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 "rx_packet.h"
87 #include <afs/rxgen_consts.h>
90 #ifdef AFS_PTHREAD_ENV
92 int (*registerProgram) (pid_t, char *) = 0;
93 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
96 int (*registerProgram) (PROCESS, char *) = 0;
97 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
101 /* Local static routines */
102 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
103 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
104 struct rx_call *, struct rx_peer *,
106 static void rxi_Resend(struct rxevent *event, void *arg0, void *arg1,
108 static void rxi_SendDelayedAck(struct rxevent *event, void *call,
109 void *dummy, int dummy2);
110 static void rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1,
111 void *dummy, int dummy2);
112 static void rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1,
113 void *unused, int unused2);
114 static void rxi_ReapConnections(struct rxevent *unused, void *unused1,
115 void *unused2, int unused3);
117 #ifdef RX_ENABLE_LOCKS
118 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
121 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
123 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
124 rx_atomic_t rxi_start_in_error;
126 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
128 /* Constant delay time before sending an acknowledge of the last packet
129 * received. This is to avoid sending an extra acknowledge when the
130 * client is about to make another call, anyway, or the server is
133 * The lastAckDelay may not exceeed 400ms without causing peers to
134 * unecessarily timeout.
136 struct clock rx_lastAckDelay = {0, 400000};
138 /* Constant delay time before sending a soft ack when none was requested.
139 * This is to make sure we send soft acks before the sender times out,
140 * Normally we wait and send a hard ack when the receiver consumes the packet
142 * This value has been 100ms in all shipping versions of OpenAFS. Changing it
143 * will require changes to the peer's RTT calculations.
145 struct clock rx_softAckDelay = {0, 100000};
148 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
149 * currently allocated within rx. This number is used to allocate the
150 * memory required to return the statistics when queried.
151 * Protected by the rx_rpc_stats mutex.
154 static unsigned int rxi_rpc_peer_stat_cnt;
157 * rxi_rpc_process_stat_cnt counts the total number of local process stat
158 * structures currently allocated within rx. The number is used to allocate
159 * the memory required to return the statistics when queried.
160 * Protected by the rx_rpc_stats mutex.
163 static unsigned int rxi_rpc_process_stat_cnt;
166 * rxi_busyChannelError is the error to return to the application when a call
167 * channel appears busy (inferred from the receipt of RX_PACKET_TYPE_BUSY
168 * packets on the channel), and there are other call channels in the
169 * connection that are not busy. If 0, we do not return errors upon receiving
170 * busy packets; we just keep trying on the same call channel until we hit a
173 static afs_int32 rxi_busyChannelError = 0;
175 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
176 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
178 #if !defined(offsetof)
179 #include <stddef.h> /* for definition of offsetof() */
182 #ifdef RX_ENABLE_LOCKS
183 afs_kmutex_t rx_atomic_mutex;
186 /* Forward prototypes */
187 static struct rx_call * rxi_NewCall(struct rx_connection *, int);
190 putConnection (struct rx_connection *conn) {
191 MUTEX_ENTER(&rx_refcnt_mutex);
193 MUTEX_EXIT(&rx_refcnt_mutex);
196 #ifdef AFS_PTHREAD_ENV
199 * Use procedural initialization of mutexes/condition variables
203 extern afs_kmutex_t rx_quota_mutex;
204 extern afs_kmutex_t rx_pthread_mutex;
205 extern afs_kmutex_t rx_packets_mutex;
206 extern afs_kmutex_t rx_refcnt_mutex;
207 extern afs_kmutex_t des_init_mutex;
208 extern afs_kmutex_t des_random_mutex;
209 extern afs_kmutex_t rx_clock_mutex;
210 extern afs_kmutex_t rxi_connCacheMutex;
211 extern afs_kmutex_t event_handler_mutex;
212 extern afs_kmutex_t listener_mutex;
213 extern afs_kmutex_t rx_if_init_mutex;
214 extern afs_kmutex_t rx_if_mutex;
216 extern afs_kcondvar_t rx_event_handler_cond;
217 extern afs_kcondvar_t rx_listener_cond;
219 static afs_kmutex_t epoch_mutex;
220 static afs_kmutex_t rx_init_mutex;
221 static afs_kmutex_t rx_debug_mutex;
222 static afs_kmutex_t rx_rpc_stats;
225 rxi_InitPthread(void)
227 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
228 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
229 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
230 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
231 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
232 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
233 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
234 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
235 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
236 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
237 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
238 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
239 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
240 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
241 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
243 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
244 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
246 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
247 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
249 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
250 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
251 #ifdef RX_ENABLE_LOCKS
254 #endif /* RX_LOCKS_DB */
255 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
256 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
258 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
260 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
262 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
264 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
265 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
266 #endif /* RX_ENABLE_LOCKS */
269 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
270 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
272 * The rx_stats_mutex mutex protects the following global variables:
273 * rxi_lowConnRefCount
274 * rxi_lowPeerRefCount
283 * The rx_quota_mutex mutex protects the following global variables:
291 * The rx_freePktQ_lock protects the following global variables:
296 * The rx_packets_mutex mutex protects the following global variables:
304 * The rx_pthread_mutex mutex protects the following global variables:
305 * rxi_fcfs_thread_num
308 #define INIT_PTHREAD_LOCKS
312 /* Variables for handling the minProcs implementation. availProcs gives the
313 * number of threads available in the pool at this moment (not counting dudes
314 * executing right now). totalMin gives the total number of procs required
315 * for handling all minProcs requests. minDeficit is a dynamic variable
316 * tracking the # of procs required to satisfy all of the remaining minProcs
318 * For fine grain locking to work, the quota check and the reservation of
319 * a server thread has to come while rxi_availProcs and rxi_minDeficit
320 * are locked. To this end, the code has been modified under #ifdef
321 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
322 * same time. A new function, ReturnToServerPool() returns the allocation.
324 * A call can be on several queue's (but only one at a time). When
325 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
326 * that no one else is touching the queue. To this end, we store the address
327 * of the queue lock in the call structure (under the call lock) when we
328 * put the call on a queue, and we clear the call_queue_lock when the
329 * call is removed from a queue (once the call lock has been obtained).
330 * This allows rxi_ResetCall to safely synchronize with others wishing
331 * to manipulate the queue.
334 #if defined(RX_ENABLE_LOCKS)
335 static afs_kmutex_t rx_rpc_stats;
338 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
339 ** pretty good that the next packet coming in is from the same connection
340 ** as the last packet, since we're send multiple packets in a transmit window.
342 struct rx_connection *rxLastConn = 0;
344 #ifdef RX_ENABLE_LOCKS
345 /* The locking hierarchy for rx fine grain locking is composed of these
348 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
349 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
350 * call->lock - locks call data fields.
351 * These are independent of each other:
352 * rx_freeCallQueue_lock
357 * serverQueueEntry->lock
358 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
360 * peer->lock - locks peer data fields.
361 * conn_data_lock - that more than one thread is not updating a conn data
362 * field at the same time.
373 * Do we need a lock to protect the peer field in the conn structure?
374 * conn->peer was previously a constant for all intents and so has no
375 * lock protecting this field. The multihomed client delta introduced
376 * a RX code change : change the peer field in the connection structure
377 * to that remote interface from which the last packet for this
378 * connection was sent out. This may become an issue if further changes
381 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
382 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
384 /* rxdb_fileID is used to identify the lock location, along with line#. */
385 static int rxdb_fileID = RXDB_FILE_RX;
386 #endif /* RX_LOCKS_DB */
387 #else /* RX_ENABLE_LOCKS */
388 #define SET_CALL_QUEUE_LOCK(C, L)
389 #define CLEAR_CALL_QUEUE_LOCK(C)
390 #endif /* RX_ENABLE_LOCKS */
391 struct rx_serverQueueEntry *rx_waitForPacket = 0;
392 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
394 /* ------------Exported Interfaces------------- */
396 /* This function allows rxkad to set the epoch to a suitably random number
397 * which rx_NewConnection will use in the future. The principle purpose is to
398 * get rxnull connections to use the same epoch as the rxkad connections do, at
399 * least once the first rxkad connection is established. This is important now
400 * that the host/port addresses aren't used in FindConnection: the uniqueness
401 * of epoch/cid matters and the start time won't do. */
403 #ifdef AFS_PTHREAD_ENV
405 * This mutex protects the following global variables:
409 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
410 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
414 #endif /* AFS_PTHREAD_ENV */
417 rx_SetEpoch(afs_uint32 epoch)
424 /* Initialize rx. A port number may be mentioned, in which case this
425 * becomes the default port number for any service installed later.
426 * If 0 is provided for the port number, a random port will be chosen
427 * by the kernel. Whether this will ever overlap anything in
428 * /etc/services is anybody's guess... Returns 0 on success, -1 on
433 int rxinit_status = 1;
434 #ifdef AFS_PTHREAD_ENV
436 * This mutex protects the following global variables:
440 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
441 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
444 #define UNLOCK_RX_INIT
448 rx_InitHost(u_int host, u_int port)
455 char *htable, *ptable;
462 if (rxinit_status == 0) {
463 tmp_status = rxinit_status;
465 return tmp_status; /* Already started; return previous error code. */
471 if (afs_winsockInit() < 0)
477 * Initialize anything necessary to provide a non-premptive threading
480 rxi_InitializeThreadSupport();
483 /* Allocate and initialize a socket for client and perhaps server
486 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
487 if (rx_socket == OSI_NULLSOCKET) {
491 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
494 #endif /* RX_LOCKS_DB */
495 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
496 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
497 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
498 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
499 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
500 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
501 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
502 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
503 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
505 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
507 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
509 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
511 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
512 #if defined(AFS_HPUX110_ENV)
514 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
515 #endif /* AFS_HPUX110_ENV */
516 #endif /* RX_ENABLE_LOCKS && KERNEL */
519 rx_connDeadTime = 12;
520 rx_tranquil = 0; /* reset flag */
521 rxi_ResetStatistics();
523 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
524 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
525 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
526 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
527 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
528 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
530 /* Malloc up a bunch of packets & buffers */
532 queue_Init(&rx_freePacketQueue);
533 rxi_NeedMorePackets = FALSE;
534 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
536 /* enforce a minimum number of allocated packets */
537 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
538 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
540 /* allocate the initial free packet pool */
541 #ifdef RX_ENABLE_TSFPQ
542 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
543 #else /* RX_ENABLE_TSFPQ */
544 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
545 #endif /* RX_ENABLE_TSFPQ */
552 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
553 tv.tv_sec = clock_now.sec;
554 tv.tv_usec = clock_now.usec;
555 srand((unsigned int)tv.tv_usec);
562 #if defined(KERNEL) && !defined(UKERNEL)
563 /* Really, this should never happen in a real kernel */
566 struct sockaddr_in addr;
568 int addrlen = sizeof(addr);
570 socklen_t addrlen = sizeof(addr);
572 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
576 rx_port = addr.sin_port;
579 rx_stats.minRtt.sec = 9999999;
581 rx_SetEpoch(tv.tv_sec | 0x80000000);
583 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
584 * will provide a randomer value. */
586 MUTEX_ENTER(&rx_quota_mutex);
587 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
588 MUTEX_EXIT(&rx_quota_mutex);
589 /* *Slightly* random start time for the cid. This is just to help
590 * out with the hashing function at the peer */
591 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
592 rx_connHashTable = (struct rx_connection **)htable;
593 rx_peerHashTable = (struct rx_peer **)ptable;
595 rx_hardAckDelay.sec = 0;
596 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
598 rxevent_Init(20, rxi_ReScheduleEvents);
600 /* Initialize various global queues */
601 queue_Init(&rx_idleServerQueue);
602 queue_Init(&rx_incomingCallQueue);
603 queue_Init(&rx_freeCallQueue);
605 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
606 /* Initialize our list of usable IP addresses. */
610 #if defined(RXK_LISTENER_ENV) || !defined(KERNEL)
611 /* Start listener process (exact function is dependent on the
612 * implementation environment--kernel or user space) */
617 tmp_status = rxinit_status = 0;
625 return rx_InitHost(htonl(INADDR_ANY), port);
631 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
632 * maintaing the round trip timer.
637 * Start a new RTT timer for a given call and packet.
639 * There must be no resendEvent already listed for this call, otherwise this
640 * will leak events - intended for internal use within the RTO code only
643 * the RX call to start the timer for
644 * @param[in] lastPacket
645 * a flag indicating whether the last packet has been sent or not
647 * @pre call must be locked before calling this function
651 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
653 struct clock now, retryTime;
658 clock_Add(&retryTime, &call->rto);
660 /* If we're sending the last packet, and we're the client, then the server
661 * may wait for an additional 400ms before returning the ACK, wait for it
662 * rather than hitting a timeout */
663 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
664 clock_Addmsec(&retryTime, 400);
666 MUTEX_ENTER(&rx_refcnt_mutex);
667 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
668 MUTEX_EXIT(&rx_refcnt_mutex);
669 call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
674 * Cancel an RTT timer for a given call.
678 * the RX call to cancel the timer for
680 * @pre call must be locked before calling this function
685 rxi_rto_cancel(struct rx_call *call)
687 rxevent_Cancel(&call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
691 * Tell the RTO timer that we have sent a packet.
693 * If the timer isn't already running, then start it. If the timer is running,
697 * the RX call that the packet has been sent on
698 * @param[in] lastPacket
699 * A flag which is true if this is the last packet for the call
701 * @pre The call must be locked before calling this function
706 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
708 if (call->resendEvent)
711 rxi_rto_startTimer(call, lastPacket, istack);
715 * Tell the RTO timer that we have received an new ACK message
717 * This function should be called whenever a call receives an ACK that
718 * acknowledges new packets. Whatever happens, we stop the current timer.
719 * If there are unacked packets in the queue which have been sent, then
720 * we restart the timer from now. Otherwise, we leave it stopped.
723 * the RX call that the ACK has been received on
727 rxi_rto_packet_acked(struct rx_call *call, int istack)
729 struct rx_packet *p, *nxp;
731 rxi_rto_cancel(call);
733 if (queue_IsEmpty(&call->tq))
736 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
737 if (p->header.seq > call->tfirst + call->twind)
740 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
741 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
749 * Set an initial round trip timeout for a peer connection
751 * @param[in] secs The timeout to set in seconds
755 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
756 peer->rtt = secs * 8000;
760 * Sets the error generated when a busy call channel is detected.
762 * @param[in] error The error to return for a call on a busy channel.
764 * @pre Neither rx_Init nor rx_InitHost have been called yet
767 rx_SetBusyChannelError(afs_int32 error)
769 osi_Assert(rxinit_status != 0);
770 rxi_busyChannelError = error;
774 * Set a delayed ack event on the specified call for the given time
776 * @param[in] call - the call on which to set the event
777 * @param[in] offset - the delay from now after which the event fires
780 rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
782 struct clock now, when;
786 clock_Add(&when, offset);
788 if (!call->delayedAckEvent
789 || clock_Gt(&call->delayedAckTime, &when)) {
791 rxevent_Cancel(&call->delayedAckEvent, call,
792 RX_CALL_REFCOUNT_DELAY);
793 MUTEX_ENTER(&rx_refcnt_mutex);
794 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
795 MUTEX_EXIT(&rx_refcnt_mutex);
797 call->delayedAckEvent = rxevent_Post(&when, &now,
800 call->delayedAckTime = when;
804 /* called with unincremented nRequestsRunning to see if it is OK to start
805 * a new thread in this service. Could be "no" for two reasons: over the
806 * max quota, or would prevent others from reaching their min quota.
808 #ifdef RX_ENABLE_LOCKS
809 /* This verion of QuotaOK reserves quota if it's ok while the
810 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
813 QuotaOK(struct rx_service *aservice)
815 /* check if over max quota */
816 if (aservice->nRequestsRunning >= aservice->maxProcs) {
820 /* under min quota, we're OK */
821 /* otherwise, can use only if there are enough to allow everyone
822 * to go to their min quota after this guy starts.
825 MUTEX_ENTER(&rx_quota_mutex);
826 if ((aservice->nRequestsRunning < aservice->minProcs)
827 || (rxi_availProcs > rxi_minDeficit)) {
828 aservice->nRequestsRunning++;
829 /* just started call in minProcs pool, need fewer to maintain
831 if (aservice->nRequestsRunning <= aservice->minProcs)
834 MUTEX_EXIT(&rx_quota_mutex);
837 MUTEX_EXIT(&rx_quota_mutex);
843 ReturnToServerPool(struct rx_service *aservice)
845 aservice->nRequestsRunning--;
846 MUTEX_ENTER(&rx_quota_mutex);
847 if (aservice->nRequestsRunning < aservice->minProcs)
850 MUTEX_EXIT(&rx_quota_mutex);
853 #else /* RX_ENABLE_LOCKS */
855 QuotaOK(struct rx_service *aservice)
858 /* under min quota, we're OK */
859 if (aservice->nRequestsRunning < aservice->minProcs)
862 /* check if over max quota */
863 if (aservice->nRequestsRunning >= aservice->maxProcs)
866 /* otherwise, can use only if there are enough to allow everyone
867 * to go to their min quota after this guy starts.
869 MUTEX_ENTER(&rx_quota_mutex);
870 if (rxi_availProcs > rxi_minDeficit)
872 MUTEX_EXIT(&rx_quota_mutex);
875 #endif /* RX_ENABLE_LOCKS */
878 /* Called by rx_StartServer to start up lwp's to service calls.
879 NExistingProcs gives the number of procs already existing, and which
880 therefore needn't be created. */
882 rxi_StartServerProcs(int nExistingProcs)
884 struct rx_service *service;
889 /* For each service, reserve N processes, where N is the "minimum"
890 * number of processes that MUST be able to execute a request in parallel,
891 * at any time, for that process. Also compute the maximum difference
892 * between any service's maximum number of processes that can run
893 * (i.e. the maximum number that ever will be run, and a guarantee
894 * that this number will run if other services aren't running), and its
895 * minimum number. The result is the extra number of processes that
896 * we need in order to provide the latter guarantee */
897 for (i = 0; i < RX_MAX_SERVICES; i++) {
899 service = rx_services[i];
900 if (service == (struct rx_service *)0)
902 nProcs += service->minProcs;
903 diff = service->maxProcs - service->minProcs;
907 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
908 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
909 for (i = 0; i < nProcs; i++) {
910 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
916 /* This routine is only required on Windows */
918 rx_StartClientThread(void)
920 #ifdef AFS_PTHREAD_ENV
922 pid = pthread_self();
923 #endif /* AFS_PTHREAD_ENV */
925 #endif /* AFS_NT40_ENV */
927 /* This routine must be called if any services are exported. If the
928 * donateMe flag is set, the calling process is donated to the server
931 rx_StartServer(int donateMe)
933 struct rx_service *service;
939 /* Start server processes, if necessary (exact function is dependent
940 * on the implementation environment--kernel or user space). DonateMe
941 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
942 * case, one less new proc will be created rx_StartServerProcs.
944 rxi_StartServerProcs(donateMe);
946 /* count up the # of threads in minProcs, and add set the min deficit to
947 * be that value, too.
949 for (i = 0; i < RX_MAX_SERVICES; i++) {
950 service = rx_services[i];
951 if (service == (struct rx_service *)0)
953 MUTEX_ENTER(&rx_quota_mutex);
954 rxi_totalMin += service->minProcs;
955 /* below works even if a thread is running, since minDeficit would
956 * still have been decremented and later re-incremented.
958 rxi_minDeficit += service->minProcs;
959 MUTEX_EXIT(&rx_quota_mutex);
962 /* Turn on reaping of idle server connections */
963 rxi_ReapConnections(NULL, NULL, NULL, 0);
972 #ifdef AFS_PTHREAD_ENV
974 pid = afs_pointer_to_int(pthread_self());
975 #else /* AFS_PTHREAD_ENV */
977 LWP_CurrentProcess(&pid);
978 #endif /* AFS_PTHREAD_ENV */
980 sprintf(name, "srv_%d", ++nProcs);
982 (*registerProgram) (pid, name);
984 #endif /* AFS_NT40_ENV */
985 rx_ServerProc(NULL); /* Never returns */
987 #ifdef RX_ENABLE_TSFPQ
988 /* no use leaving packets around in this thread's local queue if
989 * it isn't getting donated to the server thread pool.
991 rxi_FlushLocalPacketsTSFPQ();
992 #endif /* RX_ENABLE_TSFPQ */
996 /* Create a new client connection to the specified service, using the
997 * specified security object to implement the security model for this
999 struct rx_connection *
1000 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
1001 struct rx_securityClass *securityObject,
1002 int serviceSecurityIndex)
1006 struct rx_connection *conn;
1011 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
1012 "serviceSecurityIndex %d)\n",
1013 ntohl(shost), ntohs(sport), sservice, securityObject,
1014 serviceSecurityIndex));
1016 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
1017 * the case of kmem_alloc? */
1018 conn = rxi_AllocConnection();
1019 #ifdef RX_ENABLE_LOCKS
1020 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
1021 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
1022 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
1025 MUTEX_ENTER(&rx_connHashTable_lock);
1026 cid = (rx_nextCid += RX_MAXCALLS);
1027 conn->type = RX_CLIENT_CONNECTION;
1029 conn->epoch = rx_epoch;
1030 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
1031 conn->serviceId = sservice;
1032 conn->securityObject = securityObject;
1033 conn->securityData = (void *) 0;
1034 conn->securityIndex = serviceSecurityIndex;
1035 rx_SetConnDeadTime(conn, rx_connDeadTime);
1036 rx_SetConnSecondsUntilNatPing(conn, 0);
1037 conn->ackRate = RX_FAST_ACK_RATE;
1038 conn->nSpecific = 0;
1039 conn->specific = NULL;
1040 conn->challengeEvent = NULL;
1041 conn->delayedAbortEvent = NULL;
1042 conn->abortCount = 0;
1044 for (i = 0; i < RX_MAXCALLS; i++) {
1045 conn->twind[i] = rx_initSendWindow;
1046 conn->rwind[i] = rx_initReceiveWindow;
1047 conn->lastBusy[i] = 0;
1050 RXS_NewConnection(securityObject, conn);
1052 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1054 conn->refCount++; /* no lock required since only this thread knows... */
1055 conn->next = rx_connHashTable[hashindex];
1056 rx_connHashTable[hashindex] = conn;
1057 if (rx_stats_active)
1058 rx_atomic_inc(&rx_stats.nClientConns);
1059 MUTEX_EXIT(&rx_connHashTable_lock);
1065 * Ensure a connection's timeout values are valid.
1067 * @param[in] conn The connection to check
1069 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1070 * unless idleDeadTime and/or hardDeadTime are not set
1074 rxi_CheckConnTimeouts(struct rx_connection *conn)
1076 /* a connection's timeouts must have the relationship
1077 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1078 * total loss of network to a peer may cause an idle timeout instead of a
1079 * dead timeout, simply because the idle timeout gets hit first. Also set
1080 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1081 /* this logic is slightly complicated by the fact that
1082 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1084 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1085 if (conn->idleDeadTime) {
1086 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1088 if (conn->hardDeadTime) {
1089 if (conn->idleDeadTime) {
1090 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1092 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1098 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1100 /* The idea is to set the dead time to a value that allows several
1101 * keepalives to be dropped without timing out the connection. */
1102 conn->secondsUntilDead = seconds;
1103 rxi_CheckConnTimeouts(conn);
1104 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1108 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1110 conn->hardDeadTime = seconds;
1111 rxi_CheckConnTimeouts(conn);
1115 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1117 conn->idleDeadTime = seconds;
1118 rxi_CheckConnTimeouts(conn);
1121 int rxi_lowPeerRefCount = 0;
1122 int rxi_lowConnRefCount = 0;
1125 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1126 * NOTE: must not be called with rx_connHashTable_lock held.
1129 rxi_CleanupConnection(struct rx_connection *conn)
1131 /* Notify the service exporter, if requested, that this connection
1132 * is being destroyed */
1133 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1134 (*conn->service->destroyConnProc) (conn);
1136 /* Notify the security module that this connection is being destroyed */
1137 RXS_DestroyConnection(conn->securityObject, conn);
1139 /* If this is the last connection using the rx_peer struct, set its
1140 * idle time to now. rxi_ReapConnections will reap it if it's still
1141 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1143 MUTEX_ENTER(&rx_peerHashTable_lock);
1144 if (conn->peer->refCount < 2) {
1145 conn->peer->idleWhen = clock_Sec();
1146 if (conn->peer->refCount < 1) {
1147 conn->peer->refCount = 1;
1148 if (rx_stats_active) {
1149 MUTEX_ENTER(&rx_stats_mutex);
1150 rxi_lowPeerRefCount++;
1151 MUTEX_EXIT(&rx_stats_mutex);
1155 conn->peer->refCount--;
1156 MUTEX_EXIT(&rx_peerHashTable_lock);
1158 if (rx_stats_active)
1160 if (conn->type == RX_SERVER_CONNECTION)
1161 rx_atomic_dec(&rx_stats.nServerConns);
1163 rx_atomic_dec(&rx_stats.nClientConns);
1166 if (conn->specific) {
1168 for (i = 0; i < conn->nSpecific; i++) {
1169 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1170 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1171 conn->specific[i] = NULL;
1173 free(conn->specific);
1175 conn->specific = NULL;
1176 conn->nSpecific = 0;
1177 #endif /* !KERNEL */
1179 MUTEX_DESTROY(&conn->conn_call_lock);
1180 MUTEX_DESTROY(&conn->conn_data_lock);
1181 CV_DESTROY(&conn->conn_call_cv);
1183 rxi_FreeConnection(conn);
1186 /* Destroy the specified connection */
1188 rxi_DestroyConnection(struct rx_connection *conn)
1190 MUTEX_ENTER(&rx_connHashTable_lock);
1191 rxi_DestroyConnectionNoLock(conn);
1192 /* conn should be at the head of the cleanup list */
1193 if (conn == rx_connCleanup_list) {
1194 rx_connCleanup_list = rx_connCleanup_list->next;
1195 MUTEX_EXIT(&rx_connHashTable_lock);
1196 rxi_CleanupConnection(conn);
1198 #ifdef RX_ENABLE_LOCKS
1200 MUTEX_EXIT(&rx_connHashTable_lock);
1202 #endif /* RX_ENABLE_LOCKS */
1206 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1208 struct rx_connection **conn_ptr;
1210 struct rx_packet *packet;
1217 MUTEX_ENTER(&conn->conn_data_lock);
1218 MUTEX_ENTER(&rx_refcnt_mutex);
1219 if (conn->refCount > 0)
1222 if (rx_stats_active) {
1223 MUTEX_ENTER(&rx_stats_mutex);
1224 rxi_lowConnRefCount++;
1225 MUTEX_EXIT(&rx_stats_mutex);
1229 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1230 /* Busy; wait till the last guy before proceeding */
1231 MUTEX_EXIT(&rx_refcnt_mutex);
1232 MUTEX_EXIT(&conn->conn_data_lock);
1237 /* If the client previously called rx_NewCall, but it is still
1238 * waiting, treat this as a running call, and wait to destroy the
1239 * connection later when the call completes. */
1240 if ((conn->type == RX_CLIENT_CONNECTION)
1241 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1242 conn->flags |= RX_CONN_DESTROY_ME;
1243 MUTEX_EXIT(&conn->conn_data_lock);
1247 MUTEX_EXIT(&rx_refcnt_mutex);
1248 MUTEX_EXIT(&conn->conn_data_lock);
1250 /* Check for extant references to this connection */
1251 MUTEX_ENTER(&conn->conn_call_lock);
1252 for (i = 0; i < RX_MAXCALLS; i++) {
1253 struct rx_call *call = conn->call[i];
1256 if (conn->type == RX_CLIENT_CONNECTION) {
1257 MUTEX_ENTER(&call->lock);
1258 if (call->delayedAckEvent) {
1259 /* Push the final acknowledgment out now--there
1260 * won't be a subsequent call to acknowledge the
1261 * last reply packets */
1262 rxevent_Cancel(&call->delayedAckEvent, call,
1263 RX_CALL_REFCOUNT_DELAY);
1264 if (call->state == RX_STATE_PRECALL
1265 || call->state == RX_STATE_ACTIVE) {
1266 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1268 rxi_AckAll(NULL, call, 0);
1271 MUTEX_EXIT(&call->lock);
1275 MUTEX_EXIT(&conn->conn_call_lock);
1277 #ifdef RX_ENABLE_LOCKS
1279 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1280 MUTEX_EXIT(&conn->conn_data_lock);
1282 /* Someone is accessing a packet right now. */
1286 #endif /* RX_ENABLE_LOCKS */
1289 /* Don't destroy the connection if there are any call
1290 * structures still in use */
1291 MUTEX_ENTER(&conn->conn_data_lock);
1292 conn->flags |= RX_CONN_DESTROY_ME;
1293 MUTEX_EXIT(&conn->conn_data_lock);
1298 if (conn->natKeepAliveEvent) {
1299 rxi_NatKeepAliveOff(conn);
1302 if (conn->delayedAbortEvent) {
1303 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
1304 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1306 MUTEX_ENTER(&conn->conn_data_lock);
1307 rxi_SendConnectionAbort(conn, packet, 0, 1);
1308 MUTEX_EXIT(&conn->conn_data_lock);
1309 rxi_FreePacket(packet);
1313 /* Remove from connection hash table before proceeding */
1315 &rx_connHashTable[CONN_HASH
1316 (peer->host, peer->port, conn->cid, conn->epoch,
1318 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1319 if (*conn_ptr == conn) {
1320 *conn_ptr = conn->next;
1324 /* if the conn that we are destroying was the last connection, then we
1325 * clear rxLastConn as well */
1326 if (rxLastConn == conn)
1329 /* Make sure the connection is completely reset before deleting it. */
1330 /* get rid of pending events that could zap us later */
1331 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
1332 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
1333 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
1335 /* Add the connection to the list of destroyed connections that
1336 * need to be cleaned up. This is necessary to avoid deadlocks
1337 * in the routines we call to inform others that this connection is
1338 * being destroyed. */
1339 conn->next = rx_connCleanup_list;
1340 rx_connCleanup_list = conn;
1343 /* Externally available version */
1345 rx_DestroyConnection(struct rx_connection *conn)
1350 rxi_DestroyConnection(conn);
1355 rx_GetConnection(struct rx_connection *conn)
1360 MUTEX_ENTER(&rx_refcnt_mutex);
1362 MUTEX_EXIT(&rx_refcnt_mutex);
1366 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1367 /* Wait for the transmit queue to no longer be busy.
