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 interface 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);
1153 * Check if there are others waiting for a new call.
1154 * If so, let them go first to avoid starving them.
1155 * This is a fairly simple scheme, and might not be
1156 * a complete solution for large numbers of waiters.
1158 * makeCallWaiters keeps track of the number of
1159 * threads waiting to make calls and the
1160 * RX_CONN_MAKECALL_WAITING flag bit is used to
1161 * indicate that there are indeed calls waiting.
1162 * The flag is set when the waiter is incremented.
1163 * It is only cleared when makeCallWaiters is 0.
1164 * This prevents us from accidently destroying the
1165 * connection while it is potentially about to be used.
1167 MUTEX_ENTER(&conn->conn_call_lock);
1168 MUTEX_ENTER(&conn->conn_data_lock);
1169 if (conn->makeCallWaiters) {
1170 conn->makeCallWaiters++;
1171 MUTEX_EXIT(&conn->conn_data_lock);
1173 #ifdef RX_ENABLE_LOCKS
1174 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1178 MUTEX_ENTER(&conn->conn_data_lock);
1179 conn->makeCallWaiters--;
1180 if (conn->makeCallWaiters == 0)
1181 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1183 MUTEX_EXIT(&conn->conn_data_lock);
1186 for (i = 0; i < RX_MAXCALLS; i++) {
1187 call = conn->call[i];
1189 if (call->state == RX_STATE_DALLY) {
1190 MUTEX_ENTER(&call->lock);
1191 if (call->state == RX_STATE_DALLY) {
1192 call->state = RX_STATE_RESET;
1193 MUTEX_EXIT(&conn->conn_call_lock);
1194 rxi_ResetCall(call, 0);
1195 MUTEX_ENTER(&conn->conn_call_lock);
1196 (*call->callNumber)++;
1199 MUTEX_EXIT(&call->lock);
1202 /* rxi_NewCall returns with mutex locked */
1203 call = rxi_NewCall(conn, i);
1207 if (i < RX_MAXCALLS) {
1210 MUTEX_ENTER(&conn->conn_data_lock);
1211 conn->flags |= RX_CONN_MAKECALL_WAITING;
1212 conn->makeCallWaiters++;
1213 MUTEX_EXIT(&conn->conn_data_lock);
1215 #ifdef RX_ENABLE_LOCKS
1216 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1220 MUTEX_ENTER(&conn->conn_data_lock);
1221 conn->makeCallWaiters--;
1222 if (conn->makeCallWaiters == 0)
1223 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1224 MUTEX_EXIT(&conn->conn_data_lock);
1227 * Wake up anyone else who might be giving us a chance to
1228 * run (see code above that avoids resource starvation).
1230 #ifdef RX_ENABLE_LOCKS
1231 CV_BROADCAST(&conn->conn_call_cv);
1236 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1238 /* Client is initially in send mode */
1239 call->state = RX_STATE_ACTIVE;
1240 call->error = conn->error;
1242 call->mode = RX_MODE_ERROR;
1244 call->mode = RX_MODE_SENDING;
1246 /* remember start time for call in case we have hard dead time limit */
1247 call->queueTime = queueTime;
1248 clock_GetTime(&call->startTime);
1249 hzero(call->bytesSent);
1250 hzero(call->bytesRcvd);
1252 /* Turn on busy protocol. */
1253 rxi_KeepAliveOn(call);
1254 MUTEX_EXIT(&conn->conn_call_lock);
1256 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1257 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1258 osi_Panic("rx_NewCall call about to be used without an empty tq");
1260 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1262 MUTEX_EXIT(&call->lock);
1265 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1270 rxi_HasActiveCalls(struct rx_connection *aconn)
1273 struct rx_call *tcall;
1277 for (i = 0; i < RX_MAXCALLS; i++) {
1278 if ((tcall = aconn->call[i])) {
1279 if ((tcall->state == RX_STATE_ACTIVE)
1280 || (tcall->state == RX_STATE_PRECALL)) {
1291 rxi_GetCallNumberVector(struct rx_connection *aconn,
1292 afs_int32 * aint32s)
1295 struct rx_call *tcall;
1299 for (i = 0; i < RX_MAXCALLS; i++) {
1300 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1301 aint32s[i] = aconn->callNumber[i] + 1;
1303 aint32s[i] = aconn->callNumber[i];
1310 rxi_SetCallNumberVector(struct rx_connection *aconn,
1311 afs_int32 * aint32s)
1314 struct rx_call *tcall;
1318 for (i = 0; i < RX_MAXCALLS; i++) {
1319 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1320 aconn->callNumber[i] = aint32s[i] - 1;
1322 aconn->callNumber[i] = aint32s[i];
1328 /* Advertise a new service. A service is named locally by a UDP port
1329 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1332 char *serviceName; Name for identification purposes (e.g. the
1333 service name might be used for probing for
1336 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1337 char *serviceName, struct rx_securityClass **securityObjects,
1338 int nSecurityObjects,
1339 afs_int32(*serviceProc) (struct rx_call * acall))
1341 osi_socket socket = OSI_NULLSOCKET;
1342 struct rx_service *tservice;
1348 if (serviceId == 0) {
1350 "rx_NewService: service id for service %s is not non-zero.\n",
1357 "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",
1365 tservice = rxi_AllocService();
1367 for (i = 0; i < RX_MAX_SERVICES; i++) {
1368 struct rx_service *service = rx_services[i];
1370 if (port == service->servicePort && host == service->serviceHost) {
1371 if (service->serviceId == serviceId) {
1372 /* The identical service has already been
1373 * installed; if the caller was intending to
1374 * change the security classes used by this
1375 * service, he/she loses. */
1377 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1378 serviceName, serviceId, service->serviceName);
1380 rxi_FreeService(tservice);
1383 /* Different service, same port: re-use the socket
1384 * which is bound to the same port */
1385 socket = service->socket;
1388 if (socket == OSI_NULLSOCKET) {
1389 /* If we don't already have a socket (from another
1390 * service on same port) get a new one */
1391 socket = rxi_GetHostUDPSocket(host, port);
1392 if (socket == OSI_NULLSOCKET) {
1394 rxi_FreeService(tservice);
1399 service->socket = socket;
1400 service->serviceHost = host;
1401 service->servicePort = port;
1402 service->serviceId = serviceId;
1403 service->serviceName = serviceName;
1404 service->nSecurityObjects = nSecurityObjects;
1405 service->securityObjects = securityObjects;
1406 service->minProcs = 0;
1407 service->maxProcs = 1;
1408 service->idleDeadTime = 60;
1409 service->idleDeadErr = 0;
1410 service->connDeadTime = rx_connDeadTime;
1411 service->executeRequestProc = serviceProc;
1412 service->checkReach = 0;
1413 rx_services[i] = service; /* not visible until now */
1419 rxi_FreeService(tservice);
1420 (osi_Msg "rx_NewService: cannot support > %d services\n",
1425 /* Set configuration options for all of a service's security objects */
1428 rx_SetSecurityConfiguration(struct rx_service *service,
1429 rx_securityConfigVariables type,
1433 for (i = 0; i<service->nSecurityObjects; i++) {
1434 if (service->securityObjects[i]) {
1435 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1443 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1444 struct rx_securityClass **securityObjects, int nSecurityObjects,
1445 afs_int32(*serviceProc) (struct rx_call * acall))
1447 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1450 /* Generic request processing loop. This routine should be called
1451 * by the implementation dependent rx_ServerProc. If socketp is
1452 * non-null, it will be set to the file descriptor that this thread
1453 * is now listening on. If socketp is null, this routine will never
1456 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1458 struct rx_call *call;
1460 struct rx_service *tservice = NULL;
1467 call = rx_GetCall(threadID, tservice, socketp);
1468 if (socketp && *socketp != OSI_NULLSOCKET) {
1469 /* We are now a listener thread */
1474 /* if server is restarting( typically smooth shutdown) then do not
1475 * allow any new calls.
1478 if (rx_tranquil && (call != NULL)) {
1482 MUTEX_ENTER(&call->lock);
1484 rxi_CallError(call, RX_RESTARTING);
1485 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1487 MUTEX_EXIT(&call->lock);
1491 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1492 #ifdef RX_ENABLE_LOCKS
1494 #endif /* RX_ENABLE_LOCKS */
1495 afs_termState = AFSOP_STOP_AFS;
1496 afs_osi_Wakeup(&afs_termState);
1497 #ifdef RX_ENABLE_LOCKS
1499 #endif /* RX_ENABLE_LOCKS */
1504 tservice = call->conn->service;
1506 if (tservice->beforeProc)
1507 (*tservice->beforeProc) (call);
1509 code = call->conn->service->executeRequestProc(call);
1511 if (tservice->afterProc)
1512 (*tservice->afterProc) (call, code);
1514 rx_EndCall(call, code);
1515 if (rx_stats_active) {
1516 MUTEX_ENTER(&rx_stats_mutex);
1518 MUTEX_EXIT(&rx_stats_mutex);
1525 rx_WakeupServerProcs(void)
1527 struct rx_serverQueueEntry *np, *tqp;
1531 MUTEX_ENTER(&rx_serverPool_lock);
1533 #ifdef RX_ENABLE_LOCKS
1534 if (rx_waitForPacket)
1535 CV_BROADCAST(&rx_waitForPacket->cv);
1536 #else /* RX_ENABLE_LOCKS */
1537 if (rx_waitForPacket)
1538 osi_rxWakeup(rx_waitForPacket);
1539 #endif /* RX_ENABLE_LOCKS */
1540 MUTEX_ENTER(&freeSQEList_lock);
1541 for (np = rx_FreeSQEList; np; np = tqp) {
1542 tqp = *(struct rx_serverQueueEntry **)np;
1543 #ifdef RX_ENABLE_LOCKS
1544 CV_BROADCAST(&np->cv);
1545 #else /* RX_ENABLE_LOCKS */
1547 #endif /* RX_ENABLE_LOCKS */
1549 MUTEX_EXIT(&freeSQEList_lock);
1550 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1551 #ifdef RX_ENABLE_LOCKS
1552 CV_BROADCAST(&np->cv);
1553 #else /* RX_ENABLE_LOCKS */
1555 #endif /* RX_ENABLE_LOCKS */
1557 MUTEX_EXIT(&rx_serverPool_lock);
1562 * One thing that seems to happen is that all the server threads get
1563 * tied up on some empty or slow call, and then a whole bunch of calls
1564 * arrive at once, using up the packet pool, so now there are more
1565 * empty calls. The most critical resources here are server threads
1566 * and the free packet pool. The "doreclaim" code seems to help in
1567 * general. I think that eventually we arrive in this state: there
1568 * are lots of pending calls which do have all their packets present,
1569 * so they won't be reclaimed, are multi-packet calls, so they won't
1570 * be scheduled until later, and thus are tying up most of the free
1571 * packet pool for a very long time.
1573 * 1. schedule multi-packet calls if all the packets are present.
1574 * Probably CPU-bound operation, useful to return packets to pool.
1575 * Do what if there is a full window, but the last packet isn't here?
1576 * 3. preserve one thread which *only* runs "best" calls, otherwise
1577 * it sleeps and waits for that type of call.
1578 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1579 * the current dataquota business is badly broken. The quota isn't adjusted
1580 * to reflect how many packets are presently queued for a running call.
1581 * So, when we schedule a queued call with a full window of packets queued
1582 * up for it, that *should* free up a window full of packets for other 2d-class
1583 * calls to be able to use from the packet pool. But it doesn't.
1585 * NB. Most of the time, this code doesn't run -- since idle server threads
1586 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1587 * as a new call arrives.
1589 /* Sleep until a call arrives. Returns a pointer to the call, ready
1590 * for an rx_Read. */
1591 #ifdef RX_ENABLE_LOCKS
1593 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1595 struct rx_serverQueueEntry *sq;
1596 struct rx_call *call = (struct rx_call *)0;
1597 struct rx_service *service = NULL;
1600 MUTEX_ENTER(&freeSQEList_lock);
1602 if ((sq = rx_FreeSQEList)) {
1603 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1604 MUTEX_EXIT(&freeSQEList_lock);
1605 } else { /* otherwise allocate a new one and return that */
1606 MUTEX_EXIT(&freeSQEList_lock);
1607 sq = (struct rx_serverQueueEntry *)
1608 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1609 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1610 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1613 MUTEX_ENTER(&rx_serverPool_lock);
1614 if (cur_service != NULL) {
1615 ReturnToServerPool(cur_service);
1618 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1619 struct rx_call *tcall, *ncall, *choice2 = NULL;
1621 /* Scan for eligible incoming calls. A call is not eligible
1622 * if the maximum number of calls for its service type are
1623 * already executing */
1624 /* One thread will process calls FCFS (to prevent starvation),
1625 * while the other threads may run ahead looking for calls which
1626 * have all their input data available immediately. This helps
1627 * keep threads from blocking, waiting for data from the client. */
1628 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1629 service = tcall->conn->service;
1630 if (!QuotaOK(service)) {
1633 MUTEX_ENTER(&rx_pthread_mutex);
1634 if (tno == rxi_fcfs_thread_num
1635 || !tcall->queue_item_header.next) {
1636 MUTEX_EXIT(&rx_pthread_mutex);
1637 /* If we're the fcfs thread , then we'll just use
1638 * this call. If we haven't been able to find an optimal
1639 * choice, and we're at the end of the list, then use a
1640 * 2d choice if one has been identified. Otherwise... */
1641 call = (choice2 ? choice2 : tcall);
1642 service = call->conn->service;
1644 MUTEX_EXIT(&rx_pthread_mutex);
1645 if (!queue_IsEmpty(&tcall->rq)) {
1646 struct rx_packet *rp;
1647 rp = queue_First(&tcall->rq, rx_packet);
1648 if (rp->header.seq == 1) {
1650 || (rp->header.flags & RX_LAST_PACKET)) {
1652 } else if (rxi_2dchoice && !choice2
1653 && !(tcall->flags & RX_CALL_CLEARED)
1654 && (tcall->rprev > rxi_HardAckRate)) {
1664 ReturnToServerPool(service);
1671 MUTEX_EXIT(&rx_serverPool_lock);
1672 MUTEX_ENTER(&call->lock);
1674 if (call->flags & RX_CALL_WAIT_PROC) {
1675 call->flags &= ~RX_CALL_WAIT_PROC;
1676 MUTEX_ENTER(&rx_waiting_mutex);
1678 MUTEX_EXIT(&rx_waiting_mutex);
1681 if (call->state != RX_STATE_PRECALL || call->error) {
1682 MUTEX_EXIT(&call->lock);
1683 MUTEX_ENTER(&rx_serverPool_lock);
1684 ReturnToServerPool(service);
1689 if (queue_IsEmpty(&call->rq)
1690 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1691 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1693 CLEAR_CALL_QUEUE_LOCK(call);
1696 /* If there are no eligible incoming calls, add this process
1697 * to the idle server queue, to wait for one */
1701 *socketp = OSI_NULLSOCKET;
1703 sq->socketp = socketp;
1704 queue_Append(&rx_idleServerQueue, sq);
1705 #ifndef AFS_AIX41_ENV
1706 rx_waitForPacket = sq;
1708 rx_waitingForPacket = sq;
1709 #endif /* AFS_AIX41_ENV */
1711 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1713 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1714 MUTEX_EXIT(&rx_serverPool_lock);
1715 return (struct rx_call *)0;
1718 } while (!(call = sq->newcall)
1719 && !(socketp && *socketp != OSI_NULLSOCKET));
1720 MUTEX_EXIT(&rx_serverPool_lock);
1722 MUTEX_ENTER(&call->lock);
1728 MUTEX_ENTER(&freeSQEList_lock);
1729 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1730 rx_FreeSQEList = sq;
1731 MUTEX_EXIT(&freeSQEList_lock);
1734 clock_GetTime(&call->startTime);
1735 call->state = RX_STATE_ACTIVE;
1736 call->mode = RX_MODE_RECEIVING;
1737 #ifdef RX_KERNEL_TRACE
1738 if (ICL_SETACTIVE(afs_iclSetp)) {
1739 int glockOwner = ISAFS_GLOCK();
1742 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1743 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1750 rxi_calltrace(RX_CALL_START, call);
1751 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
1752 call->conn->service->servicePort, call->conn->service->serviceId,
1755 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1756 MUTEX_EXIT(&call->lock);
1758 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1763 #else /* RX_ENABLE_LOCKS */
1765 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1767 struct rx_serverQueueEntry *sq;
1768 struct rx_call *call = (struct rx_call *)0, *choice2;
1769 struct rx_service *service = NULL;
1773 MUTEX_ENTER(&freeSQEList_lock);
1775 if ((sq = rx_FreeSQEList)) {
1776 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1777 MUTEX_EXIT(&freeSQEList_lock);
1778 } else { /* otherwise allocate a new one and return that */
1779 MUTEX_EXIT(&freeSQEList_lock);
1780 sq = (struct rx_serverQueueEntry *)
1781 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1782 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1783 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1785 MUTEX_ENTER(&sq->lock);
1787 if (cur_service != NULL) {
1788 cur_service->nRequestsRunning--;
1789 if (cur_service->nRequestsRunning < cur_service->minProcs)
1793 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1794 struct rx_call *tcall, *ncall;
1795 /* Scan for eligible incoming calls. A call is not eligible
1796 * if the maximum number of calls for its service type are
1797 * already executing */
1798 /* One thread will process calls FCFS (to prevent starvation),
1799 * while the other threads may run ahead looking for calls which
1800 * have all their input data available immediately. This helps
1801 * keep threads from blocking, waiting for data from the client. */
1802 choice2 = (struct rx_call *)0;
1803 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1804 service = tcall->conn->service;
1805 if (QuotaOK(service)) {
1806 MUTEX_ENTER(&rx_pthread_mutex);
1807 if (tno == rxi_fcfs_thread_num
1808 || !tcall->queue_item_header.next) {
1809 MUTEX_EXIT(&rx_pthread_mutex);
1810 /* If we're the fcfs thread, then we'll just use
1811 * this call. If we haven't been able to find an optimal
1812 * choice, and we're at the end of the list, then use a
1813 * 2d choice if one has been identified. Otherwise... */
1814 call = (choice2 ? choice2 : tcall);
1815 service = call->conn->service;
1817 MUTEX_EXIT(&rx_pthread_mutex);
1818 if (!queue_IsEmpty(&tcall->rq)) {
1819 struct rx_packet *rp;
1820 rp = queue_First(&tcall->rq, rx_packet);
1821 if (rp->header.seq == 1
1823 || (rp->header.flags & RX_LAST_PACKET))) {
1825 } else if (rxi_2dchoice && !choice2
1826 && !(tcall->flags & RX_CALL_CLEARED)
1827 && (tcall->rprev > rxi_HardAckRate)) {
1841 /* we can't schedule a call if there's no data!!! */
1842 /* send an ack if there's no data, if we're missing the
1843 * first packet, or we're missing something between first
1844 * and last -- there's a "hole" in the incoming data. */
1845 if (queue_IsEmpty(&call->rq)
1846 || queue_First(&call->rq, rx_packet)->header.seq != 1
1847 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1848 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1850 call->flags &= (~RX_CALL_WAIT_PROC);
1851 service->nRequestsRunning++;
1852 /* just started call in minProcs pool, need fewer to maintain
1854 if (service->nRequestsRunning <= service->minProcs)
1858 /* MUTEX_EXIT(&call->lock); */
1860 /* If there are no eligible incoming calls, add this process
1861 * to the idle server queue, to wait for one */
1864 *socketp = OSI_NULLSOCKET;
1866 sq->socketp = socketp;
1867 queue_Append(&rx_idleServerQueue, sq);
1871 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1873 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1874 return (struct rx_call *)0;
1877 } while (!(call = sq->newcall)
1878 && !(socketp && *socketp != OSI_NULLSOCKET));
1880 MUTEX_EXIT(&sq->lock);
1882 MUTEX_ENTER(&freeSQEList_lock);
1883 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1884 rx_FreeSQEList = sq;
1885 MUTEX_EXIT(&freeSQEList_lock);
1888 clock_GetTime(&call->startTime);
1889 call->state = RX_STATE_ACTIVE;
1890 call->mode = RX_MODE_RECEIVING;
1891 #ifdef RX_KERNEL_TRACE
1892 if (ICL_SETACTIVE(afs_iclSetp)) {
1893 int glockOwner = ISAFS_GLOCK();
1896 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1897 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1904 rxi_calltrace(RX_CALL_START, call);
1905 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
1906 call->conn->service->servicePort, call->conn->service->serviceId,
1909 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1916 #endif /* RX_ENABLE_LOCKS */
1920 /* Establish a procedure to be called when a packet arrives for a
1921 * call. This routine will be called at most once after each call,
1922 * and will also be called if there is an error condition on the or
1923 * the call is complete. Used by multi rx to build a selection
1924 * function which determines which of several calls is likely to be a
1925 * good one to read from.
1926 * NOTE: the way this is currently implemented it is probably only a
1927 * good idea to (1) use it immediately after a newcall (clients only)
1928 * and (2) only use it once. Other uses currently void your warranty
1931 rx_SetArrivalProc(struct rx_call *call,
1932 void (*proc) (struct rx_call * call,
1935 void * handle, int arg)
1937 call->arrivalProc = proc;
1938 call->arrivalProcHandle = handle;
1939 call->arrivalProcArg = arg;
1942 /* Call is finished (possibly prematurely). Return rc to the peer, if
1943 * appropriate, and return the final error code from the conversation
1947 rx_EndCall(struct rx_call *call, afs_int32 rc)
1949 struct rx_connection *conn = call->conn;
1950 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 MUTEX_ENTER(&conn->conn_data_lock);
2023 conn->flags |= RX_CONN_BUSY;
2024 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2025 MUTEX_EXIT(&conn->conn_data_lock);
2026 #ifdef RX_ENABLE_LOCKS
2027 CV_BROADCAST(&conn->conn_call_cv);
2032 #ifdef RX_ENABLE_LOCKS
2034 MUTEX_EXIT(&conn->conn_data_lock);
2036 #endif /* RX_ENABLE_LOCKS */
2037 call->state = RX_STATE_DALLY;
2039 error = call->error;
2041 /* currentPacket, nLeft, and NFree must be zeroed here, because
2042 * ResetCall cannot: ResetCall may be called at splnet(), in the
2043 * kernel version, and may interrupt the macros rx_Read or
2044 * rx_Write, which run at normal priority for efficiency. */
2045 if (call->currentPacket) {
2046 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2047 rxi_FreePacket(call->currentPacket);
2048 call->currentPacket = (struct rx_packet *)0;
2051 call->nLeft = call->nFree = call->curlen = 0;
2053 /* Free any packets from the last call to ReadvProc/WritevProc */
2054 #ifdef RXDEBUG_PACKET
2056 #endif /* RXDEBUG_PACKET */
2057 rxi_FreePackets(0, &call->iovq);
2059 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2060 MUTEX_EXIT(&call->lock);
2061 if (conn->type == RX_CLIENT_CONNECTION) {
2062 conn->flags &= ~RX_CONN_BUSY;
2066 * Map errors to the local host's errno.h format.
2068 error = ntoh_syserr_conv(error);
2072 #if !defined(KERNEL)
2074 /* Call this routine when shutting down a server or client (especially
2075 * clients). This will allow Rx to gracefully garbage collect server
2076 * connections, and reduce the number of retries that a server might
2077 * make to a dead client.
