2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
14 #include "afs/param.h"
16 #include <afs/param.h>
21 #include "afs/sysincludes.h"
22 #include "afsincludes.h"
28 #include <net/net_globals.h>
29 #endif /* AFS_OSF_ENV */
30 #ifdef AFS_LINUX20_ENV
33 #include "netinet/in.h"
35 #include "inet/common.h"
37 #include "inet/ip_ire.h"
39 #include "afs/afs_args.h"
40 #include "afs/afs_osi.h"
41 #ifdef RX_KERNEL_TRACE
42 #include "rx_kcommon.h"
44 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
48 #undef RXDEBUG /* turn off debugging */
50 #if defined(AFS_SGI_ENV)
51 #include "sys/debug.h"
59 #endif /* AFS_OSF_ENV */
61 #include "afs/sysincludes.h"
62 #include "afsincludes.h"
65 #include "rx_kmutex.h"
66 #include "rx_kernel.h"
70 #include "rx_globals.h"
72 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
73 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
74 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
76 extern afs_int32 afs_termState;
78 #include "sys/lockl.h"
79 #include "sys/lock_def.h"
80 #endif /* AFS_AIX41_ENV */
81 # include "rxgen_consts.h"
83 # include <sys/types.h>
93 # include <afs/afsutil.h>
94 # include <WINNT\afsreg.h>
96 # include <sys/socket.h>
97 # include <sys/file.h>
99 # include <sys/stat.h>
100 # include <netinet/in.h>
101 # include <sys/time.h>
104 # include "rx_user.h"
105 # include "rx_clock.h"
106 # include "rx_queue.h"
107 # include "rx_globals.h"
108 # include "rx_trace.h"
109 # include <afs/rxgen_consts.h>
113 #ifdef AFS_PTHREAD_ENV
115 int (*registerProgram) (pid_t, char *) = 0;
116 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
119 int (*registerProgram) (PROCESS, char *) = 0;
120 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
124 /* Local static routines */
125 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
126 #ifdef RX_ENABLE_LOCKS
127 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
130 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
132 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
133 afs_int32 rxi_start_in_error;
135 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
138 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
139 * currently allocated within rx. This number is used to allocate the
140 * memory required to return the statistics when queried.
141 * Protected by the rx_rpc_stats mutex.
144 static unsigned int rxi_rpc_peer_stat_cnt;
147 * rxi_rpc_process_stat_cnt counts the total number of local process stat
148 * structures currently allocated within rx. The number is used to allocate
149 * the memory required to return the statistics when queried.
150 * Protected by the rx_rpc_stats mutex.
153 static unsigned int rxi_rpc_process_stat_cnt;
155 #if !defined(offsetof)
156 #include <stddef.h> /* for definition of offsetof() */
159 #ifdef AFS_PTHREAD_ENV
163 * Use procedural initialization of mutexes/condition variables
167 extern afs_kmutex_t rx_stats_mutex;
168 extern afs_kmutex_t rx_waiting_mutex;
169 extern afs_kmutex_t rx_quota_mutex;
170 extern afs_kmutex_t rx_pthread_mutex;
171 extern afs_kmutex_t rx_packets_mutex;
172 extern afs_kmutex_t des_init_mutex;
173 extern afs_kmutex_t des_random_mutex;
174 extern afs_kmutex_t rx_clock_mutex;
175 extern afs_kmutex_t rxi_connCacheMutex;
176 extern afs_kmutex_t rx_event_mutex;
177 extern afs_kmutex_t osi_malloc_mutex;
178 extern afs_kmutex_t event_handler_mutex;
179 extern afs_kmutex_t listener_mutex;
180 extern afs_kmutex_t rx_if_init_mutex;
181 extern afs_kmutex_t rx_if_mutex;
182 extern afs_kmutex_t rxkad_client_uid_mutex;
183 extern afs_kmutex_t rxkad_random_mutex;
185 extern afs_kcondvar_t rx_event_handler_cond;
186 extern afs_kcondvar_t rx_listener_cond;
188 static afs_kmutex_t epoch_mutex;
189 static afs_kmutex_t rx_init_mutex;
190 static afs_kmutex_t rx_debug_mutex;
191 static afs_kmutex_t rx_rpc_stats;
194 rxi_InitPthread(void)
196 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
197 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
198 MUTEX_INIT(&rx_waiting_mutex, "waiting", MUTEX_DEFAULT, 0);
199 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
200 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
201 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
202 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
203 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
204 MUTEX_INIT(&rx_event_mutex, "event", MUTEX_DEFAULT, 0);
205 MUTEX_INIT(&des_init_mutex, "des", MUTEX_DEFAULT, 0);
206 MUTEX_INIT(&des_random_mutex, "random", MUTEX_DEFAULT, 0);
207 MUTEX_INIT(&osi_malloc_mutex, "malloc", MUTEX_DEFAULT, 0);
208 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
209 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
210 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
211 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
212 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
213 MUTEX_INIT(&rxkad_client_uid_mutex, "uid", MUTEX_DEFAULT, 0);
214 MUTEX_INIT(&rxkad_random_mutex, "rxkad random", MUTEX_DEFAULT, 0);
215 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
217 assert(pthread_cond_init
218 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
219 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
221 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
222 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
224 rxkad_global_stats_init();
226 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
227 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
228 #ifdef RX_ENABLE_LOCKS
231 #endif /* RX_LOCKS_DB */
232 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
233 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
235 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
237 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
239 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
241 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
242 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
243 #endif /* RX_ENABLE_LOCKS */
246 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
247 #define INIT_PTHREAD_LOCKS \
248 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
250 * The rx_stats_mutex mutex protects the following global variables:
251 * rxi_lowConnRefCount
252 * rxi_lowPeerRefCount
261 * The rx_quota_mutex mutex protects the following global variables:
269 * The rx_freePktQ_lock protects the following global variables:
274 * The rx_packets_mutex mutex protects the following global variables:
282 * The rx_pthread_mutex mutex protects the following global variables:
286 #define INIT_PTHREAD_LOCKS
290 /* Variables for handling the minProcs implementation. availProcs gives the
291 * number of threads available in the pool at this moment (not counting dudes
292 * executing right now). totalMin gives the total number of procs required
293 * for handling all minProcs requests. minDeficit is a dynamic variable
294 * tracking the # of procs required to satisfy all of the remaining minProcs
296 * For fine grain locking to work, the quota check and the reservation of
297 * a server thread has to come while rxi_availProcs and rxi_minDeficit
298 * are locked. To this end, the code has been modified under #ifdef
299 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
300 * same time. A new function, ReturnToServerPool() returns the allocation.
302 * A call can be on several queue's (but only one at a time). When
303 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
304 * that no one else is touching the queue. To this end, we store the address
305 * of the queue lock in the call structure (under the call lock) when we
306 * put the call on a queue, and we clear the call_queue_lock when the
307 * call is removed from a queue (once the call lock has been obtained).
308 * This allows rxi_ResetCall to safely synchronize with others wishing
309 * to manipulate the queue.
312 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
313 static afs_kmutex_t rx_rpc_stats;
314 void rxi_StartUnlocked(struct rxevent *event, void *call,
315 void *arg1, int istack);
318 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
319 ** pretty good that the next packet coming in is from the same connection
320 ** as the last packet, since we're send multiple packets in a transmit window.
322 struct rx_connection *rxLastConn = 0;
324 #ifdef RX_ENABLE_LOCKS
325 /* The locking hierarchy for rx fine grain locking is composed of these
328 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
329 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
330 * call->lock - locks call data fields.
331 * These are independent of each other:
332 * rx_freeCallQueue_lock
337 * serverQueueEntry->lock
338 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
340 * peer->lock - locks peer data fields.
341 * conn_data_lock - that more than one thread is not updating a conn data
342 * field at the same time.
350 * Do we need a lock to protect the peer field in the conn structure?
351 * conn->peer was previously a constant for all intents and so has no
352 * lock protecting this field. The multihomed client delta introduced
353 * a RX code change : change the peer field in the connection structure
354 * to that remote interface from which the last packet for this
355 * connection was sent out. This may become an issue if further changes
358 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
359 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
361 /* rxdb_fileID is used to identify the lock location, along with line#. */
362 static int rxdb_fileID = RXDB_FILE_RX;
363 #endif /* RX_LOCKS_DB */
364 #else /* RX_ENABLE_LOCKS */
365 #define SET_CALL_QUEUE_LOCK(C, L)
366 #define CLEAR_CALL_QUEUE_LOCK(C)
367 #endif /* RX_ENABLE_LOCKS */
368 struct rx_serverQueueEntry *rx_waitForPacket = 0;
369 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
371 /* ------------Exported Interfaces------------- */
373 /* This function allows rxkad to set the epoch to a suitably random number
374 * which rx_NewConnection will use in the future. The principle purpose is to
375 * get rxnull connections to use the same epoch as the rxkad connections do, at
376 * least once the first rxkad connection is established. This is important now
377 * that the host/port addresses aren't used in FindConnection: the uniqueness
378 * of epoch/cid matters and the start time won't do. */
380 #ifdef AFS_PTHREAD_ENV
382 * This mutex protects the following global variables:
386 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
387 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
391 #endif /* AFS_PTHREAD_ENV */
394 rx_SetEpoch(afs_uint32 epoch)
401 /* Initialize rx. A port number may be mentioned, in which case this
402 * becomes the default port number for any service installed later.
403 * If 0 is provided for the port number, a random port will be chosen
404 * by the kernel. Whether this will ever overlap anything in
405 * /etc/services is anybody's guess... Returns 0 on success, -1 on
410 int rxinit_status = 1;
411 #ifdef AFS_PTHREAD_ENV
413 * This mutex protects the following global variables:
417 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
418 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
421 #define UNLOCK_RX_INIT
425 rx_InitHost(u_int host, u_int port)
432 char *htable, *ptable;
439 if (rxinit_status == 0) {
440 tmp_status = rxinit_status;
442 return tmp_status; /* Already started; return previous error code. */
448 if (afs_winsockInit() < 0)
454 * Initialize anything necessary to provide a non-premptive threading
457 rxi_InitializeThreadSupport();
460 /* Allocate and initialize a socket for client and perhaps server
463 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
464 if (rx_socket == OSI_NULLSOCKET) {
468 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
471 #endif /* RX_LOCKS_DB */
472 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
473 MUTEX_INIT(&rx_waiting_mutex, "rx_waiting_mutex", MUTEX_DEFAULT, 0);
474 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
475 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
476 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
477 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
478 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
479 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
480 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
482 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
484 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
486 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
488 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
489 #if defined(AFS_HPUX110_ENV)
491 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
492 #endif /* AFS_HPUX110_ENV */
493 #endif /* RX_ENABLE_LOCKS && KERNEL */
496 rx_connDeadTime = 12;
497 rx_tranquil = 0; /* reset flag */
498 memset(&rx_stats, 0, sizeof(struct rx_statistics));
500 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
501 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
502 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
503 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
504 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
505 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
507 /* Malloc up a bunch of packets & buffers */
509 queue_Init(&rx_freePacketQueue);
510 rxi_NeedMorePackets = FALSE;
511 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
513 /* enforce a minimum number of allocated packets */
514 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
515 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
517 /* allocate the initial free packet pool */
518 #ifdef RX_ENABLE_TSFPQ
519 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
520 #else /* RX_ENABLE_TSFPQ */
521 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
522 #endif /* RX_ENABLE_TSFPQ */
529 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
530 tv.tv_sec = clock_now.sec;
531 tv.tv_usec = clock_now.usec;
532 srand((unsigned int)tv.tv_usec);
539 #if defined(KERNEL) && !defined(UKERNEL)
540 /* Really, this should never happen in a real kernel */
543 struct sockaddr_in addr;
545 int addrlen = sizeof(addr);
547 socklen_t addrlen = sizeof(addr);
549 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
553 rx_port = addr.sin_port;
556 rx_stats.minRtt.sec = 9999999;
558 rx_SetEpoch(tv.tv_sec | 0x80000000);
560 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
561 * will provide a randomer value. */
563 MUTEX_ENTER(&rx_quota_mutex);
564 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
565 MUTEX_EXIT(&rx_quota_mutex);
566 /* *Slightly* random start time for the cid. This is just to help
567 * out with the hashing function at the peer */
568 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
569 rx_connHashTable = (struct rx_connection **)htable;
570 rx_peerHashTable = (struct rx_peer **)ptable;
572 rx_lastAckDelay.sec = 0;
573 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
574 rx_hardAckDelay.sec = 0;
575 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
576 rx_softAckDelay.sec = 0;
577 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
579 rxevent_Init(20, rxi_ReScheduleEvents);
581 /* Initialize various global queues */
582 queue_Init(&rx_idleServerQueue);
583 queue_Init(&rx_incomingCallQueue);
584 queue_Init(&rx_freeCallQueue);
586 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
587 /* Initialize our list of usable IP addresses. */
591 /* Start listener process (exact function is dependent on the
592 * implementation environment--kernel or user space) */
596 tmp_status = rxinit_status = 0;
604 return rx_InitHost(htonl(INADDR_ANY), port);
607 /* called with unincremented nRequestsRunning to see if it is OK to start
608 * a new thread in this service. Could be "no" for two reasons: over the
609 * max quota, or would prevent others from reaching their min quota.
611 #ifdef RX_ENABLE_LOCKS
612 /* This verion of QuotaOK reserves quota if it's ok while the
613 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
616 QuotaOK(struct rx_service *aservice)
618 /* check if over max quota */
619 if (aservice->nRequestsRunning >= aservice->maxProcs) {
623 /* under min quota, we're OK */
624 /* otherwise, can use only if there are enough to allow everyone
625 * to go to their min quota after this guy starts.
628 MUTEX_ENTER(&rx_quota_mutex);
629 if ((aservice->nRequestsRunning < aservice->minProcs)
630 || (rxi_availProcs > rxi_minDeficit)) {
631 aservice->nRequestsRunning++;
632 /* just started call in minProcs pool, need fewer to maintain
634 if (aservice->nRequestsRunning <= aservice->minProcs)
637 MUTEX_EXIT(&rx_quota_mutex);
640 MUTEX_EXIT(&rx_quota_mutex);
646 ReturnToServerPool(struct rx_service *aservice)
648 aservice->nRequestsRunning--;
649 MUTEX_ENTER(&rx_quota_mutex);
650 if (aservice->nRequestsRunning < aservice->minProcs)
653 MUTEX_EXIT(&rx_quota_mutex);
656 #else /* RX_ENABLE_LOCKS */
658 QuotaOK(struct rx_service *aservice)
661 /* under min quota, we're OK */
662 if (aservice->nRequestsRunning < aservice->minProcs)
665 /* check if over max quota */
666 if (aservice->nRequestsRunning >= aservice->maxProcs)
669 /* otherwise, can use only if there are enough to allow everyone
670 * to go to their min quota after this guy starts.
672 MUTEX_ENTER(&rx_quota_mutex);
673 if (rxi_availProcs > rxi_minDeficit)
675 MUTEX_EXIT(&rx_quota_mutex);
678 #endif /* RX_ENABLE_LOCKS */
681 /* Called by rx_StartServer to start up lwp's to service calls.
682 NExistingProcs gives the number of procs already existing, and which
683 therefore needn't be created. */
685 rxi_StartServerProcs(int nExistingProcs)
687 struct rx_service *service;
692 /* For each service, reserve N processes, where N is the "minimum"
693 * number of processes that MUST be able to execute a request in parallel,
694 * at any time, for that process. Also compute the maximum difference
695 * between any service's maximum number of processes that can run
696 * (i.e. the maximum number that ever will be run, and a guarantee
697 * that this number will run if other services aren't running), and its
698 * minimum number. The result is the extra number of processes that
699 * we need in order to provide the latter guarantee */
700 for (i = 0; i < RX_MAX_SERVICES; i++) {
702 service = rx_services[i];
703 if (service == (struct rx_service *)0)
705 nProcs += service->minProcs;
706 diff = service->maxProcs - service->minProcs;
710 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
711 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
712 for (i = 0; i < nProcs; i++) {
713 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
719 /* This routine is only required on Windows */
721 rx_StartClientThread(void)
723 #ifdef AFS_PTHREAD_ENV
725 pid = pthread_self();
726 #endif /* AFS_PTHREAD_ENV */
728 #endif /* AFS_NT40_ENV */
730 /* This routine must be called if any services are exported. If the
731 * donateMe flag is set, the calling process is donated to the server
734 rx_StartServer(int donateMe)
736 struct rx_service *service;
742 /* Start server processes, if necessary (exact function is dependent
743 * on the implementation environment--kernel or user space). DonateMe
744 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
745 * case, one less new proc will be created rx_StartServerProcs.
747 rxi_StartServerProcs(donateMe);
749 /* count up the # of threads in minProcs, and add set the min deficit to
750 * be that value, too.
752 for (i = 0; i < RX_MAX_SERVICES; i++) {
753 service = rx_services[i];
754 if (service == (struct rx_service *)0)
756 MUTEX_ENTER(&rx_quota_mutex);
757 rxi_totalMin += service->minProcs;
758 /* below works even if a thread is running, since minDeficit would
759 * still have been decremented and later re-incremented.
761 rxi_minDeficit += service->minProcs;
762 MUTEX_EXIT(&rx_quota_mutex);
765 /* Turn on reaping of idle server connections */
766 rxi_ReapConnections(NULL, NULL, NULL);
775 #ifdef AFS_PTHREAD_ENV
777 pid = afs_pointer_to_int(pthread_self());
778 #else /* AFS_PTHREAD_ENV */
780 LWP_CurrentProcess(&pid);
781 #endif /* AFS_PTHREAD_ENV */
783 sprintf(name, "srv_%d", ++nProcs);
785 (*registerProgram) (pid, name);
787 #endif /* AFS_NT40_ENV */
788 rx_ServerProc(NULL); /* Never returns */
790 #ifdef RX_ENABLE_TSFPQ
791 /* no use leaving packets around in this thread's local queue if
792 * it isn't getting donated to the server thread pool.
794 rxi_FlushLocalPacketsTSFPQ();
795 #endif /* RX_ENABLE_TSFPQ */
799 /* Create a new client connection to the specified service, using the
800 * specified security object to implement the security model for this
802 struct rx_connection *
803 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
804 struct rx_securityClass *securityObject,
805 int serviceSecurityIndex)
809 struct rx_connection *conn;
814 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
815 "serviceSecurityIndex %d)\n",
816 ntohl(shost), ntohs(sport), sservice, securityObject,
817 serviceSecurityIndex));
819 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
820 * the case of kmem_alloc? */
821 conn = rxi_AllocConnection();
822 #ifdef RX_ENABLE_LOCKS
823 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
824 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
825 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
828 MUTEX_ENTER(&rx_connHashTable_lock);
829 cid = (rx_nextCid += RX_MAXCALLS);
830 conn->type = RX_CLIENT_CONNECTION;
832 conn->epoch = rx_epoch;
833 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
834 conn->serviceId = sservice;
835 conn->securityObject = securityObject;
836 conn->securityData = (void *) 0;
837 conn->securityIndex = serviceSecurityIndex;
838 rx_SetConnDeadTime(conn, rx_connDeadTime);
839 rx_SetConnSecondsUntilNatPing(conn, 0);
840 conn->ackRate = RX_FAST_ACK_RATE;
842 conn->specific = NULL;
843 conn->challengeEvent = NULL;
844 conn->delayedAbortEvent = NULL;
845 conn->abortCount = 0;
847 for (i = 0; i < RX_MAXCALLS; i++) {
848 conn->twind[i] = rx_initSendWindow;
849 conn->rwind[i] = rx_initReceiveWindow;
852 RXS_NewConnection(securityObject, conn);
854 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
856 conn->refCount++; /* no lock required since only this thread knows... */
857 conn->next = rx_connHashTable[hashindex];
858 rx_connHashTable[hashindex] = conn;
860 rx_MutexIncrement(rx_stats.nClientConns, rx_stats_mutex);
861 MUTEX_EXIT(&rx_connHashTable_lock);
867 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
869 /* The idea is to set the dead time to a value that allows several
870 * keepalives to be dropped without timing out the connection. */
871 conn->secondsUntilDead = MAX(seconds, 6);
872 conn->secondsUntilPing = conn->secondsUntilDead / 6;
875 int rxi_lowPeerRefCount = 0;
876 int rxi_lowConnRefCount = 0;
879 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
880 * NOTE: must not be called with rx_connHashTable_lock held.
883 rxi_CleanupConnection(struct rx_connection *conn)
885 /* Notify the service exporter, if requested, that this connection
886 * is being destroyed */
887 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
888 (*conn->service->destroyConnProc) (conn);
890 /* Notify the security module that this connection is being destroyed */
891 RXS_DestroyConnection(conn->securityObject, conn);
893 /* If this is the last connection using the rx_peer struct, set its
894 * idle time to now. rxi_ReapConnections will reap it if it's still
895 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
897 MUTEX_ENTER(&rx_peerHashTable_lock);
898 if (conn->peer->refCount < 2) {
899 conn->peer->idleWhen = clock_Sec();
900 if (conn->peer->refCount < 1) {
901 conn->peer->refCount = 1;
902 if (rx_stats_active) {
903 MUTEX_ENTER(&rx_stats_mutex);
904 rxi_lowPeerRefCount++;
905 MUTEX_EXIT(&rx_stats_mutex);
909 conn->peer->refCount--;
910 MUTEX_EXIT(&rx_peerHashTable_lock);
914 if (conn->type == RX_SERVER_CONNECTION)
915 rx_MutexDecrement(rx_stats.nServerConns, rx_stats_mutex);
917 rx_MutexDecrement(rx_stats.nClientConns, rx_stats_mutex);
920 if (conn->specific) {
922 for (i = 0; i < conn->nSpecific; i++) {
923 if (conn->specific[i] && rxi_keyCreate_destructor[i])
924 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
925 conn->specific[i] = NULL;
927 free(conn->specific);
929 conn->specific = NULL;
933 MUTEX_DESTROY(&conn->conn_call_lock);
934 MUTEX_DESTROY(&conn->conn_data_lock);
935 CV_DESTROY(&conn->conn_call_cv);
937 rxi_FreeConnection(conn);
940 /* Destroy the specified connection */
942 rxi_DestroyConnection(struct rx_connection *conn)
944 MUTEX_ENTER(&rx_connHashTable_lock);
945 rxi_DestroyConnectionNoLock(conn);
946 /* conn should be at the head of the cleanup list */
947 if (conn == rx_connCleanup_list) {
948 rx_connCleanup_list = rx_connCleanup_list->next;
949 MUTEX_EXIT(&rx_connHashTable_lock);
950 rxi_CleanupConnection(conn);
952 #ifdef RX_ENABLE_LOCKS
954 MUTEX_EXIT(&rx_connHashTable_lock);
956 #endif /* RX_ENABLE_LOCKS */
960 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
962 struct rx_connection **conn_ptr;
964 struct rx_packet *packet;
971 MUTEX_ENTER(&conn->conn_data_lock);
972 if (conn->refCount > 0)
975 if (rx_stats_active) {
976 MUTEX_ENTER(&rx_stats_mutex);
977 rxi_lowConnRefCount++;
978 MUTEX_EXIT(&rx_stats_mutex);
982 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
983 /* Busy; wait till the last guy before proceeding */
984 MUTEX_EXIT(&conn->conn_data_lock);
989 /* If the client previously called rx_NewCall, but it is still
990 * waiting, treat this as a running call, and wait to destroy the
991 * connection later when the call completes. */
992 if ((conn->type == RX_CLIENT_CONNECTION)
993 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
994 conn->flags |= RX_CONN_DESTROY_ME;
995 MUTEX_EXIT(&conn->conn_data_lock);
999 MUTEX_EXIT(&conn->conn_data_lock);
1001 /* Check for extant references to this connection */
1002 for (i = 0; i < RX_MAXCALLS; i++) {
1003 struct rx_call *call = conn->call[i];
1006 if (conn->type == RX_CLIENT_CONNECTION) {
1007 MUTEX_ENTER(&call->lock);
1008 if (call->delayedAckEvent) {
1009 /* Push the final acknowledgment out now--there
1010 * won't be a subsequent call to acknowledge the
1011 * last reply packets */
1012 rxevent_Cancel(call->delayedAckEvent, call,
1013 RX_CALL_REFCOUNT_DELAY);
1014 if (call->state == RX_STATE_PRECALL
1015 || call->state == RX_STATE_ACTIVE) {
1016 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1018 rxi_AckAll(NULL, call, 0);
1021 MUTEX_EXIT(&call->lock);
1025 #ifdef RX_ENABLE_LOCKS
1027 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1028 MUTEX_EXIT(&conn->conn_data_lock);
1030 /* Someone is accessing a packet right now. */
1034 #endif /* RX_ENABLE_LOCKS */
1037 /* Don't destroy the connection if there are any call
1038 * structures still in use */
1039 MUTEX_ENTER(&conn->conn_data_lock);
1040 conn->flags |= RX_CONN_DESTROY_ME;
1041 MUTEX_EXIT(&conn->conn_data_lock);
1046 if (conn->natKeepAliveEvent) {
1047 rxi_NatKeepAliveOff(conn);
1050 if (conn->delayedAbortEvent) {
1051 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1052 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1054 MUTEX_ENTER(&conn->conn_data_lock);
1055 rxi_SendConnectionAbort(conn, packet, 0, 1);
1056 MUTEX_EXIT(&conn->conn_data_lock);
1057 rxi_FreePacket(packet);
1061 /* Remove from connection hash table before proceeding */
1063 &rx_connHashTable[CONN_HASH
1064 (peer->host, peer->port, conn->cid, conn->epoch,
1066 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1067 if (*conn_ptr == conn) {
1068 *conn_ptr = conn->next;
1072 /* if the conn that we are destroying was the last connection, then we
1073 * clear rxLastConn as well */
1074 if (rxLastConn == conn)
1077 /* Make sure the connection is completely reset before deleting it. */
1078 /* get rid of pending events that could zap us later */
1079 if (conn->challengeEvent)
1080 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1081 if (conn->checkReachEvent)
1082 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1083 if (conn->natKeepAliveEvent)
1084 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
1086 /* Add the connection to the list of destroyed connections that
1087 * need to be cleaned up. This is necessary to avoid deadlocks
1088 * in the routines we call to inform others that this connection is
1089 * being destroyed. */
1090 conn->next = rx_connCleanup_list;
1091 rx_connCleanup_list = conn;
1094 /* Externally available version */
1096 rx_DestroyConnection(struct rx_connection *conn)
1101 rxi_DestroyConnection(conn);
1106 rx_GetConnection(struct rx_connection *conn)
1111 MUTEX_ENTER(&conn->conn_data_lock);
1113 MUTEX_EXIT(&conn->conn_data_lock);
1117 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1118 /* Wait for the transmit queue to no longer be busy.
1119 * requires the call->lock to be held */
1120 static void rxi_WaitforTQBusy(struct rx_call *call) {
1121 while (call->flags & RX_CALL_TQ_BUSY) {
1122 call->flags |= RX_CALL_TQ_WAIT;
1124 #ifdef RX_ENABLE_LOCKS
1125 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1126 CV_WAIT(&call->cv_tq, &call->lock);
1127 #else /* RX_ENABLE_LOCKS */
1128 osi_rxSleep(&call->tq);
1129 #endif /* RX_ENABLE_LOCKS */
1131 if (call->tqWaiters == 0) {
1132 call->flags &= ~RX_CALL_TQ_WAIT;
1138 /* Start a new rx remote procedure call, on the specified connection.
1139 * If wait is set to 1, wait for a free call channel; otherwise return
1140 * 0. Maxtime gives the maximum number of seconds this call may take,
1141 * after rx_NewCall returns. After this time interval, a call to any
1142 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1143 * For fine grain locking, we hold the conn_call_lock in order to
1144 * to ensure that we don't get signalle after we found a call in an active
1145 * state and before we go to sleep.
1148 rx_NewCall(struct rx_connection *conn)
1151 struct rx_call *call;
1152 struct clock queueTime;
1156 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1159 clock_GetTime(&queueTime);
1161 * Check if there are others waiting for a new call.
1162 * If so, let them go first to avoid starving them.
1163 * This is a fairly simple scheme, and might not be
1164 * a complete solution for large numbers of waiters.
1166 * makeCallWaiters keeps track of the number of
1167 * threads waiting to make calls and the
1168 * RX_CONN_MAKECALL_WAITING flag bit is used to
1169 * indicate that there are indeed calls waiting.
1170 * The flag is set when the waiter is incremented.
1171 * It is only cleared when makeCallWaiters is 0.
1172 * This prevents us from accidently destroying the
1173 * connection while it is potentially about to be used.
