2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
13 #include "afs/param.h"
16 # include "afs/sysincludes.h"
17 # include "afsincludes.h"
22 # ifdef AFS_LINUX20_ENV
23 # include "h/socket.h"
25 # include "netinet/in.h"
27 # include "inet/common.h"
29 # include "inet/ip_ire.h"
31 # include "afs/afs_args.h"
32 # include "afs/afs_osi.h"
33 # ifdef RX_KERNEL_TRACE
34 # include "rx_kcommon.h"
36 # if defined(AFS_AIX_ENV)
40 # undef RXDEBUG /* turn off debugging */
42 # if defined(AFS_SGI_ENV)
43 # include "sys/debug.h"
46 # include "afs/sysincludes.h"
47 # include "afsincludes.h"
48 # endif /* !UKERNEL */
49 # include "afs/lock.h"
50 # include "rx_kmutex.h"
51 # include "rx_kernel.h"
52 # define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
53 # define AFSOP_STOP_AFS 211 /* Stop AFS process */
54 # define AFSOP_STOP_BKG 212 /* Stop BKG process */
55 extern afs_int32 afs_termState;
57 # include "sys/lockl.h"
58 # include "sys/lock_def.h"
59 # endif /* AFS_AIX41_ENV */
60 # include "afs/rxgen_consts.h"
63 # include <sys/types.h>
73 # include <afs/afsutil.h>
74 # include <WINNT\afsreg.h>
76 # include <sys/socket.h>
77 # include <sys/file.h>
79 # include <sys/stat.h>
80 # include <netinet/in.h>
81 # include <sys/time.h>
89 #include "rx_atomic.h"
90 #include "rx_globals.h"
92 #include "rx_internal.h"
95 #include <afs/rxgen_consts.h>
98 #ifdef AFS_PTHREAD_ENV
100 int (*registerProgram) (pid_t, char *) = 0;
101 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
104 int (*registerProgram) (PROCESS, char *) = 0;
105 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
109 /* Local static routines */
110 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
111 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
112 struct rx_peer *, struct clock *);
114 #ifdef RX_ENABLE_LOCKS
115 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
118 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
120 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
121 rx_atomic_t rxi_start_in_error;
123 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
126 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
127 * currently allocated within rx. This number is used to allocate the
128 * memory required to return the statistics when queried.
129 * Protected by the rx_rpc_stats mutex.
132 static unsigned int rxi_rpc_peer_stat_cnt;
135 * rxi_rpc_process_stat_cnt counts the total number of local process stat
136 * structures currently allocated within rx. The number is used to allocate
137 * the memory required to return the statistics when queried.
138 * Protected by the rx_rpc_stats mutex.
141 static unsigned int rxi_rpc_process_stat_cnt;
143 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
144 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
146 #if !defined(offsetof)
147 #include <stddef.h> /* for definition of offsetof() */
150 #ifdef RX_ENABLE_LOCKS
151 afs_kmutex_t rx_atomic_mutex;
154 #ifdef AFS_PTHREAD_ENV
157 * Use procedural initialization of mutexes/condition variables
161 extern afs_kmutex_t rx_quota_mutex;
162 extern afs_kmutex_t rx_pthread_mutex;
163 extern afs_kmutex_t rx_packets_mutex;
164 extern afs_kmutex_t rx_refcnt_mutex;
165 extern afs_kmutex_t des_init_mutex;
166 extern afs_kmutex_t des_random_mutex;
167 extern afs_kmutex_t rx_clock_mutex;
168 extern afs_kmutex_t rxi_connCacheMutex;
169 extern afs_kmutex_t rx_event_mutex;
170 extern afs_kmutex_t osi_malloc_mutex;
171 extern afs_kmutex_t event_handler_mutex;
172 extern afs_kmutex_t listener_mutex;
173 extern afs_kmutex_t rx_if_init_mutex;
174 extern afs_kmutex_t rx_if_mutex;
175 extern afs_kmutex_t rxkad_client_uid_mutex;
176 extern afs_kmutex_t rxkad_random_mutex;
178 extern afs_kcondvar_t rx_event_handler_cond;
179 extern afs_kcondvar_t rx_listener_cond;
181 static afs_kmutex_t epoch_mutex;
182 static afs_kmutex_t rx_init_mutex;
183 static afs_kmutex_t rx_debug_mutex;
184 static afs_kmutex_t rx_rpc_stats;
187 rxi_InitPthread(void)
189 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
190 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
191 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
192 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
193 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
194 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
195 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
196 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
197 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
198 MUTEX_INIT(&rx_event_mutex, "event", MUTEX_DEFAULT, 0);
199 MUTEX_INIT(&osi_malloc_mutex, "malloc", MUTEX_DEFAULT, 0);
200 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
201 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
202 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
203 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
204 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
205 MUTEX_INIT(&rxkad_client_uid_mutex, "uid", MUTEX_DEFAULT, 0);
206 MUTEX_INIT(&rxkad_random_mutex, "rxkad random", MUTEX_DEFAULT, 0);
207 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
209 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
210 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
212 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
213 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
215 rxkad_global_stats_init();
217 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
218 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
219 #ifdef RX_ENABLE_LOCKS
222 #endif /* RX_LOCKS_DB */
223 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
224 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
226 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
228 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
230 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
232 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
233 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
234 #endif /* RX_ENABLE_LOCKS */
237 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
238 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
240 * The rx_stats_mutex mutex protects the following global variables:
241 * rxi_lowConnRefCount
242 * rxi_lowPeerRefCount
251 * The rx_quota_mutex mutex protects the following global variables:
259 * The rx_freePktQ_lock protects the following global variables:
264 * The rx_packets_mutex mutex protects the following global variables:
272 * The rx_pthread_mutex mutex protects the following global variables:
273 * rxi_fcfs_thread_num
276 #define INIT_PTHREAD_LOCKS
280 /* Variables for handling the minProcs implementation. availProcs gives the
281 * number of threads available in the pool at this moment (not counting dudes
282 * executing right now). totalMin gives the total number of procs required
283 * for handling all minProcs requests. minDeficit is a dynamic variable
284 * tracking the # of procs required to satisfy all of the remaining minProcs
286 * For fine grain locking to work, the quota check and the reservation of
287 * a server thread has to come while rxi_availProcs and rxi_minDeficit
288 * are locked. To this end, the code has been modified under #ifdef
289 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
290 * same time. A new function, ReturnToServerPool() returns the allocation.
292 * A call can be on several queue's (but only one at a time). When
293 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
294 * that no one else is touching the queue. To this end, we store the address
295 * of the queue lock in the call structure (under the call lock) when we
296 * put the call on a queue, and we clear the call_queue_lock when the
297 * call is removed from a queue (once the call lock has been obtained).
298 * This allows rxi_ResetCall to safely synchronize with others wishing
299 * to manipulate the queue.
302 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
303 static afs_kmutex_t rx_rpc_stats;
304 void rxi_StartUnlocked(struct rxevent *event, void *call,
305 void *arg1, int istack);
308 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
309 ** pretty good that the next packet coming in is from the same connection
310 ** as the last packet, since we're send multiple packets in a transmit window.
312 struct rx_connection *rxLastConn = 0;
314 #ifdef RX_ENABLE_LOCKS
315 /* The locking hierarchy for rx fine grain locking is composed of these
318 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
319 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
320 * call->lock - locks call data fields.
321 * These are independent of each other:
322 * rx_freeCallQueue_lock
327 * serverQueueEntry->lock
328 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
330 * peer->lock - locks peer data fields.
331 * conn_data_lock - that more than one thread is not updating a conn data
332 * field at the same time.
343 * Do we need a lock to protect the peer field in the conn structure?
344 * conn->peer was previously a constant for all intents and so has no
345 * lock protecting this field. The multihomed client delta introduced
346 * a RX code change : change the peer field in the connection structure
347 * to that remote interface from which the last packet for this
348 * connection was sent out. This may become an issue if further changes
351 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
352 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
354 /* rxdb_fileID is used to identify the lock location, along with line#. */
355 static int rxdb_fileID = RXDB_FILE_RX;
356 #endif /* RX_LOCKS_DB */
357 #else /* RX_ENABLE_LOCKS */
358 #define SET_CALL_QUEUE_LOCK(C, L)
359 #define CLEAR_CALL_QUEUE_LOCK(C)
360 #endif /* RX_ENABLE_LOCKS */
361 struct rx_serverQueueEntry *rx_waitForPacket = 0;
362 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
364 /* ------------Exported Interfaces------------- */
366 /* This function allows rxkad to set the epoch to a suitably random number
367 * which rx_NewConnection will use in the future. The principle purpose is to
368 * get rxnull connections to use the same epoch as the rxkad connections do, at
369 * least once the first rxkad connection is established. This is important now
370 * that the host/port addresses aren't used in FindConnection: the uniqueness
371 * of epoch/cid matters and the start time won't do. */
373 #ifdef AFS_PTHREAD_ENV
375 * This mutex protects the following global variables:
379 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
380 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
384 #endif /* AFS_PTHREAD_ENV */
387 rx_SetEpoch(afs_uint32 epoch)
394 /* Initialize rx. A port number may be mentioned, in which case this
395 * becomes the default port number for any service installed later.
396 * If 0 is provided for the port number, a random port will be chosen
397 * by the kernel. Whether this will ever overlap anything in
398 * /etc/services is anybody's guess... Returns 0 on success, -1 on
403 int rxinit_status = 1;
404 #ifdef AFS_PTHREAD_ENV
406 * This mutex protects the following global variables:
410 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
411 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
414 #define UNLOCK_RX_INIT
418 rx_InitHost(u_int host, u_int port)
425 char *htable, *ptable;
432 if (rxinit_status == 0) {
433 tmp_status = rxinit_status;
435 return tmp_status; /* Already started; return previous error code. */
441 if (afs_winsockInit() < 0)
447 * Initialize anything necessary to provide a non-premptive threading
450 rxi_InitializeThreadSupport();
453 /* Allocate and initialize a socket for client and perhaps server
456 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
457 if (rx_socket == OSI_NULLSOCKET) {
461 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
464 #endif /* RX_LOCKS_DB */
465 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
466 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
467 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
468 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
469 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
470 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
471 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
472 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
473 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
475 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
477 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
479 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
481 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
482 #if defined(AFS_HPUX110_ENV)
484 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
485 #endif /* AFS_HPUX110_ENV */
486 #endif /* RX_ENABLE_LOCKS && KERNEL */
489 rx_connDeadTime = 12;
490 rx_tranquil = 0; /* reset flag */
491 rxi_ResetStatistics();
493 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
494 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
495 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
496 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
497 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
498 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
500 /* Malloc up a bunch of packets & buffers */
502 queue_Init(&rx_freePacketQueue);
503 rxi_NeedMorePackets = FALSE;
504 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
506 /* enforce a minimum number of allocated packets */
507 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
508 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
510 /* allocate the initial free packet pool */
511 #ifdef RX_ENABLE_TSFPQ
512 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
513 #else /* RX_ENABLE_TSFPQ */
514 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
515 #endif /* RX_ENABLE_TSFPQ */
522 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
523 tv.tv_sec = clock_now.sec;
524 tv.tv_usec = clock_now.usec;
525 srand((unsigned int)tv.tv_usec);
532 #if defined(KERNEL) && !defined(UKERNEL)
533 /* Really, this should never happen in a real kernel */
536 struct sockaddr_in addr;
538 int addrlen = sizeof(addr);
540 socklen_t addrlen = sizeof(addr);
542 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
546 rx_port = addr.sin_port;
549 rx_stats.minRtt.sec = 9999999;
551 rx_SetEpoch(tv.tv_sec | 0x80000000);
553 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
554 * will provide a randomer value. */
556 MUTEX_ENTER(&rx_quota_mutex);
557 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
558 MUTEX_EXIT(&rx_quota_mutex);
559 /* *Slightly* random start time for the cid. This is just to help
560 * out with the hashing function at the peer */
561 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
562 rx_connHashTable = (struct rx_connection **)htable;
563 rx_peerHashTable = (struct rx_peer **)ptable;
565 rx_lastAckDelay.sec = 0;
566 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
567 rx_hardAckDelay.sec = 0;
568 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
569 rx_softAckDelay.sec = 0;
570 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
572 rxevent_Init(20, rxi_ReScheduleEvents);
574 /* Initialize various global queues */
575 queue_Init(&rx_idleServerQueue);
576 queue_Init(&rx_incomingCallQueue);
577 queue_Init(&rx_freeCallQueue);
579 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
580 /* Initialize our list of usable IP addresses. */
584 /* Start listener process (exact function is dependent on the
585 * implementation environment--kernel or user space) */
589 tmp_status = rxinit_status = 0;
597 return rx_InitHost(htonl(INADDR_ANY), port);
600 /* called with unincremented nRequestsRunning to see if it is OK to start
601 * a new thread in this service. Could be "no" for two reasons: over the
602 * max quota, or would prevent others from reaching their min quota.
604 #ifdef RX_ENABLE_LOCKS
605 /* This verion of QuotaOK reserves quota if it's ok while the
606 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
609 QuotaOK(struct rx_service *aservice)
611 /* check if over max quota */
612 if (aservice->nRequestsRunning >= aservice->maxProcs) {
616 /* under min quota, we're OK */
617 /* otherwise, can use only if there are enough to allow everyone
618 * to go to their min quota after this guy starts.
621 MUTEX_ENTER(&rx_quota_mutex);
622 if ((aservice->nRequestsRunning < aservice->minProcs)
623 || (rxi_availProcs > rxi_minDeficit)) {
624 aservice->nRequestsRunning++;
625 /* just started call in minProcs pool, need fewer to maintain
627 if (aservice->nRequestsRunning <= aservice->minProcs)
630 MUTEX_EXIT(&rx_quota_mutex);
633 MUTEX_EXIT(&rx_quota_mutex);
639 ReturnToServerPool(struct rx_service *aservice)
641 aservice->nRequestsRunning--;
642 MUTEX_ENTER(&rx_quota_mutex);
643 if (aservice->nRequestsRunning < aservice->minProcs)
646 MUTEX_EXIT(&rx_quota_mutex);
649 #else /* RX_ENABLE_LOCKS */
651 QuotaOK(struct rx_service *aservice)
654 /* under min quota, we're OK */
655 if (aservice->nRequestsRunning < aservice->minProcs)
658 /* check if over max quota */
659 if (aservice->nRequestsRunning >= aservice->maxProcs)
662 /* otherwise, can use only if there are enough to allow everyone
663 * to go to their min quota after this guy starts.
665 MUTEX_ENTER(&rx_quota_mutex);
666 if (rxi_availProcs > rxi_minDeficit)
668 MUTEX_EXIT(&rx_quota_mutex);
671 #endif /* RX_ENABLE_LOCKS */
674 /* Called by rx_StartServer to start up lwp's to service calls.
675 NExistingProcs gives the number of procs already existing, and which
676 therefore needn't be created. */
678 rxi_StartServerProcs(int nExistingProcs)
680 struct rx_service *service;
685 /* For each service, reserve N processes, where N is the "minimum"
686 * number of processes that MUST be able to execute a request in parallel,
687 * at any time, for that process. Also compute the maximum difference
688 * between any service's maximum number of processes that can run
689 * (i.e. the maximum number that ever will be run, and a guarantee
690 * that this number will run if other services aren't running), and its
691 * minimum number. The result is the extra number of processes that
692 * we need in order to provide the latter guarantee */
693 for (i = 0; i < RX_MAX_SERVICES; i++) {
695 service = rx_services[i];
696 if (service == (struct rx_service *)0)
698 nProcs += service->minProcs;
699 diff = service->maxProcs - service->minProcs;
703 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
704 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
705 for (i = 0; i < nProcs; i++) {
706 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
712 /* This routine is only required on Windows */
714 rx_StartClientThread(void)
716 #ifdef AFS_PTHREAD_ENV
718 pid = pthread_self();
719 #endif /* AFS_PTHREAD_ENV */
721 #endif /* AFS_NT40_ENV */
723 /* This routine must be called if any services are exported. If the
724 * donateMe flag is set, the calling process is donated to the server
727 rx_StartServer(int donateMe)
729 struct rx_service *service;
735 /* Start server processes, if necessary (exact function is dependent
736 * on the implementation environment--kernel or user space). DonateMe
737 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
738 * case, one less new proc will be created rx_StartServerProcs.
740 rxi_StartServerProcs(donateMe);
742 /* count up the # of threads in minProcs, and add set the min deficit to
743 * be that value, too.
745 for (i = 0; i < RX_MAX_SERVICES; i++) {
746 service = rx_services[i];
747 if (service == (struct rx_service *)0)
749 MUTEX_ENTER(&rx_quota_mutex);
750 rxi_totalMin += service->minProcs;
751 /* below works even if a thread is running, since minDeficit would
752 * still have been decremented and later re-incremented.
754 rxi_minDeficit += service->minProcs;
755 MUTEX_EXIT(&rx_quota_mutex);
758 /* Turn on reaping of idle server connections */
759 rxi_ReapConnections(NULL, NULL, NULL);
768 #ifdef AFS_PTHREAD_ENV
770 pid = afs_pointer_to_int(pthread_self());
771 #else /* AFS_PTHREAD_ENV */
773 LWP_CurrentProcess(&pid);
774 #endif /* AFS_PTHREAD_ENV */
776 sprintf(name, "srv_%d", ++nProcs);
778 (*registerProgram) (pid, name);
780 #endif /* AFS_NT40_ENV */
781 rx_ServerProc(NULL); /* Never returns */
783 #ifdef RX_ENABLE_TSFPQ
784 /* no use leaving packets around in this thread's local queue if
785 * it isn't getting donated to the server thread pool.
787 rxi_FlushLocalPacketsTSFPQ();
788 #endif /* RX_ENABLE_TSFPQ */
792 /* Create a new client connection to the specified service, using the
793 * specified security object to implement the security model for this
795 struct rx_connection *
796 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
797 struct rx_securityClass *securityObject,
798 int serviceSecurityIndex)
802 struct rx_connection *conn;
807 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
808 "serviceSecurityIndex %d)\n",
809 ntohl(shost), ntohs(sport), sservice, securityObject,
810 serviceSecurityIndex));
812 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
813 * the case of kmem_alloc? */
814 conn = rxi_AllocConnection();
815 #ifdef RX_ENABLE_LOCKS
816 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
817 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
818 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
821 MUTEX_ENTER(&rx_connHashTable_lock);
822 cid = (rx_nextCid += RX_MAXCALLS);
823 conn->type = RX_CLIENT_CONNECTION;
825 conn->epoch = rx_epoch;
826 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
827 conn->serviceId = sservice;
828 conn->securityObject = securityObject;
829 conn->securityData = (void *) 0;
830 conn->securityIndex = serviceSecurityIndex;
831 rx_SetConnDeadTime(conn, rx_connDeadTime);
832 rx_SetConnSecondsUntilNatPing(conn, 0);
833 conn->ackRate = RX_FAST_ACK_RATE;
835 conn->specific = NULL;
836 conn->challengeEvent = NULL;
837 conn->delayedAbortEvent = NULL;
838 conn->abortCount = 0;
840 for (i = 0; i < RX_MAXCALLS; i++) {
841 conn->twind[i] = rx_initSendWindow;
842 conn->rwind[i] = rx_initReceiveWindow;
845 RXS_NewConnection(securityObject, conn);
847 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
849 conn->refCount++; /* no lock required since only this thread knows... */
850 conn->next = rx_connHashTable[hashindex];
851 rx_connHashTable[hashindex] = conn;
853 rx_atomic_inc(&rx_stats.nClientConns);
854 MUTEX_EXIT(&rx_connHashTable_lock);
860 * Ensure a connection's timeout values are valid.
862 * @param[in] conn The connection to check
864 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
865 * unless idleDeadTime and/or hardDeadTime are not set
869 rxi_CheckConnTimeouts(struct rx_connection *conn)
871 /* a connection's timeouts must have the relationship
872 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
873 * total loss of network to a peer may cause an idle timeout instead of a
874 * dead timeout, simply because the idle timeout gets hit first. Also set
875 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
876 /* this logic is slightly complicated by the fact that
877 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
879 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
880 if (conn->idleDeadTime) {
881 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
883 if (conn->hardDeadTime) {
884 if (conn->idleDeadTime) {
885 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
887 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
893 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
895 /* The idea is to set the dead time to a value that allows several
896 * keepalives to be dropped without timing out the connection. */
897 conn->secondsUntilDead = seconds;
898 rxi_CheckConnTimeouts(conn);
899 conn->secondsUntilPing = conn->secondsUntilDead / 6;
903 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
905 conn->hardDeadTime = seconds;
906 rxi_CheckConnTimeouts(conn);
910 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
912 conn->idleDeadTime = seconds;
913 rxi_CheckConnTimeouts(conn);
916 int rxi_lowPeerRefCount = 0;
917 int rxi_lowConnRefCount = 0;
920 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
921 * NOTE: must not be called with rx_connHashTable_lock held.
924 rxi_CleanupConnection(struct rx_connection *conn)
926 /* Notify the service exporter, if requested, that this connection
927 * is being destroyed */
928 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
929 (*conn->service->destroyConnProc) (conn);
931 /* Notify the security module that this connection is being destroyed */
932 RXS_DestroyConnection(conn->securityObject, conn);
934 /* If this is the last connection using the rx_peer struct, set its
935 * idle time to now. rxi_ReapConnections will reap it if it's still
936 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
938 MUTEX_ENTER(&rx_peerHashTable_lock);
939 if (conn->peer->refCount < 2) {
940 conn->peer->idleWhen = clock_Sec();
941 if (conn->peer->refCount < 1) {
942 conn->peer->refCount = 1;
943 if (rx_stats_active) {
944 MUTEX_ENTER(&rx_stats_mutex);
945 rxi_lowPeerRefCount++;
946 MUTEX_EXIT(&rx_stats_mutex);
950 conn->peer->refCount--;
951 MUTEX_EXIT(&rx_peerHashTable_lock);
955 if (conn->type == RX_SERVER_CONNECTION)
956 rx_atomic_dec(&rx_stats.nServerConns);
958 rx_atomic_dec(&rx_stats.nClientConns);
961 if (conn->specific) {
963 for (i = 0; i < conn->nSpecific; i++) {
964 if (conn->specific[i] && rxi_keyCreate_destructor[i])
965 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
966 conn->specific[i] = NULL;
968 free(conn->specific);
970 conn->specific = NULL;
974 MUTEX_DESTROY(&conn->conn_call_lock);
975 MUTEX_DESTROY(&conn->conn_data_lock);
976 CV_DESTROY(&conn->conn_call_cv);
978 rxi_FreeConnection(conn);
981 /* Destroy the specified connection */
983 rxi_DestroyConnection(struct rx_connection *conn)
985 MUTEX_ENTER(&rx_connHashTable_lock);
986 rxi_DestroyConnectionNoLock(conn);
987 /* conn should be at the head of the cleanup list */
988 if (conn == rx_connCleanup_list) {
989 rx_connCleanup_list = rx_connCleanup_list->next;
990 MUTEX_EXIT(&rx_connHashTable_lock);
991 rxi_CleanupConnection(conn);
993 #ifdef RX_ENABLE_LOCKS
995 MUTEX_EXIT(&rx_connHashTable_lock);
997 #endif /* RX_ENABLE_LOCKS */
1001 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1003 struct rx_connection **conn_ptr;
1005 struct rx_packet *packet;
1012 MUTEX_ENTER(&conn->conn_data_lock);
1013 MUTEX_ENTER(&rx_refcnt_mutex);
1014 if (conn->refCount > 0)
1017 if (rx_stats_active) {
1018 MUTEX_ENTER(&rx_stats_mutex);
1019 rxi_lowConnRefCount++;
1020 MUTEX_EXIT(&rx_stats_mutex);
1024 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1025 /* Busy; wait till the last guy before proceeding */
1026 MUTEX_EXIT(&rx_refcnt_mutex);
1027 MUTEX_EXIT(&conn->conn_data_lock);
1032 /* If the client previously called rx_NewCall, but it is still
1033 * waiting, treat this as a running call, and wait to destroy the
1034 * connection later when the call completes. */
1035 if ((conn->type == RX_CLIENT_CONNECTION)
1036 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1037 conn->flags |= RX_CONN_DESTROY_ME;
1038 MUTEX_EXIT(&conn->conn_data_lock);
1042 MUTEX_EXIT(&rx_refcnt_mutex);
1043 MUTEX_EXIT(&conn->conn_data_lock);
1045 /* Check for extant references to this connection */
1046 for (i = 0; i < RX_MAXCALLS; i++) {
1047 struct rx_call *call = conn->call[i];
1050 if (conn->type == RX_CLIENT_CONNECTION) {
1051 MUTEX_ENTER(&call->lock);
1052 if (call->delayedAckEvent) {
1053 /* Push the final acknowledgment out now--there
1054 * won't be a subsequent call to acknowledge the
1055 * last reply packets */
1056 rxevent_Cancel(call->delayedAckEvent, call,
1057 RX_CALL_REFCOUNT_DELAY);
1058 if (call->state == RX_STATE_PRECALL
1059 || call->state == RX_STATE_ACTIVE) {
1060 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1062 rxi_AckAll(NULL, call, 0);
1065 MUTEX_EXIT(&call->lock);
1069 #ifdef RX_ENABLE_LOCKS
1071 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1072 MUTEX_EXIT(&conn->conn_data_lock);
1074 /* Someone is accessing a packet right now. */
1078 #endif /* RX_ENABLE_LOCKS */
1081 /* Don't destroy the connection if there are any call
1082 * structures still in use */
1083 MUTEX_ENTER(&conn->conn_data_lock);
1084 conn->flags |= RX_CONN_DESTROY_ME;
1085 MUTEX_EXIT(&conn->conn_data_lock);
1090 if (conn->natKeepAliveEvent) {
1091 rxi_NatKeepAliveOff(conn);
1094 if (conn->delayedAbortEvent) {
1095 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1096 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1098 MUTEX_ENTER(&conn->conn_data_lock);
1099 rxi_SendConnectionAbort(conn, packet, 0, 1);
1100 MUTEX_EXIT(&conn->conn_data_lock);
1101 rxi_FreePacket(packet);
1105 /* Remove from connection hash table before proceeding */
1107 &rx_connHashTable[CONN_HASH
1108 (peer->host, peer->port, conn->cid, conn->epoch,
1110 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1111 if (*conn_ptr == conn) {
1112 *conn_ptr = conn->next;
1116 /* if the conn that we are destroying was the last connection, then we
1117 * clear rxLastConn as well */
1118 if (rxLastConn == conn)
1121 /* Make sure the connection is completely reset before deleting it. */
1122 /* get rid of pending events that could zap us later */
1123 if (conn->challengeEvent)
1124 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1125 if (conn->checkReachEvent)
1126 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1127 if (conn->natKeepAliveEvent)
1128 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
1130 /* Add the connection to the list of destroyed connections that
1131 * need to be cleaned up. This is necessary to avoid deadlocks
1132 * in the routines we call to inform others that this connection is
1133 * being destroyed. */
1134 conn->next = rx_connCleanup_list;
1135 rx_connCleanup_list = conn;
1138 /* Externally available version */
1140 rx_DestroyConnection(struct rx_connection *conn)
1145 rxi_DestroyConnection(conn);
1150 rx_GetConnection(struct rx_connection *conn)
1155 MUTEX_ENTER(&rx_refcnt_mutex);
1157 MUTEX_EXIT(&rx_refcnt_mutex);
1161 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1162 /* Wait for the transmit queue to no longer be busy.
1163 * requires the call->lock to be held */
1165 rxi_WaitforTQBusy(struct rx_call *call) {
1166 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1167 call->flags |= RX_CALL_TQ_WAIT;
1169 #ifdef RX_ENABLE_LOCKS
1170 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1171 CV_WAIT(&call->cv_tq, &call->lock);
1172 #else /* RX_ENABLE_LOCKS */
1173 osi_rxSleep(&call->tq);
1174 #endif /* RX_ENABLE_LOCKS */
1176 if (call->tqWaiters == 0) {
1177 call->flags &= ~RX_CALL_TQ_WAIT;
1184 rxi_WakeUpTransmitQueue(struct rx_call *call)
1186 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1187 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1188 call, call->tqWaiters, call->flags));
1189 #ifdef RX_ENABLE_LOCKS
1190 osirx_AssertMine(&call->lock, "rxi_Start start");
1191 CV_BROADCAST(&call->cv_tq);
1192 #else /* RX_ENABLE_LOCKS */
1193 osi_rxWakeup(&call->tq);
1194 #endif /* RX_ENABLE_LOCKS */
1198 /* Start a new rx remote procedure call, on the specified connection.
1199 * If wait is set to 1, wait for a free call channel; otherwise return
1200 * 0. Maxtime gives the maximum number of seconds this call may take,
1201 * after rx_NewCall returns. After this time interval, a call to any
1202 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1203 * For fine grain locking, we hold the conn_call_lock in order to
1204 * to ensure that we don't get signalle after we found a call in an active
1205 * state and before we go to sleep.
1208 rx_NewCall(struct rx_connection *conn)
1211 struct rx_call *call;
1212 struct clock queueTime;
1216 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1219 clock_GetTime(&queueTime);
1221 * Check if there are others waiting for a new call.
1222 * If so, let them go first to avoid starving them.
1223 * This is a fairly simple scheme, and might not be
1224 * a complete solution for large numbers of waiters.
