_CALL_NEED_START;
5368 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5370 maxXmitPackets = MIN(call->twind, call->cwind);
5371 xmitList = (struct rx_packet **)
5372 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
5373 /* XXXX else we must drop any mtx we hold */
5374 afs_osi_Alloc_NoSleep(maxXmitPackets * sizeof(struct rx_packet *));
5376 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5378 if (xmitList == NULL)
5379 osi_Panic("rxi_Start, failed to allocate xmit list");
5380 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5381 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5382 /* We shouldn't be sending packets if a thread is waiting
5383 * to initiate congestion recovery */
5387 && (call->flags & RX_CALL_FAST_RECOVER)) {
5388 /* Only send one packet during fast recovery */
5391 if ((p->flags & RX_PKTFLAG_FREE)
5392 || (!queue_IsEnd(&call->tq, nxp)
5393 && (nxp->flags & RX_PKTFLAG_FREE))
5394 || (p == (struct rx_packet *)&rx_freePacketQueue)
5395 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5396 osi_Panic("rxi_Start: xmit queue clobbered");
5398 if (p->flags & RX_PKTFLAG_ACKED) {
5399 /* Since we may block, don't trust this */
5400 usenow.sec = usenow.usec = 0;
5401 if (rx_stats_active)
5402 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5403 continue; /* Ignore this packet if it has been acknowledged */
5406 /* Turn off all flags except these ones, which are the same
5407 * on each transmission */
5408 p->header.flags &= RX_PRESET_FLAGS;
5410 if (p->header.seq >=
5411 call->tfirst + MIN((int)call->twind,
5412 (int)(call->nSoftAcked +
5414 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5415 /* Note: if we're waiting for more window space, we can
5416 * still send retransmits; hence we don't return here, but
5417 * break out to schedule a retransmit event */
5418 dpf(("call %d waiting for window",
5419 *(call->callNumber)));
5423 /* Transmit the packet if it needs to be sent. */
5424 if (!clock_Lt(&now, &p->retryTime)) {
5425 if (nXmitPackets == maxXmitPackets) {
5426 rxi_SendXmitList(call, xmitList, nXmitPackets,
5427 istack, &now, &retryTime,
5429 osi_Free(xmitList, maxXmitPackets *
5430 sizeof(struct rx_packet *));
5433 xmitList[nXmitPackets++] = p;
5437 /* xmitList now hold pointers to all of the packets that are
5438 * ready to send. Now we loop to send the packets */
5439 if (nXmitPackets > 0) {
5440 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5441 &now, &retryTime, resending);
5444 maxXmitPackets * sizeof(struct rx_packet *));
5446 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5448 * TQ references no longer protected by this flag; they must remain
5449 * protected by the global lock.
5451 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5452 call->flags &= ~RX_CALL_TQ_BUSY;
5453 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5454 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5455 #ifdef RX_ENABLE_LOCKS
5456 osirx_AssertMine(&call->lock, "rxi_Start start");
5457 CV_BROADCAST(&call->cv_tq);
5458 #else /* RX_ENABLE_LOCKS */
5459 osi_rxWakeup(&call->tq);
5460 #endif /* RX_ENABLE_LOCKS */
5465 /* We went into the error state while sending packets. Now is
5466 * the time to reset the call. This will also inform the using
5467 * process that the call is in an error state.
5469 if (rx_stats_active)
5470 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5471 call->flags &= ~RX_CALL_TQ_BUSY;
5472 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5473 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5474 #ifdef RX_ENABLE_LOCKS
5475 osirx_AssertMine(&call->lock, "rxi_Start middle");
5476 CV_BROADCAST(&call->cv_tq);
5477 #else /* RX_ENABLE_LOCKS */
5478 osi_rxWakeup(&call->tq);
5479 #endif /* RX_ENABLE_LOCKS */
5481 rxi_CallError(call, call->error);
5484 #ifdef RX_ENABLE_LOCKS
5485 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5487 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5488 /* Some packets have received acks. If they all have, we can clear
5489 * the transmit queue.
5492 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5493 if (p->header.seq < call->tfirst
5494 && (p->flags & RX_PKTFLAG_ACKED)) {
5496 p->flags &= ~RX_PKTFLAG_TQ;
5497 #ifdef RXDEBUG_PACKET
5505 call->flags |= RX_CALL_TQ_CLEARME;
5507 #endif /* RX_ENABLE_LOCKS */
5508 /* Don't bother doing retransmits if the TQ is cleared. */
5509 if (call->flags & RX_CALL_TQ_CLEARME) {
5510 rxi_ClearTransmitQueue(call, 1);
5512 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5515 /* Always post a resend event, if there is anything in the
5516 * queue, and resend is possible. There should be at least
5517 * one unacknowledged packet in the queue ... otherwise none
5518 * of these packets should be on the queue in the first place.
5520 if (call->resendEvent) {
5521 /* Cancel the existing event and post a new one */
5522 rxevent_Cancel(call->resendEvent, call,
5523 RX_CALL_REFCOUNT_RESEND);
5526 /* The retry time is the retry time on the first unacknowledged
5527 * packet inside the current window */
5529 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5530 /* Don't set timers for packets outside the window */
5531 if (p->header.seq >= call->tfirst + call->twind) {
5535 if (!(p->flags & RX_PKTFLAG_ACKED)
5536 && !clock_IsZero(&p->retryTime)) {
5538 retryTime = p->retryTime;
5543 /* Post a new event to re-run rxi_Start when retries may be needed */
5544 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5545 #ifdef RX_ENABLE_LOCKS
5546 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5548 rxevent_PostNow2(&retryTime, &usenow,
5550 (void *)call, 0, istack);
5551 #else /* RX_ENABLE_LOCKS */
5553 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5554 (void *)call, 0, istack);
5555 #endif /* RX_ENABLE_LOCKS */
5558 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5559 } while (call->flags & RX_CALL_NEED_START);
5561 * TQ references no longer protected by this flag; they must remain
5562 * protected by the global lock.
5564 call->flags &= ~RX_CALL_TQ_BUSY;
5565 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5566 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5567 #ifdef RX_ENABLE_LOCKS
5568 osirx_AssertMine(&call->lock, "rxi_Start end");
5569 CV_BROADCAST(&call->cv_tq);
5570 #else /* RX_ENABLE_LOCKS */
5571 osi_rxWakeup(&call->tq);
5572 #endif /* RX_ENABLE_LOCKS */
5575 call->flags |= RX_CALL_NEED_START;
5577 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5579 if (call->resendEvent) {
5580 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5585 /* Also adjusts the keep alive parameters for the call, to reflect
5586 * that we have just sent a packet (so keep alives aren't sent
5589 rxi_Send(struct rx_call *call, struct rx_packet *p,
5592 struct rx_connection *conn = call->conn;
5594 /* Stamp each packet with the user supplied status */
5595 p->header.userStatus = call->localStatus;
5597 /* Allow the security object controlling this call's security to
5598 * make any last-minute changes to the packet */
5599 RXS_SendPacket(conn->securityObject, call, p);
5601 /* Since we're about to send SOME sort of packet to the peer, it's
5602 * safe to nuke any scheduled end-of-packets ack */
5603 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5605 /* Actually send the packet, filling in more connection-specific fields */
5606 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5607 MUTEX_EXIT(&call->lock);
5608 rxi_SendPacket(call, conn, p, istack);
5609 MUTEX_ENTER(&call->lock);
5610 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5612 /* Update last send time for this call (for keep-alive
5613 * processing), and for the connection (so that we can discover
5614 * idle connections) */
5615 conn->lastSendTime = call->lastSendTime = clock_Sec();
5616 /* Don't count keepalives here, so idleness can be tracked. */
5617 if ((p->header.type != RX_PACKET_TYPE_ACK) || (((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING))
5618 call->lastSendData = call->lastSendTime;
5622 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5623 * that things are fine. Also called periodically to guarantee that nothing
5624 * falls through the cracks (e.g. (error + dally) connections have keepalive
5625 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5627 * haveCTLock Set if calling from rxi_ReapConnections
5629 #ifdef RX_ENABLE_LOCKS
5631 rxi_CheckCall(struct rx_call *call, int haveCTLock)
5632 #else /* RX_ENABLE_LOCKS */
5634 rxi_CheckCall(struct rx_call *call)
5635 #endif /* RX_ENABLE_LOCKS */
5637 struct rx_connection *conn = call->conn;
5639 afs_uint32 deadTime;
5641 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5642 if (call->flags & RX_CALL_TQ_BUSY) {
5643 /* Call is active and will be reset by rxi_Start if it's
5644 * in an error state.
