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>
78 #include "rx_atomic.h"
79 #include "rx_globals.h"
81 #include "rx_internal.h"
88 #include "rx_packet.h"
90 #include <afs/rxgen_consts.h>
93 #ifdef AFS_PTHREAD_ENV
95 int (*registerProgram) (pid_t, char *) = 0;
96 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
99 int (*registerProgram) (PROCESS, char *) = 0;
100 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
104 /* Local static routines */
105 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
106 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
107 struct rx_call *, struct rx_peer *,
109 static void rxi_Resend(struct rxevent *event, void *arg0, void *arg1,
111 static void rxi_SendDelayedAck(struct rxevent *event, void *call,
112 void *dummy, int dummy2);
113 static void rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1,
114 void *dummy, int dummy2);
115 static void rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1,
116 void *unused, int unused2);
117 static void rxi_ReapConnections(struct rxevent *unused, void *unused1,
118 void *unused2, int unused3);
119 static struct rx_packet *rxi_SendCallAbort(struct rx_call *call,
120 struct rx_packet *packet,
121 int istack, int force);
122 static void rxi_AckAll(struct rx_call *call);
123 static struct rx_connection
124 *rxi_FindConnection(osi_socket socket, afs_uint32 host, u_short port,
125 u_short serviceId, afs_uint32 cid,
126 afs_uint32 epoch, int type, u_int securityIndex);
127 static struct rx_packet
128 *rxi_ReceiveDataPacket(struct rx_call *call, struct rx_packet *np,
129 int istack, osi_socket socket,
130 afs_uint32 host, u_short port, int *tnop,
131 struct rx_call **newcallp);
132 static struct rx_packet
133 *rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
135 static struct rx_packet
136 *rxi_ReceiveResponsePacket(struct rx_connection *conn,
137 struct rx_packet *np, int istack);
138 static struct rx_packet
139 *rxi_ReceiveChallengePacket(struct rx_connection *conn,
140 struct rx_packet *np, int istack);
141 static void rxi_AttachServerProc(struct rx_call *call, osi_socket socket,
142 int *tnop, struct rx_call **newcallp);
143 static void rxi_ClearTransmitQueue(struct rx_call *call, int force);
144 static void rxi_ClearReceiveQueue(struct rx_call *call);
145 static void rxi_ResetCall(struct rx_call *call, int newcall);
146 static void rxi_ScheduleKeepAliveEvent(struct rx_call *call);
147 static void rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn);
148 static void rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs);
149 static void rxi_KeepAliveOn(struct rx_call *call);
150 static void rxi_GrowMTUOn(struct rx_call *call);
151 static void rxi_ChallengeOn(struct rx_connection *conn);
153 #ifdef RX_ENABLE_LOCKS
154 static int rxi_CheckCall(struct rx_call *call, int haveCTLock);
155 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
157 static int rxi_CheckCall(struct rx_call *call);
160 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
162 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
163 rx_atomic_t rxi_start_in_error;
165 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
167 /* Constant delay time before sending an acknowledge of the last packet
168 * received. This is to avoid sending an extra acknowledge when the
169 * client is about to make another call, anyway, or the server is
172 * The lastAckDelay may not exceeed 400ms without causing peers to
173 * unecessarily timeout.
175 struct clock rx_lastAckDelay = {0, 400000};
177 /* Constant delay time before sending a soft ack when none was requested.
178 * This is to make sure we send soft acks before the sender times out,
179 * Normally we wait and send a hard ack when the receiver consumes the packet
181 * This value has been 100ms in all shipping versions of OpenAFS. Changing it
182 * will require changes to the peer's RTT calculations.
184 struct clock rx_softAckDelay = {0, 100000};
187 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
188 * currently allocated within rx. This number is used to allocate the
189 * memory required to return the statistics when queried.
190 * Protected by the rx_rpc_stats mutex.
193 static unsigned int rxi_rpc_peer_stat_cnt;
196 * rxi_rpc_process_stat_cnt counts the total number of local process stat
197 * structures currently allocated within rx. The number is used to allocate
198 * the memory required to return the statistics when queried.
199 * Protected by the rx_rpc_stats mutex.
202 static unsigned int rxi_rpc_process_stat_cnt;
205 * rxi_busyChannelError is a boolean. It indicates whether or not RX_CALL_BUSY
206 * errors should be reported to the application when a call channel appears busy
207 * (inferred from the receipt of RX_PACKET_TYPE_BUSY packets on the channel),
208 * and there are other call channels in the connection that are not busy.
209 * If 0, we do not return errors upon receiving busy packets; we just keep
210 * trying on the same call channel until we hit a timeout.
212 static afs_int32 rxi_busyChannelError = 0;
214 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
215 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
217 #if !defined(offsetof)
218 #include <stddef.h> /* for definition of offsetof() */
221 #ifdef RX_ENABLE_LOCKS
222 afs_kmutex_t rx_atomic_mutex;
225 /* Forward prototypes */
226 static struct rx_call * rxi_NewCall(struct rx_connection *, int);
229 putConnection (struct rx_connection *conn) {
230 MUTEX_ENTER(&rx_refcnt_mutex);
232 MUTEX_EXIT(&rx_refcnt_mutex);
235 #ifdef AFS_PTHREAD_ENV
238 * Use procedural initialization of mutexes/condition variables
242 extern afs_kmutex_t rx_quota_mutex;
243 extern afs_kmutex_t rx_pthread_mutex;
244 extern afs_kmutex_t rx_packets_mutex;
245 extern afs_kmutex_t rx_refcnt_mutex;
246 extern afs_kmutex_t des_init_mutex;
247 extern afs_kmutex_t des_random_mutex;
248 extern afs_kmutex_t rx_clock_mutex;
249 extern afs_kmutex_t rxi_connCacheMutex;
250 extern afs_kmutex_t event_handler_mutex;
251 extern afs_kmutex_t listener_mutex;
252 extern afs_kmutex_t rx_if_init_mutex;
253 extern afs_kmutex_t rx_if_mutex;
255 extern afs_kcondvar_t rx_event_handler_cond;
256 extern afs_kcondvar_t rx_listener_cond;
258 static afs_kmutex_t epoch_mutex;
259 static afs_kmutex_t rx_init_mutex;
260 static afs_kmutex_t rx_debug_mutex;
261 static afs_kmutex_t rx_rpc_stats;
264 rxi_InitPthread(void)
266 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
267 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
268 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
269 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
270 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
271 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
272 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
273 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
274 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
275 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
276 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
277 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
278 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
279 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
280 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
282 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
283 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
285 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
286 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
288 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
289 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
290 #ifdef RX_ENABLE_LOCKS
293 #endif /* RX_LOCKS_DB */
294 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
295 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
297 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
299 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
301 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
303 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
304 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
305 #endif /* RX_ENABLE_LOCKS */
308 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
309 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
311 * The rx_stats_mutex mutex protects the following global variables:
312 * rxi_lowConnRefCount
313 * rxi_lowPeerRefCount
322 * The rx_quota_mutex mutex protects the following global variables:
330 * The rx_freePktQ_lock protects the following global variables:
335 * The rx_packets_mutex mutex protects the following global variables:
343 * The rx_pthread_mutex mutex protects the following global variables:
344 * rxi_fcfs_thread_num
347 #define INIT_PTHREAD_LOCKS
351 /* Variables for handling the minProcs implementation. availProcs gives the
352 * number of threads available in the pool at this moment (not counting dudes
353 * executing right now). totalMin gives the total number of procs required
354 * for handling all minProcs requests. minDeficit is a dynamic variable
355 * tracking the # of procs required to satisfy all of the remaining minProcs
357 * For fine grain locking to work, the quota check and the reservation of
358 * a server thread has to come while rxi_availProcs and rxi_minDeficit
359 * are locked. To this end, the code has been modified under #ifdef
360 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
361 * same time. A new function, ReturnToServerPool() returns the allocation.
363 * A call can be on several queue's (but only one at a time). When
364 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
365 * that no one else is touching the queue. To this end, we store the address
366 * of the queue lock in the call structure (under the call lock) when we
367 * put the call on a queue, and we clear the call_queue_lock when the
368 * call is removed from a queue (once the call lock has been obtained).
369 * This allows rxi_ResetCall to safely synchronize with others wishing
370 * to manipulate the queue.
373 #if defined(RX_ENABLE_LOCKS)
374 static afs_kmutex_t rx_rpc_stats;
377 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
378 ** pretty good that the next packet coming in is from the same connection
379 ** as the last packet, since we're send multiple packets in a transmit window.
381 struct rx_connection *rxLastConn = 0;
383 #ifdef RX_ENABLE_LOCKS
384 /* The locking hierarchy for rx fine grain locking is composed of these
387 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
388 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
389 * call->lock - locks call data fields.
390 * These are independent of each other:
391 * rx_freeCallQueue_lock
396 * serverQueueEntry->lock
397 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
399 * peer->lock - locks peer data fields.
400 * conn_data_lock - that more than one thread is not updating a conn data
401 * field at the same time.
412 * Do we need a lock to protect the peer field in the conn structure?
413 * conn->peer was previously a constant for all intents and so has no
414 * lock protecting this field. The multihomed client delta introduced
415 * a RX code change : change the peer field in the connection structure
416 * to that remote interface from which the last packet for this
417 * connection was sent out. This may become an issue if further changes
420 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
421 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
423 /* rxdb_fileID is used to identify the lock location, along with line#. */
424 static int rxdb_fileID = RXDB_FILE_RX;
425 #endif /* RX_LOCKS_DB */
426 #else /* RX_ENABLE_LOCKS */
427 #define SET_CALL_QUEUE_LOCK(C, L)
428 #define CLEAR_CALL_QUEUE_LOCK(C)
429 #endif /* RX_ENABLE_LOCKS */
430 struct rx_serverQueueEntry *rx_waitForPacket = 0;
431 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
433 /* ------------Exported Interfaces------------- */
435 /* This function allows rxkad to set the epoch to a suitably random number
436 * which rx_NewConnection will use in the future. The principle purpose is to
437 * get rxnull connections to use the same epoch as the rxkad connections do, at
438 * least once the first rxkad connection is established. This is important now
439 * that the host/port addresses aren't used in FindConnection: the uniqueness
440 * of epoch/cid matters and the start time won't do. */
442 #ifdef AFS_PTHREAD_ENV
444 * This mutex protects the following global variables:
448 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
449 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
453 #endif /* AFS_PTHREAD_ENV */
456 rx_SetEpoch(afs_uint32 epoch)
463 /* Initialize rx. A port number may be mentioned, in which case this
464 * becomes the default port number for any service installed later.
465 * If 0 is provided for the port number, a random port will be chosen
466 * by the kernel. Whether this will ever overlap anything in
467 * /etc/services is anybody's guess... Returns 0 on success, -1 on
472 int rxinit_status = 1;
473 #ifdef AFS_PTHREAD_ENV
475 * This mutex protects the following global variables:
479 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
480 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
483 #define UNLOCK_RX_INIT
487 rx_InitHost(u_int host, u_int port)
494 char *htable, *ptable;
501 if (rxinit_status == 0) {
502 tmp_status = rxinit_status;
504 return tmp_status; /* Already started; return previous error code. */
510 if (afs_winsockInit() < 0)
516 * Initialize anything necessary to provide a non-premptive threading
519 rxi_InitializeThreadSupport();
522 /* Allocate and initialize a socket for client and perhaps server
525 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
526 if (rx_socket == OSI_NULLSOCKET) {
530 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
533 #endif /* RX_LOCKS_DB */
534 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
535 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
536 MUTEX_INIT(&rx_atomic_mutex, "rx_atomic_mutex", MUTEX_DEFAULT, 0);
537 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
538 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
539 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
540 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
541 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
542 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
543 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
545 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
547 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
549 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
551 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
552 #if defined(AFS_HPUX110_ENV)
554 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
555 #endif /* AFS_HPUX110_ENV */
556 #endif /* RX_ENABLE_LOCKS && KERNEL */
559 rx_connDeadTime = 12;
560 rx_tranquil = 0; /* reset flag */
561 rxi_ResetStatistics();
562 htable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
563 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
564 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
565 ptable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
566 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
567 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
569 /* Malloc up a bunch of packets & buffers */
571 queue_Init(&rx_freePacketQueue);
572 rxi_NeedMorePackets = FALSE;
573 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
575 /* enforce a minimum number of allocated packets */
576 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
577 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
579 /* allocate the initial free packet pool */
580 #ifdef RX_ENABLE_TSFPQ
581 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
582 #else /* RX_ENABLE_TSFPQ */
583 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
584 #endif /* RX_ENABLE_TSFPQ */
591 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
592 tv.tv_sec = clock_now.sec;
593 tv.tv_usec = clock_now.usec;
594 srand((unsigned int)tv.tv_usec);
601 #if defined(KERNEL) && !defined(UKERNEL)
602 /* Really, this should never happen in a real kernel */
605 struct sockaddr_in addr;
607 int addrlen = sizeof(addr);
609 socklen_t addrlen = sizeof(addr);
611 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
613 osi_Free(htable, rx_hashTableSize * sizeof(struct rx_connection *));
616 rx_port = addr.sin_port;
619 rx_stats.minRtt.sec = 9999999;
621 rx_SetEpoch(tv.tv_sec | 0x80000000);
623 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
624 * will provide a randomer value. */
626 MUTEX_ENTER(&rx_quota_mutex);
627 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
628 MUTEX_EXIT(&rx_quota_mutex);
629 /* *Slightly* random start time for the cid. This is just to help
630 * out with the hashing function at the peer */
631 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
632 rx_connHashTable = (struct rx_connection **)htable;
633 rx_peerHashTable = (struct rx_peer **)ptable;
635 rx_hardAckDelay.sec = 0;
636 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
638 rxevent_Init(20, rxi_ReScheduleEvents);
640 /* Initialize various global queues */
641 queue_Init(&rx_idleServerQueue);
642 queue_Init(&rx_incomingCallQueue);
643 queue_Init(&rx_freeCallQueue);
645 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
646 /* Initialize our list of usable IP addresses. */
650 #if defined(RXK_LISTENER_ENV) || !defined(KERNEL)
651 /* Start listener process (exact function is dependent on the
652 * implementation environment--kernel or user space) */
657 tmp_status = rxinit_status = 0;
665 return rx_InitHost(htonl(INADDR_ANY), port);
671 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
672 * maintaing the round trip timer.
677 * Start a new RTT timer for a given call and packet.
679 * There must be no resendEvent already listed for this call, otherwise this
680 * will leak events - intended for internal use within the RTO code only
683 * the RX call to start the timer for
684 * @param[in] lastPacket
685 * a flag indicating whether the last packet has been sent or not
687 * @pre call must be locked before calling this function
691 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
693 struct clock now, retryTime;
698 clock_Add(&retryTime, &call->rto);
700 /* If we're sending the last packet, and we're the client, then the server
701 * may wait for an additional 400ms before returning the ACK, wait for it
702 * rather than hitting a timeout */
703 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
704 clock_Addmsec(&retryTime, 400);
706 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
707 call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
712 * Cancel an RTT timer for a given call.
716 * the RX call to cancel the timer for
718 * @pre call must be locked before calling this function
723 rxi_rto_cancel(struct rx_call *call)
725 rxevent_Cancel(&call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
729 * Tell the RTO timer that we have sent a packet.
731 * If the timer isn't already running, then start it. If the timer is running,
735 * the RX call that the packet has been sent on
736 * @param[in] lastPacket
737 * A flag which is true if this is the last packet for the call
739 * @pre The call must be locked before calling this function
744 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
746 if (call->resendEvent)
749 rxi_rto_startTimer(call, lastPacket, istack);
753 * Tell the RTO timer that we have received an new ACK message
755 * This function should be called whenever a call receives an ACK that
756 * acknowledges new packets. Whatever happens, we stop the current timer.
757 * If there are unacked packets in the queue which have been sent, then
758 * we restart the timer from now. Otherwise, we leave it stopped.
761 * the RX call that the ACK has been received on
765 rxi_rto_packet_acked(struct rx_call *call, int istack)
767 struct rx_packet *p, *nxp;
769 rxi_rto_cancel(call);
771 if (queue_IsEmpty(&call->tq))
774 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
775 if (p->header.seq > call->tfirst + call->twind)
778 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
779 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
787 * Set an initial round trip timeout for a peer connection
789 * @param[in] secs The timeout to set in seconds
793 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
794 peer->rtt = secs * 8000;
798 * Enables or disables the busy call channel error (RX_CALL_BUSY).
800 * @param[in] onoff Non-zero to enable busy call channel errors.
802 * @pre Neither rx_Init nor rx_InitHost have been called yet
805 rx_SetBusyChannelError(afs_int32 onoff)
807 osi_Assert(rxinit_status != 0);
808 rxi_busyChannelError = onoff ? 1 : 0;
812 * Set a delayed ack event on the specified call for the given time
814 * @param[in] call - the call on which to set the event
815 * @param[in] offset - the delay from now after which the event fires
818 rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
820 struct clock now, when;
824 clock_Add(&when, offset);
826 if (!call->delayedAckEvent
827 || clock_Gt(&call->delayedAckTime, &when)) {
829 rxevent_Cancel(&call->delayedAckEvent, call,
830 RX_CALL_REFCOUNT_DELAY);
831 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
833 call->delayedAckEvent = rxevent_Post(&when, &now,
836 call->delayedAckTime = when;
840 /* called with unincremented nRequestsRunning to see if it is OK to start
841 * a new thread in this service. Could be "no" for two reasons: over the
842 * max quota, or would prevent others from reaching their min quota.
844 #ifdef RX_ENABLE_LOCKS
845 /* This verion of QuotaOK reserves quota if it's ok while the
846 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
849 QuotaOK(struct rx_service *aservice)
851 /* check if over max quota */
852 if (aservice->nRequestsRunning >= aservice->maxProcs) {
856 /* under min quota, we're OK */
857 /* otherwise, can use only if there are enough to allow everyone
858 * to go to their min quota after this guy starts.
861 MUTEX_ENTER(&rx_quota_mutex);
862 if ((aservice->nRequestsRunning < aservice->minProcs)
863 || (rxi_availProcs > rxi_minDeficit)) {
864 aservice->nRequestsRunning++;
865 /* just started call in minProcs pool, need fewer to maintain
867 if (aservice->nRequestsRunning <= aservice->minProcs)
870 MUTEX_EXIT(&rx_quota_mutex);
873 MUTEX_EXIT(&rx_quota_mutex);
879 ReturnToServerPool(struct rx_service *aservice)
881 aservice->nRequestsRunning--;
882 MUTEX_ENTER(&rx_quota_mutex);
883 if (aservice->nRequestsRunning < aservice->minProcs)
886 MUTEX_EXIT(&rx_quota_mutex);
889 #else /* RX_ENABLE_LOCKS */
891 QuotaOK(struct rx_service *aservice)
894 /* under min quota, we're OK */
895 if (aservice->nRequestsRunning < aservice->minProcs)
898 /* check if over max quota */
899 if (aservice->nRequestsRunning >= aservice->maxProcs)
902 /* otherwise, can use only if there are enough to allow everyone
903 * to go to their min quota after this guy starts.
905 MUTEX_ENTER(&rx_quota_mutex);
906 if (rxi_availProcs > rxi_minDeficit)
908 MUTEX_EXIT(&rx_quota_mutex);
911 #endif /* RX_ENABLE_LOCKS */
914 /* Called by rx_StartServer to start up lwp's to service calls.
915 NExistingProcs gives the number of procs already existing, and which
916 therefore needn't be created. */
918 rxi_StartServerProcs(int nExistingProcs)
920 struct rx_service *service;
925 /* For each service, reserve N processes, where N is the "minimum"
926 * number of processes that MUST be able to execute a request in parallel,
927 * at any time, for that process. Also compute the maximum difference
928 * between any service's maximum number of processes that can run
929 * (i.e. the maximum number that ever will be run, and a guarantee
930 * that this number will run if other services aren't running), and its
931 * minimum number. The result is the extra number of processes that
932 * we need in order to provide the latter guarantee */
933 for (i = 0; i < RX_MAX_SERVICES; i++) {
935 service = rx_services[i];
936 if (service == (struct rx_service *)0)
938 nProcs += service->minProcs;
939 diff = service->maxProcs - service->minProcs;
943 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
944 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
945 for (i = 0; i < nProcs; i++) {
946 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
952 /* This routine is only required on Windows */
954 rx_StartClientThread(void)
956 #ifdef AFS_PTHREAD_ENV
958 pid = pthread_self();
959 #endif /* AFS_PTHREAD_ENV */
961 #endif /* AFS_NT40_ENV */
963 /* This routine must be called if any services are exported. If the
964 * donateMe flag is set, the calling process is donated to the server
967 rx_StartServer(int donateMe)
969 struct rx_service *service;
975 /* Start server processes, if necessary (exact function is dependent
976 * on the implementation environment--kernel or user space). DonateMe
977 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
978 * case, one less new proc will be created rx_StartServerProcs.
980 rxi_StartServerProcs(donateMe);
982 /* count up the # of threads in minProcs, and add set the min deficit to
983 * be that value, too.
985 for (i = 0; i < RX_MAX_SERVICES; i++) {
986 service = rx_services[i];
987 if (service == (struct rx_service *)0)
989 MUTEX_ENTER(&rx_quota_mutex);
990 rxi_totalMin += service->minProcs;
991 /* below works even if a thread is running, since minDeficit would
992 * still have been decremented and later re-incremented.
994 rxi_minDeficit += service->minProcs;
995 MUTEX_EXIT(&rx_quota_mutex);
998 /* Turn on reaping of idle server connections */
999 rxi_ReapConnections(NULL, NULL, NULL, 0);
1004 #ifndef AFS_NT40_ENV
1008 #ifdef AFS_PTHREAD_ENV
1010 pid = afs_pointer_to_int(pthread_self());
1011 #else /* AFS_PTHREAD_ENV */
1013 LWP_CurrentProcess(&pid);
1014 #endif /* AFS_PTHREAD_ENV */
1016 sprintf(name, "srv_%d", ++nProcs);
1017 if (registerProgram)
1018 (*registerProgram) (pid, name);
1020 #endif /* AFS_NT40_ENV */
1021 rx_ServerProc(NULL); /* Never returns */
1023 #ifdef RX_ENABLE_TSFPQ
1024 /* no use leaving packets around in this thread's local queue if
1025 * it isn't getting donated to the server thread pool.
1027 rxi_FlushLocalPacketsTSFPQ();
1028 #endif /* RX_ENABLE_TSFPQ */
1032 /* Create a new client connection to the specified service, using the
1033 * specified security object to implement the security model for this
1035 struct rx_connection *
1036 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
1037 struct rx_securityClass *securityObject,
1038 int serviceSecurityIndex)
1042 struct rx_connection *conn;
1047 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
1048 "serviceSecurityIndex %d)\n",
1049 ntohl(shost), ntohs(sport), sservice, securityObject,
1050 serviceSecurityIndex));
1052 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
1053 * the case of kmem_alloc? */
1054 conn = rxi_AllocConnection();
1055 #ifdef RX_ENABLE_LOCKS
1056 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
1057 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
1058 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
1061 MUTEX_ENTER(&rx_connHashTable_lock);
1062 cid = (rx_nextCid += RX_MAXCALLS);
1063 conn->type = RX_CLIENT_CONNECTION;
1065 conn->epoch = rx_epoch;
1066 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
1067 conn->serviceId = sservice;
1068 conn->securityObject = securityObject;
1069 conn->securityData = (void *) 0;
1070 conn->securityIndex = serviceSecurityIndex;
1071 rx_SetConnDeadTime(conn, rx_connDeadTime);
1072 rx_SetConnSecondsUntilNatPing(conn, 0);
1073 conn->ackRate = RX_FAST_ACK_RATE;
1074 conn->nSpecific = 0;
1075 conn->specific = NULL;
1076 conn->challengeEvent = NULL;
1077 conn->delayedAbortEvent = NULL;
1078 conn->abortCount = 0;
1080 for (i = 0; i < RX_MAXCALLS; i++) {
1081 conn->twind[i] = rx_initSendWindow;
1082 conn->rwind[i] = rx_initReceiveWindow;
1083 conn->lastBusy[i] = 0;
1086 RXS_NewConnection(securityObject, conn);
1088 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1090 conn->refCount++; /* no lock required since only this thread knows... */
1091 conn->next = rx_connHashTable[hashindex];
1092 rx_connHashTable[hashindex] = conn;
1093 if (rx_stats_active)
1094 rx_atomic_inc(&rx_stats.nClientConns);
1095 MUTEX_EXIT(&rx_connHashTable_lock);
1101 * Ensure a connection's timeout values are valid.
1103 * @param[in] conn The connection to check
1105 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1106 * unless idleDeadTime and/or hardDeadTime are not set
1110 rxi_CheckConnTimeouts(struct rx_connection *conn)
1112 /* a connection's timeouts must have the relationship
1113 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1114 * total loss of network to a peer may cause an idle timeout instead of a
1115 * dead timeout, simply because the idle timeout gets hit first. Also set
1116 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1117 /* this logic is slightly complicated by the fact that
1118 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1120 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1121 if (conn->idleDeadTime) {
1122 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1124 if (conn->hardDeadTime) {
1125 if (conn->idleDeadTime) {
1126 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1128 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1134 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1136 /* The idea is to set the dead time to a value that allows several
1137 * keepalives to be dropped without timing out the connection. */
1138 conn->secondsUntilDead = seconds;
1139 rxi_CheckConnTimeouts(conn);
1140 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1144 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1146 conn->hardDeadTime = seconds;
1147 rxi_CheckConnTimeouts(conn);
1151 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1153 conn->idleDeadTime = seconds;
1154 conn->idleDeadDetection = (seconds ? 1 : 0);
1155 rxi_CheckConnTimeouts(conn);
1158 int rxi_lowPeerRefCount = 0;
1159 int rxi_lowConnRefCount = 0;
1162 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1163 * NOTE: must not be called with rx_connHashTable_lock held.
1166 rxi_CleanupConnection(struct rx_connection *conn)
1168 /* Notify the service exporter, if requested, that this connection
1169 * is being destroyed */
1170 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1171 (*conn->service->destroyConnProc) (conn);
1173 /* Notify the security module that this connection is being destroyed */
1174 RXS_DestroyConnection(conn->securityObject, conn);
1176 /* If this is the last connection using the rx_peer struct, set its
1177 * idle time to now. rxi_ReapConnections will reap it if it's still
1178 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1180 MUTEX_ENTER(&rx_peerHashTable_lock);
1181 if (conn->peer->refCount < 2) {
1182 conn->peer->idleWhen = clock_Sec();
1183 if (conn->peer->refCount < 1) {
1184 conn->peer->refCount = 1;
1185 if (rx_stats_active) {
1186 MUTEX_ENTER(&rx_stats_mutex);
1187 rxi_lowPeerRefCount++;
1188 MUTEX_EXIT(&rx_stats_mutex);
1192 conn->peer->refCount--;
1193 MUTEX_EXIT(&rx_peerHashTable_lock);
1195 if (rx_stats_active)
1197 if (conn->type == RX_SERVER_CONNECTION)
1198 rx_atomic_dec(&rx_stats.nServerConns);
1200 rx_atomic_dec(&rx_stats.nClientConns);
1203 if (conn->specific) {
1205 for (i = 0; i < conn->nSpecific; i++) {
1206 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1207 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1208 conn->specific[i] = NULL;
1210 free(conn->specific);
1212 conn->specific = NULL;
1213 conn->nSpecific = 0;
1214 #endif /* !KERNEL */
1216 MUTEX_DESTROY(&conn->conn_call_lock);
1217 MUTEX_DESTROY(&conn->conn_data_lock);
1218 CV_DESTROY(&conn->conn_call_cv);
1220 rxi_FreeConnection(conn);
1223 /* Destroy the specified connection */
1225 rxi_DestroyConnection(struct rx_connection *conn)
1227 MUTEX_ENTER(&rx_connHashTable_lock);
1228 rxi_DestroyConnectionNoLock(conn);
1229 /* conn should be at the head of the cleanup list */
1230 if (conn == rx_connCleanup_list) {
1231 rx_connCleanup_list = rx_connCleanup_list->next;
1232 MUTEX_EXIT(&rx_connHashTable_lock);
1233 rxi_CleanupConnection(conn);
1235 #ifdef RX_ENABLE_LOCKS
1237 MUTEX_EXIT(&rx_connHashTable_lock);
1239 #endif /* RX_ENABLE_LOCKS */
1243 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1245 struct rx_connection **conn_ptr;
1247 struct rx_packet *packet;
1254 MUTEX_ENTER(&conn->conn_data_lock);
1255 MUTEX_ENTER(&rx_refcnt_mutex);
1256 if (conn->refCount > 0)
1259 if (rx_stats_active) {
1260 MUTEX_ENTER(&rx_stats_mutex);
1261 rxi_lowConnRefCount++;
1262 MUTEX_EXIT(&rx_stats_mutex);
1266 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1267 /* Busy; wait till the last guy before proceeding */
1268 MUTEX_EXIT(&rx_refcnt_mutex);
1269 MUTEX_EXIT(&conn->conn_data_lock);
1274 /* If the client previously called rx_NewCall, but it is still
1275 * waiting, treat this as a running call, and wait to destroy the
1276 * connection later when the call completes. */
1277 if ((conn->type == RX_CLIENT_CONNECTION)
1278 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1279 conn->flags |= RX_CONN_DESTROY_ME;
1280 MUTEX_EXIT(&conn->conn_data_lock);
1284 MUTEX_EXIT(&rx_refcnt_mutex);
1285 MUTEX_EXIT(&conn->conn_data_lock);
1287 /* Check for extant references to this connection */
1288 MUTEX_ENTER(&conn->conn_call_lock);
1289 for (i = 0; i < RX_MAXCALLS; i++) {
1290 struct rx_call *call = conn->call[i];
1293 if (conn->type == RX_CLIENT_CONNECTION) {
1294 MUTEX_ENTER(&call->lock);
1295 if (call->delayedAckEvent) {
1296 /* Push the final acknowledgment out now--there
1297 * won't be a subsequent call to acknowledge the
1298 * last reply packets */
1299 rxevent_Cancel(&call->delayedAckEvent, call,
1300 RX_CALL_REFCOUNT_DELAY);
1301 if (call->state == RX_STATE_PRECALL
1302 || call->state == RX_STATE_ACTIVE) {
1303 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1308 MUTEX_EXIT(&call->lock);
1312 MUTEX_EXIT(&conn->conn_call_lock);
1314 #ifdef RX_ENABLE_LOCKS
1316 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1317 MUTEX_EXIT(&conn->conn_data_lock);
1319 /* Someone is accessing a packet right now. */
1323 #endif /* RX_ENABLE_LOCKS */
1326 /* Don't destroy the connection if there are any call
1327 * structures still in use */
1328 MUTEX_ENTER(&conn->conn_data_lock);
1329 conn->flags |= RX_CONN_DESTROY_ME;
1330 MUTEX_EXIT(&conn->conn_data_lock);
1335 if (conn->natKeepAliveEvent) {
1336 rxi_NatKeepAliveOff(conn);
1339 if (conn->delayedAbortEvent) {
1340 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
1341 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1343 MUTEX_ENTER(&conn->conn_data_lock);
1344 rxi_SendConnectionAbort(conn, packet, 0, 1);
1345 MUTEX_EXIT(&conn->conn_data_lock);
1346 rxi_FreePacket(packet);
1350 /* Remove from connection hash table before proceeding */
1352 &rx_connHashTable[CONN_HASH
1353 (peer->host, peer->port, conn->cid, conn->epoch,
1355 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1356 if (*conn_ptr == conn) {
1357 *conn_ptr = conn->next;
1361 /* if the conn that we are destroying was the last connection, then we
1362 * clear rxLastConn as well */
1363 if (rxLastConn == conn)
1366 /* Make sure the connection is completely reset before deleting it. */
1367 /* get rid of pending events that could zap us later */
1368 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
1369 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
1370 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
1372 /* Add the connection to the list of destroyed connections that
1373 * need to be cleaned up. This is necessary to avoid deadlocks
1374 * in the routines we call to inform others that this connection is
1375 * being destroyed. */
1376 conn->next = rx_connCleanup_list;
1377 rx_connCleanup_list = conn;
1380 /* Externally available version */
1382 rx_DestroyConnection(struct rx_connection *conn)
1387 rxi_DestroyConnection(conn);
1392 rx_GetConnection(struct rx_connection *conn)
1397 MUTEX_ENTER(&rx_refcnt_mutex);
1399 MUTEX_EXIT(&rx_refcnt_mutex);
1403 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1404 /* Wait for the transmit queue to no longer be busy.
