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.
2682 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2684 int channel = call->channel;
2685 struct rx_connection *conn = call->conn;
2688 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2689 (*call->callNumber)++;
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 MUTEX_ENTER(&conn->conn_call_lock);
2703 if (call->conn->call[channel] == call)
2704 call->conn->call[channel] = 0;
2705 MUTEX_EXIT(&conn->conn_call_lock);
2707 MUTEX_ENTER(&rx_freeCallQueue_lock);
2708 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2709 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2710 /* A call may be free even though its transmit queue is still in use.
2711 * Since we search the call list from head to tail, put busy calls at
2712 * the head of the list, and idle calls at the tail.
2714 if (call->flags & RX_CALL_TQ_BUSY)
2715 queue_Prepend(&rx_freeCallQueue, call);
2717 queue_Append(&rx_freeCallQueue, call);
2718 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2719 queue_Append(&rx_freeCallQueue, call);
2720 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2721 if (rx_stats_active)
2722 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2723 MUTEX_EXIT(&rx_freeCallQueue_lock);
2725 /* Destroy the connection if it was previously slated for
2726 * destruction, i.e. the Rx client code previously called
2727 * rx_DestroyConnection (client connections), or
2728 * rxi_ReapConnections called the same routine (server
2729 * connections). Only do this, however, if there are no
2730 * outstanding calls. Note that for fine grain locking, there appears
2731 * to be a deadlock in that rxi_FreeCall has a call locked and
2732 * DestroyConnectionNoLock locks each call in the conn. But note a
2733 * few lines up where we have removed this call from the conn.
2734 * If someone else destroys a connection, they either have no
2735 * call lock held or are going through this section of code.
2737 MUTEX_ENTER(&conn->conn_data_lock);
2738 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2739 MUTEX_ENTER(&rx_refcnt_mutex);
2741 MUTEX_EXIT(&rx_refcnt_mutex);
2742 MUTEX_EXIT(&conn->conn_data_lock);
2743 #ifdef RX_ENABLE_LOCKS
2745 rxi_DestroyConnectionNoLock(conn);
2747 rxi_DestroyConnection(conn);
2748 #else /* RX_ENABLE_LOCKS */
2749 rxi_DestroyConnection(conn);
2750 #endif /* RX_ENABLE_LOCKS */
2752 MUTEX_EXIT(&conn->conn_data_lock);
2754 MUTEX_ENTER(&rx_refcnt_mutex);
2757 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2758 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2761 rxi_Alloc(size_t size)
2765 if (rx_stats_active) {
2766 rx_atomic_add(&rxi_Allocsize, (int) size);
2767 rx_atomic_inc(&rxi_Alloccnt);
2771 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2772 afs_osi_Alloc_NoSleep(size);
2777 osi_Panic("rxi_Alloc error");
2783 rxi_Free(void *addr, size_t size)
2785 if (rx_stats_active) {
2786 rx_atomic_sub(&rxi_Allocsize, (int) size);
2787 rx_atomic_dec(&rxi_Alloccnt);
2789 osi_Free(addr, size);
2793 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2795 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2796 struct rx_peer *next = NULL;
2800 MUTEX_ENTER(&rx_peerHashTable_lock);
2802 peer_ptr = &rx_peerHashTable[0];
2803 peer_end = &rx_peerHashTable[rx_hashTableSize];
2806 for ( ; peer_ptr < peer_end; peer_ptr++) {
2809 for ( ; peer; peer = next) {
2811 if (host == peer->host)
2816 hashIndex = PEER_HASH(host, port);
2817 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2818 if ((peer->host == host) && (peer->port == port))
2823 MUTEX_ENTER(&rx_peerHashTable_lock);
2828 MUTEX_EXIT(&rx_peerHashTable_lock);
2830 MUTEX_ENTER(&peer->peer_lock);
2831 /* We don't handle dropping below min, so don't */
2832 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2833 peer->ifMTU=MIN(mtu, peer->ifMTU);
2834 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2835 /* if we tweaked this down, need to tune our peer MTU too */
2836 peer->MTU = MIN(peer->MTU, peer->natMTU);
2837 /* if we discovered a sub-1500 mtu, degrade */
2838 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2839 peer->maxDgramPackets = 1;
2840 /* We no longer have valid peer packet information */
2841 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2842 peer->maxPacketSize = 0;
2843 MUTEX_EXIT(&peer->peer_lock);
2845 MUTEX_ENTER(&rx_peerHashTable_lock);
2847 if (host && !port) {
2849 /* pick up where we left off */
2853 MUTEX_EXIT(&rx_peerHashTable_lock);
2856 /* Find the peer process represented by the supplied (host,port)
2857 * combination. If there is no appropriate active peer structure, a
2858 * new one will be allocated and initialized
2859 * The origPeer, if set, is a pointer to a peer structure on which the
2860 * refcount will be be decremented. This is used to replace the peer
2861 * structure hanging off a connection structure */
2863 rxi_FindPeer(afs_uint32 host, u_short port,
2864 struct rx_peer *origPeer, int create)
2868 hashIndex = PEER_HASH(host, port);
2869 MUTEX_ENTER(&rx_peerHashTable_lock);
2870 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2871 if ((pp->host == host) && (pp->port == port))
2876 pp = rxi_AllocPeer(); /* This bzero's *pp */
2877 pp->host = host; /* set here or in InitPeerParams is zero */
2879 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2880 queue_Init(&pp->congestionQueue);
2881 queue_Init(&pp->rpcStats);
2882 pp->next = rx_peerHashTable[hashIndex];
2883 rx_peerHashTable[hashIndex] = pp;
2884 rxi_InitPeerParams(pp);
2885 if (rx_stats_active)
2886 rx_atomic_inc(&rx_stats.nPeerStructs);
2893 origPeer->refCount--;
2894 MUTEX_EXIT(&rx_peerHashTable_lock);
2899 /* Find the connection at (host, port) started at epoch, and with the
2900 * given connection id. Creates the server connection if necessary.
