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"
85 #include <afs/rxgen_consts.h>
88 #ifdef AFS_PTHREAD_ENV
90 int (*registerProgram) (pid_t, char *) = 0;
91 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
94 int (*registerProgram) (PROCESS, char *) = 0;
95 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
99 /* Local static routines */
100 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
101 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
102 struct rx_call *, struct rx_peer *,
104 static void rxi_Resend(struct rxevent *event, void *arg0, void *arg1,
106 static void rxi_SendDelayedAck(struct rxevent *event, void *call,
107 void *dummy, int dummy2);
108 static void rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1,
109 void *dummy, int dummy2);
110 static void rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1,
111 void *unused, int unused2);
112 static void rxi_ReapConnections(struct rxevent *unused, void *unused1,
113 void *unused2, int unused3);
115 #ifdef RX_ENABLE_LOCKS
116 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
119 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
121 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
122 rx_atomic_t rxi_start_in_error;
124 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
126 /* Constant delay time before sending an acknowledge of the last packet
127 * received. This is to avoid sending an extra acknowledge when the
128 * client is about to make another call, anyway, or the server is
131 * The lastAckDelay may not exceeed 400ms without causing peers to
132 * unecessarily timeout.
134 struct clock rx_lastAckDelay = {0, 400000};
136 /* Constant delay time before sending a soft ack when none was requested.
137 * This is to make sure we send soft acks before the sender times out,
138 * Normally we wait and send a hard ack when the receiver consumes the packet
140 * This value has been 100ms in all shipping versions of OpenAFS. Changing it
141 * will require changes to the peer's RTT calculations.
143 struct clock rx_softAckDelay = {0, 100000};
146 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
147 * currently allocated within rx. This number is used to allocate the
148 * memory required to return the statistics when queried.
149 * Protected by the rx_rpc_stats mutex.
152 static unsigned int rxi_rpc_peer_stat_cnt;
155 * rxi_rpc_process_stat_cnt counts the total number of local process stat
156 * structures currently allocated within rx. The number is used to allocate
157 * the memory required to return the statistics when queried.
158 * Protected by the rx_rpc_stats mutex.
161 static unsigned int rxi_rpc_process_stat_cnt;
164 * rxi_busyChannelError is the error to return to the application when a call
165 * channel appears busy (inferred from the receipt of RX_PACKET_TYPE_BUSY
166 * packets on the channel), and there are other call channels in the
167 * connection that are not busy. If 0, we do not return errors upon receiving
168 * busy packets; we just keep trying on the same call channel until we hit a
171 static afs_int32 rxi_busyChannelError = 0;
173 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
174 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
176 #if !defined(offsetof)
177 #include <stddef.h> /* for definition of offsetof() */
180 #ifdef RX_ENABLE_LOCKS
181 afs_kmutex_t rx_atomic_mutex;
184 /* Forward prototypes */
185 static struct rx_call * rxi_NewCall(struct rx_connection *, int);
187 #ifdef AFS_PTHREAD_ENV
190 * Use procedural initialization of mutexes/condition variables
194 extern afs_kmutex_t rx_quota_mutex;
195 extern afs_kmutex_t rx_pthread_mutex;
196 extern afs_kmutex_t rx_packets_mutex;
197 extern afs_kmutex_t rx_refcnt_mutex;
198 extern afs_kmutex_t des_init_mutex;
199 extern afs_kmutex_t des_random_mutex;
200 extern afs_kmutex_t rx_clock_mutex;
201 extern afs_kmutex_t rxi_connCacheMutex;
202 extern afs_kmutex_t event_handler_mutex;
203 extern afs_kmutex_t listener_mutex;
204 extern afs_kmutex_t rx_if_init_mutex;
205 extern afs_kmutex_t rx_if_mutex;
207 extern afs_kcondvar_t rx_event_handler_cond;
208 extern afs_kcondvar_t rx_listener_cond;
210 static afs_kmutex_t epoch_mutex;
211 static afs_kmutex_t rx_init_mutex;
212 static afs_kmutex_t rx_debug_mutex;
213 static afs_kmutex_t rx_rpc_stats;
216 rxi_InitPthread(void)
218 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
219 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
220 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
221 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
222 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
223 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
224 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
225 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
226 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
227 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
228 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
229 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
230 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
231 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
232 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
234 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
235 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
237 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
238 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
240 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
241 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
242 #ifdef RX_ENABLE_LOCKS
245 #endif /* RX_LOCKS_DB */
246 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
247 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
249 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
251 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
253 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
255 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
256 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
257 #endif /* RX_ENABLE_LOCKS */
260 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
261 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
263 * The rx_stats_mutex mutex protects the following global variables:
264 * rxi_lowConnRefCount
265 * rxi_lowPeerRefCount
274 * The rx_quota_mutex mutex protects the following global variables:
282 * The rx_freePktQ_lock protects the following global variables:
287 * The rx_packets_mutex mutex protects the following global variables:
295 * The rx_pthread_mutex mutex protects the following global variables:
296 * rxi_fcfs_thread_num
299 #define INIT_PTHREAD_LOCKS
303 /* Variables for handling the minProcs implementation. availProcs gives the
304 * number of threads available in the pool at this moment (not counting dudes
305 * executing right now). totalMin gives the total number of procs required
306 * for handling all minProcs requests. minDeficit is a dynamic variable
307 * tracking the # of procs required to satisfy all of the remaining minProcs
309 * For fine grain locking to work, the quota check and the reservation of
310 * a server thread has to come while rxi_availProcs and rxi_minDeficit
311 * are locked. To this end, the code has been modified under #ifdef
312 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
313 * same time. A new function, ReturnToServerPool() returns the allocation.
315 * A call can be on several queue's (but only one at a time). When
316 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
317 * that no one else is touching the queue. To this end, we store the address
318 * of the queue lock in the call structure (under the call lock) when we
319 * put the call on a queue, and we clear the call_queue_lock when the
320 * call is removed from a queue (once the call lock has been obtained).
321 * This allows rxi_ResetCall to safely synchronize with others wishing
322 * to manipulate the queue.
325 #if defined(RX_ENABLE_LOCKS)
326 static afs_kmutex_t rx_rpc_stats;
329 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
330 ** pretty good that the next packet coming in is from the same connection
331 ** as the last packet, since we're send multiple packets in a transmit window.
333 struct rx_connection *rxLastConn = 0;
335 #ifdef RX_ENABLE_LOCKS
336 /* The locking hierarchy for rx fine grain locking is composed of these
339 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
340 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
341 * call->lock - locks call data fields.
342 * These are independent of each other:
343 * rx_freeCallQueue_lock
348 * serverQueueEntry->lock
349 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
351 * peer->lock - locks peer data fields.
352 * conn_data_lock - that more than one thread is not updating a conn data
353 * field at the same time.
364 * Do we need a lock to protect the peer field in the conn structure?
365 * conn->peer was previously a constant for all intents and so has no
366 * lock protecting this field. The multihomed client delta introduced
367 * a RX code change : change the peer field in the connection structure
368 * to that remote interface from which the last packet for this
369 * connection was sent out. This may become an issue if further changes
372 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
373 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
375 /* rxdb_fileID is used to identify the lock location, along with line#. */
376 static int rxdb_fileID = RXDB_FILE_RX;
377 #endif /* RX_LOCKS_DB */
378 #else /* RX_ENABLE_LOCKS */
379 #define SET_CALL_QUEUE_LOCK(C, L)
380 #define CLEAR_CALL_QUEUE_LOCK(C)
381 #endif /* RX_ENABLE_LOCKS */
382 struct rx_serverQueueEntry *rx_waitForPacket = 0;
383 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
385 /* ------------Exported Interfaces------------- */
387 /* This function allows rxkad to set the epoch to a suitably random number
388 * which rx_NewConnection will use in the future. The principle purpose is to
389 * get rxnull connections to use the same epoch as the rxkad connections do, at
390 * least once the first rxkad connection is established. This is important now
391 * that the host/port addresses aren't used in FindConnection: the uniqueness
392 * of epoch/cid matters and the start time won't do. */
394 #ifdef AFS_PTHREAD_ENV
396 * This mutex protects the following global variables:
400 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
401 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
405 #endif /* AFS_PTHREAD_ENV */
408 rx_SetEpoch(afs_uint32 epoch)
415 /* Initialize rx. A port number may be mentioned, in which case this
416 * becomes the default port number for any service installed later.
417 * If 0 is provided for the port number, a random port will be chosen
418 * by the kernel. Whether this will ever overlap anything in
419 * /etc/services is anybody's guess... Returns 0 on success, -1 on
424 int rxinit_status = 1;
425 #ifdef AFS_PTHREAD_ENV
427 * This mutex protects the following global variables:
431 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
432 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
435 #define UNLOCK_RX_INIT
439 rx_InitHost(u_int host, u_int port)
446 char *htable, *ptable;
453 if (rxinit_status == 0) {
454 tmp_status = rxinit_status;
456 return tmp_status; /* Already started; return previous error code. */
462 if (afs_winsockInit() < 0)
468 * Initialize anything necessary to provide a non-premptive threading
471 rxi_InitializeThreadSupport();
474 /* Allocate and initialize a socket for client and perhaps server
477 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
478 if (rx_socket == OSI_NULLSOCKET) {
482 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
485 #endif /* RX_LOCKS_DB */
486 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
487 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
488 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
489 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
490 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
491 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
492 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
493 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
494 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
496 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
498 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
500 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
502 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
503 #if defined(AFS_HPUX110_ENV)
505 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
506 #endif /* AFS_HPUX110_ENV */
507 #endif /* RX_ENABLE_LOCKS && KERNEL */
510 rx_connDeadTime = 12;
511 rx_tranquil = 0; /* reset flag */
512 rxi_ResetStatistics();
514 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
515 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
516 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
517 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
518 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
519 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
521 /* Malloc up a bunch of packets & buffers */
523 queue_Init(&rx_freePacketQueue);
524 rxi_NeedMorePackets = FALSE;
525 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
527 /* enforce a minimum number of allocated packets */
528 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
529 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
531 /* allocate the initial free packet pool */
532 #ifdef RX_ENABLE_TSFPQ
533 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
534 #else /* RX_ENABLE_TSFPQ */
535 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
536 #endif /* RX_ENABLE_TSFPQ */
543 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
544 tv.tv_sec = clock_now.sec;
545 tv.tv_usec = clock_now.usec;
546 srand((unsigned int)tv.tv_usec);
553 #if defined(KERNEL) && !defined(UKERNEL)
554 /* Really, this should never happen in a real kernel */
557 struct sockaddr_in addr;
559 int addrlen = sizeof(addr);
561 socklen_t addrlen = sizeof(addr);
563 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
567 rx_port = addr.sin_port;
570 rx_stats.minRtt.sec = 9999999;
572 rx_SetEpoch(tv.tv_sec | 0x80000000);
574 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
575 * will provide a randomer value. */
577 MUTEX_ENTER(&rx_quota_mutex);
578 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
579 MUTEX_EXIT(&rx_quota_mutex);
580 /* *Slightly* random start time for the cid. This is just to help
581 * out with the hashing function at the peer */
582 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
583 rx_connHashTable = (struct rx_connection **)htable;
584 rx_peerHashTable = (struct rx_peer **)ptable;
586 rx_hardAckDelay.sec = 0;
587 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
589 rxevent_Init(20, rxi_ReScheduleEvents);
591 /* Initialize various global queues */
592 queue_Init(&rx_idleServerQueue);
593 queue_Init(&rx_incomingCallQueue);
594 queue_Init(&rx_freeCallQueue);
596 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
597 /* Initialize our list of usable IP addresses. */
601 #if defined(RXK_LISTENER_ENV) || !defined(KERNEL)
602 /* Start listener process (exact function is dependent on the
603 * implementation environment--kernel or user space) */
608 tmp_status = rxinit_status = 0;
616 return rx_InitHost(htonl(INADDR_ANY), port);
622 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
623 * maintaing the round trip timer.
628 * Start a new RTT timer for a given call and packet.
630 * There must be no resendEvent already listed for this call, otherwise this
631 * will leak events - intended for internal use within the RTO code only
634 * the RX call to start the timer for
635 * @param[in] lastPacket
636 * a flag indicating whether the last packet has been sent or not
638 * @pre call must be locked before calling this function
642 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
644 struct clock now, retryTime;
649 clock_Add(&retryTime, &call->rto);
651 /* If we're sending the last packet, and we're the client, then the server
652 * may wait for an additional 400ms before returning the ACK, wait for it
653 * rather than hitting a timeout */
654 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
655 clock_Addmsec(&retryTime, 400);
657 MUTEX_ENTER(&rx_refcnt_mutex);
658 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
659 MUTEX_EXIT(&rx_refcnt_mutex);
660 call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
665 * Cancel an RTT timer for a given call.
669 * the RX call to cancel the timer for
671 * @pre call must be locked before calling this function
676 rxi_rto_cancel(struct rx_call *call)
678 rxevent_Cancel(&call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
682 * Tell the RTO timer that we have sent a packet.
684 * If the timer isn't already running, then start it. If the timer is running,
688 * the RX call that the packet has been sent on
689 * @param[in] lastPacket
690 * A flag which is true if this is the last packet for the call
692 * @pre The call must be locked before calling this function
697 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
699 if (call->resendEvent)
702 rxi_rto_startTimer(call, lastPacket, istack);
706 * Tell the RTO timer that we have received an new ACK message
708 * This function should be called whenever a call receives an ACK that
709 * acknowledges new packets. Whatever happens, we stop the current timer.
710 * If there are unacked packets in the queue which have been sent, then
711 * we restart the timer from now. Otherwise, we leave it stopped.
714 * the RX call that the ACK has been received on
718 rxi_rto_packet_acked(struct rx_call *call, int istack)
720 struct rx_packet *p, *nxp;
722 rxi_rto_cancel(call);
724 if (queue_IsEmpty(&call->tq))
727 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
728 if (p->header.seq > call->tfirst + call->twind)
731 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
732 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
740 * Set an initial round trip timeout for a peer connection
742 * @param[in] secs The timeout to set in seconds
746 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
747 peer->rtt = secs * 8000;
751 * Sets the error generated when a busy call channel is detected.
753 * @param[in] error The error to return for a call on a busy channel.
755 * @pre Neither rx_Init nor rx_InitHost have been called yet
758 rx_SetBusyChannelError(afs_int32 error)
760 osi_Assert(rxinit_status != 0);
761 rxi_busyChannelError = error;
765 * Set a delayed ack event on the specified call for the given time
767 * @param[in] call - the call on which to set the event
768 * @param[in] offset - the delay from now after which the event fires
771 rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
773 struct clock now, when;
777 clock_Add(&when, offset);
779 if (!call->delayedAckEvent
780 || clock_Gt(&call->delayedAckTime, &when)) {
782 rxevent_Cancel(&call->delayedAckEvent, call,
783 RX_CALL_REFCOUNT_DELAY);
784 MUTEX_ENTER(&rx_refcnt_mutex);
785 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
786 MUTEX_EXIT(&rx_refcnt_mutex);
788 call->delayedAckEvent = rxevent_Post(&when, &now,
791 call->delayedAckTime = when;
795 /* called with unincremented nRequestsRunning to see if it is OK to start
796 * a new thread in this service. Could be "no" for two reasons: over the
797 * max quota, or would prevent others from reaching their min quota.
799 #ifdef RX_ENABLE_LOCKS
800 /* This verion of QuotaOK reserves quota if it's ok while the
801 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
804 QuotaOK(struct rx_service *aservice)
806 /* check if over max quota */
807 if (aservice->nRequestsRunning >= aservice->maxProcs) {
811 /* under min quota, we're OK */
812 /* otherwise, can use only if there are enough to allow everyone
813 * to go to their min quota after this guy starts.
816 MUTEX_ENTER(&rx_quota_mutex);
817 if ((aservice->nRequestsRunning < aservice->minProcs)
818 || (rxi_availProcs > rxi_minDeficit)) {
819 aservice->nRequestsRunning++;
820 /* just started call in minProcs pool, need fewer to maintain
822 if (aservice->nRequestsRunning <= aservice->minProcs)
825 MUTEX_EXIT(&rx_quota_mutex);
828 MUTEX_EXIT(&rx_quota_mutex);
834 ReturnToServerPool(struct rx_service *aservice)
836 aservice->nRequestsRunning--;
837 MUTEX_ENTER(&rx_quota_mutex);
838 if (aservice->nRequestsRunning < aservice->minProcs)
841 MUTEX_EXIT(&rx_quota_mutex);
844 #else /* RX_ENABLE_LOCKS */
846 QuotaOK(struct rx_service *aservice)
849 /* under min quota, we're OK */
850 if (aservice->nRequestsRunning < aservice->minProcs)
853 /* check if over max quota */
854 if (aservice->nRequestsRunning >= aservice->maxProcs)
857 /* otherwise, can use only if there are enough to allow everyone
858 * to go to their min quota after this guy starts.
860 MUTEX_ENTER(&rx_quota_mutex);
861 if (rxi_availProcs > rxi_minDeficit)
863 MUTEX_EXIT(&rx_quota_mutex);
866 #endif /* RX_ENABLE_LOCKS */
869 /* Called by rx_StartServer to start up lwp's to service calls.
870 NExistingProcs gives the number of procs already existing, and which
871 therefore needn't be created. */
873 rxi_StartServerProcs(int nExistingProcs)
875 struct rx_service *service;
880 /* For each service, reserve N processes, where N is the "minimum"
881 * number of processes that MUST be able to execute a request in parallel,
882 * at any time, for that process. Also compute the maximum difference
883 * between any service's maximum number of processes that can run
884 * (i.e. the maximum number that ever will be run, and a guarantee
885 * that this number will run if other services aren't running), and its
886 * minimum number. The result is the extra number of processes that
887 * we need in order to provide the latter guarantee */
888 for (i = 0; i < RX_MAX_SERVICES; i++) {
890 service = rx_services[i];
891 if (service == (struct rx_service *)0)
893 nProcs += service->minProcs;
894 diff = service->maxProcs - service->minProcs;
898 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
899 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
900 for (i = 0; i < nProcs; i++) {
901 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
907 /* This routine is only required on Windows */
909 rx_StartClientThread(void)
911 #ifdef AFS_PTHREAD_ENV
913 pid = pthread_self();
914 #endif /* AFS_PTHREAD_ENV */
916 #endif /* AFS_NT40_ENV */
918 /* This routine must be called if any services are exported. If the
919 * donateMe flag is set, the calling process is donated to the server
922 rx_StartServer(int donateMe)
924 struct rx_service *service;
930 /* Start server processes, if necessary (exact function is dependent
931 * on the implementation environment--kernel or user space). DonateMe
932 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
933 * case, one less new proc will be created rx_StartServerProcs.
935 rxi_StartServerProcs(donateMe);
937 /* count up the # of threads in minProcs, and add set the min deficit to
938 * be that value, too.
940 for (i = 0; i < RX_MAX_SERVICES; i++) {
941 service = rx_services[i];
942 if (service == (struct rx_service *)0)
944 MUTEX_ENTER(&rx_quota_mutex);
945 rxi_totalMin += service->minProcs;
946 /* below works even if a thread is running, since minDeficit would
947 * still have been decremented and later re-incremented.
949 rxi_minDeficit += service->minProcs;
950 MUTEX_EXIT(&rx_quota_mutex);
953 /* Turn on reaping of idle server connections */
954 rxi_ReapConnections(NULL, NULL, NULL, 0);
963 #ifdef AFS_PTHREAD_ENV
965 pid = afs_pointer_to_int(pthread_self());
966 #else /* AFS_PTHREAD_ENV */
968 LWP_CurrentProcess(&pid);
969 #endif /* AFS_PTHREAD_ENV */
971 sprintf(name, "srv_%d", ++nProcs);
973 (*registerProgram) (pid, name);
975 #endif /* AFS_NT40_ENV */
976 rx_ServerProc(NULL); /* Never returns */
978 #ifdef RX_ENABLE_TSFPQ
979 /* no use leaving packets around in this thread's local queue if
980 * it isn't getting donated to the server thread pool.
982 rxi_FlushLocalPacketsTSFPQ();
983 #endif /* RX_ENABLE_TSFPQ */
987 /* Create a new client connection to the specified service, using the
988 * specified security object to implement the security model for this
990 struct rx_connection *
991 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
992 struct rx_securityClass *securityObject,
993 int serviceSecurityIndex)
997 struct rx_connection *conn;
1002 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
1003 "serviceSecurityIndex %d)\n",
1004 ntohl(shost), ntohs(sport), sservice, securityObject,
1005 serviceSecurityIndex));
1007 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
1008 * the case of kmem_alloc? */
1009 conn = rxi_AllocConnection();
1010 #ifdef RX_ENABLE_LOCKS
1011 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
1012 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
1013 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
1016 MUTEX_ENTER(&rx_connHashTable_lock);
1017 cid = (rx_nextCid += RX_MAXCALLS);
1018 conn->type = RX_CLIENT_CONNECTION;
1020 conn->epoch = rx_epoch;
1021 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
1022 conn->serviceId = sservice;
1023 conn->securityObject = securityObject;
1024 conn->securityData = (void *) 0;
1025 conn->securityIndex = serviceSecurityIndex;
1026 rx_SetConnDeadTime(conn, rx_connDeadTime);
1027 rx_SetConnSecondsUntilNatPing(conn, 0);
1028 conn->ackRate = RX_FAST_ACK_RATE;
1029 conn->nSpecific = 0;
1030 conn->specific = NULL;
1031 conn->challengeEvent = NULL;
1032 conn->delayedAbortEvent = NULL;
1033 conn->abortCount = 0;
1035 for (i = 0; i < RX_MAXCALLS; i++) {
1036 conn->twind[i] = rx_initSendWindow;
1037 conn->rwind[i] = rx_initReceiveWindow;
1038 conn->lastBusy[i] = 0;
1041 RXS_NewConnection(securityObject, conn);
1043 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1045 conn->refCount++; /* no lock required since only this thread knows... */
1046 conn->next = rx_connHashTable[hashindex];
1047 rx_connHashTable[hashindex] = conn;
1048 if (rx_stats_active)
1049 rx_atomic_inc(&rx_stats.nClientConns);
1050 MUTEX_EXIT(&rx_connHashTable_lock);
1056 * Ensure a connection's timeout values are valid.
1058 * @param[in] conn The connection to check
1060 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1061 * unless idleDeadTime and/or hardDeadTime are not set
1065 rxi_CheckConnTimeouts(struct rx_connection *conn)
1067 /* a connection's timeouts must have the relationship
1068 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1069 * total loss of network to a peer may cause an idle timeout instead of a
1070 * dead timeout, simply because the idle timeout gets hit first. Also set
1071 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1072 /* this logic is slightly complicated by the fact that
1073 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1075 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1076 if (conn->idleDeadTime) {
1077 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1079 if (conn->hardDeadTime) {
1080 if (conn->idleDeadTime) {
1081 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1083 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1089 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1091 /* The idea is to set the dead time to a value that allows several
1092 * keepalives to be dropped without timing out the connection. */
1093 conn->secondsUntilDead = seconds;
1094 rxi_CheckConnTimeouts(conn);
1095 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1099 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1101 conn->hardDeadTime = seconds;
1102 rxi_CheckConnTimeouts(conn);
1106 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1108 conn->idleDeadTime = seconds;
1109 rxi_CheckConnTimeouts(conn);
1112 int rxi_lowPeerRefCount = 0;
1113 int rxi_lowConnRefCount = 0;
1116 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1117 * NOTE: must not be called with rx_connHashTable_lock held.
1120 rxi_CleanupConnection(struct rx_connection *conn)
1122 /* Notify the service exporter, if requested, that this connection
1123 * is being destroyed */
1124 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1125 (*conn->service->destroyConnProc) (conn);
1127 /* Notify the security module that this connection is being destroyed */
1128 RXS_DestroyConnection(conn->securityObject, conn);
1130 /* If this is the last connection using the rx_peer struct, set its
1131 * idle time to now. rxi_ReapConnections will reap it if it's still
1132 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1134 MUTEX_ENTER(&rx_peerHashTable_lock);
1135 if (conn->peer->refCount < 2) {
1136 conn->peer->idleWhen = clock_Sec();
1137 if (conn->peer->refCount < 1) {
1138 conn->peer->refCount = 1;
1139 if (rx_stats_active) {
1140 MUTEX_ENTER(&rx_stats_mutex);
1141 rxi_lowPeerRefCount++;
1142 MUTEX_EXIT(&rx_stats_mutex);
1146 conn->peer->refCount--;
1147 MUTEX_EXIT(&rx_peerHashTable_lock);
1149 if (rx_stats_active)
1151 if (conn->type == RX_SERVER_CONNECTION)
1152 rx_atomic_dec(&rx_stats.nServerConns);
1154 rx_atomic_dec(&rx_stats.nClientConns);
1157 if (conn->specific) {
1159 for (i = 0; i < conn->nSpecific; i++) {
1160 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1161 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1162 conn->specific[i] = NULL;
1164 free(conn->specific);
1166 conn->specific = NULL;
1167 conn->nSpecific = 0;
1168 #endif /* !KERNEL */
1170 MUTEX_DESTROY(&conn->conn_call_lock);
1171 MUTEX_DESTROY(&conn->conn_data_lock);
1172 CV_DESTROY(&conn->conn_call_cv);
1174 rxi_FreeConnection(conn);
1177 /* Destroy the specified connection */
1179 rxi_DestroyConnection(struct rx_connection *conn)
1181 MUTEX_ENTER(&rx_connHashTable_lock);
1182 rxi_DestroyConnectionNoLock(conn);
1183 /* conn should be at the head of the cleanup list */
1184 if (conn == rx_connCleanup_list) {
1185 rx_connCleanup_list = rx_connCleanup_list->next;
1186 MUTEX_EXIT(&rx_connHashTable_lock);
1187 rxi_CleanupConnection(conn);
1189 #ifdef RX_ENABLE_LOCKS
1191 MUTEX_EXIT(&rx_connHashTable_lock);
1193 #endif /* RX_ENABLE_LOCKS */
1197 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1199 struct rx_connection **conn_ptr;
1201 struct rx_packet *packet;
1208 MUTEX_ENTER(&conn->conn_data_lock);
1209 MUTEX_ENTER(&rx_refcnt_mutex);
1210 if (conn->refCount > 0)
1213 if (rx_stats_active) {
1214 MUTEX_ENTER(&rx_stats_mutex);
1215 rxi_lowConnRefCount++;
1216 MUTEX_EXIT(&rx_stats_mutex);
1220 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1221 /* Busy; wait till the last guy before proceeding */
1222 MUTEX_EXIT(&rx_refcnt_mutex);
1223 MUTEX_EXIT(&conn->conn_data_lock);
1228 /* If the client previously called rx_NewCall, but it is still
1229 * waiting, treat this as a running call, and wait to destroy the
1230 * connection later when the call completes. */
1231 if ((conn->type == RX_CLIENT_CONNECTION)
1232 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1233 conn->flags |= RX_CONN_DESTROY_ME;
1234 MUTEX_EXIT(&conn->conn_data_lock);
1238 MUTEX_EXIT(&rx_refcnt_mutex);
1239 MUTEX_EXIT(&conn->conn_data_lock);
1241 /* Check for extant references to this connection */
1242 MUTEX_ENTER(&conn->conn_call_lock);
1243 for (i = 0; i < RX_MAXCALLS; i++) {
1244 struct rx_call *call = conn->call[i];
1247 if (conn->type == RX_CLIENT_CONNECTION) {
1248 MUTEX_ENTER(&call->lock);
1249 if (call->delayedAckEvent) {
1250 /* Push the final acknowledgment out now--there
1251 * won't be a subsequent call to acknowledge the
1252 * last reply packets */
1253 rxevent_Cancel(&call->delayedAckEvent, call,
1254 RX_CALL_REFCOUNT_DELAY);
1255 if (call->state == RX_STATE_PRECALL
1256 || call->state == RX_STATE_ACTIVE) {
1257 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1259 rxi_AckAll(NULL, call, 0);
1262 MUTEX_EXIT(&call->lock);
1266 MUTEX_EXIT(&conn->conn_call_lock);
1268 #ifdef RX_ENABLE_LOCKS
1270 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1271 MUTEX_EXIT(&conn->conn_data_lock);
1273 /* Someone is accessing a packet right now. */
1277 #endif /* RX_ENABLE_LOCKS */
1280 /* Don't destroy the connection if there are any call
1281 * structures still in use */
1282 MUTEX_ENTER(&conn->conn_data_lock);
1283 conn->flags |= RX_CONN_DESTROY_ME;
1284 MUTEX_EXIT(&conn->conn_data_lock);
1289 if (conn->natKeepAliveEvent) {
1290 rxi_NatKeepAliveOff(conn);
1293 if (conn->delayedAbortEvent) {
1294 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
1295 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1297 MUTEX_ENTER(&conn->conn_data_lock);
1298 rxi_SendConnectionAbort(conn, packet, 0, 1);
1299 MUTEX_EXIT(&conn->conn_data_lock);
1300 rxi_FreePacket(packet);
1304 /* Remove from connection hash table before proceeding */
1306 &rx_connHashTable[CONN_HASH
1307 (peer->host, peer->port, conn->cid, conn->epoch,
1309 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1310 if (*conn_ptr == conn) {
1311 *conn_ptr = conn->next;
1315 /* if the conn that we are destroying was the last connection, then we
1316 * clear rxLastConn as well */
1317 if (rxLastConn == conn)
1320 /* Make sure the connection is completely reset before deleting it. */
1321 /* get rid of pending events that could zap us later */
1322 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
1323 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
1324 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
1326 /* Add the connection to the list of destroyed connections that
1327 * need to be cleaned up. This is necessary to avoid deadlocks
1328 * in the routines we call to inform others that this connection is
1329 * being destroyed. */
1330 conn->next = rx_connCleanup_list;
1331 rx_connCleanup_list = conn;
1334 /* Externally available version */
1336 rx_DestroyConnection(struct rx_connection *conn)
1341 rxi_DestroyConnection(conn);
1346 rx_GetConnection(struct rx_connection *conn)
1351 MUTEX_ENTER(&rx_refcnt_mutex);
1353 MUTEX_EXIT(&rx_refcnt_mutex);
1357 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1358 /* Wait for the transmit queue to no longer be busy.