1175 MUTEX_ENTER(&conn->conn_call_lock);
1176 MUTEX_ENTER(&conn->conn_data_lock);
1177 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1178 conn->flags |= RX_CONN_MAKECALL_WAITING;
1179 conn->makeCallWaiters++;
1180 MUTEX_EXIT(&conn->conn_data_lock);
1182 #ifdef RX_ENABLE_LOCKS
1183 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1187 MUTEX_ENTER(&conn->conn_data_lock);
1188 conn->makeCallWaiters--;
1189 if (conn->makeCallWaiters == 0)
1190 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1193 /* We are now the active thread in rx_NewCall */
1194 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1195 MUTEX_EXIT(&conn->conn_data_lock);
1200 for (i = 0; i < RX_MAXCALLS; i++) {
1201 call = conn->call[i];
1203 if (call->state == RX_STATE_DALLY) {
1204 MUTEX_ENTER(&call->lock);
1205 if (call->state == RX_STATE_DALLY) {
1207 * We are setting the state to RX_STATE_RESET to
1208 * ensure that no one else will attempt to use this
1209 * call once we drop the conn->conn_call_lock and
1210 * call->lock. We must drop the conn->conn_call_lock
1211 * before calling rxi_ResetCall because the process
1212 * of clearing the transmit queue can block for an
1213 * extended period of time. If we block while holding
1214 * the conn->conn_call_lock, then all rx_EndCall
1215 * processing will block as well. This has a detrimental
1216 * effect on overall system performance.
1218 call->state = RX_STATE_RESET;
1219 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1220 MUTEX_EXIT(&conn->conn_call_lock);
1221 rxi_ResetCall(call, 0);
1222 (*call->callNumber)++;
1223 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1227 * If we failed to be able to safely obtain the
1228 * conn->conn_call_lock we will have to drop the
1229 * call->lock to avoid a deadlock. When the call->lock
1230 * is released the state of the call can change. If it
1231 * is no longer RX_STATE_RESET then some other thread is
1234 MUTEX_EXIT(&call->lock);
1235 MUTEX_ENTER(&conn->conn_call_lock);
1236 MUTEX_ENTER(&call->lock);
1238 if (call->state == RX_STATE_RESET)
1242 * If we get here it means that after dropping
1243 * the conn->conn_call_lock and call->lock that
1244 * the call is no longer ours. If we can't find
1245 * a free call in the remaining slots we should
1246 * not go immediately to RX_CONN_MAKECALL_WAITING
1247 * because by dropping the conn->conn_call_lock
1248 * we have given up synchronization with rx_EndCall.
1249 * Instead, cycle through one more time to see if
1250 * we can find a call that can call our own.
1252 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1255 MUTEX_EXIT(&call->lock);
1258 /* rxi_NewCall returns with mutex locked */
1259 call = rxi_NewCall(conn, i);
1260 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1264 if (i < RX_MAXCALLS) {
1270 MUTEX_ENTER(&conn->conn_data_lock);
1271 conn->flags |= RX_CONN_MAKECALL_WAITING;
1272 conn->makeCallWaiters++;
1273 MUTEX_EXIT(&conn->conn_data_lock);
1275 #ifdef RX_ENABLE_LOCKS
1276 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1280 MUTEX_ENTER(&conn->conn_data_lock);
1281 conn->makeCallWaiters--;
1282 if (conn->makeCallWaiters == 0)
1283 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1284 MUTEX_EXIT(&conn->conn_data_lock);
1286 /* Client is initially in send mode */
1287 call->state = RX_STATE_ACTIVE;
1288 call->error = conn->error;
1290 call->mode = RX_MODE_ERROR;
1292 call->mode = RX_MODE_SENDING;
1294 /* remember start time for call in case we have hard dead time limit */
1295 call->queueTime = queueTime;
1296 clock_GetTime(&call->startTime);
1297 hzero(call->bytesSent);
1298 hzero(call->bytesRcvd);
1300 /* Turn on busy protocol. */
1301 rxi_KeepAliveOn(call);
1303 /* Attempt MTU discovery */
1304 rxi_GrowMTUOn(call);
1307 * We are no longer the active thread in rx_NewCall
1309 MUTEX_ENTER(&conn->conn_data_lock);
1310 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1311 MUTEX_EXIT(&conn->conn_data_lock);
1314 * Wake up anyone else who might be giving us a chance to
1315 * run (see code above that avoids resource starvation).
1317 #ifdef RX_ENABLE_LOCKS
1318 CV_BROADCAST(&conn->conn_call_cv);
1322 MUTEX_EXIT(&conn->conn_call_lock);
1324 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1325 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1326 osi_Panic("rx_NewCall call about to be used without an empty tq");
1328 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1330 MUTEX_EXIT(&call->lock);
1333 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1338 rxi_HasActiveCalls(struct rx_connection *aconn)
1341 struct rx_call *tcall;
1345 for (i = 0; i < RX_MAXCALLS; i++) {
1346 if ((tcall = aconn->call[i])) {
1347 if ((tcall->state == RX_STATE_ACTIVE)
1348 || (tcall->state == RX_STATE_PRECALL)) {
1359 rxi_GetCallNumberVector(struct rx_connection *aconn,
1360 afs_int32 * aint32s)
1363 struct rx_call *tcall;
1367 for (i = 0; i < RX_MAXCALLS; i++) {
1368 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1369 aint32s[i] = aconn->callNumber[i] + 1;
1371 aint32s[i] = aconn->callNumber[i];
1378 rxi_SetCallNumberVector(struct rx_connection *aconn,
1379 afs_int32 * aint32s)
1382 struct rx_call *tcall;
1386 for (i = 0; i < RX_MAXCALLS; i++) {
1387 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1388 aconn->callNumber[i] = aint32s[i] - 1;
1390 aconn->callNumber[i] = aint32s[i];
1396 /* Advertise a new service. A service is named locally by a UDP port
1397 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1400 char *serviceName; Name for identification purposes (e.g. the
1401 service name might be used for probing for
1404 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1405 char *serviceName, struct rx_securityClass **securityObjects,
1406 int nSecurityObjects,
1407 afs_int32(*serviceProc) (struct rx_call * acall))
1409 osi_socket socket = OSI_NULLSOCKET;
1410 struct rx_service *tservice;
1416 if (serviceId == 0) {
1418 "rx_NewService: service id for service %s is not non-zero.\n",
1425 "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",
1433 tservice = rxi_AllocService();
1436 #ifdef RX_ENABLE_LOCKS
1437 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1440 for (i = 0; i < RX_MAX_SERVICES; i++) {
1441 struct rx_service *service = rx_services[i];
1443 if (port == service->servicePort && host == service->serviceHost) {
1444 if (service->serviceId == serviceId) {
1445 /* The identical service has already been
1446 * installed; if the caller was intending to
1447 * change the security classes used by this
1448 * service, he/she loses. */
1450 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1451 serviceName, serviceId, service->serviceName);
1453 rxi_FreeService(tservice);
1456 /* Different service, same port: re-use the socket
1457 * which is bound to the same port */
1458 socket = service->socket;
1461 if (socket == OSI_NULLSOCKET) {
1462 /* If we don't already have a socket (from another
1463 * service on same port) get a new one */
1464 socket = rxi_GetHostUDPSocket(host, port);
1465 if (socket == OSI_NULLSOCKET) {
1467 rxi_FreeService(tservice);
1472 service->socket = socket;
1473 service->serviceHost = host;
1474 service->servicePort = port;
1475 service->serviceId = serviceId;
1476 service->serviceName = serviceName;
1477 service->nSecurityObjects = nSecurityObjects;
1478 service->securityObjects = securityObjects;
1479 service->minProcs = 0;
1480 service->maxProcs = 1;
1481 service->idleDeadTime = 60;
1482 service->idleDeadErr = 0;
1483 service->connDeadTime = rx_connDeadTime;
1484 service->executeRequestProc = serviceProc;
1485 service->checkReach = 0;
1486 service->nSpecific = 0;
1487 service->specific = NULL;
1488 rx_services[i] = service; /* not visible until now */
1494 rxi_FreeService(tservice);
1495 (osi_Msg "rx_NewService: cannot support > %d services\n",
1500 /* Set configuration options for all of a service's security objects */
1503 rx_SetSecurityConfiguration(struct rx_service *service,
1504 rx_securityConfigVariables type,
1508 for (i = 0; i<service->nSecurityObjects; i++) {
1509 if (service->securityObjects[i]) {
1510 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1518 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1519 struct rx_securityClass **securityObjects, int nSecurityObjects,
1520 afs_int32(*serviceProc) (struct rx_call * acall))
1522 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1525 /* Generic request processing loop. This routine should be called
1526 * by the implementation dependent rx_ServerProc. If socketp is
1527 * non-null, it will be set to the file descriptor that this thread
1528 * is now listening on. If socketp is null, this routine will never
1531 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1533 struct rx_call *call;
1535 struct rx_service *tservice = NULL;
1542 call = rx_GetCall(threadID, tservice, socketp);
1543 if (socketp && *socketp != OSI_NULLSOCKET) {
1544 /* We are now a listener thread */
1549 /* if server is restarting( typically smooth shutdown) then do not
1550 * allow any new calls.
1553 if (rx_tranquil && (call != NULL)) {
1557 MUTEX_ENTER(&call->lock);
1559 rxi_CallError(call, RX_RESTARTING);
1560 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1562 MUTEX_EXIT(&call->lock);
1566 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1567 #ifdef RX_ENABLE_LOCKS
1569 #endif /* RX_ENABLE_LOCKS */
1570 afs_termState = AFSOP_STOP_AFS;
1571 afs_osi_Wakeup(&afs_termState);
1572 #ifdef RX_ENABLE_LOCKS
1574 #endif /* RX_ENABLE_LOCKS */
1579 tservice = call->conn->service;
1581 if (tservice->beforeProc)
1582 (*tservice->beforeProc) (call);
1584 code = call->conn->service->executeRequestProc(call);
1586 if (tservice->afterProc)
1587 (*tservice->afterProc) (call, code);
1589 rx_EndCall(call, code);
1590 if (rx_stats_active) {
1591 MUTEX_ENTER(&rx_stats_mutex);
1593 MUTEX_EXIT(&rx_stats_mutex);
1600 rx_WakeupServerProcs(void)
1602 struct rx_serverQueueEntry *np, *tqp;
1606 MUTEX_ENTER(&rx_serverPool_lock);
1608 #ifdef RX_ENABLE_LOCKS
1609 if (rx_waitForPacket)
1610 CV_BROADCAST(&rx_waitForPacket->cv);
1611 #else /* RX_ENABLE_LOCKS */
1612 if (rx_waitForPacket)
1613 osi_rxWakeup(rx_waitForPacket);
1614 #endif /* RX_ENABLE_LOCKS */
1615 MUTEX_ENTER(&freeSQEList_lock);
1616 for (np = rx_FreeSQEList; np; np = tqp) {
1617 tqp = *(struct rx_serverQueueEntry **)np;
1618 #ifdef RX_ENABLE_LOCKS
1619 CV_BROADCAST(&np->cv);
1620 #else /* RX_ENABLE_LOCKS */
1622 #endif /* RX_ENABLE_LOCKS */
1624 MUTEX_EXIT(&freeSQEList_lock);
1625 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1626 #ifdef RX_ENABLE_LOCKS
1627 CV_BROADCAST(&np->cv);
1628 #else /* RX_ENABLE_LOCKS */
1630 #endif /* RX_ENABLE_LOCKS */
1632 MUTEX_EXIT(&rx_serverPool_lock);
1637 * One thing that seems to happen is that all the server threads get
1638 * tied up on some empty or slow call, and then a whole bunch of calls
1639 * arrive at once, using up the packet pool, so now there are more
1640 * empty calls. The most critical resources here are server threads
1641 * and the free packet pool. The "doreclaim" code seems to help in
1642 * general. I think that eventually we arrive in this state: there
1643 * are lots of pending calls which do have all their packets present,
1644 * so they won't be reclaimed, are multi-packet calls, so they won't
1645 * be scheduled until later, and thus are tying up most of the free
1646 * packet pool for a very long time.
1648 * 1. schedule multi-packet calls if all the packets are present.
1649 * Probably CPU-bound operation, useful to return packets to pool.
1650 * Do what if there is a full window, but the last packet isn't here?
1651 * 3. preserve one thread which *only* runs "best" calls, otherwise
1652 * it sleeps and waits for that type of call.
1653 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1654 * the current dataquota business is badly broken. The quota isn't adjusted
1655 * to reflect how many packets are presently queued for a running call.
1656 * So, when we schedule a queued call with a full window of packets queued
1657 * up for it, that *should* free up a window full of packets for other 2d-class
1658 * calls to be able to use from the packet pool. But it doesn't.
1660 * NB. Most of the time, this code doesn't run -- since idle server threads
1661 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1662 * as a new call arrives.
1664 /* Sleep until a call arrives. Returns a pointer to the call, ready
1665 * for an rx_Read. */
1666 #ifdef RX_ENABLE_LOCKS
1668 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1670 struct rx_serverQueueEntry *sq;
1671 struct rx_call *call = (struct rx_call *)0;
1672 struct rx_service *service = NULL;
1675 MUTEX_ENTER(&freeSQEList_lock);
1677 if ((sq = rx_FreeSQEList)) {
1678 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1679 MUTEX_EXIT(&freeSQEList_lock);
1680 } else { /* otherwise allocate a new one and return that */
1681 MUTEX_EXIT(&freeSQEList_lock);
1682 sq = (struct rx_serverQueueEntry *)
1683 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1684 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1685 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1688 MUTEX_ENTER(&rx_serverPool_lock);
1689 if (cur_service != NULL) {
1690 ReturnToServerPool(cur_service);
1693 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1694 struct rx_call *tcall, *ncall, *choice2 = NULL;
1696 /* Scan for eligible incoming calls. A call is not eligible
1697 * if the maximum number of calls for its service type are
1698 * already executing */
1699 /* One thread will process calls FCFS (to prevent starvation),
1700 * while the other threads may run ahead looking for calls which
1701 * have all their input data available immediately. This helps
1702 * keep threads from blocking, waiting for data from the client. */
1703 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1704 service = tcall->conn->service;
1705 if (!QuotaOK(service)) {
1708 MUTEX_ENTER(&rx_pthread_mutex);
1709 if (tno == rxi_fcfs_thread_num
1710 || !tcall->queue_item_header.next) {
1711 MUTEX_EXIT(&rx_pthread_mutex);
1712 /* If we're the fcfs thread , then we'll just use
1713 * this call. If we haven't been able to find an optimal
1714 * choice, and we're at the end of the list, then use a
1715 * 2d choice if one has been identified. Otherwise... */
1716 call = (choice2 ? choice2 : tcall);
1717 service = call->conn->service;
1719 MUTEX_EXIT(&rx_pthread_mutex);
1720 if (!queue_IsEmpty(&tcall->rq)) {
1721 struct rx_packet *rp;
1722 rp = queue_First(&tcall->rq, rx_packet);
1723 if (rp->header.seq == 1) {
1725 || (rp->header.flags & RX_LAST_PACKET)) {
1727 } else if (rxi_2dchoice && !choice2
1728 && !(tcall->flags & RX_CALL_CLEARED)
1729 && (tcall->rprev > rxi_HardAckRate)) {
1739 ReturnToServerPool(service);
1746 MUTEX_EXIT(&rx_serverPool_lock);
1747 MUTEX_ENTER(&call->lock);
1749 if (call->flags & RX_CALL_WAIT_PROC) {
1750 call->flags &= ~RX_CALL_WAIT_PROC;
1751 MUTEX_ENTER(&rx_waiting_mutex);
1753 MUTEX_EXIT(&rx_waiting_mutex);
1756 if (call->state != RX_STATE_PRECALL || call->error) {
1757 MUTEX_EXIT(&call->lock);
1758 MUTEX_ENTER(&rx_serverPool_lock);
1759 ReturnToServerPool(service);
1764 if (queue_IsEmpty(&call->rq)
1765 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1766 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1768 CLEAR_CALL_QUEUE_LOCK(call);
1771 /* If there are no eligible incoming calls, add this process
1772 * to the idle server queue, to wait for one */
1776 *socketp = OSI_NULLSOCKET;
1778 sq->socketp = socketp;
1779 queue_Append(&rx_idleServerQueue, sq);
1780 #ifndef AFS_AIX41_ENV
1781 rx_waitForPacket = sq;
1783 rx_waitingForPacket = sq;
1784 #endif /* AFS_AIX41_ENV */
1786 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1788 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1789 MUTEX_EXIT(&rx_serverPool_lock);
1790 return (struct rx_call *)0;
1793 } while (!(call = sq->newcall)
1794 && !(socketp && *socketp != OSI_NULLSOCKET));
1795 MUTEX_EXIT(&rx_serverPool_lock);
1797 MUTEX_ENTER(&call->lock);
1803 MUTEX_ENTER(&freeSQEList_lock);
1804 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1805 rx_FreeSQEList = sq;
1806 MUTEX_EXIT(&freeSQEList_lock);
1809 clock_GetTime(&call->startTime);
1810 call->state = RX_STATE_ACTIVE;
1811 call->mode = RX_MODE_RECEIVING;
1812 #ifdef RX_KERNEL_TRACE
1813 if (ICL_SETACTIVE(afs_iclSetp)) {
1814 int glockOwner = ISAFS_GLOCK();
1817 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1818 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1825 rxi_calltrace(RX_CALL_START, call);
1826 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
1827 call->conn->service->servicePort, call->conn->service->serviceId,
1830 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1831 MUTEX_EXIT(&call->lock);
1833 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1838 #else /* RX_ENABLE_LOCKS */
1840 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1842 struct rx_serverQueueEntry *sq;
1843 struct rx_call *call = (struct rx_call *)0, *choice2;
1844 struct rx_service *service = NULL;
1848 MUTEX_ENTER(&freeSQEList_lock);
1850 if ((sq = rx_FreeSQEList)) {
1851 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1852 MUTEX_EXIT(&freeSQEList_lock);
1853 } else { /* otherwise allocate a new one and return that */
1854 MUTEX_EXIT(&freeSQEList_lock);
1855 sq = (struct rx_serverQueueEntry *)
1856 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1857 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1858 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1860 MUTEX_ENTER(&sq->lock);
1862 if (cur_service != NULL) {
1863 cur_service->nRequestsRunning--;
1864 MUTEX_ENTER(&rx_quota_mutex);
1865 if (cur_service->nRequestsRunning < cur_service->minProcs)
1868 MUTEX_EXIT(&rx_quota_mutex);
1870 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1871 struct rx_call *tcall, *ncall;
1872 /* Scan for eligible incoming calls. A call is not eligible
1873 * if the maximum number of calls for its service type are
1874 * already executing */
1875 /* One thread will process calls FCFS (to prevent starvation),
1876 * while the other threads may run ahead looking for calls which
1877 * have all their input data available immediately. This helps
1878 * keep threads from blocking, waiting for data from the client. */
1879 choice2 = (struct rx_call *)0;
1880 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1881 service = tcall->conn->service;
1882 if (QuotaOK(service)) {
1883 MUTEX_ENTER(&rx_pthread_mutex);
1884 if (tno == rxi_fcfs_thread_num
1885 || !tcall->queue_item_header.next) {
1886 MUTEX_EXIT(&rx_pthread_mutex);
1887 /* If we're the fcfs thread, then we'll just use
1888 * this call. If we haven't been able to find an optimal
1889 * choice, and we're at the end of the list, then use a
1890 * 2d choice if one has been identified. Otherwise... */
1891 call = (choice2 ? choice2 : tcall);
1892 service = call->conn->service;
1894 MUTEX_EXIT(&rx_pthread_mutex);
1895 if (!queue_IsEmpty(&tcall->rq)) {
1896 struct rx_packet *rp;
1897 rp = queue_First(&tcall->rq, rx_packet);
1898 if (rp->header.seq == 1
1900 || (rp->header.flags & RX_LAST_PACKET))) {
1902 } else if (rxi_2dchoice && !choice2
1903 && !(tcall->flags & RX_CALL_CLEARED)
1904 && (tcall->rprev > rxi_HardAckRate)) {
1918 /* we can't schedule a call if there's no data!!! */
1919 /* send an ack if there's no data, if we're missing the
1920 * first packet, or we're missing something between first
1921 * and last -- there's a "hole" in the incoming data. */
1922 if (queue_IsEmpty(&call->rq)
1923 || queue_First(&call->rq, rx_packet)->header.seq != 1
1924 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1925 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1927 call->flags &= (~RX_CALL_WAIT_PROC);
1928 service->nRequestsRunning++;
1929 /* just started call in minProcs pool, need fewer to maintain
1931 MUTEX_ENTER(&rx_quota_mutex);
1932 if (service->nRequestsRunning <= service->minProcs)
1935 MUTEX_EXIT(&rx_quota_mutex);
1937 /* MUTEX_EXIT(&call->lock); */
1939 /* If there are no eligible incoming calls, add this process
1940 * to the idle server queue, to wait for one */
1943 *socketp = OSI_NULLSOCKET;
1945 sq->socketp = socketp;
1946 queue_Append(&rx_idleServerQueue, sq);
1950 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1952 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1953 return (struct rx_call *)0;
1956 } while (!(call = sq->newcall)
1957 && !(socketp && *socketp != OSI_NULLSOCKET));
1959 MUTEX_EXIT(&sq->lock);
1961 MUTEX_ENTER(&freeSQEList_lock);
1962 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1963 rx_FreeSQEList = sq;
1964 MUTEX_EXIT(&freeSQEList_lock);
1967 clock_GetTime(&call->startTime);
1968 call->state = RX_STATE_ACTIVE;
1969 call->mode = RX_MODE_RECEIVING;
1970 #ifdef RX_KERNEL_TRACE
1971 if (ICL_SETACTIVE(afs_iclSetp)) {
1972 int glockOwner = ISAFS_GLOCK();
1975 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1976 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1983 rxi_calltrace(RX_CALL_START, call);
1984 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
1985 call->conn->service->servicePort, call->conn->service->serviceId,
1988 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1995 #endif /* RX_ENABLE_LOCKS */
1999 /* Establish a procedure to be called when a packet arrives for a
2000 * call. This routine will be called at most once after each call,
2001 * and will also be called if there is an error condition on the or
2002 * the call is complete. Used by multi rx to build a selection
2003 * function which determines which of several calls is likely to be a
2004 * good one to read from.
2005 * NOTE: the way this is currently implemented it is probably only a
2006 * good idea to (1) use it immediately after a newcall (clients only)
2007 * and (2) only use it once. Other uses currently void your warranty
2010 rx_SetArrivalProc(struct rx_call *call,
2011 void (*proc) (struct rx_call * call,
2014 void * handle, int arg)
2016 call->arrivalProc = proc;
2017 call->arrivalProcHandle = handle;
2018 call->arrivalProcArg = arg;
2021 /* Call is finished (possibly prematurely). Return rc to the peer, if
2022 * appropriate, and return the final error code from the conversation
2026 rx_EndCall(struct rx_call *call, afs_int32 rc)
2028 struct rx_connection *conn = call->conn;
2032 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2033 call, rc, call->error, call->abortCode));
2036 MUTEX_ENTER(&call->lock);
2038 if (rc == 0 && call->error == 0) {
2039 call->abortCode = 0;
2040 call->abortCount = 0;
2043 call->arrivalProc = (void (*)())0;
2044 if (rc && call->error == 0) {
2045 rxi_CallError(call, rc);
2046 /* Send an abort message to the peer if this error code has
2047 * only just been set. If it was set previously, assume the
2048 * peer has already been sent the error code or will request it
2050 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2052 if (conn->type == RX_SERVER_CONNECTION) {
2053 /* Make sure reply or at least dummy reply is sent */
2054 if (call->mode == RX_MODE_RECEIVING) {
2055 rxi_WriteProc(call, 0, 0);
2057 if (call->mode == RX_MODE_SENDING) {
2058 rxi_FlushWrite(call);
2060 rxi_calltrace(RX_CALL_END, call);
2061 /* Call goes to hold state until reply packets are acknowledged */
2062 if (call->tfirst + call->nSoftAcked < call->tnext) {
2063 call->state = RX_STATE_HOLD;
2065 call->state = RX_STATE_DALLY;
2066 rxi_ClearTransmitQueue(call, 0);
2067 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
2068 rxevent_Cancel(call->keepAliveEvent, call,
2069 RX_CALL_REFCOUNT_ALIVE);
2071 } else { /* Client connection */
2073 /* Make sure server receives input packets, in the case where
2074 * no reply arguments are expected */
2075 if ((call->mode == RX_MODE_SENDING)
2076 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2077 (void)rxi_ReadProc(call, &dummy, 1);
2080 /* If we had an outstanding delayed ack, be nice to the server
2081 * and force-send it now.
2083 if (call->delayedAckEvent) {
2084 rxevent_Cancel(call->delayedAckEvent, call,
2085 RX_CALL_REFCOUNT_DELAY);
2086 call->delayedAckEvent = NULL;
2087 rxi_SendDelayedAck(NULL, call, NULL);
2090 /* We need to release the call lock since it's lower than the
2091 * conn_call_lock and we don't want to hold the conn_call_lock
2092 * over the rx_ReadProc call. The conn_call_lock needs to be held
2093 * here for the case where rx_NewCall is perusing the calls on
2094 * the connection structure. We don't want to signal until
2095 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2096 * have checked this call, found it active and by the time it
2097 * goes to sleep, will have missed the signal.
2099 MUTEX_EXIT(&call->lock);
2100 MUTEX_ENTER(&conn->conn_call_lock);
2101 MUTEX_ENTER(&call->lock);
2102 MUTEX_ENTER(&conn->conn_data_lock);
2103 conn->flags |= RX_CONN_BUSY;
2104 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2105 MUTEX_EXIT(&conn->conn_data_lock);
2106 #ifdef RX_ENABLE_LOCKS
2107 CV_BROADCAST(&conn->conn_call_cv);
2112 #ifdef RX_ENABLE_LOCKS
2114 MUTEX_EXIT(&conn->conn_data_lock);
2116 #endif /* RX_ENABLE_LOCKS */
2117 call->state = RX_STATE_DALLY;
2119 error = call->error;
2121 /* currentPacket, nLeft, and NFree must be zeroed here, because
2122 * ResetCall cannot: ResetCall may be called at splnet(), in the
2123 * kernel version, and may interrupt the macros rx_Read or
2124 * rx_Write, which run at normal priority for efficiency. */
2125 if (call->currentPacket) {
2126 #ifdef RX_TRACK_PACKETS
2127 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2129 rxi_FreePacket(call->currentPacket);
2130 call->currentPacket = (struct rx_packet *)0;
2133 call->nLeft = call->nFree = call->curlen = 0;
2135 /* Free any packets from the last call to ReadvProc/WritevProc */
2136 #ifdef RXDEBUG_PACKET
2138 #endif /* RXDEBUG_PACKET */
2139 rxi_FreePackets(0, &call->iovq);
2141 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2142 MUTEX_EXIT(&call->lock);
2143 if (conn->type == RX_CLIENT_CONNECTION) {
2144 MUTEX_ENTER(&conn->conn_data_lock);
2145 conn->flags &= ~RX_CONN_BUSY;
2146 MUTEX_EXIT(&conn->conn_data_lock);
2147 MUTEX_EXIT(&conn->conn_call_lock);
2151 * Map errors to the local host's errno.h format.
2153 error = ntoh_syserr_conv(error);
2157 #if !defined(KERNEL)
2159 /* Call this routine when shutting down a server or client (especially
2160 * clients). This will allow Rx to gracefully garbage collect server
2161 * connections, and reduce the number of retries that a server might
2162 * make to a dead client.