2901 * The type specifies whether a client connection or a server
2902 * connection is desired. In both cases, (host, port) specify the
2903 * peer's (host, pair) pair. Client connections are not made
2904 * automatically by this routine. The parameter socket gives the
2905 * socket descriptor on which the packet was received. This is used,
2906 * in the case of server connections, to check that *new* connections
2907 * come via a valid (port, serviceId). Finally, the securityIndex
2908 * parameter must match the existing index for the connection. If a
2909 * server connection is created, it will be created using the supplied
2910 * index, if the index is valid for this service */
2911 struct rx_connection *
2912 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2913 u_short port, u_short serviceId, afs_uint32 cid,
2914 afs_uint32 epoch, int type, u_int securityIndex)
2916 int hashindex, flag, i;
2917 struct rx_connection *conn;
2918 hashindex = CONN_HASH(host, port, cid, epoch, type);
2919 MUTEX_ENTER(&rx_connHashTable_lock);
2920 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2921 rx_connHashTable[hashindex],
2924 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2925 && (epoch == conn->epoch)) {
2926 struct rx_peer *pp = conn->peer;
2927 if (securityIndex != conn->securityIndex) {
2928 /* this isn't supposed to happen, but someone could forge a packet
2929 * like this, and there seems to be some CM bug that makes this
2930 * happen from time to time -- in which case, the fileserver
2932 MUTEX_EXIT(&rx_connHashTable_lock);
2933 return (struct rx_connection *)0;
2935 if (pp->host == host && pp->port == port)
2937 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2939 /* So what happens when it's a callback connection? */
2940 if ( /*type == RX_CLIENT_CONNECTION && */
2941 (conn->epoch & 0x80000000))
2945 /* the connection rxLastConn that was used the last time is not the
2946 ** one we are looking for now. Hence, start searching in the hash */
2948 conn = rx_connHashTable[hashindex];
2953 struct rx_service *service;
2954 if (type == RX_CLIENT_CONNECTION) {
2955 MUTEX_EXIT(&rx_connHashTable_lock);
2956 return (struct rx_connection *)0;
2958 service = rxi_FindService(socket, serviceId);
2959 if (!service || (securityIndex >= service->nSecurityObjects)
2960 || (service->securityObjects[securityIndex] == 0)) {
2961 MUTEX_EXIT(&rx_connHashTable_lock);
2962 return (struct rx_connection *)0;
2964 conn = rxi_AllocConnection(); /* This bzero's the connection */
2965 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2966 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2967 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2968 conn->next = rx_connHashTable[hashindex];
2969 rx_connHashTable[hashindex] = conn;
2970 conn->peer = rxi_FindPeer(host, port, 0, 1);
2971 conn->type = RX_SERVER_CONNECTION;
2972 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2973 conn->epoch = epoch;
2974 conn->cid = cid & RX_CIDMASK;
2975 /* conn->serial = conn->lastSerial = 0; */
2976 /* conn->timeout = 0; */
2977 conn->ackRate = RX_FAST_ACK_RATE;
2978 conn->service = service;
2979 conn->serviceId = serviceId;
2980 conn->securityIndex = securityIndex;
2981 conn->securityObject = service->securityObjects[securityIndex];
2982 conn->nSpecific = 0;
2983 conn->specific = NULL;
2984 rx_SetConnDeadTime(conn, service->connDeadTime);
2985 conn->idleDeadTime = service->idleDeadTime;
2986 conn->idleDeadDetection = service->idleDeadErr ? 1 : 0;
2987 for (i = 0; i < RX_MAXCALLS; i++) {
2988 conn->twind[i] = rx_initSendWindow;
2989 conn->rwind[i] = rx_initReceiveWindow;
2991 /* Notify security object of the new connection */
2992 RXS_NewConnection(conn->securityObject, conn);
2993 /* XXXX Connection timeout? */
2994 if (service->newConnProc)
2995 (*service->newConnProc) (conn);
2996 if (rx_stats_active)
2997 rx_atomic_inc(&rx_stats.nServerConns);
3000 MUTEX_ENTER(&rx_refcnt_mutex);
3002 MUTEX_EXIT(&rx_refcnt_mutex);
3004 rxLastConn = conn; /* store this connection as the last conn used */
3005 MUTEX_EXIT(&rx_connHashTable_lock);
3010 * Timeout a call on a busy call channel if appropriate.
3012 * @param[in] call The busy call.