1405 * requires the call->lock to be held */
1407 rxi_WaitforTQBusy(struct rx_call *call) {
1408 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1409 call->flags |= RX_CALL_TQ_WAIT;
1411 #ifdef RX_ENABLE_LOCKS
1412 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1413 CV_WAIT(&call->cv_tq, &call->lock);
1414 #else /* RX_ENABLE_LOCKS */
1415 osi_rxSleep(&call->tq);
1416 #endif /* RX_ENABLE_LOCKS */
1418 if (call->tqWaiters == 0) {
1419 call->flags &= ~RX_CALL_TQ_WAIT;
1426 rxi_WakeUpTransmitQueue(struct rx_call *call)
1428 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1429 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1430 call, call->tqWaiters, call->flags));
1431 #ifdef RX_ENABLE_LOCKS
1432 osirx_AssertMine(&call->lock, "rxi_Start start");
1433 CV_BROADCAST(&call->cv_tq);
1434 #else /* RX_ENABLE_LOCKS */
1435 osi_rxWakeup(&call->tq);
1436 #endif /* RX_ENABLE_LOCKS */
1440 /* Start a new rx remote procedure call, on the specified connection.
1441 * If wait is set to 1, wait for a free call channel; otherwise return
1442 * 0. Maxtime gives the maximum number of seconds this call may take,
1443 * after rx_NewCall returns. After this time interval, a call to any
1444 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1445 * For fine grain locking, we hold the conn_call_lock in order to
1446 * to ensure that we don't get signalle after we found a call in an active
1447 * state and before we go to sleep.
1450 rx_NewCall(struct rx_connection *conn)
1452 int i, wait, ignoreBusy = 1;
1453 struct rx_call *call;
1454 struct clock queueTime;
1455 afs_uint32 leastBusy = 0;
1459 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1462 clock_GetTime(&queueTime);
1464 * Check if there are others waiting for a new call.
1465 * If so, let them go first to avoid starving them.
1466 * This is a fairly simple scheme, and might not be
1467 * a complete solution for large numbers of waiters.
1469 * makeCallWaiters keeps track of the number of
1470 * threads waiting to make calls and the
1471 * RX_CONN_MAKECALL_WAITING flag bit is used to
1472 * indicate that there are indeed calls waiting.
1473 * The flag is set when the waiter is incremented.
1474 * It is only cleared when makeCallWaiters is 0.
1475 * This prevents us from accidently destroying the
1476 * connection while it is potentially about to be used.
1478 MUTEX_ENTER(&conn->conn_call_lock);
1479 MUTEX_ENTER(&conn->conn_data_lock);
1480 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1481 conn->flags |= RX_CONN_MAKECALL_WAITING;
1482 conn->makeCallWaiters++;
1483 MUTEX_EXIT(&conn->conn_data_lock);
1485 #ifdef RX_ENABLE_LOCKS
1486 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1490 MUTEX_ENTER(&conn->conn_data_lock);
1491 conn->makeCallWaiters--;
1492 if (conn->makeCallWaiters == 0)
1493 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1496 /* We are now the active thread in rx_NewCall */
1497 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1498 MUTEX_EXIT(&conn->conn_data_lock);
1503 for (i = 0; i < RX_MAXCALLS; i++) {
1504 call = conn->call[i];
1506 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1507 /* we're not ignoring busy call slots; only look at the
1508 * call slot that is the "least" busy */
1512 if (call->state == RX_STATE_DALLY) {
1513 MUTEX_ENTER(&call->lock);
1514 if (call->state == RX_STATE_DALLY) {
1515 if (ignoreBusy && conn->lastBusy[i]) {
1516 /* if we're ignoring busy call slots, skip any ones that
1517 * have lastBusy set */
1518 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1519 leastBusy = conn->lastBusy[i];
1521 MUTEX_EXIT(&call->lock);
1526 * We are setting the state to RX_STATE_RESET to
1527 * ensure that no one else will attempt to use this
1528 * call once we drop the conn->conn_call_lock and
1529 * call->lock. We must drop the conn->conn_call_lock
1530 * before calling rxi_ResetCall because the process
1531 * of clearing the transmit queue can block for an
1532 * extended period of time. If we block while holding
1533 * the conn->conn_call_lock, then all rx_EndCall
1534 * processing will block as well. This has a detrimental
1535 * effect on overall system performance.
1537 call->state = RX_STATE_RESET;
1538 (*call->callNumber)++;
1539 MUTEX_EXIT(&conn->conn_call_lock);
1540 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1541 rxi_ResetCall(call, 0);
1542 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1546 * If we failed to be able to safely obtain the
1547 * conn->conn_call_lock we will have to drop the
1548 * call->lock to avoid a deadlock. When the call->lock
1549 * is released the state of the call can change. If it
1550 * is no longer RX_STATE_RESET then some other thread is
1553 MUTEX_EXIT(&call->lock);
1554 MUTEX_ENTER(&conn->conn_call_lock);
1555 MUTEX_ENTER(&call->lock);
1557 if (call->state == RX_STATE_RESET)
1561 * If we get here it means that after dropping
1562 * the conn->conn_call_lock and call->lock that
1563 * the call is no longer ours. If we can't find
1564 * a free call in the remaining slots we should
1565 * not go immediately to RX_CONN_MAKECALL_WAITING
1566 * because by dropping the conn->conn_call_lock
1567 * we have given up synchronization with rx_EndCall.
1568 * Instead, cycle through one more time to see if
1569 * we can find a call that can call our own.
1571 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1574 MUTEX_EXIT(&call->lock);
1577 if (ignoreBusy && conn->lastBusy[i]) {
1578 /* if we're ignoring busy call slots, skip any ones that
1579 * have lastBusy set */
1580 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1581 leastBusy = conn->lastBusy[i];
1586 /* rxi_NewCall returns with mutex locked */
1587 call = rxi_NewCall(conn, i);
1588 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1592 if (i < RX_MAXCALLS) {
1593 conn->lastBusy[i] = 0;
1594 call->flags &= ~RX_CALL_PEER_BUSY;
1599 if (leastBusy && ignoreBusy) {
1600 /* we didn't find a useable call slot, but we did see at least one
1601 * 'busy' slot; look again and only use a slot with the 'least
1607 MUTEX_ENTER(&conn->conn_data_lock);
1608 conn->flags |= RX_CONN_MAKECALL_WAITING;
1609 conn->makeCallWaiters++;
1610 MUTEX_EXIT(&conn->conn_data_lock);
1612 #ifdef RX_ENABLE_LOCKS
1613 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1617 MUTEX_ENTER(&conn->conn_data_lock);
1618 conn->makeCallWaiters--;
1619 if (conn->makeCallWaiters == 0)
1620 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1621 MUTEX_EXIT(&conn->conn_data_lock);
1623 /* Client is initially in send mode */
1624 call->state = RX_STATE_ACTIVE;
1625 call->error = conn->error;
1627 call->mode = RX_MODE_ERROR;
1629 call->mode = RX_MODE_SENDING;
1631 #ifdef AFS_RXERRQ_ENV
1632 /* remember how many network errors the peer has when we started, so if
1633 * more errors are encountered after the call starts, we know the other endpoint won't be
1634 * responding to us */
1635 call->neterr_gen = rx_atomic_read(&conn->peer->neterrs);
1638 /* remember start time for call in case we have hard dead time limit */
1639 call->queueTime = queueTime;
1640 clock_GetTime(&call->startTime);
1641 call->bytesSent = 0;
1642 call->bytesRcvd = 0;
1644 /* Turn on busy protocol. */
1645 rxi_KeepAliveOn(call);
1647 /* Attempt MTU discovery */
1648 rxi_GrowMTUOn(call);
1651 * We are no longer the active thread in rx_NewCall
1653 MUTEX_ENTER(&conn->conn_data_lock);
1654 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1655 MUTEX_EXIT(&conn->conn_data_lock);
1658 * Wake up anyone else who might be giving us a chance to
1659 * run (see code above that avoids resource starvation).
1661 #ifdef RX_ENABLE_LOCKS
1662 CV_BROADCAST(&conn->conn_call_cv);
1666 MUTEX_EXIT(&conn->conn_call_lock);
1668 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1669 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1670 osi_Panic("rx_NewCall call about to be used without an empty tq");
1672 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1674 MUTEX_EXIT(&call->lock);
1677 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1682 rxi_HasActiveCalls(struct rx_connection *aconn)
1685 struct rx_call *tcall;
1689 for (i = 0; i < RX_MAXCALLS; i++) {
1690 if ((tcall = aconn->call[i])) {
1691 if ((tcall->state == RX_STATE_ACTIVE)
1692 || (tcall->state == RX_STATE_PRECALL)) {
1703 rxi_GetCallNumberVector(struct rx_connection *aconn,
1704 afs_int32 * aint32s)
1707 struct rx_call *tcall;
1711 MUTEX_ENTER(&aconn->conn_call_lock);
1712 for (i = 0; i < RX_MAXCALLS; i++) {
1713 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1714 aint32s[i] = aconn->callNumber[i] + 1;
1716 aint32s[i] = aconn->callNumber[i];
1718 MUTEX_EXIT(&aconn->conn_call_lock);
1724 rxi_SetCallNumberVector(struct rx_connection *aconn,
1725 afs_int32 * aint32s)
1728 struct rx_call *tcall;
1732 MUTEX_ENTER(&aconn->conn_call_lock);
1733 for (i = 0; i < RX_MAXCALLS; i++) {
1734 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1735 aconn->callNumber[i] = aint32s[i] - 1;
1737 aconn->callNumber[i] = aint32s[i];
1739 MUTEX_EXIT(&aconn->conn_call_lock);
1744 /* Advertise a new service. A service is named locally by a UDP port
1745 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1748 char *serviceName; Name for identification purposes (e.g. the
1749 service name might be used for probing for
1752 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1753 char *serviceName, struct rx_securityClass **securityObjects,
1754 int nSecurityObjects,
1755 afs_int32(*serviceProc) (struct rx_call * acall))
1757 osi_socket socket = OSI_NULLSOCKET;
1758 struct rx_service *tservice;
1764 if (serviceId == 0) {
1766 "rx_NewService: service id for service %s is not non-zero.\n",
1773 "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",
1781 tservice = rxi_AllocService();
1784 #ifdef RX_ENABLE_LOCKS
1785 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1788 for (i = 0; i < RX_MAX_SERVICES; i++) {
1789 struct rx_service *service = rx_services[i];
1791 if (port == service->servicePort && host == service->serviceHost) {
1792 if (service->serviceId == serviceId) {
1793 /* The identical service has already been
1794 * installed; if the caller was intending to
1795 * change the security classes used by this
1796 * service, he/she loses. */
1798 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1799 serviceName, serviceId, service->serviceName);
1801 rxi_FreeService(tservice);
1804 /* Different service, same port: re-use the socket
1805 * which is bound to the same port */
1806 socket = service->socket;
1809 if (socket == OSI_NULLSOCKET) {
1810 /* If we don't already have a socket (from another
1811 * service on same port) get a new one */
1812 socket = rxi_GetHostUDPSocket(host, port);
1813 if (socket == OSI_NULLSOCKET) {
1815 rxi_FreeService(tservice);
1820 service->socket = socket;
1821 service->serviceHost = host;
1822 service->servicePort = port;
1823 service->serviceId = serviceId;
1824 service->serviceName = serviceName;
1825 service->nSecurityObjects = nSecurityObjects;
1826 service->securityObjects = securityObjects;
1827 service->minProcs = 0;
1828 service->maxProcs = 1;
1829 service->idleDeadTime = 60;
1830 service->idleDeadErr = 0;
1831 service->connDeadTime = rx_connDeadTime;
1832 service->executeRequestProc = serviceProc;
1833 service->checkReach = 0;
1834 service->nSpecific = 0;
1835 service->specific = NULL;
1836 rx_services[i] = service; /* not visible until now */
1842 rxi_FreeService(tservice);
1843 (osi_Msg "rx_NewService: cannot support > %d services\n",
1848 /* Set configuration options for all of a service's security objects */
1851 rx_SetSecurityConfiguration(struct rx_service *service,
1852 rx_securityConfigVariables type,
1856 for (i = 0; i<service->nSecurityObjects; i++) {
1857 if (service->securityObjects[i]) {
1858 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1866 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1867 struct rx_securityClass **securityObjects, int nSecurityObjects,
1868 afs_int32(*serviceProc) (struct rx_call * acall))
1870 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1873 /* Generic request processing loop. This routine should be called
1874 * by the implementation dependent rx_ServerProc. If socketp is
1875 * non-null, it will be set to the file descriptor that this thread
1876 * is now listening on. If socketp is null, this routine will never
1879 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1881 struct rx_call *call;
1883 struct rx_service *tservice = NULL;
1890 call = rx_GetCall(threadID, tservice, socketp);
1891 if (socketp && *socketp != OSI_NULLSOCKET) {
1892 /* We are now a listener thread */
1898 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1899 #ifdef RX_ENABLE_LOCKS
1901 #endif /* RX_ENABLE_LOCKS */
1902 afs_termState = AFSOP_STOP_AFS;
1903 afs_osi_Wakeup(&afs_termState);
1904 #ifdef RX_ENABLE_LOCKS
1906 #endif /* RX_ENABLE_LOCKS */
1911 /* if server is restarting( typically smooth shutdown) then do not
1912 * allow any new calls.
1915 if (rx_tranquil && (call != NULL)) {
1919 MUTEX_ENTER(&call->lock);
1921 rxi_CallError(call, RX_RESTARTING);
1922 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1924 MUTEX_EXIT(&call->lock);
1929 tservice = call->conn->service;
1931 if (tservice->beforeProc)
1932 (*tservice->beforeProc) (call);
1934 code = tservice->executeRequestProc(call);
1936 if (tservice->afterProc)
1937 (*tservice->afterProc) (call, code);
1939 rx_EndCall(call, code);
1941 if (tservice->postProc)
1942 (*tservice->postProc) (code);
1944 if (rx_stats_active) {
1945 MUTEX_ENTER(&rx_stats_mutex);
1947 MUTEX_EXIT(&rx_stats_mutex);
1954 rx_WakeupServerProcs(void)
1956 struct rx_serverQueueEntry *np, *tqp;
1960 MUTEX_ENTER(&rx_serverPool_lock);
1962 #ifdef RX_ENABLE_LOCKS
1963 if (rx_waitForPacket)
1964 CV_BROADCAST(&rx_waitForPacket->cv);
1965 #else /* RX_ENABLE_LOCKS */
1966 if (rx_waitForPacket)
1967 osi_rxWakeup(rx_waitForPacket);
1968 #endif /* RX_ENABLE_LOCKS */
1969 MUTEX_ENTER(&freeSQEList_lock);
1970 for (np = rx_FreeSQEList; np; np = tqp) {
1971 tqp = *(struct rx_serverQueueEntry **)np;
1972 #ifdef RX_ENABLE_LOCKS
1973 CV_BROADCAST(&np->cv);
1974 #else /* RX_ENABLE_LOCKS */
1976 #endif /* RX_ENABLE_LOCKS */
1978 MUTEX_EXIT(&freeSQEList_lock);
1979 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1980 #ifdef RX_ENABLE_LOCKS
1981 CV_BROADCAST(&np->cv);
1982 #else /* RX_ENABLE_LOCKS */
1984 #endif /* RX_ENABLE_LOCKS */
1986 MUTEX_EXIT(&rx_serverPool_lock);
1991 * One thing that seems to happen is that all the server threads get
1992 * tied up on some empty or slow call, and then a whole bunch of calls
1993 * arrive at once, using up the packet pool, so now there are more
1994 * empty calls. The most critical resources here are server threads
1995 * and the free packet pool. The "doreclaim" code seems to help in
1996 * general. I think that eventually we arrive in this state: there
1997 * are lots of pending calls which do have all their packets present,
1998 * so they won't be reclaimed, are multi-packet calls, so they won't
1999 * be scheduled until later, and thus are tying up most of the free
2000 * packet pool for a very long time.
2002 * 1. schedule multi-packet calls if all the packets are present.
2003 * Probably CPU-bound operation, useful to return packets to pool.
2004 * Do what if there is a full window, but the last packet isn't here?
2005 * 3. preserve one thread which *only* runs "best" calls, otherwise
2006 * it sleeps and waits for that type of call.
2007 * 4. Don't necessarily reserve a whole window for each thread. In fact,
2008 * the current dataquota business is badly broken. The quota isn't adjusted
2009 * to reflect how many packets are presently queued for a running call.
2010 * So, when we schedule a queued call with a full window of packets queued
2011 * up for it, that *should* free up a window full of packets for other 2d-class
2012 * calls to be able to use from the packet pool. But it doesn't.
2014 * NB. Most of the time, this code doesn't run -- since idle server threads
2015 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
2016 * as a new call arrives.
2018 /* Sleep until a call arrives. Returns a pointer to the call, ready
2019 * for an rx_Read. */
2020 #ifdef RX_ENABLE_LOCKS
2022 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2024 struct rx_serverQueueEntry *sq;
2025 struct rx_call *call = (struct rx_call *)0;
2026 struct rx_service *service = NULL;
2028 MUTEX_ENTER(&freeSQEList_lock);
2030 if ((sq = rx_FreeSQEList)) {
2031 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2032 MUTEX_EXIT(&freeSQEList_lock);
2033 } else { /* otherwise allocate a new one and return that */
2034 MUTEX_EXIT(&freeSQEList_lock);
2035 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2036 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2037 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2040 MUTEX_ENTER(&rx_serverPool_lock);
2041 if (cur_service != NULL) {
2042 ReturnToServerPool(cur_service);
2045 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
2046 struct rx_call *tcall, *ncall, *choice2 = NULL;
2048 /* Scan for eligible incoming calls. A call is not eligible
2049 * if the maximum number of calls for its service type are
2050 * already executing */
2051 /* One thread will process calls FCFS (to prevent starvation),
2052 * while the other threads may run ahead looking for calls which
2053 * have all their input data available immediately. This helps
2054 * keep threads from blocking, waiting for data from the client. */
2055 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
2056 service = tcall->conn->service;
2057 if (!QuotaOK(service)) {
2060 MUTEX_ENTER(&rx_pthread_mutex);
2061 if (tno == rxi_fcfs_thread_num
2062 || queue_IsLast(&rx_incomingCallQueue, tcall)) {
2063 MUTEX_EXIT(&rx_pthread_mutex);
2064 /* If we're the fcfs thread , then we'll just use
2065 * this call. If we haven't been able to find an optimal
2066 * choice, and we're at the end of the list, then use a
2067 * 2d choice if one has been identified. Otherwise... */
2068 call = (choice2 ? choice2 : tcall);
2069 service = call->conn->service;
2071 MUTEX_EXIT(&rx_pthread_mutex);
2072 if (!queue_IsEmpty(&tcall->rq)) {
2073 struct rx_packet *rp;
2074 rp = queue_First(&tcall->rq, rx_packet);
2075 if (rp->header.seq == 1) {
2077 || (rp->header.flags & RX_LAST_PACKET)) {
2079 } else if (rxi_2dchoice && !choice2
2080 && !(tcall->flags & RX_CALL_CLEARED)
2081 && (tcall->rprev > rxi_HardAckRate)) {
2091 ReturnToServerPool(service);
2098 MUTEX_EXIT(&rx_serverPool_lock);
2099 MUTEX_ENTER(&call->lock);
2101 if (call->flags & RX_CALL_WAIT_PROC) {
2102 call->flags &= ~RX_CALL_WAIT_PROC;
2103 rx_atomic_dec(&rx_nWaiting);
2106 if (call->state != RX_STATE_PRECALL || call->error) {
2107 MUTEX_EXIT(&call->lock);
2108 MUTEX_ENTER(&rx_serverPool_lock);
2109 ReturnToServerPool(service);
2114 if (queue_IsEmpty(&call->rq)
2115 || queue_First(&call->rq, rx_packet)->header.seq != 1)
2116 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2118 CLEAR_CALL_QUEUE_LOCK(call);
2121 /* If there are no eligible incoming calls, add this process
2122 * to the idle server queue, to wait for one */
2126 *socketp = OSI_NULLSOCKET;
2128 sq->socketp = socketp;
2129 queue_Append(&rx_idleServerQueue, sq);
2130 #ifndef AFS_AIX41_ENV
2131 rx_waitForPacket = sq;
2133 rx_waitingForPacket = sq;
2134 #endif /* AFS_AIX41_ENV */
2136 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2138 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2139 MUTEX_EXIT(&rx_serverPool_lock);
2140 return (struct rx_call *)0;
2143 } while (!(call = sq->newcall)
2144 && !(socketp && *socketp != OSI_NULLSOCKET));
2145 MUTEX_EXIT(&rx_serverPool_lock);
2147 MUTEX_ENTER(&call->lock);
2153 MUTEX_ENTER(&freeSQEList_lock);
2154 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2155 rx_FreeSQEList = sq;
2156 MUTEX_EXIT(&freeSQEList_lock);
2159 clock_GetTime(&call->startTime);
2160 call->state = RX_STATE_ACTIVE;
2161 call->mode = RX_MODE_RECEIVING;
2162 #ifdef RX_KERNEL_TRACE
2163 if (ICL_SETACTIVE(afs_iclSetp)) {
2164 int glockOwner = ISAFS_GLOCK();
2167 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2168 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2175 rxi_calltrace(RX_CALL_START, call);
2176 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2177 call->conn->service->servicePort, call->conn->service->serviceId,
2180 MUTEX_EXIT(&call->lock);
2181 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2183 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2188 #else /* RX_ENABLE_LOCKS */
2190 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2192 struct rx_serverQueueEntry *sq;
2193 struct rx_call *call = (struct rx_call *)0, *choice2;
2194 struct rx_service *service = NULL;
2198 MUTEX_ENTER(&freeSQEList_lock);
2200 if ((sq = rx_FreeSQEList)) {
2201 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2202 MUTEX_EXIT(&freeSQEList_lock);
2203 } else { /* otherwise allocate a new one and return that */
2204 MUTEX_EXIT(&freeSQEList_lock);
2205 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2206 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2207 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2209 MUTEX_ENTER(&sq->lock);
2211 if (cur_service != NULL) {
2212 cur_service->nRequestsRunning--;
2213 MUTEX_ENTER(&rx_quota_mutex);
2214 if (cur_service->nRequestsRunning < cur_service->minProcs)
2217 MUTEX_EXIT(&rx_quota_mutex);
2219 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
2220 struct rx_call *tcall, *ncall;
2221 /* Scan for eligible incoming calls. A call is not eligible
2222 * if the maximum number of calls for its service type are
2223 * already executing */
2224 /* One thread will process calls FCFS (to prevent starvation),
2225 * while the other threads may run ahead looking for calls which
2226 * have all their input data available immediately. This helps
2227 * keep threads from blocking, waiting for data from the client. */
2228 choice2 = (struct rx_call *)0;
2229 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
2230 service = tcall->conn->service;
2231 if (QuotaOK(service)) {
2232 MUTEX_ENTER(&rx_pthread_mutex);
2233 if (tno == rxi_fcfs_thread_num
2234 || !tcall->queue_item_header.next) {
2235 MUTEX_EXIT(&rx_pthread_mutex);
2236 /* If we're the fcfs thread, then we'll just use
2237 * this call. If we haven't been able to find an optimal
2238 * choice, and we're at the end of the list, then use a
2239 * 2d choice if one has been identified. Otherwise... */
2240 call = (choice2 ? choice2 : tcall);
2241 service = call->conn->service;
2243 MUTEX_EXIT(&rx_pthread_mutex);
2244 if (!queue_IsEmpty(&tcall->rq)) {
2245 struct rx_packet *rp;
2246 rp = queue_First(&tcall->rq, rx_packet);
2247 if (rp->header.seq == 1
2249 || (rp->header.flags & RX_LAST_PACKET))) {
2251 } else if (rxi_2dchoice && !choice2
2252 && !(tcall->flags & RX_CALL_CLEARED)
2253 && (tcall->rprev > rxi_HardAckRate)) {
2267 /* we can't schedule a call if there's no data!!! */
2268 /* send an ack if there's no data, if we're missing the
2269 * first packet, or we're missing something between first
2270 * and last -- there's a "hole" in the incoming data. */
2271 if (queue_IsEmpty(&call->rq)
2272 || queue_First(&call->rq, rx_packet)->header.seq != 1
2273 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
2274 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2276 call->flags &= (~RX_CALL_WAIT_PROC);
2277 service->nRequestsRunning++;
2278 /* just started call in minProcs pool, need fewer to maintain
2280 MUTEX_ENTER(&rx_quota_mutex);
2281 if (service->nRequestsRunning <= service->minProcs)
2284 MUTEX_EXIT(&rx_quota_mutex);
2285 rx_atomic_dec(&rx_nWaiting);
2286 /* MUTEX_EXIT(&call->lock); */
2288 /* If there are no eligible incoming calls, add this process
2289 * to the idle server queue, to wait for one */
2292 *socketp = OSI_NULLSOCKET;
2294 sq->socketp = socketp;
2295 queue_Append(&rx_idleServerQueue, sq);
2299 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2301 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2302 return (struct rx_call *)0;
2305 } while (!(call = sq->newcall)
2306 && !(socketp && *socketp != OSI_NULLSOCKET));
2308 MUTEX_EXIT(&sq->lock);
2310 MUTEX_ENTER(&freeSQEList_lock);
2311 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2312 rx_FreeSQEList = sq;
2313 MUTEX_EXIT(&freeSQEList_lock);
2316 clock_GetTime(&call->startTime);
2317 call->state = RX_STATE_ACTIVE;
2318 call->mode = RX_MODE_RECEIVING;
2319 #ifdef RX_KERNEL_TRACE
2320 if (ICL_SETACTIVE(afs_iclSetp)) {
2321 int glockOwner = ISAFS_GLOCK();
2324 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2325 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2332 rxi_calltrace(RX_CALL_START, call);
2333 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2334 call->conn->service->servicePort, call->conn->service->serviceId,
2337 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2344 #endif /* RX_ENABLE_LOCKS */
2348 /* Establish a procedure to be called when a packet arrives for a
2349 * call. This routine will be called at most once after each call,
2350 * and will also be called if there is an error condition on the or
2351 * the call is complete. Used by multi rx to build a selection
2352 * function which determines which of several calls is likely to be a
2353 * good one to read from.
2354 * NOTE: the way this is currently implemented it is probably only a
2355 * good idea to (1) use it immediately after a newcall (clients only)
2356 * and (2) only use it once. Other uses currently void your warranty
2359 rx_SetArrivalProc(struct rx_call *call,
2360 void (*proc) (struct rx_call * call,
2363 void * handle, int arg)
2365 call->arrivalProc = proc;
2366 call->arrivalProcHandle = handle;
2367 call->arrivalProcArg = arg;
2370 /* Call is finished (possibly prematurely). Return rc to the peer, if
2371 * appropriate, and return the final error code from the conversation
2375 rx_EndCall(struct rx_call *call, afs_int32 rc)
2377 struct rx_connection *conn = call->conn;
2381 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2382 call, rc, call->error, call->abortCode));
2385 MUTEX_ENTER(&call->lock);
2387 if (rc == 0 && call->error == 0) {
2388 call->abortCode = 0;
2389 call->abortCount = 0;
2392 call->arrivalProc = (void (*)())0;
2393 if (rc && call->error == 0) {
2394 rxi_CallError(call, rc);
2395 call->mode = RX_MODE_ERROR;
2396 /* Send an abort message to the peer if this error code has
2397 * only just been set. If it was set previously, assume the
2398 * peer has already been sent the error code or will request it
2400 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2402 if (conn->type == RX_SERVER_CONNECTION) {
2403 /* Make sure reply or at least dummy reply is sent */
2404 if (call->mode == RX_MODE_RECEIVING) {
2405 MUTEX_EXIT(&call->lock);
2406 rxi_WriteProc(call, 0, 0);
2407 MUTEX_ENTER(&call->lock);
2409 if (call->mode == RX_MODE_SENDING) {
2410 MUTEX_EXIT(&call->lock);
2411 rxi_FlushWrite(call);
2412 MUTEX_ENTER(&call->lock);
2414 rxi_calltrace(RX_CALL_END, call);
2415 /* Call goes to hold state until reply packets are acknowledged */
2416 if (call->tfirst + call->nSoftAcked < call->tnext) {
2417 call->state = RX_STATE_HOLD;
2419 call->state = RX_STATE_DALLY;
2420 rxi_ClearTransmitQueue(call, 0);
2421 rxi_rto_cancel(call);
2422 rxevent_Cancel(&call->keepAliveEvent, call,
2423 RX_CALL_REFCOUNT_ALIVE);
2425 } else { /* Client connection */
2427 /* Make sure server receives input packets, in the case where
2428 * no reply arguments are expected */
2429 if ((call->mode == RX_MODE_SENDING)
2430 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2431 MUTEX_EXIT(&call->lock);
2432 (void)rxi_ReadProc(call, &dummy, 1);
2433 MUTEX_ENTER(&call->lock);
2436 /* If we had an outstanding delayed ack, be nice to the server
2437 * and force-send it now.