2163 * This is not quite right, since some calls may still be ongoing and
2164 * we can't lock them to destroy them. */
2168 struct rx_connection **conn_ptr, **conn_end;
2172 if (rxinit_status == 1) {
2174 return; /* Already shutdown. */
2176 rxi_DeleteCachedConnections();
2177 if (rx_connHashTable) {
2178 MUTEX_ENTER(&rx_connHashTable_lock);
2179 for (conn_ptr = &rx_connHashTable[0], conn_end =
2180 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2182 struct rx_connection *conn, *next;
2183 for (conn = *conn_ptr; conn; conn = next) {
2185 if (conn->type == RX_CLIENT_CONNECTION) {
2186 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2188 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2189 #ifdef RX_ENABLE_LOCKS
2190 rxi_DestroyConnectionNoLock(conn);
2191 #else /* RX_ENABLE_LOCKS */
2192 rxi_DestroyConnection(conn);
2193 #endif /* RX_ENABLE_LOCKS */
2197 #ifdef RX_ENABLE_LOCKS
2198 while (rx_connCleanup_list) {
2199 struct rx_connection *conn;
2200 conn = rx_connCleanup_list;
2201 rx_connCleanup_list = rx_connCleanup_list->next;
2202 MUTEX_EXIT(&rx_connHashTable_lock);
2203 rxi_CleanupConnection(conn);
2204 MUTEX_ENTER(&rx_connHashTable_lock);
2206 MUTEX_EXIT(&rx_connHashTable_lock);
2207 #endif /* RX_ENABLE_LOCKS */
2212 afs_winsockCleanup();
2220 /* if we wakeup packet waiter too often, can get in loop with two
2221 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2223 rxi_PacketsUnWait(void)
2225 if (!rx_waitingForPackets) {
2229 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2230 return; /* still over quota */
2233 rx_waitingForPackets = 0;
2234 #ifdef RX_ENABLE_LOCKS
2235 CV_BROADCAST(&rx_waitingForPackets_cv);
2237 osi_rxWakeup(&rx_waitingForPackets);
2243 /* ------------------Internal interfaces------------------------- */
2245 /* Return this process's service structure for the
2246 * specified socket and service */
2248 rxi_FindService(osi_socket socket, u_short serviceId)
2250 struct rx_service **sp;
2251 for (sp = &rx_services[0]; *sp; sp++) {
2252 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2258 #ifdef RXDEBUG_PACKET
2259 #ifdef KDUMP_RX_LOCK
2260 static struct rx_call_rx_lock *rx_allCallsp = 0;
2262 static struct rx_call *rx_allCallsp = 0;
2264 #endif /* RXDEBUG_PACKET */
2266 /* Allocate a call structure, for the indicated channel of the
2267 * supplied connection. The mode and state of the call must be set by
2268 * the caller. Returns the call with mutex locked. */
2270 rxi_NewCall(struct rx_connection *conn, int channel)
2272 struct rx_call *call;
2273 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2274 struct rx_call *cp; /* Call pointer temp */
2275 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2276 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2278 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2280 /* Grab an existing call structure, or allocate a new one.
2281 * Existing call structures are assumed to have been left reset by
2283 MUTEX_ENTER(&rx_freeCallQueue_lock);
2285 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2287 * EXCEPT that the TQ might not yet be cleared out.
2288 * Skip over those with in-use TQs.
2291 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2292 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2298 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2299 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2300 call = queue_First(&rx_freeCallQueue, rx_call);
2301 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2303 if (rx_stats_active)
2304 rx_MutexDecrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2305 MUTEX_EXIT(&rx_freeCallQueue_lock);
2306 MUTEX_ENTER(&call->lock);
2307 CLEAR_CALL_QUEUE_LOCK(call);
2308 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2309 /* Now, if TQ wasn't cleared earlier, do it now. */
2310 rxi_WaitforTQBusy(call);
2311 if (call->flags & RX_CALL_TQ_CLEARME) {
2312 rxi_ClearTransmitQueue(call, 1);
2313 /*queue_Init(&call->tq);*/
2315 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2316 /* Bind the call to its connection structure */
2318 rxi_ResetCall(call, 1);
2321 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2322 #ifdef RXDEBUG_PACKET
2323 call->allNextp = rx_allCallsp;
2324 rx_allCallsp = call;
2326 #endif /* RXDEBUG_PACKET */
2327 rx_MutexIncrement(rx_stats.nCallStructs, rx_stats_mutex);
2329 MUTEX_EXIT(&rx_freeCallQueue_lock);
2330 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2331 MUTEX_ENTER(&call->lock);
2332 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2333 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2334 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2336 /* Initialize once-only items */
2337 queue_Init(&call->tq);
2338 queue_Init(&call->rq);
2339 queue_Init(&call->iovq);
2340 #ifdef RXDEBUG_PACKET
2341 call->rqc = call->tqc = call->iovqc = 0;
2342 #endif /* RXDEBUG_PACKET */
2343 /* Bind the call to its connection structure (prereq for reset) */
2345 rxi_ResetCall(call, 1);
2347 call->channel = channel;
2348 call->callNumber = &conn->callNumber[channel];
2349 call->rwind = conn->rwind[channel];
2350 call->twind = conn->twind[channel];
2351 /* Note that the next expected call number is retained (in
2352 * conn->callNumber[i]), even if we reallocate the call structure
2354 conn->call[channel] = call;
2355 /* if the channel's never been used (== 0), we should start at 1, otherwise
2356 * the call number is valid from the last time this channel was used */
2357 if (*call->callNumber == 0)
2358 *call->callNumber = 1;
2363 /* A call has been inactive long enough that so we can throw away
2364 * state, including the call structure, which is placed on the call
2366 * Call is locked upon entry.
2367 * haveCTLock set if called from rxi_ReapConnections
2369 #ifdef RX_ENABLE_LOCKS
2371 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2372 #else /* RX_ENABLE_LOCKS */
2374 rxi_FreeCall(struct rx_call *call)
2375 #endif /* RX_ENABLE_LOCKS */
2377 int channel = call->channel;
2378 struct rx_connection *conn = call->conn;
2381 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2382 (*call->callNumber)++;
2383 rxi_ResetCall(call, 0);
2384 call->conn->call[channel] = (struct rx_call *)0;
2386 MUTEX_ENTER(&rx_freeCallQueue_lock);
2387 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2388 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2389 /* A call may be free even though its transmit queue is still in use.
2390 * Since we search the call list from head to tail, put busy calls at
2391 * the head of the list, and idle calls at the tail.
2393 if (call->flags & RX_CALL_TQ_BUSY)
2394 queue_Prepend(&rx_freeCallQueue, call);
2396 queue_Append(&rx_freeCallQueue, call);
2397 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2398 queue_Append(&rx_freeCallQueue, call);
2399 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2400 if (rx_stats_active)
2401 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2402 MUTEX_EXIT(&rx_freeCallQueue_lock);
2404 /* Destroy the connection if it was previously slated for
2405 * destruction, i.e. the Rx client code previously called
2406 * rx_DestroyConnection (client connections), or
2407 * rxi_ReapConnections called the same routine (server
2408 * connections). Only do this, however, if there are no
2409 * outstanding calls. Note that for fine grain locking, there appears
2410 * to be a deadlock in that rxi_FreeCall has a call locked and
2411 * DestroyConnectionNoLock locks each call in the conn. But note a
2412 * few lines up where we have removed this call from the conn.
2413 * If someone else destroys a connection, they either have no
2414 * call lock held or are going through this section of code.
2416 MUTEX_ENTER(&conn->conn_data_lock);
2417 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2419 MUTEX_EXIT(&conn->conn_data_lock);
2420 #ifdef RX_ENABLE_LOCKS
2422 rxi_DestroyConnectionNoLock(conn);
2424 rxi_DestroyConnection(conn);
2425 #else /* RX_ENABLE_LOCKS */
2426 rxi_DestroyConnection(conn);
2427 #endif /* RX_ENABLE_LOCKS */
2429 MUTEX_EXIT(&conn->conn_data_lock);
2433 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2435 rxi_Alloc(size_t size)
2439 if (rx_stats_active)
2440 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2443 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2444 afs_osi_Alloc_NoSleep(size);
2449 osi_Panic("rxi_Alloc error");
2455 rxi_Free(void *addr, size_t size)
2457 if (rx_stats_active)
2458 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2459 osi_Free(addr, size);
2463 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2465 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2466 struct rx_peer *next = NULL;
2470 MUTEX_ENTER(&rx_peerHashTable_lock);
2472 peer_ptr = &rx_peerHashTable[0];
2473 peer_end = &rx_peerHashTable[rx_hashTableSize];
2476 for ( ; peer_ptr < peer_end; peer_ptr++) {
2479 for ( ; peer; peer = next) {
2481 if (host == peer->host)
2486 hashIndex = PEER_HASH(host, port);
2487 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2488 if ((peer->host == host) && (peer->port == port))
2493 MUTEX_ENTER(&rx_peerHashTable_lock);
2498 MUTEX_EXIT(&rx_peerHashTable_lock);
2500 MUTEX_ENTER(&peer->peer_lock);
2501 /* We don't handle dropping below min, so don't */
2502 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2503 peer->ifMTU=MIN(mtu, peer->ifMTU);
2504 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2505 /* if we tweaked this down, need to tune our peer MTU too */
2506 peer->MTU = MIN(peer->MTU, peer->natMTU);
2507 /* if we discovered a sub-1500 mtu, degrade */
2508 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2509 peer->maxDgramPackets = 1;
2510 /* We no longer have valid peer packet information */
2511 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2512 peer->maxPacketSize = 0;
2513 MUTEX_EXIT(&peer->peer_lock);
2515 MUTEX_ENTER(&rx_peerHashTable_lock);
2517 if (host && !port) {
2519 /* pick up where we left off */
2523 MUTEX_EXIT(&rx_peerHashTable_lock);
2526 /* Find the peer process represented by the supplied (host,port)
2527 * combination. If there is no appropriate active peer structure, a
2528 * new one will be allocated and initialized
2529 * The origPeer, if set, is a pointer to a peer structure on which the
2530 * refcount will be be decremented. This is used to replace the peer
2531 * structure hanging off a connection structure */
2533 rxi_FindPeer(afs_uint32 host, u_short port,
2534 struct rx_peer *origPeer, int create)
2538 hashIndex = PEER_HASH(host, port);
2539 MUTEX_ENTER(&rx_peerHashTable_lock);
2540 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2541 if ((pp->host == host) && (pp->port == port))
2546 pp = rxi_AllocPeer(); /* This bzero's *pp */
2547 pp->host = host; /* set here or in InitPeerParams is zero */
2549 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2550 queue_Init(&pp->congestionQueue);
2551 queue_Init(&pp->rpcStats);
2552 pp->next = rx_peerHashTable[hashIndex];
2553 rx_peerHashTable[hashIndex] = pp;
2554 rxi_InitPeerParams(pp);
2555 if (rx_stats_active)
2556 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2563 origPeer->refCount--;
2564 MUTEX_EXIT(&rx_peerHashTable_lock);
2569 /* Find the connection at (host, port) started at epoch, and with the
2570 * given connection id. Creates the server connection if necessary.
2571 * The type specifies whether a client connection or a server
2572 * connection is desired. In both cases, (host, port) specify the
2573 * peer's (host, pair) pair. Client connections are not made
2574 * automatically by this routine. The parameter socket gives the
2575 * socket descriptor on which the packet was received. This is used,
2576 * in the case of server connections, to check that *new* connections
2577 * come via a valid (port, serviceId). Finally, the securityIndex
2578 * parameter must match the existing index for the connection. If a
2579 * server connection is created, it will be created using the supplied
2580 * index, if the index is valid for this service */
2581 struct rx_connection *
2582 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2583 u_short port, u_short serviceId, afs_uint32 cid,
2584 afs_uint32 epoch, int type, u_int securityIndex)
2586 int hashindex, flag, i;
2587 struct rx_connection *conn;
2588 hashindex = CONN_HASH(host, port, cid, epoch, type);
2589 MUTEX_ENTER(&rx_connHashTable_lock);
2590 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2591 rx_connHashTable[hashindex],
2594 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2595 && (epoch == conn->epoch)) {
2596 struct rx_peer *pp = conn->peer;
2597 if (securityIndex != conn->securityIndex) {
2598 /* this isn't supposed to happen, but someone could forge a packet
2599 * like this, and there seems to be some CM bug that makes this
2600 * happen from time to time -- in which case, the fileserver
2602 MUTEX_EXIT(&rx_connHashTable_lock);
2603 return (struct rx_connection *)0;
2605 if (pp->host == host && pp->port == port)
2607 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2609 /* So what happens when it's a callback connection? */
2610 if ( /*type == RX_CLIENT_CONNECTION && */
2611 (conn->epoch & 0x80000000))
2615 /* the connection rxLastConn that was used the last time is not the
2616 ** one we are looking for now. Hence, start searching in the hash */
2618 conn = rx_connHashTable[hashindex];
2623 struct rx_service *service;
2624 if (type == RX_CLIENT_CONNECTION) {
2625 MUTEX_EXIT(&rx_connHashTable_lock);
2626 return (struct rx_connection *)0;
2628 service = rxi_FindService(socket, serviceId);
2629 if (!service || (securityIndex >= service->nSecurityObjects)
2630 || (service->securityObjects[securityIndex] == 0)) {
2631 MUTEX_EXIT(&rx_connHashTable_lock);
2632 return (struct rx_connection *)0;
2634 conn = rxi_AllocConnection(); /* This bzero's the connection */
2635 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2636 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2637 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2638 conn->next = rx_connHashTable[hashindex];
2639 rx_connHashTable[hashindex] = conn;
2640 conn->peer = rxi_FindPeer(host, port, 0, 1);
2641 conn->type = RX_SERVER_CONNECTION;
2642 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2643 conn->epoch = epoch;
2644 conn->cid = cid & RX_CIDMASK;
2645 /* conn->serial = conn->lastSerial = 0; */
2646 /* conn->timeout = 0; */
2647 conn->ackRate = RX_FAST_ACK_RATE;
2648 conn->service = service;
2649 conn->serviceId = serviceId;
2650 conn->securityIndex = securityIndex;
2651 conn->securityObject = service->securityObjects[securityIndex];
2652 conn->nSpecific = 0;
2653 conn->specific = NULL;
2654 rx_SetConnDeadTime(conn, service->connDeadTime);
2655 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2656 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2657 for (i = 0; i < RX_MAXCALLS; i++) {
2658 conn->twind[i] = rx_initSendWindow;
2659 conn->rwind[i] = rx_initReceiveWindow;
2661 /* Notify security object of the new connection */
2662 RXS_NewConnection(conn->securityObject, conn);
2663 /* XXXX Connection timeout? */
2664 if (service->newConnProc)
2665 (*service->newConnProc) (conn);
2666 if (rx_stats_active)
2667 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2670 MUTEX_ENTER(&conn->conn_data_lock);
2672 MUTEX_EXIT(&conn->conn_data_lock);
2674 rxLastConn = conn; /* store this connection as the last conn used */
2675 MUTEX_EXIT(&rx_connHashTable_lock);
2679 /* There are two packet tracing routines available for testing and monitoring
2680 * Rx. One is called just after every packet is received and the other is
2681 * called just before every packet is sent. Received packets, have had their
2682 * headers decoded, and packets to be sent have not yet had their headers
2683 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2684 * containing the network address. Both can be modified. The return value, if
2685 * non-zero, indicates that the packet should be dropped. */
2687 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2688 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2690 /* A packet has been received off the interface. Np is the packet, socket is
2691 * the socket number it was received from (useful in determining which service
2692 * this packet corresponds to), and (host, port) reflect the host,port of the
2693 * sender. This call returns the packet to the caller if it is finished with
2694 * it, rather than de-allocating it, just as a small performance hack */
2697 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
2698 afs_uint32 host, u_short port, int *tnop,
2699 struct rx_call **newcallp)
2701 struct rx_call *call;
2702 struct rx_connection *conn;
2704 afs_uint32 currentCallNumber;
2710 struct rx_packet *tnp;
2713 /* We don't print out the packet until now because (1) the time may not be
2714 * accurate enough until now in the lwp implementation (rx_Listener only gets
2715 * the time after the packet is read) and (2) from a protocol point of view,
2716 * this is the first time the packet has been seen */
2717 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2718 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2719 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT,
2720 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2721 np->header.epoch, np->header.cid, np->header.callNumber,
2722 np->header.seq, np->header.flags, np));
2725 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2726 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2729 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2730 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2733 /* If an input tracer function is defined, call it with the packet and
2734 * network address. Note this function may modify its arguments. */
2735 if (rx_justReceived) {
2736 struct sockaddr_in addr;
2738 addr.sin_family = AF_INET;
2739 addr.sin_port = port;
2740 addr.sin_addr.s_addr = host;
2741 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2742 addr.sin_len = sizeof(addr);
2743 #endif /* AFS_OSF_ENV */
2744 drop = (*rx_justReceived) (np, &addr);
2745 /* drop packet if return value is non-zero */
2748 port = addr.sin_port; /* in case fcn changed addr */
2749 host = addr.sin_addr.s_addr;
2753 /* If packet was not sent by the client, then *we* must be the client */
2754 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2755 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2757 /* Find the connection (or fabricate one, if we're the server & if
2758 * necessary) associated with this packet */
2760 rxi_FindConnection(socket, host, port, np->header.serviceId,
2761 np->header.cid, np->header.epoch, type,
2762 np->header.securityIndex);
2765 /* If no connection found or fabricated, just ignore the packet.
2766 * (An argument could be made for sending an abort packet for
2771 MUTEX_ENTER(&conn->conn_data_lock);
2772 if (conn->maxSerial < np->header.serial)
2773 conn->maxSerial = np->header.serial;
2774 MUTEX_EXIT(&conn->conn_data_lock);
2776 /* If the connection is in an error state, send an abort packet and ignore
2777 * the incoming packet */
2779 /* Don't respond to an abort packet--we don't want loops! */
2780 MUTEX_ENTER(&conn->conn_data_lock);
2781 if (np->header.type != RX_PACKET_TYPE_ABORT)
2782 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2784 MUTEX_EXIT(&conn->conn_data_lock);
2788 /* Check for connection-only requests (i.e. not call specific). */
2789 if (np->header.callNumber == 0) {
2790 switch (np->header.type) {
2791 case RX_PACKET_TYPE_ABORT: {
2792 /* What if the supplied error is zero? */
2793 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2794 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2795 rxi_ConnectionError(conn, errcode);
2796 MUTEX_ENTER(&conn->conn_data_lock);
2798 MUTEX_EXIT(&conn->conn_data_lock);
2801 case RX_PACKET_TYPE_CHALLENGE:
2802 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2803 MUTEX_ENTER(&conn->conn_data_lock);
2805 MUTEX_EXIT(&conn->conn_data_lock);
2807 case RX_PACKET_TYPE_RESPONSE:
2808 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2809 MUTEX_ENTER(&conn->conn_data_lock);
2811 MUTEX_EXIT(&conn->conn_data_lock);
2813 case RX_PACKET_TYPE_PARAMS:
2814 case RX_PACKET_TYPE_PARAMS + 1:
2815 case RX_PACKET_TYPE_PARAMS + 2:
2816 /* ignore these packet types for now */
2817 MUTEX_ENTER(&conn->conn_data_lock);
2819 MUTEX_EXIT(&conn->conn_data_lock);
2824 /* Should not reach here, unless the peer is broken: send an
2826 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2827 MUTEX_ENTER(&conn->conn_data_lock);
2828 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2830 MUTEX_EXIT(&conn->conn_data_lock);
2835 channel = np->header.cid & RX_CHANNELMASK;
2836 call = conn->call[channel];
2837 #ifdef RX_ENABLE_LOCKS
2839 MUTEX_ENTER(&call->lock);
2840 /* Test to see if call struct is still attached to conn. */
2841 if (call != conn->call[channel]) {
2843 MUTEX_EXIT(&call->lock);
2844 if (type == RX_SERVER_CONNECTION) {
2845 call = conn->call[channel];
2846 /* If we started with no call attached and there is one now,
2847 * another thread is also running this routine and has gotten
2848 * the connection channel. We should drop this packet in the tests
2849 * below. If there was a call on this connection and it's now
2850 * gone, then we'll be making a new call below.
2851 * If there was previously a call and it's now different then
2852 * the old call was freed and another thread running this routine
2853 * has created a call on this channel. One of these two threads
2854 * has a packet for the old call and the code below handles those
2858 MUTEX_ENTER(&call->lock);
2860 /* This packet can't be for this call. If the new call address is
2861 * 0 then no call is running on this channel. If there is a call
2862 * then, since this is a client connection we're getting data for
2863 * it must be for the previous call.
2865 if (rx_stats_active)
2866 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2867 MUTEX_ENTER(&conn->conn_data_lock);
2869 MUTEX_EXIT(&conn->conn_data_lock);
2874 currentCallNumber = conn->callNumber[channel];
2876 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2877 if (np->header.callNumber < currentCallNumber) {
2878 if (rx_stats_active)
2879 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2880 #ifdef RX_ENABLE_LOCKS
2882 MUTEX_EXIT(&call->lock);
2884 MUTEX_ENTER(&conn->conn_data_lock);
2886 MUTEX_EXIT(&conn->conn_data_lock);
2890 MUTEX_ENTER(&conn->conn_call_lock);
2891 call = rxi_NewCall(conn, channel);
2892 MUTEX_EXIT(&conn->conn_call_lock);
2893 *call->callNumber = np->header.callNumber;
2895 if (np->header.callNumber == 0)
2896 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%.06d len %d",
2897 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
2898 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
2899 np->header.flags, np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2901 call->state = RX_STATE_PRECALL;
2902 clock_GetTime(&call->queueTime);
2903 hzero(call->bytesSent);
2904 hzero(call->bytesRcvd);
2906 * If the number of queued calls exceeds the overload
2907 * threshold then abort this call.
2909 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2910 struct rx_packet *tp;
2912 rxi_CallError(call, rx_BusyError);
2913 tp = rxi_SendCallAbort(call, np, 1, 0);
2914 MUTEX_EXIT(&call->lock);
2915 MUTEX_ENTER(&conn->conn_data_lock);
2917 MUTEX_EXIT(&conn->conn_data_lock);
2918 if (rx_stats_active)
2919 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2922 rxi_KeepAliveOn(call);
2923 } else if (np->header.callNumber != currentCallNumber) {
2924 /* Wait until the transmit queue is idle before deciding
2925 * whether to reset the current call. Chances are that the
2926 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2929 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2930 while ((call->state == RX_STATE_ACTIVE)
2931 && (call->flags & RX_CALL_TQ_BUSY)) {
2932 call->flags |= RX_CALL_TQ_WAIT;
2934 #ifdef RX_ENABLE_LOCKS
2935 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2936 CV_WAIT(&call->cv_tq, &call->lock);
2937 #else /* RX_ENABLE_LOCKS */
2938 osi_rxSleep(&call->tq);
2939 #endif /* RX_ENABLE_LOCKS */
2941 if (call->tqWaiters == 0)
2942 call->flags &= ~RX_CALL_TQ_WAIT;
2944 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2945 /* If the new call cannot be taken right now send a busy and set
2946 * the error condition in this call, so that it terminates as
2947 * quickly as possible */
2948 if (call->state == RX_STATE_ACTIVE) {
2949 struct rx_packet *tp;
2951 rxi_CallError(call, RX_CALL_DEAD);
2952 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2954 MUTEX_EXIT(&call->lock);
2955 MUTEX_ENTER(&conn->conn_data_lock);
2957 MUTEX_EXIT(&conn->conn_data_lock);
2960 rxi_ResetCall(call, 0);
2961 *call->callNumber = np->header.callNumber;
2963 if (np->header.callNumber == 0)
2964 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%06d len %d",
2965 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
2966 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
2967 np->header.flags, np, np->retryTime.sec, np->retryTime.usec, np->length));
2969 call->state = RX_STATE_PRECALL;
2970 clock_GetTime(&call->queueTime);
2971 hzero(call->bytesSent);
2972 hzero(call->bytesRcvd);
2974 * If the number of queued calls exceeds the overload
2975 * threshold then abort this call.
2977 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2978 struct rx_packet *tp;
2980 rxi_CallError(call, rx_BusyError);
2981 tp = rxi_SendCallAbort(call, np, 1, 0);
2982 MUTEX_EXIT(&call->lock);
2983 MUTEX_ENTER(&conn->conn_data_lock);
2985 MUTEX_EXIT(&conn->conn_data_lock);
2986 if (rx_stats_active)
2987 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2990 rxi_KeepAliveOn(call);
2992 /* Continuing call; do nothing here. */
2994 } else { /* we're the client */
2995 /* Ignore all incoming acknowledgements for calls in DALLY state */
2996 if (call && (call->state == RX_STATE_DALLY)
2997 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2998 if (rx_stats_active)
2999 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
3000 #ifdef RX_ENABLE_LOCKS
3002 MUTEX_EXIT(&call->lock);
3005 MUTEX_ENTER(&conn->conn_data_lock);
3007 MUTEX_EXIT(&conn->conn_data_lock);
3011 /* Ignore anything that's not relevant to the current call. If there
3012 * isn't a current call, then no packet is relevant. */
3013 if (!call || (np->header.callNumber != currentCallNumber)) {
3014 if (rx_stats_active)
3015 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
3016 #ifdef RX_ENABLE_LOCKS
3018 MUTEX_EXIT(&call->lock);
3021 MUTEX_ENTER(&conn->conn_data_lock);
3023 MUTEX_EXIT(&conn->conn_data_lock);
3026 /* If the service security object index stamped in the packet does not
3027 * match the connection's security index, ignore the packet */
3028 if (np->header.securityIndex != conn->securityIndex) {
3029 #ifdef RX_ENABLE_LOCKS
3030 MUTEX_EXIT(&call->lock);
3032 MUTEX_ENTER(&conn->conn_data_lock);
3034 MUTEX_EXIT(&conn->conn_data_lock);
3038 /* If we're receiving the response, then all transmit packets are
3039 * implicitly acknowledged. Get rid of them. */
3040 if (np->header.type == RX_PACKET_TYPE_DATA) {
3041 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3042 /* XXX Hack. Because we must release the global rx lock when
3043 * sending packets (osi_NetSend) we drop all acks while we're
3044 * traversing the tq in rxi_Start sending packets out because
3045 * packets may move to the freePacketQueue as result of being here!
3046 * So we drop these packets until we're safely out of the
3047 * traversing. Really ugly!
3048 * For fine grain RX locking, we set the acked field in the
3049 * packets and let rxi_Start remove them from the transmit queue.
3051 if (call->flags & RX_CALL_TQ_BUSY) {
3052 #ifdef RX_ENABLE_LOCKS
3053 rxi_SetAcksInTransmitQueue(call);
3056 return np; /* xmitting; drop packet */
3059 rxi_ClearTransmitQueue(call, 0);
3061 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3062 rxi_ClearTransmitQueue(call, 0);
3063 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3065 if (np->header.type == RX_PACKET_TYPE_ACK) {
3066 /* now check to see if this is an ack packet acknowledging that the
3067 * server actually *lost* some hard-acked data. If this happens we
3068 * ignore this packet, as it may indicate that the server restarted in
3069 * the middle of a call. It is also possible that this is an old ack
3070 * packet. We don't abort the connection in this case, because this
3071 * *might* just be an old ack packet. The right way to detect a server
3072 * restart in the midst of a call is to notice that the server epoch
3074 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3075 * XXX unacknowledged. I think that this is off-by-one, but
3076 * XXX I don't dare change it just yet, since it will
3077 * XXX interact badly with the server-restart detection
3078 * XXX code in receiveackpacket. */
3079 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3080 if (rx_stats_active)
3081 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
3082 MUTEX_EXIT(&call->lock);
3083 MUTEX_ENTER(&conn->conn_data_lock);
3085 MUTEX_EXIT(&conn->conn_data_lock);
3089 } /* else not a data packet */
3092 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3093 /* Set remote user defined status from packet */
3094 call->remoteStatus = np->header.userStatus;
3096 /* Note the gap between the expected next packet and the actual
3097 * packet that arrived, when the new packet has a smaller serial number
3098 * than expected. Rioses frequently reorder packets all by themselves,
3099 * so this will be quite important with very large window sizes.
3100 * Skew is checked against 0 here to avoid any dependence on the type of
3101 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3103 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3104 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3105 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3107 MUTEX_ENTER(&conn->conn_data_lock);
3108 skew = conn->lastSerial - np->header.serial;
3109 conn->lastSerial = np->header.serial;
3110 MUTEX_EXIT(&conn->conn_data_lock);
3112 struct rx_peer *peer;
3114 if (skew > peer->inPacketSkew) {
3115 dpf(("*** In skew changed from %d to %d\n",
3116 peer->inPacketSkew, skew));
3117 peer->inPacketSkew = skew;
3121 /* Now do packet type-specific processing */
3122 switch (np->header.type) {
3123 case RX_PACKET_TYPE_DATA:
3124 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3127 case RX_PACKET_TYPE_ACK:
3128 /* Respond immediately to ack packets requesting acknowledgement
3130 if (np->header.flags & RX_REQUEST_ACK) {
3132 (void)rxi_SendCallAbort(call, 0, 1, 0);
3134 (void)rxi_SendAck(call, 0, np->header.serial,
3135 RX_ACK_PING_RESPONSE, 1);
3137 np = rxi_ReceiveAckPacket(call, np, 1);
3139 case RX_PACKET_TYPE_ABORT: {
3140 /* An abort packet: reset the call, passing the error up to the user. */
3141 /* What if error is zero? */
3142 /* What if the error is -1? the application will treat it as a timeout. */
3143 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3144 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
3145 rxi_CallError(call, errdata);
3146 MUTEX_EXIT(&call->lock);
3147 MUTEX_ENTER(&conn->conn_data_lock);
3149 MUTEX_EXIT(&conn->conn_data_lock);
3150 return np; /* xmitting; drop packet */
3152 case RX_PACKET_TYPE_BUSY:
3155 case RX_PACKET_TYPE_ACKALL:
3156 /* All packets acknowledged, so we can drop all packets previously
3157 * readied for sending */
3158 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3159 /* XXX Hack. We because we can't release the global rx lock when
3160 * sending packets (osi_NetSend) we drop all ack pkts while we're
3161 * traversing the tq in rxi_Start sending packets out because
3162 * packets may move to the freePacketQueue as result of being
3163 * here! So we drop these packets until we're safely out of the
3164 * traversing. Really ugly!
