2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
14 #include "afs/param.h"
16 #include <afs/param.h>
21 #include "afs/sysincludes.h"
22 #include "afsincludes.h"
28 #include <net/net_globals.h>
29 #endif /* AFS_OSF_ENV */
30 #ifdef AFS_LINUX20_ENV
33 #include "netinet/in.h"
35 #include "inet/common.h"
37 #include "inet/ip_ire.h"
39 #include "afs/afs_args.h"
40 #include "afs/afs_osi.h"
41 #ifdef RX_KERNEL_TRACE
42 #include "rx_kcommon.h"
44 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
48 #undef RXDEBUG /* turn off debugging */
50 #if defined(AFS_SGI_ENV)
51 #include "sys/debug.h"
60 #endif /* AFS_OSF_ENV */
62 #include "afs/sysincludes.h"
63 #include "afsincludes.h"
66 #include "rx_kmutex.h"
67 #include "rx_kernel.h"
71 #include "rx_globals.h"
73 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
74 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
75 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
77 extern afs_int32 afs_termState;
79 #include "sys/lockl.h"
80 #include "sys/lock_def.h"
81 #endif /* AFS_AIX41_ENV */
82 # include "rxgen_consts.h"
84 # include <sys/types.h>
94 # include <afs/afsutil.h>
95 # include <WINNT\afsreg.h>
97 # include <sys/socket.h>
98 # include <sys/file.h>
100 # include <sys/stat.h>
101 # include <netinet/in.h>
102 # include <sys/time.h>
105 # include "rx_user.h"
106 # include "rx_clock.h"
107 # include "rx_queue.h"
108 # include "rx_globals.h"
109 # include "rx_trace.h"
110 # include <afs/rxgen_consts.h>
114 #ifdef AFS_PTHREAD_ENV
116 int (*registerProgram) (pid_t, char *) = 0;
117 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
120 int (*registerProgram) (PROCESS, char *) = 0;
121 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
125 /* Local static routines */
126 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
127 #ifdef RX_ENABLE_LOCKS
128 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
131 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
133 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
134 afs_int32 rxi_start_in_error;
136 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
139 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
140 * currently allocated within rx. This number is used to allocate the
141 * memory required to return the statistics when queried.
144 static unsigned int rxi_rpc_peer_stat_cnt;
147 * rxi_rpc_process_stat_cnt counts the total number of local process stat
148 * structures currently allocated within rx. The number is used to allocate
149 * the memory required to return the statistics when queried.
152 static unsigned int rxi_rpc_process_stat_cnt;
154 #if !defined(offsetof)
155 #include <stddef.h> /* for definition of offsetof() */
158 #ifdef AFS_PTHREAD_ENV
162 * Use procedural initialization of mutexes/condition variables
166 extern afs_kmutex_t rx_stats_mutex;
167 extern afs_kmutex_t rx_waiting_mutex;
168 extern afs_kmutex_t rx_quota_mutex;
169 extern afs_kmutex_t rx_pthread_mutex;
170 extern afs_kmutex_t rx_packets_mutex;
171 extern afs_kmutex_t des_init_mutex;
172 extern afs_kmutex_t des_random_mutex;
173 extern afs_kmutex_t rx_clock_mutex;
174 extern afs_kmutex_t rxi_connCacheMutex;
175 extern afs_kmutex_t rx_event_mutex;
176 extern afs_kmutex_t osi_malloc_mutex;
177 extern afs_kmutex_t event_handler_mutex;
178 extern afs_kmutex_t listener_mutex;
179 extern afs_kmutex_t rx_if_init_mutex;
180 extern afs_kmutex_t rx_if_mutex;
181 extern afs_kmutex_t rxkad_client_uid_mutex;
182 extern afs_kmutex_t rxkad_random_mutex;
184 extern afs_kcondvar_t rx_event_handler_cond;
185 extern afs_kcondvar_t rx_listener_cond;
187 static afs_kmutex_t epoch_mutex;
188 static afs_kmutex_t rx_init_mutex;
189 static afs_kmutex_t rx_debug_mutex;
190 static afs_kmutex_t rx_rpc_stats;
193 rxi_InitPthread(void)
195 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
196 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
197 MUTEX_INIT(&rx_waiting_mutex, "waiting", MUTEX_DEFAULT, 0);
198 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
199 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
200 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
201 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
202 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
203 MUTEX_INIT(&rx_event_mutex, "event", MUTEX_DEFAULT, 0);
204 MUTEX_INIT(&des_init_mutex, "des", MUTEX_DEFAULT, 0);
205 MUTEX_INIT(&des_random_mutex, "random", MUTEX_DEFAULT, 0);
206 MUTEX_INIT(&osi_malloc_mutex, "malloc", MUTEX_DEFAULT, 0);
207 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
208 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
209 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
210 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
211 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
212 MUTEX_INIT(&rxkad_client_uid_mutex, "uid", MUTEX_DEFAULT, 0);
213 MUTEX_INIT(&rxkad_random_mutex, "rxkad random", MUTEX_DEFAULT, 0);
214 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
216 assert(pthread_cond_init
217 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
218 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
220 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
221 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
223 rxkad_global_stats_init();
225 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
226 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
227 #ifdef RX_ENABLE_LOCKS
230 #endif /* RX_LOCKS_DB */
231 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
232 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
234 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
236 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
238 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
240 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
241 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
242 #endif /* RX_ENABLE_LOCKS */
245 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
246 #define INIT_PTHREAD_LOCKS \
247 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
249 * The rx_stats_mutex mutex protects the following global variables:
250 * rxi_lowConnRefCount
251 * rxi_lowPeerRefCount
260 * The rx_quota_mutex mutex protects the following global variables:
268 * The rx_freePktQ_lock protects the following global variables:
273 * The rx_packets_mutex mutex protects the following global variables:
281 * The rx_pthread_mutex mutex protects the following global variables:
285 #define INIT_PTHREAD_LOCKS
289 /* Variables for handling the minProcs implementation. availProcs gives the
290 * number of threads available in the pool at this moment (not counting dudes
291 * executing right now). totalMin gives the total number of procs required
292 * for handling all minProcs requests. minDeficit is a dynamic variable
293 * tracking the # of procs required to satisfy all of the remaining minProcs
295 * For fine grain locking to work, the quota check and the reservation of
296 * a server thread has to come while rxi_availProcs and rxi_minDeficit
297 * are locked. To this end, the code has been modified under #ifdef
298 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
299 * same time. A new function, ReturnToServerPool() returns the allocation.
301 * A call can be on several queue's (but only one at a time). When
302 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
303 * that no one else is touching the queue. To this end, we store the address
304 * of the queue lock in the call structure (under the call lock) when we
305 * put the call on a queue, and we clear the call_queue_lock when the
306 * call is removed from a queue (once the call lock has been obtained).
307 * This allows rxi_ResetCall to safely synchronize with others wishing
308 * to manipulate the queue.
311 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
312 static afs_kmutex_t rx_rpc_stats;
313 void rxi_StartUnlocked(struct rxevent *event, void *call,
314 void *arg1, int istack);
317 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
318 ** pretty good that the next packet coming in is from the same connection
319 ** as the last packet, since we're send multiple packets in a transmit window.
321 struct rx_connection *rxLastConn = 0;
323 #ifdef RX_ENABLE_LOCKS
324 /* The locking hierarchy for rx fine grain locking is composed of these
327 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
328 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
329 * call->lock - locks call data fields.
330 * These are independent of each other:
331 * rx_freeCallQueue_lock
336 * serverQueueEntry->lock
338 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
339 * peer->lock - locks peer data fields.
340 * conn_data_lock - that more than one thread is not updating a conn data
341 * field at the same time.
349 * Do we need a lock to protect the peer field in the conn structure?
350 * conn->peer was previously a constant for all intents and so has no
351 * lock protecting this field. The multihomed client delta introduced
352 * a RX code change : change the peer field in the connection structure
353 * to that remote inetrface from which the last packet for this
354 * connection was sent out. This may become an issue if further changes
357 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
358 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
360 /* rxdb_fileID is used to identify the lock location, along with line#. */
361 static int rxdb_fileID = RXDB_FILE_RX;
362 #endif /* RX_LOCKS_DB */
363 #else /* RX_ENABLE_LOCKS */
364 #define SET_CALL_QUEUE_LOCK(C, L)
365 #define CLEAR_CALL_QUEUE_LOCK(C)
366 #endif /* RX_ENABLE_LOCKS */
367 struct rx_serverQueueEntry *rx_waitForPacket = 0;
368 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
370 /* ------------Exported Interfaces------------- */
372 /* This function allows rxkad to set the epoch to a suitably random number
373 * which rx_NewConnection will use in the future. The principle purpose is to
374 * get rxnull connections to use the same epoch as the rxkad connections do, at
375 * least once the first rxkad connection is established. This is important now
376 * that the host/port addresses aren't used in FindConnection: the uniqueness
377 * of epoch/cid matters and the start time won't do. */
379 #ifdef AFS_PTHREAD_ENV
381 * This mutex protects the following global variables:
385 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
386 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
390 #endif /* AFS_PTHREAD_ENV */
393 rx_SetEpoch(afs_uint32 epoch)
400 /* Initialize rx. A port number may be mentioned, in which case this
401 * becomes the default port number for any service installed later.
402 * If 0 is provided for the port number, a random port will be chosen
403 * by the kernel. Whether this will ever overlap anything in
404 * /etc/services is anybody's guess... Returns 0 on success, -1 on
409 int rxinit_status = 1;
410 #ifdef AFS_PTHREAD_ENV
412 * This mutex protects the following global variables:
416 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
417 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
420 #define UNLOCK_RX_INIT
424 rx_InitHost(u_int host, u_int port)
431 char *htable, *ptable;
438 if (rxinit_status == 0) {
439 tmp_status = rxinit_status;
441 return tmp_status; /* Already started; return previous error code. */
447 if (afs_winsockInit() < 0)
453 * Initialize anything necessary to provide a non-premptive threading
456 rxi_InitializeThreadSupport();
459 /* Allocate and initialize a socket for client and perhaps server
462 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
463 if (rx_socket == OSI_NULLSOCKET) {
467 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
470 #endif /* RX_LOCKS_DB */
471 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
472 MUTEX_INIT(&rx_waiting_mutex, "rx_waiting_mutex", MUTEX_DEFAULT, 0);
473 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
474 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
475 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
476 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
477 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
478 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
479 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
481 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
483 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
485 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
487 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
488 #if defined(AFS_HPUX110_ENV)
490 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
491 #endif /* AFS_HPUX110_ENV */
492 #endif /* RX_ENABLE_LOCKS && KERNEL */
495 rx_connDeadTime = 12;
496 rx_tranquil = 0; /* reset flag */
497 memset(&rx_stats, 0, sizeof(struct rx_statistics));
499 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
500 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
501 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
502 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
503 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
504 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
506 /* Malloc up a bunch of packets & buffers */
508 queue_Init(&rx_freePacketQueue);
509 rxi_NeedMorePackets = FALSE;
510 #ifdef RX_ENABLE_TSFPQ
511 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
512 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
513 #else /* RX_ENABLE_TSFPQ */
514 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
515 rxi_MorePackets(rx_nPackets);
516 #endif /* RX_ENABLE_TSFPQ */
523 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
524 tv.tv_sec = clock_now.sec;
525 tv.tv_usec = clock_now.usec;
526 srand((unsigned int)tv.tv_usec);
533 #if defined(KERNEL) && !defined(UKERNEL)
534 /* Really, this should never happen in a real kernel */
537 struct sockaddr_in addr;
539 int addrlen = sizeof(addr);
541 socklen_t addrlen = sizeof(addr);
543 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
547 rx_port = addr.sin_port;
550 rx_stats.minRtt.sec = 9999999;
552 rx_SetEpoch(tv.tv_sec | 0x80000000);
554 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
555 * will provide a randomer value. */
557 MUTEX_ENTER(&rx_quota_mutex);
558 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
559 MUTEX_EXIT(&rx_quota_mutex);
560 /* *Slightly* random start time for the cid. This is just to help
561 * out with the hashing function at the peer */
562 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
563 rx_connHashTable = (struct rx_connection **)htable;
564 rx_peerHashTable = (struct rx_peer **)ptable;
566 rx_lastAckDelay.sec = 0;
567 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
568 rx_hardAckDelay.sec = 0;
569 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
570 rx_softAckDelay.sec = 0;
571 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
573 rxevent_Init(20, rxi_ReScheduleEvents);
575 /* Initialize various global queues */
576 queue_Init(&rx_idleServerQueue);
577 queue_Init(&rx_incomingCallQueue);
578 queue_Init(&rx_freeCallQueue);
580 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
581 /* Initialize our list of usable IP addresses. */
585 /* Start listener process (exact function is dependent on the
586 * implementation environment--kernel or user space) */
590 tmp_status = rxinit_status = 0;
598 return rx_InitHost(htonl(INADDR_ANY), port);
601 /* called with unincremented nRequestsRunning to see if it is OK to start
602 * a new thread in this service. Could be "no" for two reasons: over the
603 * max quota, or would prevent others from reaching their min quota.
605 #ifdef RX_ENABLE_LOCKS
606 /* This verion of QuotaOK reserves quota if it's ok while the
607 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
610 QuotaOK(struct rx_service *aservice)
612 /* check if over max quota */
613 if (aservice->nRequestsRunning >= aservice->maxProcs) {
617 /* under min quota, we're OK */
618 /* otherwise, can use only if there are enough to allow everyone
619 * to go to their min quota after this guy starts.
622 MUTEX_ENTER(&rx_quota_mutex);
623 if ((aservice->nRequestsRunning < aservice->minProcs)
624 || (rxi_availProcs > rxi_minDeficit)) {
625 aservice->nRequestsRunning++;
626 /* just started call in minProcs pool, need fewer to maintain
628 if (aservice->nRequestsRunning <= aservice->minProcs)
631 MUTEX_EXIT(&rx_quota_mutex);
634 MUTEX_EXIT(&rx_quota_mutex);
640 ReturnToServerPool(struct rx_service *aservice)
642 aservice->nRequestsRunning--;
643 MUTEX_ENTER(&rx_quota_mutex);
644 if (aservice->nRequestsRunning < aservice->minProcs)
647 MUTEX_EXIT(&rx_quota_mutex);
650 #else /* RX_ENABLE_LOCKS */
652 QuotaOK(struct rx_service *aservice)
655 /* under min quota, we're OK */
656 if (aservice->nRequestsRunning < aservice->minProcs)
659 /* check if over max quota */
660 if (aservice->nRequestsRunning >= aservice->maxProcs)
663 /* otherwise, can use only if there are enough to allow everyone
664 * to go to their min quota after this guy starts.
666 if (rxi_availProcs > rxi_minDeficit)
670 #endif /* RX_ENABLE_LOCKS */
673 /* Called by rx_StartServer to start up lwp's to service calls.
674 NExistingProcs gives the number of procs already existing, and which
675 therefore needn't be created. */
677 rxi_StartServerProcs(int nExistingProcs)
679 struct rx_service *service;
684 /* For each service, reserve N processes, where N is the "minimum"
685 * number of processes that MUST be able to execute a request in parallel,
686 * at any time, for that process. Also compute the maximum difference
687 * between any service's maximum number of processes that can run
688 * (i.e. the maximum number that ever will be run, and a guarantee
689 * that this number will run if other services aren't running), and its
690 * minimum number. The result is the extra number of processes that
691 * we need in order to provide the latter guarantee */
692 for (i = 0; i < RX_MAX_SERVICES; i++) {
694 service = rx_services[i];
695 if (service == (struct rx_service *)0)
697 nProcs += service->minProcs;
698 diff = service->maxProcs - service->minProcs;
702 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
703 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
704 for (i = 0; i < nProcs; i++) {
705 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
711 /* This routine is only required on Windows */
713 rx_StartClientThread(void)
715 #ifdef AFS_PTHREAD_ENV
717 pid = pthread_self();
718 #endif /* AFS_PTHREAD_ENV */
720 #endif /* AFS_NT40_ENV */
722 /* This routine must be called if any services are exported. If the
723 * donateMe flag is set, the calling process is donated to the server
726 rx_StartServer(int donateMe)
728 struct rx_service *service;
734 /* Start server processes, if necessary (exact function is dependent
735 * on the implementation environment--kernel or user space). DonateMe
736 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
737 * case, one less new proc will be created rx_StartServerProcs.
739 rxi_StartServerProcs(donateMe);
741 /* count up the # of threads in minProcs, and add set the min deficit to
742 * be that value, too.
744 for (i = 0; i < RX_MAX_SERVICES; i++) {
745 service = rx_services[i];
746 if (service == (struct rx_service *)0)
748 MUTEX_ENTER(&rx_quota_mutex);
749 rxi_totalMin += service->minProcs;
750 /* below works even if a thread is running, since minDeficit would
751 * still have been decremented and later re-incremented.
753 rxi_minDeficit += service->minProcs;
754 MUTEX_EXIT(&rx_quota_mutex);
757 /* Turn on reaping of idle server connections */
758 rxi_ReapConnections(NULL, NULL, NULL);
767 #ifdef AFS_PTHREAD_ENV
769 pid = afs_pointer_to_int(pthread_self());
770 #else /* AFS_PTHREAD_ENV */
772 LWP_CurrentProcess(&pid);
773 #endif /* AFS_PTHREAD_ENV */
775 sprintf(name, "srv_%d", ++nProcs);
777 (*registerProgram) (pid, name);
779 #endif /* AFS_NT40_ENV */
780 rx_ServerProc(NULL); /* Never returns */
782 #ifdef RX_ENABLE_TSFPQ
783 /* no use leaving packets around in this thread's local queue if
784 * it isn't getting donated to the server thread pool.
786 rxi_FlushLocalPacketsTSFPQ();
787 #endif /* RX_ENABLE_TSFPQ */
791 /* Create a new client connection to the specified service, using the
792 * specified security object to implement the security model for this
794 struct rx_connection *
795 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
796 struct rx_securityClass *securityObject,
797 int serviceSecurityIndex)
801 struct rx_connection *conn;
806 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
807 "serviceSecurityIndex %d)\n",
808 ntohl(shost), ntohs(sport), sservice, securityObject,
809 serviceSecurityIndex));
811 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
812 * the case of kmem_alloc? */
813 conn = rxi_AllocConnection();
814 #ifdef RX_ENABLE_LOCKS
815 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
816 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
817 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
820 MUTEX_ENTER(&rx_connHashTable_lock);
821 cid = (rx_nextCid += RX_MAXCALLS);
822 conn->type = RX_CLIENT_CONNECTION;
824 conn->epoch = rx_epoch;
825 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
826 conn->serviceId = sservice;
827 conn->securityObject = securityObject;
828 conn->securityData = (void *) 0;
829 conn->securityIndex = serviceSecurityIndex;
830 rx_SetConnDeadTime(conn, rx_connDeadTime);
831 rx_SetConnSecondsUntilNatPing(conn, 0);
832 conn->ackRate = RX_FAST_ACK_RATE;
834 conn->specific = NULL;
835 conn->challengeEvent = NULL;
836 conn->delayedAbortEvent = NULL;
837 conn->abortCount = 0;
839 for (i = 0; i < RX_MAXCALLS; i++) {
840 conn->twind[i] = rx_initSendWindow;
841 conn->rwind[i] = rx_initReceiveWindow;
844 RXS_NewConnection(securityObject, conn);
846 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
848 conn->refCount++; /* no lock required since only this thread knows... */
849 conn->next = rx_connHashTable[hashindex];
850 rx_connHashTable[hashindex] = conn;
852 rx_MutexIncrement(rx_stats.nClientConns, rx_stats_mutex);
853 MUTEX_EXIT(&rx_connHashTable_lock);
859 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
861 /* The idea is to set the dead time to a value that allows several
862 * keepalives to be dropped without timing out the connection. */
863 conn->secondsUntilDead = MAX(seconds, 6);
864 conn->secondsUntilPing = conn->secondsUntilDead / 6;
867 int rxi_lowPeerRefCount = 0;
868 int rxi_lowConnRefCount = 0;
871 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
872 * NOTE: must not be called with rx_connHashTable_lock held.
875 rxi_CleanupConnection(struct rx_connection *conn)
877 /* Notify the service exporter, if requested, that this connection
878 * is being destroyed */
879 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
880 (*conn->service->destroyConnProc) (conn);
882 /* Notify the security module that this connection is being destroyed */
883 RXS_DestroyConnection(conn->securityObject, conn);
885 /* If this is the last connection using the rx_peer struct, set its
886 * idle time to now. rxi_ReapConnections will reap it if it's still
887 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
889 MUTEX_ENTER(&rx_peerHashTable_lock);
890 if (conn->peer->refCount < 2) {
891 conn->peer->idleWhen = clock_Sec();
892 if (conn->peer->refCount < 1) {
893 conn->peer->refCount = 1;
894 if (rx_stats_active) {
895 MUTEX_ENTER(&rx_stats_mutex);
896 rxi_lowPeerRefCount++;
897 MUTEX_EXIT(&rx_stats_mutex);
901 conn->peer->refCount--;
902 MUTEX_EXIT(&rx_peerHashTable_lock);
906 if (conn->type == RX_SERVER_CONNECTION)
907 rx_MutexDecrement(rx_stats.nServerConns, rx_stats_mutex);
909 rx_MutexDecrement(rx_stats.nClientConns, rx_stats_mutex);
912 if (conn->specific) {
914 for (i = 0; i < conn->nSpecific; i++) {
915 if (conn->specific[i] && rxi_keyCreate_destructor[i])
916 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
917 conn->specific[i] = NULL;
919 free(conn->specific);
921 conn->specific = NULL;
925 MUTEX_DESTROY(&conn->conn_call_lock);
926 MUTEX_DESTROY(&conn->conn_data_lock);
927 CV_DESTROY(&conn->conn_call_cv);
929 rxi_FreeConnection(conn);
932 /* Destroy the specified connection */
934 rxi_DestroyConnection(struct rx_connection *conn)
936 MUTEX_ENTER(&rx_connHashTable_lock);
937 rxi_DestroyConnectionNoLock(conn);
938 /* conn should be at the head of the cleanup list */
939 if (conn == rx_connCleanup_list) {
940 rx_connCleanup_list = rx_connCleanup_list->next;
941 MUTEX_EXIT(&rx_connHashTable_lock);
942 rxi_CleanupConnection(conn);
944 #ifdef RX_ENABLE_LOCKS
946 MUTEX_EXIT(&rx_connHashTable_lock);
948 #endif /* RX_ENABLE_LOCKS */
952 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
954 struct rx_connection **conn_ptr;
956 struct rx_packet *packet;
963 MUTEX_ENTER(&conn->conn_data_lock);
964 if (conn->refCount > 0)
967 if (rx_stats_active) {
968 MUTEX_ENTER(&rx_stats_mutex);
969 rxi_lowConnRefCount++;
970 MUTEX_EXIT(&rx_stats_mutex);
974 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
975 /* Busy; wait till the last guy before proceeding */
976 MUTEX_EXIT(&conn->conn_data_lock);
981 /* If the client previously called rx_NewCall, but it is still
982 * waiting, treat this as a running call, and wait to destroy the
983 * connection later when the call completes. */
984 if ((conn->type == RX_CLIENT_CONNECTION)
985 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
986 conn->flags |= RX_CONN_DESTROY_ME;
987 MUTEX_EXIT(&conn->conn_data_lock);
991 MUTEX_EXIT(&conn->conn_data_lock);
993 /* Check for extant references to this connection */
994 for (i = 0; i < RX_MAXCALLS; i++) {
995 struct rx_call *call = conn->call[i];
998 if (conn->type == RX_CLIENT_CONNECTION) {
999 MUTEX_ENTER(&call->lock);
1000 if (call->delayedAckEvent) {
1001 /* Push the final acknowledgment out now--there
1002 * won't be a subsequent call to acknowledge the
1003 * last reply packets */
1004 rxevent_Cancel(call->delayedAckEvent, call,
1005 RX_CALL_REFCOUNT_DELAY);
1006 if (call->state == RX_STATE_PRECALL
1007 || call->state == RX_STATE_ACTIVE) {
1008 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1010 rxi_AckAll(NULL, call, 0);
1013 MUTEX_EXIT(&call->lock);
1017 #ifdef RX_ENABLE_LOCKS
1019 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1020 MUTEX_EXIT(&conn->conn_data_lock);
1022 /* Someone is accessing a packet right now. */
1026 #endif /* RX_ENABLE_LOCKS */
1029 /* Don't destroy the connection if there are any call
1030 * structures still in use */
1031 MUTEX_ENTER(&conn->conn_data_lock);
1032 conn->flags |= RX_CONN_DESTROY_ME;
1033 MUTEX_EXIT(&conn->conn_data_lock);
1038 if (conn->natKeepAliveEvent) {
1039 rxi_NatKeepAliveOff(conn);
1042 if (conn->delayedAbortEvent) {
1043 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1044 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1046 MUTEX_ENTER(&conn->conn_data_lock);
1047 rxi_SendConnectionAbort(conn, packet, 0, 1);
1048 MUTEX_EXIT(&conn->conn_data_lock);
1049 rxi_FreePacket(packet);
1053 /* Remove from connection hash table before proceeding */
1055 &rx_connHashTable[CONN_HASH
1056 (peer->host, peer->port, conn->cid, conn->epoch,
1058 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1059 if (*conn_ptr == conn) {
1060 *conn_ptr = conn->next;
1064 /* if the conn that we are destroying was the last connection, then we
1065 * clear rxLastConn as well */
1066 if (rxLastConn == conn)
1069 /* Make sure the connection is completely reset before deleting it. */
1070 /* get rid of pending events that could zap us later */
1071 if (conn->challengeEvent)
1072 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1073 if (conn->checkReachEvent)
1074 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1075 if (conn->natKeepAliveEvent)
1076 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
1078 /* Add the connection to the list of destroyed connections that
1079 * need to be cleaned up. This is necessary to avoid deadlocks
1080 * in the routines we call to inform others that this connection is
1081 * being destroyed. */
1082 conn->next = rx_connCleanup_list;
1083 rx_connCleanup_list = conn;
1086 /* Externally available version */
1088 rx_DestroyConnection(struct rx_connection *conn)
1093 rxi_DestroyConnection(conn);
1098 rx_GetConnection(struct rx_connection *conn)
1103 MUTEX_ENTER(&conn->conn_data_lock);
1105 MUTEX_EXIT(&conn->conn_data_lock);
1109 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1110 /* Wait for the transmit queue to no longer be busy.
1111 * requires the call->lock to be held */
1112 static void rxi_WaitforTQBusy(struct rx_call *call) {
1113 while (call->flags & RX_CALL_TQ_BUSY) {
1114 call->flags |= RX_CALL_TQ_WAIT;
1116 #ifdef RX_ENABLE_LOCKS
1117 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1118 CV_WAIT(&call->cv_tq, &call->lock);
1119 #else /* RX_ENABLE_LOCKS */
1120 osi_rxSleep(&call->tq);
1121 #endif /* RX_ENABLE_LOCKS */
1123 if (call->tqWaiters == 0) {
1124 call->flags &= ~RX_CALL_TQ_WAIT;
1130 /* Start a new rx remote procedure call, on the specified connection.
1131 * If wait is set to 1, wait for a free call channel; otherwise return
1132 * 0. Maxtime gives the maximum number of seconds this call may take,
1133 * after rx_NewCall returns. After this time interval, a call to any
1134 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1135 * For fine grain locking, we hold the conn_call_lock in order to
1136 * to ensure that we don't get signalle after we found a call in an active
1137 * state and before we go to sleep.
