2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
14 #include "afs/param.h"
16 #include <afs/param.h>
23 #include "afs/sysincludes.h"
24 #include "afsincludes.h"
30 #include <net/net_globals.h>
31 #endif /* AFS_OSF_ENV */
32 #ifdef AFS_LINUX20_ENV
35 #include "netinet/in.h"
37 #include "inet/common.h"
39 #include "inet/ip_ire.h"
41 #include "afs/afs_args.h"
42 #include "afs/afs_osi.h"
43 #ifdef RX_KERNEL_TRACE
44 #include "rx_kcommon.h"
46 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
50 #undef RXDEBUG /* turn off debugging */
52 #if defined(AFS_SGI_ENV)
53 #include "sys/debug.h"
62 #endif /* AFS_OSF_ENV */
64 #include "afs/sysincludes.h"
65 #include "afsincludes.h"
68 #include "rx_kmutex.h"
69 #include "rx_kernel.h"
73 #include "rx_globals.h"
75 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
76 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
77 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
79 extern afs_int32 afs_termState;
81 #include "sys/lockl.h"
82 #include "sys/lock_def.h"
83 #endif /* AFS_AIX41_ENV */
84 # include "rxgen_consts.h"
86 # include <sys/types.h>
93 # include <afs/afsutil.h>
94 # include <WINNT\afsreg.h>
96 # include <sys/socket.h>
97 # include <sys/file.h>
99 # include <sys/stat.h>
100 # include <netinet/in.h>
101 # include <sys/time.h>
104 # include "rx_user.h"
105 # include "rx_clock.h"
106 # include "rx_queue.h"
107 # include "rx_globals.h"
108 # include "rx_trace.h"
109 # include <afs/rxgen_consts.h>
113 #ifdef AFS_PTHREAD_ENV
115 int (*registerProgram) (pid_t, char *) = 0;
116 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
119 int (*registerProgram) (PROCESS, char *) = 0;
120 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
124 /* Local static routines */
125 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
126 #ifdef RX_ENABLE_LOCKS
127 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
130 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
132 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
133 afs_int32 rxi_start_in_error;
135 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
138 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
139 * currently allocated within rx. This number is used to allocate the
140 * memory required to return the statistics when queried.
143 static unsigned int rxi_rpc_peer_stat_cnt;
146 * rxi_rpc_process_stat_cnt counts the total number of local process stat
147 * structures currently allocated within rx. The number is used to allocate
148 * the memory required to return the statistics when queried.
151 static unsigned int rxi_rpc_process_stat_cnt;
153 #if !defined(offsetof)
154 #include <stddef.h> /* for definition of offsetof() */
157 #ifdef AFS_PTHREAD_ENV
161 * Use procedural initialization of mutexes/condition variables
165 extern afs_kmutex_t rx_stats_mutex;
166 extern afs_kmutex_t rx_waiting_mutex;
167 extern afs_kmutex_t rx_quota_mutex;
168 extern afs_kmutex_t rx_pthread_mutex;
169 extern afs_kmutex_t rx_packets_mutex;
170 extern afs_kmutex_t des_init_mutex;
171 extern afs_kmutex_t des_random_mutex;
172 extern afs_kmutex_t rx_clock_mutex;
173 extern afs_kmutex_t rxi_connCacheMutex;
174 extern afs_kmutex_t rx_event_mutex;
175 extern afs_kmutex_t osi_malloc_mutex;
176 extern afs_kmutex_t event_handler_mutex;
177 extern afs_kmutex_t listener_mutex;
178 extern afs_kmutex_t rx_if_init_mutex;
179 extern afs_kmutex_t rx_if_mutex;
180 extern afs_kmutex_t rxkad_client_uid_mutex;
181 extern afs_kmutex_t rxkad_random_mutex;
183 extern afs_kcondvar_t rx_event_handler_cond;
184 extern afs_kcondvar_t rx_listener_cond;
186 static afs_kmutex_t epoch_mutex;
187 static afs_kmutex_t rx_init_mutex;
188 static afs_kmutex_t rx_debug_mutex;
189 static afs_kmutex_t rx_rpc_stats;
192 rxi_InitPthread(void)
194 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
195 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
196 MUTEX_INIT(&rx_waiting_mutex, "waiting", MUTEX_DEFAULT, 0);
197 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
198 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
199 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
200 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
201 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
202 MUTEX_INIT(&rx_event_mutex, "event", MUTEX_DEFAULT, 0);
203 MUTEX_INIT(&des_init_mutex, "des", MUTEX_DEFAULT, 0);
204 MUTEX_INIT(&des_random_mutex, "random", MUTEX_DEFAULT, 0);
205 MUTEX_INIT(&osi_malloc_mutex, "malloc", MUTEX_DEFAULT, 0);
206 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
207 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
208 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
209 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
210 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
211 MUTEX_INIT(&rxkad_client_uid_mutex, "uid", MUTEX_DEFAULT, 0);
212 MUTEX_INIT(&rxkad_random_mutex, "rxkad random", MUTEX_DEFAULT, 0);
213 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
215 assert(pthread_cond_init
216 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
217 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
219 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
220 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
222 rxkad_global_stats_init();
224 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
225 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
226 #ifdef RX_ENABLE_LOCKS
229 #endif /* RX_LOCKS_DB */
230 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
231 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
233 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
235 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
237 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
239 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
240 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
241 #endif /* RX_ENABLE_LOCKS */
244 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
245 #define INIT_PTHREAD_LOCKS \
246 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
248 * The rx_stats_mutex mutex protects the following global variables:
249 * rxi_lowConnRefCount
250 * rxi_lowPeerRefCount
259 * The rx_quota_mutex mutex protects the following global variables:
267 * The rx_freePktQ_lock protects the following global variables:
272 * The rx_packets_mutex mutex protects the following global variables:
280 * The rx_pthread_mutex mutex protects the following global variables:
284 #define INIT_PTHREAD_LOCKS
288 /* Variables for handling the minProcs implementation. availProcs gives the
289 * number of threads available in the pool at this moment (not counting dudes
290 * executing right now). totalMin gives the total number of procs required
291 * for handling all minProcs requests. minDeficit is a dynamic variable
292 * tracking the # of procs required to satisfy all of the remaining minProcs
294 * For fine grain locking to work, the quota check and the reservation of
295 * a server thread has to come while rxi_availProcs and rxi_minDeficit
296 * are locked. To this end, the code has been modified under #ifdef
297 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
298 * same time. A new function, ReturnToServerPool() returns the allocation.
300 * A call can be on several queue's (but only one at a time). When
301 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
302 * that no one else is touching the queue. To this end, we store the address
303 * of the queue lock in the call structure (under the call lock) when we
304 * put the call on a queue, and we clear the call_queue_lock when the
305 * call is removed from a queue (once the call lock has been obtained).
306 * This allows rxi_ResetCall to safely synchronize with others wishing
307 * to manipulate the queue.
310 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
311 static afs_kmutex_t rx_rpc_stats;
312 void rxi_StartUnlocked(struct rxevent *event, void *call,
313 void *arg1, int istack);
316 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
317 ** pretty good that the next packet coming in is from the same connection
318 ** as the last packet, since we're send multiple packets in a transmit window.
320 struct rx_connection *rxLastConn = 0;
322 #ifdef RX_ENABLE_LOCKS
323 /* The locking hierarchy for rx fine grain locking is composed of these
326 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
327 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
328 * call->lock - locks call data fields.
329 * These are independent of each other:
330 * rx_freeCallQueue_lock
335 * serverQueueEntry->lock
337 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
338 * peer->lock - locks peer data fields.
339 * conn_data_lock - that more than one thread is not updating a conn data
340 * field at the same time.
348 * Do we need a lock to protect the peer field in the conn structure?
349 * conn->peer was previously a constant for all intents and so has no
350 * lock protecting this field. The multihomed client delta introduced
351 * a RX code change : change the peer field in the connection structure
352 * to that remote inetrface from which the last packet for this
353 * connection was sent out. This may become an issue if further changes
356 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
357 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
359 /* rxdb_fileID is used to identify the lock location, along with line#. */
360 static int rxdb_fileID = RXDB_FILE_RX;
361 #endif /* RX_LOCKS_DB */
362 #else /* RX_ENABLE_LOCKS */
363 #define SET_CALL_QUEUE_LOCK(C, L)
364 #define CLEAR_CALL_QUEUE_LOCK(C)
365 #endif /* RX_ENABLE_LOCKS */
366 struct rx_serverQueueEntry *rx_waitForPacket = 0;
367 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
369 /* ------------Exported Interfaces------------- */
371 /* This function allows rxkad to set the epoch to a suitably random number
372 * which rx_NewConnection will use in the future. The principle purpose is to
373 * get rxnull connections to use the same epoch as the rxkad connections do, at
374 * least once the first rxkad connection is established. This is important now
375 * that the host/port addresses aren't used in FindConnection: the uniqueness
376 * of epoch/cid matters and the start time won't do. */
378 #ifdef AFS_PTHREAD_ENV
380 * This mutex protects the following global variables:
384 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
385 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
389 #endif /* AFS_PTHREAD_ENV */
392 rx_SetEpoch(afs_uint32 epoch)
399 /* Initialize rx. A port number may be mentioned, in which case this
400 * becomes the default port number for any service installed later.
401 * If 0 is provided for the port number, a random port will be chosen
402 * by the kernel. Whether this will ever overlap anything in
403 * /etc/services is anybody's guess... Returns 0 on success, -1 on
408 int rxinit_status = 1;
409 #ifdef AFS_PTHREAD_ENV
411 * This mutex protects the following global variables:
415 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
416 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
419 #define UNLOCK_RX_INIT
423 rx_InitHost(u_int host, u_int port)
430 char *htable, *ptable;
437 if (rxinit_status == 0) {
438 tmp_status = rxinit_status;
440 return tmp_status; /* Already started; return previous error code. */
446 if (afs_winsockInit() < 0)
452 * Initialize anything necessary to provide a non-premptive threading
455 rxi_InitializeThreadSupport();
458 /* Allocate and initialize a socket for client and perhaps server
461 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
462 if (rx_socket == OSI_NULLSOCKET) {
466 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
469 #endif /* RX_LOCKS_DB */
470 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
471 MUTEX_INIT(&rx_waiting_mutex, "rx_waiting_mutex", MUTEX_DEFAULT, 0);
472 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
473 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
474 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
475 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
476 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
477 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
478 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
480 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
482 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
484 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
486 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
487 #if defined(AFS_HPUX110_ENV)
489 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
490 #endif /* AFS_HPUX110_ENV */
491 #endif /* RX_ENABLE_LOCKS && KERNEL */
494 rx_connDeadTime = 12;
495 rx_tranquil = 0; /* reset flag */
496 memset((char *)&rx_stats, 0, sizeof(struct rx_statistics));
498 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
499 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
500 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
501 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
502 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
503 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
505 /* Malloc up a bunch of packets & buffers */
507 queue_Init(&rx_freePacketQueue);
508 rxi_NeedMorePackets = FALSE;
509 #ifdef RX_ENABLE_TSFPQ
510 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
511 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
512 #else /* RX_ENABLE_TSFPQ */
513 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
514 rxi_MorePackets(rx_nPackets);
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((int)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 if (rxi_availProcs > rxi_minDeficit)
669 #endif /* RX_ENABLE_LOCKS */
672 /* Called by rx_StartServer to start up lwp's to service calls.
673 NExistingProcs gives the number of procs already existing, and which
674 therefore needn't be created. */
676 rxi_StartServerProcs(int nExistingProcs)
678 struct rx_service *service;
683 /* For each service, reserve N processes, where N is the "minimum"
684 * number of processes that MUST be able to execute a request in parallel,
685 * at any time, for that process. Also compute the maximum difference
686 * between any service's maximum number of processes that can run
687 * (i.e. the maximum number that ever will be run, and a guarantee
688 * that this number will run if other services aren't running), and its
689 * minimum number. The result is the extra number of processes that
690 * we need in order to provide the latter guarantee */
691 for (i = 0; i < RX_MAX_SERVICES; i++) {
693 service = rx_services[i];
694 if (service == (struct rx_service *)0)
696 nProcs += service->minProcs;
697 diff = service->maxProcs - service->minProcs;
701 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
702 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
703 for (i = 0; i < nProcs; i++) {
704 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
710 /* This routine is only required on Windows */
712 rx_StartClientThread(void)
714 #ifdef AFS_PTHREAD_ENV
716 pid = pthread_self();
717 #endif /* AFS_PTHREAD_ENV */
719 #endif /* AFS_NT40_ENV */
721 /* This routine must be called if any services are exported. If the
722 * donateMe flag is set, the calling process is donated to the server
725 rx_StartServer(int donateMe)
727 struct rx_service *service;
733 /* Start server processes, if necessary (exact function is dependent
734 * on the implementation environment--kernel or user space). DonateMe
735 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
736 * case, one less new proc will be created rx_StartServerProcs.
738 rxi_StartServerProcs(donateMe);
740 /* count up the # of threads in minProcs, and add set the min deficit to
741 * be that value, too.
743 for (i = 0; i < RX_MAX_SERVICES; i++) {
744 service = rx_services[i];
745 if (service == (struct rx_service *)0)
747 MUTEX_ENTER(&rx_quota_mutex);
748 rxi_totalMin += service->minProcs;
749 /* below works even if a thread is running, since minDeficit would
750 * still have been decremented and later re-incremented.
752 rxi_minDeficit += service->minProcs;
753 MUTEX_EXIT(&rx_quota_mutex);
756 /* Turn on reaping of idle server connections */
757 rxi_ReapConnections(NULL, NULL, NULL);
766 #ifdef AFS_PTHREAD_ENV
768 pid = (pid_t) pthread_self();
769 #else /* AFS_PTHREAD_ENV */
771 LWP_CurrentProcess(&pid);
772 #endif /* AFS_PTHREAD_ENV */
774 sprintf(name, "srv_%d", ++nProcs);
776 (*registerProgram) (pid, name);
778 #endif /* AFS_NT40_ENV */
779 rx_ServerProc(NULL); /* Never returns */
781 #ifdef RX_ENABLE_TSFPQ
782 /* no use leaving packets around in this thread's local queue if
783 * it isn't getting donated to the server thread pool.
785 rxi_FlushLocalPacketsTSFPQ();
786 #endif /* RX_ENABLE_TSFPQ */
790 /* Create a new client connection to the specified service, using the
791 * specified security object to implement the security model for this
793 struct rx_connection *
794 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
795 struct rx_securityClass *securityObject,
796 int serviceSecurityIndex)
800 struct rx_connection *conn;
805 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", ntohl(shost), ntohs(sport), sservice, securityObject, serviceSecurityIndex));
807 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
808 * the case of kmem_alloc? */
809 conn = rxi_AllocConnection();
810 #ifdef RX_ENABLE_LOCKS
811 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
812 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
813 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
816 MUTEX_ENTER(&rx_connHashTable_lock);
817 cid = (rx_nextCid += RX_MAXCALLS);
818 conn->type = RX_CLIENT_CONNECTION;
820 conn->epoch = rx_epoch;
821 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
822 conn->serviceId = sservice;
823 conn->securityObject = securityObject;
824 conn->securityData = (void *) 0;
825 conn->securityIndex = serviceSecurityIndex;
826 rx_SetConnDeadTime(conn, rx_connDeadTime);
827 conn->ackRate = RX_FAST_ACK_RATE;
829 conn->specific = NULL;
830 conn->challengeEvent = NULL;
831 conn->delayedAbortEvent = NULL;
832 conn->abortCount = 0;
834 for (i = 0; i < RX_MAXCALLS; i++) {
835 conn->twind[i] = rx_initSendWindow;
836 conn->rwind[i] = rx_initReceiveWindow;
839 RXS_NewConnection(securityObject, conn);
841 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
843 conn->refCount++; /* no lock required since only this thread knows... */
844 conn->next = rx_connHashTable[hashindex];
845 rx_connHashTable[hashindex] = conn;
847 rx_MutexIncrement(rx_stats.nClientConns, rx_stats_mutex);
848 MUTEX_EXIT(&rx_connHashTable_lock);
854 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
856 /* The idea is to set the dead time to a value that allows several
857 * keepalives to be dropped without timing out the connection. */
858 conn->secondsUntilDead = MAX(seconds, 6);
859 conn->secondsUntilPing = conn->secondsUntilDead / 6;
862 int rxi_lowPeerRefCount = 0;
863 int rxi_lowConnRefCount = 0;
866 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
867 * NOTE: must not be called with rx_connHashTable_lock held.
870 rxi_CleanupConnection(struct rx_connection *conn)
872 /* Notify the service exporter, if requested, that this connection
873 * is being destroyed */
874 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
875 (*conn->service->destroyConnProc) (conn);
877 /* Notify the security module that this connection is being destroyed */
878 RXS_DestroyConnection(conn->securityObject, conn);
880 /* If this is the last connection using the rx_peer struct, set its
881 * idle time to now. rxi_ReapConnections will reap it if it's still
882 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
884 MUTEX_ENTER(&rx_peerHashTable_lock);
885 if (conn->peer->refCount < 2) {
886 conn->peer->idleWhen = clock_Sec();
887 if (conn->peer->refCount < 1) {
888 conn->peer->refCount = 1;
889 if (rx_stats_active) {
890 MUTEX_ENTER(&rx_stats_mutex);
891 rxi_lowPeerRefCount++;
892 MUTEX_EXIT(&rx_stats_mutex);
896 conn->peer->refCount--;
897 MUTEX_EXIT(&rx_peerHashTable_lock);
901 if (conn->type == RX_SERVER_CONNECTION)
902 rx_MutexDecrement(rx_stats.nServerConns, rx_stats_mutex);
904 rx_MutexDecrement(rx_stats.nClientConns, rx_stats_mutex);
907 if (conn->specific) {
909 for (i = 0; i < conn->nSpecific; i++) {
910 if (conn->specific[i] && rxi_keyCreate_destructor[i])
911 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
912 conn->specific[i] = NULL;
914 free(conn->specific);
916 conn->specific = NULL;
920 MUTEX_DESTROY(&conn->conn_call_lock);
921 MUTEX_DESTROY(&conn->conn_data_lock);
922 CV_DESTROY(&conn->conn_call_cv);
924 rxi_FreeConnection(conn);
927 /* Destroy the specified connection */
929 rxi_DestroyConnection(struct rx_connection *conn)
931 MUTEX_ENTER(&rx_connHashTable_lock);
932 rxi_DestroyConnectionNoLock(conn);
933 /* conn should be at the head of the cleanup list */
934 if (conn == rx_connCleanup_list) {
935 rx_connCleanup_list = rx_connCleanup_list->next;
936 MUTEX_EXIT(&rx_connHashTable_lock);
937 rxi_CleanupConnection(conn);
939 #ifdef RX_ENABLE_LOCKS
941 MUTEX_EXIT(&rx_connHashTable_lock);
943 #endif /* RX_ENABLE_LOCKS */
947 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
949 struct rx_connection **conn_ptr;
951 struct rx_packet *packet;
958 MUTEX_ENTER(&conn->conn_data_lock);
959 if (conn->refCount > 0)
962 if (rx_stats_active) {
963 MUTEX_ENTER(&rx_stats_mutex);
964 rxi_lowConnRefCount++;
965 MUTEX_EXIT(&rx_stats_mutex);
969 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
970 /* Busy; wait till the last guy before proceeding */
971 MUTEX_EXIT(&conn->conn_data_lock);
976 /* If the client previously called rx_NewCall, but it is still
977 * waiting, treat this as a running call, and wait to destroy the
978 * connection later when the call completes. */
979 if ((conn->type == RX_CLIENT_CONNECTION)
980 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
981 conn->flags |= RX_CONN_DESTROY_ME;
982 MUTEX_EXIT(&conn->conn_data_lock);
986 MUTEX_EXIT(&conn->conn_data_lock);
988 /* Check for extant references to this connection */
989 for (i = 0; i < RX_MAXCALLS; i++) {
990 struct rx_call *call = conn->call[i];
993 if (conn->type == RX_CLIENT_CONNECTION) {
994 MUTEX_ENTER(&call->lock);
995 if (call->delayedAckEvent) {
996 /* Push the final acknowledgment out now--there
997 * won't be a subsequent call to acknowledge the
998 * last reply packets */
999 rxevent_Cancel(call->delayedAckEvent, call,
1000 RX_CALL_REFCOUNT_DELAY);
1001 if (call->state == RX_STATE_PRECALL
1002 || call->state == RX_STATE_ACTIVE) {
1003 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1005 rxi_AckAll(NULL, call, 0);
1008 MUTEX_EXIT(&call->lock);
1012 #ifdef RX_ENABLE_LOCKS
1014 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1015 MUTEX_EXIT(&conn->conn_data_lock);
1017 /* Someone is accessing a packet right now. */
1021 #endif /* RX_ENABLE_LOCKS */
1024 /* Don't destroy the connection if there are any call
1025 * structures still in use */
1026 MUTEX_ENTER(&conn->conn_data_lock);
1027 conn->flags |= RX_CONN_DESTROY_ME;
1028 MUTEX_EXIT(&conn->conn_data_lock);
1033 if (conn->delayedAbortEvent) {
1034 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1035 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1037 MUTEX_ENTER(&conn->conn_data_lock);
1038 rxi_SendConnectionAbort(conn, packet, 0, 1);
1039 MUTEX_EXIT(&conn->conn_data_lock);
1040 rxi_FreePacket(packet);
1044 /* Remove from connection hash table before proceeding */
1046 &rx_connHashTable[CONN_HASH
1047 (peer->host, peer->port, conn->cid, conn->epoch,
1049 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1050 if (*conn_ptr == conn) {
1051 *conn_ptr = conn->next;
1055 /* if the conn that we are destroying was the last connection, then we
1056 * clear rxLastConn as well */
1057 if (rxLastConn == conn)
1060 /* Make sure the connection is completely reset before deleting it. */
1061 /* get rid of pending events that could zap us later */
1062 if (conn->challengeEvent)
1063 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1064 if (conn->checkReachEvent)
1065 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1067 /* Add the connection to the list of destroyed connections that
1068 * need to be cleaned up. This is necessary to avoid deadlocks
1069 * in the routines we call to inform others that this connection is
1070 * being destroyed. */
1071 conn->next = rx_connCleanup_list;
1072 rx_connCleanup_list = conn;
1075 /* Externally available version */
1077 rx_DestroyConnection(struct rx_connection *conn)
1082 rxi_DestroyConnection(conn);
1087 rx_GetConnection(struct rx_connection *conn)
1092 MUTEX_ENTER(&conn->conn_data_lock);
1094 MUTEX_EXIT(&conn->conn_data_lock);
1098 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1099 /* Wait for the transmit queue to no longer be busy.
1100 * requires the call->lock to be held */
1101 static void rxi_WaitforTQBusy(struct rx_call *call) {
1102 while (call->flags & RX_CALL_TQ_BUSY) {
1103 call->flags |= RX_CALL_TQ_WAIT;
1105 #ifdef RX_ENABLE_LOCKS
1106 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1107 CV_WAIT(&call->cv_tq, &call->lock);
1108 #else /* RX_ENABLE_LOCKS */
1109 osi_rxSleep(&call->tq);
1110 #endif /* RX_ENABLE_LOCKS */
1112 if (call->tqWaiters == 0) {
1113 call->flags &= ~RX_CALL_TQ_WAIT;
1119 /* Start a new rx remote procedure call, on the specified connection.
1120 * If wait is set to 1, wait for a free call channel; otherwise return
1121 * 0. Maxtime gives the maximum number of seconds this call may take,
1122 * after rx_NewCall returns. After this time interval, a call to any
1123 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1124 * For fine grain locking, we hold the conn_call_lock in order to
1125 * to ensure that we don't get signalle after we found a call in an active
1126 * state and before we go to sleep.
1129 rx_NewCall(struct rx_connection *conn)
1132 struct rx_call *call;
1133 struct clock queueTime;
1137 dpf(("rx_NewCall(conn %x)\n", conn));
1140 clock_GetTime(&queueTime);
1141 MUTEX_ENTER(&conn->conn_call_lock);
1144 * Check if there are others waiting for a new call.
1145 * If so, let them go first to avoid starving them.
1146 * This is a fairly simple scheme, and might not be
1147 * a complete solution for large numbers of waiters.
1149 * makeCallWaiters keeps track of the number of
1150 * threads waiting to make calls and the
1151 * RX_CONN_MAKECALL_WAITING flag bit is used to
1152 * indicate that there are indeed calls waiting.
1153 * The flag is set when the waiter is incremented.
1154 * It is only cleared in rx_EndCall when
1155 * makeCallWaiters is 0. This prevents us from
1156 * accidently destroying the connection while it
1157 * is potentially about to be used.
1159 MUTEX_ENTER(&conn->conn_data_lock);
1160 if (conn->makeCallWaiters) {
1161 conn->flags |= RX_CONN_MAKECALL_WAITING;
1162 conn->makeCallWaiters++;
1163 MUTEX_EXIT(&conn->conn_data_lock);
1165 #ifdef RX_ENABLE_LOCKS
1166 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1170 MUTEX_ENTER(&conn->conn_data_lock);
1171 conn->makeCallWaiters--;
1173 MUTEX_EXIT(&conn->conn_data_lock);
1176 for (i = 0; i < RX_MAXCALLS; i++) {
1177 call = conn->call[i];
1179 MUTEX_ENTER(&call->lock);
1180 if (call->state == RX_STATE_DALLY) {
1181 rxi_ResetCall(call, 0);
1182 (*call->callNumber)++;
1185 MUTEX_EXIT(&call->lock);
1187 call = rxi_NewCall(conn, i);
1191 if (i < RX_MAXCALLS) {
1194 MUTEX_ENTER(&conn->conn_data_lock);
1195 conn->flags |= RX_CONN_MAKECALL_WAITING;
1196 conn->makeCallWaiters++;
1197 MUTEX_EXIT(&conn->conn_data_lock);
1199 #ifdef RX_ENABLE_LOCKS
1200 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1204 MUTEX_ENTER(&conn->conn_data_lock);
1205 conn->makeCallWaiters--;
1206 MUTEX_EXIT(&conn->conn_data_lock);
1209 * Wake up anyone else who might be giving us a chance to
1210 * run (see code above that avoids resource starvation).
