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 a boolean. It indicates whether or not RX_CALL_BUSY
167 * errors should be reported to the application when a call channel appears busy
168 * (inferred from the receipt of RX_PACKET_TYPE_BUSY packets on the channel),
169 * and there are other call channels in the connection that are not busy.
170 * If 0, we do not return errors upon receiving busy packets; we just keep
171 * trying on the same call channel until we hit a timeout.
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 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
667 call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
672 * Cancel an RTT timer for a given call.
676 * the RX call to cancel the timer for
678 * @pre call must be locked before calling this function
683 rxi_rto_cancel(struct rx_call *call)
685 rxevent_Cancel(&call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
689 * Tell the RTO timer that we have sent a packet.
691 * If the timer isn't already running, then start it. If the timer is running,
695 * the RX call that the packet has been sent on
696 * @param[in] lastPacket
697 * A flag which is true if this is the last packet for the call
699 * @pre The call must be locked before calling this function
704 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
706 if (call->resendEvent)
709 rxi_rto_startTimer(call, lastPacket, istack);
713 * Tell the RTO timer that we have received an new ACK message
715 * This function should be called whenever a call receives an ACK that
716 * acknowledges new packets. Whatever happens, we stop the current timer.
717 * If there are unacked packets in the queue which have been sent, then
718 * we restart the timer from now. Otherwise, we leave it stopped.
721 * the RX call that the ACK has been received on
725 rxi_rto_packet_acked(struct rx_call *call, int istack)
727 struct rx_packet *p, *nxp;
729 rxi_rto_cancel(call);
731 if (queue_IsEmpty(&call->tq))
734 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
735 if (p->header.seq > call->tfirst + call->twind)
738 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
739 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
747 * Set an initial round trip timeout for a peer connection
749 * @param[in] secs The timeout to set in seconds
753 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
754 peer->rtt = secs * 8000;
758 * Enables or disables the busy call channel error (RX_CALL_BUSY).
760 * @param[in] onoff Non-zero to enable busy call channel errors.
762 * @pre Neither rx_Init nor rx_InitHost have been called yet
765 rx_SetBusyChannelError(afs_int32 onoff)
767 osi_Assert(rxinit_status != 0);
768 rxi_busyChannelError = onoff ? 1 : 0;
772 * Set a delayed ack event on the specified call for the given time
774 * @param[in] call - the call on which to set the event
775 * @param[in] offset - the delay from now after which the event fires
778 rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
780 struct clock now, when;
784 clock_Add(&when, offset);
786 if (!call->delayedAckEvent
787 || clock_Gt(&call->delayedAckTime, &when)) {
789 rxevent_Cancel(&call->delayedAckEvent, call,
790 RX_CALL_REFCOUNT_DELAY);
791 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
793 call->delayedAckEvent = rxevent_Post(&when, &now,
796 call->delayedAckTime = when;
800 /* called with unincremented nRequestsRunning to see if it is OK to start
801 * a new thread in this service. Could be "no" for two reasons: over the
802 * max quota, or would prevent others from reaching their min quota.
804 #ifdef RX_ENABLE_LOCKS
805 /* This verion of QuotaOK reserves quota if it's ok while the
806 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
809 QuotaOK(struct rx_service *aservice)
811 /* check if over max quota */
812 if (aservice->nRequestsRunning >= aservice->maxProcs) {
816 /* under min quota, we're OK */
817 /* otherwise, can use only if there are enough to allow everyone
818 * to go to their min quota after this guy starts.
821 MUTEX_ENTER(&rx_quota_mutex);
822 if ((aservice->nRequestsRunning < aservice->minProcs)
823 || (rxi_availProcs > rxi_minDeficit)) {
824 aservice->nRequestsRunning++;
825 /* just started call in minProcs pool, need fewer to maintain
827 if (aservice->nRequestsRunning <= aservice->minProcs)
830 MUTEX_EXIT(&rx_quota_mutex);
833 MUTEX_EXIT(&rx_quota_mutex);
839 ReturnToServerPool(struct rx_service *aservice)
841 aservice->nRequestsRunning--;
842 MUTEX_ENTER(&rx_quota_mutex);
843 if (aservice->nRequestsRunning < aservice->minProcs)
846 MUTEX_EXIT(&rx_quota_mutex);
849 #else /* RX_ENABLE_LOCKS */
851 QuotaOK(struct rx_service *aservice)
854 /* under min quota, we're OK */
855 if (aservice->nRequestsRunning < aservice->minProcs)
858 /* check if over max quota */
859 if (aservice->nRequestsRunning >= aservice->maxProcs)
862 /* otherwise, can use only if there are enough to allow everyone
863 * to go to their min quota after this guy starts.
865 MUTEX_ENTER(&rx_quota_mutex);
866 if (rxi_availProcs > rxi_minDeficit)
868 MUTEX_EXIT(&rx_quota_mutex);
871 #endif /* RX_ENABLE_LOCKS */
874 /* Called by rx_StartServer to start up lwp's to service calls.
875 NExistingProcs gives the number of procs already existing, and which
876 therefore needn't be created. */
878 rxi_StartServerProcs(int nExistingProcs)
880 struct rx_service *service;
885 /* For each service, reserve N processes, where N is the "minimum"
886 * number of processes that MUST be able to execute a request in parallel,
887 * at any time, for that process. Also compute the maximum difference
888 * between any service's maximum number of processes that can run
889 * (i.e. the maximum number that ever will be run, and a guarantee
890 * that this number will run if other services aren't running), and its
891 * minimum number. The result is the extra number of processes that
892 * we need in order to provide the latter guarantee */
893 for (i = 0; i < RX_MAX_SERVICES; i++) {
895 service = rx_services[i];
896 if (service == (struct rx_service *)0)
898 nProcs += service->minProcs;
899 diff = service->maxProcs - service->minProcs;
903 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
904 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
905 for (i = 0; i < nProcs; i++) {
906 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
912 /* This routine is only required on Windows */
914 rx_StartClientThread(void)
916 #ifdef AFS_PTHREAD_ENV
918 pid = pthread_self();
919 #endif /* AFS_PTHREAD_ENV */
921 #endif /* AFS_NT40_ENV */
923 /* This routine must be called if any services are exported. If the
924 * donateMe flag is set, the calling process is donated to the server
927 rx_StartServer(int donateMe)
929 struct rx_service *service;
935 /* Start server processes, if necessary (exact function is dependent
936 * on the implementation environment--kernel or user space). DonateMe
937 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
938 * case, one less new proc will be created rx_StartServerProcs.
940 rxi_StartServerProcs(donateMe);
942 /* count up the # of threads in minProcs, and add set the min deficit to
943 * be that value, too.
945 for (i = 0; i < RX_MAX_SERVICES; i++) {
946 service = rx_services[i];
947 if (service == (struct rx_service *)0)
949 MUTEX_ENTER(&rx_quota_mutex);
950 rxi_totalMin += service->minProcs;
951 /* below works even if a thread is running, since minDeficit would
952 * still have been decremented and later re-incremented.
954 rxi_minDeficit += service->minProcs;
955 MUTEX_EXIT(&rx_quota_mutex);
958 /* Turn on reaping of idle server connections */
959 rxi_ReapConnections(NULL, NULL, NULL, 0);
968 #ifdef AFS_PTHREAD_ENV
970 pid = afs_pointer_to_int(pthread_self());
971 #else /* AFS_PTHREAD_ENV */
973 LWP_CurrentProcess(&pid);
974 #endif /* AFS_PTHREAD_ENV */
976 sprintf(name, "srv_%d", ++nProcs);
978 (*registerProgram) (pid, name);
980 #endif /* AFS_NT40_ENV */
981 rx_ServerProc(NULL); /* Never returns */
983 #ifdef RX_ENABLE_TSFPQ
984 /* no use leaving packets around in this thread's local queue if
985 * it isn't getting donated to the server thread pool.
987 rxi_FlushLocalPacketsTSFPQ();
988 #endif /* RX_ENABLE_TSFPQ */
992 /* Create a new client connection to the specified service, using the
993 * specified security object to implement the security model for this
995 struct rx_connection *
996 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
997 struct rx_securityClass *securityObject,
998 int serviceSecurityIndex)
1002 struct rx_connection *conn;
1007 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
1008 "serviceSecurityIndex %d)\n",
1009 ntohl(shost), ntohs(sport), sservice, securityObject,
1010 serviceSecurityIndex));
1012 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
1013 * the case of kmem_alloc? */
1014 conn = rxi_AllocConnection();
1015 #ifdef RX_ENABLE_LOCKS
1016 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
1017 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
1018 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
1021 MUTEX_ENTER(&rx_connHashTable_lock);
1022 cid = (rx_nextCid += RX_MAXCALLS);
1023 conn->type = RX_CLIENT_CONNECTION;
1025 conn->epoch = rx_epoch;
1026 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
1027 conn->serviceId = sservice;
1028 conn->securityObject = securityObject;
1029 conn->securityData = (void *) 0;
1030 conn->securityIndex = serviceSecurityIndex;
1031 rx_SetConnDeadTime(conn, rx_connDeadTime);
1032 rx_SetConnSecondsUntilNatPing(conn, 0);
1033 conn->ackRate = RX_FAST_ACK_RATE;
1034 conn->nSpecific = 0;
1035 conn->specific = NULL;
1036 conn->challengeEvent = NULL;
1037 conn->delayedAbortEvent = NULL;
1038 conn->abortCount = 0;
1040 for (i = 0; i < RX_MAXCALLS; i++) {
1041 conn->twind[i] = rx_initSendWindow;
1042 conn->rwind[i] = rx_initReceiveWindow;
1043 conn->lastBusy[i] = 0;
1046 RXS_NewConnection(securityObject, conn);
1048 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1050 conn->refCount++; /* no lock required since only this thread knows... */
1051 conn->next = rx_connHashTable[hashindex];
1052 rx_connHashTable[hashindex] = conn;
1053 if (rx_stats_active)
1054 rx_atomic_inc(&rx_stats.nClientConns);
1055 MUTEX_EXIT(&rx_connHashTable_lock);
1061 * Ensure a connection's timeout values are valid.
1063 * @param[in] conn The connection to check
1065 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1066 * unless idleDeadTime and/or hardDeadTime are not set
1070 rxi_CheckConnTimeouts(struct rx_connection *conn)
1072 /* a connection's timeouts must have the relationship
1073 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1074 * total loss of network to a peer may cause an idle timeout instead of a
1075 * dead timeout, simply because the idle timeout gets hit first. Also set
1076 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1077 /* this logic is slightly complicated by the fact that
1078 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1080 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1081 if (conn->idleDeadTime) {
1082 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1084 if (conn->hardDeadTime) {
1085 if (conn->idleDeadTime) {
1086 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1088 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1094 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1096 /* The idea is to set the dead time to a value that allows several
1097 * keepalives to be dropped without timing out the connection. */
1098 conn->secondsUntilDead = seconds;
1099 rxi_CheckConnTimeouts(conn);
1100 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1104 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1106 conn->hardDeadTime = seconds;
1107 rxi_CheckConnTimeouts(conn);
1111 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1113 conn->idleDeadTime = seconds;
1114 conn->idleDeadDetection = (seconds ? 1 : 0);
1115 rxi_CheckConnTimeouts(conn);
1118 int rxi_lowPeerRefCount = 0;
1119 int rxi_lowConnRefCount = 0;
1122 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1123 * NOTE: must not be called with rx_connHashTable_lock held.
1126 rxi_CleanupConnection(struct rx_connection *conn)
1128 /* Notify the service exporter, if requested, that this connection
1129 * is being destroyed */
1130 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1131 (*conn->service->destroyConnProc) (conn);
1133 /* Notify the security module that this connection is being destroyed */
1134 RXS_DestroyConnection(conn->securityObject, conn);
1136 /* If this is the last connection using the rx_peer struct, set its
1137 * idle time to now. rxi_ReapConnections will reap it if it's still
1138 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1140 MUTEX_ENTER(&rx_peerHashTable_lock);
1141 if (conn->peer->refCount < 2) {
1142 conn->peer->idleWhen = clock_Sec();
1143 if (conn->peer->refCount < 1) {
1144 conn->peer->refCount = 1;
1145 if (rx_stats_active) {
1146 MUTEX_ENTER(&rx_stats_mutex);
1147 rxi_lowPeerRefCount++;
1148 MUTEX_EXIT(&rx_stats_mutex);
1152 conn->peer->refCount--;
1153 MUTEX_EXIT(&rx_peerHashTable_lock);
1155 if (rx_stats_active)
1157 if (conn->type == RX_SERVER_CONNECTION)
1158 rx_atomic_dec(&rx_stats.nServerConns);
1160 rx_atomic_dec(&rx_stats.nClientConns);
1163 if (conn->specific) {
1165 for (i = 0; i < conn->nSpecific; i++) {
1166 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1167 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1168 conn->specific[i] = NULL;
1170 free(conn->specific);
1172 conn->specific = NULL;
1173 conn->nSpecific = 0;
1174 #endif /* !KERNEL */
1176 MUTEX_DESTROY(&conn->conn_call_lock);
1177 MUTEX_DESTROY(&conn->conn_data_lock);
1178 CV_DESTROY(&conn->conn_call_cv);
1180 rxi_FreeConnection(conn);
1183 /* Destroy the specified connection */
1185 rxi_DestroyConnection(struct rx_connection *conn)
1187 MUTEX_ENTER(&rx_connHashTable_lock);
1188 rxi_DestroyConnectionNoLock(conn);
1189 /* conn should be at the head of the cleanup list */
1190 if (conn == rx_connCleanup_list) {
1191 rx_connCleanup_list = rx_connCleanup_list->next;
1192 MUTEX_EXIT(&rx_connHashTable_lock);
1193 rxi_CleanupConnection(conn);
1195 #ifdef RX_ENABLE_LOCKS
1197 MUTEX_EXIT(&rx_connHashTable_lock);
1199 #endif /* RX_ENABLE_LOCKS */
1203 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1205 struct rx_connection **conn_ptr;
1207 struct rx_packet *packet;
1214 MUTEX_ENTER(&conn->conn_data_lock);
1215 MUTEX_ENTER(&rx_refcnt_mutex);
1216 if (conn->refCount > 0)
1219 if (rx_stats_active) {
1220 MUTEX_ENTER(&rx_stats_mutex);
1221 rxi_lowConnRefCount++;
1222 MUTEX_EXIT(&rx_stats_mutex);
1226 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1227 /* Busy; wait till the last guy before proceeding */
1228 MUTEX_EXIT(&rx_refcnt_mutex);
1229 MUTEX_EXIT(&conn->conn_data_lock);
1234 /* If the client previously called rx_NewCall, but it is still
1235 * waiting, treat this as a running call, and wait to destroy the
1236 * connection later when the call completes. */
1237 if ((conn->type == RX_CLIENT_CONNECTION)
1238 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1239 conn->flags |= RX_CONN_DESTROY_ME;
1240 MUTEX_EXIT(&conn->conn_data_lock);
1244 MUTEX_EXIT(&rx_refcnt_mutex);
1245 MUTEX_EXIT(&conn->conn_data_lock);
1247 /* Check for extant references to this connection */
1248 MUTEX_ENTER(&conn->conn_call_lock);
1249 for (i = 0; i < RX_MAXCALLS; i++) {
1250 struct rx_call *call = conn->call[i];
1253 if (conn->type == RX_CLIENT_CONNECTION) {
1254 MUTEX_ENTER(&call->lock);
1255 if (call->delayedAckEvent) {
1256 /* Push the final acknowledgment out now--there
1257 * won't be a subsequent call to acknowledge the
1258 * last reply packets */
1259 rxevent_Cancel(&call->delayedAckEvent, call,
1260 RX_CALL_REFCOUNT_DELAY);
1261 if (call->state == RX_STATE_PRECALL
1262 || call->state == RX_STATE_ACTIVE) {
1263 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1265 rxi_AckAll(NULL, call, 0);
1268 MUTEX_EXIT(&call->lock);
1272 MUTEX_EXIT(&conn->conn_call_lock);
1274 #ifdef RX_ENABLE_LOCKS
1276 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1277 MUTEX_EXIT(&conn->conn_data_lock);
1279 /* Someone is accessing a packet right now. */
1283 #endif /* RX_ENABLE_LOCKS */
1286 /* Don't destroy the connection if there are any call
1287 * structures still in use */
1288 MUTEX_ENTER(&conn->conn_data_lock);
1289 conn->flags |= RX_CONN_DESTROY_ME;
1290 MUTEX_EXIT(&conn->conn_data_lock);
1295 if (conn->natKeepAliveEvent) {
1296 rxi_NatKeepAliveOff(conn);
1299 if (conn->delayedAbortEvent) {
1300 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
1301 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1303 MUTEX_ENTER(&conn->conn_data_lock);
1304 rxi_SendConnectionAbort(conn, packet, 0, 1);
1305 MUTEX_EXIT(&conn->conn_data_lock);
1306 rxi_FreePacket(packet);
1310 /* Remove from connection hash table before proceeding */
1312 &rx_connHashTable[CONN_HASH
1313 (peer->host, peer->port, conn->cid, conn->epoch,
1315 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1316 if (*conn_ptr == conn) {
1317 *conn_ptr = conn->next;
1321 /* if the conn that we are destroying was the last connection, then we
1322 * clear rxLastConn as well */
1323 if (rxLastConn == conn)
1326 /* Make sure the connection is completely reset before deleting it. */
1327 /* get rid of pending events that could zap us later */
1328 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
1329 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
1330 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
1332 /* Add the connection to the list of destroyed connections that
1333 * need to be cleaned up. This is necessary to avoid deadlocks
1334 * in the routines we call to inform others that this connection is
1335 * being destroyed. */
1336 conn->next = rx_connCleanup_list;
1337 rx_connCleanup_list = conn;
1340 /* Externally available version */
1342 rx_DestroyConnection(struct rx_connection *conn)
1347 rxi_DestroyConnection(conn);
1352 rx_GetConnection(struct rx_connection *conn)
1357 MUTEX_ENTER(&rx_refcnt_mutex);
1359 MUTEX_EXIT(&rx_refcnt_mutex);
1363 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1364 /* Wait for the transmit queue to no longer be busy.
1365 * requires the call->lock to be held */
1367 rxi_WaitforTQBusy(struct rx_call *call) {
1368 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1369 call->flags |= RX_CALL_TQ_WAIT;
1371 #ifdef RX_ENABLE_LOCKS
1372 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1373 CV_WAIT(&call->cv_tq, &call->lock);
1374 #else /* RX_ENABLE_LOCKS */
1375 osi_rxSleep(&call->tq);
1376 #endif /* RX_ENABLE_LOCKS */
1378 if (call->tqWaiters == 0) {
1379 call->flags &= ~RX_CALL_TQ_WAIT;
1386 rxi_WakeUpTransmitQueue(struct rx_call *call)
1388 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1389 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1390 call, call->tqWaiters, call->flags));
1391 #ifdef RX_ENABLE_LOCKS
1392 osirx_AssertMine(&call->lock, "rxi_Start start");
1393 CV_BROADCAST(&call->cv_tq);
1394 #else /* RX_ENABLE_LOCKS */
1395 osi_rxWakeup(&call->tq);
1396 #endif /* RX_ENABLE_LOCKS */
1400 /* Start a new rx remote procedure call, on the specified connection.
1401 * If wait is set to 1, wait for a free call channel; otherwise return
1402 * 0. Maxtime gives the maximum number of seconds this call may take,
1403 * after rx_NewCall returns. After this time interval, a call to any
1404 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1405 * For fine grain locking, we hold the conn_call_lock in order to
1406 * to ensure that we don't get signalle after we found a call in an active
1407 * state and before we go to sleep.
1410 rx_NewCall(struct rx_connection *conn)
1412 int i, wait, ignoreBusy = 1;
1413 struct rx_call *call;
1414 struct clock queueTime;
1415 afs_uint32 leastBusy = 0;
1419 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1422 clock_GetTime(&queueTime);
1424 * Check if there are others waiting for a new call.
1425 * If so, let them go first to avoid starving them.
1426 * This is a fairly simple scheme, and might not be
1427 * a complete solution for large numbers of waiters.
1429 * makeCallWaiters keeps track of the number of
1430 * threads waiting to make calls and the
1431 * RX_CONN_MAKECALL_WAITING flag bit is used to
1432 * indicate that there are indeed calls waiting.
1433 * The flag is set when the waiter is incremented.
1434 * It is only cleared when makeCallWaiters is 0.
1435 * This prevents us from accidently destroying the
1436 * connection while it is potentially about to be used.
1438 MUTEX_ENTER(&conn->conn_call_lock);
1439 MUTEX_ENTER(&conn->conn_data_lock);
1440 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1441 conn->flags |= RX_CONN_MAKECALL_WAITING;
1442 conn->makeCallWaiters++;
1443 MUTEX_EXIT(&conn->conn_data_lock);
1445 #ifdef RX_ENABLE_LOCKS
1446 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1450 MUTEX_ENTER(&conn->conn_data_lock);
1451 conn->makeCallWaiters--;
1452 if (conn->makeCallWaiters == 0)
1453 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1456 /* We are now the active thread in rx_NewCall */
1457 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1458 MUTEX_EXIT(&conn->conn_data_lock);
1463 for (i = 0; i < RX_MAXCALLS; i++) {
1464 call = conn->call[i];
1466 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1467 /* we're not ignoring busy call slots; only look at the
1468 * call slot that is the "least" busy */
1472 if (call->state == RX_STATE_DALLY) {
1473 MUTEX_ENTER(&call->lock);
1474 if (call->state == RX_STATE_DALLY) {
1475 if (ignoreBusy && conn->lastBusy[i]) {
1476 /* if we're ignoring busy call slots, skip any ones that
1477 * have lastBusy set */
1478 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1479 leastBusy = conn->lastBusy[i];
1481 MUTEX_EXIT(&call->lock);
1486 * We are setting the state to RX_STATE_RESET to
1487 * ensure that no one else will attempt to use this
1488 * call once we drop the conn->conn_call_lock and
1489 * call->lock. We must drop the conn->conn_call_lock
1490 * before calling rxi_ResetCall because the process
1491 * of clearing the transmit queue can block for an
1492 * extended period of time. If we block while holding
1493 * the conn->conn_call_lock, then all rx_EndCall
1494 * processing will block as well. This has a detrimental
1495 * effect on overall system performance.
1497 call->state = RX_STATE_RESET;
1498 MUTEX_EXIT(&conn->conn_call_lock);
1499 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1500 rxi_ResetCall(call, 0);
1501 (*call->callNumber)++;
1502 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1506 * If we failed to be able to safely obtain the
1507 * conn->conn_call_lock we will have to drop the
1508 * call->lock to avoid a deadlock. When the call->lock
1509 * is released the state of the call can change. If it
1510 * is no longer RX_STATE_RESET then some other thread is
1513 MUTEX_EXIT(&call->lock);
1514 MUTEX_ENTER(&conn->conn_call_lock);
1515 MUTEX_ENTER(&call->lock);
1517 if (call->state == RX_STATE_RESET)
1521 * If we get here it means that after dropping
1522 * the conn->conn_call_lock and call->lock that
1523 * the call is no longer ours. If we can't find
1524 * a free call in the remaining slots we should
1525 * not go immediately to RX_CONN_MAKECALL_WAITING
1526 * because by dropping the conn->conn_call_lock
1527 * we have given up synchronization with rx_EndCall.
1528 * Instead, cycle through one more time to see if
1529 * we can find a call that can call our own.
1531 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1534 MUTEX_EXIT(&call->lock);
1537 if (ignoreBusy && conn->lastBusy[i]) {
1538 /* if we're ignoring busy call slots, skip any ones that
1539 * have lastBusy set */
1540 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1541 leastBusy = conn->lastBusy[i];
1546 /* rxi_NewCall returns with mutex locked */
1547 call = rxi_NewCall(conn, i);
1548 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1552 if (i < RX_MAXCALLS) {
1553 conn->lastBusy[i] = 0;
1554 call->flags &= ~RX_CALL_PEER_BUSY;
1559 if (leastBusy && ignoreBusy) {
1560 /* we didn't find a useable call slot, but we did see at least one
1561 * 'busy' slot; look again and only use a slot with the 'least
1567 MUTEX_ENTER(&conn->conn_data_lock);
1568 conn->flags |= RX_CONN_MAKECALL_WAITING;
1569 conn->makeCallWaiters++;
1570 MUTEX_EXIT(&conn->conn_data_lock);
1572 #ifdef RX_ENABLE_LOCKS
1573 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1577 MUTEX_ENTER(&conn->conn_data_lock);
1578 conn->makeCallWaiters--;
1579 if (conn->makeCallWaiters == 0)
1580 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1581 MUTEX_EXIT(&conn->conn_data_lock);
1583 /* Client is initially in send mode */
1584 call->state = RX_STATE_ACTIVE;
1585 call->error = conn->error;
1587 call->mode = RX_MODE_ERROR;
1589 call->mode = RX_MODE_SENDING;
1591 /* remember start time for call in case we have hard dead time limit */
1592 call->queueTime = queueTime;
1593 clock_GetTime(&call->startTime);
1594 hzero(call->bytesSent);
1595 hzero(call->bytesRcvd);
1597 /* Turn on busy protocol. */
1598 rxi_KeepAliveOn(call);
1600 /* Attempt MTU discovery */
1601 rxi_GrowMTUOn(call);
1604 * We are no longer the active thread in rx_NewCall
1606 MUTEX_ENTER(&conn->conn_data_lock);
1607 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1608 MUTEX_EXIT(&conn->conn_data_lock);
1611 * Wake up anyone else who might be giving us a chance to
1612 * run (see code above that avoids resource starvation).
