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>
13 #include "afs/param.h"
16 # include "afs/sysincludes.h"
17 # include "afsincludes.h"
22 # ifdef AFS_LINUX20_ENV
23 # include "h/socket.h"
25 # include "netinet/in.h"
27 # include "netinet/ip6.h"
30 # include "inet/common.h"
32 # include "inet/ip_ire.h"
34 # include "afs/afs_args.h"
35 # include "afs/afs_osi.h"
36 # ifdef RX_KERNEL_TRACE
37 # include "rx_kcommon.h"
39 # if defined(AFS_AIX_ENV)
43 # undef RXDEBUG /* turn off debugging */
45 # if defined(AFS_SGI_ENV)
46 # include "sys/debug.h"
49 # include "afs/sysincludes.h"
50 # include "afsincludes.h"
51 # endif /* !UKERNEL */
52 # include "afs/lock.h"
53 # include "rx_kmutex.h"
54 # include "rx_kernel.h"
55 # define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
56 # define AFSOP_STOP_AFS 211 /* Stop AFS process */
57 # define AFSOP_STOP_BKG 212 /* Stop BKG process */
58 extern afs_int32 afs_termState;
60 # include "sys/lockl.h"
61 # include "sys/lock_def.h"
62 # endif /* AFS_AIX41_ENV */
63 # include "afs/rxgen_consts.h"
66 # include <sys/types.h>
76 # include <afs/afsutil.h>
77 # include <WINNT\afsreg.h>
79 # include <sys/socket.h>
80 # include <sys/file.h>
82 # include <sys/stat.h>
83 # include <netinet/in.h>
84 # include <sys/time.h>
92 #include "rx_atomic.h"
93 #include "rx_globals.h"
95 #include "rx_internal.h"
98 #include <afs/rxgen_consts.h>
101 #ifdef AFS_PTHREAD_ENV
103 int (*registerProgram) (pid_t, char *) = 0;
104 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
107 int (*registerProgram) (PROCESS, char *) = 0;
108 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
112 /* Local static routines */
113 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
114 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
115 struct rx_peer *, struct clock *);
117 #ifdef RX_ENABLE_LOCKS
118 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
121 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
123 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
124 rx_atomic_t rxi_start_in_error;
126 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
129 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
130 * currently allocated within rx. This number is used to allocate the
131 * memory required to return the statistics when queried.
132 * Protected by the rx_rpc_stats mutex.
135 static unsigned int rxi_rpc_peer_stat_cnt;
138 * rxi_rpc_process_stat_cnt counts the total number of local process stat
139 * structures currently allocated within rx. The number is used to allocate
140 * the memory required to return the statistics when queried.
141 * Protected by the rx_rpc_stats mutex.
144 static unsigned int rxi_rpc_process_stat_cnt;
146 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
147 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
149 #if !defined(offsetof)
150 #include <stddef.h> /* for definition of offsetof() */
153 #ifdef RX_ENABLE_LOCKS
154 afs_kmutex_t rx_atomic_mutex;
157 #ifdef AFS_PTHREAD_ENV
160 * Use procedural initialization of mutexes/condition variables
164 extern afs_kmutex_t rx_quota_mutex;
165 extern afs_kmutex_t rx_pthread_mutex;
166 extern afs_kmutex_t rx_packets_mutex;
167 extern afs_kmutex_t rx_refcnt_mutex;
168 extern afs_kmutex_t des_init_mutex;
169 extern afs_kmutex_t des_random_mutex;
170 extern afs_kmutex_t rx_clock_mutex;
171 extern afs_kmutex_t rxi_connCacheMutex;
172 extern afs_kmutex_t rx_event_mutex;
173 extern afs_kmutex_t osi_malloc_mutex;
174 extern afs_kmutex_t event_handler_mutex;
175 extern afs_kmutex_t listener_mutex;
176 extern afs_kmutex_t rx_if_init_mutex;
177 extern afs_kmutex_t rx_if_mutex;
178 extern afs_kmutex_t rxkad_client_uid_mutex;
179 extern afs_kmutex_t rxkad_random_mutex;
181 extern afs_kcondvar_t rx_event_handler_cond;
182 extern afs_kcondvar_t rx_listener_cond;
184 static afs_kmutex_t epoch_mutex;
185 static afs_kmutex_t rx_init_mutex;
186 static afs_kmutex_t rx_debug_mutex;
187 static afs_kmutex_t rx_rpc_stats;
190 rxi_InitPthread(void)
192 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
193 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
194 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
195 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
196 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
197 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
198 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
199 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
200 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
201 MUTEX_INIT(&rx_event_mutex, "event", MUTEX_DEFAULT, 0);
202 MUTEX_INIT(&osi_malloc_mutex, "malloc", MUTEX_DEFAULT, 0);
203 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
204 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
205 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
206 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
207 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
208 MUTEX_INIT(&rxkad_client_uid_mutex, "uid", MUTEX_DEFAULT, 0);
209 MUTEX_INIT(&rxkad_random_mutex, "rxkad random", MUTEX_DEFAULT, 0);
210 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
212 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
213 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
215 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
216 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
218 rxkad_global_stats_init();
220 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
221 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
222 #ifdef RX_ENABLE_LOCKS
225 #endif /* RX_LOCKS_DB */
226 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
227 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
229 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
231 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
233 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
235 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
236 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
237 #endif /* RX_ENABLE_LOCKS */
240 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
241 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
243 * The rx_stats_mutex mutex protects the following global variables:
244 * rxi_lowConnRefCount
245 * rxi_lowPeerRefCount
254 * The rx_quota_mutex mutex protects the following global variables:
262 * The rx_freePktQ_lock protects the following global variables:
267 * The rx_packets_mutex mutex protects the following global variables:
275 * The rx_pthread_mutex mutex protects the following global variables:
276 * rxi_fcfs_thread_num
279 #define INIT_PTHREAD_LOCKS
283 /* Variables for handling the minProcs implementation. availProcs gives the
284 * number of threads available in the pool at this moment (not counting dudes
285 * executing right now). totalMin gives the total number of procs required
286 * for handling all minProcs requests. minDeficit is a dynamic variable
287 * tracking the # of procs required to satisfy all of the remaining minProcs
289 * For fine grain locking to work, the quota check and the reservation of
290 * a server thread has to come while rxi_availProcs and rxi_minDeficit
291 * are locked. To this end, the code has been modified under #ifdef
292 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
293 * same time. A new function, ReturnToServerPool() returns the allocation.
295 * A call can be on several queue's (but only one at a time). When
296 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
297 * that no one else is touching the queue. To this end, we store the address
298 * of the queue lock in the call structure (under the call lock) when we
299 * put the call on a queue, and we clear the call_queue_lock when the
300 * call is removed from a queue (once the call lock has been obtained).
301 * This allows rxi_ResetCall to safely synchronize with others wishing
302 * to manipulate the queue.
305 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
306 static afs_kmutex_t rx_rpc_stats;
307 void rxi_StartUnlocked(struct rxevent *event, void *call,
308 void *arg1, int istack);
311 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
312 ** pretty good that the next packet coming in is from the same connection
313 ** as the last packet, since we're send multiple packets in a transmit window.
315 struct rx_connection *rxLastConn = 0;
317 #ifdef RX_ENABLE_LOCKS
318 /* The locking hierarchy for rx fine grain locking is composed of these
321 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
322 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
323 * call->lock - locks call data fields.
324 * These are independent of each other:
325 * rx_freeCallQueue_lock
330 * serverQueueEntry->lock
331 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
333 * peer->lock - locks peer data fields.
334 * conn_data_lock - that more than one thread is not updating a conn data
335 * field at the same time.
346 * Do we need a lock to protect the peer field in the conn structure?
347 * conn->peer was previously a constant for all intents and so has no
348 * lock protecting this field. The multihomed client delta introduced
349 * a RX code change : change the peer field in the connection structure
350 * to that remote interface from which the last packet for this
351 * connection was sent out. This may become an issue if further changes
354 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
355 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
357 /* rxdb_fileID is used to identify the lock location, along with line#. */
358 static int rxdb_fileID = RXDB_FILE_RX;
359 #endif /* RX_LOCKS_DB */
360 #else /* RX_ENABLE_LOCKS */
361 #define SET_CALL_QUEUE_LOCK(C, L)
362 #define CLEAR_CALL_QUEUE_LOCK(C)
363 #endif /* RX_ENABLE_LOCKS */
364 struct rx_serverQueueEntry *rx_waitForPacket = 0;
365 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
367 /* ------------Exported Interfaces------------- */
369 /* This function allows rxkad to set the epoch to a suitably random number
370 * which rx_NewConnection will use in the future. The principle purpose is to
371 * get rxnull connections to use the same epoch as the rxkad connections do, at
372 * least once the first rxkad connection is established. This is important now
373 * that the host/port addresses aren't used in FindConnection: the uniqueness
374 * of epoch/cid matters and the start time won't do. */
376 #ifdef AFS_PTHREAD_ENV
378 * This mutex protects the following global variables:
382 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
383 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
387 #endif /* AFS_PTHREAD_ENV */
390 rx_SetEpoch(afs_uint32 epoch)
397 /* Initialize rx. A port number may be mentioned, in which case this
398 * becomes the default port number for any service installed later.
399 * If 0 is provided for the port number, a random port will be chosen
400 * by the kernel. Whether this will ever overlap anything in
401 * /etc/services is anybody's guess... Returns 0 on success, -1 on
406 int rxinit_status = 1;
407 #ifdef AFS_PTHREAD_ENV
409 * This mutex protects the following global variables:
413 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
414 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
417 #define UNLOCK_RX_INIT
421 rx_InitHost(u_int host, u_int port)
428 char *htable, *ptable;
435 if (rxinit_status == 0) {
436 tmp_status = rxinit_status;
438 return tmp_status; /* Already started; return previous error code. */
444 if (afs_winsockInit() < 0)
450 * Initialize anything necessary to provide a non-premptive threading
453 rxi_InitializeThreadSupport();
456 /* Allocate and initialize a socket for client and perhaps server
459 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
460 if (rx_socket == OSI_NULLSOCKET) {
464 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
467 #endif /* RX_LOCKS_DB */
468 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
469 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
470 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
471 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
472 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
473 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
474 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
475 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
476 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
478 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
480 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
482 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
484 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
485 #if defined(AFS_HPUX110_ENV)
487 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
488 #endif /* AFS_HPUX110_ENV */
489 #endif /* RX_ENABLE_LOCKS && KERNEL */
492 rx_connDeadTime = 12;
493 rx_tranquil = 0; /* reset flag */
494 rxi_ResetStatistics();
496 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
497 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
498 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
499 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
500 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
501 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
503 /* Malloc up a bunch of packets & buffers */
505 queue_Init(&rx_freePacketQueue);
506 rxi_NeedMorePackets = FALSE;
507 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
509 /* enforce a minimum number of allocated packets */
510 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
511 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
513 /* allocate the initial free packet pool */
514 #ifdef RX_ENABLE_TSFPQ
515 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
516 #else /* RX_ENABLE_TSFPQ */
517 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
518 #endif /* RX_ENABLE_TSFPQ */
525 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
526 tv.tv_sec = clock_now.sec;
527 tv.tv_usec = clock_now.usec;
528 srand((unsigned int)tv.tv_usec);
535 #if defined(KERNEL) && !defined(UKERNEL)
536 /* Really, this should never happen in a real kernel */
539 struct sockaddr_in addr;
541 int addrlen = sizeof(addr);
543 socklen_t addrlen = sizeof(addr);
545 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
549 rx_port = addr.sin_port;
552 rx_stats.minRtt.sec = 9999999;
554 rx_SetEpoch(tv.tv_sec | 0x80000000);
556 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
557 * will provide a randomer value. */
559 MUTEX_ENTER(&rx_quota_mutex);
560 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
561 MUTEX_EXIT(&rx_quota_mutex);
562 /* *Slightly* random start time for the cid. This is just to help
563 * out with the hashing function at the peer */
564 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
565 rx_connHashTable = (struct rx_connection **)htable;
566 rx_peerHashTable = (struct rx_peer **)ptable;
568 rx_lastAckDelay.sec = 0;
569 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
570 rx_hardAckDelay.sec = 0;
571 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
572 rx_softAckDelay.sec = 0;
573 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
575 rxevent_Init(20, rxi_ReScheduleEvents);
577 /* Initialize various global queues */
578 queue_Init(&rx_idleServerQueue);
579 queue_Init(&rx_incomingCallQueue);
580 queue_Init(&rx_freeCallQueue);
582 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
583 /* Initialize our list of usable IP addresses. */
587 /* Start listener process (exact function is dependent on the
588 * implementation environment--kernel or user space) */
592 tmp_status = rxinit_status = 0;
600 return rx_InitHost(htonl(INADDR_ANY), port);
603 /* called with unincremented nRequestsRunning to see if it is OK to start
604 * a new thread in this service. Could be "no" for two reasons: over the
605 * max quota, or would prevent others from reaching their min quota.
607 #ifdef RX_ENABLE_LOCKS
608 /* This verion of QuotaOK reserves quota if it's ok while the
609 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
612 QuotaOK(struct rx_service *aservice)
614 /* check if over max quota */
615 if (aservice->nRequestsRunning >= aservice->maxProcs) {
619 /* under min quota, we're OK */
620 /* otherwise, can use only if there are enough to allow everyone
621 * to go to their min quota after this guy starts.
624 MUTEX_ENTER(&rx_quota_mutex);
625 if ((aservice->nRequestsRunning < aservice->minProcs)
626 || (rxi_availProcs > rxi_minDeficit)) {
627 aservice->nRequestsRunning++;
628 /* just started call in minProcs pool, need fewer to maintain
630 if (aservice->nRequestsRunning <= aservice->minProcs)
633 MUTEX_EXIT(&rx_quota_mutex);
636 MUTEX_EXIT(&rx_quota_mutex);
642 ReturnToServerPool(struct rx_service *aservice)
644 aservice->nRequestsRunning--;
645 MUTEX_ENTER(&rx_quota_mutex);
646 if (aservice->nRequestsRunning < aservice->minProcs)
649 MUTEX_EXIT(&rx_quota_mutex);
652 #else /* RX_ENABLE_LOCKS */
654 QuotaOK(struct rx_service *aservice)
657 /* under min quota, we're OK */
658 if (aservice->nRequestsRunning < aservice->minProcs)
661 /* check if over max quota */
662 if (aservice->nRequestsRunning >= aservice->maxProcs)
665 /* otherwise, can use only if there are enough to allow everyone
666 * to go to their min quota after this guy starts.
668 MUTEX_ENTER(&rx_quota_mutex);
669 if (rxi_availProcs > rxi_minDeficit)
671 MUTEX_EXIT(&rx_quota_mutex);
674 #endif /* RX_ENABLE_LOCKS */
677 /* Called by rx_StartServer to start up lwp's to service calls.
678 NExistingProcs gives the number of procs already existing, and which
679 therefore needn't be created. */
681 rxi_StartServerProcs(int nExistingProcs)
683 struct rx_service *service;
688 /* For each service, reserve N processes, where N is the "minimum"
689 * number of processes that MUST be able to execute a request in parallel,
690 * at any time, for that process. Also compute the maximum difference
691 * between any service's maximum number of processes that can run
692 * (i.e. the maximum number that ever will be run, and a guarantee
693 * that this number will run if other services aren't running), and its
694 * minimum number. The result is the extra number of processes that
695 * we need in order to provide the latter guarantee */
696 for (i = 0; i < RX_MAX_SERVICES; i++) {
698 service = rx_services[i];
699 if (service == (struct rx_service *)0)
701 nProcs += service->minProcs;
702 diff = service->maxProcs - service->minProcs;
706 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
707 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
708 for (i = 0; i < nProcs; i++) {
709 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
715 /* This routine is only required on Windows */
717 rx_StartClientThread(void)
719 #ifdef AFS_PTHREAD_ENV
721 pid = pthread_self();
722 #endif /* AFS_PTHREAD_ENV */
724 #endif /* AFS_NT40_ENV */
726 /* This routine must be called if any services are exported. If the
727 * donateMe flag is set, the calling process is donated to the server
730 rx_StartServer(int donateMe)
732 struct rx_service *service;
738 /* Start server processes, if necessary (exact function is dependent
739 * on the implementation environment--kernel or user space). DonateMe
740 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
741 * case, one less new proc will be created rx_StartServerProcs.
743 rxi_StartServerProcs(donateMe);
745 /* count up the # of threads in minProcs, and add set the min deficit to
746 * be that value, too.
748 for (i = 0; i < RX_MAX_SERVICES; i++) {
749 service = rx_services[i];
750 if (service == (struct rx_service *)0)
752 MUTEX_ENTER(&rx_quota_mutex);
753 rxi_totalMin += service->minProcs;
754 /* below works even if a thread is running, since minDeficit would
755 * still have been decremented and later re-incremented.
757 rxi_minDeficit += service->minProcs;
758 MUTEX_EXIT(&rx_quota_mutex);
761 /* Turn on reaping of idle server connections */
762 rxi_ReapConnections(NULL, NULL, NULL);
771 #ifdef AFS_PTHREAD_ENV
773 pid = afs_pointer_to_int(pthread_self());
774 #else /* AFS_PTHREAD_ENV */
776 LWP_CurrentProcess(&pid);
777 #endif /* AFS_PTHREAD_ENV */
779 sprintf(name, "srv_%d", ++nProcs);
781 (*registerProgram) (pid, name);
783 #endif /* AFS_NT40_ENV */
784 rx_ServerProc(NULL); /* Never returns */
786 #ifdef RX_ENABLE_TSFPQ
787 /* no use leaving packets around in this thread's local queue if
788 * it isn't getting donated to the server thread pool.
790 rxi_FlushLocalPacketsTSFPQ();
791 #endif /* RX_ENABLE_TSFPQ */
795 /* Create a new client connection to the specified service, using the
796 * specified security object to implement the security model for this
798 struct rx_connection *
799 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
800 struct rx_securityClass *securityObject,
801 int serviceSecurityIndex)
805 struct rx_connection *conn;
810 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
811 "serviceSecurityIndex %d)\n",
812 ntohl(shost), ntohs(sport), sservice, securityObject,
813 serviceSecurityIndex));
815 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
816 * the case of kmem_alloc? */
817 conn = rxi_AllocConnection();
818 #ifdef RX_ENABLE_LOCKS
819 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
820 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
821 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
824 MUTEX_ENTER(&rx_connHashTable_lock);
825 cid = (rx_nextCid += RX_MAXCALLS);
826 conn->type = RX_CLIENT_CONNECTION;
828 conn->epoch = rx_epoch;
829 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
830 conn->serviceId = sservice;
831 conn->securityObject = securityObject;
832 conn->securityData = (void *) 0;
833 conn->securityIndex = serviceSecurityIndex;
834 rx_SetConnDeadTime(conn, rx_connDeadTime);
835 rx_SetConnSecondsUntilNatPing(conn, 0);
836 conn->ackRate = RX_FAST_ACK_RATE;
838 conn->specific = NULL;
839 conn->challengeEvent = NULL;
840 conn->delayedAbortEvent = NULL;
841 conn->abortCount = 0;
843 for (i = 0; i < RX_MAXCALLS; i++) {
844 conn->twind[i] = rx_initSendWindow;
845 conn->rwind[i] = rx_initReceiveWindow;
848 RXS_NewConnection(securityObject, conn);
850 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
852 conn->refCount++; /* no lock required since only this thread knows... */
853 conn->next = rx_connHashTable[hashindex];
854 rx_connHashTable[hashindex] = conn;
856 rx_atomic_inc(&rx_stats.nClientConns);
857 MUTEX_EXIT(&rx_connHashTable_lock);
863 * Ensure a connection's timeout values are valid.
865 * @param[in] conn The connection to check
867 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
868 * unless idleDeadTime and/or hardDeadTime are not set
872 rxi_CheckConnTimeouts(struct rx_connection *conn)
874 /* a connection's timeouts must have the relationship
875 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
876 * total loss of network to a peer may cause an idle timeout instead of a
877 * dead timeout, simply because the idle timeout gets hit first. Also set
878 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
879 /* this logic is slightly complicated by the fact that
880 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
882 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
883 if (conn->idleDeadTime) {
884 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
886 if (conn->hardDeadTime) {
887 if (conn->idleDeadTime) {
888 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
890 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
896 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
898 /* The idea is to set the dead time to a value that allows several
899 * keepalives to be dropped without timing out the connection. */
900 conn->secondsUntilDead = seconds;
901 rxi_CheckConnTimeouts(conn);
902 conn->secondsUntilPing = conn->secondsUntilDead / 6;
906 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
908 conn->hardDeadTime = seconds;
909 rxi_CheckConnTimeouts(conn);
913 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
915 conn->idleDeadTime = seconds;
916 rxi_CheckConnTimeouts(conn);
919 int rxi_lowPeerRefCount = 0;
920 int rxi_lowConnRefCount = 0;
923 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
924 * NOTE: must not be called with rx_connHashTable_lock held.
927 rxi_CleanupConnection(struct rx_connection *conn)
929 /* Notify the service exporter, if requested, that this connection
930 * is being destroyed */
931 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
932 (*conn->service->destroyConnProc) (conn);
934 /* Notify the security module that this connection is being destroyed */
935 RXS_DestroyConnection(conn->securityObject, conn);
937 /* If this is the last connection using the rx_peer struct, set its
938 * idle time to now. rxi_ReapConnections will reap it if it's still
939 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
941 MUTEX_ENTER(&rx_peerHashTable_lock);
942 if (conn->peer->refCount < 2) {
943 conn->peer->idleWhen = clock_Sec();
944 if (conn->peer->refCount < 1) {
945 conn->peer->refCount = 1;
946 if (rx_stats_active) {
947 MUTEX_ENTER(&rx_stats_mutex);
948 rxi_lowPeerRefCount++;
949 MUTEX_EXIT(&rx_stats_mutex);
953 conn->peer->refCount--;
954 MUTEX_EXIT(&rx_peerHashTable_lock);
958 if (conn->type == RX_SERVER_CONNECTION)
959 rx_atomic_dec(&rx_stats.nServerConns);
961 rx_atomic_dec(&rx_stats.nClientConns);
964 if (conn->specific) {
966 for (i = 0; i < conn->nSpecific; i++) {
967 if (conn->specific[i] && rxi_keyCreate_destructor[i])
968 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
969 conn->specific[i] = NULL;
971 free(conn->specific);
973 conn->specific = NULL;
977 MUTEX_DESTROY(&conn->conn_call_lock);
978 MUTEX_DESTROY(&conn->conn_data_lock);
979 CV_DESTROY(&conn->conn_call_cv);
981 rxi_FreeConnection(conn);
984 /* Destroy the specified connection */
986 rxi_DestroyConnection(struct rx_connection *conn)
988 MUTEX_ENTER(&rx_connHashTable_lock);
989 rxi_DestroyConnectionNoLock(conn);
990 /* conn should be at the head of the cleanup list */
991 if (conn == rx_connCleanup_list) {
992 rx_connCleanup_list = rx_connCleanup_list->next;
993 MUTEX_EXIT(&rx_connHashTable_lock);
994 rxi_CleanupConnection(conn);
996 #ifdef RX_ENABLE_LOCKS
998 MUTEX_EXIT(&rx_connHashTable_lock);
1000 #endif /* RX_ENABLE_LOCKS */
1004 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1006 struct rx_connection **conn_ptr;
1008 struct rx_packet *packet;
1015 MUTEX_ENTER(&conn->conn_data_lock);
1016 MUTEX_ENTER(&rx_refcnt_mutex);
1017 if (conn->refCount > 0)
1020 if (rx_stats_active) {
1021 MUTEX_ENTER(&rx_stats_mutex);
1022 rxi_lowConnRefCount++;
1023 MUTEX_EXIT(&rx_stats_mutex);
1027 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1028 /* Busy; wait till the last guy before proceeding */
1029 MUTEX_EXIT(&rx_refcnt_mutex);
1030 MUTEX_EXIT(&conn->conn_data_lock);
1035 /* If the client previously called rx_NewCall, but it is still
1036 * waiting, treat this as a running call, and wait to destroy the
1037 * connection later when the call completes. */
1038 if ((conn->type == RX_CLIENT_CONNECTION)
1039 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1040 conn->flags |= RX_CONN_DESTROY_ME;
1041 MUTEX_EXIT(&conn->conn_data_lock);
1045 MUTEX_EXIT(&rx_refcnt_mutex);
1046 MUTEX_EXIT(&conn->conn_data_lock);
1048 /* Check for extant references to this connection */
1049 for (i = 0; i < RX_MAXCALLS; i++) {
1050 struct rx_call *call = conn->call[i];
1053 if (conn->type == RX_CLIENT_CONNECTION) {
1054 MUTEX_ENTER(&call->lock);
1055 if (call->delayedAckEvent) {
1056 /* Push the final acknowledgment out now--there
1057 * won't be a subsequent call to acknowledge the
1058 * last reply packets */
1059 rxevent_Cancel(call->delayedAckEvent, call,
1060 RX_CALL_REFCOUNT_DELAY);
1061 if (call->state == RX_STATE_PRECALL
1062 || call->state == RX_STATE_ACTIVE) {
1063 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1065 rxi_AckAll(NULL, call, 0);
1068 MUTEX_EXIT(&call->lock);
1072 #ifdef RX_ENABLE_LOCKS
1074 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1075 MUTEX_EXIT(&conn->conn_data_lock);
1077 /* Someone is accessing a packet right now. */
1081 #endif /* RX_ENABLE_LOCKS */
1084 /* Don't destroy the connection if there are any call
1085 * structures still in use */
1086 MUTEX_ENTER(&conn->conn_data_lock);
1087 conn->flags |= RX_CONN_DESTROY_ME;
1088 MUTEX_EXIT(&conn->conn_data_lock);
1093 if (conn->natKeepAliveEvent) {
1094 rxi_NatKeepAliveOff(conn);
1097 if (conn->delayedAbortEvent) {
1098 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1099 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1101 MUTEX_ENTER(&conn->conn_data_lock);
1102 rxi_SendConnectionAbort(conn, packet, 0, 1);
1103 MUTEX_EXIT(&conn->conn_data_lock);
1104 rxi_FreePacket(packet);
1108 /* Remove from connection hash table before proceeding */
1110 &rx_connHashTable[CONN_HASH
1111 (peer->host, peer->port, conn->cid, conn->epoch,
1113 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1114 if (*conn_ptr == conn) {
1115 *conn_ptr = conn->next;
1119 /* if the conn that we are destroying was the last connection, then we
1120 * clear rxLastConn as well */
1121 if (rxLastConn == conn)
1124 /* Make sure the connection is completely reset before deleting it. */
1125 /* get rid of pending events that could zap us later */
1126 if (conn->challengeEvent)
1127 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1128 if (conn->checkReachEvent)
1129 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1130 if (conn->natKeepAliveEvent)
1131 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
1133 /* Add the connection to the list of destroyed connections that
1134 * need to be cleaned up. This is necessary to avoid deadlocks
1135 * in the routines we call to inform others that this connection is
1136 * being destroyed. */
1137 conn->next = rx_connCleanup_list;
1138 rx_connCleanup_list = conn;
1141 /* Externally available version */
1143 rx_DestroyConnection(struct rx_connection *conn)
1148 rxi_DestroyConnection(conn);
1153 rx_GetConnection(struct rx_connection *conn)
1158 MUTEX_ENTER(&rx_refcnt_mutex);
1160 MUTEX_EXIT(&rx_refcnt_mutex);
1164 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1165 /* Wait for the transmit queue to no longer be busy.
1166 * requires the call->lock to be held */
1168 rxi_WaitforTQBusy(struct rx_call *call) {
1169 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1170 call->flags |= RX_CALL_TQ_WAIT;
1172 #ifdef RX_ENABLE_LOCKS
1173 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1174 CV_WAIT(&call->cv_tq, &call->lock);
1175 #else /* RX_ENABLE_LOCKS */
1176 osi_rxSleep(&call->tq);
1177 #endif /* RX_ENABLE_LOCKS */
1179 if (call->tqWaiters == 0) {
1180 call->flags &= ~RX_CALL_TQ_WAIT;
1187 rxi_WakeUpTransmitQueue(struct rx_call *call)
1189 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1190 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1191 call, call->tqWaiters, call->flags));
1192 #ifdef RX_ENABLE_LOCKS
1193 osirx_AssertMine(&call->lock, "rxi_Start start");
1194 CV_BROADCAST(&call->cv_tq);
1195 #else /* RX_ENABLE_LOCKS */
1196 osi_rxWakeup(&call->tq);
1197 #endif /* RX_ENABLE_LOCKS */
1201 /* Start a new rx remote procedure call, on the specified connection.
1202 * If wait is set to 1, wait for a free call channel; otherwise return
1203 * 0. Maxtime gives the maximum number of seconds this call may take,
1204 * after rx_NewCall returns. After this time interval, a call to any
1205 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1206 * For fine grain locking, we hold the conn_call_lock in order to
1207 * to ensure that we don't get signalle after we found a call in an active
1208 * state and before we go to sleep.