1359 * requires the call->lock to be held */
1361 rxi_WaitforTQBusy(struct rx_call *call) {
1362 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1363 call->flags |= RX_CALL_TQ_WAIT;
1365 #ifdef RX_ENABLE_LOCKS
1366 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1367 CV_WAIT(&call->cv_tq, &call->lock);
1368 #else /* RX_ENABLE_LOCKS */
1369 osi_rxSleep(&call->tq);
1370 #endif /* RX_ENABLE_LOCKS */
1372 if (call->tqWaiters == 0) {
1373 call->flags &= ~RX_CALL_TQ_WAIT;
1380 rxi_WakeUpTransmitQueue(struct rx_call *call)
1382 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1383 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1384 call, call->tqWaiters, call->flags));
1385 #ifdef RX_ENABLE_LOCKS
1386 osirx_AssertMine(&call->lock, "rxi_Start start");
1387 CV_BROADCAST(&call->cv_tq);
1388 #else /* RX_ENABLE_LOCKS */
1389 osi_rxWakeup(&call->tq);
1390 #endif /* RX_ENABLE_LOCKS */
1394 /* Start a new rx remote procedure call, on the specified connection.
1395 * If wait is set to 1, wait for a free call channel; otherwise return
1396 * 0. Maxtime gives the maximum number of seconds this call may take,
1397 * after rx_NewCall returns. After this time interval, a call to any
1398 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1399 * For fine grain locking, we hold the conn_call_lock in order to
1400 * to ensure that we don't get signalle after we found a call in an active
1401 * state and before we go to sleep.
1404 rx_NewCall(struct rx_connection *conn)
1406 int i, wait, ignoreBusy = 1;
1407 struct rx_call *call;
1408 struct clock queueTime;
1409 afs_uint32 leastBusy = 0;
1413 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1416 clock_GetTime(&queueTime);
1418 * Check if there are others waiting for a new call.
1419 * If so, let them go first to avoid starving them.
1420 * This is a fairly simple scheme, and might not be
1421 * a complete solution for large numbers of waiters.
1423 * makeCallWaiters keeps track of the number of
1424 * threads waiting to make calls and the
1425 * RX_CONN_MAKECALL_WAITING flag bit is used to
1426 * indicate that there are indeed calls waiting.
1427 * The flag is set when the waiter is incremented.
1428 * It is only cleared when makeCallWaiters is 0.
1429 * This prevents us from accidently destroying the
1430 * connection while it is potentially about to be used.
1432 MUTEX_ENTER(&conn->conn_call_lock);
1433 MUTEX_ENTER(&conn->conn_data_lock);
1434 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1435 conn->flags |= RX_CONN_MAKECALL_WAITING;
1436 conn->makeCallWaiters++;
1437 MUTEX_EXIT(&conn->conn_data_lock);
1439 #ifdef RX_ENABLE_LOCKS
1440 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1444 MUTEX_ENTER(&conn->conn_data_lock);
1445 conn->makeCallWaiters--;
1446 if (conn->makeCallWaiters == 0)
1447 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1450 /* We are now the active thread in rx_NewCall */
1451 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1452 MUTEX_EXIT(&conn->conn_data_lock);
1457 for (i = 0; i < RX_MAXCALLS; i++) {
1458 call = conn->call[i];
1460 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1461 /* we're not ignoring busy call slots; only look at the
1462 * call slot that is the "least" busy */
1466 if (call->state == RX_STATE_DALLY) {
1467 MUTEX_ENTER(&call->lock);
1468 if (call->state == RX_STATE_DALLY) {
1469 if (ignoreBusy && conn->lastBusy[i]) {
1470 /* if we're ignoring busy call slots, skip any ones that
1471 * have lastBusy set */
1472 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1473 leastBusy = conn->lastBusy[i];
1475 MUTEX_EXIT(&call->lock);
1480 * We are setting the state to RX_STATE_RESET to
1481 * ensure that no one else will attempt to use this
1482 * call once we drop the conn->conn_call_lock and
1483 * call->lock. We must drop the conn->conn_call_lock
1484 * before calling rxi_ResetCall because the process
1485 * of clearing the transmit queue can block for an
1486 * extended period of time. If we block while holding
1487 * the conn->conn_call_lock, then all rx_EndCall
1488 * processing will block as well. This has a detrimental
1489 * effect on overall system performance.
1491 call->state = RX_STATE_RESET;
1492 MUTEX_EXIT(&conn->conn_call_lock);
1493 MUTEX_ENTER(&rx_refcnt_mutex);
1494 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1495 MUTEX_EXIT(&rx_refcnt_mutex);
1496 rxi_ResetCall(call, 0);
1497 (*call->callNumber)++;
1498 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1502 * If we failed to be able to safely obtain the
1503 * conn->conn_call_lock we will have to drop the
1504 * call->lock to avoid a deadlock. When the call->lock
1505 * is released the state of the call can change. If it
1506 * is no longer RX_STATE_RESET then some other thread is
1509 MUTEX_EXIT(&call->lock);
1510 MUTEX_ENTER(&conn->conn_call_lock);
1511 MUTEX_ENTER(&call->lock);
1513 if (call->state == RX_STATE_RESET)
1517 * If we get here it means that after dropping
1518 * the conn->conn_call_lock and call->lock that
1519 * the call is no longer ours. If we can't find
1520 * a free call in the remaining slots we should
1521 * not go immediately to RX_CONN_MAKECALL_WAITING
1522 * because by dropping the conn->conn_call_lock
1523 * we have given up synchronization with rx_EndCall.
1524 * Instead, cycle through one more time to see if
1525 * we can find a call that can call our own.
1527 MUTEX_ENTER(&rx_refcnt_mutex);
1528 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1529 MUTEX_EXIT(&rx_refcnt_mutex);
1532 MUTEX_EXIT(&call->lock);
1535 if (ignoreBusy && conn->lastBusy[i]) {
1536 /* if we're ignoring busy call slots, skip any ones that
1537 * have lastBusy set */
1538 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1539 leastBusy = conn->lastBusy[i];
1544 /* rxi_NewCall returns with mutex locked */
1545 call = rxi_NewCall(conn, i);
1546 MUTEX_ENTER(&rx_refcnt_mutex);
1547 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1548 MUTEX_EXIT(&rx_refcnt_mutex);
1552 if (i < RX_MAXCALLS) {
1553 conn->lastBusy[i] = 0;
1558 if (leastBusy && ignoreBusy) {
1559 /* we didn't find a useable call slot, but we did see at least one
1560 * 'busy' slot; look again and only use a slot with the 'least
1566 MUTEX_ENTER(&conn->conn_data_lock);
1567 conn->flags |= RX_CONN_MAKECALL_WAITING;
1568 conn->makeCallWaiters++;
1569 MUTEX_EXIT(&conn->conn_data_lock);
1571 #ifdef RX_ENABLE_LOCKS
1572 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1576 MUTEX_ENTER(&conn->conn_data_lock);
1577 conn->makeCallWaiters--;
1578 if (conn->makeCallWaiters == 0)
1579 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1580 MUTEX_EXIT(&conn->conn_data_lock);
1582 /* Client is initially in send mode */
1583 call->state = RX_STATE_ACTIVE;
1584 call->error = conn->error;
1586 call->mode = RX_MODE_ERROR;
1588 call->mode = RX_MODE_SENDING;
1590 /* remember start time for call in case we have hard dead time limit */
1591 call->queueTime = queueTime;
1592 clock_GetTime(&call->startTime);
1593 hzero(call->bytesSent);
1594 hzero(call->bytesRcvd);
1596 /* Turn on busy protocol. */
1597 rxi_KeepAliveOn(call);
1599 /* Attempt MTU discovery */
1600 rxi_GrowMTUOn(call);
1603 * We are no longer the active thread in rx_NewCall
1605 MUTEX_ENTER(&conn->conn_data_lock);
1606 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1607 MUTEX_EXIT(&conn->conn_data_lock);
1610 * Wake up anyone else who might be giving us a chance to
1611 * run (see code above that avoids resource starvation).
1613 #ifdef RX_ENABLE_LOCKS
1614 CV_BROADCAST(&conn->conn_call_cv);
1618 MUTEX_EXIT(&conn->conn_call_lock);
1620 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1621 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1622 osi_Panic("rx_NewCall call about to be used without an empty tq");
1624 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1626 MUTEX_EXIT(&call->lock);
1629 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1634 rxi_HasActiveCalls(struct rx_connection *aconn)
1637 struct rx_call *tcall;
1641 for (i = 0; i < RX_MAXCALLS; i++) {
1642 if ((tcall = aconn->call[i])) {
1643 if ((tcall->state == RX_STATE_ACTIVE)
1644 || (tcall->state == RX_STATE_PRECALL)) {
1655 rxi_GetCallNumberVector(struct rx_connection *aconn,
1656 afs_int32 * aint32s)
1659 struct rx_call *tcall;
1663 for (i = 0; i < RX_MAXCALLS; i++) {
1664 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1665 aint32s[i] = aconn->callNumber[i] + 1;
1667 aint32s[i] = aconn->callNumber[i];
1674 rxi_SetCallNumberVector(struct rx_connection *aconn,
1675 afs_int32 * aint32s)
1678 struct rx_call *tcall;
1682 for (i = 0; i < RX_MAXCALLS; i++) {
1683 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1684 aconn->callNumber[i] = aint32s[i] - 1;
1686 aconn->callNumber[i] = aint32s[i];
1692 /* Advertise a new service. A service is named locally by a UDP port
1693 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1696 char *serviceName; Name for identification purposes (e.g. the
1697 service name might be used for probing for
1700 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1701 char *serviceName, struct rx_securityClass **securityObjects,
1702 int nSecurityObjects,
1703 afs_int32(*serviceProc) (struct rx_call * acall))
1705 osi_socket socket = OSI_NULLSOCKET;
1706 struct rx_service *tservice;
1712 if (serviceId == 0) {
1714 "rx_NewService: service id for service %s is not non-zero.\n",
1721 "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",
1729 tservice = rxi_AllocService();
1732 #ifdef RX_ENABLE_LOCKS
1733 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1736 for (i = 0; i < RX_MAX_SERVICES; i++) {
1737 struct rx_service *service = rx_services[i];
1739 if (port == service->servicePort && host == service->serviceHost) {
1740 if (service->serviceId == serviceId) {
1741 /* The identical service has already been
1742 * installed; if the caller was intending to
1743 * change the security classes used by this
1744 * service, he/she loses. */
1746 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1747 serviceName, serviceId, service->serviceName);
1749 rxi_FreeService(tservice);
1752 /* Different service, same port: re-use the socket
1753 * which is bound to the same port */
1754 socket = service->socket;
1757 if (socket == OSI_NULLSOCKET) {
1758 /* If we don't already have a socket (from another
1759 * service on same port) get a new one */
1760 socket = rxi_GetHostUDPSocket(host, port);
1761 if (socket == OSI_NULLSOCKET) {
1763 rxi_FreeService(tservice);
1768 service->socket = socket;
1769 service->serviceHost = host;
1770 service->servicePort = port;
1771 service->serviceId = serviceId;
1772 service->serviceName = serviceName;
1773 service->nSecurityObjects = nSecurityObjects;
1774 service->securityObjects = securityObjects;
1775 service->minProcs = 0;
1776 service->maxProcs = 1;
1777 service->idleDeadTime = 60;
1778 service->idleDeadErr = 0;
1779 service->connDeadTime = rx_connDeadTime;
1780 service->executeRequestProc = serviceProc;
1781 service->checkReach = 0;
1782 service->nSpecific = 0;
1783 service->specific = NULL;
1784 rx_services[i] = service; /* not visible until now */
1790 rxi_FreeService(tservice);
1791 (osi_Msg "rx_NewService: cannot support > %d services\n",
1796 /* Set configuration options for all of a service's security objects */
1799 rx_SetSecurityConfiguration(struct rx_service *service,
1800 rx_securityConfigVariables type,
1804 for (i = 0; i<service->nSecurityObjects; i++) {
1805 if (service->securityObjects[i]) {
1806 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1814 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1815 struct rx_securityClass **securityObjects, int nSecurityObjects,
1816 afs_int32(*serviceProc) (struct rx_call * acall))
1818 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1821 /* Generic request processing loop. This routine should be called
1822 * by the implementation dependent rx_ServerProc. If socketp is
1823 * non-null, it will be set to the file descriptor that this thread
1824 * is now listening on. If socketp is null, this routine will never
1827 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1829 struct rx_call *call;
1831 struct rx_service *tservice = NULL;
1838 call = rx_GetCall(threadID, tservice, socketp);
1839 if (socketp && *socketp != OSI_NULLSOCKET) {
1840 /* We are now a listener thread */
1846 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1847 #ifdef RX_ENABLE_LOCKS
1849 #endif /* RX_ENABLE_LOCKS */
1850 afs_termState = AFSOP_STOP_AFS;
1851 afs_osi_Wakeup(&afs_termState);
1852 #ifdef RX_ENABLE_LOCKS
1854 #endif /* RX_ENABLE_LOCKS */
1859 /* if server is restarting( typically smooth shutdown) then do not
1860 * allow any new calls.
1863 if (rx_tranquil && (call != NULL)) {
1867 MUTEX_ENTER(&call->lock);
1869 rxi_CallError(call, RX_RESTARTING);
1870 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1872 MUTEX_EXIT(&call->lock);
1877 tservice = call->conn->service;
1879 if (tservice->beforeProc)
1880 (*tservice->beforeProc) (call);
1882 code = tservice->executeRequestProc(call);
1884 if (tservice->afterProc)
1885 (*tservice->afterProc) (call, code);
1887 rx_EndCall(call, code);
1889 if (tservice->postProc)
1890 (*tservice->postProc) (code);
1892 if (rx_stats_active) {
1893 MUTEX_ENTER(&rx_stats_mutex);
1895 MUTEX_EXIT(&rx_stats_mutex);
1902 rx_WakeupServerProcs(void)
1904 struct rx_serverQueueEntry *np, *tqp;
1908 MUTEX_ENTER(&rx_serverPool_lock);
1910 #ifdef RX_ENABLE_LOCKS
1911 if (rx_waitForPacket)
1912 CV_BROADCAST(&rx_waitForPacket->cv);
1913 #else /* RX_ENABLE_LOCKS */
1914 if (rx_waitForPacket)
1915 osi_rxWakeup(rx_waitForPacket);
1916 #endif /* RX_ENABLE_LOCKS */
1917 MUTEX_ENTER(&freeSQEList_lock);
1918 for (np = rx_FreeSQEList; np; np = tqp) {
1919 tqp = *(struct rx_serverQueueEntry **)np;
1920 #ifdef RX_ENABLE_LOCKS
1921 CV_BROADCAST(&np->cv);
1922 #else /* RX_ENABLE_LOCKS */
1924 #endif /* RX_ENABLE_LOCKS */
1926 MUTEX_EXIT(&freeSQEList_lock);
1927 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1928 #ifdef RX_ENABLE_LOCKS
1929 CV_BROADCAST(&np->cv);
1930 #else /* RX_ENABLE_LOCKS */
1932 #endif /* RX_ENABLE_LOCKS */
1934 MUTEX_EXIT(&rx_serverPool_lock);
1939 * One thing that seems to happen is that all the server threads get
1940 * tied up on some empty or slow call, and then a whole bunch of calls
1941 * arrive at once, using up the packet pool, so now there are more
1942 * empty calls. The most critical resources here are server threads
1943 * and the free packet pool. The "doreclaim" code seems to help in
1944 * general. I think that eventually we arrive in this state: there
1945 * are lots of pending calls which do have all their packets present,
1946 * so they won't be reclaimed, are multi-packet calls, so they won't
1947 * be scheduled until later, and thus are tying up most of the free
1948 * packet pool for a very long time.
1950 * 1. schedule multi-packet calls if all the packets are present.
1951 * Probably CPU-bound operation, useful to return packets to pool.
1952 * Do what if there is a full window, but the last packet isn't here?
1953 * 3. preserve one thread which *only* runs "best" calls, otherwise
1954 * it sleeps and waits for that type of call.
1955 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1956 * the current dataquota business is badly broken. The quota isn't adjusted
1957 * to reflect how many packets are presently queued for a running call.
1958 * So, when we schedule a queued call with a full window of packets queued
1959 * up for it, that *should* free up a window full of packets for other 2d-class
1960 * calls to be able to use from the packet pool. But it doesn't.
1962 * NB. Most of the time, this code doesn't run -- since idle server threads
1963 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1964 * as a new call arrives.
1966 /* Sleep until a call arrives. Returns a pointer to the call, ready
1967 * for an rx_Read. */
1968 #ifdef RX_ENABLE_LOCKS
1970 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1972 struct rx_serverQueueEntry *sq;
1973 struct rx_call *call = (struct rx_call *)0;
1974 struct rx_service *service = NULL;
1976 MUTEX_ENTER(&freeSQEList_lock);
1978 if ((sq = rx_FreeSQEList)) {
1979 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1980 MUTEX_EXIT(&freeSQEList_lock);
1981 } else { /* otherwise allocate a new one and return that */
1982 MUTEX_EXIT(&freeSQEList_lock);
1983 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1984 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1985 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1988 MUTEX_ENTER(&rx_serverPool_lock);
1989 if (cur_service != NULL) {
1990 ReturnToServerPool(cur_service);
1993 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1994 struct rx_call *tcall, *ncall, *choice2 = NULL;
1996 /* Scan for eligible incoming calls. A call is not eligible
1997 * if the maximum number of calls for its service type are
1998 * already executing */
1999 /* One thread will process calls FCFS (to prevent starvation),
2000 * while the other threads may run ahead looking for calls which
2001 * have all their input data available immediately. This helps
2002 * keep threads from blocking, waiting for data from the client. */
2003 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
2004 service = tcall->conn->service;
2005 if (!QuotaOK(service)) {
2008 MUTEX_ENTER(&rx_pthread_mutex);
2009 if (tno == rxi_fcfs_thread_num
2010 || !tcall->queue_item_header.next) {
2011 MUTEX_EXIT(&rx_pthread_mutex);
2012 /* If we're the fcfs thread , then we'll just use
2013 * this call. If we haven't been able to find an optimal
2014 * choice, and we're at the end of the list, then use a
2015 * 2d choice if one has been identified. Otherwise... */
2016 call = (choice2 ? choice2 : tcall);
2017 service = call->conn->service;
2019 MUTEX_EXIT(&rx_pthread_mutex);
2020 if (!queue_IsEmpty(&tcall->rq)) {
2021 struct rx_packet *rp;
2022 rp = queue_First(&tcall->rq, rx_packet);
2023 if (rp->header.seq == 1) {
2025 || (rp->header.flags & RX_LAST_PACKET)) {
2027 } else if (rxi_2dchoice && !choice2
2028 && !(tcall->flags & RX_CALL_CLEARED)
2029 && (tcall->rprev > rxi_HardAckRate)) {
2039 ReturnToServerPool(service);
2046 MUTEX_EXIT(&rx_serverPool_lock);
2047 MUTEX_ENTER(&call->lock);
2049 if (call->flags & RX_CALL_WAIT_PROC) {
2050 call->flags &= ~RX_CALL_WAIT_PROC;
2051 rx_atomic_dec(&rx_nWaiting);
2054 if (call->state != RX_STATE_PRECALL || call->error) {
2055 MUTEX_EXIT(&call->lock);
2056 MUTEX_ENTER(&rx_serverPool_lock);
2057 ReturnToServerPool(service);
2062 if (queue_IsEmpty(&call->rq)
2063 || queue_First(&call->rq, rx_packet)->header.seq != 1)
2064 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2066 CLEAR_CALL_QUEUE_LOCK(call);
2069 /* If there are no eligible incoming calls, add this process
2070 * to the idle server queue, to wait for one */
2074 *socketp = OSI_NULLSOCKET;
2076 sq->socketp = socketp;
2077 queue_Append(&rx_idleServerQueue, sq);
2078 #ifndef AFS_AIX41_ENV
2079 rx_waitForPacket = sq;
2081 rx_waitingForPacket = sq;
2082 #endif /* AFS_AIX41_ENV */
2084 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2086 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2087 MUTEX_EXIT(&rx_serverPool_lock);
2088 return (struct rx_call *)0;
2091 } while (!(call = sq->newcall)
2092 && !(socketp && *socketp != OSI_NULLSOCKET));
2093 MUTEX_EXIT(&rx_serverPool_lock);
2095 MUTEX_ENTER(&call->lock);
2101 MUTEX_ENTER(&freeSQEList_lock);
2102 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2103 rx_FreeSQEList = sq;
2104 MUTEX_EXIT(&freeSQEList_lock);
2107 clock_GetTime(&call->startTime);
2108 call->state = RX_STATE_ACTIVE;
2109 call->mode = RX_MODE_RECEIVING;
2110 #ifdef RX_KERNEL_TRACE
2111 if (ICL_SETACTIVE(afs_iclSetp)) {
2112 int glockOwner = ISAFS_GLOCK();
2115 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2116 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2123 rxi_calltrace(RX_CALL_START, call);
2124 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2125 call->conn->service->servicePort, call->conn->service->serviceId,
2128 MUTEX_EXIT(&call->lock);
2129 MUTEX_ENTER(&rx_refcnt_mutex);
2130 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2131 MUTEX_EXIT(&rx_refcnt_mutex);
2133 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2138 #else /* RX_ENABLE_LOCKS */
2140 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2142 struct rx_serverQueueEntry *sq;
2143 struct rx_call *call = (struct rx_call *)0, *choice2;
2144 struct rx_service *service = NULL;
2148 MUTEX_ENTER(&freeSQEList_lock);
2150 if ((sq = rx_FreeSQEList)) {
2151 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2152 MUTEX_EXIT(&freeSQEList_lock);
2153 } else { /* otherwise allocate a new one and return that */
2154 MUTEX_EXIT(&freeSQEList_lock);
2155 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2156 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2157 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2159 MUTEX_ENTER(&sq->lock);
2161 if (cur_service != NULL) {
2162 cur_service->nRequestsRunning--;
2163 MUTEX_ENTER(&rx_quota_mutex);
2164 if (cur_service->nRequestsRunning < cur_service->minProcs)
2167 MUTEX_EXIT(&rx_quota_mutex);
2169 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
2170 struct rx_call *tcall, *ncall;
2171 /* Scan for eligible incoming calls. A call is not eligible
2172 * if the maximum number of calls for its service type are
2173 * already executing */
2174 /* One thread will process calls FCFS (to prevent starvation),
2175 * while the other threads may run ahead looking for calls which
2176 * have all their input data available immediately. This helps
2177 * keep threads from blocking, waiting for data from the client. */
2178 choice2 = (struct rx_call *)0;
2179 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
2180 service = tcall->conn->service;
2181 if (QuotaOK(service)) {
2182 MUTEX_ENTER(&rx_pthread_mutex);
2183 if (tno == rxi_fcfs_thread_num
2184 || !tcall->queue_item_header.next) {
2185 MUTEX_EXIT(&rx_pthread_mutex);
2186 /* If we're the fcfs thread, then we'll just use
2187 * this call. If we haven't been able to find an optimal
2188 * choice, and we're at the end of the list, then use a
2189 * 2d choice if one has been identified. Otherwise... */
2190 call = (choice2 ? choice2 : tcall);
2191 service = call->conn->service;
2193 MUTEX_EXIT(&rx_pthread_mutex);
2194 if (!queue_IsEmpty(&tcall->rq)) {
2195 struct rx_packet *rp;
2196 rp = queue_First(&tcall->rq, rx_packet);
2197 if (rp->header.seq == 1
2199 || (rp->header.flags & RX_LAST_PACKET))) {
2201 } else if (rxi_2dchoice && !choice2
2202 && !(tcall->flags & RX_CALL_CLEARED)
2203 && (tcall->rprev > rxi_HardAckRate)) {
2217 /* we can't schedule a call if there's no data!!! */
2218 /* send an ack if there's no data, if we're missing the
2219 * first packet, or we're missing something between first
2220 * and last -- there's a "hole" in the incoming data. */
2221 if (queue_IsEmpty(&call->rq)
2222 || queue_First(&call->rq, rx_packet)->header.seq != 1
2223 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
2224 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2226 call->flags &= (~RX_CALL_WAIT_PROC);
2227 service->nRequestsRunning++;
2228 /* just started call in minProcs pool, need fewer to maintain
2230 MUTEX_ENTER(&rx_quota_mutex);
2231 if (service->nRequestsRunning <= service->minProcs)
2234 MUTEX_EXIT(&rx_quota_mutex);
2235 rx_atomic_dec(&rx_nWaiting);
2236 /* MUTEX_EXIT(&call->lock); */
2238 /* If there are no eligible incoming calls, add this process
2239 * to the idle server queue, to wait for one */
2242 *socketp = OSI_NULLSOCKET;
2244 sq->socketp = socketp;
2245 queue_Append(&rx_idleServerQueue, sq);
2249 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2251 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2252 return (struct rx_call *)0;
2255 } while (!(call = sq->newcall)
2256 && !(socketp && *socketp != OSI_NULLSOCKET));
2258 MUTEX_EXIT(&sq->lock);
2260 MUTEX_ENTER(&freeSQEList_lock);
2261 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2262 rx_FreeSQEList = sq;
2263 MUTEX_EXIT(&freeSQEList_lock);
2266 clock_GetTime(&call->startTime);
2267 call->state = RX_STATE_ACTIVE;
2268 call->mode = RX_MODE_RECEIVING;
2269 #ifdef RX_KERNEL_TRACE
2270 if (ICL_SETACTIVE(afs_iclSetp)) {
2271 int glockOwner = ISAFS_GLOCK();
2274 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2275 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2282 rxi_calltrace(RX_CALL_START, call);
2283 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2284 call->conn->service->servicePort, call->conn->service->serviceId,
2287 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2294 #endif /* RX_ENABLE_LOCKS */
2298 /* Establish a procedure to be called when a packet arrives for a
2299 * call. This routine will be called at most once after each call,
2300 * and will also be called if there is an error condition on the or
2301 * the call is complete. Used by multi rx to build a selection
2302 * function which determines which of several calls is likely to be a
2303 * good one to read from.
2304 * NOTE: the way this is currently implemented it is probably only a
2305 * good idea to (1) use it immediately after a newcall (clients only)
2306 * and (2) only use it once. Other uses currently void your warranty
2309 rx_SetArrivalProc(struct rx_call *call,
2310 void (*proc) (struct rx_call * call,
2313 void * handle, int arg)
2315 call->arrivalProc = proc;
2316 call->arrivalProcHandle = handle;
2317 call->arrivalProcArg = arg;
2320 /* Call is finished (possibly prematurely). Return rc to the peer, if
2321 * appropriate, and return the final error code from the conversation
2325 rx_EndCall(struct rx_call *call, afs_int32 rc)
2327 struct rx_connection *conn = call->conn;
2331 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2332 call, rc, call->error, call->abortCode));
2335 MUTEX_ENTER(&call->lock);
2337 if (rc == 0 && call->error == 0) {
2338 call->abortCode = 0;
2339 call->abortCount = 0;
2342 call->arrivalProc = (void (*)())0;
2343 if (rc && call->error == 0) {
2344 rxi_CallError(call, rc);
2345 call->mode = RX_MODE_ERROR;
2346 /* Send an abort message to the peer if this error code has
2347 * only just been set. If it was set previously, assume the
2348 * peer has already been sent the error code or will request it
2350 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2352 if (conn->type == RX_SERVER_CONNECTION) {
2353 /* Make sure reply or at least dummy reply is sent */
2354 if (call->mode == RX_MODE_RECEIVING) {
2355 MUTEX_EXIT(&call->lock);
2356 rxi_WriteProc(call, 0, 0);
2357 MUTEX_ENTER(&call->lock);
2359 if (call->mode == RX_MODE_SENDING) {
2360 MUTEX_EXIT(&call->lock);
2361 rxi_FlushWrite(call);
2362 MUTEX_ENTER(&call->lock);
2364 rxi_calltrace(RX_CALL_END, call);
2365 /* Call goes to hold state until reply packets are acknowledged */
2366 if (call->tfirst + call->nSoftAcked < call->tnext) {
2367 call->state = RX_STATE_HOLD;
2369 call->state = RX_STATE_DALLY;
2370 rxi_ClearTransmitQueue(call, 0);
2371 rxi_rto_cancel(call);
2372 rxevent_Cancel(&call->keepAliveEvent, call,
2373 RX_CALL_REFCOUNT_ALIVE);
2375 } else { /* Client connection */
2377 /* Make sure server receives input packets, in the case where
2378 * no reply arguments are expected */
2379 if ((call->mode == RX_MODE_SENDING)
2380 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2381 MUTEX_EXIT(&call->lock);
2382 (void)rxi_ReadProc(call, &dummy, 1);
2383 MUTEX_ENTER(&call->lock);
2386 /* If we had an outstanding delayed ack, be nice to the server
2387 * and force-send it now.
