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>
76 #include "rx_atomic.h"
77 #include "rx_globals.h"
79 #include "rx_internal.h"
86 #include <afs/rxgen_consts.h>
89 #ifdef AFS_PTHREAD_ENV
91 int (*registerProgram) (pid_t, char *) = 0;
92 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
95 int (*registerProgram) (PROCESS, char *) = 0;
96 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
100 /* Local static routines */
101 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
102 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
103 struct rx_call *, struct rx_peer *,
105 static void rxi_Resend(struct rxevent *event, void *arg0, void *arg1,
107 static void rxi_SendDelayedAck(struct rxevent *event, void *call,
108 void *dummy, int dummy2);
109 static void rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1,
110 void *dummy, int dummy2);
111 static void rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1,
112 void *unused, int unused2);
113 static void rxi_ReapConnections(struct rxevent *unused, void *unused1,
114 void *unused2, int unused3);
116 #ifdef RX_ENABLE_LOCKS
117 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
120 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
122 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
123 rx_atomic_t rxi_start_in_error;
125 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
127 /* Constant delay time before sending an acknowledge of the last packet
128 * received. This is to avoid sending an extra acknowledge when the
129 * client is about to make another call, anyway, or the server is
132 * The lastAckDelay may not exceeed 400ms without causing peers to
133 * unecessarily timeout.
135 struct clock rx_lastAckDelay = {0, 400000};
137 /* Constant delay time before sending a soft ack when none was requested.
138 * This is to make sure we send soft acks before the sender times out,
139 * Normally we wait and send a hard ack when the receiver consumes the packet
141 * This value has been 100ms in all shipping versions of OpenAFS. Changing it
142 * will require changes to the peer's RTT calculations.
144 struct clock rx_softAckDelay = {0, 100000};
147 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
148 * currently allocated within rx. This number is used to allocate the
149 * memory required to return the statistics when queried.
150 * Protected by the rx_rpc_stats mutex.
153 static unsigned int rxi_rpc_peer_stat_cnt;
156 * rxi_rpc_process_stat_cnt counts the total number of local process stat
157 * structures currently allocated within rx. The number is used to allocate
158 * the memory required to return the statistics when queried.
159 * Protected by the rx_rpc_stats mutex.
162 static unsigned int rxi_rpc_process_stat_cnt;
165 * rxi_busyChannelError is the error to return to the application when a call
166 * channel appears busy (inferred from the receipt of RX_PACKET_TYPE_BUSY
167 * packets on the channel), and there are other call channels in the
168 * connection that are not busy. If 0, we do not return errors upon receiving
169 * busy packets; we just keep trying on the same call channel until we hit a
172 static afs_int32 rxi_busyChannelError = 0;
174 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
175 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
177 #if !defined(offsetof)
178 #include <stddef.h> /* for definition of offsetof() */
181 #ifdef RX_ENABLE_LOCKS
182 afs_kmutex_t rx_atomic_mutex;
185 /* Forward prototypes */
186 static struct rx_call * rxi_NewCall(struct rx_connection *, int);
188 #ifdef AFS_PTHREAD_ENV
191 * Use procedural initialization of mutexes/condition variables
195 extern afs_kmutex_t rx_quota_mutex;
196 extern afs_kmutex_t rx_pthread_mutex;
197 extern afs_kmutex_t rx_packets_mutex;
198 extern afs_kmutex_t rx_refcnt_mutex;
199 extern afs_kmutex_t des_init_mutex;
200 extern afs_kmutex_t des_random_mutex;
201 extern afs_kmutex_t rx_clock_mutex;
202 extern afs_kmutex_t rxi_connCacheMutex;
203 extern afs_kmutex_t event_handler_mutex;
204 extern afs_kmutex_t listener_mutex;
205 extern afs_kmutex_t rx_if_init_mutex;
206 extern afs_kmutex_t rx_if_mutex;
208 extern afs_kcondvar_t rx_event_handler_cond;
209 extern afs_kcondvar_t rx_listener_cond;
211 static afs_kmutex_t epoch_mutex;
212 static afs_kmutex_t rx_init_mutex;
213 static afs_kmutex_t rx_debug_mutex;
214 static afs_kmutex_t rx_rpc_stats;
217 rxi_InitPthread(void)
219 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
220 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
221 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
222 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
223 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
224 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
225 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
226 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
227 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
228 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
229 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
230 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
231 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
232 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
233 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
235 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
236 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
238 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
239 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
241 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
242 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
243 #ifdef RX_ENABLE_LOCKS
246 #endif /* RX_LOCKS_DB */
247 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
248 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
250 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
252 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
254 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
256 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
257 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
258 #endif /* RX_ENABLE_LOCKS */
261 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
262 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
264 * The rx_stats_mutex mutex protects the following global variables:
265 * rxi_lowConnRefCount
266 * rxi_lowPeerRefCount
275 * The rx_quota_mutex mutex protects the following global variables:
283 * The rx_freePktQ_lock protects the following global variables:
288 * The rx_packets_mutex mutex protects the following global variables:
296 * The rx_pthread_mutex mutex protects the following global variables:
297 * rxi_fcfs_thread_num
300 #define INIT_PTHREAD_LOCKS
304 /* Variables for handling the minProcs implementation. availProcs gives the
305 * number of threads available in the pool at this moment (not counting dudes
306 * executing right now). totalMin gives the total number of procs required
307 * for handling all minProcs requests. minDeficit is a dynamic variable
308 * tracking the # of procs required to satisfy all of the remaining minProcs
310 * For fine grain locking to work, the quota check and the reservation of
311 * a server thread has to come while rxi_availProcs and rxi_minDeficit
312 * are locked. To this end, the code has been modified under #ifdef
313 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
314 * same time. A new function, ReturnToServerPool() returns the allocation.
316 * A call can be on several queue's (but only one at a time). When
317 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
318 * that no one else is touching the queue. To this end, we store the address
319 * of the queue lock in the call structure (under the call lock) when we
320 * put the call on a queue, and we clear the call_queue_lock when the
321 * call is removed from a queue (once the call lock has been obtained).
322 * This allows rxi_ResetCall to safely synchronize with others wishing
323 * to manipulate the queue.
326 #if defined(RX_ENABLE_LOCKS)
327 static afs_kmutex_t rx_rpc_stats;
330 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
331 ** pretty good that the next packet coming in is from the same connection
332 ** as the last packet, since we're send multiple packets in a transmit window.
334 struct rx_connection *rxLastConn = 0;
336 #ifdef RX_ENABLE_LOCKS
337 /* The locking hierarchy for rx fine grain locking is composed of these
340 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
341 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
342 * call->lock - locks call data fields.
343 * These are independent of each other:
344 * rx_freeCallQueue_lock
349 * serverQueueEntry->lock
350 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
352 * peer->lock - locks peer data fields.
353 * conn_data_lock - that more than one thread is not updating a conn data
354 * field at the same time.
365 * Do we need a lock to protect the peer field in the conn structure?
366 * conn->peer was previously a constant for all intents and so has no
367 * lock protecting this field. The multihomed client delta introduced
368 * a RX code change : change the peer field in the connection structure
369 * to that remote interface from which the last packet for this
370 * connection was sent out. This may become an issue if further changes
373 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
374 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
376 /* rxdb_fileID is used to identify the lock location, along with line#. */
377 static int rxdb_fileID = RXDB_FILE_RX;
378 #endif /* RX_LOCKS_DB */
379 #else /* RX_ENABLE_LOCKS */
380 #define SET_CALL_QUEUE_LOCK(C, L)
381 #define CLEAR_CALL_QUEUE_LOCK(C)
382 #endif /* RX_ENABLE_LOCKS */
383 struct rx_serverQueueEntry *rx_waitForPacket = 0;
384 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
386 /* ------------Exported Interfaces------------- */
388 /* This function allows rxkad to set the epoch to a suitably random number
389 * which rx_NewConnection will use in the future. The principle purpose is to
390 * get rxnull connections to use the same epoch as the rxkad connections do, at
391 * least once the first rxkad connection is established. This is important now
392 * that the host/port addresses aren't used in FindConnection: the uniqueness
393 * of epoch/cid matters and the start time won't do. */
395 #ifdef AFS_PTHREAD_ENV
397 * This mutex protects the following global variables:
401 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
402 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
406 #endif /* AFS_PTHREAD_ENV */
409 rx_SetEpoch(afs_uint32 epoch)
416 /* Initialize rx. A port number may be mentioned, in which case this
417 * becomes the default port number for any service installed later.
418 * If 0 is provided for the port number, a random port will be chosen
419 * by the kernel. Whether this will ever overlap anything in
420 * /etc/services is anybody's guess... Returns 0 on success, -1 on
425 int rxinit_status = 1;
426 #ifdef AFS_PTHREAD_ENV
428 * This mutex protects the following global variables:
432 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
433 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
436 #define UNLOCK_RX_INIT
440 rx_InitHost(u_int host, u_int port)
447 char *htable, *ptable;
454 if (rxinit_status == 0) {
455 tmp_status = rxinit_status;
457 return tmp_status; /* Already started; return previous error code. */
463 if (afs_winsockInit() < 0)
469 * Initialize anything necessary to provide a non-premptive threading
472 rxi_InitializeThreadSupport();
475 /* Allocate and initialize a socket for client and perhaps server
478 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
479 if (rx_socket == OSI_NULLSOCKET) {
483 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
486 #endif /* RX_LOCKS_DB */
487 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
488 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
489 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
490 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
491 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
492 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
493 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
494 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
495 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
497 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
499 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
501 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
503 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
504 #if defined(AFS_HPUX110_ENV)
506 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
507 #endif /* AFS_HPUX110_ENV */
508 #endif /* RX_ENABLE_LOCKS && KERNEL */
511 rx_connDeadTime = 12;
512 rx_tranquil = 0; /* reset flag */
513 rxi_ResetStatistics();
515 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
516 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
517 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
518 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
519 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
520 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
522 /* Malloc up a bunch of packets & buffers */
524 queue_Init(&rx_freePacketQueue);
525 rxi_NeedMorePackets = FALSE;
526 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
528 /* enforce a minimum number of allocated packets */
529 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
530 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
532 /* allocate the initial free packet pool */
533 #ifdef RX_ENABLE_TSFPQ
534 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
535 #else /* RX_ENABLE_TSFPQ */
536 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
537 #endif /* RX_ENABLE_TSFPQ */
544 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
545 tv.tv_sec = clock_now.sec;
546 tv.tv_usec = clock_now.usec;
547 srand((unsigned int)tv.tv_usec);
554 #if defined(KERNEL) && !defined(UKERNEL)
555 /* Really, this should never happen in a real kernel */
558 struct sockaddr_in addr;
560 int addrlen = sizeof(addr);
562 socklen_t addrlen = sizeof(addr);
564 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
568 rx_port = addr.sin_port;
571 rx_stats.minRtt.sec = 9999999;
573 rx_SetEpoch(tv.tv_sec | 0x80000000);
575 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
576 * will provide a randomer value. */
578 MUTEX_ENTER(&rx_quota_mutex);
579 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
580 MUTEX_EXIT(&rx_quota_mutex);
581 /* *Slightly* random start time for the cid. This is just to help
582 * out with the hashing function at the peer */
583 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
584 rx_connHashTable = (struct rx_connection **)htable;
585 rx_peerHashTable = (struct rx_peer **)ptable;
587 rx_hardAckDelay.sec = 0;
588 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
590 rxevent_Init(20, rxi_ReScheduleEvents);
592 /* Initialize various global queues */
593 queue_Init(&rx_idleServerQueue);
594 queue_Init(&rx_incomingCallQueue);
595 queue_Init(&rx_freeCallQueue);
597 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
598 /* Initialize our list of usable IP addresses. */
602 #if defined(RXK_LISTENER_ENV) || !defined(KERNEL)
603 /* Start listener process (exact function is dependent on the
604 * implementation environment--kernel or user space) */
609 tmp_status = rxinit_status = 0;
617 return rx_InitHost(htonl(INADDR_ANY), port);
623 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
624 * maintaing the round trip timer.
629 * Start a new RTT timer for a given call and packet.
631 * There must be no resendEvent already listed for this call, otherwise this
632 * will leak events - intended for internal use within the RTO code only
635 * the RX call to start the timer for
636 * @param[in] lastPacket
637 * a flag indicating whether the last packet has been sent or not
639 * @pre call must be locked before calling this function
643 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
645 struct clock now, retryTime;
650 clock_Add(&retryTime, &call->rto);
652 /* If we're sending the last packet, and we're the client, then the server
653 * may wait for an additional 400ms before returning the ACK, wait for it
654 * rather than hitting a timeout */
655 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
656 clock_Addmsec(&retryTime, 400);
658 MUTEX_ENTER(&rx_refcnt_mutex);
659 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
660 MUTEX_EXIT(&rx_refcnt_mutex);
661 call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
666 * Cancel an RTT timer for a given call.
670 * the RX call to cancel the timer for
672 * @pre call must be locked before calling this function
677 rxi_rto_cancel(struct rx_call *call)
679 rxevent_Cancel(&call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
683 * Tell the RTO timer that we have sent a packet.
685 * If the timer isn't already running, then start it. If the timer is running,
689 * the RX call that the packet has been sent on
690 * @param[in] lastPacket
691 * A flag which is true if this is the last packet for the call
693 * @pre The call must be locked before calling this function
698 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
700 if (call->resendEvent)
703 rxi_rto_startTimer(call, lastPacket, istack);
707 * Tell the RTO timer that we have received an new ACK message
709 * This function should be called whenever a call receives an ACK that
710 * acknowledges new packets. Whatever happens, we stop the current timer.
711 * If there are unacked packets in the queue which have been sent, then
712 * we restart the timer from now. Otherwise, we leave it stopped.
715 * the RX call that the ACK has been received on
719 rxi_rto_packet_acked(struct rx_call *call, int istack)
721 struct rx_packet *p, *nxp;
723 rxi_rto_cancel(call);
725 if (queue_IsEmpty(&call->tq))
728 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
729 if (p->header.seq > call->tfirst + call->twind)
732 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
733 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
741 * Set an initial round trip timeout for a peer connection
743 * @param[in] secs The timeout to set in seconds
747 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
748 peer->rtt = secs * 8000;
752 * Sets the error generated when a busy call channel is detected.
754 * @param[in] error The error to return for a call on a busy channel.
756 * @pre Neither rx_Init nor rx_InitHost have been called yet
759 rx_SetBusyChannelError(afs_int32 error)
761 osi_Assert(rxinit_status != 0);
762 rxi_busyChannelError = error;
766 * Set a delayed ack event on the specified call for the given time
768 * @param[in] call - the call on which to set the event
769 * @param[in] offset - the delay from now after which the event fires
772 rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
774 struct clock now, when;
778 clock_Add(&when, offset);
780 if (!call->delayedAckEvent
781 || clock_Gt(&call->delayedAckTime, &when)) {
783 rxevent_Cancel(&call->delayedAckEvent, call,
784 RX_CALL_REFCOUNT_DELAY);
785 MUTEX_ENTER(&rx_refcnt_mutex);
786 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
787 MUTEX_EXIT(&rx_refcnt_mutex);
789 call->delayedAckEvent = rxevent_Post(&when, &now,
792 call->delayedAckTime = when;
796 /* called with unincremented nRequestsRunning to see if it is OK to start
797 * a new thread in this service. Could be "no" for two reasons: over the
798 * max quota, or would prevent others from reaching their min quota.
800 #ifdef RX_ENABLE_LOCKS
801 /* This verion of QuotaOK reserves quota if it's ok while the
802 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
805 QuotaOK(struct rx_service *aservice)
807 /* check if over max quota */
808 if (aservice->nRequestsRunning >= aservice->maxProcs) {
812 /* under min quota, we're OK */
813 /* otherwise, can use only if there are enough to allow everyone
814 * to go to their min quota after this guy starts.
817 MUTEX_ENTER(&rx_quota_mutex);
818 if ((aservice->nRequestsRunning < aservice->minProcs)
819 || (rxi_availProcs > rxi_minDeficit)) {
820 aservice->nRequestsRunning++;
821 /* just started call in minProcs pool, need fewer to maintain
823 if (aservice->nRequestsRunning <= aservice->minProcs)
826 MUTEX_EXIT(&rx_quota_mutex);
829 MUTEX_EXIT(&rx_quota_mutex);
835 ReturnToServerPool(struct rx_service *aservice)
837 aservice->nRequestsRunning--;
838 MUTEX_ENTER(&rx_quota_mutex);
839 if (aservice->nRequestsRunning < aservice->minProcs)
842 MUTEX_EXIT(&rx_quota_mutex);
845 #else /* RX_ENABLE_LOCKS */
847 QuotaOK(struct rx_service *aservice)
850 /* under min quota, we're OK */
851 if (aservice->nRequestsRunning < aservice->minProcs)
854 /* check if over max quota */
855 if (aservice->nRequestsRunning >= aservice->maxProcs)
858 /* otherwise, can use only if there are enough to allow everyone
859 * to go to their min quota after this guy starts.
861 MUTEX_ENTER(&rx_quota_mutex);
862 if (rxi_availProcs > rxi_minDeficit)
864 MUTEX_EXIT(&rx_quota_mutex);
867 #endif /* RX_ENABLE_LOCKS */
870 /* Called by rx_StartServer to start up lwp's to service calls.
871 NExistingProcs gives the number of procs already existing, and which
872 therefore needn't be created. */
874 rxi_StartServerProcs(int nExistingProcs)
876 struct rx_service *service;
881 /* For each service, reserve N processes, where N is the "minimum"
882 * number of processes that MUST be able to execute a request in parallel,
883 * at any time, for that process. Also compute the maximum difference
884 * between any service's maximum number of processes that can run
885 * (i.e. the maximum number that ever will be run, and a guarantee
886 * that this number will run if other services aren't running), and its
887 * minimum number. The result is the extra number of processes that
888 * we need in order to provide the latter guarantee */
889 for (i = 0; i < RX_MAX_SERVICES; i++) {
891 service = rx_services[i];
892 if (service == (struct rx_service *)0)
894 nProcs += service->minProcs;
895 diff = service->maxProcs - service->minProcs;
899 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
900 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
901 for (i = 0; i < nProcs; i++) {
902 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
908 /* This routine is only required on Windows */
910 rx_StartClientThread(void)
912 #ifdef AFS_PTHREAD_ENV
914 pid = pthread_self();
915 #endif /* AFS_PTHREAD_ENV */
917 #endif /* AFS_NT40_ENV */
919 /* This routine must be called if any services are exported. If the
920 * donateMe flag is set, the calling process is donated to the server
923 rx_StartServer(int donateMe)
925 struct rx_service *service;
931 /* Start server processes, if necessary (exact function is dependent
932 * on the implementation environment--kernel or user space). DonateMe
933 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
934 * case, one less new proc will be created rx_StartServerProcs.
936 rxi_StartServerProcs(donateMe);
938 /* count up the # of threads in minProcs, and add set the min deficit to
939 * be that value, too.
941 for (i = 0; i < RX_MAX_SERVICES; i++) {
942 service = rx_services[i];
943 if (service == (struct rx_service *)0)
945 MUTEX_ENTER(&rx_quota_mutex);
946 rxi_totalMin += service->minProcs;
947 /* below works even if a thread is running, since minDeficit would
948 * still have been decremented and later re-incremented.
950 rxi_minDeficit += service->minProcs;
951 MUTEX_EXIT(&rx_quota_mutex);
954 /* Turn on reaping of idle server connections */
955 rxi_ReapConnections(NULL, NULL, NULL, 0);
964 #ifdef AFS_PTHREAD_ENV
966 pid = afs_pointer_to_int(pthread_self());
967 #else /* AFS_PTHREAD_ENV */
969 LWP_CurrentProcess(&pid);
970 #endif /* AFS_PTHREAD_ENV */
972 sprintf(name, "srv_%d", ++nProcs);
974 (*registerProgram) (pid, name);
976 #endif /* AFS_NT40_ENV */
977 rx_ServerProc(NULL); /* Never returns */
979 #ifdef RX_ENABLE_TSFPQ
980 /* no use leaving packets around in this thread's local queue if
981 * it isn't getting donated to the server thread pool.
983 rxi_FlushLocalPacketsTSFPQ();
984 #endif /* RX_ENABLE_TSFPQ */
988 /* Create a new client connection to the specified service, using the
989 * specified security object to implement the security model for this
991 struct rx_connection *
992 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
993 struct rx_securityClass *securityObject,
994 int serviceSecurityIndex)
998 struct rx_connection *conn;
1003 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
1004 "serviceSecurityIndex %d)\n",
1005 ntohl(shost), ntohs(sport), sservice, securityObject,
1006 serviceSecurityIndex));
1008 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
1009 * the case of kmem_alloc? */
1010 conn = rxi_AllocConnection();
1011 #ifdef RX_ENABLE_LOCKS
1012 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
1013 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
1014 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
1017 MUTEX_ENTER(&rx_connHashTable_lock);
1018 cid = (rx_nextCid += RX_MAXCALLS);
1019 conn->type = RX_CLIENT_CONNECTION;
1021 conn->epoch = rx_epoch;
1022 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
1023 conn->serviceId = sservice;
1024 conn->securityObject = securityObject;
1025 conn->securityData = (void *) 0;
1026 conn->securityIndex = serviceSecurityIndex;
1027 rx_SetConnDeadTime(conn, rx_connDeadTime);
1028 rx_SetConnSecondsUntilNatPing(conn, 0);
1029 conn->ackRate = RX_FAST_ACK_RATE;
1030 conn->nSpecific = 0;
1031 conn->specific = NULL;
1032 conn->challengeEvent = NULL;
1033 conn->delayedAbortEvent = NULL;
1034 conn->abortCount = 0;
1036 for (i = 0; i < RX_MAXCALLS; i++) {
1037 conn->twind[i] = rx_initSendWindow;
1038 conn->rwind[i] = rx_initReceiveWindow;
1039 conn->lastBusy[i] = 0;
1042 RXS_NewConnection(securityObject, conn);
1044 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1046 conn->refCount++; /* no lock required since only this thread knows... */
1047 conn->next = rx_connHashTable[hashindex];
1048 rx_connHashTable[hashindex] = conn;
1049 if (rx_stats_active)
1050 rx_atomic_inc(&rx_stats.nClientConns);
1051 MUTEX_EXIT(&rx_connHashTable_lock);
1057 * Ensure a connection's timeout values are valid.
1059 * @param[in] conn The connection to check
1061 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1062 * unless idleDeadTime and/or hardDeadTime are not set
1066 rxi_CheckConnTimeouts(struct rx_connection *conn)
1068 /* a connection's timeouts must have the relationship
1069 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1070 * total loss of network to a peer may cause an idle timeout instead of a
1071 * dead timeout, simply because the idle timeout gets hit first. Also set
1072 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1073 /* this logic is slightly complicated by the fact that
1074 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1076 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1077 if (conn->idleDeadTime) {
1078 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1080 if (conn->hardDeadTime) {
1081 if (conn->idleDeadTime) {
1082 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1084 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1090 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1092 /* The idea is to set the dead time to a value that allows several
1093 * keepalives to be dropped without timing out the connection. */
1094 conn->secondsUntilDead = seconds;
1095 rxi_CheckConnTimeouts(conn);
1096 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1100 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1102 conn->hardDeadTime = seconds;
1103 rxi_CheckConnTimeouts(conn);
1107 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1109 conn->idleDeadTime = seconds;
1110 rxi_CheckConnTimeouts(conn);
1113 int rxi_lowPeerRefCount = 0;
1114 int rxi_lowConnRefCount = 0;
1117 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1118 * NOTE: must not be called with rx_connHashTable_lock held.
1121 rxi_CleanupConnection(struct rx_connection *conn)
1123 /* Notify the service exporter, if requested, that this connection
1124 * is being destroyed */
1125 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1126 (*conn->service->destroyConnProc) (conn);
1128 /* Notify the security module that this connection is being destroyed */
1129 RXS_DestroyConnection(conn->securityObject, conn);
1131 /* If this is the last connection using the rx_peer struct, set its
1132 * idle time to now. rxi_ReapConnections will reap it if it's still
1133 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1135 MUTEX_ENTER(&rx_peerHashTable_lock);
1136 if (conn->peer->refCount < 2) {
1137 conn->peer->idleWhen = clock_Sec();
1138 if (conn->peer->refCount < 1) {
1139 conn->peer->refCount = 1;
1140 if (rx_stats_active) {
1141 MUTEX_ENTER(&rx_stats_mutex);
1142 rxi_lowPeerRefCount++;
1143 MUTEX_EXIT(&rx_stats_mutex);
1147 conn->peer->refCount--;
1148 MUTEX_EXIT(&rx_peerHashTable_lock);
1150 if (rx_stats_active)
1152 if (conn->type == RX_SERVER_CONNECTION)
1153 rx_atomic_dec(&rx_stats.nServerConns);
1155 rx_atomic_dec(&rx_stats.nClientConns);
1158 if (conn->specific) {
1160 for (i = 0; i < conn->nSpecific; i++) {
1161 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1162 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1163 conn->specific[i] = NULL;
1165 free(conn->specific);
1167 conn->specific = NULL;
1168 conn->nSpecific = 0;
1169 #endif /* !KERNEL */
1171 MUTEX_DESTROY(&conn->conn_call_lock);
1172 MUTEX_DESTROY(&conn->conn_data_lock);
1173 CV_DESTROY(&conn->conn_call_cv);
1175 rxi_FreeConnection(conn);
1178 /* Destroy the specified connection */
1180 rxi_DestroyConnection(struct rx_connection *conn)
1182 MUTEX_ENTER(&rx_connHashTable_lock);
1183 rxi_DestroyConnectionNoLock(conn);
1184 /* conn should be at the head of the cleanup list */
1185 if (conn == rx_connCleanup_list) {
1186 rx_connCleanup_list = rx_connCleanup_list->next;
1187 MUTEX_EXIT(&rx_connHashTable_lock);
1188 rxi_CleanupConnection(conn);
1190 #ifdef RX_ENABLE_LOCKS
1192 MUTEX_EXIT(&rx_connHashTable_lock);
1194 #endif /* RX_ENABLE_LOCKS */
1198 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1200 struct rx_connection **conn_ptr;
1202 struct rx_packet *packet;
1209 MUTEX_ENTER(&conn->conn_data_lock);
1210 MUTEX_ENTER(&rx_refcnt_mutex);
1211 if (conn->refCount > 0)
1214 if (rx_stats_active) {
1215 MUTEX_ENTER(&rx_stats_mutex);
1216 rxi_lowConnRefCount++;
1217 MUTEX_EXIT(&rx_stats_mutex);
1221 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1222 /* Busy; wait till the last guy before proceeding */
1223 MUTEX_EXIT(&rx_refcnt_mutex);
1224 MUTEX_EXIT(&conn->conn_data_lock);
1229 /* If the client previously called rx_NewCall, but it is still
1230 * waiting, treat this as a running call, and wait to destroy the
1231 * connection later when the call completes. */
1232 if ((conn->type == RX_CLIENT_CONNECTION)
1233 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1234 conn->flags |= RX_CONN_DESTROY_ME;
1235 MUTEX_EXIT(&conn->conn_data_lock);
1239 MUTEX_EXIT(&rx_refcnt_mutex);
1240 MUTEX_EXIT(&conn->conn_data_lock);
1242 /* Check for extant references to this connection */
1243 MUTEX_ENTER(&conn->conn_call_lock);
1244 for (i = 0; i < RX_MAXCALLS; i++) {
1245 struct rx_call *call = conn->call[i];
1248 if (conn->type == RX_CLIENT_CONNECTION) {
1249 MUTEX_ENTER(&call->lock);
1250 if (call->delayedAckEvent) {
1251 /* Push the final acknowledgment out now--there
1252 * won't be a subsequent call to acknowledge the
1253 * last reply packets */
1254 rxevent_Cancel(&call->delayedAckEvent, call,
1255 RX_CALL_REFCOUNT_DELAY);
1256 if (call->state == RX_STATE_PRECALL
1257 || call->state == RX_STATE_ACTIVE) {
1258 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1260 rxi_AckAll(NULL, call, 0);
1263 MUTEX_EXIT(&call->lock);
1267 MUTEX_EXIT(&conn->conn_call_lock);
1269 #ifdef RX_ENABLE_LOCKS
1271 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1272 MUTEX_EXIT(&conn->conn_data_lock);
1274 /* Someone is accessing a packet right now. */
1278 #endif /* RX_ENABLE_LOCKS */
1281 /* Don't destroy the connection if there are any call
1282 * structures still in use */
1283 MUTEX_ENTER(&conn->conn_data_lock);
1284 conn->flags |= RX_CONN_DESTROY_ME;
1285 MUTEX_EXIT(&conn->conn_data_lock);
1290 if (conn->natKeepAliveEvent) {
1291 rxi_NatKeepAliveOff(conn);
1294 if (conn->delayedAbortEvent) {
1295 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
1296 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1298 MUTEX_ENTER(&conn->conn_data_lock);
1299 rxi_SendConnectionAbort(conn, packet, 0, 1);
1300 MUTEX_EXIT(&conn->conn_data_lock);
1301 rxi_FreePacket(packet);
1305 /* Remove from connection hash table before proceeding */
1307 &rx_connHashTable[CONN_HASH
1308 (peer->host, peer->port, conn->cid, conn->epoch,
1310 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1311 if (*conn_ptr == conn) {
1312 *conn_ptr = conn->next;
1316 /* if the conn that we are destroying was the last connection, then we
1317 * clear rxLastConn as well */
1318 if (rxLastConn == conn)
1321 /* Make sure the connection is completely reset before deleting it. */
1322 /* get rid of pending events that could zap us later */
1323 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
1324 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
1325 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
1327 /* Add the connection to the list of destroyed connections that
1328 * need to be cleaned up. This is necessary to avoid deadlocks
1329 * in the routines we call to inform others that this connection is
1330 * being destroyed. */
1331 conn->next = rx_connCleanup_list;
1332 rx_connCleanup_list = conn;
1335 /* Externally available version */
1337 rx_DestroyConnection(struct rx_connection *conn)
1342 rxi_DestroyConnection(conn);
1347 rx_GetConnection(struct rx_connection *conn)
1352 MUTEX_ENTER(&rx_refcnt_mutex);
1354 MUTEX_EXIT(&rx_refcnt_mutex);
1358 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1359 /* Wait for the transmit queue to no longer be busy.
