2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
14 #include "afs/param.h"
16 #include <afs/param.h>
21 #include "afs/sysincludes.h"
22 #include "afsincludes.h"
28 #include <net/net_globals.h>
29 #endif /* AFS_OSF_ENV */
30 #ifdef AFS_LINUX20_ENV
33 #include "netinet/in.h"
35 #include "inet/common.h"
37 #include "inet/ip_ire.h"
39 #include "afs/afs_args.h"
40 #include "afs/afs_osi.h"
41 #ifdef RX_KERNEL_TRACE
42 #include "rx_kcommon.h"
44 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
48 #undef RXDEBUG /* turn off debugging */
50 #if defined(AFS_SGI_ENV)
51 #include "sys/debug.h"
59 #endif /* AFS_OSF_ENV */
61 #include "afs/sysincludes.h"
62 #include "afsincludes.h"
65 #include "rx_kmutex.h"
66 #include "rx_kernel.h"
70 #include "rx_globals.h"
72 #include "rx_atomic.h"
73 #include "rx_internal.h"
75 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
76 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
77 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
79 extern afs_int32 afs_termState;
81 #include "sys/lockl.h"
82 #include "sys/lock_def.h"
83 #endif /* AFS_AIX41_ENV */
84 # include "afs/rxgen_consts.h"
86 # include <sys/types.h>
96 # include <afs/afsutil.h>
97 # include <WINNT\afsreg.h>
99 # include <sys/socket.h>
100 # include <sys/file.h>
102 # include <sys/stat.h>
103 # include <netinet/in.h>
104 # include <sys/time.h>
107 # include "rx_user.h"
108 # include "rx_clock.h"
109 # include "rx_queue.h"
110 # include "rx_atomic.h"
111 # include "rx_globals.h"
112 # include "rx_trace.h"
113 # include "rx_internal.h"
114 # include "rx_stats.h"
115 # include <afs/rxgen_consts.h>
119 #ifdef AFS_PTHREAD_ENV
121 int (*registerProgram) (pid_t, char *) = 0;
122 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
125 int (*registerProgram) (PROCESS, char *) = 0;
126 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
130 /* Local static routines */
131 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
132 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct clock *,
133 struct rx_peer *, struct clock *);
135 #ifdef RX_ENABLE_LOCKS
136 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
139 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
141 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
142 rx_atomic_t rxi_start_in_error;
144 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
147 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
148 * currently allocated within rx. This number is used to allocate the
149 * memory required to return the statistics when queried.
150 * Protected by the rx_rpc_stats mutex.
153 static unsigned int rxi_rpc_peer_stat_cnt;
156 * rxi_rpc_process_stat_cnt counts the total number of local process stat
157 * structures currently allocated within rx. The number is used to allocate
158 * the memory required to return the statistics when queried.
159 * Protected by the rx_rpc_stats mutex.
162 static unsigned int rxi_rpc_process_stat_cnt;
164 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
165 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
167 #if !defined(offsetof)
168 #include <stddef.h> /* for definition of offsetof() */
171 #ifdef RX_ENABLE_LOCKS
172 afs_kmutex_t rx_atomic_mutex;
175 #ifdef AFS_PTHREAD_ENV
179 * Use procedural initialization of mutexes/condition variables
183 extern afs_kmutex_t rx_quota_mutex;
184 extern afs_kmutex_t rx_pthread_mutex;
185 extern afs_kmutex_t rx_packets_mutex;
186 extern afs_kmutex_t rx_refcnt_mutex;
187 extern afs_kmutex_t des_init_mutex;
188 extern afs_kmutex_t des_random_mutex;
189 extern afs_kmutex_t rx_clock_mutex;
190 extern afs_kmutex_t rxi_connCacheMutex;
191 extern afs_kmutex_t rx_event_mutex;
192 extern afs_kmutex_t osi_malloc_mutex;
193 extern afs_kmutex_t event_handler_mutex;
194 extern afs_kmutex_t listener_mutex;
195 extern afs_kmutex_t rx_if_init_mutex;
196 extern afs_kmutex_t rx_if_mutex;
197 extern afs_kmutex_t rxkad_client_uid_mutex;
198 extern afs_kmutex_t rxkad_random_mutex;
200 extern afs_kcondvar_t rx_event_handler_cond;
201 extern afs_kcondvar_t rx_listener_cond;
203 static afs_kmutex_t epoch_mutex;
204 static afs_kmutex_t rx_init_mutex;
205 static afs_kmutex_t rx_debug_mutex;
206 static afs_kmutex_t rx_rpc_stats;
209 rxi_InitPthread(void)
211 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
212 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
213 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
214 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
215 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
216 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
217 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
218 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
219 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
220 MUTEX_INIT(&rx_event_mutex, "event", MUTEX_DEFAULT, 0);
221 MUTEX_INIT(&des_init_mutex, "des", MUTEX_DEFAULT, 0);
222 MUTEX_INIT(&des_random_mutex, "random", MUTEX_DEFAULT, 0);
223 MUTEX_INIT(&osi_malloc_mutex, "malloc", MUTEX_DEFAULT, 0);
224 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
225 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
226 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
227 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
228 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
229 MUTEX_INIT(&rxkad_client_uid_mutex, "uid", MUTEX_DEFAULT, 0);
230 MUTEX_INIT(&rxkad_random_mutex, "rxkad random", MUTEX_DEFAULT, 0);
231 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
233 assert(pthread_cond_init
234 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
235 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
237 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
238 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
240 rxkad_global_stats_init();
242 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
243 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
244 #ifdef RX_ENABLE_LOCKS
247 #endif /* RX_LOCKS_DB */
248 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
249 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
251 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
253 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
255 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
257 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
258 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
259 #endif /* RX_ENABLE_LOCKS */
262 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
263 #define INIT_PTHREAD_LOCKS \
264 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
266 * The rx_stats_mutex mutex protects the following global variables:
267 * rxi_lowConnRefCount
268 * rxi_lowPeerRefCount
277 * The rx_quota_mutex mutex protects the following global variables:
285 * The rx_freePktQ_lock protects the following global variables:
290 * The rx_packets_mutex mutex protects the following global variables:
298 * The rx_pthread_mutex mutex protects the following global variables:
299 * rxi_fcfs_thread_num
302 #define INIT_PTHREAD_LOCKS
306 /* Variables for handling the minProcs implementation. availProcs gives the
307 * number of threads available in the pool at this moment (not counting dudes
308 * executing right now). totalMin gives the total number of procs required
309 * for handling all minProcs requests. minDeficit is a dynamic variable
310 * tracking the # of procs required to satisfy all of the remaining minProcs
312 * For fine grain locking to work, the quota check and the reservation of
313 * a server thread has to come while rxi_availProcs and rxi_minDeficit
314 * are locked. To this end, the code has been modified under #ifdef
315 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
316 * same time. A new function, ReturnToServerPool() returns the allocation.
318 * A call can be on several queue's (but only one at a time). When
319 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
320 * that no one else is touching the queue. To this end, we store the address
321 * of the queue lock in the call structure (under the call lock) when we
322 * put the call on a queue, and we clear the call_queue_lock when the
323 * call is removed from a queue (once the call lock has been obtained).
324 * This allows rxi_ResetCall to safely synchronize with others wishing
325 * to manipulate the queue.
328 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
329 static afs_kmutex_t rx_rpc_stats;
330 void rxi_StartUnlocked(struct rxevent *event, void *call,
331 void *arg1, int istack);
334 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
335 ** pretty good that the next packet coming in is from the same connection
336 ** as the last packet, since we're send multiple packets in a transmit window.
338 struct rx_connection *rxLastConn = 0;
340 #ifdef RX_ENABLE_LOCKS
341 /* The locking hierarchy for rx fine grain locking is composed of these
344 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
345 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
346 * call->lock - locks call data fields.
347 * These are independent of each other:
348 * rx_freeCallQueue_lock
353 * serverQueueEntry->lock
354 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
356 * peer->lock - locks peer data fields.
357 * conn_data_lock - that more than one thread is not updating a conn data
358 * field at the same time.
369 * Do we need a lock to protect the peer field in the conn structure?
370 * conn->peer was previously a constant for all intents and so has no
371 * lock protecting this field. The multihomed client delta introduced
372 * a RX code change : change the peer field in the connection structure
373 * to that remote interface from which the last packet for this
374 * connection was sent out. This may become an issue if further changes
377 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
378 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
380 /* rxdb_fileID is used to identify the lock location, along with line#. */
381 static int rxdb_fileID = RXDB_FILE_RX;
382 #endif /* RX_LOCKS_DB */
383 #else /* RX_ENABLE_LOCKS */
384 #define SET_CALL_QUEUE_LOCK(C, L)
385 #define CLEAR_CALL_QUEUE_LOCK(C)
386 #endif /* RX_ENABLE_LOCKS */
387 struct rx_serverQueueEntry *rx_waitForPacket = 0;
388 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
390 /* ------------Exported Interfaces------------- */
392 /* This function allows rxkad to set the epoch to a suitably random number
393 * which rx_NewConnection will use in the future. The principle purpose is to
394 * get rxnull connections to use the same epoch as the rxkad connections do, at
395 * least once the first rxkad connection is established. This is important now
396 * that the host/port addresses aren't used in FindConnection: the uniqueness
397 * of epoch/cid matters and the start time won't do. */
399 #ifdef AFS_PTHREAD_ENV
401 * This mutex protects the following global variables:
405 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
406 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
410 #endif /* AFS_PTHREAD_ENV */
413 rx_SetEpoch(afs_uint32 epoch)
420 /* Initialize rx. A port number may be mentioned, in which case this
421 * becomes the default port number for any service installed later.
422 * If 0 is provided for the port number, a random port will be chosen
423 * by the kernel. Whether this will ever overlap anything in
424 * /etc/services is anybody's guess... Returns 0 on success, -1 on
429 int rxinit_status = 1;
430 #ifdef AFS_PTHREAD_ENV
432 * This mutex protects the following global variables:
436 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
437 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
440 #define UNLOCK_RX_INIT
444 rx_InitHost(u_int host, u_int port)
451 char *htable, *ptable;
458 if (rxinit_status == 0) {
459 tmp_status = rxinit_status;
461 return tmp_status; /* Already started; return previous error code. */
467 if (afs_winsockInit() < 0)
473 * Initialize anything necessary to provide a non-premptive threading
476 rxi_InitializeThreadSupport();
479 /* Allocate and initialize a socket for client and perhaps server
482 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
483 if (rx_socket == OSI_NULLSOCKET) {
487 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
490 #endif /* RX_LOCKS_DB */
491 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
492 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
493 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
494 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
495 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
496 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
497 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
498 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
499 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
501 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
503 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
505 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
507 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
508 #if defined(AFS_HPUX110_ENV)
510 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
511 #endif /* AFS_HPUX110_ENV */
512 #endif /* RX_ENABLE_LOCKS && KERNEL */
515 rx_connDeadTime = 12;
516 rx_tranquil = 0; /* reset flag */
517 rxi_ResetStatistics();
519 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
520 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
521 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
522 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
523 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
524 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
526 /* Malloc up a bunch of packets & buffers */
528 queue_Init(&rx_freePacketQueue);
529 rxi_NeedMorePackets = FALSE;
530 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
532 /* enforce a minimum number of allocated packets */
533 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
534 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
536 /* allocate the initial free packet pool */
537 #ifdef RX_ENABLE_TSFPQ
538 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
539 #else /* RX_ENABLE_TSFPQ */
540 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
541 #endif /* RX_ENABLE_TSFPQ */
548 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
549 tv.tv_sec = clock_now.sec;
550 tv.tv_usec = clock_now.usec;
551 srand((unsigned int)tv.tv_usec);
558 #if defined(KERNEL) && !defined(UKERNEL)
559 /* Really, this should never happen in a real kernel */
562 struct sockaddr_in addr;
564 int addrlen = sizeof(addr);
566 socklen_t addrlen = sizeof(addr);
568 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
572 rx_port = addr.sin_port;
575 rx_stats.minRtt.sec = 9999999;
577 rx_SetEpoch(tv.tv_sec | 0x80000000);
579 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
580 * will provide a randomer value. */
582 MUTEX_ENTER(&rx_quota_mutex);
583 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
584 MUTEX_EXIT(&rx_quota_mutex);
585 /* *Slightly* random start time for the cid. This is just to help
586 * out with the hashing function at the peer */
587 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
588 rx_connHashTable = (struct rx_connection **)htable;
589 rx_peerHashTable = (struct rx_peer **)ptable;
591 rx_lastAckDelay.sec = 0;
592 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
593 rx_hardAckDelay.sec = 0;
594 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
595 rx_softAckDelay.sec = 0;
596 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
598 rxevent_Init(20, rxi_ReScheduleEvents);
600 /* Initialize various global queues */
601 queue_Init(&rx_idleServerQueue);
602 queue_Init(&rx_incomingCallQueue);
603 queue_Init(&rx_freeCallQueue);
605 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
606 /* Initialize our list of usable IP addresses. */
610 /* Start listener process (exact function is dependent on the
611 * implementation environment--kernel or user space) */
615 tmp_status = rxinit_status = 0;
623 return rx_InitHost(htonl(INADDR_ANY), port);
626 /* called with unincremented nRequestsRunning to see if it is OK to start
627 * a new thread in this service. Could be "no" for two reasons: over the
628 * max quota, or would prevent others from reaching their min quota.
630 #ifdef RX_ENABLE_LOCKS
631 /* This verion of QuotaOK reserves quota if it's ok while the
632 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
635 QuotaOK(struct rx_service *aservice)
637 /* check if over max quota */
638 if (aservice->nRequestsRunning >= aservice->maxProcs) {
642 /* under min quota, we're OK */
643 /* otherwise, can use only if there are enough to allow everyone
644 * to go to their min quota after this guy starts.
647 MUTEX_ENTER(&rx_quota_mutex);
648 if ((aservice->nRequestsRunning < aservice->minProcs)
649 || (rxi_availProcs > rxi_minDeficit)) {
650 aservice->nRequestsRunning++;
651 /* just started call in minProcs pool, need fewer to maintain
653 if (aservice->nRequestsRunning <= aservice->minProcs)
656 MUTEX_EXIT(&rx_quota_mutex);
659 MUTEX_EXIT(&rx_quota_mutex);
665 ReturnToServerPool(struct rx_service *aservice)
667 aservice->nRequestsRunning--;
668 MUTEX_ENTER(&rx_quota_mutex);
669 if (aservice->nRequestsRunning < aservice->minProcs)
672 MUTEX_EXIT(&rx_quota_mutex);
675 #else /* RX_ENABLE_LOCKS */
677 QuotaOK(struct rx_service *aservice)
680 /* under min quota, we're OK */
681 if (aservice->nRequestsRunning < aservice->minProcs)
684 /* check if over max quota */
685 if (aservice->nRequestsRunning >= aservice->maxProcs)
688 /* otherwise, can use only if there are enough to allow everyone
689 * to go to their min quota after this guy starts.
691 MUTEX_ENTER(&rx_quota_mutex);
692 if (rxi_availProcs > rxi_minDeficit)
694 MUTEX_EXIT(&rx_quota_mutex);
697 #endif /* RX_ENABLE_LOCKS */
700 /* Called by rx_StartServer to start up lwp's to service calls.
701 NExistingProcs gives the number of procs already existing, and which
702 therefore needn't be created. */
704 rxi_StartServerProcs(int nExistingProcs)
706 struct rx_service *service;
711 /* For each service, reserve N processes, where N is the "minimum"
712 * number of processes that MUST be able to execute a request in parallel,
713 * at any time, for that process. Also compute the maximum difference
714 * between any service's maximum number of processes that can run
715 * (i.e. the maximum number that ever will be run, and a guarantee
716 * that this number will run if other services aren't running), and its
717 * minimum number. The result is the extra number of processes that
718 * we need in order to provide the latter guarantee */
719 for (i = 0; i < RX_MAX_SERVICES; i++) {
721 service = rx_services[i];
722 if (service == (struct rx_service *)0)
724 nProcs += service->minProcs;
725 diff = service->maxProcs - service->minProcs;
729 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
730 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
731 for (i = 0; i < nProcs; i++) {
732 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
738 /* This routine is only required on Windows */
740 rx_StartClientThread(void)
742 #ifdef AFS_PTHREAD_ENV
744 pid = pthread_self();
745 #endif /* AFS_PTHREAD_ENV */
747 #endif /* AFS_NT40_ENV */
749 /* This routine must be called if any services are exported. If the
750 * donateMe flag is set, the calling process is donated to the server
753 rx_StartServer(int donateMe)
755 struct rx_service *service;
761 /* Start server processes, if necessary (exact function is dependent
762 * on the implementation environment--kernel or user space). DonateMe
763 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
764 * case, one less new proc will be created rx_StartServerProcs.
766 rxi_StartServerProcs(donateMe);
768 /* count up the # of threads in minProcs, and add set the min deficit to
769 * be that value, too.
771 for (i = 0; i < RX_MAX_SERVICES; i++) {
772 service = rx_services[i];
773 if (service == (struct rx_service *)0)
775 MUTEX_ENTER(&rx_quota_mutex);
776 rxi_totalMin += service->minProcs;
777 /* below works even if a thread is running, since minDeficit would
778 * still have been decremented and later re-incremented.
780 rxi_minDeficit += service->minProcs;
781 MUTEX_EXIT(&rx_quota_mutex);
784 /* Turn on reaping of idle server connections */
785 rxi_ReapConnections(NULL, NULL, NULL);
794 #ifdef AFS_PTHREAD_ENV
796 pid = afs_pointer_to_int(pthread_self());
797 #else /* AFS_PTHREAD_ENV */
799 LWP_CurrentProcess(&pid);
800 #endif /* AFS_PTHREAD_ENV */
802 sprintf(name, "srv_%d", ++nProcs);
804 (*registerProgram) (pid, name);
806 #endif /* AFS_NT40_ENV */
807 rx_ServerProc(NULL); /* Never returns */
809 #ifdef RX_ENABLE_TSFPQ
810 /* no use leaving packets around in this thread's local queue if
811 * it isn't getting donated to the server thread pool.
813 rxi_FlushLocalPacketsTSFPQ();
814 #endif /* RX_ENABLE_TSFPQ */
818 /* Create a new client connection to the specified service, using the
819 * specified security object to implement the security model for this
821 struct rx_connection *
822 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
823 struct rx_securityClass *securityObject,
824 int serviceSecurityIndex)
828 struct rx_connection *conn;
833 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
834 "serviceSecurityIndex %d)\n",
835 ntohl(shost), ntohs(sport), sservice, securityObject,
836 serviceSecurityIndex));
838 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
839 * the case of kmem_alloc? */
840 conn = rxi_AllocConnection();
841 #ifdef RX_ENABLE_LOCKS
842 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
843 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
844 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
847 MUTEX_ENTER(&rx_connHashTable_lock);
848 cid = (rx_nextCid += RX_MAXCALLS);
849 conn->type = RX_CLIENT_CONNECTION;
851 conn->epoch = rx_epoch;
852 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
853 conn->serviceId = sservice;
854 conn->securityObject = securityObject;
855 conn->securityData = (void *) 0;
856 conn->securityIndex = serviceSecurityIndex;
857 rx_SetConnDeadTime(conn, rx_connDeadTime);
858 rx_SetConnSecondsUntilNatPing(conn, 0);
859 conn->ackRate = RX_FAST_ACK_RATE;
861 conn->specific = NULL;
862 conn->challengeEvent = NULL;
863 conn->delayedAbortEvent = NULL;
864 conn->abortCount = 0;
866 for (i = 0; i < RX_MAXCALLS; i++) {
867 conn->twind[i] = rx_initSendWindow;
868 conn->rwind[i] = rx_initReceiveWindow;
871 RXS_NewConnection(securityObject, conn);
873 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
875 conn->refCount++; /* no lock required since only this thread knows... */
876 conn->next = rx_connHashTable[hashindex];
877 rx_connHashTable[hashindex] = conn;
879 rx_atomic_inc(&rx_stats.nClientConns);
880 MUTEX_EXIT(&rx_connHashTable_lock);
886 * Ensure a connection's timeout values are valid.
888 * @param[in] conn The connection to check
890 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
891 * unless idleDeadTime and/or hardDeadTime are not set
895 rxi_CheckConnTimeouts(struct rx_connection *conn)
897 /* a connection's timeouts must have the relationship
898 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
899 * total loss of network to a peer may cause an idle timeout instead of a
900 * dead timeout, simply because the idle timeout gets hit first. Also set
901 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
902 /* this logic is slightly complicated by the fact that
903 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
905 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
906 if (conn->idleDeadTime) {
907 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
909 if (conn->hardDeadTime) {
910 if (conn->idleDeadTime) {
911 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
913 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
919 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
921 /* The idea is to set the dead time to a value that allows several
922 * keepalives to be dropped without timing out the connection. */
923 conn->secondsUntilDead = seconds;
924 rxi_CheckConnTimeouts(conn);
925 conn->secondsUntilPing = conn->secondsUntilDead / 6;
929 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
931 conn->hardDeadTime = seconds;
932 rxi_CheckConnTimeouts(conn);
936 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
938 conn->idleDeadTime = seconds;
939 rxi_CheckConnTimeouts(conn);
942 int rxi_lowPeerRefCount = 0;
943 int rxi_lowConnRefCount = 0;
946 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
947 * NOTE: must not be called with rx_connHashTable_lock held.
950 rxi_CleanupConnection(struct rx_connection *conn)
952 /* Notify the service exporter, if requested, that this connection
953 * is being destroyed */
954 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
955 (*conn->service->destroyConnProc) (conn);
957 /* Notify the security module that this connection is being destroyed */
958 RXS_DestroyConnection(conn->securityObject, conn);
960 /* If this is the last connection using the rx_peer struct, set its
961 * idle time to now. rxi_ReapConnections will reap it if it's still
962 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
964 MUTEX_ENTER(&rx_peerHashTable_lock);
965 if (conn->peer->refCount < 2) {
966 conn->peer->idleWhen = clock_Sec();
967 if (conn->peer->refCount < 1) {
968 conn->peer->refCount = 1;
969 if (rx_stats_active) {
970 MUTEX_ENTER(&rx_stats_mutex);
971 rxi_lowPeerRefCount++;
972 MUTEX_EXIT(&rx_stats_mutex);
976 conn->peer->refCount--;
977 MUTEX_EXIT(&rx_peerHashTable_lock);
981 if (conn->type == RX_SERVER_CONNECTION)
982 rx_atomic_dec(&rx_stats.nServerConns);
984 rx_atomic_dec(&rx_stats.nClientConns);
987 if (conn->specific) {
989 for (i = 0; i < conn->nSpecific; i++) {
990 if (conn->specific[i] && rxi_keyCreate_destructor[i])
991 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
992 conn->specific[i] = NULL;
994 free(conn->specific);
996 conn->specific = NULL;
1000 MUTEX_DESTROY(&conn->conn_call_lock);
1001 MUTEX_DESTROY(&conn->conn_data_lock);
1002 CV_DESTROY(&conn->conn_call_cv);
1004 rxi_FreeConnection(conn);
1007 /* Destroy the specified connection */
1009 rxi_DestroyConnection(struct rx_connection *conn)
1011 MUTEX_ENTER(&rx_connHashTable_lock);
1012 rxi_DestroyConnectionNoLock(conn);
1013 /* conn should be at the head of the cleanup list */
1014 if (conn == rx_connCleanup_list) {
1015 rx_connCleanup_list = rx_connCleanup_list->next;
1016 MUTEX_EXIT(&rx_connHashTable_lock);
1017 rxi_CleanupConnection(conn);
1019 #ifdef RX_ENABLE_LOCKS
1021 MUTEX_EXIT(&rx_connHashTable_lock);
1023 #endif /* RX_ENABLE_LOCKS */
1027 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1029 struct rx_connection **conn_ptr;
1031 struct rx_packet *packet;
1038 MUTEX_ENTER(&conn->conn_data_lock);
1039 MUTEX_ENTER(&rx_refcnt_mutex);
1040 if (conn->refCount > 0)
1043 if (rx_stats_active) {
1044 MUTEX_ENTER(&rx_stats_mutex);
1045 rxi_lowConnRefCount++;
1046 MUTEX_EXIT(&rx_stats_mutex);
1050 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1051 /* Busy; wait till the last guy before proceeding */
1052 MUTEX_EXIT(&rx_refcnt_mutex);
1053 MUTEX_EXIT(&conn->conn_data_lock);
1058 /* If the client previously called rx_NewCall, but it is still
1059 * waiting, treat this as a running call, and wait to destroy the
1060 * connection later when the call completes. */
1061 if ((conn->type == RX_CLIENT_CONNECTION)
1062 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1063 conn->flags |= RX_CONN_DESTROY_ME;
1064 MUTEX_EXIT(&conn->conn_data_lock);
1068 MUTEX_EXIT(&rx_refcnt_mutex);
1069 MUTEX_EXIT(&conn->conn_data_lock);
1071 /* Check for extant references to this connection */
1072 for (i = 0; i < RX_MAXCALLS; i++) {
1073 struct rx_call *call = conn->call[i];
1076 if (conn->type == RX_CLIENT_CONNECTION) {
1077 MUTEX_ENTER(&call->lock);
1078 if (call->delayedAckEvent) {
1079 /* Push the final acknowledgment out now--there
1080 * won't be a subsequent call to acknowledge the
1081 * last reply packets */
1082 rxevent_Cancel(call->delayedAckEvent, call,
1083 RX_CALL_REFCOUNT_DELAY);
1084 if (call->state == RX_STATE_PRECALL
1085 || call->state == RX_STATE_ACTIVE) {
1086 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1088 rxi_AckAll(NULL, call, 0);
1091 MUTEX_EXIT(&call->lock);
1095 #ifdef RX_ENABLE_LOCKS
1097 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1098 MUTEX_EXIT(&conn->conn_data_lock);
1100 /* Someone is accessing a packet right now. */
1104 #endif /* RX_ENABLE_LOCKS */
1107 /* Don't destroy the connection if there are any call
1108 * structures still in use */
1109 MUTEX_ENTER(&conn->conn_data_lock);
1110 conn->flags |= RX_CONN_DESTROY_ME;
1111 MUTEX_EXIT(&conn->conn_data_lock);
1116 if (conn->natKeepAliveEvent) {
1117 rxi_NatKeepAliveOff(conn);
1120 if (conn->delayedAbortEvent) {
1121 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1122 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1124 MUTEX_ENTER(&conn->conn_data_lock);
1125 rxi_SendConnectionAbort(conn, packet, 0, 1);
1126 MUTEX_EXIT(&conn->conn_data_lock);
1127 rxi_FreePacket(packet);
1131 /* Remove from connection hash table before proceeding */
1133 &rx_connHashTable[CONN_HASH
1134 (peer->host, peer->port, conn->cid, conn->epoch,
1136 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1137 if (*conn_ptr == conn) {
1138 *conn_ptr = conn->next;
1142 /* if the conn that we are destroying was the last connection, then we
1143 * clear rxLastConn as well */
1144 if (rxLastConn == conn)
1147 /* Make sure the connection is completely reset before deleting it. */
1148 /* get rid of pending events that could zap us later */
1149 if (conn->challengeEvent)
1150 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1151 if (conn->checkReachEvent)
1152 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1153 if (conn->natKeepAliveEvent)
1154 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
1156 /* Add the connection to the list of destroyed connections that
1157 * need to be cleaned up. This is necessary to avoid deadlocks
1158 * in the routines we call to inform others that this connection is
1159 * being destroyed. */
1160 conn->next = rx_connCleanup_list;
1161 rx_connCleanup_list = conn;
1164 /* Externally available version */
1166 rx_DestroyConnection(struct rx_connection *conn)
1171 rxi_DestroyConnection(conn);
1176 rx_GetConnection(struct rx_connection *conn)
1181 MUTEX_ENTER(&rx_refcnt_mutex);
1183 MUTEX_EXIT(&rx_refcnt_mutex);
1187 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1188 /* Wait for the transmit queue to no longer be busy.
1189 * requires the call->lock to be held */
1191 rxi_WaitforTQBusy(struct rx_call *call) {
1192 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1193 call->flags |= RX_CALL_TQ_WAIT;
1195 #ifdef RX_ENABLE_LOCKS
1196 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1197 CV_WAIT(&call->cv_tq, &call->lock);
1198 #else /* RX_ENABLE_LOCKS */
1199 osi_rxSleep(&call->tq);
1200 #endif /* RX_ENABLE_LOCKS */
1202 if (call->tqWaiters == 0) {
1203 call->flags &= ~RX_CALL_TQ_WAIT;
1209 /* Start a new rx remote procedure call, on the specified connection.
1210 * If wait is set to 1, wait for a free call channel; otherwise return
1211 * 0. Maxtime gives the maximum number of seconds this call may take,
1212 * after rx_NewCall returns. After this time interval, a call to any
1213 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1214 * For fine grain locking, we hold the conn_call_lock in order to
1215 * to ensure that we don't get signalle after we found a call in an active
1216 * state and before we go to sleep.
1219 rx_NewCall(struct rx_connection *conn)
1222 struct rx_call *call;
1223 struct clock queueTime;
1227 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1230 clock_GetTime(&queueTime);
1232 * Check if there are others waiting for a new call.
1233 * If so, let them go first to avoid starving them.
1234 * This is a fairly simple scheme, and might not be
1235 * a complete solution for large numbers of waiters.
1237 * makeCallWaiters keeps track of the number of
1238 * threads waiting to make calls and the
1239 * RX_CONN_MAKECALL_WAITING flag bit is used to
1240 * indicate that there are indeed calls waiting.