3165 * For fine grain RX locking, we set the acked field in the packets
3166 * and let rxi_Start remove the packets from the transmit queue.
3168 if (call->flags & RX_CALL_TQ_BUSY) {
3169 #ifdef RX_ENABLE_LOCKS
3170 rxi_SetAcksInTransmitQueue(call);
3172 #else /* RX_ENABLE_LOCKS */
3173 MUTEX_EXIT(&call->lock);
3174 MUTEX_ENTER(&conn->conn_data_lock);
3176 MUTEX_EXIT(&conn->conn_data_lock);
3177 return np; /* xmitting; drop packet */
3178 #endif /* RX_ENABLE_LOCKS */
3180 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3181 rxi_ClearTransmitQueue(call, 0);
3182 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3185 /* Should not reach here, unless the peer is broken: send an abort
3187 rxi_CallError(call, RX_PROTOCOL_ERROR);
3188 np = rxi_SendCallAbort(call, np, 1, 0);
3191 /* Note when this last legitimate packet was received, for keep-alive
3192 * processing. Note, we delay getting the time until now in the hope that
3193 * the packet will be delivered to the user before any get time is required
3194 * (if not, then the time won't actually be re-evaluated here). */
3195 call->lastReceiveTime = clock_Sec();
3196 MUTEX_EXIT(&call->lock);
3197 MUTEX_ENTER(&conn->conn_data_lock);
3199 MUTEX_EXIT(&conn->conn_data_lock);
3203 /* return true if this is an "interesting" connection from the point of view
3204 of someone trying to debug the system */
3206 rxi_IsConnInteresting(struct rx_connection *aconn)
3209 struct rx_call *tcall;
3211 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3214 for (i = 0; i < RX_MAXCALLS; i++) {
3215 tcall = aconn->call[i];
3217 if ((tcall->state == RX_STATE_PRECALL)
3218 || (tcall->state == RX_STATE_ACTIVE))
3220 if ((tcall->mode == RX_MODE_SENDING)
3221 || (tcall->mode == RX_MODE_RECEIVING))
3229 /* if this is one of the last few packets AND it wouldn't be used by the
3230 receiving call to immediately satisfy a read request, then drop it on
3231 the floor, since accepting it might prevent a lock-holding thread from
3232 making progress in its reading. If a call has been cleared while in
3233 the precall state then ignore all subsequent packets until the call
3234 is assigned to a thread. */
3237 TooLow(struct rx_packet *ap, struct rx_call *acall)
3241 MUTEX_ENTER(&rx_quota_mutex);
3242 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3243 && (acall->state == RX_STATE_PRECALL))
3244 || ((rx_nFreePackets < rxi_dataQuota + 2)
3245 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3246 && (acall->flags & RX_CALL_READER_WAIT)))) {
3249 MUTEX_EXIT(&rx_quota_mutex);
3255 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3257 struct rx_connection *conn = arg1;
3258 struct rx_call *acall = arg2;
3259 struct rx_call *call = acall;
3260 struct clock when, now;
3263 MUTEX_ENTER(&conn->conn_data_lock);
3264 conn->checkReachEvent = NULL;
3265 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3268 MUTEX_EXIT(&conn->conn_data_lock);
3272 MUTEX_ENTER(&conn->conn_call_lock);
3273 MUTEX_ENTER(&conn->conn_data_lock);
3274 for (i = 0; i < RX_MAXCALLS; i++) {
3275 struct rx_call *tc = conn->call[i];
3276 if (tc && tc->state == RX_STATE_PRECALL) {
3282 /* Indicate that rxi_CheckReachEvent is no longer running by
3283 * clearing the flag. Must be atomic under conn_data_lock to
3284 * avoid a new call slipping by: rxi_CheckConnReach holds
3285 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3287 conn->flags &= ~RX_CONN_ATTACHWAIT;
3288 MUTEX_EXIT(&conn->conn_data_lock);
3289 MUTEX_EXIT(&conn->conn_call_lock);
3294 MUTEX_ENTER(&call->lock);
3295 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3297 MUTEX_EXIT(&call->lock);
3299 clock_GetTime(&now);
3301 when.sec += RX_CHECKREACH_TIMEOUT;
3302 MUTEX_ENTER(&conn->conn_data_lock);
3303 if (!conn->checkReachEvent) {
3305 conn->checkReachEvent =
3306 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3309 MUTEX_EXIT(&conn->conn_data_lock);
3315 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3317 struct rx_service *service = conn->service;
3318 struct rx_peer *peer = conn->peer;
3319 afs_uint32 now, lastReach;
3321 if (service->checkReach == 0)
3325 MUTEX_ENTER(&peer->peer_lock);
3326 lastReach = peer->lastReachTime;
3327 MUTEX_EXIT(&peer->peer_lock);
3328 if (now - lastReach < RX_CHECKREACH_TTL)
3331 MUTEX_ENTER(&conn->conn_data_lock);
3332 if (conn->flags & RX_CONN_ATTACHWAIT) {
3333 MUTEX_EXIT(&conn->conn_data_lock);
3336 conn->flags |= RX_CONN_ATTACHWAIT;
3337 MUTEX_EXIT(&conn->conn_data_lock);
3338 if (!conn->checkReachEvent)
3339 rxi_CheckReachEvent(NULL, conn, call);
3344 /* try to attach call, if authentication is complete */
3346 TryAttach(struct rx_call *acall, osi_socket socket,
3347 int *tnop, struct rx_call **newcallp,
3350 struct rx_connection *conn = acall->conn;
3352 if (conn->type == RX_SERVER_CONNECTION
3353 && acall->state == RX_STATE_PRECALL) {
3354 /* Don't attach until we have any req'd. authentication. */
3355 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3356 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3357 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3358 /* Note: this does not necessarily succeed; there
3359 * may not any proc available
3362 rxi_ChallengeOn(acall->conn);
3367 /* A data packet has been received off the interface. This packet is
3368 * appropriate to the call (the call is in the right state, etc.). This
3369 * routine can return a packet to the caller, for re-use */
3372 rxi_ReceiveDataPacket(struct rx_call *call,
3373 struct rx_packet *np, int istack,
3374 osi_socket socket, afs_uint32 host, u_short port,
3375 int *tnop, struct rx_call **newcallp)
3377 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3382 afs_uint32 serial=0, flags=0;
3384 struct rx_packet *tnp;
3385 struct clock when, now;
3386 if (rx_stats_active)
3387 rx_MutexIncrement(rx_stats.dataPacketsRead, rx_stats_mutex);
3390 /* If there are no packet buffers, drop this new packet, unless we can find
3391 * packet buffers from inactive calls */
3393 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3394 MUTEX_ENTER(&rx_freePktQ_lock);
3395 rxi_NeedMorePackets = TRUE;
3396 MUTEX_EXIT(&rx_freePktQ_lock);
3397 if (rx_stats_active)
3398 rx_MutexIncrement(rx_stats.noPacketBuffersOnRead, rx_stats_mutex);
3399 call->rprev = np->header.serial;
3400 rxi_calltrace(RX_TRACE_DROP, call);
3401 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems", np));
3403 rxi_ClearReceiveQueue(call);
3404 clock_GetTime(&now);
3406 clock_Add(&when, &rx_softAckDelay);
3407 if (!call->delayedAckEvent
3408 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3409 rxevent_Cancel(call->delayedAckEvent, call,
3410 RX_CALL_REFCOUNT_DELAY);
3411 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3412 call->delayedAckEvent =
3413 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3415 /* we've damaged this call already, might as well do it in. */
3421 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3422 * packet is one of several packets transmitted as a single
3423 * datagram. Do not send any soft or hard acks until all packets
3424 * in a jumbogram have been processed. Send negative acks right away.
3426 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3427 /* tnp is non-null when there are more packets in the
3428 * current jumbo gram */
3435 seq = np->header.seq;
3436 serial = np->header.serial;
3437 flags = np->header.flags;
3439 /* If the call is in an error state, send an abort message */
3441 return rxi_SendCallAbort(call, np, istack, 0);
3443 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3444 * AFS 3.5 jumbogram. */
3445 if (flags & RX_JUMBO_PACKET) {
3446 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3451 if (np->header.spare != 0) {
3452 MUTEX_ENTER(&call->conn->conn_data_lock);
3453 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3454 MUTEX_EXIT(&call->conn->conn_data_lock);
3457 /* The usual case is that this is the expected next packet */
3458 if (seq == call->rnext) {
3460 /* Check to make sure it is not a duplicate of one already queued */
3461 if (queue_IsNotEmpty(&call->rq)
3462 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3463 if (rx_stats_active)
3464 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3465 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate", np));
3466 rxevent_Cancel(call->delayedAckEvent, call,
3467 RX_CALL_REFCOUNT_DELAY);
3468 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3474 /* It's the next packet. Stick it on the receive queue
3475 * for this call. Set newPackets to make sure we wake
3476 * the reader once all packets have been processed */
3477 #ifdef RX_TRACK_PACKETS
3478 np->flags |= RX_PKTFLAG_RQ;
3480 queue_Prepend(&call->rq, np);
3481 #ifdef RXDEBUG_PACKET
3483 #endif /* RXDEBUG_PACKET */
3485 np = NULL; /* We can't use this anymore */
3488 /* If an ack is requested then set a flag to make sure we
3489 * send an acknowledgement for this packet */
3490 if (flags & RX_REQUEST_ACK) {
3491 ackNeeded = RX_ACK_REQUESTED;
3494 /* Keep track of whether we have received the last packet */
3495 if (flags & RX_LAST_PACKET) {
3496 call->flags |= RX_CALL_HAVE_LAST;
3500 /* Check whether we have all of the packets for this call */
3501 if (call->flags & RX_CALL_HAVE_LAST) {
3502 afs_uint32 tseq; /* temporary sequence number */
3503 struct rx_packet *tp; /* Temporary packet pointer */
3504 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3506 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3507 if (tseq != tp->header.seq)
3509 if (tp->header.flags & RX_LAST_PACKET) {
3510 call->flags |= RX_CALL_RECEIVE_DONE;
3517 /* Provide asynchronous notification for those who want it
3518 * (e.g. multi rx) */
3519 if (call->arrivalProc) {
3520 (*call->arrivalProc) (call, call->arrivalProcHandle,
3521 call->arrivalProcArg);
3522 call->arrivalProc = (void (*)())0;
3525 /* Update last packet received */
3528 /* If there is no server process serving this call, grab
3529 * one, if available. We only need to do this once. If a
3530 * server thread is available, this thread becomes a server
3531 * thread and the server thread becomes a listener thread. */
3533 TryAttach(call, socket, tnop, newcallp, 0);
3536 /* This is not the expected next packet. */
3538 /* Determine whether this is a new or old packet, and if it's
3539 * a new one, whether it fits into the current receive window.
3540 * Also figure out whether the packet was delivered in sequence.
3541 * We use the prev variable to determine whether the new packet
3542 * is the successor of its immediate predecessor in the
3543 * receive queue, and the missing flag to determine whether
3544 * any of this packets predecessors are missing. */
3546 afs_uint32 prev; /* "Previous packet" sequence number */
3547 struct rx_packet *tp; /* Temporary packet pointer */
3548 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3549 int missing; /* Are any predecessors missing? */
3551 /* If the new packet's sequence number has been sent to the
3552 * application already, then this is a duplicate */
3553 if (seq < call->rnext) {
3554 if (rx_stats_active)
3555 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3556 rxevent_Cancel(call->delayedAckEvent, call,
3557 RX_CALL_REFCOUNT_DELAY);
3558 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3564 /* If the sequence number is greater than what can be
3565 * accomodated by the current window, then send a negative
3566 * acknowledge and drop the packet */
3567 if ((call->rnext + call->rwind) <= seq) {
3568 rxevent_Cancel(call->delayedAckEvent, call,
3569 RX_CALL_REFCOUNT_DELAY);
3570 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3577 /* Look for the packet in the queue of old received packets */
3578 for (prev = call->rnext - 1, missing =
3579 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3580 /*Check for duplicate packet */
3581 if (seq == tp->header.seq) {
3582 if (rx_stats_active)
3583 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3584 rxevent_Cancel(call->delayedAckEvent, call,
3585 RX_CALL_REFCOUNT_DELAY);
3586 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3592 /* If we find a higher sequence packet, break out and
3593 * insert the new packet here. */
3594 if (seq < tp->header.seq)
3596 /* Check for missing packet */
3597 if (tp->header.seq != prev + 1) {
3601 prev = tp->header.seq;
3604 /* Keep track of whether we have received the last packet. */
3605 if (flags & RX_LAST_PACKET) {
3606 call->flags |= RX_CALL_HAVE_LAST;
3609 /* It's within the window: add it to the the receive queue.
3610 * tp is left by the previous loop either pointing at the
3611 * packet before which to insert the new packet, or at the
3612 * queue head if the queue is empty or the packet should be
3614 #ifdef RX_TRACK_PACKETS
3615 np->flags |= RX_PKTFLAG_RQ;
3617 #ifdef RXDEBUG_PACKET
3619 #endif /* RXDEBUG_PACKET */
3620 queue_InsertBefore(tp, np);
3624 /* Check whether we have all of the packets for this call */
3625 if ((call->flags & RX_CALL_HAVE_LAST)
3626 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3627 afs_uint32 tseq; /* temporary sequence number */
3630 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3631 if (tseq != tp->header.seq)
3633 if (tp->header.flags & RX_LAST_PACKET) {
3634 call->flags |= RX_CALL_RECEIVE_DONE;
3641 /* We need to send an ack of the packet is out of sequence,
3642 * or if an ack was requested by the peer. */
3643 if (seq != prev + 1 || missing) {
3644 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3645 } else if (flags & RX_REQUEST_ACK) {
3646 ackNeeded = RX_ACK_REQUESTED;
3649 /* Acknowledge the last packet for each call */
3650 if (flags & RX_LAST_PACKET) {
3661 * If the receiver is waiting for an iovec, fill the iovec
3662 * using the data from the receive queue */
3663 if (call->flags & RX_CALL_IOVEC_WAIT) {
3664 didHardAck = rxi_FillReadVec(call, serial);
3665 /* the call may have been aborted */
3674 /* Wakeup the reader if any */
3675 if ((call->flags & RX_CALL_READER_WAIT)
3676 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3677 || (call->iovNext >= call->iovMax)
3678 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3679 call->flags &= ~RX_CALL_READER_WAIT;
3680 #ifdef RX_ENABLE_LOCKS
3681 CV_BROADCAST(&call->cv_rq);
3683 osi_rxWakeup(&call->rq);
3689 * Send an ack when requested by the peer, or once every
3690 * rxi_SoftAckRate packets until the last packet has been
3691 * received. Always send a soft ack for the last packet in
3692 * the server's reply. */
3694 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3695 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3696 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3697 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3698 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3699 } else if (call->nSoftAcks) {
3700 clock_GetTime(&now);
3702 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3703 clock_Add(&when, &rx_lastAckDelay);
3705 clock_Add(&when, &rx_softAckDelay);
3707 if (!call->delayedAckEvent
3708 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3709 rxevent_Cancel(call->delayedAckEvent, call,
3710 RX_CALL_REFCOUNT_DELAY);
3711 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3712 call->delayedAckEvent =
3713 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3715 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3716 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3723 static void rxi_ComputeRate();
3727 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3729 struct rx_peer *peer = conn->peer;
3731 MUTEX_ENTER(&peer->peer_lock);
3732 peer->lastReachTime = clock_Sec();
3733 MUTEX_EXIT(&peer->peer_lock);
3735 MUTEX_ENTER(&conn->conn_data_lock);
3736 if (conn->flags & RX_CONN_ATTACHWAIT) {
3739 conn->flags &= ~RX_CONN_ATTACHWAIT;
3740 MUTEX_EXIT(&conn->conn_data_lock);
3742 for (i = 0; i < RX_MAXCALLS; i++) {
3743 struct rx_call *call = conn->call[i];
3746 MUTEX_ENTER(&call->lock);
3747 /* tnop can be null if newcallp is null */
3748 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3750 MUTEX_EXIT(&call->lock);
3754 MUTEX_EXIT(&conn->conn_data_lock);
3757 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
3759 rx_ack_reason(int reason)
3762 case RX_ACK_REQUESTED:
3764 case RX_ACK_DUPLICATE:
3766 case RX_ACK_OUT_OF_SEQUENCE:
3768 case RX_ACK_EXCEEDS_WINDOW:
3770 case RX_ACK_NOSPACE:
3774 case RX_ACK_PING_RESPONSE:
3787 /* rxi_ComputePeerNetStats
3789 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3790 * estimates (like RTT and throughput) based on ack packets. Caller
3791 * must ensure that the packet in question is the right one (i.e.
3792 * serial number matches).
3795 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3796 struct rx_ackPacket *ap, struct rx_packet *np)
3798 struct rx_peer *peer = call->conn->peer;
3800 /* Use RTT if not delayed by client and
3801 * ignore packets that were retransmitted. */
3802 if (!(p->flags & RX_PKTFLAG_ACKED) &&
3803 ap->reason != RX_ACK_DELAY &&
3804 clock_Eq(&p->timeSent, &p->firstSent))
3805 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3807 rxi_ComputeRate(peer, call, p, np, ap->reason);
3811 /* The real smarts of the whole thing. */
3813 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
3816 struct rx_ackPacket *ap;
3818 struct rx_packet *tp;
3819 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3820 struct rx_connection *conn = call->conn;
3821 struct rx_peer *peer = conn->peer;
3824 /* because there are CM's that are bogus, sending weird values for this. */
3825 afs_uint32 skew = 0;
3831 int newAckCount = 0;
3832 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3833 int pktsize = 0; /* Set if we need to update the peer mtu */
3835 if (rx_stats_active)
3836 rx_MutexIncrement(rx_stats.ackPacketsRead, rx_stats_mutex);
3837 ap = (struct rx_ackPacket *)rx_DataOf(np);
3838 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3840 return np; /* truncated ack packet */
3842 /* depends on ack packet struct */
3843 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3844 first = ntohl(ap->firstPacket);
3845 serial = ntohl(ap->serial);
3846 /* temporarily disabled -- needs to degrade over time
3847 * skew = ntohs(ap->maxSkew); */
3849 /* Ignore ack packets received out of order */
3850 if (first < call->tfirst) {
3854 if (np->header.flags & RX_SLOW_START_OK) {
3855 call->flags |= RX_CALL_SLOW_START_OK;
3858 if (ap->reason == RX_ACK_PING_RESPONSE)
3859 rxi_UpdatePeerReach(conn, call);
3861 if (conn->lastPacketSizeSeq) {
3862 MUTEX_ENTER(&conn->conn_data_lock);
3863 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
3864 pktsize = conn->lastPacketSize;
3865 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
3867 MUTEX_EXIT(&conn->conn_data_lock);
3869 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
3870 MUTEX_ENTER(&conn->conn_data_lock);
3871 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
3872 /* process mtu ping ack */
3873 pktsize = conn->lastPingSize;
3874 conn->lastPingSizeSer = conn->lastPingSize = 0;
3876 MUTEX_EXIT(&conn->conn_data_lock);
3880 MUTEX_ENTER(&peer->peer_lock);
3882 * Start somewhere. Can't assume we can send what we can receive,
3883 * but we are clearly receiving.
3885 if (!peer->maxPacketSize)
3886 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
3888 if (pktsize > peer->maxPacketSize) {
3889 peer->maxPacketSize = pktsize;
3890 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
3891 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
3892 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
3893 rxi_ScheduleGrowMTUEvent(call, 1);
3896 MUTEX_EXIT(&peer->peer_lock);
3901 if (rxdebug_active) {
3905 len = _snprintf(msg, sizeof(msg),
3906 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3907 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3908 ntohl(ap->serial), ntohl(ap->previousPacket),
3909 (unsigned int)np->header.seq, (unsigned int)skew,
3910 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3914 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3915 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3919 OutputDebugString(msg);
3921 #else /* AFS_NT40_ENV */
3924 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3925 ap->reason, ntohl(ap->previousPacket),
3926 (unsigned int)np->header.seq, (unsigned int)serial,
3927 (unsigned int)skew, ntohl(ap->firstPacket));
3930 for (offset = 0; offset < nAcks; offset++)
3931 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3936 #endif /* AFS_NT40_ENV */
3939 /* Update the outgoing packet skew value to the latest value of
3940 * the peer's incoming packet skew value. The ack packet, of
3941 * course, could arrive out of order, but that won't affect things
3943 MUTEX_ENTER(&peer->peer_lock);
3944 peer->outPacketSkew = skew;
3946 /* Check for packets that no longer need to be transmitted, and
3947 * discard them. This only applies to packets positively
3948 * acknowledged as having been sent to the peer's upper level.
3949 * All other packets must be retained. So only packets with
3950 * sequence numbers < ap->firstPacket are candidates. */
3951 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3952 if (tp->header.seq >= first)
3954 call->tfirst = tp->header.seq + 1;
3955 rxi_ComputePeerNetStats(call, tp, ap, np);
3956 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3959 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3960 /* XXX Hack. Because we have to release the global rx lock when sending
3961 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3962 * in rxi_Start sending packets out because packets may move to the
3963 * freePacketQueue as result of being here! So we drop these packets until
3964 * we're safely out of the traversing. Really ugly!
3965 * To make it even uglier, if we're using fine grain locking, we can
3966 * set the ack bits in the packets and have rxi_Start remove the packets
3967 * when it's done transmitting.
3969 if (call->flags & RX_CALL_TQ_BUSY) {
3970 #ifdef RX_ENABLE_LOCKS
3971 tp->flags |= RX_PKTFLAG_ACKED;
3972 call->flags |= RX_CALL_TQ_SOME_ACKED;
3973 #else /* RX_ENABLE_LOCKS */
3975 #endif /* RX_ENABLE_LOCKS */
3977 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3980 #ifdef RX_TRACK_PACKETS
3981 tp->flags &= ~RX_PKTFLAG_TQ;
3983 #ifdef RXDEBUG_PACKET
3985 #endif /* RXDEBUG_PACKET */
3986 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3991 /* Give rate detector a chance to respond to ping requests */
3992 if (ap->reason == RX_ACK_PING_RESPONSE) {
3993 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3997 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3999 /* Now go through explicit acks/nacks and record the results in
4000 * the waiting packets. These are packets that can't be released
4001 * yet, even with a positive acknowledge. This positive
4002 * acknowledge only means the packet has been received by the
4003 * peer, not that it will be retained long enough to be sent to
4004 * the peer's upper level. In addition, reset the transmit timers
4005 * of any missing packets (those packets that must be missing
4006 * because this packet was out of sequence) */
4008 call->nSoftAcked = 0;
4009 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
4010 /* Update round trip time if the ack was stimulated on receipt
4012 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4013 #ifdef RX_ENABLE_LOCKS
4014 if (tp->header.seq >= first)
4015 #endif /* RX_ENABLE_LOCKS */
4016 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4017 rxi_ComputePeerNetStats(call, tp, ap, np);
4019 /* Set the acknowledge flag per packet based on the
4020 * information in the ack packet. An acknowlegded packet can
4021 * be downgraded when the server has discarded a packet it
4022 * soacked previously, or when an ack packet is received
4023 * out of sequence. */
4024 if (tp->header.seq < first) {
4025 /* Implicit ack information */
4026 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4029 tp->flags |= RX_PKTFLAG_ACKED;
4030 } else if (tp->header.seq < first + nAcks) {
4031 /* Explicit ack information: set it in the packet appropriately */
4032 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4033 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4035 tp->flags |= RX_PKTFLAG_ACKED;
4042 } else /* RX_ACK_TYPE_NACK */ {
4043 tp->flags &= ~RX_PKTFLAG_ACKED;
4047 tp->flags &= ~RX_PKTFLAG_ACKED;
4052 * Following the suggestion of Phil Kern, we back off the peer's
4053 * timeout value for future packets until a successful response
4054 * is received for an initial transmission.
4056 if (missing && !backedOff) {
4057 struct clock c = peer->timeout;
4058 struct clock max_to = {3, 0};
4060 clock_Add(&peer->timeout, &c);
4061 if (clock_Gt(&peer->timeout, &max_to))
4062 peer->timeout = max_to;
4066 /* If packet isn't yet acked, and it has been transmitted at least
4067 * once, reset retransmit time using latest timeout
4068 * ie, this should readjust the retransmit timer for all outstanding
4069 * packets... So we don't just retransmit when we should know better*/
4071 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
4072 tp->retryTime = tp->timeSent;
4073 clock_Add(&tp->retryTime, &peer->timeout);
4074 /* shift by eight because one quarter-sec ~ 256 milliseconds */
4075 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
4079 /* If the window has been extended by this acknowledge packet,
4080 * then wakeup a sender waiting in alloc for window space, or try
4081 * sending packets now, if he's been sitting on packets due to
4082 * lack of window space */
4083 if (call->tnext < (call->tfirst + call->twind)) {
4084 #ifdef RX_ENABLE_LOCKS
4085 CV_SIGNAL(&call->cv_twind);
4087 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4088 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4089 osi_rxWakeup(&call->twind);
4092 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4093 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4097 /* if the ack packet has a receivelen field hanging off it,
4098 * update our state */
4099 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4102 /* If the ack packet has a "recommended" size that is less than
4103 * what I am using now, reduce my size to match */
4104 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4105 (int)sizeof(afs_int32), &tSize);
4106 tSize = (afs_uint32) ntohl(tSize);
4107 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4109 /* Get the maximum packet size to send to this peer */
4110 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4112 tSize = (afs_uint32) ntohl(tSize);
4113 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4114 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4116 /* sanity check - peer might have restarted with different params.
4117 * If peer says "send less", dammit, send less... Peer should never
4118 * be unable to accept packets of the size that prior AFS versions would
4119 * send without asking. */
4120 if (peer->maxMTU != tSize) {
4121 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4123 peer->maxMTU = tSize;
4124 peer->MTU = MIN(tSize, peer->MTU);
4125 call->MTU = MIN(call->MTU, tSize);
4128 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4131 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4132 (int)sizeof(afs_int32), &tSize);
4133 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4134 if (tSize < call->twind) { /* smaller than our send */
4135 call->twind = tSize; /* window, we must send less... */
4136 call->ssthresh = MIN(call->twind, call->ssthresh);
4137 call->conn->twind[call->channel] = call->twind;
4140 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4141 * network MTU confused with the loopback MTU. Calculate the
4142 * maximum MTU here for use in the slow start code below.
4144 /* Did peer restart with older RX version? */
4145 if (peer->maxDgramPackets > 1) {
4146 peer->maxDgramPackets = 1;
4148 } else if (np->length >=
4149 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4152 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4153 sizeof(afs_int32), &tSize);
4154 tSize = (afs_uint32) ntohl(tSize);
4156 * As of AFS 3.5 we set the send window to match the receive window.
4158 if (tSize < call->twind) {
4159 call->twind = tSize;
4160 call->conn->twind[call->channel] = call->twind;
4161 call->ssthresh = MIN(call->twind, call->ssthresh);
4162 } else if (tSize > call->twind) {
4163 call->twind = tSize;
4164 call->conn->twind[call->channel] = call->twind;
4168 * As of AFS 3.5, a jumbogram is more than one fixed size
4169 * packet transmitted in a single UDP datagram. If the remote
4170 * MTU is smaller than our local MTU then never send a datagram
4171 * larger than the natural MTU.