1140 rx_NewCall(struct rx_connection *conn)
1143 struct rx_call *call;
1144 struct clock queueTime;
1148 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1151 clock_GetTime(&queueTime);
1152 MUTEX_ENTER(&conn->conn_call_lock);
1155 * Check if there are others waiting for a new call.
1156 * If so, let them go first to avoid starving them.
1157 * This is a fairly simple scheme, and might not be
1158 * a complete solution for large numbers of waiters.
1160 * makeCallWaiters keeps track of the number of
1161 * threads waiting to make calls and the
1162 * RX_CONN_MAKECALL_WAITING flag bit is used to
1163 * indicate that there are indeed calls waiting.
1164 * The flag is set when the waiter is incremented.
1165 * It is only cleared in rx_EndCall when
1166 * makeCallWaiters is 0. This prevents us from
1167 * accidently destroying the connection while it
1168 * is potentially about to be used.
1170 MUTEX_ENTER(&conn->conn_data_lock);
1171 if (conn->makeCallWaiters) {
1172 conn->flags |= RX_CONN_MAKECALL_WAITING;
1173 conn->makeCallWaiters++;
1174 MUTEX_EXIT(&conn->conn_data_lock);
1176 #ifdef RX_ENABLE_LOCKS
1177 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1181 MUTEX_ENTER(&conn->conn_data_lock);
1182 conn->makeCallWaiters--;
1184 MUTEX_EXIT(&conn->conn_data_lock);
1187 for (i = 0; i < RX_MAXCALLS; i++) {
1188 call = conn->call[i];
1190 MUTEX_ENTER(&call->lock);
1191 if (call->state == RX_STATE_DALLY) {
1192 rxi_ResetCall(call, 0);
1193 (*call->callNumber)++;
1196 MUTEX_EXIT(&call->lock);
1198 call = rxi_NewCall(conn, i);
1202 if (i < RX_MAXCALLS) {
1205 MUTEX_ENTER(&conn->conn_data_lock);
1206 conn->flags |= RX_CONN_MAKECALL_WAITING;
1207 conn->makeCallWaiters++;
1208 MUTEX_EXIT(&conn->conn_data_lock);
1210 #ifdef RX_ENABLE_LOCKS
1211 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1215 MUTEX_ENTER(&conn->conn_data_lock);
1216 conn->makeCallWaiters--;
1217 MUTEX_EXIT(&conn->conn_data_lock);
1220 * Wake up anyone else who might be giving us a chance to
1221 * run (see code above that avoids resource starvation).
1223 #ifdef RX_ENABLE_LOCKS
1224 CV_BROADCAST(&conn->conn_call_cv);
1229 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1231 /* Client is initially in send mode */
1232 call->state = RX_STATE_ACTIVE;
1233 call->error = conn->error;
1235 call->mode = RX_MODE_ERROR;
1237 call->mode = RX_MODE_SENDING;
1239 /* remember start time for call in case we have hard dead time limit */
1240 call->queueTime = queueTime;
1241 clock_GetTime(&call->startTime);
1242 hzero(call->bytesSent);
1243 hzero(call->bytesRcvd);
1245 /* Turn on busy protocol. */
1246 rxi_KeepAliveOn(call);
1248 MUTEX_EXIT(&call->lock);
1249 MUTEX_EXIT(&conn->conn_call_lock);
1252 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1253 /* Now, if TQ wasn't cleared earlier, do it now. */
1254 MUTEX_ENTER(&call->lock);
1255 rxi_WaitforTQBusy(call);
1256 if (call->flags & RX_CALL_TQ_CLEARME) {
1257 rxi_ClearTransmitQueue(call, 1);
1258 /*queue_Init(&call->tq);*/
1260 MUTEX_EXIT(&call->lock);
1261 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1263 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1268 rxi_HasActiveCalls(struct rx_connection *aconn)
1271 struct rx_call *tcall;
1275 for (i = 0; i < RX_MAXCALLS; i++) {
1276 if ((tcall = aconn->call[i])) {
1277 if ((tcall->state == RX_STATE_ACTIVE)
1278 || (tcall->state == RX_STATE_PRECALL)) {
1289 rxi_GetCallNumberVector(struct rx_connection *aconn,
1290 afs_int32 * aint32s)
1293 struct rx_call *tcall;
1297 for (i = 0; i < RX_MAXCALLS; i++) {
1298 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1299 aint32s[i] = aconn->callNumber[i] + 1;
1301 aint32s[i] = aconn->callNumber[i];
1308 rxi_SetCallNumberVector(struct rx_connection *aconn,
1309 afs_int32 * aint32s)
1312 struct rx_call *tcall;
1316 for (i = 0; i < RX_MAXCALLS; i++) {
1317 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1318 aconn->callNumber[i] = aint32s[i] - 1;
1320 aconn->callNumber[i] = aint32s[i];
1326 /* Advertise a new service. A service is named locally by a UDP port
1327 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1330 char *serviceName; Name for identification purposes (e.g. the
1331 service name might be used for probing for
1334 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1335 char *serviceName, struct rx_securityClass **securityObjects,
1336 int nSecurityObjects,
1337 afs_int32(*serviceProc) (struct rx_call * acall))
1339 osi_socket socket = OSI_NULLSOCKET;
1340 struct rx_service *tservice;
1346 if (serviceId == 0) {
1348 "rx_NewService: service id for service %s is not non-zero.\n",
1355 "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",
1363 tservice = rxi_AllocService();
1365 for (i = 0; i < RX_MAX_SERVICES; i++) {
1366 struct rx_service *service = rx_services[i];
1368 if (port == service->servicePort && host == service->serviceHost) {
1369 if (service->serviceId == serviceId) {
1370 /* The identical service has already been
1371 * installed; if the caller was intending to
1372 * change the security classes used by this
1373 * service, he/she loses. */
1375 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1376 serviceName, serviceId, service->serviceName);
1378 rxi_FreeService(tservice);
1381 /* Different service, same port: re-use the socket
1382 * which is bound to the same port */
1383 socket = service->socket;
1386 if (socket == OSI_NULLSOCKET) {
1387 /* If we don't already have a socket (from another
1388 * service on same port) get a new one */
1389 socket = rxi_GetHostUDPSocket(host, port);
1390 if (socket == OSI_NULLSOCKET) {
1392 rxi_FreeService(tservice);
1397 service->socket = socket;
1398 service->serviceHost = host;
1399 service->servicePort = port;
1400 service->serviceId = serviceId;
1401 service->serviceName = serviceName;
1402 service->nSecurityObjects = nSecurityObjects;
1403 service->securityObjects = securityObjects;
1404 service->minProcs = 0;
1405 service->maxProcs = 1;
1406 service->idleDeadTime = 60;
1407 service->idleDeadErr = 0;
1408 service->connDeadTime = rx_connDeadTime;
1409 service->executeRequestProc = serviceProc;
1410 service->checkReach = 0;
1411 rx_services[i] = service; /* not visible until now */
1417 rxi_FreeService(tservice);
1418 (osi_Msg "rx_NewService: cannot support > %d services\n",
1423 /* Set configuration options for all of a service's security objects */
1426 rx_SetSecurityConfiguration(struct rx_service *service,
1427 rx_securityConfigVariables type,
1431 for (i = 0; i<service->nSecurityObjects; i++) {
1432 if (service->securityObjects[i]) {
1433 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1441 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1442 struct rx_securityClass **securityObjects, int nSecurityObjects,
1443 afs_int32(*serviceProc) (struct rx_call * acall))
1445 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1448 /* Generic request processing loop. This routine should be called
1449 * by the implementation dependent rx_ServerProc. If socketp is
1450 * non-null, it will be set to the file descriptor that this thread
1451 * is now listening on. If socketp is null, this routine will never
1454 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1456 struct rx_call *call;
1458 struct rx_service *tservice = NULL;
1465 call = rx_GetCall(threadID, tservice, socketp);
1466 if (socketp && *socketp != OSI_NULLSOCKET) {
1467 /* We are now a listener thread */
1472 /* if server is restarting( typically smooth shutdown) then do not
1473 * allow any new calls.
1476 if (rx_tranquil && (call != NULL)) {
1480 MUTEX_ENTER(&call->lock);
1482 rxi_CallError(call, RX_RESTARTING);
1483 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1485 MUTEX_EXIT(&call->lock);
1489 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1490 #ifdef RX_ENABLE_LOCKS
1492 #endif /* RX_ENABLE_LOCKS */
1493 afs_termState = AFSOP_STOP_AFS;
1494 afs_osi_Wakeup(&afs_termState);
1495 #ifdef RX_ENABLE_LOCKS
1497 #endif /* RX_ENABLE_LOCKS */
1502 tservice = call->conn->service;
1504 if (tservice->beforeProc)
1505 (*tservice->beforeProc) (call);
1507 code = call->conn->service->executeRequestProc(call);
1509 if (tservice->afterProc)
1510 (*tservice->afterProc) (call, code);
1512 rx_EndCall(call, code);
1513 if (rx_stats_active) {
1514 MUTEX_ENTER(&rx_stats_mutex);
1516 MUTEX_EXIT(&rx_stats_mutex);
1523 rx_WakeupServerProcs(void)
1525 struct rx_serverQueueEntry *np, *tqp;
1529 MUTEX_ENTER(&rx_serverPool_lock);
1531 #ifdef RX_ENABLE_LOCKS
1532 if (rx_waitForPacket)
1533 CV_BROADCAST(&rx_waitForPacket->cv);
1534 #else /* RX_ENABLE_LOCKS */
1535 if (rx_waitForPacket)
1536 osi_rxWakeup(rx_waitForPacket);
1537 #endif /* RX_ENABLE_LOCKS */
1538 MUTEX_ENTER(&freeSQEList_lock);
1539 for (np = rx_FreeSQEList; np; np = tqp) {
1540 tqp = *(struct rx_serverQueueEntry **)np;
1541 #ifdef RX_ENABLE_LOCKS
1542 CV_BROADCAST(&np->cv);
1543 #else /* RX_ENABLE_LOCKS */
1545 #endif /* RX_ENABLE_LOCKS */
1547 MUTEX_EXIT(&freeSQEList_lock);
1548 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1549 #ifdef RX_ENABLE_LOCKS
1550 CV_BROADCAST(&np->cv);
1551 #else /* RX_ENABLE_LOCKS */
1553 #endif /* RX_ENABLE_LOCKS */
1555 MUTEX_EXIT(&rx_serverPool_lock);
1560 * One thing that seems to happen is that all the server threads get
1561 * tied up on some empty or slow call, and then a whole bunch of calls
1562 * arrive at once, using up the packet pool, so now there are more
1563 * empty calls. The most critical resources here are server threads
1564 * and the free packet pool. The "doreclaim" code seems to help in
1565 * general. I think that eventually we arrive in this state: there
1566 * are lots of pending calls which do have all their packets present,
1567 * so they won't be reclaimed, are multi-packet calls, so they won't
1568 * be scheduled until later, and thus are tying up most of the free
1569 * packet pool for a very long time.
1571 * 1. schedule multi-packet calls if all the packets are present.
1572 * Probably CPU-bound operation, useful to return packets to pool.
1573 * Do what if there is a full window, but the last packet isn't here?
1574 * 3. preserve one thread which *only* runs "best" calls, otherwise
1575 * it sleeps and waits for that type of call.
1576 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1577 * the current dataquota business is badly broken. The quota isn't adjusted
1578 * to reflect how many packets are presently queued for a running call.
1579 * So, when we schedule a queued call with a full window of packets queued
1580 * up for it, that *should* free up a window full of packets for other 2d-class
1581 * calls to be able to use from the packet pool. But it doesn't.
1583 * NB. Most of the time, this code doesn't run -- since idle server threads
1584 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1585 * as a new call arrives.
1587 /* Sleep until a call arrives. Returns a pointer to the call, ready
1588 * for an rx_Read. */
1589 #ifdef RX_ENABLE_LOCKS
1591 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1593 struct rx_serverQueueEntry *sq;
1594 struct rx_call *call = (struct rx_call *)0;
1595 struct rx_service *service = NULL;
1598 MUTEX_ENTER(&freeSQEList_lock);
1600 if ((sq = rx_FreeSQEList)) {
1601 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1602 MUTEX_EXIT(&freeSQEList_lock);
1603 } else { /* otherwise allocate a new one and return that */
1604 MUTEX_EXIT(&freeSQEList_lock);
1605 sq = (struct rx_serverQueueEntry *)
1606 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1607 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1608 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1611 MUTEX_ENTER(&rx_serverPool_lock);
1612 if (cur_service != NULL) {
1613 ReturnToServerPool(cur_service);
1616 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1617 struct rx_call *tcall, *ncall, *choice2 = NULL;
1619 /* Scan for eligible incoming calls. A call is not eligible
1620 * if the maximum number of calls for its service type are
1621 * already executing */
1622 /* One thread will process calls FCFS (to prevent starvation),
1623 * while the other threads may run ahead looking for calls which
1624 * have all their input data available immediately. This helps
1625 * keep threads from blocking, waiting for data from the client. */
1626 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1627 service = tcall->conn->service;
1628 if (!QuotaOK(service)) {
1631 MUTEX_ENTER(&rx_pthread_mutex);
1632 if (tno == rxi_fcfs_thread_num
1633 || !tcall->queue_item_header.next) {
1634 MUTEX_EXIT(&rx_pthread_mutex);
1635 /* If we're the fcfs thread , then we'll just use
1636 * this call. If we haven't been able to find an optimal
1637 * choice, and we're at the end of the list, then use a
1638 * 2d choice if one has been identified. Otherwise... */
1639 call = (choice2 ? choice2 : tcall);
1640 service = call->conn->service;
1642 MUTEX_EXIT(&rx_pthread_mutex);
1643 if (!queue_IsEmpty(&tcall->rq)) {
1644 struct rx_packet *rp;
1645 rp = queue_First(&tcall->rq, rx_packet);
1646 if (rp->header.seq == 1) {
1648 || (rp->header.flags & RX_LAST_PACKET)) {
1650 } else if (rxi_2dchoice && !choice2
1651 && !(tcall->flags & RX_CALL_CLEARED)
1652 && (tcall->rprev > rxi_HardAckRate)) {
1662 ReturnToServerPool(service);
1669 MUTEX_EXIT(&rx_serverPool_lock);
1670 MUTEX_ENTER(&call->lock);
1672 if (call->flags & RX_CALL_WAIT_PROC) {
1673 call->flags &= ~RX_CALL_WAIT_PROC;
1674 MUTEX_ENTER(&rx_waiting_mutex);
1676 MUTEX_EXIT(&rx_waiting_mutex);
1679 if (call->state != RX_STATE_PRECALL || call->error) {
1680 MUTEX_EXIT(&call->lock);
1681 MUTEX_ENTER(&rx_serverPool_lock);
1682 ReturnToServerPool(service);
1687 if (queue_IsEmpty(&call->rq)
1688 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1689 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1691 CLEAR_CALL_QUEUE_LOCK(call);
1694 /* If there are no eligible incoming calls, add this process
1695 * to the idle server queue, to wait for one */
1699 *socketp = OSI_NULLSOCKET;
1701 sq->socketp = socketp;
1702 queue_Append(&rx_idleServerQueue, sq);
1703 #ifndef AFS_AIX41_ENV
1704 rx_waitForPacket = sq;
1706 rx_waitingForPacket = sq;
1707 #endif /* AFS_AIX41_ENV */
1709 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1711 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1712 MUTEX_EXIT(&rx_serverPool_lock);
1713 return (struct rx_call *)0;
1716 } while (!(call = sq->newcall)
1717 && !(socketp && *socketp != OSI_NULLSOCKET));
1718 MUTEX_EXIT(&rx_serverPool_lock);
1720 MUTEX_ENTER(&call->lock);
1726 MUTEX_ENTER(&freeSQEList_lock);
1727 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1728 rx_FreeSQEList = sq;
1729 MUTEX_EXIT(&freeSQEList_lock);
1732 clock_GetTime(&call->startTime);
1733 call->state = RX_STATE_ACTIVE;
1734 call->mode = RX_MODE_RECEIVING;
1735 #ifdef RX_KERNEL_TRACE
1736 if (ICL_SETACTIVE(afs_iclSetp)) {
1737 int glockOwner = ISAFS_GLOCK();
1740 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1741 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1748 rxi_calltrace(RX_CALL_START, call);
1749 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
1750 call->conn->service->servicePort, call->conn->service->serviceId,
1753 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1754 MUTEX_EXIT(&call->lock);
1756 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1761 #else /* RX_ENABLE_LOCKS */
1763 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1765 struct rx_serverQueueEntry *sq;
1766 struct rx_call *call = (struct rx_call *)0, *choice2;
1767 struct rx_service *service = NULL;
1771 MUTEX_ENTER(&freeSQEList_lock);
1773 if ((sq = rx_FreeSQEList)) {
1774 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1775 MUTEX_EXIT(&freeSQEList_lock);
1776 } else { /* otherwise allocate a new one and return that */
1777 MUTEX_EXIT(&freeSQEList_lock);
1778 sq = (struct rx_serverQueueEntry *)
1779 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1780 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1781 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1783 MUTEX_ENTER(&sq->lock);
1785 if (cur_service != NULL) {
1786 cur_service->nRequestsRunning--;
1787 if (cur_service->nRequestsRunning < cur_service->minProcs)
1791 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1792 struct rx_call *tcall, *ncall;
1793 /* Scan for eligible incoming calls. A call is not eligible
1794 * if the maximum number of calls for its service type are
1795 * already executing */
1796 /* One thread will process calls FCFS (to prevent starvation),
1797 * while the other threads may run ahead looking for calls which
1798 * have all their input data available immediately. This helps
1799 * keep threads from blocking, waiting for data from the client. */
1800 choice2 = (struct rx_call *)0;
1801 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1802 service = tcall->conn->service;
1803 if (QuotaOK(service)) {
1804 MUTEX_ENTER(&rx_pthread_mutex);
1805 if (tno == rxi_fcfs_thread_num
1806 || !tcall->queue_item_header.next) {
1807 MUTEX_EXIT(&rx_pthread_mutex);
1808 /* If we're the fcfs thread, then we'll just use
1809 * this call. If we haven't been able to find an optimal
1810 * choice, and we're at the end of the list, then use a
1811 * 2d choice if one has been identified. Otherwise... */
1812 call = (choice2 ? choice2 : tcall);
1813 service = call->conn->service;
1815 MUTEX_EXIT(&rx_pthread_mutex);
1816 if (!queue_IsEmpty(&tcall->rq)) {
1817 struct rx_packet *rp;
1818 rp = queue_First(&tcall->rq, rx_packet);
1819 if (rp->header.seq == 1
1821 || (rp->header.flags & RX_LAST_PACKET))) {
1823 } else if (rxi_2dchoice && !choice2
1824 && !(tcall->flags & RX_CALL_CLEARED)
1825 && (tcall->rprev > rxi_HardAckRate)) {
1839 /* we can't schedule a call if there's no data!!! */
1840 /* send an ack if there's no data, if we're missing the
1841 * first packet, or we're missing something between first
1842 * and last -- there's a "hole" in the incoming data. */
1843 if (queue_IsEmpty(&call->rq)
1844 || queue_First(&call->rq, rx_packet)->header.seq != 1
1845 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1846 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1848 call->flags &= (~RX_CALL_WAIT_PROC);
1849 service->nRequestsRunning++;
1850 /* just started call in minProcs pool, need fewer to maintain
1852 if (service->nRequestsRunning <= service->minProcs)
1856 /* MUTEX_EXIT(&call->lock); */
1858 /* If there are no eligible incoming calls, add this process
1859 * to the idle server queue, to wait for one */
1862 *socketp = OSI_NULLSOCKET;
1864 sq->socketp = socketp;
1865 queue_Append(&rx_idleServerQueue, sq);
1869 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1871 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1872 return (struct rx_call *)0;
1875 } while (!(call = sq->newcall)
1876 && !(socketp && *socketp != OSI_NULLSOCKET));
1878 MUTEX_EXIT(&sq->lock);
1880 MUTEX_ENTER(&freeSQEList_lock);
1881 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1882 rx_FreeSQEList = sq;
1883 MUTEX_EXIT(&freeSQEList_lock);
1886 clock_GetTime(&call->startTime);
1887 call->state = RX_STATE_ACTIVE;
1888 call->mode = RX_MODE_RECEIVING;
1889 #ifdef RX_KERNEL_TRACE
1890 if (ICL_SETACTIVE(afs_iclSetp)) {
1891 int glockOwner = ISAFS_GLOCK();
1894 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1895 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1902 rxi_calltrace(RX_CALL_START, call);
1903 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
1904 call->conn->service->servicePort, call->conn->service->serviceId,
1907 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1914 #endif /* RX_ENABLE_LOCKS */
1918 /* Establish a procedure to be called when a packet arrives for a
1919 * call. This routine will be called at most once after each call,
1920 * and will also be called if there is an error condition on the or
1921 * the call is complete. Used by multi rx to build a selection
1922 * function which determines which of several calls is likely to be a
1923 * good one to read from.
1924 * NOTE: the way this is currently implemented it is probably only a
1925 * good idea to (1) use it immediately after a newcall (clients only)
1926 * and (2) only use it once. Other uses currently void your warranty
1929 rx_SetArrivalProc(struct rx_call *call,
1930 void (*proc) (struct rx_call * call,
1933 void * handle, int arg)
1935 call->arrivalProc = proc;
1936 call->arrivalProcHandle = handle;
1937 call->arrivalProcArg = arg;
1940 /* Call is finished (possibly prematurely). Return rc to the peer, if
1941 * appropriate, and return the final error code from the conversation
1945 rx_EndCall(struct rx_call *call, afs_int32 rc)
1947 struct rx_connection *conn = call->conn;
1948 struct rx_service *service;
1954 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
1955 call, rc, call->error, call->abortCode));
1958 MUTEX_ENTER(&call->lock);
1960 if (rc == 0 && call->error == 0) {
1961 call->abortCode = 0;
1962 call->abortCount = 0;
1965 call->arrivalProc = (void (*)())0;
1966 if (rc && call->error == 0) {
1967 rxi_CallError(call, rc);
1968 /* Send an abort message to the peer if this error code has
1969 * only just been set. If it was set previously, assume the
1970 * peer has already been sent the error code or will request it
1972 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1974 if (conn->type == RX_SERVER_CONNECTION) {
1975 /* Make sure reply or at least dummy reply is sent */
1976 if (call->mode == RX_MODE_RECEIVING) {
1977 rxi_WriteProc(call, 0, 0);
1979 if (call->mode == RX_MODE_SENDING) {
1980 rxi_FlushWrite(call);
1982 service = conn->service;
1983 rxi_calltrace(RX_CALL_END, call);
1984 /* Call goes to hold state until reply packets are acknowledged */
1985 if (call->tfirst + call->nSoftAcked < call->tnext) {
1986 call->state = RX_STATE_HOLD;
1988 call->state = RX_STATE_DALLY;
1989 rxi_ClearTransmitQueue(call, 0);
1990 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1991 rxevent_Cancel(call->keepAliveEvent, call,
1992 RX_CALL_REFCOUNT_ALIVE);
1994 } else { /* Client connection */
1996 /* Make sure server receives input packets, in the case where
1997 * no reply arguments are expected */
1998 if ((call->mode == RX_MODE_SENDING)
1999 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2000 (void)rxi_ReadProc(call, &dummy, 1);
2003 /* If we had an outstanding delayed ack, be nice to the server
2004 * and force-send it now.
2006 if (call->delayedAckEvent) {
2007 rxevent_Cancel(call->delayedAckEvent, call,
2008 RX_CALL_REFCOUNT_DELAY);
2009 call->delayedAckEvent = NULL;
2010 rxi_SendDelayedAck(NULL, call, NULL);
2013 /* We need to release the call lock since it's lower than the
2014 * conn_call_lock and we don't want to hold the conn_call_lock
2015 * over the rx_ReadProc call. The conn_call_lock needs to be held
2016 * here for the case where rx_NewCall is perusing the calls on
2017 * the connection structure. We don't want to signal until
2018 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2019 * have checked this call, found it active and by the time it
2020 * goes to sleep, will have missed the signal.
2022 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
2023 * there are threads waiting to use the conn object.
2025 MUTEX_EXIT(&call->lock);
2026 MUTEX_ENTER(&conn->conn_call_lock);
2027 MUTEX_ENTER(&call->lock);
2028 MUTEX_ENTER(&conn->conn_data_lock);
2029 conn->flags |= RX_CONN_BUSY;
2030 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2031 if (conn->makeCallWaiters == 0)
2032 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
2033 MUTEX_EXIT(&conn->conn_data_lock);
2034 #ifdef RX_ENABLE_LOCKS
2035 CV_BROADCAST(&conn->conn_call_cv);
2040 #ifdef RX_ENABLE_LOCKS
2042 MUTEX_EXIT(&conn->conn_data_lock);
2044 #endif /* RX_ENABLE_LOCKS */
2045 call->state = RX_STATE_DALLY;
2047 error = call->error;
2049 /* currentPacket, nLeft, and NFree must be zeroed here, because
2050 * ResetCall cannot: ResetCall may be called at splnet(), in the
2051 * kernel version, and may interrupt the macros rx_Read or
2052 * rx_Write, which run at normal priority for efficiency. */
2053 if (call->currentPacket) {
2054 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2055 rxi_FreePacket(call->currentPacket);
2056 call->currentPacket = (struct rx_packet *)0;
2059 call->nLeft = call->nFree = call->curlen = 0;
2061 /* Free any packets from the last call to ReadvProc/WritevProc */
2062 #ifdef RXDEBUG_PACKET
2064 #endif /* RXDEBUG_PACKET */
2065 rxi_FreePackets(0, &call->iovq);
2067 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2068 MUTEX_EXIT(&call->lock);
2069 if (conn->type == RX_CLIENT_CONNECTION) {
2070 MUTEX_EXIT(&conn->conn_call_lock);
2071 conn->flags &= ~RX_CONN_BUSY;
2075 * Map errors to the local host's errno.h format.
2077 error = ntoh_syserr_conv(error);
2081 #if !defined(KERNEL)
2083 /* Call this routine when shutting down a server or client (especially
2084 * clients). This will allow Rx to gracefully garbage collect server
2085 * connections, and reduce the number of retries that a server might
2086 * make to a dead client.