1360 * requires the call->lock to be held */
1362 rxi_WaitforTQBusy(struct rx_call *call) {
1363 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1364 call->flags |= RX_CALL_TQ_WAIT;
1366 #ifdef RX_ENABLE_LOCKS
1367 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1368 CV_WAIT(&call->cv_tq, &call->lock);
1369 #else /* RX_ENABLE_LOCKS */
1370 osi_rxSleep(&call->tq);
1371 #endif /* RX_ENABLE_LOCKS */
1373 if (call->tqWaiters == 0) {
1374 call->flags &= ~RX_CALL_TQ_WAIT;
1381 rxi_WakeUpTransmitQueue(struct rx_call *call)
1383 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1384 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1385 call, call->tqWaiters, call->flags));
1386 #ifdef RX_ENABLE_LOCKS
1387 osirx_AssertMine(&call->lock, "rxi_Start start");
1388 CV_BROADCAST(&call->cv_tq);
1389 #else /* RX_ENABLE_LOCKS */
1390 osi_rxWakeup(&call->tq);
1391 #endif /* RX_ENABLE_LOCKS */
1395 /* Start a new rx remote procedure call, on the specified connection.
1396 * If wait is set to 1, wait for a free call channel; otherwise return
1397 * 0. Maxtime gives the maximum number of seconds this call may take,
1398 * after rx_NewCall returns. After this time interval, a call to any
1399 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1400 * For fine grain locking, we hold the conn_call_lock in order to
1401 * to ensure that we don't get signalle after we found a call in an active
1402 * state and before we go to sleep.
1405 rx_NewCall(struct rx_connection *conn)
1407 int i, wait, ignoreBusy = 1;
1408 struct rx_call *call;
1409 struct clock queueTime;
1410 afs_uint32 leastBusy = 0;
1414 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1417 clock_GetTime(&queueTime);
1419 * Check if there are others waiting for a new call.
1420 * If so, let them go first to avoid starving them.
1421 * This is a fairly simple scheme, and might not be
1422 * a complete solution for large numbers of waiters.
1424 * makeCallWaiters keeps track of the number of
1425 * threads waiting to make calls and the
1426 * RX_CONN_MAKECALL_WAITING flag bit is used to
1427 * indicate that there are indeed calls waiting.
1428 * The flag is set when the waiter is incremented.
1429 * It is only cleared when makeCallWaiters is 0.
1430 * This prevents us from accidently destroying the
1431 * connection while it is potentially about to be used.
1433 MUTEX_ENTER(&conn->conn_call_lock);
1434 MUTEX_ENTER(&conn->conn_data_lock);
1435 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1436 conn->flags |= RX_CONN_MAKECALL_WAITING;
1437 conn->makeCallWaiters++;
1438 MUTEX_EXIT(&conn->conn_data_lock);
1440 #ifdef RX_ENABLE_LOCKS
1441 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1445 MUTEX_ENTER(&conn->conn_data_lock);
1446 conn->makeCallWaiters--;
1447 if (conn->makeCallWaiters == 0)
1448 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1451 /* We are now the active thread in rx_NewCall */
1452 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1453 MUTEX_EXIT(&conn->conn_data_lock);
1458 for (i = 0; i < RX_MAXCALLS; i++) {
1459 call = conn->call[i];
1461 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1462 /* we're not ignoring busy call slots; only look at the
1463 * call slot that is the "least" busy */
1467 if (call->state == RX_STATE_DALLY) {
1468 MUTEX_ENTER(&call->lock);
1469 if (call->state == RX_STATE_DALLY) {
1470 if (ignoreBusy && conn->lastBusy[i]) {
1471 /* if we're ignoring busy call slots, skip any ones that
1472 * have lastBusy set */
1473 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1474 leastBusy = conn->lastBusy[i];
1476 MUTEX_EXIT(&call->lock);
1481 * We are setting the state to RX_STATE_RESET to
1482 * ensure that no one else will attempt to use this
1483 * call once we drop the conn->conn_call_lock and
1484 * call->lock. We must drop the conn->conn_call_lock
1485 * before calling rxi_ResetCall because the process
1486 * of clearing the transmit queue can block for an
1487 * extended period of time. If we block while holding
1488 * the conn->conn_call_lock, then all rx_EndCall
1489 * processing will block as well. This has a detrimental
1490 * effect on overall system performance.
1492 call->state = RX_STATE_RESET;
1493 MUTEX_EXIT(&conn->conn_call_lock);
1494 MUTEX_ENTER(&rx_refcnt_mutex);
1495 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1496 MUTEX_EXIT(&rx_refcnt_mutex);
1497 rxi_ResetCall(call, 0);
1498 (*call->callNumber)++;
1499 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1503 * If we failed to be able to safely obtain the
1504 * conn->conn_call_lock we will have to drop the
1505 * call->lock to avoid a deadlock. When the call->lock
1506 * is released the state of the call can change. If it
1507 * is no longer RX_STATE_RESET then some other thread is
1510 MUTEX_EXIT(&call->lock);
1511 MUTEX_ENTER(&conn->conn_call_lock);
1512 MUTEX_ENTER(&call->lock);
1514 if (call->state == RX_STATE_RESET)
1518 * If we get here it means that after dropping
1519 * the conn->conn_call_lock and call->lock that
1520 * the call is no longer ours. If we can't find
1521 * a free call in the remaining slots we should
1522 * not go immediately to RX_CONN_MAKECALL_WAITING
1523 * because by dropping the conn->conn_call_lock
1524 * we have given up synchronization with rx_EndCall.
1525 * Instead, cycle through one more time to see if
1526 * we can find a call that can call our own.
1528 MUTEX_ENTER(&rx_refcnt_mutex);
1529 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1530 MUTEX_EXIT(&rx_refcnt_mutex);
1533 MUTEX_EXIT(&call->lock);
1536 if (ignoreBusy && conn->lastBusy[i]) {
1537 /* if we're ignoring busy call slots, skip any ones that
1538 * have lastBusy set */
1539 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1540 leastBusy = conn->lastBusy[i];
1545 /* rxi_NewCall returns with mutex locked */
1546 call = rxi_NewCall(conn, i);
1547 MUTEX_ENTER(&rx_refcnt_mutex);
1548 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1549 MUTEX_EXIT(&rx_refcnt_mutex);
1553 if (i < RX_MAXCALLS) {
1554 conn->lastBusy[i] = 0;
1559 if (leastBusy && ignoreBusy) {
1560 /* we didn't find a useable call slot, but we did see at least one
1561 * 'busy' slot; look again and only use a slot with the 'least
1567 MUTEX_ENTER(&conn->conn_data_lock);
1568 conn->flags |= RX_CONN_MAKECALL_WAITING;
1569 conn->makeCallWaiters++;
1570 MUTEX_EXIT(&conn->conn_data_lock);
1572 #ifdef RX_ENABLE_LOCKS
1573 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1577 MUTEX_ENTER(&conn->conn_data_lock);
1578 conn->makeCallWaiters--;
1579 if (conn->makeCallWaiters == 0)
1580 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1581 MUTEX_EXIT(&conn->conn_data_lock);
1583 /* Client is initially in send mode */
1584 call->state = RX_STATE_ACTIVE;
1585 call->error = conn->error;
1587 call->mode = RX_MODE_ERROR;
1589 call->mode = RX_MODE_SENDING;
1591 /* remember start time for call in case we have hard dead time limit */
1592 call->queueTime = queueTime;
1593 clock_GetTime(&call->startTime);
1594 hzero(call->bytesSent);
1595 hzero(call->bytesRcvd);
1597 /* Turn on busy protocol. */
1598 rxi_KeepAliveOn(call);
1600 /* Attempt MTU discovery */
1601 rxi_GrowMTUOn(call);
1604 * We are no longer the active thread in rx_NewCall
1606 MUTEX_ENTER(&conn->conn_data_lock);
1607 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1608 MUTEX_EXIT(&conn->conn_data_lock);
1611 * Wake up anyone else who might be giving us a chance to
1612 * run (see code above that avoids resource starvation).
1614 #ifdef RX_ENABLE_LOCKS
1615 CV_BROADCAST(&conn->conn_call_cv);
1619 MUTEX_EXIT(&conn->conn_call_lock);
1621 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1622 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1623 osi_Panic("rx_NewCall call about to be used without an empty tq");
1625 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1627 MUTEX_EXIT(&call->lock);
1630 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1635 rxi_HasActiveCalls(struct rx_connection *aconn)
1638 struct rx_call *tcall;
1642 for (i = 0; i < RX_MAXCALLS; i++) {
1643 if ((tcall = aconn->call[i])) {
1644 if ((tcall->state == RX_STATE_ACTIVE)
1645 || (tcall->state == RX_STATE_PRECALL)) {
1656 rxi_GetCallNumberVector(struct rx_connection *aconn,
1657 afs_int32 * aint32s)
1660 struct rx_call *tcall;
1664 for (i = 0; i < RX_MAXCALLS; i++) {
1665 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1666 aint32s[i] = aconn->callNumber[i] + 1;
1668 aint32s[i] = aconn->callNumber[i];
1675 rxi_SetCallNumberVector(struct rx_connection *aconn,
1676 afs_int32 * aint32s)
1679 struct rx_call *tcall;
1683 for (i = 0; i < RX_MAXCALLS; i++) {
1684 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1685 aconn->callNumber[i] = aint32s[i] - 1;
1687 aconn->callNumber[i] = aint32s[i];
1693 /* Advertise a new service. A service is named locally by a UDP port
1694 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1697 char *serviceName; Name for identification purposes (e.g. the
1698 service name might be used for probing for
1701 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1702 char *serviceName, struct rx_securityClass **securityObjects,
1703 int nSecurityObjects,
1704 afs_int32(*serviceProc) (struct rx_call * acall))
1706 osi_socket socket = OSI_NULLSOCKET;
1707 struct rx_service *tservice;
1713 if (serviceId == 0) {
1715 "rx_NewService: service id for service %s is not non-zero.\n",
1722 "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",
1730 tservice = rxi_AllocService();
1733 #ifdef RX_ENABLE_LOCKS
1734 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1737 for (i = 0; i < RX_MAX_SERVICES; i++) {
1738 struct rx_service *service = rx_services[i];
1740 if (port == service->servicePort && host == service->serviceHost) {
1741 if (service->serviceId == serviceId) {
1742 /* The identical service has already been
1743 * installed; if the caller was intending to
1744 * change the security classes used by this
1745 * service, he/she loses. */
1747 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1748 serviceName, serviceId, service->serviceName);
1750 rxi_FreeService(tservice);
1753 /* Different service, same port: re-use the socket
1754 * which is bound to the same port */
1755 socket = service->socket;
1758 if (socket == OSI_NULLSOCKET) {
1759 /* If we don't already have a socket (from another
1760 * service on same port) get a new one */
1761 socket = rxi_GetHostUDPSocket(host, port);
1762 if (socket == OSI_NULLSOCKET) {
1764 rxi_FreeService(tservice);
1769 service->socket = socket;
1770 service->serviceHost = host;
1771 service->servicePort = port;
1772 service->serviceId = serviceId;
1773 service->serviceName = serviceName;
1774 service->nSecurityObjects = nSecurityObjects;
1775 service->securityObjects = securityObjects;
1776 service->minProcs = 0;
1777 service->maxProcs = 1;
1778 service->idleDeadTime = 60;
1779 service->idleDeadErr = 0;
1780 service->connDeadTime = rx_connDeadTime;
1781 service->executeRequestProc = serviceProc;
1782 service->checkReach = 0;
1783 service->nSpecific = 0;
1784 service->specific = NULL;
1785 rx_services[i] = service; /* not visible until now */
1791 rxi_FreeService(tservice);
1792 (osi_Msg "rx_NewService: cannot support > %d services\n",
1797 /* Set configuration options for all of a service's security objects */
1800 rx_SetSecurityConfiguration(struct rx_service *service,
1801 rx_securityConfigVariables type,
1805 for (i = 0; i<service->nSecurityObjects; i++) {
1806 if (service->securityObjects[i]) {
1807 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1815 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1816 struct rx_securityClass **securityObjects, int nSecurityObjects,
1817 afs_int32(*serviceProc) (struct rx_call * acall))
1819 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1822 /* Generic request processing loop. This routine should be called
1823 * by the implementation dependent rx_ServerProc. If socketp is
1824 * non-null, it will be set to the file descriptor that this thread
1825 * is now listening on. If socketp is null, this routine will never
1828 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1830 struct rx_call *call;
1832 struct rx_service *tservice = NULL;
1839 call = rx_GetCall(threadID, tservice, socketp);
1840 if (socketp && *socketp != OSI_NULLSOCKET) {
1841 /* We are now a listener thread */
1847 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1848 #ifdef RX_ENABLE_LOCKS
1850 #endif /* RX_ENABLE_LOCKS */
1851 afs_termState = AFSOP_STOP_AFS;
1852 afs_osi_Wakeup(&afs_termState);
1853 #ifdef RX_ENABLE_LOCKS
1855 #endif /* RX_ENABLE_LOCKS */
1860 /* if server is restarting( typically smooth shutdown) then do not
1861 * allow any new calls.
1864 if (rx_tranquil && (call != NULL)) {
1868 MUTEX_ENTER(&call->lock);
1870 rxi_CallError(call, RX_RESTARTING);
1871 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1873 MUTEX_EXIT(&call->lock);
1878 tservice = call->conn->service;
1880 if (tservice->beforeProc)
1881 (*tservice->beforeProc) (call);
1883 code = tservice->executeRequestProc(call);
1885 if (tservice->afterProc)
1886 (*tservice->afterProc) (call, code);
1888 rx_EndCall(call, code);
1890 if (tservice->postProc)
1891 (*tservice->postProc) (code);
1893 if (rx_stats_active) {
1894 MUTEX_ENTER(&rx_stats_mutex);
1896 MUTEX_EXIT(&rx_stats_mutex);
1903 rx_WakeupServerProcs(void)
1905 struct rx_serverQueueEntry *np, *tqp;
1909 MUTEX_ENTER(&rx_serverPool_lock);
1911 #ifdef RX_ENABLE_LOCKS
1912 if (rx_waitForPacket)
1913 CV_BROADCAST(&rx_waitForPacket->cv);
1914 #else /* RX_ENABLE_LOCKS */
1915 if (rx_waitForPacket)
1916 osi_rxWakeup(rx_waitForPacket);
1917 #endif /* RX_ENABLE_LOCKS */
1918 MUTEX_ENTER(&freeSQEList_lock);
1919 for (np = rx_FreeSQEList; np; np = tqp) {
1920 tqp = *(struct rx_serverQueueEntry **)np;
1921 #ifdef RX_ENABLE_LOCKS
1922 CV_BROADCAST(&np->cv);
1923 #else /* RX_ENABLE_LOCKS */
1925 #endif /* RX_ENABLE_LOCKS */
1927 MUTEX_EXIT(&freeSQEList_lock);
1928 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1929 #ifdef RX_ENABLE_LOCKS
1930 CV_BROADCAST(&np->cv);
1931 #else /* RX_ENABLE_LOCKS */
1933 #endif /* RX_ENABLE_LOCKS */
1935 MUTEX_EXIT(&rx_serverPool_lock);
1940 * One thing that seems to happen is that all the server threads get
1941 * tied up on some empty or slow call, and then a whole bunch of calls
1942 * arrive at once, using up the packet pool, so now there are more
1943 * empty calls. The most critical resources here are server threads
1944 * and the free packet pool. The "doreclaim" code seems to help in
1945 * general. I think that eventually we arrive in this state: there
1946 * are lots of pending calls which do have all their packets present,
1947 * so they won't be reclaimed, are multi-packet calls, so they won't
1948 * be scheduled until later, and thus are tying up most of the free
1949 * packet pool for a very long time.
1951 * 1. schedule multi-packet calls if all the packets are present.
1952 * Probably CPU-bound operation, useful to return packets to pool.
1953 * Do what if there is a full window, but the last packet isn't here?
1954 * 3. preserve one thread which *only* runs "best" calls, otherwise
1955 * it sleeps and waits for that type of call.
1956 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1957 * the current dataquota business is badly broken. The quota isn't adjusted
1958 * to reflect how many packets are presently queued for a running call.
1959 * So, when we schedule a queued call with a full window of packets queued
1960 * up for it, that *should* free up a window full of packets for other 2d-class
1961 * calls to be able to use from the packet pool. But it doesn't.
1963 * NB. Most of the time, this code doesn't run -- since idle server threads
1964 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1965 * as a new call arrives.
1967 /* Sleep until a call arrives. Returns a pointer to the call, ready
1968 * for an rx_Read. */
1969 #ifdef RX_ENABLE_LOCKS
1971 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1973 struct rx_serverQueueEntry *sq;
1974 struct rx_call *call = (struct rx_call *)0;
1975 struct rx_service *service = NULL;
1977 MUTEX_ENTER(&freeSQEList_lock);
1979 if ((sq = rx_FreeSQEList)) {
1980 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1981 MUTEX_EXIT(&freeSQEList_lock);
1982 } else { /* otherwise allocate a new one and return that */
1983 MUTEX_EXIT(&freeSQEList_lock);
1984 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1985 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1986 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1989 MUTEX_ENTER(&rx_serverPool_lock);
1990 if (cur_service != NULL) {
1991 ReturnToServerPool(cur_service);
1994 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1995 struct rx_call *tcall, *ncall, *choice2 = NULL;
1997 /* Scan for eligible incoming calls. A call is not eligible
1998 * if the maximum number of calls for its service type are
1999 * already executing */
2000 /* One thread will process calls FCFS (to prevent starvation),
2001 * while the other threads may run ahead looking for calls which
2002 * have all their input data available immediately. This helps
2003 * keep threads from blocking, waiting for data from the client. */
2004 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
2005 service = tcall->conn->service;
2006 if (!QuotaOK(service)) {
2009 MUTEX_ENTER(&rx_pthread_mutex);
2010 if (tno == rxi_fcfs_thread_num
2011 || !tcall->queue_item_header.next) {
2012 MUTEX_EXIT(&rx_pthread_mutex);
2013 /* If we're the fcfs thread , then we'll just use
2014 * this call. If we haven't been able to find an optimal
2015 * choice, and we're at the end of the list, then use a
2016 * 2d choice if one has been identified. Otherwise... */
2017 call = (choice2 ? choice2 : tcall);
2018 service = call->conn->service;
2020 MUTEX_EXIT(&rx_pthread_mutex);
2021 if (!queue_IsEmpty(&tcall->rq)) {
2022 struct rx_packet *rp;
2023 rp = queue_First(&tcall->rq, rx_packet);
2024 if (rp->header.seq == 1) {
2026 || (rp->header.flags & RX_LAST_PACKET)) {
2028 } else if (rxi_2dchoice && !choice2
2029 && !(tcall->flags & RX_CALL_CLEARED)
2030 && (tcall->rprev > rxi_HardAckRate)) {
2040 ReturnToServerPool(service);
2047 MUTEX_EXIT(&rx_serverPool_lock);
2048 MUTEX_ENTER(&call->lock);
2050 if (call->flags & RX_CALL_WAIT_PROC) {
2051 call->flags &= ~RX_CALL_WAIT_PROC;
2052 rx_atomic_dec(&rx_nWaiting);
2055 if (call->state != RX_STATE_PRECALL || call->error) {
2056 MUTEX_EXIT(&call->lock);
2057 MUTEX_ENTER(&rx_serverPool_lock);
2058 ReturnToServerPool(service);
2063 if (queue_IsEmpty(&call->rq)
2064 || queue_First(&call->rq, rx_packet)->header.seq != 1)
2065 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2067 CLEAR_CALL_QUEUE_LOCK(call);
2070 /* If there are no eligible incoming calls, add this process
2071 * to the idle server queue, to wait for one */
2075 *socketp = OSI_NULLSOCKET;
2077 sq->socketp = socketp;
2078 queue_Append(&rx_idleServerQueue, sq);
2079 #ifndef AFS_AIX41_ENV
2080 rx_waitForPacket = sq;
2082 rx_waitingForPacket = sq;
2083 #endif /* AFS_AIX41_ENV */
2085 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2087 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2088 MUTEX_EXIT(&rx_serverPool_lock);
2089 return (struct rx_call *)0;
2092 } while (!(call = sq->newcall)
2093 && !(socketp && *socketp != OSI_NULLSOCKET));
2094 MUTEX_EXIT(&rx_serverPool_lock);
2096 MUTEX_ENTER(&call->lock);
2102 MUTEX_ENTER(&freeSQEList_lock);
2103 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2104 rx_FreeSQEList = sq;
2105 MUTEX_EXIT(&freeSQEList_lock);
2108 clock_GetTime(&call->startTime);
2109 call->state = RX_STATE_ACTIVE;
2110 call->mode = RX_MODE_RECEIVING;
2111 #ifdef RX_KERNEL_TRACE
2112 if (ICL_SETACTIVE(afs_iclSetp)) {
2113 int glockOwner = ISAFS_GLOCK();
2116 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2117 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2124 rxi_calltrace(RX_CALL_START, call);
2125 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2126 call->conn->service->servicePort, call->conn->service->serviceId,
2129 MUTEX_EXIT(&call->lock);
2130 MUTEX_ENTER(&rx_refcnt_mutex);
2131 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2132 MUTEX_EXIT(&rx_refcnt_mutex);
2134 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2139 #else /* RX_ENABLE_LOCKS */
2141 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2143 struct rx_serverQueueEntry *sq;
2144 struct rx_call *call = (struct rx_call *)0, *choice2;
2145 struct rx_service *service = NULL;
2149 MUTEX_ENTER(&freeSQEList_lock);
2151 if ((sq = rx_FreeSQEList)) {
2152 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2153 MUTEX_EXIT(&freeSQEList_lock);
2154 } else { /* otherwise allocate a new one and return that */
2155 MUTEX_EXIT(&freeSQEList_lock);
2156 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2157 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2158 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2160 MUTEX_ENTER(&sq->lock);
2162 if (cur_service != NULL) {
2163 cur_service->nRequestsRunning--;
2164 MUTEX_ENTER(&rx_quota_mutex);
2165 if (cur_service->nRequestsRunning < cur_service->minProcs)
2168 MUTEX_EXIT(&rx_quota_mutex);
2170 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
2171 struct rx_call *tcall, *ncall;
2172 /* Scan for eligible incoming calls. A call is not eligible
2173 * if the maximum number of calls for its service type are
2174 * already executing */
2175 /* One thread will process calls FCFS (to prevent starvation),
2176 * while the other threads may run ahead looking for calls which
2177 * have all their input data available immediately. This helps
2178 * keep threads from blocking, waiting for data from the client. */
2179 choice2 = (struct rx_call *)0;
2180 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
2181 service = tcall->conn->service;
2182 if (QuotaOK(service)) {
2183 MUTEX_ENTER(&rx_pthread_mutex);
2184 if (tno == rxi_fcfs_thread_num
2185 || !tcall->queue_item_header.next) {
2186 MUTEX_EXIT(&rx_pthread_mutex);
2187 /* If we're the fcfs thread, then we'll just use
2188 * this call. If we haven't been able to find an optimal
2189 * choice, and we're at the end of the list, then use a
2190 * 2d choice if one has been identified. Otherwise... */
2191 call = (choice2 ? choice2 : tcall);
2192 service = call->conn->service;
2194 MUTEX_EXIT(&rx_pthread_mutex);
2195 if (!queue_IsEmpty(&tcall->rq)) {
2196 struct rx_packet *rp;
2197 rp = queue_First(&tcall->rq, rx_packet);
2198 if (rp->header.seq == 1
2200 || (rp->header.flags & RX_LAST_PACKET))) {
2202 } else if (rxi_2dchoice && !choice2
2203 && !(tcall->flags & RX_CALL_CLEARED)
2204 && (tcall->rprev > rxi_HardAckRate)) {
2218 /* we can't schedule a call if there's no data!!! */
2219 /* send an ack if there's no data, if we're missing the
2220 * first packet, or we're missing something between first
2221 * and last -- there's a "hole" in the incoming data. */
2222 if (queue_IsEmpty(&call->rq)
2223 || queue_First(&call->rq, rx_packet)->header.seq != 1
2224 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
2225 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2227 call->flags &= (~RX_CALL_WAIT_PROC);
2228 service->nRequestsRunning++;
2229 /* just started call in minProcs pool, need fewer to maintain
2231 MUTEX_ENTER(&rx_quota_mutex);
2232 if (service->nRequestsRunning <= service->minProcs)
2235 MUTEX_EXIT(&rx_quota_mutex);
2236 rx_atomic_dec(&rx_nWaiting);
2237 /* MUTEX_EXIT(&call->lock); */
2239 /* If there are no eligible incoming calls, add this process
2240 * to the idle server queue, to wait for one */
2243 *socketp = OSI_NULLSOCKET;
2245 sq->socketp = socketp;
2246 queue_Append(&rx_idleServerQueue, sq);
2250 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2252 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2253 return (struct rx_call *)0;
2256 } while (!(call = sq->newcall)
2257 && !(socketp && *socketp != OSI_NULLSOCKET));
2259 MUTEX_EXIT(&sq->lock);
2261 MUTEX_ENTER(&freeSQEList_lock);
2262 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2263 rx_FreeSQEList = sq;
2264 MUTEX_EXIT(&freeSQEList_lock);
2267 clock_GetTime(&call->startTime);
2268 call->state = RX_STATE_ACTIVE;
2269 call->mode = RX_MODE_RECEIVING;
2270 #ifdef RX_KERNEL_TRACE
2271 if (ICL_SETACTIVE(afs_iclSetp)) {
2272 int glockOwner = ISAFS_GLOCK();
2275 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2276 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2283 rxi_calltrace(RX_CALL_START, call);
2284 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2285 call->conn->service->servicePort, call->conn->service->serviceId,
2288 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2295 #endif /* RX_ENABLE_LOCKS */
2299 /* Establish a procedure to be called when a packet arrives for a
2300 * call. This routine will be called at most once after each call,
2301 * and will also be called if there is an error condition on the or
2302 * the call is complete. Used by multi rx to build a selection
2303 * function which determines which of several calls is likely to be a
2304 * good one to read from.
2305 * NOTE: the way this is currently implemented it is probably only a
2306 * good idea to (1) use it immediately after a newcall (clients only)
2307 * and (2) only use it once. Other uses currently void your warranty
2310 rx_SetArrivalProc(struct rx_call *call,
2311 void (*proc) (struct rx_call * call,
2314 void * handle, int arg)
2316 call->arrivalProc = proc;
2317 call->arrivalProcHandle = handle;
2318 call->arrivalProcArg = arg;
2321 /* Call is finished (possibly prematurely). Return rc to the peer, if
2322 * appropriate, and return the final error code from the conversation
2326 rx_EndCall(struct rx_call *call, afs_int32 rc)
2328 struct rx_connection *conn = call->conn;
2332 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2333 call, rc, call->error, call->abortCode));
2336 MUTEX_ENTER(&call->lock);
2338 if (rc == 0 && call->error == 0) {
2339 call->abortCode = 0;
2340 call->abortCount = 0;
2343 call->arrivalProc = (void (*)())0;
2344 if (rc && call->error == 0) {
2345 rxi_CallError(call, rc);
2346 call->mode = RX_MODE_ERROR;
2347 /* Send an abort message to the peer if this error code has
2348 * only just been set. If it was set previously, assume the
2349 * peer has already been sent the error code or will request it
2351 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2353 if (conn->type == RX_SERVER_CONNECTION) {
2354 /* Make sure reply or at least dummy reply is sent */
2355 if (call->mode == RX_MODE_RECEIVING) {
2356 MUTEX_EXIT(&call->lock);
2357 rxi_WriteProc(call, 0, 0);
2358 MUTEX_ENTER(&call->lock);
2360 if (call->mode == RX_MODE_SENDING) {
2361 MUTEX_EXIT(&call->lock);
2362 rxi_FlushWrite(call);
2363 MUTEX_ENTER(&call->lock);
2365 rxi_calltrace(RX_CALL_END, call);
2366 /* Call goes to hold state until reply packets are acknowledged */
2367 if (call->tfirst + call->nSoftAcked < call->tnext) {
2368 call->state = RX_STATE_HOLD;
2370 call->state = RX_STATE_DALLY;
2371 rxi_ClearTransmitQueue(call, 0);
2372 rxi_rto_cancel(call);
2373 rxevent_Cancel(&call->keepAliveEvent, call,
2374 RX_CALL_REFCOUNT_ALIVE);
2376 } else { /* Client connection */
2378 /* Make sure server receives input packets, in the case where
2379 * no reply arguments are expected */
2380 if ((call->mode == RX_MODE_SENDING)
2381 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2382 MUTEX_EXIT(&call->lock);
2383 (void)rxi_ReadProc(call, &dummy, 1);
2384 MUTEX_ENTER(&call->lock);
2387 /* If we had an outstanding delayed ack, be nice to the server
2388 * and force-send it now.