2439 if (call->delayedAckEvent) {
2440 rxevent_Cancel(&call->delayedAckEvent, call,
2441 RX_CALL_REFCOUNT_DELAY);
2442 rxi_SendDelayedAck(NULL, call, NULL, 0);
2445 /* We need to release the call lock since it's lower than the
2446 * conn_call_lock and we don't want to hold the conn_call_lock
2447 * over the rx_ReadProc call. The conn_call_lock needs to be held
2448 * here for the case where rx_NewCall is perusing the calls on
2449 * the connection structure. We don't want to signal until
2450 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2451 * have checked this call, found it active and by the time it
2452 * goes to sleep, will have missed the signal.
2454 MUTEX_EXIT(&call->lock);
2455 MUTEX_ENTER(&conn->conn_call_lock);
2456 MUTEX_ENTER(&call->lock);
2458 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2459 conn->lastBusy[call->channel] = 0;
2462 MUTEX_ENTER(&conn->conn_data_lock);
2463 conn->flags |= RX_CONN_BUSY;
2464 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2465 MUTEX_EXIT(&conn->conn_data_lock);
2466 #ifdef RX_ENABLE_LOCKS
2467 CV_BROADCAST(&conn->conn_call_cv);
2472 #ifdef RX_ENABLE_LOCKS
2474 MUTEX_EXIT(&conn->conn_data_lock);
2476 #endif /* RX_ENABLE_LOCKS */
2477 call->state = RX_STATE_DALLY;
2479 error = call->error;
2481 /* currentPacket, nLeft, and NFree must be zeroed here, because
2482 * ResetCall cannot: ResetCall may be called at splnet(), in the
2483 * kernel version, and may interrupt the macros rx_Read or
2484 * rx_Write, which run at normal priority for efficiency. */
2485 if (call->currentPacket) {
2486 #ifdef RX_TRACK_PACKETS
2487 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2489 rxi_FreePacket(call->currentPacket);
2490 call->currentPacket = (struct rx_packet *)0;
2493 call->nLeft = call->nFree = call->curlen = 0;
2495 /* Free any packets from the last call to ReadvProc/WritevProc */
2496 #ifdef RXDEBUG_PACKET
2498 #endif /* RXDEBUG_PACKET */
2499 rxi_FreePackets(0, &call->iovq);
2500 MUTEX_EXIT(&call->lock);
2502 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2503 if (conn->type == RX_CLIENT_CONNECTION) {
2504 MUTEX_ENTER(&conn->conn_data_lock);
2505 conn->flags &= ~RX_CONN_BUSY;
2506 MUTEX_EXIT(&conn->conn_data_lock);
2507 MUTEX_EXIT(&conn->conn_call_lock);
2511 * Map errors to the local host's errno.h format.
2513 error = ntoh_syserr_conv(error);
2517 #if !defined(KERNEL)
2519 /* Call this routine when shutting down a server or client (especially
2520 * clients). This will allow Rx to gracefully garbage collect server
2521 * connections, and reduce the number of retries that a server might
2522 * make to a dead client.
2523 * This is not quite right, since some calls may still be ongoing and
2524 * we can't lock them to destroy them. */
2528 struct rx_connection **conn_ptr, **conn_end;
2532 if (rxinit_status == 1) {
2534 return; /* Already shutdown. */
2536 rxi_DeleteCachedConnections();
2537 if (rx_connHashTable) {
2538 MUTEX_ENTER(&rx_connHashTable_lock);
2539 for (conn_ptr = &rx_connHashTable[0], conn_end =
2540 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2542 struct rx_connection *conn, *next;
2543 for (conn = *conn_ptr; conn; conn = next) {
2545 if (conn->type == RX_CLIENT_CONNECTION) {
2546 MUTEX_ENTER(&rx_refcnt_mutex);
2548 MUTEX_EXIT(&rx_refcnt_mutex);
2549 #ifdef RX_ENABLE_LOCKS
2550 rxi_DestroyConnectionNoLock(conn);
2551 #else /* RX_ENABLE_LOCKS */
2552 rxi_DestroyConnection(conn);
2553 #endif /* RX_ENABLE_LOCKS */
2557 #ifdef RX_ENABLE_LOCKS
2558 while (rx_connCleanup_list) {
2559 struct rx_connection *conn;
2560 conn = rx_connCleanup_list;
2561 rx_connCleanup_list = rx_connCleanup_list->next;
2562 MUTEX_EXIT(&rx_connHashTable_lock);
2563 rxi_CleanupConnection(conn);
2564 MUTEX_ENTER(&rx_connHashTable_lock);
2566 MUTEX_EXIT(&rx_connHashTable_lock);
2567 #endif /* RX_ENABLE_LOCKS */
2572 afs_winsockCleanup();
2580 /* if we wakeup packet waiter too often, can get in loop with two
2581 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2583 rxi_PacketsUnWait(void)
2585 if (!rx_waitingForPackets) {
2589 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2590 return; /* still over quota */
2593 rx_waitingForPackets = 0;
2594 #ifdef RX_ENABLE_LOCKS
2595 CV_BROADCAST(&rx_waitingForPackets_cv);
2597 osi_rxWakeup(&rx_waitingForPackets);
2603 /* ------------------Internal interfaces------------------------- */
2605 /* Return this process's service structure for the
2606 * specified socket and service */
2607 static struct rx_service *
2608 rxi_FindService(osi_socket socket, u_short serviceId)
2610 struct rx_service **sp;
2611 for (sp = &rx_services[0]; *sp; sp++) {
2612 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2618 #ifdef RXDEBUG_PACKET
2619 #ifdef KDUMP_RX_LOCK
2620 static struct rx_call_rx_lock *rx_allCallsp = 0;
2622 static struct rx_call *rx_allCallsp = 0;
2624 #endif /* RXDEBUG_PACKET */
2626 /* Allocate a call structure, for the indicated channel of the
2627 * supplied connection. The mode and state of the call must be set by
2628 * the caller. Returns the call with mutex locked. */
2629 static struct rx_call *
2630 rxi_NewCall(struct rx_connection *conn, int channel)
2632 struct rx_call *call;
2633 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2634 struct rx_call *cp; /* Call pointer temp */
2635 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2636 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2638 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2640 /* Grab an existing call structure, or allocate a new one.
2641 * Existing call structures are assumed to have been left reset by
2643 MUTEX_ENTER(&rx_freeCallQueue_lock);
2645 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2647 * EXCEPT that the TQ might not yet be cleared out.
2648 * Skip over those with in-use TQs.
2651 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2652 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2658 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2659 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2660 call = queue_First(&rx_freeCallQueue, rx_call);
2661 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2663 if (rx_stats_active)
2664 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2665 MUTEX_EXIT(&rx_freeCallQueue_lock);
2666 MUTEX_ENTER(&call->lock);
2667 CLEAR_CALL_QUEUE_LOCK(call);
2668 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2669 /* Now, if TQ wasn't cleared earlier, do it now. */
2670 rxi_WaitforTQBusy(call);
2671 if (call->flags & RX_CALL_TQ_CLEARME) {
2672 rxi_ClearTransmitQueue(call, 1);
2673 /*queue_Init(&call->tq);*/
2675 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2676 /* Bind the call to its connection structure */
2678 rxi_ResetCall(call, 1);
2681 call = rxi_Alloc(sizeof(struct rx_call));
2682 #ifdef RXDEBUG_PACKET
2683 call->allNextp = rx_allCallsp;
2684 rx_allCallsp = call;
2686 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2687 #else /* RXDEBUG_PACKET */
2688 rx_atomic_inc(&rx_stats.nCallStructs);
2689 #endif /* RXDEBUG_PACKET */
2691 MUTEX_EXIT(&rx_freeCallQueue_lock);
2692 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2693 MUTEX_ENTER(&call->lock);
2694 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2695 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2696 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2698 /* Initialize once-only items */
2699 queue_Init(&call->tq);
2700 queue_Init(&call->rq);
2701 queue_Init(&call->iovq);
2702 #ifdef RXDEBUG_PACKET
2703 call->rqc = call->tqc = call->iovqc = 0;
2704 #endif /* RXDEBUG_PACKET */
2705 /* Bind the call to its connection structure (prereq for reset) */
2707 rxi_ResetCall(call, 1);
2709 call->channel = channel;
2710 call->callNumber = &conn->callNumber[channel];
2711 call->rwind = conn->rwind[channel];
2712 call->twind = conn->twind[channel];
2713 /* Note that the next expected call number is retained (in
2714 * conn->callNumber[i]), even if we reallocate the call structure
2716 conn->call[channel] = call;
2717 /* if the channel's never been used (== 0), we should start at 1, otherwise
2718 * the call number is valid from the last time this channel was used */
2719 if (*call->callNumber == 0)
2720 *call->callNumber = 1;
2725 /* A call has been inactive long enough that so we can throw away
2726 * state, including the call structure, which is placed on the call
2729 * call->lock amd rx_refcnt_mutex are held upon entry.
2730 * haveCTLock is set when called from rxi_ReapConnections.
2732 * return 1 if the call is freed, 0 if not.
2735 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2737 int channel = call->channel;
2738 struct rx_connection *conn = call->conn;
2739 u_char state = call->state;
2742 * We are setting the state to RX_STATE_RESET to
2743 * ensure that no one else will attempt to use this
2744 * call once we drop the refcnt lock. We must drop
2745 * the refcnt lock before calling rxi_ResetCall
2746 * because it cannot be held across acquiring the
2747 * freepktQ lock. NewCall does the same.
2749 call->state = RX_STATE_RESET;
2750 MUTEX_EXIT(&rx_refcnt_mutex);
2751 rxi_ResetCall(call, 0);
2753 if (MUTEX_TRYENTER(&conn->conn_call_lock))
2755 if (state == RX_STATE_DALLY || state == RX_STATE_HOLD)
2756 (*call->callNumber)++;
2758 if (call->conn->call[channel] == call)
2759 call->conn->call[channel] = 0;
2760 MUTEX_EXIT(&conn->conn_call_lock);
2763 * We couldn't obtain the conn_call_lock so we can't
2764 * disconnect the call from the connection. Set the
2765 * call state to dally so that the call can be reused.
2767 MUTEX_ENTER(&rx_refcnt_mutex);
2768 call->state = RX_STATE_DALLY;
2772 MUTEX_ENTER(&rx_freeCallQueue_lock);
2773 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2774 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2775 /* A call may be free even though its transmit queue is still in use.
2776 * Since we search the call list from head to tail, put busy calls at
2777 * the head of the list, and idle calls at the tail.
2779 if (call->flags & RX_CALL_TQ_BUSY)
2780 queue_Prepend(&rx_freeCallQueue, call);
2782 queue_Append(&rx_freeCallQueue, call);
2783 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2784 queue_Append(&rx_freeCallQueue, call);
2785 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2786 if (rx_stats_active)
2787 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2788 MUTEX_EXIT(&rx_freeCallQueue_lock);
2790 /* Destroy the connection if it was previously slated for
2791 * destruction, i.e. the Rx client code previously called
2792 * rx_DestroyConnection (client connections), or
2793 * rxi_ReapConnections called the same routine (server
2794 * connections). Only do this, however, if there are no
2795 * outstanding calls. Note that for fine grain locking, there appears
2796 * to be a deadlock in that rxi_FreeCall has a call locked and
2797 * DestroyConnectionNoLock locks each call in the conn. But note a
2798 * few lines up where we have removed this call from the conn.
2799 * If someone else destroys a connection, they either have no
2800 * call lock held or are going through this section of code.
2802 MUTEX_ENTER(&conn->conn_data_lock);
2803 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2804 MUTEX_ENTER(&rx_refcnt_mutex);
2806 MUTEX_EXIT(&rx_refcnt_mutex);
2807 MUTEX_EXIT(&conn->conn_data_lock);
2808 #ifdef RX_ENABLE_LOCKS
2810 rxi_DestroyConnectionNoLock(conn);
2812 rxi_DestroyConnection(conn);
2813 #else /* RX_ENABLE_LOCKS */
2814 rxi_DestroyConnection(conn);
2815 #endif /* RX_ENABLE_LOCKS */
2817 MUTEX_EXIT(&conn->conn_data_lock);
2819 MUTEX_ENTER(&rx_refcnt_mutex);
2823 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2824 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2827 rxi_Alloc(size_t size)
2831 if (rx_stats_active) {
2832 rx_atomic_add(&rxi_Allocsize, (int) size);
2833 rx_atomic_inc(&rxi_Alloccnt);
2837 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2838 afs_osi_Alloc_NoSleep(size);
2843 osi_Panic("rxi_Alloc error");
2849 rxi_Free(void *addr, size_t size)
2851 if (rx_stats_active) {
2852 rx_atomic_sub(&rxi_Allocsize, (int) size);
2853 rx_atomic_dec(&rxi_Alloccnt);
2855 osi_Free(addr, size);
2859 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2861 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2862 struct rx_peer *next = NULL;
2866 MUTEX_ENTER(&rx_peerHashTable_lock);
2868 peer_ptr = &rx_peerHashTable[0];
2869 peer_end = &rx_peerHashTable[rx_hashTableSize];
2872 for ( ; peer_ptr < peer_end; peer_ptr++) {
2875 for ( ; peer; peer = next) {
2877 if (host == peer->host)
2882 hashIndex = PEER_HASH(host, port);
2883 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2884 if ((peer->host == host) && (peer->port == port))
2889 MUTEX_ENTER(&rx_peerHashTable_lock);
2894 MUTEX_EXIT(&rx_peerHashTable_lock);
2896 MUTEX_ENTER(&peer->peer_lock);
2897 /* We don't handle dropping below min, so don't */
2898 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2899 peer->ifMTU=MIN(mtu, peer->ifMTU);
2900 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2901 /* if we tweaked this down, need to tune our peer MTU too */
2902 peer->MTU = MIN(peer->MTU, peer->natMTU);
2903 /* if we discovered a sub-1500 mtu, degrade */
2904 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2905 peer->maxDgramPackets = 1;
2906 /* We no longer have valid peer packet information */
2907 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2908 peer->maxPacketSize = 0;
2909 MUTEX_EXIT(&peer->peer_lock);
2911 MUTEX_ENTER(&rx_peerHashTable_lock);
2913 if (host && !port) {
2915 /* pick up where we left off */
2919 MUTEX_EXIT(&rx_peerHashTable_lock);
2922 #ifdef AFS_RXERRQ_ENV
2924 rxi_SetPeerDead(afs_uint32 host, afs_uint16 port)
2926 int hashIndex = PEER_HASH(host, port);
2927 struct rx_peer *peer;
2929 MUTEX_ENTER(&rx_peerHashTable_lock);
2931 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2932 if (peer->host == host && peer->port == port) {
2938 rx_atomic_inc(&peer->neterrs);
2941 MUTEX_EXIT(&rx_peerHashTable_lock);
2945 rxi_ProcessNetError(struct sock_extended_err *err, afs_uint32 addr, afs_uint16 port)
2947 # ifdef AFS_ADAPT_PMTU
2948 if (err->ee_errno == EMSGSIZE && err->ee_info >= 68) {
2949 rxi_SetPeerMtu(NULL, addr, port, err->ee_info - RX_IPUDP_SIZE);
2953 if (err->ee_origin == SO_EE_ORIGIN_ICMP && err->ee_type == ICMP_DEST_UNREACH) {
2954 switch (err->ee_code) {
2955 case ICMP_NET_UNREACH:
2956 case ICMP_HOST_UNREACH:
2957 case ICMP_PORT_UNREACH:
2960 rxi_SetPeerDead(addr, port);
2965 #endif /* AFS_RXERRQ_ENV */
2967 /* Find the peer process represented by the supplied (host,port)
2968 * combination. If there is no appropriate active peer structure, a
2969 * new one will be allocated and initialized
2970 * The origPeer, if set, is a pointer to a peer structure on which the
2971 * refcount will be be decremented. This is used to replace the peer
2972 * structure hanging off a connection structure */
2974 rxi_FindPeer(afs_uint32 host, u_short port,
2975 struct rx_peer *origPeer, int create)
2979 hashIndex = PEER_HASH(host, port);
2980 MUTEX_ENTER(&rx_peerHashTable_lock);
2981 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2982 if ((pp->host == host) && (pp->port == port))
2987 pp = rxi_AllocPeer(); /* This bzero's *pp */
2988 pp->host = host; /* set here or in InitPeerParams is zero */
2990 #ifdef AFS_RXERRQ_ENV
2991 rx_atomic_set(&pp->neterrs, 0);
2993 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2994 queue_Init(&pp->rpcStats);
2995 pp->next = rx_peerHashTable[hashIndex];
2996 rx_peerHashTable[hashIndex] = pp;
2997 rxi_InitPeerParams(pp);
2998 if (rx_stats_active)
2999 rx_atomic_inc(&rx_stats.nPeerStructs);
3006 origPeer->refCount--;
3007 MUTEX_EXIT(&rx_peerHashTable_lock);
3012 /* Find the connection at (host, port) started at epoch, and with the
3013 * given connection id. Creates the server connection if necessary.
3014 * The type specifies whether a client connection or a server
3015 * connection is desired. In both cases, (host, port) specify the
3016 * peer's (host, pair) pair. Client connections are not made
3017 * automatically by this routine. The parameter socket gives the
3018 * socket descriptor on which the packet was received. This is used,
3019 * in the case of server connections, to check that *new* connections
3020 * come via a valid (port, serviceId). Finally, the securityIndex
3021 * parameter must match the existing index for the connection. If a
3022 * server connection is created, it will be created using the supplied
3023 * index, if the index is valid for this service */
3024 static struct rx_connection *
3025 rxi_FindConnection(osi_socket socket, afs_uint32 host,
3026 u_short port, u_short serviceId, afs_uint32 cid,
3027 afs_uint32 epoch, int type, u_int securityIndex)
3029 int hashindex, flag, i;
3030 struct rx_connection *conn;
3031 hashindex = CONN_HASH(host, port, cid, epoch, type);
3032 MUTEX_ENTER(&rx_connHashTable_lock);
3033 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
3034 rx_connHashTable[hashindex],
3037 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
3038 && (epoch == conn->epoch)) {
3039 struct rx_peer *pp = conn->peer;
3040 if (securityIndex != conn->securityIndex) {
3041 /* this isn't supposed to happen, but someone could forge a packet
3042 * like this, and there seems to be some CM bug that makes this
3043 * happen from time to time -- in which case, the fileserver
3045 MUTEX_EXIT(&rx_connHashTable_lock);
3046 return (struct rx_connection *)0;
3048 if (pp->host == host && pp->port == port)
3050 if (type == RX_CLIENT_CONNECTION && pp->port == port)
3052 /* So what happens when it's a callback connection? */
3053 if ( /*type == RX_CLIENT_CONNECTION && */
3054 (conn->epoch & 0x80000000))
3058 /* the connection rxLastConn that was used the last time is not the
3059 ** one we are looking for now. Hence, start searching in the hash */
3061 conn = rx_connHashTable[hashindex];
3066 struct rx_service *service;
3067 if (type == RX_CLIENT_CONNECTION) {
3068 MUTEX_EXIT(&rx_connHashTable_lock);
3069 return (struct rx_connection *)0;
3071 service = rxi_FindService(socket, serviceId);
3072 if (!service || (securityIndex >= service->nSecurityObjects)
3073 || (service->securityObjects[securityIndex] == 0)) {
3074 MUTEX_EXIT(&rx_connHashTable_lock);
3075 return (struct rx_connection *)0;
3077 conn = rxi_AllocConnection(); /* This bzero's the connection */
3078 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
3079 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
3080 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
3081 conn->next = rx_connHashTable[hashindex];
3082 rx_connHashTable[hashindex] = conn;
3083 conn->peer = rxi_FindPeer(host, port, 0, 1);
3084 conn->type = RX_SERVER_CONNECTION;
3085 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
3086 conn->epoch = epoch;
3087 conn->cid = cid & RX_CIDMASK;
3088 conn->ackRate = RX_FAST_ACK_RATE;
3089 conn->service = service;
3090 conn->serviceId = serviceId;
3091 conn->securityIndex = securityIndex;
3092 conn->securityObject = service->securityObjects[securityIndex];
3093 conn->nSpecific = 0;
3094 conn->specific = NULL;
3095 rx_SetConnDeadTime(conn, service->connDeadTime);
3096 conn->idleDeadTime = service->idleDeadTime;
3097 conn->idleDeadDetection = service->idleDeadErr ? 1 : 0;
3098 for (i = 0; i < RX_MAXCALLS; i++) {
3099 conn->twind[i] = rx_initSendWindow;
3100 conn->rwind[i] = rx_initReceiveWindow;
3102 /* Notify security object of the new connection */
3103 RXS_NewConnection(conn->securityObject, conn);
3104 /* XXXX Connection timeout? */
3105 if (service->newConnProc)
3106 (*service->newConnProc) (conn);
3107 if (rx_stats_active)
3108 rx_atomic_inc(&rx_stats.nServerConns);
3111 MUTEX_ENTER(&rx_refcnt_mutex);
3113 MUTEX_EXIT(&rx_refcnt_mutex);
3115 rxLastConn = conn; /* store this connection as the last conn used */
3116 MUTEX_EXIT(&rx_connHashTable_lock);
3121 * Timeout a call on a busy call channel if appropriate.
3123 * @param[in] call The busy call.
3125 * @pre 'call' is marked as busy (namely,
3126 * call->conn->lastBusy[call->channel] != 0)
3128 * @pre call->lock is held
3129 * @pre rxi_busyChannelError is nonzero
3131 * @note call->lock is dropped and reacquired
3134 rxi_CheckBusy(struct rx_call *call)
3136 struct rx_connection *conn = call->conn;
3137 int channel = call->channel;
3138 int freechannel = 0;
3140 afs_uint32 callNumber;
3142 MUTEX_EXIT(&call->lock);
3144 MUTEX_ENTER(&conn->conn_call_lock);
3145 callNumber = *call->callNumber;
3147 /* Are there any other call slots on this conn that we should try? Look for
3148 * slots that are empty and are either non-busy, or were marked as busy
3149 * longer than conn->secondsUntilDead seconds before this call started. */
3151 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3153 /* only look at channels that aren't us */
3157 if (conn->lastBusy[i]) {
3158 /* if this channel looked busy too recently, don't look at it */
3159 if (conn->lastBusy[i] >= call->startTime.sec) {
3162 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3167 if (conn->call[i]) {
3168 struct rx_call *tcall = conn->call[i];
3169 MUTEX_ENTER(&tcall->lock);
3170 if (tcall->state == RX_STATE_DALLY) {
3173 MUTEX_EXIT(&tcall->lock);
3179 MUTEX_ENTER(&call->lock);
3181 /* Since the call->lock and conn->conn_call_lock have been released it is
3182 * possible that (1) the call may no longer be busy and/or (2) the call may
3183 * have been reused by another waiting thread. Therefore, we must confirm
3184 * that the call state has not changed when deciding whether or not to
3185 * force this application thread to retry by forcing a Timeout error. */
3187 if (freechannel && *call->callNumber == callNumber &&
3188 (call->flags & RX_CALL_PEER_BUSY)) {
3189 /* Since 'freechannel' is set, there exists another channel in this
3190 * rx_conn that the application thread might be able to use. We know
3191 * that we have the correct call since callNumber is unchanged, and we
3192 * know that the call is still busy. So, set the call error state to
3193 * rxi_busyChannelError so the application can retry the request,
3194 * presumably on a less-busy call channel. */
3196 rxi_CallError(call, RX_CALL_BUSY);
3198 MUTEX_EXIT(&conn->conn_call_lock);
3201 /* There are two packet tracing routines available for testing and monitoring
3202 * Rx. One is called just after every packet is received and the other is
3203 * called just before every packet is sent. Received packets, have had their
3204 * headers decoded, and packets to be sent have not yet had their headers
3205 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3206 * containing the network address. Both can be modified. The return value, if
3207 * non-zero, indicates that the packet should be dropped. */
3209 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3210 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3212 /* A packet has been received off the interface. Np is the packet, socket is
3213 * the socket number it was received from (useful in determining which service
3214 * this packet corresponds to), and (host, port) reflect the host,port of the
3215 * sender. This call returns the packet to the caller if it is finished with
3216 * it, rather than de-allocating it, just as a small performance hack */
3219 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3220 afs_uint32 host, u_short port, int *tnop,
3221 struct rx_call **newcallp)
3223 struct rx_call *call;
3224 struct rx_connection *conn;
3226 afs_uint32 currentCallNumber;
3231 struct rx_packet *tnp;
3234 /* We don't print out the packet until now because (1) the time may not be
3235 * accurate enough until now in the lwp implementation (rx_Listener only gets
3236 * the time after the packet is read) and (2) from a protocol point of view,
3237 * this is the first time the packet has been seen */
3238 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3239 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3240 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3241 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3242 np->header.epoch, np->header.cid, np->header.callNumber,
3243 np->header.seq, np->header.flags, np));
3246 /* Account for connectionless packets */
3247 if (rx_stats_active &&
3248 ((np->header.type == RX_PACKET_TYPE_VERSION) ||
3249 (np->header.type == RX_PACKET_TYPE_DEBUG))) {
3250 struct rx_peer *peer;
3252 /* Try to look up the peer structure, but don't create one */
3253 peer = rxi_FindPeer(host, port, 0, 0);
3255 /* Since this may not be associated with a connection, it may have
3256 * no refCount, meaning we could race with ReapConnections
3259 if (peer && (peer->refCount > 0)) {
3260 #ifdef AFS_RXERRQ_ENV
3261 if (rx_atomic_read(&peer->neterrs)) {
3262 rx_atomic_set(&peer->neterrs, 0);
3265 MUTEX_ENTER(&peer->peer_lock);
3266 peer->bytesReceived += np->length;
3267 MUTEX_EXIT(&peer->peer_lock);
3271 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3272 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3275 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3276 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3279 /* If an input tracer function is defined, call it with the packet and
3280 * network address. Note this function may modify its arguments. */
3281 if (rx_justReceived) {
3282 struct sockaddr_in addr;
3284 addr.sin_family = AF_INET;
3285 addr.sin_port = port;
3286 addr.sin_addr.s_addr = host;
3287 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3288 addr.sin_len = sizeof(addr);
3289 #endif /* AFS_OSF_ENV */
3290 drop = (*rx_justReceived) (np, &addr);
3291 /* drop packet if return value is non-zero */
3294 port = addr.sin_port; /* in case fcn changed addr */
3295 host = addr.sin_addr.s_addr;
3299 /* If packet was not sent by the client, then *we* must be the client */
3300 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3301 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3303 /* Find the connection (or fabricate one, if we're the server & if
3304 * necessary) associated with this packet */
3306 rxi_FindConnection(socket, host, port, np->header.serviceId,
3307 np->header.cid, np->header.epoch, type,
3308 np->header.securityIndex);
3310 /* To avoid having 2 connections just abort at each other,
3311 don't abort an abort. */
3313 if (np->header.type != RX_PACKET_TYPE_ABORT)
3314 rxi_SendRawAbort(socket, host, port, RX_INVALID_OPERATION,
3319 #ifdef AFS_RXERRQ_ENV
3320 if (rx_atomic_read(&conn->peer->neterrs)) {
3321 rx_atomic_set(&conn->peer->neterrs, 0);
3325 /* If we're doing statistics, then account for the incoming packet */
3326 if (rx_stats_active) {
3327 MUTEX_ENTER(&conn->peer->peer_lock);
3328 conn->peer->bytesReceived += np->length;
3329 MUTEX_EXIT(&conn->peer->peer_lock);
3332 /* If the connection is in an error state, send an abort packet and ignore
3333 * the incoming packet */
3335 /* Don't respond to an abort packet--we don't want loops! */
3336 MUTEX_ENTER(&conn->conn_data_lock);
3337 if (np->header.type != RX_PACKET_TYPE_ABORT)
3338 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3339 putConnection(conn);
3340 MUTEX_EXIT(&conn->conn_data_lock);
3344 /* Check for connection-only requests (i.e. not call specific). */
3345 if (np->header.callNumber == 0) {
3346 switch (np->header.type) {
3347 case RX_PACKET_TYPE_ABORT: {
3348 /* What if the supplied error is zero? */
3349 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3350 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3351 rxi_ConnectionError(conn, errcode);
3352 putConnection(conn);
3355 case RX_PACKET_TYPE_CHALLENGE:
3356 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3357 putConnection(conn);
3359 case RX_PACKET_TYPE_RESPONSE:
3360 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3361 putConnection(conn);
3363 case RX_PACKET_TYPE_PARAMS:
3364 case RX_PACKET_TYPE_PARAMS + 1:
3365 case RX_PACKET_TYPE_PARAMS + 2:
3366 /* ignore these packet types for now */
3367 putConnection(conn);
3371 /* Should not reach here, unless the peer is broken: send an
3373 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3374 MUTEX_ENTER(&conn->conn_data_lock);
3375 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3376 putConnection(conn);
3377 MUTEX_EXIT(&conn->conn_data_lock);
3382 channel = np->header.cid & RX_CHANNELMASK;
3383 MUTEX_ENTER(&conn->conn_call_lock);
3384 call = conn->call[channel];
3387 MUTEX_ENTER(&call->lock);
3388 currentCallNumber = conn->callNumber[channel];
3389 MUTEX_EXIT(&conn->conn_call_lock);
3390 } else if (type == RX_SERVER_CONNECTION) { /* No call allocated */
3391 call = conn->call[channel];
3393 MUTEX_ENTER(&call->lock);
3394 currentCallNumber = conn->callNumber[channel];
3395 MUTEX_EXIT(&conn->conn_call_lock);
3397 call = rxi_NewCall(conn, channel); /* returns locked call */
3398 *call->callNumber = currentCallNumber = np->header.callNumber;
3399 MUTEX_EXIT(&conn->conn_call_lock);
3401 if (np->header.callNumber == 0)
3402 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3403 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3404 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3405 np->header.flags, np, np->length));
3407 call->state = RX_STATE_PRECALL;
3408 clock_GetTime(&call->queueTime);
3409 call->bytesSent = 0;
3410 call->bytesRcvd = 0;
3412 * If the number of queued calls exceeds the overload
3413 * threshold then abort this call.