2087 * This is not quite right, since some calls may still be ongoing and
2088 * we can't lock them to destroy them. */
2092 struct rx_connection **conn_ptr, **conn_end;
2096 if (rxinit_status == 1) {
2098 return; /* Already shutdown. */
2100 rxi_DeleteCachedConnections();
2101 if (rx_connHashTable) {
2102 MUTEX_ENTER(&rx_connHashTable_lock);
2103 for (conn_ptr = &rx_connHashTable[0], conn_end =
2104 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2106 struct rx_connection *conn, *next;
2107 for (conn = *conn_ptr; conn; conn = next) {
2109 if (conn->type == RX_CLIENT_CONNECTION) {
2110 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2112 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2113 #ifdef RX_ENABLE_LOCKS
2114 rxi_DestroyConnectionNoLock(conn);
2115 #else /* RX_ENABLE_LOCKS */
2116 rxi_DestroyConnection(conn);
2117 #endif /* RX_ENABLE_LOCKS */
2121 #ifdef RX_ENABLE_LOCKS
2122 while (rx_connCleanup_list) {
2123 struct rx_connection *conn;
2124 conn = rx_connCleanup_list;
2125 rx_connCleanup_list = rx_connCleanup_list->next;
2126 MUTEX_EXIT(&rx_connHashTable_lock);
2127 rxi_CleanupConnection(conn);
2128 MUTEX_ENTER(&rx_connHashTable_lock);
2130 MUTEX_EXIT(&rx_connHashTable_lock);
2131 #endif /* RX_ENABLE_LOCKS */
2136 afs_winsockCleanup();
2144 /* if we wakeup packet waiter too often, can get in loop with two
2145 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2147 rxi_PacketsUnWait(void)
2149 if (!rx_waitingForPackets) {
2153 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2154 return; /* still over quota */
2157 rx_waitingForPackets = 0;
2158 #ifdef RX_ENABLE_LOCKS
2159 CV_BROADCAST(&rx_waitingForPackets_cv);
2161 osi_rxWakeup(&rx_waitingForPackets);
2167 /* ------------------Internal interfaces------------------------- */
2169 /* Return this process's service structure for the
2170 * specified socket and service */
2172 rxi_FindService(osi_socket socket, u_short serviceId)
2174 struct rx_service **sp;
2175 for (sp = &rx_services[0]; *sp; sp++) {
2176 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2182 #ifdef RXDEBUG_PACKET
2183 #ifdef KDUMP_RX_LOCK
2184 static struct rx_call_rx_lock *rx_allCallsp = 0;
2186 static struct rx_call *rx_allCallsp = 0;
2188 #endif /* RXDEBUG_PACKET */
2190 /* Allocate a call structure, for the indicated channel of the
2191 * supplied connection. The mode and state of the call must be set by
2192 * the caller. Returns the call with mutex locked. */
2194 rxi_NewCall(struct rx_connection *conn, int channel)
2196 struct rx_call *call;
2197 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2198 struct rx_call *cp; /* Call pointer temp */
2199 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2200 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2202 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2204 /* Grab an existing call structure, or allocate a new one.
2205 * Existing call structures are assumed to have been left reset by
2207 MUTEX_ENTER(&rx_freeCallQueue_lock);
2209 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2211 * EXCEPT that the TQ might not yet be cleared out.
2212 * Skip over those with in-use TQs.
2215 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2216 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2222 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2223 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2224 call = queue_First(&rx_freeCallQueue, rx_call);
2225 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2227 if (rx_stats_active)
2228 rx_MutexDecrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2229 MUTEX_EXIT(&rx_freeCallQueue_lock);
2230 MUTEX_ENTER(&call->lock);
2231 CLEAR_CALL_QUEUE_LOCK(call);
2232 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2233 /* Now, if TQ wasn't cleared earlier, do it now. */
2234 rxi_WaitforTQBusy(call);
2235 if (call->flags & RX_CALL_TQ_CLEARME) {
2236 rxi_ClearTransmitQueue(call, 1);
2237 /*queue_Init(&call->tq);*/
2239 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2240 /* Bind the call to its connection structure */
2242 rxi_ResetCall(call, 1);
2245 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2246 #ifdef RXDEBUG_PACKET
2247 call->allNextp = rx_allCallsp;
2248 rx_allCallsp = call;
2250 #endif /* RXDEBUG_PACKET */
2251 rx_MutexIncrement(rx_stats.nCallStructs, rx_stats_mutex);
2253 MUTEX_EXIT(&rx_freeCallQueue_lock);
2254 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2255 MUTEX_ENTER(&call->lock);
2256 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2257 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2258 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2260 /* Initialize once-only items */
2261 queue_Init(&call->tq);
2262 queue_Init(&call->rq);
2263 queue_Init(&call->iovq);
2264 #ifdef RXDEBUG_PACKET
2265 call->rqc = call->tqc = call->iovqc = 0;
2266 #endif /* RXDEBUG_PACKET */
2267 /* Bind the call to its connection structure (prereq for reset) */
2269 rxi_ResetCall(call, 1);
2271 call->channel = channel;
2272 call->callNumber = &conn->callNumber[channel];
2273 call->rwind = conn->rwind[channel];
2274 call->twind = conn->twind[channel];
2275 /* Note that the next expected call number is retained (in
2276 * conn->callNumber[i]), even if we reallocate the call structure
2278 conn->call[channel] = call;
2279 /* if the channel's never been used (== 0), we should start at 1, otherwise
2280 * the call number is valid from the last time this channel was used */
2281 if (*call->callNumber == 0)
2282 *call->callNumber = 1;
2287 /* A call has been inactive long enough that so we can throw away
2288 * state, including the call structure, which is placed on the call
2290 * Call is locked upon entry.
2291 * haveCTLock set if called from rxi_ReapConnections
2293 #ifdef RX_ENABLE_LOCKS
2295 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2296 #else /* RX_ENABLE_LOCKS */
2298 rxi_FreeCall(struct rx_call *call)
2299 #endif /* RX_ENABLE_LOCKS */
2301 int channel = call->channel;
2302 struct rx_connection *conn = call->conn;
2305 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2306 (*call->callNumber)++;
2307 rxi_ResetCall(call, 0);
2308 call->conn->call[channel] = (struct rx_call *)0;
2310 MUTEX_ENTER(&rx_freeCallQueue_lock);
2311 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2312 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2313 /* A call may be free even though its transmit queue is still in use.
2314 * Since we search the call list from head to tail, put busy calls at
2315 * the head of the list, and idle calls at the tail.
2317 if (call->flags & RX_CALL_TQ_BUSY)
2318 queue_Prepend(&rx_freeCallQueue, call);
2320 queue_Append(&rx_freeCallQueue, call);
2321 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2322 queue_Append(&rx_freeCallQueue, call);
2323 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2324 if (rx_stats_active)
2325 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2326 MUTEX_EXIT(&rx_freeCallQueue_lock);
2328 /* Destroy the connection if it was previously slated for
2329 * destruction, i.e. the Rx client code previously called
2330 * rx_DestroyConnection (client connections), or
2331 * rxi_ReapConnections called the same routine (server
2332 * connections). Only do this, however, if there are no
2333 * outstanding calls. Note that for fine grain locking, there appears
2334 * to be a deadlock in that rxi_FreeCall has a call locked and
2335 * DestroyConnectionNoLock locks each call in the conn. But note a
2336 * few lines up where we have removed this call from the conn.
2337 * If someone else destroys a connection, they either have no
2338 * call lock held or are going through this section of code.
2340 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2341 MUTEX_ENTER(&conn->conn_data_lock);
2343 MUTEX_EXIT(&conn->conn_data_lock);
2344 #ifdef RX_ENABLE_LOCKS
2346 rxi_DestroyConnectionNoLock(conn);
2348 rxi_DestroyConnection(conn);
2349 #else /* RX_ENABLE_LOCKS */
2350 rxi_DestroyConnection(conn);
2351 #endif /* RX_ENABLE_LOCKS */
2355 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2357 rxi_Alloc(size_t size)
2361 if (rx_stats_active)
2362 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2365 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2366 afs_osi_Alloc_NoSleep(size);
2371 osi_Panic("rxi_Alloc error");
2377 rxi_Free(void *addr, size_t size)
2379 if (rx_stats_active)
2380 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2381 osi_Free(addr, size);
2385 rxi_SetPeerMtu(afs_uint32 host, afs_uint32 port, int mtu)
2387 struct rx_peer **peer_ptr, **peer_end;
2390 MUTEX_ENTER(&rx_peerHashTable_lock);
2392 for (peer_ptr = &rx_peerHashTable[0], peer_end =
2393 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
2395 struct rx_peer *peer, *next;
2396 for (peer = *peer_ptr; peer; peer = next) {
2398 if (host == peer->host) {
2399 MUTEX_ENTER(&peer->peer_lock);
2400 peer->ifMTU=MIN(mtu, peer->ifMTU);
2401 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2402 MUTEX_EXIT(&peer->peer_lock);
2407 struct rx_peer *peer;
2408 hashIndex = PEER_HASH(host, port);
2409 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2410 if ((peer->host == host) && (peer->port == port)) {
2411 MUTEX_ENTER(&peer->peer_lock);
2412 peer->ifMTU=MIN(mtu, peer->ifMTU);
2413 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2414 MUTEX_EXIT(&peer->peer_lock);
2418 MUTEX_EXIT(&rx_peerHashTable_lock);
2421 /* Find the peer process represented by the supplied (host,port)
2422 * combination. If there is no appropriate active peer structure, a
2423 * new one will be allocated and initialized
2424 * The origPeer, if set, is a pointer to a peer structure on which the
2425 * refcount will be be decremented. This is used to replace the peer
2426 * structure hanging off a connection structure */
2428 rxi_FindPeer(afs_uint32 host, u_short port,
2429 struct rx_peer *origPeer, int create)
2433 hashIndex = PEER_HASH(host, port);
2434 MUTEX_ENTER(&rx_peerHashTable_lock);
2435 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2436 if ((pp->host == host) && (pp->port == port))
2441 pp = rxi_AllocPeer(); /* This bzero's *pp */
2442 pp->host = host; /* set here or in InitPeerParams is zero */
2444 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2445 queue_Init(&pp->congestionQueue);
2446 queue_Init(&pp->rpcStats);
2447 pp->next = rx_peerHashTable[hashIndex];
2448 rx_peerHashTable[hashIndex] = pp;
2449 rxi_InitPeerParams(pp);
2450 if (rx_stats_active)
2451 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2458 origPeer->refCount--;
2459 MUTEX_EXIT(&rx_peerHashTable_lock);
2464 /* Find the connection at (host, port) started at epoch, and with the
2465 * given connection id. Creates the server connection if necessary.
2466 * The type specifies whether a client connection or a server
2467 * connection is desired. In both cases, (host, port) specify the
2468 * peer's (host, pair) pair. Client connections are not made
2469 * automatically by this routine. The parameter socket gives the
2470 * socket descriptor on which the packet was received. This is used,
2471 * in the case of server connections, to check that *new* connections
2472 * come via a valid (port, serviceId). Finally, the securityIndex
2473 * parameter must match the existing index for the connection. If a
2474 * server connection is created, it will be created using the supplied
2475 * index, if the index is valid for this service */
2476 struct rx_connection *
2477 rxi_FindConnection(osi_socket socket, afs_int32 host,
2478 u_short port, u_short serviceId, afs_uint32 cid,
2479 afs_uint32 epoch, int type, u_int securityIndex)
2481 int hashindex, flag, i;
2482 struct rx_connection *conn;
2483 hashindex = CONN_HASH(host, port, cid, epoch, type);
2484 MUTEX_ENTER(&rx_connHashTable_lock);
2485 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2486 rx_connHashTable[hashindex],
2489 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2490 && (epoch == conn->epoch)) {
2491 struct rx_peer *pp = conn->peer;
2492 if (securityIndex != conn->securityIndex) {
2493 /* this isn't supposed to happen, but someone could forge a packet
2494 * like this, and there seems to be some CM bug that makes this
2495 * happen from time to time -- in which case, the fileserver
2497 MUTEX_EXIT(&rx_connHashTable_lock);
2498 return (struct rx_connection *)0;
2500 if (pp->host == host && pp->port == port)
2502 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2504 /* So what happens when it's a callback connection? */
2505 if ( /*type == RX_CLIENT_CONNECTION && */
2506 (conn->epoch & 0x80000000))
2510 /* the connection rxLastConn that was used the last time is not the
2511 ** one we are looking for now. Hence, start searching in the hash */
2513 conn = rx_connHashTable[hashindex];
2518 struct rx_service *service;
2519 if (type == RX_CLIENT_CONNECTION) {
2520 MUTEX_EXIT(&rx_connHashTable_lock);
2521 return (struct rx_connection *)0;
2523 service = rxi_FindService(socket, serviceId);
2524 if (!service || (securityIndex >= service->nSecurityObjects)
2525 || (service->securityObjects[securityIndex] == 0)) {
2526 MUTEX_EXIT(&rx_connHashTable_lock);
2527 return (struct rx_connection *)0;
2529 conn = rxi_AllocConnection(); /* This bzero's the connection */
2530 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2531 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2532 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2533 conn->next = rx_connHashTable[hashindex];
2534 rx_connHashTable[hashindex] = conn;
2535 conn->peer = rxi_FindPeer(host, port, 0, 1);
2536 conn->type = RX_SERVER_CONNECTION;
2537 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2538 conn->epoch = epoch;
2539 conn->cid = cid & RX_CIDMASK;
2540 /* conn->serial = conn->lastSerial = 0; */
2541 /* conn->timeout = 0; */
2542 conn->ackRate = RX_FAST_ACK_RATE;
2543 conn->service = service;
2544 conn->serviceId = serviceId;
2545 conn->securityIndex = securityIndex;
2546 conn->securityObject = service->securityObjects[securityIndex];
2547 conn->nSpecific = 0;
2548 conn->specific = NULL;
2549 rx_SetConnDeadTime(conn, service->connDeadTime);
2550 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2551 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2552 for (i = 0; i < RX_MAXCALLS; i++) {
2553 conn->twind[i] = rx_initSendWindow;
2554 conn->rwind[i] = rx_initReceiveWindow;
2556 /* Notify security object of the new connection */
2557 RXS_NewConnection(conn->securityObject, conn);
2558 /* XXXX Connection timeout? */
2559 if (service->newConnProc)
2560 (*service->newConnProc) (conn);
2561 if (rx_stats_active)
2562 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2565 MUTEX_ENTER(&conn->conn_data_lock);
2567 MUTEX_EXIT(&conn->conn_data_lock);
2569 rxLastConn = conn; /* store this connection as the last conn used */
2570 MUTEX_EXIT(&rx_connHashTable_lock);
2574 /* There are two packet tracing routines available for testing and monitoring
2575 * Rx. One is called just after every packet is received and the other is
2576 * called just before every packet is sent. Received packets, have had their
2577 * headers decoded, and packets to be sent have not yet had their headers
2578 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2579 * containing the network address. Both can be modified. The return value, if
2580 * non-zero, indicates that the packet should be dropped. */
2582 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2583 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2585 /* A packet has been received off the interface. Np is the packet, socket is
2586 * the socket number it was received from (useful in determining which service
2587 * this packet corresponds to), and (host, port) reflect the host,port of the
2588 * sender. This call returns the packet to the caller if it is finished with
2589 * it, rather than de-allocating it, just as a small performance hack */
2592 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
2593 afs_uint32 host, u_short port, int *tnop,
2594 struct rx_call **newcallp)
2596 struct rx_call *call;
2597 struct rx_connection *conn;
2599 afs_uint32 currentCallNumber;
2605 struct rx_packet *tnp;
2608 /* We don't print out the packet until now because (1) the time may not be
2609 * accurate enough until now in the lwp implementation (rx_Listener only gets
2610 * the time after the packet is read) and (2) from a protocol point of view,
2611 * this is the first time the packet has been seen */
2612 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2613 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2614 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT,
2615 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2616 np->header.epoch, np->header.cid, np->header.callNumber,
2617 np->header.seq, np->header.flags, np));
2620 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2621 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2624 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2625 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2628 /* If an input tracer function is defined, call it with the packet and
2629 * network address. Note this function may modify its arguments. */
2630 if (rx_justReceived) {
2631 struct sockaddr_in addr;
2633 addr.sin_family = AF_INET;
2634 addr.sin_port = port;
2635 addr.sin_addr.s_addr = host;
2636 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2637 addr.sin_len = sizeof(addr);
2638 #endif /* AFS_OSF_ENV */
2639 drop = (*rx_justReceived) (np, &addr);
2640 /* drop packet if return value is non-zero */
2643 port = addr.sin_port; /* in case fcn changed addr */
2644 host = addr.sin_addr.s_addr;
2648 /* If packet was not sent by the client, then *we* must be the client */
2649 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2650 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2652 /* Find the connection (or fabricate one, if we're the server & if
2653 * necessary) associated with this packet */
2655 rxi_FindConnection(socket, host, port, np->header.serviceId,
2656 np->header.cid, np->header.epoch, type,
2657 np->header.securityIndex);
2660 /* If no connection found or fabricated, just ignore the packet.
2661 * (An argument could be made for sending an abort packet for
2666 MUTEX_ENTER(&conn->conn_data_lock);
2667 if (conn->maxSerial < np->header.serial)
2668 conn->maxSerial = np->header.serial;
2669 MUTEX_EXIT(&conn->conn_data_lock);
2671 /* If the connection is in an error state, send an abort packet and ignore
2672 * the incoming packet */
2674 /* Don't respond to an abort packet--we don't want loops! */
2675 MUTEX_ENTER(&conn->conn_data_lock);
2676 if (np->header.type != RX_PACKET_TYPE_ABORT)
2677 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2679 MUTEX_EXIT(&conn->conn_data_lock);
2683 /* Check for connection-only requests (i.e. not call specific). */
2684 if (np->header.callNumber == 0) {
2685 switch (np->header.type) {
2686 case RX_PACKET_TYPE_ABORT: {
2687 /* What if the supplied error is zero? */
2688 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2689 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2690 rxi_ConnectionError(conn, errcode);
2691 MUTEX_ENTER(&conn->conn_data_lock);
2693 MUTEX_EXIT(&conn->conn_data_lock);
2696 case RX_PACKET_TYPE_CHALLENGE:
2697 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2698 MUTEX_ENTER(&conn->conn_data_lock);
2700 MUTEX_EXIT(&conn->conn_data_lock);
2702 case RX_PACKET_TYPE_RESPONSE:
2703 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2704 MUTEX_ENTER(&conn->conn_data_lock);
2706 MUTEX_EXIT(&conn->conn_data_lock);
2708 case RX_PACKET_TYPE_PARAMS:
2709 case RX_PACKET_TYPE_PARAMS + 1:
2710 case RX_PACKET_TYPE_PARAMS + 2:
2711 /* ignore these packet types for now */
2712 MUTEX_ENTER(&conn->conn_data_lock);
2714 MUTEX_EXIT(&conn->conn_data_lock);
2719 /* Should not reach here, unless the peer is broken: send an
2721 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2722 MUTEX_ENTER(&conn->conn_data_lock);
2723 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2725 MUTEX_EXIT(&conn->conn_data_lock);
2730 channel = np->header.cid & RX_CHANNELMASK;
2731 call = conn->call[channel];
2732 #ifdef RX_ENABLE_LOCKS
2734 MUTEX_ENTER(&call->lock);
2735 /* Test to see if call struct is still attached to conn. */
2736 if (call != conn->call[channel]) {
2738 MUTEX_EXIT(&call->lock);
2739 if (type == RX_SERVER_CONNECTION) {
2740 call = conn->call[channel];
2741 /* If we started with no call attached and there is one now,
2742 * another thread is also running this routine and has gotten
2743 * the connection channel. We should drop this packet in the tests
2744 * below. If there was a call on this connection and it's now
2745 * gone, then we'll be making a new call below.
2746 * If there was previously a call and it's now different then
2747 * the old call was freed and another thread running this routine
2748 * has created a call on this channel. One of these two threads
2749 * has a packet for the old call and the code below handles those
2753 MUTEX_ENTER(&call->lock);
2755 /* This packet can't be for this call. If the new call address is
2756 * 0 then no call is running on this channel. If there is a call
2757 * then, since this is a client connection we're getting data for
2758 * it must be for the previous call.
2760 if (rx_stats_active)
2761 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2762 MUTEX_ENTER(&conn->conn_data_lock);
2764 MUTEX_EXIT(&conn->conn_data_lock);
2769 currentCallNumber = conn->callNumber[channel];
2771 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2772 if (np->header.callNumber < currentCallNumber) {
2773 if (rx_stats_active)
2774 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2775 #ifdef RX_ENABLE_LOCKS
2777 MUTEX_EXIT(&call->lock);
2779 MUTEX_ENTER(&conn->conn_data_lock);
2781 MUTEX_EXIT(&conn->conn_data_lock);
2785 MUTEX_ENTER(&conn->conn_call_lock);
2786 call = rxi_NewCall(conn, channel);
2787 MUTEX_EXIT(&conn->conn_call_lock);
2788 *call->callNumber = np->header.callNumber;
2790 if (np->header.callNumber == 0)
2791 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%.06d len %d",
2792 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
2793 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
2794 np->header.flags, np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2796 call->state = RX_STATE_PRECALL;
2797 clock_GetTime(&call->queueTime);
2798 hzero(call->bytesSent);
2799 hzero(call->bytesRcvd);
2801 * If the number of queued calls exceeds the overload
2802 * threshold then abort this call.
2804 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2805 struct rx_packet *tp;
2807 rxi_CallError(call, rx_BusyError);
2808 tp = rxi_SendCallAbort(call, np, 1, 0);
2809 MUTEX_EXIT(&call->lock);
2810 MUTEX_ENTER(&conn->conn_data_lock);
2812 MUTEX_EXIT(&conn->conn_data_lock);
2813 if (rx_stats_active)
2814 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2817 rxi_KeepAliveOn(call);
2818 } else if (np->header.callNumber != currentCallNumber) {
2819 /* Wait until the transmit queue is idle before deciding
2820 * whether to reset the current call. Chances are that the
2821 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2824 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2825 while ((call->state == RX_STATE_ACTIVE)
2826 && (call->flags & RX_CALL_TQ_BUSY)) {
2827 call->flags |= RX_CALL_TQ_WAIT;
2829 #ifdef RX_ENABLE_LOCKS
2830 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2831 CV_WAIT(&call->cv_tq, &call->lock);
2832 #else /* RX_ENABLE_LOCKS */
2833 osi_rxSleep(&call->tq);
2834 #endif /* RX_ENABLE_LOCKS */
2836 if (call->tqWaiters == 0)
2837 call->flags &= ~RX_CALL_TQ_WAIT;
2839 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2840 /* If the new call cannot be taken right now send a busy and set
2841 * the error condition in this call, so that it terminates as
2842 * quickly as possible */
2843 if (call->state == RX_STATE_ACTIVE) {
2844 struct rx_packet *tp;
2846 rxi_CallError(call, RX_CALL_DEAD);
2847 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2849 MUTEX_EXIT(&call->lock);
2850 MUTEX_ENTER(&conn->conn_data_lock);
2852 MUTEX_EXIT(&conn->conn_data_lock);
2855 rxi_ResetCall(call, 0);
2856 *call->callNumber = np->header.callNumber;
2858 if (np->header.callNumber == 0)
2859 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%06d len %d",
2860 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
2861 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
2862 np->header.flags, np, np->retryTime.sec, np->retryTime.usec, np->length));
2864 call->state = RX_STATE_PRECALL;
2865 clock_GetTime(&call->queueTime);
2866 hzero(call->bytesSent);
2867 hzero(call->bytesRcvd);
2869 * If the number of queued calls exceeds the overload
2870 * threshold then abort this call.
2872 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2873 struct rx_packet *tp;
2875 rxi_CallError(call, rx_BusyError);
2876 tp = rxi_SendCallAbort(call, np, 1, 0);
2877 MUTEX_EXIT(&call->lock);
2878 MUTEX_ENTER(&conn->conn_data_lock);
2880 MUTEX_EXIT(&conn->conn_data_lock);
2881 if (rx_stats_active)
2882 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2885 rxi_KeepAliveOn(call);
2887 /* Continuing call; do nothing here. */
2889 } else { /* we're the client */
2890 /* Ignore all incoming acknowledgements for calls in DALLY state */
2891 if (call && (call->state == RX_STATE_DALLY)
2892 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2893 if (rx_stats_active)
2894 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
2895 #ifdef RX_ENABLE_LOCKS
2897 MUTEX_EXIT(&call->lock);
2900 MUTEX_ENTER(&conn->conn_data_lock);
2902 MUTEX_EXIT(&conn->conn_data_lock);
2906 /* Ignore anything that's not relevant to the current call. If there
2907 * isn't a current call, then no packet is relevant. */
2908 if (!call || (np->header.callNumber != currentCallNumber)) {
2909 if (rx_stats_active)
2910 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2911 #ifdef RX_ENABLE_LOCKS
2913 MUTEX_EXIT(&call->lock);
2916 MUTEX_ENTER(&conn->conn_data_lock);
2918 MUTEX_EXIT(&conn->conn_data_lock);
2921 /* If the service security object index stamped in the packet does not
2922 * match the connection's security index, ignore the packet */
2923 if (np->header.securityIndex != conn->securityIndex) {
2924 #ifdef RX_ENABLE_LOCKS
2925 MUTEX_EXIT(&call->lock);
2927 MUTEX_ENTER(&conn->conn_data_lock);
2929 MUTEX_EXIT(&conn->conn_data_lock);
2933 /* If we're receiving the response, then all transmit packets are
2934 * implicitly acknowledged. Get rid of them. */
2935 if (np->header.type == RX_PACKET_TYPE_DATA) {
2936 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2937 /* XXX Hack. Because we must release the global rx lock when
2938 * sending packets (osi_NetSend) we drop all acks while we're
2939 * traversing the tq in rxi_Start sending packets out because
2940 * packets may move to the freePacketQueue as result of being here!
2941 * So we drop these packets until we're safely out of the
2942 * traversing. Really ugly!
2943 * For fine grain RX locking, we set the acked field in the
2944 * packets and let rxi_Start remove them from the transmit queue.
2946 if (call->flags & RX_CALL_TQ_BUSY) {
2947 #ifdef RX_ENABLE_LOCKS
2948 rxi_SetAcksInTransmitQueue(call);
2951 return np; /* xmitting; drop packet */
2954 rxi_ClearTransmitQueue(call, 0);
2956 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2957 rxi_ClearTransmitQueue(call, 0);
2958 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2960 if (np->header.type == RX_PACKET_TYPE_ACK) {
2961 /* now check to see if this is an ack packet acknowledging that the
2962 * server actually *lost* some hard-acked data. If this happens we
2963 * ignore this packet, as it may indicate that the server restarted in
2964 * the middle of a call. It is also possible that this is an old ack
2965 * packet. We don't abort the connection in this case, because this
2966 * *might* just be an old ack packet. The right way to detect a server
2967 * restart in the midst of a call is to notice that the server epoch
2969 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2970 * XXX unacknowledged. I think that this is off-by-one, but
2971 * XXX I don't dare change it just yet, since it will
2972 * XXX interact badly with the server-restart detection
2973 * XXX code in receiveackpacket. */
2974 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2975 if (rx_stats_active)
2976 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2977 MUTEX_EXIT(&call->lock);
2978 MUTEX_ENTER(&conn->conn_data_lock);
2980 MUTEX_EXIT(&conn->conn_data_lock);
2984 } /* else not a data packet */
2987 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2988 /* Set remote user defined status from packet */
2989 call->remoteStatus = np->header.userStatus;
2991 /* Note the gap between the expected next packet and the actual
2992 * packet that arrived, when the new packet has a smaller serial number
2993 * than expected. Rioses frequently reorder packets all by themselves,
2994 * so this will be quite important with very large window sizes.
