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>
79 #include "rx_atomic.h"
80 #include "rx_globals.h"
82 #include "rx_internal.h"
89 #include "rx_packet.h"
90 #include "rx_server.h"
92 #include <afs/rxgen_consts.h>
95 #ifdef AFS_PTHREAD_ENV
97 int (*registerProgram) (pid_t, char *) = 0;
98 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
101 int (*registerProgram) (PROCESS, char *) = 0;
102 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
106 /* Local static routines */
107 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
108 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
109 struct rx_call *, struct rx_peer *,
111 static void rxi_Resend(struct rxevent *event, void *arg0, void *arg1,
113 static void rxi_SendDelayedAck(struct rxevent *event, void *call,
114 void *dummy, int dummy2);
115 static void rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1,
116 void *dummy, int dummy2);
117 static void rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1,
118 void *unused, int unused2);
119 static void rxi_ReapConnections(struct rxevent *unused, void *unused1,
120 void *unused2, int unused3);
121 static struct rx_packet *rxi_SendCallAbort(struct rx_call *call,
122 struct rx_packet *packet,
123 int istack, int force);
124 static void rxi_AckAll(struct rx_call *call);
125 static struct rx_connection
126 *rxi_FindConnection(osi_socket socket, afs_uint32 host, u_short port,
127 u_short serviceId, afs_uint32 cid,
128 afs_uint32 epoch, int type, u_int securityIndex);
129 static struct rx_packet
130 *rxi_ReceiveDataPacket(struct rx_call *call, struct rx_packet *np,
131 int istack, osi_socket socket,
132 afs_uint32 host, u_short port, int *tnop,
133 struct rx_call **newcallp);
134 static struct rx_packet
135 *rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
137 static struct rx_packet
138 *rxi_ReceiveResponsePacket(struct rx_connection *conn,
139 struct rx_packet *np, int istack);
140 static struct rx_packet
141 *rxi_ReceiveChallengePacket(struct rx_connection *conn,
142 struct rx_packet *np, int istack);
143 static void rxi_AttachServerProc(struct rx_call *call, osi_socket socket,
144 int *tnop, struct rx_call **newcallp);
145 static void rxi_ClearTransmitQueue(struct rx_call *call, int force);
146 static void rxi_ClearReceiveQueue(struct rx_call *call);
147 static void rxi_ResetCall(struct rx_call *call, int newcall);
148 static void rxi_ScheduleKeepAliveEvent(struct rx_call *call);
149 static void rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn);
150 static void rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs);
151 static void rxi_KeepAliveOn(struct rx_call *call);
152 static void rxi_GrowMTUOn(struct rx_call *call);
153 static void rxi_ChallengeOn(struct rx_connection *conn);
155 #ifdef RX_ENABLE_LOCKS
156 static int rxi_CheckCall(struct rx_call *call, int haveCTLock);
157 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
159 static int rxi_CheckCall(struct rx_call *call);
162 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
164 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
165 rx_atomic_t rxi_start_in_error;
167 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
169 /* Constant delay time before sending an acknowledge of the last packet
170 * received. This is to avoid sending an extra acknowledge when the
171 * client is about to make another call, anyway, or the server is
174 * The lastAckDelay may not exceeed 400ms without causing peers to
175 * unecessarily timeout.
177 struct clock rx_lastAckDelay = {0, 400000};
179 /* Constant delay time before sending a soft ack when none was requested.
180 * This is to make sure we send soft acks before the sender times out,
181 * Normally we wait and send a hard ack when the receiver consumes the packet
183 * This value has been 100ms in all shipping versions of OpenAFS. Changing it
184 * will require changes to the peer's RTT calculations.
186 struct clock rx_softAckDelay = {0, 100000};
189 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
190 * currently allocated within rx. This number is used to allocate the
191 * memory required to return the statistics when queried.
192 * Protected by the rx_rpc_stats mutex.
195 static unsigned int rxi_rpc_peer_stat_cnt;
198 * rxi_rpc_process_stat_cnt counts the total number of local process stat
199 * structures currently allocated within rx. The number is used to allocate
200 * the memory required to return the statistics when queried.
201 * Protected by the rx_rpc_stats mutex.
204 static unsigned int rxi_rpc_process_stat_cnt;
207 * rxi_busyChannelError is a boolean. It indicates whether or not RX_CALL_BUSY
208 * errors should be reported to the application when a call channel appears busy
209 * (inferred from the receipt of RX_PACKET_TYPE_BUSY packets on the channel),
210 * and there are other call channels in the connection that are not busy.
211 * If 0, we do not return errors upon receiving busy packets; we just keep
212 * trying on the same call channel until we hit a timeout.
214 static afs_int32 rxi_busyChannelError = 0;
216 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
217 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
219 /* Incoming calls wait on this queue when there are no available
220 * server processes */
221 struct opr_queue rx_incomingCallQueue;
223 /* Server processes wait on this queue when there are no appropriate
224 * calls to process */
225 struct opr_queue rx_idleServerQueue;
227 #if !defined(offsetof)
228 #include <stddef.h> /* for definition of offsetof() */
231 #ifdef RX_ENABLE_LOCKS
232 afs_kmutex_t rx_atomic_mutex;
235 /* Forward prototypes */
236 static struct rx_call * rxi_NewCall(struct rx_connection *, int);
239 putConnection (struct rx_connection *conn) {
240 MUTEX_ENTER(&rx_refcnt_mutex);
242 MUTEX_EXIT(&rx_refcnt_mutex);
245 #ifdef AFS_PTHREAD_ENV
248 * Use procedural initialization of mutexes/condition variables
252 extern afs_kmutex_t rx_quota_mutex;
253 extern afs_kmutex_t rx_pthread_mutex;
254 extern afs_kmutex_t rx_packets_mutex;
255 extern afs_kmutex_t rx_refcnt_mutex;
256 extern afs_kmutex_t des_init_mutex;
257 extern afs_kmutex_t des_random_mutex;
258 extern afs_kmutex_t rx_clock_mutex;
259 extern afs_kmutex_t rxi_connCacheMutex;
260 extern afs_kmutex_t event_handler_mutex;
261 extern afs_kmutex_t listener_mutex;
262 extern afs_kmutex_t rx_if_init_mutex;
263 extern afs_kmutex_t rx_if_mutex;
265 extern afs_kcondvar_t rx_event_handler_cond;
266 extern afs_kcondvar_t rx_listener_cond;
268 static afs_kmutex_t epoch_mutex;
269 static afs_kmutex_t rx_init_mutex;
270 static afs_kmutex_t rx_debug_mutex;
271 static afs_kmutex_t rx_rpc_stats;
274 rxi_InitPthread(void)
276 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
277 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
278 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
279 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
280 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
281 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
282 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", 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(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
286 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
287 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
288 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
289 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
290 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
292 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
293 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
295 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
296 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
298 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
299 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
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);
314 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
315 #endif /* RX_ENABLE_LOCKS */
318 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
319 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
321 * The rx_stats_mutex mutex protects the following global variables:
322 * rxi_lowConnRefCount
323 * rxi_lowPeerRefCount
332 * The rx_quota_mutex mutex protects the following global variables:
340 * The rx_freePktQ_lock protects the following global variables:
345 * The rx_packets_mutex mutex protects the following global variables:
353 * The rx_pthread_mutex mutex protects the following global variables:
354 * rxi_fcfs_thread_num
357 #define INIT_PTHREAD_LOCKS
361 /* Variables for handling the minProcs implementation. availProcs gives the
362 * number of threads available in the pool at this moment (not counting dudes
363 * executing right now). totalMin gives the total number of procs required
364 * for handling all minProcs requests. minDeficit is a dynamic variable
365 * tracking the # of procs required to satisfy all of the remaining minProcs
367 * For fine grain locking to work, the quota check and the reservation of
368 * a server thread has to come while rxi_availProcs and rxi_minDeficit
369 * are locked. To this end, the code has been modified under #ifdef
370 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
371 * same time. A new function, ReturnToServerPool() returns the allocation.
373 * A call can be on several queue's (but only one at a time). When
374 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
375 * that no one else is touching the queue. To this end, we store the address
376 * of the queue lock in the call structure (under the call lock) when we
377 * put the call on a queue, and we clear the call_queue_lock when the
378 * call is removed from a queue (once the call lock has been obtained).
379 * This allows rxi_ResetCall to safely synchronize with others wishing
380 * to manipulate the queue.
383 #if defined(RX_ENABLE_LOCKS)
384 static afs_kmutex_t rx_rpc_stats;
387 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
388 ** pretty good that the next packet coming in is from the same connection
389 ** as the last packet, since we're send multiple packets in a transmit window.
391 struct rx_connection *rxLastConn = 0;
393 #ifdef RX_ENABLE_LOCKS
394 /* The locking hierarchy for rx fine grain locking is composed of these
397 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
398 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
399 * call->lock - locks call data fields.
400 * These are independent of each other:
401 * rx_freeCallQueue_lock
406 * serverQueueEntry->lock
407 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
409 * peer->lock - locks peer data fields.
410 * conn_data_lock - that more than one thread is not updating a conn data
411 * field at the same time.
422 * Do we need a lock to protect the peer field in the conn structure?
423 * conn->peer was previously a constant for all intents and so has no
424 * lock protecting this field. The multihomed client delta introduced
425 * a RX code change : change the peer field in the connection structure
426 * to that remote interface from which the last packet for this
427 * connection was sent out. This may become an issue if further changes
430 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
431 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
433 /* rxdb_fileID is used to identify the lock location, along with line#. */
434 static int rxdb_fileID = RXDB_FILE_RX;
435 #endif /* RX_LOCKS_DB */
436 #else /* RX_ENABLE_LOCKS */
437 #define SET_CALL_QUEUE_LOCK(C, L)
438 #define CLEAR_CALL_QUEUE_LOCK(C)
439 #endif /* RX_ENABLE_LOCKS */
440 struct rx_serverQueueEntry *rx_waitForPacket = 0;
441 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
443 /* ------------Exported Interfaces------------- */
445 /* This function allows rxkad to set the epoch to a suitably random number
446 * which rx_NewConnection will use in the future. The principle purpose is to
447 * get rxnull connections to use the same epoch as the rxkad connections do, at
448 * least once the first rxkad connection is established. This is important now
449 * that the host/port addresses aren't used in FindConnection: the uniqueness
450 * of epoch/cid matters and the start time won't do. */
452 #ifdef AFS_PTHREAD_ENV
454 * This mutex protects the following global variables:
458 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
459 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
463 #endif /* AFS_PTHREAD_ENV */
466 rx_SetEpoch(afs_uint32 epoch)
473 /* Initialize rx. A port number may be mentioned, in which case this
474 * becomes the default port number for any service installed later.
475 * If 0 is provided for the port number, a random port will be chosen
476 * by the kernel. Whether this will ever overlap anything in
477 * /etc/services is anybody's guess... Returns 0 on success, -1 on
482 int rxinit_status = 1;
483 #ifdef AFS_PTHREAD_ENV
485 * This mutex protects the following global variables:
489 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
490 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
493 #define UNLOCK_RX_INIT
497 rx_InitHost(u_int host, u_int port)
504 char *htable, *ptable;
511 if (rxinit_status == 0) {
512 tmp_status = rxinit_status;
514 return tmp_status; /* Already started; return previous error code. */
520 if (afs_winsockInit() < 0)
526 * Initialize anything necessary to provide a non-premptive threading
529 rxi_InitializeThreadSupport();
532 /* Allocate and initialize a socket for client and perhaps server
535 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
536 if (rx_socket == OSI_NULLSOCKET) {
540 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
543 #endif /* RX_LOCKS_DB */
544 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
545 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
546 MUTEX_INIT(&rx_atomic_mutex, "rx_atomic_mutex", MUTEX_DEFAULT, 0);
547 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
548 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
549 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
550 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
551 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
552 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
553 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
555 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
557 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
559 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
561 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
562 #if defined(AFS_HPUX110_ENV)
564 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
565 #endif /* AFS_HPUX110_ENV */
566 #endif /* RX_ENABLE_LOCKS && KERNEL */
569 rx_connDeadTime = 12;
570 rx_tranquil = 0; /* reset flag */
571 rxi_ResetStatistics();
572 htable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
573 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
574 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
575 ptable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
576 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
577 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
579 /* Malloc up a bunch of packets & buffers */
581 opr_queue_Init(&rx_freePacketQueue);
582 rxi_NeedMorePackets = FALSE;
583 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
585 /* enforce a minimum number of allocated packets */
586 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
587 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
589 /* allocate the initial free packet pool */
590 #ifdef RX_ENABLE_TSFPQ
591 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
592 #else /* RX_ENABLE_TSFPQ */
593 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
594 #endif /* RX_ENABLE_TSFPQ */
601 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
602 tv.tv_sec = clock_now.sec;
603 tv.tv_usec = clock_now.usec;
604 srand((unsigned int)tv.tv_usec);
611 #if defined(KERNEL) && !defined(UKERNEL)
612 /* Really, this should never happen in a real kernel */
615 struct sockaddr_in addr;
617 int addrlen = sizeof(addr);
619 socklen_t addrlen = sizeof(addr);
621 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
623 osi_Free(htable, rx_hashTableSize * sizeof(struct rx_connection *));
626 rx_port = addr.sin_port;
629 rx_stats.minRtt.sec = 9999999;
631 rx_SetEpoch(tv.tv_sec | 0x80000000);
633 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
634 * will provide a randomer value. */
636 MUTEX_ENTER(&rx_quota_mutex);
637 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
638 MUTEX_EXIT(&rx_quota_mutex);
639 /* *Slightly* random start time for the cid. This is just to help
640 * out with the hashing function at the peer */
641 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
642 rx_connHashTable = (struct rx_connection **)htable;
643 rx_peerHashTable = (struct rx_peer **)ptable;
645 rx_hardAckDelay.sec = 0;
646 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
648 rxevent_Init(20, rxi_ReScheduleEvents);
650 /* Initialize various global queues */
651 opr_queue_Init(&rx_idleServerQueue);
652 opr_queue_Init(&rx_incomingCallQueue);
653 opr_queue_Init(&rx_freeCallQueue);
655 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
656 /* Initialize our list of usable IP addresses. */
660 #if defined(RXK_LISTENER_ENV) || !defined(KERNEL)
661 /* Start listener process (exact function is dependent on the
662 * implementation environment--kernel or user space) */
667 tmp_status = rxinit_status = 0;
675 return rx_InitHost(htonl(INADDR_ANY), port);
681 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
682 * maintaing the round trip timer.
687 * Start a new RTT timer for a given call and packet.
689 * There must be no resendEvent already listed for this call, otherwise this
690 * will leak events - intended for internal use within the RTO code only
693 * the RX call to start the timer for
694 * @param[in] lastPacket
695 * a flag indicating whether the last packet has been sent or not
697 * @pre call must be locked before calling this function
701 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
703 struct clock now, retryTime;
708 clock_Add(&retryTime, &call->rto);
710 /* If we're sending the last packet, and we're the client, then the server
711 * may wait for an additional 400ms before returning the ACK, wait for it
712 * rather than hitting a timeout */
713 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
714 clock_Addmsec(&retryTime, 400);
716 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
717 call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
722 * Cancel an RTT timer for a given call.
726 * the RX call to cancel the timer for
728 * @pre call must be locked before calling this function
733 rxi_rto_cancel(struct rx_call *call)
735 rxevent_Cancel(&call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
739 * Tell the RTO timer that we have sent a packet.
741 * If the timer isn't already running, then start it. If the timer is running,
745 * the RX call that the packet has been sent on
746 * @param[in] lastPacket
747 * A flag which is true if this is the last packet for the call
749 * @pre The call must be locked before calling this function
754 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
756 if (call->resendEvent)
759 rxi_rto_startTimer(call, lastPacket, istack);
763 * Tell the RTO timer that we have received an new ACK message
765 * This function should be called whenever a call receives an ACK that
766 * acknowledges new packets. Whatever happens, we stop the current timer.
767 * If there are unacked packets in the queue which have been sent, then
768 * we restart the timer from now. Otherwise, we leave it stopped.
771 * the RX call that the ACK has been received on
775 rxi_rto_packet_acked(struct rx_call *call, int istack)
777 struct opr_queue *cursor;
779 rxi_rto_cancel(call);
781 if (opr_queue_IsEmpty(&call->tq))
784 for (opr_queue_Scan(&call->tq, cursor)) {
785 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
786 if (p->header.seq > call->tfirst + call->twind)
789 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
790 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
798 * Set an initial round trip timeout for a peer connection
800 * @param[in] secs The timeout to set in seconds
804 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
805 peer->rtt = secs * 8000;
809 * Enables or disables the busy call channel error (RX_CALL_BUSY).
811 * @param[in] onoff Non-zero to enable busy call channel errors.
813 * @pre Neither rx_Init nor rx_InitHost have been called yet
816 rx_SetBusyChannelError(afs_int32 onoff)
818 osi_Assert(rxinit_status != 0);
819 rxi_busyChannelError = onoff ? 1 : 0;
823 * Set a delayed ack event on the specified call for the given time
825 * @param[in] call - the call on which to set the event
826 * @param[in] offset - the delay from now after which the event fires
829 rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
831 struct clock now, when;
835 clock_Add(&when, offset);
837 if (!call->delayedAckEvent
838 || clock_Gt(&call->delayedAckTime, &when)) {
840 rxevent_Cancel(&call->delayedAckEvent, call,
841 RX_CALL_REFCOUNT_DELAY);
842 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
844 call->delayedAckEvent = rxevent_Post(&when, &now,
847 call->delayedAckTime = when;
851 /* called with unincremented nRequestsRunning to see if it is OK to start
852 * a new thread in this service. Could be "no" for two reasons: over the
853 * max quota, or would prevent others from reaching their min quota.
855 #ifdef RX_ENABLE_LOCKS
856 /* This verion of QuotaOK reserves quota if it's ok while the
857 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
860 QuotaOK(struct rx_service *aservice)
862 /* check if over max quota */
863 if (aservice->nRequestsRunning >= aservice->maxProcs) {
867 /* under min quota, we're OK */
868 /* otherwise, can use only if there are enough to allow everyone
869 * to go to their min quota after this guy starts.
872 MUTEX_ENTER(&rx_quota_mutex);
873 if ((aservice->nRequestsRunning < aservice->minProcs)
874 || (rxi_availProcs > rxi_minDeficit)) {
875 aservice->nRequestsRunning++;
876 /* just started call in minProcs pool, need fewer to maintain
878 if (aservice->nRequestsRunning <= aservice->minProcs)
881 MUTEX_EXIT(&rx_quota_mutex);
884 MUTEX_EXIT(&rx_quota_mutex);
890 ReturnToServerPool(struct rx_service *aservice)
892 aservice->nRequestsRunning--;
893 MUTEX_ENTER(&rx_quota_mutex);
894 if (aservice->nRequestsRunning < aservice->minProcs)
897 MUTEX_EXIT(&rx_quota_mutex);
900 #else /* RX_ENABLE_LOCKS */
902 QuotaOK(struct rx_service *aservice)
905 /* under min quota, we're OK */
906 if (aservice->nRequestsRunning < aservice->minProcs)
909 /* check if over max quota */
910 if (aservice->nRequestsRunning >= aservice->maxProcs)
913 /* otherwise, can use only if there are enough to allow everyone
914 * to go to their min quota after this guy starts.
916 MUTEX_ENTER(&rx_quota_mutex);
917 if (rxi_availProcs > rxi_minDeficit)
919 MUTEX_EXIT(&rx_quota_mutex);
922 #endif /* RX_ENABLE_LOCKS */
925 /* Called by rx_StartServer to start up lwp's to service calls.
926 NExistingProcs gives the number of procs already existing, and which
927 therefore needn't be created. */
929 rxi_StartServerProcs(int nExistingProcs)
931 struct rx_service *service;
936 /* For each service, reserve N processes, where N is the "minimum"
937 * number of processes that MUST be able to execute a request in parallel,
938 * at any time, for that process. Also compute the maximum difference
939 * between any service's maximum number of processes that can run
940 * (i.e. the maximum number that ever will be run, and a guarantee
941 * that this number will run if other services aren't running), and its
942 * minimum number. The result is the extra number of processes that
943 * we need in order to provide the latter guarantee */
944 for (i = 0; i < RX_MAX_SERVICES; i++) {
946 service = rx_services[i];
947 if (service == (struct rx_service *)0)
949 nProcs += service->minProcs;
950 diff = service->maxProcs - service->minProcs;
954 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
955 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
956 for (i = 0; i < nProcs; i++) {
957 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
963 /* This routine is only required on Windows */
965 rx_StartClientThread(void)
967 #ifdef AFS_PTHREAD_ENV
969 pid = pthread_self();
970 #endif /* AFS_PTHREAD_ENV */
972 #endif /* AFS_NT40_ENV */
974 /* This routine must be called if any services are exported. If the
975 * donateMe flag is set, the calling process is donated to the server
978 rx_StartServer(int donateMe)
980 struct rx_service *service;
986 /* Start server processes, if necessary (exact function is dependent
987 * on the implementation environment--kernel or user space). DonateMe
988 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
989 * case, one less new proc will be created rx_StartServerProcs.
991 rxi_StartServerProcs(donateMe);
993 /* count up the # of threads in minProcs, and add set the min deficit to
994 * be that value, too.
996 for (i = 0; i < RX_MAX_SERVICES; i++) {
997 service = rx_services[i];
998 if (service == (struct rx_service *)0)
1000 MUTEX_ENTER(&rx_quota_mutex);
1001 rxi_totalMin += service->minProcs;
1002 /* below works even if a thread is running, since minDeficit would
1003 * still have been decremented and later re-incremented.
1005 rxi_minDeficit += service->minProcs;
1006 MUTEX_EXIT(&rx_quota_mutex);
1009 /* Turn on reaping of idle server connections */
1010 rxi_ReapConnections(NULL, NULL, NULL, 0);
1015 #ifndef AFS_NT40_ENV
1019 #ifdef AFS_PTHREAD_ENV
1021 pid = afs_pointer_to_int(pthread_self());
1022 #else /* AFS_PTHREAD_ENV */
1024 LWP_CurrentProcess(&pid);
1025 #endif /* AFS_PTHREAD_ENV */
1027 sprintf(name, "srv_%d", ++nProcs);
1028 if (registerProgram)
1029 (*registerProgram) (pid, name);
1031 #endif /* AFS_NT40_ENV */
1032 rx_ServerProc(NULL); /* Never returns */
1034 #ifdef RX_ENABLE_TSFPQ
1035 /* no use leaving packets around in this thread's local queue if
1036 * it isn't getting donated to the server thread pool.
1038 rxi_FlushLocalPacketsTSFPQ();
1039 #endif /* RX_ENABLE_TSFPQ */
1043 /* Create a new client connection to the specified service, using the
1044 * specified security object to implement the security model for this
1046 struct rx_connection *
1047 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
1048 struct rx_securityClass *securityObject,
1049 int serviceSecurityIndex)
1053 struct rx_connection *conn;
1058 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
1059 "serviceSecurityIndex %d)\n",
1060 ntohl(shost), ntohs(sport), sservice, securityObject,
1061 serviceSecurityIndex));
1063 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
1064 * the case of kmem_alloc? */
1065 conn = rxi_AllocConnection();
1066 #ifdef RX_ENABLE_LOCKS
1067 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
1068 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
1069 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
1072 MUTEX_ENTER(&rx_connHashTable_lock);
1073 cid = (rx_nextCid += RX_MAXCALLS);
1074 conn->type = RX_CLIENT_CONNECTION;
1076 conn->epoch = rx_epoch;
1077 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
1078 conn->serviceId = sservice;
1079 conn->securityObject = securityObject;
1080 conn->securityData = (void *) 0;
1081 conn->securityIndex = serviceSecurityIndex;
1082 rx_SetConnDeadTime(conn, rx_connDeadTime);
1083 rx_SetConnSecondsUntilNatPing(conn, 0);
1084 conn->ackRate = RX_FAST_ACK_RATE;
1085 conn->nSpecific = 0;
1086 conn->specific = NULL;
1087 conn->challengeEvent = NULL;
1088 conn->delayedAbortEvent = NULL;
1089 conn->abortCount = 0;
1091 for (i = 0; i < RX_MAXCALLS; i++) {
1092 conn->twind[i] = rx_initSendWindow;
1093 conn->rwind[i] = rx_initReceiveWindow;
1094 conn->lastBusy[i] = 0;
1097 RXS_NewConnection(securityObject, conn);
1099 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1101 conn->refCount++; /* no lock required since only this thread knows... */
1102 conn->next = rx_connHashTable[hashindex];
1103 rx_connHashTable[hashindex] = conn;
1104 if (rx_stats_active)
1105 rx_atomic_inc(&rx_stats.nClientConns);
1106 MUTEX_EXIT(&rx_connHashTable_lock);
1112 * Ensure a connection's timeout values are valid.
1114 * @param[in] conn The connection to check
1116 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1117 * unless idleDeadTime and/or hardDeadTime are not set
1121 rxi_CheckConnTimeouts(struct rx_connection *conn)
1123 /* a connection's timeouts must have the relationship
1124 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1125 * total loss of network to a peer may cause an idle timeout instead of a
1126 * dead timeout, simply because the idle timeout gets hit first. Also set
1127 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1128 /* this logic is slightly complicated by the fact that
1129 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1131 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1132 if (conn->idleDeadTime) {
1133 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1135 if (conn->hardDeadTime) {
1136 if (conn->idleDeadTime) {
1137 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1139 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1145 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1147 /* The idea is to set the dead time to a value that allows several
1148 * keepalives to be dropped without timing out the connection. */
1149 conn->secondsUntilDead = seconds;
1150 rxi_CheckConnTimeouts(conn);
1151 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1155 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1157 conn->hardDeadTime = seconds;
1158 rxi_CheckConnTimeouts(conn);
1162 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1164 conn->idleDeadTime = seconds;
1165 conn->idleDeadDetection = (seconds ? 1 : 0);
1166 rxi_CheckConnTimeouts(conn);
1169 int rxi_lowPeerRefCount = 0;
1170 int rxi_lowConnRefCount = 0;
1173 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1174 * NOTE: must not be called with rx_connHashTable_lock held.
1177 rxi_CleanupConnection(struct rx_connection *conn)
1179 /* Notify the service exporter, if requested, that this connection
1180 * is being destroyed */
1181 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1182 (*conn->service->destroyConnProc) (conn);
1184 /* Notify the security module that this connection is being destroyed */
1185 RXS_DestroyConnection(conn->securityObject, conn);
1187 /* If this is the last connection using the rx_peer struct, set its
1188 * idle time to now. rxi_ReapConnections will reap it if it's still
1189 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1191 MUTEX_ENTER(&rx_peerHashTable_lock);
1192 if (conn->peer->refCount < 2) {
1193 conn->peer->idleWhen = clock_Sec();
1194 if (conn->peer->refCount < 1) {
1195 conn->peer->refCount = 1;
1196 if (rx_stats_active) {
1197 MUTEX_ENTER(&rx_stats_mutex);
1198 rxi_lowPeerRefCount++;
1199 MUTEX_EXIT(&rx_stats_mutex);
1203 conn->peer->refCount--;
1204 MUTEX_EXIT(&rx_peerHashTable_lock);
1206 if (rx_stats_active)
1208 if (conn->type == RX_SERVER_CONNECTION)
1209 rx_atomic_dec(&rx_stats.nServerConns);
1211 rx_atomic_dec(&rx_stats.nClientConns);
1214 if (conn->specific) {
1216 for (i = 0; i < conn->nSpecific; i++) {
1217 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1218 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1219 conn->specific[i] = NULL;
1221 free(conn->specific);
1223 conn->specific = NULL;
1224 conn->nSpecific = 0;
1225 #endif /* !KERNEL */
1227 MUTEX_DESTROY(&conn->conn_call_lock);
1228 MUTEX_DESTROY(&conn->conn_data_lock);
1229 CV_DESTROY(&conn->conn_call_cv);
1231 rxi_FreeConnection(conn);
1234 /* Destroy the specified connection */
1236 rxi_DestroyConnection(struct rx_connection *conn)
1238 MUTEX_ENTER(&rx_connHashTable_lock);
1239 rxi_DestroyConnectionNoLock(conn);
1240 /* conn should be at the head of the cleanup list */
1241 if (conn == rx_connCleanup_list) {
1242 rx_connCleanup_list = rx_connCleanup_list->next;
1243 MUTEX_EXIT(&rx_connHashTable_lock);
1244 rxi_CleanupConnection(conn);
1246 #ifdef RX_ENABLE_LOCKS
1248 MUTEX_EXIT(&rx_connHashTable_lock);
1250 #endif /* RX_ENABLE_LOCKS */
1254 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1256 struct rx_connection **conn_ptr;
1258 struct rx_packet *packet;
1265 MUTEX_ENTER(&conn->conn_data_lock);
1266 MUTEX_ENTER(&rx_refcnt_mutex);
1267 if (conn->refCount > 0)
1270 if (rx_stats_active) {
1271 MUTEX_ENTER(&rx_stats_mutex);
1272 rxi_lowConnRefCount++;
1273 MUTEX_EXIT(&rx_stats_mutex);
1277 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1278 /* Busy; wait till the last guy before proceeding */
1279 MUTEX_EXIT(&rx_refcnt_mutex);
1280 MUTEX_EXIT(&conn->conn_data_lock);
1285 /* If the client previously called rx_NewCall, but it is still
1286 * waiting, treat this as a running call, and wait to destroy the
1287 * connection later when the call completes. */
1288 if ((conn->type == RX_CLIENT_CONNECTION)
1289 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1290 conn->flags |= RX_CONN_DESTROY_ME;
1291 MUTEX_EXIT(&conn->conn_data_lock);
1295 MUTEX_EXIT(&rx_refcnt_mutex);
1296 MUTEX_EXIT(&conn->conn_data_lock);
1298 /* Check for extant references to this connection */
1299 MUTEX_ENTER(&conn->conn_call_lock);
1300 for (i = 0; i < RX_MAXCALLS; i++) {
1301 struct rx_call *call = conn->call[i];
1304 if (conn->type == RX_CLIENT_CONNECTION) {
1305 MUTEX_ENTER(&call->lock);
1306 if (call->delayedAckEvent) {
1307 /* Push the final acknowledgment out now--there
1308 * won't be a subsequent call to acknowledge the
1309 * last reply packets */
1310 rxevent_Cancel(&call->delayedAckEvent, call,
1311 RX_CALL_REFCOUNT_DELAY);
1312 if (call->state == RX_STATE_PRECALL
1313 || call->state == RX_STATE_ACTIVE) {
1314 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1319 MUTEX_EXIT(&call->lock);
1323 MUTEX_EXIT(&conn->conn_call_lock);
1325 #ifdef RX_ENABLE_LOCKS
1327 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1328 MUTEX_EXIT(&conn->conn_data_lock);
1330 /* Someone is accessing a packet right now. */
1334 #endif /* RX_ENABLE_LOCKS */
1337 /* Don't destroy the connection if there are any call
1338 * structures still in use */
1339 MUTEX_ENTER(&conn->conn_data_lock);
1340 conn->flags |= RX_CONN_DESTROY_ME;
1341 MUTEX_EXIT(&conn->conn_data_lock);
1346 if (conn->natKeepAliveEvent) {
1347 rxi_NatKeepAliveOff(conn);
1350 if (conn->delayedAbortEvent) {
1351 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
1352 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1354 MUTEX_ENTER(&conn->conn_data_lock);
1355 rxi_SendConnectionAbort(conn, packet, 0, 1);
1356 MUTEX_EXIT(&conn->conn_data_lock);
1357 rxi_FreePacket(packet);
1361 /* Remove from connection hash table before proceeding */
1363 &rx_connHashTable[CONN_HASH
1364 (peer->host, peer->port, conn->cid, conn->epoch,
1366 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1367 if (*conn_ptr == conn) {
1368 *conn_ptr = conn->next;
1372 /* if the conn that we are destroying was the last connection, then we
1373 * clear rxLastConn as well */
1374 if (rxLastConn == conn)
1377 /* Make sure the connection is completely reset before deleting it. */
1378 /* get rid of pending events that could zap us later */
1379 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
1380 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
1381 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
1383 /* Add the connection to the list of destroyed connections that
1384 * need to be cleaned up. This is necessary to avoid deadlocks
1385 * in the routines we call to inform others that this connection is
1386 * being destroyed. */
1387 conn->next = rx_connCleanup_list;
1388 rx_connCleanup_list = conn;
1391 /* Externally available version */
1393 rx_DestroyConnection(struct rx_connection *conn)
1398 rxi_DestroyConnection(conn);
1403 rx_GetConnection(struct rx_connection *conn)
1408 MUTEX_ENTER(&rx_refcnt_mutex);
1410 MUTEX_EXIT(&rx_refcnt_mutex);
1414 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1415 /* Wait for the transmit queue to no longer be busy.