1241 * The flag is set when the waiter is incremented.
1242 * It is only cleared when makeCallWaiters is 0.
1243 * This prevents us from accidently destroying the
1244 * connection while it is potentially about to be used.
1246 MUTEX_ENTER(&conn->conn_call_lock);
1247 MUTEX_ENTER(&conn->conn_data_lock);
1248 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1249 conn->flags |= RX_CONN_MAKECALL_WAITING;
1250 conn->makeCallWaiters++;
1251 MUTEX_EXIT(&conn->conn_data_lock);
1253 #ifdef RX_ENABLE_LOCKS
1254 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1258 MUTEX_ENTER(&conn->conn_data_lock);
1259 conn->makeCallWaiters--;
1260 if (conn->makeCallWaiters == 0)
1261 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1264 /* We are now the active thread in rx_NewCall */
1265 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1266 MUTEX_EXIT(&conn->conn_data_lock);
1271 for (i = 0; i < RX_MAXCALLS; i++) {
1272 call = conn->call[i];
1274 if (call->state == RX_STATE_DALLY) {
1275 MUTEX_ENTER(&call->lock);
1276 if (call->state == RX_STATE_DALLY) {
1278 * We are setting the state to RX_STATE_RESET to
1279 * ensure that no one else will attempt to use this
1280 * call once we drop the conn->conn_call_lock and
1281 * call->lock. We must drop the conn->conn_call_lock
1282 * before calling rxi_ResetCall because the process
1283 * of clearing the transmit queue can block for an
1284 * extended period of time. If we block while holding
1285 * the conn->conn_call_lock, then all rx_EndCall
1286 * processing will block as well. This has a detrimental
1287 * effect on overall system performance.
1289 call->state = RX_STATE_RESET;
1290 MUTEX_EXIT(&conn->conn_call_lock);
1291 MUTEX_ENTER(&rx_refcnt_mutex);
1292 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1293 MUTEX_EXIT(&rx_refcnt_mutex);
1294 rxi_ResetCall(call, 0);
1295 (*call->callNumber)++;
1296 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1300 * If we failed to be able to safely obtain the
1301 * conn->conn_call_lock we will have to drop the
1302 * call->lock to avoid a deadlock. When the call->lock
1303 * is released the state of the call can change. If it
1304 * is no longer RX_STATE_RESET then some other thread is
1307 MUTEX_EXIT(&call->lock);
1308 MUTEX_ENTER(&conn->conn_call_lock);
1309 MUTEX_ENTER(&call->lock);
1311 if (call->state == RX_STATE_RESET)
1315 * If we get here it means that after dropping
1316 * the conn->conn_call_lock and call->lock that
1317 * the call is no longer ours. If we can't find
1318 * a free call in the remaining slots we should
1319 * not go immediately to RX_CONN_MAKECALL_WAITING
1320 * because by dropping the conn->conn_call_lock
1321 * we have given up synchronization with rx_EndCall.
1322 * Instead, cycle through one more time to see if
1323 * we can find a call that can call our own.
1325 MUTEX_ENTER(&rx_refcnt_mutex);
1326 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1327 MUTEX_EXIT(&rx_refcnt_mutex);
1330 MUTEX_EXIT(&call->lock);
1333 /* rxi_NewCall returns with mutex locked */
1334 call = rxi_NewCall(conn, i);
1335 MUTEX_ENTER(&rx_refcnt_mutex);
1336 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1337 MUTEX_EXIT(&rx_refcnt_mutex);
1341 if (i < RX_MAXCALLS) {
1347 MUTEX_ENTER(&conn->conn_data_lock);
1348 conn->flags |= RX_CONN_MAKECALL_WAITING;
1349 conn->makeCallWaiters++;
1350 MUTEX_EXIT(&conn->conn_data_lock);
1352 #ifdef RX_ENABLE_LOCKS
1353 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1357 MUTEX_ENTER(&conn->conn_data_lock);
1358 conn->makeCallWaiters--;
1359 if (conn->makeCallWaiters == 0)
1360 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1361 MUTEX_EXIT(&conn->conn_data_lock);
1363 /* Client is initially in send mode */
1364 call->state = RX_STATE_ACTIVE;
1365 call->error = conn->error;
1367 call->mode = RX_MODE_ERROR;
1369 call->mode = RX_MODE_SENDING;
1371 /* remember start time for call in case we have hard dead time limit */
1372 call->queueTime = queueTime;
1373 clock_GetTime(&call->startTime);
1374 hzero(call->bytesSent);
1375 hzero(call->bytesRcvd);
1377 /* Turn on busy protocol. */
1378 rxi_KeepAliveOn(call);
1380 /* Attempt MTU discovery */
1381 rxi_GrowMTUOn(call);
1384 * We are no longer the active thread in rx_NewCall
1386 MUTEX_ENTER(&conn->conn_data_lock);
1387 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1388 MUTEX_EXIT(&conn->conn_data_lock);
1391 * Wake up anyone else who might be giving us a chance to
1392 * run (see code above that avoids resource starvation).
1394 #ifdef RX_ENABLE_LOCKS
1395 CV_BROADCAST(&conn->conn_call_cv);
1399 MUTEX_EXIT(&conn->conn_call_lock);
1401 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1402 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1403 osi_Panic("rx_NewCall call about to be used without an empty tq");
1405 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1407 MUTEX_EXIT(&call->lock);
1410 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1415 rxi_HasActiveCalls(struct rx_connection *aconn)
1418 struct rx_call *tcall;
1422 for (i = 0; i < RX_MAXCALLS; i++) {
1423 if ((tcall = aconn->call[i])) {
1424 if ((tcall->state == RX_STATE_ACTIVE)
1425 || (tcall->state == RX_STATE_PRECALL)) {
1436 rxi_GetCallNumberVector(struct rx_connection *aconn,
1437 afs_int32 * aint32s)
1440 struct rx_call *tcall;
1444 for (i = 0; i < RX_MAXCALLS; i++) {
1445 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1446 aint32s[i] = aconn->callNumber[i] + 1;
1448 aint32s[i] = aconn->callNumber[i];
1455 rxi_SetCallNumberVector(struct rx_connection *aconn,
1456 afs_int32 * aint32s)
1459 struct rx_call *tcall;
1463 for (i = 0; i < RX_MAXCALLS; i++) {
1464 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1465 aconn->callNumber[i] = aint32s[i] - 1;
1467 aconn->callNumber[i] = aint32s[i];
1473 /* Advertise a new service. A service is named locally by a UDP port
1474 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1477 char *serviceName; Name for identification purposes (e.g. the
1478 service name might be used for probing for
1481 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1482 char *serviceName, struct rx_securityClass **securityObjects,
1483 int nSecurityObjects,
1484 afs_int32(*serviceProc) (struct rx_call * acall))
1486 osi_socket socket = OSI_NULLSOCKET;
1487 struct rx_service *tservice;
1493 if (serviceId == 0) {
1495 "rx_NewService: service id for service %s is not non-zero.\n",
1502 "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",
1510 tservice = rxi_AllocService();
1513 #ifdef RX_ENABLE_LOCKS
1514 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1517 for (i = 0; i < RX_MAX_SERVICES; i++) {
1518 struct rx_service *service = rx_services[i];
1520 if (port == service->servicePort && host == service->serviceHost) {
1521 if (service->serviceId == serviceId) {
1522 /* The identical service has already been
1523 * installed; if the caller was intending to
1524 * change the security classes used by this
1525 * service, he/she loses. */
1527 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1528 serviceName, serviceId, service->serviceName);
1530 rxi_FreeService(tservice);
1533 /* Different service, same port: re-use the socket
1534 * which is bound to the same port */
1535 socket = service->socket;
1538 if (socket == OSI_NULLSOCKET) {
1539 /* If we don't already have a socket (from another
1540 * service on same port) get a new one */
1541 socket = rxi_GetHostUDPSocket(host, port);
1542 if (socket == OSI_NULLSOCKET) {
1544 rxi_FreeService(tservice);
1549 service->socket = socket;
1550 service->serviceHost = host;
1551 service->servicePort = port;
1552 service->serviceId = serviceId;
1553 service->serviceName = serviceName;
1554 service->nSecurityObjects = nSecurityObjects;
1555 service->securityObjects = securityObjects;
1556 service->minProcs = 0;
1557 service->maxProcs = 1;
1558 service->idleDeadTime = 60;
1559 service->idleDeadErr = 0;
1560 service->connDeadTime = rx_connDeadTime;
1561 service->executeRequestProc = serviceProc;
1562 service->checkReach = 0;
1563 service->nSpecific = 0;
1564 service->specific = NULL;
1565 rx_services[i] = service; /* not visible until now */
1571 rxi_FreeService(tservice);
1572 (osi_Msg "rx_NewService: cannot support > %d services\n",
1577 /* Set configuration options for all of a service's security objects */
1580 rx_SetSecurityConfiguration(struct rx_service *service,
1581 rx_securityConfigVariables type,
1585 for (i = 0; i<service->nSecurityObjects; i++) {
1586 if (service->securityObjects[i]) {
1587 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1595 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1596 struct rx_securityClass **securityObjects, int nSecurityObjects,
1597 afs_int32(*serviceProc) (struct rx_call * acall))
1599 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1602 /* Generic request processing loop. This routine should be called
1603 * by the implementation dependent rx_ServerProc. If socketp is
1604 * non-null, it will be set to the file descriptor that this thread
1605 * is now listening on. If socketp is null, this routine will never
1608 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1610 struct rx_call *call;
1612 struct rx_service *tservice = NULL;
1619 call = rx_GetCall(threadID, tservice, socketp);
1620 if (socketp && *socketp != OSI_NULLSOCKET) {
1621 /* We are now a listener thread */
1626 /* if server is restarting( typically smooth shutdown) then do not
1627 * allow any new calls.
1630 if (rx_tranquil && (call != NULL)) {
1634 MUTEX_ENTER(&call->lock);
1636 rxi_CallError(call, RX_RESTARTING);
1637 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1639 MUTEX_EXIT(&call->lock);
1643 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1644 #ifdef RX_ENABLE_LOCKS
1646 #endif /* RX_ENABLE_LOCKS */
1647 afs_termState = AFSOP_STOP_AFS;
1648 afs_osi_Wakeup(&afs_termState);
1649 #ifdef RX_ENABLE_LOCKS
1651 #endif /* RX_ENABLE_LOCKS */
1656 tservice = call->conn->service;
1658 if (tservice->beforeProc)
1659 (*tservice->beforeProc) (call);
1661 code = tservice->executeRequestProc(call);
1663 if (tservice->afterProc)
1664 (*tservice->afterProc) (call, code);
1666 rx_EndCall(call, code);
1667 if (rx_stats_active) {
1668 MUTEX_ENTER(&rx_stats_mutex);
1670 MUTEX_EXIT(&rx_stats_mutex);
1677 rx_WakeupServerProcs(void)
1679 struct rx_serverQueueEntry *np, *tqp;
1683 MUTEX_ENTER(&rx_serverPool_lock);
1685 #ifdef RX_ENABLE_LOCKS
1686 if (rx_waitForPacket)
1687 CV_BROADCAST(&rx_waitForPacket->cv);
1688 #else /* RX_ENABLE_LOCKS */
1689 if (rx_waitForPacket)
1690 osi_rxWakeup(rx_waitForPacket);
1691 #endif /* RX_ENABLE_LOCKS */
1692 MUTEX_ENTER(&freeSQEList_lock);
1693 for (np = rx_FreeSQEList; np; np = tqp) {
1694 tqp = *(struct rx_serverQueueEntry **)np;
1695 #ifdef RX_ENABLE_LOCKS
1696 CV_BROADCAST(&np->cv);
1697 #else /* RX_ENABLE_LOCKS */
1699 #endif /* RX_ENABLE_LOCKS */
1701 MUTEX_EXIT(&freeSQEList_lock);
1702 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1703 #ifdef RX_ENABLE_LOCKS
1704 CV_BROADCAST(&np->cv);
1705 #else /* RX_ENABLE_LOCKS */
1707 #endif /* RX_ENABLE_LOCKS */
1709 MUTEX_EXIT(&rx_serverPool_lock);
1714 * One thing that seems to happen is that all the server threads get
1715 * tied up on some empty or slow call, and then a whole bunch of calls
1716 * arrive at once, using up the packet pool, so now there are more
1717 * empty calls. The most critical resources here are server threads
1718 * and the free packet pool. The "doreclaim" code seems to help in
1719 * general. I think that eventually we arrive in this state: there
1720 * are lots of pending calls which do have all their packets present,
1721 * so they won't be reclaimed, are multi-packet calls, so they won't
1722 * be scheduled until later, and thus are tying up most of the free
1723 * packet pool for a very long time.
1725 * 1. schedule multi-packet calls if all the packets are present.
1726 * Probably CPU-bound operation, useful to return packets to pool.
1727 * Do what if there is a full window, but the last packet isn't here?
1728 * 3. preserve one thread which *only* runs "best" calls, otherwise
1729 * it sleeps and waits for that type of call.
1730 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1731 * the current dataquota business is badly broken. The quota isn't adjusted
1732 * to reflect how many packets are presently queued for a running call.
1733 * So, when we schedule a queued call with a full window of packets queued
1734 * up for it, that *should* free up a window full of packets for other 2d-class
1735 * calls to be able to use from the packet pool. But it doesn't.
1737 * NB. Most of the time, this code doesn't run -- since idle server threads
1738 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1739 * as a new call arrives.
1741 /* Sleep until a call arrives. Returns a pointer to the call, ready
1742 * for an rx_Read. */
1743 #ifdef RX_ENABLE_LOCKS
1745 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1747 struct rx_serverQueueEntry *sq;
1748 struct rx_call *call = (struct rx_call *)0;
1749 struct rx_service *service = NULL;
1752 MUTEX_ENTER(&freeSQEList_lock);
1754 if ((sq = rx_FreeSQEList)) {
1755 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1756 MUTEX_EXIT(&freeSQEList_lock);
1757 } else { /* otherwise allocate a new one and return that */
1758 MUTEX_EXIT(&freeSQEList_lock);
1759 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1760 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1761 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1764 MUTEX_ENTER(&rx_serverPool_lock);
1765 if (cur_service != NULL) {
1766 ReturnToServerPool(cur_service);
1769 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1770 struct rx_call *tcall, *ncall, *choice2 = NULL;
1772 /* Scan for eligible incoming calls. A call is not eligible
1773 * if the maximum number of calls for its service type are
1774 * already executing */
1775 /* One thread will process calls FCFS (to prevent starvation),
1776 * while the other threads may run ahead looking for calls which
1777 * have all their input data available immediately. This helps
1778 * keep threads from blocking, waiting for data from the client. */
1779 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1780 service = tcall->conn->service;
1781 if (!QuotaOK(service)) {
1784 MUTEX_ENTER(&rx_pthread_mutex);
1785 if (tno == rxi_fcfs_thread_num
1786 || !tcall->queue_item_header.next) {
1787 MUTEX_EXIT(&rx_pthread_mutex);
1788 /* If we're the fcfs thread , then we'll just use
1789 * this call. If we haven't been able to find an optimal
1790 * choice, and we're at the end of the list, then use a
1791 * 2d choice if one has been identified. Otherwise... */
1792 call = (choice2 ? choice2 : tcall);
1793 service = call->conn->service;
1795 MUTEX_EXIT(&rx_pthread_mutex);
1796 if (!queue_IsEmpty(&tcall->rq)) {
1797 struct rx_packet *rp;
1798 rp = queue_First(&tcall->rq, rx_packet);
1799 if (rp->header.seq == 1) {
1801 || (rp->header.flags & RX_LAST_PACKET)) {
1803 } else if (rxi_2dchoice && !choice2
1804 && !(tcall->flags & RX_CALL_CLEARED)
1805 && (tcall->rprev > rxi_HardAckRate)) {
1815 ReturnToServerPool(service);
1822 MUTEX_EXIT(&rx_serverPool_lock);
1823 MUTEX_ENTER(&call->lock);
1825 if (call->flags & RX_CALL_WAIT_PROC) {
1826 call->flags &= ~RX_CALL_WAIT_PROC;
1827 rx_atomic_dec(&rx_nWaiting);
1830 if (call->state != RX_STATE_PRECALL || call->error) {
1831 MUTEX_EXIT(&call->lock);
1832 MUTEX_ENTER(&rx_serverPool_lock);
1833 ReturnToServerPool(service);
1838 if (queue_IsEmpty(&call->rq)
1839 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1840 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1842 CLEAR_CALL_QUEUE_LOCK(call);
1845 /* If there are no eligible incoming calls, add this process
1846 * to the idle server queue, to wait for one */
1850 *socketp = OSI_NULLSOCKET;
1852 sq->socketp = socketp;
1853 queue_Append(&rx_idleServerQueue, sq);
1854 #ifndef AFS_AIX41_ENV
1855 rx_waitForPacket = sq;
1857 rx_waitingForPacket = sq;
1858 #endif /* AFS_AIX41_ENV */
1860 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1862 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1863 MUTEX_EXIT(&rx_serverPool_lock);
1864 return (struct rx_call *)0;
1867 } while (!(call = sq->newcall)
1868 && !(socketp && *socketp != OSI_NULLSOCKET));
1869 MUTEX_EXIT(&rx_serverPool_lock);
1871 MUTEX_ENTER(&call->lock);
1877 MUTEX_ENTER(&freeSQEList_lock);
1878 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1879 rx_FreeSQEList = sq;
1880 MUTEX_EXIT(&freeSQEList_lock);
1883 clock_GetTime(&call->startTime);
1884 call->state = RX_STATE_ACTIVE;
1885 call->mode = RX_MODE_RECEIVING;
1886 #ifdef RX_KERNEL_TRACE
1887 if (ICL_SETACTIVE(afs_iclSetp)) {
1888 int glockOwner = ISAFS_GLOCK();
1891 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1892 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1899 rxi_calltrace(RX_CALL_START, call);
1900 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
1901 call->conn->service->servicePort, call->conn->service->serviceId,
1904 MUTEX_EXIT(&call->lock);
1905 MUTEX_ENTER(&rx_refcnt_mutex);
1906 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1907 MUTEX_EXIT(&rx_refcnt_mutex);
1909 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1914 #else /* RX_ENABLE_LOCKS */
1916 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1918 struct rx_serverQueueEntry *sq;
1919 struct rx_call *call = (struct rx_call *)0, *choice2;
1920 struct rx_service *service = NULL;
1924 MUTEX_ENTER(&freeSQEList_lock);
1926 if ((sq = rx_FreeSQEList)) {
1927 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1928 MUTEX_EXIT(&freeSQEList_lock);
1929 } else { /* otherwise allocate a new one and return that */
1930 MUTEX_EXIT(&freeSQEList_lock);
1931 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1932 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1933 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1935 MUTEX_ENTER(&sq->lock);
1937 if (cur_service != NULL) {
1938 cur_service->nRequestsRunning--;
1939 MUTEX_ENTER(&rx_quota_mutex);
1940 if (cur_service->nRequestsRunning < cur_service->minProcs)
1943 MUTEX_EXIT(&rx_quota_mutex);
1945 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1946 struct rx_call *tcall, *ncall;
1947 /* Scan for eligible incoming calls. A call is not eligible
1948 * if the maximum number of calls for its service type are
1949 * already executing */
1950 /* One thread will process calls FCFS (to prevent starvation),
1951 * while the other threads may run ahead looking for calls which
1952 * have all their input data available immediately. This helps
1953 * keep threads from blocking, waiting for data from the client. */
1954 choice2 = (struct rx_call *)0;
1955 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1956 service = tcall->conn->service;
1957 if (QuotaOK(service)) {
1958 MUTEX_ENTER(&rx_pthread_mutex);
1959 if (tno == rxi_fcfs_thread_num
1960 || !tcall->queue_item_header.next) {
1961 MUTEX_EXIT(&rx_pthread_mutex);
1962 /* If we're the fcfs thread, then we'll just use
1963 * this call. If we haven't been able to find an optimal
1964 * choice, and we're at the end of the list, then use a
1965 * 2d choice if one has been identified. Otherwise... */
1966 call = (choice2 ? choice2 : tcall);
1967 service = call->conn->service;
1969 MUTEX_EXIT(&rx_pthread_mutex);
1970 if (!queue_IsEmpty(&tcall->rq)) {
1971 struct rx_packet *rp;
1972 rp = queue_First(&tcall->rq, rx_packet);
1973 if (rp->header.seq == 1
1975 || (rp->header.flags & RX_LAST_PACKET))) {
1977 } else if (rxi_2dchoice && !choice2
1978 && !(tcall->flags & RX_CALL_CLEARED)
1979 && (tcall->rprev > rxi_HardAckRate)) {
1993 /* we can't schedule a call if there's no data!!! */
1994 /* send an ack if there's no data, if we're missing the
1995 * first packet, or we're missing something between first
1996 * and last -- there's a "hole" in the incoming data. */
1997 if (queue_IsEmpty(&call->rq)
1998 || queue_First(&call->rq, rx_packet)->header.seq != 1
1999 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
2000 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2002 call->flags &= (~RX_CALL_WAIT_PROC);
2003 service->nRequestsRunning++;
2004 /* just started call in minProcs pool, need fewer to maintain
2006 MUTEX_ENTER(&rx_quota_mutex);
2007 if (service->nRequestsRunning <= service->minProcs)
2010 MUTEX_EXIT(&rx_quota_mutex);
2011 rx_atomic_dec(&rx_nWaiting);
2012 /* MUTEX_EXIT(&call->lock); */
2014 /* If there are no eligible incoming calls, add this process
2015 * to the idle server queue, to wait for one */
2018 *socketp = OSI_NULLSOCKET;
2020 sq->socketp = socketp;
2021 queue_Append(&rx_idleServerQueue, sq);
2025 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2027 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2028 return (struct rx_call *)0;
2031 } while (!(call = sq->newcall)
2032 && !(socketp && *socketp != OSI_NULLSOCKET));
2034 MUTEX_EXIT(&sq->lock);
2036 MUTEX_ENTER(&freeSQEList_lock);
2037 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2038 rx_FreeSQEList = sq;
2039 MUTEX_EXIT(&freeSQEList_lock);
2042 clock_GetTime(&call->startTime);
2043 call->state = RX_STATE_ACTIVE;
2044 call->mode = RX_MODE_RECEIVING;
2045 #ifdef RX_KERNEL_TRACE
2046 if (ICL_SETACTIVE(afs_iclSetp)) {
2047 int glockOwner = ISAFS_GLOCK();
2050 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2051 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2058 rxi_calltrace(RX_CALL_START, call);
2059 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2060 call->conn->service->servicePort, call->conn->service->serviceId,
2063 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2070 #endif /* RX_ENABLE_LOCKS */
2074 /* Establish a procedure to be called when a packet arrives for a
2075 * call. This routine will be called at most once after each call,
2076 * and will also be called if there is an error condition on the or
2077 * the call is complete. Used by multi rx to build a selection
2078 * function which determines which of several calls is likely to be a
2079 * good one to read from.
2080 * NOTE: the way this is currently implemented it is probably only a
2081 * good idea to (1) use it immediately after a newcall (clients only)
2082 * and (2) only use it once. Other uses currently void your warranty
2085 rx_SetArrivalProc(struct rx_call *call,
2086 void (*proc) (struct rx_call * call,
2089 void * handle, int arg)
2091 call->arrivalProc = proc;
2092 call->arrivalProcHandle = handle;
2093 call->arrivalProcArg = arg;
2096 /* Call is finished (possibly prematurely). Return rc to the peer, if
2097 * appropriate, and return the final error code from the conversation
2101 rx_EndCall(struct rx_call *call, afs_int32 rc)
2103 struct rx_connection *conn = call->conn;
2107 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2108 call, rc, call->error, call->abortCode));
2111 MUTEX_ENTER(&call->lock);
2113 if (rc == 0 && call->error == 0) {
2114 call->abortCode = 0;
2115 call->abortCount = 0;
2118 call->arrivalProc = (void (*)())0;
2119 if (rc && call->error == 0) {
2120 rxi_CallError(call, rc);
2121 call->mode = RX_MODE_ERROR;
2122 /* Send an abort message to the peer if this error code has
2123 * only just been set. If it was set previously, assume the
2124 * peer has already been sent the error code or will request it
2126 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2128 if (conn->type == RX_SERVER_CONNECTION) {
2129 /* Make sure reply or at least dummy reply is sent */
2130 if (call->mode == RX_MODE_RECEIVING) {
2131 MUTEX_EXIT(&call->lock);
2132 rxi_WriteProc(call, 0, 0);
2133 MUTEX_ENTER(&call->lock);
2135 if (call->mode == RX_MODE_SENDING) {
2136 MUTEX_EXIT(&call->lock);
2137 rxi_FlushWrite(call);
2138 MUTEX_ENTER(&call->lock);
2140 rxi_calltrace(RX_CALL_END, call);
2141 /* Call goes to hold state until reply packets are acknowledged */
2142 if (call->tfirst + call->nSoftAcked < call->tnext) {
2143 call->state = RX_STATE_HOLD;
2145 call->state = RX_STATE_DALLY;
2146 rxi_ClearTransmitQueue(call, 0);
2147 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
2148 rxevent_Cancel(call->keepAliveEvent, call,
2149 RX_CALL_REFCOUNT_ALIVE);
2151 } else { /* Client connection */
2153 /* Make sure server receives input packets, in the case where
2154 * no reply arguments are expected */
2155 if ((call->mode == RX_MODE_SENDING)
2156 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2157 MUTEX_EXIT(&call->lock);
2158 (void)rxi_ReadProc(call, &dummy, 1);
2159 MUTEX_ENTER(&call->lock);
2162 /* If we had an outstanding delayed ack, be nice to the server
2163 * and force-send it now.
2165 if (call->delayedAckEvent) {
2166 rxevent_Cancel(call->delayedAckEvent, call,
2167 RX_CALL_REFCOUNT_DELAY);
2168 call->delayedAckEvent = NULL;
2169 rxi_SendDelayedAck(NULL, call, NULL);
2172 /* We need to release the call lock since it's lower than the
2173 * conn_call_lock and we don't want to hold the conn_call_lock
2174 * over the rx_ReadProc call. The conn_call_lock needs to be held
2175 * here for the case where rx_NewCall is perusing the calls on
2176 * the connection structure. We don't want to signal until
2177 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2178 * have checked this call, found it active and by the time it
2179 * goes to sleep, will have missed the signal.
2181 MUTEX_EXIT(&call->lock);
2182 MUTEX_ENTER(&conn->conn_call_lock);
2183 MUTEX_ENTER(&call->lock);
2184 MUTEX_ENTER(&conn->conn_data_lock);
2185 conn->flags |= RX_CONN_BUSY;
2186 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2187 MUTEX_EXIT(&conn->conn_data_lock);
2188 #ifdef RX_ENABLE_LOCKS
2189 CV_BROADCAST(&conn->conn_call_cv);
2194 #ifdef RX_ENABLE_LOCKS
2196 MUTEX_EXIT(&conn->conn_data_lock);
2198 #endif /* RX_ENABLE_LOCKS */
2199 call->state = RX_STATE_DALLY;
2201 error = call->error;
2203 /* currentPacket, nLeft, and NFree must be zeroed here, because
2204 * ResetCall cannot: ResetCall may be called at splnet(), in the
2205 * kernel version, and may interrupt the macros rx_Read or
2206 * rx_Write, which run at normal priority for efficiency. */
2207 if (call->currentPacket) {
2208 #ifdef RX_TRACK_PACKETS
2209 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2211 rxi_FreePacket(call->currentPacket);
2212 call->currentPacket = (struct rx_packet *)0;
2215 call->nLeft = call->nFree = call->curlen = 0;
2217 /* Free any packets from the last call to ReadvProc/WritevProc */
2218 #ifdef RXDEBUG_PACKET
2220 #endif /* RXDEBUG_PACKET */
2221 rxi_FreePackets(0, &call->iovq);
2222 MUTEX_EXIT(&call->lock);
2224 MUTEX_ENTER(&rx_refcnt_mutex);
2225 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2226 MUTEX_EXIT(&rx_refcnt_mutex);
2227 if (conn->type == RX_CLIENT_CONNECTION) {
2228 MUTEX_ENTER(&conn->conn_data_lock);
2229 conn->flags &= ~RX_CONN_BUSY;
2230 MUTEX_EXIT(&conn->conn_data_lock);
2231 MUTEX_EXIT(&conn->conn_call_lock);
2235 * Map errors to the local host's errno.h format.
2237 error = ntoh_syserr_conv(error);
2241 #if !defined(KERNEL)
2243 /* Call this routine when shutting down a server or client (especially
2244 * clients). This will allow Rx to gracefully garbage collect server
2245 * connections, and reduce the number of retries that a server might
2246 * make to a dead client.
