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
178 * Use procedural initialization of mutexes/condition variables
182 extern afs_kmutex_t rx_quota_mutex;
183 extern afs_kmutex_t rx_pthread_mutex;
184 extern afs_kmutex_t rx_packets_mutex;
185 extern afs_kmutex_t rx_refcnt_mutex;
186 extern afs_kmutex_t des_init_mutex;
187 extern afs_kmutex_t des_random_mutex;
188 extern afs_kmutex_t rx_clock_mutex;
189 extern afs_kmutex_t rxi_connCacheMutex;
190 extern afs_kmutex_t rx_event_mutex;
191 extern afs_kmutex_t osi_malloc_mutex;
192 extern afs_kmutex_t event_handler_mutex;
193 extern afs_kmutex_t listener_mutex;
194 extern afs_kmutex_t rx_if_init_mutex;
195 extern afs_kmutex_t rx_if_mutex;
196 extern afs_kmutex_t rxkad_client_uid_mutex;
197 extern afs_kmutex_t rxkad_random_mutex;
199 extern afs_kcondvar_t rx_event_handler_cond;
200 extern afs_kcondvar_t rx_listener_cond;
202 static afs_kmutex_t epoch_mutex;
203 static afs_kmutex_t rx_init_mutex;
204 static afs_kmutex_t rx_debug_mutex;
205 static afs_kmutex_t rx_rpc_stats;
208 rxi_InitPthread(void)
210 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
211 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
212 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
213 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
214 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
215 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
216 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
217 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
218 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
219 MUTEX_INIT(&rx_event_mutex, "event", MUTEX_DEFAULT, 0);
220 MUTEX_INIT(&des_init_mutex, "des", MUTEX_DEFAULT, 0);
221 MUTEX_INIT(&des_random_mutex, "random", MUTEX_DEFAULT, 0);
222 MUTEX_INIT(&osi_malloc_mutex, "malloc", MUTEX_DEFAULT, 0);
223 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
224 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
225 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
226 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
227 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
228 MUTEX_INIT(&rxkad_client_uid_mutex, "uid", MUTEX_DEFAULT, 0);
229 MUTEX_INIT(&rxkad_random_mutex, "rxkad random", MUTEX_DEFAULT, 0);
230 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
232 osi_Assert(pthread_cond_init
233 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
234 osi_Assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
236 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
237 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
239 rxkad_global_stats_init();
241 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
242 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
243 #ifdef RX_ENABLE_LOCKS
246 #endif /* RX_LOCKS_DB */
247 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
248 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
250 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
252 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
254 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
256 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
257 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
258 #endif /* RX_ENABLE_LOCKS */
261 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
262 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
264 * The rx_stats_mutex mutex protects the following global variables:
265 * rxi_lowConnRefCount
266 * rxi_lowPeerRefCount
275 * The rx_quota_mutex mutex protects the following global variables:
283 * The rx_freePktQ_lock protects the following global variables:
288 * The rx_packets_mutex mutex protects the following global variables:
296 * The rx_pthread_mutex mutex protects the following global variables:
297 * rxi_fcfs_thread_num
300 #define INIT_PTHREAD_LOCKS
304 /* Variables for handling the minProcs implementation. availProcs gives the
305 * number of threads available in the pool at this moment (not counting dudes
306 * executing right now). totalMin gives the total number of procs required
307 * for handling all minProcs requests. minDeficit is a dynamic variable
308 * tracking the # of procs required to satisfy all of the remaining minProcs
310 * For fine grain locking to work, the quota check and the reservation of
311 * a server thread has to come while rxi_availProcs and rxi_minDeficit
312 * are locked. To this end, the code has been modified under #ifdef
313 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
314 * same time. A new function, ReturnToServerPool() returns the allocation.
316 * A call can be on several queue's (but only one at a time). When
317 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
318 * that no one else is touching the queue. To this end, we store the address
319 * of the queue lock in the call structure (under the call lock) when we
320 * put the call on a queue, and we clear the call_queue_lock when the
321 * call is removed from a queue (once the call lock has been obtained).
322 * This allows rxi_ResetCall to safely synchronize with others wishing
323 * to manipulate the queue.
326 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
327 static afs_kmutex_t rx_rpc_stats;
328 void rxi_StartUnlocked(struct rxevent *event, void *call,
329 void *arg1, int istack);
332 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
333 ** pretty good that the next packet coming in is from the same connection
334 ** as the last packet, since we're send multiple packets in a transmit window.
336 struct rx_connection *rxLastConn = 0;
338 #ifdef RX_ENABLE_LOCKS
339 /* The locking hierarchy for rx fine grain locking is composed of these
342 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
343 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
344 * call->lock - locks call data fields.
345 * These are independent of each other:
346 * rx_freeCallQueue_lock
351 * serverQueueEntry->lock
352 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
354 * peer->lock - locks peer data fields.
355 * conn_data_lock - that more than one thread is not updating a conn data
356 * field at the same time.
367 * Do we need a lock to protect the peer field in the conn structure?
368 * conn->peer was previously a constant for all intents and so has no
369 * lock protecting this field. The multihomed client delta introduced
370 * a RX code change : change the peer field in the connection structure
371 * to that remote interface from which the last packet for this
372 * connection was sent out. This may become an issue if further changes
375 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
376 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
378 /* rxdb_fileID is used to identify the lock location, along with line#. */
379 static int rxdb_fileID = RXDB_FILE_RX;
380 #endif /* RX_LOCKS_DB */
381 #else /* RX_ENABLE_LOCKS */
382 #define SET_CALL_QUEUE_LOCK(C, L)
383 #define CLEAR_CALL_QUEUE_LOCK(C)
384 #endif /* RX_ENABLE_LOCKS */
385 struct rx_serverQueueEntry *rx_waitForPacket = 0;
386 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
388 /* ------------Exported Interfaces------------- */
390 /* This function allows rxkad to set the epoch to a suitably random number
391 * which rx_NewConnection will use in the future. The principle purpose is to
392 * get rxnull connections to use the same epoch as the rxkad connections do, at
393 * least once the first rxkad connection is established. This is important now
394 * that the host/port addresses aren't used in FindConnection: the uniqueness
395 * of epoch/cid matters and the start time won't do. */
397 #ifdef AFS_PTHREAD_ENV
399 * This mutex protects the following global variables:
403 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
404 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
408 #endif /* AFS_PTHREAD_ENV */
411 rx_SetEpoch(afs_uint32 epoch)
418 /* Initialize rx. A port number may be mentioned, in which case this
419 * becomes the default port number for any service installed later.
420 * If 0 is provided for the port number, a random port will be chosen
421 * by the kernel. Whether this will ever overlap anything in
422 * /etc/services is anybody's guess... Returns 0 on success, -1 on
427 int rxinit_status = 1;
428 #ifdef AFS_PTHREAD_ENV
430 * This mutex protects the following global variables:
434 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
435 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
438 #define UNLOCK_RX_INIT
442 rx_InitHost(u_int host, u_int port)
449 char *htable, *ptable;
456 if (rxinit_status == 0) {
457 tmp_status = rxinit_status;
459 return tmp_status; /* Already started; return previous error code. */
465 if (afs_winsockInit() < 0)
471 * Initialize anything necessary to provide a non-premptive threading
474 rxi_InitializeThreadSupport();
477 /* Allocate and initialize a socket for client and perhaps server
480 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
481 if (rx_socket == OSI_NULLSOCKET) {
485 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
488 #endif /* RX_LOCKS_DB */
489 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
490 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
491 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
492 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
493 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
494 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
495 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
496 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
497 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
499 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
501 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
503 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
505 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
506 #if defined(AFS_HPUX110_ENV)
508 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
509 #endif /* AFS_HPUX110_ENV */
510 #endif /* RX_ENABLE_LOCKS && KERNEL */
513 rx_connDeadTime = 12;
514 rx_tranquil = 0; /* reset flag */
515 rxi_ResetStatistics();
517 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
518 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
519 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
520 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
521 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
522 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
524 /* Malloc up a bunch of packets & buffers */
526 queue_Init(&rx_freePacketQueue);
527 rxi_NeedMorePackets = FALSE;
528 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
530 /* enforce a minimum number of allocated packets */
531 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
532 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
534 /* allocate the initial free packet pool */
535 #ifdef RX_ENABLE_TSFPQ
536 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
537 #else /* RX_ENABLE_TSFPQ */
538 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
539 #endif /* RX_ENABLE_TSFPQ */
546 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
547 tv.tv_sec = clock_now.sec;
548 tv.tv_usec = clock_now.usec;
549 srand((unsigned int)tv.tv_usec);
556 #if defined(KERNEL) && !defined(UKERNEL)
557 /* Really, this should never happen in a real kernel */
560 struct sockaddr_in addr;
562 int addrlen = sizeof(addr);
564 socklen_t addrlen = sizeof(addr);
566 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
570 rx_port = addr.sin_port;
573 rx_stats.minRtt.sec = 9999999;
575 rx_SetEpoch(tv.tv_sec | 0x80000000);
577 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
578 * will provide a randomer value. */
580 MUTEX_ENTER(&rx_quota_mutex);
581 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
582 MUTEX_EXIT(&rx_quota_mutex);
583 /* *Slightly* random start time for the cid. This is just to help
584 * out with the hashing function at the peer */
585 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
586 rx_connHashTable = (struct rx_connection **)htable;
587 rx_peerHashTable = (struct rx_peer **)ptable;
589 rx_lastAckDelay.sec = 0;
590 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
591 rx_hardAckDelay.sec = 0;
592 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
593 rx_softAckDelay.sec = 0;
594 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
596 rxevent_Init(20, rxi_ReScheduleEvents);
598 /* Initialize various global queues */
599 queue_Init(&rx_idleServerQueue);
600 queue_Init(&rx_incomingCallQueue);
601 queue_Init(&rx_freeCallQueue);
603 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
604 /* Initialize our list of usable IP addresses. */
608 /* Start listener process (exact function is dependent on the
609 * implementation environment--kernel or user space) */
613 tmp_status = rxinit_status = 0;
621 return rx_InitHost(htonl(INADDR_ANY), port);
624 /* called with unincremented nRequestsRunning to see if it is OK to start
625 * a new thread in this service. Could be "no" for two reasons: over the
626 * max quota, or would prevent others from reaching their min quota.
628 #ifdef RX_ENABLE_LOCKS
629 /* This verion of QuotaOK reserves quota if it's ok while the
630 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
633 QuotaOK(struct rx_service *aservice)
635 /* check if over max quota */
636 if (aservice->nRequestsRunning >= aservice->maxProcs) {
640 /* under min quota, we're OK */
641 /* otherwise, can use only if there are enough to allow everyone
642 * to go to their min quota after this guy starts.
645 MUTEX_ENTER(&rx_quota_mutex);
646 if ((aservice->nRequestsRunning < aservice->minProcs)
647 || (rxi_availProcs > rxi_minDeficit)) {
648 aservice->nRequestsRunning++;
649 /* just started call in minProcs pool, need fewer to maintain
651 if (aservice->nRequestsRunning <= aservice->minProcs)
654 MUTEX_EXIT(&rx_quota_mutex);
657 MUTEX_EXIT(&rx_quota_mutex);
663 ReturnToServerPool(struct rx_service *aservice)
665 aservice->nRequestsRunning--;
666 MUTEX_ENTER(&rx_quota_mutex);
667 if (aservice->nRequestsRunning < aservice->minProcs)
670 MUTEX_EXIT(&rx_quota_mutex);
673 #else /* RX_ENABLE_LOCKS */
675 QuotaOK(struct rx_service *aservice)
678 /* under min quota, we're OK */
679 if (aservice->nRequestsRunning < aservice->minProcs)
682 /* check if over max quota */
683 if (aservice->nRequestsRunning >= aservice->maxProcs)
686 /* otherwise, can use only if there are enough to allow everyone
687 * to go to their min quota after this guy starts.
689 MUTEX_ENTER(&rx_quota_mutex);
690 if (rxi_availProcs > rxi_minDeficit)
692 MUTEX_EXIT(&rx_quota_mutex);
695 #endif /* RX_ENABLE_LOCKS */
698 /* Called by rx_StartServer to start up lwp's to service calls.
699 NExistingProcs gives the number of procs already existing, and which
700 therefore needn't be created. */
702 rxi_StartServerProcs(int nExistingProcs)
704 struct rx_service *service;
709 /* For each service, reserve N processes, where N is the "minimum"
710 * number of processes that MUST be able to execute a request in parallel,
711 * at any time, for that process. Also compute the maximum difference
712 * between any service's maximum number of processes that can run
713 * (i.e. the maximum number that ever will be run, and a guarantee
714 * that this number will run if other services aren't running), and its
715 * minimum number. The result is the extra number of processes that
716 * we need in order to provide the latter guarantee */
717 for (i = 0; i < RX_MAX_SERVICES; i++) {
719 service = rx_services[i];
720 if (service == (struct rx_service *)0)
722 nProcs += service->minProcs;
723 diff = service->maxProcs - service->minProcs;
727 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
728 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
729 for (i = 0; i < nProcs; i++) {
730 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
736 /* This routine is only required on Windows */
738 rx_StartClientThread(void)
740 #ifdef AFS_PTHREAD_ENV
742 pid = pthread_self();
743 #endif /* AFS_PTHREAD_ENV */
745 #endif /* AFS_NT40_ENV */
747 /* This routine must be called if any services are exported. If the
748 * donateMe flag is set, the calling process is donated to the server
751 rx_StartServer(int donateMe)
753 struct rx_service *service;
759 /* Start server processes, if necessary (exact function is dependent
760 * on the implementation environment--kernel or user space). DonateMe
761 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
762 * case, one less new proc will be created rx_StartServerProcs.
764 rxi_StartServerProcs(donateMe);
766 /* count up the # of threads in minProcs, and add set the min deficit to
767 * be that value, too.
769 for (i = 0; i < RX_MAX_SERVICES; i++) {
770 service = rx_services[i];
771 if (service == (struct rx_service *)0)
773 MUTEX_ENTER(&rx_quota_mutex);
774 rxi_totalMin += service->minProcs;
775 /* below works even if a thread is running, since minDeficit would
776 * still have been decremented and later re-incremented.
778 rxi_minDeficit += service->minProcs;
779 MUTEX_EXIT(&rx_quota_mutex);
782 /* Turn on reaping of idle server connections */
783 rxi_ReapConnections(NULL, NULL, NULL);
792 #ifdef AFS_PTHREAD_ENV
794 pid = afs_pointer_to_int(pthread_self());
795 #else /* AFS_PTHREAD_ENV */
797 LWP_CurrentProcess(&pid);
798 #endif /* AFS_PTHREAD_ENV */
800 sprintf(name, "srv_%d", ++nProcs);
802 (*registerProgram) (pid, name);
804 #endif /* AFS_NT40_ENV */
805 rx_ServerProc(NULL); /* Never returns */
807 #ifdef RX_ENABLE_TSFPQ
808 /* no use leaving packets around in this thread's local queue if
809 * it isn't getting donated to the server thread pool.
811 rxi_FlushLocalPacketsTSFPQ();
812 #endif /* RX_ENABLE_TSFPQ */
816 /* Create a new client connection to the specified service, using the
817 * specified security object to implement the security model for this
819 struct rx_connection *
820 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
821 struct rx_securityClass *securityObject,
822 int serviceSecurityIndex)
826 struct rx_connection *conn;
831 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
832 "serviceSecurityIndex %d)\n",
833 ntohl(shost), ntohs(sport), sservice, securityObject,
834 serviceSecurityIndex));
836 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
837 * the case of kmem_alloc? */
838 conn = rxi_AllocConnection();
839 #ifdef RX_ENABLE_LOCKS
840 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
841 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
842 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
845 MUTEX_ENTER(&rx_connHashTable_lock);
846 cid = (rx_nextCid += RX_MAXCALLS);
847 conn->type = RX_CLIENT_CONNECTION;
849 conn->epoch = rx_epoch;
850 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
851 conn->serviceId = sservice;
852 conn->securityObject = securityObject;
853 conn->securityData = (void *) 0;
854 conn->securityIndex = serviceSecurityIndex;
855 rx_SetConnDeadTime(conn, rx_connDeadTime);
856 rx_SetConnSecondsUntilNatPing(conn, 0);
857 conn->ackRate = RX_FAST_ACK_RATE;
859 conn->specific = NULL;
860 conn->challengeEvent = NULL;
861 conn->delayedAbortEvent = NULL;
862 conn->abortCount = 0;
864 for (i = 0; i < RX_MAXCALLS; i++) {
865 conn->twind[i] = rx_initSendWindow;
866 conn->rwind[i] = rx_initReceiveWindow;
869 RXS_NewConnection(securityObject, conn);
871 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
873 conn->refCount++; /* no lock required since only this thread knows... */
874 conn->next = rx_connHashTable[hashindex];
875 rx_connHashTable[hashindex] = conn;
877 rx_atomic_inc(&rx_stats.nClientConns);
878 MUTEX_EXIT(&rx_connHashTable_lock);
884 * Ensure a connection's timeout values are valid.
886 * @param[in] conn The connection to check
888 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
889 * unless idleDeadTime and/or hardDeadTime are not set
893 rxi_CheckConnTimeouts(struct rx_connection *conn)
895 /* a connection's timeouts must have the relationship
896 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
897 * total loss of network to a peer may cause an idle timeout instead of a
898 * dead timeout, simply because the idle timeout gets hit first. Also set
899 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
900 /* this logic is slightly complicated by the fact that
901 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
903 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
904 if (conn->idleDeadTime) {
905 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
907 if (conn->hardDeadTime) {
908 if (conn->idleDeadTime) {
909 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
911 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
917 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
919 /* The idea is to set the dead time to a value that allows several
920 * keepalives to be dropped without timing out the connection. */
921 conn->secondsUntilDead = seconds;
922 rxi_CheckConnTimeouts(conn);
923 conn->secondsUntilPing = conn->secondsUntilDead / 6;
927 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
929 conn->hardDeadTime = seconds;
930 rxi_CheckConnTimeouts(conn);
934 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
936 conn->idleDeadTime = seconds;
937 rxi_CheckConnTimeouts(conn);
940 int rxi_lowPeerRefCount = 0;
941 int rxi_lowConnRefCount = 0;
944 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
945 * NOTE: must not be called with rx_connHashTable_lock held.
948 rxi_CleanupConnection(struct rx_connection *conn)
950 /* Notify the service exporter, if requested, that this connection
951 * is being destroyed */
952 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
953 (*conn->service->destroyConnProc) (conn);
955 /* Notify the security module that this connection is being destroyed */
956 RXS_DestroyConnection(conn->securityObject, conn);
958 /* If this is the last connection using the rx_peer struct, set its
959 * idle time to now. rxi_ReapConnections will reap it if it's still
960 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
962 MUTEX_ENTER(&rx_peerHashTable_lock);
963 if (conn->peer->refCount < 2) {
964 conn->peer->idleWhen = clock_Sec();
965 if (conn->peer->refCount < 1) {
966 conn->peer->refCount = 1;
967 if (rx_stats_active) {
968 MUTEX_ENTER(&rx_stats_mutex);
969 rxi_lowPeerRefCount++;
970 MUTEX_EXIT(&rx_stats_mutex);
974 conn->peer->refCount--;
975 MUTEX_EXIT(&rx_peerHashTable_lock);
979 if (conn->type == RX_SERVER_CONNECTION)
980 rx_atomic_dec(&rx_stats.nServerConns);
982 rx_atomic_dec(&rx_stats.nClientConns);
985 if (conn->specific) {
987 for (i = 0; i < conn->nSpecific; i++) {
988 if (conn->specific[i] && rxi_keyCreate_destructor[i])
989 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
990 conn->specific[i] = NULL;
992 free(conn->specific);
994 conn->specific = NULL;
998 MUTEX_DESTROY(&conn->conn_call_lock);
999 MUTEX_DESTROY(&conn->conn_data_lock);
1000 CV_DESTROY(&conn->conn_call_cv);
1002 rxi_FreeConnection(conn);
1005 /* Destroy the specified connection */
1007 rxi_DestroyConnection(struct rx_connection *conn)
1009 MUTEX_ENTER(&rx_connHashTable_lock);
1010 rxi_DestroyConnectionNoLock(conn);
1011 /* conn should be at the head of the cleanup list */
1012 if (conn == rx_connCleanup_list) {
1013 rx_connCleanup_list = rx_connCleanup_list->next;
1014 MUTEX_EXIT(&rx_connHashTable_lock);
1015 rxi_CleanupConnection(conn);
1017 #ifdef RX_ENABLE_LOCKS
1019 MUTEX_EXIT(&rx_connHashTable_lock);
1021 #endif /* RX_ENABLE_LOCKS */
1025 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1027 struct rx_connection **conn_ptr;
1029 struct rx_packet *packet;
1036 MUTEX_ENTER(&conn->conn_data_lock);
1037 MUTEX_ENTER(&rx_refcnt_mutex);
1038 if (conn->refCount > 0)
1041 if (rx_stats_active) {
1042 MUTEX_ENTER(&rx_stats_mutex);
1043 rxi_lowConnRefCount++;
1044 MUTEX_EXIT(&rx_stats_mutex);
1048 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1049 /* Busy; wait till the last guy before proceeding */
1050 MUTEX_EXIT(&rx_refcnt_mutex);
1051 MUTEX_EXIT(&conn->conn_data_lock);
1056 /* If the client previously called rx_NewCall, but it is still
1057 * waiting, treat this as a running call, and wait to destroy the
1058 * connection later when the call completes. */
1059 if ((conn->type == RX_CLIENT_CONNECTION)
1060 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1061 conn->flags |= RX_CONN_DESTROY_ME;
1062 MUTEX_EXIT(&conn->conn_data_lock);
1066 MUTEX_EXIT(&rx_refcnt_mutex);
1067 MUTEX_EXIT(&conn->conn_data_lock);
1069 /* Check for extant references to this connection */
1070 for (i = 0; i < RX_MAXCALLS; i++) {
1071 struct rx_call *call = conn->call[i];
1074 if (conn->type == RX_CLIENT_CONNECTION) {
1075 MUTEX_ENTER(&call->lock);
1076 if (call->delayedAckEvent) {
1077 /* Push the final acknowledgment out now--there
1078 * won't be a subsequent call to acknowledge the
1079 * last reply packets */
1080 rxevent_Cancel(call->delayedAckEvent, call,
1081 RX_CALL_REFCOUNT_DELAY);
1082 if (call->state == RX_STATE_PRECALL
1083 || call->state == RX_STATE_ACTIVE) {
1084 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1086 rxi_AckAll(NULL, call, 0);
1089 MUTEX_EXIT(&call->lock);
1093 #ifdef RX_ENABLE_LOCKS
1095 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1096 MUTEX_EXIT(&conn->conn_data_lock);
1098 /* Someone is accessing a packet right now. */
1102 #endif /* RX_ENABLE_LOCKS */
1105 /* Don't destroy the connection if there are any call
1106 * structures still in use */
1107 MUTEX_ENTER(&conn->conn_data_lock);
1108 conn->flags |= RX_CONN_DESTROY_ME;
1109 MUTEX_EXIT(&conn->conn_data_lock);
1114 if (conn->natKeepAliveEvent) {
1115 rxi_NatKeepAliveOff(conn);
1118 if (conn->delayedAbortEvent) {
1119 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1120 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1122 MUTEX_ENTER(&conn->conn_data_lock);
1123 rxi_SendConnectionAbort(conn, packet, 0, 1);
1124 MUTEX_EXIT(&conn->conn_data_lock);
1125 rxi_FreePacket(packet);
1129 /* Remove from connection hash table before proceeding */
1131 &rx_connHashTable[CONN_HASH
1132 (peer->host, peer->port, conn->cid, conn->epoch,
1134 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1135 if (*conn_ptr == conn) {
1136 *conn_ptr = conn->next;
1140 /* if the conn that we are destroying was the last connection, then we
1141 * clear rxLastConn as well */
1142 if (rxLastConn == conn)
1145 /* Make sure the connection is completely reset before deleting it. */
1146 /* get rid of pending events that could zap us later */
1147 if (conn->challengeEvent)
1148 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1149 if (conn->checkReachEvent)
1150 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1151 if (conn->natKeepAliveEvent)
1152 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
1154 /* Add the connection to the list of destroyed connections that
1155 * need to be cleaned up. This is necessary to avoid deadlocks
1156 * in the routines we call to inform others that this connection is
1157 * being destroyed. */
1158 conn->next = rx_connCleanup_list;
1159 rx_connCleanup_list = conn;
1162 /* Externally available version */
1164 rx_DestroyConnection(struct rx_connection *conn)
1169 rxi_DestroyConnection(conn);
1174 rx_GetConnection(struct rx_connection *conn)
1179 MUTEX_ENTER(&rx_refcnt_mutex);
1181 MUTEX_EXIT(&rx_refcnt_mutex);
1185 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1186 /* Wait for the transmit queue to no longer be busy.
1187 * requires the call->lock to be held */
1189 rxi_WaitforTQBusy(struct rx_call *call) {
1190 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1191 call->flags |= RX_CALL_TQ_WAIT;
1193 #ifdef RX_ENABLE_LOCKS
1194 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1195 CV_WAIT(&call->cv_tq, &call->lock);
1196 #else /* RX_ENABLE_LOCKS */
1197 osi_rxSleep(&call->tq);
1198 #endif /* RX_ENABLE_LOCKS */
1200 if (call->tqWaiters == 0) {
1201 call->flags &= ~RX_CALL_TQ_WAIT;
1207 /* Start a new rx remote procedure call, on the specified connection.
1208 * If wait is set to 1, wait for a free call channel; otherwise return
1209 * 0. Maxtime gives the maximum number of seconds this call may take,
1210 * after rx_NewCall returns. After this time interval, a call to any
1211 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1212 * For fine grain locking, we hold the conn_call_lock in order to
1213 * to ensure that we don't get signalle after we found a call in an active
1214 * state and before we go to sleep.
1217 rx_NewCall(struct rx_connection *conn)
1220 struct rx_call *call;
1221 struct clock queueTime;
1225 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1228 clock_GetTime(&queueTime);
1230 * Check if there are others waiting for a new call.
1231 * If so, let them go first to avoid starving them.
1232 * This is a fairly simple scheme, and might not be
1233 * a complete solution for large numbers of waiters.
1235 * makeCallWaiters keeps track of the number of
1236 * threads waiting to make calls and the
1237 * RX_CONN_MAKECALL_WAITING flag bit is used to
1238 * indicate that there are indeed calls waiting.
1239 * The flag is set when the waiter is incremented.
1240 * It is only cleared when makeCallWaiters is 0.
1241 * This prevents us from accidently destroying the
1242 * connection while it is potentially about to be used.
1244 MUTEX_ENTER(&conn->conn_call_lock);
1245 MUTEX_ENTER(&conn->conn_data_lock);
1246 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1247 conn->flags |= RX_CONN_MAKECALL_WAITING;
1248 conn->makeCallWaiters++;
1249 MUTEX_EXIT(&conn->conn_data_lock);
1251 #ifdef RX_ENABLE_LOCKS
1252 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1256 MUTEX_ENTER(&conn->conn_data_lock);
1257 conn->makeCallWaiters--;
1258 if (conn->makeCallWaiters == 0)
1259 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1262 /* We are now the active thread in rx_NewCall */
1263 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1264 MUTEX_EXIT(&conn->conn_data_lock);
1269 for (i = 0; i < RX_MAXCALLS; i++) {
1270 call = conn->call[i];
1272 if (call->state == RX_STATE_DALLY) {
1273 MUTEX_ENTER(&call->lock);
1274 if (call->state == RX_STATE_DALLY) {
1276 * We are setting the state to RX_STATE_RESET to
1277 * ensure that no one else will attempt to use this
1278 * call once we drop the conn->conn_call_lock and
1279 * call->lock. We must drop the conn->conn_call_lock
1280 * before calling rxi_ResetCall because the process
1281 * of clearing the transmit queue can block for an
1282 * extended period of time. If we block while holding
1283 * the conn->conn_call_lock, then all rx_EndCall
1284 * processing will block as well. This has a detrimental
1285 * effect on overall system performance.
1287 call->state = RX_STATE_RESET;
1288 MUTEX_EXIT(&conn->conn_call_lock);
1289 MUTEX_ENTER(&rx_refcnt_mutex);
1290 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1291 MUTEX_EXIT(&rx_refcnt_mutex);
1292 rxi_ResetCall(call, 0);
1293 (*call->callNumber)++;
1294 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1298 * If we failed to be able to safely obtain the
1299 * conn->conn_call_lock we will have to drop the
1300 * call->lock to avoid a deadlock. When the call->lock
1301 * is released the state of the call can change. If it
1302 * is no longer RX_STATE_RESET then some other thread is
1305 MUTEX_EXIT(&call->lock);
1306 MUTEX_ENTER(&conn->conn_call_lock);
1307 MUTEX_ENTER(&call->lock);
1309 if (call->state == RX_STATE_RESET)
1313 * If we get here it means that after dropping
1314 * the conn->conn_call_lock and call->lock that
1315 * the call is no longer ours. If we can't find
1316 * a free call in the remaining slots we should
1317 * not go immediately to RX_CONN_MAKECALL_WAITING
1318 * because by dropping the conn->conn_call_lock
1319 * we have given up synchronization with rx_EndCall.