2390 if (call->delayedAckEvent) {
2391 rxevent_Cancel(&call->delayedAckEvent, call,
2392 RX_CALL_REFCOUNT_DELAY);
2393 rxi_SendDelayedAck(NULL, call, NULL, 0);
2396 /* We need to release the call lock since it's lower than the
2397 * conn_call_lock and we don't want to hold the conn_call_lock
2398 * over the rx_ReadProc call. The conn_call_lock needs to be held
2399 * here for the case where rx_NewCall is perusing the calls on
2400 * the connection structure. We don't want to signal until
2401 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2402 * have checked this call, found it active and by the time it
2403 * goes to sleep, will have missed the signal.
2405 MUTEX_EXIT(&call->lock);
2406 MUTEX_ENTER(&conn->conn_call_lock);
2407 MUTEX_ENTER(&call->lock);
2409 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2410 conn->lastBusy[call->channel] = 0;
2413 MUTEX_ENTER(&conn->conn_data_lock);
2414 conn->flags |= RX_CONN_BUSY;
2415 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2416 MUTEX_EXIT(&conn->conn_data_lock);
2417 #ifdef RX_ENABLE_LOCKS
2418 CV_BROADCAST(&conn->conn_call_cv);
2423 #ifdef RX_ENABLE_LOCKS
2425 MUTEX_EXIT(&conn->conn_data_lock);
2427 #endif /* RX_ENABLE_LOCKS */
2428 call->state = RX_STATE_DALLY;
2430 error = call->error;
2432 /* currentPacket, nLeft, and NFree must be zeroed here, because
2433 * ResetCall cannot: ResetCall may be called at splnet(), in the
2434 * kernel version, and may interrupt the macros rx_Read or
2435 * rx_Write, which run at normal priority for efficiency. */
2436 if (call->currentPacket) {
2437 #ifdef RX_TRACK_PACKETS
2438 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2440 rxi_FreePacket(call->currentPacket);
2441 call->currentPacket = (struct rx_packet *)0;
2444 call->nLeft = call->nFree = call->curlen = 0;
2446 /* Free any packets from the last call to ReadvProc/WritevProc */
2447 #ifdef RXDEBUG_PACKET
2449 #endif /* RXDEBUG_PACKET */
2450 rxi_FreePackets(0, &call->iovq);
2451 MUTEX_EXIT(&call->lock);
2453 MUTEX_ENTER(&rx_refcnt_mutex);
2454 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2455 MUTEX_EXIT(&rx_refcnt_mutex);
2456 if (conn->type == RX_CLIENT_CONNECTION) {
2457 MUTEX_ENTER(&conn->conn_data_lock);
2458 conn->flags &= ~RX_CONN_BUSY;
2459 MUTEX_EXIT(&conn->conn_data_lock);
2460 MUTEX_EXIT(&conn->conn_call_lock);
2464 * Map errors to the local host's errno.h format.
2466 error = ntoh_syserr_conv(error);
2470 #if !defined(KERNEL)
2472 /* Call this routine when shutting down a server or client (especially
2473 * clients). This will allow Rx to gracefully garbage collect server
2474 * connections, and reduce the number of retries that a server might
2475 * make to a dead client.
2476 * This is not quite right, since some calls may still be ongoing and
2477 * we can't lock them to destroy them. */
2481 struct rx_connection **conn_ptr, **conn_end;
2485 if (rxinit_status == 1) {
2487 return; /* Already shutdown. */
2489 rxi_DeleteCachedConnections();
2490 if (rx_connHashTable) {
2491 MUTEX_ENTER(&rx_connHashTable_lock);
2492 for (conn_ptr = &rx_connHashTable[0], conn_end =
2493 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2495 struct rx_connection *conn, *next;
2496 for (conn = *conn_ptr; conn; conn = next) {
2498 if (conn->type == RX_CLIENT_CONNECTION) {
2499 MUTEX_ENTER(&rx_refcnt_mutex);
2501 MUTEX_EXIT(&rx_refcnt_mutex);
2502 #ifdef RX_ENABLE_LOCKS
2503 rxi_DestroyConnectionNoLock(conn);
2504 #else /* RX_ENABLE_LOCKS */
2505 rxi_DestroyConnection(conn);
2506 #endif /* RX_ENABLE_LOCKS */
2510 #ifdef RX_ENABLE_LOCKS
2511 while (rx_connCleanup_list) {
2512 struct rx_connection *conn;
2513 conn = rx_connCleanup_list;
2514 rx_connCleanup_list = rx_connCleanup_list->next;
2515 MUTEX_EXIT(&rx_connHashTable_lock);
2516 rxi_CleanupConnection(conn);
2517 MUTEX_ENTER(&rx_connHashTable_lock);
2519 MUTEX_EXIT(&rx_connHashTable_lock);
2520 #endif /* RX_ENABLE_LOCKS */
2525 afs_winsockCleanup();
2533 /* if we wakeup packet waiter too often, can get in loop with two
2534 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2536 rxi_PacketsUnWait(void)
2538 if (!rx_waitingForPackets) {
2542 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2543 return; /* still over quota */
2546 rx_waitingForPackets = 0;
2547 #ifdef RX_ENABLE_LOCKS
2548 CV_BROADCAST(&rx_waitingForPackets_cv);
2550 osi_rxWakeup(&rx_waitingForPackets);
2556 /* ------------------Internal interfaces------------------------- */
2558 /* Return this process's service structure for the
2559 * specified socket and service */
2560 static struct rx_service *
2561 rxi_FindService(osi_socket socket, u_short serviceId)
2563 struct rx_service **sp;
2564 for (sp = &rx_services[0]; *sp; sp++) {
2565 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2571 #ifdef RXDEBUG_PACKET
2572 #ifdef KDUMP_RX_LOCK
2573 static struct rx_call_rx_lock *rx_allCallsp = 0;
2575 static struct rx_call *rx_allCallsp = 0;
2577 #endif /* RXDEBUG_PACKET */
2579 /* Allocate a call structure, for the indicated channel of the
2580 * supplied connection. The mode and state of the call must be set by
2581 * the caller. Returns the call with mutex locked. */
2582 static struct rx_call *
2583 rxi_NewCall(struct rx_connection *conn, int channel)
2585 struct rx_call *call;
2586 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2587 struct rx_call *cp; /* Call pointer temp */
2588 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2589 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2591 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2593 /* Grab an existing call structure, or allocate a new one.
2594 * Existing call structures are assumed to have been left reset by
2596 MUTEX_ENTER(&rx_freeCallQueue_lock);
2598 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2600 * EXCEPT that the TQ might not yet be cleared out.
2601 * Skip over those with in-use TQs.
2604 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2605 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2611 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2612 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2613 call = queue_First(&rx_freeCallQueue, rx_call);
2614 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2616 if (rx_stats_active)
2617 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2618 MUTEX_EXIT(&rx_freeCallQueue_lock);
2619 MUTEX_ENTER(&call->lock);
2620 CLEAR_CALL_QUEUE_LOCK(call);
2621 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2622 /* Now, if TQ wasn't cleared earlier, do it now. */
2623 rxi_WaitforTQBusy(call);
2624 if (call->flags & RX_CALL_TQ_CLEARME) {
2625 rxi_ClearTransmitQueue(call, 1);
2626 /*queue_Init(&call->tq);*/
2628 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2629 /* Bind the call to its connection structure */
2631 rxi_ResetCall(call, 1);
2634 call = rxi_Alloc(sizeof(struct rx_call));
2635 #ifdef RXDEBUG_PACKET
2636 call->allNextp = rx_allCallsp;
2637 rx_allCallsp = call;
2639 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2640 #else /* RXDEBUG_PACKET */
2641 rx_atomic_inc(&rx_stats.nCallStructs);
2642 #endif /* RXDEBUG_PACKET */
2644 MUTEX_EXIT(&rx_freeCallQueue_lock);
2645 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2646 MUTEX_ENTER(&call->lock);
2647 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2648 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2649 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2651 /* Initialize once-only items */
2652 queue_Init(&call->tq);
2653 queue_Init(&call->rq);
2654 queue_Init(&call->iovq);
2655 #ifdef RXDEBUG_PACKET
2656 call->rqc = call->tqc = call->iovqc = 0;
2657 #endif /* RXDEBUG_PACKET */
2658 /* Bind the call to its connection structure (prereq for reset) */
2660 rxi_ResetCall(call, 1);
2662 call->channel = channel;
2663 call->callNumber = &conn->callNumber[channel];
2664 call->rwind = conn->rwind[channel];
2665 call->twind = conn->twind[channel];
2666 /* Note that the next expected call number is retained (in
2667 * conn->callNumber[i]), even if we reallocate the call structure
2669 conn->call[channel] = call;
2670 /* if the channel's never been used (== 0), we should start at 1, otherwise
2671 * the call number is valid from the last time this channel was used */
2672 if (*call->callNumber == 0)
2673 *call->callNumber = 1;
2678 /* A call has been inactive long enough that so we can throw away
2679 * state, including the call structure, which is placed on the call
2682 * call->lock amd rx_refcnt_mutex are held upon entry.
2683 * haveCTLock is set when called from rxi_ReapConnections.
2686 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2688 int channel = call->channel;
2689 struct rx_connection *conn = call->conn;
2692 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2693 (*call->callNumber)++;
2695 * We are setting the state to RX_STATE_RESET to
2696 * ensure that no one else will attempt to use this
2697 * call once we drop the refcnt lock. We must drop
2698 * the refcnt lock before calling rxi_ResetCall
2699 * because it cannot be held across acquiring the
2700 * freepktQ lock. NewCall does the same.
2702 call->state = RX_STATE_RESET;
2703 MUTEX_EXIT(&rx_refcnt_mutex);
2704 rxi_ResetCall(call, 0);
2706 MUTEX_ENTER(&conn->conn_call_lock);
2707 if (call->conn->call[channel] == call)
2708 call->conn->call[channel] = 0;
2709 MUTEX_EXIT(&conn->conn_call_lock);
2711 MUTEX_ENTER(&rx_freeCallQueue_lock);
2712 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2713 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2714 /* A call may be free even though its transmit queue is still in use.
2715 * Since we search the call list from head to tail, put busy calls at
2716 * the head of the list, and idle calls at the tail.
2718 if (call->flags & RX_CALL_TQ_BUSY)
2719 queue_Prepend(&rx_freeCallQueue, call);
2721 queue_Append(&rx_freeCallQueue, call);
2722 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2723 queue_Append(&rx_freeCallQueue, call);
2724 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2725 if (rx_stats_active)
2726 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2727 MUTEX_EXIT(&rx_freeCallQueue_lock);
2729 /* Destroy the connection if it was previously slated for
2730 * destruction, i.e. the Rx client code previously called
2731 * rx_DestroyConnection (client connections), or
2732 * rxi_ReapConnections called the same routine (server
2733 * connections). Only do this, however, if there are no
2734 * outstanding calls. Note that for fine grain locking, there appears
2735 * to be a deadlock in that rxi_FreeCall has a call locked and
2736 * DestroyConnectionNoLock locks each call in the conn. But note a
2737 * few lines up where we have removed this call from the conn.
2738 * If someone else destroys a connection, they either have no
2739 * call lock held or are going through this section of code.
2741 MUTEX_ENTER(&conn->conn_data_lock);
2742 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2743 MUTEX_ENTER(&rx_refcnt_mutex);
2745 MUTEX_EXIT(&rx_refcnt_mutex);
2746 MUTEX_EXIT(&conn->conn_data_lock);
2747 #ifdef RX_ENABLE_LOCKS
2749 rxi_DestroyConnectionNoLock(conn);
2751 rxi_DestroyConnection(conn);
2752 #else /* RX_ENABLE_LOCKS */
2753 rxi_DestroyConnection(conn);
2754 #endif /* RX_ENABLE_LOCKS */
2756 MUTEX_EXIT(&conn->conn_data_lock);
2758 MUTEX_ENTER(&rx_refcnt_mutex);
2761 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2762 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2765 rxi_Alloc(size_t size)
2769 if (rx_stats_active) {
2770 rx_atomic_add(&rxi_Allocsize, (int) size);
2771 rx_atomic_inc(&rxi_Alloccnt);
2775 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2776 afs_osi_Alloc_NoSleep(size);
2781 osi_Panic("rxi_Alloc error");
2787 rxi_Free(void *addr, size_t size)
2789 if (rx_stats_active) {
2790 rx_atomic_sub(&rxi_Allocsize, (int) size);
2791 rx_atomic_dec(&rxi_Alloccnt);
2793 osi_Free(addr, size);
2797 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2799 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2800 struct rx_peer *next = NULL;
2804 MUTEX_ENTER(&rx_peerHashTable_lock);
2806 peer_ptr = &rx_peerHashTable[0];
2807 peer_end = &rx_peerHashTable[rx_hashTableSize];
2810 for ( ; peer_ptr < peer_end; peer_ptr++) {
2813 for ( ; peer; peer = next) {
2815 if (host == peer->host)
2820 hashIndex = PEER_HASH(host, port);
2821 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2822 if ((peer->host == host) && (peer->port == port))
2827 MUTEX_ENTER(&rx_peerHashTable_lock);
2832 MUTEX_EXIT(&rx_peerHashTable_lock);
2834 MUTEX_ENTER(&peer->peer_lock);
2835 /* We don't handle dropping below min, so don't */
2836 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2837 peer->ifMTU=MIN(mtu, peer->ifMTU);
2838 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2839 /* if we tweaked this down, need to tune our peer MTU too */
2840 peer->MTU = MIN(peer->MTU, peer->natMTU);
2841 /* if we discovered a sub-1500 mtu, degrade */
2842 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2843 peer->maxDgramPackets = 1;
2844 /* We no longer have valid peer packet information */
2845 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2846 peer->maxPacketSize = 0;
2847 MUTEX_EXIT(&peer->peer_lock);
2849 MUTEX_ENTER(&rx_peerHashTable_lock);
2851 if (host && !port) {
2853 /* pick up where we left off */
2857 MUTEX_EXIT(&rx_peerHashTable_lock);
2860 /* Find the peer process represented by the supplied (host,port)
2861 * combination. If there is no appropriate active peer structure, a
2862 * new one will be allocated and initialized
2863 * The origPeer, if set, is a pointer to a peer structure on which the
2864 * refcount will be be decremented. This is used to replace the peer
2865 * structure hanging off a connection structure */
2867 rxi_FindPeer(afs_uint32 host, u_short port,
2868 struct rx_peer *origPeer, int create)
2872 hashIndex = PEER_HASH(host, port);
2873 MUTEX_ENTER(&rx_peerHashTable_lock);
2874 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2875 if ((pp->host == host) && (pp->port == port))
2880 pp = rxi_AllocPeer(); /* This bzero's *pp */
2881 pp->host = host; /* set here or in InitPeerParams is zero */
2883 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2884 queue_Init(&pp->congestionQueue);
2885 queue_Init(&pp->rpcStats);
2886 pp->next = rx_peerHashTable[hashIndex];
2887 rx_peerHashTable[hashIndex] = pp;
2888 rxi_InitPeerParams(pp);
2889 if (rx_stats_active)
2890 rx_atomic_inc(&rx_stats.nPeerStructs);
2897 origPeer->refCount--;
2898 MUTEX_EXIT(&rx_peerHashTable_lock);
2903 /* Find the connection at (host, port) started at epoch, and with the
2904 * given connection id. Creates the server connection if necessary.
2905 * The type specifies whether a client connection or a server
2906 * connection is desired. In both cases, (host, port) specify the
2907 * peer's (host, pair) pair. Client connections are not made
2908 * automatically by this routine. The parameter socket gives the
2909 * socket descriptor on which the packet was received. This is used,
2910 * in the case of server connections, to check that *new* connections
2911 * come via a valid (port, serviceId). Finally, the securityIndex
2912 * parameter must match the existing index for the connection. If a
2913 * server connection is created, it will be created using the supplied
2914 * index, if the index is valid for this service */
2915 struct rx_connection *
2916 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2917 u_short port, u_short serviceId, afs_uint32 cid,
2918 afs_uint32 epoch, int type, u_int securityIndex)
2920 int hashindex, flag, i;
2921 struct rx_connection *conn;
2922 hashindex = CONN_HASH(host, port, cid, epoch, type);
2923 MUTEX_ENTER(&rx_connHashTable_lock);
2924 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2925 rx_connHashTable[hashindex],
2928 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2929 && (epoch == conn->epoch)) {
2930 struct rx_peer *pp = conn->peer;
2931 if (securityIndex != conn->securityIndex) {
2932 /* this isn't supposed to happen, but someone could forge a packet
2933 * like this, and there seems to be some CM bug that makes this
2934 * happen from time to time -- in which case, the fileserver
2936 MUTEX_EXIT(&rx_connHashTable_lock);
2937 return (struct rx_connection *)0;
2939 if (pp->host == host && pp->port == port)
2941 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2943 /* So what happens when it's a callback connection? */
2944 if ( /*type == RX_CLIENT_CONNECTION && */
2945 (conn->epoch & 0x80000000))
2949 /* the connection rxLastConn that was used the last time is not the
2950 ** one we are looking for now. Hence, start searching in the hash */
2952 conn = rx_connHashTable[hashindex];
2957 struct rx_service *service;
2958 if (type == RX_CLIENT_CONNECTION) {
2959 MUTEX_EXIT(&rx_connHashTable_lock);
2960 return (struct rx_connection *)0;
2962 service = rxi_FindService(socket, serviceId);
2963 if (!service || (securityIndex >= service->nSecurityObjects)
2964 || (service->securityObjects[securityIndex] == 0)) {
2965 MUTEX_EXIT(&rx_connHashTable_lock);
2966 return (struct rx_connection *)0;
2968 conn = rxi_AllocConnection(); /* This bzero's the connection */
2969 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2970 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2971 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2972 conn->next = rx_connHashTable[hashindex];
2973 rx_connHashTable[hashindex] = conn;
2974 conn->peer = rxi_FindPeer(host, port, 0, 1);
2975 conn->type = RX_SERVER_CONNECTION;
2976 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2977 conn->epoch = epoch;
2978 conn->cid = cid & RX_CIDMASK;
2979 /* conn->serial = conn->lastSerial = 0; */
2980 /* conn->timeout = 0; */
2981 conn->ackRate = RX_FAST_ACK_RATE;
2982 conn->service = service;
2983 conn->serviceId = serviceId;
2984 conn->securityIndex = securityIndex;
2985 conn->securityObject = service->securityObjects[securityIndex];
2986 conn->nSpecific = 0;
2987 conn->specific = NULL;
2988 rx_SetConnDeadTime(conn, service->connDeadTime);
2989 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2990 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2991 for (i = 0; i < RX_MAXCALLS; i++) {
2992 conn->twind[i] = rx_initSendWindow;
2993 conn->rwind[i] = rx_initReceiveWindow;
2995 /* Notify security object of the new connection */
2996 RXS_NewConnection(conn->securityObject, conn);
2997 /* XXXX Connection timeout? */
2998 if (service->newConnProc)
2999 (*service->newConnProc) (conn);
3000 if (rx_stats_active)
3001 rx_atomic_inc(&rx_stats.nServerConns);
3004 MUTEX_ENTER(&rx_refcnt_mutex);
3006 MUTEX_EXIT(&rx_refcnt_mutex);
3008 rxLastConn = conn; /* store this connection as the last conn used */
3009 MUTEX_EXIT(&rx_connHashTable_lock);
3014 * Timeout a call on a busy call channel if appropriate.
3016 * @param[in] call The busy call.
3018 * @pre 'call' is marked as busy (namely,
3019 * call->conn->lastBusy[call->channel] != 0)
3021 * @pre call->lock is held
3022 * @pre rxi_busyChannelError is nonzero
3024 * @note call->lock is dropped and reacquired
3027 rxi_CheckBusy(struct rx_call *call)
3029 struct rx_connection *conn = call->conn;
3030 int channel = call->channel;
3031 int freechannel = 0;
3033 afs_uint32 callNumber = *call->callNumber;
3035 MUTEX_EXIT(&call->lock);
3037 MUTEX_ENTER(&conn->conn_call_lock);
3039 /* Are there any other call slots on this conn that we should try? Look for
3040 * slots that are empty and are either non-busy, or were marked as busy
3041 * longer than conn->secondsUntilDead seconds before this call started. */
3043 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3045 /* only look at channels that aren't us */
3049 if (conn->lastBusy[i]) {
3050 /* if this channel looked busy too recently, don't look at it */
3051 if (conn->lastBusy[i] >= call->startTime.sec) {
3054 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3059 if (conn->call[i]) {
3060 struct rx_call *tcall = conn->call[i];
3061 MUTEX_ENTER(&tcall->lock);
3062 if (tcall->state == RX_STATE_DALLY) {
3065 MUTEX_EXIT(&tcall->lock);
3071 MUTEX_EXIT(&conn->conn_call_lock);
3073 MUTEX_ENTER(&call->lock);
3075 /* Since the call->lock and conn->conn_call_lock have been released it is
3076 * possible that (1) the call may no longer be busy and/or (2) the call may
3077 * have been reused by another waiting thread. Therefore, we must confirm
3078 * that the call state has not changed when deciding whether or not to
3079 * force this application thread to retry by forcing a Timeout error. */
3081 if (freechannel && *call->callNumber == callNumber &&
3082 (call->flags & RX_CALL_PEER_BUSY)) {
3083 /* Since 'freechannel' is set, there exists another channel in this
3084 * rx_conn that the application thread might be able to use. We know
3085 * that we have the correct call since callNumber is unchanged, and we
3086 * know that the call is still busy. So, set the call error state to
3087 * rxi_busyChannelError so the application can retry the request,
3088 * presumably on a less-busy call channel. */
3090 rxi_CallError(call, rxi_busyChannelError);
3094 /* There are two packet tracing routines available for testing and monitoring
3095 * Rx. One is called just after every packet is received and the other is
3096 * called just before every packet is sent. Received packets, have had their
3097 * headers decoded, and packets to be sent have not yet had their headers
3098 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3099 * containing the network address. Both can be modified. The return value, if
3100 * non-zero, indicates that the packet should be dropped. */
3102 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3103 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3105 /* A packet has been received off the interface. Np is the packet, socket is
3106 * the socket number it was received from (useful in determining which service
3107 * this packet corresponds to), and (host, port) reflect the host,port of the
3108 * sender. This call returns the packet to the caller if it is finished with
3109 * it, rather than de-allocating it, just as a small performance hack */
3112 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3113 afs_uint32 host, u_short port, int *tnop,
3114 struct rx_call **newcallp)
3116 struct rx_call *call;
3117 struct rx_connection *conn;
3119 afs_uint32 currentCallNumber;
3125 struct rx_packet *tnp;
3128 /* We don't print out the packet until now because (1) the time may not be
3129 * accurate enough until now in the lwp implementation (rx_Listener only gets
3130 * the time after the packet is read) and (2) from a protocol point of view,
3131 * this is the first time the packet has been seen */
3132 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3133 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3134 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3135 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3136 np->header.epoch, np->header.cid, np->header.callNumber,
3137 np->header.seq, np->header.flags, np));
3140 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3141 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3144 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3145 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3148 /* If an input tracer function is defined, call it with the packet and
3149 * network address. Note this function may modify its arguments. */
3150 if (rx_justReceived) {
3151 struct sockaddr_in addr;
3153 addr.sin_family = AF_INET;
3154 addr.sin_port = port;
3155 addr.sin_addr.s_addr = host;
3156 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3157 addr.sin_len = sizeof(addr);
3158 #endif /* AFS_OSF_ENV */
3159 drop = (*rx_justReceived) (np, &addr);
3160 /* drop packet if return value is non-zero */
3163 port = addr.sin_port; /* in case fcn changed addr */
3164 host = addr.sin_addr.s_addr;
3168 /* If packet was not sent by the client, then *we* must be the client */
3169 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3170 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3172 /* Find the connection (or fabricate one, if we're the server & if
3173 * necessary) associated with this packet */
3175 rxi_FindConnection(socket, host, port, np->header.serviceId,
3176 np->header.cid, np->header.epoch, type,
3177 np->header.securityIndex);
3180 /* If no connection found or fabricated, just ignore the packet.
3181 * (An argument could be made for sending an abort packet for
3186 /* If the connection is in an error state, send an abort packet and ignore
3187 * the incoming packet */
3189 /* Don't respond to an abort packet--we don't want loops! */
3190 MUTEX_ENTER(&conn->conn_data_lock);
3191 if (np->header.type != RX_PACKET_TYPE_ABORT)
3192 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3193 MUTEX_ENTER(&rx_refcnt_mutex);
3195 MUTEX_EXIT(&rx_refcnt_mutex);
3196 MUTEX_EXIT(&conn->conn_data_lock);
3200 /* Check for connection-only requests (i.e. not call specific). */
3201 if (np->header.callNumber == 0) {
3202 switch (np->header.type) {
3203 case RX_PACKET_TYPE_ABORT: {
3204 /* What if the supplied error is zero? */
3205 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3206 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3207 rxi_ConnectionError(conn, errcode);
3208 MUTEX_ENTER(&rx_refcnt_mutex);
3210 MUTEX_EXIT(&rx_refcnt_mutex);
3213 case RX_PACKET_TYPE_CHALLENGE:
3214 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3215 MUTEX_ENTER(&rx_refcnt_mutex);
3217 MUTEX_EXIT(&rx_refcnt_mutex);
3219 case RX_PACKET_TYPE_RESPONSE:
3220 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3221 MUTEX_ENTER(&rx_refcnt_mutex);
3223 MUTEX_EXIT(&rx_refcnt_mutex);
3225 case RX_PACKET_TYPE_PARAMS:
3226 case RX_PACKET_TYPE_PARAMS + 1:
3227 case RX_PACKET_TYPE_PARAMS + 2:
3228 /* ignore these packet types for now */
3229 MUTEX_ENTER(&rx_refcnt_mutex);
3231 MUTEX_EXIT(&rx_refcnt_mutex);
3236 /* Should not reach here, unless the peer is broken: send an
3238 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3239 MUTEX_ENTER(&conn->conn_data_lock);
3240 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3241 MUTEX_ENTER(&rx_refcnt_mutex);
3243 MUTEX_EXIT(&rx_refcnt_mutex);
3244 MUTEX_EXIT(&conn->conn_data_lock);
3249 channel = np->header.cid & RX_CHANNELMASK;
3250 call = conn->call[channel];
3253 MUTEX_ENTER(&call->lock);
3254 currentCallNumber = conn->callNumber[channel];
3255 } else if (type == RX_SERVER_CONNECTION) { /* No call allocated */
3256 MUTEX_ENTER(&conn->conn_call_lock);
3257 call = conn->call[channel];
3259 MUTEX_ENTER(&call->lock);
3260 MUTEX_EXIT(&conn->conn_call_lock);
3261 currentCallNumber = conn->callNumber[channel];
3263 call = rxi_NewCall(conn, channel); /* returns locked call */
3264 MUTEX_EXIT(&conn->conn_call_lock);
3265 *call->callNumber = currentCallNumber = np->header.callNumber;
3267 if (np->header.callNumber == 0)
3268 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3269 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3270 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3271 np->header.flags, np, np->length));
3273 call->state = RX_STATE_PRECALL;
3274 clock_GetTime(&call->queueTime);
3275 hzero(call->bytesSent);
3276 hzero(call->bytesRcvd);
3278 * If the number of queued calls exceeds the overload
3279 * threshold then abort this call.
3281 if ((rx_BusyThreshold > 0) &&
3282 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3283 struct rx_packet *tp;
3285 rxi_CallError(call, rx_BusyError);
3286 tp = rxi_SendCallAbort(call, np, 1, 0);
3287 MUTEX_EXIT(&call->lock);
3288 MUTEX_ENTER(&rx_refcnt_mutex);
3290 MUTEX_EXIT(&rx_refcnt_mutex);
3291 if (rx_stats_active)
3292 rx_atomic_inc(&rx_stats.nBusies);
3295 rxi_KeepAliveOn(call);
3297 } else { /* RX_CLIENT_CONNECTION and No call allocated */
3298 /* This packet can't be for this call. If the new call address is
3299 * 0 then no call is running on this channel. If there is a call
3300 * then, since this is a client connection we're getting data for
3301 * it must be for the previous call.
3303 if (rx_stats_active)
3304 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3305 MUTEX_ENTER(&rx_refcnt_mutex);
3307 MUTEX_EXIT(&rx_refcnt_mutex);
3311 /* There is a non-NULL locked call at this point */
3312 if (type == RX_SERVER_CONNECTION) { /* We're the server */
3313 if (np->header.callNumber < currentCallNumber) {
3314 MUTEX_EXIT(&call->lock);
3315 if (rx_stats_active)
3316 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3317 MUTEX_ENTER(&rx_refcnt_mutex);
3319 MUTEX_EXIT(&rx_refcnt_mutex);
3321 } else if (np->header.callNumber != currentCallNumber) {
3322 /* Wait until the transmit queue is idle before deciding
3323 * whether to reset the current call. Chances are that the
3324 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3327 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3328 if (call->state == RX_STATE_ACTIVE) {
3329 rxi_WaitforTQBusy(call);
3331 * If we entered error state while waiting,
3332 * must call rxi_CallError to permit rxi_ResetCall
3333 * to processed when the tqWaiter count hits zero.
