2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
13 #include <afs/param.h>
16 # include "afs/sysincludes.h"
17 # include "afsincludes.h"
22 # ifdef AFS_LINUX20_ENV
23 # include "h/socket.h"
25 # include "netinet/in.h"
27 # include "netinet/ip6.h"
28 # include "inet/common.h"
30 # include "inet/ip_ire.h"
32 # include "afs/afs_args.h"
33 # include "afs/afs_osi.h"
34 # ifdef RX_KERNEL_TRACE
35 # include "rx_kcommon.h"
37 # if defined(AFS_AIX_ENV)
41 # undef RXDEBUG /* turn off debugging */
43 # if defined(AFS_SGI_ENV)
44 # include "sys/debug.h"
47 # include "afs/sysincludes.h"
48 # include "afsincludes.h"
49 # endif /* !UKERNEL */
50 # include "afs/lock.h"
51 # include "rx_kmutex.h"
52 # include "rx_kernel.h"
53 # define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
54 # define AFSOP_STOP_AFS 211 /* Stop AFS process */
55 # define AFSOP_STOP_BKG 212 /* Stop BKG process */
56 extern afs_int32 afs_termState;
58 # include "sys/lockl.h"
59 # include "sys/lock_def.h"
60 # endif /* AFS_AIX41_ENV */
61 # include "afs/rxgen_consts.h"
66 # include <afs/afsutil.h>
67 # include <WINNT\afsreg.h>
75 #include <opr/queue.h>
76 #include <hcrypto/rand.h>
80 #include "rx_atomic.h"
81 #include "rx_globals.h"
83 #include "rx_internal.h"
90 #include "rx_packet.h"
91 #include "rx_server.h"
93 #include <afs/rxgen_consts.h>
96 #ifdef AFS_PTHREAD_ENV
98 int (*registerProgram) (pid_t, char *) = 0;
99 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
102 int (*registerProgram) (PROCESS, char *) = 0;
103 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
107 /* Local static routines */
108 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
109 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
110 struct rx_call *, struct rx_peer *,
112 static void rxi_Resend(struct rxevent *event, void *arg0, void *arg1,
114 static void rxi_SendDelayedAck(struct rxevent *event, void *call,
115 void *dummy, int dummy2);
116 static void rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1,
117 void *dummy, int dummy2);
118 static void rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1,
119 void *unused, int unused2);
120 static void rxi_ReapConnections(struct rxevent *unused, void *unused1,
121 void *unused2, int unused3);
122 static struct rx_packet *rxi_SendCallAbort(struct rx_call *call,
123 struct rx_packet *packet,
124 int istack, int force);
125 static void rxi_AckAll(struct rx_call *call);
126 static struct rx_connection
127 *rxi_FindConnection(osi_socket socket, afs_uint32 host, u_short port,
128 u_short serviceId, afs_uint32 cid,
129 afs_uint32 epoch, int type, u_int securityIndex,
130 int *unknownService);
131 static struct rx_packet
132 *rxi_ReceiveDataPacket(struct rx_call *call, struct rx_packet *np,
133 int istack, osi_socket socket,
134 afs_uint32 host, u_short port, int *tnop,
135 struct rx_call **newcallp);
136 static struct rx_packet
137 *rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
139 static struct rx_packet
140 *rxi_ReceiveResponsePacket(struct rx_connection *conn,
141 struct rx_packet *np, int istack);
142 static struct rx_packet
143 *rxi_ReceiveChallengePacket(struct rx_connection *conn,
144 struct rx_packet *np, int istack);
145 static void rxi_AttachServerProc(struct rx_call *call, osi_socket socket,
146 int *tnop, struct rx_call **newcallp);
147 static void rxi_ClearTransmitQueue(struct rx_call *call, int force);
148 static void rxi_ClearReceiveQueue(struct rx_call *call);
149 static void rxi_ResetCall(struct rx_call *call, int newcall);
150 static void rxi_ScheduleKeepAliveEvent(struct rx_call *call);
151 static void rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn);
152 static void rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs);
153 static void rxi_KeepAliveOn(struct rx_call *call);
154 static void rxi_GrowMTUOn(struct rx_call *call);
155 static void rxi_ChallengeOn(struct rx_connection *conn);
156 static int rxi_CheckCall(struct rx_call *call, int haveCTLock);
157 static void rxi_AckAllInTransmitQueue(struct rx_call *call);
158 static void rxi_CancelKeepAliveEvent(struct rx_call *call);
159 static void rxi_CancelDelayedAbortEvent(struct rx_call *call);
160 static void rxi_CancelGrowMTUEvent(struct rx_call *call);
161 static void update_nextCid(void);
163 #ifdef RX_ENABLE_LOCKS
165 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
166 rx_atomic_t rxi_start_in_error;
168 #endif /* RX_ENABLE_LOCKS */
170 /* Constant delay time before sending an acknowledge of the last packet
171 * received. This is to avoid sending an extra acknowledge when the
172 * client is about to make another call, anyway, or the server is
175 * The lastAckDelay may not exceeed 400ms without causing peers to
176 * unecessarily timeout.
178 struct clock rx_lastAckDelay = {0, 400000};
180 /* Constant delay time before sending a soft ack when none was requested.
181 * This is to make sure we send soft acks before the sender times out,
182 * Normally we wait and send a hard ack when the receiver consumes the packet
184 * This value has been 100ms in all shipping versions of OpenAFS. Changing it
185 * will require changes to the peer's RTT calculations.
187 struct clock rx_softAckDelay = {0, 100000};
190 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
191 * currently allocated within rx. This number is used to allocate the
192 * memory required to return the statistics when queried.
193 * Protected by the rx_rpc_stats mutex.
196 static unsigned int rxi_rpc_peer_stat_cnt;
199 * rxi_rpc_process_stat_cnt counts the total number of local process stat
200 * structures currently allocated within rx. The number is used to allocate
201 * the memory required to return the statistics when queried.
202 * Protected by the rx_rpc_stats mutex.
205 static unsigned int rxi_rpc_process_stat_cnt;
207 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
208 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
210 /* Incoming calls wait on this queue when there are no available
211 * server processes */
212 struct opr_queue rx_incomingCallQueue;
214 /* Server processes wait on this queue when there are no appropriate
215 * calls to process */
216 struct opr_queue rx_idleServerQueue;
218 #if !defined(offsetof)
219 #include <stddef.h> /* for definition of offsetof() */
222 #ifdef RX_ENABLE_LOCKS
223 afs_kmutex_t rx_atomic_mutex;
226 /* Forward prototypes */
227 static struct rx_call * rxi_NewCall(struct rx_connection *, int);
230 putConnection (struct rx_connection *conn) {
231 MUTEX_ENTER(&rx_refcnt_mutex);
233 MUTEX_EXIT(&rx_refcnt_mutex);
236 #ifdef AFS_PTHREAD_ENV
239 * Use procedural initialization of mutexes/condition variables
243 extern afs_kmutex_t rx_quota_mutex;
244 extern afs_kmutex_t rx_pthread_mutex;
245 extern afs_kmutex_t rx_packets_mutex;
246 extern afs_kmutex_t rx_refcnt_mutex;
247 extern afs_kmutex_t des_init_mutex;
248 extern afs_kmutex_t des_random_mutex;
250 extern afs_kmutex_t rx_clock_mutex;
251 extern afs_kmutex_t rxi_connCacheMutex;
252 extern afs_kmutex_t event_handler_mutex;
253 extern afs_kmutex_t listener_mutex;
254 extern afs_kmutex_t rx_if_init_mutex;
255 extern afs_kmutex_t rx_if_mutex;
257 extern afs_kcondvar_t rx_event_handler_cond;
258 extern afs_kcondvar_t rx_listener_cond;
261 static afs_kmutex_t epoch_mutex;
262 static afs_kmutex_t rx_init_mutex;
263 static afs_kmutex_t rx_debug_mutex;
264 static afs_kmutex_t rx_rpc_stats;
267 rxi_InitPthread(void)
269 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
270 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
271 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
272 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
274 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
275 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
276 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
277 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
278 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
279 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
281 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
282 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
283 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
284 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
285 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
288 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
289 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
292 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
293 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
295 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
296 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
297 MUTEX_INIT(&rx_mallocedPktQ_lock, "rx_mallocedPktQ_lock", MUTEX_DEFAULT,
300 #ifdef RX_ENABLE_LOCKS
303 #endif /* RX_LOCKS_DB */
304 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
305 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
307 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
309 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
311 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
313 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
315 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
317 #endif /* RX_ENABLE_LOCKS */
320 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
321 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
323 * The rx_stats_mutex mutex protects the following global variables:
324 * rxi_lowConnRefCount
325 * rxi_lowPeerRefCount
334 * The rx_quota_mutex mutex protects the following global variables:
342 * The rx_freePktQ_lock protects the following global variables:
347 * The rx_packets_mutex mutex protects the following global variables:
355 * The rx_pthread_mutex mutex protects the following global variables:
356 * rxi_fcfs_thread_num
359 #define INIT_PTHREAD_LOCKS
363 /* Variables for handling the minProcs implementation. availProcs gives the
364 * number of threads available in the pool at this moment (not counting dudes
365 * executing right now). totalMin gives the total number of procs required
366 * for handling all minProcs requests. minDeficit is a dynamic variable
367 * tracking the # of procs required to satisfy all of the remaining minProcs
369 * For fine grain locking to work, the quota check and the reservation of
370 * a server thread has to come while rxi_availProcs and rxi_minDeficit
371 * are locked. To this end, the code has been modified under #ifdef
372 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
373 * same time. A new function, ReturnToServerPool() returns the allocation.
375 * A call can be on several queue's (but only one at a time). When
376 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
377 * that no one else is touching the queue. To this end, we store the address
378 * of the queue lock in the call structure (under the call lock) when we
379 * put the call on a queue, and we clear the call_queue_lock when the
380 * call is removed from a queue (once the call lock has been obtained).
381 * This allows rxi_ResetCall to safely synchronize with others wishing
382 * to manipulate the queue.
385 #if defined(RX_ENABLE_LOCKS)
386 static afs_kmutex_t rx_rpc_stats;
389 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
390 ** pretty good that the next packet coming in is from the same connection
391 ** as the last packet, since we're send multiple packets in a transmit window.
393 struct rx_connection *rxLastConn = 0;
395 #ifdef RX_ENABLE_LOCKS
396 /* The locking hierarchy for rx fine grain locking is composed of these
399 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
400 * also protects updates to rx_nextCid
401 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
402 * call->lock - locks call data fields.
403 * These are independent of each other:
404 * rx_freeCallQueue_lock
409 * serverQueueEntry->lock
410 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
412 * peer->lock - locks peer data fields.
413 * conn_data_lock - that more than one thread is not updating a conn data
414 * field at the same time.
425 * Do we need a lock to protect the peer field in the conn structure?
426 * conn->peer was previously a constant for all intents and so has no
427 * lock protecting this field. The multihomed client delta introduced
428 * a RX code change : change the peer field in the connection structure
429 * to that remote interface from which the last packet for this
430 * connection was sent out. This may become an issue if further changes
433 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
434 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
436 /* rxdb_fileID is used to identify the lock location, along with line#. */
437 static int rxdb_fileID = RXDB_FILE_RX;
438 #endif /* RX_LOCKS_DB */
439 #else /* RX_ENABLE_LOCKS */
440 #define SET_CALL_QUEUE_LOCK(C, L)
441 #define CLEAR_CALL_QUEUE_LOCK(C)
442 #endif /* RX_ENABLE_LOCKS */
443 struct rx_serverQueueEntry *rx_waitForPacket = 0;
445 /* ------------Exported Interfaces------------- */
447 /* Initialize rx. A port number may be mentioned, in which case this
448 * becomes the default port number for any service installed later.
449 * If 0 is provided for the port number, a random port will be chosen
450 * by the kernel. Whether this will ever overlap anything in
451 * /etc/services is anybody's guess... Returns 0 on success, -1 on
453 #if !(defined(AFS_NT40_ENV) || defined(RXK_UPCALL_ENV))
456 rx_atomic_t rxinit_status = RX_ATOMIC_INIT(1);
459 rx_InitHost(u_int host, u_int port)
466 char *htable, *ptable;
471 if (!rx_atomic_test_and_clear_bit(&rxinit_status, 0))
472 return 0; /* already started */
478 if (afs_winsockInit() < 0)
484 * Initialize anything necessary to provide a non-premptive threading
487 rxi_InitializeThreadSupport();
490 /* Allocate and initialize a socket for client and perhaps server
493 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
494 if (rx_socket == OSI_NULLSOCKET) {
497 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
500 #endif /* RX_LOCKS_DB */
501 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
502 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
503 MUTEX_INIT(&rx_atomic_mutex, "rx_atomic_mutex", MUTEX_DEFAULT, 0);
504 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
505 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
506 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
507 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
508 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
509 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
510 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
512 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
514 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
516 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
518 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
519 MUTEX_INIT(&rx_mallocedPktQ_lock, "rx_mallocedPktQ_lock", MUTEX_DEFAULT,
522 #if defined(AFS_HPUX110_ENV)
524 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
525 #endif /* AFS_HPUX110_ENV */
526 #endif /* RX_ENABLE_LOCKS && KERNEL */
529 rx_connDeadTime = 12;
530 rx_tranquil = 0; /* reset flag */
531 rxi_ResetStatistics();
532 htable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
533 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
534 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
535 ptable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
536 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
537 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
539 /* Malloc up a bunch of packets & buffers */
541 opr_queue_Init(&rx_freePacketQueue);
542 rxi_NeedMorePackets = FALSE;
543 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
544 opr_queue_Init(&rx_mallocedPacketQueue);
546 /* enforce a minimum number of allocated packets */
547 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
548 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
550 /* allocate the initial free packet pool */
551 #ifdef RX_ENABLE_TSFPQ
552 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
553 #else /* RX_ENABLE_TSFPQ */
554 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
555 #endif /* RX_ENABLE_TSFPQ */
562 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
563 tv.tv_sec = clock_now.sec;
564 tv.tv_usec = clock_now.usec;
565 srand((unsigned int)tv.tv_usec);
572 #if defined(KERNEL) && !defined(UKERNEL)
573 /* Really, this should never happen in a real kernel */
576 struct sockaddr_in addr;
578 int addrlen = sizeof(addr);
580 socklen_t addrlen = sizeof(addr);
582 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
584 osi_Free(htable, rx_hashTableSize * sizeof(struct rx_connection *));
587 rx_port = addr.sin_port;
590 rx_stats.minRtt.sec = 9999999;
591 if (RAND_bytes(&rx_epoch, sizeof(rx_epoch)) != 1)
593 rx_epoch = (rx_epoch & ~0x40000000) | 0x80000000;
594 if (RAND_bytes(&rx_nextCid, sizeof(rx_nextCid)) != 1)
596 rx_nextCid &= RX_CIDMASK;
597 MUTEX_ENTER(&rx_quota_mutex);
598 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
599 MUTEX_EXIT(&rx_quota_mutex);
600 /* *Slightly* random start time for the cid. This is just to help
601 * out with the hashing function at the peer */
602 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
603 rx_connHashTable = (struct rx_connection **)htable;
604 rx_peerHashTable = (struct rx_peer **)ptable;
606 rx_hardAckDelay.sec = 0;
607 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
609 rxevent_Init(20, rxi_ReScheduleEvents);
611 /* Initialize various global queues */
612 opr_queue_Init(&rx_idleServerQueue);
613 opr_queue_Init(&rx_incomingCallQueue);
614 opr_queue_Init(&rx_freeCallQueue);
616 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
617 /* Initialize our list of usable IP addresses. */
621 /* Start listener process (exact function is dependent on the
622 * implementation environment--kernel or user space) */
626 rx_atomic_clear_bit(&rxinit_status, 0);
633 return rx_InitHost(htonl(INADDR_ANY), port);
639 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
640 * maintaing the round trip timer.
645 * Start a new RTT timer for a given call and packet.
647 * There must be no resendEvent already listed for this call, otherwise this
648 * will leak events - intended for internal use within the RTO code only
651 * the RX call to start the timer for
652 * @param[in] lastPacket
653 * a flag indicating whether the last packet has been sent or not
655 * @pre call must be locked before calling this function
659 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
661 struct clock now, retryTime;
666 clock_Add(&retryTime, &call->rto);
668 /* If we're sending the last packet, and we're the client, then the server
669 * may wait for an additional 400ms before returning the ACK, wait for it
670 * rather than hitting a timeout */
671 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
672 clock_Addmsec(&retryTime, 400);
674 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
675 call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
680 * Cancel an RTT timer for a given call.
684 * the RX call to cancel the timer for
686 * @pre call must be locked before calling this function
691 rxi_rto_cancel(struct rx_call *call)
693 if (call->resendEvent != NULL) {
694 rxevent_Cancel(&call->resendEvent);
695 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
700 * Tell the RTO timer that we have sent a packet.
702 * If the timer isn't already running, then start it. If the timer is running,
706 * the RX call that the packet has been sent on
707 * @param[in] lastPacket
708 * A flag which is true if this is the last packet for the call
710 * @pre The call must be locked before calling this function
715 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
717 if (call->resendEvent)
720 rxi_rto_startTimer(call, lastPacket, istack);
724 * Tell the RTO timer that we have received an new ACK message
726 * This function should be called whenever a call receives an ACK that
727 * acknowledges new packets. Whatever happens, we stop the current timer.
728 * If there are unacked packets in the queue which have been sent, then
729 * we restart the timer from now. Otherwise, we leave it stopped.
732 * the RX call that the ACK has been received on
736 rxi_rto_packet_acked(struct rx_call *call, int istack)
738 struct opr_queue *cursor;
740 rxi_rto_cancel(call);
742 if (opr_queue_IsEmpty(&call->tq))
745 for (opr_queue_Scan(&call->tq, cursor)) {
746 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
747 if (p->header.seq > call->tfirst + call->twind)
750 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
751 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
759 * Set an initial round trip timeout for a peer connection
761 * @param[in] secs The timeout to set in seconds
765 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
766 peer->rtt = secs * 8000;
770 * Set a delayed ack event on the specified call for the given time
772 * @param[in] call - the call on which to set the event
773 * @param[in] offset - the delay from now after which the event fires
776 rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
778 struct clock now, when;
782 clock_Add(&when, offset);
784 if (call->delayedAckEvent && clock_Gt(&call->delayedAckTime, &when)) {
785 /* The event we're cancelling already has a reference, so we don't
787 rxevent_Cancel(&call->delayedAckEvent);
788 call->delayedAckEvent = rxevent_Post(&when, &now, rxi_SendDelayedAck,
791 call->delayedAckTime = when;
792 } else if (!call->delayedAckEvent) {
793 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
794 call->delayedAckEvent = rxevent_Post(&when, &now,
797 call->delayedAckTime = when;
802 rxi_CancelDelayedAckEvent(struct rx_call *call)
804 if (call->delayedAckEvent) {
805 rxevent_Cancel(&call->delayedAckEvent);
806 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
810 /* called with unincremented nRequestsRunning to see if it is OK to start
811 * a new thread in this service. Could be "no" for two reasons: over the
812 * max quota, or would prevent others from reaching their min quota.
814 #ifdef RX_ENABLE_LOCKS
815 /* This verion of QuotaOK reserves quota if it's ok while the
816 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
819 QuotaOK(struct rx_service *aservice)
821 /* check if over max quota */
822 if (aservice->nRequestsRunning >= aservice->maxProcs) {
826 /* under min quota, we're OK */
827 /* otherwise, can use only if there are enough to allow everyone
828 * to go to their min quota after this guy starts.
831 MUTEX_ENTER(&rx_quota_mutex);
832 if ((aservice->nRequestsRunning < aservice->minProcs)
833 || (rxi_availProcs > rxi_minDeficit)) {
834 aservice->nRequestsRunning++;
835 /* just started call in minProcs pool, need fewer to maintain
837 if (aservice->nRequestsRunning <= aservice->minProcs)
840 MUTEX_EXIT(&rx_quota_mutex);
843 MUTEX_EXIT(&rx_quota_mutex);
849 ReturnToServerPool(struct rx_service *aservice)
851 aservice->nRequestsRunning--;
852 MUTEX_ENTER(&rx_quota_mutex);
853 if (aservice->nRequestsRunning < aservice->minProcs)
856 MUTEX_EXIT(&rx_quota_mutex);
859 #else /* RX_ENABLE_LOCKS */
861 QuotaOK(struct rx_service *aservice)
864 /* under min quota, we're OK */
865 if (aservice->nRequestsRunning < aservice->minProcs)
868 /* check if over max quota */
869 if (aservice->nRequestsRunning >= aservice->maxProcs)
872 /* otherwise, can use only if there are enough to allow everyone
873 * to go to their min quota after this guy starts.
875 MUTEX_ENTER(&rx_quota_mutex);
876 if (rxi_availProcs > rxi_minDeficit)
878 MUTEX_EXIT(&rx_quota_mutex);
881 #endif /* RX_ENABLE_LOCKS */
884 /* Called by rx_StartServer to start up lwp's to service calls.
885 NExistingProcs gives the number of procs already existing, and which
886 therefore needn't be created. */
888 rxi_StartServerProcs(int nExistingProcs)
890 struct rx_service *service;
895 /* For each service, reserve N processes, where N is the "minimum"
896 * number of processes that MUST be able to execute a request in parallel,
897 * at any time, for that process. Also compute the maximum difference
898 * between any service's maximum number of processes that can run
899 * (i.e. the maximum number that ever will be run, and a guarantee
900 * that this number will run if other services aren't running), and its
901 * minimum number. The result is the extra number of processes that
902 * we need in order to provide the latter guarantee */
903 for (i = 0; i < RX_MAX_SERVICES; i++) {
905 service = rx_services[i];
906 if (service == (struct rx_service *)0)
908 nProcs += service->minProcs;
909 diff = service->maxProcs - service->minProcs;
913 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
914 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
915 for (i = 0; i < nProcs; i++) {
916 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
922 /* This routine is only required on Windows */
924 rx_StartClientThread(void)
926 #ifdef AFS_PTHREAD_ENV
928 pid = pthread_self();
929 #endif /* AFS_PTHREAD_ENV */
931 #endif /* AFS_NT40_ENV */
933 /* This routine must be called if any services are exported. If the
934 * donateMe flag is set, the calling process is donated to the server
937 rx_StartServer(int donateMe)
939 struct rx_service *service;
945 /* Start server processes, if necessary (exact function is dependent
946 * on the implementation environment--kernel or user space). DonateMe
947 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
948 * case, one less new proc will be created rx_StartServerProcs.
950 rxi_StartServerProcs(donateMe);
952 /* count up the # of threads in minProcs, and add set the min deficit to
953 * be that value, too.
955 for (i = 0; i < RX_MAX_SERVICES; i++) {
956 service = rx_services[i];
957 if (service == (struct rx_service *)0)
959 MUTEX_ENTER(&rx_quota_mutex);
960 rxi_totalMin += service->minProcs;
961 /* below works even if a thread is running, since minDeficit would
962 * still have been decremented and later re-incremented.
964 rxi_minDeficit += service->minProcs;
965 MUTEX_EXIT(&rx_quota_mutex);
968 /* Turn on reaping of idle server connections */
969 rxi_ReapConnections(NULL, NULL, NULL, 0);
978 #ifdef AFS_PTHREAD_ENV
980 pid = afs_pointer_to_int(pthread_self());
981 #else /* AFS_PTHREAD_ENV */
983 LWP_CurrentProcess(&pid);
984 #endif /* AFS_PTHREAD_ENV */
986 sprintf(name, "srv_%d", ++nProcs);
988 (*registerProgram) (pid, name);
990 #endif /* AFS_NT40_ENV */
991 rx_ServerProc(NULL); /* Never returns */
993 #ifdef RX_ENABLE_TSFPQ
994 /* no use leaving packets around in this thread's local queue if
995 * it isn't getting donated to the server thread pool.
997 rxi_FlushLocalPacketsTSFPQ();
998 #endif /* RX_ENABLE_TSFPQ */
1002 /* Create a new client connection to the specified service, using the
1003 * specified security object to implement the security model for this
1005 struct rx_connection *
1006 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
1007 struct rx_securityClass *securityObject,
1008 int serviceSecurityIndex)
1011 struct rx_connection *conn;
1016 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
1017 "serviceSecurityIndex %d)\n",
1018 ntohl(shost), ntohs(sport), sservice, securityObject,
1019 serviceSecurityIndex));
1021 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
1022 * the case of kmem_alloc? */
1023 conn = rxi_AllocConnection();
1024 #ifdef RX_ENABLE_LOCKS
1025 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
1026 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
1027 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
1030 MUTEX_ENTER(&rx_connHashTable_lock);
1031 conn->type = RX_CLIENT_CONNECTION;
1032 conn->epoch = rx_epoch;
1033 conn->cid = rx_nextCid;
1035 conn->peer = rxi_FindPeer(shost, sport, 1);
1036 conn->serviceId = sservice;
1037 conn->securityObject = securityObject;
1038 conn->securityData = (void *) 0;
1039 conn->securityIndex = serviceSecurityIndex;
1040 rx_SetConnDeadTime(conn, rx_connDeadTime);
1041 rx_SetConnSecondsUntilNatPing(conn, 0);
1042 conn->ackRate = RX_FAST_ACK_RATE;
1043 conn->nSpecific = 0;
1044 conn->specific = NULL;
1045 conn->challengeEvent = NULL;
1046 conn->delayedAbortEvent = NULL;
1047 conn->abortCount = 0;
1049 for (i = 0; i < RX_MAXCALLS; i++) {
1050 conn->twind[i] = rx_initSendWindow;
1051 conn->rwind[i] = rx_initReceiveWindow;
1052 conn->lastBusy[i] = 0;
1055 RXS_NewConnection(securityObject, conn);
1057 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1059 conn->refCount++; /* no lock required since only this thread knows... */
1060 conn->next = rx_connHashTable[hashindex];
1061 rx_connHashTable[hashindex] = conn;
1062 if (rx_stats_active)
1063 rx_atomic_inc(&rx_stats.nClientConns);
1064 MUTEX_EXIT(&rx_connHashTable_lock);
1070 * Ensure a connection's timeout values are valid.
1072 * @param[in] conn The connection to check
1074 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1075 * unless idleDeadTime and/or hardDeadTime are not set
1079 rxi_CheckConnTimeouts(struct rx_connection *conn)
1081 /* a connection's timeouts must have the relationship
1082 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1083 * total loss of network to a peer may cause an idle timeout instead of a
1084 * dead timeout, simply because the idle timeout gets hit first. Also set
1085 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1086 /* this logic is slightly complicated by the fact that
1087 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1089 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1090 if (conn->idleDeadTime) {
1091 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1093 if (conn->hardDeadTime) {
1094 if (conn->idleDeadTime) {
1095 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1097 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1103 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1105 /* The idea is to set the dead time to a value that allows several
1106 * keepalives to be dropped without timing out the connection. */
1107 conn->secondsUntilDead = seconds;
1108 rxi_CheckConnTimeouts(conn);
1109 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1113 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1115 conn->hardDeadTime = seconds;
1116 rxi_CheckConnTimeouts(conn);
1120 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1122 conn->idleDeadTime = seconds;
1123 rxi_CheckConnTimeouts(conn);
1126 int rxi_lowPeerRefCount = 0;
1127 int rxi_lowConnRefCount = 0;
1130 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1131 * NOTE: must not be called with rx_connHashTable_lock held.