1320 * Instead, cycle through one more time to see if
1321 * we can find a call that can call our own.
1323 MUTEX_ENTER(&rx_refcnt_mutex);
1324 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1325 MUTEX_EXIT(&rx_refcnt_mutex);
1328 MUTEX_EXIT(&call->lock);
1331 /* rxi_NewCall returns with mutex locked */
1332 call = rxi_NewCall(conn, i);
1333 MUTEX_ENTER(&rx_refcnt_mutex);
1334 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1335 MUTEX_EXIT(&rx_refcnt_mutex);
1339 if (i < RX_MAXCALLS) {
1345 MUTEX_ENTER(&conn->conn_data_lock);
1346 conn->flags |= RX_CONN_MAKECALL_WAITING;
1347 conn->makeCallWaiters++;
1348 MUTEX_EXIT(&conn->conn_data_lock);
1350 #ifdef RX_ENABLE_LOCKS
1351 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1355 MUTEX_ENTER(&conn->conn_data_lock);
1356 conn->makeCallWaiters--;
1357 if (conn->makeCallWaiters == 0)
1358 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1359 MUTEX_EXIT(&conn->conn_data_lock);
1361 /* Client is initially in send mode */
1362 call->state = RX_STATE_ACTIVE;
1363 call->error = conn->error;
1365 call->mode = RX_MODE_ERROR;
1367 call->mode = RX_MODE_SENDING;
1369 /* remember start time for call in case we have hard dead time limit */
1370 call->queueTime = queueTime;
1371 clock_GetTime(&call->startTime);
1372 hzero(call->bytesSent);
1373 hzero(call->bytesRcvd);
1375 /* Turn on busy protocol. */
1376 rxi_KeepAliveOn(call);
1378 /* Attempt MTU discovery */
1379 rxi_GrowMTUOn(call);
1382 * We are no longer the active thread in rx_NewCall
1384 MUTEX_ENTER(&conn->conn_data_lock);
1385 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1386 MUTEX_EXIT(&conn->conn_data_lock);
1389 * Wake up anyone else who might be giving us a chance to
1390 * run (see code above that avoids resource starvation).
1392 #ifdef RX_ENABLE_LOCKS
1393 CV_BROADCAST(&conn->conn_call_cv);
1397 MUTEX_EXIT(&conn->conn_call_lock);
1399 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1400 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1401 osi_Panic("rx_NewCall call about to be used without an empty tq");
1403 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1405 MUTEX_EXIT(&call->lock);
1408 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1413 rxi_HasActiveCalls(struct rx_connection *aconn)
1416 struct rx_call *tcall;
1420 for (i = 0; i < RX_MAXCALLS; i++) {
1421 if ((tcall = aconn->call[i])) {
1422 if ((tcall->state == RX_STATE_ACTIVE)
1423 || (tcall->state == RX_STATE_PRECALL)) {
1434 rxi_GetCallNumberVector(struct rx_connection *aconn,
1435 afs_int32 * aint32s)
1438 struct rx_call *tcall;
1442 for (i = 0; i < RX_MAXCALLS; i++) {
1443 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1444 aint32s[i] = aconn->callNumber[i] + 1;
1446 aint32s[i] = aconn->callNumber[i];
1453 rxi_SetCallNumberVector(struct rx_connection *aconn,
1454 afs_int32 * aint32s)
1457 struct rx_call *tcall;
1461 for (i = 0; i < RX_MAXCALLS; i++) {
1462 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1463 aconn->callNumber[i] = aint32s[i] - 1;
1465 aconn->callNumber[i] = aint32s[i];
1471 /* Advertise a new service. A service is named locally by a UDP port
1472 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1475 char *serviceName; Name for identification purposes (e.g. the
1476 service name might be used for probing for
1479 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1480 char *serviceName, struct rx_securityClass **securityObjects,
1481 int nSecurityObjects,
1482 afs_int32(*serviceProc) (struct rx_call * acall))
1484 osi_socket socket = OSI_NULLSOCKET;
1485 struct rx_service *tservice;
1491 if (serviceId == 0) {
1493 "rx_NewService: service id for service %s is not non-zero.\n",
1500 "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",
1508 tservice = rxi_AllocService();
1511 #ifdef RX_ENABLE_LOCKS
1512 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1515 for (i = 0; i < RX_MAX_SERVICES; i++) {
1516 struct rx_service *service = rx_services[i];
1518 if (port == service->servicePort && host == service->serviceHost) {
1519 if (service->serviceId == serviceId) {
1520 /* The identical service has already been
1521 * installed; if the caller was intending to
1522 * change the security classes used by this
1523 * service, he/she loses. */
1525 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1526 serviceName, serviceId, service->serviceName);
1528 rxi_FreeService(tservice);
1531 /* Different service, same port: re-use the socket
1532 * which is bound to the same port */
1533 socket = service->socket;
1536 if (socket == OSI_NULLSOCKET) {
1537 /* If we don't already have a socket (from another
1538 * service on same port) get a new one */
1539 socket = rxi_GetHostUDPSocket(host, port);
1540 if (socket == OSI_NULLSOCKET) {
1542 rxi_FreeService(tservice);
1547 service->socket = socket;
1548 service->serviceHost = host;
1549 service->servicePort = port;
1550 service->serviceId = serviceId;
1551 service->serviceName = serviceName;
1552 service->nSecurityObjects = nSecurityObjects;
1553 service->securityObjects = securityObjects;
1554 service->minProcs = 0;
1555 service->maxProcs = 1;
1556 service->idleDeadTime = 60;
1557 service->idleDeadErr = 0;
1558 service->connDeadTime = rx_connDeadTime;
1559 service->executeRequestProc = serviceProc;
1560 service->checkReach = 0;
1561 service->nSpecific = 0;
1562 service->specific = NULL;
1563 rx_services[i] = service; /* not visible until now */
1569 rxi_FreeService(tservice);
1570 (osi_Msg "rx_NewService: cannot support > %d services\n",
1575 /* Set configuration options for all of a service's security objects */
1578 rx_SetSecurityConfiguration(struct rx_service *service,
1579 rx_securityConfigVariables type,
1583 for (i = 0; i<service->nSecurityObjects; i++) {
1584 if (service->securityObjects[i]) {
1585 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1593 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1594 struct rx_securityClass **securityObjects, int nSecurityObjects,
1595 afs_int32(*serviceProc) (struct rx_call * acall))
1597 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1600 /* Generic request processing loop. This routine should be called
1601 * by the implementation dependent rx_ServerProc. If socketp is
1602 * non-null, it will be set to the file descriptor that this thread
1603 * is now listening on. If socketp is null, this routine will never
1606 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1608 struct rx_call *call;
1610 struct rx_service *tservice = NULL;
1617 call = rx_GetCall(threadID, tservice, socketp);
1618 if (socketp && *socketp != OSI_NULLSOCKET) {
1619 /* We are now a listener thread */
1624 /* if server is restarting( typically smooth shutdown) then do not
1625 * allow any new calls.
1628 if (rx_tranquil && (call != NULL)) {
1632 MUTEX_ENTER(&call->lock);
1634 rxi_CallError(call, RX_RESTARTING);
1635 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1637 MUTEX_EXIT(&call->lock);
1641 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1642 #ifdef RX_ENABLE_LOCKS
1644 #endif /* RX_ENABLE_LOCKS */
1645 afs_termState = AFSOP_STOP_AFS;
1646 afs_osi_Wakeup(&afs_termState);
1647 #ifdef RX_ENABLE_LOCKS
1649 #endif /* RX_ENABLE_LOCKS */
1654 tservice = call->conn->service;
1656 if (tservice->beforeProc)
1657 (*tservice->beforeProc) (call);
1659 code = tservice->executeRequestProc(call);
1661 if (tservice->afterProc)
1662 (*tservice->afterProc) (call, code);
1664 rx_EndCall(call, code);
1665 if (rx_stats_active) {
1666 MUTEX_ENTER(&rx_stats_mutex);
1668 MUTEX_EXIT(&rx_stats_mutex);
1675 rx_WakeupServerProcs(void)
1677 struct rx_serverQueueEntry *np, *tqp;
1681 MUTEX_ENTER(&rx_serverPool_lock);
1683 #ifdef RX_ENABLE_LOCKS
1684 if (rx_waitForPacket)
1685 CV_BROADCAST(&rx_waitForPacket->cv);
1686 #else /* RX_ENABLE_LOCKS */
1687 if (rx_waitForPacket)
1688 osi_rxWakeup(rx_waitForPacket);
1689 #endif /* RX_ENABLE_LOCKS */
1690 MUTEX_ENTER(&freeSQEList_lock);
1691 for (np = rx_FreeSQEList; np; np = tqp) {
1692 tqp = *(struct rx_serverQueueEntry **)np;
1693 #ifdef RX_ENABLE_LOCKS
1694 CV_BROADCAST(&np->cv);
1695 #else /* RX_ENABLE_LOCKS */
1697 #endif /* RX_ENABLE_LOCKS */
1699 MUTEX_EXIT(&freeSQEList_lock);
1700 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1701 #ifdef RX_ENABLE_LOCKS
1702 CV_BROADCAST(&np->cv);
1703 #else /* RX_ENABLE_LOCKS */
1705 #endif /* RX_ENABLE_LOCKS */
1707 MUTEX_EXIT(&rx_serverPool_lock);
1712 * One thing that seems to happen is that all the server threads get
1713 * tied up on some empty or slow call, and then a whole bunch of calls
1714 * arrive at once, using up the packet pool, so now there are more
1715 * empty calls. The most critical resources here are server threads
1716 * and the free packet pool. The "doreclaim" code seems to help in
1717 * general. I think that eventually we arrive in this state: there
1718 * are lots of pending calls which do have all their packets present,
1719 * so they won't be reclaimed, are multi-packet calls, so they won't
1720 * be scheduled until later, and thus are tying up most of the free
1721 * packet pool for a very long time.
1723 * 1. schedule multi-packet calls if all the packets are present.
1724 * Probably CPU-bound operation, useful to return packets to pool.
1725 * Do what if there is a full window, but the last packet isn't here?
1726 * 3. preserve one thread which *only* runs "best" calls, otherwise
1727 * it sleeps and waits for that type of call.
1728 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1729 * the current dataquota business is badly broken. The quota isn't adjusted
1730 * to reflect how many packets are presently queued for a running call.
1731 * So, when we schedule a queued call with a full window of packets queued
1732 * up for it, that *should* free up a window full of packets for other 2d-class
1733 * calls to be able to use from the packet pool. But it doesn't.
1735 * NB. Most of the time, this code doesn't run -- since idle server threads
1736 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1737 * as a new call arrives.
1739 /* Sleep until a call arrives. Returns a pointer to the call, ready
1740 * for an rx_Read. */
1741 #ifdef RX_ENABLE_LOCKS
1743 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1745 struct rx_serverQueueEntry *sq;
1746 struct rx_call *call = (struct rx_call *)0;
1747 struct rx_service *service = NULL;
1750 MUTEX_ENTER(&freeSQEList_lock);
1752 if ((sq = rx_FreeSQEList)) {
1753 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1754 MUTEX_EXIT(&freeSQEList_lock);
1755 } else { /* otherwise allocate a new one and return that */
1756 MUTEX_EXIT(&freeSQEList_lock);
1757 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1758 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1759 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1762 MUTEX_ENTER(&rx_serverPool_lock);
1763 if (cur_service != NULL) {
1764 ReturnToServerPool(cur_service);
1767 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1768 struct rx_call *tcall, *ncall, *choice2 = NULL;
1770 /* Scan for eligible incoming calls. A call is not eligible
1771 * if the maximum number of calls for its service type are
1772 * already executing */
1773 /* One thread will process calls FCFS (to prevent starvation),
1774 * while the other threads may run ahead looking for calls which
1775 * have all their input data available immediately. This helps
1776 * keep threads from blocking, waiting for data from the client. */
1777 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1778 service = tcall->conn->service;
1779 if (!QuotaOK(service)) {
1782 MUTEX_ENTER(&rx_pthread_mutex);
1783 if (tno == rxi_fcfs_thread_num
1784 || !tcall->queue_item_header.next) {
1785 MUTEX_EXIT(&rx_pthread_mutex);
1786 /* If we're the fcfs thread , then we'll just use
1787 * this call. If we haven't been able to find an optimal
1788 * choice, and we're at the end of the list, then use a
1789 * 2d choice if one has been identified. Otherwise... */
1790 call = (choice2 ? choice2 : tcall);
1791 service = call->conn->service;
1793 MUTEX_EXIT(&rx_pthread_mutex);
1794 if (!queue_IsEmpty(&tcall->rq)) {
1795 struct rx_packet *rp;
1796 rp = queue_First(&tcall->rq, rx_packet);
1797 if (rp->header.seq == 1) {
1799 || (rp->header.flags & RX_LAST_PACKET)) {
1801 } else if (rxi_2dchoice && !choice2
1802 && !(tcall->flags & RX_CALL_CLEARED)
1803 && (tcall->rprev > rxi_HardAckRate)) {
1813 ReturnToServerPool(service);
1820 MUTEX_EXIT(&rx_serverPool_lock);
1821 MUTEX_ENTER(&call->lock);
1823 if (call->flags & RX_CALL_WAIT_PROC) {
1824 call->flags &= ~RX_CALL_WAIT_PROC;
1825 rx_atomic_dec(&rx_nWaiting);
1828 if (call->state != RX_STATE_PRECALL || call->error) {
1829 MUTEX_EXIT(&call->lock);
1830 MUTEX_ENTER(&rx_serverPool_lock);
1831 ReturnToServerPool(service);
1836 if (queue_IsEmpty(&call->rq)
1837 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1838 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1840 CLEAR_CALL_QUEUE_LOCK(call);
1843 /* If there are no eligible incoming calls, add this process
1844 * to the idle server queue, to wait for one */
1848 *socketp = OSI_NULLSOCKET;
1850 sq->socketp = socketp;
1851 queue_Append(&rx_idleServerQueue, sq);
1852 #ifndef AFS_AIX41_ENV
1853 rx_waitForPacket = sq;
1855 rx_waitingForPacket = sq;
1856 #endif /* AFS_AIX41_ENV */
1858 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1860 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1861 MUTEX_EXIT(&rx_serverPool_lock);
1862 return (struct rx_call *)0;
1865 } while (!(call = sq->newcall)
1866 && !(socketp && *socketp != OSI_NULLSOCKET));
1867 MUTEX_EXIT(&rx_serverPool_lock);
1869 MUTEX_ENTER(&call->lock);
1875 MUTEX_ENTER(&freeSQEList_lock);
1876 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1877 rx_FreeSQEList = sq;
1878 MUTEX_EXIT(&freeSQEList_lock);
1881 clock_GetTime(&call->startTime);
1882 call->state = RX_STATE_ACTIVE;
1883 call->mode = RX_MODE_RECEIVING;
1884 #ifdef RX_KERNEL_TRACE
1885 if (ICL_SETACTIVE(afs_iclSetp)) {
1886 int glockOwner = ISAFS_GLOCK();
1889 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1890 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1897 rxi_calltrace(RX_CALL_START, call);
1898 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
1899 call->conn->service->servicePort, call->conn->service->serviceId,
1902 MUTEX_EXIT(&call->lock);
1903 MUTEX_ENTER(&rx_refcnt_mutex);
1904 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1905 MUTEX_EXIT(&rx_refcnt_mutex);
1907 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1912 #else /* RX_ENABLE_LOCKS */
1914 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1916 struct rx_serverQueueEntry *sq;
1917 struct rx_call *call = (struct rx_call *)0, *choice2;
1918 struct rx_service *service = NULL;
1922 MUTEX_ENTER(&freeSQEList_lock);
1924 if ((sq = rx_FreeSQEList)) {
1925 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1926 MUTEX_EXIT(&freeSQEList_lock);
1927 } else { /* otherwise allocate a new one and return that */
1928 MUTEX_EXIT(&freeSQEList_lock);
1929 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1930 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1931 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1933 MUTEX_ENTER(&sq->lock);
1935 if (cur_service != NULL) {
1936 cur_service->nRequestsRunning--;
1937 MUTEX_ENTER(&rx_quota_mutex);
1938 if (cur_service->nRequestsRunning < cur_service->minProcs)
1941 MUTEX_EXIT(&rx_quota_mutex);
1943 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1944 struct rx_call *tcall, *ncall;
1945 /* Scan for eligible incoming calls. A call is not eligible
1946 * if the maximum number of calls for its service type are
1947 * already executing */
1948 /* One thread will process calls FCFS (to prevent starvation),
1949 * while the other threads may run ahead looking for calls which
1950 * have all their input data available immediately. This helps
1951 * keep threads from blocking, waiting for data from the client. */
1952 choice2 = (struct rx_call *)0;
1953 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1954 service = tcall->conn->service;
1955 if (QuotaOK(service)) {
1956 MUTEX_ENTER(&rx_pthread_mutex);
1957 if (tno == rxi_fcfs_thread_num
1958 || !tcall->queue_item_header.next) {
1959 MUTEX_EXIT(&rx_pthread_mutex);
1960 /* If we're the fcfs thread, then we'll just use
1961 * this call. If we haven't been able to find an optimal
1962 * choice, and we're at the end of the list, then use a
1963 * 2d choice if one has been identified. Otherwise... */
1964 call = (choice2 ? choice2 : tcall);
1965 service = call->conn->service;
1967 MUTEX_EXIT(&rx_pthread_mutex);
1968 if (!queue_IsEmpty(&tcall->rq)) {
1969 struct rx_packet *rp;
1970 rp = queue_First(&tcall->rq, rx_packet);
1971 if (rp->header.seq == 1
1973 || (rp->header.flags & RX_LAST_PACKET))) {
1975 } else if (rxi_2dchoice && !choice2
1976 && !(tcall->flags & RX_CALL_CLEARED)
1977 && (tcall->rprev > rxi_HardAckRate)) {
1991 /* we can't schedule a call if there's no data!!! */
1992 /* send an ack if there's no data, if we're missing the
1993 * first packet, or we're missing something between first
1994 * and last -- there's a "hole" in the incoming data. */
1995 if (queue_IsEmpty(&call->rq)
1996 || queue_First(&call->rq, rx_packet)->header.seq != 1
1997 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1998 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2000 call->flags &= (~RX_CALL_WAIT_PROC);
2001 service->nRequestsRunning++;
2002 /* just started call in minProcs pool, need fewer to maintain
2004 MUTEX_ENTER(&rx_quota_mutex);
2005 if (service->nRequestsRunning <= service->minProcs)
2008 MUTEX_EXIT(&rx_quota_mutex);
2009 rx_atomic_dec(&rx_nWaiting);
2010 /* MUTEX_EXIT(&call->lock); */
2012 /* If there are no eligible incoming calls, add this process
2013 * to the idle server queue, to wait for one */
2016 *socketp = OSI_NULLSOCKET;
2018 sq->socketp = socketp;
2019 queue_Append(&rx_idleServerQueue, sq);
2023 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2025 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2026 return (struct rx_call *)0;
2029 } while (!(call = sq->newcall)
2030 && !(socketp && *socketp != OSI_NULLSOCKET));
2032 MUTEX_EXIT(&sq->lock);
2034 MUTEX_ENTER(&freeSQEList_lock);
2035 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2036 rx_FreeSQEList = sq;
2037 MUTEX_EXIT(&freeSQEList_lock);
2040 clock_GetTime(&call->startTime);
2041 call->state = RX_STATE_ACTIVE;
2042 call->mode = RX_MODE_RECEIVING;
2043 #ifdef RX_KERNEL_TRACE
2044 if (ICL_SETACTIVE(afs_iclSetp)) {
2045 int glockOwner = ISAFS_GLOCK();
2048 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2049 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2056 rxi_calltrace(RX_CALL_START, call);
2057 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2058 call->conn->service->servicePort, call->conn->service->serviceId,
2061 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2068 #endif /* RX_ENABLE_LOCKS */
2072 /* Establish a procedure to be called when a packet arrives for a
2073 * call. This routine will be called at most once after each call,
2074 * and will also be called if there is an error condition on the or
2075 * the call is complete. Used by multi rx to build a selection
2076 * function which determines which of several calls is likely to be a
2077 * good one to read from.
2078 * NOTE: the way this is currently implemented it is probably only a
2079 * good idea to (1) use it immediately after a newcall (clients only)
2080 * and (2) only use it once. Other uses currently void your warranty
2083 rx_SetArrivalProc(struct rx_call *call,
2084 void (*proc) (struct rx_call * call,
2087 void * handle, int arg)
2089 call->arrivalProc = proc;
2090 call->arrivalProcHandle = handle;
2091 call->arrivalProcArg = arg;
2094 /* Call is finished (possibly prematurely). Return rc to the peer, if
2095 * appropriate, and return the final error code from the conversation
2099 rx_EndCall(struct rx_call *call, afs_int32 rc)
2101 struct rx_connection *conn = call->conn;
2105 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2106 call, rc, call->error, call->abortCode));
2109 MUTEX_ENTER(&call->lock);
2111 if (rc == 0 && call->error == 0) {
2112 call->abortCode = 0;
2113 call->abortCount = 0;
2116 call->arrivalProc = (void (*)())0;
2117 if (rc && call->error == 0) {
2118 rxi_CallError(call, rc);
2119 call->mode = RX_MODE_ERROR;
2120 /* Send an abort message to the peer if this error code has
2121 * only just been set. If it was set previously, assume the
2122 * peer has already been sent the error code or will request it
2124 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2126 if (conn->type == RX_SERVER_CONNECTION) {
2127 /* Make sure reply or at least dummy reply is sent */
2128 if (call->mode == RX_MODE_RECEIVING) {
2129 MUTEX_EXIT(&call->lock);
2130 rxi_WriteProc(call, 0, 0);
2131 MUTEX_ENTER(&call->lock);
2133 if (call->mode == RX_MODE_SENDING) {
2134 MUTEX_EXIT(&call->lock);
2135 rxi_FlushWrite(call);
2136 MUTEX_ENTER(&call->lock);
2138 rxi_calltrace(RX_CALL_END, call);
2139 /* Call goes to hold state until reply packets are acknowledged */
2140 if (call->tfirst + call->nSoftAcked < call->tnext) {
2141 call->state = RX_STATE_HOLD;
2143 call->state = RX_STATE_DALLY;
2144 rxi_ClearTransmitQueue(call, 0);
2145 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
2146 rxevent_Cancel(call->keepAliveEvent, call,
2147 RX_CALL_REFCOUNT_ALIVE);
2149 } else { /* Client connection */
2151 /* Make sure server receives input packets, in the case where
2152 * no reply arguments are expected */
2153 if ((call->mode == RX_MODE_SENDING)
2154 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2155 MUTEX_EXIT(&call->lock);
2156 (void)rxi_ReadProc(call, &dummy, 1);
2157 MUTEX_ENTER(&call->lock);
2160 /* If we had an outstanding delayed ack, be nice to the server
2161 * and force-send it now.
2163 if (call->delayedAckEvent) {
2164 rxevent_Cancel(call->delayedAckEvent, call,
2165 RX_CALL_REFCOUNT_DELAY);
2166 call->delayedAckEvent = NULL;
2167 rxi_SendDelayedAck(NULL, call, NULL);
2170 /* We need to release the call lock since it's lower than the
2171 * conn_call_lock and we don't want to hold the conn_call_lock
2172 * over the rx_ReadProc call. The conn_call_lock needs to be held
2173 * here for the case where rx_NewCall is perusing the calls on
2174 * the connection structure. We don't want to signal until
2175 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2176 * have checked this call, found it active and by the time it
2177 * goes to sleep, will have missed the signal.
2179 MUTEX_EXIT(&call->lock);
2180 MUTEX_ENTER(&conn->conn_call_lock);
2181 MUTEX_ENTER(&call->lock);
2182 MUTEX_ENTER(&conn->conn_data_lock);
2183 conn->flags |= RX_CONN_BUSY;
2184 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2185 MUTEX_EXIT(&conn->conn_data_lock);
2186 #ifdef RX_ENABLE_LOCKS
2187 CV_BROADCAST(&conn->conn_call_cv);
2192 #ifdef RX_ENABLE_LOCKS
2194 MUTEX_EXIT(&conn->conn_data_lock);
2196 #endif /* RX_ENABLE_LOCKS */
2197 call->state = RX_STATE_DALLY;
2199 error = call->error;
2201 /* currentPacket, nLeft, and NFree must be zeroed here, because
2202 * ResetCall cannot: ResetCall may be called at splnet(), in the
2203 * kernel version, and may interrupt the macros rx_Read or
2204 * rx_Write, which run at normal priority for efficiency. */
2205 if (call->currentPacket) {
2206 #ifdef RX_TRACK_PACKETS
2207 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2209 rxi_FreePacket(call->currentPacket);
2210 call->currentPacket = (struct rx_packet *)0;
2213 call->nLeft = call->nFree = call->curlen = 0;
2215 /* Free any packets from the last call to ReadvProc/WritevProc */
2216 #ifdef RXDEBUG_PACKET
2218 #endif /* RXDEBUG_PACKET */
2219 rxi_FreePackets(0, &call->iovq);
2220 MUTEX_EXIT(&call->lock);
2222 MUTEX_ENTER(&rx_refcnt_mutex);
2223 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2224 MUTEX_EXIT(&rx_refcnt_mutex);
2225 if (conn->type == RX_CLIENT_CONNECTION) {
2226 MUTEX_ENTER(&conn->conn_data_lock);
2227 conn->flags &= ~RX_CONN_BUSY;
2228 MUTEX_EXIT(&conn->conn_data_lock);
2229 MUTEX_EXIT(&conn->conn_call_lock);
2233 * Map errors to the local host's errno.h format.
2235 error = ntoh_syserr_conv(error);
2239 #if !defined(KERNEL)
2241 /* Call this routine when shutting down a server or client (especially
2242 * clients). This will allow Rx to gracefully garbage collect server
2243 * connections, and reduce the number of retries that a server might
2244 * make to a dead client.
2245 * This is not quite right, since some calls may still be ongoing and
2246 * we can't lock them to destroy them. */
2250 struct rx_connection **conn_ptr, **conn_end;
2254 if (rxinit_status == 1) {
2256 return; /* Already shutdown. */
2258 rxi_DeleteCachedConnections();
2259 if (rx_connHashTable) {
2260 MUTEX_ENTER(&rx_connHashTable_lock);
2261 for (conn_ptr = &rx_connHashTable[0], conn_end =
2262 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2264 struct rx_connection *conn, *next;
2265 for (conn = *conn_ptr; conn; conn = next) {
2267 if (conn->type == RX_CLIENT_CONNECTION) {
2268 MUTEX_ENTER(&rx_refcnt_mutex);
2270 MUTEX_EXIT(&rx_refcnt_mutex);
2271 #ifdef RX_ENABLE_LOCKS
2272 rxi_DestroyConnectionNoLock(conn);
2273 #else /* RX_ENABLE_LOCKS */
2274 rxi_DestroyConnection(conn);
2275 #endif /* RX_ENABLE_LOCKS */
2279 #ifdef RX_ENABLE_LOCKS
2280 while (rx_connCleanup_list) {
2281 struct rx_connection *conn;
2282 conn = rx_connCleanup_list;
2283 rx_connCleanup_list = rx_connCleanup_list->next;
2284 MUTEX_EXIT(&rx_connHashTable_lock);
2285 rxi_CleanupConnection(conn);
2286 MUTEX_ENTER(&rx_connHashTable_lock);
2288 MUTEX_EXIT(&rx_connHashTable_lock);
2289 #endif /* RX_ENABLE_LOCKS */
2294 afs_winsockCleanup();
2302 /* if we wakeup packet waiter too often, can get in loop with two
2303 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2305 rxi_PacketsUnWait(void)
2307 if (!rx_waitingForPackets) {
2311 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2312 return; /* still over quota */
2315 rx_waitingForPackets = 0;
2316 #ifdef RX_ENABLE_LOCKS
2317 CV_BROADCAST(&rx_waitingForPackets_cv);
2319 osi_rxWakeup(&rx_waitingForPackets);
2325 /* ------------------Internal interfaces------------------------- */
2327 /* Return this process's service structure for the
2328 * specified socket and service */
2330 rxi_FindService(osi_socket socket, u_short serviceId)
2332 struct rx_service **sp;
2333 for (sp = &rx_services[0]; *sp; sp++) {
2334 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2340 #ifdef RXDEBUG_PACKET
2341 #ifdef KDUMP_RX_LOCK
2342 static struct rx_call_rx_lock *rx_allCallsp = 0;
2344 static struct rx_call *rx_allCallsp = 0;
2346 #endif /* RXDEBUG_PACKET */
2348 /* Allocate a call structure, for the indicated channel of the
2349 * supplied connection. The mode and state of the call must be set by
2350 * the caller. Returns the call with mutex locked. */
2352 rxi_NewCall(struct rx_connection *conn, int channel)
2354 struct rx_call *call;
2355 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2356 struct rx_call *cp; /* Call pointer temp */
2357 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2358 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2360 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2362 /* Grab an existing call structure, or allocate a new one.