1614 #ifdef RX_ENABLE_LOCKS
1615 CV_BROADCAST(&conn->conn_call_cv);
1619 MUTEX_EXIT(&conn->conn_call_lock);
1621 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1622 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1623 osi_Panic("rx_NewCall call about to be used without an empty tq");
1625 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1627 MUTEX_EXIT(&call->lock);
1630 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1635 rxi_HasActiveCalls(struct rx_connection *aconn)
1638 struct rx_call *tcall;
1642 for (i = 0; i < RX_MAXCALLS; i++) {
1643 if ((tcall = aconn->call[i])) {
1644 if ((tcall->state == RX_STATE_ACTIVE)
1645 || (tcall->state == RX_STATE_PRECALL)) {
1656 rxi_GetCallNumberVector(struct rx_connection *aconn,
1657 afs_int32 * aint32s)
1660 struct rx_call *tcall;
1664 for (i = 0; i < RX_MAXCALLS; i++) {
1665 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1666 aint32s[i] = aconn->callNumber[i] + 1;
1668 aint32s[i] = aconn->callNumber[i];
1675 rxi_SetCallNumberVector(struct rx_connection *aconn,
1676 afs_int32 * aint32s)
1679 struct rx_call *tcall;
1683 for (i = 0; i < RX_MAXCALLS; i++) {
1684 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1685 aconn->callNumber[i] = aint32s[i] - 1;
1687 aconn->callNumber[i] = aint32s[i];
1693 /* Advertise a new service. A service is named locally by a UDP port
1694 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1697 char *serviceName; Name for identification purposes (e.g. the
1698 service name might be used for probing for
1701 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1702 char *serviceName, struct rx_securityClass **securityObjects,
1703 int nSecurityObjects,
1704 afs_int32(*serviceProc) (struct rx_call * acall))
1706 osi_socket socket = OSI_NULLSOCKET;
1707 struct rx_service *tservice;
1713 if (serviceId == 0) {
1715 "rx_NewService: service id for service %s is not non-zero.\n",
1722 "rx_NewService: A non-zero port must be specified on this call if a non-zero port was not provided at Rx initialization (service %s).\n",
1730 tservice = rxi_AllocService();
1733 #ifdef RX_ENABLE_LOCKS
1734 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1737 for (i = 0; i < RX_MAX_SERVICES; i++) {
1738 struct rx_service *service = rx_services[i];
1740 if (port == service->servicePort && host == service->serviceHost) {
1741 if (service->serviceId == serviceId) {
1742 /* The identical service has already been
1743 * installed; if the caller was intending to
1744 * change the security classes used by this
1745 * service, he/she loses. */
1747 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1748 serviceName, serviceId, service->serviceName);
1750 rxi_FreeService(tservice);
1753 /* Different service, same port: re-use the socket
1754 * which is bound to the same port */
1755 socket = service->socket;
1758 if (socket == OSI_NULLSOCKET) {
1759 /* If we don't already have a socket (from another
1760 * service on same port) get a new one */
1761 socket = rxi_GetHostUDPSocket(host, port);
1762 if (socket == OSI_NULLSOCKET) {
1764 rxi_FreeService(tservice);
1769 service->socket = socket;
1770 service->serviceHost = host;
1771 service->servicePort = port;
1772 service->serviceId = serviceId;
1773 service->serviceName = serviceName;
1774 service->nSecurityObjects = nSecurityObjects;
1775 service->securityObjects = securityObjects;
1776 service->minProcs = 0;
1777 service->maxProcs = 1;
1778 service->idleDeadTime = 60;
1779 service->idleDeadErr = 0;
1780 service->connDeadTime = rx_connDeadTime;
1781 service->executeRequestProc = serviceProc;
1782 service->checkReach = 0;
1783 service->nSpecific = 0;
1784 service->specific = NULL;
1785 rx_services[i] = service; /* not visible until now */
1791 rxi_FreeService(tservice);
1792 (osi_Msg "rx_NewService: cannot support > %d services\n",
1797 /* Set configuration options for all of a service's security objects */
1800 rx_SetSecurityConfiguration(struct rx_service *service,
1801 rx_securityConfigVariables type,
1805 for (i = 0; i<service->nSecurityObjects; i++) {
1806 if (service->securityObjects[i]) {
1807 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1815 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1816 struct rx_securityClass **securityObjects, int nSecurityObjects,
1817 afs_int32(*serviceProc) (struct rx_call * acall))
1819 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1822 /* Generic request processing loop. This routine should be called
1823 * by the implementation dependent rx_ServerProc. If socketp is
1824 * non-null, it will be set to the file descriptor that this thread
1825 * is now listening on. If socketp is null, this routine will never
1828 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1830 struct rx_call *call;
1832 struct rx_service *tservice = NULL;
1839 call = rx_GetCall(threadID, tservice, socketp);
1840 if (socketp && *socketp != OSI_NULLSOCKET) {
1841 /* We are now a listener thread */
1847 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1848 #ifdef RX_ENABLE_LOCKS
1850 #endif /* RX_ENABLE_LOCKS */
1851 afs_termState = AFSOP_STOP_AFS;
1852 afs_osi_Wakeup(&afs_termState);
1853 #ifdef RX_ENABLE_LOCKS
1855 #endif /* RX_ENABLE_LOCKS */
1860 /* if server is restarting( typically smooth shutdown) then do not
1861 * allow any new calls.
1864 if (rx_tranquil && (call != NULL)) {
1868 MUTEX_ENTER(&call->lock);
1870 rxi_CallError(call, RX_RESTARTING);
1871 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1873 MUTEX_EXIT(&call->lock);
1878 tservice = call->conn->service;
1880 if (tservice->beforeProc)
1881 (*tservice->beforeProc) (call);
1883 code = tservice->executeRequestProc(call);
1885 if (tservice->afterProc)
1886 (*tservice->afterProc) (call, code);
1888 rx_EndCall(call, code);
1890 if (tservice->postProc)
1891 (*tservice->postProc) (code);
1893 if (rx_stats_active) {
1894 MUTEX_ENTER(&rx_stats_mutex);
1896 MUTEX_EXIT(&rx_stats_mutex);
1903 rx_WakeupServerProcs(void)
1905 struct rx_serverQueueEntry *np, *tqp;
1909 MUTEX_ENTER(&rx_serverPool_lock);
1911 #ifdef RX_ENABLE_LOCKS
1912 if (rx_waitForPacket)
1913 CV_BROADCAST(&rx_waitForPacket->cv);
1914 #else /* RX_ENABLE_LOCKS */
1915 if (rx_waitForPacket)
1916 osi_rxWakeup(rx_waitForPacket);
1917 #endif /* RX_ENABLE_LOCKS */
1918 MUTEX_ENTER(&freeSQEList_lock);
1919 for (np = rx_FreeSQEList; np; np = tqp) {
1920 tqp = *(struct rx_serverQueueEntry **)np;
1921 #ifdef RX_ENABLE_LOCKS
1922 CV_BROADCAST(&np->cv);
1923 #else /* RX_ENABLE_LOCKS */
1925 #endif /* RX_ENABLE_LOCKS */
1927 MUTEX_EXIT(&freeSQEList_lock);
1928 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1929 #ifdef RX_ENABLE_LOCKS
1930 CV_BROADCAST(&np->cv);
1931 #else /* RX_ENABLE_LOCKS */
1933 #endif /* RX_ENABLE_LOCKS */
1935 MUTEX_EXIT(&rx_serverPool_lock);
1940 * One thing that seems to happen is that all the server threads get
1941 * tied up on some empty or slow call, and then a whole bunch of calls
1942 * arrive at once, using up the packet pool, so now there are more
1943 * empty calls. The most critical resources here are server threads
1944 * and the free packet pool. The "doreclaim" code seems to help in
1945 * general. I think that eventually we arrive in this state: there
1946 * are lots of pending calls which do have all their packets present,
1947 * so they won't be reclaimed, are multi-packet calls, so they won't
1948 * be scheduled until later, and thus are tying up most of the free
1949 * packet pool for a very long time.
1951 * 1. schedule multi-packet calls if all the packets are present.
1952 * Probably CPU-bound operation, useful to return packets to pool.
1953 * Do what if there is a full window, but the last packet isn't here?
1954 * 3. preserve one thread which *only* runs "best" calls, otherwise
1955 * it sleeps and waits for that type of call.
1956 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1957 * the current dataquota business is badly broken. The quota isn't adjusted
1958 * to reflect how many packets are presently queued for a running call.
1959 * So, when we schedule a queued call with a full window of packets queued
1960 * up for it, that *should* free up a window full of packets for other 2d-class
1961 * calls to be able to use from the packet pool. But it doesn't.
1963 * NB. Most of the time, this code doesn't run -- since idle server threads
1964 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1965 * as a new call arrives.
1967 /* Sleep until a call arrives. Returns a pointer to the call, ready
1968 * for an rx_Read. */
1969 #ifdef RX_ENABLE_LOCKS
1971 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1973 struct rx_serverQueueEntry *sq;
1974 struct rx_call *call = (struct rx_call *)0;
1975 struct rx_service *service = NULL;
1977 MUTEX_ENTER(&freeSQEList_lock);
1979 if ((sq = rx_FreeSQEList)) {
1980 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1981 MUTEX_EXIT(&freeSQEList_lock);
1982 } else { /* otherwise allocate a new one and return that */
1983 MUTEX_EXIT(&freeSQEList_lock);
1984 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1985 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1986 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1989 MUTEX_ENTER(&rx_serverPool_lock);
1990 if (cur_service != NULL) {
1991 ReturnToServerPool(cur_service);
1994 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1995 struct rx_call *tcall, *ncall, *choice2 = NULL;
1997 /* Scan for eligible incoming calls. A call is not eligible
1998 * if the maximum number of calls for its service type are
1999 * already executing */
2000 /* One thread will process calls FCFS (to prevent starvation),
2001 * while the other threads may run ahead looking for calls which
2002 * have all their input data available immediately. This helps
2003 * keep threads from blocking, waiting for data from the client. */
2004 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
2005 service = tcall->conn->service;
2006 if (!QuotaOK(service)) {
2009 MUTEX_ENTER(&rx_pthread_mutex);
2010 if (tno == rxi_fcfs_thread_num
2011 || queue_IsLast(&rx_incomingCallQueue, tcall)) {
2012 MUTEX_EXIT(&rx_pthread_mutex);
2013 /* If we're the fcfs thread , then we'll just use
2014 * this call. If we haven't been able to find an optimal
2015 * choice, and we're at the end of the list, then use a
2016 * 2d choice if one has been identified. Otherwise... */
2017 call = (choice2 ? choice2 : tcall);
2018 service = call->conn->service;
2020 MUTEX_EXIT(&rx_pthread_mutex);
2021 if (!queue_IsEmpty(&tcall->rq)) {
2022 struct rx_packet *rp;
2023 rp = queue_First(&tcall->rq, rx_packet);
2024 if (rp->header.seq == 1) {
2026 || (rp->header.flags & RX_LAST_PACKET)) {
2028 } else if (rxi_2dchoice && !choice2
2029 && !(tcall->flags & RX_CALL_CLEARED)
2030 && (tcall->rprev > rxi_HardAckRate)) {
2040 ReturnToServerPool(service);
2047 MUTEX_EXIT(&rx_serverPool_lock);
2048 MUTEX_ENTER(&call->lock);
2050 if (call->flags & RX_CALL_WAIT_PROC) {
2051 call->flags &= ~RX_CALL_WAIT_PROC;
2052 rx_atomic_dec(&rx_nWaiting);
2055 if (call->state != RX_STATE_PRECALL || call->error) {
2056 MUTEX_EXIT(&call->lock);
2057 MUTEX_ENTER(&rx_serverPool_lock);
2058 ReturnToServerPool(service);
2063 if (queue_IsEmpty(&call->rq)
2064 || queue_First(&call->rq, rx_packet)->header.seq != 1)
2065 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2067 CLEAR_CALL_QUEUE_LOCK(call);
2070 /* If there are no eligible incoming calls, add this process
2071 * to the idle server queue, to wait for one */
2075 *socketp = OSI_NULLSOCKET;
2077 sq->socketp = socketp;
2078 queue_Append(&rx_idleServerQueue, sq);
2079 #ifndef AFS_AIX41_ENV
2080 rx_waitForPacket = sq;
2082 rx_waitingForPacket = sq;
2083 #endif /* AFS_AIX41_ENV */
2085 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2087 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2088 MUTEX_EXIT(&rx_serverPool_lock);
2089 return (struct rx_call *)0;
2092 } while (!(call = sq->newcall)
2093 && !(socketp && *socketp != OSI_NULLSOCKET));
2094 MUTEX_EXIT(&rx_serverPool_lock);
2096 MUTEX_ENTER(&call->lock);
2102 MUTEX_ENTER(&freeSQEList_lock);
2103 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2104 rx_FreeSQEList = sq;
2105 MUTEX_EXIT(&freeSQEList_lock);
2108 clock_GetTime(&call->startTime);
2109 call->state = RX_STATE_ACTIVE;
2110 call->mode = RX_MODE_RECEIVING;
2111 #ifdef RX_KERNEL_TRACE
2112 if (ICL_SETACTIVE(afs_iclSetp)) {
2113 int glockOwner = ISAFS_GLOCK();
2116 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2117 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2124 rxi_calltrace(RX_CALL_START, call);
2125 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2126 call->conn->service->servicePort, call->conn->service->serviceId,
2129 MUTEX_EXIT(&call->lock);
2130 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2132 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2137 #else /* RX_ENABLE_LOCKS */
2139 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2141 struct rx_serverQueueEntry *sq;
2142 struct rx_call *call = (struct rx_call *)0, *choice2;
2143 struct rx_service *service = NULL;
2147 MUTEX_ENTER(&freeSQEList_lock);
2149 if ((sq = rx_FreeSQEList)) {
2150 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2151 MUTEX_EXIT(&freeSQEList_lock);
2152 } else { /* otherwise allocate a new one and return that */
2153 MUTEX_EXIT(&freeSQEList_lock);
2154 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2155 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2156 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2158 MUTEX_ENTER(&sq->lock);
2160 if (cur_service != NULL) {
2161 cur_service->nRequestsRunning--;
2162 MUTEX_ENTER(&rx_quota_mutex);
2163 if (cur_service->nRequestsRunning < cur_service->minProcs)
2166 MUTEX_EXIT(&rx_quota_mutex);
2168 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
2169 struct rx_call *tcall, *ncall;
2170 /* Scan for eligible incoming calls. A call is not eligible
2171 * if the maximum number of calls for its service type are
2172 * already executing */
2173 /* One thread will process calls FCFS (to prevent starvation),
2174 * while the other threads may run ahead looking for calls which
2175 * have all their input data available immediately. This helps
2176 * keep threads from blocking, waiting for data from the client. */
2177 choice2 = (struct rx_call *)0;
2178 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
2179 service = tcall->conn->service;
2180 if (QuotaOK(service)) {
2181 MUTEX_ENTER(&rx_pthread_mutex);
2182 if (tno == rxi_fcfs_thread_num
2183 || !tcall->queue_item_header.next) {
2184 MUTEX_EXIT(&rx_pthread_mutex);
2185 /* If we're the fcfs thread, then we'll just use
2186 * this call. If we haven't been able to find an optimal
2187 * choice, and we're at the end of the list, then use a
2188 * 2d choice if one has been identified. Otherwise... */
2189 call = (choice2 ? choice2 : tcall);
2190 service = call->conn->service;
2192 MUTEX_EXIT(&rx_pthread_mutex);
2193 if (!queue_IsEmpty(&tcall->rq)) {
2194 struct rx_packet *rp;
2195 rp = queue_First(&tcall->rq, rx_packet);
2196 if (rp->header.seq == 1
2198 || (rp->header.flags & RX_LAST_PACKET))) {
2200 } else if (rxi_2dchoice && !choice2
2201 && !(tcall->flags & RX_CALL_CLEARED)
2202 && (tcall->rprev > rxi_HardAckRate)) {
2216 /* we can't schedule a call if there's no data!!! */
2217 /* send an ack if there's no data, if we're missing the
2218 * first packet, or we're missing something between first
2219 * and last -- there's a "hole" in the incoming data. */
2220 if (queue_IsEmpty(&call->rq)
2221 || queue_First(&call->rq, rx_packet)->header.seq != 1
2222 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
2223 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2225 call->flags &= (~RX_CALL_WAIT_PROC);
2226 service->nRequestsRunning++;
2227 /* just started call in minProcs pool, need fewer to maintain
2229 MUTEX_ENTER(&rx_quota_mutex);
2230 if (service->nRequestsRunning <= service->minProcs)
2233 MUTEX_EXIT(&rx_quota_mutex);
2234 rx_atomic_dec(&rx_nWaiting);
2235 /* MUTEX_EXIT(&call->lock); */
2237 /* If there are no eligible incoming calls, add this process
2238 * to the idle server queue, to wait for one */
2241 *socketp = OSI_NULLSOCKET;
2243 sq->socketp = socketp;
2244 queue_Append(&rx_idleServerQueue, sq);
2248 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2250 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2251 return (struct rx_call *)0;
2254 } while (!(call = sq->newcall)
2255 && !(socketp && *socketp != OSI_NULLSOCKET));
2257 MUTEX_EXIT(&sq->lock);
2259 MUTEX_ENTER(&freeSQEList_lock);
2260 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2261 rx_FreeSQEList = sq;
2262 MUTEX_EXIT(&freeSQEList_lock);
2265 clock_GetTime(&call->startTime);
2266 call->state = RX_STATE_ACTIVE;
2267 call->mode = RX_MODE_RECEIVING;
2268 #ifdef RX_KERNEL_TRACE
2269 if (ICL_SETACTIVE(afs_iclSetp)) {
2270 int glockOwner = ISAFS_GLOCK();
2273 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2274 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2281 rxi_calltrace(RX_CALL_START, call);
2282 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2283 call->conn->service->servicePort, call->conn->service->serviceId,
2286 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2293 #endif /* RX_ENABLE_LOCKS */
2297 /* Establish a procedure to be called when a packet arrives for a
2298 * call. This routine will be called at most once after each call,
2299 * and will also be called if there is an error condition on the or
2300 * the call is complete. Used by multi rx to build a selection
2301 * function which determines which of several calls is likely to be a
2302 * good one to read from.
2303 * NOTE: the way this is currently implemented it is probably only a
2304 * good idea to (1) use it immediately after a newcall (clients only)
2305 * and (2) only use it once. Other uses currently void your warranty
2308 rx_SetArrivalProc(struct rx_call *call,
2309 void (*proc) (struct rx_call * call,
2312 void * handle, int arg)
2314 call->arrivalProc = proc;
2315 call->arrivalProcHandle = handle;
2316 call->arrivalProcArg = arg;
2319 /* Call is finished (possibly prematurely). Return rc to the peer, if
2320 * appropriate, and return the final error code from the conversation
2324 rx_EndCall(struct rx_call *call, afs_int32 rc)
2326 struct rx_connection *conn = call->conn;
2330 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2331 call, rc, call->error, call->abortCode));
2334 MUTEX_ENTER(&call->lock);
2336 if (rc == 0 && call->error == 0) {
2337 call->abortCode = 0;
2338 call->abortCount = 0;
2341 call->arrivalProc = (void (*)())0;
2342 if (rc && call->error == 0) {
2343 rxi_CallError(call, rc);
2344 call->mode = RX_MODE_ERROR;
2345 /* Send an abort message to the peer if this error code has
2346 * only just been set. If it was set previously, assume the
2347 * peer has already been sent the error code or will request it
2349 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2351 if (conn->type == RX_SERVER_CONNECTION) {
2352 /* Make sure reply or at least dummy reply is sent */
2353 if (call->mode == RX_MODE_RECEIVING) {
2354 MUTEX_EXIT(&call->lock);
2355 rxi_WriteProc(call, 0, 0);
2356 MUTEX_ENTER(&call->lock);
2358 if (call->mode == RX_MODE_SENDING) {
2359 MUTEX_EXIT(&call->lock);
2360 rxi_FlushWrite(call);
2361 MUTEX_ENTER(&call->lock);
2363 rxi_calltrace(RX_CALL_END, call);
2364 /* Call goes to hold state until reply packets are acknowledged */
2365 if (call->tfirst + call->nSoftAcked < call->tnext) {
2366 call->state = RX_STATE_HOLD;
2368 call->state = RX_STATE_DALLY;
2369 rxi_ClearTransmitQueue(call, 0);
2370 rxi_rto_cancel(call);
2371 rxevent_Cancel(&call->keepAliveEvent, call,
2372 RX_CALL_REFCOUNT_ALIVE);
2374 } else { /* Client connection */
2376 /* Make sure server receives input packets, in the case where
2377 * no reply arguments are expected */
2378 if ((call->mode == RX_MODE_SENDING)
2379 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2380 MUTEX_EXIT(&call->lock);
2381 (void)rxi_ReadProc(call, &dummy, 1);
2382 MUTEX_ENTER(&call->lock);
2385 /* If we had an outstanding delayed ack, be nice to the server
2386 * and force-send it now.
2388 if (call->delayedAckEvent) {
2389 rxevent_Cancel(&call->delayedAckEvent, call,
2390 RX_CALL_REFCOUNT_DELAY);
2391 rxi_SendDelayedAck(NULL, call, NULL, 0);
2394 /* We need to release the call lock since it's lower than the
2395 * conn_call_lock and we don't want to hold the conn_call_lock
2396 * over the rx_ReadProc call. The conn_call_lock needs to be held
2397 * here for the case where rx_NewCall is perusing the calls on
2398 * the connection structure. We don't want to signal until
2399 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2400 * have checked this call, found it active and by the time it
2401 * goes to sleep, will have missed the signal.
2403 MUTEX_EXIT(&call->lock);
2404 MUTEX_ENTER(&conn->conn_call_lock);
2405 MUTEX_ENTER(&call->lock);
2407 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2408 conn->lastBusy[call->channel] = 0;
2411 MUTEX_ENTER(&conn->conn_data_lock);
2412 conn->flags |= RX_CONN_BUSY;
2413 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2414 MUTEX_EXIT(&conn->conn_data_lock);
2415 #ifdef RX_ENABLE_LOCKS
2416 CV_BROADCAST(&conn->conn_call_cv);
2421 #ifdef RX_ENABLE_LOCKS
2423 MUTEX_EXIT(&conn->conn_data_lock);
2425 #endif /* RX_ENABLE_LOCKS */
2426 call->state = RX_STATE_DALLY;
2428 error = call->error;
2430 /* currentPacket, nLeft, and NFree must be zeroed here, because
2431 * ResetCall cannot: ResetCall may be called at splnet(), in the
2432 * kernel version, and may interrupt the macros rx_Read or
2433 * rx_Write, which run at normal priority for efficiency. */
2434 if (call->currentPacket) {
2435 #ifdef RX_TRACK_PACKETS
2436 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2438 rxi_FreePacket(call->currentPacket);
2439 call->currentPacket = (struct rx_packet *)0;
2442 call->nLeft = call->nFree = call->curlen = 0;
2444 /* Free any packets from the last call to ReadvProc/WritevProc */
2445 #ifdef RXDEBUG_PACKET
2447 #endif /* RXDEBUG_PACKET */
2448 rxi_FreePackets(0, &call->iovq);
2449 MUTEX_EXIT(&call->lock);
2451 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2452 if (conn->type == RX_CLIENT_CONNECTION) {
2453 MUTEX_ENTER(&conn->conn_data_lock);
2454 conn->flags &= ~RX_CONN_BUSY;
2455 MUTEX_EXIT(&conn->conn_data_lock);
2456 MUTEX_EXIT(&conn->conn_call_lock);
2460 * Map errors to the local host's errno.h format.
2462 error = ntoh_syserr_conv(error);
2466 #if !defined(KERNEL)
2468 /* Call this routine when shutting down a server or client (especially
2469 * clients). This will allow Rx to gracefully garbage collect server
2470 * connections, and reduce the number of retries that a server might
2471 * make to a dead client.
2472 * This is not quite right, since some calls may still be ongoing and
2473 * we can't lock them to destroy them. */
2477 struct rx_connection **conn_ptr, **conn_end;
2481 if (rxinit_status == 1) {
2483 return; /* Already shutdown. */
2485 rxi_DeleteCachedConnections();
2486 if (rx_connHashTable) {
2487 MUTEX_ENTER(&rx_connHashTable_lock);
2488 for (conn_ptr = &rx_connHashTable[0], conn_end =
2489 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2491 struct rx_connection *conn, *next;
2492 for (conn = *conn_ptr; conn; conn = next) {
2494 if (conn->type == RX_CLIENT_CONNECTION) {
2495 MUTEX_ENTER(&rx_refcnt_mutex);
2497 MUTEX_EXIT(&rx_refcnt_mutex);
2498 #ifdef RX_ENABLE_LOCKS
2499 rxi_DestroyConnectionNoLock(conn);
2500 #else /* RX_ENABLE_LOCKS */
2501 rxi_DestroyConnection(conn);
2502 #endif /* RX_ENABLE_LOCKS */
2506 #ifdef RX_ENABLE_LOCKS
2507 while (rx_connCleanup_list) {
2508 struct rx_connection *conn;
2509 conn = rx_connCleanup_list;
2510 rx_connCleanup_list = rx_connCleanup_list->next;
2511 MUTEX_EXIT(&rx_connHashTable_lock);
2512 rxi_CleanupConnection(conn);
2513 MUTEX_ENTER(&rx_connHashTable_lock);
2515 MUTEX_EXIT(&rx_connHashTable_lock);
2516 #endif /* RX_ENABLE_LOCKS */
2521 afs_winsockCleanup();
2529 /* if we wakeup packet waiter too often, can get in loop with two
2530 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2532 rxi_PacketsUnWait(void)
2534 if (!rx_waitingForPackets) {
2538 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2539 return; /* still over quota */
2542 rx_waitingForPackets = 0;
2543 #ifdef RX_ENABLE_LOCKS
2544 CV_BROADCAST(&rx_waitingForPackets_cv);
2546 osi_rxWakeup(&rx_waitingForPackets);
2552 /* ------------------Internal interfaces------------------------- */
2554 /* Return this process's service structure for the
2555 * specified socket and service */
2556 static struct rx_service *
2557 rxi_FindService(osi_socket socket, u_short serviceId)
2559 struct rx_service **sp;
2560 for (sp = &rx_services[0]; *sp; sp++) {
2561 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2567 #ifdef RXDEBUG_PACKET
2568 #ifdef KDUMP_RX_LOCK
2569 static struct rx_call_rx_lock *rx_allCallsp = 0;
2571 static struct rx_call *rx_allCallsp = 0;
2573 #endif /* RXDEBUG_PACKET */
2575 /* Allocate a call structure, for the indicated channel of the
2576 * supplied connection. The mode and state of the call must be set by
2577 * the caller. Returns the call with mutex locked. */
2578 static struct rx_call *
2579 rxi_NewCall(struct rx_connection *conn, int channel)
2581 struct rx_call *call;
2582 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2583 struct rx_call *cp; /* Call pointer temp */
2584 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2585 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2587 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2589 /* Grab an existing call structure, or allocate a new one.
2590 * Existing call structures are assumed to have been left reset by
2592 MUTEX_ENTER(&rx_freeCallQueue_lock);
2594 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2596 * EXCEPT that the TQ might not yet be cleared out.
2597 * Skip over those with in-use TQs.
2600 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2601 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2607 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2608 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2609 call = queue_First(&rx_freeCallQueue, rx_call);
2610 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2612 if (rx_stats_active)
2613 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2614 MUTEX_EXIT(&rx_freeCallQueue_lock);
2615 MUTEX_ENTER(&call->lock);
2616 CLEAR_CALL_QUEUE_LOCK(call);
2617 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2618 /* Now, if TQ wasn't cleared earlier, do it now. */
2619 rxi_WaitforTQBusy(call);
2620 if (call->flags & RX_CALL_TQ_CLEARME) {
2621 rxi_ClearTransmitQueue(call, 1);
2622 /*queue_Init(&call->tq);*/
2624 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2625 /* Bind the call to its connection structure */
2627 rxi_ResetCall(call, 1);
2630 call = rxi_Alloc(sizeof(struct rx_call));
2631 #ifdef RXDEBUG_PACKET
2632 call->allNextp = rx_allCallsp;
2633 rx_allCallsp = call;
2635 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2636 #else /* RXDEBUG_PACKET */
2637 rx_atomic_inc(&rx_stats.nCallStructs);
2638 #endif /* RXDEBUG_PACKET */
2640 MUTEX_EXIT(&rx_freeCallQueue_lock);
2641 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2642 MUTEX_ENTER(&call->lock);
2643 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2644 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2645 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2647 /* Initialize once-only items */
2648 queue_Init(&call->tq);
2649 queue_Init(&call->rq);
2650 queue_Init(&call->iovq);
2651 #ifdef RXDEBUG_PACKET
2652 call->rqc = call->tqc = call->iovqc = 0;
2653 #endif /* RXDEBUG_PACKET */
2654 /* Bind the call to its connection structure (prereq for reset) */
2656 rxi_ResetCall(call, 1);
2658 call->channel = channel;
2659 call->callNumber = &conn->callNumber[channel];
2660 call->rwind = conn->rwind[channel];
2661 call->twind = conn->twind[channel];
2662 /* Note that the next expected call number is retained (in
2663 * conn->callNumber[i]), even if we reallocate the call structure
2665 conn->call[channel] = call;
2666 /* if the channel's never been used (== 0), we should start at 1, otherwise
2667 * the call number is valid from the last time this channel was used */
2668 if (*call->callNumber == 0)
2669 *call->callNumber = 1;
2674 /* A call has been inactive long enough that so we can throw away
2675 * state, including the call structure, which is placed on the call
2678 * call->lock amd rx_refcnt_mutex are held upon entry.