2995 * Skew is checked against 0 here to avoid any dependence on the type of
2996 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2998 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2999 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3000 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3002 MUTEX_ENTER(&conn->conn_data_lock);
3003 skew = conn->lastSerial - np->header.serial;
3004 conn->lastSerial = np->header.serial;
3005 MUTEX_EXIT(&conn->conn_data_lock);
3007 struct rx_peer *peer;
3009 if (skew > peer->inPacketSkew) {
3010 dpf(("*** In skew changed from %d to %d\n",
3011 peer->inPacketSkew, skew));
3012 peer->inPacketSkew = skew;
3016 /* Now do packet type-specific processing */
3017 switch (np->header.type) {
3018 case RX_PACKET_TYPE_DATA:
3019 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3022 case RX_PACKET_TYPE_ACK:
3023 /* Respond immediately to ack packets requesting acknowledgement
3025 if (np->header.flags & RX_REQUEST_ACK) {
3027 (void)rxi_SendCallAbort(call, 0, 1, 0);
3029 (void)rxi_SendAck(call, 0, np->header.serial,
3030 RX_ACK_PING_RESPONSE, 1);
3032 np = rxi_ReceiveAckPacket(call, np, 1);
3034 case RX_PACKET_TYPE_ABORT: {
3035 /* An abort packet: reset the call, passing the error up to the user. */
3036 /* What if error is zero? */
3037 /* What if the error is -1? the application will treat it as a timeout. */
3038 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3039 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
3040 rxi_CallError(call, errdata);
3041 MUTEX_EXIT(&call->lock);
3042 MUTEX_ENTER(&conn->conn_data_lock);
3044 MUTEX_EXIT(&conn->conn_data_lock);
3045 return np; /* xmitting; drop packet */
3047 case RX_PACKET_TYPE_BUSY:
3050 case RX_PACKET_TYPE_ACKALL:
3051 /* All packets acknowledged, so we can drop all packets previously
3052 * readied for sending */
3053 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3054 /* XXX Hack. We because we can't release the global rx lock when
3055 * sending packets (osi_NetSend) we drop all ack pkts while we're
3056 * traversing the tq in rxi_Start sending packets out because
3057 * packets may move to the freePacketQueue as result of being
3058 * here! So we drop these packets until we're safely out of the
3059 * traversing. Really ugly!
3060 * For fine grain RX locking, we set the acked field in the packets
3061 * and let rxi_Start remove the packets from the transmit queue.
3063 if (call->flags & RX_CALL_TQ_BUSY) {
3064 #ifdef RX_ENABLE_LOCKS
3065 rxi_SetAcksInTransmitQueue(call);
3067 #else /* RX_ENABLE_LOCKS */
3068 MUTEX_EXIT(&call->lock);
3069 MUTEX_ENTER(&conn->conn_data_lock);
3071 MUTEX_EXIT(&conn->conn_data_lock);
3072 return np; /* xmitting; drop packet */
3073 #endif /* RX_ENABLE_LOCKS */
3075 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3076 rxi_ClearTransmitQueue(call, 0);
3077 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3080 /* Should not reach here, unless the peer is broken: send an abort
3082 rxi_CallError(call, RX_PROTOCOL_ERROR);
3083 np = rxi_SendCallAbort(call, np, 1, 0);
3086 /* Note when this last legitimate packet was received, for keep-alive
3087 * processing. Note, we delay getting the time until now in the hope that
3088 * the packet will be delivered to the user before any get time is required
3089 * (if not, then the time won't actually be re-evaluated here). */
3090 call->lastReceiveTime = clock_Sec();
3091 MUTEX_EXIT(&call->lock);
3092 MUTEX_ENTER(&conn->conn_data_lock);
3094 MUTEX_EXIT(&conn->conn_data_lock);
3098 /* return true if this is an "interesting" connection from the point of view
3099 of someone trying to debug the system */
3101 rxi_IsConnInteresting(struct rx_connection *aconn)
3104 struct rx_call *tcall;
3106 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3108 for (i = 0; i < RX_MAXCALLS; i++) {
3109 tcall = aconn->call[i];
3111 if ((tcall->state == RX_STATE_PRECALL)
3112 || (tcall->state == RX_STATE_ACTIVE))
3114 if ((tcall->mode == RX_MODE_SENDING)
3115 || (tcall->mode == RX_MODE_RECEIVING))
3123 /* if this is one of the last few packets AND it wouldn't be used by the
3124 receiving call to immediately satisfy a read request, then drop it on
3125 the floor, since accepting it might prevent a lock-holding thread from
3126 making progress in its reading. If a call has been cleared while in
3127 the precall state then ignore all subsequent packets until the call
3128 is assigned to a thread. */
3131 TooLow(struct rx_packet *ap, struct rx_call *acall)
3135 MUTEX_ENTER(&rx_quota_mutex);
3136 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3137 && (acall->state == RX_STATE_PRECALL))
3138 || ((rx_nFreePackets < rxi_dataQuota + 2)
3139 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3140 && (acall->flags & RX_CALL_READER_WAIT)))) {
3143 MUTEX_EXIT(&rx_quota_mutex);
3149 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3151 struct rx_connection *conn = arg1;
3152 struct rx_call *acall = arg2;
3153 struct rx_call *call = acall;
3154 struct clock when, now;
3157 MUTEX_ENTER(&conn->conn_data_lock);
3158 conn->checkReachEvent = NULL;
3159 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3162 MUTEX_EXIT(&conn->conn_data_lock);
3166 MUTEX_ENTER(&conn->conn_call_lock);
3167 MUTEX_ENTER(&conn->conn_data_lock);
3168 for (i = 0; i < RX_MAXCALLS; i++) {
3169 struct rx_call *tc = conn->call[i];
3170 if (tc && tc->state == RX_STATE_PRECALL) {
3176 /* Indicate that rxi_CheckReachEvent is no longer running by
3177 * clearing the flag. Must be atomic under conn_data_lock to
3178 * avoid a new call slipping by: rxi_CheckConnReach holds
3179 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3181 conn->flags &= ~RX_CONN_ATTACHWAIT;
3182 MUTEX_EXIT(&conn->conn_data_lock);
3183 MUTEX_EXIT(&conn->conn_call_lock);
3188 MUTEX_ENTER(&call->lock);
3189 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3191 MUTEX_EXIT(&call->lock);
3193 clock_GetTime(&now);
3195 when.sec += RX_CHECKREACH_TIMEOUT;
3196 MUTEX_ENTER(&conn->conn_data_lock);
3197 if (!conn->checkReachEvent) {
3199 conn->checkReachEvent =
3200 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3203 MUTEX_EXIT(&conn->conn_data_lock);
3209 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3211 struct rx_service *service = conn->service;
3212 struct rx_peer *peer = conn->peer;
3213 afs_uint32 now, lastReach;
3215 if (service->checkReach == 0)
3219 MUTEX_ENTER(&peer->peer_lock);
3220 lastReach = peer->lastReachTime;
3221 MUTEX_EXIT(&peer->peer_lock);
3222 if (now - lastReach < RX_CHECKREACH_TTL)
3225 MUTEX_ENTER(&conn->conn_data_lock);
3226 if (conn->flags & RX_CONN_ATTACHWAIT) {
3227 MUTEX_EXIT(&conn->conn_data_lock);
3230 conn->flags |= RX_CONN_ATTACHWAIT;
3231 MUTEX_EXIT(&conn->conn_data_lock);
3232 if (!conn->checkReachEvent)
3233 rxi_CheckReachEvent(NULL, conn, call);
3238 /* try to attach call, if authentication is complete */
3240 TryAttach(struct rx_call *acall, osi_socket socket,
3241 int *tnop, struct rx_call **newcallp,
3244 struct rx_connection *conn = acall->conn;
3246 if (conn->type == RX_SERVER_CONNECTION
3247 && acall->state == RX_STATE_PRECALL) {
3248 /* Don't attach until we have any req'd. authentication. */
3249 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3250 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3251 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3252 /* Note: this does not necessarily succeed; there
3253 * may not any proc available
3256 rxi_ChallengeOn(acall->conn);
3261 /* A data packet has been received off the interface. This packet is
3262 * appropriate to the call (the call is in the right state, etc.). This
3263 * routine can return a packet to the caller, for re-use */
3266 rxi_ReceiveDataPacket(struct rx_call *call,
3267 struct rx_packet *np, int istack,
3268 osi_socket socket, afs_uint32 host, u_short port,
3269 int *tnop, struct rx_call **newcallp)
3271 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3276 afs_uint32 serial=0, flags=0;
3278 struct rx_packet *tnp;
3279 struct clock when, now;
3280 if (rx_stats_active)
3281 rx_MutexIncrement(rx_stats.dataPacketsRead, rx_stats_mutex);
3284 /* If there are no packet buffers, drop this new packet, unless we can find
3285 * packet buffers from inactive calls */
3287 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3288 MUTEX_ENTER(&rx_freePktQ_lock);
3289 rxi_NeedMorePackets = TRUE;
3290 MUTEX_EXIT(&rx_freePktQ_lock);
3291 if (rx_stats_active)
3292 rx_MutexIncrement(rx_stats.noPacketBuffersOnRead, rx_stats_mutex);
3293 call->rprev = np->header.serial;
3294 rxi_calltrace(RX_TRACE_DROP, call);
3295 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems", np));
3297 rxi_ClearReceiveQueue(call);
3298 clock_GetTime(&now);
3300 clock_Add(&when, &rx_softAckDelay);
3301 if (!call->delayedAckEvent
3302 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3303 rxevent_Cancel(call->delayedAckEvent, call,
3304 RX_CALL_REFCOUNT_DELAY);
3305 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3306 call->delayedAckEvent =
3307 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3309 /* we've damaged this call already, might as well do it in. */
3315 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3316 * packet is one of several packets transmitted as a single
3317 * datagram. Do not send any soft or hard acks until all packets
3318 * in a jumbogram have been processed. Send negative acks right away.
3320 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3321 /* tnp is non-null when there are more packets in the
3322 * current jumbo gram */
3329 seq = np->header.seq;
3330 serial = np->header.serial;
3331 flags = np->header.flags;
3333 /* If the call is in an error state, send an abort message */
3335 return rxi_SendCallAbort(call, np, istack, 0);
3337 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3338 * AFS 3.5 jumbogram. */
3339 if (flags & RX_JUMBO_PACKET) {
3340 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3345 if (np->header.spare != 0) {
3346 MUTEX_ENTER(&call->conn->conn_data_lock);
3347 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3348 MUTEX_EXIT(&call->conn->conn_data_lock);
3351 /* The usual case is that this is the expected next packet */
3352 if (seq == call->rnext) {
3354 /* Check to make sure it is not a duplicate of one already queued */
3355 if (queue_IsNotEmpty(&call->rq)
3356 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3357 if (rx_stats_active)
3358 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3359 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate", np));
3360 rxevent_Cancel(call->delayedAckEvent, call,
3361 RX_CALL_REFCOUNT_DELAY);
3362 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3368 /* It's the next packet. Stick it on the receive queue
3369 * for this call. Set newPackets to make sure we wake
3370 * the reader once all packets have been processed */
3371 np->flags |= RX_PKTFLAG_RQ;
3372 queue_Prepend(&call->rq, np);
3373 #ifdef RXDEBUG_PACKET
3375 #endif /* RXDEBUG_PACKET */
3377 np = NULL; /* We can't use this anymore */
3380 /* If an ack is requested then set a flag to make sure we
3381 * send an acknowledgement for this packet */
3382 if (flags & RX_REQUEST_ACK) {
3383 ackNeeded = RX_ACK_REQUESTED;
3386 /* Keep track of whether we have received the last packet */
3387 if (flags & RX_LAST_PACKET) {
3388 call->flags |= RX_CALL_HAVE_LAST;
3392 /* Check whether we have all of the packets for this call */
3393 if (call->flags & RX_CALL_HAVE_LAST) {
3394 afs_uint32 tseq; /* temporary sequence number */
3395 struct rx_packet *tp; /* Temporary packet pointer */
3396 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3398 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3399 if (tseq != tp->header.seq)
3401 if (tp->header.flags & RX_LAST_PACKET) {
3402 call->flags |= RX_CALL_RECEIVE_DONE;
3409 /* Provide asynchronous notification for those who want it
3410 * (e.g. multi rx) */
3411 if (call->arrivalProc) {
3412 (*call->arrivalProc) (call, call->arrivalProcHandle,
3413 call->arrivalProcArg);
3414 call->arrivalProc = (void (*)())0;
3417 /* Update last packet received */
3420 /* If there is no server process serving this call, grab
3421 * one, if available. We only need to do this once. If a
3422 * server thread is available, this thread becomes a server
3423 * thread and the server thread becomes a listener thread. */
3425 TryAttach(call, socket, tnop, newcallp, 0);
3428 /* This is not the expected next packet. */
3430 /* Determine whether this is a new or old packet, and if it's
3431 * a new one, whether it fits into the current receive window.
3432 * Also figure out whether the packet was delivered in sequence.
3433 * We use the prev variable to determine whether the new packet
3434 * is the successor of its immediate predecessor in the
3435 * receive queue, and the missing flag to determine whether
3436 * any of this packets predecessors are missing. */
3438 afs_uint32 prev; /* "Previous packet" sequence number */
3439 struct rx_packet *tp; /* Temporary packet pointer */
3440 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3441 int missing; /* Are any predecessors missing? */
3443 /* If the new packet's sequence number has been sent to the
3444 * application already, then this is a duplicate */
3445 if (seq < call->rnext) {
3446 if (rx_stats_active)
3447 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3448 rxevent_Cancel(call->delayedAckEvent, call,
3449 RX_CALL_REFCOUNT_DELAY);
3450 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3456 /* If the sequence number is greater than what can be
3457 * accomodated by the current window, then send a negative
3458 * acknowledge and drop the packet */
3459 if ((call->rnext + call->rwind) <= seq) {
3460 rxevent_Cancel(call->delayedAckEvent, call,
3461 RX_CALL_REFCOUNT_DELAY);
3462 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3469 /* Look for the packet in the queue of old received packets */
3470 for (prev = call->rnext - 1, missing =
3471 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3472 /*Check for duplicate packet */
3473 if (seq == tp->header.seq) {
3474 if (rx_stats_active)
3475 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3476 rxevent_Cancel(call->delayedAckEvent, call,
3477 RX_CALL_REFCOUNT_DELAY);
3478 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3484 /* If we find a higher sequence packet, break out and
3485 * insert the new packet here. */
3486 if (seq < tp->header.seq)
3488 /* Check for missing packet */
3489 if (tp->header.seq != prev + 1) {
3493 prev = tp->header.seq;
3496 /* Keep track of whether we have received the last packet. */
3497 if (flags & RX_LAST_PACKET) {
3498 call->flags |= RX_CALL_HAVE_LAST;
3501 /* It's within the window: add it to the the receive queue.
3502 * tp is left by the previous loop either pointing at the
3503 * packet before which to insert the new packet, or at the
3504 * queue head if the queue is empty or the packet should be
3506 np->flags |= RX_PKTFLAG_RQ;
3507 #ifdef RXDEBUG_PACKET
3509 #endif /* RXDEBUG_PACKET */
3510 queue_InsertBefore(tp, np);
3514 /* Check whether we have all of the packets for this call */
3515 if ((call->flags & RX_CALL_HAVE_LAST)
3516 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3517 afs_uint32 tseq; /* temporary sequence number */
3520 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3521 if (tseq != tp->header.seq)
3523 if (tp->header.flags & RX_LAST_PACKET) {
3524 call->flags |= RX_CALL_RECEIVE_DONE;
3531 /* We need to send an ack of the packet is out of sequence,
3532 * or if an ack was requested by the peer. */
3533 if (seq != prev + 1 || missing) {
3534 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3535 } else if (flags & RX_REQUEST_ACK) {
3536 ackNeeded = RX_ACK_REQUESTED;
3539 /* Acknowledge the last packet for each call */
3540 if (flags & RX_LAST_PACKET) {
3551 * If the receiver is waiting for an iovec, fill the iovec
3552 * using the data from the receive queue */
3553 if (call->flags & RX_CALL_IOVEC_WAIT) {
3554 didHardAck = rxi_FillReadVec(call, serial);
3555 /* the call may have been aborted */
3564 /* Wakeup the reader if any */
3565 if ((call->flags & RX_CALL_READER_WAIT)
3566 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3567 || (call->iovNext >= call->iovMax)
3568 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3569 call->flags &= ~RX_CALL_READER_WAIT;
3570 #ifdef RX_ENABLE_LOCKS
3571 CV_BROADCAST(&call->cv_rq);
3573 osi_rxWakeup(&call->rq);
3579 * Send an ack when requested by the peer, or once every
3580 * rxi_SoftAckRate packets until the last packet has been
3581 * received. Always send a soft ack for the last packet in
3582 * the server's reply. */
3584 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3585 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3586 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3587 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3588 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3589 } else if (call->nSoftAcks) {
3590 clock_GetTime(&now);
3592 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3593 clock_Add(&when, &rx_lastAckDelay);
3595 clock_Add(&when, &rx_softAckDelay);
3597 if (!call->delayedAckEvent
3598 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3599 rxevent_Cancel(call->delayedAckEvent, call,
3600 RX_CALL_REFCOUNT_DELAY);
3601 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3602 call->delayedAckEvent =
3603 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3605 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3606 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3613 static void rxi_ComputeRate();
3617 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3619 struct rx_peer *peer = conn->peer;
3621 MUTEX_ENTER(&peer->peer_lock);
3622 peer->lastReachTime = clock_Sec();
3623 MUTEX_EXIT(&peer->peer_lock);
3625 MUTEX_ENTER(&conn->conn_data_lock);
3626 if (conn->flags & RX_CONN_ATTACHWAIT) {
3629 conn->flags &= ~RX_CONN_ATTACHWAIT;
3630 MUTEX_EXIT(&conn->conn_data_lock);
3632 for (i = 0; i < RX_MAXCALLS; i++) {
3633 struct rx_call *call = conn->call[i];
3636 MUTEX_ENTER(&call->lock);
3637 /* tnop can be null if newcallp is null */
3638 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3640 MUTEX_EXIT(&call->lock);
3644 MUTEX_EXIT(&conn->conn_data_lock);
3647 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
3649 rx_ack_reason(int reason)
3652 case RX_ACK_REQUESTED:
3654 case RX_ACK_DUPLICATE:
3656 case RX_ACK_OUT_OF_SEQUENCE:
3658 case RX_ACK_EXCEEDS_WINDOW:
3660 case RX_ACK_NOSPACE:
3664 case RX_ACK_PING_RESPONSE:
3677 /* rxi_ComputePeerNetStats
3679 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3680 * estimates (like RTT and throughput) based on ack packets. Caller
3681 * must ensure that the packet in question is the right one (i.e.
3682 * serial number matches).
3685 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3686 struct rx_ackPacket *ap, struct rx_packet *np)
3688 struct rx_peer *peer = call->conn->peer;
3690 /* Use RTT if not delayed by client and
3691 * ignore packets that were retransmitted. */
3692 if (!(p->flags & RX_PKTFLAG_ACKED) &&
3693 ap->reason != RX_ACK_DELAY &&
3694 clock_Eq(&p->timeSent, &p->firstSent))
3695 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3697 rxi_ComputeRate(peer, call, p, np, ap->reason);
3701 /* The real smarts of the whole thing. */
3703 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
3706 struct rx_ackPacket *ap;
3708 struct rx_packet *tp;
3709 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3710 struct rx_connection *conn = call->conn;
3711 struct rx_peer *peer = conn->peer;
3714 /* because there are CM's that are bogus, sending weird values for this. */
3715 afs_uint32 skew = 0;
3721 int newAckCount = 0;
3722 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3723 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3725 if (rx_stats_active)
3726 rx_MutexIncrement(rx_stats.ackPacketsRead, rx_stats_mutex);
3727 ap = (struct rx_ackPacket *)rx_DataOf(np);
3728 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3730 return np; /* truncated ack packet */
3732 /* depends on ack packet struct */
3733 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3734 first = ntohl(ap->firstPacket);
3735 serial = ntohl(ap->serial);
3736 /* temporarily disabled -- needs to degrade over time
3737 * skew = ntohs(ap->maxSkew); */
3739 /* Ignore ack packets received out of order */
3740 if (first < call->tfirst) {
3744 if (np->header.flags & RX_SLOW_START_OK) {
3745 call->flags |= RX_CALL_SLOW_START_OK;
3748 if (ap->reason == RX_ACK_PING_RESPONSE)
3749 rxi_UpdatePeerReach(conn, call);
3753 if (rxdebug_active) {
3757 len = _snprintf(msg, sizeof(msg),
3758 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3759 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3760 ntohl(ap->serial), ntohl(ap->previousPacket),
3761 (unsigned int)np->header.seq, (unsigned int)skew,
3762 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3766 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3767 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3771 OutputDebugString(msg);
3773 #else /* AFS_NT40_ENV */
3776 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3777 ap->reason, ntohl(ap->previousPacket),
3778 (unsigned int)np->header.seq, (unsigned int)serial,
3779 (unsigned int)skew, ntohl(ap->firstPacket));
3782 for (offset = 0; offset < nAcks; offset++)
3783 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3788 #endif /* AFS_NT40_ENV */
3791 /* Update the outgoing packet skew value to the latest value of
3792 * the peer's incoming packet skew value. The ack packet, of
3793 * course, could arrive out of order, but that won't affect things
3795 MUTEX_ENTER(&peer->peer_lock);
3796 peer->outPacketSkew = skew;
3798 /* Check for packets that no longer need to be transmitted, and
3799 * discard them. This only applies to packets positively
3800 * acknowledged as having been sent to the peer's upper level.
3801 * All other packets must be retained. So only packets with
3802 * sequence numbers < ap->firstPacket are candidates. */
3803 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3804 if (tp->header.seq >= first)
3806 call->tfirst = tp->header.seq + 1;
3807 rxi_ComputePeerNetStats(call, tp, ap, np);
3808 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3811 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3812 /* XXX Hack. Because we have to release the global rx lock when sending
3813 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3814 * in rxi_Start sending packets out because packets may move to the
3815 * freePacketQueue as result of being here! So we drop these packets until
3816 * we're safely out of the traversing. Really ugly!
3817 * To make it even uglier, if we're using fine grain locking, we can
3818 * set the ack bits in the packets and have rxi_Start remove the packets
3819 * when it's done transmitting.
3821 if (call->flags & RX_CALL_TQ_BUSY) {
3822 #ifdef RX_ENABLE_LOCKS
3823 tp->flags |= RX_PKTFLAG_ACKED;
3824 call->flags |= RX_CALL_TQ_SOME_ACKED;
3825 #else /* RX_ENABLE_LOCKS */
3827 #endif /* RX_ENABLE_LOCKS */
3829 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3832 tp->flags &= ~RX_PKTFLAG_TQ;
3833 #ifdef RXDEBUG_PACKET
3835 #endif /* RXDEBUG_PACKET */
3836 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3841 /* Give rate detector a chance to respond to ping requests */
3842 if (ap->reason == RX_ACK_PING_RESPONSE) {
3843 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3847 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3849 /* Now go through explicit acks/nacks and record the results in
3850 * the waiting packets. These are packets that can't be released
3851 * yet, even with a positive acknowledge. This positive
3852 * acknowledge only means the packet has been received by the
3853 * peer, not that it will be retained long enough to be sent to
3854 * the peer's upper level. In addition, reset the transmit timers
3855 * of any missing packets (those packets that must be missing
3856 * because this packet was out of sequence) */
3858 call->nSoftAcked = 0;
3859 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3860 /* Update round trip time if the ack was stimulated on receipt
3862 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3863 #ifdef RX_ENABLE_LOCKS
3864 if (tp->header.seq >= first)
3865 #endif /* RX_ENABLE_LOCKS */
3866 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3867 rxi_ComputePeerNetStats(call, tp, ap, np);
3869 /* Set the acknowledge flag per packet based on the
3870 * information in the ack packet. An acknowlegded packet can
3871 * be downgraded when the server has discarded a packet it
3872 * soacked previously, or when an ack packet is received
3873 * out of sequence. */
3874 if (tp->header.seq < first) {
3875 /* Implicit ack information */
3876 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3879 tp->flags |= RX_PKTFLAG_ACKED;
3880 } else if (tp->header.seq < first + nAcks) {
3881 /* Explicit ack information: set it in the packet appropriately */
3882 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3883 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3885 tp->flags |= RX_PKTFLAG_ACKED;
3892 } else /* RX_ACK_TYPE_NACK */ {
3893 tp->flags &= ~RX_PKTFLAG_ACKED;
3897 tp->flags &= ~RX_PKTFLAG_ACKED;
3902 * Following the suggestion of Phil Kern, we back off the peer's
3903 * timeout value for future packets until a successful response
3904 * is received for an initial transmission.
3906 if (missing && !backedOff) {
3907 struct clock c = peer->timeout;
3908 struct clock max_to = {3, 0};
3910 clock_Add(&peer->timeout, &c);
3911 if (clock_Gt(&peer->timeout, &max_to))
3912 peer->timeout = max_to;
3916 /* If packet isn't yet acked, and it has been transmitted at least
3917 * once, reset retransmit time using latest timeout
3918 * ie, this should readjust the retransmit timer for all outstanding
3919 * packets... So we don't just retransmit when we should know better*/
3921 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3922 tp->retryTime = tp->timeSent;
3923 clock_Add(&tp->retryTime, &peer->timeout);
3924 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3925 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3929 /* If the window has been extended by this acknowledge packet,
3930 * then wakeup a sender waiting in alloc for window space, or try
3931 * sending packets now, if he's been sitting on packets due to
3932 * lack of window space */
3933 if (call->tnext < (call->tfirst + call->twind)) {
3934 #ifdef RX_ENABLE_LOCKS
3935 CV_SIGNAL(&call->cv_twind);
3937 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3938 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3939 osi_rxWakeup(&call->twind);
3942 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3943 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3947 /* if the ack packet has a receivelen field hanging off it,
3948 * update our state */
3949 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3952 /* If the ack packet has a "recommended" size that is less than
3953 * what I am using now, reduce my size to match */
3954 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
3955 (int)sizeof(afs_int32), &tSize);
3956 tSize = (afs_uint32) ntohl(tSize);
3957 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3959 /* Get the maximum packet size to send to this peer */
3960 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3962 tSize = (afs_uint32) ntohl(tSize);
3963 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3964 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3966 /* sanity check - peer might have restarted with different params.
3967 * If peer says "send less", dammit, send less... Peer should never
3968 * be unable to accept packets of the size that prior AFS versions would
3969 * send without asking. */
3970 if (peer->maxMTU != tSize) {
3971 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3973 peer->maxMTU = tSize;
3974 peer->MTU = MIN(tSize, peer->MTU);
3975 call->MTU = MIN(call->MTU, tSize);
3978 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3981 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
3982 (int)sizeof(afs_int32), &tSize);
3983 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3984 if (tSize < call->twind) { /* smaller than our send */
3985 call->twind = tSize; /* window, we must send less... */
3986 call->ssthresh = MIN(call->twind, call->ssthresh);
3987 call->conn->twind[call->channel] = call->twind;
3990 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3991 * network MTU confused with the loopback MTU. Calculate the
3992 * maximum MTU here for use in the slow start code below.
3994 maxMTU = peer->maxMTU;
3995 /* Did peer restart with older RX version? */
3996 if (peer->maxDgramPackets > 1) {
3997 peer->maxDgramPackets = 1;
3999 } else if (np->length >=
4000 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4003 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4004 sizeof(afs_int32), &tSize);
4005 tSize = (afs_uint32) ntohl(tSize);
4007 * As of AFS 3.5 we set the send window to match the receive window.
4009 if (tSize < call->twind) {
4010 call->twind = tSize;
4011 call->conn->twind[call->channel] = call->twind;
4012 call->ssthresh = MIN(call->twind, call->ssthresh);
4013 } else if (tSize > call->twind) {
4014 call->twind = tSize;
4015 call->conn->twind[call->channel] = call->twind;
4019 * As of AFS 3.5, a jumbogram is more than one fixed size
4020 * packet transmitted in a single UDP datagram. If the remote
4021 * MTU is smaller than our local MTU then never send a datagram
4022 * larger than the natural MTU.