4174 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4175 (int)sizeof(afs_int32), &tSize);
4176 maxDgramPackets = (afs_uint32) ntohl(tSize);
4177 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4179 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4180 maxDgramPackets = MIN(maxDgramPackets, tSize);
4181 if (maxDgramPackets > 1) {
4182 peer->maxDgramPackets = maxDgramPackets;
4183 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4185 peer->maxDgramPackets = 1;
4186 call->MTU = peer->natMTU;
4188 } else if (peer->maxDgramPackets > 1) {
4189 /* Restarted with lower version of RX */
4190 peer->maxDgramPackets = 1;
4192 } else if (peer->maxDgramPackets > 1
4193 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4194 /* Restarted with lower version of RX */
4195 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4196 peer->natMTU = OLD_MAX_PACKET_SIZE;
4197 peer->MTU = OLD_MAX_PACKET_SIZE;
4198 peer->maxDgramPackets = 1;
4199 peer->nDgramPackets = 1;
4201 call->MTU = OLD_MAX_PACKET_SIZE;
4206 * Calculate how many datagrams were successfully received after
4207 * the first missing packet and adjust the negative ack counter
4212 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4213 if (call->nNacks < nNacked) {
4214 call->nNacks = nNacked;
4217 call->nAcks += newAckCount;
4221 if (call->flags & RX_CALL_FAST_RECOVER) {
4223 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4225 call->flags &= ~RX_CALL_FAST_RECOVER;
4226 call->cwind = call->nextCwind;
4227 call->nextCwind = 0;
4230 call->nCwindAcks = 0;
4231 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4232 /* Three negative acks in a row trigger congestion recovery */
4233 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4234 MUTEX_EXIT(&peer->peer_lock);
4235 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4236 /* someone else is waiting to start recovery */
4239 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4240 rxi_WaitforTQBusy(call);
4241 MUTEX_ENTER(&peer->peer_lock);
4242 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4243 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4244 call->flags |= RX_CALL_FAST_RECOVER;
4245 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4247 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4248 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4249 call->nextCwind = call->ssthresh;
4252 peer->MTU = call->MTU;
4253 peer->cwind = call->nextCwind;
4254 peer->nDgramPackets = call->nDgramPackets;
4256 call->congestSeq = peer->congestSeq;
4257 /* Reset the resend times on the packets that were nacked
4258 * so we will retransmit as soon as the window permits*/
4259 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4261 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4262 clock_Zero(&tp->retryTime);
4264 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4269 /* If cwind is smaller than ssthresh, then increase
4270 * the window one packet for each ack we receive (exponential
4272 * If cwind is greater than or equal to ssthresh then increase
4273 * the congestion window by one packet for each cwind acks we
4274 * receive (linear growth). */
4275 if (call->cwind < call->ssthresh) {
4277 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4278 call->nCwindAcks = 0;
4280 call->nCwindAcks += newAckCount;
4281 if (call->nCwindAcks >= call->cwind) {
4282 call->nCwindAcks = 0;
4283 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4287 * If we have received several acknowledgements in a row then
4288 * it is time to increase the size of our datagrams
4290 if ((int)call->nAcks > rx_nDgramThreshold) {
4291 if (peer->maxDgramPackets > 1) {
4292 if (call->nDgramPackets < peer->maxDgramPackets) {
4293 call->nDgramPackets++;
4295 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4296 } else if (call->MTU < peer->maxMTU) {
4297 /* don't upgrade if we can't handle it */
4298 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4299 call->MTU = peer->ifMTU;
4301 call->MTU += peer->natMTU;
4302 call->MTU = MIN(call->MTU, peer->maxMTU);
4309 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4311 /* Servers need to hold the call until all response packets have
4312 * been acknowledged. Soft acks are good enough since clients
4313 * are not allowed to clear their receive queues. */
4314 if (call->state == RX_STATE_HOLD
4315 && call->tfirst + call->nSoftAcked >= call->tnext) {
4316 call->state = RX_STATE_DALLY;
4317 rxi_ClearTransmitQueue(call, 0);
4318 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4319 } else if (!queue_IsEmpty(&call->tq)) {
4320 rxi_Start(0, call, 0, istack);
4325 /* Received a response to a challenge packet */
4327 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4328 struct rx_packet *np, int istack)
4332 /* Ignore the packet if we're the client */
4333 if (conn->type == RX_CLIENT_CONNECTION)
4336 /* If already authenticated, ignore the packet (it's probably a retry) */
4337 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4340 /* Otherwise, have the security object evaluate the response packet */
4341 error = RXS_CheckResponse(conn->securityObject, conn, np);
4343 /* If the response is invalid, reset the connection, sending
4344 * an abort to the peer */
4348 rxi_ConnectionError(conn, error);
4349 MUTEX_ENTER(&conn->conn_data_lock);
4350 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4351 MUTEX_EXIT(&conn->conn_data_lock);
4354 /* If the response is valid, any calls waiting to attach
4355 * servers can now do so */
4358 for (i = 0; i < RX_MAXCALLS; i++) {
4359 struct rx_call *call = conn->call[i];
4361 MUTEX_ENTER(&call->lock);
4362 if (call->state == RX_STATE_PRECALL)
4363 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4364 /* tnop can be null if newcallp is null */
4365 MUTEX_EXIT(&call->lock);
4369 /* Update the peer reachability information, just in case
4370 * some calls went into attach-wait while we were waiting
4371 * for authentication..
4373 rxi_UpdatePeerReach(conn, NULL);
4378 /* A client has received an authentication challenge: the security
4379 * object is asked to cough up a respectable response packet to send
4380 * back to the server. The server is responsible for retrying the
4381 * challenge if it fails to get a response. */
4384 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4385 struct rx_packet *np, int istack)
4389 /* Ignore the challenge if we're the server */
4390 if (conn->type == RX_SERVER_CONNECTION)
4393 /* Ignore the challenge if the connection is otherwise idle; someone's
4394 * trying to use us as an oracle. */
4395 if (!rxi_HasActiveCalls(conn))
4398 /* Send the security object the challenge packet. It is expected to fill
4399 * in the response. */
4400 error = RXS_GetResponse(conn->securityObject, conn, np);
4402 /* If the security object is unable to return a valid response, reset the
4403 * connection and send an abort to the peer. Otherwise send the response
4404 * packet to the peer connection. */
4406 rxi_ConnectionError(conn, error);
4407 MUTEX_ENTER(&conn->conn_data_lock);
4408 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4409 MUTEX_EXIT(&conn->conn_data_lock);
4411 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4412 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4418 /* Find an available server process to service the current request in
4419 * the given call structure. If one isn't available, queue up this
4420 * call so it eventually gets one */
4422 rxi_AttachServerProc(struct rx_call *call,
4423 osi_socket socket, int *tnop,
4424 struct rx_call **newcallp)
4426 struct rx_serverQueueEntry *sq;
4427 struct rx_service *service = call->conn->service;
4430 /* May already be attached */
4431 if (call->state == RX_STATE_ACTIVE)
4434 MUTEX_ENTER(&rx_serverPool_lock);
4436 haveQuota = QuotaOK(service);
4437 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4438 /* If there are no processes available to service this call,
4439 * put the call on the incoming call queue (unless it's
4440 * already on the queue).
4442 #ifdef RX_ENABLE_LOCKS
4444 ReturnToServerPool(service);
4445 #endif /* RX_ENABLE_LOCKS */
4447 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4448 call->flags |= RX_CALL_WAIT_PROC;
4449 MUTEX_ENTER(&rx_waiting_mutex);
4452 MUTEX_EXIT(&rx_waiting_mutex);
4453 rxi_calltrace(RX_CALL_ARRIVAL, call);
4454 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4455 queue_Append(&rx_incomingCallQueue, call);
4458 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4460 /* If hot threads are enabled, and both newcallp and sq->socketp
4461 * are non-null, then this thread will process the call, and the
4462 * idle server thread will start listening on this threads socket.
4465 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4468 *sq->socketp = socket;
4469 clock_GetTime(&call->startTime);
4470 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4474 if (call->flags & RX_CALL_WAIT_PROC) {
4475 /* Conservative: I don't think this should happen */
4476 call->flags &= ~RX_CALL_WAIT_PROC;
4477 if (queue_IsOnQueue(call)) {
4480 MUTEX_ENTER(&rx_waiting_mutex);
4482 MUTEX_EXIT(&rx_waiting_mutex);
4485 call->state = RX_STATE_ACTIVE;
4486 call->mode = RX_MODE_RECEIVING;
4487 #ifdef RX_KERNEL_TRACE
4489 int glockOwner = ISAFS_GLOCK();
4492 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4493 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4499 if (call->flags & RX_CALL_CLEARED) {
4500 /* send an ack now to start the packet flow up again */
4501 call->flags &= ~RX_CALL_CLEARED;
4502 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4504 #ifdef RX_ENABLE_LOCKS
4507 service->nRequestsRunning++;
4508 MUTEX_ENTER(&rx_quota_mutex);
4509 if (service->nRequestsRunning <= service->minProcs)
4512 MUTEX_EXIT(&rx_quota_mutex);
4516 MUTEX_EXIT(&rx_serverPool_lock);
4519 /* Delay the sending of an acknowledge event for a short while, while
4520 * a new call is being prepared (in the case of a client) or a reply
4521 * is being prepared (in the case of a server). Rather than sending
4522 * an ack packet, an ACKALL packet is sent. */
4524 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4526 #ifdef RX_ENABLE_LOCKS
4528 MUTEX_ENTER(&call->lock);
4529 call->delayedAckEvent = NULL;
4530 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4532 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4533 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4535 MUTEX_EXIT(&call->lock);
4536 #else /* RX_ENABLE_LOCKS */
4538 call->delayedAckEvent = NULL;
4539 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4540 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4541 #endif /* RX_ENABLE_LOCKS */
4545 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4547 struct rx_call *call = arg1;
4548 #ifdef RX_ENABLE_LOCKS
4550 MUTEX_ENTER(&call->lock);
4551 if (event == call->delayedAckEvent)
4552 call->delayedAckEvent = NULL;
4553 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4555 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4557 MUTEX_EXIT(&call->lock);
4558 #else /* RX_ENABLE_LOCKS */
4560 call->delayedAckEvent = NULL;
4561 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4562 #endif /* RX_ENABLE_LOCKS */
4566 #ifdef RX_ENABLE_LOCKS
4567 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4568 * clearing them out.
4571 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4573 struct rx_packet *p, *tp;
4576 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4577 p->flags |= RX_PKTFLAG_ACKED;
4581 call->flags |= RX_CALL_TQ_CLEARME;
4582 call->flags |= RX_CALL_TQ_SOME_ACKED;
4585 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4586 call->tfirst = call->tnext;
4587 call->nSoftAcked = 0;
4589 if (call->flags & RX_CALL_FAST_RECOVER) {
4590 call->flags &= ~RX_CALL_FAST_RECOVER;
4591 call->cwind = call->nextCwind;
4592 call->nextCwind = 0;
4595 CV_SIGNAL(&call->cv_twind);
4597 #endif /* RX_ENABLE_LOCKS */
4599 /* Clear out the transmit queue for the current call (all packets have
4600 * been received by peer) */
4602 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4604 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4605 struct rx_packet *p, *tp;
4607 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4609 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4610 p->flags |= RX_PKTFLAG_ACKED;
4614 call->flags |= RX_CALL_TQ_CLEARME;
4615 call->flags |= RX_CALL_TQ_SOME_ACKED;
4618 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4619 #ifdef RXDEBUG_PACKET
4621 #endif /* RXDEBUG_PACKET */
4622 rxi_FreePackets(0, &call->tq);
4623 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
4624 #ifdef RX_ENABLE_LOCKS
4625 CV_BROADCAST(&call->cv_tq);
4626 #else /* RX_ENABLE_LOCKS */
4627 osi_rxWakeup(&call->tq);
4628 #endif /* RX_ENABLE_LOCKS */
4630 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4631 call->flags &= ~RX_CALL_TQ_CLEARME;
4633 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4635 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4636 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4637 call->nSoftAcked = 0;
4639 if (call->flags & RX_CALL_FAST_RECOVER) {
4640 call->flags &= ~RX_CALL_FAST_RECOVER;
4641 call->cwind = call->nextCwind;
4643 #ifdef RX_ENABLE_LOCKS
4644 CV_SIGNAL(&call->cv_twind);
4646 osi_rxWakeup(&call->twind);
4651 rxi_ClearReceiveQueue(struct rx_call *call)
4653 if (queue_IsNotEmpty(&call->rq)) {
4656 count = rxi_FreePackets(0, &call->rq);
4657 rx_packetReclaims += count;
4658 #ifdef RXDEBUG_PACKET
4660 if ( call->rqc != 0 )
4661 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0", call, call->rqc));
4663 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4665 if (call->state == RX_STATE_PRECALL) {
4666 call->flags |= RX_CALL_CLEARED;
4670 /* Send an abort packet for the specified call */
4672 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
4673 int istack, int force)
4676 struct clock when, now;
4681 /* Clients should never delay abort messages */
4682 if (rx_IsClientConn(call->conn))
4685 if (call->abortCode != call->error) {
4686 call->abortCode = call->error;
4687 call->abortCount = 0;
4690 if (force || rxi_callAbortThreshhold == 0
4691 || call->abortCount < rxi_callAbortThreshhold) {
4692 if (call->delayedAbortEvent) {
4693 rxevent_Cancel(call->delayedAbortEvent, call,
4694 RX_CALL_REFCOUNT_ABORT);
4696 error = htonl(call->error);
4699 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4700 (char *)&error, sizeof(error), istack);
4701 } else if (!call->delayedAbortEvent) {
4702 clock_GetTime(&now);
4704 clock_Addmsec(&when, rxi_callAbortDelay);
4705 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4706 call->delayedAbortEvent =
4707 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4712 /* Send an abort packet for the specified connection. Packet is an
4713 * optional pointer to a packet that can be used to send the abort.
4714 * Once the number of abort messages reaches the threshhold, an
4715 * event is scheduled to send the abort. Setting the force flag
4716 * overrides sending delayed abort messages.
4718 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4719 * to send the abort packet.
4722 rxi_SendConnectionAbort(struct rx_connection *conn,
4723 struct rx_packet *packet, int istack, int force)
4726 struct clock when, now;
4731 /* Clients should never delay abort messages */
4732 if (rx_IsClientConn(conn))
4735 if (force || rxi_connAbortThreshhold == 0
4736 || conn->abortCount < rxi_connAbortThreshhold) {
4737 if (conn->delayedAbortEvent) {
4738 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4740 error = htonl(conn->error);
4742 MUTEX_EXIT(&conn->conn_data_lock);
4744 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4745 RX_PACKET_TYPE_ABORT, (char *)&error,
4746 sizeof(error), istack);
4747 MUTEX_ENTER(&conn->conn_data_lock);
4748 } else if (!conn->delayedAbortEvent) {
4749 clock_GetTime(&now);
4751 clock_Addmsec(&when, rxi_connAbortDelay);
4752 conn->delayedAbortEvent =
4753 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4758 /* Associate an error all of the calls owned by a connection. Called
4759 * with error non-zero. This is only for really fatal things, like
4760 * bad authentication responses. The connection itself is set in
4761 * error at this point, so that future packets received will be
4764 rxi_ConnectionError(struct rx_connection *conn,
4770 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d", conn, error));
4772 MUTEX_ENTER(&conn->conn_data_lock);
4773 if (conn->challengeEvent)
4774 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4775 if (conn->natKeepAliveEvent)
4776 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
4777 if (conn->checkReachEvent) {
4778 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4779 conn->checkReachEvent = 0;
4780 conn->flags &= ~RX_CONN_ATTACHWAIT;
4783 MUTEX_EXIT(&conn->conn_data_lock);
4784 for (i = 0; i < RX_MAXCALLS; i++) {
4785 struct rx_call *call = conn->call[i];
4787 MUTEX_ENTER(&call->lock);
4788 rxi_CallError(call, error);
4789 MUTEX_EXIT(&call->lock);
4792 conn->error = error;
4793 if (rx_stats_active)
4794 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4799 rxi_CallError(struct rx_call *call, afs_int32 error)
4802 osirx_AssertMine(&call->lock, "rxi_CallError");
4804 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d", call, error, call->error));
4806 error = call->error;
4808 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4809 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4810 rxi_ResetCall(call, 0);
4813 rxi_ResetCall(call, 0);
4815 call->error = error;
4816 call->mode = RX_MODE_ERROR;
4819 /* Reset various fields in a call structure, and wakeup waiting
4820 * processes. Some fields aren't changed: state & mode are not
4821 * touched (these must be set by the caller), and bufptr, nLeft, and
4822 * nFree are not reset, since these fields are manipulated by
4823 * unprotected macros, and may only be reset by non-interrupting code.
4826 /* this code requires that call->conn be set properly as a pre-condition. */
4827 #endif /* ADAPT_WINDOW */
4830 rxi_ResetCall(struct rx_call *call, int newcall)
4833 struct rx_peer *peer;
4834 struct rx_packet *packet;
4836 osirx_AssertMine(&call->lock, "rxi_ResetCall");
4838 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
4840 /* Notify anyone who is waiting for asynchronous packet arrival */
4841 if (call->arrivalProc) {
4842 (*call->arrivalProc) (call, call->arrivalProcHandle,
4843 call->arrivalProcArg);
4844 call->arrivalProc = (void (*)())0;
4847 if (call->delayedAbortEvent) {
4848 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4849 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4851 rxi_SendCallAbort(call, packet, 0, 1);
4852 rxi_FreePacket(packet);
4857 * Update the peer with the congestion information in this call
4858 * so other calls on this connection can pick up where this call
4859 * left off. If the congestion sequence numbers don't match then
4860 * another call experienced a retransmission.
4862 peer = call->conn->peer;
4863 MUTEX_ENTER(&peer->peer_lock);
4865 if (call->congestSeq == peer->congestSeq) {
4866 peer->cwind = MAX(peer->cwind, call->cwind);
4867 peer->MTU = MAX(peer->MTU, call->MTU);
4868 peer->nDgramPackets =
4869 MAX(peer->nDgramPackets, call->nDgramPackets);
4872 call->abortCode = 0;
4873 call->abortCount = 0;
4875 if (peer->maxDgramPackets > 1) {
4876 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4878 call->MTU = peer->MTU;
4880 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4881 call->ssthresh = rx_maxSendWindow;
4882 call->nDgramPackets = peer->nDgramPackets;
4883 call->congestSeq = peer->congestSeq;
4884 MUTEX_EXIT(&peer->peer_lock);
4886 flags = call->flags;
4887 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4888 rxi_WaitforTQBusy(call);
4889 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4891 rxi_ClearTransmitQueue(call, 1);
4892 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4893 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4897 rxi_ClearReceiveQueue(call);
4898 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
4900 if (call->currentPacket) {
4901 #ifdef RX_TRACK_PACKETS
4902 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
4903 call->currentPacket->flags |= RX_PKTFLAG_IOVQ;
4905 queue_Prepend(&call->iovq, call->currentPacket);
4906 #ifdef RXDEBUG_PACKET
4908 #endif /* RXDEBUG_PACKET */
4909 call->currentPacket = (struct rx_packet *)0;
4911 call->curlen = call->nLeft = call->nFree = 0;
4913 #ifdef RXDEBUG_PACKET
4916 rxi_FreePackets(0, &call->iovq);
4919 call->twind = call->conn->twind[call->channel];
4920 call->rwind = call->conn->rwind[call->channel];
4921 call->nSoftAcked = 0;
4922 call->nextCwind = 0;
4925 call->nCwindAcks = 0;
4926 call->nSoftAcks = 0;
4927 call->nHardAcks = 0;
4929 call->tfirst = call->rnext = call->tnext = 1;
4931 call->lastAcked = 0;
4932 call->localStatus = call->remoteStatus = 0;
4934 if (flags & RX_CALL_READER_WAIT) {
4935 #ifdef RX_ENABLE_LOCKS
4936 CV_BROADCAST(&call->cv_rq);
4938 osi_rxWakeup(&call->rq);
4941 if (flags & RX_CALL_WAIT_PACKETS) {
4942 MUTEX_ENTER(&rx_freePktQ_lock);
4943 rxi_PacketsUnWait(); /* XXX */
4944 MUTEX_EXIT(&rx_freePktQ_lock);
4946 #ifdef RX_ENABLE_LOCKS
4947 CV_SIGNAL(&call->cv_twind);
4949 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4950 osi_rxWakeup(&call->twind);
4953 #ifdef RX_ENABLE_LOCKS
4954 /* The following ensures that we don't mess with any queue while some
4955 * other thread might also be doing so. The call_queue_lock field is
4956 * is only modified under the call lock. If the call is in the process
4957 * of being removed from a queue, the call is not locked until the
4958 * the queue lock is dropped and only then is the call_queue_lock field
4959 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4960 * Note that any other routine which removes a call from a queue has to
4961 * obtain the queue lock before examing the queue and removing the call.
4963 if (call->call_queue_lock) {
4964 MUTEX_ENTER(call->call_queue_lock);
4965 if (queue_IsOnQueue(call)) {
4967 if (flags & RX_CALL_WAIT_PROC) {
4969 MUTEX_ENTER(&rx_waiting_mutex);
4971 MUTEX_EXIT(&rx_waiting_mutex);
4974 MUTEX_EXIT(call->call_queue_lock);
4975 CLEAR_CALL_QUEUE_LOCK(call);
4977 #else /* RX_ENABLE_LOCKS */
4978 if (queue_IsOnQueue(call)) {
4980 if (flags & RX_CALL_WAIT_PROC)
4983 #endif /* RX_ENABLE_LOCKS */
4985 rxi_KeepAliveOff(call);
4986 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4989 /* Send an acknowledge for the indicated packet (seq,serial) of the
4990 * indicated call, for the indicated reason (reason). This
4991 * acknowledge will specifically acknowledge receiving the packet, and
4992 * will also specify which other packets for this call have been
4993 * received. This routine returns the packet that was used to the
4994 * caller. The caller is responsible for freeing it or re-using it.
4995 * This acknowledgement also returns the highest sequence number
4996 * actually read out by the higher level to the sender; the sender
4997 * promises to keep around packets that have not been read by the
4998 * higher level yet (unless, of course, the sender decides to abort
4999 * the call altogether). Any of p, seq, serial, pflags, or reason may
5000 * be set to zero without ill effect. That is, if they are zero, they
5001 * will not convey any information.
5002 * NOW there is a trailer field, after the ack where it will safely be
5003 * ignored by mundanes, which indicates the maximum size packet this
5004 * host can swallow. */
5006 struct rx_packet *optionalPacket; use to send ack (or null)
5007 int seq; Sequence number of the packet we are acking
5008 int serial; Serial number of the packet
5009 int pflags; Flags field from packet header
5010 int reason; Reason an acknowledge was prompted
5014 rxi_SendAck(struct rx_call *call,
5015 struct rx_packet *optionalPacket, int serial, int reason,
5018 struct rx_ackPacket *ap;
5019 struct rx_packet *rqp;
5020 struct rx_packet *nxp; /* For queue_Scan */
5021 struct rx_packet *p;
5024 afs_uint32 padbytes = 0;
5025 #ifdef RX_ENABLE_TSFPQ
5026 struct rx_ts_info_t * rx_ts_info;
5030 * Open the receive window once a thread starts reading packets
5032 if (call->rnext > 1) {
5033 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5036 /* Don't attempt to grow MTU if this is a critical ping */
5037 if (reason == RX_ACK_MTU) {
5038 /* keep track of per-call attempts, if we're over max, do in small
5039 * otherwise in larger? set a size to increment by, decrease
5042 if (call->conn->peer->maxPacketSize &&
5043 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5045 padbytes = call->conn->peer->maxPacketSize+16;
5047 padbytes = call->conn->peer->maxMTU + 128;
5049 /* do always try a minimum size ping */
5050 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5052 /* subtract the ack payload */
5053 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5054 reason = RX_ACK_PING;
5057 call->nHardAcks = 0;
5058 call->nSoftAcks = 0;
5059 if (call->rnext > call->lastAcked)
5060 call->lastAcked = call->rnext;
5064 rx_computelen(p, p->length); /* reset length, you never know */
5065 } /* where that's been... */
5066 #ifdef RX_ENABLE_TSFPQ
5068 RX_TS_INFO_GET(rx_ts_info);
5069 if ((p = rx_ts_info->local_special_packet)) {
5070 rx_computelen(p, p->length);
5071 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5072 rx_ts_info->local_special_packet = p;
5073 } else { /* We won't send the ack, but don't panic. */
5074 return optionalPacket;
5078 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5079 /* We won't send the ack, but don't panic. */
5080 return optionalPacket;
5085 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5088 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5089 #ifndef RX_ENABLE_TSFPQ
5090 if (!optionalPacket)
5093 return optionalPacket;
5095 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5096 if (rx_Contiguous(p) < templ) {
5097 #ifndef RX_ENABLE_TSFPQ
5098 if (!optionalPacket)
5101 return optionalPacket;
5106 /* MTUXXX failing to send an ack is very serious. We should */
5107 /* try as hard as possible to send even a partial ack; it's */
5108 /* better than nothing. */
5109 ap = (struct rx_ackPacket *)rx_DataOf(p);
5110 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5111 ap->reason = reason;
5113 /* The skew computation used to be bogus, I think it's better now. */
5114 /* We should start paying attention to skew. XXX */
5115 ap->serial = htonl(serial);
5116 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5118 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
5119 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5121 /* No fear of running out of ack packet here because there can only be at most
5122 * one window full of unacknowledged packets. The window size must be constrained
5123 * to be less than the maximum ack size, of course. Also, an ack should always
5124 * fit into a single packet -- it should not ever be fragmented. */
5125 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5126 if (!rqp || !call->rq.next
5127 || (rqp->header.seq > (call->rnext + call->rwind))) {
5128 #ifndef RX_ENABLE_TSFPQ
5129 if (!optionalPacket)
5132 rxi_CallError(call, RX_CALL_DEAD);
5133 return optionalPacket;
5136 while (rqp->header.seq > call->rnext + offset)
5137 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5138 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5140 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5141 #ifndef RX_ENABLE_TSFPQ
5142 if (!optionalPacket)
5145 rxi_CallError(call, RX_CALL_DEAD);
5146 return optionalPacket;
5151 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5153 /* these are new for AFS 3.3 */
5154 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5155 templ = htonl(templ);
5156 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5157 templ = htonl(call->conn->peer->ifMTU);
5158 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5159 sizeof(afs_int32), &templ);
5161 /* new for AFS 3.4 */
5162 templ = htonl(call->rwind);
5163 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5164 sizeof(afs_int32), &templ);
5166 /* new for AFS 3.5 */
5167 templ = htonl(call->conn->peer->ifDgramPackets);
5168 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5169 sizeof(afs_int32), &templ);
5171 p->header.serviceId = call->conn->serviceId;
5172 p->header.cid = (call->conn->cid | call->channel);
5173 p->header.callNumber = *call->callNumber;
5175 p->header.securityIndex = call->conn->securityIndex;
5176 p->header.epoch = call->conn->epoch;
5177 p->header.type = RX_PACKET_TYPE_ACK;
5178 p->header.flags = RX_SLOW_START_OK;
5179 if (reason == RX_ACK_PING) {
5180 p->header.flags |= RX_REQUEST_ACK;
5182 clock_GetTime(&call->pingRequestTime);
5185 p->length = padbytes +
5186 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5189 /* not fast but we can potentially use this if truncated
5190 * fragments are delivered to figure out the mtu.