3014 * @pre 'call' is marked as busy (namely,
3015 * call->conn->lastBusy[call->channel] != 0)
3017 * @pre call->lock is held
3018 * @pre rxi_busyChannelError is nonzero
3020 * @note call->lock is dropped and reacquired
3023 rxi_CheckBusy(struct rx_call *call)
3025 struct rx_connection *conn = call->conn;
3026 int channel = call->channel;
3027 int freechannel = 0;
3029 afs_uint32 callNumber = *call->callNumber;
3031 MUTEX_EXIT(&call->lock);
3033 MUTEX_ENTER(&conn->conn_call_lock);
3035 /* Are there any other call slots on this conn that we should try? Look for
3036 * slots that are empty and are either non-busy, or were marked as busy
3037 * longer than conn->secondsUntilDead seconds before this call started. */
3039 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3041 /* only look at channels that aren't us */
3045 if (conn->lastBusy[i]) {
3046 /* if this channel looked busy too recently, don't look at it */
3047 if (conn->lastBusy[i] >= call->startTime.sec) {
3050 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3055 if (conn->call[i]) {
3056 struct rx_call *tcall = conn->call[i];
3057 MUTEX_ENTER(&tcall->lock);
3058 if (tcall->state == RX_STATE_DALLY) {
3061 MUTEX_EXIT(&tcall->lock);
3067 MUTEX_EXIT(&conn->conn_call_lock);
3069 MUTEX_ENTER(&call->lock);
3071 /* Since the call->lock and conn->conn_call_lock have been released it is
3072 * possible that (1) the call may no longer be busy and/or (2) the call may
3073 * have been reused by another waiting thread. Therefore, we must confirm
3074 * that the call state has not changed when deciding whether or not to
3075 * force this application thread to retry by forcing a Timeout error. */
3077 if (freechannel && *call->callNumber == callNumber &&
3078 (call->flags & RX_CALL_PEER_BUSY)) {
3079 /* Since 'freechannel' is set, there exists another channel in this
3080 * rx_conn that the application thread might be able to use. We know
3081 * that we have the correct call since callNumber is unchanged, and we
3082 * know that the call is still busy. So, set the call error state to
3083 * rxi_busyChannelError so the application can retry the request,
3084 * presumably on a less-busy call channel. */
3086 rxi_CallError(call, RX_CALL_BUSY);
3090 /* There are two packet tracing routines available for testing and monitoring
3091 * Rx. One is called just after every packet is received and the other is
3092 * called just before every packet is sent. Received packets, have had their
3093 * headers decoded, and packets to be sent have not yet had their headers
3094 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3095 * containing the network address. Both can be modified. The return value, if
3096 * non-zero, indicates that the packet should be dropped. */
3098 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3099 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3101 /* A packet has been received off the interface. Np is the packet, socket is
3102 * the socket number it was received from (useful in determining which service
3103 * this packet corresponds to), and (host, port) reflect the host,port of the
3104 * sender. This call returns the packet to the caller if it is finished with
3105 * it, rather than de-allocating it, just as a small performance hack */
3108 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3109 afs_uint32 host, u_short port, int *tnop,
3110 struct rx_call **newcallp)
3112 struct rx_call *call;
3113 struct rx_connection *conn;
3115 afs_uint32 currentCallNumber;
3121 struct rx_packet *tnp;
3124 /* We don't print out the packet until now because (1) the time may not be
3125 * accurate enough until now in the lwp implementation (rx_Listener only gets
3126 * the time after the packet is read) and (2) from a protocol point of view,
3127 * this is the first time the packet has been seen */
3128 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3129 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3130 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3131 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3132 np->header.epoch, np->header.cid, np->header.callNumber,
3133 np->header.seq, np->header.flags, np));
3136 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3137 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3140 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3141 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3144 /* If an input tracer function is defined, call it with the packet and
3145 * network address. Note this function may modify its arguments. */
3146 if (rx_justReceived) {
3147 struct sockaddr_in addr;
3149 addr.sin_family = AF_INET;
3150 addr.sin_port = port;
3151 addr.sin_addr.s_addr = host;
3152 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3153 addr.sin_len = sizeof(addr);
3154 #endif /* AFS_OSF_ENV */
3155 drop = (*rx_justReceived) (np, &addr);
3156 /* drop packet if return value is non-zero */
3159 port = addr.sin_port; /* in case fcn changed addr */
3160 host = addr.sin_addr.s_addr;
3164 /* If packet was not sent by the client, then *we* must be the client */
3165 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3166 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3168 /* Find the connection (or fabricate one, if we're the server & if
3169 * necessary) associated with this packet */
3171 rxi_FindConnection(socket, host, port, np->header.serviceId,
3172 np->header.cid, np->header.epoch, type,
3173 np->header.securityIndex);
3176 /* If no connection found or fabricated, just ignore the packet.
3177 * (An argument could be made for sending an abort packet for
3182 /* If the connection is in an error state, send an abort packet and ignore
3183 * the incoming packet */
3185 /* Don't respond to an abort packet--we don't want loops! */
3186 MUTEX_ENTER(&conn->conn_data_lock);
3187 if (np->header.type != RX_PACKET_TYPE_ABORT)
3188 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3189 putConnection(conn);
3190 MUTEX_EXIT(&conn->conn_data_lock);
3194 /* Check for connection-only requests (i.e. not call specific). */
3195 if (np->header.callNumber == 0) {
3196 switch (np->header.type) {
3197 case RX_PACKET_TYPE_ABORT: {
3198 /* What if the supplied error is zero? */
3199 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3200 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3201 rxi_ConnectionError(conn, errcode);
3202 putConnection(conn);
3205 case RX_PACKET_TYPE_CHALLENGE:
3206 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3207 putConnection(conn);
3209 case RX_PACKET_TYPE_RESPONSE:
3210 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3211 putConnection(conn);
3213 case RX_PACKET_TYPE_PARAMS:
3214 case RX_PACKET_TYPE_PARAMS + 1:
3215 case RX_PACKET_TYPE_PARAMS + 2:
3216 /* ignore these packet types for now */
3217 putConnection(conn);
3221 /* Should not reach here, unless the peer is broken: send an
3223 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3224 MUTEX_ENTER(&conn->conn_data_lock);
3225 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3226 putConnection(conn);
3227 MUTEX_EXIT(&conn->conn_data_lock);
3232 channel = np->header.cid & RX_CHANNELMASK;
3233 call = conn->call[channel];
3236 MUTEX_ENTER(&call->lock);
3237 currentCallNumber = conn->callNumber[channel];
3238 } else if (type == RX_SERVER_CONNECTION) { /* No call allocated */
3239 MUTEX_ENTER(&conn->conn_call_lock);
3240 call = conn->call[channel];
3242 MUTEX_ENTER(&call->lock);
3243 MUTEX_EXIT(&conn->conn_call_lock);
3244 currentCallNumber = conn->callNumber[channel];
3246 call = rxi_NewCall(conn, channel); /* returns locked call */
3247 MUTEX_EXIT(&conn->conn_call_lock);
3248 *call->callNumber = currentCallNumber = np->header.callNumber;
3250 if (np->header.callNumber == 0)
3251 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3252 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3253 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3254 np->header.flags, np, np->length));
3256 call->state = RX_STATE_PRECALL;
3257 clock_GetTime(&call->queueTime);
3258 hzero(call->bytesSent);
3259 hzero(call->bytesRcvd);
3261 * If the number of queued calls exceeds the overload
3262 * threshold then abort this call.