2363 * Existing call structures are assumed to have been left reset by
2365 MUTEX_ENTER(&rx_freeCallQueue_lock);
2367 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2369 * EXCEPT that the TQ might not yet be cleared out.
2370 * Skip over those with in-use TQs.
2373 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2374 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2380 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2381 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2382 call = queue_First(&rx_freeCallQueue, rx_call);
2383 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2385 if (rx_stats_active)
2386 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2387 MUTEX_EXIT(&rx_freeCallQueue_lock);
2388 MUTEX_ENTER(&call->lock);
2389 CLEAR_CALL_QUEUE_LOCK(call);
2390 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2391 /* Now, if TQ wasn't cleared earlier, do it now. */
2392 rxi_WaitforTQBusy(call);
2393 if (call->flags & RX_CALL_TQ_CLEARME) {
2394 rxi_ClearTransmitQueue(call, 1);
2395 /*queue_Init(&call->tq);*/
2397 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2398 /* Bind the call to its connection structure */
2400 rxi_ResetCall(call, 1);
2403 call = rxi_Alloc(sizeof(struct rx_call));
2404 #ifdef RXDEBUG_PACKET
2405 call->allNextp = rx_allCallsp;
2406 rx_allCallsp = call;
2408 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2409 #else /* RXDEBUG_PACKET */
2410 rx_atomic_inc(&rx_stats.nCallStructs);
2411 #endif /* RXDEBUG_PACKET */
2413 MUTEX_EXIT(&rx_freeCallQueue_lock);
2414 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2415 MUTEX_ENTER(&call->lock);
2416 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2417 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2418 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2420 /* Initialize once-only items */
2421 queue_Init(&call->tq);
2422 queue_Init(&call->rq);
2423 queue_Init(&call->iovq);
2424 #ifdef RXDEBUG_PACKET
2425 call->rqc = call->tqc = call->iovqc = 0;
2426 #endif /* RXDEBUG_PACKET */
2427 /* Bind the call to its connection structure (prereq for reset) */
2429 rxi_ResetCall(call, 1);
2431 call->channel = channel;
2432 call->callNumber = &conn->callNumber[channel];
2433 call->rwind = conn->rwind[channel];
2434 call->twind = conn->twind[channel];
2435 /* Note that the next expected call number is retained (in
2436 * conn->callNumber[i]), even if we reallocate the call structure
2438 conn->call[channel] = call;
2439 /* if the channel's never been used (== 0), we should start at 1, otherwise
2440 * the call number is valid from the last time this channel was used */
2441 if (*call->callNumber == 0)
2442 *call->callNumber = 1;
2447 /* A call has been inactive long enough that so we can throw away
2448 * state, including the call structure, which is placed on the call
2451 * call->lock amd rx_refcnt_mutex are held upon entry.
2452 * haveCTLock is set when called from rxi_ReapConnections.
2455 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2457 int channel = call->channel;
2458 struct rx_connection *conn = call->conn;
2461 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2462 (*call->callNumber)++;
2463 rxi_ResetCall(call, 0);
2464 call->conn->call[channel] = (struct rx_call *)0;
2465 MUTEX_EXIT(&rx_refcnt_mutex);
2467 MUTEX_ENTER(&rx_freeCallQueue_lock);
2468 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2469 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2470 /* A call may be free even though its transmit queue is still in use.
2471 * Since we search the call list from head to tail, put busy calls at
2472 * the head of the list, and idle calls at the tail.
2474 if (call->flags & RX_CALL_TQ_BUSY)
2475 queue_Prepend(&rx_freeCallQueue, call);
2477 queue_Append(&rx_freeCallQueue, call);
2478 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2479 queue_Append(&rx_freeCallQueue, call);
2480 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2481 if (rx_stats_active)
2482 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2483 MUTEX_EXIT(&rx_freeCallQueue_lock);
2485 /* Destroy the connection if it was previously slated for
2486 * destruction, i.e. the Rx client code previously called
2487 * rx_DestroyConnection (client connections), or
2488 * rxi_ReapConnections called the same routine (server
2489 * connections). Only do this, however, if there are no
2490 * outstanding calls. Note that for fine grain locking, there appears
2491 * to be a deadlock in that rxi_FreeCall has a call locked and
2492 * DestroyConnectionNoLock locks each call in the conn. But note a
2493 * few lines up where we have removed this call from the conn.
2494 * If someone else destroys a connection, they either have no
2495 * call lock held or are going through this section of code.
2497 MUTEX_ENTER(&conn->conn_data_lock);
2498 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2499 MUTEX_ENTER(&rx_refcnt_mutex);
2501 MUTEX_EXIT(&rx_refcnt_mutex);
2502 MUTEX_EXIT(&conn->conn_data_lock);
2503 #ifdef RX_ENABLE_LOCKS
2505 rxi_DestroyConnectionNoLock(conn);
2507 rxi_DestroyConnection(conn);
2508 #else /* RX_ENABLE_LOCKS */
2509 rxi_DestroyConnection(conn);
2510 #endif /* RX_ENABLE_LOCKS */
2512 MUTEX_EXIT(&conn->conn_data_lock);
2514 MUTEX_ENTER(&rx_refcnt_mutex);
2517 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2518 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2521 rxi_Alloc(size_t size)
2525 if (rx_stats_active) {
2526 rx_atomic_add(&rxi_Allocsize, (int) size);
2527 rx_atomic_inc(&rxi_Alloccnt);
2531 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2532 afs_osi_Alloc_NoSleep(size);
2537 osi_Panic("rxi_Alloc error");
2543 rxi_Free(void *addr, size_t size)
2545 if (rx_stats_active) {
2546 rx_atomic_sub(&rxi_Allocsize, (int) size);
2547 rx_atomic_dec(&rxi_Alloccnt);
2549 osi_Free(addr, size);
2553 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2555 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2556 struct rx_peer *next = NULL;
2560 MUTEX_ENTER(&rx_peerHashTable_lock);
2562 peer_ptr = &rx_peerHashTable[0];
2563 peer_end = &rx_peerHashTable[rx_hashTableSize];
2566 for ( ; peer_ptr < peer_end; peer_ptr++) {
2569 for ( ; peer; peer = next) {
2571 if (host == peer->host)
2576 hashIndex = PEER_HASH(host, port);
2577 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2578 if ((peer->host == host) && (peer->port == port))
2583 MUTEX_ENTER(&rx_peerHashTable_lock);
2588 MUTEX_EXIT(&rx_peerHashTable_lock);
2590 MUTEX_ENTER(&peer->peer_lock);
2591 /* We don't handle dropping below min, so don't */
2592 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2593 peer->ifMTU=MIN(mtu, peer->ifMTU);
2594 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2595 /* if we tweaked this down, need to tune our peer MTU too */
2596 peer->MTU = MIN(peer->MTU, peer->natMTU);
2597 /* if we discovered a sub-1500 mtu, degrade */
2598 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2599 peer->maxDgramPackets = 1;
2600 /* We no longer have valid peer packet information */
2601 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2602 peer->maxPacketSize = 0;
2603 MUTEX_EXIT(&peer->peer_lock);
2605 MUTEX_ENTER(&rx_peerHashTable_lock);
2607 if (host && !port) {
2609 /* pick up where we left off */
2613 MUTEX_EXIT(&rx_peerHashTable_lock);
2616 /* Find the peer process represented by the supplied (host,port)
2617 * combination. If there is no appropriate active peer structure, a
2618 * new one will be allocated and initialized
2619 * The origPeer, if set, is a pointer to a peer structure on which the
2620 * refcount will be be decremented. This is used to replace the peer
2621 * structure hanging off a connection structure */
2623 rxi_FindPeer(afs_uint32 host, u_short port,
2624 struct rx_peer *origPeer, int create)
2628 hashIndex = PEER_HASH(host, port);
2629 MUTEX_ENTER(&rx_peerHashTable_lock);
2630 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2631 if ((pp->host == host) && (pp->port == port))
2636 pp = rxi_AllocPeer(); /* This bzero's *pp */
2637 pp->host = host; /* set here or in InitPeerParams is zero */
2639 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2640 queue_Init(&pp->congestionQueue);
2641 queue_Init(&pp->rpcStats);
2642 pp->next = rx_peerHashTable[hashIndex];
2643 rx_peerHashTable[hashIndex] = pp;
2644 rxi_InitPeerParams(pp);
2645 if (rx_stats_active)
2646 rx_atomic_inc(&rx_stats.nPeerStructs);
2653 origPeer->refCount--;
2654 MUTEX_EXIT(&rx_peerHashTable_lock);
2659 /* Find the connection at (host, port) started at epoch, and with the
2660 * given connection id. Creates the server connection if necessary.
2661 * The type specifies whether a client connection or a server
2662 * connection is desired. In both cases, (host, port) specify the
2663 * peer's (host, pair) pair. Client connections are not made
2664 * automatically by this routine. The parameter socket gives the
2665 * socket descriptor on which the packet was received. This is used,
2666 * in the case of server connections, to check that *new* connections
2667 * come via a valid (port, serviceId). Finally, the securityIndex
2668 * parameter must match the existing index for the connection. If a
2669 * server connection is created, it will be created using the supplied
2670 * index, if the index is valid for this service */
2671 struct rx_connection *
2672 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2673 u_short port, u_short serviceId, afs_uint32 cid,
2674 afs_uint32 epoch, int type, u_int securityIndex)
2676 int hashindex, flag, i;
2677 struct rx_connection *conn;
2678 hashindex = CONN_HASH(host, port, cid, epoch, type);
2679 MUTEX_ENTER(&rx_connHashTable_lock);
2680 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2681 rx_connHashTable[hashindex],
2684 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2685 && (epoch == conn->epoch)) {
2686 struct rx_peer *pp = conn->peer;
2687 if (securityIndex != conn->securityIndex) {
2688 /* this isn't supposed to happen, but someone could forge a packet
2689 * like this, and there seems to be some CM bug that makes this
2690 * happen from time to time -- in which case, the fileserver
2692 MUTEX_EXIT(&rx_connHashTable_lock);
2693 return (struct rx_connection *)0;
2695 if (pp->host == host && pp->port == port)
2697 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2699 /* So what happens when it's a callback connection? */
2700 if ( /*type == RX_CLIENT_CONNECTION && */
2701 (conn->epoch & 0x80000000))
2705 /* the connection rxLastConn that was used the last time is not the
2706 ** one we are looking for now. Hence, start searching in the hash */
2708 conn = rx_connHashTable[hashindex];
2713 struct rx_service *service;
2714 if (type == RX_CLIENT_CONNECTION) {
2715 MUTEX_EXIT(&rx_connHashTable_lock);
2716 return (struct rx_connection *)0;
2718 service = rxi_FindService(socket, serviceId);
2719 if (!service || (securityIndex >= service->nSecurityObjects)
2720 || (service->securityObjects[securityIndex] == 0)) {
2721 MUTEX_EXIT(&rx_connHashTable_lock);
2722 return (struct rx_connection *)0;
2724 conn = rxi_AllocConnection(); /* This bzero's the connection */
2725 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2726 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2727 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2728 conn->next = rx_connHashTable[hashindex];
2729 rx_connHashTable[hashindex] = conn;
2730 conn->peer = rxi_FindPeer(host, port, 0, 1);
2731 conn->type = RX_SERVER_CONNECTION;
2732 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2733 conn->epoch = epoch;
2734 conn->cid = cid & RX_CIDMASK;
2735 /* conn->serial = conn->lastSerial = 0; */
2736 /* conn->timeout = 0; */
2737 conn->ackRate = RX_FAST_ACK_RATE;
2738 conn->service = service;
2739 conn->serviceId = serviceId;
2740 conn->securityIndex = securityIndex;
2741 conn->securityObject = service->securityObjects[securityIndex];
2742 conn->nSpecific = 0;
2743 conn->specific = NULL;
2744 rx_SetConnDeadTime(conn, service->connDeadTime);
2745 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2746 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2747 for (i = 0; i < RX_MAXCALLS; i++) {
2748 conn->twind[i] = rx_initSendWindow;
2749 conn->rwind[i] = rx_initReceiveWindow;
2751 /* Notify security object of the new connection */
2752 RXS_NewConnection(conn->securityObject, conn);
2753 /* XXXX Connection timeout? */
2754 if (service->newConnProc)
2755 (*service->newConnProc) (conn);
2756 if (rx_stats_active)
2757 rx_atomic_inc(&rx_stats.nServerConns);
2760 MUTEX_ENTER(&rx_refcnt_mutex);
2762 MUTEX_EXIT(&rx_refcnt_mutex);
2764 rxLastConn = conn; /* store this connection as the last conn used */
2765 MUTEX_EXIT(&rx_connHashTable_lock);
2769 /* There are two packet tracing routines available for testing and monitoring
2770 * Rx. One is called just after every packet is received and the other is
2771 * called just before every packet is sent. Received packets, have had their
2772 * headers decoded, and packets to be sent have not yet had their headers
2773 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2774 * containing the network address. Both can be modified. The return value, if
2775 * non-zero, indicates that the packet should be dropped. */
2777 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2778 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2780 /* A packet has been received off the interface. Np is the packet, socket is
2781 * the socket number it was received from (useful in determining which service
2782 * this packet corresponds to), and (host, port) reflect the host,port of the
2783 * sender. This call returns the packet to the caller if it is finished with
2784 * it, rather than de-allocating it, just as a small performance hack */
2787 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
2788 afs_uint32 host, u_short port, int *tnop,
2789 struct rx_call **newcallp)
2791 struct rx_call *call;
2792 struct rx_connection *conn;
2794 afs_uint32 currentCallNumber;
2800 struct rx_packet *tnp;
2803 /* We don't print out the packet until now because (1) the time may not be
2804 * accurate enough until now in the lwp implementation (rx_Listener only gets
2805 * the time after the packet is read) and (2) from a protocol point of view,
2806 * this is the first time the packet has been seen */
2807 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2808 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2809 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
2810 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2811 np->header.epoch, np->header.cid, np->header.callNumber,
2812 np->header.seq, np->header.flags, np));
2815 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2816 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2819 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2820 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2823 /* If an input tracer function is defined, call it with the packet and
2824 * network address. Note this function may modify its arguments. */
2825 if (rx_justReceived) {
2826 struct sockaddr_in addr;
2828 addr.sin_family = AF_INET;
2829 addr.sin_port = port;
2830 addr.sin_addr.s_addr = host;
2831 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2832 addr.sin_len = sizeof(addr);
2833 #endif /* AFS_OSF_ENV */
2834 drop = (*rx_justReceived) (np, &addr);
2835 /* drop packet if return value is non-zero */
2838 port = addr.sin_port; /* in case fcn changed addr */
2839 host = addr.sin_addr.s_addr;
2843 /* If packet was not sent by the client, then *we* must be the client */
2844 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2845 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2847 /* Find the connection (or fabricate one, if we're the server & if
2848 * necessary) associated with this packet */
2850 rxi_FindConnection(socket, host, port, np->header.serviceId,
2851 np->header.cid, np->header.epoch, type,
2852 np->header.securityIndex);
2855 /* If no connection found or fabricated, just ignore the packet.
2856 * (An argument could be made for sending an abort packet for
2861 MUTEX_ENTER(&conn->conn_data_lock);
2862 if (conn->maxSerial < np->header.serial)
2863 conn->maxSerial = np->header.serial;
2864 MUTEX_EXIT(&conn->conn_data_lock);
2866 /* If the connection is in an error state, send an abort packet and ignore
2867 * the incoming packet */
2869 /* Don't respond to an abort packet--we don't want loops! */
2870 MUTEX_ENTER(&conn->conn_data_lock);
2871 if (np->header.type != RX_PACKET_TYPE_ABORT)
2872 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2873 MUTEX_ENTER(&rx_refcnt_mutex);
2875 MUTEX_EXIT(&rx_refcnt_mutex);
2876 MUTEX_EXIT(&conn->conn_data_lock);
2880 /* Check for connection-only requests (i.e. not call specific). */
2881 if (np->header.callNumber == 0) {
2882 switch (np->header.type) {
2883 case RX_PACKET_TYPE_ABORT: {
2884 /* What if the supplied error is zero? */
2885 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2886 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
2887 rxi_ConnectionError(conn, errcode);
2888 MUTEX_ENTER(&rx_refcnt_mutex);
2890 MUTEX_EXIT(&rx_refcnt_mutex);
2893 case RX_PACKET_TYPE_CHALLENGE:
2894 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2895 MUTEX_ENTER(&rx_refcnt_mutex);
2897 MUTEX_EXIT(&rx_refcnt_mutex);
2899 case RX_PACKET_TYPE_RESPONSE:
2900 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2901 MUTEX_ENTER(&rx_refcnt_mutex);
2903 MUTEX_EXIT(&rx_refcnt_mutex);
2905 case RX_PACKET_TYPE_PARAMS:
2906 case RX_PACKET_TYPE_PARAMS + 1:
2907 case RX_PACKET_TYPE_PARAMS + 2:
2908 /* ignore these packet types for now */
2909 MUTEX_ENTER(&rx_refcnt_mutex);
2911 MUTEX_EXIT(&rx_refcnt_mutex);
2916 /* Should not reach here, unless the peer is broken: send an
2918 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2919 MUTEX_ENTER(&conn->conn_data_lock);
2920 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2921 MUTEX_ENTER(&rx_refcnt_mutex);
2923 MUTEX_EXIT(&rx_refcnt_mutex);
2924 MUTEX_EXIT(&conn->conn_data_lock);
2929 channel = np->header.cid & RX_CHANNELMASK;
2930 call = conn->call[channel];
2931 #ifdef RX_ENABLE_LOCKS
2933 MUTEX_ENTER(&call->lock);
2934 /* Test to see if call struct is still attached to conn. */
2935 if (call != conn->call[channel]) {
2937 MUTEX_EXIT(&call->lock);
2938 if (type == RX_SERVER_CONNECTION) {
2939 call = conn->call[channel];
2940 /* If we started with no call attached and there is one now,
2941 * another thread is also running this routine and has gotten
2942 * the connection channel. We should drop this packet in the tests
2943 * below. If there was a call on this connection and it's now
2944 * gone, then we'll be making a new call below.
2945 * If there was previously a call and it's now different then
2946 * the old call was freed and another thread running this routine
2947 * has created a call on this channel. One of these two threads
2948 * has a packet for the old call and the code below handles those
2952 MUTEX_ENTER(&call->lock);
2954 /* This packet can't be for this call. If the new call address is
2955 * 0 then no call is running on this channel. If there is a call
2956 * then, since this is a client connection we're getting data for
2957 * it must be for the previous call.
2959 if (rx_stats_active)
2960 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
2961 MUTEX_ENTER(&rx_refcnt_mutex);
2963 MUTEX_EXIT(&rx_refcnt_mutex);
2968 currentCallNumber = conn->callNumber[channel];
2970 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2971 if (np->header.callNumber < currentCallNumber) {
2972 if (rx_stats_active)
2973 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
2974 #ifdef RX_ENABLE_LOCKS
2976 MUTEX_EXIT(&call->lock);
2978 MUTEX_ENTER(&rx_refcnt_mutex);
2980 MUTEX_EXIT(&rx_refcnt_mutex);
2984 MUTEX_ENTER(&conn->conn_call_lock);
2985 call = rxi_NewCall(conn, channel);
2986 MUTEX_EXIT(&conn->conn_call_lock);
2987 *call->callNumber = np->header.callNumber;
2989 if (np->header.callNumber == 0)
2990 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",
2991 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
2992 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
2993 np->header.flags, np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2995 call->state = RX_STATE_PRECALL;
2996 clock_GetTime(&call->queueTime);
2997 hzero(call->bytesSent);
2998 hzero(call->bytesRcvd);
3000 * If the number of queued calls exceeds the overload
3001 * threshold then abort this call.
3003 if ((rx_BusyThreshold > 0) &&
3004 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3005 struct rx_packet *tp;
3007 rxi_CallError(call, rx_BusyError);
3008 tp = rxi_SendCallAbort(call, np, 1, 0);
3009 MUTEX_EXIT(&call->lock);
3010 MUTEX_ENTER(&rx_refcnt_mutex);
3012 MUTEX_EXIT(&rx_refcnt_mutex);
3013 if (rx_stats_active)
3014 rx_atomic_inc(&rx_stats.nBusies);
3017 rxi_KeepAliveOn(call);
3018 } else if (np->header.callNumber != currentCallNumber) {
3019 /* Wait until the transmit queue is idle before deciding
3020 * whether to reset the current call. Chances are that the
3021 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3024 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3025 if (call->state == RX_STATE_ACTIVE) {
3026 rxi_WaitforTQBusy(call);
3028 * If we entered error state while waiting,
3029 * must call rxi_CallError to permit rxi_ResetCall
3030 * to processed when the tqWaiter count hits zero.
3033 rxi_CallError(call, call->error);
3034 MUTEX_EXIT(&call->lock);
3035 MUTEX_ENTER(&rx_refcnt_mutex);
3037 MUTEX_EXIT(&rx_refcnt_mutex);
3041 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3042 /* If the new call cannot be taken right now send a busy and set
3043 * the error condition in this call, so that it terminates as
3044 * quickly as possible */
3045 if (call->state == RX_STATE_ACTIVE) {
3046 struct rx_packet *tp;
3048 rxi_CallError(call, RX_CALL_DEAD);
3049 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3051 MUTEX_EXIT(&call->lock);
3052 MUTEX_ENTER(&rx_refcnt_mutex);
3054 MUTEX_EXIT(&rx_refcnt_mutex);
3057 rxi_ResetCall(call, 0);
3058 *call->callNumber = np->header.callNumber;
3060 if (np->header.callNumber == 0)
3061 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",
3062 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3063 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3064 np->header.flags, np, np->retryTime.sec, np->retryTime.usec, np->length));
3066 call->state = RX_STATE_PRECALL;
3067 clock_GetTime(&call->queueTime);
3068 hzero(call->bytesSent);
3069 hzero(call->bytesRcvd);
3071 * If the number of queued calls exceeds the overload
3072 * threshold then abort this call.
3074 if ((rx_BusyThreshold > 0) &&
3075 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3076 struct rx_packet *tp;
3078 rxi_CallError(call, rx_BusyError);
3079 tp = rxi_SendCallAbort(call, np, 1, 0);
3080 MUTEX_EXIT(&call->lock);
3081 MUTEX_ENTER(&rx_refcnt_mutex);
3083 MUTEX_EXIT(&rx_refcnt_mutex);
3084 if (rx_stats_active)
3085 rx_atomic_inc(&rx_stats.nBusies);
3088 rxi_KeepAliveOn(call);
3090 /* Continuing call; do nothing here. */
3092 } else { /* we're the client */
3093 /* Ignore all incoming acknowledgements for calls in DALLY state */
3094 if (call && (call->state == RX_STATE_DALLY)
3095 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3096 if (rx_stats_active)
3097 rx_atomic_inc(&rx_stats.ignorePacketDally);
3098 #ifdef RX_ENABLE_LOCKS
3100 MUTEX_EXIT(&call->lock);
3103 MUTEX_ENTER(&rx_refcnt_mutex);
3105 MUTEX_EXIT(&rx_refcnt_mutex);
3109 /* Ignore anything that's not relevant to the current call. If there
3110 * isn't a current call, then no packet is relevant. */
3111 if (!call || (np->header.callNumber != currentCallNumber)) {
3112 if (rx_stats_active)
3113 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3114 #ifdef RX_ENABLE_LOCKS
3116 MUTEX_EXIT(&call->lock);
3119 MUTEX_ENTER(&rx_refcnt_mutex);
3121 MUTEX_EXIT(&rx_refcnt_mutex);
3124 /* If the service security object index stamped in the packet does not
3125 * match the connection's security index, ignore the packet */
3126 if (np->header.securityIndex != conn->securityIndex) {
3127 #ifdef RX_ENABLE_LOCKS
3128 MUTEX_EXIT(&call->lock);
3130 MUTEX_ENTER(&rx_refcnt_mutex);
3132 MUTEX_EXIT(&rx_refcnt_mutex);
3136 /* If we're receiving the response, then all transmit packets are
3137 * implicitly acknowledged. Get rid of them. */
3138 if (np->header.type == RX_PACKET_TYPE_DATA) {
3139 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3140 /* XXX Hack. Because we must release the global rx lock when
3141 * sending packets (osi_NetSend) we drop all acks while we're
3142 * traversing the tq in rxi_Start sending packets out because
3143 * packets may move to the freePacketQueue as result of being here!
3144 * So we drop these packets until we're safely out of the
3145 * traversing. Really ugly!
3146 * For fine grain RX locking, we set the acked field in the
3147 * packets and let rxi_Start remove them from the transmit queue.
3149 if (call->flags & RX_CALL_TQ_BUSY) {
3150 #ifdef RX_ENABLE_LOCKS
3151 rxi_SetAcksInTransmitQueue(call);
3153 MUTEX_ENTER(&rx_refcnt_mutex);
3155 MUTEX_EXIT(&rx_refcnt_mutex);
3156 return np; /* xmitting; drop packet */
3159 rxi_ClearTransmitQueue(call, 0);
3161 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3162 rxi_ClearTransmitQueue(call, 0);
3163 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3165 if (np->header.type == RX_PACKET_TYPE_ACK) {
3166 /* now check to see if this is an ack packet acknowledging that the
3167 * server actually *lost* some hard-acked data. If this happens we
3168 * ignore this packet, as it may indicate that the server restarted in
3169 * the middle of a call. It is also possible that this is an old ack
3170 * packet. We don't abort the connection in this case, because this
3171 * *might* just be an old ack packet. The right way to detect a server
3172 * restart in the midst of a call is to notice that the server epoch
3174 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3175 * XXX unacknowledged. I think that this is off-by-one, but
3176 * XXX I don't dare change it just yet, since it will
3177 * XXX interact badly with the server-restart detection
3178 * XXX code in receiveackpacket. */
3179 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3180 if (rx_stats_active)
3181 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3182 MUTEX_EXIT(&call->lock);
3183 MUTEX_ENTER(&rx_refcnt_mutex);
3185 MUTEX_EXIT(&rx_refcnt_mutex);
3189 } /* else not a data packet */
3192 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3193 /* Set remote user defined status from packet */
3194 call->remoteStatus = np->header.userStatus;
3196 /* Note the gap between the expected next packet and the actual
3197 * packet that arrived, when the new packet has a smaller serial number
3198 * than expected. Rioses frequently reorder packets all by themselves,
3199 * so this will be quite important with very large window sizes.
3200 * Skew is checked against 0 here to avoid any dependence on the type of
3201 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3203 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3204 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3205 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3207 MUTEX_ENTER(&conn->conn_data_lock);
3208 skew = conn->lastSerial - np->header.serial;
3209 conn->lastSerial = np->header.serial;
3210 MUTEX_EXIT(&conn->conn_data_lock);
3212 struct rx_peer *peer;
3214 if (skew > peer->inPacketSkew) {
3215 dpf(("*** In skew changed from %d to %d\n",
3216 peer->inPacketSkew, skew));
3217 peer->inPacketSkew = skew;
3221 /* Now do packet type-specific processing */
3222 switch (np->header.type) {
3223 case RX_PACKET_TYPE_DATA:
3224 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3227 case RX_PACKET_TYPE_ACK:
3228 /* Respond immediately to ack packets requesting acknowledgement
3230 if (np->header.flags & RX_REQUEST_ACK) {
3232 (void)rxi_SendCallAbort(call, 0, 1, 0);
3234 (void)rxi_SendAck(call, 0, np->header.serial,
3235 RX_ACK_PING_RESPONSE, 1);
3237 np = rxi_ReceiveAckPacket(call, np, 1);
3239 case RX_PACKET_TYPE_ABORT: {
3240 /* An abort packet: reset the call, passing the error up to the user. */
3241 /* What if error is zero? */
3242 /* What if the error is -1? the application will treat it as a timeout. */
3243 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3244 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3245 rxi_CallError(call, errdata);
3246 MUTEX_EXIT(&call->lock);
3247 MUTEX_ENTER(&rx_refcnt_mutex);
3249 MUTEX_EXIT(&rx_refcnt_mutex);
3250 return np; /* xmitting; drop packet */
3252 case RX_PACKET_TYPE_BUSY:
3255 case RX_PACKET_TYPE_ACKALL:
3256 /* All packets acknowledged, so we can drop all packets previously
3257 * readied for sending */
3258 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3259 /* XXX Hack. We because we can't release the global rx lock when
3260 * sending packets (osi_NetSend) we drop all ack pkts while we're
3261 * traversing the tq in rxi_Start sending packets out because
3262 * packets may move to the freePacketQueue as result of being
3263 * here! So we drop these packets until we're safely out of the
3264 * traversing. Really ugly!
3265 * For fine grain RX locking, we set the acked field in the packets
3266 * and let rxi_Start remove the packets from the transmit queue.