3415 if ((rx_BusyThreshold > 0) &&
3416 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3417 struct rx_packet *tp;
3419 rxi_CallError(call, rx_BusyError);
3420 tp = rxi_SendCallAbort(call, np, 1, 0);
3421 MUTEX_EXIT(&call->lock);
3422 putConnection(conn);
3423 if (rx_stats_active)
3424 rx_atomic_inc(&rx_stats.nBusies);
3427 rxi_KeepAliveOn(call);
3429 } else { /* RX_CLIENT_CONNECTION and No call allocated */
3430 /* This packet can't be for this call. If the new call address is
3431 * 0 then no call is running on this channel. If there is a call
3432 * then, since this is a client connection we're getting data for
3433 * it must be for the previous call.
3435 MUTEX_EXIT(&conn->conn_call_lock);
3436 if (rx_stats_active)
3437 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3438 putConnection(conn);
3442 /* There is a non-NULL locked call at this point */
3443 if (type == RX_SERVER_CONNECTION) { /* We're the server */
3444 if (np->header.callNumber < currentCallNumber) {
3445 MUTEX_EXIT(&call->lock);
3446 if (rx_stats_active)
3447 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3448 putConnection(conn);
3450 } else if (np->header.callNumber != currentCallNumber) {
3451 /* Wait until the transmit queue is idle before deciding
3452 * whether to reset the current call. Chances are that the
3453 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3456 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3457 if (call->state == RX_STATE_ACTIVE) {
3458 rxi_WaitforTQBusy(call);
3460 * If we entered error state while waiting,
3461 * must call rxi_CallError to permit rxi_ResetCall
3462 * to processed when the tqWaiter count hits zero.
3465 rxi_CallError(call, call->error);
3466 MUTEX_EXIT(&call->lock);
3467 putConnection(conn);
3471 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3472 /* If the new call cannot be taken right now send a busy and set
3473 * the error condition in this call, so that it terminates as
3474 * quickly as possible */
3475 if (call->state == RX_STATE_ACTIVE) {
3476 struct rx_packet *tp;
3478 rxi_CallError(call, RX_CALL_DEAD);
3479 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3481 MUTEX_EXIT(&call->lock);
3482 putConnection(conn);
3485 rxi_ResetCall(call, 0);
3487 * The conn_call_lock is not held but no one else should be
3488 * using this call channel while we are processing this incoming
3489 * packet. This assignment should be safe.
3491 *call->callNumber = np->header.callNumber;
3493 if (np->header.callNumber == 0)
3494 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3495 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3496 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3497 np->header.flags, np, np->length));
3499 call->state = RX_STATE_PRECALL;
3500 clock_GetTime(&call->queueTime);
3501 call->bytesSent = 0;
3502 call->bytesRcvd = 0;
3504 * If the number of queued calls exceeds the overload
3505 * threshold then abort this call.
3507 if ((rx_BusyThreshold > 0) &&
3508 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3509 struct rx_packet *tp;
3511 rxi_CallError(call, rx_BusyError);
3512 tp = rxi_SendCallAbort(call, np, 1, 0);
3513 MUTEX_EXIT(&call->lock);
3514 putConnection(conn);
3515 if (rx_stats_active)
3516 rx_atomic_inc(&rx_stats.nBusies);
3519 rxi_KeepAliveOn(call);
3521 /* Continuing call; do nothing here. */
3523 } else { /* we're the client */
3524 /* Ignore all incoming acknowledgements for calls in DALLY state */
3525 if ((call->state == RX_STATE_DALLY)
3526 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3527 if (rx_stats_active)
3528 rx_atomic_inc(&rx_stats.ignorePacketDally);
3529 MUTEX_EXIT(&call->lock);
3530 putConnection(conn);
3534 /* Ignore anything that's not relevant to the current call. If there
3535 * isn't a current call, then no packet is relevant. */
3536 if (np->header.callNumber != currentCallNumber) {
3537 if (rx_stats_active)
3538 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3539 MUTEX_EXIT(&call->lock);
3540 putConnection(conn);
3543 /* If the service security object index stamped in the packet does not
3544 * match the connection's security index, ignore the packet */
3545 if (np->header.securityIndex != conn->securityIndex) {
3546 MUTEX_EXIT(&call->lock);
3547 putConnection(conn);
3551 /* If we're receiving the response, then all transmit packets are
3552 * implicitly acknowledged. Get rid of them. */
3553 if (np->header.type == RX_PACKET_TYPE_DATA) {
3554 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3555 /* XXX Hack. Because we must release the global rx lock when
3556 * sending packets (osi_NetSend) we drop all acks while we're
3557 * traversing the tq in rxi_Start sending packets out because
3558 * packets may move to the freePacketQueue as result of being here!
3559 * So we drop these packets until we're safely out of the
3560 * traversing. Really ugly!
3561 * For fine grain RX locking, we set the acked field in the
3562 * packets and let rxi_Start remove them from the transmit queue.
3564 if (call->flags & RX_CALL_TQ_BUSY) {
3565 #ifdef RX_ENABLE_LOCKS
3566 rxi_SetAcksInTransmitQueue(call);
3568 putConnection(conn);
3569 return np; /* xmitting; drop packet */
3572 rxi_ClearTransmitQueue(call, 0);
3574 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3575 rxi_ClearTransmitQueue(call, 0);
3576 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3578 if (np->header.type == RX_PACKET_TYPE_ACK) {
3579 /* now check to see if this is an ack packet acknowledging that the
3580 * server actually *lost* some hard-acked data. If this happens we
3581 * ignore this packet, as it may indicate that the server restarted in
3582 * the middle of a call. It is also possible that this is an old ack
3583 * packet. We don't abort the connection in this case, because this
3584 * *might* just be an old ack packet. The right way to detect a server
3585 * restart in the midst of a call is to notice that the server epoch
3587 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3588 * XXX unacknowledged. I think that this is off-by-one, but
3589 * XXX I don't dare change it just yet, since it will
3590 * XXX interact badly with the server-restart detection
3591 * XXX code in receiveackpacket. */
3592 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3593 if (rx_stats_active)
3594 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3595 MUTEX_EXIT(&call->lock);
3596 putConnection(conn);
3600 } /* else not a data packet */
3603 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3604 /* Set remote user defined status from packet */
3605 call->remoteStatus = np->header.userStatus;
3607 /* Now do packet type-specific processing */
3608 switch (np->header.type) {
3609 case RX_PACKET_TYPE_DATA:
3610 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3613 case RX_PACKET_TYPE_ACK:
3614 /* Respond immediately to ack packets requesting acknowledgement
3616 if (np->header.flags & RX_REQUEST_ACK) {
3618 (void)rxi_SendCallAbort(call, 0, 1, 0);
3620 (void)rxi_SendAck(call, 0, np->header.serial,
3621 RX_ACK_PING_RESPONSE, 1);
3623 np = rxi_ReceiveAckPacket(call, np, 1);
3625 case RX_PACKET_TYPE_ABORT: {
3626 /* An abort packet: reset the call, passing the error up to the user. */
3627 /* What if error is zero? */
3628 /* What if the error is -1? the application will treat it as a timeout. */
3629 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3630 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3631 rxi_CallError(call, errdata);
3632 MUTEX_EXIT(&call->lock);
3633 putConnection(conn);
3634 return np; /* xmitting; drop packet */
3636 case RX_PACKET_TYPE_BUSY: {
3637 struct clock busyTime;
3639 clock_GetTime(&busyTime);
3641 MUTEX_EXIT(&call->lock);
3643 MUTEX_ENTER(&conn->conn_call_lock);
3644 MUTEX_ENTER(&call->lock);
3645 conn->lastBusy[call->channel] = busyTime.sec;
3646 call->flags |= RX_CALL_PEER_BUSY;
3647 MUTEX_EXIT(&call->lock);
3648 MUTEX_EXIT(&conn->conn_call_lock);
3650 putConnection(conn);
3654 case RX_PACKET_TYPE_ACKALL:
3655 /* All packets acknowledged, so we can drop all packets previously
3656 * readied for sending */
3657 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3658 /* XXX Hack. We because we can't release the global rx lock when
3659 * sending packets (osi_NetSend) we drop all ack pkts while we're
3660 * traversing the tq in rxi_Start sending packets out because
3661 * packets may move to the freePacketQueue as result of being
3662 * here! So we drop these packets until we're safely out of the
3663 * traversing. Really ugly!
3664 * For fine grain RX locking, we set the acked field in the packets
3665 * and let rxi_Start remove the packets from the transmit queue.
3667 if (call->flags & RX_CALL_TQ_BUSY) {
3668 #ifdef RX_ENABLE_LOCKS
3669 rxi_SetAcksInTransmitQueue(call);
3671 #else /* RX_ENABLE_LOCKS */
3672 MUTEX_EXIT(&call->lock);
3673 putConnection(conn);
3674 return np; /* xmitting; drop packet */
3675 #endif /* RX_ENABLE_LOCKS */
3677 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3678 rxi_ClearTransmitQueue(call, 0);
3681 /* Should not reach here, unless the peer is broken: send an abort
3683 rxi_CallError(call, RX_PROTOCOL_ERROR);
3684 np = rxi_SendCallAbort(call, np, 1, 0);
3687 /* Note when this last legitimate packet was received, for keep-alive
3688 * processing. Note, we delay getting the time until now in the hope that
3689 * the packet will be delivered to the user before any get time is required
3690 * (if not, then the time won't actually be re-evaluated here). */
3691 call->lastReceiveTime = clock_Sec();
3692 /* we've received a legit packet, so the channel is not busy */
3693 call->flags &= ~RX_CALL_PEER_BUSY;
3694 MUTEX_EXIT(&call->lock);
3695 putConnection(conn);
3699 /* return true if this is an "interesting" connection from the point of view
3700 of someone trying to debug the system */
3702 rxi_IsConnInteresting(struct rx_connection *aconn)
3705 struct rx_call *tcall;
3707 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3710 for (i = 0; i < RX_MAXCALLS; i++) {
3711 tcall = aconn->call[i];
3713 if ((tcall->state == RX_STATE_PRECALL)
3714 || (tcall->state == RX_STATE_ACTIVE))
3716 if ((tcall->mode == RX_MODE_SENDING)
3717 || (tcall->mode == RX_MODE_RECEIVING))
3725 /* if this is one of the last few packets AND it wouldn't be used by the
3726 receiving call to immediately satisfy a read request, then drop it on
3727 the floor, since accepting it might prevent a lock-holding thread from
3728 making progress in its reading. If a call has been cleared while in
3729 the precall state then ignore all subsequent packets until the call
3730 is assigned to a thread. */
3733 TooLow(struct rx_packet *ap, struct rx_call *acall)
3737 MUTEX_ENTER(&rx_quota_mutex);
3738 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3739 && (acall->state == RX_STATE_PRECALL))
3740 || ((rx_nFreePackets < rxi_dataQuota + 2)
3741 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3742 && (acall->flags & RX_CALL_READER_WAIT)))) {
3745 MUTEX_EXIT(&rx_quota_mutex);
3751 * Clear the attach wait flag on a connection and proceed.
3753 * Any processing waiting for a connection to be attached should be
3754 * unblocked. We clear the flag and do any other needed tasks.
3757 * the conn to unmark waiting for attach
3759 * @pre conn's conn_data_lock must be locked before calling this function
3763 rxi_ConnClearAttachWait(struct rx_connection *conn)
3765 /* Indicate that rxi_CheckReachEvent is no longer running by
3766 * clearing the flag. Must be atomic under conn_data_lock to
3767 * avoid a new call slipping by: rxi_CheckConnReach holds
3768 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3770 conn->flags &= ~RX_CONN_ATTACHWAIT;
3771 if (conn->flags & RX_CONN_NAT_PING) {
3772 conn->flags &= ~RX_CONN_NAT_PING;
3773 rxi_ScheduleNatKeepAliveEvent(conn);
3778 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3780 struct rx_connection *conn = arg1;
3781 struct rx_call *acall = arg2;
3782 struct rx_call *call = acall;
3783 struct clock when, now;
3786 MUTEX_ENTER(&conn->conn_data_lock);
3789 rxevent_Put(conn->checkReachEvent);
3790 conn->checkReachEvent = NULL;
3793 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3795 putConnection(conn);
3797 MUTEX_EXIT(&conn->conn_data_lock);
3801 MUTEX_ENTER(&conn->conn_call_lock);
3802 MUTEX_ENTER(&conn->conn_data_lock);
3803 for (i = 0; i < RX_MAXCALLS; i++) {
3804 struct rx_call *tc = conn->call[i];
3805 if (tc && tc->state == RX_STATE_PRECALL) {
3811 rxi_ConnClearAttachWait(conn);
3812 MUTEX_EXIT(&conn->conn_data_lock);
3813 MUTEX_EXIT(&conn->conn_call_lock);
3818 MUTEX_ENTER(&call->lock);
3819 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3821 MUTEX_EXIT(&call->lock);
3823 clock_GetTime(&now);
3825 when.sec += RX_CHECKREACH_TIMEOUT;
3826 MUTEX_ENTER(&conn->conn_data_lock);
3827 if (!conn->checkReachEvent) {
3828 MUTEX_ENTER(&rx_refcnt_mutex);
3830 MUTEX_EXIT(&rx_refcnt_mutex);
3831 conn->checkReachEvent = rxevent_Post(&when, &now,
3832 rxi_CheckReachEvent, conn,
3835 MUTEX_EXIT(&conn->conn_data_lock);
3841 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3843 struct rx_service *service = conn->service;
3844 struct rx_peer *peer = conn->peer;
3845 afs_uint32 now, lastReach;
3847 if (service->checkReach == 0)
3851 MUTEX_ENTER(&peer->peer_lock);
3852 lastReach = peer->lastReachTime;
3853 MUTEX_EXIT(&peer->peer_lock);
3854 if (now - lastReach < RX_CHECKREACH_TTL)
3857 MUTEX_ENTER(&conn->conn_data_lock);
3858 if (conn->flags & RX_CONN_ATTACHWAIT) {
3859 MUTEX_EXIT(&conn->conn_data_lock);
3862 conn->flags |= RX_CONN_ATTACHWAIT;
3863 MUTEX_EXIT(&conn->conn_data_lock);
3864 if (!conn->checkReachEvent)
3865 rxi_CheckReachEvent(NULL, conn, call, 0);
3870 /* try to attach call, if authentication is complete */
3872 TryAttach(struct rx_call *acall, osi_socket socket,
3873 int *tnop, struct rx_call **newcallp,
3876 struct rx_connection *conn = acall->conn;
3878 if (conn->type == RX_SERVER_CONNECTION
3879 && acall->state == RX_STATE_PRECALL) {
3880 /* Don't attach until we have any req'd. authentication. */
3881 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3882 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3883 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3884 /* Note: this does not necessarily succeed; there
3885 * may not any proc available
3888 rxi_ChallengeOn(acall->conn);
3893 /* A data packet has been received off the interface. This packet is
3894 * appropriate to the call (the call is in the right state, etc.). This
3895 * routine can return a packet to the caller, for re-use */
3897 static struct rx_packet *
3898 rxi_ReceiveDataPacket(struct rx_call *call,
3899 struct rx_packet *np, int istack,
3900 osi_socket socket, afs_uint32 host, u_short port,
3901 int *tnop, struct rx_call **newcallp)
3903 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3908 afs_uint32 serial=0, flags=0;
3910 struct rx_packet *tnp;
3911 if (rx_stats_active)
3912 rx_atomic_inc(&rx_stats.dataPacketsRead);
3915 /* If there are no packet buffers, drop this new packet, unless we can find
3916 * packet buffers from inactive calls */
3918 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3919 MUTEX_ENTER(&rx_freePktQ_lock);
3920 rxi_NeedMorePackets = TRUE;
3921 MUTEX_EXIT(&rx_freePktQ_lock);
3922 if (rx_stats_active)
3923 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3924 rxi_calltrace(RX_TRACE_DROP, call);
3925 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3926 /* We used to clear the receive queue here, in an attempt to free
3927 * packets. However this is unsafe if the queue has received a
3928 * soft ACK for the final packet */
3929 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3935 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3936 * packet is one of several packets transmitted as a single
3937 * datagram. Do not send any soft or hard acks until all packets
3938 * in a jumbogram have been processed. Send negative acks right away.
3940 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3941 /* tnp is non-null when there are more packets in the
3942 * current jumbo gram */
3949 seq = np->header.seq;
3950 serial = np->header.serial;
3951 flags = np->header.flags;
3953 /* If the call is in an error state, send an abort message */
3955 return rxi_SendCallAbort(call, np, istack, 0);
3957 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3958 * AFS 3.5 jumbogram. */
3959 if (flags & RX_JUMBO_PACKET) {
3960 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3965 if (np->header.spare != 0) {
3966 MUTEX_ENTER(&call->conn->conn_data_lock);
3967 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3968 MUTEX_EXIT(&call->conn->conn_data_lock);
3971 /* The usual case is that this is the expected next packet */
3972 if (seq == call->rnext) {
3974 /* Check to make sure it is not a duplicate of one already queued */
3975 if (queue_IsNotEmpty(&call->rq)
3976 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3977 if (rx_stats_active)
3978 rx_atomic_inc(&rx_stats.dupPacketsRead);
3979 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3980 rxevent_Cancel(&call->delayedAckEvent, call,
3981 RX_CALL_REFCOUNT_DELAY);
3982 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3988 /* It's the next packet. Stick it on the receive queue
3989 * for this call. Set newPackets to make sure we wake
3990 * the reader once all packets have been processed */
3991 #ifdef RX_TRACK_PACKETS
3992 np->flags |= RX_PKTFLAG_RQ;
3994 queue_Prepend(&call->rq, np);
3995 #ifdef RXDEBUG_PACKET
3997 #endif /* RXDEBUG_PACKET */
3999 np = NULL; /* We can't use this anymore */
4002 /* If an ack is requested then set a flag to make sure we
4003 * send an acknowledgement for this packet */
4004 if (flags & RX_REQUEST_ACK) {
4005 ackNeeded = RX_ACK_REQUESTED;
4008 /* Keep track of whether we have received the last packet */
4009 if (flags & RX_LAST_PACKET) {
4010 call->flags |= RX_CALL_HAVE_LAST;
4014 /* Check whether we have all of the packets for this call */
4015 if (call->flags & RX_CALL_HAVE_LAST) {
4016 afs_uint32 tseq; /* temporary sequence number */
4017 struct rx_packet *tp; /* Temporary packet pointer */
4018 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
4020 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
4021 if (tseq != tp->header.seq)
4023 if (tp->header.flags & RX_LAST_PACKET) {
4024 call->flags |= RX_CALL_RECEIVE_DONE;
4031 /* Provide asynchronous notification for those who want it
4032 * (e.g. multi rx) */
4033 if (call->arrivalProc) {
4034 (*call->arrivalProc) (call, call->arrivalProcHandle,
4035 call->arrivalProcArg);
4036 call->arrivalProc = (void (*)())0;
4039 /* Update last packet received */
4042 /* If there is no server process serving this call, grab
4043 * one, if available. We only need to do this once. If a
4044 * server thread is available, this thread becomes a server
4045 * thread and the server thread becomes a listener thread. */
4047 TryAttach(call, socket, tnop, newcallp, 0);
4050 /* This is not the expected next packet. */
4052 /* Determine whether this is a new or old packet, and if it's
4053 * a new one, whether it fits into the current receive window.
4054 * Also figure out whether the packet was delivered in sequence.
4055 * We use the prev variable to determine whether the new packet
4056 * is the successor of its immediate predecessor in the
4057 * receive queue, and the missing flag to determine whether
4058 * any of this packets predecessors are missing. */
4060 afs_uint32 prev; /* "Previous packet" sequence number */
4061 struct rx_packet *tp; /* Temporary packet pointer */
4062 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
4063 int missing; /* Are any predecessors missing? */
4065 /* If the new packet's sequence number has been sent to the
4066 * application already, then this is a duplicate */
4067 if (seq < call->rnext) {
4068 if (rx_stats_active)
4069 rx_atomic_inc(&rx_stats.dupPacketsRead);
4070 rxevent_Cancel(&call->delayedAckEvent, call,
4071 RX_CALL_REFCOUNT_DELAY);
4072 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
4078 /* If the sequence number is greater than what can be
4079 * accomodated by the current window, then send a negative
4080 * acknowledge and drop the packet */
4081 if ((call->rnext + call->rwind) <= seq) {
4082 rxevent_Cancel(&call->delayedAckEvent, call,
4083 RX_CALL_REFCOUNT_DELAY);
4084 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
4091 /* Look for the packet in the queue of old received packets */
4092 for (prev = call->rnext - 1, missing =
4093 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
4094 /*Check for duplicate packet */
4095 if (seq == tp->header.seq) {
4096 if (rx_stats_active)
4097 rx_atomic_inc(&rx_stats.dupPacketsRead);
4098 rxevent_Cancel(&call->delayedAckEvent, call,
4099 RX_CALL_REFCOUNT_DELAY);
4100 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
4106 /* If we find a higher sequence packet, break out and
4107 * insert the new packet here. */
4108 if (seq < tp->header.seq)
4110 /* Check for missing packet */
4111 if (tp->header.seq != prev + 1) {
4115 prev = tp->header.seq;
4118 /* Keep track of whether we have received the last packet. */
4119 if (flags & RX_LAST_PACKET) {
4120 call->flags |= RX_CALL_HAVE_LAST;
4123 /* It's within the window: add it to the the receive queue.
4124 * tp is left by the previous loop either pointing at the
4125 * packet before which to insert the new packet, or at the
4126 * queue head if the queue is empty or the packet should be
4128 #ifdef RX_TRACK_PACKETS
4129 np->flags |= RX_PKTFLAG_RQ;
4131 #ifdef RXDEBUG_PACKET
4133 #endif /* RXDEBUG_PACKET */
4134 queue_InsertBefore(tp, np);
4138 /* Check whether we have all of the packets for this call */
4139 if ((call->flags & RX_CALL_HAVE_LAST)
4140 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4141 afs_uint32 tseq; /* temporary sequence number */
4144 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
4145 if (tseq != tp->header.seq)
4147 if (tp->header.flags & RX_LAST_PACKET) {
4148 call->flags |= RX_CALL_RECEIVE_DONE;
4155 /* We need to send an ack of the packet is out of sequence,
4156 * or if an ack was requested by the peer. */
4157 if (seq != prev + 1 || missing) {
4158 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4159 } else if (flags & RX_REQUEST_ACK) {
4160 ackNeeded = RX_ACK_REQUESTED;
4163 /* Acknowledge the last packet for each call */
4164 if (flags & RX_LAST_PACKET) {
4175 * If the receiver is waiting for an iovec, fill the iovec
4176 * using the data from the receive queue */
4177 if (call->flags & RX_CALL_IOVEC_WAIT) {
4178 didHardAck = rxi_FillReadVec(call, serial);
4179 /* the call may have been aborted */
4188 /* Wakeup the reader if any */
4189 if ((call->flags & RX_CALL_READER_WAIT)
4190 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4191 || (call->iovNext >= call->iovMax)
4192 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4193 call->flags &= ~RX_CALL_READER_WAIT;
4194 #ifdef RX_ENABLE_LOCKS
4195 CV_BROADCAST(&call->cv_rq);
4197 osi_rxWakeup(&call->rq);
4203 * Send an ack when requested by the peer, or once every
4204 * rxi_SoftAckRate packets until the last packet has been
4205 * received. Always send a soft ack for the last packet in
4206 * the server's reply. */
4208 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4209 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4210 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4211 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4212 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4213 } else if (call->nSoftAcks) {
4214 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4215 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4217 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4218 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4219 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4226 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4228 struct rx_peer *peer = conn->peer;
4230 MUTEX_ENTER(&peer->peer_lock);
4231 peer->lastReachTime = clock_Sec();
4232 MUTEX_EXIT(&peer->peer_lock);
4234 MUTEX_ENTER(&conn->conn_data_lock);
4235 if (conn->flags & RX_CONN_ATTACHWAIT) {
4238 rxi_ConnClearAttachWait(conn);
4239 MUTEX_EXIT(&conn->conn_data_lock);
4241 for (i = 0; i < RX_MAXCALLS; i++) {
4242 struct rx_call *call = conn->call[i];
4245 MUTEX_ENTER(&call->lock);
4246 /* tnop can be null if newcallp is null */
4247 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4249 MUTEX_EXIT(&call->lock);
4253 MUTEX_EXIT(&conn->conn_data_lock);
4256 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4258 rx_ack_reason(int reason)
4261 case RX_ACK_REQUESTED:
4263 case RX_ACK_DUPLICATE:
4265 case RX_ACK_OUT_OF_SEQUENCE:
4267 case RX_ACK_EXCEEDS_WINDOW:
4269 case RX_ACK_NOSPACE:
4273 case RX_ACK_PING_RESPONSE:
4286 /* The real smarts of the whole thing. */
4287 static struct rx_packet *
4288 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4291 struct rx_ackPacket *ap;
4293 struct rx_packet *tp;
4294 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
4295 struct rx_connection *conn = call->conn;
4296 struct rx_peer *peer = conn->peer;
4297 struct clock now; /* Current time, for RTT calculations */
4305 int newAckCount = 0;
4306 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4307 int pktsize = 0; /* Set if we need to update the peer mtu */
4308 int conn_data_locked = 0;
4310 if (rx_stats_active)
4311 rx_atomic_inc(&rx_stats.ackPacketsRead);
4312 ap = (struct rx_ackPacket *)rx_DataOf(np);
4313 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4315 return np; /* truncated ack packet */
4317 /* depends on ack packet struct */
4318 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4319 first = ntohl(ap->firstPacket);
4320 prev = ntohl(ap->previousPacket);
4321 serial = ntohl(ap->serial);
4324 * Ignore ack packets received out of order while protecting
4325 * against peers that set the previousPacket field to a packet
4326 * serial number instead of a sequence number.
4328 if (first < call->tfirst ||
4329 (first == call->tfirst && prev < call->tprev && prev < call->tfirst
4336 if (np->header.flags & RX_SLOW_START_OK) {
4337 call->flags |= RX_CALL_SLOW_START_OK;
4340 if (ap->reason == RX_ACK_PING_RESPONSE)
4341 rxi_UpdatePeerReach(conn, call);
4343 if (conn->lastPacketSizeSeq) {
4344 MUTEX_ENTER(&conn->conn_data_lock);
4345 conn_data_locked = 1;
4346 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4347 pktsize = conn->lastPacketSize;
4348 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4351 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4352 if (!conn_data_locked) {
4353 MUTEX_ENTER(&conn->conn_data_lock);
4354 conn_data_locked = 1;
4356 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4357 /* process mtu ping ack */
4358 pktsize = conn->lastPingSize;
4359 conn->lastPingSizeSer = conn->lastPingSize = 0;
4363 if (conn_data_locked) {
4364 MUTEX_EXIT(&conn->conn_data_lock);
4365 conn_data_locked = 0;
4369 if (rxdebug_active) {
4373 len = _snprintf(msg, sizeof(msg),
4374 "tid[%d] RACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4375 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4376 ntohl(ap->serial), ntohl(ap->previousPacket),
4377 (unsigned int)np->header.seq, ntohl(ap->firstPacket),
4378 ap->nAcks, ntohs(ap->bufferSpace) );
4382 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4383 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4387 OutputDebugString(msg);
4389 #else /* AFS_NT40_ENV */
4392 "RACK: reason %x previous %u seq %u serial %u first %u",
4393 ap->reason, ntohl(ap->previousPacket),
4394 (unsigned int)np->header.seq, (unsigned int)serial,
4395 ntohl(ap->firstPacket));
4398 for (offset = 0; offset < nAcks; offset++)
4399 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4404 #endif /* AFS_NT40_ENV */
4407 MUTEX_ENTER(&peer->peer_lock);
4410 * Start somewhere. Can't assume we can send what we can receive,
4411 * but we are clearly receiving.
4413 if (!peer->maxPacketSize)
4414 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4416 if (pktsize > peer->maxPacketSize) {
4417 peer->maxPacketSize = pktsize;
4418 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4419 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4420 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4421 rxi_ScheduleGrowMTUEvent(call, 1);
4426 clock_GetTime(&now);
4428 /* The transmit queue splits into 4 sections.
4430 * The first section is packets which have now been acknowledged
4431 * by a window size change in the ack. These have reached the
4432 * application layer, and may be discarded. These are packets
4433 * with sequence numbers < ap->firstPacket.
4435 * The second section is packets which have sequence numbers in
4436 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4437 * contents of the packet's ack array determines whether these
4438 * packets are acknowledged or not.
4440 * The third section is packets which fall above the range
4441 * addressed in the ack packet. These have not yet been received
4444 * The four section is packets which have not yet been transmitted.
4445 * These packets will have a header.serial of 0.
4448 /* First section - implicitly acknowledged packets that can be
4452 tp = queue_First(&call->tq, rx_packet);
4453 while(!queue_IsEnd(&call->tq, tp) && tp->header.seq < first) {
4454 struct rx_packet *next;
4456 next = queue_Next(tp, rx_packet);
4457 call->tfirst = tp->header.seq + 1;
4459 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4461 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4464 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4465 /* XXX Hack. Because we have to release the global rx lock when sending
4466 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4467 * in rxi_Start sending packets out because packets may move to the
4468 * freePacketQueue as result of being here! So we drop these packets until
4469 * we're safely out of the traversing. Really ugly!
4470 * To make it even uglier, if we're using fine grain locking, we can
4471 * set the ack bits in the packets and have rxi_Start remove the packets
4472 * when it's done transmitting.