1134 rxi_CleanupConnection(struct rx_connection *conn)
1136 /* Notify the service exporter, if requested, that this connection
1137 * is being destroyed */
1138 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1139 (*conn->service->destroyConnProc) (conn);
1141 /* Notify the security module that this connection is being destroyed */
1142 RXS_DestroyConnection(conn->securityObject, conn);
1144 /* If this is the last connection using the rx_peer struct, set its
1145 * idle time to now. rxi_ReapConnections will reap it if it's still
1146 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1148 MUTEX_ENTER(&rx_peerHashTable_lock);
1149 if (conn->peer->refCount < 2) {
1150 conn->peer->idleWhen = clock_Sec();
1151 if (conn->peer->refCount < 1) {
1152 conn->peer->refCount = 1;
1153 if (rx_stats_active) {
1154 MUTEX_ENTER(&rx_stats_mutex);
1155 rxi_lowPeerRefCount++;
1156 MUTEX_EXIT(&rx_stats_mutex);
1160 conn->peer->refCount--;
1161 MUTEX_EXIT(&rx_peerHashTable_lock);
1163 if (rx_stats_active)
1165 if (conn->type == RX_SERVER_CONNECTION)
1166 rx_atomic_dec(&rx_stats.nServerConns);
1168 rx_atomic_dec(&rx_stats.nClientConns);
1171 if (conn->specific) {
1173 for (i = 0; i < conn->nSpecific; i++) {
1174 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1175 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1176 conn->specific[i] = NULL;
1178 free(conn->specific);
1180 conn->specific = NULL;
1181 conn->nSpecific = 0;
1182 #endif /* !KERNEL */
1184 MUTEX_DESTROY(&conn->conn_call_lock);
1185 MUTEX_DESTROY(&conn->conn_data_lock);
1186 CV_DESTROY(&conn->conn_call_cv);
1188 rxi_FreeConnection(conn);
1191 /* Destroy the specified connection */
1193 rxi_DestroyConnection(struct rx_connection *conn)
1195 MUTEX_ENTER(&rx_connHashTable_lock);
1196 rxi_DestroyConnectionNoLock(conn);
1197 /* conn should be at the head of the cleanup list */
1198 if (conn == rx_connCleanup_list) {
1199 rx_connCleanup_list = rx_connCleanup_list->next;
1200 MUTEX_EXIT(&rx_connHashTable_lock);
1201 rxi_CleanupConnection(conn);
1203 #ifdef RX_ENABLE_LOCKS
1205 MUTEX_EXIT(&rx_connHashTable_lock);
1207 #endif /* RX_ENABLE_LOCKS */
1211 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1213 struct rx_connection **conn_ptr;
1215 struct rx_packet *packet;
1222 MUTEX_ENTER(&conn->conn_data_lock);
1223 MUTEX_ENTER(&rx_refcnt_mutex);
1224 if (conn->refCount > 0)
1227 if (rx_stats_active) {
1228 MUTEX_ENTER(&rx_stats_mutex);
1229 rxi_lowConnRefCount++;
1230 MUTEX_EXIT(&rx_stats_mutex);
1234 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1235 /* Busy; wait till the last guy before proceeding */
1236 MUTEX_EXIT(&rx_refcnt_mutex);
1237 MUTEX_EXIT(&conn->conn_data_lock);
1242 /* If the client previously called rx_NewCall, but it is still
1243 * waiting, treat this as a running call, and wait to destroy the
1244 * connection later when the call completes. */
1245 if ((conn->type == RX_CLIENT_CONNECTION)
1246 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1247 conn->flags |= RX_CONN_DESTROY_ME;
1248 MUTEX_EXIT(&conn->conn_data_lock);
1252 MUTEX_EXIT(&rx_refcnt_mutex);
1253 MUTEX_EXIT(&conn->conn_data_lock);
1255 /* Check for extant references to this connection */
1256 MUTEX_ENTER(&conn->conn_call_lock);
1257 for (i = 0; i < RX_MAXCALLS; i++) {
1258 struct rx_call *call = conn->call[i];
1261 if (conn->type == RX_CLIENT_CONNECTION) {
1262 MUTEX_ENTER(&call->lock);
1263 if (call->delayedAckEvent) {
1264 /* Push the final acknowledgment out now--there
1265 * won't be a subsequent call to acknowledge the
1266 * last reply packets */
1267 rxi_CancelDelayedAckEvent(call);
1268 if (call->state == RX_STATE_PRECALL
1269 || call->state == RX_STATE_ACTIVE) {
1270 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1275 MUTEX_EXIT(&call->lock);
1279 MUTEX_EXIT(&conn->conn_call_lock);
1281 #ifdef RX_ENABLE_LOCKS
1283 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1284 MUTEX_EXIT(&conn->conn_data_lock);
1286 /* Someone is accessing a packet right now. */
1290 #endif /* RX_ENABLE_LOCKS */
1293 /* Don't destroy the connection if there are any call
1294 * structures still in use */
1295 MUTEX_ENTER(&conn->conn_data_lock);
1296 conn->flags |= RX_CONN_DESTROY_ME;
1297 MUTEX_EXIT(&conn->conn_data_lock);
1302 if (conn->natKeepAliveEvent) {
1303 rxi_NatKeepAliveOff(conn);
1306 if (conn->delayedAbortEvent) {
1307 rxevent_Cancel(&conn->delayedAbortEvent);
1308 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1310 MUTEX_ENTER(&conn->conn_data_lock);
1311 rxi_SendConnectionAbort(conn, packet, 0, 1);
1312 MUTEX_EXIT(&conn->conn_data_lock);
1313 rxi_FreePacket(packet);
1317 /* Remove from connection hash table before proceeding */
1319 &rx_connHashTable[CONN_HASH
1320 (peer->host, peer->port, conn->cid, conn->epoch,
1322 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1323 if (*conn_ptr == conn) {
1324 *conn_ptr = conn->next;
1328 /* if the conn that we are destroying was the last connection, then we
1329 * clear rxLastConn as well */
1330 if (rxLastConn == conn)
1333 /* Make sure the connection is completely reset before deleting it. */
1334 /* get rid of pending events that could zap us later */
1335 rxevent_Cancel(&conn->challengeEvent);
1336 rxevent_Cancel(&conn->checkReachEvent);
1337 rxevent_Cancel(&conn->natKeepAliveEvent);
1339 /* Add the connection to the list of destroyed connections that
1340 * need to be cleaned up. This is necessary to avoid deadlocks
1341 * in the routines we call to inform others that this connection is
1342 * being destroyed. */
1343 conn->next = rx_connCleanup_list;
1344 rx_connCleanup_list = conn;
1347 /* Externally available version */
1349 rx_DestroyConnection(struct rx_connection *conn)
1354 rxi_DestroyConnection(conn);
1359 rx_GetConnection(struct rx_connection *conn)
1364 MUTEX_ENTER(&rx_refcnt_mutex);
1366 MUTEX_EXIT(&rx_refcnt_mutex);
1370 #ifdef RX_ENABLE_LOCKS
1371 /* Wait for the transmit queue to no longer be busy.
1372 * requires the call->lock to be held */
1374 rxi_WaitforTQBusy(struct rx_call *call) {
1375 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1376 call->flags |= RX_CALL_TQ_WAIT;
1378 MUTEX_ASSERT(&call->lock);
1379 CV_WAIT(&call->cv_tq, &call->lock);
1381 if (call->tqWaiters == 0) {
1382 call->flags &= ~RX_CALL_TQ_WAIT;
1389 rxi_WakeUpTransmitQueue(struct rx_call *call)
1391 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1392 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1393 call, call->tqWaiters, call->flags));
1394 #ifdef RX_ENABLE_LOCKS
1395 MUTEX_ASSERT(&call->lock);
1396 CV_BROADCAST(&call->cv_tq);
1397 #else /* RX_ENABLE_LOCKS */
1398 osi_rxWakeup(&call->tq);
1399 #endif /* RX_ENABLE_LOCKS */
1403 /* Start a new rx remote procedure call, on the specified connection.
1404 * If wait is set to 1, wait for a free call channel; otherwise return
1405 * 0. Maxtime gives the maximum number of seconds this call may take,
1406 * after rx_NewCall returns. After this time interval, a call to any
1407 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1408 * For fine grain locking, we hold the conn_call_lock in order to
1409 * to ensure that we don't get signalle after we found a call in an active
1410 * state and before we go to sleep.
1413 rx_NewCall(struct rx_connection *conn)
1415 int i, wait, ignoreBusy = 1;
1416 struct rx_call *call;
1417 struct clock queueTime;
1418 afs_uint32 leastBusy = 0;
1422 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1425 clock_GetTime(&queueTime);
1427 * Check if there are others waiting for a new call.
1428 * If so, let them go first to avoid starving them.
1429 * This is a fairly simple scheme, and might not be
1430 * a complete solution for large numbers of waiters.
1432 * makeCallWaiters keeps track of the number of
1433 * threads waiting to make calls and the
1434 * RX_CONN_MAKECALL_WAITING flag bit is used to
1435 * indicate that there are indeed calls waiting.
1436 * The flag is set when the waiter is incremented.
1437 * It is only cleared when makeCallWaiters is 0.
1438 * This prevents us from accidently destroying the
1439 * connection while it is potentially about to be used.
1441 MUTEX_ENTER(&conn->conn_call_lock);
1442 MUTEX_ENTER(&conn->conn_data_lock);
1443 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1444 conn->flags |= RX_CONN_MAKECALL_WAITING;
1445 conn->makeCallWaiters++;
1446 MUTEX_EXIT(&conn->conn_data_lock);
1448 #ifdef RX_ENABLE_LOCKS
1449 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1453 MUTEX_ENTER(&conn->conn_data_lock);
1454 conn->makeCallWaiters--;
1455 if (conn->makeCallWaiters == 0)
1456 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1459 /* We are now the active thread in rx_NewCall */
1460 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1461 MUTEX_EXIT(&conn->conn_data_lock);
1466 for (i = 0; i < RX_MAXCALLS; i++) {
1467 call = conn->call[i];
1469 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1470 /* we're not ignoring busy call slots; only look at the
1471 * call slot that is the "least" busy */
1475 if (call->state == RX_STATE_DALLY) {
1476 MUTEX_ENTER(&call->lock);
1477 if (call->state == RX_STATE_DALLY) {
1478 if (ignoreBusy && conn->lastBusy[i]) {
1479 /* if we're ignoring busy call slots, skip any ones that
1480 * have lastBusy set */
1481 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1482 leastBusy = conn->lastBusy[i];
1484 MUTEX_EXIT(&call->lock);
1489 * We are setting the state to RX_STATE_RESET to
1490 * ensure that no one else will attempt to use this
1491 * call once we drop the conn->conn_call_lock and
1492 * call->lock. We must drop the conn->conn_call_lock
1493 * before calling rxi_ResetCall because the process
1494 * of clearing the transmit queue can block for an
1495 * extended period of time. If we block while holding
1496 * the conn->conn_call_lock, then all rx_EndCall
1497 * processing will block as well. This has a detrimental
1498 * effect on overall system performance.
1500 call->state = RX_STATE_RESET;
1501 (*call->callNumber)++;
1502 MUTEX_EXIT(&conn->conn_call_lock);
1503 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1504 rxi_ResetCall(call, 0);
1505 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1509 * If we failed to be able to safely obtain the
1510 * conn->conn_call_lock we will have to drop the
1511 * call->lock to avoid a deadlock. When the call->lock
1512 * is released the state of the call can change. If it
1513 * is no longer RX_STATE_RESET then some other thread is
1516 MUTEX_EXIT(&call->lock);
1517 MUTEX_ENTER(&conn->conn_call_lock);
1518 MUTEX_ENTER(&call->lock);
1520 if (call->state == RX_STATE_RESET)
1524 * If we get here it means that after dropping
1525 * the conn->conn_call_lock and call->lock that
1526 * the call is no longer ours. If we can't find
1527 * a free call in the remaining slots we should
1528 * not go immediately to RX_CONN_MAKECALL_WAITING
1529 * because by dropping the conn->conn_call_lock
1530 * we have given up synchronization with rx_EndCall.
1531 * Instead, cycle through one more time to see if
1532 * we can find a call that can call our own.
1534 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1537 MUTEX_EXIT(&call->lock);
1540 if (ignoreBusy && conn->lastBusy[i]) {
1541 /* if we're ignoring busy call slots, skip any ones that
1542 * have lastBusy set */
1543 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1544 leastBusy = conn->lastBusy[i];
1549 /* rxi_NewCall returns with mutex locked */
1550 call = rxi_NewCall(conn, i);
1551 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1555 if (i < RX_MAXCALLS) {
1556 conn->lastBusy[i] = 0;
1561 if (leastBusy && ignoreBusy) {
1562 /* we didn't find a useable call slot, but we did see at least one
1563 * 'busy' slot; look again and only use a slot with the 'least
1569 MUTEX_ENTER(&conn->conn_data_lock);
1570 conn->flags |= RX_CONN_MAKECALL_WAITING;
1571 conn->makeCallWaiters++;
1572 MUTEX_EXIT(&conn->conn_data_lock);
1574 #ifdef RX_ENABLE_LOCKS
1575 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1579 MUTEX_ENTER(&conn->conn_data_lock);
1580 conn->makeCallWaiters--;
1581 if (conn->makeCallWaiters == 0)
1582 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1583 MUTEX_EXIT(&conn->conn_data_lock);
1585 /* Client is initially in send mode */
1586 call->state = RX_STATE_ACTIVE;
1587 call->error = conn->error;
1589 call->app.mode = RX_MODE_ERROR;
1591 call->app.mode = RX_MODE_SENDING;
1593 #ifdef AFS_RXERRQ_ENV
1594 /* remember how many network errors the peer has when we started, so if
1595 * more errors are encountered after the call starts, we know the other endpoint won't be
1596 * responding to us */
1597 call->neterr_gen = rx_atomic_read(&conn->peer->neterrs);
1600 /* remember start time for call in case we have hard dead time limit */
1601 call->queueTime = queueTime;
1602 clock_GetTime(&call->startTime);
1603 call->app.bytesSent = 0;
1604 call->app.bytesRcvd = 0;
1606 /* Turn on busy protocol. */
1607 rxi_KeepAliveOn(call);
1609 /* Attempt MTU discovery */
1610 rxi_GrowMTUOn(call);
1613 * We are no longer the active thread in rx_NewCall
1615 MUTEX_ENTER(&conn->conn_data_lock);
1616 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1617 MUTEX_EXIT(&conn->conn_data_lock);
1620 * Wake up anyone else who might be giving us a chance to
1621 * run (see code above that avoids resource starvation).
1623 #ifdef RX_ENABLE_LOCKS
1624 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1625 osi_Panic("rx_NewCall call about to be used without an empty tq");
1628 CV_BROADCAST(&conn->conn_call_cv);
1632 MUTEX_EXIT(&conn->conn_call_lock);
1633 MUTEX_EXIT(&call->lock);
1636 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1641 rxi_HasActiveCalls(struct rx_connection *aconn)
1644 struct rx_call *tcall;
1648 for (i = 0; i < RX_MAXCALLS; i++) {
1649 if ((tcall = aconn->call[i])) {
1650 if ((tcall->state == RX_STATE_ACTIVE)
1651 || (tcall->state == RX_STATE_PRECALL)) {
1662 rxi_GetCallNumberVector(struct rx_connection *aconn,
1663 afs_int32 * aint32s)
1666 struct rx_call *tcall;
1670 MUTEX_ENTER(&aconn->conn_call_lock);
1671 for (i = 0; i < RX_MAXCALLS; i++) {
1672 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1673 aint32s[i] = aconn->callNumber[i] + 1;
1675 aint32s[i] = aconn->callNumber[i];
1677 MUTEX_EXIT(&aconn->conn_call_lock);
1683 rxi_SetCallNumberVector(struct rx_connection *aconn,
1684 afs_int32 * aint32s)
1687 struct rx_call *tcall;
1691 MUTEX_ENTER(&aconn->conn_call_lock);
1692 for (i = 0; i < RX_MAXCALLS; i++) {
1693 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1694 aconn->callNumber[i] = aint32s[i] - 1;
1696 aconn->callNumber[i] = aint32s[i];
1698 MUTEX_EXIT(&aconn->conn_call_lock);
1703 /* Advertise a new service. A service is named locally by a UDP port
1704 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1707 char *serviceName; Name for identification purposes (e.g. the
1708 service name might be used for probing for
1711 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1712 char *serviceName, struct rx_securityClass **securityObjects,
1713 int nSecurityObjects,
1714 afs_int32(*serviceProc) (struct rx_call * acall))
1716 osi_socket socket = OSI_NULLSOCKET;
1717 struct rx_service *tservice;
1723 if (serviceId == 0) {
1725 "rx_NewService: service id for service %s is not non-zero.\n",
1732 "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",
1740 tservice = rxi_AllocService();
1743 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1745 for (i = 0; i < RX_MAX_SERVICES; i++) {
1746 struct rx_service *service = rx_services[i];
1748 if (port == service->servicePort && host == service->serviceHost) {
1749 if (service->serviceId == serviceId) {
1750 /* The identical service has already been
1751 * installed; if the caller was intending to
1752 * change the security classes used by this
1753 * service, he/she loses. */
1755 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1756 serviceName, serviceId, service->serviceName);
1758 rxi_FreeService(tservice);
1761 /* Different service, same port: re-use the socket
1762 * which is bound to the same port */
1763 socket = service->socket;
1766 if (socket == OSI_NULLSOCKET) {
1767 /* If we don't already have a socket (from another
1768 * service on same port) get a new one */
1769 socket = rxi_GetHostUDPSocket(host, port);
1770 if (socket == OSI_NULLSOCKET) {
1772 rxi_FreeService(tservice);
1777 service->socket = socket;
1778 service->serviceHost = host;
1779 service->servicePort = port;
1780 service->serviceId = serviceId;
1781 service->serviceName = serviceName;
1782 service->nSecurityObjects = nSecurityObjects;
1783 service->securityObjects = securityObjects;
1784 service->minProcs = 0;
1785 service->maxProcs = 1;
1786 service->idleDeadTime = 60;
1787 service->connDeadTime = rx_connDeadTime;
1788 service->executeRequestProc = serviceProc;
1789 service->checkReach = 0;
1790 service->nSpecific = 0;
1791 service->specific = NULL;
1792 rx_services[i] = service; /* not visible until now */
1798 rxi_FreeService(tservice);
1799 (osi_Msg "rx_NewService: cannot support > %d services\n",
1804 /* Set configuration options for all of a service's security objects */
1807 rx_SetSecurityConfiguration(struct rx_service *service,
1808 rx_securityConfigVariables type,
1812 for (i = 0; i<service->nSecurityObjects; i++) {
1813 if (service->securityObjects[i]) {
1814 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1822 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1823 struct rx_securityClass **securityObjects, int nSecurityObjects,
1824 afs_int32(*serviceProc) (struct rx_call * acall))
1826 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1829 /* Generic request processing loop. This routine should be called
1830 * by the implementation dependent rx_ServerProc. If socketp is
1831 * non-null, it will be set to the file descriptor that this thread
1832 * is now listening on. If socketp is null, this routine will never
1835 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1837 struct rx_call *call;
1839 struct rx_service *tservice = NULL;
1846 call = rx_GetCall(threadID, tservice, socketp);
1847 if (socketp && *socketp != OSI_NULLSOCKET) {
1848 /* We are now a listener thread */
1854 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1855 #ifdef RX_ENABLE_LOCKS
1857 #endif /* RX_ENABLE_LOCKS */
1858 afs_termState = AFSOP_STOP_AFS;
1859 afs_osi_Wakeup(&afs_termState);
1860 #ifdef RX_ENABLE_LOCKS
1862 #endif /* RX_ENABLE_LOCKS */
1867 /* if server is restarting( typically smooth shutdown) then do not
1868 * allow any new calls.
1871 if (rx_tranquil && (call != NULL)) {
1875 MUTEX_ENTER(&call->lock);
1877 rxi_CallError(call, RX_RESTARTING);
1878 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1880 MUTEX_EXIT(&call->lock);
1885 tservice = call->conn->service;
1887 if (tservice->beforeProc)
1888 (*tservice->beforeProc) (call);
1890 code = tservice->executeRequestProc(call);
1892 if (tservice->afterProc)
1893 (*tservice->afterProc) (call, code);
1895 rx_EndCall(call, code);
1897 if (tservice->postProc)
1898 (*tservice->postProc) (code);
1900 if (rx_stats_active) {
1901 MUTEX_ENTER(&rx_stats_mutex);
1903 MUTEX_EXIT(&rx_stats_mutex);
1910 rx_WakeupServerProcs(void)
1912 struct rx_serverQueueEntry *np, *tqp;
1913 struct opr_queue *cursor;
1917 MUTEX_ENTER(&rx_serverPool_lock);
1919 #ifdef RX_ENABLE_LOCKS
1920 if (rx_waitForPacket)
1921 CV_BROADCAST(&rx_waitForPacket->cv);
1922 #else /* RX_ENABLE_LOCKS */
1923 if (rx_waitForPacket)
1924 osi_rxWakeup(rx_waitForPacket);
1925 #endif /* RX_ENABLE_LOCKS */
1926 MUTEX_ENTER(&freeSQEList_lock);
1927 for (np = rx_FreeSQEList; np; np = tqp) {
1928 tqp = *(struct rx_serverQueueEntry **)np;
1929 #ifdef RX_ENABLE_LOCKS
1930 CV_BROADCAST(&np->cv);
1931 #else /* RX_ENABLE_LOCKS */
1933 #endif /* RX_ENABLE_LOCKS */
1935 MUTEX_EXIT(&freeSQEList_lock);
1936 for (opr_queue_Scan(&rx_idleServerQueue, cursor)) {
1937 np = opr_queue_Entry(cursor, struct rx_serverQueueEntry, entry);
1938 #ifdef RX_ENABLE_LOCKS
1939 CV_BROADCAST(&np->cv);
1940 #else /* RX_ENABLE_LOCKS */
1942 #endif /* RX_ENABLE_LOCKS */
1944 MUTEX_EXIT(&rx_serverPool_lock);
1949 * One thing that seems to happen is that all the server threads get
1950 * tied up on some empty or slow call, and then a whole bunch of calls
1951 * arrive at once, using up the packet pool, so now there are more
1952 * empty calls. The most critical resources here are server threads
1953 * and the free packet pool. The "doreclaim" code seems to help in
1954 * general. I think that eventually we arrive in this state: there
1955 * are lots of pending calls which do have all their packets present,
1956 * so they won't be reclaimed, are multi-packet calls, so they won't
1957 * be scheduled until later, and thus are tying up most of the free
1958 * packet pool for a very long time.
1960 * 1. schedule multi-packet calls if all the packets are present.
1961 * Probably CPU-bound operation, useful to return packets to pool.
1962 * Do what if there is a full window, but the last packet isn't here?
1963 * 3. preserve one thread which *only* runs "best" calls, otherwise
1964 * it sleeps and waits for that type of call.
1965 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1966 * the current dataquota business is badly broken. The quota isn't adjusted
1967 * to reflect how many packets are presently queued for a running call.
1968 * So, when we schedule a queued call with a full window of packets queued
1969 * up for it, that *should* free up a window full of packets for other 2d-class
1970 * calls to be able to use from the packet pool. But it doesn't.
1972 * NB. Most of the time, this code doesn't run -- since idle server threads
1973 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1974 * as a new call arrives.
1976 /* Sleep until a call arrives. Returns a pointer to the call, ready
1977 * for an rx_Read. */
1978 #ifdef RX_ENABLE_LOCKS
1980 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1982 struct rx_serverQueueEntry *sq;
1983 struct rx_call *call = (struct rx_call *)0;
1984 struct rx_service *service = NULL;
1986 MUTEX_ENTER(&freeSQEList_lock);
1988 if ((sq = rx_FreeSQEList)) {
1989 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1990 MUTEX_EXIT(&freeSQEList_lock);
1991 } else { /* otherwise allocate a new one and return that */
1992 MUTEX_EXIT(&freeSQEList_lock);
1993 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1994 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1995 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1998 MUTEX_ENTER(&rx_serverPool_lock);
1999 if (cur_service != NULL) {
2000 ReturnToServerPool(cur_service);
2003 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2004 struct rx_call *tcall, *choice2 = NULL;
2005 struct opr_queue *cursor;
2007 /* Scan for eligible incoming calls. A call is not eligible
2008 * if the maximum number of calls for its service type are
2009 * already executing */
2010 /* One thread will process calls FCFS (to prevent starvation),
2011 * while the other threads may run ahead looking for calls which
2012 * have all their input data available immediately. This helps
2013 * keep threads from blocking, waiting for data from the client. */
2014 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2015 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2017 service = tcall->conn->service;
2018 if (!QuotaOK(service)) {
2021 MUTEX_ENTER(&rx_pthread_mutex);
2022 if (tno == rxi_fcfs_thread_num
2023 || opr_queue_IsEnd(&rx_incomingCallQueue, cursor)) {
2024 MUTEX_EXIT(&rx_pthread_mutex);
2025 /* If we're the fcfs thread , then we'll just use
2026 * this call. If we haven't been able to find an optimal
2027 * choice, and we're at the end of the list, then use a
2028 * 2d choice if one has been identified. Otherwise... */
2029 call = (choice2 ? choice2 : tcall);
2030 service = call->conn->service;
2032 MUTEX_EXIT(&rx_pthread_mutex);
2033 if (!opr_queue_IsEmpty(&tcall->rq)) {
2034 struct rx_packet *rp;
2035 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2037 if (rp->header.seq == 1) {
2039 || (rp->header.flags & RX_LAST_PACKET)) {
2041 } else if (rxi_2dchoice && !choice2
2042 && !(tcall->flags & RX_CALL_CLEARED)
2043 && (tcall->rprev > rxi_HardAckRate)) {
2053 ReturnToServerPool(service);
2059 opr_queue_Remove(&call->entry);
2060 MUTEX_EXIT(&rx_serverPool_lock);
2061 MUTEX_ENTER(&call->lock);
2063 if (call->flags & RX_CALL_WAIT_PROC) {
2064 call->flags &= ~RX_CALL_WAIT_PROC;
2065 rx_atomic_dec(&rx_nWaiting);
2068 if (call->state != RX_STATE_PRECALL || call->error) {
2069 MUTEX_EXIT(&call->lock);
2070 MUTEX_ENTER(&rx_serverPool_lock);
2071 ReturnToServerPool(service);
2076 if (opr_queue_IsEmpty(&call->rq)
2077 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1)
2078 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2080 CLEAR_CALL_QUEUE_LOCK(call);
2083 /* If there are no eligible incoming calls, add this process
2084 * to the idle server queue, to wait for one */
2088 *socketp = OSI_NULLSOCKET;
2090 sq->socketp = socketp;
2091 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2092 #ifndef AFS_AIX41_ENV
2093 rx_waitForPacket = sq;
2094 #endif /* AFS_AIX41_ENV */
2096 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2098 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2099 MUTEX_EXIT(&rx_serverPool_lock);
2100 return (struct rx_call *)0;
2103 } while (!(call = sq->newcall)
2104 && !(socketp && *socketp != OSI_NULLSOCKET));
2105 MUTEX_EXIT(&rx_serverPool_lock);
2107 MUTEX_ENTER(&call->lock);
2113 MUTEX_ENTER(&freeSQEList_lock);
2114 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2115 rx_FreeSQEList = sq;
2116 MUTEX_EXIT(&freeSQEList_lock);
2119 clock_GetTime(&call->startTime);
2120 call->state = RX_STATE_ACTIVE;
2121 call->app.mode = RX_MODE_RECEIVING;
2122 #ifdef RX_KERNEL_TRACE
2123 if (ICL_SETACTIVE(afs_iclSetp)) {
2124 int glockOwner = ISAFS_GLOCK();
2127 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2128 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2135 rxi_calltrace(RX_CALL_START, call);
2136 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2137 call->conn->service->servicePort, call->conn->service->serviceId,
2140 MUTEX_EXIT(&call->lock);
2141 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2143 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2148 #else /* RX_ENABLE_LOCKS */
2150 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2152 struct rx_serverQueueEntry *sq;
2153 struct rx_call *call = (struct rx_call *)0, *choice2;
2154 struct rx_service *service = NULL;
2158 MUTEX_ENTER(&freeSQEList_lock);
2160 if ((sq = rx_FreeSQEList)) {
2161 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2162 MUTEX_EXIT(&freeSQEList_lock);
2163 } else { /* otherwise allocate a new one and return that */
2164 MUTEX_EXIT(&freeSQEList_lock);
2165 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2166 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2167 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2169 MUTEX_ENTER(&sq->lock);
2171 if (cur_service != NULL) {
2172 cur_service->nRequestsRunning--;
2173 MUTEX_ENTER(&rx_quota_mutex);
2174 if (cur_service->nRequestsRunning < cur_service->minProcs)
2177 MUTEX_EXIT(&rx_quota_mutex);
2179 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2180 struct rx_call *tcall;
2181 struct opr_queue *cursor;
2182 /* Scan for eligible incoming calls. A call is not eligible
2183 * if the maximum number of calls for its service type are
2184 * already executing */
2185 /* One thread will process calls FCFS (to prevent starvation),
2186 * while the other threads may run ahead looking for calls which
2187 * have all their input data available immediately. This helps
2188 * keep threads from blocking, waiting for data from the client. */
2189 choice2 = (struct rx_call *)0;
2190 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2191 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2192 service = tcall->conn->service;
2193 if (QuotaOK(service)) {
2194 MUTEX_ENTER(&rx_pthread_mutex);
2195 /* XXX - If tcall->entry.next is NULL, then we're no longer
2196 * on a queue at all. This shouldn't happen. */
2197 if (tno == rxi_fcfs_thread_num || !tcall->entry.next) {
2198 MUTEX_EXIT(&rx_pthread_mutex);
2199 /* If we're the fcfs thread, then we'll just use
2200 * this call. If we haven't been able to find an optimal
2201 * choice, and we're at the end of the list, then use a
2202 * 2d choice if one has been identified. Otherwise... */
2203 call = (choice2 ? choice2 : tcall);
2204 service = call->conn->service;
2206 MUTEX_EXIT(&rx_pthread_mutex);
2207 if (!opr_queue_IsEmpty(&tcall->rq)) {
2208 struct rx_packet *rp;
2209 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2211 if (rp->header.seq == 1
2213 || (rp->header.flags & RX_LAST_PACKET))) {
2215 } else if (rxi_2dchoice && !choice2
2216 && !(tcall->flags & RX_CALL_CLEARED)
2217 && (tcall->rprev > rxi_HardAckRate)) {
2230 opr_queue_Remove(&call->entry);
2231 /* we can't schedule a call if there's no data!!! */
2232 /* send an ack if there's no data, if we're missing the
2233 * first packet, or we're missing something between first
2234 * and last -- there's a "hole" in the incoming data. */
2235 if (opr_queue_IsEmpty(&call->rq)
2236 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1
2237 || call->rprev != opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq)
2238 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2240 call->flags &= (~RX_CALL_WAIT_PROC);
2241 service->nRequestsRunning++;
2242 /* just started call in minProcs pool, need fewer to maintain
2244 MUTEX_ENTER(&rx_quota_mutex);
2245 if (service->nRequestsRunning <= service->minProcs)
2248 MUTEX_EXIT(&rx_quota_mutex);
2249 rx_atomic_dec(&rx_nWaiting);
2250 /* MUTEX_EXIT(&call->lock); */
2252 /* If there are no eligible incoming calls, add this process
2253 * to the idle server queue, to wait for one */
2256 *socketp = OSI_NULLSOCKET;
2258 sq->socketp = socketp;
2259 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2263 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2265 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2266 return (struct rx_call *)0;
2269 } while (!(call = sq->newcall)
2270 && !(socketp && *socketp != OSI_NULLSOCKET));
2272 MUTEX_EXIT(&sq->lock);
2274 MUTEX_ENTER(&freeSQEList_lock);
2275 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2276 rx_FreeSQEList = sq;
2277 MUTEX_EXIT(&freeSQEList_lock);
2280 clock_GetTime(&call->startTime);
2281 call->state = RX_STATE_ACTIVE;
2282 call->app.mode = RX_MODE_RECEIVING;
2283 #ifdef RX_KERNEL_TRACE
2284 if (ICL_SETACTIVE(afs_iclSetp)) {
2285 int glockOwner = ISAFS_GLOCK();
2288 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2289 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2296 rxi_calltrace(RX_CALL_START, call);
2297 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2298 call->conn->service->servicePort, call->conn->service->serviceId,
2301 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2308 #endif /* RX_ENABLE_LOCKS */
2312 /* Establish a procedure to be called when a packet arrives for a
2313 * call. This routine will be called at most once after each call,
2314 * and will also be called if there is an error condition on the or
2315 * the call is complete. Used by multi rx to build a selection
2316 * function which determines which of several calls is likely to be a
2317 * good one to read from.