2247 * This is not quite right, since some calls may still be ongoing and
2248 * we can't lock them to destroy them. */
2252 struct rx_connection **conn_ptr, **conn_end;
2256 if (rxinit_status == 1) {
2258 return; /* Already shutdown. */
2260 rxi_DeleteCachedConnections();
2261 if (rx_connHashTable) {
2262 MUTEX_ENTER(&rx_connHashTable_lock);
2263 for (conn_ptr = &rx_connHashTable[0], conn_end =
2264 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2266 struct rx_connection *conn, *next;
2267 for (conn = *conn_ptr; conn; conn = next) {
2269 if (conn->type == RX_CLIENT_CONNECTION) {
2270 MUTEX_ENTER(&rx_refcnt_mutex);
2272 MUTEX_EXIT(&rx_refcnt_mutex);
2273 #ifdef RX_ENABLE_LOCKS
2274 rxi_DestroyConnectionNoLock(conn);
2275 #else /* RX_ENABLE_LOCKS */
2276 rxi_DestroyConnection(conn);
2277 #endif /* RX_ENABLE_LOCKS */
2281 #ifdef RX_ENABLE_LOCKS
2282 while (rx_connCleanup_list) {
2283 struct rx_connection *conn;
2284 conn = rx_connCleanup_list;
2285 rx_connCleanup_list = rx_connCleanup_list->next;
2286 MUTEX_EXIT(&rx_connHashTable_lock);
2287 rxi_CleanupConnection(conn);
2288 MUTEX_ENTER(&rx_connHashTable_lock);
2290 MUTEX_EXIT(&rx_connHashTable_lock);
2291 #endif /* RX_ENABLE_LOCKS */
2296 afs_winsockCleanup();
2304 /* if we wakeup packet waiter too often, can get in loop with two
2305 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2307 rxi_PacketsUnWait(void)
2309 if (!rx_waitingForPackets) {
2313 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2314 return; /* still over quota */
2317 rx_waitingForPackets = 0;
2318 #ifdef RX_ENABLE_LOCKS
2319 CV_BROADCAST(&rx_waitingForPackets_cv);
2321 osi_rxWakeup(&rx_waitingForPackets);
2327 /* ------------------Internal interfaces------------------------- */
2329 /* Return this process's service structure for the
2330 * specified socket and service */
2332 rxi_FindService(osi_socket socket, u_short serviceId)
2334 struct rx_service **sp;
2335 for (sp = &rx_services[0]; *sp; sp++) {
2336 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2342 #ifdef RXDEBUG_PACKET
2343 #ifdef KDUMP_RX_LOCK
2344 static struct rx_call_rx_lock *rx_allCallsp = 0;
2346 static struct rx_call *rx_allCallsp = 0;
2348 #endif /* RXDEBUG_PACKET */
2350 /* Allocate a call structure, for the indicated channel of the
2351 * supplied connection. The mode and state of the call must be set by
2352 * the caller. Returns the call with mutex locked. */
2354 rxi_NewCall(struct rx_connection *conn, int channel)
2356 struct rx_call *call;
2357 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2358 struct rx_call *cp; /* Call pointer temp */
2359 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2360 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2362 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2364 /* Grab an existing call structure, or allocate a new one.
2365 * Existing call structures are assumed to have been left reset by
2367 MUTEX_ENTER(&rx_freeCallQueue_lock);
2369 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2371 * EXCEPT that the TQ might not yet be cleared out.
2372 * Skip over those with in-use TQs.
2375 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2376 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2382 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2383 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2384 call = queue_First(&rx_freeCallQueue, rx_call);
2385 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2387 if (rx_stats_active)
2388 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2389 MUTEX_EXIT(&rx_freeCallQueue_lock);
2390 MUTEX_ENTER(&call->lock);
2391 CLEAR_CALL_QUEUE_LOCK(call);
2392 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2393 /* Now, if TQ wasn't cleared earlier, do it now. */
2394 rxi_WaitforTQBusy(call);
2395 if (call->flags & RX_CALL_TQ_CLEARME) {
2396 rxi_ClearTransmitQueue(call, 1);
2397 /*queue_Init(&call->tq);*/
2399 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2400 /* Bind the call to its connection structure */
2402 rxi_ResetCall(call, 1);
2405 call = rxi_Alloc(sizeof(struct rx_call));
2406 #ifdef RXDEBUG_PACKET
2407 call->allNextp = rx_allCallsp;
2408 rx_allCallsp = call;
2410 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2411 #else /* RXDEBUG_PACKET */
2412 rx_atomic_inc(&rx_stats.nCallStructs);
2413 #endif /* RXDEBUG_PACKET */
2415 MUTEX_EXIT(&rx_freeCallQueue_lock);
2416 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2417 MUTEX_ENTER(&call->lock);
2418 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2419 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2420 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2422 /* Initialize once-only items */
2423 queue_Init(&call->tq);
2424 queue_Init(&call->rq);
2425 queue_Init(&call->iovq);
2426 #ifdef RXDEBUG_PACKET
2427 call->rqc = call->tqc = call->iovqc = 0;
2428 #endif /* RXDEBUG_PACKET */
2429 /* Bind the call to its connection structure (prereq for reset) */
2431 rxi_ResetCall(call, 1);
2433 call->channel = channel;
2434 call->callNumber = &conn->callNumber[channel];
2435 call->rwind = conn->rwind[channel];
2436 call->twind = conn->twind[channel];
2437 /* Note that the next expected call number is retained (in
2438 * conn->callNumber[i]), even if we reallocate the call structure
2440 conn->call[channel] = call;
2441 /* if the channel's never been used (== 0), we should start at 1, otherwise
2442 * the call number is valid from the last time this channel was used */
2443 if (*call->callNumber == 0)
2444 *call->callNumber = 1;
2449 /* A call has been inactive long enough that so we can throw away
2450 * state, including the call structure, which is placed on the call
2453 * call->lock amd rx_refcnt_mutex are held upon entry.
2454 * haveCTLock is set when called from rxi_ReapConnections.
2457 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2459 int channel = call->channel;
2460 struct rx_connection *conn = call->conn;
2463 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2464 (*call->callNumber)++;
2465 rxi_ResetCall(call, 0);
2466 call->conn->call[channel] = (struct rx_call *)0;
2467 MUTEX_EXIT(&rx_refcnt_mutex);
2469 MUTEX_ENTER(&rx_freeCallQueue_lock);
2470 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2471 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2472 /* A call may be free even though its transmit queue is still in use.
2473 * Since we search the call list from head to tail, put busy calls at
2474 * the head of the list, and idle calls at the tail.
2476 if (call->flags & RX_CALL_TQ_BUSY)
2477 queue_Prepend(&rx_freeCallQueue, call);
2479 queue_Append(&rx_freeCallQueue, call);
2480 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2481 queue_Append(&rx_freeCallQueue, call);
2482 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2483 if (rx_stats_active)
2484 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2485 MUTEX_EXIT(&rx_freeCallQueue_lock);
2487 /* Destroy the connection if it was previously slated for
2488 * destruction, i.e. the Rx client code previously called
2489 * rx_DestroyConnection (client connections), or
2490 * rxi_ReapConnections called the same routine (server
2491 * connections). Only do this, however, if there are no
2492 * outstanding calls. Note that for fine grain locking, there appears
2493 * to be a deadlock in that rxi_FreeCall has a call locked and
2494 * DestroyConnectionNoLock locks each call in the conn. But note a
2495 * few lines up where we have removed this call from the conn.
2496 * If someone else destroys a connection, they either have no
2497 * call lock held or are going through this section of code.
2499 MUTEX_ENTER(&conn->conn_data_lock);
2500 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2501 MUTEX_ENTER(&rx_refcnt_mutex);
2503 MUTEX_EXIT(&rx_refcnt_mutex);
2504 MUTEX_EXIT(&conn->conn_data_lock);
2505 #ifdef RX_ENABLE_LOCKS
2507 rxi_DestroyConnectionNoLock(conn);
2509 rxi_DestroyConnection(conn);
2510 #else /* RX_ENABLE_LOCKS */
2511 rxi_DestroyConnection(conn);
2512 #endif /* RX_ENABLE_LOCKS */
2514 MUTEX_EXIT(&conn->conn_data_lock);
2516 MUTEX_ENTER(&rx_refcnt_mutex);
2519 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2520 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2523 rxi_Alloc(size_t size)
2527 if (rx_stats_active) {
2528 rx_atomic_add(&rxi_Allocsize, (int) size);
2529 rx_atomic_inc(&rxi_Alloccnt);
2533 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2534 afs_osi_Alloc_NoSleep(size);
2539 osi_Panic("rxi_Alloc error");
2545 rxi_Free(void *addr, size_t size)
2547 if (rx_stats_active) {
2548 rx_atomic_sub(&rxi_Allocsize, (int) size);
2549 rx_atomic_dec(&rxi_Alloccnt);
2551 osi_Free(addr, size);
2555 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2557 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2558 struct rx_peer *next = NULL;
2562 MUTEX_ENTER(&rx_peerHashTable_lock);
2564 peer_ptr = &rx_peerHashTable[0];
2565 peer_end = &rx_peerHashTable[rx_hashTableSize];
2568 for ( ; peer_ptr < peer_end; peer_ptr++) {
2571 for ( ; peer; peer = next) {
2573 if (host == peer->host)
2578 hashIndex = PEER_HASH(host, port);
2579 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2580 if ((peer->host == host) && (peer->port == port))
2585 MUTEX_ENTER(&rx_peerHashTable_lock);
2590 MUTEX_EXIT(&rx_peerHashTable_lock);
2592 MUTEX_ENTER(&peer->peer_lock);
2593 /* We don't handle dropping below min, so don't */
2594 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2595 peer->ifMTU=MIN(mtu, peer->ifMTU);
2596 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2597 /* if we tweaked this down, need to tune our peer MTU too */
2598 peer->MTU = MIN(peer->MTU, peer->natMTU);
2599 /* if we discovered a sub-1500 mtu, degrade */
2600 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2601 peer->maxDgramPackets = 1;
2602 /* We no longer have valid peer packet information */
2603 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2604 peer->maxPacketSize = 0;
2605 MUTEX_EXIT(&peer->peer_lock);
2607 MUTEX_ENTER(&rx_peerHashTable_lock);
2609 if (host && !port) {
2611 /* pick up where we left off */
2615 MUTEX_EXIT(&rx_peerHashTable_lock);
2618 /* Find the peer process represented by the supplied (host,port)
2619 * combination. If there is no appropriate active peer structure, a
2620 * new one will be allocated and initialized
2621 * The origPeer, if set, is a pointer to a peer structure on which the
2622 * refcount will be be decremented. This is used to replace the peer
2623 * structure hanging off a connection structure */
2625 rxi_FindPeer(afs_uint32 host, u_short port,
2626 struct rx_peer *origPeer, int create)
2630 hashIndex = PEER_HASH(host, port);
2631 MUTEX_ENTER(&rx_peerHashTable_lock);
2632 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2633 if ((pp->host == host) && (pp->port == port))
2638 pp = rxi_AllocPeer(); /* This bzero's *pp */
2639 pp->host = host; /* set here or in InitPeerParams is zero */
2641 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2642 queue_Init(&pp->congestionQueue);
2643 queue_Init(&pp->rpcStats);
2644 pp->next = rx_peerHashTable[hashIndex];
2645 rx_peerHashTable[hashIndex] = pp;
2646 rxi_InitPeerParams(pp);
2647 if (rx_stats_active)
2648 rx_atomic_inc(&rx_stats.nPeerStructs);
2655 origPeer->refCount--;
2656 MUTEX_EXIT(&rx_peerHashTable_lock);
2661 /* Find the connection at (host, port) started at epoch, and with the
2662 * given connection id. Creates the server connection if necessary.
2663 * The type specifies whether a client connection or a server
2664 * connection is desired. In both cases, (host, port) specify the
2665 * peer's (host, pair) pair. Client connections are not made
2666 * automatically by this routine. The parameter socket gives the
2667 * socket descriptor on which the packet was received. This is used,
2668 * in the case of server connections, to check that *new* connections
2669 * come via a valid (port, serviceId). Finally, the securityIndex
2670 * parameter must match the existing index for the connection. If a
2671 * server connection is created, it will be created using the supplied
2672 * index, if the index is valid for this service */
2673 struct rx_connection *
2674 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2675 u_short port, u_short serviceId, afs_uint32 cid,
2676 afs_uint32 epoch, int type, u_int securityIndex)
2678 int hashindex, flag, i;
2679 struct rx_connection *conn;
2680 hashindex = CONN_HASH(host, port, cid, epoch, type);
2681 MUTEX_ENTER(&rx_connHashTable_lock);
2682 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2683 rx_connHashTable[hashindex],
2686 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2687 && (epoch == conn->epoch)) {
2688 struct rx_peer *pp = conn->peer;
2689 if (securityIndex != conn->securityIndex) {
2690 /* this isn't supposed to happen, but someone could forge a packet
2691 * like this, and there seems to be some CM bug that makes this
2692 * happen from time to time -- in which case, the fileserver
2694 MUTEX_EXIT(&rx_connHashTable_lock);
2695 return (struct rx_connection *)0;
2697 if (pp->host == host && pp->port == port)
2699 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2701 /* So what happens when it's a callback connection? */
2702 if ( /*type == RX_CLIENT_CONNECTION && */
2703 (conn->epoch & 0x80000000))
2707 /* the connection rxLastConn that was used the last time is not the
2708 ** one we are looking for now. Hence, start searching in the hash */
2710 conn = rx_connHashTable[hashindex];
2715 struct rx_service *service;
2716 if (type == RX_CLIENT_CONNECTION) {
2717 MUTEX_EXIT(&rx_connHashTable_lock);
2718 return (struct rx_connection *)0;
2720 service = rxi_FindService(socket, serviceId);
2721 if (!service || (securityIndex >= service->nSecurityObjects)
2722 || (service->securityObjects[securityIndex] == 0)) {
2723 MUTEX_EXIT(&rx_connHashTable_lock);
2724 return (struct rx_connection *)0;
2726 conn = rxi_AllocConnection(); /* This bzero's the connection */
2727 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2728 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2729 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2730 conn->next = rx_connHashTable[hashindex];
2731 rx_connHashTable[hashindex] = conn;
2732 conn->peer = rxi_FindPeer(host, port, 0, 1);
2733 conn->type = RX_SERVER_CONNECTION;
2734 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2735 conn->epoch = epoch;
2736 conn->cid = cid & RX_CIDMASK;
2737 /* conn->serial = conn->lastSerial = 0; */
2738 /* conn->timeout = 0; */
2739 conn->ackRate = RX_FAST_ACK_RATE;
2740 conn->service = service;
2741 conn->serviceId = serviceId;
2742 conn->securityIndex = securityIndex;
2743 conn->securityObject = service->securityObjects[securityIndex];
2744 conn->nSpecific = 0;
2745 conn->specific = NULL;
2746 rx_SetConnDeadTime(conn, service->connDeadTime);
2747 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2748 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2749 for (i = 0; i < RX_MAXCALLS; i++) {
2750 conn->twind[i] = rx_initSendWindow;
2751 conn->rwind[i] = rx_initReceiveWindow;
2753 /* Notify security object of the new connection */
2754 RXS_NewConnection(conn->securityObject, conn);
2755 /* XXXX Connection timeout? */
2756 if (service->newConnProc)
2757 (*service->newConnProc) (conn);
2758 if (rx_stats_active)
2759 rx_atomic_inc(&rx_stats.nServerConns);
2762 MUTEX_ENTER(&rx_refcnt_mutex);
2764 MUTEX_EXIT(&rx_refcnt_mutex);
2766 rxLastConn = conn; /* store this connection as the last conn used */
2767 MUTEX_EXIT(&rx_connHashTable_lock);
2771 /* There are two packet tracing routines available for testing and monitoring
2772 * Rx. One is called just after every packet is received and the other is
2773 * called just before every packet is sent. Received packets, have had their
2774 * headers decoded, and packets to be sent have not yet had their headers
2775 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2776 * containing the network address. Both can be modified. The return value, if
2777 * non-zero, indicates that the packet should be dropped. */
2779 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2780 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2782 /* A packet has been received off the interface. Np is the packet, socket is
2783 * the socket number it was received from (useful in determining which service
2784 * this packet corresponds to), and (host, port) reflect the host,port of the
2785 * sender. This call returns the packet to the caller if it is finished with
2786 * it, rather than de-allocating it, just as a small performance hack */
2789 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
2790 afs_uint32 host, u_short port, int *tnop,
2791 struct rx_call **newcallp)
2793 struct rx_call *call;
2794 struct rx_connection *conn;
2796 afs_uint32 currentCallNumber;
2802 struct rx_packet *tnp;
2805 /* We don't print out the packet until now because (1) the time may not be
2806 * accurate enough until now in the lwp implementation (rx_Listener only gets
2807 * the time after the packet is read) and (2) from a protocol point of view,
2808 * this is the first time the packet has been seen */
2809 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2810 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2811 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
2812 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2813 np->header.epoch, np->header.cid, np->header.callNumber,
2814 np->header.seq, np->header.flags, np));
2817 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2818 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2821 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2822 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2825 /* If an input tracer function is defined, call it with the packet and
2826 * network address. Note this function may modify its arguments. */
2827 if (rx_justReceived) {
2828 struct sockaddr_in addr;
2830 addr.sin_family = AF_INET;
2831 addr.sin_port = port;
2832 addr.sin_addr.s_addr = host;
2833 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2834 addr.sin_len = sizeof(addr);
2835 #endif /* AFS_OSF_ENV */
2836 drop = (*rx_justReceived) (np, &addr);
2837 /* drop packet if return value is non-zero */
2840 port = addr.sin_port; /* in case fcn changed addr */
2841 host = addr.sin_addr.s_addr;
2845 /* If packet was not sent by the client, then *we* must be the client */
2846 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2847 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2849 /* Find the connection (or fabricate one, if we're the server & if
2850 * necessary) associated with this packet */
2852 rxi_FindConnection(socket, host, port, np->header.serviceId,
2853 np->header.cid, np->header.epoch, type,
2854 np->header.securityIndex);
2857 /* If no connection found or fabricated, just ignore the packet.
2858 * (An argument could be made for sending an abort packet for
2863 MUTEX_ENTER(&conn->conn_data_lock);
2864 if (conn->maxSerial < np->header.serial)
2865 conn->maxSerial = np->header.serial;
2866 MUTEX_EXIT(&conn->conn_data_lock);
2868 /* If the connection is in an error state, send an abort packet and ignore
2869 * the incoming packet */
2871 /* Don't respond to an abort packet--we don't want loops! */
2872 MUTEX_ENTER(&conn->conn_data_lock);
2873 if (np->header.type != RX_PACKET_TYPE_ABORT)
2874 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2875 MUTEX_ENTER(&rx_refcnt_mutex);
2877 MUTEX_EXIT(&rx_refcnt_mutex);
2878 MUTEX_EXIT(&conn->conn_data_lock);
2882 /* Check for connection-only requests (i.e. not call specific). */
2883 if (np->header.callNumber == 0) {
2884 switch (np->header.type) {
2885 case RX_PACKET_TYPE_ABORT: {
2886 /* What if the supplied error is zero? */
2887 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2888 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
2889 rxi_ConnectionError(conn, errcode);
2890 MUTEX_ENTER(&rx_refcnt_mutex);
2892 MUTEX_EXIT(&rx_refcnt_mutex);
2895 case RX_PACKET_TYPE_CHALLENGE:
2896 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2897 MUTEX_ENTER(&rx_refcnt_mutex);
2899 MUTEX_EXIT(&rx_refcnt_mutex);
2901 case RX_PACKET_TYPE_RESPONSE:
2902 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2903 MUTEX_ENTER(&rx_refcnt_mutex);
2905 MUTEX_EXIT(&rx_refcnt_mutex);
2907 case RX_PACKET_TYPE_PARAMS:
2908 case RX_PACKET_TYPE_PARAMS + 1:
2909 case RX_PACKET_TYPE_PARAMS + 2:
2910 /* ignore these packet types for now */
2911 MUTEX_ENTER(&rx_refcnt_mutex);
2913 MUTEX_EXIT(&rx_refcnt_mutex);
2918 /* Should not reach here, unless the peer is broken: send an
2920 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2921 MUTEX_ENTER(&conn->conn_data_lock);
2922 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2923 MUTEX_ENTER(&rx_refcnt_mutex);
2925 MUTEX_EXIT(&rx_refcnt_mutex);
2926 MUTEX_EXIT(&conn->conn_data_lock);
2931 channel = np->header.cid & RX_CHANNELMASK;
2932 call = conn->call[channel];
2933 #ifdef RX_ENABLE_LOCKS
2935 MUTEX_ENTER(&call->lock);
2936 /* Test to see if call struct is still attached to conn. */
2937 if (call != conn->call[channel]) {
2939 MUTEX_EXIT(&call->lock);
2940 if (type == RX_SERVER_CONNECTION) {
2941 call = conn->call[channel];
2942 /* If we started with no call attached and there is one now,
2943 * another thread is also running this routine and has gotten
2944 * the connection channel. We should drop this packet in the tests
2945 * below. If there was a call on this connection and it's now
2946 * gone, then we'll be making a new call below.
2947 * If there was previously a call and it's now different then
2948 * the old call was freed and another thread running this routine
2949 * has created a call on this channel. One of these two threads
2950 * has a packet for the old call and the code below handles those
2954 MUTEX_ENTER(&call->lock);
2956 /* This packet can't be for this call. If the new call address is
2957 * 0 then no call is running on this channel. If there is a call
2958 * then, since this is a client connection we're getting data for
2959 * it must be for the previous call.
2961 if (rx_stats_active)
2962 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
2963 MUTEX_ENTER(&rx_refcnt_mutex);
2965 MUTEX_EXIT(&rx_refcnt_mutex);
2970 currentCallNumber = conn->callNumber[channel];
2972 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2973 if (np->header.callNumber < currentCallNumber) {
2974 if (rx_stats_active)
2975 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
2976 #ifdef RX_ENABLE_LOCKS
2978 MUTEX_EXIT(&call->lock);
2980 MUTEX_ENTER(&rx_refcnt_mutex);
2982 MUTEX_EXIT(&rx_refcnt_mutex);
2986 MUTEX_ENTER(&conn->conn_call_lock);
2987 call = rxi_NewCall(conn, channel);
2988 MUTEX_EXIT(&conn->conn_call_lock);
2989 *call->callNumber = np->header.callNumber;
2991 if (np->header.callNumber == 0)
2992 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%.06d len %d\n",
2993 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
2994 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
2995 np->header.flags, np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2997 call->state = RX_STATE_PRECALL;
2998 clock_GetTime(&call->queueTime);
2999 hzero(call->bytesSent);
3000 hzero(call->bytesRcvd);
3002 * If the number of queued calls exceeds the overload
3003 * threshold then abort this call.
3005 if ((rx_BusyThreshold > 0) &&
3006 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3007 struct rx_packet *tp;
3009 rxi_CallError(call, rx_BusyError);
3010 tp = rxi_SendCallAbort(call, np, 1, 0);
3011 MUTEX_EXIT(&call->lock);
3012 MUTEX_ENTER(&rx_refcnt_mutex);
3014 MUTEX_EXIT(&rx_refcnt_mutex);
3015 if (rx_stats_active)
3016 rx_atomic_inc(&rx_stats.nBusies);
3019 rxi_KeepAliveOn(call);
3020 } else if (np->header.callNumber != currentCallNumber) {
3021 /* Wait until the transmit queue is idle before deciding
3022 * whether to reset the current call. Chances are that the
3023 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3026 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3027 if (call->state == RX_STATE_ACTIVE) {
3028 rxi_WaitforTQBusy(call);
3030 * If we entered error state while waiting,
3031 * must call rxi_CallError to permit rxi_ResetCall
3032 * to processed when the tqWaiter count hits zero.
3035 rxi_CallError(call, call->error);
3036 MUTEX_EXIT(&call->lock);
3037 MUTEX_ENTER(&rx_refcnt_mutex);
3039 MUTEX_EXIT(&rx_refcnt_mutex);
3043 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3044 /* If the new call cannot be taken right now send a busy and set
3045 * the error condition in this call, so that it terminates as
3046 * quickly as possible */
3047 if (call->state == RX_STATE_ACTIVE) {
3048 struct rx_packet *tp;
3050 rxi_CallError(call, RX_CALL_DEAD);
3051 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3053 MUTEX_EXIT(&call->lock);
3054 MUTEX_ENTER(&rx_refcnt_mutex);
3056 MUTEX_EXIT(&rx_refcnt_mutex);
3059 rxi_ResetCall(call, 0);
3060 *call->callNumber = np->header.callNumber;
3062 if (np->header.callNumber == 0)
3063 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%06d len %d\n",
3064 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3065 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3066 np->header.flags, np, np->retryTime.sec, np->retryTime.usec, np->length));
3068 call->state = RX_STATE_PRECALL;
3069 clock_GetTime(&call->queueTime);
3070 hzero(call->bytesSent);
3071 hzero(call->bytesRcvd);
3073 * If the number of queued calls exceeds the overload
3074 * threshold then abort this call.
3076 if ((rx_BusyThreshold > 0) &&
3077 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3078 struct rx_packet *tp;
3080 rxi_CallError(call, rx_BusyError);
3081 tp = rxi_SendCallAbort(call, np, 1, 0);
3082 MUTEX_EXIT(&call->lock);
3083 MUTEX_ENTER(&rx_refcnt_mutex);
3085 MUTEX_EXIT(&rx_refcnt_mutex);
3086 if (rx_stats_active)
3087 rx_atomic_inc(&rx_stats.nBusies);
3090 rxi_KeepAliveOn(call);
3092 /* Continuing call; do nothing here. */
3094 } else { /* we're the client */
3095 /* Ignore all incoming acknowledgements for calls in DALLY state */
3096 if (call && (call->state == RX_STATE_DALLY)
3097 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3098 if (rx_stats_active)
3099 rx_atomic_inc(&rx_stats.ignorePacketDally);
3100 #ifdef RX_ENABLE_LOCKS
3102 MUTEX_EXIT(&call->lock);
3105 MUTEX_ENTER(&rx_refcnt_mutex);
3107 MUTEX_EXIT(&rx_refcnt_mutex);
3111 /* Ignore anything that's not relevant to the current call. If there
3112 * isn't a current call, then no packet is relevant. */
3113 if (!call || (np->header.callNumber != currentCallNumber)) {
3114 if (rx_stats_active)
3115 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3116 #ifdef RX_ENABLE_LOCKS
3118 MUTEX_EXIT(&call->lock);
3121 MUTEX_ENTER(&rx_refcnt_mutex);
3123 MUTEX_EXIT(&rx_refcnt_mutex);
3126 /* If the service security object index stamped in the packet does not
3127 * match the connection's security index, ignore the packet */
3128 if (np->header.securityIndex != conn->securityIndex) {
3129 #ifdef RX_ENABLE_LOCKS
3130 MUTEX_EXIT(&call->lock);
3132 MUTEX_ENTER(&rx_refcnt_mutex);
3134 MUTEX_EXIT(&rx_refcnt_mutex);
3138 /* If we're receiving the response, then all transmit packets are
3139 * implicitly acknowledged. Get rid of them. */
3140 if (np->header.type == RX_PACKET_TYPE_DATA) {
3141 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3142 /* XXX Hack. Because we must release the global rx lock when
3143 * sending packets (osi_NetSend) we drop all acks while we're
3144 * traversing the tq in rxi_Start sending packets out because
3145 * packets may move to the freePacketQueue as result of being here!
3146 * So we drop these packets until we're safely out of the
3147 * traversing. Really ugly!
3148 * For fine grain RX locking, we set the acked field in the
3149 * packets and let rxi_Start remove them from the transmit queue.
3151 if (call->flags & RX_CALL_TQ_BUSY) {
3152 #ifdef RX_ENABLE_LOCKS
3153 rxi_SetAcksInTransmitQueue(call);
3155 MUTEX_ENTER(&rx_refcnt_mutex);
3157 MUTEX_EXIT(&rx_refcnt_mutex);
3158 return np; /* xmitting; drop packet */
3161 rxi_ClearTransmitQueue(call, 0);
3163 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3164 rxi_ClearTransmitQueue(call, 0);
3165 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3167 if (np->header.type == RX_PACKET_TYPE_ACK) {
3168 /* now check to see if this is an ack packet acknowledging that the
3169 * server actually *lost* some hard-acked data. If this happens we
3170 * ignore this packet, as it may indicate that the server restarted in
3171 * the middle of a call. It is also possible that this is an old ack
3172 * packet. We don't abort the connection in this case, because this
3173 * *might* just be an old ack packet. The right way to detect a server
3174 * restart in the midst of a call is to notice that the server epoch
3176 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3177 * XXX unacknowledged. I think that this is off-by-one, but
3178 * XXX I don't dare change it just yet, since it will
3179 * XXX interact badly with the server-restart detection
3180 * XXX code in receiveackpacket. */
3181 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3182 if (rx_stats_active)
3183 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3184 MUTEX_EXIT(&call->lock);
3185 MUTEX_ENTER(&rx_refcnt_mutex);
3187 MUTEX_EXIT(&rx_refcnt_mutex);
3191 } /* else not a data packet */
3194 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3195 /* Set remote user defined status from packet */
3196 call->remoteStatus = np->header.userStatus;
3198 /* Note the gap between the expected next packet and the actual
3199 * packet that arrived, when the new packet has a smaller serial number
3200 * than expected. Rioses frequently reorder packets all by themselves,
3201 * so this will be quite important with very large window sizes.