1226 * makeCallWaiters keeps track of the number of
1227 * threads waiting to make calls and the
1228 * RX_CONN_MAKECALL_WAITING flag bit is used to
1229 * indicate that there are indeed calls waiting.
1230 * The flag is set when the waiter is incremented.
1231 * It is only cleared when makeCallWaiters is 0.
1232 * This prevents us from accidently destroying the
1233 * connection while it is potentially about to be used.
1235 MUTEX_ENTER(&conn->conn_call_lock);
1236 MUTEX_ENTER(&conn->conn_data_lock);
1237 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1238 conn->flags |= RX_CONN_MAKECALL_WAITING;
1239 conn->makeCallWaiters++;
1240 MUTEX_EXIT(&conn->conn_data_lock);
1242 #ifdef RX_ENABLE_LOCKS
1243 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1247 MUTEX_ENTER(&conn->conn_data_lock);
1248 conn->makeCallWaiters--;
1249 if (conn->makeCallWaiters == 0)
1250 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1253 /* We are now the active thread in rx_NewCall */
1254 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1255 MUTEX_EXIT(&conn->conn_data_lock);
1260 for (i = 0; i < RX_MAXCALLS; i++) {
1261 call = conn->call[i];
1263 if (call->state == RX_STATE_DALLY) {
1264 MUTEX_ENTER(&call->lock);
1265 if (call->state == RX_STATE_DALLY) {
1267 * We are setting the state to RX_STATE_RESET to
1268 * ensure that no one else will attempt to use this
1269 * call once we drop the conn->conn_call_lock and
1270 * call->lock. We must drop the conn->conn_call_lock
1271 * before calling rxi_ResetCall because the process
1272 * of clearing the transmit queue can block for an
1273 * extended period of time. If we block while holding
1274 * the conn->conn_call_lock, then all rx_EndCall
1275 * processing will block as well. This has a detrimental
1276 * effect on overall system performance.
1278 call->state = RX_STATE_RESET;
1279 MUTEX_EXIT(&conn->conn_call_lock);
1280 MUTEX_ENTER(&rx_refcnt_mutex);
1281 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1282 MUTEX_EXIT(&rx_refcnt_mutex);
1283 rxi_ResetCall(call, 0);
1284 (*call->callNumber)++;
1285 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1289 * If we failed to be able to safely obtain the
1290 * conn->conn_call_lock we will have to drop the
1291 * call->lock to avoid a deadlock. When the call->lock
1292 * is released the state of the call can change. If it
1293 * is no longer RX_STATE_RESET then some other thread is
1296 MUTEX_EXIT(&call->lock);
1297 MUTEX_ENTER(&conn->conn_call_lock);
1298 MUTEX_ENTER(&call->lock);
1300 if (call->state == RX_STATE_RESET)
1304 * If we get here it means that after dropping
1305 * the conn->conn_call_lock and call->lock that
1306 * the call is no longer ours. If we can't find
1307 * a free call in the remaining slots we should
1308 * not go immediately to RX_CONN_MAKECALL_WAITING
1309 * because by dropping the conn->conn_call_lock
1310 * we have given up synchronization with rx_EndCall.
1311 * Instead, cycle through one more time to see if
1312 * we can find a call that can call our own.
1314 MUTEX_ENTER(&rx_refcnt_mutex);
1315 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1316 MUTEX_EXIT(&rx_refcnt_mutex);
1319 MUTEX_EXIT(&call->lock);
1322 /* rxi_NewCall returns with mutex locked */
1323 call = rxi_NewCall(conn, i);
1324 MUTEX_ENTER(&rx_refcnt_mutex);
1325 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1326 MUTEX_EXIT(&rx_refcnt_mutex);
1330 if (i < RX_MAXCALLS) {
1336 MUTEX_ENTER(&conn->conn_data_lock);
1337 conn->flags |= RX_CONN_MAKECALL_WAITING;
1338 conn->makeCallWaiters++;
1339 MUTEX_EXIT(&conn->conn_data_lock);
1341 #ifdef RX_ENABLE_LOCKS
1342 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1346 MUTEX_ENTER(&conn->conn_data_lock);
1347 conn->makeCallWaiters--;
1348 if (conn->makeCallWaiters == 0)
1349 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1350 MUTEX_EXIT(&conn->conn_data_lock);
1352 /* Client is initially in send mode */
1353 call->state = RX_STATE_ACTIVE;
1354 call->error = conn->error;
1356 call->mode = RX_MODE_ERROR;
1358 call->mode = RX_MODE_SENDING;
1360 /* remember start time for call in case we have hard dead time limit */
1361 call->queueTime = queueTime;
1362 clock_GetTime(&call->startTime);
1363 hzero(call->bytesSent);
1364 hzero(call->bytesRcvd);
1366 /* Turn on busy protocol. */
1367 rxi_KeepAliveOn(call);
1369 /* Attempt MTU discovery */
1370 rxi_GrowMTUOn(call);
1373 * We are no longer the active thread in rx_NewCall
1375 MUTEX_ENTER(&conn->conn_data_lock);
1376 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1377 MUTEX_EXIT(&conn->conn_data_lock);
1380 * Wake up anyone else who might be giving us a chance to
1381 * run (see code above that avoids resource starvation).
1383 #ifdef RX_ENABLE_LOCKS
1384 CV_BROADCAST(&conn->conn_call_cv);
1388 MUTEX_EXIT(&conn->conn_call_lock);
1390 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1391 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1392 osi_Panic("rx_NewCall call about to be used without an empty tq");
1394 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1396 MUTEX_EXIT(&call->lock);
1399 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1404 rxi_HasActiveCalls(struct rx_connection *aconn)
1407 struct rx_call *tcall;
1411 for (i = 0; i < RX_MAXCALLS; i++) {
1412 if ((tcall = aconn->call[i])) {
1413 if ((tcall->state == RX_STATE_ACTIVE)
1414 || (tcall->state == RX_STATE_PRECALL)) {
1425 rxi_GetCallNumberVector(struct rx_connection *aconn,
1426 afs_int32 * aint32s)
1429 struct rx_call *tcall;
1433 for (i = 0; i < RX_MAXCALLS; i++) {
1434 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1435 aint32s[i] = aconn->callNumber[i] + 1;
1437 aint32s[i] = aconn->callNumber[i];
1444 rxi_SetCallNumberVector(struct rx_connection *aconn,
1445 afs_int32 * aint32s)
1448 struct rx_call *tcall;
1452 for (i = 0; i < RX_MAXCALLS; i++) {
1453 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1454 aconn->callNumber[i] = aint32s[i] - 1;
1456 aconn->callNumber[i] = aint32s[i];
1462 /* Advertise a new service. A service is named locally by a UDP port
1463 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1466 char *serviceName; Name for identification purposes (e.g. the
1467 service name might be used for probing for
1470 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1471 char *serviceName, struct rx_securityClass **securityObjects,
1472 int nSecurityObjects,
1473 afs_int32(*serviceProc) (struct rx_call * acall))
1475 osi_socket socket = OSI_NULLSOCKET;
1476 struct rx_service *tservice;
1482 if (serviceId == 0) {
1484 "rx_NewService: service id for service %s is not non-zero.\n",
1491 "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",
1499 tservice = rxi_AllocService();
1502 #ifdef RX_ENABLE_LOCKS
1503 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1506 for (i = 0; i < RX_MAX_SERVICES; i++) {
1507 struct rx_service *service = rx_services[i];
1509 if (port == service->servicePort && host == service->serviceHost) {
1510 if (service->serviceId == serviceId) {
1511 /* The identical service has already been
1512 * installed; if the caller was intending to
1513 * change the security classes used by this
1514 * service, he/she loses. */
1516 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1517 serviceName, serviceId, service->serviceName);
1519 rxi_FreeService(tservice);
1522 /* Different service, same port: re-use the socket
1523 * which is bound to the same port */
1524 socket = service->socket;
1527 if (socket == OSI_NULLSOCKET) {
1528 /* If we don't already have a socket (from another
1529 * service on same port) get a new one */
1530 socket = rxi_GetHostUDPSocket(host, port);
1531 if (socket == OSI_NULLSOCKET) {
1533 rxi_FreeService(tservice);
1538 service->socket = socket;
1539 service->serviceHost = host;
1540 service->servicePort = port;
1541 service->serviceId = serviceId;
1542 service->serviceName = serviceName;
1543 service->nSecurityObjects = nSecurityObjects;
1544 service->securityObjects = securityObjects;
1545 service->minProcs = 0;
1546 service->maxProcs = 1;
1547 service->idleDeadTime = 60;
1548 service->idleDeadErr = 0;
1549 service->connDeadTime = rx_connDeadTime;
1550 service->executeRequestProc = serviceProc;
1551 service->checkReach = 0;
1552 service->nSpecific = 0;
1553 service->specific = NULL;
1554 rx_services[i] = service; /* not visible until now */
1560 rxi_FreeService(tservice);
1561 (osi_Msg "rx_NewService: cannot support > %d services\n",
1566 /* Set configuration options for all of a service's security objects */
1569 rx_SetSecurityConfiguration(struct rx_service *service,
1570 rx_securityConfigVariables type,
1574 for (i = 0; i<service->nSecurityObjects; i++) {
1575 if (service->securityObjects[i]) {
1576 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1584 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1585 struct rx_securityClass **securityObjects, int nSecurityObjects,
1586 afs_int32(*serviceProc) (struct rx_call * acall))
1588 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1591 /* Generic request processing loop. This routine should be called
1592 * by the implementation dependent rx_ServerProc. If socketp is
1593 * non-null, it will be set to the file descriptor that this thread
1594 * is now listening on. If socketp is null, this routine will never
1597 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1599 struct rx_call *call;
1601 struct rx_service *tservice = NULL;
1608 call = rx_GetCall(threadID, tservice, socketp);
1609 if (socketp && *socketp != OSI_NULLSOCKET) {
1610 /* We are now a listener thread */
1615 /* if server is restarting( typically smooth shutdown) then do not
1616 * allow any new calls.
1619 if (rx_tranquil && (call != NULL)) {
1623 MUTEX_ENTER(&call->lock);
1625 rxi_CallError(call, RX_RESTARTING);
1626 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1628 MUTEX_EXIT(&call->lock);
1632 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1633 #ifdef RX_ENABLE_LOCKS
1635 #endif /* RX_ENABLE_LOCKS */
1636 afs_termState = AFSOP_STOP_AFS;
1637 afs_osi_Wakeup(&afs_termState);
1638 #ifdef RX_ENABLE_LOCKS
1640 #endif /* RX_ENABLE_LOCKS */
1645 tservice = call->conn->service;
1647 if (tservice->beforeProc)
1648 (*tservice->beforeProc) (call);
1650 code = tservice->executeRequestProc(call);
1652 if (tservice->afterProc)
1653 (*tservice->afterProc) (call, code);
1655 rx_EndCall(call, code);
1656 if (rx_stats_active) {
1657 MUTEX_ENTER(&rx_stats_mutex);
1659 MUTEX_EXIT(&rx_stats_mutex);
1666 rx_WakeupServerProcs(void)
1668 struct rx_serverQueueEntry *np, *tqp;
1672 MUTEX_ENTER(&rx_serverPool_lock);
1674 #ifdef RX_ENABLE_LOCKS
1675 if (rx_waitForPacket)
1676 CV_BROADCAST(&rx_waitForPacket->cv);
1677 #else /* RX_ENABLE_LOCKS */
1678 if (rx_waitForPacket)
1679 osi_rxWakeup(rx_waitForPacket);
1680 #endif /* RX_ENABLE_LOCKS */
1681 MUTEX_ENTER(&freeSQEList_lock);
1682 for (np = rx_FreeSQEList; np; np = tqp) {
1683 tqp = *(struct rx_serverQueueEntry **)np;
1684 #ifdef RX_ENABLE_LOCKS
1685 CV_BROADCAST(&np->cv);
1686 #else /* RX_ENABLE_LOCKS */
1688 #endif /* RX_ENABLE_LOCKS */
1690 MUTEX_EXIT(&freeSQEList_lock);
1691 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1692 #ifdef RX_ENABLE_LOCKS
1693 CV_BROADCAST(&np->cv);
1694 #else /* RX_ENABLE_LOCKS */
1696 #endif /* RX_ENABLE_LOCKS */
1698 MUTEX_EXIT(&rx_serverPool_lock);
1703 * One thing that seems to happen is that all the server threads get
1704 * tied up on some empty or slow call, and then a whole bunch of calls
1705 * arrive at once, using up the packet pool, so now there are more
1706 * empty calls. The most critical resources here are server threads
1707 * and the free packet pool. The "doreclaim" code seems to help in
1708 * general. I think that eventually we arrive in this state: there
1709 * are lots of pending calls which do have all their packets present,
1710 * so they won't be reclaimed, are multi-packet calls, so they won't
1711 * be scheduled until later, and thus are tying up most of the free
1712 * packet pool for a very long time.
1714 * 1. schedule multi-packet calls if all the packets are present.
1715 * Probably CPU-bound operation, useful to return packets to pool.
1716 * Do what if there is a full window, but the last packet isn't here?
1717 * 3. preserve one thread which *only* runs "best" calls, otherwise
1718 * it sleeps and waits for that type of call.
1719 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1720 * the current dataquota business is badly broken. The quota isn't adjusted
1721 * to reflect how many packets are presently queued for a running call.
1722 * So, when we schedule a queued call with a full window of packets queued
1723 * up for it, that *should* free up a window full of packets for other 2d-class
1724 * calls to be able to use from the packet pool. But it doesn't.
1726 * NB. Most of the time, this code doesn't run -- since idle server threads
1727 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1728 * as a new call arrives.
1730 /* Sleep until a call arrives. Returns a pointer to the call, ready
1731 * for an rx_Read. */
1732 #ifdef RX_ENABLE_LOCKS
1734 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1736 struct rx_serverQueueEntry *sq;
1737 struct rx_call *call = (struct rx_call *)0;
1738 struct rx_service *service = NULL;
1741 MUTEX_ENTER(&freeSQEList_lock);
1743 if ((sq = rx_FreeSQEList)) {
1744 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1745 MUTEX_EXIT(&freeSQEList_lock);
1746 } else { /* otherwise allocate a new one and return that */
1747 MUTEX_EXIT(&freeSQEList_lock);
1748 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1749 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1750 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1753 MUTEX_ENTER(&rx_serverPool_lock);
1754 if (cur_service != NULL) {
1755 ReturnToServerPool(cur_service);
1758 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1759 struct rx_call *tcall, *ncall, *choice2 = NULL;
1761 /* Scan for eligible incoming calls. A call is not eligible
1762 * if the maximum number of calls for its service type are
1763 * already executing */
1764 /* One thread will process calls FCFS (to prevent starvation),
1765 * while the other threads may run ahead looking for calls which
1766 * have all their input data available immediately. This helps
1767 * keep threads from blocking, waiting for data from the client. */
1768 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1769 service = tcall->conn->service;
1770 if (!QuotaOK(service)) {
1773 MUTEX_ENTER(&rx_pthread_mutex);
1774 if (tno == rxi_fcfs_thread_num
1775 || !tcall->queue_item_header.next) {
1776 MUTEX_EXIT(&rx_pthread_mutex);
1777 /* If we're the fcfs thread , then we'll just use
1778 * this call. If we haven't been able to find an optimal
1779 * choice, and we're at the end of the list, then use a
1780 * 2d choice if one has been identified. Otherwise... */
1781 call = (choice2 ? choice2 : tcall);
1782 service = call->conn->service;
1784 MUTEX_EXIT(&rx_pthread_mutex);
1785 if (!queue_IsEmpty(&tcall->rq)) {
1786 struct rx_packet *rp;
1787 rp = queue_First(&tcall->rq, rx_packet);
1788 if (rp->header.seq == 1) {
1790 || (rp->header.flags & RX_LAST_PACKET)) {
1792 } else if (rxi_2dchoice && !choice2
1793 && !(tcall->flags & RX_CALL_CLEARED)
1794 && (tcall->rprev > rxi_HardAckRate)) {
1804 ReturnToServerPool(service);
1811 MUTEX_EXIT(&rx_serverPool_lock);
1812 MUTEX_ENTER(&call->lock);
1814 if (call->flags & RX_CALL_WAIT_PROC) {
1815 call->flags &= ~RX_CALL_WAIT_PROC;
1816 rx_atomic_dec(&rx_nWaiting);
1819 if (call->state != RX_STATE_PRECALL || call->error) {
1820 MUTEX_EXIT(&call->lock);
1821 MUTEX_ENTER(&rx_serverPool_lock);
1822 ReturnToServerPool(service);
1827 if (queue_IsEmpty(&call->rq)
1828 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1829 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1831 CLEAR_CALL_QUEUE_LOCK(call);
1834 /* If there are no eligible incoming calls, add this process
1835 * to the idle server queue, to wait for one */
1839 *socketp = OSI_NULLSOCKET;
1841 sq->socketp = socketp;
1842 queue_Append(&rx_idleServerQueue, sq);
1843 #ifndef AFS_AIX41_ENV
1844 rx_waitForPacket = sq;
1846 rx_waitingForPacket = sq;
1847 #endif /* AFS_AIX41_ENV */
1849 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1851 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1852 MUTEX_EXIT(&rx_serverPool_lock);
1853 return (struct rx_call *)0;
1856 } while (!(call = sq->newcall)
1857 && !(socketp && *socketp != OSI_NULLSOCKET));
1858 MUTEX_EXIT(&rx_serverPool_lock);
1860 MUTEX_ENTER(&call->lock);
1866 MUTEX_ENTER(&freeSQEList_lock);
1867 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1868 rx_FreeSQEList = sq;
1869 MUTEX_EXIT(&freeSQEList_lock);
1872 clock_GetTime(&call->startTime);
1873 call->state = RX_STATE_ACTIVE;
1874 call->mode = RX_MODE_RECEIVING;
1875 #ifdef RX_KERNEL_TRACE
1876 if (ICL_SETACTIVE(afs_iclSetp)) {
1877 int glockOwner = ISAFS_GLOCK();
1880 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1881 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1888 rxi_calltrace(RX_CALL_START, call);
1889 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
1890 call->conn->service->servicePort, call->conn->service->serviceId,
1893 MUTEX_EXIT(&call->lock);
1894 MUTEX_ENTER(&rx_refcnt_mutex);
1895 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1896 MUTEX_EXIT(&rx_refcnt_mutex);
1898 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
1903 #else /* RX_ENABLE_LOCKS */
1905 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1907 struct rx_serverQueueEntry *sq;
1908 struct rx_call *call = (struct rx_call *)0, *choice2;
1909 struct rx_service *service = NULL;
1913 MUTEX_ENTER(&freeSQEList_lock);
1915 if ((sq = rx_FreeSQEList)) {
1916 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1917 MUTEX_EXIT(&freeSQEList_lock);
1918 } else { /* otherwise allocate a new one and return that */
1919 MUTEX_EXIT(&freeSQEList_lock);
1920 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1921 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1922 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1924 MUTEX_ENTER(&sq->lock);
1926 if (cur_service != NULL) {
1927 cur_service->nRequestsRunning--;
1928 MUTEX_ENTER(&rx_quota_mutex);
1929 if (cur_service->nRequestsRunning < cur_service->minProcs)
1932 MUTEX_EXIT(&rx_quota_mutex);
1934 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1935 struct rx_call *tcall, *ncall;
1936 /* Scan for eligible incoming calls. A call is not eligible
1937 * if the maximum number of calls for its service type are
1938 * already executing */
1939 /* One thread will process calls FCFS (to prevent starvation),
1940 * while the other threads may run ahead looking for calls which
1941 * have all their input data available immediately. This helps
1942 * keep threads from blocking, waiting for data from the client. */
1943 choice2 = (struct rx_call *)0;
1944 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1945 service = tcall->conn->service;
1946 if (QuotaOK(service)) {
1947 MUTEX_ENTER(&rx_pthread_mutex);
1948 if (tno == rxi_fcfs_thread_num
1949 || !tcall->queue_item_header.next) {
1950 MUTEX_EXIT(&rx_pthread_mutex);
1951 /* If we're the fcfs thread, then we'll just use
1952 * this call. If we haven't been able to find an optimal
1953 * choice, and we're at the end of the list, then use a
1954 * 2d choice if one has been identified. Otherwise... */
1955 call = (choice2 ? choice2 : tcall);
1956 service = call->conn->service;
1958 MUTEX_EXIT(&rx_pthread_mutex);
1959 if (!queue_IsEmpty(&tcall->rq)) {
1960 struct rx_packet *rp;
1961 rp = queue_First(&tcall->rq, rx_packet);
1962 if (rp->header.seq == 1
1964 || (rp->header.flags & RX_LAST_PACKET))) {
1966 } else if (rxi_2dchoice && !choice2
1967 && !(tcall->flags & RX_CALL_CLEARED)
1968 && (tcall->rprev > rxi_HardAckRate)) {
1982 /* we can't schedule a call if there's no data!!! */
1983 /* send an ack if there's no data, if we're missing the
1984 * first packet, or we're missing something between first
1985 * and last -- there's a "hole" in the incoming data. */
1986 if (queue_IsEmpty(&call->rq)
1987 || queue_First(&call->rq, rx_packet)->header.seq != 1
1988 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1989 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1991 call->flags &= (~RX_CALL_WAIT_PROC);
1992 service->nRequestsRunning++;
1993 /* just started call in minProcs pool, need fewer to maintain
1995 MUTEX_ENTER(&rx_quota_mutex);
1996 if (service->nRequestsRunning <= service->minProcs)
1999 MUTEX_EXIT(&rx_quota_mutex);
2000 rx_atomic_dec(&rx_nWaiting);
2001 /* MUTEX_EXIT(&call->lock); */
2003 /* If there are no eligible incoming calls, add this process
2004 * to the idle server queue, to wait for one */
2007 *socketp = OSI_NULLSOCKET;
2009 sq->socketp = socketp;
2010 queue_Append(&rx_idleServerQueue, sq);
2014 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2016 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2017 return (struct rx_call *)0;
2020 } while (!(call = sq->newcall)
2021 && !(socketp && *socketp != OSI_NULLSOCKET));
2023 MUTEX_EXIT(&sq->lock);
2025 MUTEX_ENTER(&freeSQEList_lock);
2026 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2027 rx_FreeSQEList = sq;
2028 MUTEX_EXIT(&freeSQEList_lock);
2031 clock_GetTime(&call->startTime);
2032 call->state = RX_STATE_ACTIVE;
2033 call->mode = RX_MODE_RECEIVING;
2034 #ifdef RX_KERNEL_TRACE
2035 if (ICL_SETACTIVE(afs_iclSetp)) {
2036 int glockOwner = ISAFS_GLOCK();
2039 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2040 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2047 rxi_calltrace(RX_CALL_START, call);
2048 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2049 call->conn->service->servicePort, call->conn->service->serviceId,
2052 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2059 #endif /* RX_ENABLE_LOCKS */
2063 /* Establish a procedure to be called when a packet arrives for a
2064 * call. This routine will be called at most once after each call,
2065 * and will also be called if there is an error condition on the or
2066 * the call is complete. Used by multi rx to build a selection
2067 * function which determines which of several calls is likely to be a
2068 * good one to read from.
2069 * NOTE: the way this is currently implemented it is probably only a
2070 * good idea to (1) use it immediately after a newcall (clients only)
2071 * and (2) only use it once. Other uses currently void your warranty
2074 rx_SetArrivalProc(struct rx_call *call,
2075 void (*proc) (struct rx_call * call,
2078 void * handle, int arg)
2080 call->arrivalProc = proc;
2081 call->arrivalProcHandle = handle;
2082 call->arrivalProcArg = arg;
2085 /* Call is finished (possibly prematurely). Return rc to the peer, if
2086 * appropriate, and return the final error code from the conversation
2090 rx_EndCall(struct rx_call *call, afs_int32 rc)
2092 struct rx_connection *conn = call->conn;
2096 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2097 call, rc, call->error, call->abortCode));
2100 MUTEX_ENTER(&call->lock);
2102 if (rc == 0 && call->error == 0) {
2103 call->abortCode = 0;
2104 call->abortCount = 0;
2107 call->arrivalProc = (void (*)())0;
2108 if (rc && call->error == 0) {
2109 rxi_CallError(call, rc);
2110 call->mode = RX_MODE_ERROR;
2111 /* Send an abort message to the peer if this error code has
2112 * only just been set. If it was set previously, assume the
2113 * peer has already been sent the error code or will request it
2115 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2117 if (conn->type == RX_SERVER_CONNECTION) {
2118 /* Make sure reply or at least dummy reply is sent */
2119 if (call->mode == RX_MODE_RECEIVING) {
2120 MUTEX_EXIT(&call->lock);
2121 rxi_WriteProc(call, 0, 0);
2122 MUTEX_ENTER(&call->lock);
2124 if (call->mode == RX_MODE_SENDING) {
2125 MUTEX_EXIT(&call->lock);
2126 rxi_FlushWrite(call);
2127 MUTEX_ENTER(&call->lock);
2129 rxi_calltrace(RX_CALL_END, call);
2130 /* Call goes to hold state until reply packets are acknowledged */
2131 if (call->tfirst + call->nSoftAcked < call->tnext) {
2132 call->state = RX_STATE_HOLD;
2134 call->state = RX_STATE_DALLY;
2135 rxi_ClearTransmitQueue(call, 0);
2136 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
2137 rxevent_Cancel(call->keepAliveEvent, call,
2138 RX_CALL_REFCOUNT_ALIVE);
2140 } else { /* Client connection */
2142 /* Make sure server receives input packets, in the case where
2143 * no reply arguments are expected */
2144 if ((call->mode == RX_MODE_SENDING)
2145 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2146 MUTEX_EXIT(&call->lock);
2147 (void)rxi_ReadProc(call, &dummy, 1);
2148 MUTEX_ENTER(&call->lock);
2151 /* If we had an outstanding delayed ack, be nice to the server
2152 * and force-send it now.
2154 if (call->delayedAckEvent) {
2155 rxevent_Cancel(call->delayedAckEvent, call,
2156 RX_CALL_REFCOUNT_DELAY);
2157 call->delayedAckEvent = NULL;
2158 rxi_SendDelayedAck(NULL, call, NULL);
2161 /* We need to release the call lock since it's lower than the
2162 * conn_call_lock and we don't want to hold the conn_call_lock
2163 * over the rx_ReadProc call. The conn_call_lock needs to be held
2164 * here for the case where rx_NewCall is perusing the calls on
2165 * the connection structure. We don't want to signal until
2166 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2167 * have checked this call, found it active and by the time it
2168 * goes to sleep, will have missed the signal.
2170 MUTEX_EXIT(&call->lock);
2171 MUTEX_ENTER(&conn->conn_call_lock);
2172 MUTEX_ENTER(&call->lock);
2173 MUTEX_ENTER(&conn->conn_data_lock);
2174 conn->flags |= RX_CONN_BUSY;
2175 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2176 MUTEX_EXIT(&conn->conn_data_lock);
2177 #ifdef RX_ENABLE_LOCKS
2178 CV_BROADCAST(&conn->conn_call_cv);
2183 #ifdef RX_ENABLE_LOCKS
2185 MUTEX_EXIT(&conn->conn_data_lock);
2187 #endif /* RX_ENABLE_LOCKS */
2188 call->state = RX_STATE_DALLY;
2190 error = call->error;
2192 /* currentPacket, nLeft, and NFree must be zeroed here, because
2193 * ResetCall cannot: ResetCall may be called at splnet(), in the
2194 * kernel version, and may interrupt the macros rx_Read or
2195 * rx_Write, which run at normal priority for efficiency. */
2196 if (call->currentPacket) {
2197 #ifdef RX_TRACK_PACKETS
2198 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2200 rxi_FreePacket(call->currentPacket);
2201 call->currentPacket = (struct rx_packet *)0;
2204 call->nLeft = call->nFree = call->curlen = 0;
2206 /* Free any packets from the last call to ReadvProc/WritevProc */
2207 #ifdef RXDEBUG_PACKET
2209 #endif /* RXDEBUG_PACKET */
2210 rxi_FreePackets(0, &call->iovq);
2211 MUTEX_EXIT(&call->lock);
2213 MUTEX_ENTER(&rx_refcnt_mutex);
2214 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2215 MUTEX_EXIT(&rx_refcnt_mutex);
2216 if (conn->type == RX_CLIENT_CONNECTION) {
2217 MUTEX_ENTER(&conn->conn_data_lock);
2218 conn->flags &= ~RX_CONN_BUSY;
2219 MUTEX_EXIT(&conn->conn_data_lock);
2220 MUTEX_EXIT(&conn->conn_call_lock);
2224 * Map errors to the local host's errno.h format.
2226 error = ntoh_syserr_conv(error);
2230 #if !defined(KERNEL)
2232 /* Call this routine when shutting down a server or client (especially
2233 * clients). This will allow Rx to gracefully garbage collect server
2234 * connections, and reduce the number of retries that a server might
2235 * make to a dead client.