4474 if (call->flags & RX_CALL_TQ_BUSY) {
4475 #ifdef RX_ENABLE_LOCKS
4476 tp->flags |= RX_PKTFLAG_ACKED;
4477 call->flags |= RX_CALL_TQ_SOME_ACKED;
4478 #else /* RX_ENABLE_LOCKS */
4480 #endif /* RX_ENABLE_LOCKS */
4482 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4485 #ifdef RX_TRACK_PACKETS
4486 tp->flags &= ~RX_PKTFLAG_TQ;
4488 #ifdef RXDEBUG_PACKET
4490 #endif /* RXDEBUG_PACKET */
4491 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4496 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4498 /* Second section of the queue - packets for which we are receiving
4501 * Go through the explicit acks/nacks and record the results in
4502 * the waiting packets. These are packets that can't be released
4503 * yet, even with a positive acknowledge. This positive
4504 * acknowledge only means the packet has been received by the
4505 * peer, not that it will be retained long enough to be sent to
4506 * the peer's upper level. In addition, reset the transmit timers
4507 * of any missing packets (those packets that must be missing
4508 * because this packet was out of sequence) */
4510 call->nSoftAcked = 0;
4512 while (!queue_IsEnd(&call->tq, tp) && tp->header.seq < first + nAcks) {
4513 /* Set the acknowledge flag per packet based on the
4514 * information in the ack packet. An acknowlegded packet can
4515 * be downgraded when the server has discarded a packet it
4516 * soacked previously, or when an ack packet is received
4517 * out of sequence. */
4518 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4519 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4521 tp->flags |= RX_PKTFLAG_ACKED;
4522 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4529 } else /* RX_ACK_TYPE_NACK */ {
4530 tp->flags &= ~RX_PKTFLAG_ACKED;
4534 tp = queue_Next(tp, rx_packet);
4537 /* We don't need to take any action with the 3rd or 4th section in the
4538 * queue - they're not addressed by the contents of this ACK packet.
4541 /* If the window has been extended by this acknowledge packet,
4542 * then wakeup a sender waiting in alloc for window space, or try
4543 * sending packets now, if he's been sitting on packets due to
4544 * lack of window space */
4545 if (call->tnext < (call->tfirst + call->twind)) {
4546 #ifdef RX_ENABLE_LOCKS
4547 CV_SIGNAL(&call->cv_twind);
4549 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4550 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4551 osi_rxWakeup(&call->twind);
4554 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4555 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4559 /* if the ack packet has a receivelen field hanging off it,
4560 * update our state */
4561 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4564 /* If the ack packet has a "recommended" size that is less than
4565 * what I am using now, reduce my size to match */
4566 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4567 (int)sizeof(afs_int32), &tSize);
4568 tSize = (afs_uint32) ntohl(tSize);
4569 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4571 /* Get the maximum packet size to send to this peer */
4572 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4574 tSize = (afs_uint32) ntohl(tSize);
4575 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4576 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4578 /* sanity check - peer might have restarted with different params.
4579 * If peer says "send less", dammit, send less... Peer should never
4580 * be unable to accept packets of the size that prior AFS versions would
4581 * send without asking. */
4582 if (peer->maxMTU != tSize) {
4583 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4585 peer->maxMTU = tSize;
4586 peer->MTU = MIN(tSize, peer->MTU);
4587 call->MTU = MIN(call->MTU, tSize);
4590 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4593 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4594 (int)sizeof(afs_int32), &tSize);
4595 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4596 if (tSize < call->twind) { /* smaller than our send */
4597 call->twind = tSize; /* window, we must send less... */
4598 call->ssthresh = MIN(call->twind, call->ssthresh);
4599 call->conn->twind[call->channel] = call->twind;
4602 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4603 * network MTU confused with the loopback MTU. Calculate the
4604 * maximum MTU here for use in the slow start code below.
4606 /* Did peer restart with older RX version? */
4607 if (peer->maxDgramPackets > 1) {
4608 peer->maxDgramPackets = 1;
4610 } else if (np->length >=
4611 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4614 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4615 sizeof(afs_int32), &tSize);
4616 tSize = (afs_uint32) ntohl(tSize);
4618 * As of AFS 3.5 we set the send window to match the receive window.
4620 if (tSize < call->twind) {
4621 call->twind = tSize;
4622 call->conn->twind[call->channel] = call->twind;
4623 call->ssthresh = MIN(call->twind, call->ssthresh);
4624 } else if (tSize > call->twind) {
4625 call->twind = tSize;
4626 call->conn->twind[call->channel] = call->twind;
4630 * As of AFS 3.5, a jumbogram is more than one fixed size
4631 * packet transmitted in a single UDP datagram. If the remote
4632 * MTU is smaller than our local MTU then never send a datagram
4633 * larger than the natural MTU.
4636 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4637 (int)sizeof(afs_int32), &tSize);
4638 maxDgramPackets = (afs_uint32) ntohl(tSize);
4639 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4641 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4642 if (maxDgramPackets > 1) {
4643 peer->maxDgramPackets = maxDgramPackets;
4644 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4646 peer->maxDgramPackets = 1;
4647 call->MTU = peer->natMTU;
4649 } else if (peer->maxDgramPackets > 1) {
4650 /* Restarted with lower version of RX */
4651 peer->maxDgramPackets = 1;
4653 } else if (peer->maxDgramPackets > 1
4654 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4655 /* Restarted with lower version of RX */
4656 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4657 peer->natMTU = OLD_MAX_PACKET_SIZE;
4658 peer->MTU = OLD_MAX_PACKET_SIZE;
4659 peer->maxDgramPackets = 1;
4660 peer->nDgramPackets = 1;
4662 call->MTU = OLD_MAX_PACKET_SIZE;
4667 * Calculate how many datagrams were successfully received after
4668 * the first missing packet and adjust the negative ack counter
4673 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4674 if (call->nNacks < nNacked) {
4675 call->nNacks = nNacked;
4678 call->nAcks += newAckCount;
4682 /* If the packet contained new acknowledgements, rather than just
4683 * being a duplicate of one we have previously seen, then we can restart
4686 if (newAckCount > 0)
4687 rxi_rto_packet_acked(call, istack);
4689 if (call->flags & RX_CALL_FAST_RECOVER) {
4690 if (newAckCount == 0) {
4691 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4693 call->flags &= ~RX_CALL_FAST_RECOVER;
4694 call->cwind = call->nextCwind;
4695 call->nextCwind = 0;
4698 call->nCwindAcks = 0;
4699 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4700 /* Three negative acks in a row trigger congestion recovery */
4701 call->flags |= RX_CALL_FAST_RECOVER;
4702 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4704 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4705 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4706 call->nextCwind = call->ssthresh;
4709 peer->MTU = call->MTU;
4710 peer->cwind = call->nextCwind;
4711 peer->nDgramPackets = call->nDgramPackets;
4713 call->congestSeq = peer->congestSeq;
4715 /* Reset the resend times on the packets that were nacked
4716 * so we will retransmit as soon as the window permits
4719 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4721 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4722 tp->flags &= ~RX_PKTFLAG_SENT;
4724 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4729 /* If cwind is smaller than ssthresh, then increase
4730 * the window one packet for each ack we receive (exponential
4732 * If cwind is greater than or equal to ssthresh then increase
4733 * the congestion window by one packet for each cwind acks we
4734 * receive (linear growth). */
4735 if (call->cwind < call->ssthresh) {
4737 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4738 call->nCwindAcks = 0;
4740 call->nCwindAcks += newAckCount;
4741 if (call->nCwindAcks >= call->cwind) {
4742 call->nCwindAcks = 0;
4743 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4747 * If we have received several acknowledgements in a row then
4748 * it is time to increase the size of our datagrams
4750 if ((int)call->nAcks > rx_nDgramThreshold) {
4751 if (peer->maxDgramPackets > 1) {
4752 if (call->nDgramPackets < peer->maxDgramPackets) {
4753 call->nDgramPackets++;
4755 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4756 } else if (call->MTU < peer->maxMTU) {
4757 /* don't upgrade if we can't handle it */
4758 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4759 call->MTU = peer->ifMTU;
4761 call->MTU += peer->natMTU;
4762 call->MTU = MIN(call->MTU, peer->maxMTU);
4769 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4771 /* Servers need to hold the call until all response packets have
4772 * been acknowledged. Soft acks are good enough since clients
4773 * are not allowed to clear their receive queues. */
4774 if (call->state == RX_STATE_HOLD
4775 && call->tfirst + call->nSoftAcked >= call->tnext) {
4776 call->state = RX_STATE_DALLY;
4777 rxi_ClearTransmitQueue(call, 0);
4778 rxevent_Cancel(&call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4779 } else if (!queue_IsEmpty(&call->tq)) {
4780 rxi_Start(call, istack);
4785 /* Received a response to a challenge packet */
4786 static struct rx_packet *
4787 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4788 struct rx_packet *np, int istack)
4792 /* Ignore the packet if we're the client */
4793 if (conn->type == RX_CLIENT_CONNECTION)
4796 /* If already authenticated, ignore the packet (it's probably a retry) */
4797 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4800 /* Otherwise, have the security object evaluate the response packet */
4801 error = RXS_CheckResponse(conn->securityObject, conn, np);
4803 /* If the response is invalid, reset the connection, sending
4804 * an abort to the peer */
4808 rxi_ConnectionError(conn, error);
4809 MUTEX_ENTER(&conn->conn_data_lock);
4810 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4811 MUTEX_EXIT(&conn->conn_data_lock);
4814 /* If the response is valid, any calls waiting to attach
4815 * servers can now do so */
4818 for (i = 0; i < RX_MAXCALLS; i++) {
4819 struct rx_call *call = conn->call[i];
4821 MUTEX_ENTER(&call->lock);
4822 if (call->state == RX_STATE_PRECALL)
4823 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4824 /* tnop can be null if newcallp is null */
4825 MUTEX_EXIT(&call->lock);
4829 /* Update the peer reachability information, just in case
4830 * some calls went into attach-wait while we were waiting
4831 * for authentication..
4833 rxi_UpdatePeerReach(conn, NULL);
4838 /* A client has received an authentication challenge: the security
4839 * object is asked to cough up a respectable response packet to send
4840 * back to the server. The server is responsible for retrying the
4841 * challenge if it fails to get a response. */
4843 static struct rx_packet *
4844 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4845 struct rx_packet *np, int istack)
4849 /* Ignore the challenge if we're the server */
4850 if (conn->type == RX_SERVER_CONNECTION)
4853 /* Ignore the challenge if the connection is otherwise idle; someone's
4854 * trying to use us as an oracle. */
4855 if (!rxi_HasActiveCalls(conn))
4858 /* Send the security object the challenge packet. It is expected to fill
4859 * in the response. */
4860 error = RXS_GetResponse(conn->securityObject, conn, np);
4862 /* If the security object is unable to return a valid response, reset the
4863 * connection and send an abort to the peer. Otherwise send the response
4864 * packet to the peer connection. */
4866 rxi_ConnectionError(conn, error);
4867 MUTEX_ENTER(&conn->conn_data_lock);
4868 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4869 MUTEX_EXIT(&conn->conn_data_lock);
4871 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4872 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4878 /* Find an available server process to service the current request in
4879 * the given call structure. If one isn't available, queue up this
4880 * call so it eventually gets one */
4882 rxi_AttachServerProc(struct rx_call *call,
4883 osi_socket socket, int *tnop,
4884 struct rx_call **newcallp)
4886 struct rx_serverQueueEntry *sq;
4887 struct rx_service *service = call->conn->service;
4890 /* May already be attached */
4891 if (call->state == RX_STATE_ACTIVE)
4894 MUTEX_ENTER(&rx_serverPool_lock);
4896 haveQuota = QuotaOK(service);
4897 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4898 /* If there are no processes available to service this call,
4899 * put the call on the incoming call queue (unless it's
4900 * already on the queue).
4902 #ifdef RX_ENABLE_LOCKS
4904 ReturnToServerPool(service);
4905 #endif /* RX_ENABLE_LOCKS */
4907 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4908 call->flags |= RX_CALL_WAIT_PROC;
4909 rx_atomic_inc(&rx_nWaiting);
4910 rx_atomic_inc(&rx_nWaited);
4911 rxi_calltrace(RX_CALL_ARRIVAL, call);
4912 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4913 queue_Append(&rx_incomingCallQueue, call);
4916 sq = queue_Last(&rx_idleServerQueue, rx_serverQueueEntry);
4918 /* If hot threads are enabled, and both newcallp and sq->socketp
4919 * are non-null, then this thread will process the call, and the
4920 * idle server thread will start listening on this threads socket.
4923 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4926 *sq->socketp = socket;
4927 clock_GetTime(&call->startTime);
4928 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4932 if (call->flags & RX_CALL_WAIT_PROC) {
4933 /* Conservative: I don't think this should happen */
4934 call->flags &= ~RX_CALL_WAIT_PROC;
4935 rx_atomic_dec(&rx_nWaiting);
4936 if (queue_IsOnQueue(call)) {
4940 call->state = RX_STATE_ACTIVE;
4941 call->mode = RX_MODE_RECEIVING;
4942 #ifdef RX_KERNEL_TRACE
4944 int glockOwner = ISAFS_GLOCK();
4947 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4948 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4954 if (call->flags & RX_CALL_CLEARED) {
4955 /* send an ack now to start the packet flow up again */
4956 call->flags &= ~RX_CALL_CLEARED;
4957 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4959 #ifdef RX_ENABLE_LOCKS
4962 service->nRequestsRunning++;
4963 MUTEX_ENTER(&rx_quota_mutex);
4964 if (service->nRequestsRunning <= service->minProcs)
4967 MUTEX_EXIT(&rx_quota_mutex);
4971 MUTEX_EXIT(&rx_serverPool_lock);
4974 /* Delay the sending of an acknowledge event for a short while, while
4975 * a new call is being prepared (in the case of a client) or a reply
4976 * is being prepared (in the case of a server). Rather than sending
4977 * an ack packet, an ACKALL packet is sent. */
4979 rxi_AckAll(struct rx_call *call)
4981 rxi_SendSpecial(call, call->conn, NULL, RX_PACKET_TYPE_ACKALL,
4983 call->flags |= RX_CALL_ACKALL_SENT;
4987 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4990 struct rx_call *call = arg1;
4991 #ifdef RX_ENABLE_LOCKS
4993 MUTEX_ENTER(&call->lock);
4994 if (event == call->delayedAckEvent) {
4995 rxevent_Put(call->delayedAckEvent);
4996 call->delayedAckEvent = NULL;
4998 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
5000 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
5002 MUTEX_EXIT(&call->lock);
5003 #else /* RX_ENABLE_LOCKS */
5005 rxevent_Put(call->delayedAckEvent);
5006 call->delayedAckEvent = NULL;
5008 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
5009 #endif /* RX_ENABLE_LOCKS */
5013 #ifdef RX_ENABLE_LOCKS
5014 /* Set ack in all packets in transmit queue. rxi_Start will deal with
5015 * clearing them out.
5018 rxi_SetAcksInTransmitQueue(struct rx_call *call)
5020 struct rx_packet *p, *tp;
5023 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
5024 p->flags |= RX_PKTFLAG_ACKED;
5028 call->flags |= RX_CALL_TQ_CLEARME;
5029 call->flags |= RX_CALL_TQ_SOME_ACKED;
5032 rxi_rto_cancel(call);
5034 call->tfirst = call->tnext;
5035 call->nSoftAcked = 0;
5037 if (call->flags & RX_CALL_FAST_RECOVER) {
5038 call->flags &= ~RX_CALL_FAST_RECOVER;
5039 call->cwind = call->nextCwind;
5040 call->nextCwind = 0;
5043 CV_SIGNAL(&call->cv_twind);
5045 #endif /* RX_ENABLE_LOCKS */
5047 /* Clear out the transmit queue for the current call (all packets have
5048 * been received by peer) */
5050 rxi_ClearTransmitQueue(struct rx_call *call, int force)
5052 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5053 struct rx_packet *p, *tp;
5055 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
5057 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
5058 p->flags |= RX_PKTFLAG_ACKED;
5062 call->flags |= RX_CALL_TQ_CLEARME;
5063 call->flags |= RX_CALL_TQ_SOME_ACKED;
5066 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5067 #ifdef RXDEBUG_PACKET
5069 #endif /* RXDEBUG_PACKET */
5070 rxi_FreePackets(0, &call->tq);
5071 rxi_WakeUpTransmitQueue(call);
5072 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5073 call->flags &= ~RX_CALL_TQ_CLEARME;
5075 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5077 rxi_rto_cancel(call);
5078 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5079 call->nSoftAcked = 0;
5081 if (call->flags & RX_CALL_FAST_RECOVER) {
5082 call->flags &= ~RX_CALL_FAST_RECOVER;
5083 call->cwind = call->nextCwind;
5085 #ifdef RX_ENABLE_LOCKS
5086 CV_SIGNAL(&call->cv_twind);
5088 osi_rxWakeup(&call->twind);
5093 rxi_ClearReceiveQueue(struct rx_call *call)
5095 if (queue_IsNotEmpty(&call->rq)) {
5098 count = rxi_FreePackets(0, &call->rq);
5099 rx_packetReclaims += count;
5100 #ifdef RXDEBUG_PACKET
5102 if ( call->rqc != 0 )
5103 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5105 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5107 if (call->state == RX_STATE_PRECALL) {
5108 call->flags |= RX_CALL_CLEARED;
5112 /* Send an abort packet for the specified call */
5113 static struct rx_packet *
5114 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5115 int istack, int force)
5117 afs_int32 error, cerror;
5118 struct clock when, now;
5123 switch (call->error) {
5126 cerror = RX_CALL_TIMEOUT;
5129 cerror = call->error;
5132 /* Clients should never delay abort messages */
5133 if (rx_IsClientConn(call->conn))
5136 if (call->abortCode != cerror) {
5137 call->abortCode = cerror;
5138 call->abortCount = 0;
5141 if (force || rxi_callAbortThreshhold == 0
5142 || call->abortCount < rxi_callAbortThreshhold) {
5143 if (call->delayedAbortEvent) {
5144 rxevent_Cancel(&call->delayedAbortEvent, call,
5145 RX_CALL_REFCOUNT_ABORT);
5147 error = htonl(cerror);
5150 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5151 (char *)&error, sizeof(error), istack);
5152 } else if (!call->delayedAbortEvent) {
5153 clock_GetTime(&now);
5155 clock_Addmsec(&when, rxi_callAbortDelay);
5156 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5157 call->delayedAbortEvent =
5158 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5163 /* Send an abort packet for the specified connection. Packet is an
5164 * optional pointer to a packet that can be used to send the abort.
5165 * Once the number of abort messages reaches the threshhold, an
5166 * event is scheduled to send the abort. Setting the force flag
5167 * overrides sending delayed abort messages.
5169 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5170 * to send the abort packet.
5173 rxi_SendConnectionAbort(struct rx_connection *conn,
5174 struct rx_packet *packet, int istack, int force)
5177 struct clock when, now;
5182 /* Clients should never delay abort messages */
5183 if (rx_IsClientConn(conn))
5186 if (force || rxi_connAbortThreshhold == 0
5187 || conn->abortCount < rxi_connAbortThreshhold) {
5189 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
5190 error = htonl(conn->error);
5192 MUTEX_EXIT(&conn->conn_data_lock);
5194 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5195 RX_PACKET_TYPE_ABORT, (char *)&error,
5196 sizeof(error), istack);
5197 MUTEX_ENTER(&conn->conn_data_lock);
5198 } else if (!conn->delayedAbortEvent) {
5199 clock_GetTime(&now);
5201 clock_Addmsec(&when, rxi_connAbortDelay);
5202 conn->delayedAbortEvent =
5203 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
5208 /* Associate an error all of the calls owned by a connection. Called
5209 * with error non-zero. This is only for really fatal things, like
5210 * bad authentication responses. The connection itself is set in
5211 * error at this point, so that future packets received will be
5214 rxi_ConnectionError(struct rx_connection *conn,
5220 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5222 MUTEX_ENTER(&conn->conn_data_lock);
5223 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
5224 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
5225 if (conn->checkReachEvent) {
5226 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
5227 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5228 putConnection(conn);
5230 MUTEX_EXIT(&conn->conn_data_lock);
5231 for (i = 0; i < RX_MAXCALLS; i++) {
5232 struct rx_call *call = conn->call[i];
5234 MUTEX_ENTER(&call->lock);
5235 rxi_CallError(call, error);
5236 MUTEX_EXIT(&call->lock);
5239 conn->error = error;
5240 if (rx_stats_active)
5241 rx_atomic_inc(&rx_stats.fatalErrors);
5246 * Interrupt an in-progress call with the specified error and wakeup waiters.
5248 * @param[in] call The call to interrupt
5249 * @param[in] error The error code to send to the peer
5252 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5254 MUTEX_ENTER(&call->lock);
5255 rxi_CallError(call, error);
5256 rxi_SendCallAbort(call, NULL, 0, 1);
5257 MUTEX_EXIT(&call->lock);
5261 rxi_CallError(struct rx_call *call, afs_int32 error)
5264 osirx_AssertMine(&call->lock, "rxi_CallError");
5266 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5268 error = call->error;
5270 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5271 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5272 rxi_ResetCall(call, 0);
5275 rxi_ResetCall(call, 0);
5277 call->error = error;
5280 /* Reset various fields in a call structure, and wakeup waiting
5281 * processes. Some fields aren't changed: state & mode are not
5282 * touched (these must be set by the caller), and bufptr, nLeft, and
5283 * nFree are not reset, since these fields are manipulated by
5284 * unprotected macros, and may only be reset by non-interrupting code.
5288 rxi_ResetCall(struct rx_call *call, int newcall)
5291 struct rx_peer *peer;
5292 struct rx_packet *packet;
5294 osirx_AssertMine(&call->lock, "rxi_ResetCall");
5296 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5298 /* Notify anyone who is waiting for asynchronous packet arrival */
5299 if (call->arrivalProc) {
5300 (*call->arrivalProc) (call, call->arrivalProcHandle,
5301 call->arrivalProcArg);
5302 call->arrivalProc = (void (*)())0;
5306 rxevent_Cancel(&call->growMTUEvent, call, RX_CALL_REFCOUNT_MTU);
5308 if (call->delayedAbortEvent) {
5309 rxevent_Cancel(&call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
5310 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5312 rxi_SendCallAbort(call, packet, 0, 1);
5313 rxi_FreePacket(packet);
5318 * Update the peer with the congestion information in this call
5319 * so other calls on this connection can pick up where this call
5320 * left off. If the congestion sequence numbers don't match then
5321 * another call experienced a retransmission.
5323 peer = call->conn->peer;
5324 MUTEX_ENTER(&peer->peer_lock);
5326 if (call->congestSeq == peer->congestSeq) {
5327 peer->cwind = MAX(peer->cwind, call->cwind);
5328 peer->MTU = MAX(peer->MTU, call->MTU);
5329 peer->nDgramPackets =
5330 MAX(peer->nDgramPackets, call->nDgramPackets);
5333 call->abortCode = 0;
5334 call->abortCount = 0;
5336 if (peer->maxDgramPackets > 1) {
5337 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5339 call->MTU = peer->MTU;
5341 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5342 call->ssthresh = rx_maxSendWindow;
5343 call->nDgramPackets = peer->nDgramPackets;
5344 call->congestSeq = peer->congestSeq;
5345 call->rtt = peer->rtt;
5346 call->rtt_dev = peer->rtt_dev;
5347 clock_Zero(&call->rto);
5348 clock_Addmsec(&call->rto,
5349 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5350 MUTEX_EXIT(&peer->peer_lock);
5352 flags = call->flags;
5353 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5354 rxi_WaitforTQBusy(call);
5355 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5357 rxi_ClearTransmitQueue(call, 1);
5358 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5359 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5363 if (!newcall && (flags & RX_CALL_PEER_BUSY)) {
5364 /* The call channel is still busy; resetting the call doesn't change
5365 * that. However, if 'newcall' is set, we are processing a call
5366 * structure that has either been recycled from the free list, or has
5367 * been newly allocated. So, RX_CALL_PEER_BUSY is not relevant if
5368 * 'newcall' is set, since it describes a completely different call
5369 * channel which we do not care about. */
5370 call->flags |= RX_CALL_PEER_BUSY;
5373 rxi_ClearReceiveQueue(call);
5374 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5378 call->twind = call->conn->twind[call->channel];
5379 call->rwind = call->conn->rwind[call->channel];
5380 call->nSoftAcked = 0;
5381 call->nextCwind = 0;
5384 call->nCwindAcks = 0;
5385 call->nSoftAcks = 0;
5386 call->nHardAcks = 0;
5388 call->tfirst = call->rnext = call->tnext = 1;
5391 call->lastAcked = 0;
5392 call->localStatus = call->remoteStatus = 0;
5394 if (flags & RX_CALL_READER_WAIT) {
5395 #ifdef RX_ENABLE_LOCKS
5396 CV_BROADCAST(&call->cv_rq);
5398 osi_rxWakeup(&call->rq);
5401 if (flags & RX_CALL_WAIT_PACKETS) {
5402 MUTEX_ENTER(&rx_freePktQ_lock);
5403 rxi_PacketsUnWait(); /* XXX */
5404 MUTEX_EXIT(&rx_freePktQ_lock);
5406 #ifdef RX_ENABLE_LOCKS
5407 CV_SIGNAL(&call->cv_twind);
5409 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5410 osi_rxWakeup(&call->twind);
5413 if (flags & RX_CALL_WAIT_PROC) {
5414 rx_atomic_dec(&rx_nWaiting);
5416 #ifdef RX_ENABLE_LOCKS
5417 /* The following ensures that we don't mess with any queue while some
5418 * other thread might also be doing so. The call_queue_lock field is
5419 * is only modified under the call lock. If the call is in the process
5420 * of being removed from a queue, the call is not locked until the
5421 * the queue lock is dropped and only then is the call_queue_lock field
5422 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5423 * Note that any other routine which removes a call from a queue has to
5424 * obtain the queue lock before examing the queue and removing the call.
5426 if (call->call_queue_lock) {
5427 MUTEX_ENTER(call->call_queue_lock);
5428 if (queue_IsOnQueue(call)) {
5431 MUTEX_EXIT(call->call_queue_lock);
5432 CLEAR_CALL_QUEUE_LOCK(call);
5434 #else /* RX_ENABLE_LOCKS */
5435 if (queue_IsOnQueue(call)) {
5438 #endif /* RX_ENABLE_LOCKS */
5440 rxi_KeepAliveOff(call);
5441 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5444 /* Send an acknowledge for the indicated packet (seq,serial) of the
5445 * indicated call, for the indicated reason (reason). This
5446 * acknowledge will specifically acknowledge receiving the packet, and
5447 * will also specify which other packets for this call have been
5448 * received. This routine returns the packet that was used to the
5449 * caller. The caller is responsible for freeing it or re-using it.
5450 * This acknowledgement also returns the highest sequence number
5451 * actually read out by the higher level to the sender; the sender
5452 * promises to keep around packets that have not been read by the
5453 * higher level yet (unless, of course, the sender decides to abort
5454 * the call altogether). Any of p, seq, serial, pflags, or reason may
5455 * be set to zero without ill effect. That is, if they are zero, they
5456 * will not convey any information.
5457 * NOW there is a trailer field, after the ack where it will safely be
5458 * ignored by mundanes, which indicates the maximum size packet this
5459 * host can swallow. */
5461 struct rx_packet *optionalPacket; use to send ack (or null)
5462 int seq; Sequence number of the packet we are acking
5463 int serial; Serial number of the packet
5464 int pflags; Flags field from packet header
5465 int reason; Reason an acknowledge was prompted
5469 rxi_SendAck(struct rx_call *call,
5470 struct rx_packet *optionalPacket, int serial, int reason,
5473 struct rx_ackPacket *ap;
5474 struct rx_packet *rqp;
5475 struct rx_packet *nxp; /* For queue_Scan */
5476 struct rx_packet *p;
5479 afs_uint32 padbytes = 0;
5480 #ifdef RX_ENABLE_TSFPQ
5481 struct rx_ts_info_t * rx_ts_info;
5485 * Open the receive window once a thread starts reading packets
5487 if (call->rnext > 1) {
5488 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5491 /* Don't attempt to grow MTU if this is a critical ping */
5492 if (reason == RX_ACK_MTU) {
5493 /* keep track of per-call attempts, if we're over max, do in small
5494 * otherwise in larger? set a size to increment by, decrease
5497 if (call->conn->peer->maxPacketSize &&
5498 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5500 padbytes = call->conn->peer->maxPacketSize+16;
5502 padbytes = call->conn->peer->maxMTU + 128;
5504 /* do always try a minimum size ping */
5505 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5507 /* subtract the ack payload */
5508 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5509 reason = RX_ACK_PING;
5512 call->nHardAcks = 0;
5513 call->nSoftAcks = 0;
5514 if (call->rnext > call->lastAcked)
5515 call->lastAcked = call->rnext;
5519 rx_computelen(p, p->length); /* reset length, you never know */
5520 } /* where that's been... */
5521 #ifdef RX_ENABLE_TSFPQ
5523 RX_TS_INFO_GET(rx_ts_info);
5524 if ((p = rx_ts_info->local_special_packet)) {
5525 rx_computelen(p, p->length);
5526 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5527 rx_ts_info->local_special_packet = p;
5528 } else { /* We won't send the ack, but don't panic. */
5529 return optionalPacket;
5533 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5534 /* We won't send the ack, but don't panic. */
5535 return optionalPacket;
5540 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5543 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5544 #ifndef RX_ENABLE_TSFPQ
5545 if (!optionalPacket)
5548 return optionalPacket;
5550 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5551 if (rx_Contiguous(p) < templ) {
5552 #ifndef RX_ENABLE_TSFPQ
5553 if (!optionalPacket)
5556 return optionalPacket;
5561 /* MTUXXX failing to send an ack is very serious. We should */
5562 /* try as hard as possible to send even a partial ack; it's */
5563 /* better than nothing. */
5564 ap = (struct rx_ackPacket *)rx_DataOf(p);
5565 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5566 ap->reason = reason;
5568 /* The skew computation used to be bogus, I think it's better now. */
5569 /* We should start paying attention to skew. XXX */
5570 ap->serial = htonl(serial);
5571 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5574 * First packet not yet forwarded to reader. When ACKALL has been
5575 * sent the peer has been told that all received packets will be
5576 * delivered to the reader. The value 'rnext' is used internally
5577 * to refer to the next packet in the receive queue that must be
5578 * delivered to the reader. From the perspective of the peer it
5579 * already has so report the last sequence number plus one if there
5580 * are packets in the receive queue awaiting processing.