5649 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5651 (((afs_uint32) conn->secondsUntilDead << 10) +
5652 ((afs_uint32) conn->peer->rtt >> 3) +
5653 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5655 /* These are computed to the second (+- 1 second). But that's
5656 * good enough for these values, which should be a significant
5657 * number of seconds. */
5658 if (now > (call->lastReceiveTime + deadTime)) {
5659 if (call->state == RX_STATE_ACTIVE) {
5661 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5663 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5664 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5665 ip_stack_t *ipst = ns->netstack_ip;
5667 ire = ire_cache_lookup(call->conn->peer->host
5668 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5670 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5672 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5679 if (ire && ire->ire_max_frag > 0)
5680 rxi_SetPeerMtu(call->conn->peer->host, 0, ire->ire_max_frag);
5681 #if defined(GLOBAL_NETSTACKID)
5685 #endif /* ADAPT_PMTU */
5686 rxi_CallError(call, RX_CALL_DEAD);
5689 #ifdef RX_ENABLE_LOCKS
5690 /* Cancel pending events */
5691 rxevent_Cancel(call->delayedAckEvent, call,
5692 RX_CALL_REFCOUNT_DELAY);
5693 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5694 rxevent_Cancel(call->keepAliveEvent, call,
5695 RX_CALL_REFCOUNT_ALIVE);
5696 if (call->refCount == 0) {
5697 rxi_FreeCall(call, haveCTLock);
5701 #else /* RX_ENABLE_LOCKS */
5704 #endif /* RX_ENABLE_LOCKS */
5706 /* Non-active calls are destroyed if they are not responding
5707 * to pings; active calls are simply flagged in error, so the
5708 * attached process can die reasonably gracefully. */
5710 /* see if we have a non-activity timeout */
5711 if (call->startWait && conn->idleDeadTime
5712 && ((call->startWait + conn->idleDeadTime) < now)) {
5713 if (call->state == RX_STATE_ACTIVE) {
5714 rxi_CallError(call, RX_CALL_TIMEOUT);
5718 if (call->lastSendData && conn->idleDeadTime && (conn->idleDeadErr != 0)
5719 && ((call->lastSendData + conn->idleDeadTime) < now)) {
5720 if (call->state == RX_STATE_ACTIVE) {
5721 rxi_CallError(call, conn->idleDeadErr);
5725 /* see if we have a hard timeout */
5726 if (conn->hardDeadTime
5727 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5728 if (call->state == RX_STATE_ACTIVE)
5729 rxi_CallError(call, RX_CALL_TIMEOUT);
5736 /* When a call is in progress, this routine is called occasionally to
5737 * make sure that some traffic has arrived (or been sent to) the peer.
5738 * If nothing has arrived in a reasonable amount of time, the call is
5739 * declared dead; if nothing has been sent for a while, we send a
5740 * keep-alive packet (if we're actually trying to keep the call alive)
5743 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
5745 struct rx_call *call = arg1;
5746 struct rx_connection *conn;
5749 MUTEX_ENTER(&call->lock);
5750 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5751 if (event == call->keepAliveEvent)
5752 call->keepAliveEvent = NULL;
5755 #ifdef RX_ENABLE_LOCKS
5756 if (rxi_CheckCall(call, 0)) {
5757 MUTEX_EXIT(&call->lock);
5760 #else /* RX_ENABLE_LOCKS */
5761 if (rxi_CheckCall(call))
5763 #endif /* RX_ENABLE_LOCKS */
5765 /* Don't try to keep alive dallying calls */
5766 if (call->state == RX_STATE_DALLY) {
5767 MUTEX_EXIT(&call->lock);
5772 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5773 /* Don't try to send keepalives if there is unacknowledged data */
5774 /* the rexmit code should be good enough, this little hack
5775 * doesn't quite work XXX */
5776 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5778 rxi_ScheduleKeepAliveEvent(call);
5779 MUTEX_EXIT(&call->lock);
5784 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
5786 if (!call->keepAliveEvent) {
5787 struct clock when, now;
5788 clock_GetTime(&now);
5790 when.sec += call->conn->secondsUntilPing;
5791 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5792 call->keepAliveEvent =
5793 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
5797 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5799 rxi_KeepAliveOn(struct rx_call *call)
5801 /* Pretend last packet received was received now--i.e. if another
5802 * packet isn't received within the keep alive time, then the call
5803 * will die; Initialize last send time to the current time--even
5804 * if a packet hasn't been sent yet. This will guarantee that a
5805 * keep-alive is sent within the ping time */
5806 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5807 rxi_ScheduleKeepAliveEvent(call);
5810 /* This routine is called to send connection abort messages
5811 * that have been delayed to throttle looping clients. */
5813 rxi_SendDelayedConnAbort(struct rxevent *event,
5814 void *arg1, void *unused)
5816 struct rx_connection *conn = arg1;
5819 struct rx_packet *packet;
5821 MUTEX_ENTER(&conn->conn_data_lock);
5822 conn->delayedAbortEvent = NULL;
5823 error = htonl(conn->error);
5825 MUTEX_EXIT(&conn->conn_data_lock);
5826 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5829 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5830 RX_PACKET_TYPE_ABORT, (char *)&error,
5832 rxi_FreePacket(packet);
5836 /* This routine is called to send call abort messages
5837 * that have been delayed to throttle looping clients. */
5839 rxi_SendDelayedCallAbort(struct rxevent *event,
5840 void *arg1, void *dummy)
5842 struct rx_call *call = arg1;
5845 struct rx_packet *packet;
5847 MUTEX_ENTER(&call->lock);
5848 call->delayedAbortEvent = NULL;
5849 error = htonl(call->error);
5851 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5854 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5855 (char *)&error, sizeof(error), 0);
5856 rxi_FreePacket(packet);
5858 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5859 MUTEX_EXIT(&call->lock);
5862 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5863 * seconds) to ask the client to authenticate itself. The routine
5864 * issues a challenge to the client, which is obtained from the
5865 * security object associated with the connection */
5867 rxi_ChallengeEvent(struct rxevent *event,
5868 void *arg0, void *arg1, int tries)
5870 struct rx_connection *conn = arg0;
5872 conn->challengeEvent = NULL;
5873 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5874 struct rx_packet *packet;
5875 struct clock when, now;
5878 /* We've failed to authenticate for too long.
5879 * Reset any calls waiting for authentication;
5880 * they are all in RX_STATE_PRECALL.
5884 MUTEX_ENTER(&conn->conn_call_lock);
5885 for (i = 0; i < RX_MAXCALLS; i++) {
5886 struct rx_call *call = conn->call[i];
5888 MUTEX_ENTER(&call->lock);
5889 if (call->state == RX_STATE_PRECALL) {
5890 rxi_CallError(call, RX_CALL_DEAD);
5891 rxi_SendCallAbort(call, NULL, 0, 0);
5893 MUTEX_EXIT(&call->lock);
5896 MUTEX_EXIT(&conn->conn_call_lock);
5900 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5902 /* If there's no packet available, do this later. */
5903 RXS_GetChallenge(conn->securityObject, conn, packet);
5904 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5905 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5906 rxi_FreePacket(packet);
5908 clock_GetTime(&now);
5910 when.sec += RX_CHALLENGE_TIMEOUT;
5911 conn->challengeEvent =
5912 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
5917 /* Call this routine to start requesting the client to authenticate
5918 * itself. This will continue until authentication is established,
5919 * the call times out, or an invalid response is returned. The
5920 * security object associated with the connection is asked to create
5921 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5922 * defined earlier. */
5924 rxi_ChallengeOn(struct rx_connection *conn)
5926 if (!conn->challengeEvent) {
5927 RXS_CreateChallenge(conn->securityObject, conn);
5928 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5933 /* Compute round trip time of the packet provided, in *rttp.
5936 /* rxi_ComputeRoundTripTime is called with peer locked. */
5937 /* sentp and/or peer may be null */
5939 rxi_ComputeRoundTripTime(struct rx_packet *p,
5940 struct clock *sentp,
5941 struct rx_peer *peer)
5943 struct clock thisRtt, *rttp = &thisRtt;
5947 clock_GetTime(rttp);
5949 if (clock_Lt(rttp, sentp)) {
5951 return; /* somebody set the clock back, don't count this time. */
5953 clock_Sub(rttp, sentp);
5954 if (rx_stats_active) {
5955 MUTEX_ENTER(&rx_stats_mutex);
5956 if (clock_Lt(rttp, &rx_stats.minRtt))
5957 rx_stats.minRtt = *rttp;
5958 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5959 if (rttp->sec > 60) {
5960 MUTEX_EXIT(&rx_stats_mutex);
5961 return; /* somebody set the clock ahead */
5963 rx_stats.maxRtt = *rttp;
5965 clock_Add(&rx_stats.totalRtt, rttp);
5966 rx_stats.nRttSamples++;
5967 MUTEX_EXIT(&rx_stats_mutex);
5970 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5972 /* Apply VanJacobson round-trip estimations */
5977 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5978 * srtt is stored as fixed point with 3 bits after the binary
5979 * point (i.e., scaled by 8). The following magic is
5980 * equivalent to the smoothing algorithm in rfc793 with an
5981 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5982 * srtt*8 = srtt*8 + rtt - srtt
5983 * srtt = srtt + rtt/8 - srtt/8
5986 delta = MSEC(rttp) - (peer->rtt >> 3);
5990 * We accumulate a smoothed rtt variance (actually, a smoothed
5991 * mean difference), then set the retransmit timer to smoothed
5992 * rtt + 4 times the smoothed variance (was 2x in van's original
5993 * paper, but 4x works better for me, and apparently for him as
5995 * rttvar is stored as
5996 * fixed point with 2 bits after the binary point (scaled by
5997 * 4). The following is equivalent to rfc793 smoothing with
5998 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5999 * replaces rfc793's wired-in beta.