2236 * This is not quite right, since some calls may still be ongoing and
2237 * we can't lock them to destroy them. */
2241 struct rx_connection **conn_ptr, **conn_end;
2245 if (rxinit_status == 1) {
2247 return; /* Already shutdown. */
2249 rxi_DeleteCachedConnections();
2250 if (rx_connHashTable) {
2251 MUTEX_ENTER(&rx_connHashTable_lock);
2252 for (conn_ptr = &rx_connHashTable[0], conn_end =
2253 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2255 struct rx_connection *conn, *next;
2256 for (conn = *conn_ptr; conn; conn = next) {
2258 if (conn->type == RX_CLIENT_CONNECTION) {
2259 MUTEX_ENTER(&rx_refcnt_mutex);
2261 MUTEX_EXIT(&rx_refcnt_mutex);
2262 #ifdef RX_ENABLE_LOCKS
2263 rxi_DestroyConnectionNoLock(conn);
2264 #else /* RX_ENABLE_LOCKS */
2265 rxi_DestroyConnection(conn);
2266 #endif /* RX_ENABLE_LOCKS */
2270 #ifdef RX_ENABLE_LOCKS
2271 while (rx_connCleanup_list) {
2272 struct rx_connection *conn;
2273 conn = rx_connCleanup_list;
2274 rx_connCleanup_list = rx_connCleanup_list->next;
2275 MUTEX_EXIT(&rx_connHashTable_lock);
2276 rxi_CleanupConnection(conn);
2277 MUTEX_ENTER(&rx_connHashTable_lock);
2279 MUTEX_EXIT(&rx_connHashTable_lock);
2280 #endif /* RX_ENABLE_LOCKS */
2285 afs_winsockCleanup();
2293 /* if we wakeup packet waiter too often, can get in loop with two
2294 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2296 rxi_PacketsUnWait(void)
2298 if (!rx_waitingForPackets) {
2302 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2303 return; /* still over quota */
2306 rx_waitingForPackets = 0;
2307 #ifdef RX_ENABLE_LOCKS
2308 CV_BROADCAST(&rx_waitingForPackets_cv);
2310 osi_rxWakeup(&rx_waitingForPackets);
2316 /* ------------------Internal interfaces------------------------- */
2318 /* Return this process's service structure for the
2319 * specified socket and service */
2321 rxi_FindService(osi_socket socket, u_short serviceId)
2323 struct rx_service **sp;
2324 for (sp = &rx_services[0]; *sp; sp++) {
2325 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2331 #ifdef RXDEBUG_PACKET
2332 #ifdef KDUMP_RX_LOCK
2333 static struct rx_call_rx_lock *rx_allCallsp = 0;
2335 static struct rx_call *rx_allCallsp = 0;
2337 #endif /* RXDEBUG_PACKET */
2339 /* Allocate a call structure, for the indicated channel of the
2340 * supplied connection. The mode and state of the call must be set by
2341 * the caller. Returns the call with mutex locked. */
2343 rxi_NewCall(struct rx_connection *conn, int channel)
2345 struct rx_call *call;
2346 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2347 struct rx_call *cp; /* Call pointer temp */
2348 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2349 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2351 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2353 /* Grab an existing call structure, or allocate a new one.
2354 * Existing call structures are assumed to have been left reset by
2356 MUTEX_ENTER(&rx_freeCallQueue_lock);
2358 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2360 * EXCEPT that the TQ might not yet be cleared out.
2361 * Skip over those with in-use TQs.
2364 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2365 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2371 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2372 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2373 call = queue_First(&rx_freeCallQueue, rx_call);
2374 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2376 if (rx_stats_active)
2377 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2378 MUTEX_EXIT(&rx_freeCallQueue_lock);
2379 MUTEX_ENTER(&call->lock);
2380 CLEAR_CALL_QUEUE_LOCK(call);
2381 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2382 /* Now, if TQ wasn't cleared earlier, do it now. */
2383 rxi_WaitforTQBusy(call);
2384 if (call->flags & RX_CALL_TQ_CLEARME) {
2385 rxi_ClearTransmitQueue(call, 1);
2386 /*queue_Init(&call->tq);*/
2388 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2389 /* Bind the call to its connection structure */
2391 rxi_ResetCall(call, 1);
2394 call = rxi_Alloc(sizeof(struct rx_call));
2395 #ifdef RXDEBUG_PACKET
2396 call->allNextp = rx_allCallsp;
2397 rx_allCallsp = call;
2399 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2400 #else /* RXDEBUG_PACKET */
2401 rx_atomic_inc(&rx_stats.nCallStructs);
2402 #endif /* RXDEBUG_PACKET */
2404 MUTEX_EXIT(&rx_freeCallQueue_lock);
2405 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2406 MUTEX_ENTER(&call->lock);
2407 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2408 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2409 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2411 /* Initialize once-only items */
2412 queue_Init(&call->tq);
2413 queue_Init(&call->rq);
2414 queue_Init(&call->iovq);
2415 #ifdef RXDEBUG_PACKET
2416 call->rqc = call->tqc = call->iovqc = 0;
2417 #endif /* RXDEBUG_PACKET */
2418 /* Bind the call to its connection structure (prereq for reset) */
2420 rxi_ResetCall(call, 1);
2422 call->channel = channel;
2423 call->callNumber = &conn->callNumber[channel];
2424 call->rwind = conn->rwind[channel];
2425 call->twind = conn->twind[channel];
2426 /* Note that the next expected call number is retained (in
2427 * conn->callNumber[i]), even if we reallocate the call structure
2429 conn->call[channel] = call;
2430 /* if the channel's never been used (== 0), we should start at 1, otherwise
2431 * the call number is valid from the last time this channel was used */
2432 if (*call->callNumber == 0)
2433 *call->callNumber = 1;
2438 /* A call has been inactive long enough that so we can throw away
2439 * state, including the call structure, which is placed on the call
2442 * call->lock amd rx_refcnt_mutex are held upon entry.
2443 * haveCTLock is set when called from rxi_ReapConnections.
2446 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2448 int channel = call->channel;
2449 struct rx_connection *conn = call->conn;
2452 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2453 (*call->callNumber)++;
2455 * We are setting the state to RX_STATE_RESET to
2456 * ensure that no one else will attempt to use this
2457 * call once we drop the refcnt lock. We must drop
2458 * the refcnt lock before calling rxi_ResetCall
2459 * because it cannot be held across acquiring the
2460 * freepktQ lock. NewCall does the same.
2462 call->state = RX_STATE_RESET;
2463 MUTEX_EXIT(&rx_refcnt_mutex);
2464 rxi_ResetCall(call, 0);
2465 call->conn->call[channel] = (struct rx_call *)0;
2467 MUTEX_ENTER(&rx_freeCallQueue_lock);
2468 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2469 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2470 /* A call may be free even though its transmit queue is still in use.
2471 * Since we search the call list from head to tail, put busy calls at
2472 * the head of the list, and idle calls at the tail.
2474 if (call->flags & RX_CALL_TQ_BUSY)
2475 queue_Prepend(&rx_freeCallQueue, call);
2477 queue_Append(&rx_freeCallQueue, call);
2478 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2479 queue_Append(&rx_freeCallQueue, call);
2480 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2481 if (rx_stats_active)
2482 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2483 MUTEX_EXIT(&rx_freeCallQueue_lock);
2485 /* Destroy the connection if it was previously slated for
2486 * destruction, i.e. the Rx client code previously called
2487 * rx_DestroyConnection (client connections), or
2488 * rxi_ReapConnections called the same routine (server
2489 * connections). Only do this, however, if there are no
2490 * outstanding calls. Note that for fine grain locking, there appears
2491 * to be a deadlock in that rxi_FreeCall has a call locked and
2492 * DestroyConnectionNoLock locks each call in the conn. But note a
2493 * few lines up where we have removed this call from the conn.
2494 * If someone else destroys a connection, they either have no
2495 * call lock held or are going through this section of code.
2497 MUTEX_ENTER(&conn->conn_data_lock);
2498 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2499 MUTEX_ENTER(&rx_refcnt_mutex);
2501 MUTEX_EXIT(&rx_refcnt_mutex);
2502 MUTEX_EXIT(&conn->conn_data_lock);
2503 #ifdef RX_ENABLE_LOCKS
2505 rxi_DestroyConnectionNoLock(conn);
2507 rxi_DestroyConnection(conn);
2508 #else /* RX_ENABLE_LOCKS */
2509 rxi_DestroyConnection(conn);
2510 #endif /* RX_ENABLE_LOCKS */
2512 MUTEX_EXIT(&conn->conn_data_lock);
2514 MUTEX_ENTER(&rx_refcnt_mutex);
2517 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2518 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2521 rxi_Alloc(size_t size)
2525 if (rx_stats_active) {
2526 rx_atomic_add(&rxi_Allocsize, (int) size);
2527 rx_atomic_inc(&rxi_Alloccnt);
2531 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2532 afs_osi_Alloc_NoSleep(size);
2537 osi_Panic("rxi_Alloc error");
2543 rxi_Free(void *addr, size_t size)
2545 if (rx_stats_active) {
2546 rx_atomic_sub(&rxi_Allocsize, (int) size);
2547 rx_atomic_dec(&rxi_Alloccnt);
2549 osi_Free(addr, size);
2553 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2555 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2556 struct rx_peer *next = NULL;
2560 MUTEX_ENTER(&rx_peerHashTable_lock);
2562 peer_ptr = &rx_peerHashTable[0];
2563 peer_end = &rx_peerHashTable[rx_hashTableSize];
2566 for ( ; peer_ptr < peer_end; peer_ptr++) {
2569 for ( ; peer; peer = next) {
2571 if (host == peer->host)
2576 hashIndex = PEER_HASH(host, port);
2577 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2578 if ((peer->host == host) && (peer->port == port))
2583 MUTEX_ENTER(&rx_peerHashTable_lock);
2588 MUTEX_EXIT(&rx_peerHashTable_lock);
2590 MUTEX_ENTER(&peer->peer_lock);
2591 /* We don't handle dropping below min, so don't */
2592 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2593 peer->ifMTU=MIN(mtu, peer->ifMTU);
2594 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2595 /* if we tweaked this down, need to tune our peer MTU too */
2596 peer->MTU = MIN(peer->MTU, peer->natMTU);
2597 /* if we discovered a sub-1500 mtu, degrade */
2598 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2599 peer->maxDgramPackets = 1;
2600 /* We no longer have valid peer packet information */
2601 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2602 peer->maxPacketSize = 0;
2603 MUTEX_EXIT(&peer->peer_lock);
2605 MUTEX_ENTER(&rx_peerHashTable_lock);
2607 if (host && !port) {
2609 /* pick up where we left off */
2613 MUTEX_EXIT(&rx_peerHashTable_lock);
2616 /* Find the peer process represented by the supplied (host,port)
2617 * combination. If there is no appropriate active peer structure, a
2618 * new one will be allocated and initialized
2619 * The origPeer, if set, is a pointer to a peer structure on which the
2620 * refcount will be be decremented. This is used to replace the peer
2621 * structure hanging off a connection structure */
2623 rxi_FindPeer(afs_uint32 host, u_short port,
2624 struct rx_peer *origPeer, int create)
2628 hashIndex = PEER_HASH(host, port);
2629 MUTEX_ENTER(&rx_peerHashTable_lock);
2630 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2631 if ((pp->host == host) && (pp->port == port))
2636 pp = rxi_AllocPeer(); /* This bzero's *pp */
2637 pp->host = host; /* set here or in InitPeerParams is zero */
2639 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2640 queue_Init(&pp->congestionQueue);
2641 queue_Init(&pp->rpcStats);
2642 pp->next = rx_peerHashTable[hashIndex];
2643 rx_peerHashTable[hashIndex] = pp;
2644 rxi_InitPeerParams(pp);
2645 if (rx_stats_active)
2646 rx_atomic_inc(&rx_stats.nPeerStructs);
2653 origPeer->refCount--;
2654 MUTEX_EXIT(&rx_peerHashTable_lock);
2659 /* Find the connection at (host, port) started at epoch, and with the
2660 * given connection id. Creates the server connection if necessary.
2661 * The type specifies whether a client connection or a server
2662 * connection is desired. In both cases, (host, port) specify the
2663 * peer's (host, pair) pair. Client connections are not made
2664 * automatically by this routine. The parameter socket gives the
2665 * socket descriptor on which the packet was received. This is used,
2666 * in the case of server connections, to check that *new* connections
2667 * come via a valid (port, serviceId). Finally, the securityIndex
2668 * parameter must match the existing index for the connection. If a
2669 * server connection is created, it will be created using the supplied
2670 * index, if the index is valid for this service */
2671 struct rx_connection *
2672 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2673 u_short port, u_short serviceId, afs_uint32 cid,
2674 afs_uint32 epoch, int type, u_int securityIndex)
2676 int hashindex, flag, i;
2677 struct rx_connection *conn;
2678 hashindex = CONN_HASH(host, port, cid, epoch, type);
2679 MUTEX_ENTER(&rx_connHashTable_lock);
2680 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2681 rx_connHashTable[hashindex],
2684 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2685 && (epoch == conn->epoch)) {
2686 struct rx_peer *pp = conn->peer;
2687 if (securityIndex != conn->securityIndex) {
2688 /* this isn't supposed to happen, but someone could forge a packet
2689 * like this, and there seems to be some CM bug that makes this
2690 * happen from time to time -- in which case, the fileserver
2692 MUTEX_EXIT(&rx_connHashTable_lock);
2693 return (struct rx_connection *)0;
2695 if (pp->host == host && pp->port == port)
2697 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2699 /* So what happens when it's a callback connection? */
2700 if ( /*type == RX_CLIENT_CONNECTION && */
2701 (conn->epoch & 0x80000000))
2705 /* the connection rxLastConn that was used the last time is not the
2706 ** one we are looking for now. Hence, start searching in the hash */
2708 conn = rx_connHashTable[hashindex];
2713 struct rx_service *service;
2714 if (type == RX_CLIENT_CONNECTION) {
2715 MUTEX_EXIT(&rx_connHashTable_lock);
2716 return (struct rx_connection *)0;
2718 service = rxi_FindService(socket, serviceId);
2719 if (!service || (securityIndex >= service->nSecurityObjects)
2720 || (service->securityObjects[securityIndex] == 0)) {
2721 MUTEX_EXIT(&rx_connHashTable_lock);
2722 return (struct rx_connection *)0;
2724 conn = rxi_AllocConnection(); /* This bzero's the connection */
2725 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2726 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2727 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2728 conn->next = rx_connHashTable[hashindex];
2729 rx_connHashTable[hashindex] = conn;
2730 conn->peer = rxi_FindPeer(host, port, 0, 1);
2731 conn->type = RX_SERVER_CONNECTION;
2732 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2733 conn->epoch = epoch;
2734 conn->cid = cid & RX_CIDMASK;
2735 /* conn->serial = conn->lastSerial = 0; */
2736 /* conn->timeout = 0; */
2737 conn->ackRate = RX_FAST_ACK_RATE;
2738 conn->service = service;
2739 conn->serviceId = serviceId;
2740 conn->securityIndex = securityIndex;
2741 conn->securityObject = service->securityObjects[securityIndex];
2742 conn->nSpecific = 0;
2743 conn->specific = NULL;
2744 rx_SetConnDeadTime(conn, service->connDeadTime);
2745 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2746 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2747 for (i = 0; i < RX_MAXCALLS; i++) {
2748 conn->twind[i] = rx_initSendWindow;
2749 conn->rwind[i] = rx_initReceiveWindow;
2751 /* Notify security object of the new connection */
2752 RXS_NewConnection(conn->securityObject, conn);
2753 /* XXXX Connection timeout? */
2754 if (service->newConnProc)
2755 (*service->newConnProc) (conn);
2756 if (rx_stats_active)
2757 rx_atomic_inc(&rx_stats.nServerConns);
2760 MUTEX_ENTER(&rx_refcnt_mutex);
2762 MUTEX_EXIT(&rx_refcnt_mutex);
2764 rxLastConn = conn; /* store this connection as the last conn used */
2765 MUTEX_EXIT(&rx_connHashTable_lock);
2769 /* There are two packet tracing routines available for testing and monitoring
2770 * Rx. One is called just after every packet is received and the other is
2771 * called just before every packet is sent. Received packets, have had their
2772 * headers decoded, and packets to be sent have not yet had their headers
2773 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2774 * containing the network address. Both can be modified. The return value, if
2775 * non-zero, indicates that the packet should be dropped. */
2777 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2778 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2780 /* A packet has been received off the interface. Np is the packet, socket is
2781 * the socket number it was received from (useful in determining which service
2782 * this packet corresponds to), and (host, port) reflect the host,port of the
2783 * sender. This call returns the packet to the caller if it is finished with
2784 * it, rather than de-allocating it, just as a small performance hack */
2787 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
2788 afs_uint32 host, u_short port, int *tnop,
2789 struct rx_call **newcallp)
2791 struct rx_call *call;
2792 struct rx_connection *conn;
2794 afs_uint32 currentCallNumber;
2800 struct rx_packet *tnp;
2803 /* We don't print out the packet until now because (1) the time may not be
2804 * accurate enough until now in the lwp implementation (rx_Listener only gets
2805 * the time after the packet is read) and (2) from a protocol point of view,
2806 * this is the first time the packet has been seen */
2807 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2808 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2809 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
2810 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2811 np->header.epoch, np->header.cid, np->header.callNumber,
2812 np->header.seq, np->header.flags, np));
2815 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2816 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2819 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2820 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2823 /* If an input tracer function is defined, call it with the packet and
2824 * network address. Note this function may modify its arguments. */
2825 if (rx_justReceived) {
2826 struct sockaddr_in addr;
2828 addr.sin_family = AF_INET;
2829 addr.sin_port = port;
2830 addr.sin_addr.s_addr = host;
2831 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2832 addr.sin_len = sizeof(addr);
2833 #endif /* AFS_OSF_ENV */
2834 drop = (*rx_justReceived) (np, &addr);
2835 /* drop packet if return value is non-zero */
2838 port = addr.sin_port; /* in case fcn changed addr */
2839 host = addr.sin_addr.s_addr;
2843 /* If packet was not sent by the client, then *we* must be the client */
2844 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2845 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2847 /* Find the connection (or fabricate one, if we're the server & if
2848 * necessary) associated with this packet */
2850 rxi_FindConnection(socket, host, port, np->header.serviceId,
2851 np->header.cid, np->header.epoch, type,
2852 np->header.securityIndex);
2855 /* If no connection found or fabricated, just ignore the packet.
2856 * (An argument could be made for sending an abort packet for
2861 MUTEX_ENTER(&conn->conn_data_lock);
2862 if (conn->maxSerial < np->header.serial)
2863 conn->maxSerial = np->header.serial;
2864 MUTEX_EXIT(&conn->conn_data_lock);
2866 /* If the connection is in an error state, send an abort packet and ignore
2867 * the incoming packet */
2869 /* Don't respond to an abort packet--we don't want loops! */
2870 MUTEX_ENTER(&conn->conn_data_lock);
2871 if (np->header.type != RX_PACKET_TYPE_ABORT)
2872 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2873 MUTEX_ENTER(&rx_refcnt_mutex);
2875 MUTEX_EXIT(&rx_refcnt_mutex);
2876 MUTEX_EXIT(&conn->conn_data_lock);
2880 /* Check for connection-only requests (i.e. not call specific). */
2881 if (np->header.callNumber == 0) {
2882 switch (np->header.type) {
2883 case RX_PACKET_TYPE_ABORT: {
2884 /* What if the supplied error is zero? */
2885 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2886 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
2887 rxi_ConnectionError(conn, errcode);
2888 MUTEX_ENTER(&rx_refcnt_mutex);
2890 MUTEX_EXIT(&rx_refcnt_mutex);
2893 case RX_PACKET_TYPE_CHALLENGE:
2894 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2895 MUTEX_ENTER(&rx_refcnt_mutex);
2897 MUTEX_EXIT(&rx_refcnt_mutex);
2899 case RX_PACKET_TYPE_RESPONSE:
2900 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2901 MUTEX_ENTER(&rx_refcnt_mutex);
2903 MUTEX_EXIT(&rx_refcnt_mutex);
2905 case RX_PACKET_TYPE_PARAMS:
2906 case RX_PACKET_TYPE_PARAMS + 1:
2907 case RX_PACKET_TYPE_PARAMS + 2:
2908 /* ignore these packet types for now */
2909 MUTEX_ENTER(&rx_refcnt_mutex);
2911 MUTEX_EXIT(&rx_refcnt_mutex);
2916 /* Should not reach here, unless the peer is broken: send an
2918 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2919 MUTEX_ENTER(&conn->conn_data_lock);
2920 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2921 MUTEX_ENTER(&rx_refcnt_mutex);
2923 MUTEX_EXIT(&rx_refcnt_mutex);
2924 MUTEX_EXIT(&conn->conn_data_lock);
2929 channel = np->header.cid & RX_CHANNELMASK;
2930 call = conn->call[channel];
2931 #ifdef RX_ENABLE_LOCKS
2933 MUTEX_ENTER(&call->lock);
2934 /* Test to see if call struct is still attached to conn. */
2935 if (call != conn->call[channel]) {
2937 MUTEX_EXIT(&call->lock);
2938 if (type == RX_SERVER_CONNECTION) {
2939 call = conn->call[channel];
2940 /* If we started with no call attached and there is one now,
2941 * another thread is also running this routine and has gotten
2942 * the connection channel. We should drop this packet in the tests
2943 * below. If there was a call on this connection and it's now
2944 * gone, then we'll be making a new call below.
2945 * If there was previously a call and it's now different then
2946 * the old call was freed and another thread running this routine
2947 * has created a call on this channel. One of these two threads
2948 * has a packet for the old call and the code below handles those
2952 MUTEX_ENTER(&call->lock);
2954 /* This packet can't be for this call. If the new call address is
2955 * 0 then no call is running on this channel. If there is a call
2956 * then, since this is a client connection we're getting data for
2957 * it must be for the previous call.
2959 if (rx_stats_active)
2960 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
2961 MUTEX_ENTER(&rx_refcnt_mutex);
2963 MUTEX_EXIT(&rx_refcnt_mutex);
2968 currentCallNumber = conn->callNumber[channel];
2970 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2971 if (np->header.callNumber < currentCallNumber) {
2972 if (rx_stats_active)
2973 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
2974 #ifdef RX_ENABLE_LOCKS
2976 MUTEX_EXIT(&call->lock);
2978 MUTEX_ENTER(&rx_refcnt_mutex);
2980 MUTEX_EXIT(&rx_refcnt_mutex);
2984 MUTEX_ENTER(&conn->conn_call_lock);
2985 call = rxi_NewCall(conn, channel);
2986 MUTEX_EXIT(&conn->conn_call_lock);
2987 *call->callNumber = np->header.callNumber;
2989 if (np->header.callNumber == 0)
2990 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%.06d len %d\n",
2991 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
2992 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
2993 np->header.flags, np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2995 call->state = RX_STATE_PRECALL;
2996 clock_GetTime(&call->queueTime);
2997 hzero(call->bytesSent);
2998 hzero(call->bytesRcvd);
3000 * If the number of queued calls exceeds the overload
3001 * threshold then abort this call.
3003 if ((rx_BusyThreshold > 0) &&
3004 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3005 struct rx_packet *tp;
3007 rxi_CallError(call, rx_BusyError);
3008 tp = rxi_SendCallAbort(call, np, 1, 0);
3009 MUTEX_EXIT(&call->lock);
3010 MUTEX_ENTER(&rx_refcnt_mutex);
3012 MUTEX_EXIT(&rx_refcnt_mutex);
3013 if (rx_stats_active)
3014 rx_atomic_inc(&rx_stats.nBusies);
3017 rxi_KeepAliveOn(call);
3018 } else if (np->header.callNumber != currentCallNumber) {
3019 /* Wait until the transmit queue is idle before deciding
3020 * whether to reset the current call. Chances are that the
3021 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3024 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3025 if (call->state == RX_STATE_ACTIVE) {
3026 rxi_WaitforTQBusy(call);
3028 * If we entered error state while waiting,
3029 * must call rxi_CallError to permit rxi_ResetCall
3030 * to processed when the tqWaiter count hits zero.
3033 rxi_CallError(call, call->error);
3034 MUTEX_EXIT(&call->lock);
3035 MUTEX_ENTER(&rx_refcnt_mutex);
3037 MUTEX_EXIT(&rx_refcnt_mutex);
3041 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3042 /* If the new call cannot be taken right now send a busy and set
3043 * the error condition in this call, so that it terminates as
3044 * quickly as possible */
3045 if (call->state == RX_STATE_ACTIVE) {
3046 struct rx_packet *tp;
3048 rxi_CallError(call, RX_CALL_DEAD);
3049 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3051 MUTEX_EXIT(&call->lock);
3052 MUTEX_ENTER(&rx_refcnt_mutex);
3054 MUTEX_EXIT(&rx_refcnt_mutex);
3057 rxi_ResetCall(call, 0);
3058 *call->callNumber = np->header.callNumber;
3060 if (np->header.callNumber == 0)
3061 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%06d len %d\n",
3062 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3063 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3064 np->header.flags, np, np->retryTime.sec, np->retryTime.usec, np->length));
3066 call->state = RX_STATE_PRECALL;
3067 clock_GetTime(&call->queueTime);
3068 hzero(call->bytesSent);
3069 hzero(call->bytesRcvd);
3071 * If the number of queued calls exceeds the overload
3072 * threshold then abort this call.
3074 if ((rx_BusyThreshold > 0) &&
3075 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3076 struct rx_packet *tp;
3078 rxi_CallError(call, rx_BusyError);
3079 tp = rxi_SendCallAbort(call, np, 1, 0);
3080 MUTEX_EXIT(&call->lock);
3081 MUTEX_ENTER(&rx_refcnt_mutex);
3083 MUTEX_EXIT(&rx_refcnt_mutex);
3084 if (rx_stats_active)
3085 rx_atomic_inc(&rx_stats.nBusies);
3088 rxi_KeepAliveOn(call);
3090 /* Continuing call; do nothing here. */
3092 } else { /* we're the client */
3093 /* Ignore all incoming acknowledgements for calls in DALLY state */
3094 if (call && (call->state == RX_STATE_DALLY)
3095 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3096 if (rx_stats_active)
3097 rx_atomic_inc(&rx_stats.ignorePacketDally);
3098 #ifdef RX_ENABLE_LOCKS
3100 MUTEX_EXIT(&call->lock);
3103 MUTEX_ENTER(&rx_refcnt_mutex);
3105 MUTEX_EXIT(&rx_refcnt_mutex);
3109 /* Ignore anything that's not relevant to the current call. If there
3110 * isn't a current call, then no packet is relevant. */
3111 if (!call || (np->header.callNumber != currentCallNumber)) {
3112 if (rx_stats_active)
3113 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3114 #ifdef RX_ENABLE_LOCKS
3116 MUTEX_EXIT(&call->lock);
3119 MUTEX_ENTER(&rx_refcnt_mutex);
3121 MUTEX_EXIT(&rx_refcnt_mutex);
3124 /* If the service security object index stamped in the packet does not
3125 * match the connection's security index, ignore the packet */
3126 if (np->header.securityIndex != conn->securityIndex) {
3127 #ifdef RX_ENABLE_LOCKS
3128 MUTEX_EXIT(&call->lock);
3130 MUTEX_ENTER(&rx_refcnt_mutex);
3132 MUTEX_EXIT(&rx_refcnt_mutex);
3136 /* If we're receiving the response, then all transmit packets are
3137 * implicitly acknowledged. Get rid of them. */
3138 if (np->header.type == RX_PACKET_TYPE_DATA) {
3139 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3140 /* XXX Hack. Because we must release the global rx lock when
3141 * sending packets (osi_NetSend) we drop all acks while we're
3142 * traversing the tq in rxi_Start sending packets out because
3143 * packets may move to the freePacketQueue as result of being here!
3144 * So we drop these packets until we're safely out of the
3145 * traversing. Really ugly!
3146 * For fine grain RX locking, we set the acked field in the
3147 * packets and let rxi_Start remove them from the transmit queue.
3149 if (call->flags & RX_CALL_TQ_BUSY) {
3150 #ifdef RX_ENABLE_LOCKS
3151 rxi_SetAcksInTransmitQueue(call);
3153 MUTEX_ENTER(&rx_refcnt_mutex);
3155 MUTEX_EXIT(&rx_refcnt_mutex);
3156 return np; /* xmitting; drop packet */
3159 rxi_ClearTransmitQueue(call, 0);
3161 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3162 rxi_ClearTransmitQueue(call, 0);
3163 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3165 if (np->header.type == RX_PACKET_TYPE_ACK) {
3166 /* now check to see if this is an ack packet acknowledging that the
3167 * server actually *lost* some hard-acked data. If this happens we
3168 * ignore this packet, as it may indicate that the server restarted in
3169 * the middle of a call. It is also possible that this is an old ack
3170 * packet. We don't abort the connection in this case, because this
3171 * *might* just be an old ack packet. The right way to detect a server
3172 * restart in the midst of a call is to notice that the server epoch
3174 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3175 * XXX unacknowledged. I think that this is off-by-one, but
3176 * XXX I don't dare change it just yet, since it will
3177 * XXX interact badly with the server-restart detection
3178 * XXX code in receiveackpacket. */
3179 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3180 if (rx_stats_active)
3181 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3182 MUTEX_EXIT(&call->lock);
3183 MUTEX_ENTER(&rx_refcnt_mutex);
3185 MUTEX_EXIT(&rx_refcnt_mutex);
3189 } /* else not a data packet */
3192 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3193 /* Set remote user defined status from packet */
3194 call->remoteStatus = np->header.userStatus;
3196 /* Note the gap between the expected next packet and the actual
3197 * packet that arrived, when the new packet has a smaller serial number
3198 * than expected. Rioses frequently reorder packets all by themselves,
3199 * so this will be quite important with very large window sizes.