3202 * Skew is checked against 0 here to avoid any dependence on the type of
3203 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3205 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3206 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3207 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3209 MUTEX_ENTER(&conn->conn_data_lock);
3210 skew = conn->lastSerial - np->header.serial;
3211 conn->lastSerial = np->header.serial;
3212 MUTEX_EXIT(&conn->conn_data_lock);
3214 struct rx_peer *peer;
3216 if (skew > peer->inPacketSkew) {
3217 dpf(("*** In skew changed from %d to %d\n",
3218 peer->inPacketSkew, skew));
3219 peer->inPacketSkew = skew;
3223 /* Now do packet type-specific processing */
3224 switch (np->header.type) {
3225 case RX_PACKET_TYPE_DATA:
3226 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3229 case RX_PACKET_TYPE_ACK:
3230 /* Respond immediately to ack packets requesting acknowledgement
3232 if (np->header.flags & RX_REQUEST_ACK) {
3234 (void)rxi_SendCallAbort(call, 0, 1, 0);
3236 (void)rxi_SendAck(call, 0, np->header.serial,
3237 RX_ACK_PING_RESPONSE, 1);
3239 np = rxi_ReceiveAckPacket(call, np, 1);
3241 case RX_PACKET_TYPE_ABORT: {
3242 /* An abort packet: reset the call, passing the error up to the user. */
3243 /* What if error is zero? */
3244 /* What if the error is -1? the application will treat it as a timeout. */
3245 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3246 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3247 rxi_CallError(call, errdata);
3248 MUTEX_EXIT(&call->lock);
3249 MUTEX_ENTER(&rx_refcnt_mutex);
3251 MUTEX_EXIT(&rx_refcnt_mutex);
3252 return np; /* xmitting; drop packet */
3254 case RX_PACKET_TYPE_BUSY:
3257 case RX_PACKET_TYPE_ACKALL:
3258 /* All packets acknowledged, so we can drop all packets previously
3259 * readied for sending */
3260 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3261 /* XXX Hack. We because we can't release the global rx lock when
3262 * sending packets (osi_NetSend) we drop all ack pkts while we're
3263 * traversing the tq in rxi_Start sending packets out because
3264 * packets may move to the freePacketQueue as result of being
3265 * here! So we drop these packets until we're safely out of the
3266 * traversing. Really ugly!
3267 * For fine grain RX locking, we set the acked field in the packets
3268 * and let rxi_Start remove the packets from the transmit queue.
3270 if (call->flags & RX_CALL_TQ_BUSY) {
3271 #ifdef RX_ENABLE_LOCKS
3272 rxi_SetAcksInTransmitQueue(call);
3274 #else /* RX_ENABLE_LOCKS */
3275 MUTEX_EXIT(&call->lock);
3276 MUTEX_ENTER(&rx_refcnt_mutex);
3278 MUTEX_EXIT(&rx_refcnt_mutex);
3279 return np; /* xmitting; drop packet */
3280 #endif /* RX_ENABLE_LOCKS */
3282 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3283 rxi_ClearTransmitQueue(call, 0);
3284 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3287 /* Should not reach here, unless the peer is broken: send an abort
3289 rxi_CallError(call, RX_PROTOCOL_ERROR);
3290 np = rxi_SendCallAbort(call, np, 1, 0);
3293 /* Note when this last legitimate packet was received, for keep-alive
3294 * processing. Note, we delay getting the time until now in the hope that
3295 * the packet will be delivered to the user before any get time is required
3296 * (if not, then the time won't actually be re-evaluated here). */
3297 call->lastReceiveTime = clock_Sec();
3298 MUTEX_EXIT(&call->lock);
3299 MUTEX_ENTER(&rx_refcnt_mutex);
3301 MUTEX_EXIT(&rx_refcnt_mutex);
3305 /* return true if this is an "interesting" connection from the point of view
3306 of someone trying to debug the system */
3308 rxi_IsConnInteresting(struct rx_connection *aconn)
3311 struct rx_call *tcall;
3313 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3316 for (i = 0; i < RX_MAXCALLS; i++) {
3317 tcall = aconn->call[i];
3319 if ((tcall->state == RX_STATE_PRECALL)
3320 || (tcall->state == RX_STATE_ACTIVE))
3322 if ((tcall->mode == RX_MODE_SENDING)
3323 || (tcall->mode == RX_MODE_RECEIVING))
3331 /* if this is one of the last few packets AND it wouldn't be used by the
3332 receiving call to immediately satisfy a read request, then drop it on
3333 the floor, since accepting it might prevent a lock-holding thread from
3334 making progress in its reading. If a call has been cleared while in
3335 the precall state then ignore all subsequent packets until the call
3336 is assigned to a thread. */
3339 TooLow(struct rx_packet *ap, struct rx_call *acall)
3343 MUTEX_ENTER(&rx_quota_mutex);
3344 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3345 && (acall->state == RX_STATE_PRECALL))
3346 || ((rx_nFreePackets < rxi_dataQuota + 2)
3347 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3348 && (acall->flags & RX_CALL_READER_WAIT)))) {
3351 MUTEX_EXIT(&rx_quota_mutex);
3357 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3359 struct rx_connection *conn = arg1;
3360 struct rx_call *acall = arg2;
3361 struct rx_call *call = acall;
3362 struct clock when, now;
3365 MUTEX_ENTER(&conn->conn_data_lock);
3366 conn->checkReachEvent = NULL;
3367 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3369 MUTEX_ENTER(&rx_refcnt_mutex);
3371 MUTEX_EXIT(&rx_refcnt_mutex);
3373 MUTEX_EXIT(&conn->conn_data_lock);
3377 MUTEX_ENTER(&conn->conn_call_lock);
3378 MUTEX_ENTER(&conn->conn_data_lock);
3379 for (i = 0; i < RX_MAXCALLS; i++) {
3380 struct rx_call *tc = conn->call[i];
3381 if (tc && tc->state == RX_STATE_PRECALL) {
3387 /* Indicate that rxi_CheckReachEvent is no longer running by
3388 * clearing the flag. Must be atomic under conn_data_lock to
3389 * avoid a new call slipping by: rxi_CheckConnReach holds
3390 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3392 conn->flags &= ~RX_CONN_ATTACHWAIT;
3393 MUTEX_EXIT(&conn->conn_data_lock);
3394 MUTEX_EXIT(&conn->conn_call_lock);
3399 MUTEX_ENTER(&call->lock);
3400 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3402 MUTEX_EXIT(&call->lock);
3404 clock_GetTime(&now);
3406 when.sec += RX_CHECKREACH_TIMEOUT;
3407 MUTEX_ENTER(&conn->conn_data_lock);
3408 if (!conn->checkReachEvent) {
3409 MUTEX_ENTER(&rx_refcnt_mutex);
3411 MUTEX_EXIT(&rx_refcnt_mutex);
3412 conn->checkReachEvent =
3413 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3416 MUTEX_EXIT(&conn->conn_data_lock);
3422 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3424 struct rx_service *service = conn->service;
3425 struct rx_peer *peer = conn->peer;
3426 afs_uint32 now, lastReach;
3428 if (service->checkReach == 0)
3432 MUTEX_ENTER(&peer->peer_lock);
3433 lastReach = peer->lastReachTime;
3434 MUTEX_EXIT(&peer->peer_lock);
3435 if (now - lastReach < RX_CHECKREACH_TTL)
3438 MUTEX_ENTER(&conn->conn_data_lock);
3439 if (conn->flags & RX_CONN_ATTACHWAIT) {
3440 MUTEX_EXIT(&conn->conn_data_lock);
3443 conn->flags |= RX_CONN_ATTACHWAIT;
3444 MUTEX_EXIT(&conn->conn_data_lock);
3445 if (!conn->checkReachEvent)
3446 rxi_CheckReachEvent(NULL, conn, call);
3451 /* try to attach call, if authentication is complete */
3453 TryAttach(struct rx_call *acall, osi_socket socket,
3454 int *tnop, struct rx_call **newcallp,
3457 struct rx_connection *conn = acall->conn;
3459 if (conn->type == RX_SERVER_CONNECTION
3460 && acall->state == RX_STATE_PRECALL) {
3461 /* Don't attach until we have any req'd. authentication. */
3462 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3463 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3464 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3465 /* Note: this does not necessarily succeed; there
3466 * may not any proc available
3469 rxi_ChallengeOn(acall->conn);
3474 /* A data packet has been received off the interface. This packet is
3475 * appropriate to the call (the call is in the right state, etc.). This
3476 * routine can return a packet to the caller, for re-use */
3479 rxi_ReceiveDataPacket(struct rx_call *call,
3480 struct rx_packet *np, int istack,
3481 osi_socket socket, afs_uint32 host, u_short port,
3482 int *tnop, struct rx_call **newcallp)
3484 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3489 afs_uint32 serial=0, flags=0;
3491 struct rx_packet *tnp;
3492 struct clock when, now;
3493 if (rx_stats_active)
3494 rx_atomic_inc(&rx_stats.dataPacketsRead);
3497 /* If there are no packet buffers, drop this new packet, unless we can find
3498 * packet buffers from inactive calls */
3500 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3501 MUTEX_ENTER(&rx_freePktQ_lock);
3502 rxi_NeedMorePackets = TRUE;
3503 MUTEX_EXIT(&rx_freePktQ_lock);
3504 if (rx_stats_active)
3505 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3506 call->rprev = np->header.serial;
3507 rxi_calltrace(RX_TRACE_DROP, call);
3508 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3510 rxi_ClearReceiveQueue(call);
3511 clock_GetTime(&now);
3513 clock_Add(&when, &rx_softAckDelay);
3514 if (!call->delayedAckEvent
3515 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3516 rxevent_Cancel(call->delayedAckEvent, call,
3517 RX_CALL_REFCOUNT_DELAY);
3518 MUTEX_ENTER(&rx_refcnt_mutex);
3519 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3520 MUTEX_EXIT(&rx_refcnt_mutex);
3522 call->delayedAckEvent =
3523 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3525 /* we've damaged this call already, might as well do it in. */
3531 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3532 * packet is one of several packets transmitted as a single
3533 * datagram. Do not send any soft or hard acks until all packets
3534 * in a jumbogram have been processed. Send negative acks right away.
3536 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3537 /* tnp is non-null when there are more packets in the
3538 * current jumbo gram */
3545 seq = np->header.seq;
3546 serial = np->header.serial;
3547 flags = np->header.flags;
3549 /* If the call is in an error state, send an abort message */
3551 return rxi_SendCallAbort(call, np, istack, 0);
3553 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3554 * AFS 3.5 jumbogram. */
3555 if (flags & RX_JUMBO_PACKET) {
3556 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3561 if (np->header.spare != 0) {
3562 MUTEX_ENTER(&call->conn->conn_data_lock);
3563 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3564 MUTEX_EXIT(&call->conn->conn_data_lock);
3567 /* The usual case is that this is the expected next packet */
3568 if (seq == call->rnext) {
3570 /* Check to make sure it is not a duplicate of one already queued */
3571 if (queue_IsNotEmpty(&call->rq)
3572 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3573 if (rx_stats_active)
3574 rx_atomic_inc(&rx_stats.dupPacketsRead);
3575 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3576 rxevent_Cancel(call->delayedAckEvent, call,
3577 RX_CALL_REFCOUNT_DELAY);
3578 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3584 /* It's the next packet. Stick it on the receive queue
3585 * for this call. Set newPackets to make sure we wake
3586 * the reader once all packets have been processed */
3587 #ifdef RX_TRACK_PACKETS
3588 np->flags |= RX_PKTFLAG_RQ;
3590 queue_Prepend(&call->rq, np);
3591 #ifdef RXDEBUG_PACKET
3593 #endif /* RXDEBUG_PACKET */
3595 np = NULL; /* We can't use this anymore */
3598 /* If an ack is requested then set a flag to make sure we
3599 * send an acknowledgement for this packet */
3600 if (flags & RX_REQUEST_ACK) {
3601 ackNeeded = RX_ACK_REQUESTED;
3604 /* Keep track of whether we have received the last packet */
3605 if (flags & RX_LAST_PACKET) {
3606 call->flags |= RX_CALL_HAVE_LAST;
3610 /* Check whether we have all of the packets for this call */
3611 if (call->flags & RX_CALL_HAVE_LAST) {
3612 afs_uint32 tseq; /* temporary sequence number */
3613 struct rx_packet *tp; /* Temporary packet pointer */
3614 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3616 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3617 if (tseq != tp->header.seq)
3619 if (tp->header.flags & RX_LAST_PACKET) {
3620 call->flags |= RX_CALL_RECEIVE_DONE;
3627 /* Provide asynchronous notification for those who want it
3628 * (e.g. multi rx) */
3629 if (call->arrivalProc) {
3630 (*call->arrivalProc) (call, call->arrivalProcHandle,
3631 call->arrivalProcArg);
3632 call->arrivalProc = (void (*)())0;
3635 /* Update last packet received */
3638 /* If there is no server process serving this call, grab
3639 * one, if available. We only need to do this once. If a
3640 * server thread is available, this thread becomes a server
3641 * thread and the server thread becomes a listener thread. */
3643 TryAttach(call, socket, tnop, newcallp, 0);
3646 /* This is not the expected next packet. */
3648 /* Determine whether this is a new or old packet, and if it's
3649 * a new one, whether it fits into the current receive window.
3650 * Also figure out whether the packet was delivered in sequence.
3651 * We use the prev variable to determine whether the new packet
3652 * is the successor of its immediate predecessor in the
3653 * receive queue, and the missing flag to determine whether
3654 * any of this packets predecessors are missing. */
3656 afs_uint32 prev; /* "Previous packet" sequence number */
3657 struct rx_packet *tp; /* Temporary packet pointer */
3658 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3659 int missing; /* Are any predecessors missing? */
3661 /* If the new packet's sequence number has been sent to the
3662 * application already, then this is a duplicate */
3663 if (seq < call->rnext) {
3664 if (rx_stats_active)
3665 rx_atomic_inc(&rx_stats.dupPacketsRead);
3666 rxevent_Cancel(call->delayedAckEvent, call,
3667 RX_CALL_REFCOUNT_DELAY);
3668 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3674 /* If the sequence number is greater than what can be
3675 * accomodated by the current window, then send a negative
3676 * acknowledge and drop the packet */
3677 if ((call->rnext + call->rwind) <= seq) {
3678 rxevent_Cancel(call->delayedAckEvent, call,
3679 RX_CALL_REFCOUNT_DELAY);
3680 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3687 /* Look for the packet in the queue of old received packets */
3688 for (prev = call->rnext - 1, missing =
3689 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3690 /*Check for duplicate packet */
3691 if (seq == tp->header.seq) {
3692 if (rx_stats_active)
3693 rx_atomic_inc(&rx_stats.dupPacketsRead);
3694 rxevent_Cancel(call->delayedAckEvent, call,
3695 RX_CALL_REFCOUNT_DELAY);
3696 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3702 /* If we find a higher sequence packet, break out and
3703 * insert the new packet here. */
3704 if (seq < tp->header.seq)
3706 /* Check for missing packet */
3707 if (tp->header.seq != prev + 1) {
3711 prev = tp->header.seq;
3714 /* Keep track of whether we have received the last packet. */
3715 if (flags & RX_LAST_PACKET) {
3716 call->flags |= RX_CALL_HAVE_LAST;
3719 /* It's within the window: add it to the the receive queue.
3720 * tp is left by the previous loop either pointing at the
3721 * packet before which to insert the new packet, or at the
3722 * queue head if the queue is empty or the packet should be
3724 #ifdef RX_TRACK_PACKETS
3725 np->flags |= RX_PKTFLAG_RQ;
3727 #ifdef RXDEBUG_PACKET
3729 #endif /* RXDEBUG_PACKET */
3730 queue_InsertBefore(tp, np);
3734 /* Check whether we have all of the packets for this call */
3735 if ((call->flags & RX_CALL_HAVE_LAST)
3736 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3737 afs_uint32 tseq; /* temporary sequence number */
3740 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3741 if (tseq != tp->header.seq)
3743 if (tp->header.flags & RX_LAST_PACKET) {
3744 call->flags |= RX_CALL_RECEIVE_DONE;
3751 /* We need to send an ack of the packet is out of sequence,
3752 * or if an ack was requested by the peer. */
3753 if (seq != prev + 1 || missing) {
3754 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3755 } else if (flags & RX_REQUEST_ACK) {
3756 ackNeeded = RX_ACK_REQUESTED;
3759 /* Acknowledge the last packet for each call */
3760 if (flags & RX_LAST_PACKET) {
3771 * If the receiver is waiting for an iovec, fill the iovec
3772 * using the data from the receive queue */
3773 if (call->flags & RX_CALL_IOVEC_WAIT) {
3774 didHardAck = rxi_FillReadVec(call, serial);
3775 /* the call may have been aborted */
3784 /* Wakeup the reader if any */
3785 if ((call->flags & RX_CALL_READER_WAIT)
3786 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3787 || (call->iovNext >= call->iovMax)
3788 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3789 call->flags &= ~RX_CALL_READER_WAIT;
3790 #ifdef RX_ENABLE_LOCKS
3791 CV_BROADCAST(&call->cv_rq);
3793 osi_rxWakeup(&call->rq);
3799 * Send an ack when requested by the peer, or once every
3800 * rxi_SoftAckRate packets until the last packet has been
3801 * received. Always send a soft ack for the last packet in
3802 * the server's reply.
3804 * If we have received all of the packets for the call
3805 * immediately send an RX_PACKET_TYPE_ACKALL packet so that
3806 * the peer can empty its packet queue and cancel all resend
3809 if (call->flags & RX_CALL_RECEIVE_DONE) {
3810 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3811 rxi_AckAll(NULL, call, 0);
3812 } else if (ackNeeded) {
3813 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3814 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3815 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3816 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3817 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3818 } else if (call->nSoftAcks) {
3819 clock_GetTime(&now);
3821 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3822 clock_Add(&when, &rx_lastAckDelay);
3824 clock_Add(&when, &rx_softAckDelay);
3826 if (!call->delayedAckEvent
3827 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3828 rxevent_Cancel(call->delayedAckEvent, call,
3829 RX_CALL_REFCOUNT_DELAY);
3830 MUTEX_ENTER(&rx_refcnt_mutex);
3831 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3832 MUTEX_EXIT(&rx_refcnt_mutex);
3833 call->delayedAckEvent =
3834 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3842 static void rxi_ComputeRate();
3846 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3848 struct rx_peer *peer = conn->peer;
3850 MUTEX_ENTER(&peer->peer_lock);
3851 peer->lastReachTime = clock_Sec();
3852 MUTEX_EXIT(&peer->peer_lock);
3854 MUTEX_ENTER(&conn->conn_data_lock);
3855 if (conn->flags & RX_CONN_ATTACHWAIT) {
3858 conn->flags &= ~RX_CONN_ATTACHWAIT;
3859 MUTEX_EXIT(&conn->conn_data_lock);
3861 for (i = 0; i < RX_MAXCALLS; i++) {
3862 struct rx_call *call = conn->call[i];
3865 MUTEX_ENTER(&call->lock);
3866 /* tnop can be null if newcallp is null */
3867 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3869 MUTEX_EXIT(&call->lock);
3873 MUTEX_EXIT(&conn->conn_data_lock);
3876 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
3878 rx_ack_reason(int reason)
3881 case RX_ACK_REQUESTED:
3883 case RX_ACK_DUPLICATE:
3885 case RX_ACK_OUT_OF_SEQUENCE:
3887 case RX_ACK_EXCEEDS_WINDOW:
3889 case RX_ACK_NOSPACE:
3893 case RX_ACK_PING_RESPONSE:
3906 /* rxi_ComputePeerNetStats
3908 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3909 * estimates (like RTT and throughput) based on ack packets. Caller
3910 * must ensure that the packet in question is the right one (i.e.
3911 * serial number matches).
3914 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3915 struct rx_ackPacket *ap, struct rx_packet *np,
3918 struct rx_peer *peer = call->conn->peer;
3920 /* Use RTT if not delayed by client and
3921 * ignore packets that were retransmitted. */
3922 if (!(p->flags & RX_PKTFLAG_ACKED) &&
3923 ap->reason != RX_ACK_DELAY &&
3924 clock_Eq(&p->timeSent, &p->firstSent))
3925 rxi_ComputeRoundTripTime(p, &p->timeSent, peer, now);
3927 rxi_ComputeRate(peer, call, p, np, ap->reason);
3931 /* The real smarts of the whole thing. */
3933 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
3936 struct rx_ackPacket *ap;
3938 struct rx_packet *tp;
3939 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3940 struct rx_connection *conn = call->conn;
3941 struct rx_peer *peer = conn->peer;
3942 struct clock now; /* Current time, for RTT calculations */
3945 /* because there are CM's that are bogus, sending weird values for this. */
3946 afs_uint32 skew = 0;
3951 int newAckCount = 0;
3952 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3953 int pktsize = 0; /* Set if we need to update the peer mtu */
3954 int conn_data_locked = 0;
3956 if (rx_stats_active)
3957 rx_atomic_inc(&rx_stats.ackPacketsRead);
3958 ap = (struct rx_ackPacket *)rx_DataOf(np);
3959 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3961 return np; /* truncated ack packet */
3963 /* depends on ack packet struct */
3964 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3965 first = ntohl(ap->firstPacket);
3966 serial = ntohl(ap->serial);
3967 /* temporarily disabled -- needs to degrade over time
3968 * skew = ntohs(ap->maxSkew); */
3970 /* Ignore ack packets received out of order */
3971 if (first < call->tfirst) {
3975 if (np->header.flags & RX_SLOW_START_OK) {
3976 call->flags |= RX_CALL_SLOW_START_OK;
3979 if (ap->reason == RX_ACK_PING_RESPONSE)
3980 rxi_UpdatePeerReach(conn, call);
3982 if (conn->lastPacketSizeSeq) {
3983 MUTEX_ENTER(&conn->conn_data_lock);
3984 conn_data_locked = 1;
3985 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
3986 pktsize = conn->lastPacketSize;
3987 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
3990 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
3991 if (!conn_data_locked) {
3992 MUTEX_ENTER(&conn->conn_data_lock);
3993 conn_data_locked = 1;
3995 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
3996 /* process mtu ping ack */
3997 pktsize = conn->lastPingSize;
3998 conn->lastPingSizeSer = conn->lastPingSize = 0;
4002 if (conn_data_locked) {
4003 MUTEX_EXIT(&conn->conn_data_lock);
4004 conn_data_locked = 0;
4008 if (rxdebug_active) {
4012 len = _snprintf(msg, sizeof(msg),
4013 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
4014 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4015 ntohl(ap->serial), ntohl(ap->previousPacket),
4016 (unsigned int)np->header.seq, (unsigned int)skew,
4017 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
4021 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4022 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4026 OutputDebugString(msg);
4028 #else /* AFS_NT40_ENV */
4031 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
4032 ap->reason, ntohl(ap->previousPacket),
4033 (unsigned int)np->header.seq, (unsigned int)serial,
4034 (unsigned int)skew, ntohl(ap->firstPacket));
4037 for (offset = 0; offset < nAcks; offset++)
4038 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4043 #endif /* AFS_NT40_ENV */
4046 MUTEX_ENTER(&peer->peer_lock);
4049 * Start somewhere. Can't assume we can send what we can receive,
4050 * but we are clearly receiving.
4052 if (!peer->maxPacketSize)
4053 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4055 if (pktsize > peer->maxPacketSize) {
4056 peer->maxPacketSize = pktsize;
4057 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4058 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4059 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4060 rxi_ScheduleGrowMTUEvent(call, 1);
4065 /* Update the outgoing packet skew value to the latest value of
4066 * the peer's incoming packet skew value. The ack packet, of
4067 * course, could arrive out of order, but that won't affect things
4069 peer->outPacketSkew = skew;
4071 /* Check for packets that no longer need to be transmitted, and
4072 * discard them. This only applies to packets positively
4073 * acknowledged as having been sent to the peer's upper level.
4074 * All other packets must be retained. So only packets with
4075 * sequence numbers < ap->firstPacket are candidates. */
4077 clock_GetTime(&now);
4079 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
4080 if (tp->header.seq >= first)
4082 call->tfirst = tp->header.seq + 1;
4083 rxi_ComputePeerNetStats(call, tp, ap, np, &now);
4084 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4087 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4088 /* XXX Hack. Because we have to release the global rx lock when sending
4089 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4090 * in rxi_Start sending packets out because packets may move to the
4091 * freePacketQueue as result of being here! So we drop these packets until
4092 * we're safely out of the traversing. Really ugly!
4093 * To make it even uglier, if we're using fine grain locking, we can
4094 * set the ack bits in the packets and have rxi_Start remove the packets
4095 * when it's done transmitting.
4097 if (call->flags & RX_CALL_TQ_BUSY) {
4098 #ifdef RX_ENABLE_LOCKS
4099 tp->flags |= RX_PKTFLAG_ACKED;
4100 call->flags |= RX_CALL_TQ_SOME_ACKED;
4101 #else /* RX_ENABLE_LOCKS */
4103 #endif /* RX_ENABLE_LOCKS */
4105 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4108 #ifdef RX_TRACK_PACKETS
4109 tp->flags &= ~RX_PKTFLAG_TQ;
4111 #ifdef RXDEBUG_PACKET
4113 #endif /* RXDEBUG_PACKET */
4114 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4119 /* Give rate detector a chance to respond to ping requests */
4120 if (ap->reason == RX_ACK_PING_RESPONSE) {
4121 rxi_ComputeRate(peer, call, 0, np, ap->reason);
4125 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4127 /* Now go through explicit acks/nacks and record the results in
4128 * the waiting packets. These are packets that can't be released
4129 * yet, even with a positive acknowledge. This positive
4130 * acknowledge only means the packet has been received by the
4131 * peer, not that it will be retained long enough to be sent to
4132 * the peer's upper level. In addition, reset the transmit timers
4133 * of any missing packets (those packets that must be missing
4134 * because this packet was out of sequence) */
4136 call->nSoftAcked = 0;
4137 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
4138 /* Update round trip time if the ack was stimulated on receipt
4140 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4141 #ifdef RX_ENABLE_LOCKS
4142 if (tp->header.seq >= first)
4143 #endif /* RX_ENABLE_LOCKS */
4144 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4145 rxi_ComputePeerNetStats(call, tp, ap, np, &now);
4147 /* Set the acknowledge flag per packet based on the
4148 * information in the ack packet. An acknowlegded packet can
4149 * be downgraded when the server has discarded a packet it
4150 * soacked previously, or when an ack packet is received
4151 * out of sequence. */
4152 if (tp->header.seq < first) {
4153 /* Implicit ack information */
4154 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4157 tp->flags |= RX_PKTFLAG_ACKED;
4158 } else if (tp->header.seq < first + nAcks) {
4159 /* Explicit ack information: set it in the packet appropriately */
4160 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4161 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4163 tp->flags |= RX_PKTFLAG_ACKED;
4170 } else /* RX_ACK_TYPE_NACK */ {
4171 tp->flags &= ~RX_PKTFLAG_ACKED;
4175 if (tp->flags & RX_PKTFLAG_ACKED) {
4176 tp->flags &= ~RX_PKTFLAG_ACKED;
4182 * Following the suggestion of Phil Kern, we back off the peer's
4183 * timeout value for future packets until a successful response
4184 * is received for an initial transmission.
4186 if (missing && !peer->backedOff) {
4187 struct clock c = peer->timeout;
4188 struct clock max_to = {3, 0};
4190 clock_Add(&peer->timeout, &c);
4191 if (clock_Gt(&peer->timeout, &max_to))
4192 peer->timeout = max_to;
4193 peer->backedOff = 1;
4196 /* If packet isn't yet acked, and it has been transmitted at least
4197 * once, reset retransmit time using latest timeout
4198 * ie, this should readjust the retransmit timer for all outstanding
4199 * packets... So we don't just retransmit when we should know better*/
4201 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
4202 tp->retryTime = tp->timeSent;
4203 clock_Add(&tp->retryTime, &peer->timeout);
4204 /* shift by eight because one quarter-sec ~ 256 milliseconds */
4205 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
4209 /* If the window has been extended by this acknowledge packet,
4210 * then wakeup a sender waiting in alloc for window space, or try
4211 * sending packets now, if he's been sitting on packets due to
4212 * lack of window space */
4213 if (call->tnext < (call->tfirst + call->twind)) {
4214 #ifdef RX_ENABLE_LOCKS
4215 CV_SIGNAL(&call->cv_twind);
4217 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4218 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4219 osi_rxWakeup(&call->twind);
4222 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4223 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4227 /* if the ack packet has a receivelen field hanging off it,
4228 * update our state */
4229 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4232 /* If the ack packet has a "recommended" size that is less than
4233 * what I am using now, reduce my size to match */
4234 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4235 (int)sizeof(afs_int32), &tSize);
4236 tSize = (afs_uint32) ntohl(tSize);
4237 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4239 /* Get the maximum packet size to send to this peer */
4240 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4242 tSize = (afs_uint32) ntohl(tSize);
4243 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4244 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4246 /* sanity check - peer might have restarted with different params.
4247 * If peer says "send less", dammit, send less... Peer should never
4248 * be unable to accept packets of the size that prior AFS versions would
4249 * send without asking. */
4250 if (peer->maxMTU != tSize) {
4251 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4253 peer->maxMTU = tSize;
4254 peer->MTU = MIN(tSize, peer->MTU);
4255 call->MTU = MIN(call->MTU, tSize);
4258 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4261 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4262 (int)sizeof(afs_int32), &tSize);
4263 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4264 if (tSize < call->twind) { /* smaller than our send */
4265 call->twind = tSize; /* window, we must send less... */
4266 call->ssthresh = MIN(call->twind, call->ssthresh);
4267 call->conn->twind[call->channel] = call->twind;
4270 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4271 * network MTU confused with the loopback MTU. Calculate the
4272 * maximum MTU here for use in the slow start code below.
4274 /* Did peer restart with older RX version? */
4275 if (peer->maxDgramPackets > 1) {
4276 peer->maxDgramPackets = 1;
4278 } else if (np->length >=
4279 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4282 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4283 sizeof(afs_int32), &tSize);
4284 tSize = (afs_uint32) ntohl(tSize);
4286 * As of AFS 3.5 we set the send window to match the receive window.