4025 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4026 (int)sizeof(afs_int32), &tSize);
4027 maxDgramPackets = (afs_uint32) ntohl(tSize);
4028 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4030 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4031 maxDgramPackets = MIN(maxDgramPackets, tSize);
4032 if (maxDgramPackets > 1) {
4033 peer->maxDgramPackets = maxDgramPackets;
4034 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4036 peer->maxDgramPackets = 1;
4037 call->MTU = peer->natMTU;
4039 } else if (peer->maxDgramPackets > 1) {
4040 /* Restarted with lower version of RX */
4041 peer->maxDgramPackets = 1;
4043 } else if (peer->maxDgramPackets > 1
4044 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4045 /* Restarted with lower version of RX */
4046 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4047 peer->natMTU = OLD_MAX_PACKET_SIZE;
4048 peer->MTU = OLD_MAX_PACKET_SIZE;
4049 peer->maxDgramPackets = 1;
4050 peer->nDgramPackets = 1;
4052 call->MTU = OLD_MAX_PACKET_SIZE;
4057 * Calculate how many datagrams were successfully received after
4058 * the first missing packet and adjust the negative ack counter
4063 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4064 if (call->nNacks < nNacked) {
4065 call->nNacks = nNacked;
4068 call->nAcks += newAckCount;
4072 if (call->flags & RX_CALL_FAST_RECOVER) {
4074 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4076 call->flags &= ~RX_CALL_FAST_RECOVER;
4077 call->cwind = call->nextCwind;
4078 call->nextCwind = 0;
4081 call->nCwindAcks = 0;
4082 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4083 /* Three negative acks in a row trigger congestion recovery */
4084 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4085 MUTEX_EXIT(&peer->peer_lock);
4086 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4087 /* someone else is waiting to start recovery */
4090 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4091 rxi_WaitforTQBusy(call);
4092 MUTEX_ENTER(&peer->peer_lock);
4093 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4094 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4095 call->flags |= RX_CALL_FAST_RECOVER;
4096 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4098 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4099 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4100 call->nextCwind = call->ssthresh;
4103 peer->MTU = call->MTU;
4104 peer->cwind = call->nextCwind;
4105 peer->nDgramPackets = call->nDgramPackets;
4107 call->congestSeq = peer->congestSeq;
4108 /* Reset the resend times on the packets that were nacked
4109 * so we will retransmit as soon as the window permits*/
4110 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4112 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4113 clock_Zero(&tp->retryTime);
4115 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4120 /* If cwind is smaller than ssthresh, then increase
4121 * the window one packet for each ack we receive (exponential
4123 * If cwind is greater than or equal to ssthresh then increase
4124 * the congestion window by one packet for each cwind acks we
4125 * receive (linear growth). */
4126 if (call->cwind < call->ssthresh) {
4128 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4129 call->nCwindAcks = 0;
4131 call->nCwindAcks += newAckCount;
4132 if (call->nCwindAcks >= call->cwind) {
4133 call->nCwindAcks = 0;
4134 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4138 * If we have received several acknowledgements in a row then
4139 * it is time to increase the size of our datagrams
4141 if ((int)call->nAcks > rx_nDgramThreshold) {
4142 if (peer->maxDgramPackets > 1) {
4143 if (call->nDgramPackets < peer->maxDgramPackets) {
4144 call->nDgramPackets++;
4146 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4147 } else if (call->MTU < peer->maxMTU) {
4148 call->MTU += peer->natMTU;
4149 call->MTU = MIN(call->MTU, peer->maxMTU);
4155 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4157 /* Servers need to hold the call until all response packets have
4158 * been acknowledged. Soft acks are good enough since clients
4159 * are not allowed to clear their receive queues. */
4160 if (call->state == RX_STATE_HOLD
4161 && call->tfirst + call->nSoftAcked >= call->tnext) {
4162 call->state = RX_STATE_DALLY;
4163 rxi_ClearTransmitQueue(call, 0);
4164 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4165 } else if (!queue_IsEmpty(&call->tq)) {
4166 rxi_Start(0, call, 0, istack);
4171 /* Received a response to a challenge packet */
4173 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4174 struct rx_packet *np, int istack)
4178 /* Ignore the packet if we're the client */
4179 if (conn->type == RX_CLIENT_CONNECTION)
4182 /* If already authenticated, ignore the packet (it's probably a retry) */
4183 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4186 /* Otherwise, have the security object evaluate the response packet */
4187 error = RXS_CheckResponse(conn->securityObject, conn, np);
4189 /* If the response is invalid, reset the connection, sending
4190 * an abort to the peer */
4194 rxi_ConnectionError(conn, error);
4195 MUTEX_ENTER(&conn->conn_data_lock);
4196 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4197 MUTEX_EXIT(&conn->conn_data_lock);
4200 /* If the response is valid, any calls waiting to attach
4201 * servers can now do so */
4204 for (i = 0; i < RX_MAXCALLS; i++) {
4205 struct rx_call *call = conn->call[i];
4207 MUTEX_ENTER(&call->lock);
4208 if (call->state == RX_STATE_PRECALL)
4209 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4210 /* tnop can be null if newcallp is null */
4211 MUTEX_EXIT(&call->lock);
4215 /* Update the peer reachability information, just in case
4216 * some calls went into attach-wait while we were waiting
4217 * for authentication..
4219 rxi_UpdatePeerReach(conn, NULL);
4224 /* A client has received an authentication challenge: the security
4225 * object is asked to cough up a respectable response packet to send
4226 * back to the server. The server is responsible for retrying the
4227 * challenge if it fails to get a response. */
4230 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4231 struct rx_packet *np, int istack)
4235 /* Ignore the challenge if we're the server */
4236 if (conn->type == RX_SERVER_CONNECTION)
4239 /* Ignore the challenge if the connection is otherwise idle; someone's
4240 * trying to use us as an oracle. */
4241 if (!rxi_HasActiveCalls(conn))
4244 /* Send the security object the challenge packet. It is expected to fill
4245 * in the response. */
4246 error = RXS_GetResponse(conn->securityObject, conn, np);
4248 /* If the security object is unable to return a valid response, reset the
4249 * connection and send an abort to the peer. Otherwise send the response
4250 * packet to the peer connection. */
4252 rxi_ConnectionError(conn, error);
4253 MUTEX_ENTER(&conn->conn_data_lock);
4254 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4255 MUTEX_EXIT(&conn->conn_data_lock);
4257 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4258 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4264 /* Find an available server process to service the current request in
4265 * the given call structure. If one isn't available, queue up this
4266 * call so it eventually gets one */
4268 rxi_AttachServerProc(struct rx_call *call,
4269 osi_socket socket, int *tnop,
4270 struct rx_call **newcallp)
4272 struct rx_serverQueueEntry *sq;
4273 struct rx_service *service = call->conn->service;
4276 /* May already be attached */
4277 if (call->state == RX_STATE_ACTIVE)
4280 MUTEX_ENTER(&rx_serverPool_lock);
4282 haveQuota = QuotaOK(service);
4283 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4284 /* If there are no processes available to service this call,
4285 * put the call on the incoming call queue (unless it's
4286 * already on the queue).
4288 #ifdef RX_ENABLE_LOCKS
4290 ReturnToServerPool(service);
4291 #endif /* RX_ENABLE_LOCKS */
4293 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4294 call->flags |= RX_CALL_WAIT_PROC;
4295 MUTEX_ENTER(&rx_waiting_mutex);
4298 MUTEX_EXIT(&rx_waiting_mutex);
4299 rxi_calltrace(RX_CALL_ARRIVAL, call);
4300 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4301 queue_Append(&rx_incomingCallQueue, call);
4304 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4306 /* If hot threads are enabled, and both newcallp and sq->socketp
4307 * are non-null, then this thread will process the call, and the
4308 * idle server thread will start listening on this threads socket.
4311 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4314 *sq->socketp = socket;
4315 clock_GetTime(&call->startTime);
4316 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4320 if (call->flags & RX_CALL_WAIT_PROC) {
4321 /* Conservative: I don't think this should happen */
4322 call->flags &= ~RX_CALL_WAIT_PROC;
4323 if (queue_IsOnQueue(call)) {
4326 MUTEX_ENTER(&rx_waiting_mutex);
4328 MUTEX_EXIT(&rx_waiting_mutex);
4331 call->state = RX_STATE_ACTIVE;
4332 call->mode = RX_MODE_RECEIVING;
4333 #ifdef RX_KERNEL_TRACE
4335 int glockOwner = ISAFS_GLOCK();
4338 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4339 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4345 if (call->flags & RX_CALL_CLEARED) {
4346 /* send an ack now to start the packet flow up again */
4347 call->flags &= ~RX_CALL_CLEARED;
4348 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4350 #ifdef RX_ENABLE_LOCKS
4353 service->nRequestsRunning++;
4354 if (service->nRequestsRunning <= service->minProcs)
4360 MUTEX_EXIT(&rx_serverPool_lock);
4363 /* Delay the sending of an acknowledge event for a short while, while
4364 * a new call is being prepared (in the case of a client) or a reply
4365 * is being prepared (in the case of a server). Rather than sending
4366 * an ack packet, an ACKALL packet is sent. */
4368 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4370 #ifdef RX_ENABLE_LOCKS
4372 MUTEX_ENTER(&call->lock);
4373 call->delayedAckEvent = NULL;
4374 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4376 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4377 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4379 MUTEX_EXIT(&call->lock);
4380 #else /* RX_ENABLE_LOCKS */
4382 call->delayedAckEvent = NULL;
4383 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4384 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4385 #endif /* RX_ENABLE_LOCKS */
4389 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4391 struct rx_call *call = arg1;
4392 #ifdef RX_ENABLE_LOCKS
4394 MUTEX_ENTER(&call->lock);
4395 if (event == call->delayedAckEvent)
4396 call->delayedAckEvent = NULL;
4397 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4399 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4401 MUTEX_EXIT(&call->lock);
4402 #else /* RX_ENABLE_LOCKS */
4404 call->delayedAckEvent = NULL;
4405 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4406 #endif /* RX_ENABLE_LOCKS */
4410 #ifdef RX_ENABLE_LOCKS
4411 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4412 * clearing them out.
4415 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4417 struct rx_packet *p, *tp;
4420 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4421 p->flags |= RX_PKTFLAG_ACKED;
4425 call->flags |= RX_CALL_TQ_CLEARME;
4426 call->flags |= RX_CALL_TQ_SOME_ACKED;
4429 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4430 call->tfirst = call->tnext;
4431 call->nSoftAcked = 0;
4433 if (call->flags & RX_CALL_FAST_RECOVER) {
4434 call->flags &= ~RX_CALL_FAST_RECOVER;
4435 call->cwind = call->nextCwind;
4436 call->nextCwind = 0;
4439 CV_SIGNAL(&call->cv_twind);
4441 #endif /* RX_ENABLE_LOCKS */
4443 /* Clear out the transmit queue for the current call (all packets have
4444 * been received by peer) */
4446 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4448 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4449 struct rx_packet *p, *tp;
4451 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4453 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4454 p->flags |= RX_PKTFLAG_ACKED;
4458 call->flags |= RX_CALL_TQ_CLEARME;
4459 call->flags |= RX_CALL_TQ_SOME_ACKED;
4462 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4463 #ifdef RXDEBUG_PACKET
4465 #endif /* RXDEBUG_PACKET */
4466 rxi_FreePackets(0, &call->tq);
4467 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
4468 #ifdef RX_ENABLE_LOCKS
4469 CV_BROADCAST(&call->cv_tq);
4470 #else /* RX_ENABLE_LOCKS */
4471 osi_rxWakeup(&call->tq);
4472 #endif /* RX_ENABLE_LOCKS */
4474 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4475 call->flags &= ~RX_CALL_TQ_CLEARME;
4477 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4479 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4480 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4481 call->nSoftAcked = 0;
4483 if (call->flags & RX_CALL_FAST_RECOVER) {
4484 call->flags &= ~RX_CALL_FAST_RECOVER;
4485 call->cwind = call->nextCwind;
4487 #ifdef RX_ENABLE_LOCKS
4488 CV_SIGNAL(&call->cv_twind);
4490 osi_rxWakeup(&call->twind);
4495 rxi_ClearReceiveQueue(struct rx_call *call)
4497 if (queue_IsNotEmpty(&call->rq)) {
4500 count = rxi_FreePackets(0, &call->rq);
4501 rx_packetReclaims += count;
4502 #ifdef RXDEBUG_PACKET
4504 if ( call->rqc != 0 )
4505 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0", call, call->rqc));
4507 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4509 if (call->state == RX_STATE_PRECALL) {
4510 call->flags |= RX_CALL_CLEARED;
4514 /* Send an abort packet for the specified call */
4516 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
4517 int istack, int force)
4520 struct clock when, now;
4525 /* Clients should never delay abort messages */
4526 if (rx_IsClientConn(call->conn))
4529 if (call->abortCode != call->error) {
4530 call->abortCode = call->error;
4531 call->abortCount = 0;
4534 if (force || rxi_callAbortThreshhold == 0
4535 || call->abortCount < rxi_callAbortThreshhold) {
4536 if (call->delayedAbortEvent) {
4537 rxevent_Cancel(call->delayedAbortEvent, call,
4538 RX_CALL_REFCOUNT_ABORT);
4540 error = htonl(call->error);
4543 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4544 (char *)&error, sizeof(error), istack);
4545 } else if (!call->delayedAbortEvent) {
4546 clock_GetTime(&now);
4548 clock_Addmsec(&when, rxi_callAbortDelay);
4549 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4550 call->delayedAbortEvent =
4551 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4556 /* Send an abort packet for the specified connection. Packet is an
4557 * optional pointer to a packet that can be used to send the abort.
4558 * Once the number of abort messages reaches the threshhold, an
4559 * event is scheduled to send the abort. Setting the force flag
4560 * overrides sending delayed abort messages.
4562 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4563 * to send the abort packet.
4566 rxi_SendConnectionAbort(struct rx_connection *conn,
4567 struct rx_packet *packet, int istack, int force)
4570 struct clock when, now;
4575 /* Clients should never delay abort messages */
4576 if (rx_IsClientConn(conn))
4579 if (force || rxi_connAbortThreshhold == 0
4580 || conn->abortCount < rxi_connAbortThreshhold) {
4581 if (conn->delayedAbortEvent) {
4582 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4584 error = htonl(conn->error);
4586 MUTEX_EXIT(&conn->conn_data_lock);
4588 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4589 RX_PACKET_TYPE_ABORT, (char *)&error,
4590 sizeof(error), istack);
4591 MUTEX_ENTER(&conn->conn_data_lock);
4592 } else if (!conn->delayedAbortEvent) {
4593 clock_GetTime(&now);
4595 clock_Addmsec(&when, rxi_connAbortDelay);
4596 conn->delayedAbortEvent =
4597 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4602 /* Associate an error all of the calls owned by a connection. Called
4603 * with error non-zero. This is only for really fatal things, like
4604 * bad authentication responses. The connection itself is set in
4605 * error at this point, so that future packets received will be
4608 rxi_ConnectionError(struct rx_connection *conn,
4614 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d", conn, error));
4616 MUTEX_ENTER(&conn->conn_data_lock);
4617 if (conn->challengeEvent)
4618 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4619 if (conn->natKeepAliveEvent)
4620 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
4621 if (conn->checkReachEvent) {
4622 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4623 conn->checkReachEvent = 0;
4624 conn->flags &= ~RX_CONN_ATTACHWAIT;
4627 MUTEX_EXIT(&conn->conn_data_lock);
4628 for (i = 0; i < RX_MAXCALLS; i++) {
4629 struct rx_call *call = conn->call[i];
4631 MUTEX_ENTER(&call->lock);
4632 rxi_CallError(call, error);
4633 MUTEX_EXIT(&call->lock);
4636 conn->error = error;
4637 if (rx_stats_active)
4638 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4643 rxi_CallError(struct rx_call *call, afs_int32 error)
4646 osirx_AssertMine(&call->lock, "rxi_CallError");
4648 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d", call, error, call->error));
4650 error = call->error;
4652 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4653 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4654 rxi_ResetCall(call, 0);
4657 rxi_ResetCall(call, 0);
4659 call->error = error;
4660 call->mode = RX_MODE_ERROR;
4663 /* Reset various fields in a call structure, and wakeup waiting
4664 * processes. Some fields aren't changed: state & mode are not
4665 * touched (these must be set by the caller), and bufptr, nLeft, and
4666 * nFree are not reset, since these fields are manipulated by
4667 * unprotected macros, and may only be reset by non-interrupting code.
4670 /* this code requires that call->conn be set properly as a pre-condition. */
4671 #endif /* ADAPT_WINDOW */
4674 rxi_ResetCall(struct rx_call *call, int newcall)
4677 struct rx_peer *peer;
4678 struct rx_packet *packet;
4680 osirx_AssertMine(&call->lock, "rxi_ResetCall");
4682 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
4684 /* Notify anyone who is waiting for asynchronous packet arrival */
4685 if (call->arrivalProc) {
4686 (*call->arrivalProc) (call, call->arrivalProcHandle,
4687 call->arrivalProcArg);
4688 call->arrivalProc = (void (*)())0;
4691 if (call->delayedAbortEvent) {
4692 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4693 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4695 rxi_SendCallAbort(call, packet, 0, 1);
4696 rxi_FreePacket(packet);
4701 * Update the peer with the congestion information in this call
4702 * so other calls on this connection can pick up where this call
4703 * left off. If the congestion sequence numbers don't match then
4704 * another call experienced a retransmission.
4706 peer = call->conn->peer;
4707 MUTEX_ENTER(&peer->peer_lock);
4709 if (call->congestSeq == peer->congestSeq) {
4710 peer->cwind = MAX(peer->cwind, call->cwind);
4711 peer->MTU = MAX(peer->MTU, call->MTU);
4712 peer->nDgramPackets =
4713 MAX(peer->nDgramPackets, call->nDgramPackets);
4716 call->abortCode = 0;
4717 call->abortCount = 0;
4719 if (peer->maxDgramPackets > 1) {
4720 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4722 call->MTU = peer->MTU;
4724 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4725 call->ssthresh = rx_maxSendWindow;
4726 call->nDgramPackets = peer->nDgramPackets;
4727 call->congestSeq = peer->congestSeq;
4728 MUTEX_EXIT(&peer->peer_lock);
4730 flags = call->flags;
4731 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4732 if (flags & RX_CALL_TQ_BUSY) {
4733 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4734 call->flags |= (flags & RX_CALL_TQ_WAIT);
4736 #ifdef RX_ENABLE_LOCKS
4737 CV_WAIT(&call->cv_tq, &call->lock);
4738 #else /* RX_ENABLE_LOCKS */
4739 osi_rxSleep(&call->tq);
4740 #endif /* RX_ENABLE_LOCKS */
4742 if (call->tqWaiters == 0)
4743 call->flags &= ~RX_CALL_TQ_WAIT;
4745 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4747 rxi_ClearTransmitQueue(call, 1);
4748 /* why init the queue if you just emptied it? queue_Init(&call->tq); */
4749 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4750 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4755 rxi_ClearReceiveQueue(call);
4756 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
4758 if (call->currentPacket) {
4759 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
4760 call->currentPacket->flags |= RX_PKTFLAG_IOVQ;
4761 queue_Prepend(&call->iovq, call->currentPacket);
4762 #ifdef RXDEBUG_PACKET
4764 #endif /* RXDEBUG_PACKET */
4765 call->currentPacket = (struct rx_packet *)0;
4767 call->curlen = call->nLeft = call->nFree = 0;
4769 #ifdef RXDEBUG_PACKET
4772 rxi_FreePackets(0, &call->iovq);
4775 call->twind = call->conn->twind[call->channel];
4776 call->rwind = call->conn->rwind[call->channel];
4777 call->nSoftAcked = 0;
4778 call->nextCwind = 0;
4781 call->nCwindAcks = 0;
4782 call->nSoftAcks = 0;
4783 call->nHardAcks = 0;
4785 call->tfirst = call->rnext = call->tnext = 1;
4787 call->lastAcked = 0;
4788 call->localStatus = call->remoteStatus = 0;
4790 if (flags & RX_CALL_READER_WAIT) {
4791 #ifdef RX_ENABLE_LOCKS
4792 CV_BROADCAST(&call->cv_rq);
4794 osi_rxWakeup(&call->rq);
4797 if (flags & RX_CALL_WAIT_PACKETS) {
4798 MUTEX_ENTER(&rx_freePktQ_lock);
4799 rxi_PacketsUnWait(); /* XXX */
4800 MUTEX_EXIT(&rx_freePktQ_lock);
4802 #ifdef RX_ENABLE_LOCKS
4803 CV_SIGNAL(&call->cv_twind);
4805 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4806 osi_rxWakeup(&call->twind);
4809 #ifdef RX_ENABLE_LOCKS
4810 /* The following ensures that we don't mess with any queue while some
4811 * other thread might also be doing so. The call_queue_lock field is
4812 * is only modified under the call lock. If the call is in the process
4813 * of being removed from a queue, the call is not locked until the
4814 * the queue lock is dropped and only then is the call_queue_lock field
4815 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4816 * Note that any other routine which removes a call from a queue has to
4817 * obtain the queue lock before examing the queue and removing the call.
4819 if (call->call_queue_lock) {
4820 MUTEX_ENTER(call->call_queue_lock);
4821 if (queue_IsOnQueue(call)) {
4823 if (flags & RX_CALL_WAIT_PROC) {
4825 MUTEX_ENTER(&rx_waiting_mutex);
4827 MUTEX_EXIT(&rx_waiting_mutex);
4830 MUTEX_EXIT(call->call_queue_lock);
4831 CLEAR_CALL_QUEUE_LOCK(call);
4833 #else /* RX_ENABLE_LOCKS */
4834 if (queue_IsOnQueue(call)) {
4836 if (flags & RX_CALL_WAIT_PROC)
4839 #endif /* RX_ENABLE_LOCKS */
4841 rxi_KeepAliveOff(call);
4842 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4845 /* Send an acknowledge for the indicated packet (seq,serial) of the
4846 * indicated call, for the indicated reason (reason). This
4847 * acknowledge will specifically acknowledge receiving the packet, and
4848 * will also specify which other packets for this call have been
4849 * received. This routine returns the packet that was used to the
4850 * caller. The caller is responsible for freeing it or re-using it.
4851 * This acknowledgement also returns the highest sequence number
4852 * actually read out by the higher level to the sender; the sender
4853 * promises to keep around packets that have not been read by the
4854 * higher level yet (unless, of course, the sender decides to abort
4855 * the call altogether). Any of p, seq, serial, pflags, or reason may
4856 * be set to zero without ill effect. That is, if they are zero, they
4857 * will not convey any information.
4858 * NOW there is a trailer field, after the ack where it will safely be
4859 * ignored by mundanes, which indicates the maximum size packet this
4860 * host can swallow. */
4862 struct rx_packet *optionalPacket; use to send ack (or null)
4863 int seq; Sequence number of the packet we are acking
4864 int serial; Serial number of the packet
4865 int pflags; Flags field from packet header
4866 int reason; Reason an acknowledge was prompted
4870 rxi_SendAck(struct rx_call *call,
4871 struct rx_packet *optionalPacket, int serial, int reason,
4874 struct rx_ackPacket *ap;
4875 struct rx_packet *rqp;
4876 struct rx_packet *nxp; /* For queue_Scan */
4877 struct rx_packet *p;
4880 #ifdef RX_ENABLE_TSFPQ
4881 struct rx_ts_info_t * rx_ts_info;
4885 * Open the receive window once a thread starts reading packets
4887 if (call->rnext > 1) {
4888 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
4891 call->nHardAcks = 0;
4892 call->nSoftAcks = 0;
4893 if (call->rnext > call->lastAcked)
4894 call->lastAcked = call->rnext;
4898 rx_computelen(p, p->length); /* reset length, you never know */
4899 } /* where that's been... */
4900 #ifdef RX_ENABLE_TSFPQ
4902 RX_TS_INFO_GET(rx_ts_info);
4903 if ((p = rx_ts_info->local_special_packet)) {
4904 rx_computelen(p, p->length);
4905 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4906 rx_ts_info->local_special_packet = p;
4907 } else { /* We won't send the ack, but don't panic. */
4908 return optionalPacket;
4912 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4913 /* We won't send the ack, but don't panic. */
4914 return optionalPacket;
4919 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4922 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4923 #ifndef RX_ENABLE_TSFPQ
4924 if (!optionalPacket)
4927 return optionalPacket;
4929 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4930 if (rx_Contiguous(p) < templ) {
4931 #ifndef RX_ENABLE_TSFPQ
4932 if (!optionalPacket)
4935 return optionalPacket;
4940 /* MTUXXX failing to send an ack is very serious. We should */
4941 /* try as hard as possible to send even a partial ack; it's */
4942 /* better than nothing. */
4943 ap = (struct rx_ackPacket *)rx_DataOf(p);
4944 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4945 ap->reason = reason;
4947 /* The skew computation used to be bogus, I think it's better now. */
4948 /* We should start paying attention to skew. XXX */
4949 ap->serial = htonl(serial);
4950 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4952 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4953 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4955 /* No fear of running out of ack packet here because there can only be at most
4956 * one window full of unacknowledged packets. The window size must be constrained
4957 * to be less than the maximum ack size, of course. Also, an ack should always
4958 * fit into a single packet -- it should not ever be fragmented. */
4959 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4960 if (!rqp || !call->rq.next
4961 || (rqp->header.seq > (call->rnext + call->rwind))) {
4962 #ifndef RX_ENABLE_TSFPQ
4963 if (!optionalPacket)
4966 rxi_CallError(call, RX_CALL_DEAD);
4967 return optionalPacket;
4970 while (rqp->header.seq > call->rnext + offset)
4971 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4972 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4974 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4975 #ifndef RX_ENABLE_TSFPQ
4976 if (!optionalPacket)
4979 rxi_CallError(call, RX_CALL_DEAD);
4980 return optionalPacket;
4985 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4987 /* these are new for AFS 3.3 */
4988 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4989 templ = htonl(templ);
4990 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4991 templ = htonl(call->conn->peer->ifMTU);
4992 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4993 sizeof(afs_int32), &templ);
4995 /* new for AFS 3.4 */
4996 templ = htonl(call->rwind);
4997 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4998 sizeof(afs_int32), &templ);
5000 /* new for AFS 3.5 */
5001 templ = htonl(call->conn->peer->ifDgramPackets);
5002 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5003 sizeof(afs_int32), &templ);
5005 p->header.serviceId = call->conn->serviceId;
5006 p->header.cid = (call->conn->cid | call->channel);
5007 p->header.callNumber = *call->callNumber;
5009 p->header.securityIndex = call->conn->securityIndex;
5010 p->header.epoch = call->conn->epoch;
5011 p->header.type = RX_PACKET_TYPE_ACK;
5012 p->header.flags = RX_SLOW_START_OK;
5013 if (reason == RX_ACK_PING) {
5014 p->header.flags |= RX_REQUEST_ACK;
5016 clock_GetTime(&call->pingRequestTime);
5019 if (call->conn->type == RX_CLIENT_CONNECTION)
5020 p->header.flags |= RX_CLIENT_INITIATED;
5024 if (rxdebug_active) {
5028 len = _snprintf(msg, sizeof(msg),
5029 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5030 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5031 ntohl(ap->serial), ntohl(ap->previousPacket),
5032 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5033 ap->nAcks, ntohs(ap->bufferSpace) );
5037 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5038 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5042 OutputDebugString(msg);
5044 #else /* AFS_NT40_ENV */
5046 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5047 ap->reason, ntohl(ap->previousPacket),
5048 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5050 for (offset = 0; offset < ap->nAcks; offset++)
5051 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5056 #endif /* AFS_NT40_ENV */
5059 int i, nbytes = p->length;
5061 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5062 if (nbytes <= p->wirevec[i].iov_len) {
5065 savelen = p->wirevec[i].iov_len;
5067 p->wirevec[i].iov_len = nbytes;
5069 rxi_Send(call, p, istack);
5070 p->wirevec[i].iov_len = savelen;
5074 nbytes -= p->wirevec[i].iov_len;
5077 if (rx_stats_active)
5078 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
5079 #ifndef RX_ENABLE_TSFPQ
5080 if (!optionalPacket)
5083 return optionalPacket; /* Return packet for re-use by caller */
5086 /* Send all of the packets in the list in single datagram */
5088 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
5089 int istack, int moreFlag, struct clock *now,
5090 struct clock *retryTime, int resending)
5095 struct rx_connection *conn = call->conn;
5096 struct rx_peer *peer = conn->peer;
5098 MUTEX_ENTER(&peer->peer_lock);
5101 peer->reSends += len;
5102 if (rx_stats_active)
5103 rx_MutexAdd(rx_stats.dataPacketsSent, len, rx_stats_mutex);
5104 MUTEX_EXIT(&peer->peer_lock);
5106 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
5110 /* Set the packet flags and schedule the resend events */
5111 /* Only request an ack for the last packet in the list */
5112 for (i = 0; i < len; i++) {
5113 list[i]->retryTime = *retryTime;
5114 if (list[i]->header.serial) {
5115 /* Exponentially backoff retry times */
5116 if (list[i]->backoff < MAXBACKOFF) {
5117 /* so it can't stay == 0 */
5118 list[i]->backoff = (list[i]->backoff << 1) + 1;
5121 clock_Addmsec(&(list[i]->retryTime),
5122 ((afs_uint32) list[i]->backoff) << 8);
5125 /* Wait a little extra for the ack on the last packet */
5126 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
5127 clock_Addmsec(&(list[i]->retryTime), 400);
5130 /* Record the time sent */
5131 list[i]->timeSent = *now;
5133 /* Ask for an ack on retransmitted packets, on every other packet
5134 * if the peer doesn't support slow start. Ask for an ack on every
5135 * packet until the congestion window reaches the ack rate. */
5136 if (list[i]->header.serial) {
5138 if (rx_stats_active)
5139 rx_MutexIncrement(rx_stats.dataPacketsReSent, rx_stats_mutex);
5141 /* improved RTO calculation- not Karn */
5142 list[i]->firstSent = *now;
5143 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5144 || (!(call->flags & RX_CALL_SLOW_START_OK)
5145 && (list[i]->header.seq & 1)))) {
5150 /* Tag this packet as not being the last in this group,
5151 * for the receiver's benefit */
5152 if (i < len - 1 || moreFlag) {
5153 list[i]->header.flags |= RX_MORE_PACKETS;
5156 /* Install the new retransmit time for the packet, and
5157 * record the time sent */
5158 list[i]->timeSent = *now;
5162 list[len - 1]->header.flags |= RX_REQUEST_ACK;
5165 /* Since we're about to send a data packet to the peer, it's
5166 * safe to nuke any scheduled end-of-packets ack */
5167 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5169 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5170 MUTEX_EXIT(&call->lock);
5172 rxi_SendPacketList(call, conn, list, len, istack);
5174 rxi_SendPacket(call, conn, list[0], istack);
5176 MUTEX_ENTER(&call->lock);
5177 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5179 /* Update last send time for this call (for keep-alive
5180 * processing), and for the connection (so that we can discover
5181 * idle connections) */
5182 call->lastSendData = conn->lastSendTime = call->lastSendTime = clock_Sec();
5185 /* When sending packets we need to follow these rules:
5186 * 1. Never send more than maxDgramPackets in a jumbogram.