2078 * This is not quite right, since some calls may still be ongoing and
2079 * we can't lock them to destroy them. */
2083 struct rx_connection **conn_ptr, **conn_end;
2087 if (rxinit_status == 1) {
2089 return; /* Already shutdown. */
2091 rxi_DeleteCachedConnections();
2092 if (rx_connHashTable) {
2093 MUTEX_ENTER(&rx_connHashTable_lock);
2094 for (conn_ptr = &rx_connHashTable[0], conn_end =
2095 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2097 struct rx_connection *conn, *next;
2098 for (conn = *conn_ptr; conn; conn = next) {
2100 if (conn->type == RX_CLIENT_CONNECTION) {
2101 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2103 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2104 #ifdef RX_ENABLE_LOCKS
2105 rxi_DestroyConnectionNoLock(conn);
2106 #else /* RX_ENABLE_LOCKS */
2107 rxi_DestroyConnection(conn);
2108 #endif /* RX_ENABLE_LOCKS */
2112 #ifdef RX_ENABLE_LOCKS
2113 while (rx_connCleanup_list) {
2114 struct rx_connection *conn;
2115 conn = rx_connCleanup_list;
2116 rx_connCleanup_list = rx_connCleanup_list->next;
2117 MUTEX_EXIT(&rx_connHashTable_lock);
2118 rxi_CleanupConnection(conn);
2119 MUTEX_ENTER(&rx_connHashTable_lock);
2121 MUTEX_EXIT(&rx_connHashTable_lock);
2122 #endif /* RX_ENABLE_LOCKS */
2127 afs_winsockCleanup();
2135 /* if we wakeup packet waiter too often, can get in loop with two
2136 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2138 rxi_PacketsUnWait(void)
2140 if (!rx_waitingForPackets) {
2144 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2145 return; /* still over quota */
2148 rx_waitingForPackets = 0;
2149 #ifdef RX_ENABLE_LOCKS
2150 CV_BROADCAST(&rx_waitingForPackets_cv);
2152 osi_rxWakeup(&rx_waitingForPackets);
2158 /* ------------------Internal interfaces------------------------- */
2160 /* Return this process's service structure for the
2161 * specified socket and service */
2163 rxi_FindService(osi_socket socket, u_short serviceId)
2165 struct rx_service **sp;
2166 for (sp = &rx_services[0]; *sp; sp++) {
2167 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2173 #ifdef RXDEBUG_PACKET
2174 #ifdef KDUMP_RX_LOCK
2175 static struct rx_call_rx_lock *rx_allCallsp = 0;
2177 static struct rx_call *rx_allCallsp = 0;
2179 #endif /* RXDEBUG_PACKET */
2181 /* Allocate a call structure, for the indicated channel of the
2182 * supplied connection. The mode and state of the call must be set by
2183 * the caller. Returns the call with mutex locked. */
2185 rxi_NewCall(struct rx_connection *conn, int channel)
2187 struct rx_call *call;
2188 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2189 struct rx_call *cp; /* Call pointer temp */
2190 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2191 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2193 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2195 /* Grab an existing call structure, or allocate a new one.
2196 * Existing call structures are assumed to have been left reset by
2198 MUTEX_ENTER(&rx_freeCallQueue_lock);
2200 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2202 * EXCEPT that the TQ might not yet be cleared out.
2203 * Skip over those with in-use TQs.
2206 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2207 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2213 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2214 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2215 call = queue_First(&rx_freeCallQueue, rx_call);
2216 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2218 if (rx_stats_active)
2219 rx_MutexDecrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2220 MUTEX_EXIT(&rx_freeCallQueue_lock);
2221 MUTEX_ENTER(&call->lock);
2222 CLEAR_CALL_QUEUE_LOCK(call);
2223 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2224 /* Now, if TQ wasn't cleared earlier, do it now. */
2225 rxi_WaitforTQBusy(call);
2226 if (call->flags & RX_CALL_TQ_CLEARME) {
2227 rxi_ClearTransmitQueue(call, 1);
2228 /*queue_Init(&call->tq);*/
2230 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2231 /* Bind the call to its connection structure */
2233 rxi_ResetCall(call, 1);
2236 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2237 #ifdef RXDEBUG_PACKET
2238 call->allNextp = rx_allCallsp;
2239 rx_allCallsp = call;
2241 #endif /* RXDEBUG_PACKET */
2242 rx_MutexIncrement(rx_stats.nCallStructs, rx_stats_mutex);
2244 MUTEX_EXIT(&rx_freeCallQueue_lock);
2245 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2246 MUTEX_ENTER(&call->lock);
2247 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2248 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2249 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2251 /* Initialize once-only items */
2252 queue_Init(&call->tq);
2253 queue_Init(&call->rq);
2254 queue_Init(&call->iovq);
2255 #ifdef RXDEBUG_PACKET
2256 call->rqc = call->tqc = call->iovqc = 0;
2257 #endif /* RXDEBUG_PACKET */
2258 /* Bind the call to its connection structure (prereq for reset) */
2260 rxi_ResetCall(call, 1);
2262 call->channel = channel;
2263 call->callNumber = &conn->callNumber[channel];
2264 call->rwind = conn->rwind[channel];
2265 call->twind = conn->twind[channel];
2266 /* Note that the next expected call number is retained (in
2267 * conn->callNumber[i]), even if we reallocate the call structure
2269 conn->call[channel] = call;
2270 /* if the channel's never been used (== 0), we should start at 1, otherwise
2271 * the call number is valid from the last time this channel was used */
2272 if (*call->callNumber == 0)
2273 *call->callNumber = 1;
2278 /* A call has been inactive long enough that so we can throw away
2279 * state, including the call structure, which is placed on the call
2281 * Call is locked upon entry.
2282 * haveCTLock set if called from rxi_ReapConnections
2284 #ifdef RX_ENABLE_LOCKS
2286 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2287 #else /* RX_ENABLE_LOCKS */
2289 rxi_FreeCall(struct rx_call *call)
2290 #endif /* RX_ENABLE_LOCKS */
2292 int channel = call->channel;
2293 struct rx_connection *conn = call->conn;
2296 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2297 (*call->callNumber)++;
2298 rxi_ResetCall(call, 0);
2299 call->conn->call[channel] = (struct rx_call *)0;
2301 MUTEX_ENTER(&rx_freeCallQueue_lock);
2302 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2303 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2304 /* A call may be free even though its transmit queue is still in use.
2305 * Since we search the call list from head to tail, put busy calls at
2306 * the head of the list, and idle calls at the tail.
2308 if (call->flags & RX_CALL_TQ_BUSY)
2309 queue_Prepend(&rx_freeCallQueue, call);
2311 queue_Append(&rx_freeCallQueue, call);
2312 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2313 queue_Append(&rx_freeCallQueue, call);
2314 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2315 if (rx_stats_active)
2316 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2317 MUTEX_EXIT(&rx_freeCallQueue_lock);
2319 /* Destroy the connection if it was previously slated for
2320 * destruction, i.e. the Rx client code previously called
2321 * rx_DestroyConnection (client connections), or
2322 * rxi_ReapConnections called the same routine (server
2323 * connections). Only do this, however, if there are no
2324 * outstanding calls. Note that for fine grain locking, there appears
2325 * to be a deadlock in that rxi_FreeCall has a call locked and
2326 * DestroyConnectionNoLock locks each call in the conn. But note a
2327 * few lines up where we have removed this call from the conn.
2328 * If someone else destroys a connection, they either have no
2329 * call lock held or are going through this section of code.
2331 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2332 MUTEX_ENTER(&conn->conn_data_lock);
2334 MUTEX_EXIT(&conn->conn_data_lock);
2335 #ifdef RX_ENABLE_LOCKS
2337 rxi_DestroyConnectionNoLock(conn);
2339 rxi_DestroyConnection(conn);
2340 #else /* RX_ENABLE_LOCKS */
2341 rxi_DestroyConnection(conn);
2342 #endif /* RX_ENABLE_LOCKS */
2346 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2348 rxi_Alloc(size_t size)
2352 if (rx_stats_active)
2353 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2356 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2357 afs_osi_Alloc_NoSleep(size);
2362 osi_Panic("rxi_Alloc error");
2368 rxi_Free(void *addr, size_t size)
2370 if (rx_stats_active)
2371 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2372 osi_Free(addr, size);
2376 rxi_SetPeerMtu(afs_uint32 host, afs_uint32 port, int mtu)
2378 struct rx_peer **peer_ptr, **peer_end;
2381 MUTEX_ENTER(&rx_peerHashTable_lock);
2383 for (peer_ptr = &rx_peerHashTable[0], peer_end =
2384 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
2386 struct rx_peer *peer, *next;
2387 for (peer = *peer_ptr; peer; peer = next) {
2389 if (host == peer->host) {
2390 MUTEX_ENTER(&peer->peer_lock);
2391 peer->ifMTU=MIN(mtu, peer->ifMTU);
2392 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2393 MUTEX_EXIT(&peer->peer_lock);
2398 struct rx_peer *peer;
2399 hashIndex = PEER_HASH(host, port);
2400 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2401 if ((peer->host == host) && (peer->port == port)) {
2402 MUTEX_ENTER(&peer->peer_lock);
2403 peer->ifMTU=MIN(mtu, peer->ifMTU);
2404 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2405 MUTEX_EXIT(&peer->peer_lock);
2409 MUTEX_EXIT(&rx_peerHashTable_lock);
2412 /* Find the peer process represented by the supplied (host,port)
2413 * combination. If there is no appropriate active peer structure, a
2414 * new one will be allocated and initialized
2415 * The origPeer, if set, is a pointer to a peer structure on which the
2416 * refcount will be be decremented. This is used to replace the peer
2417 * structure hanging off a connection structure */
2419 rxi_FindPeer(afs_uint32 host, u_short port,
2420 struct rx_peer *origPeer, int create)
2424 hashIndex = PEER_HASH(host, port);
2425 MUTEX_ENTER(&rx_peerHashTable_lock);
2426 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2427 if ((pp->host == host) && (pp->port == port))
2432 pp = rxi_AllocPeer(); /* This bzero's *pp */
2433 pp->host = host; /* set here or in InitPeerParams is zero */
2435 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2436 queue_Init(&pp->congestionQueue);
2437 queue_Init(&pp->rpcStats);
2438 pp->next = rx_peerHashTable[hashIndex];
2439 rx_peerHashTable[hashIndex] = pp;
2440 rxi_InitPeerParams(pp);
2441 if (rx_stats_active)
2442 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2449 origPeer->refCount--;
2450 MUTEX_EXIT(&rx_peerHashTable_lock);
2455 /* Find the connection at (host, port) started at epoch, and with the
2456 * given connection id. Creates the server connection if necessary.
2457 * The type specifies whether a client connection or a server
2458 * connection is desired. In both cases, (host, port) specify the
2459 * peer's (host, pair) pair. Client connections are not made
2460 * automatically by this routine. The parameter socket gives the
2461 * socket descriptor on which the packet was received. This is used,
2462 * in the case of server connections, to check that *new* connections
2463 * come via a valid (port, serviceId). Finally, the securityIndex
2464 * parameter must match the existing index for the connection. If a
2465 * server connection is created, it will be created using the supplied
2466 * index, if the index is valid for this service */
2467 struct rx_connection *
2468 rxi_FindConnection(osi_socket socket, afs_int32 host,
2469 u_short port, u_short serviceId, afs_uint32 cid,
2470 afs_uint32 epoch, int type, u_int securityIndex)
2472 int hashindex, flag, i;
2473 struct rx_connection *conn;
2474 hashindex = CONN_HASH(host, port, cid, epoch, type);
2475 MUTEX_ENTER(&rx_connHashTable_lock);
2476 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2477 rx_connHashTable[hashindex],
2480 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2481 && (epoch == conn->epoch)) {
2482 struct rx_peer *pp = conn->peer;
2483 if (securityIndex != conn->securityIndex) {
2484 /* this isn't supposed to happen, but someone could forge a packet
2485 * like this, and there seems to be some CM bug that makes this
2486 * happen from time to time -- in which case, the fileserver
2488 MUTEX_EXIT(&rx_connHashTable_lock);
2489 return (struct rx_connection *)0;
2491 if (pp->host == host && pp->port == port)
2493 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2495 /* So what happens when it's a callback connection? */
2496 if ( /*type == RX_CLIENT_CONNECTION && */
2497 (conn->epoch & 0x80000000))
2501 /* the connection rxLastConn that was used the last time is not the
2502 ** one we are looking for now. Hence, start searching in the hash */
2504 conn = rx_connHashTable[hashindex];
2509 struct rx_service *service;
2510 if (type == RX_CLIENT_CONNECTION) {
2511 MUTEX_EXIT(&rx_connHashTable_lock);
2512 return (struct rx_connection *)0;
2514 service = rxi_FindService(socket, serviceId);
2515 if (!service || (securityIndex >= service->nSecurityObjects)
2516 || (service->securityObjects[securityIndex] == 0)) {
2517 MUTEX_EXIT(&rx_connHashTable_lock);
2518 return (struct rx_connection *)0;
2520 conn = rxi_AllocConnection(); /* This bzero's the connection */
2521 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2522 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2523 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2524 conn->next = rx_connHashTable[hashindex];
2525 rx_connHashTable[hashindex] = conn;
2526 conn->peer = rxi_FindPeer(host, port, 0, 1);
2527 conn->type = RX_SERVER_CONNECTION;
2528 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2529 conn->epoch = epoch;
2530 conn->cid = cid & RX_CIDMASK;
2531 /* conn->serial = conn->lastSerial = 0; */
2532 /* conn->timeout = 0; */
2533 conn->ackRate = RX_FAST_ACK_RATE;
2534 conn->service = service;
2535 conn->serviceId = serviceId;
2536 conn->securityIndex = securityIndex;
2537 conn->securityObject = service->securityObjects[securityIndex];
2538 conn->nSpecific = 0;
2539 conn->specific = NULL;
2540 rx_SetConnDeadTime(conn, service->connDeadTime);
2541 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2542 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2543 for (i = 0; i < RX_MAXCALLS; i++) {
2544 conn->twind[i] = rx_initSendWindow;
2545 conn->rwind[i] = rx_initReceiveWindow;
2547 /* Notify security object of the new connection */
2548 RXS_NewConnection(conn->securityObject, conn);
2549 /* XXXX Connection timeout? */
2550 if (service->newConnProc)
2551 (*service->newConnProc) (conn);
2552 if (rx_stats_active)
2553 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2556 MUTEX_ENTER(&conn->conn_data_lock);
2558 MUTEX_EXIT(&conn->conn_data_lock);
2560 rxLastConn = conn; /* store this connection as the last conn used */
2561 MUTEX_EXIT(&rx_connHashTable_lock);
2565 /* There are two packet tracing routines available for testing and monitoring
2566 * Rx. One is called just after every packet is received and the other is
2567 * called just before every packet is sent. Received packets, have had their
2568 * headers decoded, and packets to be sent have not yet had their headers
2569 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2570 * containing the network address. Both can be modified. The return value, if
2571 * non-zero, indicates that the packet should be dropped. */
2573 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2574 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2576 /* A packet has been received off the interface. Np is the packet, socket is
2577 * the socket number it was received from (useful in determining which service
2578 * this packet corresponds to), and (host, port) reflect the host,port of the
2579 * sender. This call returns the packet to the caller if it is finished with
2580 * it, rather than de-allocating it, just as a small performance hack */
2583 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
2584 afs_uint32 host, u_short port, int *tnop,
2585 struct rx_call **newcallp)
2587 struct rx_call *call;
2588 struct rx_connection *conn;
2590 afs_uint32 currentCallNumber;
2596 struct rx_packet *tnp;
2599 /* We don't print out the packet until now because (1) the time may not be
2600 * accurate enough until now in the lwp implementation (rx_Listener only gets
2601 * the time after the packet is read) and (2) from a protocol point of view,
2602 * this is the first time the packet has been seen */
2603 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2604 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2605 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT,
2606 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2607 np->header.epoch, np->header.cid, np->header.callNumber,
2608 np->header.seq, np->header.flags, np));
2611 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2612 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2615 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2616 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2619 /* If an input tracer function is defined, call it with the packet and
2620 * network address. Note this function may modify its arguments. */
2621 if (rx_justReceived) {
2622 struct sockaddr_in addr;
2624 addr.sin_family = AF_INET;
2625 addr.sin_port = port;
2626 addr.sin_addr.s_addr = host;
2627 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2628 addr.sin_len = sizeof(addr);
2629 #endif /* AFS_OSF_ENV */
2630 drop = (*rx_justReceived) (np, &addr);
2631 /* drop packet if return value is non-zero */
2634 port = addr.sin_port; /* in case fcn changed addr */
2635 host = addr.sin_addr.s_addr;
2639 /* If packet was not sent by the client, then *we* must be the client */
2640 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2641 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2643 /* Find the connection (or fabricate one, if we're the server & if
2644 * necessary) associated with this packet */
2646 rxi_FindConnection(socket, host, port, np->header.serviceId,
2647 np->header.cid, np->header.epoch, type,
2648 np->header.securityIndex);
2651 /* If no connection found or fabricated, just ignore the packet.
2652 * (An argument could be made for sending an abort packet for
2657 MUTEX_ENTER(&conn->conn_data_lock);
2658 if (conn->maxSerial < np->header.serial)
2659 conn->maxSerial = np->header.serial;
2660 MUTEX_EXIT(&conn->conn_data_lock);
2662 /* If the connection is in an error state, send an abort packet and ignore
2663 * the incoming packet */
2665 /* Don't respond to an abort packet--we don't want loops! */
2666 MUTEX_ENTER(&conn->conn_data_lock);
2667 if (np->header.type != RX_PACKET_TYPE_ABORT)
2668 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2670 MUTEX_EXIT(&conn->conn_data_lock);
2674 /* Check for connection-only requests (i.e. not call specific). */
2675 if (np->header.callNumber == 0) {
2676 switch (np->header.type) {
2677 case RX_PACKET_TYPE_ABORT: {
2678 /* What if the supplied error is zero? */
2679 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2680 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2681 rxi_ConnectionError(conn, errcode);
2682 MUTEX_ENTER(&conn->conn_data_lock);
2684 MUTEX_EXIT(&conn->conn_data_lock);
2687 case RX_PACKET_TYPE_CHALLENGE:
2688 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2689 MUTEX_ENTER(&conn->conn_data_lock);
2691 MUTEX_EXIT(&conn->conn_data_lock);
2693 case RX_PACKET_TYPE_RESPONSE:
2694 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2695 MUTEX_ENTER(&conn->conn_data_lock);
2697 MUTEX_EXIT(&conn->conn_data_lock);
2699 case RX_PACKET_TYPE_PARAMS:
2700 case RX_PACKET_TYPE_PARAMS + 1:
2701 case RX_PACKET_TYPE_PARAMS + 2:
2702 /* ignore these packet types for now */
2703 MUTEX_ENTER(&conn->conn_data_lock);
2705 MUTEX_EXIT(&conn->conn_data_lock);
2710 /* Should not reach here, unless the peer is broken: send an
2712 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2713 MUTEX_ENTER(&conn->conn_data_lock);
2714 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2716 MUTEX_EXIT(&conn->conn_data_lock);
2721 channel = np->header.cid & RX_CHANNELMASK;
2722 call = conn->call[channel];
2723 #ifdef RX_ENABLE_LOCKS
2725 MUTEX_ENTER(&call->lock);
2726 /* Test to see if call struct is still attached to conn. */
2727 if (call != conn->call[channel]) {
2729 MUTEX_EXIT(&call->lock);
2730 if (type == RX_SERVER_CONNECTION) {
2731 call = conn->call[channel];
2732 /* If we started with no call attached and there is one now,
2733 * another thread is also running this routine and has gotten
2734 * the connection channel. We should drop this packet in the tests
2735 * below. If there was a call on this connection and it's now
2736 * gone, then we'll be making a new call below.
2737 * If there was previously a call and it's now different then
2738 * the old call was freed and another thread running this routine
2739 * has created a call on this channel. One of these two threads
2740 * has a packet for the old call and the code below handles those
2744 MUTEX_ENTER(&call->lock);
2746 /* This packet can't be for this call. If the new call address is
2747 * 0 then no call is running on this channel. If there is a call
2748 * then, since this is a client connection we're getting data for
2749 * it must be for the previous call.
2751 if (rx_stats_active)
2752 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2753 MUTEX_ENTER(&conn->conn_data_lock);
2755 MUTEX_EXIT(&conn->conn_data_lock);
2760 currentCallNumber = conn->callNumber[channel];
2762 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2763 if (np->header.callNumber < currentCallNumber) {
2764 if (rx_stats_active)
2765 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2766 #ifdef RX_ENABLE_LOCKS
2768 MUTEX_EXIT(&call->lock);
2770 MUTEX_ENTER(&conn->conn_data_lock);
2772 MUTEX_EXIT(&conn->conn_data_lock);
2776 MUTEX_ENTER(&conn->conn_call_lock);
2777 call = rxi_NewCall(conn, channel);
2778 MUTEX_EXIT(&conn->conn_call_lock);
2779 *call->callNumber = np->header.callNumber;
2781 if (np->header.callNumber == 0)
2782 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%.06d len %d",
2783 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
2784 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
2785 np->header.flags, np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2787 call->state = RX_STATE_PRECALL;
2788 clock_GetTime(&call->queueTime);
2789 hzero(call->bytesSent);
2790 hzero(call->bytesRcvd);
2792 * If the number of queued calls exceeds the overload
2793 * threshold then abort this call.
2795 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2796 struct rx_packet *tp;
2798 rxi_CallError(call, rx_BusyError);
2799 tp = rxi_SendCallAbort(call, np, 1, 0);
2800 MUTEX_EXIT(&call->lock);
2801 MUTEX_ENTER(&conn->conn_data_lock);
2803 MUTEX_EXIT(&conn->conn_data_lock);
2804 if (rx_stats_active)
2805 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2808 rxi_KeepAliveOn(call);
2809 } else if (np->header.callNumber != currentCallNumber) {
2810 /* Wait until the transmit queue is idle before deciding
2811 * whether to reset the current call. Chances are that the
2812 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2815 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2816 while ((call->state == RX_STATE_ACTIVE)
2817 && (call->flags & RX_CALL_TQ_BUSY)) {
2818 call->flags |= RX_CALL_TQ_WAIT;
2820 #ifdef RX_ENABLE_LOCKS
2821 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2822 CV_WAIT(&call->cv_tq, &call->lock);
2823 #else /* RX_ENABLE_LOCKS */
2824 osi_rxSleep(&call->tq);
2825 #endif /* RX_ENABLE_LOCKS */
2827 if (call->tqWaiters == 0)
2828 call->flags &= ~RX_CALL_TQ_WAIT;
2830 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2831 /* If the new call cannot be taken right now send a busy and set
2832 * the error condition in this call, so that it terminates as
2833 * quickly as possible */
2834 if (call->state == RX_STATE_ACTIVE) {
2835 struct rx_packet *tp;
2837 rxi_CallError(call, RX_CALL_DEAD);
2838 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2840 MUTEX_EXIT(&call->lock);
2841 MUTEX_ENTER(&conn->conn_data_lock);
2843 MUTEX_EXIT(&conn->conn_data_lock);
2846 rxi_ResetCall(call, 0);
2847 *call->callNumber = np->header.callNumber;
2849 if (np->header.callNumber == 0)
2850 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%06d len %d",
2851 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
2852 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
2853 np->header.flags, np, np->retryTime.sec, np->retryTime.usec, np->length));
2855 call->state = RX_STATE_PRECALL;
2856 clock_GetTime(&call->queueTime);
2857 hzero(call->bytesSent);
2858 hzero(call->bytesRcvd);
2860 * If the number of queued calls exceeds the overload
2861 * threshold then abort this call.