5192 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5193 sizeof(afs_int32), sizeof(afs_int32),
5197 if (call->conn->type == RX_CLIENT_CONNECTION)
5198 p->header.flags |= RX_CLIENT_INITIATED;
5202 if (rxdebug_active) {
5206 len = _snprintf(msg, sizeof(msg),
5207 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5208 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5209 ntohl(ap->serial), ntohl(ap->previousPacket),
5210 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5211 ap->nAcks, ntohs(ap->bufferSpace) );
5215 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5216 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5220 OutputDebugString(msg);
5222 #else /* AFS_NT40_ENV */
5224 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5225 ap->reason, ntohl(ap->previousPacket),
5226 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5228 for (offset = 0; offset < ap->nAcks; offset++)
5229 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5234 #endif /* AFS_NT40_ENV */
5237 int i, nbytes = p->length;
5239 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5240 if (nbytes <= p->wirevec[i].iov_len) {
5243 savelen = p->wirevec[i].iov_len;
5245 p->wirevec[i].iov_len = nbytes;
5247 rxi_Send(call, p, istack);
5248 p->wirevec[i].iov_len = savelen;
5252 nbytes -= p->wirevec[i].iov_len;
5255 if (rx_stats_active)
5256 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
5257 #ifndef RX_ENABLE_TSFPQ
5258 if (!optionalPacket)
5261 return optionalPacket; /* Return packet for re-use by caller */
5264 /* Send all of the packets in the list in single datagram */
5266 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
5267 int istack, int moreFlag, struct clock *now,
5268 struct clock *retryTime, int resending)
5273 struct rx_connection *conn = call->conn;
5274 struct rx_peer *peer = conn->peer;
5276 MUTEX_ENTER(&peer->peer_lock);
5279 peer->reSends += len;
5280 if (rx_stats_active)
5281 rx_MutexAdd(rx_stats.dataPacketsSent, len, rx_stats_mutex);
5282 MUTEX_EXIT(&peer->peer_lock);
5284 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
5288 /* Set the packet flags and schedule the resend events */
5289 /* Only request an ack for the last packet in the list */
5290 for (i = 0; i < len; i++) {
5291 list[i]->retryTime = *retryTime;
5292 if (list[i]->header.serial) {
5293 /* Exponentially backoff retry times */
5294 if (list[i]->backoff < MAXBACKOFF) {
5295 /* so it can't stay == 0 */
5296 list[i]->backoff = (list[i]->backoff << 1) + 1;
5299 clock_Addmsec(&(list[i]->retryTime),
5300 ((afs_uint32) list[i]->backoff) << 8);
5303 /* Wait a little extra for the ack on the last packet */
5304 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
5305 clock_Addmsec(&(list[i]->retryTime), 400);
5308 /* Record the time sent */
5309 list[i]->timeSent = *now;
5311 /* Ask for an ack on retransmitted packets, on every other packet
5312 * if the peer doesn't support slow start. Ask for an ack on every
5313 * packet until the congestion window reaches the ack rate. */
5314 if (list[i]->header.serial) {
5316 if (rx_stats_active)
5317 rx_MutexIncrement(rx_stats.dataPacketsReSent, rx_stats_mutex);
5319 /* improved RTO calculation- not Karn */
5320 list[i]->firstSent = *now;
5321 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5322 || (!(call->flags & RX_CALL_SLOW_START_OK)
5323 && (list[i]->header.seq & 1)))) {
5328 /* Tag this packet as not being the last in this group,
5329 * for the receiver's benefit */
5330 if (i < len - 1 || moreFlag) {
5331 list[i]->header.flags |= RX_MORE_PACKETS;
5334 /* Install the new retransmit time for the packet, and
5335 * record the time sent */
5336 list[i]->timeSent = *now;
5340 list[len - 1]->header.flags |= RX_REQUEST_ACK;
5343 /* Since we're about to send a data packet to the peer, it's
5344 * safe to nuke any scheduled end-of-packets ack */
5345 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5347 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5348 MUTEX_EXIT(&call->lock);
5350 rxi_SendPacketList(call, conn, list, len, istack);
5352 rxi_SendPacket(call, conn, list[0], istack);
5354 MUTEX_ENTER(&call->lock);
5355 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5357 /* Update last send time for this call (for keep-alive
5358 * processing), and for the connection (so that we can discover
5359 * idle connections) */
5360 conn->lastSendTime = call->lastSendTime = clock_Sec();
5361 /* Let a set of retransmits trigger an idle timeout */
5363 call->lastSendData = call->lastSendTime;
5366 /* When sending packets we need to follow these rules:
5367 * 1. Never send more than maxDgramPackets in a jumbogram.
5368 * 2. Never send a packet with more than two iovecs in a jumbogram.
5369 * 3. Never send a retransmitted packet in a jumbogram.
5370 * 4. Never send more than cwind/4 packets in a jumbogram
5371 * We always keep the last list we should have sent so we
5372 * can set the RX_MORE_PACKETS flags correctly.
5375 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5376 int istack, struct clock *now, struct clock *retryTime,
5379 int i, cnt, lastCnt = 0;
5380 struct rx_packet **listP, **lastP = 0;
5381 struct rx_peer *peer = call->conn->peer;
5382 int morePackets = 0;
5384 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
5385 /* Does the current packet force us to flush the current list? */
5387 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5388 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5390 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
5392 /* If the call enters an error state stop sending, or if
5393 * we entered congestion recovery mode, stop sending */
5394 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5402 /* Add the current packet to the list if it hasn't been acked.
5403 * Otherwise adjust the list pointer to skip the current packet. */
5404 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5406 /* Do we need to flush the list? */
5407 if (cnt >= (int)peer->maxDgramPackets
5408 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5409 || list[i]->header.serial
5410 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5412 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5413 retryTime, resending);
5414 /* If the call enters an error state stop sending, or if
5415 * we entered congestion recovery mode, stop sending */
5417 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5422 listP = &list[i + 1];
5427 osi_Panic("rxi_SendList error");
5429 listP = &list[i + 1];
5433 /* Send the whole list when the call is in receive mode, when
5434 * the call is in eof mode, when we are in fast recovery mode,
5435 * and when we have the last packet */
5436 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5437 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5438 || (call->flags & RX_CALL_FAST_RECOVER)) {
5439 /* Check for the case where the current list contains
5440 * an acked packet. Since we always send retransmissions
5441 * in a separate packet, we only need to check the first
5442 * packet in the list */
5443 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5447 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5448 retryTime, resending);
5449 /* If the call enters an error state stop sending, or if
5450 * we entered congestion recovery mode, stop sending */
5451 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5455 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5458 } else if (lastCnt > 0) {
5459 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5464 #ifdef RX_ENABLE_LOCKS
5465 /* Call rxi_Start, below, but with the call lock held. */
5467 rxi_StartUnlocked(struct rxevent *event,
5468 void *arg0, void *arg1, int istack)
5470 struct rx_call *call = arg0;
5472 MUTEX_ENTER(&call->lock);
5473 rxi_Start(event, call, arg1, istack);
5474 MUTEX_EXIT(&call->lock);
5476 #endif /* RX_ENABLE_LOCKS */
5478 /* This routine is called when new packets are readied for
5479 * transmission and when retransmission may be necessary, or when the
5480 * transmission window or burst count are favourable. This should be
5481 * better optimized for new packets, the usual case, now that we've
5482 * got rid of queues of send packets. XXXXXXXXXXX */
5484 rxi_Start(struct rxevent *event,
5485 void *arg0, void *arg1, int istack)
5487 struct rx_call *call = arg0;
5489 struct rx_packet *p;
5490 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5491 struct rx_peer *peer = call->conn->peer;
5492 struct clock now, usenow, retryTime;
5496 struct rx_packet **xmitList;
5499 /* If rxi_Start is being called as a result of a resend event,
5500 * then make sure that the event pointer is removed from the call
5501 * structure, since there is no longer a per-call retransmission
5503 if (event && event == call->resendEvent) {
5504 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5505 call->resendEvent = NULL;
5507 if (queue_IsEmpty(&call->tq)) {
5511 /* Timeouts trigger congestion recovery */
5512 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5513 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5514 /* someone else is waiting to start recovery */
5517 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5518 rxi_WaitforTQBusy(call);
5519 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5520 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5521 call->flags |= RX_CALL_FAST_RECOVER;
5522 if (peer->maxDgramPackets > 1) {
5523 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5525 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5527 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5528 call->nDgramPackets = 1;
5530 call->nextCwind = 1;
5533 MUTEX_ENTER(&peer->peer_lock);
5534 peer->MTU = call->MTU;
5535 peer->cwind = call->cwind;
5536 peer->nDgramPackets = 1;
5538 call->congestSeq = peer->congestSeq;
5539 MUTEX_EXIT(&peer->peer_lock);
5540 /* Clear retry times on packets. Otherwise, it's possible for
5541 * some packets in the queue to force resends at rates faster
5542 * than recovery rates.
5544 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5545 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5546 clock_Zero(&p->retryTime);
5551 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5552 if (rx_stats_active)
5553 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5558 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5559 /* Get clock to compute the re-transmit time for any packets
5560 * in this burst. Note, if we back off, it's reasonable to
5561 * back off all of the packets in the same manner, even if
5562 * some of them have been retransmitted more times than more
5564 * Do a dance to avoid blocking after setting now. */
5565 MUTEX_ENTER(&peer->peer_lock);
5566 retryTime = peer->timeout;
5567 MUTEX_EXIT(&peer->peer_lock);
5568 clock_GetTime(&now);
5569 clock_Add(&retryTime, &now);
5571 /* Send (or resend) any packets that need it, subject to
5572 * window restrictions and congestion burst control
5573 * restrictions. Ask for an ack on the last packet sent in
5574 * this burst. For now, we're relying upon the window being
5575 * considerably bigger than the largest number of packets that
5576 * are typically sent at once by one initial call to
5577 * rxi_Start. This is probably bogus (perhaps we should ask
5578 * for an ack when we're half way through the current
5579 * window?). Also, for non file transfer applications, this
5580 * may end up asking for an ack for every packet. Bogus. XXXX
5583 * But check whether we're here recursively, and let the other guy
5586 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5587 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5588 call->flags |= RX_CALL_TQ_BUSY;
5590 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5592 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5593 call->flags &= ~RX_CALL_NEED_START;
5594 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5596 maxXmitPackets = MIN(call->twind, call->cwind);
5597 xmitList = (struct rx_packet **)
5598 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
5599 /* XXXX else we must drop any mtx we hold */
5600 afs_osi_Alloc_NoSleep(maxXmitPackets * sizeof(struct rx_packet *));
5602 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5604 if (xmitList == NULL)
5605 osi_Panic("rxi_Start, failed to allocate xmit list");
5606 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5607 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5608 /* We shouldn't be sending packets if a thread is waiting
5609 * to initiate congestion recovery */
5610 dpf(("call %d waiting to initiate fast recovery\n",
5611 *(call->callNumber)));
5615 && (call->flags & RX_CALL_FAST_RECOVER)) {
5616 /* Only send one packet during fast recovery */
5617 dpf(("call %d restricted to one packet per send during fast recovery\n",
5618 *(call->callNumber)));
5621 #ifdef RX_TRACK_PACKETS
5622 if ((p->flags & RX_PKTFLAG_FREE)
5623 || (!queue_IsEnd(&call->tq, nxp)
5624 && (nxp->flags & RX_PKTFLAG_FREE))
5625 || (p == (struct rx_packet *)&rx_freePacketQueue)
5626 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5627 osi_Panic("rxi_Start: xmit queue clobbered");
5630 if (p->flags & RX_PKTFLAG_ACKED) {
5631 /* Since we may block, don't trust this */
5632 usenow.sec = usenow.usec = 0;
5633 if (rx_stats_active)
5634 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5635 continue; /* Ignore this packet if it has been acknowledged */
5638 /* Turn off all flags except these ones, which are the same
5639 * on each transmission */
5640 p->header.flags &= RX_PRESET_FLAGS;
5642 if (p->header.seq >=
5643 call->tfirst + MIN((int)call->twind,
5644 (int)(call->nSoftAcked +
5646 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5647 /* Note: if we're waiting for more window space, we can
5648 * still send retransmits; hence we don't return here, but
5649 * break out to schedule a retransmit event */
5650 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
5651 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
5656 /* Transmit the packet if it needs to be sent. */
5657 if (!clock_Lt(&now, &p->retryTime)) {
5658 if (nXmitPackets == maxXmitPackets) {
5659 rxi_SendXmitList(call, xmitList, nXmitPackets,
5660 istack, &now, &retryTime,
5662 osi_Free(xmitList, maxXmitPackets *
5663 sizeof(struct rx_packet *));
5666 dpf(("call %d xmit packet %"AFS_PTR_FMT" now %u.%06u retryTime %u.%06u nextRetry %u.%06u\n",
5667 *(call->callNumber), p,
5669 p->retryTime.sec, p->retryTime.usec,
5670 retryTime.sec, retryTime.usec));
5671 xmitList[nXmitPackets++] = p;
5675 /* xmitList now hold pointers to all of the packets that are
5676 * ready to send. Now we loop to send the packets */
5677 if (nXmitPackets > 0) {
5678 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5679 &now, &retryTime, resending);
5682 maxXmitPackets * sizeof(struct rx_packet *));
5684 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5686 * TQ references no longer protected by this flag; they must remain
5687 * protected by the global lock.
5689 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5690 call->flags &= ~RX_CALL_TQ_BUSY;
5691 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5692 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
5693 call, call->tqWaiters, call->flags));
5694 #ifdef RX_ENABLE_LOCKS
5695 osirx_AssertMine(&call->lock, "rxi_Start start");
5696 CV_BROADCAST(&call->cv_tq);
5697 #else /* RX_ENABLE_LOCKS */
5698 osi_rxWakeup(&call->tq);
5699 #endif /* RX_ENABLE_LOCKS */
5704 /* We went into the error state while sending packets. Now is
5705 * the time to reset the call. This will also inform the using
5706 * process that the call is in an error state.
5708 if (rx_stats_active)
5709 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5710 call->flags &= ~RX_CALL_TQ_BUSY;
5711 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5712 dpf(("call error %d while xmit %p has %d waiters and flags %d\n",
5713 call->error, call, call->tqWaiters, call->flags));
5714 #ifdef RX_ENABLE_LOCKS
5715 osirx_AssertMine(&call->lock, "rxi_Start middle");
5716 CV_BROADCAST(&call->cv_tq);
5717 #else /* RX_ENABLE_LOCKS */
5718 osi_rxWakeup(&call->tq);
5719 #endif /* RX_ENABLE_LOCKS */
5721 rxi_CallError(call, call->error);
5724 #ifdef RX_ENABLE_LOCKS
5725 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5727 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5728 /* Some packets have received acks. If they all have, we can clear
5729 * the transmit queue.
5732 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5733 if (p->header.seq < call->tfirst
5734 && (p->flags & RX_PKTFLAG_ACKED)) {
5736 #ifdef RX_TRACK_PACKETS
5737 p->flags &= ~RX_PKTFLAG_TQ;
5739 #ifdef RXDEBUG_PACKET
5747 call->flags |= RX_CALL_TQ_CLEARME;
5749 #endif /* RX_ENABLE_LOCKS */
5750 /* Don't bother doing retransmits if the TQ is cleared. */
5751 if (call->flags & RX_CALL_TQ_CLEARME) {
5752 rxi_ClearTransmitQueue(call, 1);
5754 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5757 /* Always post a resend event, if there is anything in the
5758 * queue, and resend is possible. There should be at least
5759 * one unacknowledged packet in the queue ... otherwise none
5760 * of these packets should be on the queue in the first place.
5762 if (call->resendEvent) {
5763 /* Cancel the existing event and post a new one */
5764 rxevent_Cancel(call->resendEvent, call,
5765 RX_CALL_REFCOUNT_RESEND);
5768 /* The retry time is the retry time on the first unacknowledged
5769 * packet inside the current window */
5771 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5772 /* Don't set timers for packets outside the window */
5773 if (p->header.seq >= call->tfirst + call->twind) {
5777 if (!(p->flags & RX_PKTFLAG_ACKED)
5778 && !clock_IsZero(&p->retryTime)) {
5780 retryTime = p->retryTime;
5785 /* Post a new event to re-run rxi_Start when retries may be needed */
5786 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5787 #ifdef RX_ENABLE_LOCKS
5788 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5790 rxevent_PostNow2(&retryTime, &usenow,
5792 (void *)call, 0, istack);
5793 #else /* RX_ENABLE_LOCKS */
5795 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5796 (void *)call, 0, istack);
5797 #endif /* RX_ENABLE_LOCKS */
5800 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5801 } while (call->flags & RX_CALL_NEED_START);
5803 * TQ references no longer protected by this flag; they must remain
5804 * protected by the global lock.
5806 call->flags &= ~RX_CALL_TQ_BUSY;
5807 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5808 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
5809 call, call->tqWaiters, call->flags));
5810 #ifdef RX_ENABLE_LOCKS
5811 osirx_AssertMine(&call->lock, "rxi_Start end");
5812 CV_BROADCAST(&call->cv_tq);
5813 #else /* RX_ENABLE_LOCKS */
5814 osi_rxWakeup(&call->tq);
5815 #endif /* RX_ENABLE_LOCKS */
5818 call->flags |= RX_CALL_NEED_START;
5820 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5822 if (call->resendEvent) {
5823 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5828 /* Also adjusts the keep alive parameters for the call, to reflect
5829 * that we have just sent a packet (so keep alives aren't sent
5832 rxi_Send(struct rx_call *call, struct rx_packet *p,
5835 struct rx_connection *conn = call->conn;
5837 /* Stamp each packet with the user supplied status */
5838 p->header.userStatus = call->localStatus;
5840 /* Allow the security object controlling this call's security to
5841 * make any last-minute changes to the packet */
5842 RXS_SendPacket(conn->securityObject, call, p);
5844 /* Since we're about to send SOME sort of packet to the peer, it's
5845 * safe to nuke any scheduled end-of-packets ack */
5846 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5848 /* Actually send the packet, filling in more connection-specific fields */
5849 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5850 MUTEX_EXIT(&call->lock);
5851 rxi_SendPacket(call, conn, p, istack);
5852 MUTEX_ENTER(&call->lock);
5853 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5855 /* Update last send time for this call (for keep-alive
5856 * processing), and for the connection (so that we can discover
5857 * idle connections) */
5858 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5859 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
5860 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
5862 conn->lastSendTime = call->lastSendTime = clock_Sec();
5863 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
5864 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5865 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
5866 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
5867 RX_ACK_PING_RESPONSE)))
5868 call->lastSendData = call->lastSendTime;
5872 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5873 * that things are fine. Also called periodically to guarantee that nothing
5874 * falls through the cracks (e.g. (error + dally) connections have keepalive
5875 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5877 * haveCTLock Set if calling from rxi_ReapConnections
5879 #ifdef RX_ENABLE_LOCKS
5881 rxi_CheckCall(struct rx_call *call, int haveCTLock)
5882 #else /* RX_ENABLE_LOCKS */
5884 rxi_CheckCall(struct rx_call *call)
5885 #endif /* RX_ENABLE_LOCKS */
5887 struct rx_connection *conn = call->conn;
5889 afs_uint32 deadTime;
5893 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5894 if (call->flags & RX_CALL_TQ_BUSY) {
5895 /* Call is active and will be reset by rxi_Start if it's
5896 * in an error state.
5901 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5903 (((afs_uint32) conn->secondsUntilDead << 10) +
5904 ((afs_uint32) conn->peer->rtt >> 3) +
5905 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5907 /* These are computed to the second (+- 1 second). But that's
5908 * good enough for these values, which should be a significant
5909 * number of seconds. */
5910 if (now > (call->lastReceiveTime + deadTime)) {
5911 if (call->state == RX_STATE_ACTIVE) {
5913 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5915 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5916 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5917 ip_stack_t *ipst = ns->netstack_ip;
5919 ire = ire_cache_lookup(conn->peer->host
5920 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5922 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5924 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5931 if (ire && ire->ire_max_frag > 0)
5932 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
5934 #if defined(GLOBAL_NETSTACKID)
5938 #endif /* ADAPT_PMTU */
5939 cerror = RX_CALL_DEAD;
5942 #ifdef RX_ENABLE_LOCKS
5943 /* Cancel pending events */
5944 rxevent_Cancel(call->delayedAckEvent, call,
5945 RX_CALL_REFCOUNT_DELAY);
5946 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5947 rxevent_Cancel(call->keepAliveEvent, call,
5948 RX_CALL_REFCOUNT_ALIVE);
5949 if (call->refCount == 0) {
5950 rxi_FreeCall(call, haveCTLock);
5954 #else /* RX_ENABLE_LOCKS */
5957 #endif /* RX_ENABLE_LOCKS */
5959 /* Non-active calls are destroyed if they are not responding
5960 * to pings; active calls are simply flagged in error, so the
5961 * attached process can die reasonably gracefully. */
5963 /* see if we have a non-activity timeout */
5964 if (call->startWait && conn->idleDeadTime
5965 && ((call->startWait + conn->idleDeadTime) < now) &&
5966 (call->flags & RX_CALL_READER_WAIT)) {
5967 if (call->state == RX_STATE_ACTIVE) {
5968 cerror = RX_CALL_TIMEOUT;
5972 if (call->lastSendData && conn->idleDeadTime && (conn->idleDeadErr != 0)
5973 && ((call->lastSendData + conn->idleDeadTime) < now)) {
5974 if (call->state == RX_STATE_ACTIVE) {
5975 cerror = conn->idleDeadErr;
5979 /* see if we have a hard timeout */
5980 if (conn->hardDeadTime
5981 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5982 if (call->state == RX_STATE_ACTIVE)
5983 rxi_CallError(call, RX_CALL_TIMEOUT);
5988 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT) {
5989 int oldMTU = conn->peer->ifMTU;
5991 /* if we thought we could send more, perhaps things got worse */
5992 if (call->conn->peer->maxPacketSize > conn->lastPacketSize)
5993 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
5994 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
5995 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
5997 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
5999 /* minimum capped in SetPeerMtu */
6000 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6003 conn->lastPacketSize = 0;
6005 /* needed so ResetCall doesn't clobber us. */
6006 call->MTU = conn->peer->ifMTU;
6008 /* if we never succeeded, let the error pass out as-is */
6009 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6010 cerror = conn->msgsizeRetryErr;
6013 rxi_CallError(call, cerror);
6018 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6020 struct rx_connection *conn = arg1;
6021 struct rx_header theader;
6023 struct sockaddr_in taddr;
6026 struct iovec tmpiov[2];
6029 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6032 tp = &tbuffer[sizeof(struct rx_header)];
6033 taddr.sin_family = AF_INET;
6034 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6035 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6036 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6037 taddr.sin_len = sizeof(struct sockaddr_in);
6039 memset(&theader, 0, sizeof(theader));
6040 theader.epoch = htonl(999);
6042 theader.callNumber = 0;
6045 theader.type = RX_PACKET_TYPE_VERSION;
6046 theader.flags = RX_LAST_PACKET;
6047 theader.serviceId = 0;
6049 memcpy(tbuffer, &theader, sizeof(theader));
6050 memcpy(tp, &a, sizeof(a));
6051 tmpiov[0].iov_base = tbuffer;
6052 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6054 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6056 MUTEX_ENTER(&conn->conn_data_lock);
6057 /* Only reschedule ourselves if the connection would not be destroyed */
6058 if (conn->refCount <= 1) {
6059 conn->natKeepAliveEvent = NULL;
6060 MUTEX_EXIT(&conn->conn_data_lock);
6061 rx_DestroyConnection(conn); /* drop the reference for this */
6063 conn->natKeepAliveEvent = NULL;
6064 conn->refCount--; /* drop the reference for this */
6065 rxi_ScheduleNatKeepAliveEvent(conn);
6066 MUTEX_EXIT(&conn->conn_data_lock);
6071 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6073 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6074 struct clock when, now;
6075 clock_GetTime(&now);
6077 when.sec += conn->secondsUntilNatPing;
6078 conn->refCount++; /* hold a reference for this */
6079 conn->natKeepAliveEvent =
6080 rxevent_PostNow(&when, &now, rxi_NatKeepAliveEvent, conn, 0);
6085 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6087 MUTEX_ENTER(&conn->conn_data_lock);
6088 conn->secondsUntilNatPing = seconds;
6090 rxi_ScheduleNatKeepAliveEvent(conn);
6091 MUTEX_EXIT(&conn->conn_data_lock);
6095 rxi_NatKeepAliveOn(struct rx_connection *conn)
6097 MUTEX_ENTER(&conn->conn_data_lock);
6098 rxi_ScheduleNatKeepAliveEvent(conn);
6099 MUTEX_EXIT(&conn->conn_data_lock);
6102 /* When a call is in progress, this routine is called occasionally to
6103 * make sure that some traffic has arrived (or been sent to) the peer.
6104 * If nothing has arrived in a reasonable amount of time, the call is
6105 * declared dead; if nothing has been sent for a while, we send a
6106 * keep-alive packet (if we're actually trying to keep the call alive)
6109 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6111 struct rx_call *call = arg1;
6112 struct rx_connection *conn;
6115 MUTEX_ENTER(&call->lock);
6116 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6117 if (event == call->keepAliveEvent)
6118 call->keepAliveEvent = NULL;
6121 #ifdef RX_ENABLE_LOCKS
6122 if (rxi_CheckCall(call, 0)) {
6123 MUTEX_EXIT(&call->lock);
6126 #else /* RX_ENABLE_LOCKS */
6127 if (rxi_CheckCall(call))
6129 #endif /* RX_ENABLE_LOCKS */
6131 /* Don't try to keep alive dallying calls */
6132 if (call->state == RX_STATE_DALLY) {
6133 MUTEX_EXIT(&call->lock);
6138 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6139 /* Don't try to send keepalives if there is unacknowledged data */
6140 /* the rexmit code should be good enough, this little hack
6141 * doesn't quite work XXX */
6142 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6144 rxi_ScheduleKeepAliveEvent(call);
6145 MUTEX_EXIT(&call->lock);
6148 /* Does what's on the nameplate. */
6150 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy)
6152 struct rx_call *call = arg1;
6153 struct rx_connection *conn;
6155 MUTEX_ENTER(&call->lock);
6156 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6157 if (event == call->growMTUEvent)
6158 call->growMTUEvent = NULL;
6160 #ifdef RX_ENABLE_LOCKS
6161 if (rxi_CheckCall(call, 0)) {
6162 MUTEX_EXIT(&call->lock);
6165 #else /* RX_ENABLE_LOCKS */
6166 if (rxi_CheckCall(call))
6168 #endif /* RX_ENABLE_LOCKS */
6170 /* Don't bother with dallying calls */
6171 if (call->state == RX_STATE_DALLY) {
6172 MUTEX_EXIT(&call->lock);
6179 * keep being scheduled, just don't do anything if we're at peak,
6180 * or we're not set up to be properly handled (idle timeout required)
6182 if ((conn->peer->maxPacketSize != 0) &&
6183 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6184 (conn->idleDeadErr))
6185 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6186 rxi_ScheduleGrowMTUEvent(call, 0);
6187 MUTEX_EXIT(&call->lock);
6191 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6193 if (!call->keepAliveEvent) {
6194 struct clock when, now;
6195 clock_GetTime(&now);
6197 when.sec += call->conn->secondsUntilPing;
6198 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6199 call->keepAliveEvent =
6200 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
6205 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6207 if (!call->growMTUEvent) {
6208 struct clock when, now;
6210 clock_GetTime(&now);
6213 if (call->conn->secondsUntilPing)
6214 secs = (6*call->conn->secondsUntilPing)-1;
6216 if (call->conn->secondsUntilDead)
6217 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6221 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6222 call->growMTUEvent =
6223 rxevent_PostNow(&when, &now, rxi_GrowMTUEvent, call, 0);
6227 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6229 rxi_KeepAliveOn(struct rx_call *call)
6231 /* Pretend last packet received was received now--i.e. if another
6232 * packet isn't received within the keep alive time, then the call
6233 * will die; Initialize last send time to the current time--even
6234 * if a packet hasn't been sent yet. This will guarantee that a
6235 * keep-alive is sent within the ping time */
6236 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6237 rxi_ScheduleKeepAliveEvent(call);
6241 rxi_GrowMTUOn(struct rx_call *call)
6243 struct rx_connection *conn = call->conn;
6244 MUTEX_ENTER(&conn->conn_data_lock);
6245 conn->lastPingSizeSer = conn->lastPingSize = 0;
6246 MUTEX_EXIT(&conn->conn_data_lock);
6247 rxi_ScheduleGrowMTUEvent(call, 1);
6250 /* This routine is called to send connection abort messages
6251 * that have been delayed to throttle looping clients. */
6253 rxi_SendDelayedConnAbort(struct rxevent *event,
6254 void *arg1, void *unused)
6256 struct rx_connection *conn = arg1;
6259 struct rx_packet *packet;
6261 MUTEX_ENTER(&conn->conn_data_lock);
6262 conn->delayedAbortEvent = NULL;
6263 error = htonl(conn->error);
6265 MUTEX_EXIT(&conn->conn_data_lock);
6266 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6269 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6270 RX_PACKET_TYPE_ABORT, (char *)&error,
6272 rxi_FreePacket(packet);
6276 /* This routine is called to send call abort messages
6277 * that have been delayed to throttle looping clients. */
6279 rxi_SendDelayedCallAbort(struct rxevent *event,
6280 void *arg1, void *dummy)
6282 struct rx_call *call = arg1;
6285 struct rx_packet *packet;
6287 MUTEX_ENTER(&call->lock);
6288 call->delayedAbortEvent = NULL;
6289 error = htonl(call->error);
6291 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6294 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6295 (char *)&error, sizeof(error), 0);
6296 rxi_FreePacket(packet);
6298 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6299 MUTEX_EXIT(&call->lock);
6302 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6303 * seconds) to ask the client to authenticate itself. The routine
6304 * issues a challenge to the client, which is obtained from the
6305 * security object associated with the connection */
6307 rxi_ChallengeEvent(struct rxevent *event,
6308 void *arg0, void *arg1, int tries)
6310 struct rx_connection *conn = arg0;
6312 conn->challengeEvent = NULL;
6313 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6314 struct rx_packet *packet;
6315 struct clock when, now;
6318 /* We've failed to authenticate for too long.
6319 * Reset any calls waiting for authentication;
6320 * they are all in RX_STATE_PRECALL.