2318 * NOTE: the way this is currently implemented it is probably only a
2319 * good idea to (1) use it immediately after a newcall (clients only)
2320 * and (2) only use it once. Other uses currently void your warranty
2323 rx_SetArrivalProc(struct rx_call *call,
2324 void (*proc) (struct rx_call * call,
2327 void * handle, int arg)
2329 call->arrivalProc = proc;
2330 call->arrivalProcHandle = handle;
2331 call->arrivalProcArg = arg;
2334 /* Call is finished (possibly prematurely). Return rc to the peer, if
2335 * appropriate, and return the final error code from the conversation
2339 rx_EndCall(struct rx_call *call, afs_int32 rc)
2341 struct rx_connection *conn = call->conn;
2345 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2346 call, rc, call->error, call->abortCode));
2349 MUTEX_ENTER(&call->lock);
2351 if (rc == 0 && call->error == 0) {
2352 call->abortCode = 0;
2353 call->abortCount = 0;
2356 call->arrivalProc = (void (*)())0;
2357 if (rc && call->error == 0) {
2358 rxi_CallError(call, rc);
2359 call->app.mode = RX_MODE_ERROR;
2360 /* Send an abort message to the peer if this error code has
2361 * only just been set. If it was set previously, assume the
2362 * peer has already been sent the error code or will request it
2364 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2366 if (conn->type == RX_SERVER_CONNECTION) {
2367 /* Make sure reply or at least dummy reply is sent */
2368 if (call->app.mode == RX_MODE_RECEIVING) {
2369 MUTEX_EXIT(&call->lock);
2370 rxi_WriteProc(call, 0, 0);
2371 MUTEX_ENTER(&call->lock);
2373 if (call->app.mode == RX_MODE_SENDING) {
2374 rxi_FlushWriteLocked(call);
2376 rxi_calltrace(RX_CALL_END, call);
2377 /* Call goes to hold state until reply packets are acknowledged */
2378 if (call->tfirst + call->nSoftAcked < call->tnext) {
2379 call->state = RX_STATE_HOLD;
2381 call->state = RX_STATE_DALLY;
2382 rxi_ClearTransmitQueue(call, 0);
2383 rxi_rto_cancel(call);
2384 rxi_CancelKeepAliveEvent(call);
2386 } else { /* Client connection */
2388 /* Make sure server receives input packets, in the case where
2389 * no reply arguments are expected */
2391 if ((call->app.mode == RX_MODE_SENDING)
2392 || (call->app.mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2393 MUTEX_EXIT(&call->lock);
2394 (void)rxi_ReadProc(call, &dummy, 1);
2395 MUTEX_ENTER(&call->lock);
2398 /* If we had an outstanding delayed ack, be nice to the server
2399 * and force-send it now.
2401 if (call->delayedAckEvent) {
2402 rxi_CancelDelayedAckEvent(call);
2403 rxi_SendDelayedAck(NULL, call, NULL, 0);
2406 /* We need to release the call lock since it's lower than the
2407 * conn_call_lock and we don't want to hold the conn_call_lock
2408 * over the rx_ReadProc call. The conn_call_lock needs to be held
2409 * here for the case where rx_NewCall is perusing the calls on
2410 * the connection structure. We don't want to signal until
2411 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2412 * have checked this call, found it active and by the time it
2413 * goes to sleep, will have missed the signal.
2415 MUTEX_EXIT(&call->lock);
2416 MUTEX_ENTER(&conn->conn_call_lock);
2417 MUTEX_ENTER(&call->lock);
2420 /* While there are some circumstances where a call with an error is
2421 * obviously not on a "busy" channel, be conservative (clearing
2422 * lastBusy is just best-effort to possibly speed up rx_NewCall).
2423 * The call channel is definitely not busy if we just successfully
2424 * completed a call on it. */
2425 conn->lastBusy[call->channel] = 0;
2427 } else if (call->error == RX_CALL_TIMEOUT) {
2428 /* The call is still probably running on the server side, so try to
2429 * avoid this call channel in the future. */
2430 conn->lastBusy[call->channel] = clock_Sec();
2433 MUTEX_ENTER(&conn->conn_data_lock);
2434 conn->flags |= RX_CONN_BUSY;
2435 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2436 MUTEX_EXIT(&conn->conn_data_lock);
2437 #ifdef RX_ENABLE_LOCKS
2438 CV_BROADCAST(&conn->conn_call_cv);
2443 #ifdef RX_ENABLE_LOCKS
2445 MUTEX_EXIT(&conn->conn_data_lock);
2447 #endif /* RX_ENABLE_LOCKS */
2448 call->state = RX_STATE_DALLY;
2450 error = call->error;
2452 /* currentPacket, nLeft, and NFree must be zeroed here, because
2453 * ResetCall cannot: ResetCall may be called at splnet(), in the
2454 * kernel version, and may interrupt the macros rx_Read or
2455 * rx_Write, which run at normal priority for efficiency. */
2456 if (call->app.currentPacket) {
2457 #ifdef RX_TRACK_PACKETS
2458 call->app.currentPacket->flags &= ~RX_PKTFLAG_CP;
2460 rxi_FreePacket(call->app.currentPacket);
2461 call->app.currentPacket = (struct rx_packet *)0;
2464 call->app.nLeft = call->app.nFree = call->app.curlen = 0;
2466 /* Free any packets from the last call to ReadvProc/WritevProc */
2467 #ifdef RXDEBUG_PACKET
2469 #endif /* RXDEBUG_PACKET */
2470 rxi_FreePackets(0, &call->app.iovq);
2471 MUTEX_EXIT(&call->lock);
2473 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2474 if (conn->type == RX_CLIENT_CONNECTION) {
2475 MUTEX_ENTER(&conn->conn_data_lock);
2476 conn->flags &= ~RX_CONN_BUSY;
2477 MUTEX_EXIT(&conn->conn_data_lock);
2478 MUTEX_EXIT(&conn->conn_call_lock);
2482 * Map errors to the local host's errno.h format.
2484 error = ntoh_syserr_conv(error);
2486 /* If the caller said the call failed with some error, we had better
2487 * return an error code. */
2488 osi_Assert(!rc || error);
2492 #if !defined(KERNEL)
2494 /* Call this routine when shutting down a server or client (especially
2495 * clients). This will allow Rx to gracefully garbage collect server
2496 * connections, and reduce the number of retries that a server might
2497 * make to a dead client.
2498 * This is not quite right, since some calls may still be ongoing and
2499 * we can't lock them to destroy them. */
2503 struct rx_connection **conn_ptr, **conn_end;
2506 if (rx_atomic_test_and_set_bit(&rxinit_status, 0))
2507 return; /* Already shutdown. */
2509 rxi_DeleteCachedConnections();
2510 if (rx_connHashTable) {
2511 MUTEX_ENTER(&rx_connHashTable_lock);
2512 for (conn_ptr = &rx_connHashTable[0], conn_end =
2513 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2515 struct rx_connection *conn, *next;
2516 for (conn = *conn_ptr; conn; conn = next) {
2518 if (conn->type == RX_CLIENT_CONNECTION) {
2519 MUTEX_ENTER(&rx_refcnt_mutex);
2521 MUTEX_EXIT(&rx_refcnt_mutex);
2522 #ifdef RX_ENABLE_LOCKS
2523 rxi_DestroyConnectionNoLock(conn);
2524 #else /* RX_ENABLE_LOCKS */
2525 rxi_DestroyConnection(conn);
2526 #endif /* RX_ENABLE_LOCKS */
2530 #ifdef RX_ENABLE_LOCKS
2531 while (rx_connCleanup_list) {
2532 struct rx_connection *conn;
2533 conn = rx_connCleanup_list;
2534 rx_connCleanup_list = rx_connCleanup_list->next;
2535 MUTEX_EXIT(&rx_connHashTable_lock);
2536 rxi_CleanupConnection(conn);
2537 MUTEX_ENTER(&rx_connHashTable_lock);
2539 MUTEX_EXIT(&rx_connHashTable_lock);
2540 #endif /* RX_ENABLE_LOCKS */
2545 afs_winsockCleanup();
2551 /* if we wakeup packet waiter too often, can get in loop with two
2552 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2554 rxi_PacketsUnWait(void)
2556 if (!rx_waitingForPackets) {
2560 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2561 return; /* still over quota */
2564 rx_waitingForPackets = 0;
2565 #ifdef RX_ENABLE_LOCKS
2566 CV_BROADCAST(&rx_waitingForPackets_cv);
2568 osi_rxWakeup(&rx_waitingForPackets);
2574 /* ------------------Internal interfaces------------------------- */
2576 /* Return this process's service structure for the
2577 * specified socket and service */
2578 static struct rx_service *
2579 rxi_FindService(osi_socket socket, u_short serviceId)
2581 struct rx_service **sp;
2582 for (sp = &rx_services[0]; *sp; sp++) {
2583 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2589 #ifdef RXDEBUG_PACKET
2590 #ifdef KDUMP_RX_LOCK
2591 static struct rx_call_rx_lock *rx_allCallsp = 0;
2593 static struct rx_call *rx_allCallsp = 0;
2595 #endif /* RXDEBUG_PACKET */
2597 /* Allocate a call structure, for the indicated channel of the
2598 * supplied connection. The mode and state of the call must be set by
2599 * the caller. Returns the call with mutex locked. */
2600 static struct rx_call *
2601 rxi_NewCall(struct rx_connection *conn, int channel)
2603 struct rx_call *call;
2604 #ifdef RX_ENABLE_LOCKS
2605 struct rx_call *cp; /* Call pointer temp */
2606 struct opr_queue *cursor;
2609 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2611 /* Grab an existing call structure, or allocate a new one.
2612 * Existing call structures are assumed to have been left reset by
2614 MUTEX_ENTER(&rx_freeCallQueue_lock);
2616 #ifdef RX_ENABLE_LOCKS
2618 * EXCEPT that the TQ might not yet be cleared out.
2619 * Skip over those with in-use TQs.
2622 for (opr_queue_Scan(&rx_freeCallQueue, cursor)) {
2623 cp = opr_queue_Entry(cursor, struct rx_call, entry);
2624 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2630 #else /* RX_ENABLE_LOCKS */
2631 if (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
2632 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
2633 #endif /* RX_ENABLE_LOCKS */
2634 opr_queue_Remove(&call->entry);
2635 if (rx_stats_active)
2636 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2637 MUTEX_EXIT(&rx_freeCallQueue_lock);
2638 MUTEX_ENTER(&call->lock);
2639 CLEAR_CALL_QUEUE_LOCK(call);
2640 #ifdef RX_ENABLE_LOCKS
2641 /* Now, if TQ wasn't cleared earlier, do it now. */
2642 rxi_WaitforTQBusy(call);
2643 if (call->flags & RX_CALL_TQ_CLEARME) {
2644 rxi_ClearTransmitQueue(call, 1);
2645 /*queue_Init(&call->tq);*/
2647 #endif /* RX_ENABLE_LOCKS */
2648 /* Bind the call to its connection structure */
2650 rxi_ResetCall(call, 1);
2653 call = rxi_Alloc(sizeof(struct rx_call));
2654 #ifdef RXDEBUG_PACKET
2655 call->allNextp = rx_allCallsp;
2656 rx_allCallsp = call;
2658 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2659 #else /* RXDEBUG_PACKET */
2660 rx_atomic_inc(&rx_stats.nCallStructs);
2661 #endif /* RXDEBUG_PACKET */
2663 MUTEX_EXIT(&rx_freeCallQueue_lock);
2664 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2665 MUTEX_ENTER(&call->lock);
2666 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2667 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2668 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2670 /* Initialize once-only items */
2671 opr_queue_Init(&call->tq);
2672 opr_queue_Init(&call->rq);
2673 opr_queue_Init(&call->app.iovq);
2674 #ifdef RXDEBUG_PACKET
2675 call->rqc = call->tqc = call->iovqc = 0;
2676 #endif /* RXDEBUG_PACKET */
2677 /* Bind the call to its connection structure (prereq for reset) */
2679 rxi_ResetCall(call, 1);
2681 call->channel = channel;
2682 call->callNumber = &conn->callNumber[channel];
2683 call->rwind = conn->rwind[channel];
2684 call->twind = conn->twind[channel];
2685 /* Note that the next expected call number is retained (in
2686 * conn->callNumber[i]), even if we reallocate the call structure
2688 conn->call[channel] = call;
2689 /* if the channel's never been used (== 0), we should start at 1, otherwise
2690 * the call number is valid from the last time this channel was used */
2691 if (*call->callNumber == 0)
2692 *call->callNumber = 1;
2697 /* A call has been inactive long enough that so we can throw away
2698 * state, including the call structure, which is placed on the call
2701 * call->lock amd rx_refcnt_mutex are held upon entry.
2702 * haveCTLock is set when called from rxi_ReapConnections.
2704 * return 1 if the call is freed, 0 if not.
2707 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2709 int channel = call->channel;
2710 struct rx_connection *conn = call->conn;
2711 u_char state = call->state;
2714 * We are setting the state to RX_STATE_RESET to
2715 * ensure that no one else will attempt to use this
2716 * call once we drop the refcnt lock. We must drop
2717 * the refcnt lock before calling rxi_ResetCall
2718 * because it cannot be held across acquiring the
2719 * freepktQ lock. NewCall does the same.
2721 call->state = RX_STATE_RESET;
2722 MUTEX_EXIT(&rx_refcnt_mutex);
2723 rxi_ResetCall(call, 0);
2725 if (MUTEX_TRYENTER(&conn->conn_call_lock))
2727 if (state == RX_STATE_DALLY || state == RX_STATE_HOLD)
2728 (*call->callNumber)++;
2730 if (call->conn->call[channel] == call)
2731 call->conn->call[channel] = 0;
2732 MUTEX_EXIT(&conn->conn_call_lock);
2735 * We couldn't obtain the conn_call_lock so we can't
2736 * disconnect the call from the connection. Set the
2737 * call state to dally so that the call can be reused.
2739 MUTEX_ENTER(&rx_refcnt_mutex);
2740 call->state = RX_STATE_DALLY;
2744 MUTEX_ENTER(&rx_freeCallQueue_lock);
2745 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2746 #ifdef RX_ENABLE_LOCKS
2747 /* A call may be free even though its transmit queue is still in use.
2748 * Since we search the call list from head to tail, put busy calls at
2749 * the head of the list, and idle calls at the tail.
2751 if (call->flags & RX_CALL_TQ_BUSY)
2752 opr_queue_Prepend(&rx_freeCallQueue, &call->entry);
2754 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2755 #else /* RX_ENABLE_LOCKS */
2756 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2757 #endif /* RX_ENABLE_LOCKS */
2758 if (rx_stats_active)
2759 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2760 MUTEX_EXIT(&rx_freeCallQueue_lock);
2762 /* Destroy the connection if it was previously slated for
2763 * destruction, i.e. the Rx client code previously called
2764 * rx_DestroyConnection (client connections), or
2765 * rxi_ReapConnections called the same routine (server
2766 * connections). Only do this, however, if there are no
2767 * outstanding calls. Note that for fine grain locking, there appears
2768 * to be a deadlock in that rxi_FreeCall has a call locked and
2769 * DestroyConnectionNoLock locks each call in the conn. But note a
2770 * few lines up where we have removed this call from the conn.
2771 * If someone else destroys a connection, they either have no
2772 * call lock held or are going through this section of code.
2774 MUTEX_ENTER(&conn->conn_data_lock);
2775 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2776 MUTEX_ENTER(&rx_refcnt_mutex);
2778 MUTEX_EXIT(&rx_refcnt_mutex);
2779 MUTEX_EXIT(&conn->conn_data_lock);
2780 #ifdef RX_ENABLE_LOCKS
2782 rxi_DestroyConnectionNoLock(conn);
2784 rxi_DestroyConnection(conn);
2785 #else /* RX_ENABLE_LOCKS */
2786 rxi_DestroyConnection(conn);
2787 #endif /* RX_ENABLE_LOCKS */
2789 MUTEX_EXIT(&conn->conn_data_lock);
2791 MUTEX_ENTER(&rx_refcnt_mutex);
2795 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2796 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2799 rxi_Alloc(size_t size)
2803 if (rx_stats_active) {
2804 rx_atomic_add(&rxi_Allocsize, (int) size);
2805 rx_atomic_inc(&rxi_Alloccnt);
2809 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2810 afs_osi_Alloc_NoSleep(size);
2815 osi_Panic("rxi_Alloc error");
2821 rxi_Free(void *addr, size_t size)
2823 if (rx_stats_active) {
2824 rx_atomic_sub(&rxi_Allocsize, (int) size);
2825 rx_atomic_dec(&rxi_Alloccnt);
2827 osi_Free(addr, size);
2831 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2833 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2834 struct rx_peer *next = NULL;
2838 MUTEX_ENTER(&rx_peerHashTable_lock);
2840 peer_ptr = &rx_peerHashTable[0];
2841 peer_end = &rx_peerHashTable[rx_hashTableSize];
2844 for ( ; peer_ptr < peer_end; peer_ptr++) {
2847 for ( ; peer; peer = next) {
2849 if (host == peer->host)
2854 hashIndex = PEER_HASH(host, port);
2855 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2856 if ((peer->host == host) && (peer->port == port))
2861 MUTEX_ENTER(&rx_peerHashTable_lock);
2866 MUTEX_EXIT(&rx_peerHashTable_lock);
2868 MUTEX_ENTER(&peer->peer_lock);
2869 /* We don't handle dropping below min, so don't */
2870 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2871 peer->ifMTU=MIN(mtu, peer->ifMTU);
2872 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2873 /* if we tweaked this down, need to tune our peer MTU too */
2874 peer->MTU = MIN(peer->MTU, peer->natMTU);
2875 /* if we discovered a sub-1500 mtu, degrade */
2876 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2877 peer->maxDgramPackets = 1;
2878 /* We no longer have valid peer packet information */
2879 if (peer->maxPacketSize + RX_HEADER_SIZE > peer->ifMTU)
2880 peer->maxPacketSize = 0;
2881 MUTEX_EXIT(&peer->peer_lock);
2883 MUTEX_ENTER(&rx_peerHashTable_lock);
2885 if (host && !port) {
2887 /* pick up where we left off */
2891 MUTEX_EXIT(&rx_peerHashTable_lock);
2894 #ifdef AFS_RXERRQ_ENV
2896 rxi_SetPeerDead(struct sock_extended_err *err, afs_uint32 host, afs_uint16 port)
2898 int hashIndex = PEER_HASH(host, port);
2899 struct rx_peer *peer;
2901 MUTEX_ENTER(&rx_peerHashTable_lock);
2903 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2904 if (peer->host == host && peer->port == port) {
2910 MUTEX_EXIT(&rx_peerHashTable_lock);
2913 rx_atomic_inc(&peer->neterrs);
2914 MUTEX_ENTER(&peer->peer_lock);
2915 peer->last_err_origin = RX_NETWORK_ERROR_ORIGIN_ICMP;
2916 peer->last_err_type = err->ee_type;
2917 peer->last_err_code = err->ee_code;
2918 MUTEX_EXIT(&peer->peer_lock);
2920 MUTEX_ENTER(&rx_peerHashTable_lock);
2922 MUTEX_EXIT(&rx_peerHashTable_lock);
2927 rxi_ProcessNetError(struct sock_extended_err *err, afs_uint32 addr, afs_uint16 port)
2929 # ifdef AFS_ADAPT_PMTU
2930 if (err->ee_errno == EMSGSIZE && err->ee_info >= 68) {
2931 rxi_SetPeerMtu(NULL, addr, port, err->ee_info - RX_IPUDP_SIZE);
2935 if (err->ee_origin == SO_EE_ORIGIN_ICMP && err->ee_type == ICMP_DEST_UNREACH) {
2936 switch (err->ee_code) {
2937 case ICMP_NET_UNREACH:
2938 case ICMP_HOST_UNREACH:
2939 case ICMP_PORT_UNREACH:
2942 rxi_SetPeerDead(err, addr, port);
2949 rxi_TranslateICMP(int type, int code)
2952 case ICMP_DEST_UNREACH:
2954 case ICMP_NET_UNREACH:
2955 return "Destination Net Unreachable";
2956 case ICMP_HOST_UNREACH:
2957 return "Destination Host Unreachable";
2958 case ICMP_PROT_UNREACH:
2959 return "Destination Protocol Unreachable";
2960 case ICMP_PORT_UNREACH:
2961 return "Destination Port Unreachable";
2963 return "Destination Net Prohibited";
2965 return "Destination Host Prohibited";
2971 #endif /* AFS_RXERRQ_ENV */
2974 * Get the last network error for a connection
2976 * A "network error" here means an error retrieved from ICMP, or some other
2977 * mechanism outside of Rx that informs us of errors in network reachability.
2979 * If a peer associated with the given Rx connection has received a network
2980 * error recently, this function allows the caller to know what error
2981 * specifically occurred. This can be useful to know, since e.g. ICMP errors
2982 * can cause calls to that peer to be quickly aborted. So, this function can
2983 * help see why a call was aborted due to network errors.
2985 * If we have received traffic from a peer since the last network error, we
2986 * treat that peer as if we had not received an network error for it.
2988 * @param[in] conn The Rx connection to examine
2989 * @param[out] err_origin The origin of the last network error (e.g. ICMP);
2990 * one of the RX_NETWORK_ERROR_ORIGIN_* constants
2991 * @param[out] err_type The type of the last error
2992 * @param[out] err_code The code of the last error
2993 * @param[out] msg Human-readable error message, if applicable; NULL otherwise
2995 * @return If we have an error
2996 * @retval -1 No error to get; 'out' params are undefined
2997 * @retval 0 We have an error; 'out' params contain the last error
3000 rx_GetNetworkError(struct rx_connection *conn, int *err_origin, int *err_type,
3001 int *err_code, const char **msg)
3003 #ifdef AFS_RXERRQ_ENV
3004 struct rx_peer *peer = conn->peer;
3005 if (rx_atomic_read(&peer->neterrs)) {
3006 MUTEX_ENTER(&peer->peer_lock);
3007 *err_origin = peer->last_err_origin;
3008 *err_type = peer->last_err_type;
3009 *err_code = peer->last_err_code;
3010 MUTEX_EXIT(&peer->peer_lock);
3013 if (*err_origin == RX_NETWORK_ERROR_ORIGIN_ICMP) {
3014 *msg = rxi_TranslateICMP(*err_type, *err_code);
3023 /* Find the peer process represented by the supplied (host,port)
3024 * combination. If there is no appropriate active peer structure, a
3025 * new one will be allocated and initialized
3028 rxi_FindPeer(afs_uint32 host, u_short port, int create)
3032 hashIndex = PEER_HASH(host, port);
3033 MUTEX_ENTER(&rx_peerHashTable_lock);
3034 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
3035 if ((pp->host == host) && (pp->port == port))
3040 pp = rxi_AllocPeer(); /* This bzero's *pp */
3041 pp->host = host; /* set here or in InitPeerParams is zero */
3043 #ifdef AFS_RXERRQ_ENV
3044 rx_atomic_set(&pp->neterrs, 0);
3046 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
3047 opr_queue_Init(&pp->rpcStats);
3048 pp->next = rx_peerHashTable[hashIndex];
3049 rx_peerHashTable[hashIndex] = pp;
3050 rxi_InitPeerParams(pp);
3051 if (rx_stats_active)
3052 rx_atomic_inc(&rx_stats.nPeerStructs);
3058 MUTEX_EXIT(&rx_peerHashTable_lock);
3063 /* Find the connection at (host, port) started at epoch, and with the
3064 * given connection id. Creates the server connection if necessary.
3065 * The type specifies whether a client connection or a server
3066 * connection is desired. In both cases, (host, port) specify the
3067 * peer's (host, pair) pair. Client connections are not made
3068 * automatically by this routine. The parameter socket gives the
3069 * socket descriptor on which the packet was received. This is used,
3070 * in the case of server connections, to check that *new* connections
3071 * come via a valid (port, serviceId). Finally, the securityIndex
3072 * parameter must match the existing index for the connection. If a
3073 * server connection is created, it will be created using the supplied
3074 * index, if the index is valid for this service */
3075 static struct rx_connection *
3076 rxi_FindConnection(osi_socket socket, afs_uint32 host,
3077 u_short port, u_short serviceId, afs_uint32 cid,
3078 afs_uint32 epoch, int type, u_int securityIndex,
3079 int *unknownService)
3081 int hashindex, flag, i;
3082 struct rx_connection *conn;
3083 *unknownService = 0;
3084 hashindex = CONN_HASH(host, port, cid, epoch, type);
3085 MUTEX_ENTER(&rx_connHashTable_lock);
3086 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
3087 rx_connHashTable[hashindex],
3090 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
3091 && (epoch == conn->epoch)) {
3092 struct rx_peer *pp = conn->peer;
3093 if (securityIndex != conn->securityIndex) {
3094 /* this isn't supposed to happen, but someone could forge a packet
3095 * like this, and there seems to be some CM bug that makes this
3096 * happen from time to time -- in which case, the fileserver
3098 MUTEX_EXIT(&rx_connHashTable_lock);
3099 return (struct rx_connection *)0;
3101 if (pp->host == host && pp->port == port)
3103 if (type == RX_CLIENT_CONNECTION && pp->port == port)
3105 /* So what happens when it's a callback connection? */
3106 if ( /*type == RX_CLIENT_CONNECTION && */
3107 (conn->epoch & 0x80000000))
3111 /* the connection rxLastConn that was used the last time is not the
3112 ** one we are looking for now. Hence, start searching in the hash */
3114 conn = rx_connHashTable[hashindex];
3119 struct rx_service *service;
3120 if (type == RX_CLIENT_CONNECTION) {
3121 MUTEX_EXIT(&rx_connHashTable_lock);
3122 return (struct rx_connection *)0;
3124 service = rxi_FindService(socket, serviceId);
3125 if (!service || (securityIndex >= service->nSecurityObjects)
3126 || (service->securityObjects[securityIndex] == 0)) {
3127 MUTEX_EXIT(&rx_connHashTable_lock);
3128 *unknownService = 1;
3129 return (struct rx_connection *)0;
3131 conn = rxi_AllocConnection(); /* This bzero's the connection */
3132 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
3133 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
3134 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
3135 conn->next = rx_connHashTable[hashindex];
3136 rx_connHashTable[hashindex] = conn;
3137 conn->peer = rxi_FindPeer(host, port, 1);
3138 conn->type = RX_SERVER_CONNECTION;
3139 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
3140 conn->epoch = epoch;
3141 conn->cid = cid & RX_CIDMASK;
3142 conn->ackRate = RX_FAST_ACK_RATE;
3143 conn->service = service;
3144 conn->serviceId = serviceId;
3145 conn->securityIndex = securityIndex;
3146 conn->securityObject = service->securityObjects[securityIndex];
3147 conn->nSpecific = 0;
3148 conn->specific = NULL;
3149 rx_SetConnDeadTime(conn, service->connDeadTime);
3150 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
3151 for (i = 0; i < RX_MAXCALLS; i++) {
3152 conn->twind[i] = rx_initSendWindow;
3153 conn->rwind[i] = rx_initReceiveWindow;
3155 /* Notify security object of the new connection */
3156 RXS_NewConnection(conn->securityObject, conn);
3157 /* XXXX Connection timeout? */
3158 if (service->newConnProc)
3159 (*service->newConnProc) (conn);
3160 if (rx_stats_active)
3161 rx_atomic_inc(&rx_stats.nServerConns);
3164 MUTEX_ENTER(&rx_refcnt_mutex);
3166 MUTEX_EXIT(&rx_refcnt_mutex);
3168 rxLastConn = conn; /* store this connection as the last conn used */
3169 MUTEX_EXIT(&rx_connHashTable_lock);
3174 * Abort the call if the server is over the busy threshold. This
3175 * can be used without requiring a call structure be initialised,
3176 * or connected to a particular channel
3179 rxi_AbortIfServerBusy(osi_socket socket, struct rx_connection *conn,
3180 struct rx_packet *np)
3182 if ((rx_BusyThreshold > 0) &&
3183 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3184 rxi_SendRawAbort(socket, conn->peer->host, conn->peer->port,
3185 rx_BusyError, np, 0);
3186 if (rx_stats_active)
3187 rx_atomic_inc(&rx_stats.nBusies);
3194 static_inline struct rx_call *
3195 rxi_ReceiveClientCall(struct rx_packet *np, struct rx_connection *conn)
3198 struct rx_call *call;
3200 channel = np->header.cid & RX_CHANNELMASK;
3201 MUTEX_ENTER(&conn->conn_call_lock);
3202 call = conn->call[channel];
3203 if (np->header.type == RX_PACKET_TYPE_BUSY) {
3204 conn->lastBusy[channel] = clock_Sec();
3206 if (!call || conn->callNumber[channel] != np->header.callNumber) {
3207 MUTEX_EXIT(&conn->conn_call_lock);
3208 if (rx_stats_active)
3209 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3213 MUTEX_ENTER(&call->lock);
3214 MUTEX_EXIT(&conn->conn_call_lock);
3216 if ((call->state == RX_STATE_DALLY)
3217 && np->header.type == RX_PACKET_TYPE_ACK) {
3218 if (rx_stats_active)
3219 rx_atomic_inc(&rx_stats.ignorePacketDally);
3220 MUTEX_EXIT(&call->lock);
3227 static_inline struct rx_call *
3228 rxi_ReceiveServerCall(osi_socket socket, struct rx_packet *np,
3229 struct rx_connection *conn)
3232 struct rx_call *call;
3234 channel = np->header.cid & RX_CHANNELMASK;
3235 MUTEX_ENTER(&conn->conn_call_lock);
3236 call = conn->call[channel];
3239 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3240 MUTEX_EXIT(&conn->conn_call_lock);
3244 call = rxi_NewCall(conn, channel); /* returns locked call */
3245 *call->callNumber = np->header.callNumber;
3246 MUTEX_EXIT(&conn->conn_call_lock);
3248 call->state = RX_STATE_PRECALL;
3249 clock_GetTime(&call->queueTime);
3250 call->app.bytesSent = 0;
3251 call->app.bytesRcvd = 0;
3252 rxi_KeepAliveOn(call);
3257 if (np->header.callNumber == conn->callNumber[channel]) {
3258 MUTEX_ENTER(&call->lock);
3259 MUTEX_EXIT(&conn->conn_call_lock);
3263 if (np->header.callNumber < conn->callNumber[channel]) {
3264 MUTEX_EXIT(&conn->conn_call_lock);
3265 if (rx_stats_active)
3266 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3270 MUTEX_ENTER(&call->lock);
3271 MUTEX_EXIT(&conn->conn_call_lock);
3273 /* Wait until the transmit queue is idle before deciding
3274 * whether to reset the current call. Chances are that the
3275 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3278 #ifdef RX_ENABLE_LOCKS
3279 if (call->state == RX_STATE_ACTIVE && !call->error) {
3280 rxi_WaitforTQBusy(call);
3281 /* If we entered error state while waiting,
3282 * must call rxi_CallError to permit rxi_ResetCall
3283 * to processed when the tqWaiter count hits zero.
3286 rxi_CallError(call, call->error);
3287 MUTEX_EXIT(&call->lock);
3291 #endif /* RX_ENABLE_LOCKS */
3292 /* If the new call cannot be taken right now send a busy and set
3293 * the error condition in this call, so that it terminates as
3294 * quickly as possible */
3295 if (call->state == RX_STATE_ACTIVE) {
3296 rxi_CallError(call, RX_CALL_DEAD);
3297 rxi_SendSpecial(call, conn, NULL, RX_PACKET_TYPE_BUSY,
3299 MUTEX_EXIT(&call->lock);
3303 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3304 MUTEX_EXIT(&call->lock);
3308 rxi_ResetCall(call, 0);
3309 /* The conn_call_lock is not held but no one else should be
3310 * using this call channel while we are processing this incoming
3311 * packet. This assignment should be safe.