3264 if ((rx_BusyThreshold > 0) &&
3265 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3266 struct rx_packet *tp;
3268 rxi_CallError(call, rx_BusyError);
3269 tp = rxi_SendCallAbort(call, np, 1, 0);
3270 MUTEX_EXIT(&call->lock);
3271 putConnection(conn);
3272 if (rx_stats_active)
3273 rx_atomic_inc(&rx_stats.nBusies);
3276 rxi_KeepAliveOn(call);
3278 } else { /* RX_CLIENT_CONNECTION and No call allocated */
3279 /* This packet can't be for this call. If the new call address is
3280 * 0 then no call is running on this channel. If there is a call
3281 * then, since this is a client connection we're getting data for
3282 * it must be for the previous call.
3284 if (rx_stats_active)
3285 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3286 putConnection(conn);
3290 /* There is a non-NULL locked call at this point */
3291 if (type == RX_SERVER_CONNECTION) { /* We're the server */
3292 if (np->header.callNumber < currentCallNumber) {
3293 MUTEX_EXIT(&call->lock);
3294 if (rx_stats_active)
3295 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3296 putConnection(conn);
3298 } else if (np->header.callNumber != currentCallNumber) {
3299 /* Wait until the transmit queue is idle before deciding
3300 * whether to reset the current call. Chances are that the
3301 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3304 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3305 if (call->state == RX_STATE_ACTIVE) {
3306 rxi_WaitforTQBusy(call);
3308 * If we entered error state while waiting,
3309 * must call rxi_CallError to permit rxi_ResetCall
3310 * to processed when the tqWaiter count hits zero.
3313 rxi_CallError(call, call->error);
3314 MUTEX_EXIT(&call->lock);
3315 putConnection(conn);
3319 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3320 /* If the new call cannot be taken right now send a busy and set
3321 * the error condition in this call, so that it terminates as
3322 * quickly as possible */
3323 if (call->state == RX_STATE_ACTIVE) {
3324 struct rx_packet *tp;
3326 rxi_CallError(call, RX_CALL_DEAD);
3327 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3329 MUTEX_EXIT(&call->lock);
3330 putConnection(conn);
3333 rxi_ResetCall(call, 0);
3334 *call->callNumber = np->header.callNumber;
3336 if (np->header.callNumber == 0)
3337 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3338 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3339 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3340 np->header.flags, np, np->length));
3342 call->state = RX_STATE_PRECALL;
3343 clock_GetTime(&call->queueTime);
3344 hzero(call->bytesSent);
3345 hzero(call->bytesRcvd);
3347 * If the number of queued calls exceeds the overload
3348 * threshold then abort this call.
3350 if ((rx_BusyThreshold > 0) &&
3351 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3352 struct rx_packet *tp;
3354 rxi_CallError(call, rx_BusyError);
3355 tp = rxi_SendCallAbort(call, np, 1, 0);
3356 MUTEX_EXIT(&call->lock);
3357 putConnection(conn);
3358 if (rx_stats_active)
3359 rx_atomic_inc(&rx_stats.nBusies);
3362 rxi_KeepAliveOn(call);
3364 /* Continuing call; do nothing here. */
3366 } else { /* we're the client */
3367 /* Ignore all incoming acknowledgements for calls in DALLY state */
3368 if ((call->state == RX_STATE_DALLY)
3369 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3370 if (rx_stats_active)
3371 rx_atomic_inc(&rx_stats.ignorePacketDally);
3372 MUTEX_EXIT(&call->lock);
3373 putConnection(conn);
3377 /* Ignore anything that's not relevant to the current call. If there
3378 * isn't a current call, then no packet is relevant. */
3379 if (np->header.callNumber != currentCallNumber) {
3380 if (rx_stats_active)
3381 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3382 MUTEX_EXIT(&call->lock);
3383 putConnection(conn);
3386 /* If the service security object index stamped in the packet does not
3387 * match the connection's security index, ignore the packet */
3388 if (np->header.securityIndex != conn->securityIndex) {
3389 MUTEX_EXIT(&call->lock);
3390 putConnection(conn);
3394 /* If we're receiving the response, then all transmit packets are
3395 * implicitly acknowledged. Get rid of them. */
3396 if (np->header.type == RX_PACKET_TYPE_DATA) {
3397 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3398 /* XXX Hack. Because we must release the global rx lock when
3399 * sending packets (osi_NetSend) we drop all acks while we're
3400 * traversing the tq in rxi_Start sending packets out because
3401 * packets may move to the freePacketQueue as result of being here!
3402 * So we drop these packets until we're safely out of the
3403 * traversing. Really ugly!
3404 * For fine grain RX locking, we set the acked field in the
3405 * packets and let rxi_Start remove them from the transmit queue.
3407 if (call->flags & RX_CALL_TQ_BUSY) {
3408 #ifdef RX_ENABLE_LOCKS
3409 rxi_SetAcksInTransmitQueue(call);
3411 putConnection(conn);
3412 return np; /* xmitting; drop packet */
3415 rxi_ClearTransmitQueue(call, 0);
3417 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3418 rxi_ClearTransmitQueue(call, 0);
3419 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3421 if (np->header.type == RX_PACKET_TYPE_ACK) {
3422 /* now check to see if this is an ack packet acknowledging that the
3423 * server actually *lost* some hard-acked data. If this happens we
3424 * ignore this packet, as it may indicate that the server restarted in
3425 * the middle of a call. It is also possible that this is an old ack
3426 * packet. We don't abort the connection in this case, because this
3427 * *might* just be an old ack packet. The right way to detect a server
3428 * restart in the midst of a call is to notice that the server epoch
3430 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3431 * XXX unacknowledged. I think that this is off-by-one, but
3432 * XXX I don't dare change it just yet, since it will
3433 * XXX interact badly with the server-restart detection
3434 * XXX code in receiveackpacket. */
3435 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3436 if (rx_stats_active)
3437 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3438 MUTEX_EXIT(&call->lock);
3439 putConnection(conn);
3443 } /* else not a data packet */
3446 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3447 /* Set remote user defined status from packet */
3448 call->remoteStatus = np->header.userStatus;
3450 /* Note the gap between the expected next packet and the actual
3451 * packet that arrived, when the new packet has a smaller serial number
3452 * than expected. Rioses frequently reorder packets all by themselves,
3453 * so this will be quite important with very large window sizes.