5582 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5583 !queue_IsEmpty(&call->rq)) {
5584 ap->firstPacket = htonl(queue_Last(&call->rq, rx_packet)->header.seq + 1);
5586 ap->firstPacket = htonl(call->rnext);
5588 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5590 /* No fear of running out of ack packet here because there can only be at most
5591 * one window full of unacknowledged packets. The window size must be constrained
5592 * to be less than the maximum ack size, of course. Also, an ack should always
5593 * fit into a single packet -- it should not ever be fragmented. */
5594 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5595 if (!rqp || !call->rq.next
5596 || (rqp->header.seq > (call->rnext + call->rwind))) {
5597 #ifndef RX_ENABLE_TSFPQ
5598 if (!optionalPacket)
5601 rxi_CallError(call, RX_CALL_DEAD);
5602 return optionalPacket;
5605 while (rqp->header.seq > call->rnext + offset)
5606 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5607 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5609 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5610 #ifndef RX_ENABLE_TSFPQ
5611 if (!optionalPacket)
5614 rxi_CallError(call, RX_CALL_DEAD);
5615 return optionalPacket;
5621 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5623 /* these are new for AFS 3.3 */
5624 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5625 templ = htonl(templ);
5626 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5627 templ = htonl(call->conn->peer->ifMTU);
5628 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5629 sizeof(afs_int32), &templ);
5631 /* new for AFS 3.4 */
5632 templ = htonl(call->rwind);
5633 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5634 sizeof(afs_int32), &templ);
5636 /* new for AFS 3.5 */
5637 templ = htonl(call->conn->peer->ifDgramPackets);
5638 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5639 sizeof(afs_int32), &templ);
5641 p->header.serviceId = call->conn->serviceId;
5642 p->header.cid = (call->conn->cid | call->channel);
5643 p->header.callNumber = *call->callNumber;
5645 p->header.securityIndex = call->conn->securityIndex;
5646 p->header.epoch = call->conn->epoch;
5647 p->header.type = RX_PACKET_TYPE_ACK;
5648 p->header.flags = RX_SLOW_START_OK;
5649 if (reason == RX_ACK_PING) {
5650 p->header.flags |= RX_REQUEST_ACK;
5652 p->length = padbytes +
5653 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5656 /* not fast but we can potentially use this if truncated
5657 * fragments are delivered to figure out the mtu.
5659 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5660 sizeof(afs_int32), sizeof(afs_int32),
5664 if (call->conn->type == RX_CLIENT_CONNECTION)
5665 p->header.flags |= RX_CLIENT_INITIATED;
5669 if (rxdebug_active) {
5673 len = _snprintf(msg, sizeof(msg),
5674 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5675 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5676 ntohl(ap->serial), ntohl(ap->previousPacket),
5677 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5678 ap->nAcks, ntohs(ap->bufferSpace) );
5682 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5683 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5687 OutputDebugString(msg);
5689 #else /* AFS_NT40_ENV */
5691 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5692 ap->reason, ntohl(ap->previousPacket),
5693 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5695 for (offset = 0; offset < ap->nAcks; offset++)
5696 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5701 #endif /* AFS_NT40_ENV */
5704 int i, nbytes = p->length;
5706 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5707 if (nbytes <= p->wirevec[i].iov_len) {
5710 savelen = p->wirevec[i].iov_len;
5712 p->wirevec[i].iov_len = nbytes;
5714 rxi_Send(call, p, istack);
5715 p->wirevec[i].iov_len = savelen;
5719 nbytes -= p->wirevec[i].iov_len;
5722 if (rx_stats_active)
5723 rx_atomic_inc(&rx_stats.ackPacketsSent);
5724 #ifndef RX_ENABLE_TSFPQ
5725 if (!optionalPacket)
5728 return optionalPacket; /* Return packet for re-use by caller */
5732 struct rx_packet **list;
5737 /* Send all of the packets in the list in single datagram */
5739 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5740 int istack, int moreFlag)
5746 struct rx_connection *conn = call->conn;
5747 struct rx_peer *peer = conn->peer;
5749 MUTEX_ENTER(&peer->peer_lock);
5750 peer->nSent += xmit->len;
5751 if (xmit->resending)
5752 peer->reSends += xmit->len;
5753 MUTEX_EXIT(&peer->peer_lock);
5755 if (rx_stats_active) {
5756 if (xmit->resending)
5757 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5759 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5762 clock_GetTime(&now);
5764 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5768 /* Set the packet flags and schedule the resend events */
5769 /* Only request an ack for the last packet in the list */
5770 for (i = 0; i < xmit->len; i++) {
5771 struct rx_packet *packet = xmit->list[i];
5773 /* Record the time sent */
5774 packet->timeSent = now;
5775 packet->flags |= RX_PKTFLAG_SENT;
5777 /* Ask for an ack on retransmitted packets, on every other packet
5778 * if the peer doesn't support slow start. Ask for an ack on every
5779 * packet until the congestion window reaches the ack rate. */
5780 if (packet->header.serial) {
5783 packet->firstSent = now;
5784 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5785 || (!(call->flags & RX_CALL_SLOW_START_OK)
5786 && (packet->header.seq & 1)))) {
5791 /* Tag this packet as not being the last in this group,
5792 * for the receiver's benefit */
5793 if (i < xmit->len - 1 || moreFlag) {
5794 packet->header.flags |= RX_MORE_PACKETS;
5799 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5802 /* Since we're about to send a data packet to the peer, it's
5803 * safe to nuke any scheduled end-of-packets ack */
5804 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5806 MUTEX_EXIT(&call->lock);
5807 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5808 if (xmit->len > 1) {
5809 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5811 rxi_SendPacket(call, conn, xmit->list[0], istack);
5813 MUTEX_ENTER(&call->lock);
5814 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5816 /* Tell the RTO calculation engine that we have sent a packet, and
5817 * if it was the last one */
5818 rxi_rto_packet_sent(call, lastPacket, istack);
5820 /* Update last send time for this call (for keep-alive
5821 * processing), and for the connection (so that we can discover
5822 * idle connections) */
5823 conn->lastSendTime = call->lastSendTime = clock_Sec();
5824 /* Let a set of retransmits trigger an idle timeout */
5825 if (!xmit->resending)
5826 call->lastSendData = call->lastSendTime;
5829 /* When sending packets we need to follow these rules:
5830 * 1. Never send more than maxDgramPackets in a jumbogram.
5831 * 2. Never send a packet with more than two iovecs in a jumbogram.
5832 * 3. Never send a retransmitted packet in a jumbogram.
5833 * 4. Never send more than cwind/4 packets in a jumbogram
5834 * We always keep the last list we should have sent so we
5835 * can set the RX_MORE_PACKETS flags correctly.
5839 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5844 struct xmitlist working;
5845 struct xmitlist last;
5847 struct rx_peer *peer = call->conn->peer;
5848 int morePackets = 0;
5850 memset(&last, 0, sizeof(struct xmitlist));
5851 working.list = &list[0];
5853 working.resending = 0;
5855 recovery = call->flags & RX_CALL_FAST_RECOVER;
5857 for (i = 0; i < len; i++) {
5858 /* Does the current packet force us to flush the current list? */
5860 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5861 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5863 /* This sends the 'last' list and then rolls the current working
5864 * set into the 'last' one, and resets the working set */
5867 rxi_SendList(call, &last, istack, 1);
5868 /* If the call enters an error state stop sending, or if
5869 * we entered congestion recovery mode, stop sending */
5871 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5876 working.resending = 0;
5877 working.list = &list[i];
5879 /* Add the current packet to the list if it hasn't been acked.
5880 * Otherwise adjust the list pointer to skip the current packet. */
5881 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5884 if (list[i]->header.serial)
5885 working.resending = 1;
5887 /* Do we need to flush the list? */
5888 if (working.len >= (int)peer->maxDgramPackets
5889 || working.len >= (int)call->nDgramPackets
5890 || working.len >= (int)call->cwind
5891 || list[i]->header.serial
5892 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5894 rxi_SendList(call, &last, istack, 1);
5895 /* If the call enters an error state stop sending, or if
5896 * we entered congestion recovery mode, stop sending */
5898 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5903 working.resending = 0;
5904 working.list = &list[i + 1];
5907 if (working.len != 0) {
5908 osi_Panic("rxi_SendList error");
5910 working.list = &list[i + 1];
5914 /* Send the whole list when the call is in receive mode, when
5915 * the call is in eof mode, when we are in fast recovery mode,
5916 * and when we have the last packet */
5917 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5918 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5919 || (call->flags & RX_CALL_FAST_RECOVER)) {
5920 /* Check for the case where the current list contains
5921 * an acked packet. Since we always send retransmissions
5922 * in a separate packet, we only need to check the first
5923 * packet in the list */
5924 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5928 rxi_SendList(call, &last, istack, morePackets);
5929 /* If the call enters an error state stop sending, or if
5930 * we entered congestion recovery mode, stop sending */
5932 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5936 rxi_SendList(call, &working, istack, 0);
5938 } else if (last.len > 0) {
5939 rxi_SendList(call, &last, istack, 0);
5940 /* Packets which are in 'working' are not sent by this call */
5945 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
5947 struct rx_call *call = arg0;
5948 struct rx_peer *peer;
5949 struct rx_packet *p, *nxp;
5950 struct clock maxTimeout = { 60, 0 };
5952 MUTEX_ENTER(&call->lock);
5954 peer = call->conn->peer;
5956 /* Make sure that the event pointer is removed from the call
5957 * structure, since there is no longer a per-call retransmission
5959 if (event == call->resendEvent) {
5960 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5961 rxevent_Put(call->resendEvent);
5962 call->resendEvent = NULL;
5965 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
5966 rxi_CheckBusy(call);
5969 if (queue_IsEmpty(&call->tq)) {
5970 /* Nothing to do. This means that we've been raced, and that an
5971 * ACK has come in between when we were triggered, and when we
5972 * actually got to run. */
5976 /* We're in loss recovery */
5977 call->flags |= RX_CALL_FAST_RECOVER;
5979 /* Mark all of the pending packets in the queue as being lost */
5980 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5981 if (!(p->flags & RX_PKTFLAG_ACKED))
5982 p->flags &= ~RX_PKTFLAG_SENT;
5985 /* We're resending, so we double the timeout of the call. This will be
5986 * dropped back down by the first successful ACK that we receive.
5988 * We apply a maximum value here of 60 seconds
5990 clock_Add(&call->rto, &call->rto);
5991 if (clock_Gt(&call->rto, &maxTimeout))
5992 call->rto = maxTimeout;
5994 /* Packet loss is most likely due to congestion, so drop our window size
5995 * and start again from the beginning */
5996 if (peer->maxDgramPackets >1) {
5997 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5998 call->MTU = MIN(peer->natMTU, peer->maxMTU);
6000 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
6001 call->nDgramPackets = 1;
6003 call->nextCwind = 1;
6006 MUTEX_ENTER(&peer->peer_lock);
6007 peer->MTU = call->MTU;
6008 peer->cwind = call->cwind;
6009 peer->nDgramPackets = 1;
6011 call->congestSeq = peer->congestSeq;
6012 MUTEX_EXIT(&peer->peer_lock);
6014 rxi_Start(call, istack);
6017 MUTEX_EXIT(&call->lock);
6020 /* This routine is called when new packets are readied for
6021 * transmission and when retransmission may be necessary, or when the
6022 * transmission window or burst count are favourable. This should be
6023 * better optimized for new packets, the usual case, now that we've
6024 * got rid of queues of send packets. XXXXXXXXXXX */
6026 rxi_Start(struct rx_call *call, int istack)
6029 struct rx_packet *p;
6030 struct rx_packet *nxp; /* Next pointer for queue_Scan */
6035 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6036 if (rx_stats_active)
6037 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
6042 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
6044 /* Send (or resend) any packets that need it, subject to
6045 * window restrictions and congestion burst control
6046 * restrictions. Ask for an ack on the last packet sent in
6047 * this burst. For now, we're relying upon the window being
6048 * considerably bigger than the largest number of packets that
6049 * are typically sent at once by one initial call to
6050 * rxi_Start. This is probably bogus (perhaps we should ask
6051 * for an ack when we're half way through the current
6052 * window?). Also, for non file transfer applications, this
6053 * may end up asking for an ack for every packet. Bogus. XXXX
6056 * But check whether we're here recursively, and let the other guy
6059 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6060 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6061 call->flags |= RX_CALL_TQ_BUSY;
6063 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6065 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6066 call->flags &= ~RX_CALL_NEED_START;
6067 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6069 maxXmitPackets = MIN(call->twind, call->cwind);
6070 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
6071 #ifdef RX_TRACK_PACKETS
6072 if ((p->flags & RX_PKTFLAG_FREE)
6073 || (!queue_IsEnd(&call->tq, nxp)
6074 && (nxp->flags & RX_PKTFLAG_FREE))
6075 || (p == (struct rx_packet *)&rx_freePacketQueue)
6076 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
6077 osi_Panic("rxi_Start: xmit queue clobbered");
6080 if (p->flags & RX_PKTFLAG_ACKED) {
6081 /* Since we may block, don't trust this */
6082 if (rx_stats_active)
6083 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6084 continue; /* Ignore this packet if it has been acknowledged */
6087 /* Turn off all flags except these ones, which are the same
6088 * on each transmission */
6089 p->header.flags &= RX_PRESET_FLAGS;
6091 if (p->header.seq >=
6092 call->tfirst + MIN((int)call->twind,
6093 (int)(call->nSoftAcked +
6095 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6096 /* Note: if we're waiting for more window space, we can
6097 * still send retransmits; hence we don't return here, but
6098 * break out to schedule a retransmit event */
6099 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6100 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6105 /* Transmit the packet if it needs to be sent. */
6106 if (!(p->flags & RX_PKTFLAG_SENT)) {
6107 if (nXmitPackets == maxXmitPackets) {
6108 rxi_SendXmitList(call, call->xmitList,
6109 nXmitPackets, istack);
6112 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6113 *(call->callNumber), p));
6114 call->xmitList[nXmitPackets++] = p;
6118 /* xmitList now hold pointers to all of the packets that are
6119 * ready to send. Now we loop to send the packets */
6120 if (nXmitPackets > 0) {
6121 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6125 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6127 /* We went into the error state while sending packets. Now is
6128 * the time to reset the call. This will also inform the using
6129 * process that the call is in an error state.
6131 if (rx_stats_active)
6132 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6133 call->flags &= ~RX_CALL_TQ_BUSY;
6134 rxi_WakeUpTransmitQueue(call);
6135 rxi_CallError(call, call->error);
6138 #ifdef RX_ENABLE_LOCKS
6139 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6141 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6142 /* Some packets have received acks. If they all have, we can clear
6143 * the transmit queue.
6146 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
6147 if (p->header.seq < call->tfirst
6148 && (p->flags & RX_PKTFLAG_ACKED)) {
6150 #ifdef RX_TRACK_PACKETS
6151 p->flags &= ~RX_PKTFLAG_TQ;
6153 #ifdef RXDEBUG_PACKET
6161 call->flags |= RX_CALL_TQ_CLEARME;
6163 #endif /* RX_ENABLE_LOCKS */
6164 if (call->flags & RX_CALL_TQ_CLEARME)
6165 rxi_ClearTransmitQueue(call, 1);
6166 } while (call->flags & RX_CALL_NEED_START);
6168 * TQ references no longer protected by this flag; they must remain
6169 * protected by the global lock.
6171 call->flags &= ~RX_CALL_TQ_BUSY;
6172 rxi_WakeUpTransmitQueue(call);
6174 call->flags |= RX_CALL_NEED_START;
6176 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6178 rxi_rto_cancel(call);
6182 /* Also adjusts the keep alive parameters for the call, to reflect
6183 * that we have just sent a packet (so keep alives aren't sent
6186 rxi_Send(struct rx_call *call, struct rx_packet *p,
6189 struct rx_connection *conn = call->conn;
6191 /* Stamp each packet with the user supplied status */
6192 p->header.userStatus = call->localStatus;
6194 /* Allow the security object controlling this call's security to
6195 * make any last-minute changes to the packet */
6196 RXS_SendPacket(conn->securityObject, call, p);
6198 /* Since we're about to send SOME sort of packet to the peer, it's
6199 * safe to nuke any scheduled end-of-packets ack */
6200 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
6202 /* Actually send the packet, filling in more connection-specific fields */
6203 MUTEX_EXIT(&call->lock);
6204 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6205 rxi_SendPacket(call, conn, p, istack);
6206 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6207 MUTEX_ENTER(&call->lock);
6209 /* Update last send time for this call (for keep-alive
6210 * processing), and for the connection (so that we can discover
6211 * idle connections) */
6212 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6213 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6214 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6216 conn->lastSendTime = call->lastSendTime = clock_Sec();
6217 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6218 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6219 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6220 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6221 RX_ACK_PING_RESPONSE)))
6222 call->lastSendData = call->lastSendTime;
6226 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6227 * that things are fine. Also called periodically to guarantee that nothing
6228 * falls through the cracks (e.g. (error + dally) connections have keepalive
6229 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6231 * haveCTLock Set if calling from rxi_ReapConnections
6233 #ifdef RX_ENABLE_LOCKS
6235 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6236 #else /* RX_ENABLE_LOCKS */
6238 rxi_CheckCall(struct rx_call *call)
6239 #endif /* RX_ENABLE_LOCKS */
6241 struct rx_connection *conn = call->conn;
6243 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6244 afs_uint32 fudgeFactor;
6247 int idle_timeout = 0;
6248 afs_int32 clock_diff = 0;
6250 #ifdef AFS_RXERRQ_ENV
6251 int peererrs = rx_atomic_read(&call->conn->peer->neterrs);
6252 if (call->neterr_gen < peererrs) {
6253 /* we have received network errors since this call started; kill
6255 if (call->state == RX_STATE_ACTIVE) {
6256 rxi_CallError(call, RX_CALL_DEAD);
6260 if (call->neterr_gen > peererrs) {
6261 /* someone has reset the number of peer errors; set the call error gen
6262 * so we can detect if more errors are encountered */
6263 call->neterr_gen = peererrs;
6269 /* Large swings in the clock can have a significant impact on
6270 * the performance of RX call processing. Forward clock shifts
6271 * will result in premature event triggering or timeouts.
6272 * Backward shifts can result in calls not completing until
6273 * the clock catches up with the original start clock value.
6275 * If a backward clock shift of more than five minutes is noticed,
6276 * just fail the call.
6278 if (now < call->lastSendTime)
6279 clock_diff = call->lastSendTime - now;
6280 if (now < call->startWait)
6281 clock_diff = MAX(clock_diff, call->startWait - now);
6282 if (now < call->lastReceiveTime)
6283 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6284 if (clock_diff > 5 * 60)
6286 if (call->state == RX_STATE_ACTIVE)
6287 rxi_CallError(call, RX_CALL_TIMEOUT);
6291 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6292 if (call->flags & RX_CALL_TQ_BUSY) {
6293 /* Call is active and will be reset by rxi_Start if it's
6294 * in an error state.
6299 /* RTT + 8*MDEV, rounded up to the next second. */
6300 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6301 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6303 deadTime = conn->secondsUntilDead + fudgeFactor;
6304 /* These are computed to the second (+- 1 second). But that's
6305 * good enough for these values, which should be a significant
6306 * number of seconds. */
6307 if (now > (call->lastReceiveTime + deadTime)) {
6308 if (call->state == RX_STATE_ACTIVE) {
6309 #ifdef AFS_ADAPT_PMTU
6310 # if defined(KERNEL) && defined(AFS_SUN5_ENV)
6312 # if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6313 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6314 ip_stack_t *ipst = ns->netstack_ip;
6316 ire = ire_cache_lookup(conn->peer->host
6317 # if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6319 # if defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID)
6321 # if defined(GLOBAL_NETSTACKID)
6328 if (ire && ire->ire_max_frag > 0)
6329 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6331 # if defined(GLOBAL_NETSTACKID)
6335 #endif /* AFS_ADAPT_PMTU */
6336 cerror = RX_CALL_DEAD;
6339 #ifdef RX_ENABLE_LOCKS
6340 /* Cancel pending events */
6341 rxevent_Cancel(&call->delayedAckEvent, call,
6342 RX_CALL_REFCOUNT_DELAY);
6343 rxi_rto_cancel(call);
6344 rxevent_Cancel(&call->keepAliveEvent, call,
6345 RX_CALL_REFCOUNT_ALIVE);
6346 rxevent_Cancel(&call->growMTUEvent, call,
6347 RX_CALL_REFCOUNT_MTU);
6348 MUTEX_ENTER(&rx_refcnt_mutex);
6349 /* if rxi_FreeCall returns 1 it has freed the call */
6350 if (call->refCount == 0 &&
6351 rxi_FreeCall(call, haveCTLock))
6353 MUTEX_EXIT(&rx_refcnt_mutex);
6356 MUTEX_EXIT(&rx_refcnt_mutex);
6358 #else /* RX_ENABLE_LOCKS */
6359 rxi_FreeCall(call, 0);
6361 #endif /* RX_ENABLE_LOCKS */
6363 /* Non-active calls are destroyed if they are not responding
6364 * to pings; active calls are simply flagged in error, so the
6365 * attached process can die reasonably gracefully. */
6368 if (conn->idleDeadDetection) {
6369 if (conn->idleDeadTime) {
6370 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6374 /* see if we have a non-activity timeout */
6375 if (call->startWait && ((call->startWait + idleDeadTime) < now) &&
6376 (call->flags & RX_CALL_READER_WAIT)) {
6377 if (call->state == RX_STATE_ACTIVE) {
6378 cerror = RX_CALL_TIMEOUT;
6383 if (call->lastSendData && ((call->lastSendData + idleDeadTime) < now)) {
6384 if (call->state == RX_STATE_ACTIVE) {
6385 cerror = conn->service ? conn->service->idleDeadErr : RX_CALL_IDLE;
6393 if (conn->hardDeadTime) {
6394 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6397 /* see if we have a hard timeout */
6399 && (now > (hardDeadTime + call->startTime.sec))) {
6400 if (call->state == RX_STATE_ACTIVE)
6401 rxi_CallError(call, RX_CALL_TIMEOUT);
6406 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6407 call->lastReceiveTime) {
6408 int oldMTU = conn->peer->ifMTU;
6410 /* if we thought we could send more, perhaps things got worse */
6411 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6412 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6413 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6414 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6416 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6418 /* minimum capped in SetPeerMtu */
6419 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6422 conn->lastPacketSize = 0;
6424 /* needed so ResetCall doesn't clobber us. */
6425 call->MTU = conn->peer->ifMTU;
6427 /* if we never succeeded, let the error pass out as-is */
6428 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6429 cerror = conn->msgsizeRetryErr;
6432 rxi_CallError(call, cerror);
6437 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6438 void *dummy, int dummy2)
6440 struct rx_connection *conn = arg1;
6441 struct rx_header theader;
6442 char tbuffer[1 + sizeof(struct rx_header)];
6443 struct sockaddr_in taddr;
6446 struct iovec tmpiov[2];
6449 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6452 tp = &tbuffer[sizeof(struct rx_header)];
6453 taddr.sin_family = AF_INET;
6454 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6455 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6456 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6457 taddr.sin_len = sizeof(struct sockaddr_in);
6459 memset(&theader, 0, sizeof(theader));
6460 theader.epoch = htonl(999);
6462 theader.callNumber = 0;
6465 theader.type = RX_PACKET_TYPE_VERSION;
6466 theader.flags = RX_LAST_PACKET;
6467 theader.serviceId = 0;
6469 memcpy(tbuffer, &theader, sizeof(theader));
6470 memcpy(tp, &a, sizeof(a));
6471 tmpiov[0].iov_base = tbuffer;
6472 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6474 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6476 MUTEX_ENTER(&conn->conn_data_lock);
6477 MUTEX_ENTER(&rx_refcnt_mutex);
6478 /* Only reschedule ourselves if the connection would not be destroyed */
6479 if (conn->refCount <= 1) {
6480 rxevent_Put(conn->natKeepAliveEvent);
6481 conn->natKeepAliveEvent = NULL;
6482 MUTEX_EXIT(&rx_refcnt_mutex);
6483 MUTEX_EXIT(&conn->conn_data_lock);
6484 rx_DestroyConnection(conn); /* drop the reference for this */
6486 conn->refCount--; /* drop the reference for this */
6487 MUTEX_EXIT(&rx_refcnt_mutex);
6488 rxevent_Put(conn->natKeepAliveEvent);
6489 conn->natKeepAliveEvent = NULL;
6490 rxi_ScheduleNatKeepAliveEvent(conn);
6491 MUTEX_EXIT(&conn->conn_data_lock);
6496 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6498 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6499 struct clock when, now;
6500 clock_GetTime(&now);
6502 when.sec += conn->secondsUntilNatPing;
6503 MUTEX_ENTER(&rx_refcnt_mutex);
6504 conn->refCount++; /* hold a reference for this */
6505 MUTEX_EXIT(&rx_refcnt_mutex);
6506 conn->natKeepAliveEvent =
6507 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6512 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6514 MUTEX_ENTER(&conn->conn_data_lock);
6515 conn->secondsUntilNatPing = seconds;
6517 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6518 rxi_ScheduleNatKeepAliveEvent(conn);
6520 conn->flags |= RX_CONN_NAT_PING;
6522 MUTEX_EXIT(&conn->conn_data_lock);
6525 /* When a call is in progress, this routine is called occasionally to
6526 * make sure that some traffic has arrived (or been sent to) the peer.
6527 * If nothing has arrived in a reasonable amount of time, the call is
6528 * declared dead; if nothing has been sent for a while, we send a
6529 * keep-alive packet (if we're actually trying to keep the call alive)
6532 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6535 struct rx_call *call = arg1;
6536 struct rx_connection *conn;
6539 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6540 MUTEX_ENTER(&call->lock);
6542 if (event == call->keepAliveEvent) {
6543 rxevent_Put(call->keepAliveEvent);
6544 call->keepAliveEvent = NULL;
6549 #ifdef RX_ENABLE_LOCKS
6550 if (rxi_CheckCall(call, 0)) {
6551 MUTEX_EXIT(&call->lock);
6554 #else /* RX_ENABLE_LOCKS */
6555 if (rxi_CheckCall(call))
6557 #endif /* RX_ENABLE_LOCKS */
6559 /* Don't try to keep alive dallying calls */
6560 if (call->state == RX_STATE_DALLY) {
6561 MUTEX_EXIT(&call->lock);
6566 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6567 /* Don't try to send keepalives if there is unacknowledged data */
6568 /* the rexmit code should be good enough, this little hack
6569 * doesn't quite work XXX */
6570 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6572 rxi_ScheduleKeepAliveEvent(call);
6573 MUTEX_EXIT(&call->lock);
6576 /* Does what's on the nameplate. */
6578 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6580 struct rx_call *call = arg1;
6581 struct rx_connection *conn;
6583 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6584 MUTEX_ENTER(&call->lock);
6586 if (event == call->growMTUEvent) {
6587 rxevent_Put(call->growMTUEvent);
6588 call->growMTUEvent = NULL;
6591 #ifdef RX_ENABLE_LOCKS
6592 if (rxi_CheckCall(call, 0)) {
6593 MUTEX_EXIT(&call->lock);
6596 #else /* RX_ENABLE_LOCKS */
6597 if (rxi_CheckCall(call))
6599 #endif /* RX_ENABLE_LOCKS */
6601 /* Don't bother with dallying calls */
6602 if (call->state == RX_STATE_DALLY) {
6603 MUTEX_EXIT(&call->lock);
6610 * keep being scheduled, just don't do anything if we're at peak,
6611 * or we're not set up to be properly handled (idle timeout required)
6613 if ((conn->peer->maxPacketSize != 0) &&
6614 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6615 conn->idleDeadDetection)
6616 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6617 rxi_ScheduleGrowMTUEvent(call, 0);
6618 MUTEX_EXIT(&call->lock);
6622 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6624 if (!call->keepAliveEvent) {
6625 struct clock when, now;
6626 clock_GetTime(&now);
6628 when.sec += call->conn->secondsUntilPing;
6629 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6630 call->keepAliveEvent =
6631 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6636 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6638 if (!call->growMTUEvent) {
6639 struct clock when, now;
6641 clock_GetTime(&now);
6644 if (call->conn->secondsUntilPing)
6645 secs = (6*call->conn->secondsUntilPing)-1;
6647 if (call->conn->secondsUntilDead)
6648 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6652 CALL_HOLD(call, RX_CALL_REFCOUNT_MTU);
6653 call->growMTUEvent =
6654 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6658 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6660 rxi_KeepAliveOn(struct rx_call *call)
6662 /* Pretend last packet received was received now--i.e. if another
6663 * packet isn't received within the keep alive time, then the call
6664 * will die; Initialize last send time to the current time--even
6665 * if a packet hasn't been sent yet. This will guarantee that a
6666 * keep-alive is sent within the ping time */
6667 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6668 rxi_ScheduleKeepAliveEvent(call);
6672 * Solely in order that callers not need to include rx_call.h
6675 rx_KeepAliveOff(struct rx_call *call)
6677 rxi_KeepAliveOff(call);
6680 rx_KeepAliveOn(struct rx_call *call)
6682 rxi_KeepAliveOn(call);
6686 rxi_GrowMTUOn(struct rx_call *call)
6688 struct rx_connection *conn = call->conn;
6689 MUTEX_ENTER(&conn->conn_data_lock);
6690 conn->lastPingSizeSer = conn->lastPingSize = 0;
6691 MUTEX_EXIT(&conn->conn_data_lock);
6692 rxi_ScheduleGrowMTUEvent(call, 1);
6695 /* This routine is called to send connection abort messages
6696 * that have been delayed to throttle looping clients. */
6698 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6701 struct rx_connection *conn = arg1;
6704 struct rx_packet *packet;
6706 MUTEX_ENTER(&conn->conn_data_lock);
6707 rxevent_Put(conn->delayedAbortEvent);
6708 conn->delayedAbortEvent = NULL;
6709 error = htonl(conn->error);
6711 MUTEX_EXIT(&conn->conn_data_lock);
6712 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6715 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6716 RX_PACKET_TYPE_ABORT, (char *)&error,
6718 rxi_FreePacket(packet);
6722 /* This routine is called to send call abort messages
6723 * that have been delayed to throttle looping clients. */
6725 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6728 struct rx_call *call = arg1;
6731 struct rx_packet *packet;
6733 MUTEX_ENTER(&call->lock);
6734 rxevent_Put(call->delayedAbortEvent);
6735 call->delayedAbortEvent = NULL;
6736 error = htonl(call->error);
6738 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6741 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6742 (char *)&error, sizeof(error), 0);
6743 rxi_FreePacket(packet);
6745 MUTEX_EXIT(&call->lock);
6746 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6749 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6750 * seconds) to ask the client to authenticate itself. The routine
6751 * issues a challenge to the client, which is obtained from the
6752 * security object associated with the connection */
6754 rxi_ChallengeEvent(struct rxevent *event,
6755 void *arg0, void *arg1, int tries)
6757 struct rx_connection *conn = arg0;
6760 rxevent_Put(conn->challengeEvent);
6761 conn->challengeEvent = NULL;
6764 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6765 struct rx_packet *packet;
6766 struct clock when, now;
6769 /* We've failed to authenticate for too long.