1416 * requires the call->lock to be held */
1418 rxi_WaitforTQBusy(struct rx_call *call) {
1419 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1420 call->flags |= RX_CALL_TQ_WAIT;
1422 #ifdef RX_ENABLE_LOCKS
1423 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1424 CV_WAIT(&call->cv_tq, &call->lock);
1425 #else /* RX_ENABLE_LOCKS */
1426 osi_rxSleep(&call->tq);
1427 #endif /* RX_ENABLE_LOCKS */
1429 if (call->tqWaiters == 0) {
1430 call->flags &= ~RX_CALL_TQ_WAIT;
1437 rxi_WakeUpTransmitQueue(struct rx_call *call)
1439 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1440 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1441 call, call->tqWaiters, call->flags));
1442 #ifdef RX_ENABLE_LOCKS
1443 osirx_AssertMine(&call->lock, "rxi_Start start");
1444 CV_BROADCAST(&call->cv_tq);
1445 #else /* RX_ENABLE_LOCKS */
1446 osi_rxWakeup(&call->tq);
1447 #endif /* RX_ENABLE_LOCKS */
1451 /* Start a new rx remote procedure call, on the specified connection.
1452 * If wait is set to 1, wait for a free call channel; otherwise return
1453 * 0. Maxtime gives the maximum number of seconds this call may take,
1454 * after rx_NewCall returns. After this time interval, a call to any
1455 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1456 * For fine grain locking, we hold the conn_call_lock in order to
1457 * to ensure that we don't get signalle after we found a call in an active
1458 * state and before we go to sleep.
1461 rx_NewCall(struct rx_connection *conn)
1463 int i, wait, ignoreBusy = 1;
1464 struct rx_call *call;
1465 struct clock queueTime;
1466 afs_uint32 leastBusy = 0;
1470 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1473 clock_GetTime(&queueTime);
1475 * Check if there are others waiting for a new call.
1476 * If so, let them go first to avoid starving them.
1477 * This is a fairly simple scheme, and might not be
1478 * a complete solution for large numbers of waiters.
1480 * makeCallWaiters keeps track of the number of
1481 * threads waiting to make calls and the
1482 * RX_CONN_MAKECALL_WAITING flag bit is used to
1483 * indicate that there are indeed calls waiting.
1484 * The flag is set when the waiter is incremented.
1485 * It is only cleared when makeCallWaiters is 0.
1486 * This prevents us from accidently destroying the
1487 * connection while it is potentially about to be used.
1489 MUTEX_ENTER(&conn->conn_call_lock);
1490 MUTEX_ENTER(&conn->conn_data_lock);
1491 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1492 conn->flags |= RX_CONN_MAKECALL_WAITING;
1493 conn->makeCallWaiters++;
1494 MUTEX_EXIT(&conn->conn_data_lock);
1496 #ifdef RX_ENABLE_LOCKS
1497 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1501 MUTEX_ENTER(&conn->conn_data_lock);
1502 conn->makeCallWaiters--;
1503 if (conn->makeCallWaiters == 0)
1504 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1507 /* We are now the active thread in rx_NewCall */
1508 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1509 MUTEX_EXIT(&conn->conn_data_lock);
1514 for (i = 0; i < RX_MAXCALLS; i++) {
1515 call = conn->call[i];
1517 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1518 /* we're not ignoring busy call slots; only look at the
1519 * call slot that is the "least" busy */
1523 if (call->state == RX_STATE_DALLY) {
1524 MUTEX_ENTER(&call->lock);
1525 if (call->state == RX_STATE_DALLY) {
1526 if (ignoreBusy && conn->lastBusy[i]) {
1527 /* if we're ignoring busy call slots, skip any ones that
1528 * have lastBusy set */
1529 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1530 leastBusy = conn->lastBusy[i];
1532 MUTEX_EXIT(&call->lock);
1537 * We are setting the state to RX_STATE_RESET to
1538 * ensure that no one else will attempt to use this
1539 * call once we drop the conn->conn_call_lock and
1540 * call->lock. We must drop the conn->conn_call_lock
1541 * before calling rxi_ResetCall because the process
1542 * of clearing the transmit queue can block for an
1543 * extended period of time. If we block while holding
1544 * the conn->conn_call_lock, then all rx_EndCall
1545 * processing will block as well. This has a detrimental
1546 * effect on overall system performance.
1548 call->state = RX_STATE_RESET;
1549 (*call->callNumber)++;
1550 MUTEX_EXIT(&conn->conn_call_lock);
1551 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1552 rxi_ResetCall(call, 0);
1553 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1557 * If we failed to be able to safely obtain the
1558 * conn->conn_call_lock we will have to drop the
1559 * call->lock to avoid a deadlock. When the call->lock
1560 * is released the state of the call can change. If it
1561 * is no longer RX_STATE_RESET then some other thread is
1564 MUTEX_EXIT(&call->lock);
1565 MUTEX_ENTER(&conn->conn_call_lock);
1566 MUTEX_ENTER(&call->lock);
1568 if (call->state == RX_STATE_RESET)
1572 * If we get here it means that after dropping
1573 * the conn->conn_call_lock and call->lock that
1574 * the call is no longer ours. If we can't find
1575 * a free call in the remaining slots we should
1576 * not go immediately to RX_CONN_MAKECALL_WAITING
1577 * because by dropping the conn->conn_call_lock
1578 * we have given up synchronization with rx_EndCall.
1579 * Instead, cycle through one more time to see if
1580 * we can find a call that can call our own.
1582 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1585 MUTEX_EXIT(&call->lock);
1588 if (ignoreBusy && conn->lastBusy[i]) {
1589 /* if we're ignoring busy call slots, skip any ones that
1590 * have lastBusy set */
1591 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1592 leastBusy = conn->lastBusy[i];
1597 /* rxi_NewCall returns with mutex locked */
1598 call = rxi_NewCall(conn, i);
1599 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1603 if (i < RX_MAXCALLS) {
1604 conn->lastBusy[i] = 0;
1605 call->flags &= ~RX_CALL_PEER_BUSY;
1610 if (leastBusy && ignoreBusy) {
1611 /* we didn't find a useable call slot, but we did see at least one
1612 * 'busy' slot; look again and only use a slot with the 'least
1618 MUTEX_ENTER(&conn->conn_data_lock);
1619 conn->flags |= RX_CONN_MAKECALL_WAITING;
1620 conn->makeCallWaiters++;
1621 MUTEX_EXIT(&conn->conn_data_lock);
1623 #ifdef RX_ENABLE_LOCKS
1624 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1628 MUTEX_ENTER(&conn->conn_data_lock);
1629 conn->makeCallWaiters--;
1630 if (conn->makeCallWaiters == 0)
1631 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1632 MUTEX_EXIT(&conn->conn_data_lock);
1634 /* Client is initially in send mode */
1635 call->state = RX_STATE_ACTIVE;
1636 call->error = conn->error;
1638 call->mode = RX_MODE_ERROR;
1640 call->mode = RX_MODE_SENDING;
1642 #ifdef AFS_RXERRQ_ENV
1643 /* remember how many network errors the peer has when we started, so if
1644 * more errors are encountered after the call starts, we know the other endpoint won't be
1645 * responding to us */
1646 call->neterr_gen = rx_atomic_read(&conn->peer->neterrs);
1649 /* remember start time for call in case we have hard dead time limit */
1650 call->queueTime = queueTime;
1651 clock_GetTime(&call->startTime);
1652 call->bytesSent = 0;
1653 call->bytesRcvd = 0;
1655 /* Turn on busy protocol. */
1656 rxi_KeepAliveOn(call);
1658 /* Attempt MTU discovery */
1659 rxi_GrowMTUOn(call);
1662 * We are no longer the active thread in rx_NewCall
1664 MUTEX_ENTER(&conn->conn_data_lock);
1665 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1666 MUTEX_EXIT(&conn->conn_data_lock);
1669 * Wake up anyone else who might be giving us a chance to
1670 * run (see code above that avoids resource starvation).
1672 #ifdef RX_ENABLE_LOCKS
1673 CV_BROADCAST(&conn->conn_call_cv);
1677 MUTEX_EXIT(&conn->conn_call_lock);
1679 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1680 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1681 osi_Panic("rx_NewCall call about to be used without an empty tq");
1683 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1685 MUTEX_EXIT(&call->lock);
1688 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1693 rxi_HasActiveCalls(struct rx_connection *aconn)
1696 struct rx_call *tcall;
1700 for (i = 0; i < RX_MAXCALLS; i++) {
1701 if ((tcall = aconn->call[i])) {
1702 if ((tcall->state == RX_STATE_ACTIVE)
1703 || (tcall->state == RX_STATE_PRECALL)) {
1714 rxi_GetCallNumberVector(struct rx_connection *aconn,
1715 afs_int32 * aint32s)
1718 struct rx_call *tcall;
1722 MUTEX_ENTER(&aconn->conn_call_lock);
1723 for (i = 0; i < RX_MAXCALLS; i++) {
1724 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1725 aint32s[i] = aconn->callNumber[i] + 1;
1727 aint32s[i] = aconn->callNumber[i];
1729 MUTEX_EXIT(&aconn->conn_call_lock);
1735 rxi_SetCallNumberVector(struct rx_connection *aconn,
1736 afs_int32 * aint32s)
1739 struct rx_call *tcall;
1743 MUTEX_ENTER(&aconn->conn_call_lock);
1744 for (i = 0; i < RX_MAXCALLS; i++) {
1745 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1746 aconn->callNumber[i] = aint32s[i] - 1;
1748 aconn->callNumber[i] = aint32s[i];
1750 MUTEX_EXIT(&aconn->conn_call_lock);
1755 /* Advertise a new service. A service is named locally by a UDP port
1756 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1759 char *serviceName; Name for identification purposes (e.g. the
1760 service name might be used for probing for
1763 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1764 char *serviceName, struct rx_securityClass **securityObjects,
1765 int nSecurityObjects,
1766 afs_int32(*serviceProc) (struct rx_call * acall))
1768 osi_socket socket = OSI_NULLSOCKET;
1769 struct rx_service *tservice;
1775 if (serviceId == 0) {
1777 "rx_NewService: service id for service %s is not non-zero.\n",
1784 "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",
1792 tservice = rxi_AllocService();
1795 #ifdef RX_ENABLE_LOCKS
1796 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1799 for (i = 0; i < RX_MAX_SERVICES; i++) {
1800 struct rx_service *service = rx_services[i];
1802 if (port == service->servicePort && host == service->serviceHost) {
1803 if (service->serviceId == serviceId) {
1804 /* The identical service has already been
1805 * installed; if the caller was intending to
1806 * change the security classes used by this
1807 * service, he/she loses. */
1809 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1810 serviceName, serviceId, service->serviceName);
1812 rxi_FreeService(tservice);
1815 /* Different service, same port: re-use the socket
1816 * which is bound to the same port */
1817 socket = service->socket;
1820 if (socket == OSI_NULLSOCKET) {
1821 /* If we don't already have a socket (from another
1822 * service on same port) get a new one */
1823 socket = rxi_GetHostUDPSocket(host, port);
1824 if (socket == OSI_NULLSOCKET) {
1826 rxi_FreeService(tservice);
1831 service->socket = socket;
1832 service->serviceHost = host;
1833 service->servicePort = port;
1834 service->serviceId = serviceId;
1835 service->serviceName = serviceName;
1836 service->nSecurityObjects = nSecurityObjects;
1837 service->securityObjects = securityObjects;
1838 service->minProcs = 0;
1839 service->maxProcs = 1;
1840 service->idleDeadTime = 60;
1841 service->idleDeadErr = 0;
1842 service->connDeadTime = rx_connDeadTime;
1843 service->executeRequestProc = serviceProc;
1844 service->checkReach = 0;
1845 service->nSpecific = 0;
1846 service->specific = NULL;
1847 rx_services[i] = service; /* not visible until now */
1853 rxi_FreeService(tservice);
1854 (osi_Msg "rx_NewService: cannot support > %d services\n",
1859 /* Set configuration options for all of a service's security objects */
1862 rx_SetSecurityConfiguration(struct rx_service *service,
1863 rx_securityConfigVariables type,
1867 for (i = 0; i<service->nSecurityObjects; i++) {
1868 if (service->securityObjects[i]) {
1869 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1877 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1878 struct rx_securityClass **securityObjects, int nSecurityObjects,
1879 afs_int32(*serviceProc) (struct rx_call * acall))
1881 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1884 /* Generic request processing loop. This routine should be called
1885 * by the implementation dependent rx_ServerProc. If socketp is
1886 * non-null, it will be set to the file descriptor that this thread
1887 * is now listening on. If socketp is null, this routine will never
1890 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1892 struct rx_call *call;
1894 struct rx_service *tservice = NULL;
1901 call = rx_GetCall(threadID, tservice, socketp);
1902 if (socketp && *socketp != OSI_NULLSOCKET) {
1903 /* We are now a listener thread */
1909 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1910 #ifdef RX_ENABLE_LOCKS
1912 #endif /* RX_ENABLE_LOCKS */
1913 afs_termState = AFSOP_STOP_AFS;
1914 afs_osi_Wakeup(&afs_termState);
1915 #ifdef RX_ENABLE_LOCKS
1917 #endif /* RX_ENABLE_LOCKS */
1922 /* if server is restarting( typically smooth shutdown) then do not
1923 * allow any new calls.
1926 if (rx_tranquil && (call != NULL)) {
1930 MUTEX_ENTER(&call->lock);
1932 rxi_CallError(call, RX_RESTARTING);
1933 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1935 MUTEX_EXIT(&call->lock);
1940 tservice = call->conn->service;
1942 if (tservice->beforeProc)
1943 (*tservice->beforeProc) (call);
1945 code = tservice->executeRequestProc(call);
1947 if (tservice->afterProc)
1948 (*tservice->afterProc) (call, code);
1950 rx_EndCall(call, code);
1952 if (tservice->postProc)
1953 (*tservice->postProc) (code);
1955 if (rx_stats_active) {
1956 MUTEX_ENTER(&rx_stats_mutex);
1958 MUTEX_EXIT(&rx_stats_mutex);
1965 rx_WakeupServerProcs(void)
1967 struct rx_serverQueueEntry *np, *tqp;
1968 struct opr_queue *cursor;
1972 MUTEX_ENTER(&rx_serverPool_lock);
1974 #ifdef RX_ENABLE_LOCKS
1975 if (rx_waitForPacket)
1976 CV_BROADCAST(&rx_waitForPacket->cv);
1977 #else /* RX_ENABLE_LOCKS */
1978 if (rx_waitForPacket)
1979 osi_rxWakeup(rx_waitForPacket);
1980 #endif /* RX_ENABLE_LOCKS */
1981 MUTEX_ENTER(&freeSQEList_lock);
1982 for (np = rx_FreeSQEList; np; np = tqp) {
1983 tqp = *(struct rx_serverQueueEntry **)np;
1984 #ifdef RX_ENABLE_LOCKS
1985 CV_BROADCAST(&np->cv);
1986 #else /* RX_ENABLE_LOCKS */
1988 #endif /* RX_ENABLE_LOCKS */
1990 MUTEX_EXIT(&freeSQEList_lock);
1991 for (opr_queue_Scan(&rx_idleServerQueue, cursor)) {
1992 np = opr_queue_Entry(cursor, struct rx_serverQueueEntry, entry);
1993 #ifdef RX_ENABLE_LOCKS
1994 CV_BROADCAST(&np->cv);
1995 #else /* RX_ENABLE_LOCKS */
1997 #endif /* RX_ENABLE_LOCKS */
1999 MUTEX_EXIT(&rx_serverPool_lock);
2004 * One thing that seems to happen is that all the server threads get
2005 * tied up on some empty or slow call, and then a whole bunch of calls
2006 * arrive at once, using up the packet pool, so now there are more
2007 * empty calls. The most critical resources here are server threads
2008 * and the free packet pool. The "doreclaim" code seems to help in
2009 * general. I think that eventually we arrive in this state: there
2010 * are lots of pending calls which do have all their packets present,
2011 * so they won't be reclaimed, are multi-packet calls, so they won't
2012 * be scheduled until later, and thus are tying up most of the free
2013 * packet pool for a very long time.
2015 * 1. schedule multi-packet calls if all the packets are present.
2016 * Probably CPU-bound operation, useful to return packets to pool.
2017 * Do what if there is a full window, but the last packet isn't here?
2018 * 3. preserve one thread which *only* runs "best" calls, otherwise
2019 * it sleeps and waits for that type of call.
2020 * 4. Don't necessarily reserve a whole window for each thread. In fact,
2021 * the current dataquota business is badly broken. The quota isn't adjusted
2022 * to reflect how many packets are presently queued for a running call.
2023 * So, when we schedule a queued call with a full window of packets queued
2024 * up for it, that *should* free up a window full of packets for other 2d-class
2025 * calls to be able to use from the packet pool. But it doesn't.
2027 * NB. Most of the time, this code doesn't run -- since idle server threads
2028 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
2029 * as a new call arrives.
2031 /* Sleep until a call arrives. Returns a pointer to the call, ready
2032 * for an rx_Read. */
2033 #ifdef RX_ENABLE_LOCKS
2035 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2037 struct rx_serverQueueEntry *sq;
2038 struct rx_call *call = (struct rx_call *)0;
2039 struct rx_service *service = NULL;
2041 MUTEX_ENTER(&freeSQEList_lock);
2043 if ((sq = rx_FreeSQEList)) {
2044 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2045 MUTEX_EXIT(&freeSQEList_lock);
2046 } else { /* otherwise allocate a new one and return that */
2047 MUTEX_EXIT(&freeSQEList_lock);
2048 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2049 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2050 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2053 MUTEX_ENTER(&rx_serverPool_lock);
2054 if (cur_service != NULL) {
2055 ReturnToServerPool(cur_service);
2058 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2059 struct rx_call *tcall, *choice2 = NULL;
2060 struct opr_queue *cursor;
2062 /* Scan for eligible incoming calls. A call is not eligible
2063 * if the maximum number of calls for its service type are
2064 * already executing */
2065 /* One thread will process calls FCFS (to prevent starvation),
2066 * while the other threads may run ahead looking for calls which
2067 * have all their input data available immediately. This helps
2068 * keep threads from blocking, waiting for data from the client. */
2069 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2070 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2072 service = tcall->conn->service;
2073 if (!QuotaOK(service)) {
2076 MUTEX_ENTER(&rx_pthread_mutex);
2077 if (tno == rxi_fcfs_thread_num
2078 || opr_queue_IsEnd(&rx_incomingCallQueue, cursor)) {
2079 MUTEX_EXIT(&rx_pthread_mutex);
2080 /* If we're the fcfs thread , then we'll just use
2081 * this call. If we haven't been able to find an optimal
2082 * choice, and we're at the end of the list, then use a
2083 * 2d choice if one has been identified. Otherwise... */
2084 call = (choice2 ? choice2 : tcall);
2085 service = call->conn->service;
2087 MUTEX_EXIT(&rx_pthread_mutex);
2088 if (!opr_queue_IsEmpty(&tcall->rq)) {
2089 struct rx_packet *rp;
2090 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2092 if (rp->header.seq == 1) {
2094 || (rp->header.flags & RX_LAST_PACKET)) {
2096 } else if (rxi_2dchoice && !choice2
2097 && !(tcall->flags & RX_CALL_CLEARED)
2098 && (tcall->rprev > rxi_HardAckRate)) {
2108 ReturnToServerPool(service);
2114 opr_queue_Remove(&call->entry);
2115 MUTEX_EXIT(&rx_serverPool_lock);
2116 MUTEX_ENTER(&call->lock);
2118 if (call->flags & RX_CALL_WAIT_PROC) {
2119 call->flags &= ~RX_CALL_WAIT_PROC;
2120 rx_atomic_dec(&rx_nWaiting);
2123 if (call->state != RX_STATE_PRECALL || call->error) {
2124 MUTEX_EXIT(&call->lock);
2125 MUTEX_ENTER(&rx_serverPool_lock);
2126 ReturnToServerPool(service);
2131 if (opr_queue_IsEmpty(&call->rq)
2132 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1)
2133 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2135 CLEAR_CALL_QUEUE_LOCK(call);
2138 /* If there are no eligible incoming calls, add this process
2139 * to the idle server queue, to wait for one */
2143 *socketp = OSI_NULLSOCKET;
2145 sq->socketp = socketp;
2146 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2147 #ifndef AFS_AIX41_ENV
2148 rx_waitForPacket = sq;
2150 rx_waitingForPacket = sq;
2151 #endif /* AFS_AIX41_ENV */
2153 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2155 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2156 MUTEX_EXIT(&rx_serverPool_lock);
2157 return (struct rx_call *)0;
2160 } while (!(call = sq->newcall)
2161 && !(socketp && *socketp != OSI_NULLSOCKET));
2162 MUTEX_EXIT(&rx_serverPool_lock);
2164 MUTEX_ENTER(&call->lock);
2170 MUTEX_ENTER(&freeSQEList_lock);
2171 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2172 rx_FreeSQEList = sq;
2173 MUTEX_EXIT(&freeSQEList_lock);
2176 clock_GetTime(&call->startTime);
2177 call->state = RX_STATE_ACTIVE;
2178 call->mode = RX_MODE_RECEIVING;
2179 #ifdef RX_KERNEL_TRACE
2180 if (ICL_SETACTIVE(afs_iclSetp)) {
2181 int glockOwner = ISAFS_GLOCK();
2184 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2185 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2192 rxi_calltrace(RX_CALL_START, call);
2193 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2194 call->conn->service->servicePort, call->conn->service->serviceId,
2197 MUTEX_EXIT(&call->lock);
2198 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2200 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2205 #else /* RX_ENABLE_LOCKS */
2207 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2209 struct rx_serverQueueEntry *sq;
2210 struct rx_call *call = (struct rx_call *)0, *choice2;
2211 struct rx_service *service = NULL;
2215 MUTEX_ENTER(&freeSQEList_lock);
2217 if ((sq = rx_FreeSQEList)) {
2218 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2219 MUTEX_EXIT(&freeSQEList_lock);
2220 } else { /* otherwise allocate a new one and return that */
2221 MUTEX_EXIT(&freeSQEList_lock);
2222 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2223 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2224 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2226 MUTEX_ENTER(&sq->lock);
2228 if (cur_service != NULL) {
2229 cur_service->nRequestsRunning--;
2230 MUTEX_ENTER(&rx_quota_mutex);
2231 if (cur_service->nRequestsRunning < cur_service->minProcs)
2234 MUTEX_EXIT(&rx_quota_mutex);
2236 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2237 struct rx_call *tcall;
2238 struct opr_queue *cursor;
2239 /* Scan for eligible incoming calls. A call is not eligible
2240 * if the maximum number of calls for its service type are
2241 * already executing */
2242 /* One thread will process calls FCFS (to prevent starvation),
2243 * while the other threads may run ahead looking for calls which
2244 * have all their input data available immediately. This helps
2245 * keep threads from blocking, waiting for data from the client. */
2246 choice2 = (struct rx_call *)0;
2247 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2248 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2249 service = tcall->conn->service;
2250 if (QuotaOK(service)) {
2251 MUTEX_ENTER(&rx_pthread_mutex);
2252 /* XXX - If tcall->entry.next is NULL, then we're no longer
2253 * on a queue at all. This shouldn't happen. */
2254 if (tno == rxi_fcfs_thread_num || !tcall->entry.next) {
2255 MUTEX_EXIT(&rx_pthread_mutex);
2256 /* If we're the fcfs thread, then we'll just use
2257 * this call. If we haven't been able to find an optimal
2258 * choice, and we're at the end of the list, then use a
2259 * 2d choice if one has been identified. Otherwise... */
2260 call = (choice2 ? choice2 : tcall);
2261 service = call->conn->service;
2263 MUTEX_EXIT(&rx_pthread_mutex);
2264 if (!opr_queue_IsEmpty(&tcall->rq)) {
2265 struct rx_packet *rp;
2266 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2268 if (rp->header.seq == 1
2270 || (rp->header.flags & RX_LAST_PACKET))) {
2272 } else if (rxi_2dchoice && !choice2
2273 && !(tcall->flags & RX_CALL_CLEARED)
2274 && (tcall->rprev > rxi_HardAckRate)) {
2287 opr_queue_Remove(&call->entry);
2288 /* we can't schedule a call if there's no data!!! */
2289 /* send an ack if there's no data, if we're missing the
2290 * first packet, or we're missing something between first
2291 * and last -- there's a "hole" in the incoming data. */
2292 if (opr_queue_IsEmpty(&call->rq)
2293 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1
2294 || call->rprev != opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq)
2295 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2297 call->flags &= (~RX_CALL_WAIT_PROC);
2298 service->nRequestsRunning++;
2299 /* just started call in minProcs pool, need fewer to maintain
2301 MUTEX_ENTER(&rx_quota_mutex);
2302 if (service->nRequestsRunning <= service->minProcs)
2305 MUTEX_EXIT(&rx_quota_mutex);
2306 rx_atomic_dec(&rx_nWaiting);
2307 /* MUTEX_EXIT(&call->lock); */
2309 /* If there are no eligible incoming calls, add this process
2310 * to the idle server queue, to wait for one */
2313 *socketp = OSI_NULLSOCKET;
2315 sq->socketp = socketp;
2316 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2320 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2322 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2323 return (struct rx_call *)0;
2326 } while (!(call = sq->newcall)
2327 && !(socketp && *socketp != OSI_NULLSOCKET));
2329 MUTEX_EXIT(&sq->lock);
2331 MUTEX_ENTER(&freeSQEList_lock);
2332 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2333 rx_FreeSQEList = sq;
2334 MUTEX_EXIT(&freeSQEList_lock);
2337 clock_GetTime(&call->startTime);
2338 call->state = RX_STATE_ACTIVE;
2339 call->mode = RX_MODE_RECEIVING;
2340 #ifdef RX_KERNEL_TRACE
2341 if (ICL_SETACTIVE(afs_iclSetp)) {
2342 int glockOwner = ISAFS_GLOCK();
2345 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2346 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2353 rxi_calltrace(RX_CALL_START, call);
2354 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2355 call->conn->service->servicePort, call->conn->service->serviceId,
2358 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2365 #endif /* RX_ENABLE_LOCKS */
2369 /* Establish a procedure to be called when a packet arrives for a
2370 * call. This routine will be called at most once after each call,
2371 * and will also be called if there is an error condition on the or
2372 * the call is complete. Used by multi rx to build a selection
2373 * function which determines which of several calls is likely to be a
2374 * good one to read from.
2375 * NOTE: the way this is currently implemented it is probably only a
2376 * good idea to (1) use it immediately after a newcall (clients only)
2377 * and (2) only use it once. Other uses currently void your warranty
2380 rx_SetArrivalProc(struct rx_call *call,
2381 void (*proc) (struct rx_call * call,
2384 void * handle, int arg)
2386 call->arrivalProc = proc;
2387 call->arrivalProcHandle = handle;
2388 call->arrivalProcArg = arg;
2391 /* Call is finished (possibly prematurely). Return rc to the peer, if
2392 * appropriate, and return the final error code from the conversation
2396 rx_EndCall(struct rx_call *call, afs_int32 rc)
2398 struct rx_connection *conn = call->conn;
2402 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2403 call, rc, call->error, call->abortCode));
2406 MUTEX_ENTER(&call->lock);
2408 if (rc == 0 && call->error == 0) {
2409 call->abortCode = 0;
2410 call->abortCount = 0;
2413 call->arrivalProc = (void (*)())0;
2414 if (rc && call->error == 0) {
2415 rxi_CallError(call, rc);
2416 call->mode = RX_MODE_ERROR;
2417 /* Send an abort message to the peer if this error code has
2418 * only just been set. If it was set previously, assume the
2419 * peer has already been sent the error code or will request it
2421 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2423 if (conn->type == RX_SERVER_CONNECTION) {
2424 /* Make sure reply or at least dummy reply is sent */
2425 if (call->mode == RX_MODE_RECEIVING) {
2426 MUTEX_EXIT(&call->lock);
2427 rxi_WriteProc(call, 0, 0);
2428 MUTEX_ENTER(&call->lock);
2430 if (call->mode == RX_MODE_SENDING) {
2431 MUTEX_EXIT(&call->lock);
2432 rxi_FlushWrite(call);
2433 MUTEX_ENTER(&call->lock);
2435 rxi_calltrace(RX_CALL_END, call);
2436 /* Call goes to hold state until reply packets are acknowledged */
2437 if (call->tfirst + call->nSoftAcked < call->tnext) {
2438 call->state = RX_STATE_HOLD;
2440 call->state = RX_STATE_DALLY;
2441 rxi_ClearTransmitQueue(call, 0);
2442 rxi_rto_cancel(call);
2443 rxevent_Cancel(&call->keepAliveEvent, call,
2444 RX_CALL_REFCOUNT_ALIVE);
2446 } else { /* Client connection */
2448 /* Make sure server receives input packets, in the case where
2449 * no reply arguments are expected */
2450 if ((call->mode == RX_MODE_SENDING)
2451 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2452 MUTEX_EXIT(&call->lock);
2453 (void)rxi_ReadProc(call, &dummy, 1);
2454 MUTEX_ENTER(&call->lock);
2457 /* If we had an outstanding delayed ack, be nice to the server
2458 * and force-send it now.