6324 MUTEX_ENTER(&conn->conn_call_lock);
6325 for (i = 0; i < RX_MAXCALLS; i++) {
6326 struct rx_call *call = conn->call[i];
6328 MUTEX_ENTER(&call->lock);
6329 if (call->state == RX_STATE_PRECALL) {
6330 rxi_CallError(call, RX_CALL_DEAD);
6331 rxi_SendCallAbort(call, NULL, 0, 0);
6333 MUTEX_EXIT(&call->lock);
6336 MUTEX_EXIT(&conn->conn_call_lock);
6340 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6342 /* If there's no packet available, do this later. */
6343 RXS_GetChallenge(conn->securityObject, conn, packet);
6344 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6345 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6346 rxi_FreePacket(packet);
6348 clock_GetTime(&now);
6350 when.sec += RX_CHALLENGE_TIMEOUT;
6351 conn->challengeEvent =
6352 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
6357 /* Call this routine to start requesting the client to authenticate
6358 * itself. This will continue until authentication is established,
6359 * the call times out, or an invalid response is returned. The
6360 * security object associated with the connection is asked to create
6361 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6362 * defined earlier. */
6364 rxi_ChallengeOn(struct rx_connection *conn)
6366 if (!conn->challengeEvent) {
6367 RXS_CreateChallenge(conn->securityObject, conn);
6368 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6373 /* Compute round trip time of the packet provided, in *rttp.
6376 /* rxi_ComputeRoundTripTime is called with peer locked. */
6377 /* sentp and/or peer may be null */
6379 rxi_ComputeRoundTripTime(struct rx_packet *p,
6380 struct clock *sentp,
6381 struct rx_peer *peer)
6383 struct clock thisRtt, *rttp = &thisRtt;
6387 clock_GetTime(rttp);
6389 if (clock_Lt(rttp, sentp)) {
6391 return; /* somebody set the clock back, don't count this time. */
6393 clock_Sub(rttp, sentp);
6394 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6395 p->header.callNumber, p, rttp->sec, rttp->usec));
6397 if (rttp->sec == 0 && rttp->usec == 0) {
6399 * The actual round trip time is shorter than the
6400 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6401 * Since we can't tell which at the moment we will assume 1ms.
6406 if (rx_stats_active) {
6407 MUTEX_ENTER(&rx_stats_mutex);
6408 if (clock_Lt(rttp, &rx_stats.minRtt))
6409 rx_stats.minRtt = *rttp;
6410 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
6411 if (rttp->sec > 60) {
6412 MUTEX_EXIT(&rx_stats_mutex);
6413 return; /* somebody set the clock ahead */
6415 rx_stats.maxRtt = *rttp;
6417 clock_Add(&rx_stats.totalRtt, rttp);
6418 rx_stats.nRttSamples++;
6419 MUTEX_EXIT(&rx_stats_mutex);
6422 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6424 /* Apply VanJacobson round-trip estimations */
6429 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
6430 * srtt is stored as fixed point with 3 bits after the binary
6431 * point (i.e., scaled by 8). The following magic is
6432 * equivalent to the smoothing algorithm in rfc793 with an
6433 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6434 * srtt'*8 = rtt + srtt*7
6435 * srtt'*8 = srtt*8 + rtt - srtt
6436 * srtt' = srtt + rtt/8 - srtt/8
6437 * srtt' = srtt + (rtt - srtt)/8
6440 delta = _8THMSEC(rttp) - peer->rtt;
6441 peer->rtt += (delta >> 3);
6444 * We accumulate a smoothed rtt variance (actually, a smoothed
6445 * mean difference), then set the retransmit timer to smoothed
6446 * rtt + 4 times the smoothed variance (was 2x in van's original
6447 * paper, but 4x works better for me, and apparently for him as
6449 * rttvar is stored as
6450 * fixed point with 2 bits after the binary point (scaled by
6451 * 4). The following is equivalent to rfc793 smoothing with
6452 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6453 * rttvar'*4 = rttvar*3 + |delta|
6454 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6455 * rttvar' = rttvar + |delta|/4 - rttvar/4
6456 * rttvar' = rttvar + (|delta| - rttvar)/4
6457 * This replaces rfc793's wired-in beta.
6458 * dev*4 = dev*4 + (|actual - expected| - dev)
6464 delta -= (peer->rtt_dev << 1);
6465 peer->rtt_dev += (delta >> 3);
6467 /* I don't have a stored RTT so I start with this value. Since I'm
6468 * probably just starting a call, and will be pushing more data down
6469 * this, I expect congestion to increase rapidly. So I fudge a
6470 * little, and I set deviance to half the rtt. In practice,
6471 * deviance tends to approach something a little less than
6472 * half the smoothed rtt. */
6473 peer->rtt = _8THMSEC(rttp) + 8;
6474 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
6476 /* the timeout is RTT + 4*MDEV but no less than rx_minPeerTimeout msec.
6477 * This is because one end or the other of these connections is usually
6478 * in a user process, and can be switched and/or swapped out. So on fast,
6479 * reliable networks, the timeout would otherwise be too short. */
6480 rtt_timeout = MAX(((peer->rtt >> 3) + peer->rtt_dev), rx_minPeerTimeout);
6481 clock_Zero(&(peer->timeout));
6482 clock_Addmsec(&(peer->timeout), rtt_timeout);
6484 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6485 p->header.callNumber, p, MSEC(rttp), peer->rtt >> 3, peer->rtt_dev >> 2, (peer->timeout.sec), (peer->timeout.usec)));
6489 /* Find all server connections that have not been active for a long time, and
6492 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6494 struct clock now, when;
6495 clock_GetTime(&now);
6497 /* Find server connection structures that haven't been used for
6498 * greater than rx_idleConnectionTime */
6500 struct rx_connection **conn_ptr, **conn_end;
6501 int i, havecalls = 0;
6502 MUTEX_ENTER(&rx_connHashTable_lock);
6503 for (conn_ptr = &rx_connHashTable[0], conn_end =
6504 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6506 struct rx_connection *conn, *next;
6507 struct rx_call *call;
6511 for (conn = *conn_ptr; conn; conn = next) {
6512 /* XXX -- Shouldn't the connection be locked? */
6515 for (i = 0; i < RX_MAXCALLS; i++) {
6516 call = conn->call[i];
6520 code = MUTEX_TRYENTER(&call->lock);
6523 #ifdef RX_ENABLE_LOCKS
6524 result = rxi_CheckCall(call, 1);
6525 #else /* RX_ENABLE_LOCKS */
6526 result = rxi_CheckCall(call);
6527 #endif /* RX_ENABLE_LOCKS */
6528 MUTEX_EXIT(&call->lock);
6530 /* If CheckCall freed the call, it might
6531 * have destroyed the connection as well,
6532 * which screws up the linked lists.
6538 if (conn->type == RX_SERVER_CONNECTION) {
6539 /* This only actually destroys the connection if
6540 * there are no outstanding calls */
6541 MUTEX_ENTER(&conn->conn_data_lock);
6542 if (!havecalls && !conn->refCount
6543 && ((conn->lastSendTime + rx_idleConnectionTime) <
6545 conn->refCount++; /* it will be decr in rx_DestroyConn */
6546 MUTEX_EXIT(&conn->conn_data_lock);
6547 #ifdef RX_ENABLE_LOCKS
6548 rxi_DestroyConnectionNoLock(conn);
6549 #else /* RX_ENABLE_LOCKS */
6550 rxi_DestroyConnection(conn);
6551 #endif /* RX_ENABLE_LOCKS */
6553 #ifdef RX_ENABLE_LOCKS
6555 MUTEX_EXIT(&conn->conn_data_lock);
6557 #endif /* RX_ENABLE_LOCKS */
6561 #ifdef RX_ENABLE_LOCKS
6562 while (rx_connCleanup_list) {
6563 struct rx_connection *conn;
6564 conn = rx_connCleanup_list;
6565 rx_connCleanup_list = rx_connCleanup_list->next;
6566 MUTEX_EXIT(&rx_connHashTable_lock);
6567 rxi_CleanupConnection(conn);
6568 MUTEX_ENTER(&rx_connHashTable_lock);
6570 MUTEX_EXIT(&rx_connHashTable_lock);
6571 #endif /* RX_ENABLE_LOCKS */
6574 /* Find any peer structures that haven't been used (haven't had an
6575 * associated connection) for greater than rx_idlePeerTime */
6577 struct rx_peer **peer_ptr, **peer_end;
6581 * Why do we need to hold the rx_peerHashTable_lock across
6582 * the incrementing of peer_ptr since the rx_peerHashTable
6583 * array is not changing? We don't.
6585 * By dropping the lock periodically we can permit other
6586 * activities to be performed while a rxi_ReapConnections
6587 * call is in progress. The goal of reap connections
6588 * is to clean up quickly without causing large amounts
6589 * of contention. Therefore, it is important that global
6590 * mutexes not be held for extended periods of time.
6592 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6593 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6595 struct rx_peer *peer, *next, *prev;
6597 MUTEX_ENTER(&rx_peerHashTable_lock);
6598 for (prev = peer = *peer_ptr; peer; peer = next) {
6600 code = MUTEX_TRYENTER(&peer->peer_lock);
6601 if ((code) && (peer->refCount == 0)
6602 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6603 rx_interface_stat_p rpc_stat, nrpc_stat;
6607 * now know that this peer object is one to be
6608 * removed from the hash table. Once it is removed
6609 * it can't be referenced by other threads.
6610 * Lets remove it first and decrement the struct
6611 * nPeerStructs count.
6613 if (peer == *peer_ptr) {
6619 if (rx_stats_active)
6620 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6623 * Now if we hold references on 'prev' and 'next'
6624 * we can safely drop the rx_peerHashTable_lock
6625 * while we destroy this 'peer' object.
6631 MUTEX_EXIT(&rx_peerHashTable_lock);
6633 MUTEX_EXIT(&peer->peer_lock);
6634 MUTEX_DESTROY(&peer->peer_lock);
6636 (&peer->rpcStats, rpc_stat, nrpc_stat,
6637 rx_interface_stat)) {
6638 unsigned int num_funcs;
6641 queue_Remove(&rpc_stat->queue_header);
6642 queue_Remove(&rpc_stat->all_peers);
6643 num_funcs = rpc_stat->stats[0].func_total;
6645 sizeof(rx_interface_stat_t) +
6646 rpc_stat->stats[0].func_total *
6647 sizeof(rx_function_entry_v1_t);
6649 rxi_Free(rpc_stat, space);
6651 MUTEX_ENTER(&rx_rpc_stats);
6652 rxi_rpc_peer_stat_cnt -= num_funcs;
6653 MUTEX_EXIT(&rx_rpc_stats);
6658 * Regain the rx_peerHashTable_lock and
6659 * decrement the reference count on 'prev'
6662 MUTEX_ENTER(&rx_peerHashTable_lock);
6669 MUTEX_EXIT(&peer->peer_lock);
6674 MUTEX_EXIT(&rx_peerHashTable_lock);
6678 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6679 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6680 * GC, just below. Really, we shouldn't have to keep moving packets from
6681 * one place to another, but instead ought to always know if we can
6682 * afford to hold onto a packet in its particular use. */
6683 MUTEX_ENTER(&rx_freePktQ_lock);
6684 if (rx_waitingForPackets) {
6685 rx_waitingForPackets = 0;
6686 #ifdef RX_ENABLE_LOCKS
6687 CV_BROADCAST(&rx_waitingForPackets_cv);
6689 osi_rxWakeup(&rx_waitingForPackets);
6692 MUTEX_EXIT(&rx_freePktQ_lock);
6695 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6696 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6700 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6701 * rx.h is sort of strange this is better. This is called with a security
6702 * object before it is discarded. Each connection using a security object has
6703 * its own refcount to the object so it won't actually be freed until the last
6704 * connection is destroyed.
6706 * This is the only rxs module call. A hold could also be written but no one
6710 rxs_Release(struct rx_securityClass *aobj)
6712 return RXS_Close(aobj);
6716 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6717 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6718 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6719 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6721 /* Adjust our estimate of the transmission rate to this peer, given
6722 * that the packet p was just acked. We can adjust peer->timeout and
6723 * call->twind. Pragmatically, this is called
6724 * only with packets of maximal length.
6725 * Called with peer and call locked.
6729 rxi_ComputeRate(struct rx_peer *peer, struct rx_call *call,
6730 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6732 afs_int32 xferSize, xferMs;
6736 /* Count down packets */
6737 if (peer->rateFlag > 0)
6739 /* Do nothing until we're enabled */
6740 if (peer->rateFlag != 0)
6745 /* Count only when the ack seems legitimate */
6746 switch (ackReason) {
6747 case RX_ACK_REQUESTED:
6749 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6753 case RX_ACK_PING_RESPONSE:
6754 if (p) /* want the response to ping-request, not data send */
6756 clock_GetTime(&newTO);
6757 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6758 clock_Sub(&newTO, &call->pingRequestTime);
6759 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6763 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6770 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %d.%06d, rtt %u, ps %u)",
6771 ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"),
6772 xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
6774 /* Track only packets that are big enough. */
6775 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6779 /* absorb RTT data (in milliseconds) for these big packets */
6780 if (peer->smRtt == 0) {
6781 peer->smRtt = xferMs;
6783 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6788 if (peer->countDown) {
6792 peer->countDown = 10; /* recalculate only every so often */
6794 /* In practice, we can measure only the RTT for full packets,
6795 * because of the way Rx acks the data that it receives. (If it's
6796 * smaller than a full packet, it often gets implicitly acked
6797 * either by the call response (from a server) or by the next call
6798 * (from a client), and either case confuses transmission times
6799 * with processing times.) Therefore, replace the above
6800 * more-sophisticated processing with a simpler version, where the
6801 * smoothed RTT is kept for full-size packets, and the time to
6802 * transmit a windowful of full-size packets is simply RTT *
6803 * windowSize. Again, we take two steps:
6804 - ensure the timeout is large enough for a single packet's RTT;
6805 - ensure that the window is small enough to fit in the desired timeout.*/
6807 /* First, the timeout check. */
6808 minTime = peer->smRtt;
6809 /* Get a reasonable estimate for a timeout period */
6811 newTO.sec = minTime / 1000;
6812 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6814 /* Increase the timeout period so that we can always do at least
6815 * one packet exchange */
6816 if (clock_Gt(&newTO, &peer->timeout)) {
6818 dpf(("CONG peer %lx/%u: timeout %d.%06d ==> %ld.%06d (rtt %u, ps %u)",
6819 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec,
6820 newTO.sec, newTO.usec, peer->smRtt, peer->packetSize));
6822 peer->timeout = newTO;
6825 /* Now, get an estimate for the transmit window size. */
6826 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6827 /* Now, convert to the number of full packets that could fit in a
6828 * reasonable fraction of that interval */
6829 minTime /= (peer->smRtt << 1);
6830 xferSize = minTime; /* (make a copy) */
6832 /* Now clamp the size to reasonable bounds. */
6835 else if (minTime > rx_Window)
6836 minTime = rx_Window;
6837 /* if (minTime != peer->maxWindow) {
6838 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6839 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6840 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6842 peer->maxWindow = minTime;
6843 elide... call->twind = minTime;
6847 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6848 * Discern this by calculating the timeout necessary for rx_Window
6850 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6851 /* calculate estimate for transmission interval in milliseconds */
6852 minTime = rx_Window * peer->smRtt;
6853 if (minTime < 1000) {
6854 dpf(("CONG peer %lx/%u: cut TO %d.%06d by 0.5 (rtt %u, ps %u)",
6855 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6856 peer->timeout.usec, peer->smRtt, peer->packetSize));
6858 newTO.sec = 0; /* cut back on timeout by half a second */
6859 newTO.usec = 500000;
6860 clock_Sub(&peer->timeout, &newTO);
6865 } /* end of rxi_ComputeRate */
6866 #endif /* ADAPT_WINDOW */
6874 #define TRACE_OPTION_RX_DEBUG 16
6882 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6883 0, KEY_QUERY_VALUE, &parmKey);
6884 if (code != ERROR_SUCCESS)
6887 dummyLen = sizeof(TraceOption);
6888 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6889 (BYTE *) &TraceOption, &dummyLen);
6890 if (code == ERROR_SUCCESS) {
6891 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
6893 RegCloseKey (parmKey);
6894 #endif /* AFS_NT40_ENV */
6899 rx_DebugOnOff(int on)
6903 rxdebug_active = on;
6909 rx_StatsOnOff(int on)
6912 rx_stats_active = on;
6917 /* Don't call this debugging routine directly; use dpf */
6919 rxi_DebugPrint(char *format, ...)
6928 va_start(ap, format);
6930 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6933 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
6935 if (msg[len-1] != '\n') {
6939 OutputDebugString(msg);
6946 va_start(ap, format);
6948 clock_GetTime(&now);
6949 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
6950 (unsigned int)now.usec);
6951 vfprintf(rx_Log, format, ap);
6960 * This function is used to process the rx_stats structure that is local
6961 * to a process as well as an rx_stats structure received from a remote
6962 * process (via rxdebug). Therefore, it needs to do minimal version
6966 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
6967 afs_int32 freePackets, char version)
6972 if (size != sizeof(struct rx_statistics)) {
6974 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
6975 size, sizeof(struct rx_statistics));
6978 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6981 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6982 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
6983 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6984 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6985 s->specialPktAllocFailures);
6987 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
6988 s->receivePktAllocFailures, s->sendPktAllocFailures,
6989 s->specialPktAllocFailures);
6993 " greedy %u, " "bogusReads %u (last from host %x), "
6994 "noPackets %u, " "noBuffers %u, " "selects %u, "
6995 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
6996 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6997 s->selects, s->sendSelects);
6999 fprintf(file, " packets read: ");
7000 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7001 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7003 fprintf(file, "\n");
7006 " other read counters: data %u, " "ack %u, " "dup %u "
7007 "spurious %u " "dally %u\n", s->dataPacketsRead,
7008 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7009 s->ignorePacketDally);
7011 fprintf(file, " packets sent: ");
7012 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7013 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7015 fprintf(file, "\n");
7018 " other send counters: ack %u, " "data %u (not resends), "
7019 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7020 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7021 s->dataPacketsPushed, s->ignoreAckedPacket);
7024 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7025 s->netSendFailures, (int)s->fatalErrors);
7027 if (s->nRttSamples) {
7028 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7029 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7031 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7032 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7036 " %d server connections, " "%d client connections, "
7037 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7038 s->nServerConns, s->nClientConns, s->nPeerStructs,
7039 s->nCallStructs, s->nFreeCallStructs);
7041 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7042 fprintf(file, " %d clock updates\n", clock_nUpdates);
7045 fprintf(file, "ERROR: compiled without RXDEBUG\n");
7049 /* for backward compatibility */
7051 rx_PrintStats(FILE * file)
7053 MUTEX_ENTER(&rx_stats_mutex);
7054 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
7056 MUTEX_EXIT(&rx_stats_mutex);
7060 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7062 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
7063 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
7064 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
7067 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
7068 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
7069 (int)peer->timeout.usec, peer->nSent, peer->reSends);
7072 " Packet size %d, " "max in packet skew %d, "
7073 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
7074 (int)peer->outPacketSkew);
7078 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7080 * This mutex protects the following static variables:
7084 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7085 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7087 #define LOCK_RX_DEBUG
7088 #define UNLOCK_RX_DEBUG
7089 #endif /* AFS_PTHREAD_ENV */
7093 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7094 u_char type, void *inputData, size_t inputLength,
7095 void *outputData, size_t outputLength)
7097 static afs_int32 counter = 100;
7098 time_t waitTime, waitCount;
7099 struct rx_header theader;
7102 struct timeval tv_now, tv_wake, tv_delta;
7103 struct sockaddr_in taddr, faddr;
7117 tp = &tbuffer[sizeof(struct rx_header)];
7118 taddr.sin_family = AF_INET;
7119 taddr.sin_port = remotePort;
7120 taddr.sin_addr.s_addr = remoteAddr;
7121 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7122 taddr.sin_len = sizeof(struct sockaddr_in);
7125 memset(&theader, 0, sizeof(theader));
7126 theader.epoch = htonl(999);
7128 theader.callNumber = htonl(counter);
7131 theader.type = type;
7132 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7133 theader.serviceId = 0;
7135 memcpy(tbuffer, &theader, sizeof(theader));
7136 memcpy(tp, inputData, inputLength);
7138 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7139 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7141 /* see if there's a packet available */
7142 gettimeofday(&tv_wake,0);
7143 tv_wake.tv_sec += waitTime;
7146 FD_SET(socket, &imask);
7147 tv_delta.tv_sec = tv_wake.tv_sec;
7148 tv_delta.tv_usec = tv_wake.tv_usec;
7149 gettimeofday(&tv_now, 0);
7151 if (tv_delta.tv_usec < tv_now.tv_usec) {
7153 tv_delta.tv_usec += 1000000;
7156 tv_delta.tv_usec -= tv_now.tv_usec;
7158 if (tv_delta.tv_sec < tv_now.tv_sec) {
7162 tv_delta.tv_sec -= tv_now.tv_sec;
7165 code = select(0, &imask, 0, 0, &tv_delta);
7166 #else /* AFS_NT40_ENV */
7167 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7168 #endif /* AFS_NT40_ENV */
7169 if (code == 1 && FD_ISSET(socket, &imask)) {
7170 /* now receive a packet */
7171 faddrLen = sizeof(struct sockaddr_in);
7173 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7174 (struct sockaddr *)&faddr, &faddrLen);
7177 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7178 if (counter == ntohl(theader.callNumber))
7186 /* see if we've timed out */
7194 code -= sizeof(struct rx_header);
7195 if (code > outputLength)
7196 code = outputLength;
7197 memcpy(outputData, tp, code);
7200 #endif /* RXDEBUG */
7203 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7204 afs_uint16 remotePort, struct rx_debugStats * stat,
7205 afs_uint32 * supportedValues)
7211 struct rx_debugIn in;
7213 *supportedValues = 0;
7214 in.type = htonl(RX_DEBUGI_GETSTATS);
7217 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7218 &in, sizeof(in), stat, sizeof(*stat));
7221 * If the call was successful, fixup the version and indicate
7222 * what contents of the stat structure are valid.
7223 * Also do net to host conversion of fields here.
7227 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7228 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7230 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7231 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7233 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7234 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7236 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7237 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7239 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7240 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7242 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7243 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7245 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7246 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7248 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7249 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7251 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7252 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7254 stat->nFreePackets = ntohl(stat->nFreePackets);
7255 stat->packetReclaims = ntohl(stat->packetReclaims);
7256 stat->callsExecuted = ntohl(stat->callsExecuted);
7257 stat->nWaiting = ntohl(stat->nWaiting);
7258 stat->idleThreads = ntohl(stat->idleThreads);
7259 stat->nWaited = ntohl(stat->nWaited);
7260 stat->nPackets = ntohl(stat->nPackets);
7267 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7268 afs_uint16 remotePort, struct rx_statistics * stat,
7269 afs_uint32 * supportedValues)
7275 struct rx_debugIn in;
7276 afs_int32 *lp = (afs_int32 *) stat;
7280 * supportedValues is currently unused, but added to allow future
7281 * versioning of this function.
7284 *supportedValues = 0;
7285 in.type = htonl(RX_DEBUGI_RXSTATS);
7287 memset(stat, 0, sizeof(*stat));
7289 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7290 &in, sizeof(in), stat, sizeof(*stat));
7295 * Do net to host conversion here
7298 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7307 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7308 afs_uint16 remotePort, size_t version_length,
7313 return MakeDebugCall(socket, remoteAddr, remotePort,
7314 RX_PACKET_TYPE_VERSION, a, 1, version,
7322 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7323 afs_uint16 remotePort, afs_int32 * nextConnection,
7324 int allConnections, afs_uint32 debugSupportedValues,
7325 struct rx_debugConn * conn,
7326 afs_uint32 * supportedValues)
7332 struct rx_debugIn in;
7336 * supportedValues is currently unused, but added to allow future
7337 * versioning of this function.
7340 *supportedValues = 0;
7341 if (allConnections) {
7342 in.type = htonl(RX_DEBUGI_GETALLCONN);
7344 in.type = htonl(RX_DEBUGI_GETCONN);
7346 in.index = htonl(*nextConnection);
7347 memset(conn, 0, sizeof(*conn));
7349 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7350 &in, sizeof(in), conn, sizeof(*conn));
7353 *nextConnection += 1;
7356 * Convert old connection format to new structure.
7359 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7360 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7361 #define MOVEvL(a) (conn->a = vL->a)
7363 /* any old or unrecognized version... */
7364 for (i = 0; i < RX_MAXCALLS; i++) {
7365 MOVEvL(callState[i]);
7366 MOVEvL(callMode[i]);
7367 MOVEvL(callFlags[i]);
7368 MOVEvL(callOther[i]);
7370 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7371 MOVEvL(secStats.type);
7372 MOVEvL(secStats.level);
7373 MOVEvL(secStats.flags);
7374 MOVEvL(secStats.expires);
7375 MOVEvL(secStats.packetsReceived);
7376 MOVEvL(secStats.packetsSent);
7377 MOVEvL(secStats.bytesReceived);
7378 MOVEvL(secStats.bytesSent);
7383 * Do net to host conversion here
7385 * I don't convert host or port since we are most likely
7386 * going to want these in NBO.
7388 conn->cid = ntohl(conn->cid);
7389 conn->serial = ntohl(conn->serial);
7390 for (i = 0; i < RX_MAXCALLS; i++) {
7391 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7393 conn->error = ntohl(conn->error);
7394 conn->secStats.flags = ntohl(conn->secStats.flags);
7395 conn->secStats.expires = ntohl(conn->secStats.expires);
7396 conn->secStats.packetsReceived =
7397 ntohl(conn->secStats.packetsReceived);
7398 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7399 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7400 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7401 conn->epoch = ntohl(conn->epoch);
7402 conn->natMTU = ntohl(conn->natMTU);
7409 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7410 afs_uint16 remotePort, afs_int32 * nextPeer,
7411 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7412 afs_uint32 * supportedValues)
7418 struct rx_debugIn in;
7421 * supportedValues is currently unused, but added to allow future
7422 * versioning of this function.
7425 *supportedValues = 0;
7426 in.type = htonl(RX_DEBUGI_GETPEER);
7427 in.index = htonl(*nextPeer);
7428 memset(peer, 0, sizeof(*peer));
7430 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7431 &in, sizeof(in), peer, sizeof(*peer));
7437 * Do net to host conversion here
7439 * I don't convert host or port since we are most likely
7440 * going to want these in NBO.