2679 * haveCTLock is set when called from rxi_ReapConnections.
2681 * return 1 if the call is freed, 0 if not.
2684 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2686 int channel = call->channel;
2687 struct rx_connection *conn = call->conn;
2688 u_char state = call->state;
2691 * We are setting the state to RX_STATE_RESET to
2692 * ensure that no one else will attempt to use this
2693 * call once we drop the refcnt lock. We must drop
2694 * the refcnt lock before calling rxi_ResetCall
2695 * because it cannot be held across acquiring the
2696 * freepktQ lock. NewCall does the same.
2698 call->state = RX_STATE_RESET;
2699 MUTEX_EXIT(&rx_refcnt_mutex);
2700 rxi_ResetCall(call, 0);
2702 if (MUTEX_TRYENTER(&conn->conn_call_lock))
2704 if (state == RX_STATE_DALLY || state == RX_STATE_HOLD)
2705 (*call->callNumber)++;
2707 if (call->conn->call[channel] == call)
2708 call->conn->call[channel] = 0;
2709 MUTEX_EXIT(&conn->conn_call_lock);
2712 * We couldn't obtain the conn_call_lock so we can't
2713 * disconnect the call from the connection. Set the
2714 * call state to dally so that the call can be reused.
2716 MUTEX_ENTER(&rx_refcnt_mutex);
2717 call->state = RX_STATE_DALLY;
2721 MUTEX_ENTER(&rx_freeCallQueue_lock);
2722 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2723 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2724 /* A call may be free even though its transmit queue is still in use.
2725 * Since we search the call list from head to tail, put busy calls at
2726 * the head of the list, and idle calls at the tail.
2728 if (call->flags & RX_CALL_TQ_BUSY)
2729 queue_Prepend(&rx_freeCallQueue, call);
2731 queue_Append(&rx_freeCallQueue, call);
2732 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2733 queue_Append(&rx_freeCallQueue, call);
2734 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2735 if (rx_stats_active)
2736 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2737 MUTEX_EXIT(&rx_freeCallQueue_lock);
2739 /* Destroy the connection if it was previously slated for
2740 * destruction, i.e. the Rx client code previously called
2741 * rx_DestroyConnection (client connections), or
2742 * rxi_ReapConnections called the same routine (server
2743 * connections). Only do this, however, if there are no
2744 * outstanding calls. Note that for fine grain locking, there appears
2745 * to be a deadlock in that rxi_FreeCall has a call locked and
2746 * DestroyConnectionNoLock locks each call in the conn. But note a
2747 * few lines up where we have removed this call from the conn.
2748 * If someone else destroys a connection, they either have no
2749 * call lock held or are going through this section of code.
2751 MUTEX_ENTER(&conn->conn_data_lock);
2752 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2753 MUTEX_ENTER(&rx_refcnt_mutex);
2755 MUTEX_EXIT(&rx_refcnt_mutex);
2756 MUTEX_EXIT(&conn->conn_data_lock);
2757 #ifdef RX_ENABLE_LOCKS
2759 rxi_DestroyConnectionNoLock(conn);
2761 rxi_DestroyConnection(conn);
2762 #else /* RX_ENABLE_LOCKS */
2763 rxi_DestroyConnection(conn);
2764 #endif /* RX_ENABLE_LOCKS */
2766 MUTEX_EXIT(&conn->conn_data_lock);
2768 MUTEX_ENTER(&rx_refcnt_mutex);
2772 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2773 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2776 rxi_Alloc(size_t size)
2780 if (rx_stats_active) {
2781 rx_atomic_add(&rxi_Allocsize, (int) size);
2782 rx_atomic_inc(&rxi_Alloccnt);
2786 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2787 afs_osi_Alloc_NoSleep(size);
2792 osi_Panic("rxi_Alloc error");
2798 rxi_Free(void *addr, size_t size)
2800 if (rx_stats_active) {
2801 rx_atomic_sub(&rxi_Allocsize, (int) size);
2802 rx_atomic_dec(&rxi_Alloccnt);
2804 osi_Free(addr, size);
2808 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2810 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2811 struct rx_peer *next = NULL;
2815 MUTEX_ENTER(&rx_peerHashTable_lock);
2817 peer_ptr = &rx_peerHashTable[0];
2818 peer_end = &rx_peerHashTable[rx_hashTableSize];
2821 for ( ; peer_ptr < peer_end; peer_ptr++) {
2824 for ( ; peer; peer = next) {
2826 if (host == peer->host)
2831 hashIndex = PEER_HASH(host, port);
2832 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2833 if ((peer->host == host) && (peer->port == port))
2838 MUTEX_ENTER(&rx_peerHashTable_lock);
2843 MUTEX_EXIT(&rx_peerHashTable_lock);
2845 MUTEX_ENTER(&peer->peer_lock);
2846 /* We don't handle dropping below min, so don't */
2847 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2848 peer->ifMTU=MIN(mtu, peer->ifMTU);
2849 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2850 /* if we tweaked this down, need to tune our peer MTU too */
2851 peer->MTU = MIN(peer->MTU, peer->natMTU);
2852 /* if we discovered a sub-1500 mtu, degrade */
2853 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2854 peer->maxDgramPackets = 1;
2855 /* We no longer have valid peer packet information */
2856 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2857 peer->maxPacketSize = 0;
2858 MUTEX_EXIT(&peer->peer_lock);
2860 MUTEX_ENTER(&rx_peerHashTable_lock);
2862 if (host && !port) {
2864 /* pick up where we left off */
2868 MUTEX_EXIT(&rx_peerHashTable_lock);
2871 /* Find the peer process represented by the supplied (host,port)
2872 * combination. If there is no appropriate active peer structure, a
2873 * new one will be allocated and initialized
2874 * The origPeer, if set, is a pointer to a peer structure on which the
2875 * refcount will be be decremented. This is used to replace the peer
2876 * structure hanging off a connection structure */
2878 rxi_FindPeer(afs_uint32 host, u_short port,
2879 struct rx_peer *origPeer, int create)
2883 hashIndex = PEER_HASH(host, port);
2884 MUTEX_ENTER(&rx_peerHashTable_lock);
2885 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2886 if ((pp->host == host) && (pp->port == port))
2891 pp = rxi_AllocPeer(); /* This bzero's *pp */
2892 pp->host = host; /* set here or in InitPeerParams is zero */
2894 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2895 queue_Init(&pp->congestionQueue);
2896 queue_Init(&pp->rpcStats);
2897 pp->next = rx_peerHashTable[hashIndex];
2898 rx_peerHashTable[hashIndex] = pp;
2899 rxi_InitPeerParams(pp);
2900 if (rx_stats_active)
2901 rx_atomic_inc(&rx_stats.nPeerStructs);
2908 origPeer->refCount--;
2909 MUTEX_EXIT(&rx_peerHashTable_lock);
2914 /* Find the connection at (host, port) started at epoch, and with the
2915 * given connection id. Creates the server connection if necessary.
2916 * The type specifies whether a client connection or a server
2917 * connection is desired. In both cases, (host, port) specify the
2918 * peer's (host, pair) pair. Client connections are not made
2919 * automatically by this routine. The parameter socket gives the
2920 * socket descriptor on which the packet was received. This is used,
2921 * in the case of server connections, to check that *new* connections
2922 * come via a valid (port, serviceId). Finally, the securityIndex
2923 * parameter must match the existing index for the connection. If a
2924 * server connection is created, it will be created using the supplied
2925 * index, if the index is valid for this service */
2926 struct rx_connection *
2927 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2928 u_short port, u_short serviceId, afs_uint32 cid,
2929 afs_uint32 epoch, int type, u_int securityIndex)
2931 int hashindex, flag, i;
2932 struct rx_connection *conn;
2933 hashindex = CONN_HASH(host, port, cid, epoch, type);
2934 MUTEX_ENTER(&rx_connHashTable_lock);
2935 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2936 rx_connHashTable[hashindex],
2939 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2940 && (epoch == conn->epoch)) {
2941 struct rx_peer *pp = conn->peer;
2942 if (securityIndex != conn->securityIndex) {
2943 /* this isn't supposed to happen, but someone could forge a packet
2944 * like this, and there seems to be some CM bug that makes this
2945 * happen from time to time -- in which case, the fileserver
2947 MUTEX_EXIT(&rx_connHashTable_lock);
2948 return (struct rx_connection *)0;
2950 if (pp->host == host && pp->port == port)
2952 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2954 /* So what happens when it's a callback connection? */
2955 if ( /*type == RX_CLIENT_CONNECTION && */
2956 (conn->epoch & 0x80000000))
2960 /* the connection rxLastConn that was used the last time is not the
2961 ** one we are looking for now. Hence, start searching in the hash */
2963 conn = rx_connHashTable[hashindex];
2968 struct rx_service *service;
2969 if (type == RX_CLIENT_CONNECTION) {
2970 MUTEX_EXIT(&rx_connHashTable_lock);
2971 return (struct rx_connection *)0;
2973 service = rxi_FindService(socket, serviceId);
2974 if (!service || (securityIndex >= service->nSecurityObjects)
2975 || (service->securityObjects[securityIndex] == 0)) {
2976 MUTEX_EXIT(&rx_connHashTable_lock);
2977 return (struct rx_connection *)0;
2979 conn = rxi_AllocConnection(); /* This bzero's the connection */
2980 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2981 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2982 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2983 conn->next = rx_connHashTable[hashindex];
2984 rx_connHashTable[hashindex] = conn;
2985 conn->peer = rxi_FindPeer(host, port, 0, 1);
2986 conn->type = RX_SERVER_CONNECTION;
2987 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2988 conn->epoch = epoch;
2989 conn->cid = cid & RX_CIDMASK;
2990 /* conn->serial = conn->lastSerial = 0; */
2991 /* conn->timeout = 0; */
2992 conn->ackRate = RX_FAST_ACK_RATE;
2993 conn->service = service;
2994 conn->serviceId = serviceId;
2995 conn->securityIndex = securityIndex;
2996 conn->securityObject = service->securityObjects[securityIndex];
2997 conn->nSpecific = 0;
2998 conn->specific = NULL;
2999 rx_SetConnDeadTime(conn, service->connDeadTime);
3000 conn->idleDeadTime = service->idleDeadTime;
3001 conn->idleDeadDetection = service->idleDeadErr ? 1 : 0;
3002 for (i = 0; i < RX_MAXCALLS; i++) {
3003 conn->twind[i] = rx_initSendWindow;
3004 conn->rwind[i] = rx_initReceiveWindow;
3006 /* Notify security object of the new connection */
3007 RXS_NewConnection(conn->securityObject, conn);
3008 /* XXXX Connection timeout? */
3009 if (service->newConnProc)
3010 (*service->newConnProc) (conn);
3011 if (rx_stats_active)
3012 rx_atomic_inc(&rx_stats.nServerConns);
3015 MUTEX_ENTER(&rx_refcnt_mutex);
3017 MUTEX_EXIT(&rx_refcnt_mutex);
3019 rxLastConn = conn; /* store this connection as the last conn used */
3020 MUTEX_EXIT(&rx_connHashTable_lock);
3025 * Timeout a call on a busy call channel if appropriate.
3027 * @param[in] call The busy call.
3029 * @pre 'call' is marked as busy (namely,
3030 * call->conn->lastBusy[call->channel] != 0)
3032 * @pre call->lock is held
3033 * @pre rxi_busyChannelError is nonzero
3035 * @note call->lock is dropped and reacquired
3038 rxi_CheckBusy(struct rx_call *call)
3040 struct rx_connection *conn = call->conn;
3041 int channel = call->channel;
3042 int freechannel = 0;
3044 afs_uint32 callNumber = *call->callNumber;
3046 MUTEX_EXIT(&call->lock);
3048 MUTEX_ENTER(&conn->conn_call_lock);
3050 /* Are there any other call slots on this conn that we should try? Look for
3051 * slots that are empty and are either non-busy, or were marked as busy
3052 * longer than conn->secondsUntilDead seconds before this call started. */
3054 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3056 /* only look at channels that aren't us */
3060 if (conn->lastBusy[i]) {
3061 /* if this channel looked busy too recently, don't look at it */
3062 if (conn->lastBusy[i] >= call->startTime.sec) {
3065 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3070 if (conn->call[i]) {
3071 struct rx_call *tcall = conn->call[i];
3072 MUTEX_ENTER(&tcall->lock);
3073 if (tcall->state == RX_STATE_DALLY) {
3076 MUTEX_EXIT(&tcall->lock);
3082 MUTEX_EXIT(&conn->conn_call_lock);
3084 MUTEX_ENTER(&call->lock);
3086 /* Since the call->lock and conn->conn_call_lock have been released it is
3087 * possible that (1) the call may no longer be busy and/or (2) the call may
3088 * have been reused by another waiting thread. Therefore, we must confirm
3089 * that the call state has not changed when deciding whether or not to
3090 * force this application thread to retry by forcing a Timeout error. */
3092 if (freechannel && *call->callNumber == callNumber &&
3093 (call->flags & RX_CALL_PEER_BUSY)) {
3094 /* Since 'freechannel' is set, there exists another channel in this
3095 * rx_conn that the application thread might be able to use. We know
3096 * that we have the correct call since callNumber is unchanged, and we
3097 * know that the call is still busy. So, set the call error state to
3098 * rxi_busyChannelError so the application can retry the request,
3099 * presumably on a less-busy call channel. */
3101 rxi_CallError(call, RX_CALL_BUSY);
3105 /* There are two packet tracing routines available for testing and monitoring
3106 * Rx. One is called just after every packet is received and the other is
3107 * called just before every packet is sent. Received packets, have had their
3108 * headers decoded, and packets to be sent have not yet had their headers
3109 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3110 * containing the network address. Both can be modified. The return value, if
3111 * non-zero, indicates that the packet should be dropped. */
3113 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3114 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3116 /* A packet has been received off the interface. Np is the packet, socket is
3117 * the socket number it was received from (useful in determining which service
3118 * this packet corresponds to), and (host, port) reflect the host,port of the
3119 * sender. This call returns the packet to the caller if it is finished with
3120 * it, rather than de-allocating it, just as a small performance hack */
3123 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3124 afs_uint32 host, u_short port, int *tnop,
3125 struct rx_call **newcallp)
3127 struct rx_call *call;
3128 struct rx_connection *conn;
3130 afs_uint32 currentCallNumber;
3136 struct rx_packet *tnp;
3139 /* We don't print out the packet until now because (1) the time may not be
3140 * accurate enough until now in the lwp implementation (rx_Listener only gets
3141 * the time after the packet is read) and (2) from a protocol point of view,
3142 * this is the first time the packet has been seen */
3143 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3144 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3145 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3146 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3147 np->header.epoch, np->header.cid, np->header.callNumber,
3148 np->header.seq, np->header.flags, np));
3151 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3152 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3155 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3156 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3159 /* If an input tracer function is defined, call it with the packet and
3160 * network address. Note this function may modify its arguments. */
3161 if (rx_justReceived) {
3162 struct sockaddr_in addr;
3164 addr.sin_family = AF_INET;
3165 addr.sin_port = port;
3166 addr.sin_addr.s_addr = host;
3167 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3168 addr.sin_len = sizeof(addr);
3169 #endif /* AFS_OSF_ENV */
3170 drop = (*rx_justReceived) (np, &addr);
3171 /* drop packet if return value is non-zero */
3174 port = addr.sin_port; /* in case fcn changed addr */
3175 host = addr.sin_addr.s_addr;
3179 /* If packet was not sent by the client, then *we* must be the client */
3180 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3181 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3183 /* Find the connection (or fabricate one, if we're the server & if
3184 * necessary) associated with this packet */
3186 rxi_FindConnection(socket, host, port, np->header.serviceId,
3187 np->header.cid, np->header.epoch, type,
3188 np->header.securityIndex);
3191 /* If no connection found or fabricated, just ignore the packet.
3192 * (An argument could be made for sending an abort packet for
3197 /* If the connection is in an error state, send an abort packet and ignore
3198 * the incoming packet */
3200 /* Don't respond to an abort packet--we don't want loops! */
3201 MUTEX_ENTER(&conn->conn_data_lock);
3202 if (np->header.type != RX_PACKET_TYPE_ABORT)
3203 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3204 putConnection(conn);
3205 MUTEX_EXIT(&conn->conn_data_lock);
3209 /* Check for connection-only requests (i.e. not call specific). */
3210 if (np->header.callNumber == 0) {
3211 switch (np->header.type) {
3212 case RX_PACKET_TYPE_ABORT: {
3213 /* What if the supplied error is zero? */
3214 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3215 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3216 rxi_ConnectionError(conn, errcode);
3217 putConnection(conn);
3220 case RX_PACKET_TYPE_CHALLENGE:
3221 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3222 putConnection(conn);
3224 case RX_PACKET_TYPE_RESPONSE:
3225 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3226 putConnection(conn);
3228 case RX_PACKET_TYPE_PARAMS:
3229 case RX_PACKET_TYPE_PARAMS + 1:
3230 case RX_PACKET_TYPE_PARAMS + 2:
3231 /* ignore these packet types for now */
3232 putConnection(conn);
3236 /* Should not reach here, unless the peer is broken: send an
3238 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3239 MUTEX_ENTER(&conn->conn_data_lock);
3240 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3241 putConnection(conn);
3242 MUTEX_EXIT(&conn->conn_data_lock);
3247 channel = np->header.cid & RX_CHANNELMASK;
3248 call = conn->call[channel];
3251 MUTEX_ENTER(&call->lock);
3252 currentCallNumber = conn->callNumber[channel];
3253 } else if (type == RX_SERVER_CONNECTION) { /* No call allocated */
3254 MUTEX_ENTER(&conn->conn_call_lock);
3255 call = conn->call[channel];
3257 MUTEX_ENTER(&call->lock);
3258 MUTEX_EXIT(&conn->conn_call_lock);
3259 currentCallNumber = conn->callNumber[channel];
3261 call = rxi_NewCall(conn, channel); /* returns locked call */
3262 MUTEX_EXIT(&conn->conn_call_lock);
3263 *call->callNumber = currentCallNumber = np->header.callNumber;
3265 if (np->header.callNumber == 0)
3266 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3267 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3268 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3269 np->header.flags, np, np->length));
3271 call->state = RX_STATE_PRECALL;
3272 clock_GetTime(&call->queueTime);
3273 hzero(call->bytesSent);
3274 hzero(call->bytesRcvd);
3276 * If the number of queued calls exceeds the overload
3277 * threshold then abort this call.
3279 if ((rx_BusyThreshold > 0) &&
3280 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3281 struct rx_packet *tp;
3283 rxi_CallError(call, rx_BusyError);
3284 tp = rxi_SendCallAbort(call, np, 1, 0);
3285 MUTEX_EXIT(&call->lock);
3286 putConnection(conn);
3287 if (rx_stats_active)
3288 rx_atomic_inc(&rx_stats.nBusies);
3291 rxi_KeepAliveOn(call);
3293 } else { /* RX_CLIENT_CONNECTION and No call allocated */
3294 /* This packet can't be for this call. If the new call address is
3295 * 0 then no call is running on this channel. If there is a call
3296 * then, since this is a client connection we're getting data for
3297 * it must be for the previous call.
3299 if (rx_stats_active)
3300 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3301 putConnection(conn);
3305 /* There is a non-NULL locked call at this point */
3306 if (type == RX_SERVER_CONNECTION) { /* We're the server */
3307 if (np->header.callNumber < currentCallNumber) {
3308 MUTEX_EXIT(&call->lock);
3309 if (rx_stats_active)
3310 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3311 putConnection(conn);
3313 } else if (np->header.callNumber != currentCallNumber) {
3314 /* Wait until the transmit queue is idle before deciding
3315 * whether to reset the current call. Chances are that the
3316 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3319 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3320 if (call->state == RX_STATE_ACTIVE) {
3321 rxi_WaitforTQBusy(call);
3323 * If we entered error state while waiting,
3324 * must call rxi_CallError to permit rxi_ResetCall
3325 * to processed when the tqWaiter count hits zero.
3328 rxi_CallError(call, call->error);
3329 MUTEX_EXIT(&call->lock);
3330 putConnection(conn);
3334 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3335 /* If the new call cannot be taken right now send a busy and set
3336 * the error condition in this call, so that it terminates as
3337 * quickly as possible */
3338 if (call->state == RX_STATE_ACTIVE) {
3339 struct rx_packet *tp;
3341 rxi_CallError(call, RX_CALL_DEAD);
3342 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3344 MUTEX_EXIT(&call->lock);
3345 putConnection(conn);
3348 rxi_ResetCall(call, 0);
3349 *call->callNumber = np->header.callNumber;
3351 if (np->header.callNumber == 0)
3352 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3353 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3354 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3355 np->header.flags, np, np->length));
3357 call->state = RX_STATE_PRECALL;
3358 clock_GetTime(&call->queueTime);
3359 hzero(call->bytesSent);
3360 hzero(call->bytesRcvd);
3362 * If the number of queued calls exceeds the overload
3363 * threshold then abort this call.
3365 if ((rx_BusyThreshold > 0) &&
3366 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3367 struct rx_packet *tp;
3369 rxi_CallError(call, rx_BusyError);
3370 tp = rxi_SendCallAbort(call, np, 1, 0);
3371 MUTEX_EXIT(&call->lock);
3372 putConnection(conn);
3373 if (rx_stats_active)
3374 rx_atomic_inc(&rx_stats.nBusies);
3377 rxi_KeepAliveOn(call);
3379 /* Continuing call; do nothing here. */
3381 } else { /* we're the client */
3382 /* Ignore all incoming acknowledgements for calls in DALLY state */
3383 if ((call->state == RX_STATE_DALLY)
3384 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3385 if (rx_stats_active)
3386 rx_atomic_inc(&rx_stats.ignorePacketDally);
3387 MUTEX_EXIT(&call->lock);
3388 putConnection(conn);
3392 /* Ignore anything that's not relevant to the current call. If there
3393 * isn't a current call, then no packet is relevant. */
3394 if (np->header.callNumber != currentCallNumber) {
3395 if (rx_stats_active)
3396 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3397 MUTEX_EXIT(&call->lock);
3398 putConnection(conn);
3401 /* If the service security object index stamped in the packet does not
3402 * match the connection's security index, ignore the packet */
3403 if (np->header.securityIndex != conn->securityIndex) {
3404 MUTEX_EXIT(&call->lock);
3405 putConnection(conn);
3409 /* If we're receiving the response, then all transmit packets are
3410 * implicitly acknowledged. Get rid of them. */
3411 if (np->header.type == RX_PACKET_TYPE_DATA) {
3412 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3413 /* XXX Hack. Because we must release the global rx lock when
3414 * sending packets (osi_NetSend) we drop all acks while we're
3415 * traversing the tq in rxi_Start sending packets out because
3416 * packets may move to the freePacketQueue as result of being here!
3417 * So we drop these packets until we're safely out of the
3418 * traversing. Really ugly!
3419 * For fine grain RX locking, we set the acked field in the
3420 * packets and let rxi_Start remove them from the transmit queue.
3422 if (call->flags & RX_CALL_TQ_BUSY) {
3423 #ifdef RX_ENABLE_LOCKS
3424 rxi_SetAcksInTransmitQueue(call);
3426 putConnection(conn);
3427 return np; /* xmitting; drop packet */
3430 rxi_ClearTransmitQueue(call, 0);
3432 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3433 rxi_ClearTransmitQueue(call, 0);
3434 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3436 if (np->header.type == RX_PACKET_TYPE_ACK) {
3437 /* now check to see if this is an ack packet acknowledging that the
3438 * server actually *lost* some hard-acked data. If this happens we
3439 * ignore this packet, as it may indicate that the server restarted in
3440 * the middle of a call. It is also possible that this is an old ack
3441 * packet. We don't abort the connection in this case, because this
3442 * *might* just be an old ack packet. The right way to detect a server
3443 * restart in the midst of a call is to notice that the server epoch
3445 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3446 * XXX unacknowledged. I think that this is off-by-one, but
3447 * XXX I don't dare change it just yet, since it will
3448 * XXX interact badly with the server-restart detection
3449 * XXX code in receiveackpacket. */
3450 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3451 if (rx_stats_active)
3452 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3453 MUTEX_EXIT(&call->lock);
3454 putConnection(conn);
3458 } /* else not a data packet */
3461 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3462 /* Set remote user defined status from packet */
3463 call->remoteStatus = np->header.userStatus;
3465 /* Note the gap between the expected next packet and the actual
3466 * packet that arrived, when the new packet has a smaller serial number
3467 * than expected. Rioses frequently reorder packets all by themselves,
3468 * so this will be quite important with very large window sizes.
3469 * Skew is checked against 0 here to avoid any dependence on the type of
3470 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3472 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3473 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3474 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3476 MUTEX_ENTER(&conn->conn_data_lock);
3477 skew = conn->lastSerial - np->header.serial;
3478 conn->lastSerial = np->header.serial;
3479 MUTEX_EXIT(&conn->conn_data_lock);
3481 struct rx_peer *peer;
3483 if (skew > peer->inPacketSkew) {
3484 dpf(("*** In skew changed from %d to %d\n",
3485 peer->inPacketSkew, skew));
3486 peer->inPacketSkew = skew;
3490 /* Now do packet type-specific processing */
3491 switch (np->header.type) {
3492 case RX_PACKET_TYPE_DATA:
3493 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3496 case RX_PACKET_TYPE_ACK:
3497 /* Respond immediately to ack packets requesting acknowledgement
3499 if (np->header.flags & RX_REQUEST_ACK) {
3501 (void)rxi_SendCallAbort(call, 0, 1, 0);
3503 (void)rxi_SendAck(call, 0, np->header.serial,
3504 RX_ACK_PING_RESPONSE, 1);
3506 np = rxi_ReceiveAckPacket(call, np, 1);
3508 case RX_PACKET_TYPE_ABORT: {
3509 /* An abort packet: reset the call, passing the error up to the user. */
3510 /* What if error is zero? */
3511 /* What if the error is -1? the application will treat it as a timeout. */
3512 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3513 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3514 rxi_CallError(call, errdata);
3515 MUTEX_EXIT(&call->lock);
3516 putConnection(conn);
3517 return np; /* xmitting; drop packet */
3519 case RX_PACKET_TYPE_BUSY: {
3520 struct clock busyTime;
3522 clock_GetTime(&busyTime);
3524 MUTEX_EXIT(&call->lock);
3526 MUTEX_ENTER(&conn->conn_call_lock);
3527 MUTEX_ENTER(&call->lock);
3528 conn->lastBusy[call->channel] = busyTime.sec;
3529 call->flags |= RX_CALL_PEER_BUSY;
3530 MUTEX_EXIT(&call->lock);
3531 MUTEX_EXIT(&conn->conn_call_lock);
3533 putConnection(conn);
3537 case RX_PACKET_TYPE_ACKALL:
3538 /* All packets acknowledged, so we can drop all packets previously
3539 * readied for sending */
3540 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3541 /* XXX Hack. We because we can't release the global rx lock when
3542 * sending packets (osi_NetSend) we drop all ack pkts while we're
3543 * traversing the tq in rxi_Start sending packets out because
3544 * packets may move to the freePacketQueue as result of being
3545 * here! So we drop these packets until we're safely out of the
3546 * traversing. Really ugly!