3200 * Skew is checked against 0 here to avoid any dependence on the type of
3201 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3203 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3204 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3205 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3207 MUTEX_ENTER(&conn->conn_data_lock);
3208 skew = conn->lastSerial - np->header.serial;
3209 conn->lastSerial = np->header.serial;
3210 MUTEX_EXIT(&conn->conn_data_lock);
3212 struct rx_peer *peer;
3214 if (skew > peer->inPacketSkew) {
3215 dpf(("*** In skew changed from %d to %d\n",
3216 peer->inPacketSkew, skew));
3217 peer->inPacketSkew = skew;
3221 /* Now do packet type-specific processing */
3222 switch (np->header.type) {
3223 case RX_PACKET_TYPE_DATA:
3224 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3227 case RX_PACKET_TYPE_ACK:
3228 /* Respond immediately to ack packets requesting acknowledgement
3230 if (np->header.flags & RX_REQUEST_ACK) {
3232 (void)rxi_SendCallAbort(call, 0, 1, 0);
3234 (void)rxi_SendAck(call, 0, np->header.serial,
3235 RX_ACK_PING_RESPONSE, 1);
3237 np = rxi_ReceiveAckPacket(call, np, 1);
3239 case RX_PACKET_TYPE_ABORT: {
3240 /* An abort packet: reset the call, passing the error up to the user. */
3241 /* What if error is zero? */
3242 /* What if the error is -1? the application will treat it as a timeout. */
3243 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3244 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3245 rxi_CallError(call, errdata);
3246 MUTEX_EXIT(&call->lock);
3247 MUTEX_ENTER(&rx_refcnt_mutex);
3249 MUTEX_EXIT(&rx_refcnt_mutex);
3250 return np; /* xmitting; drop packet */
3252 case RX_PACKET_TYPE_BUSY:
3255 case RX_PACKET_TYPE_ACKALL:
3256 /* All packets acknowledged, so we can drop all packets previously
3257 * readied for sending */
3258 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3259 /* XXX Hack. We because we can't release the global rx lock when
3260 * sending packets (osi_NetSend) we drop all ack pkts while we're
3261 * traversing the tq in rxi_Start sending packets out because
3262 * packets may move to the freePacketQueue as result of being
3263 * here! So we drop these packets until we're safely out of the
3264 * traversing. Really ugly!
3265 * For fine grain RX locking, we set the acked field in the packets
3266 * and let rxi_Start remove the packets from the transmit queue.
3268 if (call->flags & RX_CALL_TQ_BUSY) {
3269 #ifdef RX_ENABLE_LOCKS
3270 rxi_SetAcksInTransmitQueue(call);
3272 #else /* RX_ENABLE_LOCKS */
3273 MUTEX_EXIT(&call->lock);
3274 MUTEX_ENTER(&rx_refcnt_mutex);
3276 MUTEX_EXIT(&rx_refcnt_mutex);
3277 return np; /* xmitting; drop packet */
3278 #endif /* RX_ENABLE_LOCKS */
3280 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3281 rxi_ClearTransmitQueue(call, 0);
3282 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3285 /* Should not reach here, unless the peer is broken: send an abort
3287 rxi_CallError(call, RX_PROTOCOL_ERROR);
3288 np = rxi_SendCallAbort(call, np, 1, 0);
3291 /* Note when this last legitimate packet was received, for keep-alive
3292 * processing. Note, we delay getting the time until now in the hope that
3293 * the packet will be delivered to the user before any get time is required
3294 * (if not, then the time won't actually be re-evaluated here). */
3295 call->lastReceiveTime = clock_Sec();
3296 MUTEX_EXIT(&call->lock);
3297 MUTEX_ENTER(&rx_refcnt_mutex);
3299 MUTEX_EXIT(&rx_refcnt_mutex);
3303 /* return true if this is an "interesting" connection from the point of view
3304 of someone trying to debug the system */
3306 rxi_IsConnInteresting(struct rx_connection *aconn)
3309 struct rx_call *tcall;
3311 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3314 for (i = 0; i < RX_MAXCALLS; i++) {
3315 tcall = aconn->call[i];
3317 if ((tcall->state == RX_STATE_PRECALL)
3318 || (tcall->state == RX_STATE_ACTIVE))
3320 if ((tcall->mode == RX_MODE_SENDING)
3321 || (tcall->mode == RX_MODE_RECEIVING))
3329 /* if this is one of the last few packets AND it wouldn't be used by the
3330 receiving call to immediately satisfy a read request, then drop it on
3331 the floor, since accepting it might prevent a lock-holding thread from
3332 making progress in its reading. If a call has been cleared while in
3333 the precall state then ignore all subsequent packets until the call
3334 is assigned to a thread. */
3337 TooLow(struct rx_packet *ap, struct rx_call *acall)
3341 MUTEX_ENTER(&rx_quota_mutex);
3342 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3343 && (acall->state == RX_STATE_PRECALL))
3344 || ((rx_nFreePackets < rxi_dataQuota + 2)
3345 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3346 && (acall->flags & RX_CALL_READER_WAIT)))) {
3349 MUTEX_EXIT(&rx_quota_mutex);
3355 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3357 struct rx_connection *conn = arg1;
3358 struct rx_call *acall = arg2;
3359 struct rx_call *call = acall;
3360 struct clock when, now;
3363 MUTEX_ENTER(&conn->conn_data_lock);
3364 conn->checkReachEvent = NULL;
3365 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3367 MUTEX_ENTER(&rx_refcnt_mutex);
3369 MUTEX_EXIT(&rx_refcnt_mutex);
3371 MUTEX_EXIT(&conn->conn_data_lock);
3375 MUTEX_ENTER(&conn->conn_call_lock);
3376 MUTEX_ENTER(&conn->conn_data_lock);
3377 for (i = 0; i < RX_MAXCALLS; i++) {
3378 struct rx_call *tc = conn->call[i];
3379 if (tc && tc->state == RX_STATE_PRECALL) {
3385 /* Indicate that rxi_CheckReachEvent is no longer running by
3386 * clearing the flag. Must be atomic under conn_data_lock to
3387 * avoid a new call slipping by: rxi_CheckConnReach holds
3388 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3390 conn->flags &= ~RX_CONN_ATTACHWAIT;
3391 MUTEX_EXIT(&conn->conn_data_lock);
3392 MUTEX_EXIT(&conn->conn_call_lock);
3397 MUTEX_ENTER(&call->lock);
3398 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3400 MUTEX_EXIT(&call->lock);
3402 clock_GetTime(&now);
3404 when.sec += RX_CHECKREACH_TIMEOUT;
3405 MUTEX_ENTER(&conn->conn_data_lock);
3406 if (!conn->checkReachEvent) {
3407 MUTEX_ENTER(&rx_refcnt_mutex);
3409 MUTEX_EXIT(&rx_refcnt_mutex);
3410 conn->checkReachEvent =
3411 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3414 MUTEX_EXIT(&conn->conn_data_lock);
3420 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3422 struct rx_service *service = conn->service;
3423 struct rx_peer *peer = conn->peer;
3424 afs_uint32 now, lastReach;
3426 if (service->checkReach == 0)
3430 MUTEX_ENTER(&peer->peer_lock);
3431 lastReach = peer->lastReachTime;
3432 MUTEX_EXIT(&peer->peer_lock);
3433 if (now - lastReach < RX_CHECKREACH_TTL)
3436 MUTEX_ENTER(&conn->conn_data_lock);
3437 if (conn->flags & RX_CONN_ATTACHWAIT) {
3438 MUTEX_EXIT(&conn->conn_data_lock);
3441 conn->flags |= RX_CONN_ATTACHWAIT;
3442 MUTEX_EXIT(&conn->conn_data_lock);
3443 if (!conn->checkReachEvent)
3444 rxi_CheckReachEvent(NULL, conn, call);
3449 /* try to attach call, if authentication is complete */
3451 TryAttach(struct rx_call *acall, osi_socket socket,
3452 int *tnop, struct rx_call **newcallp,
3455 struct rx_connection *conn = acall->conn;
3457 if (conn->type == RX_SERVER_CONNECTION
3458 && acall->state == RX_STATE_PRECALL) {
3459 /* Don't attach until we have any req'd. authentication. */
3460 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3461 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3462 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3463 /* Note: this does not necessarily succeed; there
3464 * may not any proc available
3467 rxi_ChallengeOn(acall->conn);
3472 /* A data packet has been received off the interface. This packet is
3473 * appropriate to the call (the call is in the right state, etc.). This
3474 * routine can return a packet to the caller, for re-use */
3477 rxi_ReceiveDataPacket(struct rx_call *call,
3478 struct rx_packet *np, int istack,
3479 osi_socket socket, afs_uint32 host, u_short port,
3480 int *tnop, struct rx_call **newcallp)
3482 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3487 afs_uint32 serial=0, flags=0;
3489 struct rx_packet *tnp;
3490 struct clock when, now;
3491 if (rx_stats_active)
3492 rx_atomic_inc(&rx_stats.dataPacketsRead);
3495 /* If there are no packet buffers, drop this new packet, unless we can find
3496 * packet buffers from inactive calls */
3498 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3499 MUTEX_ENTER(&rx_freePktQ_lock);
3500 rxi_NeedMorePackets = TRUE;
3501 MUTEX_EXIT(&rx_freePktQ_lock);
3502 if (rx_stats_active)
3503 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3504 call->rprev = np->header.serial;
3505 rxi_calltrace(RX_TRACE_DROP, call);
3506 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3508 rxi_ClearReceiveQueue(call);
3509 clock_GetTime(&now);
3511 clock_Add(&when, &rx_softAckDelay);
3512 if (!call->delayedAckEvent
3513 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3514 rxevent_Cancel(call->delayedAckEvent, call,
3515 RX_CALL_REFCOUNT_DELAY);
3516 MUTEX_ENTER(&rx_refcnt_mutex);
3517 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3518 MUTEX_EXIT(&rx_refcnt_mutex);
3520 call->delayedAckEvent =
3521 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3523 /* we've damaged this call already, might as well do it in. */
3529 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3530 * packet is one of several packets transmitted as a single
3531 * datagram. Do not send any soft or hard acks until all packets
3532 * in a jumbogram have been processed. Send negative acks right away.
3534 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3535 /* tnp is non-null when there are more packets in the
3536 * current jumbo gram */
3543 seq = np->header.seq;
3544 serial = np->header.serial;
3545 flags = np->header.flags;
3547 /* If the call is in an error state, send an abort message */
3549 return rxi_SendCallAbort(call, np, istack, 0);
3551 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3552 * AFS 3.5 jumbogram. */
3553 if (flags & RX_JUMBO_PACKET) {
3554 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3559 if (np->header.spare != 0) {
3560 MUTEX_ENTER(&call->conn->conn_data_lock);
3561 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3562 MUTEX_EXIT(&call->conn->conn_data_lock);
3565 /* The usual case is that this is the expected next packet */
3566 if (seq == call->rnext) {
3568 /* Check to make sure it is not a duplicate of one already queued */
3569 if (queue_IsNotEmpty(&call->rq)
3570 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3571 if (rx_stats_active)
3572 rx_atomic_inc(&rx_stats.dupPacketsRead);
3573 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3574 rxevent_Cancel(call->delayedAckEvent, call,
3575 RX_CALL_REFCOUNT_DELAY);
3576 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3582 /* It's the next packet. Stick it on the receive queue
3583 * for this call. Set newPackets to make sure we wake
3584 * the reader once all packets have been processed */
3585 #ifdef RX_TRACK_PACKETS
3586 np->flags |= RX_PKTFLAG_RQ;
3588 queue_Prepend(&call->rq, np);
3589 #ifdef RXDEBUG_PACKET
3591 #endif /* RXDEBUG_PACKET */
3593 np = NULL; /* We can't use this anymore */
3596 /* If an ack is requested then set a flag to make sure we
3597 * send an acknowledgement for this packet */
3598 if (flags & RX_REQUEST_ACK) {
3599 ackNeeded = RX_ACK_REQUESTED;
3602 /* Keep track of whether we have received the last packet */
3603 if (flags & RX_LAST_PACKET) {
3604 call->flags |= RX_CALL_HAVE_LAST;
3608 /* Check whether we have all of the packets for this call */
3609 if (call->flags & RX_CALL_HAVE_LAST) {
3610 afs_uint32 tseq; /* temporary sequence number */
3611 struct rx_packet *tp; /* Temporary packet pointer */
3612 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3614 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3615 if (tseq != tp->header.seq)
3617 if (tp->header.flags & RX_LAST_PACKET) {
3618 call->flags |= RX_CALL_RECEIVE_DONE;
3625 /* Provide asynchronous notification for those who want it
3626 * (e.g. multi rx) */
3627 if (call->arrivalProc) {
3628 (*call->arrivalProc) (call, call->arrivalProcHandle,
3629 call->arrivalProcArg);
3630 call->arrivalProc = (void (*)())0;
3633 /* Update last packet received */
3636 /* If there is no server process serving this call, grab
3637 * one, if available. We only need to do this once. If a
3638 * server thread is available, this thread becomes a server
3639 * thread and the server thread becomes a listener thread. */
3641 TryAttach(call, socket, tnop, newcallp, 0);
3644 /* This is not the expected next packet. */
3646 /* Determine whether this is a new or old packet, and if it's
3647 * a new one, whether it fits into the current receive window.
3648 * Also figure out whether the packet was delivered in sequence.
3649 * We use the prev variable to determine whether the new packet
3650 * is the successor of its immediate predecessor in the
3651 * receive queue, and the missing flag to determine whether
3652 * any of this packets predecessors are missing. */
3654 afs_uint32 prev; /* "Previous packet" sequence number */
3655 struct rx_packet *tp; /* Temporary packet pointer */
3656 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3657 int missing; /* Are any predecessors missing? */
3659 /* If the new packet's sequence number has been sent to the
3660 * application already, then this is a duplicate */
3661 if (seq < call->rnext) {
3662 if (rx_stats_active)
3663 rx_atomic_inc(&rx_stats.dupPacketsRead);
3664 rxevent_Cancel(call->delayedAckEvent, call,
3665 RX_CALL_REFCOUNT_DELAY);
3666 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3672 /* If the sequence number is greater than what can be
3673 * accomodated by the current window, then send a negative
3674 * acknowledge and drop the packet */
3675 if ((call->rnext + call->rwind) <= seq) {
3676 rxevent_Cancel(call->delayedAckEvent, call,
3677 RX_CALL_REFCOUNT_DELAY);
3678 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3685 /* Look for the packet in the queue of old received packets */
3686 for (prev = call->rnext - 1, missing =
3687 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3688 /*Check for duplicate packet */
3689 if (seq == tp->header.seq) {
3690 if (rx_stats_active)
3691 rx_atomic_inc(&rx_stats.dupPacketsRead);
3692 rxevent_Cancel(call->delayedAckEvent, call,
3693 RX_CALL_REFCOUNT_DELAY);
3694 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3700 /* If we find a higher sequence packet, break out and
3701 * insert the new packet here. */
3702 if (seq < tp->header.seq)
3704 /* Check for missing packet */
3705 if (tp->header.seq != prev + 1) {
3709 prev = tp->header.seq;
3712 /* Keep track of whether we have received the last packet. */
3713 if (flags & RX_LAST_PACKET) {
3714 call->flags |= RX_CALL_HAVE_LAST;
3717 /* It's within the window: add it to the the receive queue.
3718 * tp is left by the previous loop either pointing at the
3719 * packet before which to insert the new packet, or at the
3720 * queue head if the queue is empty or the packet should be
3722 #ifdef RX_TRACK_PACKETS
3723 np->flags |= RX_PKTFLAG_RQ;
3725 #ifdef RXDEBUG_PACKET
3727 #endif /* RXDEBUG_PACKET */
3728 queue_InsertBefore(tp, np);
3732 /* Check whether we have all of the packets for this call */
3733 if ((call->flags & RX_CALL_HAVE_LAST)
3734 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3735 afs_uint32 tseq; /* temporary sequence number */
3738 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3739 if (tseq != tp->header.seq)
3741 if (tp->header.flags & RX_LAST_PACKET) {
3742 call->flags |= RX_CALL_RECEIVE_DONE;
3749 /* We need to send an ack of the packet is out of sequence,
3750 * or if an ack was requested by the peer. */
3751 if (seq != prev + 1 || missing) {
3752 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3753 } else if (flags & RX_REQUEST_ACK) {
3754 ackNeeded = RX_ACK_REQUESTED;
3757 /* Acknowledge the last packet for each call */
3758 if (flags & RX_LAST_PACKET) {
3769 * If the receiver is waiting for an iovec, fill the iovec
3770 * using the data from the receive queue */
3771 if (call->flags & RX_CALL_IOVEC_WAIT) {
3772 didHardAck = rxi_FillReadVec(call, serial);
3773 /* the call may have been aborted */
3782 /* Wakeup the reader if any */
3783 if ((call->flags & RX_CALL_READER_WAIT)
3784 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3785 || (call->iovNext >= call->iovMax)
3786 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3787 call->flags &= ~RX_CALL_READER_WAIT;
3788 #ifdef RX_ENABLE_LOCKS
3789 CV_BROADCAST(&call->cv_rq);
3791 osi_rxWakeup(&call->rq);
3797 * Send an ack when requested by the peer, or once every
3798 * rxi_SoftAckRate packets until the last packet has been
3799 * received. Always send a soft ack for the last packet in
3800 * the server's reply.
3802 * If we have received all of the packets for the call
3803 * immediately send an RX_PACKET_TYPE_ACKALL packet so that
3804 * the peer can empty its packet queue and cancel all resend
3807 if (call->flags & RX_CALL_RECEIVE_DONE) {
3808 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3809 rxi_AckAll(NULL, call, 0);
3810 } else if (ackNeeded) {
3811 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3812 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3813 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3814 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3815 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3816 } else if (call->nSoftAcks) {
3817 clock_GetTime(&now);
3819 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3820 clock_Add(&when, &rx_lastAckDelay);
3822 clock_Add(&when, &rx_softAckDelay);
3824 if (!call->delayedAckEvent
3825 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3826 rxevent_Cancel(call->delayedAckEvent, call,
3827 RX_CALL_REFCOUNT_DELAY);
3828 MUTEX_ENTER(&rx_refcnt_mutex);
3829 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3830 MUTEX_EXIT(&rx_refcnt_mutex);
3831 call->delayedAckEvent =
3832 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3840 static void rxi_ComputeRate();
3844 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3846 struct rx_peer *peer = conn->peer;
3848 MUTEX_ENTER(&peer->peer_lock);
3849 peer->lastReachTime = clock_Sec();
3850 MUTEX_EXIT(&peer->peer_lock);
3852 MUTEX_ENTER(&conn->conn_data_lock);
3853 if (conn->flags & RX_CONN_ATTACHWAIT) {
3856 conn->flags &= ~RX_CONN_ATTACHWAIT;
3857 MUTEX_EXIT(&conn->conn_data_lock);
3859 for (i = 0; i < RX_MAXCALLS; i++) {
3860 struct rx_call *call = conn->call[i];
3863 MUTEX_ENTER(&call->lock);
3864 /* tnop can be null if newcallp is null */
3865 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3867 MUTEX_EXIT(&call->lock);
3871 MUTEX_EXIT(&conn->conn_data_lock);
3874 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
3876 rx_ack_reason(int reason)
3879 case RX_ACK_REQUESTED:
3881 case RX_ACK_DUPLICATE:
3883 case RX_ACK_OUT_OF_SEQUENCE:
3885 case RX_ACK_EXCEEDS_WINDOW:
3887 case RX_ACK_NOSPACE:
3891 case RX_ACK_PING_RESPONSE:
3904 /* The real smarts of the whole thing. */
3906 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
3909 struct rx_ackPacket *ap;
3911 struct rx_packet *tp;
3912 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3913 struct rx_connection *conn = call->conn;
3914 struct rx_peer *peer = conn->peer;
3915 struct clock now; /* Current time, for RTT calculations */
3919 /* because there are CM's that are bogus, sending weird values for this. */
3920 afs_uint32 skew = 0;
3925 int newAckCount = 0;
3926 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3927 int pktsize = 0; /* Set if we need to update the peer mtu */
3928 int conn_data_locked = 0;
3930 if (rx_stats_active)
3931 rx_atomic_inc(&rx_stats.ackPacketsRead);
3932 ap = (struct rx_ackPacket *)rx_DataOf(np);
3933 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3935 return np; /* truncated ack packet */
3937 /* depends on ack packet struct */
3938 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3939 first = ntohl(ap->firstPacket);
3940 prev = ntohl(ap->previousPacket);
3941 serial = ntohl(ap->serial);
3942 /* temporarily disabled -- needs to degrade over time
3943 * skew = ntohs(ap->maxSkew); */
3945 /* Ignore ack packets received out of order */
3946 if (first < call->tfirst ||
3947 (first == call->tfirst && prev < call->tprev)) {
3953 if (np->header.flags & RX_SLOW_START_OK) {
3954 call->flags |= RX_CALL_SLOW_START_OK;
3957 if (ap->reason == RX_ACK_PING_RESPONSE)
3958 rxi_UpdatePeerReach(conn, call);
3960 if (conn->lastPacketSizeSeq) {
3961 MUTEX_ENTER(&conn->conn_data_lock);
3962 conn_data_locked = 1;
3963 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
3964 pktsize = conn->lastPacketSize;
3965 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
3968 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
3969 if (!conn_data_locked) {
3970 MUTEX_ENTER(&conn->conn_data_lock);
3971 conn_data_locked = 1;
3973 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
3974 /* process mtu ping ack */
3975 pktsize = conn->lastPingSize;
3976 conn->lastPingSizeSer = conn->lastPingSize = 0;
3980 if (conn_data_locked) {
3981 MUTEX_EXIT(&conn->conn_data_lock);
3982 conn_data_locked = 0;
3986 if (rxdebug_active) {
3990 len = _snprintf(msg, sizeof(msg),
3991 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3992 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3993 ntohl(ap->serial), ntohl(ap->previousPacket),
3994 (unsigned int)np->header.seq, (unsigned int)skew,
3995 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3999 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4000 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4004 OutputDebugString(msg);
4006 #else /* AFS_NT40_ENV */
4009 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
4010 ap->reason, ntohl(ap->previousPacket),
4011 (unsigned int)np->header.seq, (unsigned int)serial,
4012 (unsigned int)skew, ntohl(ap->firstPacket));
4015 for (offset = 0; offset < nAcks; offset++)
4016 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4021 #endif /* AFS_NT40_ENV */
4024 MUTEX_ENTER(&peer->peer_lock);
4027 * Start somewhere. Can't assume we can send what we can receive,
4028 * but we are clearly receiving.
4030 if (!peer->maxPacketSize)
4031 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4033 if (pktsize > peer->maxPacketSize) {
4034 peer->maxPacketSize = pktsize;
4035 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4036 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4037 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4038 rxi_ScheduleGrowMTUEvent(call, 1);
4043 /* Update the outgoing packet skew value to the latest value of
4044 * the peer's incoming packet skew value. The ack packet, of
4045 * course, could arrive out of order, but that won't affect things
4047 peer->outPacketSkew = skew;
4049 /* Check for packets that no longer need to be transmitted, and
4050 * discard them. This only applies to packets positively
4051 * acknowledged as having been sent to the peer's upper level.
4052 * All other packets must be retained. So only packets with
4053 * sequence numbers < ap->firstPacket are candidates. */
4055 clock_GetTime(&now);
4057 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
4058 if (tp->header.seq >= first)
4060 call->tfirst = tp->header.seq + 1;
4062 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4065 rxi_ComputeRoundTripTime(tp, ap, call->conn->peer, &now);
4069 rxi_ComputeRate(call->conn->peer, call, p, np, ap->reason);
4072 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4073 /* XXX Hack. Because we have to release the global rx lock when sending
4074 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4075 * in rxi_Start sending packets out because packets may move to the
4076 * freePacketQueue as result of being here! So we drop these packets until
4077 * we're safely out of the traversing. Really ugly!
4078 * To make it even uglier, if we're using fine grain locking, we can
4079 * set the ack bits in the packets and have rxi_Start remove the packets
4080 * when it's done transmitting.
4082 if (call->flags & RX_CALL_TQ_BUSY) {
4083 #ifdef RX_ENABLE_LOCKS
4084 tp->flags |= RX_PKTFLAG_ACKED;
4085 call->flags |= RX_CALL_TQ_SOME_ACKED;
4086 #else /* RX_ENABLE_LOCKS */
4088 #endif /* RX_ENABLE_LOCKS */
4090 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4093 #ifdef RX_TRACK_PACKETS
4094 tp->flags &= ~RX_PKTFLAG_TQ;
4096 #ifdef RXDEBUG_PACKET
4098 #endif /* RXDEBUG_PACKET */
4099 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4104 /* Give rate detector a chance to respond to ping requests */
4105 if (ap->reason == RX_ACK_PING_RESPONSE) {
4106 rxi_ComputeRate(peer, call, 0, np, ap->reason);
4110 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4112 /* Now go through explicit acks/nacks and record the results in
4113 * the waiting packets. These are packets that can't be released
4114 * yet, even with a positive acknowledge. This positive
4115 * acknowledge only means the packet has been received by the
4116 * peer, not that it will be retained long enough to be sent to
4117 * the peer's upper level. In addition, reset the transmit timers
4118 * of any missing packets (those packets that must be missing
4119 * because this packet was out of sequence) */
4121 call->nSoftAcked = 0;
4122 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
4124 /* Set the acknowledge flag per packet based on the
4125 * information in the ack packet. An acknowlegded packet can
4126 * be downgraded when the server has discarded a packet it
4127 * soacked previously, or when an ack packet is received
4128 * out of sequence. */
4129 if (tp->header.seq < first) {
4130 /* Implicit ack information */
4131 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4134 tp->flags |= RX_PKTFLAG_ACKED;
4135 } else if (tp->header.seq < first + nAcks) {
4136 /* Explicit ack information: set it in the packet appropriately */
4137 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4138 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4140 tp->flags |= RX_PKTFLAG_ACKED;
4142 rxi_ComputeRoundTripTime(tp, ap, call->conn->peer, &now);
4144 rxi_ComputeRate(call->conn->peer, call, tp, np,
4153 } else /* RX_ACK_TYPE_NACK */ {
4154 tp->flags &= ~RX_PKTFLAG_ACKED;
4158 if (tp->flags & RX_PKTFLAG_ACKED) {
4159 tp->flags &= ~RX_PKTFLAG_ACKED;
4165 * Following the suggestion of Phil Kern, we back off the peer's
4166 * timeout value for future packets until a successful response
4167 * is received for an initial transmission.
4169 if (missing && !peer->backedOff) {
4170 struct clock c = peer->timeout;
4171 struct clock max_to = {3, 0};
4173 clock_Add(&peer->timeout, &c);
4174 if (clock_Gt(&peer->timeout, &max_to))
4175 peer->timeout = max_to;
4176 peer->backedOff = 1;
4179 /* If packet isn't yet acked, and it has been transmitted at least
4180 * once, reset retransmit time using latest timeout
4181 * ie, this should readjust the retransmit timer for all outstanding
4182 * packets... So we don't just retransmit when we should know better*/
4184 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
4185 tp->retryTime = tp->timeSent;
4186 clock_Add(&tp->retryTime, &peer->timeout);
4187 /* shift by eight because one quarter-sec ~ 256 milliseconds */
4188 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
4192 /* If the window has been extended by this acknowledge packet,
4193 * then wakeup a sender waiting in alloc for window space, or try
4194 * sending packets now, if he's been sitting on packets due to
4195 * lack of window space */
4196 if (call->tnext < (call->tfirst + call->twind)) {
4197 #ifdef RX_ENABLE_LOCKS
4198 CV_SIGNAL(&call->cv_twind);
4200 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4201 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4202 osi_rxWakeup(&call->twind);
4205 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4206 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4210 /* if the ack packet has a receivelen field hanging off it,
4211 * update our state */
4212 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4215 /* If the ack packet has a "recommended" size that is less than
4216 * what I am using now, reduce my size to match */
4217 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4218 (int)sizeof(afs_int32), &tSize);
4219 tSize = (afs_uint32) ntohl(tSize);
4220 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4222 /* Get the maximum packet size to send to this peer */
4223 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4225 tSize = (afs_uint32) ntohl(tSize);
4226 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4227 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4229 /* sanity check - peer might have restarted with different params.
4230 * If peer says "send less", dammit, send less... Peer should never
4231 * be unable to accept packets of the size that prior AFS versions would
4232 * send without asking. */
4233 if (peer->maxMTU != tSize) {
4234 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4236 peer->maxMTU = tSize;
4237 peer->MTU = MIN(tSize, peer->MTU);
4238 call->MTU = MIN(call->MTU, tSize);
4241 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4244 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4245 (int)sizeof(afs_int32), &tSize);
4246 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4247 if (tSize < call->twind) { /* smaller than our send */
4248 call->twind = tSize; /* window, we must send less... */
4249 call->ssthresh = MIN(call->twind, call->ssthresh);
4250 call->conn->twind[call->channel] = call->twind;
4253 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4254 * network MTU confused with the loopback MTU. Calculate the
4255 * maximum MTU here for use in the slow start code below.