4288 if (tSize < call->twind) {
4289 call->twind = tSize;
4290 call->conn->twind[call->channel] = call->twind;
4291 call->ssthresh = MIN(call->twind, call->ssthresh);
4292 } else if (tSize > call->twind) {
4293 call->twind = tSize;
4294 call->conn->twind[call->channel] = call->twind;
4298 * As of AFS 3.5, a jumbogram is more than one fixed size
4299 * packet transmitted in a single UDP datagram. If the remote
4300 * MTU is smaller than our local MTU then never send a datagram
4301 * larger than the natural MTU.
4304 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4305 (int)sizeof(afs_int32), &tSize);
4306 maxDgramPackets = (afs_uint32) ntohl(tSize);
4307 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4309 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4310 maxDgramPackets = MIN(maxDgramPackets, tSize);
4311 if (maxDgramPackets > 1) {
4312 peer->maxDgramPackets = maxDgramPackets;
4313 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4315 peer->maxDgramPackets = 1;
4316 call->MTU = peer->natMTU;
4318 } else if (peer->maxDgramPackets > 1) {
4319 /* Restarted with lower version of RX */
4320 peer->maxDgramPackets = 1;
4322 } else if (peer->maxDgramPackets > 1
4323 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4324 /* Restarted with lower version of RX */
4325 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4326 peer->natMTU = OLD_MAX_PACKET_SIZE;
4327 peer->MTU = OLD_MAX_PACKET_SIZE;
4328 peer->maxDgramPackets = 1;
4329 peer->nDgramPackets = 1;
4331 call->MTU = OLD_MAX_PACKET_SIZE;
4336 * Calculate how many datagrams were successfully received after
4337 * the first missing packet and adjust the negative ack counter
4342 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4343 if (call->nNacks < nNacked) {
4344 call->nNacks = nNacked;
4347 call->nAcks += newAckCount;
4351 if (call->flags & RX_CALL_FAST_RECOVER) {
4353 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4355 call->flags &= ~RX_CALL_FAST_RECOVER;
4356 call->cwind = call->nextCwind;
4357 call->nextCwind = 0;
4360 call->nCwindAcks = 0;
4361 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4362 /* Three negative acks in a row trigger congestion recovery */
4363 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4364 MUTEX_EXIT(&peer->peer_lock);
4365 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4366 /* someone else is waiting to start recovery */
4369 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4370 rxi_WaitforTQBusy(call);
4371 MUTEX_ENTER(&peer->peer_lock);
4372 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4373 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4374 call->flags |= RX_CALL_FAST_RECOVER;
4375 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4377 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4378 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4379 call->nextCwind = call->ssthresh;
4382 peer->MTU = call->MTU;
4383 peer->cwind = call->nextCwind;
4384 peer->nDgramPackets = call->nDgramPackets;
4386 call->congestSeq = peer->congestSeq;
4387 /* Reset the resend times on the packets that were nacked
4388 * so we will retransmit as soon as the window permits*/
4389 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4391 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4392 clock_Zero(&tp->retryTime);
4394 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4399 /* If cwind is smaller than ssthresh, then increase
4400 * the window one packet for each ack we receive (exponential
4402 * If cwind is greater than or equal to ssthresh then increase
4403 * the congestion window by one packet for each cwind acks we
4404 * receive (linear growth). */
4405 if (call->cwind < call->ssthresh) {
4407 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4408 call->nCwindAcks = 0;
4410 call->nCwindAcks += newAckCount;
4411 if (call->nCwindAcks >= call->cwind) {
4412 call->nCwindAcks = 0;
4413 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4417 * If we have received several acknowledgements in a row then
4418 * it is time to increase the size of our datagrams
4420 if ((int)call->nAcks > rx_nDgramThreshold) {
4421 if (peer->maxDgramPackets > 1) {
4422 if (call->nDgramPackets < peer->maxDgramPackets) {
4423 call->nDgramPackets++;
4425 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4426 } else if (call->MTU < peer->maxMTU) {
4427 /* don't upgrade if we can't handle it */
4428 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4429 call->MTU = peer->ifMTU;
4431 call->MTU += peer->natMTU;
4432 call->MTU = MIN(call->MTU, peer->maxMTU);
4439 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4441 /* Servers need to hold the call until all response packets have
4442 * been acknowledged. Soft acks are good enough since clients
4443 * are not allowed to clear their receive queues. */
4444 if (call->state == RX_STATE_HOLD
4445 && call->tfirst + call->nSoftAcked >= call->tnext) {
4446 call->state = RX_STATE_DALLY;
4447 rxi_ClearTransmitQueue(call, 0);
4448 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4449 } else if (!queue_IsEmpty(&call->tq)) {
4450 rxi_Start(0, call, 0, istack);
4455 /* Received a response to a challenge packet */
4457 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4458 struct rx_packet *np, int istack)
4462 /* Ignore the packet if we're the client */
4463 if (conn->type == RX_CLIENT_CONNECTION)
4466 /* If already authenticated, ignore the packet (it's probably a retry) */
4467 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4470 /* Otherwise, have the security object evaluate the response packet */
4471 error = RXS_CheckResponse(conn->securityObject, conn, np);
4473 /* If the response is invalid, reset the connection, sending
4474 * an abort to the peer */
4478 rxi_ConnectionError(conn, error);
4479 MUTEX_ENTER(&conn->conn_data_lock);
4480 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4481 MUTEX_EXIT(&conn->conn_data_lock);
4484 /* If the response is valid, any calls waiting to attach
4485 * servers can now do so */
4488 for (i = 0; i < RX_MAXCALLS; i++) {
4489 struct rx_call *call = conn->call[i];
4491 MUTEX_ENTER(&call->lock);
4492 if (call->state == RX_STATE_PRECALL)
4493 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4494 /* tnop can be null if newcallp is null */
4495 MUTEX_EXIT(&call->lock);
4499 /* Update the peer reachability information, just in case
4500 * some calls went into attach-wait while we were waiting
4501 * for authentication..
4503 rxi_UpdatePeerReach(conn, NULL);
4508 /* A client has received an authentication challenge: the security
4509 * object is asked to cough up a respectable response packet to send
4510 * back to the server. The server is responsible for retrying the
4511 * challenge if it fails to get a response. */
4514 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4515 struct rx_packet *np, int istack)
4519 /* Ignore the challenge if we're the server */
4520 if (conn->type == RX_SERVER_CONNECTION)
4523 /* Ignore the challenge if the connection is otherwise idle; someone's
4524 * trying to use us as an oracle. */
4525 if (!rxi_HasActiveCalls(conn))
4528 /* Send the security object the challenge packet. It is expected to fill
4529 * in the response. */
4530 error = RXS_GetResponse(conn->securityObject, conn, np);
4532 /* If the security object is unable to return a valid response, reset the
4533 * connection and send an abort to the peer. Otherwise send the response
4534 * packet to the peer connection. */
4536 rxi_ConnectionError(conn, error);
4537 MUTEX_ENTER(&conn->conn_data_lock);
4538 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4539 MUTEX_EXIT(&conn->conn_data_lock);
4541 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4542 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4548 /* Find an available server process to service the current request in
4549 * the given call structure. If one isn't available, queue up this
4550 * call so it eventually gets one */
4552 rxi_AttachServerProc(struct rx_call *call,
4553 osi_socket socket, int *tnop,
4554 struct rx_call **newcallp)
4556 struct rx_serverQueueEntry *sq;
4557 struct rx_service *service = call->conn->service;
4560 /* May already be attached */
4561 if (call->state == RX_STATE_ACTIVE)
4564 MUTEX_ENTER(&rx_serverPool_lock);
4566 haveQuota = QuotaOK(service);
4567 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4568 /* If there are no processes available to service this call,
4569 * put the call on the incoming call queue (unless it's
4570 * already on the queue).
4572 #ifdef RX_ENABLE_LOCKS
4574 ReturnToServerPool(service);
4575 #endif /* RX_ENABLE_LOCKS */
4577 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4578 call->flags |= RX_CALL_WAIT_PROC;
4579 rx_atomic_inc(&rx_nWaiting);
4580 rx_atomic_inc(&rx_nWaited);
4581 rxi_calltrace(RX_CALL_ARRIVAL, call);
4582 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4583 queue_Append(&rx_incomingCallQueue, call);
4586 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4588 /* If hot threads are enabled, and both newcallp and sq->socketp
4589 * are non-null, then this thread will process the call, and the
4590 * idle server thread will start listening on this threads socket.
4593 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4596 *sq->socketp = socket;
4597 clock_GetTime(&call->startTime);
4598 MUTEX_ENTER(&rx_refcnt_mutex);
4599 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4600 MUTEX_EXIT(&rx_refcnt_mutex);
4604 if (call->flags & RX_CALL_WAIT_PROC) {
4605 /* Conservative: I don't think this should happen */
4606 call->flags &= ~RX_CALL_WAIT_PROC;
4607 if (queue_IsOnQueue(call)) {
4610 rx_atomic_dec(&rx_nWaiting);
4613 call->state = RX_STATE_ACTIVE;
4614 call->mode = RX_MODE_RECEIVING;
4615 #ifdef RX_KERNEL_TRACE
4617 int glockOwner = ISAFS_GLOCK();
4620 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4621 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4627 if (call->flags & RX_CALL_CLEARED) {
4628 /* send an ack now to start the packet flow up again */
4629 call->flags &= ~RX_CALL_CLEARED;
4630 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4632 #ifdef RX_ENABLE_LOCKS
4635 service->nRequestsRunning++;
4636 MUTEX_ENTER(&rx_quota_mutex);
4637 if (service->nRequestsRunning <= service->minProcs)
4640 MUTEX_EXIT(&rx_quota_mutex);
4644 MUTEX_EXIT(&rx_serverPool_lock);
4647 /* Delay the sending of an acknowledge event for a short while, while
4648 * a new call is being prepared (in the case of a client) or a reply
4649 * is being prepared (in the case of a server). Rather than sending
4650 * an ack packet, an ACKALL packet is sent. */
4652 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4654 #ifdef RX_ENABLE_LOCKS
4656 MUTEX_ENTER(&call->lock);
4657 call->delayedAckEvent = NULL;
4658 MUTEX_ENTER(&rx_refcnt_mutex);
4659 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4660 MUTEX_EXIT(&rx_refcnt_mutex);
4662 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4663 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4665 MUTEX_EXIT(&call->lock);
4666 #else /* RX_ENABLE_LOCKS */
4668 call->delayedAckEvent = NULL;
4669 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4670 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4671 #endif /* RX_ENABLE_LOCKS */
4675 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4677 struct rx_call *call = arg1;
4678 #ifdef RX_ENABLE_LOCKS
4680 MUTEX_ENTER(&call->lock);
4681 if (event == call->delayedAckEvent)
4682 call->delayedAckEvent = NULL;
4683 MUTEX_ENTER(&rx_refcnt_mutex);
4684 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4685 MUTEX_EXIT(&rx_refcnt_mutex);
4687 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4689 MUTEX_EXIT(&call->lock);
4690 #else /* RX_ENABLE_LOCKS */
4692 call->delayedAckEvent = NULL;
4693 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4694 #endif /* RX_ENABLE_LOCKS */
4698 #ifdef RX_ENABLE_LOCKS
4699 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4700 * clearing them out.
4703 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4705 struct rx_packet *p, *tp;
4708 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4709 p->flags |= RX_PKTFLAG_ACKED;
4713 call->flags |= RX_CALL_TQ_CLEARME;
4714 call->flags |= RX_CALL_TQ_SOME_ACKED;
4717 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4718 call->tfirst = call->tnext;
4719 call->nSoftAcked = 0;
4721 if (call->flags & RX_CALL_FAST_RECOVER) {
4722 call->flags &= ~RX_CALL_FAST_RECOVER;
4723 call->cwind = call->nextCwind;
4724 call->nextCwind = 0;
4727 CV_SIGNAL(&call->cv_twind);
4729 #endif /* RX_ENABLE_LOCKS */
4731 /* Clear out the transmit queue for the current call (all packets have
4732 * been received by peer) */
4734 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4736 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4737 struct rx_packet *p, *tp;
4739 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4741 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4742 p->flags |= RX_PKTFLAG_ACKED;
4746 call->flags |= RX_CALL_TQ_CLEARME;
4747 call->flags |= RX_CALL_TQ_SOME_ACKED;
4750 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4751 #ifdef RXDEBUG_PACKET
4753 #endif /* RXDEBUG_PACKET */
4754 rxi_FreePackets(0, &call->tq);
4755 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
4756 #ifdef RX_ENABLE_LOCKS
4757 CV_BROADCAST(&call->cv_tq);
4758 #else /* RX_ENABLE_LOCKS */
4759 osi_rxWakeup(&call->tq);
4760 #endif /* RX_ENABLE_LOCKS */
4762 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4763 call->flags &= ~RX_CALL_TQ_CLEARME;
4765 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4767 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4768 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4769 call->nSoftAcked = 0;
4771 if (call->flags & RX_CALL_FAST_RECOVER) {
4772 call->flags &= ~RX_CALL_FAST_RECOVER;
4773 call->cwind = call->nextCwind;
4775 #ifdef RX_ENABLE_LOCKS
4776 CV_SIGNAL(&call->cv_twind);
4778 osi_rxWakeup(&call->twind);
4783 rxi_ClearReceiveQueue(struct rx_call *call)
4785 if (queue_IsNotEmpty(&call->rq)) {
4788 count = rxi_FreePackets(0, &call->rq);
4789 rx_packetReclaims += count;
4790 #ifdef RXDEBUG_PACKET
4792 if ( call->rqc != 0 )
4793 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
4795 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4797 if (call->state == RX_STATE_PRECALL) {
4798 call->flags |= RX_CALL_CLEARED;
4802 /* Send an abort packet for the specified call */
4804 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
4805 int istack, int force)
4808 struct clock when, now;
4813 /* Clients should never delay abort messages */
4814 if (rx_IsClientConn(call->conn))
4817 if (call->abortCode != call->error) {
4818 call->abortCode = call->error;
4819 call->abortCount = 0;
4822 if (force || rxi_callAbortThreshhold == 0
4823 || call->abortCount < rxi_callAbortThreshhold) {
4824 if (call->delayedAbortEvent) {
4825 rxevent_Cancel(call->delayedAbortEvent, call,
4826 RX_CALL_REFCOUNT_ABORT);
4828 error = htonl(call->error);
4831 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4832 (char *)&error, sizeof(error), istack);
4833 } else if (!call->delayedAbortEvent) {
4834 clock_GetTime(&now);
4836 clock_Addmsec(&when, rxi_callAbortDelay);
4837 MUTEX_ENTER(&rx_refcnt_mutex);
4838 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4839 MUTEX_EXIT(&rx_refcnt_mutex);
4840 call->delayedAbortEvent =
4841 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4846 /* Send an abort packet for the specified connection. Packet is an
4847 * optional pointer to a packet that can be used to send the abort.
4848 * Once the number of abort messages reaches the threshhold, an
4849 * event is scheduled to send the abort. Setting the force flag
4850 * overrides sending delayed abort messages.
4852 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4853 * to send the abort packet.
4856 rxi_SendConnectionAbort(struct rx_connection *conn,
4857 struct rx_packet *packet, int istack, int force)
4860 struct clock when, now;
4865 /* Clients should never delay abort messages */
4866 if (rx_IsClientConn(conn))
4869 if (force || rxi_connAbortThreshhold == 0
4870 || conn->abortCount < rxi_connAbortThreshhold) {
4871 if (conn->delayedAbortEvent) {
4872 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4874 error = htonl(conn->error);
4876 MUTEX_EXIT(&conn->conn_data_lock);
4878 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4879 RX_PACKET_TYPE_ABORT, (char *)&error,
4880 sizeof(error), istack);
4881 MUTEX_ENTER(&conn->conn_data_lock);
4882 } else if (!conn->delayedAbortEvent) {
4883 clock_GetTime(&now);
4885 clock_Addmsec(&when, rxi_connAbortDelay);
4886 conn->delayedAbortEvent =
4887 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4892 /* Associate an error all of the calls owned by a connection. Called
4893 * with error non-zero. This is only for really fatal things, like
4894 * bad authentication responses. The connection itself is set in
4895 * error at this point, so that future packets received will be
4898 rxi_ConnectionError(struct rx_connection *conn,
4904 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
4906 MUTEX_ENTER(&conn->conn_data_lock);
4907 if (conn->challengeEvent)
4908 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4909 if (conn->natKeepAliveEvent)
4910 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
4911 if (conn->checkReachEvent) {
4912 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4913 conn->checkReachEvent = 0;
4914 conn->flags &= ~RX_CONN_ATTACHWAIT;
4915 MUTEX_ENTER(&rx_refcnt_mutex);
4917 MUTEX_EXIT(&rx_refcnt_mutex);
4919 MUTEX_EXIT(&conn->conn_data_lock);
4920 for (i = 0; i < RX_MAXCALLS; i++) {
4921 struct rx_call *call = conn->call[i];
4923 MUTEX_ENTER(&call->lock);
4924 rxi_CallError(call, error);
4925 MUTEX_EXIT(&call->lock);
4928 conn->error = error;
4929 if (rx_stats_active)
4930 rx_atomic_inc(&rx_stats.fatalErrors);
4935 rxi_CallError(struct rx_call *call, afs_int32 error)
4938 osirx_AssertMine(&call->lock, "rxi_CallError");
4940 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
4942 error = call->error;
4944 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4945 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4946 rxi_ResetCall(call, 0);
4949 rxi_ResetCall(call, 0);
4951 call->error = error;
4954 /* Reset various fields in a call structure, and wakeup waiting
4955 * processes. Some fields aren't changed: state & mode are not
4956 * touched (these must be set by the caller), and bufptr, nLeft, and
4957 * nFree are not reset, since these fields are manipulated by
4958 * unprotected macros, and may only be reset by non-interrupting code.
4961 /* this code requires that call->conn be set properly as a pre-condition. */
4962 #endif /* ADAPT_WINDOW */
4965 rxi_ResetCall(struct rx_call *call, int newcall)
4968 struct rx_peer *peer;
4969 struct rx_packet *packet;
4971 osirx_AssertMine(&call->lock, "rxi_ResetCall");
4973 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
4975 /* Notify anyone who is waiting for asynchronous packet arrival */
4976 if (call->arrivalProc) {
4977 (*call->arrivalProc) (call, call->arrivalProcHandle,
4978 call->arrivalProcArg);
4979 call->arrivalProc = (void (*)())0;
4982 if (call->delayedAbortEvent) {
4983 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4984 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4986 rxi_SendCallAbort(call, packet, 0, 1);
4987 rxi_FreePacket(packet);
4992 * Update the peer with the congestion information in this call
4993 * so other calls on this connection can pick up where this call
4994 * left off. If the congestion sequence numbers don't match then
4995 * another call experienced a retransmission.
4997 peer = call->conn->peer;
4998 MUTEX_ENTER(&peer->peer_lock);
5000 if (call->congestSeq == peer->congestSeq) {
5001 peer->cwind = MAX(peer->cwind, call->cwind);
5002 peer->MTU = MAX(peer->MTU, call->MTU);
5003 peer->nDgramPackets =
5004 MAX(peer->nDgramPackets, call->nDgramPackets);
5007 call->abortCode = 0;
5008 call->abortCount = 0;
5010 if (peer->maxDgramPackets > 1) {
5011 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5013 call->MTU = peer->MTU;
5015 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5016 call->ssthresh = rx_maxSendWindow;
5017 call->nDgramPackets = peer->nDgramPackets;
5018 call->congestSeq = peer->congestSeq;
5019 MUTEX_EXIT(&peer->peer_lock);
5021 flags = call->flags;
5022 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5023 rxi_WaitforTQBusy(call);
5024 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5026 rxi_ClearTransmitQueue(call, 1);
5027 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5028 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5032 rxi_ClearReceiveQueue(call);
5033 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5037 call->twind = call->conn->twind[call->channel];
5038 call->rwind = call->conn->rwind[call->channel];
5039 call->nSoftAcked = 0;
5040 call->nextCwind = 0;
5043 call->nCwindAcks = 0;
5044 call->nSoftAcks = 0;
5045 call->nHardAcks = 0;
5047 call->tfirst = call->rnext = call->tnext = 1;
5049 call->lastAcked = 0;
5050 call->localStatus = call->remoteStatus = 0;
5052 if (flags & RX_CALL_READER_WAIT) {
5053 #ifdef RX_ENABLE_LOCKS
5054 CV_BROADCAST(&call->cv_rq);
5056 osi_rxWakeup(&call->rq);
5059 if (flags & RX_CALL_WAIT_PACKETS) {
5060 MUTEX_ENTER(&rx_freePktQ_lock);
5061 rxi_PacketsUnWait(); /* XXX */
5062 MUTEX_EXIT(&rx_freePktQ_lock);
5064 #ifdef RX_ENABLE_LOCKS
5065 CV_SIGNAL(&call->cv_twind);
5067 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5068 osi_rxWakeup(&call->twind);
5071 #ifdef RX_ENABLE_LOCKS
5072 /* The following ensures that we don't mess with any queue while some
5073 * other thread might also be doing so. The call_queue_lock field is
5074 * is only modified under the call lock. If the call is in the process
5075 * of being removed from a queue, the call is not locked until the
5076 * the queue lock is dropped and only then is the call_queue_lock field
5077 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5078 * Note that any other routine which removes a call from a queue has to
5079 * obtain the queue lock before examing the queue and removing the call.
5081 if (call->call_queue_lock) {
5082 MUTEX_ENTER(call->call_queue_lock);
5083 if (queue_IsOnQueue(call)) {
5085 if (flags & RX_CALL_WAIT_PROC) {
5086 rx_atomic_dec(&rx_nWaiting);
5089 MUTEX_EXIT(call->call_queue_lock);
5090 CLEAR_CALL_QUEUE_LOCK(call);
5092 #else /* RX_ENABLE_LOCKS */
5093 if (queue_IsOnQueue(call)) {
5095 if (flags & RX_CALL_WAIT_PROC)
5096 rx_atomic_dec(&rx_nWaiting);
5098 #endif /* RX_ENABLE_LOCKS */
5100 rxi_KeepAliveOff(call);
5101 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5104 /* Send an acknowledge for the indicated packet (seq,serial) of the
5105 * indicated call, for the indicated reason (reason). This
5106 * acknowledge will specifically acknowledge receiving the packet, and
5107 * will also specify which other packets for this call have been
5108 * received. This routine returns the packet that was used to the
5109 * caller. The caller is responsible for freeing it or re-using it.
5110 * This acknowledgement also returns the highest sequence number
5111 * actually read out by the higher level to the sender; the sender
5112 * promises to keep around packets that have not been read by the
5113 * higher level yet (unless, of course, the sender decides to abort
5114 * the call altogether). Any of p, seq, serial, pflags, or reason may
5115 * be set to zero without ill effect. That is, if they are zero, they
5116 * will not convey any information.
5117 * NOW there is a trailer field, after the ack where it will safely be
5118 * ignored by mundanes, which indicates the maximum size packet this
5119 * host can swallow. */
5121 struct rx_packet *optionalPacket; use to send ack (or null)
5122 int seq; Sequence number of the packet we are acking
5123 int serial; Serial number of the packet
5124 int pflags; Flags field from packet header
5125 int reason; Reason an acknowledge was prompted
5129 rxi_SendAck(struct rx_call *call,
5130 struct rx_packet *optionalPacket, int serial, int reason,
5133 struct rx_ackPacket *ap;
5134 struct rx_packet *rqp;
5135 struct rx_packet *nxp; /* For queue_Scan */
5136 struct rx_packet *p;
5139 afs_uint32 padbytes = 0;
5140 #ifdef RX_ENABLE_TSFPQ
5141 struct rx_ts_info_t * rx_ts_info;
5145 * Open the receive window once a thread starts reading packets
5147 if (call->rnext > 1) {
5148 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5151 /* Don't attempt to grow MTU if this is a critical ping */
5152 if (reason == RX_ACK_MTU) {
5153 /* keep track of per-call attempts, if we're over max, do in small
5154 * otherwise in larger? set a size to increment by, decrease
5157 if (call->conn->peer->maxPacketSize &&
5158 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5160 padbytes = call->conn->peer->maxPacketSize+16;
5162 padbytes = call->conn->peer->maxMTU + 128;
5164 /* do always try a minimum size ping */
5165 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5167 /* subtract the ack payload */
5168 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5169 reason = RX_ACK_PING;
5172 call->nHardAcks = 0;
5173 call->nSoftAcks = 0;
5174 if (call->rnext > call->lastAcked)
5175 call->lastAcked = call->rnext;
5179 rx_computelen(p, p->length); /* reset length, you never know */
5180 } /* where that's been... */
5181 #ifdef RX_ENABLE_TSFPQ
5183 RX_TS_INFO_GET(rx_ts_info);
5184 if ((p = rx_ts_info->local_special_packet)) {
5185 rx_computelen(p, p->length);
5186 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5187 rx_ts_info->local_special_packet = p;
5188 } else { /* We won't send the ack, but don't panic. */
5189 return optionalPacket;
5193 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5194 /* We won't send the ack, but don't panic. */
5195 return optionalPacket;
5200 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5203 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5204 #ifndef RX_ENABLE_TSFPQ
5205 if (!optionalPacket)
5208 return optionalPacket;
5210 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5211 if (rx_Contiguous(p) < templ) {
5212 #ifndef RX_ENABLE_TSFPQ
5213 if (!optionalPacket)
5216 return optionalPacket;
5221 /* MTUXXX failing to send an ack is very serious. We should */
5222 /* try as hard as possible to send even a partial ack; it's */
5223 /* better than nothing. */
5224 ap = (struct rx_ackPacket *)rx_DataOf(p);
5225 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5226 ap->reason = reason;
5228 /* The skew computation used to be bogus, I think it's better now. */
5229 /* We should start paying attention to skew. XXX */
5230 ap->serial = htonl(serial);
5231 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5233 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
5234 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5236 /* No fear of running out of ack packet here because there can only be at most
5237 * one window full of unacknowledged packets. The window size must be constrained
5238 * to be less than the maximum ack size, of course. Also, an ack should always
5239 * fit into a single packet -- it should not ever be fragmented. */
5240 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5241 if (!rqp || !call->rq.next
5242 || (rqp->header.seq > (call->rnext + call->rwind))) {
5243 #ifndef RX_ENABLE_TSFPQ
5244 if (!optionalPacket)
5247 rxi_CallError(call, RX_CALL_DEAD);
5248 return optionalPacket;
5251 while (rqp->header.seq > call->rnext + offset)
5252 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5253 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5255 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5256 #ifndef RX_ENABLE_TSFPQ
5257 if (!optionalPacket)
5260 rxi_CallError(call, RX_CALL_DEAD);
5261 return optionalPacket;
5266 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5268 /* these are new for AFS 3.3 */
5269 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5270 templ = htonl(templ);
5271 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5272 templ = htonl(call->conn->peer->ifMTU);
5273 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5274 sizeof(afs_int32), &templ);
5276 /* new for AFS 3.4 */
5277 templ = htonl(call->rwind);
5278 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5279 sizeof(afs_int32), &templ);
5281 /* new for AFS 3.5 */
5282 templ = htonl(call->conn->peer->ifDgramPackets);
5283 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5284 sizeof(afs_int32), &templ);
5286 p->header.serviceId = call->conn->serviceId;
5287 p->header.cid = (call->conn->cid | call->channel);
5288 p->header.callNumber = *call->callNumber;
5290 p->header.securityIndex = call->conn->securityIndex;
5291 p->header.epoch = call->conn->epoch;
5292 p->header.type = RX_PACKET_TYPE_ACK;
5293 p->header.flags = RX_SLOW_START_OK;
5294 if (reason == RX_ACK_PING) {
5295 p->header.flags |= RX_REQUEST_ACK;
5297 clock_GetTime(&call->pingRequestTime);
5300 p->length = padbytes +
5301 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5304 /* not fast but we can potentially use this if truncated
5305 * fragments are delivered to figure out the mtu.