6000 * dev*4 = dev*4 + (|actual - expected| - dev)
6006 delta -= (peer->rtt_dev >> 2);
6007 peer->rtt_dev += delta;
6009 /* I don't have a stored RTT so I start with this value. Since I'm
6010 * probably just starting a call, and will be pushing more data down
6011 * this, I expect congestion to increase rapidly. So I fudge a
6012 * little, and I set deviance to half the rtt. In practice,
6013 * deviance tends to approach something a little less than
6014 * half the smoothed rtt. */
6015 peer->rtt = (MSEC(rttp) << 3) + 8;
6016 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
6018 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
6019 * the other of these connections is usually in a user process, and can
6020 * be switched and/or swapped out. So on fast, reliable networks, the
6021 * timeout would otherwise be too short.
6023 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
6024 clock_Zero(&(peer->timeout));
6025 clock_Addmsec(&(peer->timeout), rtt_timeout);
6027 dpf(("rxi_ComputeRoundTripTime(rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n", MSEC(rttp), peer->rtt >> 3, peer->rtt_dev >> 2, (peer->timeout.sec), (peer->timeout.usec)));
6031 /* Find all server connections that have not been active for a long time, and
6034 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6036 struct clock now, when;
6037 clock_GetTime(&now);
6039 /* Find server connection structures that haven't been used for
6040 * greater than rx_idleConnectionTime */
6042 struct rx_connection **conn_ptr, **conn_end;
6043 int i, havecalls = 0;
6044 MUTEX_ENTER(&rx_connHashTable_lock);
6045 for (conn_ptr = &rx_connHashTable[0], conn_end =
6046 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6048 struct rx_connection *conn, *next;
6049 struct rx_call *call;
6053 for (conn = *conn_ptr; conn; conn = next) {
6054 /* XXX -- Shouldn't the connection be locked? */
6057 for (i = 0; i < RX_MAXCALLS; i++) {
6058 call = conn->call[i];
6062 code = MUTEX_TRYENTER(&call->lock);
6065 #ifdef RX_ENABLE_LOCKS
6066 result = rxi_CheckCall(call, 1);
6067 #else /* RX_ENABLE_LOCKS */
6068 result = rxi_CheckCall(call);
6069 #endif /* RX_ENABLE_LOCKS */
6070 MUTEX_EXIT(&call->lock);
6072 /* If CheckCall freed the call, it might
6073 * have destroyed the connection as well,
6074 * which screws up the linked lists.
6080 if (conn->type == RX_SERVER_CONNECTION) {
6081 /* This only actually destroys the connection if
6082 * there are no outstanding calls */
6083 MUTEX_ENTER(&conn->conn_data_lock);
6084 if (!havecalls && !conn->refCount
6085 && ((conn->lastSendTime + rx_idleConnectionTime) <
6087 conn->refCount++; /* it will be decr in rx_DestroyConn */
6088 MUTEX_EXIT(&conn->conn_data_lock);
6089 #ifdef RX_ENABLE_LOCKS
6090 rxi_DestroyConnectionNoLock(conn);
6091 #else /* RX_ENABLE_LOCKS */
6092 rxi_DestroyConnection(conn);
6093 #endif /* RX_ENABLE_LOCKS */
6095 #ifdef RX_ENABLE_LOCKS
6097 MUTEX_EXIT(&conn->conn_data_lock);
6099 #endif /* RX_ENABLE_LOCKS */
6103 #ifdef RX_ENABLE_LOCKS
6104 while (rx_connCleanup_list) {
6105 struct rx_connection *conn;
6106 conn = rx_connCleanup_list;
6107 rx_connCleanup_list = rx_connCleanup_list->next;
6108 MUTEX_EXIT(&rx_connHashTable_lock);
6109 rxi_CleanupConnection(conn);
6110 MUTEX_ENTER(&rx_connHashTable_lock);
6112 MUTEX_EXIT(&rx_connHashTable_lock);
6113 #endif /* RX_ENABLE_LOCKS */
6116 /* Find any peer structures that haven't been used (haven't had an
6117 * associated connection) for greater than rx_idlePeerTime */
6119 struct rx_peer **peer_ptr, **peer_end;
6121 MUTEX_ENTER(&rx_rpc_stats);
6122 MUTEX_ENTER(&rx_peerHashTable_lock);
6123 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6124 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6126 struct rx_peer *peer, *next, *prev;
6127 for (prev = peer = *peer_ptr; peer; peer = next) {
6129 code = MUTEX_TRYENTER(&peer->peer_lock);
6130 if ((code) && (peer->refCount == 0)
6131 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6132 rx_interface_stat_p rpc_stat, nrpc_stat;
6134 MUTEX_EXIT(&peer->peer_lock);
6135 MUTEX_DESTROY(&peer->peer_lock);
6137 (&peer->rpcStats, rpc_stat, nrpc_stat,
6138 rx_interface_stat)) {
6139 unsigned int num_funcs;
6142 queue_Remove(&rpc_stat->queue_header);
6143 queue_Remove(&rpc_stat->all_peers);
6144 num_funcs = rpc_stat->stats[0].func_total;
6146 sizeof(rx_interface_stat_t) +
6147 rpc_stat->stats[0].func_total *
6148 sizeof(rx_function_entry_v1_t);
6150 rxi_Free(rpc_stat, space);
6151 rxi_rpc_peer_stat_cnt -= num_funcs;
6154 if (rx_stats_active)
6155 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6156 if (peer == *peer_ptr) {
6163 MUTEX_EXIT(&peer->peer_lock);
6169 MUTEX_EXIT(&rx_peerHashTable_lock);
6170 MUTEX_EXIT(&rx_rpc_stats);
6173 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6174 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6175 * GC, just below. Really, we shouldn't have to keep moving packets from
6176 * one place to another, but instead ought to always know if we can
6177 * afford to hold onto a packet in its particular use. */
6178 MUTEX_ENTER(&rx_freePktQ_lock);
6179 if (rx_waitingForPackets) {
6180 rx_waitingForPackets = 0;
6181 #ifdef RX_ENABLE_LOCKS
6182 CV_BROADCAST(&rx_waitingForPackets_cv);
6184 osi_rxWakeup(&rx_waitingForPackets);
6187 MUTEX_EXIT(&rx_freePktQ_lock);
6190 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6191 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6195 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6196 * rx.h is sort of strange this is better. This is called with a security
6197 * object before it is discarded. Each connection using a security object has
6198 * its own refcount to the object so it won't actually be freed until the last
6199 * connection is destroyed.
6201 * This is the only rxs module call. A hold could also be written but no one
6205 rxs_Release(struct rx_securityClass *aobj)
6207 return RXS_Close(aobj);
6211 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6212 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6213 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6214 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6216 /* Adjust our estimate of the transmission rate to this peer, given
6217 * that the packet p was just acked. We can adjust peer->timeout and
6218 * call->twind. Pragmatically, this is called
6219 * only with packets of maximal length.
6220 * Called with peer and call locked.