5187 * 2. Never send a packet with more than two iovecs in a jumbogram.
5188 * 3. Never send a retransmitted packet in a jumbogram.
5189 * 4. Never send more than cwind/4 packets in a jumbogram
5190 * We always keep the last list we should have sent so we
5191 * can set the RX_MORE_PACKETS flags correctly.
5194 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5195 int istack, struct clock *now, struct clock *retryTime,
5198 int i, cnt, lastCnt = 0;
5199 struct rx_packet **listP, **lastP = 0;
5200 struct rx_peer *peer = call->conn->peer;
5201 int morePackets = 0;
5203 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
5204 /* Does the current packet force us to flush the current list? */
5206 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5207 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5209 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
5211 /* If the call enters an error state stop sending, or if
5212 * we entered congestion recovery mode, stop sending */
5213 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5221 /* Add the current packet to the list if it hasn't been acked.
5222 * Otherwise adjust the list pointer to skip the current packet. */
5223 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5225 /* Do we need to flush the list? */
5226 if (cnt >= (int)peer->maxDgramPackets
5227 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5228 || list[i]->header.serial
5229 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5231 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5232 retryTime, resending);
5233 /* If the call enters an error state stop sending, or if
5234 * we entered congestion recovery mode, stop sending */
5236 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5241 listP = &list[i + 1];
5246 osi_Panic("rxi_SendList error");
5248 listP = &list[i + 1];
5252 /* Send the whole list when the call is in receive mode, when
5253 * the call is in eof mode, when we are in fast recovery mode,
5254 * and when we have the last packet */
5255 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5256 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5257 || (call->flags & RX_CALL_FAST_RECOVER)) {
5258 /* Check for the case where the current list contains
5259 * an acked packet. Since we always send retransmissions
5260 * in a separate packet, we only need to check the first
5261 * packet in the list */
5262 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5266 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5267 retryTime, resending);
5268 /* If the call enters an error state stop sending, or if
5269 * we entered congestion recovery mode, stop sending */
5270 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5274 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5277 } else if (lastCnt > 0) {
5278 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5283 #ifdef RX_ENABLE_LOCKS
5284 /* Call rxi_Start, below, but with the call lock held. */
5286 rxi_StartUnlocked(struct rxevent *event,
5287 void *arg0, void *arg1, int istack)
5289 struct rx_call *call = arg0;
5291 MUTEX_ENTER(&call->lock);
5292 rxi_Start(event, call, arg1, istack);
5293 MUTEX_EXIT(&call->lock);
5295 #endif /* RX_ENABLE_LOCKS */
5297 /* This routine is called when new packets are readied for
5298 * transmission and when retransmission may be necessary, or when the
5299 * transmission window or burst count are favourable. This should be
5300 * better optimized for new packets, the usual case, now that we've
5301 * got rid of queues of send packets. XXXXXXXXXXX */
5303 rxi_Start(struct rxevent *event,
5304 void *arg0, void *arg1, int istack)
5306 struct rx_call *call = arg0;
5308 struct rx_packet *p;
5309 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5310 struct rx_peer *peer = call->conn->peer;
5311 struct clock now, usenow, retryTime;
5315 struct rx_packet **xmitList;
5318 /* If rxi_Start is being called as a result of a resend event,
5319 * then make sure that the event pointer is removed from the call
5320 * structure, since there is no longer a per-call retransmission
5322 if (event && event == call->resendEvent) {
5323 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5324 call->resendEvent = NULL;
5326 if (queue_IsEmpty(&call->tq)) {
5330 /* Timeouts trigger congestion recovery */
5331 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5332 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5333 /* someone else is waiting to start recovery */
5336 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5337 rxi_WaitforTQBusy(call);
5338 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5339 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5340 call->flags |= RX_CALL_FAST_RECOVER;
5341 if (peer->maxDgramPackets > 1) {
5342 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5344 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5346 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5347 call->nDgramPackets = 1;
5349 call->nextCwind = 1;
5352 MUTEX_ENTER(&peer->peer_lock);
5353 peer->MTU = call->MTU;
5354 peer->cwind = call->cwind;
5355 peer->nDgramPackets = 1;
5357 call->congestSeq = peer->congestSeq;
5358 MUTEX_EXIT(&peer->peer_lock);
5359 /* Clear retry times on packets. Otherwise, it's possible for
5360 * some packets in the queue to force resends at rates faster
5361 * than recovery rates.
5363 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5364 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5365 clock_Zero(&p->retryTime);
5370 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5371 if (rx_stats_active)
5372 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5377 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5378 /* Get clock to compute the re-transmit time for any packets
5379 * in this burst. Note, if we back off, it's reasonable to
5380 * back off all of the packets in the same manner, even if
5381 * some of them have been retransmitted more times than more
5383 * Do a dance to avoid blocking after setting now. */
5384 MUTEX_ENTER(&peer->peer_lock);
5385 retryTime = peer->timeout;
5386 MUTEX_EXIT(&peer->peer_lock);
5387 clock_GetTime(&now);
5388 clock_Add(&retryTime, &now);
5390 /* Send (or resend) any packets that need it, subject to
5391 * window restrictions and congestion burst control
5392 * restrictions. Ask for an ack on the last packet sent in
5393 * this burst. For now, we're relying upon the window being
5394 * considerably bigger than the largest number of packets that
5395 * are typically sent at once by one initial call to
5396 * rxi_Start. This is probably bogus (perhaps we should ask
5397 * for an ack when we're half way through the current
5398 * window?). Also, for non file transfer applications, this
5399 * may end up asking for an ack for every packet. Bogus. XXXX
5402 * But check whether we're here recursively, and let the other guy
5405 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5406 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5407 call->flags |= RX_CALL_TQ_BUSY;
5409 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5411 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5412 call->flags &= ~RX_CALL_NEED_START;
5413 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5415 maxXmitPackets = MIN(call->twind, call->cwind);
5416 xmitList = (struct rx_packet **)
5417 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
5418 /* XXXX else we must drop any mtx we hold */
5419 afs_osi_Alloc_NoSleep(maxXmitPackets * sizeof(struct rx_packet *));
5421 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5423 if (xmitList == NULL)
5424 osi_Panic("rxi_Start, failed to allocate xmit list");
5425 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5426 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5427 /* We shouldn't be sending packets if a thread is waiting
5428 * to initiate congestion recovery */
5429 dpf(("call %d waiting to initiate fast recovery\n",
5430 *(call->callNumber)));
5434 && (call->flags & RX_CALL_FAST_RECOVER)) {
5435 /* Only send one packet during fast recovery */
5436 dpf(("call %d restricted to one packet per send during fast recovery\n",
5437 *(call->callNumber)));
5440 if ((p->flags & RX_PKTFLAG_FREE)
5441 || (!queue_IsEnd(&call->tq, nxp)
5442 && (nxp->flags & RX_PKTFLAG_FREE))
5443 || (p == (struct rx_packet *)&rx_freePacketQueue)
5444 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5445 osi_Panic("rxi_Start: xmit queue clobbered");
5447 if (p->flags & RX_PKTFLAG_ACKED) {
5448 /* Since we may block, don't trust this */
5449 usenow.sec = usenow.usec = 0;
5450 if (rx_stats_active)
5451 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5452 continue; /* Ignore this packet if it has been acknowledged */
5455 /* Turn off all flags except these ones, which are the same
5456 * on each transmission */
5457 p->header.flags &= RX_PRESET_FLAGS;
5459 if (p->header.seq >=
5460 call->tfirst + MIN((int)call->twind,
5461 (int)(call->nSoftAcked +
5463 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5464 /* Note: if we're waiting for more window space, we can
5465 * still send retransmits; hence we don't return here, but
5466 * break out to schedule a retransmit event */
5467 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
5468 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
5473 /* Transmit the packet if it needs to be sent. */
5474 if (!clock_Lt(&now, &p->retryTime)) {
5475 if (nXmitPackets == maxXmitPackets) {
5476 rxi_SendXmitList(call, xmitList, nXmitPackets,
5477 istack, &now, &retryTime,
5479 osi_Free(xmitList, maxXmitPackets *
5480 sizeof(struct rx_packet *));
5483 dpf(("call %d xmit packet %"AFS_PTR_FMT" now %u.%06u retryTime %u.%06u nextRetry %u.%06u\n",
5484 *(call->callNumber), p,
5486 p->retryTime.sec, p->retryTime.usec,
5487 retryTime.sec, retryTime.usec));
5488 xmitList[nXmitPackets++] = p;
5492 /* xmitList now hold pointers to all of the packets that are
5493 * ready to send. Now we loop to send the packets */
5494 if (nXmitPackets > 0) {
5495 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5496 &now, &retryTime, resending);
5499 maxXmitPackets * sizeof(struct rx_packet *));
5501 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5503 * TQ references no longer protected by this flag; they must remain
5504 * protected by the global lock.
5506 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5507 call->flags &= ~RX_CALL_TQ_BUSY;
5508 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5509 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
5510 call, call->tqWaiters, call->flags));
5511 #ifdef RX_ENABLE_LOCKS
5512 osirx_AssertMine(&call->lock, "rxi_Start start");
5513 CV_BROADCAST(&call->cv_tq);
5514 #else /* RX_ENABLE_LOCKS */
5515 osi_rxWakeup(&call->tq);
5516 #endif /* RX_ENABLE_LOCKS */
5521 /* We went into the error state while sending packets. Now is
5522 * the time to reset the call. This will also inform the using
5523 * process that the call is in an error state.
5525 if (rx_stats_active)
5526 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5527 call->flags &= ~RX_CALL_TQ_BUSY;
5528 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5529 dpf(("call error %d while xmit %p has %d waiters and flags %d\n",
5530 call->error, call, call->tqWaiters, call->flags));
5531 #ifdef RX_ENABLE_LOCKS
5532 osirx_AssertMine(&call->lock, "rxi_Start middle");
5533 CV_BROADCAST(&call->cv_tq);
5534 #else /* RX_ENABLE_LOCKS */
5535 osi_rxWakeup(&call->tq);
5536 #endif /* RX_ENABLE_LOCKS */
5538 rxi_CallError(call, call->error);
5541 #ifdef RX_ENABLE_LOCKS
5542 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5544 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5545 /* Some packets have received acks. If they all have, we can clear
5546 * the transmit queue.
5549 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5550 if (p->header.seq < call->tfirst
5551 && (p->flags & RX_PKTFLAG_ACKED)) {
5553 p->flags &= ~RX_PKTFLAG_TQ;
5554 #ifdef RXDEBUG_PACKET
5562 call->flags |= RX_CALL_TQ_CLEARME;
5564 #endif /* RX_ENABLE_LOCKS */
5565 /* Don't bother doing retransmits if the TQ is cleared. */
5566 if (call->flags & RX_CALL_TQ_CLEARME) {
5567 rxi_ClearTransmitQueue(call, 1);
5569 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5572 /* Always post a resend event, if there is anything in the
5573 * queue, and resend is possible. There should be at least
5574 * one unacknowledged packet in the queue ... otherwise none
5575 * of these packets should be on the queue in the first place.
5577 if (call->resendEvent) {
5578 /* Cancel the existing event and post a new one */
5579 rxevent_Cancel(call->resendEvent, call,
5580 RX_CALL_REFCOUNT_RESEND);
5583 /* The retry time is the retry time on the first unacknowledged
5584 * packet inside the current window */
5586 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5587 /* Don't set timers for packets outside the window */
5588 if (p->header.seq >= call->tfirst + call->twind) {
5592 if (!(p->flags & RX_PKTFLAG_ACKED)
5593 && !clock_IsZero(&p->retryTime)) {
5595 retryTime = p->retryTime;
5600 /* Post a new event to re-run rxi_Start when retries may be needed */
5601 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5602 #ifdef RX_ENABLE_LOCKS
5603 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5605 rxevent_PostNow2(&retryTime, &usenow,
5607 (void *)call, 0, istack);
5608 #else /* RX_ENABLE_LOCKS */
5610 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5611 (void *)call, 0, istack);
5612 #endif /* RX_ENABLE_LOCKS */
5615 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5616 } while (call->flags & RX_CALL_NEED_START);
5618 * TQ references no longer protected by this flag; they must remain
5619 * protected by the global lock.
5621 call->flags &= ~RX_CALL_TQ_BUSY;
5622 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5623 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
5624 call, call->tqWaiters, call->flags));
5625 #ifdef RX_ENABLE_LOCKS
5626 osirx_AssertMine(&call->lock, "rxi_Start end");
5627 CV_BROADCAST(&call->cv_tq);
5628 #else /* RX_ENABLE_LOCKS */
5629 osi_rxWakeup(&call->tq);
5630 #endif /* RX_ENABLE_LOCKS */
5633 call->flags |= RX_CALL_NEED_START;
5635 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5637 if (call->resendEvent) {
5638 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5643 /* Also adjusts the keep alive parameters for the call, to reflect
5644 * that we have just sent a packet (so keep alives aren't sent
5647 rxi_Send(struct rx_call *call, struct rx_packet *p,
5650 struct rx_connection *conn = call->conn;
5652 /* Stamp each packet with the user supplied status */
5653 p->header.userStatus = call->localStatus;
5655 /* Allow the security object controlling this call's security to
5656 * make any last-minute changes to the packet */
5657 RXS_SendPacket(conn->securityObject, call, p);
5659 /* Since we're about to send SOME sort of packet to the peer, it's
5660 * safe to nuke any scheduled end-of-packets ack */
5661 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5663 /* Actually send the packet, filling in more connection-specific fields */
5664 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5665 MUTEX_EXIT(&call->lock);
5666 rxi_SendPacket(call, conn, p, istack);
5667 MUTEX_ENTER(&call->lock);
5668 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5670 /* Update last send time for this call (for keep-alive
5671 * processing), and for the connection (so that we can discover
5672 * idle connections) */
5673 conn->lastSendTime = call->lastSendTime = clock_Sec();
5674 /* Don't count keepalives here, so idleness can be tracked. */
5675 if ((p->header.type != RX_PACKET_TYPE_ACK) || (((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING))
5676 call->lastSendData = call->lastSendTime;
5680 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5681 * that things are fine. Also called periodically to guarantee that nothing
5682 * falls through the cracks (e.g. (error + dally) connections have keepalive
5683 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5685 * haveCTLock Set if calling from rxi_ReapConnections
5687 #ifdef RX_ENABLE_LOCKS
5689 rxi_CheckCall(struct rx_call *call, int haveCTLock)
5690 #else /* RX_ENABLE_LOCKS */
5692 rxi_CheckCall(struct rx_call *call)
5693 #endif /* RX_ENABLE_LOCKS */
5695 struct rx_connection *conn = call->conn;
5697 afs_uint32 deadTime;
5699 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5700 if (call->flags & RX_CALL_TQ_BUSY) {
5701 /* Call is active and will be reset by rxi_Start if it's
5702 * in an error state.
5707 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5709 (((afs_uint32) conn->secondsUntilDead << 10) +
5710 ((afs_uint32) conn->peer->rtt >> 3) +
5711 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5713 /* These are computed to the second (+- 1 second). But that's
5714 * good enough for these values, which should be a significant
5715 * number of seconds. */
5716 if (now > (call->lastReceiveTime + deadTime)) {
5717 if (call->state == RX_STATE_ACTIVE) {
5719 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5721 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5722 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5723 ip_stack_t *ipst = ns->netstack_ip;
5725 ire = ire_cache_lookup(call->conn->peer->host
5726 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5728 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5730 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5737 if (ire && ire->ire_max_frag > 0)
5738 rxi_SetPeerMtu(call->conn->peer->host, 0, ire->ire_max_frag);
5739 #if defined(GLOBAL_NETSTACKID)
5743 #endif /* ADAPT_PMTU */
5744 rxi_CallError(call, RX_CALL_DEAD);
5747 #ifdef RX_ENABLE_LOCKS
5748 /* Cancel pending events */
5749 rxevent_Cancel(call->delayedAckEvent, call,
5750 RX_CALL_REFCOUNT_DELAY);
5751 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5752 rxevent_Cancel(call->keepAliveEvent, call,
5753 RX_CALL_REFCOUNT_ALIVE);
5754 if (call->refCount == 0) {
5755 rxi_FreeCall(call, haveCTLock);
5759 #else /* RX_ENABLE_LOCKS */
5762 #endif /* RX_ENABLE_LOCKS */
5764 /* Non-active calls are destroyed if they are not responding
5765 * to pings; active calls are simply flagged in error, so the
5766 * attached process can die reasonably gracefully. */
5768 /* see if we have a non-activity timeout */
5769 if (call->startWait && conn->idleDeadTime
5770 && ((call->startWait + conn->idleDeadTime) < now) &&
5771 (call->flags & RX_CALL_READER_WAIT)) {
5772 if (call->state == RX_STATE_ACTIVE) {
5773 rxi_CallError(call, RX_CALL_TIMEOUT);
5777 if (call->lastSendData && conn->idleDeadTime && (conn->idleDeadErr != 0)
5778 && ((call->lastSendData + conn->idleDeadTime) < now)) {
5779 if (call->state == RX_STATE_ACTIVE) {
5780 rxi_CallError(call, conn->idleDeadErr);
5784 /* see if we have a hard timeout */
5785 if (conn->hardDeadTime
5786 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5787 if (call->state == RX_STATE_ACTIVE)
5788 rxi_CallError(call, RX_CALL_TIMEOUT);
5795 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
5797 struct rx_connection *conn = arg1;
5798 struct rx_header theader;
5800 struct sockaddr_in taddr;
5803 struct iovec tmpiov[2];
5806 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
5809 tp = &tbuffer[sizeof(struct rx_header)];
5810 taddr.sin_family = AF_INET;
5811 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
5812 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
5813 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
5814 taddr.sin_len = sizeof(struct sockaddr_in);
5816 memset(&theader, 0, sizeof(theader));
5817 theader.epoch = htonl(999);
5819 theader.callNumber = 0;
5822 theader.type = RX_PACKET_TYPE_VERSION;
5823 theader.flags = RX_LAST_PACKET;
5824 theader.serviceId = 0;
5826 memcpy(tbuffer, &theader, sizeof(theader));
5827 memcpy(tp, &a, sizeof(a));
5828 tmpiov[0].iov_base = tbuffer;
5829 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
5831 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
5833 MUTEX_ENTER(&conn->conn_data_lock);
5834 /* Only reschedule ourselves if the connection would not be destroyed */
5835 if (conn->refCount <= 1) {
5836 conn->natKeepAliveEvent = NULL;
5837 MUTEX_EXIT(&conn->conn_data_lock);
5838 rx_DestroyConnection(conn); /* drop the reference for this */
5840 conn->natKeepAliveEvent = NULL;
5841 conn->refCount--; /* drop the reference for this */
5842 rxi_ScheduleNatKeepAliveEvent(conn);
5843 MUTEX_EXIT(&conn->conn_data_lock);
5848 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
5850 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
5851 struct clock when, now;
5852 clock_GetTime(&now);
5854 when.sec += conn->secondsUntilNatPing;
5855 conn->refCount++; /* hold a reference for this */
5856 conn->natKeepAliveEvent =
5857 rxevent_PostNow(&when, &now, rxi_NatKeepAliveEvent, conn, 0);
5862 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
5864 MUTEX_ENTER(&conn->conn_data_lock);
5865 conn->secondsUntilNatPing = seconds;
5867 rxi_ScheduleNatKeepAliveEvent(conn);
5868 MUTEX_EXIT(&conn->conn_data_lock);
5872 rxi_NatKeepAliveOn(struct rx_connection *conn)
5874 MUTEX_ENTER(&conn->conn_data_lock);
5875 rxi_ScheduleNatKeepAliveEvent(conn);
5876 MUTEX_EXIT(&conn->conn_data_lock);
5879 /* When a call is in progress, this routine is called occasionally to
5880 * make sure that some traffic has arrived (or been sent to) the peer.
5881 * If nothing has arrived in a reasonable amount of time, the call is
5882 * declared dead; if nothing has been sent for a while, we send a
5883 * keep-alive packet (if we're actually trying to keep the call alive)
5886 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
5888 struct rx_call *call = arg1;
5889 struct rx_connection *conn;
5892 MUTEX_ENTER(&call->lock);
5893 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5894 if (event == call->keepAliveEvent)
5895 call->keepAliveEvent = NULL;
5898 #ifdef RX_ENABLE_LOCKS
5899 if (rxi_CheckCall(call, 0)) {
5900 MUTEX_EXIT(&call->lock);
5903 #else /* RX_ENABLE_LOCKS */
5904 if (rxi_CheckCall(call))
5906 #endif /* RX_ENABLE_LOCKS */
5908 /* Don't try to keep alive dallying calls */
5909 if (call->state == RX_STATE_DALLY) {
5910 MUTEX_EXIT(&call->lock);
5915 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5916 /* Don't try to send keepalives if there is unacknowledged data */
5917 /* the rexmit code should be good enough, this little hack
5918 * doesn't quite work XXX */
5919 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5921 rxi_ScheduleKeepAliveEvent(call);
5922 MUTEX_EXIT(&call->lock);
5927 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
5929 if (!call->keepAliveEvent) {
5930 struct clock when, now;
5931 clock_GetTime(&now);
5933 when.sec += call->conn->secondsUntilPing;
5934 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5935 call->keepAliveEvent =
5936 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
5940 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5942 rxi_KeepAliveOn(struct rx_call *call)
5944 /* Pretend last packet received was received now--i.e. if another
5945 * packet isn't received within the keep alive time, then the call
5946 * will die; Initialize last send time to the current time--even
5947 * if a packet hasn't been sent yet. This will guarantee that a
5948 * keep-alive is sent within the ping time */
5949 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5950 rxi_ScheduleKeepAliveEvent(call);
5953 /* This routine is called to send connection abort messages
5954 * that have been delayed to throttle looping clients. */
5956 rxi_SendDelayedConnAbort(struct rxevent *event,
5957 void *arg1, void *unused)
5959 struct rx_connection *conn = arg1;
5962 struct rx_packet *packet;
5964 MUTEX_ENTER(&conn->conn_data_lock);
5965 conn->delayedAbortEvent = NULL;
5966 error = htonl(conn->error);
5968 MUTEX_EXIT(&conn->conn_data_lock);
5969 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5972 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5973 RX_PACKET_TYPE_ABORT, (char *)&error,
5975 rxi_FreePacket(packet);
5979 /* This routine is called to send call abort messages
5980 * that have been delayed to throttle looping clients. */
5982 rxi_SendDelayedCallAbort(struct rxevent *event,
5983 void *arg1, void *dummy)
5985 struct rx_call *call = arg1;
5988 struct rx_packet *packet;
5990 MUTEX_ENTER(&call->lock);
5991 call->delayedAbortEvent = NULL;
5992 error = htonl(call->error);
5994 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5997 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5998 (char *)&error, sizeof(error), 0);
5999 rxi_FreePacket(packet);
6001 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6002 MUTEX_EXIT(&call->lock);
6005 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6006 * seconds) to ask the client to authenticate itself. The routine
6007 * issues a challenge to the client, which is obtained from the
6008 * security object associated with the connection */
6010 rxi_ChallengeEvent(struct rxevent *event,
6011 void *arg0, void *arg1, int tries)
6013 struct rx_connection *conn = arg0;
6015 conn->challengeEvent = NULL;
6016 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6017 struct rx_packet *packet;
6018 struct clock when, now;
6021 /* We've failed to authenticate for too long.