3313 *call->callNumber = np->header.callNumber;
3314 call->state = RX_STATE_PRECALL;
3315 clock_GetTime(&call->queueTime);
3316 call->app.bytesSent = 0;
3317 call->app.bytesRcvd = 0;
3318 rxi_KeepAliveOn(call);
3324 /* There are two packet tracing routines available for testing and monitoring
3325 * Rx. One is called just after every packet is received and the other is
3326 * called just before every packet is sent. Received packets, have had their
3327 * headers decoded, and packets to be sent have not yet had their headers
3328 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3329 * containing the network address. Both can be modified. The return value, if
3330 * non-zero, indicates that the packet should be dropped. */
3332 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3333 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3335 /* A packet has been received off the interface. Np is the packet, socket is
3336 * the socket number it was received from (useful in determining which service
3337 * this packet corresponds to), and (host, port) reflect the host,port of the
3338 * sender. This call returns the packet to the caller if it is finished with
3339 * it, rather than de-allocating it, just as a small performance hack */
3342 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3343 afs_uint32 host, u_short port, int *tnop,
3344 struct rx_call **newcallp)
3346 struct rx_call *call;
3347 struct rx_connection *conn;
3349 int unknownService = 0;
3353 struct rx_packet *tnp;
3356 /* We don't print out the packet until now because (1) the time may not be
3357 * accurate enough until now in the lwp implementation (rx_Listener only gets
3358 * the time after the packet is read) and (2) from a protocol point of view,
3359 * this is the first time the packet has been seen */
3360 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3361 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3362 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3363 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3364 np->header.epoch, np->header.cid, np->header.callNumber,
3365 np->header.seq, np->header.flags, np));
3368 /* Account for connectionless packets */
3369 if (rx_stats_active &&
3370 ((np->header.type == RX_PACKET_TYPE_VERSION) ||
3371 (np->header.type == RX_PACKET_TYPE_DEBUG))) {
3372 struct rx_peer *peer;
3374 /* Try to look up the peer structure, but don't create one */
3375 peer = rxi_FindPeer(host, port, 0);
3377 /* Since this may not be associated with a connection, it may have
3378 * no refCount, meaning we could race with ReapConnections
3381 if (peer && (peer->refCount > 0)) {
3382 #ifdef AFS_RXERRQ_ENV
3383 if (rx_atomic_read(&peer->neterrs)) {
3384 rx_atomic_set(&peer->neterrs, 0);
3387 MUTEX_ENTER(&peer->peer_lock);
3388 peer->bytesReceived += np->length;
3389 MUTEX_EXIT(&peer->peer_lock);
3393 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3394 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3397 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3398 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3401 /* If an input tracer function is defined, call it with the packet and
3402 * network address. Note this function may modify its arguments. */
3403 if (rx_justReceived) {
3404 struct sockaddr_in addr;
3406 addr.sin_family = AF_INET;
3407 addr.sin_port = port;
3408 addr.sin_addr.s_addr = host;
3409 memset(&addr.sin_zero, 0, sizeof(addr.sin_zero));
3410 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3411 addr.sin_len = sizeof(addr);
3412 #endif /* AFS_OSF_ENV */
3413 drop = (*rx_justReceived) (np, &addr);
3414 /* drop packet if return value is non-zero */
3417 port = addr.sin_port; /* in case fcn changed addr */
3418 host = addr.sin_addr.s_addr;
3422 /* If packet was not sent by the client, then *we* must be the client */
3423 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3424 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3426 /* Find the connection (or fabricate one, if we're the server & if
3427 * necessary) associated with this packet */
3429 rxi_FindConnection(socket, host, port, np->header.serviceId,
3430 np->header.cid, np->header.epoch, type,
3431 np->header.securityIndex, &unknownService);
3433 /* To avoid having 2 connections just abort at each other,
3434 don't abort an abort. */
3436 if (unknownService && (np->header.type != RX_PACKET_TYPE_ABORT))
3437 rxi_SendRawAbort(socket, host, port, RX_INVALID_OPERATION,
3442 #ifdef AFS_RXERRQ_ENV
3443 if (rx_atomic_read(&conn->peer->neterrs)) {
3444 rx_atomic_set(&conn->peer->neterrs, 0);
3448 /* If we're doing statistics, then account for the incoming packet */
3449 if (rx_stats_active) {
3450 MUTEX_ENTER(&conn->peer->peer_lock);
3451 conn->peer->bytesReceived += np->length;
3452 MUTEX_EXIT(&conn->peer->peer_lock);
3455 /* If the connection is in an error state, send an abort packet and ignore
3456 * the incoming packet */
3458 /* Don't respond to an abort packet--we don't want loops! */
3459 MUTEX_ENTER(&conn->conn_data_lock);
3460 if (np->header.type != RX_PACKET_TYPE_ABORT)
3461 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3462 putConnection(conn);
3463 MUTEX_EXIT(&conn->conn_data_lock);
3467 /* Check for connection-only requests (i.e. not call specific). */
3468 if (np->header.callNumber == 0) {
3469 switch (np->header.type) {
3470 case RX_PACKET_TYPE_ABORT: {
3471 /* What if the supplied error is zero? */
3472 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3473 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3474 rxi_ConnectionError(conn, errcode);
3475 putConnection(conn);
3478 case RX_PACKET_TYPE_CHALLENGE:
3479 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3480 putConnection(conn);
3482 case RX_PACKET_TYPE_RESPONSE:
3483 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3484 putConnection(conn);
3486 case RX_PACKET_TYPE_PARAMS:
3487 case RX_PACKET_TYPE_PARAMS + 1:
3488 case RX_PACKET_TYPE_PARAMS + 2:
3489 /* ignore these packet types for now */
3490 putConnection(conn);
3494 /* Should not reach here, unless the peer is broken: send an
3496 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3497 MUTEX_ENTER(&conn->conn_data_lock);
3498 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3499 putConnection(conn);
3500 MUTEX_EXIT(&conn->conn_data_lock);
3505 if (type == RX_SERVER_CONNECTION)
3506 call = rxi_ReceiveServerCall(socket, np, conn);
3508 call = rxi_ReceiveClientCall(np, conn);
3511 putConnection(conn);
3515 MUTEX_ASSERT(&call->lock);
3516 /* Set remote user defined status from packet */
3517 call->remoteStatus = np->header.userStatus;
3519 /* Now do packet type-specific processing */
3520 switch (np->header.type) {
3521 case RX_PACKET_TYPE_DATA:
3522 /* If we're a client, and receiving a response, then all the packets
3523 * we transmitted packets are implicitly acknowledged. */
3524 if (type == RX_CLIENT_CONNECTION && !opr_queue_IsEmpty(&call->tq))
3525 rxi_AckAllInTransmitQueue(call);
3527 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3530 case RX_PACKET_TYPE_ACK:
3531 /* Respond immediately to ack packets requesting acknowledgement
3533 if (np->header.flags & RX_REQUEST_ACK) {
3535 (void)rxi_SendCallAbort(call, 0, 1, 0);
3537 (void)rxi_SendAck(call, 0, np->header.serial,
3538 RX_ACK_PING_RESPONSE, 1);
3540 np = rxi_ReceiveAckPacket(call, np, 1);
3542 case RX_PACKET_TYPE_ABORT: {
3543 /* An abort packet: reset the call, passing the error up to the user. */
3544 /* What if error is zero? */
3545 /* What if the error is -1? the application will treat it as a timeout. */
3546 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3547 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3548 rxi_CallError(call, errdata);
3549 MUTEX_EXIT(&call->lock);
3550 putConnection(conn);
3551 return np; /* xmitting; drop packet */
3553 case RX_PACKET_TYPE_BUSY:
3554 /* Mostly ignore BUSY packets. We will update lastReceiveTime below,
3555 * so we don't think the endpoint is completely dead, but otherwise
3556 * just act as if we never saw anything. If all we get are BUSY packets
3557 * back, then we will eventually error out with RX_CALL_TIMEOUT if the
3558 * connection is configured with idle/hard timeouts. */
3561 case RX_PACKET_TYPE_ACKALL:
3562 /* All packets acknowledged, so we can drop all packets previously
3563 * readied for sending */
3564 rxi_AckAllInTransmitQueue(call);
3567 /* Should not reach here, unless the peer is broken: send an abort
3569 rxi_CallError(call, RX_PROTOCOL_ERROR);
3570 np = rxi_SendCallAbort(call, np, 1, 0);
3573 /* Note when this last legitimate packet was received, for keep-alive
3574 * processing. Note, we delay getting the time until now in the hope that
3575 * the packet will be delivered to the user before any get time is required
3576 * (if not, then the time won't actually be re-evaluated here). */
3577 call->lastReceiveTime = clock_Sec();
3578 MUTEX_EXIT(&call->lock);
3579 putConnection(conn);
3583 /* return true if this is an "interesting" connection from the point of view
3584 of someone trying to debug the system */
3586 rxi_IsConnInteresting(struct rx_connection *aconn)
3589 struct rx_call *tcall;
3591 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3594 for (i = 0; i < RX_MAXCALLS; i++) {
3595 tcall = aconn->call[i];
3597 if ((tcall->state == RX_STATE_PRECALL)
3598 || (tcall->state == RX_STATE_ACTIVE))
3600 if ((tcall->app.mode == RX_MODE_SENDING)
3601 || (tcall->app.mode == RX_MODE_RECEIVING))
3609 /* if this is one of the last few packets AND it wouldn't be used by the
3610 receiving call to immediately satisfy a read request, then drop it on
3611 the floor, since accepting it might prevent a lock-holding thread from
3612 making progress in its reading. If a call has been cleared while in
3613 the precall state then ignore all subsequent packets until the call
3614 is assigned to a thread. */
3617 TooLow(struct rx_packet *ap, struct rx_call *acall)
3621 MUTEX_ENTER(&rx_quota_mutex);
3622 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3623 && (acall->state == RX_STATE_PRECALL))
3624 || ((rx_nFreePackets < rxi_dataQuota + 2)
3625 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3626 && (acall->flags & RX_CALL_READER_WAIT)))) {
3629 MUTEX_EXIT(&rx_quota_mutex);
3635 * Clear the attach wait flag on a connection and proceed.
3637 * Any processing waiting for a connection to be attached should be
3638 * unblocked. We clear the flag and do any other needed tasks.
3641 * the conn to unmark waiting for attach
3643 * @pre conn's conn_data_lock must be locked before calling this function
3647 rxi_ConnClearAttachWait(struct rx_connection *conn)
3649 /* Indicate that rxi_CheckReachEvent is no longer running by
3650 * clearing the flag. Must be atomic under conn_data_lock to
3651 * avoid a new call slipping by: rxi_CheckConnReach holds
3652 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3654 conn->flags &= ~RX_CONN_ATTACHWAIT;
3655 if (conn->flags & RX_CONN_NAT_PING) {
3656 conn->flags &= ~RX_CONN_NAT_PING;
3657 rxi_ScheduleNatKeepAliveEvent(conn);
3662 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3664 struct rx_connection *conn = arg1;
3665 struct rx_call *acall = arg2;
3666 struct rx_call *call = acall;
3667 struct clock when, now;
3670 MUTEX_ENTER(&conn->conn_data_lock);
3673 rxevent_Put(&conn->checkReachEvent);
3675 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3677 putConnection(conn);
3679 MUTEX_EXIT(&conn->conn_data_lock);
3683 MUTEX_ENTER(&conn->conn_call_lock);
3684 MUTEX_ENTER(&conn->conn_data_lock);
3685 for (i = 0; i < RX_MAXCALLS; i++) {
3686 struct rx_call *tc = conn->call[i];
3687 if (tc && tc->state == RX_STATE_PRECALL) {
3693 rxi_ConnClearAttachWait(conn);
3694 MUTEX_EXIT(&conn->conn_data_lock);
3695 MUTEX_EXIT(&conn->conn_call_lock);
3700 MUTEX_ENTER(&call->lock);
3701 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3703 MUTEX_EXIT(&call->lock);
3705 clock_GetTime(&now);
3707 when.sec += RX_CHECKREACH_TIMEOUT;
3708 MUTEX_ENTER(&conn->conn_data_lock);
3709 if (!conn->checkReachEvent) {
3710 MUTEX_ENTER(&rx_refcnt_mutex);
3712 MUTEX_EXIT(&rx_refcnt_mutex);
3713 conn->checkReachEvent = rxevent_Post(&when, &now,
3714 rxi_CheckReachEvent, conn,
3717 MUTEX_EXIT(&conn->conn_data_lock);
3723 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3725 struct rx_service *service = conn->service;
3726 struct rx_peer *peer = conn->peer;
3727 afs_uint32 now, lastReach;
3729 if (service->checkReach == 0)
3733 MUTEX_ENTER(&peer->peer_lock);
3734 lastReach = peer->lastReachTime;
3735 MUTEX_EXIT(&peer->peer_lock);
3736 if (now - lastReach < RX_CHECKREACH_TTL)
3739 MUTEX_ENTER(&conn->conn_data_lock);
3740 if (conn->flags & RX_CONN_ATTACHWAIT) {
3741 MUTEX_EXIT(&conn->conn_data_lock);
3744 conn->flags |= RX_CONN_ATTACHWAIT;
3745 MUTEX_EXIT(&conn->conn_data_lock);
3746 if (!conn->checkReachEvent)
3747 rxi_CheckReachEvent(NULL, conn, call, 0);
3752 /* try to attach call, if authentication is complete */
3754 TryAttach(struct rx_call *acall, osi_socket socket,
3755 int *tnop, struct rx_call **newcallp,
3758 struct rx_connection *conn = acall->conn;
3760 if (conn->type == RX_SERVER_CONNECTION
3761 && acall->state == RX_STATE_PRECALL) {
3762 /* Don't attach until we have any req'd. authentication. */
3763 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3764 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3765 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3766 /* Note: this does not necessarily succeed; there
3767 * may not any proc available
3770 rxi_ChallengeOn(acall->conn);
3775 /* A data packet has been received off the interface. This packet is
3776 * appropriate to the call (the call is in the right state, etc.). This
3777 * routine can return a packet to the caller, for re-use */
3779 static struct rx_packet *
3780 rxi_ReceiveDataPacket(struct rx_call *call,
3781 struct rx_packet *np, int istack,
3782 osi_socket socket, afs_uint32 host, u_short port,
3783 int *tnop, struct rx_call **newcallp)
3785 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3790 afs_uint32 serial=0, flags=0;
3792 struct rx_packet *tnp;
3793 if (rx_stats_active)
3794 rx_atomic_inc(&rx_stats.dataPacketsRead);
3797 /* If there are no packet buffers, drop this new packet, unless we can find
3798 * packet buffers from inactive calls */
3800 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3801 MUTEX_ENTER(&rx_freePktQ_lock);
3802 rxi_NeedMorePackets = TRUE;
3803 MUTEX_EXIT(&rx_freePktQ_lock);
3804 if (rx_stats_active)
3805 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3806 rxi_calltrace(RX_TRACE_DROP, call);
3807 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3808 /* We used to clear the receive queue here, in an attempt to free
3809 * packets. However this is unsafe if the queue has received a
3810 * soft ACK for the final packet */
3811 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3817 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3818 * packet is one of several packets transmitted as a single
3819 * datagram. Do not send any soft or hard acks until all packets
3820 * in a jumbogram have been processed. Send negative acks right away.
3822 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3823 /* tnp is non-null when there are more packets in the
3824 * current jumbo gram */
3831 seq = np->header.seq;
3832 serial = np->header.serial;
3833 flags = np->header.flags;
3835 /* If the call is in an error state, send an abort message */
3837 return rxi_SendCallAbort(call, np, istack, 0);
3839 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3840 * AFS 3.5 jumbogram. */
3841 if (flags & RX_JUMBO_PACKET) {
3842 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3847 if (np->header.spare != 0) {
3848 MUTEX_ENTER(&call->conn->conn_data_lock);
3849 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3850 MUTEX_EXIT(&call->conn->conn_data_lock);
3853 /* The usual case is that this is the expected next packet */
3854 if (seq == call->rnext) {
3856 /* Check to make sure it is not a duplicate of one already queued */
3857 if (!opr_queue_IsEmpty(&call->rq)
3858 && opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq == seq) {
3859 if (rx_stats_active)
3860 rx_atomic_inc(&rx_stats.dupPacketsRead);
3861 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3862 rxi_CancelDelayedAckEvent(call);
3863 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3869 /* It's the next packet. Stick it on the receive queue
3870 * for this call. Set newPackets to make sure we wake
3871 * the reader once all packets have been processed */
3872 #ifdef RX_TRACK_PACKETS
3873 np->flags |= RX_PKTFLAG_RQ;
3875 opr_queue_Prepend(&call->rq, &np->entry);
3876 #ifdef RXDEBUG_PACKET
3878 #endif /* RXDEBUG_PACKET */
3880 np = NULL; /* We can't use this anymore */
3883 /* If an ack is requested then set a flag to make sure we
3884 * send an acknowledgement for this packet */
3885 if (flags & RX_REQUEST_ACK) {
3886 ackNeeded = RX_ACK_REQUESTED;
3889 /* Keep track of whether we have received the last packet */
3890 if (flags & RX_LAST_PACKET) {
3891 call->flags |= RX_CALL_HAVE_LAST;
3895 /* Check whether we have all of the packets for this call */
3896 if (call->flags & RX_CALL_HAVE_LAST) {
3897 afs_uint32 tseq; /* temporary sequence number */
3898 struct opr_queue *cursor;
3900 for (tseq = seq, opr_queue_Scan(&call->rq, cursor)) {
3901 struct rx_packet *tp;
3903 tp = opr_queue_Entry(cursor, struct rx_packet, entry);
3904 if (tseq != tp->header.seq)
3906 if (tp->header.flags & RX_LAST_PACKET) {
3907 call->flags |= RX_CALL_RECEIVE_DONE;
3914 /* Provide asynchronous notification for those who want it
3915 * (e.g. multi rx) */
3916 if (call->arrivalProc) {
3917 (*call->arrivalProc) (call, call->arrivalProcHandle,
3918 call->arrivalProcArg);
3919 call->arrivalProc = (void (*)())0;
3922 /* Update last packet received */
3925 /* If there is no server process serving this call, grab
3926 * one, if available. We only need to do this once. If a
3927 * server thread is available, this thread becomes a server
3928 * thread and the server thread becomes a listener thread. */
3930 TryAttach(call, socket, tnop, newcallp, 0);
3933 /* This is not the expected next packet. */
3935 /* Determine whether this is a new or old packet, and if it's
3936 * a new one, whether it fits into the current receive window.
3937 * Also figure out whether the packet was delivered in sequence.
3938 * We use the prev variable to determine whether the new packet
3939 * is the successor of its immediate predecessor in the
3940 * receive queue, and the missing flag to determine whether
3941 * any of this packets predecessors are missing. */
3943 afs_uint32 prev; /* "Previous packet" sequence number */
3944 struct opr_queue *cursor;
3945 int missing; /* Are any predecessors missing? */
3947 /* If the new packet's sequence number has been sent to the
3948 * application already, then this is a duplicate */
3949 if (seq < call->rnext) {
3950 if (rx_stats_active)
3951 rx_atomic_inc(&rx_stats.dupPacketsRead);
3952 rxi_CancelDelayedAckEvent(call);
3953 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3959 /* If the sequence number is greater than what can be
3960 * accomodated by the current window, then send a negative
3961 * acknowledge and drop the packet */
3962 if ((call->rnext + call->rwind) <= seq) {
3963 rxi_CancelDelayedAckEvent(call);
3964 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3971 /* Look for the packet in the queue of old received packets */
3972 prev = call->rnext - 1;
3974 for (opr_queue_Scan(&call->rq, cursor)) {
3975 struct rx_packet *tp
3976 = opr_queue_Entry(cursor, struct rx_packet, entry);
3978 /*Check for duplicate packet */
3979 if (seq == tp->header.seq) {
3980 if (rx_stats_active)
3981 rx_atomic_inc(&rx_stats.dupPacketsRead);
3982 rxi_CancelDelayedAckEvent(call);
3983 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3989 /* If we find a higher sequence packet, break out and
3990 * insert the new packet here. */
3991 if (seq < tp->header.seq)
3993 /* Check for missing packet */
3994 if (tp->header.seq != prev + 1) {
3998 prev = tp->header.seq;
4001 /* Keep track of whether we have received the last packet. */
4002 if (flags & RX_LAST_PACKET) {
4003 call->flags |= RX_CALL_HAVE_LAST;
4006 /* It's within the window: add it to the the receive queue.
4007 * tp is left by the previous loop either pointing at the
4008 * packet before which to insert the new packet, or at the
4009 * queue head if the queue is empty or the packet should be
4011 #ifdef RX_TRACK_PACKETS
4012 np->flags |= RX_PKTFLAG_RQ;
4014 #ifdef RXDEBUG_PACKET
4016 #endif /* RXDEBUG_PACKET */
4017 opr_queue_InsertBefore(cursor, &np->entry);
4021 /* Check whether we have all of the packets for this call */
4022 if ((call->flags & RX_CALL_HAVE_LAST)
4023 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4024 afs_uint32 tseq; /* temporary sequence number */
4027 for (opr_queue_Scan(&call->rq, cursor)) {
4028 struct rx_packet *tp
4029 = opr_queue_Entry(cursor, struct rx_packet, entry);
4030 if (tseq != tp->header.seq)
4032 if (tp->header.flags & RX_LAST_PACKET) {
4033 call->flags |= RX_CALL_RECEIVE_DONE;
4040 /* We need to send an ack of the packet is out of sequence,
4041 * or if an ack was requested by the peer. */
4042 if (seq != prev + 1 || missing) {
4043 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4044 } else if (flags & RX_REQUEST_ACK) {
4045 ackNeeded = RX_ACK_REQUESTED;
4048 /* Acknowledge the last packet for each call */
4049 if (flags & RX_LAST_PACKET) {
4060 * If the receiver is waiting for an iovec, fill the iovec
4061 * using the data from the receive queue */
4062 if (call->flags & RX_CALL_IOVEC_WAIT) {
4063 didHardAck = rxi_FillReadVec(call, serial);
4064 /* the call may have been aborted */
4073 /* Wakeup the reader if any */
4074 if ((call->flags & RX_CALL_READER_WAIT)
4075 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4076 || (call->iovNext >= call->iovMax)
4077 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4078 call->flags &= ~RX_CALL_READER_WAIT;
4079 #ifdef RX_ENABLE_LOCKS
4080 CV_BROADCAST(&call->cv_rq);
4082 osi_rxWakeup(&call->rq);
4088 * Send an ack when requested by the peer, or once every
4089 * rxi_SoftAckRate packets until the last packet has been
4090 * received. Always send a soft ack for the last packet in
4091 * the server's reply. */
4093 rxi_CancelDelayedAckEvent(call);
4094 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4095 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4096 rxi_CancelDelayedAckEvent(call);
4097 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4098 } else if (call->nSoftAcks) {
4099 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4100 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4102 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4103 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4104 rxi_CancelDelayedAckEvent(call);
4111 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4113 struct rx_peer *peer = conn->peer;
4115 MUTEX_ENTER(&peer->peer_lock);
4116 peer->lastReachTime = clock_Sec();
4117 MUTEX_EXIT(&peer->peer_lock);
4119 MUTEX_ENTER(&conn->conn_data_lock);
4120 if (conn->flags & RX_CONN_ATTACHWAIT) {
4123 rxi_ConnClearAttachWait(conn);
4124 MUTEX_EXIT(&conn->conn_data_lock);
4126 for (i = 0; i < RX_MAXCALLS; i++) {
4127 struct rx_call *call = conn->call[i];
4130 MUTEX_ENTER(&call->lock);
4131 /* tnop can be null if newcallp is null */
4132 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4134 MUTEX_EXIT(&call->lock);
4138 MUTEX_EXIT(&conn->conn_data_lock);
4141 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4143 rx_ack_reason(int reason)
4146 case RX_ACK_REQUESTED:
4148 case RX_ACK_DUPLICATE:
4150 case RX_ACK_OUT_OF_SEQUENCE:
4152 case RX_ACK_EXCEEDS_WINDOW:
4154 case RX_ACK_NOSPACE:
4158 case RX_ACK_PING_RESPONSE:
4171 /* The real smarts of the whole thing. */
4172 static struct rx_packet *
4173 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4176 struct rx_ackPacket *ap;
4178 struct rx_packet *tp;
4179 struct rx_connection *conn = call->conn;
4180 struct rx_peer *peer = conn->peer;
4181 struct opr_queue *cursor;
4182 struct clock now; /* Current time, for RTT calculations */
4190 int newAckCount = 0;
4191 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4192 int pktsize = 0; /* Set if we need to update the peer mtu */
4193 int conn_data_locked = 0;
4195 if (rx_stats_active)
4196 rx_atomic_inc(&rx_stats.ackPacketsRead);
4197 ap = (struct rx_ackPacket *)rx_DataOf(np);
4198 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4200 return np; /* truncated ack packet */
4202 /* depends on ack packet struct */
4203 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4204 first = ntohl(ap->firstPacket);
4205 prev = ntohl(ap->previousPacket);
4206 serial = ntohl(ap->serial);
4209 * Ignore ack packets received out of order while protecting
4210 * against peers that set the previousPacket field to a packet
4211 * serial number instead of a sequence number.
4213 if (first < call->tfirst ||
4214 (first == call->tfirst && prev < call->tprev && prev < call->tfirst
4221 if (np->header.flags & RX_SLOW_START_OK) {
4222 call->flags |= RX_CALL_SLOW_START_OK;
4225 if (ap->reason == RX_ACK_PING_RESPONSE)
4226 rxi_UpdatePeerReach(conn, call);
4228 if (conn->lastPacketSizeSeq) {
4229 MUTEX_ENTER(&conn->conn_data_lock);
4230 conn_data_locked = 1;
4231 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4232 pktsize = conn->lastPacketSize;
4233 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4236 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4237 if (!conn_data_locked) {
4238 MUTEX_ENTER(&conn->conn_data_lock);
4239 conn_data_locked = 1;
4241 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4242 /* process mtu ping ack */
4243 pktsize = conn->lastPingSize;
4244 conn->lastPingSizeSer = conn->lastPingSize = 0;
4248 if (conn_data_locked) {
4249 MUTEX_EXIT(&conn->conn_data_lock);
4250 conn_data_locked = 0;
4254 if (rxdebug_active) {
4258 len = _snprintf(msg, sizeof(msg),
4259 "tid[%d] RACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4260 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4261 ntohl(ap->serial), ntohl(ap->previousPacket),
4262 (unsigned int)np->header.seq, ntohl(ap->firstPacket),
4263 ap->nAcks, ntohs(ap->bufferSpace) );
4267 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4268 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4272 OutputDebugString(msg);
4274 #else /* AFS_NT40_ENV */
4277 "RACK: reason %x previous %u seq %u serial %u first %u",
4278 ap->reason, ntohl(ap->previousPacket),
4279 (unsigned int)np->header.seq, (unsigned int)serial,
4280 ntohl(ap->firstPacket));
4283 for (offset = 0; offset < nAcks; offset++)
4284 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4289 #endif /* AFS_NT40_ENV */
4292 MUTEX_ENTER(&peer->peer_lock);
4295 * Start somewhere. Can't assume we can send what we can receive,
4296 * but we are clearly receiving.
4298 if (!peer->maxPacketSize)
4299 peer->maxPacketSize = RX_MIN_PACKET_SIZE - RX_HEADER_SIZE;
4301 if (pktsize > peer->maxPacketSize) {
4302 peer->maxPacketSize = pktsize;
4303 if ((pktsize + RX_HEADER_SIZE > peer->ifMTU)) {
4304 peer->ifMTU = pktsize + RX_HEADER_SIZE;
4305 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4306 rxi_ScheduleGrowMTUEvent(call, 1);
4311 clock_GetTime(&now);
4313 /* The transmit queue splits into 4 sections.
4315 * The first section is packets which have now been acknowledged
4316 * by a window size change in the ack. These have reached the
4317 * application layer, and may be discarded. These are packets
4318 * with sequence numbers < ap->firstPacket.
4320 * The second section is packets which have sequence numbers in
4321 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4322 * contents of the packet's ack array determines whether these
4323 * packets are acknowledged or not.
4325 * The third section is packets which fall above the range
4326 * addressed in the ack packet. These have not yet been received
4329 * The four section is packets which have not yet been transmitted.
4330 * These packets will have a header.serial of 0.
4333 /* First section - implicitly acknowledged packets that can be
4337 tp = opr_queue_First(&call->tq, struct rx_packet, entry);
4338 while(!opr_queue_IsEnd(&call->tq, &tp->entry) && tp->header.seq < first) {
4339 struct rx_packet *next;
4341 next = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4342 call->tfirst = tp->header.seq + 1;
4344 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4346 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4349 #ifdef RX_ENABLE_LOCKS
4350 /* XXX Hack. Because we have to release the global call lock when sending
4351 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4352 * in rxi_Start sending packets out because packets may move to the
4353 * freePacketQueue as result of being here! So we drop these packets until
4354 * we're safely out of the traversing. Really ugly!
4355 * To make it even uglier, if we're using fine grain locking, we can
4356 * set the ack bits in the packets and have rxi_Start remove the packets
4357 * when it's done transmitting.
4359 if (call->flags & RX_CALL_TQ_BUSY) {
4360 tp->flags |= RX_PKTFLAG_ACKED;
4361 call->flags |= RX_CALL_TQ_SOME_ACKED;
4363 #endif /* RX_ENABLE_LOCKS */
4365 opr_queue_Remove(&tp->entry);
4366 #ifdef RX_TRACK_PACKETS
4367 tp->flags &= ~RX_PKTFLAG_TQ;
4369 #ifdef RXDEBUG_PACKET
4371 #endif /* RXDEBUG_PACKET */
4372 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4377 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4379 /* Second section of the queue - packets for which we are receiving
4382 * Go through the explicit acks/nacks and record the results in
4383 * the waiting packets. These are packets that can't be released
4384 * yet, even with a positive acknowledge. This positive
4385 * acknowledge only means the packet has been received by the
4386 * peer, not that it will be retained long enough to be sent to
4387 * the peer's upper level. In addition, reset the transmit timers
4388 * of any missing packets (those packets that must be missing
4389 * because this packet was out of sequence) */
4391 call->nSoftAcked = 0;
4393 while (!opr_queue_IsEnd(&call->tq, &tp->entry)
4394 && tp->header.seq < first + nAcks) {
4395 /* Set the acknowledge flag per packet based on the
4396 * information in the ack packet. An acknowlegded packet can
4397 * be downgraded when the server has discarded a packet it
4398 * soacked previously, or when an ack packet is received
4399 * out of sequence. */
4400 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4401 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4403 tp->flags |= RX_PKTFLAG_ACKED;
4404 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4411 } else /* RX_ACK_TYPE_NACK */ {
4412 tp->flags &= ~RX_PKTFLAG_ACKED;
4416 tp = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4419 /* We don't need to take any action with the 3rd or 4th section in the
4420 * queue - they're not addressed by the contents of this ACK packet.