4257 /* Did peer restart with older RX version? */
4258 if (peer->maxDgramPackets > 1) {
4259 peer->maxDgramPackets = 1;
4261 } else if (np->length >=
4262 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4265 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4266 sizeof(afs_int32), &tSize);
4267 tSize = (afs_uint32) ntohl(tSize);
4269 * As of AFS 3.5 we set the send window to match the receive window.
4271 if (tSize < call->twind) {
4272 call->twind = tSize;
4273 call->conn->twind[call->channel] = call->twind;
4274 call->ssthresh = MIN(call->twind, call->ssthresh);
4275 } else if (tSize > call->twind) {
4276 call->twind = tSize;
4277 call->conn->twind[call->channel] = call->twind;
4281 * As of AFS 3.5, a jumbogram is more than one fixed size
4282 * packet transmitted in a single UDP datagram. If the remote
4283 * MTU is smaller than our local MTU then never send a datagram
4284 * larger than the natural MTU.
4287 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4288 (int)sizeof(afs_int32), &tSize);
4289 maxDgramPackets = (afs_uint32) ntohl(tSize);
4290 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4292 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4293 if (maxDgramPackets > 1) {
4294 peer->maxDgramPackets = maxDgramPackets;
4295 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4297 peer->maxDgramPackets = 1;
4298 call->MTU = peer->natMTU;
4300 } else if (peer->maxDgramPackets > 1) {
4301 /* Restarted with lower version of RX */
4302 peer->maxDgramPackets = 1;
4304 } else if (peer->maxDgramPackets > 1
4305 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4306 /* Restarted with lower version of RX */
4307 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4308 peer->natMTU = OLD_MAX_PACKET_SIZE;
4309 peer->MTU = OLD_MAX_PACKET_SIZE;
4310 peer->maxDgramPackets = 1;
4311 peer->nDgramPackets = 1;
4313 call->MTU = OLD_MAX_PACKET_SIZE;
4318 * Calculate how many datagrams were successfully received after
4319 * the first missing packet and adjust the negative ack counter
4324 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4325 if (call->nNacks < nNacked) {
4326 call->nNacks = nNacked;
4329 call->nAcks += newAckCount;
4333 if (call->flags & RX_CALL_FAST_RECOVER) {
4335 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4337 call->flags &= ~RX_CALL_FAST_RECOVER;
4338 call->cwind = call->nextCwind;
4339 call->nextCwind = 0;
4342 call->nCwindAcks = 0;
4343 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4344 /* Three negative acks in a row trigger congestion recovery */
4345 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4346 MUTEX_EXIT(&peer->peer_lock);
4347 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4348 /* someone else is waiting to start recovery */
4351 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4352 rxi_WaitforTQBusy(call);
4353 MUTEX_ENTER(&peer->peer_lock);
4354 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4355 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4356 call->flags |= RX_CALL_FAST_RECOVER;
4357 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4359 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4360 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4361 call->nextCwind = call->ssthresh;
4364 peer->MTU = call->MTU;
4365 peer->cwind = call->nextCwind;
4366 peer->nDgramPackets = call->nDgramPackets;
4368 call->congestSeq = peer->congestSeq;
4369 /* Reset the resend times on the packets that were nacked
4370 * so we will retransmit as soon as the window permits*/
4371 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4373 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4374 clock_Zero(&tp->retryTime);
4376 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4381 /* If cwind is smaller than ssthresh, then increase
4382 * the window one packet for each ack we receive (exponential
4384 * If cwind is greater than or equal to ssthresh then increase
4385 * the congestion window by one packet for each cwind acks we
4386 * receive (linear growth). */
4387 if (call->cwind < call->ssthresh) {
4389 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4390 call->nCwindAcks = 0;
4392 call->nCwindAcks += newAckCount;
4393 if (call->nCwindAcks >= call->cwind) {
4394 call->nCwindAcks = 0;
4395 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4399 * If we have received several acknowledgements in a row then
4400 * it is time to increase the size of our datagrams
4402 if ((int)call->nAcks > rx_nDgramThreshold) {
4403 if (peer->maxDgramPackets > 1) {
4404 if (call->nDgramPackets < peer->maxDgramPackets) {
4405 call->nDgramPackets++;
4407 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4408 } else if (call->MTU < peer->maxMTU) {
4409 /* don't upgrade if we can't handle it */
4410 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4411 call->MTU = peer->ifMTU;
4413 call->MTU += peer->natMTU;
4414 call->MTU = MIN(call->MTU, peer->maxMTU);
4421 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4423 /* Servers need to hold the call until all response packets have
4424 * been acknowledged. Soft acks are good enough since clients
4425 * are not allowed to clear their receive queues. */
4426 if (call->state == RX_STATE_HOLD
4427 && call->tfirst + call->nSoftAcked >= call->tnext) {
4428 call->state = RX_STATE_DALLY;
4429 rxi_ClearTransmitQueue(call, 0);
4430 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4431 } else if (!queue_IsEmpty(&call->tq)) {
4432 rxi_Start(0, call, 0, istack);
4437 /* Received a response to a challenge packet */
4439 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4440 struct rx_packet *np, int istack)
4444 /* Ignore the packet if we're the client */
4445 if (conn->type == RX_CLIENT_CONNECTION)
4448 /* If already authenticated, ignore the packet (it's probably a retry) */
4449 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4452 /* Otherwise, have the security object evaluate the response packet */
4453 error = RXS_CheckResponse(conn->securityObject, conn, np);
4455 /* If the response is invalid, reset the connection, sending
4456 * an abort to the peer */
4460 rxi_ConnectionError(conn, error);
4461 MUTEX_ENTER(&conn->conn_data_lock);
4462 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4463 MUTEX_EXIT(&conn->conn_data_lock);
4466 /* If the response is valid, any calls waiting to attach
4467 * servers can now do so */
4470 for (i = 0; i < RX_MAXCALLS; i++) {
4471 struct rx_call *call = conn->call[i];
4473 MUTEX_ENTER(&call->lock);
4474 if (call->state == RX_STATE_PRECALL)
4475 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4476 /* tnop can be null if newcallp is null */
4477 MUTEX_EXIT(&call->lock);
4481 /* Update the peer reachability information, just in case
4482 * some calls went into attach-wait while we were waiting
4483 * for authentication..
4485 rxi_UpdatePeerReach(conn, NULL);
4490 /* A client has received an authentication challenge: the security
4491 * object is asked to cough up a respectable response packet to send
4492 * back to the server. The server is responsible for retrying the
4493 * challenge if it fails to get a response. */
4496 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4497 struct rx_packet *np, int istack)
4501 /* Ignore the challenge if we're the server */
4502 if (conn->type == RX_SERVER_CONNECTION)
4505 /* Ignore the challenge if the connection is otherwise idle; someone's
4506 * trying to use us as an oracle. */
4507 if (!rxi_HasActiveCalls(conn))
4510 /* Send the security object the challenge packet. It is expected to fill
4511 * in the response. */
4512 error = RXS_GetResponse(conn->securityObject, conn, np);
4514 /* If the security object is unable to return a valid response, reset the
4515 * connection and send an abort to the peer. Otherwise send the response
4516 * packet to the peer connection. */
4518 rxi_ConnectionError(conn, error);
4519 MUTEX_ENTER(&conn->conn_data_lock);
4520 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4521 MUTEX_EXIT(&conn->conn_data_lock);
4523 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4524 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4530 /* Find an available server process to service the current request in
4531 * the given call structure. If one isn't available, queue up this
4532 * call so it eventually gets one */
4534 rxi_AttachServerProc(struct rx_call *call,
4535 osi_socket socket, int *tnop,
4536 struct rx_call **newcallp)
4538 struct rx_serverQueueEntry *sq;
4539 struct rx_service *service = call->conn->service;
4542 /* May already be attached */
4543 if (call->state == RX_STATE_ACTIVE)
4546 MUTEX_ENTER(&rx_serverPool_lock);
4548 haveQuota = QuotaOK(service);
4549 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4550 /* If there are no processes available to service this call,
4551 * put the call on the incoming call queue (unless it's
4552 * already on the queue).
4554 #ifdef RX_ENABLE_LOCKS
4556 ReturnToServerPool(service);
4557 #endif /* RX_ENABLE_LOCKS */
4559 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4560 call->flags |= RX_CALL_WAIT_PROC;
4561 rx_atomic_inc(&rx_nWaiting);
4562 rx_atomic_inc(&rx_nWaited);
4563 rxi_calltrace(RX_CALL_ARRIVAL, call);
4564 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4565 queue_Append(&rx_incomingCallQueue, call);
4568 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4570 /* If hot threads are enabled, and both newcallp and sq->socketp
4571 * are non-null, then this thread will process the call, and the
4572 * idle server thread will start listening on this threads socket.
4575 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4578 *sq->socketp = socket;
4579 clock_GetTime(&call->startTime);
4580 MUTEX_ENTER(&rx_refcnt_mutex);
4581 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4582 MUTEX_EXIT(&rx_refcnt_mutex);
4586 if (call->flags & RX_CALL_WAIT_PROC) {
4587 /* Conservative: I don't think this should happen */
4588 call->flags &= ~RX_CALL_WAIT_PROC;
4589 if (queue_IsOnQueue(call)) {
4592 rx_atomic_dec(&rx_nWaiting);
4595 call->state = RX_STATE_ACTIVE;
4596 call->mode = RX_MODE_RECEIVING;
4597 #ifdef RX_KERNEL_TRACE
4599 int glockOwner = ISAFS_GLOCK();
4602 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4603 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4609 if (call->flags & RX_CALL_CLEARED) {
4610 /* send an ack now to start the packet flow up again */
4611 call->flags &= ~RX_CALL_CLEARED;
4612 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4614 #ifdef RX_ENABLE_LOCKS
4617 service->nRequestsRunning++;
4618 MUTEX_ENTER(&rx_quota_mutex);
4619 if (service->nRequestsRunning <= service->minProcs)
4622 MUTEX_EXIT(&rx_quota_mutex);
4626 MUTEX_EXIT(&rx_serverPool_lock);
4629 /* Delay the sending of an acknowledge event for a short while, while
4630 * a new call is being prepared (in the case of a client) or a reply
4631 * is being prepared (in the case of a server). Rather than sending
4632 * an ack packet, an ACKALL packet is sent. */
4634 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4636 #ifdef RX_ENABLE_LOCKS
4638 MUTEX_ENTER(&call->lock);
4639 call->delayedAckEvent = NULL;
4640 MUTEX_ENTER(&rx_refcnt_mutex);
4641 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4642 MUTEX_EXIT(&rx_refcnt_mutex);
4644 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4645 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4647 MUTEX_EXIT(&call->lock);
4648 #else /* RX_ENABLE_LOCKS */
4650 call->delayedAckEvent = NULL;
4651 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4652 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4653 #endif /* RX_ENABLE_LOCKS */
4657 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4659 struct rx_call *call = arg1;
4660 #ifdef RX_ENABLE_LOCKS
4662 MUTEX_ENTER(&call->lock);
4663 if (event == call->delayedAckEvent)
4664 call->delayedAckEvent = NULL;
4665 MUTEX_ENTER(&rx_refcnt_mutex);
4666 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4667 MUTEX_EXIT(&rx_refcnt_mutex);
4669 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4671 MUTEX_EXIT(&call->lock);
4672 #else /* RX_ENABLE_LOCKS */
4674 call->delayedAckEvent = NULL;
4675 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4676 #endif /* RX_ENABLE_LOCKS */
4680 #ifdef RX_ENABLE_LOCKS
4681 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4682 * clearing them out.
4685 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4687 struct rx_packet *p, *tp;
4690 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4691 p->flags |= RX_PKTFLAG_ACKED;
4695 call->flags |= RX_CALL_TQ_CLEARME;
4696 call->flags |= RX_CALL_TQ_SOME_ACKED;
4699 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4700 call->tfirst = call->tnext;
4701 call->nSoftAcked = 0;
4703 if (call->flags & RX_CALL_FAST_RECOVER) {
4704 call->flags &= ~RX_CALL_FAST_RECOVER;
4705 call->cwind = call->nextCwind;
4706 call->nextCwind = 0;
4709 CV_SIGNAL(&call->cv_twind);
4711 #endif /* RX_ENABLE_LOCKS */
4713 /* Clear out the transmit queue for the current call (all packets have
4714 * been received by peer) */
4716 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4718 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4719 struct rx_packet *p, *tp;
4721 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4723 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4724 p->flags |= RX_PKTFLAG_ACKED;
4728 call->flags |= RX_CALL_TQ_CLEARME;
4729 call->flags |= RX_CALL_TQ_SOME_ACKED;
4732 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4733 #ifdef RXDEBUG_PACKET
4735 #endif /* RXDEBUG_PACKET */
4736 rxi_FreePackets(0, &call->tq);
4737 rxi_WakeUpTransmitQueue(call);
4738 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4739 call->flags &= ~RX_CALL_TQ_CLEARME;
4741 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4743 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4744 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4745 call->nSoftAcked = 0;
4747 if (call->flags & RX_CALL_FAST_RECOVER) {
4748 call->flags &= ~RX_CALL_FAST_RECOVER;
4749 call->cwind = call->nextCwind;
4751 #ifdef RX_ENABLE_LOCKS
4752 CV_SIGNAL(&call->cv_twind);
4754 osi_rxWakeup(&call->twind);
4759 rxi_ClearReceiveQueue(struct rx_call *call)
4761 if (queue_IsNotEmpty(&call->rq)) {
4764 count = rxi_FreePackets(0, &call->rq);
4765 rx_packetReclaims += count;
4766 #ifdef RXDEBUG_PACKET
4768 if ( call->rqc != 0 )
4769 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
4771 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4773 if (call->state == RX_STATE_PRECALL) {
4774 call->flags |= RX_CALL_CLEARED;
4778 /* Send an abort packet for the specified call */
4780 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
4781 int istack, int force)
4784 struct clock when, now;
4789 /* Clients should never delay abort messages */
4790 if (rx_IsClientConn(call->conn))
4793 if (call->abortCode != call->error) {
4794 call->abortCode = call->error;
4795 call->abortCount = 0;
4798 if (force || rxi_callAbortThreshhold == 0
4799 || call->abortCount < rxi_callAbortThreshhold) {
4800 if (call->delayedAbortEvent) {
4801 rxevent_Cancel(call->delayedAbortEvent, call,
4802 RX_CALL_REFCOUNT_ABORT);
4804 error = htonl(call->error);
4807 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4808 (char *)&error, sizeof(error), istack);
4809 } else if (!call->delayedAbortEvent) {
4810 clock_GetTime(&now);
4812 clock_Addmsec(&when, rxi_callAbortDelay);
4813 MUTEX_ENTER(&rx_refcnt_mutex);
4814 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4815 MUTEX_EXIT(&rx_refcnt_mutex);
4816 call->delayedAbortEvent =
4817 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4822 /* Send an abort packet for the specified connection. Packet is an
4823 * optional pointer to a packet that can be used to send the abort.
4824 * Once the number of abort messages reaches the threshhold, an
4825 * event is scheduled to send the abort. Setting the force flag
4826 * overrides sending delayed abort messages.
4828 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4829 * to send the abort packet.
4832 rxi_SendConnectionAbort(struct rx_connection *conn,
4833 struct rx_packet *packet, int istack, int force)
4836 struct clock when, now;
4841 /* Clients should never delay abort messages */
4842 if (rx_IsClientConn(conn))
4845 if (force || rxi_connAbortThreshhold == 0
4846 || conn->abortCount < rxi_connAbortThreshhold) {
4847 if (conn->delayedAbortEvent) {
4848 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4850 error = htonl(conn->error);
4852 MUTEX_EXIT(&conn->conn_data_lock);
4854 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4855 RX_PACKET_TYPE_ABORT, (char *)&error,
4856 sizeof(error), istack);
4857 MUTEX_ENTER(&conn->conn_data_lock);
4858 } else if (!conn->delayedAbortEvent) {
4859 clock_GetTime(&now);
4861 clock_Addmsec(&when, rxi_connAbortDelay);
4862 conn->delayedAbortEvent =
4863 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4868 /* Associate an error all of the calls owned by a connection. Called
4869 * with error non-zero. This is only for really fatal things, like
4870 * bad authentication responses. The connection itself is set in
4871 * error at this point, so that future packets received will be
4874 rxi_ConnectionError(struct rx_connection *conn,
4880 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
4882 MUTEX_ENTER(&conn->conn_data_lock);
4883 if (conn->challengeEvent)
4884 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4885 if (conn->natKeepAliveEvent)
4886 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
4887 if (conn->checkReachEvent) {
4888 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4889 conn->checkReachEvent = 0;
4890 conn->flags &= ~RX_CONN_ATTACHWAIT;
4891 MUTEX_ENTER(&rx_refcnt_mutex);
4893 MUTEX_EXIT(&rx_refcnt_mutex);
4895 MUTEX_EXIT(&conn->conn_data_lock);
4896 for (i = 0; i < RX_MAXCALLS; i++) {
4897 struct rx_call *call = conn->call[i];
4899 MUTEX_ENTER(&call->lock);
4900 rxi_CallError(call, error);
4901 MUTEX_EXIT(&call->lock);
4904 conn->error = error;
4905 if (rx_stats_active)
4906 rx_atomic_inc(&rx_stats.fatalErrors);
4911 * Interrupt an in-progress call with the specified error and wakeup waiters.
4913 * @param[in] call The call to interrupt
4914 * @param[in] error The error code to send to the peer
4917 rx_InterruptCall(struct rx_call *call, afs_int32 error)
4919 MUTEX_ENTER(&call->lock);
4920 rxi_CallError(call, error);
4921 rxi_SendCallAbort(call, NULL, 0, 1);
4922 MUTEX_EXIT(&call->lock);
4926 rxi_CallError(struct rx_call *call, afs_int32 error)
4929 osirx_AssertMine(&call->lock, "rxi_CallError");
4931 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
4933 error = call->error;
4935 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4936 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4937 rxi_ResetCall(call, 0);
4940 rxi_ResetCall(call, 0);
4942 call->error = error;
4945 /* Reset various fields in a call structure, and wakeup waiting
4946 * processes. Some fields aren't changed: state & mode are not
4947 * touched (these must be set by the caller), and bufptr, nLeft, and
4948 * nFree are not reset, since these fields are manipulated by
4949 * unprotected macros, and may only be reset by non-interrupting code.
4952 /* this code requires that call->conn be set properly as a pre-condition. */
4953 #endif /* ADAPT_WINDOW */
4956 rxi_ResetCall(struct rx_call *call, int newcall)
4959 struct rx_peer *peer;
4960 struct rx_packet *packet;
4962 osirx_AssertMine(&call->lock, "rxi_ResetCall");
4964 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
4966 /* Notify anyone who is waiting for asynchronous packet arrival */
4967 if (call->arrivalProc) {
4968 (*call->arrivalProc) (call, call->arrivalProcHandle,
4969 call->arrivalProcArg);
4970 call->arrivalProc = (void (*)())0;
4973 if (call->delayedAbortEvent) {
4974 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4975 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4977 rxi_SendCallAbort(call, packet, 0, 1);
4978 rxi_FreePacket(packet);
4983 * Update the peer with the congestion information in this call
4984 * so other calls on this connection can pick up where this call
4985 * left off. If the congestion sequence numbers don't match then
4986 * another call experienced a retransmission.
4988 peer = call->conn->peer;
4989 MUTEX_ENTER(&peer->peer_lock);
4991 if (call->congestSeq == peer->congestSeq) {
4992 peer->cwind = MAX(peer->cwind, call->cwind);
4993 peer->MTU = MAX(peer->MTU, call->MTU);
4994 peer->nDgramPackets =
4995 MAX(peer->nDgramPackets, call->nDgramPackets);
4998 call->abortCode = 0;
4999 call->abortCount = 0;
5001 if (peer->maxDgramPackets > 1) {
5002 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5004 call->MTU = peer->MTU;
5006 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5007 call->ssthresh = rx_maxSendWindow;
5008 call->nDgramPackets = peer->nDgramPackets;
5009 call->congestSeq = peer->congestSeq;
5010 MUTEX_EXIT(&peer->peer_lock);
5012 flags = call->flags;
5013 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5014 rxi_WaitforTQBusy(call);
5015 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5017 rxi_ClearTransmitQueue(call, 1);
5018 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5019 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5023 rxi_ClearReceiveQueue(call);
5024 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5028 call->twind = call->conn->twind[call->channel];
5029 call->rwind = call->conn->rwind[call->channel];
5030 call->nSoftAcked = 0;
5031 call->nextCwind = 0;
5034 call->nCwindAcks = 0;
5035 call->nSoftAcks = 0;
5036 call->nHardAcks = 0;
5038 call->tfirst = call->rnext = call->tnext = 1;
5041 call->lastAcked = 0;
5042 call->localStatus = call->remoteStatus = 0;
5044 if (flags & RX_CALL_READER_WAIT) {
5045 #ifdef RX_ENABLE_LOCKS
5046 CV_BROADCAST(&call->cv_rq);
5048 osi_rxWakeup(&call->rq);
5051 if (flags & RX_CALL_WAIT_PACKETS) {
5052 MUTEX_ENTER(&rx_freePktQ_lock);
5053 rxi_PacketsUnWait(); /* XXX */
5054 MUTEX_EXIT(&rx_freePktQ_lock);
5056 #ifdef RX_ENABLE_LOCKS
5057 CV_SIGNAL(&call->cv_twind);
5059 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5060 osi_rxWakeup(&call->twind);
5063 #ifdef RX_ENABLE_LOCKS
5064 /* The following ensures that we don't mess with any queue while some
5065 * other thread might also be doing so. The call_queue_lock field is
5066 * is only modified under the call lock. If the call is in the process
5067 * of being removed from a queue, the call is not locked until the
5068 * the queue lock is dropped and only then is the call_queue_lock field
5069 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5070 * Note that any other routine which removes a call from a queue has to
5071 * obtain the queue lock before examing the queue and removing the call.
5073 if (call->call_queue_lock) {
5074 MUTEX_ENTER(call->call_queue_lock);
5075 if (queue_IsOnQueue(call)) {
5077 if (flags & RX_CALL_WAIT_PROC) {
5078 rx_atomic_dec(&rx_nWaiting);
5081 MUTEX_EXIT(call->call_queue_lock);
5082 CLEAR_CALL_QUEUE_LOCK(call);
5084 #else /* RX_ENABLE_LOCKS */
5085 if (queue_IsOnQueue(call)) {
5087 if (flags & RX_CALL_WAIT_PROC)
5088 rx_atomic_dec(&rx_nWaiting);
5090 #endif /* RX_ENABLE_LOCKS */
5092 rxi_KeepAliveOff(call);
5093 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5096 /* Send an acknowledge for the indicated packet (seq,serial) of the
5097 * indicated call, for the indicated reason (reason). This
5098 * acknowledge will specifically acknowledge receiving the packet, and
5099 * will also specify which other packets for this call have been
5100 * received. This routine returns the packet that was used to the
5101 * caller. The caller is responsible for freeing it or re-using it.
5102 * This acknowledgement also returns the highest sequence number
5103 * actually read out by the higher level to the sender; the sender
5104 * promises to keep around packets that have not been read by the
5105 * higher level yet (unless, of course, the sender decides to abort
5106 * the call altogether). Any of p, seq, serial, pflags, or reason may
5107 * be set to zero without ill effect. That is, if they are zero, they
5108 * will not convey any information.
5109 * NOW there is a trailer field, after the ack where it will safely be
5110 * ignored by mundanes, which indicates the maximum size packet this
5111 * host can swallow. */
5113 struct rx_packet *optionalPacket; use to send ack (or null)
5114 int seq; Sequence number of the packet we are acking
5115 int serial; Serial number of the packet
5116 int pflags; Flags field from packet header
5117 int reason; Reason an acknowledge was prompted
5121 rxi_SendAck(struct rx_call *call,
5122 struct rx_packet *optionalPacket, int serial, int reason,
5125 struct rx_ackPacket *ap;
5126 struct rx_packet *rqp;
5127 struct rx_packet *nxp; /* For queue_Scan */
5128 struct rx_packet *p;
5131 afs_uint32 padbytes = 0;
5132 #ifdef RX_ENABLE_TSFPQ
5133 struct rx_ts_info_t * rx_ts_info;
5137 * Open the receive window once a thread starts reading packets
5139 if (call->rnext > 1) {
5140 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5143 /* Don't attempt to grow MTU if this is a critical ping */
5144 if (reason == RX_ACK_MTU) {
5145 /* keep track of per-call attempts, if we're over max, do in small
5146 * otherwise in larger? set a size to increment by, decrease
5149 if (call->conn->peer->maxPacketSize &&
5150 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5152 padbytes = call->conn->peer->maxPacketSize+16;
5154 padbytes = call->conn->peer->maxMTU + 128;
5156 /* do always try a minimum size ping */
5157 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5159 /* subtract the ack payload */
5160 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5161 reason = RX_ACK_PING;
5164 call->nHardAcks = 0;
5165 call->nSoftAcks = 0;
5166 if (call->rnext > call->lastAcked)
5167 call->lastAcked = call->rnext;
5171 rx_computelen(p, p->length); /* reset length, you never know */
5172 } /* where that's been... */
5173 #ifdef RX_ENABLE_TSFPQ
5175 RX_TS_INFO_GET(rx_ts_info);
5176 if ((p = rx_ts_info->local_special_packet)) {
5177 rx_computelen(p, p->length);
5178 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5179 rx_ts_info->local_special_packet = p;
5180 } else { /* We won't send the ack, but don't panic. */
5181 return optionalPacket;
5185 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5186 /* We won't send the ack, but don't panic. */
5187 return optionalPacket;
5192 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5195 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5196 #ifndef RX_ENABLE_TSFPQ
5197 if (!optionalPacket)
5200 return optionalPacket;
5202 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5203 if (rx_Contiguous(p) < templ) {
5204 #ifndef RX_ENABLE_TSFPQ
5205 if (!optionalPacket)
5208 return optionalPacket;
5213 /* MTUXXX failing to send an ack is very serious. We should */
5214 /* try as hard as possible to send even a partial ack; it's */
5215 /* better than nothing. */
5216 ap = (struct rx_ackPacket *)rx_DataOf(p);
5217 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5218 ap->reason = reason;
5220 /* The skew computation used to be bogus, I think it's better now. */
5221 /* We should start paying attention to skew. XXX */
5222 ap->serial = htonl(serial);
5223 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5225 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
5226 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5228 /* No fear of running out of ack packet here because there can only be at most
5229 * one window full of unacknowledged packets. The window size must be constrained
5230 * to be less than the maximum ack size, of course. Also, an ack should always
5231 * fit into a single packet -- it should not ever be fragmented. */
5232 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5233 if (!rqp || !call->rq.next
5234 || (rqp->header.seq > (call->rnext + call->rwind))) {
5235 #ifndef RX_ENABLE_TSFPQ
5236 if (!optionalPacket)
5239 rxi_CallError(call, RX_CALL_DEAD);
5240 return optionalPacket;
5243 while (rqp->header.seq > call->rnext + offset)
5244 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5245 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5247 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5248 #ifndef RX_ENABLE_TSFPQ
5249 if (!optionalPacket)
5252 rxi_CallError(call, RX_CALL_DEAD);
5253 return optionalPacket;
5258 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5260 /* these are new for AFS 3.3 */
5261 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5262 templ = htonl(templ);
5263 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5264 templ = htonl(call->conn->peer->ifMTU);
5265 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5266 sizeof(afs_int32), &templ);
5268 /* new for AFS 3.4 */
5269 templ = htonl(call->rwind);
5270 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5271 sizeof(afs_int32), &templ);
5273 /* new for AFS 3.5 */
5274 templ = htonl(call->conn->peer->ifDgramPackets);
5275 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5276 sizeof(afs_int32), &templ);
5278 p->header.serviceId = call->conn->serviceId;
5279 p->header.cid = (call->conn->cid | call->channel);
5280 p->header.callNumber = *call->callNumber;
5282 p->header.securityIndex = call->conn->securityIndex;
5283 p->header.epoch = call->conn->epoch;
5284 p->header.type = RX_PACKET_TYPE_ACK;
5285 p->header.flags = RX_SLOW_START_OK;
5286 if (reason == RX_ACK_PING) {
5287 p->header.flags |= RX_REQUEST_ACK;
5289 clock_GetTime(&call->pingRequestTime);
5292 p->length = padbytes +
5293 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5296 /* not fast but we can potentially use this if truncated
5297 * fragments are delivered to figure out the mtu.