3268 if (call->flags & RX_CALL_TQ_BUSY) {
3269 #ifdef RX_ENABLE_LOCKS
3270 rxi_SetAcksInTransmitQueue(call);
3272 #else /* RX_ENABLE_LOCKS */
3273 MUTEX_EXIT(&call->lock);
3274 MUTEX_ENTER(&rx_refcnt_mutex);
3276 MUTEX_EXIT(&rx_refcnt_mutex);
3277 return np; /* xmitting; drop packet */
3278 #endif /* RX_ENABLE_LOCKS */
3280 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3281 rxi_ClearTransmitQueue(call, 0);
3282 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3285 /* Should not reach here, unless the peer is broken: send an abort
3287 rxi_CallError(call, RX_PROTOCOL_ERROR);
3288 np = rxi_SendCallAbort(call, np, 1, 0);
3291 /* Note when this last legitimate packet was received, for keep-alive
3292 * processing. Note, we delay getting the time until now in the hope that
3293 * the packet will be delivered to the user before any get time is required
3294 * (if not, then the time won't actually be re-evaluated here). */
3295 call->lastReceiveTime = clock_Sec();
3296 MUTEX_EXIT(&call->lock);
3297 MUTEX_ENTER(&rx_refcnt_mutex);
3299 MUTEX_EXIT(&rx_refcnt_mutex);
3303 /* return true if this is an "interesting" connection from the point of view
3304 of someone trying to debug the system */
3306 rxi_IsConnInteresting(struct rx_connection *aconn)
3309 struct rx_call *tcall;
3311 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3314 for (i = 0; i < RX_MAXCALLS; i++) {
3315 tcall = aconn->call[i];
3317 if ((tcall->state == RX_STATE_PRECALL)
3318 || (tcall->state == RX_STATE_ACTIVE))
3320 if ((tcall->mode == RX_MODE_SENDING)
3321 || (tcall->mode == RX_MODE_RECEIVING))
3329 /* if this is one of the last few packets AND it wouldn't be used by the
3330 receiving call to immediately satisfy a read request, then drop it on
3331 the floor, since accepting it might prevent a lock-holding thread from
3332 making progress in its reading. If a call has been cleared while in
3333 the precall state then ignore all subsequent packets until the call
3334 is assigned to a thread. */
3337 TooLow(struct rx_packet *ap, struct rx_call *acall)
3341 MUTEX_ENTER(&rx_quota_mutex);
3342 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3343 && (acall->state == RX_STATE_PRECALL))
3344 || ((rx_nFreePackets < rxi_dataQuota + 2)
3345 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3346 && (acall->flags & RX_CALL_READER_WAIT)))) {
3349 MUTEX_EXIT(&rx_quota_mutex);
3355 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3357 struct rx_connection *conn = arg1;
3358 struct rx_call *acall = arg2;
3359 struct rx_call *call = acall;
3360 struct clock when, now;
3363 MUTEX_ENTER(&conn->conn_data_lock);
3364 conn->checkReachEvent = NULL;
3365 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3367 MUTEX_ENTER(&rx_refcnt_mutex);
3369 MUTEX_EXIT(&rx_refcnt_mutex);
3371 MUTEX_EXIT(&conn->conn_data_lock);
3375 MUTEX_ENTER(&conn->conn_call_lock);
3376 MUTEX_ENTER(&conn->conn_data_lock);
3377 for (i = 0; i < RX_MAXCALLS; i++) {
3378 struct rx_call *tc = conn->call[i];
3379 if (tc && tc->state == RX_STATE_PRECALL) {
3385 /* Indicate that rxi_CheckReachEvent is no longer running by
3386 * clearing the flag. Must be atomic under conn_data_lock to
3387 * avoid a new call slipping by: rxi_CheckConnReach holds
3388 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3390 conn->flags &= ~RX_CONN_ATTACHWAIT;
3391 MUTEX_EXIT(&conn->conn_data_lock);
3392 MUTEX_EXIT(&conn->conn_call_lock);
3397 MUTEX_ENTER(&call->lock);
3398 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3400 MUTEX_EXIT(&call->lock);
3402 clock_GetTime(&now);
3404 when.sec += RX_CHECKREACH_TIMEOUT;
3405 MUTEX_ENTER(&conn->conn_data_lock);
3406 if (!conn->checkReachEvent) {
3407 MUTEX_ENTER(&rx_refcnt_mutex);
3409 MUTEX_EXIT(&rx_refcnt_mutex);
3410 conn->checkReachEvent =
3411 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3414 MUTEX_EXIT(&conn->conn_data_lock);
3420 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3422 struct rx_service *service = conn->service;
3423 struct rx_peer *peer = conn->peer;
3424 afs_uint32 now, lastReach;
3426 if (service->checkReach == 0)
3430 MUTEX_ENTER(&peer->peer_lock);
3431 lastReach = peer->lastReachTime;
3432 MUTEX_EXIT(&peer->peer_lock);
3433 if (now - lastReach < RX_CHECKREACH_TTL)
3436 MUTEX_ENTER(&conn->conn_data_lock);
3437 if (conn->flags & RX_CONN_ATTACHWAIT) {
3438 MUTEX_EXIT(&conn->conn_data_lock);
3441 conn->flags |= RX_CONN_ATTACHWAIT;
3442 MUTEX_EXIT(&conn->conn_data_lock);
3443 if (!conn->checkReachEvent)
3444 rxi_CheckReachEvent(NULL, conn, call);
3449 /* try to attach call, if authentication is complete */
3451 TryAttach(struct rx_call *acall, osi_socket socket,
3452 int *tnop, struct rx_call **newcallp,
3455 struct rx_connection *conn = acall->conn;
3457 if (conn->type == RX_SERVER_CONNECTION
3458 && acall->state == RX_STATE_PRECALL) {
3459 /* Don't attach until we have any req'd. authentication. */
3460 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3461 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3462 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3463 /* Note: this does not necessarily succeed; there
3464 * may not any proc available
3467 rxi_ChallengeOn(acall->conn);
3472 /* A data packet has been received off the interface. This packet is
3473 * appropriate to the call (the call is in the right state, etc.). This
3474 * routine can return a packet to the caller, for re-use */
3477 rxi_ReceiveDataPacket(struct rx_call *call,
3478 struct rx_packet *np, int istack,
3479 osi_socket socket, afs_uint32 host, u_short port,
3480 int *tnop, struct rx_call **newcallp)
3482 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3487 afs_uint32 serial=0, flags=0;
3489 struct rx_packet *tnp;
3490 struct clock when, now;
3491 if (rx_stats_active)
3492 rx_atomic_inc(&rx_stats.dataPacketsRead);
3495 /* If there are no packet buffers, drop this new packet, unless we can find
3496 * packet buffers from inactive calls */
3498 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3499 MUTEX_ENTER(&rx_freePktQ_lock);
3500 rxi_NeedMorePackets = TRUE;
3501 MUTEX_EXIT(&rx_freePktQ_lock);
3502 if (rx_stats_active)
3503 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3504 call->rprev = np->header.serial;
3505 rxi_calltrace(RX_TRACE_DROP, call);
3506 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3508 rxi_ClearReceiveQueue(call);
3509 clock_GetTime(&now);
3511 clock_Add(&when, &rx_softAckDelay);
3512 if (!call->delayedAckEvent
3513 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3514 rxevent_Cancel(call->delayedAckEvent, call,
3515 RX_CALL_REFCOUNT_DELAY);
3516 MUTEX_ENTER(&rx_refcnt_mutex);
3517 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3518 MUTEX_EXIT(&rx_refcnt_mutex);
3520 call->delayedAckEvent =
3521 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3523 /* we've damaged this call already, might as well do it in. */
3529 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3530 * packet is one of several packets transmitted as a single
3531 * datagram. Do not send any soft or hard acks until all packets
3532 * in a jumbogram have been processed. Send negative acks right away.
3534 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3535 /* tnp is non-null when there are more packets in the
3536 * current jumbo gram */
3543 seq = np->header.seq;
3544 serial = np->header.serial;
3545 flags = np->header.flags;
3547 /* If the call is in an error state, send an abort message */
3549 return rxi_SendCallAbort(call, np, istack, 0);
3551 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3552 * AFS 3.5 jumbogram. */
3553 if (flags & RX_JUMBO_PACKET) {
3554 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3559 if (np->header.spare != 0) {
3560 MUTEX_ENTER(&call->conn->conn_data_lock);
3561 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3562 MUTEX_EXIT(&call->conn->conn_data_lock);
3565 /* The usual case is that this is the expected next packet */
3566 if (seq == call->rnext) {
3568 /* Check to make sure it is not a duplicate of one already queued */
3569 if (queue_IsNotEmpty(&call->rq)
3570 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3571 if (rx_stats_active)
3572 rx_atomic_inc(&rx_stats.dupPacketsRead);
3573 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3574 rxevent_Cancel(call->delayedAckEvent, call,
3575 RX_CALL_REFCOUNT_DELAY);
3576 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3582 /* It's the next packet. Stick it on the receive queue
3583 * for this call. Set newPackets to make sure we wake
3584 * the reader once all packets have been processed */
3585 #ifdef RX_TRACK_PACKETS
3586 np->flags |= RX_PKTFLAG_RQ;
3588 queue_Prepend(&call->rq, np);
3589 #ifdef RXDEBUG_PACKET
3591 #endif /* RXDEBUG_PACKET */
3593 np = NULL; /* We can't use this anymore */
3596 /* If an ack is requested then set a flag to make sure we
3597 * send an acknowledgement for this packet */
3598 if (flags & RX_REQUEST_ACK) {
3599 ackNeeded = RX_ACK_REQUESTED;
3602 /* Keep track of whether we have received the last packet */
3603 if (flags & RX_LAST_PACKET) {
3604 call->flags |= RX_CALL_HAVE_LAST;
3608 /* Check whether we have all of the packets for this call */
3609 if (call->flags & RX_CALL_HAVE_LAST) {
3610 afs_uint32 tseq; /* temporary sequence number */
3611 struct rx_packet *tp; /* Temporary packet pointer */
3612 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3614 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3615 if (tseq != tp->header.seq)
3617 if (tp->header.flags & RX_LAST_PACKET) {
3618 call->flags |= RX_CALL_RECEIVE_DONE;
3625 /* Provide asynchronous notification for those who want it
3626 * (e.g. multi rx) */
3627 if (call->arrivalProc) {
3628 (*call->arrivalProc) (call, call->arrivalProcHandle,
3629 call->arrivalProcArg);
3630 call->arrivalProc = (void (*)())0;
3633 /* Update last packet received */
3636 /* If there is no server process serving this call, grab
3637 * one, if available. We only need to do this once. If a
3638 * server thread is available, this thread becomes a server
3639 * thread and the server thread becomes a listener thread. */
3641 TryAttach(call, socket, tnop, newcallp, 0);
3644 /* This is not the expected next packet. */
3646 /* Determine whether this is a new or old packet, and if it's
3647 * a new one, whether it fits into the current receive window.
3648 * Also figure out whether the packet was delivered in sequence.
3649 * We use the prev variable to determine whether the new packet
3650 * is the successor of its immediate predecessor in the
3651 * receive queue, and the missing flag to determine whether
3652 * any of this packets predecessors are missing. */
3654 afs_uint32 prev; /* "Previous packet" sequence number */
3655 struct rx_packet *tp; /* Temporary packet pointer */
3656 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3657 int missing; /* Are any predecessors missing? */
3659 /* If the new packet's sequence number has been sent to the
3660 * application already, then this is a duplicate */
3661 if (seq < call->rnext) {
3662 if (rx_stats_active)
3663 rx_atomic_inc(&rx_stats.dupPacketsRead);
3664 rxevent_Cancel(call->delayedAckEvent, call,
3665 RX_CALL_REFCOUNT_DELAY);
3666 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3672 /* If the sequence number is greater than what can be
3673 * accomodated by the current window, then send a negative
3674 * acknowledge and drop the packet */
3675 if ((call->rnext + call->rwind) <= seq) {
3676 rxevent_Cancel(call->delayedAckEvent, call,
3677 RX_CALL_REFCOUNT_DELAY);
3678 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3685 /* Look for the packet in the queue of old received packets */
3686 for (prev = call->rnext - 1, missing =
3687 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3688 /*Check for duplicate packet */
3689 if (seq == tp->header.seq) {
3690 if (rx_stats_active)
3691 rx_atomic_inc(&rx_stats.dupPacketsRead);
3692 rxevent_Cancel(call->delayedAckEvent, call,
3693 RX_CALL_REFCOUNT_DELAY);
3694 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3700 /* If we find a higher sequence packet, break out and
3701 * insert the new packet here. */
3702 if (seq < tp->header.seq)
3704 /* Check for missing packet */
3705 if (tp->header.seq != prev + 1) {
3709 prev = tp->header.seq;
3712 /* Keep track of whether we have received the last packet. */
3713 if (flags & RX_LAST_PACKET) {
3714 call->flags |= RX_CALL_HAVE_LAST;
3717 /* It's within the window: add it to the the receive queue.
3718 * tp is left by the previous loop either pointing at the
3719 * packet before which to insert the new packet, or at the
3720 * queue head if the queue is empty or the packet should be
3722 #ifdef RX_TRACK_PACKETS
3723 np->flags |= RX_PKTFLAG_RQ;
3725 #ifdef RXDEBUG_PACKET
3727 #endif /* RXDEBUG_PACKET */
3728 queue_InsertBefore(tp, np);
3732 /* Check whether we have all of the packets for this call */
3733 if ((call->flags & RX_CALL_HAVE_LAST)
3734 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3735 afs_uint32 tseq; /* temporary sequence number */
3738 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3739 if (tseq != tp->header.seq)
3741 if (tp->header.flags & RX_LAST_PACKET) {
3742 call->flags |= RX_CALL_RECEIVE_DONE;
3749 /* We need to send an ack of the packet is out of sequence,
3750 * or if an ack was requested by the peer. */
3751 if (seq != prev + 1 || missing) {
3752 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3753 } else if (flags & RX_REQUEST_ACK) {
3754 ackNeeded = RX_ACK_REQUESTED;
3757 /* Acknowledge the last packet for each call */
3758 if (flags & RX_LAST_PACKET) {
3769 * If the receiver is waiting for an iovec, fill the iovec
3770 * using the data from the receive queue */
3771 if (call->flags & RX_CALL_IOVEC_WAIT) {
3772 didHardAck = rxi_FillReadVec(call, serial);
3773 /* the call may have been aborted */
3782 /* Wakeup the reader if any */
3783 if ((call->flags & RX_CALL_READER_WAIT)
3784 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3785 || (call->iovNext >= call->iovMax)
3786 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3787 call->flags &= ~RX_CALL_READER_WAIT;
3788 #ifdef RX_ENABLE_LOCKS
3789 CV_BROADCAST(&call->cv_rq);
3791 osi_rxWakeup(&call->rq);
3797 * Send an ack when requested by the peer, or once every
3798 * rxi_SoftAckRate packets until the last packet has been
3799 * received. Always send a soft ack for the last packet in
3800 * the server's reply.
3802 * If we have received all of the packets for the call
3803 * immediately send an RX_PACKET_TYPE_ACKALL packet so that
3804 * the peer can empty its packet queue and cancel all resend
3807 if (call->flags & RX_CALL_RECEIVE_DONE) {
3808 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3809 rxi_AckAll(NULL, call, 0);
3810 } else if (ackNeeded) {
3811 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3812 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3813 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3814 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3815 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3816 } else if (call->nSoftAcks) {
3817 clock_GetTime(&now);
3819 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3820 clock_Add(&when, &rx_lastAckDelay);
3822 clock_Add(&when, &rx_softAckDelay);
3824 if (!call->delayedAckEvent
3825 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3826 rxevent_Cancel(call->delayedAckEvent, call,
3827 RX_CALL_REFCOUNT_DELAY);
3828 MUTEX_ENTER(&rx_refcnt_mutex);
3829 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3830 MUTEX_EXIT(&rx_refcnt_mutex);
3831 call->delayedAckEvent =
3832 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3840 static void rxi_ComputeRate();
3844 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3846 struct rx_peer *peer = conn->peer;
3848 MUTEX_ENTER(&peer->peer_lock);
3849 peer->lastReachTime = clock_Sec();
3850 MUTEX_EXIT(&peer->peer_lock);
3852 MUTEX_ENTER(&conn->conn_data_lock);
3853 if (conn->flags & RX_CONN_ATTACHWAIT) {
3856 conn->flags &= ~RX_CONN_ATTACHWAIT;
3857 MUTEX_EXIT(&conn->conn_data_lock);
3859 for (i = 0; i < RX_MAXCALLS; i++) {
3860 struct rx_call *call = conn->call[i];
3863 MUTEX_ENTER(&call->lock);
3864 /* tnop can be null if newcallp is null */
3865 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3867 MUTEX_EXIT(&call->lock);
3871 MUTEX_EXIT(&conn->conn_data_lock);
3874 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
3876 rx_ack_reason(int reason)
3879 case RX_ACK_REQUESTED:
3881 case RX_ACK_DUPLICATE:
3883 case RX_ACK_OUT_OF_SEQUENCE:
3885 case RX_ACK_EXCEEDS_WINDOW:
3887 case RX_ACK_NOSPACE:
3891 case RX_ACK_PING_RESPONSE:
3904 /* The real smarts of the whole thing. */
3906 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
3909 struct rx_ackPacket *ap;
3911 struct rx_packet *tp;
3912 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3913 struct rx_connection *conn = call->conn;
3914 struct rx_peer *peer = conn->peer;
3915 struct clock now; /* Current time, for RTT calculations */
3919 /* because there are CM's that are bogus, sending weird values for this. */
3920 afs_uint32 skew = 0;
3925 int newAckCount = 0;
3926 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3927 int pktsize = 0; /* Set if we need to update the peer mtu */
3928 int conn_data_locked = 0;
3930 if (rx_stats_active)
3931 rx_atomic_inc(&rx_stats.ackPacketsRead);
3932 ap = (struct rx_ackPacket *)rx_DataOf(np);
3933 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3935 return np; /* truncated ack packet */
3937 /* depends on ack packet struct */
3938 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3939 first = ntohl(ap->firstPacket);
3940 prev = ntohl(ap->previousPacket);
3941 serial = ntohl(ap->serial);
3942 /* temporarily disabled -- needs to degrade over time
3943 * skew = ntohs(ap->maxSkew); */
3945 /* Ignore ack packets received out of order */
3946 if (first < call->tfirst ||
3947 (first == call->tfirst && prev < call->tprev)) {
3953 if (np->header.flags & RX_SLOW_START_OK) {
3954 call->flags |= RX_CALL_SLOW_START_OK;
3957 if (ap->reason == RX_ACK_PING_RESPONSE)
3958 rxi_UpdatePeerReach(conn, call);
3960 if (conn->lastPacketSizeSeq) {
3961 MUTEX_ENTER(&conn->conn_data_lock);
3962 conn_data_locked = 1;
3963 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
3964 pktsize = conn->lastPacketSize;
3965 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
3968 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
3969 if (!conn_data_locked) {
3970 MUTEX_ENTER(&conn->conn_data_lock);
3971 conn_data_locked = 1;
3973 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
3974 /* process mtu ping ack */
3975 pktsize = conn->lastPingSize;
3976 conn->lastPingSizeSer = conn->lastPingSize = 0;
3980 if (conn_data_locked) {
3981 MUTEX_EXIT(&conn->conn_data_lock);
3982 conn_data_locked = 0;
3986 if (rxdebug_active) {
3990 len = _snprintf(msg, sizeof(msg),
3991 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3992 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3993 ntohl(ap->serial), ntohl(ap->previousPacket),
3994 (unsigned int)np->header.seq, (unsigned int)skew,
3995 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3999 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4000 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4004 OutputDebugString(msg);
4006 #else /* AFS_NT40_ENV */
4009 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
4010 ap->reason, ntohl(ap->previousPacket),
4011 (unsigned int)np->header.seq, (unsigned int)serial,
4012 (unsigned int)skew, ntohl(ap->firstPacket));
4015 for (offset = 0; offset < nAcks; offset++)
4016 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4021 #endif /* AFS_NT40_ENV */
4024 MUTEX_ENTER(&peer->peer_lock);
4027 * Start somewhere. Can't assume we can send what we can receive,
4028 * but we are clearly receiving.
4030 if (!peer->maxPacketSize)
4031 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4033 if (pktsize > peer->maxPacketSize) {
4034 peer->maxPacketSize = pktsize;
4035 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4036 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4037 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4038 rxi_ScheduleGrowMTUEvent(call, 1);
4043 /* Update the outgoing packet skew value to the latest value of
4044 * the peer's incoming packet skew value. The ack packet, of
4045 * course, could arrive out of order, but that won't affect things
4047 peer->outPacketSkew = skew;
4049 /* Check for packets that no longer need to be transmitted, and
4050 * discard them. This only applies to packets positively
4051 * acknowledged as having been sent to the peer's upper level.
4052 * All other packets must be retained. So only packets with
4053 * sequence numbers < ap->firstPacket are candidates. */
4055 clock_GetTime(&now);
4057 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
4058 if (tp->header.seq >= first)
4060 call->tfirst = tp->header.seq + 1;
4062 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4064 if (ap->reason != RX_ACK_DELAY &&
4065 clock_Eq(&tp->timeSent, &tp->firstSent)) {
4066 rxi_ComputeRoundTripTime(tp, &tp->timeSent, call->conn->peer,
4072 rxi_ComputeRate(call->conn->peer, call, p, np, ap->reason);
4075 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4076 /* XXX Hack. Because we have to release the global rx lock when sending
4077 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4078 * in rxi_Start sending packets out because packets may move to the
4079 * freePacketQueue as result of being here! So we drop these packets until
4080 * we're safely out of the traversing. Really ugly!
4081 * To make it even uglier, if we're using fine grain locking, we can
4082 * set the ack bits in the packets and have rxi_Start remove the packets
4083 * when it's done transmitting.
4085 if (call->flags & RX_CALL_TQ_BUSY) {
4086 #ifdef RX_ENABLE_LOCKS
4087 tp->flags |= RX_PKTFLAG_ACKED;
4088 call->flags |= RX_CALL_TQ_SOME_ACKED;
4089 #else /* RX_ENABLE_LOCKS */
4091 #endif /* RX_ENABLE_LOCKS */
4093 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4096 #ifdef RX_TRACK_PACKETS
4097 tp->flags &= ~RX_PKTFLAG_TQ;
4099 #ifdef RXDEBUG_PACKET
4101 #endif /* RXDEBUG_PACKET */
4102 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4107 /* Give rate detector a chance to respond to ping requests */
4108 if (ap->reason == RX_ACK_PING_RESPONSE) {
4109 rxi_ComputeRate(peer, call, 0, np, ap->reason);
4113 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4115 /* Now go through explicit acks/nacks and record the results in
4116 * the waiting packets. These are packets that can't be released
4117 * yet, even with a positive acknowledge. This positive
4118 * acknowledge only means the packet has been received by the
4119 * peer, not that it will be retained long enough to be sent to
4120 * the peer's upper level. In addition, reset the transmit timers
4121 * of any missing packets (those packets that must be missing
4122 * because this packet was out of sequence) */
4124 call->nSoftAcked = 0;
4125 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
4127 /* Set the acknowledge flag per packet based on the
4128 * information in the ack packet. An acknowlegded packet can
4129 * be downgraded when the server has discarded a packet it
4130 * soacked previously, or when an ack packet is received
4131 * out of sequence. */
4132 if (tp->header.seq < first) {
4133 /* Implicit ack information */
4134 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4137 tp->flags |= RX_PKTFLAG_ACKED;
4138 } else if (tp->header.seq < first + nAcks) {
4139 /* Explicit ack information: set it in the packet appropriately */
4140 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4141 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4143 tp->flags |= RX_PKTFLAG_ACKED;
4145 if (ap->reason != RX_ACK_DELAY &&
4146 clock_Eq(&tp->timeSent, &tp->firstSent)) {
4147 rxi_ComputeRoundTripTime(tp, &tp->timeSent,
4148 call->conn->peer, &now);
4151 rxi_ComputeRate(call->conn->peer, call, tp, np,
4160 } else /* RX_ACK_TYPE_NACK */ {
4161 tp->flags &= ~RX_PKTFLAG_ACKED;
4165 if (tp->flags & RX_PKTFLAG_ACKED) {
4166 tp->flags &= ~RX_PKTFLAG_ACKED;
4172 * Following the suggestion of Phil Kern, we back off the peer's
4173 * timeout value for future packets until a successful response
4174 * is received for an initial transmission.
4176 if (missing && !peer->backedOff) {
4177 struct clock c = peer->timeout;
4178 struct clock max_to = {3, 0};
4180 clock_Add(&peer->timeout, &c);
4181 if (clock_Gt(&peer->timeout, &max_to))
4182 peer->timeout = max_to;
4183 peer->backedOff = 1;
4186 /* If packet isn't yet acked, and it has been transmitted at least
4187 * once, reset retransmit time using latest timeout
4188 * ie, this should readjust the retransmit timer for all outstanding
4189 * packets... So we don't just retransmit when we should know better*/
4191 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
4192 tp->retryTime = tp->timeSent;
4193 clock_Add(&tp->retryTime, &peer->timeout);
4194 /* shift by eight because one quarter-sec ~ 256 milliseconds */
4195 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
4199 /* If the window has been extended by this acknowledge packet,
4200 * then wakeup a sender waiting in alloc for window space, or try
4201 * sending packets now, if he's been sitting on packets due to
4202 * lack of window space */
4203 if (call->tnext < (call->tfirst + call->twind)) {
4204 #ifdef RX_ENABLE_LOCKS
4205 CV_SIGNAL(&call->cv_twind);
4207 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4208 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4209 osi_rxWakeup(&call->twind);
4212 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4213 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4217 /* if the ack packet has a receivelen field hanging off it,
4218 * update our state */
4219 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4222 /* If the ack packet has a "recommended" size that is less than
4223 * what I am using now, reduce my size to match */
4224 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4225 (int)sizeof(afs_int32), &tSize);
4226 tSize = (afs_uint32) ntohl(tSize);
4227 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4229 /* Get the maximum packet size to send to this peer */
4230 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4232 tSize = (afs_uint32) ntohl(tSize);
4233 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4234 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4236 /* sanity check - peer might have restarted with different params.
4237 * If peer says "send less", dammit, send less... Peer should never
4238 * be unable to accept packets of the size that prior AFS versions would
4239 * send without asking. */
4240 if (peer->maxMTU != tSize) {
4241 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4243 peer->maxMTU = tSize;
4244 peer->MTU = MIN(tSize, peer->MTU);
4245 call->MTU = MIN(call->MTU, tSize);
4248 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4251 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4252 (int)sizeof(afs_int32), &tSize);
4253 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4254 if (tSize < call->twind) { /* smaller than our send */
4255 call->twind = tSize; /* window, we must send less... */
4256 call->ssthresh = MIN(call->twind, call->ssthresh);
4257 call->conn->twind[call->channel] = call->twind;
4260 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4261 * network MTU confused with the loopback MTU. Calculate the
4262 * maximum MTU here for use in the slow start code below.
4264 /* Did peer restart with older RX version? */
4265 if (peer->maxDgramPackets > 1) {
4266 peer->maxDgramPackets = 1;
4268 } else if (np->length >=
4269 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4272 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4273 sizeof(afs_int32), &tSize);
4274 tSize = (afs_uint32) ntohl(tSize);
4276 * As of AFS 3.5 we set the send window to match the receive window.
4278 if (tSize < call->twind) {
4279 call->twind = tSize;
4280 call->conn->twind[call->channel] = call->twind;
4281 call->ssthresh = MIN(call->twind, call->ssthresh);
4282 } else if (tSize > call->twind) {
4283 call->twind = tSize;
4284 call->conn->twind[call->channel] = call->twind;
4288 * As of AFS 3.5, a jumbogram is more than one fixed size
4289 * packet transmitted in a single UDP datagram. If the remote
4290 * MTU is smaller than our local MTU then never send a datagram
4291 * larger than the natural MTU.