1368 * requires the call->lock to be held */
1370 rxi_WaitforTQBusy(struct rx_call *call) {
1371 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1372 call->flags |= RX_CALL_TQ_WAIT;
1374 #ifdef RX_ENABLE_LOCKS
1375 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1376 CV_WAIT(&call->cv_tq, &call->lock);
1377 #else /* RX_ENABLE_LOCKS */
1378 osi_rxSleep(&call->tq);
1379 #endif /* RX_ENABLE_LOCKS */
1381 if (call->tqWaiters == 0) {
1382 call->flags &= ~RX_CALL_TQ_WAIT;
1389 rxi_WakeUpTransmitQueue(struct rx_call *call)
1391 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1392 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1393 call, call->tqWaiters, call->flags));
1394 #ifdef RX_ENABLE_LOCKS
1395 osirx_AssertMine(&call->lock, "rxi_Start start");
1396 CV_BROADCAST(&call->cv_tq);
1397 #else /* RX_ENABLE_LOCKS */
1398 osi_rxWakeup(&call->tq);
1399 #endif /* RX_ENABLE_LOCKS */
1403 /* Start a new rx remote procedure call, on the specified connection.
1404 * If wait is set to 1, wait for a free call channel; otherwise return
1405 * 0. Maxtime gives the maximum number of seconds this call may take,
1406 * after rx_NewCall returns. After this time interval, a call to any
1407 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1408 * For fine grain locking, we hold the conn_call_lock in order to
1409 * to ensure that we don't get signalle after we found a call in an active
1410 * state and before we go to sleep.
1413 rx_NewCall(struct rx_connection *conn)
1415 int i, wait, ignoreBusy = 1;
1416 struct rx_call *call;
1417 struct clock queueTime;
1418 afs_uint32 leastBusy = 0;
1422 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1425 clock_GetTime(&queueTime);
1427 * Check if there are others waiting for a new call.
1428 * If so, let them go first to avoid starving them.
1429 * This is a fairly simple scheme, and might not be
1430 * a complete solution for large numbers of waiters.
1432 * makeCallWaiters keeps track of the number of
1433 * threads waiting to make calls and the
1434 * RX_CONN_MAKECALL_WAITING flag bit is used to
1435 * indicate that there are indeed calls waiting.
1436 * The flag is set when the waiter is incremented.
1437 * It is only cleared when makeCallWaiters is 0.
1438 * This prevents us from accidently destroying the
1439 * connection while it is potentially about to be used.
1441 MUTEX_ENTER(&conn->conn_call_lock);
1442 MUTEX_ENTER(&conn->conn_data_lock);
1443 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1444 conn->flags |= RX_CONN_MAKECALL_WAITING;
1445 conn->makeCallWaiters++;
1446 MUTEX_EXIT(&conn->conn_data_lock);
1448 #ifdef RX_ENABLE_LOCKS
1449 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1453 MUTEX_ENTER(&conn->conn_data_lock);
1454 conn->makeCallWaiters--;
1455 if (conn->makeCallWaiters == 0)
1456 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1459 /* We are now the active thread in rx_NewCall */
1460 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1461 MUTEX_EXIT(&conn->conn_data_lock);
1466 for (i = 0; i < RX_MAXCALLS; i++) {
1467 call = conn->call[i];
1469 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1470 /* we're not ignoring busy call slots; only look at the
1471 * call slot that is the "least" busy */
1475 if (call->state == RX_STATE_DALLY) {
1476 MUTEX_ENTER(&call->lock);
1477 if (call->state == RX_STATE_DALLY) {
1478 if (ignoreBusy && conn->lastBusy[i]) {
1479 /* if we're ignoring busy call slots, skip any ones that
1480 * have lastBusy set */
1481 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1482 leastBusy = conn->lastBusy[i];
1484 MUTEX_EXIT(&call->lock);
1489 * We are setting the state to RX_STATE_RESET to
1490 * ensure that no one else will attempt to use this
1491 * call once we drop the conn->conn_call_lock and
1492 * call->lock. We must drop the conn->conn_call_lock
1493 * before calling rxi_ResetCall because the process
1494 * of clearing the transmit queue can block for an
1495 * extended period of time. If we block while holding
1496 * the conn->conn_call_lock, then all rx_EndCall
1497 * processing will block as well. This has a detrimental
1498 * effect on overall system performance.
1500 call->state = RX_STATE_RESET;
1501 MUTEX_EXIT(&conn->conn_call_lock);
1502 MUTEX_ENTER(&rx_refcnt_mutex);
1503 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1504 MUTEX_EXIT(&rx_refcnt_mutex);
1505 rxi_ResetCall(call, 0);
1506 (*call->callNumber)++;
1507 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1511 * If we failed to be able to safely obtain the
1512 * conn->conn_call_lock we will have to drop the
1513 * call->lock to avoid a deadlock. When the call->lock
1514 * is released the state of the call can change. If it
1515 * is no longer RX_STATE_RESET then some other thread is
1518 MUTEX_EXIT(&call->lock);
1519 MUTEX_ENTER(&conn->conn_call_lock);
1520 MUTEX_ENTER(&call->lock);
1522 if (call->state == RX_STATE_RESET)
1526 * If we get here it means that after dropping
1527 * the conn->conn_call_lock and call->lock that
1528 * the call is no longer ours. If we can't find
1529 * a free call in the remaining slots we should
1530 * not go immediately to RX_CONN_MAKECALL_WAITING
1531 * because by dropping the conn->conn_call_lock
1532 * we have given up synchronization with rx_EndCall.
1533 * Instead, cycle through one more time to see if
1534 * we can find a call that can call our own.
1536 MUTEX_ENTER(&rx_refcnt_mutex);
1537 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1538 MUTEX_EXIT(&rx_refcnt_mutex);
1541 MUTEX_EXIT(&call->lock);
1544 if (ignoreBusy && conn->lastBusy[i]) {
1545 /* if we're ignoring busy call slots, skip any ones that
1546 * have lastBusy set */
1547 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1548 leastBusy = conn->lastBusy[i];
1553 /* rxi_NewCall returns with mutex locked */
1554 call = rxi_NewCall(conn, i);
1555 MUTEX_ENTER(&rx_refcnt_mutex);
1556 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1557 MUTEX_EXIT(&rx_refcnt_mutex);
1561 if (i < RX_MAXCALLS) {
1562 conn->lastBusy[i] = 0;
1567 if (leastBusy && ignoreBusy) {
1568 /* we didn't find a useable call slot, but we did see at least one
1569 * 'busy' slot; look again and only use a slot with the 'least
1575 MUTEX_ENTER(&conn->conn_data_lock);
1576 conn->flags |= RX_CONN_MAKECALL_WAITING;
1577 conn->makeCallWaiters++;
1578 MUTEX_EXIT(&conn->conn_data_lock);
1580 #ifdef RX_ENABLE_LOCKS
1581 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1585 MUTEX_ENTER(&conn->conn_data_lock);
1586 conn->makeCallWaiters--;
1587 if (conn->makeCallWaiters == 0)
1588 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1589 MUTEX_EXIT(&conn->conn_data_lock);
1591 /* Client is initially in send mode */
1592 call->state = RX_STATE_ACTIVE;
1593 call->error = conn->error;
1595 call->mode = RX_MODE_ERROR;
1597 call->mode = RX_MODE_SENDING;
1599 /* remember start time for call in case we have hard dead time limit */
1600 call->queueTime = queueTime;
1601 clock_GetTime(&call->startTime);
1602 hzero(call->bytesSent);
1603 hzero(call->bytesRcvd);
1605 /* Turn on busy protocol. */
1606 rxi_KeepAliveOn(call);
1608 /* Attempt MTU discovery */
1609 rxi_GrowMTUOn(call);
1612 * We are no longer the active thread in rx_NewCall
1614 MUTEX_ENTER(&conn->conn_data_lock);
1615 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1616 MUTEX_EXIT(&conn->conn_data_lock);
1619 * Wake up anyone else who might be giving us a chance to
1620 * run (see code above that avoids resource starvation).
1622 #ifdef RX_ENABLE_LOCKS
1623 CV_BROADCAST(&conn->conn_call_cv);
1627 MUTEX_EXIT(&conn->conn_call_lock);
1629 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1630 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1631 osi_Panic("rx_NewCall call about to be used without an empty tq");
1633 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1635 MUTEX_EXIT(&call->lock);
1638 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1643 rxi_HasActiveCalls(struct rx_connection *aconn)
1646 struct rx_call *tcall;
1650 for (i = 0; i < RX_MAXCALLS; i++) {
1651 if ((tcall = aconn->call[i])) {
1652 if ((tcall->state == RX_STATE_ACTIVE)
1653 || (tcall->state == RX_STATE_PRECALL)) {
1664 rxi_GetCallNumberVector(struct rx_connection *aconn,
1665 afs_int32 * aint32s)
1668 struct rx_call *tcall;
1672 for (i = 0; i < RX_MAXCALLS; i++) {
1673 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1674 aint32s[i] = aconn->callNumber[i] + 1;
1676 aint32s[i] = aconn->callNumber[i];
1683 rxi_SetCallNumberVector(struct rx_connection *aconn,
1684 afs_int32 * aint32s)
1687 struct rx_call *tcall;
1691 for (i = 0; i < RX_MAXCALLS; i++) {
1692 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1693 aconn->callNumber[i] = aint32s[i] - 1;
1695 aconn->callNumber[i] = aint32s[i];
1701 /* Advertise a new service. A service is named locally by a UDP port
1702 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1705 char *serviceName; Name for identification purposes (e.g. the
1706 service name might be used for probing for
1709 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1710 char *serviceName, struct rx_securityClass **securityObjects,
1711 int nSecurityObjects,
1712 afs_int32(*serviceProc) (struct rx_call * acall))
1714 osi_socket socket = OSI_NULLSOCKET;
1715 struct rx_service *tservice;
1721 if (serviceId == 0) {
1723 "rx_NewService: service id for service %s is not non-zero.\n",
1730 "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",
1738 tservice = rxi_AllocService();
1741 #ifdef RX_ENABLE_LOCKS
1742 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1745 for (i = 0; i < RX_MAX_SERVICES; i++) {
1746 struct rx_service *service = rx_services[i];
1748 if (port == service->servicePort && host == service->serviceHost) {
1749 if (service->serviceId == serviceId) {
1750 /* The identical service has already been
1751 * installed; if the caller was intending to
1752 * change the security classes used by this
1753 * service, he/she loses. */
1755 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1756 serviceName, serviceId, service->serviceName);
1758 rxi_FreeService(tservice);
1761 /* Different service, same port: re-use the socket
1762 * which is bound to the same port */
1763 socket = service->socket;
1766 if (socket == OSI_NULLSOCKET) {
1767 /* If we don't already have a socket (from another
1768 * service on same port) get a new one */
1769 socket = rxi_GetHostUDPSocket(host, port);
1770 if (socket == OSI_NULLSOCKET) {
1772 rxi_FreeService(tservice);
1777 service->socket = socket;
1778 service->serviceHost = host;
1779 service->servicePort = port;
1780 service->serviceId = serviceId;
1781 service->serviceName = serviceName;
1782 service->nSecurityObjects = nSecurityObjects;
1783 service->securityObjects = securityObjects;
1784 service->minProcs = 0;
1785 service->maxProcs = 1;
1786 service->idleDeadTime = 60;
1787 service->idleDeadErr = 0;
1788 service->connDeadTime = rx_connDeadTime;
1789 service->executeRequestProc = serviceProc;
1790 service->checkReach = 0;
1791 service->nSpecific = 0;
1792 service->specific = NULL;
1793 rx_services[i] = service; /* not visible until now */
1799 rxi_FreeService(tservice);
1800 (osi_Msg "rx_NewService: cannot support > %d services\n",
1805 /* Set configuration options for all of a service's security objects */
1808 rx_SetSecurityConfiguration(struct rx_service *service,
1809 rx_securityConfigVariables type,
1813 for (i = 0; i<service->nSecurityObjects; i++) {
1814 if (service->securityObjects[i]) {
1815 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1823 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1824 struct rx_securityClass **securityObjects, int nSecurityObjects,
1825 afs_int32(*serviceProc) (struct rx_call * acall))
1827 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1830 /* Generic request processing loop. This routine should be called
1831 * by the implementation dependent rx_ServerProc. If socketp is
1832 * non-null, it will be set to the file descriptor that this thread
1833 * is now listening on. If socketp is null, this routine will never
1836 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1838 struct rx_call *call;
1840 struct rx_service *tservice = NULL;
1847 call = rx_GetCall(threadID, tservice, socketp);
1848 if (socketp && *socketp != OSI_NULLSOCKET) {
1849 /* We are now a listener thread */
1855 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1856 #ifdef RX_ENABLE_LOCKS
1858 #endif /* RX_ENABLE_LOCKS */
1859 afs_termState = AFSOP_STOP_AFS;
1860 afs_osi_Wakeup(&afs_termState);
1861 #ifdef RX_ENABLE_LOCKS
1863 #endif /* RX_ENABLE_LOCKS */
1868 /* if server is restarting( typically smooth shutdown) then do not
1869 * allow any new calls.
1872 if (rx_tranquil && (call != NULL)) {
1876 MUTEX_ENTER(&call->lock);
1878 rxi_CallError(call, RX_RESTARTING);
1879 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1881 MUTEX_EXIT(&call->lock);
1886 tservice = call->conn->service;
1888 if (tservice->beforeProc)
1889 (*tservice->beforeProc) (call);
1891 code = tservice->executeRequestProc(call);
1893 if (tservice->afterProc)
1894 (*tservice->afterProc) (call, code);
1896 rx_EndCall(call, code);
1898 if (tservice->postProc)
1899 (*tservice->postProc) (code);
1901 if (rx_stats_active) {
1902 MUTEX_ENTER(&rx_stats_mutex);
1904 MUTEX_EXIT(&rx_stats_mutex);
1911 rx_WakeupServerProcs(void)
1913 struct rx_serverQueueEntry *np, *tqp;
1917 MUTEX_ENTER(&rx_serverPool_lock);
1919 #ifdef RX_ENABLE_LOCKS
1920 if (rx_waitForPacket)
1921 CV_BROADCAST(&rx_waitForPacket->cv);
1922 #else /* RX_ENABLE_LOCKS */
1923 if (rx_waitForPacket)
1924 osi_rxWakeup(rx_waitForPacket);
1925 #endif /* RX_ENABLE_LOCKS */
1926 MUTEX_ENTER(&freeSQEList_lock);
1927 for (np = rx_FreeSQEList; np; np = tqp) {
1928 tqp = *(struct rx_serverQueueEntry **)np;
1929 #ifdef RX_ENABLE_LOCKS
1930 CV_BROADCAST(&np->cv);
1931 #else /* RX_ENABLE_LOCKS */
1933 #endif /* RX_ENABLE_LOCKS */
1935 MUTEX_EXIT(&freeSQEList_lock);
1936 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1937 #ifdef RX_ENABLE_LOCKS
1938 CV_BROADCAST(&np->cv);
1939 #else /* RX_ENABLE_LOCKS */
1941 #endif /* RX_ENABLE_LOCKS */
1943 MUTEX_EXIT(&rx_serverPool_lock);
1948 * One thing that seems to happen is that all the server threads get
1949 * tied up on some empty or slow call, and then a whole bunch of calls
1950 * arrive at once, using up the packet pool, so now there are more
1951 * empty calls. The most critical resources here are server threads
1952 * and the free packet pool. The "doreclaim" code seems to help in
1953 * general. I think that eventually we arrive in this state: there
1954 * are lots of pending calls which do have all their packets present,
1955 * so they won't be reclaimed, are multi-packet calls, so they won't
1956 * be scheduled until later, and thus are tying up most of the free
1957 * packet pool for a very long time.
1959 * 1. schedule multi-packet calls if all the packets are present.
1960 * Probably CPU-bound operation, useful to return packets to pool.
1961 * Do what if there is a full window, but the last packet isn't here?
1962 * 3. preserve one thread which *only* runs "best" calls, otherwise
1963 * it sleeps and waits for that type of call.
1964 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1965 * the current dataquota business is badly broken. The quota isn't adjusted
1966 * to reflect how many packets are presently queued for a running call.
1967 * So, when we schedule a queued call with a full window of packets queued
1968 * up for it, that *should* free up a window full of packets for other 2d-class
1969 * calls to be able to use from the packet pool. But it doesn't.
1971 * NB. Most of the time, this code doesn't run -- since idle server threads
1972 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1973 * as a new call arrives.
1975 /* Sleep until a call arrives. Returns a pointer to the call, ready
1976 * for an rx_Read. */
1977 #ifdef RX_ENABLE_LOCKS
1979 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1981 struct rx_serverQueueEntry *sq;
1982 struct rx_call *call = (struct rx_call *)0;
1983 struct rx_service *service = NULL;
1985 MUTEX_ENTER(&freeSQEList_lock);
1987 if ((sq = rx_FreeSQEList)) {
1988 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1989 MUTEX_EXIT(&freeSQEList_lock);
1990 } else { /* otherwise allocate a new one and return that */
1991 MUTEX_EXIT(&freeSQEList_lock);
1992 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1993 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1994 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1997 MUTEX_ENTER(&rx_serverPool_lock);
1998 if (cur_service != NULL) {
1999 ReturnToServerPool(cur_service);
2002 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
2003 struct rx_call *tcall, *ncall, *choice2 = NULL;
2005 /* Scan for eligible incoming calls. A call is not eligible
2006 * if the maximum number of calls for its service type are
2007 * already executing */
2008 /* One thread will process calls FCFS (to prevent starvation),
2009 * while the other threads may run ahead looking for calls which
2010 * have all their input data available immediately. This helps
2011 * keep threads from blocking, waiting for data from the client. */
2012 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
2013 service = tcall->conn->service;
2014 if (!QuotaOK(service)) {
2017 MUTEX_ENTER(&rx_pthread_mutex);
2018 if (tno == rxi_fcfs_thread_num
2019 || !tcall->queue_item_header.next) {
2020 MUTEX_EXIT(&rx_pthread_mutex);
2021 /* If we're the fcfs thread , then we'll just use
2022 * this call. If we haven't been able to find an optimal
2023 * choice, and we're at the end of the list, then use a
2024 * 2d choice if one has been identified. Otherwise... */
2025 call = (choice2 ? choice2 : tcall);
2026 service = call->conn->service;
2028 MUTEX_EXIT(&rx_pthread_mutex);
2029 if (!queue_IsEmpty(&tcall->rq)) {
2030 struct rx_packet *rp;
2031 rp = queue_First(&tcall->rq, rx_packet);
2032 if (rp->header.seq == 1) {
2034 || (rp->header.flags & RX_LAST_PACKET)) {
2036 } else if (rxi_2dchoice && !choice2
2037 && !(tcall->flags & RX_CALL_CLEARED)
2038 && (tcall->rprev > rxi_HardAckRate)) {
2048 ReturnToServerPool(service);
2055 MUTEX_EXIT(&rx_serverPool_lock);
2056 MUTEX_ENTER(&call->lock);
2058 if (call->flags & RX_CALL_WAIT_PROC) {
2059 call->flags &= ~RX_CALL_WAIT_PROC;
2060 rx_atomic_dec(&rx_nWaiting);
2063 if (call->state != RX_STATE_PRECALL || call->error) {
2064 MUTEX_EXIT(&call->lock);
2065 MUTEX_ENTER(&rx_serverPool_lock);
2066 ReturnToServerPool(service);
2071 if (queue_IsEmpty(&call->rq)
2072 || queue_First(&call->rq, rx_packet)->header.seq != 1)
2073 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2075 CLEAR_CALL_QUEUE_LOCK(call);
2078 /* If there are no eligible incoming calls, add this process
2079 * to the idle server queue, to wait for one */
2083 *socketp = OSI_NULLSOCKET;
2085 sq->socketp = socketp;
2086 queue_Append(&rx_idleServerQueue, sq);
2087 #ifndef AFS_AIX41_ENV
2088 rx_waitForPacket = sq;
2090 rx_waitingForPacket = sq;
2091 #endif /* AFS_AIX41_ENV */
2093 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2095 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2096 MUTEX_EXIT(&rx_serverPool_lock);
2097 return (struct rx_call *)0;
2100 } while (!(call = sq->newcall)
2101 && !(socketp && *socketp != OSI_NULLSOCKET));
2102 MUTEX_EXIT(&rx_serverPool_lock);
2104 MUTEX_ENTER(&call->lock);
2110 MUTEX_ENTER(&freeSQEList_lock);
2111 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2112 rx_FreeSQEList = sq;
2113 MUTEX_EXIT(&freeSQEList_lock);
2116 clock_GetTime(&call->startTime);
2117 call->state = RX_STATE_ACTIVE;
2118 call->mode = RX_MODE_RECEIVING;
2119 #ifdef RX_KERNEL_TRACE
2120 if (ICL_SETACTIVE(afs_iclSetp)) {
2121 int glockOwner = ISAFS_GLOCK();
2124 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2125 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2132 rxi_calltrace(RX_CALL_START, call);
2133 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2134 call->conn->service->servicePort, call->conn->service->serviceId,
2137 MUTEX_EXIT(&call->lock);
2138 MUTEX_ENTER(&rx_refcnt_mutex);
2139 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2140 MUTEX_EXIT(&rx_refcnt_mutex);
2142 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2147 #else /* RX_ENABLE_LOCKS */
2149 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2151 struct rx_serverQueueEntry *sq;
2152 struct rx_call *call = (struct rx_call *)0, *choice2;
2153 struct rx_service *service = NULL;
2157 MUTEX_ENTER(&freeSQEList_lock);
2159 if ((sq = rx_FreeSQEList)) {
2160 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2161 MUTEX_EXIT(&freeSQEList_lock);
2162 } else { /* otherwise allocate a new one and return that */
2163 MUTEX_EXIT(&freeSQEList_lock);
2164 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2165 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2166 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2168 MUTEX_ENTER(&sq->lock);
2170 if (cur_service != NULL) {
2171 cur_service->nRequestsRunning--;
2172 MUTEX_ENTER(&rx_quota_mutex);
2173 if (cur_service->nRequestsRunning < cur_service->minProcs)
2176 MUTEX_EXIT(&rx_quota_mutex);
2178 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
2179 struct rx_call *tcall, *ncall;
2180 /* Scan for eligible incoming calls. A call is not eligible
2181 * if the maximum number of calls for its service type are
2182 * already executing */
2183 /* One thread will process calls FCFS (to prevent starvation),
2184 * while the other threads may run ahead looking for calls which
2185 * have all their input data available immediately. This helps
2186 * keep threads from blocking, waiting for data from the client. */
2187 choice2 = (struct rx_call *)0;
2188 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
2189 service = tcall->conn->service;
2190 if (QuotaOK(service)) {
2191 MUTEX_ENTER(&rx_pthread_mutex);
2192 if (tno == rxi_fcfs_thread_num
2193 || !tcall->queue_item_header.next) {
2194 MUTEX_EXIT(&rx_pthread_mutex);
2195 /* If we're the fcfs thread, then we'll just use
2196 * this call. If we haven't been able to find an optimal
2197 * choice, and we're at the end of the list, then use a
2198 * 2d choice if one has been identified. Otherwise... */
2199 call = (choice2 ? choice2 : tcall);
2200 service = call->conn->service;
2202 MUTEX_EXIT(&rx_pthread_mutex);
2203 if (!queue_IsEmpty(&tcall->rq)) {
2204 struct rx_packet *rp;
2205 rp = queue_First(&tcall->rq, rx_packet);
2206 if (rp->header.seq == 1
2208 || (rp->header.flags & RX_LAST_PACKET))) {
2210 } else if (rxi_2dchoice && !choice2
2211 && !(tcall->flags & RX_CALL_CLEARED)
2212 && (tcall->rprev > rxi_HardAckRate)) {
2226 /* we can't schedule a call if there's no data!!! */
2227 /* send an ack if there's no data, if we're missing the
2228 * first packet, or we're missing something between first
2229 * and last -- there's a "hole" in the incoming data. */
2230 if (queue_IsEmpty(&call->rq)
2231 || queue_First(&call->rq, rx_packet)->header.seq != 1
2232 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
2233 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2235 call->flags &= (~RX_CALL_WAIT_PROC);
2236 service->nRequestsRunning++;
2237 /* just started call in minProcs pool, need fewer to maintain
2239 MUTEX_ENTER(&rx_quota_mutex);
2240 if (service->nRequestsRunning <= service->minProcs)
2243 MUTEX_EXIT(&rx_quota_mutex);
2244 rx_atomic_dec(&rx_nWaiting);
2245 /* MUTEX_EXIT(&call->lock); */
2247 /* If there are no eligible incoming calls, add this process
2248 * to the idle server queue, to wait for one */
2251 *socketp = OSI_NULLSOCKET;
2253 sq->socketp = socketp;
2254 queue_Append(&rx_idleServerQueue, sq);
2258 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2260 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2261 return (struct rx_call *)0;
2264 } while (!(call = sq->newcall)
2265 && !(socketp && *socketp != OSI_NULLSOCKET));
2267 MUTEX_EXIT(&sq->lock);
2269 MUTEX_ENTER(&freeSQEList_lock);
2270 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2271 rx_FreeSQEList = sq;
2272 MUTEX_EXIT(&freeSQEList_lock);
2275 clock_GetTime(&call->startTime);
2276 call->state = RX_STATE_ACTIVE;
2277 call->mode = RX_MODE_RECEIVING;
2278 #ifdef RX_KERNEL_TRACE
2279 if (ICL_SETACTIVE(afs_iclSetp)) {
2280 int glockOwner = ISAFS_GLOCK();
2283 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2284 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2291 rxi_calltrace(RX_CALL_START, call);
2292 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2293 call->conn->service->servicePort, call->conn->service->serviceId,
2296 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2303 #endif /* RX_ENABLE_LOCKS */
2307 /* Establish a procedure to be called when a packet arrives for a
2308 * call. This routine will be called at most once after each call,
2309 * and will also be called if there is an error condition on the or
2310 * the call is complete. Used by multi rx to build a selection
2311 * function which determines which of several calls is likely to be a
2312 * good one to read from.
2313 * NOTE: the way this is currently implemented it is probably only a
2314 * good idea to (1) use it immediately after a newcall (clients only)
2315 * and (2) only use it once. Other uses currently void your warranty
2318 rx_SetArrivalProc(struct rx_call *call,
2319 void (*proc) (struct rx_call * call,
2322 void * handle, int arg)
2324 call->arrivalProc = proc;
2325 call->arrivalProcHandle = handle;
2326 call->arrivalProcArg = arg;
2329 /* Call is finished (possibly prematurely). Return rc to the peer, if
2330 * appropriate, and return the final error code from the conversation
2334 rx_EndCall(struct rx_call *call, afs_int32 rc)
2336 struct rx_connection *conn = call->conn;
2340 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2341 call, rc, call->error, call->abortCode));
2344 MUTEX_ENTER(&call->lock);
2346 if (rc == 0 && call->error == 0) {
2347 call->abortCode = 0;
2348 call->abortCount = 0;
2351 call->arrivalProc = (void (*)())0;
2352 if (rc && call->error == 0) {
2353 rxi_CallError(call, rc);
2354 call->mode = RX_MODE_ERROR;
2355 /* Send an abort message to the peer if this error code has
2356 * only just been set. If it was set previously, assume the
2357 * peer has already been sent the error code or will request it
2359 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2361 if (conn->type == RX_SERVER_CONNECTION) {
2362 /* Make sure reply or at least dummy reply is sent */
2363 if (call->mode == RX_MODE_RECEIVING) {
2364 MUTEX_EXIT(&call->lock);
2365 rxi_WriteProc(call, 0, 0);
2366 MUTEX_ENTER(&call->lock);
2368 if (call->mode == RX_MODE_SENDING) {
2369 MUTEX_EXIT(&call->lock);
2370 rxi_FlushWrite(call);
2371 MUTEX_ENTER(&call->lock);
2373 rxi_calltrace(RX_CALL_END, call);
2374 /* Call goes to hold state until reply packets are acknowledged */
2375 if (call->tfirst + call->nSoftAcked < call->tnext) {
2376 call->state = RX_STATE_HOLD;
2378 call->state = RX_STATE_DALLY;
2379 rxi_ClearTransmitQueue(call, 0);
2380 rxi_rto_cancel(call);
2381 rxevent_Cancel(&call->keepAliveEvent, call,
2382 RX_CALL_REFCOUNT_ALIVE);
2384 } else { /* Client connection */
2386 /* Make sure server receives input packets, in the case where
2387 * no reply arguments are expected */
2388 if ((call->mode == RX_MODE_SENDING)
2389 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2390 MUTEX_EXIT(&call->lock);
2391 (void)rxi_ReadProc(call, &dummy, 1);
2392 MUTEX_ENTER(&call->lock);
2395 /* If we had an outstanding delayed ack, be nice to the server
2396 * and force-send it now.