3454 * Skew is checked against 0 here to avoid any dependence on the type of
3455 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3457 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3458 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3459 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3461 MUTEX_ENTER(&conn->conn_data_lock);
3462 skew = conn->lastSerial - np->header.serial;
3463 conn->lastSerial = np->header.serial;
3464 MUTEX_EXIT(&conn->conn_data_lock);
3466 struct rx_peer *peer;
3468 if (skew > peer->inPacketSkew) {
3469 dpf(("*** In skew changed from %d to %d\n",
3470 peer->inPacketSkew, skew));
3471 peer->inPacketSkew = skew;
3475 /* Now do packet type-specific processing */
3476 switch (np->header.type) {
3477 case RX_PACKET_TYPE_DATA:
3478 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3481 case RX_PACKET_TYPE_ACK:
3482 /* Respond immediately to ack packets requesting acknowledgement
3484 if (np->header.flags & RX_REQUEST_ACK) {
3486 (void)rxi_SendCallAbort(call, 0, 1, 0);
3488 (void)rxi_SendAck(call, 0, np->header.serial,
3489 RX_ACK_PING_RESPONSE, 1);
3491 np = rxi_ReceiveAckPacket(call, np, 1);
3493 case RX_PACKET_TYPE_ABORT: {
3494 /* An abort packet: reset the call, passing the error up to the user. */
3495 /* What if error is zero? */
3496 /* What if the error is -1? the application will treat it as a timeout. */
3497 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3498 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3499 rxi_CallError(call, errdata);
3500 MUTEX_EXIT(&call->lock);
3501 putConnection(conn);
3502 return np; /* xmitting; drop packet */
3504 case RX_PACKET_TYPE_BUSY: {
3505 struct clock busyTime;
3507 clock_GetTime(&busyTime);
3509 MUTEX_EXIT(&call->lock);
3511 MUTEX_ENTER(&conn->conn_call_lock);
3512 MUTEX_ENTER(&call->lock);
3513 conn->lastBusy[call->channel] = busyTime.sec;
3514 call->flags |= RX_CALL_PEER_BUSY;
3515 MUTEX_EXIT(&call->lock);
3516 MUTEX_EXIT(&conn->conn_call_lock);
3518 putConnection(conn);
3522 case RX_PACKET_TYPE_ACKALL:
3523 /* All packets acknowledged, so we can drop all packets previously
3524 * readied for sending */
3525 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3526 /* XXX Hack. We because we can't release the global rx lock when
3527 * sending packets (osi_NetSend) we drop all ack pkts while we're
3528 * traversing the tq in rxi_Start sending packets out because
3529 * packets may move to the freePacketQueue as result of being
3530 * here! So we drop these packets until we're safely out of the
3531 * traversing. Really ugly!
3532 * For fine grain RX locking, we set the acked field in the packets
3533 * and let rxi_Start remove the packets from the transmit queue.
3535 if (call->flags & RX_CALL_TQ_BUSY) {
3536 #ifdef RX_ENABLE_LOCKS
3537 rxi_SetAcksInTransmitQueue(call);
3539 #else /* RX_ENABLE_LOCKS */
3540 MUTEX_EXIT(&call->lock);
3541 putConnection(conn);
3542 return np; /* xmitting; drop packet */
3543 #endif /* RX_ENABLE_LOCKS */
3545 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3546 rxi_ClearTransmitQueue(call, 0);
3549 /* Should not reach here, unless the peer is broken: send an abort
3551 rxi_CallError(call, RX_PROTOCOL_ERROR);
3552 np = rxi_SendCallAbort(call, np, 1, 0);
3555 /* Note when this last legitimate packet was received, for keep-alive
3556 * processing. Note, we delay getting the time until now in the hope that
3557 * the packet will be delivered to the user before any get time is required
3558 * (if not, then the time won't actually be re-evaluated here). */
3559 call->lastReceiveTime = clock_Sec();
3560 /* we've received a legit packet, so the channel is not busy */
3561 call->flags &= ~RX_CALL_PEER_BUSY;
3562 MUTEX_EXIT(&call->lock);
3563 putConnection(conn);
3567 /* return true if this is an "interesting" connection from the point of view
3568 of someone trying to debug the system */
3570 rxi_IsConnInteresting(struct rx_connection *aconn)
3573 struct rx_call *tcall;
3575 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3578 for (i = 0; i < RX_MAXCALLS; i++) {
3579 tcall = aconn->call[i];
3581 if ((tcall->state == RX_STATE_PRECALL)
3582 || (tcall->state == RX_STATE_ACTIVE))
3584 if ((tcall->mode == RX_MODE_SENDING)
3585 || (tcall->mode == RX_MODE_RECEIVING))
3593 /* if this is one of the last few packets AND it wouldn't be used by the
3594 receiving call to immediately satisfy a read request, then drop it on
3595 the floor, since accepting it might prevent a lock-holding thread from
3596 making progress in its reading. If a call has been cleared while in
3597 the precall state then ignore all subsequent packets until the call
3598 is assigned to a thread. */
3601 TooLow(struct rx_packet *ap, struct rx_call *acall)
3605 MUTEX_ENTER(&rx_quota_mutex);
3606 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3607 && (acall->state == RX_STATE_PRECALL))
3608 || ((rx_nFreePackets < rxi_dataQuota + 2)
3609 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3610 && (acall->flags & RX_CALL_READER_WAIT)))) {
3613 MUTEX_EXIT(&rx_quota_mutex);
3619 * Clear the attach wait flag on a connection and proceed.
3621 * Any processing waiting for a connection to be attached should be
3622 * unblocked. We clear the flag and do any other needed tasks.