2863 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2864 struct rx_packet *tp;
2866 rxi_CallError(call, rx_BusyError);
2867 tp = rxi_SendCallAbort(call, np, 1, 0);
2868 MUTEX_EXIT(&call->lock);
2869 MUTEX_ENTER(&conn->conn_data_lock);
2871 MUTEX_EXIT(&conn->conn_data_lock);
2872 if (rx_stats_active)
2873 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2876 rxi_KeepAliveOn(call);
2878 /* Continuing call; do nothing here. */
2880 } else { /* we're the client */
2881 /* Ignore all incoming acknowledgements for calls in DALLY state */
2882 if (call && (call->state == RX_STATE_DALLY)
2883 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2884 if (rx_stats_active)
2885 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
2886 #ifdef RX_ENABLE_LOCKS
2888 MUTEX_EXIT(&call->lock);
2891 MUTEX_ENTER(&conn->conn_data_lock);
2893 MUTEX_EXIT(&conn->conn_data_lock);
2897 /* Ignore anything that's not relevant to the current call. If there
2898 * isn't a current call, then no packet is relevant. */
2899 if (!call || (np->header.callNumber != currentCallNumber)) {
2900 if (rx_stats_active)
2901 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2902 #ifdef RX_ENABLE_LOCKS
2904 MUTEX_EXIT(&call->lock);
2907 MUTEX_ENTER(&conn->conn_data_lock);
2909 MUTEX_EXIT(&conn->conn_data_lock);
2912 /* If the service security object index stamped in the packet does not
2913 * match the connection's security index, ignore the packet */
2914 if (np->header.securityIndex != conn->securityIndex) {
2915 #ifdef RX_ENABLE_LOCKS
2916 MUTEX_EXIT(&call->lock);
2918 MUTEX_ENTER(&conn->conn_data_lock);
2920 MUTEX_EXIT(&conn->conn_data_lock);
2924 /* If we're receiving the response, then all transmit packets are
2925 * implicitly acknowledged. Get rid of them. */
2926 if (np->header.type == RX_PACKET_TYPE_DATA) {
2927 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2928 /* XXX Hack. Because we must release the global rx lock when
2929 * sending packets (osi_NetSend) we drop all acks while we're
2930 * traversing the tq in rxi_Start sending packets out because
2931 * packets may move to the freePacketQueue as result of being here!
2932 * So we drop these packets until we're safely out of the
2933 * traversing. Really ugly!
2934 * For fine grain RX locking, we set the acked field in the
2935 * packets and let rxi_Start remove them from the transmit queue.
2937 if (call->flags & RX_CALL_TQ_BUSY) {
2938 #ifdef RX_ENABLE_LOCKS
2939 rxi_SetAcksInTransmitQueue(call);
2942 return np; /* xmitting; drop packet */
2945 rxi_ClearTransmitQueue(call, 0);
2947 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2948 rxi_ClearTransmitQueue(call, 0);
2949 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2951 if (np->header.type == RX_PACKET_TYPE_ACK) {
2952 /* now check to see if this is an ack packet acknowledging that the
2953 * server actually *lost* some hard-acked data. If this happens we
2954 * ignore this packet, as it may indicate that the server restarted in
2955 * the middle of a call. It is also possible that this is an old ack
2956 * packet. We don't abort the connection in this case, because this
2957 * *might* just be an old ack packet. The right way to detect a server
2958 * restart in the midst of a call is to notice that the server epoch
2960 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2961 * XXX unacknowledged. I think that this is off-by-one, but
2962 * XXX I don't dare change it just yet, since it will
2963 * XXX interact badly with the server-restart detection
2964 * XXX code in receiveackpacket. */
2965 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2966 if (rx_stats_active)
2967 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2968 MUTEX_EXIT(&call->lock);
2969 MUTEX_ENTER(&conn->conn_data_lock);
2971 MUTEX_EXIT(&conn->conn_data_lock);
2975 } /* else not a data packet */
2978 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2979 /* Set remote user defined status from packet */
2980 call->remoteStatus = np->header.userStatus;
2982 /* Note the gap between the expected next packet and the actual
2983 * packet that arrived, when the new packet has a smaller serial number
2984 * than expected. Rioses frequently reorder packets all by themselves,
2985 * so this will be quite important with very large window sizes.
2986 * Skew is checked against 0 here to avoid any dependence on the type of
2987 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2989 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2990 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2991 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2993 MUTEX_ENTER(&conn->conn_data_lock);
2994 skew = conn->lastSerial - np->header.serial;
2995 conn->lastSerial = np->header.serial;
2996 MUTEX_EXIT(&conn->conn_data_lock);
2998 struct rx_peer *peer;
3000 if (skew > peer->inPacketSkew) {
3001 dpf(("*** In skew changed from %d to %d\n",
3002 peer->inPacketSkew, skew));
3003 peer->inPacketSkew = skew;
3007 /* Now do packet type-specific processing */
3008 switch (np->header.type) {
3009 case RX_PACKET_TYPE_DATA:
3010 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3013 case RX_PACKET_TYPE_ACK:
3014 /* Respond immediately to ack packets requesting acknowledgement
3016 if (np->header.flags & RX_REQUEST_ACK) {
3018 (void)rxi_SendCallAbort(call, 0, 1, 0);
3020 (void)rxi_SendAck(call, 0, np->header.serial,
3021 RX_ACK_PING_RESPONSE, 1);
3023 np = rxi_ReceiveAckPacket(call, np, 1);
3025 case RX_PACKET_TYPE_ABORT: {
3026 /* An abort packet: reset the call, passing the error up to the user. */
3027 /* What if error is zero? */
3028 /* What if the error is -1? the application will treat it as a timeout. */
3029 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3030 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
3031 rxi_CallError(call, errdata);
3032 MUTEX_EXIT(&call->lock);
3033 MUTEX_ENTER(&conn->conn_data_lock);
3035 MUTEX_EXIT(&conn->conn_data_lock);
3036 return np; /* xmitting; drop packet */
3038 case RX_PACKET_TYPE_BUSY:
3041 case RX_PACKET_TYPE_ACKALL:
3042 /* All packets acknowledged, so we can drop all packets previously
3043 * readied for sending */
3044 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3045 /* XXX Hack. We because we can't release the global rx lock when
3046 * sending packets (osi_NetSend) we drop all ack pkts while we're
3047 * traversing the tq in rxi_Start sending packets out because
3048 * packets may move to the freePacketQueue as result of being
3049 * here! So we drop these packets until we're safely out of the
3050 * traversing. Really ugly!
3051 * For fine grain RX locking, we set the acked field in the packets
3052 * and let rxi_Start remove the packets from the transmit queue.
3054 if (call->flags & RX_CALL_TQ_BUSY) {
3055 #ifdef RX_ENABLE_LOCKS
3056 rxi_SetAcksInTransmitQueue(call);
3058 #else /* RX_ENABLE_LOCKS */
3059 MUTEX_EXIT(&call->lock);
3060 MUTEX_ENTER(&conn->conn_data_lock);
3062 MUTEX_EXIT(&conn->conn_data_lock);
3063 return np; /* xmitting; drop packet */
3064 #endif /* RX_ENABLE_LOCKS */
3066 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3067 rxi_ClearTransmitQueue(call, 0);
3068 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3071 /* Should not reach here, unless the peer is broken: send an abort
3073 rxi_CallError(call, RX_PROTOCOL_ERROR);
3074 np = rxi_SendCallAbort(call, np, 1, 0);
3077 /* Note when this last legitimate packet was received, for keep-alive
3078 * processing. Note, we delay getting the time until now in the hope that
3079 * the packet will be delivered to the user before any get time is required
3080 * (if not, then the time won't actually be re-evaluated here). */
3081 call->lastReceiveTime = clock_Sec();
3082 MUTEX_EXIT(&call->lock);
3083 MUTEX_ENTER(&conn->conn_data_lock);
3085 MUTEX_EXIT(&conn->conn_data_lock);
3089 /* return true if this is an "interesting" connection from the point of view
3090 of someone trying to debug the system */
3092 rxi_IsConnInteresting(struct rx_connection *aconn)
3095 struct rx_call *tcall;
3097 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3099 for (i = 0; i < RX_MAXCALLS; i++) {
3100 tcall = aconn->call[i];
3102 if ((tcall->state == RX_STATE_PRECALL)
3103 || (tcall->state == RX_STATE_ACTIVE))
3105 if ((tcall->mode == RX_MODE_SENDING)
3106 || (tcall->mode == RX_MODE_RECEIVING))
3114 /* if this is one of the last few packets AND it wouldn't be used by the
3115 receiving call to immediately satisfy a read request, then drop it on
3116 the floor, since accepting it might prevent a lock-holding thread from
3117 making progress in its reading. If a call has been cleared while in
3118 the precall state then ignore all subsequent packets until the call
3119 is assigned to a thread. */
3122 TooLow(struct rx_packet *ap, struct rx_call *acall)
3126 MUTEX_ENTER(&rx_quota_mutex);
3127 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3128 && (acall->state == RX_STATE_PRECALL))
3129 || ((rx_nFreePackets < rxi_dataQuota + 2)
3130 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3131 && (acall->flags & RX_CALL_READER_WAIT)))) {
3134 MUTEX_EXIT(&rx_quota_mutex);
3140 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3142 struct rx_connection *conn = arg1;
3143 struct rx_call *acall = arg2;
3144 struct rx_call *call = acall;
3145 struct clock when, now;
3148 MUTEX_ENTER(&conn->conn_data_lock);
3149 conn->checkReachEvent = NULL;
3150 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3153 MUTEX_EXIT(&conn->conn_data_lock);
3157 MUTEX_ENTER(&conn->conn_call_lock);
3158 MUTEX_ENTER(&conn->conn_data_lock);
3159 for (i = 0; i < RX_MAXCALLS; i++) {
3160 struct rx_call *tc = conn->call[i];
3161 if (tc && tc->state == RX_STATE_PRECALL) {
3167 /* Indicate that rxi_CheckReachEvent is no longer running by
3168 * clearing the flag. Must be atomic under conn_data_lock to
3169 * avoid a new call slipping by: rxi_CheckConnReach holds
3170 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3172 conn->flags &= ~RX_CONN_ATTACHWAIT;
3173 MUTEX_EXIT(&conn->conn_data_lock);
3174 MUTEX_EXIT(&conn->conn_call_lock);
3179 MUTEX_ENTER(&call->lock);
3180 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3182 MUTEX_EXIT(&call->lock);
3184 clock_GetTime(&now);
3186 when.sec += RX_CHECKREACH_TIMEOUT;
3187 MUTEX_ENTER(&conn->conn_data_lock);
3188 if (!conn->checkReachEvent) {
3190 conn->checkReachEvent =
3191 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3194 MUTEX_EXIT(&conn->conn_data_lock);
3200 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3202 struct rx_service *service = conn->service;
3203 struct rx_peer *peer = conn->peer;
3204 afs_uint32 now, lastReach;
3206 if (service->checkReach == 0)
3210 MUTEX_ENTER(&peer->peer_lock);
3211 lastReach = peer->lastReachTime;
3212 MUTEX_EXIT(&peer->peer_lock);
3213 if (now - lastReach < RX_CHECKREACH_TTL)
3216 MUTEX_ENTER(&conn->conn_data_lock);
3217 if (conn->flags & RX_CONN_ATTACHWAIT) {
3218 MUTEX_EXIT(&conn->conn_data_lock);
3221 conn->flags |= RX_CONN_ATTACHWAIT;
3222 MUTEX_EXIT(&conn->conn_data_lock);
3223 if (!conn->checkReachEvent)
3224 rxi_CheckReachEvent(NULL, conn, call);
3229 /* try to attach call, if authentication is complete */
3231 TryAttach(struct rx_call *acall, osi_socket socket,
3232 int *tnop, struct rx_call **newcallp,
3235 struct rx_connection *conn = acall->conn;
3237 if (conn->type == RX_SERVER_CONNECTION
3238 && acall->state == RX_STATE_PRECALL) {
3239 /* Don't attach until we have any req'd. authentication. */
3240 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3241 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3242 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3243 /* Note: this does not necessarily succeed; there
3244 * may not any proc available
3247 rxi_ChallengeOn(acall->conn);
3252 /* A data packet has been received off the interface. This packet is
3253 * appropriate to the call (the call is in the right state, etc.). This
3254 * routine can return a packet to the caller, for re-use */
3257 rxi_ReceiveDataPacket(struct rx_call *call,
3258 struct rx_packet *np, int istack,
3259 osi_socket socket, afs_uint32 host, u_short port,
3260 int *tnop, struct rx_call **newcallp)
3262 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3267 afs_uint32 serial=0, flags=0;
3269 struct rx_packet *tnp;
3270 struct clock when, now;
3271 if (rx_stats_active)
3272 rx_MutexIncrement(rx_stats.dataPacketsRead, rx_stats_mutex);
3275 /* If there are no packet buffers, drop this new packet, unless we can find
3276 * packet buffers from inactive calls */
3278 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3279 MUTEX_ENTER(&rx_freePktQ_lock);
3280 rxi_NeedMorePackets = TRUE;
3281 MUTEX_EXIT(&rx_freePktQ_lock);
3282 if (rx_stats_active)
3283 rx_MutexIncrement(rx_stats.noPacketBuffersOnRead, rx_stats_mutex);
3284 call->rprev = np->header.serial;
3285 rxi_calltrace(RX_TRACE_DROP, call);
3286 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems", np));
3288 rxi_ClearReceiveQueue(call);
3289 clock_GetTime(&now);
3291 clock_Add(&when, &rx_softAckDelay);
3292 if (!call->delayedAckEvent
3293 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3294 rxevent_Cancel(call->delayedAckEvent, call,
3295 RX_CALL_REFCOUNT_DELAY);
3296 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3297 call->delayedAckEvent =
3298 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3300 /* we've damaged this call already, might as well do it in. */
3306 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3307 * packet is one of several packets transmitted as a single
3308 * datagram. Do not send any soft or hard acks until all packets
3309 * in a jumbogram have been processed. Send negative acks right away.
3311 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3312 /* tnp is non-null when there are more packets in the
3313 * current jumbo gram */
3320 seq = np->header.seq;
3321 serial = np->header.serial;
3322 flags = np->header.flags;
3324 /* If the call is in an error state, send an abort message */
3326 return rxi_SendCallAbort(call, np, istack, 0);
3328 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3329 * AFS 3.5 jumbogram. */
3330 if (flags & RX_JUMBO_PACKET) {
3331 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3336 if (np->header.spare != 0) {
3337 MUTEX_ENTER(&call->conn->conn_data_lock);
3338 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3339 MUTEX_EXIT(&call->conn->conn_data_lock);
3342 /* The usual case is that this is the expected next packet */
3343 if (seq == call->rnext) {
3345 /* Check to make sure it is not a duplicate of one already queued */
3346 if (queue_IsNotEmpty(&call->rq)
3347 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3348 if (rx_stats_active)
3349 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3350 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate", np));
3351 rxevent_Cancel(call->delayedAckEvent, call,
3352 RX_CALL_REFCOUNT_DELAY);
3353 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3359 /* It's the next packet. Stick it on the receive queue
3360 * for this call. Set newPackets to make sure we wake
3361 * the reader once all packets have been processed */
3362 np->flags |= RX_PKTFLAG_RQ;
3363 queue_Prepend(&call->rq, np);
3364 #ifdef RXDEBUG_PACKET
3366 #endif /* RXDEBUG_PACKET */
3368 np = NULL; /* We can't use this anymore */
3371 /* If an ack is requested then set a flag to make sure we
3372 * send an acknowledgement for this packet */
3373 if (flags & RX_REQUEST_ACK) {
3374 ackNeeded = RX_ACK_REQUESTED;
3377 /* Keep track of whether we have received the last packet */
3378 if (flags & RX_LAST_PACKET) {
3379 call->flags |= RX_CALL_HAVE_LAST;
3383 /* Check whether we have all of the packets for this call */
3384 if (call->flags & RX_CALL_HAVE_LAST) {
3385 afs_uint32 tseq; /* temporary sequence number */
3386 struct rx_packet *tp; /* Temporary packet pointer */
3387 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3389 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3390 if (tseq != tp->header.seq)
3392 if (tp->header.flags & RX_LAST_PACKET) {
3393 call->flags |= RX_CALL_RECEIVE_DONE;
3400 /* Provide asynchronous notification for those who want it
3401 * (e.g. multi rx) */
3402 if (call->arrivalProc) {
3403 (*call->arrivalProc) (call, call->arrivalProcHandle,
3404 call->arrivalProcArg);
3405 call->arrivalProc = (void (*)())0;
3408 /* Update last packet received */
3411 /* If there is no server process serving this call, grab
3412 * one, if available. We only need to do this once. If a
3413 * server thread is available, this thread becomes a server
3414 * thread and the server thread becomes a listener thread. */
3416 TryAttach(call, socket, tnop, newcallp, 0);
3419 /* This is not the expected next packet. */
3421 /* Determine whether this is a new or old packet, and if it's
3422 * a new one, whether it fits into the current receive window.
3423 * Also figure out whether the packet was delivered in sequence.
3424 * We use the prev variable to determine whether the new packet
3425 * is the successor of its immediate predecessor in the
3426 * receive queue, and the missing flag to determine whether
3427 * any of this packets predecessors are missing. */
3429 afs_uint32 prev; /* "Previous packet" sequence number */
3430 struct rx_packet *tp; /* Temporary packet pointer */
3431 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3432 int missing; /* Are any predecessors missing? */
3434 /* If the new packet's sequence number has been sent to the
3435 * application already, then this is a duplicate */
3436 if (seq < call->rnext) {
3437 if (rx_stats_active)
3438 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3439 rxevent_Cancel(call->delayedAckEvent, call,
3440 RX_CALL_REFCOUNT_DELAY);
3441 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3447 /* If the sequence number is greater than what can be
3448 * accomodated by the current window, then send a negative
3449 * acknowledge and drop the packet */
3450 if ((call->rnext + call->rwind) <= seq) {
3451 rxevent_Cancel(call->delayedAckEvent, call,
3452 RX_CALL_REFCOUNT_DELAY);
3453 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3460 /* Look for the packet in the queue of old received packets */
3461 for (prev = call->rnext - 1, missing =
3462 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3463 /*Check for duplicate packet */
3464 if (seq == tp->header.seq) {
3465 if (rx_stats_active)
3466 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3467 rxevent_Cancel(call->delayedAckEvent, call,
3468 RX_CALL_REFCOUNT_DELAY);
3469 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3475 /* If we find a higher sequence packet, break out and
3476 * insert the new packet here. */
3477 if (seq < tp->header.seq)
3479 /* Check for missing packet */
3480 if (tp->header.seq != prev + 1) {
3484 prev = tp->header.seq;
3487 /* Keep track of whether we have received the last packet. */
3488 if (flags & RX_LAST_PACKET) {
3489 call->flags |= RX_CALL_HAVE_LAST;
3492 /* It's within the window: add it to the the receive queue.
3493 * tp is left by the previous loop either pointing at the
3494 * packet before which to insert the new packet, or at the
3495 * queue head if the queue is empty or the packet should be
3497 np->flags |= RX_PKTFLAG_RQ;
3498 #ifdef RXDEBUG_PACKET
3500 #endif /* RXDEBUG_PACKET */
3501 queue_InsertBefore(tp, np);
3505 /* Check whether we have all of the packets for this call */
3506 if ((call->flags & RX_CALL_HAVE_LAST)
3507 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3508 afs_uint32 tseq; /* temporary sequence number */
3511 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3512 if (tseq != tp->header.seq)
3514 if (tp->header.flags & RX_LAST_PACKET) {
3515 call->flags |= RX_CALL_RECEIVE_DONE;
3522 /* We need to send an ack of the packet is out of sequence,
3523 * or if an ack was requested by the peer. */
3524 if (seq != prev + 1 || missing) {
3525 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3526 } else if (flags & RX_REQUEST_ACK) {
3527 ackNeeded = RX_ACK_REQUESTED;
3530 /* Acknowledge the last packet for each call */
3531 if (flags & RX_LAST_PACKET) {
3542 * If the receiver is waiting for an iovec, fill the iovec
3543 * using the data from the receive queue */
3544 if (call->flags & RX_CALL_IOVEC_WAIT) {
3545 didHardAck = rxi_FillReadVec(call, serial);
3546 /* the call may have been aborted */
3555 /* Wakeup the reader if any */
3556 if ((call->flags & RX_CALL_READER_WAIT)
3557 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3558 || (call->iovNext >= call->iovMax)
3559 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3560 call->flags &= ~RX_CALL_READER_WAIT;
3561 #ifdef RX_ENABLE_LOCKS
3562 CV_BROADCAST(&call->cv_rq);
3564 osi_rxWakeup(&call->rq);
3570 * Send an ack when requested by the peer, or once every
3571 * rxi_SoftAckRate packets until the last packet has been
3572 * received. Always send a soft ack for the last packet in
3573 * the server's reply. */
3575 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3576 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3577 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3578 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3579 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3580 } else if (call->nSoftAcks) {
3581 clock_GetTime(&now);
3583 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3584 clock_Add(&when, &rx_lastAckDelay);
3586 clock_Add(&when, &rx_softAckDelay);
3588 if (!call->delayedAckEvent
3589 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3590 rxevent_Cancel(call->delayedAckEvent, call,
3591 RX_CALL_REFCOUNT_DELAY);
3592 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3593 call->delayedAckEvent =
3594 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3596 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3597 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3604 static void rxi_ComputeRate();
3608 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3610 struct rx_peer *peer = conn->peer;
3612 MUTEX_ENTER(&peer->peer_lock);
3613 peer->lastReachTime = clock_Sec();
3614 MUTEX_EXIT(&peer->peer_lock);
3616 MUTEX_ENTER(&conn->conn_data_lock);
3617 if (conn->flags & RX_CONN_ATTACHWAIT) {
3620 conn->flags &= ~RX_CONN_ATTACHWAIT;
3621 MUTEX_EXIT(&conn->conn_data_lock);
3623 for (i = 0; i < RX_MAXCALLS; i++) {
3624 struct rx_call *call = conn->call[i];
3627 MUTEX_ENTER(&call->lock);
3628 /* tnop can be null if newcallp is null */
3629 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3631 MUTEX_EXIT(&call->lock);
3635 MUTEX_EXIT(&conn->conn_data_lock);
3638 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
3640 rx_ack_reason(int reason)
3643 case RX_ACK_REQUESTED:
3645 case RX_ACK_DUPLICATE:
3647 case RX_ACK_OUT_OF_SEQUENCE:
3649 case RX_ACK_EXCEEDS_WINDOW:
3651 case RX_ACK_NOSPACE:
3655 case RX_ACK_PING_RESPONSE:
3668 /* rxi_ComputePeerNetStats
3670 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3671 * estimates (like RTT and throughput) based on ack packets. Caller
3672 * must ensure that the packet in question is the right one (i.e.
3673 * serial number matches).