3336 rxi_CallError(call, call->error);
3337 MUTEX_EXIT(&call->lock);
3338 MUTEX_ENTER(&rx_refcnt_mutex);
3340 MUTEX_EXIT(&rx_refcnt_mutex);
3344 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3345 /* If the new call cannot be taken right now send a busy and set
3346 * the error condition in this call, so that it terminates as
3347 * quickly as possible */
3348 if (call->state == RX_STATE_ACTIVE) {
3349 struct rx_packet *tp;
3351 rxi_CallError(call, RX_CALL_DEAD);
3352 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3354 MUTEX_EXIT(&call->lock);
3355 MUTEX_ENTER(&rx_refcnt_mutex);
3357 MUTEX_EXIT(&rx_refcnt_mutex);
3360 rxi_ResetCall(call, 0);
3361 *call->callNumber = np->header.callNumber;
3363 if (np->header.callNumber == 0)
3364 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3365 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3366 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3367 np->header.flags, np, np->length));
3369 call->state = RX_STATE_PRECALL;
3370 clock_GetTime(&call->queueTime);
3371 hzero(call->bytesSent);
3372 hzero(call->bytesRcvd);
3374 * If the number of queued calls exceeds the overload
3375 * threshold then abort this call.
3377 if ((rx_BusyThreshold > 0) &&
3378 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3379 struct rx_packet *tp;
3381 rxi_CallError(call, rx_BusyError);
3382 tp = rxi_SendCallAbort(call, np, 1, 0);
3383 MUTEX_EXIT(&call->lock);
3384 MUTEX_ENTER(&rx_refcnt_mutex);
3386 MUTEX_EXIT(&rx_refcnt_mutex);
3387 if (rx_stats_active)
3388 rx_atomic_inc(&rx_stats.nBusies);
3391 rxi_KeepAliveOn(call);
3393 /* Continuing call; do nothing here. */
3395 } else { /* we're the client */
3396 /* Ignore all incoming acknowledgements for calls in DALLY state */
3397 if ((call->state == RX_STATE_DALLY)
3398 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3399 if (rx_stats_active)
3400 rx_atomic_inc(&rx_stats.ignorePacketDally);
3401 MUTEX_EXIT(&call->lock);
3402 MUTEX_ENTER(&rx_refcnt_mutex);
3404 MUTEX_EXIT(&rx_refcnt_mutex);
3408 /* Ignore anything that's not relevant to the current call. If there
3409 * isn't a current call, then no packet is relevant. */
3410 if (np->header.callNumber != currentCallNumber) {
3411 if (rx_stats_active)
3412 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3413 MUTEX_EXIT(&call->lock);
3414 MUTEX_ENTER(&rx_refcnt_mutex);
3416 MUTEX_EXIT(&rx_refcnt_mutex);
3419 /* If the service security object index stamped in the packet does not
3420 * match the connection's security index, ignore the packet */
3421 if (np->header.securityIndex != conn->securityIndex) {
3422 MUTEX_EXIT(&call->lock);
3423 MUTEX_ENTER(&rx_refcnt_mutex);
3425 MUTEX_EXIT(&rx_refcnt_mutex);
3429 /* If we're receiving the response, then all transmit packets are
3430 * implicitly acknowledged. Get rid of them. */
3431 if (np->header.type == RX_PACKET_TYPE_DATA) {
3432 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3433 /* XXX Hack. Because we must release the global rx lock when
3434 * sending packets (osi_NetSend) we drop all acks while we're
3435 * traversing the tq in rxi_Start sending packets out because
3436 * packets may move to the freePacketQueue as result of being here!
3437 * So we drop these packets until we're safely out of the
3438 * traversing. Really ugly!
3439 * For fine grain RX locking, we set the acked field in the
3440 * packets and let rxi_Start remove them from the transmit queue.
3442 if (call->flags & RX_CALL_TQ_BUSY) {
3443 #ifdef RX_ENABLE_LOCKS
3444 rxi_SetAcksInTransmitQueue(call);
3446 MUTEX_ENTER(&rx_refcnt_mutex);
3448 MUTEX_EXIT(&rx_refcnt_mutex);
3449 return np; /* xmitting; drop packet */
3452 rxi_ClearTransmitQueue(call, 0);
3454 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3455 rxi_ClearTransmitQueue(call, 0);
3456 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3458 if (np->header.type == RX_PACKET_TYPE_ACK) {
3459 /* now check to see if this is an ack packet acknowledging that the
3460 * server actually *lost* some hard-acked data. If this happens we
3461 * ignore this packet, as it may indicate that the server restarted in
3462 * the middle of a call. It is also possible that this is an old ack
3463 * packet. We don't abort the connection in this case, because this
3464 * *might* just be an old ack packet. The right way to detect a server
3465 * restart in the midst of a call is to notice that the server epoch
3467 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3468 * XXX unacknowledged. I think that this is off-by-one, but
3469 * XXX I don't dare change it just yet, since it will
3470 * XXX interact badly with the server-restart detection
3471 * XXX code in receiveackpacket. */
3472 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3473 if (rx_stats_active)
3474 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3475 MUTEX_EXIT(&call->lock);
3476 MUTEX_ENTER(&rx_refcnt_mutex);
3478 MUTEX_EXIT(&rx_refcnt_mutex);
3482 } /* else not a data packet */
3485 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3486 /* Set remote user defined status from packet */
3487 call->remoteStatus = np->header.userStatus;
3489 /* Note the gap between the expected next packet and the actual
3490 * packet that arrived, when the new packet has a smaller serial number
3491 * than expected. Rioses frequently reorder packets all by themselves,
3492 * so this will be quite important with very large window sizes.
3493 * Skew is checked against 0 here to avoid any dependence on the type of
3494 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3496 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3497 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3498 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3500 MUTEX_ENTER(&conn->conn_data_lock);
3501 skew = conn->lastSerial - np->header.serial;
3502 conn->lastSerial = np->header.serial;
3503 MUTEX_EXIT(&conn->conn_data_lock);
3505 struct rx_peer *peer;
3507 if (skew > peer->inPacketSkew) {
3508 dpf(("*** In skew changed from %d to %d\n",
3509 peer->inPacketSkew, skew));
3510 peer->inPacketSkew = skew;
3514 /* Now do packet type-specific processing */
3515 switch (np->header.type) {
3516 case RX_PACKET_TYPE_DATA:
3517 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3520 case RX_PACKET_TYPE_ACK:
3521 /* Respond immediately to ack packets requesting acknowledgement
3523 if (np->header.flags & RX_REQUEST_ACK) {
3525 (void)rxi_SendCallAbort(call, 0, 1, 0);
3527 (void)rxi_SendAck(call, 0, np->header.serial,
3528 RX_ACK_PING_RESPONSE, 1);
3530 np = rxi_ReceiveAckPacket(call, np, 1);
3532 case RX_PACKET_TYPE_ABORT: {
3533 /* An abort packet: reset the call, passing the error up to the user. */
3534 /* What if error is zero? */
3535 /* What if the error is -1? the application will treat it as a timeout. */
3536 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3537 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3538 rxi_CallError(call, errdata);
3539 MUTEX_EXIT(&call->lock);
3540 MUTEX_ENTER(&rx_refcnt_mutex);
3542 MUTEX_EXIT(&rx_refcnt_mutex);
3543 return np; /* xmitting; drop packet */
3545 case RX_PACKET_TYPE_BUSY: {
3546 struct clock busyTime;
3548 clock_GetTime(&busyTime);
3550 MUTEX_EXIT(&call->lock);
3552 MUTEX_ENTER(&conn->conn_call_lock);
3553 MUTEX_ENTER(&call->lock);
3554 conn->lastBusy[call->channel] = busyTime.sec;
3555 call->flags |= RX_CALL_PEER_BUSY;
3556 MUTEX_EXIT(&call->lock);
3557 MUTEX_EXIT(&conn->conn_call_lock);
3559 MUTEX_ENTER(&rx_refcnt_mutex);
3561 MUTEX_EXIT(&rx_refcnt_mutex);
3565 case RX_PACKET_TYPE_ACKALL:
3566 /* All packets acknowledged, so we can drop all packets previously
3567 * readied for sending */
3568 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3569 /* XXX Hack. We because we can't release the global rx lock when
3570 * sending packets (osi_NetSend) we drop all ack pkts while we're
3571 * traversing the tq in rxi_Start sending packets out because
3572 * packets may move to the freePacketQueue as result of being
3573 * here! So we drop these packets until we're safely out of the
3574 * traversing. Really ugly!
3575 * For fine grain RX locking, we set the acked field in the packets
3576 * and let rxi_Start remove the packets from the transmit queue.
3578 if (call->flags & RX_CALL_TQ_BUSY) {
3579 #ifdef RX_ENABLE_LOCKS
3580 rxi_SetAcksInTransmitQueue(call);
3582 #else /* RX_ENABLE_LOCKS */
3583 MUTEX_EXIT(&call->lock);
3584 MUTEX_ENTER(&rx_refcnt_mutex);
3586 MUTEX_EXIT(&rx_refcnt_mutex);
3587 return np; /* xmitting; drop packet */
3588 #endif /* RX_ENABLE_LOCKS */
3590 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3591 rxi_ClearTransmitQueue(call, 0);
3594 /* Should not reach here, unless the peer is broken: send an abort
3596 rxi_CallError(call, RX_PROTOCOL_ERROR);
3597 np = rxi_SendCallAbort(call, np, 1, 0);
3600 /* Note when this last legitimate packet was received, for keep-alive
3601 * processing. Note, we delay getting the time until now in the hope that
3602 * the packet will be delivered to the user before any get time is required
3603 * (if not, then the time won't actually be re-evaluated here). */
3604 call->lastReceiveTime = clock_Sec();
3605 /* we've received a legit packet, so the channel is not busy */
3606 call->flags &= ~RX_CALL_PEER_BUSY;
3607 MUTEX_EXIT(&call->lock);
3608 MUTEX_ENTER(&rx_refcnt_mutex);
3610 MUTEX_EXIT(&rx_refcnt_mutex);
3614 /* return true if this is an "interesting" connection from the point of view
3615 of someone trying to debug the system */
3617 rxi_IsConnInteresting(struct rx_connection *aconn)
3620 struct rx_call *tcall;
3622 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3625 for (i = 0; i < RX_MAXCALLS; i++) {
3626 tcall = aconn->call[i];
3628 if ((tcall->state == RX_STATE_PRECALL)
3629 || (tcall->state == RX_STATE_ACTIVE))
3631 if ((tcall->mode == RX_MODE_SENDING)
3632 || (tcall->mode == RX_MODE_RECEIVING))
3640 /* if this is one of the last few packets AND it wouldn't be used by the
3641 receiving call to immediately satisfy a read request, then drop it on
3642 the floor, since accepting it might prevent a lock-holding thread from
3643 making progress in its reading. If a call has been cleared while in
3644 the precall state then ignore all subsequent packets until the call
3645 is assigned to a thread. */
3648 TooLow(struct rx_packet *ap, struct rx_call *acall)
3652 MUTEX_ENTER(&rx_quota_mutex);
3653 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3654 && (acall->state == RX_STATE_PRECALL))
3655 || ((rx_nFreePackets < rxi_dataQuota + 2)
3656 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3657 && (acall->flags & RX_CALL_READER_WAIT)))) {
3660 MUTEX_EXIT(&rx_quota_mutex);
3666 * Clear the attach wait flag on a connection and proceed.
3668 * Any processing waiting for a connection to be attached should be
3669 * unblocked. We clear the flag and do any other needed tasks.
3672 * the conn to unmark waiting for attach
3674 * @pre conn's conn_data_lock must be locked before calling this function
3678 rxi_ConnClearAttachWait(struct rx_connection *conn)
3680 /* Indicate that rxi_CheckReachEvent is no longer running by
3681 * clearing the flag. Must be atomic under conn_data_lock to
3682 * avoid a new call slipping by: rxi_CheckConnReach holds
3683 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3685 conn->flags &= ~RX_CONN_ATTACHWAIT;
3686 if (conn->flags & RX_CONN_NAT_PING) {
3687 conn->flags &= ~RX_CONN_NAT_PING;
3688 rxi_ScheduleNatKeepAliveEvent(conn);
3693 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3695 struct rx_connection *conn = arg1;
3696 struct rx_call *acall = arg2;
3697 struct rx_call *call = acall;
3698 struct clock when, now;
3701 MUTEX_ENTER(&conn->conn_data_lock);
3704 rxevent_Put(conn->checkReachEvent);
3705 conn->checkReachEvent = NULL;
3708 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3710 MUTEX_ENTER(&rx_refcnt_mutex);
3712 MUTEX_EXIT(&rx_refcnt_mutex);
3714 MUTEX_EXIT(&conn->conn_data_lock);
3718 MUTEX_ENTER(&conn->conn_call_lock);
3719 MUTEX_ENTER(&conn->conn_data_lock);
3720 for (i = 0; i < RX_MAXCALLS; i++) {
3721 struct rx_call *tc = conn->call[i];
3722 if (tc && tc->state == RX_STATE_PRECALL) {
3728 rxi_ConnClearAttachWait(conn);
3729 MUTEX_EXIT(&conn->conn_data_lock);
3730 MUTEX_EXIT(&conn->conn_call_lock);
3735 MUTEX_ENTER(&call->lock);
3736 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3738 MUTEX_EXIT(&call->lock);
3740 clock_GetTime(&now);
3742 when.sec += RX_CHECKREACH_TIMEOUT;
3743 MUTEX_ENTER(&conn->conn_data_lock);
3744 if (!conn->checkReachEvent) {
3745 MUTEX_ENTER(&rx_refcnt_mutex);
3747 MUTEX_EXIT(&rx_refcnt_mutex);
3748 conn->checkReachEvent = rxevent_Post(&when, &now,
3749 rxi_CheckReachEvent, conn,
3752 MUTEX_EXIT(&conn->conn_data_lock);
3758 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3760 struct rx_service *service = conn->service;
3761 struct rx_peer *peer = conn->peer;
3762 afs_uint32 now, lastReach;
3764 if (service->checkReach == 0)
3768 MUTEX_ENTER(&peer->peer_lock);
3769 lastReach = peer->lastReachTime;
3770 MUTEX_EXIT(&peer->peer_lock);
3771 if (now - lastReach < RX_CHECKREACH_TTL)
3774 MUTEX_ENTER(&conn->conn_data_lock);
3775 if (conn->flags & RX_CONN_ATTACHWAIT) {
3776 MUTEX_EXIT(&conn->conn_data_lock);
3779 conn->flags |= RX_CONN_ATTACHWAIT;
3780 MUTEX_EXIT(&conn->conn_data_lock);
3781 if (!conn->checkReachEvent)
3782 rxi_CheckReachEvent(NULL, conn, call, 0);
3787 /* try to attach call, if authentication is complete */
3789 TryAttach(struct rx_call *acall, osi_socket socket,
3790 int *tnop, struct rx_call **newcallp,
3793 struct rx_connection *conn = acall->conn;
3795 if (conn->type == RX_SERVER_CONNECTION
3796 && acall->state == RX_STATE_PRECALL) {
3797 /* Don't attach until we have any req'd. authentication. */
3798 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3799 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3800 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3801 /* Note: this does not necessarily succeed; there
3802 * may not any proc available
3805 rxi_ChallengeOn(acall->conn);
3810 /* A data packet has been received off the interface. This packet is
3811 * appropriate to the call (the call is in the right state, etc.). This
3812 * routine can return a packet to the caller, for re-use */
3815 rxi_ReceiveDataPacket(struct rx_call *call,
3816 struct rx_packet *np, int istack,
3817 osi_socket socket, afs_uint32 host, u_short port,
3818 int *tnop, struct rx_call **newcallp)
3820 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3825 afs_uint32 serial=0, flags=0;
3827 struct rx_packet *tnp;
3828 if (rx_stats_active)
3829 rx_atomic_inc(&rx_stats.dataPacketsRead);
3832 /* If there are no packet buffers, drop this new packet, unless we can find
3833 * packet buffers from inactive calls */
3835 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3836 MUTEX_ENTER(&rx_freePktQ_lock);
3837 rxi_NeedMorePackets = TRUE;
3838 MUTEX_EXIT(&rx_freePktQ_lock);
3839 if (rx_stats_active)
3840 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3841 call->rprev = np->header.serial;
3842 rxi_calltrace(RX_TRACE_DROP, call);
3843 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3844 /* We used to clear the receive queue here, in an attempt to free
3845 * packets. However this is unsafe if the queue has received a
3846 * soft ACK for the final packet */
3847 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3849 /* we've damaged this call already, might as well do it in. */
3855 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3856 * packet is one of several packets transmitted as a single
3857 * datagram. Do not send any soft or hard acks until all packets
3858 * in a jumbogram have been processed. Send negative acks right away.
3860 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3861 /* tnp is non-null when there are more packets in the
3862 * current jumbo gram */
3869 seq = np->header.seq;
3870 serial = np->header.serial;
3871 flags = np->header.flags;
3873 /* If the call is in an error state, send an abort message */
3875 return rxi_SendCallAbort(call, np, istack, 0);
3877 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3878 * AFS 3.5 jumbogram. */
3879 if (flags & RX_JUMBO_PACKET) {
3880 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3885 if (np->header.spare != 0) {
3886 MUTEX_ENTER(&call->conn->conn_data_lock);
3887 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3888 MUTEX_EXIT(&call->conn->conn_data_lock);
3891 /* The usual case is that this is the expected next packet */
3892 if (seq == call->rnext) {
3894 /* Check to make sure it is not a duplicate of one already queued */
3895 if (queue_IsNotEmpty(&call->rq)
3896 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3897 if (rx_stats_active)
3898 rx_atomic_inc(&rx_stats.dupPacketsRead);
3899 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3900 rxevent_Cancel(&call->delayedAckEvent, call,
3901 RX_CALL_REFCOUNT_DELAY);
3902 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3908 /* It's the next packet. Stick it on the receive queue
3909 * for this call. Set newPackets to make sure we wake
3910 * the reader once all packets have been processed */
3911 #ifdef RX_TRACK_PACKETS
3912 np->flags |= RX_PKTFLAG_RQ;
3914 queue_Prepend(&call->rq, np);
3915 #ifdef RXDEBUG_PACKET
3917 #endif /* RXDEBUG_PACKET */
3919 np = NULL; /* We can't use this anymore */
3922 /* If an ack is requested then set a flag to make sure we
3923 * send an acknowledgement for this packet */
3924 if (flags & RX_REQUEST_ACK) {
3925 ackNeeded = RX_ACK_REQUESTED;
3928 /* Keep track of whether we have received the last packet */
3929 if (flags & RX_LAST_PACKET) {
3930 call->flags |= RX_CALL_HAVE_LAST;
3934 /* Check whether we have all of the packets for this call */
3935 if (call->flags & RX_CALL_HAVE_LAST) {
3936 afs_uint32 tseq; /* temporary sequence number */
3937 struct rx_packet *tp; /* Temporary packet pointer */
3938 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3940 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3941 if (tseq != tp->header.seq)
3943 if (tp->header.flags & RX_LAST_PACKET) {
3944 call->flags |= RX_CALL_RECEIVE_DONE;
3951 /* Provide asynchronous notification for those who want it
3952 * (e.g. multi rx) */
3953 if (call->arrivalProc) {
3954 (*call->arrivalProc) (call, call->arrivalProcHandle,
3955 call->arrivalProcArg);
3956 call->arrivalProc = (void (*)())0;
3959 /* Update last packet received */
3962 /* If there is no server process serving this call, grab
3963 * one, if available. We only need to do this once. If a
3964 * server thread is available, this thread becomes a server
3965 * thread and the server thread becomes a listener thread. */
3967 TryAttach(call, socket, tnop, newcallp, 0);
3970 /* This is not the expected next packet. */
3972 /* Determine whether this is a new or old packet, and if it's
3973 * a new one, whether it fits into the current receive window.
3974 * Also figure out whether the packet was delivered in sequence.
3975 * We use the prev variable to determine whether the new packet
3976 * is the successor of its immediate predecessor in the
3977 * receive queue, and the missing flag to determine whether
3978 * any of this packets predecessors are missing. */
3980 afs_uint32 prev; /* "Previous packet" sequence number */
3981 struct rx_packet *tp; /* Temporary packet pointer */
3982 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3983 int missing; /* Are any predecessors missing? */
3985 /* If the new packet's sequence number has been sent to the
3986 * application already, then this is a duplicate */
3987 if (seq < call->rnext) {
3988 if (rx_stats_active)
3989 rx_atomic_inc(&rx_stats.dupPacketsRead);
3990 rxevent_Cancel(&call->delayedAckEvent, call,
3991 RX_CALL_REFCOUNT_DELAY);
3992 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3998 /* If the sequence number is greater than what can be
3999 * accomodated by the current window, then send a negative
4000 * acknowledge and drop the packet */
4001 if ((call->rnext + call->rwind) <= seq) {
4002 rxevent_Cancel(&call->delayedAckEvent, call,
4003 RX_CALL_REFCOUNT_DELAY);
4004 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
4011 /* Look for the packet in the queue of old received packets */
4012 for (prev = call->rnext - 1, missing =
4013 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
4014 /*Check for duplicate packet */
4015 if (seq == tp->header.seq) {
4016 if (rx_stats_active)
4017 rx_atomic_inc(&rx_stats.dupPacketsRead);
4018 rxevent_Cancel(&call->delayedAckEvent, call,
4019 RX_CALL_REFCOUNT_DELAY);
4020 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
4026 /* If we find a higher sequence packet, break out and
4027 * insert the new packet here. */
4028 if (seq < tp->header.seq)
4030 /* Check for missing packet */
4031 if (tp->header.seq != prev + 1) {
4035 prev = tp->header.seq;
4038 /* Keep track of whether we have received the last packet. */
4039 if (flags & RX_LAST_PACKET) {
4040 call->flags |= RX_CALL_HAVE_LAST;
4043 /* It's within the window: add it to the the receive queue.
4044 * tp is left by the previous loop either pointing at the
4045 * packet before which to insert the new packet, or at the
4046 * queue head if the queue is empty or the packet should be
4048 #ifdef RX_TRACK_PACKETS
4049 np->flags |= RX_PKTFLAG_RQ;
4051 #ifdef RXDEBUG_PACKET
4053 #endif /* RXDEBUG_PACKET */
4054 queue_InsertBefore(tp, np);
4058 /* Check whether we have all of the packets for this call */
4059 if ((call->flags & RX_CALL_HAVE_LAST)
4060 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4061 afs_uint32 tseq; /* temporary sequence number */
4064 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
4065 if (tseq != tp->header.seq)
4067 if (tp->header.flags & RX_LAST_PACKET) {
4068 call->flags |= RX_CALL_RECEIVE_DONE;
4075 /* We need to send an ack of the packet is out of sequence,
4076 * or if an ack was requested by the peer. */
4077 if (seq != prev + 1 || missing) {
4078 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4079 } else if (flags & RX_REQUEST_ACK) {
4080 ackNeeded = RX_ACK_REQUESTED;
4083 /* Acknowledge the last packet for each call */
4084 if (flags & RX_LAST_PACKET) {
4095 * If the receiver is waiting for an iovec, fill the iovec
4096 * using the data from the receive queue */
4097 if (call->flags & RX_CALL_IOVEC_WAIT) {
4098 didHardAck = rxi_FillReadVec(call, serial);
4099 /* the call may have been aborted */
4108 /* Wakeup the reader if any */
4109 if ((call->flags & RX_CALL_READER_WAIT)
4110 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4111 || (call->iovNext >= call->iovMax)
4112 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4113 call->flags &= ~RX_CALL_READER_WAIT;
4114 #ifdef RX_ENABLE_LOCKS
4115 CV_BROADCAST(&call->cv_rq);
4117 osi_rxWakeup(&call->rq);
4123 * Send an ack when requested by the peer, or once every
4124 * rxi_SoftAckRate packets until the last packet has been
4125 * received. Always send a soft ack for the last packet in
4126 * the server's reply. */
4128 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4129 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4130 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4131 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4132 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4133 } else if (call->nSoftAcks) {
4134 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4135 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4137 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4138 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4139 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4146 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4148 struct rx_peer *peer = conn->peer;
4150 MUTEX_ENTER(&peer->peer_lock);
4151 peer->lastReachTime = clock_Sec();
4152 MUTEX_EXIT(&peer->peer_lock);
4154 MUTEX_ENTER(&conn->conn_data_lock);
4155 if (conn->flags & RX_CONN_ATTACHWAIT) {
4158 rxi_ConnClearAttachWait(conn);
4159 MUTEX_EXIT(&conn->conn_data_lock);
4161 for (i = 0; i < RX_MAXCALLS; i++) {
4162 struct rx_call *call = conn->call[i];
4165 MUTEX_ENTER(&call->lock);
4166 /* tnop can be null if newcallp is null */
4167 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4169 MUTEX_EXIT(&call->lock);
4173 MUTEX_EXIT(&conn->conn_data_lock);
4176 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4178 rx_ack_reason(int reason)
4181 case RX_ACK_REQUESTED:
4183 case RX_ACK_DUPLICATE:
4185 case RX_ACK_OUT_OF_SEQUENCE:
4187 case RX_ACK_EXCEEDS_WINDOW:
4189 case RX_ACK_NOSPACE:
4193 case RX_ACK_PING_RESPONSE:
4206 /* The real smarts of the whole thing. */
4208 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4211 struct rx_ackPacket *ap;
4213 struct rx_packet *tp;
4214 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
4215 struct rx_connection *conn = call->conn;
4216 struct rx_peer *peer = conn->peer;
4217 struct clock now; /* Current time, for RTT calculations */
4221 /* because there are CM's that are bogus, sending weird values for this. */
4222 afs_uint32 skew = 0;
4227 int newAckCount = 0;
4228 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4229 int pktsize = 0; /* Set if we need to update the peer mtu */
4230 int conn_data_locked = 0;
4232 if (rx_stats_active)
4233 rx_atomic_inc(&rx_stats.ackPacketsRead);
4234 ap = (struct rx_ackPacket *)rx_DataOf(np);
4235 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4237 return np; /* truncated ack packet */
4239 /* depends on ack packet struct */
4240 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4241 first = ntohl(ap->firstPacket);
4242 prev = ntohl(ap->previousPacket);
4243 serial = ntohl(ap->serial);
4244 /* temporarily disabled -- needs to degrade over time
4245 * skew = ntohs(ap->maxSkew); */
4247 /* Ignore ack packets received out of order */
4248 if (first < call->tfirst ||
4249 (first == call->tfirst && prev < call->tprev)) {
4255 if (np->header.flags & RX_SLOW_START_OK) {
4256 call->flags |= RX_CALL_SLOW_START_OK;
4259 if (ap->reason == RX_ACK_PING_RESPONSE)
4260 rxi_UpdatePeerReach(conn, call);
4262 if (conn->lastPacketSizeSeq) {
4263 MUTEX_ENTER(&conn->conn_data_lock);
4264 conn_data_locked = 1;
4265 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4266 pktsize = conn->lastPacketSize;
4267 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4270 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4271 if (!conn_data_locked) {
4272 MUTEX_ENTER(&conn->conn_data_lock);
4273 conn_data_locked = 1;
4275 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4276 /* process mtu ping ack */
4277 pktsize = conn->lastPingSize;
4278 conn->lastPingSizeSer = conn->lastPingSize = 0;
4282 if (conn_data_locked) {
4283 MUTEX_EXIT(&conn->conn_data_lock);
4284 conn_data_locked = 0;
4288 if (rxdebug_active) {
4292 len = _snprintf(msg, sizeof(msg),
4293 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
4294 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4295 ntohl(ap->serial), ntohl(ap->previousPacket),
4296 (unsigned int)np->header.seq, (unsigned int)skew,
4297 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
4301 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4302 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4306 OutputDebugString(msg);
4308 #else /* AFS_NT40_ENV */
4311 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
4312 ap->reason, ntohl(ap->previousPacket),
4313 (unsigned int)np->header.seq, (unsigned int)serial,
4314 (unsigned int)skew, ntohl(ap->firstPacket));
4317 for (offset = 0; offset < nAcks; offset++)
4318 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4323 #endif /* AFS_NT40_ENV */
4326 MUTEX_ENTER(&peer->peer_lock);
4329 * Start somewhere. Can't assume we can send what we can receive,
4330 * but we are clearly receiving.
4332 if (!peer->maxPacketSize)
4333 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4335 if (pktsize > peer->maxPacketSize) {
4336 peer->maxPacketSize = pktsize;
4337 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4338 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4339 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4340 rxi_ScheduleGrowMTUEvent(call, 1);
4345 /* Update the outgoing packet skew value to the latest value of
4346 * the peer's incoming packet skew value. The ack packet, of
4347 * course, could arrive out of order, but that won't affect things
4349 peer->outPacketSkew = skew;
4352 clock_GetTime(&now);
4354 /* The transmit queue splits into 4 sections.
4356 * The first section is packets which have now been acknowledged
4357 * by a window size change in the ack. These have reached the
4358 * application layer, and may be discarded. These are packets
4359 * with sequence numbers < ap->firstPacket.
4361 * The second section is packets which have sequence numbers in
4362 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4363 * contents of the packet's ack array determines whether these
4364 * packets are acknowledged or not.