1211 rx_NewCall(struct rx_connection *conn)
1214 struct rx_call *call;
1215 struct clock queueTime;
1219 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1222 clock_GetTime(&queueTime);
1224 * Check if there are others waiting for a new call.
1225 * If so, let them go first to avoid starving them.
1226 * This is a fairly simple scheme, and might not be
1227 * a complete solution for large numbers of waiters.
1229 * makeCallWaiters keeps track of the number of
1230 * threads waiting to make calls and the
1231 * RX_CONN_MAKECALL_WAITING flag bit is used to
1232 * indicate that there are indeed calls waiting.
1233 * The flag is set when the waiter is incremented.
1234 * It is only cleared when makeCallWaiters is 0.
1235 * This prevents us from accidently destroying the
1236 * connection while it is potentially about to be used.
1238 MUTEX_ENTER(&conn->conn_call_lock);
1239 MUTEX_ENTER(&conn->conn_data_lock);
1240 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1241 conn->flags |= RX_CONN_MAKECALL_WAITING;
1242 conn->makeCallWaiters++;
1243 MUTEX_EXIT(&conn->conn_data_lock);
1245 #ifdef RX_ENABLE_LOCKS
1246 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1250 MUTEX_ENTER(&conn->conn_data_lock);
1251 conn->makeCallWaiters--;
1252 if (conn->makeCallWaiters == 0)
1253 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1256 /* We are now the active thread in rx_NewCall */
1257 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1258 MUTEX_EXIT(&conn->conn_data_lock);
1263 for (i = 0; i < RX_MAXCALLS; i++) {
1264 call = conn->call[i];
1266 if (call->state == RX_STATE_DALLY) {
1267 MUTEX_ENTER(&call->lock);
1268 if (call->state == RX_STATE_DALLY) {
1270 * We are setting the state to RX_STATE_RESET to
1271 * ensure that no one else will attempt to use this
1272 * call once we drop the conn->conn_call_lock and
1273 * call->lock. We must drop the conn->conn_call_lock
1274 * before calling rxi_ResetCall because the process
1275 * of clearing the transmit queue can block for an
1276 * extended period of time. If we block while holding
1277 * the conn->conn_call_lock, then all rx_EndCall
1278 * processing will block as well. This has a detrimental
1279 * effect on overall system performance.
1281 call->state = RX_STATE_RESET;
1282 MUTEX_EXIT(&conn->conn_call_lock);
1283 MUTEX_ENTER(&rx_refcnt_mutex);
1284 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1285 MUTEX_EXIT(&rx_refcnt_mutex);
1286 rxi_ResetCall(call, 0);
1287 (*call->callNumber)++;
1288 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1292 * If we failed to be able to safely obtain the
1293 * conn->conn_call_lock we will have to drop the
1294 * call->lock to avoid a deadlock. When the call->lock
1295 * is released the state of the call can change. If it
1296 * is no longer RX_STATE_RESET then some other thread is
1299 MUTEX_EXIT(&call->lock);
1300 MUTEX_ENTER(&conn->conn_call_lock);
1301 MUTEX_ENTER(&call->lock);
1303 if (call->state == RX_STATE_RESET)
1307 * If we get here it means that after dropping
1308 * the conn->conn_call_lock and call->lock that
1309 * the call is no longer ours. If we can't find
1310 * a free call in the remaining slots we should
1311 * not go immediately to RX_CONN_MAKECALL_WAITING
1312 * because by dropping the conn->conn_call_lock
1313 * we have given up synchronization with rx_EndCall.
1314 * Instead, cycle through one more time to see if
1315 * we can find a call that can call our own.
1317 MUTEX_ENTER(&rx_refcnt_mutex);
1318 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1319 MUTEX_EXIT(&rx_refcnt_mutex);
1322 MUTEX_EXIT(&call->lock);
1325 /* rxi_NewCall returns with mutex locked */
1326 call = rxi_NewCall(conn, i);
1327 MUTEX_ENTER(&rx_refcnt_mutex);
1328 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1329 MUTEX_EXIT(&rx_refcnt_mutex);
1333 if (i < RX_MAXCALLS) {
1339 MUTEX_ENTER(&conn->conn_data_lock);
1340 conn->flags |= RX_CONN_MAKECALL_WAITING;
1341 conn->makeCallWaiters++;
1342 MUTEX_EXIT(&conn->conn_data_lock);
1344 #ifdef RX_ENABLE_LOCKS
1345 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1349 MUTEX_ENTER(&conn->conn_data_lock);
1350 conn->makeCallWaiters--;
1351 if (conn->makeCallWaiters == 0)
1352 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1353 MUTEX_EXIT(&conn->conn_data_lock);
1355 /* Client is initially in send mode */
1356 call->state = RX_STATE_ACTIVE;
1357 call->error = conn->error;
1359 call->mode = RX_MODE_ERROR;
1361 call->mode = RX_MODE_SENDING;
1363 /* remember start time for call in case we have hard dead time limit */
1364 call->queueTime = queueTime;
1365 clock_GetTime(&call->startTime);
1366 hzero(call->bytesSent);
1367 hzero(call->bytesRcvd);
1369 /* Turn on busy protocol. */
1370 rxi_KeepAliveOn(call);
1372 /* Attempt MTU discovery */
1373 rxi_GrowMTUOn(call);
1376 * We are no longer the active thread in rx_NewCall
1378 MUTEX_ENTER(&conn->conn_data_lock);
1379 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1380 MUTEX_EXIT(&conn->conn_data_lock);
1383 * Wake up anyone else who might be giving us a chance to
1384 * run (see code above that avoids resource starvation).
1386 #ifdef RX_ENABLE_LOCKS
1387 CV_BROADCAST(&conn->conn_call_cv);
1391 MUTEX_EXIT(&conn->conn_call_lock);
1393 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1394 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1395 osi_Panic("rx_NewCall call about to be used without an empty tq");
1397 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1399 MUTEX_EXIT(&call->lock);
1402 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1407 rxi_HasActiveCalls(struct rx_connection *aconn)
1410 struct rx_call *tcall;
1414 for (i = 0; i < RX_MAXCALLS; i++) {
1415 if ((tcall = aconn->call[i])) {
1416 if ((tcall->state == RX_STATE_ACTIVE)
1417 || (tcall->state == RX_STATE_PRECALL)) {
1428 rxi_GetCallNumberVector(struct rx_connection *aconn,
1429 afs_int32 * aint32s)
1432 struct rx_call *tcall;
1436 for (i = 0; i < RX_MAXCALLS; i++) {
1437 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1438 aint32s[i] = aconn->callNumber[i] + 1;
1440 aint32s[i] = aconn->callNumber[i];
1447 rxi_SetCallNumberVector(struct rx_connection *aconn,
1448 afs_int32 * aint32s)
1451 struct rx_call *tcall;
1455 for (i = 0; i < RX_MAXCALLS; i++) {
1456 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1457 aconn->callNumber[i] = aint32s[i] - 1;
1459 aconn->callNumber[i] = aint32s[i];
1465 /* Advertise a new service. A service is named locally by a UDP port
1466 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1469 char *serviceName; Name for identification purposes (e.g. the
1470 service name might be used for probing for
1473 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1474 char *serviceName, struct rx_securityClass **securityObjects,
1475 int nSecurityObjects,
1476 afs_int32(*serviceProc) (struct rx_call * acall))
1478 osi_socket socket = OSI_NULLSOCKET;
1479 struct rx_service *tservice;
1485 if (serviceId == 0) {
1487 "rx_NewService: service id for service %s is not non-zero.\n",
1494 "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",
1502 tservice = rxi_AllocService();
1505 #ifdef RX_ENABLE_LOCKS
1506 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1509 for (i = 0; i < RX_MAX_SERVICES; i++) {
1510 struct rx_service *service = rx_services[i];
1512 if (port == service->servicePort && host == service->serviceHost) {
1513 if (service->serviceId == serviceId) {
1514 /* The identical service has already been
1515 * installed; if the caller was intending to
1516 * change the security classes used by this
1517 * service, he/she loses. */
1519 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1520 serviceName, serviceId, service->serviceName);
1522 rxi_FreeService(tservice);
1525 /* Different service, same port: re-use the socket
1526 * which is bound to the same port */
1527 socket = service->socket;
1530 if (socket == OSI_NULLSOCKET) {
1531 /* If we don't already have a socket (from another
1532 * service on same port) get a new one */
1533 socket = rxi_GetHostUDPSocket(host, port);
1534 if (socket == OSI_NULLSOCKET) {
1536 rxi_FreeService(tservice);
1541 service->socket = socket;
1542 service->serviceHost = host;
1543 service->servicePort = port;
1544 service->serviceId = serviceId;
1545 service->serviceName = serviceName;
1546 service->nSecurityObjects = nSecurityObjects;
1547 service->securityObjects = securityObjects;
1548 service->minProcs = 0;
1549 service->maxProcs = 1;
1550 service->idleDeadTime = 60;
1551 service->idleDeadErr = 0;
1552 service->connDeadTime = rx_connDeadTime;
1553 service->executeRequestProc = serviceProc;
1554 service->checkReach = 0;
1555 service->nSpecific = 0;
1556 service->specific = NULL;
1557 rx_services[i] = service; /* not visible until now */
1563 rxi_FreeService(tservice);
1564 (osi_Msg "rx_NewService: cannot support > %d services\n",
1569 /* Set configuration options for all of a service's security objects */
1572 rx_SetSecurityConfiguration(struct rx_service *service,
1573 rx_securityConfigVariables type,
1577 for (i = 0; i<service->nSecurityObjects; i++) {
1578 if (service->securityObjects[i]) {
1579 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1587 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1588 struct rx_securityClass **securityObjects, int nSecurityObjects,
1589 afs_int32(*serviceProc) (struct rx_call * acall))
1591 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1594 /* Generic request processing loop. This routine should be called
1595 * by the implementation dependent rx_ServerProc. If socketp is
1596 * non-null, it will be set to the file descriptor that this thread
1597 * is now listening on. If socketp is null, this routine will never
1600 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1602 struct rx_call *call;
1604 struct rx_service *tservice = NULL;
1611 call = rx_GetCall(threadID, tservice, socketp);
1612 if (socketp && *socketp != OSI_NULLSOCKET) {
1613 /* We are now a listener thread */
1618 /* if server is restarting( typically smooth shutdown) then do not
1619 * allow any new calls.
1622 if (rx_tranquil && (call != NULL)) {
1626 MUTEX_ENTER(&call->lock);
1628 rxi_CallError(call, RX_RESTARTING);
1629 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1631 MUTEX_EXIT(&call->lock);
1635 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1636 #ifdef RX_ENABLE_LOCKS
1638 #endif /* RX_ENABLE_LOCKS */
1639 afs_termState = AFSOP_STOP_AFS;
1640 afs_osi_Wakeup(&afs_termState);
1641 #ifdef RX_ENABLE_LOCKS
1643 #endif /* RX_ENABLE_LOCKS */
1648 tservice = call->conn->service;
1650 if (tservice->beforeProc)
1651 (*tservice->beforeProc) (call);
1653 code = tservice->executeRequestProc(call);
1655 if (tservice->afterProc)
1656 (*tservice->afterProc) (call, code);
1658 rx_EndCall(call, code);
1659 if (rx_stats_active) {
1660 MUTEX_ENTER(&rx_stats_mutex);
1662 MUTEX_EXIT(&rx_stats_mutex);
1669 rx_WakeupServerProcs(void)
1671 struct rx_serverQueueEntry *np, *tqp;
1675 MUTEX_ENTER(&rx_serverPool_lock);
1677 #ifdef RX_ENABLE_LOCKS
1678 if (rx_waitForPacket)
1679 CV_BROADCAST(&rx_waitForPacket->cv);
1680 #else /* RX_ENABLE_LOCKS */
1681 if (rx_waitForPacket)
1682 osi_rxWakeup(rx_waitForPacket);
1683 #endif /* RX_ENABLE_LOCKS */
1684 MUTEX_ENTER(&freeSQEList_lock);
1685 for (np = rx_FreeSQEList; np; np = tqp) {
1686 tqp = *(struct rx_serverQueueEntry **)np;
1687 #ifdef RX_ENABLE_LOCKS
1688 CV_BROADCAST(&np->cv);
1689 #else /* RX_ENABLE_LOCKS */
1691 #endif /* RX_ENABLE_LOCKS */
1693 MUTEX_EXIT(&freeSQEList_lock);
1694 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1695 #ifdef RX_ENABLE_LOCKS
1696 CV_BROADCAST(&np->cv);
1697 #else /* RX_ENABLE_LOCKS */
1699 #endif /* RX_ENABLE_LOCKS */
1701 MUTEX_EXIT(&rx_serverPool_lock);
1706 * One thing that seems to happen is that all the server threads get
1707 * tied up on some empty or slow call, and then a whole bunch of calls
1708 * arrive at once, using up the packet pool, so now there are more
1709 * empty calls. The most critical resources here are server threads
1710 * and the free packet pool. The "doreclaim" code seems to help in
1711 * general. I think that eventually we arrive in this state: there
1712 * are lots of pending calls which do have all their packets present,
1713 * so they won't be reclaimed, are multi-packet calls, so they won't
1714 * be scheduled until later, and thus are tying up most of the free
1715 * packet pool for a very long time.
1717 * 1. schedule multi-packet calls if all the packets are present.
1718 * Probably CPU-bound operation, useful to return packets to pool.
1719 * Do what if there is a full window, but the last packet isn't here?
1720 * 3. preserve one thread which *only* runs "best" calls, otherwise
1721 * it sleeps and waits for that type of call.
1722 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1723 * the current dataquota business is badly broken. The quota isn't adjusted
1724 * to reflect how many packets are presently queued for a running call.
1725 * So, when we schedule a queued call with a full window of packets queued
1726 * up for it, that *should* free up a window full of packets for other 2d-class
1727 * calls to be able to use from the packet pool. But it doesn't.
1729 * NB. Most of the time, this code doesn't run -- since idle server threads
1730 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1731 * as a new call arrives.
1733 /* Sleep until a call arrives. Returns a pointer to the call, ready
1734 * for an rx_Read. */
1735 #ifdef RX_ENABLE_LOCKS
1737 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1739 struct rx_serverQueueEntry *sq;
1740 struct rx_call *call = (struct rx_call *)0;
1741 struct rx_service *service = NULL;
1744 MUTEX_ENTER(&freeSQEList_lock);
1746 if ((sq = rx_FreeSQEList)) {
1747 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1748 MUTEX_EXIT(&freeSQEList_lock);
1749 } else { /* otherwise allocate a new one and return that */
1750 MUTEX_EXIT(&freeSQEList_lock);
1751 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1752 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1753 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1756 MUTEX_ENTER(&rx_serverPool_lock);
1757 if (cur_service != NULL) {
1758 ReturnToServerPool(cur_service);
1761 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1762 struct rx_call *tcall, *ncall, *choice2 = NULL;
1764 /* Scan for eligible incoming calls. A call is not eligible
1765 * if the maximum number of calls for its service type are
1766 * already executing */
1767 /* One thread will process calls FCFS (to prevent starvation),
1768 * while the other threads may run ahead looking for calls which
1769 * have all their input data available immediately. This helps
1770 * keep threads from blocking, waiting for data from the client. */
1771 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1772 service = tcall->conn->service;
1773 if (!QuotaOK(service)) {
1776 MUTEX_ENTER(&rx_pthread_mutex);
1777 if (tno == rxi_fcfs_thread_num
1778 || !tcall->queue_item_header.next) {
1779 MUTEX_EXIT(&rx_pthread_mutex);
1780 /* If we're the fcfs thread , then we'll just use
1781 * this call. If we haven't been able to find an optimal
1782 * choice, and we're at the end of the list, then use a
1783 * 2d choice if one has been identified. Otherwise... */
1784 call = (choice2 ? choice2 : tcall);
1785 service = call->conn->service;
1787 MUTEX_EXIT(&rx_pthread_mutex);
1788 if (!queue_IsEmpty(&tcall->rq)) {
1789 struct rx_packet *rp;
1790 rp = queue_First(&tcall->rq, rx_packet);
1791 if (rp->header.seq == 1) {
1793 || (rp->header.flags & RX_LAST_PACKET)) {
1795 } else if (rxi_2dchoice && !choice2
1796 && !(tcall->flags & RX_CALL_CLEARED)
1797 && (tcall->rprev > rxi_HardAckRate)) {
1807 ReturnToServerPool(service);
1814 MUTEX_EXIT(&rx_serverPool_lock);
1815 MUTEX_ENTER(&call->lock);
1817 if (call->flags & RX_CALL_WAIT_PROC) {
1818 call->flags &= ~RX_CALL_WAIT_PROC;
1819 rx_atomic_dec(&rx_nWaiting);
1822 if (call->state != RX_STATE_PRECALL || call->error) {
1823 MUTEX_EXIT(&call->lock);
1824 MUTEX_ENTER(&rx_serverPool_lock);
1825 ReturnToServerPool(service);
1830 if (queue_IsEmpty(&call->rq)
1831 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1832 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1834 CLEAR_CALL_QUEUE_LOCK(call);
1837 /* If there are no eligible incoming calls, add this process
1838 * to the idle server queue, to wait for one */
1842 *socketp = OSI_NULLSOCKET;
1844 sq->socketp = socketp;
1845 queue_Append(&rx_idleServerQueue, sq);
1846 #ifndef AFS_AIX41_ENV
1847 rx_waitForPacket = sq;
1849 rx_waitingForPacket = sq;
1850 #endif /* AFS_AIX41_ENV */
1852 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1854 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1855 MUTEX_EXIT(&rx_serverPool_lock);
1856 return (struct rx_call *)0;
1859 } while (!(call = sq->newcall)
1860 && !(socketp && *socketp != OSI_NULLSOCKET));
1861 MUTEX_EXIT(&rx_serverPool_lock);
1863 MUTEX_ENTER(&call->lock);
1869 MUTEX_ENTER(&freeSQEList_lock);
1870 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1871 rx_FreeSQEList = sq;
1872 MUTEX_EXIT(&freeSQEList_lock);
1875 clock_GetTime(&call->startTime);
1876 call->state = RX_STATE_ACTIVE;
1877 call->mode = RX_MODE_RECEIVING;
1878 #ifdef RX_KERNEL_TRACE
1879 if (ICL_SETACTIVE(afs_iclSetp)) {
1880 int glockOwner = ISAFS_GLOCK();
1883 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1884 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1891 rxi_calltrace(RX_CALL_START, call);
1892 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
1893 call->conn->service->servicePort, call->conn->service->serviceId,
1896 MUTEX_EXIT(&call->lock);
1897 MUTEX_ENTER(&rx_refcnt_mutex);
1898 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1899 MUTEX_EXIT(&rx_refcnt_mutex);
1901 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1906 #else /* RX_ENABLE_LOCKS */
1908 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1910 struct rx_serverQueueEntry *sq;
1911 struct rx_call *call = (struct rx_call *)0, *choice2;
1912 struct rx_service *service = NULL;
1916 MUTEX_ENTER(&freeSQEList_lock);
1918 if ((sq = rx_FreeSQEList)) {
1919 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1920 MUTEX_EXIT(&freeSQEList_lock);
1921 } else { /* otherwise allocate a new one and return that */
1922 MUTEX_EXIT(&freeSQEList_lock);
1923 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1924 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1925 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1927 MUTEX_ENTER(&sq->lock);
1929 if (cur_service != NULL) {
1930 cur_service->nRequestsRunning--;
1931 MUTEX_ENTER(&rx_quota_mutex);
1932 if (cur_service->nRequestsRunning < cur_service->minProcs)
1935 MUTEX_EXIT(&rx_quota_mutex);
1937 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1938 struct rx_call *tcall, *ncall;
1939 /* Scan for eligible incoming calls. A call is not eligible
1940 * if the maximum number of calls for its service type are
1941 * already executing */
1942 /* One thread will process calls FCFS (to prevent starvation),
1943 * while the other threads may run ahead looking for calls which
1944 * have all their input data available immediately. This helps
1945 * keep threads from blocking, waiting for data from the client. */
1946 choice2 = (struct rx_call *)0;
1947 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1948 service = tcall->conn->service;
1949 if (QuotaOK(service)) {
1950 MUTEX_ENTER(&rx_pthread_mutex);
1951 if (tno == rxi_fcfs_thread_num
1952 || !tcall->queue_item_header.next) {
1953 MUTEX_EXIT(&rx_pthread_mutex);
1954 /* If we're the fcfs thread, then we'll just use
1955 * this call. If we haven't been able to find an optimal
1956 * choice, and we're at the end of the list, then use a
1957 * 2d choice if one has been identified. Otherwise... */
1958 call = (choice2 ? choice2 : tcall);
1959 service = call->conn->service;
1961 MUTEX_EXIT(&rx_pthread_mutex);
1962 if (!queue_IsEmpty(&tcall->rq)) {
1963 struct rx_packet *rp;
1964 rp = queue_First(&tcall->rq, rx_packet);
1965 if (rp->header.seq == 1
1967 || (rp->header.flags & RX_LAST_PACKET))) {
1969 } else if (rxi_2dchoice && !choice2
1970 && !(tcall->flags & RX_CALL_CLEARED)
1971 && (tcall->rprev > rxi_HardAckRate)) {
1985 /* we can't schedule a call if there's no data!!! */
1986 /* send an ack if there's no data, if we're missing the
1987 * first packet, or we're missing something between first
1988 * and last -- there's a "hole" in the incoming data. */
1989 if (queue_IsEmpty(&call->rq)
1990 || queue_First(&call->rq, rx_packet)->header.seq != 1
1991 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1992 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1994 call->flags &= (~RX_CALL_WAIT_PROC);
1995 service->nRequestsRunning++;
1996 /* just started call in minProcs pool, need fewer to maintain
1998 MUTEX_ENTER(&rx_quota_mutex);
1999 if (service->nRequestsRunning <= service->minProcs)
2002 MUTEX_EXIT(&rx_quota_mutex);
2003 rx_atomic_dec(&rx_nWaiting);
2004 /* MUTEX_EXIT(&call->lock); */
2006 /* If there are no eligible incoming calls, add this process
2007 * to the idle server queue, to wait for one */
2010 *socketp = OSI_NULLSOCKET;
2012 sq->socketp = socketp;
2013 queue_Append(&rx_idleServerQueue, sq);
2017 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2019 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2020 return (struct rx_call *)0;
2023 } while (!(call = sq->newcall)
2024 && !(socketp && *socketp != OSI_NULLSOCKET));
2026 MUTEX_EXIT(&sq->lock);
2028 MUTEX_ENTER(&freeSQEList_lock);
2029 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2030 rx_FreeSQEList = sq;
2031 MUTEX_EXIT(&freeSQEList_lock);
2034 clock_GetTime(&call->startTime);
2035 call->state = RX_STATE_ACTIVE;
2036 call->mode = RX_MODE_RECEIVING;
2037 #ifdef RX_KERNEL_TRACE
2038 if (ICL_SETACTIVE(afs_iclSetp)) {
2039 int glockOwner = ISAFS_GLOCK();
2042 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2043 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2050 rxi_calltrace(RX_CALL_START, call);
2051 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2052 call->conn->service->servicePort, call->conn->service->serviceId,
2055 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2062 #endif /* RX_ENABLE_LOCKS */
2066 /* Establish a procedure to be called when a packet arrives for a
2067 * call. This routine will be called at most once after each call,
2068 * and will also be called if there is an error condition on the or
2069 * the call is complete. Used by multi rx to build a selection
2070 * function which determines which of several calls is likely to be a
2071 * good one to read from.
2072 * NOTE: the way this is currently implemented it is probably only a
2073 * good idea to (1) use it immediately after a newcall (clients only)
2074 * and (2) only use it once. Other uses currently void your warranty
2077 rx_SetArrivalProc(struct rx_call *call,
2078 void (*proc) (struct rx_call * call,
2081 void * handle, int arg)
2083 call->arrivalProc = proc;
2084 call->arrivalProcHandle = handle;
2085 call->arrivalProcArg = arg;
2088 /* Call is finished (possibly prematurely). Return rc to the peer, if
2089 * appropriate, and return the final error code from the conversation
2093 rx_EndCall(struct rx_call *call, afs_int32 rc)
2095 struct rx_connection *conn = call->conn;
2099 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2100 call, rc, call->error, call->abortCode));
2103 MUTEX_ENTER(&call->lock);
2105 if (rc == 0 && call->error == 0) {
2106 call->abortCode = 0;
2107 call->abortCount = 0;
2110 call->arrivalProc = (void (*)())0;
2111 if (rc && call->error == 0) {
2112 rxi_CallError(call, rc);
2113 call->mode = RX_MODE_ERROR;
2114 /* Send an abort message to the peer if this error code has
2115 * only just been set. If it was set previously, assume the
2116 * peer has already been sent the error code or will request it
2118 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2120 if (conn->type == RX_SERVER_CONNECTION) {
2121 /* Make sure reply or at least dummy reply is sent */
2122 if (call->mode == RX_MODE_RECEIVING) {
2123 MUTEX_EXIT(&call->lock);
2124 rxi_WriteProc(call, 0, 0);
2125 MUTEX_ENTER(&call->lock);
2127 if (call->mode == RX_MODE_SENDING) {
2128 MUTEX_EXIT(&call->lock);
2129 rxi_FlushWrite(call);
2130 MUTEX_ENTER(&call->lock);
2132 rxi_calltrace(RX_CALL_END, call);
2133 /* Call goes to hold state until reply packets are acknowledged */
2134 if (call->tfirst + call->nSoftAcked < call->tnext) {
2135 call->state = RX_STATE_HOLD;
2137 call->state = RX_STATE_DALLY;
2138 rxi_ClearTransmitQueue(call, 0);
2139 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
2140 rxevent_Cancel(call->keepAliveEvent, call,
2141 RX_CALL_REFCOUNT_ALIVE);
2143 } else { /* Client connection */
2145 /* Make sure server receives input packets, in the case where
2146 * no reply arguments are expected */
2147 if ((call->mode == RX_MODE_SENDING)
2148 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2149 MUTEX_EXIT(&call->lock);
2150 (void)rxi_ReadProc(call, &dummy, 1);
2151 MUTEX_ENTER(&call->lock);
2154 /* If we had an outstanding delayed ack, be nice to the server
2155 * and force-send it now.
2157 if (call->delayedAckEvent) {
2158 rxevent_Cancel(call->delayedAckEvent, call,
2159 RX_CALL_REFCOUNT_DELAY);
2160 call->delayedAckEvent = NULL;
2161 rxi_SendDelayedAck(NULL, call, NULL);
2164 /* We need to release the call lock since it's lower than the
2165 * conn_call_lock and we don't want to hold the conn_call_lock
2166 * over the rx_ReadProc call. The conn_call_lock needs to be held
2167 * here for the case where rx_NewCall is perusing the calls on
2168 * the connection structure. We don't want to signal until
2169 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2170 * have checked this call, found it active and by the time it
2171 * goes to sleep, will have missed the signal.
2173 MUTEX_EXIT(&call->lock);
2174 MUTEX_ENTER(&conn->conn_call_lock);
2175 MUTEX_ENTER(&call->lock);
2176 MUTEX_ENTER(&conn->conn_data_lock);
2177 conn->flags |= RX_CONN_BUSY;
2178 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2179 MUTEX_EXIT(&conn->conn_data_lock);
2180 #ifdef RX_ENABLE_LOCKS
2181 CV_BROADCAST(&conn->conn_call_cv);
2186 #ifdef RX_ENABLE_LOCKS
2188 MUTEX_EXIT(&conn->conn_data_lock);
2190 #endif /* RX_ENABLE_LOCKS */
2191 call->state = RX_STATE_DALLY;
2193 error = call->error;
2195 /* currentPacket, nLeft, and NFree must be zeroed here, because
2196 * ResetCall cannot: ResetCall may be called at splnet(), in the
2197 * kernel version, and may interrupt the macros rx_Read or
2198 * rx_Write, which run at normal priority for efficiency. */
2199 if (call->currentPacket) {
2200 #ifdef RX_TRACK_PACKETS
2201 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2203 rxi_FreePacket(call->currentPacket);
2204 call->currentPacket = (struct rx_packet *)0;
2207 call->nLeft = call->nFree = call->curlen = 0;
2209 /* Free any packets from the last call to ReadvProc/WritevProc */
2210 #ifdef RXDEBUG_PACKET
2212 #endif /* RXDEBUG_PACKET */
2213 rxi_FreePackets(0, &call->iovq);
2214 MUTEX_EXIT(&call->lock);
2216 MUTEX_ENTER(&rx_refcnt_mutex);
2217 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2218 MUTEX_EXIT(&rx_refcnt_mutex);
2219 if (conn->type == RX_CLIENT_CONNECTION) {
2220 MUTEX_ENTER(&conn->conn_data_lock);
2221 conn->flags &= ~RX_CONN_BUSY;
2222 MUTEX_EXIT(&conn->conn_data_lock);
2223 MUTEX_EXIT(&conn->conn_call_lock);
2227 * Map errors to the local host's errno.h format.