2389 if (call->delayedAckEvent) {
2390 rxevent_Cancel(&call->delayedAckEvent, call,
2391 RX_CALL_REFCOUNT_DELAY);
2392 rxi_SendDelayedAck(NULL, call, NULL, 0);
2395 /* We need to release the call lock since it's lower than the
2396 * conn_call_lock and we don't want to hold the conn_call_lock
2397 * over the rx_ReadProc call. The conn_call_lock needs to be held
2398 * here for the case where rx_NewCall is perusing the calls on
2399 * the connection structure. We don't want to signal until
2400 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2401 * have checked this call, found it active and by the time it
2402 * goes to sleep, will have missed the signal.
2404 MUTEX_EXIT(&call->lock);
2405 MUTEX_ENTER(&conn->conn_call_lock);
2406 MUTEX_ENTER(&call->lock);
2408 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2409 conn->lastBusy[call->channel] = 0;
2412 MUTEX_ENTER(&conn->conn_data_lock);
2413 conn->flags |= RX_CONN_BUSY;
2414 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2415 MUTEX_EXIT(&conn->conn_data_lock);
2416 #ifdef RX_ENABLE_LOCKS
2417 CV_BROADCAST(&conn->conn_call_cv);
2422 #ifdef RX_ENABLE_LOCKS
2424 MUTEX_EXIT(&conn->conn_data_lock);
2426 #endif /* RX_ENABLE_LOCKS */
2427 call->state = RX_STATE_DALLY;
2429 error = call->error;
2431 /* currentPacket, nLeft, and NFree must be zeroed here, because
2432 * ResetCall cannot: ResetCall may be called at splnet(), in the
2433 * kernel version, and may interrupt the macros rx_Read or
2434 * rx_Write, which run at normal priority for efficiency. */
2435 if (call->currentPacket) {
2436 #ifdef RX_TRACK_PACKETS
2437 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2439 rxi_FreePacket(call->currentPacket);
2440 call->currentPacket = (struct rx_packet *)0;
2443 call->nLeft = call->nFree = call->curlen = 0;
2445 /* Free any packets from the last call to ReadvProc/WritevProc */
2446 #ifdef RXDEBUG_PACKET
2448 #endif /* RXDEBUG_PACKET */
2449 rxi_FreePackets(0, &call->iovq);
2450 MUTEX_EXIT(&call->lock);
2452 MUTEX_ENTER(&rx_refcnt_mutex);
2453 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2454 MUTEX_EXIT(&rx_refcnt_mutex);
2455 if (conn->type == RX_CLIENT_CONNECTION) {
2456 MUTEX_ENTER(&conn->conn_data_lock);
2457 conn->flags &= ~RX_CONN_BUSY;
2458 MUTEX_EXIT(&conn->conn_data_lock);
2459 MUTEX_EXIT(&conn->conn_call_lock);
2463 * Map errors to the local host's errno.h format.
2465 error = ntoh_syserr_conv(error);
2469 #if !defined(KERNEL)
2471 /* Call this routine when shutting down a server or client (especially
2472 * clients). This will allow Rx to gracefully garbage collect server
2473 * connections, and reduce the number of retries that a server might
2474 * make to a dead client.
2475 * This is not quite right, since some calls may still be ongoing and
2476 * we can't lock them to destroy them. */
2480 struct rx_connection **conn_ptr, **conn_end;
2484 if (rxinit_status == 1) {
2486 return; /* Already shutdown. */
2488 rxi_DeleteCachedConnections();
2489 if (rx_connHashTable) {
2490 MUTEX_ENTER(&rx_connHashTable_lock);
2491 for (conn_ptr = &rx_connHashTable[0], conn_end =
2492 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2494 struct rx_connection *conn, *next;
2495 for (conn = *conn_ptr; conn; conn = next) {
2497 if (conn->type == RX_CLIENT_CONNECTION) {
2498 MUTEX_ENTER(&rx_refcnt_mutex);
2500 MUTEX_EXIT(&rx_refcnt_mutex);
2501 #ifdef RX_ENABLE_LOCKS
2502 rxi_DestroyConnectionNoLock(conn);
2503 #else /* RX_ENABLE_LOCKS */
2504 rxi_DestroyConnection(conn);
2505 #endif /* RX_ENABLE_LOCKS */
2509 #ifdef RX_ENABLE_LOCKS
2510 while (rx_connCleanup_list) {
2511 struct rx_connection *conn;
2512 conn = rx_connCleanup_list;
2513 rx_connCleanup_list = rx_connCleanup_list->next;
2514 MUTEX_EXIT(&rx_connHashTable_lock);
2515 rxi_CleanupConnection(conn);
2516 MUTEX_ENTER(&rx_connHashTable_lock);
2518 MUTEX_EXIT(&rx_connHashTable_lock);
2519 #endif /* RX_ENABLE_LOCKS */
2524 afs_winsockCleanup();
2532 /* if we wakeup packet waiter too often, can get in loop with two
2533 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2535 rxi_PacketsUnWait(void)
2537 if (!rx_waitingForPackets) {
2541 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2542 return; /* still over quota */
2545 rx_waitingForPackets = 0;
2546 #ifdef RX_ENABLE_LOCKS
2547 CV_BROADCAST(&rx_waitingForPackets_cv);
2549 osi_rxWakeup(&rx_waitingForPackets);
2555 /* ------------------Internal interfaces------------------------- */
2557 /* Return this process's service structure for the
2558 * specified socket and service */
2559 static struct rx_service *
2560 rxi_FindService(osi_socket socket, u_short serviceId)
2562 struct rx_service **sp;
2563 for (sp = &rx_services[0]; *sp; sp++) {
2564 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2570 #ifdef RXDEBUG_PACKET
2571 #ifdef KDUMP_RX_LOCK
2572 static struct rx_call_rx_lock *rx_allCallsp = 0;
2574 static struct rx_call *rx_allCallsp = 0;
2576 #endif /* RXDEBUG_PACKET */
2578 /* Allocate a call structure, for the indicated channel of the
2579 * supplied connection. The mode and state of the call must be set by
2580 * the caller. Returns the call with mutex locked. */
2581 static struct rx_call *
2582 rxi_NewCall(struct rx_connection *conn, int channel)
2584 struct rx_call *call;
2585 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2586 struct rx_call *cp; /* Call pointer temp */
2587 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2588 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2590 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2592 /* Grab an existing call structure, or allocate a new one.
2593 * Existing call structures are assumed to have been left reset by
2595 MUTEX_ENTER(&rx_freeCallQueue_lock);
2597 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2599 * EXCEPT that the TQ might not yet be cleared out.
2600 * Skip over those with in-use TQs.
2603 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2604 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2610 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2611 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2612 call = queue_First(&rx_freeCallQueue, rx_call);
2613 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2615 if (rx_stats_active)
2616 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2617 MUTEX_EXIT(&rx_freeCallQueue_lock);
2618 MUTEX_ENTER(&call->lock);
2619 CLEAR_CALL_QUEUE_LOCK(call);
2620 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2621 /* Now, if TQ wasn't cleared earlier, do it now. */
2622 rxi_WaitforTQBusy(call);
2623 if (call->flags & RX_CALL_TQ_CLEARME) {
2624 rxi_ClearTransmitQueue(call, 1);
2625 /*queue_Init(&call->tq);*/
2627 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2628 /* Bind the call to its connection structure */
2630 rxi_ResetCall(call, 1);
2633 call = rxi_Alloc(sizeof(struct rx_call));
2634 #ifdef RXDEBUG_PACKET
2635 call->allNextp = rx_allCallsp;
2636 rx_allCallsp = call;
2638 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2639 #else /* RXDEBUG_PACKET */
2640 rx_atomic_inc(&rx_stats.nCallStructs);
2641 #endif /* RXDEBUG_PACKET */
2643 MUTEX_EXIT(&rx_freeCallQueue_lock);
2644 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2645 MUTEX_ENTER(&call->lock);
2646 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2647 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2648 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2650 /* Initialize once-only items */
2651 queue_Init(&call->tq);
2652 queue_Init(&call->rq);
2653 queue_Init(&call->iovq);
2654 #ifdef RXDEBUG_PACKET
2655 call->rqc = call->tqc = call->iovqc = 0;
2656 #endif /* RXDEBUG_PACKET */
2657 /* Bind the call to its connection structure (prereq for reset) */
2659 rxi_ResetCall(call, 1);
2661 call->channel = channel;
2662 call->callNumber = &conn->callNumber[channel];
2663 call->rwind = conn->rwind[channel];
2664 call->twind = conn->twind[channel];
2665 /* Note that the next expected call number is retained (in
2666 * conn->callNumber[i]), even if we reallocate the call structure
2668 conn->call[channel] = call;
2669 /* if the channel's never been used (== 0), we should start at 1, otherwise
2670 * the call number is valid from the last time this channel was used */
2671 if (*call->callNumber == 0)
2672 *call->callNumber = 1;
2677 /* A call has been inactive long enough that so we can throw away
2678 * state, including the call structure, which is placed on the call
2681 * call->lock amd rx_refcnt_mutex are held upon entry.
2682 * haveCTLock is set when called from rxi_ReapConnections.
2685 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2687 int channel = call->channel;
2688 struct rx_connection *conn = call->conn;
2691 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2692 (*call->callNumber)++;
2694 * We are setting the state to RX_STATE_RESET to
2695 * ensure that no one else will attempt to use this
2696 * call once we drop the refcnt lock. We must drop
2697 * the refcnt lock before calling rxi_ResetCall
2698 * because it cannot be held across acquiring the
2699 * freepktQ lock. NewCall does the same.
2701 call->state = RX_STATE_RESET;
2702 MUTEX_EXIT(&rx_refcnt_mutex);
2703 rxi_ResetCall(call, 0);
2705 MUTEX_ENTER(&conn->conn_call_lock);
2706 if (call->conn->call[channel] == call)
2707 call->conn->call[channel] = 0;
2708 MUTEX_EXIT(&conn->conn_call_lock);
2710 MUTEX_ENTER(&rx_freeCallQueue_lock);
2711 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2712 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2713 /* A call may be free even though its transmit queue is still in use.
2714 * Since we search the call list from head to tail, put busy calls at
2715 * the head of the list, and idle calls at the tail.
2717 if (call->flags & RX_CALL_TQ_BUSY)
2718 queue_Prepend(&rx_freeCallQueue, call);
2720 queue_Append(&rx_freeCallQueue, call);
2721 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2722 queue_Append(&rx_freeCallQueue, call);
2723 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2724 if (rx_stats_active)
2725 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2726 MUTEX_EXIT(&rx_freeCallQueue_lock);
2728 /* Destroy the connection if it was previously slated for
2729 * destruction, i.e. the Rx client code previously called
2730 * rx_DestroyConnection (client connections), or
2731 * rxi_ReapConnections called the same routine (server
2732 * connections). Only do this, however, if there are no
2733 * outstanding calls. Note that for fine grain locking, there appears
2734 * to be a deadlock in that rxi_FreeCall has a call locked and
2735 * DestroyConnectionNoLock locks each call in the conn. But note a
2736 * few lines up where we have removed this call from the conn.
2737 * If someone else destroys a connection, they either have no
2738 * call lock held or are going through this section of code.
2740 MUTEX_ENTER(&conn->conn_data_lock);
2741 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2742 MUTEX_ENTER(&rx_refcnt_mutex);
2744 MUTEX_EXIT(&rx_refcnt_mutex);
2745 MUTEX_EXIT(&conn->conn_data_lock);
2746 #ifdef RX_ENABLE_LOCKS
2748 rxi_DestroyConnectionNoLock(conn);
2750 rxi_DestroyConnection(conn);
2751 #else /* RX_ENABLE_LOCKS */
2752 rxi_DestroyConnection(conn);
2753 #endif /* RX_ENABLE_LOCKS */
2755 MUTEX_EXIT(&conn->conn_data_lock);
2757 MUTEX_ENTER(&rx_refcnt_mutex);
2760 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2761 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2764 rxi_Alloc(size_t size)
2768 if (rx_stats_active) {
2769 rx_atomic_add(&rxi_Allocsize, (int) size);
2770 rx_atomic_inc(&rxi_Alloccnt);
2774 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2775 afs_osi_Alloc_NoSleep(size);
2780 osi_Panic("rxi_Alloc error");
2786 rxi_Free(void *addr, size_t size)
2788 if (rx_stats_active) {
2789 rx_atomic_sub(&rxi_Allocsize, (int) size);
2790 rx_atomic_dec(&rxi_Alloccnt);
2792 osi_Free(addr, size);
2796 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2798 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2799 struct rx_peer *next = NULL;
2803 MUTEX_ENTER(&rx_peerHashTable_lock);
2805 peer_ptr = &rx_peerHashTable[0];
2806 peer_end = &rx_peerHashTable[rx_hashTableSize];
2809 for ( ; peer_ptr < peer_end; peer_ptr++) {
2812 for ( ; peer; peer = next) {
2814 if (host == peer->host)
2819 hashIndex = PEER_HASH(host, port);
2820 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2821 if ((peer->host == host) && (peer->port == port))
2826 MUTEX_ENTER(&rx_peerHashTable_lock);
2831 MUTEX_EXIT(&rx_peerHashTable_lock);
2833 MUTEX_ENTER(&peer->peer_lock);
2834 /* We don't handle dropping below min, so don't */
2835 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2836 peer->ifMTU=MIN(mtu, peer->ifMTU);
2837 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2838 /* if we tweaked this down, need to tune our peer MTU too */
2839 peer->MTU = MIN(peer->MTU, peer->natMTU);
2840 /* if we discovered a sub-1500 mtu, degrade */
2841 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2842 peer->maxDgramPackets = 1;
2843 /* We no longer have valid peer packet information */
2844 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2845 peer->maxPacketSize = 0;
2846 MUTEX_EXIT(&peer->peer_lock);
2848 MUTEX_ENTER(&rx_peerHashTable_lock);
2850 if (host && !port) {
2852 /* pick up where we left off */
2856 MUTEX_EXIT(&rx_peerHashTable_lock);
2859 /* Find the peer process represented by the supplied (host,port)
2860 * combination. If there is no appropriate active peer structure, a
2861 * new one will be allocated and initialized
2862 * The origPeer, if set, is a pointer to a peer structure on which the
2863 * refcount will be be decremented. This is used to replace the peer
2864 * structure hanging off a connection structure */
2866 rxi_FindPeer(afs_uint32 host, u_short port,
2867 struct rx_peer *origPeer, int create)
2871 hashIndex = PEER_HASH(host, port);
2872 MUTEX_ENTER(&rx_peerHashTable_lock);
2873 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2874 if ((pp->host == host) && (pp->port == port))
2879 pp = rxi_AllocPeer(); /* This bzero's *pp */
2880 pp->host = host; /* set here or in InitPeerParams is zero */
2882 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2883 queue_Init(&pp->congestionQueue);
2884 queue_Init(&pp->rpcStats);
2885 pp->next = rx_peerHashTable[hashIndex];
2886 rx_peerHashTable[hashIndex] = pp;
2887 rxi_InitPeerParams(pp);
2888 if (rx_stats_active)
2889 rx_atomic_inc(&rx_stats.nPeerStructs);
2896 origPeer->refCount--;
2897 MUTEX_EXIT(&rx_peerHashTable_lock);
2902 /* Find the connection at (host, port) started at epoch, and with the
2903 * given connection id. Creates the server connection if necessary.
2904 * The type specifies whether a client connection or a server
2905 * connection is desired. In both cases, (host, port) specify the
2906 * peer's (host, pair) pair. Client connections are not made
2907 * automatically by this routine. The parameter socket gives the
2908 * socket descriptor on which the packet was received. This is used,
2909 * in the case of server connections, to check that *new* connections
2910 * come via a valid (port, serviceId). Finally, the securityIndex
2911 * parameter must match the existing index for the connection. If a
2912 * server connection is created, it will be created using the supplied
2913 * index, if the index is valid for this service */
2914 struct rx_connection *
2915 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2916 u_short port, u_short serviceId, afs_uint32 cid,
2917 afs_uint32 epoch, int type, u_int securityIndex)
2919 int hashindex, flag, i;
2920 struct rx_connection *conn;
2921 hashindex = CONN_HASH(host, port, cid, epoch, type);
2922 MUTEX_ENTER(&rx_connHashTable_lock);
2923 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2924 rx_connHashTable[hashindex],
2927 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2928 && (epoch == conn->epoch)) {
2929 struct rx_peer *pp = conn->peer;
2930 if (securityIndex != conn->securityIndex) {
2931 /* this isn't supposed to happen, but someone could forge a packet
2932 * like this, and there seems to be some CM bug that makes this
2933 * happen from time to time -- in which case, the fileserver
2935 MUTEX_EXIT(&rx_connHashTable_lock);
2936 return (struct rx_connection *)0;
2938 if (pp->host == host && pp->port == port)
2940 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2942 /* So what happens when it's a callback connection? */
2943 if ( /*type == RX_CLIENT_CONNECTION && */
2944 (conn->epoch & 0x80000000))
2948 /* the connection rxLastConn that was used the last time is not the
2949 ** one we are looking for now. Hence, start searching in the hash */
2951 conn = rx_connHashTable[hashindex];
2956 struct rx_service *service;
2957 if (type == RX_CLIENT_CONNECTION) {
2958 MUTEX_EXIT(&rx_connHashTable_lock);
2959 return (struct rx_connection *)0;
2961 service = rxi_FindService(socket, serviceId);
2962 if (!service || (securityIndex >= service->nSecurityObjects)
2963 || (service->securityObjects[securityIndex] == 0)) {
2964 MUTEX_EXIT(&rx_connHashTable_lock);
2965 return (struct rx_connection *)0;
2967 conn = rxi_AllocConnection(); /* This bzero's the connection */
2968 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2969 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2970 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2971 conn->next = rx_connHashTable[hashindex];
2972 rx_connHashTable[hashindex] = conn;
2973 conn->peer = rxi_FindPeer(host, port, 0, 1);
2974 conn->type = RX_SERVER_CONNECTION;
2975 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2976 conn->epoch = epoch;
2977 conn->cid = cid & RX_CIDMASK;
2978 /* conn->serial = conn->lastSerial = 0; */
2979 /* conn->timeout = 0; */
2980 conn->ackRate = RX_FAST_ACK_RATE;
2981 conn->service = service;
2982 conn->serviceId = serviceId;
2983 conn->securityIndex = securityIndex;
2984 conn->securityObject = service->securityObjects[securityIndex];
2985 conn->nSpecific = 0;
2986 conn->specific = NULL;
2987 rx_SetConnDeadTime(conn, service->connDeadTime);
2988 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2989 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2990 for (i = 0; i < RX_MAXCALLS; i++) {
2991 conn->twind[i] = rx_initSendWindow;
2992 conn->rwind[i] = rx_initReceiveWindow;
2994 /* Notify security object of the new connection */
2995 RXS_NewConnection(conn->securityObject, conn);
2996 /* XXXX Connection timeout? */
2997 if (service->newConnProc)
2998 (*service->newConnProc) (conn);
2999 if (rx_stats_active)
3000 rx_atomic_inc(&rx_stats.nServerConns);
3003 MUTEX_ENTER(&rx_refcnt_mutex);
3005 MUTEX_EXIT(&rx_refcnt_mutex);
3007 rxLastConn = conn; /* store this connection as the last conn used */
3008 MUTEX_EXIT(&rx_connHashTable_lock);
3013 * Timeout a call on a busy call channel if appropriate.
3015 * @param[in] call The busy call.
3017 * @pre 'call' is marked as busy (namely,
3018 * call->conn->lastBusy[call->channel] != 0)
3020 * @pre call->lock is held
3021 * @pre rxi_busyChannelError is nonzero
3023 * @note call->lock is dropped and reacquired
3026 rxi_CheckBusy(struct rx_call *call)
3028 struct rx_connection *conn = call->conn;
3029 int channel = call->channel;
3030 int freechannel = 0;
3032 afs_uint32 callNumber = *call->callNumber;
3034 MUTEX_EXIT(&call->lock);
3036 MUTEX_ENTER(&conn->conn_call_lock);
3038 /* Are there any other call slots on this conn that we should try? Look for
3039 * slots that are empty and are either non-busy, or were marked as busy
3040 * longer than conn->secondsUntilDead seconds before this call started. */
3042 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3044 /* only look at channels that aren't us */
3048 if (conn->lastBusy[i]) {
3049 /* if this channel looked busy too recently, don't look at it */
3050 if (conn->lastBusy[i] >= call->startTime.sec) {
3053 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3058 if (conn->call[i]) {
3059 struct rx_call *tcall = conn->call[i];
3060 MUTEX_ENTER(&tcall->lock);
3061 if (tcall->state == RX_STATE_DALLY) {
3064 MUTEX_EXIT(&tcall->lock);
3070 MUTEX_EXIT(&conn->conn_call_lock);
3072 MUTEX_ENTER(&call->lock);
3074 /* Since the call->lock and conn->conn_call_lock have been released it is
3075 * possible that (1) the call may no longer be busy and/or (2) the call may
3076 * have been reused by another waiting thread. Therefore, we must confirm
3077 * that the call state has not changed when deciding whether or not to
3078 * force this application thread to retry by forcing a Timeout error. */
3080 if (freechannel && *call->callNumber == callNumber &&
3081 (call->flags & RX_CALL_PEER_BUSY)) {
3082 /* Since 'freechannel' is set, there exists another channel in this
3083 * rx_conn that the application thread might be able to use. We know
3084 * that we have the correct call since callNumber is unchanged, and we
3085 * know that the call is still busy. So, set the call error state to
3086 * rxi_busyChannelError so the application can retry the request,
3087 * presumably on a less-busy call channel. */
3089 rxi_CallError(call, rxi_busyChannelError);
3093 /* There are two packet tracing routines available for testing and monitoring
3094 * Rx. One is called just after every packet is received and the other is
3095 * called just before every packet is sent. Received packets, have had their
3096 * headers decoded, and packets to be sent have not yet had their headers
3097 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3098 * containing the network address. Both can be modified. The return value, if
3099 * non-zero, indicates that the packet should be dropped. */
3101 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3102 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3104 /* A packet has been received off the interface. Np is the packet, socket is
3105 * the socket number it was received from (useful in determining which service
3106 * this packet corresponds to), and (host, port) reflect the host,port of the
3107 * sender. This call returns the packet to the caller if it is finished with
3108 * it, rather than de-allocating it, just as a small performance hack */
3111 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3112 afs_uint32 host, u_short port, int *tnop,
3113 struct rx_call **newcallp)
3115 struct rx_call *call;
3116 struct rx_connection *conn;
3118 afs_uint32 currentCallNumber;
3124 struct rx_packet *tnp;
3127 /* We don't print out the packet until now because (1) the time may not be
3128 * accurate enough until now in the lwp implementation (rx_Listener only gets
3129 * the time after the packet is read) and (2) from a protocol point of view,
3130 * this is the first time the packet has been seen */
3131 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3132 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3133 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3134 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3135 np->header.epoch, np->header.cid, np->header.callNumber,
3136 np->header.seq, np->header.flags, np));
3139 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3140 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3143 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3144 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3147 /* If an input tracer function is defined, call it with the packet and
3148 * network address. Note this function may modify its arguments. */
3149 if (rx_justReceived) {
3150 struct sockaddr_in addr;
3152 addr.sin_family = AF_INET;
3153 addr.sin_port = port;
3154 addr.sin_addr.s_addr = host;
3155 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3156 addr.sin_len = sizeof(addr);
3157 #endif /* AFS_OSF_ENV */
3158 drop = (*rx_justReceived) (np, &addr);
3159 /* drop packet if return value is non-zero */
3162 port = addr.sin_port; /* in case fcn changed addr */
3163 host = addr.sin_addr.s_addr;
3167 /* If packet was not sent by the client, then *we* must be the client */
3168 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3169 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3171 /* Find the connection (or fabricate one, if we're the server & if
3172 * necessary) associated with this packet */
3174 rxi_FindConnection(socket, host, port, np->header.serviceId,
3175 np->header.cid, np->header.epoch, type,
3176 np->header.securityIndex);
3179 /* If no connection found or fabricated, just ignore the packet.
3180 * (An argument could be made for sending an abort packet for
3185 /* If the connection is in an error state, send an abort packet and ignore
3186 * the incoming packet */
3188 /* Don't respond to an abort packet--we don't want loops! */
3189 MUTEX_ENTER(&conn->conn_data_lock);
3190 if (np->header.type != RX_PACKET_TYPE_ABORT)
3191 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3192 MUTEX_ENTER(&rx_refcnt_mutex);
3194 MUTEX_EXIT(&rx_refcnt_mutex);
3195 MUTEX_EXIT(&conn->conn_data_lock);
3199 /* Check for connection-only requests (i.e. not call specific). */
3200 if (np->header.callNumber == 0) {
3201 switch (np->header.type) {
3202 case RX_PACKET_TYPE_ABORT: {
3203 /* What if the supplied error is zero? */
3204 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3205 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3206 rxi_ConnectionError(conn, errcode);
3207 MUTEX_ENTER(&rx_refcnt_mutex);
3209 MUTEX_EXIT(&rx_refcnt_mutex);
3212 case RX_PACKET_TYPE_CHALLENGE:
3213 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3214 MUTEX_ENTER(&rx_refcnt_mutex);
3216 MUTEX_EXIT(&rx_refcnt_mutex);
3218 case RX_PACKET_TYPE_RESPONSE:
3219 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3220 MUTEX_ENTER(&rx_refcnt_mutex);
3222 MUTEX_EXIT(&rx_refcnt_mutex);
3224 case RX_PACKET_TYPE_PARAMS:
3225 case RX_PACKET_TYPE_PARAMS + 1:
3226 case RX_PACKET_TYPE_PARAMS + 2:
3227 /* ignore these packet types for now */
3228 MUTEX_ENTER(&rx_refcnt_mutex);
3230 MUTEX_EXIT(&rx_refcnt_mutex);
3235 /* Should not reach here, unless the peer is broken: send an
3237 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3238 MUTEX_ENTER(&conn->conn_data_lock);
3239 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3240 MUTEX_ENTER(&rx_refcnt_mutex);
3242 MUTEX_EXIT(&rx_refcnt_mutex);
3243 MUTEX_EXIT(&conn->conn_data_lock);
3248 channel = np->header.cid & RX_CHANNELMASK;
3249 call = conn->call[channel];
3252 MUTEX_ENTER(&call->lock);
3253 currentCallNumber = conn->callNumber[channel];
3254 } else if (type == RX_SERVER_CONNECTION) { /* No call allocated */
3255 MUTEX_ENTER(&conn->conn_call_lock);
3256 call = conn->call[channel];
3258 MUTEX_ENTER(&call->lock);
3259 MUTEX_EXIT(&conn->conn_call_lock);
3260 currentCallNumber = conn->callNumber[channel];
3262 call = rxi_NewCall(conn, channel); /* returns locked call */
3263 MUTEX_EXIT(&conn->conn_call_lock);
3264 *call->callNumber = currentCallNumber = np->header.callNumber;
3266 if (np->header.callNumber == 0)
3267 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3268 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3269 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3270 np->header.flags, np, np->length));
3272 call->state = RX_STATE_PRECALL;
3273 clock_GetTime(&call->queueTime);
3274 hzero(call->bytesSent);
3275 hzero(call->bytesRcvd);
3277 * If the number of queued calls exceeds the overload
3278 * threshold then abort this call.
3280 if ((rx_BusyThreshold > 0) &&
3281 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3282 struct rx_packet *tp;
3284 rxi_CallError(call, rx_BusyError);
3285 tp = rxi_SendCallAbort(call, np, 1, 0);
3286 MUTEX_EXIT(&call->lock);
3287 MUTEX_ENTER(&rx_refcnt_mutex);
3289 MUTEX_EXIT(&rx_refcnt_mutex);
3290 if (rx_stats_active)
3291 rx_atomic_inc(&rx_stats.nBusies);
3294 rxi_KeepAliveOn(call);
3296 } else { /* RX_CLIENT_CONNECTION and No call allocated */
3297 /* This packet can't be for this call. If the new call address is
3298 * 0 then no call is running on this channel. If there is a call
3299 * then, since this is a client connection we're getting data for
3300 * it must be for the previous call.
3302 if (rx_stats_active)
3303 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3304 MUTEX_ENTER(&rx_refcnt_mutex);
3306 MUTEX_EXIT(&rx_refcnt_mutex);
3310 /* There is a non-NULL locked call at this point */
3311 if (type == RX_SERVER_CONNECTION) { /* We're the server */
3312 if (np->header.callNumber < currentCallNumber) {
3313 MUTEX_EXIT(&call->lock);
3314 if (rx_stats_active)
3315 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3316 MUTEX_ENTER(&rx_refcnt_mutex);
3318 MUTEX_EXIT(&rx_refcnt_mutex);
3320 } else if (np->header.callNumber != currentCallNumber) {
3321 /* Wait until the transmit queue is idle before deciding
3322 * whether to reset the current call. Chances are that the
3323 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3326 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3327 if (call->state == RX_STATE_ACTIVE) {
3328 rxi_WaitforTQBusy(call);
3330 * If we entered error state while waiting,
3331 * must call rxi_CallError to permit rxi_ResetCall
3332 * to processed when the tqWaiter count hits zero.