2460 if (call->delayedAckEvent) {
2461 rxevent_Cancel(&call->delayedAckEvent, call,
2462 RX_CALL_REFCOUNT_DELAY);
2463 rxi_SendDelayedAck(NULL, call, NULL, 0);
2466 /* We need to release the call lock since it's lower than the
2467 * conn_call_lock and we don't want to hold the conn_call_lock
2468 * over the rx_ReadProc call. The conn_call_lock needs to be held
2469 * here for the case where rx_NewCall is perusing the calls on
2470 * the connection structure. We don't want to signal until
2471 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2472 * have checked this call, found it active and by the time it
2473 * goes to sleep, will have missed the signal.
2475 MUTEX_EXIT(&call->lock);
2476 MUTEX_ENTER(&conn->conn_call_lock);
2477 MUTEX_ENTER(&call->lock);
2479 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2480 conn->lastBusy[call->channel] = 0;
2483 MUTEX_ENTER(&conn->conn_data_lock);
2484 conn->flags |= RX_CONN_BUSY;
2485 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2486 MUTEX_EXIT(&conn->conn_data_lock);
2487 #ifdef RX_ENABLE_LOCKS
2488 CV_BROADCAST(&conn->conn_call_cv);
2493 #ifdef RX_ENABLE_LOCKS
2495 MUTEX_EXIT(&conn->conn_data_lock);
2497 #endif /* RX_ENABLE_LOCKS */
2498 call->state = RX_STATE_DALLY;
2500 error = call->error;
2502 /* currentPacket, nLeft, and NFree must be zeroed here, because
2503 * ResetCall cannot: ResetCall may be called at splnet(), in the
2504 * kernel version, and may interrupt the macros rx_Read or
2505 * rx_Write, which run at normal priority for efficiency. */
2506 if (call->currentPacket) {
2507 #ifdef RX_TRACK_PACKETS
2508 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2510 rxi_FreePacket(call->currentPacket);
2511 call->currentPacket = (struct rx_packet *)0;
2514 call->nLeft = call->nFree = call->curlen = 0;
2516 /* Free any packets from the last call to ReadvProc/WritevProc */
2517 #ifdef RXDEBUG_PACKET
2519 #endif /* RXDEBUG_PACKET */
2520 rxi_FreePackets(0, &call->iovq);
2521 MUTEX_EXIT(&call->lock);
2523 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2524 if (conn->type == RX_CLIENT_CONNECTION) {
2525 MUTEX_ENTER(&conn->conn_data_lock);
2526 conn->flags &= ~RX_CONN_BUSY;
2527 MUTEX_EXIT(&conn->conn_data_lock);
2528 MUTEX_EXIT(&conn->conn_call_lock);
2532 * Map errors to the local host's errno.h format.
2534 error = ntoh_syserr_conv(error);
2538 #if !defined(KERNEL)
2540 /* Call this routine when shutting down a server or client (especially
2541 * clients). This will allow Rx to gracefully garbage collect server
2542 * connections, and reduce the number of retries that a server might
2543 * make to a dead client.
2544 * This is not quite right, since some calls may still be ongoing and
2545 * we can't lock them to destroy them. */
2549 struct rx_connection **conn_ptr, **conn_end;
2553 if (rxinit_status == 1) {
2555 return; /* Already shutdown. */
2557 rxi_DeleteCachedConnections();
2558 if (rx_connHashTable) {
2559 MUTEX_ENTER(&rx_connHashTable_lock);
2560 for (conn_ptr = &rx_connHashTable[0], conn_end =
2561 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2563 struct rx_connection *conn, *next;
2564 for (conn = *conn_ptr; conn; conn = next) {
2566 if (conn->type == RX_CLIENT_CONNECTION) {
2567 MUTEX_ENTER(&rx_refcnt_mutex);
2569 MUTEX_EXIT(&rx_refcnt_mutex);
2570 #ifdef RX_ENABLE_LOCKS
2571 rxi_DestroyConnectionNoLock(conn);
2572 #else /* RX_ENABLE_LOCKS */
2573 rxi_DestroyConnection(conn);
2574 #endif /* RX_ENABLE_LOCKS */
2578 #ifdef RX_ENABLE_LOCKS
2579 while (rx_connCleanup_list) {
2580 struct rx_connection *conn;
2581 conn = rx_connCleanup_list;
2582 rx_connCleanup_list = rx_connCleanup_list->next;
2583 MUTEX_EXIT(&rx_connHashTable_lock);
2584 rxi_CleanupConnection(conn);
2585 MUTEX_ENTER(&rx_connHashTable_lock);
2587 MUTEX_EXIT(&rx_connHashTable_lock);
2588 #endif /* RX_ENABLE_LOCKS */
2593 afs_winsockCleanup();
2601 /* if we wakeup packet waiter too often, can get in loop with two
2602 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2604 rxi_PacketsUnWait(void)
2606 if (!rx_waitingForPackets) {
2610 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2611 return; /* still over quota */
2614 rx_waitingForPackets = 0;
2615 #ifdef RX_ENABLE_LOCKS
2616 CV_BROADCAST(&rx_waitingForPackets_cv);
2618 osi_rxWakeup(&rx_waitingForPackets);
2624 /* ------------------Internal interfaces------------------------- */
2626 /* Return this process's service structure for the
2627 * specified socket and service */
2628 static struct rx_service *
2629 rxi_FindService(osi_socket socket, u_short serviceId)
2631 struct rx_service **sp;
2632 for (sp = &rx_services[0]; *sp; sp++) {
2633 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2639 #ifdef RXDEBUG_PACKET
2640 #ifdef KDUMP_RX_LOCK
2641 static struct rx_call_rx_lock *rx_allCallsp = 0;
2643 static struct rx_call *rx_allCallsp = 0;
2645 #endif /* RXDEBUG_PACKET */
2647 /* Allocate a call structure, for the indicated channel of the
2648 * supplied connection. The mode and state of the call must be set by
2649 * the caller. Returns the call with mutex locked. */
2650 static struct rx_call *
2651 rxi_NewCall(struct rx_connection *conn, int channel)
2653 struct rx_call *call;
2654 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2655 struct rx_call *cp; /* Call pointer temp */
2656 struct opr_queue *cursor;
2657 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2659 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2661 /* Grab an existing call structure, or allocate a new one.
2662 * Existing call structures are assumed to have been left reset by
2664 MUTEX_ENTER(&rx_freeCallQueue_lock);
2666 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2668 * EXCEPT that the TQ might not yet be cleared out.
2669 * Skip over those with in-use TQs.
2672 for (opr_queue_Scan(&rx_freeCallQueue, cursor)) {
2673 cp = opr_queue_Entry(cursor, struct rx_call, entry);
2674 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2680 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2681 if (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
2682 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
2683 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2684 opr_queue_Remove(&call->entry);
2685 if (rx_stats_active)
2686 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2687 MUTEX_EXIT(&rx_freeCallQueue_lock);
2688 MUTEX_ENTER(&call->lock);
2689 CLEAR_CALL_QUEUE_LOCK(call);
2690 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2691 /* Now, if TQ wasn't cleared earlier, do it now. */
2692 rxi_WaitforTQBusy(call);
2693 if (call->flags & RX_CALL_TQ_CLEARME) {
2694 rxi_ClearTransmitQueue(call, 1);
2695 /*queue_Init(&call->tq);*/
2697 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2698 /* Bind the call to its connection structure */
2700 rxi_ResetCall(call, 1);
2703 call = rxi_Alloc(sizeof(struct rx_call));
2704 #ifdef RXDEBUG_PACKET
2705 call->allNextp = rx_allCallsp;
2706 rx_allCallsp = call;
2708 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2709 #else /* RXDEBUG_PACKET */
2710 rx_atomic_inc(&rx_stats.nCallStructs);
2711 #endif /* RXDEBUG_PACKET */
2713 MUTEX_EXIT(&rx_freeCallQueue_lock);
2714 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2715 MUTEX_ENTER(&call->lock);
2716 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2717 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2718 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2720 /* Initialize once-only items */
2721 opr_queue_Init(&call->tq);
2722 opr_queue_Init(&call->rq);
2723 opr_queue_Init(&call->iovq);
2724 #ifdef RXDEBUG_PACKET
2725 call->rqc = call->tqc = call->iovqc = 0;
2726 #endif /* RXDEBUG_PACKET */
2727 /* Bind the call to its connection structure (prereq for reset) */
2729 rxi_ResetCall(call, 1);
2731 call->channel = channel;
2732 call->callNumber = &conn->callNumber[channel];
2733 call->rwind = conn->rwind[channel];
2734 call->twind = conn->twind[channel];
2735 /* Note that the next expected call number is retained (in
2736 * conn->callNumber[i]), even if we reallocate the call structure
2738 conn->call[channel] = call;
2739 /* if the channel's never been used (== 0), we should start at 1, otherwise
2740 * the call number is valid from the last time this channel was used */
2741 if (*call->callNumber == 0)
2742 *call->callNumber = 1;
2747 /* A call has been inactive long enough that so we can throw away
2748 * state, including the call structure, which is placed on the call
2751 * call->lock amd rx_refcnt_mutex are held upon entry.
2752 * haveCTLock is set when called from rxi_ReapConnections.
2754 * return 1 if the call is freed, 0 if not.
2757 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2759 int channel = call->channel;
2760 struct rx_connection *conn = call->conn;
2761 u_char state = call->state;
2764 * We are setting the state to RX_STATE_RESET to
2765 * ensure that no one else will attempt to use this
2766 * call once we drop the refcnt lock. We must drop
2767 * the refcnt lock before calling rxi_ResetCall
2768 * because it cannot be held across acquiring the
2769 * freepktQ lock. NewCall does the same.
2771 call->state = RX_STATE_RESET;
2772 MUTEX_EXIT(&rx_refcnt_mutex);
2773 rxi_ResetCall(call, 0);
2775 if (MUTEX_TRYENTER(&conn->conn_call_lock))
2777 if (state == RX_STATE_DALLY || state == RX_STATE_HOLD)
2778 (*call->callNumber)++;
2780 if (call->conn->call[channel] == call)
2781 call->conn->call[channel] = 0;
2782 MUTEX_EXIT(&conn->conn_call_lock);
2785 * We couldn't obtain the conn_call_lock so we can't
2786 * disconnect the call from the connection. Set the
2787 * call state to dally so that the call can be reused.
2789 MUTEX_ENTER(&rx_refcnt_mutex);
2790 call->state = RX_STATE_DALLY;
2794 MUTEX_ENTER(&rx_freeCallQueue_lock);
2795 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2796 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2797 /* A call may be free even though its transmit queue is still in use.
2798 * Since we search the call list from head to tail, put busy calls at
2799 * the head of the list, and idle calls at the tail.
2801 if (call->flags & RX_CALL_TQ_BUSY)
2802 opr_queue_Prepend(&rx_freeCallQueue, &call->entry);
2804 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2805 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2806 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2807 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2808 if (rx_stats_active)
2809 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2810 MUTEX_EXIT(&rx_freeCallQueue_lock);
2812 /* Destroy the connection if it was previously slated for
2813 * destruction, i.e. the Rx client code previously called
2814 * rx_DestroyConnection (client connections), or
2815 * rxi_ReapConnections called the same routine (server
2816 * connections). Only do this, however, if there are no
2817 * outstanding calls. Note that for fine grain locking, there appears
2818 * to be a deadlock in that rxi_FreeCall has a call locked and
2819 * DestroyConnectionNoLock locks each call in the conn. But note a
2820 * few lines up where we have removed this call from the conn.
2821 * If someone else destroys a connection, they either have no
2822 * call lock held or are going through this section of code.
2824 MUTEX_ENTER(&conn->conn_data_lock);
2825 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2826 MUTEX_ENTER(&rx_refcnt_mutex);
2828 MUTEX_EXIT(&rx_refcnt_mutex);
2829 MUTEX_EXIT(&conn->conn_data_lock);
2830 #ifdef RX_ENABLE_LOCKS
2832 rxi_DestroyConnectionNoLock(conn);
2834 rxi_DestroyConnection(conn);
2835 #else /* RX_ENABLE_LOCKS */
2836 rxi_DestroyConnection(conn);
2837 #endif /* RX_ENABLE_LOCKS */
2839 MUTEX_EXIT(&conn->conn_data_lock);
2841 MUTEX_ENTER(&rx_refcnt_mutex);
2845 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2846 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2849 rxi_Alloc(size_t size)
2853 if (rx_stats_active) {
2854 rx_atomic_add(&rxi_Allocsize, (int) size);
2855 rx_atomic_inc(&rxi_Alloccnt);
2859 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2860 afs_osi_Alloc_NoSleep(size);
2865 osi_Panic("rxi_Alloc error");
2871 rxi_Free(void *addr, size_t size)
2873 if (rx_stats_active) {
2874 rx_atomic_sub(&rxi_Allocsize, (int) size);
2875 rx_atomic_dec(&rxi_Alloccnt);
2877 osi_Free(addr, size);
2881 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2883 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2884 struct rx_peer *next = NULL;
2888 MUTEX_ENTER(&rx_peerHashTable_lock);
2890 peer_ptr = &rx_peerHashTable[0];
2891 peer_end = &rx_peerHashTable[rx_hashTableSize];
2894 for ( ; peer_ptr < peer_end; peer_ptr++) {
2897 for ( ; peer; peer = next) {
2899 if (host == peer->host)
2904 hashIndex = PEER_HASH(host, port);
2905 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2906 if ((peer->host == host) && (peer->port == port))
2911 MUTEX_ENTER(&rx_peerHashTable_lock);
2916 MUTEX_EXIT(&rx_peerHashTable_lock);
2918 MUTEX_ENTER(&peer->peer_lock);
2919 /* We don't handle dropping below min, so don't */
2920 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2921 peer->ifMTU=MIN(mtu, peer->ifMTU);
2922 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2923 /* if we tweaked this down, need to tune our peer MTU too */
2924 peer->MTU = MIN(peer->MTU, peer->natMTU);
2925 /* if we discovered a sub-1500 mtu, degrade */
2926 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2927 peer->maxDgramPackets = 1;
2928 /* We no longer have valid peer packet information */
2929 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2930 peer->maxPacketSize = 0;
2931 MUTEX_EXIT(&peer->peer_lock);
2933 MUTEX_ENTER(&rx_peerHashTable_lock);
2935 if (host && !port) {
2937 /* pick up where we left off */
2941 MUTEX_EXIT(&rx_peerHashTable_lock);
2944 #ifdef AFS_RXERRQ_ENV
2946 rxi_SetPeerDead(afs_uint32 host, afs_uint16 port)
2948 int hashIndex = PEER_HASH(host, port);
2949 struct rx_peer *peer;
2951 MUTEX_ENTER(&rx_peerHashTable_lock);
2953 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2954 if (peer->host == host && peer->port == port) {
2960 rx_atomic_inc(&peer->neterrs);
2963 MUTEX_EXIT(&rx_peerHashTable_lock);
2967 rxi_ProcessNetError(struct sock_extended_err *err, afs_uint32 addr, afs_uint16 port)
2969 # ifdef AFS_ADAPT_PMTU
2970 if (err->ee_errno == EMSGSIZE && err->ee_info >= 68) {
2971 rxi_SetPeerMtu(NULL, addr, port, err->ee_info - RX_IPUDP_SIZE);
2975 if (err->ee_origin == SO_EE_ORIGIN_ICMP && err->ee_type == ICMP_DEST_UNREACH) {
2976 switch (err->ee_code) {
2977 case ICMP_NET_UNREACH:
2978 case ICMP_HOST_UNREACH:
2979 case ICMP_PORT_UNREACH:
2982 rxi_SetPeerDead(addr, port);
2987 #endif /* AFS_RXERRQ_ENV */
2989 /* Find the peer process represented by the supplied (host,port)
2990 * combination. If there is no appropriate active peer structure, a
2991 * new one will be allocated and initialized
2992 * The origPeer, if set, is a pointer to a peer structure on which the
2993 * refcount will be be decremented. This is used to replace the peer
2994 * structure hanging off a connection structure */
2996 rxi_FindPeer(afs_uint32 host, u_short port,
2997 struct rx_peer *origPeer, int create)
3001 hashIndex = PEER_HASH(host, port);
3002 MUTEX_ENTER(&rx_peerHashTable_lock);
3003 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
3004 if ((pp->host == host) && (pp->port == port))
3009 pp = rxi_AllocPeer(); /* This bzero's *pp */
3010 pp->host = host; /* set here or in InitPeerParams is zero */
3012 #ifdef AFS_RXERRQ_ENV
3013 rx_atomic_set(&pp->neterrs, 0);
3015 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
3016 opr_queue_Init(&pp->rpcStats);
3017 pp->next = rx_peerHashTable[hashIndex];
3018 rx_peerHashTable[hashIndex] = pp;
3019 rxi_InitPeerParams(pp);
3020 if (rx_stats_active)
3021 rx_atomic_inc(&rx_stats.nPeerStructs);
3028 origPeer->refCount--;
3029 MUTEX_EXIT(&rx_peerHashTable_lock);
3034 /* Find the connection at (host, port) started at epoch, and with the
3035 * given connection id. Creates the server connection if necessary.
3036 * The type specifies whether a client connection or a server
3037 * connection is desired. In both cases, (host, port) specify the
3038 * peer's (host, pair) pair. Client connections are not made
3039 * automatically by this routine. The parameter socket gives the
3040 * socket descriptor on which the packet was received. This is used,
3041 * in the case of server connections, to check that *new* connections
3042 * come via a valid (port, serviceId). Finally, the securityIndex
3043 * parameter must match the existing index for the connection. If a
3044 * server connection is created, it will be created using the supplied
3045 * index, if the index is valid for this service */
3046 static struct rx_connection *
3047 rxi_FindConnection(osi_socket socket, afs_uint32 host,
3048 u_short port, u_short serviceId, afs_uint32 cid,
3049 afs_uint32 epoch, int type, u_int securityIndex)
3051 int hashindex, flag, i;
3052 struct rx_connection *conn;
3053 hashindex = CONN_HASH(host, port, cid, epoch, type);
3054 MUTEX_ENTER(&rx_connHashTable_lock);
3055 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
3056 rx_connHashTable[hashindex],
3059 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
3060 && (epoch == conn->epoch)) {
3061 struct rx_peer *pp = conn->peer;
3062 if (securityIndex != conn->securityIndex) {
3063 /* this isn't supposed to happen, but someone could forge a packet
3064 * like this, and there seems to be some CM bug that makes this
3065 * happen from time to time -- in which case, the fileserver
3067 MUTEX_EXIT(&rx_connHashTable_lock);
3068 return (struct rx_connection *)0;
3070 if (pp->host == host && pp->port == port)
3072 if (type == RX_CLIENT_CONNECTION && pp->port == port)
3074 /* So what happens when it's a callback connection? */
3075 if ( /*type == RX_CLIENT_CONNECTION && */
3076 (conn->epoch & 0x80000000))
3080 /* the connection rxLastConn that was used the last time is not the
3081 ** one we are looking for now. Hence, start searching in the hash */
3083 conn = rx_connHashTable[hashindex];
3088 struct rx_service *service;
3089 if (type == RX_CLIENT_CONNECTION) {
3090 MUTEX_EXIT(&rx_connHashTable_lock);
3091 return (struct rx_connection *)0;
3093 service = rxi_FindService(socket, serviceId);
3094 if (!service || (securityIndex >= service->nSecurityObjects)
3095 || (service->securityObjects[securityIndex] == 0)) {
3096 MUTEX_EXIT(&rx_connHashTable_lock);
3097 return (struct rx_connection *)0;
3099 conn = rxi_AllocConnection(); /* This bzero's the connection */
3100 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
3101 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
3102 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
3103 conn->next = rx_connHashTable[hashindex];
3104 rx_connHashTable[hashindex] = conn;
3105 conn->peer = rxi_FindPeer(host, port, 0, 1);
3106 conn->type = RX_SERVER_CONNECTION;
3107 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
3108 conn->epoch = epoch;
3109 conn->cid = cid & RX_CIDMASK;
3110 conn->ackRate = RX_FAST_ACK_RATE;
3111 conn->service = service;
3112 conn->serviceId = serviceId;
3113 conn->securityIndex = securityIndex;
3114 conn->securityObject = service->securityObjects[securityIndex];
3115 conn->nSpecific = 0;
3116 conn->specific = NULL;
3117 rx_SetConnDeadTime(conn, service->connDeadTime);
3118 conn->idleDeadTime = service->idleDeadTime;
3119 conn->idleDeadDetection = service->idleDeadErr ? 1 : 0;
3120 for (i = 0; i < RX_MAXCALLS; i++) {
3121 conn->twind[i] = rx_initSendWindow;
3122 conn->rwind[i] = rx_initReceiveWindow;
3124 /* Notify security object of the new connection */
3125 RXS_NewConnection(conn->securityObject, conn);
3126 /* XXXX Connection timeout? */
3127 if (service->newConnProc)
3128 (*service->newConnProc) (conn);
3129 if (rx_stats_active)
3130 rx_atomic_inc(&rx_stats.nServerConns);
3133 MUTEX_ENTER(&rx_refcnt_mutex);
3135 MUTEX_EXIT(&rx_refcnt_mutex);
3137 rxLastConn = conn; /* store this connection as the last conn used */
3138 MUTEX_EXIT(&rx_connHashTable_lock);
3143 * Timeout a call on a busy call channel if appropriate.
3145 * @param[in] call The busy call.
3147 * @pre 'call' is marked as busy (namely,
3148 * call->conn->lastBusy[call->channel] != 0)
3150 * @pre call->lock is held
3151 * @pre rxi_busyChannelError is nonzero
3153 * @note call->lock is dropped and reacquired
3156 rxi_CheckBusy(struct rx_call *call)
3158 struct rx_connection *conn = call->conn;
3159 int channel = call->channel;
3160 int freechannel = 0;
3162 afs_uint32 callNumber;
3164 MUTEX_EXIT(&call->lock);
3166 MUTEX_ENTER(&conn->conn_call_lock);
3167 callNumber = *call->callNumber;
3169 /* Are there any other call slots on this conn that we should try? Look for
3170 * slots that are empty and are either non-busy, or were marked as busy
3171 * longer than conn->secondsUntilDead seconds before this call started. */
3173 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3175 /* only look at channels that aren't us */
3179 if (conn->lastBusy[i]) {
3180 /* if this channel looked busy too recently, don't look at it */
3181 if (conn->lastBusy[i] >= call->startTime.sec) {
3184 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3189 if (conn->call[i]) {
3190 struct rx_call *tcall = conn->call[i];
3191 MUTEX_ENTER(&tcall->lock);
3192 if (tcall->state == RX_STATE_DALLY) {
3195 MUTEX_EXIT(&tcall->lock);
3201 MUTEX_ENTER(&call->lock);
3203 /* Since the call->lock and conn->conn_call_lock have been released it is
3204 * possible that (1) the call may no longer be busy and/or (2) the call may
3205 * have been reused by another waiting thread. Therefore, we must confirm
3206 * that the call state has not changed when deciding whether or not to
3207 * force this application thread to retry by forcing a Timeout error. */
3209 if (freechannel && *call->callNumber == callNumber &&
3210 (call->flags & RX_CALL_PEER_BUSY)) {
3211 /* Since 'freechannel' is set, there exists another channel in this
3212 * rx_conn that the application thread might be able to use. We know
3213 * that we have the correct call since callNumber is unchanged, and we
3214 * know that the call is still busy. So, set the call error state to
3215 * rxi_busyChannelError so the application can retry the request,
3216 * presumably on a less-busy call channel. */
3218 rxi_CallError(call, RX_CALL_BUSY);
3220 MUTEX_EXIT(&conn->conn_call_lock);
3223 /* There are two packet tracing routines available for testing and monitoring
3224 * Rx. One is called just after every packet is received and the other is
3225 * called just before every packet is sent. Received packets, have had their
3226 * headers decoded, and packets to be sent have not yet had their headers
3227 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3228 * containing the network address. Both can be modified. The return value, if
3229 * non-zero, indicates that the packet should be dropped. */
3231 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3232 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3234 /* A packet has been received off the interface. Np is the packet, socket is
3235 * the socket number it was received from (useful in determining which service
3236 * this packet corresponds to), and (host, port) reflect the host,port of the
3237 * sender. This call returns the packet to the caller if it is finished with
3238 * it, rather than de-allocating it, just as a small performance hack */
3241 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3242 afs_uint32 host, u_short port, int *tnop,
3243 struct rx_call **newcallp)
3245 struct rx_call *call;
3246 struct rx_connection *conn;
3248 afs_uint32 currentCallNumber;
3253 struct rx_packet *tnp;
3256 /* We don't print out the packet until now because (1) the time may not be
3257 * accurate enough until now in the lwp implementation (rx_Listener only gets
3258 * the time after the packet is read) and (2) from a protocol point of view,
3259 * this is the first time the packet has been seen */
3260 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3261 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3262 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3263 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3264 np->header.epoch, np->header.cid, np->header.callNumber,
3265 np->header.seq, np->header.flags, np));
3268 /* Account for connectionless packets */
3269 if (rx_stats_active &&
3270 ((np->header.type == RX_PACKET_TYPE_VERSION) ||
3271 (np->header.type == RX_PACKET_TYPE_DEBUG))) {
3272 struct rx_peer *peer;
3274 /* Try to look up the peer structure, but don't create one */
3275 peer = rxi_FindPeer(host, port, 0, 0);
3277 /* Since this may not be associated with a connection, it may have
3278 * no refCount, meaning we could race with ReapConnections
3281 if (peer && (peer->refCount > 0)) {
3282 #ifdef AFS_RXERRQ_ENV
3283 if (rx_atomic_read(&peer->neterrs)) {
3284 rx_atomic_set(&peer->neterrs, 0);
3287 MUTEX_ENTER(&peer->peer_lock);
3288 peer->bytesReceived += np->length;
3289 MUTEX_EXIT(&peer->peer_lock);
3293 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3294 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3297 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3298 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3301 /* If an input tracer function is defined, call it with the packet and
3302 * network address. Note this function may modify its arguments. */
3303 if (rx_justReceived) {
3304 struct sockaddr_in addr;
3306 addr.sin_family = AF_INET;
3307 addr.sin_port = port;
3308 addr.sin_addr.s_addr = host;
3309 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3310 addr.sin_len = sizeof(addr);
3311 #endif /* AFS_OSF_ENV */
3312 drop = (*rx_justReceived) (np, &addr);
3313 /* drop packet if return value is non-zero */
3316 port = addr.sin_port; /* in case fcn changed addr */
3317 host = addr.sin_addr.s_addr;
3321 /* If packet was not sent by the client, then *we* must be the client */
3322 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3323 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3325 /* Find the connection (or fabricate one, if we're the server & if
3326 * necessary) associated with this packet */
3328 rxi_FindConnection(socket, host, port, np->header.serviceId,
3329 np->header.cid, np->header.epoch, type,
3330 np->header.securityIndex);
3332 /* To avoid having 2 connections just abort at each other,
3333 don't abort an abort. */
3335 if (np->header.type != RX_PACKET_TYPE_ABORT)
3336 rxi_SendRawAbort(socket, host, port, RX_INVALID_OPERATION,
3341 #ifdef AFS_RXERRQ_ENV
3342 if (rx_atomic_read(&conn->peer->neterrs)) {
3343 rx_atomic_set(&conn->peer->neterrs, 0);
3347 /* If we're doing statistics, then account for the incoming packet */
3348 if (rx_stats_active) {
3349 MUTEX_ENTER(&conn->peer->peer_lock);
3350 conn->peer->bytesReceived += np->length;
3351 MUTEX_EXIT(&conn->peer->peer_lock);
3354 /* If the connection is in an error state, send an abort packet and ignore
3355 * the incoming packet */
3357 /* Don't respond to an abort packet--we don't want loops! */
3358 MUTEX_ENTER(&conn->conn_data_lock);
3359 if (np->header.type != RX_PACKET_TYPE_ABORT)
3360 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3361 putConnection(conn);
3362 MUTEX_EXIT(&conn->conn_data_lock);
3366 /* Check for connection-only requests (i.e. not call specific). */
3367 if (np->header.callNumber == 0) {
3368 switch (np->header.type) {
3369 case RX_PACKET_TYPE_ABORT: {
3370 /* What if the supplied error is zero? */
3371 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3372 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3373 rxi_ConnectionError(conn, errcode);
3374 putConnection(conn);
3377 case RX_PACKET_TYPE_CHALLENGE:
3378 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3379 putConnection(conn);
3381 case RX_PACKET_TYPE_RESPONSE:
3382 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3383 putConnection(conn);
3385 case RX_PACKET_TYPE_PARAMS:
3386 case RX_PACKET_TYPE_PARAMS + 1:
3387 case RX_PACKET_TYPE_PARAMS + 2:
3388 /* ignore these packet types for now */
3389 putConnection(conn);
3393 /* Should not reach here, unless the peer is broken: send an
3395 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3396 MUTEX_ENTER(&conn->conn_data_lock);
3397 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3398 putConnection(conn);
3399 MUTEX_EXIT(&conn->conn_data_lock);
3404 channel = np->header.cid & RX_CHANNELMASK;
3405 MUTEX_ENTER(&conn->conn_call_lock);
3406 call = conn->call[channel];
3409 MUTEX_ENTER(&call->lock);
3410 currentCallNumber = conn->callNumber[channel];
3411 MUTEX_EXIT(&conn->conn_call_lock);
3412 } else if (type == RX_SERVER_CONNECTION) { /* No call allocated */
3413 call = conn->call[channel];
3415 MUTEX_ENTER(&call->lock);
3416 currentCallNumber = conn->callNumber[channel];
3417 MUTEX_EXIT(&conn->conn_call_lock);
3419 call = rxi_NewCall(conn, channel); /* returns locked call */
3420 *call->callNumber = currentCallNumber = np->header.callNumber;
3421 MUTEX_EXIT(&conn->conn_call_lock);
3423 if (np->header.callNumber == 0)
3424 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3425 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3426 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3427 np->header.flags, np, np->length));
3429 call->state = RX_STATE_PRECALL;
3430 clock_GetTime(&call->queueTime);
3431 call->bytesSent = 0;
3432 call->bytesRcvd = 0;
3434 * If the number of queued calls exceeds the overload
3435 * threshold then abort this call.