1212 #ifdef RX_ENABLE_LOCKS
1213 CV_BROADCAST(&conn->conn_call_cv);
1218 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1220 /* Client is initially in send mode */
1221 call->state = RX_STATE_ACTIVE;
1222 call->error = conn->error;
1224 call->mode = RX_MODE_ERROR;
1226 call->mode = RX_MODE_SENDING;
1228 /* remember start time for call in case we have hard dead time limit */
1229 call->queueTime = queueTime;
1230 clock_GetTime(&call->startTime);
1231 hzero(call->bytesSent);
1232 hzero(call->bytesRcvd);
1234 /* Turn on busy protocol. */
1235 rxi_KeepAliveOn(call);
1237 MUTEX_EXIT(&call->lock);
1238 MUTEX_EXIT(&conn->conn_call_lock);
1241 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1242 /* Now, if TQ wasn't cleared earlier, do it now. */
1243 MUTEX_ENTER(&call->lock);
1244 rxi_WaitforTQBusy(call);
1245 if (call->flags & RX_CALL_TQ_CLEARME) {
1246 rxi_ClearTransmitQueue(call, 1);
1247 /*queue_Init(&call->tq);*/
1249 MUTEX_EXIT(&call->lock);
1250 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1252 dpf(("rx_NewCall(call %x)\n", call));
1257 rxi_HasActiveCalls(struct rx_connection *aconn)
1260 struct rx_call *tcall;
1264 for (i = 0; i < RX_MAXCALLS; i++) {
1265 if ((tcall = aconn->call[i])) {
1266 if ((tcall->state == RX_STATE_ACTIVE)
1267 || (tcall->state == RX_STATE_PRECALL)) {
1278 rxi_GetCallNumberVector(struct rx_connection *aconn,
1279 afs_int32 * aint32s)
1282 struct rx_call *tcall;
1286 for (i = 0; i < RX_MAXCALLS; i++) {
1287 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1288 aint32s[i] = aconn->callNumber[i] + 1;
1290 aint32s[i] = aconn->callNumber[i];
1297 rxi_SetCallNumberVector(struct rx_connection *aconn,
1298 afs_int32 * aint32s)
1301 struct rx_call *tcall;
1305 for (i = 0; i < RX_MAXCALLS; i++) {
1306 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1307 aconn->callNumber[i] = aint32s[i] - 1;
1309 aconn->callNumber[i] = aint32s[i];
1315 /* Advertise a new service. A service is named locally by a UDP port
1316 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1319 char *serviceName; Name for identification purposes (e.g. the
1320 service name might be used for probing for
1323 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1324 char *serviceName, struct rx_securityClass **securityObjects,
1325 int nSecurityObjects,
1326 afs_int32(*serviceProc) (struct rx_call * acall))
1328 osi_socket socket = OSI_NULLSOCKET;
1329 struct rx_service *tservice;
1335 if (serviceId == 0) {
1337 "rx_NewService: service id for service %s is not non-zero.\n",
1344 "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",
1352 tservice = rxi_AllocService();
1354 for (i = 0; i < RX_MAX_SERVICES; i++) {
1355 struct rx_service *service = rx_services[i];
1357 if (port == service->servicePort && host == service->serviceHost) {
1358 if (service->serviceId == serviceId) {
1359 /* The identical service has already been
1360 * installed; if the caller was intending to
1361 * change the security classes used by this
1362 * service, he/she loses. */
1364 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1365 serviceName, serviceId, service->serviceName);
1367 rxi_FreeService(tservice);
1370 /* Different service, same port: re-use the socket
1371 * which is bound to the same port */
1372 socket = service->socket;
1375 if (socket == OSI_NULLSOCKET) {
1376 /* If we don't already have a socket (from another
1377 * service on same port) get a new one */
1378 socket = rxi_GetHostUDPSocket(host, port);
1379 if (socket == OSI_NULLSOCKET) {
1381 rxi_FreeService(tservice);
1386 service->socket = socket;
1387 service->serviceHost = host;
1388 service->servicePort = port;
1389 service->serviceId = serviceId;
1390 service->serviceName = serviceName;
1391 service->nSecurityObjects = nSecurityObjects;
1392 service->securityObjects = securityObjects;
1393 service->minProcs = 0;
1394 service->maxProcs = 1;
1395 service->idleDeadTime = 60;
1396 service->idleDeadErr = 0;
1397 service->connDeadTime = rx_connDeadTime;
1398 service->executeRequestProc = serviceProc;
1399 service->checkReach = 0;
1400 rx_services[i] = service; /* not visible until now */
1406 rxi_FreeService(tservice);
1407 (osi_Msg "rx_NewService: cannot support > %d services\n",
1412 /* Set configuration options for all of a service's security objects */
1415 rx_SetSecurityConfiguration(struct rx_service *service,
1416 rx_securityConfigVariables type,
1420 for (i = 0; i<service->nSecurityObjects; i++) {
1421 if (service->securityObjects[i]) {
1422 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1430 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1431 struct rx_securityClass **securityObjects, int nSecurityObjects,
1432 afs_int32(*serviceProc) (struct rx_call * acall))
1434 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1437 /* Generic request processing loop. This routine should be called
1438 * by the implementation dependent rx_ServerProc. If socketp is
1439 * non-null, it will be set to the file descriptor that this thread
1440 * is now listening on. If socketp is null, this routine will never
1443 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1445 struct rx_call *call;
1447 struct rx_service *tservice = NULL;
1454 call = rx_GetCall(threadID, tservice, socketp);
1455 if (socketp && *socketp != OSI_NULLSOCKET) {
1456 /* We are now a listener thread */
1461 /* if server is restarting( typically smooth shutdown) then do not
1462 * allow any new calls.
1465 if (rx_tranquil && (call != NULL)) {
1469 MUTEX_ENTER(&call->lock);
1471 rxi_CallError(call, RX_RESTARTING);
1472 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1474 MUTEX_EXIT(&call->lock);
1478 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1479 #ifdef RX_ENABLE_LOCKS
1481 #endif /* RX_ENABLE_LOCKS */
1482 afs_termState = AFSOP_STOP_AFS;
1483 afs_osi_Wakeup(&afs_termState);
1484 #ifdef RX_ENABLE_LOCKS
1486 #endif /* RX_ENABLE_LOCKS */
1491 tservice = call->conn->service;
1493 if (tservice->beforeProc)
1494 (*tservice->beforeProc) (call);
1496 code = call->conn->service->executeRequestProc(call);
1498 if (tservice->afterProc)
1499 (*tservice->afterProc) (call, code);
1501 rx_EndCall(call, code);
1502 if (rx_stats_active) {
1503 MUTEX_ENTER(&rx_stats_mutex);
1505 MUTEX_EXIT(&rx_stats_mutex);
1512 rx_WakeupServerProcs(void)
1514 struct rx_serverQueueEntry *np, *tqp;
1518 MUTEX_ENTER(&rx_serverPool_lock);
1520 #ifdef RX_ENABLE_LOCKS
1521 if (rx_waitForPacket)
1522 CV_BROADCAST(&rx_waitForPacket->cv);
1523 #else /* RX_ENABLE_LOCKS */
1524 if (rx_waitForPacket)
1525 osi_rxWakeup(rx_waitForPacket);
1526 #endif /* RX_ENABLE_LOCKS */
1527 MUTEX_ENTER(&freeSQEList_lock);
1528 for (np = rx_FreeSQEList; np; np = tqp) {
1529 tqp = *(struct rx_serverQueueEntry **)np;
1530 #ifdef RX_ENABLE_LOCKS
1531 CV_BROADCAST(&np->cv);
1532 #else /* RX_ENABLE_LOCKS */
1534 #endif /* RX_ENABLE_LOCKS */
1536 MUTEX_EXIT(&freeSQEList_lock);
1537 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1538 #ifdef RX_ENABLE_LOCKS
1539 CV_BROADCAST(&np->cv);
1540 #else /* RX_ENABLE_LOCKS */
1542 #endif /* RX_ENABLE_LOCKS */
1544 MUTEX_EXIT(&rx_serverPool_lock);
1549 * One thing that seems to happen is that all the server threads get
1550 * tied up on some empty or slow call, and then a whole bunch of calls
1551 * arrive at once, using up the packet pool, so now there are more
1552 * empty calls. The most critical resources here are server threads
1553 * and the free packet pool. The "doreclaim" code seems to help in
1554 * general. I think that eventually we arrive in this state: there
1555 * are lots of pending calls which do have all their packets present,
1556 * so they won't be reclaimed, are multi-packet calls, so they won't
1557 * be scheduled until later, and thus are tying up most of the free
1558 * packet pool for a very long time.
1560 * 1. schedule multi-packet calls if all the packets are present.
1561 * Probably CPU-bound operation, useful to return packets to pool.
1562 * Do what if there is a full window, but the last packet isn't here?
1563 * 3. preserve one thread which *only* runs "best" calls, otherwise
1564 * it sleeps and waits for that type of call.
1565 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1566 * the current dataquota business is badly broken. The quota isn't adjusted
1567 * to reflect how many packets are presently queued for a running call.
1568 * So, when we schedule a queued call with a full window of packets queued
1569 * up for it, that *should* free up a window full of packets for other 2d-class
1570 * calls to be able to use from the packet pool. But it doesn't.
1572 * NB. Most of the time, this code doesn't run -- since idle server threads
1573 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1574 * as a new call arrives.
1576 /* Sleep until a call arrives. Returns a pointer to the call, ready
1577 * for an rx_Read. */
1578 #ifdef RX_ENABLE_LOCKS
1580 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1582 struct rx_serverQueueEntry *sq;
1583 struct rx_call *call = (struct rx_call *)0;
1584 struct rx_service *service = NULL;
1587 MUTEX_ENTER(&freeSQEList_lock);
1589 if ((sq = rx_FreeSQEList)) {
1590 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1591 MUTEX_EXIT(&freeSQEList_lock);
1592 } else { /* otherwise allocate a new one and return that */
1593 MUTEX_EXIT(&freeSQEList_lock);
1594 sq = (struct rx_serverQueueEntry *)
1595 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1596 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1597 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1600 MUTEX_ENTER(&rx_serverPool_lock);
1601 if (cur_service != NULL) {
1602 ReturnToServerPool(cur_service);
1605 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1606 struct rx_call *tcall, *ncall, *choice2 = NULL;
1608 /* Scan for eligible incoming calls. A call is not eligible
1609 * if the maximum number of calls for its service type are
1610 * already executing */
1611 /* One thread will process calls FCFS (to prevent starvation),
1612 * while the other threads may run ahead looking for calls which
1613 * have all their input data available immediately. This helps
1614 * keep threads from blocking, waiting for data from the client. */
1615 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1616 service = tcall->conn->service;
1617 if (!QuotaOK(service)) {
1620 MUTEX_ENTER(&rx_pthread_mutex);
1621 if (tno == rxi_fcfs_thread_num
1622 || !tcall->queue_item_header.next) {
1623 MUTEX_EXIT(&rx_pthread_mutex);
1624 /* If we're the fcfs thread , then we'll just use
1625 * this call. If we haven't been able to find an optimal
1626 * choice, and we're at the end of the list, then use a
1627 * 2d choice if one has been identified. Otherwise... */
1628 call = (choice2 ? choice2 : tcall);
1629 service = call->conn->service;
1631 MUTEX_EXIT(&rx_pthread_mutex);
1632 if (!queue_IsEmpty(&tcall->rq)) {
1633 struct rx_packet *rp;
1634 rp = queue_First(&tcall->rq, rx_packet);
1635 if (rp->header.seq == 1) {
1637 || (rp->header.flags & RX_LAST_PACKET)) {
1639 } else if (rxi_2dchoice && !choice2
1640 && !(tcall->flags & RX_CALL_CLEARED)
1641 && (tcall->rprev > rxi_HardAckRate)) {
1651 ReturnToServerPool(service);
1658 MUTEX_EXIT(&rx_serverPool_lock);
1659 MUTEX_ENTER(&call->lock);
1661 if (call->flags & RX_CALL_WAIT_PROC) {
1662 call->flags &= ~RX_CALL_WAIT_PROC;
1663 MUTEX_ENTER(&rx_waiting_mutex);
1665 MUTEX_EXIT(&rx_waiting_mutex);
1668 if (call->state != RX_STATE_PRECALL || call->error) {
1669 MUTEX_EXIT(&call->lock);
1670 MUTEX_ENTER(&rx_serverPool_lock);
1671 ReturnToServerPool(service);
1676 if (queue_IsEmpty(&call->rq)
1677 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1678 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1680 CLEAR_CALL_QUEUE_LOCK(call);
1683 /* If there are no eligible incoming calls, add this process
1684 * to the idle server queue, to wait for one */
1688 *socketp = OSI_NULLSOCKET;
1690 sq->socketp = socketp;
1691 queue_Append(&rx_idleServerQueue, sq);
1692 #ifndef AFS_AIX41_ENV
1693 rx_waitForPacket = sq;
1695 rx_waitingForPacket = sq;
1696 #endif /* AFS_AIX41_ENV */
1698 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1700 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1701 MUTEX_EXIT(&rx_serverPool_lock);
1702 return (struct rx_call *)0;
1705 } while (!(call = sq->newcall)
1706 && !(socketp && *socketp != OSI_NULLSOCKET));
1707 MUTEX_EXIT(&rx_serverPool_lock);
1709 MUTEX_ENTER(&call->lock);
1715 MUTEX_ENTER(&freeSQEList_lock);
1716 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1717 rx_FreeSQEList = sq;
1718 MUTEX_EXIT(&freeSQEList_lock);
1721 clock_GetTime(&call->startTime);
1722 call->state = RX_STATE_ACTIVE;
1723 call->mode = RX_MODE_RECEIVING;
1724 #ifdef RX_KERNEL_TRACE
1725 if (ICL_SETACTIVE(afs_iclSetp)) {
1726 int glockOwner = ISAFS_GLOCK();
1729 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1730 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1737 rxi_calltrace(RX_CALL_START, call);
1738 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1739 call->conn->service->servicePort, call->conn->service->serviceId,
1742 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1743 MUTEX_EXIT(&call->lock);
1745 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1750 #else /* RX_ENABLE_LOCKS */
1752 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1754 struct rx_serverQueueEntry *sq;
1755 struct rx_call *call = (struct rx_call *)0, *choice2;
1756 struct rx_service *service = NULL;
1760 MUTEX_ENTER(&freeSQEList_lock);
1762 if ((sq = rx_FreeSQEList)) {
1763 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1764 MUTEX_EXIT(&freeSQEList_lock);
1765 } else { /* otherwise allocate a new one and return that */
1766 MUTEX_EXIT(&freeSQEList_lock);
1767 sq = (struct rx_serverQueueEntry *)
1768 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1769 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1770 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1772 MUTEX_ENTER(&sq->lock);
1774 if (cur_service != NULL) {
1775 cur_service->nRequestsRunning--;
1776 if (cur_service->nRequestsRunning < cur_service->minProcs)
1780 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1781 struct rx_call *tcall, *ncall;
1782 /* Scan for eligible incoming calls. A call is not eligible
1783 * if the maximum number of calls for its service type are
1784 * already executing */
1785 /* One thread will process calls FCFS (to prevent starvation),
1786 * while the other threads may run ahead looking for calls which
1787 * have all their input data available immediately. This helps
1788 * keep threads from blocking, waiting for data from the client. */
1789 choice2 = (struct rx_call *)0;
1790 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1791 service = tcall->conn->service;
1792 if (QuotaOK(service)) {
1793 MUTEX_ENTER(&rx_pthread_mutex);
1794 if (tno == rxi_fcfs_thread_num
1795 || !tcall->queue_item_header.next) {
1796 MUTEX_EXIT(&rx_pthread_mutex);
1797 /* If we're the fcfs thread, then we'll just use
1798 * this call. If we haven't been able to find an optimal
1799 * choice, and we're at the end of the list, then use a
1800 * 2d choice if one has been identified. Otherwise... */
1801 call = (choice2 ? choice2 : tcall);
1802 service = call->conn->service;
1804 MUTEX_EXIT(&rx_pthread_mutex);
1805 if (!queue_IsEmpty(&tcall->rq)) {
1806 struct rx_packet *rp;
1807 rp = queue_First(&tcall->rq, rx_packet);
1808 if (rp->header.seq == 1
1810 || (rp->header.flags & RX_LAST_PACKET))) {
1812 } else if (rxi_2dchoice && !choice2
1813 && !(tcall->flags & RX_CALL_CLEARED)
1814 && (tcall->rprev > rxi_HardAckRate)) {
1828 /* we can't schedule a call if there's no data!!! */
1829 /* send an ack if there's no data, if we're missing the
1830 * first packet, or we're missing something between first
1831 * and last -- there's a "hole" in the incoming data. */
1832 if (queue_IsEmpty(&call->rq)
1833 || queue_First(&call->rq, rx_packet)->header.seq != 1
1834 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1835 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1837 call->flags &= (~RX_CALL_WAIT_PROC);
1838 service->nRequestsRunning++;
1839 /* just started call in minProcs pool, need fewer to maintain
1841 if (service->nRequestsRunning <= service->minProcs)
1845 /* MUTEX_EXIT(&call->lock); */
1847 /* If there are no eligible incoming calls, add this process
1848 * to the idle server queue, to wait for one */
1851 *socketp = OSI_NULLSOCKET;
1853 sq->socketp = socketp;
1854 queue_Append(&rx_idleServerQueue, sq);
1858 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1860 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1861 return (struct rx_call *)0;
1864 } while (!(call = sq->newcall)
1865 && !(socketp && *socketp != OSI_NULLSOCKET));
1867 MUTEX_EXIT(&sq->lock);
1869 MUTEX_ENTER(&freeSQEList_lock);
1870 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1871 rx_FreeSQEList = sq;
1872 MUTEX_EXIT(&freeSQEList_lock);
1875 clock_GetTime(&call->startTime);
1876 call->state = RX_STATE_ACTIVE;
1877 call->mode = RX_MODE_RECEIVING;
1878 #ifdef RX_KERNEL_TRACE
1879 if (ICL_SETACTIVE(afs_iclSetp)) {
1880 int glockOwner = ISAFS_GLOCK();
1883 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1884 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1891 rxi_calltrace(RX_CALL_START, call);
1892 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1893 call->conn->service->servicePort, call->conn->service->serviceId,
1896 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1903 #endif /* RX_ENABLE_LOCKS */
1907 /* Establish a procedure to be called when a packet arrives for a
1908 * call. This routine will be called at most once after each call,
1909 * and will also be called if there is an error condition on the or
1910 * the call is complete. Used by multi rx to build a selection
1911 * function which determines which of several calls is likely to be a
1912 * good one to read from.
1913 * NOTE: the way this is currently implemented it is probably only a
1914 * good idea to (1) use it immediately after a newcall (clients only)
1915 * and (2) only use it once. Other uses currently void your warranty
1918 rx_SetArrivalProc(struct rx_call *call,
1919 void (*proc) (struct rx_call * call,
1922 void * handle, int arg)
1924 call->arrivalProc = proc;
1925 call->arrivalProcHandle = handle;
1926 call->arrivalProcArg = arg;
1929 /* Call is finished (possibly prematurely). Return rc to the peer, if
1930 * appropriate, and return the final error code from the conversation
1934 rx_EndCall(struct rx_call *call, afs_int32 rc)
1936 struct rx_connection *conn = call->conn;
1937 struct rx_service *service;
1943 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1946 MUTEX_ENTER(&call->lock);
1948 if (rc == 0 && call->error == 0) {
1949 call->abortCode = 0;
1950 call->abortCount = 0;
1953 call->arrivalProc = (void (*)())0;
1954 if (rc && call->error == 0) {
1955 rxi_CallError(call, rc);
1956 /* Send an abort message to the peer if this error code has
1957 * only just been set. If it was set previously, assume the
1958 * peer has already been sent the error code or will request it
1960 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1962 if (conn->type == RX_SERVER_CONNECTION) {
1963 /* Make sure reply or at least dummy reply is sent */
1964 if (call->mode == RX_MODE_RECEIVING) {
1965 rxi_WriteProc(call, 0, 0);
1967 if (call->mode == RX_MODE_SENDING) {
1968 rxi_FlushWrite(call);
1970 service = conn->service;
1971 rxi_calltrace(RX_CALL_END, call);
1972 /* Call goes to hold state until reply packets are acknowledged */
1973 if (call->tfirst + call->nSoftAcked < call->tnext) {
1974 call->state = RX_STATE_HOLD;
1976 call->state = RX_STATE_DALLY;
1977 rxi_ClearTransmitQueue(call, 0);
1978 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1979 rxevent_Cancel(call->keepAliveEvent, call,
1980 RX_CALL_REFCOUNT_ALIVE);
1982 } else { /* Client connection */
1984 /* Make sure server receives input packets, in the case where
1985 * no reply arguments are expected */
1986 if ((call->mode == RX_MODE_SENDING)
1987 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1988 (void)rxi_ReadProc(call, &dummy, 1);
1991 /* If we had an outstanding delayed ack, be nice to the server
1992 * and force-send it now.
1994 if (call->delayedAckEvent) {
1995 rxevent_Cancel(call->delayedAckEvent, call,
1996 RX_CALL_REFCOUNT_DELAY);
1997 call->delayedAckEvent = NULL;
1998 rxi_SendDelayedAck(NULL, call, NULL);
2001 /* We need to release the call lock since it's lower than the
2002 * conn_call_lock and we don't want to hold the conn_call_lock
2003 * over the rx_ReadProc call. The conn_call_lock needs to be held
2004 * here for the case where rx_NewCall is perusing the calls on
2005 * the connection structure. We don't want to signal until
2006 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2007 * have checked this call, found it active and by the time it
2008 * goes to sleep, will have missed the signal.
2010 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
2011 * there are threads waiting to use the conn object.
2013 MUTEX_EXIT(&call->lock);
2014 MUTEX_ENTER(&conn->conn_call_lock);
2015 MUTEX_ENTER(&call->lock);
2016 MUTEX_ENTER(&conn->conn_data_lock);
2017 conn->flags |= RX_CONN_BUSY;
2018 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2019 if (conn->makeCallWaiters == 0)
2020 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
2021 MUTEX_EXIT(&conn->conn_data_lock);
2022 #ifdef RX_ENABLE_LOCKS
2023 CV_BROADCAST(&conn->conn_call_cv);
2028 #ifdef RX_ENABLE_LOCKS
2030 MUTEX_EXIT(&conn->conn_data_lock);
2032 #endif /* RX_ENABLE_LOCKS */
2033 call->state = RX_STATE_DALLY;
2035 error = call->error;
2037 /* currentPacket, nLeft, and NFree must be zeroed here, because
2038 * ResetCall cannot: ResetCall may be called at splnet(), in the
2039 * kernel version, and may interrupt the macros rx_Read or
2040 * rx_Write, which run at normal priority for efficiency. */
2041 if (call->currentPacket) {
2042 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2043 rxi_FreePacket(call->currentPacket);
2044 call->currentPacket = (struct rx_packet *)0;
2047 call->nLeft = call->nFree = call->curlen = 0;
2049 /* Free any packets from the last call to ReadvProc/WritevProc */
2050 #ifdef RXDEBUG_PACKET
2052 #endif /* RXDEBUG_PACKET */
2053 rxi_FreePackets(0, &call->iovq);
2055 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2056 MUTEX_EXIT(&call->lock);
2057 if (conn->type == RX_CLIENT_CONNECTION) {
2058 MUTEX_EXIT(&conn->conn_call_lock);
2059 conn->flags &= ~RX_CONN_BUSY;
2063 * Map errors to the local host's errno.h format.
2065 error = ntoh_syserr_conv(error);
2069 #if !defined(KERNEL)
2071 /* Call this routine when shutting down a server or client (especially
2072 * clients). This will allow Rx to gracefully garbage collect server
2073 * connections, and reduce the number of retries that a server might
2074 * make to a dead client.
2075 * This is not quite right, since some calls may still be ongoing and
2076 * we can't lock them to destroy them. */
2080 struct rx_connection **conn_ptr, **conn_end;
2084 if (rxinit_status == 1) {
2086 return; /* Already shutdown. */
2088 rxi_DeleteCachedConnections();
2089 if (rx_connHashTable) {
2090 MUTEX_ENTER(&rx_connHashTable_lock);
2091 for (conn_ptr = &rx_connHashTable[0], conn_end =
2092 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2094 struct rx_connection *conn, *next;
2095 for (conn = *conn_ptr; conn; conn = next) {
2097 if (conn->type == RX_CLIENT_CONNECTION) {
2098 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2100 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2101 #ifdef RX_ENABLE_LOCKS
2102 rxi_DestroyConnectionNoLock(conn);
2103 #else /* RX_ENABLE_LOCKS */
2104 rxi_DestroyConnection(conn);
2105 #endif /* RX_ENABLE_LOCKS */
2109 #ifdef RX_ENABLE_LOCKS
2110 while (rx_connCleanup_list) {
2111 struct rx_connection *conn;
2112 conn = rx_connCleanup_list;
2113 rx_connCleanup_list = rx_connCleanup_list->next;
2114 MUTEX_EXIT(&rx_connHashTable_lock);
2115 rxi_CleanupConnection(conn);
2116 MUTEX_ENTER(&rx_connHashTable_lock);
2118 MUTEX_EXIT(&rx_connHashTable_lock);
2119 #endif /* RX_ENABLE_LOCKS */
2124 afs_winsockCleanup();
2132 /* if we wakeup packet waiter too often, can get in loop with two
2133 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2135 rxi_PacketsUnWait(void)
2137 if (!rx_waitingForPackets) {
2141 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2142 return; /* still over quota */
2145 rx_waitingForPackets = 0;
2146 #ifdef RX_ENABLE_LOCKS
2147 CV_BROADCAST(&rx_waitingForPackets_cv);
2149 osi_rxWakeup(&rx_waitingForPackets);
2155 /* ------------------Internal interfaces------------------------- */
2157 /* Return this process's service structure for the
2158 * specified socket and service */
2160 rxi_FindService(osi_socket socket, u_short serviceId)
2162 struct rx_service **sp;
2163 for (sp = &rx_services[0]; *sp; sp++) {
2164 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2171 #ifdef KDUMP_RX_LOCK
2172 static struct rx_call_rx_lock *rx_allCallsp = 0;
2174 static struct rx_call *rx_allCallsp = 0;
2178 /* Allocate a call structure, for the indicated channel of the
2179 * supplied connection. The mode and state of the call must be set by
2180 * the caller. Returns the call with mutex locked. */
2182 rxi_NewCall(struct rx_connection *conn, int channel)
2184 struct rx_call *call;
2185 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2186 struct rx_call *cp; /* Call pointer temp */
2187 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2188 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2190 dpf(("rxi_NewCall(conn %x, channel %d)\n", conn, channel));
2192 /* Grab an existing call structure, or allocate a new one.