6022 * Reset any calls waiting for authentication;
6023 * they are all in RX_STATE_PRECALL.
6027 MUTEX_ENTER(&conn->conn_call_lock);
6028 for (i = 0; i < RX_MAXCALLS; i++) {
6029 struct rx_call *call = conn->call[i];
6031 MUTEX_ENTER(&call->lock);
6032 if (call->state == RX_STATE_PRECALL) {
6033 rxi_CallError(call, RX_CALL_DEAD);
6034 rxi_SendCallAbort(call, NULL, 0, 0);
6036 MUTEX_EXIT(&call->lock);
6039 MUTEX_EXIT(&conn->conn_call_lock);
6043 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6045 /* If there's no packet available, do this later. */
6046 RXS_GetChallenge(conn->securityObject, conn, packet);
6047 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6048 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6049 rxi_FreePacket(packet);
6051 clock_GetTime(&now);
6053 when.sec += RX_CHALLENGE_TIMEOUT;
6054 conn->challengeEvent =
6055 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
6060 /* Call this routine to start requesting the client to authenticate
6061 * itself. This will continue until authentication is established,
6062 * the call times out, or an invalid response is returned. The
6063 * security object associated with the connection is asked to create
6064 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6065 * defined earlier. */
6067 rxi_ChallengeOn(struct rx_connection *conn)
6069 if (!conn->challengeEvent) {
6070 RXS_CreateChallenge(conn->securityObject, conn);
6071 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6076 /* Compute round trip time of the packet provided, in *rttp.
6079 /* rxi_ComputeRoundTripTime is called with peer locked. */
6080 /* sentp and/or peer may be null */
6082 rxi_ComputeRoundTripTime(struct rx_packet *p,
6083 struct clock *sentp,
6084 struct rx_peer *peer)
6086 struct clock thisRtt, *rttp = &thisRtt;
6090 clock_GetTime(rttp);
6092 if (clock_Lt(rttp, sentp)) {
6094 return; /* somebody set the clock back, don't count this time. */
6096 clock_Sub(rttp, sentp);
6097 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6098 p->header.callNumber, p, rttp->sec, rttp->usec));
6100 if (rttp->sec == 0 && rttp->usec == 0) {
6102 * The actual round trip time is shorter than the
6103 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6104 * Since we can't tell which at the moment we will assume 1ms.
6109 if (rx_stats_active) {
6110 MUTEX_ENTER(&rx_stats_mutex);
6111 if (clock_Lt(rttp, &rx_stats.minRtt))
6112 rx_stats.minRtt = *rttp;
6113 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
6114 if (rttp->sec > 60) {
6115 MUTEX_EXIT(&rx_stats_mutex);
6116 return; /* somebody set the clock ahead */
6118 rx_stats.maxRtt = *rttp;
6120 clock_Add(&rx_stats.totalRtt, rttp);
6121 rx_stats.nRttSamples++;
6122 MUTEX_EXIT(&rx_stats_mutex);
6125 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6127 /* Apply VanJacobson round-trip estimations */
6132 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
6133 * srtt is stored as fixed point with 3 bits after the binary
6134 * point (i.e., scaled by 8). The following magic is
6135 * equivalent to the smoothing algorithm in rfc793 with an
6136 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6137 * srtt'*8 = rtt + srtt*7
6138 * srtt'*8 = srtt*8 + rtt - srtt
6139 * srtt' = srtt + rtt/8 - srtt/8
6140 * srtt' = srtt + (rtt - srtt)/8
6143 delta = _8THMSEC(rttp) - peer->rtt;
6144 peer->rtt += (delta >> 3);
6147 * We accumulate a smoothed rtt variance (actually, a smoothed
6148 * mean difference), then set the retransmit timer to smoothed
6149 * rtt + 4 times the smoothed variance (was 2x in van's original
6150 * paper, but 4x works better for me, and apparently for him as
6152 * rttvar is stored as
6153 * fixed point with 2 bits after the binary point (scaled by
6154 * 4). The following is equivalent to rfc793 smoothing with
6155 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6156 * rttvar'*4 = rttvar*3 + |delta|
6157 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6158 * rttvar' = rttvar + |delta|/4 - rttvar/4
6159 * rttvar' = rttvar + (|delta| - rttvar)/4
6160 * This replaces rfc793's wired-in beta.
6161 * dev*4 = dev*4 + (|actual - expected| - dev)
6167 delta -= (peer->rtt_dev << 1);
6168 peer->rtt_dev += (delta >> 3);
6170 /* I don't have a stored RTT so I start with this value. Since I'm
6171 * probably just starting a call, and will be pushing more data down
6172 * this, I expect congestion to increase rapidly. So I fudge a
6173 * little, and I set deviance to half the rtt. In practice,
6174 * deviance tends to approach something a little less than
6175 * half the smoothed rtt. */
6176 peer->rtt = _8THMSEC(rttp) + 8;
6177 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
6179 /* the timeout is RTT + 4*MDEV but no less than rx_minPeerTimeout msec.
6180 * This is because one end or the other of these connections is usually
6181 * in a user process, and can be switched and/or swapped out. So on fast,
6182 * reliable networks, the timeout would otherwise be too short. */
6183 rtt_timeout = MAX(((peer->rtt >> 3) + peer->rtt_dev), rx_minPeerTimeout);
6184 clock_Zero(&(peer->timeout));
6185 clock_Addmsec(&(peer->timeout), rtt_timeout);
6187 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6188 p->header.callNumber, p, MSEC(rttp), peer->rtt >> 3, peer->rtt_dev >> 2, (peer->timeout.sec), (peer->timeout.usec)));
6192 /* Find all server connections that have not been active for a long time, and
6195 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6197 struct clock now, when;
6198 clock_GetTime(&now);
6200 /* Find server connection structures that haven't been used for
6201 * greater than rx_idleConnectionTime */
6203 struct rx_connection **conn_ptr, **conn_end;
6204 int i, havecalls = 0;
6205 MUTEX_ENTER(&rx_connHashTable_lock);
6206 for (conn_ptr = &rx_connHashTable[0], conn_end =
6207 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6209 struct rx_connection *conn, *next;
6210 struct rx_call *call;
6214 for (conn = *conn_ptr; conn; conn = next) {
6215 /* XXX -- Shouldn't the connection be locked? */
6218 for (i = 0; i < RX_MAXCALLS; i++) {
6219 call = conn->call[i];
6223 code = MUTEX_TRYENTER(&call->lock);
6226 #ifdef RX_ENABLE_LOCKS
6227 result = rxi_CheckCall(call, 1);
6228 #else /* RX_ENABLE_LOCKS */
6229 result = rxi_CheckCall(call);
6230 #endif /* RX_ENABLE_LOCKS */
6231 MUTEX_EXIT(&call->lock);
6233 /* If CheckCall freed the call, it might
6234 * have destroyed the connection as well,
6235 * which screws up the linked lists.
6241 if (conn->type == RX_SERVER_CONNECTION) {
6242 /* This only actually destroys the connection if
6243 * there are no outstanding calls */
6244 MUTEX_ENTER(&conn->conn_data_lock);
6245 if (!havecalls && !conn->refCount
6246 && ((conn->lastSendTime + rx_idleConnectionTime) <
6248 conn->refCount++; /* it will be decr in rx_DestroyConn */
6249 MUTEX_EXIT(&conn->conn_data_lock);
6250 #ifdef RX_ENABLE_LOCKS
6251 rxi_DestroyConnectionNoLock(conn);
6252 #else /* RX_ENABLE_LOCKS */
6253 rxi_DestroyConnection(conn);
6254 #endif /* RX_ENABLE_LOCKS */
6256 #ifdef RX_ENABLE_LOCKS
6258 MUTEX_EXIT(&conn->conn_data_lock);
6260 #endif /* RX_ENABLE_LOCKS */
6264 #ifdef RX_ENABLE_LOCKS
6265 while (rx_connCleanup_list) {
6266 struct rx_connection *conn;
6267 conn = rx_connCleanup_list;
6268 rx_connCleanup_list = rx_connCleanup_list->next;
6269 MUTEX_EXIT(&rx_connHashTable_lock);
6270 rxi_CleanupConnection(conn);
6271 MUTEX_ENTER(&rx_connHashTable_lock);
6273 MUTEX_EXIT(&rx_connHashTable_lock);
6274 #endif /* RX_ENABLE_LOCKS */
6277 /* Find any peer structures that haven't been used (haven't had an
6278 * associated connection) for greater than rx_idlePeerTime */
6280 struct rx_peer **peer_ptr, **peer_end;
6282 MUTEX_ENTER(&rx_rpc_stats);
6283 MUTEX_ENTER(&rx_peerHashTable_lock);
6284 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6285 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6287 struct rx_peer *peer, *next, *prev;
6288 for (prev = peer = *peer_ptr; peer; peer = next) {
6290 code = MUTEX_TRYENTER(&peer->peer_lock);
6291 if ((code) && (peer->refCount == 0)
6292 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6293 rx_interface_stat_p rpc_stat, nrpc_stat;
6295 MUTEX_EXIT(&peer->peer_lock);
6296 MUTEX_DESTROY(&peer->peer_lock);
6298 (&peer->rpcStats, rpc_stat, nrpc_stat,
6299 rx_interface_stat)) {
6300 unsigned int num_funcs;
6303 queue_Remove(&rpc_stat->queue_header);
6304 queue_Remove(&rpc_stat->all_peers);
6305 num_funcs = rpc_stat->stats[0].func_total;
6307 sizeof(rx_interface_stat_t) +
6308 rpc_stat->stats[0].func_total *
6309 sizeof(rx_function_entry_v1_t);
6311 rxi_Free(rpc_stat, space);
6312 rxi_rpc_peer_stat_cnt -= num_funcs;
6315 if (rx_stats_active)
6316 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6317 if (peer == *peer_ptr) {
6324 MUTEX_EXIT(&peer->peer_lock);
6330 MUTEX_EXIT(&rx_peerHashTable_lock);
6331 MUTEX_EXIT(&rx_rpc_stats);
6334 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6335 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6336 * GC, just below. Really, we shouldn't have to keep moving packets from
6337 * one place to another, but instead ought to always know if we can
6338 * afford to hold onto a packet in its particular use. */
6339 MUTEX_ENTER(&rx_freePktQ_lock);
6340 if (rx_waitingForPackets) {
6341 rx_waitingForPackets = 0;
6342 #ifdef RX_ENABLE_LOCKS
6343 CV_BROADCAST(&rx_waitingForPackets_cv);
6345 osi_rxWakeup(&rx_waitingForPackets);
6348 MUTEX_EXIT(&rx_freePktQ_lock);
6351 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6352 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6356 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6357 * rx.h is sort of strange this is better. This is called with a security
6358 * object before it is discarded. Each connection using a security object has
6359 * its own refcount to the object so it won't actually be freed until the last
6360 * connection is destroyed.
6362 * This is the only rxs module call. A hold could also be written but no one
6366 rxs_Release(struct rx_securityClass *aobj)
6368 return RXS_Close(aobj);
6372 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6373 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6374 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6375 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6377 /* Adjust our estimate of the transmission rate to this peer, given
6378 * that the packet p was just acked. We can adjust peer->timeout and
6379 * call->twind. Pragmatically, this is called
6380 * only with packets of maximal length.
6381 * Called with peer and call locked.
6385 rxi_ComputeRate(struct rx_peer *peer, struct rx_call *call,
6386 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6388 afs_int32 xferSize, xferMs;
6392 /* Count down packets */
6393 if (peer->rateFlag > 0)
6395 /* Do nothing until we're enabled */
6396 if (peer->rateFlag != 0)
6401 /* Count only when the ack seems legitimate */
6402 switch (ackReason) {
6403 case RX_ACK_REQUESTED:
6405 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6409 case RX_ACK_PING_RESPONSE:
6410 if (p) /* want the response to ping-request, not data send */
6412 clock_GetTime(&newTO);
6413 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6414 clock_Sub(&newTO, &call->pingRequestTime);
6415 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6419 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6426 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %d.%06d, rtt %u, ps %u)",
6427 ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"),
6428 xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
6430 /* Track only packets that are big enough. */
6431 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6435 /* absorb RTT data (in milliseconds) for these big packets */
6436 if (peer->smRtt == 0) {
6437 peer->smRtt = xferMs;
6439 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6444 if (peer->countDown) {
6448 peer->countDown = 10; /* recalculate only every so often */
6450 /* In practice, we can measure only the RTT for full packets,
6451 * because of the way Rx acks the data that it receives. (If it's
6452 * smaller than a full packet, it often gets implicitly acked
6453 * either by the call response (from a server) or by the next call
6454 * (from a client), and either case confuses transmission times
6455 * with processing times.) Therefore, replace the above
6456 * more-sophisticated processing with a simpler version, where the
6457 * smoothed RTT is kept for full-size packets, and the time to
6458 * transmit a windowful of full-size packets is simply RTT *
6459 * windowSize. Again, we take two steps:
6460 - ensure the timeout is large enough for a single packet's RTT;
6461 - ensure that the window is small enough to fit in the desired timeout.*/
6463 /* First, the timeout check. */
6464 minTime = peer->smRtt;
6465 /* Get a reasonable estimate for a timeout period */
6467 newTO.sec = minTime / 1000;
6468 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6470 /* Increase the timeout period so that we can always do at least
6471 * one packet exchange */
6472 if (clock_Gt(&newTO, &peer->timeout)) {
6474 dpf(("CONG peer %lx/%u: timeout %d.%06d ==> %ld.%06d (rtt %u, ps %u)",
6475 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec,
6476 newTO.sec, newTO.usec, peer->smRtt, peer->packetSize));
6478 peer->timeout = newTO;
6481 /* Now, get an estimate for the transmit window size. */
6482 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6483 /* Now, convert to the number of full packets that could fit in a
6484 * reasonable fraction of that interval */
6485 minTime /= (peer->smRtt << 1);
6486 xferSize = minTime; /* (make a copy) */
6488 /* Now clamp the size to reasonable bounds. */
6491 else if (minTime > rx_Window)
6492 minTime = rx_Window;
6493 /* if (minTime != peer->maxWindow) {
6494 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6495 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6496 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6498 peer->maxWindow = minTime;
6499 elide... call->twind = minTime;
6503 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6504 * Discern this by calculating the timeout necessary for rx_Window
6506 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6507 /* calculate estimate for transmission interval in milliseconds */
6508 minTime = rx_Window * peer->smRtt;
6509 if (minTime < 1000) {
6510 dpf(("CONG peer %lx/%u: cut TO %d.%06d by 0.5 (rtt %u, ps %u)",
6511 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6512 peer->timeout.usec, peer->smRtt, peer->packetSize));
6514 newTO.sec = 0; /* cut back on timeout by half a second */
6515 newTO.usec = 500000;
6516 clock_Sub(&peer->timeout, &newTO);
6521 } /* end of rxi_ComputeRate */
6522 #endif /* ADAPT_WINDOW */
6530 #define TRACE_OPTION_RX_DEBUG 16
6538 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6539 0, KEY_QUERY_VALUE, &parmKey);
6540 if (code != ERROR_SUCCESS)
6543 dummyLen = sizeof(TraceOption);
6544 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6545 (BYTE *) &TraceOption, &dummyLen);
6546 if (code == ERROR_SUCCESS) {
6547 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
6549 RegCloseKey (parmKey);
6550 #endif /* AFS_NT40_ENV */
6555 rx_DebugOnOff(int on)
6559 rxdebug_active = on;
6565 rx_StatsOnOff(int on)
6568 rx_stats_active = on;
6573 /* Don't call this debugging routine directly; use dpf */
6575 rxi_DebugPrint(char *format, ...)
6584 va_start(ap, format);
6586 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6589 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
6591 if (msg[len-1] != '\n') {
6595 OutputDebugString(msg);
6602 va_start(ap, format);
6604 clock_GetTime(&now);
6605 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
6606 (unsigned int)now.usec);
6607 vfprintf(rx_Log, format, ap);
6616 * This function is used to process the rx_stats structure that is local
6617 * to a process as well as an rx_stats structure received from a remote
6618 * process (via rxdebug). Therefore, it needs to do minimal version
6622 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
6623 afs_int32 freePackets, char version)
6628 if (size != sizeof(struct rx_statistics)) {
6630 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
6631 size, sizeof(struct rx_statistics));
6634 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6637 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6638 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
6639 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6640 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6641 s->specialPktAllocFailures);
6643 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
6644 s->receivePktAllocFailures, s->sendPktAllocFailures,
6645 s->specialPktAllocFailures);
6649 " greedy %u, " "bogusReads %u (last from host %x), "
6650 "noPackets %u, " "noBuffers %u, " "selects %u, "
6651 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
6652 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6653 s->selects, s->sendSelects);
6655 fprintf(file, " packets read: ");
6656 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6657 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
6659 fprintf(file, "\n");
6662 " other read counters: data %u, " "ack %u, " "dup %u "
6663 "spurious %u " "dally %u\n", s->dataPacketsRead,
6664 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6665 s->ignorePacketDally);
6667 fprintf(file, " packets sent: ");
6668 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6669 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
6671 fprintf(file, "\n");
6674 " other send counters: ack %u, " "data %u (not resends), "
6675 "resends %u, " "pushed %u, " "acked&ignored %u\n",
6676 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6677 s->dataPacketsPushed, s->ignoreAckedPacket);
6680 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
6681 s->netSendFailures, (int)s->fatalErrors);
6683 if (s->nRttSamples) {
6684 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6685 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6687 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6688 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6692 " %d server connections, " "%d client connections, "
6693 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6694 s->nServerConns, s->nClientConns, s->nPeerStructs,
6695 s->nCallStructs, s->nFreeCallStructs);
6697 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6698 fprintf(file, " %d clock updates\n", clock_nUpdates);
6701 fprintf(file, "ERROR: compiled without RXDEBUG\n");
6705 /* for backward compatibility */
6707 rx_PrintStats(FILE * file)
6709 MUTEX_ENTER(&rx_stats_mutex);
6710 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6712 MUTEX_EXIT(&rx_stats_mutex);
6716 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6718 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
6719 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6720 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6723 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6724 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6725 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6728 " Packet size %d, " "max in packet skew %d, "
6729 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6730 (int)peer->outPacketSkew);
6734 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
6736 * This mutex protects the following static variables:
6740 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
6741 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
6743 #define LOCK_RX_DEBUG
6744 #define UNLOCK_RX_DEBUG
6745 #endif /* AFS_PTHREAD_ENV */
6749 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6750 u_char type, void *inputData, size_t inputLength,
6751 void *outputData, size_t outputLength)
6753 static afs_int32 counter = 100;
6754 time_t waitTime, waitCount, startTime;
6755 struct rx_header theader;
6758 struct timeval tv_now, tv_wake, tv_delta;
6759 struct sockaddr_in taddr, faddr;
6768 startTime = time(0);
6774 tp = &tbuffer[sizeof(struct rx_header)];
6775 taddr.sin_family = AF_INET;
6776 taddr.sin_port = remotePort;
6777 taddr.sin_addr.s_addr = remoteAddr;
6778 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6779 taddr.sin_len = sizeof(struct sockaddr_in);
6782 memset(&theader, 0, sizeof(theader));
6783 theader.epoch = htonl(999);
6785 theader.callNumber = htonl(counter);
6788 theader.type = type;
6789 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6790 theader.serviceId = 0;
6792 memcpy(tbuffer, &theader, sizeof(theader));
6793 memcpy(tp, inputData, inputLength);
6795 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6796 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6798 /* see if there's a packet available */
6799 gettimeofday(&tv_wake,0);
6800 tv_wake.tv_sec += waitTime;
6803 FD_SET(socket, &imask);
6804 tv_delta.tv_sec = tv_wake.tv_sec;
6805 tv_delta.tv_usec = tv_wake.tv_usec;
6806 gettimeofday(&tv_now, 0);
6808 if (tv_delta.tv_usec < tv_now.tv_usec) {
6810 tv_delta.tv_usec += 1000000;
6813 tv_delta.tv_usec -= tv_now.tv_usec;
6815 if (tv_delta.tv_sec < tv_now.tv_sec) {
6819 tv_delta.tv_sec -= tv_now.tv_sec;
6822 code = select(0, &imask, 0, 0, &tv_delta);
6823 #else /* AFS_NT40_ENV */
6824 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6825 #endif /* AFS_NT40_ENV */
6826 if (code == 1 && FD_ISSET(socket, &imask)) {
6827 /* now receive a packet */
6828 faddrLen = sizeof(struct sockaddr_in);
6830 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6831 (struct sockaddr *)&faddr, &faddrLen);
6834 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6835 if (counter == ntohl(theader.callNumber))
6843 /* see if we've timed out */
6851 code -= sizeof(struct rx_header);
6852 if (code > outputLength)
6853 code = outputLength;
6854 memcpy(outputData, tp, code);
6857 #endif /* RXDEBUG */
6860 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6861 afs_uint16 remotePort, struct rx_debugStats * stat,
6862 afs_uint32 * supportedValues)
6868 struct rx_debugIn in;
6870 *supportedValues = 0;
6871 in.type = htonl(RX_DEBUGI_GETSTATS);
6874 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6875 &in, sizeof(in), stat, sizeof(*stat));
6878 * If the call was successful, fixup the version and indicate
6879 * what contents of the stat structure are valid.
6880 * Also do net to host conversion of fields here.
6884 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6885 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6887 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6888 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6890 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6891 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6893 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6894 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6896 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6897 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6899 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6900 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6902 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6903 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6905 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6906 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6908 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
6909 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
6911 stat->nFreePackets = ntohl(stat->nFreePackets);
6912 stat->packetReclaims = ntohl(stat->packetReclaims);
6913 stat->callsExecuted = ntohl(stat->callsExecuted);
6914 stat->nWaiting = ntohl(stat->nWaiting);
6915 stat->idleThreads = ntohl(stat->idleThreads);
6916 stat->nWaited = ntohl(stat->nWaited);
6917 stat->nPackets = ntohl(stat->nPackets);
6924 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6925 afs_uint16 remotePort, struct rx_statistics * stat,
6926 afs_uint32 * supportedValues)
6932 struct rx_debugIn in;
6933 afs_int32 *lp = (afs_int32 *) stat;
6937 * supportedValues is currently unused, but added to allow future
6938 * versioning of this function.
6941 *supportedValues = 0;
6942 in.type = htonl(RX_DEBUGI_RXSTATS);
6944 memset(stat, 0, sizeof(*stat));
6946 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6947 &in, sizeof(in), stat, sizeof(*stat));
6952 * Do net to host conversion here
6955 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6964 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6965 afs_uint16 remotePort, size_t version_length,
6970 return MakeDebugCall(socket, remoteAddr, remotePort,
6971 RX_PACKET_TYPE_VERSION, a, 1, version,
6979 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6980 afs_uint16 remotePort, afs_int32 * nextConnection,
6981 int allConnections, afs_uint32 debugSupportedValues,
6982 struct rx_debugConn * conn,
6983 afs_uint32 * supportedValues)
6989 struct rx_debugIn in;
6993 * supportedValues is currently unused, but added to allow future
6994 * versioning of this function.
6997 *supportedValues = 0;
6998 if (allConnections) {
6999 in.type = htonl(RX_DEBUGI_GETALLCONN);
7001 in.type = htonl(RX_DEBUGI_GETCONN);
7003 in.index = htonl(*nextConnection);
7004 memset(conn, 0, sizeof(*conn));
7006 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7007 &in, sizeof(in), conn, sizeof(*conn));
7010 *nextConnection += 1;
7013 * Convert old connection format to new structure.
7016 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7017 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7018 #define MOVEvL(a) (conn->a = vL->a)
7020 /* any old or unrecognized version... */
7021 for (i = 0; i < RX_MAXCALLS; i++) {
7022 MOVEvL(callState[i]);
7023 MOVEvL(callMode[i]);
7024 MOVEvL(callFlags[i]);
7025 MOVEvL(callOther[i]);
7027 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7028 MOVEvL(secStats.type);
7029 MOVEvL(secStats.level);
7030 MOVEvL(secStats.flags);
7031 MOVEvL(secStats.expires);
7032 MOVEvL(secStats.packetsReceived);
7033 MOVEvL(secStats.packetsSent);
7034 MOVEvL(secStats.bytesReceived);
7035 MOVEvL(secStats.bytesSent);
7040 * Do net to host conversion here
7042 * I don't convert host or port since we are most likely
7043 * going to want these in NBO.
7045 conn->cid = ntohl(conn->cid);
7046 conn->serial = ntohl(conn->serial);
7047 for (i = 0; i < RX_MAXCALLS; i++) {
7048 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7050 conn->error = ntohl(conn->error);
7051 conn->secStats.flags = ntohl(conn->secStats.flags);
7052 conn->secStats.expires = ntohl(conn->secStats.expires);
7053 conn->secStats.packetsReceived =
7054 ntohl(conn->secStats.packetsReceived);
7055 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7056 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7057 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7058 conn->epoch = ntohl(conn->epoch);
7059 conn->natMTU = ntohl(conn->natMTU);
7066 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7067 afs_uint16 remotePort, afs_int32 * nextPeer,
7068 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7069 afs_uint32 * supportedValues)
7075 struct rx_debugIn in;
7078 * supportedValues is currently unused, but added to allow future
7079 * versioning of this function.
7082 *supportedValues = 0;
7083 in.type = htonl(RX_DEBUGI_GETPEER);
7084 in.index = htonl(*nextPeer);
7085 memset(peer, 0, sizeof(*peer));
7087 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7088 &in, sizeof(in), peer, sizeof(*peer));
7094 * Do net to host conversion here
7096 * I don't convert host or port since we are most likely
7097 * going to want these in NBO.