3437 if ((rx_BusyThreshold > 0) &&
3438 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3439 struct rx_packet *tp;
3441 rxi_CallError(call, rx_BusyError);
3442 tp = rxi_SendCallAbort(call, np, 1, 0);
3443 MUTEX_EXIT(&call->lock);
3444 putConnection(conn);
3445 if (rx_stats_active)
3446 rx_atomic_inc(&rx_stats.nBusies);
3449 rxi_KeepAliveOn(call);
3451 } else { /* RX_CLIENT_CONNECTION and No call allocated */
3452 /* This packet can't be for this call. If the new call address is
3453 * 0 then no call is running on this channel. If there is a call
3454 * then, since this is a client connection we're getting data for
3455 * it must be for the previous call.
3457 MUTEX_EXIT(&conn->conn_call_lock);
3458 if (rx_stats_active)
3459 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3460 putConnection(conn);
3464 /* There is a non-NULL locked call at this point */
3465 if (type == RX_SERVER_CONNECTION) { /* We're the server */
3466 if (np->header.callNumber < currentCallNumber) {
3467 MUTEX_EXIT(&call->lock);
3468 if (rx_stats_active)
3469 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3470 putConnection(conn);
3472 } else if (np->header.callNumber != currentCallNumber) {
3473 /* Wait until the transmit queue is idle before deciding
3474 * whether to reset the current call. Chances are that the
3475 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3478 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3479 if (call->state == RX_STATE_ACTIVE) {
3480 rxi_WaitforTQBusy(call);
3482 * If we entered error state while waiting,
3483 * must call rxi_CallError to permit rxi_ResetCall
3484 * to processed when the tqWaiter count hits zero.
3487 rxi_CallError(call, call->error);
3488 MUTEX_EXIT(&call->lock);
3489 putConnection(conn);
3493 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3494 /* If the new call cannot be taken right now send a busy and set
3495 * the error condition in this call, so that it terminates as
3496 * quickly as possible */
3497 if (call->state == RX_STATE_ACTIVE) {
3498 struct rx_packet *tp;
3500 rxi_CallError(call, RX_CALL_DEAD);
3501 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3503 MUTEX_EXIT(&call->lock);
3504 putConnection(conn);
3507 rxi_ResetCall(call, 0);
3509 * The conn_call_lock is not held but no one else should be
3510 * using this call channel while we are processing this incoming
3511 * packet. This assignment should be safe.
3513 *call->callNumber = np->header.callNumber;
3515 if (np->header.callNumber == 0)
3516 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3517 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3518 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3519 np->header.flags, np, np->length));
3521 call->state = RX_STATE_PRECALL;
3522 clock_GetTime(&call->queueTime);
3523 call->bytesSent = 0;
3524 call->bytesRcvd = 0;
3526 * If the number of queued calls exceeds the overload
3527 * threshold then abort this call.
3529 if ((rx_BusyThreshold > 0) &&
3530 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3531 struct rx_packet *tp;
3533 rxi_CallError(call, rx_BusyError);
3534 tp = rxi_SendCallAbort(call, np, 1, 0);
3535 MUTEX_EXIT(&call->lock);
3536 putConnection(conn);
3537 if (rx_stats_active)
3538 rx_atomic_inc(&rx_stats.nBusies);
3541 rxi_KeepAliveOn(call);
3543 /* Continuing call; do nothing here. */
3545 } else { /* we're the client */
3546 /* Ignore all incoming acknowledgements for calls in DALLY state */
3547 if ((call->state == RX_STATE_DALLY)
3548 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3549 if (rx_stats_active)
3550 rx_atomic_inc(&rx_stats.ignorePacketDally);
3551 MUTEX_EXIT(&call->lock);
3552 putConnection(conn);
3556 /* Ignore anything that's not relevant to the current call. If there
3557 * isn't a current call, then no packet is relevant. */
3558 if (np->header.callNumber != currentCallNumber) {
3559 if (rx_stats_active)
3560 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3561 MUTEX_EXIT(&call->lock);
3562 putConnection(conn);
3565 /* If the service security object index stamped in the packet does not
3566 * match the connection's security index, ignore the packet */
3567 if (np->header.securityIndex != conn->securityIndex) {
3568 MUTEX_EXIT(&call->lock);
3569 putConnection(conn);
3573 /* If we're receiving the response, then all transmit packets are
3574 * implicitly acknowledged. Get rid of them. */
3575 if (np->header.type == RX_PACKET_TYPE_DATA) {
3576 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3577 /* XXX Hack. Because we must release the global rx lock when
3578 * sending packets (osi_NetSend) we drop all acks while we're
3579 * traversing the tq in rxi_Start sending packets out because
3580 * packets may move to the freePacketQueue as result of being here!
3581 * So we drop these packets until we're safely out of the
3582 * traversing. Really ugly!
3583 * For fine grain RX locking, we set the acked field in the
3584 * packets and let rxi_Start remove them from the transmit queue.
3586 if (call->flags & RX_CALL_TQ_BUSY) {
3587 #ifdef RX_ENABLE_LOCKS
3588 rxi_SetAcksInTransmitQueue(call);
3590 putConnection(conn);
3591 return np; /* xmitting; drop packet */
3594 rxi_ClearTransmitQueue(call, 0);
3596 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3597 rxi_ClearTransmitQueue(call, 0);
3598 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3600 if (np->header.type == RX_PACKET_TYPE_ACK) {
3601 /* now check to see if this is an ack packet acknowledging that the
3602 * server actually *lost* some hard-acked data. If this happens we
3603 * ignore this packet, as it may indicate that the server restarted in
3604 * the middle of a call. It is also possible that this is an old ack
3605 * packet. We don't abort the connection in this case, because this
3606 * *might* just be an old ack packet. The right way to detect a server
3607 * restart in the midst of a call is to notice that the server epoch
3609 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3610 * XXX unacknowledged. I think that this is off-by-one, but
3611 * XXX I don't dare change it just yet, since it will
3612 * XXX interact badly with the server-restart detection
3613 * XXX code in receiveackpacket. */
3614 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3615 if (rx_stats_active)
3616 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3617 MUTEX_EXIT(&call->lock);
3618 putConnection(conn);
3622 } /* else not a data packet */
3625 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3626 /* Set remote user defined status from packet */
3627 call->remoteStatus = np->header.userStatus;
3629 /* Now do packet type-specific processing */
3630 switch (np->header.type) {
3631 case RX_PACKET_TYPE_DATA:
3632 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3635 case RX_PACKET_TYPE_ACK:
3636 /* Respond immediately to ack packets requesting acknowledgement
3638 if (np->header.flags & RX_REQUEST_ACK) {
3640 (void)rxi_SendCallAbort(call, 0, 1, 0);
3642 (void)rxi_SendAck(call, 0, np->header.serial,
3643 RX_ACK_PING_RESPONSE, 1);
3645 np = rxi_ReceiveAckPacket(call, np, 1);
3647 case RX_PACKET_TYPE_ABORT: {
3648 /* An abort packet: reset the call, passing the error up to the user. */
3649 /* What if error is zero? */
3650 /* What if the error is -1? the application will treat it as a timeout. */
3651 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3652 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3653 rxi_CallError(call, errdata);
3654 MUTEX_EXIT(&call->lock);
3655 putConnection(conn);
3656 return np; /* xmitting; drop packet */
3658 case RX_PACKET_TYPE_BUSY: {
3659 struct clock busyTime;
3661 clock_GetTime(&busyTime);
3663 MUTEX_EXIT(&call->lock);
3665 MUTEX_ENTER(&conn->conn_call_lock);
3666 MUTEX_ENTER(&call->lock);
3667 conn->lastBusy[call->channel] = busyTime.sec;
3668 call->flags |= RX_CALL_PEER_BUSY;
3669 MUTEX_EXIT(&call->lock);
3670 MUTEX_EXIT(&conn->conn_call_lock);
3672 putConnection(conn);
3676 case RX_PACKET_TYPE_ACKALL:
3677 /* All packets acknowledged, so we can drop all packets previously
3678 * readied for sending */
3679 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3680 /* XXX Hack. We because we can't release the global rx lock when
3681 * sending packets (osi_NetSend) we drop all ack pkts while we're
3682 * traversing the tq in rxi_Start sending packets out because
3683 * packets may move to the freePacketQueue as result of being
3684 * here! So we drop these packets until we're safely out of the
3685 * traversing. Really ugly!
3686 * For fine grain RX locking, we set the acked field in the packets
3687 * and let rxi_Start remove the packets from the transmit queue.
3689 if (call->flags & RX_CALL_TQ_BUSY) {
3690 #ifdef RX_ENABLE_LOCKS
3691 rxi_SetAcksInTransmitQueue(call);
3693 #else /* RX_ENABLE_LOCKS */
3694 MUTEX_EXIT(&call->lock);
3695 putConnection(conn);
3696 return np; /* xmitting; drop packet */
3697 #endif /* RX_ENABLE_LOCKS */
3699 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3700 rxi_ClearTransmitQueue(call, 0);
3703 /* Should not reach here, unless the peer is broken: send an abort
3705 rxi_CallError(call, RX_PROTOCOL_ERROR);
3706 np = rxi_SendCallAbort(call, np, 1, 0);
3709 /* Note when this last legitimate packet was received, for keep-alive
3710 * processing. Note, we delay getting the time until now in the hope that
3711 * the packet will be delivered to the user before any get time is required
3712 * (if not, then the time won't actually be re-evaluated here). */
3713 call->lastReceiveTime = clock_Sec();
3714 /* we've received a legit packet, so the channel is not busy */
3715 call->flags &= ~RX_CALL_PEER_BUSY;
3716 MUTEX_EXIT(&call->lock);
3717 putConnection(conn);
3721 /* return true if this is an "interesting" connection from the point of view
3722 of someone trying to debug the system */
3724 rxi_IsConnInteresting(struct rx_connection *aconn)
3727 struct rx_call *tcall;
3729 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3732 for (i = 0; i < RX_MAXCALLS; i++) {
3733 tcall = aconn->call[i];
3735 if ((tcall->state == RX_STATE_PRECALL)
3736 || (tcall->state == RX_STATE_ACTIVE))
3738 if ((tcall->mode == RX_MODE_SENDING)
3739 || (tcall->mode == RX_MODE_RECEIVING))
3747 /* if this is one of the last few packets AND it wouldn't be used by the
3748 receiving call to immediately satisfy a read request, then drop it on
3749 the floor, since accepting it might prevent a lock-holding thread from
3750 making progress in its reading. If a call has been cleared while in
3751 the precall state then ignore all subsequent packets until the call
3752 is assigned to a thread. */
3755 TooLow(struct rx_packet *ap, struct rx_call *acall)
3759 MUTEX_ENTER(&rx_quota_mutex);
3760 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3761 && (acall->state == RX_STATE_PRECALL))
3762 || ((rx_nFreePackets < rxi_dataQuota + 2)
3763 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3764 && (acall->flags & RX_CALL_READER_WAIT)))) {
3767 MUTEX_EXIT(&rx_quota_mutex);
3773 * Clear the attach wait flag on a connection and proceed.
3775 * Any processing waiting for a connection to be attached should be
3776 * unblocked. We clear the flag and do any other needed tasks.
3779 * the conn to unmark waiting for attach
3781 * @pre conn's conn_data_lock must be locked before calling this function
3785 rxi_ConnClearAttachWait(struct rx_connection *conn)
3787 /* Indicate that rxi_CheckReachEvent is no longer running by
3788 * clearing the flag. Must be atomic under conn_data_lock to
3789 * avoid a new call slipping by: rxi_CheckConnReach holds
3790 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3792 conn->flags &= ~RX_CONN_ATTACHWAIT;
3793 if (conn->flags & RX_CONN_NAT_PING) {
3794 conn->flags &= ~RX_CONN_NAT_PING;
3795 rxi_ScheduleNatKeepAliveEvent(conn);
3800 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3802 struct rx_connection *conn = arg1;
3803 struct rx_call *acall = arg2;
3804 struct rx_call *call = acall;
3805 struct clock when, now;
3808 MUTEX_ENTER(&conn->conn_data_lock);
3811 rxevent_Put(conn->checkReachEvent);
3812 conn->checkReachEvent = NULL;
3815 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3817 putConnection(conn);
3819 MUTEX_EXIT(&conn->conn_data_lock);
3823 MUTEX_ENTER(&conn->conn_call_lock);
3824 MUTEX_ENTER(&conn->conn_data_lock);
3825 for (i = 0; i < RX_MAXCALLS; i++) {
3826 struct rx_call *tc = conn->call[i];
3827 if (tc && tc->state == RX_STATE_PRECALL) {
3833 rxi_ConnClearAttachWait(conn);
3834 MUTEX_EXIT(&conn->conn_data_lock);
3835 MUTEX_EXIT(&conn->conn_call_lock);
3840 MUTEX_ENTER(&call->lock);
3841 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3843 MUTEX_EXIT(&call->lock);
3845 clock_GetTime(&now);
3847 when.sec += RX_CHECKREACH_TIMEOUT;
3848 MUTEX_ENTER(&conn->conn_data_lock);
3849 if (!conn->checkReachEvent) {
3850 MUTEX_ENTER(&rx_refcnt_mutex);
3852 MUTEX_EXIT(&rx_refcnt_mutex);
3853 conn->checkReachEvent = rxevent_Post(&when, &now,
3854 rxi_CheckReachEvent, conn,
3857 MUTEX_EXIT(&conn->conn_data_lock);
3863 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3865 struct rx_service *service = conn->service;
3866 struct rx_peer *peer = conn->peer;
3867 afs_uint32 now, lastReach;
3869 if (service->checkReach == 0)
3873 MUTEX_ENTER(&peer->peer_lock);
3874 lastReach = peer->lastReachTime;
3875 MUTEX_EXIT(&peer->peer_lock);
3876 if (now - lastReach < RX_CHECKREACH_TTL)
3879 MUTEX_ENTER(&conn->conn_data_lock);
3880 if (conn->flags & RX_CONN_ATTACHWAIT) {
3881 MUTEX_EXIT(&conn->conn_data_lock);
3884 conn->flags |= RX_CONN_ATTACHWAIT;
3885 MUTEX_EXIT(&conn->conn_data_lock);
3886 if (!conn->checkReachEvent)
3887 rxi_CheckReachEvent(NULL, conn, call, 0);
3892 /* try to attach call, if authentication is complete */
3894 TryAttach(struct rx_call *acall, osi_socket socket,
3895 int *tnop, struct rx_call **newcallp,
3898 struct rx_connection *conn = acall->conn;
3900 if (conn->type == RX_SERVER_CONNECTION
3901 && acall->state == RX_STATE_PRECALL) {
3902 /* Don't attach until we have any req'd. authentication. */
3903 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3904 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3905 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3906 /* Note: this does not necessarily succeed; there
3907 * may not any proc available
3910 rxi_ChallengeOn(acall->conn);
3915 /* A data packet has been received off the interface. This packet is
3916 * appropriate to the call (the call is in the right state, etc.). This
3917 * routine can return a packet to the caller, for re-use */
3919 static struct rx_packet *
3920 rxi_ReceiveDataPacket(struct rx_call *call,
3921 struct rx_packet *np, int istack,
3922 osi_socket socket, afs_uint32 host, u_short port,
3923 int *tnop, struct rx_call **newcallp)
3925 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3930 afs_uint32 serial=0, flags=0;
3932 struct rx_packet *tnp;
3933 if (rx_stats_active)
3934 rx_atomic_inc(&rx_stats.dataPacketsRead);
3937 /* If there are no packet buffers, drop this new packet, unless we can find
3938 * packet buffers from inactive calls */
3940 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3941 MUTEX_ENTER(&rx_freePktQ_lock);
3942 rxi_NeedMorePackets = TRUE;
3943 MUTEX_EXIT(&rx_freePktQ_lock);
3944 if (rx_stats_active)
3945 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3946 rxi_calltrace(RX_TRACE_DROP, call);
3947 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3948 /* We used to clear the receive queue here, in an attempt to free
3949 * packets. However this is unsafe if the queue has received a
3950 * soft ACK for the final packet */
3951 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3957 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3958 * packet is one of several packets transmitted as a single
3959 * datagram. Do not send any soft or hard acks until all packets
3960 * in a jumbogram have been processed. Send negative acks right away.
3962 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3963 /* tnp is non-null when there are more packets in the
3964 * current jumbo gram */
3971 seq = np->header.seq;
3972 serial = np->header.serial;
3973 flags = np->header.flags;
3975 /* If the call is in an error state, send an abort message */
3977 return rxi_SendCallAbort(call, np, istack, 0);
3979 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3980 * AFS 3.5 jumbogram. */
3981 if (flags & RX_JUMBO_PACKET) {
3982 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3987 if (np->header.spare != 0) {
3988 MUTEX_ENTER(&call->conn->conn_data_lock);
3989 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3990 MUTEX_EXIT(&call->conn->conn_data_lock);
3993 /* The usual case is that this is the expected next packet */
3994 if (seq == call->rnext) {
3996 /* Check to make sure it is not a duplicate of one already queued */
3997 if (!opr_queue_IsEmpty(&call->rq)
3998 && opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq == seq) {
3999 if (rx_stats_active)
4000 rx_atomic_inc(&rx_stats.dupPacketsRead);
4001 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
4002 rxevent_Cancel(&call->delayedAckEvent, call,
4003 RX_CALL_REFCOUNT_DELAY);
4004 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
4010 /* It's the next packet. Stick it on the receive queue
4011 * for this call. Set newPackets to make sure we wake
4012 * the reader once all packets have been processed */
4013 #ifdef RX_TRACK_PACKETS
4014 np->flags |= RX_PKTFLAG_RQ;
4016 opr_queue_Prepend(&call->rq, &np->entry);
4017 #ifdef RXDEBUG_PACKET
4019 #endif /* RXDEBUG_PACKET */
4021 np = NULL; /* We can't use this anymore */
4024 /* If an ack is requested then set a flag to make sure we
4025 * send an acknowledgement for this packet */
4026 if (flags & RX_REQUEST_ACK) {
4027 ackNeeded = RX_ACK_REQUESTED;
4030 /* Keep track of whether we have received the last packet */
4031 if (flags & RX_LAST_PACKET) {
4032 call->flags |= RX_CALL_HAVE_LAST;
4036 /* Check whether we have all of the packets for this call */
4037 if (call->flags & RX_CALL_HAVE_LAST) {
4038 afs_uint32 tseq; /* temporary sequence number */
4039 struct opr_queue *cursor;
4041 for (tseq = seq, opr_queue_Scan(&call->rq, cursor)) {
4042 struct rx_packet *tp;
4044 tp = opr_queue_Entry(cursor, struct rx_packet, entry);
4045 if (tseq != tp->header.seq)
4047 if (tp->header.flags & RX_LAST_PACKET) {
4048 call->flags |= RX_CALL_RECEIVE_DONE;
4055 /* Provide asynchronous notification for those who want it
4056 * (e.g. multi rx) */
4057 if (call->arrivalProc) {
4058 (*call->arrivalProc) (call, call->arrivalProcHandle,
4059 call->arrivalProcArg);
4060 call->arrivalProc = (void (*)())0;
4063 /* Update last packet received */
4066 /* If there is no server process serving this call, grab
4067 * one, if available. We only need to do this once. If a
4068 * server thread is available, this thread becomes a server
4069 * thread and the server thread becomes a listener thread. */
4071 TryAttach(call, socket, tnop, newcallp, 0);
4074 /* This is not the expected next packet. */
4076 /* Determine whether this is a new or old packet, and if it's
4077 * a new one, whether it fits into the current receive window.
4078 * Also figure out whether the packet was delivered in sequence.
4079 * We use the prev variable to determine whether the new packet
4080 * is the successor of its immediate predecessor in the
4081 * receive queue, and the missing flag to determine whether
4082 * any of this packets predecessors are missing. */
4084 afs_uint32 prev; /* "Previous packet" sequence number */
4085 struct opr_queue *cursor;
4086 int missing; /* Are any predecessors missing? */
4088 /* If the new packet's sequence number has been sent to the
4089 * application already, then this is a duplicate */
4090 if (seq < call->rnext) {
4091 if (rx_stats_active)
4092 rx_atomic_inc(&rx_stats.dupPacketsRead);
4093 rxevent_Cancel(&call->delayedAckEvent, call,
4094 RX_CALL_REFCOUNT_DELAY);
4095 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
4101 /* If the sequence number is greater than what can be
4102 * accomodated by the current window, then send a negative
4103 * acknowledge and drop the packet */
4104 if ((call->rnext + call->rwind) <= seq) {
4105 rxevent_Cancel(&call->delayedAckEvent, call,
4106 RX_CALL_REFCOUNT_DELAY);
4107 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
4114 /* Look for the packet in the queue of old received packets */
4115 prev = call->rnext - 1;
4117 for (opr_queue_Scan(&call->rq, cursor)) {
4118 struct rx_packet *tp
4119 = opr_queue_Entry(cursor, struct rx_packet, entry);
4121 /*Check for duplicate packet */
4122 if (seq == tp->header.seq) {
4123 if (rx_stats_active)
4124 rx_atomic_inc(&rx_stats.dupPacketsRead);
4125 rxevent_Cancel(&call->delayedAckEvent, call,
4126 RX_CALL_REFCOUNT_DELAY);
4127 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
4133 /* If we find a higher sequence packet, break out and
4134 * insert the new packet here. */
4135 if (seq < tp->header.seq)
4137 /* Check for missing packet */
4138 if (tp->header.seq != prev + 1) {
4142 prev = tp->header.seq;
4145 /* Keep track of whether we have received the last packet. */
4146 if (flags & RX_LAST_PACKET) {
4147 call->flags |= RX_CALL_HAVE_LAST;
4150 /* It's within the window: add it to the the receive queue.
4151 * tp is left by the previous loop either pointing at the
4152 * packet before which to insert the new packet, or at the
4153 * queue head if the queue is empty or the packet should be
4155 #ifdef RX_TRACK_PACKETS
4156 np->flags |= RX_PKTFLAG_RQ;
4158 #ifdef RXDEBUG_PACKET
4160 #endif /* RXDEBUG_PACKET */
4161 opr_queue_InsertBefore(cursor, &np->entry);
4165 /* Check whether we have all of the packets for this call */
4166 if ((call->flags & RX_CALL_HAVE_LAST)
4167 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4168 afs_uint32 tseq; /* temporary sequence number */
4171 for (opr_queue_Scan(&call->rq, cursor)) {
4172 struct rx_packet *tp
4173 = opr_queue_Entry(cursor, struct rx_packet, entry);
4174 if (tseq != tp->header.seq)
4176 if (tp->header.flags & RX_LAST_PACKET) {
4177 call->flags |= RX_CALL_RECEIVE_DONE;
4184 /* We need to send an ack of the packet is out of sequence,
4185 * or if an ack was requested by the peer. */
4186 if (seq != prev + 1 || missing) {
4187 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4188 } else if (flags & RX_REQUEST_ACK) {
4189 ackNeeded = RX_ACK_REQUESTED;
4192 /* Acknowledge the last packet for each call */
4193 if (flags & RX_LAST_PACKET) {
4204 * If the receiver is waiting for an iovec, fill the iovec
4205 * using the data from the receive queue */
4206 if (call->flags & RX_CALL_IOVEC_WAIT) {
4207 didHardAck = rxi_FillReadVec(call, serial);
4208 /* the call may have been aborted */
4217 /* Wakeup the reader if any */
4218 if ((call->flags & RX_CALL_READER_WAIT)
4219 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4220 || (call->iovNext >= call->iovMax)
4221 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4222 call->flags &= ~RX_CALL_READER_WAIT;
4223 #ifdef RX_ENABLE_LOCKS
4224 CV_BROADCAST(&call->cv_rq);
4226 osi_rxWakeup(&call->rq);
4232 * Send an ack when requested by the peer, or once every
4233 * rxi_SoftAckRate packets until the last packet has been
4234 * received. Always send a soft ack for the last packet in
4235 * the server's reply. */
4237 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4238 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4239 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4240 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4241 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4242 } else if (call->nSoftAcks) {
4243 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4244 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4246 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4247 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4248 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4255 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4257 struct rx_peer *peer = conn->peer;
4259 MUTEX_ENTER(&peer->peer_lock);
4260 peer->lastReachTime = clock_Sec();
4261 MUTEX_EXIT(&peer->peer_lock);
4263 MUTEX_ENTER(&conn->conn_data_lock);
4264 if (conn->flags & RX_CONN_ATTACHWAIT) {
4267 rxi_ConnClearAttachWait(conn);
4268 MUTEX_EXIT(&conn->conn_data_lock);
4270 for (i = 0; i < RX_MAXCALLS; i++) {
4271 struct rx_call *call = conn->call[i];
4274 MUTEX_ENTER(&call->lock);
4275 /* tnop can be null if newcallp is null */
4276 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4278 MUTEX_EXIT(&call->lock);
4282 MUTEX_EXIT(&conn->conn_data_lock);
4285 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4287 rx_ack_reason(int reason)
4290 case RX_ACK_REQUESTED:
4292 case RX_ACK_DUPLICATE:
4294 case RX_ACK_OUT_OF_SEQUENCE:
4296 case RX_ACK_EXCEEDS_WINDOW:
4298 case RX_ACK_NOSPACE:
4302 case RX_ACK_PING_RESPONSE:
4315 /* The real smarts of the whole thing. */
4316 static struct rx_packet *
4317 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4320 struct rx_ackPacket *ap;
4322 struct rx_packet *tp;
4323 struct rx_connection *conn = call->conn;
4324 struct rx_peer *peer = conn->peer;
4325 struct opr_queue *cursor;
4326 struct clock now; /* Current time, for RTT calculations */
4334 int newAckCount = 0;
4335 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4336 int pktsize = 0; /* Set if we need to update the peer mtu */
4337 int conn_data_locked = 0;
4339 if (rx_stats_active)
4340 rx_atomic_inc(&rx_stats.ackPacketsRead);
4341 ap = (struct rx_ackPacket *)rx_DataOf(np);
4342 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4344 return np; /* truncated ack packet */
4346 /* depends on ack packet struct */
4347 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4348 first = ntohl(ap->firstPacket);
4349 prev = ntohl(ap->previousPacket);
4350 serial = ntohl(ap->serial);
4353 * Ignore ack packets received out of order while protecting
4354 * against peers that set the previousPacket field to a packet
4355 * serial number instead of a sequence number.
4357 if (first < call->tfirst ||
4358 (first == call->tfirst && prev < call->tprev && prev < call->tfirst
4365 if (np->header.flags & RX_SLOW_START_OK) {
4366 call->flags |= RX_CALL_SLOW_START_OK;
4369 if (ap->reason == RX_ACK_PING_RESPONSE)
4370 rxi_UpdatePeerReach(conn, call);
4372 if (conn->lastPacketSizeSeq) {
4373 MUTEX_ENTER(&conn->conn_data_lock);
4374 conn_data_locked = 1;
4375 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4376 pktsize = conn->lastPacketSize;
4377 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4380 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4381 if (!conn_data_locked) {
4382 MUTEX_ENTER(&conn->conn_data_lock);
4383 conn_data_locked = 1;
4385 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4386 /* process mtu ping ack */
4387 pktsize = conn->lastPingSize;
4388 conn->lastPingSizeSer = conn->lastPingSize = 0;
4392 if (conn_data_locked) {
4393 MUTEX_EXIT(&conn->conn_data_lock);
4394 conn_data_locked = 0;
4398 if (rxdebug_active) {
4402 len = _snprintf(msg, sizeof(msg),
4403 "tid[%d] RACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4404 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4405 ntohl(ap->serial), ntohl(ap->previousPacket),
4406 (unsigned int)np->header.seq, ntohl(ap->firstPacket),
4407 ap->nAcks, ntohs(ap->bufferSpace) );
4411 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4412 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4416 OutputDebugString(msg);
4418 #else /* AFS_NT40_ENV */
4421 "RACK: reason %x previous %u seq %u serial %u first %u",
4422 ap->reason, ntohl(ap->previousPacket),
4423 (unsigned int)np->header.seq, (unsigned int)serial,
4424 ntohl(ap->firstPacket));
4427 for (offset = 0; offset < nAcks; offset++)
4428 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4433 #endif /* AFS_NT40_ENV */
4436 MUTEX_ENTER(&peer->peer_lock);
4439 * Start somewhere. Can't assume we can send what we can receive,
4440 * but we are clearly receiving.
4442 if (!peer->maxPacketSize)
4443 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4445 if (pktsize > peer->maxPacketSize) {
4446 peer->maxPacketSize = pktsize;
4447 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4448 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4449 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4450 rxi_ScheduleGrowMTUEvent(call, 1);
4455 clock_GetTime(&now);
4457 /* The transmit queue splits into 4 sections.
4459 * The first section is packets which have now been acknowledged
4460 * by a window size change in the ack. These have reached the
4461 * application layer, and may be discarded. These are packets
4462 * with sequence numbers < ap->firstPacket.
4464 * The second section is packets which have sequence numbers in
4465 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4466 * contents of the packet's ack array determines whether these
4467 * packets are acknowledged or not.
4469 * The third section is packets which fall above the range
4470 * addressed in the ack packet. These have not yet been received
4473 * The four section is packets which have not yet been transmitted.
4474 * These packets will have a header.serial of 0.
4477 /* First section - implicitly acknowledged packets that can be
4481 tp = opr_queue_First(&call->tq, struct rx_packet, entry);
4482 while(!opr_queue_IsEnd(&call->tq, &tp->entry) && tp->header.seq < first) {
4483 struct rx_packet *next;
4485 next = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4486 call->tfirst = tp->header.seq + 1;
4488 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4490 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4493 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4494 /* XXX Hack. Because we have to release the global rx lock when sending
4495 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4496 * in rxi_Start sending packets out because packets may move to the
4497 * freePacketQueue as result of being here! So we drop these packets until
4498 * we're safely out of the traversing. Really ugly!
4499 * To make it even uglier, if we're using fine grain locking, we can
4500 * set the ack bits in the packets and have rxi_Start remove the packets
4501 * when it's done transmitting.