3335 rxi_CallError(call, call->error);
3336 MUTEX_EXIT(&call->lock);
3337 MUTEX_ENTER(&rx_refcnt_mutex);
3339 MUTEX_EXIT(&rx_refcnt_mutex);
3343 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3344 /* If the new call cannot be taken right now send a busy and set
3345 * the error condition in this call, so that it terminates as
3346 * quickly as possible */
3347 if (call->state == RX_STATE_ACTIVE) {
3348 struct rx_packet *tp;
3350 rxi_CallError(call, RX_CALL_DEAD);
3351 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3353 MUTEX_EXIT(&call->lock);
3354 MUTEX_ENTER(&rx_refcnt_mutex);
3356 MUTEX_EXIT(&rx_refcnt_mutex);
3359 rxi_ResetCall(call, 0);
3360 *call->callNumber = np->header.callNumber;
3362 if (np->header.callNumber == 0)
3363 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3364 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3365 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3366 np->header.flags, np, np->length));
3368 call->state = RX_STATE_PRECALL;
3369 clock_GetTime(&call->queueTime);
3370 hzero(call->bytesSent);
3371 hzero(call->bytesRcvd);
3373 * If the number of queued calls exceeds the overload
3374 * threshold then abort this call.
3376 if ((rx_BusyThreshold > 0) &&
3377 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3378 struct rx_packet *tp;
3380 rxi_CallError(call, rx_BusyError);
3381 tp = rxi_SendCallAbort(call, np, 1, 0);
3382 MUTEX_EXIT(&call->lock);
3383 MUTEX_ENTER(&rx_refcnt_mutex);
3385 MUTEX_EXIT(&rx_refcnt_mutex);
3386 if (rx_stats_active)
3387 rx_atomic_inc(&rx_stats.nBusies);
3390 rxi_KeepAliveOn(call);
3392 /* Continuing call; do nothing here. */
3394 } else { /* we're the client */
3395 /* Ignore all incoming acknowledgements for calls in DALLY state */
3396 if ((call->state == RX_STATE_DALLY)
3397 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3398 if (rx_stats_active)
3399 rx_atomic_inc(&rx_stats.ignorePacketDally);
3400 MUTEX_EXIT(&call->lock);
3401 MUTEX_ENTER(&rx_refcnt_mutex);
3403 MUTEX_EXIT(&rx_refcnt_mutex);
3407 /* Ignore anything that's not relevant to the current call. If there
3408 * isn't a current call, then no packet is relevant. */
3409 if (np->header.callNumber != currentCallNumber) {
3410 if (rx_stats_active)
3411 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3412 MUTEX_EXIT(&call->lock);
3413 MUTEX_ENTER(&rx_refcnt_mutex);
3415 MUTEX_EXIT(&rx_refcnt_mutex);
3418 /* If the service security object index stamped in the packet does not
3419 * match the connection's security index, ignore the packet */
3420 if (np->header.securityIndex != conn->securityIndex) {
3421 MUTEX_EXIT(&call->lock);
3422 MUTEX_ENTER(&rx_refcnt_mutex);
3424 MUTEX_EXIT(&rx_refcnt_mutex);
3428 /* If we're receiving the response, then all transmit packets are
3429 * implicitly acknowledged. Get rid of them. */
3430 if (np->header.type == RX_PACKET_TYPE_DATA) {
3431 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3432 /* XXX Hack. Because we must release the global rx lock when
3433 * sending packets (osi_NetSend) we drop all acks while we're
3434 * traversing the tq in rxi_Start sending packets out because
3435 * packets may move to the freePacketQueue as result of being here!
3436 * So we drop these packets until we're safely out of the
3437 * traversing. Really ugly!
3438 * For fine grain RX locking, we set the acked field in the
3439 * packets and let rxi_Start remove them from the transmit queue.
3441 if (call->flags & RX_CALL_TQ_BUSY) {
3442 #ifdef RX_ENABLE_LOCKS
3443 rxi_SetAcksInTransmitQueue(call);
3445 MUTEX_ENTER(&rx_refcnt_mutex);
3447 MUTEX_EXIT(&rx_refcnt_mutex);
3448 return np; /* xmitting; drop packet */
3451 rxi_ClearTransmitQueue(call, 0);
3453 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3454 rxi_ClearTransmitQueue(call, 0);
3455 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3457 if (np->header.type == RX_PACKET_TYPE_ACK) {
3458 /* now check to see if this is an ack packet acknowledging that the
3459 * server actually *lost* some hard-acked data. If this happens we
3460 * ignore this packet, as it may indicate that the server restarted in
3461 * the middle of a call. It is also possible that this is an old ack
3462 * packet. We don't abort the connection in this case, because this
3463 * *might* just be an old ack packet. The right way to detect a server
3464 * restart in the midst of a call is to notice that the server epoch
3466 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3467 * XXX unacknowledged. I think that this is off-by-one, but
3468 * XXX I don't dare change it just yet, since it will
3469 * XXX interact badly with the server-restart detection
3470 * XXX code in receiveackpacket. */
3471 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3472 if (rx_stats_active)
3473 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3474 MUTEX_EXIT(&call->lock);
3475 MUTEX_ENTER(&rx_refcnt_mutex);
3477 MUTEX_EXIT(&rx_refcnt_mutex);
3481 } /* else not a data packet */
3484 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3485 /* Set remote user defined status from packet */
3486 call->remoteStatus = np->header.userStatus;
3488 /* Note the gap between the expected next packet and the actual
3489 * packet that arrived, when the new packet has a smaller serial number
3490 * than expected. Rioses frequently reorder packets all by themselves,
3491 * so this will be quite important with very large window sizes.
3492 * Skew is checked against 0 here to avoid any dependence on the type of
3493 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3495 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3496 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3497 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3499 MUTEX_ENTER(&conn->conn_data_lock);
3500 skew = conn->lastSerial - np->header.serial;
3501 conn->lastSerial = np->header.serial;
3502 MUTEX_EXIT(&conn->conn_data_lock);
3504 struct rx_peer *peer;
3506 if (skew > peer->inPacketSkew) {
3507 dpf(("*** In skew changed from %d to %d\n",
3508 peer->inPacketSkew, skew));
3509 peer->inPacketSkew = skew;
3513 /* Now do packet type-specific processing */
3514 switch (np->header.type) {
3515 case RX_PACKET_TYPE_DATA:
3516 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3519 case RX_PACKET_TYPE_ACK:
3520 /* Respond immediately to ack packets requesting acknowledgement
3522 if (np->header.flags & RX_REQUEST_ACK) {
3524 (void)rxi_SendCallAbort(call, 0, 1, 0);
3526 (void)rxi_SendAck(call, 0, np->header.serial,
3527 RX_ACK_PING_RESPONSE, 1);
3529 np = rxi_ReceiveAckPacket(call, np, 1);
3531 case RX_PACKET_TYPE_ABORT: {
3532 /* An abort packet: reset the call, passing the error up to the user. */
3533 /* What if error is zero? */
3534 /* What if the error is -1? the application will treat it as a timeout. */
3535 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3536 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3537 rxi_CallError(call, errdata);
3538 MUTEX_EXIT(&call->lock);
3539 MUTEX_ENTER(&rx_refcnt_mutex);
3541 MUTEX_EXIT(&rx_refcnt_mutex);
3542 return np; /* xmitting; drop packet */
3544 case RX_PACKET_TYPE_BUSY: {
3545 struct clock busyTime;
3547 clock_GetTime(&busyTime);
3549 MUTEX_EXIT(&call->lock);
3551 MUTEX_ENTER(&conn->conn_call_lock);
3552 MUTEX_ENTER(&call->lock);
3553 conn->lastBusy[call->channel] = busyTime.sec;
3554 call->flags |= RX_CALL_PEER_BUSY;
3555 MUTEX_EXIT(&call->lock);
3556 MUTEX_EXIT(&conn->conn_call_lock);
3558 MUTEX_ENTER(&rx_refcnt_mutex);
3560 MUTEX_EXIT(&rx_refcnt_mutex);
3564 case RX_PACKET_TYPE_ACKALL:
3565 /* All packets acknowledged, so we can drop all packets previously
3566 * readied for sending */
3567 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3568 /* XXX Hack. We because we can't release the global rx lock when
3569 * sending packets (osi_NetSend) we drop all ack pkts while we're
3570 * traversing the tq in rxi_Start sending packets out because
3571 * packets may move to the freePacketQueue as result of being
3572 * here! So we drop these packets until we're safely out of the
3573 * traversing. Really ugly!
3574 * For fine grain RX locking, we set the acked field in the packets
3575 * and let rxi_Start remove the packets from the transmit queue.
3577 if (call->flags & RX_CALL_TQ_BUSY) {
3578 #ifdef RX_ENABLE_LOCKS
3579 rxi_SetAcksInTransmitQueue(call);
3581 #else /* RX_ENABLE_LOCKS */
3582 MUTEX_EXIT(&call->lock);
3583 MUTEX_ENTER(&rx_refcnt_mutex);
3585 MUTEX_EXIT(&rx_refcnt_mutex);
3586 return np; /* xmitting; drop packet */
3587 #endif /* RX_ENABLE_LOCKS */
3589 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3590 rxi_ClearTransmitQueue(call, 0);
3593 /* Should not reach here, unless the peer is broken: send an abort
3595 rxi_CallError(call, RX_PROTOCOL_ERROR);
3596 np = rxi_SendCallAbort(call, np, 1, 0);
3599 /* Note when this last legitimate packet was received, for keep-alive
3600 * processing. Note, we delay getting the time until now in the hope that
3601 * the packet will be delivered to the user before any get time is required
3602 * (if not, then the time won't actually be re-evaluated here). */
3603 call->lastReceiveTime = clock_Sec();
3604 /* we've received a legit packet, so the channel is not busy */
3605 call->flags &= ~RX_CALL_PEER_BUSY;
3606 MUTEX_EXIT(&call->lock);
3607 MUTEX_ENTER(&rx_refcnt_mutex);
3609 MUTEX_EXIT(&rx_refcnt_mutex);
3613 /* return true if this is an "interesting" connection from the point of view
3614 of someone trying to debug the system */
3616 rxi_IsConnInteresting(struct rx_connection *aconn)
3619 struct rx_call *tcall;
3621 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3624 for (i = 0; i < RX_MAXCALLS; i++) {
3625 tcall = aconn->call[i];
3627 if ((tcall->state == RX_STATE_PRECALL)
3628 || (tcall->state == RX_STATE_ACTIVE))
3630 if ((tcall->mode == RX_MODE_SENDING)
3631 || (tcall->mode == RX_MODE_RECEIVING))
3639 /* if this is one of the last few packets AND it wouldn't be used by the
3640 receiving call to immediately satisfy a read request, then drop it on
3641 the floor, since accepting it might prevent a lock-holding thread from
3642 making progress in its reading. If a call has been cleared while in
3643 the precall state then ignore all subsequent packets until the call
3644 is assigned to a thread. */
3647 TooLow(struct rx_packet *ap, struct rx_call *acall)
3651 MUTEX_ENTER(&rx_quota_mutex);
3652 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3653 && (acall->state == RX_STATE_PRECALL))
3654 || ((rx_nFreePackets < rxi_dataQuota + 2)
3655 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3656 && (acall->flags & RX_CALL_READER_WAIT)))) {
3659 MUTEX_EXIT(&rx_quota_mutex);
3665 * Clear the attach wait flag on a connection and proceed.
3667 * Any processing waiting for a connection to be attached should be
3668 * unblocked. We clear the flag and do any other needed tasks.
3671 * the conn to unmark waiting for attach
3673 * @pre conn's conn_data_lock must be locked before calling this function
3677 rxi_ConnClearAttachWait(struct rx_connection *conn)
3679 /* Indicate that rxi_CheckReachEvent is no longer running by
3680 * clearing the flag. Must be atomic under conn_data_lock to
3681 * avoid a new call slipping by: rxi_CheckConnReach holds
3682 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3684 conn->flags &= ~RX_CONN_ATTACHWAIT;
3685 if (conn->flags & RX_CONN_NAT_PING) {
3686 conn->flags &= ~RX_CONN_NAT_PING;
3687 rxi_ScheduleNatKeepAliveEvent(conn);
3692 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3694 struct rx_connection *conn = arg1;
3695 struct rx_call *acall = arg2;
3696 struct rx_call *call = acall;
3697 struct clock when, now;
3700 MUTEX_ENTER(&conn->conn_data_lock);
3703 rxevent_Put(conn->checkReachEvent);
3704 conn->checkReachEvent = NULL;
3707 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3709 MUTEX_ENTER(&rx_refcnt_mutex);
3711 MUTEX_EXIT(&rx_refcnt_mutex);
3713 MUTEX_EXIT(&conn->conn_data_lock);
3717 MUTEX_ENTER(&conn->conn_call_lock);
3718 MUTEX_ENTER(&conn->conn_data_lock);
3719 for (i = 0; i < RX_MAXCALLS; i++) {
3720 struct rx_call *tc = conn->call[i];
3721 if (tc && tc->state == RX_STATE_PRECALL) {
3727 rxi_ConnClearAttachWait(conn);
3728 MUTEX_EXIT(&conn->conn_data_lock);
3729 MUTEX_EXIT(&conn->conn_call_lock);
3734 MUTEX_ENTER(&call->lock);
3735 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3737 MUTEX_EXIT(&call->lock);
3739 clock_GetTime(&now);
3741 when.sec += RX_CHECKREACH_TIMEOUT;
3742 MUTEX_ENTER(&conn->conn_data_lock);
3743 if (!conn->checkReachEvent) {
3744 MUTEX_ENTER(&rx_refcnt_mutex);
3746 MUTEX_EXIT(&rx_refcnt_mutex);
3747 conn->checkReachEvent = rxevent_Post(&when, &now,
3748 rxi_CheckReachEvent, conn,
3751 MUTEX_EXIT(&conn->conn_data_lock);
3757 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3759 struct rx_service *service = conn->service;
3760 struct rx_peer *peer = conn->peer;
3761 afs_uint32 now, lastReach;
3763 if (service->checkReach == 0)
3767 MUTEX_ENTER(&peer->peer_lock);
3768 lastReach = peer->lastReachTime;
3769 MUTEX_EXIT(&peer->peer_lock);
3770 if (now - lastReach < RX_CHECKREACH_TTL)
3773 MUTEX_ENTER(&conn->conn_data_lock);
3774 if (conn->flags & RX_CONN_ATTACHWAIT) {
3775 MUTEX_EXIT(&conn->conn_data_lock);
3778 conn->flags |= RX_CONN_ATTACHWAIT;
3779 MUTEX_EXIT(&conn->conn_data_lock);
3780 if (!conn->checkReachEvent)
3781 rxi_CheckReachEvent(NULL, conn, call, 0);
3786 /* try to attach call, if authentication is complete */
3788 TryAttach(struct rx_call *acall, osi_socket socket,
3789 int *tnop, struct rx_call **newcallp,
3792 struct rx_connection *conn = acall->conn;
3794 if (conn->type == RX_SERVER_CONNECTION
3795 && acall->state == RX_STATE_PRECALL) {
3796 /* Don't attach until we have any req'd. authentication. */
3797 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3798 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3799 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3800 /* Note: this does not necessarily succeed; there
3801 * may not any proc available
3804 rxi_ChallengeOn(acall->conn);
3809 /* A data packet has been received off the interface. This packet is
3810 * appropriate to the call (the call is in the right state, etc.). This
3811 * routine can return a packet to the caller, for re-use */
3814 rxi_ReceiveDataPacket(struct rx_call *call,
3815 struct rx_packet *np, int istack,
3816 osi_socket socket, afs_uint32 host, u_short port,
3817 int *tnop, struct rx_call **newcallp)
3819 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3824 afs_uint32 serial=0, flags=0;
3826 struct rx_packet *tnp;
3827 if (rx_stats_active)
3828 rx_atomic_inc(&rx_stats.dataPacketsRead);
3831 /* If there are no packet buffers, drop this new packet, unless we can find
3832 * packet buffers from inactive calls */
3834 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3835 MUTEX_ENTER(&rx_freePktQ_lock);
3836 rxi_NeedMorePackets = TRUE;
3837 MUTEX_EXIT(&rx_freePktQ_lock);
3838 if (rx_stats_active)
3839 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3840 call->rprev = np->header.serial;
3841 rxi_calltrace(RX_TRACE_DROP, call);
3842 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3843 /* We used to clear the receive queue here, in an attempt to free
3844 * packets. However this is unsafe if the queue has received a
3845 * soft ACK for the final packet */
3846 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3848 /* we've damaged this call already, might as well do it in. */
3854 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3855 * packet is one of several packets transmitted as a single
3856 * datagram. Do not send any soft or hard acks until all packets
3857 * in a jumbogram have been processed. Send negative acks right away.
3859 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3860 /* tnp is non-null when there are more packets in the
3861 * current jumbo gram */
3868 seq = np->header.seq;
3869 serial = np->header.serial;
3870 flags = np->header.flags;
3872 /* If the call is in an error state, send an abort message */
3874 return rxi_SendCallAbort(call, np, istack, 0);
3876 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3877 * AFS 3.5 jumbogram. */
3878 if (flags & RX_JUMBO_PACKET) {
3879 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3884 if (np->header.spare != 0) {
3885 MUTEX_ENTER(&call->conn->conn_data_lock);
3886 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3887 MUTEX_EXIT(&call->conn->conn_data_lock);
3890 /* The usual case is that this is the expected next packet */
3891 if (seq == call->rnext) {
3893 /* Check to make sure it is not a duplicate of one already queued */
3894 if (queue_IsNotEmpty(&call->rq)
3895 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3896 if (rx_stats_active)
3897 rx_atomic_inc(&rx_stats.dupPacketsRead);
3898 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3899 rxevent_Cancel(&call->delayedAckEvent, call,
3900 RX_CALL_REFCOUNT_DELAY);
3901 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3907 /* It's the next packet. Stick it on the receive queue
3908 * for this call. Set newPackets to make sure we wake
3909 * the reader once all packets have been processed */
3910 #ifdef RX_TRACK_PACKETS
3911 np->flags |= RX_PKTFLAG_RQ;
3913 queue_Prepend(&call->rq, np);
3914 #ifdef RXDEBUG_PACKET
3916 #endif /* RXDEBUG_PACKET */
3918 np = NULL; /* We can't use this anymore */
3921 /* If an ack is requested then set a flag to make sure we
3922 * send an acknowledgement for this packet */
3923 if (flags & RX_REQUEST_ACK) {
3924 ackNeeded = RX_ACK_REQUESTED;
3927 /* Keep track of whether we have received the last packet */
3928 if (flags & RX_LAST_PACKET) {
3929 call->flags |= RX_CALL_HAVE_LAST;
3933 /* Check whether we have all of the packets for this call */
3934 if (call->flags & RX_CALL_HAVE_LAST) {
3935 afs_uint32 tseq; /* temporary sequence number */
3936 struct rx_packet *tp; /* Temporary packet pointer */
3937 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3939 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3940 if (tseq != tp->header.seq)
3942 if (tp->header.flags & RX_LAST_PACKET) {
3943 call->flags |= RX_CALL_RECEIVE_DONE;
3950 /* Provide asynchronous notification for those who want it
3951 * (e.g. multi rx) */
3952 if (call->arrivalProc) {
3953 (*call->arrivalProc) (call, call->arrivalProcHandle,
3954 call->arrivalProcArg);
3955 call->arrivalProc = (void (*)())0;
3958 /* Update last packet received */
3961 /* If there is no server process serving this call, grab
3962 * one, if available. We only need to do this once. If a
3963 * server thread is available, this thread becomes a server
3964 * thread and the server thread becomes a listener thread. */
3966 TryAttach(call, socket, tnop, newcallp, 0);
3969 /* This is not the expected next packet. */
3971 /* Determine whether this is a new or old packet, and if it's
3972 * a new one, whether it fits into the current receive window.
3973 * Also figure out whether the packet was delivered in sequence.
3974 * We use the prev variable to determine whether the new packet
3975 * is the successor of its immediate predecessor in the
3976 * receive queue, and the missing flag to determine whether
3977 * any of this packets predecessors are missing. */
3979 afs_uint32 prev; /* "Previous packet" sequence number */
3980 struct rx_packet *tp; /* Temporary packet pointer */
3981 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3982 int missing; /* Are any predecessors missing? */
3984 /* If the new packet's sequence number has been sent to the
3985 * application already, then this is a duplicate */
3986 if (seq < call->rnext) {
3987 if (rx_stats_active)
3988 rx_atomic_inc(&rx_stats.dupPacketsRead);
3989 rxevent_Cancel(&call->delayedAckEvent, call,
3990 RX_CALL_REFCOUNT_DELAY);
3991 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3997 /* If the sequence number is greater than what can be
3998 * accomodated by the current window, then send a negative
3999 * acknowledge and drop the packet */
4000 if ((call->rnext + call->rwind) <= seq) {
4001 rxevent_Cancel(&call->delayedAckEvent, call,
4002 RX_CALL_REFCOUNT_DELAY);
4003 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
4010 /* Look for the packet in the queue of old received packets */
4011 for (prev = call->rnext - 1, missing =
4012 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
4013 /*Check for duplicate packet */
4014 if (seq == tp->header.seq) {
4015 if (rx_stats_active)
4016 rx_atomic_inc(&rx_stats.dupPacketsRead);
4017 rxevent_Cancel(&call->delayedAckEvent, call,
4018 RX_CALL_REFCOUNT_DELAY);
4019 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
4025 /* If we find a higher sequence packet, break out and
4026 * insert the new packet here. */
4027 if (seq < tp->header.seq)
4029 /* Check for missing packet */
4030 if (tp->header.seq != prev + 1) {
4034 prev = tp->header.seq;
4037 /* Keep track of whether we have received the last packet. */
4038 if (flags & RX_LAST_PACKET) {
4039 call->flags |= RX_CALL_HAVE_LAST;
4042 /* It's within the window: add it to the the receive queue.
4043 * tp is left by the previous loop either pointing at the
4044 * packet before which to insert the new packet, or at the
4045 * queue head if the queue is empty or the packet should be
4047 #ifdef RX_TRACK_PACKETS
4048 np->flags |= RX_PKTFLAG_RQ;
4050 #ifdef RXDEBUG_PACKET
4052 #endif /* RXDEBUG_PACKET */
4053 queue_InsertBefore(tp, np);
4057 /* Check whether we have all of the packets for this call */
4058 if ((call->flags & RX_CALL_HAVE_LAST)
4059 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4060 afs_uint32 tseq; /* temporary sequence number */
4063 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
4064 if (tseq != tp->header.seq)
4066 if (tp->header.flags & RX_LAST_PACKET) {
4067 call->flags |= RX_CALL_RECEIVE_DONE;
4074 /* We need to send an ack of the packet is out of sequence,
4075 * or if an ack was requested by the peer. */
4076 if (seq != prev + 1 || missing) {
4077 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4078 } else if (flags & RX_REQUEST_ACK) {
4079 ackNeeded = RX_ACK_REQUESTED;
4082 /* Acknowledge the last packet for each call */
4083 if (flags & RX_LAST_PACKET) {
4094 * If the receiver is waiting for an iovec, fill the iovec
4095 * using the data from the receive queue */
4096 if (call->flags & RX_CALL_IOVEC_WAIT) {
4097 didHardAck = rxi_FillReadVec(call, serial);
4098 /* the call may have been aborted */
4107 /* Wakeup the reader if any */
4108 if ((call->flags & RX_CALL_READER_WAIT)
4109 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4110 || (call->iovNext >= call->iovMax)
4111 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4112 call->flags &= ~RX_CALL_READER_WAIT;
4113 #ifdef RX_ENABLE_LOCKS
4114 CV_BROADCAST(&call->cv_rq);
4116 osi_rxWakeup(&call->rq);
4122 * Send an ack when requested by the peer, or once every
4123 * rxi_SoftAckRate packets until the last packet has been
4124 * received. Always send a soft ack for the last packet in
4125 * the server's reply. */
4127 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4128 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4129 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4130 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4131 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4132 } else if (call->nSoftAcks) {
4133 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4134 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4136 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4137 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4138 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4145 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4147 struct rx_peer *peer = conn->peer;
4149 MUTEX_ENTER(&peer->peer_lock);
4150 peer->lastReachTime = clock_Sec();
4151 MUTEX_EXIT(&peer->peer_lock);
4153 MUTEX_ENTER(&conn->conn_data_lock);
4154 if (conn->flags & RX_CONN_ATTACHWAIT) {
4157 rxi_ConnClearAttachWait(conn);
4158 MUTEX_EXIT(&conn->conn_data_lock);
4160 for (i = 0; i < RX_MAXCALLS; i++) {
4161 struct rx_call *call = conn->call[i];
4164 MUTEX_ENTER(&call->lock);
4165 /* tnop can be null if newcallp is null */
4166 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4168 MUTEX_EXIT(&call->lock);
4172 MUTEX_EXIT(&conn->conn_data_lock);
4175 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4177 rx_ack_reason(int reason)
4180 case RX_ACK_REQUESTED:
4182 case RX_ACK_DUPLICATE:
4184 case RX_ACK_OUT_OF_SEQUENCE:
4186 case RX_ACK_EXCEEDS_WINDOW:
4188 case RX_ACK_NOSPACE:
4192 case RX_ACK_PING_RESPONSE:
4205 /* The real smarts of the whole thing. */
4207 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4210 struct rx_ackPacket *ap;
4212 struct rx_packet *tp;
4213 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
4214 struct rx_connection *conn = call->conn;
4215 struct rx_peer *peer = conn->peer;
4216 struct clock now; /* Current time, for RTT calculations */
4220 /* because there are CM's that are bogus, sending weird values for this. */
4221 afs_uint32 skew = 0;
4226 int newAckCount = 0;
4227 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4228 int pktsize = 0; /* Set if we need to update the peer mtu */
4229 int conn_data_locked = 0;
4231 if (rx_stats_active)
4232 rx_atomic_inc(&rx_stats.ackPacketsRead);
4233 ap = (struct rx_ackPacket *)rx_DataOf(np);
4234 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4236 return np; /* truncated ack packet */
4238 /* depends on ack packet struct */
4239 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4240 first = ntohl(ap->firstPacket);
4241 prev = ntohl(ap->previousPacket);
4242 serial = ntohl(ap->serial);
4243 /* temporarily disabled -- needs to degrade over time
4244 * skew = ntohs(ap->maxSkew); */
4246 /* Ignore ack packets received out of order */
4247 if (first < call->tfirst ||
4248 (first == call->tfirst && prev < call->tprev)) {
4254 if (np->header.flags & RX_SLOW_START_OK) {
4255 call->flags |= RX_CALL_SLOW_START_OK;
4258 if (ap->reason == RX_ACK_PING_RESPONSE)
4259 rxi_UpdatePeerReach(conn, call);
4261 if (conn->lastPacketSizeSeq) {
4262 MUTEX_ENTER(&conn->conn_data_lock);
4263 conn_data_locked = 1;
4264 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4265 pktsize = conn->lastPacketSize;
4266 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4269 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4270 if (!conn_data_locked) {
4271 MUTEX_ENTER(&conn->conn_data_lock);
4272 conn_data_locked = 1;
4274 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4275 /* process mtu ping ack */
4276 pktsize = conn->lastPingSize;
4277 conn->lastPingSizeSer = conn->lastPingSize = 0;
4281 if (conn_data_locked) {
4282 MUTEX_EXIT(&conn->conn_data_lock);
4283 conn_data_locked = 0;
4287 if (rxdebug_active) {
4291 len = _snprintf(msg, sizeof(msg),
4292 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
4293 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4294 ntohl(ap->serial), ntohl(ap->previousPacket),
4295 (unsigned int)np->header.seq, (unsigned int)skew,
4296 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
4300 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4301 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4305 OutputDebugString(msg);
4307 #else /* AFS_NT40_ENV */
4310 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
4311 ap->reason, ntohl(ap->previousPacket),
4312 (unsigned int)np->header.seq, (unsigned int)serial,
4313 (unsigned int)skew, ntohl(ap->firstPacket));
4316 for (offset = 0; offset < nAcks; offset++)
4317 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4322 #endif /* AFS_NT40_ENV */
4325 MUTEX_ENTER(&peer->peer_lock);
4328 * Start somewhere. Can't assume we can send what we can receive,
4329 * but we are clearly receiving.
4331 if (!peer->maxPacketSize)
4332 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4334 if (pktsize > peer->maxPacketSize) {
4335 peer->maxPacketSize = pktsize;
4336 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4337 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4338 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4339 rxi_ScheduleGrowMTUEvent(call, 1);
4344 /* Update the outgoing packet skew value to the latest value of
4345 * the peer's incoming packet skew value. The ack packet, of
4346 * course, could arrive out of order, but that won't affect things
4348 peer->outPacketSkew = skew;
4351 clock_GetTime(&now);
4353 /* The transmit queue splits into 4 sections.
4355 * The first section is packets which have now been acknowledged
4356 * by a window size change in the ack. These have reached the
4357 * application layer, and may be discarded. These are packets
4358 * with sequence numbers < ap->firstPacket.
4360 * The second section is packets which have sequence numbers in
4361 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4362 * contents of the packet's ack array determines whether these
4363 * packets are acknowledged or not.