4423 /* if the ack packet has a receivelen field hanging off it,
4424 * update our state */
4425 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4428 /* If the ack packet has a "recommended" size that is less than
4429 * what I am using now, reduce my size to match */
4430 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4431 (int)sizeof(afs_int32), &tSize);
4432 tSize = (afs_uint32) ntohl(tSize);
4433 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4435 /* Get the maximum packet size to send to this peer */
4436 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4438 tSize = (afs_uint32) ntohl(tSize);
4439 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4440 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4442 /* sanity check - peer might have restarted with different params.
4443 * If peer says "send less", dammit, send less... Peer should never
4444 * be unable to accept packets of the size that prior AFS versions would
4445 * send without asking. */
4446 if (peer->maxMTU != tSize) {
4447 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4449 peer->maxMTU = tSize;
4450 peer->MTU = MIN(tSize, peer->MTU);
4451 call->MTU = MIN(call->MTU, tSize);
4454 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4457 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4458 (int)sizeof(afs_int32), &tSize);
4459 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4460 if (tSize < call->twind) { /* smaller than our send */
4461 call->twind = tSize; /* window, we must send less... */
4462 call->ssthresh = MIN(call->twind, call->ssthresh);
4463 call->conn->twind[call->channel] = call->twind;
4466 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4467 * network MTU confused with the loopback MTU. Calculate the
4468 * maximum MTU here for use in the slow start code below.
4470 /* Did peer restart with older RX version? */
4471 if (peer->maxDgramPackets > 1) {
4472 peer->maxDgramPackets = 1;
4474 } else if (np->length >=
4475 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4478 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4479 sizeof(afs_int32), &tSize);
4480 tSize = (afs_uint32) ntohl(tSize);
4482 * As of AFS 3.5 we set the send window to match the receive window.
4484 if (tSize < call->twind) {
4485 call->twind = tSize;
4486 call->conn->twind[call->channel] = call->twind;
4487 call->ssthresh = MIN(call->twind, call->ssthresh);
4488 } else if (tSize > call->twind) {
4489 call->twind = tSize;
4490 call->conn->twind[call->channel] = call->twind;
4494 * As of AFS 3.5, a jumbogram is more than one fixed size
4495 * packet transmitted in a single UDP datagram. If the remote
4496 * MTU is smaller than our local MTU then never send a datagram
4497 * larger than the natural MTU.
4500 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4501 (int)sizeof(afs_int32), &tSize);
4502 maxDgramPackets = (afs_uint32) ntohl(tSize);
4503 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4505 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4506 if (maxDgramPackets > 1) {
4507 peer->maxDgramPackets = maxDgramPackets;
4508 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4510 peer->maxDgramPackets = 1;
4511 call->MTU = peer->natMTU;
4513 } else if (peer->maxDgramPackets > 1) {
4514 /* Restarted with lower version of RX */
4515 peer->maxDgramPackets = 1;
4517 } else if (peer->maxDgramPackets > 1
4518 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4519 /* Restarted with lower version of RX */
4520 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4521 peer->natMTU = OLD_MAX_PACKET_SIZE;
4522 peer->MTU = OLD_MAX_PACKET_SIZE;
4523 peer->maxDgramPackets = 1;
4524 peer->nDgramPackets = 1;
4526 call->MTU = OLD_MAX_PACKET_SIZE;
4529 /* If the window has been extended by this acknowledge packet,
4530 * then wakeup a sender waiting in alloc for window space, or try
4531 * sending packets now, if he's been sitting on packets due to
4532 * lack of window space */
4533 if (call->tnext < (call->tfirst + call->twind)) {
4534 #ifdef RX_ENABLE_LOCKS
4535 CV_SIGNAL(&call->cv_twind);
4537 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4538 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4539 osi_rxWakeup(&call->twind);
4542 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4543 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4549 * Calculate how many datagrams were successfully received after
4550 * the first missing packet and adjust the negative ack counter
4555 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4556 if (call->nNacks < nNacked) {
4557 call->nNacks = nNacked;
4560 call->nAcks += newAckCount;
4564 /* If the packet contained new acknowledgements, rather than just
4565 * being a duplicate of one we have previously seen, then we can restart
4568 if (newAckCount > 0)
4569 rxi_rto_packet_acked(call, istack);
4571 if (call->flags & RX_CALL_FAST_RECOVER) {
4572 if (newAckCount == 0) {
4573 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4575 call->flags &= ~RX_CALL_FAST_RECOVER;
4576 call->cwind = call->nextCwind;
4577 call->nextCwind = 0;
4580 call->nCwindAcks = 0;
4581 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4582 /* Three negative acks in a row trigger congestion recovery */
4583 call->flags |= RX_CALL_FAST_RECOVER;
4584 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4586 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4587 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4588 call->nextCwind = call->ssthresh;
4591 peer->MTU = call->MTU;
4592 peer->cwind = call->nextCwind;
4593 peer->nDgramPackets = call->nDgramPackets;
4595 call->congestSeq = peer->congestSeq;
4597 /* Reset the resend times on the packets that were nacked
4598 * so we will retransmit as soon as the window permits
4602 for (opr_queue_ScanBackwards(&call->tq, cursor)) {
4603 struct rx_packet *tp =
4604 opr_queue_Entry(cursor, struct rx_packet, entry);
4606 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4607 tp->flags &= ~RX_PKTFLAG_SENT;
4609 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4614 /* If cwind is smaller than ssthresh, then increase
4615 * the window one packet for each ack we receive (exponential
4617 * If cwind is greater than or equal to ssthresh then increase
4618 * the congestion window by one packet for each cwind acks we
4619 * receive (linear growth). */
4620 if (call->cwind < call->ssthresh) {
4622 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4623 call->nCwindAcks = 0;
4625 call->nCwindAcks += newAckCount;
4626 if (call->nCwindAcks >= call->cwind) {
4627 call->nCwindAcks = 0;
4628 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4632 * If we have received several acknowledgements in a row then
4633 * it is time to increase the size of our datagrams
4635 if ((int)call->nAcks > rx_nDgramThreshold) {
4636 if (peer->maxDgramPackets > 1) {
4637 if (call->nDgramPackets < peer->maxDgramPackets) {
4638 call->nDgramPackets++;
4640 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4641 } else if (call->MTU < peer->maxMTU) {
4642 /* don't upgrade if we can't handle it */
4643 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4644 call->MTU = peer->ifMTU;
4646 call->MTU += peer->natMTU;
4647 call->MTU = MIN(call->MTU, peer->maxMTU);
4654 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4656 /* Servers need to hold the call until all response packets have
4657 * been acknowledged. Soft acks are good enough since clients
4658 * are not allowed to clear their receive queues. */
4659 if (call->state == RX_STATE_HOLD
4660 && call->tfirst + call->nSoftAcked >= call->tnext) {
4661 call->state = RX_STATE_DALLY;
4662 rxi_ClearTransmitQueue(call, 0);
4663 rxi_CancelKeepAliveEvent(call);
4664 } else if (!opr_queue_IsEmpty(&call->tq)) {
4665 rxi_Start(call, istack);
4671 * Schedule a connection abort to be sent after some delay.
4673 * @param[in] conn The connection to send the abort on.
4674 * @param[in] msec The number of milliseconds to wait before sending.
4676 * @pre conn_data_lock must be held
4679 rxi_SendConnectionAbortLater(struct rx_connection *conn, int msec)
4681 struct clock when, now;
4685 if (!conn->delayedAbortEvent) {
4686 clock_GetTime(&now);
4688 clock_Addmsec(&when, msec);
4689 conn->delayedAbortEvent =
4690 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
4694 /* Received a response to a challenge packet */
4695 static struct rx_packet *
4696 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4697 struct rx_packet *np, int istack)
4701 /* Ignore the packet if we're the client */
4702 if (conn->type == RX_CLIENT_CONNECTION)
4705 /* If already authenticated, ignore the packet (it's probably a retry) */
4706 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4709 if (!conn->securityChallengeSent) {
4710 /* We've never sent out a challenge for this connection, so this
4711 * response cannot possibly be correct; ignore it. This can happen
4712 * if we sent a challenge to the client, then we were restarted, and
4713 * then the client sent us a response. If we ignore the response, the
4714 * client will eventually resend a data packet, causing us to send a
4715 * new challenge and the client to send a new response. */
4719 /* Otherwise, have the security object evaluate the response packet */
4720 error = RXS_CheckResponse(conn->securityObject, conn, np);
4722 /* If the response is invalid, reset the connection, sending
4723 * an abort to the peer. Send the abort with a 1 second delay,
4724 * to avoid a peer hammering us by constantly recreating a
4725 * connection with bad credentials. */
4726 rxi_ConnectionError(conn, error);
4727 MUTEX_ENTER(&conn->conn_data_lock);
4728 rxi_SendConnectionAbortLater(conn, 1000);
4729 MUTEX_EXIT(&conn->conn_data_lock);
4732 /* If the response is valid, any calls waiting to attach
4733 * servers can now do so */
4736 for (i = 0; i < RX_MAXCALLS; i++) {
4737 struct rx_call *call = conn->call[i];
4739 MUTEX_ENTER(&call->lock);
4740 if (call->state == RX_STATE_PRECALL)
4741 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4742 /* tnop can be null if newcallp is null */
4743 MUTEX_EXIT(&call->lock);
4747 /* Update the peer reachability information, just in case
4748 * some calls went into attach-wait while we were waiting
4749 * for authentication..
4751 rxi_UpdatePeerReach(conn, NULL);
4756 /* A client has received an authentication challenge: the security
4757 * object is asked to cough up a respectable response packet to send
4758 * back to the server. The server is responsible for retrying the
4759 * challenge if it fails to get a response. */
4761 static struct rx_packet *
4762 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4763 struct rx_packet *np, int istack)
4767 /* Ignore the challenge if we're the server */
4768 if (conn->type == RX_SERVER_CONNECTION)
4771 /* Ignore the challenge if the connection is otherwise idle; someone's
4772 * trying to use us as an oracle. */
4773 if (!rxi_HasActiveCalls(conn))
4776 /* Send the security object the challenge packet. It is expected to fill
4777 * in the response. */
4778 error = RXS_GetResponse(conn->securityObject, conn, np);
4780 /* If the security object is unable to return a valid response, reset the
4781 * connection and send an abort to the peer. Otherwise send the response
4782 * packet to the peer connection. */
4784 rxi_ConnectionError(conn, error);
4785 MUTEX_ENTER(&conn->conn_data_lock);
4786 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4787 MUTEX_EXIT(&conn->conn_data_lock);
4789 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4790 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4796 /* Find an available server process to service the current request in
4797 * the given call structure. If one isn't available, queue up this
4798 * call so it eventually gets one */
4800 rxi_AttachServerProc(struct rx_call *call,
4801 osi_socket socket, int *tnop,
4802 struct rx_call **newcallp)
4804 struct rx_serverQueueEntry *sq;
4805 struct rx_service *service = call->conn->service;
4808 /* May already be attached */
4809 if (call->state == RX_STATE_ACTIVE)
4812 MUTEX_ENTER(&rx_serverPool_lock);
4814 haveQuota = QuotaOK(service);
4815 if ((!haveQuota) || opr_queue_IsEmpty(&rx_idleServerQueue)) {
4816 /* If there are no processes available to service this call,
4817 * put the call on the incoming call queue (unless it's
4818 * already on the queue).
4820 #ifdef RX_ENABLE_LOCKS
4822 ReturnToServerPool(service);
4823 #endif /* RX_ENABLE_LOCKS */
4825 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4826 call->flags |= RX_CALL_WAIT_PROC;
4827 rx_atomic_inc(&rx_nWaiting);
4828 rx_atomic_inc(&rx_nWaited);
4829 rxi_calltrace(RX_CALL_ARRIVAL, call);
4830 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4831 opr_queue_Append(&rx_incomingCallQueue, &call->entry);
4834 sq = opr_queue_Last(&rx_idleServerQueue,
4835 struct rx_serverQueueEntry, entry);
4837 /* If hot threads are enabled, and both newcallp and sq->socketp
4838 * are non-null, then this thread will process the call, and the
4839 * idle server thread will start listening on this threads socket.
4841 opr_queue_Remove(&sq->entry);
4843 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4846 *sq->socketp = socket;
4847 clock_GetTime(&call->startTime);
4848 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4852 if (call->flags & RX_CALL_WAIT_PROC) {
4853 /* Conservative: I don't think this should happen */
4854 call->flags &= ~RX_CALL_WAIT_PROC;
4855 rx_atomic_dec(&rx_nWaiting);
4856 if (opr_queue_IsOnQueue(&call->entry)) {
4857 opr_queue_Remove(&call->entry);
4860 call->state = RX_STATE_ACTIVE;
4861 call->app.mode = RX_MODE_RECEIVING;
4862 #ifdef RX_KERNEL_TRACE
4864 int glockOwner = ISAFS_GLOCK();
4867 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4868 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4874 if (call->flags & RX_CALL_CLEARED) {
4875 /* send an ack now to start the packet flow up again */
4876 call->flags &= ~RX_CALL_CLEARED;
4877 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4879 #ifdef RX_ENABLE_LOCKS
4882 service->nRequestsRunning++;
4883 MUTEX_ENTER(&rx_quota_mutex);
4884 if (service->nRequestsRunning <= service->minProcs)
4887 MUTEX_EXIT(&rx_quota_mutex);
4891 MUTEX_EXIT(&rx_serverPool_lock);
4894 /* Delay the sending of an acknowledge event for a short while, while
4895 * a new call is being prepared (in the case of a client) or a reply
4896 * is being prepared (in the case of a server). Rather than sending
4897 * an ack packet, an ACKALL packet is sent. */
4899 rxi_AckAll(struct rx_call *call)
4901 rxi_SendSpecial(call, call->conn, NULL, RX_PACKET_TYPE_ACKALL,
4903 call->flags |= RX_CALL_ACKALL_SENT;
4907 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4910 struct rx_call *call = arg1;
4911 #ifdef RX_ENABLE_LOCKS
4913 MUTEX_ENTER(&call->lock);
4914 if (event == call->delayedAckEvent)
4915 rxevent_Put(&call->delayedAckEvent);
4916 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4918 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4920 MUTEX_EXIT(&call->lock);
4921 #else /* RX_ENABLE_LOCKS */
4923 rxevent_Put(&call->delayedAckEvent);
4924 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4925 #endif /* RX_ENABLE_LOCKS */
4928 #ifdef RX_ENABLE_LOCKS
4929 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4930 * clearing them out.
4933 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4935 struct opr_queue *cursor;
4938 for (opr_queue_Scan(&call->tq, cursor)) {
4940 = opr_queue_Entry(cursor, struct rx_packet, entry);
4942 p->flags |= RX_PKTFLAG_ACKED;
4947 call->flags |= RX_CALL_TQ_CLEARME;
4948 call->flags |= RX_CALL_TQ_SOME_ACKED;
4951 rxi_rto_cancel(call);
4953 call->tfirst = call->tnext;
4954 call->nSoftAcked = 0;
4956 if (call->flags & RX_CALL_FAST_RECOVER) {
4957 call->flags &= ~RX_CALL_FAST_RECOVER;
4958 call->cwind = call->nextCwind;
4959 call->nextCwind = 0;
4962 CV_SIGNAL(&call->cv_twind);
4964 #endif /* RX_ENABLE_LOCKS */
4967 * Acknowledge the whole transmit queue.
4969 * If we're running without locks, or the transmit queue isn't busy, then
4970 * we can just clear the queue now. Otherwise, we have to mark all of the
4971 * packets as acknowledged, and let rxi_Start clear it later on
4974 rxi_AckAllInTransmitQueue(struct rx_call *call)
4976 #ifdef RX_ENABLE_LOCKS
4977 if (call->flags & RX_CALL_TQ_BUSY) {
4978 rxi_SetAcksInTransmitQueue(call);
4982 rxi_ClearTransmitQueue(call, 0);
4984 /* Clear out the transmit queue for the current call (all packets have
4985 * been received by peer) */
4987 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4989 #ifdef RX_ENABLE_LOCKS
4990 struct opr_queue *cursor;
4991 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4993 for (opr_queue_Scan(&call->tq, cursor)) {
4995 = opr_queue_Entry(cursor, struct rx_packet, entry);
4997 p->flags |= RX_PKTFLAG_ACKED;
5001 call->flags |= RX_CALL_TQ_CLEARME;
5002 call->flags |= RX_CALL_TQ_SOME_ACKED;
5005 #endif /* RX_ENABLE_LOCKS */
5006 #ifdef RXDEBUG_PACKET
5008 #endif /* RXDEBUG_PACKET */
5009 rxi_FreePackets(0, &call->tq);
5010 rxi_WakeUpTransmitQueue(call);
5011 #ifdef RX_ENABLE_LOCKS
5012 call->flags &= ~RX_CALL_TQ_CLEARME;
5016 rxi_rto_cancel(call);
5017 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5018 call->nSoftAcked = 0;
5020 if (call->flags & RX_CALL_FAST_RECOVER) {
5021 call->flags &= ~RX_CALL_FAST_RECOVER;
5022 call->cwind = call->nextCwind;
5024 #ifdef RX_ENABLE_LOCKS
5025 CV_SIGNAL(&call->cv_twind);
5027 osi_rxWakeup(&call->twind);
5032 rxi_ClearReceiveQueue(struct rx_call *call)
5034 if (!opr_queue_IsEmpty(&call->rq)) {
5037 count = rxi_FreePackets(0, &call->rq);
5038 rx_packetReclaims += count;
5039 #ifdef RXDEBUG_PACKET
5041 if ( call->rqc != 0 )
5042 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5044 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5046 if (call->state == RX_STATE_PRECALL) {
5047 call->flags |= RX_CALL_CLEARED;
5051 /* Send an abort packet for the specified call */
5052 static struct rx_packet *
5053 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5054 int istack, int force)
5057 struct clock when, now;
5062 /* Clients should never delay abort messages */
5063 if (rx_IsClientConn(call->conn))
5066 if (call->abortCode != call->error) {
5067 call->abortCode = call->error;
5068 call->abortCount = 0;
5071 if (force || rxi_callAbortThreshhold == 0
5072 || call->abortCount < rxi_callAbortThreshhold) {
5073 rxi_CancelDelayedAbortEvent(call);
5074 error = htonl(call->error);
5077 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5078 (char *)&error, sizeof(error), istack);
5079 } else if (!call->delayedAbortEvent) {
5080 clock_GetTime(&now);
5082 clock_Addmsec(&when, rxi_callAbortDelay);
5083 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5084 call->delayedAbortEvent =
5085 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5091 rxi_CancelDelayedAbortEvent(struct rx_call *call)
5093 if (call->delayedAbortEvent) {
5094 rxevent_Cancel(&call->delayedAbortEvent);
5095 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5099 /* Send an abort packet for the specified connection. Packet is an
5100 * optional pointer to a packet that can be used to send the abort.
5101 * Once the number of abort messages reaches the threshhold, an
5102 * event is scheduled to send the abort. Setting the force flag
5103 * overrides sending delayed abort messages.
5105 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5106 * to send the abort packet.
5109 rxi_SendConnectionAbort(struct rx_connection *conn,
5110 struct rx_packet *packet, int istack, int force)
5117 /* Clients should never delay abort messages */
5118 if (rx_IsClientConn(conn))
5121 if (force || rxi_connAbortThreshhold == 0
5122 || conn->abortCount < rxi_connAbortThreshhold) {
5124 rxevent_Cancel(&conn->delayedAbortEvent);
5125 error = htonl(conn->error);
5127 MUTEX_EXIT(&conn->conn_data_lock);
5129 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5130 RX_PACKET_TYPE_ABORT, (char *)&error,
5131 sizeof(error), istack);
5132 MUTEX_ENTER(&conn->conn_data_lock);
5134 rxi_SendConnectionAbortLater(conn, rxi_connAbortDelay);
5139 /* Associate an error all of the calls owned by a connection. Called
5140 * with error non-zero. This is only for really fatal things, like
5141 * bad authentication responses. The connection itself is set in
5142 * error at this point, so that future packets received will be
5145 rxi_ConnectionError(struct rx_connection *conn,
5151 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5153 MUTEX_ENTER(&conn->conn_data_lock);
5154 rxevent_Cancel(&conn->challengeEvent);
5155 rxevent_Cancel(&conn->natKeepAliveEvent);
5156 if (conn->checkReachEvent) {
5157 rxevent_Cancel(&conn->checkReachEvent);
5158 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5159 putConnection(conn);
5161 MUTEX_EXIT(&conn->conn_data_lock);
5162 for (i = 0; i < RX_MAXCALLS; i++) {
5163 struct rx_call *call = conn->call[i];
5165 MUTEX_ENTER(&call->lock);
5166 rxi_CallError(call, error);
5167 MUTEX_EXIT(&call->lock);
5170 conn->error = error;
5171 if (rx_stats_active)
5172 rx_atomic_inc(&rx_stats.fatalErrors);
5177 * Interrupt an in-progress call with the specified error and wakeup waiters.
5179 * @param[in] call The call to interrupt
5180 * @param[in] error The error code to send to the peer
5183 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5185 MUTEX_ENTER(&call->lock);
5186 rxi_CallError(call, error);
5187 rxi_SendCallAbort(call, NULL, 0, 1);
5188 MUTEX_EXIT(&call->lock);
5192 rxi_CallError(struct rx_call *call, afs_int32 error)
5194 MUTEX_ASSERT(&call->lock);
5195 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5197 error = call->error;
5199 #ifdef RX_ENABLE_LOCKS
5200 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5201 rxi_ResetCall(call, 0);
5204 rxi_ResetCall(call, 0);
5206 call->error = error;
5209 /* Reset various fields in a call structure, and wakeup waiting
5210 * processes. Some fields aren't changed: state & mode are not
5211 * touched (these must be set by the caller), and bufptr, nLeft, and
5212 * nFree are not reset, since these fields are manipulated by
5213 * unprotected macros, and may only be reset by non-interrupting code.
5217 rxi_ResetCall(struct rx_call *call, int newcall)
5220 struct rx_peer *peer;
5221 struct rx_packet *packet;
5223 MUTEX_ASSERT(&call->lock);
5224 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5226 /* Notify anyone who is waiting for asynchronous packet arrival */
5227 if (call->arrivalProc) {
5228 (*call->arrivalProc) (call, call->arrivalProcHandle,
5229 call->arrivalProcArg);
5230 call->arrivalProc = (void (*)())0;
5234 rxi_CancelGrowMTUEvent(call);
5236 if (call->delayedAbortEvent) {
5237 rxi_CancelDelayedAbortEvent(call);
5238 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5240 rxi_SendCallAbort(call, packet, 0, 1);
5241 rxi_FreePacket(packet);
5246 * Update the peer with the congestion information in this call
5247 * so other calls on this connection can pick up where this call
5248 * left off. If the congestion sequence numbers don't match then
5249 * another call experienced a retransmission.
5251 peer = call->conn->peer;
5252 MUTEX_ENTER(&peer->peer_lock);
5254 if (call->congestSeq == peer->congestSeq) {
5255 peer->cwind = MAX(peer->cwind, call->cwind);
5256 peer->MTU = MAX(peer->MTU, call->MTU);
5257 peer->nDgramPackets =
5258 MAX(peer->nDgramPackets, call->nDgramPackets);
5261 call->abortCode = 0;
5262 call->abortCount = 0;
5264 if (peer->maxDgramPackets > 1) {
5265 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5267 call->MTU = peer->MTU;
5269 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5270 call->ssthresh = rx_maxSendWindow;
5271 call->nDgramPackets = peer->nDgramPackets;
5272 call->congestSeq = peer->congestSeq;
5273 call->rtt = peer->rtt;
5274 call->rtt_dev = peer->rtt_dev;
5275 clock_Zero(&call->rto);
5276 clock_Addmsec(&call->rto,
5277 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5278 MUTEX_EXIT(&peer->peer_lock);
5280 flags = call->flags;
5281 rxi_WaitforTQBusy(call);
5283 rxi_ClearTransmitQueue(call, 1);
5284 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5285 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5289 rxi_ClearReceiveQueue(call);
5290 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5294 call->twind = call->conn->twind[call->channel];
5295 call->rwind = call->conn->rwind[call->channel];
5296 call->nSoftAcked = 0;
5297 call->nextCwind = 0;
5300 call->nCwindAcks = 0;
5301 call->nSoftAcks = 0;
5302 call->nHardAcks = 0;
5304 call->tfirst = call->rnext = call->tnext = 1;
5307 call->lastAcked = 0;
5308 call->localStatus = call->remoteStatus = 0;
5310 if (flags & RX_CALL_READER_WAIT) {
5311 #ifdef RX_ENABLE_LOCKS
5312 CV_BROADCAST(&call->cv_rq);
5314 osi_rxWakeup(&call->rq);
5317 if (flags & RX_CALL_WAIT_PACKETS) {
5318 MUTEX_ENTER(&rx_freePktQ_lock);
5319 rxi_PacketsUnWait(); /* XXX */
5320 MUTEX_EXIT(&rx_freePktQ_lock);
5322 #ifdef RX_ENABLE_LOCKS
5323 CV_SIGNAL(&call->cv_twind);
5325 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5326 osi_rxWakeup(&call->twind);
5329 if (flags & RX_CALL_WAIT_PROC) {
5330 rx_atomic_dec(&rx_nWaiting);
5332 #ifdef RX_ENABLE_LOCKS
5333 /* The following ensures that we don't mess with any queue while some
5334 * other thread might also be doing so. The call_queue_lock field is
5335 * is only modified under the call lock. If the call is in the process
5336 * of being removed from a queue, the call is not locked until the
5337 * the queue lock is dropped and only then is the call_queue_lock field
5338 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5339 * Note that any other routine which removes a call from a queue has to
5340 * obtain the queue lock before examing the queue and removing the call.
5342 if (call->call_queue_lock) {
5343 MUTEX_ENTER(call->call_queue_lock);
5344 if (opr_queue_IsOnQueue(&call->entry)) {
5345 opr_queue_Remove(&call->entry);
5347 MUTEX_EXIT(call->call_queue_lock);
5348 CLEAR_CALL_QUEUE_LOCK(call);
5350 #else /* RX_ENABLE_LOCKS */
5351 if (opr_queue_IsOnQueue(&call->entry)) {
5352 opr_queue_Remove(&call->entry);
5354 #endif /* RX_ENABLE_LOCKS */
5356 rxi_CancelKeepAliveEvent(call);
5357 rxi_CancelDelayedAckEvent(call);
5360 /* Send an acknowledge for the indicated packet (seq,serial) of the
5361 * indicated call, for the indicated reason (reason). This
5362 * acknowledge will specifically acknowledge receiving the packet, and
5363 * will also specify which other packets for this call have been
5364 * received. This routine returns the packet that was used to the
5365 * caller. The caller is responsible for freeing it or re-using it.
5366 * This acknowledgement also returns the highest sequence number
5367 * actually read out by the higher level to the sender; the sender
5368 * promises to keep around packets that have not been read by the
5369 * higher level yet (unless, of course, the sender decides to abort
5370 * the call altogether). Any of p, seq, serial, pflags, or reason may
5371 * be set to zero without ill effect. That is, if they are zero, they
5372 * will not convey any information.
5373 * NOW there is a trailer field, after the ack where it will safely be
5374 * ignored by mundanes, which indicates the maximum size packet this
5375 * host can swallow. */
5377 struct rx_packet *optionalPacket; use to send ack (or null)
5378 int seq; Sequence number of the packet we are acking
5379 int serial; Serial number of the packet
5380 int pflags; Flags field from packet header
5381 int reason; Reason an acknowledge was prompted
5384 #define RX_ZEROS 1024
5385 static char rx_zeros[RX_ZEROS];
5388 rxi_SendAck(struct rx_call *call,
5389 struct rx_packet *optionalPacket, int serial, int reason,
5392 struct rx_ackPacket *ap;
5393 struct rx_packet *p;
5394 struct opr_queue *cursor;
5397 afs_uint32 padbytes = 0;
5398 #ifdef RX_ENABLE_TSFPQ
5399 struct rx_ts_info_t * rx_ts_info;
5403 * Open the receive window once a thread starts reading packets
5405 if (call->rnext > 1) {
5406 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5409 /* Don't attempt to grow MTU if this is a critical ping */
5410 if (reason == RX_ACK_MTU) {
5411 /* keep track of per-call attempts, if we're over max, do in small
5412 * otherwise in larger? set a size to increment by, decrease
5415 if (call->conn->peer->maxPacketSize &&
5416 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5418 padbytes = call->conn->peer->maxPacketSize+16;
5420 padbytes = call->conn->peer->maxMTU + 128;
5422 /* do always try a minimum size ping */
5423 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5425 /* subtract the ack payload */
5426 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5427 reason = RX_ACK_PING;
5430 call->nHardAcks = 0;
5431 call->nSoftAcks = 0;
5432 if (call->rnext > call->lastAcked)
5433 call->lastAcked = call->rnext;
5437 rx_computelen(p, p->length); /* reset length, you never know */
5438 } /* where that's been... */
5439 #ifdef RX_ENABLE_TSFPQ
5441 RX_TS_INFO_GET(rx_ts_info);
5442 if ((p = rx_ts_info->local_special_packet)) {
5443 rx_computelen(p, p->length);
5444 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5445 rx_ts_info->local_special_packet = p;
5446 } else { /* We won't send the ack, but don't panic. */
5447 return optionalPacket;
5451 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5452 /* We won't send the ack, but don't panic. */
5453 return optionalPacket;
5458 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5461 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5462 #ifndef RX_ENABLE_TSFPQ
5463 if (!optionalPacket)
5466 return optionalPacket;
5468 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5469 if (rx_Contiguous(p) < templ) {
5470 #ifndef RX_ENABLE_TSFPQ
5471 if (!optionalPacket)
5474 return optionalPacket;
5479 /* MTUXXX failing to send an ack is very serious. We should */
5480 /* try as hard as possible to send even a partial ack; it's */
5481 /* better than nothing. */
5482 ap = (struct rx_ackPacket *)rx_DataOf(p);
5483 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5484 ap->reason = reason;
5486 /* The skew computation used to be bogus, I think it's better now. */
5487 /* We should start paying attention to skew. XXX */
5488 ap->serial = htonl(serial);
5489 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5492 * First packet not yet forwarded to reader. When ACKALL has been
5493 * sent the peer has been told that all received packets will be
5494 * delivered to the reader. The value 'rnext' is used internally
5495 * to refer to the next packet in the receive queue that must be
5496 * delivered to the reader. From the perspective of the peer it
5497 * already has so report the last sequence number plus one if there
5498 * are packets in the receive queue awaiting processing.