4503 if (call->flags & RX_CALL_TQ_BUSY) {
4504 #ifdef RX_ENABLE_LOCKS
4505 tp->flags |= RX_PKTFLAG_ACKED;
4506 call->flags |= RX_CALL_TQ_SOME_ACKED;
4507 #else /* RX_ENABLE_LOCKS */
4509 #endif /* RX_ENABLE_LOCKS */
4511 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4513 opr_queue_Remove(&tp->entry);
4514 #ifdef RX_TRACK_PACKETS
4515 tp->flags &= ~RX_PKTFLAG_TQ;
4517 #ifdef RXDEBUG_PACKET
4519 #endif /* RXDEBUG_PACKET */
4520 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4525 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4527 /* Second section of the queue - packets for which we are receiving
4530 * Go through the explicit acks/nacks and record the results in
4531 * the waiting packets. These are packets that can't be released
4532 * yet, even with a positive acknowledge. This positive
4533 * acknowledge only means the packet has been received by the
4534 * peer, not that it will be retained long enough to be sent to
4535 * the peer's upper level. In addition, reset the transmit timers
4536 * of any missing packets (those packets that must be missing
4537 * because this packet was out of sequence) */
4539 call->nSoftAcked = 0;
4541 while (!opr_queue_IsEnd(&call->tq, &tp->entry)
4542 && tp->header.seq < first + nAcks) {
4543 /* Set the acknowledge flag per packet based on the
4544 * information in the ack packet. An acknowlegded packet can
4545 * be downgraded when the server has discarded a packet it
4546 * soacked previously, or when an ack packet is received
4547 * out of sequence. */
4548 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4549 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4551 tp->flags |= RX_PKTFLAG_ACKED;
4552 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4559 } else /* RX_ACK_TYPE_NACK */ {
4560 tp->flags &= ~RX_PKTFLAG_ACKED;
4564 tp = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4567 /* We don't need to take any action with the 3rd or 4th section in the
4568 * queue - they're not addressed by the contents of this ACK packet.
4571 /* If the window has been extended by this acknowledge packet,
4572 * then wakeup a sender waiting in alloc for window space, or try
4573 * sending packets now, if he's been sitting on packets due to
4574 * lack of window space */
4575 if (call->tnext < (call->tfirst + call->twind)) {
4576 #ifdef RX_ENABLE_LOCKS
4577 CV_SIGNAL(&call->cv_twind);
4579 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4580 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4581 osi_rxWakeup(&call->twind);
4584 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4585 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4589 /* if the ack packet has a receivelen field hanging off it,
4590 * update our state */
4591 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4594 /* If the ack packet has a "recommended" size that is less than
4595 * what I am using now, reduce my size to match */
4596 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4597 (int)sizeof(afs_int32), &tSize);
4598 tSize = (afs_uint32) ntohl(tSize);
4599 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4601 /* Get the maximum packet size to send to this peer */
4602 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4604 tSize = (afs_uint32) ntohl(tSize);
4605 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4606 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4608 /* sanity check - peer might have restarted with different params.
4609 * If peer says "send less", dammit, send less... Peer should never
4610 * be unable to accept packets of the size that prior AFS versions would
4611 * send without asking. */
4612 if (peer->maxMTU != tSize) {
4613 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4615 peer->maxMTU = tSize;
4616 peer->MTU = MIN(tSize, peer->MTU);
4617 call->MTU = MIN(call->MTU, tSize);
4620 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4623 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4624 (int)sizeof(afs_int32), &tSize);
4625 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4626 if (tSize < call->twind) { /* smaller than our send */
4627 call->twind = tSize; /* window, we must send less... */
4628 call->ssthresh = MIN(call->twind, call->ssthresh);
4629 call->conn->twind[call->channel] = call->twind;
4632 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4633 * network MTU confused with the loopback MTU. Calculate the
4634 * maximum MTU here for use in the slow start code below.
4636 /* Did peer restart with older RX version? */
4637 if (peer->maxDgramPackets > 1) {
4638 peer->maxDgramPackets = 1;
4640 } else if (np->length >=
4641 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4644 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4645 sizeof(afs_int32), &tSize);
4646 tSize = (afs_uint32) ntohl(tSize);
4648 * As of AFS 3.5 we set the send window to match the receive window.
4650 if (tSize < call->twind) {
4651 call->twind = tSize;
4652 call->conn->twind[call->channel] = call->twind;
4653 call->ssthresh = MIN(call->twind, call->ssthresh);
4654 } else if (tSize > call->twind) {
4655 call->twind = tSize;
4656 call->conn->twind[call->channel] = call->twind;
4660 * As of AFS 3.5, a jumbogram is more than one fixed size
4661 * packet transmitted in a single UDP datagram. If the remote
4662 * MTU is smaller than our local MTU then never send a datagram
4663 * larger than the natural MTU.
4666 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4667 (int)sizeof(afs_int32), &tSize);
4668 maxDgramPackets = (afs_uint32) ntohl(tSize);
4669 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4671 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4672 if (maxDgramPackets > 1) {
4673 peer->maxDgramPackets = maxDgramPackets;
4674 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4676 peer->maxDgramPackets = 1;
4677 call->MTU = peer->natMTU;
4679 } else if (peer->maxDgramPackets > 1) {
4680 /* Restarted with lower version of RX */
4681 peer->maxDgramPackets = 1;
4683 } else if (peer->maxDgramPackets > 1
4684 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4685 /* Restarted with lower version of RX */
4686 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4687 peer->natMTU = OLD_MAX_PACKET_SIZE;
4688 peer->MTU = OLD_MAX_PACKET_SIZE;
4689 peer->maxDgramPackets = 1;
4690 peer->nDgramPackets = 1;
4692 call->MTU = OLD_MAX_PACKET_SIZE;
4697 * Calculate how many datagrams were successfully received after
4698 * the first missing packet and adjust the negative ack counter
4703 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4704 if (call->nNacks < nNacked) {
4705 call->nNacks = nNacked;
4708 call->nAcks += newAckCount;
4712 /* If the packet contained new acknowledgements, rather than just
4713 * being a duplicate of one we have previously seen, then we can restart
4716 if (newAckCount > 0)
4717 rxi_rto_packet_acked(call, istack);
4719 if (call->flags & RX_CALL_FAST_RECOVER) {
4720 if (newAckCount == 0) {
4721 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4723 call->flags &= ~RX_CALL_FAST_RECOVER;
4724 call->cwind = call->nextCwind;
4725 call->nextCwind = 0;
4728 call->nCwindAcks = 0;
4729 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4730 /* Three negative acks in a row trigger congestion recovery */
4731 call->flags |= RX_CALL_FAST_RECOVER;
4732 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4734 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4735 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4736 call->nextCwind = call->ssthresh;
4739 peer->MTU = call->MTU;
4740 peer->cwind = call->nextCwind;
4741 peer->nDgramPackets = call->nDgramPackets;
4743 call->congestSeq = peer->congestSeq;
4745 /* Reset the resend times on the packets that were nacked
4746 * so we will retransmit as soon as the window permits
4750 for (opr_queue_ScanBackwards(&call->tq, cursor)) {
4751 struct rx_packet *tp =
4752 opr_queue_Entry(cursor, struct rx_packet, entry);
4754 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4755 tp->flags &= ~RX_PKTFLAG_SENT;
4757 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4762 /* If cwind is smaller than ssthresh, then increase
4763 * the window one packet for each ack we receive (exponential
4765 * If cwind is greater than or equal to ssthresh then increase
4766 * the congestion window by one packet for each cwind acks we
4767 * receive (linear growth). */
4768 if (call->cwind < call->ssthresh) {
4770 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4771 call->nCwindAcks = 0;
4773 call->nCwindAcks += newAckCount;
4774 if (call->nCwindAcks >= call->cwind) {
4775 call->nCwindAcks = 0;
4776 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4780 * If we have received several acknowledgements in a row then
4781 * it is time to increase the size of our datagrams
4783 if ((int)call->nAcks > rx_nDgramThreshold) {
4784 if (peer->maxDgramPackets > 1) {
4785 if (call->nDgramPackets < peer->maxDgramPackets) {
4786 call->nDgramPackets++;
4788 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4789 } else if (call->MTU < peer->maxMTU) {
4790 /* don't upgrade if we can't handle it */
4791 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4792 call->MTU = peer->ifMTU;
4794 call->MTU += peer->natMTU;
4795 call->MTU = MIN(call->MTU, peer->maxMTU);
4802 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4804 /* Servers need to hold the call until all response packets have
4805 * been acknowledged. Soft acks are good enough since clients
4806 * are not allowed to clear their receive queues. */
4807 if (call->state == RX_STATE_HOLD
4808 && call->tfirst + call->nSoftAcked >= call->tnext) {
4809 call->state = RX_STATE_DALLY;
4810 rxi_ClearTransmitQueue(call, 0);
4811 rxevent_Cancel(&call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4812 } else if (!opr_queue_IsEmpty(&call->tq)) {
4813 rxi_Start(call, istack);
4818 /* Received a response to a challenge packet */
4819 static struct rx_packet *
4820 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4821 struct rx_packet *np, int istack)
4825 /* Ignore the packet if we're the client */
4826 if (conn->type == RX_CLIENT_CONNECTION)
4829 /* If already authenticated, ignore the packet (it's probably a retry) */
4830 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4833 /* Otherwise, have the security object evaluate the response packet */
4834 error = RXS_CheckResponse(conn->securityObject, conn, np);
4836 /* If the response is invalid, reset the connection, sending
4837 * an abort to the peer */
4841 rxi_ConnectionError(conn, error);
4842 MUTEX_ENTER(&conn->conn_data_lock);
4843 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4844 MUTEX_EXIT(&conn->conn_data_lock);
4847 /* If the response is valid, any calls waiting to attach
4848 * servers can now do so */
4851 for (i = 0; i < RX_MAXCALLS; i++) {
4852 struct rx_call *call = conn->call[i];
4854 MUTEX_ENTER(&call->lock);
4855 if (call->state == RX_STATE_PRECALL)
4856 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4857 /* tnop can be null if newcallp is null */
4858 MUTEX_EXIT(&call->lock);
4862 /* Update the peer reachability information, just in case
4863 * some calls went into attach-wait while we were waiting
4864 * for authentication..
4866 rxi_UpdatePeerReach(conn, NULL);
4871 /* A client has received an authentication challenge: the security
4872 * object is asked to cough up a respectable response packet to send
4873 * back to the server. The server is responsible for retrying the
4874 * challenge if it fails to get a response. */
4876 static struct rx_packet *
4877 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4878 struct rx_packet *np, int istack)
4882 /* Ignore the challenge if we're the server */
4883 if (conn->type == RX_SERVER_CONNECTION)
4886 /* Ignore the challenge if the connection is otherwise idle; someone's
4887 * trying to use us as an oracle. */
4888 if (!rxi_HasActiveCalls(conn))
4891 /* Send the security object the challenge packet. It is expected to fill
4892 * in the response. */
4893 error = RXS_GetResponse(conn->securityObject, conn, np);
4895 /* If the security object is unable to return a valid response, reset the
4896 * connection and send an abort to the peer. Otherwise send the response
4897 * packet to the peer connection. */
4899 rxi_ConnectionError(conn, error);
4900 MUTEX_ENTER(&conn->conn_data_lock);
4901 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4902 MUTEX_EXIT(&conn->conn_data_lock);
4904 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4905 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4911 /* Find an available server process to service the current request in
4912 * the given call structure. If one isn't available, queue up this
4913 * call so it eventually gets one */
4915 rxi_AttachServerProc(struct rx_call *call,
4916 osi_socket socket, int *tnop,
4917 struct rx_call **newcallp)
4919 struct rx_serverQueueEntry *sq;
4920 struct rx_service *service = call->conn->service;
4923 /* May already be attached */
4924 if (call->state == RX_STATE_ACTIVE)
4927 MUTEX_ENTER(&rx_serverPool_lock);
4929 haveQuota = QuotaOK(service);
4930 if ((!haveQuota) || opr_queue_IsEmpty(&rx_idleServerQueue)) {
4931 /* If there are no processes available to service this call,
4932 * put the call on the incoming call queue (unless it's
4933 * already on the queue).
4935 #ifdef RX_ENABLE_LOCKS
4937 ReturnToServerPool(service);
4938 #endif /* RX_ENABLE_LOCKS */
4940 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4941 call->flags |= RX_CALL_WAIT_PROC;
4942 rx_atomic_inc(&rx_nWaiting);
4943 rx_atomic_inc(&rx_nWaited);
4944 rxi_calltrace(RX_CALL_ARRIVAL, call);
4945 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4946 opr_queue_Append(&rx_incomingCallQueue, &call->entry);
4949 sq = opr_queue_Last(&rx_idleServerQueue,
4950 struct rx_serverQueueEntry, entry);
4952 /* If hot threads are enabled, and both newcallp and sq->socketp
4953 * are non-null, then this thread will process the call, and the
4954 * idle server thread will start listening on this threads socket.
4956 opr_queue_Remove(&sq->entry);
4958 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4961 *sq->socketp = socket;
4962 clock_GetTime(&call->startTime);
4963 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4967 if (call->flags & RX_CALL_WAIT_PROC) {
4968 /* Conservative: I don't think this should happen */
4969 call->flags &= ~RX_CALL_WAIT_PROC;
4970 rx_atomic_dec(&rx_nWaiting);
4971 if (opr_queue_IsOnQueue(&call->entry)) {
4972 opr_queue_Remove(&call->entry);
4975 call->state = RX_STATE_ACTIVE;
4976 call->mode = RX_MODE_RECEIVING;
4977 #ifdef RX_KERNEL_TRACE
4979 int glockOwner = ISAFS_GLOCK();
4982 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4983 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4989 if (call->flags & RX_CALL_CLEARED) {
4990 /* send an ack now to start the packet flow up again */
4991 call->flags &= ~RX_CALL_CLEARED;
4992 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4994 #ifdef RX_ENABLE_LOCKS
4997 service->nRequestsRunning++;
4998 MUTEX_ENTER(&rx_quota_mutex);
4999 if (service->nRequestsRunning <= service->minProcs)
5002 MUTEX_EXIT(&rx_quota_mutex);
5006 MUTEX_EXIT(&rx_serverPool_lock);
5009 /* Delay the sending of an acknowledge event for a short while, while
5010 * a new call is being prepared (in the case of a client) or a reply
5011 * is being prepared (in the case of a server). Rather than sending
5012 * an ack packet, an ACKALL packet is sent. */
5014 rxi_AckAll(struct rx_call *call)
5016 rxi_SendSpecial(call, call->conn, NULL, RX_PACKET_TYPE_ACKALL,
5018 call->flags |= RX_CALL_ACKALL_SENT;
5022 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
5025 struct rx_call *call = arg1;
5026 #ifdef RX_ENABLE_LOCKS
5028 MUTEX_ENTER(&call->lock);
5029 if (event == call->delayedAckEvent) {
5030 rxevent_Put(call->delayedAckEvent);
5031 call->delayedAckEvent = NULL;
5033 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
5035 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
5037 MUTEX_EXIT(&call->lock);
5038 #else /* RX_ENABLE_LOCKS */
5040 rxevent_Put(call->delayedAckEvent);
5041 call->delayedAckEvent = NULL;
5043 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
5044 #endif /* RX_ENABLE_LOCKS */
5048 #ifdef RX_ENABLE_LOCKS
5049 /* Set ack in all packets in transmit queue. rxi_Start will deal with
5050 * clearing them out.
5053 rxi_SetAcksInTransmitQueue(struct rx_call *call)
5055 struct opr_queue *cursor;
5058 for (opr_queue_Scan(&call->tq, cursor)) {
5060 = opr_queue_Entry(cursor, struct rx_packet, entry);
5062 p->flags |= RX_PKTFLAG_ACKED;
5067 call->flags |= RX_CALL_TQ_CLEARME;
5068 call->flags |= RX_CALL_TQ_SOME_ACKED;
5071 rxi_rto_cancel(call);
5073 call->tfirst = call->tnext;
5074 call->nSoftAcked = 0;
5076 if (call->flags & RX_CALL_FAST_RECOVER) {
5077 call->flags &= ~RX_CALL_FAST_RECOVER;
5078 call->cwind = call->nextCwind;
5079 call->nextCwind = 0;
5082 CV_SIGNAL(&call->cv_twind);
5084 #endif /* RX_ENABLE_LOCKS */
5086 /* Clear out the transmit queue for the current call (all packets have
5087 * been received by peer) */
5089 rxi_ClearTransmitQueue(struct rx_call *call, int force)
5091 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5092 struct opr_queue *cursor;
5093 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
5095 for (opr_queue_Scan(&call->tq, cursor)) {
5097 = opr_queue_Entry(cursor, struct rx_packet, entry);
5099 p->flags |= RX_PKTFLAG_ACKED;
5103 call->flags |= RX_CALL_TQ_CLEARME;
5104 call->flags |= RX_CALL_TQ_SOME_ACKED;
5107 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5108 #ifdef RXDEBUG_PACKET
5110 #endif /* RXDEBUG_PACKET */
5111 rxi_FreePackets(0, &call->tq);
5112 rxi_WakeUpTransmitQueue(call);
5113 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5114 call->flags &= ~RX_CALL_TQ_CLEARME;
5116 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5118 rxi_rto_cancel(call);
5119 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5120 call->nSoftAcked = 0;
5122 if (call->flags & RX_CALL_FAST_RECOVER) {
5123 call->flags &= ~RX_CALL_FAST_RECOVER;
5124 call->cwind = call->nextCwind;
5126 #ifdef RX_ENABLE_LOCKS
5127 CV_SIGNAL(&call->cv_twind);
5129 osi_rxWakeup(&call->twind);
5134 rxi_ClearReceiveQueue(struct rx_call *call)
5136 if (!opr_queue_IsEmpty(&call->rq)) {
5139 count = rxi_FreePackets(0, &call->rq);
5140 rx_packetReclaims += count;
5141 #ifdef RXDEBUG_PACKET
5143 if ( call->rqc != 0 )
5144 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5146 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5148 if (call->state == RX_STATE_PRECALL) {
5149 call->flags |= RX_CALL_CLEARED;
5153 /* Send an abort packet for the specified call */
5154 static struct rx_packet *
5155 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5156 int istack, int force)
5158 afs_int32 error, cerror;
5159 struct clock when, now;
5164 switch (call->error) {
5167 cerror = RX_CALL_TIMEOUT;
5170 cerror = call->error;
5173 /* Clients should never delay abort messages */
5174 if (rx_IsClientConn(call->conn))
5177 if (call->abortCode != cerror) {
5178 call->abortCode = cerror;
5179 call->abortCount = 0;
5182 if (force || rxi_callAbortThreshhold == 0
5183 || call->abortCount < rxi_callAbortThreshhold) {
5184 if (call->delayedAbortEvent) {
5185 rxevent_Cancel(&call->delayedAbortEvent, call,
5186 RX_CALL_REFCOUNT_ABORT);
5188 error = htonl(cerror);
5191 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5192 (char *)&error, sizeof(error), istack);
5193 } else if (!call->delayedAbortEvent) {
5194 clock_GetTime(&now);
5196 clock_Addmsec(&when, rxi_callAbortDelay);
5197 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5198 call->delayedAbortEvent =
5199 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5204 /* Send an abort packet for the specified connection. Packet is an
5205 * optional pointer to a packet that can be used to send the abort.
5206 * Once the number of abort messages reaches the threshhold, an
5207 * event is scheduled to send the abort. Setting the force flag
5208 * overrides sending delayed abort messages.
5210 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5211 * to send the abort packet.
5214 rxi_SendConnectionAbort(struct rx_connection *conn,
5215 struct rx_packet *packet, int istack, int force)
5218 struct clock when, now;
5223 /* Clients should never delay abort messages */
5224 if (rx_IsClientConn(conn))
5227 if (force || rxi_connAbortThreshhold == 0
5228 || conn->abortCount < rxi_connAbortThreshhold) {
5230 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
5231 error = htonl(conn->error);
5233 MUTEX_EXIT(&conn->conn_data_lock);
5235 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5236 RX_PACKET_TYPE_ABORT, (char *)&error,
5237 sizeof(error), istack);
5238 MUTEX_ENTER(&conn->conn_data_lock);
5239 } else if (!conn->delayedAbortEvent) {
5240 clock_GetTime(&now);
5242 clock_Addmsec(&when, rxi_connAbortDelay);
5243 conn->delayedAbortEvent =
5244 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
5249 /* Associate an error all of the calls owned by a connection. Called
5250 * with error non-zero. This is only for really fatal things, like
5251 * bad authentication responses. The connection itself is set in
5252 * error at this point, so that future packets received will be
5255 rxi_ConnectionError(struct rx_connection *conn,
5261 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5263 MUTEX_ENTER(&conn->conn_data_lock);
5264 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
5265 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
5266 if (conn->checkReachEvent) {
5267 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
5268 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5269 putConnection(conn);
5271 MUTEX_EXIT(&conn->conn_data_lock);
5272 for (i = 0; i < RX_MAXCALLS; i++) {
5273 struct rx_call *call = conn->call[i];
5275 MUTEX_ENTER(&call->lock);
5276 rxi_CallError(call, error);
5277 MUTEX_EXIT(&call->lock);
5280 conn->error = error;
5281 if (rx_stats_active)
5282 rx_atomic_inc(&rx_stats.fatalErrors);
5287 * Interrupt an in-progress call with the specified error and wakeup waiters.
5289 * @param[in] call The call to interrupt
5290 * @param[in] error The error code to send to the peer
5293 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5295 MUTEX_ENTER(&call->lock);
5296 rxi_CallError(call, error);
5297 rxi_SendCallAbort(call, NULL, 0, 1);
5298 MUTEX_EXIT(&call->lock);
5302 rxi_CallError(struct rx_call *call, afs_int32 error)
5305 osirx_AssertMine(&call->lock, "rxi_CallError");
5307 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5309 error = call->error;
5311 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5312 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5313 rxi_ResetCall(call, 0);
5316 rxi_ResetCall(call, 0);
5318 call->error = error;
5321 /* Reset various fields in a call structure, and wakeup waiting
5322 * processes. Some fields aren't changed: state & mode are not
5323 * touched (these must be set by the caller), and bufptr, nLeft, and
5324 * nFree are not reset, since these fields are manipulated by
5325 * unprotected macros, and may only be reset by non-interrupting code.
5329 rxi_ResetCall(struct rx_call *call, int newcall)
5332 struct rx_peer *peer;
5333 struct rx_packet *packet;
5335 osirx_AssertMine(&call->lock, "rxi_ResetCall");
5337 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5339 /* Notify anyone who is waiting for asynchronous packet arrival */
5340 if (call->arrivalProc) {
5341 (*call->arrivalProc) (call, call->arrivalProcHandle,
5342 call->arrivalProcArg);
5343 call->arrivalProc = (void (*)())0;
5347 rxevent_Cancel(&call->growMTUEvent, call, RX_CALL_REFCOUNT_MTU);
5349 if (call->delayedAbortEvent) {
5350 rxevent_Cancel(&call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
5351 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5353 rxi_SendCallAbort(call, packet, 0, 1);
5354 rxi_FreePacket(packet);
5359 * Update the peer with the congestion information in this call
5360 * so other calls on this connection can pick up where this call
5361 * left off. If the congestion sequence numbers don't match then
5362 * another call experienced a retransmission.
5364 peer = call->conn->peer;
5365 MUTEX_ENTER(&peer->peer_lock);
5367 if (call->congestSeq == peer->congestSeq) {
5368 peer->cwind = MAX(peer->cwind, call->cwind);
5369 peer->MTU = MAX(peer->MTU, call->MTU);
5370 peer->nDgramPackets =
5371 MAX(peer->nDgramPackets, call->nDgramPackets);
5374 call->abortCode = 0;
5375 call->abortCount = 0;
5377 if (peer->maxDgramPackets > 1) {
5378 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5380 call->MTU = peer->MTU;
5382 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5383 call->ssthresh = rx_maxSendWindow;
5384 call->nDgramPackets = peer->nDgramPackets;
5385 call->congestSeq = peer->congestSeq;
5386 call->rtt = peer->rtt;
5387 call->rtt_dev = peer->rtt_dev;
5388 clock_Zero(&call->rto);
5389 clock_Addmsec(&call->rto,
5390 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5391 MUTEX_EXIT(&peer->peer_lock);
5393 flags = call->flags;
5394 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5395 rxi_WaitforTQBusy(call);
5396 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5398 rxi_ClearTransmitQueue(call, 1);
5399 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5400 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5404 if (!newcall && (flags & RX_CALL_PEER_BUSY)) {
5405 /* The call channel is still busy; resetting the call doesn't change
5406 * that. However, if 'newcall' is set, we are processing a call
5407 * structure that has either been recycled from the free list, or has
5408 * been newly allocated. So, RX_CALL_PEER_BUSY is not relevant if
5409 * 'newcall' is set, since it describes a completely different call
5410 * channel which we do not care about. */
5411 call->flags |= RX_CALL_PEER_BUSY;
5414 rxi_ClearReceiveQueue(call);
5415 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5419 call->twind = call->conn->twind[call->channel];
5420 call->rwind = call->conn->rwind[call->channel];
5421 call->nSoftAcked = 0;
5422 call->nextCwind = 0;
5425 call->nCwindAcks = 0;
5426 call->nSoftAcks = 0;
5427 call->nHardAcks = 0;
5429 call->tfirst = call->rnext = call->tnext = 1;
5432 call->lastAcked = 0;
5433 call->localStatus = call->remoteStatus = 0;
5435 if (flags & RX_CALL_READER_WAIT) {
5436 #ifdef RX_ENABLE_LOCKS
5437 CV_BROADCAST(&call->cv_rq);
5439 osi_rxWakeup(&call->rq);
5442 if (flags & RX_CALL_WAIT_PACKETS) {
5443 MUTEX_ENTER(&rx_freePktQ_lock);
5444 rxi_PacketsUnWait(); /* XXX */
5445 MUTEX_EXIT(&rx_freePktQ_lock);
5447 #ifdef RX_ENABLE_LOCKS
5448 CV_SIGNAL(&call->cv_twind);
5450 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5451 osi_rxWakeup(&call->twind);
5454 if (flags & RX_CALL_WAIT_PROC) {
5455 rx_atomic_dec(&rx_nWaiting);
5457 #ifdef RX_ENABLE_LOCKS
5458 /* The following ensures that we don't mess with any queue while some
5459 * other thread might also be doing so. The call_queue_lock field is
5460 * is only modified under the call lock. If the call is in the process
5461 * of being removed from a queue, the call is not locked until the
5462 * the queue lock is dropped and only then is the call_queue_lock field
5463 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5464 * Note that any other routine which removes a call from a queue has to
5465 * obtain the queue lock before examing the queue and removing the call.
5467 if (call->call_queue_lock) {
5468 MUTEX_ENTER(call->call_queue_lock);
5469 if (opr_queue_IsOnQueue(&call->entry)) {
5470 opr_queue_Remove(&call->entry);
5472 MUTEX_EXIT(call->call_queue_lock);
5473 CLEAR_CALL_QUEUE_LOCK(call);
5475 #else /* RX_ENABLE_LOCKS */
5476 if (opr_queue_IsOnQueue(&call->entry)) {
5477 opr_queue_Remove(&call->entry);
5479 #endif /* RX_ENABLE_LOCKS */
5481 rxi_KeepAliveOff(call);
5482 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5485 /* Send an acknowledge for the indicated packet (seq,serial) of the
5486 * indicated call, for the indicated reason (reason). This
5487 * acknowledge will specifically acknowledge receiving the packet, and
5488 * will also specify which other packets for this call have been
5489 * received. This routine returns the packet that was used to the
5490 * caller. The caller is responsible for freeing it or re-using it.
5491 * This acknowledgement also returns the highest sequence number
5492 * actually read out by the higher level to the sender; the sender
5493 * promises to keep around packets that have not been read by the
5494 * higher level yet (unless, of course, the sender decides to abort
5495 * the call altogether). Any of p, seq, serial, pflags, or reason may
5496 * be set to zero without ill effect. That is, if they are zero, they
5497 * will not convey any information.
5498 * NOW there is a trailer field, after the ack where it will safely be
5499 * ignored by mundanes, which indicates the maximum size packet this
5500 * host can swallow. */
5502 struct rx_packet *optionalPacket; use to send ack (or null)
5503 int seq; Sequence number of the packet we are acking
5504 int serial; Serial number of the packet
5505 int pflags; Flags field from packet header
5506 int reason; Reason an acknowledge was prompted
5510 rxi_SendAck(struct rx_call *call,
5511 struct rx_packet *optionalPacket, int serial, int reason,
5514 struct rx_ackPacket *ap;
5515 struct rx_packet *p;
5516 struct opr_queue *cursor;
5519 afs_uint32 padbytes = 0;
5520 #ifdef RX_ENABLE_TSFPQ
5521 struct rx_ts_info_t * rx_ts_info;
5525 * Open the receive window once a thread starts reading packets
5527 if (call->rnext > 1) {
5528 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5531 /* Don't attempt to grow MTU if this is a critical ping */
5532 if (reason == RX_ACK_MTU) {
5533 /* keep track of per-call attempts, if we're over max, do in small
5534 * otherwise in larger? set a size to increment by, decrease
5537 if (call->conn->peer->maxPacketSize &&
5538 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5540 padbytes = call->conn->peer->maxPacketSize+16;
5542 padbytes = call->conn->peer->maxMTU + 128;
5544 /* do always try a minimum size ping */
5545 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5547 /* subtract the ack payload */
5548 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5549 reason = RX_ACK_PING;
5552 call->nHardAcks = 0;
5553 call->nSoftAcks = 0;
5554 if (call->rnext > call->lastAcked)
5555 call->lastAcked = call->rnext;
5559 rx_computelen(p, p->length); /* reset length, you never know */
5560 } /* where that's been... */
5561 #ifdef RX_ENABLE_TSFPQ
5563 RX_TS_INFO_GET(rx_ts_info);
5564 if ((p = rx_ts_info->local_special_packet)) {
5565 rx_computelen(p, p->length);
5566 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5567 rx_ts_info->local_special_packet = p;
5568 } else { /* We won't send the ack, but don't panic. */
5569 return optionalPacket;
5573 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5574 /* We won't send the ack, but don't panic. */
5575 return optionalPacket;
5580 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5583 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5584 #ifndef RX_ENABLE_TSFPQ
5585 if (!optionalPacket)
5588 return optionalPacket;
5590 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5591 if (rx_Contiguous(p) < templ) {
5592 #ifndef RX_ENABLE_TSFPQ
5593 if (!optionalPacket)
5596 return optionalPacket;
5601 /* MTUXXX failing to send an ack is very serious. We should */
5602 /* try as hard as possible to send even a partial ack; it's */
5603 /* better than nothing. */
5604 ap = (struct rx_ackPacket *)rx_DataOf(p);
5605 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5606 ap->reason = reason;
5608 /* The skew computation used to be bogus, I think it's better now. */
5609 /* We should start paying attention to skew. XXX */
5610 ap->serial = htonl(serial);
5611 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5614 * First packet not yet forwarded to reader. When ACKALL has been
5615 * sent the peer has been told that all received packets will be
5616 * delivered to the reader. The value 'rnext' is used internally
5617 * to refer to the next packet in the receive queue that must be
5618 * delivered to the reader. From the perspective of the peer it
5619 * already has so report the last sequence number plus one if there
5620 * are packets in the receive queue awaiting processing.