3625 * the conn to unmark waiting for attach
3627 * @pre conn's conn_data_lock must be locked before calling this function
3631 rxi_ConnClearAttachWait(struct rx_connection *conn)
3633 /* Indicate that rxi_CheckReachEvent is no longer running by
3634 * clearing the flag. Must be atomic under conn_data_lock to
3635 * avoid a new call slipping by: rxi_CheckConnReach holds
3636 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3638 conn->flags &= ~RX_CONN_ATTACHWAIT;
3639 if (conn->flags & RX_CONN_NAT_PING) {
3640 conn->flags &= ~RX_CONN_NAT_PING;
3641 rxi_ScheduleNatKeepAliveEvent(conn);
3646 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3648 struct rx_connection *conn = arg1;
3649 struct rx_call *acall = arg2;
3650 struct rx_call *call = acall;
3651 struct clock when, now;
3654 MUTEX_ENTER(&conn->conn_data_lock);
3657 rxevent_Put(conn->checkReachEvent);
3658 conn->checkReachEvent = NULL;
3661 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3663 putConnection(conn);
3665 MUTEX_EXIT(&conn->conn_data_lock);
3669 MUTEX_ENTER(&conn->conn_call_lock);
3670 MUTEX_ENTER(&conn->conn_data_lock);
3671 for (i = 0; i < RX_MAXCALLS; i++) {
3672 struct rx_call *tc = conn->call[i];
3673 if (tc && tc->state == RX_STATE_PRECALL) {
3679 rxi_ConnClearAttachWait(conn);
3680 MUTEX_EXIT(&conn->conn_data_lock);
3681 MUTEX_EXIT(&conn->conn_call_lock);
3686 MUTEX_ENTER(&call->lock);
3687 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3689 MUTEX_EXIT(&call->lock);
3691 clock_GetTime(&now);
3693 when.sec += RX_CHECKREACH_TIMEOUT;
3694 MUTEX_ENTER(&conn->conn_data_lock);
3695 if (!conn->checkReachEvent) {
3696 MUTEX_ENTER(&rx_refcnt_mutex);
3698 MUTEX_EXIT(&rx_refcnt_mutex);
3699 conn->checkReachEvent = rxevent_Post(&when, &now,
3700 rxi_CheckReachEvent, conn,
3703 MUTEX_EXIT(&conn->conn_data_lock);
3709 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3711 struct rx_service *service = conn->service;
3712 struct rx_peer *peer = conn->peer;
3713 afs_uint32 now, lastReach;
3715 if (service->checkReach == 0)
3719 MUTEX_ENTER(&peer->peer_lock);
3720 lastReach = peer->lastReachTime;
3721 MUTEX_EXIT(&peer->peer_lock);
3722 if (now - lastReach < RX_CHECKREACH_TTL)
3725 MUTEX_ENTER(&conn->conn_data_lock);
3726 if (conn->flags & RX_CONN_ATTACHWAIT) {
3727 MUTEX_EXIT(&conn->conn_data_lock);
3730 conn->flags |= RX_CONN_ATTACHWAIT;
3731 MUTEX_EXIT(&conn->conn_data_lock);
3732 if (!conn->checkReachEvent)
3733 rxi_CheckReachEvent(NULL, conn, call, 0);
3738 /* try to attach call, if authentication is complete */
3740 TryAttach(struct rx_call *acall, osi_socket socket,
3741 int *tnop, struct rx_call **newcallp,
3744 struct rx_connection *conn = acall->conn;
3746 if (conn->type == RX_SERVER_CONNECTION
3747 && acall->state == RX_STATE_PRECALL) {
3748 /* Don't attach until we have any req'd. authentication. */
3749 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3750 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3751 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3752 /* Note: this does not necessarily succeed; there
3753 * may not any proc available
3756 rxi_ChallengeOn(acall->conn);
3761 /* A data packet has been received off the interface. This packet is
3762 * appropriate to the call (the call is in the right state, etc.). This
3763 * routine can return a packet to the caller, for re-use */
3766 rxi_ReceiveDataPacket(struct rx_call *call,
3767 struct rx_packet *np, int istack,
3768 osi_socket socket, afs_uint32 host, u_short port,
3769 int *tnop, struct rx_call **newcallp)
3771 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3776 afs_uint32 serial=0, flags=0;
3778 struct rx_packet *tnp;
3779 if (rx_stats_active)
3780 rx_atomic_inc(&rx_stats.dataPacketsRead);
3783 /* If there are no packet buffers, drop this new packet, unless we can find
3784 * packet buffers from inactive calls */
3786 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3787 MUTEX_ENTER(&rx_freePktQ_lock);
3788 rxi_NeedMorePackets = TRUE;
3789 MUTEX_EXIT(&rx_freePktQ_lock);
3790 if (rx_stats_active)
3791 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3792 call->rprev = np->header.serial;
3793 rxi_calltrace(RX_TRACE_DROP, call);
3794 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3795 /* We used to clear the receive queue here, in an attempt to free
3796 * packets. However this is unsafe if the queue has received a
3797 * soft ACK for the final packet */
3798 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3800 /* we've damaged this call already, might as well do it in. */
3806 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3807 * packet is one of several packets transmitted as a single
3808 * datagram. Do not send any soft or hard acks until all packets
3809 * in a jumbogram have been processed. Send negative acks right away.