4365 * The third section is packets which fall above the range
4366 * addressed in the ack packet. These have not yet been received
4369 * The four section is packets which have not yet been transmitted.
4370 * These packets will have a header.serial of 0.
4373 /* First section - implicitly acknowledged packets that can be
4377 tp = queue_First(&call->tq, rx_packet);
4378 while(!queue_IsEnd(&call->tq, tp) && tp->header.seq < first) {
4379 struct rx_packet *next;
4381 next = queue_Next(tp, rx_packet);
4382 call->tfirst = tp->header.seq + 1;
4384 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4386 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4389 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4390 /* XXX Hack. Because we have to release the global rx lock when sending
4391 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4392 * in rxi_Start sending packets out because packets may move to the
4393 * freePacketQueue as result of being here! So we drop these packets until
4394 * we're safely out of the traversing. Really ugly!
4395 * To make it even uglier, if we're using fine grain locking, we can
4396 * set the ack bits in the packets and have rxi_Start remove the packets
4397 * when it's done transmitting.
4399 if (call->flags & RX_CALL_TQ_BUSY) {
4400 #ifdef RX_ENABLE_LOCKS
4401 tp->flags |= RX_PKTFLAG_ACKED;
4402 call->flags |= RX_CALL_TQ_SOME_ACKED;
4403 #else /* RX_ENABLE_LOCKS */
4405 #endif /* RX_ENABLE_LOCKS */
4407 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4410 #ifdef RX_TRACK_PACKETS
4411 tp->flags &= ~RX_PKTFLAG_TQ;
4413 #ifdef RXDEBUG_PACKET
4415 #endif /* RXDEBUG_PACKET */
4416 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4421 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4423 /* Second section of the queue - packets for which we are receiving
4426 * Go through the explicit acks/nacks and record the results in
4427 * the waiting packets. These are packets that can't be released
4428 * yet, even with a positive acknowledge. This positive
4429 * acknowledge only means the packet has been received by the
4430 * peer, not that it will be retained long enough to be sent to
4431 * the peer's upper level. In addition, reset the transmit timers
4432 * of any missing packets (those packets that must be missing
4433 * because this packet was out of sequence) */
4435 call->nSoftAcked = 0;
4437 while (!queue_IsEnd(&call->tq, tp) && tp->header.seq < first + nAcks) {
4438 /* Set the acknowledge flag per packet based on the
4439 * information in the ack packet. An acknowlegded packet can
4440 * be downgraded when the server has discarded a packet it
4441 * soacked previously, or when an ack packet is received
4442 * out of sequence. */
4443 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4444 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4446 tp->flags |= RX_PKTFLAG_ACKED;
4447 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4454 } else /* RX_ACK_TYPE_NACK */ {
4455 tp->flags &= ~RX_PKTFLAG_ACKED;
4459 tp = queue_Next(tp, rx_packet);
4462 /* We don't need to take any action with the 3rd or 4th section in the
4463 * queue - they're not addressed by the contents of this ACK packet.
4466 /* If the window has been extended by this acknowledge packet,
4467 * then wakeup a sender waiting in alloc for window space, or try
4468 * sending packets now, if he's been sitting on packets due to
4469 * lack of window space */
4470 if (call->tnext < (call->tfirst + call->twind)) {
4471 #ifdef RX_ENABLE_LOCKS
4472 CV_SIGNAL(&call->cv_twind);
4474 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4475 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4476 osi_rxWakeup(&call->twind);
4479 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4480 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4484 /* if the ack packet has a receivelen field hanging off it,
4485 * update our state */
4486 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4489 /* If the ack packet has a "recommended" size that is less than
4490 * what I am using now, reduce my size to match */
4491 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4492 (int)sizeof(afs_int32), &tSize);
4493 tSize = (afs_uint32) ntohl(tSize);
4494 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4496 /* Get the maximum packet size to send to this peer */
4497 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4499 tSize = (afs_uint32) ntohl(tSize);
4500 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4501 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4503 /* sanity check - peer might have restarted with different params.
4504 * If peer says "send less", dammit, send less... Peer should never
4505 * be unable to accept packets of the size that prior AFS versions would
4506 * send without asking. */
4507 if (peer->maxMTU != tSize) {
4508 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4510 peer->maxMTU = tSize;
4511 peer->MTU = MIN(tSize, peer->MTU);
4512 call->MTU = MIN(call->MTU, tSize);
4515 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4518 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4519 (int)sizeof(afs_int32), &tSize);
4520 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4521 if (tSize < call->twind) { /* smaller than our send */
4522 call->twind = tSize; /* window, we must send less... */
4523 call->ssthresh = MIN(call->twind, call->ssthresh);
4524 call->conn->twind[call->channel] = call->twind;
4527 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4528 * network MTU confused with the loopback MTU. Calculate the
4529 * maximum MTU here for use in the slow start code below.
4531 /* Did peer restart with older RX version? */
4532 if (peer->maxDgramPackets > 1) {
4533 peer->maxDgramPackets = 1;
4535 } else if (np->length >=
4536 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4539 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4540 sizeof(afs_int32), &tSize);
4541 tSize = (afs_uint32) ntohl(tSize);
4543 * As of AFS 3.5 we set the send window to match the receive window.
4545 if (tSize < call->twind) {
4546 call->twind = tSize;
4547 call->conn->twind[call->channel] = call->twind;
4548 call->ssthresh = MIN(call->twind, call->ssthresh);
4549 } else if (tSize > call->twind) {
4550 call->twind = tSize;
4551 call->conn->twind[call->channel] = call->twind;
4555 * As of AFS 3.5, a jumbogram is more than one fixed size
4556 * packet transmitted in a single UDP datagram. If the remote
4557 * MTU is smaller than our local MTU then never send a datagram
4558 * larger than the natural MTU.
4561 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4562 (int)sizeof(afs_int32), &tSize);
4563 maxDgramPackets = (afs_uint32) ntohl(tSize);
4564 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4566 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4567 if (maxDgramPackets > 1) {
4568 peer->maxDgramPackets = maxDgramPackets;
4569 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4571 peer->maxDgramPackets = 1;
4572 call->MTU = peer->natMTU;
4574 } else if (peer->maxDgramPackets > 1) {
4575 /* Restarted with lower version of RX */
4576 peer->maxDgramPackets = 1;
4578 } else if (peer->maxDgramPackets > 1
4579 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4580 /* Restarted with lower version of RX */
4581 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4582 peer->natMTU = OLD_MAX_PACKET_SIZE;
4583 peer->MTU = OLD_MAX_PACKET_SIZE;
4584 peer->maxDgramPackets = 1;
4585 peer->nDgramPackets = 1;
4587 call->MTU = OLD_MAX_PACKET_SIZE;
4592 * Calculate how many datagrams were successfully received after
4593 * the first missing packet and adjust the negative ack counter
4598 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4599 if (call->nNacks < nNacked) {
4600 call->nNacks = nNacked;
4603 call->nAcks += newAckCount;
4607 /* If the packet contained new acknowledgements, rather than just
4608 * being a duplicate of one we have previously seen, then we can restart
4611 if (newAckCount > 0)
4612 rxi_rto_packet_acked(call, istack);
4614 if (call->flags & RX_CALL_FAST_RECOVER) {
4615 if (newAckCount == 0) {
4616 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4618 call->flags &= ~RX_CALL_FAST_RECOVER;
4619 call->cwind = call->nextCwind;
4620 call->nextCwind = 0;
4623 call->nCwindAcks = 0;
4624 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4625 /* Three negative acks in a row trigger congestion recovery */
4626 call->flags |= RX_CALL_FAST_RECOVER;
4627 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4629 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4630 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4631 call->nextCwind = call->ssthresh;
4634 peer->MTU = call->MTU;
4635 peer->cwind = call->nextCwind;
4636 peer->nDgramPackets = call->nDgramPackets;
4638 call->congestSeq = peer->congestSeq;
4640 /* Reset the resend times on the packets that were nacked
4641 * so we will retransmit as soon as the window permits
4644 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4646 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4647 tp->flags &= ~RX_PKTFLAG_SENT;
4649 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4654 /* If cwind is smaller than ssthresh, then increase
4655 * the window one packet for each ack we receive (exponential
4657 * If cwind is greater than or equal to ssthresh then increase
4658 * the congestion window by one packet for each cwind acks we
4659 * receive (linear growth). */
4660 if (call->cwind < call->ssthresh) {
4662 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4663 call->nCwindAcks = 0;
4665 call->nCwindAcks += newAckCount;
4666 if (call->nCwindAcks >= call->cwind) {
4667 call->nCwindAcks = 0;
4668 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4672 * If we have received several acknowledgements in a row then
4673 * it is time to increase the size of our datagrams
4675 if ((int)call->nAcks > rx_nDgramThreshold) {
4676 if (peer->maxDgramPackets > 1) {
4677 if (call->nDgramPackets < peer->maxDgramPackets) {
4678 call->nDgramPackets++;
4680 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4681 } else if (call->MTU < peer->maxMTU) {
4682 /* don't upgrade if we can't handle it */
4683 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4684 call->MTU = peer->ifMTU;
4686 call->MTU += peer->natMTU;
4687 call->MTU = MIN(call->MTU, peer->maxMTU);
4694 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4696 /* Servers need to hold the call until all response packets have
4697 * been acknowledged. Soft acks are good enough since clients
4698 * are not allowed to clear their receive queues. */
4699 if (call->state == RX_STATE_HOLD
4700 && call->tfirst + call->nSoftAcked >= call->tnext) {
4701 call->state = RX_STATE_DALLY;
4702 rxi_ClearTransmitQueue(call, 0);
4703 rxevent_Cancel(&call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4704 } else if (!queue_IsEmpty(&call->tq)) {
4705 rxi_Start(call, istack);
4710 /* Received a response to a challenge packet */
4712 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4713 struct rx_packet *np, int istack)
4717 /* Ignore the packet if we're the client */
4718 if (conn->type == RX_CLIENT_CONNECTION)
4721 /* If already authenticated, ignore the packet (it's probably a retry) */
4722 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4725 /* Otherwise, have the security object evaluate the response packet */
4726 error = RXS_CheckResponse(conn->securityObject, conn, np);
4728 /* If the response is invalid, reset the connection, sending
4729 * an abort to the peer */
4733 rxi_ConnectionError(conn, error);
4734 MUTEX_ENTER(&conn->conn_data_lock);
4735 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4736 MUTEX_EXIT(&conn->conn_data_lock);
4739 /* If the response is valid, any calls waiting to attach
4740 * servers can now do so */
4743 for (i = 0; i < RX_MAXCALLS; i++) {
4744 struct rx_call *call = conn->call[i];
4746 MUTEX_ENTER(&call->lock);
4747 if (call->state == RX_STATE_PRECALL)
4748 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4749 /* tnop can be null if newcallp is null */
4750 MUTEX_EXIT(&call->lock);
4754 /* Update the peer reachability information, just in case
4755 * some calls went into attach-wait while we were waiting
4756 * for authentication..
4758 rxi_UpdatePeerReach(conn, NULL);
4763 /* A client has received an authentication challenge: the security
4764 * object is asked to cough up a respectable response packet to send
4765 * back to the server. The server is responsible for retrying the
4766 * challenge if it fails to get a response. */
4769 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4770 struct rx_packet *np, int istack)
4774 /* Ignore the challenge if we're the server */
4775 if (conn->type == RX_SERVER_CONNECTION)
4778 /* Ignore the challenge if the connection is otherwise idle; someone's
4779 * trying to use us as an oracle. */
4780 if (!rxi_HasActiveCalls(conn))
4783 /* Send the security object the challenge packet. It is expected to fill
4784 * in the response. */
4785 error = RXS_GetResponse(conn->securityObject, conn, np);
4787 /* If the security object is unable to return a valid response, reset the
4788 * connection and send an abort to the peer. Otherwise send the response
4789 * packet to the peer connection. */
4791 rxi_ConnectionError(conn, error);
4792 MUTEX_ENTER(&conn->conn_data_lock);
4793 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4794 MUTEX_EXIT(&conn->conn_data_lock);
4796 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4797 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4803 /* Find an available server process to service the current request in
4804 * the given call structure. If one isn't available, queue up this
4805 * call so it eventually gets one */
4807 rxi_AttachServerProc(struct rx_call *call,
4808 osi_socket socket, int *tnop,
4809 struct rx_call **newcallp)
4811 struct rx_serverQueueEntry *sq;
4812 struct rx_service *service = call->conn->service;
4815 /* May already be attached */
4816 if (call->state == RX_STATE_ACTIVE)
4819 MUTEX_ENTER(&rx_serverPool_lock);
4821 haveQuota = QuotaOK(service);
4822 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4823 /* If there are no processes available to service this call,
4824 * put the call on the incoming call queue (unless it's
4825 * already on the queue).
4827 #ifdef RX_ENABLE_LOCKS
4829 ReturnToServerPool(service);
4830 #endif /* RX_ENABLE_LOCKS */
4832 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4833 call->flags |= RX_CALL_WAIT_PROC;
4834 rx_atomic_inc(&rx_nWaiting);
4835 rx_atomic_inc(&rx_nWaited);
4836 rxi_calltrace(RX_CALL_ARRIVAL, call);
4837 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4838 queue_Append(&rx_incomingCallQueue, call);
4841 sq = queue_Last(&rx_idleServerQueue, rx_serverQueueEntry);
4843 /* If hot threads are enabled, and both newcallp and sq->socketp
4844 * are non-null, then this thread will process the call, and the
4845 * idle server thread will start listening on this threads socket.
4848 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4851 *sq->socketp = socket;
4852 clock_GetTime(&call->startTime);
4853 MUTEX_ENTER(&rx_refcnt_mutex);
4854 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4855 MUTEX_EXIT(&rx_refcnt_mutex);
4859 if (call->flags & RX_CALL_WAIT_PROC) {
4860 /* Conservative: I don't think this should happen */
4861 call->flags &= ~RX_CALL_WAIT_PROC;
4862 if (queue_IsOnQueue(call)) {
4865 rx_atomic_dec(&rx_nWaiting);
4868 call->state = RX_STATE_ACTIVE;
4869 call->mode = RX_MODE_RECEIVING;
4870 #ifdef RX_KERNEL_TRACE
4872 int glockOwner = ISAFS_GLOCK();
4875 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4876 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4882 if (call->flags & RX_CALL_CLEARED) {
4883 /* send an ack now to start the packet flow up again */
4884 call->flags &= ~RX_CALL_CLEARED;
4885 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4887 #ifdef RX_ENABLE_LOCKS
4890 service->nRequestsRunning++;
4891 MUTEX_ENTER(&rx_quota_mutex);
4892 if (service->nRequestsRunning <= service->minProcs)
4895 MUTEX_EXIT(&rx_quota_mutex);
4899 MUTEX_EXIT(&rx_serverPool_lock);
4902 /* Delay the sending of an acknowledge event for a short while, while
4903 * a new call is being prepared (in the case of a client) or a reply
4904 * is being prepared (in the case of a server). Rather than sending
4905 * an ack packet, an ACKALL packet is sent. */
4907 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4909 #ifdef RX_ENABLE_LOCKS
4911 MUTEX_ENTER(&call->lock);
4912 rxevent_Put(call->delayedAckEvent);
4913 call->delayedAckEvent = NULL;
4914 MUTEX_ENTER(&rx_refcnt_mutex);
4915 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4916 MUTEX_EXIT(&rx_refcnt_mutex);
4918 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4919 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4920 call->flags |= RX_CALL_ACKALL_SENT;
4922 MUTEX_EXIT(&call->lock);
4923 #else /* RX_ENABLE_LOCKS */
4925 rxevent_Put(call->delayedAckEvent);
4926 call->delayedAckEvent = NULL;
4928 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4929 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4930 call->flags |= RX_CALL_ACKALL_SENT;
4931 #endif /* RX_ENABLE_LOCKS */
4935 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4938 struct rx_call *call = arg1;
4939 #ifdef RX_ENABLE_LOCKS
4941 MUTEX_ENTER(&call->lock);
4942 if (event == call->delayedAckEvent) {
4943 rxevent_Put(call->delayedAckEvent);
4944 call->delayedAckEvent = NULL;
4946 MUTEX_ENTER(&rx_refcnt_mutex);
4947 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4948 MUTEX_EXIT(&rx_refcnt_mutex);
4950 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4952 MUTEX_EXIT(&call->lock);
4953 #else /* RX_ENABLE_LOCKS */
4955 rxevent_Put(call->delayedAckEvent);
4956 call->delayedAckEvent = NULL;
4958 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4959 #endif /* RX_ENABLE_LOCKS */
4963 #ifdef RX_ENABLE_LOCKS
4964 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4965 * clearing them out.
4968 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4970 struct rx_packet *p, *tp;
4973 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4974 p->flags |= RX_PKTFLAG_ACKED;
4978 call->flags |= RX_CALL_TQ_CLEARME;
4979 call->flags |= RX_CALL_TQ_SOME_ACKED;
4982 rxi_rto_cancel(call);
4984 call->tfirst = call->tnext;
4985 call->nSoftAcked = 0;
4987 if (call->flags & RX_CALL_FAST_RECOVER) {
4988 call->flags &= ~RX_CALL_FAST_RECOVER;
4989 call->cwind = call->nextCwind;
4990 call->nextCwind = 0;
4993 CV_SIGNAL(&call->cv_twind);
4995 #endif /* RX_ENABLE_LOCKS */
4997 /* Clear out the transmit queue for the current call (all packets have
4998 * been received by peer) */
5000 rxi_ClearTransmitQueue(struct rx_call *call, int force)
5002 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5003 struct rx_packet *p, *tp;
5005 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
5007 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
5008 p->flags |= RX_PKTFLAG_ACKED;
5012 call->flags |= RX_CALL_TQ_CLEARME;
5013 call->flags |= RX_CALL_TQ_SOME_ACKED;
5016 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5017 #ifdef RXDEBUG_PACKET
5019 #endif /* RXDEBUG_PACKET */
5020 rxi_FreePackets(0, &call->tq);
5021 rxi_WakeUpTransmitQueue(call);
5022 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5023 call->flags &= ~RX_CALL_TQ_CLEARME;
5025 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5027 rxi_rto_cancel(call);
5028 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5029 call->nSoftAcked = 0;
5031 if (call->flags & RX_CALL_FAST_RECOVER) {
5032 call->flags &= ~RX_CALL_FAST_RECOVER;
5033 call->cwind = call->nextCwind;
5035 #ifdef RX_ENABLE_LOCKS
5036 CV_SIGNAL(&call->cv_twind);
5038 osi_rxWakeup(&call->twind);
5043 rxi_ClearReceiveQueue(struct rx_call *call)
5045 if (queue_IsNotEmpty(&call->rq)) {
5048 count = rxi_FreePackets(0, &call->rq);
5049 rx_packetReclaims += count;
5050 #ifdef RXDEBUG_PACKET
5052 if ( call->rqc != 0 )
5053 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5055 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5057 if (call->state == RX_STATE_PRECALL) {
5058 call->flags |= RX_CALL_CLEARED;
5062 /* Send an abort packet for the specified call */
5064 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5065 int istack, int force)
5068 struct clock when, now;
5073 /* Clients should never delay abort messages */
5074 if (rx_IsClientConn(call->conn))
5077 if (call->abortCode != call->error) {
5078 call->abortCode = call->error;
5079 call->abortCount = 0;
5082 if (force || rxi_callAbortThreshhold == 0
5083 || call->abortCount < rxi_callAbortThreshhold) {
5084 if (call->delayedAbortEvent) {
5085 rxevent_Cancel(&call->delayedAbortEvent, call,
5086 RX_CALL_REFCOUNT_ABORT);
5088 error = htonl(call->error);
5091 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5092 (char *)&error, sizeof(error), istack);
5093 } else if (!call->delayedAbortEvent) {
5094 clock_GetTime(&now);
5096 clock_Addmsec(&when, rxi_callAbortDelay);
5097 MUTEX_ENTER(&rx_refcnt_mutex);
5098 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5099 MUTEX_EXIT(&rx_refcnt_mutex);
5100 call->delayedAbortEvent =
5101 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5106 /* Send an abort packet for the specified connection. Packet is an
5107 * optional pointer to a packet that can be used to send the abort.
5108 * Once the number of abort messages reaches the threshhold, an
5109 * event is scheduled to send the abort. Setting the force flag
5110 * overrides sending delayed abort messages.
5112 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5113 * to send the abort packet.
5116 rxi_SendConnectionAbort(struct rx_connection *conn,
5117 struct rx_packet *packet, int istack, int force)
5120 struct clock when, now;
5125 /* Clients should never delay abort messages */
5126 if (rx_IsClientConn(conn))
5129 if (force || rxi_connAbortThreshhold == 0
5130 || conn->abortCount < rxi_connAbortThreshhold) {
5132 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
5133 error = htonl(conn->error);
5135 MUTEX_EXIT(&conn->conn_data_lock);
5137 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5138 RX_PACKET_TYPE_ABORT, (char *)&error,
5139 sizeof(error), istack);
5140 MUTEX_ENTER(&conn->conn_data_lock);
5141 } else if (!conn->delayedAbortEvent) {
5142 clock_GetTime(&now);
5144 clock_Addmsec(&when, rxi_connAbortDelay);
5145 conn->delayedAbortEvent =
5146 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
5151 /* Associate an error all of the calls owned by a connection. Called
5152 * with error non-zero. This is only for really fatal things, like
5153 * bad authentication responses. The connection itself is set in
5154 * error at this point, so that future packets received will be
5157 rxi_ConnectionError(struct rx_connection *conn,
5163 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5165 MUTEX_ENTER(&conn->conn_data_lock);
5166 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
5167 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
5168 if (conn->checkReachEvent) {
5169 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
5170 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5171 MUTEX_ENTER(&rx_refcnt_mutex);
5173 MUTEX_EXIT(&rx_refcnt_mutex);
5175 MUTEX_EXIT(&conn->conn_data_lock);
5176 for (i = 0; i < RX_MAXCALLS; i++) {
5177 struct rx_call *call = conn->call[i];
5179 MUTEX_ENTER(&call->lock);
5180 rxi_CallError(call, error);
5181 MUTEX_EXIT(&call->lock);
5184 conn->error = error;
5185 if (rx_stats_active)
5186 rx_atomic_inc(&rx_stats.fatalErrors);
5191 * Interrupt an in-progress call with the specified error and wakeup waiters.
5193 * @param[in] call The call to interrupt
5194 * @param[in] error The error code to send to the peer
5197 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5199 MUTEX_ENTER(&call->lock);
5200 rxi_CallError(call, error);
5201 rxi_SendCallAbort(call, NULL, 0, 1);
5202 MUTEX_EXIT(&call->lock);
5206 rxi_CallError(struct rx_call *call, afs_int32 error)
5209 osirx_AssertMine(&call->lock, "rxi_CallError");
5211 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5213 error = call->error;
5215 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5216 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5217 rxi_ResetCall(call, 0);
5220 rxi_ResetCall(call, 0);
5222 call->error = error;
5225 /* Reset various fields in a call structure, and wakeup waiting
5226 * processes. Some fields aren't changed: state & mode are not
5227 * touched (these must be set by the caller), and bufptr, nLeft, and
5228 * nFree are not reset, since these fields are manipulated by
5229 * unprotected macros, and may only be reset by non-interrupting code.
5233 rxi_ResetCall(struct rx_call *call, int newcall)
5236 struct rx_peer *peer;
5237 struct rx_packet *packet;
5239 osirx_AssertMine(&call->lock, "rxi_ResetCall");
5241 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5243 /* Notify anyone who is waiting for asynchronous packet arrival */
5244 if (call->arrivalProc) {
5245 (*call->arrivalProc) (call, call->arrivalProcHandle,
5246 call->arrivalProcArg);
5247 call->arrivalProc = (void (*)())0;
5251 rxevent_Cancel(&call->growMTUEvent, call, RX_CALL_REFCOUNT_ALIVE);
5253 if (call->delayedAbortEvent) {
5254 rxevent_Cancel(&call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
5255 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5257 rxi_SendCallAbort(call, packet, 0, 1);
5258 rxi_FreePacket(packet);
5263 * Update the peer with the congestion information in this call
5264 * so other calls on this connection can pick up where this call
5265 * left off. If the congestion sequence numbers don't match then
5266 * another call experienced a retransmission.
5268 peer = call->conn->peer;
5269 MUTEX_ENTER(&peer->peer_lock);
5271 if (call->congestSeq == peer->congestSeq) {
5272 peer->cwind = MAX(peer->cwind, call->cwind);
5273 peer->MTU = MAX(peer->MTU, call->MTU);
5274 peer->nDgramPackets =
5275 MAX(peer->nDgramPackets, call->nDgramPackets);
5278 call->abortCode = 0;
5279 call->abortCount = 0;
5281 if (peer->maxDgramPackets > 1) {
5282 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5284 call->MTU = peer->MTU;
5286 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5287 call->ssthresh = rx_maxSendWindow;
5288 call->nDgramPackets = peer->nDgramPackets;
5289 call->congestSeq = peer->congestSeq;
5290 call->rtt = peer->rtt;
5291 call->rtt_dev = peer->rtt_dev;
5292 clock_Zero(&call->rto);
5293 clock_Addmsec(&call->rto,
5294 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5295 MUTEX_EXIT(&peer->peer_lock);
5297 flags = call->flags;
5298 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5299 rxi_WaitforTQBusy(call);
5300 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5302 rxi_ClearTransmitQueue(call, 1);
5303 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5304 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5308 if ((flags & RX_CALL_PEER_BUSY)) {
5309 /* The call channel is still busy; resetting the call doesn't change
5311 call->flags |= RX_CALL_PEER_BUSY;
5314 rxi_ClearReceiveQueue(call);
5315 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5319 call->twind = call->conn->twind[call->channel];
5320 call->rwind = call->conn->rwind[call->channel];
5321 call->nSoftAcked = 0;
5322 call->nextCwind = 0;
5325 call->nCwindAcks = 0;
5326 call->nSoftAcks = 0;
5327 call->nHardAcks = 0;
5329 call->tfirst = call->rnext = call->tnext = 1;
5332 call->lastAcked = 0;
5333 call->localStatus = call->remoteStatus = 0;
5335 if (flags & RX_CALL_READER_WAIT) {
5336 #ifdef RX_ENABLE_LOCKS
5337 CV_BROADCAST(&call->cv_rq);
5339 osi_rxWakeup(&call->rq);
5342 if (flags & RX_CALL_WAIT_PACKETS) {
5343 MUTEX_ENTER(&rx_freePktQ_lock);
5344 rxi_PacketsUnWait(); /* XXX */
5345 MUTEX_EXIT(&rx_freePktQ_lock);
5347 #ifdef RX_ENABLE_LOCKS
5348 CV_SIGNAL(&call->cv_twind);
5350 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5351 osi_rxWakeup(&call->twind);
5354 #ifdef RX_ENABLE_LOCKS
5355 /* The following ensures that we don't mess with any queue while some
5356 * other thread might also be doing so. The call_queue_lock field is
5357 * is only modified under the call lock. If the call is in the process
5358 * of being removed from a queue, the call is not locked until the
5359 * the queue lock is dropped and only then is the call_queue_lock field
5360 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5361 * Note that any other routine which removes a call from a queue has to
5362 * obtain the queue lock before examing the queue and removing the call.
5364 if (call->call_queue_lock) {
5365 MUTEX_ENTER(call->call_queue_lock);
5366 if (queue_IsOnQueue(call)) {
5368 if (flags & RX_CALL_WAIT_PROC) {
5369 rx_atomic_dec(&rx_nWaiting);
5372 MUTEX_EXIT(call->call_queue_lock);
5373 CLEAR_CALL_QUEUE_LOCK(call);
5375 #else /* RX_ENABLE_LOCKS */
5376 if (queue_IsOnQueue(call)) {
5378 if (flags & RX_CALL_WAIT_PROC)
5379 rx_atomic_dec(&rx_nWaiting);
5381 #endif /* RX_ENABLE_LOCKS */
5383 rxi_KeepAliveOff(call);
5384 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5387 /* Send an acknowledge for the indicated packet (seq,serial) of the
5388 * indicated call, for the indicated reason (reason). This
5389 * acknowledge will specifically acknowledge receiving the packet, and
5390 * will also specify which other packets for this call have been
5391 * received. This routine returns the packet that was used to the
5392 * caller. The caller is responsible for freeing it or re-using it.
5393 * This acknowledgement also returns the highest sequence number
5394 * actually read out by the higher level to the sender; the sender
5395 * promises to keep around packets that have not been read by the
5396 * higher level yet (unless, of course, the sender decides to abort
5397 * the call altogether). Any of p, seq, serial, pflags, or reason may
5398 * be set to zero without ill effect. That is, if they are zero, they
5399 * will not convey any information.