5622 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5623 !opr_queue_IsEmpty(&call->rq)) {
5624 ap->firstPacket = htonl(opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq + 1);
5626 ap->firstPacket = htonl(call->rnext);
5628 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5630 /* No fear of running out of ack packet here because there can only
5631 * be at most one window full of unacknowledged packets. The window
5632 * size must be constrained to be less than the maximum ack size,
5633 * of course. Also, an ack should always fit into a single packet
5634 * -- it should not ever be fragmented. */
5636 for (opr_queue_Scan(&call->rq, cursor)) {
5637 struct rx_packet *rqp
5638 = opr_queue_Entry(cursor, struct rx_packet, entry);
5640 if (!rqp || !call->rq.next
5641 || (rqp->header.seq > (call->rnext + call->rwind))) {
5642 #ifndef RX_ENABLE_TSFPQ
5643 if (!optionalPacket)
5646 rxi_CallError(call, RX_CALL_DEAD);
5647 return optionalPacket;
5650 while (rqp->header.seq > call->rnext + offset)
5651 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5652 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5654 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5655 #ifndef RX_ENABLE_TSFPQ
5656 if (!optionalPacket)
5659 rxi_CallError(call, RX_CALL_DEAD);
5660 return optionalPacket;
5666 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5668 /* these are new for AFS 3.3 */
5669 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5670 templ = htonl(templ);
5671 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5672 templ = htonl(call->conn->peer->ifMTU);
5673 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5674 sizeof(afs_int32), &templ);
5676 /* new for AFS 3.4 */
5677 templ = htonl(call->rwind);
5678 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5679 sizeof(afs_int32), &templ);
5681 /* new for AFS 3.5 */
5682 templ = htonl(call->conn->peer->ifDgramPackets);
5683 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5684 sizeof(afs_int32), &templ);
5686 p->header.serviceId = call->conn->serviceId;
5687 p->header.cid = (call->conn->cid | call->channel);
5688 p->header.callNumber = *call->callNumber;
5690 p->header.securityIndex = call->conn->securityIndex;
5691 p->header.epoch = call->conn->epoch;
5692 p->header.type = RX_PACKET_TYPE_ACK;
5693 p->header.flags = RX_SLOW_START_OK;
5694 if (reason == RX_ACK_PING) {
5695 p->header.flags |= RX_REQUEST_ACK;
5697 p->length = padbytes +
5698 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5701 /* not fast but we can potentially use this if truncated
5702 * fragments are delivered to figure out the mtu.
5704 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5705 sizeof(afs_int32), sizeof(afs_int32),
5709 if (call->conn->type == RX_CLIENT_CONNECTION)
5710 p->header.flags |= RX_CLIENT_INITIATED;
5714 if (rxdebug_active) {
5718 len = _snprintf(msg, sizeof(msg),
5719 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5720 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5721 ntohl(ap->serial), ntohl(ap->previousPacket),
5722 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5723 ap->nAcks, ntohs(ap->bufferSpace) );
5727 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5728 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5732 OutputDebugString(msg);
5734 #else /* AFS_NT40_ENV */
5736 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5737 ap->reason, ntohl(ap->previousPacket),
5738 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5740 for (offset = 0; offset < ap->nAcks; offset++)
5741 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5746 #endif /* AFS_NT40_ENV */
5749 int i, nbytes = p->length;
5751 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5752 if (nbytes <= p->wirevec[i].iov_len) {
5755 savelen = p->wirevec[i].iov_len;
5757 p->wirevec[i].iov_len = nbytes;
5759 rxi_Send(call, p, istack);
5760 p->wirevec[i].iov_len = savelen;
5764 nbytes -= p->wirevec[i].iov_len;
5767 if (rx_stats_active)
5768 rx_atomic_inc(&rx_stats.ackPacketsSent);
5769 #ifndef RX_ENABLE_TSFPQ
5770 if (!optionalPacket)
5773 return optionalPacket; /* Return packet for re-use by caller */
5777 struct rx_packet **list;
5782 /* Send all of the packets in the list in single datagram */
5784 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5785 int istack, int moreFlag)
5791 struct rx_connection *conn = call->conn;
5792 struct rx_peer *peer = conn->peer;
5794 MUTEX_ENTER(&peer->peer_lock);
5795 peer->nSent += xmit->len;
5796 if (xmit->resending)
5797 peer->reSends += xmit->len;
5798 MUTEX_EXIT(&peer->peer_lock);
5800 if (rx_stats_active) {
5801 if (xmit->resending)
5802 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5804 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5807 clock_GetTime(&now);
5809 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5813 /* Set the packet flags and schedule the resend events */
5814 /* Only request an ack for the last packet in the list */
5815 for (i = 0; i < xmit->len; i++) {
5816 struct rx_packet *packet = xmit->list[i];
5818 /* Record the time sent */
5819 packet->timeSent = now;
5820 packet->flags |= RX_PKTFLAG_SENT;
5822 /* Ask for an ack on retransmitted packets, on every other packet
5823 * if the peer doesn't support slow start. Ask for an ack on every
5824 * packet until the congestion window reaches the ack rate. */
5825 if (packet->header.serial) {
5828 packet->firstSent = now;
5829 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5830 || (!(call->flags & RX_CALL_SLOW_START_OK)
5831 && (packet->header.seq & 1)))) {
5836 /* Tag this packet as not being the last in this group,
5837 * for the receiver's benefit */
5838 if (i < xmit->len - 1 || moreFlag) {
5839 packet->header.flags |= RX_MORE_PACKETS;
5844 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5847 /* Since we're about to send a data packet to the peer, it's
5848 * safe to nuke any scheduled end-of-packets ack */
5849 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5851 MUTEX_EXIT(&call->lock);
5852 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5853 if (xmit->len > 1) {
5854 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5856 rxi_SendPacket(call, conn, xmit->list[0], istack);
5858 MUTEX_ENTER(&call->lock);
5859 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5861 /* Tell the RTO calculation engine that we have sent a packet, and
5862 * if it was the last one */
5863 rxi_rto_packet_sent(call, lastPacket, istack);
5865 /* Update last send time for this call (for keep-alive
5866 * processing), and for the connection (so that we can discover
5867 * idle connections) */
5868 conn->lastSendTime = call->lastSendTime = clock_Sec();
5869 /* Let a set of retransmits trigger an idle timeout */
5870 if (!xmit->resending)
5871 call->lastSendData = call->lastSendTime;
5874 /* When sending packets we need to follow these rules:
5875 * 1. Never send more than maxDgramPackets in a jumbogram.
5876 * 2. Never send a packet with more than two iovecs in a jumbogram.
5877 * 3. Never send a retransmitted packet in a jumbogram.
5878 * 4. Never send more than cwind/4 packets in a jumbogram
5879 * We always keep the last list we should have sent so we
5880 * can set the RX_MORE_PACKETS flags correctly.
5884 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5889 struct xmitlist working;
5890 struct xmitlist last;
5892 struct rx_peer *peer = call->conn->peer;
5893 int morePackets = 0;
5895 memset(&last, 0, sizeof(struct xmitlist));
5896 working.list = &list[0];
5898 working.resending = 0;
5900 recovery = call->flags & RX_CALL_FAST_RECOVER;
5902 for (i = 0; i < len; i++) {
5903 /* Does the current packet force us to flush the current list? */
5905 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5906 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5908 /* This sends the 'last' list and then rolls the current working
5909 * set into the 'last' one, and resets the working set */
5912 rxi_SendList(call, &last, istack, 1);
5913 /* If the call enters an error state stop sending, or if
5914 * we entered congestion recovery mode, stop sending */
5916 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5921 working.resending = 0;
5922 working.list = &list[i];
5924 /* Add the current packet to the list if it hasn't been acked.
5925 * Otherwise adjust the list pointer to skip the current packet. */
5926 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5929 if (list[i]->header.serial)
5930 working.resending = 1;
5932 /* Do we need to flush the list? */
5933 if (working.len >= (int)peer->maxDgramPackets
5934 || working.len >= (int)call->nDgramPackets
5935 || working.len >= (int)call->cwind
5936 || list[i]->header.serial
5937 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5939 rxi_SendList(call, &last, istack, 1);
5940 /* If the call enters an error state stop sending, or if
5941 * we entered congestion recovery mode, stop sending */
5943 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5948 working.resending = 0;
5949 working.list = &list[i + 1];
5952 if (working.len != 0) {
5953 osi_Panic("rxi_SendList error");
5955 working.list = &list[i + 1];
5959 /* Send the whole list when the call is in receive mode, when
5960 * the call is in eof mode, when we are in fast recovery mode,
5961 * and when we have the last packet */
5962 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5963 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5964 || (call->flags & RX_CALL_FAST_RECOVER)) {
5965 /* Check for the case where the current list contains
5966 * an acked packet. Since we always send retransmissions
5967 * in a separate packet, we only need to check the first
5968 * packet in the list */
5969 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5973 rxi_SendList(call, &last, istack, morePackets);
5974 /* If the call enters an error state stop sending, or if
5975 * we entered congestion recovery mode, stop sending */
5977 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5981 rxi_SendList(call, &working, istack, 0);
5983 } else if (last.len > 0) {
5984 rxi_SendList(call, &last, istack, 0);
5985 /* Packets which are in 'working' are not sent by this call */
5990 * Check if the peer for the given call is known to be dead
5992 * If the call's peer appears dead (it has encountered fatal network errors
5993 * since the call started) the call is killed with RX_CALL_DEAD if the call
5994 * is active. Otherwise, we do nothing.
5996 * @param[in] call The call to check
5999 * @retval 0 The call is fine, and we haven't done anything to the call
6000 * @retval nonzero The call's peer appears dead, and the call has been
6001 * terminated if it was active
6003 * @pre call->lock must be locked
6006 rxi_CheckPeerDead(struct rx_call *call)
6008 #ifdef AFS_RXERRQ_ENV
6011 if (call->state == RX_STATE_DALLY) {
6015 peererrs = rx_atomic_read(&call->conn->peer->neterrs);
6016 if (call->neterr_gen < peererrs) {
6017 /* we have received network errors since this call started; kill
6019 if (call->state == RX_STATE_ACTIVE) {
6020 rxi_CallError(call, RX_CALL_DEAD);
6024 if (call->neterr_gen > peererrs) {
6025 /* someone has reset the number of peer errors; set the call error gen
6026 * so we can detect if more errors are encountered */
6027 call->neterr_gen = peererrs;
6034 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
6036 struct rx_call *call = arg0;
6037 struct rx_peer *peer;
6038 struct opr_queue *cursor;
6039 struct clock maxTimeout = { 60, 0 };
6041 MUTEX_ENTER(&call->lock);
6043 peer = call->conn->peer;
6045 /* Make sure that the event pointer is removed from the call
6046 * structure, since there is no longer a per-call retransmission
6048 if (event == call->resendEvent) {
6049 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
6050 rxevent_Put(call->resendEvent);
6051 call->resendEvent = NULL;
6054 rxi_CheckPeerDead(call);
6056 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
6057 rxi_CheckBusy(call);
6060 if (opr_queue_IsEmpty(&call->tq)) {
6061 /* Nothing to do. This means that we've been raced, and that an
6062 * ACK has come in between when we were triggered, and when we
6063 * actually got to run. */
6067 /* We're in loss recovery */
6068 call->flags |= RX_CALL_FAST_RECOVER;
6070 /* Mark all of the pending packets in the queue as being lost */
6071 for (opr_queue_Scan(&call->tq, cursor)) {
6072 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
6073 if (!(p->flags & RX_PKTFLAG_ACKED))
6074 p->flags &= ~RX_PKTFLAG_SENT;
6077 /* We're resending, so we double the timeout of the call. This will be
6078 * dropped back down by the first successful ACK that we receive.
6080 * We apply a maximum value here of 60 seconds
6082 clock_Add(&call->rto, &call->rto);
6083 if (clock_Gt(&call->rto, &maxTimeout))
6084 call->rto = maxTimeout;
6086 /* Packet loss is most likely due to congestion, so drop our window size
6087 * and start again from the beginning */
6088 if (peer->maxDgramPackets >1) {
6089 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
6090 call->MTU = MIN(peer->natMTU, peer->maxMTU);
6092 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
6093 call->nDgramPackets = 1;
6095 call->nextCwind = 1;
6098 MUTEX_ENTER(&peer->peer_lock);
6099 peer->MTU = call->MTU;
6100 peer->cwind = call->cwind;
6101 peer->nDgramPackets = 1;
6103 call->congestSeq = peer->congestSeq;
6104 MUTEX_EXIT(&peer->peer_lock);
6106 rxi_Start(call, istack);
6109 MUTEX_EXIT(&call->lock);
6112 /* This routine is called when new packets are readied for
6113 * transmission and when retransmission may be necessary, or when the
6114 * transmission window or burst count are favourable. This should be
6115 * better optimized for new packets, the usual case, now that we've
6116 * got rid of queues of send packets. XXXXXXXXXXX */
6118 rxi_Start(struct rx_call *call, int istack)
6120 struct opr_queue *cursor;
6121 #ifdef RX_ENABLE_LOCKS
6122 struct opr_queue *store;
6128 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6129 if (rx_stats_active)
6130 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
6135 if (!opr_queue_IsEmpty(&call->tq)) { /* If we have anything to send */
6136 /* Send (or resend) any packets that need it, subject to
6137 * window restrictions and congestion burst control
6138 * restrictions. Ask for an ack on the last packet sent in
6139 * this burst. For now, we're relying upon the window being
6140 * considerably bigger than the largest number of packets that
6141 * are typically sent at once by one initial call to
6142 * rxi_Start. This is probably bogus (perhaps we should ask
6143 * for an ack when we're half way through the current
6144 * window?). Also, for non file transfer applications, this
6145 * may end up asking for an ack for every packet. Bogus. XXXX
6148 * But check whether we're here recursively, and let the other guy
6151 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6152 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6153 call->flags |= RX_CALL_TQ_BUSY;
6155 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6157 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6158 call->flags &= ~RX_CALL_NEED_START;
6159 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6161 maxXmitPackets = MIN(call->twind, call->cwind);
6162 for (opr_queue_Scan(&call->tq, cursor)) {
6164 = opr_queue_Entry(cursor, struct rx_packet, entry);
6166 if (p->flags & RX_PKTFLAG_ACKED) {
6167 /* Since we may block, don't trust this */
6168 if (rx_stats_active)
6169 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6170 continue; /* Ignore this packet if it has been acknowledged */
6173 /* Turn off all flags except these ones, which are the same
6174 * on each transmission */
6175 p->header.flags &= RX_PRESET_FLAGS;
6177 if (p->header.seq >=
6178 call->tfirst + MIN((int)call->twind,
6179 (int)(call->nSoftAcked +
6181 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6182 /* Note: if we're waiting for more window space, we can
6183 * still send retransmits; hence we don't return here, but
6184 * break out to schedule a retransmit event */
6185 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6186 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6191 /* Transmit the packet if it needs to be sent. */
6192 if (!(p->flags & RX_PKTFLAG_SENT)) {
6193 if (nXmitPackets == maxXmitPackets) {
6194 rxi_SendXmitList(call, call->xmitList,
6195 nXmitPackets, istack);
6198 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6199 *(call->callNumber), p));
6200 call->xmitList[nXmitPackets++] = p;
6202 } /* end of the queue_Scan */
6204 /* xmitList now hold pointers to all of the packets that are
6205 * ready to send. Now we loop to send the packets */
6206 if (nXmitPackets > 0) {
6207 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6211 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6213 /* We went into the error state while sending packets. Now is
6214 * the time to reset the call. This will also inform the using
6215 * process that the call is in an error state.
6217 if (rx_stats_active)
6218 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6219 call->flags &= ~RX_CALL_TQ_BUSY;
6220 rxi_WakeUpTransmitQueue(call);
6221 rxi_CallError(call, call->error);
6224 #ifdef RX_ENABLE_LOCKS
6225 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6227 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6228 /* Some packets have received acks. If they all have, we can clear
6229 * the transmit queue.
6232 for (opr_queue_ScanSafe(&call->tq, cursor, store)) {
6234 = opr_queue_Entry(cursor, struct rx_packet, entry);
6236 if (p->header.seq < call->tfirst
6237 && (p->flags & RX_PKTFLAG_ACKED)) {
6238 opr_queue_Remove(&p->entry);
6239 #ifdef RX_TRACK_PACKETS
6240 p->flags &= ~RX_PKTFLAG_TQ;
6242 #ifdef RXDEBUG_PACKET
6250 call->flags |= RX_CALL_TQ_CLEARME;
6252 #endif /* RX_ENABLE_LOCKS */
6253 if (call->flags & RX_CALL_TQ_CLEARME)
6254 rxi_ClearTransmitQueue(call, 1);
6255 } while (call->flags & RX_CALL_NEED_START);
6257 * TQ references no longer protected by this flag; they must remain
6258 * protected by the global lock.
6260 call->flags &= ~RX_CALL_TQ_BUSY;
6261 rxi_WakeUpTransmitQueue(call);
6263 call->flags |= RX_CALL_NEED_START;
6265 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6267 rxi_rto_cancel(call);
6271 /* Also adjusts the keep alive parameters for the call, to reflect
6272 * that we have just sent a packet (so keep alives aren't sent
6275 rxi_Send(struct rx_call *call, struct rx_packet *p,
6278 struct rx_connection *conn = call->conn;
6280 /* Stamp each packet with the user supplied status */
6281 p->header.userStatus = call->localStatus;
6283 /* Allow the security object controlling this call's security to
6284 * make any last-minute changes to the packet */
6285 RXS_SendPacket(conn->securityObject, call, p);
6287 /* Since we're about to send SOME sort of packet to the peer, it's
6288 * safe to nuke any scheduled end-of-packets ack */
6289 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
6291 /* Actually send the packet, filling in more connection-specific fields */
6292 MUTEX_EXIT(&call->lock);
6293 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6294 rxi_SendPacket(call, conn, p, istack);
6295 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6296 MUTEX_ENTER(&call->lock);
6298 /* Update last send time for this call (for keep-alive
6299 * processing), and for the connection (so that we can discover
6300 * idle connections) */
6301 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6302 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6303 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6305 conn->lastSendTime = call->lastSendTime = clock_Sec();
6306 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6307 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6308 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6309 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6310 RX_ACK_PING_RESPONSE)))
6311 call->lastSendData = call->lastSendTime;
6315 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6316 * that things are fine. Also called periodically to guarantee that nothing
6317 * falls through the cracks (e.g. (error + dally) connections have keepalive
6318 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6320 * haveCTLock Set if calling from rxi_ReapConnections
6322 #ifdef RX_ENABLE_LOCKS
6324 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6325 #else /* RX_ENABLE_LOCKS */
6327 rxi_CheckCall(struct rx_call *call)
6328 #endif /* RX_ENABLE_LOCKS */
6330 struct rx_connection *conn = call->conn;
6332 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6333 afs_uint32 fudgeFactor;
6336 int idle_timeout = 0;
6337 afs_int32 clock_diff = 0;
6339 if (rxi_CheckPeerDead(call)) {
6345 /* Large swings in the clock can have a significant impact on
6346 * the performance of RX call processing. Forward clock shifts
6347 * will result in premature event triggering or timeouts.
6348 * Backward shifts can result in calls not completing until
6349 * the clock catches up with the original start clock value.
6351 * If a backward clock shift of more than five minutes is noticed,
6352 * just fail the call.
6354 if (now < call->lastSendTime)
6355 clock_diff = call->lastSendTime - now;
6356 if (now < call->startWait)
6357 clock_diff = MAX(clock_diff, call->startWait - now);
6358 if (now < call->lastReceiveTime)
6359 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6360 if (clock_diff > 5 * 60)
6362 if (call->state == RX_STATE_ACTIVE)
6363 rxi_CallError(call, RX_CALL_TIMEOUT);
6367 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6368 if (call->flags & RX_CALL_TQ_BUSY) {
6369 /* Call is active and will be reset by rxi_Start if it's
6370 * in an error state.
6375 /* RTT + 8*MDEV, rounded up to the next second. */
6376 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6377 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6379 deadTime = conn->secondsUntilDead + fudgeFactor;
6380 /* These are computed to the second (+- 1 second). But that's
6381 * good enough for these values, which should be a significant
6382 * number of seconds. */
6383 if (now > (call->lastReceiveTime + deadTime)) {
6384 if (call->state == RX_STATE_ACTIVE) {
6385 #ifdef AFS_ADAPT_PMTU
6386 # if defined(KERNEL) && defined(AFS_SUN5_ENV)
6388 # if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6389 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6390 ip_stack_t *ipst = ns->netstack_ip;
6392 ire = ire_cache_lookup(conn->peer->host
6393 # if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6395 # if defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID)
6397 # if defined(GLOBAL_NETSTACKID)
6404 if (ire && ire->ire_max_frag > 0)
6405 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6407 # if defined(GLOBAL_NETSTACKID)
6411 #endif /* AFS_ADAPT_PMTU */
6412 cerror = RX_CALL_DEAD;
6415 #ifdef RX_ENABLE_LOCKS
6416 /* Cancel pending events */
6417 rxevent_Cancel(&call->delayedAckEvent, call,
6418 RX_CALL_REFCOUNT_DELAY);
6419 rxi_rto_cancel(call);
6420 rxevent_Cancel(&call->keepAliveEvent, call,
6421 RX_CALL_REFCOUNT_ALIVE);
6422 rxevent_Cancel(&call->growMTUEvent, call,
6423 RX_CALL_REFCOUNT_MTU);
6424 MUTEX_ENTER(&rx_refcnt_mutex);
6425 /* if rxi_FreeCall returns 1 it has freed the call */
6426 if (call->refCount == 0 &&
6427 rxi_FreeCall(call, haveCTLock))
6429 MUTEX_EXIT(&rx_refcnt_mutex);
6432 MUTEX_EXIT(&rx_refcnt_mutex);
6434 #else /* RX_ENABLE_LOCKS */
6435 rxi_FreeCall(call, 0);
6437 #endif /* RX_ENABLE_LOCKS */
6439 /* Non-active calls are destroyed if they are not responding
6440 * to pings; active calls are simply flagged in error, so the
6441 * attached process can die reasonably gracefully. */
6444 if (conn->idleDeadDetection) {
6445 if (conn->idleDeadTime) {
6446 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6450 /* see if we have a non-activity timeout */
6451 if (call->startWait && ((call->startWait + idleDeadTime) < now) &&
6452 (call->flags & RX_CALL_READER_WAIT)) {
6453 if (call->state == RX_STATE_ACTIVE) {
6454 cerror = RX_CALL_TIMEOUT;
6459 if (call->lastSendData && ((call->lastSendData + idleDeadTime) < now)) {
6460 if (call->state == RX_STATE_ACTIVE) {
6461 cerror = conn->service ? conn->service->idleDeadErr : RX_CALL_IDLE;
6469 if (conn->hardDeadTime) {
6470 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6473 /* see if we have a hard timeout */
6475 && (now > (hardDeadTime + call->startTime.sec))) {
6476 if (call->state == RX_STATE_ACTIVE)
6477 rxi_CallError(call, RX_CALL_TIMEOUT);
6482 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6483 call->lastReceiveTime) {
6484 int oldMTU = conn->peer->ifMTU;
6486 /* if we thought we could send more, perhaps things got worse */
6487 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6488 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6489 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6490 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6492 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6494 /* minimum capped in SetPeerMtu */
6495 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6498 conn->lastPacketSize = 0;
6500 /* needed so ResetCall doesn't clobber us. */
6501 call->MTU = conn->peer->ifMTU;
6503 /* if we never succeeded, let the error pass out as-is */
6504 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6505 cerror = conn->msgsizeRetryErr;
6508 rxi_CallError(call, cerror);
6513 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6514 void *dummy, int dummy2)
6516 struct rx_connection *conn = arg1;
6517 struct rx_header theader;
6518 char tbuffer[1 + sizeof(struct rx_header)];
6519 struct sockaddr_in taddr;
6522 struct iovec tmpiov[2];
6525 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6528 tp = &tbuffer[sizeof(struct rx_header)];
6529 taddr.sin_family = AF_INET;
6530 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6531 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6532 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6533 taddr.sin_len = sizeof(struct sockaddr_in);
6535 memset(&theader, 0, sizeof(theader));
6536 theader.epoch = htonl(999);
6538 theader.callNumber = 0;
6541 theader.type = RX_PACKET_TYPE_VERSION;
6542 theader.flags = RX_LAST_PACKET;
6543 theader.serviceId = 0;
6545 memcpy(tbuffer, &theader, sizeof(theader));
6546 memcpy(tp, &a, sizeof(a));
6547 tmpiov[0].iov_base = tbuffer;
6548 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6550 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6552 MUTEX_ENTER(&conn->conn_data_lock);
6553 MUTEX_ENTER(&rx_refcnt_mutex);
6554 /* Only reschedule ourselves if the connection would not be destroyed */
6555 if (conn->refCount <= 1) {
6556 rxevent_Put(conn->natKeepAliveEvent);
6557 conn->natKeepAliveEvent = NULL;
6558 MUTEX_EXIT(&rx_refcnt_mutex);
6559 MUTEX_EXIT(&conn->conn_data_lock);
6560 rx_DestroyConnection(conn); /* drop the reference for this */
6562 conn->refCount--; /* drop the reference for this */
6563 MUTEX_EXIT(&rx_refcnt_mutex);
6564 rxevent_Put(conn->natKeepAliveEvent);
6565 conn->natKeepAliveEvent = NULL;
6566 rxi_ScheduleNatKeepAliveEvent(conn);
6567 MUTEX_EXIT(&conn->conn_data_lock);
6572 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6574 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6575 struct clock when, now;
6576 clock_GetTime(&now);
6578 when.sec += conn->secondsUntilNatPing;
6579 MUTEX_ENTER(&rx_refcnt_mutex);
6580 conn->refCount++; /* hold a reference for this */
6581 MUTEX_EXIT(&rx_refcnt_mutex);
6582 conn->natKeepAliveEvent =
6583 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6588 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6590 MUTEX_ENTER(&conn->conn_data_lock);
6591 conn->secondsUntilNatPing = seconds;
6593 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6594 rxi_ScheduleNatKeepAliveEvent(conn);
6596 conn->flags |= RX_CONN_NAT_PING;
6598 MUTEX_EXIT(&conn->conn_data_lock);
6601 /* When a call is in progress, this routine is called occasionally to
6602 * make sure that some traffic has arrived (or been sent to) the peer.
6603 * If nothing has arrived in a reasonable amount of time, the call is
6604 * declared dead; if nothing has been sent for a while, we send a
6605 * keep-alive packet (if we're actually trying to keep the call alive)
6608 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6611 struct rx_call *call = arg1;
6612 struct rx_connection *conn;
6615 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6616 MUTEX_ENTER(&call->lock);
6618 if (event == call->keepAliveEvent) {
6619 rxevent_Put(call->keepAliveEvent);
6620 call->keepAliveEvent = NULL;
6625 #ifdef RX_ENABLE_LOCKS
6626 if (rxi_CheckCall(call, 0)) {
6627 MUTEX_EXIT(&call->lock);
6630 #else /* RX_ENABLE_LOCKS */
6631 if (rxi_CheckCall(call))
6633 #endif /* RX_ENABLE_LOCKS */
6635 /* Don't try to keep alive dallying calls */
6636 if (call->state == RX_STATE_DALLY) {
6637 MUTEX_EXIT(&call->lock);
6642 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6643 /* Don't try to send keepalives if there is unacknowledged data */
6644 /* the rexmit code should be good enough, this little hack
6645 * doesn't quite work XXX */
6646 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6648 rxi_ScheduleKeepAliveEvent(call);
6649 MUTEX_EXIT(&call->lock);
6652 /* Does what's on the nameplate. */
6654 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6656 struct rx_call *call = arg1;
6657 struct rx_connection *conn;
6659 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6660 MUTEX_ENTER(&call->lock);
6662 if (event == call->growMTUEvent) {
6663 rxevent_Put(call->growMTUEvent);
6664 call->growMTUEvent = NULL;
6667 #ifdef RX_ENABLE_LOCKS
6668 if (rxi_CheckCall(call, 0)) {
6669 MUTEX_EXIT(&call->lock);
6672 #else /* RX_ENABLE_LOCKS */
6673 if (rxi_CheckCall(call))
6675 #endif /* RX_ENABLE_LOCKS */
6677 /* Don't bother with dallying calls */
6678 if (call->state == RX_STATE_DALLY) {
6679 MUTEX_EXIT(&call->lock);
6686 * keep being scheduled, just don't do anything if we're at peak,
6687 * or we're not set up to be properly handled (idle timeout required)
6689 if ((conn->peer->maxPacketSize != 0) &&
6690 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6691 conn->idleDeadDetection)
6692 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6693 rxi_ScheduleGrowMTUEvent(call, 0);
6694 MUTEX_EXIT(&call->lock);
6698 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6700 if (!call->keepAliveEvent) {
6701 struct clock when, now;
6702 clock_GetTime(&now);
6704 when.sec += call->conn->secondsUntilPing;
6705 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6706 call->keepAliveEvent =
6707 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6712 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6714 if (!call->growMTUEvent) {
6715 struct clock when, now;
6717 clock_GetTime(&now);
6720 if (call->conn->secondsUntilPing)
6721 secs = (6*call->conn->secondsUntilPing)-1;
6723 if (call->conn->secondsUntilDead)
6724 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6728 CALL_HOLD(call, RX_CALL_REFCOUNT_MTU);
6729 call->growMTUEvent =
6730 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6734 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6736 rxi_KeepAliveOn(struct rx_call *call)
6738 /* Pretend last packet received was received now--i.e. if another
6739 * packet isn't received within the keep alive time, then the call
6740 * will die; Initialize last send time to the current time--even
6741 * if a packet hasn't been sent yet. This will guarantee that a
6742 * keep-alive is sent within the ping time */
6743 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6744 rxi_ScheduleKeepAliveEvent(call);
6748 * Solely in order that callers not need to include rx_call.h
6751 rx_KeepAliveOff(struct rx_call *call)
6753 rxi_KeepAliveOff(call);
6756 rx_KeepAliveOn(struct rx_call *call)
6758 rxi_KeepAliveOn(call);
6762 rxi_GrowMTUOn(struct rx_call *call)
6764 struct rx_connection *conn = call->conn;
6765 MUTEX_ENTER(&conn->conn_data_lock);
6766 conn->lastPingSizeSer = conn->lastPingSize = 0;
6767 MUTEX_EXIT(&conn->conn_data_lock);
6768 rxi_ScheduleGrowMTUEvent(call, 1);
6771 /* This routine is called to send connection abort messages
6772 * that have been delayed to throttle looping clients. */
6774 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6777 struct rx_connection *conn = arg1;
6780 struct rx_packet *packet;
6782 MUTEX_ENTER(&conn->conn_data_lock);
6783 rxevent_Put(conn->delayedAbortEvent);
6784 conn->delayedAbortEvent = NULL;
6785 error = htonl(conn->error);
6787 MUTEX_EXIT(&conn->conn_data_lock);
6788 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6791 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6792 RX_PACKET_TYPE_ABORT, (char *)&error,
6794 rxi_FreePacket(packet);
6798 /* This routine is called to send call abort messages
6799 * that have been delayed to throttle looping clients. */
6801 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6804 struct rx_call *call = arg1;
6807 struct rx_packet *packet;
6809 MUTEX_ENTER(&call->lock);
6810 rxevent_Put(call->delayedAbortEvent);
6811 call->delayedAbortEvent = NULL;
6812 error = htonl(call->error);
6814 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6817 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6818 (char *)&error, sizeof(error), 0);
6819 rxi_FreePacket(packet);
6821 MUTEX_EXIT(&call->lock);
6822 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6825 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6826 * seconds) to ask the client to authenticate itself. The routine
6827 * issues a challenge to the client, which is obtained from the
6828 * security object associated with the connection */
6830 rxi_ChallengeEvent(struct rxevent *event,
6831 void *arg0, void *arg1, int tries)
6833 struct rx_connection *conn = arg0;
6836 rxevent_Put(conn->challengeEvent);
6837 conn->challengeEvent = NULL;
6840 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6841 struct rx_packet *packet;
6842 struct clock when, now;
6845 /* We've failed to authenticate for too long.