3811 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3812 /* tnp is non-null when there are more packets in the
3813 * current jumbo gram */
3820 seq = np->header.seq;
3821 serial = np->header.serial;
3822 flags = np->header.flags;
3824 /* If the call is in an error state, send an abort message */
3826 return rxi_SendCallAbort(call, np, istack, 0);
3828 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3829 * AFS 3.5 jumbogram. */
3830 if (flags & RX_JUMBO_PACKET) {
3831 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3836 if (np->header.spare != 0) {
3837 MUTEX_ENTER(&call->conn->conn_data_lock);
3838 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3839 MUTEX_EXIT(&call->conn->conn_data_lock);
3842 /* The usual case is that this is the expected next packet */
3843 if (seq == call->rnext) {
3845 /* Check to make sure it is not a duplicate of one already queued */
3846 if (queue_IsNotEmpty(&call->rq)
3847 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3848 if (rx_stats_active)
3849 rx_atomic_inc(&rx_stats.dupPacketsRead);
3850 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3851 rxevent_Cancel(&call->delayedAckEvent, call,
3852 RX_CALL_REFCOUNT_DELAY);
3853 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3859 /* It's the next packet. Stick it on the receive queue
3860 * for this call. Set newPackets to make sure we wake
3861 * the reader once all packets have been processed */
3862 #ifdef RX_TRACK_PACKETS
3863 np->flags |= RX_PKTFLAG_RQ;
3865 queue_Prepend(&call->rq, np);
3866 #ifdef RXDEBUG_PACKET
3868 #endif /* RXDEBUG_PACKET */
3870 np = NULL; /* We can't use this anymore */
3873 /* If an ack is requested then set a flag to make sure we
3874 * send an acknowledgement for this packet */
3875 if (flags & RX_REQUEST_ACK) {
3876 ackNeeded = RX_ACK_REQUESTED;
3879 /* Keep track of whether we have received the last packet */
3880 if (flags & RX_LAST_PACKET) {
3881 call->flags |= RX_CALL_HAVE_LAST;
3885 /* Check whether we have all of the packets for this call */
3886 if (call->flags & RX_CALL_HAVE_LAST) {
3887 afs_uint32 tseq; /* temporary sequence number */
3888 struct rx_packet *tp; /* Temporary packet pointer */
3889 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3891 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3892 if (tseq != tp->header.seq)
3894 if (tp->header.flags & RX_LAST_PACKET) {
3895 call->flags |= RX_CALL_RECEIVE_DONE;
3902 /* Provide asynchronous notification for those who want it
3903 * (e.g. multi rx) */
3904 if (call->arrivalProc) {
3905 (*call->arrivalProc) (call, call->arrivalProcHandle,
3906 call->arrivalProcArg);
3907 call->arrivalProc = (void (*)())0;
3910 /* Update last packet received */
3913 /* If there is no server process serving this call, grab
3914 * one, if available. We only need to do this once. If a
3915 * server thread is available, this thread becomes a server
3916 * thread and the server thread becomes a listener thread. */
3918 TryAttach(call, socket, tnop, newcallp, 0);
3921 /* This is not the expected next packet. */
3923 /* Determine whether this is a new or old packet, and if it's
3924 * a new one, whether it fits into the current receive window.
3925 * Also figure out whether the packet was delivered in sequence.
3926 * We use the prev variable to determine whether the new packet
3927 * is the successor of its immediate predecessor in the
3928 * receive queue, and the missing flag to determine whether
3929 * any of this packets predecessors are missing. */
3931 afs_uint32 prev; /* "Previous packet" sequence number */
3932 struct rx_packet *tp; /* Temporary packet pointer */
3933 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3934 int missing; /* Are any predecessors missing? */
3936 /* If the new packet's sequence number has been sent to the
3937 * application already, then this is a duplicate */
3938 if (seq < call->rnext) {
3939 if (rx_stats_active)
3940 rx_atomic_inc(&rx_stats.dupPacketsRead);
3941 rxevent_Cancel(&call->delayedAckEvent, call,
3942 RX_CALL_REFCOUNT_DELAY);
3943 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3949 /* If the sequence number is greater than what can be
3950 * accomodated by the current window, then send a negative
3951 * acknowledge and drop the packet */
3952 if ((call->rnext + call->rwind) <= seq) {
3953 rxevent_Cancel(&call->delayedAckEvent, call,
3954 RX_CALL_REFCOUNT_DELAY);
3955 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3962 /* Look for the packet in the queue of old received packets */
3963 for (prev = call->rnext - 1, missing =
3964 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3965 /*Check for duplicate packet */
3966 if (seq == tp->header.seq) {
3967 if (rx_stats_active)
3968 rx_atomic_inc(&rx_stats.dupPacketsRead);
3969 rxevent_Cancel(&call->delayedAckEvent, call,
3970 RX_CALL_REFCOUNT_DELAY);
3971 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3977 /* If we find a higher sequence packet, break out and
3978 * insert the new packet here. */
3979 if (seq < tp->header.seq)
3981 /* Check for missing packet */
3982 if (tp->header.seq != prev + 1) {
3986 prev = tp->header.seq;
3989 /* Keep track of whether we have received the last packet. */
3990 if (flags & RX_LAST_PACKET) {
3991 call->flags |= RX_CALL_HAVE_LAST;
3994 /* It's within the window: add it to the the receive queue.