5400 * NOW there is a trailer field, after the ack where it will safely be
5401 * ignored by mundanes, which indicates the maximum size packet this
5402 * host can swallow. */
5404 struct rx_packet *optionalPacket; use to send ack (or null)
5405 int seq; Sequence number of the packet we are acking
5406 int serial; Serial number of the packet
5407 int pflags; Flags field from packet header
5408 int reason; Reason an acknowledge was prompted
5412 rxi_SendAck(struct rx_call *call,
5413 struct rx_packet *optionalPacket, int serial, int reason,
5416 struct rx_ackPacket *ap;
5417 struct rx_packet *rqp;
5418 struct rx_packet *nxp; /* For queue_Scan */
5419 struct rx_packet *p;
5422 afs_uint32 padbytes = 0;
5423 #ifdef RX_ENABLE_TSFPQ
5424 struct rx_ts_info_t * rx_ts_info;
5428 * Open the receive window once a thread starts reading packets
5430 if (call->rnext > 1) {
5431 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5434 /* Don't attempt to grow MTU if this is a critical ping */
5435 if (reason == RX_ACK_MTU) {
5436 /* keep track of per-call attempts, if we're over max, do in small
5437 * otherwise in larger? set a size to increment by, decrease
5440 if (call->conn->peer->maxPacketSize &&
5441 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5443 padbytes = call->conn->peer->maxPacketSize+16;
5445 padbytes = call->conn->peer->maxMTU + 128;
5447 /* do always try a minimum size ping */
5448 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5450 /* subtract the ack payload */
5451 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5452 reason = RX_ACK_PING;
5455 call->nHardAcks = 0;
5456 call->nSoftAcks = 0;
5457 if (call->rnext > call->lastAcked)
5458 call->lastAcked = call->rnext;
5462 rx_computelen(p, p->length); /* reset length, you never know */
5463 } /* where that's been... */
5464 #ifdef RX_ENABLE_TSFPQ
5466 RX_TS_INFO_GET(rx_ts_info);
5467 if ((p = rx_ts_info->local_special_packet)) {
5468 rx_computelen(p, p->length);
5469 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5470 rx_ts_info->local_special_packet = p;
5471 } else { /* We won't send the ack, but don't panic. */
5472 return optionalPacket;
5476 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5477 /* We won't send the ack, but don't panic. */
5478 return optionalPacket;
5483 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5486 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5487 #ifndef RX_ENABLE_TSFPQ
5488 if (!optionalPacket)
5491 return optionalPacket;
5493 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5494 if (rx_Contiguous(p) < templ) {
5495 #ifndef RX_ENABLE_TSFPQ
5496 if (!optionalPacket)
5499 return optionalPacket;
5504 /* MTUXXX failing to send an ack is very serious. We should */
5505 /* try as hard as possible to send even a partial ack; it's */
5506 /* better than nothing. */
5507 ap = (struct rx_ackPacket *)rx_DataOf(p);
5508 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5509 ap->reason = reason;
5511 /* The skew computation used to be bogus, I think it's better now. */
5512 /* We should start paying attention to skew. XXX */
5513 ap->serial = htonl(serial);
5514 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5517 * First packet not yet forwarded to reader. When ACKALL has been
5518 * sent the peer has been told that all received packets will be
5519 * delivered to the reader. The value 'rnext' is used internally
5520 * to refer to the next packet in the receive queue that must be
5521 * delivered to the reader. From the perspective of the peer it
5522 * already has so report the last sequence number plus one if there
5523 * are packets in the receive queue awaiting processing.
5525 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5526 !queue_IsEmpty(&call->rq)) {
5527 ap->firstPacket = htonl(queue_Last(&call->rq, rx_packet)->header.seq + 1);
5529 ap->firstPacket = htonl(call->rnext);
5531 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5533 /* No fear of running out of ack packet here because there can only be at most
5534 * one window full of unacknowledged packets. The window size must be constrained
5535 * to be less than the maximum ack size, of course. Also, an ack should always
5536 * fit into a single packet -- it should not ever be fragmented. */
5537 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5538 if (!rqp || !call->rq.next
5539 || (rqp->header.seq > (call->rnext + call->rwind))) {
5540 #ifndef RX_ENABLE_TSFPQ
5541 if (!optionalPacket)
5544 rxi_CallError(call, RX_CALL_DEAD);
5545 return optionalPacket;
5548 while (rqp->header.seq > call->rnext + offset)
5549 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5550 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5552 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5553 #ifndef RX_ENABLE_TSFPQ
5554 if (!optionalPacket)
5557 rxi_CallError(call, RX_CALL_DEAD);
5558 return optionalPacket;
5564 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5566 /* these are new for AFS 3.3 */
5567 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5568 templ = htonl(templ);
5569 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5570 templ = htonl(call->conn->peer->ifMTU);
5571 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5572 sizeof(afs_int32), &templ);
5574 /* new for AFS 3.4 */
5575 templ = htonl(call->rwind);
5576 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5577 sizeof(afs_int32), &templ);
5579 /* new for AFS 3.5 */
5580 templ = htonl(call->conn->peer->ifDgramPackets);
5581 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5582 sizeof(afs_int32), &templ);
5584 p->header.serviceId = call->conn->serviceId;
5585 p->header.cid = (call->conn->cid | call->channel);
5586 p->header.callNumber = *call->callNumber;
5588 p->header.securityIndex = call->conn->securityIndex;
5589 p->header.epoch = call->conn->epoch;
5590 p->header.type = RX_PACKET_TYPE_ACK;
5591 p->header.flags = RX_SLOW_START_OK;
5592 if (reason == RX_ACK_PING) {
5593 p->header.flags |= RX_REQUEST_ACK;
5595 p->length = padbytes +
5596 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5599 /* not fast but we can potentially use this if truncated
5600 * fragments are delivered to figure out the mtu.
5602 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5603 sizeof(afs_int32), sizeof(afs_int32),
5607 if (call->conn->type == RX_CLIENT_CONNECTION)
5608 p->header.flags |= RX_CLIENT_INITIATED;
5612 if (rxdebug_active) {
5616 len = _snprintf(msg, sizeof(msg),
5617 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5618 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5619 ntohl(ap->serial), ntohl(ap->previousPacket),
5620 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5621 ap->nAcks, ntohs(ap->bufferSpace) );
5625 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5626 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5630 OutputDebugString(msg);
5632 #else /* AFS_NT40_ENV */
5634 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5635 ap->reason, ntohl(ap->previousPacket),
5636 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5638 for (offset = 0; offset < ap->nAcks; offset++)
5639 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5644 #endif /* AFS_NT40_ENV */
5647 int i, nbytes = p->length;
5649 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5650 if (nbytes <= p->wirevec[i].iov_len) {
5653 savelen = p->wirevec[i].iov_len;
5655 p->wirevec[i].iov_len = nbytes;
5657 rxi_Send(call, p, istack);
5658 p->wirevec[i].iov_len = savelen;
5662 nbytes -= p->wirevec[i].iov_len;
5665 if (rx_stats_active)
5666 rx_atomic_inc(&rx_stats.ackPacketsSent);
5667 #ifndef RX_ENABLE_TSFPQ
5668 if (!optionalPacket)
5671 return optionalPacket; /* Return packet for re-use by caller */
5675 struct rx_packet **list;
5680 /* Send all of the packets in the list in single datagram */
5682 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5683 int istack, int moreFlag)
5689 struct rx_connection *conn = call->conn;
5690 struct rx_peer *peer = conn->peer;
5692 MUTEX_ENTER(&peer->peer_lock);
5693 peer->nSent += xmit->len;
5694 if (xmit->resending)
5695 peer->reSends += xmit->len;
5696 MUTEX_EXIT(&peer->peer_lock);
5698 if (rx_stats_active) {
5699 if (xmit->resending)
5700 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5702 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5705 clock_GetTime(&now);
5707 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5711 /* Set the packet flags and schedule the resend events */
5712 /* Only request an ack for the last packet in the list */
5713 for (i = 0; i < xmit->len; i++) {
5714 struct rx_packet *packet = xmit->list[i];
5716 /* Record the time sent */
5717 packet->timeSent = now;
5718 packet->flags |= RX_PKTFLAG_SENT;
5720 /* Ask for an ack on retransmitted packets, on every other packet
5721 * if the peer doesn't support slow start. Ask for an ack on every
5722 * packet until the congestion window reaches the ack rate. */
5723 if (packet->header.serial) {
5726 packet->firstSent = now;
5727 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5728 || (!(call->flags & RX_CALL_SLOW_START_OK)
5729 && (packet->header.seq & 1)))) {
5734 /* Tag this packet as not being the last in this group,
5735 * for the receiver's benefit */
5736 if (i < xmit->len - 1 || moreFlag) {
5737 packet->header.flags |= RX_MORE_PACKETS;
5742 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5745 /* Since we're about to send a data packet to the peer, it's
5746 * safe to nuke any scheduled end-of-packets ack */
5747 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5749 MUTEX_EXIT(&call->lock);
5750 MUTEX_ENTER(&rx_refcnt_mutex);
5751 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5752 MUTEX_EXIT(&rx_refcnt_mutex);
5753 if (xmit->len > 1) {
5754 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5756 rxi_SendPacket(call, conn, xmit->list[0], istack);
5758 MUTEX_ENTER(&call->lock);
5759 MUTEX_ENTER(&rx_refcnt_mutex);
5760 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5761 MUTEX_EXIT(&rx_refcnt_mutex);
5763 /* Tell the RTO calculation engine that we have sent a packet, and
5764 * if it was the last one */
5765 rxi_rto_packet_sent(call, lastPacket, istack);
5767 /* Update last send time for this call (for keep-alive
5768 * processing), and for the connection (so that we can discover
5769 * idle connections) */
5770 conn->lastSendTime = call->lastSendTime = clock_Sec();
5771 /* Let a set of retransmits trigger an idle timeout */
5772 if (!xmit->resending)
5773 call->lastSendData = call->lastSendTime;
5776 /* When sending packets we need to follow these rules:
5777 * 1. Never send more than maxDgramPackets in a jumbogram.
5778 * 2. Never send a packet with more than two iovecs in a jumbogram.
5779 * 3. Never send a retransmitted packet in a jumbogram.
5780 * 4. Never send more than cwind/4 packets in a jumbogram
5781 * We always keep the last list we should have sent so we
5782 * can set the RX_MORE_PACKETS flags correctly.
5786 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5791 struct xmitlist working;
5792 struct xmitlist last;
5794 struct rx_peer *peer = call->conn->peer;
5795 int morePackets = 0;
5797 memset(&last, 0, sizeof(struct xmitlist));
5798 working.list = &list[0];
5800 working.resending = 0;
5802 recovery = call->flags & RX_CALL_FAST_RECOVER;
5804 for (i = 0; i < len; i++) {
5805 /* Does the current packet force us to flush the current list? */
5807 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5808 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5810 /* This sends the 'last' list and then rolls the current working
5811 * set into the 'last' one, and resets the working set */
5814 rxi_SendList(call, &last, istack, 1);
5815 /* If the call enters an error state stop sending, or if
5816 * we entered congestion recovery mode, stop sending */
5818 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5823 working.resending = 0;
5824 working.list = &list[i];
5826 /* Add the current packet to the list if it hasn't been acked.
5827 * Otherwise adjust the list pointer to skip the current packet. */
5828 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5831 if (list[i]->header.serial)
5832 working.resending = 1;
5834 /* Do we need to flush the list? */
5835 if (working.len >= (int)peer->maxDgramPackets
5836 || working.len >= (int)call->nDgramPackets
5837 || working.len >= (int)call->cwind
5838 || list[i]->header.serial
5839 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5841 rxi_SendList(call, &last, istack, 1);
5842 /* If the call enters an error state stop sending, or if
5843 * we entered congestion recovery mode, stop sending */
5845 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5850 working.resending = 0;
5851 working.list = &list[i + 1];
5854 if (working.len != 0) {
5855 osi_Panic("rxi_SendList error");
5857 working.list = &list[i + 1];
5861 /* Send the whole list when the call is in receive mode, when
5862 * the call is in eof mode, when we are in fast recovery mode,
5863 * and when we have the last packet */
5864 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5865 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5866 || (call->flags & RX_CALL_FAST_RECOVER)) {
5867 /* Check for the case where the current list contains
5868 * an acked packet. Since we always send retransmissions
5869 * in a separate packet, we only need to check the first
5870 * packet in the list */
5871 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5875 rxi_SendList(call, &last, istack, morePackets);
5876 /* If the call enters an error state stop sending, or if
5877 * we entered congestion recovery mode, stop sending */
5879 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5883 rxi_SendList(call, &working, istack, 0);
5885 } else if (last.len > 0) {
5886 rxi_SendList(call, &last, istack, 0);
5887 /* Packets which are in 'working' are not sent by this call */
5892 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
5894 struct rx_call *call = arg0;
5895 struct rx_peer *peer;
5896 struct rx_packet *p, *nxp;
5897 struct clock maxTimeout = { 60, 0 };
5899 MUTEX_ENTER(&call->lock);
5901 peer = call->conn->peer;
5903 /* Make sure that the event pointer is removed from the call
5904 * structure, since there is no longer a per-call retransmission
5906 if (event == call->resendEvent) {
5907 MUTEX_ENTER(&rx_refcnt_mutex);
5908 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5909 MUTEX_EXIT(&rx_refcnt_mutex);
5910 rxevent_Put(call->resendEvent);
5911 call->resendEvent = NULL;
5914 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
5915 rxi_CheckBusy(call);
5918 if (queue_IsEmpty(&call->tq)) {
5919 /* Nothing to do. This means that we've been raced, and that an
5920 * ACK has come in between when we were triggered, and when we
5921 * actually got to run. */
5925 /* We're in loss recovery */
5926 call->flags |= RX_CALL_FAST_RECOVER;
5928 /* Mark all of the pending packets in the queue as being lost */
5929 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5930 if (!(p->flags & RX_PKTFLAG_ACKED))
5931 p->flags &= ~RX_PKTFLAG_SENT;
5934 /* We're resending, so we double the timeout of the call. This will be
5935 * dropped back down by the first successful ACK that we receive.
5937 * We apply a maximum value here of 60 seconds
5939 clock_Add(&call->rto, &call->rto);
5940 if (clock_Gt(&call->rto, &maxTimeout))
5941 call->rto = maxTimeout;
5943 /* Packet loss is most likely due to congestion, so drop our window size
5944 * and start again from the beginning */
5945 if (peer->maxDgramPackets >1) {
5946 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5947 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5949 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5950 call->nDgramPackets = 1;
5952 call->nextCwind = 1;
5955 MUTEX_ENTER(&peer->peer_lock);
5956 peer->MTU = call->MTU;
5957 peer->cwind = call->cwind;
5958 peer->nDgramPackets = 1;
5960 call->congestSeq = peer->congestSeq;
5961 MUTEX_EXIT(&peer->peer_lock);
5963 rxi_Start(call, istack);
5966 MUTEX_EXIT(&call->lock);
5969 /* This routine is called when new packets are readied for
5970 * transmission and when retransmission may be necessary, or when the
5971 * transmission window or burst count are favourable. This should be
5972 * better optimized for new packets, the usual case, now that we've
5973 * got rid of queues of send packets. XXXXXXXXXXX */
5975 rxi_Start(struct rx_call *call, int istack)
5978 struct rx_packet *p;
5979 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5984 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5985 if (rx_stats_active)
5986 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5991 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5993 /* Send (or resend) any packets that need it, subject to
5994 * window restrictions and congestion burst control
5995 * restrictions. Ask for an ack on the last packet sent in
5996 * this burst. For now, we're relying upon the window being
5997 * considerably bigger than the largest number of packets that
5998 * are typically sent at once by one initial call to
5999 * rxi_Start. This is probably bogus (perhaps we should ask
6000 * for an ack when we're half way through the current
6001 * window?). Also, for non file transfer applications, this
6002 * may end up asking for an ack for every packet. Bogus. XXXX
6005 * But check whether we're here recursively, and let the other guy
6008 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6009 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6010 call->flags |= RX_CALL_TQ_BUSY;
6012 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6014 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6015 call->flags &= ~RX_CALL_NEED_START;
6016 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6018 maxXmitPackets = MIN(call->twind, call->cwind);
6019 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
6020 #ifdef RX_TRACK_PACKETS
6021 if ((p->flags & RX_PKTFLAG_FREE)
6022 || (!queue_IsEnd(&call->tq, nxp)
6023 && (nxp->flags & RX_PKTFLAG_FREE))
6024 || (p == (struct rx_packet *)&rx_freePacketQueue)
6025 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
6026 osi_Panic("rxi_Start: xmit queue clobbered");
6029 if (p->flags & RX_PKTFLAG_ACKED) {
6030 /* Since we may block, don't trust this */
6031 if (rx_stats_active)
6032 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6033 continue; /* Ignore this packet if it has been acknowledged */
6036 /* Turn off all flags except these ones, which are the same
6037 * on each transmission */
6038 p->header.flags &= RX_PRESET_FLAGS;
6040 if (p->header.seq >=
6041 call->tfirst + MIN((int)call->twind,
6042 (int)(call->nSoftAcked +
6044 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6045 /* Note: if we're waiting for more window space, we can
6046 * still send retransmits; hence we don't return here, but
6047 * break out to schedule a retransmit event */
6048 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6049 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6054 /* Transmit the packet if it needs to be sent. */
6055 if (!(p->flags & RX_PKTFLAG_SENT)) {
6056 if (nXmitPackets == maxXmitPackets) {
6057 rxi_SendXmitList(call, call->xmitList,
6058 nXmitPackets, istack);
6061 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6062 *(call->callNumber), p));
6063 call->xmitList[nXmitPackets++] = p;
6067 /* xmitList now hold pointers to all of the packets that are
6068 * ready to send. Now we loop to send the packets */
6069 if (nXmitPackets > 0) {
6070 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6074 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6076 /* We went into the error state while sending packets. Now is
6077 * the time to reset the call. This will also inform the using
6078 * process that the call is in an error state.
6080 if (rx_stats_active)
6081 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6082 call->flags &= ~RX_CALL_TQ_BUSY;
6083 rxi_WakeUpTransmitQueue(call);
6084 rxi_CallError(call, call->error);
6087 #ifdef RX_ENABLE_LOCKS
6088 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6090 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6091 /* Some packets have received acks. If they all have, we can clear
6092 * the transmit queue.
6095 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
6096 if (p->header.seq < call->tfirst
6097 && (p->flags & RX_PKTFLAG_ACKED)) {
6099 #ifdef RX_TRACK_PACKETS
6100 p->flags &= ~RX_PKTFLAG_TQ;
6102 #ifdef RXDEBUG_PACKET
6110 call->flags |= RX_CALL_TQ_CLEARME;
6112 #endif /* RX_ENABLE_LOCKS */
6113 if (call->flags & RX_CALL_TQ_CLEARME)
6114 rxi_ClearTransmitQueue(call, 1);
6115 } while (call->flags & RX_CALL_NEED_START);
6117 * TQ references no longer protected by this flag; they must remain
6118 * protected by the global lock.
6120 call->flags &= ~RX_CALL_TQ_BUSY;
6121 rxi_WakeUpTransmitQueue(call);
6123 call->flags |= RX_CALL_NEED_START;
6125 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6127 rxi_rto_cancel(call);
6131 /* Also adjusts the keep alive parameters for the call, to reflect
6132 * that we have just sent a packet (so keep alives aren't sent
6135 rxi_Send(struct rx_call *call, struct rx_packet *p,
6138 struct rx_connection *conn = call->conn;
6140 /* Stamp each packet with the user supplied status */
6141 p->header.userStatus = call->localStatus;
6143 /* Allow the security object controlling this call's security to
6144 * make any last-minute changes to the packet */
6145 RXS_SendPacket(conn->securityObject, call, p);
6147 /* Since we're about to send SOME sort of packet to the peer, it's
6148 * safe to nuke any scheduled end-of-packets ack */
6149 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
6151 /* Actually send the packet, filling in more connection-specific fields */
6152 MUTEX_EXIT(&call->lock);
6153 MUTEX_ENTER(&rx_refcnt_mutex);
6154 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6155 MUTEX_EXIT(&rx_refcnt_mutex);
6156 rxi_SendPacket(call, conn, p, istack);
6157 MUTEX_ENTER(&rx_refcnt_mutex);
6158 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6159 MUTEX_EXIT(&rx_refcnt_mutex);
6160 MUTEX_ENTER(&call->lock);
6162 /* Update last send time for this call (for keep-alive
6163 * processing), and for the connection (so that we can discover
6164 * idle connections) */
6165 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6166 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6167 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6169 conn->lastSendTime = call->lastSendTime = clock_Sec();
6170 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6171 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6172 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6173 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6174 RX_ACK_PING_RESPONSE)))
6175 call->lastSendData = call->lastSendTime;
6179 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6180 * that things are fine. Also called periodically to guarantee that nothing
6181 * falls through the cracks (e.g. (error + dally) connections have keepalive
6182 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6184 * haveCTLock Set if calling from rxi_ReapConnections
6186 #ifdef RX_ENABLE_LOCKS
6188 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6189 #else /* RX_ENABLE_LOCKS */
6191 rxi_CheckCall(struct rx_call *call)
6192 #endif /* RX_ENABLE_LOCKS */
6194 struct rx_connection *conn = call->conn;
6196 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6197 afs_uint32 fudgeFactor;
6201 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6202 if (call->flags & RX_CALL_TQ_BUSY) {
6203 /* Call is active and will be reset by rxi_Start if it's
6204 * in an error state.
6209 /* RTT + 8*MDEV, rounded up to the next second. */
6210 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6211 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6213 deadTime = conn->secondsUntilDead + fudgeFactor;
6215 /* These are computed to the second (+- 1 second). But that's
6216 * good enough for these values, which should be a significant
6217 * number of seconds. */
6218 if (now > (call->lastReceiveTime + deadTime)) {
6219 if (call->state == RX_STATE_ACTIVE) {
6221 #if defined(KERNEL) && defined(AFS_SUN5_ENV)
6223 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6224 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6225 ip_stack_t *ipst = ns->netstack_ip;
6227 ire = ire_cache_lookup(conn->peer->host
6228 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6230 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
6232 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6239 if (ire && ire->ire_max_frag > 0)
6240 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6242 #if defined(GLOBAL_NETSTACKID)
6246 #endif /* ADAPT_PMTU */
6247 cerror = RX_CALL_DEAD;
6250 #ifdef RX_ENABLE_LOCKS
6251 /* Cancel pending events */
6252 rxevent_Cancel(&call->delayedAckEvent, call,
6253 RX_CALL_REFCOUNT_DELAY);
6254 rxi_rto_cancel(call);
6255 rxevent_Cancel(&call->keepAliveEvent, call,
6256 RX_CALL_REFCOUNT_ALIVE);
6257 rxevent_Cancel(&call->growMTUEvent, call,
6258 RX_CALL_REFCOUNT_ALIVE);
6259 MUTEX_ENTER(&rx_refcnt_mutex);
6260 if (call->refCount == 0) {
6261 rxi_FreeCall(call, haveCTLock);
6262 MUTEX_EXIT(&rx_refcnt_mutex);
6265 MUTEX_EXIT(&rx_refcnt_mutex);
6267 #else /* RX_ENABLE_LOCKS */
6268 rxi_FreeCall(call, 0);
6270 #endif /* RX_ENABLE_LOCKS */
6272 /* Non-active calls are destroyed if they are not responding
6273 * to pings; active calls are simply flagged in error, so the
6274 * attached process can die reasonably gracefully. */
6277 if (conn->idleDeadTime) {
6278 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6281 /* see if we have a non-activity timeout */
6282 if (call->startWait && idleDeadTime
6283 && ((call->startWait + idleDeadTime) < now) &&
6284 (call->flags & RX_CALL_READER_WAIT)) {
6285 if (call->state == RX_STATE_ACTIVE) {
6286 cerror = RX_CALL_TIMEOUT;
6290 if (call->lastSendData && idleDeadTime && (conn->idleDeadErr != 0)
6291 && ((call->lastSendData + idleDeadTime) < now)) {
6292 if (call->state == RX_STATE_ACTIVE) {
6293 cerror = conn->idleDeadErr;
6298 if (conn->hardDeadTime) {
6299 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6302 /* see if we have a hard timeout */
6304 && (now > (hardDeadTime + call->startTime.sec))) {
6305 if (call->state == RX_STATE_ACTIVE)
6306 rxi_CallError(call, RX_CALL_TIMEOUT);
6311 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT
6312 && call->lastReceiveTime) {
6313 int oldMTU = conn->peer->ifMTU;
6315 /* if we thought we could send more, perhaps things got worse */
6316 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6317 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6318 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6319 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6321 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6323 /* minimum capped in SetPeerMtu */
6324 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6327 conn->lastPacketSize = 0;
6329 /* needed so ResetCall doesn't clobber us. */
6330 call->MTU = conn->peer->ifMTU;
6332 /* if we never succeeded, let the error pass out as-is */
6333 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6334 cerror = conn->msgsizeRetryErr;
6337 rxi_CallError(call, cerror);
6342 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6343 void *dummy, int dummy2)
6345 struct rx_connection *conn = arg1;
6346 struct rx_header theader;
6347 char tbuffer[1 + sizeof(struct rx_header)];
6348 struct sockaddr_in taddr;
6351 struct iovec tmpiov[2];
6354 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6357 tp = &tbuffer[sizeof(struct rx_header)];
6358 taddr.sin_family = AF_INET;
6359 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6360 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6361 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6362 taddr.sin_len = sizeof(struct sockaddr_in);
6364 memset(&theader, 0, sizeof(theader));
6365 theader.epoch = htonl(999);
6367 theader.callNumber = 0;
6370 theader.type = RX_PACKET_TYPE_VERSION;
6371 theader.flags = RX_LAST_PACKET;
6372 theader.serviceId = 0;
6374 memcpy(tbuffer, &theader, sizeof(theader));
6375 memcpy(tp, &a, sizeof(a));
6376 tmpiov[0].iov_base = tbuffer;
6377 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6379 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6381 MUTEX_ENTER(&conn->conn_data_lock);
6382 MUTEX_ENTER(&rx_refcnt_mutex);
6383 /* Only reschedule ourselves if the connection would not be destroyed */
6384 if (conn->refCount <= 1) {
6385 rxevent_Put(conn->natKeepAliveEvent);
6386 conn->natKeepAliveEvent = NULL;
6387 MUTEX_EXIT(&rx_refcnt_mutex);
6388 MUTEX_EXIT(&conn->conn_data_lock);
6389 rx_DestroyConnection(conn); /* drop the reference for this */
6391 conn->refCount--; /* drop the reference for this */
6392 MUTEX_EXIT(&rx_refcnt_mutex);
6393 rxevent_Put(conn->natKeepAliveEvent);
6394 conn->natKeepAliveEvent = NULL;
6395 rxi_ScheduleNatKeepAliveEvent(conn);
6396 MUTEX_EXIT(&conn->conn_data_lock);
6401 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6403 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6404 struct clock when, now;
6405 clock_GetTime(&now);
6407 when.sec += conn->secondsUntilNatPing;
6408 MUTEX_ENTER(&rx_refcnt_mutex);
6409 conn->refCount++; /* hold a reference for this */
6410 MUTEX_EXIT(&rx_refcnt_mutex);
6411 conn->natKeepAliveEvent =
6412 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6417 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6419 MUTEX_ENTER(&conn->conn_data_lock);
6420 conn->secondsUntilNatPing = seconds;
6422 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6423 rxi_ScheduleNatKeepAliveEvent(conn);
6425 conn->flags |= RX_CONN_NAT_PING;
6427 MUTEX_EXIT(&conn->conn_data_lock);
6431 rxi_NatKeepAliveOn(struct rx_connection *conn)
6433 MUTEX_ENTER(&conn->conn_data_lock);
6434 /* if it's already attached */
6435 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6436 rxi_ScheduleNatKeepAliveEvent(conn);
6438 conn->flags |= RX_CONN_NAT_PING;
6439 MUTEX_EXIT(&conn->conn_data_lock);
6442 /* When a call is in progress, this routine is called occasionally to
6443 * make sure that some traffic has arrived (or been sent to) the peer.