6846 * Reset any calls waiting for authentication;
6847 * they are all in RX_STATE_PRECALL.
6851 MUTEX_ENTER(&conn->conn_call_lock);
6852 for (i = 0; i < RX_MAXCALLS; i++) {
6853 struct rx_call *call = conn->call[i];
6855 MUTEX_ENTER(&call->lock);
6856 if (call->state == RX_STATE_PRECALL) {
6857 rxi_CallError(call, RX_CALL_DEAD);
6858 rxi_SendCallAbort(call, NULL, 0, 0);
6860 MUTEX_EXIT(&call->lock);
6863 MUTEX_EXIT(&conn->conn_call_lock);
6867 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6869 /* If there's no packet available, do this later. */
6870 RXS_GetChallenge(conn->securityObject, conn, packet);
6871 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6872 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6873 rxi_FreePacket(packet);
6875 clock_GetTime(&now);
6877 when.sec += RX_CHALLENGE_TIMEOUT;
6878 conn->challengeEvent =
6879 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6884 /* Call this routine to start requesting the client to authenticate
6885 * itself. This will continue until authentication is established,
6886 * the call times out, or an invalid response is returned. The
6887 * security object associated with the connection is asked to create
6888 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6889 * defined earlier. */
6891 rxi_ChallengeOn(struct rx_connection *conn)
6893 if (!conn->challengeEvent) {
6894 RXS_CreateChallenge(conn->securityObject, conn);
6895 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6900 /* rxi_ComputeRoundTripTime is called with peer locked. */
6901 /* peer may be null */
6903 rxi_ComputeRoundTripTime(struct rx_packet *p,
6904 struct rx_ackPacket *ack,
6905 struct rx_call *call,
6906 struct rx_peer *peer,
6909 struct clock thisRtt, *sentp;
6913 /* If the ACK is delayed, then do nothing */
6914 if (ack->reason == RX_ACK_DELAY)
6917 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6918 * their RTT multiple times, so only include the RTT of the last packet
6920 if (p->flags & RX_JUMBO_PACKET)
6923 /* Use the serial number to determine which transmission the ACK is for,
6924 * and set the sent time to match this. If we have no serial number, then
6925 * only use the ACK for RTT calculations if the packet has not been
6929 serial = ntohl(ack->serial);
6931 if (serial == p->header.serial) {
6932 sentp = &p->timeSent;
6933 } else if (serial == p->firstSerial) {
6934 sentp = &p->firstSent;
6935 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6936 sentp = &p->firstSent;
6940 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6941 sentp = &p->firstSent;
6948 if (clock_Lt(&thisRtt, sentp))
6949 return; /* somebody set the clock back, don't count this time. */
6951 clock_Sub(&thisRtt, sentp);
6952 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6953 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6955 if (clock_IsZero(&thisRtt)) {
6957 * The actual round trip time is shorter than the
6958 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6959 * Since we can't tell which at the moment we will assume 1ms.
6961 thisRtt.usec = 1000;
6964 if (rx_stats_active) {
6965 MUTEX_ENTER(&rx_stats_mutex);
6966 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6967 rx_stats.minRtt = thisRtt;
6968 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6969 if (thisRtt.sec > 60) {
6970 MUTEX_EXIT(&rx_stats_mutex);
6971 return; /* somebody set the clock ahead */
6973 rx_stats.maxRtt = thisRtt;
6975 clock_Add(&rx_stats.totalRtt, &thisRtt);
6976 rx_atomic_inc(&rx_stats.nRttSamples);
6977 MUTEX_EXIT(&rx_stats_mutex);
6980 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6982 /* Apply VanJacobson round-trip estimations */
6987 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6988 * srtt is stored as fixed point with 3 bits after the binary
6989 * point (i.e., scaled by 8). The following magic is
6990 * equivalent to the smoothing algorithm in rfc793 with an
6991 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6992 * srtt'*8 = rtt + srtt*7
6993 * srtt'*8 = srtt*8 + rtt - srtt
6994 * srtt' = srtt + rtt/8 - srtt/8
6995 * srtt' = srtt + (rtt - srtt)/8
6998 delta = _8THMSEC(&thisRtt) - call->rtt;
6999 call->rtt += (delta >> 3);
7002 * We accumulate a smoothed rtt variance (actually, a smoothed
7003 * mean difference), then set the retransmit timer to smoothed
7004 * rtt + 4 times the smoothed variance (was 2x in van's original
7005 * paper, but 4x works better for me, and apparently for him as
7007 * rttvar is stored as
7008 * fixed point with 2 bits after the binary point (scaled by
7009 * 4). The following is equivalent to rfc793 smoothing with
7010 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
7011 * rttvar'*4 = rttvar*3 + |delta|
7012 * rttvar'*4 = rttvar*4 + |delta| - rttvar
7013 * rttvar' = rttvar + |delta|/4 - rttvar/4
7014 * rttvar' = rttvar + (|delta| - rttvar)/4
7015 * This replaces rfc793's wired-in beta.
7016 * dev*4 = dev*4 + (|actual - expected| - dev)
7022 delta -= (call->rtt_dev << 1);
7023 call->rtt_dev += (delta >> 3);
7025 /* I don't have a stored RTT so I start with this value. Since I'm
7026 * probably just starting a call, and will be pushing more data down
7027 * this, I expect congestion to increase rapidly. So I fudge a
7028 * little, and I set deviance to half the rtt. In practice,
7029 * deviance tends to approach something a little less than
7030 * half the smoothed rtt. */
7031 call->rtt = _8THMSEC(&thisRtt) + 8;
7032 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
7034 /* the smoothed RTT time is RTT + 4*MDEV
7036 * We allow a user specified minimum to be set for this, to allow clamping
7037 * at a minimum value in the same way as TCP. In addition, we have to allow
7038 * for the possibility that this packet is answered by a delayed ACK, so we
7039 * add on a fixed 200ms to account for that timer expiring.
7042 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
7043 rx_minPeerTimeout) + 200;
7044 clock_Zero(&call->rto);
7045 clock_Addmsec(&call->rto, rtt_timeout);
7047 /* Update the peer, so any new calls start with our values */
7048 peer->rtt_dev = call->rtt_dev;
7049 peer->rtt = call->rtt;
7051 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
7052 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
7056 /* Find all server connections that have not been active for a long time, and
7059 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
7062 struct clock now, when;
7063 clock_GetTime(&now);
7065 /* Find server connection structures that haven't been used for
7066 * greater than rx_idleConnectionTime */
7068 struct rx_connection **conn_ptr, **conn_end;
7069 int i, havecalls = 0;
7070 MUTEX_ENTER(&rx_connHashTable_lock);
7071 for (conn_ptr = &rx_connHashTable[0], conn_end =
7072 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
7074 struct rx_connection *conn, *next;
7075 struct rx_call *call;
7079 for (conn = *conn_ptr; conn; conn = next) {
7080 /* XXX -- Shouldn't the connection be locked? */
7083 for (i = 0; i < RX_MAXCALLS; i++) {
7084 call = conn->call[i];
7088 code = MUTEX_TRYENTER(&call->lock);
7091 #ifdef RX_ENABLE_LOCKS
7092 result = rxi_CheckCall(call, 1);
7093 #else /* RX_ENABLE_LOCKS */
7094 result = rxi_CheckCall(call);
7095 #endif /* RX_ENABLE_LOCKS */
7096 MUTEX_EXIT(&call->lock);
7098 /* If CheckCall freed the call, it might
7099 * have destroyed the connection as well,
7100 * which screws up the linked lists.
7106 if (conn->type == RX_SERVER_CONNECTION) {
7107 /* This only actually destroys the connection if
7108 * there are no outstanding calls */
7109 MUTEX_ENTER(&conn->conn_data_lock);
7110 MUTEX_ENTER(&rx_refcnt_mutex);
7111 if (!havecalls && !conn->refCount
7112 && ((conn->lastSendTime + rx_idleConnectionTime) <
7114 conn->refCount++; /* it will be decr in rx_DestroyConn */
7115 MUTEX_EXIT(&rx_refcnt_mutex);
7116 MUTEX_EXIT(&conn->conn_data_lock);
7117 #ifdef RX_ENABLE_LOCKS
7118 rxi_DestroyConnectionNoLock(conn);
7119 #else /* RX_ENABLE_LOCKS */
7120 rxi_DestroyConnection(conn);
7121 #endif /* RX_ENABLE_LOCKS */
7123 #ifdef RX_ENABLE_LOCKS
7125 MUTEX_EXIT(&rx_refcnt_mutex);
7126 MUTEX_EXIT(&conn->conn_data_lock);
7128 #endif /* RX_ENABLE_LOCKS */
7132 #ifdef RX_ENABLE_LOCKS
7133 while (rx_connCleanup_list) {
7134 struct rx_connection *conn;
7135 conn = rx_connCleanup_list;
7136 rx_connCleanup_list = rx_connCleanup_list->next;
7137 MUTEX_EXIT(&rx_connHashTable_lock);
7138 rxi_CleanupConnection(conn);
7139 MUTEX_ENTER(&rx_connHashTable_lock);
7141 MUTEX_EXIT(&rx_connHashTable_lock);
7142 #endif /* RX_ENABLE_LOCKS */
7145 /* Find any peer structures that haven't been used (haven't had an
7146 * associated connection) for greater than rx_idlePeerTime */
7148 struct rx_peer **peer_ptr, **peer_end;
7152 * Why do we need to hold the rx_peerHashTable_lock across
7153 * the incrementing of peer_ptr since the rx_peerHashTable
7154 * array is not changing? We don't.
7156 * By dropping the lock periodically we can permit other
7157 * activities to be performed while a rxi_ReapConnections
7158 * call is in progress. The goal of reap connections
7159 * is to clean up quickly without causing large amounts
7160 * of contention. Therefore, it is important that global
7161 * mutexes not be held for extended periods of time.
7163 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7164 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7166 struct rx_peer *peer, *next, *prev;
7168 MUTEX_ENTER(&rx_peerHashTable_lock);
7169 for (prev = peer = *peer_ptr; peer; peer = next) {
7171 code = MUTEX_TRYENTER(&peer->peer_lock);
7172 if ((code) && (peer->refCount == 0)
7173 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7174 struct opr_queue *cursor, *store;
7178 * now know that this peer object is one to be
7179 * removed from the hash table. Once it is removed
7180 * it can't be referenced by other threads.
7181 * Lets remove it first and decrement the struct
7182 * nPeerStructs count.
7184 if (peer == *peer_ptr) {
7190 if (rx_stats_active)
7191 rx_atomic_dec(&rx_stats.nPeerStructs);
7194 * Now if we hold references on 'prev' and 'next'
7195 * we can safely drop the rx_peerHashTable_lock
7196 * while we destroy this 'peer' object.
7202 MUTEX_EXIT(&rx_peerHashTable_lock);
7204 MUTEX_EXIT(&peer->peer_lock);
7205 MUTEX_DESTROY(&peer->peer_lock);
7207 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7208 unsigned int num_funcs;
7209 struct rx_interface_stat *rpc_stat
7210 = opr_queue_Entry(cursor, struct rx_interface_stat,
7215 opr_queue_Remove(&rpc_stat->entry);
7216 opr_queue_Remove(&rpc_stat->entryPeers);
7218 num_funcs = rpc_stat->stats[0].func_total;
7220 sizeof(rx_interface_stat_t) +
7221 rpc_stat->stats[0].func_total *
7222 sizeof(rx_function_entry_v1_t);
7224 rxi_Free(rpc_stat, space);
7226 MUTEX_ENTER(&rx_rpc_stats);
7227 rxi_rpc_peer_stat_cnt -= num_funcs;
7228 MUTEX_EXIT(&rx_rpc_stats);
7233 * Regain the rx_peerHashTable_lock and
7234 * decrement the reference count on 'prev'
7237 MUTEX_ENTER(&rx_peerHashTable_lock);
7244 MUTEX_EXIT(&peer->peer_lock);
7249 MUTEX_EXIT(&rx_peerHashTable_lock);
7253 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7254 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7255 * GC, just below. Really, we shouldn't have to keep moving packets from
7256 * one place to another, but instead ought to always know if we can
7257 * afford to hold onto a packet in its particular use. */
7258 MUTEX_ENTER(&rx_freePktQ_lock);
7259 if (rx_waitingForPackets) {
7260 rx_waitingForPackets = 0;
7261 #ifdef RX_ENABLE_LOCKS
7262 CV_BROADCAST(&rx_waitingForPackets_cv);
7264 osi_rxWakeup(&rx_waitingForPackets);
7267 MUTEX_EXIT(&rx_freePktQ_lock);
7270 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7271 rxevent_Put(rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0));
7275 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7276 * rx.h is sort of strange this is better. This is called with a security
7277 * object before it is discarded. Each connection using a security object has
7278 * its own refcount to the object so it won't actually be freed until the last
7279 * connection is destroyed.
7281 * This is the only rxs module call. A hold could also be written but no one
7285 rxs_Release(struct rx_securityClass *aobj)
7287 return RXS_Close(aobj);
7295 #define TRACE_OPTION_RX_DEBUG 16
7303 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7304 0, KEY_QUERY_VALUE, &parmKey);
7305 if (code != ERROR_SUCCESS)
7308 dummyLen = sizeof(TraceOption);
7309 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7310 (BYTE *) &TraceOption, &dummyLen);
7311 if (code == ERROR_SUCCESS) {
7312 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7314 RegCloseKey (parmKey);
7315 #endif /* AFS_NT40_ENV */
7320 rx_DebugOnOff(int on)
7324 rxdebug_active = on;
7330 rx_StatsOnOff(int on)
7332 rx_stats_active = on;
7336 /* Don't call this debugging routine directly; use dpf */
7338 rxi_DebugPrint(char *format, ...)
7347 va_start(ap, format);
7349 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7352 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7354 OutputDebugString(msg);
7360 va_start(ap, format);
7362 clock_GetTime(&now);
7363 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7364 (unsigned int)now.usec);
7365 vfprintf(rx_Log, format, ap);
7373 * This function is used to process the rx_stats structure that is local
7374 * to a process as well as an rx_stats structure received from a remote
7375 * process (via rxdebug). Therefore, it needs to do minimal version
7379 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7380 afs_int32 freePackets, char version)
7384 if (size != sizeof(struct rx_statistics)) {
7386 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7387 size, sizeof(struct rx_statistics));
7390 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7393 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7394 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7395 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7396 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7397 s->specialPktAllocFailures);
7399 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7400 s->receivePktAllocFailures, s->sendPktAllocFailures,
7401 s->specialPktAllocFailures);
7405 " greedy %u, " "bogusReads %u (last from host %x), "
7406 "noPackets %u, " "noBuffers %u, " "selects %u, "
7407 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7408 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7409 s->selects, s->sendSelects);
7411 fprintf(file, " packets read: ");
7412 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7413 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7415 fprintf(file, "\n");
7418 " other read counters: data %u, " "ack %u, " "dup %u "
7419 "spurious %u " "dally %u\n", s->dataPacketsRead,
7420 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7421 s->ignorePacketDally);
7423 fprintf(file, " packets sent: ");
7424 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7425 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7427 fprintf(file, "\n");
7430 " other send counters: ack %u, " "data %u (not resends), "
7431 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7432 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7433 s->dataPacketsPushed, s->ignoreAckedPacket);
7436 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7437 s->netSendFailures, (int)s->fatalErrors);
7439 if (s->nRttSamples) {
7440 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7441 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7443 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7444 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7448 " %d server connections, " "%d client connections, "
7449 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7450 s->nServerConns, s->nClientConns, s->nPeerStructs,
7451 s->nCallStructs, s->nFreeCallStructs);
7453 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7454 fprintf(file, " %d clock updates\n", clock_nUpdates);
7458 /* for backward compatibility */
7460 rx_PrintStats(FILE * file)
7462 MUTEX_ENTER(&rx_stats_mutex);
7463 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7464 sizeof(rx_stats), rx_nFreePackets,
7466 MUTEX_EXIT(&rx_stats_mutex);
7470 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7472 fprintf(file, "Peer %x.%d.\n",
7473 ntohl(peer->host), (int)ntohs(peer->port));
7476 " Rtt %d, " "total sent %d, " "resent %d\n",
7477 peer->rtt, peer->nSent, peer->reSends);
7479 fprintf(file, " Packet size %d\n", peer->ifMTU);
7483 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7485 * This mutex protects the following static variables:
7489 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7490 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7492 #define LOCK_RX_DEBUG
7493 #define UNLOCK_RX_DEBUG
7494 #endif /* AFS_PTHREAD_ENV */
7496 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7498 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7499 u_char type, void *inputData, size_t inputLength,
7500 void *outputData, size_t outputLength)
7502 static afs_int32 counter = 100;
7503 time_t waitTime, waitCount;
7504 struct rx_header theader;
7507 struct timeval tv_now, tv_wake, tv_delta;
7508 struct sockaddr_in taddr, faddr;
7522 tp = &tbuffer[sizeof(struct rx_header)];
7523 taddr.sin_family = AF_INET;
7524 taddr.sin_port = remotePort;
7525 taddr.sin_addr.s_addr = remoteAddr;
7526 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7527 taddr.sin_len = sizeof(struct sockaddr_in);
7530 memset(&theader, 0, sizeof(theader));
7531 theader.epoch = htonl(999);
7533 theader.callNumber = htonl(counter);
7536 theader.type = type;
7537 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7538 theader.serviceId = 0;
7540 memcpy(tbuffer, &theader, sizeof(theader));
7541 memcpy(tp, inputData, inputLength);
7543 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7544 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7546 /* see if there's a packet available */
7547 gettimeofday(&tv_wake, NULL);
7548 tv_wake.tv_sec += waitTime;
7551 FD_SET(socket, &imask);
7552 tv_delta.tv_sec = tv_wake.tv_sec;
7553 tv_delta.tv_usec = tv_wake.tv_usec;
7554 gettimeofday(&tv_now, NULL);
7556 if (tv_delta.tv_usec < tv_now.tv_usec) {
7558 tv_delta.tv_usec += 1000000;
7561 tv_delta.tv_usec -= tv_now.tv_usec;
7563 if (tv_delta.tv_sec < tv_now.tv_sec) {
7567 tv_delta.tv_sec -= tv_now.tv_sec;
7570 code = select(0, &imask, 0, 0, &tv_delta);
7571 #else /* AFS_NT40_ENV */
7572 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7573 #endif /* AFS_NT40_ENV */
7574 if (code == 1 && FD_ISSET(socket, &imask)) {
7575 /* now receive a packet */
7576 faddrLen = sizeof(struct sockaddr_in);
7578 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7579 (struct sockaddr *)&faddr, &faddrLen);
7582 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7583 if (counter == ntohl(theader.callNumber))
7591 /* see if we've timed out */
7599 code -= sizeof(struct rx_header);
7600 if (code > outputLength)
7601 code = outputLength;
7602 memcpy(outputData, tp, code);
7605 #endif /* RXDEBUG */
7608 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7609 afs_uint16 remotePort, struct rx_debugStats * stat,
7610 afs_uint32 * supportedValues)
7612 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7614 struct rx_debugIn in;
7616 *supportedValues = 0;
7617 in.type = htonl(RX_DEBUGI_GETSTATS);
7620 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7621 &in, sizeof(in), stat, sizeof(*stat));
7624 * If the call was successful, fixup the version and indicate
7625 * what contents of the stat structure are valid.
7626 * Also do net to host conversion of fields here.
7630 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7631 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7633 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7634 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7636 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7637 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7639 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7640 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7642 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7643 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7645 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7646 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7648 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7649 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7651 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7652 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7654 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7655 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7657 stat->nFreePackets = ntohl(stat->nFreePackets);
7658 stat->packetReclaims = ntohl(stat->packetReclaims);
7659 stat->callsExecuted = ntohl(stat->callsExecuted);
7660 stat->nWaiting = ntohl(stat->nWaiting);
7661 stat->idleThreads = ntohl(stat->idleThreads);
7662 stat->nWaited = ntohl(stat->nWaited);
7663 stat->nPackets = ntohl(stat->nPackets);
7672 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7673 afs_uint16 remotePort, struct rx_statistics * stat,
7674 afs_uint32 * supportedValues)
7676 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7678 struct rx_debugIn in;
7679 afs_int32 *lp = (afs_int32 *) stat;
7683 * supportedValues is currently unused, but added to allow future
7684 * versioning of this function.
7687 *supportedValues = 0;
7688 in.type = htonl(RX_DEBUGI_RXSTATS);
7690 memset(stat, 0, sizeof(*stat));
7692 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7693 &in, sizeof(in), stat, sizeof(*stat));
7698 * Do net to host conversion here
7701 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7712 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7713 afs_uint16 remotePort, size_t version_length,
7716 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7718 return MakeDebugCall(socket, remoteAddr, remotePort,
7719 RX_PACKET_TYPE_VERSION, a, 1, version,
7727 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7728 afs_uint16 remotePort, afs_int32 * nextConnection,
7729 int allConnections, afs_uint32 debugSupportedValues,
7730 struct rx_debugConn * conn,
7731 afs_uint32 * supportedValues)
7733 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7735 struct rx_debugIn in;
7739 * supportedValues is currently unused, but added to allow future
7740 * versioning of this function.
7743 *supportedValues = 0;
7744 if (allConnections) {
7745 in.type = htonl(RX_DEBUGI_GETALLCONN);
7747 in.type = htonl(RX_DEBUGI_GETCONN);
7749 in.index = htonl(*nextConnection);
7750 memset(conn, 0, sizeof(*conn));
7752 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7753 &in, sizeof(in), conn, sizeof(*conn));
7756 *nextConnection += 1;
7759 * Convert old connection format to new structure.
7762 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7763 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7764 #define MOVEvL(a) (conn->a = vL->a)
7766 /* any old or unrecognized version... */
7767 for (i = 0; i < RX_MAXCALLS; i++) {
7768 MOVEvL(callState[i]);
7769 MOVEvL(callMode[i]);
7770 MOVEvL(callFlags[i]);
7771 MOVEvL(callOther[i]);
7773 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7774 MOVEvL(secStats.type);
7775 MOVEvL(secStats.level);
7776 MOVEvL(secStats.flags);
7777 MOVEvL(secStats.expires);
7778 MOVEvL(secStats.packetsReceived);
7779 MOVEvL(secStats.packetsSent);
7780 MOVEvL(secStats.bytesReceived);
7781 MOVEvL(secStats.bytesSent);
7786 * Do net to host conversion here
7788 * I don't convert host or port since we are most likely
7789 * going to want these in NBO.
7791 conn->cid = ntohl(conn->cid);
7792 conn->serial = ntohl(conn->serial);
7793 for (i = 0; i < RX_MAXCALLS; i++) {
7794 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7796 conn->error = ntohl(conn->error);
7797 conn->secStats.flags = ntohl(conn->secStats.flags);
7798 conn->secStats.expires = ntohl(conn->secStats.expires);
7799 conn->secStats.packetsReceived =
7800 ntohl(conn->secStats.packetsReceived);
7801 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7802 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7803 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7804 conn->epoch = ntohl(conn->epoch);
7805 conn->natMTU = ntohl(conn->natMTU);
7814 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7815 afs_uint16 remotePort, afs_int32 * nextPeer,
7816 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7817 afs_uint32 * supportedValues)
7819 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7821 struct rx_debugIn in;
7824 * supportedValues is currently unused, but added to allow future
7825 * versioning of this function.
7828 *supportedValues = 0;
7829 in.type = htonl(RX_DEBUGI_GETPEER);
7830 in.index = htonl(*nextPeer);
7831 memset(peer, 0, sizeof(*peer));
7833 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7834 &in, sizeof(in), peer, sizeof(*peer));
7840 * Do net to host conversion here
7842 * I don't convert host or port since we are most likely
7843 * going to want these in NBO.
7845 peer->ifMTU = ntohs(peer->ifMTU);
7846 peer->idleWhen = ntohl(peer->idleWhen);
7847 peer->refCount = ntohs(peer->refCount);
7848 peer->rtt = ntohl(peer->rtt);
7849 peer->rtt_dev = ntohl(peer->rtt_dev);
7850 peer->timeout.sec = 0;
7851 peer->timeout.usec = 0;
7852 peer->nSent = ntohl(peer->nSent);
7853 peer->reSends = ntohl(peer->reSends);
7854 peer->natMTU = ntohs(peer->natMTU);
7855 peer->maxMTU = ntohs(peer->maxMTU);
7856 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7857 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7858 peer->MTU = ntohs(peer->MTU);
7859 peer->cwind = ntohs(peer->cwind);
7860 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7861 peer->congestSeq = ntohs(peer->congestSeq);
7862 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7863 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7864 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7865 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7874 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7875 struct rx_debugPeer * peerStats)
7878 afs_int32 error = 1; /* default to "did not succeed" */
7879 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7881 MUTEX_ENTER(&rx_peerHashTable_lock);
7882 for(tp = rx_peerHashTable[hashValue];
7883 tp != NULL; tp = tp->next) {
7884 if (tp->host == peerHost)
7890 MUTEX_EXIT(&rx_peerHashTable_lock);
7894 MUTEX_ENTER(&tp->peer_lock);
7895 peerStats->host = tp->host;
7896 peerStats->port = tp->port;
7897 peerStats->ifMTU = tp->ifMTU;
7898 peerStats->idleWhen = tp->idleWhen;
7899 peerStats->refCount = tp->refCount;
7900 peerStats->burstSize = 0;
7901 peerStats->burst = 0;
7902 peerStats->burstWait.sec = 0;
7903 peerStats->burstWait.usec = 0;
7904 peerStats->rtt = tp->rtt;
7905 peerStats->rtt_dev = tp->rtt_dev;
7906 peerStats->timeout.sec = 0;
7907 peerStats->timeout.usec = 0;
7908 peerStats->nSent = tp->nSent;
7909 peerStats->reSends = tp->reSends;
7910 peerStats->natMTU = tp->natMTU;
7911 peerStats->maxMTU = tp->maxMTU;
7912 peerStats->maxDgramPackets = tp->maxDgramPackets;
7913 peerStats->ifDgramPackets = tp->ifDgramPackets;
7914 peerStats->MTU = tp->MTU;
7915 peerStats->cwind = tp->cwind;
7916 peerStats->nDgramPackets = tp->nDgramPackets;
7917 peerStats->congestSeq = tp->congestSeq;
7918 peerStats->bytesSent.high = tp->bytesSent >> 32;
7919 peerStats->bytesSent.low = tp->bytesSent & MAX_AFS_UINT32;
7920 peerStats->bytesReceived.high = tp->bytesReceived >> 32;
7921 peerStats->bytesReceived.low
7922 = tp->bytesReceived & MAX_AFS_UINT32;
7923 MUTEX_EXIT(&tp->peer_lock);
7925 MUTEX_ENTER(&rx_peerHashTable_lock);
7928 MUTEX_EXIT(&rx_peerHashTable_lock);
7936 struct rx_serverQueueEntry *np;
7939 struct rx_call *call;
7940 struct rx_serverQueueEntry *sq;
7944 if (rxinit_status == 1) {
7946 return; /* Already shutdown. */
7950 #ifndef AFS_PTHREAD_ENV
7951 FD_ZERO(&rx_selectMask);
7952 #endif /* AFS_PTHREAD_ENV */
7953 rxi_dataQuota = RX_MAX_QUOTA;
7954 #ifndef AFS_PTHREAD_ENV
7956 #endif /* AFS_PTHREAD_ENV */
7959 #ifndef AFS_PTHREAD_ENV
7960 #ifndef AFS_USE_GETTIMEOFDAY
7962 #endif /* AFS_USE_GETTIMEOFDAY */
7963 #endif /* AFS_PTHREAD_ENV */
7965 while (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
7966 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
7967 opr_queue_Remove(&call->entry);
7968 rxi_Free(call, sizeof(struct rx_call));
7971 while (!opr_queue_IsEmpty(&rx_idleServerQueue)) {
7972 sq = opr_queue_First(&rx_idleServerQueue, struct rx_serverQueueEntry,
7974 opr_queue_Remove(&sq->entry);
7979 struct rx_peer **peer_ptr, **peer_end;
7980 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7981 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7983 struct rx_peer *peer, *next;
7985 MUTEX_ENTER(&rx_peerHashTable_lock);
7986 for (peer = *peer_ptr; peer; peer = next) {
7987 struct opr_queue *cursor, *store;
7990 MUTEX_ENTER(&rx_rpc_stats);
7991 MUTEX_ENTER(&peer->peer_lock);
7992 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7993 unsigned int num_funcs;
7994 struct rx_interface_stat *rpc_stat
7995 = opr_queue_Entry(cursor, struct rx_interface_stat,
7999 opr_queue_Remove(&rpc_stat->entry);
8000 opr_queue_Remove(&rpc_stat->entryPeers);
8001 num_funcs = rpc_stat->stats[0].func_total;
8003 sizeof(rx_interface_stat_t) +
8004 rpc_stat->stats[0].func_total *
8005 sizeof(rx_function_entry_v1_t);
8007 rxi_Free(rpc_stat, space);
8009 /* rx_rpc_stats must be held */
8010 rxi_rpc_peer_stat_cnt -= num_funcs;
8012 MUTEX_EXIT(&peer->peer_lock);
8013 MUTEX_EXIT(&rx_rpc_stats);
8017 if (rx_stats_active)
8018 rx_atomic_dec(&rx_stats.nPeerStructs);
8020 MUTEX_EXIT(&rx_peerHashTable_lock);
8023 for (i = 0; i < RX_MAX_SERVICES; i++) {
8025 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
8027 for (i = 0; i < rx_hashTableSize; i++) {
8028 struct rx_connection *tc, *ntc;
8029 MUTEX_ENTER(&rx_connHashTable_lock);
8030 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
8032 for (j = 0; j < RX_MAXCALLS; j++) {
8034 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
8037 rxi_Free(tc, sizeof(*tc));
8039 MUTEX_EXIT(&rx_connHashTable_lock);
8042 MUTEX_ENTER(&freeSQEList_lock);
8044 while ((np = rx_FreeSQEList)) {
8045 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
8046 MUTEX_DESTROY(&np->lock);
8047 rxi_Free(np, sizeof(*np));
8050 MUTEX_EXIT(&freeSQEList_lock);
8051 MUTEX_DESTROY(&freeSQEList_lock);
8052 MUTEX_DESTROY(&rx_freeCallQueue_lock);
8053 MUTEX_DESTROY(&rx_connHashTable_lock);
8054 MUTEX_DESTROY(&rx_peerHashTable_lock);
8055 MUTEX_DESTROY(&rx_serverPool_lock);
8057 osi_Free(rx_connHashTable,
8058 rx_hashTableSize * sizeof(struct rx_connection *));
8059 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8061 UNPIN(rx_connHashTable,
8062 rx_hashTableSize * sizeof(struct rx_connection *));
8063 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8065 rxi_FreeAllPackets();
8067 MUTEX_ENTER(&rx_quota_mutex);
8068 rxi_dataQuota = RX_MAX_QUOTA;
8069 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
8070 MUTEX_EXIT(&rx_quota_mutex);
8075 #ifdef RX_ENABLE_LOCKS
8077 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
8079 if (!MUTEX_ISMINE(lockaddr))
8080 osi_Panic("Lock not held: %s", msg);
8082 #endif /* RX_ENABLE_LOCKS */
8087 * Routines to implement connection specific data.