4294 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4295 (int)sizeof(afs_int32), &tSize);
4296 maxDgramPackets = (afs_uint32) ntohl(tSize);
4297 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4299 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4300 if (maxDgramPackets > 1) {
4301 peer->maxDgramPackets = maxDgramPackets;
4302 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4304 peer->maxDgramPackets = 1;
4305 call->MTU = peer->natMTU;
4307 } else if (peer->maxDgramPackets > 1) {
4308 /* Restarted with lower version of RX */
4309 peer->maxDgramPackets = 1;
4311 } else if (peer->maxDgramPackets > 1
4312 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4313 /* Restarted with lower version of RX */
4314 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4315 peer->natMTU = OLD_MAX_PACKET_SIZE;
4316 peer->MTU = OLD_MAX_PACKET_SIZE;
4317 peer->maxDgramPackets = 1;
4318 peer->nDgramPackets = 1;
4320 call->MTU = OLD_MAX_PACKET_SIZE;
4325 * Calculate how many datagrams were successfully received after
4326 * the first missing packet and adjust the negative ack counter
4331 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4332 if (call->nNacks < nNacked) {
4333 call->nNacks = nNacked;
4336 call->nAcks += newAckCount;
4340 if (call->flags & RX_CALL_FAST_RECOVER) {
4342 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4344 call->flags &= ~RX_CALL_FAST_RECOVER;
4345 call->cwind = call->nextCwind;
4346 call->nextCwind = 0;
4349 call->nCwindAcks = 0;
4350 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4351 /* Three negative acks in a row trigger congestion recovery */
4352 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4353 MUTEX_EXIT(&peer->peer_lock);
4354 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4355 /* someone else is waiting to start recovery */
4358 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4359 rxi_WaitforTQBusy(call);
4360 MUTEX_ENTER(&peer->peer_lock);
4361 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4362 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4363 call->flags |= RX_CALL_FAST_RECOVER;
4364 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4366 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4367 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4368 call->nextCwind = call->ssthresh;
4371 peer->MTU = call->MTU;
4372 peer->cwind = call->nextCwind;
4373 peer->nDgramPackets = call->nDgramPackets;
4375 call->congestSeq = peer->congestSeq;
4376 /* Reset the resend times on the packets that were nacked
4377 * so we will retransmit as soon as the window permits*/
4378 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4380 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4381 clock_Zero(&tp->retryTime);
4383 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4388 /* If cwind is smaller than ssthresh, then increase
4389 * the window one packet for each ack we receive (exponential
4391 * If cwind is greater than or equal to ssthresh then increase
4392 * the congestion window by one packet for each cwind acks we
4393 * receive (linear growth). */
4394 if (call->cwind < call->ssthresh) {
4396 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4397 call->nCwindAcks = 0;
4399 call->nCwindAcks += newAckCount;
4400 if (call->nCwindAcks >= call->cwind) {
4401 call->nCwindAcks = 0;
4402 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4406 * If we have received several acknowledgements in a row then
4407 * it is time to increase the size of our datagrams
4409 if ((int)call->nAcks > rx_nDgramThreshold) {
4410 if (peer->maxDgramPackets > 1) {
4411 if (call->nDgramPackets < peer->maxDgramPackets) {
4412 call->nDgramPackets++;
4414 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4415 } else if (call->MTU < peer->maxMTU) {
4416 /* don't upgrade if we can't handle it */
4417 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4418 call->MTU = peer->ifMTU;
4420 call->MTU += peer->natMTU;
4421 call->MTU = MIN(call->MTU, peer->maxMTU);
4428 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4430 /* Servers need to hold the call until all response packets have
4431 * been acknowledged. Soft acks are good enough since clients
4432 * are not allowed to clear their receive queues. */
4433 if (call->state == RX_STATE_HOLD
4434 && call->tfirst + call->nSoftAcked >= call->tnext) {
4435 call->state = RX_STATE_DALLY;
4436 rxi_ClearTransmitQueue(call, 0);
4437 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4438 } else if (!queue_IsEmpty(&call->tq)) {
4439 rxi_Start(0, call, 0, istack);
4444 /* Received a response to a challenge packet */
4446 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4447 struct rx_packet *np, int istack)
4451 /* Ignore the packet if we're the client */
4452 if (conn->type == RX_CLIENT_CONNECTION)
4455 /* If already authenticated, ignore the packet (it's probably a retry) */
4456 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4459 /* Otherwise, have the security object evaluate the response packet */
4460 error = RXS_CheckResponse(conn->securityObject, conn, np);
4462 /* If the response is invalid, reset the connection, sending
4463 * an abort to the peer */
4467 rxi_ConnectionError(conn, error);
4468 MUTEX_ENTER(&conn->conn_data_lock);
4469 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4470 MUTEX_EXIT(&conn->conn_data_lock);
4473 /* If the response is valid, any calls waiting to attach
4474 * servers can now do so */
4477 for (i = 0; i < RX_MAXCALLS; i++) {
4478 struct rx_call *call = conn->call[i];
4480 MUTEX_ENTER(&call->lock);
4481 if (call->state == RX_STATE_PRECALL)
4482 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4483 /* tnop can be null if newcallp is null */
4484 MUTEX_EXIT(&call->lock);
4488 /* Update the peer reachability information, just in case
4489 * some calls went into attach-wait while we were waiting
4490 * for authentication..
4492 rxi_UpdatePeerReach(conn, NULL);
4497 /* A client has received an authentication challenge: the security
4498 * object is asked to cough up a respectable response packet to send
4499 * back to the server. The server is responsible for retrying the
4500 * challenge if it fails to get a response. */
4503 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4504 struct rx_packet *np, int istack)
4508 /* Ignore the challenge if we're the server */
4509 if (conn->type == RX_SERVER_CONNECTION)
4512 /* Ignore the challenge if the connection is otherwise idle; someone's
4513 * trying to use us as an oracle. */
4514 if (!rxi_HasActiveCalls(conn))
4517 /* Send the security object the challenge packet. It is expected to fill
4518 * in the response. */
4519 error = RXS_GetResponse(conn->securityObject, conn, np);
4521 /* If the security object is unable to return a valid response, reset the
4522 * connection and send an abort to the peer. Otherwise send the response
4523 * packet to the peer connection. */
4525 rxi_ConnectionError(conn, error);
4526 MUTEX_ENTER(&conn->conn_data_lock);
4527 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4528 MUTEX_EXIT(&conn->conn_data_lock);
4530 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4531 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4537 /* Find an available server process to service the current request in
4538 * the given call structure. If one isn't available, queue up this
4539 * call so it eventually gets one */
4541 rxi_AttachServerProc(struct rx_call *call,
4542 osi_socket socket, int *tnop,
4543 struct rx_call **newcallp)
4545 struct rx_serverQueueEntry *sq;
4546 struct rx_service *service = call->conn->service;
4549 /* May already be attached */
4550 if (call->state == RX_STATE_ACTIVE)
4553 MUTEX_ENTER(&rx_serverPool_lock);
4555 haveQuota = QuotaOK(service);
4556 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4557 /* If there are no processes available to service this call,
4558 * put the call on the incoming call queue (unless it's
4559 * already on the queue).
4561 #ifdef RX_ENABLE_LOCKS
4563 ReturnToServerPool(service);
4564 #endif /* RX_ENABLE_LOCKS */
4566 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4567 call->flags |= RX_CALL_WAIT_PROC;
4568 rx_atomic_inc(&rx_nWaiting);
4569 rx_atomic_inc(&rx_nWaited);
4570 rxi_calltrace(RX_CALL_ARRIVAL, call);
4571 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4572 queue_Append(&rx_incomingCallQueue, call);
4575 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4577 /* If hot threads are enabled, and both newcallp and sq->socketp
4578 * are non-null, then this thread will process the call, and the
4579 * idle server thread will start listening on this threads socket.
4582 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4585 *sq->socketp = socket;
4586 clock_GetTime(&call->startTime);
4587 MUTEX_ENTER(&rx_refcnt_mutex);
4588 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4589 MUTEX_EXIT(&rx_refcnt_mutex);
4593 if (call->flags & RX_CALL_WAIT_PROC) {
4594 /* Conservative: I don't think this should happen */
4595 call->flags &= ~RX_CALL_WAIT_PROC;
4596 if (queue_IsOnQueue(call)) {
4599 rx_atomic_dec(&rx_nWaiting);
4602 call->state = RX_STATE_ACTIVE;
4603 call->mode = RX_MODE_RECEIVING;
4604 #ifdef RX_KERNEL_TRACE
4606 int glockOwner = ISAFS_GLOCK();
4609 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4610 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4616 if (call->flags & RX_CALL_CLEARED) {
4617 /* send an ack now to start the packet flow up again */
4618 call->flags &= ~RX_CALL_CLEARED;
4619 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4621 #ifdef RX_ENABLE_LOCKS
4624 service->nRequestsRunning++;
4625 MUTEX_ENTER(&rx_quota_mutex);
4626 if (service->nRequestsRunning <= service->minProcs)
4629 MUTEX_EXIT(&rx_quota_mutex);
4633 MUTEX_EXIT(&rx_serverPool_lock);
4636 /* Delay the sending of an acknowledge event for a short while, while
4637 * a new call is being prepared (in the case of a client) or a reply
4638 * is being prepared (in the case of a server). Rather than sending
4639 * an ack packet, an ACKALL packet is sent. */
4641 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4643 #ifdef RX_ENABLE_LOCKS
4645 MUTEX_ENTER(&call->lock);
4646 call->delayedAckEvent = NULL;
4647 MUTEX_ENTER(&rx_refcnt_mutex);
4648 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4649 MUTEX_EXIT(&rx_refcnt_mutex);
4651 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4652 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4654 MUTEX_EXIT(&call->lock);
4655 #else /* RX_ENABLE_LOCKS */
4657 call->delayedAckEvent = NULL;
4658 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4659 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4660 #endif /* RX_ENABLE_LOCKS */
4664 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4666 struct rx_call *call = arg1;
4667 #ifdef RX_ENABLE_LOCKS
4669 MUTEX_ENTER(&call->lock);
4670 if (event == call->delayedAckEvent)
4671 call->delayedAckEvent = NULL;
4672 MUTEX_ENTER(&rx_refcnt_mutex);
4673 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4674 MUTEX_EXIT(&rx_refcnt_mutex);
4676 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4678 MUTEX_EXIT(&call->lock);
4679 #else /* RX_ENABLE_LOCKS */
4681 call->delayedAckEvent = NULL;
4682 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4683 #endif /* RX_ENABLE_LOCKS */
4687 #ifdef RX_ENABLE_LOCKS
4688 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4689 * clearing them out.
4692 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4694 struct rx_packet *p, *tp;
4697 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4698 p->flags |= RX_PKTFLAG_ACKED;
4702 call->flags |= RX_CALL_TQ_CLEARME;
4703 call->flags |= RX_CALL_TQ_SOME_ACKED;
4706 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4707 call->tfirst = call->tnext;
4708 call->nSoftAcked = 0;
4710 if (call->flags & RX_CALL_FAST_RECOVER) {
4711 call->flags &= ~RX_CALL_FAST_RECOVER;
4712 call->cwind = call->nextCwind;
4713 call->nextCwind = 0;
4716 CV_SIGNAL(&call->cv_twind);
4718 #endif /* RX_ENABLE_LOCKS */
4720 /* Clear out the transmit queue for the current call (all packets have
4721 * been received by peer) */
4723 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4725 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4726 struct rx_packet *p, *tp;
4728 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4730 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4731 p->flags |= RX_PKTFLAG_ACKED;
4735 call->flags |= RX_CALL_TQ_CLEARME;
4736 call->flags |= RX_CALL_TQ_SOME_ACKED;
4739 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4740 #ifdef RXDEBUG_PACKET
4742 #endif /* RXDEBUG_PACKET */
4743 rxi_FreePackets(0, &call->tq);
4744 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
4745 #ifdef RX_ENABLE_LOCKS
4746 CV_BROADCAST(&call->cv_tq);
4747 #else /* RX_ENABLE_LOCKS */
4748 osi_rxWakeup(&call->tq);
4749 #endif /* RX_ENABLE_LOCKS */
4751 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4752 call->flags &= ~RX_CALL_TQ_CLEARME;
4754 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4756 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4757 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4758 call->nSoftAcked = 0;
4760 if (call->flags & RX_CALL_FAST_RECOVER) {
4761 call->flags &= ~RX_CALL_FAST_RECOVER;
4762 call->cwind = call->nextCwind;
4764 #ifdef RX_ENABLE_LOCKS
4765 CV_SIGNAL(&call->cv_twind);
4767 osi_rxWakeup(&call->twind);
4772 rxi_ClearReceiveQueue(struct rx_call *call)
4774 if (queue_IsNotEmpty(&call->rq)) {
4777 count = rxi_FreePackets(0, &call->rq);
4778 rx_packetReclaims += count;
4779 #ifdef RXDEBUG_PACKET
4781 if ( call->rqc != 0 )
4782 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
4784 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4786 if (call->state == RX_STATE_PRECALL) {
4787 call->flags |= RX_CALL_CLEARED;
4791 /* Send an abort packet for the specified call */
4793 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
4794 int istack, int force)
4797 struct clock when, now;
4802 /* Clients should never delay abort messages */
4803 if (rx_IsClientConn(call->conn))
4806 if (call->abortCode != call->error) {
4807 call->abortCode = call->error;
4808 call->abortCount = 0;
4811 if (force || rxi_callAbortThreshhold == 0
4812 || call->abortCount < rxi_callAbortThreshhold) {
4813 if (call->delayedAbortEvent) {
4814 rxevent_Cancel(call->delayedAbortEvent, call,
4815 RX_CALL_REFCOUNT_ABORT);
4817 error = htonl(call->error);
4820 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4821 (char *)&error, sizeof(error), istack);
4822 } else if (!call->delayedAbortEvent) {
4823 clock_GetTime(&now);
4825 clock_Addmsec(&when, rxi_callAbortDelay);
4826 MUTEX_ENTER(&rx_refcnt_mutex);
4827 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4828 MUTEX_EXIT(&rx_refcnt_mutex);
4829 call->delayedAbortEvent =
4830 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4835 /* Send an abort packet for the specified connection. Packet is an
4836 * optional pointer to a packet that can be used to send the abort.
4837 * Once the number of abort messages reaches the threshhold, an
4838 * event is scheduled to send the abort. Setting the force flag
4839 * overrides sending delayed abort messages.
4841 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4842 * to send the abort packet.
4845 rxi_SendConnectionAbort(struct rx_connection *conn,
4846 struct rx_packet *packet, int istack, int force)
4849 struct clock when, now;
4854 /* Clients should never delay abort messages */
4855 if (rx_IsClientConn(conn))
4858 if (force || rxi_connAbortThreshhold == 0
4859 || conn->abortCount < rxi_connAbortThreshhold) {
4860 if (conn->delayedAbortEvent) {
4861 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4863 error = htonl(conn->error);
4865 MUTEX_EXIT(&conn->conn_data_lock);
4867 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4868 RX_PACKET_TYPE_ABORT, (char *)&error,
4869 sizeof(error), istack);
4870 MUTEX_ENTER(&conn->conn_data_lock);
4871 } else if (!conn->delayedAbortEvent) {
4872 clock_GetTime(&now);
4874 clock_Addmsec(&when, rxi_connAbortDelay);
4875 conn->delayedAbortEvent =
4876 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4881 /* Associate an error all of the calls owned by a connection. Called
4882 * with error non-zero. This is only for really fatal things, like
4883 * bad authentication responses. The connection itself is set in
4884 * error at this point, so that future packets received will be
4887 rxi_ConnectionError(struct rx_connection *conn,
4893 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
4895 MUTEX_ENTER(&conn->conn_data_lock);
4896 if (conn->challengeEvent)
4897 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4898 if (conn->natKeepAliveEvent)
4899 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
4900 if (conn->checkReachEvent) {
4901 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4902 conn->checkReachEvent = 0;
4903 conn->flags &= ~RX_CONN_ATTACHWAIT;
4904 MUTEX_ENTER(&rx_refcnt_mutex);
4906 MUTEX_EXIT(&rx_refcnt_mutex);
4908 MUTEX_EXIT(&conn->conn_data_lock);
4909 for (i = 0; i < RX_MAXCALLS; i++) {
4910 struct rx_call *call = conn->call[i];
4912 MUTEX_ENTER(&call->lock);
4913 rxi_CallError(call, error);
4914 MUTEX_EXIT(&call->lock);
4917 conn->error = error;
4918 if (rx_stats_active)
4919 rx_atomic_inc(&rx_stats.fatalErrors);
4924 * Interrupt an in-progress call with the specified error and wakeup waiters.
4926 * @param[in] call The call to interrupt
4927 * @param[in] error The error code to send to the peer
4930 rx_InterruptCall(struct rx_call *call, afs_int32 error)
4932 MUTEX_ENTER(&call->lock);
4933 rxi_CallError(call, error);
4934 rxi_SendCallAbort(call, NULL, 0, 1);
4935 MUTEX_EXIT(&call->lock);
4939 rxi_CallError(struct rx_call *call, afs_int32 error)
4942 osirx_AssertMine(&call->lock, "rxi_CallError");
4944 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
4946 error = call->error;
4948 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4949 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4950 rxi_ResetCall(call, 0);
4953 rxi_ResetCall(call, 0);
4955 call->error = error;
4958 /* Reset various fields in a call structure, and wakeup waiting
4959 * processes. Some fields aren't changed: state & mode are not
4960 * touched (these must be set by the caller), and bufptr, nLeft, and
4961 * nFree are not reset, since these fields are manipulated by
4962 * unprotected macros, and may only be reset by non-interrupting code.
4965 /* this code requires that call->conn be set properly as a pre-condition. */
4966 #endif /* ADAPT_WINDOW */
4969 rxi_ResetCall(struct rx_call *call, int newcall)
4972 struct rx_peer *peer;
4973 struct rx_packet *packet;
4975 osirx_AssertMine(&call->lock, "rxi_ResetCall");
4977 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
4979 /* Notify anyone who is waiting for asynchronous packet arrival */
4980 if (call->arrivalProc) {
4981 (*call->arrivalProc) (call, call->arrivalProcHandle,
4982 call->arrivalProcArg);
4983 call->arrivalProc = (void (*)())0;
4986 if (call->delayedAbortEvent) {
4987 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4988 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4990 rxi_SendCallAbort(call, packet, 0, 1);
4991 rxi_FreePacket(packet);
4996 * Update the peer with the congestion information in this call
4997 * so other calls on this connection can pick up where this call
4998 * left off. If the congestion sequence numbers don't match then
4999 * another call experienced a retransmission.
5001 peer = call->conn->peer;
5002 MUTEX_ENTER(&peer->peer_lock);
5004 if (call->congestSeq == peer->congestSeq) {
5005 peer->cwind = MAX(peer->cwind, call->cwind);
5006 peer->MTU = MAX(peer->MTU, call->MTU);
5007 peer->nDgramPackets =
5008 MAX(peer->nDgramPackets, call->nDgramPackets);
5011 call->abortCode = 0;
5012 call->abortCount = 0;
5014 if (peer->maxDgramPackets > 1) {
5015 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5017 call->MTU = peer->MTU;
5019 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5020 call->ssthresh = rx_maxSendWindow;
5021 call->nDgramPackets = peer->nDgramPackets;
5022 call->congestSeq = peer->congestSeq;
5023 MUTEX_EXIT(&peer->peer_lock);
5025 flags = call->flags;
5026 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5027 rxi_WaitforTQBusy(call);
5028 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5030 rxi_ClearTransmitQueue(call, 1);
5031 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5032 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5036 rxi_ClearReceiveQueue(call);
5037 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5041 call->twind = call->conn->twind[call->channel];
5042 call->rwind = call->conn->rwind[call->channel];
5043 call->nSoftAcked = 0;
5044 call->nextCwind = 0;
5047 call->nCwindAcks = 0;
5048 call->nSoftAcks = 0;
5049 call->nHardAcks = 0;
5051 call->tfirst = call->rnext = call->tnext = 1;
5054 call->lastAcked = 0;
5055 call->localStatus = call->remoteStatus = 0;
5057 if (flags & RX_CALL_READER_WAIT) {
5058 #ifdef RX_ENABLE_LOCKS
5059 CV_BROADCAST(&call->cv_rq);
5061 osi_rxWakeup(&call->rq);
5064 if (flags & RX_CALL_WAIT_PACKETS) {
5065 MUTEX_ENTER(&rx_freePktQ_lock);
5066 rxi_PacketsUnWait(); /* XXX */
5067 MUTEX_EXIT(&rx_freePktQ_lock);
5069 #ifdef RX_ENABLE_LOCKS
5070 CV_SIGNAL(&call->cv_twind);
5072 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5073 osi_rxWakeup(&call->twind);
5076 #ifdef RX_ENABLE_LOCKS
5077 /* The following ensures that we don't mess with any queue while some
5078 * other thread might also be doing so. The call_queue_lock field is
5079 * is only modified under the call lock. If the call is in the process
5080 * of being removed from a queue, the call is not locked until the
5081 * the queue lock is dropped and only then is the call_queue_lock field
5082 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5083 * Note that any other routine which removes a call from a queue has to
5084 * obtain the queue lock before examing the queue and removing the call.
5086 if (call->call_queue_lock) {
5087 MUTEX_ENTER(call->call_queue_lock);
5088 if (queue_IsOnQueue(call)) {
5090 if (flags & RX_CALL_WAIT_PROC) {
5091 rx_atomic_dec(&rx_nWaiting);
5094 MUTEX_EXIT(call->call_queue_lock);
5095 CLEAR_CALL_QUEUE_LOCK(call);
5097 #else /* RX_ENABLE_LOCKS */
5098 if (queue_IsOnQueue(call)) {
5100 if (flags & RX_CALL_WAIT_PROC)
5101 rx_atomic_dec(&rx_nWaiting);
5103 #endif /* RX_ENABLE_LOCKS */
5105 rxi_KeepAliveOff(call);
5106 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5109 /* Send an acknowledge for the indicated packet (seq,serial) of the
5110 * indicated call, for the indicated reason (reason). This
5111 * acknowledge will specifically acknowledge receiving the packet, and
5112 * will also specify which other packets for this call have been
5113 * received. This routine returns the packet that was used to the
5114 * caller. The caller is responsible for freeing it or re-using it.
5115 * This acknowledgement also returns the highest sequence number
5116 * actually read out by the higher level to the sender; the sender
5117 * promises to keep around packets that have not been read by the
5118 * higher level yet (unless, of course, the sender decides to abort
5119 * the call altogether). Any of p, seq, serial, pflags, or reason may
5120 * be set to zero without ill effect. That is, if they are zero, they
5121 * will not convey any information.
5122 * NOW there is a trailer field, after the ack where it will safely be
5123 * ignored by mundanes, which indicates the maximum size packet this
5124 * host can swallow. */
5126 struct rx_packet *optionalPacket; use to send ack (or null)
5127 int seq; Sequence number of the packet we are acking
5128 int serial; Serial number of the packet
5129 int pflags; Flags field from packet header
5130 int reason; Reason an acknowledge was prompted
5134 rxi_SendAck(struct rx_call *call,
5135 struct rx_packet *optionalPacket, int serial, int reason,
5138 struct rx_ackPacket *ap;
5139 struct rx_packet *rqp;
5140 struct rx_packet *nxp; /* For queue_Scan */
5141 struct rx_packet *p;
5144 afs_uint32 padbytes = 0;
5145 #ifdef RX_ENABLE_TSFPQ
5146 struct rx_ts_info_t * rx_ts_info;
5150 * Open the receive window once a thread starts reading packets
5152 if (call->rnext > 1) {
5153 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5156 /* Don't attempt to grow MTU if this is a critical ping */
5157 if (reason == RX_ACK_MTU) {
5158 /* keep track of per-call attempts, if we're over max, do in small
5159 * otherwise in larger? set a size to increment by, decrease
5162 if (call->conn->peer->maxPacketSize &&
5163 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5165 padbytes = call->conn->peer->maxPacketSize+16;
5167 padbytes = call->conn->peer->maxMTU + 128;
5169 /* do always try a minimum size ping */
5170 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5172 /* subtract the ack payload */
5173 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5174 reason = RX_ACK_PING;
5177 call->nHardAcks = 0;
5178 call->nSoftAcks = 0;
5179 if (call->rnext > call->lastAcked)
5180 call->lastAcked = call->rnext;
5184 rx_computelen(p, p->length); /* reset length, you never know */
5185 } /* where that's been... */
5186 #ifdef RX_ENABLE_TSFPQ
5188 RX_TS_INFO_GET(rx_ts_info);
5189 if ((p = rx_ts_info->local_special_packet)) {
5190 rx_computelen(p, p->length);
5191 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5192 rx_ts_info->local_special_packet = p;
5193 } else { /* We won't send the ack, but don't panic. */
5194 return optionalPacket;
5198 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5199 /* We won't send the ack, but don't panic. */
5200 return optionalPacket;
5205 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5208 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5209 #ifndef RX_ENABLE_TSFPQ
5210 if (!optionalPacket)
5213 return optionalPacket;
5215 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5216 if (rx_Contiguous(p) < templ) {
5217 #ifndef RX_ENABLE_TSFPQ
5218 if (!optionalPacket)
5221 return optionalPacket;
5226 /* MTUXXX failing to send an ack is very serious. We should */
5227 /* try as hard as possible to send even a partial ack; it's */
5228 /* better than nothing. */
5229 ap = (struct rx_ackPacket *)rx_DataOf(p);
5230 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5231 ap->reason = reason;
5233 /* The skew computation used to be bogus, I think it's better now. */
5234 /* We should start paying attention to skew. XXX */
5235 ap->serial = htonl(serial);
5236 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5238 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
5239 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5241 /* No fear of running out of ack packet here because there can only be at most
5242 * one window full of unacknowledged packets. The window size must be constrained
5243 * to be less than the maximum ack size, of course. Also, an ack should always
5244 * fit into a single packet -- it should not ever be fragmented. */
5245 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5246 if (!rqp || !call->rq.next
5247 || (rqp->header.seq > (call->rnext + call->rwind))) {
5248 #ifndef RX_ENABLE_TSFPQ
5249 if (!optionalPacket)
5252 rxi_CallError(call, RX_CALL_DEAD);
5253 return optionalPacket;
5256 while (rqp->header.seq > call->rnext + offset)
5257 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5258 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5260 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5261 #ifndef RX_ENABLE_TSFPQ
5262 if (!optionalPacket)
5265 rxi_CallError(call, RX_CALL_DEAD);
5266 return optionalPacket;
5271 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5273 /* these are new for AFS 3.3 */
5274 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5275 templ = htonl(templ);
5276 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5277 templ = htonl(call->conn->peer->ifMTU);
5278 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5279 sizeof(afs_int32), &templ);
5281 /* new for AFS 3.4 */
5282 templ = htonl(call->rwind);
5283 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5284 sizeof(afs_int32), &templ);
5286 /* new for AFS 3.5 */
5287 templ = htonl(call->conn->peer->ifDgramPackets);
5288 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5289 sizeof(afs_int32), &templ);
5291 p->header.serviceId = call->conn->serviceId;
5292 p->header.cid = (call->conn->cid | call->channel);
5293 p->header.callNumber = *call->callNumber;
5295 p->header.securityIndex = call->conn->securityIndex;
5296 p->header.epoch = call->conn->epoch;
5297 p->header.type = RX_PACKET_TYPE_ACK;
5298 p->header.flags = RX_SLOW_START_OK;
5299 if (reason == RX_ACK_PING) {
5300 p->header.flags |= RX_REQUEST_ACK;
5302 clock_GetTime(&call->pingRequestTime);
5305 p->length = padbytes +
5306 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5309 /* not fast but we can potentially use this if truncated
5310 * fragments are delivered to figure out the mtu.
5312 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5313 sizeof(afs_int32), sizeof(afs_int32),
5317 if (call->conn->type == RX_CLIENT_CONNECTION)
5318 p->header.flags |= RX_CLIENT_INITIATED;
5322 if (rxdebug_active) {
5326 len = _snprintf(msg, sizeof(msg),
5327 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5328 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5329 ntohl(ap->serial), ntohl(ap->previousPacket),
5330 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5331 ap->nAcks, ntohs(ap->bufferSpace) );
5335 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5336 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5340 OutputDebugString(msg);
5342 #else /* AFS_NT40_ENV */
5344 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5345 ap->reason, ntohl(ap->previousPacket),
5346 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5348 for (offset = 0; offset < ap->nAcks; offset++)
5349 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5354 #endif /* AFS_NT40_ENV */
5357 int i, nbytes = p->length;
5359 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5360 if (nbytes <= p->wirevec[i].iov_len) {
5363 savelen = p->wirevec[i].iov_len;
5365 p->wirevec[i].iov_len = nbytes;
5367 rxi_Send(call, p, istack);
5368 p->wirevec[i].iov_len = savelen;
5372 nbytes -= p->wirevec[i].iov_len;
5375 if (rx_stats_active)
5376 rx_atomic_inc(&rx_stats.ackPacketsSent);
5377 #ifndef RX_ENABLE_TSFPQ
5378 if (!optionalPacket)
5381 return optionalPacket; /* Return packet for re-use by caller */
5384 /* Send all of the packets in the list in single datagram */
5386 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
5387 int istack, int moreFlag, struct clock *now,
5388 struct clock *retryTime, int resending)
5393 struct rx_connection *conn = call->conn;
5394 struct rx_peer *peer = conn->peer;
5396 MUTEX_ENTER(&peer->peer_lock);
5399 peer->reSends += len;
5400 MUTEX_EXIT(&peer->peer_lock);
5402 if (rx_stats_active) {
5404 rx_atomic_add(&rx_stats.dataPacketsReSent, len);
5406 rx_atomic_add(&rx_stats.dataPacketsSent, len);
5409 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
5413 /* Set the packet flags and schedule the resend events */
5414 /* Only request an ack for the last packet in the list */
5415 for (i = 0; i < len; i++) {
5416 list[i]->retryTime = *retryTime;
5417 if (list[i]->header.serial) {
5418 /* Exponentially backoff retry times */
5419 if (list[i]->backoff < MAXBACKOFF) {
5420 /* so it can't stay == 0 */
5421 list[i]->backoff = (list[i]->backoff << 1) + 1;
5424 clock_Addmsec(&(list[i]->retryTime),
5425 ((afs_uint32) list[i]->backoff) << 8);
5428 /* Wait a little extra for the ack on the last packet */
5429 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
5430 clock_Addmsec(&(list[i]->retryTime), 400);
5433 /* Record the time sent */
5434 list[i]->timeSent = *now;
5436 /* Ask for an ack on retransmitted packets, on every other packet
5437 * if the peer doesn't support slow start. Ask for an ack on every
5438 * packet until the congestion window reaches the ack rate. */
5439 if (list[i]->header.serial) {
5442 /* improved RTO calculation- not Karn */
5443 list[i]->firstSent = *now;
5444 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5445 || (!(call->flags & RX_CALL_SLOW_START_OK)
5446 && (list[i]->header.seq & 1)))) {
5451 /* Tag this packet as not being the last in this group,
5452 * for the receiver's benefit */
5453 if (i < len - 1 || moreFlag) {
5454 list[i]->header.flags |= RX_MORE_PACKETS;
5457 /* Install the new retransmit time for the packet, and
5458 * record the time sent */
5459 list[i]->timeSent = *now;
5463 list[len - 1]->header.flags |= RX_REQUEST_ACK;
5466 /* Since we're about to send a data packet to the peer, it's
5467 * safe to nuke any scheduled end-of-packets ack */
5468 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5470 MUTEX_EXIT(&call->lock);
5471 MUTEX_ENTER(&rx_refcnt_mutex);
5472 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5473 MUTEX_EXIT(&rx_refcnt_mutex);
5475 rxi_SendPacketList(call, conn, list, len, istack);
5477 rxi_SendPacket(call, conn, list[0], istack);
5479 MUTEX_ENTER(&call->lock);
5480 MUTEX_ENTER(&rx_refcnt_mutex);
5481 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5482 MUTEX_EXIT(&rx_refcnt_mutex);
5484 /* Update last send time for this call (for keep-alive
5485 * processing), and for the connection (so that we can discover
5486 * idle connections) */
5487 conn->lastSendTime = call->lastSendTime = clock_Sec();
5488 /* Let a set of retransmits trigger an idle timeout */
5490 call->lastSendData = call->lastSendTime;
5493 /* When sending packets we need to follow these rules:
5494 * 1. Never send more than maxDgramPackets in a jumbogram.