2193 * Existing call structures are assumed to have been left reset by
2195 MUTEX_ENTER(&rx_freeCallQueue_lock);
2197 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2199 * EXCEPT that the TQ might not yet be cleared out.
2200 * Skip over those with in-use TQs.
2203 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2204 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2210 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2211 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2212 call = queue_First(&rx_freeCallQueue, rx_call);
2213 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2215 if (rx_stats_active)
2216 rx_MutexDecrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2217 MUTEX_EXIT(&rx_freeCallQueue_lock);
2218 MUTEX_ENTER(&call->lock);
2219 CLEAR_CALL_QUEUE_LOCK(call);
2220 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2221 /* Now, if TQ wasn't cleared earlier, do it now. */
2222 if (call->flags & RX_CALL_TQ_CLEARME) {
2223 rxi_ClearTransmitQueue(call, 1);
2224 /*queue_Init(&call->tq);*/
2226 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2227 /* Bind the call to its connection structure */
2229 rxi_ResetCall(call, 1);
2232 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2233 #ifdef RXDEBUG_PACKET
2234 call->allNextp = rx_allCallsp;
2235 rx_allCallsp = call;
2237 #endif /* RXDEBUG_PACKET */
2238 rx_MutexIncrement(rx_stats.nCallStructs, rx_stats_mutex);
2240 MUTEX_EXIT(&rx_freeCallQueue_lock);
2241 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2242 MUTEX_ENTER(&call->lock);
2243 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2244 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2245 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2247 /* Initialize once-only items */
2248 queue_Init(&call->tq);
2249 queue_Init(&call->rq);
2250 queue_Init(&call->iovq);
2251 #ifdef RXDEBUG_PACKET
2252 call->rqc = call->tqc = call->iovqc = 0;
2253 #endif /* RXDEBUG_PACKET */
2254 /* Bind the call to its connection structure (prereq for reset) */
2256 rxi_ResetCall(call, 1);
2258 call->channel = channel;
2259 call->callNumber = &conn->callNumber[channel];
2260 call->rwind = conn->rwind[channel];
2261 call->twind = conn->twind[channel];
2262 /* Note that the next expected call number is retained (in
2263 * conn->callNumber[i]), even if we reallocate the call structure
2265 conn->call[channel] = call;
2266 /* if the channel's never been used (== 0), we should start at 1, otherwise
2267 * the call number is valid from the last time this channel was used */
2268 if (*call->callNumber == 0)
2269 *call->callNumber = 1;
2274 /* A call has been inactive long enough that so we can throw away
2275 * state, including the call structure, which is placed on the call
2277 * Call is locked upon entry.
2278 * haveCTLock set if called from rxi_ReapConnections
2280 #ifdef RX_ENABLE_LOCKS
2282 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2283 #else /* RX_ENABLE_LOCKS */
2285 rxi_FreeCall(struct rx_call *call)
2286 #endif /* RX_ENABLE_LOCKS */
2288 int channel = call->channel;
2289 struct rx_connection *conn = call->conn;
2292 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2293 (*call->callNumber)++;
2294 rxi_ResetCall(call, 0);
2295 call->conn->call[channel] = (struct rx_call *)0;
2297 MUTEX_ENTER(&rx_freeCallQueue_lock);
2298 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2299 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2300 /* A call may be free even though its transmit queue is still in use.
2301 * Since we search the call list from head to tail, put busy calls at
2302 * the head of the list, and idle calls at the tail.
2304 if (call->flags & RX_CALL_TQ_BUSY)
2305 queue_Prepend(&rx_freeCallQueue, call);
2307 queue_Append(&rx_freeCallQueue, call);
2308 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2309 queue_Append(&rx_freeCallQueue, call);
2310 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2311 if (rx_stats_active)
2312 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2313 MUTEX_EXIT(&rx_freeCallQueue_lock);
2315 /* Destroy the connection if it was previously slated for
2316 * destruction, i.e. the Rx client code previously called
2317 * rx_DestroyConnection (client connections), or
2318 * rxi_ReapConnections called the same routine (server
2319 * connections). Only do this, however, if there are no
2320 * outstanding calls. Note that for fine grain locking, there appears
2321 * to be a deadlock in that rxi_FreeCall has a call locked and
2322 * DestroyConnectionNoLock locks each call in the conn. But note a
2323 * few lines up where we have removed this call from the conn.
2324 * If someone else destroys a connection, they either have no
2325 * call lock held or are going through this section of code.
2327 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2328 MUTEX_ENTER(&conn->conn_data_lock);
2330 MUTEX_EXIT(&conn->conn_data_lock);
2331 #ifdef RX_ENABLE_LOCKS
2333 rxi_DestroyConnectionNoLock(conn);
2335 rxi_DestroyConnection(conn);
2336 #else /* RX_ENABLE_LOCKS */
2337 rxi_DestroyConnection(conn);
2338 #endif /* RX_ENABLE_LOCKS */
2342 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2344 rxi_Alloc(size_t size)
2348 if (rx_stats_active)
2349 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2352 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2353 afs_osi_Alloc_NoSleep(size);
2358 osi_Panic("rxi_Alloc error");
2364 rxi_Free(void *addr, size_t size)
2366 if (rx_stats_active)
2367 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2368 osi_Free(addr, size);
2372 rxi_SetPeerMtu(afs_uint32 host, afs_uint32 port, int mtu)
2374 struct rx_peer **peer_ptr, **peer_end;
2377 MUTEX_ENTER(&rx_peerHashTable_lock);
2379 for (peer_ptr = &rx_peerHashTable[0], peer_end =
2380 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
2382 struct rx_peer *peer, *next;
2383 for (peer = *peer_ptr; peer; peer = next) {
2385 if (host == peer->host) {
2386 MUTEX_ENTER(&peer->peer_lock);
2387 peer->ifMTU=MIN(mtu, peer->ifMTU);
2388 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2389 MUTEX_EXIT(&peer->peer_lock);
2394 struct rx_peer *peer;
2395 hashIndex = PEER_HASH(host, port);
2396 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2397 if ((peer->host == host) && (peer->port == port)) {
2398 MUTEX_ENTER(&peer->peer_lock);
2399 peer->ifMTU=MIN(mtu, peer->ifMTU);
2400 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2401 MUTEX_EXIT(&peer->peer_lock);
2405 MUTEX_EXIT(&rx_peerHashTable_lock);
2408 /* Find the peer process represented by the supplied (host,port)
2409 * combination. If there is no appropriate active peer structure, a
2410 * new one will be allocated and initialized
2411 * The origPeer, if set, is a pointer to a peer structure on which the
2412 * refcount will be be decremented. This is used to replace the peer
2413 * structure hanging off a connection structure */
2415 rxi_FindPeer(afs_uint32 host, u_short port,
2416 struct rx_peer *origPeer, int create)
2420 hashIndex = PEER_HASH(host, port);
2421 MUTEX_ENTER(&rx_peerHashTable_lock);
2422 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2423 if ((pp->host == host) && (pp->port == port))
2428 pp = rxi_AllocPeer(); /* This bzero's *pp */
2429 pp->host = host; /* set here or in InitPeerParams is zero */
2431 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2432 queue_Init(&pp->congestionQueue);
2433 queue_Init(&pp->rpcStats);
2434 pp->next = rx_peerHashTable[hashIndex];
2435 rx_peerHashTable[hashIndex] = pp;
2436 rxi_InitPeerParams(pp);
2437 if (rx_stats_active)
2438 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2445 origPeer->refCount--;
2446 MUTEX_EXIT(&rx_peerHashTable_lock);
2451 /* Find the connection at (host, port) started at epoch, and with the
2452 * given connection id. Creates the server connection if necessary.
2453 * The type specifies whether a client connection or a server
2454 * connection is desired. In both cases, (host, port) specify the
2455 * peer's (host, pair) pair. Client connections are not made
2456 * automatically by this routine. The parameter socket gives the
2457 * socket descriptor on which the packet was received. This is used,
2458 * in the case of server connections, to check that *new* connections
2459 * come via a valid (port, serviceId). Finally, the securityIndex
2460 * parameter must match the existing index for the connection. If a
2461 * server connection is created, it will be created using the supplied
2462 * index, if the index is valid for this service */
2463 struct rx_connection *
2464 rxi_FindConnection(osi_socket socket, afs_int32 host,
2465 u_short port, u_short serviceId, afs_uint32 cid,
2466 afs_uint32 epoch, int type, u_int securityIndex)
2468 int hashindex, flag, i;
2469 struct rx_connection *conn;
2470 hashindex = CONN_HASH(host, port, cid, epoch, type);
2471 MUTEX_ENTER(&rx_connHashTable_lock);
2472 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2473 rx_connHashTable[hashindex],
2476 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2477 && (epoch == conn->epoch)) {
2478 struct rx_peer *pp = conn->peer;
2479 if (securityIndex != conn->securityIndex) {
2480 /* this isn't supposed to happen, but someone could forge a packet
2481 * like this, and there seems to be some CM bug that makes this
2482 * happen from time to time -- in which case, the fileserver
2484 MUTEX_EXIT(&rx_connHashTable_lock);
2485 return (struct rx_connection *)0;
2487 if (pp->host == host && pp->port == port)
2489 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2491 /* So what happens when it's a callback connection? */
2492 if ( /*type == RX_CLIENT_CONNECTION && */
2493 (conn->epoch & 0x80000000))
2497 /* the connection rxLastConn that was used the last time is not the
2498 ** one we are looking for now. Hence, start searching in the hash */
2500 conn = rx_connHashTable[hashindex];
2505 struct rx_service *service;
2506 if (type == RX_CLIENT_CONNECTION) {
2507 MUTEX_EXIT(&rx_connHashTable_lock);
2508 return (struct rx_connection *)0;
2510 service = rxi_FindService(socket, serviceId);
2511 if (!service || (securityIndex >= service->nSecurityObjects)
2512 || (service->securityObjects[securityIndex] == 0)) {
2513 MUTEX_EXIT(&rx_connHashTable_lock);
2514 return (struct rx_connection *)0;
2516 conn = rxi_AllocConnection(); /* This bzero's the connection */
2517 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2518 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2519 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2520 conn->next = rx_connHashTable[hashindex];
2521 rx_connHashTable[hashindex] = conn;
2522 conn->peer = rxi_FindPeer(host, port, 0, 1);
2523 conn->type = RX_SERVER_CONNECTION;
2524 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2525 conn->epoch = epoch;
2526 conn->cid = cid & RX_CIDMASK;
2527 /* conn->serial = conn->lastSerial = 0; */
2528 /* conn->timeout = 0; */
2529 conn->ackRate = RX_FAST_ACK_RATE;
2530 conn->service = service;
2531 conn->serviceId = serviceId;
2532 conn->securityIndex = securityIndex;
2533 conn->securityObject = service->securityObjects[securityIndex];
2534 conn->nSpecific = 0;
2535 conn->specific = NULL;
2536 rx_SetConnDeadTime(conn, service->connDeadTime);
2537 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2538 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2539 for (i = 0; i < RX_MAXCALLS; i++) {
2540 conn->twind[i] = rx_initSendWindow;
2541 conn->rwind[i] = rx_initReceiveWindow;
2543 /* Notify security object of the new connection */
2544 RXS_NewConnection(conn->securityObject, conn);
2545 /* XXXX Connection timeout? */
2546 if (service->newConnProc)
2547 (*service->newConnProc) (conn);
2548 if (rx_stats_active)
2549 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2552 MUTEX_ENTER(&conn->conn_data_lock);
2554 MUTEX_EXIT(&conn->conn_data_lock);
2556 rxLastConn = conn; /* store this connection as the last conn used */
2557 MUTEX_EXIT(&rx_connHashTable_lock);
2561 /* There are two packet tracing routines available for testing and monitoring
2562 * Rx. One is called just after every packet is received and the other is
2563 * called just before every packet is sent. Received packets, have had their
2564 * headers decoded, and packets to be sent have not yet had their headers
2565 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2566 * containing the network address. Both can be modified. The return value, if
2567 * non-zero, indicates that the packet should be dropped. */
2569 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2570 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2572 /* A packet has been received off the interface. Np is the packet, socket is
2573 * the socket number it was received from (useful in determining which service
2574 * this packet corresponds to), and (host, port) reflect the host,port of the
2575 * sender. This call returns the packet to the caller if it is finished with
2576 * it, rather than de-allocating it, just as a small performance hack */
2579 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
2580 afs_uint32 host, u_short port, int *tnop,
2581 struct rx_call **newcallp)
2583 struct rx_call *call;
2584 struct rx_connection *conn;
2586 afs_uint32 currentCallNumber;
2592 struct rx_packet *tnp;
2595 /* We don't print out the packet until now because (1) the time may not be
2596 * accurate enough until now in the lwp implementation (rx_Listener only gets
2597 * the time after the packet is read) and (2) from a protocol point of view,
2598 * this is the first time the packet has been seen */
2599 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2600 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2601 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2602 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2603 np->header.epoch, np->header.cid, np->header.callNumber,
2604 np->header.seq, np->header.flags, np));
2607 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2608 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2611 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2612 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2615 /* If an input tracer function is defined, call it with the packet and
2616 * network address. Note this function may modify its arguments. */
2617 if (rx_justReceived) {
2618 struct sockaddr_in addr;
2620 addr.sin_family = AF_INET;
2621 addr.sin_port = port;
2622 addr.sin_addr.s_addr = host;
2623 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2624 addr.sin_len = sizeof(addr);
2625 #endif /* AFS_OSF_ENV */
2626 drop = (*rx_justReceived) (np, &addr);
2627 /* drop packet if return value is non-zero */
2630 port = addr.sin_port; /* in case fcn changed addr */
2631 host = addr.sin_addr.s_addr;
2635 /* If packet was not sent by the client, then *we* must be the client */
2636 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2637 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2639 /* Find the connection (or fabricate one, if we're the server & if
2640 * necessary) associated with this packet */
2642 rxi_FindConnection(socket, host, port, np->header.serviceId,
2643 np->header.cid, np->header.epoch, type,
2644 np->header.securityIndex);
2647 /* If no connection found or fabricated, just ignore the packet.
2648 * (An argument could be made for sending an abort packet for
2653 MUTEX_ENTER(&conn->conn_data_lock);
2654 if (conn->maxSerial < np->header.serial)
2655 conn->maxSerial = np->header.serial;
2656 MUTEX_EXIT(&conn->conn_data_lock);
2658 /* If the connection is in an error state, send an abort packet and ignore
2659 * the incoming packet */
2661 /* Don't respond to an abort packet--we don't want loops! */
2662 MUTEX_ENTER(&conn->conn_data_lock);
2663 if (np->header.type != RX_PACKET_TYPE_ABORT)
2664 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2666 MUTEX_EXIT(&conn->conn_data_lock);
2670 /* Check for connection-only requests (i.e. not call specific). */
2671 if (np->header.callNumber == 0) {
2672 switch (np->header.type) {
2673 case RX_PACKET_TYPE_ABORT: {
2674 /* What if the supplied error is zero? */
2675 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2676 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2677 rxi_ConnectionError(conn, errcode);
2678 MUTEX_ENTER(&conn->conn_data_lock);
2680 MUTEX_EXIT(&conn->conn_data_lock);
2683 case RX_PACKET_TYPE_CHALLENGE:
2684 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2685 MUTEX_ENTER(&conn->conn_data_lock);
2687 MUTEX_EXIT(&conn->conn_data_lock);
2689 case RX_PACKET_TYPE_RESPONSE:
2690 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2691 MUTEX_ENTER(&conn->conn_data_lock);
2693 MUTEX_EXIT(&conn->conn_data_lock);
2695 case RX_PACKET_TYPE_PARAMS:
2696 case RX_PACKET_TYPE_PARAMS + 1:
2697 case RX_PACKET_TYPE_PARAMS + 2:
2698 /* ignore these packet types for now */
2699 MUTEX_ENTER(&conn->conn_data_lock);
2701 MUTEX_EXIT(&conn->conn_data_lock);
2706 /* Should not reach here, unless the peer is broken: send an
2708 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2709 MUTEX_ENTER(&conn->conn_data_lock);
2710 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2712 MUTEX_EXIT(&conn->conn_data_lock);
2717 channel = np->header.cid & RX_CHANNELMASK;
2718 call = conn->call[channel];
2719 #ifdef RX_ENABLE_LOCKS
2721 MUTEX_ENTER(&call->lock);
2722 /* Test to see if call struct is still attached to conn. */
2723 if (call != conn->call[channel]) {
2725 MUTEX_EXIT(&call->lock);
2726 if (type == RX_SERVER_CONNECTION) {
2727 call = conn->call[channel];
2728 /* If we started with no call attached and there is one now,
2729 * another thread is also running this routine and has gotten
2730 * the connection channel. We should drop this packet in the tests
2731 * below. If there was a call on this connection and it's now
2732 * gone, then we'll be making a new call below.
2733 * If there was previously a call and it's now different then
2734 * the old call was freed and another thread running this routine
2735 * has created a call on this channel. One of these two threads
2736 * has a packet for the old call and the code below handles those
2740 MUTEX_ENTER(&call->lock);
2742 /* This packet can't be for this call. If the new call address is
2743 * 0 then no call is running on this channel. If there is a call
2744 * then, since this is a client connection we're getting data for
2745 * it must be for the previous call.
2747 if (rx_stats_active)
2748 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2749 MUTEX_ENTER(&conn->conn_data_lock);
2751 MUTEX_EXIT(&conn->conn_data_lock);
2756 currentCallNumber = conn->callNumber[channel];
2758 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2759 if (np->header.callNumber < currentCallNumber) {
2760 if (rx_stats_active)
2761 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2762 #ifdef RX_ENABLE_LOCKS
2764 MUTEX_EXIT(&call->lock);
2766 MUTEX_ENTER(&conn->conn_data_lock);
2768 MUTEX_EXIT(&conn->conn_data_lock);
2772 MUTEX_ENTER(&conn->conn_call_lock);
2773 call = rxi_NewCall(conn, channel);
2774 MUTEX_EXIT(&conn->conn_call_lock);
2775 *call->callNumber = np->header.callNumber;
2777 if (np->header.callNumber == 0)
2778 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %lx resend %d.%0.3d len %d", np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port), np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq, np->header.flags, (unsigned long)np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2780 call->state = RX_STATE_PRECALL;
2781 clock_GetTime(&call->queueTime);
2782 hzero(call->bytesSent);
2783 hzero(call->bytesRcvd);
2785 * If the number of queued calls exceeds the overload
2786 * threshold then abort this call.
2788 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2789 struct rx_packet *tp;
2791 rxi_CallError(call, rx_BusyError);
2792 tp = rxi_SendCallAbort(call, np, 1, 0);
2793 MUTEX_EXIT(&call->lock);
2794 MUTEX_ENTER(&conn->conn_data_lock);
2796 MUTEX_EXIT(&conn->conn_data_lock);
2797 if (rx_stats_active)
2798 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2801 rxi_KeepAliveOn(call);
2802 } else if (np->header.callNumber != currentCallNumber) {
2803 /* Wait until the transmit queue is idle before deciding
2804 * whether to reset the current call. Chances are that the
2805 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2808 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2809 while ((call->state == RX_STATE_ACTIVE)
2810 && (call->flags & RX_CALL_TQ_BUSY)) {
2811 call->flags |= RX_CALL_TQ_WAIT;
2813 #ifdef RX_ENABLE_LOCKS
2814 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2815 CV_WAIT(&call->cv_tq, &call->lock);
2816 #else /* RX_ENABLE_LOCKS */
2817 osi_rxSleep(&call->tq);
2818 #endif /* RX_ENABLE_LOCKS */
2820 if (call->tqWaiters == 0)
2821 call->flags &= ~RX_CALL_TQ_WAIT;
2823 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2824 /* If the new call cannot be taken right now send a busy and set
2825 * the error condition in this call, so that it terminates as
2826 * quickly as possible */
2827 if (call->state == RX_STATE_ACTIVE) {
2828 struct rx_packet *tp;
2830 rxi_CallError(call, RX_CALL_DEAD);
2831 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2833 MUTEX_EXIT(&call->lock);
2834 MUTEX_ENTER(&conn->conn_data_lock);
2836 MUTEX_EXIT(&conn->conn_data_lock);
2839 rxi_ResetCall(call, 0);
2840 *call->callNumber = np->header.callNumber;
2842 if (np->header.callNumber == 0)
2843 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %lx resend %d.%0.3d len %d", np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port), np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq, np->header.flags, (unsigned long)np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2845 call->state = RX_STATE_PRECALL;
2846 clock_GetTime(&call->queueTime);
2847 hzero(call->bytesSent);
2848 hzero(call->bytesRcvd);
2850 * If the number of queued calls exceeds the overload
2851 * threshold then abort this call.
2853 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2854 struct rx_packet *tp;
2856 rxi_CallError(call, rx_BusyError);
2857 tp = rxi_SendCallAbort(call, np, 1, 0);
2858 MUTEX_EXIT(&call->lock);
2859 MUTEX_ENTER(&conn->conn_data_lock);
2861 MUTEX_EXIT(&conn->conn_data_lock);
2862 if (rx_stats_active)
2863 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2866 rxi_KeepAliveOn(call);
2868 /* Continuing call; do nothing here. */
2870 } else { /* we're the client */
2871 /* Ignore all incoming acknowledgements for calls in DALLY state */
2872 if (call && (call->state == RX_STATE_DALLY)
2873 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2874 if (rx_stats_active)
2875 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
2876 #ifdef RX_ENABLE_LOCKS
2878 MUTEX_EXIT(&call->lock);
2881 MUTEX_ENTER(&conn->conn_data_lock);
2883 MUTEX_EXIT(&conn->conn_data_lock);
2887 /* Ignore anything that's not relevant to the current call. If there
2888 * isn't a current call, then no packet is relevant. */
2889 if (!call || (np->header.callNumber != currentCallNumber)) {
2890 if (rx_stats_active)
2891 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2892 #ifdef RX_ENABLE_LOCKS
2894 MUTEX_EXIT(&call->lock);
2897 MUTEX_ENTER(&conn->conn_data_lock);
2899 MUTEX_EXIT(&conn->conn_data_lock);
2902 /* If the service security object index stamped in the packet does not
2903 * match the connection's security index, ignore the packet */
2904 if (np->header.securityIndex != conn->securityIndex) {
2905 #ifdef RX_ENABLE_LOCKS
2906 MUTEX_EXIT(&call->lock);
2908 MUTEX_ENTER(&conn->conn_data_lock);
2910 MUTEX_EXIT(&conn->conn_data_lock);
2914 /* If we're receiving the response, then all transmit packets are
2915 * implicitly acknowledged. Get rid of them. */
2916 if (np->header.type == RX_PACKET_TYPE_DATA) {
2917 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2918 /* XXX Hack. Because we must release the global rx lock when
2919 * sending packets (osi_NetSend) we drop all acks while we're
2920 * traversing the tq in rxi_Start sending packets out because
2921 * packets may move to the freePacketQueue as result of being here!
2922 * So we drop these packets until we're safely out of the
2923 * traversing. Really ugly!
2924 * For fine grain RX locking, we set the acked field in the
2925 * packets and let rxi_Start remove them from the transmit queue.
2927 if (call->flags & RX_CALL_TQ_BUSY) {
2928 #ifdef RX_ENABLE_LOCKS
2929 rxi_SetAcksInTransmitQueue(call);
2932 return np; /* xmitting; drop packet */
2935 rxi_ClearTransmitQueue(call, 0);
2937 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2938 rxi_ClearTransmitQueue(call, 0);
2939 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2941 if (np->header.type == RX_PACKET_TYPE_ACK) {
2942 /* now check to see if this is an ack packet acknowledging that the
2943 * server actually *lost* some hard-acked data. If this happens we
2944 * ignore this packet, as it may indicate that the server restarted in
2945 * the middle of a call. It is also possible that this is an old ack
2946 * packet. We don't abort the connection in this case, because this
2947 * *might* just be an old ack packet. The right way to detect a server
2948 * restart in the midst of a call is to notice that the server epoch
2950 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2951 * XXX unacknowledged. I think that this is off-by-one, but
2952 * XXX I don't dare change it just yet, since it will
2953 * XXX interact badly with the server-restart detection
2954 * XXX code in receiveackpacket. */
2955 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2956 if (rx_stats_active)
2957 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2958 MUTEX_EXIT(&call->lock);
2959 MUTEX_ENTER(&conn->conn_data_lock);
2961 MUTEX_EXIT(&conn->conn_data_lock);
2965 } /* else not a data packet */
2968 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2969 /* Set remote user defined status from packet */
2970 call->remoteStatus = np->header.userStatus;
2972 /* Note the gap between the expected next packet and the actual
2973 * packet that arrived, when the new packet has a smaller serial number
2974 * than expected. Rioses frequently reorder packets all by themselves,
2975 * so this will be quite important with very large window sizes.