5299 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5300 sizeof(afs_int32), sizeof(afs_int32),
5304 if (call->conn->type == RX_CLIENT_CONNECTION)
5305 p->header.flags |= RX_CLIENT_INITIATED;
5309 if (rxdebug_active) {
5313 len = _snprintf(msg, sizeof(msg),
5314 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5315 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5316 ntohl(ap->serial), ntohl(ap->previousPacket),
5317 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5318 ap->nAcks, ntohs(ap->bufferSpace) );
5322 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5323 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5327 OutputDebugString(msg);
5329 #else /* AFS_NT40_ENV */
5331 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5332 ap->reason, ntohl(ap->previousPacket),
5333 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5335 for (offset = 0; offset < ap->nAcks; offset++)
5336 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5341 #endif /* AFS_NT40_ENV */
5344 int i, nbytes = p->length;
5346 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5347 if (nbytes <= p->wirevec[i].iov_len) {
5350 savelen = p->wirevec[i].iov_len;
5352 p->wirevec[i].iov_len = nbytes;
5354 rxi_Send(call, p, istack);
5355 p->wirevec[i].iov_len = savelen;
5359 nbytes -= p->wirevec[i].iov_len;
5362 if (rx_stats_active)
5363 rx_atomic_inc(&rx_stats.ackPacketsSent);
5364 #ifndef RX_ENABLE_TSFPQ
5365 if (!optionalPacket)
5368 return optionalPacket; /* Return packet for re-use by caller */
5372 struct rx_packet **list;
5377 /* Send all of the packets in the list in single datagram */
5379 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5380 int istack, int moreFlag)
5385 struct clock now, retryTime;
5386 struct rx_connection *conn = call->conn;
5387 struct rx_peer *peer = conn->peer;
5389 MUTEX_ENTER(&peer->peer_lock);
5390 peer->nSent += xmit->len;
5391 if (xmit->resending)
5392 peer->reSends += xmit->len;
5393 retryTime = peer->timeout;
5394 MUTEX_EXIT(&peer->peer_lock);
5396 if (rx_stats_active) {
5397 if (xmit->resending)
5398 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5400 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5403 clock_GetTime(&now);
5404 clock_Add(&retryTime, &now);
5406 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5410 /* Set the packet flags and schedule the resend events */
5411 /* Only request an ack for the last packet in the list */
5412 for (i = 0; i < xmit->len; i++) {
5413 struct rx_packet *packet = xmit->list[i];
5415 packet->retryTime = retryTime;
5416 if (packet->header.serial) {
5417 /* Exponentially backoff retry times */
5418 if (packet->backoff < MAXBACKOFF) {
5419 /* so it can't stay == 0 */
5420 packet->backoff = (packet->backoff << 1) + 1;
5423 clock_Addmsec(&(packet->retryTime),
5424 ((afs_uint32) packet->backoff) << 8);
5427 /* Wait a little extra for the ack on the last packet */
5429 && !(packet->header.flags & RX_CLIENT_INITIATED)) {
5430 clock_Addmsec(&(packet->retryTime), 400);
5433 /* Record the time sent */
5434 packet->timeSent = now;
5436 /* Ask for an ack on retransmitted packets, on every other packet
5437 * if the peer doesn't support slow start. Ask for an ack on every
5438 * packet until the congestion window reaches the ack rate. */
5439 if (packet->header.serial) {
5442 /* improved RTO calculation- not Karn */
5443 packet->firstSent = now;
5444 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5445 || (!(call->flags & RX_CALL_SLOW_START_OK)
5446 && (packet->header.seq & 1)))) {
5451 /* Tag this packet as not being the last in this group,
5452 * for the receiver's benefit */
5453 if (i < xmit->len - 1 || moreFlag) {
5454 packet->header.flags |= RX_MORE_PACKETS;
5459 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5462 /* Since we're about to send a data packet to the peer, it's
5463 * safe to nuke any scheduled end-of-packets ack */
5464 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5466 MUTEX_EXIT(&call->lock);
5467 MUTEX_ENTER(&rx_refcnt_mutex);
5468 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5469 MUTEX_EXIT(&rx_refcnt_mutex);
5470 if (xmit->len > 1) {
5471 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5473 rxi_SendPacket(call, conn, xmit->list[0], istack);
5475 MUTEX_ENTER(&call->lock);
5476 MUTEX_ENTER(&rx_refcnt_mutex);
5477 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5478 MUTEX_EXIT(&rx_refcnt_mutex);
5480 /* Update last send time for this call (for keep-alive
5481 * processing), and for the connection (so that we can discover
5482 * idle connections) */
5483 conn->lastSendTime = call->lastSendTime = clock_Sec();
5484 /* Let a set of retransmits trigger an idle timeout */
5485 if (!xmit->resending)
5486 call->lastSendData = call->lastSendTime;
5489 /* When sending packets we need to follow these rules:
5490 * 1. Never send more than maxDgramPackets in a jumbogram.
5491 * 2. Never send a packet with more than two iovecs in a jumbogram.
5492 * 3. Never send a retransmitted packet in a jumbogram.
5493 * 4. Never send more than cwind/4 packets in a jumbogram
5494 * We always keep the last list we should have sent so we
5495 * can set the RX_MORE_PACKETS flags correctly.
5499 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5503 struct xmitlist working;
5504 struct xmitlist last;
5506 struct rx_peer *peer = call->conn->peer;
5507 int morePackets = 0;
5509 memset(&last, 0, sizeof(struct xmitlist));
5510 working.list = &list[0];
5512 working.resending = 0;
5514 for (i = 0; i < len; i++) {
5515 /* Does the current packet force us to flush the current list? */
5517 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5518 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5520 /* This sends the 'last' list and then rolls the current working
5521 * set into the 'last' one, and resets the working set */
5524 rxi_SendList(call, &last, istack, 1);
5525 /* If the call enters an error state stop sending, or if
5526 * we entered congestion recovery mode, stop sending */
5527 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5532 working.resending = 0;
5533 working.list = &list[i];
5535 /* Add the current packet to the list if it hasn't been acked.
5536 * Otherwise adjust the list pointer to skip the current packet. */
5537 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5540 if (list[i]->header.serial)
5541 working.resending = 1;
5543 /* Do we need to flush the list? */
5544 if (working.len >= (int)peer->maxDgramPackets
5545 || working.len >= (int)call->nDgramPackets
5546 || working.len >= (int)call->cwind
5547 || list[i]->header.serial
5548 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5550 rxi_SendList(call, &last, istack, 1);
5551 /* If the call enters an error state stop sending, or if
5552 * we entered congestion recovery mode, stop sending */
5554 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5559 working.resending = 0;
5560 working.list = &list[i + 1];
5563 if (working.len != 0) {
5564 osi_Panic("rxi_SendList error");
5566 working.list = &list[i + 1];
5570 /* Send the whole list when the call is in receive mode, when
5571 * the call is in eof mode, when we are in fast recovery mode,
5572 * and when we have the last packet */
5573 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5574 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5575 || (call->flags & RX_CALL_FAST_RECOVER)) {
5576 /* Check for the case where the current list contains
5577 * an acked packet. Since we always send retransmissions
5578 * in a separate packet, we only need to check the first
5579 * packet in the list */
5580 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5584 rxi_SendList(call, &last, istack, morePackets);
5585 /* If the call enters an error state stop sending, or if
5586 * we entered congestion recovery mode, stop sending */
5587 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5591 rxi_SendList(call, &working, istack, 0);
5593 } else if (last.len > 0) {
5594 rxi_SendList(call, &last, istack, 0);
5595 /* Packets which are in 'working' are not sent by this call */
5599 #ifdef RX_ENABLE_LOCKS
5600 /* Call rxi_Start, below, but with the call lock held. */
5602 rxi_StartUnlocked(struct rxevent *event,
5603 void *arg0, void *arg1, int istack)
5605 struct rx_call *call = arg0;
5607 MUTEX_ENTER(&call->lock);
5608 rxi_Start(event, call, arg1, istack);
5609 MUTEX_EXIT(&call->lock);
5611 #endif /* RX_ENABLE_LOCKS */
5613 /* This routine is called when new packets are readied for
5614 * transmission and when retransmission may be necessary, or when the
5615 * transmission window or burst count are favourable. This should be
5616 * better optimized for new packets, the usual case, now that we've
5617 * got rid of queues of send packets. XXXXXXXXXXX */
5619 rxi_Start(struct rxevent *event,
5620 void *arg0, void *arg1, int istack)
5622 struct rx_call *call = arg0;
5624 struct rx_packet *p;
5625 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5626 struct clock now, usenow, retryTime;
5631 /* If rxi_Start is being called as a result of a resend event,
5632 * then make sure that the event pointer is removed from the call
5633 * structure, since there is no longer a per-call retransmission
5635 if (event && event == call->resendEvent) {
5636 MUTEX_ENTER(&rx_refcnt_mutex);
5637 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5638 MUTEX_EXIT(&rx_refcnt_mutex);
5639 call->resendEvent = NULL;
5640 if (queue_IsEmpty(&call->tq)) {
5647 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5648 if (rx_stats_active)
5649 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5654 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5656 clock_GetTime(&now);
5659 /* Send (or resend) any packets that need it, subject to
5660 * window restrictions and congestion burst control
5661 * restrictions. Ask for an ack on the last packet sent in
5662 * this burst. For now, we're relying upon the window being
5663 * considerably bigger than the largest number of packets that
5664 * are typically sent at once by one initial call to
5665 * rxi_Start. This is probably bogus (perhaps we should ask
5666 * for an ack when we're half way through the current
5667 * window?). Also, for non file transfer applications, this
5668 * may end up asking for an ack for every packet. Bogus. XXXX
5671 * But check whether we're here recursively, and let the other guy
5674 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5675 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5676 call->flags |= RX_CALL_TQ_BUSY;
5678 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5680 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5681 call->flags &= ~RX_CALL_NEED_START;
5682 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5684 maxXmitPackets = MIN(call->twind, call->cwind);
5685 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5686 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5687 /* We shouldn't be sending packets if a thread is waiting
5688 * to initiate congestion recovery */
5689 dpf(("call %d waiting to initiate fast recovery\n",
5690 *(call->callNumber)));
5694 && (call->flags & RX_CALL_FAST_RECOVER)) {
5695 /* Only send one packet during fast recovery */
5696 dpf(("call %d restricted to one packet per send during fast recovery\n",
5697 *(call->callNumber)));
5700 #ifdef RX_TRACK_PACKETS
5701 if ((p->flags & RX_PKTFLAG_FREE)
5702 || (!queue_IsEnd(&call->tq, nxp)
5703 && (nxp->flags & RX_PKTFLAG_FREE))
5704 || (p == (struct rx_packet *)&rx_freePacketQueue)
5705 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5706 osi_Panic("rxi_Start: xmit queue clobbered");
5709 if (p->flags & RX_PKTFLAG_ACKED) {
5710 /* Since we may block, don't trust this */
5711 usenow.sec = usenow.usec = 0;
5712 if (rx_stats_active)
5713 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
5714 continue; /* Ignore this packet if it has been acknowledged */
5717 /* Turn off all flags except these ones, which are the same
5718 * on each transmission */
5719 p->header.flags &= RX_PRESET_FLAGS;
5721 if (p->header.seq >=
5722 call->tfirst + MIN((int)call->twind,
5723 (int)(call->nSoftAcked +
5725 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5726 /* Note: if we're waiting for more window space, we can
5727 * still send retransmits; hence we don't return here, but
5728 * break out to schedule a retransmit event */
5729 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
5730 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
5735 /* Transmit the packet if it needs to be sent. */
5736 if (!clock_Lt(&now, &p->retryTime)) {
5737 if (nXmitPackets == maxXmitPackets) {
5738 rxi_SendXmitList(call, call->xmitList,
5739 nXmitPackets, istack);
5742 dpf(("call %d xmit packet %"AFS_PTR_FMT" now %u.%06u retryTime %u.%06u\n",
5743 *(call->callNumber), p,
5745 p->retryTime.sec, p->retryTime.usec));
5746 call->xmitList[nXmitPackets++] = p;
5750 /* xmitList now hold pointers to all of the packets that are
5751 * ready to send. Now we loop to send the packets */
5752 if (nXmitPackets > 0) {
5753 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
5757 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5759 * TQ references no longer protected by this flag; they must remain
5760 * protected by the global lock.
5762 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5763 call->flags &= ~RX_CALL_TQ_BUSY;
5764 rxi_WakeUpTransmitQueue(call);
5768 /* We went into the error state while sending packets. Now is
5769 * the time to reset the call. This will also inform the using
5770 * process that the call is in an error state.
5772 if (rx_stats_active)
5773 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
5774 call->flags &= ~RX_CALL_TQ_BUSY;
5775 rxi_WakeUpTransmitQueue(call);
5776 rxi_CallError(call, call->error);
5779 #ifdef RX_ENABLE_LOCKS
5780 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5782 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5783 /* Some packets have received acks. If they all have, we can clear
5784 * the transmit queue.
5787 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5788 if (p->header.seq < call->tfirst
5789 && (p->flags & RX_PKTFLAG_ACKED)) {
5791 #ifdef RX_TRACK_PACKETS
5792 p->flags &= ~RX_PKTFLAG_TQ;
5794 #ifdef RXDEBUG_PACKET
5802 call->flags |= RX_CALL_TQ_CLEARME;
5804 #endif /* RX_ENABLE_LOCKS */
5805 /* Don't bother doing retransmits if the TQ is cleared. */
5806 if (call->flags & RX_CALL_TQ_CLEARME) {
5807 rxi_ClearTransmitQueue(call, 1);
5809 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5812 /* Always post a resend event, if there is anything in the
5813 * queue, and resend is possible. There should be at least
5814 * one unacknowledged packet in the queue ... otherwise none
5815 * of these packets should be on the queue in the first place.
5817 if (call->resendEvent) {
5818 /* Cancel the existing event and post a new one */
5819 rxevent_Cancel(call->resendEvent, call,
5820 RX_CALL_REFCOUNT_RESEND);
5823 /* The retry time is the retry time on the first unacknowledged
5824 * packet inside the current window */
5826 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5827 /* Don't set timers for packets outside the window */
5828 if (p->header.seq >= call->tfirst + call->twind) {
5832 if (!(p->flags & RX_PKTFLAG_ACKED)
5833 && !clock_IsZero(&p->retryTime)) {
5835 retryTime = p->retryTime;
5840 /* Post a new event to re-run rxi_Start when retries may be needed */
5841 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5842 #ifdef RX_ENABLE_LOCKS
5843 MUTEX_ENTER(&rx_refcnt_mutex);
5844 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5845 MUTEX_EXIT(&rx_refcnt_mutex);
5847 rxevent_PostNow2(&retryTime, &usenow,
5849 (void *)call, 0, istack);
5850 #else /* RX_ENABLE_LOCKS */
5852 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5853 (void *)call, 0, istack);
5854 #endif /* RX_ENABLE_LOCKS */
5857 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5858 } while (call->flags & RX_CALL_NEED_START);
5860 * TQ references no longer protected by this flag; they must remain
5861 * protected by the global lock.
5863 call->flags &= ~RX_CALL_TQ_BUSY;
5864 rxi_WakeUpTransmitQueue(call);
5866 call->flags |= RX_CALL_NEED_START;
5868 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5870 if (call->resendEvent) {
5871 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5876 /* Also adjusts the keep alive parameters for the call, to reflect
5877 * that we have just sent a packet (so keep alives aren't sent
5880 rxi_Send(struct rx_call *call, struct rx_packet *p,
5883 struct rx_connection *conn = call->conn;
5885 /* Stamp each packet with the user supplied status */
5886 p->header.userStatus = call->localStatus;
5888 /* Allow the security object controlling this call's security to
5889 * make any last-minute changes to the packet */
5890 RXS_SendPacket(conn->securityObject, call, p);
5892 /* Since we're about to send SOME sort of packet to the peer, it's
5893 * safe to nuke any scheduled end-of-packets ack */
5894 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5896 /* Actually send the packet, filling in more connection-specific fields */
5897 MUTEX_EXIT(&call->lock);
5898 MUTEX_ENTER(&rx_refcnt_mutex);
5899 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5900 MUTEX_EXIT(&rx_refcnt_mutex);
5901 rxi_SendPacket(call, conn, p, istack);
5902 MUTEX_ENTER(&rx_refcnt_mutex);
5903 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5904 MUTEX_EXIT(&rx_refcnt_mutex);
5905 MUTEX_ENTER(&call->lock);
5907 /* Update last send time for this call (for keep-alive
5908 * processing), and for the connection (so that we can discover
5909 * idle connections) */
5910 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5911 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
5912 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
5914 conn->lastSendTime = call->lastSendTime = clock_Sec();
5915 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
5916 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
5917 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
5918 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
5919 RX_ACK_PING_RESPONSE)))
5920 call->lastSendData = call->lastSendTime;
5924 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5925 * that things are fine. Also called periodically to guarantee that nothing
5926 * falls through the cracks (e.g. (error + dally) connections have keepalive
5927 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5929 * haveCTLock Set if calling from rxi_ReapConnections
5931 #ifdef RX_ENABLE_LOCKS
5933 rxi_CheckCall(struct rx_call *call, int haveCTLock)
5934 #else /* RX_ENABLE_LOCKS */
5936 rxi_CheckCall(struct rx_call *call)
5937 #endif /* RX_ENABLE_LOCKS */
5939 struct rx_connection *conn = call->conn;
5941 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
5942 afs_uint32 fudgeFactor;
5946 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5947 if (call->flags & RX_CALL_TQ_BUSY) {
5948 /* Call is active and will be reset by rxi_Start if it's
5949 * in an error state.
5954 /* RTT + 8*MDEV, rounded up to the next second. */
5955 fudgeFactor = (((afs_uint32) conn->peer->rtt >> 3) +
5956 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5958 deadTime = conn->secondsUntilDead + fudgeFactor;
5960 /* These are computed to the second (+- 1 second). But that's
5961 * good enough for these values, which should be a significant
5962 * number of seconds. */
5963 if (now > (call->lastReceiveTime + deadTime)) {
5964 if (call->state == RX_STATE_ACTIVE) {
5966 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5968 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5969 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5970 ip_stack_t *ipst = ns->netstack_ip;
5972 ire = ire_cache_lookup(conn->peer->host
5973 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5975 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5977 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5984 if (ire && ire->ire_max_frag > 0)
5985 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
5987 #if defined(GLOBAL_NETSTACKID)
5991 #endif /* ADAPT_PMTU */
5992 cerror = RX_CALL_DEAD;
5995 #ifdef RX_ENABLE_LOCKS
5996 /* Cancel pending events */
5997 rxevent_Cancel(call->delayedAckEvent, call,
5998 RX_CALL_REFCOUNT_DELAY);
5999 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
6000 rxevent_Cancel(call->keepAliveEvent, call,
6001 RX_CALL_REFCOUNT_ALIVE);
6002 MUTEX_ENTER(&rx_refcnt_mutex);
6003 if (call->refCount == 0) {
6004 rxi_FreeCall(call, haveCTLock);
6005 MUTEX_EXIT(&rx_refcnt_mutex);
6008 MUTEX_EXIT(&rx_refcnt_mutex);
6010 #else /* RX_ENABLE_LOCKS */
6011 rxi_FreeCall(call, 0);
6013 #endif /* RX_ENABLE_LOCKS */
6015 /* Non-active calls are destroyed if they are not responding
6016 * to pings; active calls are simply flagged in error, so the
6017 * attached process can die reasonably gracefully. */
6020 if (conn->idleDeadTime) {
6021 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6024 /* see if we have a non-activity timeout */
6025 if (call->startWait && idleDeadTime
6026 && ((call->startWait + idleDeadTime) < now) &&
6027 (call->flags & RX_CALL_READER_WAIT)) {
6028 if (call->state == RX_STATE_ACTIVE) {
6029 cerror = RX_CALL_TIMEOUT;
6033 if (call->lastSendData && idleDeadTime && (conn->idleDeadErr != 0)
6034 && ((call->lastSendData + idleDeadTime) < now)) {
6035 if (call->state == RX_STATE_ACTIVE) {
6036 cerror = conn->idleDeadErr;
6042 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6045 /* see if we have a hard timeout */
6047 && (now > (hardDeadTime + call->startTime.sec))) {
6048 if (call->state == RX_STATE_ACTIVE)
6049 rxi_CallError(call, RX_CALL_TIMEOUT);
6054 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT
6055 && call->lastReceiveTime) {
6056 int oldMTU = conn->peer->ifMTU;
6058 /* if we thought we could send more, perhaps things got worse */
6059 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6060 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6061 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6062 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6064 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6066 /* minimum capped in SetPeerMtu */
6067 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6070 conn->lastPacketSize = 0;
6072 /* needed so ResetCall doesn't clobber us. */
6073 call->MTU = conn->peer->ifMTU;
6075 /* if we never succeeded, let the error pass out as-is */
6076 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6077 cerror = conn->msgsizeRetryErr;
6080 rxi_CallError(call, cerror);
6085 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6087 struct rx_connection *conn = arg1;
6088 struct rx_header theader;
6090 struct sockaddr_in taddr;
6093 struct iovec tmpiov[2];
6096 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6099 tp = &tbuffer[sizeof(struct rx_header)];
6100 taddr.sin_family = AF_INET;
6101 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6102 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6103 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6104 taddr.sin_len = sizeof(struct sockaddr_in);
6106 memset(&theader, 0, sizeof(theader));
6107 theader.epoch = htonl(999);
6109 theader.callNumber = 0;
6112 theader.type = RX_PACKET_TYPE_VERSION;
6113 theader.flags = RX_LAST_PACKET;
6114 theader.serviceId = 0;
6116 memcpy(tbuffer, &theader, sizeof(theader));
6117 memcpy(tp, &a, sizeof(a));
6118 tmpiov[0].iov_base = tbuffer;
6119 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6121 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6123 MUTEX_ENTER(&conn->conn_data_lock);
6124 MUTEX_ENTER(&rx_refcnt_mutex);
6125 /* Only reschedule ourselves if the connection would not be destroyed */
6126 if (conn->refCount <= 1) {
6127 conn->natKeepAliveEvent = NULL;
6128 MUTEX_EXIT(&rx_refcnt_mutex);
6129 MUTEX_EXIT(&conn->conn_data_lock);
6130 rx_DestroyConnection(conn); /* drop the reference for this */
6132 conn->refCount--; /* drop the reference for this */
6133 MUTEX_EXIT(&rx_refcnt_mutex);
6134 conn->natKeepAliveEvent = NULL;
6135 rxi_ScheduleNatKeepAliveEvent(conn);
6136 MUTEX_EXIT(&conn->conn_data_lock);
6141 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6143 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6144 struct clock when, now;
6145 clock_GetTime(&now);
6147 when.sec += conn->secondsUntilNatPing;
6148 MUTEX_ENTER(&rx_refcnt_mutex);
6149 conn->refCount++; /* hold a reference for this */
6150 MUTEX_EXIT(&rx_refcnt_mutex);
6151 conn->natKeepAliveEvent =
6152 rxevent_PostNow(&when, &now, rxi_NatKeepAliveEvent, conn, 0);
6157 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6159 MUTEX_ENTER(&conn->conn_data_lock);
6160 conn->secondsUntilNatPing = seconds;
6162 rxi_ScheduleNatKeepAliveEvent(conn);
6163 MUTEX_EXIT(&conn->conn_data_lock);
6167 rxi_NatKeepAliveOn(struct rx_connection *conn)
6169 MUTEX_ENTER(&conn->conn_data_lock);
6170 rxi_ScheduleNatKeepAliveEvent(conn);
6171 MUTEX_EXIT(&conn->conn_data_lock);
6174 /* When a call is in progress, this routine is called occasionally to
6175 * make sure that some traffic has arrived (or been sent to) the peer.
6176 * If nothing has arrived in a reasonable amount of time, the call is
6177 * declared dead; if nothing has been sent for a while, we send a
6178 * keep-alive packet (if we're actually trying to keep the call alive)
6181 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6183 struct rx_call *call = arg1;
6184 struct rx_connection *conn;
6187 MUTEX_ENTER(&rx_refcnt_mutex);
6188 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6189 MUTEX_EXIT(&rx_refcnt_mutex);
6190 MUTEX_ENTER(&call->lock);
6191 if (event == call->keepAliveEvent)
6192 call->keepAliveEvent = NULL;
6195 #ifdef RX_ENABLE_LOCKS
6196 if (rxi_CheckCall(call, 0)) {
6197 MUTEX_EXIT(&call->lock);
6200 #else /* RX_ENABLE_LOCKS */
6201 if (rxi_CheckCall(call))
6203 #endif /* RX_ENABLE_LOCKS */
6205 /* Don't try to keep alive dallying calls */
6206 if (call->state == RX_STATE_DALLY) {
6207 MUTEX_EXIT(&call->lock);
6212 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6213 /* Don't try to send keepalives if there is unacknowledged data */
6214 /* the rexmit code should be good enough, this little hack
6215 * doesn't quite work XXX */
6216 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6218 rxi_ScheduleKeepAliveEvent(call);
6219 MUTEX_EXIT(&call->lock);
6222 /* Does what's on the nameplate. */
6224 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy)
6226 struct rx_call *call = arg1;
6227 struct rx_connection *conn;
6229 MUTEX_ENTER(&rx_refcnt_mutex);
6230 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6231 MUTEX_EXIT(&rx_refcnt_mutex);
6232 MUTEX_ENTER(&call->lock);
6234 if (event == call->growMTUEvent)
6235 call->growMTUEvent = NULL;
6237 #ifdef RX_ENABLE_LOCKS
6238 if (rxi_CheckCall(call, 0)) {
6239 MUTEX_EXIT(&call->lock);
6242 #else /* RX_ENABLE_LOCKS */
6243 if (rxi_CheckCall(call))
6245 #endif /* RX_ENABLE_LOCKS */
6247 /* Don't bother with dallying calls */
6248 if (call->state == RX_STATE_DALLY) {
6249 MUTEX_EXIT(&call->lock);
6256 * keep being scheduled, just don't do anything if we're at peak,
6257 * or we're not set up to be properly handled (idle timeout required)
6259 if ((conn->peer->maxPacketSize != 0) &&
6260 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6261 (conn->idleDeadErr))
6262 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6263 rxi_ScheduleGrowMTUEvent(call, 0);
6264 MUTEX_EXIT(&call->lock);
6268 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6270 if (!call->keepAliveEvent) {
6271 struct clock when, now;
6272 clock_GetTime(&now);
6274 when.sec += call->conn->secondsUntilPing;
6275 MUTEX_ENTER(&rx_refcnt_mutex);
6276 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6277 MUTEX_EXIT(&rx_refcnt_mutex);
6278 call->keepAliveEvent =
6279 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
6284 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6286 if (!call->growMTUEvent) {
6287 struct clock when, now;
6289 clock_GetTime(&now);
6292 if (call->conn->secondsUntilPing)
6293 secs = (6*call->conn->secondsUntilPing)-1;
6295 if (call->conn->secondsUntilDead)
6296 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6300 MUTEX_ENTER(&rx_refcnt_mutex);
6301 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6302 MUTEX_EXIT(&rx_refcnt_mutex);
6303 call->growMTUEvent =
6304 rxevent_PostNow(&when, &now, rxi_GrowMTUEvent, call, 0);
6308 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6310 rxi_KeepAliveOn(struct rx_call *call)
6312 /* Pretend last packet received was received now--i.e. if another
6313 * packet isn't received within the keep alive time, then the call
6314 * will die; Initialize last send time to the current time--even
6315 * if a packet hasn't been sent yet. This will guarantee that a
6316 * keep-alive is sent within the ping time */
6317 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6318 rxi_ScheduleKeepAliveEvent(call);
6322 rxi_GrowMTUOn(struct rx_call *call)
6324 struct rx_connection *conn = call->conn;
6325 MUTEX_ENTER(&conn->conn_data_lock);
6326 conn->lastPingSizeSer = conn->lastPingSize = 0;
6327 MUTEX_EXIT(&conn->conn_data_lock);
6328 rxi_ScheduleGrowMTUEvent(call, 1);
6331 /* This routine is called to send connection abort messages
6332 * that have been delayed to throttle looping clients. */
6334 rxi_SendDelayedConnAbort(struct rxevent *event,
6335 void *arg1, void *unused)
6337 struct rx_connection *conn = arg1;
6340 struct rx_packet *packet;
6342 MUTEX_ENTER(&conn->conn_data_lock);
6343 conn->delayedAbortEvent = NULL;
6344 error = htonl(conn->error);
6346 MUTEX_EXIT(&conn->conn_data_lock);
6347 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6350 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6351 RX_PACKET_TYPE_ABORT, (char *)&error,
6353 rxi_FreePacket(packet);
6357 /* This routine is called to send call abort messages
6358 * that have been delayed to throttle looping clients. */
6360 rxi_SendDelayedCallAbort(struct rxevent *event,
6361 void *arg1, void *dummy)
6363 struct rx_call *call = arg1;
6366 struct rx_packet *packet;
6368 MUTEX_ENTER(&call->lock);
6369 call->delayedAbortEvent = NULL;
6370 error = htonl(call->error);
6372 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6375 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6376 (char *)&error, sizeof(error), 0);
6377 rxi_FreePacket(packet);
6379 MUTEX_EXIT(&call->lock);
6380 MUTEX_ENTER(&rx_refcnt_mutex);
6381 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6382 MUTEX_EXIT(&rx_refcnt_mutex);
6385 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6386 * seconds) to ask the client to authenticate itself. The routine
6387 * issues a challenge to the client, which is obtained from the
6388 * security object associated with the connection */
6390 rxi_ChallengeEvent(struct rxevent *event,
6391 void *arg0, void *arg1, int tries)
6393 struct rx_connection *conn = arg0;
6395 conn->challengeEvent = NULL;
6396 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6397 struct rx_packet *packet;
6398 struct clock when, now;
6401 /* We've failed to authenticate for too long.
6402 * Reset any calls waiting for authentication;
6403 * they are all in RX_STATE_PRECALL.