2229 error = ntoh_syserr_conv(error);
2233 #if !defined(KERNEL)
2235 /* Call this routine when shutting down a server or client (especially
2236 * clients). This will allow Rx to gracefully garbage collect server
2237 * connections, and reduce the number of retries that a server might
2238 * make to a dead client.
2239 * This is not quite right, since some calls may still be ongoing and
2240 * we can't lock them to destroy them. */
2244 struct rx_connection **conn_ptr, **conn_end;
2248 if (rxinit_status == 1) {
2250 return; /* Already shutdown. */
2252 rxi_DeleteCachedConnections();
2253 if (rx_connHashTable) {
2254 MUTEX_ENTER(&rx_connHashTable_lock);
2255 for (conn_ptr = &rx_connHashTable[0], conn_end =
2256 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2258 struct rx_connection *conn, *next;
2259 for (conn = *conn_ptr; conn; conn = next) {
2261 if (conn->type == RX_CLIENT_CONNECTION) {
2262 MUTEX_ENTER(&rx_refcnt_mutex);
2264 MUTEX_EXIT(&rx_refcnt_mutex);
2265 #ifdef RX_ENABLE_LOCKS
2266 rxi_DestroyConnectionNoLock(conn);
2267 #else /* RX_ENABLE_LOCKS */
2268 rxi_DestroyConnection(conn);
2269 #endif /* RX_ENABLE_LOCKS */
2273 #ifdef RX_ENABLE_LOCKS
2274 while (rx_connCleanup_list) {
2275 struct rx_connection *conn;
2276 conn = rx_connCleanup_list;
2277 rx_connCleanup_list = rx_connCleanup_list->next;
2278 MUTEX_EXIT(&rx_connHashTable_lock);
2279 rxi_CleanupConnection(conn);
2280 MUTEX_ENTER(&rx_connHashTable_lock);
2282 MUTEX_EXIT(&rx_connHashTable_lock);
2283 #endif /* RX_ENABLE_LOCKS */
2288 afs_winsockCleanup();
2296 /* if we wakeup packet waiter too often, can get in loop with two
2297 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2299 rxi_PacketsUnWait(void)
2301 if (!rx_waitingForPackets) {
2305 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2306 return; /* still over quota */
2309 rx_waitingForPackets = 0;
2310 #ifdef RX_ENABLE_LOCKS
2311 CV_BROADCAST(&rx_waitingForPackets_cv);
2313 osi_rxWakeup(&rx_waitingForPackets);
2319 /* ------------------Internal interfaces------------------------- */
2321 /* Return this process's service structure for the
2322 * specified socket and service */
2324 rxi_FindService(osi_socket socket, u_short serviceId)
2326 struct rx_service **sp;
2327 for (sp = &rx_services[0]; *sp; sp++) {
2328 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2334 #ifdef RXDEBUG_PACKET
2335 #ifdef KDUMP_RX_LOCK
2336 static struct rx_call_rx_lock *rx_allCallsp = 0;
2338 static struct rx_call *rx_allCallsp = 0;
2340 #endif /* RXDEBUG_PACKET */
2342 /* Allocate a call structure, for the indicated channel of the
2343 * supplied connection. The mode and state of the call must be set by
2344 * the caller. Returns the call with mutex locked. */
2346 rxi_NewCall(struct rx_connection *conn, int channel)
2348 struct rx_call *call;
2349 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2350 struct rx_call *cp; /* Call pointer temp */
2351 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2352 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2354 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2356 /* Grab an existing call structure, or allocate a new one.
2357 * Existing call structures are assumed to have been left reset by
2359 MUTEX_ENTER(&rx_freeCallQueue_lock);
2361 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2363 * EXCEPT that the TQ might not yet be cleared out.
2364 * Skip over those with in-use TQs.
2367 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2368 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2374 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2375 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2376 call = queue_First(&rx_freeCallQueue, rx_call);
2377 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2379 if (rx_stats_active)
2380 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2381 MUTEX_EXIT(&rx_freeCallQueue_lock);
2382 MUTEX_ENTER(&call->lock);
2383 CLEAR_CALL_QUEUE_LOCK(call);
2384 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2385 /* Now, if TQ wasn't cleared earlier, do it now. */
2386 rxi_WaitforTQBusy(call);
2387 if (call->flags & RX_CALL_TQ_CLEARME) {
2388 rxi_ClearTransmitQueue(call, 1);
2389 /*queue_Init(&call->tq);*/
2391 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2392 /* Bind the call to its connection structure */
2394 rxi_ResetCall(call, 1);
2397 call = rxi_Alloc(sizeof(struct rx_call));
2398 #ifdef RXDEBUG_PACKET
2399 call->allNextp = rx_allCallsp;
2400 rx_allCallsp = call;
2402 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2403 #else /* RXDEBUG_PACKET */
2404 rx_atomic_inc(&rx_stats.nCallStructs);
2405 #endif /* RXDEBUG_PACKET */
2407 MUTEX_EXIT(&rx_freeCallQueue_lock);
2408 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2409 MUTEX_ENTER(&call->lock);
2410 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2411 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2412 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2414 /* Initialize once-only items */
2415 queue_Init(&call->tq);
2416 queue_Init(&call->rq);
2417 queue_Init(&call->iovq);
2418 #ifdef RXDEBUG_PACKET
2419 call->rqc = call->tqc = call->iovqc = 0;
2420 #endif /* RXDEBUG_PACKET */
2421 /* Bind the call to its connection structure (prereq for reset) */
2423 rxi_ResetCall(call, 1);
2425 call->channel = channel;
2426 call->callNumber = &conn->callNumber[channel];
2427 call->rwind = conn->rwind[channel];
2428 call->twind = conn->twind[channel];
2429 /* Note that the next expected call number is retained (in
2430 * conn->callNumber[i]), even if we reallocate the call structure
2432 conn->call[channel] = call;
2433 /* if the channel's never been used (== 0), we should start at 1, otherwise
2434 * the call number is valid from the last time this channel was used */
2435 if (*call->callNumber == 0)
2436 *call->callNumber = 1;
2441 /* A call has been inactive long enough that so we can throw away
2442 * state, including the call structure, which is placed on the call
2445 * call->lock amd rx_refcnt_mutex are held upon entry.
2446 * haveCTLock is set when called from rxi_ReapConnections.
2449 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2451 int channel = call->channel;
2452 struct rx_connection *conn = call->conn;
2455 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2456 (*call->callNumber)++;
2458 * We are setting the state to RX_STATE_RESET to
2459 * ensure that no one else will attempt to use this
2460 * call once we drop the refcnt lock. We must drop
2461 * the refcnt lock before calling rxi_ResetCall
2462 * because it cannot be held across acquiring the
2463 * freepktQ lock. NewCall does the same.
2465 call->state = RX_STATE_RESET;
2466 MUTEX_EXIT(&rx_refcnt_mutex);
2467 rxi_ResetCall(call, 0);
2468 call->conn->call[channel] = (struct rx_call *)0;
2470 MUTEX_ENTER(&rx_freeCallQueue_lock);
2471 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2472 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2473 /* A call may be free even though its transmit queue is still in use.
2474 * Since we search the call list from head to tail, put busy calls at
2475 * the head of the list, and idle calls at the tail.
2477 if (call->flags & RX_CALL_TQ_BUSY)
2478 queue_Prepend(&rx_freeCallQueue, call);
2480 queue_Append(&rx_freeCallQueue, call);
2481 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2482 queue_Append(&rx_freeCallQueue, call);
2483 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2484 if (rx_stats_active)
2485 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2486 MUTEX_EXIT(&rx_freeCallQueue_lock);
2488 /* Destroy the connection if it was previously slated for
2489 * destruction, i.e. the Rx client code previously called
2490 * rx_DestroyConnection (client connections), or
2491 * rxi_ReapConnections called the same routine (server
2492 * connections). Only do this, however, if there are no
2493 * outstanding calls. Note that for fine grain locking, there appears
2494 * to be a deadlock in that rxi_FreeCall has a call locked and
2495 * DestroyConnectionNoLock locks each call in the conn. But note a
2496 * few lines up where we have removed this call from the conn.
2497 * If someone else destroys a connection, they either have no
2498 * call lock held or are going through this section of code.
2500 MUTEX_ENTER(&conn->conn_data_lock);
2501 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2502 MUTEX_ENTER(&rx_refcnt_mutex);
2504 MUTEX_EXIT(&rx_refcnt_mutex);
2505 MUTEX_EXIT(&conn->conn_data_lock);
2506 #ifdef RX_ENABLE_LOCKS
2508 rxi_DestroyConnectionNoLock(conn);
2510 rxi_DestroyConnection(conn);
2511 #else /* RX_ENABLE_LOCKS */
2512 rxi_DestroyConnection(conn);
2513 #endif /* RX_ENABLE_LOCKS */
2515 MUTEX_EXIT(&conn->conn_data_lock);
2517 MUTEX_ENTER(&rx_refcnt_mutex);
2520 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2521 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2524 rxi_Alloc(size_t size)
2528 if (rx_stats_active) {
2529 rx_atomic_add(&rxi_Allocsize, (int) size);
2530 rx_atomic_inc(&rxi_Alloccnt);
2534 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2535 afs_osi_Alloc_NoSleep(size);
2540 osi_Panic("rxi_Alloc error");
2546 rxi_Free(void *addr, size_t size)
2548 if (rx_stats_active) {
2549 rx_atomic_sub(&rxi_Allocsize, (int) size);
2550 rx_atomic_dec(&rxi_Alloccnt);
2552 osi_Free(addr, size);
2556 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2558 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2559 struct rx_peer *next = NULL;
2563 MUTEX_ENTER(&rx_peerHashTable_lock);
2565 peer_ptr = &rx_peerHashTable[0];
2566 peer_end = &rx_peerHashTable[rx_hashTableSize];
2569 for ( ; peer_ptr < peer_end; peer_ptr++) {
2572 for ( ; peer; peer = next) {
2574 if (host == peer->host)
2579 hashIndex = PEER_HASH(host, port);
2580 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2581 if ((peer->host == host) && (peer->port == port))
2586 MUTEX_ENTER(&rx_peerHashTable_lock);
2591 MUTEX_EXIT(&rx_peerHashTable_lock);
2593 MUTEX_ENTER(&peer->peer_lock);
2594 /* We don't handle dropping below min, so don't */
2595 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2596 peer->ifMTU=MIN(mtu, peer->ifMTU);
2597 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2598 /* if we tweaked this down, need to tune our peer MTU too */
2599 peer->MTU = MIN(peer->MTU, peer->natMTU);
2600 /* if we discovered a sub-1500 mtu, degrade */
2601 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2602 peer->maxDgramPackets = 1;
2603 /* We no longer have valid peer packet information */
2604 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2605 peer->maxPacketSize = 0;
2606 MUTEX_EXIT(&peer->peer_lock);
2608 MUTEX_ENTER(&rx_peerHashTable_lock);
2610 if (host && !port) {
2612 /* pick up where we left off */
2616 MUTEX_EXIT(&rx_peerHashTable_lock);
2619 /* Find the peer process represented by the supplied (host,port)
2620 * combination. If there is no appropriate active peer structure, a
2621 * new one will be allocated and initialized
2622 * The origPeer, if set, is a pointer to a peer structure on which the
2623 * refcount will be be decremented. This is used to replace the peer
2624 * structure hanging off a connection structure */
2626 rxi_FindPeer(afs_uint32 host, u_short port,
2627 struct rx_peer *origPeer, int create)
2631 hashIndex = PEER_HASH(host, port);
2632 MUTEX_ENTER(&rx_peerHashTable_lock);
2633 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2634 if ((pp->host == host) && (pp->port == port))
2639 pp = rxi_AllocPeer(); /* This bzero's *pp */
2640 pp->host = host; /* set here or in InitPeerParams is zero */
2642 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2643 queue_Init(&pp->congestionQueue);
2644 queue_Init(&pp->rpcStats);
2645 pp->next = rx_peerHashTable[hashIndex];
2646 rx_peerHashTable[hashIndex] = pp;
2647 rxi_InitPeerParams(pp);
2648 if (rx_stats_active)
2649 rx_atomic_inc(&rx_stats.nPeerStructs);
2656 origPeer->refCount--;
2657 MUTEX_EXIT(&rx_peerHashTable_lock);
2662 /* Find the connection at (host, port) started at epoch, and with the
2663 * given connection id. Creates the server connection if necessary.
2664 * The type specifies whether a client connection or a server
2665 * connection is desired. In both cases, (host, port) specify the
2666 * peer's (host, pair) pair. Client connections are not made
2667 * automatically by this routine. The parameter socket gives the
2668 * socket descriptor on which the packet was received. This is used,
2669 * in the case of server connections, to check that *new* connections
2670 * come via a valid (port, serviceId). Finally, the securityIndex
2671 * parameter must match the existing index for the connection. If a
2672 * server connection is created, it will be created using the supplied
2673 * index, if the index is valid for this service */
2674 struct rx_connection *
2675 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2676 u_short port, u_short serviceId, afs_uint32 cid,
2677 afs_uint32 epoch, int type, u_int securityIndex)
2679 int hashindex, flag, i;
2680 struct rx_connection *conn;
2681 hashindex = CONN_HASH(host, port, cid, epoch, type);
2682 MUTEX_ENTER(&rx_connHashTable_lock);
2683 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2684 rx_connHashTable[hashindex],
2687 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2688 && (epoch == conn->epoch)) {
2689 struct rx_peer *pp = conn->peer;
2690 if (securityIndex != conn->securityIndex) {
2691 /* this isn't supposed to happen, but someone could forge a packet
2692 * like this, and there seems to be some CM bug that makes this
2693 * happen from time to time -- in which case, the fileserver
2695 MUTEX_EXIT(&rx_connHashTable_lock);
2696 return (struct rx_connection *)0;
2698 if (pp->host == host && pp->port == port)
2700 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2702 /* So what happens when it's a callback connection? */
2703 if ( /*type == RX_CLIENT_CONNECTION && */
2704 (conn->epoch & 0x80000000))
2708 /* the connection rxLastConn that was used the last time is not the
2709 ** one we are looking for now. Hence, start searching in the hash */
2711 conn = rx_connHashTable[hashindex];
2716 struct rx_service *service;
2717 if (type == RX_CLIENT_CONNECTION) {
2718 MUTEX_EXIT(&rx_connHashTable_lock);
2719 return (struct rx_connection *)0;
2721 service = rxi_FindService(socket, serviceId);
2722 if (!service || (securityIndex >= service->nSecurityObjects)
2723 || (service->securityObjects[securityIndex] == 0)) {
2724 MUTEX_EXIT(&rx_connHashTable_lock);
2725 return (struct rx_connection *)0;
2727 conn = rxi_AllocConnection(); /* This bzero's the connection */
2728 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2729 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2730 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2731 conn->next = rx_connHashTable[hashindex];
2732 rx_connHashTable[hashindex] = conn;
2733 conn->peer = rxi_FindPeer(host, port, 0, 1);
2734 conn->type = RX_SERVER_CONNECTION;
2735 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2736 conn->epoch = epoch;
2737 conn->cid = cid & RX_CIDMASK;
2738 /* conn->serial = conn->lastSerial = 0; */
2739 /* conn->timeout = 0; */
2740 conn->ackRate = RX_FAST_ACK_RATE;
2741 conn->service = service;
2742 conn->serviceId = serviceId;
2743 conn->securityIndex = securityIndex;
2744 conn->securityObject = service->securityObjects[securityIndex];
2745 conn->nSpecific = 0;
2746 conn->specific = NULL;
2747 rx_SetConnDeadTime(conn, service->connDeadTime);
2748 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2749 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2750 for (i = 0; i < RX_MAXCALLS; i++) {
2751 conn->twind[i] = rx_initSendWindow;
2752 conn->rwind[i] = rx_initReceiveWindow;
2754 /* Notify security object of the new connection */
2755 RXS_NewConnection(conn->securityObject, conn);
2756 /* XXXX Connection timeout? */
2757 if (service->newConnProc)
2758 (*service->newConnProc) (conn);
2759 if (rx_stats_active)
2760 rx_atomic_inc(&rx_stats.nServerConns);
2763 MUTEX_ENTER(&rx_refcnt_mutex);
2765 MUTEX_EXIT(&rx_refcnt_mutex);
2767 rxLastConn = conn; /* store this connection as the last conn used */
2768 MUTEX_EXIT(&rx_connHashTable_lock);
2772 /* There are two packet tracing routines available for testing and monitoring
2773 * Rx. One is called just after every packet is received and the other is
2774 * called just before every packet is sent. Received packets, have had their
2775 * headers decoded, and packets to be sent have not yet had their headers
2776 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2777 * containing the network address. Both can be modified. The return value, if
2778 * non-zero, indicates that the packet should be dropped. */
2780 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2781 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2783 /* A packet has been received off the interface. Np is the packet, socket is
2784 * the socket number it was received from (useful in determining which service
2785 * this packet corresponds to), and (host, port) reflect the host,port of the
2786 * sender. This call returns the packet to the caller if it is finished with
2787 * it, rather than de-allocating it, just as a small performance hack */
2790 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
2791 afs_uint32 host, u_short port, int *tnop,
2792 struct rx_call **newcallp)
2794 struct rx_call *call;
2795 struct rx_connection *conn;
2797 afs_uint32 currentCallNumber;
2803 struct rx_packet *tnp;
2806 /* We don't print out the packet until now because (1) the time may not be
2807 * accurate enough until now in the lwp implementation (rx_Listener only gets
2808 * the time after the packet is read) and (2) from a protocol point of view,
2809 * this is the first time the packet has been seen */
2810 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2811 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2812 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
2813 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2814 np->header.epoch, np->header.cid, np->header.callNumber,
2815 np->header.seq, np->header.flags, np));
2818 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2819 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2822 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2823 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2826 /* If an input tracer function is defined, call it with the packet and
2827 * network address. Note this function may modify its arguments. */
2828 if (rx_justReceived) {
2829 struct sockaddr_in addr;
2831 addr.sin_family = AF_INET;
2832 addr.sin_port = port;
2833 addr.sin_addr.s_addr = host;
2834 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2835 addr.sin_len = sizeof(addr);
2836 #endif /* AFS_OSF_ENV */
2837 drop = (*rx_justReceived) (np, &addr);
2838 /* drop packet if return value is non-zero */
2841 port = addr.sin_port; /* in case fcn changed addr */
2842 host = addr.sin_addr.s_addr;
2846 /* If packet was not sent by the client, then *we* must be the client */
2847 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2848 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2850 /* Find the connection (or fabricate one, if we're the server & if
2851 * necessary) associated with this packet */
2853 rxi_FindConnection(socket, host, port, np->header.serviceId,
2854 np->header.cid, np->header.epoch, type,
2855 np->header.securityIndex);
2858 /* If no connection found or fabricated, just ignore the packet.
2859 * (An argument could be made for sending an abort packet for
2864 MUTEX_ENTER(&conn->conn_data_lock);
2865 if (conn->maxSerial < np->header.serial)
2866 conn->maxSerial = np->header.serial;
2867 MUTEX_EXIT(&conn->conn_data_lock);
2869 /* If the connection is in an error state, send an abort packet and ignore
2870 * the incoming packet */
2872 /* Don't respond to an abort packet--we don't want loops! */
2873 MUTEX_ENTER(&conn->conn_data_lock);
2874 if (np->header.type != RX_PACKET_TYPE_ABORT)
2875 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2876 MUTEX_ENTER(&rx_refcnt_mutex);
2878 MUTEX_EXIT(&rx_refcnt_mutex);
2879 MUTEX_EXIT(&conn->conn_data_lock);
2883 /* Check for connection-only requests (i.e. not call specific). */
2884 if (np->header.callNumber == 0) {
2885 switch (np->header.type) {
2886 case RX_PACKET_TYPE_ABORT: {
2887 /* What if the supplied error is zero? */
2888 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2889 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
2890 rxi_ConnectionError(conn, errcode);
2891 MUTEX_ENTER(&rx_refcnt_mutex);
2893 MUTEX_EXIT(&rx_refcnt_mutex);
2896 case RX_PACKET_TYPE_CHALLENGE:
2897 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2898 MUTEX_ENTER(&rx_refcnt_mutex);
2900 MUTEX_EXIT(&rx_refcnt_mutex);
2902 case RX_PACKET_TYPE_RESPONSE:
2903 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2904 MUTEX_ENTER(&rx_refcnt_mutex);
2906 MUTEX_EXIT(&rx_refcnt_mutex);
2908 case RX_PACKET_TYPE_PARAMS:
2909 case RX_PACKET_TYPE_PARAMS + 1:
2910 case RX_PACKET_TYPE_PARAMS + 2:
2911 /* ignore these packet types for now */
2912 MUTEX_ENTER(&rx_refcnt_mutex);
2914 MUTEX_EXIT(&rx_refcnt_mutex);
2919 /* Should not reach here, unless the peer is broken: send an
2921 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2922 MUTEX_ENTER(&conn->conn_data_lock);
2923 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2924 MUTEX_ENTER(&rx_refcnt_mutex);
2926 MUTEX_EXIT(&rx_refcnt_mutex);
2927 MUTEX_EXIT(&conn->conn_data_lock);
2932 channel = np->header.cid & RX_CHANNELMASK;
2933 call = conn->call[channel];
2934 #ifdef RX_ENABLE_LOCKS
2936 MUTEX_ENTER(&call->lock);
2937 /* Test to see if call struct is still attached to conn. */
2938 if (call != conn->call[channel]) {
2940 MUTEX_EXIT(&call->lock);
2941 if (type == RX_SERVER_CONNECTION) {
2942 call = conn->call[channel];
2943 /* If we started with no call attached and there is one now,
2944 * another thread is also running this routine and has gotten
2945 * the connection channel. We should drop this packet in the tests
2946 * below. If there was a call on this connection and it's now
2947 * gone, then we'll be making a new call below.
2948 * If there was previously a call and it's now different then
2949 * the old call was freed and another thread running this routine
2950 * has created a call on this channel. One of these two threads
2951 * has a packet for the old call and the code below handles those
2955 MUTEX_ENTER(&call->lock);
2957 /* This packet can't be for this call. If the new call address is
2958 * 0 then no call is running on this channel. If there is a call
2959 * then, since this is a client connection we're getting data for
2960 * it must be for the previous call.
2962 if (rx_stats_active)
2963 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
2964 MUTEX_ENTER(&rx_refcnt_mutex);
2966 MUTEX_EXIT(&rx_refcnt_mutex);
2971 currentCallNumber = conn->callNumber[channel];
2973 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2974 if (np->header.callNumber < currentCallNumber) {
2975 if (rx_stats_active)
2976 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
2977 #ifdef RX_ENABLE_LOCKS
2979 MUTEX_EXIT(&call->lock);
2981 MUTEX_ENTER(&rx_refcnt_mutex);
2983 MUTEX_EXIT(&rx_refcnt_mutex);
2987 MUTEX_ENTER(&conn->conn_call_lock);
2988 call = rxi_NewCall(conn, channel);
2989 MUTEX_EXIT(&conn->conn_call_lock);
2990 *call->callNumber = np->header.callNumber;
2992 if (np->header.callNumber == 0)
2993 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",
2994 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
2995 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
2996 np->header.flags, np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2998 call->state = RX_STATE_PRECALL;
2999 clock_GetTime(&call->queueTime);
3000 hzero(call->bytesSent);
3001 hzero(call->bytesRcvd);
3003 * If the number of queued calls exceeds the overload
3004 * threshold then abort this call.
3006 if ((rx_BusyThreshold > 0) &&
3007 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3008 struct rx_packet *tp;
3010 rxi_CallError(call, rx_BusyError);
3011 tp = rxi_SendCallAbort(call, np, 1, 0);
3012 MUTEX_EXIT(&call->lock);
3013 MUTEX_ENTER(&rx_refcnt_mutex);
3015 MUTEX_EXIT(&rx_refcnt_mutex);
3016 if (rx_stats_active)
3017 rx_atomic_inc(&rx_stats.nBusies);
3020 rxi_KeepAliveOn(call);
3021 } else if (np->header.callNumber != currentCallNumber) {
3022 /* Wait until the transmit queue is idle before deciding
3023 * whether to reset the current call. Chances are that the
3024 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3027 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3028 if (call->state == RX_STATE_ACTIVE) {
3029 rxi_WaitforTQBusy(call);
3031 * If we entered error state while waiting,
3032 * must call rxi_CallError to permit rxi_ResetCall
3033 * to processed when the tqWaiter count hits zero.
3036 rxi_CallError(call, call->error);
3037 MUTEX_EXIT(&call->lock);
3038 MUTEX_ENTER(&rx_refcnt_mutex);
3040 MUTEX_EXIT(&rx_refcnt_mutex);
3044 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3045 /* If the new call cannot be taken right now send a busy and set
3046 * the error condition in this call, so that it terminates as
3047 * quickly as possible */
3048 if (call->state == RX_STATE_ACTIVE) {
3049 struct rx_packet *tp;
3051 rxi_CallError(call, RX_CALL_DEAD);
3052 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3054 MUTEX_EXIT(&call->lock);
3055 MUTEX_ENTER(&rx_refcnt_mutex);
3057 MUTEX_EXIT(&rx_refcnt_mutex);
3060 rxi_ResetCall(call, 0);
3061 *call->callNumber = np->header.callNumber;
3063 if (np->header.callNumber == 0)
3064 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",
3065 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3066 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3067 np->header.flags, np, np->retryTime.sec, np->retryTime.usec, np->length));
3069 call->state = RX_STATE_PRECALL;
3070 clock_GetTime(&call->queueTime);
3071 hzero(call->bytesSent);
3072 hzero(call->bytesRcvd);
3074 * If the number of queued calls exceeds the overload
3075 * threshold then abort this call.
3077 if ((rx_BusyThreshold > 0) &&
3078 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3079 struct rx_packet *tp;
3081 rxi_CallError(call, rx_BusyError);
3082 tp = rxi_SendCallAbort(call, np, 1, 0);
3083 MUTEX_EXIT(&call->lock);
3084 MUTEX_ENTER(&rx_refcnt_mutex);
3086 MUTEX_EXIT(&rx_refcnt_mutex);
3087 if (rx_stats_active)
3088 rx_atomic_inc(&rx_stats.nBusies);
3091 rxi_KeepAliveOn(call);
3093 /* Continuing call; do nothing here. */
3095 } else { /* we're the client */
3096 /* Ignore all incoming acknowledgements for calls in DALLY state */
3097 if (call && (call->state == RX_STATE_DALLY)
3098 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3099 if (rx_stats_active)
3100 rx_atomic_inc(&rx_stats.ignorePacketDally);
3101 #ifdef RX_ENABLE_LOCKS
3103 MUTEX_EXIT(&call->lock);
3106 MUTEX_ENTER(&rx_refcnt_mutex);
3108 MUTEX_EXIT(&rx_refcnt_mutex);
3112 /* Ignore anything that's not relevant to the current call. If there
3113 * isn't a current call, then no packet is relevant. */
3114 if (!call || (np->header.callNumber != currentCallNumber)) {
3115 if (rx_stats_active)
3116 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3117 #ifdef RX_ENABLE_LOCKS
3119 MUTEX_EXIT(&call->lock);
3122 MUTEX_ENTER(&rx_refcnt_mutex);
3124 MUTEX_EXIT(&rx_refcnt_mutex);
3127 /* If the service security object index stamped in the packet does not
3128 * match the connection's security index, ignore the packet */
3129 if (np->header.securityIndex != conn->securityIndex) {
3130 #ifdef RX_ENABLE_LOCKS
3131 MUTEX_EXIT(&call->lock);
3133 MUTEX_ENTER(&rx_refcnt_mutex);
3135 MUTEX_EXIT(&rx_refcnt_mutex);
3139 /* If we're receiving the response, then all transmit packets are
3140 * implicitly acknowledged. Get rid of them. */
3141 if (np->header.type == RX_PACKET_TYPE_DATA) {
3142 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3143 /* XXX Hack. Because we must release the global rx lock when
3144 * sending packets (osi_NetSend) we drop all acks while we're
3145 * traversing the tq in rxi_Start sending packets out because
3146 * packets may move to the freePacketQueue as result of being here!
3147 * So we drop these packets until we're safely out of the
3148 * traversing. Really ugly!
3149 * For fine grain RX locking, we set the acked field in the
3150 * packets and let rxi_Start remove them from the transmit queue.