5495 * 2. Never send a packet with more than two iovecs in a jumbogram.
5496 * 3. Never send a retransmitted packet in a jumbogram.
5497 * 4. Never send more than cwind/4 packets in a jumbogram
5498 * We always keep the last list we should have sent so we
5499 * can set the RX_MORE_PACKETS flags correctly.
5502 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5503 int istack, struct clock *now, struct clock *retryTime,
5506 int i, cnt, lastCnt = 0;
5507 struct rx_packet **listP, **lastP = 0;
5508 struct rx_peer *peer = call->conn->peer;
5509 int morePackets = 0;
5511 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
5512 /* Does the current packet force us to flush the current list? */
5514 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5515 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5517 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
5519 /* If the call enters an error state stop sending, or if
5520 * we entered congestion recovery mode, stop sending */
5521 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5529 /* Add the current packet to the list if it hasn't been acked.
5530 * Otherwise adjust the list pointer to skip the current packet. */
5531 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5533 /* Do we need to flush the list? */
5534 if (cnt >= (int)peer->maxDgramPackets
5535 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5536 || list[i]->header.serial
5537 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5539 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5540 retryTime, resending);
5541 /* If the call enters an error state stop sending, or if
5542 * we entered congestion recovery mode, stop sending */
5544 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5549 listP = &list[i + 1];
5554 osi_Panic("rxi_SendList error");
5556 listP = &list[i + 1];
5560 /* Send the whole list when the call is in receive mode, when
5561 * the call is in eof mode, when we are in fast recovery mode,
5562 * and when we have the last packet */
5563 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5564 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5565 || (call->flags & RX_CALL_FAST_RECOVER)) {
5566 /* Check for the case where the current list contains
5567 * an acked packet. Since we always send retransmissions
5568 * in a separate packet, we only need to check the first
5569 * packet in the list */
5570 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5574 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5575 retryTime, resending);
5576 /* If the call enters an error state stop sending, or if
5577 * we entered congestion recovery mode, stop sending */
5578 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5582 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5585 } else if (lastCnt > 0) {
5586 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5591 #ifdef RX_ENABLE_LOCKS
5592 /* Call rxi_Start, below, but with the call lock held. */
5594 rxi_StartUnlocked(struct rxevent *event,
5595 void *arg0, void *arg1, int istack)
5597 struct rx_call *call = arg0;
5599 MUTEX_ENTER(&call->lock);
5600 rxi_Start(event, call, arg1, istack);
5601 MUTEX_EXIT(&call->lock);
5603 #endif /* RX_ENABLE_LOCKS */
5605 /* This routine is called when new packets are readied for
5606 * transmission and when retransmission may be necessary, or when the
5607 * transmission window or burst count are favourable. This should be
5608 * better optimized for new packets, the usual case, now that we've
5609 * got rid of queues of send packets. XXXXXXXXXXX */
5611 rxi_Start(struct rxevent *event,
5612 void *arg0, void *arg1, int istack)
5614 struct rx_call *call = arg0;
5616 struct rx_packet *p;
5617 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5618 struct rx_peer *peer = call->conn->peer;
5619 struct clock now, usenow, retryTime;
5625 /* If rxi_Start is being called as a result of a resend event,
5626 * then make sure that the event pointer is removed from the call
5627 * structure, since there is no longer a per-call retransmission
5629 if (event && event == call->resendEvent) {
5630 MUTEX_ENTER(&rx_refcnt_mutex);
5631 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5632 MUTEX_EXIT(&rx_refcnt_mutex);
5633 call->resendEvent = NULL;
5635 if (queue_IsEmpty(&call->tq)) {
5639 /* Timeouts trigger congestion recovery */
5640 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5641 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5642 /* someone else is waiting to start recovery */
5645 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5646 rxi_WaitforTQBusy(call);
5647 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5648 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5649 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5651 if (rx_stats_active)
5652 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5656 call->flags |= RX_CALL_FAST_RECOVER;
5658 if (peer->maxDgramPackets > 1) {
5659 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5661 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5663 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5664 call->nDgramPackets = 1;
5666 call->nextCwind = 1;
5669 MUTEX_ENTER(&peer->peer_lock);
5670 peer->MTU = call->MTU;
5671 peer->cwind = call->cwind;
5672 peer->nDgramPackets = 1;
5674 call->congestSeq = peer->congestSeq;
5675 MUTEX_EXIT(&peer->peer_lock);
5676 /* Clear retry times on packets. Otherwise, it's possible for
5677 * some packets in the queue to force resends at rates faster
5678 * than recovery rates.
5680 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5681 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5682 clock_Zero(&p->retryTime);
5687 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5688 if (rx_stats_active)
5689 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5694 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5695 /* Get clock to compute the re-transmit time for any packets
5696 * in this burst. Note, if we back off, it's reasonable to
5697 * back off all of the packets in the same manner, even if
5698 * some of them have been retransmitted more times than more
5700 * Do a dance to avoid blocking after setting now. */
5701 MUTEX_ENTER(&peer->peer_lock);
5702 retryTime = peer->timeout;
5703 MUTEX_EXIT(&peer->peer_lock);
5705 clock_GetTime(&now);
5706 clock_Add(&retryTime, &now);
5708 /* Send (or resend) any packets that need it, subject to
5709 * window restrictions and congestion burst control
5710 * restrictions. Ask for an ack on the last packet sent in
5711 * this burst. For now, we're relying upon the window being
5712 * considerably bigger than the largest number of packets that
5713 * are typically sent at once by one initial call to
5714 * rxi_Start. This is probably bogus (perhaps we should ask
5715 * for an ack when we're half way through the current
5716 * window?). Also, for non file transfer applications, this
5717 * may end up asking for an ack for every packet. Bogus. XXXX
5720 * But check whether we're here recursively, and let the other guy
5723 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5724 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5725 call->flags |= RX_CALL_TQ_BUSY;
5727 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5729 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5730 call->flags &= ~RX_CALL_NEED_START;
5731 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5733 maxXmitPackets = MIN(call->twind, call->cwind);
5734 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5735 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5736 /* We shouldn't be sending packets if a thread is waiting
5737 * to initiate congestion recovery */
5738 dpf(("call %d waiting to initiate fast recovery\n",
5739 *(call->callNumber)));
5743 && (call->flags & RX_CALL_FAST_RECOVER)) {
5744 /* Only send one packet during fast recovery */
5745 dpf(("call %d restricted to one packet per send during fast recovery\n",
5746 *(call->callNumber)));
5749 #ifdef RX_TRACK_PACKETS
5750 if ((p->flags & RX_PKTFLAG_FREE)
5751 || (!queue_IsEnd(&call->tq, nxp)
5752 && (nxp->flags & RX_PKTFLAG_FREE))
5753 || (p == (struct rx_packet *)&rx_freePacketQueue)
5754 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5755 osi_Panic("rxi_Start: xmit queue clobbered");
5758 if (p->flags & RX_PKTFLAG_ACKED) {
5759 /* Since we may block, don't trust this */
5760 usenow.sec = usenow.usec = 0;
5761 if (rx_stats_active)
5762 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
5763 continue; /* Ignore this packet if it has been acknowledged */
5766 /* Turn off all flags except these ones, which are the same
5767 * on each transmission */
5768 p->header.flags &= RX_PRESET_FLAGS;
5770 if (p->header.seq >=
5771 call->tfirst + MIN((int)call->twind,
5772 (int)(call->nSoftAcked +
5774 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5775 /* Note: if we're waiting for more window space, we can
5776 * still send retransmits; hence we don't return here, but
5777 * break out to schedule a retransmit event */
5778 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
5779 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
5784 /* Transmit the packet if it needs to be sent. */
5785 if (!clock_Lt(&now, &p->retryTime)) {
5786 if (nXmitPackets == maxXmitPackets) {
5787 rxi_SendXmitList(call, call->xmitList,
5788 nXmitPackets, istack, &now,
5789 &retryTime, resending);
5792 dpf(("call %d xmit packet %"AFS_PTR_FMT" now %u.%06u retryTime %u.%06u nextRetry %u.%06u\n",
5793 *(call->callNumber), p,
5795 p->retryTime.sec, p->retryTime.usec,
5796 retryTime.sec, retryTime.usec));
5797 call->xmitList[nXmitPackets++] = p;
5801 /* xmitList now hold pointers to all of the packets that are
5802 * ready to send. Now we loop to send the packets */
5803 if (nXmitPackets > 0) {
5804 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
5805 istack, &now, &retryTime, resending);
5808 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5810 * TQ references no longer protected by this flag; they must remain
5811 * protected by the global lock.
5813 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5814 call->flags &= ~RX_CALL_TQ_BUSY;
5815 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5816 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
5817 call, call->tqWaiters, call->flags));
5818 #ifdef RX_ENABLE_LOCKS
5819 osirx_AssertMine(&call->lock, "rxi_Start start");
5820 CV_BROADCAST(&call->cv_tq);
5821 #else /* RX_ENABLE_LOCKS */
5822 osi_rxWakeup(&call->tq);
5823 #endif /* RX_ENABLE_LOCKS */
5828 /* We went into the error state while sending packets. Now is
5829 * the time to reset the call. This will also inform the using
5830 * process that the call is in an error state.
5832 if (rx_stats_active)
5833 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
5834 call->flags &= ~RX_CALL_TQ_BUSY;
5835 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5836 dpf(("call error %d while xmit %p has %d waiters and flags %d\n",
5837 call->error, call, call->tqWaiters, call->flags));
5838 #ifdef RX_ENABLE_LOCKS
5839 osirx_AssertMine(&call->lock, "rxi_Start middle");
5840 CV_BROADCAST(&call->cv_tq);
5841 #else /* RX_ENABLE_LOCKS */
5842 osi_rxWakeup(&call->tq);
5843 #endif /* RX_ENABLE_LOCKS */
5845 rxi_CallError(call, call->error);
5848 #ifdef RX_ENABLE_LOCKS
5849 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5851 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5852 /* Some packets have received acks. If they all have, we can clear
5853 * the transmit queue.
5856 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5857 if (p->header.seq < call->tfirst
5858 && (p->flags & RX_PKTFLAG_ACKED)) {
5860 #ifdef RX_TRACK_PACKETS
5861 p->flags &= ~RX_PKTFLAG_TQ;
5863 #ifdef RXDEBUG_PACKET
5871 call->flags |= RX_CALL_TQ_CLEARME;
5873 #endif /* RX_ENABLE_LOCKS */
5874 /* Don't bother doing retransmits if the TQ is cleared. */
5875 if (call->flags & RX_CALL_TQ_CLEARME) {
5876 rxi_ClearTransmitQueue(call, 1);
5878 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5881 /* Always post a resend event, if there is anything in the
5882 * queue, and resend is possible. There should be at least
5883 * one unacknowledged packet in the queue ... otherwise none
5884 * of these packets should be on the queue in the first place.
5886 if (call->resendEvent) {
5887 /* Cancel the existing event and post a new one */
5888 rxevent_Cancel(call->resendEvent, call,
5889 RX_CALL_REFCOUNT_RESEND);
5892 /* The retry time is the retry time on the first unacknowledged
5893 * packet inside the current window */
5895 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5896 /* Don't set timers for packets outside the window */
5897 if (p->header.seq >= call->tfirst + call->twind) {
5901 if (!(p->flags & RX_PKTFLAG_ACKED)
5902 && !clock_IsZero(&p->retryTime)) {
5904 retryTime = p->retryTime;
5909 /* Post a new event to re-run rxi_Start when retries may be needed */
5910 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5911 #ifdef RX_ENABLE_LOCKS
5912 MUTEX_ENTER(&rx_refcnt_mutex);
5913 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5914 MUTEX_EXIT(&rx_refcnt_mutex);
5916 rxevent_PostNow2(&retryTime, &usenow,
5918 (void *)call, 0, istack);
5919 #else /* RX_ENABLE_LOCKS */
5921 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5922 (void *)call, 0, istack);
5923 #endif /* RX_ENABLE_LOCKS */
5926 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5927 } while (call->flags & RX_CALL_NEED_START);
5929 * TQ references no longer protected by this flag; they must remain
5930 * protected by the global lock.
5932 call->flags &= ~RX_CALL_TQ_BUSY;
5933 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5934 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
5935 call, call->tqWaiters, call->flags));
5936 #ifdef RX_ENABLE_LOCKS
5937 osirx_AssertMine(&call->lock, "rxi_Start end");
5938 CV_BROADCAST(&call->cv_tq);
5939 #else /* RX_ENABLE_LOCKS */
5940 osi_rxWakeup(&call->tq);
5941 #endif /* RX_ENABLE_LOCKS */
5944 call->flags |= RX_CALL_NEED_START;
5946 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5948 if (call->resendEvent) {
5949 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5954 /* Also adjusts the keep alive parameters for the call, to reflect
5955 * that we have just sent a packet (so keep alives aren't sent
5958 rxi_Send(struct rx_call *call, struct rx_packet *p,
5961 struct rx_connection *conn = call->conn;
5963 /* Stamp each packet with the user supplied status */
5964 p->header.userStatus = call->localStatus;
5966 /* Allow the security object controlling this call's security to
5967 * make any last-minute changes to the packet */
5968 RXS_SendPacket(conn->securityObject, call, p);
5970 /* Since we're about to send SOME sort of packet to the peer, it's
5971 * safe to nuke any scheduled end-of-packets ack */
5972 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5974 /* Actually send the packet, filling in more connection-specific fields */
5975 MUTEX_EXIT(&call->lock);
5976 MUTEX_ENTER(&rx_refcnt_mutex);
5977 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5978 MUTEX_EXIT(&rx_refcnt_mutex);
5979 rxi_SendPacket(call, conn, p, istack);
5980 MUTEX_ENTER(&rx_refcnt_mutex);
5981 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5982 MUTEX_EXIT(&rx_refcnt_mutex);
5983 MUTEX_ENTER(&call->lock);
5985 /* Update last send time for this call (for keep-alive
5986 * processing), and for the connection (so that we can discover
5987 * idle connections) */
5988 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5989 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
5990 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
5992 conn->lastSendTime = call->lastSendTime = clock_Sec();
5993 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
5994 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5995 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
5996 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
5997 RX_ACK_PING_RESPONSE)))
5998 call->lastSendData = call->lastSendTime;
6002 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6003 * that things are fine. Also called periodically to guarantee that nothing
6004 * falls through the cracks (e.g. (error + dally) connections have keepalive
6005 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6007 * haveCTLock Set if calling from rxi_ReapConnections
6009 #ifdef RX_ENABLE_LOCKS
6011 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6012 #else /* RX_ENABLE_LOCKS */
6014 rxi_CheckCall(struct rx_call *call)
6015 #endif /* RX_ENABLE_LOCKS */
6017 struct rx_connection *conn = call->conn;
6019 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6020 afs_uint32 fudgeFactor;
6024 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6025 if (call->flags & RX_CALL_TQ_BUSY) {
6026 /* Call is active and will be reset by rxi_Start if it's
6027 * in an error state.
6032 /* RTT + 8*MDEV, rounded up to the next second. */
6033 fudgeFactor = (((afs_uint32) conn->peer->rtt >> 3) +
6034 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
6036 deadTime = conn->secondsUntilDead + fudgeFactor;
6038 /* These are computed to the second (+- 1 second). But that's
6039 * good enough for these values, which should be a significant
6040 * number of seconds. */
6041 if (now > (call->lastReceiveTime + deadTime)) {
6042 if (call->state == RX_STATE_ACTIVE) {
6044 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
6046 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6047 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6048 ip_stack_t *ipst = ns->netstack_ip;
6050 ire = ire_cache_lookup(conn->peer->host
6051 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6053 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
6055 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6062 if (ire && ire->ire_max_frag > 0)
6063 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6065 #if defined(GLOBAL_NETSTACKID)
6069 #endif /* ADAPT_PMTU */
6070 cerror = RX_CALL_DEAD;
6073 #ifdef RX_ENABLE_LOCKS
6074 /* Cancel pending events */
6075 rxevent_Cancel(call->delayedAckEvent, call,
6076 RX_CALL_REFCOUNT_DELAY);
6077 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
6078 rxevent_Cancel(call->keepAliveEvent, call,
6079 RX_CALL_REFCOUNT_ALIVE);
6080 MUTEX_ENTER(&rx_refcnt_mutex);
6081 if (call->refCount == 0) {
6082 rxi_FreeCall(call, haveCTLock);
6083 MUTEX_EXIT(&rx_refcnt_mutex);
6086 MUTEX_EXIT(&rx_refcnt_mutex);
6088 #else /* RX_ENABLE_LOCKS */
6089 rxi_FreeCall(call, 0);
6091 #endif /* RX_ENABLE_LOCKS */
6093 /* Non-active calls are destroyed if they are not responding
6094 * to pings; active calls are simply flagged in error, so the
6095 * attached process can die reasonably gracefully. */
6098 if (conn->idleDeadTime) {
6099 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6102 /* see if we have a non-activity timeout */
6103 if (call->startWait && idleDeadTime
6104 && ((call->startWait + idleDeadTime) < now) &&
6105 (call->flags & RX_CALL_READER_WAIT)) {
6106 if (call->state == RX_STATE_ACTIVE) {
6107 cerror = RX_CALL_TIMEOUT;
6111 if (call->lastSendData && idleDeadTime && (conn->idleDeadErr != 0)
6112 && ((call->lastSendData + idleDeadTime) < now)) {
6113 if (call->state == RX_STATE_ACTIVE) {
6114 cerror = conn->idleDeadErr;
6120 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6123 /* see if we have a hard timeout */
6125 && (now > (hardDeadTime + call->startTime.sec))) {
6126 if (call->state == RX_STATE_ACTIVE)
6127 rxi_CallError(call, RX_CALL_TIMEOUT);
6132 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT
6133 && call->lastReceiveTime) {
6134 int oldMTU = conn->peer->ifMTU;
6136 /* if we thought we could send more, perhaps things got worse */
6137 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6138 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6139 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6140 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6142 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6144 /* minimum capped in SetPeerMtu */
6145 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6148 conn->lastPacketSize = 0;
6150 /* needed so ResetCall doesn't clobber us. */
6151 call->MTU = conn->peer->ifMTU;
6153 /* if we never succeeded, let the error pass out as-is */
6154 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6155 cerror = conn->msgsizeRetryErr;
6158 rxi_CallError(call, cerror);
6163 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6165 struct rx_connection *conn = arg1;
6166 struct rx_header theader;
6168 struct sockaddr_in taddr;
6171 struct iovec tmpiov[2];
6174 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6177 tp = &tbuffer[sizeof(struct rx_header)];
6178 taddr.sin_family = AF_INET;
6179 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6180 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6181 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6182 taddr.sin_len = sizeof(struct sockaddr_in);
6184 memset(&theader, 0, sizeof(theader));
6185 theader.epoch = htonl(999);
6187 theader.callNumber = 0;
6190 theader.type = RX_PACKET_TYPE_VERSION;
6191 theader.flags = RX_LAST_PACKET;
6192 theader.serviceId = 0;
6194 memcpy(tbuffer, &theader, sizeof(theader));
6195 memcpy(tp, &a, sizeof(a));
6196 tmpiov[0].iov_base = tbuffer;
6197 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6199 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6201 MUTEX_ENTER(&conn->conn_data_lock);
6202 MUTEX_ENTER(&rx_refcnt_mutex);
6203 /* Only reschedule ourselves if the connection would not be destroyed */
6204 if (conn->refCount <= 1) {
6205 conn->natKeepAliveEvent = NULL;
6206 MUTEX_EXIT(&rx_refcnt_mutex);
6207 MUTEX_EXIT(&conn->conn_data_lock);
6208 rx_DestroyConnection(conn); /* drop the reference for this */
6210 conn->refCount--; /* drop the reference for this */
6211 MUTEX_EXIT(&rx_refcnt_mutex);
6212 conn->natKeepAliveEvent = NULL;
6213 rxi_ScheduleNatKeepAliveEvent(conn);
6214 MUTEX_EXIT(&conn->conn_data_lock);
6219 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6221 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6222 struct clock when, now;
6223 clock_GetTime(&now);
6225 when.sec += conn->secondsUntilNatPing;
6226 MUTEX_ENTER(&rx_refcnt_mutex);
6227 conn->refCount++; /* hold a reference for this */
6228 MUTEX_EXIT(&rx_refcnt_mutex);
6229 conn->natKeepAliveEvent =
6230 rxevent_PostNow(&when, &now, rxi_NatKeepAliveEvent, conn, 0);
6235 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6237 MUTEX_ENTER(&conn->conn_data_lock);
6238 conn->secondsUntilNatPing = seconds;
6240 rxi_ScheduleNatKeepAliveEvent(conn);
6241 MUTEX_EXIT(&conn->conn_data_lock);
6245 rxi_NatKeepAliveOn(struct rx_connection *conn)
6247 MUTEX_ENTER(&conn->conn_data_lock);
6248 rxi_ScheduleNatKeepAliveEvent(conn);
6249 MUTEX_EXIT(&conn->conn_data_lock);
6252 /* When a call is in progress, this routine is called occasionally to
6253 * make sure that some traffic has arrived (or been sent to) the peer.
6254 * If nothing has arrived in a reasonable amount of time, the call is
6255 * declared dead; if nothing has been sent for a while, we send a
6256 * keep-alive packet (if we're actually trying to keep the call alive)
6259 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6261 struct rx_call *call = arg1;
6262 struct rx_connection *conn;
6265 MUTEX_ENTER(&rx_refcnt_mutex);
6266 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6267 MUTEX_EXIT(&rx_refcnt_mutex);
6268 MUTEX_ENTER(&call->lock);
6269 if (event == call->keepAliveEvent)
6270 call->keepAliveEvent = NULL;
6273 #ifdef RX_ENABLE_LOCKS
6274 if (rxi_CheckCall(call, 0)) {
6275 MUTEX_EXIT(&call->lock);
6278 #else /* RX_ENABLE_LOCKS */
6279 if (rxi_CheckCall(call))
6281 #endif /* RX_ENABLE_LOCKS */
6283 /* Don't try to keep alive dallying calls */
6284 if (call->state == RX_STATE_DALLY) {
6285 MUTEX_EXIT(&call->lock);
6290 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6291 /* Don't try to send keepalives if there is unacknowledged data */
6292 /* the rexmit code should be good enough, this little hack
6293 * doesn't quite work XXX */
6294 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6296 rxi_ScheduleKeepAliveEvent(call);
6297 MUTEX_EXIT(&call->lock);
6300 /* Does what's on the nameplate. */
6302 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy)
6304 struct rx_call *call = arg1;
6305 struct rx_connection *conn;
6307 MUTEX_ENTER(&rx_refcnt_mutex);
6308 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6309 MUTEX_EXIT(&rx_refcnt_mutex);
6310 MUTEX_ENTER(&call->lock);
6312 if (event == call->growMTUEvent)
6313 call->growMTUEvent = NULL;
6315 #ifdef RX_ENABLE_LOCKS
6316 if (rxi_CheckCall(call, 0)) {
6317 MUTEX_EXIT(&call->lock);
6320 #else /* RX_ENABLE_LOCKS */
6321 if (rxi_CheckCall(call))
6323 #endif /* RX_ENABLE_LOCKS */
6325 /* Don't bother with dallying calls */
6326 if (call->state == RX_STATE_DALLY) {
6327 MUTEX_EXIT(&call->lock);
6334 * keep being scheduled, just don't do anything if we're at peak,
6335 * or we're not set up to be properly handled (idle timeout required)
6337 if ((conn->peer->maxPacketSize != 0) &&
6338 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6339 (conn->idleDeadErr))
6340 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6341 rxi_ScheduleGrowMTUEvent(call, 0);
6342 MUTEX_EXIT(&call->lock);
6346 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6348 if (!call->keepAliveEvent) {
6349 struct clock when, now;
6350 clock_GetTime(&now);
6352 when.sec += call->conn->secondsUntilPing;
6353 MUTEX_ENTER(&rx_refcnt_mutex);
6354 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6355 MUTEX_EXIT(&rx_refcnt_mutex);
6356 call->keepAliveEvent =
6357 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
6362 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6364 if (!call->growMTUEvent) {
6365 struct clock when, now;
6367 clock_GetTime(&now);
6370 if (call->conn->secondsUntilPing)
6371 secs = (6*call->conn->secondsUntilPing)-1;
6373 if (call->conn->secondsUntilDead)
6374 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6378 MUTEX_ENTER(&rx_refcnt_mutex);
6379 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6380 MUTEX_EXIT(&rx_refcnt_mutex);
6381 call->growMTUEvent =
6382 rxevent_PostNow(&when, &now, rxi_GrowMTUEvent, call, 0);
6386 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6388 rxi_KeepAliveOn(struct rx_call *call)
6390 /* Pretend last packet received was received now--i.e. if another
6391 * packet isn't received within the keep alive time, then the call
6392 * will die; Initialize last send time to the current time--even
6393 * if a packet hasn't been sent yet. This will guarantee that a
6394 * keep-alive is sent within the ping time */
6395 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6396 rxi_ScheduleKeepAliveEvent(call);
6400 rxi_GrowMTUOn(struct rx_call *call)
6402 struct rx_connection *conn = call->conn;
6403 MUTEX_ENTER(&conn->conn_data_lock);
6404 conn->lastPingSizeSer = conn->lastPingSize = 0;
6405 MUTEX_EXIT(&conn->conn_data_lock);
6406 rxi_ScheduleGrowMTUEvent(call, 1);
6409 /* This routine is called to send connection abort messages
6410 * that have been delayed to throttle looping clients. */
6412 rxi_SendDelayedConnAbort(struct rxevent *event,
6413 void *arg1, void *unused)
6415 struct rx_connection *conn = arg1;
6418 struct rx_packet *packet;
6420 MUTEX_ENTER(&conn->conn_data_lock);
6421 conn->delayedAbortEvent = NULL;
6422 error = htonl(conn->error);
6424 MUTEX_EXIT(&conn->conn_data_lock);
6425 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6428 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6429 RX_PACKET_TYPE_ABORT, (char *)&error,
6431 rxi_FreePacket(packet);
6435 /* This routine is called to send call abort messages
6436 * that have been delayed to throttle looping clients. */
6438 rxi_SendDelayedCallAbort(struct rxevent *event,
6439 void *arg1, void *dummy)
6441 struct rx_call *call = arg1;
6444 struct rx_packet *packet;
6446 MUTEX_ENTER(&call->lock);
6447 call->delayedAbortEvent = NULL;
6448 error = htonl(call->error);
6450 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6453 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6454 (char *)&error, sizeof(error), 0);
6455 rxi_FreePacket(packet);
6457 MUTEX_EXIT(&call->lock);
6458 MUTEX_ENTER(&rx_refcnt_mutex);
6459 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6460 MUTEX_EXIT(&rx_refcnt_mutex);
6463 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6464 * seconds) to ask the client to authenticate itself. The routine
6465 * issues a challenge to the client, which is obtained from the
6466 * security object associated with the connection */
6468 rxi_ChallengeEvent(struct rxevent *event,
6469 void *arg0, void *arg1, int tries)
6471 struct rx_connection *conn = arg0;
6473 conn->challengeEvent = NULL;
6474 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6475 struct rx_packet *packet;
6476 struct clock when, now;
6479 /* We've failed to authenticate for too long.
6480 * Reset any calls waiting for authentication;
6481 * they are all in RX_STATE_PRECALL.