3676 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3677 struct rx_ackPacket *ap, struct rx_packet *np)
3679 struct rx_peer *peer = call->conn->peer;
3681 /* Use RTT if not delayed by client and
3682 * ignore packets that were retransmitted. */
3683 if (!(p->flags & RX_PKTFLAG_ACKED) &&
3684 ap->reason != RX_ACK_DELAY &&
3685 clock_Eq(&p->timeSent, &p->firstSent))
3686 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3688 rxi_ComputeRate(peer, call, p, np, ap->reason);
3692 /* The real smarts of the whole thing. */
3694 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
3697 struct rx_ackPacket *ap;
3699 struct rx_packet *tp;
3700 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3701 struct rx_connection *conn = call->conn;
3702 struct rx_peer *peer = conn->peer;
3705 /* because there are CM's that are bogus, sending weird values for this. */
3706 afs_uint32 skew = 0;
3712 int newAckCount = 0;
3713 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3714 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3716 if (rx_stats_active)
3717 rx_MutexIncrement(rx_stats.ackPacketsRead, rx_stats_mutex);
3718 ap = (struct rx_ackPacket *)rx_DataOf(np);
3719 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3721 return np; /* truncated ack packet */
3723 /* depends on ack packet struct */
3724 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3725 first = ntohl(ap->firstPacket);
3726 serial = ntohl(ap->serial);
3727 /* temporarily disabled -- needs to degrade over time
3728 * skew = ntohs(ap->maxSkew); */
3730 /* Ignore ack packets received out of order */
3731 if (first < call->tfirst) {
3735 if (np->header.flags & RX_SLOW_START_OK) {
3736 call->flags |= RX_CALL_SLOW_START_OK;
3739 if (ap->reason == RX_ACK_PING_RESPONSE)
3740 rxi_UpdatePeerReach(conn, call);
3744 if (rxdebug_active) {
3748 len = _snprintf(msg, sizeof(msg),
3749 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3750 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3751 ntohl(ap->serial), ntohl(ap->previousPacket),
3752 (unsigned int)np->header.seq, (unsigned int)skew,
3753 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3757 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3758 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3762 OutputDebugString(msg);
3764 #else /* AFS_NT40_ENV */
3767 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3768 ap->reason, ntohl(ap->previousPacket),
3769 (unsigned int)np->header.seq, (unsigned int)serial,
3770 (unsigned int)skew, ntohl(ap->firstPacket));
3773 for (offset = 0; offset < nAcks; offset++)
3774 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3779 #endif /* AFS_NT40_ENV */
3782 /* Update the outgoing packet skew value to the latest value of
3783 * the peer's incoming packet skew value. The ack packet, of
3784 * course, could arrive out of order, but that won't affect things
3786 MUTEX_ENTER(&peer->peer_lock);
3787 peer->outPacketSkew = skew;
3789 /* Check for packets that no longer need to be transmitted, and
3790 * discard them. This only applies to packets positively
3791 * acknowledged as having been sent to the peer's upper level.
3792 * All other packets must be retained. So only packets with
3793 * sequence numbers < ap->firstPacket are candidates. */
3794 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3795 if (tp->header.seq >= first)
3797 call->tfirst = tp->header.seq + 1;
3798 rxi_ComputePeerNetStats(call, tp, ap, np);
3799 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3802 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3803 /* XXX Hack. Because we have to release the global rx lock when sending
3804 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3805 * in rxi_Start sending packets out because packets may move to the
3806 * freePacketQueue as result of being here! So we drop these packets until
3807 * we're safely out of the traversing. Really ugly!
3808 * To make it even uglier, if we're using fine grain locking, we can
3809 * set the ack bits in the packets and have rxi_Start remove the packets
3810 * when it's done transmitting.
3812 if (call->flags & RX_CALL_TQ_BUSY) {
3813 #ifdef RX_ENABLE_LOCKS
3814 tp->flags |= RX_PKTFLAG_ACKED;
3815 call->flags |= RX_CALL_TQ_SOME_ACKED;
3816 #else /* RX_ENABLE_LOCKS */
3818 #endif /* RX_ENABLE_LOCKS */
3820 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3823 tp->flags &= ~RX_PKTFLAG_TQ;
3824 #ifdef RXDEBUG_PACKET
3826 #endif /* RXDEBUG_PACKET */
3827 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3832 /* Give rate detector a chance to respond to ping requests */
3833 if (ap->reason == RX_ACK_PING_RESPONSE) {
3834 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3838 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3840 /* Now go through explicit acks/nacks and record the results in
3841 * the waiting packets. These are packets that can't be released
3842 * yet, even with a positive acknowledge. This positive
3843 * acknowledge only means the packet has been received by the
3844 * peer, not that it will be retained long enough to be sent to
3845 * the peer's upper level. In addition, reset the transmit timers
3846 * of any missing packets (those packets that must be missing
3847 * because this packet was out of sequence) */
3849 call->nSoftAcked = 0;
3850 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3851 /* Update round trip time if the ack was stimulated on receipt
3853 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3854 #ifdef RX_ENABLE_LOCKS
3855 if (tp->header.seq >= first)
3856 #endif /* RX_ENABLE_LOCKS */
3857 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3858 rxi_ComputePeerNetStats(call, tp, ap, np);
3860 /* Set the acknowledge flag per packet based on the
3861 * information in the ack packet. An acknowlegded packet can
3862 * be downgraded when the server has discarded a packet it
3863 * soacked previously, or when an ack packet is received
3864 * out of sequence. */
3865 if (tp->header.seq < first) {
3866 /* Implicit ack information */
3867 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3870 tp->flags |= RX_PKTFLAG_ACKED;
3871 } else if (tp->header.seq < first + nAcks) {
3872 /* Explicit ack information: set it in the packet appropriately */
3873 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3874 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3876 tp->flags |= RX_PKTFLAG_ACKED;
3883 } else /* RX_ACK_TYPE_NACK */ {
3884 tp->flags &= ~RX_PKTFLAG_ACKED;
3888 tp->flags &= ~RX_PKTFLAG_ACKED;
3893 * Following the suggestion of Phil Kern, we back off the peer's
3894 * timeout value for future packets until a successful response
3895 * is received for an initial transmission.
3897 if (missing && !backedOff) {
3898 struct clock c = peer->timeout;
3899 struct clock max_to = {3, 0};
3901 clock_Add(&peer->timeout, &c);
3902 if (clock_Gt(&peer->timeout, &max_to))
3903 peer->timeout = max_to;
3907 /* If packet isn't yet acked, and it has been transmitted at least
3908 * once, reset retransmit time using latest timeout
3909 * ie, this should readjust the retransmit timer for all outstanding
3910 * packets... So we don't just retransmit when we should know better*/
3912 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3913 tp->retryTime = tp->timeSent;
3914 clock_Add(&tp->retryTime, &peer->timeout);
3915 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3916 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3920 /* If the window has been extended by this acknowledge packet,
3921 * then wakeup a sender waiting in alloc for window space, or try
3922 * sending packets now, if he's been sitting on packets due to
3923 * lack of window space */
3924 if (call->tnext < (call->tfirst + call->twind)) {
3925 #ifdef RX_ENABLE_LOCKS
3926 CV_SIGNAL(&call->cv_twind);
3928 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3929 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3930 osi_rxWakeup(&call->twind);
3933 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3934 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3938 /* if the ack packet has a receivelen field hanging off it,
3939 * update our state */
3940 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3943 /* If the ack packet has a "recommended" size that is less than
3944 * what I am using now, reduce my size to match */
3945 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
3946 (int)sizeof(afs_int32), &tSize);
3947 tSize = (afs_uint32) ntohl(tSize);
3948 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3950 /* Get the maximum packet size to send to this peer */
3951 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3953 tSize = (afs_uint32) ntohl(tSize);
3954 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3955 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3957 /* sanity check - peer might have restarted with different params.
3958 * If peer says "send less", dammit, send less... Peer should never
3959 * be unable to accept packets of the size that prior AFS versions would
3960 * send without asking. */
3961 if (peer->maxMTU != tSize) {
3962 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3964 peer->maxMTU = tSize;
3965 peer->MTU = MIN(tSize, peer->MTU);
3966 call->MTU = MIN(call->MTU, tSize);
3969 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3972 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
3973 (int)sizeof(afs_int32), &tSize);
3974 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3975 if (tSize < call->twind) { /* smaller than our send */
3976 call->twind = tSize; /* window, we must send less... */
3977 call->ssthresh = MIN(call->twind, call->ssthresh);
3978 call->conn->twind[call->channel] = call->twind;
3981 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3982 * network MTU confused with the loopback MTU. Calculate the
3983 * maximum MTU here for use in the slow start code below.
3985 maxMTU = peer->maxMTU;
3986 /* Did peer restart with older RX version? */
3987 if (peer->maxDgramPackets > 1) {
3988 peer->maxDgramPackets = 1;
3990 } else if (np->length >=
3991 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3994 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
3995 sizeof(afs_int32), &tSize);
3996 tSize = (afs_uint32) ntohl(tSize);
3998 * As of AFS 3.5 we set the send window to match the receive window.
4000 if (tSize < call->twind) {
4001 call->twind = tSize;
4002 call->conn->twind[call->channel] = call->twind;
4003 call->ssthresh = MIN(call->twind, call->ssthresh);
4004 } else if (tSize > call->twind) {
4005 call->twind = tSize;
4006 call->conn->twind[call->channel] = call->twind;
4010 * As of AFS 3.5, a jumbogram is more than one fixed size
4011 * packet transmitted in a single UDP datagram. If the remote
4012 * MTU is smaller than our local MTU then never send a datagram
4013 * larger than the natural MTU.
4016 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4017 (int)sizeof(afs_int32), &tSize);
4018 maxDgramPackets = (afs_uint32) ntohl(tSize);
4019 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4021 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4022 maxDgramPackets = MIN(maxDgramPackets, tSize);
4023 if (maxDgramPackets > 1) {
4024 peer->maxDgramPackets = maxDgramPackets;
4025 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4027 peer->maxDgramPackets = 1;
4028 call->MTU = peer->natMTU;
4030 } else if (peer->maxDgramPackets > 1) {
4031 /* Restarted with lower version of RX */
4032 peer->maxDgramPackets = 1;
4034 } else if (peer->maxDgramPackets > 1
4035 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4036 /* Restarted with lower version of RX */
4037 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4038 peer->natMTU = OLD_MAX_PACKET_SIZE;
4039 peer->MTU = OLD_MAX_PACKET_SIZE;
4040 peer->maxDgramPackets = 1;
4041 peer->nDgramPackets = 1;
4043 call->MTU = OLD_MAX_PACKET_SIZE;
4048 * Calculate how many datagrams were successfully received after
4049 * the first missing packet and adjust the negative ack counter
4054 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4055 if (call->nNacks < nNacked) {
4056 call->nNacks = nNacked;
4059 call->nAcks += newAckCount;
4063 if (call->flags & RX_CALL_FAST_RECOVER) {
4065 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4067 call->flags &= ~RX_CALL_FAST_RECOVER;
4068 call->cwind = call->nextCwind;
4069 call->nextCwind = 0;
4072 call->nCwindAcks = 0;
4073 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4074 /* Three negative acks in a row trigger congestion recovery */
4075 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4076 MUTEX_EXIT(&peer->peer_lock);
4077 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4078 /* someone else is waiting to start recovery */
4081 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4082 rxi_WaitforTQBusy(call);
4083 MUTEX_ENTER(&peer->peer_lock);
4084 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4085 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4086 call->flags |= RX_CALL_FAST_RECOVER;
4087 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4089 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4090 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4091 call->nextCwind = call->ssthresh;
4094 peer->MTU = call->MTU;
4095 peer->cwind = call->nextCwind;
4096 peer->nDgramPackets = call->nDgramPackets;
4098 call->congestSeq = peer->congestSeq;
4099 /* Reset the resend times on the packets that were nacked
4100 * so we will retransmit as soon as the window permits*/
4101 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4103 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4104 clock_Zero(&tp->retryTime);
4106 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4111 /* If cwind is smaller than ssthresh, then increase
4112 * the window one packet for each ack we receive (exponential
4114 * If cwind is greater than or equal to ssthresh then increase
4115 * the congestion window by one packet for each cwind acks we
4116 * receive (linear growth). */
4117 if (call->cwind < call->ssthresh) {
4119 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4120 call->nCwindAcks = 0;
4122 call->nCwindAcks += newAckCount;
4123 if (call->nCwindAcks >= call->cwind) {
4124 call->nCwindAcks = 0;
4125 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4129 * If we have received several acknowledgements in a row then
4130 * it is time to increase the size of our datagrams
4132 if ((int)call->nAcks > rx_nDgramThreshold) {
4133 if (peer->maxDgramPackets > 1) {
4134 if (call->nDgramPackets < peer->maxDgramPackets) {
4135 call->nDgramPackets++;
4137 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4138 } else if (call->MTU < peer->maxMTU) {
4139 call->MTU += peer->natMTU;
4140 call->MTU = MIN(call->MTU, peer->maxMTU);
4146 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4148 /* Servers need to hold the call until all response packets have
4149 * been acknowledged. Soft acks are good enough since clients
4150 * are not allowed to clear their receive queues. */
4151 if (call->state == RX_STATE_HOLD
4152 && call->tfirst + call->nSoftAcked >= call->tnext) {
4153 call->state = RX_STATE_DALLY;
4154 rxi_ClearTransmitQueue(call, 0);
4155 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4156 } else if (!queue_IsEmpty(&call->tq)) {
4157 rxi_Start(0, call, 0, istack);
4162 /* Received a response to a challenge packet */
4164 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4165 struct rx_packet *np, int istack)
4169 /* Ignore the packet if we're the client */
4170 if (conn->type == RX_CLIENT_CONNECTION)
4173 /* If already authenticated, ignore the packet (it's probably a retry) */
4174 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4177 /* Otherwise, have the security object evaluate the response packet */
4178 error = RXS_CheckResponse(conn->securityObject, conn, np);
4180 /* If the response is invalid, reset the connection, sending
4181 * an abort to the peer */
4185 rxi_ConnectionError(conn, error);
4186 MUTEX_ENTER(&conn->conn_data_lock);
4187 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4188 MUTEX_EXIT(&conn->conn_data_lock);
4191 /* If the response is valid, any calls waiting to attach
4192 * servers can now do so */
4195 for (i = 0; i < RX_MAXCALLS; i++) {
4196 struct rx_call *call = conn->call[i];
4198 MUTEX_ENTER(&call->lock);
4199 if (call->state == RX_STATE_PRECALL)
4200 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4201 /* tnop can be null if newcallp is null */
4202 MUTEX_EXIT(&call->lock);
4206 /* Update the peer reachability information, just in case
4207 * some calls went into attach-wait while we were waiting
4208 * for authentication..
4210 rxi_UpdatePeerReach(conn, NULL);
4215 /* A client has received an authentication challenge: the security
4216 * object is asked to cough up a respectable response packet to send
4217 * back to the server. The server is responsible for retrying the
4218 * challenge if it fails to get a response. */
4221 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4222 struct rx_packet *np, int istack)
4226 /* Ignore the challenge if we're the server */
4227 if (conn->type == RX_SERVER_CONNECTION)
4230 /* Ignore the challenge if the connection is otherwise idle; someone's
4231 * trying to use us as an oracle. */
4232 if (!rxi_HasActiveCalls(conn))
4235 /* Send the security object the challenge packet. It is expected to fill
4236 * in the response. */
4237 error = RXS_GetResponse(conn->securityObject, conn, np);
4239 /* If the security object is unable to return a valid response, reset the
4240 * connection and send an abort to the peer. Otherwise send the response
4241 * packet to the peer connection. */
4243 rxi_ConnectionError(conn, error);
4244 MUTEX_ENTER(&conn->conn_data_lock);
4245 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4246 MUTEX_EXIT(&conn->conn_data_lock);
4248 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4249 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4255 /* Find an available server process to service the current request in
4256 * the given call structure. If one isn't available, queue up this
4257 * call so it eventually gets one */
4259 rxi_AttachServerProc(struct rx_call *call,
4260 osi_socket socket, int *tnop,
4261 struct rx_call **newcallp)
4263 struct rx_serverQueueEntry *sq;
4264 struct rx_service *service = call->conn->service;
4267 /* May already be attached */
4268 if (call->state == RX_STATE_ACTIVE)
4271 MUTEX_ENTER(&rx_serverPool_lock);
4273 haveQuota = QuotaOK(service);
4274 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4275 /* If there are no processes available to service this call,
4276 * put the call on the incoming call queue (unless it's
4277 * already on the queue).
4279 #ifdef RX_ENABLE_LOCKS
4281 ReturnToServerPool(service);
4282 #endif /* RX_ENABLE_LOCKS */
4284 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4285 call->flags |= RX_CALL_WAIT_PROC;
4286 MUTEX_ENTER(&rx_waiting_mutex);
4289 MUTEX_EXIT(&rx_waiting_mutex);
4290 rxi_calltrace(RX_CALL_ARRIVAL, call);
4291 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4292 queue_Append(&rx_incomingCallQueue, call);
4295 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4297 /* If hot threads are enabled, and both newcallp and sq->socketp
4298 * are non-null, then this thread will process the call, and the
4299 * idle server thread will start listening on this threads socket.
4302 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4305 *sq->socketp = socket;
4306 clock_GetTime(&call->startTime);
4307 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4311 if (call->flags & RX_CALL_WAIT_PROC) {
4312 /* Conservative: I don't think this should happen */
4313 call->flags &= ~RX_CALL_WAIT_PROC;
4314 if (queue_IsOnQueue(call)) {
4317 MUTEX_ENTER(&rx_waiting_mutex);
4319 MUTEX_EXIT(&rx_waiting_mutex);
4322 call->state = RX_STATE_ACTIVE;
4323 call->mode = RX_MODE_RECEIVING;
4324 #ifdef RX_KERNEL_TRACE
4326 int glockOwner = ISAFS_GLOCK();
4329 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4330 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4336 if (call->flags & RX_CALL_CLEARED) {
4337 /* send an ack now to start the packet flow up again */
4338 call->flags &= ~RX_CALL_CLEARED;
4339 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4341 #ifdef RX_ENABLE_LOCKS
4344 service->nRequestsRunning++;
4345 if (service->nRequestsRunning <= service->minProcs)
4351 MUTEX_EXIT(&rx_serverPool_lock);
4354 /* Delay the sending of an acknowledge event for a short while, while
4355 * a new call is being prepared (in the case of a client) or a reply
4356 * is being prepared (in the case of a server). Rather than sending
4357 * an ack packet, an ACKALL packet is sent. */
4359 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4361 #ifdef RX_ENABLE_LOCKS
4363 MUTEX_ENTER(&call->lock);
4364 call->delayedAckEvent = NULL;
4365 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4367 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4368 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4370 MUTEX_EXIT(&call->lock);
4371 #else /* RX_ENABLE_LOCKS */
4373 call->delayedAckEvent = NULL;
4374 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4375 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4376 #endif /* RX_ENABLE_LOCKS */
4380 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4382 struct rx_call *call = arg1;
4383 #ifdef RX_ENABLE_LOCKS
4385 MUTEX_ENTER(&call->lock);
4386 if (event == call->delayedAckEvent)
4387 call->delayedAckEvent = NULL;
4388 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4390 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4392 MUTEX_EXIT(&call->lock);
4393 #else /* RX_ENABLE_LOCKS */
4395 call->delayedAckEvent = NULL;
4396 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4397 #endif /* RX_ENABLE_LOCKS */
4401 #ifdef RX_ENABLE_LOCKS
4402 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4403 * clearing them out.
4406 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4408 struct rx_packet *p, *tp;
4411 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4412 p->flags |= RX_PKTFLAG_ACKED;
4416 call->flags |= RX_CALL_TQ_CLEARME;
4417 call->flags |= RX_CALL_TQ_SOME_ACKED;
4420 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4421 call->tfirst = call->tnext;
4422 call->nSoftAcked = 0;
4424 if (call->flags & RX_CALL_FAST_RECOVER) {
4425 call->flags &= ~RX_CALL_FAST_RECOVER;
4426 call->cwind = call->nextCwind;
4427 call->nextCwind = 0;
4430 CV_SIGNAL(&call->cv_twind);
4432 #endif /* RX_ENABLE_LOCKS */
4434 /* Clear out the transmit queue for the current call (all packets have
4435 * been received by peer) */
4437 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4439 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4440 struct rx_packet *p, *tp;
4442 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4444 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4445 p->flags |= RX_PKTFLAG_ACKED;
4449 call->flags |= RX_CALL_TQ_CLEARME;
4450 call->flags |= RX_CALL_TQ_SOME_ACKED;
4453 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4454 #ifdef RXDEBUG_PACKET
4456 #endif /* RXDEBUG_PACKET */
4457 rxi_FreePackets(0, &call->tq);
4458 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
4459 #ifdef RX_ENABLE_LOCKS
4460 CV_BROADCAST(&call->cv_tq);
4461 #else /* RX_ENABLE_LOCKS */
4462 osi_rxWakeup(&call->tq);
4463 #endif /* RX_ENABLE_LOCKS */
4465 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4466 call->flags &= ~RX_CALL_TQ_CLEARME;
4468 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4470 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4471 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4472 call->nSoftAcked = 0;
4474 if (call->flags & RX_CALL_FAST_RECOVER) {
4475 call->flags &= ~RX_CALL_FAST_RECOVER;
4476 call->cwind = call->nextCwind;
4478 #ifdef RX_ENABLE_LOCKS
4479 CV_SIGNAL(&call->cv_twind);
4481 osi_rxWakeup(&call->twind);
4486 rxi_ClearReceiveQueue(struct rx_call *call)
4488 if (queue_IsNotEmpty(&call->rq)) {
4491 count = rxi_FreePackets(0, &call->rq);
4492 rx_packetReclaims += count;
4493 #ifdef RXDEBUG_PACKET
4495 if ( call->rqc != 0 )
4496 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0", call, call->rqc));
4498 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4500 if (call->state == RX_STATE_PRECALL) {
4501 call->flags |= RX_CALL_CLEARED;
4505 /* Send an abort packet for the specified call */
4507 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
4508 int istack, int force)
4511 struct clock when, now;
4516 /* Clients should never delay abort messages */
4517 if (rx_IsClientConn(call->conn))
4520 if (call->abortCode != call->error) {
4521 call->abortCode = call->error;
4522 call->abortCount = 0;
4525 if (force || rxi_callAbortThreshhold == 0
4526 || call->abortCount < rxi_callAbortThreshhold) {
4527 if (call->delayedAbortEvent) {
4528 rxevent_Cancel(call->delayedAbortEvent, call,
4529 RX_CALL_REFCOUNT_ABORT);
4531 error = htonl(call->error);
4534 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4535 (char *)&error, sizeof(error), istack);
4536 } else if (!call->delayedAbortEvent) {
4537 clock_GetTime(&now);
4539 clock_Addmsec(&when, rxi_callAbortDelay);
4540 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4541 call->delayedAbortEvent =
4542 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4547 /* Send an abort packet for the specified connection. Packet is an
4548 * optional pointer to a packet that can be used to send the abort.
4549 * Once the number of abort messages reaches the threshhold, an
4550 * event is scheduled to send the abort. Setting the force flag
4551 * overrides sending delayed abort messages.
4553 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4554 * to send the abort packet.