3152 if (call->flags & RX_CALL_TQ_BUSY) {
3153 #ifdef RX_ENABLE_LOCKS
3154 rxi_SetAcksInTransmitQueue(call);
3156 MUTEX_ENTER(&rx_refcnt_mutex);
3158 MUTEX_EXIT(&rx_refcnt_mutex);
3159 return np; /* xmitting; drop packet */
3162 rxi_ClearTransmitQueue(call, 0);
3164 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3165 rxi_ClearTransmitQueue(call, 0);
3166 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3168 if (np->header.type == RX_PACKET_TYPE_ACK) {
3169 /* now check to see if this is an ack packet acknowledging that the
3170 * server actually *lost* some hard-acked data. If this happens we
3171 * ignore this packet, as it may indicate that the server restarted in
3172 * the middle of a call. It is also possible that this is an old ack
3173 * packet. We don't abort the connection in this case, because this
3174 * *might* just be an old ack packet. The right way to detect a server
3175 * restart in the midst of a call is to notice that the server epoch
3177 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3178 * XXX unacknowledged. I think that this is off-by-one, but
3179 * XXX I don't dare change it just yet, since it will
3180 * XXX interact badly with the server-restart detection
3181 * XXX code in receiveackpacket. */
3182 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3183 if (rx_stats_active)
3184 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3185 MUTEX_EXIT(&call->lock);
3186 MUTEX_ENTER(&rx_refcnt_mutex);
3188 MUTEX_EXIT(&rx_refcnt_mutex);
3192 } /* else not a data packet */
3195 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3196 /* Set remote user defined status from packet */
3197 call->remoteStatus = np->header.userStatus;
3199 /* Note the gap between the expected next packet and the actual
3200 * packet that arrived, when the new packet has a smaller serial number
3201 * than expected. Rioses frequently reorder packets all by themselves,
3202 * so this will be quite important with very large window sizes.
3203 * Skew is checked against 0 here to avoid any dependence on the type of
3204 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3206 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3207 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3208 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3210 MUTEX_ENTER(&conn->conn_data_lock);
3211 skew = conn->lastSerial - np->header.serial;
3212 conn->lastSerial = np->header.serial;
3213 MUTEX_EXIT(&conn->conn_data_lock);
3215 struct rx_peer *peer;
3217 if (skew > peer->inPacketSkew) {
3218 dpf(("*** In skew changed from %d to %d\n",
3219 peer->inPacketSkew, skew));
3220 peer->inPacketSkew = skew;
3224 /* Now do packet type-specific processing */
3225 switch (np->header.type) {
3226 case RX_PACKET_TYPE_DATA:
3227 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3230 case RX_PACKET_TYPE_ACK:
3231 /* Respond immediately to ack packets requesting acknowledgement
3233 if (np->header.flags & RX_REQUEST_ACK) {
3235 (void)rxi_SendCallAbort(call, 0, 1, 0);
3237 (void)rxi_SendAck(call, 0, np->header.serial,
3238 RX_ACK_PING_RESPONSE, 1);
3240 np = rxi_ReceiveAckPacket(call, np, 1);
3242 case RX_PACKET_TYPE_ABORT: {
3243 /* An abort packet: reset the call, passing the error up to the user. */
3244 /* What if error is zero? */
3245 /* What if the error is -1? the application will treat it as a timeout. */
3246 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3247 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3248 rxi_CallError(call, errdata);
3249 MUTEX_EXIT(&call->lock);
3250 MUTEX_ENTER(&rx_refcnt_mutex);
3252 MUTEX_EXIT(&rx_refcnt_mutex);
3253 return np; /* xmitting; drop packet */
3255 case RX_PACKET_TYPE_BUSY:
3258 case RX_PACKET_TYPE_ACKALL:
3259 /* All packets acknowledged, so we can drop all packets previously
3260 * readied for sending */
3261 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3262 /* XXX Hack. We because we can't release the global rx lock when
3263 * sending packets (osi_NetSend) we drop all ack pkts while we're
3264 * traversing the tq in rxi_Start sending packets out because
3265 * packets may move to the freePacketQueue as result of being
3266 * here! So we drop these packets until we're safely out of the
3267 * traversing. Really ugly!
3268 * For fine grain RX locking, we set the acked field in the packets
3269 * and let rxi_Start remove the packets from the transmit queue.
3271 if (call->flags & RX_CALL_TQ_BUSY) {
3272 #ifdef RX_ENABLE_LOCKS
3273 rxi_SetAcksInTransmitQueue(call);
3275 #else /* RX_ENABLE_LOCKS */
3276 MUTEX_EXIT(&call->lock);
3277 MUTEX_ENTER(&rx_refcnt_mutex);
3279 MUTEX_EXIT(&rx_refcnt_mutex);
3280 return np; /* xmitting; drop packet */
3281 #endif /* RX_ENABLE_LOCKS */
3283 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3284 rxi_ClearTransmitQueue(call, 0);
3285 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3288 /* Should not reach here, unless the peer is broken: send an abort
3290 rxi_CallError(call, RX_PROTOCOL_ERROR);
3291 np = rxi_SendCallAbort(call, np, 1, 0);
3294 /* Note when this last legitimate packet was received, for keep-alive
3295 * processing. Note, we delay getting the time until now in the hope that
3296 * the packet will be delivered to the user before any get time is required
3297 * (if not, then the time won't actually be re-evaluated here). */
3298 call->lastReceiveTime = clock_Sec();
3299 MUTEX_EXIT(&call->lock);
3300 MUTEX_ENTER(&rx_refcnt_mutex);
3302 MUTEX_EXIT(&rx_refcnt_mutex);
3306 /* return true if this is an "interesting" connection from the point of view
3307 of someone trying to debug the system */
3309 rxi_IsConnInteresting(struct rx_connection *aconn)
3312 struct rx_call *tcall;
3314 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3317 for (i = 0; i < RX_MAXCALLS; i++) {
3318 tcall = aconn->call[i];
3320 if ((tcall->state == RX_STATE_PRECALL)
3321 || (tcall->state == RX_STATE_ACTIVE))
3323 if ((tcall->mode == RX_MODE_SENDING)
3324 || (tcall->mode == RX_MODE_RECEIVING))
3332 /* if this is one of the last few packets AND it wouldn't be used by the
3333 receiving call to immediately satisfy a read request, then drop it on
3334 the floor, since accepting it might prevent a lock-holding thread from
3335 making progress in its reading. If a call has been cleared while in
3336 the precall state then ignore all subsequent packets until the call
3337 is assigned to a thread. */
3340 TooLow(struct rx_packet *ap, struct rx_call *acall)
3344 MUTEX_ENTER(&rx_quota_mutex);
3345 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3346 && (acall->state == RX_STATE_PRECALL))
3347 || ((rx_nFreePackets < rxi_dataQuota + 2)
3348 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3349 && (acall->flags & RX_CALL_READER_WAIT)))) {
3352 MUTEX_EXIT(&rx_quota_mutex);
3358 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3360 struct rx_connection *conn = arg1;
3361 struct rx_call *acall = arg2;
3362 struct rx_call *call = acall;
3363 struct clock when, now;
3366 MUTEX_ENTER(&conn->conn_data_lock);
3367 conn->checkReachEvent = NULL;
3368 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3370 MUTEX_ENTER(&rx_refcnt_mutex);
3372 MUTEX_EXIT(&rx_refcnt_mutex);
3374 MUTEX_EXIT(&conn->conn_data_lock);
3378 MUTEX_ENTER(&conn->conn_call_lock);
3379 MUTEX_ENTER(&conn->conn_data_lock);
3380 for (i = 0; i < RX_MAXCALLS; i++) {
3381 struct rx_call *tc = conn->call[i];
3382 if (tc && tc->state == RX_STATE_PRECALL) {
3388 /* Indicate that rxi_CheckReachEvent is no longer running by
3389 * clearing the flag. Must be atomic under conn_data_lock to
3390 * avoid a new call slipping by: rxi_CheckConnReach holds
3391 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3393 conn->flags &= ~RX_CONN_ATTACHWAIT;
3394 MUTEX_EXIT(&conn->conn_data_lock);
3395 MUTEX_EXIT(&conn->conn_call_lock);
3400 MUTEX_ENTER(&call->lock);
3401 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3403 MUTEX_EXIT(&call->lock);
3405 clock_GetTime(&now);
3407 when.sec += RX_CHECKREACH_TIMEOUT;
3408 MUTEX_ENTER(&conn->conn_data_lock);
3409 if (!conn->checkReachEvent) {
3410 MUTEX_ENTER(&rx_refcnt_mutex);
3412 MUTEX_EXIT(&rx_refcnt_mutex);
3413 conn->checkReachEvent =
3414 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3417 MUTEX_EXIT(&conn->conn_data_lock);
3423 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3425 struct rx_service *service = conn->service;
3426 struct rx_peer *peer = conn->peer;
3427 afs_uint32 now, lastReach;
3429 if (service->checkReach == 0)
3433 MUTEX_ENTER(&peer->peer_lock);
3434 lastReach = peer->lastReachTime;
3435 MUTEX_EXIT(&peer->peer_lock);
3436 if (now - lastReach < RX_CHECKREACH_TTL)
3439 MUTEX_ENTER(&conn->conn_data_lock);
3440 if (conn->flags & RX_CONN_ATTACHWAIT) {
3441 MUTEX_EXIT(&conn->conn_data_lock);
3444 conn->flags |= RX_CONN_ATTACHWAIT;
3445 MUTEX_EXIT(&conn->conn_data_lock);
3446 if (!conn->checkReachEvent)
3447 rxi_CheckReachEvent(NULL, conn, call);
3452 /* try to attach call, if authentication is complete */
3454 TryAttach(struct rx_call *acall, osi_socket socket,
3455 int *tnop, struct rx_call **newcallp,
3458 struct rx_connection *conn = acall->conn;
3460 if (conn->type == RX_SERVER_CONNECTION
3461 && acall->state == RX_STATE_PRECALL) {
3462 /* Don't attach until we have any req'd. authentication. */
3463 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3464 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3465 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3466 /* Note: this does not necessarily succeed; there
3467 * may not any proc available
3470 rxi_ChallengeOn(acall->conn);
3475 /* A data packet has been received off the interface. This packet is
3476 * appropriate to the call (the call is in the right state, etc.). This
3477 * routine can return a packet to the caller, for re-use */
3480 rxi_ReceiveDataPacket(struct rx_call *call,
3481 struct rx_packet *np, int istack,
3482 osi_socket socket, afs_uint32 host, u_short port,
3483 int *tnop, struct rx_call **newcallp)
3485 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3490 afs_uint32 serial=0, flags=0;
3492 struct rx_packet *tnp;
3493 struct clock when, now;
3494 if (rx_stats_active)
3495 rx_atomic_inc(&rx_stats.dataPacketsRead);
3498 /* If there are no packet buffers, drop this new packet, unless we can find
3499 * packet buffers from inactive calls */
3501 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3502 MUTEX_ENTER(&rx_freePktQ_lock);
3503 rxi_NeedMorePackets = TRUE;
3504 MUTEX_EXIT(&rx_freePktQ_lock);
3505 if (rx_stats_active)
3506 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3507 call->rprev = np->header.serial;
3508 rxi_calltrace(RX_TRACE_DROP, call);
3509 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3511 rxi_ClearReceiveQueue(call);
3512 clock_GetTime(&now);
3514 clock_Add(&when, &rx_softAckDelay);
3515 if (!call->delayedAckEvent
3516 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3517 rxevent_Cancel(call->delayedAckEvent, call,
3518 RX_CALL_REFCOUNT_DELAY);
3519 MUTEX_ENTER(&rx_refcnt_mutex);
3520 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3521 MUTEX_EXIT(&rx_refcnt_mutex);
3523 call->delayedAckEvent =
3524 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3526 /* we've damaged this call already, might as well do it in. */
3532 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3533 * packet is one of several packets transmitted as a single
3534 * datagram. Do not send any soft or hard acks until all packets
3535 * in a jumbogram have been processed. Send negative acks right away.
3537 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3538 /* tnp is non-null when there are more packets in the
3539 * current jumbo gram */
3546 seq = np->header.seq;
3547 serial = np->header.serial;
3548 flags = np->header.flags;
3550 /* If the call is in an error state, send an abort message */
3552 return rxi_SendCallAbort(call, np, istack, 0);
3554 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3555 * AFS 3.5 jumbogram. */
3556 if (flags & RX_JUMBO_PACKET) {
3557 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3562 if (np->header.spare != 0) {
3563 MUTEX_ENTER(&call->conn->conn_data_lock);
3564 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3565 MUTEX_EXIT(&call->conn->conn_data_lock);
3568 /* The usual case is that this is the expected next packet */
3569 if (seq == call->rnext) {
3571 /* Check to make sure it is not a duplicate of one already queued */
3572 if (queue_IsNotEmpty(&call->rq)
3573 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3574 if (rx_stats_active)
3575 rx_atomic_inc(&rx_stats.dupPacketsRead);
3576 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3577 rxevent_Cancel(call->delayedAckEvent, call,
3578 RX_CALL_REFCOUNT_DELAY);
3579 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3585 /* It's the next packet. Stick it on the receive queue
3586 * for this call. Set newPackets to make sure we wake
3587 * the reader once all packets have been processed */
3588 #ifdef RX_TRACK_PACKETS
3589 np->flags |= RX_PKTFLAG_RQ;
3591 queue_Prepend(&call->rq, np);
3592 #ifdef RXDEBUG_PACKET
3594 #endif /* RXDEBUG_PACKET */
3596 np = NULL; /* We can't use this anymore */
3599 /* If an ack is requested then set a flag to make sure we
3600 * send an acknowledgement for this packet */
3601 if (flags & RX_REQUEST_ACK) {
3602 ackNeeded = RX_ACK_REQUESTED;
3605 /* Keep track of whether we have received the last packet */
3606 if (flags & RX_LAST_PACKET) {
3607 call->flags |= RX_CALL_HAVE_LAST;
3611 /* Check whether we have all of the packets for this call */
3612 if (call->flags & RX_CALL_HAVE_LAST) {
3613 afs_uint32 tseq; /* temporary sequence number */
3614 struct rx_packet *tp; /* Temporary packet pointer */
3615 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3617 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3618 if (tseq != tp->header.seq)
3620 if (tp->header.flags & RX_LAST_PACKET) {
3621 call->flags |= RX_CALL_RECEIVE_DONE;
3628 /* Provide asynchronous notification for those who want it
3629 * (e.g. multi rx) */
3630 if (call->arrivalProc) {
3631 (*call->arrivalProc) (call, call->arrivalProcHandle,
3632 call->arrivalProcArg);
3633 call->arrivalProc = (void (*)())0;
3636 /* Update last packet received */
3639 /* If there is no server process serving this call, grab
3640 * one, if available. We only need to do this once. If a
3641 * server thread is available, this thread becomes a server
3642 * thread and the server thread becomes a listener thread. */
3644 TryAttach(call, socket, tnop, newcallp, 0);
3647 /* This is not the expected next packet. */
3649 /* Determine whether this is a new or old packet, and if it's
3650 * a new one, whether it fits into the current receive window.
3651 * Also figure out whether the packet was delivered in sequence.
3652 * We use the prev variable to determine whether the new packet
3653 * is the successor of its immediate predecessor in the
3654 * receive queue, and the missing flag to determine whether
3655 * any of this packets predecessors are missing. */
3657 afs_uint32 prev; /* "Previous packet" sequence number */
3658 struct rx_packet *tp; /* Temporary packet pointer */
3659 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3660 int missing; /* Are any predecessors missing? */
3662 /* If the new packet's sequence number has been sent to the
3663 * application already, then this is a duplicate */
3664 if (seq < call->rnext) {
3665 if (rx_stats_active)
3666 rx_atomic_inc(&rx_stats.dupPacketsRead);
3667 rxevent_Cancel(call->delayedAckEvent, call,
3668 RX_CALL_REFCOUNT_DELAY);
3669 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3675 /* If the sequence number is greater than what can be
3676 * accomodated by the current window, then send a negative
3677 * acknowledge and drop the packet */
3678 if ((call->rnext + call->rwind) <= seq) {
3679 rxevent_Cancel(call->delayedAckEvent, call,
3680 RX_CALL_REFCOUNT_DELAY);
3681 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3688 /* Look for the packet in the queue of old received packets */
3689 for (prev = call->rnext - 1, missing =
3690 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3691 /*Check for duplicate packet */
3692 if (seq == tp->header.seq) {
3693 if (rx_stats_active)
3694 rx_atomic_inc(&rx_stats.dupPacketsRead);
3695 rxevent_Cancel(call->delayedAckEvent, call,
3696 RX_CALL_REFCOUNT_DELAY);
3697 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3703 /* If we find a higher sequence packet, break out and
3704 * insert the new packet here. */
3705 if (seq < tp->header.seq)
3707 /* Check for missing packet */
3708 if (tp->header.seq != prev + 1) {
3712 prev = tp->header.seq;
3715 /* Keep track of whether we have received the last packet. */
3716 if (flags & RX_LAST_PACKET) {
3717 call->flags |= RX_CALL_HAVE_LAST;
3720 /* It's within the window: add it to the the receive queue.
3721 * tp is left by the previous loop either pointing at the
3722 * packet before which to insert the new packet, or at the
3723 * queue head if the queue is empty or the packet should be
3725 #ifdef RX_TRACK_PACKETS
3726 np->flags |= RX_PKTFLAG_RQ;
3728 #ifdef RXDEBUG_PACKET
3730 #endif /* RXDEBUG_PACKET */
3731 queue_InsertBefore(tp, np);
3735 /* Check whether we have all of the packets for this call */
3736 if ((call->flags & RX_CALL_HAVE_LAST)
3737 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3738 afs_uint32 tseq; /* temporary sequence number */
3741 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3742 if (tseq != tp->header.seq)
3744 if (tp->header.flags & RX_LAST_PACKET) {
3745 call->flags |= RX_CALL_RECEIVE_DONE;
3752 /* We need to send an ack of the packet is out of sequence,
3753 * or if an ack was requested by the peer. */
3754 if (seq != prev + 1 || missing) {
3755 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3756 } else if (flags & RX_REQUEST_ACK) {
3757 ackNeeded = RX_ACK_REQUESTED;
3760 /* Acknowledge the last packet for each call */
3761 if (flags & RX_LAST_PACKET) {
3772 * If the receiver is waiting for an iovec, fill the iovec
3773 * using the data from the receive queue */
3774 if (call->flags & RX_CALL_IOVEC_WAIT) {
3775 didHardAck = rxi_FillReadVec(call, serial);
3776 /* the call may have been aborted */
3785 /* Wakeup the reader if any */
3786 if ((call->flags & RX_CALL_READER_WAIT)
3787 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3788 || (call->iovNext >= call->iovMax)
3789 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3790 call->flags &= ~RX_CALL_READER_WAIT;
3791 #ifdef RX_ENABLE_LOCKS
3792 CV_BROADCAST(&call->cv_rq);
3794 osi_rxWakeup(&call->rq);
3800 * Send an ack when requested by the peer, or once every
3801 * rxi_SoftAckRate packets until the last packet has been
3802 * received. Always send a soft ack for the last packet in
3803 * the server's reply.
3805 * If we have received all of the packets for the call
3806 * immediately send an RX_PACKET_TYPE_ACKALL packet so that
3807 * the peer can empty its packet queue and cancel all resend
3810 if (call->flags & RX_CALL_RECEIVE_DONE) {
3811 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3812 rxi_AckAll(NULL, call, 0);
3813 } else if (ackNeeded) {
3814 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3815 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3816 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3817 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3818 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3819 } else if (call->nSoftAcks) {
3820 clock_GetTime(&now);
3822 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3823 clock_Add(&when, &rx_lastAckDelay);
3825 clock_Add(&when, &rx_softAckDelay);
3827 if (!call->delayedAckEvent
3828 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3829 rxevent_Cancel(call->delayedAckEvent, call,
3830 RX_CALL_REFCOUNT_DELAY);
3831 MUTEX_ENTER(&rx_refcnt_mutex);
3832 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3833 MUTEX_EXIT(&rx_refcnt_mutex);
3834 call->delayedAckEvent =
3835 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3843 static void rxi_ComputeRate();
3847 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3849 struct rx_peer *peer = conn->peer;
3851 MUTEX_ENTER(&peer->peer_lock);
3852 peer->lastReachTime = clock_Sec();
3853 MUTEX_EXIT(&peer->peer_lock);
3855 MUTEX_ENTER(&conn->conn_data_lock);
3856 if (conn->flags & RX_CONN_ATTACHWAIT) {
3859 conn->flags &= ~RX_CONN_ATTACHWAIT;
3860 MUTEX_EXIT(&conn->conn_data_lock);
3862 for (i = 0; i < RX_MAXCALLS; i++) {
3863 struct rx_call *call = conn->call[i];
3866 MUTEX_ENTER(&call->lock);
3867 /* tnop can be null if newcallp is null */
3868 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3870 MUTEX_EXIT(&call->lock);
3874 MUTEX_EXIT(&conn->conn_data_lock);
3877 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
3879 rx_ack_reason(int reason)
3882 case RX_ACK_REQUESTED:
3884 case RX_ACK_DUPLICATE:
3886 case RX_ACK_OUT_OF_SEQUENCE:
3888 case RX_ACK_EXCEEDS_WINDOW:
3890 case RX_ACK_NOSPACE:
3894 case RX_ACK_PING_RESPONSE:
3907 /* The real smarts of the whole thing. */
3909 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
3912 struct rx_ackPacket *ap;
3914 struct rx_packet *tp;
3915 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3916 struct rx_connection *conn = call->conn;
3917 struct rx_peer *peer = conn->peer;
3918 struct clock now; /* Current time, for RTT calculations */
3922 /* because there are CM's that are bogus, sending weird values for this. */
3923 afs_uint32 skew = 0;
3928 int newAckCount = 0;
3929 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3930 int pktsize = 0; /* Set if we need to update the peer mtu */
3931 int conn_data_locked = 0;
3933 if (rx_stats_active)
3934 rx_atomic_inc(&rx_stats.ackPacketsRead);
3935 ap = (struct rx_ackPacket *)rx_DataOf(np);
3936 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3938 return np; /* truncated ack packet */
3940 /* depends on ack packet struct */
3941 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3942 first = ntohl(ap->firstPacket);
3943 prev = ntohl(ap->previousPacket);
3944 serial = ntohl(ap->serial);
3945 /* temporarily disabled -- needs to degrade over time
3946 * skew = ntohs(ap->maxSkew); */
3948 /* Ignore ack packets received out of order */
3949 if (first < call->tfirst ||
3950 (first == call->tfirst && prev < call->tprev)) {
3956 if (np->header.flags & RX_SLOW_START_OK) {
3957 call->flags |= RX_CALL_SLOW_START_OK;
3960 if (ap->reason == RX_ACK_PING_RESPONSE)
3961 rxi_UpdatePeerReach(conn, call);
3963 if (conn->lastPacketSizeSeq) {
3964 MUTEX_ENTER(&conn->conn_data_lock);
3965 conn_data_locked = 1;
3966 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
3967 pktsize = conn->lastPacketSize;
3968 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
3971 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
3972 if (!conn_data_locked) {
3973 MUTEX_ENTER(&conn->conn_data_lock);
3974 conn_data_locked = 1;
3976 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
3977 /* process mtu ping ack */
3978 pktsize = conn->lastPingSize;
3979 conn->lastPingSizeSer = conn->lastPingSize = 0;
3983 if (conn_data_locked) {
3984 MUTEX_EXIT(&conn->conn_data_lock);
3985 conn_data_locked = 0;
3989 if (rxdebug_active) {
3993 len = _snprintf(msg, sizeof(msg),
3994 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3995 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3996 ntohl(ap->serial), ntohl(ap->previousPacket),
3997 (unsigned int)np->header.seq, (unsigned int)skew,
3998 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
4002 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4003 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4007 OutputDebugString(msg);
4009 #else /* AFS_NT40_ENV */
4012 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
4013 ap->reason, ntohl(ap->previousPacket),
4014 (unsigned int)np->header.seq, (unsigned int)serial,
4015 (unsigned int)skew, ntohl(ap->firstPacket));
4018 for (offset = 0; offset < nAcks; offset++)
4019 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4024 #endif /* AFS_NT40_ENV */
4027 MUTEX_ENTER(&peer->peer_lock);
4030 * Start somewhere. Can't assume we can send what we can receive,
4031 * but we are clearly receiving.
4033 if (!peer->maxPacketSize)
4034 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4036 if (pktsize > peer->maxPacketSize) {
4037 peer->maxPacketSize = pktsize;
4038 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4039 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4040 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4041 rxi_ScheduleGrowMTUEvent(call, 1);
4046 /* Update the outgoing packet skew value to the latest value of
4047 * the peer's incoming packet skew value. The ack packet, of
4048 * course, could arrive out of order, but that won't affect things
4050 peer->outPacketSkew = skew;
4052 /* Check for packets that no longer need to be transmitted, and
4053 * discard them. This only applies to packets positively
4054 * acknowledged as having been sent to the peer's upper level.
4055 * All other packets must be retained. So only packets with
4056 * sequence numbers < ap->firstPacket are candidates. */
4058 clock_GetTime(&now);
4060 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
4061 if (tp->header.seq >= first)
4063 call->tfirst = tp->header.seq + 1;
4065 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4068 rxi_ComputeRoundTripTime(tp, ap, call->conn->peer, &now);
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 rxi_ComputeRoundTripTime(tp, ap, call->conn->peer, &now);
4147 rxi_ComputeRate(call->conn->peer, call, tp, np,
4156 } else /* RX_ACK_TYPE_NACK */ {
4157 tp->flags &= ~RX_PKTFLAG_ACKED;
4161 if (tp->flags & RX_PKTFLAG_ACKED) {
4162 tp->flags &= ~RX_PKTFLAG_ACKED;
4168 * Following the suggestion of Phil Kern, we back off the peer's
4169 * timeout value for future packets until a successful response
4170 * is received for an initial transmission.
4172 if (missing && !peer->backedOff) {
4173 struct clock c = peer->timeout;
4174 struct clock max_to = {3, 0};
4176 clock_Add(&peer->timeout, &c);
4177 if (clock_Gt(&peer->timeout, &max_to))
4178 peer->timeout = max_to;
4179 peer->backedOff = 1;
4182 /* If packet isn't yet acked, and it has been transmitted at least
4183 * once, reset retransmit time using latest timeout
4184 * ie, this should readjust the retransmit timer for all outstanding
4185 * packets... So we don't just retransmit when we should know better*/
4187 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
4188 tp->retryTime = tp->timeSent;
4189 clock_Add(&tp->retryTime, &peer->timeout);
4190 /* shift by eight because one quarter-sec ~ 256 milliseconds */
4191 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
4195 /* If the window has been extended by this acknowledge packet,
4196 * then wakeup a sender waiting in alloc for window space, or try
4197 * sending packets now, if he's been sitting on packets due to
4198 * lack of window space */
4199 if (call->tnext < (call->tfirst + call->twind)) {
4200 #ifdef RX_ENABLE_LOCKS
4201 CV_SIGNAL(&call->cv_twind);
4203 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4204 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4205 osi_rxWakeup(&call->twind);
4208 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4209 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4213 /* if the ack packet has a receivelen field hanging off it,
4214 * update our state */
4215 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4218 /* If the ack packet has a "recommended" size that is less than
4219 * what I am using now, reduce my size to match */
4220 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4221 (int)sizeof(afs_int32), &tSize);
4222 tSize = (afs_uint32) ntohl(tSize);
4223 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4225 /* Get the maximum packet size to send to this peer */
4226 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4228 tSize = (afs_uint32) ntohl(tSize);
4229 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4230 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4232 /* sanity check - peer might have restarted with different params.
4233 * If peer says "send less", dammit, send less... Peer should never
4234 * be unable to accept packets of the size that prior AFS versions would
4235 * send without asking. */
4236 if (peer->maxMTU != tSize) {
4237 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4239 peer->maxMTU = tSize;
4240 peer->MTU = MIN(tSize, peer->MTU);
4241 call->MTU = MIN(call->MTU, tSize);
4244 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4247 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4248 (int)sizeof(afs_int32), &tSize);
4249 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4250 if (tSize < call->twind) { /* smaller than our send */
4251 call->twind = tSize; /* window, we must send less... */
4252 call->ssthresh = MIN(call->twind, call->ssthresh);
4253 call->conn->twind[call->channel] = call->twind;
4256 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4257 * network MTU confused with the loopback MTU. Calculate the
4258 * maximum MTU here for use in the slow start code below.
4260 /* Did peer restart with older RX version? */
4261 if (peer->maxDgramPackets > 1) {
4262 peer->maxDgramPackets = 1;
4264 } else if (np->length >=
4265 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4268 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4269 sizeof(afs_int32), &tSize);
4270 tSize = (afs_uint32) ntohl(tSize);
4272 * As of AFS 3.5 we set the send window to match the receive window.
4274 if (tSize < call->twind) {
4275 call->twind = tSize;
4276 call->conn->twind[call->channel] = call->twind;
4277 call->ssthresh = MIN(call->twind, call->ssthresh);
4278 } else if (tSize > call->twind) {
4279 call->twind = tSize;
4280 call->conn->twind[call->channel] = call->twind;
4284 * As of AFS 3.5, a jumbogram is more than one fixed size
4285 * packet transmitted in a single UDP datagram. If the remote
4286 * MTU is smaller than our local MTU then never send a datagram
4287 * larger than the natural MTU.