3547 * For fine grain RX locking, we set the acked field in the packets
3548 * and let rxi_Start remove the packets from the transmit queue.
3550 if (call->flags & RX_CALL_TQ_BUSY) {
3551 #ifdef RX_ENABLE_LOCKS
3552 rxi_SetAcksInTransmitQueue(call);
3554 #else /* RX_ENABLE_LOCKS */
3555 MUTEX_EXIT(&call->lock);
3556 putConnection(conn);
3557 return np; /* xmitting; drop packet */
3558 #endif /* RX_ENABLE_LOCKS */
3560 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3561 rxi_ClearTransmitQueue(call, 0);
3564 /* Should not reach here, unless the peer is broken: send an abort
3566 rxi_CallError(call, RX_PROTOCOL_ERROR);
3567 np = rxi_SendCallAbort(call, np, 1, 0);
3570 /* Note when this last legitimate packet was received, for keep-alive
3571 * processing. Note, we delay getting the time until now in the hope that
3572 * the packet will be delivered to the user before any get time is required
3573 * (if not, then the time won't actually be re-evaluated here). */
3574 call->lastReceiveTime = clock_Sec();
3575 /* we've received a legit packet, so the channel is not busy */
3576 call->flags &= ~RX_CALL_PEER_BUSY;
3577 MUTEX_EXIT(&call->lock);
3578 putConnection(conn);
3582 /* return true if this is an "interesting" connection from the point of view
3583 of someone trying to debug the system */
3585 rxi_IsConnInteresting(struct rx_connection *aconn)
3588 struct rx_call *tcall;
3590 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3593 for (i = 0; i < RX_MAXCALLS; i++) {
3594 tcall = aconn->call[i];
3596 if ((tcall->state == RX_STATE_PRECALL)
3597 || (tcall->state == RX_STATE_ACTIVE))
3599 if ((tcall->mode == RX_MODE_SENDING)
3600 || (tcall->mode == RX_MODE_RECEIVING))
3608 /* if this is one of the last few packets AND it wouldn't be used by the
3609 receiving call to immediately satisfy a read request, then drop it on
3610 the floor, since accepting it might prevent a lock-holding thread from
3611 making progress in its reading. If a call has been cleared while in
3612 the precall state then ignore all subsequent packets until the call
3613 is assigned to a thread. */
3616 TooLow(struct rx_packet *ap, struct rx_call *acall)
3620 MUTEX_ENTER(&rx_quota_mutex);
3621 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3622 && (acall->state == RX_STATE_PRECALL))
3623 || ((rx_nFreePackets < rxi_dataQuota + 2)
3624 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3625 && (acall->flags & RX_CALL_READER_WAIT)))) {
3628 MUTEX_EXIT(&rx_quota_mutex);
3634 * Clear the attach wait flag on a connection and proceed.
3636 * Any processing waiting for a connection to be attached should be
3637 * unblocked. We clear the flag and do any other needed tasks.
3640 * the conn to unmark waiting for attach
3642 * @pre conn's conn_data_lock must be locked before calling this function
3646 rxi_ConnClearAttachWait(struct rx_connection *conn)
3648 /* Indicate that rxi_CheckReachEvent is no longer running by
3649 * clearing the flag. Must be atomic under conn_data_lock to
3650 * avoid a new call slipping by: rxi_CheckConnReach holds
3651 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3653 conn->flags &= ~RX_CONN_ATTACHWAIT;
3654 if (conn->flags & RX_CONN_NAT_PING) {
3655 conn->flags &= ~RX_CONN_NAT_PING;
3656 rxi_ScheduleNatKeepAliveEvent(conn);
3661 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3663 struct rx_connection *conn = arg1;
3664 struct rx_call *acall = arg2;
3665 struct rx_call *call = acall;
3666 struct clock when, now;
3669 MUTEX_ENTER(&conn->conn_data_lock);
3672 rxevent_Put(conn->checkReachEvent);
3673 conn->checkReachEvent = NULL;
3676 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3678 putConnection(conn);
3680 MUTEX_EXIT(&conn->conn_data_lock);
3684 MUTEX_ENTER(&conn->conn_call_lock);
3685 MUTEX_ENTER(&conn->conn_data_lock);
3686 for (i = 0; i < RX_MAXCALLS; i++) {
3687 struct rx_call *tc = conn->call[i];
3688 if (tc && tc->state == RX_STATE_PRECALL) {
3694 rxi_ConnClearAttachWait(conn);
3695 MUTEX_EXIT(&conn->conn_data_lock);
3696 MUTEX_EXIT(&conn->conn_call_lock);
3701 MUTEX_ENTER(&call->lock);
3702 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3704 MUTEX_EXIT(&call->lock);
3706 clock_GetTime(&now);
3708 when.sec += RX_CHECKREACH_TIMEOUT;
3709 MUTEX_ENTER(&conn->conn_data_lock);
3710 if (!conn->checkReachEvent) {
3711 MUTEX_ENTER(&rx_refcnt_mutex);
3713 MUTEX_EXIT(&rx_refcnt_mutex);
3714 conn->checkReachEvent = rxevent_Post(&when, &now,
3715 rxi_CheckReachEvent, conn,
3718 MUTEX_EXIT(&conn->conn_data_lock);
3724 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3726 struct rx_service *service = conn->service;
3727 struct rx_peer *peer = conn->peer;
3728 afs_uint32 now, lastReach;
3730 if (service->checkReach == 0)
3734 MUTEX_ENTER(&peer->peer_lock);
3735 lastReach = peer->lastReachTime;
3736 MUTEX_EXIT(&peer->peer_lock);
3737 if (now - lastReach < RX_CHECKREACH_TTL)
3740 MUTEX_ENTER(&conn->conn_data_lock);
3741 if (conn->flags & RX_CONN_ATTACHWAIT) {
3742 MUTEX_EXIT(&conn->conn_data_lock);
3745 conn->flags |= RX_CONN_ATTACHWAIT;
3746 MUTEX_EXIT(&conn->conn_data_lock);
3747 if (!conn->checkReachEvent)
3748 rxi_CheckReachEvent(NULL, conn, call, 0);
3753 /* try to attach call, if authentication is complete */
3755 TryAttach(struct rx_call *acall, osi_socket socket,
3756 int *tnop, struct rx_call **newcallp,
3759 struct rx_connection *conn = acall->conn;
3761 if (conn->type == RX_SERVER_CONNECTION
3762 && acall->state == RX_STATE_PRECALL) {
3763 /* Don't attach until we have any req'd. authentication. */
3764 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3765 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3766 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3767 /* Note: this does not necessarily succeed; there
3768 * may not any proc available
3771 rxi_ChallengeOn(acall->conn);
3776 /* A data packet has been received off the interface. This packet is
3777 * appropriate to the call (the call is in the right state, etc.). This
3778 * routine can return a packet to the caller, for re-use */
3781 rxi_ReceiveDataPacket(struct rx_call *call,
3782 struct rx_packet *np, int istack,
3783 osi_socket socket, afs_uint32 host, u_short port,
3784 int *tnop, struct rx_call **newcallp)
3786 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3791 afs_uint32 serial=0, flags=0;
3793 struct rx_packet *tnp;
3794 if (rx_stats_active)
3795 rx_atomic_inc(&rx_stats.dataPacketsRead);
3798 /* If there are no packet buffers, drop this new packet, unless we can find
3799 * packet buffers from inactive calls */
3801 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3802 MUTEX_ENTER(&rx_freePktQ_lock);
3803 rxi_NeedMorePackets = TRUE;
3804 MUTEX_EXIT(&rx_freePktQ_lock);
3805 if (rx_stats_active)
3806 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3807 call->rprev = np->header.serial;
3808 rxi_calltrace(RX_TRACE_DROP, call);
3809 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3810 /* We used to clear the receive queue here, in an attempt to free
3811 * packets. However this is unsafe if the queue has received a
3812 * soft ACK for the final packet */
3813 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3815 /* we've damaged this call already, might as well do it in. */
3821 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3822 * packet is one of several packets transmitted as a single
3823 * datagram. Do not send any soft or hard acks until all packets
3824 * in a jumbogram have been processed. Send negative acks right away.
3826 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3827 /* tnp is non-null when there are more packets in the
3828 * current jumbo gram */
3835 seq = np->header.seq;
3836 serial = np->header.serial;
3837 flags = np->header.flags;
3839 /* If the call is in an error state, send an abort message */
3841 return rxi_SendCallAbort(call, np, istack, 0);
3843 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3844 * AFS 3.5 jumbogram. */
3845 if (flags & RX_JUMBO_PACKET) {
3846 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3851 if (np->header.spare != 0) {
3852 MUTEX_ENTER(&call->conn->conn_data_lock);
3853 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3854 MUTEX_EXIT(&call->conn->conn_data_lock);
3857 /* The usual case is that this is the expected next packet */
3858 if (seq == call->rnext) {
3860 /* Check to make sure it is not a duplicate of one already queued */
3861 if (queue_IsNotEmpty(&call->rq)
3862 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3863 if (rx_stats_active)
3864 rx_atomic_inc(&rx_stats.dupPacketsRead);
3865 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3866 rxevent_Cancel(&call->delayedAckEvent, call,
3867 RX_CALL_REFCOUNT_DELAY);
3868 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3874 /* It's the next packet. Stick it on the receive queue
3875 * for this call. Set newPackets to make sure we wake
3876 * the reader once all packets have been processed */
3877 #ifdef RX_TRACK_PACKETS
3878 np->flags |= RX_PKTFLAG_RQ;
3880 queue_Prepend(&call->rq, np);
3881 #ifdef RXDEBUG_PACKET
3883 #endif /* RXDEBUG_PACKET */
3885 np = NULL; /* We can't use this anymore */
3888 /* If an ack is requested then set a flag to make sure we
3889 * send an acknowledgement for this packet */
3890 if (flags & RX_REQUEST_ACK) {
3891 ackNeeded = RX_ACK_REQUESTED;
3894 /* Keep track of whether we have received the last packet */
3895 if (flags & RX_LAST_PACKET) {
3896 call->flags |= RX_CALL_HAVE_LAST;
3900 /* Check whether we have all of the packets for this call */
3901 if (call->flags & RX_CALL_HAVE_LAST) {
3902 afs_uint32 tseq; /* temporary sequence number */
3903 struct rx_packet *tp; /* Temporary packet pointer */
3904 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3906 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3907 if (tseq != tp->header.seq)
3909 if (tp->header.flags & RX_LAST_PACKET) {
3910 call->flags |= RX_CALL_RECEIVE_DONE;
3917 /* Provide asynchronous notification for those who want it
3918 * (e.g. multi rx) */
3919 if (call->arrivalProc) {
3920 (*call->arrivalProc) (call, call->arrivalProcHandle,
3921 call->arrivalProcArg);
3922 call->arrivalProc = (void (*)())0;
3925 /* Update last packet received */
3928 /* If there is no server process serving this call, grab
3929 * one, if available. We only need to do this once. If a
3930 * server thread is available, this thread becomes a server
3931 * thread and the server thread becomes a listener thread. */
3933 TryAttach(call, socket, tnop, newcallp, 0);
3936 /* This is not the expected next packet. */
3938 /* Determine whether this is a new or old packet, and if it's
3939 * a new one, whether it fits into the current receive window.
3940 * Also figure out whether the packet was delivered in sequence.
3941 * We use the prev variable to determine whether the new packet
3942 * is the successor of its immediate predecessor in the
3943 * receive queue, and the missing flag to determine whether
3944 * any of this packets predecessors are missing. */
3946 afs_uint32 prev; /* "Previous packet" sequence number */
3947 struct rx_packet *tp; /* Temporary packet pointer */
3948 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3949 int missing; /* Are any predecessors missing? */
3951 /* If the new packet's sequence number has been sent to the
3952 * application already, then this is a duplicate */
3953 if (seq < call->rnext) {
3954 if (rx_stats_active)
3955 rx_atomic_inc(&rx_stats.dupPacketsRead);
3956 rxevent_Cancel(&call->delayedAckEvent, call,
3957 RX_CALL_REFCOUNT_DELAY);
3958 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3964 /* If the sequence number is greater than what can be
3965 * accomodated by the current window, then send a negative
3966 * acknowledge and drop the packet */
3967 if ((call->rnext + call->rwind) <= seq) {
3968 rxevent_Cancel(&call->delayedAckEvent, call,
3969 RX_CALL_REFCOUNT_DELAY);
3970 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3977 /* Look for the packet in the queue of old received packets */
3978 for (prev = call->rnext - 1, missing =
3979 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3980 /*Check for duplicate packet */
3981 if (seq == tp->header.seq) {
3982 if (rx_stats_active)
3983 rx_atomic_inc(&rx_stats.dupPacketsRead);
3984 rxevent_Cancel(&call->delayedAckEvent, call,
3985 RX_CALL_REFCOUNT_DELAY);
3986 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3992 /* If we find a higher sequence packet, break out and
3993 * insert the new packet here. */
3994 if (seq < tp->header.seq)
3996 /* Check for missing packet */
3997 if (tp->header.seq != prev + 1) {
4001 prev = tp->header.seq;
4004 /* Keep track of whether we have received the last packet. */
4005 if (flags & RX_LAST_PACKET) {
4006 call->flags |= RX_CALL_HAVE_LAST;
4009 /* It's within the window: add it to the the receive queue.
4010 * tp is left by the previous loop either pointing at the
4011 * packet before which to insert the new packet, or at the
4012 * queue head if the queue is empty or the packet should be
4014 #ifdef RX_TRACK_PACKETS
4015 np->flags |= RX_PKTFLAG_RQ;
4017 #ifdef RXDEBUG_PACKET
4019 #endif /* RXDEBUG_PACKET */
4020 queue_InsertBefore(tp, np);
4024 /* Check whether we have all of the packets for this call */
4025 if ((call->flags & RX_CALL_HAVE_LAST)
4026 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4027 afs_uint32 tseq; /* temporary sequence number */
4030 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
4031 if (tseq != tp->header.seq)
4033 if (tp->header.flags & RX_LAST_PACKET) {
4034 call->flags |= RX_CALL_RECEIVE_DONE;
4041 /* We need to send an ack of the packet is out of sequence,
4042 * or if an ack was requested by the peer. */
4043 if (seq != prev + 1 || missing) {
4044 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4045 } else if (flags & RX_REQUEST_ACK) {
4046 ackNeeded = RX_ACK_REQUESTED;
4049 /* Acknowledge the last packet for each call */
4050 if (flags & RX_LAST_PACKET) {
4061 * If the receiver is waiting for an iovec, fill the iovec
4062 * using the data from the receive queue */
4063 if (call->flags & RX_CALL_IOVEC_WAIT) {
4064 didHardAck = rxi_FillReadVec(call, serial);
4065 /* the call may have been aborted */
4074 /* Wakeup the reader if any */
4075 if ((call->flags & RX_CALL_READER_WAIT)
4076 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4077 || (call->iovNext >= call->iovMax)
4078 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4079 call->flags &= ~RX_CALL_READER_WAIT;
4080 #ifdef RX_ENABLE_LOCKS
4081 CV_BROADCAST(&call->cv_rq);
4083 osi_rxWakeup(&call->rq);
4089 * Send an ack when requested by the peer, or once every
4090 * rxi_SoftAckRate packets until the last packet has been
4091 * received. Always send a soft ack for the last packet in
4092 * the server's reply. */
4094 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4095 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4096 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4097 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4098 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4099 } else if (call->nSoftAcks) {
4100 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4101 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4103 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4104 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4105 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4112 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4114 struct rx_peer *peer = conn->peer;
4116 MUTEX_ENTER(&peer->peer_lock);
4117 peer->lastReachTime = clock_Sec();
4118 MUTEX_EXIT(&peer->peer_lock);
4120 MUTEX_ENTER(&conn->conn_data_lock);
4121 if (conn->flags & RX_CONN_ATTACHWAIT) {
4124 rxi_ConnClearAttachWait(conn);
4125 MUTEX_EXIT(&conn->conn_data_lock);
4127 for (i = 0; i < RX_MAXCALLS; i++) {
4128 struct rx_call *call = conn->call[i];
4131 MUTEX_ENTER(&call->lock);
4132 /* tnop can be null if newcallp is null */
4133 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4135 MUTEX_EXIT(&call->lock);
4139 MUTEX_EXIT(&conn->conn_data_lock);
4142 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4144 rx_ack_reason(int reason)
4147 case RX_ACK_REQUESTED:
4149 case RX_ACK_DUPLICATE:
4151 case RX_ACK_OUT_OF_SEQUENCE:
4153 case RX_ACK_EXCEEDS_WINDOW:
4155 case RX_ACK_NOSPACE:
4159 case RX_ACK_PING_RESPONSE:
4172 /* The real smarts of the whole thing. */
4174 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4177 struct rx_ackPacket *ap;
4179 struct rx_packet *tp;
4180 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
4181 struct rx_connection *conn = call->conn;
4182 struct rx_peer *peer = conn->peer;
4183 struct clock now; /* Current time, for RTT calculations */
4187 /* because there are CM's that are bogus, sending weird values for this. */
4188 afs_uint32 skew = 0;
4193 int newAckCount = 0;
4194 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4195 int pktsize = 0; /* Set if we need to update the peer mtu */
4196 int conn_data_locked = 0;
4198 if (rx_stats_active)
4199 rx_atomic_inc(&rx_stats.ackPacketsRead);
4200 ap = (struct rx_ackPacket *)rx_DataOf(np);
4201 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4203 return np; /* truncated ack packet */
4205 /* depends on ack packet struct */
4206 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4207 first = ntohl(ap->firstPacket);
4208 prev = ntohl(ap->previousPacket);
4209 serial = ntohl(ap->serial);
4210 /* temporarily disabled -- needs to degrade over time
4211 * skew = ntohs(ap->maxSkew); */
4213 /* Ignore ack packets received out of order */
4214 if (first < call->tfirst ||
4215 (first == call->tfirst && prev < call->tprev)) {
4221 if (np->header.flags & RX_SLOW_START_OK) {
4222 call->flags |= RX_CALL_SLOW_START_OK;
4225 if (ap->reason == RX_ACK_PING_RESPONSE)
4226 rxi_UpdatePeerReach(conn, call);
4228 if (conn->lastPacketSizeSeq) {
4229 MUTEX_ENTER(&conn->conn_data_lock);
4230 conn_data_locked = 1;
4231 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4232 pktsize = conn->lastPacketSize;
4233 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4236 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4237 if (!conn_data_locked) {
4238 MUTEX_ENTER(&conn->conn_data_lock);
4239 conn_data_locked = 1;
4241 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4242 /* process mtu ping ack */
4243 pktsize = conn->lastPingSize;
4244 conn->lastPingSizeSer = conn->lastPingSize = 0;
4248 if (conn_data_locked) {
4249 MUTEX_EXIT(&conn->conn_data_lock);
4250 conn_data_locked = 0;
4254 if (rxdebug_active) {
4258 len = _snprintf(msg, sizeof(msg),
4259 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
4260 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4261 ntohl(ap->serial), ntohl(ap->previousPacket),
4262 (unsigned int)np->header.seq, (unsigned int)skew,
4263 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
4267 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4268 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4272 OutputDebugString(msg);
4274 #else /* AFS_NT40_ENV */
4277 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
4278 ap->reason, ntohl(ap->previousPacket),
4279 (unsigned int)np->header.seq, (unsigned int)serial,
4280 (unsigned int)skew, ntohl(ap->firstPacket));
4283 for (offset = 0; offset < nAcks; offset++)
4284 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4289 #endif /* AFS_NT40_ENV */
4292 MUTEX_ENTER(&peer->peer_lock);
4295 * Start somewhere. Can't assume we can send what we can receive,
4296 * but we are clearly receiving.
4298 if (!peer->maxPacketSize)
4299 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4301 if (pktsize > peer->maxPacketSize) {
4302 peer->maxPacketSize = pktsize;
4303 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4304 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4305 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4306 rxi_ScheduleGrowMTUEvent(call, 1);
4311 /* Update the outgoing packet skew value to the latest value of
4312 * the peer's incoming packet skew value. The ack packet, of
4313 * course, could arrive out of order, but that won't affect things
4315 peer->outPacketSkew = skew;
4318 clock_GetTime(&now);
4320 /* The transmit queue splits into 4 sections.
4322 * The first section is packets which have now been acknowledged
4323 * by a window size change in the ack. These have reached the
4324 * application layer, and may be discarded. These are packets
4325 * with sequence numbers < ap->firstPacket.
4327 * The second section is packets which have sequence numbers in
4328 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4329 * contents of the packet's ack array determines whether these
4330 * packets are acknowledged or not.
4332 * The third section is packets which fall above the range
4333 * addressed in the ack packet. These have not yet been received
4336 * The four section is packets which have not yet been transmitted.
4337 * These packets will have a header.serial of 0.
4340 /* First section - implicitly acknowledged packets that can be
4344 tp = queue_First(&call->tq, rx_packet);
4345 while(!queue_IsEnd(&call->tq, tp) && tp->header.seq < first) {
4346 struct rx_packet *next;
4348 next = queue_Next(tp, rx_packet);
4349 call->tfirst = tp->header.seq + 1;
4351 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4353 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4356 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4357 /* XXX Hack. Because we have to release the global rx lock when sending
4358 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4359 * in rxi_Start sending packets out because packets may move to the
4360 * freePacketQueue as result of being here! So we drop these packets until
4361 * we're safely out of the traversing. Really ugly!
4362 * To make it even uglier, if we're using fine grain locking, we can
4363 * set the ack bits in the packets and have rxi_Start remove the packets
4364 * when it's done transmitting.
4366 if (call->flags & RX_CALL_TQ_BUSY) {
4367 #ifdef RX_ENABLE_LOCKS
4368 tp->flags |= RX_PKTFLAG_ACKED;
4369 call->flags |= RX_CALL_TQ_SOME_ACKED;
4370 #else /* RX_ENABLE_LOCKS */
4372 #endif /* RX_ENABLE_LOCKS */
4374 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4377 #ifdef RX_TRACK_PACKETS
4378 tp->flags &= ~RX_PKTFLAG_TQ;
4380 #ifdef RXDEBUG_PACKET
4382 #endif /* RXDEBUG_PACKET */
4383 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4388 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4390 /* Second section of the queue - packets for which we are receiving
4393 * Go through the explicit acks/nacks and record the results in
4394 * the waiting packets. These are packets that can't be released
4395 * yet, even with a positive acknowledge. This positive
4396 * acknowledge only means the packet has been received by the
4397 * peer, not that it will be retained long enough to be sent to
4398 * the peer's upper level. In addition, reset the transmit timers
4399 * of any missing packets (those packets that must be missing
4400 * because this packet was out of sequence) */
4402 call->nSoftAcked = 0;
4404 while (!queue_IsEnd(&call->tq, tp) && tp->header.seq < first + nAcks) {
4405 /* Set the acknowledge flag per packet based on the
4406 * information in the ack packet. An acknowlegded packet can
4407 * be downgraded when the server has discarded a packet it
4408 * soacked previously, or when an ack packet is received
4409 * out of sequence. */
4410 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4411 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4413 tp->flags |= RX_PKTFLAG_ACKED;
4414 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4421 } else /* RX_ACK_TYPE_NACK */ {
4422 tp->flags &= ~RX_PKTFLAG_ACKED;
4426 tp = queue_Next(tp, rx_packet);
4429 /* We don't need to take any action with the 3rd or 4th section in the
4430 * queue - they're not addressed by the contents of this ACK packet.
4433 /* If the window has been extended by this acknowledge packet,
4434 * then wakeup a sender waiting in alloc for window space, or try
4435 * sending packets now, if he's been sitting on packets due to
4436 * lack of window space */
4437 if (call->tnext < (call->tfirst + call->twind)) {
4438 #ifdef RX_ENABLE_LOCKS
4439 CV_SIGNAL(&call->cv_twind);
4441 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4442 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4443 osi_rxWakeup(&call->twind);
4446 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4447 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4451 /* if the ack packet has a receivelen field hanging off it,
4452 * update our state */
4453 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4456 /* If the ack packet has a "recommended" size that is less than
4457 * what I am using now, reduce my size to match */
4458 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4459 (int)sizeof(afs_int32), &tSize);
4460 tSize = (afs_uint32) ntohl(tSize);
4461 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4463 /* Get the maximum packet size to send to this peer */
4464 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4466 tSize = (afs_uint32) ntohl(tSize);
4467 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4468 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4470 /* sanity check - peer might have restarted with different params.
4471 * If peer says "send less", dammit, send less... Peer should never
4472 * be unable to accept packets of the size that prior AFS versions would
4473 * send without asking. */
4474 if (peer->maxMTU != tSize) {
4475 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4477 peer->maxMTU = tSize;
4478 peer->MTU = MIN(tSize, peer->MTU);
4479 call->MTU = MIN(call->MTU, tSize);
4482 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4485 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4486 (int)sizeof(afs_int32), &tSize);
4487 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4488 if (tSize < call->twind) { /* smaller than our send */
4489 call->twind = tSize; /* window, we must send less... */
4490 call->ssthresh = MIN(call->twind, call->ssthresh);
4491 call->conn->twind[call->channel] = call->twind;
4494 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4495 * network MTU confused with the loopback MTU. Calculate the
4496 * maximum MTU here for use in the slow start code below.
4498 /* Did peer restart with older RX version? */
4499 if (peer->maxDgramPackets > 1) {
4500 peer->maxDgramPackets = 1;
4502 } else if (np->length >=
4503 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4506 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4507 sizeof(afs_int32), &tSize);
4508 tSize = (afs_uint32) ntohl(tSize);
4510 * As of AFS 3.5 we set the send window to match the receive window.
4512 if (tSize < call->twind) {
4513 call->twind = tSize;
4514 call->conn->twind[call->channel] = call->twind;
4515 call->ssthresh = MIN(call->twind, call->ssthresh);
4516 } else if (tSize > call->twind) {
4517 call->twind = tSize;
4518 call->conn->twind[call->channel] = call->twind;
4522 * As of AFS 3.5, a jumbogram is more than one fixed size
4523 * packet transmitted in a single UDP datagram. If the remote
4524 * MTU is smaller than our local MTU then never send a datagram
4525 * larger than the natural MTU.