4557 rxi_SendConnectionAbort(struct rx_connection *conn,
4558 struct rx_packet *packet, int istack, int force)
4561 struct clock when, now;
4566 /* Clients should never delay abort messages */
4567 if (rx_IsClientConn(conn))
4570 if (force || rxi_connAbortThreshhold == 0
4571 || conn->abortCount < rxi_connAbortThreshhold) {
4572 if (conn->delayedAbortEvent) {
4573 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4575 error = htonl(conn->error);
4577 MUTEX_EXIT(&conn->conn_data_lock);
4579 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4580 RX_PACKET_TYPE_ABORT, (char *)&error,
4581 sizeof(error), istack);
4582 MUTEX_ENTER(&conn->conn_data_lock);
4583 } else if (!conn->delayedAbortEvent) {
4584 clock_GetTime(&now);
4586 clock_Addmsec(&when, rxi_connAbortDelay);
4587 conn->delayedAbortEvent =
4588 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4593 /* Associate an error all of the calls owned by a connection. Called
4594 * with error non-zero. This is only for really fatal things, like
4595 * bad authentication responses. The connection itself is set in
4596 * error at this point, so that future packets received will be
4599 rxi_ConnectionError(struct rx_connection *conn,
4605 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d", conn, error));
4607 MUTEX_ENTER(&conn->conn_data_lock);
4608 if (conn->challengeEvent)
4609 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4610 if (conn->natKeepAliveEvent)
4611 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
4612 if (conn->checkReachEvent) {
4613 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4614 conn->checkReachEvent = 0;
4615 conn->flags &= ~RX_CONN_ATTACHWAIT;
4618 MUTEX_EXIT(&conn->conn_data_lock);
4619 for (i = 0; i < RX_MAXCALLS; i++) {
4620 struct rx_call *call = conn->call[i];
4622 MUTEX_ENTER(&call->lock);
4623 rxi_CallError(call, error);
4624 MUTEX_EXIT(&call->lock);
4627 conn->error = error;
4628 if (rx_stats_active)
4629 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4634 rxi_CallError(struct rx_call *call, afs_int32 error)
4637 osirx_AssertMine(&call->lock, "rxi_CallError");
4639 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d", call, error, call->error));
4641 error = call->error;
4643 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4644 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4645 rxi_ResetCall(call, 0);
4648 rxi_ResetCall(call, 0);
4650 call->error = error;
4651 call->mode = RX_MODE_ERROR;
4654 /* Reset various fields in a call structure, and wakeup waiting
4655 * processes. Some fields aren't changed: state & mode are not
4656 * touched (these must be set by the caller), and bufptr, nLeft, and
4657 * nFree are not reset, since these fields are manipulated by
4658 * unprotected macros, and may only be reset by non-interrupting code.
4661 /* this code requires that call->conn be set properly as a pre-condition. */
4662 #endif /* ADAPT_WINDOW */
4665 rxi_ResetCall(struct rx_call *call, int newcall)
4668 struct rx_peer *peer;
4669 struct rx_packet *packet;
4671 osirx_AssertMine(&call->lock, "rxi_ResetCall");
4673 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
4675 /* Notify anyone who is waiting for asynchronous packet arrival */
4676 if (call->arrivalProc) {
4677 (*call->arrivalProc) (call, call->arrivalProcHandle,
4678 call->arrivalProcArg);
4679 call->arrivalProc = (void (*)())0;
4682 if (call->delayedAbortEvent) {
4683 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4684 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4686 rxi_SendCallAbort(call, packet, 0, 1);
4687 rxi_FreePacket(packet);
4692 * Update the peer with the congestion information in this call
4693 * so other calls on this connection can pick up where this call
4694 * left off. If the congestion sequence numbers don't match then
4695 * another call experienced a retransmission.
4697 peer = call->conn->peer;
4698 MUTEX_ENTER(&peer->peer_lock);
4700 if (call->congestSeq == peer->congestSeq) {
4701 peer->cwind = MAX(peer->cwind, call->cwind);
4702 peer->MTU = MAX(peer->MTU, call->MTU);
4703 peer->nDgramPackets =
4704 MAX(peer->nDgramPackets, call->nDgramPackets);
4707 call->abortCode = 0;
4708 call->abortCount = 0;
4710 if (peer->maxDgramPackets > 1) {
4711 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4713 call->MTU = peer->MTU;
4715 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4716 call->ssthresh = rx_maxSendWindow;
4717 call->nDgramPackets = peer->nDgramPackets;
4718 call->congestSeq = peer->congestSeq;
4719 MUTEX_EXIT(&peer->peer_lock);
4721 flags = call->flags;
4722 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4723 rxi_WaitforTQBusy(call);
4724 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4726 rxi_ClearTransmitQueue(call, 1);
4727 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4728 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4732 rxi_ClearReceiveQueue(call);
4733 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
4735 if (call->currentPacket) {
4736 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
4737 call->currentPacket->flags |= RX_PKTFLAG_IOVQ;
4738 queue_Prepend(&call->iovq, call->currentPacket);
4739 #ifdef RXDEBUG_PACKET
4741 #endif /* RXDEBUG_PACKET */
4742 call->currentPacket = (struct rx_packet *)0;
4744 call->curlen = call->nLeft = call->nFree = 0;
4746 #ifdef RXDEBUG_PACKET
4749 rxi_FreePackets(0, &call->iovq);
4752 call->twind = call->conn->twind[call->channel];
4753 call->rwind = call->conn->rwind[call->channel];
4754 call->nSoftAcked = 0;
4755 call->nextCwind = 0;
4758 call->nCwindAcks = 0;
4759 call->nSoftAcks = 0;
4760 call->nHardAcks = 0;
4762 call->tfirst = call->rnext = call->tnext = 1;
4764 call->lastAcked = 0;
4765 call->localStatus = call->remoteStatus = 0;
4767 if (flags & RX_CALL_READER_WAIT) {
4768 #ifdef RX_ENABLE_LOCKS
4769 CV_BROADCAST(&call->cv_rq);
4771 osi_rxWakeup(&call->rq);
4774 if (flags & RX_CALL_WAIT_PACKETS) {
4775 MUTEX_ENTER(&rx_freePktQ_lock);
4776 rxi_PacketsUnWait(); /* XXX */
4777 MUTEX_EXIT(&rx_freePktQ_lock);
4779 #ifdef RX_ENABLE_LOCKS
4780 CV_SIGNAL(&call->cv_twind);
4782 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4783 osi_rxWakeup(&call->twind);
4786 #ifdef RX_ENABLE_LOCKS
4787 /* The following ensures that we don't mess with any queue while some
4788 * other thread might also be doing so. The call_queue_lock field is
4789 * is only modified under the call lock. If the call is in the process
4790 * of being removed from a queue, the call is not locked until the
4791 * the queue lock is dropped and only then is the call_queue_lock field
4792 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4793 * Note that any other routine which removes a call from a queue has to
4794 * obtain the queue lock before examing the queue and removing the call.
4796 if (call->call_queue_lock) {
4797 MUTEX_ENTER(call->call_queue_lock);
4798 if (queue_IsOnQueue(call)) {
4800 if (flags & RX_CALL_WAIT_PROC) {
4802 MUTEX_ENTER(&rx_waiting_mutex);
4804 MUTEX_EXIT(&rx_waiting_mutex);
4807 MUTEX_EXIT(call->call_queue_lock);
4808 CLEAR_CALL_QUEUE_LOCK(call);
4810 #else /* RX_ENABLE_LOCKS */
4811 if (queue_IsOnQueue(call)) {
4813 if (flags & RX_CALL_WAIT_PROC)
4816 #endif /* RX_ENABLE_LOCKS */
4818 rxi_KeepAliveOff(call);
4819 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4822 /* Send an acknowledge for the indicated packet (seq,serial) of the
4823 * indicated call, for the indicated reason (reason). This
4824 * acknowledge will specifically acknowledge receiving the packet, and
4825 * will also specify which other packets for this call have been
4826 * received. This routine returns the packet that was used to the
4827 * caller. The caller is responsible for freeing it or re-using it.
4828 * This acknowledgement also returns the highest sequence number
4829 * actually read out by the higher level to the sender; the sender
4830 * promises to keep around packets that have not been read by the
4831 * higher level yet (unless, of course, the sender decides to abort
4832 * the call altogether). Any of p, seq, serial, pflags, or reason may
4833 * be set to zero without ill effect. That is, if they are zero, they
4834 * will not convey any information.
4835 * NOW there is a trailer field, after the ack where it will safely be
4836 * ignored by mundanes, which indicates the maximum size packet this
4837 * host can swallow. */
4839 struct rx_packet *optionalPacket; use to send ack (or null)
4840 int seq; Sequence number of the packet we are acking
4841 int serial; Serial number of the packet
4842 int pflags; Flags field from packet header
4843 int reason; Reason an acknowledge was prompted
4847 rxi_SendAck(struct rx_call *call,
4848 struct rx_packet *optionalPacket, int serial, int reason,
4851 struct rx_ackPacket *ap;
4852 struct rx_packet *rqp;
4853 struct rx_packet *nxp; /* For queue_Scan */
4854 struct rx_packet *p;
4857 #ifdef RX_ENABLE_TSFPQ
4858 struct rx_ts_info_t * rx_ts_info;
4862 * Open the receive window once a thread starts reading packets
4864 if (call->rnext > 1) {
4865 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
4868 call->nHardAcks = 0;
4869 call->nSoftAcks = 0;
4870 if (call->rnext > call->lastAcked)
4871 call->lastAcked = call->rnext;
4875 rx_computelen(p, p->length); /* reset length, you never know */
4876 } /* where that's been... */
4877 #ifdef RX_ENABLE_TSFPQ
4879 RX_TS_INFO_GET(rx_ts_info);
4880 if ((p = rx_ts_info->local_special_packet)) {
4881 rx_computelen(p, p->length);
4882 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4883 rx_ts_info->local_special_packet = p;
4884 } else { /* We won't send the ack, but don't panic. */
4885 return optionalPacket;
4889 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4890 /* We won't send the ack, but don't panic. */
4891 return optionalPacket;
4896 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4899 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4900 #ifndef RX_ENABLE_TSFPQ
4901 if (!optionalPacket)
4904 return optionalPacket;
4906 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4907 if (rx_Contiguous(p) < templ) {
4908 #ifndef RX_ENABLE_TSFPQ
4909 if (!optionalPacket)
4912 return optionalPacket;
4917 /* MTUXXX failing to send an ack is very serious. We should */
4918 /* try as hard as possible to send even a partial ack; it's */
4919 /* better than nothing. */
4920 ap = (struct rx_ackPacket *)rx_DataOf(p);
4921 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4922 ap->reason = reason;
4924 /* The skew computation used to be bogus, I think it's better now. */
4925 /* We should start paying attention to skew. XXX */
4926 ap->serial = htonl(serial);
4927 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4929 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4930 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4932 /* No fear of running out of ack packet here because there can only be at most
4933 * one window full of unacknowledged packets. The window size must be constrained
4934 * to be less than the maximum ack size, of course. Also, an ack should always
4935 * fit into a single packet -- it should not ever be fragmented. */
4936 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4937 if (!rqp || !call->rq.next
4938 || (rqp->header.seq > (call->rnext + call->rwind))) {
4939 #ifndef RX_ENABLE_TSFPQ
4940 if (!optionalPacket)
4943 rxi_CallError(call, RX_CALL_DEAD);
4944 return optionalPacket;
4947 while (rqp->header.seq > call->rnext + offset)
4948 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4949 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4951 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4952 #ifndef RX_ENABLE_TSFPQ
4953 if (!optionalPacket)
4956 rxi_CallError(call, RX_CALL_DEAD);
4957 return optionalPacket;
4962 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4964 /* these are new for AFS 3.3 */
4965 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4966 templ = htonl(templ);
4967 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4968 templ = htonl(call->conn->peer->ifMTU);
4969 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4970 sizeof(afs_int32), &templ);
4972 /* new for AFS 3.4 */
4973 templ = htonl(call->rwind);
4974 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4975 sizeof(afs_int32), &templ);
4977 /* new for AFS 3.5 */
4978 templ = htonl(call->conn->peer->ifDgramPackets);
4979 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4980 sizeof(afs_int32), &templ);
4982 p->header.serviceId = call->conn->serviceId;
4983 p->header.cid = (call->conn->cid | call->channel);
4984 p->header.callNumber = *call->callNumber;
4986 p->header.securityIndex = call->conn->securityIndex;
4987 p->header.epoch = call->conn->epoch;
4988 p->header.type = RX_PACKET_TYPE_ACK;
4989 p->header.flags = RX_SLOW_START_OK;
4990 if (reason == RX_ACK_PING) {
4991 p->header.flags |= RX_REQUEST_ACK;
4993 clock_GetTime(&call->pingRequestTime);
4996 if (call->conn->type == RX_CLIENT_CONNECTION)
4997 p->header.flags |= RX_CLIENT_INITIATED;
5001 if (rxdebug_active) {
5005 len = _snprintf(msg, sizeof(msg),
5006 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5007 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5008 ntohl(ap->serial), ntohl(ap->previousPacket),
5009 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5010 ap->nAcks, ntohs(ap->bufferSpace) );
5014 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5015 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5019 OutputDebugString(msg);
5021 #else /* AFS_NT40_ENV */
5023 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5024 ap->reason, ntohl(ap->previousPacket),
5025 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5027 for (offset = 0; offset < ap->nAcks; offset++)
5028 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5033 #endif /* AFS_NT40_ENV */
5036 int i, nbytes = p->length;
5038 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5039 if (nbytes <= p->wirevec[i].iov_len) {
5042 savelen = p->wirevec[i].iov_len;
5044 p->wirevec[i].iov_len = nbytes;
5046 rxi_Send(call, p, istack);
5047 p->wirevec[i].iov_len = savelen;
5051 nbytes -= p->wirevec[i].iov_len;
5054 if (rx_stats_active)
5055 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
5056 #ifndef RX_ENABLE_TSFPQ
5057 if (!optionalPacket)
5060 return optionalPacket; /* Return packet for re-use by caller */
5063 /* Send all of the packets in the list in single datagram */
5065 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
5066 int istack, int moreFlag, struct clock *now,
5067 struct clock *retryTime, int resending)
5072 struct rx_connection *conn = call->conn;
5073 struct rx_peer *peer = conn->peer;
5075 MUTEX_ENTER(&peer->peer_lock);
5078 peer->reSends += len;
5079 if (rx_stats_active)
5080 rx_MutexAdd(rx_stats.dataPacketsSent, len, rx_stats_mutex);
5081 MUTEX_EXIT(&peer->peer_lock);
5083 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
5087 /* Set the packet flags and schedule the resend events */
5088 /* Only request an ack for the last packet in the list */
5089 for (i = 0; i < len; i++) {
5090 list[i]->retryTime = *retryTime;
5091 if (list[i]->header.serial) {
5092 /* Exponentially backoff retry times */
5093 if (list[i]->backoff < MAXBACKOFF) {
5094 /* so it can't stay == 0 */
5095 list[i]->backoff = (list[i]->backoff << 1) + 1;
5098 clock_Addmsec(&(list[i]->retryTime),
5099 ((afs_uint32) list[i]->backoff) << 8);
5102 /* Wait a little extra for the ack on the last packet */
5103 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
5104 clock_Addmsec(&(list[i]->retryTime), 400);
5107 /* Record the time sent */
5108 list[i]->timeSent = *now;
5110 /* Ask for an ack on retransmitted packets, on every other packet
5111 * if the peer doesn't support slow start. Ask for an ack on every
5112 * packet until the congestion window reaches the ack rate. */
5113 if (list[i]->header.serial) {
5115 if (rx_stats_active)
5116 rx_MutexIncrement(rx_stats.dataPacketsReSent, rx_stats_mutex);
5118 /* improved RTO calculation- not Karn */
5119 list[i]->firstSent = *now;
5120 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5121 || (!(call->flags & RX_CALL_SLOW_START_OK)
5122 && (list[i]->header.seq & 1)))) {
5127 /* Tag this packet as not being the last in this group,
5128 * for the receiver's benefit */
5129 if (i < len - 1 || moreFlag) {
5130 list[i]->header.flags |= RX_MORE_PACKETS;
5133 /* Install the new retransmit time for the packet, and
5134 * record the time sent */
5135 list[i]->timeSent = *now;
5139 list[len - 1]->header.flags |= RX_REQUEST_ACK;
5142 /* Since we're about to send a data packet to the peer, it's
5143 * safe to nuke any scheduled end-of-packets ack */
5144 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5146 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5147 MUTEX_EXIT(&call->lock);
5149 rxi_SendPacketList(call, conn, list, len, istack);
5151 rxi_SendPacket(call, conn, list[0], istack);
5153 MUTEX_ENTER(&call->lock);
5154 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5156 /* Update last send time for this call (for keep-alive
5157 * processing), and for the connection (so that we can discover
5158 * idle connections) */
5159 call->lastSendData = conn->lastSendTime = call->lastSendTime = clock_Sec();
5162 /* When sending packets we need to follow these rules:
5163 * 1. Never send more than maxDgramPackets in a jumbogram.
5164 * 2. Never send a packet with more than two iovecs in a jumbogram.
5165 * 3. Never send a retransmitted packet in a jumbogram.
5166 * 4. Never send more than cwind/4 packets in a jumbogram
5167 * We always keep the last list we should have sent so we
5168 * can set the RX_MORE_PACKETS flags correctly.
5171 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5172 int istack, struct clock *now, struct clock *retryTime,
5175 int i, cnt, lastCnt = 0;
5176 struct rx_packet **listP, **lastP = 0;
5177 struct rx_peer *peer = call->conn->peer;
5178 int morePackets = 0;
5180 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
5181 /* Does the current packet force us to flush the current list? */
5183 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5184 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5186 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
5188 /* If the call enters an error state stop sending, or if
5189 * we entered congestion recovery mode, stop sending */
5190 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5198 /* Add the current packet to the list if it hasn't been acked.
5199 * Otherwise adjust the list pointer to skip the current packet. */
5200 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5202 /* Do we need to flush the list? */
5203 if (cnt >= (int)peer->maxDgramPackets
5204 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5205 || list[i]->header.serial
5206 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5208 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5209 retryTime, resending);
5210 /* If the call enters an error state stop sending, or if
5211 * we entered congestion recovery mode, stop sending */
5213 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5218 listP = &list[i + 1];
5223 osi_Panic("rxi_SendList error");
5225 listP = &list[i + 1];
5229 /* Send the whole list when the call is in receive mode, when
5230 * the call is in eof mode, when we are in fast recovery mode,
5231 * and when we have the last packet */
5232 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5233 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5234 || (call->flags & RX_CALL_FAST_RECOVER)) {
5235 /* Check for the case where the current list contains
5236 * an acked packet. Since we always send retransmissions
5237 * in a separate packet, we only need to check the first
5238 * packet in the list */
5239 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5243 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5244 retryTime, resending);
5245 /* If the call enters an error state stop sending, or if
5246 * we entered congestion recovery mode, stop sending */
5247 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5251 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5254 } else if (lastCnt > 0) {
5255 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5260 #ifdef RX_ENABLE_LOCKS
5261 /* Call rxi_Start, below, but with the call lock held. */
5263 rxi_StartUnlocked(struct rxevent *event,
5264 void *arg0, void *arg1, int istack)
5266 struct rx_call *call = arg0;
5268 MUTEX_ENTER(&call->lock);
5269 rxi_Start(event, call, arg1, istack);
5270 MUTEX_EXIT(&call->lock);
5272 #endif /* RX_ENABLE_LOCKS */
5274 /* This routine is called when new packets are readied for
5275 * transmission and when retransmission may be necessary, or when the
5276 * transmission window or burst count are favourable. This should be
5277 * better optimized for new packets, the usual case, now that we've
5278 * got rid of queues of send packets. XXXXXXXXXXX */
5280 rxi_Start(struct rxevent *event,
5281 void *arg0, void *arg1, int istack)
5283 struct rx_call *call = arg0;
5285 struct rx_packet *p;
5286 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5287 struct rx_peer *peer = call->conn->peer;
5288 struct clock now, usenow, retryTime;
5292 struct rx_packet **xmitList;
5295 /* If rxi_Start is being called as a result of a resend event,
5296 * then make sure that the event pointer is removed from the call
5297 * structure, since there is no longer a per-call retransmission
5299 if (event && event == call->resendEvent) {
5300 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5301 call->resendEvent = NULL;
5303 if (queue_IsEmpty(&call->tq)) {
5307 /* Timeouts trigger congestion recovery */
5308 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5309 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5310 /* someone else is waiting to start recovery */
5313 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5314 rxi_WaitforTQBusy(call);
5315 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5316 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5317 call->flags |= RX_CALL_FAST_RECOVER;
5318 if (peer->maxDgramPackets > 1) {
5319 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5321 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5323 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5324 call->nDgramPackets = 1;
5326 call->nextCwind = 1;
5329 MUTEX_ENTER(&peer->peer_lock);
5330 peer->MTU = call->MTU;
5331 peer->cwind = call->cwind;
5332 peer->nDgramPackets = 1;
5334 call->congestSeq = peer->congestSeq;
5335 MUTEX_EXIT(&peer->peer_lock);
5336 /* Clear retry times on packets. Otherwise, it's possible for
5337 * some packets in the queue to force resends at rates faster
5338 * than recovery rates.
5340 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5341 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5342 clock_Zero(&p->retryTime);
5347 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5348 if (rx_stats_active)
5349 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5354 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5355 /* Get clock to compute the re-transmit time for any packets
5356 * in this burst. Note, if we back off, it's reasonable to
5357 * back off all of the packets in the same manner, even if
5358 * some of them have been retransmitted more times than more
5360 * Do a dance to avoid blocking after setting now. */
5361 MUTEX_ENTER(&peer->peer_lock);
5362 retryTime = peer->timeout;
5363 MUTEX_EXIT(&peer->peer_lock);
5364 clock_GetTime(&now);
5365 clock_Add(&retryTime, &now);
5367 /* Send (or resend) any packets that need it, subject to
5368 * window restrictions and congestion burst control
5369 * restrictions. Ask for an ack on the last packet sent in
5370 * this burst. For now, we're relying upon the window being
5371 * considerably bigger than the largest number of packets that
5372 * are typically sent at once by one initial call to
5373 * rxi_Start. This is probably bogus (perhaps we should ask
5374 * for an ack when we're half way through the current
5375 * window?). Also, for non file transfer applications, this
5376 * may end up asking for an ack for every packet. Bogus. XXXX
5379 * But check whether we're here recursively, and let the other guy
5382 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5383 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5384 call->flags |= RX_CALL_TQ_BUSY;
5386 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5388 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5389 call->flags &= ~RX_CALL_NEED_START;
5390 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5392 maxXmitPackets = MIN(call->twind, call->cwind);
5393 xmitList = (struct rx_packet **)
5394 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
5395 /* XXXX else we must drop any mtx we hold */
5396 afs_osi_Alloc_NoSleep(maxXmitPackets * sizeof(struct rx_packet *));
5398 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5400 if (xmitList == NULL)
5401 osi_Panic("rxi_Start, failed to allocate xmit list");
5402 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5403 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5404 /* We shouldn't be sending packets if a thread is waiting
5405 * to initiate congestion recovery */
5406 dpf(("call %d waiting to initiate fast recovery\n",
5407 *(call->callNumber)));
5411 && (call->flags & RX_CALL_FAST_RECOVER)) {
5412 /* Only send one packet during fast recovery */
5413 dpf(("call %d restricted to one packet per send during fast recovery\n",
5414 *(call->callNumber)));
5417 if ((p->flags & RX_PKTFLAG_FREE)
5418 || (!queue_IsEnd(&call->tq, nxp)
5419 && (nxp->flags & RX_PKTFLAG_FREE))
5420 || (p == (struct rx_packet *)&rx_freePacketQueue)
5421 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5422 osi_Panic("rxi_Start: xmit queue clobbered");
5424 if (p->flags & RX_PKTFLAG_ACKED) {
5425 /* Since we may block, don't trust this */
5426 usenow.sec = usenow.usec = 0;
5427 if (rx_stats_active)
5428 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5429 continue; /* Ignore this packet if it has been acknowledged */
5432 /* Turn off all flags except these ones, which are the same
5433 * on each transmission */
5434 p->header.flags &= RX_PRESET_FLAGS;
5436 if (p->header.seq >=
5437 call->tfirst + MIN((int)call->twind,
5438 (int)(call->nSoftAcked +
5440 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5441 /* Note: if we're waiting for more window space, we can
5442 * still send retransmits; hence we don't return here, but
5443 * break out to schedule a retransmit event */
5444 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
5445 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
5450 /* Transmit the packet if it needs to be sent. */
5451 if (!clock_Lt(&now, &p->retryTime)) {
5452 if (nXmitPackets == maxXmitPackets) {
5453 rxi_SendXmitList(call, xmitList, nXmitPackets,
5454 istack, &now, &retryTime,
5456 osi_Free(xmitList, maxXmitPackets *
5457 sizeof(struct rx_packet *));
5460 dpf(("call %d xmit packet %"AFS_PTR_FMT" now %u.%06u retryTime %u.%06u nextRetry %u.%06u\n",
5461 *(call->callNumber), p,
5463 p->retryTime.sec, p->retryTime.usec,
5464 retryTime.sec, retryTime.usec));
5465 xmitList[nXmitPackets++] = p;
5469 /* xmitList now hold pointers to all of the packets that are
5470 * ready to send. Now we loop to send the packets */
5471 if (nXmitPackets > 0) {
5472 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5473 &now, &retryTime, resending);
5476 maxXmitPackets * sizeof(struct rx_packet *));
5478 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5480 * TQ references no longer protected by this flag; they must remain
5481 * protected by the global lock.