6224 rxi_ComputeRate(struct rx_peer *peer, struct rx_call *call,
6225 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6227 afs_int32 xferSize, xferMs;
6231 /* Count down packets */
6232 if (peer->rateFlag > 0)
6234 /* Do nothing until we're enabled */
6235 if (peer->rateFlag != 0)
6240 /* Count only when the ack seems legitimate */
6241 switch (ackReason) {
6242 case RX_ACK_REQUESTED:
6244 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6248 case RX_ACK_PING_RESPONSE:
6249 if (p) /* want the response to ping-request, not data send */
6251 clock_GetTime(&newTO);
6252 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6253 clock_Sub(&newTO, &call->pingRequestTime);
6254 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6258 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6265 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %d.%06d, rtt %u, ps %u)", ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"), xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
6267 /* Track only packets that are big enough. */
6268 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6272 /* absorb RTT data (in milliseconds) for these big packets */
6273 if (peer->smRtt == 0) {
6274 peer->smRtt = xferMs;
6276 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6281 if (peer->countDown) {
6285 peer->countDown = 10; /* recalculate only every so often */
6287 /* In practice, we can measure only the RTT for full packets,
6288 * because of the way Rx acks the data that it receives. (If it's
6289 * smaller than a full packet, it often gets implicitly acked
6290 * either by the call response (from a server) or by the next call
6291 * (from a client), and either case confuses transmission times
6292 * with processing times.) Therefore, replace the above
6293 * more-sophisticated processing with a simpler version, where the
6294 * smoothed RTT is kept for full-size packets, and the time to
6295 * transmit a windowful of full-size packets is simply RTT *
6296 * windowSize. Again, we take two steps:
6297 - ensure the timeout is large enough for a single packet's RTT;
6298 - ensure that the window is small enough to fit in the desired timeout.*/
6300 /* First, the timeout check. */
6301 minTime = peer->smRtt;
6302 /* Get a reasonable estimate for a timeout period */
6304 newTO.sec = minTime / 1000;
6305 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6307 /* Increase the timeout period so that we can always do at least
6308 * one packet exchange */
6309 if (clock_Gt(&newTO, &peer->timeout)) {
6311 dpf(("CONG peer %lx/%u: timeout %d.%06d ==> %ld.%06d (rtt %u, ps %u)", ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec, newTO.sec, newTO.usec, peer->smRtt, peer->packetSize));
6313 peer->timeout = newTO;
6316 /* Now, get an estimate for the transmit window size. */
6317 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6318 /* Now, convert to the number of full packets that could fit in a
6319 * reasonable fraction of that interval */
6320 minTime /= (peer->smRtt << 1);
6321 xferSize = minTime; /* (make a copy) */
6323 /* Now clamp the size to reasonable bounds. */
6326 else if (minTime > rx_Window)
6327 minTime = rx_Window;
6328 /* if (minTime != peer->maxWindow) {
6329 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6330 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6331 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6333 peer->maxWindow = minTime;
6334 elide... call->twind = minTime;
6338 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6339 * Discern this by calculating the timeout necessary for rx_Window
6341 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6342 /* calculate estimate for transmission interval in milliseconds */
6343 minTime = rx_Window * peer->smRtt;
6344 if (minTime < 1000) {
6345 dpf(("CONG peer %lx/%u: cut TO %d.%06d by 0.5 (rtt %u, ps %u)",
6346 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6347 peer->timeout.usec, peer->smRtt, peer->packetSize));
6349 newTO.sec = 0; /* cut back on timeout by half a second */
6350 newTO.usec = 500000;
6351 clock_Sub(&peer->timeout, &newTO);
6356 } /* end of rxi_ComputeRate */
6357 #endif /* ADAPT_WINDOW */
6365 #define TRACE_OPTION_RX_DEBUG 16
6373 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6374 0, KEY_QUERY_VALUE, &parmKey);
6375 if (code != ERROR_SUCCESS)
6378 dummyLen = sizeof(TraceOption);
6379 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6380 (BYTE *) &TraceOption, &dummyLen);
6381 if (code == ERROR_SUCCESS) {
6382 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
6384 RegCloseKey (parmKey);
6385 #endif /* AFS_NT40_ENV */
6390 rx_DebugOnOff(int on)
6394 rxdebug_active = on;
6400 rx_StatsOnOff(int on)
6403 rx_stats_active = on;
6408 /* Don't call this debugging routine directly; use dpf */
6410 rxi_DebugPrint(char *format, ...)
6419 va_start(ap, format);
6421 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6424 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
6426 if (msg[len-1] != '\n') {
6430 OutputDebugString(msg);
6437 va_start(ap, format);
6439 clock_GetTime(&now);
6440 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
6441 (unsigned int)now.usec);
6442 vfprintf(rx_Log, format, ap);
6451 * This function is used to process the rx_stats structure that is local
6452 * to a process as well as an rx_stats structure received from a remote
6453 * process (via rxdebug). Therefore, it needs to do minimal version
6457 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
6458 afs_int32 freePackets, char version)
6463 if (size != sizeof(struct rx_statistics)) {
6465 "Unexpected size of stats structure: was %d, expected %lud\n",
6466 size, sizeof(struct rx_statistics));
6469 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6472 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6473 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
6474 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6475 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6476 s->specialPktAllocFailures);
6478 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
6479 s->receivePktAllocFailures, s->sendPktAllocFailures,
6480 s->specialPktAllocFailures);
6484 " greedy %u, " "bogusReads %u (last from host %x), "
6485 "noPackets %u, " "noBuffers %u, " "selects %u, "
6486 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
6487 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6488 s->selects, s->sendSelects);
6490 fprintf(file, " packets read: ");
6491 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6492 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
6494 fprintf(file, "\n");
6497 " other read counters: data %u, " "ack %u, " "dup %u "
6498 "spurious %u " "dally %u\n", s->dataPacketsRead,
6499 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6500 s->ignorePacketDally);
6502 fprintf(file, " packets sent: ");
6503 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6504 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
6506 fprintf(file, "\n");
6509 " other send counters: ack %u, " "data %u (not resends), "
6510 "resends %u, " "pushed %u, " "acked&ignored %u\n",
6511 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6512 s->dataPacketsPushed, s->ignoreAckedPacket);
6515 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
6516 s->netSendFailures, (int)s->fatalErrors);
6518 if (s->nRttSamples) {
6519 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6520 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6522 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6523 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6527 " %d server connections, " "%d client connections, "
6528 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6529 s->nServerConns, s->nClientConns, s->nPeerStructs,
6530 s->nCallStructs, s->nFreeCallStructs);
6532 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6533 fprintf(file, " %d clock updates\n", clock_nUpdates);
6536 fprintf(file, "ERROR: compiled without RXDEBUG\n");
6540 /* for backward compatibility */
6542 rx_PrintStats(FILE * file)
6544 MUTEX_ENTER(&rx_stats_mutex);
6545 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6547 MUTEX_EXIT(&rx_stats_mutex);
6551 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6553 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
6554 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6555 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6558 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6559 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6560 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6563 " Packet size %d, " "max in packet skew %d, "
6564 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6565 (int)peer->outPacketSkew);
6569 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
6571 * This mutex protects the following static variables:
6575 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
6576 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
6578 #define LOCK_RX_DEBUG
6579 #define UNLOCK_RX_DEBUG
6580 #endif /* AFS_PTHREAD_ENV */
6584 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6585 u_char type, void *inputData, size_t inputLength,
6586 void *outputData, size_t outputLength)
6588 static afs_int32 counter = 100;
6589 time_t waitTime, waitCount, startTime;
6590 struct rx_header theader;
6593 struct timeval tv_now, tv_wake, tv_delta;
6594 struct sockaddr_in taddr, faddr;
6603 startTime = time(0);
6609 tp = &tbuffer[sizeof(struct rx_header)];
6610 taddr.sin_family = AF_INET;
6611 taddr.sin_port = remotePort;
6612 taddr.sin_addr.s_addr = remoteAddr;
6613 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6614 taddr.sin_len = sizeof(struct sockaddr_in);
6617 memset(&theader, 0, sizeof(theader));
6618 theader.epoch = htonl(999);
6620 theader.callNumber = htonl(counter);
6623 theader.type = type;
6624 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6625 theader.serviceId = 0;
6627 memcpy(tbuffer, &theader, sizeof(theader));
6628 memcpy(tp, inputData, inputLength);
6630 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6631 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6633 /* see if there's a packet available */
6634 gettimeofday(&tv_wake,0);
6635 tv_wake.tv_sec += waitTime;
6638 FD_SET(socket, &imask);
6639 tv_delta.tv_sec = tv_wake.tv_sec;
6640 tv_delta.tv_usec = tv_wake.tv_usec;
6641 gettimeofday(&tv_now, 0);
6643 if (tv_delta.tv_usec < tv_now.tv_usec) {
6645 tv_delta.tv_usec += 1000000;
6648 tv_delta.tv_usec -= tv_now.tv_usec;
6650 if (tv_delta.tv_sec < tv_now.tv_sec) {
6654 tv_delta.tv_sec -= tv_now.tv_sec;
6656 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6657 if (code == 1 && FD_ISSET(socket, &imask)) {
6658 /* now receive a packet */
6659 faddrLen = sizeof(struct sockaddr_in);
6661 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6662 (struct sockaddr *)&faddr, &faddrLen);
6665 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6666 if (counter == ntohl(theader.callNumber))
6674 /* see if we've timed out */
6682 code -= sizeof(struct rx_header);
6683 if (code > outputLength)
6684 code = outputLength;
6685 memcpy(outputData, tp, code);
6688 #endif /* RXDEBUG */
6691 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6692 afs_uint16 remotePort, struct rx_debugStats * stat,
6693 afs_uint32 * supportedValues)
6699 struct rx_debugIn in;
6701 *supportedValues = 0;
6702 in.type = htonl(RX_DEBUGI_GETSTATS);
6705 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6706 &in, sizeof(in), stat, sizeof(*stat));
6709 * If the call was successful, fixup the version and indicate
6710 * what contents of the stat structure are valid.