4366 * The third section is packets which fall above the range
4367 * addressed in the ack packet. These have not yet been received
4370 * The four section is packets which have not yet been transmitted.
4371 * These packets will have a header.serial of 0.
4374 /* First section - implicitly acknowledged packets that can be
4378 tp = queue_First(&call->tq, rx_packet);
4379 while(!queue_IsEnd(&call->tq, tp) && tp->header.seq < first) {
4380 struct rx_packet *next;
4382 next = queue_Next(tp, rx_packet);
4383 call->tfirst = tp->header.seq + 1;
4385 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4387 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4390 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4391 /* XXX Hack. Because we have to release the global rx lock when sending
4392 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4393 * in rxi_Start sending packets out because packets may move to the
4394 * freePacketQueue as result of being here! So we drop these packets until
4395 * we're safely out of the traversing. Really ugly!
4396 * To make it even uglier, if we're using fine grain locking, we can
4397 * set the ack bits in the packets and have rxi_Start remove the packets
4398 * when it's done transmitting.
4400 if (call->flags & RX_CALL_TQ_BUSY) {
4401 #ifdef RX_ENABLE_LOCKS
4402 tp->flags |= RX_PKTFLAG_ACKED;
4403 call->flags |= RX_CALL_TQ_SOME_ACKED;
4404 #else /* RX_ENABLE_LOCKS */
4406 #endif /* RX_ENABLE_LOCKS */
4408 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4411 #ifdef RX_TRACK_PACKETS
4412 tp->flags &= ~RX_PKTFLAG_TQ;
4414 #ifdef RXDEBUG_PACKET
4416 #endif /* RXDEBUG_PACKET */
4417 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4422 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4424 /* Second section of the queue - packets for which we are receiving
4427 * Go through the explicit acks/nacks and record the results in
4428 * the waiting packets. These are packets that can't be released
4429 * yet, even with a positive acknowledge. This positive
4430 * acknowledge only means the packet has been received by the
4431 * peer, not that it will be retained long enough to be sent to
4432 * the peer's upper level. In addition, reset the transmit timers
4433 * of any missing packets (those packets that must be missing
4434 * because this packet was out of sequence) */
4436 call->nSoftAcked = 0;
4438 while (!queue_IsEnd(&call->tq, tp) && tp->header.seq < first + nAcks) {
4439 /* Set the acknowledge flag per packet based on the
4440 * information in the ack packet. An acknowlegded packet can
4441 * be downgraded when the server has discarded a packet it
4442 * soacked previously, or when an ack packet is received
4443 * out of sequence. */
4444 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4445 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4447 tp->flags |= RX_PKTFLAG_ACKED;
4448 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4455 } else /* RX_ACK_TYPE_NACK */ {
4456 tp->flags &= ~RX_PKTFLAG_ACKED;
4460 tp = queue_Next(tp, rx_packet);
4463 /* We don't need to take any action with the 3rd or 4th section in the
4464 * queue - they're not addressed by the contents of this ACK packet.
4467 /* If the window has been extended by this acknowledge packet,
4468 * then wakeup a sender waiting in alloc for window space, or try
4469 * sending packets now, if he's been sitting on packets due to
4470 * lack of window space */
4471 if (call->tnext < (call->tfirst + call->twind)) {
4472 #ifdef RX_ENABLE_LOCKS
4473 CV_SIGNAL(&call->cv_twind);
4475 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4476 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4477 osi_rxWakeup(&call->twind);
4480 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4481 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4485 /* if the ack packet has a receivelen field hanging off it,
4486 * update our state */
4487 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4490 /* If the ack packet has a "recommended" size that is less than
4491 * what I am using now, reduce my size to match */
4492 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4493 (int)sizeof(afs_int32), &tSize);
4494 tSize = (afs_uint32) ntohl(tSize);
4495 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4497 /* Get the maximum packet size to send to this peer */
4498 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4500 tSize = (afs_uint32) ntohl(tSize);
4501 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4502 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4504 /* sanity check - peer might have restarted with different params.
4505 * If peer says "send less", dammit, send less... Peer should never
4506 * be unable to accept packets of the size that prior AFS versions would
4507 * send without asking. */
4508 if (peer->maxMTU != tSize) {
4509 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4511 peer->maxMTU = tSize;
4512 peer->MTU = MIN(tSize, peer->MTU);
4513 call->MTU = MIN(call->MTU, tSize);
4516 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4519 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4520 (int)sizeof(afs_int32), &tSize);
4521 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4522 if (tSize < call->twind) { /* smaller than our send */
4523 call->twind = tSize; /* window, we must send less... */
4524 call->ssthresh = MIN(call->twind, call->ssthresh);
4525 call->conn->twind[call->channel] = call->twind;
4528 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4529 * network MTU confused with the loopback MTU. Calculate the
4530 * maximum MTU here for use in the slow start code below.
4532 /* Did peer restart with older RX version? */
4533 if (peer->maxDgramPackets > 1) {
4534 peer->maxDgramPackets = 1;
4536 } else if (np->length >=
4537 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4540 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4541 sizeof(afs_int32), &tSize);
4542 tSize = (afs_uint32) ntohl(tSize);
4544 * As of AFS 3.5 we set the send window to match the receive window.
4546 if (tSize < call->twind) {
4547 call->twind = tSize;
4548 call->conn->twind[call->channel] = call->twind;
4549 call->ssthresh = MIN(call->twind, call->ssthresh);
4550 } else if (tSize > call->twind) {
4551 call->twind = tSize;
4552 call->conn->twind[call->channel] = call->twind;
4556 * As of AFS 3.5, a jumbogram is more than one fixed size
4557 * packet transmitted in a single UDP datagram. If the remote
4558 * MTU is smaller than our local MTU then never send a datagram
4559 * larger than the natural MTU.
4562 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4563 (int)sizeof(afs_int32), &tSize);
4564 maxDgramPackets = (afs_uint32) ntohl(tSize);
4565 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4567 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4568 if (maxDgramPackets > 1) {
4569 peer->maxDgramPackets = maxDgramPackets;
4570 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4572 peer->maxDgramPackets = 1;
4573 call->MTU = peer->natMTU;
4575 } else if (peer->maxDgramPackets > 1) {
4576 /* Restarted with lower version of RX */
4577 peer->maxDgramPackets = 1;
4579 } else if (peer->maxDgramPackets > 1
4580 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4581 /* Restarted with lower version of RX */
4582 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4583 peer->natMTU = OLD_MAX_PACKET_SIZE;
4584 peer->MTU = OLD_MAX_PACKET_SIZE;
4585 peer->maxDgramPackets = 1;
4586 peer->nDgramPackets = 1;
4588 call->MTU = OLD_MAX_PACKET_SIZE;
4593 * Calculate how many datagrams were successfully received after
4594 * the first missing packet and adjust the negative ack counter
4599 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4600 if (call->nNacks < nNacked) {
4601 call->nNacks = nNacked;
4604 call->nAcks += newAckCount;
4608 /* If the packet contained new acknowledgements, rather than just
4609 * being a duplicate of one we have previously seen, then we can restart
4612 if (newAckCount > 0)
4613 rxi_rto_packet_acked(call, istack);
4615 if (call->flags & RX_CALL_FAST_RECOVER) {
4616 if (newAckCount == 0) {
4617 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4619 call->flags &= ~RX_CALL_FAST_RECOVER;
4620 call->cwind = call->nextCwind;
4621 call->nextCwind = 0;
4624 call->nCwindAcks = 0;
4625 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4626 /* Three negative acks in a row trigger congestion recovery */
4627 call->flags |= RX_CALL_FAST_RECOVER;
4628 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4630 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4631 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4632 call->nextCwind = call->ssthresh;
4635 peer->MTU = call->MTU;
4636 peer->cwind = call->nextCwind;
4637 peer->nDgramPackets = call->nDgramPackets;
4639 call->congestSeq = peer->congestSeq;
4641 /* Reset the resend times on the packets that were nacked
4642 * so we will retransmit as soon as the window permits
4645 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4647 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4648 tp->flags &= ~RX_PKTFLAG_SENT;
4650 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4655 /* If cwind is smaller than ssthresh, then increase
4656 * the window one packet for each ack we receive (exponential
4658 * If cwind is greater than or equal to ssthresh then increase
4659 * the congestion window by one packet for each cwind acks we
4660 * receive (linear growth). */
4661 if (call->cwind < call->ssthresh) {
4663 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4664 call->nCwindAcks = 0;
4666 call->nCwindAcks += newAckCount;
4667 if (call->nCwindAcks >= call->cwind) {
4668 call->nCwindAcks = 0;
4669 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4673 * If we have received several acknowledgements in a row then
4674 * it is time to increase the size of our datagrams
4676 if ((int)call->nAcks > rx_nDgramThreshold) {
4677 if (peer->maxDgramPackets > 1) {
4678 if (call->nDgramPackets < peer->maxDgramPackets) {
4679 call->nDgramPackets++;
4681 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4682 } else if (call->MTU < peer->maxMTU) {
4683 /* don't upgrade if we can't handle it */
4684 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4685 call->MTU = peer->ifMTU;
4687 call->MTU += peer->natMTU;
4688 call->MTU = MIN(call->MTU, peer->maxMTU);
4695 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4697 /* Servers need to hold the call until all response packets have
4698 * been acknowledged. Soft acks are good enough since clients
4699 * are not allowed to clear their receive queues. */
4700 if (call->state == RX_STATE_HOLD
4701 && call->tfirst + call->nSoftAcked >= call->tnext) {
4702 call->state = RX_STATE_DALLY;
4703 rxi_ClearTransmitQueue(call, 0);
4704 rxevent_Cancel(&call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4705 } else if (!queue_IsEmpty(&call->tq)) {
4706 rxi_Start(call, istack);
4711 /* Received a response to a challenge packet */
4713 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4714 struct rx_packet *np, int istack)
4718 /* Ignore the packet if we're the client */
4719 if (conn->type == RX_CLIENT_CONNECTION)
4722 /* If already authenticated, ignore the packet (it's probably a retry) */
4723 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4726 /* Otherwise, have the security object evaluate the response packet */
4727 error = RXS_CheckResponse(conn->securityObject, conn, np);
4729 /* If the response is invalid, reset the connection, sending
4730 * an abort to the peer */
4734 rxi_ConnectionError(conn, error);
4735 MUTEX_ENTER(&conn->conn_data_lock);
4736 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4737 MUTEX_EXIT(&conn->conn_data_lock);
4740 /* If the response is valid, any calls waiting to attach
4741 * servers can now do so */
4744 for (i = 0; i < RX_MAXCALLS; i++) {
4745 struct rx_call *call = conn->call[i];
4747 MUTEX_ENTER(&call->lock);
4748 if (call->state == RX_STATE_PRECALL)
4749 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4750 /* tnop can be null if newcallp is null */
4751 MUTEX_EXIT(&call->lock);
4755 /* Update the peer reachability information, just in case
4756 * some calls went into attach-wait while we were waiting
4757 * for authentication..
4759 rxi_UpdatePeerReach(conn, NULL);
4764 /* A client has received an authentication challenge: the security
4765 * object is asked to cough up a respectable response packet to send
4766 * back to the server. The server is responsible for retrying the
4767 * challenge if it fails to get a response. */
4770 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4771 struct rx_packet *np, int istack)
4775 /* Ignore the challenge if we're the server */
4776 if (conn->type == RX_SERVER_CONNECTION)
4779 /* Ignore the challenge if the connection is otherwise idle; someone's
4780 * trying to use us as an oracle. */
4781 if (!rxi_HasActiveCalls(conn))
4784 /* Send the security object the challenge packet. It is expected to fill
4785 * in the response. */
4786 error = RXS_GetResponse(conn->securityObject, conn, np);
4788 /* If the security object is unable to return a valid response, reset the
4789 * connection and send an abort to the peer. Otherwise send the response
4790 * packet to the peer connection. */
4792 rxi_ConnectionError(conn, error);
4793 MUTEX_ENTER(&conn->conn_data_lock);
4794 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4795 MUTEX_EXIT(&conn->conn_data_lock);
4797 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4798 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4804 /* Find an available server process to service the current request in
4805 * the given call structure. If one isn't available, queue up this
4806 * call so it eventually gets one */
4808 rxi_AttachServerProc(struct rx_call *call,
4809 osi_socket socket, int *tnop,
4810 struct rx_call **newcallp)
4812 struct rx_serverQueueEntry *sq;
4813 struct rx_service *service = call->conn->service;
4816 /* May already be attached */
4817 if (call->state == RX_STATE_ACTIVE)
4820 MUTEX_ENTER(&rx_serverPool_lock);
4822 haveQuota = QuotaOK(service);
4823 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4824 /* If there are no processes available to service this call,
4825 * put the call on the incoming call queue (unless it's
4826 * already on the queue).
4828 #ifdef RX_ENABLE_LOCKS
4830 ReturnToServerPool(service);
4831 #endif /* RX_ENABLE_LOCKS */
4833 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4834 call->flags |= RX_CALL_WAIT_PROC;
4835 rx_atomic_inc(&rx_nWaiting);
4836 rx_atomic_inc(&rx_nWaited);
4837 rxi_calltrace(RX_CALL_ARRIVAL, call);
4838 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4839 queue_Append(&rx_incomingCallQueue, call);
4842 sq = queue_Last(&rx_idleServerQueue, rx_serverQueueEntry);
4844 /* If hot threads are enabled, and both newcallp and sq->socketp
4845 * are non-null, then this thread will process the call, and the
4846 * idle server thread will start listening on this threads socket.
4849 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4852 *sq->socketp = socket;
4853 clock_GetTime(&call->startTime);
4854 MUTEX_ENTER(&rx_refcnt_mutex);
4855 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4856 MUTEX_EXIT(&rx_refcnt_mutex);
4860 if (call->flags & RX_CALL_WAIT_PROC) {
4861 /* Conservative: I don't think this should happen */
4862 call->flags &= ~RX_CALL_WAIT_PROC;
4863 if (queue_IsOnQueue(call)) {
4866 rx_atomic_dec(&rx_nWaiting);
4869 call->state = RX_STATE_ACTIVE;
4870 call->mode = RX_MODE_RECEIVING;
4871 #ifdef RX_KERNEL_TRACE
4873 int glockOwner = ISAFS_GLOCK();
4876 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4877 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4883 if (call->flags & RX_CALL_CLEARED) {
4884 /* send an ack now to start the packet flow up again */
4885 call->flags &= ~RX_CALL_CLEARED;
4886 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4888 #ifdef RX_ENABLE_LOCKS
4891 service->nRequestsRunning++;
4892 MUTEX_ENTER(&rx_quota_mutex);
4893 if (service->nRequestsRunning <= service->minProcs)
4896 MUTEX_EXIT(&rx_quota_mutex);
4900 MUTEX_EXIT(&rx_serverPool_lock);
4903 /* Delay the sending of an acknowledge event for a short while, while
4904 * a new call is being prepared (in the case of a client) or a reply
4905 * is being prepared (in the case of a server). Rather than sending
4906 * an ack packet, an ACKALL packet is sent. */
4908 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4910 #ifdef RX_ENABLE_LOCKS
4912 MUTEX_ENTER(&call->lock);
4913 rxevent_Put(call->delayedAckEvent);
4914 call->delayedAckEvent = NULL;
4915 MUTEX_ENTER(&rx_refcnt_mutex);
4916 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4917 MUTEX_EXIT(&rx_refcnt_mutex);
4919 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4920 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4921 call->flags |= RX_CALL_ACKALL_SENT;
4923 MUTEX_EXIT(&call->lock);
4924 #else /* RX_ENABLE_LOCKS */
4926 rxevent_Put(call->delayedAckEvent);
4927 call->delayedAckEvent = NULL;
4929 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4930 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4931 call->flags |= RX_CALL_ACKALL_SENT;
4932 #endif /* RX_ENABLE_LOCKS */
4936 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4939 struct rx_call *call = arg1;
4940 #ifdef RX_ENABLE_LOCKS
4942 MUTEX_ENTER(&call->lock);
4943 if (event == call->delayedAckEvent) {
4944 rxevent_Put(call->delayedAckEvent);
4945 call->delayedAckEvent = NULL;
4947 MUTEX_ENTER(&rx_refcnt_mutex);
4948 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4949 MUTEX_EXIT(&rx_refcnt_mutex);
4951 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4953 MUTEX_EXIT(&call->lock);
4954 #else /* RX_ENABLE_LOCKS */
4956 rxevent_Put(call->delayedAckEvent);
4957 call->delayedAckEvent = NULL;
4959 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4960 #endif /* RX_ENABLE_LOCKS */
4964 #ifdef RX_ENABLE_LOCKS
4965 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4966 * clearing them out.
4969 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4971 struct rx_packet *p, *tp;
4974 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4975 p->flags |= RX_PKTFLAG_ACKED;
4979 call->flags |= RX_CALL_TQ_CLEARME;
4980 call->flags |= RX_CALL_TQ_SOME_ACKED;
4983 rxi_rto_cancel(call);
4985 call->tfirst = call->tnext;
4986 call->nSoftAcked = 0;
4988 if (call->flags & RX_CALL_FAST_RECOVER) {
4989 call->flags &= ~RX_CALL_FAST_RECOVER;
4990 call->cwind = call->nextCwind;
4991 call->nextCwind = 0;
4994 CV_SIGNAL(&call->cv_twind);
4996 #endif /* RX_ENABLE_LOCKS */
4998 /* Clear out the transmit queue for the current call (all packets have
4999 * been received by peer) */
5001 rxi_ClearTransmitQueue(struct rx_call *call, int force)
5003 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5004 struct rx_packet *p, *tp;
5006 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
5008 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
5009 p->flags |= RX_PKTFLAG_ACKED;
5013 call->flags |= RX_CALL_TQ_CLEARME;
5014 call->flags |= RX_CALL_TQ_SOME_ACKED;
5017 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5018 #ifdef RXDEBUG_PACKET
5020 #endif /* RXDEBUG_PACKET */
5021 rxi_FreePackets(0, &call->tq);
5022 rxi_WakeUpTransmitQueue(call);
5023 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5024 call->flags &= ~RX_CALL_TQ_CLEARME;
5026 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5028 rxi_rto_cancel(call);
5029 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5030 call->nSoftAcked = 0;
5032 if (call->flags & RX_CALL_FAST_RECOVER) {
5033 call->flags &= ~RX_CALL_FAST_RECOVER;
5034 call->cwind = call->nextCwind;
5036 #ifdef RX_ENABLE_LOCKS
5037 CV_SIGNAL(&call->cv_twind);
5039 osi_rxWakeup(&call->twind);
5044 rxi_ClearReceiveQueue(struct rx_call *call)
5046 if (queue_IsNotEmpty(&call->rq)) {
5049 count = rxi_FreePackets(0, &call->rq);
5050 rx_packetReclaims += count;
5051 #ifdef RXDEBUG_PACKET
5053 if ( call->rqc != 0 )
5054 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5056 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5058 if (call->state == RX_STATE_PRECALL) {
5059 call->flags |= RX_CALL_CLEARED;
5063 /* Send an abort packet for the specified call */
5065 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5066 int istack, int force)
5069 struct clock when, now;
5074 /* Clients should never delay abort messages */
5075 if (rx_IsClientConn(call->conn))
5078 if (call->abortCode != call->error) {
5079 call->abortCode = call->error;
5080 call->abortCount = 0;
5083 if (force || rxi_callAbortThreshhold == 0
5084 || call->abortCount < rxi_callAbortThreshhold) {
5085 if (call->delayedAbortEvent) {
5086 rxevent_Cancel(&call->delayedAbortEvent, call,
5087 RX_CALL_REFCOUNT_ABORT);
5089 error = htonl(call->error);
5092 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5093 (char *)&error, sizeof(error), istack);
5094 } else if (!call->delayedAbortEvent) {
5095 clock_GetTime(&now);
5097 clock_Addmsec(&when, rxi_callAbortDelay);
5098 MUTEX_ENTER(&rx_refcnt_mutex);
5099 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5100 MUTEX_EXIT(&rx_refcnt_mutex);
5101 call->delayedAbortEvent =
5102 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5107 /* Send an abort packet for the specified connection. Packet is an
5108 * optional pointer to a packet that can be used to send the abort.
5109 * Once the number of abort messages reaches the threshhold, an
5110 * event is scheduled to send the abort. Setting the force flag
5111 * overrides sending delayed abort messages.
5113 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5114 * to send the abort packet.
5117 rxi_SendConnectionAbort(struct rx_connection *conn,
5118 struct rx_packet *packet, int istack, int force)
5121 struct clock when, now;
5126 /* Clients should never delay abort messages */
5127 if (rx_IsClientConn(conn))
5130 if (force || rxi_connAbortThreshhold == 0
5131 || conn->abortCount < rxi_connAbortThreshhold) {
5133 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
5134 error = htonl(conn->error);
5136 MUTEX_EXIT(&conn->conn_data_lock);
5138 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5139 RX_PACKET_TYPE_ABORT, (char *)&error,
5140 sizeof(error), istack);
5141 MUTEX_ENTER(&conn->conn_data_lock);
5142 } else if (!conn->delayedAbortEvent) {
5143 clock_GetTime(&now);
5145 clock_Addmsec(&when, rxi_connAbortDelay);
5146 conn->delayedAbortEvent =
5147 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
5152 /* Associate an error all of the calls owned by a connection. Called
5153 * with error non-zero. This is only for really fatal things, like
5154 * bad authentication responses. The connection itself is set in
5155 * error at this point, so that future packets received will be
5158 rxi_ConnectionError(struct rx_connection *conn,
5164 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5166 MUTEX_ENTER(&conn->conn_data_lock);
5167 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
5168 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
5169 if (conn->checkReachEvent) {
5170 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
5171 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5172 MUTEX_ENTER(&rx_refcnt_mutex);
5174 MUTEX_EXIT(&rx_refcnt_mutex);
5176 MUTEX_EXIT(&conn->conn_data_lock);
5177 for (i = 0; i < RX_MAXCALLS; i++) {
5178 struct rx_call *call = conn->call[i];
5180 MUTEX_ENTER(&call->lock);
5181 rxi_CallError(call, error);
5182 MUTEX_EXIT(&call->lock);
5185 conn->error = error;
5186 if (rx_stats_active)
5187 rx_atomic_inc(&rx_stats.fatalErrors);
5192 * Interrupt an in-progress call with the specified error and wakeup waiters.
5194 * @param[in] call The call to interrupt
5195 * @param[in] error The error code to send to the peer
5198 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5200 MUTEX_ENTER(&call->lock);
5201 rxi_CallError(call, error);
5202 rxi_SendCallAbort(call, NULL, 0, 1);
5203 MUTEX_EXIT(&call->lock);
5207 rxi_CallError(struct rx_call *call, afs_int32 error)
5210 osirx_AssertMine(&call->lock, "rxi_CallError");
5212 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5214 error = call->error;
5216 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5217 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5218 rxi_ResetCall(call, 0);
5221 rxi_ResetCall(call, 0);
5223 call->error = error;
5226 /* Reset various fields in a call structure, and wakeup waiting
5227 * processes. Some fields aren't changed: state & mode are not
5228 * touched (these must be set by the caller), and bufptr, nLeft, and
5229 * nFree are not reset, since these fields are manipulated by
5230 * unprotected macros, and may only be reset by non-interrupting code.
5234 rxi_ResetCall(struct rx_call *call, int newcall)
5237 struct rx_peer *peer;
5238 struct rx_packet *packet;
5240 osirx_AssertMine(&call->lock, "rxi_ResetCall");
5242 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5244 /* Notify anyone who is waiting for asynchronous packet arrival */
5245 if (call->arrivalProc) {
5246 (*call->arrivalProc) (call, call->arrivalProcHandle,
5247 call->arrivalProcArg);
5248 call->arrivalProc = (void (*)())0;
5252 rxevent_Cancel(&call->growMTUEvent, call, RX_CALL_REFCOUNT_ALIVE);
5254 if (call->delayedAbortEvent) {
5255 rxevent_Cancel(&call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
5256 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5258 rxi_SendCallAbort(call, packet, 0, 1);
5259 rxi_FreePacket(packet);
5264 * Update the peer with the congestion information in this call
5265 * so other calls on this connection can pick up where this call
5266 * left off. If the congestion sequence numbers don't match then
5267 * another call experienced a retransmission.
5269 peer = call->conn->peer;
5270 MUTEX_ENTER(&peer->peer_lock);
5272 if (call->congestSeq == peer->congestSeq) {
5273 peer->cwind = MAX(peer->cwind, call->cwind);
5274 peer->MTU = MAX(peer->MTU, call->MTU);
5275 peer->nDgramPackets =
5276 MAX(peer->nDgramPackets, call->nDgramPackets);
5279 call->abortCode = 0;
5280 call->abortCount = 0;
5282 if (peer->maxDgramPackets > 1) {
5283 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5285 call->MTU = peer->MTU;
5287 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5288 call->ssthresh = rx_maxSendWindow;
5289 call->nDgramPackets = peer->nDgramPackets;
5290 call->congestSeq = peer->congestSeq;
5291 call->rtt = peer->rtt;
5292 call->rtt_dev = peer->rtt_dev;
5293 clock_Zero(&call->rto);
5294 clock_Addmsec(&call->rto,
5295 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5296 MUTEX_EXIT(&peer->peer_lock);
5298 flags = call->flags;
5299 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5300 rxi_WaitforTQBusy(call);
5301 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5303 rxi_ClearTransmitQueue(call, 1);
5304 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5305 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5309 if ((flags & RX_CALL_PEER_BUSY)) {
5310 /* The call channel is still busy; resetting the call doesn't change
5312 call->flags |= RX_CALL_PEER_BUSY;
5315 rxi_ClearReceiveQueue(call);
5316 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5320 call->twind = call->conn->twind[call->channel];
5321 call->rwind = call->conn->rwind[call->channel];
5322 call->nSoftAcked = 0;
5323 call->nextCwind = 0;
5326 call->nCwindAcks = 0;
5327 call->nSoftAcks = 0;
5328 call->nHardAcks = 0;
5330 call->tfirst = call->rnext = call->tnext = 1;
5333 call->lastAcked = 0;
5334 call->localStatus = call->remoteStatus = 0;
5336 if (flags & RX_CALL_READER_WAIT) {
5337 #ifdef RX_ENABLE_LOCKS
5338 CV_BROADCAST(&call->cv_rq);
5340 osi_rxWakeup(&call->rq);
5343 if (flags & RX_CALL_WAIT_PACKETS) {
5344 MUTEX_ENTER(&rx_freePktQ_lock);
5345 rxi_PacketsUnWait(); /* XXX */
5346 MUTEX_EXIT(&rx_freePktQ_lock);
5348 #ifdef RX_ENABLE_LOCKS
5349 CV_SIGNAL(&call->cv_twind);
5351 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5352 osi_rxWakeup(&call->twind);
5355 #ifdef RX_ENABLE_LOCKS
5356 /* The following ensures that we don't mess with any queue while some
5357 * other thread might also be doing so. The call_queue_lock field is
5358 * is only modified under the call lock. If the call is in the process
5359 * of being removed from a queue, the call is not locked until the
5360 * the queue lock is dropped and only then is the call_queue_lock field
5361 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5362 * Note that any other routine which removes a call from a queue has to
5363 * obtain the queue lock before examing the queue and removing the call.
5365 if (call->call_queue_lock) {
5366 MUTEX_ENTER(call->call_queue_lock);
5367 if (queue_IsOnQueue(call)) {
5369 if (flags & RX_CALL_WAIT_PROC) {
5370 rx_atomic_dec(&rx_nWaiting);
5373 MUTEX_EXIT(call->call_queue_lock);
5374 CLEAR_CALL_QUEUE_LOCK(call);
5376 #else /* RX_ENABLE_LOCKS */
5377 if (queue_IsOnQueue(call)) {
5379 if (flags & RX_CALL_WAIT_PROC)
5380 rx_atomic_dec(&rx_nWaiting);
5382 #endif /* RX_ENABLE_LOCKS */
5384 rxi_KeepAliveOff(call);
5385 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5388 /* Send an acknowledge for the indicated packet (seq,serial) of the
5389 * indicated call, for the indicated reason (reason). This
5390 * acknowledge will specifically acknowledge receiving the packet, and
5391 * will also specify which other packets for this call have been
5392 * received. This routine returns the packet that was used to the
5393 * caller. The caller is responsible for freeing it or re-using it.
5394 * This acknowledgement also returns the highest sequence number
5395 * actually read out by the higher level to the sender; the sender
5396 * promises to keep around packets that have not been read by the
5397 * higher level yet (unless, of course, the sender decides to abort
5398 * the call altogether). Any of p, seq, serial, pflags, or reason may
5399 * be set to zero without ill effect. That is, if they are zero, they
5400 * will not convey any information.