6485 MUTEX_ENTER(&conn->conn_call_lock);
6486 for (i = 0; i < RX_MAXCALLS; i++) {
6487 struct rx_call *call = conn->call[i];
6489 MUTEX_ENTER(&call->lock);
6490 if (call->state == RX_STATE_PRECALL) {
6491 rxi_CallError(call, RX_CALL_DEAD);
6492 rxi_SendCallAbort(call, NULL, 0, 0);
6494 MUTEX_EXIT(&call->lock);
6497 MUTEX_EXIT(&conn->conn_call_lock);
6501 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6503 /* If there's no packet available, do this later. */
6504 RXS_GetChallenge(conn->securityObject, conn, packet);
6505 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6506 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6507 rxi_FreePacket(packet);
6509 clock_GetTime(&now);
6511 when.sec += RX_CHALLENGE_TIMEOUT;
6512 conn->challengeEvent =
6513 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
6518 /* Call this routine to start requesting the client to authenticate
6519 * itself. This will continue until authentication is established,
6520 * the call times out, or an invalid response is returned. The
6521 * security object associated with the connection is asked to create
6522 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6523 * defined earlier. */
6525 rxi_ChallengeOn(struct rx_connection *conn)
6527 if (!conn->challengeEvent) {
6528 RXS_CreateChallenge(conn->securityObject, conn);
6529 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6534 /* Compute round trip time of the packet provided, in *rttp.
6537 /* rxi_ComputeRoundTripTime is called with peer locked. */
6538 /* sentp and/or peer may be null */
6540 rxi_ComputeRoundTripTime(struct rx_packet *p,
6541 struct clock *sentp,
6542 struct rx_peer *peer,
6545 struct clock thisRtt, *rttp = &thisRtt;
6550 if (clock_Lt(rttp, sentp))
6551 return; /* somebody set the clock back, don't count this time. */
6553 clock_Sub(rttp, sentp);
6554 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6555 p->header.callNumber, p, rttp->sec, rttp->usec));
6557 if (rttp->sec == 0 && rttp->usec == 0) {
6559 * The actual round trip time is shorter than the
6560 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6561 * Since we can't tell which at the moment we will assume 1ms.
6566 if (rx_stats_active) {
6567 MUTEX_ENTER(&rx_stats_mutex);
6568 if (clock_Lt(rttp, &rx_stats.minRtt))
6569 rx_stats.minRtt = *rttp;
6570 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
6571 if (rttp->sec > 60) {
6572 MUTEX_EXIT(&rx_stats_mutex);
6573 return; /* somebody set the clock ahead */
6575 rx_stats.maxRtt = *rttp;
6577 clock_Add(&rx_stats.totalRtt, rttp);
6578 rx_atomic_inc(&rx_stats.nRttSamples);
6579 MUTEX_EXIT(&rx_stats_mutex);
6582 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6584 /* Apply VanJacobson round-trip estimations */
6589 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
6590 * srtt is stored as fixed point with 3 bits after the binary
6591 * point (i.e., scaled by 8). The following magic is
6592 * equivalent to the smoothing algorithm in rfc793 with an
6593 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6594 * srtt'*8 = rtt + srtt*7
6595 * srtt'*8 = srtt*8 + rtt - srtt
6596 * srtt' = srtt + rtt/8 - srtt/8
6597 * srtt' = srtt + (rtt - srtt)/8
6600 delta = _8THMSEC(rttp) - peer->rtt;
6601 peer->rtt += (delta >> 3);
6604 * We accumulate a smoothed rtt variance (actually, a smoothed
6605 * mean difference), then set the retransmit timer to smoothed
6606 * rtt + 4 times the smoothed variance (was 2x in van's original
6607 * paper, but 4x works better for me, and apparently for him as
6609 * rttvar is stored as
6610 * fixed point with 2 bits after the binary point (scaled by
6611 * 4). The following is equivalent to rfc793 smoothing with
6612 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6613 * rttvar'*4 = rttvar*3 + |delta|
6614 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6615 * rttvar' = rttvar + |delta|/4 - rttvar/4
6616 * rttvar' = rttvar + (|delta| - rttvar)/4
6617 * This replaces rfc793's wired-in beta.
6618 * dev*4 = dev*4 + (|actual - expected| - dev)
6624 delta -= (peer->rtt_dev << 1);
6625 peer->rtt_dev += (delta >> 3);
6627 /* I don't have a stored RTT so I start with this value. Since I'm
6628 * probably just starting a call, and will be pushing more data down
6629 * this, I expect congestion to increase rapidly. So I fudge a
6630 * little, and I set deviance to half the rtt. In practice,
6631 * deviance tends to approach something a little less than
6632 * half the smoothed rtt. */
6633 peer->rtt = _8THMSEC(rttp) + 8;
6634 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
6636 /* the timeout is RTT + 4*MDEV but no less than rx_minPeerTimeout msec.
6637 * This is because one end or the other of these connections is usually
6638 * in a user process, and can be switched and/or swapped out. So on fast,
6639 * reliable networks, the timeout would otherwise be too short. */
6640 rtt_timeout = MAX(((peer->rtt >> 3) + peer->rtt_dev), rx_minPeerTimeout);
6641 clock_Zero(&(peer->timeout));
6642 clock_Addmsec(&(peer->timeout), rtt_timeout);
6644 /* Reset the backedOff flag since we just computed a new timeout value */
6645 peer->backedOff = 0;
6647 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6648 p->header.callNumber, p, MSEC(rttp), peer->rtt >> 3, peer->rtt_dev >> 2, (peer->timeout.sec), (peer->timeout.usec)));
6652 /* Find all server connections that have not been active for a long time, and
6655 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6657 struct clock now, when;
6658 clock_GetTime(&now);
6660 /* Find server connection structures that haven't been used for
6661 * greater than rx_idleConnectionTime */
6663 struct rx_connection **conn_ptr, **conn_end;
6664 int i, havecalls = 0;
6665 MUTEX_ENTER(&rx_connHashTable_lock);
6666 for (conn_ptr = &rx_connHashTable[0], conn_end =
6667 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6669 struct rx_connection *conn, *next;
6670 struct rx_call *call;
6674 for (conn = *conn_ptr; conn; conn = next) {
6675 /* XXX -- Shouldn't the connection be locked? */
6678 for (i = 0; i < RX_MAXCALLS; i++) {
6679 call = conn->call[i];
6683 code = MUTEX_TRYENTER(&call->lock);
6686 #ifdef RX_ENABLE_LOCKS
6687 result = rxi_CheckCall(call, 1);
6688 #else /* RX_ENABLE_LOCKS */
6689 result = rxi_CheckCall(call);
6690 #endif /* RX_ENABLE_LOCKS */
6691 MUTEX_EXIT(&call->lock);
6693 /* If CheckCall freed the call, it might
6694 * have destroyed the connection as well,
6695 * which screws up the linked lists.
6701 if (conn->type == RX_SERVER_CONNECTION) {
6702 /* This only actually destroys the connection if
6703 * there are no outstanding calls */
6704 MUTEX_ENTER(&conn->conn_data_lock);
6705 MUTEX_ENTER(&rx_refcnt_mutex);
6706 if (!havecalls && !conn->refCount
6707 && ((conn->lastSendTime + rx_idleConnectionTime) <
6709 conn->refCount++; /* it will be decr in rx_DestroyConn */
6710 MUTEX_EXIT(&rx_refcnt_mutex);
6711 MUTEX_EXIT(&conn->conn_data_lock);
6712 #ifdef RX_ENABLE_LOCKS
6713 rxi_DestroyConnectionNoLock(conn);
6714 #else /* RX_ENABLE_LOCKS */
6715 rxi_DestroyConnection(conn);
6716 #endif /* RX_ENABLE_LOCKS */
6718 #ifdef RX_ENABLE_LOCKS
6720 MUTEX_EXIT(&rx_refcnt_mutex);
6721 MUTEX_EXIT(&conn->conn_data_lock);
6723 #endif /* RX_ENABLE_LOCKS */
6727 #ifdef RX_ENABLE_LOCKS
6728 while (rx_connCleanup_list) {
6729 struct rx_connection *conn;
6730 conn = rx_connCleanup_list;
6731 rx_connCleanup_list = rx_connCleanup_list->next;
6732 MUTEX_EXIT(&rx_connHashTable_lock);
6733 rxi_CleanupConnection(conn);
6734 MUTEX_ENTER(&rx_connHashTable_lock);
6736 MUTEX_EXIT(&rx_connHashTable_lock);
6737 #endif /* RX_ENABLE_LOCKS */
6740 /* Find any peer structures that haven't been used (haven't had an
6741 * associated connection) for greater than rx_idlePeerTime */
6743 struct rx_peer **peer_ptr, **peer_end;
6747 * Why do we need to hold the rx_peerHashTable_lock across
6748 * the incrementing of peer_ptr since the rx_peerHashTable
6749 * array is not changing? We don't.
6751 * By dropping the lock periodically we can permit other
6752 * activities to be performed while a rxi_ReapConnections
6753 * call is in progress. The goal of reap connections
6754 * is to clean up quickly without causing large amounts
6755 * of contention. Therefore, it is important that global
6756 * mutexes not be held for extended periods of time.
6758 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6759 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6761 struct rx_peer *peer, *next, *prev;
6763 MUTEX_ENTER(&rx_peerHashTable_lock);
6764 for (prev = peer = *peer_ptr; peer; peer = next) {
6766 code = MUTEX_TRYENTER(&peer->peer_lock);
6767 if ((code) && (peer->refCount == 0)
6768 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6769 rx_interface_stat_p rpc_stat, nrpc_stat;
6773 * now know that this peer object is one to be
6774 * removed from the hash table. Once it is removed
6775 * it can't be referenced by other threads.
6776 * Lets remove it first and decrement the struct
6777 * nPeerStructs count.
6779 if (peer == *peer_ptr) {
6785 if (rx_stats_active)
6786 rx_atomic_dec(&rx_stats.nPeerStructs);
6789 * Now if we hold references on 'prev' and 'next'
6790 * we can safely drop the rx_peerHashTable_lock
6791 * while we destroy this 'peer' object.
6797 MUTEX_EXIT(&rx_peerHashTable_lock);
6799 MUTEX_EXIT(&peer->peer_lock);
6800 MUTEX_DESTROY(&peer->peer_lock);
6802 (&peer->rpcStats, rpc_stat, nrpc_stat,
6803 rx_interface_stat)) {
6804 unsigned int num_funcs;
6807 queue_Remove(&rpc_stat->queue_header);
6808 queue_Remove(&rpc_stat->all_peers);
6809 num_funcs = rpc_stat->stats[0].func_total;
6811 sizeof(rx_interface_stat_t) +
6812 rpc_stat->stats[0].func_total *
6813 sizeof(rx_function_entry_v1_t);
6815 rxi_Free(rpc_stat, space);
6817 MUTEX_ENTER(&rx_rpc_stats);
6818 rxi_rpc_peer_stat_cnt -= num_funcs;
6819 MUTEX_EXIT(&rx_rpc_stats);
6824 * Regain the rx_peerHashTable_lock and
6825 * decrement the reference count on 'prev'
6828 MUTEX_ENTER(&rx_peerHashTable_lock);
6835 MUTEX_EXIT(&peer->peer_lock);
6840 MUTEX_EXIT(&rx_peerHashTable_lock);
6844 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6845 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6846 * GC, just below. Really, we shouldn't have to keep moving packets from
6847 * one place to another, but instead ought to always know if we can
6848 * afford to hold onto a packet in its particular use. */
6849 MUTEX_ENTER(&rx_freePktQ_lock);
6850 if (rx_waitingForPackets) {
6851 rx_waitingForPackets = 0;
6852 #ifdef RX_ENABLE_LOCKS
6853 CV_BROADCAST(&rx_waitingForPackets_cv);
6855 osi_rxWakeup(&rx_waitingForPackets);
6858 MUTEX_EXIT(&rx_freePktQ_lock);
6861 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6862 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6866 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6867 * rx.h is sort of strange this is better. This is called with a security
6868 * object before it is discarded. Each connection using a security object has
6869 * its own refcount to the object so it won't actually be freed until the last
6870 * connection is destroyed.
6872 * This is the only rxs module call. A hold could also be written but no one
6876 rxs_Release(struct rx_securityClass *aobj)
6878 return RXS_Close(aobj);
6882 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6883 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6884 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6885 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6887 /* Adjust our estimate of the transmission rate to this peer, given
6888 * that the packet p was just acked. We can adjust peer->timeout and
6889 * call->twind. Pragmatically, this is called
6890 * only with packets of maximal length.
6891 * Called with peer and call locked.
6895 rxi_ComputeRate(struct rx_peer *peer, struct rx_call *call,
6896 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6898 afs_int32 xferSize, xferMs;
6902 /* Count down packets */
6903 if (peer->rateFlag > 0)
6905 /* Do nothing until we're enabled */
6906 if (peer->rateFlag != 0)
6911 /* Count only when the ack seems legitimate */
6912 switch (ackReason) {
6913 case RX_ACK_REQUESTED:
6915 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6919 case RX_ACK_PING_RESPONSE:
6920 if (p) /* want the response to ping-request, not data send */
6922 clock_GetTime(&newTO);
6923 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6924 clock_Sub(&newTO, &call->pingRequestTime);
6925 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6929 xferSize = rx_AckDataSize(rx_maxSendWindow) + RX_HEADER_SIZE;
6936 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %d.%06d, rtt %u, ps %u)\n",
6937 ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"),
6938 xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
6940 /* Track only packets that are big enough. */
6941 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6945 /* absorb RTT data (in milliseconds) for these big packets */
6946 if (peer->smRtt == 0) {
6947 peer->smRtt = xferMs;
6949 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6954 if (peer->countDown) {
6958 peer->countDown = 10; /* recalculate only every so often */
6960 /* In practice, we can measure only the RTT for full packets,
6961 * because of the way Rx acks the data that it receives. (If it's
6962 * smaller than a full packet, it often gets implicitly acked
6963 * either by the call response (from a server) or by the next call
6964 * (from a client), and either case confuses transmission times
6965 * with processing times.) Therefore, replace the above
6966 * more-sophisticated processing with a simpler version, where the
6967 * smoothed RTT is kept for full-size packets, and the time to
6968 * transmit a windowful of full-size packets is simply RTT *
6969 * windowSize. Again, we take two steps:
6970 - ensure the timeout is large enough for a single packet's RTT;
6971 - ensure that the window is small enough to fit in the desired timeout.*/
6973 /* First, the timeout check. */
6974 minTime = peer->smRtt;
6975 /* Get a reasonable estimate for a timeout period */
6977 newTO.sec = minTime / 1000;
6978 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6980 /* Increase the timeout period so that we can always do at least
6981 * one packet exchange */
6982 if (clock_Gt(&newTO, &peer->timeout)) {
6984 dpf(("CONG peer %lx/%u: timeout %d.%06d ==> %ld.%06d (rtt %u)\n",
6985 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec,
6986 newTO.sec, newTO.usec, peer->smRtt));
6988 peer->timeout = newTO;
6991 /* Now, get an estimate for the transmit window size. */
6992 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6993 /* Now, convert to the number of full packets that could fit in a
6994 * reasonable fraction of that interval */
6995 minTime /= (peer->smRtt << 1);
6996 minTime = MAX(minTime, rx_minPeerTimeout);
6997 xferSize = minTime; /* (make a copy) */
6999 /* Now clamp the size to reasonable bounds. */
7002 else if (minTime > rx_maxSendWindow)
7003 minTime = rx_maxSendWindow;
7004 /* if (minTime != peer->maxWindow) {
7005 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u)\n",
7006 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
7007 peer->timeout.sec, peer->timeout.usec, peer->smRtt));
7008 peer->maxWindow = minTime;
7009 elide... call->twind = minTime;
7013 /* Cut back on the peer timeout if it had earlier grown unreasonably.
7014 * Discern this by calculating the timeout necessary for rx_Window
7016 if ((xferSize > rx_maxSendWindow) && (peer->timeout.sec >= 3)) {
7017 /* calculate estimate for transmission interval in milliseconds */
7018 minTime = rx_maxSendWindow * peer->smRtt;
7019 if (minTime < 1000) {
7020 dpf(("CONG peer %lx/%u: cut TO %d.%06d by 0.5 (rtt %u)\n",
7021 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
7022 peer->timeout.usec, peer->smRtt));
7024 newTO.sec = 0; /* cut back on timeout by half a second */
7025 newTO.usec = 500000;
7026 clock_Sub(&peer->timeout, &newTO);
7031 } /* end of rxi_ComputeRate */
7032 #endif /* ADAPT_WINDOW */
7040 #define TRACE_OPTION_RX_DEBUG 16
7048 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7049 0, KEY_QUERY_VALUE, &parmKey);
7050 if (code != ERROR_SUCCESS)
7053 dummyLen = sizeof(TraceOption);
7054 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7055 (BYTE *) &TraceOption, &dummyLen);
7056 if (code == ERROR_SUCCESS) {
7057 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7059 RegCloseKey (parmKey);
7060 #endif /* AFS_NT40_ENV */
7065 rx_DebugOnOff(int on)
7069 rxdebug_active = on;
7075 rx_StatsOnOff(int on)
7078 rx_stats_active = on;
7083 /* Don't call this debugging routine directly; use dpf */
7085 rxi_DebugPrint(char *format, ...)
7094 va_start(ap, format);
7096 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7099 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7101 OutputDebugString(msg);
7107 va_start(ap, format);
7109 clock_GetTime(&now);
7110 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7111 (unsigned int)now.usec);
7112 vfprintf(rx_Log, format, ap);
7120 * This function is used to process the rx_stats structure that is local
7121 * to a process as well as an rx_stats structure received from a remote
7122 * process (via rxdebug). Therefore, it needs to do minimal version
7126 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7127 afs_int32 freePackets, char version)
7131 if (size != sizeof(struct rx_statistics)) {
7133 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7134 size, sizeof(struct rx_statistics));
7137 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7140 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7141 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7142 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7143 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7144 s->specialPktAllocFailures);
7146 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7147 s->receivePktAllocFailures, s->sendPktAllocFailures,
7148 s->specialPktAllocFailures);
7152 " greedy %u, " "bogusReads %u (last from host %x), "
7153 "noPackets %u, " "noBuffers %u, " "selects %u, "
7154 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7155 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7156 s->selects, s->sendSelects);
7158 fprintf(file, " packets read: ");
7159 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7160 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7162 fprintf(file, "\n");
7165 " other read counters: data %u, " "ack %u, " "dup %u "
7166 "spurious %u " "dally %u\n", s->dataPacketsRead,
7167 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7168 s->ignorePacketDally);
7170 fprintf(file, " packets sent: ");
7171 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7172 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7174 fprintf(file, "\n");
7177 " other send counters: ack %u, " "data %u (not resends), "
7178 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7179 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7180 s->dataPacketsPushed, s->ignoreAckedPacket);
7183 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7184 s->netSendFailures, (int)s->fatalErrors);
7186 if (s->nRttSamples) {
7187 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7188 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7190 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7191 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7195 " %d server connections, " "%d client connections, "
7196 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7197 s->nServerConns, s->nClientConns, s->nPeerStructs,
7198 s->nCallStructs, s->nFreeCallStructs);
7200 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7201 fprintf(file, " %d clock updates\n", clock_nUpdates);
7205 /* for backward compatibility */
7207 rx_PrintStats(FILE * file)
7209 MUTEX_ENTER(&rx_stats_mutex);
7210 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7211 sizeof(rx_stats), rx_nFreePackets,
7213 MUTEX_EXIT(&rx_stats_mutex);
7217 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7219 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
7220 ntohl(peer->host), (int)ntohs(peer->port), (int)peer->burstSize,
7221 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
7224 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
7225 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
7226 (int)peer->timeout.usec, peer->nSent, peer->reSends);
7229 " Packet size %d, " "max in packet skew %d, "
7230 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
7231 (int)peer->outPacketSkew);
7235 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7237 * This mutex protects the following static variables:
7241 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7242 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7244 #define LOCK_RX_DEBUG
7245 #define UNLOCK_RX_DEBUG
7246 #endif /* AFS_PTHREAD_ENV */
7248 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7250 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7251 u_char type, void *inputData, size_t inputLength,
7252 void *outputData, size_t outputLength)
7254 static afs_int32 counter = 100;
7255 time_t waitTime, waitCount;
7256 struct rx_header theader;
7259 struct timeval tv_now, tv_wake, tv_delta;
7260 struct sockaddr_in taddr, faddr;
7274 tp = &tbuffer[sizeof(struct rx_header)];
7275 taddr.sin_family = AF_INET;
7276 taddr.sin_port = remotePort;
7277 taddr.sin_addr.s_addr = remoteAddr;
7278 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7279 taddr.sin_len = sizeof(struct sockaddr_in);
7282 memset(&theader, 0, sizeof(theader));
7283 theader.epoch = htonl(999);
7285 theader.callNumber = htonl(counter);
7288 theader.type = type;
7289 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7290 theader.serviceId = 0;
7292 memcpy(tbuffer, &theader, sizeof(theader));
7293 memcpy(tp, inputData, inputLength);
7295 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7296 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7298 /* see if there's a packet available */
7299 gettimeofday(&tv_wake,0);
7300 tv_wake.tv_sec += waitTime;
7303 FD_SET(socket, &imask);
7304 tv_delta.tv_sec = tv_wake.tv_sec;
7305 tv_delta.tv_usec = tv_wake.tv_usec;
7306 gettimeofday(&tv_now, 0);
7308 if (tv_delta.tv_usec < tv_now.tv_usec) {
7310 tv_delta.tv_usec += 1000000;
7313 tv_delta.tv_usec -= tv_now.tv_usec;
7315 if (tv_delta.tv_sec < tv_now.tv_sec) {
7319 tv_delta.tv_sec -= tv_now.tv_sec;
7322 code = select(0, &imask, 0, 0, &tv_delta);
7323 #else /* AFS_NT40_ENV */
7324 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7325 #endif /* AFS_NT40_ENV */
7326 if (code == 1 && FD_ISSET(socket, &imask)) {
7327 /* now receive a packet */
7328 faddrLen = sizeof(struct sockaddr_in);
7330 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7331 (struct sockaddr *)&faddr, &faddrLen);
7334 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7335 if (counter == ntohl(theader.callNumber))
7343 /* see if we've timed out */
7351 code -= sizeof(struct rx_header);
7352 if (code > outputLength)
7353 code = outputLength;
7354 memcpy(outputData, tp, code);
7357 #endif /* RXDEBUG */
7360 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7361 afs_uint16 remotePort, struct rx_debugStats * stat,
7362 afs_uint32 * supportedValues)
7364 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7366 struct rx_debugIn in;
7368 *supportedValues = 0;
7369 in.type = htonl(RX_DEBUGI_GETSTATS);
7372 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7373 &in, sizeof(in), stat, sizeof(*stat));
7376 * If the call was successful, fixup the version and indicate
7377 * what contents of the stat structure are valid.
7378 * Also do net to host conversion of fields here.
7382 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7383 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7385 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7386 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7388 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7389 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7391 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7392 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7394 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7395 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7397 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7398 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7400 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7401 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7403 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7404 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7406 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7407 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7409 stat->nFreePackets = ntohl(stat->nFreePackets);
7410 stat->packetReclaims = ntohl(stat->packetReclaims);
7411 stat->callsExecuted = ntohl(stat->callsExecuted);
7412 stat->nWaiting = ntohl(stat->nWaiting);
7413 stat->idleThreads = ntohl(stat->idleThreads);
7414 stat->nWaited = ntohl(stat->nWaited);
7415 stat->nPackets = ntohl(stat->nPackets);
7424 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7425 afs_uint16 remotePort, struct rx_statistics * stat,
7426 afs_uint32 * supportedValues)
7428 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7430 struct rx_debugIn in;
7431 afs_int32 *lp = (afs_int32 *) stat;
7435 * supportedValues is currently unused, but added to allow future
7436 * versioning of this function.
7439 *supportedValues = 0;
7440 in.type = htonl(RX_DEBUGI_RXSTATS);
7442 memset(stat, 0, sizeof(*stat));
7444 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7445 &in, sizeof(in), stat, sizeof(*stat));
7450 * Do net to host conversion here
7453 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7464 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7465 afs_uint16 remotePort, size_t version_length,
7468 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7470 return MakeDebugCall(socket, remoteAddr, remotePort,
7471 RX_PACKET_TYPE_VERSION, a, 1, version,
7479 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7480 afs_uint16 remotePort, afs_int32 * nextConnection,
7481 int allConnections, afs_uint32 debugSupportedValues,
7482 struct rx_debugConn * conn,
7483 afs_uint32 * supportedValues)
7485 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7487 struct rx_debugIn in;
7491 * supportedValues is currently unused, but added to allow future
7492 * versioning of this function.
7495 *supportedValues = 0;
7496 if (allConnections) {
7497 in.type = htonl(RX_DEBUGI_GETALLCONN);
7499 in.type = htonl(RX_DEBUGI_GETCONN);
7501 in.index = htonl(*nextConnection);
7502 memset(conn, 0, sizeof(*conn));
7504 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7505 &in, sizeof(in), conn, sizeof(*conn));
7508 *nextConnection += 1;
7511 * Convert old connection format to new structure.
7514 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7515 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7516 #define MOVEvL(a) (conn->a = vL->a)
7518 /* any old or unrecognized version... */
7519 for (i = 0; i < RX_MAXCALLS; i++) {
7520 MOVEvL(callState[i]);
7521 MOVEvL(callMode[i]);
7522 MOVEvL(callFlags[i]);
7523 MOVEvL(callOther[i]);
7525 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7526 MOVEvL(secStats.type);
7527 MOVEvL(secStats.level);
7528 MOVEvL(secStats.flags);
7529 MOVEvL(secStats.expires);
7530 MOVEvL(secStats.packetsReceived);
7531 MOVEvL(secStats.packetsSent);
7532 MOVEvL(secStats.bytesReceived);
7533 MOVEvL(secStats.bytesSent);
7538 * Do net to host conversion here
7540 * I don't convert host or port since we are most likely
7541 * going to want these in NBO.
7543 conn->cid = ntohl(conn->cid);
7544 conn->serial = ntohl(conn->serial);
7545 for (i = 0; i < RX_MAXCALLS; i++) {
7546 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7548 conn->error = ntohl(conn->error);
7549 conn->secStats.flags = ntohl(conn->secStats.flags);
7550 conn->secStats.expires = ntohl(conn->secStats.expires);
7551 conn->secStats.packetsReceived =
7552 ntohl(conn->secStats.packetsReceived);
7553 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7554 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7555 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7556 conn->epoch = ntohl(conn->epoch);
7557 conn->natMTU = ntohl(conn->natMTU);
7566 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7567 afs_uint16 remotePort, afs_int32 * nextPeer,
7568 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7569 afs_uint32 * supportedValues)
7571 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7573 struct rx_debugIn in;
7576 * supportedValues is currently unused, but added to allow future
7577 * versioning of this function.
7580 *supportedValues = 0;
7581 in.type = htonl(RX_DEBUGI_GETPEER);
7582 in.index = htonl(*nextPeer);
7583 memset(peer, 0, sizeof(*peer));
7585 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7586 &in, sizeof(in), peer, sizeof(*peer));
7592 * Do net to host conversion here
7594 * I don't convert host or port since we are most likely
7595 * going to want these in NBO.