2398 if (call->delayedAckEvent) {
2399 rxevent_Cancel(&call->delayedAckEvent, call,
2400 RX_CALL_REFCOUNT_DELAY);
2401 rxi_SendDelayedAck(NULL, call, NULL, 0);
2404 /* We need to release the call lock since it's lower than the
2405 * conn_call_lock and we don't want to hold the conn_call_lock
2406 * over the rx_ReadProc call. The conn_call_lock needs to be held
2407 * here for the case where rx_NewCall is perusing the calls on
2408 * the connection structure. We don't want to signal until
2409 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2410 * have checked this call, found it active and by the time it
2411 * goes to sleep, will have missed the signal.
2413 MUTEX_EXIT(&call->lock);
2414 MUTEX_ENTER(&conn->conn_call_lock);
2415 MUTEX_ENTER(&call->lock);
2417 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2418 conn->lastBusy[call->channel] = 0;
2421 MUTEX_ENTER(&conn->conn_data_lock);
2422 conn->flags |= RX_CONN_BUSY;
2423 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2424 MUTEX_EXIT(&conn->conn_data_lock);
2425 #ifdef RX_ENABLE_LOCKS
2426 CV_BROADCAST(&conn->conn_call_cv);
2431 #ifdef RX_ENABLE_LOCKS
2433 MUTEX_EXIT(&conn->conn_data_lock);
2435 #endif /* RX_ENABLE_LOCKS */
2436 call->state = RX_STATE_DALLY;
2438 error = call->error;
2440 /* currentPacket, nLeft, and NFree must be zeroed here, because
2441 * ResetCall cannot: ResetCall may be called at splnet(), in the
2442 * kernel version, and may interrupt the macros rx_Read or
2443 * rx_Write, which run at normal priority for efficiency. */
2444 if (call->currentPacket) {
2445 #ifdef RX_TRACK_PACKETS
2446 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2448 rxi_FreePacket(call->currentPacket);
2449 call->currentPacket = (struct rx_packet *)0;
2452 call->nLeft = call->nFree = call->curlen = 0;
2454 /* Free any packets from the last call to ReadvProc/WritevProc */
2455 #ifdef RXDEBUG_PACKET
2457 #endif /* RXDEBUG_PACKET */
2458 rxi_FreePackets(0, &call->iovq);
2459 MUTEX_EXIT(&call->lock);
2461 MUTEX_ENTER(&rx_refcnt_mutex);
2462 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2463 MUTEX_EXIT(&rx_refcnt_mutex);
2464 if (conn->type == RX_CLIENT_CONNECTION) {
2465 MUTEX_ENTER(&conn->conn_data_lock);
2466 conn->flags &= ~RX_CONN_BUSY;
2467 MUTEX_EXIT(&conn->conn_data_lock);
2468 MUTEX_EXIT(&conn->conn_call_lock);
2472 * Map errors to the local host's errno.h format.
2474 error = ntoh_syserr_conv(error);
2478 #if !defined(KERNEL)
2480 /* Call this routine when shutting down a server or client (especially
2481 * clients). This will allow Rx to gracefully garbage collect server
2482 * connections, and reduce the number of retries that a server might
2483 * make to a dead client.
2484 * This is not quite right, since some calls may still be ongoing and
2485 * we can't lock them to destroy them. */
2489 struct rx_connection **conn_ptr, **conn_end;
2493 if (rxinit_status == 1) {
2495 return; /* Already shutdown. */
2497 rxi_DeleteCachedConnections();
2498 if (rx_connHashTable) {
2499 MUTEX_ENTER(&rx_connHashTable_lock);
2500 for (conn_ptr = &rx_connHashTable[0], conn_end =
2501 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2503 struct rx_connection *conn, *next;
2504 for (conn = *conn_ptr; conn; conn = next) {
2506 if (conn->type == RX_CLIENT_CONNECTION) {
2507 MUTEX_ENTER(&rx_refcnt_mutex);
2509 MUTEX_EXIT(&rx_refcnt_mutex);
2510 #ifdef RX_ENABLE_LOCKS
2511 rxi_DestroyConnectionNoLock(conn);
2512 #else /* RX_ENABLE_LOCKS */
2513 rxi_DestroyConnection(conn);
2514 #endif /* RX_ENABLE_LOCKS */
2518 #ifdef RX_ENABLE_LOCKS
2519 while (rx_connCleanup_list) {
2520 struct rx_connection *conn;
2521 conn = rx_connCleanup_list;
2522 rx_connCleanup_list = rx_connCleanup_list->next;
2523 MUTEX_EXIT(&rx_connHashTable_lock);
2524 rxi_CleanupConnection(conn);
2525 MUTEX_ENTER(&rx_connHashTable_lock);
2527 MUTEX_EXIT(&rx_connHashTable_lock);
2528 #endif /* RX_ENABLE_LOCKS */
2533 afs_winsockCleanup();
2541 /* if we wakeup packet waiter too often, can get in loop with two
2542 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2544 rxi_PacketsUnWait(void)
2546 if (!rx_waitingForPackets) {
2550 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2551 return; /* still over quota */
2554 rx_waitingForPackets = 0;
2555 #ifdef RX_ENABLE_LOCKS
2556 CV_BROADCAST(&rx_waitingForPackets_cv);
2558 osi_rxWakeup(&rx_waitingForPackets);
2564 /* ------------------Internal interfaces------------------------- */
2566 /* Return this process's service structure for the
2567 * specified socket and service */
2568 static struct rx_service *
2569 rxi_FindService(osi_socket socket, u_short serviceId)
2571 struct rx_service **sp;
2572 for (sp = &rx_services[0]; *sp; sp++) {
2573 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2579 #ifdef RXDEBUG_PACKET
2580 #ifdef KDUMP_RX_LOCK
2581 static struct rx_call_rx_lock *rx_allCallsp = 0;
2583 static struct rx_call *rx_allCallsp = 0;
2585 #endif /* RXDEBUG_PACKET */
2587 /* Allocate a call structure, for the indicated channel of the
2588 * supplied connection. The mode and state of the call must be set by
2589 * the caller. Returns the call with mutex locked. */
2590 static struct rx_call *
2591 rxi_NewCall(struct rx_connection *conn, int channel)
2593 struct rx_call *call;
2594 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2595 struct rx_call *cp; /* Call pointer temp */
2596 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2597 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2599 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2601 /* Grab an existing call structure, or allocate a new one.
2602 * Existing call structures are assumed to have been left reset by
2604 MUTEX_ENTER(&rx_freeCallQueue_lock);
2606 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2608 * EXCEPT that the TQ might not yet be cleared out.
2609 * Skip over those with in-use TQs.
2612 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2613 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2619 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2620 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2621 call = queue_First(&rx_freeCallQueue, rx_call);
2622 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2624 if (rx_stats_active)
2625 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2626 MUTEX_EXIT(&rx_freeCallQueue_lock);
2627 MUTEX_ENTER(&call->lock);
2628 CLEAR_CALL_QUEUE_LOCK(call);
2629 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2630 /* Now, if TQ wasn't cleared earlier, do it now. */
2631 rxi_WaitforTQBusy(call);
2632 if (call->flags & RX_CALL_TQ_CLEARME) {
2633 rxi_ClearTransmitQueue(call, 1);
2634 /*queue_Init(&call->tq);*/
2636 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2637 /* Bind the call to its connection structure */
2639 rxi_ResetCall(call, 1);
2642 call = rxi_Alloc(sizeof(struct rx_call));
2643 #ifdef RXDEBUG_PACKET
2644 call->allNextp = rx_allCallsp;
2645 rx_allCallsp = call;
2647 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2648 #else /* RXDEBUG_PACKET */
2649 rx_atomic_inc(&rx_stats.nCallStructs);
2650 #endif /* RXDEBUG_PACKET */
2652 MUTEX_EXIT(&rx_freeCallQueue_lock);
2653 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2654 MUTEX_ENTER(&call->lock);
2655 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2656 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2657 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2659 /* Initialize once-only items */
2660 queue_Init(&call->tq);
2661 queue_Init(&call->rq);
2662 queue_Init(&call->iovq);
2663 #ifdef RXDEBUG_PACKET
2664 call->rqc = call->tqc = call->iovqc = 0;
2665 #endif /* RXDEBUG_PACKET */
2666 /* Bind the call to its connection structure (prereq for reset) */
2668 rxi_ResetCall(call, 1);
2670 call->channel = channel;
2671 call->callNumber = &conn->callNumber[channel];
2672 call->rwind = conn->rwind[channel];
2673 call->twind = conn->twind[channel];
2674 /* Note that the next expected call number is retained (in
2675 * conn->callNumber[i]), even if we reallocate the call structure
2677 conn->call[channel] = call;
2678 /* if the channel's never been used (== 0), we should start at 1, otherwise
2679 * the call number is valid from the last time this channel was used */
2680 if (*call->callNumber == 0)
2681 *call->callNumber = 1;
2686 /* A call has been inactive long enough that so we can throw away
2687 * state, including the call structure, which is placed on the call
2690 * call->lock amd rx_refcnt_mutex are held upon entry.
2691 * haveCTLock is set when called from rxi_ReapConnections.
2694 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2696 int channel = call->channel;
2697 struct rx_connection *conn = call->conn;
2700 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2701 (*call->callNumber)++;
2703 * We are setting the state to RX_STATE_RESET to
2704 * ensure that no one else will attempt to use this
2705 * call once we drop the refcnt lock. We must drop
2706 * the refcnt lock before calling rxi_ResetCall
2707 * because it cannot be held across acquiring the
2708 * freepktQ lock. NewCall does the same.
2710 call->state = RX_STATE_RESET;
2711 MUTEX_EXIT(&rx_refcnt_mutex);
2712 rxi_ResetCall(call, 0);
2714 MUTEX_ENTER(&conn->conn_call_lock);
2715 if (call->conn->call[channel] == call)
2716 call->conn->call[channel] = 0;
2717 MUTEX_EXIT(&conn->conn_call_lock);
2719 MUTEX_ENTER(&rx_freeCallQueue_lock);
2720 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2721 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2722 /* A call may be free even though its transmit queue is still in use.
2723 * Since we search the call list from head to tail, put busy calls at
2724 * the head of the list, and idle calls at the tail.
2726 if (call->flags & RX_CALL_TQ_BUSY)
2727 queue_Prepend(&rx_freeCallQueue, call);
2729 queue_Append(&rx_freeCallQueue, call);
2730 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2731 queue_Append(&rx_freeCallQueue, call);
2732 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2733 if (rx_stats_active)
2734 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2735 MUTEX_EXIT(&rx_freeCallQueue_lock);
2737 /* Destroy the connection if it was previously slated for
2738 * destruction, i.e. the Rx client code previously called
2739 * rx_DestroyConnection (client connections), or
2740 * rxi_ReapConnections called the same routine (server
2741 * connections). Only do this, however, if there are no
2742 * outstanding calls. Note that for fine grain locking, there appears
2743 * to be a deadlock in that rxi_FreeCall has a call locked and
2744 * DestroyConnectionNoLock locks each call in the conn. But note a
2745 * few lines up where we have removed this call from the conn.
2746 * If someone else destroys a connection, they either have no
2747 * call lock held or are going through this section of code.
2749 MUTEX_ENTER(&conn->conn_data_lock);
2750 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2751 MUTEX_ENTER(&rx_refcnt_mutex);
2753 MUTEX_EXIT(&rx_refcnt_mutex);
2754 MUTEX_EXIT(&conn->conn_data_lock);
2755 #ifdef RX_ENABLE_LOCKS
2757 rxi_DestroyConnectionNoLock(conn);
2759 rxi_DestroyConnection(conn);
2760 #else /* RX_ENABLE_LOCKS */
2761 rxi_DestroyConnection(conn);
2762 #endif /* RX_ENABLE_LOCKS */
2764 MUTEX_EXIT(&conn->conn_data_lock);
2766 MUTEX_ENTER(&rx_refcnt_mutex);
2769 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2770 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2773 rxi_Alloc(size_t size)
2777 if (rx_stats_active) {
2778 rx_atomic_add(&rxi_Allocsize, (int) size);
2779 rx_atomic_inc(&rxi_Alloccnt);
2783 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2784 afs_osi_Alloc_NoSleep(size);
2789 osi_Panic("rxi_Alloc error");
2795 rxi_Free(void *addr, size_t size)
2797 if (rx_stats_active) {
2798 rx_atomic_sub(&rxi_Allocsize, (int) size);
2799 rx_atomic_dec(&rxi_Alloccnt);
2801 osi_Free(addr, size);
2805 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2807 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2808 struct rx_peer *next = NULL;
2812 MUTEX_ENTER(&rx_peerHashTable_lock);
2814 peer_ptr = &rx_peerHashTable[0];
2815 peer_end = &rx_peerHashTable[rx_hashTableSize];
2818 for ( ; peer_ptr < peer_end; peer_ptr++) {
2821 for ( ; peer; peer = next) {
2823 if (host == peer->host)
2828 hashIndex = PEER_HASH(host, port);
2829 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2830 if ((peer->host == host) && (peer->port == port))
2835 MUTEX_ENTER(&rx_peerHashTable_lock);
2840 MUTEX_EXIT(&rx_peerHashTable_lock);
2842 MUTEX_ENTER(&peer->peer_lock);
2843 /* We don't handle dropping below min, so don't */
2844 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2845 peer->ifMTU=MIN(mtu, peer->ifMTU);
2846 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2847 /* if we tweaked this down, need to tune our peer MTU too */
2848 peer->MTU = MIN(peer->MTU, peer->natMTU);
2849 /* if we discovered a sub-1500 mtu, degrade */
2850 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2851 peer->maxDgramPackets = 1;
2852 /* We no longer have valid peer packet information */
2853 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2854 peer->maxPacketSize = 0;
2855 MUTEX_EXIT(&peer->peer_lock);
2857 MUTEX_ENTER(&rx_peerHashTable_lock);
2859 if (host && !port) {
2861 /* pick up where we left off */
2865 MUTEX_EXIT(&rx_peerHashTable_lock);
2868 /* Find the peer process represented by the supplied (host,port)
2869 * combination. If there is no appropriate active peer structure, a
2870 * new one will be allocated and initialized
2871 * The origPeer, if set, is a pointer to a peer structure on which the
2872 * refcount will be be decremented. This is used to replace the peer
2873 * structure hanging off a connection structure */
2875 rxi_FindPeer(afs_uint32 host, u_short port,
2876 struct rx_peer *origPeer, int create)
2880 hashIndex = PEER_HASH(host, port);
2881 MUTEX_ENTER(&rx_peerHashTable_lock);
2882 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2883 if ((pp->host == host) && (pp->port == port))
2888 pp = rxi_AllocPeer(); /* This bzero's *pp */
2889 pp->host = host; /* set here or in InitPeerParams is zero */
2891 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2892 queue_Init(&pp->congestionQueue);
2893 queue_Init(&pp->rpcStats);
2894 pp->next = rx_peerHashTable[hashIndex];
2895 rx_peerHashTable[hashIndex] = pp;
2896 rxi_InitPeerParams(pp);
2897 if (rx_stats_active)
2898 rx_atomic_inc(&rx_stats.nPeerStructs);
2905 origPeer->refCount--;
2906 MUTEX_EXIT(&rx_peerHashTable_lock);
2911 /* Find the connection at (host, port) started at epoch, and with the
2912 * given connection id. Creates the server connection if necessary.
2913 * The type specifies whether a client connection or a server
2914 * connection is desired. In both cases, (host, port) specify the
2915 * peer's (host, pair) pair. Client connections are not made
2916 * automatically by this routine. The parameter socket gives the
2917 * socket descriptor on which the packet was received. This is used,
2918 * in the case of server connections, to check that *new* connections
2919 * come via a valid (port, serviceId). Finally, the securityIndex
2920 * parameter must match the existing index for the connection. If a
2921 * server connection is created, it will be created using the supplied
2922 * index, if the index is valid for this service */
2923 struct rx_connection *
2924 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2925 u_short port, u_short serviceId, afs_uint32 cid,
2926 afs_uint32 epoch, int type, u_int securityIndex)
2928 int hashindex, flag, i;
2929 struct rx_connection *conn;
2930 hashindex = CONN_HASH(host, port, cid, epoch, type);
2931 MUTEX_ENTER(&rx_connHashTable_lock);
2932 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2933 rx_connHashTable[hashindex],
2936 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2937 && (epoch == conn->epoch)) {
2938 struct rx_peer *pp = conn->peer;
2939 if (securityIndex != conn->securityIndex) {
2940 /* this isn't supposed to happen, but someone could forge a packet
2941 * like this, and there seems to be some CM bug that makes this
2942 * happen from time to time -- in which case, the fileserver
2944 MUTEX_EXIT(&rx_connHashTable_lock);
2945 return (struct rx_connection *)0;
2947 if (pp->host == host && pp->port == port)
2949 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2951 /* So what happens when it's a callback connection? */
2952 if ( /*type == RX_CLIENT_CONNECTION && */
2953 (conn->epoch & 0x80000000))
2957 /* the connection rxLastConn that was used the last time is not the
2958 ** one we are looking for now. Hence, start searching in the hash */
2960 conn = rx_connHashTable[hashindex];
2965 struct rx_service *service;
2966 if (type == RX_CLIENT_CONNECTION) {
2967 MUTEX_EXIT(&rx_connHashTable_lock);
2968 return (struct rx_connection *)0;
2970 service = rxi_FindService(socket, serviceId);
2971 if (!service || (securityIndex >= service->nSecurityObjects)
2972 || (service->securityObjects[securityIndex] == 0)) {
2973 MUTEX_EXIT(&rx_connHashTable_lock);
2974 return (struct rx_connection *)0;
2976 conn = rxi_AllocConnection(); /* This bzero's the connection */
2977 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2978 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2979 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2980 conn->next = rx_connHashTable[hashindex];
2981 rx_connHashTable[hashindex] = conn;
2982 conn->peer = rxi_FindPeer(host, port, 0, 1);
2983 conn->type = RX_SERVER_CONNECTION;
2984 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2985 conn->epoch = epoch;
2986 conn->cid = cid & RX_CIDMASK;
2987 /* conn->serial = conn->lastSerial = 0; */
2988 /* conn->timeout = 0; */
2989 conn->ackRate = RX_FAST_ACK_RATE;
2990 conn->service = service;
2991 conn->serviceId = serviceId;
2992 conn->securityIndex = securityIndex;
2993 conn->securityObject = service->securityObjects[securityIndex];
2994 conn->nSpecific = 0;
2995 conn->specific = NULL;
2996 rx_SetConnDeadTime(conn, service->connDeadTime);
2997 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2998 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2999 for (i = 0; i < RX_MAXCALLS; i++) {
3000 conn->twind[i] = rx_initSendWindow;
3001 conn->rwind[i] = rx_initReceiveWindow;
3003 /* Notify security object of the new connection */
3004 RXS_NewConnection(conn->securityObject, conn);
3005 /* XXXX Connection timeout? */
3006 if (service->newConnProc)
3007 (*service->newConnProc) (conn);
3008 if (rx_stats_active)
3009 rx_atomic_inc(&rx_stats.nServerConns);
3012 MUTEX_ENTER(&rx_refcnt_mutex);
3014 MUTEX_EXIT(&rx_refcnt_mutex);
3016 rxLastConn = conn; /* store this connection as the last conn used */
3017 MUTEX_EXIT(&rx_connHashTable_lock);
3022 * Timeout a call on a busy call channel if appropriate.
3024 * @param[in] call The busy call.
3026 * @pre 'call' is marked as busy (namely,
3027 * call->conn->lastBusy[call->channel] != 0)
3029 * @pre call->lock is held
3030 * @pre rxi_busyChannelError is nonzero
3032 * @note call->lock is dropped and reacquired
3035 rxi_CheckBusy(struct rx_call *call)
3037 struct rx_connection *conn = call->conn;
3038 int channel = call->channel;
3039 int freechannel = 0;
3041 afs_uint32 callNumber = *call->callNumber;
3043 MUTEX_EXIT(&call->lock);
3045 MUTEX_ENTER(&conn->conn_call_lock);
3047 /* Are there any other call slots on this conn that we should try? Look for
3048 * slots that are empty and are either non-busy, or were marked as busy
3049 * longer than conn->secondsUntilDead seconds before this call started. */
3051 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3053 /* only look at channels that aren't us */
3057 if (conn->lastBusy[i]) {
3058 /* if this channel looked busy too recently, don't look at it */
3059 if (conn->lastBusy[i] >= call->startTime.sec) {
3062 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3067 if (conn->call[i]) {
3068 struct rx_call *tcall = conn->call[i];
3069 MUTEX_ENTER(&tcall->lock);
3070 if (tcall->state == RX_STATE_DALLY) {
3073 MUTEX_EXIT(&tcall->lock);
3079 MUTEX_EXIT(&conn->conn_call_lock);
3081 MUTEX_ENTER(&call->lock);
3083 /* Since the call->lock and conn->conn_call_lock have been released it is
3084 * possible that (1) the call may no longer be busy and/or (2) the call may
3085 * have been reused by another waiting thread. Therefore, we must confirm
3086 * that the call state has not changed when deciding whether or not to
3087 * force this application thread to retry by forcing a Timeout error. */
3089 if (freechannel && *call->callNumber == callNumber &&
3090 (call->flags & RX_CALL_PEER_BUSY)) {
3091 /* Since 'freechannel' is set, there exists another channel in this
3092 * rx_conn that the application thread might be able to use. We know
3093 * that we have the correct call since callNumber is unchanged, and we
3094 * know that the call is still busy. So, set the call error state to
3095 * rxi_busyChannelError so the application can retry the request,
3096 * presumably on a less-busy call channel. */
3098 rxi_CallError(call, rxi_busyChannelError);
3102 /* There are two packet tracing routines available for testing and monitoring
3103 * Rx. One is called just after every packet is received and the other is
3104 * called just before every packet is sent. Received packets, have had their
3105 * headers decoded, and packets to be sent have not yet had their headers
3106 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3107 * containing the network address. Both can be modified. The return value, if
3108 * non-zero, indicates that the packet should be dropped. */
3110 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3111 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3113 /* A packet has been received off the interface. Np is the packet, socket is
3114 * the socket number it was received from (useful in determining which service
3115 * this packet corresponds to), and (host, port) reflect the host,port of the
3116 * sender. This call returns the packet to the caller if it is finished with
3117 * it, rather than de-allocating it, just as a small performance hack */
3120 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3121 afs_uint32 host, u_short port, int *tnop,
3122 struct rx_call **newcallp)
3124 struct rx_call *call;
3125 struct rx_connection *conn;
3127 afs_uint32 currentCallNumber;
3133 struct rx_packet *tnp;
3136 /* We don't print out the packet until now because (1) the time may not be
3137 * accurate enough until now in the lwp implementation (rx_Listener only gets
3138 * the time after the packet is read) and (2) from a protocol point of view,
3139 * this is the first time the packet has been seen */
3140 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3141 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3142 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3143 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3144 np->header.epoch, np->header.cid, np->header.callNumber,
3145 np->header.seq, np->header.flags, np));
3148 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3149 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3152 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3153 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3156 /* If an input tracer function is defined, call it with the packet and
3157 * network address. Note this function may modify its arguments. */
3158 if (rx_justReceived) {
3159 struct sockaddr_in addr;
3161 addr.sin_family = AF_INET;
3162 addr.sin_port = port;
3163 addr.sin_addr.s_addr = host;
3164 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3165 addr.sin_len = sizeof(addr);
3166 #endif /* AFS_OSF_ENV */
3167 drop = (*rx_justReceived) (np, &addr);
3168 /* drop packet if return value is non-zero */
3171 port = addr.sin_port; /* in case fcn changed addr */
3172 host = addr.sin_addr.s_addr;
3176 /* If packet was not sent by the client, then *we* must be the client */
3177 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3178 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3180 /* Find the connection (or fabricate one, if we're the server & if
3181 * necessary) associated with this packet */
3183 rxi_FindConnection(socket, host, port, np->header.serviceId,
3184 np->header.cid, np->header.epoch, type,
3185 np->header.securityIndex);
3188 /* If no connection found or fabricated, just ignore the packet.
3189 * (An argument could be made for sending an abort packet for
3194 /* If the connection is in an error state, send an abort packet and ignore
3195 * the incoming packet */
3197 /* Don't respond to an abort packet--we don't want loops! */
3198 MUTEX_ENTER(&conn->conn_data_lock);
3199 if (np->header.type != RX_PACKET_TYPE_ABORT)
3200 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3201 putConnection(conn);
3202 MUTEX_EXIT(&conn->conn_data_lock);
3206 /* Check for connection-only requests (i.e. not call specific). */
3207 if (np->header.callNumber == 0) {
3208 switch (np->header.type) {
3209 case RX_PACKET_TYPE_ABORT: {
3210 /* What if the supplied error is zero? */
3211 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3212 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3213 rxi_ConnectionError(conn, errcode);
3214 putConnection(conn);
3217 case RX_PACKET_TYPE_CHALLENGE:
3218 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3219 putConnection(conn);
3221 case RX_PACKET_TYPE_RESPONSE:
3222 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3223 putConnection(conn);
3225 case RX_PACKET_TYPE_PARAMS:
3226 case RX_PACKET_TYPE_PARAMS + 1:
3227 case RX_PACKET_TYPE_PARAMS + 2:
3228 /* ignore these packet types for now */
3229 putConnection(conn);
3233 /* Should not reach here, unless the peer is broken: send an
3235 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3236 MUTEX_ENTER(&conn->conn_data_lock);
3237 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3238 putConnection(conn);
3239 MUTEX_EXIT(&conn->conn_data_lock);
3244 channel = np->header.cid & RX_CHANNELMASK;
3245 call = conn->call[channel];
3248 MUTEX_ENTER(&call->lock);
3249 currentCallNumber = conn->callNumber[channel];
3250 } else if (type == RX_SERVER_CONNECTION) { /* No call allocated */
3251 MUTEX_ENTER(&conn->conn_call_lock);
3252 call = conn->call[channel];
3254 MUTEX_ENTER(&call->lock);
3255 MUTEX_EXIT(&conn->conn_call_lock);
3256 currentCallNumber = conn->callNumber[channel];
3258 call = rxi_NewCall(conn, channel); /* returns locked call */
3259 MUTEX_EXIT(&conn->conn_call_lock);
3260 *call->callNumber = currentCallNumber = np->header.callNumber;
3262 if (np->header.callNumber == 0)
3263 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3264 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3265 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3266 np->header.flags, np, np->length));
3268 call->state = RX_STATE_PRECALL;
3269 clock_GetTime(&call->queueTime);
3270 hzero(call->bytesSent);
3271 hzero(call->bytesRcvd);
3273 * If the number of queued calls exceeds the overload
3274 * threshold then abort this call.
3276 if ((rx_BusyThreshold > 0) &&
3277 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3278 struct rx_packet *tp;
3280 rxi_CallError(call, rx_BusyError);
3281 tp = rxi_SendCallAbort(call, np, 1, 0);
3282 MUTEX_EXIT(&call->lock);
3283 putConnection(conn);
3284 if (rx_stats_active)
3285 rx_atomic_inc(&rx_stats.nBusies);
3288 rxi_KeepAliveOn(call);
3290 } else { /* RX_CLIENT_CONNECTION and No call allocated */
3291 /* This packet can't be for this call. If the new call address is
3292 * 0 then no call is running on this channel. If there is a call
3293 * then, since this is a client connection we're getting data for
3294 * it must be for the previous call.
3296 if (rx_stats_active)
3297 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3298 putConnection(conn);
3302 /* There is a non-NULL locked call at this point */
3303 if (type == RX_SERVER_CONNECTION) { /* We're the server */
3304 if (np->header.callNumber < currentCallNumber) {
3305 MUTEX_EXIT(&call->lock);
3306 if (rx_stats_active)
3307 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3308 putConnection(conn);
3310 } else if (np->header.callNumber != currentCallNumber) {
3311 /* Wait until the transmit queue is idle before deciding
3312 * whether to reset the current call. Chances are that the
3313 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3316 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3317 if (call->state == RX_STATE_ACTIVE) {
3318 rxi_WaitforTQBusy(call);
3320 * If we entered error state while waiting,
3321 * must call rxi_CallError to permit rxi_ResetCall
3322 * to processed when the tqWaiter count hits zero.