4528 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4529 (int)sizeof(afs_int32), &tSize);
4530 maxDgramPackets = (afs_uint32) ntohl(tSize);
4531 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4533 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4534 if (maxDgramPackets > 1) {
4535 peer->maxDgramPackets = maxDgramPackets;
4536 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4538 peer->maxDgramPackets = 1;
4539 call->MTU = peer->natMTU;
4541 } else if (peer->maxDgramPackets > 1) {
4542 /* Restarted with lower version of RX */
4543 peer->maxDgramPackets = 1;
4545 } else if (peer->maxDgramPackets > 1
4546 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4547 /* Restarted with lower version of RX */
4548 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4549 peer->natMTU = OLD_MAX_PACKET_SIZE;
4550 peer->MTU = OLD_MAX_PACKET_SIZE;
4551 peer->maxDgramPackets = 1;
4552 peer->nDgramPackets = 1;
4554 call->MTU = OLD_MAX_PACKET_SIZE;
4559 * Calculate how many datagrams were successfully received after
4560 * the first missing packet and adjust the negative ack counter
4565 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4566 if (call->nNacks < nNacked) {
4567 call->nNacks = nNacked;
4570 call->nAcks += newAckCount;
4574 /* If the packet contained new acknowledgements, rather than just
4575 * being a duplicate of one we have previously seen, then we can restart
4578 if (newAckCount > 0)
4579 rxi_rto_packet_acked(call, istack);
4581 if (call->flags & RX_CALL_FAST_RECOVER) {
4582 if (newAckCount == 0) {
4583 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4585 call->flags &= ~RX_CALL_FAST_RECOVER;
4586 call->cwind = call->nextCwind;
4587 call->nextCwind = 0;
4590 call->nCwindAcks = 0;
4591 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4592 /* Three negative acks in a row trigger congestion recovery */
4593 call->flags |= RX_CALL_FAST_RECOVER;
4594 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4596 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4597 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4598 call->nextCwind = call->ssthresh;
4601 peer->MTU = call->MTU;
4602 peer->cwind = call->nextCwind;
4603 peer->nDgramPackets = call->nDgramPackets;
4605 call->congestSeq = peer->congestSeq;
4607 /* Reset the resend times on the packets that were nacked
4608 * so we will retransmit as soon as the window permits
4611 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4613 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4614 tp->flags &= ~RX_PKTFLAG_SENT;
4616 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4621 /* If cwind is smaller than ssthresh, then increase
4622 * the window one packet for each ack we receive (exponential
4624 * If cwind is greater than or equal to ssthresh then increase
4625 * the congestion window by one packet for each cwind acks we
4626 * receive (linear growth). */
4627 if (call->cwind < call->ssthresh) {
4629 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4630 call->nCwindAcks = 0;
4632 call->nCwindAcks += newAckCount;
4633 if (call->nCwindAcks >= call->cwind) {
4634 call->nCwindAcks = 0;
4635 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4639 * If we have received several acknowledgements in a row then
4640 * it is time to increase the size of our datagrams
4642 if ((int)call->nAcks > rx_nDgramThreshold) {
4643 if (peer->maxDgramPackets > 1) {
4644 if (call->nDgramPackets < peer->maxDgramPackets) {
4645 call->nDgramPackets++;
4647 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4648 } else if (call->MTU < peer->maxMTU) {
4649 /* don't upgrade if we can't handle it */
4650 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4651 call->MTU = peer->ifMTU;
4653 call->MTU += peer->natMTU;
4654 call->MTU = MIN(call->MTU, peer->maxMTU);
4661 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4663 /* Servers need to hold the call until all response packets have
4664 * been acknowledged. Soft acks are good enough since clients
4665 * are not allowed to clear their receive queues. */
4666 if (call->state == RX_STATE_HOLD
4667 && call->tfirst + call->nSoftAcked >= call->tnext) {
4668 call->state = RX_STATE_DALLY;
4669 rxi_ClearTransmitQueue(call, 0);
4670 rxevent_Cancel(&call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4671 } else if (!queue_IsEmpty(&call->tq)) {
4672 rxi_Start(call, istack);
4677 /* Received a response to a challenge packet */
4679 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4680 struct rx_packet *np, int istack)
4684 /* Ignore the packet if we're the client */
4685 if (conn->type == RX_CLIENT_CONNECTION)
4688 /* If already authenticated, ignore the packet (it's probably a retry) */
4689 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4692 /* Otherwise, have the security object evaluate the response packet */
4693 error = RXS_CheckResponse(conn->securityObject, conn, np);
4695 /* If the response is invalid, reset the connection, sending
4696 * an abort to the peer */
4700 rxi_ConnectionError(conn, error);
4701 MUTEX_ENTER(&conn->conn_data_lock);
4702 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4703 MUTEX_EXIT(&conn->conn_data_lock);
4706 /* If the response is valid, any calls waiting to attach
4707 * servers can now do so */
4710 for (i = 0; i < RX_MAXCALLS; i++) {
4711 struct rx_call *call = conn->call[i];
4713 MUTEX_ENTER(&call->lock);
4714 if (call->state == RX_STATE_PRECALL)
4715 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4716 /* tnop can be null if newcallp is null */
4717 MUTEX_EXIT(&call->lock);
4721 /* Update the peer reachability information, just in case
4722 * some calls went into attach-wait while we were waiting
4723 * for authentication..
4725 rxi_UpdatePeerReach(conn, NULL);
4730 /* A client has received an authentication challenge: the security
4731 * object is asked to cough up a respectable response packet to send
4732 * back to the server. The server is responsible for retrying the
4733 * challenge if it fails to get a response. */
4736 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4737 struct rx_packet *np, int istack)
4741 /* Ignore the challenge if we're the server */
4742 if (conn->type == RX_SERVER_CONNECTION)
4745 /* Ignore the challenge if the connection is otherwise idle; someone's
4746 * trying to use us as an oracle. */
4747 if (!rxi_HasActiveCalls(conn))
4750 /* Send the security object the challenge packet. It is expected to fill
4751 * in the response. */
4752 error = RXS_GetResponse(conn->securityObject, conn, np);
4754 /* If the security object is unable to return a valid response, reset the
4755 * connection and send an abort to the peer. Otherwise send the response
4756 * packet to the peer connection. */
4758 rxi_ConnectionError(conn, error);
4759 MUTEX_ENTER(&conn->conn_data_lock);
4760 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4761 MUTEX_EXIT(&conn->conn_data_lock);
4763 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4764 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4770 /* Find an available server process to service the current request in
4771 * the given call structure. If one isn't available, queue up this
4772 * call so it eventually gets one */
4774 rxi_AttachServerProc(struct rx_call *call,
4775 osi_socket socket, int *tnop,
4776 struct rx_call **newcallp)
4778 struct rx_serverQueueEntry *sq;
4779 struct rx_service *service = call->conn->service;
4782 /* May already be attached */
4783 if (call->state == RX_STATE_ACTIVE)
4786 MUTEX_ENTER(&rx_serverPool_lock);
4788 haveQuota = QuotaOK(service);
4789 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4790 /* If there are no processes available to service this call,
4791 * put the call on the incoming call queue (unless it's
4792 * already on the queue).
4794 #ifdef RX_ENABLE_LOCKS
4796 ReturnToServerPool(service);
4797 #endif /* RX_ENABLE_LOCKS */
4799 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4800 call->flags |= RX_CALL_WAIT_PROC;
4801 rx_atomic_inc(&rx_nWaiting);
4802 rx_atomic_inc(&rx_nWaited);
4803 rxi_calltrace(RX_CALL_ARRIVAL, call);
4804 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4805 queue_Append(&rx_incomingCallQueue, call);
4808 sq = queue_Last(&rx_idleServerQueue, rx_serverQueueEntry);
4810 /* If hot threads are enabled, and both newcallp and sq->socketp
4811 * are non-null, then this thread will process the call, and the
4812 * idle server thread will start listening on this threads socket.
4815 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4818 *sq->socketp = socket;
4819 clock_GetTime(&call->startTime);
4820 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4824 if (call->flags & RX_CALL_WAIT_PROC) {
4825 /* Conservative: I don't think this should happen */
4826 call->flags &= ~RX_CALL_WAIT_PROC;
4827 if (queue_IsOnQueue(call)) {
4830 rx_atomic_dec(&rx_nWaiting);
4833 call->state = RX_STATE_ACTIVE;
4834 call->mode = RX_MODE_RECEIVING;
4835 #ifdef RX_KERNEL_TRACE
4837 int glockOwner = ISAFS_GLOCK();
4840 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4841 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4847 if (call->flags & RX_CALL_CLEARED) {
4848 /* send an ack now to start the packet flow up again */
4849 call->flags &= ~RX_CALL_CLEARED;
4850 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4852 #ifdef RX_ENABLE_LOCKS
4855 service->nRequestsRunning++;
4856 MUTEX_ENTER(&rx_quota_mutex);
4857 if (service->nRequestsRunning <= service->minProcs)
4860 MUTEX_EXIT(&rx_quota_mutex);
4864 MUTEX_EXIT(&rx_serverPool_lock);
4867 /* Delay the sending of an acknowledge event for a short while, while
4868 * a new call is being prepared (in the case of a client) or a reply
4869 * is being prepared (in the case of a server). Rather than sending
4870 * an ack packet, an ACKALL packet is sent. */
4872 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4874 #ifdef RX_ENABLE_LOCKS
4876 MUTEX_ENTER(&call->lock);
4877 rxevent_Put(call->delayedAckEvent);
4878 call->delayedAckEvent = NULL;
4879 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4881 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4882 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4883 call->flags |= RX_CALL_ACKALL_SENT;
4885 MUTEX_EXIT(&call->lock);
4886 #else /* RX_ENABLE_LOCKS */
4888 rxevent_Put(call->delayedAckEvent);
4889 call->delayedAckEvent = NULL;
4891 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4892 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4893 call->flags |= RX_CALL_ACKALL_SENT;
4894 #endif /* RX_ENABLE_LOCKS */
4898 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4901 struct rx_call *call = arg1;
4902 #ifdef RX_ENABLE_LOCKS
4904 MUTEX_ENTER(&call->lock);
4905 if (event == call->delayedAckEvent) {
4906 rxevent_Put(call->delayedAckEvent);
4907 call->delayedAckEvent = NULL;
4909 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4911 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4913 MUTEX_EXIT(&call->lock);
4914 #else /* RX_ENABLE_LOCKS */
4916 rxevent_Put(call->delayedAckEvent);
4917 call->delayedAckEvent = NULL;
4919 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4920 #endif /* RX_ENABLE_LOCKS */
4924 #ifdef RX_ENABLE_LOCKS
4925 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4926 * clearing them out.
4929 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4931 struct rx_packet *p, *tp;
4934 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4935 p->flags |= RX_PKTFLAG_ACKED;
4939 call->flags |= RX_CALL_TQ_CLEARME;
4940 call->flags |= RX_CALL_TQ_SOME_ACKED;
4943 rxi_rto_cancel(call);
4945 call->tfirst = call->tnext;
4946 call->nSoftAcked = 0;
4948 if (call->flags & RX_CALL_FAST_RECOVER) {
4949 call->flags &= ~RX_CALL_FAST_RECOVER;
4950 call->cwind = call->nextCwind;
4951 call->nextCwind = 0;
4954 CV_SIGNAL(&call->cv_twind);
4956 #endif /* RX_ENABLE_LOCKS */
4958 /* Clear out the transmit queue for the current call (all packets have
4959 * been received by peer) */
4961 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4963 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4964 struct rx_packet *p, *tp;
4966 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4968 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4969 p->flags |= RX_PKTFLAG_ACKED;
4973 call->flags |= RX_CALL_TQ_CLEARME;
4974 call->flags |= RX_CALL_TQ_SOME_ACKED;
4977 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4978 #ifdef RXDEBUG_PACKET
4980 #endif /* RXDEBUG_PACKET */
4981 rxi_FreePackets(0, &call->tq);
4982 rxi_WakeUpTransmitQueue(call);
4983 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4984 call->flags &= ~RX_CALL_TQ_CLEARME;
4986 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4988 rxi_rto_cancel(call);
4989 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4990 call->nSoftAcked = 0;
4992 if (call->flags & RX_CALL_FAST_RECOVER) {
4993 call->flags &= ~RX_CALL_FAST_RECOVER;
4994 call->cwind = call->nextCwind;
4996 #ifdef RX_ENABLE_LOCKS
4997 CV_SIGNAL(&call->cv_twind);
4999 osi_rxWakeup(&call->twind);
5004 rxi_ClearReceiveQueue(struct rx_call *call)
5006 if (queue_IsNotEmpty(&call->rq)) {
5009 count = rxi_FreePackets(0, &call->rq);
5010 rx_packetReclaims += count;
5011 #ifdef RXDEBUG_PACKET
5013 if ( call->rqc != 0 )
5014 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5016 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5018 if (call->state == RX_STATE_PRECALL) {
5019 call->flags |= RX_CALL_CLEARED;
5023 /* Send an abort packet for the specified call */
5025 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5026 int istack, int force)
5028 afs_int32 error, cerror;
5029 struct clock when, now;
5034 switch (call->error) {
5037 cerror = RX_CALL_TIMEOUT;
5040 cerror = call->error;
5043 /* Clients should never delay abort messages */
5044 if (rx_IsClientConn(call->conn))
5047 if (call->abortCode != cerror) {
5048 call->abortCode = cerror;
5049 call->abortCount = 0;
5052 if (force || rxi_callAbortThreshhold == 0
5053 || call->abortCount < rxi_callAbortThreshhold) {
5054 if (call->delayedAbortEvent) {
5055 rxevent_Cancel(&call->delayedAbortEvent, call,
5056 RX_CALL_REFCOUNT_ABORT);
5058 error = htonl(cerror);
5061 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5062 (char *)&error, sizeof(error), istack);
5063 } else if (!call->delayedAbortEvent) {
5064 clock_GetTime(&now);
5066 clock_Addmsec(&when, rxi_callAbortDelay);
5067 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5068 call->delayedAbortEvent =
5069 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5074 /* Send an abort packet for the specified connection. Packet is an
5075 * optional pointer to a packet that can be used to send the abort.
5076 * Once the number of abort messages reaches the threshhold, an
5077 * event is scheduled to send the abort. Setting the force flag
5078 * overrides sending delayed abort messages.
5080 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5081 * to send the abort packet.
5084 rxi_SendConnectionAbort(struct rx_connection *conn,
5085 struct rx_packet *packet, int istack, int force)
5088 struct clock when, now;
5093 /* Clients should never delay abort messages */
5094 if (rx_IsClientConn(conn))
5097 if (force || rxi_connAbortThreshhold == 0
5098 || conn->abortCount < rxi_connAbortThreshhold) {
5100 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
5101 error = htonl(conn->error);
5103 MUTEX_EXIT(&conn->conn_data_lock);
5105 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5106 RX_PACKET_TYPE_ABORT, (char *)&error,
5107 sizeof(error), istack);
5108 MUTEX_ENTER(&conn->conn_data_lock);
5109 } else if (!conn->delayedAbortEvent) {
5110 clock_GetTime(&now);
5112 clock_Addmsec(&when, rxi_connAbortDelay);
5113 conn->delayedAbortEvent =
5114 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
5119 /* Associate an error all of the calls owned by a connection. Called
5120 * with error non-zero. This is only for really fatal things, like
5121 * bad authentication responses. The connection itself is set in
5122 * error at this point, so that future packets received will be
5125 rxi_ConnectionError(struct rx_connection *conn,
5131 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5133 MUTEX_ENTER(&conn->conn_data_lock);
5134 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
5135 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
5136 if (conn->checkReachEvent) {
5137 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
5138 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5139 putConnection(conn);
5141 MUTEX_EXIT(&conn->conn_data_lock);
5142 for (i = 0; i < RX_MAXCALLS; i++) {
5143 struct rx_call *call = conn->call[i];
5145 MUTEX_ENTER(&call->lock);
5146 rxi_CallError(call, error);
5147 MUTEX_EXIT(&call->lock);
5150 conn->error = error;
5151 if (rx_stats_active)
5152 rx_atomic_inc(&rx_stats.fatalErrors);
5157 * Interrupt an in-progress call with the specified error and wakeup waiters.
5159 * @param[in] call The call to interrupt
5160 * @param[in] error The error code to send to the peer
5163 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5165 MUTEX_ENTER(&call->lock);
5166 rxi_CallError(call, error);
5167 rxi_SendCallAbort(call, NULL, 0, 1);
5168 MUTEX_EXIT(&call->lock);
5172 rxi_CallError(struct rx_call *call, afs_int32 error)
5175 osirx_AssertMine(&call->lock, "rxi_CallError");
5177 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5179 error = call->error;
5181 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5182 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5183 rxi_ResetCall(call, 0);
5186 rxi_ResetCall(call, 0);
5188 call->error = error;
5191 /* Reset various fields in a call structure, and wakeup waiting
5192 * processes. Some fields aren't changed: state & mode are not
5193 * touched (these must be set by the caller), and bufptr, nLeft, and
5194 * nFree are not reset, since these fields are manipulated by
5195 * unprotected macros, and may only be reset by non-interrupting code.
5199 rxi_ResetCall(struct rx_call *call, int newcall)
5202 struct rx_peer *peer;
5203 struct rx_packet *packet;
5205 osirx_AssertMine(&call->lock, "rxi_ResetCall");
5207 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5209 /* Notify anyone who is waiting for asynchronous packet arrival */
5210 if (call->arrivalProc) {
5211 (*call->arrivalProc) (call, call->arrivalProcHandle,
5212 call->arrivalProcArg);
5213 call->arrivalProc = (void (*)())0;
5217 rxevent_Cancel(&call->growMTUEvent, call, RX_CALL_REFCOUNT_ALIVE);
5219 if (call->delayedAbortEvent) {
5220 rxevent_Cancel(&call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
5221 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5223 rxi_SendCallAbort(call, packet, 0, 1);
5224 rxi_FreePacket(packet);
5229 * Update the peer with the congestion information in this call
5230 * so other calls on this connection can pick up where this call
5231 * left off. If the congestion sequence numbers don't match then
5232 * another call experienced a retransmission.
5234 peer = call->conn->peer;
5235 MUTEX_ENTER(&peer->peer_lock);
5237 if (call->congestSeq == peer->congestSeq) {
5238 peer->cwind = MAX(peer->cwind, call->cwind);
5239 peer->MTU = MAX(peer->MTU, call->MTU);
5240 peer->nDgramPackets =
5241 MAX(peer->nDgramPackets, call->nDgramPackets);
5244 call->abortCode = 0;
5245 call->abortCount = 0;
5247 if (peer->maxDgramPackets > 1) {
5248 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5250 call->MTU = peer->MTU;
5252 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5253 call->ssthresh = rx_maxSendWindow;
5254 call->nDgramPackets = peer->nDgramPackets;
5255 call->congestSeq = peer->congestSeq;
5256 call->rtt = peer->rtt;
5257 call->rtt_dev = peer->rtt_dev;
5258 clock_Zero(&call->rto);
5259 clock_Addmsec(&call->rto,
5260 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5261 MUTEX_EXIT(&peer->peer_lock);
5263 flags = call->flags;
5264 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5265 rxi_WaitforTQBusy(call);
5266 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5268 rxi_ClearTransmitQueue(call, 1);
5269 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5270 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5274 if (!newcall && (flags & RX_CALL_PEER_BUSY)) {
5275 /* The call channel is still busy; resetting the call doesn't change
5276 * that. However, if 'newcall' is set, we are processing a call
5277 * structure that has either been recycled from the free list, or has
5278 * been newly allocated. So, RX_CALL_PEER_BUSY is not relevant if
5279 * 'newcall' is set, since it describes a completely different call
5280 * channel which we do not care about. */
5281 call->flags |= RX_CALL_PEER_BUSY;
5284 rxi_ClearReceiveQueue(call);
5285 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5289 call->twind = call->conn->twind[call->channel];
5290 call->rwind = call->conn->rwind[call->channel];
5291 call->nSoftAcked = 0;
5292 call->nextCwind = 0;
5295 call->nCwindAcks = 0;
5296 call->nSoftAcks = 0;
5297 call->nHardAcks = 0;
5299 call->tfirst = call->rnext = call->tnext = 1;
5302 call->lastAcked = 0;
5303 call->localStatus = call->remoteStatus = 0;
5305 if (flags & RX_CALL_READER_WAIT) {
5306 #ifdef RX_ENABLE_LOCKS
5307 CV_BROADCAST(&call->cv_rq);
5309 osi_rxWakeup(&call->rq);
5312 if (flags & RX_CALL_WAIT_PACKETS) {
5313 MUTEX_ENTER(&rx_freePktQ_lock);
5314 rxi_PacketsUnWait(); /* XXX */
5315 MUTEX_EXIT(&rx_freePktQ_lock);
5317 #ifdef RX_ENABLE_LOCKS
5318 CV_SIGNAL(&call->cv_twind);
5320 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5321 osi_rxWakeup(&call->twind);
5324 #ifdef RX_ENABLE_LOCKS
5325 /* The following ensures that we don't mess with any queue while some
5326 * other thread might also be doing so. The call_queue_lock field is
5327 * is only modified under the call lock. If the call is in the process
5328 * of being removed from a queue, the call is not locked until the
5329 * the queue lock is dropped and only then is the call_queue_lock field
5330 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5331 * Note that any other routine which removes a call from a queue has to
5332 * obtain the queue lock before examing the queue and removing the call.
5334 if (call->call_queue_lock) {
5335 MUTEX_ENTER(call->call_queue_lock);
5336 if (queue_IsOnQueue(call)) {
5338 if (flags & RX_CALL_WAIT_PROC) {
5339 rx_atomic_dec(&rx_nWaiting);
5342 MUTEX_EXIT(call->call_queue_lock);
5343 CLEAR_CALL_QUEUE_LOCK(call);
5345 #else /* RX_ENABLE_LOCKS */
5346 if (queue_IsOnQueue(call)) {
5348 if (flags & RX_CALL_WAIT_PROC)
5349 rx_atomic_dec(&rx_nWaiting);
5351 #endif /* RX_ENABLE_LOCKS */
5353 rxi_KeepAliveOff(call);
5354 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5357 /* Send an acknowledge for the indicated packet (seq,serial) of the
5358 * indicated call, for the indicated reason (reason). This
5359 * acknowledge will specifically acknowledge receiving the packet, and
5360 * will also specify which other packets for this call have been
5361 * received. This routine returns the packet that was used to the
5362 * caller. The caller is responsible for freeing it or re-using it.
5363 * This acknowledgement also returns the highest sequence number
5364 * actually read out by the higher level to the sender; the sender
5365 * promises to keep around packets that have not been read by the
5366 * higher level yet (unless, of course, the sender decides to abort
5367 * the call altogether). Any of p, seq, serial, pflags, or reason may
5368 * be set to zero without ill effect. That is, if they are zero, they
5369 * will not convey any information.
5370 * NOW there is a trailer field, after the ack where it will safely be
5371 * ignored by mundanes, which indicates the maximum size packet this
5372 * host can swallow. */
5374 struct rx_packet *optionalPacket; use to send ack (or null)
5375 int seq; Sequence number of the packet we are acking
5376 int serial; Serial number of the packet
5377 int pflags; Flags field from packet header
5378 int reason; Reason an acknowledge was prompted
5382 rxi_SendAck(struct rx_call *call,
5383 struct rx_packet *optionalPacket, int serial, int reason,
5386 struct rx_ackPacket *ap;
5387 struct rx_packet *rqp;
5388 struct rx_packet *nxp; /* For queue_Scan */
5389 struct rx_packet *p;
5392 afs_uint32 padbytes = 0;
5393 #ifdef RX_ENABLE_TSFPQ
5394 struct rx_ts_info_t * rx_ts_info;
5398 * Open the receive window once a thread starts reading packets
5400 if (call->rnext > 1) {
5401 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5404 /* Don't attempt to grow MTU if this is a critical ping */
5405 if (reason == RX_ACK_MTU) {
5406 /* keep track of per-call attempts, if we're over max, do in small
5407 * otherwise in larger? set a size to increment by, decrease
5410 if (call->conn->peer->maxPacketSize &&
5411 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5413 padbytes = call->conn->peer->maxPacketSize+16;
5415 padbytes = call->conn->peer->maxMTU + 128;
5417 /* do always try a minimum size ping */
5418 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5420 /* subtract the ack payload */
5421 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5422 reason = RX_ACK_PING;
5425 call->nHardAcks = 0;
5426 call->nSoftAcks = 0;
5427 if (call->rnext > call->lastAcked)
5428 call->lastAcked = call->rnext;
5432 rx_computelen(p, p->length); /* reset length, you never know */
5433 } /* where that's been... */
5434 #ifdef RX_ENABLE_TSFPQ
5436 RX_TS_INFO_GET(rx_ts_info);
5437 if ((p = rx_ts_info->local_special_packet)) {
5438 rx_computelen(p, p->length);
5439 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5440 rx_ts_info->local_special_packet = p;
5441 } else { /* We won't send the ack, but don't panic. */
5442 return optionalPacket;
5446 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5447 /* We won't send the ack, but don't panic. */
5448 return optionalPacket;
5453 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5456 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5457 #ifndef RX_ENABLE_TSFPQ
5458 if (!optionalPacket)
5461 return optionalPacket;
5463 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5464 if (rx_Contiguous(p) < templ) {
5465 #ifndef RX_ENABLE_TSFPQ
5466 if (!optionalPacket)
5469 return optionalPacket;
5474 /* MTUXXX failing to send an ack is very serious. We should */
5475 /* try as hard as possible to send even a partial ack; it's */
5476 /* better than nothing. */
5477 ap = (struct rx_ackPacket *)rx_DataOf(p);
5478 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5479 ap->reason = reason;
5481 /* The skew computation used to be bogus, I think it's better now. */
5482 /* We should start paying attention to skew. XXX */
5483 ap->serial = htonl(serial);
5484 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5487 * First packet not yet forwarded to reader. When ACKALL has been
5488 * sent the peer has been told that all received packets will be
5489 * delivered to the reader. The value 'rnext' is used internally
5490 * to refer to the next packet in the receive queue that must be
5491 * delivered to the reader. From the perspective of the peer it
5492 * already has so report the last sequence number plus one if there
5493 * are packets in the receive queue awaiting processing.
5495 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5496 !queue_IsEmpty(&call->rq)) {
5497 ap->firstPacket = htonl(queue_Last(&call->rq, rx_packet)->header.seq + 1);
5499 ap->firstPacket = htonl(call->rnext);
5501 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5503 /* No fear of running out of ack packet here because there can only be at most
5504 * one window full of unacknowledged packets. The window size must be constrained
5505 * to be less than the maximum ack size, of course. Also, an ack should always
5506 * fit into a single packet -- it should not ever be fragmented. */
5507 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5508 if (!rqp || !call->rq.next
5509 || (rqp->header.seq > (call->rnext + call->rwind))) {
5510 #ifndef RX_ENABLE_TSFPQ
5511 if (!optionalPacket)
5514 rxi_CallError(call, RX_CALL_DEAD);
5515 return optionalPacket;
5518 while (rqp->header.seq > call->rnext + offset)
5519 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5520 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5522 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5523 #ifndef RX_ENABLE_TSFPQ
5524 if (!optionalPacket)
5527 rxi_CallError(call, RX_CALL_DEAD);
5528 return optionalPacket;
5534 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5536 /* these are new for AFS 3.3 */
5537 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5538 templ = htonl(templ);
5539 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5540 templ = htonl(call->conn->peer->ifMTU);
5541 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5542 sizeof(afs_int32), &templ);
5544 /* new for AFS 3.4 */
5545 templ = htonl(call->rwind);
5546 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5547 sizeof(afs_int32), &templ);
5549 /* new for AFS 3.5 */
5550 templ = htonl(call->conn->peer->ifDgramPackets);
5551 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5552 sizeof(afs_int32), &templ);
5554 p->header.serviceId = call->conn->serviceId;
5555 p->header.cid = (call->conn->cid | call->channel);
5556 p->header.callNumber = *call->callNumber;
5558 p->header.securityIndex = call->conn->securityIndex;
5559 p->header.epoch = call->conn->epoch;
5560 p->header.type = RX_PACKET_TYPE_ACK;
5561 p->header.flags = RX_SLOW_START_OK;
5562 if (reason == RX_ACK_PING) {
5563 p->header.flags |= RX_REQUEST_ACK;
5565 p->length = padbytes +
5566 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5569 /* not fast but we can potentially use this if truncated
5570 * fragments are delivered to figure out the mtu.