6711 * Also do net to host conversion of fields here.
6715 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6716 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6718 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6719 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6721 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6722 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6724 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6725 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6727 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6728 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6730 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6731 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6733 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6734 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6736 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6737 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6739 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
6740 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
6742 stat->nFreePackets = ntohl(stat->nFreePackets);
6743 stat->packetReclaims = ntohl(stat->packetReclaims);
6744 stat->callsExecuted = ntohl(stat->callsExecuted);
6745 stat->nWaiting = ntohl(stat->nWaiting);
6746 stat->idleThreads = ntohl(stat->idleThreads);
6747 stat->nWaited = ntohl(stat->nWaited);
6748 stat->nPackets = ntohl(stat->nPackets);
6755 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6756 afs_uint16 remotePort, struct rx_statistics * stat,
6757 afs_uint32 * supportedValues)
6763 struct rx_debugIn in;
6764 afs_int32 *lp = (afs_int32 *) stat;
6768 * supportedValues is currently unused, but added to allow future
6769 * versioning of this function.
6772 *supportedValues = 0;
6773 in.type = htonl(RX_DEBUGI_RXSTATS);
6775 memset(stat, 0, sizeof(*stat));
6777 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6778 &in, sizeof(in), stat, sizeof(*stat));
6783 * Do net to host conversion here
6786 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6795 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6796 afs_uint16 remotePort, size_t version_length,
6801 return MakeDebugCall(socket, remoteAddr, remotePort,
6802 RX_PACKET_TYPE_VERSION, a, 1, version,
6810 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6811 afs_uint16 remotePort, afs_int32 * nextConnection,
6812 int allConnections, afs_uint32 debugSupportedValues,
6813 struct rx_debugConn * conn,
6814 afs_uint32 * supportedValues)
6820 struct rx_debugIn in;
6824 * supportedValues is currently unused, but added to allow future
6825 * versioning of this function.
6828 *supportedValues = 0;
6829 if (allConnections) {
6830 in.type = htonl(RX_DEBUGI_GETALLCONN);
6832 in.type = htonl(RX_DEBUGI_GETCONN);
6834 in.index = htonl(*nextConnection);
6835 memset(conn, 0, sizeof(*conn));
6837 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6838 &in, sizeof(in), conn, sizeof(*conn));
6841 *nextConnection += 1;
6844 * Convert old connection format to new structure.
6847 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6848 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6849 #define MOVEvL(a) (conn->a = vL->a)
6851 /* any old or unrecognized version... */
6852 for (i = 0; i < RX_MAXCALLS; i++) {
6853 MOVEvL(callState[i]);
6854 MOVEvL(callMode[i]);
6855 MOVEvL(callFlags[i]);
6856 MOVEvL(callOther[i]);
6858 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6859 MOVEvL(secStats.type);
6860 MOVEvL(secStats.level);
6861 MOVEvL(secStats.flags);
6862 MOVEvL(secStats.expires);
6863 MOVEvL(secStats.packetsReceived);
6864 MOVEvL(secStats.packetsSent);
6865 MOVEvL(secStats.bytesReceived);
6866 MOVEvL(secStats.bytesSent);
6871 * Do net to host conversion here
6873 * I don't convert host or port since we are most likely
6874 * going to want these in NBO.
6876 conn->cid = ntohl(conn->cid);
6877 conn->serial = ntohl(conn->serial);
6878 for (i = 0; i < RX_MAXCALLS; i++) {
6879 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6881 conn->error = ntohl(conn->error);
6882 conn->secStats.flags = ntohl(conn->secStats.flags);
6883 conn->secStats.expires = ntohl(conn->secStats.expires);
6884 conn->secStats.packetsReceived =
6885 ntohl(conn->secStats.packetsReceived);
6886 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6887 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6888 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6889 conn->epoch = ntohl(conn->epoch);
6890 conn->natMTU = ntohl(conn->natMTU);
6897 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6898 afs_uint16 remotePort, afs_int32 * nextPeer,
6899 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6900 afs_uint32 * supportedValues)
6906 struct rx_debugIn in;
6909 * supportedValues is currently unused, but added to allow future
6910 * versioning of this function.
6913 *supportedValues = 0;
6914 in.type = htonl(RX_DEBUGI_GETPEER);
6915 in.index = htonl(*nextPeer);
6916 memset(peer, 0, sizeof(*peer));
6918 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6919 &in, sizeof(in), peer, sizeof(*peer));
6925 * Do net to host conversion here
6927 * I don't convert host or port since we are most likely
6928 * going to want these in NBO.
6930 peer->ifMTU = ntohs(peer->ifMTU);
6931 peer->idleWhen = ntohl(peer->idleWhen);
6932 peer->refCount = ntohs(peer->refCount);
6933 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6934 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6935 peer->rtt = ntohl(peer->rtt);
6936 peer->rtt_dev = ntohl(peer->rtt_dev);
6937 peer->timeout.sec = ntohl(peer->timeout.sec);
6938 peer->timeout.usec = ntohl(peer->timeout.usec);
6939 peer->nSent = ntohl(peer->nSent);
6940 peer->reSends = ntohl(peer->reSends);
6941 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6942 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6943 peer->rateFlag = ntohl(peer->rateFlag);
6944 peer->natMTU = ntohs(peer->natMTU);
6945 peer->maxMTU = ntohs(peer->maxMTU);
6946 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6947 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6948 peer->MTU = ntohs(peer->MTU);
6949 peer->cwind = ntohs(peer->cwind);
6950 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6951 peer->congestSeq = ntohs(peer->congestSeq);
6952 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6953 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6954 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6955 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6962 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
6963 struct rx_debugPeer * peerStats)
6966 afs_int32 error = 1; /* default to "did not succeed" */
6967 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
6969 MUTEX_ENTER(&rx_peerHashTable_lock);
6970 for(tp = rx_peerHashTable[hashValue];
6971 tp != NULL; tp = tp->next) {
6972 if (tp->host == peerHost)
6979 peerStats->host = tp->host;
6980 peerStats->port = tp->port;
6981 peerStats->ifMTU = tp->ifMTU;
6982 peerStats->idleWhen = tp->idleWhen;
6983 peerStats->refCount = tp->refCount;
6984 peerStats->burstSize = tp->burstSize;
6985 peerStats->burst = tp->burst;
6986 peerStats->burstWait.sec = tp->burstWait.sec;
6987 peerStats->burstWait.usec = tp->burstWait.usec;
6988 peerStats->rtt = tp->rtt;
6989 peerStats->rtt_dev = tp->rtt_dev;
6990 peerStats->timeout.sec = tp->timeout.sec;
6991 peerStats->timeout.usec = tp->timeout.usec;
6992 peerStats->nSent = tp->nSent;
6993 peerStats->reSends = tp->reSends;
6994 peerStats->inPacketSkew = tp->inPacketSkew;
6995 peerStats->outPacketSkew = tp->outPacketSkew;
6996 peerStats->rateFlag = tp->rateFlag;
6997 peerStats->natMTU = tp->natMTU;
6998 peerStats->maxMTU = tp->maxMTU;
6999 peerStats->maxDgramPackets = tp->maxDgramPackets;
7000 peerStats->ifDgramPackets = tp->ifDgramPackets;
7001 peerStats->MTU = tp->MTU;
7002 peerStats->cwind = tp->cwind;
7003 peerStats->nDgramPackets = tp->nDgramPackets;
7004 peerStats->congestSeq = tp->congestSeq;
7005 peerStats->bytesSent.high = tp->bytesSent.high;
7006 peerStats->bytesSent.low = tp->bytesSent.low;
7007 peerStats->bytesReceived.high = tp->bytesReceived.high;
7008 peerStats->bytesReceived.low = tp->bytesReceived.low;
7010 MUTEX_EXIT(&rx_peerHashTable_lock);
7018 struct rx_serverQueueEntry *np;
7021 struct rx_call *call;
7022 struct rx_serverQueueEntry *sq;
7026 if (rxinit_status == 1) {
7028 return; /* Already shutdown. */
7032 #ifndef AFS_PTHREAD_ENV
7033 FD_ZERO(&rx_selectMask);
7034 #endif /* AFS_PTHREAD_ENV */
7035 rxi_dataQuota = RX_MAX_QUOTA;
7036 #ifndef AFS_PTHREAD_ENV
7038 #endif /* AFS_PTHREAD_ENV */
7041 #ifndef AFS_PTHREAD_ENV
7042 #ifndef AFS_USE_GETTIMEOFDAY
7044 #endif /* AFS_USE_GETTIMEOFDAY */
7045 #endif /* AFS_PTHREAD_ENV */
7047 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7048 call = queue_First(&rx_freeCallQueue, rx_call);
7050 rxi_Free(call, sizeof(struct rx_call));
7053 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7054 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7060 struct rx_peer **peer_ptr, **peer_end;
7061 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7062 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7064 struct rx_peer *peer, *next;
7065 for (peer = *peer_ptr; peer; peer = next) {
7066 rx_interface_stat_p rpc_stat, nrpc_stat;
7069 (&peer->rpcStats, rpc_stat, nrpc_stat,
7070 rx_interface_stat)) {
7071 unsigned int num_funcs;
7074 queue_Remove(&rpc_stat->queue_header);
7075 queue_Remove(&rpc_stat->all_peers);
7076 num_funcs = rpc_stat->stats[0].func_total;
7078 sizeof(rx_interface_stat_t) +
7079 rpc_stat->stats[0].func_total *
7080 sizeof(rx_function_entry_v1_t);
7082 rxi_Free(rpc_stat, space);
7083 MUTEX_ENTER(&rx_rpc_stats);
7084 rxi_rpc_peer_stat_cnt -= num_funcs;
7085 MUTEX_EXIT(&rx_rpc_stats);
7089 if (rx_stats_active)
7090 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
7094 for (i = 0; i < RX_MAX_SERVICES; i++) {
7096 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7098 for (i = 0; i < rx_hashTableSize; i++) {