7442 peer->ifMTU = ntohs(peer->ifMTU);
7443 peer->idleWhen = ntohl(peer->idleWhen);
7444 peer->refCount = ntohs(peer->refCount);
7445 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7446 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7447 peer->rtt = ntohl(peer->rtt);
7448 peer->rtt_dev = ntohl(peer->rtt_dev);
7449 peer->timeout.sec = ntohl(peer->timeout.sec);
7450 peer->timeout.usec = ntohl(peer->timeout.usec);
7451 peer->nSent = ntohl(peer->nSent);
7452 peer->reSends = ntohl(peer->reSends);
7453 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7454 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7455 peer->rateFlag = ntohl(peer->rateFlag);
7456 peer->natMTU = ntohs(peer->natMTU);
7457 peer->maxMTU = ntohs(peer->maxMTU);
7458 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7459 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7460 peer->MTU = ntohs(peer->MTU);
7461 peer->cwind = ntohs(peer->cwind);
7462 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7463 peer->congestSeq = ntohs(peer->congestSeq);
7464 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7465 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7466 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7467 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7474 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7475 struct rx_debugPeer * peerStats)
7478 afs_int32 error = 1; /* default to "did not succeed" */
7479 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7481 MUTEX_ENTER(&rx_peerHashTable_lock);
7482 for(tp = rx_peerHashTable[hashValue];
7483 tp != NULL; tp = tp->next) {
7484 if (tp->host == peerHost)
7490 MUTEX_EXIT(&rx_peerHashTable_lock);
7494 MUTEX_ENTER(&tp->peer_lock);
7495 peerStats->host = tp->host;
7496 peerStats->port = tp->port;
7497 peerStats->ifMTU = tp->ifMTU;
7498 peerStats->idleWhen = tp->idleWhen;
7499 peerStats->refCount = tp->refCount;
7500 peerStats->burstSize = tp->burstSize;
7501 peerStats->burst = tp->burst;
7502 peerStats->burstWait.sec = tp->burstWait.sec;
7503 peerStats->burstWait.usec = tp->burstWait.usec;
7504 peerStats->rtt = tp->rtt;
7505 peerStats->rtt_dev = tp->rtt_dev;
7506 peerStats->timeout.sec = tp->timeout.sec;
7507 peerStats->timeout.usec = tp->timeout.usec;
7508 peerStats->nSent = tp->nSent;
7509 peerStats->reSends = tp->reSends;
7510 peerStats->inPacketSkew = tp->inPacketSkew;
7511 peerStats->outPacketSkew = tp->outPacketSkew;
7512 peerStats->rateFlag = tp->rateFlag;
7513 peerStats->natMTU = tp->natMTU;
7514 peerStats->maxMTU = tp->maxMTU;
7515 peerStats->maxDgramPackets = tp->maxDgramPackets;
7516 peerStats->ifDgramPackets = tp->ifDgramPackets;
7517 peerStats->MTU = tp->MTU;
7518 peerStats->cwind = tp->cwind;
7519 peerStats->nDgramPackets = tp->nDgramPackets;
7520 peerStats->congestSeq = tp->congestSeq;
7521 peerStats->bytesSent.high = tp->bytesSent.high;
7522 peerStats->bytesSent.low = tp->bytesSent.low;
7523 peerStats->bytesReceived.high = tp->bytesReceived.high;
7524 peerStats->bytesReceived.low = tp->bytesReceived.low;
7525 MUTEX_EXIT(&tp->peer_lock);
7527 MUTEX_ENTER(&rx_peerHashTable_lock);
7530 MUTEX_EXIT(&rx_peerHashTable_lock);
7538 struct rx_serverQueueEntry *np;
7541 struct rx_call *call;
7542 struct rx_serverQueueEntry *sq;
7546 if (rxinit_status == 1) {
7548 return; /* Already shutdown. */
7552 #ifndef AFS_PTHREAD_ENV
7553 FD_ZERO(&rx_selectMask);
7554 #endif /* AFS_PTHREAD_ENV */
7555 rxi_dataQuota = RX_MAX_QUOTA;
7556 #ifndef AFS_PTHREAD_ENV
7558 #endif /* AFS_PTHREAD_ENV */
7561 #ifndef AFS_PTHREAD_ENV
7562 #ifndef AFS_USE_GETTIMEOFDAY
7564 #endif /* AFS_USE_GETTIMEOFDAY */
7565 #endif /* AFS_PTHREAD_ENV */
7567 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7568 call = queue_First(&rx_freeCallQueue, rx_call);
7570 rxi_Free(call, sizeof(struct rx_call));
7573 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7574 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7580 struct rx_peer **peer_ptr, **peer_end;
7581 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7582 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7584 struct rx_peer *peer, *next;
7586 MUTEX_ENTER(&rx_peerHashTable_lock);
7587 for (peer = *peer_ptr; peer; peer = next) {
7588 rx_interface_stat_p rpc_stat, nrpc_stat;
7591 MUTEX_ENTER(&rx_rpc_stats);
7592 MUTEX_ENTER(&peer->peer_lock);
7594 (&peer->rpcStats, rpc_stat, nrpc_stat,
7595 rx_interface_stat)) {
7596 unsigned int num_funcs;
7599 queue_Remove(&rpc_stat->queue_header);
7600 queue_Remove(&rpc_stat->all_peers);
7601 num_funcs = rpc_stat->stats[0].func_total;
7603 sizeof(rx_interface_stat_t) +
7604 rpc_stat->stats[0].func_total *
7605 sizeof(rx_function_entry_v1_t);
7607 rxi_Free(rpc_stat, space);
7609 /* rx_rpc_stats must be held */
7610 rxi_rpc_peer_stat_cnt -= num_funcs;
7612 MUTEX_EXIT(&peer->peer_lock);
7613 MUTEX_EXIT(&rx_rpc_stats);
7617 if (rx_stats_active)
7618 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
7620 MUTEX_EXIT(&rx_peerHashTable_lock);
7623 for (i = 0; i < RX_MAX_SERVICES; i++) {
7625 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7627 for (i = 0; i < rx_hashTableSize; i++) {
7628 struct rx_connection *tc, *ntc;
7629 MUTEX_ENTER(&rx_connHashTable_lock);
7630 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7632 for (j = 0; j < RX_MAXCALLS; j++) {
7634 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7637 rxi_Free(tc, sizeof(*tc));
7639 MUTEX_EXIT(&rx_connHashTable_lock);
7642 MUTEX_ENTER(&freeSQEList_lock);
7644 while ((np = rx_FreeSQEList)) {
7645 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7646 MUTEX_DESTROY(&np->lock);
7647 rxi_Free(np, sizeof(*np));
7650 MUTEX_EXIT(&freeSQEList_lock);
7651 MUTEX_DESTROY(&freeSQEList_lock);
7652 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7653 MUTEX_DESTROY(&rx_connHashTable_lock);
7654 MUTEX_DESTROY(&rx_peerHashTable_lock);
7655 MUTEX_DESTROY(&rx_serverPool_lock);
7657 osi_Free(rx_connHashTable,
7658 rx_hashTableSize * sizeof(struct rx_connection *));
7659 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7661 UNPIN(rx_connHashTable,
7662 rx_hashTableSize * sizeof(struct rx_connection *));
7663 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7665 rxi_FreeAllPackets();
7667 MUTEX_ENTER(&rx_quota_mutex);
7668 rxi_dataQuota = RX_MAX_QUOTA;
7669 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7670 MUTEX_EXIT(&rx_quota_mutex);
7675 #ifdef RX_ENABLE_LOCKS
7677 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7679 if (!MUTEX_ISMINE(lockaddr))
7680 osi_Panic("Lock not held: %s", msg);
7682 #endif /* RX_ENABLE_LOCKS */
7687 * Routines to implement connection specific data.
7691 rx_KeyCreate(rx_destructor_t rtn)
7694 MUTEX_ENTER(&rxi_keyCreate_lock);
7695 key = rxi_keyCreate_counter++;
7696 rxi_keyCreate_destructor = (rx_destructor_t *)
7697 realloc((void *)rxi_keyCreate_destructor,
7698 (key + 1) * sizeof(rx_destructor_t));
7699 rxi_keyCreate_destructor[key] = rtn;
7700 MUTEX_EXIT(&rxi_keyCreate_lock);
7705 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7708 MUTEX_ENTER(&conn->conn_data_lock);
7709 if (!conn->specific) {
7710 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7711 for (i = 0; i < key; i++)
7712 conn->specific[i] = NULL;
7713 conn->nSpecific = key + 1;
7714 conn->specific[key] = ptr;
7715 } else if (key >= conn->nSpecific) {
7716 conn->specific = (void **)
7717 realloc(conn->specific, (key + 1) * sizeof(void *));
7718 for (i = conn->nSpecific; i < key; i++)
7719 conn->specific[i] = NULL;
7720 conn->nSpecific = key + 1;
7721 conn->specific[key] = ptr;
7723 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7724 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7725 conn->specific[key] = ptr;
7727 MUTEX_EXIT(&conn->conn_data_lock);
7731 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7734 MUTEX_ENTER(&svc->svc_data_lock);
7735 if (!svc->specific) {
7736 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
7737 for (i = 0; i < key; i++)
7738 svc->specific[i] = NULL;
7739 svc->nSpecific = key + 1;
7740 svc->specific[key] = ptr;
7741 } else if (key >= svc->nSpecific) {
7742 svc->specific = (void **)
7743 realloc(svc->specific, (key + 1) * sizeof(void *));
7744 for (i = svc->nSpecific; i < key; i++)
7745 svc->specific[i] = NULL;
7746 svc->nSpecific = key + 1;
7747 svc->specific[key] = ptr;
7749 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7750 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7751 svc->specific[key] = ptr;
7753 MUTEX_EXIT(&svc->svc_data_lock);
7757 rx_GetSpecific(struct rx_connection *conn, int key)
7760 MUTEX_ENTER(&conn->conn_data_lock);
7761 if (key >= conn->nSpecific)
7764 ptr = conn->specific[key];
7765 MUTEX_EXIT(&conn->conn_data_lock);
7770 rx_GetServiceSpecific(struct rx_service *svc, int key)
7773 MUTEX_ENTER(&svc->svc_data_lock);
7774 if (key >= svc->nSpecific)
7777 ptr = svc->specific[key];
7778 MUTEX_EXIT(&svc->svc_data_lock);
7783 #endif /* !KERNEL */
7786 * processStats is a queue used to store the statistics for the local
7787 * process. Its contents are similar to the contents of the rpcStats
7788 * queue on a rx_peer structure, but the actual data stored within
7789 * this queue contains totals across the lifetime of the process (assuming
7790 * the stats have not been reset) - unlike the per peer structures
7791 * which can come and go based upon the peer lifetime.
7794 static struct rx_queue processStats = { &processStats, &processStats };
7797 * peerStats is a queue used to store the statistics for all peer structs.
7798 * Its contents are the union of all the peer rpcStats queues.
7801 static struct rx_queue peerStats = { &peerStats, &peerStats };
7804 * rxi_monitor_processStats is used to turn process wide stat collection
7808 static int rxi_monitor_processStats = 0;
7811 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7814 static int rxi_monitor_peerStats = 0;
7817 * rxi_AddRpcStat - given all of the information for a particular rpc
7818 * call, create (if needed) and update the stat totals for the rpc.
7822 * IN stats - the queue of stats that will be updated with the new value
7824 * IN rxInterface - a unique number that identifies the rpc interface
7826 * IN currentFunc - the index of the function being invoked
7828 * IN totalFunc - the total number of functions in this interface
7830 * IN queueTime - the amount of time this function waited for a thread
7832 * IN execTime - the amount of time this function invocation took to execute
7834 * IN bytesSent - the number bytes sent by this invocation
7836 * IN bytesRcvd - the number bytes received by this invocation
7838 * IN isServer - if true, this invocation was made to a server
7840 * IN remoteHost - the ip address of the remote host
7842 * IN remotePort - the port of the remote host
7844 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7846 * INOUT counter - if a new stats structure is allocated, the counter will
7847 * be updated with the new number of allocated stat structures
7855 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
7856 afs_uint32 currentFunc, afs_uint32 totalFunc,
7857 struct clock *queueTime, struct clock *execTime,
7858 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
7859 afs_uint32 remoteHost, afs_uint32 remotePort,
7860 int addToPeerList, unsigned int *counter)
7863 rx_interface_stat_p rpc_stat, nrpc_stat;
7866 * See if there's already a structure for this interface
7869 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7870 if ((rpc_stat->stats[0].interfaceId == rxInterface)
7871 && (rpc_stat->stats[0].remote_is_server == isServer))
7876 * Didn't find a match so allocate a new structure and add it to the
7880 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
7881 || (rpc_stat->stats[0].interfaceId != rxInterface)
7882 || (rpc_stat->stats[0].remote_is_server != isServer)) {
7887 sizeof(rx_interface_stat_t) +
7888 totalFunc * sizeof(rx_function_entry_v1_t);
7890 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
7891 if (rpc_stat == NULL) {
7895 *counter += totalFunc;
7896 for (i = 0; i < totalFunc; i++) {
7897 rpc_stat->stats[i].remote_peer = remoteHost;
7898 rpc_stat->stats[i].remote_port = remotePort;
7899 rpc_stat->stats[i].remote_is_server = isServer;
7900 rpc_stat->stats[i].interfaceId = rxInterface;
7901 rpc_stat->stats[i].func_total = totalFunc;
7902 rpc_stat->stats[i].func_index = i;
7903 hzero(rpc_stat->stats[i].invocations);
7904 hzero(rpc_stat->stats[i].bytes_sent);
7905 hzero(rpc_stat->stats[i].bytes_rcvd);
7906 rpc_stat->stats[i].queue_time_sum.sec = 0;
7907 rpc_stat->stats[i].queue_time_sum.usec = 0;
7908 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7909 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7910 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7911 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7912 rpc_stat->stats[i].queue_time_max.sec = 0;
7913 rpc_stat->stats[i].queue_time_max.usec = 0;
7914 rpc_stat->stats[i].execution_time_sum.sec = 0;
7915 rpc_stat->stats[i].execution_time_sum.usec = 0;
7916 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7917 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7918 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7919 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7920 rpc_stat->stats[i].execution_time_max.sec = 0;
7921 rpc_stat->stats[i].execution_time_max.usec = 0;
7923 queue_Prepend(stats, rpc_stat);
7924 if (addToPeerList) {
7925 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7930 * Increment the stats for this function
7933 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7934 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7935 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7936 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7937 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7938 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7939 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7941 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7942 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7944 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7945 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7947 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7948 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7950 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7951 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7959 * rx_IncrementTimeAndCount - increment the times and count for a particular
7964 * IN peer - the peer who invoked the rpc
7966 * IN rxInterface - a unique number that identifies the rpc interface
7968 * IN currentFunc - the index of the function being invoked
7970 * IN totalFunc - the total number of functions in this interface
7972 * IN queueTime - the amount of time this function waited for a thread
7974 * IN execTime - the amount of time this function invocation took to execute
7976 * IN bytesSent - the number bytes sent by this invocation
7978 * IN bytesRcvd - the number bytes received by this invocation
7980 * IN isServer - if true, this invocation was made to a server
7988 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7989 afs_uint32 currentFunc, afs_uint32 totalFunc,
7990 struct clock *queueTime, struct clock *execTime,
7991 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7995 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7998 MUTEX_ENTER(&rx_rpc_stats);
8000 if (rxi_monitor_peerStats) {
8001 MUTEX_ENTER(&peer->peer_lock);
8002 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8003 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8004 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8005 MUTEX_EXIT(&peer->peer_lock);
8008 if (rxi_monitor_processStats) {
8009 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8010 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8011 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8014 MUTEX_EXIT(&rx_rpc_stats);
8019 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8023 * IN callerVersion - the rpc stat version of the caller.
8025 * IN count - the number of entries to marshall.
8027 * IN stats - pointer to stats to be marshalled.
8029 * OUT ptr - Where to store the marshalled data.
8036 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8037 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8043 * We only support the first version
8045 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8046 *(ptr++) = stats->remote_peer;
8047 *(ptr++) = stats->remote_port;
8048 *(ptr++) = stats->remote_is_server;
8049 *(ptr++) = stats->interfaceId;
8050 *(ptr++) = stats->func_total;
8051 *(ptr++) = stats->func_index;
8052 *(ptr++) = hgethi(stats->invocations);
8053 *(ptr++) = hgetlo(stats->invocations);
8054 *(ptr++) = hgethi(stats->bytes_sent);
8055 *(ptr++) = hgetlo(stats->bytes_sent);
8056 *(ptr++) = hgethi(stats->bytes_rcvd);
8057 *(ptr++) = hgetlo(stats->bytes_rcvd);
8058 *(ptr++) = stats->queue_time_sum.sec;
8059 *(ptr++) = stats->queue_time_sum.usec;
8060 *(ptr++) = stats->queue_time_sum_sqr.sec;
8061 *(ptr++) = stats->queue_time_sum_sqr.usec;
8062 *(ptr++) = stats->queue_time_min.sec;
8063 *(ptr++) = stats->queue_time_min.usec;
8064 *(ptr++) = stats->queue_time_max.sec;
8065 *(ptr++) = stats->queue_time_max.usec;
8066 *(ptr++) = stats->execution_time_sum.sec;
8067 *(ptr++) = stats->execution_time_sum.usec;
8068 *(ptr++) = stats->execution_time_sum_sqr.sec;
8069 *(ptr++) = stats->execution_time_sum_sqr.usec;
8070 *(ptr++) = stats->execution_time_min.sec;
8071 *(ptr++) = stats->execution_time_min.usec;
8072 *(ptr++) = stats->execution_time_max.sec;
8073 *(ptr++) = stats->execution_time_max.usec;
8079 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8084 * IN callerVersion - the rpc stat version of the caller
8086 * OUT myVersion - the rpc stat version of this function
8088 * OUT clock_sec - local time seconds
8090 * OUT clock_usec - local time microseconds
8092 * OUT allocSize - the number of bytes allocated to contain stats
8094 * OUT statCount - the number stats retrieved from this process.
8096 * OUT stats - the actual stats retrieved from this process.
8100 * Returns void. If successful, stats will != NULL.
8104 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8105 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8106 size_t * allocSize, afs_uint32 * statCount,
8107 afs_uint32 ** stats)
8117 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8120 * Check to see if stats are enabled
8123 MUTEX_ENTER(&rx_rpc_stats);
8124 if (!rxi_monitor_processStats) {
8125 MUTEX_EXIT(&rx_rpc_stats);
8129 clock_GetTime(&now);
8130 *clock_sec = now.sec;
8131 *clock_usec = now.usec;
8134 * Allocate the space based upon the caller version
8136 * If the client is at an older version than we are,
8137 * we return the statistic data in the older data format, but
8138 * we still return our version number so the client knows we
8139 * are maintaining more data than it can retrieve.
8142 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8143 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8144 *statCount = rxi_rpc_process_stat_cnt;
8147 * This can't happen yet, but in the future version changes
8148 * can be handled by adding additional code here
8152 if (space > (size_t) 0) {
8154 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
8157 rx_interface_stat_p rpc_stat, nrpc_stat;
8161 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8163 * Copy the data based upon the caller version
8165 rx_MarshallProcessRPCStats(callerVersion,
8166 rpc_stat->stats[0].func_total,
8167 rpc_stat->stats, &ptr);
8173 MUTEX_EXIT(&rx_rpc_stats);
8178 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8182 * IN callerVersion - the rpc stat version of the caller
8184 * OUT myVersion - the rpc stat version of this function
8186 * OUT clock_sec - local time seconds
8188 * OUT clock_usec - local time microseconds
8190 * OUT allocSize - the number of bytes allocated to contain stats
8192 * OUT statCount - the number of stats retrieved from the individual
8195 * OUT stats - the actual stats retrieved from the individual peer structures.
8199 * Returns void. If successful, stats will != NULL.
8203 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8204 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8205 size_t * allocSize, afs_uint32 * statCount,
8206 afs_uint32 ** stats)
8216 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8219 * Check to see if stats are enabled
8222 MUTEX_ENTER(&rx_rpc_stats);
8223 if (!rxi_monitor_peerStats) {
8224 MUTEX_EXIT(&rx_rpc_stats);
8228 clock_GetTime(&now);
8229 *clock_sec = now.sec;
8230 *clock_usec = now.usec;
8233 * Allocate the space based upon the caller version
8235 * If the client is at an older version than we are,
8236 * we return the statistic data in the older data format, but
8237 * we still return our version number so the client knows we
8238 * are maintaining more data than it can retrieve.
8241 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8242 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8243 *statCount = rxi_rpc_peer_stat_cnt;
8246 * This can't happen yet, but in the future version changes
8247 * can be handled by adding additional code here
8251 if (space > (size_t) 0) {
8253 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
8256 rx_interface_stat_p rpc_stat, nrpc_stat;
8260 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8262 * We have to fix the offset of rpc_stat since we are
8263 * keeping this structure on two rx_queues. The rx_queue
8264 * package assumes that the rx_queue member is the first
8265 * member of the structure. That is, rx_queue assumes that
8266 * any one item is only on one queue at a time. We are
8267 * breaking that assumption and so we have to do a little
8268 * math to fix our pointers.
8271 fix_offset = (char *)rpc_stat;
8272 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8273 rpc_stat = (rx_interface_stat_p) fix_offset;
8276 * Copy the data based upon the caller version
8278 rx_MarshallProcessRPCStats(callerVersion,
8279 rpc_stat->stats[0].func_total,
8280 rpc_stat->stats, &ptr);
8286 MUTEX_EXIT(&rx_rpc_stats);
8291 * rx_FreeRPCStats - free memory allocated by
8292 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8296 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8297 * rx_RetrievePeerRPCStats
8299 * IN allocSize - the number of bytes in stats.
8307 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8309 rxi_Free(stats, allocSize);
8313 * rx_queryProcessRPCStats - see if process rpc stat collection is
8314 * currently enabled.
8320 * Returns 0 if stats are not enabled != 0 otherwise
8324 rx_queryProcessRPCStats(void)
8327 MUTEX_ENTER(&rx_rpc_stats);
8328 rc = rxi_monitor_processStats;
8329 MUTEX_EXIT(&rx_rpc_stats);
8334 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8340 * Returns 0 if stats are not enabled != 0 otherwise
8344 rx_queryPeerRPCStats(void)
8347 MUTEX_ENTER(&rx_rpc_stats);
8348 rc = rxi_monitor_peerStats;
8349 MUTEX_EXIT(&rx_rpc_stats);
8354 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8364 rx_enableProcessRPCStats(void)
8366 MUTEX_ENTER(&rx_rpc_stats);
8367 rx_enable_stats = 1;
8368 rxi_monitor_processStats = 1;
8369 MUTEX_EXIT(&rx_rpc_stats);
8373 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8383 rx_enablePeerRPCStats(void)
8385 MUTEX_ENTER(&rx_rpc_stats);
8386 rx_enable_stats = 1;
8387 rxi_monitor_peerStats = 1;
8388 MUTEX_EXIT(&rx_rpc_stats);
8392 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8402 rx_disableProcessRPCStats(void)
8404 rx_interface_stat_p rpc_stat, nrpc_stat;
8407 MUTEX_ENTER(&rx_rpc_stats);
8410 * Turn off process statistics and if peer stats is also off, turn
8414 rxi_monitor_processStats = 0;
8415 if (rxi_monitor_peerStats == 0) {
8416 rx_enable_stats = 0;
8419 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8420 unsigned int num_funcs = 0;
8423 queue_Remove(rpc_stat);
8424 num_funcs = rpc_stat->stats[0].func_total;
8426 sizeof(rx_interface_stat_t) +
8427 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8429 rxi_Free(rpc_stat, space);
8430 rxi_rpc_process_stat_cnt -= num_funcs;
8432 MUTEX_EXIT(&rx_rpc_stats);
8436 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8446 rx_disablePeerRPCStats(void)
8448 struct rx_peer **peer_ptr, **peer_end;
8452 * Turn off peer statistics and if process stats is also off, turn
8456 rxi_monitor_peerStats = 0;
8457 if (rxi_monitor_processStats == 0) {
8458 rx_enable_stats = 0;
8461 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8462 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8464 struct rx_peer *peer, *next, *prev;
8466 MUTEX_ENTER(&rx_peerHashTable_lock);
8467 MUTEX_ENTER(&rx_rpc_stats);
8468 for (prev = peer = *peer_ptr; peer; peer = next) {
8470 code = MUTEX_TRYENTER(&peer->peer_lock);
8472 rx_interface_stat_p rpc_stat, nrpc_stat;
8475 if (prev == *peer_ptr) {
8486 MUTEX_EXIT(&rx_peerHashTable_lock);
8489 (&peer->rpcStats, rpc_stat, nrpc_stat,
8490 rx_interface_stat)) {
8491 unsigned int num_funcs = 0;
8494 queue_Remove(&rpc_stat->queue_header);
8495 queue_Remove(&rpc_stat->all_peers);
8496 num_funcs = rpc_stat->stats[0].func_total;
8498 sizeof(rx_interface_stat_t) +
8499 rpc_stat->stats[0].func_total *
8500 sizeof(rx_function_entry_v1_t);
8502 rxi_Free(rpc_stat, space);
8503 rxi_rpc_peer_stat_cnt -= num_funcs;
8505 MUTEX_EXIT(&peer->peer_lock);
8507 MUTEX_ENTER(&rx_peerHashTable_lock);
8517 MUTEX_EXIT(&rx_rpc_stats);
8518 MUTEX_EXIT(&rx_peerHashTable_lock);
8523 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8528 * IN clearFlag - flag indicating which stats to clear
8536 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8538 rx_interface_stat_p rpc_stat, nrpc_stat;
8540 MUTEX_ENTER(&rx_rpc_stats);
8542 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8543 unsigned int num_funcs = 0, i;
8544 num_funcs = rpc_stat->stats[0].func_total;
8545 for (i = 0; i < num_funcs; i++) {
8546 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8547 hzero(rpc_stat->stats[i].invocations);
8549 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8550 hzero(rpc_stat->stats[i].bytes_sent);
8552 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8553 hzero(rpc_stat->stats[i].bytes_rcvd);
8555 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8556 rpc_stat->stats[i].queue_time_sum.sec = 0;
8557 rpc_stat->stats[i].queue_time_sum.usec = 0;
8559 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8560 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8561 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8563 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8564 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8565 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8567 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8568 rpc_stat->stats[i].queue_time_max.sec = 0;
8569 rpc_stat->stats[i].queue_time_max.usec = 0;
8571 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8572 rpc_stat->stats[i].execution_time_sum.sec = 0;
8573 rpc_stat->stats[i].execution_time_sum.usec = 0;
8575 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8576 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8577 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8579 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8580 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8581 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8583 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8584 rpc_stat->stats[i].execution_time_max.sec = 0;
8585 rpc_stat->stats[i].execution_time_max.usec = 0;
8590 MUTEX_EXIT(&rx_rpc_stats);
8594 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8599 * IN clearFlag - flag indicating which stats to clear
8607 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8609 rx_interface_stat_p rpc_stat, nrpc_stat;
8611 MUTEX_ENTER(&rx_rpc_stats);
8613 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8614 unsigned int num_funcs = 0, i;
8617 * We have to fix the offset of rpc_stat since we are
8618 * keeping this structure on two rx_queues. The rx_queue
8619 * package assumes that the rx_queue member is the first
8620 * member of the structure. That is, rx_queue assumes that
8621 * any one item is only on one queue at a time. We are
8622 * breaking that assumption and so we have to do a little
8623 * math to fix our pointers.
8626 fix_offset = (char *)rpc_stat;
8627 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8628 rpc_stat = (rx_interface_stat_p) fix_offset;
8630 num_funcs = rpc_stat->stats[0].func_total;
8631 for (i = 0; i < num_funcs; i++) {
8632 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8633 hzero(rpc_stat->stats[i].invocations);
8635 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8636 hzero(rpc_stat->stats[i].bytes_sent);
8638 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8639 hzero(rpc_stat->stats[i].bytes_rcvd);
8641 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8642 rpc_stat->stats[i].queue_time_sum.sec = 0;
8643 rpc_stat->stats[i].queue_time_sum.usec = 0;
8645 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8646 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8647 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8649 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8650 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8651 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8653 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8654 rpc_stat->stats[i].queue_time_max.sec = 0;
8655 rpc_stat->stats[i].queue_time_max.usec = 0;
8657 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8658 rpc_stat->stats[i].execution_time_sum.sec = 0;
8659 rpc_stat->stats[i].execution_time_sum.usec = 0;
8661 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8662 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8663 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8665 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8666 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8667 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8669 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8670 rpc_stat->stats[i].execution_time_max.sec = 0;
8671 rpc_stat->stats[i].execution_time_max.usec = 0;
8676 MUTEX_EXIT(&rx_rpc_stats);
8680 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8681 * is authorized to enable/disable/clear RX statistics.
8683 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8686 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8688 rxi_rxstat_userok = proc;
8692 rx_RxStatUserOk(struct rx_call *call)
8694 if (!rxi_rxstat_userok)
8696 return rxi_rxstat_userok(call);
8701 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8702 * function in the MSVC runtime DLL (msvcrt.dll).
8704 * Note: the system serializes calls to this function.
8707 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8708 DWORD reason, /* reason function is being called */
8709 LPVOID reserved) /* reserved for future use */
8712 case DLL_PROCESS_ATTACH:
8713 /* library is being attached to a process */
8717 case DLL_PROCESS_DETACH:
8724 #endif /* AFS_NT40_ENV */
8727 int rx_DumpCalls(FILE *outputFile, char *cookie)
8729 #ifdef RXDEBUG_PACKET
8730 #ifdef KDUMP_RX_LOCK
8731 struct rx_call_rx_lock *c;
8738 #define RXDPRINTF sprintf
8739 #define RXDPRINTOUT output
8741 #define RXDPRINTF fprintf
8742 #define RXDPRINTOUT outputFile
8745 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8747 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8750 for (c = rx_allCallsp; c; c = c->allNextp) {
8751 u_short rqc, tqc, iovqc;
8752 struct rx_packet *p, *np;
8754 MUTEX_ENTER(&c->lock);
8755 queue_Count(&c->rq, p, np, rx_packet, rqc);
8756 queue_Count(&c->tq, p, np, rx_packet, tqc);
8757 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8759 RXDPRINTF(RXDPRINTOUT, "%s - call=0x%p, id=%u, state=%u, mode=%u, conn=%p, epoch=%u, cid=%u, callNum=%u, connFlags=0x%x, flags=0x%x, "
8760 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8761 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8762 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8763 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8764 #ifdef RX_ENABLE_LOCKS
8767 #ifdef RX_REFCOUNT_CHECK
8768 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
8769 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
8772 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,
8773 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
8774 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
8775 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
8776 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
8777 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
8778 #ifdef RX_ENABLE_LOCKS
8779 , (afs_uint32)c->refCount
8781 #ifdef RX_REFCOUNT_CHECK
8782 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
8785 MUTEX_EXIT(&c->lock);
8788 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8791 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
8793 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8795 #endif /* RXDEBUG_PACKET */