5572 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5573 sizeof(afs_int32), sizeof(afs_int32),
5577 if (call->conn->type == RX_CLIENT_CONNECTION)
5578 p->header.flags |= RX_CLIENT_INITIATED;
5582 if (rxdebug_active) {
5586 len = _snprintf(msg, sizeof(msg),
5587 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5588 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5589 ntohl(ap->serial), ntohl(ap->previousPacket),
5590 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5591 ap->nAcks, ntohs(ap->bufferSpace) );
5595 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5596 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5600 OutputDebugString(msg);
5602 #else /* AFS_NT40_ENV */
5604 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5605 ap->reason, ntohl(ap->previousPacket),
5606 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5608 for (offset = 0; offset < ap->nAcks; offset++)
5609 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5614 #endif /* AFS_NT40_ENV */
5617 int i, nbytes = p->length;
5619 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5620 if (nbytes <= p->wirevec[i].iov_len) {
5623 savelen = p->wirevec[i].iov_len;
5625 p->wirevec[i].iov_len = nbytes;
5627 rxi_Send(call, p, istack);
5628 p->wirevec[i].iov_len = savelen;
5632 nbytes -= p->wirevec[i].iov_len;
5635 if (rx_stats_active)
5636 rx_atomic_inc(&rx_stats.ackPacketsSent);
5637 #ifndef RX_ENABLE_TSFPQ
5638 if (!optionalPacket)
5641 return optionalPacket; /* Return packet for re-use by caller */
5645 struct rx_packet **list;
5650 /* Send all of the packets in the list in single datagram */
5652 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5653 int istack, int moreFlag)
5659 struct rx_connection *conn = call->conn;
5660 struct rx_peer *peer = conn->peer;
5662 MUTEX_ENTER(&peer->peer_lock);
5663 peer->nSent += xmit->len;
5664 if (xmit->resending)
5665 peer->reSends += xmit->len;
5666 MUTEX_EXIT(&peer->peer_lock);
5668 if (rx_stats_active) {
5669 if (xmit->resending)
5670 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5672 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5675 clock_GetTime(&now);
5677 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5681 /* Set the packet flags and schedule the resend events */
5682 /* Only request an ack for the last packet in the list */
5683 for (i = 0; i < xmit->len; i++) {
5684 struct rx_packet *packet = xmit->list[i];
5686 /* Record the time sent */
5687 packet->timeSent = now;
5688 packet->flags |= RX_PKTFLAG_SENT;
5690 /* Ask for an ack on retransmitted packets, on every other packet
5691 * if the peer doesn't support slow start. Ask for an ack on every
5692 * packet until the congestion window reaches the ack rate. */
5693 if (packet->header.serial) {
5696 packet->firstSent = now;
5697 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5698 || (!(call->flags & RX_CALL_SLOW_START_OK)
5699 && (packet->header.seq & 1)))) {
5704 /* Tag this packet as not being the last in this group,
5705 * for the receiver's benefit */
5706 if (i < xmit->len - 1 || moreFlag) {
5707 packet->header.flags |= RX_MORE_PACKETS;
5712 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5715 /* Since we're about to send a data packet to the peer, it's
5716 * safe to nuke any scheduled end-of-packets ack */
5717 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5719 MUTEX_EXIT(&call->lock);
5720 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5721 if (xmit->len > 1) {
5722 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5724 rxi_SendPacket(call, conn, xmit->list[0], istack);
5726 MUTEX_ENTER(&call->lock);
5727 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5729 /* Tell the RTO calculation engine that we have sent a packet, and
5730 * if it was the last one */
5731 rxi_rto_packet_sent(call, lastPacket, istack);
5733 /* Update last send time for this call (for keep-alive
5734 * processing), and for the connection (so that we can discover
5735 * idle connections) */
5736 conn->lastSendTime = call->lastSendTime = clock_Sec();
5737 /* Let a set of retransmits trigger an idle timeout */
5738 if (!xmit->resending)
5739 call->lastSendData = call->lastSendTime;
5742 /* When sending packets we need to follow these rules:
5743 * 1. Never send more than maxDgramPackets in a jumbogram.
5744 * 2. Never send a packet with more than two iovecs in a jumbogram.
5745 * 3. Never send a retransmitted packet in a jumbogram.
5746 * 4. Never send more than cwind/4 packets in a jumbogram
5747 * We always keep the last list we should have sent so we
5748 * can set the RX_MORE_PACKETS flags correctly.
5752 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5757 struct xmitlist working;
5758 struct xmitlist last;
5760 struct rx_peer *peer = call->conn->peer;
5761 int morePackets = 0;
5763 memset(&last, 0, sizeof(struct xmitlist));
5764 working.list = &list[0];
5766 working.resending = 0;
5768 recovery = call->flags & RX_CALL_FAST_RECOVER;
5770 for (i = 0; i < len; i++) {
5771 /* Does the current packet force us to flush the current list? */
5773 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5774 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5776 /* This sends the 'last' list and then rolls the current working
5777 * set into the 'last' one, and resets the working set */
5780 rxi_SendList(call, &last, istack, 1);
5781 /* If the call enters an error state stop sending, or if
5782 * we entered congestion recovery mode, stop sending */
5784 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5789 working.resending = 0;
5790 working.list = &list[i];
5792 /* Add the current packet to the list if it hasn't been acked.
5793 * Otherwise adjust the list pointer to skip the current packet. */
5794 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5797 if (list[i]->header.serial)
5798 working.resending = 1;
5800 /* Do we need to flush the list? */
5801 if (working.len >= (int)peer->maxDgramPackets
5802 || working.len >= (int)call->nDgramPackets
5803 || working.len >= (int)call->cwind
5804 || list[i]->header.serial
5805 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5807 rxi_SendList(call, &last, istack, 1);
5808 /* If the call enters an error state stop sending, or if
5809 * we entered congestion recovery mode, stop sending */
5811 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5816 working.resending = 0;
5817 working.list = &list[i + 1];
5820 if (working.len != 0) {
5821 osi_Panic("rxi_SendList error");
5823 working.list = &list[i + 1];
5827 /* Send the whole list when the call is in receive mode, when
5828 * the call is in eof mode, when we are in fast recovery mode,
5829 * and when we have the last packet */
5830 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5831 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5832 || (call->flags & RX_CALL_FAST_RECOVER)) {
5833 /* Check for the case where the current list contains
5834 * an acked packet. Since we always send retransmissions
5835 * in a separate packet, we only need to check the first
5836 * packet in the list */
5837 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5841 rxi_SendList(call, &last, istack, morePackets);
5842 /* If the call enters an error state stop sending, or if
5843 * we entered congestion recovery mode, stop sending */
5845 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5849 rxi_SendList(call, &working, istack, 0);
5851 } else if (last.len > 0) {
5852 rxi_SendList(call, &last, istack, 0);
5853 /* Packets which are in 'working' are not sent by this call */
5858 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
5860 struct rx_call *call = arg0;
5861 struct rx_peer *peer;
5862 struct rx_packet *p, *nxp;
5863 struct clock maxTimeout = { 60, 0 };
5865 MUTEX_ENTER(&call->lock);
5867 peer = call->conn->peer;
5869 /* Make sure that the event pointer is removed from the call
5870 * structure, since there is no longer a per-call retransmission
5872 if (event == call->resendEvent) {
5873 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5874 rxevent_Put(call->resendEvent);
5875 call->resendEvent = NULL;
5878 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
5879 rxi_CheckBusy(call);
5882 if (queue_IsEmpty(&call->tq)) {
5883 /* Nothing to do. This means that we've been raced, and that an
5884 * ACK has come in between when we were triggered, and when we
5885 * actually got to run. */
5889 /* We're in loss recovery */
5890 call->flags |= RX_CALL_FAST_RECOVER;
5892 /* Mark all of the pending packets in the queue as being lost */
5893 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5894 if (!(p->flags & RX_PKTFLAG_ACKED))
5895 p->flags &= ~RX_PKTFLAG_SENT;
5898 /* We're resending, so we double the timeout of the call. This will be
5899 * dropped back down by the first successful ACK that we receive.
5901 * We apply a maximum value here of 60 seconds
5903 clock_Add(&call->rto, &call->rto);
5904 if (clock_Gt(&call->rto, &maxTimeout))
5905 call->rto = maxTimeout;
5907 /* Packet loss is most likely due to congestion, so drop our window size
5908 * and start again from the beginning */
5909 if (peer->maxDgramPackets >1) {
5910 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5911 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5913 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5914 call->nDgramPackets = 1;
5916 call->nextCwind = 1;
5919 MUTEX_ENTER(&peer->peer_lock);
5920 peer->MTU = call->MTU;
5921 peer->cwind = call->cwind;
5922 peer->nDgramPackets = 1;
5924 call->congestSeq = peer->congestSeq;
5925 MUTEX_EXIT(&peer->peer_lock);
5927 rxi_Start(call, istack);
5930 MUTEX_EXIT(&call->lock);
5933 /* This routine is called when new packets are readied for
5934 * transmission and when retransmission may be necessary, or when the
5935 * transmission window or burst count are favourable. This should be
5936 * better optimized for new packets, the usual case, now that we've
5937 * got rid of queues of send packets. XXXXXXXXXXX */
5939 rxi_Start(struct rx_call *call, int istack)
5942 struct rx_packet *p;
5943 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5948 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5949 if (rx_stats_active)
5950 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5955 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5957 /* Send (or resend) any packets that need it, subject to
5958 * window restrictions and congestion burst control
5959 * restrictions. Ask for an ack on the last packet sent in
5960 * this burst. For now, we're relying upon the window being
5961 * considerably bigger than the largest number of packets that
5962 * are typically sent at once by one initial call to
5963 * rxi_Start. This is probably bogus (perhaps we should ask
5964 * for an ack when we're half way through the current
5965 * window?). Also, for non file transfer applications, this
5966 * may end up asking for an ack for every packet. Bogus. XXXX
5969 * But check whether we're here recursively, and let the other guy
5972 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5973 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5974 call->flags |= RX_CALL_TQ_BUSY;
5976 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5978 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5979 call->flags &= ~RX_CALL_NEED_START;
5980 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5982 maxXmitPackets = MIN(call->twind, call->cwind);
5983 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5984 #ifdef RX_TRACK_PACKETS
5985 if ((p->flags & RX_PKTFLAG_FREE)
5986 || (!queue_IsEnd(&call->tq, nxp)
5987 && (nxp->flags & RX_PKTFLAG_FREE))
5988 || (p == (struct rx_packet *)&rx_freePacketQueue)
5989 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5990 osi_Panic("rxi_Start: xmit queue clobbered");
5993 if (p->flags & RX_PKTFLAG_ACKED) {
5994 /* Since we may block, don't trust this */
5995 if (rx_stats_active)
5996 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
5997 continue; /* Ignore this packet if it has been acknowledged */
6000 /* Turn off all flags except these ones, which are the same
6001 * on each transmission */
6002 p->header.flags &= RX_PRESET_FLAGS;
6004 if (p->header.seq >=
6005 call->tfirst + MIN((int)call->twind,
6006 (int)(call->nSoftAcked +
6008 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6009 /* Note: if we're waiting for more window space, we can
6010 * still send retransmits; hence we don't return here, but
6011 * break out to schedule a retransmit event */
6012 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6013 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6018 /* Transmit the packet if it needs to be sent. */
6019 if (!(p->flags & RX_PKTFLAG_SENT)) {
6020 if (nXmitPackets == maxXmitPackets) {
6021 rxi_SendXmitList(call, call->xmitList,
6022 nXmitPackets, istack);
6025 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6026 *(call->callNumber), p));
6027 call->xmitList[nXmitPackets++] = p;
6031 /* xmitList now hold pointers to all of the packets that are
6032 * ready to send. Now we loop to send the packets */
6033 if (nXmitPackets > 0) {
6034 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6038 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6040 /* We went into the error state while sending packets. Now is
6041 * the time to reset the call. This will also inform the using
6042 * process that the call is in an error state.
6044 if (rx_stats_active)
6045 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6046 call->flags &= ~RX_CALL_TQ_BUSY;
6047 rxi_WakeUpTransmitQueue(call);
6048 rxi_CallError(call, call->error);
6051 #ifdef RX_ENABLE_LOCKS
6052 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6054 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6055 /* Some packets have received acks. If they all have, we can clear
6056 * the transmit queue.
6059 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
6060 if (p->header.seq < call->tfirst
6061 && (p->flags & RX_PKTFLAG_ACKED)) {
6063 #ifdef RX_TRACK_PACKETS
6064 p->flags &= ~RX_PKTFLAG_TQ;
6066 #ifdef RXDEBUG_PACKET
6074 call->flags |= RX_CALL_TQ_CLEARME;
6076 #endif /* RX_ENABLE_LOCKS */
6077 if (call->flags & RX_CALL_TQ_CLEARME)
6078 rxi_ClearTransmitQueue(call, 1);
6079 } while (call->flags & RX_CALL_NEED_START);
6081 * TQ references no longer protected by this flag; they must remain
6082 * protected by the global lock.
6084 call->flags &= ~RX_CALL_TQ_BUSY;
6085 rxi_WakeUpTransmitQueue(call);
6087 call->flags |= RX_CALL_NEED_START;
6089 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6091 rxi_rto_cancel(call);
6095 /* Also adjusts the keep alive parameters for the call, to reflect
6096 * that we have just sent a packet (so keep alives aren't sent
6099 rxi_Send(struct rx_call *call, struct rx_packet *p,
6102 struct rx_connection *conn = call->conn;
6104 /* Stamp each packet with the user supplied status */
6105 p->header.userStatus = call->localStatus;
6107 /* Allow the security object controlling this call's security to
6108 * make any last-minute changes to the packet */
6109 RXS_SendPacket(conn->securityObject, call, p);
6111 /* Since we're about to send SOME sort of packet to the peer, it's
6112 * safe to nuke any scheduled end-of-packets ack */
6113 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
6115 /* Actually send the packet, filling in more connection-specific fields */
6116 MUTEX_EXIT(&call->lock);
6117 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6118 rxi_SendPacket(call, conn, p, istack);
6119 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6120 MUTEX_ENTER(&call->lock);
6122 /* Update last send time for this call (for keep-alive
6123 * processing), and for the connection (so that we can discover
6124 * idle connections) */
6125 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6126 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6127 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6129 conn->lastSendTime = call->lastSendTime = clock_Sec();
6130 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6131 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6132 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6133 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6134 RX_ACK_PING_RESPONSE)))
6135 call->lastSendData = call->lastSendTime;
6139 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6140 * that things are fine. Also called periodically to guarantee that nothing
6141 * falls through the cracks (e.g. (error + dally) connections have keepalive
6142 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6144 * haveCTLock Set if calling from rxi_ReapConnections
6146 #ifdef RX_ENABLE_LOCKS
6148 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6149 #else /* RX_ENABLE_LOCKS */
6151 rxi_CheckCall(struct rx_call *call)
6152 #endif /* RX_ENABLE_LOCKS */
6154 struct rx_connection *conn = call->conn;
6156 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6157 afs_uint32 fudgeFactor;
6160 int idle_timeout = 0;
6161 afs_int32 clock_diff = 0;
6165 /* Large swings in the clock can have a significant impact on
6166 * the performance of RX call processing. Forward clock shifts
6167 * will result in premature event triggering or timeouts.
6168 * Backward shifts can result in calls not completing until
6169 * the clock catches up with the original start clock value.
6171 * If a backward clock shift of more than five minutes is noticed,
6172 * just fail the call.
6174 if (now < call->lastSendTime)
6175 clock_diff = call->lastSendTime - now;
6176 if (now < call->startWait)
6177 clock_diff = MAX(clock_diff, call->startWait - now);
6178 if (now < call->lastReceiveTime)
6179 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6180 if (clock_diff > 5 * 60)
6182 if (call->state == RX_STATE_ACTIVE)
6183 rxi_CallError(call, RX_CALL_TIMEOUT);
6187 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6188 if (call->flags & RX_CALL_TQ_BUSY) {
6189 /* Call is active and will be reset by rxi_Start if it's
6190 * in an error state.
6195 /* RTT + 8*MDEV, rounded up to the next second. */
6196 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6197 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6199 deadTime = conn->secondsUntilDead + fudgeFactor;
6200 /* These are computed to the second (+- 1 second). But that's
6201 * good enough for these values, which should be a significant
6202 * number of seconds. */
6203 if (now > (call->lastReceiveTime + deadTime)) {
6204 if (call->state == RX_STATE_ACTIVE) {
6206 #if defined(KERNEL) && defined(AFS_SUN5_ENV)
6208 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6209 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6210 ip_stack_t *ipst = ns->netstack_ip;
6212 ire = ire_cache_lookup(conn->peer->host
6213 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6215 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
6217 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6224 if (ire && ire->ire_max_frag > 0)
6225 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6227 #if defined(GLOBAL_NETSTACKID)
6231 #endif /* ADAPT_PMTU */
6232 cerror = RX_CALL_DEAD;
6235 #ifdef RX_ENABLE_LOCKS
6236 /* Cancel pending events */
6237 rxevent_Cancel(&call->delayedAckEvent, call,
6238 RX_CALL_REFCOUNT_DELAY);
6239 rxi_rto_cancel(call);
6240 rxevent_Cancel(&call->keepAliveEvent, call,
6241 RX_CALL_REFCOUNT_ALIVE);
6242 rxevent_Cancel(&call->growMTUEvent, call,
6243 RX_CALL_REFCOUNT_ALIVE);
6244 MUTEX_ENTER(&rx_refcnt_mutex);
6245 /* if rxi_FreeCall returns 1 it has freed the call */
6246 if (call->refCount == 0 &&
6247 rxi_FreeCall(call, haveCTLock))
6249 MUTEX_EXIT(&rx_refcnt_mutex);
6252 MUTEX_EXIT(&rx_refcnt_mutex);
6254 #else /* RX_ENABLE_LOCKS */
6255 rxi_FreeCall(call, 0);
6257 #endif /* RX_ENABLE_LOCKS */
6259 /* Non-active calls are destroyed if they are not responding
6260 * to pings; active calls are simply flagged in error, so the
6261 * attached process can die reasonably gracefully. */
6264 if (conn->idleDeadDetection) {
6265 if (conn->idleDeadTime) {
6266 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6270 /* see if we have a non-activity timeout */
6271 if (call->startWait && ((call->startWait + idleDeadTime) < now) &&
6272 (call->flags & RX_CALL_READER_WAIT)) {
6273 if (call->state == RX_STATE_ACTIVE) {
6274 cerror = RX_CALL_TIMEOUT;
6279 if (call->lastSendData && ((call->lastSendData + idleDeadTime) < now)) {
6280 if (call->state == RX_STATE_ACTIVE) {
6281 cerror = conn->service ? conn->service->idleDeadErr : RX_CALL_IDLE;
6289 if (conn->hardDeadTime) {
6290 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6293 /* see if we have a hard timeout */
6295 && (now > (hardDeadTime + call->startTime.sec))) {
6296 if (call->state == RX_STATE_ACTIVE)
6297 rxi_CallError(call, RX_CALL_TIMEOUT);
6302 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6303 call->lastReceiveTime) {
6304 int oldMTU = conn->peer->ifMTU;
6306 /* if we thought we could send more, perhaps things got worse */
6307 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6308 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6309 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6310 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6312 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6314 /* minimum capped in SetPeerMtu */
6315 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6318 conn->lastPacketSize = 0;
6320 /* needed so ResetCall doesn't clobber us. */
6321 call->MTU = conn->peer->ifMTU;
6323 /* if we never succeeded, let the error pass out as-is */
6324 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6325 cerror = conn->msgsizeRetryErr;
6328 rxi_CallError(call, cerror);
6333 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6334 void *dummy, int dummy2)
6336 struct rx_connection *conn = arg1;
6337 struct rx_header theader;
6338 char tbuffer[1 + sizeof(struct rx_header)];
6339 struct sockaddr_in taddr;
6342 struct iovec tmpiov[2];
6345 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6348 tp = &tbuffer[sizeof(struct rx_header)];
6349 taddr.sin_family = AF_INET;
6350 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6351 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6352 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6353 taddr.sin_len = sizeof(struct sockaddr_in);
6355 memset(&theader, 0, sizeof(theader));
6356 theader.epoch = htonl(999);
6358 theader.callNumber = 0;
6361 theader.type = RX_PACKET_TYPE_VERSION;
6362 theader.flags = RX_LAST_PACKET;
6363 theader.serviceId = 0;
6365 memcpy(tbuffer, &theader, sizeof(theader));
6366 memcpy(tp, &a, sizeof(a));
6367 tmpiov[0].iov_base = tbuffer;
6368 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6370 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6372 MUTEX_ENTER(&conn->conn_data_lock);
6373 MUTEX_ENTER(&rx_refcnt_mutex);
6374 /* Only reschedule ourselves if the connection would not be destroyed */
6375 if (conn->refCount <= 1) {
6376 rxevent_Put(conn->natKeepAliveEvent);
6377 conn->natKeepAliveEvent = NULL;
6378 MUTEX_EXIT(&rx_refcnt_mutex);
6379 MUTEX_EXIT(&conn->conn_data_lock);
6380 rx_DestroyConnection(conn); /* drop the reference for this */
6382 conn->refCount--; /* drop the reference for this */
6383 MUTEX_EXIT(&rx_refcnt_mutex);
6384 rxevent_Put(conn->natKeepAliveEvent);
6385 conn->natKeepAliveEvent = NULL;
6386 rxi_ScheduleNatKeepAliveEvent(conn);
6387 MUTEX_EXIT(&conn->conn_data_lock);
6392 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6394 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6395 struct clock when, now;
6396 clock_GetTime(&now);
6398 when.sec += conn->secondsUntilNatPing;
6399 MUTEX_ENTER(&rx_refcnt_mutex);
6400 conn->refCount++; /* hold a reference for this */
6401 MUTEX_EXIT(&rx_refcnt_mutex);
6402 conn->natKeepAliveEvent =
6403 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6408 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6410 MUTEX_ENTER(&conn->conn_data_lock);
6411 conn->secondsUntilNatPing = seconds;
6413 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6414 rxi_ScheduleNatKeepAliveEvent(conn);
6416 conn->flags |= RX_CONN_NAT_PING;
6418 MUTEX_EXIT(&conn->conn_data_lock);
6422 rxi_NatKeepAliveOn(struct rx_connection *conn)
6424 MUTEX_ENTER(&conn->conn_data_lock);
6425 /* if it's already attached */
6426 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6427 rxi_ScheduleNatKeepAliveEvent(conn);
6429 conn->flags |= RX_CONN_NAT_PING;
6430 MUTEX_EXIT(&conn->conn_data_lock);
6433 /* When a call is in progress, this routine is called occasionally to
6434 * make sure that some traffic has arrived (or been sent to) the peer.
6435 * If nothing has arrived in a reasonable amount of time, the call is
6436 * declared dead; if nothing has been sent for a while, we send a
6437 * keep-alive packet (if we're actually trying to keep the call alive)
6440 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6443 struct rx_call *call = arg1;
6444 struct rx_connection *conn;
6447 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6448 MUTEX_ENTER(&call->lock);
6450 if (event == call->keepAliveEvent) {
6451 rxevent_Put(call->keepAliveEvent);
6452 call->keepAliveEvent = NULL;
6457 #ifdef RX_ENABLE_LOCKS
6458 if (rxi_CheckCall(call, 0)) {
6459 MUTEX_EXIT(&call->lock);
6462 #else /* RX_ENABLE_LOCKS */
6463 if (rxi_CheckCall(call))
6465 #endif /* RX_ENABLE_LOCKS */
6467 /* Don't try to keep alive dallying calls */
6468 if (call->state == RX_STATE_DALLY) {
6469 MUTEX_EXIT(&call->lock);
6474 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6475 /* Don't try to send keepalives if there is unacknowledged data */
6476 /* the rexmit code should be good enough, this little hack
6477 * doesn't quite work XXX */
6478 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6480 rxi_ScheduleKeepAliveEvent(call);
6481 MUTEX_EXIT(&call->lock);
6484 /* Does what's on the nameplate. */
6486 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6488 struct rx_call *call = arg1;
6489 struct rx_connection *conn;
6491 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6492 MUTEX_ENTER(&call->lock);
6494 if (event == call->growMTUEvent) {
6495 rxevent_Put(call->growMTUEvent);
6496 call->growMTUEvent = NULL;
6499 #ifdef RX_ENABLE_LOCKS
6500 if (rxi_CheckCall(call, 0)) {
6501 MUTEX_EXIT(&call->lock);
6504 #else /* RX_ENABLE_LOCKS */
6505 if (rxi_CheckCall(call))
6507 #endif /* RX_ENABLE_LOCKS */
6509 /* Don't bother with dallying calls */
6510 if (call->state == RX_STATE_DALLY) {
6511 MUTEX_EXIT(&call->lock);
6518 * keep being scheduled, just don't do anything if we're at peak,
6519 * or we're not set up to be properly handled (idle timeout required)
6521 if ((conn->peer->maxPacketSize != 0) &&
6522 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6523 conn->idleDeadDetection)
6524 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6525 rxi_ScheduleGrowMTUEvent(call, 0);
6526 MUTEX_EXIT(&call->lock);
6530 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6532 if (!call->keepAliveEvent) {
6533 struct clock when, now;
6534 clock_GetTime(&now);
6536 when.sec += call->conn->secondsUntilPing;
6537 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6538 call->keepAliveEvent =
6539 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6544 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6546 if (!call->growMTUEvent) {
6547 struct clock when, now;
6549 clock_GetTime(&now);
6552 if (call->conn->secondsUntilPing)
6553 secs = (6*call->conn->secondsUntilPing)-1;
6555 if (call->conn->secondsUntilDead)
6556 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6560 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6561 call->growMTUEvent =
6562 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6566 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6568 rxi_KeepAliveOn(struct rx_call *call)
6570 /* Pretend last packet received was received now--i.e. if another
6571 * packet isn't received within the keep alive time, then the call
6572 * will die; Initialize last send time to the current time--even
6573 * if a packet hasn't been sent yet. This will guarantee that a
6574 * keep-alive is sent within the ping time */
6575 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6576 rxi_ScheduleKeepAliveEvent(call);
6580 * Solely in order that callers not need to include rx_call.h
6583 rx_KeepAliveOff(struct rx_call *call)
6585 rxi_KeepAliveOff(call);
6588 rx_KeepAliveOn(struct rx_call *call)
6590 rxi_KeepAliveOn(call);
6594 rxi_GrowMTUOn(struct rx_call *call)
6596 struct rx_connection *conn = call->conn;
6597 MUTEX_ENTER(&conn->conn_data_lock);
6598 conn->lastPingSizeSer = conn->lastPingSize = 0;
6599 MUTEX_EXIT(&conn->conn_data_lock);
6600 rxi_ScheduleGrowMTUEvent(call, 1);
6603 /* This routine is called to send connection abort messages
6604 * that have been delayed to throttle looping clients. */
6606 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6609 struct rx_connection *conn = arg1;
6612 struct rx_packet *packet;
6614 MUTEX_ENTER(&conn->conn_data_lock);
6615 rxevent_Put(conn->delayedAbortEvent);
6616 conn->delayedAbortEvent = NULL;
6617 error = htonl(conn->error);
6619 MUTEX_EXIT(&conn->conn_data_lock);
6620 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6623 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6624 RX_PACKET_TYPE_ABORT, (char *)&error,
6626 rxi_FreePacket(packet);
6630 /* This routine is called to send call abort messages
6631 * that have been delayed to throttle looping clients. */
6633 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6636 struct rx_call *call = arg1;
6639 struct rx_packet *packet;
6641 MUTEX_ENTER(&call->lock);
6642 rxevent_Put(call->delayedAbortEvent);
6643 call->delayedAbortEvent = NULL;
6644 error = htonl(call->error);
6646 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6649 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6650 (char *)&error, sizeof(error), 0);
6651 rxi_FreePacket(packet);
6653 MUTEX_EXIT(&call->lock);
6654 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6657 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6658 * seconds) to ask the client to authenticate itself. The routine
6659 * issues a challenge to the client, which is obtained from the
6660 * security object associated with the connection */
6662 rxi_ChallengeEvent(struct rxevent *event,
6663 void *arg0, void *arg1, int tries)
6665 struct rx_connection *conn = arg0;
6668 rxevent_Put(conn->challengeEvent);
6669 conn->challengeEvent = NULL;
6672 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6673 struct rx_packet *packet;
6674 struct clock when, now;
6677 /* We've failed to authenticate for too long.