6770 * Reset any calls waiting for authentication;
6771 * they are all in RX_STATE_PRECALL.
6775 MUTEX_ENTER(&conn->conn_call_lock);
6776 for (i = 0; i < RX_MAXCALLS; i++) {
6777 struct rx_call *call = conn->call[i];
6779 MUTEX_ENTER(&call->lock);
6780 if (call->state == RX_STATE_PRECALL) {
6781 rxi_CallError(call, RX_CALL_DEAD);
6782 rxi_SendCallAbort(call, NULL, 0, 0);
6784 MUTEX_EXIT(&call->lock);
6787 MUTEX_EXIT(&conn->conn_call_lock);
6791 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6793 /* If there's no packet available, do this later. */
6794 RXS_GetChallenge(conn->securityObject, conn, packet);
6795 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6796 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6797 rxi_FreePacket(packet);
6799 clock_GetTime(&now);
6801 when.sec += RX_CHALLENGE_TIMEOUT;
6802 conn->challengeEvent =
6803 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6808 /* Call this routine to start requesting the client to authenticate
6809 * itself. This will continue until authentication is established,
6810 * the call times out, or an invalid response is returned. The
6811 * security object associated with the connection is asked to create
6812 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6813 * defined earlier. */
6815 rxi_ChallengeOn(struct rx_connection *conn)
6817 if (!conn->challengeEvent) {
6818 RXS_CreateChallenge(conn->securityObject, conn);
6819 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6824 /* rxi_ComputeRoundTripTime is called with peer locked. */
6825 /* peer may be null */
6827 rxi_ComputeRoundTripTime(struct rx_packet *p,
6828 struct rx_ackPacket *ack,
6829 struct rx_call *call,
6830 struct rx_peer *peer,
6833 struct clock thisRtt, *sentp;
6837 /* If the ACK is delayed, then do nothing */
6838 if (ack->reason == RX_ACK_DELAY)
6841 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6842 * their RTT multiple times, so only include the RTT of the last packet
6844 if (p->flags & RX_JUMBO_PACKET)
6847 /* Use the serial number to determine which transmission the ACK is for,
6848 * and set the sent time to match this. If we have no serial number, then
6849 * only use the ACK for RTT calculations if the packet has not been
6853 serial = ntohl(ack->serial);
6855 if (serial == p->header.serial) {
6856 sentp = &p->timeSent;
6857 } else if (serial == p->firstSerial) {
6858 sentp = &p->firstSent;
6859 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6860 sentp = &p->firstSent;
6864 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6865 sentp = &p->firstSent;
6872 if (clock_Lt(&thisRtt, sentp))
6873 return; /* somebody set the clock back, don't count this time. */
6875 clock_Sub(&thisRtt, sentp);
6876 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6877 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6879 if (clock_IsZero(&thisRtt)) {
6881 * The actual round trip time is shorter than the
6882 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6883 * Since we can't tell which at the moment we will assume 1ms.
6885 thisRtt.usec = 1000;
6888 if (rx_stats_active) {
6889 MUTEX_ENTER(&rx_stats_mutex);
6890 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6891 rx_stats.minRtt = thisRtt;
6892 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6893 if (thisRtt.sec > 60) {
6894 MUTEX_EXIT(&rx_stats_mutex);
6895 return; /* somebody set the clock ahead */
6897 rx_stats.maxRtt = thisRtt;
6899 clock_Add(&rx_stats.totalRtt, &thisRtt);
6900 rx_atomic_inc(&rx_stats.nRttSamples);
6901 MUTEX_EXIT(&rx_stats_mutex);
6904 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6906 /* Apply VanJacobson round-trip estimations */
6911 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6912 * srtt is stored as fixed point with 3 bits after the binary
6913 * point (i.e., scaled by 8). The following magic is
6914 * equivalent to the smoothing algorithm in rfc793 with an
6915 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6916 * srtt'*8 = rtt + srtt*7
6917 * srtt'*8 = srtt*8 + rtt - srtt
6918 * srtt' = srtt + rtt/8 - srtt/8
6919 * srtt' = srtt + (rtt - srtt)/8
6922 delta = _8THMSEC(&thisRtt) - call->rtt;
6923 call->rtt += (delta >> 3);
6926 * We accumulate a smoothed rtt variance (actually, a smoothed
6927 * mean difference), then set the retransmit timer to smoothed
6928 * rtt + 4 times the smoothed variance (was 2x in van's original
6929 * paper, but 4x works better for me, and apparently for him as
6931 * rttvar is stored as
6932 * fixed point with 2 bits after the binary point (scaled by
6933 * 4). The following is equivalent to rfc793 smoothing with
6934 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6935 * rttvar'*4 = rttvar*3 + |delta|
6936 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6937 * rttvar' = rttvar + |delta|/4 - rttvar/4
6938 * rttvar' = rttvar + (|delta| - rttvar)/4
6939 * This replaces rfc793's wired-in beta.
6940 * dev*4 = dev*4 + (|actual - expected| - dev)
6946 delta -= (call->rtt_dev << 1);
6947 call->rtt_dev += (delta >> 3);
6949 /* I don't have a stored RTT so I start with this value. Since I'm
6950 * probably just starting a call, and will be pushing more data down
6951 * this, I expect congestion to increase rapidly. So I fudge a
6952 * little, and I set deviance to half the rtt. In practice,
6953 * deviance tends to approach something a little less than
6954 * half the smoothed rtt. */
6955 call->rtt = _8THMSEC(&thisRtt) + 8;
6956 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6958 /* the smoothed RTT time is RTT + 4*MDEV
6960 * We allow a user specified minimum to be set for this, to allow clamping
6961 * at a minimum value in the same way as TCP. In addition, we have to allow
6962 * for the possibility that this packet is answered by a delayed ACK, so we
6963 * add on a fixed 200ms to account for that timer expiring.
6966 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6967 rx_minPeerTimeout) + 200;
6968 clock_Zero(&call->rto);
6969 clock_Addmsec(&call->rto, rtt_timeout);
6971 /* Update the peer, so any new calls start with our values */
6972 peer->rtt_dev = call->rtt_dev;
6973 peer->rtt = call->rtt;
6975 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6976 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6980 /* Find all server connections that have not been active for a long time, and
6983 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
6986 struct clock now, when;
6987 clock_GetTime(&now);
6989 /* Find server connection structures that haven't been used for
6990 * greater than rx_idleConnectionTime */
6992 struct rx_connection **conn_ptr, **conn_end;
6993 int i, havecalls = 0;
6994 MUTEX_ENTER(&rx_connHashTable_lock);
6995 for (conn_ptr = &rx_connHashTable[0], conn_end =
6996 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6998 struct rx_connection *conn, *next;
6999 struct rx_call *call;
7003 for (conn = *conn_ptr; conn; conn = next) {
7004 /* XXX -- Shouldn't the connection be locked? */
7007 for (i = 0; i < RX_MAXCALLS; i++) {
7008 call = conn->call[i];
7012 code = MUTEX_TRYENTER(&call->lock);
7015 #ifdef RX_ENABLE_LOCKS
7016 result = rxi_CheckCall(call, 1);
7017 #else /* RX_ENABLE_LOCKS */
7018 result = rxi_CheckCall(call);
7019 #endif /* RX_ENABLE_LOCKS */
7020 MUTEX_EXIT(&call->lock);
7022 /* If CheckCall freed the call, it might
7023 * have destroyed the connection as well,
7024 * which screws up the linked lists.
7030 if (conn->type == RX_SERVER_CONNECTION) {
7031 /* This only actually destroys the connection if
7032 * there are no outstanding calls */
7033 MUTEX_ENTER(&conn->conn_data_lock);
7034 MUTEX_ENTER(&rx_refcnt_mutex);
7035 if (!havecalls && !conn->refCount
7036 && ((conn->lastSendTime + rx_idleConnectionTime) <
7038 conn->refCount++; /* it will be decr in rx_DestroyConn */
7039 MUTEX_EXIT(&rx_refcnt_mutex);
7040 MUTEX_EXIT(&conn->conn_data_lock);
7041 #ifdef RX_ENABLE_LOCKS
7042 rxi_DestroyConnectionNoLock(conn);
7043 #else /* RX_ENABLE_LOCKS */
7044 rxi_DestroyConnection(conn);
7045 #endif /* RX_ENABLE_LOCKS */
7047 #ifdef RX_ENABLE_LOCKS
7049 MUTEX_EXIT(&rx_refcnt_mutex);
7050 MUTEX_EXIT(&conn->conn_data_lock);
7052 #endif /* RX_ENABLE_LOCKS */
7056 #ifdef RX_ENABLE_LOCKS
7057 while (rx_connCleanup_list) {
7058 struct rx_connection *conn;
7059 conn = rx_connCleanup_list;
7060 rx_connCleanup_list = rx_connCleanup_list->next;
7061 MUTEX_EXIT(&rx_connHashTable_lock);
7062 rxi_CleanupConnection(conn);
7063 MUTEX_ENTER(&rx_connHashTable_lock);
7065 MUTEX_EXIT(&rx_connHashTable_lock);
7066 #endif /* RX_ENABLE_LOCKS */
7069 /* Find any peer structures that haven't been used (haven't had an
7070 * associated connection) for greater than rx_idlePeerTime */
7072 struct rx_peer **peer_ptr, **peer_end;
7076 * Why do we need to hold the rx_peerHashTable_lock across
7077 * the incrementing of peer_ptr since the rx_peerHashTable
7078 * array is not changing? We don't.
7080 * By dropping the lock periodically we can permit other
7081 * activities to be performed while a rxi_ReapConnections
7082 * call is in progress. The goal of reap connections
7083 * is to clean up quickly without causing large amounts
7084 * of contention. Therefore, it is important that global
7085 * mutexes not be held for extended periods of time.
7087 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7088 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7090 struct rx_peer *peer, *next, *prev;
7092 MUTEX_ENTER(&rx_peerHashTable_lock);
7093 for (prev = peer = *peer_ptr; peer; peer = next) {
7095 code = MUTEX_TRYENTER(&peer->peer_lock);
7096 if ((code) && (peer->refCount == 0)
7097 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7098 rx_interface_stat_p rpc_stat, nrpc_stat;
7102 * now know that this peer object is one to be
7103 * removed from the hash table. Once it is removed
7104 * it can't be referenced by other threads.
7105 * Lets remove it first and decrement the struct
7106 * nPeerStructs count.
7108 if (peer == *peer_ptr) {
7114 if (rx_stats_active)
7115 rx_atomic_dec(&rx_stats.nPeerStructs);
7118 * Now if we hold references on 'prev' and 'next'
7119 * we can safely drop the rx_peerHashTable_lock
7120 * while we destroy this 'peer' object.
7126 MUTEX_EXIT(&rx_peerHashTable_lock);
7128 MUTEX_EXIT(&peer->peer_lock);
7129 MUTEX_DESTROY(&peer->peer_lock);
7131 (&peer->rpcStats, rpc_stat, nrpc_stat,
7132 rx_interface_stat)) {
7133 unsigned int num_funcs;
7136 queue_Remove(&rpc_stat->queue_header);
7137 queue_Remove(&rpc_stat->all_peers);
7138 num_funcs = rpc_stat->stats[0].func_total;
7140 sizeof(rx_interface_stat_t) +
7141 rpc_stat->stats[0].func_total *
7142 sizeof(rx_function_entry_v1_t);
7144 rxi_Free(rpc_stat, space);
7146 MUTEX_ENTER(&rx_rpc_stats);
7147 rxi_rpc_peer_stat_cnt -= num_funcs;
7148 MUTEX_EXIT(&rx_rpc_stats);
7153 * Regain the rx_peerHashTable_lock and
7154 * decrement the reference count on 'prev'
7157 MUTEX_ENTER(&rx_peerHashTable_lock);
7164 MUTEX_EXIT(&peer->peer_lock);
7169 MUTEX_EXIT(&rx_peerHashTable_lock);
7173 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7174 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7175 * GC, just below. Really, we shouldn't have to keep moving packets from
7176 * one place to another, but instead ought to always know if we can
7177 * afford to hold onto a packet in its particular use. */
7178 MUTEX_ENTER(&rx_freePktQ_lock);
7179 if (rx_waitingForPackets) {
7180 rx_waitingForPackets = 0;
7181 #ifdef RX_ENABLE_LOCKS
7182 CV_BROADCAST(&rx_waitingForPackets_cv);
7184 osi_rxWakeup(&rx_waitingForPackets);
7187 MUTEX_EXIT(&rx_freePktQ_lock);
7190 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7191 rxevent_Put(rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0));
7195 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7196 * rx.h is sort of strange this is better. This is called with a security
7197 * object before it is discarded. Each connection using a security object has
7198 * its own refcount to the object so it won't actually be freed until the last
7199 * connection is destroyed.
7201 * This is the only rxs module call. A hold could also be written but no one
7205 rxs_Release(struct rx_securityClass *aobj)
7207 return RXS_Close(aobj);
7215 #define TRACE_OPTION_RX_DEBUG 16
7223 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7224 0, KEY_QUERY_VALUE, &parmKey);
7225 if (code != ERROR_SUCCESS)
7228 dummyLen = sizeof(TraceOption);
7229 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7230 (BYTE *) &TraceOption, &dummyLen);
7231 if (code == ERROR_SUCCESS) {
7232 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7234 RegCloseKey (parmKey);
7235 #endif /* AFS_NT40_ENV */
7240 rx_DebugOnOff(int on)
7244 rxdebug_active = on;
7250 rx_StatsOnOff(int on)
7252 rx_stats_active = on;
7256 /* Don't call this debugging routine directly; use dpf */
7258 rxi_DebugPrint(char *format, ...)
7267 va_start(ap, format);
7269 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7272 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7274 OutputDebugString(msg);
7280 va_start(ap, format);
7282 clock_GetTime(&now);
7283 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7284 (unsigned int)now.usec);
7285 vfprintf(rx_Log, format, ap);
7293 * This function is used to process the rx_stats structure that is local
7294 * to a process as well as an rx_stats structure received from a remote
7295 * process (via rxdebug). Therefore, it needs to do minimal version
7299 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7300 afs_int32 freePackets, char version)
7304 if (size != sizeof(struct rx_statistics)) {
7306 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7307 size, sizeof(struct rx_statistics));
7310 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7313 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7314 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7315 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7316 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7317 s->specialPktAllocFailures);
7319 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7320 s->receivePktAllocFailures, s->sendPktAllocFailures,
7321 s->specialPktAllocFailures);
7325 " greedy %u, " "bogusReads %u (last from host %x), "
7326 "noPackets %u, " "noBuffers %u, " "selects %u, "
7327 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7328 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7329 s->selects, s->sendSelects);
7331 fprintf(file, " packets read: ");
7332 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7333 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7335 fprintf(file, "\n");
7338 " other read counters: data %u, " "ack %u, " "dup %u "
7339 "spurious %u " "dally %u\n", s->dataPacketsRead,
7340 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7341 s->ignorePacketDally);
7343 fprintf(file, " packets sent: ");
7344 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7345 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7347 fprintf(file, "\n");
7350 " other send counters: ack %u, " "data %u (not resends), "
7351 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7352 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7353 s->dataPacketsPushed, s->ignoreAckedPacket);
7356 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7357 s->netSendFailures, (int)s->fatalErrors);
7359 if (s->nRttSamples) {
7360 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7361 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7363 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7364 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7368 " %d server connections, " "%d client connections, "
7369 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7370 s->nServerConns, s->nClientConns, s->nPeerStructs,
7371 s->nCallStructs, s->nFreeCallStructs);
7373 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7374 fprintf(file, " %d clock updates\n", clock_nUpdates);
7378 /* for backward compatibility */
7380 rx_PrintStats(FILE * file)
7382 MUTEX_ENTER(&rx_stats_mutex);
7383 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7384 sizeof(rx_stats), rx_nFreePackets,
7386 MUTEX_EXIT(&rx_stats_mutex);
7390 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7392 fprintf(file, "Peer %x.%d.\n",
7393 ntohl(peer->host), (int)ntohs(peer->port));
7396 " Rtt %d, " "total sent %d, " "resent %d\n",
7397 peer->rtt, peer->nSent, peer->reSends);
7399 fprintf(file, " Packet size %d\n", peer->ifMTU);
7403 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7405 * This mutex protects the following static variables:
7409 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7410 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7412 #define LOCK_RX_DEBUG
7413 #define UNLOCK_RX_DEBUG
7414 #endif /* AFS_PTHREAD_ENV */
7416 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7418 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7419 u_char type, void *inputData, size_t inputLength,
7420 void *outputData, size_t outputLength)
7422 static afs_int32 counter = 100;
7423 time_t waitTime, waitCount;
7424 struct rx_header theader;
7427 struct timeval tv_now, tv_wake, tv_delta;
7428 struct sockaddr_in taddr, faddr;
7442 tp = &tbuffer[sizeof(struct rx_header)];
7443 taddr.sin_family = AF_INET;
7444 taddr.sin_port = remotePort;
7445 taddr.sin_addr.s_addr = remoteAddr;
7446 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7447 taddr.sin_len = sizeof(struct sockaddr_in);
7450 memset(&theader, 0, sizeof(theader));
7451 theader.epoch = htonl(999);
7453 theader.callNumber = htonl(counter);
7456 theader.type = type;
7457 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7458 theader.serviceId = 0;
7460 memcpy(tbuffer, &theader, sizeof(theader));
7461 memcpy(tp, inputData, inputLength);
7463 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7464 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7466 /* see if there's a packet available */
7467 gettimeofday(&tv_wake, NULL);
7468 tv_wake.tv_sec += waitTime;
7471 FD_SET(socket, &imask);
7472 tv_delta.tv_sec = tv_wake.tv_sec;
7473 tv_delta.tv_usec = tv_wake.tv_usec;
7474 gettimeofday(&tv_now, NULL);
7476 if (tv_delta.tv_usec < tv_now.tv_usec) {
7478 tv_delta.tv_usec += 1000000;
7481 tv_delta.tv_usec -= tv_now.tv_usec;
7483 if (tv_delta.tv_sec < tv_now.tv_sec) {
7487 tv_delta.tv_sec -= tv_now.tv_sec;
7490 code = select(0, &imask, 0, 0, &tv_delta);
7491 #else /* AFS_NT40_ENV */
7492 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7493 #endif /* AFS_NT40_ENV */
7494 if (code == 1 && FD_ISSET(socket, &imask)) {
7495 /* now receive a packet */
7496 faddrLen = sizeof(struct sockaddr_in);
7498 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7499 (struct sockaddr *)&faddr, &faddrLen);
7502 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7503 if (counter == ntohl(theader.callNumber))
7511 /* see if we've timed out */
7519 code -= sizeof(struct rx_header);
7520 if (code > outputLength)
7521 code = outputLength;
7522 memcpy(outputData, tp, code);
7525 #endif /* RXDEBUG */
7528 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7529 afs_uint16 remotePort, struct rx_debugStats * stat,
7530 afs_uint32 * supportedValues)
7532 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7534 struct rx_debugIn in;
7536 *supportedValues = 0;
7537 in.type = htonl(RX_DEBUGI_GETSTATS);
7540 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7541 &in, sizeof(in), stat, sizeof(*stat));
7544 * If the call was successful, fixup the version and indicate
7545 * what contents of the stat structure are valid.
7546 * Also do net to host conversion of fields here.
7550 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7551 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7553 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7554 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7556 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7557 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7559 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7560 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7562 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7563 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7565 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7566 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7568 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7569 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7571 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7572 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7574 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7575 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7577 stat->nFreePackets = ntohl(stat->nFreePackets);
7578 stat->packetReclaims = ntohl(stat->packetReclaims);
7579 stat->callsExecuted = ntohl(stat->callsExecuted);
7580 stat->nWaiting = ntohl(stat->nWaiting);
7581 stat->idleThreads = ntohl(stat->idleThreads);
7582 stat->nWaited = ntohl(stat->nWaited);
7583 stat->nPackets = ntohl(stat->nPackets);
7592 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7593 afs_uint16 remotePort, struct rx_statistics * stat,
7594 afs_uint32 * supportedValues)
7596 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7598 struct rx_debugIn in;
7599 afs_int32 *lp = (afs_int32 *) stat;
7603 * supportedValues is currently unused, but added to allow future
7604 * versioning of this function.
7607 *supportedValues = 0;
7608 in.type = htonl(RX_DEBUGI_RXSTATS);
7610 memset(stat, 0, sizeof(*stat));
7612 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7613 &in, sizeof(in), stat, sizeof(*stat));
7618 * Do net to host conversion here
7621 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7632 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7633 afs_uint16 remotePort, size_t version_length,
7636 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7638 return MakeDebugCall(socket, remoteAddr, remotePort,
7639 RX_PACKET_TYPE_VERSION, a, 1, version,
7647 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7648 afs_uint16 remotePort, afs_int32 * nextConnection,
7649 int allConnections, afs_uint32 debugSupportedValues,
7650 struct rx_debugConn * conn,
7651 afs_uint32 * supportedValues)
7653 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7655 struct rx_debugIn in;
7659 * supportedValues is currently unused, but added to allow future
7660 * versioning of this function.
7663 *supportedValues = 0;
7664 if (allConnections) {
7665 in.type = htonl(RX_DEBUGI_GETALLCONN);
7667 in.type = htonl(RX_DEBUGI_GETCONN);
7669 in.index = htonl(*nextConnection);
7670 memset(conn, 0, sizeof(*conn));
7672 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7673 &in, sizeof(in), conn, sizeof(*conn));
7676 *nextConnection += 1;
7679 * Convert old connection format to new structure.
7682 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7683 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7684 #define MOVEvL(a) (conn->a = vL->a)
7686 /* any old or unrecognized version... */
7687 for (i = 0; i < RX_MAXCALLS; i++) {
7688 MOVEvL(callState[i]);
7689 MOVEvL(callMode[i]);
7690 MOVEvL(callFlags[i]);
7691 MOVEvL(callOther[i]);
7693 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7694 MOVEvL(secStats.type);
7695 MOVEvL(secStats.level);
7696 MOVEvL(secStats.flags);
7697 MOVEvL(secStats.expires);
7698 MOVEvL(secStats.packetsReceived);
7699 MOVEvL(secStats.packetsSent);
7700 MOVEvL(secStats.bytesReceived);
7701 MOVEvL(secStats.bytesSent);
7706 * Do net to host conversion here
7708 * I don't convert host or port since we are most likely
7709 * going to want these in NBO.
7711 conn->cid = ntohl(conn->cid);
7712 conn->serial = ntohl(conn->serial);
7713 for (i = 0; i < RX_MAXCALLS; i++) {
7714 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7716 conn->error = ntohl(conn->error);
7717 conn->secStats.flags = ntohl(conn->secStats.flags);
7718 conn->secStats.expires = ntohl(conn->secStats.expires);
7719 conn->secStats.packetsReceived =
7720 ntohl(conn->secStats.packetsReceived);
7721 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7722 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7723 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7724 conn->epoch = ntohl(conn->epoch);
7725 conn->natMTU = ntohl(conn->natMTU);
7734 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7735 afs_uint16 remotePort, afs_int32 * nextPeer,
7736 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7737 afs_uint32 * supportedValues)
7739 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7741 struct rx_debugIn in;
7744 * supportedValues is currently unused, but added to allow future
7745 * versioning of this function.
7748 *supportedValues = 0;
7749 in.type = htonl(RX_DEBUGI_GETPEER);
7750 in.index = htonl(*nextPeer);
7751 memset(peer, 0, sizeof(*peer));
7753 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7754 &in, sizeof(in), peer, sizeof(*peer));
7760 * Do net to host conversion here
7762 * I don't convert host or port since we are most likely
7763 * going to want these in NBO.
7765 peer->ifMTU = ntohs(peer->ifMTU);
7766 peer->idleWhen = ntohl(peer->idleWhen);
7767 peer->refCount = ntohs(peer->refCount);
7768 peer->rtt = ntohl(peer->rtt);
7769 peer->rtt_dev = ntohl(peer->rtt_dev);
7770 peer->timeout.sec = 0;
7771 peer->timeout.usec = 0;
7772 peer->nSent = ntohl(peer->nSent);
7773 peer->reSends = ntohl(peer->reSends);
7774 peer->natMTU = ntohs(peer->natMTU);
7775 peer->maxMTU = ntohs(peer->maxMTU);
7776 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7777 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7778 peer->MTU = ntohs(peer->MTU);
7779 peer->cwind = ntohs(peer->cwind);
7780 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7781 peer->congestSeq = ntohs(peer->congestSeq);
7782 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7783 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7784 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7785 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7794 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7795 struct rx_debugPeer * peerStats)
7798 afs_int32 error = 1; /* default to "did not succeed" */
7799 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7801 MUTEX_ENTER(&rx_peerHashTable_lock);
7802 for(tp = rx_peerHashTable[hashValue];
7803 tp != NULL; tp = tp->next) {
7804 if (tp->host == peerHost)
7810 MUTEX_EXIT(&rx_peerHashTable_lock);
7814 MUTEX_ENTER(&tp->peer_lock);
7815 peerStats->host = tp->host;
7816 peerStats->port = tp->port;
7817 peerStats->ifMTU = tp->ifMTU;
7818 peerStats->idleWhen = tp->idleWhen;
7819 peerStats->refCount = tp->refCount;
7820 peerStats->burstSize = 0;
7821 peerStats->burst = 0;
7822 peerStats->burstWait.sec = 0;
7823 peerStats->burstWait.usec = 0;
7824 peerStats->rtt = tp->rtt;
7825 peerStats->rtt_dev = tp->rtt_dev;
7826 peerStats->timeout.sec = 0;
7827 peerStats->timeout.usec = 0;
7828 peerStats->nSent = tp->nSent;
7829 peerStats->reSends = tp->reSends;
7830 peerStats->natMTU = tp->natMTU;
7831 peerStats->maxMTU = tp->maxMTU;
7832 peerStats->maxDgramPackets = tp->maxDgramPackets;
7833 peerStats->ifDgramPackets = tp->ifDgramPackets;
7834 peerStats->MTU = tp->MTU;
7835 peerStats->cwind = tp->cwind;
7836 peerStats->nDgramPackets = tp->nDgramPackets;
7837 peerStats->congestSeq = tp->congestSeq;
7838 peerStats->bytesSent.high = tp->bytesSent >> 32;
7839 peerStats->bytesSent.low = tp->bytesSent & MAX_AFS_UINT32;
7840 peerStats->bytesReceived.high = tp->bytesReceived >> 32;
7841 peerStats->bytesReceived.low
7842 = tp->bytesReceived & MAX_AFS_UINT32;
7843 MUTEX_EXIT(&tp->peer_lock);
7845 MUTEX_ENTER(&rx_peerHashTable_lock);
7848 MUTEX_EXIT(&rx_peerHashTable_lock);
7856 struct rx_serverQueueEntry *np;
7859 struct rx_call *call;
7860 struct rx_serverQueueEntry *sq;
7864 if (rxinit_status == 1) {
7866 return; /* Already shutdown. */
7870 #ifndef AFS_PTHREAD_ENV
7871 FD_ZERO(&rx_selectMask);
7872 #endif /* AFS_PTHREAD_ENV */
7873 rxi_dataQuota = RX_MAX_QUOTA;
7874 #ifndef AFS_PTHREAD_ENV
7876 #endif /* AFS_PTHREAD_ENV */
7879 #ifndef AFS_PTHREAD_ENV
7880 #ifndef AFS_USE_GETTIMEOFDAY
7882 #endif /* AFS_USE_GETTIMEOFDAY */
7883 #endif /* AFS_PTHREAD_ENV */
7885 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7886 call = queue_First(&rx_freeCallQueue, rx_call);
7888 rxi_Free(call, sizeof(struct rx_call));
7891 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7892 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7898 struct rx_peer **peer_ptr, **peer_end;
7899 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7900 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7902 struct rx_peer *peer, *next;
7904 MUTEX_ENTER(&rx_peerHashTable_lock);
7905 for (peer = *peer_ptr; peer; peer = next) {
7906 rx_interface_stat_p rpc_stat, nrpc_stat;
7909 MUTEX_ENTER(&rx_rpc_stats);
7910 MUTEX_ENTER(&peer->peer_lock);
7912 (&peer->rpcStats, rpc_stat, nrpc_stat,
7913 rx_interface_stat)) {
7914 unsigned int num_funcs;
7917 queue_Remove(&rpc_stat->queue_header);
7918 queue_Remove(&rpc_stat->all_peers);
7919 num_funcs = rpc_stat->stats[0].func_total;
7921 sizeof(rx_interface_stat_t) +
7922 rpc_stat->stats[0].func_total *
7923 sizeof(rx_function_entry_v1_t);
7925 rxi_Free(rpc_stat, space);
7927 /* rx_rpc_stats must be held */
7928 rxi_rpc_peer_stat_cnt -= num_funcs;
7930 MUTEX_EXIT(&peer->peer_lock);
7931 MUTEX_EXIT(&rx_rpc_stats);
7935 if (rx_stats_active)
7936 rx_atomic_dec(&rx_stats.nPeerStructs);
7938 MUTEX_EXIT(&rx_peerHashTable_lock);
7941 for (i = 0; i < RX_MAX_SERVICES; i++) {
7943 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7945 for (i = 0; i < rx_hashTableSize; i++) {
7946 struct rx_connection *tc, *ntc;
7947 MUTEX_ENTER(&rx_connHashTable_lock);
7948 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7950 for (j = 0; j < RX_MAXCALLS; j++) {
7952 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7955 rxi_Free(tc, sizeof(*tc));
7957 MUTEX_EXIT(&rx_connHashTable_lock);
7960 MUTEX_ENTER(&freeSQEList_lock);
7962 while ((np = rx_FreeSQEList)) {
7963 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7964 MUTEX_DESTROY(&np->lock);
7965 rxi_Free(np, sizeof(*np));
7968 MUTEX_EXIT(&freeSQEList_lock);
7969 MUTEX_DESTROY(&freeSQEList_lock);
7970 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7971 MUTEX_DESTROY(&rx_connHashTable_lock);
7972 MUTEX_DESTROY(&rx_peerHashTable_lock);
7973 MUTEX_DESTROY(&rx_serverPool_lock);
7975 osi_Free(rx_connHashTable,
7976 rx_hashTableSize * sizeof(struct rx_connection *));
7977 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7979 UNPIN(rx_connHashTable,
7980 rx_hashTableSize * sizeof(struct rx_connection *));
7981 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7983 rxi_FreeAllPackets();
7985 MUTEX_ENTER(&rx_quota_mutex);
7986 rxi_dataQuota = RX_MAX_QUOTA;
7987 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7988 MUTEX_EXIT(&rx_quota_mutex);
7993 #ifdef RX_ENABLE_LOCKS
7995 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7997 if (!MUTEX_ISMINE(lockaddr))
7998 osi_Panic("Lock not held: %s", msg);
8000 #endif /* RX_ENABLE_LOCKS */
8005 * Routines to implement connection specific data.