3325 rxi_CallError(call, call->error);
3326 MUTEX_EXIT(&call->lock);
3327 putConnection(conn);
3331 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3332 /* If the new call cannot be taken right now send a busy and set
3333 * the error condition in this call, so that it terminates as
3334 * quickly as possible */
3335 if (call->state == RX_STATE_ACTIVE) {
3336 struct rx_packet *tp;
3338 rxi_CallError(call, RX_CALL_DEAD);
3339 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3341 MUTEX_EXIT(&call->lock);
3342 putConnection(conn);
3345 rxi_ResetCall(call, 0);
3346 *call->callNumber = np->header.callNumber;
3348 if (np->header.callNumber == 0)
3349 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3350 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3351 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3352 np->header.flags, np, np->length));
3354 call->state = RX_STATE_PRECALL;
3355 clock_GetTime(&call->queueTime);
3356 hzero(call->bytesSent);
3357 hzero(call->bytesRcvd);
3359 * If the number of queued calls exceeds the overload
3360 * threshold then abort this call.
3362 if ((rx_BusyThreshold > 0) &&
3363 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3364 struct rx_packet *tp;
3366 rxi_CallError(call, rx_BusyError);
3367 tp = rxi_SendCallAbort(call, np, 1, 0);
3368 MUTEX_EXIT(&call->lock);
3369 putConnection(conn);
3370 if (rx_stats_active)
3371 rx_atomic_inc(&rx_stats.nBusies);
3374 rxi_KeepAliveOn(call);
3376 /* Continuing call; do nothing here. */
3378 } else { /* we're the client */
3379 /* Ignore all incoming acknowledgements for calls in DALLY state */
3380 if ((call->state == RX_STATE_DALLY)
3381 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3382 if (rx_stats_active)
3383 rx_atomic_inc(&rx_stats.ignorePacketDally);
3384 MUTEX_EXIT(&call->lock);
3385 putConnection(conn);
3389 /* Ignore anything that's not relevant to the current call. If there
3390 * isn't a current call, then no packet is relevant. */
3391 if (np->header.callNumber != currentCallNumber) {
3392 if (rx_stats_active)
3393 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3394 MUTEX_EXIT(&call->lock);
3395 putConnection(conn);
3398 /* If the service security object index stamped in the packet does not
3399 * match the connection's security index, ignore the packet */
3400 if (np->header.securityIndex != conn->securityIndex) {
3401 MUTEX_EXIT(&call->lock);
3402 putConnection(conn);
3406 /* If we're receiving the response, then all transmit packets are
3407 * implicitly acknowledged. Get rid of them. */
3408 if (np->header.type == RX_PACKET_TYPE_DATA) {
3409 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3410 /* XXX Hack. Because we must release the global rx lock when
3411 * sending packets (osi_NetSend) we drop all acks while we're
3412 * traversing the tq in rxi_Start sending packets out because
3413 * packets may move to the freePacketQueue as result of being here!
3414 * So we drop these packets until we're safely out of the
3415 * traversing. Really ugly!
3416 * For fine grain RX locking, we set the acked field in the
3417 * packets and let rxi_Start remove them from the transmit queue.
3419 if (call->flags & RX_CALL_TQ_BUSY) {
3420 #ifdef RX_ENABLE_LOCKS
3421 rxi_SetAcksInTransmitQueue(call);
3423 putConnection(conn);
3424 return np; /* xmitting; drop packet */
3427 rxi_ClearTransmitQueue(call, 0);
3429 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3430 rxi_ClearTransmitQueue(call, 0);
3431 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3433 if (np->header.type == RX_PACKET_TYPE_ACK) {
3434 /* now check to see if this is an ack packet acknowledging that the
3435 * server actually *lost* some hard-acked data. If this happens we
3436 * ignore this packet, as it may indicate that the server restarted in
3437 * the middle of a call. It is also possible that this is an old ack
3438 * packet. We don't abort the connection in this case, because this
3439 * *might* just be an old ack packet. The right way to detect a server
3440 * restart in the midst of a call is to notice that the server epoch
3442 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3443 * XXX unacknowledged. I think that this is off-by-one, but
3444 * XXX I don't dare change it just yet, since it will
3445 * XXX interact badly with the server-restart detection
3446 * XXX code in receiveackpacket. */
3447 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3448 if (rx_stats_active)
3449 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3450 MUTEX_EXIT(&call->lock);
3451 putConnection(conn);
3455 } /* else not a data packet */
3458 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3459 /* Set remote user defined status from packet */
3460 call->remoteStatus = np->header.userStatus;
3462 /* Note the gap between the expected next packet and the actual
3463 * packet that arrived, when the new packet has a smaller serial number
3464 * than expected. Rioses frequently reorder packets all by themselves,
3465 * so this will be quite important with very large window sizes.
3466 * Skew is checked against 0 here to avoid any dependence on the type of
3467 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3469 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3470 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3471 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3473 MUTEX_ENTER(&conn->conn_data_lock);
3474 skew = conn->lastSerial - np->header.serial;
3475 conn->lastSerial = np->header.serial;
3476 MUTEX_EXIT(&conn->conn_data_lock);
3478 struct rx_peer *peer;
3480 if (skew > peer->inPacketSkew) {
3481 dpf(("*** In skew changed from %d to %d\n",
3482 peer->inPacketSkew, skew));
3483 peer->inPacketSkew = skew;
3487 /* Now do packet type-specific processing */
3488 switch (np->header.type) {
3489 case RX_PACKET_TYPE_DATA:
3490 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3493 case RX_PACKET_TYPE_ACK:
3494 /* Respond immediately to ack packets requesting acknowledgement
3496 if (np->header.flags & RX_REQUEST_ACK) {
3498 (void)rxi_SendCallAbort(call, 0, 1, 0);
3500 (void)rxi_SendAck(call, 0, np->header.serial,
3501 RX_ACK_PING_RESPONSE, 1);
3503 np = rxi_ReceiveAckPacket(call, np, 1);
3505 case RX_PACKET_TYPE_ABORT: {
3506 /* An abort packet: reset the call, passing the error up to the user. */
3507 /* What if error is zero? */
3508 /* What if the error is -1? the application will treat it as a timeout. */
3509 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3510 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3511 rxi_CallError(call, errdata);
3512 MUTEX_EXIT(&call->lock);
3513 putConnection(conn);
3514 return np; /* xmitting; drop packet */
3516 case RX_PACKET_TYPE_BUSY: {
3517 struct clock busyTime;
3519 clock_GetTime(&busyTime);
3521 MUTEX_EXIT(&call->lock);
3523 MUTEX_ENTER(&conn->conn_call_lock);
3524 MUTEX_ENTER(&call->lock);
3525 conn->lastBusy[call->channel] = busyTime.sec;
3526 call->flags |= RX_CALL_PEER_BUSY;
3527 MUTEX_EXIT(&call->lock);
3528 MUTEX_EXIT(&conn->conn_call_lock);
3530 putConnection(conn);
3534 case RX_PACKET_TYPE_ACKALL:
3535 /* All packets acknowledged, so we can drop all packets previously
3536 * readied for sending */
3537 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3538 /* XXX Hack. We because we can't release the global rx lock when
3539 * sending packets (osi_NetSend) we drop all ack pkts while we're
3540 * traversing the tq in rxi_Start sending packets out because
3541 * packets may move to the freePacketQueue as result of being
3542 * here! So we drop these packets until we're safely out of the
3543 * traversing. Really ugly!
3544 * For fine grain RX locking, we set the acked field in the packets
3545 * and let rxi_Start remove the packets from the transmit queue.
3547 if (call->flags & RX_CALL_TQ_BUSY) {
3548 #ifdef RX_ENABLE_LOCKS
3549 rxi_SetAcksInTransmitQueue(call);
3551 #else /* RX_ENABLE_LOCKS */
3552 MUTEX_EXIT(&call->lock);
3553 putConnection(conn);
3554 return np; /* xmitting; drop packet */
3555 #endif /* RX_ENABLE_LOCKS */
3557 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3558 rxi_ClearTransmitQueue(call, 0);
3561 /* Should not reach here, unless the peer is broken: send an abort
3563 rxi_CallError(call, RX_PROTOCOL_ERROR);
3564 np = rxi_SendCallAbort(call, np, 1, 0);
3567 /* Note when this last legitimate packet was received, for keep-alive
3568 * processing. Note, we delay getting the time until now in the hope that
3569 * the packet will be delivered to the user before any get time is required
3570 * (if not, then the time won't actually be re-evaluated here). */
3571 call->lastReceiveTime = clock_Sec();
3572 /* we've received a legit packet, so the channel is not busy */
3573 call->flags &= ~RX_CALL_PEER_BUSY;
3574 MUTEX_EXIT(&call->lock);
3575 putConnection(conn);
3579 /* return true if this is an "interesting" connection from the point of view
3580 of someone trying to debug the system */
3582 rxi_IsConnInteresting(struct rx_connection *aconn)
3585 struct rx_call *tcall;
3587 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3590 for (i = 0; i < RX_MAXCALLS; i++) {
3591 tcall = aconn->call[i];
3593 if ((tcall->state == RX_STATE_PRECALL)
3594 || (tcall->state == RX_STATE_ACTIVE))
3596 if ((tcall->mode == RX_MODE_SENDING)
3597 || (tcall->mode == RX_MODE_RECEIVING))
3605 /* if this is one of the last few packets AND it wouldn't be used by the
3606 receiving call to immediately satisfy a read request, then drop it on
3607 the floor, since accepting it might prevent a lock-holding thread from
3608 making progress in its reading. If a call has been cleared while in
3609 the precall state then ignore all subsequent packets until the call
3610 is assigned to a thread. */
3613 TooLow(struct rx_packet *ap, struct rx_call *acall)
3617 MUTEX_ENTER(&rx_quota_mutex);
3618 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3619 && (acall->state == RX_STATE_PRECALL))
3620 || ((rx_nFreePackets < rxi_dataQuota + 2)
3621 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3622 && (acall->flags & RX_CALL_READER_WAIT)))) {
3625 MUTEX_EXIT(&rx_quota_mutex);
3631 * Clear the attach wait flag on a connection and proceed.
3633 * Any processing waiting for a connection to be attached should be
3634 * unblocked. We clear the flag and do any other needed tasks.
3637 * the conn to unmark waiting for attach
3639 * @pre conn's conn_data_lock must be locked before calling this function
3643 rxi_ConnClearAttachWait(struct rx_connection *conn)
3645 /* Indicate that rxi_CheckReachEvent is no longer running by
3646 * clearing the flag. Must be atomic under conn_data_lock to
3647 * avoid a new call slipping by: rxi_CheckConnReach holds
3648 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3650 conn->flags &= ~RX_CONN_ATTACHWAIT;
3651 if (conn->flags & RX_CONN_NAT_PING) {
3652 conn->flags &= ~RX_CONN_NAT_PING;
3653 rxi_ScheduleNatKeepAliveEvent(conn);
3658 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3660 struct rx_connection *conn = arg1;
3661 struct rx_call *acall = arg2;
3662 struct rx_call *call = acall;
3663 struct clock when, now;
3666 MUTEX_ENTER(&conn->conn_data_lock);
3669 rxevent_Put(conn->checkReachEvent);
3670 conn->checkReachEvent = NULL;
3673 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3675 putConnection(conn);
3677 MUTEX_EXIT(&conn->conn_data_lock);
3681 MUTEX_ENTER(&conn->conn_call_lock);
3682 MUTEX_ENTER(&conn->conn_data_lock);
3683 for (i = 0; i < RX_MAXCALLS; i++) {
3684 struct rx_call *tc = conn->call[i];
3685 if (tc && tc->state == RX_STATE_PRECALL) {
3691 rxi_ConnClearAttachWait(conn);
3692 MUTEX_EXIT(&conn->conn_data_lock);
3693 MUTEX_EXIT(&conn->conn_call_lock);
3698 MUTEX_ENTER(&call->lock);
3699 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3701 MUTEX_EXIT(&call->lock);
3703 clock_GetTime(&now);
3705 when.sec += RX_CHECKREACH_TIMEOUT;
3706 MUTEX_ENTER(&conn->conn_data_lock);
3707 if (!conn->checkReachEvent) {
3708 MUTEX_ENTER(&rx_refcnt_mutex);
3710 MUTEX_EXIT(&rx_refcnt_mutex);
3711 conn->checkReachEvent = rxevent_Post(&when, &now,
3712 rxi_CheckReachEvent, conn,
3715 MUTEX_EXIT(&conn->conn_data_lock);
3721 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3723 struct rx_service *service = conn->service;
3724 struct rx_peer *peer = conn->peer;
3725 afs_uint32 now, lastReach;
3727 if (service->checkReach == 0)
3731 MUTEX_ENTER(&peer->peer_lock);
3732 lastReach = peer->lastReachTime;
3733 MUTEX_EXIT(&peer->peer_lock);
3734 if (now - lastReach < RX_CHECKREACH_TTL)
3737 MUTEX_ENTER(&conn->conn_data_lock);
3738 if (conn->flags & RX_CONN_ATTACHWAIT) {
3739 MUTEX_EXIT(&conn->conn_data_lock);
3742 conn->flags |= RX_CONN_ATTACHWAIT;
3743 MUTEX_EXIT(&conn->conn_data_lock);
3744 if (!conn->checkReachEvent)
3745 rxi_CheckReachEvent(NULL, conn, call, 0);
3750 /* try to attach call, if authentication is complete */
3752 TryAttach(struct rx_call *acall, osi_socket socket,
3753 int *tnop, struct rx_call **newcallp,
3756 struct rx_connection *conn = acall->conn;
3758 if (conn->type == RX_SERVER_CONNECTION
3759 && acall->state == RX_STATE_PRECALL) {
3760 /* Don't attach until we have any req'd. authentication. */
3761 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3762 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3763 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3764 /* Note: this does not necessarily succeed; there
3765 * may not any proc available
3768 rxi_ChallengeOn(acall->conn);
3773 /* A data packet has been received off the interface. This packet is
3774 * appropriate to the call (the call is in the right state, etc.). This
3775 * routine can return a packet to the caller, for re-use */
3778 rxi_ReceiveDataPacket(struct rx_call *call,
3779 struct rx_packet *np, int istack,
3780 osi_socket socket, afs_uint32 host, u_short port,
3781 int *tnop, struct rx_call **newcallp)
3783 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3788 afs_uint32 serial=0, flags=0;
3790 struct rx_packet *tnp;
3791 if (rx_stats_active)
3792 rx_atomic_inc(&rx_stats.dataPacketsRead);
3795 /* If there are no packet buffers, drop this new packet, unless we can find
3796 * packet buffers from inactive calls */
3798 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3799 MUTEX_ENTER(&rx_freePktQ_lock);
3800 rxi_NeedMorePackets = TRUE;
3801 MUTEX_EXIT(&rx_freePktQ_lock);
3802 if (rx_stats_active)
3803 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3804 call->rprev = np->header.serial;
3805 rxi_calltrace(RX_TRACE_DROP, call);
3806 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3807 /* We used to clear the receive queue here, in an attempt to free
3808 * packets. However this is unsafe if the queue has received a
3809 * soft ACK for the final packet */
3810 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3812 /* we've damaged this call already, might as well do it in. */
3818 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3819 * packet is one of several packets transmitted as a single
3820 * datagram. Do not send any soft or hard acks until all packets
3821 * in a jumbogram have been processed. Send negative acks right away.
3823 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3824 /* tnp is non-null when there are more packets in the
3825 * current jumbo gram */
3832 seq = np->header.seq;
3833 serial = np->header.serial;
3834 flags = np->header.flags;
3836 /* If the call is in an error state, send an abort message */
3838 return rxi_SendCallAbort(call, np, istack, 0);
3840 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3841 * AFS 3.5 jumbogram. */
3842 if (flags & RX_JUMBO_PACKET) {
3843 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3848 if (np->header.spare != 0) {
3849 MUTEX_ENTER(&call->conn->conn_data_lock);
3850 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3851 MUTEX_EXIT(&call->conn->conn_data_lock);
3854 /* The usual case is that this is the expected next packet */
3855 if (seq == call->rnext) {
3857 /* Check to make sure it is not a duplicate of one already queued */
3858 if (queue_IsNotEmpty(&call->rq)
3859 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3860 if (rx_stats_active)
3861 rx_atomic_inc(&rx_stats.dupPacketsRead);
3862 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3863 rxevent_Cancel(&call->delayedAckEvent, call,
3864 RX_CALL_REFCOUNT_DELAY);
3865 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3871 /* It's the next packet. Stick it on the receive queue
3872 * for this call. Set newPackets to make sure we wake
3873 * the reader once all packets have been processed */
3874 #ifdef RX_TRACK_PACKETS
3875 np->flags |= RX_PKTFLAG_RQ;
3877 queue_Prepend(&call->rq, np);
3878 #ifdef RXDEBUG_PACKET
3880 #endif /* RXDEBUG_PACKET */
3882 np = NULL; /* We can't use this anymore */
3885 /* If an ack is requested then set a flag to make sure we
3886 * send an acknowledgement for this packet */
3887 if (flags & RX_REQUEST_ACK) {
3888 ackNeeded = RX_ACK_REQUESTED;
3891 /* Keep track of whether we have received the last packet */
3892 if (flags & RX_LAST_PACKET) {
3893 call->flags |= RX_CALL_HAVE_LAST;
3897 /* Check whether we have all of the packets for this call */
3898 if (call->flags & RX_CALL_HAVE_LAST) {
3899 afs_uint32 tseq; /* temporary sequence number */
3900 struct rx_packet *tp; /* Temporary packet pointer */
3901 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3903 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3904 if (tseq != tp->header.seq)
3906 if (tp->header.flags & RX_LAST_PACKET) {
3907 call->flags |= RX_CALL_RECEIVE_DONE;
3914 /* Provide asynchronous notification for those who want it
3915 * (e.g. multi rx) */
3916 if (call->arrivalProc) {
3917 (*call->arrivalProc) (call, call->arrivalProcHandle,
3918 call->arrivalProcArg);
3919 call->arrivalProc = (void (*)())0;
3922 /* Update last packet received */
3925 /* If there is no server process serving this call, grab
3926 * one, if available. We only need to do this once. If a
3927 * server thread is available, this thread becomes a server
3928 * thread and the server thread becomes a listener thread. */
3930 TryAttach(call, socket, tnop, newcallp, 0);
3933 /* This is not the expected next packet. */
3935 /* Determine whether this is a new or old packet, and if it's
3936 * a new one, whether it fits into the current receive window.
3937 * Also figure out whether the packet was delivered in sequence.
3938 * We use the prev variable to determine whether the new packet
3939 * is the successor of its immediate predecessor in the
3940 * receive queue, and the missing flag to determine whether
3941 * any of this packets predecessors are missing. */
3943 afs_uint32 prev; /* "Previous packet" sequence number */
3944 struct rx_packet *tp; /* Temporary packet pointer */
3945 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3946 int missing; /* Are any predecessors missing? */
3948 /* If the new packet's sequence number has been sent to the
3949 * application already, then this is a duplicate */
3950 if (seq < call->rnext) {
3951 if (rx_stats_active)
3952 rx_atomic_inc(&rx_stats.dupPacketsRead);
3953 rxevent_Cancel(&call->delayedAckEvent, call,
3954 RX_CALL_REFCOUNT_DELAY);
3955 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3961 /* If the sequence number is greater than what can be
3962 * accomodated by the current window, then send a negative
3963 * acknowledge and drop the packet */
3964 if ((call->rnext + call->rwind) <= seq) {
3965 rxevent_Cancel(&call->delayedAckEvent, call,
3966 RX_CALL_REFCOUNT_DELAY);
3967 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3974 /* Look for the packet in the queue of old received packets */
3975 for (prev = call->rnext - 1, missing =
3976 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3977 /*Check for duplicate packet */
3978 if (seq == tp->header.seq) {
3979 if (rx_stats_active)
3980 rx_atomic_inc(&rx_stats.dupPacketsRead);
3981 rxevent_Cancel(&call->delayedAckEvent, call,
3982 RX_CALL_REFCOUNT_DELAY);
3983 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3989 /* If we find a higher sequence packet, break out and
3990 * insert the new packet here. */
3991 if (seq < tp->header.seq)
3993 /* Check for missing packet */
3994 if (tp->header.seq != prev + 1) {
3998 prev = tp->header.seq;
4001 /* Keep track of whether we have received the last packet. */
4002 if (flags & RX_LAST_PACKET) {
4003 call->flags |= RX_CALL_HAVE_LAST;
4006 /* It's within the window: add it to the the receive queue.
4007 * tp is left by the previous loop either pointing at the
4008 * packet before which to insert the new packet, or at the
4009 * queue head if the queue is empty or the packet should be
4011 #ifdef RX_TRACK_PACKETS
4012 np->flags |= RX_PKTFLAG_RQ;
4014 #ifdef RXDEBUG_PACKET
4016 #endif /* RXDEBUG_PACKET */
4017 queue_InsertBefore(tp, np);
4021 /* Check whether we have all of the packets for this call */
4022 if ((call->flags & RX_CALL_HAVE_LAST)
4023 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4024 afs_uint32 tseq; /* temporary sequence number */
4027 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
4028 if (tseq != tp->header.seq)
4030 if (tp->header.flags & RX_LAST_PACKET) {
4031 call->flags |= RX_CALL_RECEIVE_DONE;
4038 /* We need to send an ack of the packet is out of sequence,
4039 * or if an ack was requested by the peer. */
4040 if (seq != prev + 1 || missing) {
4041 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4042 } else if (flags & RX_REQUEST_ACK) {
4043 ackNeeded = RX_ACK_REQUESTED;
4046 /* Acknowledge the last packet for each call */
4047 if (flags & RX_LAST_PACKET) {
4058 * If the receiver is waiting for an iovec, fill the iovec
4059 * using the data from the receive queue */
4060 if (call->flags & RX_CALL_IOVEC_WAIT) {
4061 didHardAck = rxi_FillReadVec(call, serial);
4062 /* the call may have been aborted */
4071 /* Wakeup the reader if any */
4072 if ((call->flags & RX_CALL_READER_WAIT)
4073 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4074 || (call->iovNext >= call->iovMax)
4075 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4076 call->flags &= ~RX_CALL_READER_WAIT;
4077 #ifdef RX_ENABLE_LOCKS
4078 CV_BROADCAST(&call->cv_rq);
4080 osi_rxWakeup(&call->rq);
4086 * Send an ack when requested by the peer, or once every
4087 * rxi_SoftAckRate packets until the last packet has been
4088 * received. Always send a soft ack for the last packet in
4089 * the server's reply. */
4091 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4092 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4093 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4094 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4095 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4096 } else if (call->nSoftAcks) {
4097 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4098 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4100 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4101 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4102 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4109 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4111 struct rx_peer *peer = conn->peer;
4113 MUTEX_ENTER(&peer->peer_lock);
4114 peer->lastReachTime = clock_Sec();
4115 MUTEX_EXIT(&peer->peer_lock);
4117 MUTEX_ENTER(&conn->conn_data_lock);
4118 if (conn->flags & RX_CONN_ATTACHWAIT) {
4121 rxi_ConnClearAttachWait(conn);
4122 MUTEX_EXIT(&conn->conn_data_lock);
4124 for (i = 0; i < RX_MAXCALLS; i++) {
4125 struct rx_call *call = conn->call[i];
4128 MUTEX_ENTER(&call->lock);
4129 /* tnop can be null if newcallp is null */
4130 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4132 MUTEX_EXIT(&call->lock);
4136 MUTEX_EXIT(&conn->conn_data_lock);
4139 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4141 rx_ack_reason(int reason)
4144 case RX_ACK_REQUESTED:
4146 case RX_ACK_DUPLICATE:
4148 case RX_ACK_OUT_OF_SEQUENCE:
4150 case RX_ACK_EXCEEDS_WINDOW:
4152 case RX_ACK_NOSPACE:
4156 case RX_ACK_PING_RESPONSE:
4169 /* The real smarts of the whole thing. */
4171 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4174 struct rx_ackPacket *ap;
4176 struct rx_packet *tp;
4177 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
4178 struct rx_connection *conn = call->conn;
4179 struct rx_peer *peer = conn->peer;
4180 struct clock now; /* Current time, for RTT calculations */
4184 /* because there are CM's that are bogus, sending weird values for this. */
4185 afs_uint32 skew = 0;
4190 int newAckCount = 0;
4191 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4192 int pktsize = 0; /* Set if we need to update the peer mtu */
4193 int conn_data_locked = 0;
4195 if (rx_stats_active)
4196 rx_atomic_inc(&rx_stats.ackPacketsRead);
4197 ap = (struct rx_ackPacket *)rx_DataOf(np);
4198 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4200 return np; /* truncated ack packet */
4202 /* depends on ack packet struct */
4203 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4204 first = ntohl(ap->firstPacket);
4205 prev = ntohl(ap->previousPacket);
4206 serial = ntohl(ap->serial);
4207 /* temporarily disabled -- needs to degrade over time
4208 * skew = ntohs(ap->maxSkew); */
4210 /* Ignore ack packets received out of order */
4211 if (first < call->tfirst ||
4212 (first == call->tfirst && prev < call->tprev)) {
4218 if (np->header.flags & RX_SLOW_START_OK) {
4219 call->flags |= RX_CALL_SLOW_START_OK;
4222 if (ap->reason == RX_ACK_PING_RESPONSE)
4223 rxi_UpdatePeerReach(conn, call);
4225 if (conn->lastPacketSizeSeq) {
4226 MUTEX_ENTER(&conn->conn_data_lock);
4227 conn_data_locked = 1;
4228 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4229 pktsize = conn->lastPacketSize;
4230 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4233 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4234 if (!conn_data_locked) {
4235 MUTEX_ENTER(&conn->conn_data_lock);
4236 conn_data_locked = 1;
4238 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4239 /* process mtu ping ack */
4240 pktsize = conn->lastPingSize;
4241 conn->lastPingSizeSer = conn->lastPingSize = 0;
4245 if (conn_data_locked) {
4246 MUTEX_EXIT(&conn->conn_data_lock);
4247 conn_data_locked = 0;
4251 if (rxdebug_active) {
4255 len = _snprintf(msg, sizeof(msg),
4256 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
4257 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4258 ntohl(ap->serial), ntohl(ap->previousPacket),
4259 (unsigned int)np->header.seq, (unsigned int)skew,
4260 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
4264 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4265 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4269 OutputDebugString(msg);
4271 #else /* AFS_NT40_ENV */
4274 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
4275 ap->reason, ntohl(ap->previousPacket),
4276 (unsigned int)np->header.seq, (unsigned int)serial,
4277 (unsigned int)skew, ntohl(ap->firstPacket));
4280 for (offset = 0; offset < nAcks; offset++)
4281 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4286 #endif /* AFS_NT40_ENV */
4289 MUTEX_ENTER(&peer->peer_lock);
4292 * Start somewhere. Can't assume we can send what we can receive,
4293 * but we are clearly receiving.
4295 if (!peer->maxPacketSize)
4296 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4298 if (pktsize > peer->maxPacketSize) {
4299 peer->maxPacketSize = pktsize;
4300 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4301 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4302 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4303 rxi_ScheduleGrowMTUEvent(call, 1);
4308 /* Update the outgoing packet skew value to the latest value of
4309 * the peer's incoming packet skew value. The ack packet, of
4310 * course, could arrive out of order, but that won't affect things
4312 peer->outPacketSkew = skew;
4315 clock_GetTime(&now);
4317 /* The transmit queue splits into 4 sections.
4319 * The first section is packets which have now been acknowledged
4320 * by a window size change in the ack. These have reached the
4321 * application layer, and may be discarded. These are packets
4322 * with sequence numbers < ap->firstPacket.
4324 * The second section is packets which have sequence numbers in
4325 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4326 * contents of the packet's ack array determines whether these
4327 * packets are acknowledged or not.
4329 * The third section is packets which fall above the range
4330 * addressed in the ack packet. These have not yet been received
4333 * The four section is packets which have not yet been transmitted.
4334 * These packets will have a header.serial of 0.
4337 /* First section - implicitly acknowledged packets that can be
4341 tp = queue_First(&call->tq, rx_packet);
4342 while(!queue_IsEnd(&call->tq, tp) && tp->header.seq < first) {
4343 struct rx_packet *next;
4345 next = queue_Next(tp, rx_packet);
4346 call->tfirst = tp->header.seq + 1;
4348 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4350 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4353 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4354 /* XXX Hack. Because we have to release the global rx lock when sending
4355 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4356 * in rxi_Start sending packets out because packets may move to the
4357 * freePacketQueue as result of being here! So we drop these packets until
4358 * we're safely out of the traversing. Really ugly!
4359 * To make it even uglier, if we're using fine grain locking, we can
4360 * set the ack bits in the packets and have rxi_Start remove the packets
4361 * when it's done transmitting.