5401 * NOW there is a trailer field, after the ack where it will safely be
5402 * ignored by mundanes, which indicates the maximum size packet this
5403 * host can swallow. */
5405 struct rx_packet *optionalPacket; use to send ack (or null)
5406 int seq; Sequence number of the packet we are acking
5407 int serial; Serial number of the packet
5408 int pflags; Flags field from packet header
5409 int reason; Reason an acknowledge was prompted
5413 rxi_SendAck(struct rx_call *call,
5414 struct rx_packet *optionalPacket, int serial, int reason,
5417 struct rx_ackPacket *ap;
5418 struct rx_packet *rqp;
5419 struct rx_packet *nxp; /* For queue_Scan */
5420 struct rx_packet *p;
5423 afs_uint32 padbytes = 0;
5424 #ifdef RX_ENABLE_TSFPQ
5425 struct rx_ts_info_t * rx_ts_info;
5429 * Open the receive window once a thread starts reading packets
5431 if (call->rnext > 1) {
5432 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5435 /* Don't attempt to grow MTU if this is a critical ping */
5436 if (reason == RX_ACK_MTU) {
5437 /* keep track of per-call attempts, if we're over max, do in small
5438 * otherwise in larger? set a size to increment by, decrease
5441 if (call->conn->peer->maxPacketSize &&
5442 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5444 padbytes = call->conn->peer->maxPacketSize+16;
5446 padbytes = call->conn->peer->maxMTU + 128;
5448 /* do always try a minimum size ping */
5449 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5451 /* subtract the ack payload */
5452 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5453 reason = RX_ACK_PING;
5456 call->nHardAcks = 0;
5457 call->nSoftAcks = 0;
5458 if (call->rnext > call->lastAcked)
5459 call->lastAcked = call->rnext;
5463 rx_computelen(p, p->length); /* reset length, you never know */
5464 } /* where that's been... */
5465 #ifdef RX_ENABLE_TSFPQ
5467 RX_TS_INFO_GET(rx_ts_info);
5468 if ((p = rx_ts_info->local_special_packet)) {
5469 rx_computelen(p, p->length);
5470 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5471 rx_ts_info->local_special_packet = p;
5472 } else { /* We won't send the ack, but don't panic. */
5473 return optionalPacket;
5477 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5478 /* We won't send the ack, but don't panic. */
5479 return optionalPacket;
5484 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5487 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5488 #ifndef RX_ENABLE_TSFPQ
5489 if (!optionalPacket)
5492 return optionalPacket;
5494 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5495 if (rx_Contiguous(p) < templ) {
5496 #ifndef RX_ENABLE_TSFPQ
5497 if (!optionalPacket)
5500 return optionalPacket;
5505 /* MTUXXX failing to send an ack is very serious. We should */
5506 /* try as hard as possible to send even a partial ack; it's */
5507 /* better than nothing. */
5508 ap = (struct rx_ackPacket *)rx_DataOf(p);
5509 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5510 ap->reason = reason;
5512 /* The skew computation used to be bogus, I think it's better now. */
5513 /* We should start paying attention to skew. XXX */
5514 ap->serial = htonl(serial);
5515 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5518 * First packet not yet forwarded to reader. When ACKALL has been
5519 * sent the peer has been told that all received packets will be
5520 * delivered to the reader. The value 'rnext' is used internally
5521 * to refer to the next packet in the receive queue that must be
5522 * delivered to the reader. From the perspective of the peer it
5523 * already has so report the last sequence number plus one if there
5524 * are packets in the receive queue awaiting processing.
5526 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5527 !queue_IsEmpty(&call->rq)) {
5528 ap->firstPacket = htonl(queue_Last(&call->rq, rx_packet)->header.seq + 1);
5530 ap->firstPacket = htonl(call->rnext);
5532 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5534 /* No fear of running out of ack packet here because there can only be at most
5535 * one window full of unacknowledged packets. The window size must be constrained
5536 * to be less than the maximum ack size, of course. Also, an ack should always
5537 * fit into a single packet -- it should not ever be fragmented. */
5538 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5539 if (!rqp || !call->rq.next
5540 || (rqp->header.seq > (call->rnext + call->rwind))) {
5541 #ifndef RX_ENABLE_TSFPQ
5542 if (!optionalPacket)
5545 rxi_CallError(call, RX_CALL_DEAD);
5546 return optionalPacket;
5549 while (rqp->header.seq > call->rnext + offset)
5550 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5551 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5553 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5554 #ifndef RX_ENABLE_TSFPQ
5555 if (!optionalPacket)
5558 rxi_CallError(call, RX_CALL_DEAD);
5559 return optionalPacket;
5565 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5567 /* these are new for AFS 3.3 */
5568 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5569 templ = htonl(templ);
5570 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5571 templ = htonl(call->conn->peer->ifMTU);
5572 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5573 sizeof(afs_int32), &templ);
5575 /* new for AFS 3.4 */
5576 templ = htonl(call->rwind);
5577 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5578 sizeof(afs_int32), &templ);
5580 /* new for AFS 3.5 */
5581 templ = htonl(call->conn->peer->ifDgramPackets);
5582 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5583 sizeof(afs_int32), &templ);
5585 p->header.serviceId = call->conn->serviceId;
5586 p->header.cid = (call->conn->cid | call->channel);
5587 p->header.callNumber = *call->callNumber;
5589 p->header.securityIndex = call->conn->securityIndex;
5590 p->header.epoch = call->conn->epoch;
5591 p->header.type = RX_PACKET_TYPE_ACK;
5592 p->header.flags = RX_SLOW_START_OK;
5593 if (reason == RX_ACK_PING) {
5594 p->header.flags |= RX_REQUEST_ACK;
5596 p->length = padbytes +
5597 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5600 /* not fast but we can potentially use this if truncated
5601 * fragments are delivered to figure out the mtu.
5603 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5604 sizeof(afs_int32), sizeof(afs_int32),
5608 if (call->conn->type == RX_CLIENT_CONNECTION)
5609 p->header.flags |= RX_CLIENT_INITIATED;
5613 if (rxdebug_active) {
5617 len = _snprintf(msg, sizeof(msg),
5618 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5619 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5620 ntohl(ap->serial), ntohl(ap->previousPacket),
5621 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5622 ap->nAcks, ntohs(ap->bufferSpace) );
5626 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5627 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5631 OutputDebugString(msg);
5633 #else /* AFS_NT40_ENV */
5635 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5636 ap->reason, ntohl(ap->previousPacket),
5637 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5639 for (offset = 0; offset < ap->nAcks; offset++)
5640 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5645 #endif /* AFS_NT40_ENV */
5648 int i, nbytes = p->length;
5650 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5651 if (nbytes <= p->wirevec[i].iov_len) {
5654 savelen = p->wirevec[i].iov_len;
5656 p->wirevec[i].iov_len = nbytes;
5658 rxi_Send(call, p, istack);
5659 p->wirevec[i].iov_len = savelen;
5663 nbytes -= p->wirevec[i].iov_len;
5666 if (rx_stats_active)
5667 rx_atomic_inc(&rx_stats.ackPacketsSent);
5668 #ifndef RX_ENABLE_TSFPQ
5669 if (!optionalPacket)
5672 return optionalPacket; /* Return packet for re-use by caller */
5676 struct rx_packet **list;
5681 /* Send all of the packets in the list in single datagram */
5683 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5684 int istack, int moreFlag)
5690 struct rx_connection *conn = call->conn;
5691 struct rx_peer *peer = conn->peer;
5693 MUTEX_ENTER(&peer->peer_lock);
5694 peer->nSent += xmit->len;
5695 if (xmit->resending)
5696 peer->reSends += xmit->len;
5697 MUTEX_EXIT(&peer->peer_lock);
5699 if (rx_stats_active) {
5700 if (xmit->resending)
5701 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5703 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5706 clock_GetTime(&now);
5708 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5712 /* Set the packet flags and schedule the resend events */
5713 /* Only request an ack for the last packet in the list */
5714 for (i = 0; i < xmit->len; i++) {
5715 struct rx_packet *packet = xmit->list[i];
5717 /* Record the time sent */
5718 packet->timeSent = now;
5719 packet->flags |= RX_PKTFLAG_SENT;
5721 /* Ask for an ack on retransmitted packets, on every other packet
5722 * if the peer doesn't support slow start. Ask for an ack on every
5723 * packet until the congestion window reaches the ack rate. */
5724 if (packet->header.serial) {
5727 packet->firstSent = now;
5728 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5729 || (!(call->flags & RX_CALL_SLOW_START_OK)
5730 && (packet->header.seq & 1)))) {
5735 /* Tag this packet as not being the last in this group,
5736 * for the receiver's benefit */
5737 if (i < xmit->len - 1 || moreFlag) {
5738 packet->header.flags |= RX_MORE_PACKETS;
5743 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5746 /* Since we're about to send a data packet to the peer, it's
5747 * safe to nuke any scheduled end-of-packets ack */
5748 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5750 MUTEX_EXIT(&call->lock);
5751 MUTEX_ENTER(&rx_refcnt_mutex);
5752 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5753 MUTEX_EXIT(&rx_refcnt_mutex);
5754 if (xmit->len > 1) {
5755 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5757 rxi_SendPacket(call, conn, xmit->list[0], istack);
5759 MUTEX_ENTER(&call->lock);
5760 MUTEX_ENTER(&rx_refcnt_mutex);
5761 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5762 MUTEX_EXIT(&rx_refcnt_mutex);
5764 /* Tell the RTO calculation engine that we have sent a packet, and
5765 * if it was the last one */
5766 rxi_rto_packet_sent(call, lastPacket, istack);
5768 /* Update last send time for this call (for keep-alive
5769 * processing), and for the connection (so that we can discover
5770 * idle connections) */
5771 conn->lastSendTime = call->lastSendTime = clock_Sec();
5772 /* Let a set of retransmits trigger an idle timeout */
5773 if (!xmit->resending)
5774 call->lastSendData = call->lastSendTime;
5777 /* When sending packets we need to follow these rules:
5778 * 1. Never send more than maxDgramPackets in a jumbogram.
5779 * 2. Never send a packet with more than two iovecs in a jumbogram.
5780 * 3. Never send a retransmitted packet in a jumbogram.
5781 * 4. Never send more than cwind/4 packets in a jumbogram
5782 * We always keep the last list we should have sent so we
5783 * can set the RX_MORE_PACKETS flags correctly.
5787 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5792 struct xmitlist working;
5793 struct xmitlist last;
5795 struct rx_peer *peer = call->conn->peer;
5796 int morePackets = 0;
5798 memset(&last, 0, sizeof(struct xmitlist));
5799 working.list = &list[0];
5801 working.resending = 0;
5803 recovery = call->flags & RX_CALL_FAST_RECOVER;
5805 for (i = 0; i < len; i++) {
5806 /* Does the current packet force us to flush the current list? */
5808 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5809 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5811 /* This sends the 'last' list and then rolls the current working
5812 * set into the 'last' one, and resets the working set */
5815 rxi_SendList(call, &last, istack, 1);
5816 /* If the call enters an error state stop sending, or if
5817 * we entered congestion recovery mode, stop sending */
5819 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5824 working.resending = 0;
5825 working.list = &list[i];
5827 /* Add the current packet to the list if it hasn't been acked.
5828 * Otherwise adjust the list pointer to skip the current packet. */
5829 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5832 if (list[i]->header.serial)
5833 working.resending = 1;
5835 /* Do we need to flush the list? */
5836 if (working.len >= (int)peer->maxDgramPackets
5837 || working.len >= (int)call->nDgramPackets
5838 || working.len >= (int)call->cwind
5839 || list[i]->header.serial
5840 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5842 rxi_SendList(call, &last, istack, 1);
5843 /* If the call enters an error state stop sending, or if
5844 * we entered congestion recovery mode, stop sending */
5846 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5851 working.resending = 0;
5852 working.list = &list[i + 1];
5855 if (working.len != 0) {
5856 osi_Panic("rxi_SendList error");
5858 working.list = &list[i + 1];
5862 /* Send the whole list when the call is in receive mode, when
5863 * the call is in eof mode, when we are in fast recovery mode,
5864 * and when we have the last packet */
5865 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5866 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5867 || (call->flags & RX_CALL_FAST_RECOVER)) {
5868 /* Check for the case where the current list contains
5869 * an acked packet. Since we always send retransmissions
5870 * in a separate packet, we only need to check the first
5871 * packet in the list */
5872 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5876 rxi_SendList(call, &last, istack, morePackets);
5877 /* If the call enters an error state stop sending, or if
5878 * we entered congestion recovery mode, stop sending */
5880 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5884 rxi_SendList(call, &working, istack, 0);
5886 } else if (last.len > 0) {
5887 rxi_SendList(call, &last, istack, 0);
5888 /* Packets which are in 'working' are not sent by this call */
5893 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
5895 struct rx_call *call = arg0;
5896 struct rx_peer *peer;
5897 struct rx_packet *p, *nxp;
5898 struct clock maxTimeout = { 60, 0 };
5900 MUTEX_ENTER(&call->lock);
5902 peer = call->conn->peer;
5904 /* Make sure that the event pointer is removed from the call
5905 * structure, since there is no longer a per-call retransmission
5907 if (event == call->resendEvent) {
5908 MUTEX_ENTER(&rx_refcnt_mutex);
5909 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5910 MUTEX_EXIT(&rx_refcnt_mutex);
5911 rxevent_Put(call->resendEvent);
5912 call->resendEvent = NULL;
5915 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
5916 rxi_CheckBusy(call);
5919 if (queue_IsEmpty(&call->tq)) {
5920 /* Nothing to do. This means that we've been raced, and that an
5921 * ACK has come in between when we were triggered, and when we
5922 * actually got to run. */
5926 /* We're in loss recovery */
5927 call->flags |= RX_CALL_FAST_RECOVER;
5929 /* Mark all of the pending packets in the queue as being lost */
5930 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5931 if (!(p->flags & RX_PKTFLAG_ACKED))
5932 p->flags &= ~RX_PKTFLAG_SENT;
5935 /* We're resending, so we double the timeout of the call. This will be
5936 * dropped back down by the first successful ACK that we receive.
5938 * We apply a maximum value here of 60 seconds
5940 clock_Add(&call->rto, &call->rto);
5941 if (clock_Gt(&call->rto, &maxTimeout))
5942 call->rto = maxTimeout;
5944 /* Packet loss is most likely due to congestion, so drop our window size
5945 * and start again from the beginning */
5946 if (peer->maxDgramPackets >1) {
5947 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5948 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5950 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5951 call->nDgramPackets = 1;
5953 call->nextCwind = 1;
5956 MUTEX_ENTER(&peer->peer_lock);
5957 peer->MTU = call->MTU;
5958 peer->cwind = call->cwind;
5959 peer->nDgramPackets = 1;
5961 call->congestSeq = peer->congestSeq;
5962 MUTEX_EXIT(&peer->peer_lock);
5964 rxi_Start(call, istack);
5967 MUTEX_EXIT(&call->lock);
5970 /* This routine is called when new packets are readied for
5971 * transmission and when retransmission may be necessary, or when the
5972 * transmission window or burst count are favourable. This should be
5973 * better optimized for new packets, the usual case, now that we've
5974 * got rid of queues of send packets. XXXXXXXXXXX */
5976 rxi_Start(struct rx_call *call, int istack)
5979 struct rx_packet *p;
5980 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5985 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5986 if (rx_stats_active)
5987 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5992 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5994 /* Send (or resend) any packets that need it, subject to
5995 * window restrictions and congestion burst control
5996 * restrictions. Ask for an ack on the last packet sent in
5997 * this burst. For now, we're relying upon the window being
5998 * considerably bigger than the largest number of packets that
5999 * are typically sent at once by one initial call to
6000 * rxi_Start. This is probably bogus (perhaps we should ask
6001 * for an ack when we're half way through the current
6002 * window?). Also, for non file transfer applications, this
6003 * may end up asking for an ack for every packet. Bogus. XXXX
6006 * But check whether we're here recursively, and let the other guy
6009 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6010 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6011 call->flags |= RX_CALL_TQ_BUSY;
6013 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6015 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6016 call->flags &= ~RX_CALL_NEED_START;
6017 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6019 maxXmitPackets = MIN(call->twind, call->cwind);
6020 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
6021 #ifdef RX_TRACK_PACKETS
6022 if ((p->flags & RX_PKTFLAG_FREE)
6023 || (!queue_IsEnd(&call->tq, nxp)
6024 && (nxp->flags & RX_PKTFLAG_FREE))
6025 || (p == (struct rx_packet *)&rx_freePacketQueue)
6026 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
6027 osi_Panic("rxi_Start: xmit queue clobbered");
6030 if (p->flags & RX_PKTFLAG_ACKED) {
6031 /* Since we may block, don't trust this */
6032 if (rx_stats_active)
6033 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6034 continue; /* Ignore this packet if it has been acknowledged */
6037 /* Turn off all flags except these ones, which are the same
6038 * on each transmission */
6039 p->header.flags &= RX_PRESET_FLAGS;
6041 if (p->header.seq >=
6042 call->tfirst + MIN((int)call->twind,
6043 (int)(call->nSoftAcked +
6045 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6046 /* Note: if we're waiting for more window space, we can
6047 * still send retransmits; hence we don't return here, but
6048 * break out to schedule a retransmit event */
6049 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6050 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6055 /* Transmit the packet if it needs to be sent. */
6056 if (!(p->flags & RX_PKTFLAG_SENT)) {
6057 if (nXmitPackets == maxXmitPackets) {
6058 rxi_SendXmitList(call, call->xmitList,
6059 nXmitPackets, istack);
6062 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6063 *(call->callNumber), p));
6064 call->xmitList[nXmitPackets++] = p;
6068 /* xmitList now hold pointers to all of the packets that are
6069 * ready to send. Now we loop to send the packets */
6070 if (nXmitPackets > 0) {
6071 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6075 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6077 /* We went into the error state while sending packets. Now is
6078 * the time to reset the call. This will also inform the using
6079 * process that the call is in an error state.
6081 if (rx_stats_active)
6082 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6083 call->flags &= ~RX_CALL_TQ_BUSY;
6084 rxi_WakeUpTransmitQueue(call);
6085 rxi_CallError(call, call->error);
6088 #ifdef RX_ENABLE_LOCKS
6089 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6091 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6092 /* Some packets have received acks. If they all have, we can clear
6093 * the transmit queue.
6096 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
6097 if (p->header.seq < call->tfirst
6098 && (p->flags & RX_PKTFLAG_ACKED)) {
6100 #ifdef RX_TRACK_PACKETS
6101 p->flags &= ~RX_PKTFLAG_TQ;
6103 #ifdef RXDEBUG_PACKET
6111 call->flags |= RX_CALL_TQ_CLEARME;
6113 #endif /* RX_ENABLE_LOCKS */
6114 if (call->flags & RX_CALL_TQ_CLEARME)
6115 rxi_ClearTransmitQueue(call, 1);
6116 } while (call->flags & RX_CALL_NEED_START);
6118 * TQ references no longer protected by this flag; they must remain
6119 * protected by the global lock.
6121 call->flags &= ~RX_CALL_TQ_BUSY;
6122 rxi_WakeUpTransmitQueue(call);
6124 call->flags |= RX_CALL_NEED_START;
6126 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6128 rxi_rto_cancel(call);
6132 /* Also adjusts the keep alive parameters for the call, to reflect
6133 * that we have just sent a packet (so keep alives aren't sent
6136 rxi_Send(struct rx_call *call, struct rx_packet *p,
6139 struct rx_connection *conn = call->conn;
6141 /* Stamp each packet with the user supplied status */
6142 p->header.userStatus = call->localStatus;
6144 /* Allow the security object controlling this call's security to
6145 * make any last-minute changes to the packet */
6146 RXS_SendPacket(conn->securityObject, call, p);
6148 /* Since we're about to send SOME sort of packet to the peer, it's
6149 * safe to nuke any scheduled end-of-packets ack */
6150 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
6152 /* Actually send the packet, filling in more connection-specific fields */
6153 MUTEX_EXIT(&call->lock);
6154 MUTEX_ENTER(&rx_refcnt_mutex);
6155 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6156 MUTEX_EXIT(&rx_refcnt_mutex);
6157 rxi_SendPacket(call, conn, p, istack);
6158 MUTEX_ENTER(&rx_refcnt_mutex);
6159 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6160 MUTEX_EXIT(&rx_refcnt_mutex);
6161 MUTEX_ENTER(&call->lock);
6163 /* Update last send time for this call (for keep-alive
6164 * processing), and for the connection (so that we can discover
6165 * idle connections) */
6166 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6167 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6168 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6170 conn->lastSendTime = call->lastSendTime = clock_Sec();
6171 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6172 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6173 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6174 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6175 RX_ACK_PING_RESPONSE)))
6176 call->lastSendData = call->lastSendTime;
6180 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6181 * that things are fine. Also called periodically to guarantee that nothing
6182 * falls through the cracks (e.g. (error + dally) connections have keepalive
6183 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6185 * haveCTLock Set if calling from rxi_ReapConnections
6187 #ifdef RX_ENABLE_LOCKS
6189 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6190 #else /* RX_ENABLE_LOCKS */
6192 rxi_CheckCall(struct rx_call *call)
6193 #endif /* RX_ENABLE_LOCKS */
6195 struct rx_connection *conn = call->conn;
6197 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6198 afs_uint32 fudgeFactor;
6202 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6203 if (call->flags & RX_CALL_TQ_BUSY) {
6204 /* Call is active and will be reset by rxi_Start if it's
6205 * in an error state.
6210 /* RTT + 8*MDEV, rounded up to the next second. */
6211 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6212 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6214 deadTime = conn->secondsUntilDead + fudgeFactor;
6216 /* These are computed to the second (+- 1 second). But that's
6217 * good enough for these values, which should be a significant
6218 * number of seconds. */
6219 if (now > (call->lastReceiveTime + deadTime)) {
6220 if (call->state == RX_STATE_ACTIVE) {
6222 #if defined(KERNEL) && defined(AFS_SUN5_ENV)
6224 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6225 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6226 ip_stack_t *ipst = ns->netstack_ip;
6228 ire = ire_cache_lookup(conn->peer->host
6229 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6231 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
6233 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6240 if (ire && ire->ire_max_frag > 0)
6241 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6243 #if defined(GLOBAL_NETSTACKID)
6247 #endif /* ADAPT_PMTU */
6248 cerror = RX_CALL_DEAD;
6251 #ifdef RX_ENABLE_LOCKS
6252 /* Cancel pending events */
6253 rxevent_Cancel(&call->delayedAckEvent, call,
6254 RX_CALL_REFCOUNT_DELAY);
6255 rxi_rto_cancel(call);
6256 rxevent_Cancel(&call->keepAliveEvent, call,
6257 RX_CALL_REFCOUNT_ALIVE);
6258 rxevent_Cancel(&call->growMTUEvent, call,
6259 RX_CALL_REFCOUNT_ALIVE);
6260 MUTEX_ENTER(&rx_refcnt_mutex);
6261 if (call->refCount == 0) {
6262 rxi_FreeCall(call, haveCTLock);
6263 MUTEX_EXIT(&rx_refcnt_mutex);
6266 MUTEX_EXIT(&rx_refcnt_mutex);
6268 #else /* RX_ENABLE_LOCKS */
6269 rxi_FreeCall(call, 0);
6271 #endif /* RX_ENABLE_LOCKS */
6273 /* Non-active calls are destroyed if they are not responding
6274 * to pings; active calls are simply flagged in error, so the
6275 * attached process can die reasonably gracefully. */
6278 if (conn->idleDeadTime) {
6279 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6282 /* see if we have a non-activity timeout */
6283 if (call->startWait && idleDeadTime
6284 && ((call->startWait + idleDeadTime) < now) &&
6285 (call->flags & RX_CALL_READER_WAIT)) {
6286 if (call->state == RX_STATE_ACTIVE) {
6287 cerror = RX_CALL_TIMEOUT;
6291 if (call->lastSendData && idleDeadTime && (conn->idleDeadErr != 0)
6292 && ((call->lastSendData + idleDeadTime) < now)) {
6293 if (call->state == RX_STATE_ACTIVE) {
6294 cerror = conn->idleDeadErr;
6299 if (conn->hardDeadTime) {
6300 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6303 /* see if we have a hard timeout */
6305 && (now > (hardDeadTime + call->startTime.sec))) {
6306 if (call->state == RX_STATE_ACTIVE)
6307 rxi_CallError(call, RX_CALL_TIMEOUT);
6312 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT
6313 && call->lastReceiveTime) {
6314 int oldMTU = conn->peer->ifMTU;
6316 /* if we thought we could send more, perhaps things got worse */
6317 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6318 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6319 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6320 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6322 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6324 /* minimum capped in SetPeerMtu */
6325 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6328 conn->lastPacketSize = 0;
6330 /* needed so ResetCall doesn't clobber us. */
6331 call->MTU = conn->peer->ifMTU;
6333 /* if we never succeeded, let the error pass out as-is */
6334 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6335 cerror = conn->msgsizeRetryErr;
6338 rxi_CallError(call, cerror);
6343 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6344 void *dummy, int dummy2)
6346 struct rx_connection *conn = arg1;
6347 struct rx_header theader;
6348 char tbuffer[1 + sizeof(struct rx_header)];
6349 struct sockaddr_in taddr;
6352 struct iovec tmpiov[2];
6355 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6358 tp = &tbuffer[sizeof(struct rx_header)];
6359 taddr.sin_family = AF_INET;
6360 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6361 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6362 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6363 taddr.sin_len = sizeof(struct sockaddr_in);
6365 memset(&theader, 0, sizeof(theader));
6366 theader.epoch = htonl(999);
6368 theader.callNumber = 0;
6371 theader.type = RX_PACKET_TYPE_VERSION;
6372 theader.flags = RX_LAST_PACKET;
6373 theader.serviceId = 0;
6375 memcpy(tbuffer, &theader, sizeof(theader));
6376 memcpy(tp, &a, sizeof(a));
6377 tmpiov[0].iov_base = tbuffer;
6378 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6380 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6382 MUTEX_ENTER(&conn->conn_data_lock);
6383 MUTEX_ENTER(&rx_refcnt_mutex);
6384 /* Only reschedule ourselves if the connection would not be destroyed */
6385 if (conn->refCount <= 1) {
6386 rxevent_Put(conn->natKeepAliveEvent);
6387 conn->natKeepAliveEvent = NULL;
6388 MUTEX_EXIT(&rx_refcnt_mutex);
6389 MUTEX_EXIT(&conn->conn_data_lock);
6390 rx_DestroyConnection(conn); /* drop the reference for this */
6392 conn->refCount--; /* drop the reference for this */
6393 MUTEX_EXIT(&rx_refcnt_mutex);
6394 rxevent_Put(conn->natKeepAliveEvent);
6395 conn->natKeepAliveEvent = NULL;
6396 rxi_ScheduleNatKeepAliveEvent(conn);
6397 MUTEX_EXIT(&conn->conn_data_lock);
6402 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6404 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6405 struct clock when, now;
6406 clock_GetTime(&now);
6408 when.sec += conn->secondsUntilNatPing;
6409 MUTEX_ENTER(&rx_refcnt_mutex);
6410 conn->refCount++; /* hold a reference for this */
6411 MUTEX_EXIT(&rx_refcnt_mutex);
6412 conn->natKeepAliveEvent =
6413 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6418 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6420 MUTEX_ENTER(&conn->conn_data_lock);
6421 conn->secondsUntilNatPing = seconds;
6423 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6424 rxi_ScheduleNatKeepAliveEvent(conn);
6426 conn->flags |= RX_CONN_NAT_PING;
6428 MUTEX_EXIT(&conn->conn_data_lock);
6432 rxi_NatKeepAliveOn(struct rx_connection *conn)
6434 MUTEX_ENTER(&conn->conn_data_lock);
6435 /* if it's already attached */
6436 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6437 rxi_ScheduleNatKeepAliveEvent(conn);
6439 conn->flags |= RX_CONN_NAT_PING;
6440 MUTEX_EXIT(&conn->conn_data_lock);
6443 /* When a call is in progress, this routine is called occasionally to
6444 * make sure that some traffic has arrived (or been sent to) the peer.