6407 MUTEX_ENTER(&conn->conn_call_lock);
6408 for (i = 0; i < RX_MAXCALLS; i++) {
6409 struct rx_call *call = conn->call[i];
6411 MUTEX_ENTER(&call->lock);
6412 if (call->state == RX_STATE_PRECALL) {
6413 rxi_CallError(call, RX_CALL_DEAD);
6414 rxi_SendCallAbort(call, NULL, 0, 0);
6416 MUTEX_EXIT(&call->lock);
6419 MUTEX_EXIT(&conn->conn_call_lock);
6423 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6425 /* If there's no packet available, do this later. */
6426 RXS_GetChallenge(conn->securityObject, conn, packet);
6427 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6428 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6429 rxi_FreePacket(packet);
6431 clock_GetTime(&now);
6433 when.sec += RX_CHALLENGE_TIMEOUT;
6434 conn->challengeEvent =
6435 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
6440 /* Call this routine to start requesting the client to authenticate
6441 * itself. This will continue until authentication is established,
6442 * the call times out, or an invalid response is returned. The
6443 * security object associated with the connection is asked to create
6444 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6445 * defined earlier. */
6447 rxi_ChallengeOn(struct rx_connection *conn)
6449 if (!conn->challengeEvent) {
6450 RXS_CreateChallenge(conn->securityObject, conn);
6451 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6456 /* rxi_ComputeRoundTripTime is called with peer locked. */
6457 /* peer may be null */
6459 rxi_ComputeRoundTripTime(struct rx_packet *p,
6460 struct rx_ackPacket *ack,
6461 struct rx_peer *peer,
6464 struct clock thisRtt, *sentp;
6468 /* If the ACK is delayed, then do nothing */
6469 if (ack->reason == RX_ACK_DELAY)
6472 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6473 * their RTT multiple times, so only include the RTT of the last packet
6475 if (p->flags & RX_JUMBO_PACKET)
6478 /* Use the serial number to determine which transmission the ACK is for,
6479 * and set the sent time to match this. If we have no serial number, then
6480 * only use the ACK for RTT calculations if the packet has not been
6484 serial = ntohl(ack->serial);
6486 if (serial == p->header.serial) {
6487 sentp = &p->timeSent;
6488 } else if (serial == p->firstSerial) {
6489 sentp = &p->firstSent;
6490 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6491 sentp = &p->firstSent;
6495 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6496 sentp = &p->firstSent;
6503 if (clock_Lt(&thisRtt, sentp))
6504 return; /* somebody set the clock back, don't count this time. */
6506 clock_Sub(&thisRtt, sentp);
6507 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6508 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6510 if (clock_IsZero(&thisRtt)) {
6512 * The actual round trip time is shorter than the
6513 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6514 * Since we can't tell which at the moment we will assume 1ms.
6516 thisRtt.usec = 1000;
6519 if (rx_stats_active) {
6520 MUTEX_ENTER(&rx_stats_mutex);
6521 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6522 rx_stats.minRtt = thisRtt;
6523 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6524 if (thisRtt.sec > 60) {
6525 MUTEX_EXIT(&rx_stats_mutex);
6526 return; /* somebody set the clock ahead */
6528 rx_stats.maxRtt = thisRtt;
6530 clock_Add(&rx_stats.totalRtt, &thisRtt);
6531 rx_atomic_inc(&rx_stats.nRttSamples);
6532 MUTEX_EXIT(&rx_stats_mutex);
6535 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6537 /* Apply VanJacobson round-trip estimations */
6542 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
6543 * srtt is stored as fixed point with 3 bits after the binary
6544 * point (i.e., scaled by 8). The following magic is
6545 * equivalent to the smoothing algorithm in rfc793 with an
6546 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6547 * srtt'*8 = rtt + srtt*7
6548 * srtt'*8 = srtt*8 + rtt - srtt
6549 * srtt' = srtt + rtt/8 - srtt/8
6550 * srtt' = srtt + (rtt - srtt)/8
6553 delta = _8THMSEC(&thisRtt) - peer->rtt;
6554 peer->rtt += (delta >> 3);
6557 * We accumulate a smoothed rtt variance (actually, a smoothed
6558 * mean difference), then set the retransmit timer to smoothed
6559 * rtt + 4 times the smoothed variance (was 2x in van's original
6560 * paper, but 4x works better for me, and apparently for him as
6562 * rttvar is stored as
6563 * fixed point with 2 bits after the binary point (scaled by
6564 * 4). The following is equivalent to rfc793 smoothing with
6565 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6566 * rttvar'*4 = rttvar*3 + |delta|
6567 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6568 * rttvar' = rttvar + |delta|/4 - rttvar/4
6569 * rttvar' = rttvar + (|delta| - rttvar)/4
6570 * This replaces rfc793's wired-in beta.
6571 * dev*4 = dev*4 + (|actual - expected| - dev)
6577 delta -= (peer->rtt_dev << 1);
6578 peer->rtt_dev += (delta >> 3);
6580 /* I don't have a stored RTT so I start with this value. Since I'm
6581 * probably just starting a call, and will be pushing more data down
6582 * this, I expect congestion to increase rapidly. So I fudge a
6583 * little, and I set deviance to half the rtt. In practice,
6584 * deviance tends to approach something a little less than
6585 * half the smoothed rtt. */
6586 peer->rtt = _8THMSEC(&thisRtt) + 8;
6587 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
6589 /* the timeout is RTT + 4*MDEV + rx_minPeerTimeout msec.
6590 * This is because one end or the other of these connections is usually
6591 * in a user process, and can be switched and/or swapped out. So on fast,
6592 * reliable networks, the timeout would otherwise be too short. */
6593 rtt_timeout = ((peer->rtt >> 3) + peer->rtt_dev) + rx_minPeerTimeout;
6594 clock_Zero(&(peer->timeout));
6595 clock_Addmsec(&(peer->timeout), rtt_timeout);
6597 /* Reset the backedOff flag since we just computed a new timeout value */
6598 peer->backedOff = 0;
6600 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6601 p->header.callNumber, p, MSEC(&thisRtt), peer->rtt >> 3, peer->rtt_dev >> 2, (peer->timeout.sec), (peer->timeout.usec)));
6605 /* Find all server connections that have not been active for a long time, and
6608 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6610 struct clock now, when;
6611 clock_GetTime(&now);
6613 /* Find server connection structures that haven't been used for
6614 * greater than rx_idleConnectionTime */
6616 struct rx_connection **conn_ptr, **conn_end;
6617 int i, havecalls = 0;
6618 MUTEX_ENTER(&rx_connHashTable_lock);
6619 for (conn_ptr = &rx_connHashTable[0], conn_end =
6620 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6622 struct rx_connection *conn, *next;
6623 struct rx_call *call;
6627 for (conn = *conn_ptr; conn; conn = next) {
6628 /* XXX -- Shouldn't the connection be locked? */
6631 for (i = 0; i < RX_MAXCALLS; i++) {
6632 call = conn->call[i];
6636 code = MUTEX_TRYENTER(&call->lock);
6639 #ifdef RX_ENABLE_LOCKS
6640 result = rxi_CheckCall(call, 1);
6641 #else /* RX_ENABLE_LOCKS */
6642 result = rxi_CheckCall(call);
6643 #endif /* RX_ENABLE_LOCKS */
6644 MUTEX_EXIT(&call->lock);
6646 /* If CheckCall freed the call, it might
6647 * have destroyed the connection as well,
6648 * which screws up the linked lists.
6654 if (conn->type == RX_SERVER_CONNECTION) {
6655 /* This only actually destroys the connection if
6656 * there are no outstanding calls */
6657 MUTEX_ENTER(&conn->conn_data_lock);
6658 MUTEX_ENTER(&rx_refcnt_mutex);
6659 if (!havecalls && !conn->refCount
6660 && ((conn->lastSendTime + rx_idleConnectionTime) <
6662 conn->refCount++; /* it will be decr in rx_DestroyConn */
6663 MUTEX_EXIT(&rx_refcnt_mutex);
6664 MUTEX_EXIT(&conn->conn_data_lock);
6665 #ifdef RX_ENABLE_LOCKS
6666 rxi_DestroyConnectionNoLock(conn);
6667 #else /* RX_ENABLE_LOCKS */
6668 rxi_DestroyConnection(conn);
6669 #endif /* RX_ENABLE_LOCKS */
6671 #ifdef RX_ENABLE_LOCKS
6673 MUTEX_EXIT(&rx_refcnt_mutex);
6674 MUTEX_EXIT(&conn->conn_data_lock);
6676 #endif /* RX_ENABLE_LOCKS */
6680 #ifdef RX_ENABLE_LOCKS
6681 while (rx_connCleanup_list) {
6682 struct rx_connection *conn;
6683 conn = rx_connCleanup_list;
6684 rx_connCleanup_list = rx_connCleanup_list->next;
6685 MUTEX_EXIT(&rx_connHashTable_lock);
6686 rxi_CleanupConnection(conn);
6687 MUTEX_ENTER(&rx_connHashTable_lock);
6689 MUTEX_EXIT(&rx_connHashTable_lock);
6690 #endif /* RX_ENABLE_LOCKS */
6693 /* Find any peer structures that haven't been used (haven't had an
6694 * associated connection) for greater than rx_idlePeerTime */
6696 struct rx_peer **peer_ptr, **peer_end;
6700 * Why do we need to hold the rx_peerHashTable_lock across
6701 * the incrementing of peer_ptr since the rx_peerHashTable
6702 * array is not changing? We don't.
6704 * By dropping the lock periodically we can permit other
6705 * activities to be performed while a rxi_ReapConnections
6706 * call is in progress. The goal of reap connections
6707 * is to clean up quickly without causing large amounts
6708 * of contention. Therefore, it is important that global
6709 * mutexes not be held for extended periods of time.
6711 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6712 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6714 struct rx_peer *peer, *next, *prev;
6716 MUTEX_ENTER(&rx_peerHashTable_lock);
6717 for (prev = peer = *peer_ptr; peer; peer = next) {
6719 code = MUTEX_TRYENTER(&peer->peer_lock);
6720 if ((code) && (peer->refCount == 0)
6721 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6722 rx_interface_stat_p rpc_stat, nrpc_stat;
6726 * now know that this peer object is one to be
6727 * removed from the hash table. Once it is removed
6728 * it can't be referenced by other threads.
6729 * Lets remove it first and decrement the struct
6730 * nPeerStructs count.
6732 if (peer == *peer_ptr) {
6738 if (rx_stats_active)
6739 rx_atomic_dec(&rx_stats.nPeerStructs);
6742 * Now if we hold references on 'prev' and 'next'
6743 * we can safely drop the rx_peerHashTable_lock
6744 * while we destroy this 'peer' object.
6750 MUTEX_EXIT(&rx_peerHashTable_lock);
6752 MUTEX_EXIT(&peer->peer_lock);
6753 MUTEX_DESTROY(&peer->peer_lock);
6755 (&peer->rpcStats, rpc_stat, nrpc_stat,
6756 rx_interface_stat)) {
6757 unsigned int num_funcs;
6760 queue_Remove(&rpc_stat->queue_header);
6761 queue_Remove(&rpc_stat->all_peers);
6762 num_funcs = rpc_stat->stats[0].func_total;
6764 sizeof(rx_interface_stat_t) +
6765 rpc_stat->stats[0].func_total *
6766 sizeof(rx_function_entry_v1_t);
6768 rxi_Free(rpc_stat, space);
6770 MUTEX_ENTER(&rx_rpc_stats);
6771 rxi_rpc_peer_stat_cnt -= num_funcs;
6772 MUTEX_EXIT(&rx_rpc_stats);
6777 * Regain the rx_peerHashTable_lock and
6778 * decrement the reference count on 'prev'
6781 MUTEX_ENTER(&rx_peerHashTable_lock);
6788 MUTEX_EXIT(&peer->peer_lock);
6793 MUTEX_EXIT(&rx_peerHashTable_lock);
6797 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6798 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6799 * GC, just below. Really, we shouldn't have to keep moving packets from
6800 * one place to another, but instead ought to always know if we can
6801 * afford to hold onto a packet in its particular use. */
6802 MUTEX_ENTER(&rx_freePktQ_lock);
6803 if (rx_waitingForPackets) {
6804 rx_waitingForPackets = 0;
6805 #ifdef RX_ENABLE_LOCKS
6806 CV_BROADCAST(&rx_waitingForPackets_cv);
6808 osi_rxWakeup(&rx_waitingForPackets);
6811 MUTEX_EXIT(&rx_freePktQ_lock);
6814 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6815 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6819 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6820 * rx.h is sort of strange this is better. This is called with a security
6821 * object before it is discarded. Each connection using a security object has
6822 * its own refcount to the object so it won't actually be freed until the last
6823 * connection is destroyed.
6825 * This is the only rxs module call. A hold could also be written but no one
6829 rxs_Release(struct rx_securityClass *aobj)
6831 return RXS_Close(aobj);
6835 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6836 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6837 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6838 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6840 /* Adjust our estimate of the transmission rate to this peer, given
6841 * that the packet p was just acked. We can adjust peer->timeout and
6842 * call->twind. Pragmatically, this is called
6843 * only with packets of maximal length.
6844 * Called with peer and call locked.
6848 rxi_ComputeRate(struct rx_peer *peer, struct rx_call *call,
6849 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6851 afs_int32 xferSize, xferMs;
6855 /* Count down packets */
6856 if (peer->rateFlag > 0)
6858 /* Do nothing until we're enabled */
6859 if (peer->rateFlag != 0)
6864 /* Count only when the ack seems legitimate */
6865 switch (ackReason) {
6866 case RX_ACK_REQUESTED:
6868 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6872 case RX_ACK_PING_RESPONSE:
6873 if (p) /* want the response to ping-request, not data send */
6875 clock_GetTime(&newTO);
6876 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6877 clock_Sub(&newTO, &call->pingRequestTime);
6878 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6882 xferSize = rx_AckDataSize(rx_maxSendWindow) + RX_HEADER_SIZE;
6889 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %d.%06d, rtt %u, ps %u)\n",
6890 ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"),
6891 xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
6893 /* Track only packets that are big enough. */
6894 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6898 /* absorb RTT data (in milliseconds) for these big packets */
6899 if (peer->smRtt == 0) {
6900 peer->smRtt = xferMs;
6902 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6907 if (peer->countDown) {
6911 peer->countDown = 10; /* recalculate only every so often */
6913 /* In practice, we can measure only the RTT for full packets,
6914 * because of the way Rx acks the data that it receives. (If it's
6915 * smaller than a full packet, it often gets implicitly acked
6916 * either by the call response (from a server) or by the next call
6917 * (from a client), and either case confuses transmission times
6918 * with processing times.) Therefore, replace the above
6919 * more-sophisticated processing with a simpler version, where the
6920 * smoothed RTT is kept for full-size packets, and the time to
6921 * transmit a windowful of full-size packets is simply RTT *
6922 * windowSize. Again, we take two steps:
6923 - ensure the timeout is large enough for a single packet's RTT;
6924 - ensure that the window is small enough to fit in the desired timeout.*/
6926 /* First, the timeout check. */
6927 minTime = peer->smRtt;
6928 /* Get a reasonable estimate for a timeout period */
6930 newTO.sec = minTime / 1000;
6931 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6933 /* Increase the timeout period so that we can always do at least
6934 * one packet exchange */
6935 if (clock_Gt(&newTO, &peer->timeout)) {
6937 dpf(("CONG peer %lx/%u: timeout %d.%06d ==> %ld.%06d (rtt %u)\n",
6938 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec,
6939 newTO.sec, newTO.usec, peer->smRtt));
6941 peer->timeout = newTO;
6944 /* Now, get an estimate for the transmit window size. */
6945 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6946 /* Now, convert to the number of full packets that could fit in a
6947 * reasonable fraction of that interval */
6948 minTime /= (peer->smRtt << 1);
6949 minTime = MAX(minTime, rx_minPeerTimeout);
6950 xferSize = minTime; /* (make a copy) */
6952 /* Now clamp the size to reasonable bounds. */
6955 else if (minTime > rx_maxSendWindow)
6956 minTime = rx_maxSendWindow;
6957 /* if (minTime != peer->maxWindow) {
6958 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u)\n",
6959 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6960 peer->timeout.sec, peer->timeout.usec, peer->smRtt));
6961 peer->maxWindow = minTime;
6962 elide... call->twind = minTime;
6966 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6967 * Discern this by calculating the timeout necessary for rx_Window
6969 if ((xferSize > rx_maxSendWindow) && (peer->timeout.sec >= 3)) {
6970 /* calculate estimate for transmission interval in milliseconds */
6971 minTime = rx_maxSendWindow * peer->smRtt;
6972 if (minTime < 1000) {
6973 dpf(("CONG peer %lx/%u: cut TO %d.%06d by 0.5 (rtt %u)\n",
6974 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6975 peer->timeout.usec, peer->smRtt));
6977 newTO.sec = 0; /* cut back on timeout by half a second */
6978 newTO.usec = 500000;
6979 clock_Sub(&peer->timeout, &newTO);
6984 } /* end of rxi_ComputeRate */
6985 #endif /* ADAPT_WINDOW */
6993 #define TRACE_OPTION_RX_DEBUG 16
7001 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7002 0, KEY_QUERY_VALUE, &parmKey);
7003 if (code != ERROR_SUCCESS)
7006 dummyLen = sizeof(TraceOption);
7007 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7008 (BYTE *) &TraceOption, &dummyLen);
7009 if (code == ERROR_SUCCESS) {
7010 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7012 RegCloseKey (parmKey);
7013 #endif /* AFS_NT40_ENV */
7018 rx_DebugOnOff(int on)
7022 rxdebug_active = on;
7028 rx_StatsOnOff(int on)
7030 rx_stats_active = on;
7034 /* Don't call this debugging routine directly; use dpf */
7036 rxi_DebugPrint(char *format, ...)
7045 va_start(ap, format);
7047 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7050 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7052 OutputDebugString(msg);
7058 va_start(ap, format);
7060 clock_GetTime(&now);
7061 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7062 (unsigned int)now.usec);
7063 vfprintf(rx_Log, format, ap);
7071 * This function is used to process the rx_stats structure that is local
7072 * to a process as well as an rx_stats structure received from a remote
7073 * process (via rxdebug). Therefore, it needs to do minimal version
7077 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7078 afs_int32 freePackets, char version)
7082 if (size != sizeof(struct rx_statistics)) {
7084 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7085 size, sizeof(struct rx_statistics));
7088 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7091 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7092 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7093 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7094 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7095 s->specialPktAllocFailures);
7097 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7098 s->receivePktAllocFailures, s->sendPktAllocFailures,
7099 s->specialPktAllocFailures);
7103 " greedy %u, " "bogusReads %u (last from host %x), "
7104 "noPackets %u, " "noBuffers %u, " "selects %u, "
7105 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7106 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7107 s->selects, s->sendSelects);
7109 fprintf(file, " packets read: ");
7110 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7111 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7113 fprintf(file, "\n");
7116 " other read counters: data %u, " "ack %u, " "dup %u "
7117 "spurious %u " "dally %u\n", s->dataPacketsRead,
7118 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7119 s->ignorePacketDally);
7121 fprintf(file, " packets sent: ");
7122 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7123 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7125 fprintf(file, "\n");
7128 " other send counters: ack %u, " "data %u (not resends), "
7129 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7130 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7131 s->dataPacketsPushed, s->ignoreAckedPacket);
7134 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7135 s->netSendFailures, (int)s->fatalErrors);
7137 if (s->nRttSamples) {
7138 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7139 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7141 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7142 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7146 " %d server connections, " "%d client connections, "
7147 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7148 s->nServerConns, s->nClientConns, s->nPeerStructs,
7149 s->nCallStructs, s->nFreeCallStructs);
7151 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7152 fprintf(file, " %d clock updates\n", clock_nUpdates);
7156 /* for backward compatibility */
7158 rx_PrintStats(FILE * file)
7160 MUTEX_ENTER(&rx_stats_mutex);
7161 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7162 sizeof(rx_stats), rx_nFreePackets,
7164 MUTEX_EXIT(&rx_stats_mutex);
7168 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7170 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
7171 ntohl(peer->host), (int)ntohs(peer->port), (int)peer->burstSize,
7172 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
7175 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
7176 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
7177 (int)peer->timeout.usec, peer->nSent, peer->reSends);
7180 " Packet size %d, " "max in packet skew %d, "
7181 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
7182 (int)peer->outPacketSkew);
7186 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7188 * This mutex protects the following static variables:
7192 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7193 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7195 #define LOCK_RX_DEBUG
7196 #define UNLOCK_RX_DEBUG
7197 #endif /* AFS_PTHREAD_ENV */
7199 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7201 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7202 u_char type, void *inputData, size_t inputLength,
7203 void *outputData, size_t outputLength)
7205 static afs_int32 counter = 100;
7206 time_t waitTime, waitCount;
7207 struct rx_header theader;
7210 struct timeval tv_now, tv_wake, tv_delta;
7211 struct sockaddr_in taddr, faddr;
7225 tp = &tbuffer[sizeof(struct rx_header)];
7226 taddr.sin_family = AF_INET;
7227 taddr.sin_port = remotePort;
7228 taddr.sin_addr.s_addr = remoteAddr;
7229 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7230 taddr.sin_len = sizeof(struct sockaddr_in);
7233 memset(&theader, 0, sizeof(theader));
7234 theader.epoch = htonl(999);
7236 theader.callNumber = htonl(counter);
7239 theader.type = type;
7240 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7241 theader.serviceId = 0;
7243 memcpy(tbuffer, &theader, sizeof(theader));
7244 memcpy(tp, inputData, inputLength);
7246 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7247 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7249 /* see if there's a packet available */
7250 gettimeofday(&tv_wake,0);
7251 tv_wake.tv_sec += waitTime;
7254 FD_SET(socket, &imask);
7255 tv_delta.tv_sec = tv_wake.tv_sec;
7256 tv_delta.tv_usec = tv_wake.tv_usec;
7257 gettimeofday(&tv_now, 0);
7259 if (tv_delta.tv_usec < tv_now.tv_usec) {
7261 tv_delta.tv_usec += 1000000;
7264 tv_delta.tv_usec -= tv_now.tv_usec;
7266 if (tv_delta.tv_sec < tv_now.tv_sec) {
7270 tv_delta.tv_sec -= tv_now.tv_sec;
7273 code = select(0, &imask, 0, 0, &tv_delta);
7274 #else /* AFS_NT40_ENV */
7275 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7276 #endif /* AFS_NT40_ENV */
7277 if (code == 1 && FD_ISSET(socket, &imask)) {
7278 /* now receive a packet */
7279 faddrLen = sizeof(struct sockaddr_in);
7281 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7282 (struct sockaddr *)&faddr, &faddrLen);
7285 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7286 if (counter == ntohl(theader.callNumber))
7294 /* see if we've timed out */
7302 code -= sizeof(struct rx_header);
7303 if (code > outputLength)
7304 code = outputLength;
7305 memcpy(outputData, tp, code);
7308 #endif /* RXDEBUG */
7311 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7312 afs_uint16 remotePort, struct rx_debugStats * stat,
7313 afs_uint32 * supportedValues)
7315 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7317 struct rx_debugIn in;
7319 *supportedValues = 0;
7320 in.type = htonl(RX_DEBUGI_GETSTATS);
7323 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7324 &in, sizeof(in), stat, sizeof(*stat));
7327 * If the call was successful, fixup the version and indicate
7328 * what contents of the stat structure are valid.
7329 * Also do net to host conversion of fields here.
7333 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7334 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7336 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7337 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7339 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7340 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7342 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7343 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7345 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7346 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7348 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7349 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7351 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7352 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7354 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7355 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7357 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7358 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7360 stat->nFreePackets = ntohl(stat->nFreePackets);
7361 stat->packetReclaims = ntohl(stat->packetReclaims);
7362 stat->callsExecuted = ntohl(stat->callsExecuted);
7363 stat->nWaiting = ntohl(stat->nWaiting);
7364 stat->idleThreads = ntohl(stat->idleThreads);
7365 stat->nWaited = ntohl(stat->nWaited);
7366 stat->nPackets = ntohl(stat->nPackets);
7375 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7376 afs_uint16 remotePort, struct rx_statistics * stat,
7377 afs_uint32 * supportedValues)
7379 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7381 struct rx_debugIn in;
7382 afs_int32 *lp = (afs_int32 *) stat;
7386 * supportedValues is currently unused, but added to allow future
7387 * versioning of this function.
7390 *supportedValues = 0;
7391 in.type = htonl(RX_DEBUGI_RXSTATS);
7393 memset(stat, 0, sizeof(*stat));
7395 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7396 &in, sizeof(in), stat, sizeof(*stat));
7401 * Do net to host conversion here
7404 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7415 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7416 afs_uint16 remotePort, size_t version_length,
7419 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7421 return MakeDebugCall(socket, remoteAddr, remotePort,
7422 RX_PACKET_TYPE_VERSION, a, 1, version,
7430 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7431 afs_uint16 remotePort, afs_int32 * nextConnection,
7432 int allConnections, afs_uint32 debugSupportedValues,
7433 struct rx_debugConn * conn,
7434 afs_uint32 * supportedValues)
7436 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7438 struct rx_debugIn in;
7442 * supportedValues is currently unused, but added to allow future
7443 * versioning of this function.
7446 *supportedValues = 0;
7447 if (allConnections) {
7448 in.type = htonl(RX_DEBUGI_GETALLCONN);
7450 in.type = htonl(RX_DEBUGI_GETCONN);
7452 in.index = htonl(*nextConnection);
7453 memset(conn, 0, sizeof(*conn));
7455 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7456 &in, sizeof(in), conn, sizeof(*conn));
7459 *nextConnection += 1;
7462 * Convert old connection format to new structure.
7465 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7466 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7467 #define MOVEvL(a) (conn->a = vL->a)
7469 /* any old or unrecognized version... */
7470 for (i = 0; i < RX_MAXCALLS; i++) {
7471 MOVEvL(callState[i]);
7472 MOVEvL(callMode[i]);
7473 MOVEvL(callFlags[i]);
7474 MOVEvL(callOther[i]);
7476 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7477 MOVEvL(secStats.type);
7478 MOVEvL(secStats.level);
7479 MOVEvL(secStats.flags);
7480 MOVEvL(secStats.expires);
7481 MOVEvL(secStats.packetsReceived);
7482 MOVEvL(secStats.packetsSent);
7483 MOVEvL(secStats.bytesReceived);
7484 MOVEvL(secStats.bytesSent);
7489 * Do net to host conversion here
7491 * I don't convert host or port since we are most likely
7492 * going to want these in NBO.
7494 conn->cid = ntohl(conn->cid);
7495 conn->serial = ntohl(conn->serial);
7496 for (i = 0; i < RX_MAXCALLS; i++) {
7497 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7499 conn->error = ntohl(conn->error);
7500 conn->secStats.flags = ntohl(conn->secStats.flags);
7501 conn->secStats.expires = ntohl(conn->secStats.expires);
7502 conn->secStats.packetsReceived =
7503 ntohl(conn->secStats.packetsReceived);
7504 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7505 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7506 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7507 conn->epoch = ntohl(conn->epoch);
7508 conn->natMTU = ntohl(conn->natMTU);
7517 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7518 afs_uint16 remotePort, afs_int32 * nextPeer,
7519 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7520 afs_uint32 * supportedValues)
7522 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7524 struct rx_debugIn in;
7527 * supportedValues is currently unused, but added to allow future
7528 * versioning of this function.
7531 *supportedValues = 0;
7532 in.type = htonl(RX_DEBUGI_GETPEER);
7533 in.index = htonl(*nextPeer);
7534 memset(peer, 0, sizeof(*peer));
7536 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7537 &in, sizeof(in), peer, sizeof(*peer));
7543 * Do net to host conversion here
7545 * I don't convert host or port since we are most likely
7546 * going to want these in NBO.