4363 if (call->flags & RX_CALL_TQ_BUSY) {
4364 #ifdef RX_ENABLE_LOCKS
4365 tp->flags |= RX_PKTFLAG_ACKED;
4366 call->flags |= RX_CALL_TQ_SOME_ACKED;
4367 #else /* RX_ENABLE_LOCKS */
4369 #endif /* RX_ENABLE_LOCKS */
4371 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4374 #ifdef RX_TRACK_PACKETS
4375 tp->flags &= ~RX_PKTFLAG_TQ;
4377 #ifdef RXDEBUG_PACKET
4379 #endif /* RXDEBUG_PACKET */
4380 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4385 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4387 /* Second section of the queue - packets for which we are receiving
4390 * Go through the explicit acks/nacks and record the results in
4391 * the waiting packets. These are packets that can't be released
4392 * yet, even with a positive acknowledge. This positive
4393 * acknowledge only means the packet has been received by the
4394 * peer, not that it will be retained long enough to be sent to
4395 * the peer's upper level. In addition, reset the transmit timers
4396 * of any missing packets (those packets that must be missing
4397 * because this packet was out of sequence) */
4399 call->nSoftAcked = 0;
4401 while (!queue_IsEnd(&call->tq, tp) && tp->header.seq < first + nAcks) {
4402 /* Set the acknowledge flag per packet based on the
4403 * information in the ack packet. An acknowlegded packet can
4404 * be downgraded when the server has discarded a packet it
4405 * soacked previously, or when an ack packet is received
4406 * out of sequence. */
4407 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4408 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4410 tp->flags |= RX_PKTFLAG_ACKED;
4411 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4418 } else /* RX_ACK_TYPE_NACK */ {
4419 tp->flags &= ~RX_PKTFLAG_ACKED;
4423 tp = queue_Next(tp, rx_packet);
4426 /* We don't need to take any action with the 3rd or 4th section in the
4427 * queue - they're not addressed by the contents of this ACK packet.
4430 /* If the window has been extended by this acknowledge packet,
4431 * then wakeup a sender waiting in alloc for window space, or try
4432 * sending packets now, if he's been sitting on packets due to
4433 * lack of window space */
4434 if (call->tnext < (call->tfirst + call->twind)) {
4435 #ifdef RX_ENABLE_LOCKS
4436 CV_SIGNAL(&call->cv_twind);
4438 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4439 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4440 osi_rxWakeup(&call->twind);
4443 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4444 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4448 /* if the ack packet has a receivelen field hanging off it,
4449 * update our state */
4450 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4453 /* If the ack packet has a "recommended" size that is less than
4454 * what I am using now, reduce my size to match */
4455 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4456 (int)sizeof(afs_int32), &tSize);
4457 tSize = (afs_uint32) ntohl(tSize);
4458 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4460 /* Get the maximum packet size to send to this peer */
4461 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4463 tSize = (afs_uint32) ntohl(tSize);
4464 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4465 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4467 /* sanity check - peer might have restarted with different params.
4468 * If peer says "send less", dammit, send less... Peer should never
4469 * be unable to accept packets of the size that prior AFS versions would
4470 * send without asking. */
4471 if (peer->maxMTU != tSize) {
4472 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4474 peer->maxMTU = tSize;
4475 peer->MTU = MIN(tSize, peer->MTU);
4476 call->MTU = MIN(call->MTU, tSize);
4479 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4482 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4483 (int)sizeof(afs_int32), &tSize);
4484 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4485 if (tSize < call->twind) { /* smaller than our send */
4486 call->twind = tSize; /* window, we must send less... */
4487 call->ssthresh = MIN(call->twind, call->ssthresh);
4488 call->conn->twind[call->channel] = call->twind;
4491 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4492 * network MTU confused with the loopback MTU. Calculate the
4493 * maximum MTU here for use in the slow start code below.
4495 /* Did peer restart with older RX version? */
4496 if (peer->maxDgramPackets > 1) {
4497 peer->maxDgramPackets = 1;
4499 } else if (np->length >=
4500 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4503 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4504 sizeof(afs_int32), &tSize);
4505 tSize = (afs_uint32) ntohl(tSize);
4507 * As of AFS 3.5 we set the send window to match the receive window.
4509 if (tSize < call->twind) {
4510 call->twind = tSize;
4511 call->conn->twind[call->channel] = call->twind;
4512 call->ssthresh = MIN(call->twind, call->ssthresh);
4513 } else if (tSize > call->twind) {
4514 call->twind = tSize;
4515 call->conn->twind[call->channel] = call->twind;
4519 * As of AFS 3.5, a jumbogram is more than one fixed size
4520 * packet transmitted in a single UDP datagram. If the remote
4521 * MTU is smaller than our local MTU then never send a datagram
4522 * larger than the natural MTU.
4525 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4526 (int)sizeof(afs_int32), &tSize);
4527 maxDgramPackets = (afs_uint32) ntohl(tSize);
4528 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4530 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4531 if (maxDgramPackets > 1) {
4532 peer->maxDgramPackets = maxDgramPackets;
4533 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4535 peer->maxDgramPackets = 1;
4536 call->MTU = peer->natMTU;
4538 } else if (peer->maxDgramPackets > 1) {
4539 /* Restarted with lower version of RX */
4540 peer->maxDgramPackets = 1;
4542 } else if (peer->maxDgramPackets > 1
4543 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4544 /* Restarted with lower version of RX */
4545 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4546 peer->natMTU = OLD_MAX_PACKET_SIZE;
4547 peer->MTU = OLD_MAX_PACKET_SIZE;
4548 peer->maxDgramPackets = 1;
4549 peer->nDgramPackets = 1;
4551 call->MTU = OLD_MAX_PACKET_SIZE;
4556 * Calculate how many datagrams were successfully received after
4557 * the first missing packet and adjust the negative ack counter
4562 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4563 if (call->nNacks < nNacked) {
4564 call->nNacks = nNacked;
4567 call->nAcks += newAckCount;
4571 /* If the packet contained new acknowledgements, rather than just
4572 * being a duplicate of one we have previously seen, then we can restart
4575 if (newAckCount > 0)
4576 rxi_rto_packet_acked(call, istack);
4578 if (call->flags & RX_CALL_FAST_RECOVER) {
4579 if (newAckCount == 0) {
4580 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4582 call->flags &= ~RX_CALL_FAST_RECOVER;
4583 call->cwind = call->nextCwind;
4584 call->nextCwind = 0;
4587 call->nCwindAcks = 0;
4588 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4589 /* Three negative acks in a row trigger congestion recovery */
4590 call->flags |= RX_CALL_FAST_RECOVER;
4591 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4593 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4594 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4595 call->nextCwind = call->ssthresh;
4598 peer->MTU = call->MTU;
4599 peer->cwind = call->nextCwind;
4600 peer->nDgramPackets = call->nDgramPackets;
4602 call->congestSeq = peer->congestSeq;
4604 /* Reset the resend times on the packets that were nacked
4605 * so we will retransmit as soon as the window permits
4608 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4610 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4611 tp->flags &= ~RX_PKTFLAG_SENT;
4613 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4618 /* If cwind is smaller than ssthresh, then increase
4619 * the window one packet for each ack we receive (exponential
4621 * If cwind is greater than or equal to ssthresh then increase
4622 * the congestion window by one packet for each cwind acks we
4623 * receive (linear growth). */
4624 if (call->cwind < call->ssthresh) {
4626 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4627 call->nCwindAcks = 0;
4629 call->nCwindAcks += newAckCount;
4630 if (call->nCwindAcks >= call->cwind) {
4631 call->nCwindAcks = 0;
4632 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4636 * If we have received several acknowledgements in a row then
4637 * it is time to increase the size of our datagrams
4639 if ((int)call->nAcks > rx_nDgramThreshold) {
4640 if (peer->maxDgramPackets > 1) {
4641 if (call->nDgramPackets < peer->maxDgramPackets) {
4642 call->nDgramPackets++;
4644 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4645 } else if (call->MTU < peer->maxMTU) {
4646 /* don't upgrade if we can't handle it */
4647 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4648 call->MTU = peer->ifMTU;
4650 call->MTU += peer->natMTU;
4651 call->MTU = MIN(call->MTU, peer->maxMTU);
4658 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4660 /* Servers need to hold the call until all response packets have
4661 * been acknowledged. Soft acks are good enough since clients
4662 * are not allowed to clear their receive queues. */
4663 if (call->state == RX_STATE_HOLD
4664 && call->tfirst + call->nSoftAcked >= call->tnext) {
4665 call->state = RX_STATE_DALLY;
4666 rxi_ClearTransmitQueue(call, 0);
4667 rxevent_Cancel(&call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4668 } else if (!queue_IsEmpty(&call->tq)) {
4669 rxi_Start(call, istack);
4674 /* Received a response to a challenge packet */
4676 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4677 struct rx_packet *np, int istack)
4681 /* Ignore the packet if we're the client */
4682 if (conn->type == RX_CLIENT_CONNECTION)
4685 /* If already authenticated, ignore the packet (it's probably a retry) */
4686 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4689 /* Otherwise, have the security object evaluate the response packet */
4690 error = RXS_CheckResponse(conn->securityObject, conn, np);
4692 /* If the response is invalid, reset the connection, sending
4693 * an abort to the peer */
4697 rxi_ConnectionError(conn, error);
4698 MUTEX_ENTER(&conn->conn_data_lock);
4699 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4700 MUTEX_EXIT(&conn->conn_data_lock);
4703 /* If the response is valid, any calls waiting to attach
4704 * servers can now do so */
4707 for (i = 0; i < RX_MAXCALLS; i++) {
4708 struct rx_call *call = conn->call[i];
4710 MUTEX_ENTER(&call->lock);
4711 if (call->state == RX_STATE_PRECALL)
4712 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4713 /* tnop can be null if newcallp is null */
4714 MUTEX_EXIT(&call->lock);
4718 /* Update the peer reachability information, just in case
4719 * some calls went into attach-wait while we were waiting
4720 * for authentication..
4722 rxi_UpdatePeerReach(conn, NULL);
4727 /* A client has received an authentication challenge: the security
4728 * object is asked to cough up a respectable response packet to send
4729 * back to the server. The server is responsible for retrying the
4730 * challenge if it fails to get a response. */
4733 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4734 struct rx_packet *np, int istack)
4738 /* Ignore the challenge if we're the server */
4739 if (conn->type == RX_SERVER_CONNECTION)
4742 /* Ignore the challenge if the connection is otherwise idle; someone's
4743 * trying to use us as an oracle. */
4744 if (!rxi_HasActiveCalls(conn))
4747 /* Send the security object the challenge packet. It is expected to fill
4748 * in the response. */
4749 error = RXS_GetResponse(conn->securityObject, conn, np);
4751 /* If the security object is unable to return a valid response, reset the
4752 * connection and send an abort to the peer. Otherwise send the response
4753 * packet to the peer connection. */
4755 rxi_ConnectionError(conn, error);
4756 MUTEX_ENTER(&conn->conn_data_lock);
4757 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4758 MUTEX_EXIT(&conn->conn_data_lock);
4760 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4761 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4767 /* Find an available server process to service the current request in
4768 * the given call structure. If one isn't available, queue up this
4769 * call so it eventually gets one */
4771 rxi_AttachServerProc(struct rx_call *call,
4772 osi_socket socket, int *tnop,
4773 struct rx_call **newcallp)
4775 struct rx_serverQueueEntry *sq;
4776 struct rx_service *service = call->conn->service;
4779 /* May already be attached */
4780 if (call->state == RX_STATE_ACTIVE)
4783 MUTEX_ENTER(&rx_serverPool_lock);
4785 haveQuota = QuotaOK(service);
4786 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4787 /* If there are no processes available to service this call,
4788 * put the call on the incoming call queue (unless it's
4789 * already on the queue).
4791 #ifdef RX_ENABLE_LOCKS
4793 ReturnToServerPool(service);
4794 #endif /* RX_ENABLE_LOCKS */
4796 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4797 call->flags |= RX_CALL_WAIT_PROC;
4798 rx_atomic_inc(&rx_nWaiting);
4799 rx_atomic_inc(&rx_nWaited);
4800 rxi_calltrace(RX_CALL_ARRIVAL, call);
4801 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4802 queue_Append(&rx_incomingCallQueue, call);
4805 sq = queue_Last(&rx_idleServerQueue, rx_serverQueueEntry);
4807 /* If hot threads are enabled, and both newcallp and sq->socketp
4808 * are non-null, then this thread will process the call, and the
4809 * idle server thread will start listening on this threads socket.
4812 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4815 *sq->socketp = socket;
4816 clock_GetTime(&call->startTime);
4817 MUTEX_ENTER(&rx_refcnt_mutex);
4818 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4819 MUTEX_EXIT(&rx_refcnt_mutex);
4823 if (call->flags & RX_CALL_WAIT_PROC) {
4824 /* Conservative: I don't think this should happen */
4825 call->flags &= ~RX_CALL_WAIT_PROC;
4826 if (queue_IsOnQueue(call)) {
4829 rx_atomic_dec(&rx_nWaiting);
4832 call->state = RX_STATE_ACTIVE;
4833 call->mode = RX_MODE_RECEIVING;
4834 #ifdef RX_KERNEL_TRACE
4836 int glockOwner = ISAFS_GLOCK();
4839 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4840 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4846 if (call->flags & RX_CALL_CLEARED) {
4847 /* send an ack now to start the packet flow up again */
4848 call->flags &= ~RX_CALL_CLEARED;
4849 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4851 #ifdef RX_ENABLE_LOCKS
4854 service->nRequestsRunning++;
4855 MUTEX_ENTER(&rx_quota_mutex);
4856 if (service->nRequestsRunning <= service->minProcs)
4859 MUTEX_EXIT(&rx_quota_mutex);
4863 MUTEX_EXIT(&rx_serverPool_lock);
4866 /* Delay the sending of an acknowledge event for a short while, while
4867 * a new call is being prepared (in the case of a client) or a reply
4868 * is being prepared (in the case of a server). Rather than sending
4869 * an ack packet, an ACKALL packet is sent. */
4871 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4873 #ifdef RX_ENABLE_LOCKS
4875 MUTEX_ENTER(&call->lock);
4876 rxevent_Put(call->delayedAckEvent);
4877 call->delayedAckEvent = NULL;
4878 MUTEX_ENTER(&rx_refcnt_mutex);
4879 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4880 MUTEX_EXIT(&rx_refcnt_mutex);
4882 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4883 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4884 call->flags |= RX_CALL_ACKALL_SENT;
4886 MUTEX_EXIT(&call->lock);
4887 #else /* RX_ENABLE_LOCKS */
4889 rxevent_Put(call->delayedAckEvent);
4890 call->delayedAckEvent = NULL;
4892 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4893 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4894 call->flags |= RX_CALL_ACKALL_SENT;
4895 #endif /* RX_ENABLE_LOCKS */
4899 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4902 struct rx_call *call = arg1;
4903 #ifdef RX_ENABLE_LOCKS
4905 MUTEX_ENTER(&call->lock);
4906 if (event == call->delayedAckEvent) {
4907 rxevent_Put(call->delayedAckEvent);
4908 call->delayedAckEvent = NULL;
4910 MUTEX_ENTER(&rx_refcnt_mutex);
4911 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4912 MUTEX_EXIT(&rx_refcnt_mutex);
4914 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4916 MUTEX_EXIT(&call->lock);
4917 #else /* RX_ENABLE_LOCKS */
4919 rxevent_Put(call->delayedAckEvent);
4920 call->delayedAckEvent = NULL;
4922 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4923 #endif /* RX_ENABLE_LOCKS */
4927 #ifdef RX_ENABLE_LOCKS
4928 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4929 * clearing them out.
4932 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4934 struct rx_packet *p, *tp;
4937 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4938 p->flags |= RX_PKTFLAG_ACKED;
4942 call->flags |= RX_CALL_TQ_CLEARME;
4943 call->flags |= RX_CALL_TQ_SOME_ACKED;
4946 rxi_rto_cancel(call);
4948 call->tfirst = call->tnext;
4949 call->nSoftAcked = 0;
4951 if (call->flags & RX_CALL_FAST_RECOVER) {
4952 call->flags &= ~RX_CALL_FAST_RECOVER;
4953 call->cwind = call->nextCwind;
4954 call->nextCwind = 0;
4957 CV_SIGNAL(&call->cv_twind);
4959 #endif /* RX_ENABLE_LOCKS */
4961 /* Clear out the transmit queue for the current call (all packets have
4962 * been received by peer) */
4964 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4966 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4967 struct rx_packet *p, *tp;
4969 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4971 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4972 p->flags |= RX_PKTFLAG_ACKED;
4976 call->flags |= RX_CALL_TQ_CLEARME;
4977 call->flags |= RX_CALL_TQ_SOME_ACKED;
4980 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4981 #ifdef RXDEBUG_PACKET
4983 #endif /* RXDEBUG_PACKET */
4984 rxi_FreePackets(0, &call->tq);
4985 rxi_WakeUpTransmitQueue(call);
4986 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4987 call->flags &= ~RX_CALL_TQ_CLEARME;
4989 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4991 rxi_rto_cancel(call);
4992 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4993 call->nSoftAcked = 0;
4995 if (call->flags & RX_CALL_FAST_RECOVER) {
4996 call->flags &= ~RX_CALL_FAST_RECOVER;
4997 call->cwind = call->nextCwind;
4999 #ifdef RX_ENABLE_LOCKS
5000 CV_SIGNAL(&call->cv_twind);
5002 osi_rxWakeup(&call->twind);
5007 rxi_ClearReceiveQueue(struct rx_call *call)
5009 if (queue_IsNotEmpty(&call->rq)) {
5012 count = rxi_FreePackets(0, &call->rq);
5013 rx_packetReclaims += count;
5014 #ifdef RXDEBUG_PACKET
5016 if ( call->rqc != 0 )
5017 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5019 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5021 if (call->state == RX_STATE_PRECALL) {
5022 call->flags |= RX_CALL_CLEARED;
5026 /* Send an abort packet for the specified call */
5028 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5029 int istack, int force)
5032 struct clock when, now;
5037 /* Clients should never delay abort messages */
5038 if (rx_IsClientConn(call->conn))
5041 if (call->abortCode != call->error) {
5042 call->abortCode = call->error;
5043 call->abortCount = 0;
5046 if (force || rxi_callAbortThreshhold == 0
5047 || call->abortCount < rxi_callAbortThreshhold) {
5048 if (call->delayedAbortEvent) {
5049 rxevent_Cancel(&call->delayedAbortEvent, call,
5050 RX_CALL_REFCOUNT_ABORT);
5052 error = htonl(call->error);
5055 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5056 (char *)&error, sizeof(error), istack);
5057 } else if (!call->delayedAbortEvent) {
5058 clock_GetTime(&now);
5060 clock_Addmsec(&when, rxi_callAbortDelay);
5061 MUTEX_ENTER(&rx_refcnt_mutex);
5062 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5063 MUTEX_EXIT(&rx_refcnt_mutex);
5064 call->delayedAbortEvent =
5065 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5070 /* Send an abort packet for the specified connection. Packet is an
5071 * optional pointer to a packet that can be used to send the abort.
5072 * Once the number of abort messages reaches the threshhold, an
5073 * event is scheduled to send the abort. Setting the force flag
5074 * overrides sending delayed abort messages.
5076 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5077 * to send the abort packet.
5080 rxi_SendConnectionAbort(struct rx_connection *conn,
5081 struct rx_packet *packet, int istack, int force)
5084 struct clock when, now;
5089 /* Clients should never delay abort messages */
5090 if (rx_IsClientConn(conn))
5093 if (force || rxi_connAbortThreshhold == 0
5094 || conn->abortCount < rxi_connAbortThreshhold) {
5096 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
5097 error = htonl(conn->error);
5099 MUTEX_EXIT(&conn->conn_data_lock);
5101 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5102 RX_PACKET_TYPE_ABORT, (char *)&error,
5103 sizeof(error), istack);
5104 MUTEX_ENTER(&conn->conn_data_lock);
5105 } else if (!conn->delayedAbortEvent) {
5106 clock_GetTime(&now);
5108 clock_Addmsec(&when, rxi_connAbortDelay);
5109 conn->delayedAbortEvent =
5110 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
5115 /* Associate an error all of the calls owned by a connection. Called
5116 * with error non-zero. This is only for really fatal things, like
5117 * bad authentication responses. The connection itself is set in
5118 * error at this point, so that future packets received will be
5121 rxi_ConnectionError(struct rx_connection *conn,
5127 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5129 MUTEX_ENTER(&conn->conn_data_lock);
5130 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
5131 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
5132 if (conn->checkReachEvent) {
5133 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
5134 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5135 putConnection(conn);
5137 MUTEX_EXIT(&conn->conn_data_lock);
5138 for (i = 0; i < RX_MAXCALLS; i++) {
5139 struct rx_call *call = conn->call[i];
5141 MUTEX_ENTER(&call->lock);
5142 rxi_CallError(call, error);
5143 MUTEX_EXIT(&call->lock);
5146 conn->error = error;
5147 if (rx_stats_active)
5148 rx_atomic_inc(&rx_stats.fatalErrors);
5153 * Interrupt an in-progress call with the specified error and wakeup waiters.
5155 * @param[in] call The call to interrupt
5156 * @param[in] error The error code to send to the peer
5159 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5161 MUTEX_ENTER(&call->lock);
5162 rxi_CallError(call, error);
5163 rxi_SendCallAbort(call, NULL, 0, 1);
5164 MUTEX_EXIT(&call->lock);
5168 rxi_CallError(struct rx_call *call, afs_int32 error)
5171 osirx_AssertMine(&call->lock, "rxi_CallError");
5173 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5175 error = call->error;
5177 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5178 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5179 rxi_ResetCall(call, 0);
5182 rxi_ResetCall(call, 0);
5184 call->error = error;
5187 /* Reset various fields in a call structure, and wakeup waiting
5188 * processes. Some fields aren't changed: state & mode are not
5189 * touched (these must be set by the caller), and bufptr, nLeft, and
5190 * nFree are not reset, since these fields are manipulated by
5191 * unprotected macros, and may only be reset by non-interrupting code.
5195 rxi_ResetCall(struct rx_call *call, int newcall)
5198 struct rx_peer *peer;
5199 struct rx_packet *packet;
5201 osirx_AssertMine(&call->lock, "rxi_ResetCall");
5203 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5205 /* Notify anyone who is waiting for asynchronous packet arrival */
5206 if (call->arrivalProc) {
5207 (*call->arrivalProc) (call, call->arrivalProcHandle,
5208 call->arrivalProcArg);
5209 call->arrivalProc = (void (*)())0;
5213 rxevent_Cancel(&call->growMTUEvent, call, RX_CALL_REFCOUNT_ALIVE);
5215 if (call->delayedAbortEvent) {
5216 rxevent_Cancel(&call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
5217 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5219 rxi_SendCallAbort(call, packet, 0, 1);
5220 rxi_FreePacket(packet);
5225 * Update the peer with the congestion information in this call
5226 * so other calls on this connection can pick up where this call
5227 * left off. If the congestion sequence numbers don't match then
5228 * another call experienced a retransmission.
5230 peer = call->conn->peer;
5231 MUTEX_ENTER(&peer->peer_lock);
5233 if (call->congestSeq == peer->congestSeq) {
5234 peer->cwind = MAX(peer->cwind, call->cwind);
5235 peer->MTU = MAX(peer->MTU, call->MTU);
5236 peer->nDgramPackets =
5237 MAX(peer->nDgramPackets, call->nDgramPackets);
5240 call->abortCode = 0;
5241 call->abortCount = 0;
5243 if (peer->maxDgramPackets > 1) {
5244 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5246 call->MTU = peer->MTU;
5248 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5249 call->ssthresh = rx_maxSendWindow;
5250 call->nDgramPackets = peer->nDgramPackets;
5251 call->congestSeq = peer->congestSeq;
5252 call->rtt = peer->rtt;
5253 call->rtt_dev = peer->rtt_dev;
5254 clock_Zero(&call->rto);
5255 clock_Addmsec(&call->rto,
5256 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5257 MUTEX_EXIT(&peer->peer_lock);
5259 flags = call->flags;
5260 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5261 rxi_WaitforTQBusy(call);
5262 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5264 rxi_ClearTransmitQueue(call, 1);
5265 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5266 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5270 if ((flags & RX_CALL_PEER_BUSY)) {
5271 /* The call channel is still busy; resetting the call doesn't change
5273 call->flags |= RX_CALL_PEER_BUSY;
5276 rxi_ClearReceiveQueue(call);
5277 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5281 call->twind = call->conn->twind[call->channel];
5282 call->rwind = call->conn->rwind[call->channel];
5283 call->nSoftAcked = 0;
5284 call->nextCwind = 0;
5287 call->nCwindAcks = 0;
5288 call->nSoftAcks = 0;
5289 call->nHardAcks = 0;
5291 call->tfirst = call->rnext = call->tnext = 1;
5294 call->lastAcked = 0;
5295 call->localStatus = call->remoteStatus = 0;
5297 if (flags & RX_CALL_READER_WAIT) {
5298 #ifdef RX_ENABLE_LOCKS
5299 CV_BROADCAST(&call->cv_rq);
5301 osi_rxWakeup(&call->rq);
5304 if (flags & RX_CALL_WAIT_PACKETS) {
5305 MUTEX_ENTER(&rx_freePktQ_lock);
5306 rxi_PacketsUnWait(); /* XXX */
5307 MUTEX_EXIT(&rx_freePktQ_lock);
5309 #ifdef RX_ENABLE_LOCKS
5310 CV_SIGNAL(&call->cv_twind);
5312 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5313 osi_rxWakeup(&call->twind);
5316 #ifdef RX_ENABLE_LOCKS
5317 /* The following ensures that we don't mess with any queue while some
5318 * other thread might also be doing so. The call_queue_lock field is
5319 * is only modified under the call lock. If the call is in the process
5320 * of being removed from a queue, the call is not locked until the
5321 * the queue lock is dropped and only then is the call_queue_lock field
5322 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5323 * Note that any other routine which removes a call from a queue has to
5324 * obtain the queue lock before examing the queue and removing the call.
5326 if (call->call_queue_lock) {
5327 MUTEX_ENTER(call->call_queue_lock);
5328 if (queue_IsOnQueue(call)) {
5330 if (flags & RX_CALL_WAIT_PROC) {
5331 rx_atomic_dec(&rx_nWaiting);
5334 MUTEX_EXIT(call->call_queue_lock);
5335 CLEAR_CALL_QUEUE_LOCK(call);
5337 #else /* RX_ENABLE_LOCKS */
5338 if (queue_IsOnQueue(call)) {
5340 if (flags & RX_CALL_WAIT_PROC)
5341 rx_atomic_dec(&rx_nWaiting);
5343 #endif /* RX_ENABLE_LOCKS */
5345 rxi_KeepAliveOff(call);
5346 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5349 /* Send an acknowledge for the indicated packet (seq,serial) of the
5350 * indicated call, for the indicated reason (reason). This
5351 * acknowledge will specifically acknowledge receiving the packet, and
5352 * will also specify which other packets for this call have been
5353 * received. This routine returns the packet that was used to the
5354 * caller. The caller is responsible for freeing it or re-using it.
5355 * This acknowledgement also returns the highest sequence number
5356 * actually read out by the higher level to the sender; the sender
5357 * promises to keep around packets that have not been read by the
5358 * higher level yet (unless, of course, the sender decides to abort
5359 * the call altogether). Any of p, seq, serial, pflags, or reason may
5360 * be set to zero without ill effect. That is, if they are zero, they
5361 * will not convey any information.