5500 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5501 !opr_queue_IsEmpty(&call->rq)) {
5502 ap->firstPacket = htonl(opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq + 1);
5504 ap->firstPacket = htonl(call->rnext);
5506 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5508 /* No fear of running out of ack packet here because there can only
5509 * be at most one window full of unacknowledged packets. The window
5510 * size must be constrained to be less than the maximum ack size,
5511 * of course. Also, an ack should always fit into a single packet
5512 * -- it should not ever be fragmented. */
5514 for (opr_queue_Scan(&call->rq, cursor)) {
5515 struct rx_packet *rqp
5516 = opr_queue_Entry(cursor, struct rx_packet, entry);
5518 if (!rqp || !call->rq.next
5519 || (rqp->header.seq > (call->rnext + call->rwind))) {
5520 #ifndef RX_ENABLE_TSFPQ
5521 if (!optionalPacket)
5524 rxi_CallError(call, RX_CALL_DEAD);
5525 return optionalPacket;
5528 while (rqp->header.seq > call->rnext + offset)
5529 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5530 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5532 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5533 #ifndef RX_ENABLE_TSFPQ
5534 if (!optionalPacket)
5537 rxi_CallError(call, RX_CALL_DEAD);
5538 return optionalPacket;
5544 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5546 /* Must zero the 3 octets that rx_AckDataSize skips at the end of the
5549 rx_packetwrite(p, rx_AckDataSize(offset) - 3, 3, rx_zeros);
5551 /* these are new for AFS 3.3 */
5552 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5553 templ = htonl(templ);
5554 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5555 templ = htonl(call->conn->peer->ifMTU);
5556 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5557 sizeof(afs_int32), &templ);
5559 /* new for AFS 3.4 */
5560 templ = htonl(call->rwind);
5561 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5562 sizeof(afs_int32), &templ);
5564 /* new for AFS 3.5 */
5565 templ = htonl(call->conn->peer->ifDgramPackets);
5566 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5567 sizeof(afs_int32), &templ);
5569 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5571 p->header.serviceId = call->conn->serviceId;
5572 p->header.cid = (call->conn->cid | call->channel);
5573 p->header.callNumber = *call->callNumber;
5575 p->header.securityIndex = call->conn->securityIndex;
5576 p->header.epoch = call->conn->epoch;
5577 p->header.type = RX_PACKET_TYPE_ACK;
5578 p->header.flags = RX_SLOW_START_OK;
5579 if (reason == RX_ACK_PING)
5580 p->header.flags |= RX_REQUEST_ACK;
5582 while (padbytes > 0) {
5583 if (padbytes > RX_ZEROS) {
5584 rx_packetwrite(p, p->length, RX_ZEROS, rx_zeros);
5585 p->length += RX_ZEROS;
5586 padbytes -= RX_ZEROS;
5588 rx_packetwrite(p, p->length, padbytes, rx_zeros);
5589 p->length += padbytes;
5594 if (call->conn->type == RX_CLIENT_CONNECTION)
5595 p->header.flags |= RX_CLIENT_INITIATED;
5599 if (rxdebug_active) {
5603 len = _snprintf(msg, sizeof(msg),
5604 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5605 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5606 ntohl(ap->serial), ntohl(ap->previousPacket),
5607 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5608 ap->nAcks, ntohs(ap->bufferSpace) );
5612 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5613 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5617 OutputDebugString(msg);
5619 #else /* AFS_NT40_ENV */
5621 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5622 ap->reason, ntohl(ap->previousPacket),
5623 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5625 for (offset = 0; offset < ap->nAcks; offset++)
5626 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5631 #endif /* AFS_NT40_ENV */
5634 int i, nbytes = p->length;
5636 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5637 if (nbytes <= p->wirevec[i].iov_len) {
5640 savelen = p->wirevec[i].iov_len;
5642 p->wirevec[i].iov_len = nbytes;
5644 rxi_Send(call, p, istack);
5645 p->wirevec[i].iov_len = savelen;
5649 nbytes -= p->wirevec[i].iov_len;
5652 if (rx_stats_active)
5653 rx_atomic_inc(&rx_stats.ackPacketsSent);
5654 #ifndef RX_ENABLE_TSFPQ
5655 if (!optionalPacket)
5658 return optionalPacket; /* Return packet for re-use by caller */
5662 struct rx_packet **list;
5667 /* Send all of the packets in the list in single datagram */
5669 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5670 int istack, int moreFlag)
5676 struct rx_connection *conn = call->conn;
5677 struct rx_peer *peer = conn->peer;
5679 MUTEX_ENTER(&peer->peer_lock);
5680 peer->nSent += xmit->len;
5681 if (xmit->resending)
5682 peer->reSends += xmit->len;
5683 MUTEX_EXIT(&peer->peer_lock);
5685 if (rx_stats_active) {
5686 if (xmit->resending)
5687 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5689 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5692 clock_GetTime(&now);
5694 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5698 /* Set the packet flags and schedule the resend events */
5699 /* Only request an ack for the last packet in the list */
5700 for (i = 0; i < xmit->len; i++) {
5701 struct rx_packet *packet = xmit->list[i];
5703 /* Record the time sent */
5704 packet->timeSent = now;
5705 packet->flags |= RX_PKTFLAG_SENT;
5707 /* Ask for an ack on retransmitted packets, on every other packet
5708 * if the peer doesn't support slow start. Ask for an ack on every
5709 * packet until the congestion window reaches the ack rate. */
5710 if (packet->header.serial) {
5713 packet->firstSent = now;
5714 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5715 || (!(call->flags & RX_CALL_SLOW_START_OK)
5716 && (packet->header.seq & 1)))) {
5721 /* Tag this packet as not being the last in this group,
5722 * for the receiver's benefit */
5723 if (i < xmit->len - 1 || moreFlag) {
5724 packet->header.flags |= RX_MORE_PACKETS;
5729 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5732 /* Since we're about to send a data packet to the peer, it's
5733 * safe to nuke any scheduled end-of-packets ack */
5734 rxi_CancelDelayedAckEvent(call);
5736 MUTEX_EXIT(&call->lock);
5737 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5738 if (xmit->len > 1) {
5739 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5741 rxi_SendPacket(call, conn, xmit->list[0], istack);
5743 MUTEX_ENTER(&call->lock);
5744 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5746 /* Tell the RTO calculation engine that we have sent a packet, and
5747 * if it was the last one */
5748 rxi_rto_packet_sent(call, lastPacket, istack);
5750 /* Update last send time for this call (for keep-alive
5751 * processing), and for the connection (so that we can discover
5752 * idle connections) */
5753 conn->lastSendTime = call->lastSendTime = clock_Sec();
5756 /* When sending packets we need to follow these rules:
5757 * 1. Never send more than maxDgramPackets in a jumbogram.
5758 * 2. Never send a packet with more than two iovecs in a jumbogram.
5759 * 3. Never send a retransmitted packet in a jumbogram.
5760 * 4. Never send more than cwind/4 packets in a jumbogram
5761 * We always keep the last list we should have sent so we
5762 * can set the RX_MORE_PACKETS flags correctly.
5766 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5771 struct xmitlist working;
5772 struct xmitlist last;
5774 struct rx_peer *peer = call->conn->peer;
5775 int morePackets = 0;
5777 memset(&last, 0, sizeof(struct xmitlist));
5778 working.list = &list[0];
5780 working.resending = 0;
5782 recovery = call->flags & RX_CALL_FAST_RECOVER;
5784 for (i = 0; i < len; i++) {
5785 /* Does the current packet force us to flush the current list? */
5787 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5788 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5790 /* This sends the 'last' list and then rolls the current working
5791 * set into the 'last' one, and resets the working set */
5794 rxi_SendList(call, &last, istack, 1);
5795 /* If the call enters an error state stop sending, or if
5796 * we entered congestion recovery mode, stop sending */
5798 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5803 working.resending = 0;
5804 working.list = &list[i];
5806 /* Add the current packet to the list if it hasn't been acked.
5807 * Otherwise adjust the list pointer to skip the current packet. */
5808 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5811 if (list[i]->header.serial)
5812 working.resending = 1;
5814 /* Do we need to flush the list? */
5815 if (working.len >= (int)peer->maxDgramPackets
5816 || working.len >= (int)call->nDgramPackets
5817 || working.len >= (int)call->cwind
5818 || list[i]->header.serial
5819 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5821 rxi_SendList(call, &last, istack, 1);
5822 /* If the call enters an error state stop sending, or if
5823 * we entered congestion recovery mode, stop sending */
5825 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5830 working.resending = 0;
5831 working.list = &list[i + 1];
5834 if (working.len != 0) {
5835 osi_Panic("rxi_SendList error");
5837 working.list = &list[i + 1];
5841 /* Send the whole list when the call is in receive mode, when
5842 * the call is in eof mode, when we are in fast recovery mode,
5843 * and when we have the last packet */
5844 /* XXX - The accesses to app.mode aren't safe, as this may be called by
5845 * the listener or event threads
5847 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5848 || (call->flags & RX_CALL_FLUSH)
5849 || (call->flags & RX_CALL_FAST_RECOVER)) {
5850 /* Check for the case where the current list contains
5851 * an acked packet. Since we always send retransmissions
5852 * in a separate packet, we only need to check the first
5853 * packet in the list */
5854 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5858 rxi_SendList(call, &last, istack, morePackets);
5859 /* If the call enters an error state stop sending, or if
5860 * we entered congestion recovery mode, stop sending */
5862 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5866 rxi_SendList(call, &working, istack, 0);
5868 } else if (last.len > 0) {
5869 rxi_SendList(call, &last, istack, 0);
5870 /* Packets which are in 'working' are not sent by this call */
5875 * Check if the peer for the given call is known to be dead
5877 * If the call's peer appears dead (it has encountered fatal network errors
5878 * since the call started) the call is killed with RX_CALL_DEAD if the call
5879 * is active. Otherwise, we do nothing.
5881 * @param[in] call The call to check
5884 * @retval 0 The call is fine, and we haven't done anything to the call
5885 * @retval nonzero The call's peer appears dead, and the call has been
5886 * terminated if it was active
5888 * @pre call->lock must be locked
5891 rxi_CheckPeerDead(struct rx_call *call)
5893 #ifdef AFS_RXERRQ_ENV
5896 if (call->state == RX_STATE_DALLY) {
5900 peererrs = rx_atomic_read(&call->conn->peer->neterrs);
5901 if (call->neterr_gen < peererrs) {
5902 /* we have received network errors since this call started; kill
5904 if (call->state == RX_STATE_ACTIVE) {
5905 rxi_CallError(call, RX_CALL_DEAD);
5909 if (call->neterr_gen > peererrs) {
5910 /* someone has reset the number of peer errors; set the call error gen
5911 * so we can detect if more errors are encountered */
5912 call->neterr_gen = peererrs;
5919 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
5921 struct rx_call *call = arg0;
5922 struct rx_peer *peer;
5923 struct opr_queue *cursor;
5924 struct clock maxTimeout = { 60, 0 };
5926 MUTEX_ENTER(&call->lock);
5928 peer = call->conn->peer;
5930 /* Make sure that the event pointer is removed from the call
5931 * structure, since there is no longer a per-call retransmission
5933 if (event == call->resendEvent) {
5934 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5935 rxevent_Put(&call->resendEvent);
5938 rxi_CheckPeerDead(call);
5940 if (opr_queue_IsEmpty(&call->tq)) {
5941 /* Nothing to do. This means that we've been raced, and that an
5942 * ACK has come in between when we were triggered, and when we
5943 * actually got to run. */
5947 /* We're in loss recovery */
5948 call->flags |= RX_CALL_FAST_RECOVER;
5950 /* Mark all of the pending packets in the queue as being lost */
5951 for (opr_queue_Scan(&call->tq, cursor)) {
5952 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
5953 if (!(p->flags & RX_PKTFLAG_ACKED))
5954 p->flags &= ~RX_PKTFLAG_SENT;
5957 /* We're resending, so we double the timeout of the call. This will be
5958 * dropped back down by the first successful ACK that we receive.
5960 * We apply a maximum value here of 60 seconds
5962 clock_Add(&call->rto, &call->rto);
5963 if (clock_Gt(&call->rto, &maxTimeout))
5964 call->rto = maxTimeout;
5966 /* Packet loss is most likely due to congestion, so drop our window size
5967 * and start again from the beginning */
5968 if (peer->maxDgramPackets >1) {
5969 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5970 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5972 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5973 call->nDgramPackets = 1;
5975 call->nextCwind = 1;
5978 MUTEX_ENTER(&peer->peer_lock);
5979 peer->MTU = call->MTU;
5980 peer->cwind = call->cwind;
5981 peer->nDgramPackets = 1;
5983 call->congestSeq = peer->congestSeq;
5984 MUTEX_EXIT(&peer->peer_lock);
5986 rxi_Start(call, istack);
5989 MUTEX_EXIT(&call->lock);
5992 /* This routine is called when new packets are readied for
5993 * transmission and when retransmission may be necessary, or when the
5994 * transmission window or burst count are favourable. This should be
5995 * better optimized for new packets, the usual case, now that we've
5996 * got rid of queues of send packets. XXXXXXXXXXX */
5998 rxi_Start(struct rx_call *call, int istack)
6000 struct opr_queue *cursor;
6001 #ifdef RX_ENABLE_LOCKS
6002 struct opr_queue *store;
6008 #ifdef RX_ENABLE_LOCKS
6009 if (rx_stats_active)
6010 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
6015 if (!opr_queue_IsEmpty(&call->tq)) { /* If we have anything to send */
6016 /* Send (or resend) any packets that need it, subject to
6017 * window restrictions and congestion burst control
6018 * restrictions. Ask for an ack on the last packet sent in
6019 * this burst. For now, we're relying upon the window being
6020 * considerably bigger than the largest number of packets that
6021 * are typically sent at once by one initial call to
6022 * rxi_Start. This is probably bogus (perhaps we should ask
6023 * for an ack when we're half way through the current
6024 * window?). Also, for non file transfer applications, this
6025 * may end up asking for an ack for every packet. Bogus. XXXX
6028 * But check whether we're here recursively, and let the other guy
6031 #ifdef RX_ENABLE_LOCKS
6032 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6033 call->flags |= RX_CALL_TQ_BUSY;
6035 #endif /* RX_ENABLE_LOCKS */
6037 #ifdef RX_ENABLE_LOCKS
6038 call->flags &= ~RX_CALL_NEED_START;
6039 #endif /* RX_ENABLE_LOCKS */
6041 maxXmitPackets = MIN(call->twind, call->cwind);
6042 for (opr_queue_Scan(&call->tq, cursor)) {
6044 = opr_queue_Entry(cursor, struct rx_packet, entry);
6046 if (p->flags & RX_PKTFLAG_ACKED) {
6047 /* Since we may block, don't trust this */
6048 if (rx_stats_active)
6049 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6050 continue; /* Ignore this packet if it has been acknowledged */
6053 /* Turn off all flags except these ones, which are the same
6054 * on each transmission */
6055 p->header.flags &= RX_PRESET_FLAGS;
6057 if (p->header.seq >=
6058 call->tfirst + MIN((int)call->twind,
6059 (int)(call->nSoftAcked +
6061 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6062 /* Note: if we're waiting for more window space, we can
6063 * still send retransmits; hence we don't return here, but
6064 * break out to schedule a retransmit event */
6065 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6066 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6071 /* Transmit the packet if it needs to be sent. */
6072 if (!(p->flags & RX_PKTFLAG_SENT)) {
6073 if (nXmitPackets == maxXmitPackets) {
6074 rxi_SendXmitList(call, call->xmitList,
6075 nXmitPackets, istack);
6078 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6079 *(call->callNumber), p));
6080 call->xmitList[nXmitPackets++] = p;
6082 } /* end of the queue_Scan */
6084 /* xmitList now hold pointers to all of the packets that are
6085 * ready to send. Now we loop to send the packets */
6086 if (nXmitPackets > 0) {
6087 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6091 #ifdef RX_ENABLE_LOCKS
6093 /* We went into the error state while sending packets. Now is
6094 * the time to reset the call. This will also inform the using
6095 * process that the call is in an error state.
6097 if (rx_stats_active)
6098 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6099 call->flags &= ~RX_CALL_TQ_BUSY;
6100 rxi_WakeUpTransmitQueue(call);
6101 rxi_CallError(call, call->error);
6105 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6107 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6108 /* Some packets have received acks. If they all have, we can clear
6109 * the transmit queue.
6112 for (opr_queue_ScanSafe(&call->tq, cursor, store)) {
6114 = opr_queue_Entry(cursor, struct rx_packet, entry);
6116 if (p->header.seq < call->tfirst
6117 && (p->flags & RX_PKTFLAG_ACKED)) {
6118 opr_queue_Remove(&p->entry);
6119 #ifdef RX_TRACK_PACKETS
6120 p->flags &= ~RX_PKTFLAG_TQ;
6122 #ifdef RXDEBUG_PACKET
6130 call->flags |= RX_CALL_TQ_CLEARME;
6132 if (call->flags & RX_CALL_TQ_CLEARME)
6133 rxi_ClearTransmitQueue(call, 1);
6134 } while (call->flags & RX_CALL_NEED_START);
6136 * TQ references no longer protected by this flag; they must remain
6137 * protected by the call lock.
6139 call->flags &= ~RX_CALL_TQ_BUSY;
6140 rxi_WakeUpTransmitQueue(call);
6142 call->flags |= RX_CALL_NEED_START;
6144 #endif /* RX_ENABLE_LOCKS */
6146 rxi_rto_cancel(call);
6150 /* Also adjusts the keep alive parameters for the call, to reflect
6151 * that we have just sent a packet (so keep alives aren't sent
6154 rxi_Send(struct rx_call *call, struct rx_packet *p,
6157 struct rx_connection *conn = call->conn;
6159 /* Stamp each packet with the user supplied status */
6160 p->header.userStatus = call->localStatus;
6162 /* Allow the security object controlling this call's security to
6163 * make any last-minute changes to the packet */
6164 RXS_SendPacket(conn->securityObject, call, p);
6166 /* Since we're about to send SOME sort of packet to the peer, it's
6167 * safe to nuke any scheduled end-of-packets ack */
6168 rxi_CancelDelayedAckEvent(call);
6170 /* Actually send the packet, filling in more connection-specific fields */
6171 MUTEX_EXIT(&call->lock);
6172 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6173 rxi_SendPacket(call, conn, p, istack);
6174 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6175 MUTEX_ENTER(&call->lock);
6177 /* Update last send time for this call (for keep-alive
6178 * processing), and for the connection (so that we can discover
6179 * idle connections) */
6180 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6181 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6182 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6184 conn->lastSendTime = call->lastSendTime = clock_Sec();
6188 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6189 * that things are fine. Also called periodically to guarantee that nothing
6190 * falls through the cracks (e.g. (error + dally) connections have keepalive
6191 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6193 * haveCTLock Set if calling from rxi_ReapConnections
6196 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6198 struct rx_connection *conn = call->conn;
6200 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6201 afs_uint32 fudgeFactor;
6204 int idle_timeout = 0;
6205 afs_int32 clock_diff = 0;
6207 if (rxi_CheckPeerDead(call)) {
6213 /* Large swings in the clock can have a significant impact on
6214 * the performance of RX call processing. Forward clock shifts
6215 * will result in premature event triggering or timeouts.
6216 * Backward shifts can result in calls not completing until
6217 * the clock catches up with the original start clock value.
6219 * If a backward clock shift of more than five minutes is noticed,
6220 * just fail the call.
6222 if (now < call->lastSendTime)
6223 clock_diff = call->lastSendTime - now;
6224 if (now < call->startWait)
6225 clock_diff = MAX(clock_diff, call->startWait - now);
6226 if (now < call->lastReceiveTime)
6227 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6228 if (clock_diff > 5 * 60)
6230 if (call->state == RX_STATE_ACTIVE)
6231 rxi_CallError(call, RX_CALL_TIMEOUT);
6235 #ifdef RX_ENABLE_LOCKS
6236 if (call->flags & RX_CALL_TQ_BUSY) {
6237 /* Call is active and will be reset by rxi_Start if it's
6238 * in an error state.
6243 /* RTT + 8*MDEV, rounded up to the next second. */
6244 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6245 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6247 deadTime = conn->secondsUntilDead + fudgeFactor;
6248 /* These are computed to the second (+- 1 second). But that's
6249 * good enough for these values, which should be a significant
6250 * number of seconds. */
6251 if (now > (call->lastReceiveTime + deadTime)) {
6252 if (call->state == RX_STATE_ACTIVE) {
6253 cerror = RX_CALL_DEAD;
6256 #ifdef RX_ENABLE_LOCKS
6257 /* Cancel pending events */
6258 rxi_CancelDelayedAckEvent(call);
6259 rxi_rto_cancel(call);
6260 rxi_CancelKeepAliveEvent(call);
6261 rxi_CancelGrowMTUEvent(call);
6262 MUTEX_ENTER(&rx_refcnt_mutex);
6263 /* if rxi_FreeCall returns 1 it has freed the call */
6264 if (call->refCount == 0 &&
6265 rxi_FreeCall(call, haveCTLock))
6267 MUTEX_EXIT(&rx_refcnt_mutex);
6270 MUTEX_EXIT(&rx_refcnt_mutex);
6272 #else /* RX_ENABLE_LOCKS */
6273 rxi_FreeCall(call, 0);
6275 #endif /* RX_ENABLE_LOCKS */
6277 /* Non-active calls are destroyed if they are not responding
6278 * to pings; active calls are simply flagged in error, so the
6279 * attached process can die reasonably gracefully. */
6282 if (conn->idleDeadTime) {
6283 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6287 /* see if we have a non-activity timeout */
6288 if (call->startWait && ((call->startWait + idleDeadTime) < now)) {
6289 if (call->state == RX_STATE_ACTIVE) {
6290 cerror = RX_CALL_TIMEOUT;
6296 if (conn->hardDeadTime) {
6297 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6300 /* see if we have a hard timeout */
6302 && (now > (hardDeadTime + call->startTime.sec))) {
6303 if (call->state == RX_STATE_ACTIVE)
6304 rxi_CallError(call, RX_CALL_TIMEOUT);
6309 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6310 call->lastReceiveTime) {
6311 int oldMTU = conn->peer->ifMTU;
6313 /* If we thought we could send more, perhaps things got worse.
6314 * Shrink by 128 bytes and try again. */
6315 if (conn->peer->maxPacketSize < conn->lastPacketSize)
6316 /* maxPacketSize will be cleared in rxi_SetPeerMtu */
6317 newmtu = MAX(conn->peer->maxPacketSize + RX_HEADER_SIZE,
6318 conn->lastPacketSize - 128 + RX_HEADER_SIZE);
6320 newmtu = conn->lastPacketSize - 128 + RX_HEADER_SIZE;
6322 /* minimum capped in SetPeerMtu */
6323 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6326 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
6328 /* needed so ResetCall doesn't clobber us. */
6329 call->MTU = conn->peer->ifMTU;
6331 /* if we never succeeded, let the error pass out as-is */
6332 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6333 cerror = conn->msgsizeRetryErr;
6336 rxi_CallError(call, cerror);
6341 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6342 void *dummy, int dummy2)
6344 struct rx_connection *conn = arg1;
6345 struct rx_header theader;
6346 char tbuffer[1 + sizeof(struct rx_header)];
6347 struct sockaddr_in taddr;
6350 struct iovec tmpiov[2];
6353 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6356 tp = &tbuffer[sizeof(struct rx_header)];
6357 taddr.sin_family = AF_INET;
6358 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6359 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6360 memset(&taddr.sin_zero, 0, sizeof(taddr.sin_zero));
6361 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6362 taddr.sin_len = sizeof(struct sockaddr_in);
6364 memset(&theader, 0, sizeof(theader));
6365 theader.epoch = htonl(999);
6367 theader.callNumber = 0;
6370 theader.type = RX_PACKET_TYPE_VERSION;
6371 theader.flags = RX_LAST_PACKET;
6372 theader.serviceId = 0;
6374 memcpy(tbuffer, &theader, sizeof(theader));
6375 memcpy(tp, &a, sizeof(a));
6376 tmpiov[0].iov_base = tbuffer;
6377 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6379 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6381 MUTEX_ENTER(&conn->conn_data_lock);
6382 MUTEX_ENTER(&rx_refcnt_mutex);
6383 /* Only reschedule ourselves if the connection would not be destroyed */
6384 if (conn->refCount <= 1) {
6385 rxevent_Put(&conn->natKeepAliveEvent);
6386 MUTEX_EXIT(&rx_refcnt_mutex);
6387 MUTEX_EXIT(&conn->conn_data_lock);
6388 rx_DestroyConnection(conn); /* drop the reference for this */
6390 conn->refCount--; /* drop the reference for this */
6391 MUTEX_EXIT(&rx_refcnt_mutex);
6392 rxevent_Put(&conn->natKeepAliveEvent);
6393 rxi_ScheduleNatKeepAliveEvent(conn);
6394 MUTEX_EXIT(&conn->conn_data_lock);
6399 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6401 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6402 struct clock when, now;
6403 clock_GetTime(&now);
6405 when.sec += conn->secondsUntilNatPing;
6406 MUTEX_ENTER(&rx_refcnt_mutex);
6407 conn->refCount++; /* hold a reference for this */
6408 MUTEX_EXIT(&rx_refcnt_mutex);
6409 conn->natKeepAliveEvent =
6410 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6415 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6417 MUTEX_ENTER(&conn->conn_data_lock);
6418 conn->secondsUntilNatPing = seconds;
6420 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6421 rxi_ScheduleNatKeepAliveEvent(conn);
6423 conn->flags |= RX_CONN_NAT_PING;
6425 MUTEX_EXIT(&conn->conn_data_lock);
6428 /* When a call is in progress, this routine is called occasionally to
6429 * make sure that some traffic has arrived (or been sent to) the peer.
6430 * If nothing has arrived in a reasonable amount of time, the call is
6431 * declared dead; if nothing has been sent for a while, we send a
6432 * keep-alive packet (if we're actually trying to keep the call alive)
6435 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6438 struct rx_call *call = arg1;
6439 struct rx_connection *conn;
6442 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6443 MUTEX_ENTER(&call->lock);
6445 if (event == call->keepAliveEvent)
6446 rxevent_Put(&call->keepAliveEvent);
6450 if (rxi_CheckCall(call, 0)) {
6451 MUTEX_EXIT(&call->lock);
6455 /* Don't try to keep alive dallying calls */
6456 if (call->state == RX_STATE_DALLY) {
6457 MUTEX_EXIT(&call->lock);
6462 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6463 /* Don't try to send keepalives if there is unacknowledged data */
6464 /* the rexmit code should be good enough, this little hack
6465 * doesn't quite work XXX */
6466 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6468 rxi_ScheduleKeepAliveEvent(call);
6469 MUTEX_EXIT(&call->lock);
6472 /* Does what's on the nameplate. */
6474 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6476 struct rx_call *call = arg1;
6477 struct rx_connection *conn;
6479 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6480 MUTEX_ENTER(&call->lock);
6482 if (event == call->growMTUEvent)
6483 rxevent_Put(&call->growMTUEvent);
6485 if (rxi_CheckCall(call, 0)) {
6486 MUTEX_EXIT(&call->lock);
6490 /* Don't bother with dallying calls */
6491 if (call->state == RX_STATE_DALLY) {
6492 MUTEX_EXIT(&call->lock);
6499 * keep being scheduled, just don't do anything if we're at peak,
6500 * or we're not set up to be properly handled (idle timeout required)
6502 if ((conn->peer->maxPacketSize != 0) &&
6503 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6505 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6506 rxi_ScheduleGrowMTUEvent(call, 0);
6507 MUTEX_EXIT(&call->lock);
6511 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6513 if (!call->keepAliveEvent) {
6514 struct clock when, now;
6515 clock_GetTime(&now);
6517 when.sec += call->conn->secondsUntilPing;
6518 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6519 call->keepAliveEvent =
6520 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6525 rxi_CancelKeepAliveEvent(struct rx_call *call) {
6526 if (call->keepAliveEvent) {
6527 rxevent_Cancel(&call->keepAliveEvent);
6528 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6533 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6535 if (!call->growMTUEvent) {
6536 struct clock when, now;
6538 clock_GetTime(&now);
6541 if (call->conn->secondsUntilPing)
6542 secs = (6*call->conn->secondsUntilPing)-1;
6544 if (call->conn->secondsUntilDead)
6545 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6549 CALL_HOLD(call, RX_CALL_REFCOUNT_MTU);
6550 call->growMTUEvent =
6551 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6556 rxi_CancelGrowMTUEvent(struct rx_call *call)
6558 if (call->growMTUEvent) {
6559 rxevent_Cancel(&call->growMTUEvent);
6560 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6565 * Increment the counter for the next connection ID, handling overflow.
6568 update_nextCid(void)
6570 /* Overflow is technically undefined behavior; avoid it. */
6571 if (rx_nextCid > MAX_AFS_INT32 - (1 << RX_CIDSHIFT))
6572 rx_nextCid = -1 * ((MAX_AFS_INT32 / RX_CIDSHIFT) * RX_CIDSHIFT);
6574 rx_nextCid += 1 << RX_CIDSHIFT;
6578 rxi_KeepAliveOn(struct rx_call *call)
6580 /* Pretend last packet received was received now--i.e. if another
6581 * packet isn't received within the keep alive time, then the call
6582 * will die; Initialize last send time to the current time--even
6583 * if a packet hasn't been sent yet. This will guarantee that a
6584 * keep-alive is sent within the ping time */
6585 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6586 rxi_ScheduleKeepAliveEvent(call);
6590 rxi_GrowMTUOn(struct rx_call *call)
6592 struct rx_connection *conn = call->conn;
6593 MUTEX_ENTER(&conn->conn_data_lock);
6594 conn->lastPingSizeSer = conn->lastPingSize = 0;
6595 MUTEX_EXIT(&conn->conn_data_lock);
6596 rxi_ScheduleGrowMTUEvent(call, 1);
6599 /* This routine is called to send connection abort messages
6600 * that have been delayed to throttle looping clients. */
6602 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6605 struct rx_connection *conn = arg1;
6608 struct rx_packet *packet;
6610 MUTEX_ENTER(&conn->conn_data_lock);
6611 rxevent_Put(&conn->delayedAbortEvent);
6612 error = htonl(conn->error);
6614 MUTEX_EXIT(&conn->conn_data_lock);
6615 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6618 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6619 RX_PACKET_TYPE_ABORT, (char *)&error,
6621 rxi_FreePacket(packet);
6625 /* This routine is called to send call abort messages
6626 * that have been delayed to throttle looping clients. */
6628 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6631 struct rx_call *call = arg1;
6634 struct rx_packet *packet;
6636 MUTEX_ENTER(&call->lock);
6637 rxevent_Put(&call->delayedAbortEvent);
6638 error = htonl(call->error);
6640 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6643 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6644 (char *)&error, sizeof(error), 0);
6645 rxi_FreePacket(packet);
6647 MUTEX_EXIT(&call->lock);
6648 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6651 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6652 * seconds) to ask the client to authenticate itself. The routine
6653 * issues a challenge to the client, which is obtained from the
6654 * security object associated with the connection */
6656 rxi_ChallengeEvent(struct rxevent *event,
6657 void *arg0, void *arg1, int tries)
6659 struct rx_connection *conn = arg0;
6662 rxevent_Put(&conn->challengeEvent);
6664 /* If there are no active calls it is not worth re-issuing the
6665 * challenge. If the client issues another call on this connection
6666 * the challenge can be requested at that time.