2976 * Skew is checked against 0 here to avoid any dependence on the type of
2977 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2979 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2980 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2981 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2983 MUTEX_ENTER(&conn->conn_data_lock);
2984 skew = conn->lastSerial - np->header.serial;
2985 conn->lastSerial = np->header.serial;
2986 MUTEX_EXIT(&conn->conn_data_lock);
2988 struct rx_peer *peer;
2990 if (skew > peer->inPacketSkew) {
2991 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2993 peer->inPacketSkew = skew;
2997 /* Now do packet type-specific processing */
2998 switch (np->header.type) {
2999 case RX_PACKET_TYPE_DATA:
3000 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3003 case RX_PACKET_TYPE_ACK:
3004 /* Respond immediately to ack packets requesting acknowledgement
3006 if (np->header.flags & RX_REQUEST_ACK) {
3008 (void)rxi_SendCallAbort(call, 0, 1, 0);
3010 (void)rxi_SendAck(call, 0, np->header.serial,
3011 RX_ACK_PING_RESPONSE, 1);
3013 np = rxi_ReceiveAckPacket(call, np, 1);
3015 case RX_PACKET_TYPE_ABORT: {
3016 /* An abort packet: reset the call, passing the error up to the user. */
3017 /* What if error is zero? */
3018 /* What if the error is -1? the application will treat it as a timeout. */
3019 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3020 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
3021 rxi_CallError(call, errdata);
3022 MUTEX_EXIT(&call->lock);
3023 MUTEX_ENTER(&conn->conn_data_lock);
3025 MUTEX_EXIT(&conn->conn_data_lock);
3026 return np; /* xmitting; drop packet */
3028 case RX_PACKET_TYPE_BUSY:
3031 case RX_PACKET_TYPE_ACKALL:
3032 /* All packets acknowledged, so we can drop all packets previously
3033 * readied for sending */
3034 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3035 /* XXX Hack. We because we can't release the global rx lock when
3036 * sending packets (osi_NetSend) we drop all ack pkts while we're
3037 * traversing the tq in rxi_Start sending packets out because
3038 * packets may move to the freePacketQueue as result of being
3039 * here! So we drop these packets until we're safely out of the
3040 * traversing. Really ugly!
3041 * For fine grain RX locking, we set the acked field in the packets
3042 * and let rxi_Start remove the packets from the transmit queue.
3044 if (call->flags & RX_CALL_TQ_BUSY) {
3045 #ifdef RX_ENABLE_LOCKS
3046 rxi_SetAcksInTransmitQueue(call);
3048 #else /* RX_ENABLE_LOCKS */
3049 MUTEX_EXIT(&call->lock);
3050 MUTEX_ENTER(&conn->conn_data_lock);
3052 MUTEX_EXIT(&conn->conn_data_lock);
3053 return np; /* xmitting; drop packet */
3054 #endif /* RX_ENABLE_LOCKS */
3056 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3057 rxi_ClearTransmitQueue(call, 0);
3058 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3061 /* Should not reach here, unless the peer is broken: send an abort
3063 rxi_CallError(call, RX_PROTOCOL_ERROR);
3064 np = rxi_SendCallAbort(call, np, 1, 0);
3067 /* Note when this last legitimate packet was received, for keep-alive
3068 * processing. Note, we delay getting the time until now in the hope that
3069 * the packet will be delivered to the user before any get time is required
3070 * (if not, then the time won't actually be re-evaluated here). */
3071 call->lastReceiveTime = clock_Sec();
3072 MUTEX_EXIT(&call->lock);
3073 MUTEX_ENTER(&conn->conn_data_lock);
3075 MUTEX_EXIT(&conn->conn_data_lock);
3079 /* return true if this is an "interesting" connection from the point of view
3080 of someone trying to debug the system */
3082 rxi_IsConnInteresting(struct rx_connection *aconn)
3085 struct rx_call *tcall;
3087 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3089 for (i = 0; i < RX_MAXCALLS; i++) {
3090 tcall = aconn->call[i];
3092 if ((tcall->state == RX_STATE_PRECALL)
3093 || (tcall->state == RX_STATE_ACTIVE))
3095 if ((tcall->mode == RX_MODE_SENDING)
3096 || (tcall->mode == RX_MODE_RECEIVING))
3104 /* if this is one of the last few packets AND it wouldn't be used by the
3105 receiving call to immediately satisfy a read request, then drop it on
3106 the floor, since accepting it might prevent a lock-holding thread from
3107 making progress in its reading. If a call has been cleared while in
3108 the precall state then ignore all subsequent packets until the call
3109 is assigned to a thread. */
3112 TooLow(struct rx_packet *ap, struct rx_call *acall)
3116 MUTEX_ENTER(&rx_quota_mutex);
3117 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3118 && (acall->state == RX_STATE_PRECALL))
3119 || ((rx_nFreePackets < rxi_dataQuota + 2)
3120 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3121 && (acall->flags & RX_CALL_READER_WAIT)))) {
3124 MUTEX_EXIT(&rx_quota_mutex);
3130 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3132 struct rx_connection *conn = arg1;
3133 struct rx_call *acall = arg2;
3134 struct rx_call *call = acall;
3135 struct clock when, now;
3138 MUTEX_ENTER(&conn->conn_data_lock);
3139 conn->checkReachEvent = NULL;
3140 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3143 MUTEX_EXIT(&conn->conn_data_lock);
3147 MUTEX_ENTER(&conn->conn_call_lock);
3148 MUTEX_ENTER(&conn->conn_data_lock);
3149 for (i = 0; i < RX_MAXCALLS; i++) {
3150 struct rx_call *tc = conn->call[i];
3151 if (tc && tc->state == RX_STATE_PRECALL) {
3157 /* Indicate that rxi_CheckReachEvent is no longer running by
3158 * clearing the flag. Must be atomic under conn_data_lock to
3159 * avoid a new call slipping by: rxi_CheckConnReach holds
3160 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3162 conn->flags &= ~RX_CONN_ATTACHWAIT;
3163 MUTEX_EXIT(&conn->conn_data_lock);
3164 MUTEX_EXIT(&conn->conn_call_lock);
3169 MUTEX_ENTER(&call->lock);
3170 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3172 MUTEX_EXIT(&call->lock);
3174 clock_GetTime(&now);
3176 when.sec += RX_CHECKREACH_TIMEOUT;
3177 MUTEX_ENTER(&conn->conn_data_lock);
3178 if (!conn->checkReachEvent) {
3180 conn->checkReachEvent =
3181 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3184 MUTEX_EXIT(&conn->conn_data_lock);
3190 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3192 struct rx_service *service = conn->service;
3193 struct rx_peer *peer = conn->peer;
3194 afs_uint32 now, lastReach;
3196 if (service->checkReach == 0)
3200 MUTEX_ENTER(&peer->peer_lock);
3201 lastReach = peer->lastReachTime;
3202 MUTEX_EXIT(&peer->peer_lock);
3203 if (now - lastReach < RX_CHECKREACH_TTL)
3206 MUTEX_ENTER(&conn->conn_data_lock);
3207 if (conn->flags & RX_CONN_ATTACHWAIT) {
3208 MUTEX_EXIT(&conn->conn_data_lock);
3211 conn->flags |= RX_CONN_ATTACHWAIT;
3212 MUTEX_EXIT(&conn->conn_data_lock);
3213 if (!conn->checkReachEvent)
3214 rxi_CheckReachEvent(NULL, conn, call);
3219 /* try to attach call, if authentication is complete */
3221 TryAttach(struct rx_call *acall, osi_socket socket,
3222 int *tnop, struct rx_call **newcallp,
3225 struct rx_connection *conn = acall->conn;
3227 if (conn->type == RX_SERVER_CONNECTION
3228 && acall->state == RX_STATE_PRECALL) {
3229 /* Don't attach until we have any req'd. authentication. */
3230 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3231 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3232 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3233 /* Note: this does not necessarily succeed; there
3234 * may not any proc available
3237 rxi_ChallengeOn(acall->conn);
3242 /* A data packet has been received off the interface. This packet is
3243 * appropriate to the call (the call is in the right state, etc.). This
3244 * routine can return a packet to the caller, for re-use */
3247 rxi_ReceiveDataPacket(struct rx_call *call,
3248 struct rx_packet *np, int istack,
3249 osi_socket socket, afs_uint32 host, u_short port,
3250 int *tnop, struct rx_call **newcallp)
3252 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3257 afs_uint32 serial=0, flags=0;
3259 struct rx_packet *tnp;
3260 struct clock when, now;
3261 if (rx_stats_active)
3262 rx_MutexIncrement(rx_stats.dataPacketsRead, rx_stats_mutex);
3265 /* If there are no packet buffers, drop this new packet, unless we can find
3266 * packet buffers from inactive calls */
3268 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3269 MUTEX_ENTER(&rx_freePktQ_lock);
3270 rxi_NeedMorePackets = TRUE;
3271 MUTEX_EXIT(&rx_freePktQ_lock);
3272 if (rx_stats_active)
3273 rx_MutexIncrement(rx_stats.noPacketBuffersOnRead, rx_stats_mutex);
3274 call->rprev = np->header.serial;
3275 rxi_calltrace(RX_TRACE_DROP, call);
3276 dpf(("packet %x dropped on receipt - quota problems", np));
3278 rxi_ClearReceiveQueue(call);
3279 clock_GetTime(&now);
3281 clock_Add(&when, &rx_softAckDelay);
3282 if (!call->delayedAckEvent
3283 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3284 rxevent_Cancel(call->delayedAckEvent, call,
3285 RX_CALL_REFCOUNT_DELAY);
3286 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3287 call->delayedAckEvent =
3288 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3290 /* we've damaged this call already, might as well do it in. */
3296 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3297 * packet is one of several packets transmitted as a single
3298 * datagram. Do not send any soft or hard acks until all packets
3299 * in a jumbogram have been processed. Send negative acks right away.
3301 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3302 /* tnp is non-null when there are more packets in the
3303 * current jumbo gram */
3310 seq = np->header.seq;
3311 serial = np->header.serial;
3312 flags = np->header.flags;
3314 /* If the call is in an error state, send an abort message */
3316 return rxi_SendCallAbort(call, np, istack, 0);
3318 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3319 * AFS 3.5 jumbogram. */
3320 if (flags & RX_JUMBO_PACKET) {
3321 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3326 if (np->header.spare != 0) {
3327 MUTEX_ENTER(&call->conn->conn_data_lock);
3328 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3329 MUTEX_EXIT(&call->conn->conn_data_lock);
3332 /* The usual case is that this is the expected next packet */
3333 if (seq == call->rnext) {
3335 /* Check to make sure it is not a duplicate of one already queued */
3336 if (queue_IsNotEmpty(&call->rq)
3337 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3338 if (rx_stats_active)
3339 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3340 dpf(("packet %x dropped on receipt - duplicate", np));
3341 rxevent_Cancel(call->delayedAckEvent, call,
3342 RX_CALL_REFCOUNT_DELAY);
3343 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3349 /* It's the next packet. Stick it on the receive queue
3350 * for this call. Set newPackets to make sure we wake
3351 * the reader once all packets have been processed */
3352 np->flags |= RX_PKTFLAG_RQ;
3353 queue_Prepend(&call->rq, np);
3354 #ifdef RXDEBUG_PACKET
3356 #endif /* RXDEBUG_PACKET */
3358 np = NULL; /* We can't use this anymore */
3361 /* If an ack is requested then set a flag to make sure we
3362 * send an acknowledgement for this packet */
3363 if (flags & RX_REQUEST_ACK) {
3364 ackNeeded = RX_ACK_REQUESTED;
3367 /* Keep track of whether we have received the last packet */
3368 if (flags & RX_LAST_PACKET) {
3369 call->flags |= RX_CALL_HAVE_LAST;
3373 /* Check whether we have all of the packets for this call */
3374 if (call->flags & RX_CALL_HAVE_LAST) {
3375 afs_uint32 tseq; /* temporary sequence number */
3376 struct rx_packet *tp; /* Temporary packet pointer */
3377 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3379 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3380 if (tseq != tp->header.seq)
3382 if (tp->header.flags & RX_LAST_PACKET) {
3383 call->flags |= RX_CALL_RECEIVE_DONE;
3390 /* Provide asynchronous notification for those who want it
3391 * (e.g. multi rx) */
3392 if (call->arrivalProc) {
3393 (*call->arrivalProc) (call, call->arrivalProcHandle,
3394 call->arrivalProcArg);
3395 call->arrivalProc = (void (*)())0;
3398 /* Update last packet received */
3401 /* If there is no server process serving this call, grab
3402 * one, if available. We only need to do this once. If a
3403 * server thread is available, this thread becomes a server
3404 * thread and the server thread becomes a listener thread. */
3406 TryAttach(call, socket, tnop, newcallp, 0);
3409 /* This is not the expected next packet. */
3411 /* Determine whether this is a new or old packet, and if it's
3412 * a new one, whether it fits into the current receive window.
3413 * Also figure out whether the packet was delivered in sequence.
3414 * We use the prev variable to determine whether the new packet
3415 * is the successor of its immediate predecessor in the
3416 * receive queue, and the missing flag to determine whether
3417 * any of this packets predecessors are missing. */
3419 afs_uint32 prev; /* "Previous packet" sequence number */
3420 struct rx_packet *tp; /* Temporary packet pointer */
3421 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3422 int missing; /* Are any predecessors missing? */
3424 /* If the new packet's sequence number has been sent to the
3425 * application already, then this is a duplicate */
3426 if (seq < call->rnext) {
3427 if (rx_stats_active)
3428 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3429 rxevent_Cancel(call->delayedAckEvent, call,
3430 RX_CALL_REFCOUNT_DELAY);
3431 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3437 /* If the sequence number is greater than what can be
3438 * accomodated by the current window, then send a negative
3439 * acknowledge and drop the packet */
3440 if ((call->rnext + call->rwind) <= seq) {
3441 rxevent_Cancel(call->delayedAckEvent, call,
3442 RX_CALL_REFCOUNT_DELAY);
3443 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3450 /* Look for the packet in the queue of old received packets */
3451 for (prev = call->rnext - 1, missing =
3452 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3453 /*Check for duplicate packet */
3454 if (seq == tp->header.seq) {
3455 if (rx_stats_active)
3456 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3457 rxevent_Cancel(call->delayedAckEvent, call,
3458 RX_CALL_REFCOUNT_DELAY);
3459 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3465 /* If we find a higher sequence packet, break out and
3466 * insert the new packet here. */
3467 if (seq < tp->header.seq)
3469 /* Check for missing packet */
3470 if (tp->header.seq != prev + 1) {
3474 prev = tp->header.seq;
3477 /* Keep track of whether we have received the last packet. */
3478 if (flags & RX_LAST_PACKET) {
3479 call->flags |= RX_CALL_HAVE_LAST;
3482 /* It's within the window: add it to the the receive queue.
3483 * tp is left by the previous loop either pointing at the
3484 * packet before which to insert the new packet, or at the
3485 * queue head if the queue is empty or the packet should be
3487 np->flags |= RX_PKTFLAG_RQ;
3488 #ifdef RXDEBUG_PACKET
3490 #endif /* RXDEBUG_PACKET */
3491 queue_InsertBefore(tp, np);
3495 /* Check whether we have all of the packets for this call */
3496 if ((call->flags & RX_CALL_HAVE_LAST)
3497 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3498 afs_uint32 tseq; /* temporary sequence number */
3501 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3502 if (tseq != tp->header.seq)
3504 if (tp->header.flags & RX_LAST_PACKET) {
3505 call->flags |= RX_CALL_RECEIVE_DONE;
3512 /* We need to send an ack of the packet is out of sequence,
3513 * or if an ack was requested by the peer. */
3514 if (seq != prev + 1 || missing) {
3515 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3516 } else if (flags & RX_REQUEST_ACK) {
3517 ackNeeded = RX_ACK_REQUESTED;
3520 /* Acknowledge the last packet for each call */
3521 if (flags & RX_LAST_PACKET) {
3532 * If the receiver is waiting for an iovec, fill the iovec
3533 * using the data from the receive queue */
3534 if (call->flags & RX_CALL_IOVEC_WAIT) {
3535 didHardAck = rxi_FillReadVec(call, serial);
3536 /* the call may have been aborted */
3545 /* Wakeup the reader if any */
3546 if ((call->flags & RX_CALL_READER_WAIT)
3547 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3548 || (call->iovNext >= call->iovMax)
3549 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3550 call->flags &= ~RX_CALL_READER_WAIT;
3551 #ifdef RX_ENABLE_LOCKS
3552 CV_BROADCAST(&call->cv_rq);
3554 osi_rxWakeup(&call->rq);
3560 * Send an ack when requested by the peer, or once every
3561 * rxi_SoftAckRate packets until the last packet has been
3562 * received. Always send a soft ack for the last packet in
3563 * the server's reply. */
3565 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3566 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3567 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3568 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3569 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3570 } else if (call->nSoftAcks) {
3571 clock_GetTime(&now);
3573 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3574 clock_Add(&when, &rx_lastAckDelay);
3576 clock_Add(&when, &rx_softAckDelay);
3578 if (!call->delayedAckEvent
3579 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3580 rxevent_Cancel(call->delayedAckEvent, call,
3581 RX_CALL_REFCOUNT_DELAY);
3582 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3583 call->delayedAckEvent =
3584 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3586 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3587 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3594 static void rxi_ComputeRate();
3598 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3600 struct rx_peer *peer = conn->peer;
3602 MUTEX_ENTER(&peer->peer_lock);
3603 peer->lastReachTime = clock_Sec();
3604 MUTEX_EXIT(&peer->peer_lock);
3606 MUTEX_ENTER(&conn->conn_data_lock);
3607 if (conn->flags & RX_CONN_ATTACHWAIT) {
3610 conn->flags &= ~RX_CONN_ATTACHWAIT;
3611 MUTEX_EXIT(&conn->conn_data_lock);
3613 for (i = 0; i < RX_MAXCALLS; i++) {
3614 struct rx_call *call = conn->call[i];
3617 MUTEX_ENTER(&call->lock);
3618 /* tnop can be null if newcallp is null */
3619 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3621 MUTEX_EXIT(&call->lock);
3625 MUTEX_EXIT(&conn->conn_data_lock);
3629 rx_ack_reason(int reason)
3632 case RX_ACK_REQUESTED:
3634 case RX_ACK_DUPLICATE:
3636 case RX_ACK_OUT_OF_SEQUENCE:
3638 case RX_ACK_EXCEEDS_WINDOW:
3640 case RX_ACK_NOSPACE:
3644 case RX_ACK_PING_RESPONSE:
3656 /* rxi_ComputePeerNetStats
3658 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3659 * estimates (like RTT and throughput) based on ack packets. Caller
3660 * must ensure that the packet in question is the right one (i.e.
3661 * serial number matches).
3664 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3665 struct rx_ackPacket *ap, struct rx_packet *np)
3667 struct rx_peer *peer = call->conn->peer;
3669 /* Use RTT if not delayed by client. */
3670 if (ap->reason != RX_ACK_DELAY)
3671 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3673 rxi_ComputeRate(peer, call, p, np, ap->reason);
3677 /* The real smarts of the whole thing. */
3679 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
3682 struct rx_ackPacket *ap;
3684 struct rx_packet *tp;
3685 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3686 struct rx_connection *conn = call->conn;
3687 struct rx_peer *peer = conn->peer;
3690 /* because there are CM's that are bogus, sending weird values for this. */
3691 afs_uint32 skew = 0;
3696 int newAckCount = 0;
3697 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3698 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3700 if (rx_stats_active)
3701 rx_MutexIncrement(rx_stats.ackPacketsRead, rx_stats_mutex);
3702 ap = (struct rx_ackPacket *)rx_DataOf(np);
3703 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3705 return np; /* truncated ack packet */
3707 /* depends on ack packet struct */
3708 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3709 first = ntohl(ap->firstPacket);
3710 serial = ntohl(ap->serial);
3711 /* temporarily disabled -- needs to degrade over time
3712 * skew = ntohs(ap->maxSkew); */
3714 /* Ignore ack packets received out of order */
3715 if (first < call->tfirst) {
3719 if (np->header.flags & RX_SLOW_START_OK) {
3720 call->flags |= RX_CALL_SLOW_START_OK;
3723 if (ap->reason == RX_ACK_PING_RESPONSE)
3724 rxi_UpdatePeerReach(conn, call);
3728 if (rxdebug_active) {
3732 len = _snprintf(msg, sizeof(msg),
3733 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3734 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3735 ntohl(ap->serial), ntohl(ap->previousPacket),
3736 (unsigned int)np->header.seq, (unsigned int)skew,
3737 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3741 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3742 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3746 OutputDebugString(msg);
3748 #else /* AFS_NT40_ENV */
3751 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3752 ap->reason, ntohl(ap->previousPacket),
3753 (unsigned int)np->header.seq, (unsigned int)serial,
3754 (unsigned int)skew, ntohl(ap->firstPacket));
3757 for (offset = 0; offset < nAcks; offset++)
3758 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3763 #endif /* AFS_NT40_ENV */
3766 /* Update the outgoing packet skew value to the latest value of
3767 * the peer's incoming packet skew value. The ack packet, of
3768 * course, could arrive out of order, but that won't affect things
3770 MUTEX_ENTER(&peer->peer_lock);
3771 peer->outPacketSkew = skew;
3773 /* Check for packets that no longer need to be transmitted, and
3774 * discard them. This only applies to packets positively
3775 * acknowledged as having been sent to the peer's upper level.
3776 * All other packets must be retained. So only packets with
3777 * sequence numbers < ap->firstPacket are candidates. */
3778 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3779 if (tp->header.seq >= first)
3781 call->tfirst = tp->header.seq + 1;
3783 && (tp->header.serial == serial || tp->firstSerial == serial))
3784 rxi_ComputePeerNetStats(call, tp, ap, np);
3785 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3788 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3789 /* XXX Hack. Because we have to release the global rx lock when sending
3790 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3791 * in rxi_Start sending packets out because packets may move to the
3792 * freePacketQueue as result of being here! So we drop these packets until
3793 * we're safely out of the traversing. Really ugly!
3794 * To make it even uglier, if we're using fine grain locking, we can
3795 * set the ack bits in the packets and have rxi_Start remove the packets
3796 * when it's done transmitting.
3798 if (call->flags & RX_CALL_TQ_BUSY) {
3799 #ifdef RX_ENABLE_LOCKS
3800 tp->flags |= RX_PKTFLAG_ACKED;
3801 call->flags |= RX_CALL_TQ_SOME_ACKED;
3802 #else /* RX_ENABLE_LOCKS */
3804 #endif /* RX_ENABLE_LOCKS */
3806 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3809 tp->flags &= ~RX_PKTFLAG_TQ;
3810 #ifdef RXDEBUG_PACKET
3812 #endif /* RXDEBUG_PACKET */
3813 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3818 /* Give rate detector a chance to respond to ping requests */
3819 if (ap->reason == RX_ACK_PING_RESPONSE) {
3820 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3824 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3826 /* Now go through explicit acks/nacks and record the results in
3827 * the waiting packets. These are packets that can't be released
3828 * yet, even with a positive acknowledge. This positive
3829 * acknowledge only means the packet has been received by the
3830 * peer, not that it will be retained long enough to be sent to
3831 * the peer's upper level. In addition, reset the transmit timers
3832 * of any missing packets (those packets that must be missing
3833 * because this packet was out of sequence) */
3835 call->nSoftAcked = 0;
3836 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3837 /* Update round trip time if the ack was stimulated on receipt
3839 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3840 #ifdef RX_ENABLE_LOCKS
3841 if (tp->header.seq >= first)
3842 #endif /* RX_ENABLE_LOCKS */
3843 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3845 && (tp->header.serial == serial || tp->firstSerial == serial))
3846 rxi_ComputePeerNetStats(call, tp, ap, np);
3848 /* Set the acknowledge flag per packet based on the
3849 * information in the ack packet. An acknowlegded packet can
3850 * be downgraded when the server has discarded a packet it
3851 * soacked previously, or when an ack packet is received
3852 * out of sequence. */
3853 if (tp->header.seq < first) {
3854 /* Implicit ack information */
3855 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3858 tp->flags |= RX_PKTFLAG_ACKED;
3859 } else if (tp->header.seq < first + nAcks) {
3860 /* Explicit ack information: set it in the packet appropriately */
3861 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3862 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3864 tp->flags |= RX_PKTFLAG_ACKED;
3871 } else /* RX_ACK_TYPE_NACK */ {
3872 tp->flags &= ~RX_PKTFLAG_ACKED;
3876 tp->flags &= ~RX_PKTFLAG_ACKED;
3880 /* If packet isn't yet acked, and it has been transmitted at least
3881 * once, reset retransmit time using latest timeout
3882 * ie, this should readjust the retransmit timer for all outstanding
3883 * packets... So we don't just retransmit when we should know better*/
3885 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3886 tp->retryTime = tp->timeSent;
3887 clock_Add(&tp->retryTime, &peer->timeout);
3888 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3889 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3893 /* If the window has been extended by this acknowledge packet,
3894 * then wakeup a sender waiting in alloc for window space, or try
3895 * sending packets now, if he's been sitting on packets due to
3896 * lack of window space */
3897 if (call->tnext < (call->tfirst + call->twind)) {
3898 #ifdef RX_ENABLE_LOCKS
3899 CV_SIGNAL(&call->cv_twind);
3901 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3902 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3903 osi_rxWakeup(&call->twind);
3906 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3907 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3911 /* if the ack packet has a receivelen field hanging off it,
3912 * update our state */
3913 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3916 /* If the ack packet has a "recommended" size that is less than
3917 * what I am using now, reduce my size to match */
3918 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3919 (int)sizeof(afs_int32), &tSize);
3920 tSize = (afs_uint32) ntohl(tSize);
3921 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3923 /* Get the maximum packet size to send to this peer */
3924 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3926 tSize = (afs_uint32) ntohl(tSize);
3927 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3928 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3930 /* sanity check - peer might have restarted with different params.
3931 * If peer says "send less", dammit, send less... Peer should never
3932 * be unable to accept packets of the size that prior AFS versions would
3933 * send without asking. */
3934 if (peer->maxMTU != tSize) {
3935 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3937 peer->maxMTU = tSize;
3938 peer->MTU = MIN(tSize, peer->MTU);
3939 call->MTU = MIN(call->MTU, tSize);
3942 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3945 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3946 (int)sizeof(afs_int32), &tSize);
3947 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3948 if (tSize < call->twind) { /* smaller than our send */
3949 call->twind = tSize; /* window, we must send less... */
3950 call->ssthresh = MIN(call->twind, call->ssthresh);
3951 call->conn->twind[call->channel] = call->twind;
3954 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3955 * network MTU confused with the loopback MTU. Calculate the
3956 * maximum MTU here for use in the slow start code below.