5307 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5308 sizeof(afs_int32), sizeof(afs_int32),
5312 if (call->conn->type == RX_CLIENT_CONNECTION)
5313 p->header.flags |= RX_CLIENT_INITIATED;
5317 if (rxdebug_active) {
5321 len = _snprintf(msg, sizeof(msg),
5322 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5323 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5324 ntohl(ap->serial), ntohl(ap->previousPacket),
5325 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5326 ap->nAcks, ntohs(ap->bufferSpace) );
5330 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5331 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5335 OutputDebugString(msg);
5337 #else /* AFS_NT40_ENV */
5339 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5340 ap->reason, ntohl(ap->previousPacket),
5341 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5343 for (offset = 0; offset < ap->nAcks; offset++)
5344 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5349 #endif /* AFS_NT40_ENV */
5352 int i, nbytes = p->length;
5354 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5355 if (nbytes <= p->wirevec[i].iov_len) {
5358 savelen = p->wirevec[i].iov_len;
5360 p->wirevec[i].iov_len = nbytes;
5362 rxi_Send(call, p, istack);
5363 p->wirevec[i].iov_len = savelen;
5367 nbytes -= p->wirevec[i].iov_len;
5370 if (rx_stats_active)
5371 rx_atomic_inc(&rx_stats.ackPacketsSent);
5372 #ifndef RX_ENABLE_TSFPQ
5373 if (!optionalPacket)
5376 return optionalPacket; /* Return packet for re-use by caller */
5379 /* Send all of the packets in the list in single datagram */
5381 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
5382 int istack, int moreFlag, struct clock *now,
5383 struct clock *retryTime, int resending)
5388 struct rx_connection *conn = call->conn;
5389 struct rx_peer *peer = conn->peer;
5391 MUTEX_ENTER(&peer->peer_lock);
5394 peer->reSends += len;
5395 MUTEX_EXIT(&peer->peer_lock);
5397 if (rx_stats_active) {
5399 rx_atomic_add(&rx_stats.dataPacketsReSent, len);
5401 rx_atomic_add(&rx_stats.dataPacketsSent, len);
5404 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
5408 /* Set the packet flags and schedule the resend events */
5409 /* Only request an ack for the last packet in the list */
5410 for (i = 0; i < len; i++) {
5411 list[i]->retryTime = *retryTime;
5412 if (list[i]->header.serial) {
5413 /* Exponentially backoff retry times */
5414 if (list[i]->backoff < MAXBACKOFF) {
5415 /* so it can't stay == 0 */
5416 list[i]->backoff = (list[i]->backoff << 1) + 1;
5419 clock_Addmsec(&(list[i]->retryTime),
5420 ((afs_uint32) list[i]->backoff) << 8);
5423 /* Wait a little extra for the ack on the last packet */
5424 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
5425 clock_Addmsec(&(list[i]->retryTime), 400);
5428 /* Record the time sent */
5429 list[i]->timeSent = *now;
5431 /* Ask for an ack on retransmitted packets, on every other packet
5432 * if the peer doesn't support slow start. Ask for an ack on every
5433 * packet until the congestion window reaches the ack rate. */
5434 if (list[i]->header.serial) {
5437 /* improved RTO calculation- not Karn */
5438 list[i]->firstSent = *now;
5439 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5440 || (!(call->flags & RX_CALL_SLOW_START_OK)
5441 && (list[i]->header.seq & 1)))) {
5446 /* Tag this packet as not being the last in this group,
5447 * for the receiver's benefit */
5448 if (i < len - 1 || moreFlag) {
5449 list[i]->header.flags |= RX_MORE_PACKETS;
5452 /* Install the new retransmit time for the packet, and
5453 * record the time sent */
5454 list[i]->timeSent = *now;
5458 list[len - 1]->header.flags |= RX_REQUEST_ACK;
5461 /* Since we're about to send a data packet to the peer, it's
5462 * safe to nuke any scheduled end-of-packets ack */
5463 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5465 MUTEX_EXIT(&call->lock);
5466 MUTEX_ENTER(&rx_refcnt_mutex);
5467 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5468 MUTEX_EXIT(&rx_refcnt_mutex);
5470 rxi_SendPacketList(call, conn, list, len, istack);
5472 rxi_SendPacket(call, conn, list[0], istack);
5474 MUTEX_ENTER(&call->lock);
5475 MUTEX_ENTER(&rx_refcnt_mutex);
5476 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5477 MUTEX_EXIT(&rx_refcnt_mutex);
5479 /* Update last send time for this call (for keep-alive
5480 * processing), and for the connection (so that we can discover
5481 * idle connections) */
5482 conn->lastSendTime = call->lastSendTime = clock_Sec();
5483 /* Let a set of retransmits trigger an idle timeout */
5485 call->lastSendData = call->lastSendTime;
5488 /* When sending packets we need to follow these rules:
5489 * 1. Never send more than maxDgramPackets in a jumbogram.
5490 * 2. Never send a packet with more than two iovecs in a jumbogram.
5491 * 3. Never send a retransmitted packet in a jumbogram.
5492 * 4. Never send more than cwind/4 packets in a jumbogram
5493 * We always keep the last list we should have sent so we
5494 * can set the RX_MORE_PACKETS flags correctly.
5497 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5498 int istack, struct clock *now, struct clock *retryTime,
5501 int i, cnt, lastCnt = 0;
5502 struct rx_packet **listP, **lastP = 0;
5503 struct rx_peer *peer = call->conn->peer;
5504 int morePackets = 0;
5506 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
5507 /* Does the current packet force us to flush the current list? */
5509 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5510 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5512 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
5514 /* If the call enters an error state stop sending, or if
5515 * we entered congestion recovery mode, stop sending */
5516 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5524 /* Add the current packet to the list if it hasn't been acked.
5525 * Otherwise adjust the list pointer to skip the current packet. */
5526 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5528 /* Do we need to flush the list? */
5529 if (cnt >= (int)peer->maxDgramPackets
5530 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5531 || list[i]->header.serial
5532 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5534 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5535 retryTime, resending);
5536 /* If the call enters an error state stop sending, or if
5537 * we entered congestion recovery mode, stop sending */
5539 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5544 listP = &list[i + 1];
5549 osi_Panic("rxi_SendList error");
5551 listP = &list[i + 1];
5555 /* Send the whole list when the call is in receive mode, when
5556 * the call is in eof mode, when we are in fast recovery mode,
5557 * and when we have the last packet */
5558 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5559 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5560 || (call->flags & RX_CALL_FAST_RECOVER)) {
5561 /* Check for the case where the current list contains
5562 * an acked packet. Since we always send retransmissions
5563 * in a separate packet, we only need to check the first
5564 * packet in the list */
5565 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5569 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5570 retryTime, resending);
5571 /* If the call enters an error state stop sending, or if
5572 * we entered congestion recovery mode, stop sending */
5573 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5577 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5580 } else if (lastCnt > 0) {
5581 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5586 #ifdef RX_ENABLE_LOCKS
5587 /* Call rxi_Start, below, but with the call lock held. */
5589 rxi_StartUnlocked(struct rxevent *event,
5590 void *arg0, void *arg1, int istack)
5592 struct rx_call *call = arg0;
5594 MUTEX_ENTER(&call->lock);
5595 rxi_Start(event, call, arg1, istack);
5596 MUTEX_EXIT(&call->lock);
5598 #endif /* RX_ENABLE_LOCKS */
5600 /* This routine is called when new packets are readied for
5601 * transmission and when retransmission may be necessary, or when the
5602 * transmission window or burst count are favourable. This should be
5603 * better optimized for new packets, the usual case, now that we've
5604 * got rid of queues of send packets. XXXXXXXXXXX */
5606 rxi_Start(struct rxevent *event,
5607 void *arg0, void *arg1, int istack)
5609 struct rx_call *call = arg0;
5611 struct rx_packet *p;
5612 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5613 struct rx_peer *peer = call->conn->peer;
5614 struct clock now, usenow, retryTime;
5620 /* If rxi_Start is being called as a result of a resend event,
5621 * then make sure that the event pointer is removed from the call
5622 * structure, since there is no longer a per-call retransmission
5624 if (event && event == call->resendEvent) {
5625 MUTEX_ENTER(&rx_refcnt_mutex);
5626 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5627 MUTEX_EXIT(&rx_refcnt_mutex);
5628 call->resendEvent = NULL;
5630 if (queue_IsEmpty(&call->tq)) {
5634 /* Timeouts trigger congestion recovery */
5635 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5636 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5637 /* someone else is waiting to start recovery */
5640 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5641 rxi_WaitforTQBusy(call);
5642 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5643 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5644 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5646 if (rx_stats_active)
5647 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5651 call->flags |= RX_CALL_FAST_RECOVER;
5653 if (peer->maxDgramPackets > 1) {
5654 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5656 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5658 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5659 call->nDgramPackets = 1;
5661 call->nextCwind = 1;
5664 MUTEX_ENTER(&peer->peer_lock);
5665 peer->MTU = call->MTU;
5666 peer->cwind = call->cwind;
5667 peer->nDgramPackets = 1;
5669 call->congestSeq = peer->congestSeq;
5670 MUTEX_EXIT(&peer->peer_lock);
5671 /* Clear retry times on packets. Otherwise, it's possible for
5672 * some packets in the queue to force resends at rates faster
5673 * than recovery rates.
5675 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5676 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5677 clock_Zero(&p->retryTime);
5682 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5683 if (rx_stats_active)
5684 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5689 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5690 /* Get clock to compute the re-transmit time for any packets
5691 * in this burst. Note, if we back off, it's reasonable to
5692 * back off all of the packets in the same manner, even if
5693 * some of them have been retransmitted more times than more
5695 * Do a dance to avoid blocking after setting now. */
5696 MUTEX_ENTER(&peer->peer_lock);
5697 retryTime = peer->timeout;
5698 MUTEX_EXIT(&peer->peer_lock);
5700 clock_GetTime(&now);
5701 clock_Add(&retryTime, &now);
5703 /* Send (or resend) any packets that need it, subject to
5704 * window restrictions and congestion burst control
5705 * restrictions. Ask for an ack on the last packet sent in
5706 * this burst. For now, we're relying upon the window being
5707 * considerably bigger than the largest number of packets that
5708 * are typically sent at once by one initial call to
5709 * rxi_Start. This is probably bogus (perhaps we should ask
5710 * for an ack when we're half way through the current
5711 * window?). Also, for non file transfer applications, this
5712 * may end up asking for an ack for every packet. Bogus. XXXX
5715 * But check whether we're here recursively, and let the other guy
5718 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5719 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5720 call->flags |= RX_CALL_TQ_BUSY;
5722 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5724 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5725 call->flags &= ~RX_CALL_NEED_START;
5726 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5728 maxXmitPackets = MIN(call->twind, call->cwind);
5729 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5730 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5731 /* We shouldn't be sending packets if a thread is waiting
5732 * to initiate congestion recovery */
5733 dpf(("call %d waiting to initiate fast recovery\n",
5734 *(call->callNumber)));
5738 && (call->flags & RX_CALL_FAST_RECOVER)) {
5739 /* Only send one packet during fast recovery */
5740 dpf(("call %d restricted to one packet per send during fast recovery\n",
5741 *(call->callNumber)));
5744 #ifdef RX_TRACK_PACKETS
5745 if ((p->flags & RX_PKTFLAG_FREE)
5746 || (!queue_IsEnd(&call->tq, nxp)
5747 && (nxp->flags & RX_PKTFLAG_FREE))
5748 || (p == (struct rx_packet *)&rx_freePacketQueue)
5749 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5750 osi_Panic("rxi_Start: xmit queue clobbered");
5753 if (p->flags & RX_PKTFLAG_ACKED) {
5754 /* Since we may block, don't trust this */
5755 usenow.sec = usenow.usec = 0;
5756 if (rx_stats_active)
5757 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
5758 continue; /* Ignore this packet if it has been acknowledged */
5761 /* Turn off all flags except these ones, which are the same
5762 * on each transmission */
5763 p->header.flags &= RX_PRESET_FLAGS;
5765 if (p->header.seq >=
5766 call->tfirst + MIN((int)call->twind,
5767 (int)(call->nSoftAcked +
5769 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5770 /* Note: if we're waiting for more window space, we can
5771 * still send retransmits; hence we don't return here, but
5772 * break out to schedule a retransmit event */
5773 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
5774 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
5779 /* Transmit the packet if it needs to be sent. */
5780 if (!clock_Lt(&now, &p->retryTime)) {
5781 if (nXmitPackets == maxXmitPackets) {
5782 rxi_SendXmitList(call, call->xmitList,
5783 nXmitPackets, istack, &now,
5784 &retryTime, resending);
5787 dpf(("call %d xmit packet %"AFS_PTR_FMT" now %u.%06u retryTime %u.%06u nextRetry %u.%06u\n",
5788 *(call->callNumber), p,
5790 p->retryTime.sec, p->retryTime.usec,
5791 retryTime.sec, retryTime.usec));
5792 call->xmitList[nXmitPackets++] = p;
5796 /* xmitList now hold pointers to all of the packets that are
5797 * ready to send. Now we loop to send the packets */
5798 if (nXmitPackets > 0) {
5799 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
5800 istack, &now, &retryTime, resending);
5803 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5805 * TQ references no longer protected by this flag; they must remain
5806 * protected by the global lock.
5808 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5809 call->flags &= ~RX_CALL_TQ_BUSY;
5810 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5811 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
5812 call, call->tqWaiters, call->flags));
5813 #ifdef RX_ENABLE_LOCKS
5814 osirx_AssertMine(&call->lock, "rxi_Start start");
5815 CV_BROADCAST(&call->cv_tq);
5816 #else /* RX_ENABLE_LOCKS */
5817 osi_rxWakeup(&call->tq);
5818 #endif /* RX_ENABLE_LOCKS */
5823 /* We went into the error state while sending packets. Now is
5824 * the time to reset the call. This will also inform the using
5825 * process that the call is in an error state.
5827 if (rx_stats_active)
5828 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
5829 call->flags &= ~RX_CALL_TQ_BUSY;
5830 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5831 dpf(("call error %d while xmit %p has %d waiters and flags %d\n",
5832 call->error, call, call->tqWaiters, call->flags));
5833 #ifdef RX_ENABLE_LOCKS
5834 osirx_AssertMine(&call->lock, "rxi_Start middle");
5835 CV_BROADCAST(&call->cv_tq);
5836 #else /* RX_ENABLE_LOCKS */
5837 osi_rxWakeup(&call->tq);
5838 #endif /* RX_ENABLE_LOCKS */
5840 rxi_CallError(call, call->error);
5843 #ifdef RX_ENABLE_LOCKS
5844 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5846 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5847 /* Some packets have received acks. If they all have, we can clear
5848 * the transmit queue.
5851 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5852 if (p->header.seq < call->tfirst
5853 && (p->flags & RX_PKTFLAG_ACKED)) {
5855 #ifdef RX_TRACK_PACKETS
5856 p->flags &= ~RX_PKTFLAG_TQ;
5858 #ifdef RXDEBUG_PACKET
5866 call->flags |= RX_CALL_TQ_CLEARME;
5868 #endif /* RX_ENABLE_LOCKS */
5869 /* Don't bother doing retransmits if the TQ is cleared. */
5870 if (call->flags & RX_CALL_TQ_CLEARME) {
5871 rxi_ClearTransmitQueue(call, 1);
5873 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5876 /* Always post a resend event, if there is anything in the
5877 * queue, and resend is possible. There should be at least
5878 * one unacknowledged packet in the queue ... otherwise none
5879 * of these packets should be on the queue in the first place.
5881 if (call->resendEvent) {
5882 /* Cancel the existing event and post a new one */
5883 rxevent_Cancel(call->resendEvent, call,
5884 RX_CALL_REFCOUNT_RESEND);
5887 /* The retry time is the retry time on the first unacknowledged
5888 * packet inside the current window */
5890 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5891 /* Don't set timers for packets outside the window */
5892 if (p->header.seq >= call->tfirst + call->twind) {
5896 if (!(p->flags & RX_PKTFLAG_ACKED)
5897 && !clock_IsZero(&p->retryTime)) {
5899 retryTime = p->retryTime;
5904 /* Post a new event to re-run rxi_Start when retries may be needed */
5905 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5906 #ifdef RX_ENABLE_LOCKS
5907 MUTEX_ENTER(&rx_refcnt_mutex);
5908 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5909 MUTEX_EXIT(&rx_refcnt_mutex);
5911 rxevent_PostNow2(&retryTime, &usenow,
5913 (void *)call, 0, istack);
5914 #else /* RX_ENABLE_LOCKS */
5916 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5917 (void *)call, 0, istack);
5918 #endif /* RX_ENABLE_LOCKS */
5921 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5922 } while (call->flags & RX_CALL_NEED_START);
5924 * TQ references no longer protected by this flag; they must remain
5925 * protected by the global lock.
5927 call->flags &= ~RX_CALL_TQ_BUSY;
5928 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5929 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
5930 call, call->tqWaiters, call->flags));
5931 #ifdef RX_ENABLE_LOCKS
5932 osirx_AssertMine(&call->lock, "rxi_Start end");
5933 CV_BROADCAST(&call->cv_tq);
5934 #else /* RX_ENABLE_LOCKS */
5935 osi_rxWakeup(&call->tq);
5936 #endif /* RX_ENABLE_LOCKS */
5939 call->flags |= RX_CALL_NEED_START;
5941 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5943 if (call->resendEvent) {
5944 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5949 /* Also adjusts the keep alive parameters for the call, to reflect
5950 * that we have just sent a packet (so keep alives aren't sent
5953 rxi_Send(struct rx_call *call, struct rx_packet *p,
5956 struct rx_connection *conn = call->conn;
5958 /* Stamp each packet with the user supplied status */
5959 p->header.userStatus = call->localStatus;
5961 /* Allow the security object controlling this call's security to
5962 * make any last-minute changes to the packet */
5963 RXS_SendPacket(conn->securityObject, call, p);
5965 /* Since we're about to send SOME sort of packet to the peer, it's
5966 * safe to nuke any scheduled end-of-packets ack */
5967 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5969 /* Actually send the packet, filling in more connection-specific fields */
5970 MUTEX_EXIT(&call->lock);
5971 MUTEX_ENTER(&rx_refcnt_mutex);
5972 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5973 MUTEX_EXIT(&rx_refcnt_mutex);
5974 rxi_SendPacket(call, conn, p, istack);
5975 MUTEX_ENTER(&rx_refcnt_mutex);
5976 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5977 MUTEX_EXIT(&rx_refcnt_mutex);
5978 MUTEX_ENTER(&call->lock);
5980 /* Update last send time for this call (for keep-alive
5981 * processing), and for the connection (so that we can discover
5982 * idle connections) */
5983 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5984 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
5985 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
5987 conn->lastSendTime = call->lastSendTime = clock_Sec();
5988 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
5989 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5990 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
5991 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
5992 RX_ACK_PING_RESPONSE)))
5993 call->lastSendData = call->lastSendTime;
5997 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5998 * that things are fine. Also called periodically to guarantee that nothing
5999 * falls through the cracks (e.g. (error + dally) connections have keepalive
6000 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6002 * haveCTLock Set if calling from rxi_ReapConnections
6004 #ifdef RX_ENABLE_LOCKS
6006 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6007 #else /* RX_ENABLE_LOCKS */
6009 rxi_CheckCall(struct rx_call *call)
6010 #endif /* RX_ENABLE_LOCKS */
6012 struct rx_connection *conn = call->conn;
6014 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6015 afs_uint32 fudgeFactor;
6019 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6020 if (call->flags & RX_CALL_TQ_BUSY) {
6021 /* Call is active and will be reset by rxi_Start if it's
6022 * in an error state.
6027 /* RTT + 8*MDEV, rounded up to the next second. */
6028 fudgeFactor = (((afs_uint32) conn->peer->rtt >> 3) +
6029 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
6031 deadTime = conn->secondsUntilDead + fudgeFactor;
6033 /* These are computed to the second (+- 1 second). But that's
6034 * good enough for these values, which should be a significant
6035 * number of seconds. */
6036 if (now > (call->lastReceiveTime + deadTime)) {
6037 if (call->state == RX_STATE_ACTIVE) {
6039 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
6041 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6042 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6043 ip_stack_t *ipst = ns->netstack_ip;
6045 ire = ire_cache_lookup(conn->peer->host
6046 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6048 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
6050 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6057 if (ire && ire->ire_max_frag > 0)
6058 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6060 #if defined(GLOBAL_NETSTACKID)
6064 #endif /* ADAPT_PMTU */
6065 cerror = RX_CALL_DEAD;
6068 #ifdef RX_ENABLE_LOCKS
6069 /* Cancel pending events */
6070 rxevent_Cancel(call->delayedAckEvent, call,
6071 RX_CALL_REFCOUNT_DELAY);
6072 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
6073 rxevent_Cancel(call->keepAliveEvent, call,
6074 RX_CALL_REFCOUNT_ALIVE);
6075 MUTEX_ENTER(&rx_refcnt_mutex);
6076 if (call->refCount == 0) {
6077 rxi_FreeCall(call, haveCTLock);
6078 MUTEX_EXIT(&rx_refcnt_mutex);
6081 MUTEX_EXIT(&rx_refcnt_mutex);
6083 #else /* RX_ENABLE_LOCKS */
6084 rxi_FreeCall(call, 0);
6086 #endif /* RX_ENABLE_LOCKS */
6088 /* Non-active calls are destroyed if they are not responding
6089 * to pings; active calls are simply flagged in error, so the
6090 * attached process can die reasonably gracefully. */
6093 if (conn->idleDeadTime) {
6094 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6097 /* see if we have a non-activity timeout */
6098 if (call->startWait && idleDeadTime
6099 && ((call->startWait + idleDeadTime) < now) &&
6100 (call->flags & RX_CALL_READER_WAIT)) {
6101 if (call->state == RX_STATE_ACTIVE) {
6102 cerror = RX_CALL_TIMEOUT;
6106 if (call->lastSendData && idleDeadTime && (conn->idleDeadErr != 0)
6107 && ((call->lastSendData + idleDeadTime) < now)) {
6108 if (call->state == RX_STATE_ACTIVE) {
6109 cerror = conn->idleDeadErr;
6115 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6118 /* see if we have a hard timeout */
6120 && (now > (hardDeadTime + call->startTime.sec))) {
6121 if (call->state == RX_STATE_ACTIVE)
6122 rxi_CallError(call, RX_CALL_TIMEOUT);
6127 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT
6128 && call->lastReceiveTime) {
6129 int oldMTU = conn->peer->ifMTU;
6131 /* if we thought we could send more, perhaps things got worse */
6132 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6133 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6134 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6135 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6137 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6139 /* minimum capped in SetPeerMtu */
6140 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6143 conn->lastPacketSize = 0;
6145 /* needed so ResetCall doesn't clobber us. */
6146 call->MTU = conn->peer->ifMTU;
6148 /* if we never succeeded, let the error pass out as-is */
6149 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6150 cerror = conn->msgsizeRetryErr;
6153 rxi_CallError(call, cerror);
6158 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6160 struct rx_connection *conn = arg1;
6161 struct rx_header theader;
6163 struct sockaddr_in taddr;
6166 struct iovec tmpiov[2];
6169 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6172 tp = &tbuffer[sizeof(struct rx_header)];
6173 taddr.sin_family = AF_INET;
6174 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6175 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6176 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6177 taddr.sin_len = sizeof(struct sockaddr_in);
6179 memset(&theader, 0, sizeof(theader));
6180 theader.epoch = htonl(999);
6182 theader.callNumber = 0;
6185 theader.type = RX_PACKET_TYPE_VERSION;
6186 theader.flags = RX_LAST_PACKET;
6187 theader.serviceId = 0;
6189 memcpy(tbuffer, &theader, sizeof(theader));
6190 memcpy(tp, &a, sizeof(a));
6191 tmpiov[0].iov_base = tbuffer;
6192 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6194 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6196 MUTEX_ENTER(&conn->conn_data_lock);
6197 MUTEX_ENTER(&rx_refcnt_mutex);
6198 /* Only reschedule ourselves if the connection would not be destroyed */
6199 if (conn->refCount <= 1) {
6200 conn->natKeepAliveEvent = NULL;
6201 MUTEX_EXIT(&rx_refcnt_mutex);
6202 MUTEX_EXIT(&conn->conn_data_lock);
6203 rx_DestroyConnection(conn); /* drop the reference for this */
6205 conn->refCount--; /* drop the reference for this */
6206 MUTEX_EXIT(&rx_refcnt_mutex);
6207 conn->natKeepAliveEvent = NULL;
6208 rxi_ScheduleNatKeepAliveEvent(conn);
6209 MUTEX_EXIT(&conn->conn_data_lock);
6214 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6216 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6217 struct clock when, now;
6218 clock_GetTime(&now);
6220 when.sec += conn->secondsUntilNatPing;
6221 MUTEX_ENTER(&rx_refcnt_mutex);
6222 conn->refCount++; /* hold a reference for this */
6223 MUTEX_EXIT(&rx_refcnt_mutex);
6224 conn->natKeepAliveEvent =
6225 rxevent_PostNow(&when, &now, rxi_NatKeepAliveEvent, conn, 0);
6230 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6232 MUTEX_ENTER(&conn->conn_data_lock);
6233 conn->secondsUntilNatPing = seconds;
6235 rxi_ScheduleNatKeepAliveEvent(conn);
6236 MUTEX_EXIT(&conn->conn_data_lock);
6240 rxi_NatKeepAliveOn(struct rx_connection *conn)
6242 MUTEX_ENTER(&conn->conn_data_lock);
6243 rxi_ScheduleNatKeepAliveEvent(conn);
6244 MUTEX_EXIT(&conn->conn_data_lock);
6247 /* When a call is in progress, this routine is called occasionally to
6248 * make sure that some traffic has arrived (or been sent to) the peer.
6249 * If nothing has arrived in a reasonable amount of time, the call is
6250 * declared dead; if nothing has been sent for a while, we send a
6251 * keep-alive packet (if we're actually trying to keep the call alive)
6254 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6256 struct rx_call *call = arg1;
6257 struct rx_connection *conn;
6260 MUTEX_ENTER(&rx_refcnt_mutex);
6261 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6262 MUTEX_EXIT(&rx_refcnt_mutex);
6263 MUTEX_ENTER(&call->lock);
6264 if (event == call->keepAliveEvent)
6265 call->keepAliveEvent = NULL;
6268 #ifdef RX_ENABLE_LOCKS
6269 if (rxi_CheckCall(call, 0)) {
6270 MUTEX_EXIT(&call->lock);
6273 #else /* RX_ENABLE_LOCKS */
6274 if (rxi_CheckCall(call))
6276 #endif /* RX_ENABLE_LOCKS */
6278 /* Don't try to keep alive dallying calls */
6279 if (call->state == RX_STATE_DALLY) {
6280 MUTEX_EXIT(&call->lock);
6285 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6286 /* Don't try to send keepalives if there is unacknowledged data */
6287 /* the rexmit code should be good enough, this little hack
6288 * doesn't quite work XXX */
6289 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6291 rxi_ScheduleKeepAliveEvent(call);
6292 MUTEX_EXIT(&call->lock);
6295 /* Does what's on the nameplate. */
6297 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy)
6299 struct rx_call *call = arg1;
6300 struct rx_connection *conn;
6302 MUTEX_ENTER(&rx_refcnt_mutex);
6303 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6304 MUTEX_EXIT(&rx_refcnt_mutex);
6305 MUTEX_ENTER(&call->lock);
6307 if (event == call->growMTUEvent)
6308 call->growMTUEvent = NULL;
6310 #ifdef RX_ENABLE_LOCKS
6311 if (rxi_CheckCall(call, 0)) {
6312 MUTEX_EXIT(&call->lock);
6315 #else /* RX_ENABLE_LOCKS */
6316 if (rxi_CheckCall(call))
6318 #endif /* RX_ENABLE_LOCKS */
6320 /* Don't bother with dallying calls */
6321 if (call->state == RX_STATE_DALLY) {
6322 MUTEX_EXIT(&call->lock);
6329 * keep being scheduled, just don't do anything if we're at peak,
6330 * or we're not set up to be properly handled (idle timeout required)
6332 if ((conn->peer->maxPacketSize != 0) &&
6333 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6334 (conn->idleDeadErr))
6335 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6336 rxi_ScheduleGrowMTUEvent(call, 0);
6337 MUTEX_EXIT(&call->lock);
6341 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6343 if (!call->keepAliveEvent) {
6344 struct clock when, now;
6345 clock_GetTime(&now);
6347 when.sec += call->conn->secondsUntilPing;
6348 MUTEX_ENTER(&rx_refcnt_mutex);
6349 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6350 MUTEX_EXIT(&rx_refcnt_mutex);
6351 call->keepAliveEvent =
6352 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
6357 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6359 if (!call->growMTUEvent) {
6360 struct clock when, now;
6362 clock_GetTime(&now);
6365 if (call->conn->secondsUntilPing)
6366 secs = (6*call->conn->secondsUntilPing)-1;
6368 if (call->conn->secondsUntilDead)
6369 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6373 MUTEX_ENTER(&rx_refcnt_mutex);
6374 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6375 MUTEX_EXIT(&rx_refcnt_mutex);
6376 call->growMTUEvent =
6377 rxevent_PostNow(&when, &now, rxi_GrowMTUEvent, call, 0);
6381 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6383 rxi_KeepAliveOn(struct rx_call *call)
6385 /* Pretend last packet received was received now--i.e. if another
6386 * packet isn't received within the keep alive time, then the call
6387 * will die; Initialize last send time to the current time--even
6388 * if a packet hasn't been sent yet. This will guarantee that a
6389 * keep-alive is sent within the ping time */
6390 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6391 rxi_ScheduleKeepAliveEvent(call);
6395 rxi_GrowMTUOn(struct rx_call *call)
6397 struct rx_connection *conn = call->conn;
6398 MUTEX_ENTER(&conn->conn_data_lock);
6399 conn->lastPingSizeSer = conn->lastPingSize = 0;
6400 MUTEX_EXIT(&conn->conn_data_lock);
6401 rxi_ScheduleGrowMTUEvent(call, 1);
6404 /* This routine is called to send connection abort messages
6405 * that have been delayed to throttle looping clients. */
6407 rxi_SendDelayedConnAbort(struct rxevent *event,
6408 void *arg1, void *unused)
6410 struct rx_connection *conn = arg1;
6413 struct rx_packet *packet;
6415 MUTEX_ENTER(&conn->conn_data_lock);
6416 conn->delayedAbortEvent = NULL;
6417 error = htonl(conn->error);
6419 MUTEX_EXIT(&conn->conn_data_lock);
6420 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6423 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6424 RX_PACKET_TYPE_ABORT, (char *)&error,
6426 rxi_FreePacket(packet);
6430 /* This routine is called to send call abort messages
6431 * that have been delayed to throttle looping clients. */
6433 rxi_SendDelayedCallAbort(struct rxevent *event,
6434 void *arg1, void *dummy)
6436 struct rx_call *call = arg1;
6439 struct rx_packet *packet;
6441 MUTEX_ENTER(&call->lock);
6442 call->delayedAbortEvent = NULL;
6443 error = htonl(call->error);
6445 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6448 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6449 (char *)&error, sizeof(error), 0);
6450 rxi_FreePacket(packet);
6452 MUTEX_EXIT(&call->lock);
6453 MUTEX_ENTER(&rx_refcnt_mutex);
6454 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6455 MUTEX_EXIT(&rx_refcnt_mutex);
6458 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6459 * seconds) to ask the client to authenticate itself. The routine
6460 * issues a challenge to the client, which is obtained from the
6461 * security object associated with the connection */
6463 rxi_ChallengeEvent(struct rxevent *event,
6464 void *arg0, void *arg1, int tries)
6466 struct rx_connection *conn = arg0;
6468 conn->challengeEvent = NULL;
6469 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6470 struct rx_packet *packet;
6471 struct clock when, now;
6474 /* We've failed to authenticate for too long.