6445 * If nothing has arrived in a reasonable amount of time, the call is
6446 * declared dead; if nothing has been sent for a while, we send a
6447 * keep-alive packet (if we're actually trying to keep the call alive)
6450 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6453 struct rx_call *call = arg1;
6454 struct rx_connection *conn;
6457 MUTEX_ENTER(&rx_refcnt_mutex);
6458 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6459 MUTEX_EXIT(&rx_refcnt_mutex);
6460 MUTEX_ENTER(&call->lock);
6462 if (event == call->keepAliveEvent) {
6463 rxevent_Put(call->keepAliveEvent);
6464 call->keepAliveEvent = NULL;
6469 #ifdef RX_ENABLE_LOCKS
6470 if (rxi_CheckCall(call, 0)) {
6471 MUTEX_EXIT(&call->lock);
6474 #else /* RX_ENABLE_LOCKS */
6475 if (rxi_CheckCall(call))
6477 #endif /* RX_ENABLE_LOCKS */
6479 /* Don't try to keep alive dallying calls */
6480 if (call->state == RX_STATE_DALLY) {
6481 MUTEX_EXIT(&call->lock);
6486 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6487 /* Don't try to send keepalives if there is unacknowledged data */
6488 /* the rexmit code should be good enough, this little hack
6489 * doesn't quite work XXX */
6490 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6492 rxi_ScheduleKeepAliveEvent(call);
6493 MUTEX_EXIT(&call->lock);
6496 /* Does what's on the nameplate. */
6498 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6500 struct rx_call *call = arg1;
6501 struct rx_connection *conn;
6503 MUTEX_ENTER(&rx_refcnt_mutex);
6504 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6505 MUTEX_EXIT(&rx_refcnt_mutex);
6506 MUTEX_ENTER(&call->lock);
6508 if (event == call->growMTUEvent) {
6509 rxevent_Put(call->growMTUEvent);
6510 call->growMTUEvent = NULL;
6513 #ifdef RX_ENABLE_LOCKS
6514 if (rxi_CheckCall(call, 0)) {
6515 MUTEX_EXIT(&call->lock);
6518 #else /* RX_ENABLE_LOCKS */
6519 if (rxi_CheckCall(call))
6521 #endif /* RX_ENABLE_LOCKS */
6523 /* Don't bother with dallying calls */
6524 if (call->state == RX_STATE_DALLY) {
6525 MUTEX_EXIT(&call->lock);
6532 * keep being scheduled, just don't do anything if we're at peak,
6533 * or we're not set up to be properly handled (idle timeout required)
6535 if ((conn->peer->maxPacketSize != 0) &&
6536 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6537 (conn->idleDeadErr))
6538 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6539 rxi_ScheduleGrowMTUEvent(call, 0);
6540 MUTEX_EXIT(&call->lock);
6544 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6546 if (!call->keepAliveEvent) {
6547 struct clock when, now;
6548 clock_GetTime(&now);
6550 when.sec += call->conn->secondsUntilPing;
6551 MUTEX_ENTER(&rx_refcnt_mutex);
6552 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6553 MUTEX_EXIT(&rx_refcnt_mutex);
6554 call->keepAliveEvent =
6555 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6560 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6562 if (!call->growMTUEvent) {
6563 struct clock when, now;
6565 clock_GetTime(&now);
6568 if (call->conn->secondsUntilPing)
6569 secs = (6*call->conn->secondsUntilPing)-1;
6571 if (call->conn->secondsUntilDead)
6572 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6576 MUTEX_ENTER(&rx_refcnt_mutex);
6577 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6578 MUTEX_EXIT(&rx_refcnt_mutex);
6579 call->growMTUEvent =
6580 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6584 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6586 rxi_KeepAliveOn(struct rx_call *call)
6588 /* Pretend last packet received was received now--i.e. if another
6589 * packet isn't received within the keep alive time, then the call
6590 * will die; Initialize last send time to the current time--even
6591 * if a packet hasn't been sent yet. This will guarantee that a
6592 * keep-alive is sent within the ping time */
6593 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6594 rxi_ScheduleKeepAliveEvent(call);
6598 rxi_GrowMTUOn(struct rx_call *call)
6600 struct rx_connection *conn = call->conn;
6601 MUTEX_ENTER(&conn->conn_data_lock);
6602 conn->lastPingSizeSer = conn->lastPingSize = 0;
6603 MUTEX_EXIT(&conn->conn_data_lock);
6604 rxi_ScheduleGrowMTUEvent(call, 1);
6607 /* This routine is called to send connection abort messages
6608 * that have been delayed to throttle looping clients. */
6610 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6613 struct rx_connection *conn = arg1;
6616 struct rx_packet *packet;
6618 MUTEX_ENTER(&conn->conn_data_lock);
6619 rxevent_Put(conn->delayedAbortEvent);
6620 conn->delayedAbortEvent = NULL;
6621 error = htonl(conn->error);
6623 MUTEX_EXIT(&conn->conn_data_lock);
6624 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6627 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6628 RX_PACKET_TYPE_ABORT, (char *)&error,
6630 rxi_FreePacket(packet);
6634 /* This routine is called to send call abort messages
6635 * that have been delayed to throttle looping clients. */
6637 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6640 struct rx_call *call = arg1;
6643 struct rx_packet *packet;
6645 MUTEX_ENTER(&call->lock);
6646 rxevent_Put(call->delayedAbortEvent);
6647 call->delayedAbortEvent = NULL;
6648 error = htonl(call->error);
6650 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6653 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6654 (char *)&error, sizeof(error), 0);
6655 rxi_FreePacket(packet);
6657 MUTEX_EXIT(&call->lock);
6658 MUTEX_ENTER(&rx_refcnt_mutex);
6659 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6660 MUTEX_EXIT(&rx_refcnt_mutex);
6663 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6664 * seconds) to ask the client to authenticate itself. The routine
6665 * issues a challenge to the client, which is obtained from the
6666 * security object associated with the connection */
6668 rxi_ChallengeEvent(struct rxevent *event,
6669 void *arg0, void *arg1, int tries)
6671 struct rx_connection *conn = arg0;
6674 rxevent_Put(conn->challengeEvent);
6675 conn->challengeEvent = NULL;
6678 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6679 struct rx_packet *packet;
6680 struct clock when, now;
6683 /* We've failed to authenticate for too long.
6684 * Reset any calls waiting for authentication;
6685 * they are all in RX_STATE_PRECALL.
6689 MUTEX_ENTER(&conn->conn_call_lock);
6690 for (i = 0; i < RX_MAXCALLS; i++) {
6691 struct rx_call *call = conn->call[i];
6693 MUTEX_ENTER(&call->lock);
6694 if (call->state == RX_STATE_PRECALL) {
6695 rxi_CallError(call, RX_CALL_DEAD);
6696 rxi_SendCallAbort(call, NULL, 0, 0);
6698 MUTEX_EXIT(&call->lock);
6701 MUTEX_EXIT(&conn->conn_call_lock);
6705 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6707 /* If there's no packet available, do this later. */
6708 RXS_GetChallenge(conn->securityObject, conn, packet);
6709 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6710 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6711 rxi_FreePacket(packet);
6713 clock_GetTime(&now);
6715 when.sec += RX_CHALLENGE_TIMEOUT;
6716 conn->challengeEvent =
6717 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6722 /* Call this routine to start requesting the client to authenticate
6723 * itself. This will continue until authentication is established,
6724 * the call times out, or an invalid response is returned. The
6725 * security object associated with the connection is asked to create
6726 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6727 * defined earlier. */
6729 rxi_ChallengeOn(struct rx_connection *conn)
6731 if (!conn->challengeEvent) {
6732 RXS_CreateChallenge(conn->securityObject, conn);
6733 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6738 /* rxi_ComputeRoundTripTime is called with peer locked. */
6739 /* peer may be null */
6741 rxi_ComputeRoundTripTime(struct rx_packet *p,
6742 struct rx_ackPacket *ack,
6743 struct rx_call *call,
6744 struct rx_peer *peer,
6747 struct clock thisRtt, *sentp;
6751 /* If the ACK is delayed, then do nothing */
6752 if (ack->reason == RX_ACK_DELAY)
6755 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6756 * their RTT multiple times, so only include the RTT of the last packet
6758 if (p->flags & RX_JUMBO_PACKET)
6761 /* Use the serial number to determine which transmission the ACK is for,
6762 * and set the sent time to match this. If we have no serial number, then
6763 * only use the ACK for RTT calculations if the packet has not been
6767 serial = ntohl(ack->serial);
6769 if (serial == p->header.serial) {
6770 sentp = &p->timeSent;
6771 } else if (serial == p->firstSerial) {
6772 sentp = &p->firstSent;
6773 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6774 sentp = &p->firstSent;
6778 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6779 sentp = &p->firstSent;
6786 if (clock_Lt(&thisRtt, sentp))
6787 return; /* somebody set the clock back, don't count this time. */
6789 clock_Sub(&thisRtt, sentp);
6790 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6791 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6793 if (clock_IsZero(&thisRtt)) {
6795 * The actual round trip time is shorter than the
6796 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6797 * Since we can't tell which at the moment we will assume 1ms.
6799 thisRtt.usec = 1000;
6802 if (rx_stats_active) {
6803 MUTEX_ENTER(&rx_stats_mutex);
6804 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6805 rx_stats.minRtt = thisRtt;
6806 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6807 if (thisRtt.sec > 60) {
6808 MUTEX_EXIT(&rx_stats_mutex);
6809 return; /* somebody set the clock ahead */
6811 rx_stats.maxRtt = thisRtt;
6813 clock_Add(&rx_stats.totalRtt, &thisRtt);
6814 rx_atomic_inc(&rx_stats.nRttSamples);
6815 MUTEX_EXIT(&rx_stats_mutex);
6818 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6820 /* Apply VanJacobson round-trip estimations */
6825 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6826 * srtt is stored as fixed point with 3 bits after the binary
6827 * point (i.e., scaled by 8). The following magic is
6828 * equivalent to the smoothing algorithm in rfc793 with an
6829 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6830 * srtt'*8 = rtt + srtt*7
6831 * srtt'*8 = srtt*8 + rtt - srtt
6832 * srtt' = srtt + rtt/8 - srtt/8
6833 * srtt' = srtt + (rtt - srtt)/8
6836 delta = _8THMSEC(&thisRtt) - call->rtt;
6837 call->rtt += (delta >> 3);
6840 * We accumulate a smoothed rtt variance (actually, a smoothed
6841 * mean difference), then set the retransmit timer to smoothed
6842 * rtt + 4 times the smoothed variance (was 2x in van's original
6843 * paper, but 4x works better for me, and apparently for him as
6845 * rttvar is stored as
6846 * fixed point with 2 bits after the binary point (scaled by
6847 * 4). The following is equivalent to rfc793 smoothing with
6848 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6849 * rttvar'*4 = rttvar*3 + |delta|
6850 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6851 * rttvar' = rttvar + |delta|/4 - rttvar/4
6852 * rttvar' = rttvar + (|delta| - rttvar)/4
6853 * This replaces rfc793's wired-in beta.
6854 * dev*4 = dev*4 + (|actual - expected| - dev)
6860 delta -= (call->rtt_dev << 1);
6861 call->rtt_dev += (delta >> 3);
6863 /* I don't have a stored RTT so I start with this value. Since I'm
6864 * probably just starting a call, and will be pushing more data down
6865 * this, I expect congestion to increase rapidly. So I fudge a
6866 * little, and I set deviance to half the rtt. In practice,
6867 * deviance tends to approach something a little less than
6868 * half the smoothed rtt. */
6869 call->rtt = _8THMSEC(&thisRtt) + 8;
6870 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6872 /* the smoothed RTT time is RTT + 4*MDEV
6874 * We allow a user specified minimum to be set for this, to allow clamping
6875 * at a minimum value in the same way as TCP. In addition, we have to allow
6876 * for the possibility that this packet is answered by a delayed ACK, so we
6877 * add on a fixed 200ms to account for that timer expiring.
6880 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6881 rx_minPeerTimeout) + 200;
6882 clock_Zero(&call->rto);
6883 clock_Addmsec(&call->rto, rtt_timeout);
6885 /* Update the peer, so any new calls start with our values */
6886 peer->rtt_dev = call->rtt_dev;
6887 peer->rtt = call->rtt;
6889 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6890 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6894 /* Find all server connections that have not been active for a long time, and
6897 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
6900 struct clock now, when;
6901 clock_GetTime(&now);
6903 /* Find server connection structures that haven't been used for
6904 * greater than rx_idleConnectionTime */
6906 struct rx_connection **conn_ptr, **conn_end;
6907 int i, havecalls = 0;
6908 MUTEX_ENTER(&rx_connHashTable_lock);
6909 for (conn_ptr = &rx_connHashTable[0], conn_end =
6910 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6912 struct rx_connection *conn, *next;
6913 struct rx_call *call;
6917 for (conn = *conn_ptr; conn; conn = next) {
6918 /* XXX -- Shouldn't the connection be locked? */
6921 for (i = 0; i < RX_MAXCALLS; i++) {
6922 call = conn->call[i];
6926 code = MUTEX_TRYENTER(&call->lock);
6929 #ifdef RX_ENABLE_LOCKS
6930 result = rxi_CheckCall(call, 1);
6931 #else /* RX_ENABLE_LOCKS */
6932 result = rxi_CheckCall(call);
6933 #endif /* RX_ENABLE_LOCKS */
6934 MUTEX_EXIT(&call->lock);
6936 /* If CheckCall freed the call, it might
6937 * have destroyed the connection as well,
6938 * which screws up the linked lists.
6944 if (conn->type == RX_SERVER_CONNECTION) {
6945 /* This only actually destroys the connection if
6946 * there are no outstanding calls */
6947 MUTEX_ENTER(&conn->conn_data_lock);
6948 MUTEX_ENTER(&rx_refcnt_mutex);
6949 if (!havecalls && !conn->refCount
6950 && ((conn->lastSendTime + rx_idleConnectionTime) <
6952 conn->refCount++; /* it will be decr in rx_DestroyConn */
6953 MUTEX_EXIT(&rx_refcnt_mutex);
6954 MUTEX_EXIT(&conn->conn_data_lock);
6955 #ifdef RX_ENABLE_LOCKS
6956 rxi_DestroyConnectionNoLock(conn);
6957 #else /* RX_ENABLE_LOCKS */
6958 rxi_DestroyConnection(conn);
6959 #endif /* RX_ENABLE_LOCKS */
6961 #ifdef RX_ENABLE_LOCKS
6963 MUTEX_EXIT(&rx_refcnt_mutex);
6964 MUTEX_EXIT(&conn->conn_data_lock);
6966 #endif /* RX_ENABLE_LOCKS */
6970 #ifdef RX_ENABLE_LOCKS
6971 while (rx_connCleanup_list) {
6972 struct rx_connection *conn;
6973 conn = rx_connCleanup_list;
6974 rx_connCleanup_list = rx_connCleanup_list->next;
6975 MUTEX_EXIT(&rx_connHashTable_lock);
6976 rxi_CleanupConnection(conn);
6977 MUTEX_ENTER(&rx_connHashTable_lock);
6979 MUTEX_EXIT(&rx_connHashTable_lock);
6980 #endif /* RX_ENABLE_LOCKS */
6983 /* Find any peer structures that haven't been used (haven't had an
6984 * associated connection) for greater than rx_idlePeerTime */
6986 struct rx_peer **peer_ptr, **peer_end;
6990 * Why do we need to hold the rx_peerHashTable_lock across
6991 * the incrementing of peer_ptr since the rx_peerHashTable
6992 * array is not changing? We don't.
6994 * By dropping the lock periodically we can permit other
6995 * activities to be performed while a rxi_ReapConnections
6996 * call is in progress. The goal of reap connections
6997 * is to clean up quickly without causing large amounts
6998 * of contention. Therefore, it is important that global
6999 * mutexes not be held for extended periods of time.
7001 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7002 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7004 struct rx_peer *peer, *next, *prev;
7006 MUTEX_ENTER(&rx_peerHashTable_lock);
7007 for (prev = peer = *peer_ptr; peer; peer = next) {
7009 code = MUTEX_TRYENTER(&peer->peer_lock);
7010 if ((code) && (peer->refCount == 0)
7011 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7012 rx_interface_stat_p rpc_stat, nrpc_stat;
7016 * now know that this peer object is one to be
7017 * removed from the hash table. Once it is removed
7018 * it can't be referenced by other threads.
7019 * Lets remove it first and decrement the struct
7020 * nPeerStructs count.
7022 if (peer == *peer_ptr) {
7028 if (rx_stats_active)
7029 rx_atomic_dec(&rx_stats.nPeerStructs);
7032 * Now if we hold references on 'prev' and 'next'
7033 * we can safely drop the rx_peerHashTable_lock
7034 * while we destroy this 'peer' object.
7040 MUTEX_EXIT(&rx_peerHashTable_lock);
7042 MUTEX_EXIT(&peer->peer_lock);
7043 MUTEX_DESTROY(&peer->peer_lock);
7045 (&peer->rpcStats, rpc_stat, nrpc_stat,
7046 rx_interface_stat)) {
7047 unsigned int num_funcs;
7050 queue_Remove(&rpc_stat->queue_header);
7051 queue_Remove(&rpc_stat->all_peers);
7052 num_funcs = rpc_stat->stats[0].func_total;
7054 sizeof(rx_interface_stat_t) +
7055 rpc_stat->stats[0].func_total *
7056 sizeof(rx_function_entry_v1_t);
7058 rxi_Free(rpc_stat, space);
7060 MUTEX_ENTER(&rx_rpc_stats);
7061 rxi_rpc_peer_stat_cnt -= num_funcs;
7062 MUTEX_EXIT(&rx_rpc_stats);
7067 * Regain the rx_peerHashTable_lock and
7068 * decrement the reference count on 'prev'
7071 MUTEX_ENTER(&rx_peerHashTable_lock);
7078 MUTEX_EXIT(&peer->peer_lock);
7083 MUTEX_EXIT(&rx_peerHashTable_lock);
7087 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7088 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7089 * GC, just below. Really, we shouldn't have to keep moving packets from
7090 * one place to another, but instead ought to always know if we can
7091 * afford to hold onto a packet in its particular use. */
7092 MUTEX_ENTER(&rx_freePktQ_lock);
7093 if (rx_waitingForPackets) {
7094 rx_waitingForPackets = 0;
7095 #ifdef RX_ENABLE_LOCKS
7096 CV_BROADCAST(&rx_waitingForPackets_cv);
7098 osi_rxWakeup(&rx_waitingForPackets);
7101 MUTEX_EXIT(&rx_freePktQ_lock);
7104 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7105 rxevent_Put(rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0));
7109 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7110 * rx.h is sort of strange this is better. This is called with a security
7111 * object before it is discarded. Each connection using a security object has
7112 * its own refcount to the object so it won't actually be freed until the last
7113 * connection is destroyed.
7115 * This is the only rxs module call. A hold could also be written but no one
7119 rxs_Release(struct rx_securityClass *aobj)
7121 return RXS_Close(aobj);
7129 #define TRACE_OPTION_RX_DEBUG 16
7137 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7138 0, KEY_QUERY_VALUE, &parmKey);
7139 if (code != ERROR_SUCCESS)
7142 dummyLen = sizeof(TraceOption);
7143 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7144 (BYTE *) &TraceOption, &dummyLen);
7145 if (code == ERROR_SUCCESS) {
7146 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7148 RegCloseKey (parmKey);
7149 #endif /* AFS_NT40_ENV */
7154 rx_DebugOnOff(int on)
7158 rxdebug_active = on;
7164 rx_StatsOnOff(int on)
7166 rx_stats_active = on;
7170 /* Don't call this debugging routine directly; use dpf */
7172 rxi_DebugPrint(char *format, ...)
7181 va_start(ap, format);
7183 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7186 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7188 OutputDebugString(msg);
7194 va_start(ap, format);
7196 clock_GetTime(&now);
7197 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7198 (unsigned int)now.usec);
7199 vfprintf(rx_Log, format, ap);
7207 * This function is used to process the rx_stats structure that is local
7208 * to a process as well as an rx_stats structure received from a remote
7209 * process (via rxdebug). Therefore, it needs to do minimal version
7213 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7214 afs_int32 freePackets, char version)
7218 if (size != sizeof(struct rx_statistics)) {
7220 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7221 size, sizeof(struct rx_statistics));
7224 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7227 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7228 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7229 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7230 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7231 s->specialPktAllocFailures);
7233 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7234 s->receivePktAllocFailures, s->sendPktAllocFailures,
7235 s->specialPktAllocFailures);
7239 " greedy %u, " "bogusReads %u (last from host %x), "
7240 "noPackets %u, " "noBuffers %u, " "selects %u, "
7241 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7242 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7243 s->selects, s->sendSelects);
7245 fprintf(file, " packets read: ");
7246 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7247 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7249 fprintf(file, "\n");
7252 " other read counters: data %u, " "ack %u, " "dup %u "
7253 "spurious %u " "dally %u\n", s->dataPacketsRead,
7254 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7255 s->ignorePacketDally);
7257 fprintf(file, " packets sent: ");
7258 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7259 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7261 fprintf(file, "\n");
7264 " other send counters: ack %u, " "data %u (not resends), "
7265 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7266 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7267 s->dataPacketsPushed, s->ignoreAckedPacket);
7270 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7271 s->netSendFailures, (int)s->fatalErrors);
7273 if (s->nRttSamples) {
7274 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7275 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7277 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7278 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7282 " %d server connections, " "%d client connections, "
7283 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7284 s->nServerConns, s->nClientConns, s->nPeerStructs,
7285 s->nCallStructs, s->nFreeCallStructs);
7287 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7288 fprintf(file, " %d clock updates\n", clock_nUpdates);
7292 /* for backward compatibility */
7294 rx_PrintStats(FILE * file)
7296 MUTEX_ENTER(&rx_stats_mutex);
7297 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7298 sizeof(rx_stats), rx_nFreePackets,
7300 MUTEX_EXIT(&rx_stats_mutex);
7304 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7306 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
7307 ntohl(peer->host), (int)ntohs(peer->port), (int)peer->burstSize,
7308 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
7311 " Rtt %d, " "total sent %d, " "resent %d\n",
7312 peer->rtt, peer->nSent, peer->reSends);
7315 " Packet size %d, " "max in packet skew %d, "
7316 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
7317 (int)peer->outPacketSkew);
7321 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7323 * This mutex protects the following static variables:
7327 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7328 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7330 #define LOCK_RX_DEBUG
7331 #define UNLOCK_RX_DEBUG
7332 #endif /* AFS_PTHREAD_ENV */
7334 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7336 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7337 u_char type, void *inputData, size_t inputLength,
7338 void *outputData, size_t outputLength)
7340 static afs_int32 counter = 100;
7341 time_t waitTime, waitCount;
7342 struct rx_header theader;
7345 struct timeval tv_now, tv_wake, tv_delta;
7346 struct sockaddr_in taddr, faddr;
7360 tp = &tbuffer[sizeof(struct rx_header)];
7361 taddr.sin_family = AF_INET;
7362 taddr.sin_port = remotePort;
7363 taddr.sin_addr.s_addr = remoteAddr;
7364 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7365 taddr.sin_len = sizeof(struct sockaddr_in);
7368 memset(&theader, 0, sizeof(theader));
7369 theader.epoch = htonl(999);
7371 theader.callNumber = htonl(counter);
7374 theader.type = type;
7375 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7376 theader.serviceId = 0;
7378 memcpy(tbuffer, &theader, sizeof(theader));
7379 memcpy(tp, inputData, inputLength);
7381 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7382 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7384 /* see if there's a packet available */
7385 gettimeofday(&tv_wake, NULL);
7386 tv_wake.tv_sec += waitTime;
7389 FD_SET(socket, &imask);
7390 tv_delta.tv_sec = tv_wake.tv_sec;
7391 tv_delta.tv_usec = tv_wake.tv_usec;
7392 gettimeofday(&tv_now, NULL);
7394 if (tv_delta.tv_usec < tv_now.tv_usec) {
7396 tv_delta.tv_usec += 1000000;
7399 tv_delta.tv_usec -= tv_now.tv_usec;
7401 if (tv_delta.tv_sec < tv_now.tv_sec) {
7405 tv_delta.tv_sec -= tv_now.tv_sec;
7408 code = select(0, &imask, 0, 0, &tv_delta);
7409 #else /* AFS_NT40_ENV */
7410 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7411 #endif /* AFS_NT40_ENV */
7412 if (code == 1 && FD_ISSET(socket, &imask)) {
7413 /* now receive a packet */
7414 faddrLen = sizeof(struct sockaddr_in);
7416 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7417 (struct sockaddr *)&faddr, &faddrLen);
7420 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7421 if (counter == ntohl(theader.callNumber))
7429 /* see if we've timed out */
7437 code -= sizeof(struct rx_header);
7438 if (code > outputLength)
7439 code = outputLength;
7440 memcpy(outputData, tp, code);
7443 #endif /* RXDEBUG */
7446 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7447 afs_uint16 remotePort, struct rx_debugStats * stat,
7448 afs_uint32 * supportedValues)
7450 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7452 struct rx_debugIn in;
7454 *supportedValues = 0;
7455 in.type = htonl(RX_DEBUGI_GETSTATS);
7458 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7459 &in, sizeof(in), stat, sizeof(*stat));
7462 * If the call was successful, fixup the version and indicate
7463 * what contents of the stat structure are valid.
7464 * Also do net to host conversion of fields here.
7468 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7469 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7471 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7472 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7474 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7475 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7477 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7478 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7480 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7481 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7483 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7484 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7486 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7487 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7489 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7490 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7492 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7493 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7495 stat->nFreePackets = ntohl(stat->nFreePackets);
7496 stat->packetReclaims = ntohl(stat->packetReclaims);
7497 stat->callsExecuted = ntohl(stat->callsExecuted);
7498 stat->nWaiting = ntohl(stat->nWaiting);
7499 stat->idleThreads = ntohl(stat->idleThreads);
7500 stat->nWaited = ntohl(stat->nWaited);
7501 stat->nPackets = ntohl(stat->nPackets);
7510 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7511 afs_uint16 remotePort, struct rx_statistics * stat,
7512 afs_uint32 * supportedValues)
7514 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7516 struct rx_debugIn in;
7517 afs_int32 *lp = (afs_int32 *) stat;
7521 * supportedValues is currently unused, but added to allow future
7522 * versioning of this function.
7525 *supportedValues = 0;
7526 in.type = htonl(RX_DEBUGI_RXSTATS);
7528 memset(stat, 0, sizeof(*stat));
7530 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7531 &in, sizeof(in), stat, sizeof(*stat));
7536 * Do net to host conversion here
7539 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7550 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7551 afs_uint16 remotePort, size_t version_length,
7554 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7556 return MakeDebugCall(socket, remoteAddr, remotePort,
7557 RX_PACKET_TYPE_VERSION, a, 1, version,
7565 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7566 afs_uint16 remotePort, afs_int32 * nextConnection,
7567 int allConnections, afs_uint32 debugSupportedValues,
7568 struct rx_debugConn * conn,
7569 afs_uint32 * supportedValues)
7571 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7573 struct rx_debugIn in;
7577 * supportedValues is currently unused, but added to allow future
7578 * versioning of this function.
7581 *supportedValues = 0;
7582 if (allConnections) {
7583 in.type = htonl(RX_DEBUGI_GETALLCONN);
7585 in.type = htonl(RX_DEBUGI_GETCONN);
7587 in.index = htonl(*nextConnection);
7588 memset(conn, 0, sizeof(*conn));
7590 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7591 &in, sizeof(in), conn, sizeof(*conn));
7594 *nextConnection += 1;
7597 * Convert old connection format to new structure.
7600 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7601 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7602 #define MOVEvL(a) (conn->a = vL->a)
7604 /* any old or unrecognized version... */
7605 for (i = 0; i < RX_MAXCALLS; i++) {
7606 MOVEvL(callState[i]);
7607 MOVEvL(callMode[i]);
7608 MOVEvL(callFlags[i]);
7609 MOVEvL(callOther[i]);
7611 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7612 MOVEvL(secStats.type);
7613 MOVEvL(secStats.level);
7614 MOVEvL(secStats.flags);
7615 MOVEvL(secStats.expires);
7616 MOVEvL(secStats.packetsReceived);
7617 MOVEvL(secStats.packetsSent);
7618 MOVEvL(secStats.bytesReceived);
7619 MOVEvL(secStats.bytesSent);
7624 * Do net to host conversion here
7626 * I don't convert host or port since we are most likely
7627 * going to want these in NBO.
7629 conn->cid = ntohl(conn->cid);
7630 conn->serial = ntohl(conn->serial);
7631 for (i = 0; i < RX_MAXCALLS; i++) {
7632 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7634 conn->error = ntohl(conn->error);
7635 conn->secStats.flags = ntohl(conn->secStats.flags);
7636 conn->secStats.expires = ntohl(conn->secStats.expires);
7637 conn->secStats.packetsReceived =
7638 ntohl(conn->secStats.packetsReceived);
7639 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7640 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7641 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7642 conn->epoch = ntohl(conn->epoch);
7643 conn->natMTU = ntohl(conn->natMTU);
7652 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7653 afs_uint16 remotePort, afs_int32 * nextPeer,
7654 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7655 afs_uint32 * supportedValues)
7657 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7659 struct rx_debugIn in;
7662 * supportedValues is currently unused, but added to allow future
7663 * versioning of this function.
7666 *supportedValues = 0;
7667 in.type = htonl(RX_DEBUGI_GETPEER);
7668 in.index = htonl(*nextPeer);
7669 memset(peer, 0, sizeof(*peer));
7671 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7672 &in, sizeof(in), peer, sizeof(*peer));
7678 * Do net to host conversion here
7680 * I don't convert host or port since we are most likely
7681 * going to want these in NBO.