3958 maxMTU = peer->maxMTU;
3959 /* Did peer restart with older RX version? */
3960 if (peer->maxDgramPackets > 1) {
3961 peer->maxDgramPackets = 1;
3963 } else if (np->length >=
3964 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3967 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3968 sizeof(afs_int32), &tSize);
3969 tSize = (afs_uint32) ntohl(tSize);
3971 * As of AFS 3.5 we set the send window to match the receive window.
3973 if (tSize < call->twind) {
3974 call->twind = tSize;
3975 call->conn->twind[call->channel] = call->twind;
3976 call->ssthresh = MIN(call->twind, call->ssthresh);
3977 } else if (tSize > call->twind) {
3978 call->twind = tSize;
3979 call->conn->twind[call->channel] = call->twind;
3983 * As of AFS 3.5, a jumbogram is more than one fixed size
3984 * packet transmitted in a single UDP datagram. If the remote
3985 * MTU is smaller than our local MTU then never send a datagram
3986 * larger than the natural MTU.
3989 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3990 sizeof(afs_int32), &tSize);
3991 maxDgramPackets = (afs_uint32) ntohl(tSize);
3992 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3994 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3995 maxDgramPackets = MIN(maxDgramPackets, tSize);
3996 if (maxDgramPackets > 1) {
3997 peer->maxDgramPackets = maxDgramPackets;
3998 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4000 peer->maxDgramPackets = 1;
4001 call->MTU = peer->natMTU;
4003 } else if (peer->maxDgramPackets > 1) {
4004 /* Restarted with lower version of RX */
4005 peer->maxDgramPackets = 1;
4007 } else if (peer->maxDgramPackets > 1
4008 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4009 /* Restarted with lower version of RX */
4010 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4011 peer->natMTU = OLD_MAX_PACKET_SIZE;
4012 peer->MTU = OLD_MAX_PACKET_SIZE;
4013 peer->maxDgramPackets = 1;
4014 peer->nDgramPackets = 1;
4016 call->MTU = OLD_MAX_PACKET_SIZE;
4021 * Calculate how many datagrams were successfully received after
4022 * the first missing packet and adjust the negative ack counter
4027 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4028 if (call->nNacks < nNacked) {
4029 call->nNacks = nNacked;
4032 call->nAcks += newAckCount;
4036 if (call->flags & RX_CALL_FAST_RECOVER) {
4038 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4040 call->flags &= ~RX_CALL_FAST_RECOVER;
4041 call->cwind = call->nextCwind;
4042 call->nextCwind = 0;
4045 call->nCwindAcks = 0;
4046 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4047 /* Three negative acks in a row trigger congestion recovery */
4048 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4049 MUTEX_EXIT(&peer->peer_lock);
4050 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4051 /* someone else is waiting to start recovery */
4054 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4055 rxi_WaitforTQBusy(call);
4056 MUTEX_ENTER(&peer->peer_lock);
4057 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4058 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4059 call->flags |= RX_CALL_FAST_RECOVER;
4060 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4062 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4063 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4064 call->nextCwind = call->ssthresh;
4067 peer->MTU = call->MTU;
4068 peer->cwind = call->nextCwind;
4069 peer->nDgramPackets = call->nDgramPackets;
4071 call->congestSeq = peer->congestSeq;
4072 /* Reset the resend times on the packets that were nacked
4073 * so we will retransmit as soon as the window permits*/
4074 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4076 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4077 clock_Zero(&tp->retryTime);
4079 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4084 /* If cwind is smaller than ssthresh, then increase
4085 * the window one packet for each ack we receive (exponential
4087 * If cwind is greater than or equal to ssthresh then increase
4088 * the congestion window by one packet for each cwind acks we
4089 * receive (linear growth). */
4090 if (call->cwind < call->ssthresh) {
4092 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4093 call->nCwindAcks = 0;
4095 call->nCwindAcks += newAckCount;
4096 if (call->nCwindAcks >= call->cwind) {
4097 call->nCwindAcks = 0;
4098 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4102 * If we have received several acknowledgements in a row then
4103 * it is time to increase the size of our datagrams
4105 if ((int)call->nAcks > rx_nDgramThreshold) {
4106 if (peer->maxDgramPackets > 1) {
4107 if (call->nDgramPackets < peer->maxDgramPackets) {
4108 call->nDgramPackets++;
4110 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4111 } else if (call->MTU < peer->maxMTU) {
4112 call->MTU += peer->natMTU;
4113 call->MTU = MIN(call->MTU, peer->maxMTU);
4119 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4121 /* Servers need to hold the call until all response packets have
4122 * been acknowledged. Soft acks are good enough since clients
4123 * are not allowed to clear their receive queues. */
4124 if (call->state == RX_STATE_HOLD
4125 && call->tfirst + call->nSoftAcked >= call->tnext) {
4126 call->state = RX_STATE_DALLY;
4127 rxi_ClearTransmitQueue(call, 0);
4128 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4129 } else if (!queue_IsEmpty(&call->tq)) {
4130 rxi_Start(0, call, 0, istack);
4135 /* Received a response to a challenge packet */
4137 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4138 struct rx_packet *np, int istack)
4142 /* Ignore the packet if we're the client */
4143 if (conn->type == RX_CLIENT_CONNECTION)
4146 /* If already authenticated, ignore the packet (it's probably a retry) */
4147 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4150 /* Otherwise, have the security object evaluate the response packet */
4151 error = RXS_CheckResponse(conn->securityObject, conn, np);
4153 /* If the response is invalid, reset the connection, sending
4154 * an abort to the peer */
4158 rxi_ConnectionError(conn, error);
4159 MUTEX_ENTER(&conn->conn_data_lock);
4160 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4161 MUTEX_EXIT(&conn->conn_data_lock);
4164 /* If the response is valid, any calls waiting to attach
4165 * servers can now do so */
4168 for (i = 0; i < RX_MAXCALLS; i++) {
4169 struct rx_call *call = conn->call[i];
4171 MUTEX_ENTER(&call->lock);
4172 if (call->state == RX_STATE_PRECALL)
4173 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4174 /* tnop can be null if newcallp is null */
4175 MUTEX_EXIT(&call->lock);
4179 /* Update the peer reachability information, just in case
4180 * some calls went into attach-wait while we were waiting
4181 * for authentication..
4183 rxi_UpdatePeerReach(conn, NULL);
4188 /* A client has received an authentication challenge: the security
4189 * object is asked to cough up a respectable response packet to send
4190 * back to the server. The server is responsible for retrying the
4191 * challenge if it fails to get a response. */
4194 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4195 struct rx_packet *np, int istack)
4199 /* Ignore the challenge if we're the server */
4200 if (conn->type == RX_SERVER_CONNECTION)
4203 /* Ignore the challenge if the connection is otherwise idle; someone's
4204 * trying to use us as an oracle. */
4205 if (!rxi_HasActiveCalls(conn))
4208 /* Send the security object the challenge packet. It is expected to fill
4209 * in the response. */
4210 error = RXS_GetResponse(conn->securityObject, conn, np);
4212 /* If the security object is unable to return a valid response, reset the
4213 * connection and send an abort to the peer. Otherwise send the response
4214 * packet to the peer connection. */
4216 rxi_ConnectionError(conn, error);
4217 MUTEX_ENTER(&conn->conn_data_lock);
4218 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4219 MUTEX_EXIT(&conn->conn_data_lock);
4221 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4222 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4228 /* Find an available server process to service the current request in
4229 * the given call structure. If one isn't available, queue up this
4230 * call so it eventually gets one */
4232 rxi_AttachServerProc(struct rx_call *call,
4233 osi_socket socket, int *tnop,
4234 struct rx_call **newcallp)
4236 struct rx_serverQueueEntry *sq;
4237 struct rx_service *service = call->conn->service;
4240 /* May already be attached */
4241 if (call->state == RX_STATE_ACTIVE)
4244 MUTEX_ENTER(&rx_serverPool_lock);
4246 haveQuota = QuotaOK(service);
4247 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4248 /* If there are no processes available to service this call,
4249 * put the call on the incoming call queue (unless it's
4250 * already on the queue).
4252 #ifdef RX_ENABLE_LOCKS
4254 ReturnToServerPool(service);
4255 #endif /* RX_ENABLE_LOCKS */
4257 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4258 call->flags |= RX_CALL_WAIT_PROC;
4259 MUTEX_ENTER(&rx_waiting_mutex);
4262 MUTEX_EXIT(&rx_waiting_mutex);
4263 rxi_calltrace(RX_CALL_ARRIVAL, call);
4264 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4265 queue_Append(&rx_incomingCallQueue, call);
4268 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4270 /* If hot threads are enabled, and both newcallp and sq->socketp
4271 * are non-null, then this thread will process the call, and the
4272 * idle server thread will start listening on this threads socket.
4275 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4278 *sq->socketp = socket;
4279 clock_GetTime(&call->startTime);
4280 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4284 if (call->flags & RX_CALL_WAIT_PROC) {
4285 /* Conservative: I don't think this should happen */
4286 call->flags &= ~RX_CALL_WAIT_PROC;
4287 if (queue_IsOnQueue(call)) {
4290 MUTEX_ENTER(&rx_waiting_mutex);
4292 MUTEX_EXIT(&rx_waiting_mutex);
4295 call->state = RX_STATE_ACTIVE;
4296 call->mode = RX_MODE_RECEIVING;
4297 #ifdef RX_KERNEL_TRACE
4299 int glockOwner = ISAFS_GLOCK();
4302 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4303 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4309 if (call->flags & RX_CALL_CLEARED) {
4310 /* send an ack now to start the packet flow up again */
4311 call->flags &= ~RX_CALL_CLEARED;
4312 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4314 #ifdef RX_ENABLE_LOCKS
4317 service->nRequestsRunning++;
4318 if (service->nRequestsRunning <= service->minProcs)
4324 MUTEX_EXIT(&rx_serverPool_lock);
4327 /* Delay the sending of an acknowledge event for a short while, while
4328 * a new call is being prepared (in the case of a client) or a reply
4329 * is being prepared (in the case of a server). Rather than sending
4330 * an ack packet, an ACKALL packet is sent. */
4332 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4334 #ifdef RX_ENABLE_LOCKS
4336 MUTEX_ENTER(&call->lock);
4337 call->delayedAckEvent = NULL;
4338 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4340 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4341 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4343 MUTEX_EXIT(&call->lock);
4344 #else /* RX_ENABLE_LOCKS */
4346 call->delayedAckEvent = NULL;
4347 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4348 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4349 #endif /* RX_ENABLE_LOCKS */
4353 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4355 struct rx_call *call = arg1;
4356 #ifdef RX_ENABLE_LOCKS
4358 MUTEX_ENTER(&call->lock);
4359 if (event == call->delayedAckEvent)
4360 call->delayedAckEvent = NULL;
4361 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4363 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4365 MUTEX_EXIT(&call->lock);
4366 #else /* RX_ENABLE_LOCKS */
4368 call->delayedAckEvent = NULL;
4369 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4370 #endif /* RX_ENABLE_LOCKS */
4374 #ifdef RX_ENABLE_LOCKS
4375 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4376 * clearing them out.
4379 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4381 struct rx_packet *p, *tp;
4384 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4385 p->flags |= RX_PKTFLAG_ACKED;
4389 call->flags |= RX_CALL_TQ_CLEARME;
4390 call->flags |= RX_CALL_TQ_SOME_ACKED;
4393 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4394 call->tfirst = call->tnext;
4395 call->nSoftAcked = 0;
4397 if (call->flags & RX_CALL_FAST_RECOVER) {
4398 call->flags &= ~RX_CALL_FAST_RECOVER;
4399 call->cwind = call->nextCwind;
4400 call->nextCwind = 0;
4403 CV_SIGNAL(&call->cv_twind);
4405 #endif /* RX_ENABLE_LOCKS */
4407 /* Clear out the transmit queue for the current call (all packets have
4408 * been received by peer) */
4410 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4412 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4413 struct rx_packet *p, *tp;
4415 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4417 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4418 p->flags |= RX_PKTFLAG_ACKED;
4422 call->flags |= RX_CALL_TQ_CLEARME;
4423 call->flags |= RX_CALL_TQ_SOME_ACKED;
4426 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4427 #ifdef RXDEBUG_PACKET
4429 #endif /* RXDEBUG_PACKET */
4430 rxi_FreePackets(0, &call->tq);
4431 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4432 call->flags &= ~RX_CALL_TQ_CLEARME;
4434 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4436 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4437 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4438 call->nSoftAcked = 0;
4440 if (call->flags & RX_CALL_FAST_RECOVER) {
4441 call->flags &= ~RX_CALL_FAST_RECOVER;
4442 call->cwind = call->nextCwind;
4444 #ifdef RX_ENABLE_LOCKS
4445 CV_SIGNAL(&call->cv_twind);
4447 osi_rxWakeup(&call->twind);
4452 rxi_ClearReceiveQueue(struct rx_call *call)
4454 if (queue_IsNotEmpty(&call->rq)) {
4457 count = rxi_FreePackets(0, &call->rq);
4458 rx_packetReclaims += count;
4459 #ifdef RXDEBUG_PACKET
4461 if ( call->rqc != 0 )
4462 dpf(("rxi_ClearReceiveQueue call %x rqc %u != 0", call, call->rqc));
4464 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4466 if (call->state == RX_STATE_PRECALL) {
4467 call->flags |= RX_CALL_CLEARED;
4471 /* Send an abort packet for the specified call */
4473 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
4474 int istack, int force)
4477 struct clock when, now;
4482 /* Clients should never delay abort messages */
4483 if (rx_IsClientConn(call->conn))
4486 if (call->abortCode != call->error) {
4487 call->abortCode = call->error;
4488 call->abortCount = 0;
4491 if (force || rxi_callAbortThreshhold == 0
4492 || call->abortCount < rxi_callAbortThreshhold) {
4493 if (call->delayedAbortEvent) {
4494 rxevent_Cancel(call->delayedAbortEvent, call,
4495 RX_CALL_REFCOUNT_ABORT);
4497 error = htonl(call->error);
4500 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4501 (char *)&error, sizeof(error), istack);
4502 } else if (!call->delayedAbortEvent) {
4503 clock_GetTime(&now);
4505 clock_Addmsec(&when, rxi_callAbortDelay);
4506 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4507 call->delayedAbortEvent =
4508 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4513 /* Send an abort packet for the specified connection. Packet is an
4514 * optional pointer to a packet that can be used to send the abort.
4515 * Once the number of abort messages reaches the threshhold, an
4516 * event is scheduled to send the abort. Setting the force flag
4517 * overrides sending delayed abort messages.
4519 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4520 * to send the abort packet.
4523 rxi_SendConnectionAbort(struct rx_connection *conn,
4524 struct rx_packet *packet, int istack, int force)
4527 struct clock when, now;
4532 /* Clients should never delay abort messages */
4533 if (rx_IsClientConn(conn))
4536 if (force || rxi_connAbortThreshhold == 0
4537 || conn->abortCount < rxi_connAbortThreshhold) {
4538 if (conn->delayedAbortEvent) {
4539 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4541 error = htonl(conn->error);
4543 MUTEX_EXIT(&conn->conn_data_lock);
4545 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4546 RX_PACKET_TYPE_ABORT, (char *)&error,
4547 sizeof(error), istack);
4548 MUTEX_ENTER(&conn->conn_data_lock);
4549 } else if (!conn->delayedAbortEvent) {
4550 clock_GetTime(&now);
4552 clock_Addmsec(&when, rxi_connAbortDelay);
4553 conn->delayedAbortEvent =
4554 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4559 /* Associate an error all of the calls owned by a connection. Called
4560 * with error non-zero. This is only for really fatal things, like
4561 * bad authentication responses. The connection itself is set in
4562 * error at this point, so that future packets received will be
4565 rxi_ConnectionError(struct rx_connection *conn,
4571 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4573 MUTEX_ENTER(&conn->conn_data_lock);
4574 if (conn->challengeEvent)
4575 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4576 if (conn->checkReachEvent) {
4577 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4578 conn->checkReachEvent = 0;
4579 conn->flags &= ~RX_CONN_ATTACHWAIT;
4582 MUTEX_EXIT(&conn->conn_data_lock);
4583 for (i = 0; i < RX_MAXCALLS; i++) {
4584 struct rx_call *call = conn->call[i];
4586 MUTEX_ENTER(&call->lock);
4587 rxi_CallError(call, error);
4588 MUTEX_EXIT(&call->lock);
4591 conn->error = error;
4592 if (rx_stats_active)
4593 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4598 rxi_CallError(struct rx_call *call, afs_int32 error)
4601 osirx_AssertMine(&call->lock, "rxi_CallError");
4603 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4605 error = call->error;
4607 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4608 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4609 rxi_ResetCall(call, 0);
4612 rxi_ResetCall(call, 0);
4614 call->error = error;
4615 call->mode = RX_MODE_ERROR;
4618 /* Reset various fields in a call structure, and wakeup waiting
4619 * processes. Some fields aren't changed: state & mode are not
4620 * touched (these must be set by the caller), and bufptr, nLeft, and
4621 * nFree are not reset, since these fields are manipulated by
4622 * unprotected macros, and may only be reset by non-interrupting code.
4625 /* this code requires that call->conn be set properly as a pre-condition. */
4626 #endif /* ADAPT_WINDOW */
4629 rxi_ResetCall(struct rx_call *call, int newcall)
4632 struct rx_peer *peer;
4633 struct rx_packet *packet;
4635 osirx_AssertMine(&call->lock, "rxi_ResetCall");
4637 dpf(("rxi_ResetCall(call %x, newcall %d)\n", call, newcall));
4639 /* Notify anyone who is waiting for asynchronous packet arrival */
4640 if (call->arrivalProc) {
4641 (*call->arrivalProc) (call, call->arrivalProcHandle,
4642 call->arrivalProcArg);
4643 call->arrivalProc = (void (*)())0;
4646 if (call->delayedAbortEvent) {
4647 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4648 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4650 rxi_SendCallAbort(call, packet, 0, 1);
4651 rxi_FreePacket(packet);
4656 * Update the peer with the congestion information in this call
4657 * so other calls on this connection can pick up where this call
4658 * left off. If the congestion sequence numbers don't match then
4659 * another call experienced a retransmission.
4661 peer = call->conn->peer;
4662 MUTEX_ENTER(&peer->peer_lock);
4664 if (call->congestSeq == peer->congestSeq) {
4665 peer->cwind = MAX(peer->cwind, call->cwind);
4666 peer->MTU = MAX(peer->MTU, call->MTU);
4667 peer->nDgramPackets =
4668 MAX(peer->nDgramPackets, call->nDgramPackets);
4671 call->abortCode = 0;
4672 call->abortCount = 0;
4674 if (peer->maxDgramPackets > 1) {
4675 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4677 call->MTU = peer->MTU;
4679 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4680 call->ssthresh = rx_maxSendWindow;
4681 call->nDgramPackets = peer->nDgramPackets;
4682 call->congestSeq = peer->congestSeq;
4683 MUTEX_EXIT(&peer->peer_lock);
4685 flags = call->flags;
4686 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4687 if (flags & RX_CALL_TQ_BUSY) {
4688 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4689 call->flags |= (flags & RX_CALL_TQ_WAIT);
4691 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4693 rxi_ClearTransmitQueue(call, 1);
4694 /* why init the queue if you just emptied it? queue_Init(&call->tq); */
4695 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4696 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4699 while (call->tqWaiters) {
4700 #ifdef RX_ENABLE_LOCKS
4701 CV_BROADCAST(&call->cv_tq);
4702 #else /* RX_ENABLE_LOCKS */
4703 osi_rxWakeup(&call->tq);
4704 #endif /* RX_ENABLE_LOCKS */
4709 rxi_ClearReceiveQueue(call);
4710 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
4712 if (call->currentPacket) {
4713 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
4714 call->currentPacket->flags |= RX_PKTFLAG_IOVQ;
4715 queue_Prepend(&call->iovq, call->currentPacket);
4716 #ifdef RXDEBUG_PACKET
4718 #endif /* RXDEBUG_PACKET */
4719 call->currentPacket = (struct rx_packet *)0;
4721 call->curlen = call->nLeft = call->nFree = 0;
4723 #ifdef RXDEBUG_PACKET
4726 rxi_FreePackets(0, &call->iovq);
4729 call->twind = call->conn->twind[call->channel];
4730 call->rwind = call->conn->rwind[call->channel];
4731 call->nSoftAcked = 0;
4732 call->nextCwind = 0;
4735 call->nCwindAcks = 0;
4736 call->nSoftAcks = 0;
4737 call->nHardAcks = 0;
4739 call->tfirst = call->rnext = call->tnext = 1;
4741 call->lastAcked = 0;
4742 call->localStatus = call->remoteStatus = 0;
4744 if (flags & RX_CALL_READER_WAIT) {
4745 #ifdef RX_ENABLE_LOCKS
4746 CV_BROADCAST(&call->cv_rq);
4748 osi_rxWakeup(&call->rq);
4751 if (flags & RX_CALL_WAIT_PACKETS) {
4752 MUTEX_ENTER(&rx_freePktQ_lock);
4753 rxi_PacketsUnWait(); /* XXX */
4754 MUTEX_EXIT(&rx_freePktQ_lock);
4756 #ifdef RX_ENABLE_LOCKS
4757 CV_SIGNAL(&call->cv_twind);
4759 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4760 osi_rxWakeup(&call->twind);
4763 #ifdef RX_ENABLE_LOCKS
4764 /* The following ensures that we don't mess with any queue while some
4765 * other thread might also be doing so. The call_queue_lock field is
4766 * is only modified under the call lock. If the call is in the process
4767 * of being removed from a queue, the call is not locked until the
4768 * the queue lock is dropped and only then is the call_queue_lock field
4769 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4770 * Note that any other routine which removes a call from a queue has to
4771 * obtain the queue lock before examing the queue and removing the call.
4773 if (call->call_queue_lock) {
4774 MUTEX_ENTER(call->call_queue_lock);
4775 if (queue_IsOnQueue(call)) {
4777 if (flags & RX_CALL_WAIT_PROC) {
4779 MUTEX_ENTER(&rx_waiting_mutex);
4781 MUTEX_EXIT(&rx_waiting_mutex);
4784 MUTEX_EXIT(call->call_queue_lock);
4785 CLEAR_CALL_QUEUE_LOCK(call);
4787 #else /* RX_ENABLE_LOCKS */
4788 if (queue_IsOnQueue(call)) {
4790 if (flags & RX_CALL_WAIT_PROC)
4793 #endif /* RX_ENABLE_LOCKS */
4795 rxi_KeepAliveOff(call);
4796 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4799 /* Send an acknowledge for the indicated packet (seq,serial) of the
4800 * indicated call, for the indicated reason (reason). This
4801 * acknowledge will specifically acknowledge receiving the packet, and
4802 * will also specify which other packets for this call have been
4803 * received. This routine returns the packet that was used to the
4804 * caller. The caller is responsible for freeing it or re-using it.
4805 * This acknowledgement also returns the highest sequence number
4806 * actually read out by the higher level to the sender; the sender
4807 * promises to keep around packets that have not been read by the
4808 * higher level yet (unless, of course, the sender decides to abort
4809 * the call altogether). Any of p, seq, serial, pflags, or reason may
4810 * be set to zero without ill effect. That is, if they are zero, they
4811 * will not convey any information.