7099 peer->ifMTU = ntohs(peer->ifMTU);
7100 peer->idleWhen = ntohl(peer->idleWhen);
7101 peer->refCount = ntohs(peer->refCount);
7102 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7103 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7104 peer->rtt = ntohl(peer->rtt);
7105 peer->rtt_dev = ntohl(peer->rtt_dev);
7106 peer->timeout.sec = ntohl(peer->timeout.sec);
7107 peer->timeout.usec = ntohl(peer->timeout.usec);
7108 peer->nSent = ntohl(peer->nSent);
7109 peer->reSends = ntohl(peer->reSends);
7110 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7111 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7112 peer->rateFlag = ntohl(peer->rateFlag);
7113 peer->natMTU = ntohs(peer->natMTU);
7114 peer->maxMTU = ntohs(peer->maxMTU);
7115 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7116 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7117 peer->MTU = ntohs(peer->MTU);
7118 peer->cwind = ntohs(peer->cwind);
7119 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7120 peer->congestSeq = ntohs(peer->congestSeq);
7121 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7122 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7123 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7124 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7131 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7132 struct rx_debugPeer * peerStats)
7135 afs_int32 error = 1; /* default to "did not succeed" */
7136 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7138 MUTEX_ENTER(&rx_peerHashTable_lock);
7139 for(tp = rx_peerHashTable[hashValue];
7140 tp != NULL; tp = tp->next) {
7141 if (tp->host == peerHost)
7148 peerStats->host = tp->host;
7149 peerStats->port = tp->port;
7150 peerStats->ifMTU = tp->ifMTU;
7151 peerStats->idleWhen = tp->idleWhen;
7152 peerStats->refCount = tp->refCount;
7153 peerStats->burstSize = tp->burstSize;
7154 peerStats->burst = tp->burst;
7155 peerStats->burstWait.sec = tp->burstWait.sec;
7156 peerStats->burstWait.usec = tp->burstWait.usec;
7157 peerStats->rtt = tp->rtt;
7158 peerStats->rtt_dev = tp->rtt_dev;
7159 peerStats->timeout.sec = tp->timeout.sec;
7160 peerStats->timeout.usec = tp->timeout.usec;
7161 peerStats->nSent = tp->nSent;
7162 peerStats->reSends = tp->reSends;
7163 peerStats->inPacketSkew = tp->inPacketSkew;
7164 peerStats->outPacketSkew = tp->outPacketSkew;
7165 peerStats->rateFlag = tp->rateFlag;
7166 peerStats->natMTU = tp->natMTU;
7167 peerStats->maxMTU = tp->maxMTU;
7168 peerStats->maxDgramPackets = tp->maxDgramPackets;
7169 peerStats->ifDgramPackets = tp->ifDgramPackets;
7170 peerStats->MTU = tp->MTU;
7171 peerStats->cwind = tp->cwind;
7172 peerStats->nDgramPackets = tp->nDgramPackets;
7173 peerStats->congestSeq = tp->congestSeq;
7174 peerStats->bytesSent.high = tp->bytesSent.high;
7175 peerStats->bytesSent.low = tp->bytesSent.low;
7176 peerStats->bytesReceived.high = tp->bytesReceived.high;
7177 peerStats->bytesReceived.low = tp->bytesReceived.low;
7179 MUTEX_EXIT(&rx_peerHashTable_lock);
7187 struct rx_serverQueueEntry *np;
7190 struct rx_call *call;
7191 struct rx_serverQueueEntry *sq;
7195 if (rxinit_status == 1) {
7197 return; /* Already shutdown. */
7201 #ifndef AFS_PTHREAD_ENV
7202 FD_ZERO(&rx_selectMask);
7203 #endif /* AFS_PTHREAD_ENV */
7204 rxi_dataQuota = RX_MAX_QUOTA;
7205 #ifndef AFS_PTHREAD_ENV
7207 #endif /* AFS_PTHREAD_ENV */
7210 #ifndef AFS_PTHREAD_ENV
7211 #ifndef AFS_USE_GETTIMEOFDAY
7213 #endif /* AFS_USE_GETTIMEOFDAY */
7214 #endif /* AFS_PTHREAD_ENV */
7216 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7217 call = queue_First(&rx_freeCallQueue, rx_call);
7219 rxi_Free(call, sizeof(struct rx_call));
7222 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7223 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7229 struct rx_peer **peer_ptr, **peer_end;
7230 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7231 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7233 struct rx_peer *peer, *next;
7234 for (peer = *peer_ptr; peer; peer = next) {
7235 rx_interface_stat_p rpc_stat, nrpc_stat;
7238 (&peer->rpcStats, rpc_stat, nrpc_stat,
7239 rx_interface_stat)) {
7240 unsigned int num_funcs;
7243 queue_Remove(&rpc_stat->queue_header);
7244 queue_Remove(&rpc_stat->all_peers);
7245 num_funcs = rpc_stat->stats[0].func_total;
7247 sizeof(rx_interface_stat_t) +
7248 rpc_stat->stats[0].func_total *
7249 sizeof(rx_function_entry_v1_t);
7251 rxi_Free(rpc_stat, space);
7252 MUTEX_ENTER(&rx_rpc_stats);
7253 rxi_rpc_peer_stat_cnt -= num_funcs;
7254 MUTEX_EXIT(&rx_rpc_stats);
7258 if (rx_stats_active)
7259 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
7263 for (i = 0; i < RX_MAX_SERVICES; i++) {
7265 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7267 for (i = 0; i < rx_hashTableSize; i++) {
7268 struct rx_connection *tc, *ntc;
7269 MUTEX_ENTER(&rx_connHashTable_lock);
7270 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7272 for (j = 0; j < RX_MAXCALLS; j++) {
7274 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7277 rxi_Free(tc, sizeof(*tc));
7279 MUTEX_EXIT(&rx_connHashTable_lock);
7282 MUTEX_ENTER(&freeSQEList_lock);
7284 while ((np = rx_FreeSQEList)) {
7285 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7286 MUTEX_DESTROY(&np->lock);
7287 rxi_Free(np, sizeof(*np));
7290 MUTEX_EXIT(&freeSQEList_lock);
7291 MUTEX_DESTROY(&freeSQEList_lock);
7292 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7293 MUTEX_DESTROY(&rx_connHashTable_lock);
7294 MUTEX_DESTROY(&rx_peerHashTable_lock);
7295 MUTEX_DESTROY(&rx_serverPool_lock);
7297 osi_Free(rx_connHashTable,
7298 rx_hashTableSize * sizeof(struct rx_connection *));
7299 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7301 UNPIN(rx_connHashTable,
7302 rx_hashTableSize * sizeof(struct rx_connection *));
7303 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7305 rxi_FreeAllPackets();
7307 MUTEX_ENTER(&rx_quota_mutex);
7308 rxi_dataQuota = RX_MAX_QUOTA;
7309 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7310 MUTEX_EXIT(&rx_quota_mutex);
7315 #ifdef RX_ENABLE_LOCKS
7317 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7319 if (!MUTEX_ISMINE(lockaddr))
7320 osi_Panic("Lock not held: %s", msg);
7322 #endif /* RX_ENABLE_LOCKS */
7327 * Routines to implement connection specific data.
7331 rx_KeyCreate(rx_destructor_t rtn)
7334 MUTEX_ENTER(&rxi_keyCreate_lock);
7335 key = rxi_keyCreate_counter++;
7336 rxi_keyCreate_destructor = (rx_destructor_t *)
7337 realloc((void *)rxi_keyCreate_destructor,
7338 (key + 1) * sizeof(rx_destructor_t));
7339 rxi_keyCreate_destructor[key] = rtn;
7340 MUTEX_EXIT(&rxi_keyCreate_lock);
7345 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7348 MUTEX_ENTER(&conn->conn_data_lock);
7349 if (!conn->specific) {
7350 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7351 for (i = 0; i < key; i++)
7352 conn->specific[i] = NULL;
7353 conn->nSpecific = key + 1;
7354 conn->specific[key] = ptr;
7355 } else if (key >= conn->nSpecific) {
7356 conn->specific = (void **)
7357 realloc(conn->specific, (key + 1) * sizeof(void *));
7358 for (i = conn->nSpecific; i < key; i++)
7359 conn->specific[i] = NULL;
7360 conn->nSpecific = key + 1;
7361 conn->specific[key] = ptr;
7363 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7364 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7365 conn->specific[key] = ptr;
7367 MUTEX_EXIT(&conn->conn_data_lock);
7371 rx_GetSpecific(struct rx_connection *conn, int key)
7374 MUTEX_ENTER(&conn->conn_data_lock);
7375 if (key >= conn->nSpecific)
7378 ptr = conn->specific[key];
7379 MUTEX_EXIT(&conn->conn_data_lock);
7383 #endif /* !KERNEL */
7386 * processStats is a queue used to store the statistics for the local
7387 * process. Its contents are similar to the contents of the rpcStats
7388 * queue on a rx_peer structure, but the actual data stored within
7389 * this queue contains totals across the lifetime of the process (assuming
7390 * the stats have not been reset) - unlike the per peer structures
7391 * which can come and go based upon the peer lifetime.
7394 static struct rx_queue processStats = { &processStats, &processStats };
7397 * peerStats is a queue used to store the statistics for all peer structs.
7398 * Its contents are the union of all the peer rpcStats queues.
7401 static struct rx_queue peerStats = { &peerStats, &peerStats };
7404 * rxi_monitor_processStats is used to turn process wide stat collection
7408 static int rxi_monitor_processStats = 0;
7411 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7414 static int rxi_monitor_peerStats = 0;
7417 * rxi_AddRpcStat - given all of the information for a particular rpc
7418 * call, create (if needed) and update the stat totals for the rpc.
7422 * IN stats - the queue of stats that will be updated with the new value
7424 * IN rxInterface - a unique number that identifies the rpc interface
7426 * IN currentFunc - the index of the function being invoked
7428 * IN totalFunc - the total number of functions in this interface
7430 * IN queueTime - the amount of time this function waited for a thread
7432 * IN execTime - the amount of time this function invocation took to execute
7434 * IN bytesSent - the number bytes sent by this invocation
7436 * IN bytesRcvd - the number bytes received by this invocation
7438 * IN isServer - if true, this invocation was made to a server
7440 * IN remoteHost - the ip address of the remote host
7442 * IN remotePort - the port of the remote host
7444 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7446 * INOUT counter - if a new stats structure is allocated, the counter will
7447 * be updated with the new number of allocated stat structures
7455 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
7456 afs_uint32 currentFunc, afs_uint32 totalFunc,
7457 struct clock *queueTime, struct clock *execTime,
7458 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
7459 afs_uint32 remoteHost, afs_uint32 remotePort,
7460 int addToPeerList, unsigned int *counter)
7463 rx_interface_stat_p rpc_stat, nrpc_stat;
7466 * See if there's already a structure for this interface
7469 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7470 if ((rpc_stat->stats[0].interfaceId == rxInterface)
7471 && (rpc_stat->stats[0].remote_is_server == isServer))
7476 * Didn't find a match so allocate a new structure and add it to the
7480 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
7481 || (rpc_stat->stats[0].interfaceId != rxInterface)
7482 || (rpc_stat->stats[0].remote_is_server != isServer)) {
7487 sizeof(rx_interface_stat_t) +
7488 totalFunc * sizeof(rx_function_entry_v1_t);
7490 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
7491 if (rpc_stat == NULL) {
7495 *counter += totalFunc;
7496 for (i = 0; i < totalFunc; i++) {
7497 rpc_stat->stats[i].remote_peer = remoteHost;
7498 rpc_stat->stats[i].remote_port = remotePort;
7499 rpc_stat->stats[i].remote_is_server = isServer;
7500 rpc_stat->stats[i].interfaceId = rxInterface;
7501 rpc_stat->stats[i].func_total = totalFunc;
7502 rpc_stat->stats[i].func_index = i;
7503 hzero(rpc_stat->stats[i].invocations);
7504 hzero(rpc_stat->stats[i].bytes_sent);
7505 hzero(rpc_stat->stats[i].bytes_rcvd);
7506 rpc_stat->stats[i].queue_time_sum.sec = 0;
7507 rpc_stat->stats[i].queue_time_sum.usec = 0;
7508 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7509 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7510 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7511 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7512 rpc_stat->stats[i].queue_time_max.sec = 0;
7513 rpc_stat->stats[i].queue_time_max.usec = 0;
7514 rpc_stat->stats[i].execution_time_sum.sec = 0;
7515 rpc_stat->stats[i].execution_time_sum.usec = 0;
7516 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7517 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7518 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7519 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7520 rpc_stat->stats[i].execution_time_max.sec = 0;
7521 rpc_stat->stats[i].execution_time_max.usec = 0;
7523 queue_Prepend(stats, rpc_stat);
7524 if (addToPeerList) {
7525 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7530 * Increment the stats for this function
7533 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7534 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7535 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7536 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7537 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7538 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7539 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7541 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7542 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7544 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7545 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7547 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7548 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7550 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7551 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7559 * rx_IncrementTimeAndCount - increment the times and count for a particular
7564 * IN peer - the peer who invoked the rpc
7566 * IN rxInterface - a unique number that identifies the rpc interface
7568 * IN currentFunc - the index of the function being invoked
7570 * IN totalFunc - the total number of functions in this interface
7572 * IN queueTime - the amount of time this function waited for a thread
7574 * IN execTime - the amount of time this function invocation took to execute
7576 * IN bytesSent - the number bytes sent by this invocation
7578 * IN bytesRcvd - the number bytes received by this invocation
7580 * IN isServer - if true, this invocation was made to a server
7588 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7589 afs_uint32 currentFunc, afs_uint32 totalFunc,
7590 struct clock *queueTime, struct clock *execTime,
7591 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7595 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7598 MUTEX_ENTER(&rx_rpc_stats);
7599 MUTEX_ENTER(&peer->peer_lock);
7601 if (rxi_monitor_peerStats) {
7602 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7603 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7604 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7607 if (rxi_monitor_processStats) {
7608 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7609 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7610 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7613 MUTEX_EXIT(&peer->peer_lock);
7614 MUTEX_EXIT(&rx_rpc_stats);
7619 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7623 * IN callerVersion - the rpc stat version of the caller.
7625 * IN count - the number of entries to marshall.
7627 * IN stats - pointer to stats to be marshalled.
7629 * OUT ptr - Where to store the marshalled data.
7636 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7637 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7643 * We only support the first version
7645 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7646 *(ptr++) = stats->remote_peer;
7647 *(ptr++) = stats->remote_port;
7648 *(ptr++) = stats->remote_is_server;
7649 *(ptr++) = stats->interfaceId;
7650 *(ptr++) = stats->func_total;
7651 *(ptr++) = stats->func_index;
7652 *(ptr++) = hgethi(stats->invocations);
7653 *(ptr++) = hgetlo(stats->invocations);
7654 *(ptr++) = hgethi(stats->bytes_sent);
7655 *(ptr++) = hgetlo(stats->bytes_sent);
7656 *(ptr++) = hgethi(stats->bytes_rcvd);
7657 *(ptr++) = hgetlo(stats->bytes_rcvd);
7658 *(ptr++) = stats->queue_time_sum.sec;
7659 *(ptr++) = stats->queue_time_sum.usec;
7660 *(ptr++) = stats->queue_time_sum_sqr.sec;
7661 *(ptr++) = stats->queue_time_sum_sqr.usec;
7662 *(ptr++) = stats->queue_time_min.sec;
7663 *(ptr++) = stats->queue_time_min.usec;
7664 *(ptr++) = stats->queue_time_max.sec;
7665 *(ptr++) = stats->queue_time_max.usec;
7666 *(ptr++) = stats->execution_time_sum.sec;
7667 *(ptr++) = stats->execution_time_sum.usec;
7668 *(ptr++) = stats->execution_time_sum_sqr.sec;
7669 *(ptr++) = stats->execution_time_sum_sqr.usec;
7670 *(ptr++) = stats->execution_time_min.sec;
7671 *(ptr++) = stats->execution_time_min.usec;
7672 *(ptr++) = stats->execution_time_max.sec;
7673 *(ptr++) = stats->execution_time_max.usec;
7679 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7684 * IN callerVersion - the rpc stat version of the caller
7686 * OUT myVersion - the rpc stat version of this function
7688 * OUT clock_sec - local time seconds
7690 * OUT clock_usec - local time microseconds
7692 * OUT allocSize - the number of bytes allocated to contain stats
7694 * OUT statCount - the number stats retrieved from this process.
7696 * OUT stats - the actual stats retrieved from this process.
7700 * Returns void. If successful, stats will != NULL.
7704 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7705 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7706 size_t * allocSize, afs_uint32 * statCount,
7707 afs_uint32 ** stats)
7717 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7720 * Check to see if stats are enabled
7723 MUTEX_ENTER(&rx_rpc_stats);
7724 if (!rxi_monitor_processStats) {
7725 MUTEX_EXIT(&rx_rpc_stats);
7729 clock_GetTime(&now);
7730 *clock_sec = now.sec;
7731 *clock_usec = now.usec;
7734 * Allocate the space based upon the caller version
7736 * If the client is at an older version than we are,
7737 * we return the statistic data in the older data format, but
7738 * we still return our version number so the client knows we
7739 * are maintaining more data than it can retrieve.
7742 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7743 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7744 *statCount = rxi_rpc_process_stat_cnt;
7747 * This can't happen yet, but in the future version changes
7748 * can be handled by adding additional code here
7752 if (space > (size_t) 0) {
7754 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7757 rx_interface_stat_p rpc_stat, nrpc_stat;
7761 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7763 * Copy the data based upon the caller version
7765 rx_MarshallProcessRPCStats(callerVersion,
7766 rpc_stat->stats[0].func_total,
7767 rpc_stat->stats, &ptr);
7773 MUTEX_EXIT(&rx_rpc_stats);
7778 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7782 * IN callerVersion - the rpc stat version of the caller
7784 * OUT myVersion - the rpc stat version of this function
7786 * OUT clock_sec - local time seconds
7788 * OUT clock_usec - local time microseconds
7790 * OUT allocSize - the number of bytes allocated to contain stats
7792 * OUT statCount - the number of stats retrieved from the individual
7795 * OUT stats - the actual stats retrieved from the individual peer structures.
7799 * Returns void. If successful, stats will != NULL.
7803 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7804 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7805 size_t * allocSize, afs_uint32 * statCount,
7806 afs_uint32 ** stats)
7816 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7819 * Check to see if stats are enabled
7822 MUTEX_ENTER(&rx_rpc_stats);
7823 if (!rxi_monitor_peerStats) {
7824 MUTEX_EXIT(&rx_rpc_stats);
7828 clock_GetTime(&now);
7829 *clock_sec = now.sec;
7830 *clock_usec = now.usec;
7833 * Allocate the space based upon the caller version
7835 * If the client is at an older version than we are,
7836 * we return the statistic data in the older data format, but
7837 * we still return our version number so the client knows we
7838 * are maintaining more data than it can retrieve.
7841 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7842 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7843 *statCount = rxi_rpc_peer_stat_cnt;
7846 * This can't happen yet, but in the future version changes
7847 * can be handled by adding additional code here
7851 if (space > (size_t) 0) {
7853 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7856 rx_interface_stat_p rpc_stat, nrpc_stat;
7860 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7862 * We have to fix the offset of rpc_stat since we are
7863 * keeping this structure on two rx_queues. The rx_queue
7864 * package assumes that the rx_queue member is the first
7865 * member of the structure. That is, rx_queue assumes that
7866 * any one item is only on one queue at a time. We are
7867 * breaking that assumption and so we have to do a little
7868 * math to fix our pointers.
7871 fix_offset = (char *)rpc_stat;
7872 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7873 rpc_stat = (rx_interface_stat_p) fix_offset;
7876 * Copy the data based upon the caller version
7878 rx_MarshallProcessRPCStats(callerVersion,
7879 rpc_stat->stats[0].func_total,
7880 rpc_stat->stats, &ptr);
7886 MUTEX_EXIT(&rx_rpc_stats);
7891 * rx_FreeRPCStats - free memory allocated by
7892 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7896 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7897 * rx_RetrievePeerRPCStats
7899 * IN allocSize - the number of bytes in stats.
7907 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7909 rxi_Free(stats, allocSize);
7913 * rx_queryProcessRPCStats - see if process rpc stat collection is
7914 * currently enabled.
7920 * Returns 0 if stats are not enabled != 0 otherwise
7924 rx_queryProcessRPCStats(void)
7927 MUTEX_ENTER(&rx_rpc_stats);
7928 rc = rxi_monitor_processStats;
7929 MUTEX_EXIT(&rx_rpc_stats);
7934 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7940 * Returns 0 if stats are not enabled != 0 otherwise
7944 rx_queryPeerRPCStats(void)
7947 MUTEX_ENTER(&rx_rpc_stats);
7948 rc = rxi_monitor_peerStats;
7949 MUTEX_EXIT(&rx_rpc_stats);
7954 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7964 rx_enableProcessRPCStats(void)
7966 MUTEX_ENTER(&rx_rpc_stats);
7967 rx_enable_stats = 1;
7968 rxi_monitor_processStats = 1;
7969 MUTEX_EXIT(&rx_rpc_stats);
7973 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7983 rx_enablePeerRPCStats(void)
7985 MUTEX_ENTER(&rx_rpc_stats);
7986 rx_enable_stats = 1;
7987 rxi_monitor_peerStats = 1;
7988 MUTEX_EXIT(&rx_rpc_stats);
7992 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8002 rx_disableProcessRPCStats(void)
8004 rx_interface_stat_p rpc_stat, nrpc_stat;
8007 MUTEX_ENTER(&rx_rpc_stats);
8010 * Turn off process statistics and if peer stats is also off, turn
8014 rxi_monitor_processStats = 0;
8015 if (rxi_monitor_peerStats == 0) {
8016 rx_enable_stats = 0;
8019 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8020 unsigned int num_funcs = 0;
8023 queue_Remove(rpc_stat);
8024 num_funcs = rpc_stat->stats[0].func_total;
8026 sizeof(rx_interface_stat_t) +
8027 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8029 rxi_Free(rpc_stat, space);
8030 rxi_rpc_process_stat_cnt -= num_funcs;
8032 MUTEX_EXIT(&rx_rpc_stats);
8036 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8046 rx_disablePeerRPCStats(void)
8048 struct rx_peer **peer_ptr, **peer_end;
8051 MUTEX_ENTER(&rx_rpc_stats);
8054 * Turn off peer statistics and if process stats is also off, turn
8058 rxi_monitor_peerStats = 0;
8059 if (rxi_monitor_processStats == 0) {
8060 rx_enable_stats = 0;
8063 MUTEX_ENTER(&rx_peerHashTable_lock);
8064 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8065 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8067 struct rx_peer *peer, *next, *prev;
8068 for (prev = peer = *peer_ptr; peer; peer = next) {
8070 code = MUTEX_TRYENTER(&peer->peer_lock);
8072 rx_interface_stat_p rpc_stat, nrpc_stat;
8075 (&peer->rpcStats, rpc_stat, nrpc_stat,
8076 rx_interface_stat)) {
8077 unsigned int num_funcs = 0;
8080 queue_Remove(&rpc_stat->queue_header);
8081 queue_Remove(&rpc_stat->all_peers);
8082 num_funcs = rpc_stat->stats[0].func_total;
8084 sizeof(rx_interface_stat_t) +
8085 rpc_stat->stats[0].func_total *
8086 sizeof(rx_function_entry_v1_t);
8088 rxi_Free(rpc_stat, space);
8089 rxi_rpc_peer_stat_cnt -= num_funcs;
8091 MUTEX_EXIT(&peer->peer_lock);
8092 if (prev == *peer_ptr) {
8102 MUTEX_EXIT(&rx_peerHashTable_lock);
8103 MUTEX_EXIT(&rx_rpc_stats);
8107 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8112 * IN clearFlag - flag indicating which stats to clear
8120 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8122 rx_interface_stat_p rpc_stat, nrpc_stat;
8124 MUTEX_ENTER(&rx_rpc_stats);
8126 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8127 unsigned int num_funcs = 0, i;
8128 num_funcs = rpc_stat->stats[0].func_total;
8129 for (i = 0; i < num_funcs; i++) {
8130 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8131 hzero(rpc_stat->stats[i].invocations);
8133 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8134 hzero(rpc_stat->stats[i].bytes_sent);
8136 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8137 hzero(rpc_stat->stats[i].bytes_rcvd);
8139 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8140 rpc_stat->stats[i].queue_time_sum.sec = 0;
8141 rpc_stat->stats[i].queue_time_sum.usec = 0;
8143 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8144 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8145 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8147 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8148 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8149 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8151 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8152 rpc_stat->stats[i].queue_time_max.sec = 0;
8153 rpc_stat->stats[i].queue_time_max.usec = 0;
8155 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8156 rpc_stat->stats[i].execution_time_sum.sec = 0;
8157 rpc_stat->stats[i].execution_time_sum.usec = 0;
8159 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8160 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8161 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8163 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8164 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8165 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8167 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8168 rpc_stat->stats[i].execution_time_max.sec = 0;
8169 rpc_stat->stats[i].execution_time_max.usec = 0;
8174 MUTEX_EXIT(&rx_rpc_stats);
8178 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8183 * IN clearFlag - flag indicating which stats to clear
8191 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8193 rx_interface_stat_p rpc_stat, nrpc_stat;
8195 MUTEX_ENTER(&rx_rpc_stats);
8197 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8198 unsigned int num_funcs = 0, i;
8201 * We have to fix the offset of rpc_stat since we are
8202 * keeping this structure on two rx_queues. The rx_queue
8203 * package assumes that the rx_queue member is the first
8204 * member of the structure. That is, rx_queue assumes that
8205 * any one item is only on one queue at a time. We are
8206 * breaking that assumption and so we have to do a little
8207 * math to fix our pointers.
8210 fix_offset = (char *)rpc_stat;
8211 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8212 rpc_stat = (rx_interface_stat_p) fix_offset;
8214 num_funcs = rpc_stat->stats[0].func_total;
8215 for (i = 0; i < num_funcs; i++) {
8216 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8217 hzero(rpc_stat->stats[i].invocations);
8219 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8220 hzero(rpc_stat->stats[i].bytes_sent);
8222 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8223 hzero(rpc_stat->stats[i].bytes_rcvd);
8225 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8226 rpc_stat->stats[i].queue_time_sum.sec = 0;
8227 rpc_stat->stats[i].queue_time_sum.usec = 0;
8229 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8230 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8231 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8233 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8234 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8235 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8237 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8238 rpc_stat->stats[i].queue_time_max.sec = 0;
8239 rpc_stat->stats[i].queue_time_max.usec = 0;
8241 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8242 rpc_stat->stats[i].execution_time_sum.sec = 0;
8243 rpc_stat->stats[i].execution_time_sum.usec = 0;
8245 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8246 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8247 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8249 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8250 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8251 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8253 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8254 rpc_stat->stats[i].execution_time_max.sec = 0;
8255 rpc_stat->stats[i].execution_time_max.usec = 0;
8260 MUTEX_EXIT(&rx_rpc_stats);
8264 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8265 * is authorized to enable/disable/clear RX statistics.
8267 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8270 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8272 rxi_rxstat_userok = proc;
8276 rx_RxStatUserOk(struct rx_call *call)
8278 if (!rxi_rxstat_userok)
8280 return rxi_rxstat_userok(call);
8285 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8286 * function in the MSVC runtime DLL (msvcrt.dll).
8288 * Note: the system serializes calls to this function.
8291 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8292 DWORD reason, /* reason function is being called */
8293 LPVOID reserved) /* reserved for future use */
8296 case DLL_PROCESS_ATTACH:
8297 /* library is being attached to a process */
8301 case DLL_PROCESS_DETACH:
8308 #endif /* AFS_NT40_ENV */
8311 int rx_DumpCalls(FILE *outputFile, char *cookie)
8313 #ifdef RXDEBUG_PACKET
8314 #ifdef KDUMP_RX_LOCK
8315 struct rx_call_rx_lock *c;
8322 #define RXDPRINTF sprintf
8323 #define RXDPRINTOUT output
8325 #define RXDPRINTF fprintf
8326 #define RXDPRINTOUT outputFile
8329 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8331 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8334 for (c = rx_allCallsp; c; c = c->allNextp) {
8335 u_short rqc, tqc, iovqc;
8336 struct rx_packet *p, *np;
8338 MUTEX_ENTER(&c->lock);
8339 queue_Count(&c->rq, p, np, rx_packet, rqc);
8340 queue_Count(&c->tq, p, np, rx_packet, tqc);
8341 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8343 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, "
8344 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8345 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8346 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8347 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8348 #ifdef RX_ENABLE_LOCKS
8351 #ifdef RX_REFCOUNT_CHECK
8352 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
8353 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
8356 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,
8357 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
8358 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
8359 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
8360 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
8361 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
8362 #ifdef RX_ENABLE_LOCKS
8363 , (afs_uint32)c->refCount
8365 #ifdef RX_REFCOUNT_CHECK
8366 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
8369 MUTEX_EXIT(&c->lock);
8372 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8375 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
8377 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8379 #endif /* RXDEBUG_PACKET */