7597 peer->ifMTU = ntohs(peer->ifMTU);
7598 peer->idleWhen = ntohl(peer->idleWhen);
7599 peer->refCount = ntohs(peer->refCount);
7600 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7601 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7602 peer->rtt = ntohl(peer->rtt);
7603 peer->rtt_dev = ntohl(peer->rtt_dev);
7604 peer->timeout.sec = ntohl(peer->timeout.sec);
7605 peer->timeout.usec = ntohl(peer->timeout.usec);
7606 peer->nSent = ntohl(peer->nSent);
7607 peer->reSends = ntohl(peer->reSends);
7608 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7609 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7610 peer->rateFlag = ntohl(peer->rateFlag);
7611 peer->natMTU = ntohs(peer->natMTU);
7612 peer->maxMTU = ntohs(peer->maxMTU);
7613 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7614 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7615 peer->MTU = ntohs(peer->MTU);
7616 peer->cwind = ntohs(peer->cwind);
7617 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7618 peer->congestSeq = ntohs(peer->congestSeq);
7619 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7620 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7621 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7622 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7631 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7632 struct rx_debugPeer * peerStats)
7635 afs_int32 error = 1; /* default to "did not succeed" */
7636 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7638 MUTEX_ENTER(&rx_peerHashTable_lock);
7639 for(tp = rx_peerHashTable[hashValue];
7640 tp != NULL; tp = tp->next) {
7641 if (tp->host == peerHost)
7647 MUTEX_EXIT(&rx_peerHashTable_lock);
7651 MUTEX_ENTER(&tp->peer_lock);
7652 peerStats->host = tp->host;
7653 peerStats->port = tp->port;
7654 peerStats->ifMTU = tp->ifMTU;
7655 peerStats->idleWhen = tp->idleWhen;
7656 peerStats->refCount = tp->refCount;
7657 peerStats->burstSize = tp->burstSize;
7658 peerStats->burst = tp->burst;
7659 peerStats->burstWait.sec = tp->burstWait.sec;
7660 peerStats->burstWait.usec = tp->burstWait.usec;
7661 peerStats->rtt = tp->rtt;
7662 peerStats->rtt_dev = tp->rtt_dev;
7663 peerStats->timeout.sec = tp->timeout.sec;
7664 peerStats->timeout.usec = tp->timeout.usec;
7665 peerStats->nSent = tp->nSent;
7666 peerStats->reSends = tp->reSends;
7667 peerStats->inPacketSkew = tp->inPacketSkew;
7668 peerStats->outPacketSkew = tp->outPacketSkew;
7669 peerStats->rateFlag = tp->rateFlag;
7670 peerStats->natMTU = tp->natMTU;
7671 peerStats->maxMTU = tp->maxMTU;
7672 peerStats->maxDgramPackets = tp->maxDgramPackets;
7673 peerStats->ifDgramPackets = tp->ifDgramPackets;
7674 peerStats->MTU = tp->MTU;
7675 peerStats->cwind = tp->cwind;
7676 peerStats->nDgramPackets = tp->nDgramPackets;
7677 peerStats->congestSeq = tp->congestSeq;
7678 peerStats->bytesSent.high = tp->bytesSent.high;
7679 peerStats->bytesSent.low = tp->bytesSent.low;
7680 peerStats->bytesReceived.high = tp->bytesReceived.high;
7681 peerStats->bytesReceived.low = tp->bytesReceived.low;
7682 MUTEX_EXIT(&tp->peer_lock);
7684 MUTEX_ENTER(&rx_peerHashTable_lock);
7687 MUTEX_EXIT(&rx_peerHashTable_lock);
7695 struct rx_serverQueueEntry *np;
7698 struct rx_call *call;
7699 struct rx_serverQueueEntry *sq;
7703 if (rxinit_status == 1) {
7705 return; /* Already shutdown. */
7709 #ifndef AFS_PTHREAD_ENV
7710 FD_ZERO(&rx_selectMask);
7711 #endif /* AFS_PTHREAD_ENV */
7712 rxi_dataQuota = RX_MAX_QUOTA;
7713 #ifndef AFS_PTHREAD_ENV
7715 #endif /* AFS_PTHREAD_ENV */
7718 #ifndef AFS_PTHREAD_ENV
7719 #ifndef AFS_USE_GETTIMEOFDAY
7721 #endif /* AFS_USE_GETTIMEOFDAY */
7722 #endif /* AFS_PTHREAD_ENV */
7724 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7725 call = queue_First(&rx_freeCallQueue, rx_call);
7727 rxi_Free(call, sizeof(struct rx_call));
7730 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7731 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7737 struct rx_peer **peer_ptr, **peer_end;
7738 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7739 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7741 struct rx_peer *peer, *next;
7743 MUTEX_ENTER(&rx_peerHashTable_lock);
7744 for (peer = *peer_ptr; peer; peer = next) {
7745 rx_interface_stat_p rpc_stat, nrpc_stat;
7748 MUTEX_ENTER(&rx_rpc_stats);
7749 MUTEX_ENTER(&peer->peer_lock);
7751 (&peer->rpcStats, rpc_stat, nrpc_stat,
7752 rx_interface_stat)) {
7753 unsigned int num_funcs;
7756 queue_Remove(&rpc_stat->queue_header);
7757 queue_Remove(&rpc_stat->all_peers);
7758 num_funcs = rpc_stat->stats[0].func_total;
7760 sizeof(rx_interface_stat_t) +
7761 rpc_stat->stats[0].func_total *
7762 sizeof(rx_function_entry_v1_t);
7764 rxi_Free(rpc_stat, space);
7766 /* rx_rpc_stats must be held */
7767 rxi_rpc_peer_stat_cnt -= num_funcs;
7769 MUTEX_EXIT(&peer->peer_lock);
7770 MUTEX_EXIT(&rx_rpc_stats);
7774 if (rx_stats_active)
7775 rx_atomic_dec(&rx_stats.nPeerStructs);
7777 MUTEX_EXIT(&rx_peerHashTable_lock);
7780 for (i = 0; i < RX_MAX_SERVICES; i++) {
7782 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7784 for (i = 0; i < rx_hashTableSize; i++) {
7785 struct rx_connection *tc, *ntc;
7786 MUTEX_ENTER(&rx_connHashTable_lock);
7787 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7789 for (j = 0; j < RX_MAXCALLS; j++) {
7791 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7794 rxi_Free(tc, sizeof(*tc));
7796 MUTEX_EXIT(&rx_connHashTable_lock);
7799 MUTEX_ENTER(&freeSQEList_lock);
7801 while ((np = rx_FreeSQEList)) {
7802 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7803 MUTEX_DESTROY(&np->lock);
7804 rxi_Free(np, sizeof(*np));
7807 MUTEX_EXIT(&freeSQEList_lock);
7808 MUTEX_DESTROY(&freeSQEList_lock);
7809 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7810 MUTEX_DESTROY(&rx_connHashTable_lock);
7811 MUTEX_DESTROY(&rx_peerHashTable_lock);
7812 MUTEX_DESTROY(&rx_serverPool_lock);
7814 osi_Free(rx_connHashTable,
7815 rx_hashTableSize * sizeof(struct rx_connection *));
7816 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7818 UNPIN(rx_connHashTable,
7819 rx_hashTableSize * sizeof(struct rx_connection *));
7820 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7822 rxi_FreeAllPackets();
7824 MUTEX_ENTER(&rx_quota_mutex);
7825 rxi_dataQuota = RX_MAX_QUOTA;
7826 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7827 MUTEX_EXIT(&rx_quota_mutex);
7832 #ifdef RX_ENABLE_LOCKS
7834 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7836 if (!MUTEX_ISMINE(lockaddr))
7837 osi_Panic("Lock not held: %s", msg);
7839 #endif /* RX_ENABLE_LOCKS */
7844 * Routines to implement connection specific data.
7848 rx_KeyCreate(rx_destructor_t rtn)
7851 MUTEX_ENTER(&rxi_keyCreate_lock);
7852 key = rxi_keyCreate_counter++;
7853 rxi_keyCreate_destructor = (rx_destructor_t *)
7854 realloc((void *)rxi_keyCreate_destructor,
7855 (key + 1) * sizeof(rx_destructor_t));
7856 rxi_keyCreate_destructor[key] = rtn;
7857 MUTEX_EXIT(&rxi_keyCreate_lock);
7862 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7865 MUTEX_ENTER(&conn->conn_data_lock);
7866 if (!conn->specific) {
7867 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7868 for (i = 0; i < key; i++)
7869 conn->specific[i] = NULL;
7870 conn->nSpecific = key + 1;
7871 conn->specific[key] = ptr;
7872 } else if (key >= conn->nSpecific) {
7873 conn->specific = (void **)
7874 realloc(conn->specific, (key + 1) * sizeof(void *));
7875 for (i = conn->nSpecific; i < key; i++)
7876 conn->specific[i] = NULL;
7877 conn->nSpecific = key + 1;
7878 conn->specific[key] = ptr;
7880 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7881 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7882 conn->specific[key] = ptr;
7884 MUTEX_EXIT(&conn->conn_data_lock);
7888 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7891 MUTEX_ENTER(&svc->svc_data_lock);
7892 if (!svc->specific) {
7893 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
7894 for (i = 0; i < key; i++)
7895 svc->specific[i] = NULL;
7896 svc->nSpecific = key + 1;
7897 svc->specific[key] = ptr;
7898 } else if (key >= svc->nSpecific) {
7899 svc->specific = (void **)
7900 realloc(svc->specific, (key + 1) * sizeof(void *));
7901 for (i = svc->nSpecific; i < key; i++)
7902 svc->specific[i] = NULL;
7903 svc->nSpecific = key + 1;
7904 svc->specific[key] = ptr;
7906 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7907 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7908 svc->specific[key] = ptr;
7910 MUTEX_EXIT(&svc->svc_data_lock);
7914 rx_GetSpecific(struct rx_connection *conn, int key)
7917 MUTEX_ENTER(&conn->conn_data_lock);
7918 if (key >= conn->nSpecific)
7921 ptr = conn->specific[key];
7922 MUTEX_EXIT(&conn->conn_data_lock);
7927 rx_GetServiceSpecific(struct rx_service *svc, int key)
7930 MUTEX_ENTER(&svc->svc_data_lock);
7931 if (key >= svc->nSpecific)
7934 ptr = svc->specific[key];
7935 MUTEX_EXIT(&svc->svc_data_lock);
7940 #endif /* !KERNEL */
7943 * processStats is a queue used to store the statistics for the local
7944 * process. Its contents are similar to the contents of the rpcStats
7945 * queue on a rx_peer structure, but the actual data stored within
7946 * this queue contains totals across the lifetime of the process (assuming
7947 * the stats have not been reset) - unlike the per peer structures
7948 * which can come and go based upon the peer lifetime.
7951 static struct rx_queue processStats = { &processStats, &processStats };
7954 * peerStats is a queue used to store the statistics for all peer structs.
7955 * Its contents are the union of all the peer rpcStats queues.
7958 static struct rx_queue peerStats = { &peerStats, &peerStats };
7961 * rxi_monitor_processStats is used to turn process wide stat collection
7965 static int rxi_monitor_processStats = 0;
7968 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7971 static int rxi_monitor_peerStats = 0;
7974 * rxi_AddRpcStat - given all of the information for a particular rpc
7975 * call, create (if needed) and update the stat totals for the rpc.
7979 * IN stats - the queue of stats that will be updated with the new value
7981 * IN rxInterface - a unique number that identifies the rpc interface
7983 * IN currentFunc - the index of the function being invoked
7985 * IN totalFunc - the total number of functions in this interface
7987 * IN queueTime - the amount of time this function waited for a thread
7989 * IN execTime - the amount of time this function invocation took to execute
7991 * IN bytesSent - the number bytes sent by this invocation
7993 * IN bytesRcvd - the number bytes received by this invocation
7995 * IN isServer - if true, this invocation was made to a server
7997 * IN remoteHost - the ip address of the remote host
7999 * IN remotePort - the port of the remote host
8001 * IN addToPeerList - if != 0, add newly created stat to the global peer list
8003 * INOUT counter - if a new stats structure is allocated, the counter will
8004 * be updated with the new number of allocated stat structures
8012 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
8013 afs_uint32 currentFunc, afs_uint32 totalFunc,
8014 struct clock *queueTime, struct clock *execTime,
8015 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
8016 afs_uint32 remoteHost, afs_uint32 remotePort,
8017 int addToPeerList, unsigned int *counter)
8020 rx_interface_stat_p rpc_stat, nrpc_stat;
8023 * See if there's already a structure for this interface
8026 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8027 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8028 && (rpc_stat->stats[0].remote_is_server == isServer))
8033 * Didn't find a match so allocate a new structure and add it to the
8037 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
8038 || (rpc_stat->stats[0].interfaceId != rxInterface)
8039 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8044 sizeof(rx_interface_stat_t) +
8045 totalFunc * sizeof(rx_function_entry_v1_t);
8047 rpc_stat = rxi_Alloc(space);
8048 if (rpc_stat == NULL) {
8052 *counter += totalFunc;
8053 for (i = 0; i < totalFunc; i++) {
8054 rpc_stat->stats[i].remote_peer = remoteHost;
8055 rpc_stat->stats[i].remote_port = remotePort;
8056 rpc_stat->stats[i].remote_is_server = isServer;
8057 rpc_stat->stats[i].interfaceId = rxInterface;
8058 rpc_stat->stats[i].func_total = totalFunc;
8059 rpc_stat->stats[i].func_index = i;
8060 hzero(rpc_stat->stats[i].invocations);
8061 hzero(rpc_stat->stats[i].bytes_sent);
8062 hzero(rpc_stat->stats[i].bytes_rcvd);
8063 rpc_stat->stats[i].queue_time_sum.sec = 0;
8064 rpc_stat->stats[i].queue_time_sum.usec = 0;
8065 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8066 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8067 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8068 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8069 rpc_stat->stats[i].queue_time_max.sec = 0;
8070 rpc_stat->stats[i].queue_time_max.usec = 0;
8071 rpc_stat->stats[i].execution_time_sum.sec = 0;
8072 rpc_stat->stats[i].execution_time_sum.usec = 0;
8073 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8074 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8075 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8076 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8077 rpc_stat->stats[i].execution_time_max.sec = 0;
8078 rpc_stat->stats[i].execution_time_max.usec = 0;
8080 queue_Prepend(stats, rpc_stat);
8081 if (addToPeerList) {
8082 queue_Prepend(&peerStats, &rpc_stat->all_peers);
8087 * Increment the stats for this function
8090 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
8091 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
8092 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
8093 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8094 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8095 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8096 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8098 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8099 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8101 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8102 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8104 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8105 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8107 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8108 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8116 * rx_IncrementTimeAndCount - increment the times and count for a particular
8121 * IN peer - the peer who invoked the rpc
8123 * IN rxInterface - a unique number that identifies the rpc interface
8125 * IN currentFunc - the index of the function being invoked
8127 * IN totalFunc - the total number of functions in this interface
8129 * IN queueTime - the amount of time this function waited for a thread
8131 * IN execTime - the amount of time this function invocation took to execute
8133 * IN bytesSent - the number bytes sent by this invocation
8135 * IN bytesRcvd - the number bytes received by this invocation
8137 * IN isServer - if true, this invocation was made to a server
8145 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8146 afs_uint32 currentFunc, afs_uint32 totalFunc,
8147 struct clock *queueTime, struct clock *execTime,
8148 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8152 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8155 MUTEX_ENTER(&rx_rpc_stats);
8157 if (rxi_monitor_peerStats) {
8158 MUTEX_ENTER(&peer->peer_lock);
8159 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8160 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8161 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8162 MUTEX_EXIT(&peer->peer_lock);
8165 if (rxi_monitor_processStats) {
8166 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8167 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8168 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8171 MUTEX_EXIT(&rx_rpc_stats);
8176 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8180 * IN callerVersion - the rpc stat version of the caller.
8182 * IN count - the number of entries to marshall.
8184 * IN stats - pointer to stats to be marshalled.
8186 * OUT ptr - Where to store the marshalled data.
8193 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8194 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8200 * We only support the first version
8202 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8203 *(ptr++) = stats->remote_peer;
8204 *(ptr++) = stats->remote_port;
8205 *(ptr++) = stats->remote_is_server;
8206 *(ptr++) = stats->interfaceId;
8207 *(ptr++) = stats->func_total;
8208 *(ptr++) = stats->func_index;
8209 *(ptr++) = hgethi(stats->invocations);
8210 *(ptr++) = hgetlo(stats->invocations);
8211 *(ptr++) = hgethi(stats->bytes_sent);
8212 *(ptr++) = hgetlo(stats->bytes_sent);
8213 *(ptr++) = hgethi(stats->bytes_rcvd);
8214 *(ptr++) = hgetlo(stats->bytes_rcvd);
8215 *(ptr++) = stats->queue_time_sum.sec;
8216 *(ptr++) = stats->queue_time_sum.usec;
8217 *(ptr++) = stats->queue_time_sum_sqr.sec;
8218 *(ptr++) = stats->queue_time_sum_sqr.usec;
8219 *(ptr++) = stats->queue_time_min.sec;
8220 *(ptr++) = stats->queue_time_min.usec;
8221 *(ptr++) = stats->queue_time_max.sec;
8222 *(ptr++) = stats->queue_time_max.usec;
8223 *(ptr++) = stats->execution_time_sum.sec;
8224 *(ptr++) = stats->execution_time_sum.usec;
8225 *(ptr++) = stats->execution_time_sum_sqr.sec;
8226 *(ptr++) = stats->execution_time_sum_sqr.usec;
8227 *(ptr++) = stats->execution_time_min.sec;
8228 *(ptr++) = stats->execution_time_min.usec;
8229 *(ptr++) = stats->execution_time_max.sec;
8230 *(ptr++) = stats->execution_time_max.usec;
8236 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8241 * IN callerVersion - the rpc stat version of the caller
8243 * OUT myVersion - the rpc stat version of this function
8245 * OUT clock_sec - local time seconds
8247 * OUT clock_usec - local time microseconds
8249 * OUT allocSize - the number of bytes allocated to contain stats
8251 * OUT statCount - the number stats retrieved from this process.
8253 * OUT stats - the actual stats retrieved from this process.
8257 * Returns void. If successful, stats will != NULL.
8261 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8262 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8263 size_t * allocSize, afs_uint32 * statCount,
8264 afs_uint32 ** stats)
8274 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8277 * Check to see if stats are enabled
8280 MUTEX_ENTER(&rx_rpc_stats);
8281 if (!rxi_monitor_processStats) {
8282 MUTEX_EXIT(&rx_rpc_stats);
8286 clock_GetTime(&now);
8287 *clock_sec = now.sec;
8288 *clock_usec = now.usec;
8291 * Allocate the space based upon the caller version
8293 * If the client is at an older version than we are,
8294 * we return the statistic data in the older data format, but
8295 * we still return our version number so the client knows we
8296 * are maintaining more data than it can retrieve.
8299 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8300 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8301 *statCount = rxi_rpc_process_stat_cnt;
8304 * This can't happen yet, but in the future version changes
8305 * can be handled by adding additional code here
8309 if (space > (size_t) 0) {
8311 ptr = *stats = rxi_Alloc(space);
8314 rx_interface_stat_p rpc_stat, nrpc_stat;
8318 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8320 * Copy the data based upon the caller version
8322 rx_MarshallProcessRPCStats(callerVersion,
8323 rpc_stat->stats[0].func_total,
8324 rpc_stat->stats, &ptr);
8330 MUTEX_EXIT(&rx_rpc_stats);
8335 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8339 * IN callerVersion - the rpc stat version of the caller
8341 * OUT myVersion - the rpc stat version of this function
8343 * OUT clock_sec - local time seconds
8345 * OUT clock_usec - local time microseconds
8347 * OUT allocSize - the number of bytes allocated to contain stats
8349 * OUT statCount - the number of stats retrieved from the individual
8352 * OUT stats - the actual stats retrieved from the individual peer structures.
8356 * Returns void. If successful, stats will != NULL.
8360 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8361 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8362 size_t * allocSize, afs_uint32 * statCount,
8363 afs_uint32 ** stats)
8373 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8376 * Check to see if stats are enabled
8379 MUTEX_ENTER(&rx_rpc_stats);
8380 if (!rxi_monitor_peerStats) {
8381 MUTEX_EXIT(&rx_rpc_stats);
8385 clock_GetTime(&now);
8386 *clock_sec = now.sec;
8387 *clock_usec = now.usec;
8390 * Allocate the space based upon the caller version
8392 * If the client is at an older version than we are,
8393 * we return the statistic data in the older data format, but
8394 * we still return our version number so the client knows we
8395 * are maintaining more data than it can retrieve.
8398 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8399 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8400 *statCount = rxi_rpc_peer_stat_cnt;
8403 * This can't happen yet, but in the future version changes
8404 * can be handled by adding additional code here
8408 if (space > (size_t) 0) {
8410 ptr = *stats = rxi_Alloc(space);
8413 rx_interface_stat_p rpc_stat, nrpc_stat;
8417 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8419 * We have to fix the offset of rpc_stat since we are
8420 * keeping this structure on two rx_queues. The rx_queue
8421 * package assumes that the rx_queue member is the first
8422 * member of the structure. That is, rx_queue assumes that
8423 * any one item is only on one queue at a time. We are
8424 * breaking that assumption and so we have to do a little
8425 * math to fix our pointers.
8428 fix_offset = (char *)rpc_stat;
8429 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8430 rpc_stat = (rx_interface_stat_p) fix_offset;
8433 * Copy the data based upon the caller version
8435 rx_MarshallProcessRPCStats(callerVersion,
8436 rpc_stat->stats[0].func_total,
8437 rpc_stat->stats, &ptr);
8443 MUTEX_EXIT(&rx_rpc_stats);
8448 * rx_FreeRPCStats - free memory allocated by
8449 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8453 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8454 * rx_RetrievePeerRPCStats
8456 * IN allocSize - the number of bytes in stats.
8464 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8466 rxi_Free(stats, allocSize);
8470 * rx_queryProcessRPCStats - see if process rpc stat collection is
8471 * currently enabled.
8477 * Returns 0 if stats are not enabled != 0 otherwise
8481 rx_queryProcessRPCStats(void)
8484 MUTEX_ENTER(&rx_rpc_stats);
8485 rc = rxi_monitor_processStats;
8486 MUTEX_EXIT(&rx_rpc_stats);
8491 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8497 * Returns 0 if stats are not enabled != 0 otherwise
8501 rx_queryPeerRPCStats(void)
8504 MUTEX_ENTER(&rx_rpc_stats);
8505 rc = rxi_monitor_peerStats;
8506 MUTEX_EXIT(&rx_rpc_stats);
8511 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8521 rx_enableProcessRPCStats(void)
8523 MUTEX_ENTER(&rx_rpc_stats);
8524 rx_enable_stats = 1;
8525 rxi_monitor_processStats = 1;
8526 MUTEX_EXIT(&rx_rpc_stats);
8530 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8540 rx_enablePeerRPCStats(void)
8542 MUTEX_ENTER(&rx_rpc_stats);
8543 rx_enable_stats = 1;
8544 rxi_monitor_peerStats = 1;
8545 MUTEX_EXIT(&rx_rpc_stats);
8549 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8559 rx_disableProcessRPCStats(void)
8561 rx_interface_stat_p rpc_stat, nrpc_stat;
8564 MUTEX_ENTER(&rx_rpc_stats);
8567 * Turn off process statistics and if peer stats is also off, turn
8571 rxi_monitor_processStats = 0;
8572 if (rxi_monitor_peerStats == 0) {
8573 rx_enable_stats = 0;
8576 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8577 unsigned int num_funcs = 0;
8580 queue_Remove(rpc_stat);
8581 num_funcs = rpc_stat->stats[0].func_total;
8583 sizeof(rx_interface_stat_t) +
8584 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8586 rxi_Free(rpc_stat, space);
8587 rxi_rpc_process_stat_cnt -= num_funcs;
8589 MUTEX_EXIT(&rx_rpc_stats);
8593 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8603 rx_disablePeerRPCStats(void)
8605 struct rx_peer **peer_ptr, **peer_end;
8609 * Turn off peer statistics and if process stats is also off, turn
8613 rxi_monitor_peerStats = 0;
8614 if (rxi_monitor_processStats == 0) {
8615 rx_enable_stats = 0;
8618 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8619 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8621 struct rx_peer *peer, *next, *prev;
8623 MUTEX_ENTER(&rx_peerHashTable_lock);
8624 MUTEX_ENTER(&rx_rpc_stats);
8625 for (prev = peer = *peer_ptr; peer; peer = next) {
8627 code = MUTEX_TRYENTER(&peer->peer_lock);
8629 rx_interface_stat_p rpc_stat, nrpc_stat;
8632 if (prev == *peer_ptr) {
8643 MUTEX_EXIT(&rx_peerHashTable_lock);
8646 (&peer->rpcStats, rpc_stat, nrpc_stat,
8647 rx_interface_stat)) {
8648 unsigned int num_funcs = 0;
8651 queue_Remove(&rpc_stat->queue_header);
8652 queue_Remove(&rpc_stat->all_peers);
8653 num_funcs = rpc_stat->stats[0].func_total;
8655 sizeof(rx_interface_stat_t) +
8656 rpc_stat->stats[0].func_total *
8657 sizeof(rx_function_entry_v1_t);
8659 rxi_Free(rpc_stat, space);
8660 rxi_rpc_peer_stat_cnt -= num_funcs;
8662 MUTEX_EXIT(&peer->peer_lock);
8664 MUTEX_ENTER(&rx_peerHashTable_lock);
8674 MUTEX_EXIT(&rx_rpc_stats);
8675 MUTEX_EXIT(&rx_peerHashTable_lock);
8680 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8685 * IN clearFlag - flag indicating which stats to clear
8693 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8695 rx_interface_stat_p rpc_stat, nrpc_stat;
8697 MUTEX_ENTER(&rx_rpc_stats);
8699 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8700 unsigned int num_funcs = 0, i;
8701 num_funcs = rpc_stat->stats[0].func_total;
8702 for (i = 0; i < num_funcs; i++) {
8703 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8704 hzero(rpc_stat->stats[i].invocations);
8706 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8707 hzero(rpc_stat->stats[i].bytes_sent);
8709 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8710 hzero(rpc_stat->stats[i].bytes_rcvd);
8712 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8713 rpc_stat->stats[i].queue_time_sum.sec = 0;
8714 rpc_stat->stats[i].queue_time_sum.usec = 0;
8716 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8717 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8718 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8720 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8721 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8722 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8724 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8725 rpc_stat->stats[i].queue_time_max.sec = 0;
8726 rpc_stat->stats[i].queue_time_max.usec = 0;
8728 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8729 rpc_stat->stats[i].execution_time_sum.sec = 0;
8730 rpc_stat->stats[i].execution_time_sum.usec = 0;
8732 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8733 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8734 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8736 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8737 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8738 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8740 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8741 rpc_stat->stats[i].execution_time_max.sec = 0;
8742 rpc_stat->stats[i].execution_time_max.usec = 0;
8747 MUTEX_EXIT(&rx_rpc_stats);
8751 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8756 * IN clearFlag - flag indicating which stats to clear
8764 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8766 rx_interface_stat_p rpc_stat, nrpc_stat;
8768 MUTEX_ENTER(&rx_rpc_stats);
8770 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8771 unsigned int num_funcs = 0, i;
8774 * We have to fix the offset of rpc_stat since we are
8775 * keeping this structure on two rx_queues. The rx_queue
8776 * package assumes that the rx_queue member is the first
8777 * member of the structure. That is, rx_queue assumes that
8778 * any one item is only on one queue at a time. We are
8779 * breaking that assumption and so we have to do a little
8780 * math to fix our pointers.
8783 fix_offset = (char *)rpc_stat;
8784 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8785 rpc_stat = (rx_interface_stat_p) fix_offset;
8787 num_funcs = rpc_stat->stats[0].func_total;
8788 for (i = 0; i < num_funcs; i++) {
8789 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8790 hzero(rpc_stat->stats[i].invocations);
8792 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8793 hzero(rpc_stat->stats[i].bytes_sent);
8795 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8796 hzero(rpc_stat->stats[i].bytes_rcvd);
8798 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8799 rpc_stat->stats[i].queue_time_sum.sec = 0;
8800 rpc_stat->stats[i].queue_time_sum.usec = 0;
8802 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8803 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8804 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8806 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8807 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8808 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8810 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8811 rpc_stat->stats[i].queue_time_max.sec = 0;
8812 rpc_stat->stats[i].queue_time_max.usec = 0;
8814 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8815 rpc_stat->stats[i].execution_time_sum.sec = 0;
8816 rpc_stat->stats[i].execution_time_sum.usec = 0;
8818 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8819 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8820 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8822 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8823 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8824 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8826 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8827 rpc_stat->stats[i].execution_time_max.sec = 0;
8828 rpc_stat->stats[i].execution_time_max.usec = 0;
8833 MUTEX_EXIT(&rx_rpc_stats);
8837 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8838 * is authorized to enable/disable/clear RX statistics.
8840 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8843 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8845 rxi_rxstat_userok = proc;
8849 rx_RxStatUserOk(struct rx_call *call)
8851 if (!rxi_rxstat_userok)
8853 return rxi_rxstat_userok(call);
8858 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8859 * function in the MSVC runtime DLL (msvcrt.dll).
8861 * Note: the system serializes calls to this function.
8864 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8865 DWORD reason, /* reason function is being called */
8866 LPVOID reserved) /* reserved for future use */
8869 case DLL_PROCESS_ATTACH:
8870 /* library is being attached to a process */
8874 case DLL_PROCESS_DETACH:
8881 #endif /* AFS_NT40_ENV */
8884 int rx_DumpCalls(FILE *outputFile, char *cookie)
8886 #ifdef RXDEBUG_PACKET
8887 #ifdef KDUMP_RX_LOCK
8888 struct rx_call_rx_lock *c;
8895 #define RXDPRINTF sprintf
8896 #define RXDPRINTOUT output
8898 #define RXDPRINTF fprintf
8899 #define RXDPRINTOUT outputFile
8902 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8904 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8907 for (c = rx_allCallsp; c; c = c->allNextp) {
8908 u_short rqc, tqc, iovqc;
8909 struct rx_packet *p, *np;
8911 MUTEX_ENTER(&c->lock);
8912 queue_Count(&c->rq, p, np, rx_packet, rqc);
8913 queue_Count(&c->tq, p, np, rx_packet, tqc);
8914 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8916 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, "
8917 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8918 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8919 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8920 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8921 #ifdef RX_ENABLE_LOCKS
8924 #ifdef RX_REFCOUNT_CHECK
8925 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
8926 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
8929 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,
8930 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
8931 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
8932 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
8933 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
8934 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
8935 #ifdef RX_ENABLE_LOCKS
8936 , (afs_uint32)c->refCount
8938 #ifdef RX_REFCOUNT_CHECK
8939 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
8942 MUTEX_EXIT(&c->lock);
8945 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8948 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
8950 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8952 #endif /* RXDEBUG_PACKET */