4812 * NOW there is a trailer field, after the ack where it will safely be
4813 * ignored by mundanes, which indicates the maximum size packet this
4814 * host can swallow. */
4816 struct rx_packet *optionalPacket; use to send ack (or null)
4817 int seq; Sequence number of the packet we are acking
4818 int serial; Serial number of the packet
4819 int pflags; Flags field from packet header
4820 int reason; Reason an acknowledge was prompted
4824 rxi_SendAck(struct rx_call *call,
4825 struct rx_packet *optionalPacket, int serial, int reason,
4828 struct rx_ackPacket *ap;
4829 struct rx_packet *rqp;
4830 struct rx_packet *nxp; /* For queue_Scan */
4831 struct rx_packet *p;
4834 #ifdef RX_ENABLE_TSFPQ
4835 struct rx_ts_info_t * rx_ts_info;
4839 * Open the receive window once a thread starts reading packets
4841 if (call->rnext > 1) {
4842 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
4845 call->nHardAcks = 0;
4846 call->nSoftAcks = 0;
4847 if (call->rnext > call->lastAcked)
4848 call->lastAcked = call->rnext;
4852 rx_computelen(p, p->length); /* reset length, you never know */
4853 } /* where that's been... */
4854 #ifdef RX_ENABLE_TSFPQ
4856 RX_TS_INFO_GET(rx_ts_info);
4857 if ((p = rx_ts_info->local_special_packet)) {
4858 rx_computelen(p, p->length);
4859 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4860 rx_ts_info->local_special_packet = p;
4861 } else { /* We won't send the ack, but don't panic. */
4862 return optionalPacket;
4866 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4867 /* We won't send the ack, but don't panic. */
4868 return optionalPacket;
4873 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4876 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4877 #ifndef RX_ENABLE_TSFPQ
4878 if (!optionalPacket)
4881 return optionalPacket;
4883 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4884 if (rx_Contiguous(p) < templ) {
4885 #ifndef RX_ENABLE_TSFPQ
4886 if (!optionalPacket)
4889 return optionalPacket;
4894 /* MTUXXX failing to send an ack is very serious. We should */
4895 /* try as hard as possible to send even a partial ack; it's */
4896 /* better than nothing. */
4897 ap = (struct rx_ackPacket *)rx_DataOf(p);
4898 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4899 ap->reason = reason;
4901 /* The skew computation used to be bogus, I think it's better now. */
4902 /* We should start paying attention to skew. XXX */
4903 ap->serial = htonl(serial);
4904 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4906 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4907 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4909 /* No fear of running out of ack packet here because there can only be at most
4910 * one window full of unacknowledged packets. The window size must be constrained
4911 * to be less than the maximum ack size, of course. Also, an ack should always
4912 * fit into a single packet -- it should not ever be fragmented. */
4913 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4914 if (!rqp || !call->rq.next
4915 || (rqp->header.seq > (call->rnext + call->rwind))) {
4916 #ifndef RX_ENABLE_TSFPQ
4917 if (!optionalPacket)
4920 rxi_CallError(call, RX_CALL_DEAD);
4921 return optionalPacket;
4924 while (rqp->header.seq > call->rnext + offset)
4925 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4926 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4928 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4929 #ifndef RX_ENABLE_TSFPQ
4930 if (!optionalPacket)
4933 rxi_CallError(call, RX_CALL_DEAD);
4934 return optionalPacket;
4939 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4941 /* these are new for AFS 3.3 */
4942 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4943 templ = htonl(templ);
4944 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4945 templ = htonl(call->conn->peer->ifMTU);
4946 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4947 sizeof(afs_int32), &templ);
4949 /* new for AFS 3.4 */
4950 templ = htonl(call->rwind);
4951 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4952 sizeof(afs_int32), &templ);
4954 /* new for AFS 3.5 */
4955 templ = htonl(call->conn->peer->ifDgramPackets);
4956 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4957 sizeof(afs_int32), &templ);
4959 p->header.serviceId = call->conn->serviceId;
4960 p->header.cid = (call->conn->cid | call->channel);
4961 p->header.callNumber = *call->callNumber;
4963 p->header.securityIndex = call->conn->securityIndex;
4964 p->header.epoch = call->conn->epoch;
4965 p->header.type = RX_PACKET_TYPE_ACK;
4966 p->header.flags = RX_SLOW_START_OK;
4967 if (reason == RX_ACK_PING) {
4968 p->header.flags |= RX_REQUEST_ACK;
4970 clock_GetTime(&call->pingRequestTime);
4973 if (call->conn->type == RX_CLIENT_CONNECTION)
4974 p->header.flags |= RX_CLIENT_INITIATED;
4978 if (rxdebug_active) {
4982 len = _snprintf(msg, sizeof(msg),
4983 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4984 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4985 ntohl(ap->serial), ntohl(ap->previousPacket),
4986 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4987 ap->nAcks, ntohs(ap->bufferSpace) );
4991 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4992 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4996 OutputDebugString(msg);
4998 #else /* AFS_NT40_ENV */
5000 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5001 ap->reason, ntohl(ap->previousPacket),
5002 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5004 for (offset = 0; offset < ap->nAcks; offset++)
5005 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5010 #endif /* AFS_NT40_ENV */
5013 int i, nbytes = p->length;
5015 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5016 if (nbytes <= p->wirevec[i].iov_len) {
5019 savelen = p->wirevec[i].iov_len;
5021 p->wirevec[i].iov_len = nbytes;
5023 rxi_Send(call, p, istack);
5024 p->wirevec[i].iov_len = savelen;
5028 nbytes -= p->wirevec[i].iov_len;
5031 if (rx_stats_active)
5032 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
5033 #ifndef RX_ENABLE_TSFPQ
5034 if (!optionalPacket)
5037 return optionalPacket; /* Return packet for re-use by caller */
5040 /* Send all of the packets in the list in single datagram */
5042 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
5043 int istack, int moreFlag, struct clock *now,
5044 struct clock *retryTime, int resending)
5049 struct rx_connection *conn = call->conn;
5050 struct rx_peer *peer = conn->peer;
5052 MUTEX_ENTER(&peer->peer_lock);
5055 peer->reSends += len;
5056 if (rx_stats_active)
5057 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
5058 MUTEX_EXIT(&peer->peer_lock);
5060 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
5064 /* Set the packet flags and schedule the resend events */
5065 /* Only request an ack for the last packet in the list */
5066 for (i = 0; i < len; i++) {
5067 list[i]->retryTime = *retryTime;
5068 if (list[i]->header.serial) {
5069 /* Exponentially backoff retry times */
5070 if (list[i]->backoff < MAXBACKOFF) {
5071 /* so it can't stay == 0 */
5072 list[i]->backoff = (list[i]->backoff << 1) + 1;
5075 clock_Addmsec(&(list[i]->retryTime),
5076 ((afs_uint32) list[i]->backoff) << 8);
5079 /* Wait a little extra for the ack on the last packet */
5080 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
5081 clock_Addmsec(&(list[i]->retryTime), 400);
5084 /* Record the time sent */
5085 list[i]->timeSent = *now;
5087 /* Ask for an ack on retransmitted packets, on every other packet
5088 * if the peer doesn't support slow start. Ask for an ack on every
5089 * packet until the congestion window reaches the ack rate. */
5090 if (list[i]->header.serial) {
5092 if (rx_stats_active)
5093 rx_MutexIncrement(rx_stats.dataPacketsReSent, rx_stats_mutex);
5095 /* improved RTO calculation- not Karn */
5096 list[i]->firstSent = *now;
5097 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5098 || (!(call->flags & RX_CALL_SLOW_START_OK)
5099 && (list[i]->header.seq & 1)))) {
5104 MUTEX_ENTER(&peer->peer_lock);
5108 if (rx_stats_active)
5109 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
5110 MUTEX_EXIT(&peer->peer_lock);
5112 /* Tag this packet as not being the last in this group,
5113 * for the receiver's benefit */
5114 if (i < len - 1 || moreFlag) {
5115 list[i]->header.flags |= RX_MORE_PACKETS;
5118 /* Install the new retransmit time for the packet, and
5119 * record the time sent */
5120 list[i]->timeSent = *now;
5124 list[len - 1]->header.flags |= RX_REQUEST_ACK;
5127 /* Since we're about to send a data packet to the peer, it's
5128 * safe to nuke any scheduled end-of-packets ack */
5129 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5131 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5132 MUTEX_EXIT(&call->lock);
5134 rxi_SendPacketList(call, conn, list, len, istack);
5136 rxi_SendPacket(call, conn, list[0], istack);
5138 MUTEX_ENTER(&call->lock);
5139 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5141 /* Update last send time for this call (for keep-alive
5142 * processing), and for the connection (so that we can discover
5143 * idle connections) */
5144 call->lastSendData = conn->lastSendTime = call->lastSendTime = clock_Sec();
5147 /* When sending packets we need to follow these rules:
5148 * 1. Never send more than maxDgramPackets in a jumbogram.
5149 * 2. Never send a packet with more than two iovecs in a jumbogram.
5150 * 3. Never send a retransmitted packet in a jumbogram.
5151 * 4. Never send more than cwind/4 packets in a jumbogram
5152 * We always keep the last list we should have sent so we
5153 * can set the RX_MORE_PACKETS flags correctly.
5156 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5157 int istack, struct clock *now, struct clock *retryTime,
5160 int i, cnt, lastCnt = 0;
5161 struct rx_packet **listP, **lastP = 0;
5162 struct rx_peer *peer = call->conn->peer;
5163 int morePackets = 0;
5165 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
5166 /* Does the current packet force us to flush the current list? */
5168 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5169 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5171 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
5173 /* If the call enters an error state stop sending, or if
5174 * we entered congestion recovery mode, stop sending */
5175 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5183 /* Add the current packet to the list if it hasn't been acked.
5184 * Otherwise adjust the list pointer to skip the current packet. */
5185 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5187 /* Do we need to flush the list? */
5188 if (cnt >= (int)peer->maxDgramPackets
5189 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5190 || list[i]->header.serial
5191 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5193 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5194 retryTime, resending);
5195 /* If the call enters an error state stop sending, or if
5196 * we entered congestion recovery mode, stop sending */
5198 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5203 listP = &list[i + 1];
5208 osi_Panic("rxi_SendList error");
5210 listP = &list[i + 1];
5214 /* Send the whole list when the call is in receive mode, when
5215 * the call is in eof mode, when we are in fast recovery mode,
5216 * and when we have the last packet */
5217 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5218 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5219 || (call->flags & RX_CALL_FAST_RECOVER)) {
5220 /* Check for the case where the current list contains
5221 * an acked packet. Since we always send retransmissions
5222 * in a separate packet, we only need to check the first
5223 * packet in the list */
5224 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5228 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5229 retryTime, resending);
5230 /* If the call enters an error state stop sending, or if
5231 * we entered congestion recovery mode, stop sending */
5232 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5236 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5239 } else if (lastCnt > 0) {
5240 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5245 #ifdef RX_ENABLE_LOCKS
5246 /* Call rxi_Start, below, but with the call lock held. */
5248 rxi_StartUnlocked(struct rxevent *event,
5249 void *arg0, void *arg1, int istack)
5251 struct rx_call *call = arg0;
5253 MUTEX_ENTER(&call->lock);
5254 rxi_Start(event, call, arg1, istack);
5255 MUTEX_EXIT(&call->lock);
5257 #endif /* RX_ENABLE_LOCKS */
5259 /* This routine is called when new packets are readied for
5260 * transmission and when retransmission may be necessary, or when the
5261 * transmission window or burst count are favourable. This should be
5262 * better optimized for new packets, the usual case, now that we've
5263 * got rid of queues of send packets. XXXXXXXXXXX */
5265 rxi_Start(struct rxevent *event,
5266 void *arg0, void *arg1, int istack)
5268 struct rx_call *call = arg0;
5270 struct rx_packet *p;
5271 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5272 struct rx_peer *peer = call->conn->peer;
5273 struct clock now, usenow, retryTime;
5277 struct rx_packet **xmitList;
5280 /* If rxi_Start is being called as a result of a resend event,
5281 * then make sure that the event pointer is removed from the call
5282 * structure, since there is no longer a per-call retransmission
5284 if (event && event == call->resendEvent) {
5285 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5286 call->resendEvent = NULL;
5288 if (queue_IsEmpty(&call->tq)) {
5292 /* Timeouts trigger congestion recovery */
5293 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5294 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5295 /* someone else is waiting to start recovery */
5298 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5299 rxi_WaitforTQBusy(call);
5300 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5301 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5302 call->flags |= RX_CALL_FAST_RECOVER;
5303 if (peer->maxDgramPackets > 1) {
5304 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5306 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5308 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5309 call->nDgramPackets = 1;
5311 call->nextCwind = 1;
5314 MUTEX_ENTER(&peer->peer_lock);
5315 peer->MTU = call->MTU;
5316 peer->cwind = call->cwind;
5317 peer->nDgramPackets = 1;
5319 call->congestSeq = peer->congestSeq;
5320 MUTEX_EXIT(&peer->peer_lock);
5321 /* Clear retry times on packets. Otherwise, it's possible for
5322 * some packets in the queue to force resends at rates faster
5323 * than recovery rates.
5325 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5326 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5327 clock_Zero(&p->retryTime);
5332 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5333 if (rx_stats_active)
5334 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5339 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5340 /* Get clock to compute the re-transmit time for any packets
5341 * in this burst. Note, if we back off, it's reasonable to
5342 * back off all of the packets in the same manner, even if
5343 * some of them have been retransmitted more times than more
5345 * Do a dance to avoid blocking after setting now. */
5346 clock_Zero(&retryTime);
5347 MUTEX_ENTER(&peer->peer_lock);
5348 clock_Add(&retryTime, &peer->timeout);
5349 MUTEX_EXIT(&peer->peer_lock);
5350 clock_GetTime(&now);
5351 clock_Add(&retryTime, &now);
5353 /* Send (or resend) any packets that need it, subject to
5354 * window restrictions and congestion burst control
5355 * restrictions. Ask for an ack on the last packet sent in
5356 * this burst. For now, we're relying upon the window being
5357 * considerably bigger than the largest number of packets that
5358 * are typically sent at once by one initial call to
5359 * rxi_Start. This is probably bogus (perhaps we should ask
5360 * for an ack when we're half way through the current
5361 * window?). Also, for non file transfer applications, this
5362 * may end up asking for an ack for every packet. Bogus. XXXX
5365 * But check whether we're here recursively, and let the other guy
5368 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5369 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5370 call->flags |= RX_CALL_TQ_BUSY;
5372 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5374 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5375 call->flags &= ~RX_CALL_NEED_START;
5376 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5378 maxXmitPackets = MIN(call->twind, call->cwind);
5379 xmitList = (struct rx_packet **)
5380 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
5381 /* XXXX else we must drop any mtx we hold */
5382 afs_osi_Alloc_NoSleep(maxXmitPackets * sizeof(struct rx_packet *));
5384 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5386 if (xmitList == NULL)
5387 osi_Panic("rxi_Start, failed to allocate xmit list");
5388 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5389 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5390 /* We shouldn't be sending packets if a thread is waiting
5391 * to initiate congestion recovery */
5395 && (call->flags & RX_CALL_FAST_RECOVER)) {
5396 /* Only send one packet during fast recovery */
5399 if ((p->flags & RX_PKTFLAG_FREE)
5400 || (!queue_IsEnd(&call->tq, nxp)
5401 && (nxp->flags & RX_PKTFLAG_FREE))
5402 || (p == (struct rx_packet *)&rx_freePacketQueue)
5403 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5404 osi_Panic("rxi_Start: xmit queue clobbered");
5406 if (p->flags & RX_PKTFLAG_ACKED) {
5407 /* Since we may block, don't trust this */
5408 usenow.sec = usenow.usec = 0;
5409 if (rx_stats_active)
5410 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5411 continue; /* Ignore this packet if it has been acknowledged */
5414 /* Turn off all flags except these ones, which are the same
5415 * on each transmission */
5416 p->header.flags &= RX_PRESET_FLAGS;
5418 if (p->header.seq >=
5419 call->tfirst + MIN((int)call->twind,
5420 (int)(call->nSoftAcked +
5422 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5423 /* Note: if we're waiting for more window space, we can
5424 * still send retransmits; hence we don't return here, but
5425 * break out to schedule a retransmit event */
5426 dpf(("call %d waiting for window",
5427 *(call->callNumber)));
5431 /* Transmit the packet if it needs to be sent. */
5432 if (!clock_Lt(&now, &p->retryTime)) {
5433 if (nXmitPackets == maxXmitPackets) {
5434 rxi_SendXmitList(call, xmitList, nXmitPackets,
5435 istack, &now, &retryTime,
5437 osi_Free(xmitList, maxXmitPackets *
5438 sizeof(struct rx_packet *));
5441 xmitList[nXmitPackets++] = p;
5445 /* xmitList now hold pointers to all of the packets that are
5446 * ready to send. Now we loop to send the packets */
5447 if (nXmitPackets > 0) {
5448 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5449 &now, &retryTime, resending);
5452 maxXmitPackets * sizeof(struct rx_packet *));
5454 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5456 * TQ references no longer protected by this flag; they must remain
5457 * protected by the global lock.
5459 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5460 call->flags &= ~RX_CALL_TQ_BUSY;
5461 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5462 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5463 #ifdef RX_ENABLE_LOCKS
5464 osirx_AssertMine(&call->lock, "rxi_Start start");
5465 CV_BROADCAST(&call->cv_tq);
5466 #else /* RX_ENABLE_LOCKS */
5467 osi_rxWakeup(&call->tq);
5468 #endif /* RX_ENABLE_LOCKS */
5473 /* We went into the error state while sending packets. Now is
5474 * the time to reset the call. This will also inform the using
5475 * process that the call is in an error state.
5477 if (rx_stats_active)
5478 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5479 call->flags &= ~RX_CALL_TQ_BUSY;
5480 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5481 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5482 #ifdef RX_ENABLE_LOCKS
5483 osirx_AssertMine(&call->lock, "rxi_Start middle");
5484 CV_BROADCAST(&call->cv_tq);
5485 #else /* RX_ENABLE_LOCKS */
5486 osi_rxWakeup(&call->tq);
5487 #endif /* RX_ENABLE_LOCKS */
5489 rxi_CallError(call, call->error);
5492 #ifdef RX_ENABLE_LOCKS
5493 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5495 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5496 /* Some packets have received acks. If they all have, we can clear
5497 * the transmit queue.
5500 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5501 if (p->header.seq < call->tfirst
5502 && (p->flags & RX_PKTFLAG_ACKED)) {
5504 p->flags &= ~RX_PKTFLAG_TQ;
5505 #ifdef RXDEBUG_PACKET
5513 call->flags |= RX_CALL_TQ_CLEARME;
5515 #endif /* RX_ENABLE_LOCKS */
5516 /* Don't bother doing retransmits if the TQ is cleared. */
5517 if (call->flags & RX_CALL_TQ_CLEARME) {
5518 rxi_ClearTransmitQueue(call, 1);
5520 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5523 /* Always post a resend event, if there is anything in the
5524 * queue, and resend is possible. There should be at least
5525 * one unacknowledged packet in the queue ... otherwise none
5526 * of these packets should be on the queue in the first place.
5528 if (call->resendEvent) {
5529 /* Cancel the existing event and post a new one */
5530 rxevent_Cancel(call->resendEvent, call,
5531 RX_CALL_REFCOUNT_RESEND);
5534 /* The retry time is the retry time on the first unacknowledged
5535 * packet inside the current window */
5537 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5538 /* Don't set timers for packets outside the window */
5539 if (p->header.seq >= call->tfirst + call->twind) {
5543 if (!(p->flags & RX_PKTFLAG_ACKED)
5544 && !clock_IsZero(&p->retryTime)) {
5546 retryTime = p->retryTime;
5551 /* Post a new event to re-run rxi_Start when retries may be needed */
5552 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5553 #ifdef RX_ENABLE_LOCKS
5554 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5556 rxevent_PostNow2(&retryTime, &usenow,
5558 (void *)call, 0, istack);
5559 #else /* RX_ENABLE_LOCKS */
5561 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5562 (void *)call, 0, istack);
5563 #endif /* RX_ENABLE_LOCKS */
5566 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5567 } while (call->flags & RX_CALL_NEED_START);
5569 * TQ references no longer protected by this flag; they must remain
5570 * protected by the global lock.
5572 call->flags &= ~RX_CALL_TQ_BUSY;
5573 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5574 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5575 #ifdef RX_ENABLE_LOCKS
5576 osirx_AssertMine(&call->lock, "rxi_Start end");
5577 CV_BROADCAST(&call->cv_tq);
5578 #else /* RX_ENABLE_LOCKS */
5579 osi_rxWakeup(&call->tq);
5580 #endif /* RX_ENABLE_LOCKS */
5583 call->flags |= RX_CALL_NEED_START;
5585 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5587 if (call->resendEvent) {
5588 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5593 /* Also adjusts the keep alive parameters for the call, to reflect
5594 * that we have just sent a packet (so keep alives aren't sent
5597 rxi_Send(struct rx_call *call, struct rx_packet *p,
5600 struct rx_connection *conn = call->conn;
5602 /* Stamp each packet with the user supplied status */
5603 p->header.userStatus = call->localStatus;
5605 /* Allow the security object controlling this call's security to
5606 * make any last-minute changes to the packet */
5607 RXS_SendPacket(conn->securityObject, call, p);
5609 /* Since we're about to send SOME sort of packet to the peer, it's
5610 * safe to nuke any scheduled end-of-packets ack */
5611 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5613 /* Actually send the packet, filling in more connection-specific fields */
5614 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5615 MUTEX_EXIT(&call->lock);
5616 rxi_SendPacket(call, conn, p, istack);
5617 MUTEX_ENTER(&call->lock);
5618 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5620 /* Update last send time for this call (for keep-alive
5621 * processing), and for the connection (so that we can discover
5622 * idle connections) */
5623 conn->lastSendTime = call->lastSendTime = clock_Sec();
5624 /* Don't count keepalives here, so idleness can be tracked. */
5625 if ((p->header.type != RX_PACKET_TYPE_ACK) || (((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING))
5626 call->lastSendData = call->lastSendTime;
5630 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5631 * that things are fine. Also called periodically to guarantee that nothing
5632 * falls through the cracks (e.g. (error + dally) connections have keepalive
5633 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5635 * haveCTLock Set if calling from rxi_ReapConnections
5637 #ifdef RX_ENABLE_LOCKS
5639 rxi_CheckCall(struct rx_call *call, int haveCTLock)
5640 #else /* RX_ENABLE_LOCKS */
5642 rxi_CheckCall(struct rx_call *call)
5643 #endif /* RX_ENABLE_LOCKS */
5645 struct rx_connection *conn = call->conn;
5647 afs_uint32 deadTime;
5649 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5650 if (call->flags & RX_CALL_TQ_BUSY) {
5651 /* Call is active and will be reset by rxi_Start if it's
5652 * in an error state.
5657 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5659 (((afs_uint32) conn->secondsUntilDead << 10) +
5660 ((afs_uint32) conn->peer->rtt >> 3) +
5661 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5663 /* These are computed to the second (+- 1 second). But that's
5664 * good enough for these values, which should be a significant
5665 * number of seconds. */
5666 if (now > (call->lastReceiveTime + deadTime)) {
5667 if (call->state == RX_STATE_ACTIVE) {
5669 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5671 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5672 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5673 ip_stack_t *ipst = ns->netstack_ip;
5675 ire = ire_cache_lookup(call->conn->peer->host
5676 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5678 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5680 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5687 if (ire && ire->ire_max_frag > 0)
5688 rxi_SetPeerMtu(call->conn->peer->host, 0, ire->ire_max_frag);
5689 #if defined(GLOBAL_NETSTACKID)
5693 #endif /* ADAPT_PMTU */
5694 rxi_CallError(call, RX_CALL_DEAD);
5697 #ifdef RX_ENABLE_LOCKS
5698 /* Cancel pending events */
5699 rxevent_Cancel(call->delayedAckEvent, call,
5700 RX_CALL_REFCOUNT_DELAY);
5701 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5702 rxevent_Cancel(call->keepAliveEvent, call,
5703 RX_CALL_REFCOUNT_ALIVE);
5704 if (call->refCount == 0) {
5705 rxi_FreeCall(call, haveCTLock);
5709 #else /* RX_ENABLE_LOCKS */
5712 #endif /* RX_ENABLE_LOCKS */
5714 /* Non-active calls are destroyed if they are not responding
5715 * to pings; active calls are simply flagged in error, so the
5716 * attached process can die reasonably gracefully. */
5718 /* see if we have a non-activity timeout */
5719 if (call->startWait && conn->idleDeadTime
5720 && ((call->startWait + conn->idleDeadTime) < now)) {
5721 if (call->state == RX_STATE_ACTIVE) {
5722 rxi_CallError(call, RX_CALL_TIMEOUT);
5726 if (call->lastSendData && conn->idleDeadTime && (conn->idleDeadErr != 0)
5727 && ((call->lastSendData + conn->idleDeadTime) < now)) {
5728 if (call->state == RX_STATE_ACTIVE) {
5729 rxi_CallError(call, conn->idleDeadErr);
5733 /* see if we have a hard timeout */
5734 if (conn->hardDeadTime
5735 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5736 if (call->state == RX_STATE_ACTIVE)
5737 rxi_CallError(call, RX_CALL_TIMEOUT);
5744 /* When a call is in progress, this routine is called occasionally to
5745 * make sure that some traffic has arrived (or been sent to) the peer.
5746 * If nothing has arrived in a reasonable amount of time, the call is
5747 * declared dead; if nothing has been sent for a while, we send a
5748 * keep-alive packet (if we're actually trying to keep the call alive)
5751 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
5753 struct rx_call *call = arg1;
5754 struct rx_connection *conn;
5757 MUTEX_ENTER(&call->lock);
5758 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5759 if (event == call->keepAliveEvent)
5760 call->keepAliveEvent = NULL;
5763 #ifdef RX_ENABLE_LOCKS
5764 if (rxi_CheckCall(call, 0)) {
5765 MUTEX_EXIT(&call->lock);
5768 #else /* RX_ENABLE_LOCKS */
5769 if (rxi_CheckCall(call))
5771 #endif /* RX_ENABLE_LOCKS */
5773 /* Don't try to keep alive dallying calls */
5774 if (call->state == RX_STATE_DALLY) {
5775 MUTEX_EXIT(&call->lock);
5780 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5781 /* Don't try to send keepalives if there is unacknowledged data */
5782 /* the rexmit code should be good enough, this little hack
5783 * doesn't quite work XXX */
5784 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5786 rxi_ScheduleKeepAliveEvent(call);
5787 MUTEX_EXIT(&call->lock);
5792 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
5794 if (!call->keepAliveEvent) {
5795 struct clock when, now;
5796 clock_GetTime(&now);
5798 when.sec += call->conn->secondsUntilPing;
5799 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5800 call->keepAliveEvent =
5801 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
5805 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5807 rxi_KeepAliveOn(struct rx_call *call)
5809 /* Pretend last packet received was received now--i.e. if another
5810 * packet isn't received within the keep alive time, then the call
5811 * will die; Initialize last send time to the current time--even
5812 * if a packet hasn't been sent yet. This will guarantee that a
5813 * keep-alive is sent within the ping time */
5814 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5815 rxi_ScheduleKeepAliveEvent(call);
5818 /* This routine is called to send connection abort messages
5819 * that have been delayed to throttle looping clients. */
5821 rxi_SendDelayedConnAbort(struct rxevent *event,
5822 void *arg1, void *unused)
5824 struct rx_connection *conn = arg1;
5827 struct rx_packet *packet;
5829 MUTEX_ENTER(&conn->conn_data_lock);
5830 conn->delayedAbortEvent = NULL;
5831 error = htonl(conn->error);
5833 MUTEX_EXIT(&conn->conn_data_lock);
5834 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5837 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5838 RX_PACKET_TYPE_ABORT, (char *)&error,
5840 rxi_FreePacket(packet);
5844 /* This routine is called to send call abort messages
5845 * that have been delayed to throttle looping clients. */
5847 rxi_SendDelayedCallAbort(struct rxevent *event,
5848 void *arg1, void *dummy)
5850 struct rx_call *call = arg1;
5853 struct rx_packet *packet;
5855 MUTEX_ENTER(&call->lock);
5856 call->delayedAbortEvent = NULL;
5857 error = htonl(call->error);
5859 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5862 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5863 (char *)&error, sizeof(error), 0);
5864 rxi_FreePacket(packet);
5866 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5867 MUTEX_EXIT(&call->lock);
5870 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5871 * seconds) to ask the client to authenticate itself. The routine
5872 * issues a challenge to the client, which is obtained from the
5873 * security object associated with the connection */
5875 rxi_ChallengeEvent(struct rxevent *event,
5876 void *arg0, void *arg1, int tries)
5878 struct rx_connection *conn = arg0;
5880 conn->challengeEvent = NULL;
5881 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5882 struct rx_packet *packet;
5883 struct clock when, now;
5886 /* We've failed to authenticate for too long.