6475 * Reset any calls waiting for authentication;
6476 * they are all in RX_STATE_PRECALL.
6480 MUTEX_ENTER(&conn->conn_call_lock);
6481 for (i = 0; i < RX_MAXCALLS; i++) {
6482 struct rx_call *call = conn->call[i];
6484 MUTEX_ENTER(&call->lock);
6485 if (call->state == RX_STATE_PRECALL) {
6486 rxi_CallError(call, RX_CALL_DEAD);
6487 rxi_SendCallAbort(call, NULL, 0, 0);
6489 MUTEX_EXIT(&call->lock);
6492 MUTEX_EXIT(&conn->conn_call_lock);
6496 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6498 /* If there's no packet available, do this later. */
6499 RXS_GetChallenge(conn->securityObject, conn, packet);
6500 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6501 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6502 rxi_FreePacket(packet);
6504 clock_GetTime(&now);
6506 when.sec += RX_CHALLENGE_TIMEOUT;
6507 conn->challengeEvent =
6508 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
6513 /* Call this routine to start requesting the client to authenticate
6514 * itself. This will continue until authentication is established,
6515 * the call times out, or an invalid response is returned. The
6516 * security object associated with the connection is asked to create
6517 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6518 * defined earlier. */
6520 rxi_ChallengeOn(struct rx_connection *conn)
6522 if (!conn->challengeEvent) {
6523 RXS_CreateChallenge(conn->securityObject, conn);
6524 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6529 /* Compute round trip time of the packet provided, in *rttp.
6532 /* rxi_ComputeRoundTripTime is called with peer locked. */
6533 /* sentp and/or peer may be null */
6535 rxi_ComputeRoundTripTime(struct rx_packet *p,
6536 struct clock *sentp,
6537 struct rx_peer *peer,
6540 struct clock thisRtt, *rttp = &thisRtt;
6545 if (clock_Lt(rttp, sentp))
6546 return; /* somebody set the clock back, don't count this time. */
6548 clock_Sub(rttp, sentp);
6549 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6550 p->header.callNumber, p, rttp->sec, rttp->usec));
6552 if (rttp->sec == 0 && rttp->usec == 0) {
6554 * The actual round trip time is shorter than the
6555 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6556 * Since we can't tell which at the moment we will assume 1ms.
6561 if (rx_stats_active) {
6562 MUTEX_ENTER(&rx_stats_mutex);
6563 if (clock_Lt(rttp, &rx_stats.minRtt))
6564 rx_stats.minRtt = *rttp;
6565 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
6566 if (rttp->sec > 60) {
6567 MUTEX_EXIT(&rx_stats_mutex);
6568 return; /* somebody set the clock ahead */
6570 rx_stats.maxRtt = *rttp;
6572 clock_Add(&rx_stats.totalRtt, rttp);
6573 rx_atomic_inc(&rx_stats.nRttSamples);
6574 MUTEX_EXIT(&rx_stats_mutex);
6577 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6579 /* Apply VanJacobson round-trip estimations */
6584 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
6585 * srtt is stored as fixed point with 3 bits after the binary
6586 * point (i.e., scaled by 8). The following magic is
6587 * equivalent to the smoothing algorithm in rfc793 with an
6588 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6589 * srtt'*8 = rtt + srtt*7
6590 * srtt'*8 = srtt*8 + rtt - srtt
6591 * srtt' = srtt + rtt/8 - srtt/8
6592 * srtt' = srtt + (rtt - srtt)/8
6595 delta = _8THMSEC(rttp) - peer->rtt;
6596 peer->rtt += (delta >> 3);
6599 * We accumulate a smoothed rtt variance (actually, a smoothed
6600 * mean difference), then set the retransmit timer to smoothed
6601 * rtt + 4 times the smoothed variance (was 2x in van's original
6602 * paper, but 4x works better for me, and apparently for him as
6604 * rttvar is stored as
6605 * fixed point with 2 bits after the binary point (scaled by
6606 * 4). The following is equivalent to rfc793 smoothing with
6607 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6608 * rttvar'*4 = rttvar*3 + |delta|
6609 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6610 * rttvar' = rttvar + |delta|/4 - rttvar/4
6611 * rttvar' = rttvar + (|delta| - rttvar)/4
6612 * This replaces rfc793's wired-in beta.
6613 * dev*4 = dev*4 + (|actual - expected| - dev)
6619 delta -= (peer->rtt_dev << 1);
6620 peer->rtt_dev += (delta >> 3);
6622 /* I don't have a stored RTT so I start with this value. Since I'm
6623 * probably just starting a call, and will be pushing more data down
6624 * this, I expect congestion to increase rapidly. So I fudge a
6625 * little, and I set deviance to half the rtt. In practice,
6626 * deviance tends to approach something a little less than
6627 * half the smoothed rtt. */
6628 peer->rtt = _8THMSEC(rttp) + 8;
6629 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
6631 /* the timeout is RTT + 4*MDEV but no less than rx_minPeerTimeout msec.
6632 * This is because one end or the other of these connections is usually
6633 * in a user process, and can be switched and/or swapped out. So on fast,
6634 * reliable networks, the timeout would otherwise be too short. */
6635 rtt_timeout = MAX(((peer->rtt >> 3) + peer->rtt_dev), rx_minPeerTimeout);
6636 clock_Zero(&(peer->timeout));
6637 clock_Addmsec(&(peer->timeout), rtt_timeout);
6639 /* Reset the backedOff flag since we just computed a new timeout value */
6640 peer->backedOff = 0;
6642 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6643 p->header.callNumber, p, MSEC(rttp), peer->rtt >> 3, peer->rtt_dev >> 2, (peer->timeout.sec), (peer->timeout.usec)));
6647 /* Find all server connections that have not been active for a long time, and
6650 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6652 struct clock now, when;
6653 clock_GetTime(&now);
6655 /* Find server connection structures that haven't been used for
6656 * greater than rx_idleConnectionTime */
6658 struct rx_connection **conn_ptr, **conn_end;
6659 int i, havecalls = 0;
6660 MUTEX_ENTER(&rx_connHashTable_lock);
6661 for (conn_ptr = &rx_connHashTable[0], conn_end =
6662 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6664 struct rx_connection *conn, *next;
6665 struct rx_call *call;
6669 for (conn = *conn_ptr; conn; conn = next) {
6670 /* XXX -- Shouldn't the connection be locked? */
6673 for (i = 0; i < RX_MAXCALLS; i++) {
6674 call = conn->call[i];
6678 code = MUTEX_TRYENTER(&call->lock);
6681 #ifdef RX_ENABLE_LOCKS
6682 result = rxi_CheckCall(call, 1);
6683 #else /* RX_ENABLE_LOCKS */
6684 result = rxi_CheckCall(call);
6685 #endif /* RX_ENABLE_LOCKS */
6686 MUTEX_EXIT(&call->lock);
6688 /* If CheckCall freed the call, it might
6689 * have destroyed the connection as well,
6690 * which screws up the linked lists.
6696 if (conn->type == RX_SERVER_CONNECTION) {
6697 /* This only actually destroys the connection if
6698 * there are no outstanding calls */
6699 MUTEX_ENTER(&conn->conn_data_lock);
6700 MUTEX_ENTER(&rx_refcnt_mutex);
6701 if (!havecalls && !conn->refCount
6702 && ((conn->lastSendTime + rx_idleConnectionTime) <
6704 conn->refCount++; /* it will be decr in rx_DestroyConn */
6705 MUTEX_EXIT(&rx_refcnt_mutex);
6706 MUTEX_EXIT(&conn->conn_data_lock);
6707 #ifdef RX_ENABLE_LOCKS
6708 rxi_DestroyConnectionNoLock(conn);
6709 #else /* RX_ENABLE_LOCKS */
6710 rxi_DestroyConnection(conn);
6711 #endif /* RX_ENABLE_LOCKS */
6713 #ifdef RX_ENABLE_LOCKS
6715 MUTEX_EXIT(&rx_refcnt_mutex);
6716 MUTEX_EXIT(&conn->conn_data_lock);
6718 #endif /* RX_ENABLE_LOCKS */
6722 #ifdef RX_ENABLE_LOCKS
6723 while (rx_connCleanup_list) {
6724 struct rx_connection *conn;
6725 conn = rx_connCleanup_list;
6726 rx_connCleanup_list = rx_connCleanup_list->next;
6727 MUTEX_EXIT(&rx_connHashTable_lock);
6728 rxi_CleanupConnection(conn);
6729 MUTEX_ENTER(&rx_connHashTable_lock);
6731 MUTEX_EXIT(&rx_connHashTable_lock);
6732 #endif /* RX_ENABLE_LOCKS */
6735 /* Find any peer structures that haven't been used (haven't had an
6736 * associated connection) for greater than rx_idlePeerTime */
6738 struct rx_peer **peer_ptr, **peer_end;
6742 * Why do we need to hold the rx_peerHashTable_lock across
6743 * the incrementing of peer_ptr since the rx_peerHashTable
6744 * array is not changing? We don't.
6746 * By dropping the lock periodically we can permit other
6747 * activities to be performed while a rxi_ReapConnections
6748 * call is in progress. The goal of reap connections
6749 * is to clean up quickly without causing large amounts
6750 * of contention. Therefore, it is important that global
6751 * mutexes not be held for extended periods of time.
6753 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6754 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6756 struct rx_peer *peer, *next, *prev;
6758 MUTEX_ENTER(&rx_peerHashTable_lock);
6759 for (prev = peer = *peer_ptr; peer; peer = next) {
6761 code = MUTEX_TRYENTER(&peer->peer_lock);
6762 if ((code) && (peer->refCount == 0)
6763 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6764 rx_interface_stat_p rpc_stat, nrpc_stat;
6768 * now know that this peer object is one to be
6769 * removed from the hash table. Once it is removed
6770 * it can't be referenced by other threads.
6771 * Lets remove it first and decrement the struct
6772 * nPeerStructs count.
6774 if (peer == *peer_ptr) {
6780 if (rx_stats_active)
6781 rx_atomic_dec(&rx_stats.nPeerStructs);
6784 * Now if we hold references on 'prev' and 'next'
6785 * we can safely drop the rx_peerHashTable_lock
6786 * while we destroy this 'peer' object.
6792 MUTEX_EXIT(&rx_peerHashTable_lock);
6794 MUTEX_EXIT(&peer->peer_lock);
6795 MUTEX_DESTROY(&peer->peer_lock);
6797 (&peer->rpcStats, rpc_stat, nrpc_stat,
6798 rx_interface_stat)) {
6799 unsigned int num_funcs;
6802 queue_Remove(&rpc_stat->queue_header);
6803 queue_Remove(&rpc_stat->all_peers);
6804 num_funcs = rpc_stat->stats[0].func_total;
6806 sizeof(rx_interface_stat_t) +
6807 rpc_stat->stats[0].func_total *
6808 sizeof(rx_function_entry_v1_t);
6810 rxi_Free(rpc_stat, space);
6812 MUTEX_ENTER(&rx_rpc_stats);
6813 rxi_rpc_peer_stat_cnt -= num_funcs;
6814 MUTEX_EXIT(&rx_rpc_stats);
6819 * Regain the rx_peerHashTable_lock and
6820 * decrement the reference count on 'prev'
6823 MUTEX_ENTER(&rx_peerHashTable_lock);
6830 MUTEX_EXIT(&peer->peer_lock);
6835 MUTEX_EXIT(&rx_peerHashTable_lock);
6839 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6840 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6841 * GC, just below. Really, we shouldn't have to keep moving packets from
6842 * one place to another, but instead ought to always know if we can
6843 * afford to hold onto a packet in its particular use. */
6844 MUTEX_ENTER(&rx_freePktQ_lock);
6845 if (rx_waitingForPackets) {
6846 rx_waitingForPackets = 0;
6847 #ifdef RX_ENABLE_LOCKS
6848 CV_BROADCAST(&rx_waitingForPackets_cv);
6850 osi_rxWakeup(&rx_waitingForPackets);
6853 MUTEX_EXIT(&rx_freePktQ_lock);
6856 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6857 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6861 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6862 * rx.h is sort of strange this is better. This is called with a security
6863 * object before it is discarded. Each connection using a security object has
6864 * its own refcount to the object so it won't actually be freed until the last
6865 * connection is destroyed.
6867 * This is the only rxs module call. A hold could also be written but no one
6871 rxs_Release(struct rx_securityClass *aobj)
6873 return RXS_Close(aobj);
6877 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6878 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6879 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6880 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6882 /* Adjust our estimate of the transmission rate to this peer, given
6883 * that the packet p was just acked. We can adjust peer->timeout and
6884 * call->twind. Pragmatically, this is called
6885 * only with packets of maximal length.
6886 * Called with peer and call locked.
6890 rxi_ComputeRate(struct rx_peer *peer, struct rx_call *call,
6891 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6893 afs_int32 xferSize, xferMs;
6897 /* Count down packets */
6898 if (peer->rateFlag > 0)
6900 /* Do nothing until we're enabled */
6901 if (peer->rateFlag != 0)
6906 /* Count only when the ack seems legitimate */
6907 switch (ackReason) {
6908 case RX_ACK_REQUESTED:
6910 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6914 case RX_ACK_PING_RESPONSE:
6915 if (p) /* want the response to ping-request, not data send */
6917 clock_GetTime(&newTO);
6918 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6919 clock_Sub(&newTO, &call->pingRequestTime);
6920 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6924 xferSize = rx_AckDataSize(rx_maxSendWindow) + RX_HEADER_SIZE;
6931 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %d.%06d, rtt %u, ps %u)\n",
6932 ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"),
6933 xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
6935 /* Track only packets that are big enough. */
6936 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6940 /* absorb RTT data (in milliseconds) for these big packets */
6941 if (peer->smRtt == 0) {
6942 peer->smRtt = xferMs;
6944 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6949 if (peer->countDown) {
6953 peer->countDown = 10; /* recalculate only every so often */
6955 /* In practice, we can measure only the RTT for full packets,
6956 * because of the way Rx acks the data that it receives. (If it's
6957 * smaller than a full packet, it often gets implicitly acked
6958 * either by the call response (from a server) or by the next call
6959 * (from a client), and either case confuses transmission times
6960 * with processing times.) Therefore, replace the above
6961 * more-sophisticated processing with a simpler version, where the
6962 * smoothed RTT is kept for full-size packets, and the time to
6963 * transmit a windowful of full-size packets is simply RTT *
6964 * windowSize. Again, we take two steps:
6965 - ensure the timeout is large enough for a single packet's RTT;
6966 - ensure that the window is small enough to fit in the desired timeout.*/
6968 /* First, the timeout check. */
6969 minTime = peer->smRtt;
6970 /* Get a reasonable estimate for a timeout period */
6972 newTO.sec = minTime / 1000;
6973 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6975 /* Increase the timeout period so that we can always do at least
6976 * one packet exchange */
6977 if (clock_Gt(&newTO, &peer->timeout)) {
6979 dpf(("CONG peer %lx/%u: timeout %d.%06d ==> %ld.%06d (rtt %u)\n",
6980 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec,
6981 newTO.sec, newTO.usec, peer->smRtt));
6983 peer->timeout = newTO;
6986 /* Now, get an estimate for the transmit window size. */
6987 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6988 /* Now, convert to the number of full packets that could fit in a
6989 * reasonable fraction of that interval */
6990 minTime /= (peer->smRtt << 1);
6991 minTime = MAX(minTime, rx_minPeerTimeout);
6992 xferSize = minTime; /* (make a copy) */
6994 /* Now clamp the size to reasonable bounds. */
6997 else if (minTime > rx_maxSendWindow)
6998 minTime = rx_maxSendWindow;
6999 /* if (minTime != peer->maxWindow) {
7000 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u)\n",
7001 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
7002 peer->timeout.sec, peer->timeout.usec, peer->smRtt));
7003 peer->maxWindow = minTime;
7004 elide... call->twind = minTime;
7008 /* Cut back on the peer timeout if it had earlier grown unreasonably.
7009 * Discern this by calculating the timeout necessary for rx_Window
7011 if ((xferSize > rx_maxSendWindow) && (peer->timeout.sec >= 3)) {
7012 /* calculate estimate for transmission interval in milliseconds */
7013 minTime = rx_maxSendWindow * peer->smRtt;
7014 if (minTime < 1000) {
7015 dpf(("CONG peer %lx/%u: cut TO %d.%06d by 0.5 (rtt %u)\n",
7016 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
7017 peer->timeout.usec, peer->smRtt));
7019 newTO.sec = 0; /* cut back on timeout by half a second */
7020 newTO.usec = 500000;
7021 clock_Sub(&peer->timeout, &newTO);
7026 } /* end of rxi_ComputeRate */
7027 #endif /* ADAPT_WINDOW */
7035 #define TRACE_OPTION_RX_DEBUG 16
7043 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7044 0, KEY_QUERY_VALUE, &parmKey);
7045 if (code != ERROR_SUCCESS)
7048 dummyLen = sizeof(TraceOption);
7049 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7050 (BYTE *) &TraceOption, &dummyLen);
7051 if (code == ERROR_SUCCESS) {
7052 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7054 RegCloseKey (parmKey);
7055 #endif /* AFS_NT40_ENV */
7060 rx_DebugOnOff(int on)
7064 rxdebug_active = on;
7070 rx_StatsOnOff(int on)
7073 rx_stats_active = on;
7078 /* Don't call this debugging routine directly; use dpf */
7080 rxi_DebugPrint(char *format, ...)
7089 va_start(ap, format);
7091 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7094 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7096 OutputDebugString(msg);
7102 va_start(ap, format);
7104 clock_GetTime(&now);
7105 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7106 (unsigned int)now.usec);
7107 vfprintf(rx_Log, format, ap);
7115 * This function is used to process the rx_stats structure that is local
7116 * to a process as well as an rx_stats structure received from a remote
7117 * process (via rxdebug). Therefore, it needs to do minimal version
7121 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7122 afs_int32 freePackets, char version)
7126 if (size != sizeof(struct rx_statistics)) {
7128 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7129 size, sizeof(struct rx_statistics));
7132 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7135 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7136 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7137 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7138 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7139 s->specialPktAllocFailures);
7141 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7142 s->receivePktAllocFailures, s->sendPktAllocFailures,
7143 s->specialPktAllocFailures);
7147 " greedy %u, " "bogusReads %u (last from host %x), "
7148 "noPackets %u, " "noBuffers %u, " "selects %u, "
7149 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7150 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7151 s->selects, s->sendSelects);
7153 fprintf(file, " packets read: ");
7154 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7155 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7157 fprintf(file, "\n");
7160 " other read counters: data %u, " "ack %u, " "dup %u "
7161 "spurious %u " "dally %u\n", s->dataPacketsRead,
7162 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7163 s->ignorePacketDally);
7165 fprintf(file, " packets sent: ");
7166 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7167 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7169 fprintf(file, "\n");
7172 " other send counters: ack %u, " "data %u (not resends), "
7173 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7174 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7175 s->dataPacketsPushed, s->ignoreAckedPacket);
7178 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7179 s->netSendFailures, (int)s->fatalErrors);
7181 if (s->nRttSamples) {
7182 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7183 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7185 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7186 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7190 " %d server connections, " "%d client connections, "
7191 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7192 s->nServerConns, s->nClientConns, s->nPeerStructs,
7193 s->nCallStructs, s->nFreeCallStructs);
7195 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7196 fprintf(file, " %d clock updates\n", clock_nUpdates);
7200 /* for backward compatibility */
7202 rx_PrintStats(FILE * file)
7204 MUTEX_ENTER(&rx_stats_mutex);
7205 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7206 sizeof(rx_stats), rx_nFreePackets,
7208 MUTEX_EXIT(&rx_stats_mutex);
7212 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7214 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
7215 ntohl(peer->host), (int)ntohs(peer->port), (int)peer->burstSize,
7216 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
7219 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
7220 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
7221 (int)peer->timeout.usec, peer->nSent, peer->reSends);
7224 " Packet size %d, " "max in packet skew %d, "
7225 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
7226 (int)peer->outPacketSkew);
7230 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7232 * This mutex protects the following static variables:
7236 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7237 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7239 #define LOCK_RX_DEBUG
7240 #define UNLOCK_RX_DEBUG
7241 #endif /* AFS_PTHREAD_ENV */
7243 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7245 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7246 u_char type, void *inputData, size_t inputLength,
7247 void *outputData, size_t outputLength)
7249 static afs_int32 counter = 100;
7250 time_t waitTime, waitCount;
7251 struct rx_header theader;
7254 struct timeval tv_now, tv_wake, tv_delta;
7255 struct sockaddr_in taddr, faddr;
7269 tp = &tbuffer[sizeof(struct rx_header)];
7270 taddr.sin_family = AF_INET;
7271 taddr.sin_port = remotePort;
7272 taddr.sin_addr.s_addr = remoteAddr;
7273 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7274 taddr.sin_len = sizeof(struct sockaddr_in);
7277 memset(&theader, 0, sizeof(theader));
7278 theader.epoch = htonl(999);
7280 theader.callNumber = htonl(counter);
7283 theader.type = type;
7284 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7285 theader.serviceId = 0;
7287 memcpy(tbuffer, &theader, sizeof(theader));
7288 memcpy(tp, inputData, inputLength);
7290 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7291 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7293 /* see if there's a packet available */
7294 gettimeofday(&tv_wake,0);
7295 tv_wake.tv_sec += waitTime;
7298 FD_SET(socket, &imask);
7299 tv_delta.tv_sec = tv_wake.tv_sec;
7300 tv_delta.tv_usec = tv_wake.tv_usec;
7301 gettimeofday(&tv_now, 0);
7303 if (tv_delta.tv_usec < tv_now.tv_usec) {
7305 tv_delta.tv_usec += 1000000;
7308 tv_delta.tv_usec -= tv_now.tv_usec;
7310 if (tv_delta.tv_sec < tv_now.tv_sec) {
7314 tv_delta.tv_sec -= tv_now.tv_sec;
7317 code = select(0, &imask, 0, 0, &tv_delta);
7318 #else /* AFS_NT40_ENV */
7319 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7320 #endif /* AFS_NT40_ENV */
7321 if (code == 1 && FD_ISSET(socket, &imask)) {
7322 /* now receive a packet */
7323 faddrLen = sizeof(struct sockaddr_in);
7325 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7326 (struct sockaddr *)&faddr, &faddrLen);
7329 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7330 if (counter == ntohl(theader.callNumber))
7338 /* see if we've timed out */
7346 code -= sizeof(struct rx_header);
7347 if (code > outputLength)
7348 code = outputLength;
7349 memcpy(outputData, tp, code);
7352 #endif /* RXDEBUG */
7355 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7356 afs_uint16 remotePort, struct rx_debugStats * stat,
7357 afs_uint32 * supportedValues)
7359 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7361 struct rx_debugIn in;
7363 *supportedValues = 0;
7364 in.type = htonl(RX_DEBUGI_GETSTATS);
7367 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7368 &in, sizeof(in), stat, sizeof(*stat));
7371 * If the call was successful, fixup the version and indicate
7372 * what contents of the stat structure are valid.
7373 * Also do net to host conversion of fields here.
7377 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7378 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7380 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7381 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7383 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7384 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7386 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7387 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7389 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7390 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7392 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7393 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7395 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7396 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7398 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7399 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7401 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7402 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7404 stat->nFreePackets = ntohl(stat->nFreePackets);
7405 stat->packetReclaims = ntohl(stat->packetReclaims);
7406 stat->callsExecuted = ntohl(stat->callsExecuted);
7407 stat->nWaiting = ntohl(stat->nWaiting);
7408 stat->idleThreads = ntohl(stat->idleThreads);
7409 stat->nWaited = ntohl(stat->nWaited);
7410 stat->nPackets = ntohl(stat->nPackets);
7419 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7420 afs_uint16 remotePort, struct rx_statistics * stat,
7421 afs_uint32 * supportedValues)
7423 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7425 struct rx_debugIn in;
7426 afs_int32 *lp = (afs_int32 *) stat;
7430 * supportedValues is currently unused, but added to allow future
7431 * versioning of this function.
7434 *supportedValues = 0;
7435 in.type = htonl(RX_DEBUGI_RXSTATS);
7437 memset(stat, 0, sizeof(*stat));
7439 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7440 &in, sizeof(in), stat, sizeof(*stat));
7445 * Do net to host conversion here
7448 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7459 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7460 afs_uint16 remotePort, size_t version_length,
7463 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7465 return MakeDebugCall(socket, remoteAddr, remotePort,
7466 RX_PACKET_TYPE_VERSION, a, 1, version,
7474 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7475 afs_uint16 remotePort, afs_int32 * nextConnection,
7476 int allConnections, afs_uint32 debugSupportedValues,
7477 struct rx_debugConn * conn,
7478 afs_uint32 * supportedValues)
7480 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7482 struct rx_debugIn in;
7486 * supportedValues is currently unused, but added to allow future
7487 * versioning of this function.
7490 *supportedValues = 0;
7491 if (allConnections) {
7492 in.type = htonl(RX_DEBUGI_GETALLCONN);
7494 in.type = htonl(RX_DEBUGI_GETCONN);
7496 in.index = htonl(*nextConnection);
7497 memset(conn, 0, sizeof(*conn));
7499 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7500 &in, sizeof(in), conn, sizeof(*conn));
7503 *nextConnection += 1;
7506 * Convert old connection format to new structure.
7509 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7510 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7511 #define MOVEvL(a) (conn->a = vL->a)
7513 /* any old or unrecognized version... */
7514 for (i = 0; i < RX_MAXCALLS; i++) {
7515 MOVEvL(callState[i]);
7516 MOVEvL(callMode[i]);
7517 MOVEvL(callFlags[i]);
7518 MOVEvL(callOther[i]);
7520 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7521 MOVEvL(secStats.type);
7522 MOVEvL(secStats.level);
7523 MOVEvL(secStats.flags);
7524 MOVEvL(secStats.expires);
7525 MOVEvL(secStats.packetsReceived);
7526 MOVEvL(secStats.packetsSent);
7527 MOVEvL(secStats.bytesReceived);
7528 MOVEvL(secStats.bytesSent);
7533 * Do net to host conversion here
7535 * I don't convert host or port since we are most likely
7536 * going to want these in NBO.
7538 conn->cid = ntohl(conn->cid);
7539 conn->serial = ntohl(conn->serial);
7540 for (i = 0; i < RX_MAXCALLS; i++) {
7541 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7543 conn->error = ntohl(conn->error);
7544 conn->secStats.flags = ntohl(conn->secStats.flags);
7545 conn->secStats.expires = ntohl(conn->secStats.expires);
7546 conn->secStats.packetsReceived =
7547 ntohl(conn->secStats.packetsReceived);
7548 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7549 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7550 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7551 conn->epoch = ntohl(conn->epoch);
7552 conn->natMTU = ntohl(conn->natMTU);
7561 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7562 afs_uint16 remotePort, afs_int32 * nextPeer,
7563 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7564 afs_uint32 * supportedValues)
7566 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7568 struct rx_debugIn in;
7571 * supportedValues is currently unused, but added to allow future
7572 * versioning of this function.
7575 *supportedValues = 0;
7576 in.type = htonl(RX_DEBUGI_GETPEER);
7577 in.index = htonl(*nextPeer);
7578 memset(peer, 0, sizeof(*peer));
7580 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7581 &in, sizeof(in), peer, sizeof(*peer));
7587 * Do net to host conversion here
7589 * I don't convert host or port since we are most likely
7590 * going to want these in NBO.