6444 * If nothing has arrived in a reasonable amount of time, the call is
6445 * declared dead; if nothing has been sent for a while, we send a
6446 * keep-alive packet (if we're actually trying to keep the call alive)
6449 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6452 struct rx_call *call = arg1;
6453 struct rx_connection *conn;
6456 MUTEX_ENTER(&rx_refcnt_mutex);
6457 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6458 MUTEX_EXIT(&rx_refcnt_mutex);
6459 MUTEX_ENTER(&call->lock);
6461 if (event == call->keepAliveEvent) {
6462 rxevent_Put(call->keepAliveEvent);
6463 call->keepAliveEvent = NULL;
6468 #ifdef RX_ENABLE_LOCKS
6469 if (rxi_CheckCall(call, 0)) {
6470 MUTEX_EXIT(&call->lock);
6473 #else /* RX_ENABLE_LOCKS */
6474 if (rxi_CheckCall(call))
6476 #endif /* RX_ENABLE_LOCKS */
6478 /* Don't try to keep alive dallying calls */
6479 if (call->state == RX_STATE_DALLY) {
6480 MUTEX_EXIT(&call->lock);
6485 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6486 /* Don't try to send keepalives if there is unacknowledged data */
6487 /* the rexmit code should be good enough, this little hack
6488 * doesn't quite work XXX */
6489 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6491 rxi_ScheduleKeepAliveEvent(call);
6492 MUTEX_EXIT(&call->lock);
6495 /* Does what's on the nameplate. */
6497 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6499 struct rx_call *call = arg1;
6500 struct rx_connection *conn;
6502 MUTEX_ENTER(&rx_refcnt_mutex);
6503 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6504 MUTEX_EXIT(&rx_refcnt_mutex);
6505 MUTEX_ENTER(&call->lock);
6507 if (event == call->growMTUEvent) {
6508 rxevent_Put(call->growMTUEvent);
6509 call->growMTUEvent = NULL;
6512 #ifdef RX_ENABLE_LOCKS
6513 if (rxi_CheckCall(call, 0)) {
6514 MUTEX_EXIT(&call->lock);
6517 #else /* RX_ENABLE_LOCKS */
6518 if (rxi_CheckCall(call))
6520 #endif /* RX_ENABLE_LOCKS */
6522 /* Don't bother with dallying calls */
6523 if (call->state == RX_STATE_DALLY) {
6524 MUTEX_EXIT(&call->lock);
6531 * keep being scheduled, just don't do anything if we're at peak,
6532 * or we're not set up to be properly handled (idle timeout required)
6534 if ((conn->peer->maxPacketSize != 0) &&
6535 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6536 (conn->idleDeadErr))
6537 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6538 rxi_ScheduleGrowMTUEvent(call, 0);
6539 MUTEX_EXIT(&call->lock);
6543 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6545 if (!call->keepAliveEvent) {
6546 struct clock when, now;
6547 clock_GetTime(&now);
6549 when.sec += call->conn->secondsUntilPing;
6550 MUTEX_ENTER(&rx_refcnt_mutex);
6551 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6552 MUTEX_EXIT(&rx_refcnt_mutex);
6553 call->keepAliveEvent =
6554 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6559 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6561 if (!call->growMTUEvent) {
6562 struct clock when, now;
6564 clock_GetTime(&now);
6567 if (call->conn->secondsUntilPing)
6568 secs = (6*call->conn->secondsUntilPing)-1;
6570 if (call->conn->secondsUntilDead)
6571 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6575 MUTEX_ENTER(&rx_refcnt_mutex);
6576 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6577 MUTEX_EXIT(&rx_refcnt_mutex);
6578 call->growMTUEvent =
6579 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6583 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6585 rxi_KeepAliveOn(struct rx_call *call)
6587 /* Pretend last packet received was received now--i.e. if another
6588 * packet isn't received within the keep alive time, then the call
6589 * will die; Initialize last send time to the current time--even
6590 * if a packet hasn't been sent yet. This will guarantee that a
6591 * keep-alive is sent within the ping time */
6592 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6593 rxi_ScheduleKeepAliveEvent(call);
6597 rxi_GrowMTUOn(struct rx_call *call)
6599 struct rx_connection *conn = call->conn;
6600 MUTEX_ENTER(&conn->conn_data_lock);
6601 conn->lastPingSizeSer = conn->lastPingSize = 0;
6602 MUTEX_EXIT(&conn->conn_data_lock);
6603 rxi_ScheduleGrowMTUEvent(call, 1);
6606 /* This routine is called to send connection abort messages
6607 * that have been delayed to throttle looping clients. */
6609 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6612 struct rx_connection *conn = arg1;
6615 struct rx_packet *packet;
6617 MUTEX_ENTER(&conn->conn_data_lock);
6618 rxevent_Put(conn->delayedAbortEvent);
6619 conn->delayedAbortEvent = NULL;
6620 error = htonl(conn->error);
6622 MUTEX_EXIT(&conn->conn_data_lock);
6623 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6626 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6627 RX_PACKET_TYPE_ABORT, (char *)&error,
6629 rxi_FreePacket(packet);
6633 /* This routine is called to send call abort messages
6634 * that have been delayed to throttle looping clients. */
6636 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6639 struct rx_call *call = arg1;
6642 struct rx_packet *packet;
6644 MUTEX_ENTER(&call->lock);
6645 rxevent_Put(call->delayedAbortEvent);
6646 call->delayedAbortEvent = NULL;
6647 error = htonl(call->error);
6649 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6652 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6653 (char *)&error, sizeof(error), 0);
6654 rxi_FreePacket(packet);
6656 MUTEX_EXIT(&call->lock);
6657 MUTEX_ENTER(&rx_refcnt_mutex);
6658 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6659 MUTEX_EXIT(&rx_refcnt_mutex);
6662 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6663 * seconds) to ask the client to authenticate itself. The routine
6664 * issues a challenge to the client, which is obtained from the
6665 * security object associated with the connection */
6667 rxi_ChallengeEvent(struct rxevent *event,
6668 void *arg0, void *arg1, int tries)
6670 struct rx_connection *conn = arg0;
6673 rxevent_Put(conn->challengeEvent);
6674 conn->challengeEvent = NULL;
6677 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6678 struct rx_packet *packet;
6679 struct clock when, now;
6682 /* We've failed to authenticate for too long.
6683 * Reset any calls waiting for authentication;
6684 * they are all in RX_STATE_PRECALL.
6688 MUTEX_ENTER(&conn->conn_call_lock);
6689 for (i = 0; i < RX_MAXCALLS; i++) {
6690 struct rx_call *call = conn->call[i];
6692 MUTEX_ENTER(&call->lock);
6693 if (call->state == RX_STATE_PRECALL) {
6694 rxi_CallError(call, RX_CALL_DEAD);
6695 rxi_SendCallAbort(call, NULL, 0, 0);
6697 MUTEX_EXIT(&call->lock);
6700 MUTEX_EXIT(&conn->conn_call_lock);
6704 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6706 /* If there's no packet available, do this later. */
6707 RXS_GetChallenge(conn->securityObject, conn, packet);
6708 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6709 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6710 rxi_FreePacket(packet);
6712 clock_GetTime(&now);
6714 when.sec += RX_CHALLENGE_TIMEOUT;
6715 conn->challengeEvent =
6716 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6721 /* Call this routine to start requesting the client to authenticate
6722 * itself. This will continue until authentication is established,
6723 * the call times out, or an invalid response is returned. The
6724 * security object associated with the connection is asked to create
6725 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6726 * defined earlier. */
6728 rxi_ChallengeOn(struct rx_connection *conn)
6730 if (!conn->challengeEvent) {
6731 RXS_CreateChallenge(conn->securityObject, conn);
6732 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6737 /* rxi_ComputeRoundTripTime is called with peer locked. */
6738 /* peer may be null */
6740 rxi_ComputeRoundTripTime(struct rx_packet *p,
6741 struct rx_ackPacket *ack,
6742 struct rx_call *call,
6743 struct rx_peer *peer,
6746 struct clock thisRtt, *sentp;
6750 /* If the ACK is delayed, then do nothing */
6751 if (ack->reason == RX_ACK_DELAY)
6754 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6755 * their RTT multiple times, so only include the RTT of the last packet
6757 if (p->flags & RX_JUMBO_PACKET)
6760 /* Use the serial number to determine which transmission the ACK is for,
6761 * and set the sent time to match this. If we have no serial number, then
6762 * only use the ACK for RTT calculations if the packet has not been
6766 serial = ntohl(ack->serial);
6768 if (serial == p->header.serial) {
6769 sentp = &p->timeSent;
6770 } else if (serial == p->firstSerial) {
6771 sentp = &p->firstSent;
6772 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6773 sentp = &p->firstSent;
6777 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6778 sentp = &p->firstSent;
6785 if (clock_Lt(&thisRtt, sentp))
6786 return; /* somebody set the clock back, don't count this time. */
6788 clock_Sub(&thisRtt, sentp);
6789 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6790 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6792 if (clock_IsZero(&thisRtt)) {
6794 * The actual round trip time is shorter than the
6795 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6796 * Since we can't tell which at the moment we will assume 1ms.
6798 thisRtt.usec = 1000;
6801 if (rx_stats_active) {
6802 MUTEX_ENTER(&rx_stats_mutex);
6803 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6804 rx_stats.minRtt = thisRtt;
6805 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6806 if (thisRtt.sec > 60) {
6807 MUTEX_EXIT(&rx_stats_mutex);
6808 return; /* somebody set the clock ahead */
6810 rx_stats.maxRtt = thisRtt;
6812 clock_Add(&rx_stats.totalRtt, &thisRtt);
6813 rx_atomic_inc(&rx_stats.nRttSamples);
6814 MUTEX_EXIT(&rx_stats_mutex);
6817 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6819 /* Apply VanJacobson round-trip estimations */
6824 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6825 * srtt is stored as fixed point with 3 bits after the binary
6826 * point (i.e., scaled by 8). The following magic is
6827 * equivalent to the smoothing algorithm in rfc793 with an
6828 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6829 * srtt'*8 = rtt + srtt*7
6830 * srtt'*8 = srtt*8 + rtt - srtt
6831 * srtt' = srtt + rtt/8 - srtt/8
6832 * srtt' = srtt + (rtt - srtt)/8
6835 delta = _8THMSEC(&thisRtt) - call->rtt;
6836 call->rtt += (delta >> 3);
6839 * We accumulate a smoothed rtt variance (actually, a smoothed
6840 * mean difference), then set the retransmit timer to smoothed
6841 * rtt + 4 times the smoothed variance (was 2x in van's original
6842 * paper, but 4x works better for me, and apparently for him as
6844 * rttvar is stored as
6845 * fixed point with 2 bits after the binary point (scaled by
6846 * 4). The following is equivalent to rfc793 smoothing with
6847 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6848 * rttvar'*4 = rttvar*3 + |delta|
6849 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6850 * rttvar' = rttvar + |delta|/4 - rttvar/4
6851 * rttvar' = rttvar + (|delta| - rttvar)/4
6852 * This replaces rfc793's wired-in beta.
6853 * dev*4 = dev*4 + (|actual - expected| - dev)
6859 delta -= (call->rtt_dev << 1);
6860 call->rtt_dev += (delta >> 3);
6862 /* I don't have a stored RTT so I start with this value. Since I'm
6863 * probably just starting a call, and will be pushing more data down
6864 * this, I expect congestion to increase rapidly. So I fudge a
6865 * little, and I set deviance to half the rtt. In practice,
6866 * deviance tends to approach something a little less than
6867 * half the smoothed rtt. */
6868 call->rtt = _8THMSEC(&thisRtt) + 8;
6869 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6871 /* the smoothed RTT time is RTT + 4*MDEV
6873 * We allow a user specified minimum to be set for this, to allow clamping
6874 * at a minimum value in the same way as TCP. In addition, we have to allow
6875 * for the possibility that this packet is answered by a delayed ACK, so we
6876 * add on a fixed 200ms to account for that timer expiring.
6879 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6880 rx_minPeerTimeout) + 200;
6881 clock_Zero(&call->rto);
6882 clock_Addmsec(&call->rto, rtt_timeout);
6884 /* Update the peer, so any new calls start with our values */
6885 peer->rtt_dev = call->rtt_dev;
6886 peer->rtt = call->rtt;
6888 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6889 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6893 /* Find all server connections that have not been active for a long time, and
6896 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
6899 struct clock now, when;
6900 clock_GetTime(&now);
6902 /* Find server connection structures that haven't been used for
6903 * greater than rx_idleConnectionTime */
6905 struct rx_connection **conn_ptr, **conn_end;
6906 int i, havecalls = 0;
6907 MUTEX_ENTER(&rx_connHashTable_lock);
6908 for (conn_ptr = &rx_connHashTable[0], conn_end =
6909 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6911 struct rx_connection *conn, *next;
6912 struct rx_call *call;
6916 for (conn = *conn_ptr; conn; conn = next) {
6917 /* XXX -- Shouldn't the connection be locked? */
6920 for (i = 0; i < RX_MAXCALLS; i++) {
6921 call = conn->call[i];
6925 code = MUTEX_TRYENTER(&call->lock);
6928 #ifdef RX_ENABLE_LOCKS
6929 result = rxi_CheckCall(call, 1);
6930 #else /* RX_ENABLE_LOCKS */
6931 result = rxi_CheckCall(call);
6932 #endif /* RX_ENABLE_LOCKS */
6933 MUTEX_EXIT(&call->lock);
6935 /* If CheckCall freed the call, it might
6936 * have destroyed the connection as well,
6937 * which screws up the linked lists.
6943 if (conn->type == RX_SERVER_CONNECTION) {
6944 /* This only actually destroys the connection if
6945 * there are no outstanding calls */
6946 MUTEX_ENTER(&conn->conn_data_lock);
6947 MUTEX_ENTER(&rx_refcnt_mutex);
6948 if (!havecalls && !conn->refCount
6949 && ((conn->lastSendTime + rx_idleConnectionTime) <
6951 conn->refCount++; /* it will be decr in rx_DestroyConn */
6952 MUTEX_EXIT(&rx_refcnt_mutex);
6953 MUTEX_EXIT(&conn->conn_data_lock);
6954 #ifdef RX_ENABLE_LOCKS
6955 rxi_DestroyConnectionNoLock(conn);
6956 #else /* RX_ENABLE_LOCKS */
6957 rxi_DestroyConnection(conn);
6958 #endif /* RX_ENABLE_LOCKS */
6960 #ifdef RX_ENABLE_LOCKS
6962 MUTEX_EXIT(&rx_refcnt_mutex);
6963 MUTEX_EXIT(&conn->conn_data_lock);
6965 #endif /* RX_ENABLE_LOCKS */
6969 #ifdef RX_ENABLE_LOCKS
6970 while (rx_connCleanup_list) {
6971 struct rx_connection *conn;
6972 conn = rx_connCleanup_list;
6973 rx_connCleanup_list = rx_connCleanup_list->next;
6974 MUTEX_EXIT(&rx_connHashTable_lock);
6975 rxi_CleanupConnection(conn);
6976 MUTEX_ENTER(&rx_connHashTable_lock);
6978 MUTEX_EXIT(&rx_connHashTable_lock);
6979 #endif /* RX_ENABLE_LOCKS */
6982 /* Find any peer structures that haven't been used (haven't had an
6983 * associated connection) for greater than rx_idlePeerTime */
6985 struct rx_peer **peer_ptr, **peer_end;
6989 * Why do we need to hold the rx_peerHashTable_lock across
6990 * the incrementing of peer_ptr since the rx_peerHashTable
6991 * array is not changing? We don't.
6993 * By dropping the lock periodically we can permit other
6994 * activities to be performed while a rxi_ReapConnections
6995 * call is in progress. The goal of reap connections
6996 * is to clean up quickly without causing large amounts
6997 * of contention. Therefore, it is important that global
6998 * mutexes not be held for extended periods of time.
7000 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7001 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7003 struct rx_peer *peer, *next, *prev;
7005 MUTEX_ENTER(&rx_peerHashTable_lock);
7006 for (prev = peer = *peer_ptr; peer; peer = next) {
7008 code = MUTEX_TRYENTER(&peer->peer_lock);
7009 if ((code) && (peer->refCount == 0)
7010 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7011 rx_interface_stat_p rpc_stat, nrpc_stat;
7015 * now know that this peer object is one to be
7016 * removed from the hash table. Once it is removed
7017 * it can't be referenced by other threads.
7018 * Lets remove it first and decrement the struct
7019 * nPeerStructs count.
7021 if (peer == *peer_ptr) {
7027 if (rx_stats_active)
7028 rx_atomic_dec(&rx_stats.nPeerStructs);
7031 * Now if we hold references on 'prev' and 'next'
7032 * we can safely drop the rx_peerHashTable_lock
7033 * while we destroy this 'peer' object.
7039 MUTEX_EXIT(&rx_peerHashTable_lock);
7041 MUTEX_EXIT(&peer->peer_lock);
7042 MUTEX_DESTROY(&peer->peer_lock);
7044 (&peer->rpcStats, rpc_stat, nrpc_stat,
7045 rx_interface_stat)) {
7046 unsigned int num_funcs;
7049 queue_Remove(&rpc_stat->queue_header);
7050 queue_Remove(&rpc_stat->all_peers);
7051 num_funcs = rpc_stat->stats[0].func_total;
7053 sizeof(rx_interface_stat_t) +
7054 rpc_stat->stats[0].func_total *
7055 sizeof(rx_function_entry_v1_t);
7057 rxi_Free(rpc_stat, space);
7059 MUTEX_ENTER(&rx_rpc_stats);
7060 rxi_rpc_peer_stat_cnt -= num_funcs;
7061 MUTEX_EXIT(&rx_rpc_stats);
7066 * Regain the rx_peerHashTable_lock and
7067 * decrement the reference count on 'prev'
7070 MUTEX_ENTER(&rx_peerHashTable_lock);
7077 MUTEX_EXIT(&peer->peer_lock);
7082 MUTEX_EXIT(&rx_peerHashTable_lock);
7086 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7087 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7088 * GC, just below. Really, we shouldn't have to keep moving packets from
7089 * one place to another, but instead ought to always know if we can
7090 * afford to hold onto a packet in its particular use. */
7091 MUTEX_ENTER(&rx_freePktQ_lock);
7092 if (rx_waitingForPackets) {
7093 rx_waitingForPackets = 0;
7094 #ifdef RX_ENABLE_LOCKS
7095 CV_BROADCAST(&rx_waitingForPackets_cv);
7097 osi_rxWakeup(&rx_waitingForPackets);
7100 MUTEX_EXIT(&rx_freePktQ_lock);
7103 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7104 rxevent_Put(rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0));
7108 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7109 * rx.h is sort of strange this is better. This is called with a security
7110 * object before it is discarded. Each connection using a security object has
7111 * its own refcount to the object so it won't actually be freed until the last
7112 * connection is destroyed.
7114 * This is the only rxs module call. A hold could also be written but no one
7118 rxs_Release(struct rx_securityClass *aobj)
7120 return RXS_Close(aobj);
7128 #define TRACE_OPTION_RX_DEBUG 16
7136 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7137 0, KEY_QUERY_VALUE, &parmKey);
7138 if (code != ERROR_SUCCESS)
7141 dummyLen = sizeof(TraceOption);
7142 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7143 (BYTE *) &TraceOption, &dummyLen);
7144 if (code == ERROR_SUCCESS) {
7145 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7147 RegCloseKey (parmKey);
7148 #endif /* AFS_NT40_ENV */
7153 rx_DebugOnOff(int on)
7157 rxdebug_active = on;
7163 rx_StatsOnOff(int on)
7165 rx_stats_active = on;
7169 /* Don't call this debugging routine directly; use dpf */
7171 rxi_DebugPrint(char *format, ...)
7180 va_start(ap, format);
7182 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7185 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7187 OutputDebugString(msg);
7193 va_start(ap, format);
7195 clock_GetTime(&now);
7196 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7197 (unsigned int)now.usec);
7198 vfprintf(rx_Log, format, ap);
7206 * This function is used to process the rx_stats structure that is local
7207 * to a process as well as an rx_stats structure received from a remote
7208 * process (via rxdebug). Therefore, it needs to do minimal version
7212 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7213 afs_int32 freePackets, char version)
7217 if (size != sizeof(struct rx_statistics)) {
7219 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7220 size, sizeof(struct rx_statistics));
7223 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7226 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7227 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7228 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7229 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7230 s->specialPktAllocFailures);
7232 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7233 s->receivePktAllocFailures, s->sendPktAllocFailures,
7234 s->specialPktAllocFailures);
7238 " greedy %u, " "bogusReads %u (last from host %x), "
7239 "noPackets %u, " "noBuffers %u, " "selects %u, "
7240 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7241 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7242 s->selects, s->sendSelects);
7244 fprintf(file, " packets read: ");
7245 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7246 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7248 fprintf(file, "\n");
7251 " other read counters: data %u, " "ack %u, " "dup %u "
7252 "spurious %u " "dally %u\n", s->dataPacketsRead,
7253 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7254 s->ignorePacketDally);
7256 fprintf(file, " packets sent: ");
7257 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7258 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7260 fprintf(file, "\n");
7263 " other send counters: ack %u, " "data %u (not resends), "
7264 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7265 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7266 s->dataPacketsPushed, s->ignoreAckedPacket);
7269 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7270 s->netSendFailures, (int)s->fatalErrors);
7272 if (s->nRttSamples) {
7273 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7274 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7276 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7277 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7281 " %d server connections, " "%d client connections, "
7282 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7283 s->nServerConns, s->nClientConns, s->nPeerStructs,
7284 s->nCallStructs, s->nFreeCallStructs);
7286 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7287 fprintf(file, " %d clock updates\n", clock_nUpdates);
7291 /* for backward compatibility */
7293 rx_PrintStats(FILE * file)
7295 MUTEX_ENTER(&rx_stats_mutex);
7296 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7297 sizeof(rx_stats), rx_nFreePackets,
7299 MUTEX_EXIT(&rx_stats_mutex);
7303 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7305 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
7306 ntohl(peer->host), (int)ntohs(peer->port), (int)peer->burstSize,
7307 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
7310 " Rtt %d, " "total sent %d, " "resent %d\n",
7311 peer->rtt, peer->nSent, peer->reSends);
7314 " Packet size %d, " "max in packet skew %d, "
7315 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
7316 (int)peer->outPacketSkew);
7320 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7322 * This mutex protects the following static variables:
7326 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7327 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7329 #define LOCK_RX_DEBUG
7330 #define UNLOCK_RX_DEBUG
7331 #endif /* AFS_PTHREAD_ENV */
7333 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7335 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7336 u_char type, void *inputData, size_t inputLength,
7337 void *outputData, size_t outputLength)
7339 static afs_int32 counter = 100;
7340 time_t waitTime, waitCount;
7341 struct rx_header theader;
7344 struct timeval tv_now, tv_wake, tv_delta;
7345 struct sockaddr_in taddr, faddr;
7359 tp = &tbuffer[sizeof(struct rx_header)];
7360 taddr.sin_family = AF_INET;
7361 taddr.sin_port = remotePort;
7362 taddr.sin_addr.s_addr = remoteAddr;
7363 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7364 taddr.sin_len = sizeof(struct sockaddr_in);
7367 memset(&theader, 0, sizeof(theader));
7368 theader.epoch = htonl(999);
7370 theader.callNumber = htonl(counter);
7373 theader.type = type;
7374 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7375 theader.serviceId = 0;
7377 memcpy(tbuffer, &theader, sizeof(theader));
7378 memcpy(tp, inputData, inputLength);
7380 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7381 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7383 /* see if there's a packet available */
7384 gettimeofday(&tv_wake, NULL);
7385 tv_wake.tv_sec += waitTime;
7388 FD_SET(socket, &imask);
7389 tv_delta.tv_sec = tv_wake.tv_sec;
7390 tv_delta.tv_usec = tv_wake.tv_usec;
7391 gettimeofday(&tv_now, NULL);
7393 if (tv_delta.tv_usec < tv_now.tv_usec) {
7395 tv_delta.tv_usec += 1000000;
7398 tv_delta.tv_usec -= tv_now.tv_usec;
7400 if (tv_delta.tv_sec < tv_now.tv_sec) {
7404 tv_delta.tv_sec -= tv_now.tv_sec;
7407 code = select(0, &imask, 0, 0, &tv_delta);
7408 #else /* AFS_NT40_ENV */
7409 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7410 #endif /* AFS_NT40_ENV */
7411 if (code == 1 && FD_ISSET(socket, &imask)) {
7412 /* now receive a packet */
7413 faddrLen = sizeof(struct sockaddr_in);
7415 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7416 (struct sockaddr *)&faddr, &faddrLen);
7419 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7420 if (counter == ntohl(theader.callNumber))
7428 /* see if we've timed out */
7436 code -= sizeof(struct rx_header);
7437 if (code > outputLength)
7438 code = outputLength;
7439 memcpy(outputData, tp, code);
7442 #endif /* RXDEBUG */
7445 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7446 afs_uint16 remotePort, struct rx_debugStats * stat,
7447 afs_uint32 * supportedValues)
7449 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7451 struct rx_debugIn in;
7453 *supportedValues = 0;
7454 in.type = htonl(RX_DEBUGI_GETSTATS);
7457 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7458 &in, sizeof(in), stat, sizeof(*stat));
7461 * If the call was successful, fixup the version and indicate
7462 * what contents of the stat structure are valid.
7463 * Also do net to host conversion of fields here.
7467 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7468 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7470 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7471 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7473 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7474 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7476 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7477 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7479 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7480 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7482 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7483 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7485 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7486 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7488 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7489 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7491 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7492 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7494 stat->nFreePackets = ntohl(stat->nFreePackets);
7495 stat->packetReclaims = ntohl(stat->packetReclaims);
7496 stat->callsExecuted = ntohl(stat->callsExecuted);
7497 stat->nWaiting = ntohl(stat->nWaiting);
7498 stat->idleThreads = ntohl(stat->idleThreads);
7499 stat->nWaited = ntohl(stat->nWaited);
7500 stat->nPackets = ntohl(stat->nPackets);
7509 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7510 afs_uint16 remotePort, struct rx_statistics * stat,
7511 afs_uint32 * supportedValues)
7513 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7515 struct rx_debugIn in;
7516 afs_int32 *lp = (afs_int32 *) stat;
7520 * supportedValues is currently unused, but added to allow future
7521 * versioning of this function.
7524 *supportedValues = 0;
7525 in.type = htonl(RX_DEBUGI_RXSTATS);
7527 memset(stat, 0, sizeof(*stat));
7529 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7530 &in, sizeof(in), stat, sizeof(*stat));
7535 * Do net to host conversion here
7538 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7549 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7550 afs_uint16 remotePort, size_t version_length,
7553 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7555 return MakeDebugCall(socket, remoteAddr, remotePort,
7556 RX_PACKET_TYPE_VERSION, a, 1, version,
7564 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7565 afs_uint16 remotePort, afs_int32 * nextConnection,
7566 int allConnections, afs_uint32 debugSupportedValues,
7567 struct rx_debugConn * conn,
7568 afs_uint32 * supportedValues)
7570 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7572 struct rx_debugIn in;
7576 * supportedValues is currently unused, but added to allow future
7577 * versioning of this function.
7580 *supportedValues = 0;
7581 if (allConnections) {
7582 in.type = htonl(RX_DEBUGI_GETALLCONN);
7584 in.type = htonl(RX_DEBUGI_GETCONN);
7586 in.index = htonl(*nextConnection);
7587 memset(conn, 0, sizeof(*conn));
7589 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7590 &in, sizeof(in), conn, sizeof(*conn));
7593 *nextConnection += 1;
7596 * Convert old connection format to new structure.
7599 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7600 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7601 #define MOVEvL(a) (conn->a = vL->a)
7603 /* any old or unrecognized version... */
7604 for (i = 0; i < RX_MAXCALLS; i++) {
7605 MOVEvL(callState[i]);
7606 MOVEvL(callMode[i]);
7607 MOVEvL(callFlags[i]);
7608 MOVEvL(callOther[i]);
7610 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7611 MOVEvL(secStats.type);
7612 MOVEvL(secStats.level);
7613 MOVEvL(secStats.flags);
7614 MOVEvL(secStats.expires);
7615 MOVEvL(secStats.packetsReceived);
7616 MOVEvL(secStats.packetsSent);
7617 MOVEvL(secStats.bytesReceived);
7618 MOVEvL(secStats.bytesSent);
7623 * Do net to host conversion here
7625 * I don't convert host or port since we are most likely
7626 * going to want these in NBO.
7628 conn->cid = ntohl(conn->cid);
7629 conn->serial = ntohl(conn->serial);
7630 for (i = 0; i < RX_MAXCALLS; i++) {
7631 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7633 conn->error = ntohl(conn->error);
7634 conn->secStats.flags = ntohl(conn->secStats.flags);
7635 conn->secStats.expires = ntohl(conn->secStats.expires);
7636 conn->secStats.packetsReceived =
7637 ntohl(conn->secStats.packetsReceived);
7638 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7639 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7640 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7641 conn->epoch = ntohl(conn->epoch);
7642 conn->natMTU = ntohl(conn->natMTU);
7651 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7652 afs_uint16 remotePort, afs_int32 * nextPeer,
7653 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7654 afs_uint32 * supportedValues)
7656 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7658 struct rx_debugIn in;
7661 * supportedValues is currently unused, but added to allow future
7662 * versioning of this function.
7665 *supportedValues = 0;
7666 in.type = htonl(RX_DEBUGI_GETPEER);
7667 in.index = htonl(*nextPeer);
7668 memset(peer, 0, sizeof(*peer));
7670 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7671 &in, sizeof(in), peer, sizeof(*peer));
7677 * Do net to host conversion here
7679 * I don't convert host or port since we are most likely
7680 * going to want these in NBO.