6668 if (!rxi_HasActiveCalls(conn))
6671 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6672 struct rx_packet *packet;
6673 struct clock when, now;
6676 /* We've failed to authenticate for too long.
6677 * Reset any calls waiting for authentication;
6678 * they are all in RX_STATE_PRECALL.
6682 MUTEX_ENTER(&conn->conn_call_lock);
6683 for (i = 0; i < RX_MAXCALLS; i++) {
6684 struct rx_call *call = conn->call[i];
6686 MUTEX_ENTER(&call->lock);
6687 if (call->state == RX_STATE_PRECALL) {
6688 rxi_CallError(call, RX_CALL_DEAD);
6689 rxi_SendCallAbort(call, NULL, 0, 0);
6691 MUTEX_EXIT(&call->lock);
6694 MUTEX_EXIT(&conn->conn_call_lock);
6698 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6700 /* If there's no packet available, do this later. */
6701 RXS_GetChallenge(conn->securityObject, conn, packet);
6702 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6703 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6704 rxi_FreePacket(packet);
6705 conn->securityChallengeSent = 1;
6707 clock_GetTime(&now);
6709 when.sec += RX_CHALLENGE_TIMEOUT;
6710 conn->challengeEvent =
6711 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6716 /* Call this routine to start requesting the client to authenticate
6717 * itself. This will continue until authentication is established,
6718 * the call times out, or an invalid response is returned. The
6719 * security object associated with the connection is asked to create
6720 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6721 * defined earlier. */
6723 rxi_ChallengeOn(struct rx_connection *conn)
6725 if (!conn->challengeEvent) {
6726 RXS_CreateChallenge(conn->securityObject, conn);
6727 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6732 /* rxi_ComputeRoundTripTime is called with peer locked. */
6733 /* peer may be null */
6735 rxi_ComputeRoundTripTime(struct rx_packet *p,
6736 struct rx_ackPacket *ack,
6737 struct rx_call *call,
6738 struct rx_peer *peer,
6741 struct clock thisRtt, *sentp;
6745 /* If the ACK is delayed, then do nothing */
6746 if (ack->reason == RX_ACK_DELAY)
6749 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6750 * their RTT multiple times, so only include the RTT of the last packet
6752 if (p->flags & RX_JUMBO_PACKET)
6755 /* Use the serial number to determine which transmission the ACK is for,
6756 * and set the sent time to match this. If we have no serial number, then
6757 * only use the ACK for RTT calculations if the packet has not been
6761 serial = ntohl(ack->serial);
6763 if (serial == p->header.serial) {
6764 sentp = &p->timeSent;
6765 } else if (serial == p->firstSerial) {
6766 sentp = &p->firstSent;
6767 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6768 sentp = &p->firstSent;
6772 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6773 sentp = &p->firstSent;
6780 if (clock_Lt(&thisRtt, sentp))
6781 return; /* somebody set the clock back, don't count this time. */
6783 clock_Sub(&thisRtt, sentp);
6784 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6785 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6787 if (clock_IsZero(&thisRtt)) {
6789 * The actual round trip time is shorter than the
6790 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6791 * Since we can't tell which at the moment we will assume 1ms.
6793 thisRtt.usec = 1000;
6796 if (rx_stats_active) {
6797 MUTEX_ENTER(&rx_stats_mutex);
6798 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6799 rx_stats.minRtt = thisRtt;
6800 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6801 if (thisRtt.sec > 60) {
6802 MUTEX_EXIT(&rx_stats_mutex);
6803 return; /* somebody set the clock ahead */
6805 rx_stats.maxRtt = thisRtt;
6807 clock_Add(&rx_stats.totalRtt, &thisRtt);
6808 rx_atomic_inc(&rx_stats.nRttSamples);
6809 MUTEX_EXIT(&rx_stats_mutex);
6812 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6814 /* Apply VanJacobson round-trip estimations */
6819 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6820 * srtt is stored as fixed point with 3 bits after the binary
6821 * point (i.e., scaled by 8). The following magic is
6822 * equivalent to the smoothing algorithm in rfc793 with an
6823 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6824 * srtt'*8 = rtt + srtt*7
6825 * srtt'*8 = srtt*8 + rtt - srtt
6826 * srtt' = srtt + rtt/8 - srtt/8
6827 * srtt' = srtt + (rtt - srtt)/8
6830 delta = _8THMSEC(&thisRtt) - call->rtt;
6831 call->rtt += (delta >> 3);
6834 * We accumulate a smoothed rtt variance (actually, a smoothed
6835 * mean difference), then set the retransmit timer to smoothed
6836 * rtt + 4 times the smoothed variance (was 2x in van's original
6837 * paper, but 4x works better for me, and apparently for him as
6839 * rttvar is stored as
6840 * fixed point with 2 bits after the binary point (scaled by
6841 * 4). The following is equivalent to rfc793 smoothing with
6842 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6843 * rttvar'*4 = rttvar*3 + |delta|
6844 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6845 * rttvar' = rttvar + |delta|/4 - rttvar/4
6846 * rttvar' = rttvar + (|delta| - rttvar)/4
6847 * This replaces rfc793's wired-in beta.
6848 * dev*4 = dev*4 + (|actual - expected| - dev)
6854 delta -= (call->rtt_dev << 1);
6855 call->rtt_dev += (delta >> 3);
6857 /* I don't have a stored RTT so I start with this value. Since I'm
6858 * probably just starting a call, and will be pushing more data down
6859 * this, I expect congestion to increase rapidly. So I fudge a
6860 * little, and I set deviance to half the rtt. In practice,
6861 * deviance tends to approach something a little less than
6862 * half the smoothed rtt. */
6863 call->rtt = _8THMSEC(&thisRtt) + 8;
6864 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6866 /* the smoothed RTT time is RTT + 4*MDEV
6868 * We allow a user specified minimum to be set for this, to allow clamping
6869 * at a minimum value in the same way as TCP. In addition, we have to allow
6870 * for the possibility that this packet is answered by a delayed ACK, so we
6871 * add on a fixed 200ms to account for that timer expiring.
6874 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6875 rx_minPeerTimeout) + 200;
6876 clock_Zero(&call->rto);
6877 clock_Addmsec(&call->rto, rtt_timeout);
6879 /* Update the peer, so any new calls start with our values */
6880 peer->rtt_dev = call->rtt_dev;
6881 peer->rtt = call->rtt;
6883 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6884 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6888 /* Find all server connections that have not been active for a long time, and
6891 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
6894 struct clock now, when;
6895 struct rxevent *event;
6896 clock_GetTime(&now);
6898 /* Find server connection structures that haven't been used for
6899 * greater than rx_idleConnectionTime */
6901 struct rx_connection **conn_ptr, **conn_end;
6902 int i, havecalls = 0;
6903 MUTEX_ENTER(&rx_connHashTable_lock);
6904 for (conn_ptr = &rx_connHashTable[0], conn_end =
6905 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6907 struct rx_connection *conn, *next;
6908 struct rx_call *call;
6912 for (conn = *conn_ptr; conn; conn = next) {
6913 /* XXX -- Shouldn't the connection be locked? */
6916 for (i = 0; i < RX_MAXCALLS; i++) {
6917 call = conn->call[i];
6921 code = MUTEX_TRYENTER(&call->lock);
6924 result = rxi_CheckCall(call, 1);
6925 MUTEX_EXIT(&call->lock);
6927 /* If CheckCall freed the call, it might
6928 * have destroyed the connection as well,
6929 * which screws up the linked lists.
6935 if (conn->type == RX_SERVER_CONNECTION) {
6936 /* This only actually destroys the connection if
6937 * there are no outstanding calls */
6938 MUTEX_ENTER(&conn->conn_data_lock);
6939 MUTEX_ENTER(&rx_refcnt_mutex);
6940 if (!havecalls && !conn->refCount
6941 && ((conn->lastSendTime + rx_idleConnectionTime) <
6943 conn->refCount++; /* it will be decr in rx_DestroyConn */
6944 MUTEX_EXIT(&rx_refcnt_mutex);
6945 MUTEX_EXIT(&conn->conn_data_lock);
6946 #ifdef RX_ENABLE_LOCKS
6947 rxi_DestroyConnectionNoLock(conn);
6948 #else /* RX_ENABLE_LOCKS */
6949 rxi_DestroyConnection(conn);
6950 #endif /* RX_ENABLE_LOCKS */
6952 #ifdef RX_ENABLE_LOCKS
6954 MUTEX_EXIT(&rx_refcnt_mutex);
6955 MUTEX_EXIT(&conn->conn_data_lock);
6957 #endif /* RX_ENABLE_LOCKS */
6961 #ifdef RX_ENABLE_LOCKS
6962 while (rx_connCleanup_list) {
6963 struct rx_connection *conn;
6964 conn = rx_connCleanup_list;
6965 rx_connCleanup_list = rx_connCleanup_list->next;
6966 MUTEX_EXIT(&rx_connHashTable_lock);
6967 rxi_CleanupConnection(conn);
6968 MUTEX_ENTER(&rx_connHashTable_lock);
6970 MUTEX_EXIT(&rx_connHashTable_lock);
6971 #endif /* RX_ENABLE_LOCKS */
6974 /* Find any peer structures that haven't been used (haven't had an
6975 * associated connection) for greater than rx_idlePeerTime */
6977 struct rx_peer **peer_ptr, **peer_end;
6981 * Why do we need to hold the rx_peerHashTable_lock across
6982 * the incrementing of peer_ptr since the rx_peerHashTable
6983 * array is not changing? We don't.
6985 * By dropping the lock periodically we can permit other
6986 * activities to be performed while a rxi_ReapConnections
6987 * call is in progress. The goal of reap connections
6988 * is to clean up quickly without causing large amounts
6989 * of contention. Therefore, it is important that global
6990 * mutexes not be held for extended periods of time.
6992 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6993 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6995 struct rx_peer *peer, *next, *prev;
6997 MUTEX_ENTER(&rx_peerHashTable_lock);
6998 for (prev = peer = *peer_ptr; peer; peer = next) {
7000 code = MUTEX_TRYENTER(&peer->peer_lock);
7001 if ((code) && (peer->refCount == 0)
7002 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7003 struct opr_queue *cursor, *store;
7007 * now know that this peer object is one to be
7008 * removed from the hash table. Once it is removed
7009 * it can't be referenced by other threads.
7010 * Lets remove it first and decrement the struct
7011 * nPeerStructs count.
7013 if (peer == *peer_ptr) {
7019 if (rx_stats_active)
7020 rx_atomic_dec(&rx_stats.nPeerStructs);
7023 * Now if we hold references on 'prev' and 'next'
7024 * we can safely drop the rx_peerHashTable_lock
7025 * while we destroy this 'peer' object.
7031 MUTEX_EXIT(&rx_peerHashTable_lock);
7033 MUTEX_EXIT(&peer->peer_lock);
7034 MUTEX_DESTROY(&peer->peer_lock);
7036 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7037 unsigned int num_funcs;
7038 struct rx_interface_stat *rpc_stat
7039 = opr_queue_Entry(cursor, struct rx_interface_stat,
7044 opr_queue_Remove(&rpc_stat->entry);
7045 opr_queue_Remove(&rpc_stat->entryPeers);
7047 num_funcs = rpc_stat->stats[0].func_total;
7049 sizeof(rx_interface_stat_t) +
7050 rpc_stat->stats[0].func_total *
7051 sizeof(rx_function_entry_v1_t);
7053 rxi_Free(rpc_stat, space);
7055 MUTEX_ENTER(&rx_rpc_stats);
7056 rxi_rpc_peer_stat_cnt -= num_funcs;
7057 MUTEX_EXIT(&rx_rpc_stats);
7062 * Regain the rx_peerHashTable_lock and
7063 * decrement the reference count on 'prev'
7066 MUTEX_ENTER(&rx_peerHashTable_lock);
7073 MUTEX_EXIT(&peer->peer_lock);
7078 MUTEX_EXIT(&rx_peerHashTable_lock);
7082 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7083 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7084 * GC, just below. Really, we shouldn't have to keep moving packets from
7085 * one place to another, but instead ought to always know if we can
7086 * afford to hold onto a packet in its particular use. */
7087 MUTEX_ENTER(&rx_freePktQ_lock);
7088 if (rx_waitingForPackets) {
7089 rx_waitingForPackets = 0;
7090 #ifdef RX_ENABLE_LOCKS
7091 CV_BROADCAST(&rx_waitingForPackets_cv);
7093 osi_rxWakeup(&rx_waitingForPackets);
7096 MUTEX_EXIT(&rx_freePktQ_lock);
7099 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7100 event = rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0);
7101 rxevent_Put(&event);
7105 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7106 * rx.h is sort of strange this is better. This is called with a security
7107 * object before it is discarded. Each connection using a security object has
7108 * its own refcount to the object so it won't actually be freed until the last
7109 * connection is destroyed.
7111 * This is the only rxs module call. A hold could also be written but no one
7115 rxs_Release(struct rx_securityClass *aobj)
7117 return RXS_Close(aobj);
7125 #define TRACE_OPTION_RX_DEBUG 16
7133 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7134 0, KEY_QUERY_VALUE, &parmKey);
7135 if (code != ERROR_SUCCESS)
7138 dummyLen = sizeof(TraceOption);
7139 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7140 (BYTE *) &TraceOption, &dummyLen);
7141 if (code == ERROR_SUCCESS) {
7142 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7144 RegCloseKey (parmKey);
7145 #endif /* AFS_NT40_ENV */
7150 rx_DebugOnOff(int on)
7154 rxdebug_active = on;
7160 rx_StatsOnOff(int on)
7162 rx_stats_active = on;
7166 /* Don't call this debugging routine directly; use dpf */
7168 rxi_DebugPrint(char *format, ...)
7177 va_start(ap, format);
7179 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7182 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7184 OutputDebugString(msg);
7190 va_start(ap, format);
7192 clock_GetTime(&now);
7193 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7194 (unsigned int)now.usec);
7195 vfprintf(rx_Log, format, ap);
7203 * This function is used to process the rx_stats structure that is local
7204 * to a process as well as an rx_stats structure received from a remote
7205 * process (via rxdebug). Therefore, it needs to do minimal version
7209 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7210 afs_int32 freePackets, char version)
7214 if (size != sizeof(struct rx_statistics)) {
7216 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7217 size, sizeof(struct rx_statistics));
7220 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7223 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7224 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7225 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7226 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7227 s->specialPktAllocFailures);
7229 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7230 s->receivePktAllocFailures, s->sendPktAllocFailures,
7231 s->specialPktAllocFailures);
7235 " greedy %u, " "bogusReads %u (last from host %x), "
7236 "noPackets %u, " "noBuffers %u, " "selects %u, "
7237 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7238 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7239 s->selects, s->sendSelects);
7241 fprintf(file, " packets read: ");
7242 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7243 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7245 fprintf(file, "\n");
7248 " other read counters: data %u, " "ack %u, " "dup %u "
7249 "spurious %u " "dally %u\n", s->dataPacketsRead,
7250 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7251 s->ignorePacketDally);
7253 fprintf(file, " packets sent: ");
7254 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7255 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7257 fprintf(file, "\n");
7260 " other send counters: ack %u, " "data %u (not resends), "
7261 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7262 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7263 s->dataPacketsPushed, s->ignoreAckedPacket);
7266 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7267 s->netSendFailures, (int)s->fatalErrors);
7269 if (s->nRttSamples) {
7270 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7271 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7273 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7274 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7278 " %d server connections, " "%d client connections, "
7279 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7280 s->nServerConns, s->nClientConns, s->nPeerStructs,
7281 s->nCallStructs, s->nFreeCallStructs);
7283 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7284 fprintf(file, " %d clock updates\n", clock_nUpdates);
7288 /* for backward compatibility */
7290 rx_PrintStats(FILE * file)
7292 MUTEX_ENTER(&rx_stats_mutex);
7293 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7294 sizeof(rx_stats), rx_nFreePackets,
7296 MUTEX_EXIT(&rx_stats_mutex);
7300 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7302 fprintf(file, "Peer %x.%d.\n",
7303 ntohl(peer->host), (int)ntohs(peer->port));
7306 " Rtt %d, " "total sent %d, " "resent %d\n",
7307 peer->rtt, peer->nSent, peer->reSends);
7309 fprintf(file, " Packet size %d\n", peer->ifMTU);
7313 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7315 * This mutex protects the following static variables:
7319 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7320 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7322 #define LOCK_RX_DEBUG
7323 #define UNLOCK_RX_DEBUG
7324 #endif /* AFS_PTHREAD_ENV */
7326 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7328 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7329 u_char type, void *inputData, size_t inputLength,
7330 void *outputData, size_t outputLength)
7332 static afs_int32 counter = 100;
7333 time_t waitTime, waitCount;
7334 struct rx_header theader;
7337 struct timeval tv_now, tv_wake, tv_delta;
7338 struct sockaddr_in taddr, faddr;
7352 tp = &tbuffer[sizeof(struct rx_header)];
7353 taddr.sin_family = AF_INET;
7354 taddr.sin_port = remotePort;
7355 taddr.sin_addr.s_addr = remoteAddr;
7356 memset(&taddr.sin_zero, 0, sizeof(taddr.sin_zero));
7357 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7358 taddr.sin_len = sizeof(struct sockaddr_in);
7361 memset(&theader, 0, sizeof(theader));
7362 theader.epoch = htonl(999);
7364 theader.callNumber = htonl(counter);
7367 theader.type = type;
7368 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7369 theader.serviceId = 0;
7371 memcpy(tbuffer, &theader, sizeof(theader));
7372 memcpy(tp, inputData, inputLength);
7374 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7375 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7377 /* see if there's a packet available */
7378 gettimeofday(&tv_wake, NULL);
7379 tv_wake.tv_sec += waitTime;
7382 FD_SET(socket, &imask);
7383 tv_delta.tv_sec = tv_wake.tv_sec;
7384 tv_delta.tv_usec = tv_wake.tv_usec;
7385 gettimeofday(&tv_now, NULL);
7387 if (tv_delta.tv_usec < tv_now.tv_usec) {
7389 tv_delta.tv_usec += 1000000;
7392 tv_delta.tv_usec -= tv_now.tv_usec;
7394 if (tv_delta.tv_sec < tv_now.tv_sec) {
7398 tv_delta.tv_sec -= tv_now.tv_sec;
7401 code = select(0, &imask, 0, 0, &tv_delta);
7402 #else /* AFS_NT40_ENV */
7403 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7404 #endif /* AFS_NT40_ENV */
7405 if (code == 1 && FD_ISSET(socket, &imask)) {
7406 /* now receive a packet */
7407 faddrLen = sizeof(struct sockaddr_in);
7409 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7410 (struct sockaddr *)&faddr, &faddrLen);
7413 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7414 if (counter == ntohl(theader.callNumber))
7422 /* see if we've timed out */
7430 code -= sizeof(struct rx_header);
7431 if (code > outputLength)
7432 code = outputLength;
7433 memcpy(outputData, tp, code);
7436 #endif /* RXDEBUG */
7439 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7440 afs_uint16 remotePort, struct rx_debugStats * stat,
7441 afs_uint32 * supportedValues)
7443 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7445 struct rx_debugIn in;
7447 *supportedValues = 0;
7448 in.type = htonl(RX_DEBUGI_GETSTATS);
7451 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7452 &in, sizeof(in), stat, sizeof(*stat));
7455 * If the call was successful, fixup the version and indicate
7456 * what contents of the stat structure are valid.
7457 * Also do net to host conversion of fields here.
7461 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7462 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7464 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7465 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7467 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7468 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7470 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7471 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7473 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7474 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7476 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7477 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7479 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7480 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7482 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7483 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7485 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7486 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7488 stat->nFreePackets = ntohl(stat->nFreePackets);
7489 stat->packetReclaims = ntohl(stat->packetReclaims);
7490 stat->callsExecuted = ntohl(stat->callsExecuted);
7491 stat->nWaiting = ntohl(stat->nWaiting);
7492 stat->idleThreads = ntohl(stat->idleThreads);
7493 stat->nWaited = ntohl(stat->nWaited);
7494 stat->nPackets = ntohl(stat->nPackets);
7503 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7504 afs_uint16 remotePort, struct rx_statistics * stat,
7505 afs_uint32 * supportedValues)
7507 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7509 struct rx_debugIn in;
7510 afs_int32 *lp = (afs_int32 *) stat;
7514 * supportedValues is currently unused, but added to allow future
7515 * versioning of this function.
7518 *supportedValues = 0;
7519 in.type = htonl(RX_DEBUGI_RXSTATS);
7521 memset(stat, 0, sizeof(*stat));
7523 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7524 &in, sizeof(in), stat, sizeof(*stat));
7529 * Do net to host conversion here
7532 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7543 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7544 afs_uint16 remotePort, size_t version_length,
7547 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7549 return MakeDebugCall(socket, remoteAddr, remotePort,
7550 RX_PACKET_TYPE_VERSION, a, 1, version,
7558 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7559 afs_uint16 remotePort, afs_int32 * nextConnection,
7560 int allConnections, afs_uint32 debugSupportedValues,
7561 struct rx_debugConn * conn,
7562 afs_uint32 * supportedValues)
7564 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7566 struct rx_debugIn in;
7570 * supportedValues is currently unused, but added to allow future
7571 * versioning of this function.
7574 *supportedValues = 0;
7575 if (allConnections) {
7576 in.type = htonl(RX_DEBUGI_GETALLCONN);
7578 in.type = htonl(RX_DEBUGI_GETCONN);
7580 in.index = htonl(*nextConnection);
7581 memset(conn, 0, sizeof(*conn));
7583 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7584 &in, sizeof(in), conn, sizeof(*conn));
7587 *nextConnection += 1;
7590 * Convert old connection format to new structure.
7593 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7594 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7595 #define MOVEvL(a) (conn->a = vL->a)
7597 /* any old or unrecognized version... */
7598 for (i = 0; i < RX_MAXCALLS; i++) {
7599 MOVEvL(callState[i]);
7600 MOVEvL(callMode[i]);
7601 MOVEvL(callFlags[i]);
7602 MOVEvL(callOther[i]);
7604 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7605 MOVEvL(secStats.type);
7606 MOVEvL(secStats.level);
7607 MOVEvL(secStats.flags);
7608 MOVEvL(secStats.expires);
7609 MOVEvL(secStats.packetsReceived);
7610 MOVEvL(secStats.packetsSent);
7611 MOVEvL(secStats.bytesReceived);
7612 MOVEvL(secStats.bytesSent);
7617 * Do net to host conversion here
7619 * I don't convert host or port since we are most likely
7620 * going to want these in NBO.
7622 conn->cid = ntohl(conn->cid);
7623 conn->serial = ntohl(conn->serial);
7624 for (i = 0; i < RX_MAXCALLS; i++) {
7625 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7627 conn->error = ntohl(conn->error);
7628 conn->secStats.flags = ntohl(conn->secStats.flags);
7629 conn->secStats.expires = ntohl(conn->secStats.expires);
7630 conn->secStats.packetsReceived =
7631 ntohl(conn->secStats.packetsReceived);
7632 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7633 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7634 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7635 conn->epoch = ntohl(conn->epoch);
7636 conn->natMTU = ntohl(conn->natMTU);
7645 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7646 afs_uint16 remotePort, afs_int32 * nextPeer,
7647 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7648 afs_uint32 * supportedValues)
7650 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7652 struct rx_debugIn in;
7655 * supportedValues is currently unused, but added to allow future
7656 * versioning of this function.
7659 *supportedValues = 0;
7660 in.type = htonl(RX_DEBUGI_GETPEER);
7661 in.index = htonl(*nextPeer);
7662 memset(peer, 0, sizeof(*peer));
7664 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7665 &in, sizeof(in), peer, sizeof(*peer));
7671 * Do net to host conversion here
7673 * I don't convert host or port since we are most likely
7674 * going to want these in NBO.
7676 peer->ifMTU = ntohs(peer->ifMTU);
7677 peer->idleWhen = ntohl(peer->idleWhen);
7678 peer->refCount = ntohs(peer->refCount);
7679 peer->rtt = ntohl(peer->rtt);
7680 peer->rtt_dev = ntohl(peer->rtt_dev);
7681 peer->timeout.sec = 0;
7682 peer->timeout.usec = 0;
7683 peer->nSent = ntohl(peer->nSent);
7684 peer->reSends = ntohl(peer->reSends);
7685 peer->natMTU = ntohs(peer->natMTU);
7686 peer->maxMTU = ntohs(peer->maxMTU);
7687 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7688 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7689 peer->MTU = ntohs(peer->MTU);
7690 peer->cwind = ntohs(peer->cwind);
7691 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7692 peer->congestSeq = ntohs(peer->congestSeq);
7693 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7694 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7695 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7696 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7705 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7706 struct rx_debugPeer * peerStats)
7709 afs_int32 error = 1; /* default to "did not succeed" */
7710 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7712 MUTEX_ENTER(&rx_peerHashTable_lock);
7713 for(tp = rx_peerHashTable[hashValue];
7714 tp != NULL; tp = tp->next) {
7715 if (tp->host == peerHost)
7721 MUTEX_EXIT(&rx_peerHashTable_lock);
7725 MUTEX_ENTER(&tp->peer_lock);
7726 peerStats->host = tp->host;
7727 peerStats->port = tp->port;
7728 peerStats->ifMTU = tp->ifMTU;
7729 peerStats->idleWhen = tp->idleWhen;
7730 peerStats->refCount = tp->refCount;
7731 peerStats->burstSize = 0;
7732 peerStats->burst = 0;
7733 peerStats->burstWait.sec = 0;
7734 peerStats->burstWait.usec = 0;
7735 peerStats->rtt = tp->rtt;
7736 peerStats->rtt_dev = tp->rtt_dev;
7737 peerStats->timeout.sec = 0;
7738 peerStats->timeout.usec = 0;
7739 peerStats->nSent = tp->nSent;
7740 peerStats->reSends = tp->reSends;
7741 peerStats->natMTU = tp->natMTU;
7742 peerStats->maxMTU = tp->maxMTU;
7743 peerStats->maxDgramPackets = tp->maxDgramPackets;
7744 peerStats->ifDgramPackets = tp->ifDgramPackets;
7745 peerStats->MTU = tp->MTU;
7746 peerStats->cwind = tp->cwind;
7747 peerStats->nDgramPackets = tp->nDgramPackets;
7748 peerStats->congestSeq = tp->congestSeq;
7749 peerStats->bytesSent.high = tp->bytesSent >> 32;
7750 peerStats->bytesSent.low = tp->bytesSent & MAX_AFS_UINT32;
7751 peerStats->bytesReceived.high = tp->bytesReceived >> 32;
7752 peerStats->bytesReceived.low
7753 = tp->bytesReceived & MAX_AFS_UINT32;
7754 MUTEX_EXIT(&tp->peer_lock);
7756 MUTEX_ENTER(&rx_peerHashTable_lock);
7759 MUTEX_EXIT(&rx_peerHashTable_lock);
7767 struct rx_serverQueueEntry *np;
7770 struct rx_call *call;
7771 struct rx_serverQueueEntry *sq;
7774 if (rx_atomic_test_and_set_bit(&rxinit_status, 0))
7775 return; /* Already shutdown. */
7779 #ifndef AFS_PTHREAD_ENV
7780 FD_ZERO(&rx_selectMask);
7781 #endif /* AFS_PTHREAD_ENV */
7782 rxi_dataQuota = RX_MAX_QUOTA;
7783 #ifndef AFS_PTHREAD_ENV
7785 #endif /* AFS_PTHREAD_ENV */
7788 #ifndef AFS_PTHREAD_ENV
7789 #ifndef AFS_USE_GETTIMEOFDAY
7791 #endif /* AFS_USE_GETTIMEOFDAY */
7792 #endif /* AFS_PTHREAD_ENV */
7794 while (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
7795 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
7796 opr_queue_Remove(&call->entry);
7797 rxi_Free(call, sizeof(struct rx_call));
7800 while (!opr_queue_IsEmpty(&rx_idleServerQueue)) {
7801 sq = opr_queue_First(&rx_idleServerQueue, struct rx_serverQueueEntry,
7803 opr_queue_Remove(&sq->entry);
7808 struct rx_peer **peer_ptr, **peer_end;
7809 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7810 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7812 struct rx_peer *peer, *next;
7814 MUTEX_ENTER(&rx_peerHashTable_lock);
7815 for (peer = *peer_ptr; peer; peer = next) {
7816 struct opr_queue *cursor, *store;
7819 MUTEX_ENTER(&rx_rpc_stats);
7820 MUTEX_ENTER(&peer->peer_lock);
7821 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7822 unsigned int num_funcs;
7823 struct rx_interface_stat *rpc_stat
7824 = opr_queue_Entry(cursor, struct rx_interface_stat,
7828 opr_queue_Remove(&rpc_stat->entry);
7829 opr_queue_Remove(&rpc_stat->entryPeers);
7830 num_funcs = rpc_stat->stats[0].func_total;
7832 sizeof(rx_interface_stat_t) +
7833 rpc_stat->stats[0].func_total *
7834 sizeof(rx_function_entry_v1_t);
7836 rxi_Free(rpc_stat, space);
7838 /* rx_rpc_stats must be held */
7839 rxi_rpc_peer_stat_cnt -= num_funcs;
7841 MUTEX_EXIT(&peer->peer_lock);
7842 MUTEX_EXIT(&rx_rpc_stats);
7846 if (rx_stats_active)
7847 rx_atomic_dec(&rx_stats.nPeerStructs);
7849 MUTEX_EXIT(&rx_peerHashTable_lock);
7852 for (i = 0; i < RX_MAX_SERVICES; i++) {
7854 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7856 for (i = 0; i < rx_hashTableSize; i++) {
7857 struct rx_connection *tc, *ntc;
7858 MUTEX_ENTER(&rx_connHashTable_lock);
7859 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7861 for (j = 0; j < RX_MAXCALLS; j++) {
7863 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7866 rxi_Free(tc, sizeof(*tc));
7868 MUTEX_EXIT(&rx_connHashTable_lock);
7871 MUTEX_ENTER(&freeSQEList_lock);
7873 while ((np = rx_FreeSQEList)) {
7874 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7875 MUTEX_DESTROY(&np->lock);
7876 rxi_Free(np, sizeof(*np));
7879 MUTEX_EXIT(&freeSQEList_lock);
7880 MUTEX_DESTROY(&freeSQEList_lock);
7881 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7882 MUTEX_DESTROY(&rx_connHashTable_lock);
7883 MUTEX_DESTROY(&rx_peerHashTable_lock);
7884 MUTEX_DESTROY(&rx_serverPool_lock);
7886 osi_Free(rx_connHashTable,
7887 rx_hashTableSize * sizeof(struct rx_connection *));
7888 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7890 UNPIN(rx_connHashTable,
7891 rx_hashTableSize * sizeof(struct rx_connection *));
7892 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7894 MUTEX_ENTER(&rx_quota_mutex);
7895 rxi_dataQuota = RX_MAX_QUOTA;
7896 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7897 MUTEX_EXIT(&rx_quota_mutex);
7903 * Routines to implement connection specific data.