7099 struct rx_connection *tc, *ntc;
7100 MUTEX_ENTER(&rx_connHashTable_lock);
7101 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7103 for (j = 0; j < RX_MAXCALLS; j++) {
7105 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7108 rxi_Free(tc, sizeof(*tc));
7110 MUTEX_EXIT(&rx_connHashTable_lock);
7113 MUTEX_ENTER(&freeSQEList_lock);
7115 while ((np = rx_FreeSQEList)) {
7116 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7117 MUTEX_DESTROY(&np->lock);
7118 rxi_Free(np, sizeof(*np));
7121 MUTEX_EXIT(&freeSQEList_lock);
7122 MUTEX_DESTROY(&freeSQEList_lock);
7123 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7124 MUTEX_DESTROY(&rx_connHashTable_lock);
7125 MUTEX_DESTROY(&rx_peerHashTable_lock);
7126 MUTEX_DESTROY(&rx_serverPool_lock);
7128 osi_Free(rx_connHashTable,
7129 rx_hashTableSize * sizeof(struct rx_connection *));
7130 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7132 UNPIN(rx_connHashTable,
7133 rx_hashTableSize * sizeof(struct rx_connection *));
7134 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7136 rxi_FreeAllPackets();
7138 MUTEX_ENTER(&rx_quota_mutex);
7139 rxi_dataQuota = RX_MAX_QUOTA;
7140 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7141 MUTEX_EXIT(&rx_quota_mutex);
7146 #ifdef RX_ENABLE_LOCKS
7148 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7150 if (!MUTEX_ISMINE(lockaddr))
7151 osi_Panic("Lock not held: %s", msg);
7153 #endif /* RX_ENABLE_LOCKS */
7158 * Routines to implement connection specific data.
7162 rx_KeyCreate(rx_destructor_t rtn)
7165 MUTEX_ENTER(&rxi_keyCreate_lock);
7166 key = rxi_keyCreate_counter++;
7167 rxi_keyCreate_destructor = (rx_destructor_t *)
7168 realloc((void *)rxi_keyCreate_destructor,
7169 (key + 1) * sizeof(rx_destructor_t));
7170 rxi_keyCreate_destructor[key] = rtn;
7171 MUTEX_EXIT(&rxi_keyCreate_lock);
7176 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7179 MUTEX_ENTER(&conn->conn_data_lock);
7180 if (!conn->specific) {
7181 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7182 for (i = 0; i < key; i++)
7183 conn->specific[i] = NULL;
7184 conn->nSpecific = key + 1;
7185 conn->specific[key] = ptr;
7186 } else if (key >= conn->nSpecific) {
7187 conn->specific = (void **)
7188 realloc(conn->specific, (key + 1) * sizeof(void *));
7189 for (i = conn->nSpecific; i < key; i++)
7190 conn->specific[i] = NULL;
7191 conn->nSpecific = key + 1;
7192 conn->specific[key] = ptr;
7194 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7195 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7196 conn->specific[key] = ptr;
7198 MUTEX_EXIT(&conn->conn_data_lock);
7202 rx_GetSpecific(struct rx_connection *conn, int key)
7205 MUTEX_ENTER(&conn->conn_data_lock);
7206 if (key >= conn->nSpecific)
7209 ptr = conn->specific[key];
7210 MUTEX_EXIT(&conn->conn_data_lock);
7214 #endif /* !KERNEL */
7217 * processStats is a queue used to store the statistics for the local
7218 * process. Its contents are similar to the contents of the rpcStats
7219 * queue on a rx_peer structure, but the actual data stored within
7220 * this queue contains totals across the lifetime of the process (assuming
7221 * the stats have not been reset) - unlike the per peer structures
7222 * which can come and go based upon the peer lifetime.
7225 static struct rx_queue processStats = { &processStats, &processStats };
7228 * peerStats is a queue used to store the statistics for all peer structs.
7229 * Its contents are the union of all the peer rpcStats queues.
7232 static struct rx_queue peerStats = { &peerStats, &peerStats };
7235 * rxi_monitor_processStats is used to turn process wide stat collection
7239 static int rxi_monitor_processStats = 0;
7242 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7245 static int rxi_monitor_peerStats = 0;
7248 * rxi_AddRpcStat - given all of the information for a particular rpc
7249 * call, create (if needed) and update the stat totals for the rpc.
7253 * IN stats - the queue of stats that will be updated with the new value
7255 * IN rxInterface - a unique number that identifies the rpc interface
7257 * IN currentFunc - the index of the function being invoked
7259 * IN totalFunc - the total number of functions in this interface
7261 * IN queueTime - the amount of time this function waited for a thread
7263 * IN execTime - the amount of time this function invocation took to execute
7265 * IN bytesSent - the number bytes sent by this invocation
7267 * IN bytesRcvd - the number bytes received by this invocation
7269 * IN isServer - if true, this invocation was made to a server
7271 * IN remoteHost - the ip address of the remote host
7273 * IN remotePort - the port of the remote host
7275 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7277 * INOUT counter - if a new stats structure is allocated, the counter will
7278 * be updated with the new number of allocated stat structures
7286 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
7287 afs_uint32 currentFunc, afs_uint32 totalFunc,
7288 struct clock *queueTime, struct clock *execTime,
7289 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
7290 afs_uint32 remoteHost, afs_uint32 remotePort,
7291 int addToPeerList, unsigned int *counter)
7294 rx_interface_stat_p rpc_stat, nrpc_stat;
7297 * See if there's already a structure for this interface
7300 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7301 if ((rpc_stat->stats[0].interfaceId == rxInterface)
7302 && (rpc_stat->stats[0].remote_is_server == isServer))
7307 * Didn't find a match so allocate a new structure and add it to the
7311 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
7312 || (rpc_stat->stats[0].interfaceId != rxInterface)
7313 || (rpc_stat->stats[0].remote_is_server != isServer)) {
7318 sizeof(rx_interface_stat_t) +
7319 totalFunc * sizeof(rx_function_entry_v1_t);
7321 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
7322 if (rpc_stat == NULL) {
7326 *counter += totalFunc;
7327 for (i = 0; i < totalFunc; i++) {
7328 rpc_stat->stats[i].remote_peer = remoteHost;
7329 rpc_stat->stats[i].remote_port = remotePort;
7330 rpc_stat->stats[i].remote_is_server = isServer;
7331 rpc_stat->stats[i].interfaceId = rxInterface;
7332 rpc_stat->stats[i].func_total = totalFunc;
7333 rpc_stat->stats[i].func_index = i;
7334 hzero(rpc_stat->stats[i].invocations);
7335 hzero(rpc_stat->stats[i].bytes_sent);
7336 hzero(rpc_stat->stats[i].bytes_rcvd);
7337 rpc_stat->stats[i].queue_time_sum.sec = 0;
7338 rpc_stat->stats[i].queue_time_sum.usec = 0;
7339 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7340 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7341 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7342 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7343 rpc_stat->stats[i].queue_time_max.sec = 0;
7344 rpc_stat->stats[i].queue_time_max.usec = 0;
7345 rpc_stat->stats[i].execution_time_sum.sec = 0;
7346 rpc_stat->stats[i].execution_time_sum.usec = 0;
7347 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7348 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7349 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7350 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7351 rpc_stat->stats[i].execution_time_max.sec = 0;
7352 rpc_stat->stats[i].execution_time_max.usec = 0;
7354 queue_Prepend(stats, rpc_stat);
7355 if (addToPeerList) {
7356 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7361 * Increment the stats for this function
7364 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7365 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7366 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7367 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7368 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7369 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7370 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7372 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7373 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7375 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7376 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7378 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7379 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7381 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7382 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7390 * rx_IncrementTimeAndCount - increment the times and count for a particular
7395 * IN peer - the peer who invoked the rpc
7397 * IN rxInterface - a unique number that identifies the rpc interface
7399 * IN currentFunc - the index of the function being invoked
7401 * IN totalFunc - the total number of functions in this interface
7403 * IN queueTime - the amount of time this function waited for a thread
7405 * IN execTime - the amount of time this function invocation took to execute
7407 * IN bytesSent - the number bytes sent by this invocation
7409 * IN bytesRcvd - the number bytes received by this invocation
7411 * IN isServer - if true, this invocation was made to a server
7419 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7420 afs_uint32 currentFunc, afs_uint32 totalFunc,
7421 struct clock *queueTime, struct clock *execTime,
7422 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7426 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7429 MUTEX_ENTER(&rx_rpc_stats);
7430 MUTEX_ENTER(&peer->peer_lock);
7432 if (rxi_monitor_peerStats) {
7433 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7434 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7435 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7438 if (rxi_monitor_processStats) {
7439 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7440 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7441 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7444 MUTEX_EXIT(&peer->peer_lock);
7445 MUTEX_EXIT(&rx_rpc_stats);
7450 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7454 * IN callerVersion - the rpc stat version of the caller.