6678 * Reset any calls waiting for authentication;
6679 * they are all in RX_STATE_PRECALL.
6683 MUTEX_ENTER(&conn->conn_call_lock);
6684 for (i = 0; i < RX_MAXCALLS; i++) {
6685 struct rx_call *call = conn->call[i];
6687 MUTEX_ENTER(&call->lock);
6688 if (call->state == RX_STATE_PRECALL) {
6689 rxi_CallError(call, RX_CALL_DEAD);
6690 rxi_SendCallAbort(call, NULL, 0, 0);
6692 MUTEX_EXIT(&call->lock);
6695 MUTEX_EXIT(&conn->conn_call_lock);
6699 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6701 /* If there's no packet available, do this later. */
6702 RXS_GetChallenge(conn->securityObject, conn, packet);
6703 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6704 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6705 rxi_FreePacket(packet);
6707 clock_GetTime(&now);
6709 when.sec += RX_CHALLENGE_TIMEOUT;
6710 conn->challengeEvent =
6711 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6716 /* Call this routine to start requesting the client to authenticate
6717 * itself. This will continue until authentication is established,
6718 * the call times out, or an invalid response is returned. The
6719 * security object associated with the connection is asked to create
6720 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6721 * defined earlier. */
6723 rxi_ChallengeOn(struct rx_connection *conn)
6725 if (!conn->challengeEvent) {
6726 RXS_CreateChallenge(conn->securityObject, conn);
6727 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6732 /* rxi_ComputeRoundTripTime is called with peer locked. */
6733 /* peer may be null */
6735 rxi_ComputeRoundTripTime(struct rx_packet *p,
6736 struct rx_ackPacket *ack,
6737 struct rx_call *call,
6738 struct rx_peer *peer,
6741 struct clock thisRtt, *sentp;
6745 /* If the ACK is delayed, then do nothing */
6746 if (ack->reason == RX_ACK_DELAY)
6749 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6750 * their RTT multiple times, so only include the RTT of the last packet
6752 if (p->flags & RX_JUMBO_PACKET)
6755 /* Use the serial number to determine which transmission the ACK is for,
6756 * and set the sent time to match this. If we have no serial number, then
6757 * only use the ACK for RTT calculations if the packet has not been
6761 serial = ntohl(ack->serial);
6763 if (serial == p->header.serial) {
6764 sentp = &p->timeSent;
6765 } else if (serial == p->firstSerial) {
6766 sentp = &p->firstSent;
6767 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6768 sentp = &p->firstSent;
6772 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6773 sentp = &p->firstSent;
6780 if (clock_Lt(&thisRtt, sentp))
6781 return; /* somebody set the clock back, don't count this time. */
6783 clock_Sub(&thisRtt, sentp);
6784 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6785 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6787 if (clock_IsZero(&thisRtt)) {
6789 * The actual round trip time is shorter than the
6790 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6791 * Since we can't tell which at the moment we will assume 1ms.
6793 thisRtt.usec = 1000;
6796 if (rx_stats_active) {
6797 MUTEX_ENTER(&rx_stats_mutex);
6798 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6799 rx_stats.minRtt = thisRtt;
6800 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6801 if (thisRtt.sec > 60) {
6802 MUTEX_EXIT(&rx_stats_mutex);
6803 return; /* somebody set the clock ahead */
6805 rx_stats.maxRtt = thisRtt;
6807 clock_Add(&rx_stats.totalRtt, &thisRtt);
6808 rx_atomic_inc(&rx_stats.nRttSamples);
6809 MUTEX_EXIT(&rx_stats_mutex);
6812 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6814 /* Apply VanJacobson round-trip estimations */
6819 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6820 * srtt is stored as fixed point with 3 bits after the binary
6821 * point (i.e., scaled by 8). The following magic is
6822 * equivalent to the smoothing algorithm in rfc793 with an
6823 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6824 * srtt'*8 = rtt + srtt*7
6825 * srtt'*8 = srtt*8 + rtt - srtt
6826 * srtt' = srtt + rtt/8 - srtt/8
6827 * srtt' = srtt + (rtt - srtt)/8
6830 delta = _8THMSEC(&thisRtt) - call->rtt;
6831 call->rtt += (delta >> 3);
6834 * We accumulate a smoothed rtt variance (actually, a smoothed
6835 * mean difference), then set the retransmit timer to smoothed
6836 * rtt + 4 times the smoothed variance (was 2x in van's original
6837 * paper, but 4x works better for me, and apparently for him as
6839 * rttvar is stored as
6840 * fixed point with 2 bits after the binary point (scaled by
6841 * 4). The following is equivalent to rfc793 smoothing with
6842 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6843 * rttvar'*4 = rttvar*3 + |delta|
6844 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6845 * rttvar' = rttvar + |delta|/4 - rttvar/4
6846 * rttvar' = rttvar + (|delta| - rttvar)/4
6847 * This replaces rfc793's wired-in beta.
6848 * dev*4 = dev*4 + (|actual - expected| - dev)
6854 delta -= (call->rtt_dev << 1);
6855 call->rtt_dev += (delta >> 3);
6857 /* I don't have a stored RTT so I start with this value. Since I'm
6858 * probably just starting a call, and will be pushing more data down
6859 * this, I expect congestion to increase rapidly. So I fudge a
6860 * little, and I set deviance to half the rtt. In practice,
6861 * deviance tends to approach something a little less than
6862 * half the smoothed rtt. */
6863 call->rtt = _8THMSEC(&thisRtt) + 8;
6864 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6866 /* the smoothed RTT time is RTT + 4*MDEV
6868 * We allow a user specified minimum to be set for this, to allow clamping
6869 * at a minimum value in the same way as TCP. In addition, we have to allow
6870 * for the possibility that this packet is answered by a delayed ACK, so we
6871 * add on a fixed 200ms to account for that timer expiring.
6874 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6875 rx_minPeerTimeout) + 200;
6876 clock_Zero(&call->rto);
6877 clock_Addmsec(&call->rto, rtt_timeout);
6879 /* Update the peer, so any new calls start with our values */
6880 peer->rtt_dev = call->rtt_dev;
6881 peer->rtt = call->rtt;
6883 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6884 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6888 /* Find all server connections that have not been active for a long time, and
6891 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
6894 struct clock now, when;
6895 clock_GetTime(&now);
6897 /* Find server connection structures that haven't been used for
6898 * greater than rx_idleConnectionTime */
6900 struct rx_connection **conn_ptr, **conn_end;
6901 int i, havecalls = 0;
6902 MUTEX_ENTER(&rx_connHashTable_lock);
6903 for (conn_ptr = &rx_connHashTable[0], conn_end =
6904 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6906 struct rx_connection *conn, *next;
6907 struct rx_call *call;
6911 for (conn = *conn_ptr; conn; conn = next) {
6912 /* XXX -- Shouldn't the connection be locked? */
6915 for (i = 0; i < RX_MAXCALLS; i++) {
6916 call = conn->call[i];
6920 code = MUTEX_TRYENTER(&call->lock);
6923 #ifdef RX_ENABLE_LOCKS
6924 result = rxi_CheckCall(call, 1);
6925 #else /* RX_ENABLE_LOCKS */
6926 result = rxi_CheckCall(call);
6927 #endif /* RX_ENABLE_LOCKS */
6928 MUTEX_EXIT(&call->lock);
6930 /* If CheckCall freed the call, it might
6931 * have destroyed the connection as well,
6932 * which screws up the linked lists.
6938 if (conn->type == RX_SERVER_CONNECTION) {
6939 /* This only actually destroys the connection if
6940 * there are no outstanding calls */
6941 MUTEX_ENTER(&conn->conn_data_lock);
6942 MUTEX_ENTER(&rx_refcnt_mutex);
6943 if (!havecalls && !conn->refCount
6944 && ((conn->lastSendTime + rx_idleConnectionTime) <
6946 conn->refCount++; /* it will be decr in rx_DestroyConn */
6947 MUTEX_EXIT(&rx_refcnt_mutex);
6948 MUTEX_EXIT(&conn->conn_data_lock);
6949 #ifdef RX_ENABLE_LOCKS
6950 rxi_DestroyConnectionNoLock(conn);
6951 #else /* RX_ENABLE_LOCKS */
6952 rxi_DestroyConnection(conn);
6953 #endif /* RX_ENABLE_LOCKS */
6955 #ifdef RX_ENABLE_LOCKS
6957 MUTEX_EXIT(&rx_refcnt_mutex);
6958 MUTEX_EXIT(&conn->conn_data_lock);
6960 #endif /* RX_ENABLE_LOCKS */
6964 #ifdef RX_ENABLE_LOCKS
6965 while (rx_connCleanup_list) {
6966 struct rx_connection *conn;
6967 conn = rx_connCleanup_list;
6968 rx_connCleanup_list = rx_connCleanup_list->next;
6969 MUTEX_EXIT(&rx_connHashTable_lock);
6970 rxi_CleanupConnection(conn);
6971 MUTEX_ENTER(&rx_connHashTable_lock);
6973 MUTEX_EXIT(&rx_connHashTable_lock);
6974 #endif /* RX_ENABLE_LOCKS */
6977 /* Find any peer structures that haven't been used (haven't had an
6978 * associated connection) for greater than rx_idlePeerTime */
6980 struct rx_peer **peer_ptr, **peer_end;
6984 * Why do we need to hold the rx_peerHashTable_lock across
6985 * the incrementing of peer_ptr since the rx_peerHashTable
6986 * array is not changing? We don't.
6988 * By dropping the lock periodically we can permit other
6989 * activities to be performed while a rxi_ReapConnections
6990 * call is in progress. The goal of reap connections
6991 * is to clean up quickly without causing large amounts
6992 * of contention. Therefore, it is important that global
6993 * mutexes not be held for extended periods of time.
6995 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6996 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6998 struct rx_peer *peer, *next, *prev;
7000 MUTEX_ENTER(&rx_peerHashTable_lock);
7001 for (prev = peer = *peer_ptr; peer; peer = next) {
7003 code = MUTEX_TRYENTER(&peer->peer_lock);
7004 if ((code) && (peer->refCount == 0)
7005 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7006 rx_interface_stat_p rpc_stat, nrpc_stat;
7010 * now know that this peer object is one to be
7011 * removed from the hash table. Once it is removed
7012 * it can't be referenced by other threads.
7013 * Lets remove it first and decrement the struct
7014 * nPeerStructs count.
7016 if (peer == *peer_ptr) {
7022 if (rx_stats_active)
7023 rx_atomic_dec(&rx_stats.nPeerStructs);
7026 * Now if we hold references on 'prev' and 'next'
7027 * we can safely drop the rx_peerHashTable_lock
7028 * while we destroy this 'peer' object.
7034 MUTEX_EXIT(&rx_peerHashTable_lock);
7036 MUTEX_EXIT(&peer->peer_lock);
7037 MUTEX_DESTROY(&peer->peer_lock);
7039 (&peer->rpcStats, rpc_stat, nrpc_stat,
7040 rx_interface_stat)) {
7041 unsigned int num_funcs;
7044 queue_Remove(&rpc_stat->queue_header);
7045 queue_Remove(&rpc_stat->all_peers);
7046 num_funcs = rpc_stat->stats[0].func_total;
7048 sizeof(rx_interface_stat_t) +
7049 rpc_stat->stats[0].func_total *
7050 sizeof(rx_function_entry_v1_t);
7052 rxi_Free(rpc_stat, space);
7054 MUTEX_ENTER(&rx_rpc_stats);
7055 rxi_rpc_peer_stat_cnt -= num_funcs;
7056 MUTEX_EXIT(&rx_rpc_stats);
7061 * Regain the rx_peerHashTable_lock and
7062 * decrement the reference count on 'prev'
7065 MUTEX_ENTER(&rx_peerHashTable_lock);
7072 MUTEX_EXIT(&peer->peer_lock);
7077 MUTEX_EXIT(&rx_peerHashTable_lock);
7081 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7082 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7083 * GC, just below. Really, we shouldn't have to keep moving packets from
7084 * one place to another, but instead ought to always know if we can
7085 * afford to hold onto a packet in its particular use. */
7086 MUTEX_ENTER(&rx_freePktQ_lock);
7087 if (rx_waitingForPackets) {
7088 rx_waitingForPackets = 0;
7089 #ifdef RX_ENABLE_LOCKS
7090 CV_BROADCAST(&rx_waitingForPackets_cv);
7092 osi_rxWakeup(&rx_waitingForPackets);
7095 MUTEX_EXIT(&rx_freePktQ_lock);
7098 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7099 rxevent_Put(rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0));
7103 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7104 * rx.h is sort of strange this is better. This is called with a security
7105 * object before it is discarded. Each connection using a security object has
7106 * its own refcount to the object so it won't actually be freed until the last
7107 * connection is destroyed.
7109 * This is the only rxs module call. A hold could also be written but no one
7113 rxs_Release(struct rx_securityClass *aobj)
7115 return RXS_Close(aobj);
7123 #define TRACE_OPTION_RX_DEBUG 16
7131 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7132 0, KEY_QUERY_VALUE, &parmKey);
7133 if (code != ERROR_SUCCESS)
7136 dummyLen = sizeof(TraceOption);
7137 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7138 (BYTE *) &TraceOption, &dummyLen);
7139 if (code == ERROR_SUCCESS) {
7140 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7142 RegCloseKey (parmKey);
7143 #endif /* AFS_NT40_ENV */
7148 rx_DebugOnOff(int on)
7152 rxdebug_active = on;
7158 rx_StatsOnOff(int on)
7160 rx_stats_active = on;
7164 /* Don't call this debugging routine directly; use dpf */
7166 rxi_DebugPrint(char *format, ...)
7175 va_start(ap, format);
7177 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7180 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7182 OutputDebugString(msg);
7188 va_start(ap, format);
7190 clock_GetTime(&now);
7191 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7192 (unsigned int)now.usec);
7193 vfprintf(rx_Log, format, ap);
7201 * This function is used to process the rx_stats structure that is local
7202 * to a process as well as an rx_stats structure received from a remote
7203 * process (via rxdebug). Therefore, it needs to do minimal version
7207 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7208 afs_int32 freePackets, char version)
7212 if (size != sizeof(struct rx_statistics)) {
7214 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7215 size, sizeof(struct rx_statistics));
7218 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7221 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7222 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7223 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7224 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7225 s->specialPktAllocFailures);
7227 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7228 s->receivePktAllocFailures, s->sendPktAllocFailures,
7229 s->specialPktAllocFailures);
7233 " greedy %u, " "bogusReads %u (last from host %x), "
7234 "noPackets %u, " "noBuffers %u, " "selects %u, "
7235 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7236 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7237 s->selects, s->sendSelects);
7239 fprintf(file, " packets read: ");
7240 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7241 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7243 fprintf(file, "\n");
7246 " other read counters: data %u, " "ack %u, " "dup %u "
7247 "spurious %u " "dally %u\n", s->dataPacketsRead,
7248 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7249 s->ignorePacketDally);
7251 fprintf(file, " packets sent: ");
7252 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7253 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7255 fprintf(file, "\n");
7258 " other send counters: ack %u, " "data %u (not resends), "
7259 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7260 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7261 s->dataPacketsPushed, s->ignoreAckedPacket);
7264 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7265 s->netSendFailures, (int)s->fatalErrors);
7267 if (s->nRttSamples) {
7268 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7269 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7271 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7272 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7276 " %d server connections, " "%d client connections, "
7277 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7278 s->nServerConns, s->nClientConns, s->nPeerStructs,
7279 s->nCallStructs, s->nFreeCallStructs);
7281 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7282 fprintf(file, " %d clock updates\n", clock_nUpdates);
7286 /* for backward compatibility */
7288 rx_PrintStats(FILE * file)
7290 MUTEX_ENTER(&rx_stats_mutex);
7291 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7292 sizeof(rx_stats), rx_nFreePackets,
7294 MUTEX_EXIT(&rx_stats_mutex);
7298 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7300 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
7301 ntohl(peer->host), (int)ntohs(peer->port), (int)peer->burstSize,
7302 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
7305 " Rtt %d, " "total sent %d, " "resent %d\n",
7306 peer->rtt, peer->nSent, peer->reSends);
7309 " Packet size %d, " "max in packet skew %d, "
7310 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
7311 (int)peer->outPacketSkew);
7315 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7317 * This mutex protects the following static variables:
7321 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7322 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7324 #define LOCK_RX_DEBUG
7325 #define UNLOCK_RX_DEBUG
7326 #endif /* AFS_PTHREAD_ENV */
7328 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7330 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7331 u_char type, void *inputData, size_t inputLength,
7332 void *outputData, size_t outputLength)
7334 static afs_int32 counter = 100;
7335 time_t waitTime, waitCount;
7336 struct rx_header theader;
7339 struct timeval tv_now, tv_wake, tv_delta;
7340 struct sockaddr_in taddr, faddr;
7354 tp = &tbuffer[sizeof(struct rx_header)];
7355 taddr.sin_family = AF_INET;
7356 taddr.sin_port = remotePort;
7357 taddr.sin_addr.s_addr = remoteAddr;
7358 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7359 taddr.sin_len = sizeof(struct sockaddr_in);
7362 memset(&theader, 0, sizeof(theader));
7363 theader.epoch = htonl(999);
7365 theader.callNumber = htonl(counter);
7368 theader.type = type;
7369 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7370 theader.serviceId = 0;
7372 memcpy(tbuffer, &theader, sizeof(theader));
7373 memcpy(tp, inputData, inputLength);
7375 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7376 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7378 /* see if there's a packet available */
7379 gettimeofday(&tv_wake, NULL);
7380 tv_wake.tv_sec += waitTime;
7383 FD_SET(socket, &imask);
7384 tv_delta.tv_sec = tv_wake.tv_sec;
7385 tv_delta.tv_usec = tv_wake.tv_usec;
7386 gettimeofday(&tv_now, NULL);
7388 if (tv_delta.tv_usec < tv_now.tv_usec) {
7390 tv_delta.tv_usec += 1000000;
7393 tv_delta.tv_usec -= tv_now.tv_usec;
7395 if (tv_delta.tv_sec < tv_now.tv_sec) {
7399 tv_delta.tv_sec -= tv_now.tv_sec;
7402 code = select(0, &imask, 0, 0, &tv_delta);
7403 #else /* AFS_NT40_ENV */
7404 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7405 #endif /* AFS_NT40_ENV */
7406 if (code == 1 && FD_ISSET(socket, &imask)) {
7407 /* now receive a packet */
7408 faddrLen = sizeof(struct sockaddr_in);
7410 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7411 (struct sockaddr *)&faddr, &faddrLen);
7414 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7415 if (counter == ntohl(theader.callNumber))
7423 /* see if we've timed out */
7431 code -= sizeof(struct rx_header);
7432 if (code > outputLength)
7433 code = outputLength;
7434 memcpy(outputData, tp, code);
7437 #endif /* RXDEBUG */
7440 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7441 afs_uint16 remotePort, struct rx_debugStats * stat,
7442 afs_uint32 * supportedValues)
7444 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7446 struct rx_debugIn in;
7448 *supportedValues = 0;
7449 in.type = htonl(RX_DEBUGI_GETSTATS);
7452 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7453 &in, sizeof(in), stat, sizeof(*stat));
7456 * If the call was successful, fixup the version and indicate
7457 * what contents of the stat structure are valid.
7458 * Also do net to host conversion of fields here.
7462 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7463 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7465 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7466 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7468 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7469 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7471 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7472 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7474 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7475 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7477 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7478 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7480 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7481 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7483 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7484 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7486 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7487 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7489 stat->nFreePackets = ntohl(stat->nFreePackets);
7490 stat->packetReclaims = ntohl(stat->packetReclaims);
7491 stat->callsExecuted = ntohl(stat->callsExecuted);
7492 stat->nWaiting = ntohl(stat->nWaiting);
7493 stat->idleThreads = ntohl(stat->idleThreads);
7494 stat->nWaited = ntohl(stat->nWaited);
7495 stat->nPackets = ntohl(stat->nPackets);
7504 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7505 afs_uint16 remotePort, struct rx_statistics * stat,
7506 afs_uint32 * supportedValues)
7508 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7510 struct rx_debugIn in;
7511 afs_int32 *lp = (afs_int32 *) stat;
7515 * supportedValues is currently unused, but added to allow future
7516 * versioning of this function.
7519 *supportedValues = 0;
7520 in.type = htonl(RX_DEBUGI_RXSTATS);
7522 memset(stat, 0, sizeof(*stat));
7524 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7525 &in, sizeof(in), stat, sizeof(*stat));
7530 * Do net to host conversion here
7533 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7544 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7545 afs_uint16 remotePort, size_t version_length,
7548 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7550 return MakeDebugCall(socket, remoteAddr, remotePort,
7551 RX_PACKET_TYPE_VERSION, a, 1, version,
7559 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7560 afs_uint16 remotePort, afs_int32 * nextConnection,
7561 int allConnections, afs_uint32 debugSupportedValues,
7562 struct rx_debugConn * conn,
7563 afs_uint32 * supportedValues)
7565 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7567 struct rx_debugIn in;
7571 * supportedValues is currently unused, but added to allow future
7572 * versioning of this function.
7575 *supportedValues = 0;
7576 if (allConnections) {
7577 in.type = htonl(RX_DEBUGI_GETALLCONN);
7579 in.type = htonl(RX_DEBUGI_GETCONN);
7581 in.index = htonl(*nextConnection);
7582 memset(conn, 0, sizeof(*conn));
7584 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7585 &in, sizeof(in), conn, sizeof(*conn));
7588 *nextConnection += 1;
7591 * Convert old connection format to new structure.
7594 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7595 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7596 #define MOVEvL(a) (conn->a = vL->a)
7598 /* any old or unrecognized version... */
7599 for (i = 0; i < RX_MAXCALLS; i++) {
7600 MOVEvL(callState[i]);
7601 MOVEvL(callMode[i]);
7602 MOVEvL(callFlags[i]);
7603 MOVEvL(callOther[i]);
7605 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7606 MOVEvL(secStats.type);
7607 MOVEvL(secStats.level);
7608 MOVEvL(secStats.flags);
7609 MOVEvL(secStats.expires);
7610 MOVEvL(secStats.packetsReceived);
7611 MOVEvL(secStats.packetsSent);
7612 MOVEvL(secStats.bytesReceived);
7613 MOVEvL(secStats.bytesSent);
7618 * Do net to host conversion here
7620 * I don't convert host or port since we are most likely
7621 * going to want these in NBO.
7623 conn->cid = ntohl(conn->cid);
7624 conn->serial = ntohl(conn->serial);
7625 for (i = 0; i < RX_MAXCALLS; i++) {
7626 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7628 conn->error = ntohl(conn->error);
7629 conn->secStats.flags = ntohl(conn->secStats.flags);
7630 conn->secStats.expires = ntohl(conn->secStats.expires);
7631 conn->secStats.packetsReceived =
7632 ntohl(conn->secStats.packetsReceived);
7633 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7634 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7635 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7636 conn->epoch = ntohl(conn->epoch);
7637 conn->natMTU = ntohl(conn->natMTU);
7646 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7647 afs_uint16 remotePort, afs_int32 * nextPeer,
7648 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7649 afs_uint32 * supportedValues)
7651 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7653 struct rx_debugIn in;
7656 * supportedValues is currently unused, but added to allow future
7657 * versioning of this function.
7660 *supportedValues = 0;
7661 in.type = htonl(RX_DEBUGI_GETPEER);
7662 in.index = htonl(*nextPeer);
7663 memset(peer, 0, sizeof(*peer));
7665 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7666 &in, sizeof(in), peer, sizeof(*peer));
7672 * Do net to host conversion here
7674 * I don't convert host or port since we are most likely
7675 * going to want these in NBO.