8009 rx_KeyCreate(rx_destructor_t rtn)
8012 MUTEX_ENTER(&rxi_keyCreate_lock);
8013 key = rxi_keyCreate_counter++;
8014 rxi_keyCreate_destructor = (rx_destructor_t *)
8015 realloc((void *)rxi_keyCreate_destructor,
8016 (key + 1) * sizeof(rx_destructor_t));
8017 rxi_keyCreate_destructor[key] = rtn;
8018 MUTEX_EXIT(&rxi_keyCreate_lock);
8023 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
8026 MUTEX_ENTER(&conn->conn_data_lock);
8027 if (!conn->specific) {
8028 conn->specific = malloc((key + 1) * sizeof(void *));
8029 for (i = 0; i < key; i++)
8030 conn->specific[i] = NULL;
8031 conn->nSpecific = key + 1;
8032 conn->specific[key] = ptr;
8033 } else if (key >= conn->nSpecific) {
8034 conn->specific = (void **)
8035 realloc(conn->specific, (key + 1) * sizeof(void *));
8036 for (i = conn->nSpecific; i < key; i++)
8037 conn->specific[i] = NULL;
8038 conn->nSpecific = key + 1;
8039 conn->specific[key] = ptr;
8041 if (conn->specific[key] && rxi_keyCreate_destructor[key])
8042 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
8043 conn->specific[key] = ptr;
8045 MUTEX_EXIT(&conn->conn_data_lock);
8049 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
8052 MUTEX_ENTER(&svc->svc_data_lock);
8053 if (!svc->specific) {
8054 svc->specific = malloc((key + 1) * sizeof(void *));
8055 for (i = 0; i < key; i++)
8056 svc->specific[i] = NULL;
8057 svc->nSpecific = key + 1;
8058 svc->specific[key] = ptr;
8059 } else if (key >= svc->nSpecific) {
8060 svc->specific = (void **)
8061 realloc(svc->specific, (key + 1) * sizeof(void *));
8062 for (i = svc->nSpecific; i < key; i++)
8063 svc->specific[i] = NULL;
8064 svc->nSpecific = key + 1;
8065 svc->specific[key] = ptr;
8067 if (svc->specific[key] && rxi_keyCreate_destructor[key])
8068 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
8069 svc->specific[key] = ptr;
8071 MUTEX_EXIT(&svc->svc_data_lock);
8075 rx_GetSpecific(struct rx_connection *conn, int key)
8078 MUTEX_ENTER(&conn->conn_data_lock);
8079 if (key >= conn->nSpecific)
8082 ptr = conn->specific[key];
8083 MUTEX_EXIT(&conn->conn_data_lock);
8088 rx_GetServiceSpecific(struct rx_service *svc, int key)
8091 MUTEX_ENTER(&svc->svc_data_lock);
8092 if (key >= svc->nSpecific)
8095 ptr = svc->specific[key];
8096 MUTEX_EXIT(&svc->svc_data_lock);
8101 #endif /* !KERNEL */
8104 * processStats is a queue used to store the statistics for the local
8105 * process. Its contents are similar to the contents of the rpcStats
8106 * queue on a rx_peer structure, but the actual data stored within
8107 * this queue contains totals across the lifetime of the process (assuming
8108 * the stats have not been reset) - unlike the per peer structures
8109 * which can come and go based upon the peer lifetime.
8112 static struct rx_queue processStats = { &processStats, &processStats };
8115 * peerStats is a queue used to store the statistics for all peer structs.
8116 * Its contents are the union of all the peer rpcStats queues.
8119 static struct rx_queue peerStats = { &peerStats, &peerStats };
8122 * rxi_monitor_processStats is used to turn process wide stat collection
8126 static int rxi_monitor_processStats = 0;
8129 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8132 static int rxi_monitor_peerStats = 0;
8136 rxi_ClearRPCOpStat(rx_function_entry_v1_p rpc_stat)
8138 rpc_stat->invocations = 0;
8139 rpc_stat->bytes_sent = 0;
8140 rpc_stat->bytes_rcvd = 0;
8141 rpc_stat->queue_time_sum.sec = 0;
8142 rpc_stat->queue_time_sum.usec = 0;
8143 rpc_stat->queue_time_sum_sqr.sec = 0;
8144 rpc_stat->queue_time_sum_sqr.usec = 0;
8145 rpc_stat->queue_time_min.sec = 9999999;
8146 rpc_stat->queue_time_min.usec = 9999999;
8147 rpc_stat->queue_time_max.sec = 0;
8148 rpc_stat->queue_time_max.usec = 0;
8149 rpc_stat->execution_time_sum.sec = 0;
8150 rpc_stat->execution_time_sum.usec = 0;
8151 rpc_stat->execution_time_sum_sqr.sec = 0;
8152 rpc_stat->execution_time_sum_sqr.usec = 0;
8153 rpc_stat->execution_time_min.sec = 9999999;
8154 rpc_stat->execution_time_min.usec = 9999999;
8155 rpc_stat->execution_time_max.sec = 0;
8156 rpc_stat->execution_time_max.usec = 0;
8160 * Given all of the information for a particular rpc
8161 * call, find or create (if requested) the stat structure for the rpc.
8164 * the queue of stats that will be updated with the new value
8166 * @param rxInterface
8167 * a unique number that identifies the rpc interface
8170 * the total number of functions in this interface. this is only
8171 * required if create is true
8174 * if true, this invocation was made to a server
8177 * the ip address of the remote host. this is only required if create
8178 * and addToPeerList are true
8181 * the port of the remote host. this is only required if create
8182 * and addToPeerList are true
8184 * @param addToPeerList
8185 * if != 0, add newly created stat to the global peer list
8188 * if a new stats structure is allocated, the counter will
8189 * be updated with the new number of allocated stat structures.
8190 * only required if create is true
8193 * if no stats structure exists, allocate one
8197 static rx_interface_stat_p
8198 rxi_FindRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
8199 afs_uint32 totalFunc, int isServer, afs_uint32 remoteHost,
8200 afs_uint32 remotePort, int addToPeerList,
8201 unsigned int *counter, int create)
8203 rx_interface_stat_p rpc_stat, nrpc_stat;
8206 * See if there's already a structure for this interface
8209 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8210 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8211 && (rpc_stat->stats[0].remote_is_server == isServer))
8215 /* if they didn't ask us to create, we're done */
8217 if (queue_IsEnd(stats, rpc_stat))
8223 /* can't proceed without these */
8224 if (!totalFunc || !counter)
8228 * Didn't find a match so allocate a new structure and add it to the
8232 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
8233 || (rpc_stat->stats[0].interfaceId != rxInterface)
8234 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8239 sizeof(rx_interface_stat_t) +
8240 totalFunc * sizeof(rx_function_entry_v1_t);
8242 rpc_stat = rxi_Alloc(space);
8243 if (rpc_stat == NULL)
8246 *counter += totalFunc;
8247 for (i = 0; i < totalFunc; i++) {
8248 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8249 rpc_stat->stats[i].remote_peer = remoteHost;
8250 rpc_stat->stats[i].remote_port = remotePort;
8251 rpc_stat->stats[i].remote_is_server = isServer;
8252 rpc_stat->stats[i].interfaceId = rxInterface;
8253 rpc_stat->stats[i].func_total = totalFunc;
8254 rpc_stat->stats[i].func_index = i;
8256 queue_Prepend(stats, rpc_stat);
8257 if (addToPeerList) {
8258 queue_Prepend(&peerStats, &rpc_stat->all_peers);
8265 rx_ClearProcessRPCStats(afs_int32 rxInterface)
8267 rx_interface_stat_p rpc_stat;
8270 if (rxInterface == -1)
8273 MUTEX_ENTER(&rx_rpc_stats);
8274 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8277 totalFunc = rpc_stat->stats[0].func_total;
8278 for (i = 0; i < totalFunc; i++)
8279 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8281 MUTEX_EXIT(&rx_rpc_stats);
8286 rx_ClearPeerRPCStats(afs_int32 rxInterface, afs_uint32 peerHost, afs_uint16 peerPort)
8288 rx_interface_stat_p rpc_stat;
8290 struct rx_peer * peer;
8292 if (rxInterface == -1)
8295 peer = rxi_FindPeer(peerHost, peerPort, 0, 0);
8299 MUTEX_ENTER(&rx_rpc_stats);
8300 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8303 totalFunc = rpc_stat->stats[0].func_total;
8304 for (i = 0; i < totalFunc; i++)
8305 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8307 MUTEX_EXIT(&rx_rpc_stats);
8312 rx_CopyProcessRPCStats(afs_uint64 op)
8314 rx_interface_stat_p rpc_stat;
8315 rx_function_entry_v1_p rpcop_stat =
8316 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8317 int currentFunc = (op & MAX_AFS_UINT32);
8318 afs_int32 rxInterface = (op >> 32);
8320 if (!rxi_monitor_processStats)
8323 if (rxInterface == -1)
8326 if (rpcop_stat == NULL)
8329 MUTEX_ENTER(&rx_rpc_stats);
8330 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8333 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8334 sizeof(rx_function_entry_v1_t));
8335 MUTEX_EXIT(&rx_rpc_stats);
8337 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8344 rx_CopyPeerRPCStats(afs_uint64 op, afs_uint32 peerHost, afs_uint16 peerPort)
8346 rx_interface_stat_p rpc_stat;
8347 rx_function_entry_v1_p rpcop_stat =
8348 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8349 int currentFunc = (op & MAX_AFS_UINT32);
8350 afs_int32 rxInterface = (op >> 32);
8351 struct rx_peer *peer;
8353 if (!rxi_monitor_peerStats)
8356 if (rxInterface == -1)
8359 if (rpcop_stat == NULL)
8362 peer = rxi_FindPeer(peerHost, peerPort, 0, 0);
8366 MUTEX_ENTER(&rx_rpc_stats);
8367 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8370 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8371 sizeof(rx_function_entry_v1_t));
8372 MUTEX_EXIT(&rx_rpc_stats);
8374 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8381 rx_ReleaseRPCStats(void *stats)
8384 rxi_Free(stats, sizeof(rx_function_entry_v1_t));
8388 * Given all of the information for a particular rpc
8389 * call, create (if needed) and update the stat totals for the rpc.
8392 * the queue of stats that will be updated with the new value
8394 * @param rxInterface
8395 * a unique number that identifies the rpc interface
8397 * @param currentFunc
8398 * the index of the function being invoked
8401 * the total number of functions in this interface
8404 * the amount of time this function waited for a thread
8407 * the amount of time this function invocation took to execute
8410 * the number bytes sent by this invocation
8413 * the number bytes received by this invocation
8416 * if true, this invocation was made to a server
8419 * the ip address of the remote host
8422 * the port of the remote host
8424 * @param addToPeerList
8425 * if != 0, add newly created stat to the global peer list
8428 * if a new stats structure is allocated, the counter will
8429 * be updated with the new number of allocated stat structures
8434 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
8435 afs_uint32 currentFunc, afs_uint32 totalFunc,
8436 struct clock *queueTime, struct clock *execTime,
8437 afs_uint64 bytesSent, afs_uint64 bytesRcvd, int isServer,
8438 afs_uint32 remoteHost, afs_uint32 remotePort,
8439 int addToPeerList, unsigned int *counter)
8442 rx_interface_stat_p rpc_stat;
8444 rpc_stat = rxi_FindRpcStat(stats, rxInterface, totalFunc, isServer,
8445 remoteHost, remotePort, addToPeerList, counter,
8453 * Increment the stats for this function
8456 rpc_stat->stats[currentFunc].invocations++;
8457 rpc_stat->stats[currentFunc].bytes_sent += bytesSent;
8458 rpc_stat->stats[currentFunc].bytes_rcvd += bytesRcvd;
8459 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8460 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8461 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8462 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8464 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8465 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8467 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8468 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8470 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8471 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8473 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8474 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8482 rxi_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8483 afs_uint32 currentFunc, afs_uint32 totalFunc,
8484 struct clock *queueTime, struct clock *execTime,
8485 afs_uint64 bytesSent, afs_uint64 bytesRcvd,
8489 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8492 MUTEX_ENTER(&rx_rpc_stats);
8494 if (rxi_monitor_peerStats) {
8495 MUTEX_ENTER(&peer->peer_lock);
8496 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8497 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8498 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8499 MUTEX_EXIT(&peer->peer_lock);
8502 if (rxi_monitor_processStats) {
8503 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8504 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8505 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8508 MUTEX_EXIT(&rx_rpc_stats);
8512 * Increment the times and count for a particular rpc function.
8514 * Traditionally this call was invoked from rxgen stubs. Modern stubs
8515 * call rx_RecordCallStatistics instead, so the public version of this
8516 * function is left purely for legacy callers.
8519 * The peer who invoked the rpc
8521 * @param rxInterface
8522 * A unique number that identifies the rpc interface
8524 * @param currentFunc
8525 * The index of the function being invoked
8528 * The total number of functions in this interface
8531 * The amount of time this function waited for a thread
8534 * The amount of time this function invocation took to execute
8537 * The number bytes sent by this invocation
8540 * The number bytes received by this invocation
8543 * If true, this invocation was made to a server
8547 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8548 afs_uint32 currentFunc, afs_uint32 totalFunc,
8549 struct clock *queueTime, struct clock *execTime,
8550 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8556 sent64 = ((afs_uint64)bytesSent->high << 32) + bytesSent->low;
8557 rcvd64 = ((afs_uint64)bytesRcvd->high << 32) + bytesRcvd->low;
8559 rxi_IncrementTimeAndCount(peer, rxInterface, currentFunc, totalFunc,
8560 queueTime, execTime, sent64, rcvd64,
8567 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8571 * IN callerVersion - the rpc stat version of the caller.
8573 * IN count - the number of entries to marshall.
8575 * IN stats - pointer to stats to be marshalled.
8577 * OUT ptr - Where to store the marshalled data.
8584 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8585 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8591 * We only support the first version
8593 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8594 *(ptr++) = stats->remote_peer;
8595 *(ptr++) = stats->remote_port;
8596 *(ptr++) = stats->remote_is_server;
8597 *(ptr++) = stats->interfaceId;
8598 *(ptr++) = stats->func_total;
8599 *(ptr++) = stats->func_index;
8600 *(ptr++) = stats->invocations >> 32;
8601 *(ptr++) = stats->invocations & MAX_AFS_UINT32;
8602 *(ptr++) = stats->bytes_sent >> 32;
8603 *(ptr++) = stats->bytes_sent & MAX_AFS_UINT32;
8604 *(ptr++) = stats->bytes_rcvd >> 32;
8605 *(ptr++) = stats->bytes_rcvd & MAX_AFS_UINT32;
8606 *(ptr++) = stats->queue_time_sum.sec;
8607 *(ptr++) = stats->queue_time_sum.usec;
8608 *(ptr++) = stats->queue_time_sum_sqr.sec;
8609 *(ptr++) = stats->queue_time_sum_sqr.usec;
8610 *(ptr++) = stats->queue_time_min.sec;
8611 *(ptr++) = stats->queue_time_min.usec;
8612 *(ptr++) = stats->queue_time_max.sec;
8613 *(ptr++) = stats->queue_time_max.usec;
8614 *(ptr++) = stats->execution_time_sum.sec;
8615 *(ptr++) = stats->execution_time_sum.usec;
8616 *(ptr++) = stats->execution_time_sum_sqr.sec;
8617 *(ptr++) = stats->execution_time_sum_sqr.usec;
8618 *(ptr++) = stats->execution_time_min.sec;
8619 *(ptr++) = stats->execution_time_min.usec;
8620 *(ptr++) = stats->execution_time_max.sec;
8621 *(ptr++) = stats->execution_time_max.usec;
8627 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8632 * IN callerVersion - the rpc stat version of the caller
8634 * OUT myVersion - the rpc stat version of this function
8636 * OUT clock_sec - local time seconds
8638 * OUT clock_usec - local time microseconds
8640 * OUT allocSize - the number of bytes allocated to contain stats
8642 * OUT statCount - the number stats retrieved from this process.
8644 * OUT stats - the actual stats retrieved from this process.
8648 * Returns void. If successful, stats will != NULL.
8652 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8653 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8654 size_t * allocSize, afs_uint32 * statCount,
8655 afs_uint32 ** stats)
8665 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8668 * Check to see if stats are enabled
8671 MUTEX_ENTER(&rx_rpc_stats);
8672 if (!rxi_monitor_processStats) {
8673 MUTEX_EXIT(&rx_rpc_stats);
8677 clock_GetTime(&now);
8678 *clock_sec = now.sec;
8679 *clock_usec = now.usec;
8682 * Allocate the space based upon the caller version
8684 * If the client is at an older version than we are,
8685 * we return the statistic data in the older data format, but
8686 * we still return our version number so the client knows we
8687 * are maintaining more data than it can retrieve.
8690 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8691 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8692 *statCount = rxi_rpc_process_stat_cnt;
8695 * This can't happen yet, but in the future version changes
8696 * can be handled by adding additional code here
8700 if (space > (size_t) 0) {
8702 ptr = *stats = rxi_Alloc(space);
8705 rx_interface_stat_p rpc_stat, nrpc_stat;
8709 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8711 * Copy the data based upon the caller version
8713 rx_MarshallProcessRPCStats(callerVersion,
8714 rpc_stat->stats[0].func_total,
8715 rpc_stat->stats, &ptr);
8721 MUTEX_EXIT(&rx_rpc_stats);
8726 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8730 * IN callerVersion - the rpc stat version of the caller
8732 * OUT myVersion - the rpc stat version of this function
8734 * OUT clock_sec - local time seconds
8736 * OUT clock_usec - local time microseconds
8738 * OUT allocSize - the number of bytes allocated to contain stats
8740 * OUT statCount - the number of stats retrieved from the individual
8743 * OUT stats - the actual stats retrieved from the individual peer structures.
8747 * Returns void. If successful, stats will != NULL.
8751 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8752 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8753 size_t * allocSize, afs_uint32 * statCount,
8754 afs_uint32 ** stats)
8764 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8767 * Check to see if stats are enabled
8770 MUTEX_ENTER(&rx_rpc_stats);
8771 if (!rxi_monitor_peerStats) {
8772 MUTEX_EXIT(&rx_rpc_stats);
8776 clock_GetTime(&now);
8777 *clock_sec = now.sec;
8778 *clock_usec = now.usec;
8781 * Allocate the space based upon the caller version
8783 * If the client is at an older version than we are,
8784 * we return the statistic data in the older data format, but
8785 * we still return our version number so the client knows we
8786 * are maintaining more data than it can retrieve.
8789 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8790 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8791 *statCount = rxi_rpc_peer_stat_cnt;
8794 * This can't happen yet, but in the future version changes
8795 * can be handled by adding additional code here
8799 if (space > (size_t) 0) {
8801 ptr = *stats = rxi_Alloc(space);
8804 rx_interface_stat_p rpc_stat, nrpc_stat;
8808 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8810 * We have to fix the offset of rpc_stat since we are
8811 * keeping this structure on two rx_queues. The rx_queue
8812 * package assumes that the rx_queue member is the first
8813 * member of the structure. That is, rx_queue assumes that
8814 * any one item is only on one queue at a time. We are
8815 * breaking that assumption and so we have to do a little
8816 * math to fix our pointers.
8819 fix_offset = (char *)rpc_stat;
8820 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8821 rpc_stat = (rx_interface_stat_p) fix_offset;
8824 * Copy the data based upon the caller version
8826 rx_MarshallProcessRPCStats(callerVersion,
8827 rpc_stat->stats[0].func_total,
8828 rpc_stat->stats, &ptr);
8834 MUTEX_EXIT(&rx_rpc_stats);
8839 * rx_FreeRPCStats - free memory allocated by
8840 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8844 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8845 * rx_RetrievePeerRPCStats
8847 * IN allocSize - the number of bytes in stats.
8855 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8857 rxi_Free(stats, allocSize);
8861 * rx_queryProcessRPCStats - see if process rpc stat collection is
8862 * currently enabled.
8868 * Returns 0 if stats are not enabled != 0 otherwise
8872 rx_queryProcessRPCStats(void)
8875 MUTEX_ENTER(&rx_rpc_stats);
8876 rc = rxi_monitor_processStats;
8877 MUTEX_EXIT(&rx_rpc_stats);
8882 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8888 * Returns 0 if stats are not enabled != 0 otherwise
8892 rx_queryPeerRPCStats(void)
8895 MUTEX_ENTER(&rx_rpc_stats);
8896 rc = rxi_monitor_peerStats;
8897 MUTEX_EXIT(&rx_rpc_stats);
8902 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8912 rx_enableProcessRPCStats(void)
8914 MUTEX_ENTER(&rx_rpc_stats);
8915 rx_enable_stats = 1;
8916 rxi_monitor_processStats = 1;
8917 MUTEX_EXIT(&rx_rpc_stats);
8921 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8931 rx_enablePeerRPCStats(void)
8933 MUTEX_ENTER(&rx_rpc_stats);
8934 rx_enable_stats = 1;
8935 rxi_monitor_peerStats = 1;
8936 MUTEX_EXIT(&rx_rpc_stats);
8940 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8950 rx_disableProcessRPCStats(void)
8952 rx_interface_stat_p rpc_stat, nrpc_stat;
8955 MUTEX_ENTER(&rx_rpc_stats);
8958 * Turn off process statistics and if peer stats is also off, turn
8962 rxi_monitor_processStats = 0;
8963 if (rxi_monitor_peerStats == 0) {
8964 rx_enable_stats = 0;
8967 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8968 unsigned int num_funcs = 0;
8971 queue_Remove(rpc_stat);
8972 num_funcs = rpc_stat->stats[0].func_total;
8974 sizeof(rx_interface_stat_t) +
8975 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8977 rxi_Free(rpc_stat, space);
8978 rxi_rpc_process_stat_cnt -= num_funcs;
8980 MUTEX_EXIT(&rx_rpc_stats);
8984 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8994 rx_disablePeerRPCStats(void)
8996 struct rx_peer **peer_ptr, **peer_end;
9000 * Turn off peer statistics and if process stats is also off, turn
9004 rxi_monitor_peerStats = 0;
9005 if (rxi_monitor_processStats == 0) {
9006 rx_enable_stats = 0;
9009 for (peer_ptr = &rx_peerHashTable[0], peer_end =
9010 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
9012 struct rx_peer *peer, *next, *prev;
9014 MUTEX_ENTER(&rx_peerHashTable_lock);
9015 MUTEX_ENTER(&rx_rpc_stats);
9016 for (prev = peer = *peer_ptr; peer; peer = next) {
9018 code = MUTEX_TRYENTER(&peer->peer_lock);
9020 rx_interface_stat_p rpc_stat, nrpc_stat;
9023 if (prev == *peer_ptr) {
9034 MUTEX_EXIT(&rx_peerHashTable_lock);
9037 (&peer->rpcStats, rpc_stat, nrpc_stat,
9038 rx_interface_stat)) {
9039 unsigned int num_funcs = 0;
9042 queue_Remove(&rpc_stat->queue_header);
9043 queue_Remove(&rpc_stat->all_peers);
9044 num_funcs = rpc_stat->stats[0].func_total;
9046 sizeof(rx_interface_stat_t) +
9047 rpc_stat->stats[0].func_total *
9048 sizeof(rx_function_entry_v1_t);
9050 rxi_Free(rpc_stat, space);
9051 rxi_rpc_peer_stat_cnt -= num_funcs;
9053 MUTEX_EXIT(&peer->peer_lock);
9055 MUTEX_ENTER(&rx_peerHashTable_lock);
9065 MUTEX_EXIT(&rx_rpc_stats);
9066 MUTEX_EXIT(&rx_peerHashTable_lock);
9071 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
9076 * IN clearFlag - flag indicating which stats to clear
9084 rx_clearProcessRPCStats(afs_uint32 clearFlag)
9086 rx_interface_stat_p rpc_stat, nrpc_stat;
9088 MUTEX_ENTER(&rx_rpc_stats);
9090 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
9091 unsigned int num_funcs = 0, i;
9092 num_funcs = rpc_stat->stats[0].func_total;
9093 for (i = 0; i < num_funcs; i++) {
9094 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9095 rpc_stat->stats[i].invocations = 0;
9097 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9098 rpc_stat->stats[i].bytes_sent = 0;
9100 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9101 rpc_stat->stats[i].bytes_rcvd = 0;
9103 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9104 rpc_stat->stats[i].queue_time_sum.sec = 0;
9105 rpc_stat->stats[i].queue_time_sum.usec = 0;
9107 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9108 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9109 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9111 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9112 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9113 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9115 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9116 rpc_stat->stats[i].queue_time_max.sec = 0;
9117 rpc_stat->stats[i].queue_time_max.usec = 0;
9119 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9120 rpc_stat->stats[i].execution_time_sum.sec = 0;
9121 rpc_stat->stats[i].execution_time_sum.usec = 0;
9123 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9124 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9125 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9127 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9128 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9129 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9131 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9132 rpc_stat->stats[i].execution_time_max.sec = 0;
9133 rpc_stat->stats[i].execution_time_max.usec = 0;
9138 MUTEX_EXIT(&rx_rpc_stats);
9142 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
9147 * IN clearFlag - flag indicating which stats to clear
9155 rx_clearPeerRPCStats(afs_uint32 clearFlag)
9157 rx_interface_stat_p rpc_stat, nrpc_stat;
9159 MUTEX_ENTER(&rx_rpc_stats);
9161 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
9162 unsigned int num_funcs = 0, i;
9165 * We have to fix the offset of rpc_stat since we are
9166 * keeping this structure on two rx_queues. The rx_queue
9167 * package assumes that the rx_queue member is the first
9168 * member of the structure. That is, rx_queue assumes that
9169 * any one item is only on one queue at a time. We are
9170 * breaking that assumption and so we have to do a little
9171 * math to fix our pointers.
9174 fix_offset = (char *)rpc_stat;
9175 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
9176 rpc_stat = (rx_interface_stat_p) fix_offset;
9178 num_funcs = rpc_stat->stats[0].func_total;
9179 for (i = 0; i < num_funcs; i++) {
9180 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9181 rpc_stat->stats[i].invocations = 0;
9183 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9184 rpc_stat->stats[i].bytes_sent = 0;
9186 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9187 rpc_stat->stats[i].bytes_rcvd = 0;
9189 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9190 rpc_stat->stats[i].queue_time_sum.sec = 0;
9191 rpc_stat->stats[i].queue_time_sum.usec = 0;
9193 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9194 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9195 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9197 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9198 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9199 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9201 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9202 rpc_stat->stats[i].queue_time_max.sec = 0;
9203 rpc_stat->stats[i].queue_time_max.usec = 0;
9205 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9206 rpc_stat->stats[i].execution_time_sum.sec = 0;
9207 rpc_stat->stats[i].execution_time_sum.usec = 0;
9209 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9210 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9211 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9213 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9214 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9215 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9217 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9218 rpc_stat->stats[i].execution_time_max.sec = 0;
9219 rpc_stat->stats[i].execution_time_max.usec = 0;
9224 MUTEX_EXIT(&rx_rpc_stats);
9228 * rxi_rxstat_userok points to a routine that returns 1 if the caller
9229 * is authorized to enable/disable/clear RX statistics.
9231 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
9234 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
9236 rxi_rxstat_userok = proc;
9240 rx_RxStatUserOk(struct rx_call *call)
9242 if (!rxi_rxstat_userok)
9244 return rxi_rxstat_userok(call);
9249 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
9250 * function in the MSVC runtime DLL (msvcrt.dll).
9252 * Note: the system serializes calls to this function.
9255 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
9256 DWORD reason, /* reason function is being called */
9257 LPVOID reserved) /* reserved for future use */
9260 case DLL_PROCESS_ATTACH:
9261 /* library is being attached to a process */
9265 case DLL_PROCESS_DETACH:
9272 #endif /* AFS_NT40_ENV */
9275 int rx_DumpCalls(FILE *outputFile, char *cookie)
9277 #ifdef RXDEBUG_PACKET
9278 #ifdef KDUMP_RX_LOCK
9279 struct rx_call_rx_lock *c;
9286 #define RXDPRINTF sprintf
9287 #define RXDPRINTOUT output
9289 #define RXDPRINTF fprintf
9290 #define RXDPRINTOUT outputFile
9293 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9295 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9298 for (c = rx_allCallsp; c; c = c->allNextp) {
9299 u_short rqc, tqc, iovqc;
9300 struct rx_packet *p, *np;
9302 MUTEX_ENTER(&c->lock);
9303 queue_Count(&c->rq, p, np, rx_packet, rqc);
9304 queue_Count(&c->tq, p, np, rx_packet, tqc);
9305 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
9307 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, "
9308 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9309 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9310 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9311 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
9312 #ifdef RX_ENABLE_LOCKS
9315 #ifdef RX_REFCOUNT_CHECK
9316 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9317 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9320 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,
9321 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9322 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9323 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9324 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9325 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
9326 #ifdef RX_ENABLE_LOCKS
9327 , (afs_uint32)c->refCount
9329 #ifdef RX_REFCOUNT_CHECK
9330 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9333 MUTEX_EXIT(&c->lock);
9336 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9339 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9341 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9343 #endif /* RXDEBUG_PACKET */