7456 * IN count - the number of entries to marshall.
7458 * IN stats - pointer to stats to be marshalled.
7460 * OUT ptr - Where to store the marshalled data.
7467 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7468 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7474 * We only support the first version
7476 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7477 *(ptr++) = stats->remote_peer;
7478 *(ptr++) = stats->remote_port;
7479 *(ptr++) = stats->remote_is_server;
7480 *(ptr++) = stats->interfaceId;
7481 *(ptr++) = stats->func_total;
7482 *(ptr++) = stats->func_index;
7483 *(ptr++) = hgethi(stats->invocations);
7484 *(ptr++) = hgetlo(stats->invocations);
7485 *(ptr++) = hgethi(stats->bytes_sent);
7486 *(ptr++) = hgetlo(stats->bytes_sent);
7487 *(ptr++) = hgethi(stats->bytes_rcvd);
7488 *(ptr++) = hgetlo(stats->bytes_rcvd);
7489 *(ptr++) = stats->queue_time_sum.sec;
7490 *(ptr++) = stats->queue_time_sum.usec;
7491 *(ptr++) = stats->queue_time_sum_sqr.sec;
7492 *(ptr++) = stats->queue_time_sum_sqr.usec;
7493 *(ptr++) = stats->queue_time_min.sec;
7494 *(ptr++) = stats->queue_time_min.usec;
7495 *(ptr++) = stats->queue_time_max.sec;
7496 *(ptr++) = stats->queue_time_max.usec;
7497 *(ptr++) = stats->execution_time_sum.sec;
7498 *(ptr++) = stats->execution_time_sum.usec;
7499 *(ptr++) = stats->execution_time_sum_sqr.sec;
7500 *(ptr++) = stats->execution_time_sum_sqr.usec;
7501 *(ptr++) = stats->execution_time_min.sec;
7502 *(ptr++) = stats->execution_time_min.usec;
7503 *(ptr++) = stats->execution_time_max.sec;
7504 *(ptr++) = stats->execution_time_max.usec;
7510 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7515 * IN callerVersion - the rpc stat version of the caller
7517 * OUT myVersion - the rpc stat version of this function
7519 * OUT clock_sec - local time seconds
7521 * OUT clock_usec - local time microseconds
7523 * OUT allocSize - the number of bytes allocated to contain stats
7525 * OUT statCount - the number stats retrieved from this process.
7527 * OUT stats - the actual stats retrieved from this process.
7531 * Returns void. If successful, stats will != NULL.
7535 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7536 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7537 size_t * allocSize, afs_uint32 * statCount,
7538 afs_uint32 ** stats)
7548 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7551 * Check to see if stats are enabled
7554 MUTEX_ENTER(&rx_rpc_stats);
7555 if (!rxi_monitor_processStats) {
7556 MUTEX_EXIT(&rx_rpc_stats);
7560 clock_GetTime(&now);
7561 *clock_sec = now.sec;
7562 *clock_usec = now.usec;
7565 * Allocate the space based upon the caller version
7567 * If the client is at an older version than we are,
7568 * we return the statistic data in the older data format, but
7569 * we still return our version number so the client knows we
7570 * are maintaining more data than it can retrieve.
7573 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7574 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7575 *statCount = rxi_rpc_process_stat_cnt;
7578 * This can't happen yet, but in the future version changes
7579 * can be handled by adding additional code here
7583 if (space > (size_t) 0) {
7585 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7588 rx_interface_stat_p rpc_stat, nrpc_stat;
7592 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7594 * Copy the data based upon the caller version
7596 rx_MarshallProcessRPCStats(callerVersion,
7597 rpc_stat->stats[0].func_total,
7598 rpc_stat->stats, &ptr);
7604 MUTEX_EXIT(&rx_rpc_stats);
7609 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7613 * IN callerVersion - the rpc stat version of the caller
7615 * OUT myVersion - the rpc stat version of this function
7617 * OUT clock_sec - local time seconds
7619 * OUT clock_usec - local time microseconds
7621 * OUT allocSize - the number of bytes allocated to contain stats
7623 * OUT statCount - the number of stats retrieved from the individual
7626 * OUT stats - the actual stats retrieved from the individual peer structures.
7630 * Returns void. If successful, stats will != NULL.
7634 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7635 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7636 size_t * allocSize, afs_uint32 * statCount,
7637 afs_uint32 ** stats)
7647 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7650 * Check to see if stats are enabled
7653 MUTEX_ENTER(&rx_rpc_stats);
7654 if (!rxi_monitor_peerStats) {
7655 MUTEX_EXIT(&rx_rpc_stats);
7659 clock_GetTime(&now);
7660 *clock_sec = now.sec;
7661 *clock_usec = now.usec;
7664 * Allocate the space based upon the caller version
7666 * If the client is at an older version than we are,
7667 * we return the statistic data in the older data format, but
7668 * we still return our version number so the client knows we
7669 * are maintaining more data than it can retrieve.
7672 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7673 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7674 *statCount = rxi_rpc_peer_stat_cnt;
7677 * This can't happen yet, but in the future version changes
7678 * can be handled by adding additional code here
7682 if (space > (size_t) 0) {
7684 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7687 rx_interface_stat_p rpc_stat, nrpc_stat;
7691 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7693 * We have to fix the offset of rpc_stat since we are
7694 * keeping this structure on two rx_queues. The rx_queue
7695 * package assumes that the rx_queue member is the first
7696 * member of the structure. That is, rx_queue assumes that
7697 * any one item is only on one queue at a time. We are
7698 * breaking that assumption and so we have to do a little
7699 * math to fix our pointers.
7702 fix_offset = (char *)rpc_stat;
7703 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7704 rpc_stat = (rx_interface_stat_p) fix_offset;
7707 * Copy the data based upon the caller version
7709 rx_MarshallProcessRPCStats(callerVersion,
7710 rpc_stat->stats[0].func_total,
7711 rpc_stat->stats, &ptr);
7717 MUTEX_EXIT(&rx_rpc_stats);
7722 * rx_FreeRPCStats - free memory allocated by
7723 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7727 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7728 * rx_RetrievePeerRPCStats
7730 * IN allocSize - the number of bytes in stats.
7738 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7740 rxi_Free(stats, allocSize);
7744 * rx_queryProcessRPCStats - see if process rpc stat collection is
7745 * currently enabled.