5362 * NOW there is a trailer field, after the ack where it will safely be
5363 * ignored by mundanes, which indicates the maximum size packet this
5364 * host can swallow. */
5366 struct rx_packet *optionalPacket; use to send ack (or null)
5367 int seq; Sequence number of the packet we are acking
5368 int serial; Serial number of the packet
5369 int pflags; Flags field from packet header
5370 int reason; Reason an acknowledge was prompted
5374 rxi_SendAck(struct rx_call *call,
5375 struct rx_packet *optionalPacket, int serial, int reason,
5378 struct rx_ackPacket *ap;
5379 struct rx_packet *rqp;
5380 struct rx_packet *nxp; /* For queue_Scan */
5381 struct rx_packet *p;
5384 afs_uint32 padbytes = 0;
5385 #ifdef RX_ENABLE_TSFPQ
5386 struct rx_ts_info_t * rx_ts_info;
5390 * Open the receive window once a thread starts reading packets
5392 if (call->rnext > 1) {
5393 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5396 /* Don't attempt to grow MTU if this is a critical ping */
5397 if (reason == RX_ACK_MTU) {
5398 /* keep track of per-call attempts, if we're over max, do in small
5399 * otherwise in larger? set a size to increment by, decrease
5402 if (call->conn->peer->maxPacketSize &&
5403 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5405 padbytes = call->conn->peer->maxPacketSize+16;
5407 padbytes = call->conn->peer->maxMTU + 128;
5409 /* do always try a minimum size ping */
5410 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5412 /* subtract the ack payload */
5413 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5414 reason = RX_ACK_PING;
5417 call->nHardAcks = 0;
5418 call->nSoftAcks = 0;
5419 if (call->rnext > call->lastAcked)
5420 call->lastAcked = call->rnext;
5424 rx_computelen(p, p->length); /* reset length, you never know */
5425 } /* where that's been... */
5426 #ifdef RX_ENABLE_TSFPQ
5428 RX_TS_INFO_GET(rx_ts_info);
5429 if ((p = rx_ts_info->local_special_packet)) {
5430 rx_computelen(p, p->length);
5431 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5432 rx_ts_info->local_special_packet = p;
5433 } else { /* We won't send the ack, but don't panic. */
5434 return optionalPacket;
5438 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5439 /* We won't send the ack, but don't panic. */
5440 return optionalPacket;
5445 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5448 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5449 #ifndef RX_ENABLE_TSFPQ
5450 if (!optionalPacket)
5453 return optionalPacket;
5455 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5456 if (rx_Contiguous(p) < templ) {
5457 #ifndef RX_ENABLE_TSFPQ
5458 if (!optionalPacket)
5461 return optionalPacket;
5466 /* MTUXXX failing to send an ack is very serious. We should */
5467 /* try as hard as possible to send even a partial ack; it's */
5468 /* better than nothing. */
5469 ap = (struct rx_ackPacket *)rx_DataOf(p);
5470 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5471 ap->reason = reason;
5473 /* The skew computation used to be bogus, I think it's better now. */
5474 /* We should start paying attention to skew. XXX */
5475 ap->serial = htonl(serial);
5476 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5479 * First packet not yet forwarded to reader. When ACKALL has been
5480 * sent the peer has been told that all received packets will be
5481 * delivered to the reader. The value 'rnext' is used internally
5482 * to refer to the next packet in the receive queue that must be
5483 * delivered to the reader. From the perspective of the peer it
5484 * already has so report the last sequence number plus one if there
5485 * are packets in the receive queue awaiting processing.
5487 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5488 !queue_IsEmpty(&call->rq)) {
5489 ap->firstPacket = htonl(queue_Last(&call->rq, rx_packet)->header.seq + 1);
5491 ap->firstPacket = htonl(call->rnext);
5493 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5495 /* No fear of running out of ack packet here because there can only be at most
5496 * one window full of unacknowledged packets. The window size must be constrained
5497 * to be less than the maximum ack size, of course. Also, an ack should always
5498 * fit into a single packet -- it should not ever be fragmented. */
5499 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5500 if (!rqp || !call->rq.next
5501 || (rqp->header.seq > (call->rnext + call->rwind))) {
5502 #ifndef RX_ENABLE_TSFPQ
5503 if (!optionalPacket)
5506 rxi_CallError(call, RX_CALL_DEAD);
5507 return optionalPacket;
5510 while (rqp->header.seq > call->rnext + offset)
5511 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5512 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5514 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5515 #ifndef RX_ENABLE_TSFPQ
5516 if (!optionalPacket)
5519 rxi_CallError(call, RX_CALL_DEAD);
5520 return optionalPacket;
5526 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5528 /* these are new for AFS 3.3 */
5529 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5530 templ = htonl(templ);
5531 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5532 templ = htonl(call->conn->peer->ifMTU);
5533 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5534 sizeof(afs_int32), &templ);
5536 /* new for AFS 3.4 */
5537 templ = htonl(call->rwind);
5538 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5539 sizeof(afs_int32), &templ);
5541 /* new for AFS 3.5 */
5542 templ = htonl(call->conn->peer->ifDgramPackets);
5543 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5544 sizeof(afs_int32), &templ);
5546 p->header.serviceId = call->conn->serviceId;
5547 p->header.cid = (call->conn->cid | call->channel);
5548 p->header.callNumber = *call->callNumber;
5550 p->header.securityIndex = call->conn->securityIndex;
5551 p->header.epoch = call->conn->epoch;
5552 p->header.type = RX_PACKET_TYPE_ACK;
5553 p->header.flags = RX_SLOW_START_OK;
5554 if (reason == RX_ACK_PING) {
5555 p->header.flags |= RX_REQUEST_ACK;
5557 p->length = padbytes +
5558 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5561 /* not fast but we can potentially use this if truncated
5562 * fragments are delivered to figure out the mtu.
5564 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5565 sizeof(afs_int32), sizeof(afs_int32),
5569 if (call->conn->type == RX_CLIENT_CONNECTION)
5570 p->header.flags |= RX_CLIENT_INITIATED;
5574 if (rxdebug_active) {
5578 len = _snprintf(msg, sizeof(msg),
5579 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5580 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5581 ntohl(ap->serial), ntohl(ap->previousPacket),
5582 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5583 ap->nAcks, ntohs(ap->bufferSpace) );
5587 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5588 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5592 OutputDebugString(msg);
5594 #else /* AFS_NT40_ENV */
5596 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5597 ap->reason, ntohl(ap->previousPacket),
5598 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5600 for (offset = 0; offset < ap->nAcks; offset++)
5601 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5606 #endif /* AFS_NT40_ENV */
5609 int i, nbytes = p->length;
5611 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5612 if (nbytes <= p->wirevec[i].iov_len) {
5615 savelen = p->wirevec[i].iov_len;
5617 p->wirevec[i].iov_len = nbytes;
5619 rxi_Send(call, p, istack);
5620 p->wirevec[i].iov_len = savelen;
5624 nbytes -= p->wirevec[i].iov_len;
5627 if (rx_stats_active)
5628 rx_atomic_inc(&rx_stats.ackPacketsSent);
5629 #ifndef RX_ENABLE_TSFPQ
5630 if (!optionalPacket)
5633 return optionalPacket; /* Return packet for re-use by caller */
5637 struct rx_packet **list;
5642 /* Send all of the packets in the list in single datagram */
5644 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5645 int istack, int moreFlag)
5651 struct rx_connection *conn = call->conn;
5652 struct rx_peer *peer = conn->peer;
5654 MUTEX_ENTER(&peer->peer_lock);
5655 peer->nSent += xmit->len;
5656 if (xmit->resending)
5657 peer->reSends += xmit->len;
5658 MUTEX_EXIT(&peer->peer_lock);
5660 if (rx_stats_active) {
5661 if (xmit->resending)
5662 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5664 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5667 clock_GetTime(&now);
5669 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5673 /* Set the packet flags and schedule the resend events */
5674 /* Only request an ack for the last packet in the list */
5675 for (i = 0; i < xmit->len; i++) {
5676 struct rx_packet *packet = xmit->list[i];
5678 /* Record the time sent */
5679 packet->timeSent = now;
5680 packet->flags |= RX_PKTFLAG_SENT;
5682 /* Ask for an ack on retransmitted packets, on every other packet
5683 * if the peer doesn't support slow start. Ask for an ack on every
5684 * packet until the congestion window reaches the ack rate. */
5685 if (packet->header.serial) {
5688 packet->firstSent = now;
5689 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5690 || (!(call->flags & RX_CALL_SLOW_START_OK)
5691 && (packet->header.seq & 1)))) {
5696 /* Tag this packet as not being the last in this group,
5697 * for the receiver's benefit */
5698 if (i < xmit->len - 1 || moreFlag) {
5699 packet->header.flags |= RX_MORE_PACKETS;
5704 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5707 /* Since we're about to send a data packet to the peer, it's
5708 * safe to nuke any scheduled end-of-packets ack */
5709 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5711 MUTEX_EXIT(&call->lock);
5712 MUTEX_ENTER(&rx_refcnt_mutex);
5713 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5714 MUTEX_EXIT(&rx_refcnt_mutex);
5715 if (xmit->len > 1) {
5716 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5718 rxi_SendPacket(call, conn, xmit->list[0], istack);
5720 MUTEX_ENTER(&call->lock);
5721 MUTEX_ENTER(&rx_refcnt_mutex);
5722 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5723 MUTEX_EXIT(&rx_refcnt_mutex);
5725 /* Tell the RTO calculation engine that we have sent a packet, and
5726 * if it was the last one */
5727 rxi_rto_packet_sent(call, lastPacket, istack);
5729 /* Update last send time for this call (for keep-alive
5730 * processing), and for the connection (so that we can discover
5731 * idle connections) */
5732 conn->lastSendTime = call->lastSendTime = clock_Sec();
5733 /* Let a set of retransmits trigger an idle timeout */
5734 if (!xmit->resending)
5735 call->lastSendData = call->lastSendTime;
5738 /* When sending packets we need to follow these rules:
5739 * 1. Never send more than maxDgramPackets in a jumbogram.
5740 * 2. Never send a packet with more than two iovecs in a jumbogram.
5741 * 3. Never send a retransmitted packet in a jumbogram.
5742 * 4. Never send more than cwind/4 packets in a jumbogram
5743 * We always keep the last list we should have sent so we
5744 * can set the RX_MORE_PACKETS flags correctly.
5748 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5753 struct xmitlist working;
5754 struct xmitlist last;
5756 struct rx_peer *peer = call->conn->peer;
5757 int morePackets = 0;
5759 memset(&last, 0, sizeof(struct xmitlist));
5760 working.list = &list[0];
5762 working.resending = 0;
5764 recovery = call->flags & RX_CALL_FAST_RECOVER;
5766 for (i = 0; i < len; i++) {
5767 /* Does the current packet force us to flush the current list? */
5769 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5770 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5772 /* This sends the 'last' list and then rolls the current working
5773 * set into the 'last' one, and resets the working set */
5776 rxi_SendList(call, &last, istack, 1);
5777 /* If the call enters an error state stop sending, or if
5778 * we entered congestion recovery mode, stop sending */
5780 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5785 working.resending = 0;
5786 working.list = &list[i];
5788 /* Add the current packet to the list if it hasn't been acked.
5789 * Otherwise adjust the list pointer to skip the current packet. */
5790 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5793 if (list[i]->header.serial)
5794 working.resending = 1;
5796 /* Do we need to flush the list? */
5797 if (working.len >= (int)peer->maxDgramPackets
5798 || working.len >= (int)call->nDgramPackets
5799 || working.len >= (int)call->cwind
5800 || list[i]->header.serial
5801 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5803 rxi_SendList(call, &last, istack, 1);
5804 /* If the call enters an error state stop sending, or if
5805 * we entered congestion recovery mode, stop sending */
5807 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5812 working.resending = 0;
5813 working.list = &list[i + 1];
5816 if (working.len != 0) {
5817 osi_Panic("rxi_SendList error");
5819 working.list = &list[i + 1];
5823 /* Send the whole list when the call is in receive mode, when
5824 * the call is in eof mode, when we are in fast recovery mode,
5825 * and when we have the last packet */
5826 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5827 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5828 || (call->flags & RX_CALL_FAST_RECOVER)) {
5829 /* Check for the case where the current list contains
5830 * an acked packet. Since we always send retransmissions
5831 * in a separate packet, we only need to check the first
5832 * packet in the list */
5833 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5837 rxi_SendList(call, &last, istack, morePackets);
5838 /* If the call enters an error state stop sending, or if
5839 * we entered congestion recovery mode, stop sending */
5841 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5845 rxi_SendList(call, &working, istack, 0);
5847 } else if (last.len > 0) {
5848 rxi_SendList(call, &last, istack, 0);
5849 /* Packets which are in 'working' are not sent by this call */
5854 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
5856 struct rx_call *call = arg0;
5857 struct rx_peer *peer;
5858 struct rx_packet *p, *nxp;
5859 struct clock maxTimeout = { 60, 0 };
5861 MUTEX_ENTER(&call->lock);
5863 peer = call->conn->peer;
5865 /* Make sure that the event pointer is removed from the call
5866 * structure, since there is no longer a per-call retransmission
5868 if (event == call->resendEvent) {
5869 MUTEX_ENTER(&rx_refcnt_mutex);
5870 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5871 MUTEX_EXIT(&rx_refcnt_mutex);
5872 rxevent_Put(call->resendEvent);
5873 call->resendEvent = NULL;
5876 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
5877 rxi_CheckBusy(call);
5880 if (queue_IsEmpty(&call->tq)) {
5881 /* Nothing to do. This means that we've been raced, and that an
5882 * ACK has come in between when we were triggered, and when we
5883 * actually got to run. */
5887 /* We're in loss recovery */
5888 call->flags |= RX_CALL_FAST_RECOVER;
5890 /* Mark all of the pending packets in the queue as being lost */
5891 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5892 if (!(p->flags & RX_PKTFLAG_ACKED))
5893 p->flags &= ~RX_PKTFLAG_SENT;
5896 /* We're resending, so we double the timeout of the call. This will be
5897 * dropped back down by the first successful ACK that we receive.
5899 * We apply a maximum value here of 60 seconds
5901 clock_Add(&call->rto, &call->rto);
5902 if (clock_Gt(&call->rto, &maxTimeout))
5903 call->rto = maxTimeout;
5905 /* Packet loss is most likely due to congestion, so drop our window size
5906 * and start again from the beginning */
5907 if (peer->maxDgramPackets >1) {
5908 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5909 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5911 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5912 call->nDgramPackets = 1;
5914 call->nextCwind = 1;
5917 MUTEX_ENTER(&peer->peer_lock);
5918 peer->MTU = call->MTU;
5919 peer->cwind = call->cwind;
5920 peer->nDgramPackets = 1;
5922 call->congestSeq = peer->congestSeq;
5923 MUTEX_EXIT(&peer->peer_lock);
5925 rxi_Start(call, istack);
5928 MUTEX_EXIT(&call->lock);
5931 /* This routine is called when new packets are readied for
5932 * transmission and when retransmission may be necessary, or when the
5933 * transmission window or burst count are favourable. This should be
5934 * better optimized for new packets, the usual case, now that we've
5935 * got rid of queues of send packets. XXXXXXXXXXX */
5937 rxi_Start(struct rx_call *call, int istack)
5940 struct rx_packet *p;
5941 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5946 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5947 if (rx_stats_active)
5948 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5953 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5955 /* Send (or resend) any packets that need it, subject to
5956 * window restrictions and congestion burst control
5957 * restrictions. Ask for an ack on the last packet sent in
5958 * this burst. For now, we're relying upon the window being
5959 * considerably bigger than the largest number of packets that
5960 * are typically sent at once by one initial call to
5961 * rxi_Start. This is probably bogus (perhaps we should ask
5962 * for an ack when we're half way through the current
5963 * window?). Also, for non file transfer applications, this
5964 * may end up asking for an ack for every packet. Bogus. XXXX
5967 * But check whether we're here recursively, and let the other guy
5970 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5971 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5972 call->flags |= RX_CALL_TQ_BUSY;
5974 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5976 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5977 call->flags &= ~RX_CALL_NEED_START;
5978 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5980 maxXmitPackets = MIN(call->twind, call->cwind);
5981 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5982 #ifdef RX_TRACK_PACKETS
5983 if ((p->flags & RX_PKTFLAG_FREE)
5984 || (!queue_IsEnd(&call->tq, nxp)
5985 && (nxp->flags & RX_PKTFLAG_FREE))
5986 || (p == (struct rx_packet *)&rx_freePacketQueue)
5987 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5988 osi_Panic("rxi_Start: xmit queue clobbered");
5991 if (p->flags & RX_PKTFLAG_ACKED) {
5992 /* Since we may block, don't trust this */
5993 if (rx_stats_active)
5994 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
5995 continue; /* Ignore this packet if it has been acknowledged */
5998 /* Turn off all flags except these ones, which are the same
5999 * on each transmission */
6000 p->header.flags &= RX_PRESET_FLAGS;
6002 if (p->header.seq >=
6003 call->tfirst + MIN((int)call->twind,
6004 (int)(call->nSoftAcked +
6006 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6007 /* Note: if we're waiting for more window space, we can
6008 * still send retransmits; hence we don't return here, but
6009 * break out to schedule a retransmit event */
6010 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6011 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6016 /* Transmit the packet if it needs to be sent. */
6017 if (!(p->flags & RX_PKTFLAG_SENT)) {
6018 if (nXmitPackets == maxXmitPackets) {
6019 rxi_SendXmitList(call, call->xmitList,
6020 nXmitPackets, istack);
6023 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6024 *(call->callNumber), p));
6025 call->xmitList[nXmitPackets++] = p;
6029 /* xmitList now hold pointers to all of the packets that are
6030 * ready to send. Now we loop to send the packets */
6031 if (nXmitPackets > 0) {
6032 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6036 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6038 /* We went into the error state while sending packets. Now is
6039 * the time to reset the call. This will also inform the using
6040 * process that the call is in an error state.
6042 if (rx_stats_active)
6043 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6044 call->flags &= ~RX_CALL_TQ_BUSY;
6045 rxi_WakeUpTransmitQueue(call);
6046 rxi_CallError(call, call->error);
6049 #ifdef RX_ENABLE_LOCKS
6050 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6052 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6053 /* Some packets have received acks. If they all have, we can clear
6054 * the transmit queue.
6057 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
6058 if (p->header.seq < call->tfirst
6059 && (p->flags & RX_PKTFLAG_ACKED)) {
6061 #ifdef RX_TRACK_PACKETS
6062 p->flags &= ~RX_PKTFLAG_TQ;
6064 #ifdef RXDEBUG_PACKET
6072 call->flags |= RX_CALL_TQ_CLEARME;
6074 #endif /* RX_ENABLE_LOCKS */
6075 if (call->flags & RX_CALL_TQ_CLEARME)
6076 rxi_ClearTransmitQueue(call, 1);
6077 } while (call->flags & RX_CALL_NEED_START);
6079 * TQ references no longer protected by this flag; they must remain
6080 * protected by the global lock.
6082 call->flags &= ~RX_CALL_TQ_BUSY;
6083 rxi_WakeUpTransmitQueue(call);
6085 call->flags |= RX_CALL_NEED_START;
6087 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6089 rxi_rto_cancel(call);
6093 /* Also adjusts the keep alive parameters for the call, to reflect
6094 * that we have just sent a packet (so keep alives aren't sent
6097 rxi_Send(struct rx_call *call, struct rx_packet *p,
6100 struct rx_connection *conn = call->conn;
6102 /* Stamp each packet with the user supplied status */
6103 p->header.userStatus = call->localStatus;
6105 /* Allow the security object controlling this call's security to
6106 * make any last-minute changes to the packet */
6107 RXS_SendPacket(conn->securityObject, call, p);
6109 /* Since we're about to send SOME sort of packet to the peer, it's
6110 * safe to nuke any scheduled end-of-packets ack */
6111 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
6113 /* Actually send the packet, filling in more connection-specific fields */
6114 MUTEX_EXIT(&call->lock);
6115 MUTEX_ENTER(&rx_refcnt_mutex);
6116 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6117 MUTEX_EXIT(&rx_refcnt_mutex);
6118 rxi_SendPacket(call, conn, p, istack);
6119 MUTEX_ENTER(&rx_refcnt_mutex);
6120 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6121 MUTEX_EXIT(&rx_refcnt_mutex);
6122 MUTEX_ENTER(&call->lock);
6124 /* Update last send time for this call (for keep-alive
6125 * processing), and for the connection (so that we can discover
6126 * idle connections) */
6127 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6128 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6129 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6131 conn->lastSendTime = call->lastSendTime = clock_Sec();
6132 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6133 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6134 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6135 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6136 RX_ACK_PING_RESPONSE)))
6137 call->lastSendData = call->lastSendTime;
6141 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6142 * that things are fine. Also called periodically to guarantee that nothing
6143 * falls through the cracks (e.g. (error + dally) connections have keepalive
6144 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6146 * haveCTLock Set if calling from rxi_ReapConnections
6148 #ifdef RX_ENABLE_LOCKS
6150 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6151 #else /* RX_ENABLE_LOCKS */
6153 rxi_CheckCall(struct rx_call *call)
6154 #endif /* RX_ENABLE_LOCKS */
6156 struct rx_connection *conn = call->conn;
6158 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6159 afs_uint32 fudgeFactor;
6163 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6164 if (call->flags & RX_CALL_TQ_BUSY) {
6165 /* Call is active and will be reset by rxi_Start if it's
6166 * in an error state.
6171 /* RTT + 8*MDEV, rounded up to the next second. */
6172 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6173 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6175 deadTime = conn->secondsUntilDead + fudgeFactor;
6177 /* These are computed to the second (+- 1 second). But that's
6178 * good enough for these values, which should be a significant
6179 * number of seconds. */
6180 if (now > (call->lastReceiveTime + deadTime)) {
6181 if (call->state == RX_STATE_ACTIVE) {
6183 #if defined(KERNEL) && defined(AFS_SUN5_ENV)
6185 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6186 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6187 ip_stack_t *ipst = ns->netstack_ip;
6189 ire = ire_cache_lookup(conn->peer->host
6190 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6192 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
6194 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6201 if (ire && ire->ire_max_frag > 0)
6202 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6204 #if defined(GLOBAL_NETSTACKID)
6208 #endif /* ADAPT_PMTU */
6209 cerror = RX_CALL_DEAD;
6212 #ifdef RX_ENABLE_LOCKS
6213 /* Cancel pending events */
6214 rxevent_Cancel(&call->delayedAckEvent, call,
6215 RX_CALL_REFCOUNT_DELAY);
6216 rxi_rto_cancel(call);
6217 rxevent_Cancel(&call->keepAliveEvent, call,
6218 RX_CALL_REFCOUNT_ALIVE);
6219 rxevent_Cancel(&call->growMTUEvent, call,
6220 RX_CALL_REFCOUNT_ALIVE);
6221 MUTEX_ENTER(&rx_refcnt_mutex);
6222 if (call->refCount == 0) {
6223 rxi_FreeCall(call, haveCTLock);
6224 MUTEX_EXIT(&rx_refcnt_mutex);
6227 MUTEX_EXIT(&rx_refcnt_mutex);
6229 #else /* RX_ENABLE_LOCKS */
6230 rxi_FreeCall(call, 0);
6232 #endif /* RX_ENABLE_LOCKS */
6234 /* Non-active calls are destroyed if they are not responding
6235 * to pings; active calls are simply flagged in error, so the
6236 * attached process can die reasonably gracefully. */
6239 if (conn->idleDeadTime) {
6240 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6243 /* see if we have a non-activity timeout */
6244 if (call->startWait && idleDeadTime
6245 && ((call->startWait + idleDeadTime) < now) &&
6246 (call->flags & RX_CALL_READER_WAIT)) {
6247 if (call->state == RX_STATE_ACTIVE) {
6248 cerror = RX_CALL_TIMEOUT;
6252 if (call->lastSendData && idleDeadTime && (conn->idleDeadErr != 0)
6253 && ((call->lastSendData + idleDeadTime) < now)) {
6254 if (call->state == RX_STATE_ACTIVE) {
6255 cerror = conn->idleDeadErr;
6260 if (conn->hardDeadTime) {
6261 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6264 /* see if we have a hard timeout */
6266 && (now > (hardDeadTime + call->startTime.sec))) {
6267 if (call->state == RX_STATE_ACTIVE)
6268 rxi_CallError(call, RX_CALL_TIMEOUT);
6273 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT
6274 && call->lastReceiveTime) {
6275 int oldMTU = conn->peer->ifMTU;
6277 /* if we thought we could send more, perhaps things got worse */
6278 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6279 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6280 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6281 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6283 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6285 /* minimum capped in SetPeerMtu */
6286 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6289 conn->lastPacketSize = 0;
6291 /* needed so ResetCall doesn't clobber us. */
6292 call->MTU = conn->peer->ifMTU;
6294 /* if we never succeeded, let the error pass out as-is */
6295 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6296 cerror = conn->msgsizeRetryErr;
6299 rxi_CallError(call, cerror);
6304 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6305 void *dummy, int dummy2)
6307 struct rx_connection *conn = arg1;
6308 struct rx_header theader;
6309 char tbuffer[1 + sizeof(struct rx_header)];
6310 struct sockaddr_in taddr;
6313 struct iovec tmpiov[2];
6316 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6319 tp = &tbuffer[sizeof(struct rx_header)];
6320 taddr.sin_family = AF_INET;
6321 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6322 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6323 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6324 taddr.sin_len = sizeof(struct sockaddr_in);
6326 memset(&theader, 0, sizeof(theader));
6327 theader.epoch = htonl(999);
6329 theader.callNumber = 0;
6332 theader.type = RX_PACKET_TYPE_VERSION;
6333 theader.flags = RX_LAST_PACKET;
6334 theader.serviceId = 0;
6336 memcpy(tbuffer, &theader, sizeof(theader));
6337 memcpy(tp, &a, sizeof(a));
6338 tmpiov[0].iov_base = tbuffer;
6339 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6341 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6343 MUTEX_ENTER(&conn->conn_data_lock);
6344 MUTEX_ENTER(&rx_refcnt_mutex);
6345 /* Only reschedule ourselves if the connection would not be destroyed */
6346 if (conn->refCount <= 1) {
6347 rxevent_Put(conn->natKeepAliveEvent);
6348 conn->natKeepAliveEvent = NULL;
6349 MUTEX_EXIT(&rx_refcnt_mutex);
6350 MUTEX_EXIT(&conn->conn_data_lock);
6351 rx_DestroyConnection(conn); /* drop the reference for this */
6353 conn->refCount--; /* drop the reference for this */
6354 MUTEX_EXIT(&rx_refcnt_mutex);
6355 rxevent_Put(conn->natKeepAliveEvent);
6356 conn->natKeepAliveEvent = NULL;
6357 rxi_ScheduleNatKeepAliveEvent(conn);
6358 MUTEX_EXIT(&conn->conn_data_lock);
6363 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6365 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6366 struct clock when, now;
6367 clock_GetTime(&now);
6369 when.sec += conn->secondsUntilNatPing;
6370 MUTEX_ENTER(&rx_refcnt_mutex);
6371 conn->refCount++; /* hold a reference for this */
6372 MUTEX_EXIT(&rx_refcnt_mutex);
6373 conn->natKeepAliveEvent =
6374 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6379 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6381 MUTEX_ENTER(&conn->conn_data_lock);
6382 conn->secondsUntilNatPing = seconds;
6384 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6385 rxi_ScheduleNatKeepAliveEvent(conn);
6387 conn->flags |= RX_CONN_NAT_PING;
6389 MUTEX_EXIT(&conn->conn_data_lock);
6393 rxi_NatKeepAliveOn(struct rx_connection *conn)
6395 MUTEX_ENTER(&conn->conn_data_lock);
6396 /* if it's already attached */
6397 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6398 rxi_ScheduleNatKeepAliveEvent(conn);
6400 conn->flags |= RX_CONN_NAT_PING;
6401 MUTEX_EXIT(&conn->conn_data_lock);
6404 /* When a call is in progress, this routine is called occasionally to
6405 * make sure that some traffic has arrived (or been sent to) the peer.