8091 rx_KeyCreate(rx_destructor_t rtn)
8094 MUTEX_ENTER(&rxi_keyCreate_lock);
8095 key = rxi_keyCreate_counter++;
8096 rxi_keyCreate_destructor = (rx_destructor_t *)
8097 realloc((void *)rxi_keyCreate_destructor,
8098 (key + 1) * sizeof(rx_destructor_t));
8099 rxi_keyCreate_destructor[key] = rtn;
8100 MUTEX_EXIT(&rxi_keyCreate_lock);
8105 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
8108 MUTEX_ENTER(&conn->conn_data_lock);
8109 if (!conn->specific) {
8110 conn->specific = malloc((key + 1) * sizeof(void *));
8111 for (i = 0; i < key; i++)
8112 conn->specific[i] = NULL;
8113 conn->nSpecific = key + 1;
8114 conn->specific[key] = ptr;
8115 } else if (key >= conn->nSpecific) {
8116 conn->specific = (void **)
8117 realloc(conn->specific, (key + 1) * sizeof(void *));
8118 for (i = conn->nSpecific; i < key; i++)
8119 conn->specific[i] = NULL;
8120 conn->nSpecific = key + 1;
8121 conn->specific[key] = ptr;
8123 if (conn->specific[key] && rxi_keyCreate_destructor[key])
8124 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
8125 conn->specific[key] = ptr;
8127 MUTEX_EXIT(&conn->conn_data_lock);
8131 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
8134 MUTEX_ENTER(&svc->svc_data_lock);
8135 if (!svc->specific) {
8136 svc->specific = malloc((key + 1) * sizeof(void *));
8137 for (i = 0; i < key; i++)
8138 svc->specific[i] = NULL;
8139 svc->nSpecific = key + 1;
8140 svc->specific[key] = ptr;
8141 } else if (key >= svc->nSpecific) {
8142 svc->specific = (void **)
8143 realloc(svc->specific, (key + 1) * sizeof(void *));
8144 for (i = svc->nSpecific; i < key; i++)
8145 svc->specific[i] = NULL;
8146 svc->nSpecific = key + 1;
8147 svc->specific[key] = ptr;
8149 if (svc->specific[key] && rxi_keyCreate_destructor[key])
8150 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
8151 svc->specific[key] = ptr;
8153 MUTEX_EXIT(&svc->svc_data_lock);
8157 rx_GetSpecific(struct rx_connection *conn, int key)
8160 MUTEX_ENTER(&conn->conn_data_lock);
8161 if (key >= conn->nSpecific)
8164 ptr = conn->specific[key];
8165 MUTEX_EXIT(&conn->conn_data_lock);
8170 rx_GetServiceSpecific(struct rx_service *svc, int key)
8173 MUTEX_ENTER(&svc->svc_data_lock);
8174 if (key >= svc->nSpecific)
8177 ptr = svc->specific[key];
8178 MUTEX_EXIT(&svc->svc_data_lock);
8183 #endif /* !KERNEL */
8186 * processStats is a queue used to store the statistics for the local
8187 * process. Its contents are similar to the contents of the rpcStats
8188 * queue on a rx_peer structure, but the actual data stored within
8189 * this queue contains totals across the lifetime of the process (assuming
8190 * the stats have not been reset) - unlike the per peer structures
8191 * which can come and go based upon the peer lifetime.
8194 static struct opr_queue processStats = { &processStats, &processStats };
8197 * peerStats is a queue used to store the statistics for all peer structs.
8198 * Its contents are the union of all the peer rpcStats queues.
8201 static struct opr_queue peerStats = { &peerStats, &peerStats };
8204 * rxi_monitor_processStats is used to turn process wide stat collection
8208 static int rxi_monitor_processStats = 0;
8211 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8214 static int rxi_monitor_peerStats = 0;
8218 rxi_ClearRPCOpStat(rx_function_entry_v1_p rpc_stat)
8220 rpc_stat->invocations = 0;
8221 rpc_stat->bytes_sent = 0;
8222 rpc_stat->bytes_rcvd = 0;
8223 rpc_stat->queue_time_sum.sec = 0;
8224 rpc_stat->queue_time_sum.usec = 0;
8225 rpc_stat->queue_time_sum_sqr.sec = 0;
8226 rpc_stat->queue_time_sum_sqr.usec = 0;
8227 rpc_stat->queue_time_min.sec = 9999999;
8228 rpc_stat->queue_time_min.usec = 9999999;
8229 rpc_stat->queue_time_max.sec = 0;
8230 rpc_stat->queue_time_max.usec = 0;
8231 rpc_stat->execution_time_sum.sec = 0;
8232 rpc_stat->execution_time_sum.usec = 0;
8233 rpc_stat->execution_time_sum_sqr.sec = 0;
8234 rpc_stat->execution_time_sum_sqr.usec = 0;
8235 rpc_stat->execution_time_min.sec = 9999999;
8236 rpc_stat->execution_time_min.usec = 9999999;
8237 rpc_stat->execution_time_max.sec = 0;
8238 rpc_stat->execution_time_max.usec = 0;
8242 * Given all of the information for a particular rpc
8243 * call, find or create (if requested) the stat structure for the rpc.
8246 * the queue of stats that will be updated with the new value
8248 * @param rxInterface
8249 * a unique number that identifies the rpc interface
8252 * the total number of functions in this interface. this is only
8253 * required if create is true
8256 * if true, this invocation was made to a server
8259 * the ip address of the remote host. this is only required if create
8260 * and addToPeerList are true
8263 * the port of the remote host. this is only required if create
8264 * and addToPeerList are true
8266 * @param addToPeerList
8267 * if != 0, add newly created stat to the global peer list
8270 * if a new stats structure is allocated, the counter will
8271 * be updated with the new number of allocated stat structures.
8272 * only required if create is true
8275 * if no stats structure exists, allocate one
8279 static rx_interface_stat_p
8280 rxi_FindRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8281 afs_uint32 totalFunc, int isServer, afs_uint32 remoteHost,
8282 afs_uint32 remotePort, int addToPeerList,
8283 unsigned int *counter, int create)
8285 rx_interface_stat_p rpc_stat = NULL;
8286 struct opr_queue *cursor;
8289 * See if there's already a structure for this interface
8292 for (opr_queue_Scan(stats, cursor)) {
8293 rpc_stat = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8295 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8296 && (rpc_stat->stats[0].remote_is_server == isServer))
8300 /* if they didn't ask us to create, we're done */
8302 if (opr_queue_IsEnd(stats, cursor))
8308 /* can't proceed without these */
8309 if (!totalFunc || !counter)
8313 * Didn't find a match so allocate a new structure and add it to the
8317 if (opr_queue_IsEnd(stats, cursor) || (rpc_stat == NULL)
8318 || (rpc_stat->stats[0].interfaceId != rxInterface)
8319 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8324 sizeof(rx_interface_stat_t) +
8325 totalFunc * sizeof(rx_function_entry_v1_t);
8327 rpc_stat = rxi_Alloc(space);
8328 if (rpc_stat == NULL)
8331 *counter += totalFunc;
8332 for (i = 0; i < totalFunc; i++) {
8333 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8334 rpc_stat->stats[i].remote_peer = remoteHost;
8335 rpc_stat->stats[i].remote_port = remotePort;
8336 rpc_stat->stats[i].remote_is_server = isServer;
8337 rpc_stat->stats[i].interfaceId = rxInterface;
8338 rpc_stat->stats[i].func_total = totalFunc;
8339 rpc_stat->stats[i].func_index = i;
8341 opr_queue_Prepend(stats, &rpc_stat->entry);
8342 if (addToPeerList) {
8343 opr_queue_Prepend(&peerStats, &rpc_stat->entryPeers);
8350 rx_ClearProcessRPCStats(afs_int32 rxInterface)
8352 rx_interface_stat_p rpc_stat;
8355 if (rxInterface == -1)
8358 MUTEX_ENTER(&rx_rpc_stats);
8359 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8362 totalFunc = rpc_stat->stats[0].func_total;
8363 for (i = 0; i < totalFunc; i++)
8364 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8366 MUTEX_EXIT(&rx_rpc_stats);
8371 rx_ClearPeerRPCStats(afs_int32 rxInterface, afs_uint32 peerHost, afs_uint16 peerPort)
8373 rx_interface_stat_p rpc_stat;
8375 struct rx_peer * peer;
8377 if (rxInterface == -1)
8380 peer = rxi_FindPeer(peerHost, peerPort, 0, 0);
8384 MUTEX_ENTER(&rx_rpc_stats);
8385 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8388 totalFunc = rpc_stat->stats[0].func_total;
8389 for (i = 0; i < totalFunc; i++)
8390 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8392 MUTEX_EXIT(&rx_rpc_stats);
8397 rx_CopyProcessRPCStats(afs_uint64 op)
8399 rx_interface_stat_p rpc_stat;
8400 rx_function_entry_v1_p rpcop_stat =
8401 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8402 int currentFunc = (op & MAX_AFS_UINT32);
8403 afs_int32 rxInterface = (op >> 32);
8405 if (!rxi_monitor_processStats)
8408 if (rxInterface == -1)
8411 if (rpcop_stat == NULL)
8414 MUTEX_ENTER(&rx_rpc_stats);
8415 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8418 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8419 sizeof(rx_function_entry_v1_t));
8420 MUTEX_EXIT(&rx_rpc_stats);
8422 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8429 rx_CopyPeerRPCStats(afs_uint64 op, afs_uint32 peerHost, afs_uint16 peerPort)
8431 rx_interface_stat_p rpc_stat;
8432 rx_function_entry_v1_p rpcop_stat =
8433 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8434 int currentFunc = (op & MAX_AFS_UINT32);
8435 afs_int32 rxInterface = (op >> 32);
8436 struct rx_peer *peer;
8438 if (!rxi_monitor_peerStats)
8441 if (rxInterface == -1)
8444 if (rpcop_stat == NULL)
8447 peer = rxi_FindPeer(peerHost, peerPort, 0, 0);
8451 MUTEX_ENTER(&rx_rpc_stats);
8452 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8455 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8456 sizeof(rx_function_entry_v1_t));
8457 MUTEX_EXIT(&rx_rpc_stats);
8459 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8466 rx_ReleaseRPCStats(void *stats)
8469 rxi_Free(stats, sizeof(rx_function_entry_v1_t));
8473 * Given all of the information for a particular rpc
8474 * call, create (if needed) and update the stat totals for the rpc.
8477 * the queue of stats that will be updated with the new value
8479 * @param rxInterface
8480 * a unique number that identifies the rpc interface
8482 * @param currentFunc
8483 * the index of the function being invoked
8486 * the total number of functions in this interface
8489 * the amount of time this function waited for a thread
8492 * the amount of time this function invocation took to execute
8495 * the number bytes sent by this invocation
8498 * the number bytes received by this invocation
8501 * if true, this invocation was made to a server
8504 * the ip address of the remote host
8507 * the port of the remote host
8509 * @param addToPeerList
8510 * if != 0, add newly created stat to the global peer list
8513 * if a new stats structure is allocated, the counter will
8514 * be updated with the new number of allocated stat structures
8519 rxi_AddRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8520 afs_uint32 currentFunc, afs_uint32 totalFunc,
8521 struct clock *queueTime, struct clock *execTime,
8522 afs_uint64 bytesSent, afs_uint64 bytesRcvd, int isServer,
8523 afs_uint32 remoteHost, afs_uint32 remotePort,
8524 int addToPeerList, unsigned int *counter)
8527 rx_interface_stat_p rpc_stat;
8529 rpc_stat = rxi_FindRpcStat(stats, rxInterface, totalFunc, isServer,
8530 remoteHost, remotePort, addToPeerList, counter,
8538 * Increment the stats for this function
8541 rpc_stat->stats[currentFunc].invocations++;
8542 rpc_stat->stats[currentFunc].bytes_sent += bytesSent;
8543 rpc_stat->stats[currentFunc].bytes_rcvd += bytesRcvd;
8544 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8545 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8546 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8547 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8549 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8550 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8552 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8553 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8555 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8556 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8558 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8559 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8567 rxi_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8568 afs_uint32 currentFunc, afs_uint32 totalFunc,
8569 struct clock *queueTime, struct clock *execTime,
8570 afs_uint64 bytesSent, afs_uint64 bytesRcvd,
8574 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8577 MUTEX_ENTER(&rx_rpc_stats);
8579 if (rxi_monitor_peerStats) {
8580 MUTEX_ENTER(&peer->peer_lock);
8581 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8582 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8583 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8584 MUTEX_EXIT(&peer->peer_lock);
8587 if (rxi_monitor_processStats) {
8588 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8589 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8590 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8593 MUTEX_EXIT(&rx_rpc_stats);
8597 * Increment the times and count for a particular rpc function.
8599 * Traditionally this call was invoked from rxgen stubs. Modern stubs
8600 * call rx_RecordCallStatistics instead, so the public version of this
8601 * function is left purely for legacy callers.
8604 * The peer who invoked the rpc
8606 * @param rxInterface
8607 * A unique number that identifies the rpc interface
8609 * @param currentFunc
8610 * The index of the function being invoked
8613 * The total number of functions in this interface
8616 * The amount of time this function waited for a thread
8619 * The amount of time this function invocation took to execute
8622 * The number bytes sent by this invocation
8625 * The number bytes received by this invocation
8628 * If true, this invocation was made to a server
8632 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8633 afs_uint32 currentFunc, afs_uint32 totalFunc,
8634 struct clock *queueTime, struct clock *execTime,
8635 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8641 sent64 = ((afs_uint64)bytesSent->high << 32) + bytesSent->low;
8642 rcvd64 = ((afs_uint64)bytesRcvd->high << 32) + bytesRcvd->low;
8644 rxi_IncrementTimeAndCount(peer, rxInterface, currentFunc, totalFunc,
8645 queueTime, execTime, sent64, rcvd64,
8652 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8656 * IN callerVersion - the rpc stat version of the caller.
8658 * IN count - the number of entries to marshall.
8660 * IN stats - pointer to stats to be marshalled.
8662 * OUT ptr - Where to store the marshalled data.
8669 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8670 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8676 * We only support the first version
8678 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8679 *(ptr++) = stats->remote_peer;
8680 *(ptr++) = stats->remote_port;
8681 *(ptr++) = stats->remote_is_server;
8682 *(ptr++) = stats->interfaceId;
8683 *(ptr++) = stats->func_total;
8684 *(ptr++) = stats->func_index;
8685 *(ptr++) = stats->invocations >> 32;
8686 *(ptr++) = stats->invocations & MAX_AFS_UINT32;
8687 *(ptr++) = stats->bytes_sent >> 32;
8688 *(ptr++) = stats->bytes_sent & MAX_AFS_UINT32;
8689 *(ptr++) = stats->bytes_rcvd >> 32;
8690 *(ptr++) = stats->bytes_rcvd & MAX_AFS_UINT32;
8691 *(ptr++) = stats->queue_time_sum.sec;
8692 *(ptr++) = stats->queue_time_sum.usec;
8693 *(ptr++) = stats->queue_time_sum_sqr.sec;
8694 *(ptr++) = stats->queue_time_sum_sqr.usec;
8695 *(ptr++) = stats->queue_time_min.sec;
8696 *(ptr++) = stats->queue_time_min.usec;
8697 *(ptr++) = stats->queue_time_max.sec;
8698 *(ptr++) = stats->queue_time_max.usec;
8699 *(ptr++) = stats->execution_time_sum.sec;
8700 *(ptr++) = stats->execution_time_sum.usec;
8701 *(ptr++) = stats->execution_time_sum_sqr.sec;
8702 *(ptr++) = stats->execution_time_sum_sqr.usec;
8703 *(ptr++) = stats->execution_time_min.sec;
8704 *(ptr++) = stats->execution_time_min.usec;
8705 *(ptr++) = stats->execution_time_max.sec;
8706 *(ptr++) = stats->execution_time_max.usec;
8712 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8717 * IN callerVersion - the rpc stat version of the caller
8719 * OUT myVersion - the rpc stat version of this function
8721 * OUT clock_sec - local time seconds
8723 * OUT clock_usec - local time microseconds
8725 * OUT allocSize - the number of bytes allocated to contain stats
8727 * OUT statCount - the number stats retrieved from this process.
8729 * OUT stats - the actual stats retrieved from this process.
8733 * Returns void. If successful, stats will != NULL.
8737 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8738 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8739 size_t * allocSize, afs_uint32 * statCount,
8740 afs_uint32 ** stats)
8750 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8753 * Check to see if stats are enabled
8756 MUTEX_ENTER(&rx_rpc_stats);
8757 if (!rxi_monitor_processStats) {
8758 MUTEX_EXIT(&rx_rpc_stats);
8762 clock_GetTime(&now);
8763 *clock_sec = now.sec;
8764 *clock_usec = now.usec;
8767 * Allocate the space based upon the caller version
8769 * If the client is at an older version than we are,
8770 * we return the statistic data in the older data format, but
8771 * we still return our version number so the client knows we
8772 * are maintaining more data than it can retrieve.
8775 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8776 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8777 *statCount = rxi_rpc_process_stat_cnt;
8780 * This can't happen yet, but in the future version changes
8781 * can be handled by adding additional code here
8785 if (space > (size_t) 0) {
8787 ptr = *stats = rxi_Alloc(space);
8790 struct opr_queue *cursor;
8792 for (opr_queue_Scan(&processStats, cursor)) {
8793 struct rx_interface_stat *rpc_stat =
8794 opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8796 * Copy the data based upon the caller version
8798 rx_MarshallProcessRPCStats(callerVersion,
8799 rpc_stat->stats[0].func_total,
8800 rpc_stat->stats, &ptr);
8806 MUTEX_EXIT(&rx_rpc_stats);
8811 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8815 * IN callerVersion - the rpc stat version of the caller
8817 * OUT myVersion - the rpc stat version of this function
8819 * OUT clock_sec - local time seconds
8821 * OUT clock_usec - local time microseconds
8823 * OUT allocSize - the number of bytes allocated to contain stats
8825 * OUT statCount - the number of stats retrieved from the individual
8828 * OUT stats - the actual stats retrieved from the individual peer structures.
8832 * Returns void. If successful, stats will != NULL.
8836 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8837 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8838 size_t * allocSize, afs_uint32 * statCount,
8839 afs_uint32 ** stats)
8849 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8852 * Check to see if stats are enabled
8855 MUTEX_ENTER(&rx_rpc_stats);
8856 if (!rxi_monitor_peerStats) {
8857 MUTEX_EXIT(&rx_rpc_stats);
8861 clock_GetTime(&now);
8862 *clock_sec = now.sec;
8863 *clock_usec = now.usec;
8866 * Allocate the space based upon the caller version
8868 * If the client is at an older version than we are,
8869 * we return the statistic data in the older data format, but
8870 * we still return our version number so the client knows we
8871 * are maintaining more data than it can retrieve.
8874 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8875 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8876 *statCount = rxi_rpc_peer_stat_cnt;
8879 * This can't happen yet, but in the future version changes
8880 * can be handled by adding additional code here
8884 if (space > (size_t) 0) {
8886 ptr = *stats = rxi_Alloc(space);
8889 struct opr_queue *cursor;
8891 for (opr_queue_Scan(&peerStats, cursor)) {
8892 struct rx_interface_stat *rpc_stat
8893 = opr_queue_Entry(cursor, struct rx_interface_stat,
8897 * Copy the data based upon the caller version
8899 rx_MarshallProcessRPCStats(callerVersion,
8900 rpc_stat->stats[0].func_total,
8901 rpc_stat->stats, &ptr);
8907 MUTEX_EXIT(&rx_rpc_stats);
8912 * rx_FreeRPCStats - free memory allocated by
8913 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8917 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8918 * rx_RetrievePeerRPCStats
8920 * IN allocSize - the number of bytes in stats.
8928 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8930 rxi_Free(stats, allocSize);
8934 * rx_queryProcessRPCStats - see if process rpc stat collection is
8935 * currently enabled.
8941 * Returns 0 if stats are not enabled != 0 otherwise
8945 rx_queryProcessRPCStats(void)
8948 MUTEX_ENTER(&rx_rpc_stats);
8949 rc = rxi_monitor_processStats;
8950 MUTEX_EXIT(&rx_rpc_stats);
8955 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8961 * Returns 0 if stats are not enabled != 0 otherwise
8965 rx_queryPeerRPCStats(void)
8968 MUTEX_ENTER(&rx_rpc_stats);
8969 rc = rxi_monitor_peerStats;
8970 MUTEX_EXIT(&rx_rpc_stats);
8975 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8985 rx_enableProcessRPCStats(void)
8987 MUTEX_ENTER(&rx_rpc_stats);
8988 rx_enable_stats = 1;
8989 rxi_monitor_processStats = 1;
8990 MUTEX_EXIT(&rx_rpc_stats);
8994 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
9004 rx_enablePeerRPCStats(void)
9006 MUTEX_ENTER(&rx_rpc_stats);
9007 rx_enable_stats = 1;
9008 rxi_monitor_peerStats = 1;
9009 MUTEX_EXIT(&rx_rpc_stats);
9013 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
9023 rx_disableProcessRPCStats(void)
9025 struct opr_queue *cursor, *store;
9028 MUTEX_ENTER(&rx_rpc_stats);
9031 * Turn off process statistics and if peer stats is also off, turn
9035 rxi_monitor_processStats = 0;
9036 if (rxi_monitor_peerStats == 0) {
9037 rx_enable_stats = 0;
9040 for (opr_queue_ScanSafe(&processStats, cursor, store)) {
9041 unsigned int num_funcs = 0;
9042 struct rx_interface_stat *rpc_stat
9043 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
9045 opr_queue_Remove(&rpc_stat->entry);
9047 num_funcs = rpc_stat->stats[0].func_total;
9049 sizeof(rx_interface_stat_t) +
9050 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
9052 rxi_Free(rpc_stat, space);
9053 rxi_rpc_process_stat_cnt -= num_funcs;
9055 MUTEX_EXIT(&rx_rpc_stats);
9059 * rx_disablePeerRPCStats - stop rpc stat collection for peers
9069 rx_disablePeerRPCStats(void)
9071 struct rx_peer **peer_ptr, **peer_end;
9075 * Turn off peer statistics and if process stats is also off, turn
9079 rxi_monitor_peerStats = 0;
9080 if (rxi_monitor_processStats == 0) {
9081 rx_enable_stats = 0;
9084 for (peer_ptr = &rx_peerHashTable[0], peer_end =
9085 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
9087 struct rx_peer *peer, *next, *prev;
9089 MUTEX_ENTER(&rx_peerHashTable_lock);
9090 MUTEX_ENTER(&rx_rpc_stats);
9091 for (prev = peer = *peer_ptr; peer; peer = next) {
9093 code = MUTEX_TRYENTER(&peer->peer_lock);
9096 struct opr_queue *cursor, *store;
9098 if (prev == *peer_ptr) {
9109 MUTEX_EXIT(&rx_peerHashTable_lock);
9111 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
9112 unsigned int num_funcs = 0;
9113 struct rx_interface_stat *rpc_stat
9114 = opr_queue_Entry(cursor, struct rx_interface_stat,
9117 opr_queue_Remove(&rpc_stat->entry);
9118 opr_queue_Remove(&rpc_stat->entryPeers);
9119 num_funcs = rpc_stat->stats[0].func_total;
9121 sizeof(rx_interface_stat_t) +
9122 rpc_stat->stats[0].func_total *
9123 sizeof(rx_function_entry_v1_t);
9125 rxi_Free(rpc_stat, space);
9126 rxi_rpc_peer_stat_cnt -= num_funcs;
9128 MUTEX_EXIT(&peer->peer_lock);
9130 MUTEX_ENTER(&rx_peerHashTable_lock);
9140 MUTEX_EXIT(&rx_rpc_stats);
9141 MUTEX_EXIT(&rx_peerHashTable_lock);
9146 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
9151 * IN clearFlag - flag indicating which stats to clear
9159 rx_clearProcessRPCStats(afs_uint32 clearFlag)
9161 struct opr_queue *cursor;
9163 MUTEX_ENTER(&rx_rpc_stats);
9165 for (opr_queue_Scan(&processStats, cursor)) {
9166 unsigned int num_funcs = 0, i;
9167 struct rx_interface_stat *rpc_stat
9168 = opr_queue_Entry(rpc_stat, struct rx_interface_stat, entry);
9170 num_funcs = rpc_stat->stats[0].func_total;
9171 for (i = 0; i < num_funcs; i++) {
9172 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9173 rpc_stat->stats[i].invocations = 0;
9175 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9176 rpc_stat->stats[i].bytes_sent = 0;
9178 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9179 rpc_stat->stats[i].bytes_rcvd = 0;
9181 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9182 rpc_stat->stats[i].queue_time_sum.sec = 0;
9183 rpc_stat->stats[i].queue_time_sum.usec = 0;
9185 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9186 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9187 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9189 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9190 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9191 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9193 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9194 rpc_stat->stats[i].queue_time_max.sec = 0;
9195 rpc_stat->stats[i].queue_time_max.usec = 0;
9197 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9198 rpc_stat->stats[i].execution_time_sum.sec = 0;
9199 rpc_stat->stats[i].execution_time_sum.usec = 0;
9201 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9202 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9203 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9205 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9206 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9207 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9209 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9210 rpc_stat->stats[i].execution_time_max.sec = 0;
9211 rpc_stat->stats[i].execution_time_max.usec = 0;
9216 MUTEX_EXIT(&rx_rpc_stats);
9220 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
9225 * IN clearFlag - flag indicating which stats to clear
9233 rx_clearPeerRPCStats(afs_uint32 clearFlag)
9235 struct opr_queue *cursor;
9237 MUTEX_ENTER(&rx_rpc_stats);
9239 for (opr_queue_Scan(&peerStats, cursor)) {
9240 unsigned int num_funcs, i;
9241 struct rx_interface_stat *rpc_stat
9242 = opr_queue_Entry(cursor, struct rx_interface_stat, entryPeers);
9244 num_funcs = rpc_stat->stats[0].func_total;
9245 for (i = 0; i < num_funcs; i++) {
9246 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9247 rpc_stat->stats[i].invocations = 0;
9249 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9250 rpc_stat->stats[i].bytes_sent = 0;
9252 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9253 rpc_stat->stats[i].bytes_rcvd = 0;
9255 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9256 rpc_stat->stats[i].queue_time_sum.sec = 0;
9257 rpc_stat->stats[i].queue_time_sum.usec = 0;
9259 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9260 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9261 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9263 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9264 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9265 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9267 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9268 rpc_stat->stats[i].queue_time_max.sec = 0;
9269 rpc_stat->stats[i].queue_time_max.usec = 0;
9271 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9272 rpc_stat->stats[i].execution_time_sum.sec = 0;
9273 rpc_stat->stats[i].execution_time_sum.usec = 0;
9275 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9276 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9277 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9279 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9280 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9281 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9283 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9284 rpc_stat->stats[i].execution_time_max.sec = 0;
9285 rpc_stat->stats[i].execution_time_max.usec = 0;
9290 MUTEX_EXIT(&rx_rpc_stats);
9294 * rxi_rxstat_userok points to a routine that returns 1 if the caller
9295 * is authorized to enable/disable/clear RX statistics.
9297 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
9300 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
9302 rxi_rxstat_userok = proc;
9306 rx_RxStatUserOk(struct rx_call *call)
9308 if (!rxi_rxstat_userok)
9310 return rxi_rxstat_userok(call);
9315 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
9316 * function in the MSVC runtime DLL (msvcrt.dll).
9318 * Note: the system serializes calls to this function.
9321 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
9322 DWORD reason, /* reason function is being called */
9323 LPVOID reserved) /* reserved for future use */
9326 case DLL_PROCESS_ATTACH:
9327 /* library is being attached to a process */
9331 case DLL_PROCESS_DETACH:
9338 #endif /* AFS_NT40_ENV */
9341 int rx_DumpCalls(FILE *outputFile, char *cookie)
9343 #ifdef RXDEBUG_PACKET
9344 #ifdef KDUMP_RX_LOCK
9345 struct rx_call_rx_lock *c;
9352 #define RXDPRINTF sprintf
9353 #define RXDPRINTOUT output
9355 #define RXDPRINTF fprintf
9356 #define RXDPRINTOUT outputFile
9359 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9361 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9364 for (c = rx_allCallsp; c; c = c->allNextp) {
9365 u_short rqc, tqc, iovqc;
9367 MUTEX_ENTER(&c->lock);
9368 rqc = opr_queue_Count(&c->rq);
9369 tqc = opr_queue_Count(&c->tq);
9370 iovqc = opr_queue_Count(&c->iovq);
9372 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, "
9373 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9374 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9375 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9376 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
9377 #ifdef RX_ENABLE_LOCKS
9380 #ifdef RX_REFCOUNT_CHECK
9381 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9382 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9385 cookie, c, c->call_id, (afs_uint32)c->state, (afs_uint32)c->mode, c->conn, c->conn?c->conn->epoch:0, c->conn?c->conn->cid:0,
9386 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9387 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9388 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9389 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9390 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
9391 #ifdef RX_ENABLE_LOCKS
9392 , (afs_uint32)c->refCount
9394 #ifdef RX_REFCOUNT_CHECK
9395 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9398 MUTEX_EXIT(&c->lock);
9401 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9404 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9406 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9408 #endif /* RXDEBUG_PACKET */