7751 * Returns 0 if stats are not enabled != 0 otherwise
7755 rx_queryProcessRPCStats(void)
7758 MUTEX_ENTER(&rx_rpc_stats);
7759 rc = rxi_monitor_processStats;
7760 MUTEX_EXIT(&rx_rpc_stats);
7765 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7771 * Returns 0 if stats are not enabled != 0 otherwise
7775 rx_queryPeerRPCStats(void)
7778 MUTEX_ENTER(&rx_rpc_stats);
7779 rc = rxi_monitor_peerStats;
7780 MUTEX_EXIT(&rx_rpc_stats);
7785 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7795 rx_enableProcessRPCStats(void)
7797 MUTEX_ENTER(&rx_rpc_stats);
7798 rx_enable_stats = 1;
7799 rxi_monitor_processStats = 1;
7800 MUTEX_EXIT(&rx_rpc_stats);
7804 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7814 rx_enablePeerRPCStats(void)
7816 MUTEX_ENTER(&rx_rpc_stats);
7817 rx_enable_stats = 1;
7818 rxi_monitor_peerStats = 1;
7819 MUTEX_EXIT(&rx_rpc_stats);
7823 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7833 rx_disableProcessRPCStats(void)
7835 rx_interface_stat_p rpc_stat, nrpc_stat;
7838 MUTEX_ENTER(&rx_rpc_stats);
7841 * Turn off process statistics and if peer stats is also off, turn
7845 rxi_monitor_processStats = 0;
7846 if (rxi_monitor_peerStats == 0) {
7847 rx_enable_stats = 0;
7850 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7851 unsigned int num_funcs = 0;
7854 queue_Remove(rpc_stat);
7855 num_funcs = rpc_stat->stats[0].func_total;
7857 sizeof(rx_interface_stat_t) +
7858 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7860 rxi_Free(rpc_stat, space);
7861 rxi_rpc_process_stat_cnt -= num_funcs;
7863 MUTEX_EXIT(&rx_rpc_stats);
7867 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7877 rx_disablePeerRPCStats(void)
7879 struct rx_peer **peer_ptr, **peer_end;
7882 MUTEX_ENTER(&rx_rpc_stats);
7885 * Turn off peer statistics and if process stats is also off, turn
7889 rxi_monitor_peerStats = 0;
7890 if (rxi_monitor_processStats == 0) {
7891 rx_enable_stats = 0;
7894 MUTEX_ENTER(&rx_peerHashTable_lock);
7895 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7896 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7898 struct rx_peer *peer, *next, *prev;
7899 for (prev = peer = *peer_ptr; peer; peer = next) {
7901 code = MUTEX_TRYENTER(&peer->peer_lock);
7903 rx_interface_stat_p rpc_stat, nrpc_stat;
7906 (&peer->rpcStats, rpc_stat, nrpc_stat,
7907 rx_interface_stat)) {
7908 unsigned int num_funcs = 0;
7911 queue_Remove(&rpc_stat->queue_header);
7912 queue_Remove(&rpc_stat->all_peers);
7913 num_funcs = rpc_stat->stats[0].func_total;
7915 sizeof(rx_interface_stat_t) +
7916 rpc_stat->stats[0].func_total *
7917 sizeof(rx_function_entry_v1_t);
7919 rxi_Free(rpc_stat, space);
7920 rxi_rpc_peer_stat_cnt -= num_funcs;
7922 MUTEX_EXIT(&peer->peer_lock);
7923 if (prev == *peer_ptr) {
7933 MUTEX_EXIT(&rx_peerHashTable_lock);
7934 MUTEX_EXIT(&rx_rpc_stats);
7938 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7943 * IN clearFlag - flag indicating which stats to clear
7951 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7953 rx_interface_stat_p rpc_stat, nrpc_stat;
7955 MUTEX_ENTER(&rx_rpc_stats);
7957 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7958 unsigned int num_funcs = 0, i;
7959 num_funcs = rpc_stat->stats[0].func_total;
7960 for (i = 0; i < num_funcs; i++) {
7961 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7962 hzero(rpc_stat->stats[i].invocations);
7964 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7965 hzero(rpc_stat->stats[i].bytes_sent);
7967 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7968 hzero(rpc_stat->stats[i].bytes_rcvd);
7970 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7971 rpc_stat->stats[i].queue_time_sum.sec = 0;
7972 rpc_stat->stats[i].queue_time_sum.usec = 0;
7974 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7975 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7976 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7978 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7979 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7980 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7982 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7983 rpc_stat->stats[i].queue_time_max.sec = 0;
7984 rpc_stat->stats[i].queue_time_max.usec = 0;
7986 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7987 rpc_stat->stats[i].execution_time_sum.sec = 0;
7988 rpc_stat->stats[i].execution_time_sum.usec = 0;
7990 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7991 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7992 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7994 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7995 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7996 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7998 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7999 rpc_stat->stats[i].execution_time_max.sec = 0;
8000 rpc_stat->stats[i].execution_time_max.usec = 0;
8005 MUTEX_EXIT(&rx_rpc_stats);
8009 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8014 * IN clearFlag - flag indicating which stats to clear
8022 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8024 rx_interface_stat_p rpc_stat, nrpc_stat;
8026 MUTEX_ENTER(&rx_rpc_stats);
8028 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8029 unsigned int num_funcs = 0, i;
8032 * We have to fix the offset of rpc_stat since we are
8033 * keeping this structure on two rx_queues. The rx_queue
8034 * package assumes that the rx_queue member is the first
8035 * member of the structure. That is, rx_queue assumes that
8036 * any one item is only on one queue at a time. We are
8037 * breaking that assumption and so we have to do a little
8038 * math to fix our pointers.
8041 fix_offset = (char *)rpc_stat;
8042 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8043 rpc_stat = (rx_interface_stat_p) fix_offset;
8045 num_funcs = rpc_stat->stats[0].func_total;
8046 for (i = 0; i < num_funcs; i++) {
8047 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8048 hzero(rpc_stat->stats[i].invocations);
8050 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8051 hzero(rpc_stat->stats[i].bytes_sent);
8053 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8054 hzero(rpc_stat->stats[i].bytes_rcvd);
8056 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8057 rpc_stat->stats[i].queue_time_sum.sec = 0;
8058 rpc_stat->stats[i].queue_time_sum.usec = 0;
8060 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8061 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8062 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8064 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8065 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8066 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8068 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8069 rpc_stat->stats[i].queue_time_max.sec = 0;
8070 rpc_stat->stats[i].queue_time_max.usec = 0;
8072 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8073 rpc_stat->stats[i].execution_time_sum.sec = 0;
8074 rpc_stat->stats[i].execution_time_sum.usec = 0;
8076 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8077 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8078 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8080 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8081 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8082 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8084 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8085 rpc_stat->stats[i].execution_time_max.sec = 0;
8086 rpc_stat->stats[i].execution_time_max.usec = 0;
8091 MUTEX_EXIT(&rx_rpc_stats);
8095 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8096 * is authorized to enable/disable/clear RX statistics.
8098 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8101 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8103 rxi_rxstat_userok = proc;
8107 rx_RxStatUserOk(struct rx_call *call)
8109 if (!rxi_rxstat_userok)
8111 return rxi_rxstat_userok(call);
8116 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8117 * function in the MSVC runtime DLL (msvcrt.dll).
8119 * Note: the system serializes calls to this function.
8122 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8123 DWORD reason, /* reason function is being called */
8124 LPVOID reserved) /* reserved for future use */
8127 case DLL_PROCESS_ATTACH:
8128 /* library is being attached to a process */
8132 case DLL_PROCESS_DETACH:
8140 int rx_DumpCalls(FILE *outputFile, char *cookie)
8142 #ifdef RXDEBUG_PACKET
8144 #ifdef KDUMP_RX_LOCK
8145 struct rx_call_rx_lock *c;
8151 sprintf(output, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8152 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8154 for (c = rx_allCallsp; c; c = c->allNextp) {
8155 u_short rqc, tqc, iovqc;
8156 struct rx_packet *p, *np;
8158 MUTEX_ENTER(&c->lock);
8159 queue_Count(&c->rq, p, np, rx_packet, rqc);
8160 queue_Count(&c->tq, p, np, rx_packet, tqc);
8161 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8163 sprintf(output, "%s - call=0x%p, id=%u, state=%u, mode=%u, conn=%p, epoch=%u, cid=%u, callNum=%u, connFlags=0x%x, flags=0x%x, "
8164 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8165 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8166 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8167 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8168 #ifdef RX_ENABLE_LOCKS
8171 #ifdef RX_REFCOUNT_CHECK
8172 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
8173 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
8176 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,
8177 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
8178 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
8179 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
8180 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
8181 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
8182 #ifdef RX_ENABLE_LOCKS
8183 , (afs_uint32)c->refCount
8185 #ifdef RX_REFCOUNT_CHECK
8186 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
8189 MUTEX_EXIT(&c->lock);
8191 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8193 sprintf(output, "%s - End dumping all Rx Calls\r\n", cookie);
8194 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8195 #endif /* RXDEBUG_PACKET */
8198 #endif /* AFS_NT40_ENV */