6406 * If nothing has arrived in a reasonable amount of time, the call is
6407 * declared dead; if nothing has been sent for a while, we send a
6408 * keep-alive packet (if we're actually trying to keep the call alive)
6411 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6414 struct rx_call *call = arg1;
6415 struct rx_connection *conn;
6418 MUTEX_ENTER(&rx_refcnt_mutex);
6419 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6420 MUTEX_EXIT(&rx_refcnt_mutex);
6421 MUTEX_ENTER(&call->lock);
6423 if (event == call->keepAliveEvent) {
6424 rxevent_Put(call->keepAliveEvent);
6425 call->keepAliveEvent = NULL;
6430 #ifdef RX_ENABLE_LOCKS
6431 if (rxi_CheckCall(call, 0)) {
6432 MUTEX_EXIT(&call->lock);
6435 #else /* RX_ENABLE_LOCKS */
6436 if (rxi_CheckCall(call))
6438 #endif /* RX_ENABLE_LOCKS */
6440 /* Don't try to keep alive dallying calls */
6441 if (call->state == RX_STATE_DALLY) {
6442 MUTEX_EXIT(&call->lock);
6447 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6448 /* Don't try to send keepalives if there is unacknowledged data */
6449 /* the rexmit code should be good enough, this little hack
6450 * doesn't quite work XXX */
6451 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6453 rxi_ScheduleKeepAliveEvent(call);
6454 MUTEX_EXIT(&call->lock);
6457 /* Does what's on the nameplate. */
6459 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6461 struct rx_call *call = arg1;
6462 struct rx_connection *conn;
6464 MUTEX_ENTER(&rx_refcnt_mutex);
6465 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6466 MUTEX_EXIT(&rx_refcnt_mutex);
6467 MUTEX_ENTER(&call->lock);
6469 if (event == call->growMTUEvent) {
6470 rxevent_Put(call->growMTUEvent);
6471 call->growMTUEvent = NULL;
6474 #ifdef RX_ENABLE_LOCKS
6475 if (rxi_CheckCall(call, 0)) {
6476 MUTEX_EXIT(&call->lock);
6479 #else /* RX_ENABLE_LOCKS */
6480 if (rxi_CheckCall(call))
6482 #endif /* RX_ENABLE_LOCKS */
6484 /* Don't bother with dallying calls */
6485 if (call->state == RX_STATE_DALLY) {
6486 MUTEX_EXIT(&call->lock);
6493 * keep being scheduled, just don't do anything if we're at peak,
6494 * or we're not set up to be properly handled (idle timeout required)
6496 if ((conn->peer->maxPacketSize != 0) &&
6497 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6498 (conn->idleDeadErr))
6499 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6500 rxi_ScheduleGrowMTUEvent(call, 0);
6501 MUTEX_EXIT(&call->lock);
6505 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6507 if (!call->keepAliveEvent) {
6508 struct clock when, now;
6509 clock_GetTime(&now);
6511 when.sec += call->conn->secondsUntilPing;
6512 MUTEX_ENTER(&rx_refcnt_mutex);
6513 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6514 MUTEX_EXIT(&rx_refcnt_mutex);
6515 call->keepAliveEvent =
6516 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6521 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6523 if (!call->growMTUEvent) {
6524 struct clock when, now;
6526 clock_GetTime(&now);
6529 if (call->conn->secondsUntilPing)
6530 secs = (6*call->conn->secondsUntilPing)-1;
6532 if (call->conn->secondsUntilDead)
6533 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6537 MUTEX_ENTER(&rx_refcnt_mutex);
6538 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6539 MUTEX_EXIT(&rx_refcnt_mutex);
6540 call->growMTUEvent =
6541 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6545 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6547 rxi_KeepAliveOn(struct rx_call *call)
6549 /* Pretend last packet received was received now--i.e. if another
6550 * packet isn't received within the keep alive time, then the call
6551 * will die; Initialize last send time to the current time--even
6552 * if a packet hasn't been sent yet. This will guarantee that a
6553 * keep-alive is sent within the ping time */
6554 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6555 rxi_ScheduleKeepAliveEvent(call);
6559 rxi_GrowMTUOn(struct rx_call *call)
6561 struct rx_connection *conn = call->conn;
6562 MUTEX_ENTER(&conn->conn_data_lock);
6563 conn->lastPingSizeSer = conn->lastPingSize = 0;
6564 MUTEX_EXIT(&conn->conn_data_lock);
6565 rxi_ScheduleGrowMTUEvent(call, 1);
6568 /* This routine is called to send connection abort messages
6569 * that have been delayed to throttle looping clients. */
6571 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6574 struct rx_connection *conn = arg1;
6577 struct rx_packet *packet;
6579 MUTEX_ENTER(&conn->conn_data_lock);
6580 rxevent_Put(conn->delayedAbortEvent);
6581 conn->delayedAbortEvent = NULL;
6582 error = htonl(conn->error);
6584 MUTEX_EXIT(&conn->conn_data_lock);
6585 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6588 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6589 RX_PACKET_TYPE_ABORT, (char *)&error,
6591 rxi_FreePacket(packet);
6595 /* This routine is called to send call abort messages
6596 * that have been delayed to throttle looping clients. */
6598 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6601 struct rx_call *call = arg1;
6604 struct rx_packet *packet;
6606 MUTEX_ENTER(&call->lock);
6607 rxevent_Put(call->delayedAbortEvent);
6608 call->delayedAbortEvent = NULL;
6609 error = htonl(call->error);
6611 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6614 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6615 (char *)&error, sizeof(error), 0);
6616 rxi_FreePacket(packet);
6618 MUTEX_EXIT(&call->lock);
6619 MUTEX_ENTER(&rx_refcnt_mutex);
6620 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6621 MUTEX_EXIT(&rx_refcnt_mutex);
6624 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6625 * seconds) to ask the client to authenticate itself. The routine
6626 * issues a challenge to the client, which is obtained from the
6627 * security object associated with the connection */
6629 rxi_ChallengeEvent(struct rxevent *event,
6630 void *arg0, void *arg1, int tries)
6632 struct rx_connection *conn = arg0;
6635 rxevent_Put(conn->challengeEvent);
6636 conn->challengeEvent = NULL;
6639 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6640 struct rx_packet *packet;
6641 struct clock when, now;
6644 /* We've failed to authenticate for too long.
6645 * Reset any calls waiting for authentication;
6646 * they are all in RX_STATE_PRECALL.
6650 MUTEX_ENTER(&conn->conn_call_lock);
6651 for (i = 0; i < RX_MAXCALLS; i++) {
6652 struct rx_call *call = conn->call[i];
6654 MUTEX_ENTER(&call->lock);
6655 if (call->state == RX_STATE_PRECALL) {
6656 rxi_CallError(call, RX_CALL_DEAD);
6657 rxi_SendCallAbort(call, NULL, 0, 0);
6659 MUTEX_EXIT(&call->lock);
6662 MUTEX_EXIT(&conn->conn_call_lock);
6666 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6668 /* If there's no packet available, do this later. */
6669 RXS_GetChallenge(conn->securityObject, conn, packet);
6670 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6671 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6672 rxi_FreePacket(packet);
6674 clock_GetTime(&now);
6676 when.sec += RX_CHALLENGE_TIMEOUT;
6677 conn->challengeEvent =
6678 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6683 /* Call this routine to start requesting the client to authenticate
6684 * itself. This will continue until authentication is established,
6685 * the call times out, or an invalid response is returned. The
6686 * security object associated with the connection is asked to create
6687 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6688 * defined earlier. */
6690 rxi_ChallengeOn(struct rx_connection *conn)
6692 if (!conn->challengeEvent) {
6693 RXS_CreateChallenge(conn->securityObject, conn);
6694 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6699 /* rxi_ComputeRoundTripTime is called with peer locked. */
6700 /* peer may be null */
6702 rxi_ComputeRoundTripTime(struct rx_packet *p,
6703 struct rx_ackPacket *ack,
6704 struct rx_call *call,
6705 struct rx_peer *peer,
6708 struct clock thisRtt, *sentp;
6712 /* If the ACK is delayed, then do nothing */
6713 if (ack->reason == RX_ACK_DELAY)
6716 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6717 * their RTT multiple times, so only include the RTT of the last packet
6719 if (p->flags & RX_JUMBO_PACKET)
6722 /* Use the serial number to determine which transmission the ACK is for,
6723 * and set the sent time to match this. If we have no serial number, then
6724 * only use the ACK for RTT calculations if the packet has not been
6728 serial = ntohl(ack->serial);
6730 if (serial == p->header.serial) {
6731 sentp = &p->timeSent;
6732 } else if (serial == p->firstSerial) {
6733 sentp = &p->firstSent;
6734 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6735 sentp = &p->firstSent;
6739 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6740 sentp = &p->firstSent;
6747 if (clock_Lt(&thisRtt, sentp))
6748 return; /* somebody set the clock back, don't count this time. */
6750 clock_Sub(&thisRtt, sentp);
6751 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6752 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6754 if (clock_IsZero(&thisRtt)) {
6756 * The actual round trip time is shorter than the
6757 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6758 * Since we can't tell which at the moment we will assume 1ms.
6760 thisRtt.usec = 1000;
6763 if (rx_stats_active) {
6764 MUTEX_ENTER(&rx_stats_mutex);
6765 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6766 rx_stats.minRtt = thisRtt;
6767 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6768 if (thisRtt.sec > 60) {
6769 MUTEX_EXIT(&rx_stats_mutex);
6770 return; /* somebody set the clock ahead */
6772 rx_stats.maxRtt = thisRtt;
6774 clock_Add(&rx_stats.totalRtt, &thisRtt);
6775 rx_atomic_inc(&rx_stats.nRttSamples);
6776 MUTEX_EXIT(&rx_stats_mutex);
6779 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6781 /* Apply VanJacobson round-trip estimations */
6786 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6787 * srtt is stored as fixed point with 3 bits after the binary
6788 * point (i.e., scaled by 8). The following magic is
6789 * equivalent to the smoothing algorithm in rfc793 with an
6790 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6791 * srtt'*8 = rtt + srtt*7
6792 * srtt'*8 = srtt*8 + rtt - srtt
6793 * srtt' = srtt + rtt/8 - srtt/8
6794 * srtt' = srtt + (rtt - srtt)/8
6797 delta = _8THMSEC(&thisRtt) - call->rtt;
6798 call->rtt += (delta >> 3);
6801 * We accumulate a smoothed rtt variance (actually, a smoothed
6802 * mean difference), then set the retransmit timer to smoothed
6803 * rtt + 4 times the smoothed variance (was 2x in van's original
6804 * paper, but 4x works better for me, and apparently for him as
6806 * rttvar is stored as
6807 * fixed point with 2 bits after the binary point (scaled by
6808 * 4). The following is equivalent to rfc793 smoothing with
6809 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6810 * rttvar'*4 = rttvar*3 + |delta|
6811 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6812 * rttvar' = rttvar + |delta|/4 - rttvar/4
6813 * rttvar' = rttvar + (|delta| - rttvar)/4
6814 * This replaces rfc793's wired-in beta.
6815 * dev*4 = dev*4 + (|actual - expected| - dev)
6821 delta -= (call->rtt_dev << 1);
6822 call->rtt_dev += (delta >> 3);
6824 /* I don't have a stored RTT so I start with this value. Since I'm
6825 * probably just starting a call, and will be pushing more data down
6826 * this, I expect congestion to increase rapidly. So I fudge a
6827 * little, and I set deviance to half the rtt. In practice,
6828 * deviance tends to approach something a little less than
6829 * half the smoothed rtt. */
6830 call->rtt = _8THMSEC(&thisRtt) + 8;
6831 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6833 /* the smoothed RTT time is RTT + 4*MDEV
6835 * We allow a user specified minimum to be set for this, to allow clamping
6836 * at a minimum value in the same way as TCP. In addition, we have to allow
6837 * for the possibility that this packet is answered by a delayed ACK, so we
6838 * add on a fixed 200ms to account for that timer expiring.
6841 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6842 rx_minPeerTimeout) + 200;
6843 clock_Zero(&call->rto);
6844 clock_Addmsec(&call->rto, rtt_timeout);
6846 /* Update the peer, so any new calls start with our values */
6847 peer->rtt_dev = call->rtt_dev;
6848 peer->rtt = call->rtt;
6850 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6851 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6855 /* Find all server connections that have not been active for a long time, and
6858 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
6861 struct clock now, when;
6862 clock_GetTime(&now);
6864 /* Find server connection structures that haven't been used for
6865 * greater than rx_idleConnectionTime */
6867 struct rx_connection **conn_ptr, **conn_end;
6868 int i, havecalls = 0;
6869 MUTEX_ENTER(&rx_connHashTable_lock);
6870 for (conn_ptr = &rx_connHashTable[0], conn_end =
6871 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6873 struct rx_connection *conn, *next;
6874 struct rx_call *call;
6878 for (conn = *conn_ptr; conn; conn = next) {
6879 /* XXX -- Shouldn't the connection be locked? */
6882 for (i = 0; i < RX_MAXCALLS; i++) {
6883 call = conn->call[i];
6887 code = MUTEX_TRYENTER(&call->lock);
6890 #ifdef RX_ENABLE_LOCKS
6891 result = rxi_CheckCall(call, 1);
6892 #else /* RX_ENABLE_LOCKS */
6893 result = rxi_CheckCall(call);
6894 #endif /* RX_ENABLE_LOCKS */
6895 MUTEX_EXIT(&call->lock);
6897 /* If CheckCall freed the call, it might
6898 * have destroyed the connection as well,
6899 * which screws up the linked lists.
6905 if (conn->type == RX_SERVER_CONNECTION) {
6906 /* This only actually destroys the connection if
6907 * there are no outstanding calls */
6908 MUTEX_ENTER(&conn->conn_data_lock);
6909 MUTEX_ENTER(&rx_refcnt_mutex);
6910 if (!havecalls && !conn->refCount
6911 && ((conn->lastSendTime + rx_idleConnectionTime) <
6913 conn->refCount++; /* it will be decr in rx_DestroyConn */
6914 MUTEX_EXIT(&rx_refcnt_mutex);
6915 MUTEX_EXIT(&conn->conn_data_lock);
6916 #ifdef RX_ENABLE_LOCKS
6917 rxi_DestroyConnectionNoLock(conn);
6918 #else /* RX_ENABLE_LOCKS */
6919 rxi_DestroyConnection(conn);
6920 #endif /* RX_ENABLE_LOCKS */
6922 #ifdef RX_ENABLE_LOCKS
6924 MUTEX_EXIT(&rx_refcnt_mutex);
6925 MUTEX_EXIT(&conn->conn_data_lock);
6927 #endif /* RX_ENABLE_LOCKS */
6931 #ifdef RX_ENABLE_LOCKS
6932 while (rx_connCleanup_list) {
6933 struct rx_connection *conn;
6934 conn = rx_connCleanup_list;
6935 rx_connCleanup_list = rx_connCleanup_list->next;
6936 MUTEX_EXIT(&rx_connHashTable_lock);
6937 rxi_CleanupConnection(conn);
6938 MUTEX_ENTER(&rx_connHashTable_lock);
6940 MUTEX_EXIT(&rx_connHashTable_lock);
6941 #endif /* RX_ENABLE_LOCKS */
6944 /* Find any peer structures that haven't been used (haven't had an
6945 * associated connection) for greater than rx_idlePeerTime */
6947 struct rx_peer **peer_ptr, **peer_end;
6951 * Why do we need to hold the rx_peerHashTable_lock across
6952 * the incrementing of peer_ptr since the rx_peerHashTable
6953 * array is not changing? We don't.
6955 * By dropping the lock periodically we can permit other
6956 * activities to be performed while a rxi_ReapConnections
6957 * call is in progress. The goal of reap connections
6958 * is to clean up quickly without causing large amounts
6959 * of contention. Therefore, it is important that global
6960 * mutexes not be held for extended periods of time.
6962 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6963 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6965 struct rx_peer *peer, *next, *prev;
6967 MUTEX_ENTER(&rx_peerHashTable_lock);
6968 for (prev = peer = *peer_ptr; peer; peer = next) {
6970 code = MUTEX_TRYENTER(&peer->peer_lock);
6971 if ((code) && (peer->refCount == 0)
6972 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6973 rx_interface_stat_p rpc_stat, nrpc_stat;
6977 * now know that this peer object is one to be
6978 * removed from the hash table. Once it is removed
6979 * it can't be referenced by other threads.
6980 * Lets remove it first and decrement the struct
6981 * nPeerStructs count.
6983 if (peer == *peer_ptr) {
6989 if (rx_stats_active)
6990 rx_atomic_dec(&rx_stats.nPeerStructs);
6993 * Now if we hold references on 'prev' and 'next'
6994 * we can safely drop the rx_peerHashTable_lock
6995 * while we destroy this 'peer' object.
7001 MUTEX_EXIT(&rx_peerHashTable_lock);
7003 MUTEX_EXIT(&peer->peer_lock);
7004 MUTEX_DESTROY(&peer->peer_lock);
7006 (&peer->rpcStats, rpc_stat, nrpc_stat,
7007 rx_interface_stat)) {
7008 unsigned int num_funcs;
7011 queue_Remove(&rpc_stat->queue_header);
7012 queue_Remove(&rpc_stat->all_peers);
7013 num_funcs = rpc_stat->stats[0].func_total;
7015 sizeof(rx_interface_stat_t) +
7016 rpc_stat->stats[0].func_total *
7017 sizeof(rx_function_entry_v1_t);
7019 rxi_Free(rpc_stat, space);
7021 MUTEX_ENTER(&rx_rpc_stats);
7022 rxi_rpc_peer_stat_cnt -= num_funcs;
7023 MUTEX_EXIT(&rx_rpc_stats);
7028 * Regain the rx_peerHashTable_lock and
7029 * decrement the reference count on 'prev'
7032 MUTEX_ENTER(&rx_peerHashTable_lock);
7039 MUTEX_EXIT(&peer->peer_lock);
7044 MUTEX_EXIT(&rx_peerHashTable_lock);
7048 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7049 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7050 * GC, just below. Really, we shouldn't have to keep moving packets from
7051 * one place to another, but instead ought to always know if we can
7052 * afford to hold onto a packet in its particular use. */
7053 MUTEX_ENTER(&rx_freePktQ_lock);
7054 if (rx_waitingForPackets) {
7055 rx_waitingForPackets = 0;
7056 #ifdef RX_ENABLE_LOCKS
7057 CV_BROADCAST(&rx_waitingForPackets_cv);
7059 osi_rxWakeup(&rx_waitingForPackets);
7062 MUTEX_EXIT(&rx_freePktQ_lock);
7065 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7066 rxevent_Put(rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0));
7070 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7071 * rx.h is sort of strange this is better. This is called with a security
7072 * object before it is discarded. Each connection using a security object has
7073 * its own refcount to the object so it won't actually be freed until the last
7074 * connection is destroyed.
7076 * This is the only rxs module call. A hold could also be written but no one
7080 rxs_Release(struct rx_securityClass *aobj)
7082 return RXS_Close(aobj);
7090 #define TRACE_OPTION_RX_DEBUG 16
7098 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7099 0, KEY_QUERY_VALUE, &parmKey);
7100 if (code != ERROR_SUCCESS)
7103 dummyLen = sizeof(TraceOption);
7104 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7105 (BYTE *) &TraceOption, &dummyLen);
7106 if (code == ERROR_SUCCESS) {
7107 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7109 RegCloseKey (parmKey);
7110 #endif /* AFS_NT40_ENV */
7115 rx_DebugOnOff(int on)
7119 rxdebug_active = on;
7125 rx_StatsOnOff(int on)
7127 rx_stats_active = on;
7131 /* Don't call this debugging routine directly; use dpf */
7133 rxi_DebugPrint(char *format, ...)
7142 va_start(ap, format);
7144 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7147 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7149 OutputDebugString(msg);
7155 va_start(ap, format);
7157 clock_GetTime(&now);
7158 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7159 (unsigned int)now.usec);
7160 vfprintf(rx_Log, format, ap);
7168 * This function is used to process the rx_stats structure that is local
7169 * to a process as well as an rx_stats structure received from a remote
7170 * process (via rxdebug). Therefore, it needs to do minimal version
7174 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7175 afs_int32 freePackets, char version)
7179 if (size != sizeof(struct rx_statistics)) {
7181 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7182 size, sizeof(struct rx_statistics));
7185 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7188 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7189 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7190 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7191 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7192 s->specialPktAllocFailures);
7194 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7195 s->receivePktAllocFailures, s->sendPktAllocFailures,
7196 s->specialPktAllocFailures);
7200 " greedy %u, " "bogusReads %u (last from host %x), "
7201 "noPackets %u, " "noBuffers %u, " "selects %u, "
7202 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7203 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7204 s->selects, s->sendSelects);
7206 fprintf(file, " packets read: ");
7207 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7208 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7210 fprintf(file, "\n");
7213 " other read counters: data %u, " "ack %u, " "dup %u "
7214 "spurious %u " "dally %u\n", s->dataPacketsRead,
7215 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7216 s->ignorePacketDally);
7218 fprintf(file, " packets sent: ");
7219 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7220 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7222 fprintf(file, "\n");
7225 " other send counters: ack %u, " "data %u (not resends), "
7226 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7227 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7228 s->dataPacketsPushed, s->ignoreAckedPacket);
7231 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7232 s->netSendFailures, (int)s->fatalErrors);
7234 if (s->nRttSamples) {
7235 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7236 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7238 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7239 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7243 " %d server connections, " "%d client connections, "
7244 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7245 s->nServerConns, s->nClientConns, s->nPeerStructs,
7246 s->nCallStructs, s->nFreeCallStructs);
7248 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7249 fprintf(file, " %d clock updates\n", clock_nUpdates);
7253 /* for backward compatibility */
7255 rx_PrintStats(FILE * file)
7257 MUTEX_ENTER(&rx_stats_mutex);
7258 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7259 sizeof(rx_stats), rx_nFreePackets,
7261 MUTEX_EXIT(&rx_stats_mutex);
7265 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7267 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
7268 ntohl(peer->host), (int)ntohs(peer->port), (int)peer->burstSize,
7269 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
7272 " Rtt %d, " "total sent %d, " "resent %d\n",
7273 peer->rtt, peer->nSent, peer->reSends);
7276 " Packet size %d, " "max in packet skew %d, "
7277 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
7278 (int)peer->outPacketSkew);
7282 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7284 * This mutex protects the following static variables:
7288 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7289 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7291 #define LOCK_RX_DEBUG
7292 #define UNLOCK_RX_DEBUG
7293 #endif /* AFS_PTHREAD_ENV */
7295 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7297 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7298 u_char type, void *inputData, size_t inputLength,
7299 void *outputData, size_t outputLength)
7301 static afs_int32 counter = 100;
7302 time_t waitTime, waitCount;
7303 struct rx_header theader;
7306 struct timeval tv_now, tv_wake, tv_delta;
7307 struct sockaddr_in taddr, faddr;
7321 tp = &tbuffer[sizeof(struct rx_header)];
7322 taddr.sin_family = AF_INET;
7323 taddr.sin_port = remotePort;
7324 taddr.sin_addr.s_addr = remoteAddr;
7325 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7326 taddr.sin_len = sizeof(struct sockaddr_in);
7329 memset(&theader, 0, sizeof(theader));
7330 theader.epoch = htonl(999);
7332 theader.callNumber = htonl(counter);
7335 theader.type = type;
7336 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7337 theader.serviceId = 0;
7339 memcpy(tbuffer, &theader, sizeof(theader));
7340 memcpy(tp, inputData, inputLength);
7342 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7343 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7345 /* see if there's a packet available */
7346 gettimeofday(&tv_wake, NULL);
7347 tv_wake.tv_sec += waitTime;
7350 FD_SET(socket, &imask);
7351 tv_delta.tv_sec = tv_wake.tv_sec;
7352 tv_delta.tv_usec = tv_wake.tv_usec;
7353 gettimeofday(&tv_now, NULL);
7355 if (tv_delta.tv_usec < tv_now.tv_usec) {
7357 tv_delta.tv_usec += 1000000;
7360 tv_delta.tv_usec -= tv_now.tv_usec;
7362 if (tv_delta.tv_sec < tv_now.tv_sec) {
7366 tv_delta.tv_sec -= tv_now.tv_sec;
7369 code = select(0, &imask, 0, 0, &tv_delta);
7370 #else /* AFS_NT40_ENV */
7371 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7372 #endif /* AFS_NT40_ENV */
7373 if (code == 1 && FD_ISSET(socket, &imask)) {
7374 /* now receive a packet */
7375 faddrLen = sizeof(struct sockaddr_in);
7377 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7378 (struct sockaddr *)&faddr, &faddrLen);
7381 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7382 if (counter == ntohl(theader.callNumber))
7390 /* see if we've timed out */
7398 code -= sizeof(struct rx_header);
7399 if (code > outputLength)
7400 code = outputLength;
7401 memcpy(outputData, tp, code);
7404 #endif /* RXDEBUG */
7407 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7408 afs_uint16 remotePort, struct rx_debugStats * stat,
7409 afs_uint32 * supportedValues)
7411 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7413 struct rx_debugIn in;
7415 *supportedValues = 0;
7416 in.type = htonl(RX_DEBUGI_GETSTATS);
7419 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7420 &in, sizeof(in), stat, sizeof(*stat));
7423 * If the call was successful, fixup the version and indicate
7424 * what contents of the stat structure are valid.
7425 * Also do net to host conversion of fields here.
7429 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7430 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7432 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7433 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7435 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7436 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7438 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7439 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7441 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7442 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7444 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7445 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7447 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7448 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7450 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7451 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7453 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7454 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7456 stat->nFreePackets = ntohl(stat->nFreePackets);
7457 stat->packetReclaims = ntohl(stat->packetReclaims);
7458 stat->callsExecuted = ntohl(stat->callsExecuted);
7459 stat->nWaiting = ntohl(stat->nWaiting);
7460 stat->idleThreads = ntohl(stat->idleThreads);
7461 stat->nWaited = ntohl(stat->nWaited);
7462 stat->nPackets = ntohl(stat->nPackets);
7471 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7472 afs_uint16 remotePort, struct rx_statistics * stat,
7473 afs_uint32 * supportedValues)
7475 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7477 struct rx_debugIn in;
7478 afs_int32 *lp = (afs_int32 *) stat;
7482 * supportedValues is currently unused, but added to allow future
7483 * versioning of this function.
7486 *supportedValues = 0;
7487 in.type = htonl(RX_DEBUGI_RXSTATS);
7489 memset(stat, 0, sizeof(*stat));
7491 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7492 &in, sizeof(in), stat, sizeof(*stat));
7497 * Do net to host conversion here
7500 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7511 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7512 afs_uint16 remotePort, size_t version_length,
7515 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7517 return MakeDebugCall(socket, remoteAddr, remotePort,
7518 RX_PACKET_TYPE_VERSION, a, 1, version,
7526 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7527 afs_uint16 remotePort, afs_int32 * nextConnection,
7528 int allConnections, afs_uint32 debugSupportedValues,
7529 struct rx_debugConn * conn,
7530 afs_uint32 * supportedValues)
7532 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7534 struct rx_debugIn in;
7538 * supportedValues is currently unused, but added to allow future
7539 * versioning of this function.
7542 *supportedValues = 0;
7543 if (allConnections) {
7544 in.type = htonl(RX_DEBUGI_GETALLCONN);
7546 in.type = htonl(RX_DEBUGI_GETCONN);
7548 in.index = htonl(*nextConnection);
7549 memset(conn, 0, sizeof(*conn));
7551 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7552 &in, sizeof(in), conn, sizeof(*conn));
7555 *nextConnection += 1;
7558 * Convert old connection format to new structure.
7561 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7562 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7563 #define MOVEvL(a) (conn->a = vL->a)
7565 /* any old or unrecognized version... */
7566 for (i = 0; i < RX_MAXCALLS; i++) {
7567 MOVEvL(callState[i]);
7568 MOVEvL(callMode[i]);
7569 MOVEvL(callFlags[i]);
7570 MOVEvL(callOther[i]);
7572 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7573 MOVEvL(secStats.type);
7574 MOVEvL(secStats.level);
7575 MOVEvL(secStats.flags);
7576 MOVEvL(secStats.expires);
7577 MOVEvL(secStats.packetsReceived);
7578 MOVEvL(secStats.packetsSent);
7579 MOVEvL(secStats.bytesReceived);
7580 MOVEvL(secStats.bytesSent);
7585 * Do net to host conversion here
7587 * I don't convert host or port since we are most likely
7588 * going to want these in NBO.
7590 conn->cid = ntohl(conn->cid);
7591 conn->serial = ntohl(conn->serial);
7592 for (i = 0; i < RX_MAXCALLS; i++) {
7593 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7595 conn->error = ntohl(conn->error);
7596 conn->secStats.flags = ntohl(conn->secStats.flags);
7597 conn->secStats.expires = ntohl(conn->secStats.expires);
7598 conn->secStats.packetsReceived =
7599 ntohl(conn->secStats.packetsReceived);
7600 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7601 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7602 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7603 conn->epoch = ntohl(conn->epoch);
7604 conn->natMTU = ntohl(conn->natMTU);
7613 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7614 afs_uint16 remotePort, afs_int32 * nextPeer,
7615 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7616 afs_uint32 * supportedValues)
7618 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7620 struct rx_debugIn in;
7623 * supportedValues is currently unused, but added to allow future
7624 * versioning of this function.
7627 *supportedValues = 0;
7628 in.type = htonl(RX_DEBUGI_GETPEER);
7629 in.index = htonl(*nextPeer);
7630 memset(peer, 0, sizeof(*peer));
7632 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7633 &in, sizeof(in), peer, sizeof(*peer));
7639 * Do net to host conversion here
7641 * I don't convert host or port since we are most likely
7642 * going to want these in NBO.