5483 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5484 call->flags &= ~RX_CALL_TQ_BUSY;
5485 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5486 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
5487 call, call->tqWaiters, call->flags));
5488 #ifdef RX_ENABLE_LOCKS
5489 osirx_AssertMine(&call->lock, "rxi_Start start");
5490 CV_BROADCAST(&call->cv_tq);
5491 #else /* RX_ENABLE_LOCKS */
5492 osi_rxWakeup(&call->tq);
5493 #endif /* RX_ENABLE_LOCKS */
5498 /* We went into the error state while sending packets. Now is
5499 * the time to reset the call. This will also inform the using
5500 * process that the call is in an error state.
5502 if (rx_stats_active)
5503 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5504 call->flags &= ~RX_CALL_TQ_BUSY;
5505 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5506 dpf(("call error %d while xmit %p has %d waiters and flags %d\n",
5507 call->error, call, call->tqWaiters, call->flags));
5508 #ifdef RX_ENABLE_LOCKS
5509 osirx_AssertMine(&call->lock, "rxi_Start middle");
5510 CV_BROADCAST(&call->cv_tq);
5511 #else /* RX_ENABLE_LOCKS */
5512 osi_rxWakeup(&call->tq);
5513 #endif /* RX_ENABLE_LOCKS */
5515 rxi_CallError(call, call->error);
5518 #ifdef RX_ENABLE_LOCKS
5519 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5521 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5522 /* Some packets have received acks. If they all have, we can clear
5523 * the transmit queue.
5526 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5527 if (p->header.seq < call->tfirst
5528 && (p->flags & RX_PKTFLAG_ACKED)) {
5530 p->flags &= ~RX_PKTFLAG_TQ;
5531 #ifdef RXDEBUG_PACKET
5539 call->flags |= RX_CALL_TQ_CLEARME;
5541 #endif /* RX_ENABLE_LOCKS */
5542 /* Don't bother doing retransmits if the TQ is cleared. */
5543 if (call->flags & RX_CALL_TQ_CLEARME) {
5544 rxi_ClearTransmitQueue(call, 1);
5546 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5549 /* Always post a resend event, if there is anything in the
5550 * queue, and resend is possible. There should be at least
5551 * one unacknowledged packet in the queue ... otherwise none
5552 * of these packets should be on the queue in the first place.
5554 if (call->resendEvent) {
5555 /* Cancel the existing event and post a new one */
5556 rxevent_Cancel(call->resendEvent, call,
5557 RX_CALL_REFCOUNT_RESEND);
5560 /* The retry time is the retry time on the first unacknowledged
5561 * packet inside the current window */
5563 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5564 /* Don't set timers for packets outside the window */
5565 if (p->header.seq >= call->tfirst + call->twind) {
5569 if (!(p->flags & RX_PKTFLAG_ACKED)
5570 && !clock_IsZero(&p->retryTime)) {
5572 retryTime = p->retryTime;
5577 /* Post a new event to re-run rxi_Start when retries may be needed */
5578 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5579 #ifdef RX_ENABLE_LOCKS
5580 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5582 rxevent_PostNow2(&retryTime, &usenow,
5584 (void *)call, 0, istack);
5585 #else /* RX_ENABLE_LOCKS */
5587 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5588 (void *)call, 0, istack);
5589 #endif /* RX_ENABLE_LOCKS */
5592 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5593 } while (call->flags & RX_CALL_NEED_START);
5595 * TQ references no longer protected by this flag; they must remain
5596 * protected by the global lock.
5598 call->flags &= ~RX_CALL_TQ_BUSY;
5599 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5600 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
5601 call, call->tqWaiters, call->flags));
5602 #ifdef RX_ENABLE_LOCKS
5603 osirx_AssertMine(&call->lock, "rxi_Start end");
5604 CV_BROADCAST(&call->cv_tq);
5605 #else /* RX_ENABLE_LOCKS */
5606 osi_rxWakeup(&call->tq);
5607 #endif /* RX_ENABLE_LOCKS */
5610 call->flags |= RX_CALL_NEED_START;
5612 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5614 if (call->resendEvent) {
5615 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5620 /* Also adjusts the keep alive parameters for the call, to reflect
5621 * that we have just sent a packet (so keep alives aren't sent
5624 rxi_Send(struct rx_call *call, struct rx_packet *p,
5627 struct rx_connection *conn = call->conn;
5629 /* Stamp each packet with the user supplied status */
5630 p->header.userStatus = call->localStatus;
5632 /* Allow the security object controlling this call's security to
5633 * make any last-minute changes to the packet */
5634 RXS_SendPacket(conn->securityObject, call, p);
5636 /* Since we're about to send SOME sort of packet to the peer, it's
5637 * safe to nuke any scheduled end-of-packets ack */
5638 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5640 /* Actually send the packet, filling in more connection-specific fields */
5641 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5642 MUTEX_EXIT(&call->lock);
5643 rxi_SendPacket(call, conn, p, istack);
5644 MUTEX_ENTER(&call->lock);
5645 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5647 /* Update last send time for this call (for keep-alive
5648 * processing), and for the connection (so that we can discover
5649 * idle connections) */
5650 conn->lastSendTime = call->lastSendTime = clock_Sec();
5651 /* Don't count keepalives here, so idleness can be tracked. */
5652 if ((p->header.type != RX_PACKET_TYPE_ACK) || (((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING))
5653 call->lastSendData = call->lastSendTime;
5657 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5658 * that things are fine. Also called periodically to guarantee that nothing
5659 * falls through the cracks (e.g. (error + dally) connections have keepalive
5660 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5662 * haveCTLock Set if calling from rxi_ReapConnections
5664 #ifdef RX_ENABLE_LOCKS
5666 rxi_CheckCall(struct rx_call *call, int haveCTLock)
5667 #else /* RX_ENABLE_LOCKS */
5669 rxi_CheckCall(struct rx_call *call)
5670 #endif /* RX_ENABLE_LOCKS */
5672 struct rx_connection *conn = call->conn;
5674 afs_uint32 deadTime;
5676 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5677 if (call->flags & RX_CALL_TQ_BUSY) {
5678 /* Call is active and will be reset by rxi_Start if it's
5679 * in an error state.
5684 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5686 (((afs_uint32) conn->secondsUntilDead << 10) +
5687 ((afs_uint32) conn->peer->rtt >> 3) +
5688 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5690 /* These are computed to the second (+- 1 second). But that's
5691 * good enough for these values, which should be a significant
5692 * number of seconds. */
5693 if (now > (call->lastReceiveTime + deadTime)) {
5694 if (call->state == RX_STATE_ACTIVE) {
5696 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5698 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5699 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5700 ip_stack_t *ipst = ns->netstack_ip;
5702 ire = ire_cache_lookup(call->conn->peer->host
5703 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5705 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5707 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5714 if (ire && ire->ire_max_frag > 0)
5715 rxi_SetPeerMtu(call->conn->peer->host, 0, ire->ire_max_frag);
5716 #if defined(GLOBAL_NETSTACKID)
5720 #endif /* ADAPT_PMTU */
5721 rxi_CallError(call, RX_CALL_DEAD);
5724 #ifdef RX_ENABLE_LOCKS
5725 /* Cancel pending events */
5726 rxevent_Cancel(call->delayedAckEvent, call,
5727 RX_CALL_REFCOUNT_DELAY);
5728 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5729 rxevent_Cancel(call->keepAliveEvent, call,
5730 RX_CALL_REFCOUNT_ALIVE);
5731 if (call->refCount == 0) {
5732 rxi_FreeCall(call, haveCTLock);
5736 #else /* RX_ENABLE_LOCKS */
5739 #endif /* RX_ENABLE_LOCKS */
5741 /* Non-active calls are destroyed if they are not responding
5742 * to pings; active calls are simply flagged in error, so the
5743 * attached process can die reasonably gracefully. */
5745 /* see if we have a non-activity timeout */
5746 if (call->startWait && conn->idleDeadTime
5747 && ((call->startWait + conn->idleDeadTime) < now) &&
5748 (call->flags & RX_CALL_READER_WAIT)) {
5749 if (call->state == RX_STATE_ACTIVE) {
5750 rxi_CallError(call, RX_CALL_TIMEOUT);
5754 if (call->lastSendData && conn->idleDeadTime && (conn->idleDeadErr != 0)
5755 && ((call->lastSendData + conn->idleDeadTime) < now)) {
5756 if (call->state == RX_STATE_ACTIVE) {
5757 rxi_CallError(call, conn->idleDeadErr);
5761 /* see if we have a hard timeout */
5762 if (conn->hardDeadTime
5763 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5764 if (call->state == RX_STATE_ACTIVE)
5765 rxi_CallError(call, RX_CALL_TIMEOUT);
5772 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
5774 struct rx_connection *conn = arg1;
5775 struct rx_header theader;
5777 struct sockaddr_in taddr;
5780 struct iovec tmpiov[2];
5783 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
5786 tp = &tbuffer[sizeof(struct rx_header)];
5787 taddr.sin_family = AF_INET;
5788 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
5789 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
5790 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
5791 taddr.sin_len = sizeof(struct sockaddr_in);
5793 memset(&theader, 0, sizeof(theader));
5794 theader.epoch = htonl(999);
5796 theader.callNumber = 0;
5799 theader.type = RX_PACKET_TYPE_VERSION;
5800 theader.flags = RX_LAST_PACKET;
5801 theader.serviceId = 0;
5803 memcpy(tbuffer, &theader, sizeof(theader));
5804 memcpy(tp, &a, sizeof(a));
5805 tmpiov[0].iov_base = tbuffer;
5806 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
5808 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
5810 MUTEX_ENTER(&conn->conn_data_lock);
5811 /* Only reschedule ourselves if the connection would not be destroyed */
5812 if (conn->refCount <= 1) {
5813 conn->natKeepAliveEvent = NULL;
5814 MUTEX_EXIT(&conn->conn_data_lock);
5815 rx_DestroyConnection(conn); /* drop the reference for this */
5817 conn->natKeepAliveEvent = NULL;
5818 conn->refCount--; /* drop the reference for this */
5819 rxi_ScheduleNatKeepAliveEvent(conn);
5820 MUTEX_EXIT(&conn->conn_data_lock);
5825 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
5827 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
5828 struct clock when, now;
5829 clock_GetTime(&now);
5831 when.sec += conn->secondsUntilNatPing;
5832 conn->refCount++; /* hold a reference for this */
5833 conn->natKeepAliveEvent =
5834 rxevent_PostNow(&when, &now, rxi_NatKeepAliveEvent, conn, 0);
5839 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
5841 MUTEX_ENTER(&conn->conn_data_lock);
5842 conn->secondsUntilNatPing = seconds;
5844 rxi_ScheduleNatKeepAliveEvent(conn);
5845 MUTEX_EXIT(&conn->conn_data_lock);
5849 rxi_NatKeepAliveOn(struct rx_connection *conn)
5851 MUTEX_ENTER(&conn->conn_data_lock);
5852 rxi_ScheduleNatKeepAliveEvent(conn);
5853 MUTEX_EXIT(&conn->conn_data_lock);
5856 /* When a call is in progress, this routine is called occasionally to
5857 * make sure that some traffic has arrived (or been sent to) the peer.
5858 * If nothing has arrived in a reasonable amount of time, the call is
5859 * declared dead; if nothing has been sent for a while, we send a
5860 * keep-alive packet (if we're actually trying to keep the call alive)
5863 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
5865 struct rx_call *call = arg1;
5866 struct rx_connection *conn;
5869 MUTEX_ENTER(&call->lock);
5870 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5871 if (event == call->keepAliveEvent)
5872 call->keepAliveEvent = NULL;
5875 #ifdef RX_ENABLE_LOCKS
5876 if (rxi_CheckCall(call, 0)) {
5877 MUTEX_EXIT(&call->lock);
5880 #else /* RX_ENABLE_LOCKS */
5881 if (rxi_CheckCall(call))
5883 #endif /* RX_ENABLE_LOCKS */
5885 /* Don't try to keep alive dallying calls */
5886 if (call->state == RX_STATE_DALLY) {
5887 MUTEX_EXIT(&call->lock);
5892 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5893 /* Don't try to send keepalives if there is unacknowledged data */
5894 /* the rexmit code should be good enough, this little hack
5895 * doesn't quite work XXX */
5896 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5898 rxi_ScheduleKeepAliveEvent(call);
5899 MUTEX_EXIT(&call->lock);
5904 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
5906 if (!call->keepAliveEvent) {
5907 struct clock when, now;
5908 clock_GetTime(&now);
5910 when.sec += call->conn->secondsUntilPing;
5911 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5912 call->keepAliveEvent =
5913 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
5917 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5919 rxi_KeepAliveOn(struct rx_call *call)
5921 /* Pretend last packet received was received now--i.e. if another
5922 * packet isn't received within the keep alive time, then the call
5923 * will die; Initialize last send time to the current time--even
5924 * if a packet hasn't been sent yet. This will guarantee that a
5925 * keep-alive is sent within the ping time */
5926 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5927 rxi_ScheduleKeepAliveEvent(call);
5930 /* This routine is called to send connection abort messages
5931 * that have been delayed to throttle looping clients. */
5933 rxi_SendDelayedConnAbort(struct rxevent *event,
5934 void *arg1, void *unused)
5936 struct rx_connection *conn = arg1;
5939 struct rx_packet *packet;
5941 MUTEX_ENTER(&conn->conn_data_lock);
5942 conn->delayedAbortEvent = NULL;
5943 error = htonl(conn->error);
5945 MUTEX_EXIT(&conn->conn_data_lock);
5946 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5949 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5950 RX_PACKET_TYPE_ABORT, (char *)&error,
5952 rxi_FreePacket(packet);
5956 /* This routine is called to send call abort messages
5957 * that have been delayed to throttle looping clients. */
5959 rxi_SendDelayedCallAbort(struct rxevent *event,
5960 void *arg1, void *dummy)
5962 struct rx_call *call = arg1;
5965 struct rx_packet *packet;
5967 MUTEX_ENTER(&call->lock);
5968 call->delayedAbortEvent = NULL;
5969 error = htonl(call->error);
5971 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5974 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5975 (char *)&error, sizeof(error), 0);
5976 rxi_FreePacket(packet);
5978 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5979 MUTEX_EXIT(&call->lock);
5982 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5983 * seconds) to ask the client to authenticate itself. The routine
5984 * issues a challenge to the client, which is obtained from the
5985 * security object associated with the connection */
5987 rxi_ChallengeEvent(struct rxevent *event,
5988 void *arg0, void *arg1, int tries)
5990 struct rx_connection *conn = arg0;
5992 conn->challengeEvent = NULL;
5993 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5994 struct rx_packet *packet;
5995 struct clock when, now;
5998 /* We've failed to authenticate for too long.
5999 * Reset any calls waiting for authentication;
6000 * they are all in RX_STATE_PRECALL.
6004 MUTEX_ENTER(&conn->conn_call_lock);
6005 for (i = 0; i < RX_MAXCALLS; i++) {
6006 struct rx_call *call = conn->call[i];
6008 MUTEX_ENTER(&call->lock);
6009 if (call->state == RX_STATE_PRECALL) {
6010 rxi_CallError(call, RX_CALL_DEAD);
6011 rxi_SendCallAbort(call, NULL, 0, 0);
6013 MUTEX_EXIT(&call->lock);
6016 MUTEX_EXIT(&conn->conn_call_lock);
6020 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6022 /* If there's no packet available, do this later. */
6023 RXS_GetChallenge(conn->securityObject, conn, packet);
6024 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6025 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6026 rxi_FreePacket(packet);
6028 clock_GetTime(&now);
6030 when.sec += RX_CHALLENGE_TIMEOUT;
6031 conn->challengeEvent =
6032 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
6037 /* Call this routine to start requesting the client to authenticate
6038 * itself. This will continue until authentication is established,
6039 * the call times out, or an invalid response is returned. The
6040 * security object associated with the connection is asked to create
6041 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6042 * defined earlier. */
6044 rxi_ChallengeOn(struct rx_connection *conn)
6046 if (!conn->challengeEvent) {
6047 RXS_CreateChallenge(conn->securityObject, conn);
6048 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6053 /* Compute round trip time of the packet provided, in *rttp.
6056 /* rxi_ComputeRoundTripTime is called with peer locked. */
6057 /* sentp and/or peer may be null */
6059 rxi_ComputeRoundTripTime(struct rx_packet *p,
6060 struct clock *sentp,
6061 struct rx_peer *peer)
6063 struct clock thisRtt, *rttp = &thisRtt;
6067 clock_GetTime(rttp);
6069 if (clock_Lt(rttp, sentp)) {
6071 return; /* somebody set the clock back, don't count this time. */
6073 clock_Sub(rttp, sentp);
6074 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6075 p->header.callNumber, p, rttp->sec, rttp->usec));
6077 if (rttp->sec == 0 && rttp->usec == 0) {
6079 * The actual round trip time is shorter than the
6080 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6081 * Since we can't tell which at the moment we will assume 1ms.
6086 if (rx_stats_active) {
6087 MUTEX_ENTER(&rx_stats_mutex);
6088 if (clock_Lt(rttp, &rx_stats.minRtt))
6089 rx_stats.minRtt = *rttp;
6090 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
6091 if (rttp->sec > 60) {
6092 MUTEX_EXIT(&rx_stats_mutex);
6093 return; /* somebody set the clock ahead */
6095 rx_stats.maxRtt = *rttp;
6097 clock_Add(&rx_stats.totalRtt, rttp);
6098 rx_stats.nRttSamples++;
6099 MUTEX_EXIT(&rx_stats_mutex);
6102 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6104 /* Apply VanJacobson round-trip estimations */
6109 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
6110 * srtt is stored as fixed point with 3 bits after the binary
6111 * point (i.e., scaled by 8). The following magic is
6112 * equivalent to the smoothing algorithm in rfc793 with an
6113 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6114 * srtt'*8 = rtt + srtt*7
6115 * srtt'*8 = srtt*8 + rtt - srtt
6116 * srtt' = srtt + rtt/8 - srtt/8
6117 * srtt' = srtt + (rtt - srtt)/8
6120 delta = _8THMSEC(rttp) - peer->rtt;
6121 peer->rtt += (delta >> 3);
6124 * We accumulate a smoothed rtt variance (actually, a smoothed
6125 * mean difference), then set the retransmit timer to smoothed
6126 * rtt + 4 times the smoothed variance (was 2x in van's original
6127 * paper, but 4x works better for me, and apparently for him as
6129 * rttvar is stored as
6130 * fixed point with 2 bits after the binary point (scaled by
6131 * 4). The following is equivalent to rfc793 smoothing with
6132 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6133 * rttvar'*4 = rttvar*3 + |delta|
6134 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6135 * rttvar' = rttvar + |delta|/4 - rttvar/4
6136 * rttvar' = rttvar + (|delta| - rttvar)/4
6137 * This replaces rfc793's wired-in beta.
6138 * dev*4 = dev*4 + (|actual - expected| - dev)
6144 delta -= (peer->rtt_dev << 1);
6145 peer->rtt_dev += (delta >> 3);
6147 /* I don't have a stored RTT so I start with this value. Since I'm
6148 * probably just starting a call, and will be pushing more data down
6149 * this, I expect congestion to increase rapidly. So I fudge a
6150 * little, and I set deviance to half the rtt. In practice,
6151 * deviance tends to approach something a little less than
6152 * half the smoothed rtt. */
6153 peer->rtt = _8THMSEC(rttp) + 8;
6154 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
6156 /* the timeout is RTT + 4*MDEV but no less than rx_minPeerTimeout msec.
6157 * This is because one end or the other of these connections is usually
6158 * in a user process, and can be switched and/or swapped out. So on fast,
6159 * reliable networks, the timeout would otherwise be too short. */
6160 rtt_timeout = MAX(((peer->rtt >> 3) + peer->rtt_dev), rx_minPeerTimeout);
6161 clock_Zero(&(peer->timeout));
6162 clock_Addmsec(&(peer->timeout), rtt_timeout);
6164 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6165 p->header.callNumber, p, MSEC(rttp), peer->rtt >> 3, peer->rtt_dev >> 2, (peer->timeout.sec), (peer->timeout.usec)));
6169 /* Find all server connections that have not been active for a long time, and
6172 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6174 struct clock now, when;
6175 clock_GetTime(&now);
6177 /* Find server connection structures that haven't been used for
6178 * greater than rx_idleConnectionTime */
6180 struct rx_connection **conn_ptr, **conn_end;
6181 int i, havecalls = 0;
6182 MUTEX_ENTER(&rx_connHashTable_lock);
6183 for (conn_ptr = &rx_connHashTable[0], conn_end =
6184 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6186 struct rx_connection *conn, *next;
6187 struct rx_call *call;
6191 for (conn = *conn_ptr; conn; conn = next) {
6192 /* XXX -- Shouldn't the connection be locked? */
6195 for (i = 0; i < RX_MAXCALLS; i++) {
6196 call = conn->call[i];
6200 code = MUTEX_TRYENTER(&call->lock);
6203 #ifdef RX_ENABLE_LOCKS
6204 result = rxi_CheckCall(call, 1);
6205 #else /* RX_ENABLE_LOCKS */
6206 result = rxi_CheckCall(call);
6207 #endif /* RX_ENABLE_LOCKS */
6208 MUTEX_EXIT(&call->lock);
6210 /* If CheckCall freed the call, it might
6211 * have destroyed the connection as well,
6212 * which screws up the linked lists.