5887 * Reset any calls waiting for authentication;
5888 * they are all in RX_STATE_PRECALL.
5892 MUTEX_ENTER(&conn->conn_call_lock);
5893 for (i = 0; i < RX_MAXCALLS; i++) {
5894 struct rx_call *call = conn->call[i];
5896 MUTEX_ENTER(&call->lock);
5897 if (call->state == RX_STATE_PRECALL) {
5898 rxi_CallError(call, RX_CALL_DEAD);
5899 rxi_SendCallAbort(call, NULL, 0, 0);
5901 MUTEX_EXIT(&call->lock);
5904 MUTEX_EXIT(&conn->conn_call_lock);
5908 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5910 /* If there's no packet available, do this later. */
5911 RXS_GetChallenge(conn->securityObject, conn, packet);
5912 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5913 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5914 rxi_FreePacket(packet);
5916 clock_GetTime(&now);
5918 when.sec += RX_CHALLENGE_TIMEOUT;
5919 conn->challengeEvent =
5920 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
5925 /* Call this routine to start requesting the client to authenticate
5926 * itself. This will continue until authentication is established,
5927 * the call times out, or an invalid response is returned. The
5928 * security object associated with the connection is asked to create
5929 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5930 * defined earlier. */
5932 rxi_ChallengeOn(struct rx_connection *conn)
5934 if (!conn->challengeEvent) {
5935 RXS_CreateChallenge(conn->securityObject, conn);
5936 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5941 /* Compute round trip time of the packet provided, in *rttp.
5944 /* rxi_ComputeRoundTripTime is called with peer locked. */
5945 /* sentp and/or peer may be null */
5947 rxi_ComputeRoundTripTime(struct rx_packet *p,
5948 struct clock *sentp,
5949 struct rx_peer *peer)
5951 struct clock thisRtt, *rttp = &thisRtt;
5955 clock_GetTime(rttp);
5957 if (clock_Lt(rttp, sentp)) {
5959 return; /* somebody set the clock back, don't count this time. */
5961 clock_Sub(rttp, sentp);
5962 if (rx_stats_active) {
5963 MUTEX_ENTER(&rx_stats_mutex);
5964 if (clock_Lt(rttp, &rx_stats.minRtt))
5965 rx_stats.minRtt = *rttp;
5966 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5967 if (rttp->sec > 60) {
5968 MUTEX_EXIT(&rx_stats_mutex);
5969 return; /* somebody set the clock ahead */
5971 rx_stats.maxRtt = *rttp;
5973 clock_Add(&rx_stats.totalRtt, rttp);
5974 rx_stats.nRttSamples++;
5975 MUTEX_EXIT(&rx_stats_mutex);
5978 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5980 /* Apply VanJacobson round-trip estimations */
5985 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5986 * srtt is stored as fixed point with 3 bits after the binary
5987 * point (i.e., scaled by 8). The following magic is
5988 * equivalent to the smoothing algorithm in rfc793 with an
5989 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5990 * srtt*8 = srtt*8 + rtt - srtt
5991 * srtt = srtt + rtt/8 - srtt/8
5994 delta = MSEC(rttp) - (peer->rtt >> 3);
5998 * We accumulate a smoothed rtt variance (actually, a smoothed
5999 * mean difference), then set the retransmit timer to smoothed
6000 * rtt + 4 times the smoothed variance (was 2x in van's original
6001 * paper, but 4x works better for me, and apparently for him as
6003 * rttvar is stored as
6004 * fixed point with 2 bits after the binary point (scaled by
6005 * 4). The following is equivalent to rfc793 smoothing with
6006 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
6007 * replaces rfc793's wired-in beta.
6008 * dev*4 = dev*4 + (|actual - expected| - dev)
6014 delta -= (peer->rtt_dev >> 2);
6015 peer->rtt_dev += delta;
6017 /* I don't have a stored RTT so I start with this value. Since I'm
6018 * probably just starting a call, and will be pushing more data down
6019 * this, I expect congestion to increase rapidly. So I fudge a
6020 * little, and I set deviance to half the rtt. In practice,
6021 * deviance tends to approach something a little less than
6022 * half the smoothed rtt. */
6023 peer->rtt = (MSEC(rttp) << 3) + 8;
6024 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
6026 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
6027 * the other of these connections is usually in a user process, and can
6028 * be switched and/or swapped out. So on fast, reliable networks, the
6029 * timeout would otherwise be too short.
6031 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
6032 clock_Zero(&(peer->timeout));
6033 clock_Addmsec(&(peer->timeout), rtt_timeout);
6035 dpf(("rxi_ComputeRoundTripTime(rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%0.3d sec)\n", MSEC(rttp), peer->rtt >> 3, peer->rtt_dev >> 2, (peer->timeout.sec), (peer->timeout.usec)));
6039 /* Find all server connections that have not been active for a long time, and
6042 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6044 struct clock now, when;
6045 clock_GetTime(&now);
6047 /* Find server connection structures that haven't been used for
6048 * greater than rx_idleConnectionTime */
6050 struct rx_connection **conn_ptr, **conn_end;
6051 int i, havecalls = 0;
6052 MUTEX_ENTER(&rx_connHashTable_lock);
6053 for (conn_ptr = &rx_connHashTable[0], conn_end =
6054 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6056 struct rx_connection *conn, *next;
6057 struct rx_call *call;
6061 for (conn = *conn_ptr; conn; conn = next) {
6062 /* XXX -- Shouldn't the connection be locked? */
6065 for (i = 0; i < RX_MAXCALLS; i++) {
6066 call = conn->call[i];
6070 code = MUTEX_TRYENTER(&call->lock);
6073 #ifdef RX_ENABLE_LOCKS
6074 result = rxi_CheckCall(call, 1);
6075 #else /* RX_ENABLE_LOCKS */
6076 result = rxi_CheckCall(call);
6077 #endif /* RX_ENABLE_LOCKS */
6078 MUTEX_EXIT(&call->lock);
6080 /* If CheckCall freed the call, it might
6081 * have destroyed the connection as well,
6082 * which screws up the linked lists.
6088 if (conn->type == RX_SERVER_CONNECTION) {
6089 /* This only actually destroys the connection if
6090 * there are no outstanding calls */
6091 MUTEX_ENTER(&conn->conn_data_lock);
6092 if (!havecalls && !conn->refCount
6093 && ((conn->lastSendTime + rx_idleConnectionTime) <
6095 conn->refCount++; /* it will be decr in rx_DestroyConn */
6096 MUTEX_EXIT(&conn->conn_data_lock);
6097 #ifdef RX_ENABLE_LOCKS
6098 rxi_DestroyConnectionNoLock(conn);
6099 #else /* RX_ENABLE_LOCKS */
6100 rxi_DestroyConnection(conn);
6101 #endif /* RX_ENABLE_LOCKS */
6103 #ifdef RX_ENABLE_LOCKS
6105 MUTEX_EXIT(&conn->conn_data_lock);
6107 #endif /* RX_ENABLE_LOCKS */
6111 #ifdef RX_ENABLE_LOCKS
6112 while (rx_connCleanup_list) {
6113 struct rx_connection *conn;
6114 conn = rx_connCleanup_list;
6115 rx_connCleanup_list = rx_connCleanup_list->next;
6116 MUTEX_EXIT(&rx_connHashTable_lock);
6117 rxi_CleanupConnection(conn);
6118 MUTEX_ENTER(&rx_connHashTable_lock);
6120 MUTEX_EXIT(&rx_connHashTable_lock);
6121 #endif /* RX_ENABLE_LOCKS */
6124 /* Find any peer structures that haven't been used (haven't had an
6125 * associated connection) for greater than rx_idlePeerTime */
6127 struct rx_peer **peer_ptr, **peer_end;
6129 MUTEX_ENTER(&rx_rpc_stats);
6130 MUTEX_ENTER(&rx_peerHashTable_lock);
6131 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6132 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6134 struct rx_peer *peer, *next, *prev;
6135 for (prev = peer = *peer_ptr; peer; peer = next) {
6137 code = MUTEX_TRYENTER(&peer->peer_lock);
6138 if ((code) && (peer->refCount == 0)
6139 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6140 rx_interface_stat_p rpc_stat, nrpc_stat;
6142 MUTEX_EXIT(&peer->peer_lock);
6143 MUTEX_DESTROY(&peer->peer_lock);
6145 (&peer->rpcStats, rpc_stat, nrpc_stat,
6146 rx_interface_stat)) {
6147 unsigned int num_funcs;
6150 queue_Remove(&rpc_stat->queue_header);
6151 queue_Remove(&rpc_stat->all_peers);
6152 num_funcs = rpc_stat->stats[0].func_total;
6154 sizeof(rx_interface_stat_t) +
6155 rpc_stat->stats[0].func_total *
6156 sizeof(rx_function_entry_v1_t);
6158 rxi_Free(rpc_stat, space);
6159 rxi_rpc_peer_stat_cnt -= num_funcs;
6162 if (rx_stats_active)
6163 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6164 if (peer == *peer_ptr) {
6171 MUTEX_EXIT(&peer->peer_lock);
6177 MUTEX_EXIT(&rx_peerHashTable_lock);
6178 MUTEX_EXIT(&rx_rpc_stats);
6181 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6182 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6183 * GC, just below. Really, we shouldn't have to keep moving packets from
6184 * one place to another, but instead ought to always know if we can
6185 * afford to hold onto a packet in its particular use. */
6186 MUTEX_ENTER(&rx_freePktQ_lock);
6187 if (rx_waitingForPackets) {
6188 rx_waitingForPackets = 0;
6189 #ifdef RX_ENABLE_LOCKS
6190 CV_BROADCAST(&rx_waitingForPackets_cv);
6192 osi_rxWakeup(&rx_waitingForPackets);
6195 MUTEX_EXIT(&rx_freePktQ_lock);
6198 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6199 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6203 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6204 * rx.h is sort of strange this is better. This is called with a security
6205 * object before it is discarded. Each connection using a security object has
6206 * its own refcount to the object so it won't actually be freed until the last
6207 * connection is destroyed.
6209 * This is the only rxs module call. A hold could also be written but no one
6213 rxs_Release(struct rx_securityClass *aobj)
6215 return RXS_Close(aobj);
6219 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6220 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6221 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6222 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6224 /* Adjust our estimate of the transmission rate to this peer, given
6225 * that the packet p was just acked. We can adjust peer->timeout and
6226 * call->twind. Pragmatically, this is called
6227 * only with packets of maximal length.
6228 * Called with peer and call locked.
6232 rxi_ComputeRate(struct rx_peer *peer, struct rx_call *call,
6233 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6235 afs_int32 xferSize, xferMs;
6239 /* Count down packets */
6240 if (peer->rateFlag > 0)
6242 /* Do nothing until we're enabled */
6243 if (peer->rateFlag != 0)
6248 /* Count only when the ack seems legitimate */
6249 switch (ackReason) {
6250 case RX_ACK_REQUESTED:
6252 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6256 case RX_ACK_PING_RESPONSE:
6257 if (p) /* want the response to ping-request, not data send */
6259 clock_GetTime(&newTO);
6260 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6261 clock_Sub(&newTO, &call->pingRequestTime);
6262 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6266 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6273 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %lu.%06lu, rtt %u, ps %u)", ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"), xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
6275 /* Track only packets that are big enough. */
6276 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6280 /* absorb RTT data (in milliseconds) for these big packets */
6281 if (peer->smRtt == 0) {
6282 peer->smRtt = xferMs;
6284 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6289 if (peer->countDown) {
6293 peer->countDown = 10; /* recalculate only every so often */
6295 /* In practice, we can measure only the RTT for full packets,
6296 * because of the way Rx acks the data that it receives. (If it's
6297 * smaller than a full packet, it often gets implicitly acked
6298 * either by the call response (from a server) or by the next call
6299 * (from a client), and either case confuses transmission times
6300 * with processing times.) Therefore, replace the above
6301 * more-sophisticated processing with a simpler version, where the
6302 * smoothed RTT is kept for full-size packets, and the time to
6303 * transmit a windowful of full-size packets is simply RTT *
6304 * windowSize. Again, we take two steps:
6305 - ensure the timeout is large enough for a single packet's RTT;
6306 - ensure that the window is small enough to fit in the desired timeout.*/
6308 /* First, the timeout check. */
6309 minTime = peer->smRtt;
6310 /* Get a reasonable estimate for a timeout period */
6312 newTO.sec = minTime / 1000;
6313 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6315 /* Increase the timeout period so that we can always do at least
6316 * one packet exchange */
6317 if (clock_Gt(&newTO, &peer->timeout)) {
6319 dpf(("CONG peer %lx/%u: timeout %lu.%06lu ==> %lu.%06lu (rtt %u, ps %u)", ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec, newTO.sec, newTO.usec, peer->smRtt, peer->packetSize));
6321 peer->timeout = newTO;
6324 /* Now, get an estimate for the transmit window size. */
6325 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6326 /* Now, convert to the number of full packets that could fit in a
6327 * reasonable fraction of that interval */
6328 minTime /= (peer->smRtt << 1);
6329 xferSize = minTime; /* (make a copy) */
6331 /* Now clamp the size to reasonable bounds. */
6334 else if (minTime > rx_Window)
6335 minTime = rx_Window;
6336 /* if (minTime != peer->maxWindow) {
6337 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6338 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6339 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6341 peer->maxWindow = minTime;
6342 elide... call->twind = minTime;
6346 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6347 * Discern this by calculating the timeout necessary for rx_Window
6349 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6350 /* calculate estimate for transmission interval in milliseconds */
6351 minTime = rx_Window * peer->smRtt;
6352 if (minTime < 1000) {
6353 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6354 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6355 peer->timeout.usec, peer->smRtt, peer->packetSize));
6357 newTO.sec = 0; /* cut back on timeout by half a second */
6358 newTO.usec = 500000;
6359 clock_Sub(&peer->timeout, &newTO);
6364 } /* end of rxi_ComputeRate */
6365 #endif /* ADAPT_WINDOW */
6373 #define TRACE_OPTION_RX_DEBUG 16
6381 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6382 0, KEY_QUERY_VALUE, &parmKey);
6383 if (code != ERROR_SUCCESS)
6386 dummyLen = sizeof(TraceOption);
6387 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6388 (BYTE *) &TraceOption, &dummyLen);
6389 if (code == ERROR_SUCCESS) {
6390 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
6392 RegCloseKey (parmKey);
6393 #endif /* AFS_NT40_ENV */
6398 rx_DebugOnOff(int on)
6402 rxdebug_active = on;
6408 rx_StatsOnOff(int on)
6411 rx_stats_active = on;
6416 /* Don't call this debugging routine directly; use dpf */
6418 rxi_DebugPrint(char *format, ...)
6427 va_start(ap, format);
6429 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6432 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
6434 if (msg[len-1] != '\n') {
6438 OutputDebugString(msg);
6445 va_start(ap, format);
6447 clock_GetTime(&now);
6448 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6449 (unsigned int)now.usec / 1000);
6450 vfprintf(rx_Log, format, ap);
6459 * This function is used to process the rx_stats structure that is local
6460 * to a process as well as an rx_stats structure received from a remote
6461 * process (via rxdebug). Therefore, it needs to do minimal version
6465 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
6466 afs_int32 freePackets, char version)
6471 if (size != sizeof(struct rx_statistics)) {
6473 "Unexpected size of stats structure: was %d, expected %lud\n",
6474 size, sizeof(struct rx_statistics));
6477 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6480 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6481 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
6482 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6483 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6484 s->specialPktAllocFailures);
6486 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
6487 s->receivePktAllocFailures, s->sendPktAllocFailures,
6488 s->specialPktAllocFailures);
6492 " greedy %u, " "bogusReads %u (last from host %x), "
6493 "noPackets %u, " "noBuffers %u, " "selects %u, "
6494 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
6495 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6496 s->selects, s->sendSelects);
6498 fprintf(file, " packets read: ");
6499 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6500 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
6502 fprintf(file, "\n");
6505 " other read counters: data %u, " "ack %u, " "dup %u "
6506 "spurious %u " "dally %u\n", s->dataPacketsRead,
6507 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6508 s->ignorePacketDally);
6510 fprintf(file, " packets sent: ");
6511 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6512 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
6514 fprintf(file, "\n");
6517 " other send counters: ack %u, " "data %u (not resends), "
6518 "resends %u, " "pushed %u, " "acked&ignored %u\n",
6519 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6520 s->dataPacketsPushed, s->ignoreAckedPacket);
6523 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
6524 s->netSendFailures, (int)s->fatalErrors);
6526 if (s->nRttSamples) {
6527 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6528 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6530 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6531 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6535 " %d server connections, " "%d client connections, "
6536 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6537 s->nServerConns, s->nClientConns, s->nPeerStructs,
6538 s->nCallStructs, s->nFreeCallStructs);
6540 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6541 fprintf(file, " %d clock updates\n", clock_nUpdates);
6544 fprintf(file, "ERROR: compiled without RXDEBUG\n");
6548 /* for backward compatibility */
6550 rx_PrintStats(FILE * file)
6552 MUTEX_ENTER(&rx_stats_mutex);
6553 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6555 MUTEX_EXIT(&rx_stats_mutex);
6559 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6561 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6562 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6563 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6566 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6567 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6568 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6571 " Packet size %d, " "max in packet skew %d, "
6572 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6573 (int)peer->outPacketSkew);
6577 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
6579 * This mutex protects the following static variables:
6583 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
6584 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
6586 #define LOCK_RX_DEBUG
6587 #define UNLOCK_RX_DEBUG
6588 #endif /* AFS_PTHREAD_ENV */
6592 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6593 u_char type, void *inputData, size_t inputLength,
6594 void *outputData, size_t outputLength)
6596 static afs_int32 counter = 100;
6597 time_t waitTime, waitCount, startTime;
6598 struct rx_header theader;
6601 struct timeval tv_now, tv_wake, tv_delta;
6602 struct sockaddr_in taddr, faddr;
6611 startTime = time(0);
6617 tp = &tbuffer[sizeof(struct rx_header)];
6618 taddr.sin_family = AF_INET;
6619 taddr.sin_port = remotePort;
6620 taddr.sin_addr.s_addr = remoteAddr;
6621 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6622 taddr.sin_len = sizeof(struct sockaddr_in);
6625 memset(&theader, 0, sizeof(theader));
6626 theader.epoch = htonl(999);
6628 theader.callNumber = htonl(counter);
6631 theader.type = type;
6632 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6633 theader.serviceId = 0;
6635 memcpy(tbuffer, &theader, sizeof(theader));
6636 memcpy(tp, inputData, inputLength);
6638 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6639 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6641 /* see if there's a packet available */
6642 gettimeofday(&tv_wake,0);
6643 tv_wake.tv_sec += waitTime;
6646 FD_SET(socket, &imask);
6647 tv_delta.tv_sec = tv_wake.tv_sec;
6648 tv_delta.tv_usec = tv_wake.tv_usec;
6649 gettimeofday(&tv_now, 0);
6651 if (tv_delta.tv_usec < tv_now.tv_usec) {
6653 tv_delta.tv_usec += 1000000;
6656 tv_delta.tv_usec -= tv_now.tv_usec;
6658 if (tv_delta.tv_sec < tv_now.tv_sec) {
6662 tv_delta.tv_sec -= tv_now.tv_sec;
6664 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6665 if (code == 1 && FD_ISSET(socket, &imask)) {
6666 /* now receive a packet */
6667 faddrLen = sizeof(struct sockaddr_in);
6669 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6670 (struct sockaddr *)&faddr, &faddrLen);
6673 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6674 if (counter == ntohl(theader.callNumber))
6682 /* see if we've timed out */
6690 code -= sizeof(struct rx_header);
6691 if (code > outputLength)
6692 code = outputLength;
6693 memcpy(outputData, tp, code);
6696 #endif /* RXDEBUG */
6699 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6700 afs_uint16 remotePort, struct rx_debugStats * stat,
6701 afs_uint32 * supportedValues)
6707 struct rx_debugIn in;
6708 afs_int32 *lp = (afs_int32 *) stat;
6710 *supportedValues = 0;
6711 in.type = htonl(RX_DEBUGI_GETSTATS);
6714 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6715 &in, sizeof(in), stat, sizeof(*stat));
6718 * If the call was successful, fixup the version and indicate
6719 * what contents of the stat structure are valid.
6720 * Also do net to host conversion of fields here.
6724 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6725 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6727 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6728 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6730 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6731 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6733 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6734 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6736 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6737 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6739 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6740 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6742 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6743 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6745 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6746 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6748 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
6749 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
6751 stat->nFreePackets = ntohl(stat->nFreePackets);
6752 stat->packetReclaims = ntohl(stat->packetReclaims);
6753 stat->callsExecuted = ntohl(stat->callsExecuted);
6754 stat->nWaiting = ntohl(stat->nWaiting);
6755 stat->idleThreads = ntohl(stat->idleThreads);
6756 stat->nWaited = ntohl(stat->nWaited);
6757 stat->nPackets = ntohl(stat->nPackets);
6764 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6765 afs_uint16 remotePort, struct rx_statistics * stat,
6766 afs_uint32 * supportedValues)
6772 struct rx_debugIn in;
6773 afs_int32 *lp = (afs_int32 *) stat;
6777 * supportedValues is currently unused, but added to allow future
6778 * versioning of this function.
6781 *supportedValues = 0;
6782 in.type = htonl(RX_DEBUGI_RXSTATS);
6784 memset(stat, 0, sizeof(*stat));
6786 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6787 &in, sizeof(in), stat, sizeof(*stat));
6792 * Do net to host conversion here
6795 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6804 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6805 afs_uint16 remotePort, size_t version_length,
6810 return MakeDebugCall(socket, remoteAddr, remotePort,
6811 RX_PACKET_TYPE_VERSION, a, 1, version,
6819 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6820 afs_uint16 remotePort, afs_int32 * nextConnection,
6821 int allConnections, afs_uint32 debugSupportedValues,
6822 struct rx_debugConn * conn,
6823 afs_uint32 * supportedValues)
6829 struct rx_debugIn in;
6833 * supportedValues is currently unused, but added to allow future
6834 * versioning of this function.
6837 *supportedValues = 0;
6838 if (allConnections) {
6839 in.type = htonl(RX_DEBUGI_GETALLCONN);
6841 in.type = htonl(RX_DEBUGI_GETCONN);
6843 in.index = htonl(*nextConnection);
6844 memset(conn, 0, sizeof(*conn));
6846 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6847 &in, sizeof(in), conn, sizeof(*conn));
6850 *nextConnection += 1;
6853 * Convert old connection format to new structure.
6856 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6857 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6858 #define MOVEvL(a) (conn->a = vL->a)
6860 /* any old or unrecognized version... */
6861 for (i = 0; i < RX_MAXCALLS; i++) {
6862 MOVEvL(callState[i]);
6863 MOVEvL(callMode[i]);
6864 MOVEvL(callFlags[i]);
6865 MOVEvL(callOther[i]);
6867 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6868 MOVEvL(secStats.type);
6869 MOVEvL(secStats.level);
6870 MOVEvL(secStats.flags);
6871 MOVEvL(secStats.expires);
6872 MOVEvL(secStats.packetsReceived);
6873 MOVEvL(secStats.packetsSent);
6874 MOVEvL(secStats.bytesReceived);
6875 MOVEvL(secStats.bytesSent);
6880 * Do net to host conversion here
6882 * I don't convert host or port since we are most likely
6883 * going to want these in NBO.
6885 conn->cid = ntohl(conn->cid);
6886 conn->serial = ntohl(conn->serial);
6887 for (i = 0; i < RX_MAXCALLS; i++) {
6888 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6890 conn->error = ntohl(conn->error);
6891 conn->secStats.flags = ntohl(conn->secStats.flags);
6892 conn->secStats.expires = ntohl(conn->secStats.expires);
6893 conn->secStats.packetsReceived =
6894 ntohl(conn->secStats.packetsReceived);
6895 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6896 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6897 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6898 conn->epoch = ntohl(conn->epoch);
6899 conn->natMTU = ntohl(conn->natMTU);
6906 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6907 afs_uint16 remotePort, afs_int32 * nextPeer,
6908 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6909 afs_uint32 * supportedValues)
6915 struct rx_debugIn in;
6918 * supportedValues is currently unused, but added to allow future
6919 * versioning of this function.
6922 *supportedValues = 0;
6923 in.type = htonl(RX_DEBUGI_GETPEER);
6924 in.index = htonl(*nextPeer);
6925 memset(peer, 0, sizeof(*peer));
6927 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6928 &in, sizeof(in), peer, sizeof(*peer));
6934 * Do net to host conversion here
6936 * I don't convert host or port since we are most likely
6937 * going to want these in NBO.