7683 peer->ifMTU = ntohs(peer->ifMTU);
7684 peer->idleWhen = ntohl(peer->idleWhen);
7685 peer->refCount = ntohs(peer->refCount);
7686 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7687 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7688 peer->rtt = ntohl(peer->rtt);
7689 peer->rtt_dev = ntohl(peer->rtt_dev);
7690 peer->timeout.sec = 0;
7691 peer->timeout.usec = 0;
7692 peer->nSent = ntohl(peer->nSent);
7693 peer->reSends = ntohl(peer->reSends);
7694 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7695 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7696 peer->natMTU = ntohs(peer->natMTU);
7697 peer->maxMTU = ntohs(peer->maxMTU);
7698 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7699 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7700 peer->MTU = ntohs(peer->MTU);
7701 peer->cwind = ntohs(peer->cwind);
7702 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7703 peer->congestSeq = ntohs(peer->congestSeq);
7704 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7705 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7706 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7707 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7716 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7717 struct rx_debugPeer * peerStats)
7720 afs_int32 error = 1; /* default to "did not succeed" */
7721 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7723 MUTEX_ENTER(&rx_peerHashTable_lock);
7724 for(tp = rx_peerHashTable[hashValue];
7725 tp != NULL; tp = tp->next) {
7726 if (tp->host == peerHost)
7732 MUTEX_EXIT(&rx_peerHashTable_lock);
7736 MUTEX_ENTER(&tp->peer_lock);
7737 peerStats->host = tp->host;
7738 peerStats->port = tp->port;
7739 peerStats->ifMTU = tp->ifMTU;
7740 peerStats->idleWhen = tp->idleWhen;
7741 peerStats->refCount = tp->refCount;
7742 peerStats->burstSize = tp->burstSize;
7743 peerStats->burst = tp->burst;
7744 peerStats->burstWait.sec = tp->burstWait.sec;
7745 peerStats->burstWait.usec = tp->burstWait.usec;
7746 peerStats->rtt = tp->rtt;
7747 peerStats->rtt_dev = tp->rtt_dev;
7748 peerStats->timeout.sec = 0;
7749 peerStats->timeout.usec = 0;
7750 peerStats->nSent = tp->nSent;
7751 peerStats->reSends = tp->reSends;
7752 peerStats->inPacketSkew = tp->inPacketSkew;
7753 peerStats->outPacketSkew = tp->outPacketSkew;
7754 peerStats->natMTU = tp->natMTU;
7755 peerStats->maxMTU = tp->maxMTU;
7756 peerStats->maxDgramPackets = tp->maxDgramPackets;
7757 peerStats->ifDgramPackets = tp->ifDgramPackets;
7758 peerStats->MTU = tp->MTU;
7759 peerStats->cwind = tp->cwind;
7760 peerStats->nDgramPackets = tp->nDgramPackets;
7761 peerStats->congestSeq = tp->congestSeq;
7762 peerStats->bytesSent.high = tp->bytesSent.high;
7763 peerStats->bytesSent.low = tp->bytesSent.low;
7764 peerStats->bytesReceived.high = tp->bytesReceived.high;
7765 peerStats->bytesReceived.low = tp->bytesReceived.low;
7766 MUTEX_EXIT(&tp->peer_lock);
7768 MUTEX_ENTER(&rx_peerHashTable_lock);
7771 MUTEX_EXIT(&rx_peerHashTable_lock);
7779 struct rx_serverQueueEntry *np;
7782 struct rx_call *call;
7783 struct rx_serverQueueEntry *sq;
7787 if (rxinit_status == 1) {
7789 return; /* Already shutdown. */
7793 #ifndef AFS_PTHREAD_ENV
7794 FD_ZERO(&rx_selectMask);
7795 #endif /* AFS_PTHREAD_ENV */
7796 rxi_dataQuota = RX_MAX_QUOTA;
7797 #ifndef AFS_PTHREAD_ENV
7799 #endif /* AFS_PTHREAD_ENV */
7802 #ifndef AFS_PTHREAD_ENV
7803 #ifndef AFS_USE_GETTIMEOFDAY
7805 #endif /* AFS_USE_GETTIMEOFDAY */
7806 #endif /* AFS_PTHREAD_ENV */
7808 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7809 call = queue_First(&rx_freeCallQueue, rx_call);
7811 rxi_Free(call, sizeof(struct rx_call));
7814 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7815 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7821 struct rx_peer **peer_ptr, **peer_end;
7822 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7823 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7825 struct rx_peer *peer, *next;
7827 MUTEX_ENTER(&rx_peerHashTable_lock);
7828 for (peer = *peer_ptr; peer; peer = next) {
7829 rx_interface_stat_p rpc_stat, nrpc_stat;
7832 MUTEX_ENTER(&rx_rpc_stats);
7833 MUTEX_ENTER(&peer->peer_lock);
7835 (&peer->rpcStats, rpc_stat, nrpc_stat,
7836 rx_interface_stat)) {
7837 unsigned int num_funcs;
7840 queue_Remove(&rpc_stat->queue_header);
7841 queue_Remove(&rpc_stat->all_peers);
7842 num_funcs = rpc_stat->stats[0].func_total;
7844 sizeof(rx_interface_stat_t) +
7845 rpc_stat->stats[0].func_total *
7846 sizeof(rx_function_entry_v1_t);
7848 rxi_Free(rpc_stat, space);
7850 /* rx_rpc_stats must be held */
7851 rxi_rpc_peer_stat_cnt -= num_funcs;
7853 MUTEX_EXIT(&peer->peer_lock);
7854 MUTEX_EXIT(&rx_rpc_stats);
7858 if (rx_stats_active)
7859 rx_atomic_dec(&rx_stats.nPeerStructs);
7861 MUTEX_EXIT(&rx_peerHashTable_lock);
7864 for (i = 0; i < RX_MAX_SERVICES; i++) {
7866 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7868 for (i = 0; i < rx_hashTableSize; i++) {
7869 struct rx_connection *tc, *ntc;
7870 MUTEX_ENTER(&rx_connHashTable_lock);
7871 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7873 for (j = 0; j < RX_MAXCALLS; j++) {
7875 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7878 rxi_Free(tc, sizeof(*tc));
7880 MUTEX_EXIT(&rx_connHashTable_lock);
7883 MUTEX_ENTER(&freeSQEList_lock);
7885 while ((np = rx_FreeSQEList)) {
7886 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7887 MUTEX_DESTROY(&np->lock);
7888 rxi_Free(np, sizeof(*np));
7891 MUTEX_EXIT(&freeSQEList_lock);
7892 MUTEX_DESTROY(&freeSQEList_lock);
7893 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7894 MUTEX_DESTROY(&rx_connHashTable_lock);
7895 MUTEX_DESTROY(&rx_peerHashTable_lock);
7896 MUTEX_DESTROY(&rx_serverPool_lock);
7898 osi_Free(rx_connHashTable,
7899 rx_hashTableSize * sizeof(struct rx_connection *));
7900 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7902 UNPIN(rx_connHashTable,
7903 rx_hashTableSize * sizeof(struct rx_connection *));
7904 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7906 rxi_FreeAllPackets();
7908 MUTEX_ENTER(&rx_quota_mutex);
7909 rxi_dataQuota = RX_MAX_QUOTA;
7910 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7911 MUTEX_EXIT(&rx_quota_mutex);
7916 #ifdef RX_ENABLE_LOCKS
7918 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7920 if (!MUTEX_ISMINE(lockaddr))
7921 osi_Panic("Lock not held: %s", msg);
7923 #endif /* RX_ENABLE_LOCKS */
7928 * Routines to implement connection specific data.
7932 rx_KeyCreate(rx_destructor_t rtn)
7935 MUTEX_ENTER(&rxi_keyCreate_lock);
7936 key = rxi_keyCreate_counter++;
7937 rxi_keyCreate_destructor = (rx_destructor_t *)
7938 realloc((void *)rxi_keyCreate_destructor,
7939 (key + 1) * sizeof(rx_destructor_t));
7940 rxi_keyCreate_destructor[key] = rtn;
7941 MUTEX_EXIT(&rxi_keyCreate_lock);
7946 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7949 MUTEX_ENTER(&conn->conn_data_lock);
7950 if (!conn->specific) {
7951 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7952 for (i = 0; i < key; i++)
7953 conn->specific[i] = NULL;
7954 conn->nSpecific = key + 1;
7955 conn->specific[key] = ptr;
7956 } else if (key >= conn->nSpecific) {
7957 conn->specific = (void **)
7958 realloc(conn->specific, (key + 1) * sizeof(void *));
7959 for (i = conn->nSpecific; i < key; i++)
7960 conn->specific[i] = NULL;
7961 conn->nSpecific = key + 1;
7962 conn->specific[key] = ptr;
7964 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7965 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7966 conn->specific[key] = ptr;
7968 MUTEX_EXIT(&conn->conn_data_lock);
7972 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7975 MUTEX_ENTER(&svc->svc_data_lock);
7976 if (!svc->specific) {
7977 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
7978 for (i = 0; i < key; i++)
7979 svc->specific[i] = NULL;
7980 svc->nSpecific = key + 1;
7981 svc->specific[key] = ptr;
7982 } else if (key >= svc->nSpecific) {
7983 svc->specific = (void **)
7984 realloc(svc->specific, (key + 1) * sizeof(void *));
7985 for (i = svc->nSpecific; i < key; i++)
7986 svc->specific[i] = NULL;
7987 svc->nSpecific = key + 1;
7988 svc->specific[key] = ptr;
7990 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7991 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7992 svc->specific[key] = ptr;
7994 MUTEX_EXIT(&svc->svc_data_lock);
7998 rx_GetSpecific(struct rx_connection *conn, int key)
8001 MUTEX_ENTER(&conn->conn_data_lock);
8002 if (key >= conn->nSpecific)
8005 ptr = conn->specific[key];
8006 MUTEX_EXIT(&conn->conn_data_lock);
8011 rx_GetServiceSpecific(struct rx_service *svc, int key)
8014 MUTEX_ENTER(&svc->svc_data_lock);
8015 if (key >= svc->nSpecific)
8018 ptr = svc->specific[key];
8019 MUTEX_EXIT(&svc->svc_data_lock);
8024 #endif /* !KERNEL */
8027 * processStats is a queue used to store the statistics for the local
8028 * process. Its contents are similar to the contents of the rpcStats
8029 * queue on a rx_peer structure, but the actual data stored within
8030 * this queue contains totals across the lifetime of the process (assuming
8031 * the stats have not been reset) - unlike the per peer structures
8032 * which can come and go based upon the peer lifetime.
8035 static struct rx_queue processStats = { &processStats, &processStats };
8038 * peerStats is a queue used to store the statistics for all peer structs.
8039 * Its contents are the union of all the peer rpcStats queues.
8042 static struct rx_queue peerStats = { &peerStats, &peerStats };
8045 * rxi_monitor_processStats is used to turn process wide stat collection
8049 static int rxi_monitor_processStats = 0;
8052 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8055 static int rxi_monitor_peerStats = 0;
8058 * rxi_AddRpcStat - given all of the information for a particular rpc
8059 * call, create (if needed) and update the stat totals for the rpc.
8063 * IN stats - the queue of stats that will be updated with the new value
8065 * IN rxInterface - a unique number that identifies the rpc interface
8067 * IN currentFunc - the index of the function being invoked
8069 * IN totalFunc - the total number of functions in this interface
8071 * IN queueTime - the amount of time this function waited for a thread
8073 * IN execTime - the amount of time this function invocation took to execute
8075 * IN bytesSent - the number bytes sent by this invocation
8077 * IN bytesRcvd - the number bytes received by this invocation
8079 * IN isServer - if true, this invocation was made to a server
8081 * IN remoteHost - the ip address of the remote host
8083 * IN remotePort - the port of the remote host
8085 * IN addToPeerList - if != 0, add newly created stat to the global peer list
8087 * INOUT counter - if a new stats structure is allocated, the counter will
8088 * be updated with the new number of allocated stat structures
8096 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
8097 afs_uint32 currentFunc, afs_uint32 totalFunc,
8098 struct clock *queueTime, struct clock *execTime,
8099 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
8100 afs_uint32 remoteHost, afs_uint32 remotePort,
8101 int addToPeerList, unsigned int *counter)
8104 rx_interface_stat_p rpc_stat, nrpc_stat;
8107 * See if there's already a structure for this interface
8110 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8111 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8112 && (rpc_stat->stats[0].remote_is_server == isServer))
8117 * Didn't find a match so allocate a new structure and add it to the
8121 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
8122 || (rpc_stat->stats[0].interfaceId != rxInterface)
8123 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8128 sizeof(rx_interface_stat_t) +
8129 totalFunc * sizeof(rx_function_entry_v1_t);
8131 rpc_stat = rxi_Alloc(space);
8132 if (rpc_stat == NULL) {
8136 *counter += totalFunc;
8137 for (i = 0; i < totalFunc; i++) {
8138 rpc_stat->stats[i].remote_peer = remoteHost;
8139 rpc_stat->stats[i].remote_port = remotePort;
8140 rpc_stat->stats[i].remote_is_server = isServer;
8141 rpc_stat->stats[i].interfaceId = rxInterface;
8142 rpc_stat->stats[i].func_total = totalFunc;
8143 rpc_stat->stats[i].func_index = i;
8144 hzero(rpc_stat->stats[i].invocations);
8145 hzero(rpc_stat->stats[i].bytes_sent);
8146 hzero(rpc_stat->stats[i].bytes_rcvd);
8147 rpc_stat->stats[i].queue_time_sum.sec = 0;
8148 rpc_stat->stats[i].queue_time_sum.usec = 0;
8149 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8150 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8151 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8152 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8153 rpc_stat->stats[i].queue_time_max.sec = 0;
8154 rpc_stat->stats[i].queue_time_max.usec = 0;
8155 rpc_stat->stats[i].execution_time_sum.sec = 0;
8156 rpc_stat->stats[i].execution_time_sum.usec = 0;
8157 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8158 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8159 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8160 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8161 rpc_stat->stats[i].execution_time_max.sec = 0;
8162 rpc_stat->stats[i].execution_time_max.usec = 0;
8164 queue_Prepend(stats, rpc_stat);
8165 if (addToPeerList) {
8166 queue_Prepend(&peerStats, &rpc_stat->all_peers);
8171 * Increment the stats for this function
8174 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
8175 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
8176 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
8177 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8178 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8179 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8180 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8182 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8183 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8185 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8186 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8188 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8189 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8191 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8192 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8200 * rx_IncrementTimeAndCount - increment the times and count for a particular
8205 * IN peer - the peer who invoked the rpc
8207 * IN rxInterface - a unique number that identifies the rpc interface
8209 * IN currentFunc - the index of the function being invoked
8211 * IN totalFunc - the total number of functions in this interface
8213 * IN queueTime - the amount of time this function waited for a thread
8215 * IN execTime - the amount of time this function invocation took to execute
8217 * IN bytesSent - the number bytes sent by this invocation
8219 * IN bytesRcvd - the number bytes received by this invocation
8221 * IN isServer - if true, this invocation was made to a server
8229 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8230 afs_uint32 currentFunc, afs_uint32 totalFunc,
8231 struct clock *queueTime, struct clock *execTime,
8232 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8236 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8239 MUTEX_ENTER(&rx_rpc_stats);
8241 if (rxi_monitor_peerStats) {
8242 MUTEX_ENTER(&peer->peer_lock);
8243 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8244 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8245 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8246 MUTEX_EXIT(&peer->peer_lock);
8249 if (rxi_monitor_processStats) {
8250 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8251 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8252 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8255 MUTEX_EXIT(&rx_rpc_stats);
8260 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8264 * IN callerVersion - the rpc stat version of the caller.
8266 * IN count - the number of entries to marshall.
8268 * IN stats - pointer to stats to be marshalled.
8270 * OUT ptr - Where to store the marshalled data.
8277 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8278 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8284 * We only support the first version
8286 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8287 *(ptr++) = stats->remote_peer;
8288 *(ptr++) = stats->remote_port;
8289 *(ptr++) = stats->remote_is_server;
8290 *(ptr++) = stats->interfaceId;
8291 *(ptr++) = stats->func_total;
8292 *(ptr++) = stats->func_index;
8293 *(ptr++) = hgethi(stats->invocations);
8294 *(ptr++) = hgetlo(stats->invocations);
8295 *(ptr++) = hgethi(stats->bytes_sent);
8296 *(ptr++) = hgetlo(stats->bytes_sent);
8297 *(ptr++) = hgethi(stats->bytes_rcvd);
8298 *(ptr++) = hgetlo(stats->bytes_rcvd);
8299 *(ptr++) = stats->queue_time_sum.sec;
8300 *(ptr++) = stats->queue_time_sum.usec;
8301 *(ptr++) = stats->queue_time_sum_sqr.sec;
8302 *(ptr++) = stats->queue_time_sum_sqr.usec;
8303 *(ptr++) = stats->queue_time_min.sec;
8304 *(ptr++) = stats->queue_time_min.usec;
8305 *(ptr++) = stats->queue_time_max.sec;
8306 *(ptr++) = stats->queue_time_max.usec;
8307 *(ptr++) = stats->execution_time_sum.sec;
8308 *(ptr++) = stats->execution_time_sum.usec;
8309 *(ptr++) = stats->execution_time_sum_sqr.sec;
8310 *(ptr++) = stats->execution_time_sum_sqr.usec;
8311 *(ptr++) = stats->execution_time_min.sec;
8312 *(ptr++) = stats->execution_time_min.usec;
8313 *(ptr++) = stats->execution_time_max.sec;
8314 *(ptr++) = stats->execution_time_max.usec;
8320 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8325 * IN callerVersion - the rpc stat version of the caller
8327 * OUT myVersion - the rpc stat version of this function
8329 * OUT clock_sec - local time seconds
8331 * OUT clock_usec - local time microseconds
8333 * OUT allocSize - the number of bytes allocated to contain stats
8335 * OUT statCount - the number stats retrieved from this process.
8337 * OUT stats - the actual stats retrieved from this process.
8341 * Returns void. If successful, stats will != NULL.
8345 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8346 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8347 size_t * allocSize, afs_uint32 * statCount,
8348 afs_uint32 ** stats)
8358 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8361 * Check to see if stats are enabled
8364 MUTEX_ENTER(&rx_rpc_stats);
8365 if (!rxi_monitor_processStats) {
8366 MUTEX_EXIT(&rx_rpc_stats);
8370 clock_GetTime(&now);
8371 *clock_sec = now.sec;
8372 *clock_usec = now.usec;
8375 * Allocate the space based upon the caller version
8377 * If the client is at an older version than we are,
8378 * we return the statistic data in the older data format, but
8379 * we still return our version number so the client knows we
8380 * are maintaining more data than it can retrieve.
8383 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8384 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8385 *statCount = rxi_rpc_process_stat_cnt;
8388 * This can't happen yet, but in the future version changes
8389 * can be handled by adding additional code here
8393 if (space > (size_t) 0) {
8395 ptr = *stats = rxi_Alloc(space);
8398 rx_interface_stat_p rpc_stat, nrpc_stat;
8402 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8404 * Copy the data based upon the caller version
8406 rx_MarshallProcessRPCStats(callerVersion,
8407 rpc_stat->stats[0].func_total,
8408 rpc_stat->stats, &ptr);
8414 MUTEX_EXIT(&rx_rpc_stats);
8419 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8423 * IN callerVersion - the rpc stat version of the caller
8425 * OUT myVersion - the rpc stat version of this function
8427 * OUT clock_sec - local time seconds
8429 * OUT clock_usec - local time microseconds
8431 * OUT allocSize - the number of bytes allocated to contain stats
8433 * OUT statCount - the number of stats retrieved from the individual
8436 * OUT stats - the actual stats retrieved from the individual peer structures.
8440 * Returns void. If successful, stats will != NULL.
8444 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8445 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8446 size_t * allocSize, afs_uint32 * statCount,
8447 afs_uint32 ** stats)
8457 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8460 * Check to see if stats are enabled
8463 MUTEX_ENTER(&rx_rpc_stats);
8464 if (!rxi_monitor_peerStats) {
8465 MUTEX_EXIT(&rx_rpc_stats);
8469 clock_GetTime(&now);
8470 *clock_sec = now.sec;
8471 *clock_usec = now.usec;
8474 * Allocate the space based upon the caller version
8476 * If the client is at an older version than we are,
8477 * we return the statistic data in the older data format, but
8478 * we still return our version number so the client knows we
8479 * are maintaining more data than it can retrieve.
8482 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8483 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8484 *statCount = rxi_rpc_peer_stat_cnt;
8487 * This can't happen yet, but in the future version changes
8488 * can be handled by adding additional code here
8492 if (space > (size_t) 0) {
8494 ptr = *stats = rxi_Alloc(space);
8497 rx_interface_stat_p rpc_stat, nrpc_stat;
8501 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8503 * We have to fix the offset of rpc_stat since we are
8504 * keeping this structure on two rx_queues. The rx_queue
8505 * package assumes that the rx_queue member is the first
8506 * member of the structure. That is, rx_queue assumes that
8507 * any one item is only on one queue at a time. We are
8508 * breaking that assumption and so we have to do a little
8509 * math to fix our pointers.
8512 fix_offset = (char *)rpc_stat;
8513 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8514 rpc_stat = (rx_interface_stat_p) fix_offset;
8517 * Copy the data based upon the caller version
8519 rx_MarshallProcessRPCStats(callerVersion,
8520 rpc_stat->stats[0].func_total,
8521 rpc_stat->stats, &ptr);
8527 MUTEX_EXIT(&rx_rpc_stats);
8532 * rx_FreeRPCStats - free memory allocated by
8533 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8537 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8538 * rx_RetrievePeerRPCStats
8540 * IN allocSize - the number of bytes in stats.
8548 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8550 rxi_Free(stats, allocSize);
8554 * rx_queryProcessRPCStats - see if process rpc stat collection is
8555 * currently enabled.
8561 * Returns 0 if stats are not enabled != 0 otherwise
8565 rx_queryProcessRPCStats(void)
8568 MUTEX_ENTER(&rx_rpc_stats);
8569 rc = rxi_monitor_processStats;
8570 MUTEX_EXIT(&rx_rpc_stats);
8575 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8581 * Returns 0 if stats are not enabled != 0 otherwise
8585 rx_queryPeerRPCStats(void)
8588 MUTEX_ENTER(&rx_rpc_stats);
8589 rc = rxi_monitor_peerStats;
8590 MUTEX_EXIT(&rx_rpc_stats);
8595 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8605 rx_enableProcessRPCStats(void)
8607 MUTEX_ENTER(&rx_rpc_stats);
8608 rx_enable_stats = 1;
8609 rxi_monitor_processStats = 1;
8610 MUTEX_EXIT(&rx_rpc_stats);
8614 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8624 rx_enablePeerRPCStats(void)
8626 MUTEX_ENTER(&rx_rpc_stats);
8627 rx_enable_stats = 1;
8628 rxi_monitor_peerStats = 1;
8629 MUTEX_EXIT(&rx_rpc_stats);
8633 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8643 rx_disableProcessRPCStats(void)
8645 rx_interface_stat_p rpc_stat, nrpc_stat;
8648 MUTEX_ENTER(&rx_rpc_stats);
8651 * Turn off process statistics and if peer stats is also off, turn
8655 rxi_monitor_processStats = 0;
8656 if (rxi_monitor_peerStats == 0) {
8657 rx_enable_stats = 0;
8660 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8661 unsigned int num_funcs = 0;
8664 queue_Remove(rpc_stat);
8665 num_funcs = rpc_stat->stats[0].func_total;
8667 sizeof(rx_interface_stat_t) +
8668 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8670 rxi_Free(rpc_stat, space);
8671 rxi_rpc_process_stat_cnt -= num_funcs;
8673 MUTEX_EXIT(&rx_rpc_stats);
8677 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8687 rx_disablePeerRPCStats(void)
8689 struct rx_peer **peer_ptr, **peer_end;
8693 * Turn off peer statistics and if process stats is also off, turn
8697 rxi_monitor_peerStats = 0;
8698 if (rxi_monitor_processStats == 0) {
8699 rx_enable_stats = 0;
8702 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8703 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8705 struct rx_peer *peer, *next, *prev;
8707 MUTEX_ENTER(&rx_peerHashTable_lock);
8708 MUTEX_ENTER(&rx_rpc_stats);
8709 for (prev = peer = *peer_ptr; peer; peer = next) {
8711 code = MUTEX_TRYENTER(&peer->peer_lock);
8713 rx_interface_stat_p rpc_stat, nrpc_stat;
8716 if (prev == *peer_ptr) {
8727 MUTEX_EXIT(&rx_peerHashTable_lock);
8730 (&peer->rpcStats, rpc_stat, nrpc_stat,
8731 rx_interface_stat)) {
8732 unsigned int num_funcs = 0;
8735 queue_Remove(&rpc_stat->queue_header);
8736 queue_Remove(&rpc_stat->all_peers);
8737 num_funcs = rpc_stat->stats[0].func_total;
8739 sizeof(rx_interface_stat_t) +
8740 rpc_stat->stats[0].func_total *
8741 sizeof(rx_function_entry_v1_t);
8743 rxi_Free(rpc_stat, space);
8744 rxi_rpc_peer_stat_cnt -= num_funcs;
8746 MUTEX_EXIT(&peer->peer_lock);
8748 MUTEX_ENTER(&rx_peerHashTable_lock);
8758 MUTEX_EXIT(&rx_rpc_stats);
8759 MUTEX_EXIT(&rx_peerHashTable_lock);
8764 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8769 * IN clearFlag - flag indicating which stats to clear
8777 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8779 rx_interface_stat_p rpc_stat, nrpc_stat;
8781 MUTEX_ENTER(&rx_rpc_stats);
8783 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8784 unsigned int num_funcs = 0, i;
8785 num_funcs = rpc_stat->stats[0].func_total;
8786 for (i = 0; i < num_funcs; i++) {
8787 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8788 hzero(rpc_stat->stats[i].invocations);
8790 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8791 hzero(rpc_stat->stats[i].bytes_sent);
8793 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8794 hzero(rpc_stat->stats[i].bytes_rcvd);
8796 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8797 rpc_stat->stats[i].queue_time_sum.sec = 0;
8798 rpc_stat->stats[i].queue_time_sum.usec = 0;
8800 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8801 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8802 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8804 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8805 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8806 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8808 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8809 rpc_stat->stats[i].queue_time_max.sec = 0;
8810 rpc_stat->stats[i].queue_time_max.usec = 0;
8812 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8813 rpc_stat->stats[i].execution_time_sum.sec = 0;
8814 rpc_stat->stats[i].execution_time_sum.usec = 0;
8816 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8817 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8818 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8820 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8821 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8822 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8824 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8825 rpc_stat->stats[i].execution_time_max.sec = 0;
8826 rpc_stat->stats[i].execution_time_max.usec = 0;
8831 MUTEX_EXIT(&rx_rpc_stats);
8835 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8840 * IN clearFlag - flag indicating which stats to clear
8848 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8850 rx_interface_stat_p rpc_stat, nrpc_stat;
8852 MUTEX_ENTER(&rx_rpc_stats);
8854 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8855 unsigned int num_funcs = 0, i;
8858 * We have to fix the offset of rpc_stat since we are
8859 * keeping this structure on two rx_queues. The rx_queue
8860 * package assumes that the rx_queue member is the first
8861 * member of the structure. That is, rx_queue assumes that
8862 * any one item is only on one queue at a time. We are
8863 * breaking that assumption and so we have to do a little
8864 * math to fix our pointers.
8867 fix_offset = (char *)rpc_stat;
8868 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8869 rpc_stat = (rx_interface_stat_p) fix_offset;
8871 num_funcs = rpc_stat->stats[0].func_total;
8872 for (i = 0; i < num_funcs; i++) {
8873 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8874 hzero(rpc_stat->stats[i].invocations);
8876 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8877 hzero(rpc_stat->stats[i].bytes_sent);
8879 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8880 hzero(rpc_stat->stats[i].bytes_rcvd);
8882 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8883 rpc_stat->stats[i].queue_time_sum.sec = 0;
8884 rpc_stat->stats[i].queue_time_sum.usec = 0;
8886 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8887 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8888 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8890 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8891 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8892 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8894 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8895 rpc_stat->stats[i].queue_time_max.sec = 0;
8896 rpc_stat->stats[i].queue_time_max.usec = 0;
8898 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8899 rpc_stat->stats[i].execution_time_sum.sec = 0;
8900 rpc_stat->stats[i].execution_time_sum.usec = 0;
8902 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8903 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8904 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8906 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8907 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8908 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8910 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8911 rpc_stat->stats[i].execution_time_max.sec = 0;
8912 rpc_stat->stats[i].execution_time_max.usec = 0;
8917 MUTEX_EXIT(&rx_rpc_stats);
8921 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8922 * is authorized to enable/disable/clear RX statistics.
8924 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8927 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8929 rxi_rxstat_userok = proc;
8933 rx_RxStatUserOk(struct rx_call *call)
8935 if (!rxi_rxstat_userok)
8937 return rxi_rxstat_userok(call);
8942 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8943 * function in the MSVC runtime DLL (msvcrt.dll).
8945 * Note: the system serializes calls to this function.
8948 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8949 DWORD reason, /* reason function is being called */
8950 LPVOID reserved) /* reserved for future use */
8953 case DLL_PROCESS_ATTACH:
8954 /* library is being attached to a process */
8958 case DLL_PROCESS_DETACH:
8965 #endif /* AFS_NT40_ENV */
8968 int rx_DumpCalls(FILE *outputFile, char *cookie)
8970 #ifdef RXDEBUG_PACKET
8971 #ifdef KDUMP_RX_LOCK
8972 struct rx_call_rx_lock *c;
8979 #define RXDPRINTF sprintf
8980 #define RXDPRINTOUT output
8982 #define RXDPRINTF fprintf
8983 #define RXDPRINTOUT outputFile
8986 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8988 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8991 for (c = rx_allCallsp; c; c = c->allNextp) {
8992 u_short rqc, tqc, iovqc;
8993 struct rx_packet *p, *np;
8995 MUTEX_ENTER(&c->lock);
8996 queue_Count(&c->rq, p, np, rx_packet, rqc);
8997 queue_Count(&c->tq, p, np, rx_packet, tqc);
8998 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
9000 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, "
9001 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9002 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9003 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9004 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
9005 #ifdef RX_ENABLE_LOCKS
9008 #ifdef RX_REFCOUNT_CHECK
9009 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9010 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9013 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,
9014 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9015 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9016 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9017 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9018 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
9019 #ifdef RX_ENABLE_LOCKS
9020 , (afs_uint32)c->refCount
9022 #ifdef RX_REFCOUNT_CHECK
9023 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9026 MUTEX_EXIT(&c->lock);
9029 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9032 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9034 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9036 #endif /* RXDEBUG_PACKET */