7592 peer->ifMTU = ntohs(peer->ifMTU);
7593 peer->idleWhen = ntohl(peer->idleWhen);
7594 peer->refCount = ntohs(peer->refCount);
7595 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7596 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7597 peer->rtt = ntohl(peer->rtt);
7598 peer->rtt_dev = ntohl(peer->rtt_dev);
7599 peer->timeout.sec = ntohl(peer->timeout.sec);
7600 peer->timeout.usec = ntohl(peer->timeout.usec);
7601 peer->nSent = ntohl(peer->nSent);
7602 peer->reSends = ntohl(peer->reSends);
7603 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7604 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7605 peer->rateFlag = ntohl(peer->rateFlag);
7606 peer->natMTU = ntohs(peer->natMTU);
7607 peer->maxMTU = ntohs(peer->maxMTU);
7608 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7609 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7610 peer->MTU = ntohs(peer->MTU);
7611 peer->cwind = ntohs(peer->cwind);
7612 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7613 peer->congestSeq = ntohs(peer->congestSeq);
7614 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7615 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7616 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7617 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7626 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7627 struct rx_debugPeer * peerStats)
7630 afs_int32 error = 1; /* default to "did not succeed" */
7631 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7633 MUTEX_ENTER(&rx_peerHashTable_lock);
7634 for(tp = rx_peerHashTable[hashValue];
7635 tp != NULL; tp = tp->next) {
7636 if (tp->host == peerHost)
7642 MUTEX_EXIT(&rx_peerHashTable_lock);
7646 MUTEX_ENTER(&tp->peer_lock);
7647 peerStats->host = tp->host;
7648 peerStats->port = tp->port;
7649 peerStats->ifMTU = tp->ifMTU;
7650 peerStats->idleWhen = tp->idleWhen;
7651 peerStats->refCount = tp->refCount;
7652 peerStats->burstSize = tp->burstSize;
7653 peerStats->burst = tp->burst;
7654 peerStats->burstWait.sec = tp->burstWait.sec;
7655 peerStats->burstWait.usec = tp->burstWait.usec;
7656 peerStats->rtt = tp->rtt;
7657 peerStats->rtt_dev = tp->rtt_dev;
7658 peerStats->timeout.sec = tp->timeout.sec;
7659 peerStats->timeout.usec = tp->timeout.usec;
7660 peerStats->nSent = tp->nSent;
7661 peerStats->reSends = tp->reSends;
7662 peerStats->inPacketSkew = tp->inPacketSkew;
7663 peerStats->outPacketSkew = tp->outPacketSkew;
7664 peerStats->rateFlag = tp->rateFlag;
7665 peerStats->natMTU = tp->natMTU;
7666 peerStats->maxMTU = tp->maxMTU;
7667 peerStats->maxDgramPackets = tp->maxDgramPackets;
7668 peerStats->ifDgramPackets = tp->ifDgramPackets;
7669 peerStats->MTU = tp->MTU;
7670 peerStats->cwind = tp->cwind;
7671 peerStats->nDgramPackets = tp->nDgramPackets;
7672 peerStats->congestSeq = tp->congestSeq;
7673 peerStats->bytesSent.high = tp->bytesSent.high;
7674 peerStats->bytesSent.low = tp->bytesSent.low;
7675 peerStats->bytesReceived.high = tp->bytesReceived.high;
7676 peerStats->bytesReceived.low = tp->bytesReceived.low;
7677 MUTEX_EXIT(&tp->peer_lock);
7679 MUTEX_ENTER(&rx_peerHashTable_lock);
7682 MUTEX_EXIT(&rx_peerHashTable_lock);
7690 struct rx_serverQueueEntry *np;
7693 struct rx_call *call;
7694 struct rx_serverQueueEntry *sq;
7698 if (rxinit_status == 1) {
7700 return; /* Already shutdown. */
7704 #ifndef AFS_PTHREAD_ENV
7705 FD_ZERO(&rx_selectMask);
7706 #endif /* AFS_PTHREAD_ENV */
7707 rxi_dataQuota = RX_MAX_QUOTA;
7708 #ifndef AFS_PTHREAD_ENV
7710 #endif /* AFS_PTHREAD_ENV */
7713 #ifndef AFS_PTHREAD_ENV
7714 #ifndef AFS_USE_GETTIMEOFDAY
7716 #endif /* AFS_USE_GETTIMEOFDAY */
7717 #endif /* AFS_PTHREAD_ENV */
7719 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7720 call = queue_First(&rx_freeCallQueue, rx_call);
7722 rxi_Free(call, sizeof(struct rx_call));
7725 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7726 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7732 struct rx_peer **peer_ptr, **peer_end;
7733 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7734 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7736 struct rx_peer *peer, *next;
7738 MUTEX_ENTER(&rx_peerHashTable_lock);
7739 for (peer = *peer_ptr; peer; peer = next) {
7740 rx_interface_stat_p rpc_stat, nrpc_stat;
7743 MUTEX_ENTER(&rx_rpc_stats);
7744 MUTEX_ENTER(&peer->peer_lock);
7746 (&peer->rpcStats, rpc_stat, nrpc_stat,
7747 rx_interface_stat)) {
7748 unsigned int num_funcs;
7751 queue_Remove(&rpc_stat->queue_header);
7752 queue_Remove(&rpc_stat->all_peers);
7753 num_funcs = rpc_stat->stats[0].func_total;
7755 sizeof(rx_interface_stat_t) +
7756 rpc_stat->stats[0].func_total *
7757 sizeof(rx_function_entry_v1_t);
7759 rxi_Free(rpc_stat, space);
7761 /* rx_rpc_stats must be held */
7762 rxi_rpc_peer_stat_cnt -= num_funcs;
7764 MUTEX_EXIT(&peer->peer_lock);
7765 MUTEX_EXIT(&rx_rpc_stats);
7769 if (rx_stats_active)
7770 rx_atomic_dec(&rx_stats.nPeerStructs);
7772 MUTEX_EXIT(&rx_peerHashTable_lock);
7775 for (i = 0; i < RX_MAX_SERVICES; i++) {
7777 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7779 for (i = 0; i < rx_hashTableSize; i++) {
7780 struct rx_connection *tc, *ntc;
7781 MUTEX_ENTER(&rx_connHashTable_lock);
7782 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7784 for (j = 0; j < RX_MAXCALLS; j++) {
7786 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7789 rxi_Free(tc, sizeof(*tc));
7791 MUTEX_EXIT(&rx_connHashTable_lock);
7794 MUTEX_ENTER(&freeSQEList_lock);
7796 while ((np = rx_FreeSQEList)) {
7797 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7798 MUTEX_DESTROY(&np->lock);
7799 rxi_Free(np, sizeof(*np));
7802 MUTEX_EXIT(&freeSQEList_lock);
7803 MUTEX_DESTROY(&freeSQEList_lock);
7804 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7805 MUTEX_DESTROY(&rx_connHashTable_lock);
7806 MUTEX_DESTROY(&rx_peerHashTable_lock);
7807 MUTEX_DESTROY(&rx_serverPool_lock);
7809 osi_Free(rx_connHashTable,
7810 rx_hashTableSize * sizeof(struct rx_connection *));
7811 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7813 UNPIN(rx_connHashTable,
7814 rx_hashTableSize * sizeof(struct rx_connection *));
7815 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7817 rxi_FreeAllPackets();
7819 MUTEX_ENTER(&rx_quota_mutex);
7820 rxi_dataQuota = RX_MAX_QUOTA;
7821 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7822 MUTEX_EXIT(&rx_quota_mutex);
7827 #ifdef RX_ENABLE_LOCKS
7829 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7831 if (!MUTEX_ISMINE(lockaddr))
7832 osi_Panic("Lock not held: %s", msg);
7834 #endif /* RX_ENABLE_LOCKS */
7839 * Routines to implement connection specific data.
7843 rx_KeyCreate(rx_destructor_t rtn)
7846 MUTEX_ENTER(&rxi_keyCreate_lock);
7847 key = rxi_keyCreate_counter++;
7848 rxi_keyCreate_destructor = (rx_destructor_t *)
7849 realloc((void *)rxi_keyCreate_destructor,
7850 (key + 1) * sizeof(rx_destructor_t));
7851 rxi_keyCreate_destructor[key] = rtn;
7852 MUTEX_EXIT(&rxi_keyCreate_lock);
7857 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7860 MUTEX_ENTER(&conn->conn_data_lock);
7861 if (!conn->specific) {
7862 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7863 for (i = 0; i < key; i++)
7864 conn->specific[i] = NULL;
7865 conn->nSpecific = key + 1;
7866 conn->specific[key] = ptr;
7867 } else if (key >= conn->nSpecific) {
7868 conn->specific = (void **)
7869 realloc(conn->specific, (key + 1) * sizeof(void *));
7870 for (i = conn->nSpecific; i < key; i++)
7871 conn->specific[i] = NULL;
7872 conn->nSpecific = key + 1;
7873 conn->specific[key] = ptr;
7875 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7876 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7877 conn->specific[key] = ptr;
7879 MUTEX_EXIT(&conn->conn_data_lock);
7883 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7886 MUTEX_ENTER(&svc->svc_data_lock);
7887 if (!svc->specific) {
7888 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
7889 for (i = 0; i < key; i++)
7890 svc->specific[i] = NULL;
7891 svc->nSpecific = key + 1;
7892 svc->specific[key] = ptr;
7893 } else if (key >= svc->nSpecific) {
7894 svc->specific = (void **)
7895 realloc(svc->specific, (key + 1) * sizeof(void *));
7896 for (i = svc->nSpecific; i < key; i++)
7897 svc->specific[i] = NULL;
7898 svc->nSpecific = key + 1;
7899 svc->specific[key] = ptr;
7901 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7902 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7903 svc->specific[key] = ptr;
7905 MUTEX_EXIT(&svc->svc_data_lock);
7909 rx_GetSpecific(struct rx_connection *conn, int key)
7912 MUTEX_ENTER(&conn->conn_data_lock);
7913 if (key >= conn->nSpecific)
7916 ptr = conn->specific[key];
7917 MUTEX_EXIT(&conn->conn_data_lock);
7922 rx_GetServiceSpecific(struct rx_service *svc, int key)
7925 MUTEX_ENTER(&svc->svc_data_lock);
7926 if (key >= svc->nSpecific)
7929 ptr = svc->specific[key];
7930 MUTEX_EXIT(&svc->svc_data_lock);
7935 #endif /* !KERNEL */
7938 * processStats is a queue used to store the statistics for the local
7939 * process. Its contents are similar to the contents of the rpcStats
7940 * queue on a rx_peer structure, but the actual data stored within
7941 * this queue contains totals across the lifetime of the process (assuming
7942 * the stats have not been reset) - unlike the per peer structures
7943 * which can come and go based upon the peer lifetime.
7946 static struct rx_queue processStats = { &processStats, &processStats };
7949 * peerStats is a queue used to store the statistics for all peer structs.
7950 * Its contents are the union of all the peer rpcStats queues.
7953 static struct rx_queue peerStats = { &peerStats, &peerStats };
7956 * rxi_monitor_processStats is used to turn process wide stat collection
7960 static int rxi_monitor_processStats = 0;
7963 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7966 static int rxi_monitor_peerStats = 0;
7969 * rxi_AddRpcStat - given all of the information for a particular rpc
7970 * call, create (if needed) and update the stat totals for the rpc.
7974 * IN stats - the queue of stats that will be updated with the new value
7976 * IN rxInterface - a unique number that identifies the rpc interface
7978 * IN currentFunc - the index of the function being invoked
7980 * IN totalFunc - the total number of functions in this interface
7982 * IN queueTime - the amount of time this function waited for a thread
7984 * IN execTime - the amount of time this function invocation took to execute
7986 * IN bytesSent - the number bytes sent by this invocation
7988 * IN bytesRcvd - the number bytes received by this invocation
7990 * IN isServer - if true, this invocation was made to a server
7992 * IN remoteHost - the ip address of the remote host
7994 * IN remotePort - the port of the remote host
7996 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7998 * INOUT counter - if a new stats structure is allocated, the counter will
7999 * be updated with the new number of allocated stat structures
8007 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
8008 afs_uint32 currentFunc, afs_uint32 totalFunc,
8009 struct clock *queueTime, struct clock *execTime,
8010 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
8011 afs_uint32 remoteHost, afs_uint32 remotePort,
8012 int addToPeerList, unsigned int *counter)
8015 rx_interface_stat_p rpc_stat, nrpc_stat;
8018 * See if there's already a structure for this interface
8021 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8022 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8023 && (rpc_stat->stats[0].remote_is_server == isServer))
8028 * Didn't find a match so allocate a new structure and add it to the
8032 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
8033 || (rpc_stat->stats[0].interfaceId != rxInterface)
8034 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8039 sizeof(rx_interface_stat_t) +
8040 totalFunc * sizeof(rx_function_entry_v1_t);
8042 rpc_stat = rxi_Alloc(space);
8043 if (rpc_stat == NULL) {
8047 *counter += totalFunc;
8048 for (i = 0; i < totalFunc; i++) {
8049 rpc_stat->stats[i].remote_peer = remoteHost;
8050 rpc_stat->stats[i].remote_port = remotePort;
8051 rpc_stat->stats[i].remote_is_server = isServer;
8052 rpc_stat->stats[i].interfaceId = rxInterface;
8053 rpc_stat->stats[i].func_total = totalFunc;
8054 rpc_stat->stats[i].func_index = i;
8055 hzero(rpc_stat->stats[i].invocations);
8056 hzero(rpc_stat->stats[i].bytes_sent);
8057 hzero(rpc_stat->stats[i].bytes_rcvd);
8058 rpc_stat->stats[i].queue_time_sum.sec = 0;
8059 rpc_stat->stats[i].queue_time_sum.usec = 0;
8060 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8061 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8062 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8063 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8064 rpc_stat->stats[i].queue_time_max.sec = 0;
8065 rpc_stat->stats[i].queue_time_max.usec = 0;
8066 rpc_stat->stats[i].execution_time_sum.sec = 0;
8067 rpc_stat->stats[i].execution_time_sum.usec = 0;
8068 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8069 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8070 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8071 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8072 rpc_stat->stats[i].execution_time_max.sec = 0;
8073 rpc_stat->stats[i].execution_time_max.usec = 0;
8075 queue_Prepend(stats, rpc_stat);
8076 if (addToPeerList) {
8077 queue_Prepend(&peerStats, &rpc_stat->all_peers);
8082 * Increment the stats for this function
8085 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
8086 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
8087 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
8088 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8089 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8090 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8091 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8093 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8094 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8096 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8097 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8099 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8100 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8102 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8103 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8111 * rx_IncrementTimeAndCount - increment the times and count for a particular
8116 * IN peer - the peer who invoked the rpc
8118 * IN rxInterface - a unique number that identifies the rpc interface
8120 * IN currentFunc - the index of the function being invoked
8122 * IN totalFunc - the total number of functions in this interface
8124 * IN queueTime - the amount of time this function waited for a thread
8126 * IN execTime - the amount of time this function invocation took to execute
8128 * IN bytesSent - the number bytes sent by this invocation
8130 * IN bytesRcvd - the number bytes received by this invocation
8132 * IN isServer - if true, this invocation was made to a server
8140 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8141 afs_uint32 currentFunc, afs_uint32 totalFunc,
8142 struct clock *queueTime, struct clock *execTime,
8143 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8147 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8150 MUTEX_ENTER(&rx_rpc_stats);
8152 if (rxi_monitor_peerStats) {
8153 MUTEX_ENTER(&peer->peer_lock);
8154 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8155 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8156 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8157 MUTEX_EXIT(&peer->peer_lock);
8160 if (rxi_monitor_processStats) {
8161 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8162 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8163 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8166 MUTEX_EXIT(&rx_rpc_stats);
8171 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8175 * IN callerVersion - the rpc stat version of the caller.
8177 * IN count - the number of entries to marshall.
8179 * IN stats - pointer to stats to be marshalled.
8181 * OUT ptr - Where to store the marshalled data.
8188 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8189 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8195 * We only support the first version
8197 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8198 *(ptr++) = stats->remote_peer;
8199 *(ptr++) = stats->remote_port;
8200 *(ptr++) = stats->remote_is_server;
8201 *(ptr++) = stats->interfaceId;
8202 *(ptr++) = stats->func_total;
8203 *(ptr++) = stats->func_index;
8204 *(ptr++) = hgethi(stats->invocations);
8205 *(ptr++) = hgetlo(stats->invocations);
8206 *(ptr++) = hgethi(stats->bytes_sent);
8207 *(ptr++) = hgetlo(stats->bytes_sent);
8208 *(ptr++) = hgethi(stats->bytes_rcvd);
8209 *(ptr++) = hgetlo(stats->bytes_rcvd);
8210 *(ptr++) = stats->queue_time_sum.sec;
8211 *(ptr++) = stats->queue_time_sum.usec;
8212 *(ptr++) = stats->queue_time_sum_sqr.sec;
8213 *(ptr++) = stats->queue_time_sum_sqr.usec;
8214 *(ptr++) = stats->queue_time_min.sec;
8215 *(ptr++) = stats->queue_time_min.usec;
8216 *(ptr++) = stats->queue_time_max.sec;
8217 *(ptr++) = stats->queue_time_max.usec;
8218 *(ptr++) = stats->execution_time_sum.sec;
8219 *(ptr++) = stats->execution_time_sum.usec;
8220 *(ptr++) = stats->execution_time_sum_sqr.sec;
8221 *(ptr++) = stats->execution_time_sum_sqr.usec;
8222 *(ptr++) = stats->execution_time_min.sec;
8223 *(ptr++) = stats->execution_time_min.usec;
8224 *(ptr++) = stats->execution_time_max.sec;
8225 *(ptr++) = stats->execution_time_max.usec;
8231 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8236 * IN callerVersion - the rpc stat version of the caller
8238 * OUT myVersion - the rpc stat version of this function
8240 * OUT clock_sec - local time seconds
8242 * OUT clock_usec - local time microseconds
8244 * OUT allocSize - the number of bytes allocated to contain stats
8246 * OUT statCount - the number stats retrieved from this process.
8248 * OUT stats - the actual stats retrieved from this process.
8252 * Returns void. If successful, stats will != NULL.
8256 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8257 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8258 size_t * allocSize, afs_uint32 * statCount,
8259 afs_uint32 ** stats)
8269 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8272 * Check to see if stats are enabled
8275 MUTEX_ENTER(&rx_rpc_stats);
8276 if (!rxi_monitor_processStats) {
8277 MUTEX_EXIT(&rx_rpc_stats);
8281 clock_GetTime(&now);
8282 *clock_sec = now.sec;
8283 *clock_usec = now.usec;
8286 * Allocate the space based upon the caller version
8288 * If the client is at an older version than we are,
8289 * we return the statistic data in the older data format, but
8290 * we still return our version number so the client knows we
8291 * are maintaining more data than it can retrieve.
8294 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8295 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8296 *statCount = rxi_rpc_process_stat_cnt;
8299 * This can't happen yet, but in the future version changes
8300 * can be handled by adding additional code here
8304 if (space > (size_t) 0) {
8306 ptr = *stats = rxi_Alloc(space);
8309 rx_interface_stat_p rpc_stat, nrpc_stat;
8313 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8315 * Copy the data based upon the caller version
8317 rx_MarshallProcessRPCStats(callerVersion,
8318 rpc_stat->stats[0].func_total,
8319 rpc_stat->stats, &ptr);
8325 MUTEX_EXIT(&rx_rpc_stats);
8330 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8334 * IN callerVersion - the rpc stat version of the caller
8336 * OUT myVersion - the rpc stat version of this function
8338 * OUT clock_sec - local time seconds
8340 * OUT clock_usec - local time microseconds
8342 * OUT allocSize - the number of bytes allocated to contain stats
8344 * OUT statCount - the number of stats retrieved from the individual
8347 * OUT stats - the actual stats retrieved from the individual peer structures.
8351 * Returns void. If successful, stats will != NULL.
8355 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8356 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8357 size_t * allocSize, afs_uint32 * statCount,
8358 afs_uint32 ** stats)
8368 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8371 * Check to see if stats are enabled
8374 MUTEX_ENTER(&rx_rpc_stats);
8375 if (!rxi_monitor_peerStats) {
8376 MUTEX_EXIT(&rx_rpc_stats);
8380 clock_GetTime(&now);
8381 *clock_sec = now.sec;
8382 *clock_usec = now.usec;
8385 * Allocate the space based upon the caller version
8387 * If the client is at an older version than we are,
8388 * we return the statistic data in the older data format, but
8389 * we still return our version number so the client knows we
8390 * are maintaining more data than it can retrieve.
8393 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8394 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8395 *statCount = rxi_rpc_peer_stat_cnt;
8398 * This can't happen yet, but in the future version changes
8399 * can be handled by adding additional code here
8403 if (space > (size_t) 0) {
8405 ptr = *stats = rxi_Alloc(space);
8408 rx_interface_stat_p rpc_stat, nrpc_stat;
8412 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8414 * We have to fix the offset of rpc_stat since we are
8415 * keeping this structure on two rx_queues. The rx_queue
8416 * package assumes that the rx_queue member is the first
8417 * member of the structure. That is, rx_queue assumes that
8418 * any one item is only on one queue at a time. We are
8419 * breaking that assumption and so we have to do a little
8420 * math to fix our pointers.
8423 fix_offset = (char *)rpc_stat;
8424 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8425 rpc_stat = (rx_interface_stat_p) fix_offset;
8428 * Copy the data based upon the caller version
8430 rx_MarshallProcessRPCStats(callerVersion,
8431 rpc_stat->stats[0].func_total,
8432 rpc_stat->stats, &ptr);
8438 MUTEX_EXIT(&rx_rpc_stats);
8443 * rx_FreeRPCStats - free memory allocated by
8444 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8448 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8449 * rx_RetrievePeerRPCStats
8451 * IN allocSize - the number of bytes in stats.
8459 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8461 rxi_Free(stats, allocSize);
8465 * rx_queryProcessRPCStats - see if process rpc stat collection is
8466 * currently enabled.
8472 * Returns 0 if stats are not enabled != 0 otherwise
8476 rx_queryProcessRPCStats(void)
8479 MUTEX_ENTER(&rx_rpc_stats);
8480 rc = rxi_monitor_processStats;
8481 MUTEX_EXIT(&rx_rpc_stats);
8486 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8492 * Returns 0 if stats are not enabled != 0 otherwise
8496 rx_queryPeerRPCStats(void)
8499 MUTEX_ENTER(&rx_rpc_stats);
8500 rc = rxi_monitor_peerStats;
8501 MUTEX_EXIT(&rx_rpc_stats);
8506 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8516 rx_enableProcessRPCStats(void)
8518 MUTEX_ENTER(&rx_rpc_stats);
8519 rx_enable_stats = 1;
8520 rxi_monitor_processStats = 1;
8521 MUTEX_EXIT(&rx_rpc_stats);
8525 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8535 rx_enablePeerRPCStats(void)
8537 MUTEX_ENTER(&rx_rpc_stats);
8538 rx_enable_stats = 1;
8539 rxi_monitor_peerStats = 1;
8540 MUTEX_EXIT(&rx_rpc_stats);
8544 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8554 rx_disableProcessRPCStats(void)
8556 rx_interface_stat_p rpc_stat, nrpc_stat;
8559 MUTEX_ENTER(&rx_rpc_stats);
8562 * Turn off process statistics and if peer stats is also off, turn
8566 rxi_monitor_processStats = 0;
8567 if (rxi_monitor_peerStats == 0) {
8568 rx_enable_stats = 0;
8571 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8572 unsigned int num_funcs = 0;
8575 queue_Remove(rpc_stat);
8576 num_funcs = rpc_stat->stats[0].func_total;
8578 sizeof(rx_interface_stat_t) +
8579 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8581 rxi_Free(rpc_stat, space);
8582 rxi_rpc_process_stat_cnt -= num_funcs;
8584 MUTEX_EXIT(&rx_rpc_stats);
8588 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8598 rx_disablePeerRPCStats(void)
8600 struct rx_peer **peer_ptr, **peer_end;
8604 * Turn off peer statistics and if process stats is also off, turn
8608 rxi_monitor_peerStats = 0;
8609 if (rxi_monitor_processStats == 0) {
8610 rx_enable_stats = 0;
8613 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8614 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8616 struct rx_peer *peer, *next, *prev;
8618 MUTEX_ENTER(&rx_peerHashTable_lock);
8619 MUTEX_ENTER(&rx_rpc_stats);
8620 for (prev = peer = *peer_ptr; peer; peer = next) {
8622 code = MUTEX_TRYENTER(&peer->peer_lock);
8624 rx_interface_stat_p rpc_stat, nrpc_stat;
8627 if (prev == *peer_ptr) {
8638 MUTEX_EXIT(&rx_peerHashTable_lock);
8641 (&peer->rpcStats, rpc_stat, nrpc_stat,
8642 rx_interface_stat)) {
8643 unsigned int num_funcs = 0;
8646 queue_Remove(&rpc_stat->queue_header);
8647 queue_Remove(&rpc_stat->all_peers);
8648 num_funcs = rpc_stat->stats[0].func_total;
8650 sizeof(rx_interface_stat_t) +
8651 rpc_stat->stats[0].func_total *
8652 sizeof(rx_function_entry_v1_t);
8654 rxi_Free(rpc_stat, space);
8655 rxi_rpc_peer_stat_cnt -= num_funcs;
8657 MUTEX_EXIT(&peer->peer_lock);
8659 MUTEX_ENTER(&rx_peerHashTable_lock);
8669 MUTEX_EXIT(&rx_rpc_stats);
8670 MUTEX_EXIT(&rx_peerHashTable_lock);
8675 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8680 * IN clearFlag - flag indicating which stats to clear
8688 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8690 rx_interface_stat_p rpc_stat, nrpc_stat;
8692 MUTEX_ENTER(&rx_rpc_stats);
8694 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8695 unsigned int num_funcs = 0, i;
8696 num_funcs = rpc_stat->stats[0].func_total;
8697 for (i = 0; i < num_funcs; i++) {
8698 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8699 hzero(rpc_stat->stats[i].invocations);
8701 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8702 hzero(rpc_stat->stats[i].bytes_sent);
8704 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8705 hzero(rpc_stat->stats[i].bytes_rcvd);
8707 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8708 rpc_stat->stats[i].queue_time_sum.sec = 0;
8709 rpc_stat->stats[i].queue_time_sum.usec = 0;
8711 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8712 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8713 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8715 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8716 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8717 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8719 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8720 rpc_stat->stats[i].queue_time_max.sec = 0;
8721 rpc_stat->stats[i].queue_time_max.usec = 0;
8723 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8724 rpc_stat->stats[i].execution_time_sum.sec = 0;
8725 rpc_stat->stats[i].execution_time_sum.usec = 0;
8727 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8728 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8729 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8731 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8732 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8733 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8735 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8736 rpc_stat->stats[i].execution_time_max.sec = 0;
8737 rpc_stat->stats[i].execution_time_max.usec = 0;
8742 MUTEX_EXIT(&rx_rpc_stats);
8746 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8751 * IN clearFlag - flag indicating which stats to clear
8759 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8761 rx_interface_stat_p rpc_stat, nrpc_stat;
8763 MUTEX_ENTER(&rx_rpc_stats);
8765 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8766 unsigned int num_funcs = 0, i;
8769 * We have to fix the offset of rpc_stat since we are
8770 * keeping this structure on two rx_queues. The rx_queue
8771 * package assumes that the rx_queue member is the first
8772 * member of the structure. That is, rx_queue assumes that
8773 * any one item is only on one queue at a time. We are
8774 * breaking that assumption and so we have to do a little
8775 * math to fix our pointers.
8778 fix_offset = (char *)rpc_stat;
8779 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8780 rpc_stat = (rx_interface_stat_p) fix_offset;
8782 num_funcs = rpc_stat->stats[0].func_total;
8783 for (i = 0; i < num_funcs; i++) {
8784 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8785 hzero(rpc_stat->stats[i].invocations);
8787 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8788 hzero(rpc_stat->stats[i].bytes_sent);
8790 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8791 hzero(rpc_stat->stats[i].bytes_rcvd);
8793 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8794 rpc_stat->stats[i].queue_time_sum.sec = 0;
8795 rpc_stat->stats[i].queue_time_sum.usec = 0;
8797 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8798 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8799 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8801 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8802 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8803 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8805 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8806 rpc_stat->stats[i].queue_time_max.sec = 0;
8807 rpc_stat->stats[i].queue_time_max.usec = 0;
8809 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8810 rpc_stat->stats[i].execution_time_sum.sec = 0;
8811 rpc_stat->stats[i].execution_time_sum.usec = 0;
8813 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8814 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8815 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8817 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8818 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8819 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8821 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8822 rpc_stat->stats[i].execution_time_max.sec = 0;
8823 rpc_stat->stats[i].execution_time_max.usec = 0;
8828 MUTEX_EXIT(&rx_rpc_stats);
8832 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8833 * is authorized to enable/disable/clear RX statistics.
8835 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8838 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8840 rxi_rxstat_userok = proc;
8844 rx_RxStatUserOk(struct rx_call *call)
8846 if (!rxi_rxstat_userok)
8848 return rxi_rxstat_userok(call);
8853 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8854 * function in the MSVC runtime DLL (msvcrt.dll).
8856 * Note: the system serializes calls to this function.
8859 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8860 DWORD reason, /* reason function is being called */
8861 LPVOID reserved) /* reserved for future use */
8864 case DLL_PROCESS_ATTACH:
8865 /* library is being attached to a process */
8869 case DLL_PROCESS_DETACH:
8876 #endif /* AFS_NT40_ENV */
8879 int rx_DumpCalls(FILE *outputFile, char *cookie)
8881 #ifdef RXDEBUG_PACKET
8882 #ifdef KDUMP_RX_LOCK
8883 struct rx_call_rx_lock *c;
8890 #define RXDPRINTF sprintf
8891 #define RXDPRINTOUT output
8893 #define RXDPRINTF fprintf
8894 #define RXDPRINTOUT outputFile
8897 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8899 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8902 for (c = rx_allCallsp; c; c = c->allNextp) {
8903 u_short rqc, tqc, iovqc;
8904 struct rx_packet *p, *np;
8906 MUTEX_ENTER(&c->lock);
8907 queue_Count(&c->rq, p, np, rx_packet, rqc);
8908 queue_Count(&c->tq, p, np, rx_packet, tqc);
8909 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8911 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, "
8912 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8913 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8914 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8915 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8916 #ifdef RX_ENABLE_LOCKS
8919 #ifdef RX_REFCOUNT_CHECK
8920 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
8921 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
8924 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,
8925 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
8926 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
8927 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
8928 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
8929 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
8930 #ifdef RX_ENABLE_LOCKS
8931 , (afs_uint32)c->refCount
8933 #ifdef RX_REFCOUNT_CHECK
8934 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
8937 MUTEX_EXIT(&c->lock);
8940 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8943 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
8945 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8947 #endif /* RXDEBUG_PACKET */