6939 peer->ifMTU = ntohs(peer->ifMTU);
6940 peer->idleWhen = ntohl(peer->idleWhen);
6941 peer->refCount = ntohs(peer->refCount);
6942 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6943 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6944 peer->rtt = ntohl(peer->rtt);
6945 peer->rtt_dev = ntohl(peer->rtt_dev);
6946 peer->timeout.sec = ntohl(peer->timeout.sec);
6947 peer->timeout.usec = ntohl(peer->timeout.usec);
6948 peer->nSent = ntohl(peer->nSent);
6949 peer->reSends = ntohl(peer->reSends);
6950 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6951 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6952 peer->rateFlag = ntohl(peer->rateFlag);
6953 peer->natMTU = ntohs(peer->natMTU);
6954 peer->maxMTU = ntohs(peer->maxMTU);
6955 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6956 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6957 peer->MTU = ntohs(peer->MTU);
6958 peer->cwind = ntohs(peer->cwind);
6959 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6960 peer->congestSeq = ntohs(peer->congestSeq);
6961 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6962 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6963 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6964 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6973 struct rx_serverQueueEntry *np;
6976 struct rx_call *call;
6977 struct rx_serverQueueEntry *sq;
6981 if (rxinit_status == 1) {
6983 return; /* Already shutdown. */
6987 #ifndef AFS_PTHREAD_ENV
6988 FD_ZERO(&rx_selectMask);
6989 #endif /* AFS_PTHREAD_ENV */
6990 rxi_dataQuota = RX_MAX_QUOTA;
6991 #ifndef AFS_PTHREAD_ENV
6993 #endif /* AFS_PTHREAD_ENV */
6996 #ifndef AFS_PTHREAD_ENV
6997 #ifndef AFS_USE_GETTIMEOFDAY
6999 #endif /* AFS_USE_GETTIMEOFDAY */
7000 #endif /* AFS_PTHREAD_ENV */
7002 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7003 call = queue_First(&rx_freeCallQueue, rx_call);
7005 rxi_Free(call, sizeof(struct rx_call));
7008 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7009 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7015 struct rx_peer **peer_ptr, **peer_end;
7016 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7017 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7019 struct rx_peer *peer, *next;
7020 for (peer = *peer_ptr; peer; peer = next) {
7021 rx_interface_stat_p rpc_stat, nrpc_stat;
7024 (&peer->rpcStats, rpc_stat, nrpc_stat,
7025 rx_interface_stat)) {
7026 unsigned int num_funcs;
7029 queue_Remove(&rpc_stat->queue_header);
7030 queue_Remove(&rpc_stat->all_peers);
7031 num_funcs = rpc_stat->stats[0].func_total;
7033 sizeof(rx_interface_stat_t) +
7034 rpc_stat->stats[0].func_total *
7035 sizeof(rx_function_entry_v1_t);
7037 rxi_Free(rpc_stat, space);
7038 MUTEX_ENTER(&rx_rpc_stats);
7039 rxi_rpc_peer_stat_cnt -= num_funcs;
7040 MUTEX_EXIT(&rx_rpc_stats);
7044 if (rx_stats_active)
7045 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
7049 for (i = 0; i < RX_MAX_SERVICES; i++) {
7051 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7053 for (i = 0; i < rx_hashTableSize; i++) {
7054 struct rx_connection *tc, *ntc;
7055 MUTEX_ENTER(&rx_connHashTable_lock);
7056 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7058 for (j = 0; j < RX_MAXCALLS; j++) {
7060 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7063 rxi_Free(tc, sizeof(*tc));
7065 MUTEX_EXIT(&rx_connHashTable_lock);
7068 MUTEX_ENTER(&freeSQEList_lock);
7070 while ((np = rx_FreeSQEList)) {
7071 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7072 MUTEX_DESTROY(&np->lock);
7073 rxi_Free(np, sizeof(*np));
7076 MUTEX_EXIT(&freeSQEList_lock);
7077 MUTEX_DESTROY(&freeSQEList_lock);
7078 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7079 MUTEX_DESTROY(&rx_connHashTable_lock);
7080 MUTEX_DESTROY(&rx_peerHashTable_lock);
7081 MUTEX_DESTROY(&rx_serverPool_lock);
7083 osi_Free(rx_connHashTable,
7084 rx_hashTableSize * sizeof(struct rx_connection *));
7085 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7087 UNPIN(rx_connHashTable,
7088 rx_hashTableSize * sizeof(struct rx_connection *));
7089 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7091 rxi_FreeAllPackets();
7093 MUTEX_ENTER(&rx_quota_mutex);
7094 rxi_dataQuota = RX_MAX_QUOTA;
7095 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7096 MUTEX_EXIT(&rx_quota_mutex);
7101 #ifdef RX_ENABLE_LOCKS
7103 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7105 if (!MUTEX_ISMINE(lockaddr))
7106 osi_Panic("Lock not held: %s", msg);
7108 #endif /* RX_ENABLE_LOCKS */
7113 * Routines to implement connection specific data.
7117 rx_KeyCreate(rx_destructor_t rtn)
7120 MUTEX_ENTER(&rxi_keyCreate_lock);
7121 key = rxi_keyCreate_counter++;
7122 rxi_keyCreate_destructor = (rx_destructor_t *)
7123 realloc((void *)rxi_keyCreate_destructor,
7124 (key + 1) * sizeof(rx_destructor_t));
7125 rxi_keyCreate_destructor[key] = rtn;
7126 MUTEX_EXIT(&rxi_keyCreate_lock);
7131 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7134 MUTEX_ENTER(&conn->conn_data_lock);
7135 if (!conn->specific) {
7136 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7137 for (i = 0; i < key; i++)
7138 conn->specific[i] = NULL;
7139 conn->nSpecific = key + 1;
7140 conn->specific[key] = ptr;
7141 } else if (key >= conn->nSpecific) {
7142 conn->specific = (void **)
7143 realloc(conn->specific, (key + 1) * sizeof(void *));
7144 for (i = conn->nSpecific; i < key; i++)
7145 conn->specific[i] = NULL;
7146 conn->nSpecific = key + 1;
7147 conn->specific[key] = ptr;
7149 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7150 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7151 conn->specific[key] = ptr;
7153 MUTEX_EXIT(&conn->conn_data_lock);
7157 rx_GetSpecific(struct rx_connection *conn, int key)
7160 MUTEX_ENTER(&conn->conn_data_lock);
7161 if (key >= conn->nSpecific)
7164 ptr = conn->specific[key];
7165 MUTEX_EXIT(&conn->conn_data_lock);
7169 #endif /* !KERNEL */
7172 * processStats is a queue used to store the statistics for the local
7173 * process. Its contents are similar to the contents of the rpcStats
7174 * queue on a rx_peer structure, but the actual data stored within
7175 * this queue contains totals across the lifetime of the process (assuming
7176 * the stats have not been reset) - unlike the per peer structures
7177 * which can come and go based upon the peer lifetime.
7180 static struct rx_queue processStats = { &processStats, &processStats };
7183 * peerStats is a queue used to store the statistics for all peer structs.
7184 * Its contents are the union of all the peer rpcStats queues.
7187 static struct rx_queue peerStats = { &peerStats, &peerStats };
7190 * rxi_monitor_processStats is used to turn process wide stat collection
7194 static int rxi_monitor_processStats = 0;
7197 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7200 static int rxi_monitor_peerStats = 0;
7203 * rxi_AddRpcStat - given all of the information for a particular rpc
7204 * call, create (if needed) and update the stat totals for the rpc.
7208 * IN stats - the queue of stats that will be updated with the new value
7210 * IN rxInterface - a unique number that identifies the rpc interface
7212 * IN currentFunc - the index of the function being invoked
7214 * IN totalFunc - the total number of functions in this interface
7216 * IN queueTime - the amount of time this function waited for a thread
7218 * IN execTime - the amount of time this function invocation took to execute
7220 * IN bytesSent - the number bytes sent by this invocation
7222 * IN bytesRcvd - the number bytes received by this invocation
7224 * IN isServer - if true, this invocation was made to a server
7226 * IN remoteHost - the ip address of the remote host
7228 * IN remotePort - the port of the remote host
7230 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7232 * INOUT counter - if a new stats structure is allocated, the counter will
7233 * be updated with the new number of allocated stat structures
7241 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
7242 afs_uint32 currentFunc, afs_uint32 totalFunc,
7243 struct clock *queueTime, struct clock *execTime,
7244 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
7245 afs_uint32 remoteHost, afs_uint32 remotePort,
7246 int addToPeerList, unsigned int *counter)
7249 rx_interface_stat_p rpc_stat, nrpc_stat;
7252 * See if there's already a structure for this interface
7255 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7256 if ((rpc_stat->stats[0].interfaceId == rxInterface)
7257 && (rpc_stat->stats[0].remote_is_server == isServer))
7262 * Didn't find a match so allocate a new structure and add it to the
7266 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
7267 || (rpc_stat->stats[0].interfaceId != rxInterface)
7268 || (rpc_stat->stats[0].remote_is_server != isServer)) {
7273 sizeof(rx_interface_stat_t) +
7274 totalFunc * sizeof(rx_function_entry_v1_t);
7276 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
7277 if (rpc_stat == NULL) {
7281 *counter += totalFunc;
7282 for (i = 0; i < totalFunc; i++) {
7283 rpc_stat->stats[i].remote_peer = remoteHost;
7284 rpc_stat->stats[i].remote_port = remotePort;
7285 rpc_stat->stats[i].remote_is_server = isServer;
7286 rpc_stat->stats[i].interfaceId = rxInterface;
7287 rpc_stat->stats[i].func_total = totalFunc;
7288 rpc_stat->stats[i].func_index = i;
7289 hzero(rpc_stat->stats[i].invocations);
7290 hzero(rpc_stat->stats[i].bytes_sent);
7291 hzero(rpc_stat->stats[i].bytes_rcvd);
7292 rpc_stat->stats[i].queue_time_sum.sec = 0;
7293 rpc_stat->stats[i].queue_time_sum.usec = 0;
7294 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7295 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7296 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7297 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7298 rpc_stat->stats[i].queue_time_max.sec = 0;
7299 rpc_stat->stats[i].queue_time_max.usec = 0;
7300 rpc_stat->stats[i].execution_time_sum.sec = 0;
7301 rpc_stat->stats[i].execution_time_sum.usec = 0;
7302 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7303 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7304 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7305 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7306 rpc_stat->stats[i].execution_time_max.sec = 0;
7307 rpc_stat->stats[i].execution_time_max.usec = 0;
7309 queue_Prepend(stats, rpc_stat);
7310 if (addToPeerList) {
7311 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7316 * Increment the stats for this function
7319 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7320 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7321 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7322 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7323 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7324 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7325 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7327 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7328 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7330 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7331 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7333 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7334 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7336 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7337 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7345 * rx_IncrementTimeAndCount - increment the times and count for a particular
7350 * IN peer - the peer who invoked the rpc
7352 * IN rxInterface - a unique number that identifies the rpc interface
7354 * IN currentFunc - the index of the function being invoked
7356 * IN totalFunc - the total number of functions in this interface
7358 * IN queueTime - the amount of time this function waited for a thread
7360 * IN execTime - the amount of time this function invocation took to execute
7362 * IN bytesSent - the number bytes sent by this invocation
7364 * IN bytesRcvd - the number bytes received by this invocation
7366 * IN isServer - if true, this invocation was made to a server
7374 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7375 afs_uint32 currentFunc, afs_uint32 totalFunc,
7376 struct clock *queueTime, struct clock *execTime,
7377 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7381 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7384 MUTEX_ENTER(&rx_rpc_stats);
7385 MUTEX_ENTER(&peer->peer_lock);
7387 if (rxi_monitor_peerStats) {
7388 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7389 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7390 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7393 if (rxi_monitor_processStats) {
7394 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7395 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7396 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7399 MUTEX_EXIT(&peer->peer_lock);
7400 MUTEX_EXIT(&rx_rpc_stats);
7405 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7409 * IN callerVersion - the rpc stat version of the caller.
7411 * IN count - the number of entries to marshall.
7413 * IN stats - pointer to stats to be marshalled.
7415 * OUT ptr - Where to store the marshalled data.
7422 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7423 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7429 * We only support the first version
7431 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7432 *(ptr++) = stats->remote_peer;
7433 *(ptr++) = stats->remote_port;
7434 *(ptr++) = stats->remote_is_server;
7435 *(ptr++) = stats->interfaceId;
7436 *(ptr++) = stats->func_total;
7437 *(ptr++) = stats->func_index;
7438 *(ptr++) = hgethi(stats->invocations);
7439 *(ptr++) = hgetlo(stats->invocations);
7440 *(ptr++) = hgethi(stats->bytes_sent);
7441 *(ptr++) = hgetlo(stats->bytes_sent);
7442 *(ptr++) = hgethi(stats->bytes_rcvd);
7443 *(ptr++) = hgetlo(stats->bytes_rcvd);
7444 *(ptr++) = stats->queue_time_sum.sec;
7445 *(ptr++) = stats->queue_time_sum.usec;
7446 *(ptr++) = stats->queue_time_sum_sqr.sec;
7447 *(ptr++) = stats->queue_time_sum_sqr.usec;
7448 *(ptr++) = stats->queue_time_min.sec;
7449 *(ptr++) = stats->queue_time_min.usec;
7450 *(ptr++) = stats->queue_time_max.sec;
7451 *(ptr++) = stats->queue_time_max.usec;
7452 *(ptr++) = stats->execution_time_sum.sec;
7453 *(ptr++) = stats->execution_time_sum.usec;
7454 *(ptr++) = stats->execution_time_sum_sqr.sec;
7455 *(ptr++) = stats->execution_time_sum_sqr.usec;
7456 *(ptr++) = stats->execution_time_min.sec;
7457 *(ptr++) = stats->execution_time_min.usec;
7458 *(ptr++) = stats->execution_time_max.sec;
7459 *(ptr++) = stats->execution_time_max.usec;
7465 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7470 * IN callerVersion - the rpc stat version of the caller
7472 * OUT myVersion - the rpc stat version of this function
7474 * OUT clock_sec - local time seconds
7476 * OUT clock_usec - local time microseconds
7478 * OUT allocSize - the number of bytes allocated to contain stats
7480 * OUT statCount - the number stats retrieved from this process.
7482 * OUT stats - the actual stats retrieved from this process.
7486 * Returns void. If successful, stats will != NULL.
7490 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7491 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7492 size_t * allocSize, afs_uint32 * statCount,
7493 afs_uint32 ** stats)
7503 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7506 * Check to see if stats are enabled
7509 MUTEX_ENTER(&rx_rpc_stats);
7510 if (!rxi_monitor_processStats) {
7511 MUTEX_EXIT(&rx_rpc_stats);
7515 clock_GetTime(&now);
7516 *clock_sec = now.sec;
7517 *clock_usec = now.usec;
7520 * Allocate the space based upon the caller version
7522 * If the client is at an older version than we are,
7523 * we return the statistic data in the older data format, but
7524 * we still return our version number so the client knows we
7525 * are maintaining more data than it can retrieve.
7528 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7529 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7530 *statCount = rxi_rpc_process_stat_cnt;
7533 * This can't happen yet, but in the future version changes
7534 * can be handled by adding additional code here
7538 if (space > (size_t) 0) {
7540 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7543 rx_interface_stat_p rpc_stat, nrpc_stat;
7547 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7549 * Copy the data based upon the caller version
7551 rx_MarshallProcessRPCStats(callerVersion,
7552 rpc_stat->stats[0].func_total,
7553 rpc_stat->stats, &ptr);
7559 MUTEX_EXIT(&rx_rpc_stats);
7564 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7568 * IN callerVersion - the rpc stat version of the caller
7570 * OUT myVersion - the rpc stat version of this function
7572 * OUT clock_sec - local time seconds
7574 * OUT clock_usec - local time microseconds
7576 * OUT allocSize - the number of bytes allocated to contain stats
7578 * OUT statCount - the number of stats retrieved from the individual
7581 * OUT stats - the actual stats retrieved from the individual peer structures.
7585 * Returns void. If successful, stats will != NULL.
7589 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7590 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7591 size_t * allocSize, afs_uint32 * statCount,
7592 afs_uint32 ** stats)
7602 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7605 * Check to see if stats are enabled
7608 MUTEX_ENTER(&rx_rpc_stats);
7609 if (!rxi_monitor_peerStats) {
7610 MUTEX_EXIT(&rx_rpc_stats);
7614 clock_GetTime(&now);
7615 *clock_sec = now.sec;
7616 *clock_usec = now.usec;
7619 * Allocate the space based upon the caller version
7621 * If the client is at an older version than we are,
7622 * we return the statistic data in the older data format, but
7623 * we still return our version number so the client knows we
7624 * are maintaining more data than it can retrieve.
7627 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7628 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7629 *statCount = rxi_rpc_peer_stat_cnt;
7632 * This can't happen yet, but in the future version changes
7633 * can be handled by adding additional code here
7637 if (space > (size_t) 0) {
7639 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7642 rx_interface_stat_p rpc_stat, nrpc_stat;
7646 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7648 * We have to fix the offset of rpc_stat since we are
7649 * keeping this structure on two rx_queues. The rx_queue
7650 * package assumes that the rx_queue member is the first
7651 * member of the structure. That is, rx_queue assumes that
7652 * any one item is only on one queue at a time. We are
7653 * breaking that assumption and so we have to do a little
7654 * math to fix our pointers.
7657 fix_offset = (char *)rpc_stat;
7658 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7659 rpc_stat = (rx_interface_stat_p) fix_offset;
7662 * Copy the data based upon the caller version
7664 rx_MarshallProcessRPCStats(callerVersion,
7665 rpc_stat->stats[0].func_total,
7666 rpc_stat->stats, &ptr);
7672 MUTEX_EXIT(&rx_rpc_stats);
7677 * rx_FreeRPCStats - free memory allocated by
7678 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7682 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7683 * rx_RetrievePeerRPCStats
7685 * IN allocSize - the number of bytes in stats.
7693 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7695 rxi_Free(stats, allocSize);
7699 * rx_queryProcessRPCStats - see if process rpc stat collection is
7700 * currently enabled.
7706 * Returns 0 if stats are not enabled != 0 otherwise
7710 rx_queryProcessRPCStats(void)
7713 MUTEX_ENTER(&rx_rpc_stats);
7714 rc = rxi_monitor_processStats;
7715 MUTEX_EXIT(&rx_rpc_stats);
7720 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7726 * Returns 0 if stats are not enabled != 0 otherwise
7730 rx_queryPeerRPCStats(void)
7733 MUTEX_ENTER(&rx_rpc_stats);
7734 rc = rxi_monitor_peerStats;
7735 MUTEX_EXIT(&rx_rpc_stats);
7740 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7750 rx_enableProcessRPCStats(void)
7752 MUTEX_ENTER(&rx_rpc_stats);
7753 rx_enable_stats = 1;
7754 rxi_monitor_processStats = 1;
7755 MUTEX_EXIT(&rx_rpc_stats);
7759 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7769 rx_enablePeerRPCStats(void)
7771 MUTEX_ENTER(&rx_rpc_stats);
7772 rx_enable_stats = 1;
7773 rxi_monitor_peerStats = 1;
7774 MUTEX_EXIT(&rx_rpc_stats);
7778 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7788 rx_disableProcessRPCStats(void)
7790 rx_interface_stat_p rpc_stat, nrpc_stat;
7793 MUTEX_ENTER(&rx_rpc_stats);
7796 * Turn off process statistics and if peer stats is also off, turn
7800 rxi_monitor_processStats = 0;
7801 if (rxi_monitor_peerStats == 0) {
7802 rx_enable_stats = 0;
7805 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7806 unsigned int num_funcs = 0;
7809 queue_Remove(rpc_stat);
7810 num_funcs = rpc_stat->stats[0].func_total;
7812 sizeof(rx_interface_stat_t) +
7813 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7815 rxi_Free(rpc_stat, space);
7816 rxi_rpc_process_stat_cnt -= num_funcs;
7818 MUTEX_EXIT(&rx_rpc_stats);
7822 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7832 rx_disablePeerRPCStats(void)
7834 struct rx_peer **peer_ptr, **peer_end;
7837 MUTEX_ENTER(&rx_rpc_stats);
7840 * Turn off peer statistics and if process stats is also off, turn
7844 rxi_monitor_peerStats = 0;
7845 if (rxi_monitor_processStats == 0) {
7846 rx_enable_stats = 0;
7849 MUTEX_ENTER(&rx_peerHashTable_lock);
7850 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7851 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7853 struct rx_peer *peer, *next, *prev;
7854 for (prev = peer = *peer_ptr; peer; peer = next) {
7856 code = MUTEX_TRYENTER(&peer->peer_lock);
7858 rx_interface_stat_p rpc_stat, nrpc_stat;
7861 (&peer->rpcStats, rpc_stat, nrpc_stat,
7862 rx_interface_stat)) {
7863 unsigned int num_funcs = 0;
7866 queue_Remove(&rpc_stat->queue_header);
7867 queue_Remove(&rpc_stat->all_peers);
7868 num_funcs = rpc_stat->stats[0].func_total;
7870 sizeof(rx_interface_stat_t) +
7871 rpc_stat->stats[0].func_total *
7872 sizeof(rx_function_entry_v1_t);
7874 rxi_Free(rpc_stat, space);
7875 rxi_rpc_peer_stat_cnt -= num_funcs;
7877 MUTEX_EXIT(&peer->peer_lock);
7878 if (prev == *peer_ptr) {
7888 MUTEX_EXIT(&rx_peerHashTable_lock);
7889 MUTEX_EXIT(&rx_rpc_stats);
7893 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7898 * IN clearFlag - flag indicating which stats to clear
7906 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7908 rx_interface_stat_p rpc_stat, nrpc_stat;
7910 MUTEX_ENTER(&rx_rpc_stats);
7912 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7913 unsigned int num_funcs = 0, i;
7914 num_funcs = rpc_stat->stats[0].func_total;
7915 for (i = 0; i < num_funcs; i++) {
7916 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7917 hzero(rpc_stat->stats[i].invocations);
7919 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7920 hzero(rpc_stat->stats[i].bytes_sent);
7922 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7923 hzero(rpc_stat->stats[i].bytes_rcvd);
7925 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7926 rpc_stat->stats[i].queue_time_sum.sec = 0;
7927 rpc_stat->stats[i].queue_time_sum.usec = 0;
7929 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7930 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7931 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7933 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7934 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7935 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7937 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7938 rpc_stat->stats[i].queue_time_max.sec = 0;
7939 rpc_stat->stats[i].queue_time_max.usec = 0;
7941 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7942 rpc_stat->stats[i].execution_time_sum.sec = 0;
7943 rpc_stat->stats[i].execution_time_sum.usec = 0;
7945 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7946 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7947 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7949 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7950 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7951 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7953 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7954 rpc_stat->stats[i].execution_time_max.sec = 0;
7955 rpc_stat->stats[i].execution_time_max.usec = 0;
7960 MUTEX_EXIT(&rx_rpc_stats);
7964 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7969 * IN clearFlag - flag indicating which stats to clear
7977 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7979 rx_interface_stat_p rpc_stat, nrpc_stat;
7981 MUTEX_ENTER(&rx_rpc_stats);
7983 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7984 unsigned int num_funcs = 0, i;
7987 * We have to fix the offset of rpc_stat since we are
7988 * keeping this structure on two rx_queues. The rx_queue
7989 * package assumes that the rx_queue member is the first
7990 * member of the structure. That is, rx_queue assumes that
7991 * any one item is only on one queue at a time. We are
7992 * breaking that assumption and so we have to do a little
7993 * math to fix our pointers.
7996 fix_offset = (char *)rpc_stat;
7997 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7998 rpc_stat = (rx_interface_stat_p) fix_offset;
8000 num_funcs = rpc_stat->stats[0].func_total;
8001 for (i = 0; i < num_funcs; i++) {
8002 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8003 hzero(rpc_stat->stats[i].invocations);
8005 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8006 hzero(rpc_stat->stats[i].bytes_sent);
8008 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8009 hzero(rpc_stat->stats[i].bytes_rcvd);
8011 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8012 rpc_stat->stats[i].queue_time_sum.sec = 0;
8013 rpc_stat->stats[i].queue_time_sum.usec = 0;
8015 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8016 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8017 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8019 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8020 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8021 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8023 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8024 rpc_stat->stats[i].queue_time_max.sec = 0;
8025 rpc_stat->stats[i].queue_time_max.usec = 0;
8027 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8028 rpc_stat->stats[i].execution_time_sum.sec = 0;
8029 rpc_stat->stats[i].execution_time_sum.usec = 0;
8031 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8032 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8033 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8035 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8036 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8037 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8039 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8040 rpc_stat->stats[i].execution_time_max.sec = 0;
8041 rpc_stat->stats[i].execution_time_max.usec = 0;
8046 MUTEX_EXIT(&rx_rpc_stats);
8050 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8051 * is authorized to enable/disable/clear RX statistics.
8053 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8056 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8058 rxi_rxstat_userok = proc;
8062 rx_RxStatUserOk(struct rx_call *call)
8064 if (!rxi_rxstat_userok)
8066 return rxi_rxstat_userok(call);
8071 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8072 * function in the MSVC runtime DLL (msvcrt.dll).
8074 * Note: the system serializes calls to this function.
8077 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8078 DWORD reason, /* reason function is being called */
8079 LPVOID reserved) /* reserved for future use */
8082 case DLL_PROCESS_ATTACH:
8083 /* library is being attached to a process */
8087 case DLL_PROCESS_DETACH:
8096 int rx_DumpCalls(FILE *outputFile, char *cookie)
8098 #ifdef RXDEBUG_PACKET
8100 #ifdef KDUMP_RX_LOCK
8101 struct rx_call_rx_lock *c;
8107 sprintf(output, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8108 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8110 for (c = rx_allCallsp; c; c = c->allNextp) {
8111 u_short rqc, tqc, iovqc;
8112 struct rx_packet *p, *np;
8114 MUTEX_ENTER(&c->lock);
8115 queue_Count(&c->rq, p, np, rx_packet, rqc);
8116 queue_Count(&c->tq, p, np, rx_packet, tqc);
8117 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8119 sprintf(output, "%s - call=0x%p, id=%u, state=%u, mode=%u, conn=%p, epoch=%u, cid=%u, callNum=%u, connFlags=0x%x, flags=0x%x, "
8120 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8121 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8122 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8123 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8124 #ifdef RX_ENABLE_LOCKS
8127 #ifdef RX_REFCOUNT_CHECK
8128 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
8129 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
8132 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,
8133 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
8134 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
8135 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
8136 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
8137 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
8138 #ifdef RX_ENABLE_LOCKS
8139 , (afs_uint32)c->refCount
8141 #ifdef RX_REFCOUNT_CHECK
8142 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
8145 MUTEX_EXIT(&c->lock);
8147 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8149 sprintf(output, "%s - End dumping all Rx Calls\r\n", cookie);
8150 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8151 #endif /* RXDEBUG_PACKET */
8154 #endif /* AFS_NT40_ENV */