4290 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4291 (int)sizeof(afs_int32), &tSize);
4292 maxDgramPackets = (afs_uint32) ntohl(tSize);
4293 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4295 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4296 if (maxDgramPackets > 1) {
4297 peer->maxDgramPackets = maxDgramPackets;
4298 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4300 peer->maxDgramPackets = 1;
4301 call->MTU = peer->natMTU;
4303 } else if (peer->maxDgramPackets > 1) {
4304 /* Restarted with lower version of RX */
4305 peer->maxDgramPackets = 1;
4307 } else if (peer->maxDgramPackets > 1
4308 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4309 /* Restarted with lower version of RX */
4310 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4311 peer->natMTU = OLD_MAX_PACKET_SIZE;
4312 peer->MTU = OLD_MAX_PACKET_SIZE;
4313 peer->maxDgramPackets = 1;
4314 peer->nDgramPackets = 1;
4316 call->MTU = OLD_MAX_PACKET_SIZE;
4321 * Calculate how many datagrams were successfully received after
4322 * the first missing packet and adjust the negative ack counter
4327 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4328 if (call->nNacks < nNacked) {
4329 call->nNacks = nNacked;
4332 call->nAcks += newAckCount;
4336 if (call->flags & RX_CALL_FAST_RECOVER) {
4338 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4340 call->flags &= ~RX_CALL_FAST_RECOVER;
4341 call->cwind = call->nextCwind;
4342 call->nextCwind = 0;
4345 call->nCwindAcks = 0;
4346 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4347 /* Three negative acks in a row trigger congestion recovery */
4348 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4349 MUTEX_EXIT(&peer->peer_lock);
4350 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4351 /* someone else is waiting to start recovery */
4354 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4355 rxi_WaitforTQBusy(call);
4356 MUTEX_ENTER(&peer->peer_lock);
4357 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4358 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4359 call->flags |= RX_CALL_FAST_RECOVER;
4360 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4362 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4363 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4364 call->nextCwind = call->ssthresh;
4367 peer->MTU = call->MTU;
4368 peer->cwind = call->nextCwind;
4369 peer->nDgramPackets = call->nDgramPackets;
4371 call->congestSeq = peer->congestSeq;
4372 /* Reset the resend times on the packets that were nacked
4373 * so we will retransmit as soon as the window permits*/
4374 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4376 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4377 clock_Zero(&tp->retryTime);
4379 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4384 /* If cwind is smaller than ssthresh, then increase
4385 * the window one packet for each ack we receive (exponential
4387 * If cwind is greater than or equal to ssthresh then increase
4388 * the congestion window by one packet for each cwind acks we
4389 * receive (linear growth). */
4390 if (call->cwind < call->ssthresh) {
4392 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4393 call->nCwindAcks = 0;
4395 call->nCwindAcks += newAckCount;
4396 if (call->nCwindAcks >= call->cwind) {
4397 call->nCwindAcks = 0;
4398 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4402 * If we have received several acknowledgements in a row then
4403 * it is time to increase the size of our datagrams
4405 if ((int)call->nAcks > rx_nDgramThreshold) {
4406 if (peer->maxDgramPackets > 1) {
4407 if (call->nDgramPackets < peer->maxDgramPackets) {
4408 call->nDgramPackets++;
4410 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4411 } else if (call->MTU < peer->maxMTU) {
4412 /* don't upgrade if we can't handle it */
4413 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4414 call->MTU = peer->ifMTU;
4416 call->MTU += peer->natMTU;
4417 call->MTU = MIN(call->MTU, peer->maxMTU);
4424 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4426 /* Servers need to hold the call until all response packets have
4427 * been acknowledged. Soft acks are good enough since clients
4428 * are not allowed to clear their receive queues. */
4429 if (call->state == RX_STATE_HOLD
4430 && call->tfirst + call->nSoftAcked >= call->tnext) {
4431 call->state = RX_STATE_DALLY;
4432 rxi_ClearTransmitQueue(call, 0);
4433 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4434 } else if (!queue_IsEmpty(&call->tq)) {
4435 rxi_Start(0, call, 0, istack);
4440 /* Received a response to a challenge packet */
4442 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4443 struct rx_packet *np, int istack)
4447 /* Ignore the packet if we're the client */
4448 if (conn->type == RX_CLIENT_CONNECTION)
4451 /* If already authenticated, ignore the packet (it's probably a retry) */
4452 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4455 /* Otherwise, have the security object evaluate the response packet */
4456 error = RXS_CheckResponse(conn->securityObject, conn, np);
4458 /* If the response is invalid, reset the connection, sending
4459 * an abort to the peer */
4463 rxi_ConnectionError(conn, error);
4464 MUTEX_ENTER(&conn->conn_data_lock);
4465 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4466 MUTEX_EXIT(&conn->conn_data_lock);
4469 /* If the response is valid, any calls waiting to attach
4470 * servers can now do so */
4473 for (i = 0; i < RX_MAXCALLS; i++) {
4474 struct rx_call *call = conn->call[i];
4476 MUTEX_ENTER(&call->lock);
4477 if (call->state == RX_STATE_PRECALL)
4478 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4479 /* tnop can be null if newcallp is null */
4480 MUTEX_EXIT(&call->lock);
4484 /* Update the peer reachability information, just in case
4485 * some calls went into attach-wait while we were waiting
4486 * for authentication..
4488 rxi_UpdatePeerReach(conn, NULL);
4493 /* A client has received an authentication challenge: the security
4494 * object is asked to cough up a respectable response packet to send
4495 * back to the server. The server is responsible for retrying the
4496 * challenge if it fails to get a response. */
4499 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4500 struct rx_packet *np, int istack)
4504 /* Ignore the challenge if we're the server */
4505 if (conn->type == RX_SERVER_CONNECTION)
4508 /* Ignore the challenge if the connection is otherwise idle; someone's
4509 * trying to use us as an oracle. */
4510 if (!rxi_HasActiveCalls(conn))
4513 /* Send the security object the challenge packet. It is expected to fill
4514 * in the response. */
4515 error = RXS_GetResponse(conn->securityObject, conn, np);
4517 /* If the security object is unable to return a valid response, reset the
4518 * connection and send an abort to the peer. Otherwise send the response
4519 * packet to the peer connection. */
4521 rxi_ConnectionError(conn, error);
4522 MUTEX_ENTER(&conn->conn_data_lock);
4523 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4524 MUTEX_EXIT(&conn->conn_data_lock);
4526 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4527 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4533 /* Find an available server process to service the current request in
4534 * the given call structure. If one isn't available, queue up this
4535 * call so it eventually gets one */
4537 rxi_AttachServerProc(struct rx_call *call,
4538 osi_socket socket, int *tnop,
4539 struct rx_call **newcallp)
4541 struct rx_serverQueueEntry *sq;
4542 struct rx_service *service = call->conn->service;
4545 /* May already be attached */
4546 if (call->state == RX_STATE_ACTIVE)
4549 MUTEX_ENTER(&rx_serverPool_lock);
4551 haveQuota = QuotaOK(service);
4552 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4553 /* If there are no processes available to service this call,
4554 * put the call on the incoming call queue (unless it's
4555 * already on the queue).
4557 #ifdef RX_ENABLE_LOCKS
4559 ReturnToServerPool(service);
4560 #endif /* RX_ENABLE_LOCKS */
4562 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4563 call->flags |= RX_CALL_WAIT_PROC;
4564 rx_atomic_inc(&rx_nWaiting);
4565 rx_atomic_inc(&rx_nWaited);
4566 rxi_calltrace(RX_CALL_ARRIVAL, call);
4567 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4568 queue_Append(&rx_incomingCallQueue, call);
4571 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4573 /* If hot threads are enabled, and both newcallp and sq->socketp
4574 * are non-null, then this thread will process the call, and the
4575 * idle server thread will start listening on this threads socket.
4578 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4581 *sq->socketp = socket;
4582 clock_GetTime(&call->startTime);
4583 MUTEX_ENTER(&rx_refcnt_mutex);
4584 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4585 MUTEX_EXIT(&rx_refcnt_mutex);
4589 if (call->flags & RX_CALL_WAIT_PROC) {
4590 /* Conservative: I don't think this should happen */
4591 call->flags &= ~RX_CALL_WAIT_PROC;
4592 if (queue_IsOnQueue(call)) {
4595 rx_atomic_dec(&rx_nWaiting);
4598 call->state = RX_STATE_ACTIVE;
4599 call->mode = RX_MODE_RECEIVING;
4600 #ifdef RX_KERNEL_TRACE
4602 int glockOwner = ISAFS_GLOCK();
4605 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4606 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4612 if (call->flags & RX_CALL_CLEARED) {
4613 /* send an ack now to start the packet flow up again */
4614 call->flags &= ~RX_CALL_CLEARED;
4615 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4617 #ifdef RX_ENABLE_LOCKS
4620 service->nRequestsRunning++;
4621 MUTEX_ENTER(&rx_quota_mutex);
4622 if (service->nRequestsRunning <= service->minProcs)
4625 MUTEX_EXIT(&rx_quota_mutex);
4629 MUTEX_EXIT(&rx_serverPool_lock);
4632 /* Delay the sending of an acknowledge event for a short while, while
4633 * a new call is being prepared (in the case of a client) or a reply
4634 * is being prepared (in the case of a server). Rather than sending
4635 * an ack packet, an ACKALL packet is sent. */
4637 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4639 #ifdef RX_ENABLE_LOCKS
4641 MUTEX_ENTER(&call->lock);
4642 call->delayedAckEvent = NULL;
4643 MUTEX_ENTER(&rx_refcnt_mutex);
4644 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4645 MUTEX_EXIT(&rx_refcnt_mutex);
4647 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4648 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4650 MUTEX_EXIT(&call->lock);
4651 #else /* RX_ENABLE_LOCKS */
4653 call->delayedAckEvent = NULL;
4654 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4655 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4656 #endif /* RX_ENABLE_LOCKS */
4660 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4662 struct rx_call *call = arg1;
4663 #ifdef RX_ENABLE_LOCKS
4665 MUTEX_ENTER(&call->lock);
4666 if (event == call->delayedAckEvent)
4667 call->delayedAckEvent = NULL;
4668 MUTEX_ENTER(&rx_refcnt_mutex);
4669 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4670 MUTEX_EXIT(&rx_refcnt_mutex);
4672 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4674 MUTEX_EXIT(&call->lock);
4675 #else /* RX_ENABLE_LOCKS */
4677 call->delayedAckEvent = NULL;
4678 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4679 #endif /* RX_ENABLE_LOCKS */
4683 #ifdef RX_ENABLE_LOCKS
4684 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4685 * clearing them out.
4688 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4690 struct rx_packet *p, *tp;
4693 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4694 p->flags |= RX_PKTFLAG_ACKED;
4698 call->flags |= RX_CALL_TQ_CLEARME;
4699 call->flags |= RX_CALL_TQ_SOME_ACKED;
4702 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4703 call->tfirst = call->tnext;
4704 call->nSoftAcked = 0;
4706 if (call->flags & RX_CALL_FAST_RECOVER) {
4707 call->flags &= ~RX_CALL_FAST_RECOVER;
4708 call->cwind = call->nextCwind;
4709 call->nextCwind = 0;
4712 CV_SIGNAL(&call->cv_twind);
4714 #endif /* RX_ENABLE_LOCKS */
4716 /* Clear out the transmit queue for the current call (all packets have
4717 * been received by peer) */
4719 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4721 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4722 struct rx_packet *p, *tp;
4724 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4726 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4727 p->flags |= RX_PKTFLAG_ACKED;
4731 call->flags |= RX_CALL_TQ_CLEARME;
4732 call->flags |= RX_CALL_TQ_SOME_ACKED;
4735 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4736 #ifdef RXDEBUG_PACKET
4738 #endif /* RXDEBUG_PACKET */
4739 rxi_FreePackets(0, &call->tq);
4740 rxi_WakeUpTransmitQueue(call);
4741 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4742 call->flags &= ~RX_CALL_TQ_CLEARME;
4744 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4746 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4747 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4748 call->nSoftAcked = 0;
4750 if (call->flags & RX_CALL_FAST_RECOVER) {
4751 call->flags &= ~RX_CALL_FAST_RECOVER;
4752 call->cwind = call->nextCwind;
4754 #ifdef RX_ENABLE_LOCKS
4755 CV_SIGNAL(&call->cv_twind);
4757 osi_rxWakeup(&call->twind);
4762 rxi_ClearReceiveQueue(struct rx_call *call)
4764 if (queue_IsNotEmpty(&call->rq)) {
4767 count = rxi_FreePackets(0, &call->rq);
4768 rx_packetReclaims += count;
4769 #ifdef RXDEBUG_PACKET
4771 if ( call->rqc != 0 )
4772 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
4774 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4776 if (call->state == RX_STATE_PRECALL) {
4777 call->flags |= RX_CALL_CLEARED;
4781 /* Send an abort packet for the specified call */
4783 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
4784 int istack, int force)
4787 struct clock when, now;
4792 /* Clients should never delay abort messages */
4793 if (rx_IsClientConn(call->conn))
4796 if (call->abortCode != call->error) {
4797 call->abortCode = call->error;
4798 call->abortCount = 0;
4801 if (force || rxi_callAbortThreshhold == 0
4802 || call->abortCount < rxi_callAbortThreshhold) {
4803 if (call->delayedAbortEvent) {
4804 rxevent_Cancel(call->delayedAbortEvent, call,
4805 RX_CALL_REFCOUNT_ABORT);
4807 error = htonl(call->error);
4810 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4811 (char *)&error, sizeof(error), istack);
4812 } else if (!call->delayedAbortEvent) {
4813 clock_GetTime(&now);
4815 clock_Addmsec(&when, rxi_callAbortDelay);
4816 MUTEX_ENTER(&rx_refcnt_mutex);
4817 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4818 MUTEX_EXIT(&rx_refcnt_mutex);
4819 call->delayedAbortEvent =
4820 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4825 /* Send an abort packet for the specified connection. Packet is an
4826 * optional pointer to a packet that can be used to send the abort.
4827 * Once the number of abort messages reaches the threshhold, an
4828 * event is scheduled to send the abort. Setting the force flag
4829 * overrides sending delayed abort messages.
4831 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4832 * to send the abort packet.
4835 rxi_SendConnectionAbort(struct rx_connection *conn,
4836 struct rx_packet *packet, int istack, int force)
4839 struct clock when, now;
4844 /* Clients should never delay abort messages */
4845 if (rx_IsClientConn(conn))
4848 if (force || rxi_connAbortThreshhold == 0
4849 || conn->abortCount < rxi_connAbortThreshhold) {
4850 if (conn->delayedAbortEvent) {
4851 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4853 error = htonl(conn->error);
4855 MUTEX_EXIT(&conn->conn_data_lock);
4857 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4858 RX_PACKET_TYPE_ABORT, (char *)&error,
4859 sizeof(error), istack);
4860 MUTEX_ENTER(&conn->conn_data_lock);
4861 } else if (!conn->delayedAbortEvent) {
4862 clock_GetTime(&now);
4864 clock_Addmsec(&when, rxi_connAbortDelay);
4865 conn->delayedAbortEvent =
4866 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4871 /* Associate an error all of the calls owned by a connection. Called
4872 * with error non-zero. This is only for really fatal things, like
4873 * bad authentication responses. The connection itself is set in
4874 * error at this point, so that future packets received will be
4877 rxi_ConnectionError(struct rx_connection *conn,
4883 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
4885 MUTEX_ENTER(&conn->conn_data_lock);
4886 if (conn->challengeEvent)
4887 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4888 if (conn->natKeepAliveEvent)
4889 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
4890 if (conn->checkReachEvent) {
4891 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4892 conn->checkReachEvent = 0;
4893 conn->flags &= ~RX_CONN_ATTACHWAIT;
4894 MUTEX_ENTER(&rx_refcnt_mutex);
4896 MUTEX_EXIT(&rx_refcnt_mutex);
4898 MUTEX_EXIT(&conn->conn_data_lock);
4899 for (i = 0; i < RX_MAXCALLS; i++) {
4900 struct rx_call *call = conn->call[i];
4902 MUTEX_ENTER(&call->lock);
4903 rxi_CallError(call, error);
4904 MUTEX_EXIT(&call->lock);
4907 conn->error = error;
4908 if (rx_stats_active)
4909 rx_atomic_inc(&rx_stats.fatalErrors);
4914 * Interrupt an in-progress call with the specified error and wakeup waiters.
4916 * @param[in] call The call to interrupt
4917 * @param[in] error The error code to send to the peer
4920 rx_InterruptCall(struct rx_call *call, afs_int32 error)
4922 MUTEX_ENTER(&call->lock);
4923 rxi_CallError(call, error);
4924 rxi_SendCallAbort(call, NULL, 0, 1);
4925 MUTEX_EXIT(&call->lock);
4929 rxi_CallError(struct rx_call *call, afs_int32 error)
4932 osirx_AssertMine(&call->lock, "rxi_CallError");
4934 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
4936 error = call->error;
4938 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4939 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4940 rxi_ResetCall(call, 0);
4943 rxi_ResetCall(call, 0);
4945 call->error = error;
4948 /* Reset various fields in a call structure, and wakeup waiting
4949 * processes. Some fields aren't changed: state & mode are not
4950 * touched (these must be set by the caller), and bufptr, nLeft, and
4951 * nFree are not reset, since these fields are manipulated by
4952 * unprotected macros, and may only be reset by non-interrupting code.
4955 /* this code requires that call->conn be set properly as a pre-condition. */
4956 #endif /* ADAPT_WINDOW */
4959 rxi_ResetCall(struct rx_call *call, int newcall)
4962 struct rx_peer *peer;
4963 struct rx_packet *packet;
4965 osirx_AssertMine(&call->lock, "rxi_ResetCall");
4967 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
4969 /* Notify anyone who is waiting for asynchronous packet arrival */
4970 if (call->arrivalProc) {
4971 (*call->arrivalProc) (call, call->arrivalProcHandle,
4972 call->arrivalProcArg);
4973 call->arrivalProc = (void (*)())0;
4976 if (call->delayedAbortEvent) {
4977 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4978 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4980 rxi_SendCallAbort(call, packet, 0, 1);
4981 rxi_FreePacket(packet);
4986 * Update the peer with the congestion information in this call
4987 * so other calls on this connection can pick up where this call
4988 * left off. If the congestion sequence numbers don't match then
4989 * another call experienced a retransmission.
4991 peer = call->conn->peer;
4992 MUTEX_ENTER(&peer->peer_lock);
4994 if (call->congestSeq == peer->congestSeq) {
4995 peer->cwind = MAX(peer->cwind, call->cwind);
4996 peer->MTU = MAX(peer->MTU, call->MTU);
4997 peer->nDgramPackets =
4998 MAX(peer->nDgramPackets, call->nDgramPackets);
5001 call->abortCode = 0;
5002 call->abortCount = 0;
5004 if (peer->maxDgramPackets > 1) {
5005 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5007 call->MTU = peer->MTU;
5009 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5010 call->ssthresh = rx_maxSendWindow;
5011 call->nDgramPackets = peer->nDgramPackets;
5012 call->congestSeq = peer->congestSeq;
5013 MUTEX_EXIT(&peer->peer_lock);
5015 flags = call->flags;
5016 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5017 rxi_WaitforTQBusy(call);
5018 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5020 rxi_ClearTransmitQueue(call, 1);
5021 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5022 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5026 rxi_ClearReceiveQueue(call);
5027 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5031 call->twind = call->conn->twind[call->channel];
5032 call->rwind = call->conn->rwind[call->channel];
5033 call->nSoftAcked = 0;
5034 call->nextCwind = 0;
5037 call->nCwindAcks = 0;
5038 call->nSoftAcks = 0;
5039 call->nHardAcks = 0;
5041 call->tfirst = call->rnext = call->tnext = 1;
5044 call->lastAcked = 0;
5045 call->localStatus = call->remoteStatus = 0;
5047 if (flags & RX_CALL_READER_WAIT) {
5048 #ifdef RX_ENABLE_LOCKS
5049 CV_BROADCAST(&call->cv_rq);
5051 osi_rxWakeup(&call->rq);
5054 if (flags & RX_CALL_WAIT_PACKETS) {
5055 MUTEX_ENTER(&rx_freePktQ_lock);
5056 rxi_PacketsUnWait(); /* XXX */
5057 MUTEX_EXIT(&rx_freePktQ_lock);
5059 #ifdef RX_ENABLE_LOCKS
5060 CV_SIGNAL(&call->cv_twind);
5062 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5063 osi_rxWakeup(&call->twind);
5066 #ifdef RX_ENABLE_LOCKS
5067 /* The following ensures that we don't mess with any queue while some
5068 * other thread might also be doing so. The call_queue_lock field is
5069 * is only modified under the call lock. If the call is in the process
5070 * of being removed from a queue, the call is not locked until the
5071 * the queue lock is dropped and only then is the call_queue_lock field
5072 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5073 * Note that any other routine which removes a call from a queue has to
5074 * obtain the queue lock before examing the queue and removing the call.
5076 if (call->call_queue_lock) {
5077 MUTEX_ENTER(call->call_queue_lock);
5078 if (queue_IsOnQueue(call)) {
5080 if (flags & RX_CALL_WAIT_PROC) {
5081 rx_atomic_dec(&rx_nWaiting);
5084 MUTEX_EXIT(call->call_queue_lock);
5085 CLEAR_CALL_QUEUE_LOCK(call);
5087 #else /* RX_ENABLE_LOCKS */
5088 if (queue_IsOnQueue(call)) {
5090 if (flags & RX_CALL_WAIT_PROC)
5091 rx_atomic_dec(&rx_nWaiting);
5093 #endif /* RX_ENABLE_LOCKS */
5095 rxi_KeepAliveOff(call);
5096 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5099 /* Send an acknowledge for the indicated packet (seq,serial) of the
5100 * indicated call, for the indicated reason (reason). This
5101 * acknowledge will specifically acknowledge receiving the packet, and
5102 * will also specify which other packets for this call have been
5103 * received. This routine returns the packet that was used to the
5104 * caller. The caller is responsible for freeing it or re-using it.
5105 * This acknowledgement also returns the highest sequence number
5106 * actually read out by the higher level to the sender; the sender
5107 * promises to keep around packets that have not been read by the
5108 * higher level yet (unless, of course, the sender decides to abort
5109 * the call altogether). Any of p, seq, serial, pflags, or reason may
5110 * be set to zero without ill effect. That is, if they are zero, they
5111 * will not convey any information.
5112 * NOW there is a trailer field, after the ack where it will safely be
5113 * ignored by mundanes, which indicates the maximum size packet this
5114 * host can swallow. */
5116 struct rx_packet *optionalPacket; use to send ack (or null)
5117 int seq; Sequence number of the packet we are acking
5118 int serial; Serial number of the packet
5119 int pflags; Flags field from packet header
5120 int reason; Reason an acknowledge was prompted
5124 rxi_SendAck(struct rx_call *call,
5125 struct rx_packet *optionalPacket, int serial, int reason,
5128 struct rx_ackPacket *ap;
5129 struct rx_packet *rqp;
5130 struct rx_packet *nxp; /* For queue_Scan */
5131 struct rx_packet *p;
5134 afs_uint32 padbytes = 0;
5135 #ifdef RX_ENABLE_TSFPQ
5136 struct rx_ts_info_t * rx_ts_info;
5140 * Open the receive window once a thread starts reading packets
5142 if (call->rnext > 1) {
5143 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5146 /* Don't attempt to grow MTU if this is a critical ping */
5147 if (reason == RX_ACK_MTU) {
5148 /* keep track of per-call attempts, if we're over max, do in small
5149 * otherwise in larger? set a size to increment by, decrease
5152 if (call->conn->peer->maxPacketSize &&
5153 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5155 padbytes = call->conn->peer->maxPacketSize+16;
5157 padbytes = call->conn->peer->maxMTU + 128;
5159 /* do always try a minimum size ping */
5160 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5162 /* subtract the ack payload */
5163 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5164 reason = RX_ACK_PING;
5167 call->nHardAcks = 0;
5168 call->nSoftAcks = 0;
5169 if (call->rnext > call->lastAcked)
5170 call->lastAcked = call->rnext;
5174 rx_computelen(p, p->length); /* reset length, you never know */
5175 } /* where that's been... */
5176 #ifdef RX_ENABLE_TSFPQ
5178 RX_TS_INFO_GET(rx_ts_info);
5179 if ((p = rx_ts_info->local_special_packet)) {
5180 rx_computelen(p, p->length);
5181 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5182 rx_ts_info->local_special_packet = p;
5183 } else { /* We won't send the ack, but don't panic. */
5184 return optionalPacket;
5188 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5189 /* We won't send the ack, but don't panic. */
5190 return optionalPacket;
5195 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5198 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5199 #ifndef RX_ENABLE_TSFPQ
5200 if (!optionalPacket)
5203 return optionalPacket;
5205 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5206 if (rx_Contiguous(p) < templ) {
5207 #ifndef RX_ENABLE_TSFPQ
5208 if (!optionalPacket)
5211 return optionalPacket;
5216 /* MTUXXX failing to send an ack is very serious. We should */
5217 /* try as hard as possible to send even a partial ack; it's */
5218 /* better than nothing. */
5219 ap = (struct rx_ackPacket *)rx_DataOf(p);
5220 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5221 ap->reason = reason;
5223 /* The skew computation used to be bogus, I think it's better now. */
5224 /* We should start paying attention to skew. XXX */
5225 ap->serial = htonl(serial);
5226 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5228 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
5229 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5231 /* No fear of running out of ack packet here because there can only be at most
5232 * one window full of unacknowledged packets. The window size must be constrained
5233 * to be less than the maximum ack size, of course. Also, an ack should always
5234 * fit into a single packet -- it should not ever be fragmented. */
5235 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5236 if (!rqp || !call->rq.next
5237 || (rqp->header.seq > (call->rnext + call->rwind))) {
5238 #ifndef RX_ENABLE_TSFPQ
5239 if (!optionalPacket)
5242 rxi_CallError(call, RX_CALL_DEAD);
5243 return optionalPacket;
5246 while (rqp->header.seq > call->rnext + offset)
5247 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5248 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5250 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5251 #ifndef RX_ENABLE_TSFPQ
5252 if (!optionalPacket)
5255 rxi_CallError(call, RX_CALL_DEAD);
5256 return optionalPacket;
5261 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5263 /* these are new for AFS 3.3 */
5264 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5265 templ = htonl(templ);
5266 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5267 templ = htonl(call->conn->peer->ifMTU);
5268 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5269 sizeof(afs_int32), &templ);
5271 /* new for AFS 3.4 */
5272 templ = htonl(call->rwind);
5273 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5274 sizeof(afs_int32), &templ);
5276 /* new for AFS 3.5 */
5277 templ = htonl(call->conn->peer->ifDgramPackets);
5278 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5279 sizeof(afs_int32), &templ);
5281 p->header.serviceId = call->conn->serviceId;
5282 p->header.cid = (call->conn->cid | call->channel);
5283 p->header.callNumber = *call->callNumber;
5285 p->header.securityIndex = call->conn->securityIndex;
5286 p->header.epoch = call->conn->epoch;
5287 p->header.type = RX_PACKET_TYPE_ACK;
5288 p->header.flags = RX_SLOW_START_OK;
5289 if (reason == RX_ACK_PING) {
5290 p->header.flags |= RX_REQUEST_ACK;
5292 clock_GetTime(&call->pingRequestTime);
5295 p->length = padbytes +
5296 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5299 /* not fast but we can potentially use this if truncated
5300 * fragments are delivered to figure out the mtu.