7682 peer->ifMTU = ntohs(peer->ifMTU);
7683 peer->idleWhen = ntohl(peer->idleWhen);
7684 peer->refCount = ntohs(peer->refCount);
7685 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7686 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7687 peer->rtt = ntohl(peer->rtt);
7688 peer->rtt_dev = ntohl(peer->rtt_dev);
7689 peer->timeout.sec = 0;
7690 peer->timeout.usec = 0;
7691 peer->nSent = ntohl(peer->nSent);
7692 peer->reSends = ntohl(peer->reSends);
7693 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7694 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7695 peer->natMTU = ntohs(peer->natMTU);
7696 peer->maxMTU = ntohs(peer->maxMTU);
7697 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7698 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7699 peer->MTU = ntohs(peer->MTU);
7700 peer->cwind = ntohs(peer->cwind);
7701 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7702 peer->congestSeq = ntohs(peer->congestSeq);
7703 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7704 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7705 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7706 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7715 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7716 struct rx_debugPeer * peerStats)
7719 afs_int32 error = 1; /* default to "did not succeed" */
7720 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7722 MUTEX_ENTER(&rx_peerHashTable_lock);
7723 for(tp = rx_peerHashTable[hashValue];
7724 tp != NULL; tp = tp->next) {
7725 if (tp->host == peerHost)
7731 MUTEX_EXIT(&rx_peerHashTable_lock);
7735 MUTEX_ENTER(&tp->peer_lock);
7736 peerStats->host = tp->host;
7737 peerStats->port = tp->port;
7738 peerStats->ifMTU = tp->ifMTU;
7739 peerStats->idleWhen = tp->idleWhen;
7740 peerStats->refCount = tp->refCount;
7741 peerStats->burstSize = tp->burstSize;
7742 peerStats->burst = tp->burst;
7743 peerStats->burstWait.sec = tp->burstWait.sec;
7744 peerStats->burstWait.usec = tp->burstWait.usec;
7745 peerStats->rtt = tp->rtt;
7746 peerStats->rtt_dev = tp->rtt_dev;
7747 peerStats->timeout.sec = 0;
7748 peerStats->timeout.usec = 0;
7749 peerStats->nSent = tp->nSent;
7750 peerStats->reSends = tp->reSends;
7751 peerStats->inPacketSkew = tp->inPacketSkew;
7752 peerStats->outPacketSkew = tp->outPacketSkew;
7753 peerStats->natMTU = tp->natMTU;
7754 peerStats->maxMTU = tp->maxMTU;
7755 peerStats->maxDgramPackets = tp->maxDgramPackets;
7756 peerStats->ifDgramPackets = tp->ifDgramPackets;
7757 peerStats->MTU = tp->MTU;
7758 peerStats->cwind = tp->cwind;
7759 peerStats->nDgramPackets = tp->nDgramPackets;
7760 peerStats->congestSeq = tp->congestSeq;
7761 peerStats->bytesSent.high = tp->bytesSent.high;
7762 peerStats->bytesSent.low = tp->bytesSent.low;
7763 peerStats->bytesReceived.high = tp->bytesReceived.high;
7764 peerStats->bytesReceived.low = tp->bytesReceived.low;
7765 MUTEX_EXIT(&tp->peer_lock);
7767 MUTEX_ENTER(&rx_peerHashTable_lock);
7770 MUTEX_EXIT(&rx_peerHashTable_lock);
7778 struct rx_serverQueueEntry *np;
7781 struct rx_call *call;
7782 struct rx_serverQueueEntry *sq;
7786 if (rxinit_status == 1) {
7788 return; /* Already shutdown. */
7792 #ifndef AFS_PTHREAD_ENV
7793 FD_ZERO(&rx_selectMask);
7794 #endif /* AFS_PTHREAD_ENV */
7795 rxi_dataQuota = RX_MAX_QUOTA;
7796 #ifndef AFS_PTHREAD_ENV
7798 #endif /* AFS_PTHREAD_ENV */
7801 #ifndef AFS_PTHREAD_ENV
7802 #ifndef AFS_USE_GETTIMEOFDAY
7804 #endif /* AFS_USE_GETTIMEOFDAY */
7805 #endif /* AFS_PTHREAD_ENV */
7807 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7808 call = queue_First(&rx_freeCallQueue, rx_call);
7810 rxi_Free(call, sizeof(struct rx_call));
7813 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7814 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7820 struct rx_peer **peer_ptr, **peer_end;
7821 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7822 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7824 struct rx_peer *peer, *next;
7826 MUTEX_ENTER(&rx_peerHashTable_lock);
7827 for (peer = *peer_ptr; peer; peer = next) {
7828 rx_interface_stat_p rpc_stat, nrpc_stat;
7831 MUTEX_ENTER(&rx_rpc_stats);
7832 MUTEX_ENTER(&peer->peer_lock);
7834 (&peer->rpcStats, rpc_stat, nrpc_stat,
7835 rx_interface_stat)) {
7836 unsigned int num_funcs;
7839 queue_Remove(&rpc_stat->queue_header);
7840 queue_Remove(&rpc_stat->all_peers);
7841 num_funcs = rpc_stat->stats[0].func_total;
7843 sizeof(rx_interface_stat_t) +
7844 rpc_stat->stats[0].func_total *
7845 sizeof(rx_function_entry_v1_t);
7847 rxi_Free(rpc_stat, space);
7849 /* rx_rpc_stats must be held */
7850 rxi_rpc_peer_stat_cnt -= num_funcs;
7852 MUTEX_EXIT(&peer->peer_lock);
7853 MUTEX_EXIT(&rx_rpc_stats);
7857 if (rx_stats_active)
7858 rx_atomic_dec(&rx_stats.nPeerStructs);
7860 MUTEX_EXIT(&rx_peerHashTable_lock);
7863 for (i = 0; i < RX_MAX_SERVICES; i++) {
7865 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7867 for (i = 0; i < rx_hashTableSize; i++) {
7868 struct rx_connection *tc, *ntc;
7869 MUTEX_ENTER(&rx_connHashTable_lock);
7870 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7872 for (j = 0; j < RX_MAXCALLS; j++) {
7874 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7877 rxi_Free(tc, sizeof(*tc));
7879 MUTEX_EXIT(&rx_connHashTable_lock);
7882 MUTEX_ENTER(&freeSQEList_lock);
7884 while ((np = rx_FreeSQEList)) {
7885 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7886 MUTEX_DESTROY(&np->lock);
7887 rxi_Free(np, sizeof(*np));
7890 MUTEX_EXIT(&freeSQEList_lock);
7891 MUTEX_DESTROY(&freeSQEList_lock);
7892 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7893 MUTEX_DESTROY(&rx_connHashTable_lock);
7894 MUTEX_DESTROY(&rx_peerHashTable_lock);
7895 MUTEX_DESTROY(&rx_serverPool_lock);
7897 osi_Free(rx_connHashTable,
7898 rx_hashTableSize * sizeof(struct rx_connection *));
7899 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7901 UNPIN(rx_connHashTable,
7902 rx_hashTableSize * sizeof(struct rx_connection *));
7903 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7905 rxi_FreeAllPackets();
7907 MUTEX_ENTER(&rx_quota_mutex);
7908 rxi_dataQuota = RX_MAX_QUOTA;
7909 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7910 MUTEX_EXIT(&rx_quota_mutex);
7915 #ifdef RX_ENABLE_LOCKS
7917 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7919 if (!MUTEX_ISMINE(lockaddr))
7920 osi_Panic("Lock not held: %s", msg);
7922 #endif /* RX_ENABLE_LOCKS */
7927 * Routines to implement connection specific data.
7931 rx_KeyCreate(rx_destructor_t rtn)
7934 MUTEX_ENTER(&rxi_keyCreate_lock);
7935 key = rxi_keyCreate_counter++;
7936 rxi_keyCreate_destructor = (rx_destructor_t *)
7937 realloc((void *)rxi_keyCreate_destructor,
7938 (key + 1) * sizeof(rx_destructor_t));
7939 rxi_keyCreate_destructor[key] = rtn;
7940 MUTEX_EXIT(&rxi_keyCreate_lock);
7945 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7948 MUTEX_ENTER(&conn->conn_data_lock);
7949 if (!conn->specific) {
7950 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7951 for (i = 0; i < key; i++)
7952 conn->specific[i] = NULL;
7953 conn->nSpecific = key + 1;
7954 conn->specific[key] = ptr;
7955 } else if (key >= conn->nSpecific) {
7956 conn->specific = (void **)
7957 realloc(conn->specific, (key + 1) * sizeof(void *));
7958 for (i = conn->nSpecific; i < key; i++)
7959 conn->specific[i] = NULL;
7960 conn->nSpecific = key + 1;
7961 conn->specific[key] = ptr;
7963 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7964 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7965 conn->specific[key] = ptr;
7967 MUTEX_EXIT(&conn->conn_data_lock);
7971 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7974 MUTEX_ENTER(&svc->svc_data_lock);
7975 if (!svc->specific) {
7976 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
7977 for (i = 0; i < key; i++)
7978 svc->specific[i] = NULL;
7979 svc->nSpecific = key + 1;
7980 svc->specific[key] = ptr;
7981 } else if (key >= svc->nSpecific) {
7982 svc->specific = (void **)
7983 realloc(svc->specific, (key + 1) * sizeof(void *));
7984 for (i = svc->nSpecific; i < key; i++)
7985 svc->specific[i] = NULL;
7986 svc->nSpecific = key + 1;
7987 svc->specific[key] = ptr;
7989 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7990 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7991 svc->specific[key] = ptr;
7993 MUTEX_EXIT(&svc->svc_data_lock);
7997 rx_GetSpecific(struct rx_connection *conn, int key)
8000 MUTEX_ENTER(&conn->conn_data_lock);
8001 if (key >= conn->nSpecific)
8004 ptr = conn->specific[key];
8005 MUTEX_EXIT(&conn->conn_data_lock);
8010 rx_GetServiceSpecific(struct rx_service *svc, int key)
8013 MUTEX_ENTER(&svc->svc_data_lock);
8014 if (key >= svc->nSpecific)
8017 ptr = svc->specific[key];
8018 MUTEX_EXIT(&svc->svc_data_lock);
8023 #endif /* !KERNEL */
8026 * processStats is a queue used to store the statistics for the local
8027 * process. Its contents are similar to the contents of the rpcStats
8028 * queue on a rx_peer structure, but the actual data stored within
8029 * this queue contains totals across the lifetime of the process (assuming
8030 * the stats have not been reset) - unlike the per peer structures
8031 * which can come and go based upon the peer lifetime.
8034 static struct rx_queue processStats = { &processStats, &processStats };
8037 * peerStats is a queue used to store the statistics for all peer structs.
8038 * Its contents are the union of all the peer rpcStats queues.
8041 static struct rx_queue peerStats = { &peerStats, &peerStats };
8044 * rxi_monitor_processStats is used to turn process wide stat collection
8048 static int rxi_monitor_processStats = 0;
8051 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8054 static int rxi_monitor_peerStats = 0;
8057 * rxi_AddRpcStat - given all of the information for a particular rpc
8058 * call, create (if needed) and update the stat totals for the rpc.
8062 * IN stats - the queue of stats that will be updated with the new value
8064 * IN rxInterface - a unique number that identifies the rpc interface
8066 * IN currentFunc - the index of the function being invoked
8068 * IN totalFunc - the total number of functions in this interface
8070 * IN queueTime - the amount of time this function waited for a thread
8072 * IN execTime - the amount of time this function invocation took to execute
8074 * IN bytesSent - the number bytes sent by this invocation
8076 * IN bytesRcvd - the number bytes received by this invocation
8078 * IN isServer - if true, this invocation was made to a server
8080 * IN remoteHost - the ip address of the remote host
8082 * IN remotePort - the port of the remote host
8084 * IN addToPeerList - if != 0, add newly created stat to the global peer list
8086 * INOUT counter - if a new stats structure is allocated, the counter will
8087 * be updated with the new number of allocated stat structures
8095 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
8096 afs_uint32 currentFunc, afs_uint32 totalFunc,
8097 struct clock *queueTime, struct clock *execTime,
8098 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
8099 afs_uint32 remoteHost, afs_uint32 remotePort,
8100 int addToPeerList, unsigned int *counter)
8103 rx_interface_stat_p rpc_stat, nrpc_stat;
8106 * See if there's already a structure for this interface
8109 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8110 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8111 && (rpc_stat->stats[0].remote_is_server == isServer))
8116 * Didn't find a match so allocate a new structure and add it to the
8120 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
8121 || (rpc_stat->stats[0].interfaceId != rxInterface)
8122 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8127 sizeof(rx_interface_stat_t) +
8128 totalFunc * sizeof(rx_function_entry_v1_t);
8130 rpc_stat = rxi_Alloc(space);
8131 if (rpc_stat == NULL) {
8135 *counter += totalFunc;
8136 for (i = 0; i < totalFunc; i++) {
8137 rpc_stat->stats[i].remote_peer = remoteHost;
8138 rpc_stat->stats[i].remote_port = remotePort;
8139 rpc_stat->stats[i].remote_is_server = isServer;
8140 rpc_stat->stats[i].interfaceId = rxInterface;
8141 rpc_stat->stats[i].func_total = totalFunc;
8142 rpc_stat->stats[i].func_index = i;
8143 hzero(rpc_stat->stats[i].invocations);
8144 hzero(rpc_stat->stats[i].bytes_sent);
8145 hzero(rpc_stat->stats[i].bytes_rcvd);
8146 rpc_stat->stats[i].queue_time_sum.sec = 0;
8147 rpc_stat->stats[i].queue_time_sum.usec = 0;
8148 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8149 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8150 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8151 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8152 rpc_stat->stats[i].queue_time_max.sec = 0;
8153 rpc_stat->stats[i].queue_time_max.usec = 0;
8154 rpc_stat->stats[i].execution_time_sum.sec = 0;
8155 rpc_stat->stats[i].execution_time_sum.usec = 0;
8156 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8157 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8158 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8159 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8160 rpc_stat->stats[i].execution_time_max.sec = 0;
8161 rpc_stat->stats[i].execution_time_max.usec = 0;
8163 queue_Prepend(stats, rpc_stat);
8164 if (addToPeerList) {
8165 queue_Prepend(&peerStats, &rpc_stat->all_peers);
8170 * Increment the stats for this function
8173 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
8174 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
8175 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
8176 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8177 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8178 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8179 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8181 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8182 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8184 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8185 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8187 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8188 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8190 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8191 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8199 * rx_IncrementTimeAndCount - increment the times and count for a particular
8204 * IN peer - the peer who invoked the rpc
8206 * IN rxInterface - a unique number that identifies the rpc interface
8208 * IN currentFunc - the index of the function being invoked
8210 * IN totalFunc - the total number of functions in this interface
8212 * IN queueTime - the amount of time this function waited for a thread
8214 * IN execTime - the amount of time this function invocation took to execute
8216 * IN bytesSent - the number bytes sent by this invocation
8218 * IN bytesRcvd - the number bytes received by this invocation
8220 * IN isServer - if true, this invocation was made to a server
8228 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8229 afs_uint32 currentFunc, afs_uint32 totalFunc,
8230 struct clock *queueTime, struct clock *execTime,
8231 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8235 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8238 MUTEX_ENTER(&rx_rpc_stats);
8240 if (rxi_monitor_peerStats) {
8241 MUTEX_ENTER(&peer->peer_lock);
8242 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8243 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8244 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8245 MUTEX_EXIT(&peer->peer_lock);
8248 if (rxi_monitor_processStats) {
8249 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8250 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8251 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8254 MUTEX_EXIT(&rx_rpc_stats);
8259 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8263 * IN callerVersion - the rpc stat version of the caller.
8265 * IN count - the number of entries to marshall.
8267 * IN stats - pointer to stats to be marshalled.
8269 * OUT ptr - Where to store the marshalled data.
8276 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8277 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8283 * We only support the first version
8285 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8286 *(ptr++) = stats->remote_peer;
8287 *(ptr++) = stats->remote_port;
8288 *(ptr++) = stats->remote_is_server;
8289 *(ptr++) = stats->interfaceId;
8290 *(ptr++) = stats->func_total;
8291 *(ptr++) = stats->func_index;
8292 *(ptr++) = hgethi(stats->invocations);
8293 *(ptr++) = hgetlo(stats->invocations);
8294 *(ptr++) = hgethi(stats->bytes_sent);
8295 *(ptr++) = hgetlo(stats->bytes_sent);
8296 *(ptr++) = hgethi(stats->bytes_rcvd);
8297 *(ptr++) = hgetlo(stats->bytes_rcvd);
8298 *(ptr++) = stats->queue_time_sum.sec;
8299 *(ptr++) = stats->queue_time_sum.usec;
8300 *(ptr++) = stats->queue_time_sum_sqr.sec;
8301 *(ptr++) = stats->queue_time_sum_sqr.usec;
8302 *(ptr++) = stats->queue_time_min.sec;
8303 *(ptr++) = stats->queue_time_min.usec;
8304 *(ptr++) = stats->queue_time_max.sec;
8305 *(ptr++) = stats->queue_time_max.usec;
8306 *(ptr++) = stats->execution_time_sum.sec;
8307 *(ptr++) = stats->execution_time_sum.usec;
8308 *(ptr++) = stats->execution_time_sum_sqr.sec;
8309 *(ptr++) = stats->execution_time_sum_sqr.usec;
8310 *(ptr++) = stats->execution_time_min.sec;
8311 *(ptr++) = stats->execution_time_min.usec;
8312 *(ptr++) = stats->execution_time_max.sec;
8313 *(ptr++) = stats->execution_time_max.usec;
8319 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8324 * IN callerVersion - the rpc stat version of the caller
8326 * OUT myVersion - the rpc stat version of this function
8328 * OUT clock_sec - local time seconds
8330 * OUT clock_usec - local time microseconds
8332 * OUT allocSize - the number of bytes allocated to contain stats
8334 * OUT statCount - the number stats retrieved from this process.
8336 * OUT stats - the actual stats retrieved from this process.
8340 * Returns void. If successful, stats will != NULL.
8344 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8345 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8346 size_t * allocSize, afs_uint32 * statCount,
8347 afs_uint32 ** stats)
8357 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8360 * Check to see if stats are enabled
8363 MUTEX_ENTER(&rx_rpc_stats);
8364 if (!rxi_monitor_processStats) {
8365 MUTEX_EXIT(&rx_rpc_stats);
8369 clock_GetTime(&now);
8370 *clock_sec = now.sec;
8371 *clock_usec = now.usec;
8374 * Allocate the space based upon the caller version
8376 * If the client is at an older version than we are,
8377 * we return the statistic data in the older data format, but
8378 * we still return our version number so the client knows we
8379 * are maintaining more data than it can retrieve.
8382 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8383 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8384 *statCount = rxi_rpc_process_stat_cnt;
8387 * This can't happen yet, but in the future version changes
8388 * can be handled by adding additional code here
8392 if (space > (size_t) 0) {
8394 ptr = *stats = rxi_Alloc(space);
8397 rx_interface_stat_p rpc_stat, nrpc_stat;
8401 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8403 * Copy the data based upon the caller version
8405 rx_MarshallProcessRPCStats(callerVersion,
8406 rpc_stat->stats[0].func_total,
8407 rpc_stat->stats, &ptr);
8413 MUTEX_EXIT(&rx_rpc_stats);
8418 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8422 * IN callerVersion - the rpc stat version of the caller
8424 * OUT myVersion - the rpc stat version of this function
8426 * OUT clock_sec - local time seconds
8428 * OUT clock_usec - local time microseconds
8430 * OUT allocSize - the number of bytes allocated to contain stats
8432 * OUT statCount - the number of stats retrieved from the individual
8435 * OUT stats - the actual stats retrieved from the individual peer structures.
8439 * Returns void. If successful, stats will != NULL.
8443 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8444 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8445 size_t * allocSize, afs_uint32 * statCount,
8446 afs_uint32 ** stats)
8456 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8459 * Check to see if stats are enabled
8462 MUTEX_ENTER(&rx_rpc_stats);
8463 if (!rxi_monitor_peerStats) {
8464 MUTEX_EXIT(&rx_rpc_stats);
8468 clock_GetTime(&now);
8469 *clock_sec = now.sec;
8470 *clock_usec = now.usec;
8473 * Allocate the space based upon the caller version
8475 * If the client is at an older version than we are,
8476 * we return the statistic data in the older data format, but
8477 * we still return our version number so the client knows we
8478 * are maintaining more data than it can retrieve.
8481 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8482 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8483 *statCount = rxi_rpc_peer_stat_cnt;
8486 * This can't happen yet, but in the future version changes
8487 * can be handled by adding additional code here
8491 if (space > (size_t) 0) {
8493 ptr = *stats = rxi_Alloc(space);
8496 rx_interface_stat_p rpc_stat, nrpc_stat;
8500 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8502 * We have to fix the offset of rpc_stat since we are
8503 * keeping this structure on two rx_queues. The rx_queue
8504 * package assumes that the rx_queue member is the first
8505 * member of the structure. That is, rx_queue assumes that
8506 * any one item is only on one queue at a time. We are
8507 * breaking that assumption and so we have to do a little
8508 * math to fix our pointers.
8511 fix_offset = (char *)rpc_stat;
8512 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8513 rpc_stat = (rx_interface_stat_p) fix_offset;
8516 * Copy the data based upon the caller version
8518 rx_MarshallProcessRPCStats(callerVersion,
8519 rpc_stat->stats[0].func_total,
8520 rpc_stat->stats, &ptr);
8526 MUTEX_EXIT(&rx_rpc_stats);
8531 * rx_FreeRPCStats - free memory allocated by
8532 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8536 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8537 * rx_RetrievePeerRPCStats
8539 * IN allocSize - the number of bytes in stats.
8547 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8549 rxi_Free(stats, allocSize);
8553 * rx_queryProcessRPCStats - see if process rpc stat collection is
8554 * currently enabled.
8560 * Returns 0 if stats are not enabled != 0 otherwise
8564 rx_queryProcessRPCStats(void)
8567 MUTEX_ENTER(&rx_rpc_stats);
8568 rc = rxi_monitor_processStats;
8569 MUTEX_EXIT(&rx_rpc_stats);
8574 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8580 * Returns 0 if stats are not enabled != 0 otherwise
8584 rx_queryPeerRPCStats(void)
8587 MUTEX_ENTER(&rx_rpc_stats);
8588 rc = rxi_monitor_peerStats;
8589 MUTEX_EXIT(&rx_rpc_stats);
8594 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8604 rx_enableProcessRPCStats(void)
8606 MUTEX_ENTER(&rx_rpc_stats);
8607 rx_enable_stats = 1;
8608 rxi_monitor_processStats = 1;
8609 MUTEX_EXIT(&rx_rpc_stats);
8613 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8623 rx_enablePeerRPCStats(void)
8625 MUTEX_ENTER(&rx_rpc_stats);
8626 rx_enable_stats = 1;
8627 rxi_monitor_peerStats = 1;
8628 MUTEX_EXIT(&rx_rpc_stats);
8632 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8642 rx_disableProcessRPCStats(void)
8644 rx_interface_stat_p rpc_stat, nrpc_stat;
8647 MUTEX_ENTER(&rx_rpc_stats);
8650 * Turn off process statistics and if peer stats is also off, turn
8654 rxi_monitor_processStats = 0;
8655 if (rxi_monitor_peerStats == 0) {
8656 rx_enable_stats = 0;
8659 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8660 unsigned int num_funcs = 0;
8663 queue_Remove(rpc_stat);
8664 num_funcs = rpc_stat->stats[0].func_total;
8666 sizeof(rx_interface_stat_t) +
8667 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8669 rxi_Free(rpc_stat, space);
8670 rxi_rpc_process_stat_cnt -= num_funcs;
8672 MUTEX_EXIT(&rx_rpc_stats);
8676 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8686 rx_disablePeerRPCStats(void)
8688 struct rx_peer **peer_ptr, **peer_end;
8692 * Turn off peer statistics and if process stats is also off, turn
8696 rxi_monitor_peerStats = 0;
8697 if (rxi_monitor_processStats == 0) {
8698 rx_enable_stats = 0;
8701 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8702 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8704 struct rx_peer *peer, *next, *prev;
8706 MUTEX_ENTER(&rx_peerHashTable_lock);
8707 MUTEX_ENTER(&rx_rpc_stats);
8708 for (prev = peer = *peer_ptr; peer; peer = next) {
8710 code = MUTEX_TRYENTER(&peer->peer_lock);
8712 rx_interface_stat_p rpc_stat, nrpc_stat;
8715 if (prev == *peer_ptr) {
8726 MUTEX_EXIT(&rx_peerHashTable_lock);
8729 (&peer->rpcStats, rpc_stat, nrpc_stat,
8730 rx_interface_stat)) {
8731 unsigned int num_funcs = 0;
8734 queue_Remove(&rpc_stat->queue_header);
8735 queue_Remove(&rpc_stat->all_peers);
8736 num_funcs = rpc_stat->stats[0].func_total;
8738 sizeof(rx_interface_stat_t) +
8739 rpc_stat->stats[0].func_total *
8740 sizeof(rx_function_entry_v1_t);
8742 rxi_Free(rpc_stat, space);
8743 rxi_rpc_peer_stat_cnt -= num_funcs;
8745 MUTEX_EXIT(&peer->peer_lock);
8747 MUTEX_ENTER(&rx_peerHashTable_lock);
8757 MUTEX_EXIT(&rx_rpc_stats);
8758 MUTEX_EXIT(&rx_peerHashTable_lock);
8763 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8768 * IN clearFlag - flag indicating which stats to clear
8776 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8778 rx_interface_stat_p rpc_stat, nrpc_stat;
8780 MUTEX_ENTER(&rx_rpc_stats);
8782 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8783 unsigned int num_funcs = 0, i;
8784 num_funcs = rpc_stat->stats[0].func_total;
8785 for (i = 0; i < num_funcs; i++) {
8786 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8787 hzero(rpc_stat->stats[i].invocations);
8789 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8790 hzero(rpc_stat->stats[i].bytes_sent);
8792 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8793 hzero(rpc_stat->stats[i].bytes_rcvd);
8795 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8796 rpc_stat->stats[i].queue_time_sum.sec = 0;
8797 rpc_stat->stats[i].queue_time_sum.usec = 0;
8799 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8800 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8801 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8803 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8804 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8805 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8807 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8808 rpc_stat->stats[i].queue_time_max.sec = 0;
8809 rpc_stat->stats[i].queue_time_max.usec = 0;
8811 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8812 rpc_stat->stats[i].execution_time_sum.sec = 0;
8813 rpc_stat->stats[i].execution_time_sum.usec = 0;
8815 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8816 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8817 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8819 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8820 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8821 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8823 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8824 rpc_stat->stats[i].execution_time_max.sec = 0;
8825 rpc_stat->stats[i].execution_time_max.usec = 0;
8830 MUTEX_EXIT(&rx_rpc_stats);
8834 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8839 * IN clearFlag - flag indicating which stats to clear
8847 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8849 rx_interface_stat_p rpc_stat, nrpc_stat;
8851 MUTEX_ENTER(&rx_rpc_stats);
8853 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8854 unsigned int num_funcs = 0, i;
8857 * We have to fix the offset of rpc_stat since we are
8858 * keeping this structure on two rx_queues. The rx_queue
8859 * package assumes that the rx_queue member is the first
8860 * member of the structure. That is, rx_queue assumes that
8861 * any one item is only on one queue at a time. We are
8862 * breaking that assumption and so we have to do a little
8863 * math to fix our pointers.
8866 fix_offset = (char *)rpc_stat;
8867 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8868 rpc_stat = (rx_interface_stat_p) fix_offset;
8870 num_funcs = rpc_stat->stats[0].func_total;
8871 for (i = 0; i < num_funcs; i++) {
8872 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8873 hzero(rpc_stat->stats[i].invocations);
8875 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8876 hzero(rpc_stat->stats[i].bytes_sent);
8878 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8879 hzero(rpc_stat->stats[i].bytes_rcvd);
8881 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8882 rpc_stat->stats[i].queue_time_sum.sec = 0;
8883 rpc_stat->stats[i].queue_time_sum.usec = 0;
8885 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8886 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8887 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8889 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8890 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8891 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8893 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8894 rpc_stat->stats[i].queue_time_max.sec = 0;
8895 rpc_stat->stats[i].queue_time_max.usec = 0;
8897 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8898 rpc_stat->stats[i].execution_time_sum.sec = 0;
8899 rpc_stat->stats[i].execution_time_sum.usec = 0;
8901 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8902 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8903 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8905 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8906 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8907 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8909 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8910 rpc_stat->stats[i].execution_time_max.sec = 0;
8911 rpc_stat->stats[i].execution_time_max.usec = 0;
8916 MUTEX_EXIT(&rx_rpc_stats);
8920 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8921 * is authorized to enable/disable/clear RX statistics.
8923 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8926 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8928 rxi_rxstat_userok = proc;
8932 rx_RxStatUserOk(struct rx_call *call)
8934 if (!rxi_rxstat_userok)
8936 return rxi_rxstat_userok(call);
8941 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8942 * function in the MSVC runtime DLL (msvcrt.dll).
8944 * Note: the system serializes calls to this function.
8947 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8948 DWORD reason, /* reason function is being called */
8949 LPVOID reserved) /* reserved for future use */
8952 case DLL_PROCESS_ATTACH:
8953 /* library is being attached to a process */
8957 case DLL_PROCESS_DETACH:
8964 #endif /* AFS_NT40_ENV */
8967 int rx_DumpCalls(FILE *outputFile, char *cookie)
8969 #ifdef RXDEBUG_PACKET
8970 #ifdef KDUMP_RX_LOCK
8971 struct rx_call_rx_lock *c;
8978 #define RXDPRINTF sprintf
8979 #define RXDPRINTOUT output
8981 #define RXDPRINTF fprintf
8982 #define RXDPRINTOUT outputFile
8985 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8987 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8990 for (c = rx_allCallsp; c; c = c->allNextp) {
8991 u_short rqc, tqc, iovqc;
8992 struct rx_packet *p, *np;
8994 MUTEX_ENTER(&c->lock);
8995 queue_Count(&c->rq, p, np, rx_packet, rqc);
8996 queue_Count(&c->tq, p, np, rx_packet, tqc);
8997 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8999 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, "
9000 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9001 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9002 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9003 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
9004 #ifdef RX_ENABLE_LOCKS
9007 #ifdef RX_REFCOUNT_CHECK
9008 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9009 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9012 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,
9013 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9014 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9015 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9016 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9017 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
9018 #ifdef RX_ENABLE_LOCKS
9019 , (afs_uint32)c->refCount
9021 #ifdef RX_REFCOUNT_CHECK
9022 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9025 MUTEX_EXIT(&c->lock);
9028 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9031 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9033 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9035 #endif /* RXDEBUG_PACKET */