7677 peer->ifMTU = ntohs(peer->ifMTU);
7678 peer->idleWhen = ntohl(peer->idleWhen);
7679 peer->refCount = ntohs(peer->refCount);
7680 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7681 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7682 peer->rtt = ntohl(peer->rtt);
7683 peer->rtt_dev = ntohl(peer->rtt_dev);
7684 peer->timeout.sec = 0;
7685 peer->timeout.usec = 0;
7686 peer->nSent = ntohl(peer->nSent);
7687 peer->reSends = ntohl(peer->reSends);
7688 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7689 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7690 peer->natMTU = ntohs(peer->natMTU);
7691 peer->maxMTU = ntohs(peer->maxMTU);
7692 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7693 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7694 peer->MTU = ntohs(peer->MTU);
7695 peer->cwind = ntohs(peer->cwind);
7696 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7697 peer->congestSeq = ntohs(peer->congestSeq);
7698 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7699 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7700 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7701 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7710 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7711 struct rx_debugPeer * peerStats)
7714 afs_int32 error = 1; /* default to "did not succeed" */
7715 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7717 MUTEX_ENTER(&rx_peerHashTable_lock);
7718 for(tp = rx_peerHashTable[hashValue];
7719 tp != NULL; tp = tp->next) {
7720 if (tp->host == peerHost)
7726 MUTEX_EXIT(&rx_peerHashTable_lock);
7730 MUTEX_ENTER(&tp->peer_lock);
7731 peerStats->host = tp->host;
7732 peerStats->port = tp->port;
7733 peerStats->ifMTU = tp->ifMTU;
7734 peerStats->idleWhen = tp->idleWhen;
7735 peerStats->refCount = tp->refCount;
7736 peerStats->burstSize = tp->burstSize;
7737 peerStats->burst = tp->burst;
7738 peerStats->burstWait.sec = tp->burstWait.sec;
7739 peerStats->burstWait.usec = tp->burstWait.usec;
7740 peerStats->rtt = tp->rtt;
7741 peerStats->rtt_dev = tp->rtt_dev;
7742 peerStats->timeout.sec = 0;
7743 peerStats->timeout.usec = 0;
7744 peerStats->nSent = tp->nSent;
7745 peerStats->reSends = tp->reSends;
7746 peerStats->inPacketSkew = tp->inPacketSkew;
7747 peerStats->outPacketSkew = tp->outPacketSkew;
7748 peerStats->natMTU = tp->natMTU;
7749 peerStats->maxMTU = tp->maxMTU;
7750 peerStats->maxDgramPackets = tp->maxDgramPackets;
7751 peerStats->ifDgramPackets = tp->ifDgramPackets;
7752 peerStats->MTU = tp->MTU;
7753 peerStats->cwind = tp->cwind;
7754 peerStats->nDgramPackets = tp->nDgramPackets;
7755 peerStats->congestSeq = tp->congestSeq;
7756 peerStats->bytesSent.high = tp->bytesSent.high;
7757 peerStats->bytesSent.low = tp->bytesSent.low;
7758 peerStats->bytesReceived.high = tp->bytesReceived.high;
7759 peerStats->bytesReceived.low = tp->bytesReceived.low;
7760 MUTEX_EXIT(&tp->peer_lock);
7762 MUTEX_ENTER(&rx_peerHashTable_lock);
7765 MUTEX_EXIT(&rx_peerHashTable_lock);
7773 struct rx_serverQueueEntry *np;
7776 struct rx_call *call;
7777 struct rx_serverQueueEntry *sq;
7781 if (rxinit_status == 1) {
7783 return; /* Already shutdown. */
7787 #ifndef AFS_PTHREAD_ENV
7788 FD_ZERO(&rx_selectMask);
7789 #endif /* AFS_PTHREAD_ENV */
7790 rxi_dataQuota = RX_MAX_QUOTA;
7791 #ifndef AFS_PTHREAD_ENV
7793 #endif /* AFS_PTHREAD_ENV */
7796 #ifndef AFS_PTHREAD_ENV
7797 #ifndef AFS_USE_GETTIMEOFDAY
7799 #endif /* AFS_USE_GETTIMEOFDAY */
7800 #endif /* AFS_PTHREAD_ENV */
7802 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7803 call = queue_First(&rx_freeCallQueue, rx_call);
7805 rxi_Free(call, sizeof(struct rx_call));
7808 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7809 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7815 struct rx_peer **peer_ptr, **peer_end;
7816 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7817 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7819 struct rx_peer *peer, *next;
7821 MUTEX_ENTER(&rx_peerHashTable_lock);
7822 for (peer = *peer_ptr; peer; peer = next) {
7823 rx_interface_stat_p rpc_stat, nrpc_stat;
7826 MUTEX_ENTER(&rx_rpc_stats);
7827 MUTEX_ENTER(&peer->peer_lock);
7829 (&peer->rpcStats, rpc_stat, nrpc_stat,
7830 rx_interface_stat)) {
7831 unsigned int num_funcs;
7834 queue_Remove(&rpc_stat->queue_header);
7835 queue_Remove(&rpc_stat->all_peers);
7836 num_funcs = rpc_stat->stats[0].func_total;
7838 sizeof(rx_interface_stat_t) +
7839 rpc_stat->stats[0].func_total *
7840 sizeof(rx_function_entry_v1_t);
7842 rxi_Free(rpc_stat, space);
7844 /* rx_rpc_stats must be held */
7845 rxi_rpc_peer_stat_cnt -= num_funcs;
7847 MUTEX_EXIT(&peer->peer_lock);
7848 MUTEX_EXIT(&rx_rpc_stats);
7852 if (rx_stats_active)
7853 rx_atomic_dec(&rx_stats.nPeerStructs);
7855 MUTEX_EXIT(&rx_peerHashTable_lock);
7858 for (i = 0; i < RX_MAX_SERVICES; i++) {
7860 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7862 for (i = 0; i < rx_hashTableSize; i++) {
7863 struct rx_connection *tc, *ntc;
7864 MUTEX_ENTER(&rx_connHashTable_lock);
7865 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7867 for (j = 0; j < RX_MAXCALLS; j++) {
7869 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7872 rxi_Free(tc, sizeof(*tc));
7874 MUTEX_EXIT(&rx_connHashTable_lock);
7877 MUTEX_ENTER(&freeSQEList_lock);
7879 while ((np = rx_FreeSQEList)) {
7880 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7881 MUTEX_DESTROY(&np->lock);
7882 rxi_Free(np, sizeof(*np));
7885 MUTEX_EXIT(&freeSQEList_lock);
7886 MUTEX_DESTROY(&freeSQEList_lock);
7887 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7888 MUTEX_DESTROY(&rx_connHashTable_lock);
7889 MUTEX_DESTROY(&rx_peerHashTable_lock);
7890 MUTEX_DESTROY(&rx_serverPool_lock);
7892 osi_Free(rx_connHashTable,
7893 rx_hashTableSize * sizeof(struct rx_connection *));
7894 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7896 UNPIN(rx_connHashTable,
7897 rx_hashTableSize * sizeof(struct rx_connection *));
7898 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7900 rxi_FreeAllPackets();
7902 MUTEX_ENTER(&rx_quota_mutex);
7903 rxi_dataQuota = RX_MAX_QUOTA;
7904 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7905 MUTEX_EXIT(&rx_quota_mutex);
7910 #ifdef RX_ENABLE_LOCKS
7912 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7914 if (!MUTEX_ISMINE(lockaddr))
7915 osi_Panic("Lock not held: %s", msg);
7917 #endif /* RX_ENABLE_LOCKS */
7922 * Routines to implement connection specific data.
7926 rx_KeyCreate(rx_destructor_t rtn)
7929 MUTEX_ENTER(&rxi_keyCreate_lock);
7930 key = rxi_keyCreate_counter++;
7931 rxi_keyCreate_destructor = (rx_destructor_t *)
7932 realloc((void *)rxi_keyCreate_destructor,
7933 (key + 1) * sizeof(rx_destructor_t));
7934 rxi_keyCreate_destructor[key] = rtn;
7935 MUTEX_EXIT(&rxi_keyCreate_lock);
7940 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7943 MUTEX_ENTER(&conn->conn_data_lock);
7944 if (!conn->specific) {
7945 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7946 for (i = 0; i < key; i++)
7947 conn->specific[i] = NULL;
7948 conn->nSpecific = key + 1;
7949 conn->specific[key] = ptr;
7950 } else if (key >= conn->nSpecific) {
7951 conn->specific = (void **)
7952 realloc(conn->specific, (key + 1) * sizeof(void *));
7953 for (i = conn->nSpecific; i < key; i++)
7954 conn->specific[i] = NULL;
7955 conn->nSpecific = key + 1;
7956 conn->specific[key] = ptr;
7958 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7959 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7960 conn->specific[key] = ptr;
7962 MUTEX_EXIT(&conn->conn_data_lock);
7966 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7969 MUTEX_ENTER(&svc->svc_data_lock);
7970 if (!svc->specific) {
7971 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
7972 for (i = 0; i < key; i++)
7973 svc->specific[i] = NULL;
7974 svc->nSpecific = key + 1;
7975 svc->specific[key] = ptr;
7976 } else if (key >= svc->nSpecific) {
7977 svc->specific = (void **)
7978 realloc(svc->specific, (key + 1) * sizeof(void *));
7979 for (i = svc->nSpecific; i < key; i++)
7980 svc->specific[i] = NULL;
7981 svc->nSpecific = key + 1;
7982 svc->specific[key] = ptr;
7984 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7985 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7986 svc->specific[key] = ptr;
7988 MUTEX_EXIT(&svc->svc_data_lock);
7992 rx_GetSpecific(struct rx_connection *conn, int key)
7995 MUTEX_ENTER(&conn->conn_data_lock);
7996 if (key >= conn->nSpecific)
7999 ptr = conn->specific[key];
8000 MUTEX_EXIT(&conn->conn_data_lock);
8005 rx_GetServiceSpecific(struct rx_service *svc, int key)
8008 MUTEX_ENTER(&svc->svc_data_lock);
8009 if (key >= svc->nSpecific)
8012 ptr = svc->specific[key];
8013 MUTEX_EXIT(&svc->svc_data_lock);
8018 #endif /* !KERNEL */
8021 * processStats is a queue used to store the statistics for the local
8022 * process. Its contents are similar to the contents of the rpcStats
8023 * queue on a rx_peer structure, but the actual data stored within
8024 * this queue contains totals across the lifetime of the process (assuming
8025 * the stats have not been reset) - unlike the per peer structures
8026 * which can come and go based upon the peer lifetime.
8029 static struct rx_queue processStats = { &processStats, &processStats };
8032 * peerStats is a queue used to store the statistics for all peer structs.
8033 * Its contents are the union of all the peer rpcStats queues.
8036 static struct rx_queue peerStats = { &peerStats, &peerStats };
8039 * rxi_monitor_processStats is used to turn process wide stat collection
8043 static int rxi_monitor_processStats = 0;
8046 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8049 static int rxi_monitor_peerStats = 0;
8052 * rxi_AddRpcStat - given all of the information for a particular rpc
8053 * call, create (if needed) and update the stat totals for the rpc.
8057 * IN stats - the queue of stats that will be updated with the new value
8059 * IN rxInterface - a unique number that identifies the rpc interface
8061 * IN currentFunc - the index of the function being invoked
8063 * IN totalFunc - the total number of functions in this interface
8065 * IN queueTime - the amount of time this function waited for a thread
8067 * IN execTime - the amount of time this function invocation took to execute
8069 * IN bytesSent - the number bytes sent by this invocation
8071 * IN bytesRcvd - the number bytes received by this invocation
8073 * IN isServer - if true, this invocation was made to a server
8075 * IN remoteHost - the ip address of the remote host
8077 * IN remotePort - the port of the remote host
8079 * IN addToPeerList - if != 0, add newly created stat to the global peer list
8081 * INOUT counter - if a new stats structure is allocated, the counter will
8082 * be updated with the new number of allocated stat structures
8090 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
8091 afs_uint32 currentFunc, afs_uint32 totalFunc,
8092 struct clock *queueTime, struct clock *execTime,
8093 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
8094 afs_uint32 remoteHost, afs_uint32 remotePort,
8095 int addToPeerList, unsigned int *counter)
8098 rx_interface_stat_p rpc_stat, nrpc_stat;
8101 * See if there's already a structure for this interface
8104 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8105 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8106 && (rpc_stat->stats[0].remote_is_server == isServer))
8111 * Didn't find a match so allocate a new structure and add it to the
8115 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
8116 || (rpc_stat->stats[0].interfaceId != rxInterface)
8117 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8122 sizeof(rx_interface_stat_t) +
8123 totalFunc * sizeof(rx_function_entry_v1_t);
8125 rpc_stat = rxi_Alloc(space);
8126 if (rpc_stat == NULL) {
8130 *counter += totalFunc;
8131 for (i = 0; i < totalFunc; i++) {
8132 rpc_stat->stats[i].remote_peer = remoteHost;
8133 rpc_stat->stats[i].remote_port = remotePort;
8134 rpc_stat->stats[i].remote_is_server = isServer;
8135 rpc_stat->stats[i].interfaceId = rxInterface;
8136 rpc_stat->stats[i].func_total = totalFunc;
8137 rpc_stat->stats[i].func_index = i;
8138 hzero(rpc_stat->stats[i].invocations);
8139 hzero(rpc_stat->stats[i].bytes_sent);
8140 hzero(rpc_stat->stats[i].bytes_rcvd);
8141 rpc_stat->stats[i].queue_time_sum.sec = 0;
8142 rpc_stat->stats[i].queue_time_sum.usec = 0;
8143 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8144 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8145 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8146 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8147 rpc_stat->stats[i].queue_time_max.sec = 0;
8148 rpc_stat->stats[i].queue_time_max.usec = 0;
8149 rpc_stat->stats[i].execution_time_sum.sec = 0;
8150 rpc_stat->stats[i].execution_time_sum.usec = 0;
8151 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8152 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8153 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8154 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8155 rpc_stat->stats[i].execution_time_max.sec = 0;
8156 rpc_stat->stats[i].execution_time_max.usec = 0;
8158 queue_Prepend(stats, rpc_stat);
8159 if (addToPeerList) {
8160 queue_Prepend(&peerStats, &rpc_stat->all_peers);
8165 * Increment the stats for this function
8168 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
8169 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
8170 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
8171 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8172 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8173 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8174 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8176 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8177 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8179 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8180 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8182 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8183 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8185 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8186 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8194 * rx_IncrementTimeAndCount - increment the times and count for a particular
8199 * IN peer - the peer who invoked the rpc
8201 * IN rxInterface - a unique number that identifies the rpc interface
8203 * IN currentFunc - the index of the function being invoked
8205 * IN totalFunc - the total number of functions in this interface
8207 * IN queueTime - the amount of time this function waited for a thread
8209 * IN execTime - the amount of time this function invocation took to execute
8211 * IN bytesSent - the number bytes sent by this invocation
8213 * IN bytesRcvd - the number bytes received by this invocation
8215 * IN isServer - if true, this invocation was made to a server
8223 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8224 afs_uint32 currentFunc, afs_uint32 totalFunc,
8225 struct clock *queueTime, struct clock *execTime,
8226 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8230 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8233 MUTEX_ENTER(&rx_rpc_stats);
8235 if (rxi_monitor_peerStats) {
8236 MUTEX_ENTER(&peer->peer_lock);
8237 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8238 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8239 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8240 MUTEX_EXIT(&peer->peer_lock);
8243 if (rxi_monitor_processStats) {
8244 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8245 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8246 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8249 MUTEX_EXIT(&rx_rpc_stats);
8254 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8258 * IN callerVersion - the rpc stat version of the caller.
8260 * IN count - the number of entries to marshall.
8262 * IN stats - pointer to stats to be marshalled.
8264 * OUT ptr - Where to store the marshalled data.
8271 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8272 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8278 * We only support the first version
8280 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8281 *(ptr++) = stats->remote_peer;
8282 *(ptr++) = stats->remote_port;
8283 *(ptr++) = stats->remote_is_server;
8284 *(ptr++) = stats->interfaceId;
8285 *(ptr++) = stats->func_total;
8286 *(ptr++) = stats->func_index;
8287 *(ptr++) = hgethi(stats->invocations);
8288 *(ptr++) = hgetlo(stats->invocations);
8289 *(ptr++) = hgethi(stats->bytes_sent);
8290 *(ptr++) = hgetlo(stats->bytes_sent);
8291 *(ptr++) = hgethi(stats->bytes_rcvd);
8292 *(ptr++) = hgetlo(stats->bytes_rcvd);
8293 *(ptr++) = stats->queue_time_sum.sec;
8294 *(ptr++) = stats->queue_time_sum.usec;
8295 *(ptr++) = stats->queue_time_sum_sqr.sec;
8296 *(ptr++) = stats->queue_time_sum_sqr.usec;
8297 *(ptr++) = stats->queue_time_min.sec;
8298 *(ptr++) = stats->queue_time_min.usec;
8299 *(ptr++) = stats->queue_time_max.sec;
8300 *(ptr++) = stats->queue_time_max.usec;
8301 *(ptr++) = stats->execution_time_sum.sec;
8302 *(ptr++) = stats->execution_time_sum.usec;
8303 *(ptr++) = stats->execution_time_sum_sqr.sec;
8304 *(ptr++) = stats->execution_time_sum_sqr.usec;
8305 *(ptr++) = stats->execution_time_min.sec;
8306 *(ptr++) = stats->execution_time_min.usec;
8307 *(ptr++) = stats->execution_time_max.sec;
8308 *(ptr++) = stats->execution_time_max.usec;
8314 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8319 * IN callerVersion - the rpc stat version of the caller
8321 * OUT myVersion - the rpc stat version of this function
8323 * OUT clock_sec - local time seconds
8325 * OUT clock_usec - local time microseconds
8327 * OUT allocSize - the number of bytes allocated to contain stats
8329 * OUT statCount - the number stats retrieved from this process.
8331 * OUT stats - the actual stats retrieved from this process.
8335 * Returns void. If successful, stats will != NULL.
8339 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8340 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8341 size_t * allocSize, afs_uint32 * statCount,
8342 afs_uint32 ** stats)
8352 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8355 * Check to see if stats are enabled
8358 MUTEX_ENTER(&rx_rpc_stats);
8359 if (!rxi_monitor_processStats) {
8360 MUTEX_EXIT(&rx_rpc_stats);
8364 clock_GetTime(&now);
8365 *clock_sec = now.sec;
8366 *clock_usec = now.usec;
8369 * Allocate the space based upon the caller version
8371 * If the client is at an older version than we are,
8372 * we return the statistic data in the older data format, but
8373 * we still return our version number so the client knows we
8374 * are maintaining more data than it can retrieve.
8377 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8378 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8379 *statCount = rxi_rpc_process_stat_cnt;
8382 * This can't happen yet, but in the future version changes
8383 * can be handled by adding additional code here
8387 if (space > (size_t) 0) {
8389 ptr = *stats = rxi_Alloc(space);
8392 rx_interface_stat_p rpc_stat, nrpc_stat;
8396 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8398 * Copy the data based upon the caller version
8400 rx_MarshallProcessRPCStats(callerVersion,
8401 rpc_stat->stats[0].func_total,
8402 rpc_stat->stats, &ptr);
8408 MUTEX_EXIT(&rx_rpc_stats);
8413 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8417 * IN callerVersion - the rpc stat version of the caller
8419 * OUT myVersion - the rpc stat version of this function
8421 * OUT clock_sec - local time seconds
8423 * OUT clock_usec - local time microseconds
8425 * OUT allocSize - the number of bytes allocated to contain stats
8427 * OUT statCount - the number of stats retrieved from the individual
8430 * OUT stats - the actual stats retrieved from the individual peer structures.
8434 * Returns void. If successful, stats will != NULL.
8438 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8439 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8440 size_t * allocSize, afs_uint32 * statCount,
8441 afs_uint32 ** stats)
8451 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8454 * Check to see if stats are enabled
8457 MUTEX_ENTER(&rx_rpc_stats);
8458 if (!rxi_monitor_peerStats) {
8459 MUTEX_EXIT(&rx_rpc_stats);
8463 clock_GetTime(&now);
8464 *clock_sec = now.sec;
8465 *clock_usec = now.usec;
8468 * Allocate the space based upon the caller version
8470 * If the client is at an older version than we are,
8471 * we return the statistic data in the older data format, but
8472 * we still return our version number so the client knows we
8473 * are maintaining more data than it can retrieve.
8476 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8477 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8478 *statCount = rxi_rpc_peer_stat_cnt;
8481 * This can't happen yet, but in the future version changes
8482 * can be handled by adding additional code here
8486 if (space > (size_t) 0) {
8488 ptr = *stats = rxi_Alloc(space);
8491 rx_interface_stat_p rpc_stat, nrpc_stat;
8495 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8497 * We have to fix the offset of rpc_stat since we are
8498 * keeping this structure on two rx_queues. The rx_queue
8499 * package assumes that the rx_queue member is the first
8500 * member of the structure. That is, rx_queue assumes that
8501 * any one item is only on one queue at a time. We are
8502 * breaking that assumption and so we have to do a little
8503 * math to fix our pointers.
8506 fix_offset = (char *)rpc_stat;
8507 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8508 rpc_stat = (rx_interface_stat_p) fix_offset;
8511 * Copy the data based upon the caller version
8513 rx_MarshallProcessRPCStats(callerVersion,
8514 rpc_stat->stats[0].func_total,
8515 rpc_stat->stats, &ptr);
8521 MUTEX_EXIT(&rx_rpc_stats);
8526 * rx_FreeRPCStats - free memory allocated by
8527 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8531 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8532 * rx_RetrievePeerRPCStats
8534 * IN allocSize - the number of bytes in stats.
8542 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8544 rxi_Free(stats, allocSize);
8548 * rx_queryProcessRPCStats - see if process rpc stat collection is
8549 * currently enabled.
8555 * Returns 0 if stats are not enabled != 0 otherwise
8559 rx_queryProcessRPCStats(void)
8562 MUTEX_ENTER(&rx_rpc_stats);
8563 rc = rxi_monitor_processStats;
8564 MUTEX_EXIT(&rx_rpc_stats);
8569 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8575 * Returns 0 if stats are not enabled != 0 otherwise
8579 rx_queryPeerRPCStats(void)
8582 MUTEX_ENTER(&rx_rpc_stats);
8583 rc = rxi_monitor_peerStats;
8584 MUTEX_EXIT(&rx_rpc_stats);
8589 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8599 rx_enableProcessRPCStats(void)
8601 MUTEX_ENTER(&rx_rpc_stats);
8602 rx_enable_stats = 1;
8603 rxi_monitor_processStats = 1;
8604 MUTEX_EXIT(&rx_rpc_stats);
8608 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8618 rx_enablePeerRPCStats(void)
8620 MUTEX_ENTER(&rx_rpc_stats);
8621 rx_enable_stats = 1;
8622 rxi_monitor_peerStats = 1;
8623 MUTEX_EXIT(&rx_rpc_stats);
8627 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8637 rx_disableProcessRPCStats(void)
8639 rx_interface_stat_p rpc_stat, nrpc_stat;
8642 MUTEX_ENTER(&rx_rpc_stats);
8645 * Turn off process statistics and if peer stats is also off, turn
8649 rxi_monitor_processStats = 0;
8650 if (rxi_monitor_peerStats == 0) {
8651 rx_enable_stats = 0;
8654 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8655 unsigned int num_funcs = 0;
8658 queue_Remove(rpc_stat);
8659 num_funcs = rpc_stat->stats[0].func_total;
8661 sizeof(rx_interface_stat_t) +
8662 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8664 rxi_Free(rpc_stat, space);
8665 rxi_rpc_process_stat_cnt -= num_funcs;
8667 MUTEX_EXIT(&rx_rpc_stats);
8671 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8681 rx_disablePeerRPCStats(void)
8683 struct rx_peer **peer_ptr, **peer_end;
8687 * Turn off peer statistics and if process stats is also off, turn
8691 rxi_monitor_peerStats = 0;
8692 if (rxi_monitor_processStats == 0) {
8693 rx_enable_stats = 0;
8696 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8697 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8699 struct rx_peer *peer, *next, *prev;
8701 MUTEX_ENTER(&rx_peerHashTable_lock);
8702 MUTEX_ENTER(&rx_rpc_stats);
8703 for (prev = peer = *peer_ptr; peer; peer = next) {
8705 code = MUTEX_TRYENTER(&peer->peer_lock);
8707 rx_interface_stat_p rpc_stat, nrpc_stat;
8710 if (prev == *peer_ptr) {
8721 MUTEX_EXIT(&rx_peerHashTable_lock);
8724 (&peer->rpcStats, rpc_stat, nrpc_stat,
8725 rx_interface_stat)) {
8726 unsigned int num_funcs = 0;
8729 queue_Remove(&rpc_stat->queue_header);
8730 queue_Remove(&rpc_stat->all_peers);
8731 num_funcs = rpc_stat->stats[0].func_total;
8733 sizeof(rx_interface_stat_t) +
8734 rpc_stat->stats[0].func_total *
8735 sizeof(rx_function_entry_v1_t);
8737 rxi_Free(rpc_stat, space);
8738 rxi_rpc_peer_stat_cnt -= num_funcs;
8740 MUTEX_EXIT(&peer->peer_lock);
8742 MUTEX_ENTER(&rx_peerHashTable_lock);
8752 MUTEX_EXIT(&rx_rpc_stats);
8753 MUTEX_EXIT(&rx_peerHashTable_lock);
8758 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8763 * IN clearFlag - flag indicating which stats to clear
8771 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8773 rx_interface_stat_p rpc_stat, nrpc_stat;
8775 MUTEX_ENTER(&rx_rpc_stats);
8777 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8778 unsigned int num_funcs = 0, i;
8779 num_funcs = rpc_stat->stats[0].func_total;
8780 for (i = 0; i < num_funcs; i++) {
8781 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8782 hzero(rpc_stat->stats[i].invocations);
8784 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8785 hzero(rpc_stat->stats[i].bytes_sent);
8787 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8788 hzero(rpc_stat->stats[i].bytes_rcvd);
8790 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8791 rpc_stat->stats[i].queue_time_sum.sec = 0;
8792 rpc_stat->stats[i].queue_time_sum.usec = 0;
8794 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8795 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8796 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8798 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8799 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8800 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8802 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8803 rpc_stat->stats[i].queue_time_max.sec = 0;
8804 rpc_stat->stats[i].queue_time_max.usec = 0;
8806 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8807 rpc_stat->stats[i].execution_time_sum.sec = 0;
8808 rpc_stat->stats[i].execution_time_sum.usec = 0;
8810 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8811 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8812 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8814 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8815 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8816 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8818 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8819 rpc_stat->stats[i].execution_time_max.sec = 0;
8820 rpc_stat->stats[i].execution_time_max.usec = 0;
8825 MUTEX_EXIT(&rx_rpc_stats);
8829 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8834 * IN clearFlag - flag indicating which stats to clear
8842 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8844 rx_interface_stat_p rpc_stat, nrpc_stat;
8846 MUTEX_ENTER(&rx_rpc_stats);
8848 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8849 unsigned int num_funcs = 0, i;
8852 * We have to fix the offset of rpc_stat since we are
8853 * keeping this structure on two rx_queues. The rx_queue
8854 * package assumes that the rx_queue member is the first
8855 * member of the structure. That is, rx_queue assumes that
8856 * any one item is only on one queue at a time. We are
8857 * breaking that assumption and so we have to do a little
8858 * math to fix our pointers.
8861 fix_offset = (char *)rpc_stat;
8862 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8863 rpc_stat = (rx_interface_stat_p) fix_offset;
8865 num_funcs = rpc_stat->stats[0].func_total;
8866 for (i = 0; i < num_funcs; i++) {
8867 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8868 hzero(rpc_stat->stats[i].invocations);
8870 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8871 hzero(rpc_stat->stats[i].bytes_sent);
8873 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8874 hzero(rpc_stat->stats[i].bytes_rcvd);
8876 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8877 rpc_stat->stats[i].queue_time_sum.sec = 0;
8878 rpc_stat->stats[i].queue_time_sum.usec = 0;
8880 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8881 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8882 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8884 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8885 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8886 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8888 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8889 rpc_stat->stats[i].queue_time_max.sec = 0;
8890 rpc_stat->stats[i].queue_time_max.usec = 0;
8892 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8893 rpc_stat->stats[i].execution_time_sum.sec = 0;
8894 rpc_stat->stats[i].execution_time_sum.usec = 0;
8896 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8897 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8898 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8900 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8901 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8902 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8904 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8905 rpc_stat->stats[i].execution_time_max.sec = 0;
8906 rpc_stat->stats[i].execution_time_max.usec = 0;
8911 MUTEX_EXIT(&rx_rpc_stats);
8915 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8916 * is authorized to enable/disable/clear RX statistics.
8918 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8921 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8923 rxi_rxstat_userok = proc;
8927 rx_RxStatUserOk(struct rx_call *call)
8929 if (!rxi_rxstat_userok)
8931 return rxi_rxstat_userok(call);
8936 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8937 * function in the MSVC runtime DLL (msvcrt.dll).
8939 * Note: the system serializes calls to this function.
8942 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8943 DWORD reason, /* reason function is being called */
8944 LPVOID reserved) /* reserved for future use */
8947 case DLL_PROCESS_ATTACH:
8948 /* library is being attached to a process */
8952 case DLL_PROCESS_DETACH:
8959 #endif /* AFS_NT40_ENV */
8962 int rx_DumpCalls(FILE *outputFile, char *cookie)
8964 #ifdef RXDEBUG_PACKET
8965 #ifdef KDUMP_RX_LOCK
8966 struct rx_call_rx_lock *c;
8973 #define RXDPRINTF sprintf
8974 #define RXDPRINTOUT output
8976 #define RXDPRINTF fprintf
8977 #define RXDPRINTOUT outputFile
8980 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8982 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8985 for (c = rx_allCallsp; c; c = c->allNextp) {
8986 u_short rqc, tqc, iovqc;
8987 struct rx_packet *p, *np;
8989 MUTEX_ENTER(&c->lock);
8990 queue_Count(&c->rq, p, np, rx_packet, rqc);
8991 queue_Count(&c->tq, p, np, rx_packet, tqc);
8992 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8994 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, "
8995 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8996 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8997 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8998 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8999 #ifdef RX_ENABLE_LOCKS
9002 #ifdef RX_REFCOUNT_CHECK
9003 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9004 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9007 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,
9008 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9009 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9010 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9011 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9012 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
9013 #ifdef RX_ENABLE_LOCKS
9014 , (afs_uint32)c->refCount
9016 #ifdef RX_REFCOUNT_CHECK
9017 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9020 MUTEX_EXIT(&c->lock);
9023 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9026 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9028 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9030 #endif /* RXDEBUG_PACKET */