5302 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5303 sizeof(afs_int32), sizeof(afs_int32),
5307 if (call->conn->type == RX_CLIENT_CONNECTION)
5308 p->header.flags |= RX_CLIENT_INITIATED;
5312 if (rxdebug_active) {
5316 len = _snprintf(msg, sizeof(msg),
5317 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5318 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5319 ntohl(ap->serial), ntohl(ap->previousPacket),
5320 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5321 ap->nAcks, ntohs(ap->bufferSpace) );
5325 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5326 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5330 OutputDebugString(msg);
5332 #else /* AFS_NT40_ENV */
5334 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5335 ap->reason, ntohl(ap->previousPacket),
5336 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5338 for (offset = 0; offset < ap->nAcks; offset++)
5339 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5344 #endif /* AFS_NT40_ENV */
5347 int i, nbytes = p->length;
5349 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5350 if (nbytes <= p->wirevec[i].iov_len) {
5353 savelen = p->wirevec[i].iov_len;
5355 p->wirevec[i].iov_len = nbytes;
5357 rxi_Send(call, p, istack);
5358 p->wirevec[i].iov_len = savelen;
5362 nbytes -= p->wirevec[i].iov_len;
5365 if (rx_stats_active)
5366 rx_atomic_inc(&rx_stats.ackPacketsSent);
5367 #ifndef RX_ENABLE_TSFPQ
5368 if (!optionalPacket)
5371 return optionalPacket; /* Return packet for re-use by caller */
5375 struct rx_packet **list;
5380 /* Send all of the packets in the list in single datagram */
5382 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5383 int istack, int moreFlag)
5388 struct clock now, retryTime;
5389 struct rx_connection *conn = call->conn;
5390 struct rx_peer *peer = conn->peer;
5392 MUTEX_ENTER(&peer->peer_lock);
5393 peer->nSent += xmit->len;
5394 if (xmit->resending)
5395 peer->reSends += xmit->len;
5396 retryTime = peer->timeout;
5397 MUTEX_EXIT(&peer->peer_lock);
5399 if (rx_stats_active) {
5400 if (xmit->resending)
5401 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5403 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5406 clock_GetTime(&now);
5407 clock_Add(&retryTime, &now);
5409 if (xmit->list[xmit->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 < xmit->len; i++) {
5416 struct rx_packet *packet = xmit->list[i];
5418 packet->retryTime = retryTime;
5419 if (packet->header.serial) {
5420 /* Exponentially backoff retry times */
5421 if (packet->backoff < MAXBACKOFF) {
5422 /* so it can't stay == 0 */
5423 packet->backoff = (packet->backoff << 1) + 1;
5426 clock_Addmsec(&(packet->retryTime),
5427 ((afs_uint32) packet->backoff) << 8);
5430 /* Wait a little extra for the ack on the last packet */
5432 && !(packet->header.flags & RX_CLIENT_INITIATED)) {
5433 clock_Addmsec(&(packet->retryTime), 400);
5436 /* Record the time sent */
5437 packet->timeSent = now;
5439 /* Ask for an ack on retransmitted packets, on every other packet
5440 * if the peer doesn't support slow start. Ask for an ack on every
5441 * packet until the congestion window reaches the ack rate. */
5442 if (packet->header.serial) {
5445 /* improved RTO calculation- not Karn */
5446 packet->firstSent = now;
5447 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5448 || (!(call->flags & RX_CALL_SLOW_START_OK)
5449 && (packet->header.seq & 1)))) {
5454 /* Tag this packet as not being the last in this group,
5455 * for the receiver's benefit */
5456 if (i < xmit->len - 1 || moreFlag) {
5457 packet->header.flags |= RX_MORE_PACKETS;
5462 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5465 /* Since we're about to send a data packet to the peer, it's
5466 * safe to nuke any scheduled end-of-packets ack */
5467 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5469 MUTEX_EXIT(&call->lock);
5470 MUTEX_ENTER(&rx_refcnt_mutex);
5471 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5472 MUTEX_EXIT(&rx_refcnt_mutex);
5473 if (xmit->len > 1) {
5474 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5476 rxi_SendPacket(call, conn, xmit->list[0], istack);
5478 MUTEX_ENTER(&call->lock);
5479 MUTEX_ENTER(&rx_refcnt_mutex);
5480 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5481 MUTEX_EXIT(&rx_refcnt_mutex);
5483 /* Update last send time for this call (for keep-alive
5484 * processing), and for the connection (so that we can discover
5485 * idle connections) */
5486 conn->lastSendTime = call->lastSendTime = clock_Sec();
5487 /* Let a set of retransmits trigger an idle timeout */
5488 if (!xmit->resending)
5489 call->lastSendData = call->lastSendTime;
5492 /* When sending packets we need to follow these rules:
5493 * 1. Never send more than maxDgramPackets in a jumbogram.
5494 * 2. Never send a packet with more than two iovecs in a jumbogram.
5495 * 3. Never send a retransmitted packet in a jumbogram.
5496 * 4. Never send more than cwind/4 packets in a jumbogram
5497 * We always keep the last list we should have sent so we
5498 * can set the RX_MORE_PACKETS flags correctly.
5502 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5506 struct xmitlist working;
5507 struct xmitlist last;
5509 struct rx_peer *peer = call->conn->peer;
5510 int morePackets = 0;
5512 memset(&last, 0, sizeof(struct xmitlist));
5513 working.list = &list[0];
5515 working.resending = 0;
5517 for (i = 0; i < len; i++) {
5518 /* Does the current packet force us to flush the current list? */
5520 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5521 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5523 /* This sends the 'last' list and then rolls the current working
5524 * set into the 'last' one, and resets the working set */
5527 rxi_SendList(call, &last, istack, 1);
5528 /* If the call enters an error state stop sending, or if
5529 * we entered congestion recovery mode, stop sending */
5530 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5535 working.resending = 0;
5536 working.list = &list[i];
5538 /* Add the current packet to the list if it hasn't been acked.
5539 * Otherwise adjust the list pointer to skip the current packet. */
5540 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5543 if (list[i]->header.serial)
5544 working.resending = 1;
5546 /* Do we need to flush the list? */
5547 if (working.len >= (int)peer->maxDgramPackets
5548 || working.len >= (int)call->nDgramPackets
5549 || working.len >= (int)call->cwind
5550 || list[i]->header.serial
5551 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5553 rxi_SendList(call, &last, istack, 1);
5554 /* If the call enters an error state stop sending, or if
5555 * we entered congestion recovery mode, stop sending */
5557 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5562 working.resending = 0;
5563 working.list = &list[i + 1];
5566 if (working.len != 0) {
5567 osi_Panic("rxi_SendList error");
5569 working.list = &list[i + 1];
5573 /* Send the whole list when the call is in receive mode, when
5574 * the call is in eof mode, when we are in fast recovery mode,
5575 * and when we have the last packet */
5576 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5577 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5578 || (call->flags & RX_CALL_FAST_RECOVER)) {
5579 /* Check for the case where the current list contains
5580 * an acked packet. Since we always send retransmissions
5581 * in a separate packet, we only need to check the first
5582 * packet in the list */
5583 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5587 rxi_SendList(call, &last, istack, morePackets);
5588 /* If the call enters an error state stop sending, or if
5589 * we entered congestion recovery mode, stop sending */
5590 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5594 rxi_SendList(call, &working, istack, 0);
5596 } else if (last.len > 0) {
5597 rxi_SendList(call, &last, istack, 0);
5598 /* Packets which are in 'working' are not sent by this call */
5602 #ifdef RX_ENABLE_LOCKS
5603 /* Call rxi_Start, below, but with the call lock held. */
5605 rxi_StartUnlocked(struct rxevent *event,
5606 void *arg0, void *arg1, int istack)
5608 struct rx_call *call = arg0;
5610 MUTEX_ENTER(&call->lock);
5611 rxi_Start(event, call, arg1, istack);
5612 MUTEX_EXIT(&call->lock);
5614 #endif /* RX_ENABLE_LOCKS */
5616 /* This routine is called when new packets are readied for
5617 * transmission and when retransmission may be necessary, or when the
5618 * transmission window or burst count are favourable. This should be
5619 * better optimized for new packets, the usual case, now that we've
5620 * got rid of queues of send packets. XXXXXXXXXXX */
5622 rxi_Start(struct rxevent *event,
5623 void *arg0, void *arg1, int istack)
5625 struct rx_call *call = arg0;
5627 struct rx_packet *p;
5628 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5629 struct clock now, usenow, retryTime;
5634 /* If rxi_Start is being called as a result of a resend event,
5635 * then make sure that the event pointer is removed from the call
5636 * structure, since there is no longer a per-call retransmission
5638 if (event && event == call->resendEvent) {
5639 MUTEX_ENTER(&rx_refcnt_mutex);
5640 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5641 MUTEX_EXIT(&rx_refcnt_mutex);
5642 call->resendEvent = NULL;
5643 if (queue_IsEmpty(&call->tq)) {
5650 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5651 if (rx_stats_active)
5652 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5657 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5659 clock_GetTime(&now);
5662 /* Send (or resend) any packets that need it, subject to
5663 * window restrictions and congestion burst control
5664 * restrictions. Ask for an ack on the last packet sent in
5665 * this burst. For now, we're relying upon the window being
5666 * considerably bigger than the largest number of packets that
5667 * are typically sent at once by one initial call to
5668 * rxi_Start. This is probably bogus (perhaps we should ask
5669 * for an ack when we're half way through the current
5670 * window?). Also, for non file transfer applications, this
5671 * may end up asking for an ack for every packet. Bogus. XXXX
5674 * But check whether we're here recursively, and let the other guy
5677 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5678 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5679 call->flags |= RX_CALL_TQ_BUSY;
5681 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5683 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5684 call->flags &= ~RX_CALL_NEED_START;
5685 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5687 maxXmitPackets = MIN(call->twind, call->cwind);
5688 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5689 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5690 /* We shouldn't be sending packets if a thread is waiting
5691 * to initiate congestion recovery */
5692 dpf(("call %d waiting to initiate fast recovery\n",
5693 *(call->callNumber)));
5697 && (call->flags & RX_CALL_FAST_RECOVER)) {
5698 /* Only send one packet during fast recovery */
5699 dpf(("call %d restricted to one packet per send during fast recovery\n",
5700 *(call->callNumber)));
5703 #ifdef RX_TRACK_PACKETS
5704 if ((p->flags & RX_PKTFLAG_FREE)
5705 || (!queue_IsEnd(&call->tq, nxp)
5706 && (nxp->flags & RX_PKTFLAG_FREE))
5707 || (p == (struct rx_packet *)&rx_freePacketQueue)
5708 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5709 osi_Panic("rxi_Start: xmit queue clobbered");
5712 if (p->flags & RX_PKTFLAG_ACKED) {
5713 /* Since we may block, don't trust this */
5714 usenow.sec = usenow.usec = 0;
5715 if (rx_stats_active)
5716 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
5717 continue; /* Ignore this packet if it has been acknowledged */
5720 /* Turn off all flags except these ones, which are the same
5721 * on each transmission */
5722 p->header.flags &= RX_PRESET_FLAGS;
5724 if (p->header.seq >=
5725 call->tfirst + MIN((int)call->twind,
5726 (int)(call->nSoftAcked +
5728 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5729 /* Note: if we're waiting for more window space, we can
5730 * still send retransmits; hence we don't return here, but
5731 * break out to schedule a retransmit event */
5732 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
5733 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
5738 /* Transmit the packet if it needs to be sent. */
5739 if (!clock_Lt(&now, &p->retryTime)) {
5740 if (nXmitPackets == maxXmitPackets) {
5741 rxi_SendXmitList(call, call->xmitList,
5742 nXmitPackets, istack);
5745 dpf(("call %d xmit packet %"AFS_PTR_FMT" now %u.%06u retryTime %u.%06u\n",
5746 *(call->callNumber), p,
5748 p->retryTime.sec, p->retryTime.usec));
5749 call->xmitList[nXmitPackets++] = p;
5753 /* xmitList now hold pointers to all of the packets that are
5754 * ready to send. Now we loop to send the packets */
5755 if (nXmitPackets > 0) {
5756 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
5760 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5762 * TQ references no longer protected by this flag; they must remain
5763 * protected by the global lock.
5765 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5766 call->flags &= ~RX_CALL_TQ_BUSY;
5767 rxi_WakeUpTransmitQueue(call);
5771 /* We went into the error state while sending packets. Now is
5772 * the time to reset the call. This will also inform the using
5773 * process that the call is in an error state.
5775 if (rx_stats_active)
5776 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
5777 call->flags &= ~RX_CALL_TQ_BUSY;
5778 rxi_WakeUpTransmitQueue(call);
5779 rxi_CallError(call, call->error);
5782 #ifdef RX_ENABLE_LOCKS
5783 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5785 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5786 /* Some packets have received acks. If they all have, we can clear
5787 * the transmit queue.
5790 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5791 if (p->header.seq < call->tfirst
5792 && (p->flags & RX_PKTFLAG_ACKED)) {
5794 #ifdef RX_TRACK_PACKETS
5795 p->flags &= ~RX_PKTFLAG_TQ;
5797 #ifdef RXDEBUG_PACKET
5805 call->flags |= RX_CALL_TQ_CLEARME;
5807 #endif /* RX_ENABLE_LOCKS */
5808 /* Don't bother doing retransmits if the TQ is cleared. */
5809 if (call->flags & RX_CALL_TQ_CLEARME) {
5810 rxi_ClearTransmitQueue(call, 1);
5812 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5815 /* Always post a resend event, if there is anything in the
5816 * queue, and resend is possible. There should be at least
5817 * one unacknowledged packet in the queue ... otherwise none
5818 * of these packets should be on the queue in the first place.
5820 if (call->resendEvent) {
5821 /* Cancel the existing event and post a new one */
5822 rxevent_Cancel(call->resendEvent, call,
5823 RX_CALL_REFCOUNT_RESEND);
5826 /* The retry time is the retry time on the first unacknowledged
5827 * packet inside the current window */
5829 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5830 /* Don't set timers for packets outside the window */
5831 if (p->header.seq >= call->tfirst + call->twind) {
5835 if (!(p->flags & RX_PKTFLAG_ACKED)
5836 && !clock_IsZero(&p->retryTime)) {
5838 retryTime = p->retryTime;
5843 /* Post a new event to re-run rxi_Start when retries may be needed */
5844 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5845 #ifdef RX_ENABLE_LOCKS
5846 MUTEX_ENTER(&rx_refcnt_mutex);
5847 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5848 MUTEX_EXIT(&rx_refcnt_mutex);
5850 rxevent_PostNow2(&retryTime, &usenow,
5852 (void *)call, 0, istack);
5853 #else /* RX_ENABLE_LOCKS */
5855 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5856 (void *)call, 0, istack);
5857 #endif /* RX_ENABLE_LOCKS */
5860 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5861 } while (call->flags & RX_CALL_NEED_START);
5863 * TQ references no longer protected by this flag; they must remain
5864 * protected by the global lock.
5866 call->flags &= ~RX_CALL_TQ_BUSY;
5867 rxi_WakeUpTransmitQueue(call);
5869 call->flags |= RX_CALL_NEED_START;
5871 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5873 if (call->resendEvent) {
5874 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5879 /* Also adjusts the keep alive parameters for the call, to reflect
5880 * that we have just sent a packet (so keep alives aren't sent
5883 rxi_Send(struct rx_call *call, struct rx_packet *p,
5886 struct rx_connection *conn = call->conn;
5888 /* Stamp each packet with the user supplied status */
5889 p->header.userStatus = call->localStatus;
5891 /* Allow the security object controlling this call's security to
5892 * make any last-minute changes to the packet */
5893 RXS_SendPacket(conn->securityObject, call, p);
5895 /* Since we're about to send SOME sort of packet to the peer, it's
5896 * safe to nuke any scheduled end-of-packets ack */
5897 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5899 /* Actually send the packet, filling in more connection-specific fields */
5900 MUTEX_EXIT(&call->lock);
5901 MUTEX_ENTER(&rx_refcnt_mutex);
5902 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5903 MUTEX_EXIT(&rx_refcnt_mutex);
5904 rxi_SendPacket(call, conn, p, istack);
5905 MUTEX_ENTER(&rx_refcnt_mutex);
5906 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5907 MUTEX_EXIT(&rx_refcnt_mutex);
5908 MUTEX_ENTER(&call->lock);
5910 /* Update last send time for this call (for keep-alive
5911 * processing), and for the connection (so that we can discover
5912 * idle connections) */
5913 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5914 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
5915 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
5917 conn->lastSendTime = call->lastSendTime = clock_Sec();
5918 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
5919 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5920 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
5921 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
5922 RX_ACK_PING_RESPONSE)))
5923 call->lastSendData = call->lastSendTime;
5927 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5928 * that things are fine. Also called periodically to guarantee that nothing
5929 * falls through the cracks (e.g. (error + dally) connections have keepalive
5930 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5932 * haveCTLock Set if calling from rxi_ReapConnections
5934 #ifdef RX_ENABLE_LOCKS
5936 rxi_CheckCall(struct rx_call *call, int haveCTLock)
5937 #else /* RX_ENABLE_LOCKS */
5939 rxi_CheckCall(struct rx_call *call)
5940 #endif /* RX_ENABLE_LOCKS */
5942 struct rx_connection *conn = call->conn;
5944 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
5945 afs_uint32 fudgeFactor;
5949 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5950 if (call->flags & RX_CALL_TQ_BUSY) {
5951 /* Call is active and will be reset by rxi_Start if it's
5952 * in an error state.
5957 /* RTT + 8*MDEV, rounded up to the next second. */
5958 fudgeFactor = (((afs_uint32) conn->peer->rtt >> 3) +
5959 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5961 deadTime = conn->secondsUntilDead + fudgeFactor;
5963 /* These are computed to the second (+- 1 second). But that's
5964 * good enough for these values, which should be a significant
5965 * number of seconds. */
5966 if (now > (call->lastReceiveTime + deadTime)) {
5967 if (call->state == RX_STATE_ACTIVE) {
5969 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5971 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5972 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5973 ip_stack_t *ipst = ns->netstack_ip;
5975 ire = ire_cache_lookup(conn->peer->host
5976 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5978 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5980 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5987 if (ire && ire->ire_max_frag > 0)
5988 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
5990 #if defined(GLOBAL_NETSTACKID)
5994 #endif /* ADAPT_PMTU */
5995 cerror = RX_CALL_DEAD;
5998 #ifdef RX_ENABLE_LOCKS
5999 /* Cancel pending events */
6000 rxevent_Cancel(call->delayedAckEvent, call,
6001 RX_CALL_REFCOUNT_DELAY);
6002 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
6003 rxevent_Cancel(call->keepAliveEvent, call,
6004 RX_CALL_REFCOUNT_ALIVE);
6005 MUTEX_ENTER(&rx_refcnt_mutex);
6006 if (call->refCount == 0) {
6007 rxi_FreeCall(call, haveCTLock);
6008 MUTEX_EXIT(&rx_refcnt_mutex);
6011 MUTEX_EXIT(&rx_refcnt_mutex);
6013 #else /* RX_ENABLE_LOCKS */
6014 rxi_FreeCall(call, 0);
6016 #endif /* RX_ENABLE_LOCKS */
6018 /* Non-active calls are destroyed if they are not responding
6019 * to pings; active calls are simply flagged in error, so the
6020 * attached process can die reasonably gracefully. */
6023 if (conn->idleDeadTime) {
6024 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6027 /* see if we have a non-activity timeout */
6028 if (call->startWait && idleDeadTime
6029 && ((call->startWait + idleDeadTime) < now) &&
6030 (call->flags & RX_CALL_READER_WAIT)) {
6031 if (call->state == RX_STATE_ACTIVE) {
6032 cerror = RX_CALL_TIMEOUT;
6036 if (call->lastSendData && idleDeadTime && (conn->idleDeadErr != 0)
6037 && ((call->lastSendData + idleDeadTime) < now)) {
6038 if (call->state == RX_STATE_ACTIVE) {
6039 cerror = conn->idleDeadErr;
6044 if (conn->hardDeadTime) {
6045 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6048 /* see if we have a hard timeout */
6050 && (now > (hardDeadTime + call->startTime.sec))) {
6051 if (call->state == RX_STATE_ACTIVE)
6052 rxi_CallError(call, RX_CALL_TIMEOUT);
6057 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT
6058 && call->lastReceiveTime) {
6059 int oldMTU = conn->peer->ifMTU;
6061 /* if we thought we could send more, perhaps things got worse */
6062 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6063 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6064 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6065 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6067 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6069 /* minimum capped in SetPeerMtu */
6070 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6073 conn->lastPacketSize = 0;
6075 /* needed so ResetCall doesn't clobber us. */
6076 call->MTU = conn->peer->ifMTU;
6078 /* if we never succeeded, let the error pass out as-is */
6079 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6080 cerror = conn->msgsizeRetryErr;
6083 rxi_CallError(call, cerror);
6088 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6090 struct rx_connection *conn = arg1;
6091 struct rx_header theader;
6092 char tbuffer[1 + sizeof(struct rx_header)];
6093 struct sockaddr_in taddr;
6096 struct iovec tmpiov[2];
6099 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6102 tp = &tbuffer[sizeof(struct rx_header)];
6103 taddr.sin_family = AF_INET;
6104 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6105 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6106 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6107 taddr.sin_len = sizeof(struct sockaddr_in);
6109 memset(&theader, 0, sizeof(theader));
6110 theader.epoch = htonl(999);
6112 theader.callNumber = 0;
6115 theader.type = RX_PACKET_TYPE_VERSION;
6116 theader.flags = RX_LAST_PACKET;
6117 theader.serviceId = 0;
6119 memcpy(tbuffer, &theader, sizeof(theader));
6120 memcpy(tp, &a, sizeof(a));
6121 tmpiov[0].iov_base = tbuffer;
6122 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6124 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6126 MUTEX_ENTER(&conn->conn_data_lock);
6127 MUTEX_ENTER(&rx_refcnt_mutex);
6128 /* Only reschedule ourselves if the connection would not be destroyed */
6129 if (conn->refCount <= 1) {
6130 conn->natKeepAliveEvent = NULL;
6131 MUTEX_EXIT(&rx_refcnt_mutex);
6132 MUTEX_EXIT(&conn->conn_data_lock);
6133 rx_DestroyConnection(conn); /* drop the reference for this */
6135 conn->refCount--; /* drop the reference for this */
6136 MUTEX_EXIT(&rx_refcnt_mutex);
6137 conn->natKeepAliveEvent = NULL;
6138 rxi_ScheduleNatKeepAliveEvent(conn);
6139 MUTEX_EXIT(&conn->conn_data_lock);
6144 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6146 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6147 struct clock when, now;
6148 clock_GetTime(&now);
6150 when.sec += conn->secondsUntilNatPing;
6151 MUTEX_ENTER(&rx_refcnt_mutex);
6152 conn->refCount++; /* hold a reference for this */
6153 MUTEX_EXIT(&rx_refcnt_mutex);
6154 conn->natKeepAliveEvent =
6155 rxevent_PostNow(&when, &now, rxi_NatKeepAliveEvent, conn, 0);
6160 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6162 MUTEX_ENTER(&conn->conn_data_lock);
6163 conn->secondsUntilNatPing = seconds;
6165 rxi_ScheduleNatKeepAliveEvent(conn);
6166 MUTEX_EXIT(&conn->conn_data_lock);
6170 rxi_NatKeepAliveOn(struct rx_connection *conn)
6172 MUTEX_ENTER(&conn->conn_data_lock);
6173 rxi_ScheduleNatKeepAliveEvent(conn);
6174 MUTEX_EXIT(&conn->conn_data_lock);
6177 /* When a call is in progress, this routine is called occasionally to
6178 * make sure that some traffic has arrived (or been sent to) the peer.
6179 * If nothing has arrived in a reasonable amount of time, the call is
6180 * declared dead; if nothing has been sent for a while, we send a
6181 * keep-alive packet (if we're actually trying to keep the call alive)
6184 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6186 struct rx_call *call = arg1;
6187 struct rx_connection *conn;
6190 MUTEX_ENTER(&rx_refcnt_mutex);
6191 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6192 MUTEX_EXIT(&rx_refcnt_mutex);
6193 MUTEX_ENTER(&call->lock);
6194 if (event == call->keepAliveEvent)
6195 call->keepAliveEvent = NULL;
6198 #ifdef RX_ENABLE_LOCKS
6199 if (rxi_CheckCall(call, 0)) {
6200 MUTEX_EXIT(&call->lock);
6203 #else /* RX_ENABLE_LOCKS */
6204 if (rxi_CheckCall(call))
6206 #endif /* RX_ENABLE_LOCKS */
6208 /* Don't try to keep alive dallying calls */
6209 if (call->state == RX_STATE_DALLY) {
6210 MUTEX_EXIT(&call->lock);
6215 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6216 /* Don't try to send keepalives if there is unacknowledged data */
6217 /* the rexmit code should be good enough, this little hack
6218 * doesn't quite work XXX */
6219 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6221 rxi_ScheduleKeepAliveEvent(call);
6222 MUTEX_EXIT(&call->lock);
6225 /* Does what's on the nameplate. */
6227 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy)
6229 struct rx_call *call = arg1;
6230 struct rx_connection *conn;
6232 MUTEX_ENTER(&rx_refcnt_mutex);
6233 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6234 MUTEX_EXIT(&rx_refcnt_mutex);
6235 MUTEX_ENTER(&call->lock);
6237 if (event == call->growMTUEvent)
6238 call->growMTUEvent = NULL;
6240 #ifdef RX_ENABLE_LOCKS
6241 if (rxi_CheckCall(call, 0)) {
6242 MUTEX_EXIT(&call->lock);
6245 #else /* RX_ENABLE_LOCKS */
6246 if (rxi_CheckCall(call))
6248 #endif /* RX_ENABLE_LOCKS */
6250 /* Don't bother with dallying calls */
6251 if (call->state == RX_STATE_DALLY) {
6252 MUTEX_EXIT(&call->lock);
6259 * keep being scheduled, just don't do anything if we're at peak,
6260 * or we're not set up to be properly handled (idle timeout required)
6262 if ((conn->peer->maxPacketSize != 0) &&
6263 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6264 (conn->idleDeadErr))
6265 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6266 rxi_ScheduleGrowMTUEvent(call, 0);
6267 MUTEX_EXIT(&call->lock);
6271 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6273 if (!call->keepAliveEvent) {
6274 struct clock when, now;
6275 clock_GetTime(&now);
6277 when.sec += call->conn->secondsUntilPing;
6278 MUTEX_ENTER(&rx_refcnt_mutex);
6279 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6280 MUTEX_EXIT(&rx_refcnt_mutex);
6281 call->keepAliveEvent =
6282 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
6287 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6289 if (!call->growMTUEvent) {
6290 struct clock when, now;
6292 clock_GetTime(&now);
6295 if (call->conn->secondsUntilPing)
6296 secs = (6*call->conn->secondsUntilPing)-1;
6298 if (call->conn->secondsUntilDead)
6299 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6303 MUTEX_ENTER(&rx_refcnt_mutex);
6304 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6305 MUTEX_EXIT(&rx_refcnt_mutex);
6306 call->growMTUEvent =
6307 rxevent_PostNow(&when, &now, rxi_GrowMTUEvent, call, 0);
6311 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6313 rxi_KeepAliveOn(struct rx_call *call)
6315 /* Pretend last packet received was received now--i.e. if another
6316 * packet isn't received within the keep alive time, then the call
6317 * will die; Initialize last send time to the current time--even
6318 * if a packet hasn't been sent yet. This will guarantee that a
6319 * keep-alive is sent within the ping time */
6320 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6321 rxi_ScheduleKeepAliveEvent(call);
6325 rxi_GrowMTUOn(struct rx_call *call)
6327 struct rx_connection *conn = call->conn;
6328 MUTEX_ENTER(&conn->conn_data_lock);
6329 conn->lastPingSizeSer = conn->lastPingSize = 0;
6330 MUTEX_EXIT(&conn->conn_data_lock);
6331 rxi_ScheduleGrowMTUEvent(call, 1);
6334 /* This routine is called to send connection abort messages
6335 * that have been delayed to throttle looping clients. */
6337 rxi_SendDelayedConnAbort(struct rxevent *event,
6338 void *arg1, void *unused)
6340 struct rx_connection *conn = arg1;
6343 struct rx_packet *packet;
6345 MUTEX_ENTER(&conn->conn_data_lock);
6346 conn->delayedAbortEvent = NULL;
6347 error = htonl(conn->error);
6349 MUTEX_EXIT(&conn->conn_data_lock);
6350 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6353 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6354 RX_PACKET_TYPE_ABORT, (char *)&error,
6356 rxi_FreePacket(packet);
6360 /* This routine is called to send call abort messages
6361 * that have been delayed to throttle looping clients. */
6363 rxi_SendDelayedCallAbort(struct rxevent *event,
6364 void *arg1, void *dummy)
6366 struct rx_call *call = arg1;
6369 struct rx_packet *packet;
6371 MUTEX_ENTER(&call->lock);
6372 call->delayedAbortEvent = NULL;
6373 error = htonl(call->error);
6375 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6378 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6379 (char *)&error, sizeof(error), 0);
6380 rxi_FreePacket(packet);
6382 MUTEX_EXIT(&call->lock);
6383 MUTEX_ENTER(&rx_refcnt_mutex);
6384 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6385 MUTEX_EXIT(&rx_refcnt_mutex);
6388 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6389 * seconds) to ask the client to authenticate itself. The routine
6390 * issues a challenge to the client, which is obtained from the
6391 * security object associated with the connection */
6393 rxi_ChallengeEvent(struct rxevent *event,
6394 void *arg0, void *arg1, int tries)
6396 struct rx_connection *conn = arg0;
6398 conn->challengeEvent = NULL;
6399 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6400 struct rx_packet *packet;
6401 struct clock when, now;
6404 /* We've failed to authenticate for too long.
6405 * Reset any calls waiting for authentication;
6406 * they are all in RX_STATE_PRECALL.