6218 if (conn->type == RX_SERVER_CONNECTION) {
6219 /* This only actually destroys the connection if
6220 * there are no outstanding calls */
6221 MUTEX_ENTER(&conn->conn_data_lock);
6222 if (!havecalls && !conn->refCount
6223 && ((conn->lastSendTime + rx_idleConnectionTime) <
6225 conn->refCount++; /* it will be decr in rx_DestroyConn */
6226 MUTEX_EXIT(&conn->conn_data_lock);
6227 #ifdef RX_ENABLE_LOCKS
6228 rxi_DestroyConnectionNoLock(conn);
6229 #else /* RX_ENABLE_LOCKS */
6230 rxi_DestroyConnection(conn);
6231 #endif /* RX_ENABLE_LOCKS */
6233 #ifdef RX_ENABLE_LOCKS
6235 MUTEX_EXIT(&conn->conn_data_lock);
6237 #endif /* RX_ENABLE_LOCKS */
6241 #ifdef RX_ENABLE_LOCKS
6242 while (rx_connCleanup_list) {
6243 struct rx_connection *conn;
6244 conn = rx_connCleanup_list;
6245 rx_connCleanup_list = rx_connCleanup_list->next;
6246 MUTEX_EXIT(&rx_connHashTable_lock);
6247 rxi_CleanupConnection(conn);
6248 MUTEX_ENTER(&rx_connHashTable_lock);
6250 MUTEX_EXIT(&rx_connHashTable_lock);
6251 #endif /* RX_ENABLE_LOCKS */
6254 /* Find any peer structures that haven't been used (haven't had an
6255 * associated connection) for greater than rx_idlePeerTime */
6257 struct rx_peer **peer_ptr, **peer_end;
6259 MUTEX_ENTER(&rx_rpc_stats);
6260 MUTEX_ENTER(&rx_peerHashTable_lock);
6261 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6262 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6264 struct rx_peer *peer, *next, *prev;
6265 for (prev = peer = *peer_ptr; peer; peer = next) {
6267 code = MUTEX_TRYENTER(&peer->peer_lock);
6268 if ((code) && (peer->refCount == 0)
6269 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6270 rx_interface_stat_p rpc_stat, nrpc_stat;
6272 MUTEX_EXIT(&peer->peer_lock);
6273 MUTEX_DESTROY(&peer->peer_lock);
6275 (&peer->rpcStats, rpc_stat, nrpc_stat,
6276 rx_interface_stat)) {
6277 unsigned int num_funcs;
6280 queue_Remove(&rpc_stat->queue_header);
6281 queue_Remove(&rpc_stat->all_peers);
6282 num_funcs = rpc_stat->stats[0].func_total;
6284 sizeof(rx_interface_stat_t) +
6285 rpc_stat->stats[0].func_total *
6286 sizeof(rx_function_entry_v1_t);
6288 rxi_Free(rpc_stat, space);
6289 rxi_rpc_peer_stat_cnt -= num_funcs;
6292 if (rx_stats_active)
6293 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6294 if (peer == *peer_ptr) {
6301 MUTEX_EXIT(&peer->peer_lock);
6307 MUTEX_EXIT(&rx_peerHashTable_lock);
6308 MUTEX_EXIT(&rx_rpc_stats);
6311 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6312 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6313 * GC, just below. Really, we shouldn't have to keep moving packets from
6314 * one place to another, but instead ought to always know if we can
6315 * afford to hold onto a packet in its particular use. */
6316 MUTEX_ENTER(&rx_freePktQ_lock);
6317 if (rx_waitingForPackets) {
6318 rx_waitingForPackets = 0;
6319 #ifdef RX_ENABLE_LOCKS
6320 CV_BROADCAST(&rx_waitingForPackets_cv);
6322 osi_rxWakeup(&rx_waitingForPackets);
6325 MUTEX_EXIT(&rx_freePktQ_lock);
6328 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6329 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6333 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6334 * rx.h is sort of strange this is better. This is called with a security
6335 * object before it is discarded. Each connection using a security object has
6336 * its own refcount to the object so it won't actually be freed until the last
6337 * connection is destroyed.
6339 * This is the only rxs module call. A hold could also be written but no one
6343 rxs_Release(struct rx_securityClass *aobj)
6345 return RXS_Close(aobj);
6349 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6350 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6351 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6352 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6354 /* Adjust our estimate of the transmission rate to this peer, given
6355 * that the packet p was just acked. We can adjust peer->timeout and
6356 * call->twind. Pragmatically, this is called
6357 * only with packets of maximal length.
6358 * Called with peer and call locked.
6362 rxi_ComputeRate(struct rx_peer *peer, struct rx_call *call,
6363 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6365 afs_int32 xferSize, xferMs;
6369 /* Count down packets */
6370 if (peer->rateFlag > 0)
6372 /* Do nothing until we're enabled */
6373 if (peer->rateFlag != 0)
6378 /* Count only when the ack seems legitimate */
6379 switch (ackReason) {
6380 case RX_ACK_REQUESTED:
6382 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6386 case RX_ACK_PING_RESPONSE:
6387 if (p) /* want the response to ping-request, not data send */
6389 clock_GetTime(&newTO);
6390 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6391 clock_Sub(&newTO, &call->pingRequestTime);
6392 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6396 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6403 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %d.%06d, rtt %u, ps %u)",
6404 ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"),
6405 xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
6407 /* Track only packets that are big enough. */
6408 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6412 /* absorb RTT data (in milliseconds) for these big packets */
6413 if (peer->smRtt == 0) {
6414 peer->smRtt = xferMs;
6416 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6421 if (peer->countDown) {
6425 peer->countDown = 10; /* recalculate only every so often */
6427 /* In practice, we can measure only the RTT for full packets,
6428 * because of the way Rx acks the data that it receives. (If it's
6429 * smaller than a full packet, it often gets implicitly acked
6430 * either by the call response (from a server) or by the next call
6431 * (from a client), and either case confuses transmission times
6432 * with processing times.) Therefore, replace the above
6433 * more-sophisticated processing with a simpler version, where the
6434 * smoothed RTT is kept for full-size packets, and the time to
6435 * transmit a windowful of full-size packets is simply RTT *
6436 * windowSize. Again, we take two steps:
6437 - ensure the timeout is large enough for a single packet's RTT;
6438 - ensure that the window is small enough to fit in the desired timeout.*/
6440 /* First, the timeout check. */
6441 minTime = peer->smRtt;
6442 /* Get a reasonable estimate for a timeout period */
6444 newTO.sec = minTime / 1000;
6445 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6447 /* Increase the timeout period so that we can always do at least
6448 * one packet exchange */
6449 if (clock_Gt(&newTO, &peer->timeout)) {
6451 dpf(("CONG peer %lx/%u: timeout %d.%06d ==> %ld.%06d (rtt %u, ps %u)",
6452 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec,
6453 newTO.sec, newTO.usec, peer->smRtt, peer->packetSize));
6455 peer->timeout = newTO;
6458 /* Now, get an estimate for the transmit window size. */
6459 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6460 /* Now, convert to the number of full packets that could fit in a
6461 * reasonable fraction of that interval */
6462 minTime /= (peer->smRtt << 1);
6463 xferSize = minTime; /* (make a copy) */
6465 /* Now clamp the size to reasonable bounds. */
6468 else if (minTime > rx_Window)
6469 minTime = rx_Window;
6470 /* if (minTime != peer->maxWindow) {
6471 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6472 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6473 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6475 peer->maxWindow = minTime;
6476 elide... call->twind = minTime;
6480 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6481 * Discern this by calculating the timeout necessary for rx_Window
6483 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6484 /* calculate estimate for transmission interval in milliseconds */
6485 minTime = rx_Window * peer->smRtt;
6486 if (minTime < 1000) {
6487 dpf(("CONG peer %lx/%u: cut TO %d.%06d by 0.5 (rtt %u, ps %u)",
6488 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6489 peer->timeout.usec, peer->smRtt, peer->packetSize));
6491 newTO.sec = 0; /* cut back on timeout by half a second */
6492 newTO.usec = 500000;
6493 clock_Sub(&peer->timeout, &newTO);
6498 } /* end of rxi_ComputeRate */
6499 #endif /* ADAPT_WINDOW */
6507 #define TRACE_OPTION_RX_DEBUG 16
6515 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6516 0, KEY_QUERY_VALUE, &parmKey);
6517 if (code != ERROR_SUCCESS)
6520 dummyLen = sizeof(TraceOption);
6521 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6522 (BYTE *) &TraceOption, &dummyLen);
6523 if (code == ERROR_SUCCESS) {
6524 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
6526 RegCloseKey (parmKey);
6527 #endif /* AFS_NT40_ENV */
6532 rx_DebugOnOff(int on)
6536 rxdebug_active = on;
6542 rx_StatsOnOff(int on)
6545 rx_stats_active = on;
6550 /* Don't call this debugging routine directly; use dpf */
6552 rxi_DebugPrint(char *format, ...)
6561 va_start(ap, format);
6563 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6566 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
6568 if (msg[len-1] != '\n') {
6572 OutputDebugString(msg);
6579 va_start(ap, format);
6581 clock_GetTime(&now);
6582 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
6583 (unsigned int)now.usec);
6584 vfprintf(rx_Log, format, ap);
6593 * This function is used to process the rx_stats structure that is local
6594 * to a process as well as an rx_stats structure received from a remote
6595 * process (via rxdebug). Therefore, it needs to do minimal version
6599 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
6600 afs_int32 freePackets, char version)
6605 if (size != sizeof(struct rx_statistics)) {
6607 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
6608 size, sizeof(struct rx_statistics));
6611 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6614 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6615 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
6616 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6617 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6618 s->specialPktAllocFailures);
6620 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
6621 s->receivePktAllocFailures, s->sendPktAllocFailures,
6622 s->specialPktAllocFailures);
6626 " greedy %u, " "bogusReads %u (last from host %x), "
6627 "noPackets %u, " "noBuffers %u, " "selects %u, "
6628 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
6629 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6630 s->selects, s->sendSelects);
6632 fprintf(file, " packets read: ");
6633 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6634 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
6636 fprintf(file, "\n");
6639 " other read counters: data %u, " "ack %u, " "dup %u "
6640 "spurious %u " "dally %u\n", s->dataPacketsRead,
6641 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6642 s->ignorePacketDally);
6644 fprintf(file, " packets sent: ");
6645 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6646 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
6648 fprintf(file, "\n");
6651 " other send counters: ack %u, " "data %u (not resends), "
6652 "resends %u, " "pushed %u, " "acked&ignored %u\n",
6653 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6654 s->dataPacketsPushed, s->ignoreAckedPacket);
6657 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
6658 s->netSendFailures, (int)s->fatalErrors);
6660 if (s->nRttSamples) {
6661 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6662 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6664 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6665 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6669 " %d server connections, " "%d client connections, "
6670 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6671 s->nServerConns, s->nClientConns, s->nPeerStructs,
6672 s->nCallStructs, s->nFreeCallStructs);
6674 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6675 fprintf(file, " %d clock updates\n", clock_nUpdates);
6678 fprintf(file, "ERROR: compiled without RXDEBUG\n");
6682 /* for backward compatibility */
6684 rx_PrintStats(FILE * file)
6686 MUTEX_ENTER(&rx_stats_mutex);
6687 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6689 MUTEX_EXIT(&rx_stats_mutex);
6693 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6695 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
6696 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6697 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6700 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6701 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6702 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6705 " Packet size %d, " "max in packet skew %d, "
6706 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6707 (int)peer->outPacketSkew);
6711 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
6713 * This mutex protects the following static variables:
6717 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
6718 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
6720 #define LOCK_RX_DEBUG
6721 #define UNLOCK_RX_DEBUG
6722 #endif /* AFS_PTHREAD_ENV */
6726 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6727 u_char type, void *inputData, size_t inputLength,
6728 void *outputData, size_t outputLength)
6730 static afs_int32 counter = 100;
6731 time_t waitTime, waitCount, startTime;
6732 struct rx_header theader;
6735 struct timeval tv_now, tv_wake, tv_delta;
6736 struct sockaddr_in taddr, faddr;
6745 startTime = time(0);
6751 tp = &tbuffer[sizeof(struct rx_header)];
6752 taddr.sin_family = AF_INET;
6753 taddr.sin_port = remotePort;
6754 taddr.sin_addr.s_addr = remoteAddr;
6755 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6756 taddr.sin_len = sizeof(struct sockaddr_in);
6759 memset(&theader, 0, sizeof(theader));
6760 theader.epoch = htonl(999);
6762 theader.callNumber = htonl(counter);
6765 theader.type = type;
6766 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6767 theader.serviceId = 0;
6769 memcpy(tbuffer, &theader, sizeof(theader));
6770 memcpy(tp, inputData, inputLength);
6772 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6773 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6775 /* see if there's a packet available */
6776 gettimeofday(&tv_wake,0);
6777 tv_wake.tv_sec += waitTime;
6780 FD_SET(socket, &imask);
6781 tv_delta.tv_sec = tv_wake.tv_sec;
6782 tv_delta.tv_usec = tv_wake.tv_usec;
6783 gettimeofday(&tv_now, 0);
6785 if (tv_delta.tv_usec < tv_now.tv_usec) {
6787 tv_delta.tv_usec += 1000000;
6790 tv_delta.tv_usec -= tv_now.tv_usec;
6792 if (tv_delta.tv_sec < tv_now.tv_sec) {
6796 tv_delta.tv_sec -= tv_now.tv_sec;
6799 code = select(0, &imask, 0, 0, &tv_delta);
6800 #else /* AFS_NT40_ENV */
6801 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6802 #endif /* AFS_NT40_ENV */
6803 if (code == 1 && FD_ISSET(socket, &imask)) {
6804 /* now receive a packet */
6805 faddrLen = sizeof(struct sockaddr_in);
6807 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6808 (struct sockaddr *)&faddr, &faddrLen);
6811 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6812 if (counter == ntohl(theader.callNumber))
6820 /* see if we've timed out */
6828 code -= sizeof(struct rx_header);
6829 if (code > outputLength)
6830 code = outputLength;
6831 memcpy(outputData, tp, code);
6834 #endif /* RXDEBUG */
6837 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6838 afs_uint16 remotePort, struct rx_debugStats * stat,
6839 afs_uint32 * supportedValues)
6845 struct rx_debugIn in;
6847 *supportedValues = 0;
6848 in.type = htonl(RX_DEBUGI_GETSTATS);
6851 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6852 &in, sizeof(in), stat, sizeof(*stat));
6855 * If the call was successful, fixup the version and indicate
6856 * what contents of the stat structure are valid.
6857 * Also do net to host conversion of fields here.
6861 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6862 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6864 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6865 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6867 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6868 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6870 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6871 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6873 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6874 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6876 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6877 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6879 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6880 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6882 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6883 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6885 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
6886 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
6888 stat->nFreePackets = ntohl(stat->nFreePackets);
6889 stat->packetReclaims = ntohl(stat->packetReclaims);
6890 stat->callsExecuted = ntohl(stat->callsExecuted);
6891 stat->nWaiting = ntohl(stat->nWaiting);
6892 stat->idleThreads = ntohl(stat->idleThreads);
6893 stat->nWaited = ntohl(stat->nWaited);
6894 stat->nPackets = ntohl(stat->nPackets);
6901 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6902 afs_uint16 remotePort, struct rx_statistics * stat,
6903 afs_uint32 * supportedValues)
6909 struct rx_debugIn in;
6910 afs_int32 *lp = (afs_int32 *) stat;
6914 * supportedValues is currently unused, but added to allow future
6915 * versioning of this function.
6918 *supportedValues = 0;
6919 in.type = htonl(RX_DEBUGI_RXSTATS);
6921 memset(stat, 0, sizeof(*stat));
6923 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6924 &in, sizeof(in), stat, sizeof(*stat));
6929 * Do net to host conversion here
6932 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6941 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6942 afs_uint16 remotePort, size_t version_length,
6947 return MakeDebugCall(socket, remoteAddr, remotePort,
6948 RX_PACKET_TYPE_VERSION, a, 1, version,
6956 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6957 afs_uint16 remotePort, afs_int32 * nextConnection,
6958 int allConnections, afs_uint32 debugSupportedValues,
6959 struct rx_debugConn * conn,
6960 afs_uint32 * supportedValues)
6966 struct rx_debugIn in;
6970 * supportedValues is currently unused, but added to allow future
6971 * versioning of this function.
6974 *supportedValues = 0;
6975 if (allConnections) {
6976 in.type = htonl(RX_DEBUGI_GETALLCONN);
6978 in.type = htonl(RX_DEBUGI_GETCONN);
6980 in.index = htonl(*nextConnection);
6981 memset(conn, 0, sizeof(*conn));
6983 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6984 &in, sizeof(in), conn, sizeof(*conn));
6987 *nextConnection += 1;
6990 * Convert old connection format to new structure.
6993 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6994 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6995 #define MOVEvL(a) (conn->a = vL->a)
6997 /* any old or unrecognized version... */
6998 for (i = 0; i < RX_MAXCALLS; i++) {
6999 MOVEvL(callState[i]);
7000 MOVEvL(callMode[i]);
7001 MOVEvL(callFlags[i]);
7002 MOVEvL(callOther[i]);
7004 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7005 MOVEvL(secStats.type);
7006 MOVEvL(secStats.level);
7007 MOVEvL(secStats.flags);
7008 MOVEvL(secStats.expires);
7009 MOVEvL(secStats.packetsReceived);
7010 MOVEvL(secStats.packetsSent);
7011 MOVEvL(secStats.bytesReceived);
7012 MOVEvL(secStats.bytesSent);
7017 * Do net to host conversion here
7019 * I don't convert host or port since we are most likely
7020 * going to want these in NBO.
7022 conn->cid = ntohl(conn->cid);
7023 conn->serial = ntohl(conn->serial);
7024 for (i = 0; i < RX_MAXCALLS; i++) {
7025 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7027 conn->error = ntohl(conn->error);
7028 conn->secStats.flags = ntohl(conn->secStats.flags);
7029 conn->secStats.expires = ntohl(conn->secStats.expires);
7030 conn->secStats.packetsReceived =
7031 ntohl(conn->secStats.packetsReceived);
7032 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7033 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7034 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7035 conn->epoch = ntohl(conn->epoch);
7036 conn->natMTU = ntohl(conn->natMTU);
7043 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7044 afs_uint16 remotePort, afs_int32 * nextPeer,
7045 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7046 afs_uint32 * supportedValues)
7052 struct rx_debugIn in;
7055 * supportedValues is currently unused, but added to allow future
7056 * versioning of this function.
7059 *supportedValues = 0;
7060 in.type = htonl(RX_DEBUGI_GETPEER);
7061 in.index = htonl(*nextPeer);
7062 memset(peer, 0, sizeof(*peer));
7064 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7065 &in, sizeof(in), peer, sizeof(*peer));
7071 * Do net to host conversion here
7073 * I don't convert host or port since we are most likely
7074 * going to want these in NBO.
7076 peer->ifMTU = ntohs(peer->ifMTU);
7077 peer->idleWhen = ntohl(peer->idleWhen);
7078 peer->refCount = ntohs(peer->refCount);
7079 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7080 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7081 peer->rtt = ntohl(peer->rtt);
7082 peer->rtt_dev = ntohl(peer->rtt_dev);
7083 peer->timeout.sec = ntohl(peer->timeout.sec);
7084 peer->timeout.usec = ntohl(peer->timeout.usec);
7085 peer->nSent = ntohl(peer->nSent);
7086 peer->reSends = ntohl(peer->reSends);
7087 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7088 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7089 peer->rateFlag = ntohl(peer->rateFlag);
7090 peer->natMTU = ntohs(peer->natMTU);
7091 peer->maxMTU = ntohs(peer->maxMTU);
7092 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7093 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7094 peer->MTU = ntohs(peer->MTU);
7095 peer->cwind = ntohs(peer->cwind);
7096 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7097 peer->congestSeq = ntohs(peer->congestSeq);
7098 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7099 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7100 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7101 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7108 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7109 struct rx_debugPeer * peerStats)
7112 afs_int32 error = 1; /* default to "did not succeed" */
7113 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7115 MUTEX_ENTER(&rx_peerHashTable_lock);
7116 for(tp = rx_peerHashTable[hashValue];
7117 tp != NULL; tp = tp->next) {
7118 if (tp->host == peerHost)
7125 peerStats->host = tp->host;
7126 peerStats->port = tp->port;
7127 peerStats->ifMTU = tp->ifMTU;
7128 peerStats->idleWhen = tp->idleWhen;
7129 peerStats->refCount = tp->refCount;
7130 peerStats->burstSize = tp->burstSize;
7131 peerStats->burst = tp->burst;
7132 peerStats->burstWait.sec = tp->burstWait.sec;
7133 peerStats->burstWait.usec = tp->burstWait.usec;
7134 peerStats->rtt = tp->rtt;
7135 peerStats->rtt_dev = tp->rtt_dev;
7136 peerStats->timeout.sec = tp->timeout.sec;
7137 peerStats->timeout.usec = tp->timeout.usec;
7138 peerStats->nSent = tp->nSent;
7139 peerStats->reSends = tp->reSends;
7140 peerStats->inPacketSkew = tp->inPacketSkew;
7141 peerStats->outPacketSkew = tp->outPacketSkew;
7142 peerStats->rateFlag = tp->rateFlag;
7143 peerStats->natMTU = tp->natMTU;
7144 peerStats->maxMTU = tp->maxMTU;
7145 peerStats->maxDgramPackets = tp->maxDgramPackets;
7146 peerStats->ifDgramPackets = tp->ifDgramPackets;
7147 peerStats->MTU = tp->MTU;
7148 peerStats->cwind = tp->cwind;
7149 peerStats->nDgramPackets = tp->nDgramPackets;
7150 peerStats->congestSeq = tp->congestSeq;
7151 peerStats->bytesSent.high = tp->bytesSent.high;
7152 peerStats->bytesSent.low = tp->bytesSent.low;
7153 peerStats->bytesReceived.high = tp->bytesReceived.high;
7154 peerStats->bytesReceived.low = tp->bytesReceived.low;
7156 MUTEX_EXIT(&rx_peerHashTable_lock);
7164 struct rx_serverQueueEntry *np;
7167 struct rx_call *call;
7168 struct rx_serverQueueEntry *sq;
7172 if (rxinit_status == 1) {
7174 return; /* Already shutdown. */
7178 #ifndef AFS_PTHREAD_ENV
7179 FD_ZERO(&rx_selectMask);
7180 #endif /* AFS_PTHREAD_ENV */
7181 rxi_dataQuota = RX_MAX_QUOTA;
7182 #ifndef AFS_PTHREAD_ENV
7184 #endif /* AFS_PTHREAD_ENV */
7187 #ifndef AFS_PTHREAD_ENV
7188 #ifndef AFS_USE_GETTIMEOFDAY
7190 #endif /* AFS_USE_GETTIMEOFDAY */
7191 #endif /* AFS_PTHREAD_ENV */
7193 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7194 call = queue_First(&rx_freeCallQueue, rx_call);
7196 rxi_Free(call, sizeof(struct rx_call));
7199 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7200 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7206 struct rx_peer **peer_ptr, **peer_end;
7207 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7208 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7210 struct rx_peer *peer, *next;
7211 for (peer = *peer_ptr; peer; peer = next) {
7212 rx_interface_stat_p rpc_stat, nrpc_stat;
7215 (&peer->rpcStats, rpc_stat, nrpc_stat,
7216 rx_interface_stat)) {
7217 unsigned int num_funcs;
7220 queue_Remove(&rpc_stat->queue_header);
7221 queue_Remove(&rpc_stat->all_peers);
7222 num_funcs = rpc_stat->stats[0].func_total;
7224 sizeof(rx_interface_stat_t) +
7225 rpc_stat->stats[0].func_total *
7226 sizeof(rx_function_entry_v1_t);
7228 rxi_Free(rpc_stat, space);
7229 MUTEX_ENTER(&rx_rpc_stats);
7230 rxi_rpc_peer_stat_cnt -= num_funcs;
7231 MUTEX_EXIT(&rx_rpc_stats);
7235 if (rx_stats_active)
7236 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
7240 for (i = 0; i < RX_MAX_SERVICES; i++) {
7242 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7244 for (i = 0; i < rx_hashTableSize; i++) {
7245 struct rx_connection *tc, *ntc;
7246 MUTEX_ENTER(&rx_connHashTable_lock);
7247 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7249 for (j = 0; j < RX_MAXCALLS; j++) {
7251 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7254 rxi_Free(tc, sizeof(*tc));
7256 MUTEX_EXIT(&rx_connHashTable_lock);
7259 MUTEX_ENTER(&freeSQEList_lock);
7261 while ((np = rx_FreeSQEList)) {
7262 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7263 MUTEX_DESTROY(&np->lock);
7264 rxi_Free(np, sizeof(*np));
7267 MUTEX_EXIT(&freeSQEList_lock);
7268 MUTEX_DESTROY(&freeSQEList_lock);
7269 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7270 MUTEX_DESTROY(&rx_connHashTable_lock);
7271 MUTEX_DESTROY(&rx_peerHashTable_lock);
7272 MUTEX_DESTROY(&rx_serverPool_lock);
7274 osi_Free(rx_connHashTable,
7275 rx_hashTableSize * sizeof(struct rx_connection *));
7276 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7278 UNPIN(rx_connHashTable,
7279 rx_hashTableSize * sizeof(struct rx_connection *));
7280 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7282 rxi_FreeAllPackets();
7284 MUTEX_ENTER(&rx_quota_mutex);
7285 rxi_dataQuota = RX_MAX_QUOTA;
7286 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7287 MUTEX_EXIT(&rx_quota_mutex);
7292 #ifdef RX_ENABLE_LOCKS
7294 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7296 if (!MUTEX_ISMINE(lockaddr))
7297 osi_Panic("Lock not held: %s", msg);
7299 #endif /* RX_ENABLE_LOCKS */
7304 * Routines to implement connection specific data.