7907 rx_KeyCreate(rx_destructor_t rtn)
7910 MUTEX_ENTER(&rxi_keyCreate_lock);
7911 key = rxi_keyCreate_counter++;
7912 rxi_keyCreate_destructor = (rx_destructor_t *)
7913 realloc((void *)rxi_keyCreate_destructor,
7914 (key + 1) * sizeof(rx_destructor_t));
7915 rxi_keyCreate_destructor[key] = rtn;
7916 MUTEX_EXIT(&rxi_keyCreate_lock);
7921 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7924 MUTEX_ENTER(&conn->conn_data_lock);
7925 if (!conn->specific) {
7926 conn->specific = malloc((key + 1) * sizeof(void *));
7927 for (i = 0; i < key; i++)
7928 conn->specific[i] = NULL;
7929 conn->nSpecific = key + 1;
7930 conn->specific[key] = ptr;
7931 } else if (key >= conn->nSpecific) {
7932 conn->specific = (void **)
7933 realloc(conn->specific, (key + 1) * sizeof(void *));
7934 for (i = conn->nSpecific; i < key; i++)
7935 conn->specific[i] = NULL;
7936 conn->nSpecific = key + 1;
7937 conn->specific[key] = ptr;
7939 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7940 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7941 conn->specific[key] = ptr;
7943 MUTEX_EXIT(&conn->conn_data_lock);
7947 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7950 MUTEX_ENTER(&svc->svc_data_lock);
7951 if (!svc->specific) {
7952 svc->specific = malloc((key + 1) * sizeof(void *));
7953 for (i = 0; i < key; i++)
7954 svc->specific[i] = NULL;
7955 svc->nSpecific = key + 1;
7956 svc->specific[key] = ptr;
7957 } else if (key >= svc->nSpecific) {
7958 svc->specific = (void **)
7959 realloc(svc->specific, (key + 1) * sizeof(void *));
7960 for (i = svc->nSpecific; i < key; i++)
7961 svc->specific[i] = NULL;
7962 svc->nSpecific = key + 1;
7963 svc->specific[key] = ptr;
7965 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7966 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7967 svc->specific[key] = ptr;
7969 MUTEX_EXIT(&svc->svc_data_lock);
7973 rx_GetSpecific(struct rx_connection *conn, int key)
7976 MUTEX_ENTER(&conn->conn_data_lock);
7977 if (key >= conn->nSpecific)
7980 ptr = conn->specific[key];
7981 MUTEX_EXIT(&conn->conn_data_lock);
7986 rx_GetServiceSpecific(struct rx_service *svc, int key)
7989 MUTEX_ENTER(&svc->svc_data_lock);
7990 if (key >= svc->nSpecific)
7993 ptr = svc->specific[key];
7994 MUTEX_EXIT(&svc->svc_data_lock);
7999 #endif /* !KERNEL */
8002 * processStats is a queue used to store the statistics for the local
8003 * process. Its contents are similar to the contents of the rpcStats
8004 * queue on a rx_peer structure, but the actual data stored within
8005 * this queue contains totals across the lifetime of the process (assuming
8006 * the stats have not been reset) - unlike the per peer structures
8007 * which can come and go based upon the peer lifetime.
8010 static struct opr_queue processStats = { &processStats, &processStats };
8013 * peerStats is a queue used to store the statistics for all peer structs.
8014 * Its contents are the union of all the peer rpcStats queues.
8017 static struct opr_queue peerStats = { &peerStats, &peerStats };
8020 * rxi_monitor_processStats is used to turn process wide stat collection
8024 static int rxi_monitor_processStats = 0;
8027 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8030 static int rxi_monitor_peerStats = 0;
8034 rxi_ClearRPCOpStat(rx_function_entry_v1_p rpc_stat)
8036 rpc_stat->invocations = 0;
8037 rpc_stat->bytes_sent = 0;
8038 rpc_stat->bytes_rcvd = 0;
8039 rpc_stat->queue_time_sum.sec = 0;
8040 rpc_stat->queue_time_sum.usec = 0;
8041 rpc_stat->queue_time_sum_sqr.sec = 0;
8042 rpc_stat->queue_time_sum_sqr.usec = 0;
8043 rpc_stat->queue_time_min.sec = 9999999;
8044 rpc_stat->queue_time_min.usec = 9999999;
8045 rpc_stat->queue_time_max.sec = 0;
8046 rpc_stat->queue_time_max.usec = 0;
8047 rpc_stat->execution_time_sum.sec = 0;
8048 rpc_stat->execution_time_sum.usec = 0;
8049 rpc_stat->execution_time_sum_sqr.sec = 0;
8050 rpc_stat->execution_time_sum_sqr.usec = 0;
8051 rpc_stat->execution_time_min.sec = 9999999;
8052 rpc_stat->execution_time_min.usec = 9999999;
8053 rpc_stat->execution_time_max.sec = 0;
8054 rpc_stat->execution_time_max.usec = 0;
8058 * Given all of the information for a particular rpc
8059 * call, find or create (if requested) the stat structure for the rpc.
8062 * the queue of stats that will be updated with the new value
8064 * @param rxInterface
8065 * a unique number that identifies the rpc interface
8068 * the total number of functions in this interface. this is only
8069 * required if create is true
8072 * if true, this invocation was made to a server
8075 * the ip address of the remote host. this is only required if create
8076 * and addToPeerList are true
8079 * the port of the remote host. this is only required if create
8080 * and addToPeerList are true
8082 * @param addToPeerList
8083 * if != 0, add newly created stat to the global peer list
8086 * if a new stats structure is allocated, the counter will
8087 * be updated with the new number of allocated stat structures.
8088 * only required if create is true
8091 * if no stats structure exists, allocate one
8095 static rx_interface_stat_p
8096 rxi_FindRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8097 afs_uint32 totalFunc, int isServer, afs_uint32 remoteHost,
8098 afs_uint32 remotePort, int addToPeerList,
8099 unsigned int *counter, int create)
8101 rx_interface_stat_p rpc_stat = NULL;
8102 struct opr_queue *cursor;
8105 * See if there's already a structure for this interface
8108 for (opr_queue_Scan(stats, cursor)) {
8109 rpc_stat = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8111 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8112 && (rpc_stat->stats[0].remote_is_server == isServer))
8116 /* if they didn't ask us to create, we're done */
8118 if (opr_queue_IsEnd(stats, cursor))
8124 /* can't proceed without these */
8125 if (!totalFunc || !counter)
8129 * Didn't find a match so allocate a new structure and add it to the
8133 if (opr_queue_IsEnd(stats, cursor) || (rpc_stat == NULL)
8134 || (rpc_stat->stats[0].interfaceId != rxInterface)
8135 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8140 sizeof(rx_interface_stat_t) +
8141 totalFunc * sizeof(rx_function_entry_v1_t);
8143 rpc_stat = rxi_Alloc(space);
8144 if (rpc_stat == NULL)
8147 *counter += totalFunc;
8148 for (i = 0; i < totalFunc; i++) {
8149 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8150 rpc_stat->stats[i].remote_peer = remoteHost;
8151 rpc_stat->stats[i].remote_port = remotePort;
8152 rpc_stat->stats[i].remote_is_server = isServer;
8153 rpc_stat->stats[i].interfaceId = rxInterface;
8154 rpc_stat->stats[i].func_total = totalFunc;
8155 rpc_stat->stats[i].func_index = i;
8157 opr_queue_Prepend(stats, &rpc_stat->entry);
8158 if (addToPeerList) {
8159 opr_queue_Prepend(&peerStats, &rpc_stat->entryPeers);
8166 rx_ClearProcessRPCStats(afs_int32 rxInterface)
8168 rx_interface_stat_p rpc_stat;
8171 if (rxInterface == -1)
8174 MUTEX_ENTER(&rx_rpc_stats);
8175 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8178 totalFunc = rpc_stat->stats[0].func_total;
8179 for (i = 0; i < totalFunc; i++)
8180 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8182 MUTEX_EXIT(&rx_rpc_stats);
8187 rx_ClearPeerRPCStats(afs_int32 rxInterface, afs_uint32 peerHost, afs_uint16 peerPort)
8189 rx_interface_stat_p rpc_stat;
8191 struct rx_peer * peer;
8193 if (rxInterface == -1)
8196 peer = rxi_FindPeer(peerHost, peerPort, 0);
8200 MUTEX_ENTER(&rx_rpc_stats);
8201 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8204 totalFunc = rpc_stat->stats[0].func_total;
8205 for (i = 0; i < totalFunc; i++)
8206 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8208 MUTEX_EXIT(&rx_rpc_stats);
8213 rx_CopyProcessRPCStats(afs_uint64 op)
8215 rx_interface_stat_p rpc_stat;
8216 rx_function_entry_v1_p rpcop_stat =
8217 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8218 int currentFunc = (op & MAX_AFS_UINT32);
8219 afs_int32 rxInterface = (op >> 32);
8221 if (!rxi_monitor_processStats)
8224 if (rxInterface == -1)
8227 if (rpcop_stat == NULL)
8230 MUTEX_ENTER(&rx_rpc_stats);
8231 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8234 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8235 sizeof(rx_function_entry_v1_t));
8236 MUTEX_EXIT(&rx_rpc_stats);
8238 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8245 rx_CopyPeerRPCStats(afs_uint64 op, afs_uint32 peerHost, afs_uint16 peerPort)
8247 rx_interface_stat_p rpc_stat;
8248 rx_function_entry_v1_p rpcop_stat =
8249 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8250 int currentFunc = (op & MAX_AFS_UINT32);
8251 afs_int32 rxInterface = (op >> 32);
8252 struct rx_peer *peer;
8254 if (!rxi_monitor_peerStats)
8257 if (rxInterface == -1)
8260 if (rpcop_stat == NULL)
8263 peer = rxi_FindPeer(peerHost, peerPort, 0);
8267 MUTEX_ENTER(&rx_rpc_stats);
8268 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8271 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8272 sizeof(rx_function_entry_v1_t));
8273 MUTEX_EXIT(&rx_rpc_stats);
8275 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8282 rx_ReleaseRPCStats(void *stats)
8285 rxi_Free(stats, sizeof(rx_function_entry_v1_t));
8289 * Given all of the information for a particular rpc
8290 * call, create (if needed) and update the stat totals for the rpc.
8293 * the queue of stats that will be updated with the new value
8295 * @param rxInterface
8296 * a unique number that identifies the rpc interface
8298 * @param currentFunc
8299 * the index of the function being invoked
8302 * the total number of functions in this interface
8305 * the amount of time this function waited for a thread
8308 * the amount of time this function invocation took to execute
8311 * the number bytes sent by this invocation
8314 * the number bytes received by this invocation
8317 * if true, this invocation was made to a server
8320 * the ip address of the remote host
8323 * the port of the remote host
8325 * @param addToPeerList
8326 * if != 0, add newly created stat to the global peer list
8329 * if a new stats structure is allocated, the counter will
8330 * be updated with the new number of allocated stat structures
8335 rxi_AddRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8336 afs_uint32 currentFunc, afs_uint32 totalFunc,
8337 struct clock *queueTime, struct clock *execTime,
8338 afs_uint64 bytesSent, afs_uint64 bytesRcvd, int isServer,
8339 afs_uint32 remoteHost, afs_uint32 remotePort,
8340 int addToPeerList, unsigned int *counter)
8343 rx_interface_stat_p rpc_stat;
8345 rpc_stat = rxi_FindRpcStat(stats, rxInterface, totalFunc, isServer,
8346 remoteHost, remotePort, addToPeerList, counter,
8354 * Increment the stats for this function
8357 rpc_stat->stats[currentFunc].invocations++;
8358 rpc_stat->stats[currentFunc].bytes_sent += bytesSent;
8359 rpc_stat->stats[currentFunc].bytes_rcvd += bytesRcvd;
8360 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8361 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8362 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8363 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8365 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8366 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8368 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8369 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8371 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8372 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8374 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8375 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8383 rxi_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8384 afs_uint32 currentFunc, afs_uint32 totalFunc,
8385 struct clock *queueTime, struct clock *execTime,
8386 afs_uint64 bytesSent, afs_uint64 bytesRcvd,
8390 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8393 MUTEX_ENTER(&rx_rpc_stats);
8395 if (rxi_monitor_peerStats) {
8396 MUTEX_ENTER(&peer->peer_lock);
8397 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8398 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8399 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8400 MUTEX_EXIT(&peer->peer_lock);
8403 if (rxi_monitor_processStats) {
8404 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8405 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8406 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8409 MUTEX_EXIT(&rx_rpc_stats);
8413 * Increment the times and count for a particular rpc function.
8415 * Traditionally this call was invoked from rxgen stubs. Modern stubs
8416 * call rx_RecordCallStatistics instead, so the public version of this
8417 * function is left purely for legacy callers.
8420 * The peer who invoked the rpc
8422 * @param rxInterface
8423 * A unique number that identifies the rpc interface
8425 * @param currentFunc
8426 * The index of the function being invoked
8429 * The total number of functions in this interface
8432 * The amount of time this function waited for a thread
8435 * The amount of time this function invocation took to execute
8438 * The number bytes sent by this invocation
8441 * The number bytes received by this invocation
8444 * If true, this invocation was made to a server
8448 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8449 afs_uint32 currentFunc, afs_uint32 totalFunc,
8450 struct clock *queueTime, struct clock *execTime,
8451 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8457 sent64 = ((afs_uint64)bytesSent->high << 32) + bytesSent->low;
8458 rcvd64 = ((afs_uint64)bytesRcvd->high << 32) + bytesRcvd->low;
8460 rxi_IncrementTimeAndCount(peer, rxInterface, currentFunc, totalFunc,
8461 queueTime, execTime, sent64, rcvd64,
8468 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8472 * IN callerVersion - the rpc stat version of the caller.
8474 * IN count - the number of entries to marshall.
8476 * IN stats - pointer to stats to be marshalled.
8478 * OUT ptr - Where to store the marshalled data.
8485 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8486 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8492 * We only support the first version
8494 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8495 *(ptr++) = stats->remote_peer;
8496 *(ptr++) = stats->remote_port;
8497 *(ptr++) = stats->remote_is_server;
8498 *(ptr++) = stats->interfaceId;
8499 *(ptr++) = stats->func_total;
8500 *(ptr++) = stats->func_index;
8501 *(ptr++) = stats->invocations >> 32;
8502 *(ptr++) = stats->invocations & MAX_AFS_UINT32;
8503 *(ptr++) = stats->bytes_sent >> 32;
8504 *(ptr++) = stats->bytes_sent & MAX_AFS_UINT32;
8505 *(ptr++) = stats->bytes_rcvd >> 32;
8506 *(ptr++) = stats->bytes_rcvd & MAX_AFS_UINT32;
8507 *(ptr++) = stats->queue_time_sum.sec;
8508 *(ptr++) = stats->queue_time_sum.usec;
8509 *(ptr++) = stats->queue_time_sum_sqr.sec;
8510 *(ptr++) = stats->queue_time_sum_sqr.usec;
8511 *(ptr++) = stats->queue_time_min.sec;
8512 *(ptr++) = stats->queue_time_min.usec;
8513 *(ptr++) = stats->queue_time_max.sec;
8514 *(ptr++) = stats->queue_time_max.usec;
8515 *(ptr++) = stats->execution_time_sum.sec;
8516 *(ptr++) = stats->execution_time_sum.usec;
8517 *(ptr++) = stats->execution_time_sum_sqr.sec;
8518 *(ptr++) = stats->execution_time_sum_sqr.usec;
8519 *(ptr++) = stats->execution_time_min.sec;
8520 *(ptr++) = stats->execution_time_min.usec;
8521 *(ptr++) = stats->execution_time_max.sec;
8522 *(ptr++) = stats->execution_time_max.usec;
8528 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8533 * IN callerVersion - the rpc stat version of the caller
8535 * OUT myVersion - the rpc stat version of this function
8537 * OUT clock_sec - local time seconds
8539 * OUT clock_usec - local time microseconds
8541 * OUT allocSize - the number of bytes allocated to contain stats
8543 * OUT statCount - the number stats retrieved from this process.
8545 * OUT stats - the actual stats retrieved from this process.
8549 * Returns void. If successful, stats will != NULL.
8553 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8554 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8555 size_t * allocSize, afs_uint32 * statCount,
8556 afs_uint32 ** stats)
8566 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8569 * Check to see if stats are enabled
8572 MUTEX_ENTER(&rx_rpc_stats);
8573 if (!rxi_monitor_processStats) {
8574 MUTEX_EXIT(&rx_rpc_stats);
8578 clock_GetTime(&now);
8579 *clock_sec = now.sec;
8580 *clock_usec = now.usec;
8583 * Allocate the space based upon the caller version
8585 * If the client is at an older version than we are,
8586 * we return the statistic data in the older data format, but
8587 * we still return our version number so the client knows we
8588 * are maintaining more data than it can retrieve.
8591 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8592 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8593 *statCount = rxi_rpc_process_stat_cnt;
8596 * This can't happen yet, but in the future version changes
8597 * can be handled by adding additional code here
8601 if (space > (size_t) 0) {
8603 ptr = *stats = rxi_Alloc(space);
8606 struct opr_queue *cursor;
8608 for (opr_queue_Scan(&processStats, cursor)) {
8609 struct rx_interface_stat *rpc_stat =
8610 opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8612 * Copy the data based upon the caller version
8614 rx_MarshallProcessRPCStats(callerVersion,
8615 rpc_stat->stats[0].func_total,
8616 rpc_stat->stats, &ptr);
8622 MUTEX_EXIT(&rx_rpc_stats);
8627 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8631 * IN callerVersion - the rpc stat version of the caller
8633 * OUT myVersion - the rpc stat version of this function
8635 * OUT clock_sec - local time seconds
8637 * OUT clock_usec - local time microseconds
8639 * OUT allocSize - the number of bytes allocated to contain stats
8641 * OUT statCount - the number of stats retrieved from the individual
8644 * OUT stats - the actual stats retrieved from the individual peer structures.
8648 * Returns void. If successful, stats will != NULL.
8652 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8653 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8654 size_t * allocSize, afs_uint32 * statCount,
8655 afs_uint32 ** stats)
8665 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8668 * Check to see if stats are enabled
8671 MUTEX_ENTER(&rx_rpc_stats);
8672 if (!rxi_monitor_peerStats) {
8673 MUTEX_EXIT(&rx_rpc_stats);
8677 clock_GetTime(&now);
8678 *clock_sec = now.sec;
8679 *clock_usec = now.usec;
8682 * Allocate the space based upon the caller version
8684 * If the client is at an older version than we are,
8685 * we return the statistic data in the older data format, but
8686 * we still return our version number so the client knows we
8687 * are maintaining more data than it can retrieve.
8690 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8691 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8692 *statCount = rxi_rpc_peer_stat_cnt;
8695 * This can't happen yet, but in the future version changes
8696 * can be handled by adding additional code here
8700 if (space > (size_t) 0) {
8702 ptr = *stats = rxi_Alloc(space);
8705 struct opr_queue *cursor;
8707 for (opr_queue_Scan(&peerStats, cursor)) {
8708 struct rx_interface_stat *rpc_stat
8709 = opr_queue_Entry(cursor, struct rx_interface_stat,
8713 * Copy the data based upon the caller version
8715 rx_MarshallProcessRPCStats(callerVersion,
8716 rpc_stat->stats[0].func_total,
8717 rpc_stat->stats, &ptr);
8723 MUTEX_EXIT(&rx_rpc_stats);
8728 * rx_FreeRPCStats - free memory allocated by
8729 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8733 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8734 * rx_RetrievePeerRPCStats
8736 * IN allocSize - the number of bytes in stats.
8744 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8746 rxi_Free(stats, allocSize);
8750 * rx_queryProcessRPCStats - see if process rpc stat collection is
8751 * currently enabled.
8757 * Returns 0 if stats are not enabled != 0 otherwise
8761 rx_queryProcessRPCStats(void)
8764 MUTEX_ENTER(&rx_rpc_stats);
8765 rc = rxi_monitor_processStats;
8766 MUTEX_EXIT(&rx_rpc_stats);
8771 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8777 * Returns 0 if stats are not enabled != 0 otherwise
8781 rx_queryPeerRPCStats(void)
8784 MUTEX_ENTER(&rx_rpc_stats);
8785 rc = rxi_monitor_peerStats;
8786 MUTEX_EXIT(&rx_rpc_stats);
8791 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8801 rx_enableProcessRPCStats(void)
8803 MUTEX_ENTER(&rx_rpc_stats);
8804 rx_enable_stats = 1;
8805 rxi_monitor_processStats = 1;
8806 MUTEX_EXIT(&rx_rpc_stats);
8810 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8820 rx_enablePeerRPCStats(void)
8822 MUTEX_ENTER(&rx_rpc_stats);
8823 rx_enable_stats = 1;
8824 rxi_monitor_peerStats = 1;
8825 MUTEX_EXIT(&rx_rpc_stats);
8829 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8839 rx_disableProcessRPCStats(void)
8841 struct opr_queue *cursor, *store;
8844 MUTEX_ENTER(&rx_rpc_stats);
8847 * Turn off process statistics and if peer stats is also off, turn
8851 rxi_monitor_processStats = 0;
8852 if (rxi_monitor_peerStats == 0) {
8853 rx_enable_stats = 0;
8856 for (opr_queue_ScanSafe(&processStats, cursor, store)) {
8857 unsigned int num_funcs = 0;
8858 struct rx_interface_stat *rpc_stat
8859 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8861 opr_queue_Remove(&rpc_stat->entry);
8863 num_funcs = rpc_stat->stats[0].func_total;
8865 sizeof(rx_interface_stat_t) +
8866 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8868 rxi_Free(rpc_stat, space);
8869 rxi_rpc_process_stat_cnt -= num_funcs;
8871 MUTEX_EXIT(&rx_rpc_stats);
8875 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8885 rx_disablePeerRPCStats(void)
8887 struct rx_peer **peer_ptr, **peer_end;
8891 * Turn off peer statistics and if process stats is also off, turn
8895 rxi_monitor_peerStats = 0;
8896 if (rxi_monitor_processStats == 0) {
8897 rx_enable_stats = 0;
8900 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8901 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8903 struct rx_peer *peer, *next, *prev;
8905 MUTEX_ENTER(&rx_peerHashTable_lock);
8906 MUTEX_ENTER(&rx_rpc_stats);
8907 for (prev = peer = *peer_ptr; peer; peer = next) {
8909 code = MUTEX_TRYENTER(&peer->peer_lock);
8912 struct opr_queue *cursor, *store;
8914 if (prev == *peer_ptr) {
8925 MUTEX_EXIT(&rx_peerHashTable_lock);
8927 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
8928 unsigned int num_funcs = 0;
8929 struct rx_interface_stat *rpc_stat
8930 = opr_queue_Entry(cursor, struct rx_interface_stat,
8933 opr_queue_Remove(&rpc_stat->entry);
8934 opr_queue_Remove(&rpc_stat->entryPeers);
8935 num_funcs = rpc_stat->stats[0].func_total;
8937 sizeof(rx_interface_stat_t) +
8938 rpc_stat->stats[0].func_total *
8939 sizeof(rx_function_entry_v1_t);
8941 rxi_Free(rpc_stat, space);
8942 rxi_rpc_peer_stat_cnt -= num_funcs;
8944 MUTEX_EXIT(&peer->peer_lock);
8946 MUTEX_ENTER(&rx_peerHashTable_lock);
8956 MUTEX_EXIT(&rx_rpc_stats);
8957 MUTEX_EXIT(&rx_peerHashTable_lock);
8962 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8967 * IN clearFlag - flag indicating which stats to clear
8975 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8977 struct opr_queue *cursor;
8979 MUTEX_ENTER(&rx_rpc_stats);
8981 for (opr_queue_Scan(&processStats, cursor)) {
8982 unsigned int num_funcs = 0, i;
8983 struct rx_interface_stat *rpc_stat
8984 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8986 num_funcs = rpc_stat->stats[0].func_total;
8987 for (i = 0; i < num_funcs; i++) {
8988 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8989 rpc_stat->stats[i].invocations = 0;
8991 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8992 rpc_stat->stats[i].bytes_sent = 0;
8994 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8995 rpc_stat->stats[i].bytes_rcvd = 0;
8997 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8998 rpc_stat->stats[i].queue_time_sum.sec = 0;
8999 rpc_stat->stats[i].queue_time_sum.usec = 0;
9001 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9002 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9003 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9005 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9006 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9007 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9009 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9010 rpc_stat->stats[i].queue_time_max.sec = 0;
9011 rpc_stat->stats[i].queue_time_max.usec = 0;
9013 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9014 rpc_stat->stats[i].execution_time_sum.sec = 0;
9015 rpc_stat->stats[i].execution_time_sum.usec = 0;
9017 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9018 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9019 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9021 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9022 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9023 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9025 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9026 rpc_stat->stats[i].execution_time_max.sec = 0;
9027 rpc_stat->stats[i].execution_time_max.usec = 0;
9032 MUTEX_EXIT(&rx_rpc_stats);
9036 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
9041 * IN clearFlag - flag indicating which stats to clear
9049 rx_clearPeerRPCStats(afs_uint32 clearFlag)
9051 struct opr_queue *cursor;
9053 MUTEX_ENTER(&rx_rpc_stats);
9055 for (opr_queue_Scan(&peerStats, cursor)) {
9056 unsigned int num_funcs, i;
9057 struct rx_interface_stat *rpc_stat
9058 = opr_queue_Entry(cursor, struct rx_interface_stat, entryPeers);
9060 num_funcs = rpc_stat->stats[0].func_total;
9061 for (i = 0; i < num_funcs; i++) {
9062 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9063 rpc_stat->stats[i].invocations = 0;
9065 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9066 rpc_stat->stats[i].bytes_sent = 0;
9068 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9069 rpc_stat->stats[i].bytes_rcvd = 0;
9071 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9072 rpc_stat->stats[i].queue_time_sum.sec = 0;
9073 rpc_stat->stats[i].queue_time_sum.usec = 0;
9075 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9076 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9077 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9079 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9080 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9081 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9083 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9084 rpc_stat->stats[i].queue_time_max.sec = 0;
9085 rpc_stat->stats[i].queue_time_max.usec = 0;
9087 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9088 rpc_stat->stats[i].execution_time_sum.sec = 0;
9089 rpc_stat->stats[i].execution_time_sum.usec = 0;
9091 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9092 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9093 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9095 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9096 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9097 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9099 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9100 rpc_stat->stats[i].execution_time_max.sec = 0;
9101 rpc_stat->stats[i].execution_time_max.usec = 0;
9106 MUTEX_EXIT(&rx_rpc_stats);
9110 * rxi_rxstat_userok points to a routine that returns 1 if the caller
9111 * is authorized to enable/disable/clear RX statistics.
9113 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
9116 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
9118 rxi_rxstat_userok = proc;
9122 rx_RxStatUserOk(struct rx_call *call)
9124 if (!rxi_rxstat_userok)
9126 return rxi_rxstat_userok(call);
9131 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
9132 * function in the MSVC runtime DLL (msvcrt.dll).
9134 * Note: the system serializes calls to this function.
9137 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
9138 DWORD reason, /* reason function is being called */
9139 LPVOID reserved) /* reserved for future use */
9142 case DLL_PROCESS_ATTACH:
9143 /* library is being attached to a process */
9147 case DLL_PROCESS_DETACH:
9154 #endif /* AFS_NT40_ENV */
9157 int rx_DumpCalls(FILE *outputFile, char *cookie)
9159 #ifdef RXDEBUG_PACKET
9160 #ifdef KDUMP_RX_LOCK
9161 struct rx_call_rx_lock *c;
9168 #define RXDPRINTF sprintf
9169 #define RXDPRINTOUT output
9171 #define RXDPRINTF fprintf
9172 #define RXDPRINTOUT outputFile
9175 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9177 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9180 for (c = rx_allCallsp; c; c = c->allNextp) {
9181 u_short rqc, tqc, iovqc;
9183 MUTEX_ENTER(&c->lock);
9184 rqc = opr_queue_Count(&c->rq);
9185 tqc = opr_queue_Count(&c->tq);
9186 iovqc = opr_queue_Count(&c->app.iovq);
9188 RXDPRINTF(RXDPRINTOUT, "%s - call=0x%p, id=%u, state=%u, mode=%u, conn=%p, epoch=%u, cid=%u, callNum=%u, connFlags=0x%x, flags=0x%x, "
9189 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9190 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9191 "resendEvent=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9192 "lastSendTime=%u, lastRecvTime=%u"
9193 #ifdef RX_ENABLE_LOCKS
9196 #ifdef RX_REFCOUNT_CHECK
9197 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9198 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9201 cookie, c, c->call_id, (afs_uint32)c->state, (afs_uint32)c->app.mode, c->conn, c->conn?c->conn->epoch:0, c->conn?c->conn->cid:0,
9202 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9203 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9204 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9205 c->resendEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9206 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime
9207 #ifdef RX_ENABLE_LOCKS
9208 , (afs_uint32)c->refCount
9210 #ifdef RX_REFCOUNT_CHECK
9211 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9214 MUTEX_EXIT(&c->lock);
9217 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9220 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9222 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9224 #endif /* RXDEBUG_PACKET */