6410 MUTEX_ENTER(&conn->conn_call_lock);
6411 for (i = 0; i < RX_MAXCALLS; i++) {
6412 struct rx_call *call = conn->call[i];
6414 MUTEX_ENTER(&call->lock);
6415 if (call->state == RX_STATE_PRECALL) {
6416 rxi_CallError(call, RX_CALL_DEAD);
6417 rxi_SendCallAbort(call, NULL, 0, 0);
6419 MUTEX_EXIT(&call->lock);
6422 MUTEX_EXIT(&conn->conn_call_lock);
6426 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6428 /* If there's no packet available, do this later. */
6429 RXS_GetChallenge(conn->securityObject, conn, packet);
6430 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6431 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6432 rxi_FreePacket(packet);
6434 clock_GetTime(&now);
6436 when.sec += RX_CHALLENGE_TIMEOUT;
6437 conn->challengeEvent =
6438 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
6443 /* Call this routine to start requesting the client to authenticate
6444 * itself. This will continue until authentication is established,
6445 * the call times out, or an invalid response is returned. The
6446 * security object associated with the connection is asked to create
6447 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6448 * defined earlier. */
6450 rxi_ChallengeOn(struct rx_connection *conn)
6452 if (!conn->challengeEvent) {
6453 RXS_CreateChallenge(conn->securityObject, conn);
6454 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6459 /* rxi_ComputeRoundTripTime is called with peer locked. */
6460 /* peer may be null */
6462 rxi_ComputeRoundTripTime(struct rx_packet *p,
6463 struct rx_ackPacket *ack,
6464 struct rx_peer *peer,
6467 struct clock thisRtt, *sentp;
6471 /* If the ACK is delayed, then do nothing */
6472 if (ack->reason == RX_ACK_DELAY)
6475 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6476 * their RTT multiple times, so only include the RTT of the last packet
6478 if (p->flags & RX_JUMBO_PACKET)
6481 /* Use the serial number to determine which transmission the ACK is for,
6482 * and set the sent time to match this. If we have no serial number, then
6483 * only use the ACK for RTT calculations if the packet has not been
6487 serial = ntohl(ack->serial);
6489 if (serial == p->header.serial) {
6490 sentp = &p->timeSent;
6491 } else if (serial == p->firstSerial) {
6492 sentp = &p->firstSent;
6493 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6494 sentp = &p->firstSent;
6498 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6499 sentp = &p->firstSent;
6506 if (clock_Lt(&thisRtt, sentp))
6507 return; /* somebody set the clock back, don't count this time. */
6509 clock_Sub(&thisRtt, sentp);
6510 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6511 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6513 if (clock_IsZero(&thisRtt)) {
6515 * The actual round trip time is shorter than the
6516 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6517 * Since we can't tell which at the moment we will assume 1ms.
6519 thisRtt.usec = 1000;
6522 if (rx_stats_active) {
6523 MUTEX_ENTER(&rx_stats_mutex);
6524 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6525 rx_stats.minRtt = thisRtt;
6526 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6527 if (thisRtt.sec > 60) {
6528 MUTEX_EXIT(&rx_stats_mutex);
6529 return; /* somebody set the clock ahead */
6531 rx_stats.maxRtt = thisRtt;
6533 clock_Add(&rx_stats.totalRtt, &thisRtt);
6534 rx_atomic_inc(&rx_stats.nRttSamples);
6535 MUTEX_EXIT(&rx_stats_mutex);
6538 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6540 /* Apply VanJacobson round-trip estimations */
6545 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
6546 * srtt is stored as fixed point with 3 bits after the binary
6547 * point (i.e., scaled by 8). The following magic is
6548 * equivalent to the smoothing algorithm in rfc793 with an
6549 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6550 * srtt'*8 = rtt + srtt*7
6551 * srtt'*8 = srtt*8 + rtt - srtt
6552 * srtt' = srtt + rtt/8 - srtt/8
6553 * srtt' = srtt + (rtt - srtt)/8
6556 delta = _8THMSEC(&thisRtt) - peer->rtt;
6557 peer->rtt += (delta >> 3);
6560 * We accumulate a smoothed rtt variance (actually, a smoothed
6561 * mean difference), then set the retransmit timer to smoothed
6562 * rtt + 4 times the smoothed variance (was 2x in van's original
6563 * paper, but 4x works better for me, and apparently for him as
6565 * rttvar is stored as
6566 * fixed point with 2 bits after the binary point (scaled by
6567 * 4). The following is equivalent to rfc793 smoothing with
6568 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6569 * rttvar'*4 = rttvar*3 + |delta|
6570 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6571 * rttvar' = rttvar + |delta|/4 - rttvar/4
6572 * rttvar' = rttvar + (|delta| - rttvar)/4
6573 * This replaces rfc793's wired-in beta.
6574 * dev*4 = dev*4 + (|actual - expected| - dev)
6580 delta -= (peer->rtt_dev << 1);
6581 peer->rtt_dev += (delta >> 3);
6583 /* I don't have a stored RTT so I start with this value. Since I'm
6584 * probably just starting a call, and will be pushing more data down
6585 * this, I expect congestion to increase rapidly. So I fudge a
6586 * little, and I set deviance to half the rtt. In practice,
6587 * deviance tends to approach something a little less than
6588 * half the smoothed rtt. */
6589 peer->rtt = _8THMSEC(&thisRtt) + 8;
6590 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
6592 /* the timeout is RTT + 4*MDEV + rx_minPeerTimeout msec.
6593 * This is because one end or the other of these connections is usually
6594 * in a user process, and can be switched and/or swapped out. So on fast,
6595 * reliable networks, the timeout would otherwise be too short. */
6596 rtt_timeout = ((peer->rtt >> 3) + peer->rtt_dev) + rx_minPeerTimeout;
6597 clock_Zero(&(peer->timeout));
6598 clock_Addmsec(&(peer->timeout), rtt_timeout);
6600 /* Reset the backedOff flag since we just computed a new timeout value */
6601 peer->backedOff = 0;
6603 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6604 p->header.callNumber, p, MSEC(&thisRtt), peer->rtt >> 3, peer->rtt_dev >> 2, (peer->timeout.sec), (peer->timeout.usec)));
6608 /* Find all server connections that have not been active for a long time, and
6611 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6613 struct clock now, when;
6614 clock_GetTime(&now);
6616 /* Find server connection structures that haven't been used for
6617 * greater than rx_idleConnectionTime */
6619 struct rx_connection **conn_ptr, **conn_end;
6620 int i, havecalls = 0;
6621 MUTEX_ENTER(&rx_connHashTable_lock);
6622 for (conn_ptr = &rx_connHashTable[0], conn_end =
6623 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6625 struct rx_connection *conn, *next;
6626 struct rx_call *call;
6630 for (conn = *conn_ptr; conn; conn = next) {
6631 /* XXX -- Shouldn't the connection be locked? */
6634 for (i = 0; i < RX_MAXCALLS; i++) {
6635 call = conn->call[i];
6639 code = MUTEX_TRYENTER(&call->lock);
6642 #ifdef RX_ENABLE_LOCKS
6643 result = rxi_CheckCall(call, 1);
6644 #else /* RX_ENABLE_LOCKS */
6645 result = rxi_CheckCall(call);
6646 #endif /* RX_ENABLE_LOCKS */
6647 MUTEX_EXIT(&call->lock);
6649 /* If CheckCall freed the call, it might
6650 * have destroyed the connection as well,
6651 * which screws up the linked lists.
6657 if (conn->type == RX_SERVER_CONNECTION) {
6658 /* This only actually destroys the connection if
6659 * there are no outstanding calls */
6660 MUTEX_ENTER(&conn->conn_data_lock);
6661 MUTEX_ENTER(&rx_refcnt_mutex);
6662 if (!havecalls && !conn->refCount
6663 && ((conn->lastSendTime + rx_idleConnectionTime) <
6665 conn->refCount++; /* it will be decr in rx_DestroyConn */
6666 MUTEX_EXIT(&rx_refcnt_mutex);
6667 MUTEX_EXIT(&conn->conn_data_lock);
6668 #ifdef RX_ENABLE_LOCKS
6669 rxi_DestroyConnectionNoLock(conn);
6670 #else /* RX_ENABLE_LOCKS */
6671 rxi_DestroyConnection(conn);
6672 #endif /* RX_ENABLE_LOCKS */
6674 #ifdef RX_ENABLE_LOCKS
6676 MUTEX_EXIT(&rx_refcnt_mutex);
6677 MUTEX_EXIT(&conn->conn_data_lock);
6679 #endif /* RX_ENABLE_LOCKS */
6683 #ifdef RX_ENABLE_LOCKS
6684 while (rx_connCleanup_list) {
6685 struct rx_connection *conn;
6686 conn = rx_connCleanup_list;
6687 rx_connCleanup_list = rx_connCleanup_list->next;
6688 MUTEX_EXIT(&rx_connHashTable_lock);
6689 rxi_CleanupConnection(conn);
6690 MUTEX_ENTER(&rx_connHashTable_lock);
6692 MUTEX_EXIT(&rx_connHashTable_lock);
6693 #endif /* RX_ENABLE_LOCKS */
6696 /* Find any peer structures that haven't been used (haven't had an
6697 * associated connection) for greater than rx_idlePeerTime */
6699 struct rx_peer **peer_ptr, **peer_end;
6703 * Why do we need to hold the rx_peerHashTable_lock across
6704 * the incrementing of peer_ptr since the rx_peerHashTable
6705 * array is not changing? We don't.
6707 * By dropping the lock periodically we can permit other
6708 * activities to be performed while a rxi_ReapConnections
6709 * call is in progress. The goal of reap connections
6710 * is to clean up quickly without causing large amounts
6711 * of contention. Therefore, it is important that global
6712 * mutexes not be held for extended periods of time.
6714 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6715 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6717 struct rx_peer *peer, *next, *prev;
6719 MUTEX_ENTER(&rx_peerHashTable_lock);
6720 for (prev = peer = *peer_ptr; peer; peer = next) {
6722 code = MUTEX_TRYENTER(&peer->peer_lock);
6723 if ((code) && (peer->refCount == 0)
6724 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6725 rx_interface_stat_p rpc_stat, nrpc_stat;
6729 * now know that this peer object is one to be
6730 * removed from the hash table. Once it is removed
6731 * it can't be referenced by other threads.
6732 * Lets remove it first and decrement the struct
6733 * nPeerStructs count.
6735 if (peer == *peer_ptr) {
6741 if (rx_stats_active)
6742 rx_atomic_dec(&rx_stats.nPeerStructs);
6745 * Now if we hold references on 'prev' and 'next'
6746 * we can safely drop the rx_peerHashTable_lock
6747 * while we destroy this 'peer' object.
6753 MUTEX_EXIT(&rx_peerHashTable_lock);
6755 MUTEX_EXIT(&peer->peer_lock);
6756 MUTEX_DESTROY(&peer->peer_lock);
6758 (&peer->rpcStats, rpc_stat, nrpc_stat,
6759 rx_interface_stat)) {
6760 unsigned int num_funcs;
6763 queue_Remove(&rpc_stat->queue_header);
6764 queue_Remove(&rpc_stat->all_peers);
6765 num_funcs = rpc_stat->stats[0].func_total;
6767 sizeof(rx_interface_stat_t) +
6768 rpc_stat->stats[0].func_total *
6769 sizeof(rx_function_entry_v1_t);
6771 rxi_Free(rpc_stat, space);
6773 MUTEX_ENTER(&rx_rpc_stats);
6774 rxi_rpc_peer_stat_cnt -= num_funcs;
6775 MUTEX_EXIT(&rx_rpc_stats);
6780 * Regain the rx_peerHashTable_lock and
6781 * decrement the reference count on 'prev'
6784 MUTEX_ENTER(&rx_peerHashTable_lock);
6791 MUTEX_EXIT(&peer->peer_lock);
6796 MUTEX_EXIT(&rx_peerHashTable_lock);
6800 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6801 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6802 * GC, just below. Really, we shouldn't have to keep moving packets from
6803 * one place to another, but instead ought to always know if we can
6804 * afford to hold onto a packet in its particular use. */
6805 MUTEX_ENTER(&rx_freePktQ_lock);
6806 if (rx_waitingForPackets) {
6807 rx_waitingForPackets = 0;
6808 #ifdef RX_ENABLE_LOCKS
6809 CV_BROADCAST(&rx_waitingForPackets_cv);
6811 osi_rxWakeup(&rx_waitingForPackets);
6814 MUTEX_EXIT(&rx_freePktQ_lock);
6817 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6818 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6822 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6823 * rx.h is sort of strange this is better. This is called with a security
6824 * object before it is discarded. Each connection using a security object has
6825 * its own refcount to the object so it won't actually be freed until the last
6826 * connection is destroyed.
6828 * This is the only rxs module call. A hold could also be written but no one
6832 rxs_Release(struct rx_securityClass *aobj)
6834 return RXS_Close(aobj);
6838 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6839 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6840 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6841 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6843 /* Adjust our estimate of the transmission rate to this peer, given
6844 * that the packet p was just acked. We can adjust peer->timeout and
6845 * call->twind. Pragmatically, this is called
6846 * only with packets of maximal length.
6847 * Called with peer and call locked.
6851 rxi_ComputeRate(struct rx_peer *peer, struct rx_call *call,
6852 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6854 afs_int32 xferSize, xferMs;
6858 /* Count down packets */
6859 if (peer->rateFlag > 0)
6861 /* Do nothing until we're enabled */
6862 if (peer->rateFlag != 0)
6867 /* Count only when the ack seems legitimate */
6868 switch (ackReason) {
6869 case RX_ACK_REQUESTED:
6871 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6875 case RX_ACK_PING_RESPONSE:
6876 if (p) /* want the response to ping-request, not data send */
6878 clock_GetTime(&newTO);
6879 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6880 clock_Sub(&newTO, &call->pingRequestTime);
6881 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6885 xferSize = rx_AckDataSize(rx_maxSendWindow) + RX_HEADER_SIZE;
6892 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %d.%06d, rtt %u, ps %u)\n",
6893 ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"),
6894 xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
6896 /* Track only packets that are big enough. */
6897 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6901 /* absorb RTT data (in milliseconds) for these big packets */
6902 if (peer->smRtt == 0) {
6903 peer->smRtt = xferMs;
6905 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6910 if (peer->countDown) {
6914 peer->countDown = 10; /* recalculate only every so often */
6916 /* In practice, we can measure only the RTT for full packets,
6917 * because of the way Rx acks the data that it receives. (If it's
6918 * smaller than a full packet, it often gets implicitly acked
6919 * either by the call response (from a server) or by the next call
6920 * (from a client), and either case confuses transmission times
6921 * with processing times.) Therefore, replace the above
6922 * more-sophisticated processing with a simpler version, where the
6923 * smoothed RTT is kept for full-size packets, and the time to
6924 * transmit a windowful of full-size packets is simply RTT *
6925 * windowSize. Again, we take two steps:
6926 - ensure the timeout is large enough for a single packet's RTT;
6927 - ensure that the window is small enough to fit in the desired timeout.*/
6929 /* First, the timeout check. */
6930 minTime = peer->smRtt;
6931 /* Get a reasonable estimate for a timeout period */
6933 newTO.sec = minTime / 1000;
6934 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6936 /* Increase the timeout period so that we can always do at least
6937 * one packet exchange */
6938 if (clock_Gt(&newTO, &peer->timeout)) {
6940 dpf(("CONG peer %lx/%u: timeout %d.%06d ==> %ld.%06d (rtt %u)\n",
6941 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec,
6942 newTO.sec, newTO.usec, peer->smRtt));
6944 peer->timeout = newTO;
6947 /* Now, get an estimate for the transmit window size. */
6948 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6949 /* Now, convert to the number of full packets that could fit in a
6950 * reasonable fraction of that interval */
6951 minTime /= (peer->smRtt << 1);
6952 minTime = MAX(minTime, rx_minPeerTimeout);
6953 xferSize = minTime; /* (make a copy) */
6955 /* Now clamp the size to reasonable bounds. */
6958 else if (minTime > rx_maxSendWindow)
6959 minTime = rx_maxSendWindow;
6960 /* if (minTime != peer->maxWindow) {
6961 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u)\n",
6962 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6963 peer->timeout.sec, peer->timeout.usec, peer->smRtt));
6964 peer->maxWindow = minTime;
6965 elide... call->twind = minTime;
6969 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6970 * Discern this by calculating the timeout necessary for rx_Window
6972 if ((xferSize > rx_maxSendWindow) && (peer->timeout.sec >= 3)) {
6973 /* calculate estimate for transmission interval in milliseconds */
6974 minTime = rx_maxSendWindow * peer->smRtt;
6975 if (minTime < 1000) {
6976 dpf(("CONG peer %lx/%u: cut TO %d.%06d by 0.5 (rtt %u)\n",
6977 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6978 peer->timeout.usec, peer->smRtt));
6980 newTO.sec = 0; /* cut back on timeout by half a second */
6981 newTO.usec = 500000;
6982 clock_Sub(&peer->timeout, &newTO);
6987 } /* end of rxi_ComputeRate */
6988 #endif /* ADAPT_WINDOW */
6996 #define TRACE_OPTION_RX_DEBUG 16
7004 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7005 0, KEY_QUERY_VALUE, &parmKey);
7006 if (code != ERROR_SUCCESS)
7009 dummyLen = sizeof(TraceOption);
7010 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7011 (BYTE *) &TraceOption, &dummyLen);
7012 if (code == ERROR_SUCCESS) {
7013 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7015 RegCloseKey (parmKey);
7016 #endif /* AFS_NT40_ENV */
7021 rx_DebugOnOff(int on)
7025 rxdebug_active = on;
7031 rx_StatsOnOff(int on)
7033 rx_stats_active = on;
7037 /* Don't call this debugging routine directly; use dpf */
7039 rxi_DebugPrint(char *format, ...)
7048 va_start(ap, format);
7050 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7053 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7055 OutputDebugString(msg);
7061 va_start(ap, format);
7063 clock_GetTime(&now);
7064 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7065 (unsigned int)now.usec);
7066 vfprintf(rx_Log, format, ap);
7074 * This function is used to process the rx_stats structure that is local
7075 * to a process as well as an rx_stats structure received from a remote
7076 * process (via rxdebug). Therefore, it needs to do minimal version
7080 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7081 afs_int32 freePackets, char version)
7085 if (size != sizeof(struct rx_statistics)) {
7087 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7088 size, sizeof(struct rx_statistics));
7091 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7094 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7095 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7096 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7097 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7098 s->specialPktAllocFailures);
7100 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7101 s->receivePktAllocFailures, s->sendPktAllocFailures,
7102 s->specialPktAllocFailures);
7106 " greedy %u, " "bogusReads %u (last from host %x), "
7107 "noPackets %u, " "noBuffers %u, " "selects %u, "
7108 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7109 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7110 s->selects, s->sendSelects);
7112 fprintf(file, " packets read: ");
7113 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7114 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7116 fprintf(file, "\n");
7119 " other read counters: data %u, " "ack %u, " "dup %u "
7120 "spurious %u " "dally %u\n", s->dataPacketsRead,
7121 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7122 s->ignorePacketDally);
7124 fprintf(file, " packets sent: ");
7125 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7126 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7128 fprintf(file, "\n");
7131 " other send counters: ack %u, " "data %u (not resends), "
7132 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7133 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7134 s->dataPacketsPushed, s->ignoreAckedPacket);
7137 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7138 s->netSendFailures, (int)s->fatalErrors);
7140 if (s->nRttSamples) {
7141 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7142 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7144 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7145 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7149 " %d server connections, " "%d client connections, "
7150 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7151 s->nServerConns, s->nClientConns, s->nPeerStructs,
7152 s->nCallStructs, s->nFreeCallStructs);
7154 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7155 fprintf(file, " %d clock updates\n", clock_nUpdates);
7159 /* for backward compatibility */
7161 rx_PrintStats(FILE * file)
7163 MUTEX_ENTER(&rx_stats_mutex);
7164 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7165 sizeof(rx_stats), rx_nFreePackets,
7167 MUTEX_EXIT(&rx_stats_mutex);
7171 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7173 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
7174 ntohl(peer->host), (int)ntohs(peer->port), (int)peer->burstSize,
7175 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
7178 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
7179 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
7180 (int)peer->timeout.usec, peer->nSent, peer->reSends);
7183 " Packet size %d, " "max in packet skew %d, "
7184 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
7185 (int)peer->outPacketSkew);
7189 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7191 * This mutex protects the following static variables:
7195 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7196 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7198 #define LOCK_RX_DEBUG
7199 #define UNLOCK_RX_DEBUG
7200 #endif /* AFS_PTHREAD_ENV */
7202 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7204 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7205 u_char type, void *inputData, size_t inputLength,
7206 void *outputData, size_t outputLength)
7208 static afs_int32 counter = 100;
7209 time_t waitTime, waitCount;
7210 struct rx_header theader;
7213 struct timeval tv_now, tv_wake, tv_delta;
7214 struct sockaddr_in taddr, faddr;
7228 tp = &tbuffer[sizeof(struct rx_header)];
7229 taddr.sin_family = AF_INET;
7230 taddr.sin_port = remotePort;
7231 taddr.sin_addr.s_addr = remoteAddr;
7232 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7233 taddr.sin_len = sizeof(struct sockaddr_in);
7236 memset(&theader, 0, sizeof(theader));
7237 theader.epoch = htonl(999);
7239 theader.callNumber = htonl(counter);
7242 theader.type = type;
7243 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7244 theader.serviceId = 0;
7246 memcpy(tbuffer, &theader, sizeof(theader));
7247 memcpy(tp, inputData, inputLength);
7249 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7250 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7252 /* see if there's a packet available */
7253 gettimeofday(&tv_wake,0);
7254 tv_wake.tv_sec += waitTime;
7257 FD_SET(socket, &imask);
7258 tv_delta.tv_sec = tv_wake.tv_sec;
7259 tv_delta.tv_usec = tv_wake.tv_usec;
7260 gettimeofday(&tv_now, 0);
7262 if (tv_delta.tv_usec < tv_now.tv_usec) {
7264 tv_delta.tv_usec += 1000000;
7267 tv_delta.tv_usec -= tv_now.tv_usec;
7269 if (tv_delta.tv_sec < tv_now.tv_sec) {
7273 tv_delta.tv_sec -= tv_now.tv_sec;
7276 code = select(0, &imask, 0, 0, &tv_delta);
7277 #else /* AFS_NT40_ENV */
7278 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7279 #endif /* AFS_NT40_ENV */
7280 if (code == 1 && FD_ISSET(socket, &imask)) {
7281 /* now receive a packet */
7282 faddrLen = sizeof(struct sockaddr_in);
7284 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7285 (struct sockaddr *)&faddr, &faddrLen);
7288 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7289 if (counter == ntohl(theader.callNumber))
7297 /* see if we've timed out */
7305 code -= sizeof(struct rx_header);
7306 if (code > outputLength)
7307 code = outputLength;
7308 memcpy(outputData, tp, code);
7311 #endif /* RXDEBUG */
7314 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7315 afs_uint16 remotePort, struct rx_debugStats * stat,
7316 afs_uint32 * supportedValues)
7318 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7320 struct rx_debugIn in;
7322 *supportedValues = 0;
7323 in.type = htonl(RX_DEBUGI_GETSTATS);
7326 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7327 &in, sizeof(in), stat, sizeof(*stat));
7330 * If the call was successful, fixup the version and indicate
7331 * what contents of the stat structure are valid.
7332 * Also do net to host conversion of fields here.
7336 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7337 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7339 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7340 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7342 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7343 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7345 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7346 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7348 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7349 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7351 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7352 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7354 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7355 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7357 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7358 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7360 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7361 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7363 stat->nFreePackets = ntohl(stat->nFreePackets);
7364 stat->packetReclaims = ntohl(stat->packetReclaims);
7365 stat->callsExecuted = ntohl(stat->callsExecuted);
7366 stat->nWaiting = ntohl(stat->nWaiting);
7367 stat->idleThreads = ntohl(stat->idleThreads);
7368 stat->nWaited = ntohl(stat->nWaited);
7369 stat->nPackets = ntohl(stat->nPackets);
7378 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7379 afs_uint16 remotePort, struct rx_statistics * stat,
7380 afs_uint32 * supportedValues)
7382 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7384 struct rx_debugIn in;
7385 afs_int32 *lp = (afs_int32 *) stat;
7389 * supportedValues is currently unused, but added to allow future
7390 * versioning of this function.
7393 *supportedValues = 0;
7394 in.type = htonl(RX_DEBUGI_RXSTATS);
7396 memset(stat, 0, sizeof(*stat));
7398 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7399 &in, sizeof(in), stat, sizeof(*stat));
7404 * Do net to host conversion here
7407 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7418 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7419 afs_uint16 remotePort, size_t version_length,
7422 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7424 return MakeDebugCall(socket, remoteAddr, remotePort,
7425 RX_PACKET_TYPE_VERSION, a, 1, version,
7433 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7434 afs_uint16 remotePort, afs_int32 * nextConnection,
7435 int allConnections, afs_uint32 debugSupportedValues,
7436 struct rx_debugConn * conn,
7437 afs_uint32 * supportedValues)
7439 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7441 struct rx_debugIn in;
7445 * supportedValues is currently unused, but added to allow future
7446 * versioning of this function.
7449 *supportedValues = 0;
7450 if (allConnections) {
7451 in.type = htonl(RX_DEBUGI_GETALLCONN);
7453 in.type = htonl(RX_DEBUGI_GETCONN);
7455 in.index = htonl(*nextConnection);
7456 memset(conn, 0, sizeof(*conn));
7458 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7459 &in, sizeof(in), conn, sizeof(*conn));
7462 *nextConnection += 1;
7465 * Convert old connection format to new structure.
7468 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7469 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7470 #define MOVEvL(a) (conn->a = vL->a)
7472 /* any old or unrecognized version... */
7473 for (i = 0; i < RX_MAXCALLS; i++) {
7474 MOVEvL(callState[i]);
7475 MOVEvL(callMode[i]);
7476 MOVEvL(callFlags[i]);
7477 MOVEvL(callOther[i]);
7479 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7480 MOVEvL(secStats.type);
7481 MOVEvL(secStats.level);
7482 MOVEvL(secStats.flags);
7483 MOVEvL(secStats.expires);
7484 MOVEvL(secStats.packetsReceived);
7485 MOVEvL(secStats.packetsSent);
7486 MOVEvL(secStats.bytesReceived);
7487 MOVEvL(secStats.bytesSent);
7492 * Do net to host conversion here
7494 * I don't convert host or port since we are most likely
7495 * going to want these in NBO.
7497 conn->cid = ntohl(conn->cid);
7498 conn->serial = ntohl(conn->serial);
7499 for (i = 0; i < RX_MAXCALLS; i++) {
7500 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7502 conn->error = ntohl(conn->error);
7503 conn->secStats.flags = ntohl(conn->secStats.flags);
7504 conn->secStats.expires = ntohl(conn->secStats.expires);
7505 conn->secStats.packetsReceived =
7506 ntohl(conn->secStats.packetsReceived);
7507 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7508 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7509 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7510 conn->epoch = ntohl(conn->epoch);
7511 conn->natMTU = ntohl(conn->natMTU);
7520 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7521 afs_uint16 remotePort, afs_int32 * nextPeer,
7522 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7523 afs_uint32 * supportedValues)
7525 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7527 struct rx_debugIn in;
7530 * supportedValues is currently unused, but added to allow future
7531 * versioning of this function.
7534 *supportedValues = 0;
7535 in.type = htonl(RX_DEBUGI_GETPEER);
7536 in.index = htonl(*nextPeer);
7537 memset(peer, 0, sizeof(*peer));
7539 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7540 &in, sizeof(in), peer, sizeof(*peer));
7546 * Do net to host conversion here
7548 * I don't convert host or port since we are most likely
7549 * going to want these in NBO.