7308 rx_KeyCreate(rx_destructor_t rtn)
7311 MUTEX_ENTER(&rxi_keyCreate_lock);
7312 key = rxi_keyCreate_counter++;
7313 rxi_keyCreate_destructor = (rx_destructor_t *)
7314 realloc((void *)rxi_keyCreate_destructor,
7315 (key + 1) * sizeof(rx_destructor_t));
7316 rxi_keyCreate_destructor[key] = rtn;
7317 MUTEX_EXIT(&rxi_keyCreate_lock);
7322 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7325 MUTEX_ENTER(&conn->conn_data_lock);
7326 if (!conn->specific) {
7327 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7328 for (i = 0; i < key; i++)
7329 conn->specific[i] = NULL;
7330 conn->nSpecific = key + 1;
7331 conn->specific[key] = ptr;
7332 } else if (key >= conn->nSpecific) {
7333 conn->specific = (void **)
7334 realloc(conn->specific, (key + 1) * sizeof(void *));
7335 for (i = conn->nSpecific; i < key; i++)
7336 conn->specific[i] = NULL;
7337 conn->nSpecific = key + 1;
7338 conn->specific[key] = ptr;
7340 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7341 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7342 conn->specific[key] = ptr;
7344 MUTEX_EXIT(&conn->conn_data_lock);
7348 rx_GetSpecific(struct rx_connection *conn, int key)
7351 MUTEX_ENTER(&conn->conn_data_lock);
7352 if (key >= conn->nSpecific)
7355 ptr = conn->specific[key];
7356 MUTEX_EXIT(&conn->conn_data_lock);
7360 #endif /* !KERNEL */
7363 * processStats is a queue used to store the statistics for the local
7364 * process. Its contents are similar to the contents of the rpcStats
7365 * queue on a rx_peer structure, but the actual data stored within
7366 * this queue contains totals across the lifetime of the process (assuming
7367 * the stats have not been reset) - unlike the per peer structures
7368 * which can come and go based upon the peer lifetime.
7371 static struct rx_queue processStats = { &processStats, &processStats };
7374 * peerStats is a queue used to store the statistics for all peer structs.
7375 * Its contents are the union of all the peer rpcStats queues.
7378 static struct rx_queue peerStats = { &peerStats, &peerStats };
7381 * rxi_monitor_processStats is used to turn process wide stat collection
7385 static int rxi_monitor_processStats = 0;
7388 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7391 static int rxi_monitor_peerStats = 0;
7394 * rxi_AddRpcStat - given all of the information for a particular rpc
7395 * call, create (if needed) and update the stat totals for the rpc.
7399 * IN stats - the queue of stats that will be updated with the new value
7401 * IN rxInterface - a unique number that identifies the rpc interface
7403 * IN currentFunc - the index of the function being invoked
7405 * IN totalFunc - the total number of functions in this interface
7407 * IN queueTime - the amount of time this function waited for a thread
7409 * IN execTime - the amount of time this function invocation took to execute
7411 * IN bytesSent - the number bytes sent by this invocation
7413 * IN bytesRcvd - the number bytes received by this invocation
7415 * IN isServer - if true, this invocation was made to a server
7417 * IN remoteHost - the ip address of the remote host
7419 * IN remotePort - the port of the remote host
7421 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7423 * INOUT counter - if a new stats structure is allocated, the counter will
7424 * be updated with the new number of allocated stat structures
7432 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
7433 afs_uint32 currentFunc, afs_uint32 totalFunc,
7434 struct clock *queueTime, struct clock *execTime,
7435 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
7436 afs_uint32 remoteHost, afs_uint32 remotePort,
7437 int addToPeerList, unsigned int *counter)
7440 rx_interface_stat_p rpc_stat, nrpc_stat;
7443 * See if there's already a structure for this interface
7446 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7447 if ((rpc_stat->stats[0].interfaceId == rxInterface)
7448 && (rpc_stat->stats[0].remote_is_server == isServer))
7453 * Didn't find a match so allocate a new structure and add it to the
7457 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
7458 || (rpc_stat->stats[0].interfaceId != rxInterface)
7459 || (rpc_stat->stats[0].remote_is_server != isServer)) {
7464 sizeof(rx_interface_stat_t) +
7465 totalFunc * sizeof(rx_function_entry_v1_t);
7467 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
7468 if (rpc_stat == NULL) {
7472 *counter += totalFunc;
7473 for (i = 0; i < totalFunc; i++) {
7474 rpc_stat->stats[i].remote_peer = remoteHost;
7475 rpc_stat->stats[i].remote_port = remotePort;
7476 rpc_stat->stats[i].remote_is_server = isServer;
7477 rpc_stat->stats[i].interfaceId = rxInterface;
7478 rpc_stat->stats[i].func_total = totalFunc;
7479 rpc_stat->stats[i].func_index = i;
7480 hzero(rpc_stat->stats[i].invocations);
7481 hzero(rpc_stat->stats[i].bytes_sent);
7482 hzero(rpc_stat->stats[i].bytes_rcvd);
7483 rpc_stat->stats[i].queue_time_sum.sec = 0;
7484 rpc_stat->stats[i].queue_time_sum.usec = 0;
7485 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7486 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7487 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7488 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7489 rpc_stat->stats[i].queue_time_max.sec = 0;
7490 rpc_stat->stats[i].queue_time_max.usec = 0;
7491 rpc_stat->stats[i].execution_time_sum.sec = 0;
7492 rpc_stat->stats[i].execution_time_sum.usec = 0;
7493 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7494 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7495 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7496 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7497 rpc_stat->stats[i].execution_time_max.sec = 0;
7498 rpc_stat->stats[i].execution_time_max.usec = 0;
7500 queue_Prepend(stats, rpc_stat);
7501 if (addToPeerList) {
7502 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7507 * Increment the stats for this function
7510 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7511 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7512 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7513 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7514 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7515 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7516 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7518 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7519 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7521 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7522 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7524 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7525 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7527 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7528 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7536 * rx_IncrementTimeAndCount - increment the times and count for a particular
7541 * IN peer - the peer who invoked the rpc
7543 * IN rxInterface - a unique number that identifies the rpc interface
7545 * IN currentFunc - the index of the function being invoked
7547 * IN totalFunc - the total number of functions in this interface
7549 * IN queueTime - the amount of time this function waited for a thread
7551 * IN execTime - the amount of time this function invocation took to execute
7553 * IN bytesSent - the number bytes sent by this invocation
7555 * IN bytesRcvd - the number bytes received by this invocation
7557 * IN isServer - if true, this invocation was made to a server
7565 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7566 afs_uint32 currentFunc, afs_uint32 totalFunc,
7567 struct clock *queueTime, struct clock *execTime,
7568 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7572 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7575 MUTEX_ENTER(&rx_rpc_stats);
7576 MUTEX_ENTER(&peer->peer_lock);
7578 if (rxi_monitor_peerStats) {
7579 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7580 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7581 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7584 if (rxi_monitor_processStats) {
7585 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7586 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7587 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7590 MUTEX_EXIT(&peer->peer_lock);
7591 MUTEX_EXIT(&rx_rpc_stats);
7596 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7600 * IN callerVersion - the rpc stat version of the caller.
7602 * IN count - the number of entries to marshall.
7604 * IN stats - pointer to stats to be marshalled.
7606 * OUT ptr - Where to store the marshalled data.
7613 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7614 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7620 * We only support the first version
7622 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7623 *(ptr++) = stats->remote_peer;
7624 *(ptr++) = stats->remote_port;
7625 *(ptr++) = stats->remote_is_server;
7626 *(ptr++) = stats->interfaceId;
7627 *(ptr++) = stats->func_total;
7628 *(ptr++) = stats->func_index;
7629 *(ptr++) = hgethi(stats->invocations);
7630 *(ptr++) = hgetlo(stats->invocations);
7631 *(ptr++) = hgethi(stats->bytes_sent);
7632 *(ptr++) = hgetlo(stats->bytes_sent);
7633 *(ptr++) = hgethi(stats->bytes_rcvd);
7634 *(ptr++) = hgetlo(stats->bytes_rcvd);
7635 *(ptr++) = stats->queue_time_sum.sec;
7636 *(ptr++) = stats->queue_time_sum.usec;
7637 *(ptr++) = stats->queue_time_sum_sqr.sec;
7638 *(ptr++) = stats->queue_time_sum_sqr.usec;
7639 *(ptr++) = stats->queue_time_min.sec;
7640 *(ptr++) = stats->queue_time_min.usec;
7641 *(ptr++) = stats->queue_time_max.sec;
7642 *(ptr++) = stats->queue_time_max.usec;
7643 *(ptr++) = stats->execution_time_sum.sec;
7644 *(ptr++) = stats->execution_time_sum.usec;
7645 *(ptr++) = stats->execution_time_sum_sqr.sec;
7646 *(ptr++) = stats->execution_time_sum_sqr.usec;
7647 *(ptr++) = stats->execution_time_min.sec;
7648 *(ptr++) = stats->execution_time_min.usec;
7649 *(ptr++) = stats->execution_time_max.sec;
7650 *(ptr++) = stats->execution_time_max.usec;
7656 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7661 * IN callerVersion - the rpc stat version of the caller
7663 * OUT myVersion - the rpc stat version of this function
7665 * OUT clock_sec - local time seconds
7667 * OUT clock_usec - local time microseconds
7669 * OUT allocSize - the number of bytes allocated to contain stats
7671 * OUT statCount - the number stats retrieved from this process.
7673 * OUT stats - the actual stats retrieved from this process.
7677 * Returns void. If successful, stats will != NULL.
7681 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7682 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7683 size_t * allocSize, afs_uint32 * statCount,
7684 afs_uint32 ** stats)
7694 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7697 * Check to see if stats are enabled
7700 MUTEX_ENTER(&rx_rpc_stats);
7701 if (!rxi_monitor_processStats) {
7702 MUTEX_EXIT(&rx_rpc_stats);
7706 clock_GetTime(&now);
7707 *clock_sec = now.sec;
7708 *clock_usec = now.usec;
7711 * Allocate the space based upon the caller version
7713 * If the client is at an older version than we are,
7714 * we return the statistic data in the older data format, but
7715 * we still return our version number so the client knows we
7716 * are maintaining more data than it can retrieve.
7719 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7720 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7721 *statCount = rxi_rpc_process_stat_cnt;
7724 * This can't happen yet, but in the future version changes
7725 * can be handled by adding additional code here
7729 if (space > (size_t) 0) {
7731 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7734 rx_interface_stat_p rpc_stat, nrpc_stat;
7738 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7740 * Copy the data based upon the caller version
7742 rx_MarshallProcessRPCStats(callerVersion,
7743 rpc_stat->stats[0].func_total,
7744 rpc_stat->stats, &ptr);
7750 MUTEX_EXIT(&rx_rpc_stats);
7755 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7759 * IN callerVersion - the rpc stat version of the caller
7761 * OUT myVersion - the rpc stat version of this function
7763 * OUT clock_sec - local time seconds
7765 * OUT clock_usec - local time microseconds
7767 * OUT allocSize - the number of bytes allocated to contain stats
7769 * OUT statCount - the number of stats retrieved from the individual
7772 * OUT stats - the actual stats retrieved from the individual peer structures.
7776 * Returns void. If successful, stats will != NULL.
7780 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7781 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7782 size_t * allocSize, afs_uint32 * statCount,
7783 afs_uint32 ** stats)
7793 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7796 * Check to see if stats are enabled
7799 MUTEX_ENTER(&rx_rpc_stats);
7800 if (!rxi_monitor_peerStats) {
7801 MUTEX_EXIT(&rx_rpc_stats);
7805 clock_GetTime(&now);
7806 *clock_sec = now.sec;
7807 *clock_usec = now.usec;
7810 * Allocate the space based upon the caller version
7812 * If the client is at an older version than we are,
7813 * we return the statistic data in the older data format, but
7814 * we still return our version number so the client knows we
7815 * are maintaining more data than it can retrieve.
7818 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7819 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7820 *statCount = rxi_rpc_peer_stat_cnt;
7823 * This can't happen yet, but in the future version changes
7824 * can be handled by adding additional code here
7828 if (space > (size_t) 0) {
7830 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7833 rx_interface_stat_p rpc_stat, nrpc_stat;
7837 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7839 * We have to fix the offset of rpc_stat since we are
7840 * keeping this structure on two rx_queues. The rx_queue
7841 * package assumes that the rx_queue member is the first
7842 * member of the structure. That is, rx_queue assumes that
7843 * any one item is only on one queue at a time. We are
7844 * breaking that assumption and so we have to do a little
7845 * math to fix our pointers.
7848 fix_offset = (char *)rpc_stat;
7849 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7850 rpc_stat = (rx_interface_stat_p) fix_offset;
7853 * Copy the data based upon the caller version
7855 rx_MarshallProcessRPCStats(callerVersion,
7856 rpc_stat->stats[0].func_total,
7857 rpc_stat->stats, &ptr);
7863 MUTEX_EXIT(&rx_rpc_stats);
7868 * rx_FreeRPCStats - free memory allocated by
7869 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7873 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7874 * rx_RetrievePeerRPCStats
7876 * IN allocSize - the number of bytes in stats.
7884 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7886 rxi_Free(stats, allocSize);
7890 * rx_queryProcessRPCStats - see if process rpc stat collection is
7891 * currently enabled.
7897 * Returns 0 if stats are not enabled != 0 otherwise
7901 rx_queryProcessRPCStats(void)
7904 MUTEX_ENTER(&rx_rpc_stats);
7905 rc = rxi_monitor_processStats;
7906 MUTEX_EXIT(&rx_rpc_stats);
7911 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7917 * Returns 0 if stats are not enabled != 0 otherwise
7921 rx_queryPeerRPCStats(void)
7924 MUTEX_ENTER(&rx_rpc_stats);
7925 rc = rxi_monitor_peerStats;
7926 MUTEX_EXIT(&rx_rpc_stats);
7931 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7941 rx_enableProcessRPCStats(void)
7943 MUTEX_ENTER(&rx_rpc_stats);
7944 rx_enable_stats = 1;
7945 rxi_monitor_processStats = 1;
7946 MUTEX_EXIT(&rx_rpc_stats);
7950 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7960 rx_enablePeerRPCStats(void)
7962 MUTEX_ENTER(&rx_rpc_stats);
7963 rx_enable_stats = 1;
7964 rxi_monitor_peerStats = 1;
7965 MUTEX_EXIT(&rx_rpc_stats);
7969 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7979 rx_disableProcessRPCStats(void)
7981 rx_interface_stat_p rpc_stat, nrpc_stat;
7984 MUTEX_ENTER(&rx_rpc_stats);
7987 * Turn off process statistics and if peer stats is also off, turn
7991 rxi_monitor_processStats = 0;
7992 if (rxi_monitor_peerStats == 0) {
7993 rx_enable_stats = 0;
7996 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7997 unsigned int num_funcs = 0;
8000 queue_Remove(rpc_stat);
8001 num_funcs = rpc_stat->stats[0].func_total;
8003 sizeof(rx_interface_stat_t) +
8004 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8006 rxi_Free(rpc_stat, space);
8007 rxi_rpc_process_stat_cnt -= num_funcs;
8009 MUTEX_EXIT(&rx_rpc_stats);
8013 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8023 rx_disablePeerRPCStats(void)
8025 struct rx_peer **peer_ptr, **peer_end;
8028 MUTEX_ENTER(&rx_rpc_stats);
8031 * Turn off peer statistics and if process stats is also off, turn
8035 rxi_monitor_peerStats = 0;
8036 if (rxi_monitor_processStats == 0) {
8037 rx_enable_stats = 0;
8040 MUTEX_ENTER(&rx_peerHashTable_lock);
8041 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8042 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8044 struct rx_peer *peer, *next, *prev;
8045 for (prev = peer = *peer_ptr; peer; peer = next) {
8047 code = MUTEX_TRYENTER(&peer->peer_lock);
8049 rx_interface_stat_p rpc_stat, nrpc_stat;
8052 (&peer->rpcStats, rpc_stat, nrpc_stat,
8053 rx_interface_stat)) {
8054 unsigned int num_funcs = 0;
8057 queue_Remove(&rpc_stat->queue_header);
8058 queue_Remove(&rpc_stat->all_peers);
8059 num_funcs = rpc_stat->stats[0].func_total;
8061 sizeof(rx_interface_stat_t) +
8062 rpc_stat->stats[0].func_total *
8063 sizeof(rx_function_entry_v1_t);
8065 rxi_Free(rpc_stat, space);
8066 rxi_rpc_peer_stat_cnt -= num_funcs;
8068 MUTEX_EXIT(&peer->peer_lock);
8069 if (prev == *peer_ptr) {
8079 MUTEX_EXIT(&rx_peerHashTable_lock);
8080 MUTEX_EXIT(&rx_rpc_stats);
8084 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8089 * IN clearFlag - flag indicating which stats to clear
8097 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8099 rx_interface_stat_p rpc_stat, nrpc_stat;
8101 MUTEX_ENTER(&rx_rpc_stats);
8103 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8104 unsigned int num_funcs = 0, i;
8105 num_funcs = rpc_stat->stats[0].func_total;
8106 for (i = 0; i < num_funcs; i++) {
8107 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8108 hzero(rpc_stat->stats[i].invocations);
8110 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8111 hzero(rpc_stat->stats[i].bytes_sent);
8113 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8114 hzero(rpc_stat->stats[i].bytes_rcvd);
8116 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8117 rpc_stat->stats[i].queue_time_sum.sec = 0;
8118 rpc_stat->stats[i].queue_time_sum.usec = 0;
8120 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8121 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8122 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8124 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8125 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8126 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8128 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8129 rpc_stat->stats[i].queue_time_max.sec = 0;
8130 rpc_stat->stats[i].queue_time_max.usec = 0;
8132 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8133 rpc_stat->stats[i].execution_time_sum.sec = 0;
8134 rpc_stat->stats[i].execution_time_sum.usec = 0;
8136 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8137 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8138 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8140 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8141 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8142 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8144 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8145 rpc_stat->stats[i].execution_time_max.sec = 0;
8146 rpc_stat->stats[i].execution_time_max.usec = 0;
8151 MUTEX_EXIT(&rx_rpc_stats);
8155 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8160 * IN clearFlag - flag indicating which stats to clear
8168 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8170 rx_interface_stat_p rpc_stat, nrpc_stat;
8172 MUTEX_ENTER(&rx_rpc_stats);
8174 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8175 unsigned int num_funcs = 0, i;
8178 * We have to fix the offset of rpc_stat since we are
8179 * keeping this structure on two rx_queues. The rx_queue
8180 * package assumes that the rx_queue member is the first
8181 * member of the structure. That is, rx_queue assumes that
8182 * any one item is only on one queue at a time. We are
8183 * breaking that assumption and so we have to do a little
8184 * math to fix our pointers.
8187 fix_offset = (char *)rpc_stat;
8188 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8189 rpc_stat = (rx_interface_stat_p) fix_offset;
8191 num_funcs = rpc_stat->stats[0].func_total;
8192 for (i = 0; i < num_funcs; i++) {
8193 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8194 hzero(rpc_stat->stats[i].invocations);
8196 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8197 hzero(rpc_stat->stats[i].bytes_sent);
8199 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8200 hzero(rpc_stat->stats[i].bytes_rcvd);
8202 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8203 rpc_stat->stats[i].queue_time_sum.sec = 0;
8204 rpc_stat->stats[i].queue_time_sum.usec = 0;
8206 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8207 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8208 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8210 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8211 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8212 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8214 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8215 rpc_stat->stats[i].queue_time_max.sec = 0;
8216 rpc_stat->stats[i].queue_time_max.usec = 0;
8218 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8219 rpc_stat->stats[i].execution_time_sum.sec = 0;
8220 rpc_stat->stats[i].execution_time_sum.usec = 0;
8222 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8223 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8224 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8226 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8227 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8228 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8230 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8231 rpc_stat->stats[i].execution_time_max.sec = 0;
8232 rpc_stat->stats[i].execution_time_max.usec = 0;
8237 MUTEX_EXIT(&rx_rpc_stats);
8241 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8242 * is authorized to enable/disable/clear RX statistics.
8244 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8247 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8249 rxi_rxstat_userok = proc;
8253 rx_RxStatUserOk(struct rx_call *call)
8255 if (!rxi_rxstat_userok)
8257 return rxi_rxstat_userok(call);
8262 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8263 * function in the MSVC runtime DLL (msvcrt.dll).
8265 * Note: the system serializes calls to this function.
8268 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8269 DWORD reason, /* reason function is being called */
8270 LPVOID reserved) /* reserved for future use */
8273 case DLL_PROCESS_ATTACH:
8274 /* library is being attached to a process */
8278 case DLL_PROCESS_DETACH:
8285 #endif /* AFS_NT40_ENV */
8288 int rx_DumpCalls(FILE *outputFile, char *cookie)
8290 #ifdef RXDEBUG_PACKET
8291 #ifdef KDUMP_RX_LOCK
8292 struct rx_call_rx_lock *c;
8299 #define RXDPRINTF sprintf
8300 #define RXDPRINTOUT output
8302 #define RXDPRINTF fprintf
8303 #define RXDPRINTOUT outputFile
8306 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8308 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8311 for (c = rx_allCallsp; c; c = c->allNextp) {
8312 u_short rqc, tqc, iovqc;
8313 struct rx_packet *p, *np;
8315 MUTEX_ENTER(&c->lock);
8316 queue_Count(&c->rq, p, np, rx_packet, rqc);
8317 queue_Count(&c->tq, p, np, rx_packet, tqc);
8318 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8320 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, "
8321 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8322 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8323 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8324 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8325 #ifdef RX_ENABLE_LOCKS
8328 #ifdef RX_REFCOUNT_CHECK
8329 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
8330 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
8333 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,
8334 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
8335 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
8336 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
8337 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
8338 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
8339 #ifdef RX_ENABLE_LOCKS
8340 , (afs_uint32)c->refCount
8342 #ifdef RX_REFCOUNT_CHECK
8343 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
8346 MUTEX_EXIT(&c->lock);
8349 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8352 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
8354 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8356 #endif /* RXDEBUG_PACKET */