7644 peer->ifMTU = ntohs(peer->ifMTU);
7645 peer->idleWhen = ntohl(peer->idleWhen);
7646 peer->refCount = ntohs(peer->refCount);
7647 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7648 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7649 peer->rtt = ntohl(peer->rtt);
7650 peer->rtt_dev = ntohl(peer->rtt_dev);
7651 peer->timeout.sec = 0;
7652 peer->timeout.usec = 0;
7653 peer->nSent = ntohl(peer->nSent);
7654 peer->reSends = ntohl(peer->reSends);
7655 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7656 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7657 peer->natMTU = ntohs(peer->natMTU);
7658 peer->maxMTU = ntohs(peer->maxMTU);
7659 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7660 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7661 peer->MTU = ntohs(peer->MTU);
7662 peer->cwind = ntohs(peer->cwind);
7663 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7664 peer->congestSeq = ntohs(peer->congestSeq);
7665 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7666 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7667 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7668 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7677 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7678 struct rx_debugPeer * peerStats)
7681 afs_int32 error = 1; /* default to "did not succeed" */
7682 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7684 MUTEX_ENTER(&rx_peerHashTable_lock);
7685 for(tp = rx_peerHashTable[hashValue];
7686 tp != NULL; tp = tp->next) {
7687 if (tp->host == peerHost)
7693 MUTEX_EXIT(&rx_peerHashTable_lock);
7697 MUTEX_ENTER(&tp->peer_lock);
7698 peerStats->host = tp->host;
7699 peerStats->port = tp->port;
7700 peerStats->ifMTU = tp->ifMTU;
7701 peerStats->idleWhen = tp->idleWhen;
7702 peerStats->refCount = tp->refCount;
7703 peerStats->burstSize = tp->burstSize;
7704 peerStats->burst = tp->burst;
7705 peerStats->burstWait.sec = tp->burstWait.sec;
7706 peerStats->burstWait.usec = tp->burstWait.usec;
7707 peerStats->rtt = tp->rtt;
7708 peerStats->rtt_dev = tp->rtt_dev;
7709 peerStats->timeout.sec = 0;
7710 peerStats->timeout.usec = 0;
7711 peerStats->nSent = tp->nSent;
7712 peerStats->reSends = tp->reSends;
7713 peerStats->inPacketSkew = tp->inPacketSkew;
7714 peerStats->outPacketSkew = tp->outPacketSkew;
7715 peerStats->natMTU = tp->natMTU;
7716 peerStats->maxMTU = tp->maxMTU;
7717 peerStats->maxDgramPackets = tp->maxDgramPackets;
7718 peerStats->ifDgramPackets = tp->ifDgramPackets;
7719 peerStats->MTU = tp->MTU;
7720 peerStats->cwind = tp->cwind;
7721 peerStats->nDgramPackets = tp->nDgramPackets;
7722 peerStats->congestSeq = tp->congestSeq;
7723 peerStats->bytesSent.high = tp->bytesSent.high;
7724 peerStats->bytesSent.low = tp->bytesSent.low;
7725 peerStats->bytesReceived.high = tp->bytesReceived.high;
7726 peerStats->bytesReceived.low = tp->bytesReceived.low;
7727 MUTEX_EXIT(&tp->peer_lock);
7729 MUTEX_ENTER(&rx_peerHashTable_lock);
7732 MUTEX_EXIT(&rx_peerHashTable_lock);
7740 struct rx_serverQueueEntry *np;
7743 struct rx_call *call;
7744 struct rx_serverQueueEntry *sq;
7748 if (rxinit_status == 1) {
7750 return; /* Already shutdown. */
7754 #ifndef AFS_PTHREAD_ENV
7755 FD_ZERO(&rx_selectMask);
7756 #endif /* AFS_PTHREAD_ENV */
7757 rxi_dataQuota = RX_MAX_QUOTA;
7758 #ifndef AFS_PTHREAD_ENV
7760 #endif /* AFS_PTHREAD_ENV */
7763 #ifndef AFS_PTHREAD_ENV
7764 #ifndef AFS_USE_GETTIMEOFDAY
7766 #endif /* AFS_USE_GETTIMEOFDAY */
7767 #endif /* AFS_PTHREAD_ENV */
7769 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7770 call = queue_First(&rx_freeCallQueue, rx_call);
7772 rxi_Free(call, sizeof(struct rx_call));
7775 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7776 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7782 struct rx_peer **peer_ptr, **peer_end;
7783 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7784 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7786 struct rx_peer *peer, *next;
7788 MUTEX_ENTER(&rx_peerHashTable_lock);
7789 for (peer = *peer_ptr; peer; peer = next) {
7790 rx_interface_stat_p rpc_stat, nrpc_stat;
7793 MUTEX_ENTER(&rx_rpc_stats);
7794 MUTEX_ENTER(&peer->peer_lock);
7796 (&peer->rpcStats, rpc_stat, nrpc_stat,
7797 rx_interface_stat)) {
7798 unsigned int num_funcs;
7801 queue_Remove(&rpc_stat->queue_header);
7802 queue_Remove(&rpc_stat->all_peers);
7803 num_funcs = rpc_stat->stats[0].func_total;
7805 sizeof(rx_interface_stat_t) +
7806 rpc_stat->stats[0].func_total *
7807 sizeof(rx_function_entry_v1_t);
7809 rxi_Free(rpc_stat, space);
7811 /* rx_rpc_stats must be held */
7812 rxi_rpc_peer_stat_cnt -= num_funcs;
7814 MUTEX_EXIT(&peer->peer_lock);
7815 MUTEX_EXIT(&rx_rpc_stats);
7819 if (rx_stats_active)
7820 rx_atomic_dec(&rx_stats.nPeerStructs);
7822 MUTEX_EXIT(&rx_peerHashTable_lock);
7825 for (i = 0; i < RX_MAX_SERVICES; i++) {
7827 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7829 for (i = 0; i < rx_hashTableSize; i++) {
7830 struct rx_connection *tc, *ntc;
7831 MUTEX_ENTER(&rx_connHashTable_lock);
7832 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7834 for (j = 0; j < RX_MAXCALLS; j++) {
7836 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7839 rxi_Free(tc, sizeof(*tc));
7841 MUTEX_EXIT(&rx_connHashTable_lock);
7844 MUTEX_ENTER(&freeSQEList_lock);
7846 while ((np = rx_FreeSQEList)) {
7847 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7848 MUTEX_DESTROY(&np->lock);
7849 rxi_Free(np, sizeof(*np));
7852 MUTEX_EXIT(&freeSQEList_lock);
7853 MUTEX_DESTROY(&freeSQEList_lock);
7854 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7855 MUTEX_DESTROY(&rx_connHashTable_lock);
7856 MUTEX_DESTROY(&rx_peerHashTable_lock);
7857 MUTEX_DESTROY(&rx_serverPool_lock);
7859 osi_Free(rx_connHashTable,
7860 rx_hashTableSize * sizeof(struct rx_connection *));
7861 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7863 UNPIN(rx_connHashTable,
7864 rx_hashTableSize * sizeof(struct rx_connection *));
7865 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7867 rxi_FreeAllPackets();
7869 MUTEX_ENTER(&rx_quota_mutex);
7870 rxi_dataQuota = RX_MAX_QUOTA;
7871 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7872 MUTEX_EXIT(&rx_quota_mutex);
7877 #ifdef RX_ENABLE_LOCKS
7879 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7881 if (!MUTEX_ISMINE(lockaddr))
7882 osi_Panic("Lock not held: %s", msg);
7884 #endif /* RX_ENABLE_LOCKS */
7889 * Routines to implement connection specific data.
7893 rx_KeyCreate(rx_destructor_t rtn)
7896 MUTEX_ENTER(&rxi_keyCreate_lock);
7897 key = rxi_keyCreate_counter++;
7898 rxi_keyCreate_destructor = (rx_destructor_t *)
7899 realloc((void *)rxi_keyCreate_destructor,
7900 (key + 1) * sizeof(rx_destructor_t));
7901 rxi_keyCreate_destructor[key] = rtn;
7902 MUTEX_EXIT(&rxi_keyCreate_lock);
7907 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7910 MUTEX_ENTER(&conn->conn_data_lock);
7911 if (!conn->specific) {
7912 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7913 for (i = 0; i < key; i++)
7914 conn->specific[i] = NULL;
7915 conn->nSpecific = key + 1;
7916 conn->specific[key] = ptr;
7917 } else if (key >= conn->nSpecific) {
7918 conn->specific = (void **)
7919 realloc(conn->specific, (key + 1) * sizeof(void *));
7920 for (i = conn->nSpecific; i < key; i++)
7921 conn->specific[i] = NULL;
7922 conn->nSpecific = key + 1;
7923 conn->specific[key] = ptr;
7925 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7926 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7927 conn->specific[key] = ptr;
7929 MUTEX_EXIT(&conn->conn_data_lock);
7933 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7936 MUTEX_ENTER(&svc->svc_data_lock);
7937 if (!svc->specific) {
7938 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
7939 for (i = 0; i < key; i++)
7940 svc->specific[i] = NULL;
7941 svc->nSpecific = key + 1;
7942 svc->specific[key] = ptr;
7943 } else if (key >= svc->nSpecific) {
7944 svc->specific = (void **)
7945 realloc(svc->specific, (key + 1) * sizeof(void *));
7946 for (i = svc->nSpecific; i < key; i++)
7947 svc->specific[i] = NULL;
7948 svc->nSpecific = key + 1;
7949 svc->specific[key] = ptr;
7951 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7952 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7953 svc->specific[key] = ptr;
7955 MUTEX_EXIT(&svc->svc_data_lock);
7959 rx_GetSpecific(struct rx_connection *conn, int key)
7962 MUTEX_ENTER(&conn->conn_data_lock);
7963 if (key >= conn->nSpecific)
7966 ptr = conn->specific[key];
7967 MUTEX_EXIT(&conn->conn_data_lock);
7972 rx_GetServiceSpecific(struct rx_service *svc, int key)
7975 MUTEX_ENTER(&svc->svc_data_lock);
7976 if (key >= svc->nSpecific)
7979 ptr = svc->specific[key];
7980 MUTEX_EXIT(&svc->svc_data_lock);
7985 #endif /* !KERNEL */
7988 * processStats is a queue used to store the statistics for the local
7989 * process. Its contents are similar to the contents of the rpcStats
7990 * queue on a rx_peer structure, but the actual data stored within
7991 * this queue contains totals across the lifetime of the process (assuming
7992 * the stats have not been reset) - unlike the per peer structures
7993 * which can come and go based upon the peer lifetime.
7996 static struct rx_queue processStats = { &processStats, &processStats };
7999 * peerStats is a queue used to store the statistics for all peer structs.
8000 * Its contents are the union of all the peer rpcStats queues.
8003 static struct rx_queue peerStats = { &peerStats, &peerStats };
8006 * rxi_monitor_processStats is used to turn process wide stat collection
8010 static int rxi_monitor_processStats = 0;
8013 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8016 static int rxi_monitor_peerStats = 0;
8019 * rxi_AddRpcStat - given all of the information for a particular rpc
8020 * call, create (if needed) and update the stat totals for the rpc.
8024 * IN stats - the queue of stats that will be updated with the new value
8026 * IN rxInterface - a unique number that identifies the rpc interface
8028 * IN currentFunc - the index of the function being invoked
8030 * IN totalFunc - the total number of functions in this interface
8032 * IN queueTime - the amount of time this function waited for a thread
8034 * IN execTime - the amount of time this function invocation took to execute
8036 * IN bytesSent - the number bytes sent by this invocation
8038 * IN bytesRcvd - the number bytes received by this invocation
8040 * IN isServer - if true, this invocation was made to a server
8042 * IN remoteHost - the ip address of the remote host
8044 * IN remotePort - the port of the remote host
8046 * IN addToPeerList - if != 0, add newly created stat to the global peer list
8048 * INOUT counter - if a new stats structure is allocated, the counter will
8049 * be updated with the new number of allocated stat structures
8057 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
8058 afs_uint32 currentFunc, afs_uint32 totalFunc,
8059 struct clock *queueTime, struct clock *execTime,
8060 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
8061 afs_uint32 remoteHost, afs_uint32 remotePort,
8062 int addToPeerList, unsigned int *counter)
8065 rx_interface_stat_p rpc_stat, nrpc_stat;
8068 * See if there's already a structure for this interface
8071 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8072 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8073 && (rpc_stat->stats[0].remote_is_server == isServer))
8078 * Didn't find a match so allocate a new structure and add it to the
8082 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
8083 || (rpc_stat->stats[0].interfaceId != rxInterface)
8084 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8089 sizeof(rx_interface_stat_t) +
8090 totalFunc * sizeof(rx_function_entry_v1_t);
8092 rpc_stat = rxi_Alloc(space);
8093 if (rpc_stat == NULL) {
8097 *counter += totalFunc;
8098 for (i = 0; i < totalFunc; i++) {
8099 rpc_stat->stats[i].remote_peer = remoteHost;
8100 rpc_stat->stats[i].remote_port = remotePort;
8101 rpc_stat->stats[i].remote_is_server = isServer;
8102 rpc_stat->stats[i].interfaceId = rxInterface;
8103 rpc_stat->stats[i].func_total = totalFunc;
8104 rpc_stat->stats[i].func_index = i;
8105 hzero(rpc_stat->stats[i].invocations);
8106 hzero(rpc_stat->stats[i].bytes_sent);
8107 hzero(rpc_stat->stats[i].bytes_rcvd);
8108 rpc_stat->stats[i].queue_time_sum.sec = 0;
8109 rpc_stat->stats[i].queue_time_sum.usec = 0;
8110 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8111 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8112 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8113 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8114 rpc_stat->stats[i].queue_time_max.sec = 0;
8115 rpc_stat->stats[i].queue_time_max.usec = 0;
8116 rpc_stat->stats[i].execution_time_sum.sec = 0;
8117 rpc_stat->stats[i].execution_time_sum.usec = 0;
8118 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8119 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8120 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8121 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8122 rpc_stat->stats[i].execution_time_max.sec = 0;
8123 rpc_stat->stats[i].execution_time_max.usec = 0;
8125 queue_Prepend(stats, rpc_stat);
8126 if (addToPeerList) {
8127 queue_Prepend(&peerStats, &rpc_stat->all_peers);
8132 * Increment the stats for this function
8135 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
8136 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
8137 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
8138 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8139 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8140 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8141 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8143 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8144 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8146 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8147 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8149 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8150 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8152 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8153 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8161 * rx_IncrementTimeAndCount - increment the times and count for a particular
8166 * IN peer - the peer who invoked the rpc
8168 * IN rxInterface - a unique number that identifies the rpc interface
8170 * IN currentFunc - the index of the function being invoked
8172 * IN totalFunc - the total number of functions in this interface
8174 * IN queueTime - the amount of time this function waited for a thread
8176 * IN execTime - the amount of time this function invocation took to execute
8178 * IN bytesSent - the number bytes sent by this invocation
8180 * IN bytesRcvd - the number bytes received by this invocation
8182 * IN isServer - if true, this invocation was made to a server
8190 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8191 afs_uint32 currentFunc, afs_uint32 totalFunc,
8192 struct clock *queueTime, struct clock *execTime,
8193 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8197 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8200 MUTEX_ENTER(&rx_rpc_stats);
8202 if (rxi_monitor_peerStats) {
8203 MUTEX_ENTER(&peer->peer_lock);
8204 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8205 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8206 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8207 MUTEX_EXIT(&peer->peer_lock);
8210 if (rxi_monitor_processStats) {
8211 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8212 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8213 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8216 MUTEX_EXIT(&rx_rpc_stats);
8221 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8225 * IN callerVersion - the rpc stat version of the caller.
8227 * IN count - the number of entries to marshall.
8229 * IN stats - pointer to stats to be marshalled.
8231 * OUT ptr - Where to store the marshalled data.
8238 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8239 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8245 * We only support the first version
8247 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8248 *(ptr++) = stats->remote_peer;
8249 *(ptr++) = stats->remote_port;
8250 *(ptr++) = stats->remote_is_server;
8251 *(ptr++) = stats->interfaceId;
8252 *(ptr++) = stats->func_total;
8253 *(ptr++) = stats->func_index;
8254 *(ptr++) = hgethi(stats->invocations);
8255 *(ptr++) = hgetlo(stats->invocations);
8256 *(ptr++) = hgethi(stats->bytes_sent);
8257 *(ptr++) = hgetlo(stats->bytes_sent);
8258 *(ptr++) = hgethi(stats->bytes_rcvd);
8259 *(ptr++) = hgetlo(stats->bytes_rcvd);
8260 *(ptr++) = stats->queue_time_sum.sec;
8261 *(ptr++) = stats->queue_time_sum.usec;
8262 *(ptr++) = stats->queue_time_sum_sqr.sec;
8263 *(ptr++) = stats->queue_time_sum_sqr.usec;
8264 *(ptr++) = stats->queue_time_min.sec;
8265 *(ptr++) = stats->queue_time_min.usec;
8266 *(ptr++) = stats->queue_time_max.sec;
8267 *(ptr++) = stats->queue_time_max.usec;
8268 *(ptr++) = stats->execution_time_sum.sec;
8269 *(ptr++) = stats->execution_time_sum.usec;
8270 *(ptr++) = stats->execution_time_sum_sqr.sec;
8271 *(ptr++) = stats->execution_time_sum_sqr.usec;
8272 *(ptr++) = stats->execution_time_min.sec;
8273 *(ptr++) = stats->execution_time_min.usec;
8274 *(ptr++) = stats->execution_time_max.sec;
8275 *(ptr++) = stats->execution_time_max.usec;
8281 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8286 * IN callerVersion - the rpc stat version of the caller
8288 * OUT myVersion - the rpc stat version of this function
8290 * OUT clock_sec - local time seconds
8292 * OUT clock_usec - local time microseconds
8294 * OUT allocSize - the number of bytes allocated to contain stats
8296 * OUT statCount - the number stats retrieved from this process.
8298 * OUT stats - the actual stats retrieved from this process.
8302 * Returns void. If successful, stats will != NULL.
8306 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8307 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8308 size_t * allocSize, afs_uint32 * statCount,
8309 afs_uint32 ** stats)
8319 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8322 * Check to see if stats are enabled
8325 MUTEX_ENTER(&rx_rpc_stats);
8326 if (!rxi_monitor_processStats) {
8327 MUTEX_EXIT(&rx_rpc_stats);
8331 clock_GetTime(&now);
8332 *clock_sec = now.sec;
8333 *clock_usec = now.usec;
8336 * Allocate the space based upon the caller version
8338 * If the client is at an older version than we are,
8339 * we return the statistic data in the older data format, but
8340 * we still return our version number so the client knows we
8341 * are maintaining more data than it can retrieve.
8344 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8345 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8346 *statCount = rxi_rpc_process_stat_cnt;
8349 * This can't happen yet, but in the future version changes
8350 * can be handled by adding additional code here
8354 if (space > (size_t) 0) {
8356 ptr = *stats = rxi_Alloc(space);
8359 rx_interface_stat_p rpc_stat, nrpc_stat;
8363 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8365 * Copy the data based upon the caller version
8367 rx_MarshallProcessRPCStats(callerVersion,
8368 rpc_stat->stats[0].func_total,
8369 rpc_stat->stats, &ptr);
8375 MUTEX_EXIT(&rx_rpc_stats);
8380 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8384 * IN callerVersion - the rpc stat version of the caller
8386 * OUT myVersion - the rpc stat version of this function
8388 * OUT clock_sec - local time seconds
8390 * OUT clock_usec - local time microseconds
8392 * OUT allocSize - the number of bytes allocated to contain stats
8394 * OUT statCount - the number of stats retrieved from the individual
8397 * OUT stats - the actual stats retrieved from the individual peer structures.
8401 * Returns void. If successful, stats will != NULL.
8405 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8406 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8407 size_t * allocSize, afs_uint32 * statCount,
8408 afs_uint32 ** stats)
8418 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8421 * Check to see if stats are enabled
8424 MUTEX_ENTER(&rx_rpc_stats);
8425 if (!rxi_monitor_peerStats) {
8426 MUTEX_EXIT(&rx_rpc_stats);
8430 clock_GetTime(&now);
8431 *clock_sec = now.sec;
8432 *clock_usec = now.usec;
8435 * Allocate the space based upon the caller version
8437 * If the client is at an older version than we are,
8438 * we return the statistic data in the older data format, but
8439 * we still return our version number so the client knows we
8440 * are maintaining more data than it can retrieve.
8443 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8444 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8445 *statCount = rxi_rpc_peer_stat_cnt;
8448 * This can't happen yet, but in the future version changes
8449 * can be handled by adding additional code here
8453 if (space > (size_t) 0) {
8455 ptr = *stats = rxi_Alloc(space);
8458 rx_interface_stat_p rpc_stat, nrpc_stat;
8462 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8464 * We have to fix the offset of rpc_stat since we are
8465 * keeping this structure on two rx_queues. The rx_queue
8466 * package assumes that the rx_queue member is the first
8467 * member of the structure. That is, rx_queue assumes that
8468 * any one item is only on one queue at a time. We are
8469 * breaking that assumption and so we have to do a little
8470 * math to fix our pointers.
8473 fix_offset = (char *)rpc_stat;
8474 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8475 rpc_stat = (rx_interface_stat_p) fix_offset;
8478 * Copy the data based upon the caller version
8480 rx_MarshallProcessRPCStats(callerVersion,
8481 rpc_stat->stats[0].func_total,
8482 rpc_stat->stats, &ptr);
8488 MUTEX_EXIT(&rx_rpc_stats);
8493 * rx_FreeRPCStats - free memory allocated by
8494 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8498 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8499 * rx_RetrievePeerRPCStats
8501 * IN allocSize - the number of bytes in stats.
8509 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8511 rxi_Free(stats, allocSize);
8515 * rx_queryProcessRPCStats - see if process rpc stat collection is
8516 * currently enabled.
8522 * Returns 0 if stats are not enabled != 0 otherwise
8526 rx_queryProcessRPCStats(void)
8529 MUTEX_ENTER(&rx_rpc_stats);
8530 rc = rxi_monitor_processStats;
8531 MUTEX_EXIT(&rx_rpc_stats);
8536 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8542 * Returns 0 if stats are not enabled != 0 otherwise
8546 rx_queryPeerRPCStats(void)
8549 MUTEX_ENTER(&rx_rpc_stats);
8550 rc = rxi_monitor_peerStats;
8551 MUTEX_EXIT(&rx_rpc_stats);
8556 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8566 rx_enableProcessRPCStats(void)
8568 MUTEX_ENTER(&rx_rpc_stats);
8569 rx_enable_stats = 1;
8570 rxi_monitor_processStats = 1;
8571 MUTEX_EXIT(&rx_rpc_stats);
8575 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8585 rx_enablePeerRPCStats(void)
8587 MUTEX_ENTER(&rx_rpc_stats);
8588 rx_enable_stats = 1;
8589 rxi_monitor_peerStats = 1;
8590 MUTEX_EXIT(&rx_rpc_stats);
8594 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8604 rx_disableProcessRPCStats(void)
8606 rx_interface_stat_p rpc_stat, nrpc_stat;
8609 MUTEX_ENTER(&rx_rpc_stats);
8612 * Turn off process statistics and if peer stats is also off, turn
8616 rxi_monitor_processStats = 0;
8617 if (rxi_monitor_peerStats == 0) {
8618 rx_enable_stats = 0;
8621 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8622 unsigned int num_funcs = 0;
8625 queue_Remove(rpc_stat);
8626 num_funcs = rpc_stat->stats[0].func_total;
8628 sizeof(rx_interface_stat_t) +
8629 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8631 rxi_Free(rpc_stat, space);
8632 rxi_rpc_process_stat_cnt -= num_funcs;
8634 MUTEX_EXIT(&rx_rpc_stats);
8638 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8648 rx_disablePeerRPCStats(void)
8650 struct rx_peer **peer_ptr, **peer_end;
8654 * Turn off peer statistics and if process stats is also off, turn
8658 rxi_monitor_peerStats = 0;
8659 if (rxi_monitor_processStats == 0) {
8660 rx_enable_stats = 0;
8663 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8664 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8666 struct rx_peer *peer, *next, *prev;
8668 MUTEX_ENTER(&rx_peerHashTable_lock);
8669 MUTEX_ENTER(&rx_rpc_stats);
8670 for (prev = peer = *peer_ptr; peer; peer = next) {
8672 code = MUTEX_TRYENTER(&peer->peer_lock);
8674 rx_interface_stat_p rpc_stat, nrpc_stat;
8677 if (prev == *peer_ptr) {
8688 MUTEX_EXIT(&rx_peerHashTable_lock);
8691 (&peer->rpcStats, rpc_stat, nrpc_stat,
8692 rx_interface_stat)) {
8693 unsigned int num_funcs = 0;
8696 queue_Remove(&rpc_stat->queue_header);
8697 queue_Remove(&rpc_stat->all_peers);
8698 num_funcs = rpc_stat->stats[0].func_total;
8700 sizeof(rx_interface_stat_t) +
8701 rpc_stat->stats[0].func_total *
8702 sizeof(rx_function_entry_v1_t);
8704 rxi_Free(rpc_stat, space);
8705 rxi_rpc_peer_stat_cnt -= num_funcs;
8707 MUTEX_EXIT(&peer->peer_lock);
8709 MUTEX_ENTER(&rx_peerHashTable_lock);
8719 MUTEX_EXIT(&rx_rpc_stats);
8720 MUTEX_EXIT(&rx_peerHashTable_lock);
8725 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8730 * IN clearFlag - flag indicating which stats to clear
8738 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8740 rx_interface_stat_p rpc_stat, nrpc_stat;
8742 MUTEX_ENTER(&rx_rpc_stats);
8744 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8745 unsigned int num_funcs = 0, i;
8746 num_funcs = rpc_stat->stats[0].func_total;
8747 for (i = 0; i < num_funcs; i++) {
8748 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8749 hzero(rpc_stat->stats[i].invocations);
8751 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8752 hzero(rpc_stat->stats[i].bytes_sent);
8754 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8755 hzero(rpc_stat->stats[i].bytes_rcvd);
8757 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8758 rpc_stat->stats[i].queue_time_sum.sec = 0;
8759 rpc_stat->stats[i].queue_time_sum.usec = 0;
8761 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8762 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8763 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8765 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8766 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8767 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8769 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8770 rpc_stat->stats[i].queue_time_max.sec = 0;
8771 rpc_stat->stats[i].queue_time_max.usec = 0;
8773 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8774 rpc_stat->stats[i].execution_time_sum.sec = 0;
8775 rpc_stat->stats[i].execution_time_sum.usec = 0;
8777 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8778 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8779 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8781 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8782 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8783 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8785 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8786 rpc_stat->stats[i].execution_time_max.sec = 0;
8787 rpc_stat->stats[i].execution_time_max.usec = 0;
8792 MUTEX_EXIT(&rx_rpc_stats);
8796 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8801 * IN clearFlag - flag indicating which stats to clear
8809 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8811 rx_interface_stat_p rpc_stat, nrpc_stat;
8813 MUTEX_ENTER(&rx_rpc_stats);
8815 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8816 unsigned int num_funcs = 0, i;
8819 * We have to fix the offset of rpc_stat since we are
8820 * keeping this structure on two rx_queues. The rx_queue
8821 * package assumes that the rx_queue member is the first
8822 * member of the structure. That is, rx_queue assumes that
8823 * any one item is only on one queue at a time. We are
8824 * breaking that assumption and so we have to do a little
8825 * math to fix our pointers.
8828 fix_offset = (char *)rpc_stat;
8829 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8830 rpc_stat = (rx_interface_stat_p) fix_offset;
8832 num_funcs = rpc_stat->stats[0].func_total;
8833 for (i = 0; i < num_funcs; i++) {
8834 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8835 hzero(rpc_stat->stats[i].invocations);
8837 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8838 hzero(rpc_stat->stats[i].bytes_sent);
8840 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8841 hzero(rpc_stat->stats[i].bytes_rcvd);
8843 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8844 rpc_stat->stats[i].queue_time_sum.sec = 0;
8845 rpc_stat->stats[i].queue_time_sum.usec = 0;
8847 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8848 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8849 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8851 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8852 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8853 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8855 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8856 rpc_stat->stats[i].queue_time_max.sec = 0;
8857 rpc_stat->stats[i].queue_time_max.usec = 0;
8859 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8860 rpc_stat->stats[i].execution_time_sum.sec = 0;
8861 rpc_stat->stats[i].execution_time_sum.usec = 0;
8863 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8864 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8865 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8867 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8868 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8869 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8871 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8872 rpc_stat->stats[i].execution_time_max.sec = 0;
8873 rpc_stat->stats[i].execution_time_max.usec = 0;
8878 MUTEX_EXIT(&rx_rpc_stats);
8882 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8883 * is authorized to enable/disable/clear RX statistics.
8885 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8888 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8890 rxi_rxstat_userok = proc;
8894 rx_RxStatUserOk(struct rx_call *call)
8896 if (!rxi_rxstat_userok)
8898 return rxi_rxstat_userok(call);
8903 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8904 * function in the MSVC runtime DLL (msvcrt.dll).
8906 * Note: the system serializes calls to this function.
8909 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8910 DWORD reason, /* reason function is being called */
8911 LPVOID reserved) /* reserved for future use */
8914 case DLL_PROCESS_ATTACH:
8915 /* library is being attached to a process */
8919 case DLL_PROCESS_DETACH:
8926 #endif /* AFS_NT40_ENV */
8929 int rx_DumpCalls(FILE *outputFile, char *cookie)
8931 #ifdef RXDEBUG_PACKET
8932 #ifdef KDUMP_RX_LOCK
8933 struct rx_call_rx_lock *c;
8940 #define RXDPRINTF sprintf
8941 #define RXDPRINTOUT output
8943 #define RXDPRINTF fprintf
8944 #define RXDPRINTOUT outputFile
8947 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8949 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8952 for (c = rx_allCallsp; c; c = c->allNextp) {
8953 u_short rqc, tqc, iovqc;
8954 struct rx_packet *p, *np;
8956 MUTEX_ENTER(&c->lock);
8957 queue_Count(&c->rq, p, np, rx_packet, rqc);
8958 queue_Count(&c->tq, p, np, rx_packet, tqc);
8959 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8961 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, "
8962 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8963 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8964 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8965 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8966 #ifdef RX_ENABLE_LOCKS
8969 #ifdef RX_REFCOUNT_CHECK
8970 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
8971 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
8974 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,
8975 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
8976 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
8977 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
8978 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
8979 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
8980 #ifdef RX_ENABLE_LOCKS
8981 , (afs_uint32)c->refCount
8983 #ifdef RX_REFCOUNT_CHECK
8984 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
8987 MUTEX_EXIT(&c->lock);
8990 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8993 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
8995 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8997 #endif /* RXDEBUG_PACKET */