7551 peer->ifMTU = ntohs(peer->ifMTU);
7552 peer->idleWhen = ntohl(peer->idleWhen);
7553 peer->refCount = ntohs(peer->refCount);
7554 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7555 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7556 peer->rtt = ntohl(peer->rtt);
7557 peer->rtt_dev = ntohl(peer->rtt_dev);
7558 peer->timeout.sec = ntohl(peer->timeout.sec);
7559 peer->timeout.usec = ntohl(peer->timeout.usec);
7560 peer->nSent = ntohl(peer->nSent);
7561 peer->reSends = ntohl(peer->reSends);
7562 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7563 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7564 peer->rateFlag = ntohl(peer->rateFlag);
7565 peer->natMTU = ntohs(peer->natMTU);
7566 peer->maxMTU = ntohs(peer->maxMTU);
7567 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7568 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7569 peer->MTU = ntohs(peer->MTU);
7570 peer->cwind = ntohs(peer->cwind);
7571 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7572 peer->congestSeq = ntohs(peer->congestSeq);
7573 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7574 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7575 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7576 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7585 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7586 struct rx_debugPeer * peerStats)
7589 afs_int32 error = 1; /* default to "did not succeed" */
7590 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7592 MUTEX_ENTER(&rx_peerHashTable_lock);
7593 for(tp = rx_peerHashTable[hashValue];
7594 tp != NULL; tp = tp->next) {
7595 if (tp->host == peerHost)
7601 MUTEX_EXIT(&rx_peerHashTable_lock);
7605 MUTEX_ENTER(&tp->peer_lock);
7606 peerStats->host = tp->host;
7607 peerStats->port = tp->port;
7608 peerStats->ifMTU = tp->ifMTU;
7609 peerStats->idleWhen = tp->idleWhen;
7610 peerStats->refCount = tp->refCount;
7611 peerStats->burstSize = tp->burstSize;
7612 peerStats->burst = tp->burst;
7613 peerStats->burstWait.sec = tp->burstWait.sec;
7614 peerStats->burstWait.usec = tp->burstWait.usec;
7615 peerStats->rtt = tp->rtt;
7616 peerStats->rtt_dev = tp->rtt_dev;
7617 peerStats->timeout.sec = tp->timeout.sec;
7618 peerStats->timeout.usec = tp->timeout.usec;
7619 peerStats->nSent = tp->nSent;
7620 peerStats->reSends = tp->reSends;
7621 peerStats->inPacketSkew = tp->inPacketSkew;
7622 peerStats->outPacketSkew = tp->outPacketSkew;
7623 peerStats->rateFlag = tp->rateFlag;
7624 peerStats->natMTU = tp->natMTU;
7625 peerStats->maxMTU = tp->maxMTU;
7626 peerStats->maxDgramPackets = tp->maxDgramPackets;
7627 peerStats->ifDgramPackets = tp->ifDgramPackets;
7628 peerStats->MTU = tp->MTU;
7629 peerStats->cwind = tp->cwind;
7630 peerStats->nDgramPackets = tp->nDgramPackets;
7631 peerStats->congestSeq = tp->congestSeq;
7632 peerStats->bytesSent.high = tp->bytesSent.high;
7633 peerStats->bytesSent.low = tp->bytesSent.low;
7634 peerStats->bytesReceived.high = tp->bytesReceived.high;
7635 peerStats->bytesReceived.low = tp->bytesReceived.low;
7636 MUTEX_EXIT(&tp->peer_lock);
7638 MUTEX_ENTER(&rx_peerHashTable_lock);
7641 MUTEX_EXIT(&rx_peerHashTable_lock);
7649 struct rx_serverQueueEntry *np;
7652 struct rx_call *call;
7653 struct rx_serverQueueEntry *sq;
7657 if (rxinit_status == 1) {
7659 return; /* Already shutdown. */
7663 #ifndef AFS_PTHREAD_ENV
7664 FD_ZERO(&rx_selectMask);
7665 #endif /* AFS_PTHREAD_ENV */
7666 rxi_dataQuota = RX_MAX_QUOTA;
7667 #ifndef AFS_PTHREAD_ENV
7669 #endif /* AFS_PTHREAD_ENV */
7672 #ifndef AFS_PTHREAD_ENV
7673 #ifndef AFS_USE_GETTIMEOFDAY
7675 #endif /* AFS_USE_GETTIMEOFDAY */
7676 #endif /* AFS_PTHREAD_ENV */
7678 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7679 call = queue_First(&rx_freeCallQueue, rx_call);
7681 rxi_Free(call, sizeof(struct rx_call));
7684 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7685 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7691 struct rx_peer **peer_ptr, **peer_end;
7692 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7693 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7695 struct rx_peer *peer, *next;
7697 MUTEX_ENTER(&rx_peerHashTable_lock);
7698 for (peer = *peer_ptr; peer; peer = next) {
7699 rx_interface_stat_p rpc_stat, nrpc_stat;
7702 MUTEX_ENTER(&rx_rpc_stats);
7703 MUTEX_ENTER(&peer->peer_lock);
7705 (&peer->rpcStats, rpc_stat, nrpc_stat,
7706 rx_interface_stat)) {
7707 unsigned int num_funcs;
7710 queue_Remove(&rpc_stat->queue_header);
7711 queue_Remove(&rpc_stat->all_peers);
7712 num_funcs = rpc_stat->stats[0].func_total;
7714 sizeof(rx_interface_stat_t) +
7715 rpc_stat->stats[0].func_total *
7716 sizeof(rx_function_entry_v1_t);
7718 rxi_Free(rpc_stat, space);
7720 /* rx_rpc_stats must be held */
7721 rxi_rpc_peer_stat_cnt -= num_funcs;
7723 MUTEX_EXIT(&peer->peer_lock);
7724 MUTEX_EXIT(&rx_rpc_stats);
7728 if (rx_stats_active)
7729 rx_atomic_dec(&rx_stats.nPeerStructs);
7731 MUTEX_EXIT(&rx_peerHashTable_lock);
7734 for (i = 0; i < RX_MAX_SERVICES; i++) {
7736 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7738 for (i = 0; i < rx_hashTableSize; i++) {
7739 struct rx_connection *tc, *ntc;
7740 MUTEX_ENTER(&rx_connHashTable_lock);
7741 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7743 for (j = 0; j < RX_MAXCALLS; j++) {
7745 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7748 rxi_Free(tc, sizeof(*tc));
7750 MUTEX_EXIT(&rx_connHashTable_lock);
7753 MUTEX_ENTER(&freeSQEList_lock);
7755 while ((np = rx_FreeSQEList)) {
7756 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7757 MUTEX_DESTROY(&np->lock);
7758 rxi_Free(np, sizeof(*np));
7761 MUTEX_EXIT(&freeSQEList_lock);
7762 MUTEX_DESTROY(&freeSQEList_lock);
7763 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7764 MUTEX_DESTROY(&rx_connHashTable_lock);
7765 MUTEX_DESTROY(&rx_peerHashTable_lock);
7766 MUTEX_DESTROY(&rx_serverPool_lock);
7768 osi_Free(rx_connHashTable,
7769 rx_hashTableSize * sizeof(struct rx_connection *));
7770 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7772 UNPIN(rx_connHashTable,
7773 rx_hashTableSize * sizeof(struct rx_connection *));
7774 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7776 rxi_FreeAllPackets();
7778 MUTEX_ENTER(&rx_quota_mutex);
7779 rxi_dataQuota = RX_MAX_QUOTA;
7780 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7781 MUTEX_EXIT(&rx_quota_mutex);
7786 #ifdef RX_ENABLE_LOCKS
7788 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7790 if (!MUTEX_ISMINE(lockaddr))
7791 osi_Panic("Lock not held: %s", msg);
7793 #endif /* RX_ENABLE_LOCKS */
7798 * Routines to implement connection specific data.
7802 rx_KeyCreate(rx_destructor_t rtn)
7805 MUTEX_ENTER(&rxi_keyCreate_lock);
7806 key = rxi_keyCreate_counter++;
7807 rxi_keyCreate_destructor = (rx_destructor_t *)
7808 realloc((void *)rxi_keyCreate_destructor,
7809 (key + 1) * sizeof(rx_destructor_t));
7810 rxi_keyCreate_destructor[key] = rtn;
7811 MUTEX_EXIT(&rxi_keyCreate_lock);
7816 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7819 MUTEX_ENTER(&conn->conn_data_lock);
7820 if (!conn->specific) {
7821 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7822 for (i = 0; i < key; i++)
7823 conn->specific[i] = NULL;
7824 conn->nSpecific = key + 1;
7825 conn->specific[key] = ptr;
7826 } else if (key >= conn->nSpecific) {
7827 conn->specific = (void **)
7828 realloc(conn->specific, (key + 1) * sizeof(void *));
7829 for (i = conn->nSpecific; i < key; i++)
7830 conn->specific[i] = NULL;
7831 conn->nSpecific = key + 1;
7832 conn->specific[key] = ptr;
7834 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7835 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7836 conn->specific[key] = ptr;
7838 MUTEX_EXIT(&conn->conn_data_lock);
7842 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7845 MUTEX_ENTER(&svc->svc_data_lock);
7846 if (!svc->specific) {
7847 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
7848 for (i = 0; i < key; i++)
7849 svc->specific[i] = NULL;
7850 svc->nSpecific = key + 1;
7851 svc->specific[key] = ptr;
7852 } else if (key >= svc->nSpecific) {
7853 svc->specific = (void **)
7854 realloc(svc->specific, (key + 1) * sizeof(void *));
7855 for (i = svc->nSpecific; i < key; i++)
7856 svc->specific[i] = NULL;
7857 svc->nSpecific = key + 1;
7858 svc->specific[key] = ptr;
7860 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7861 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7862 svc->specific[key] = ptr;
7864 MUTEX_EXIT(&svc->svc_data_lock);
7868 rx_GetSpecific(struct rx_connection *conn, int key)
7871 MUTEX_ENTER(&conn->conn_data_lock);
7872 if (key >= conn->nSpecific)
7875 ptr = conn->specific[key];
7876 MUTEX_EXIT(&conn->conn_data_lock);
7881 rx_GetServiceSpecific(struct rx_service *svc, int key)
7884 MUTEX_ENTER(&svc->svc_data_lock);
7885 if (key >= svc->nSpecific)
7888 ptr = svc->specific[key];
7889 MUTEX_EXIT(&svc->svc_data_lock);
7894 #endif /* !KERNEL */
7897 * processStats is a queue used to store the statistics for the local
7898 * process. Its contents are similar to the contents of the rpcStats
7899 * queue on a rx_peer structure, but the actual data stored within
7900 * this queue contains totals across the lifetime of the process (assuming
7901 * the stats have not been reset) - unlike the per peer structures
7902 * which can come and go based upon the peer lifetime.
7905 static struct rx_queue processStats = { &processStats, &processStats };
7908 * peerStats is a queue used to store the statistics for all peer structs.
7909 * Its contents are the union of all the peer rpcStats queues.
7912 static struct rx_queue peerStats = { &peerStats, &peerStats };
7915 * rxi_monitor_processStats is used to turn process wide stat collection
7919 static int rxi_monitor_processStats = 0;
7922 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7925 static int rxi_monitor_peerStats = 0;
7928 * rxi_AddRpcStat - given all of the information for a particular rpc
7929 * call, create (if needed) and update the stat totals for the rpc.
7933 * IN stats - the queue of stats that will be updated with the new value
7935 * IN rxInterface - a unique number that identifies the rpc interface
7937 * IN currentFunc - the index of the function being invoked
7939 * IN totalFunc - the total number of functions in this interface
7941 * IN queueTime - the amount of time this function waited for a thread
7943 * IN execTime - the amount of time this function invocation took to execute
7945 * IN bytesSent - the number bytes sent by this invocation
7947 * IN bytesRcvd - the number bytes received by this invocation
7949 * IN isServer - if true, this invocation was made to a server
7951 * IN remoteHost - the ip address of the remote host
7953 * IN remotePort - the port of the remote host
7955 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7957 * INOUT counter - if a new stats structure is allocated, the counter will
7958 * be updated with the new number of allocated stat structures
7966 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
7967 afs_uint32 currentFunc, afs_uint32 totalFunc,
7968 struct clock *queueTime, struct clock *execTime,
7969 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
7970 afs_uint32 remoteHost, afs_uint32 remotePort,
7971 int addToPeerList, unsigned int *counter)
7974 rx_interface_stat_p rpc_stat, nrpc_stat;
7977 * See if there's already a structure for this interface
7980 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7981 if ((rpc_stat->stats[0].interfaceId == rxInterface)
7982 && (rpc_stat->stats[0].remote_is_server == isServer))
7987 * Didn't find a match so allocate a new structure and add it to the
7991 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
7992 || (rpc_stat->stats[0].interfaceId != rxInterface)
7993 || (rpc_stat->stats[0].remote_is_server != isServer)) {
7998 sizeof(rx_interface_stat_t) +
7999 totalFunc * sizeof(rx_function_entry_v1_t);
8001 rpc_stat = rxi_Alloc(space);
8002 if (rpc_stat == NULL) {
8006 *counter += totalFunc;
8007 for (i = 0; i < totalFunc; i++) {
8008 rpc_stat->stats[i].remote_peer = remoteHost;
8009 rpc_stat->stats[i].remote_port = remotePort;
8010 rpc_stat->stats[i].remote_is_server = isServer;
8011 rpc_stat->stats[i].interfaceId = rxInterface;
8012 rpc_stat->stats[i].func_total = totalFunc;
8013 rpc_stat->stats[i].func_index = i;
8014 hzero(rpc_stat->stats[i].invocations);
8015 hzero(rpc_stat->stats[i].bytes_sent);
8016 hzero(rpc_stat->stats[i].bytes_rcvd);
8017 rpc_stat->stats[i].queue_time_sum.sec = 0;
8018 rpc_stat->stats[i].queue_time_sum.usec = 0;
8019 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8020 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8021 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8022 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8023 rpc_stat->stats[i].queue_time_max.sec = 0;
8024 rpc_stat->stats[i].queue_time_max.usec = 0;
8025 rpc_stat->stats[i].execution_time_sum.sec = 0;
8026 rpc_stat->stats[i].execution_time_sum.usec = 0;
8027 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8028 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8029 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8030 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8031 rpc_stat->stats[i].execution_time_max.sec = 0;
8032 rpc_stat->stats[i].execution_time_max.usec = 0;
8034 queue_Prepend(stats, rpc_stat);
8035 if (addToPeerList) {
8036 queue_Prepend(&peerStats, &rpc_stat->all_peers);
8041 * Increment the stats for this function
8044 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
8045 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
8046 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
8047 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8048 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8049 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8050 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8052 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8053 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8055 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8056 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8058 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8059 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8061 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8062 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8070 * rx_IncrementTimeAndCount - increment the times and count for a particular
8075 * IN peer - the peer who invoked the rpc
8077 * IN rxInterface - a unique number that identifies the rpc interface
8079 * IN currentFunc - the index of the function being invoked
8081 * IN totalFunc - the total number of functions in this interface
8083 * IN queueTime - the amount of time this function waited for a thread
8085 * IN execTime - the amount of time this function invocation took to execute
8087 * IN bytesSent - the number bytes sent by this invocation
8089 * IN bytesRcvd - the number bytes received by this invocation
8091 * IN isServer - if true, this invocation was made to a server
8099 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8100 afs_uint32 currentFunc, afs_uint32 totalFunc,
8101 struct clock *queueTime, struct clock *execTime,
8102 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8106 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8109 MUTEX_ENTER(&rx_rpc_stats);
8111 if (rxi_monitor_peerStats) {
8112 MUTEX_ENTER(&peer->peer_lock);
8113 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8114 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8115 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8116 MUTEX_EXIT(&peer->peer_lock);
8119 if (rxi_monitor_processStats) {
8120 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8121 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8122 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8125 MUTEX_EXIT(&rx_rpc_stats);
8130 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8134 * IN callerVersion - the rpc stat version of the caller.
8136 * IN count - the number of entries to marshall.
8138 * IN stats - pointer to stats to be marshalled.
8140 * OUT ptr - Where to store the marshalled data.
8147 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8148 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8154 * We only support the first version
8156 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8157 *(ptr++) = stats->remote_peer;
8158 *(ptr++) = stats->remote_port;
8159 *(ptr++) = stats->remote_is_server;
8160 *(ptr++) = stats->interfaceId;
8161 *(ptr++) = stats->func_total;
8162 *(ptr++) = stats->func_index;
8163 *(ptr++) = hgethi(stats->invocations);
8164 *(ptr++) = hgetlo(stats->invocations);
8165 *(ptr++) = hgethi(stats->bytes_sent);
8166 *(ptr++) = hgetlo(stats->bytes_sent);
8167 *(ptr++) = hgethi(stats->bytes_rcvd);
8168 *(ptr++) = hgetlo(stats->bytes_rcvd);
8169 *(ptr++) = stats->queue_time_sum.sec;
8170 *(ptr++) = stats->queue_time_sum.usec;
8171 *(ptr++) = stats->queue_time_sum_sqr.sec;
8172 *(ptr++) = stats->queue_time_sum_sqr.usec;
8173 *(ptr++) = stats->queue_time_min.sec;
8174 *(ptr++) = stats->queue_time_min.usec;
8175 *(ptr++) = stats->queue_time_max.sec;
8176 *(ptr++) = stats->queue_time_max.usec;
8177 *(ptr++) = stats->execution_time_sum.sec;
8178 *(ptr++) = stats->execution_time_sum.usec;
8179 *(ptr++) = stats->execution_time_sum_sqr.sec;
8180 *(ptr++) = stats->execution_time_sum_sqr.usec;
8181 *(ptr++) = stats->execution_time_min.sec;
8182 *(ptr++) = stats->execution_time_min.usec;
8183 *(ptr++) = stats->execution_time_max.sec;
8184 *(ptr++) = stats->execution_time_max.usec;
8190 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8195 * IN callerVersion - the rpc stat version of the caller
8197 * OUT myVersion - the rpc stat version of this function
8199 * OUT clock_sec - local time seconds
8201 * OUT clock_usec - local time microseconds
8203 * OUT allocSize - the number of bytes allocated to contain stats
8205 * OUT statCount - the number stats retrieved from this process.
8207 * OUT stats - the actual stats retrieved from this process.
8211 * Returns void. If successful, stats will != NULL.
8215 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8216 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8217 size_t * allocSize, afs_uint32 * statCount,
8218 afs_uint32 ** stats)
8228 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8231 * Check to see if stats are enabled
8234 MUTEX_ENTER(&rx_rpc_stats);
8235 if (!rxi_monitor_processStats) {
8236 MUTEX_EXIT(&rx_rpc_stats);
8240 clock_GetTime(&now);
8241 *clock_sec = now.sec;
8242 *clock_usec = now.usec;
8245 * Allocate the space based upon the caller version
8247 * If the client is at an older version than we are,
8248 * we return the statistic data in the older data format, but
8249 * we still return our version number so the client knows we
8250 * are maintaining more data than it can retrieve.
8253 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8254 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8255 *statCount = rxi_rpc_process_stat_cnt;
8258 * This can't happen yet, but in the future version changes
8259 * can be handled by adding additional code here
8263 if (space > (size_t) 0) {
8265 ptr = *stats = rxi_Alloc(space);
8268 rx_interface_stat_p rpc_stat, nrpc_stat;
8272 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8274 * Copy the data based upon the caller version
8276 rx_MarshallProcessRPCStats(callerVersion,
8277 rpc_stat->stats[0].func_total,
8278 rpc_stat->stats, &ptr);
8284 MUTEX_EXIT(&rx_rpc_stats);
8289 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8293 * IN callerVersion - the rpc stat version of the caller
8295 * OUT myVersion - the rpc stat version of this function
8297 * OUT clock_sec - local time seconds
8299 * OUT clock_usec - local time microseconds
8301 * OUT allocSize - the number of bytes allocated to contain stats
8303 * OUT statCount - the number of stats retrieved from the individual
8306 * OUT stats - the actual stats retrieved from the individual peer structures.
8310 * Returns void. If successful, stats will != NULL.
8314 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8315 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8316 size_t * allocSize, afs_uint32 * statCount,
8317 afs_uint32 ** stats)
8327 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8330 * Check to see if stats are enabled
8333 MUTEX_ENTER(&rx_rpc_stats);
8334 if (!rxi_monitor_peerStats) {
8335 MUTEX_EXIT(&rx_rpc_stats);
8339 clock_GetTime(&now);
8340 *clock_sec = now.sec;
8341 *clock_usec = now.usec;
8344 * Allocate the space based upon the caller version
8346 * If the client is at an older version than we are,
8347 * we return the statistic data in the older data format, but
8348 * we still return our version number so the client knows we
8349 * are maintaining more data than it can retrieve.
8352 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8353 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8354 *statCount = rxi_rpc_peer_stat_cnt;
8357 * This can't happen yet, but in the future version changes
8358 * can be handled by adding additional code here
8362 if (space > (size_t) 0) {
8364 ptr = *stats = rxi_Alloc(space);
8367 rx_interface_stat_p rpc_stat, nrpc_stat;
8371 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8373 * We have to fix the offset of rpc_stat since we are
8374 * keeping this structure on two rx_queues. The rx_queue
8375 * package assumes that the rx_queue member is the first
8376 * member of the structure. That is, rx_queue assumes that
8377 * any one item is only on one queue at a time. We are
8378 * breaking that assumption and so we have to do a little
8379 * math to fix our pointers.
8382 fix_offset = (char *)rpc_stat;
8383 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8384 rpc_stat = (rx_interface_stat_p) fix_offset;
8387 * Copy the data based upon the caller version
8389 rx_MarshallProcessRPCStats(callerVersion,
8390 rpc_stat->stats[0].func_total,
8391 rpc_stat->stats, &ptr);
8397 MUTEX_EXIT(&rx_rpc_stats);
8402 * rx_FreeRPCStats - free memory allocated by
8403 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8407 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8408 * rx_RetrievePeerRPCStats
8410 * IN allocSize - the number of bytes in stats.
8418 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8420 rxi_Free(stats, allocSize);
8424 * rx_queryProcessRPCStats - see if process rpc stat collection is
8425 * currently enabled.
8431 * Returns 0 if stats are not enabled != 0 otherwise
8435 rx_queryProcessRPCStats(void)
8438 MUTEX_ENTER(&rx_rpc_stats);
8439 rc = rxi_monitor_processStats;
8440 MUTEX_EXIT(&rx_rpc_stats);
8445 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8451 * Returns 0 if stats are not enabled != 0 otherwise
8455 rx_queryPeerRPCStats(void)
8458 MUTEX_ENTER(&rx_rpc_stats);
8459 rc = rxi_monitor_peerStats;
8460 MUTEX_EXIT(&rx_rpc_stats);
8465 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8475 rx_enableProcessRPCStats(void)
8477 MUTEX_ENTER(&rx_rpc_stats);
8478 rx_enable_stats = 1;
8479 rxi_monitor_processStats = 1;
8480 MUTEX_EXIT(&rx_rpc_stats);
8484 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8494 rx_enablePeerRPCStats(void)
8496 MUTEX_ENTER(&rx_rpc_stats);
8497 rx_enable_stats = 1;
8498 rxi_monitor_peerStats = 1;
8499 MUTEX_EXIT(&rx_rpc_stats);
8503 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8513 rx_disableProcessRPCStats(void)
8515 rx_interface_stat_p rpc_stat, nrpc_stat;
8518 MUTEX_ENTER(&rx_rpc_stats);
8521 * Turn off process statistics and if peer stats is also off, turn
8525 rxi_monitor_processStats = 0;
8526 if (rxi_monitor_peerStats == 0) {
8527 rx_enable_stats = 0;
8530 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8531 unsigned int num_funcs = 0;
8534 queue_Remove(rpc_stat);
8535 num_funcs = rpc_stat->stats[0].func_total;
8537 sizeof(rx_interface_stat_t) +
8538 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8540 rxi_Free(rpc_stat, space);
8541 rxi_rpc_process_stat_cnt -= num_funcs;
8543 MUTEX_EXIT(&rx_rpc_stats);
8547 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8557 rx_disablePeerRPCStats(void)
8559 struct rx_peer **peer_ptr, **peer_end;
8563 * Turn off peer statistics and if process stats is also off, turn
8567 rxi_monitor_peerStats = 0;
8568 if (rxi_monitor_processStats == 0) {
8569 rx_enable_stats = 0;
8572 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8573 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8575 struct rx_peer *peer, *next, *prev;
8577 MUTEX_ENTER(&rx_peerHashTable_lock);
8578 MUTEX_ENTER(&rx_rpc_stats);
8579 for (prev = peer = *peer_ptr; peer; peer = next) {
8581 code = MUTEX_TRYENTER(&peer->peer_lock);
8583 rx_interface_stat_p rpc_stat, nrpc_stat;
8586 if (prev == *peer_ptr) {
8597 MUTEX_EXIT(&rx_peerHashTable_lock);
8600 (&peer->rpcStats, rpc_stat, nrpc_stat,
8601 rx_interface_stat)) {
8602 unsigned int num_funcs = 0;
8605 queue_Remove(&rpc_stat->queue_header);
8606 queue_Remove(&rpc_stat->all_peers);
8607 num_funcs = rpc_stat->stats[0].func_total;
8609 sizeof(rx_interface_stat_t) +
8610 rpc_stat->stats[0].func_total *
8611 sizeof(rx_function_entry_v1_t);
8613 rxi_Free(rpc_stat, space);
8614 rxi_rpc_peer_stat_cnt -= num_funcs;
8616 MUTEX_EXIT(&peer->peer_lock);
8618 MUTEX_ENTER(&rx_peerHashTable_lock);
8628 MUTEX_EXIT(&rx_rpc_stats);
8629 MUTEX_EXIT(&rx_peerHashTable_lock);
8634 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8639 * IN clearFlag - flag indicating which stats to clear
8647 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8649 rx_interface_stat_p rpc_stat, nrpc_stat;
8651 MUTEX_ENTER(&rx_rpc_stats);
8653 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8654 unsigned int num_funcs = 0, i;
8655 num_funcs = rpc_stat->stats[0].func_total;
8656 for (i = 0; i < num_funcs; i++) {
8657 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8658 hzero(rpc_stat->stats[i].invocations);
8660 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8661 hzero(rpc_stat->stats[i].bytes_sent);
8663 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8664 hzero(rpc_stat->stats[i].bytes_rcvd);
8666 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8667 rpc_stat->stats[i].queue_time_sum.sec = 0;
8668 rpc_stat->stats[i].queue_time_sum.usec = 0;
8670 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8671 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8672 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8674 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8675 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8676 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8678 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8679 rpc_stat->stats[i].queue_time_max.sec = 0;
8680 rpc_stat->stats[i].queue_time_max.usec = 0;
8682 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8683 rpc_stat->stats[i].execution_time_sum.sec = 0;
8684 rpc_stat->stats[i].execution_time_sum.usec = 0;
8686 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8687 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8688 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8690 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8691 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8692 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8694 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8695 rpc_stat->stats[i].execution_time_max.sec = 0;
8696 rpc_stat->stats[i].execution_time_max.usec = 0;
8701 MUTEX_EXIT(&rx_rpc_stats);
8705 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8710 * IN clearFlag - flag indicating which stats to clear
8718 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8720 rx_interface_stat_p rpc_stat, nrpc_stat;
8722 MUTEX_ENTER(&rx_rpc_stats);
8724 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8725 unsigned int num_funcs = 0, i;
8728 * We have to fix the offset of rpc_stat since we are
8729 * keeping this structure on two rx_queues. The rx_queue
8730 * package assumes that the rx_queue member is the first
8731 * member of the structure. That is, rx_queue assumes that
8732 * any one item is only on one queue at a time. We are
8733 * breaking that assumption and so we have to do a little
8734 * math to fix our pointers.
8737 fix_offset = (char *)rpc_stat;
8738 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8739 rpc_stat = (rx_interface_stat_p) fix_offset;
8741 num_funcs = rpc_stat->stats[0].func_total;
8742 for (i = 0; i < num_funcs; i++) {
8743 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8744 hzero(rpc_stat->stats[i].invocations);
8746 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8747 hzero(rpc_stat->stats[i].bytes_sent);
8749 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8750 hzero(rpc_stat->stats[i].bytes_rcvd);
8752 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8753 rpc_stat->stats[i].queue_time_sum.sec = 0;
8754 rpc_stat->stats[i].queue_time_sum.usec = 0;
8756 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8757 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8758 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8760 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8761 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8762 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8764 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8765 rpc_stat->stats[i].queue_time_max.sec = 0;
8766 rpc_stat->stats[i].queue_time_max.usec = 0;
8768 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8769 rpc_stat->stats[i].execution_time_sum.sec = 0;
8770 rpc_stat->stats[i].execution_time_sum.usec = 0;
8772 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8773 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8774 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8776 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8777 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8778 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8780 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8781 rpc_stat->stats[i].execution_time_max.sec = 0;
8782 rpc_stat->stats[i].execution_time_max.usec = 0;
8787 MUTEX_EXIT(&rx_rpc_stats);
8791 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8792 * is authorized to enable/disable/clear RX statistics.
8794 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8797 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8799 rxi_rxstat_userok = proc;
8803 rx_RxStatUserOk(struct rx_call *call)
8805 if (!rxi_rxstat_userok)
8807 return rxi_rxstat_userok(call);
8812 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8813 * function in the MSVC runtime DLL (msvcrt.dll).
8815 * Note: the system serializes calls to this function.
8818 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8819 DWORD reason, /* reason function is being called */
8820 LPVOID reserved) /* reserved for future use */
8823 case DLL_PROCESS_ATTACH:
8824 /* library is being attached to a process */
8828 case DLL_PROCESS_DETACH:
8835 #endif /* AFS_NT40_ENV */
8838 int rx_DumpCalls(FILE *outputFile, char *cookie)
8840 #ifdef RXDEBUG_PACKET
8841 #ifdef KDUMP_RX_LOCK
8842 struct rx_call_rx_lock *c;
8849 #define RXDPRINTF sprintf
8850 #define RXDPRINTOUT output
8852 #define RXDPRINTF fprintf
8853 #define RXDPRINTOUT outputFile
8856 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8858 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8861 for (c = rx_allCallsp; c; c = c->allNextp) {
8862 u_short rqc, tqc, iovqc;
8863 struct rx_packet *p, *np;
8865 MUTEX_ENTER(&c->lock);
8866 queue_Count(&c->rq, p, np, rx_packet, rqc);
8867 queue_Count(&c->tq, p, np, rx_packet, tqc);
8868 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8870 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, "
8871 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8872 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8873 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8874 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8875 #ifdef RX_ENABLE_LOCKS
8878 #ifdef RX_REFCOUNT_CHECK
8879 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
8880 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
8883 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,
8884 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
8885 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
8886 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
8887 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
8888 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
8889 #ifdef RX_ENABLE_LOCKS
8890 , (afs_uint32)c->refCount
8892 #ifdef RX_REFCOUNT_CHECK
8893 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
8896 MUTEX_EXIT(&c->lock);
8899 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8902 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
8904 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8906 #endif /* RXDEBUG_PACKET */