7548 peer->ifMTU = ntohs(peer->ifMTU);
7549 peer->idleWhen = ntohl(peer->idleWhen);
7550 peer->refCount = ntohs(peer->refCount);
7551 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7552 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7553 peer->rtt = ntohl(peer->rtt);
7554 peer->rtt_dev = ntohl(peer->rtt_dev);
7555 peer->timeout.sec = ntohl(peer->timeout.sec);
7556 peer->timeout.usec = ntohl(peer->timeout.usec);
7557 peer->nSent = ntohl(peer->nSent);
7558 peer->reSends = ntohl(peer->reSends);
7559 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7560 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7561 peer->rateFlag = ntohl(peer->rateFlag);
7562 peer->natMTU = ntohs(peer->natMTU);
7563 peer->maxMTU = ntohs(peer->maxMTU);
7564 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7565 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7566 peer->MTU = ntohs(peer->MTU);
7567 peer->cwind = ntohs(peer->cwind);
7568 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7569 peer->congestSeq = ntohs(peer->congestSeq);
7570 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7571 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7572 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7573 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7582 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7583 struct rx_debugPeer * peerStats)
7586 afs_int32 error = 1; /* default to "did not succeed" */
7587 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7589 MUTEX_ENTER(&rx_peerHashTable_lock);
7590 for(tp = rx_peerHashTable[hashValue];
7591 tp != NULL; tp = tp->next) {
7592 if (tp->host == peerHost)
7598 MUTEX_EXIT(&rx_peerHashTable_lock);
7602 MUTEX_ENTER(&tp->peer_lock);
7603 peerStats->host = tp->host;
7604 peerStats->port = tp->port;
7605 peerStats->ifMTU = tp->ifMTU;
7606 peerStats->idleWhen = tp->idleWhen;
7607 peerStats->refCount = tp->refCount;
7608 peerStats->burstSize = tp->burstSize;
7609 peerStats->burst = tp->burst;
7610 peerStats->burstWait.sec = tp->burstWait.sec;
7611 peerStats->burstWait.usec = tp->burstWait.usec;
7612 peerStats->rtt = tp->rtt;
7613 peerStats->rtt_dev = tp->rtt_dev;
7614 peerStats->timeout.sec = tp->timeout.sec;
7615 peerStats->timeout.usec = tp->timeout.usec;
7616 peerStats->nSent = tp->nSent;
7617 peerStats->reSends = tp->reSends;
7618 peerStats->inPacketSkew = tp->inPacketSkew;
7619 peerStats->outPacketSkew = tp->outPacketSkew;
7620 peerStats->rateFlag = tp->rateFlag;
7621 peerStats->natMTU = tp->natMTU;
7622 peerStats->maxMTU = tp->maxMTU;
7623 peerStats->maxDgramPackets = tp->maxDgramPackets;
7624 peerStats->ifDgramPackets = tp->ifDgramPackets;
7625 peerStats->MTU = tp->MTU;
7626 peerStats->cwind = tp->cwind;
7627 peerStats->nDgramPackets = tp->nDgramPackets;
7628 peerStats->congestSeq = tp->congestSeq;
7629 peerStats->bytesSent.high = tp->bytesSent.high;
7630 peerStats->bytesSent.low = tp->bytesSent.low;
7631 peerStats->bytesReceived.high = tp->bytesReceived.high;
7632 peerStats->bytesReceived.low = tp->bytesReceived.low;
7633 MUTEX_EXIT(&tp->peer_lock);
7635 MUTEX_ENTER(&rx_peerHashTable_lock);
7638 MUTEX_EXIT(&rx_peerHashTable_lock);
7646 struct rx_serverQueueEntry *np;
7649 struct rx_call *call;
7650 struct rx_serverQueueEntry *sq;
7654 if (rxinit_status == 1) {
7656 return; /* Already shutdown. */
7660 #ifndef AFS_PTHREAD_ENV
7661 FD_ZERO(&rx_selectMask);
7662 #endif /* AFS_PTHREAD_ENV */
7663 rxi_dataQuota = RX_MAX_QUOTA;
7664 #ifndef AFS_PTHREAD_ENV
7666 #endif /* AFS_PTHREAD_ENV */
7669 #ifndef AFS_PTHREAD_ENV
7670 #ifndef AFS_USE_GETTIMEOFDAY
7672 #endif /* AFS_USE_GETTIMEOFDAY */
7673 #endif /* AFS_PTHREAD_ENV */
7675 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7676 call = queue_First(&rx_freeCallQueue, rx_call);
7678 rxi_Free(call, sizeof(struct rx_call));
7681 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7682 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7688 struct rx_peer **peer_ptr, **peer_end;
7689 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7690 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7692 struct rx_peer *peer, *next;
7694 MUTEX_ENTER(&rx_peerHashTable_lock);
7695 for (peer = *peer_ptr; peer; peer = next) {
7696 rx_interface_stat_p rpc_stat, nrpc_stat;
7699 MUTEX_ENTER(&rx_rpc_stats);
7700 MUTEX_ENTER(&peer->peer_lock);
7702 (&peer->rpcStats, rpc_stat, nrpc_stat,
7703 rx_interface_stat)) {
7704 unsigned int num_funcs;
7707 queue_Remove(&rpc_stat->queue_header);
7708 queue_Remove(&rpc_stat->all_peers);
7709 num_funcs = rpc_stat->stats[0].func_total;
7711 sizeof(rx_interface_stat_t) +
7712 rpc_stat->stats[0].func_total *
7713 sizeof(rx_function_entry_v1_t);
7715 rxi_Free(rpc_stat, space);
7717 /* rx_rpc_stats must be held */
7718 rxi_rpc_peer_stat_cnt -= num_funcs;
7720 MUTEX_EXIT(&peer->peer_lock);
7721 MUTEX_EXIT(&rx_rpc_stats);
7725 if (rx_stats_active)
7726 rx_atomic_dec(&rx_stats.nPeerStructs);
7728 MUTEX_EXIT(&rx_peerHashTable_lock);
7731 for (i = 0; i < RX_MAX_SERVICES; i++) {
7733 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7735 for (i = 0; i < rx_hashTableSize; i++) {
7736 struct rx_connection *tc, *ntc;
7737 MUTEX_ENTER(&rx_connHashTable_lock);
7738 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7740 for (j = 0; j < RX_MAXCALLS; j++) {
7742 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7745 rxi_Free(tc, sizeof(*tc));
7747 MUTEX_EXIT(&rx_connHashTable_lock);
7750 MUTEX_ENTER(&freeSQEList_lock);
7752 while ((np = rx_FreeSQEList)) {
7753 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7754 MUTEX_DESTROY(&np->lock);
7755 rxi_Free(np, sizeof(*np));
7758 MUTEX_EXIT(&freeSQEList_lock);
7759 MUTEX_DESTROY(&freeSQEList_lock);
7760 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7761 MUTEX_DESTROY(&rx_connHashTable_lock);
7762 MUTEX_DESTROY(&rx_peerHashTable_lock);
7763 MUTEX_DESTROY(&rx_serverPool_lock);
7765 osi_Free(rx_connHashTable,
7766 rx_hashTableSize * sizeof(struct rx_connection *));
7767 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7769 UNPIN(rx_connHashTable,
7770 rx_hashTableSize * sizeof(struct rx_connection *));
7771 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7773 rxi_FreeAllPackets();
7775 MUTEX_ENTER(&rx_quota_mutex);
7776 rxi_dataQuota = RX_MAX_QUOTA;
7777 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7778 MUTEX_EXIT(&rx_quota_mutex);
7783 #ifdef RX_ENABLE_LOCKS
7785 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7787 if (!MUTEX_ISMINE(lockaddr))
7788 osi_Panic("Lock not held: %s", msg);
7790 #endif /* RX_ENABLE_LOCKS */
7795 * Routines to implement connection specific data.
7799 rx_KeyCreate(rx_destructor_t rtn)
7802 MUTEX_ENTER(&rxi_keyCreate_lock);
7803 key = rxi_keyCreate_counter++;
7804 rxi_keyCreate_destructor = (rx_destructor_t *)
7805 realloc((void *)rxi_keyCreate_destructor,
7806 (key + 1) * sizeof(rx_destructor_t));
7807 rxi_keyCreate_destructor[key] = rtn;
7808 MUTEX_EXIT(&rxi_keyCreate_lock);
7813 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7816 MUTEX_ENTER(&conn->conn_data_lock);
7817 if (!conn->specific) {
7818 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7819 for (i = 0; i < key; i++)
7820 conn->specific[i] = NULL;
7821 conn->nSpecific = key + 1;
7822 conn->specific[key] = ptr;
7823 } else if (key >= conn->nSpecific) {
7824 conn->specific = (void **)
7825 realloc(conn->specific, (key + 1) * sizeof(void *));
7826 for (i = conn->nSpecific; i < key; i++)
7827 conn->specific[i] = NULL;
7828 conn->nSpecific = key + 1;
7829 conn->specific[key] = ptr;
7831 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7832 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7833 conn->specific[key] = ptr;
7835 MUTEX_EXIT(&conn->conn_data_lock);
7839 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7842 MUTEX_ENTER(&svc->svc_data_lock);
7843 if (!svc->specific) {
7844 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
7845 for (i = 0; i < key; i++)
7846 svc->specific[i] = NULL;
7847 svc->nSpecific = key + 1;
7848 svc->specific[key] = ptr;
7849 } else if (key >= svc->nSpecific) {
7850 svc->specific = (void **)
7851 realloc(svc->specific, (key + 1) * sizeof(void *));
7852 for (i = svc->nSpecific; i < key; i++)
7853 svc->specific[i] = NULL;
7854 svc->nSpecific = key + 1;
7855 svc->specific[key] = ptr;
7857 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7858 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7859 svc->specific[key] = ptr;
7861 MUTEX_EXIT(&svc->svc_data_lock);
7865 rx_GetSpecific(struct rx_connection *conn, int key)
7868 MUTEX_ENTER(&conn->conn_data_lock);
7869 if (key >= conn->nSpecific)
7872 ptr = conn->specific[key];
7873 MUTEX_EXIT(&conn->conn_data_lock);
7878 rx_GetServiceSpecific(struct rx_service *svc, int key)
7881 MUTEX_ENTER(&svc->svc_data_lock);
7882 if (key >= svc->nSpecific)
7885 ptr = svc->specific[key];
7886 MUTEX_EXIT(&svc->svc_data_lock);
7891 #endif /* !KERNEL */
7894 * processStats is a queue used to store the statistics for the local
7895 * process. Its contents are similar to the contents of the rpcStats
7896 * queue on a rx_peer structure, but the actual data stored within
7897 * this queue contains totals across the lifetime of the process (assuming
7898 * the stats have not been reset) - unlike the per peer structures
7899 * which can come and go based upon the peer lifetime.
7902 static struct rx_queue processStats = { &processStats, &processStats };
7905 * peerStats is a queue used to store the statistics for all peer structs.
7906 * Its contents are the union of all the peer rpcStats queues.
7909 static struct rx_queue peerStats = { &peerStats, &peerStats };
7912 * rxi_monitor_processStats is used to turn process wide stat collection
7916 static int rxi_monitor_processStats = 0;
7919 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7922 static int rxi_monitor_peerStats = 0;
7925 * rxi_AddRpcStat - given all of the information for a particular rpc
7926 * call, create (if needed) and update the stat totals for the rpc.
7930 * IN stats - the queue of stats that will be updated with the new value
7932 * IN rxInterface - a unique number that identifies the rpc interface
7934 * IN currentFunc - the index of the function being invoked
7936 * IN totalFunc - the total number of functions in this interface
7938 * IN queueTime - the amount of time this function waited for a thread
7940 * IN execTime - the amount of time this function invocation took to execute
7942 * IN bytesSent - the number bytes sent by this invocation
7944 * IN bytesRcvd - the number bytes received by this invocation
7946 * IN isServer - if true, this invocation was made to a server
7948 * IN remoteHost - the ip address of the remote host
7950 * IN remotePort - the port of the remote host
7952 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7954 * INOUT counter - if a new stats structure is allocated, the counter will
7955 * be updated with the new number of allocated stat structures
7963 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
7964 afs_uint32 currentFunc, afs_uint32 totalFunc,
7965 struct clock *queueTime, struct clock *execTime,
7966 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
7967 afs_uint32 remoteHost, afs_uint32 remotePort,
7968 int addToPeerList, unsigned int *counter)
7971 rx_interface_stat_p rpc_stat, nrpc_stat;
7974 * See if there's already a structure for this interface
7977 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7978 if ((rpc_stat->stats[0].interfaceId == rxInterface)
7979 && (rpc_stat->stats[0].remote_is_server == isServer))
7984 * Didn't find a match so allocate a new structure and add it to the
7988 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
7989 || (rpc_stat->stats[0].interfaceId != rxInterface)
7990 || (rpc_stat->stats[0].remote_is_server != isServer)) {
7995 sizeof(rx_interface_stat_t) +
7996 totalFunc * sizeof(rx_function_entry_v1_t);
7998 rpc_stat = rxi_Alloc(space);
7999 if (rpc_stat == NULL) {
8003 *counter += totalFunc;
8004 for (i = 0; i < totalFunc; i++) {
8005 rpc_stat->stats[i].remote_peer = remoteHost;
8006 rpc_stat->stats[i].remote_port = remotePort;
8007 rpc_stat->stats[i].remote_is_server = isServer;
8008 rpc_stat->stats[i].interfaceId = rxInterface;
8009 rpc_stat->stats[i].func_total = totalFunc;
8010 rpc_stat->stats[i].func_index = i;
8011 hzero(rpc_stat->stats[i].invocations);
8012 hzero(rpc_stat->stats[i].bytes_sent);
8013 hzero(rpc_stat->stats[i].bytes_rcvd);
8014 rpc_stat->stats[i].queue_time_sum.sec = 0;
8015 rpc_stat->stats[i].queue_time_sum.usec = 0;
8016 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8017 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8018 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8019 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8020 rpc_stat->stats[i].queue_time_max.sec = 0;
8021 rpc_stat->stats[i].queue_time_max.usec = 0;
8022 rpc_stat->stats[i].execution_time_sum.sec = 0;
8023 rpc_stat->stats[i].execution_time_sum.usec = 0;
8024 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8025 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8026 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8027 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8028 rpc_stat->stats[i].execution_time_max.sec = 0;
8029 rpc_stat->stats[i].execution_time_max.usec = 0;
8031 queue_Prepend(stats, rpc_stat);
8032 if (addToPeerList) {
8033 queue_Prepend(&peerStats, &rpc_stat->all_peers);
8038 * Increment the stats for this function
8041 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
8042 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
8043 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
8044 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8045 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8046 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8047 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8049 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8050 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8052 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8053 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8055 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8056 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8058 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8059 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8067 * rx_IncrementTimeAndCount - increment the times and count for a particular
8072 * IN peer - the peer who invoked the rpc
8074 * IN rxInterface - a unique number that identifies the rpc interface
8076 * IN currentFunc - the index of the function being invoked
8078 * IN totalFunc - the total number of functions in this interface
8080 * IN queueTime - the amount of time this function waited for a thread
8082 * IN execTime - the amount of time this function invocation took to execute
8084 * IN bytesSent - the number bytes sent by this invocation
8086 * IN bytesRcvd - the number bytes received by this invocation
8088 * IN isServer - if true, this invocation was made to a server
8096 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8097 afs_uint32 currentFunc, afs_uint32 totalFunc,
8098 struct clock *queueTime, struct clock *execTime,
8099 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8103 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8106 MUTEX_ENTER(&rx_rpc_stats);
8108 if (rxi_monitor_peerStats) {
8109 MUTEX_ENTER(&peer->peer_lock);
8110 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8111 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8112 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8113 MUTEX_EXIT(&peer->peer_lock);
8116 if (rxi_monitor_processStats) {
8117 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8118 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8119 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8122 MUTEX_EXIT(&rx_rpc_stats);
8127 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8131 * IN callerVersion - the rpc stat version of the caller.
8133 * IN count - the number of entries to marshall.
8135 * IN stats - pointer to stats to be marshalled.
8137 * OUT ptr - Where to store the marshalled data.
8144 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8145 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8151 * We only support the first version
8153 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8154 *(ptr++) = stats->remote_peer;
8155 *(ptr++) = stats->remote_port;
8156 *(ptr++) = stats->remote_is_server;
8157 *(ptr++) = stats->interfaceId;
8158 *(ptr++) = stats->func_total;
8159 *(ptr++) = stats->func_index;
8160 *(ptr++) = hgethi(stats->invocations);
8161 *(ptr++) = hgetlo(stats->invocations);
8162 *(ptr++) = hgethi(stats->bytes_sent);
8163 *(ptr++) = hgetlo(stats->bytes_sent);
8164 *(ptr++) = hgethi(stats->bytes_rcvd);
8165 *(ptr++) = hgetlo(stats->bytes_rcvd);
8166 *(ptr++) = stats->queue_time_sum.sec;
8167 *(ptr++) = stats->queue_time_sum.usec;
8168 *(ptr++) = stats->queue_time_sum_sqr.sec;
8169 *(ptr++) = stats->queue_time_sum_sqr.usec;
8170 *(ptr++) = stats->queue_time_min.sec;
8171 *(ptr++) = stats->queue_time_min.usec;
8172 *(ptr++) = stats->queue_time_max.sec;
8173 *(ptr++) = stats->queue_time_max.usec;
8174 *(ptr++) = stats->execution_time_sum.sec;
8175 *(ptr++) = stats->execution_time_sum.usec;
8176 *(ptr++) = stats->execution_time_sum_sqr.sec;
8177 *(ptr++) = stats->execution_time_sum_sqr.usec;
8178 *(ptr++) = stats->execution_time_min.sec;
8179 *(ptr++) = stats->execution_time_min.usec;
8180 *(ptr++) = stats->execution_time_max.sec;
8181 *(ptr++) = stats->execution_time_max.usec;
8187 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8192 * IN callerVersion - the rpc stat version of the caller
8194 * OUT myVersion - the rpc stat version of this function
8196 * OUT clock_sec - local time seconds
8198 * OUT clock_usec - local time microseconds
8200 * OUT allocSize - the number of bytes allocated to contain stats
8202 * OUT statCount - the number stats retrieved from this process.
8204 * OUT stats - the actual stats retrieved from this process.
8208 * Returns void. If successful, stats will != NULL.
8212 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8213 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8214 size_t * allocSize, afs_uint32 * statCount,
8215 afs_uint32 ** stats)
8225 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8228 * Check to see if stats are enabled
8231 MUTEX_ENTER(&rx_rpc_stats);
8232 if (!rxi_monitor_processStats) {
8233 MUTEX_EXIT(&rx_rpc_stats);
8237 clock_GetTime(&now);
8238 *clock_sec = now.sec;
8239 *clock_usec = now.usec;
8242 * Allocate the space based upon the caller version
8244 * If the client is at an older version than we are,
8245 * we return the statistic data in the older data format, but
8246 * we still return our version number so the client knows we
8247 * are maintaining more data than it can retrieve.
8250 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8251 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8252 *statCount = rxi_rpc_process_stat_cnt;
8255 * This can't happen yet, but in the future version changes
8256 * can be handled by adding additional code here
8260 if (space > (size_t) 0) {
8262 ptr = *stats = rxi_Alloc(space);
8265 rx_interface_stat_p rpc_stat, nrpc_stat;
8269 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8271 * Copy the data based upon the caller version
8273 rx_MarshallProcessRPCStats(callerVersion,
8274 rpc_stat->stats[0].func_total,
8275 rpc_stat->stats, &ptr);
8281 MUTEX_EXIT(&rx_rpc_stats);
8286 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8290 * IN callerVersion - the rpc stat version of the caller
8292 * OUT myVersion - the rpc stat version of this function
8294 * OUT clock_sec - local time seconds
8296 * OUT clock_usec - local time microseconds
8298 * OUT allocSize - the number of bytes allocated to contain stats
8300 * OUT statCount - the number of stats retrieved from the individual
8303 * OUT stats - the actual stats retrieved from the individual peer structures.
8307 * Returns void. If successful, stats will != NULL.
8311 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8312 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8313 size_t * allocSize, afs_uint32 * statCount,
8314 afs_uint32 ** stats)
8324 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8327 * Check to see if stats are enabled
8330 MUTEX_ENTER(&rx_rpc_stats);
8331 if (!rxi_monitor_peerStats) {
8332 MUTEX_EXIT(&rx_rpc_stats);
8336 clock_GetTime(&now);
8337 *clock_sec = now.sec;
8338 *clock_usec = now.usec;
8341 * Allocate the space based upon the caller version
8343 * If the client is at an older version than we are,
8344 * we return the statistic data in the older data format, but
8345 * we still return our version number so the client knows we
8346 * are maintaining more data than it can retrieve.
8349 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8350 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8351 *statCount = rxi_rpc_peer_stat_cnt;
8354 * This can't happen yet, but in the future version changes
8355 * can be handled by adding additional code here
8359 if (space > (size_t) 0) {
8361 ptr = *stats = rxi_Alloc(space);
8364 rx_interface_stat_p rpc_stat, nrpc_stat;
8368 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8370 * We have to fix the offset of rpc_stat since we are
8371 * keeping this structure on two rx_queues. The rx_queue
8372 * package assumes that the rx_queue member is the first
8373 * member of the structure. That is, rx_queue assumes that
8374 * any one item is only on one queue at a time. We are
8375 * breaking that assumption and so we have to do a little
8376 * math to fix our pointers.
8379 fix_offset = (char *)rpc_stat;
8380 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8381 rpc_stat = (rx_interface_stat_p) fix_offset;
8384 * Copy the data based upon the caller version
8386 rx_MarshallProcessRPCStats(callerVersion,
8387 rpc_stat->stats[0].func_total,
8388 rpc_stat->stats, &ptr);
8394 MUTEX_EXIT(&rx_rpc_stats);
8399 * rx_FreeRPCStats - free memory allocated by
8400 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8404 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8405 * rx_RetrievePeerRPCStats
8407 * IN allocSize - the number of bytes in stats.
8415 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8417 rxi_Free(stats, allocSize);
8421 * rx_queryProcessRPCStats - see if process rpc stat collection is
8422 * currently enabled.
8428 * Returns 0 if stats are not enabled != 0 otherwise
8432 rx_queryProcessRPCStats(void)
8435 MUTEX_ENTER(&rx_rpc_stats);
8436 rc = rxi_monitor_processStats;
8437 MUTEX_EXIT(&rx_rpc_stats);
8442 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8448 * Returns 0 if stats are not enabled != 0 otherwise
8452 rx_queryPeerRPCStats(void)
8455 MUTEX_ENTER(&rx_rpc_stats);
8456 rc = rxi_monitor_peerStats;
8457 MUTEX_EXIT(&rx_rpc_stats);
8462 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8472 rx_enableProcessRPCStats(void)
8474 MUTEX_ENTER(&rx_rpc_stats);
8475 rx_enable_stats = 1;
8476 rxi_monitor_processStats = 1;
8477 MUTEX_EXIT(&rx_rpc_stats);
8481 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8491 rx_enablePeerRPCStats(void)
8493 MUTEX_ENTER(&rx_rpc_stats);
8494 rx_enable_stats = 1;
8495 rxi_monitor_peerStats = 1;
8496 MUTEX_EXIT(&rx_rpc_stats);
8500 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8510 rx_disableProcessRPCStats(void)
8512 rx_interface_stat_p rpc_stat, nrpc_stat;
8515 MUTEX_ENTER(&rx_rpc_stats);
8518 * Turn off process statistics and if peer stats is also off, turn
8522 rxi_monitor_processStats = 0;
8523 if (rxi_monitor_peerStats == 0) {
8524 rx_enable_stats = 0;
8527 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8528 unsigned int num_funcs = 0;
8531 queue_Remove(rpc_stat);
8532 num_funcs = rpc_stat->stats[0].func_total;
8534 sizeof(rx_interface_stat_t) +
8535 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8537 rxi_Free(rpc_stat, space);
8538 rxi_rpc_process_stat_cnt -= num_funcs;
8540 MUTEX_EXIT(&rx_rpc_stats);
8544 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8554 rx_disablePeerRPCStats(void)
8556 struct rx_peer **peer_ptr, **peer_end;
8560 * Turn off peer statistics and if process stats is also off, turn
8564 rxi_monitor_peerStats = 0;
8565 if (rxi_monitor_processStats == 0) {
8566 rx_enable_stats = 0;
8569 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8570 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8572 struct rx_peer *peer, *next, *prev;
8574 MUTEX_ENTER(&rx_peerHashTable_lock);
8575 MUTEX_ENTER(&rx_rpc_stats);
8576 for (prev = peer = *peer_ptr; peer; peer = next) {
8578 code = MUTEX_TRYENTER(&peer->peer_lock);
8580 rx_interface_stat_p rpc_stat, nrpc_stat;
8583 if (prev == *peer_ptr) {
8594 MUTEX_EXIT(&rx_peerHashTable_lock);
8597 (&peer->rpcStats, rpc_stat, nrpc_stat,
8598 rx_interface_stat)) {
8599 unsigned int num_funcs = 0;
8602 queue_Remove(&rpc_stat->queue_header);
8603 queue_Remove(&rpc_stat->all_peers);
8604 num_funcs = rpc_stat->stats[0].func_total;
8606 sizeof(rx_interface_stat_t) +
8607 rpc_stat->stats[0].func_total *
8608 sizeof(rx_function_entry_v1_t);
8610 rxi_Free(rpc_stat, space);
8611 rxi_rpc_peer_stat_cnt -= num_funcs;
8613 MUTEX_EXIT(&peer->peer_lock);
8615 MUTEX_ENTER(&rx_peerHashTable_lock);
8625 MUTEX_EXIT(&rx_rpc_stats);
8626 MUTEX_EXIT(&rx_peerHashTable_lock);
8631 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8636 * IN clearFlag - flag indicating which stats to clear
8644 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8646 rx_interface_stat_p rpc_stat, nrpc_stat;
8648 MUTEX_ENTER(&rx_rpc_stats);
8650 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8651 unsigned int num_funcs = 0, i;
8652 num_funcs = rpc_stat->stats[0].func_total;
8653 for (i = 0; i < num_funcs; i++) {
8654 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8655 hzero(rpc_stat->stats[i].invocations);
8657 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8658 hzero(rpc_stat->stats[i].bytes_sent);
8660 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8661 hzero(rpc_stat->stats[i].bytes_rcvd);
8663 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8664 rpc_stat->stats[i].queue_time_sum.sec = 0;
8665 rpc_stat->stats[i].queue_time_sum.usec = 0;
8667 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8668 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8669 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8671 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8672 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8673 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8675 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8676 rpc_stat->stats[i].queue_time_max.sec = 0;
8677 rpc_stat->stats[i].queue_time_max.usec = 0;
8679 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8680 rpc_stat->stats[i].execution_time_sum.sec = 0;
8681 rpc_stat->stats[i].execution_time_sum.usec = 0;
8683 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8684 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8685 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8687 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8688 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8689 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8691 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8692 rpc_stat->stats[i].execution_time_max.sec = 0;
8693 rpc_stat->stats[i].execution_time_max.usec = 0;
8698 MUTEX_EXIT(&rx_rpc_stats);
8702 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8707 * IN clearFlag - flag indicating which stats to clear
8715 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8717 rx_interface_stat_p rpc_stat, nrpc_stat;
8719 MUTEX_ENTER(&rx_rpc_stats);
8721 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8722 unsigned int num_funcs = 0, i;
8725 * We have to fix the offset of rpc_stat since we are
8726 * keeping this structure on two rx_queues. The rx_queue
8727 * package assumes that the rx_queue member is the first
8728 * member of the structure. That is, rx_queue assumes that
8729 * any one item is only on one queue at a time. We are
8730 * breaking that assumption and so we have to do a little
8731 * math to fix our pointers.
8734 fix_offset = (char *)rpc_stat;
8735 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8736 rpc_stat = (rx_interface_stat_p) fix_offset;
8738 num_funcs = rpc_stat->stats[0].func_total;
8739 for (i = 0; i < num_funcs; i++) {
8740 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8741 hzero(rpc_stat->stats[i].invocations);
8743 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8744 hzero(rpc_stat->stats[i].bytes_sent);
8746 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8747 hzero(rpc_stat->stats[i].bytes_rcvd);
8749 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8750 rpc_stat->stats[i].queue_time_sum.sec = 0;
8751 rpc_stat->stats[i].queue_time_sum.usec = 0;
8753 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8754 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8755 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8757 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8758 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8759 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8761 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8762 rpc_stat->stats[i].queue_time_max.sec = 0;
8763 rpc_stat->stats[i].queue_time_max.usec = 0;
8765 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8766 rpc_stat->stats[i].execution_time_sum.sec = 0;
8767 rpc_stat->stats[i].execution_time_sum.usec = 0;
8769 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8770 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8771 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8773 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8774 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8775 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8777 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8778 rpc_stat->stats[i].execution_time_max.sec = 0;
8779 rpc_stat->stats[i].execution_time_max.usec = 0;
8784 MUTEX_EXIT(&rx_rpc_stats);
8788 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8789 * is authorized to enable/disable/clear RX statistics.
8791 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8794 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8796 rxi_rxstat_userok = proc;
8800 rx_RxStatUserOk(struct rx_call *call)
8802 if (!rxi_rxstat_userok)
8804 return rxi_rxstat_userok(call);
8809 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8810 * function in the MSVC runtime DLL (msvcrt.dll).
8812 * Note: the system serializes calls to this function.
8815 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8816 DWORD reason, /* reason function is being called */
8817 LPVOID reserved) /* reserved for future use */
8820 case DLL_PROCESS_ATTACH:
8821 /* library is being attached to a process */
8825 case DLL_PROCESS_DETACH:
8832 #endif /* AFS_NT40_ENV */
8835 int rx_DumpCalls(FILE *outputFile, char *cookie)
8837 #ifdef RXDEBUG_PACKET
8838 #ifdef KDUMP_RX_LOCK
8839 struct rx_call_rx_lock *c;
8846 #define RXDPRINTF sprintf
8847 #define RXDPRINTOUT output
8849 #define RXDPRINTF fprintf
8850 #define RXDPRINTOUT outputFile
8853 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8855 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8858 for (c = rx_allCallsp; c; c = c->allNextp) {
8859 u_short rqc, tqc, iovqc;
8860 struct rx_packet *p, *np;
8862 MUTEX_ENTER(&c->lock);
8863 queue_Count(&c->rq, p, np, rx_packet, rqc);
8864 queue_Count(&c->tq, p, np, rx_packet, tqc);
8865 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8867 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, "
8868 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8869 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8870 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8871 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8872 #ifdef RX_ENABLE_LOCKS
8875 #ifdef RX_REFCOUNT_CHECK
8876 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
8877 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
8880 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,
8881 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
8882 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
8883 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
8884 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
8885 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
8886 #ifdef RX_ENABLE_LOCKS
8887 , (afs_uint32)c->refCount
8889 #ifdef RX_REFCOUNT_CHECK
8890 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
8893 MUTEX_EXIT(&c->lock);
8896 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8899 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
8901 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8903 #endif /* RXDEBUG_PACKET */