3995 * tp is left by the previous loop either pointing at the
3996 * packet before which to insert the new packet, or at the
3997 * queue head if the queue is empty or the packet should be
3999 #ifdef RX_TRACK_PACKETS
4000 np->flags |= RX_PKTFLAG_RQ;
4002 #ifdef RXDEBUG_PACKET
4004 #endif /* RXDEBUG_PACKET */
4005 queue_InsertBefore(tp, np);
4009 /* Check whether we have all of the packets for this call */
4010 if ((call->flags & RX_CALL_HAVE_LAST)
4011 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4012 afs_uint32 tseq; /* temporary sequence number */
4015 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
4016 if (tseq != tp->header.seq)
4018 if (tp->header.flags & RX_LAST_PACKET) {
4019 call->flags |= RX_CALL_RECEIVE_DONE;
4026 /* We need to send an ack of the packet is out of sequence,
4027 * or if an ack was requested by the peer. */
4028 if (seq != prev + 1 || missing) {
4029 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4030 } else if (flags & RX_REQUEST_ACK) {
4031 ackNeeded = RX_ACK_REQUESTED;
4034 /* Acknowledge the last packet for each call */
4035 if (flags & RX_LAST_PACKET) {
4046 * If the receiver is waiting for an iovec, fill the iovec
4047 * using the data from the receive queue */
4048 if (call->flags & RX_CALL_IOVEC_WAIT) {
4049 didHardAck = rxi_FillReadVec(call, serial);
4050 /* the call may have been aborted */
4059 /* Wakeup the reader if any */
4060 if ((call->flags & RX_CALL_READER_WAIT)
4061 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4062 || (call->iovNext >= call->iovMax)
4063 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4064 call->flags &= ~RX_CALL_READER_WAIT;
4065 #ifdef RX_ENABLE_LOCKS
4066 CV_BROADCAST(&call->cv_rq);
4068 osi_rxWakeup(&call->rq);
4074 * Send an ack when requested by the peer, or once every
4075 * rxi_SoftAckRate packets until the last packet has been
4076 * received. Always send a soft ack for the last packet in
4077 * the server's reply. */
4079 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4080 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4081 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4082 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4083 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4084 } else if (call->nSoftAcks) {
4085 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4086 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4088 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4089 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4090 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4097 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4099 struct rx_peer *peer = conn->peer;
4101 MUTEX_ENTER(&peer->peer_lock);
4102 peer->lastReachTime = clock_Sec();
4103 MUTEX_EXIT(&peer->peer_lock);
4105 MUTEX_ENTER(&conn->conn_data_lock);
4106 if (conn->flags & RX_CONN_ATTACHWAIT) {
4109 rxi_ConnClearAttachWait(conn);
4110 MUTEX_EXIT(&conn->conn_data_lock);
4112 for (i = 0; i < RX_MAXCALLS; i++) {
4113 struct rx_call *call = conn->call[i];
4116 MUTEX_ENTER(&call->lock);
4117 /* tnop can be null if newcallp is null */
4118 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4120 MUTEX_EXIT(&call->lock);
4124 MUTEX_EXIT(&conn->conn_data_lock);
4127 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4129 rx_ack_reason(int reason)
4132 case RX_ACK_REQUESTED:
4134 case RX_ACK_DUPLICATE:
4136 case RX_ACK_OUT_OF_SEQUENCE:
4138 case RX_ACK_EXCEEDS_WINDOW:
4140 case RX_ACK_NOSPACE:
4144 case RX_ACK_PING_RESPONSE:
4157 /* The real smarts of the whole thing. */
4159 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4162 struct rx_ackPacket *ap;
4164 struct rx_packet *tp;
4165 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
4166 struct rx_connection *conn = call->conn;
4167 struct rx_peer *peer = conn->peer;
4168 struct clock now; /* Current time, for RTT calculations */
4172 /* because there are CM's that are bogus, sending weird values for this. */
4173 afs_uint32 skew = 0;
4178 int newAckCount = 0;
4179 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4180 int pktsize = 0; /* Set if we need to update the peer mtu */
4181 int conn_data_locked = 0;
4183 if (rx_stats_active)
4184 rx_atomic_inc(&rx_stats.ackPacketsRead);
4185 ap = (struct rx_ackPacket *)rx_DataOf(np);
4186 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4188 return np; /* truncated ack packet */
4190 /* depends on ack packet struct */
4191 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4192 first = ntohl(ap->firstPacket);
4193 prev = ntohl(ap->previousPacket);
4194 serial = ntohl(ap->serial);
4195 /* temporarily disabled -- needs to degrade over time
4196 * skew = ntohs(ap->maxSkew); */
4198 /* Ignore ack packets received out of order */
4199 if (first < call->tfirst ||
4200 (first == call->tfirst && prev < call->tprev)) {
4206 if (np->header.flags & RX_SLOW_START_OK) {
4207 call->flags |= RX_CALL_SLOW_START_OK;
4210 if (ap->reason == RX_ACK_PING_RESPONSE)
4211 rxi_UpdatePeerReach(conn, call);
4213 if (conn->lastPacketSizeSeq) {
4214 MUTEX_ENTER(&conn->conn_data_lock);
4215 conn_data_locked = 1;
4216 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4217 pktsize = conn->lastPacketSize;
4218 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4221 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4222 if (!conn_data_locked) {
4223 MUTEX_ENTER(&conn->conn_data_lock);
4224 conn_data_locked = 1;
4226 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4227 /* process mtu ping ack */
4228 pktsize = conn->lastPingSize;
4229 conn->lastPingSizeSer = conn->lastPingSize = 0;
4233 if (conn_data_locked) {
4234 MUTEX_EXIT(&conn->conn_data_lock);
4235 conn_data_locked = 0;
4239 if (rxdebug_active) {
4243 len = _snprintf(msg, sizeof(msg),
4244 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
4245 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4246 ntohl(ap->serial), ntohl(ap->previousPacket),
4247 (unsigned int)np->header.seq, (unsigned int)skew,
4248 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
4252 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4253 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4257 OutputDebugString(msg);
4259 #else /* AFS_NT40_ENV */
4262 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
4263 ap->reason, ntohl(ap->previousPacket),
4264 (unsigned int)np->header.seq, (unsigned int)serial,
4265 (unsigned int)skew, ntohl(ap->firstPacket));
4268 for (offset = 0; offset < nAcks; offset++)
4269 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4274 #endif /* AFS_NT40_ENV */
4277 MUTEX_ENTER(&peer->peer_lock);
4280 * Start somewhere. Can't assume we can send what we can receive,
4281 * but we are clearly receiving.
4283 if (!peer->maxPacketSize)
4284 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4286 if (pktsize > peer->maxPacketSize) {
4287 peer->maxPacketSize = pktsize;
4288 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4289 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4290 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4291 rxi_ScheduleGrowMTUEvent(call, 1);
4296 /* Update the outgoing packet skew value to the latest value of
4297 * the peer's incoming packet skew value. The ack packet, of
4298 * course, could arrive out of order, but that won't affect things
4300 peer->outPacketSkew = skew;
4303 clock_GetTime(&now);
4305 /* The transmit queue splits into 4 sections.
4307 * The first section is packets which have now been acknowledged
4308 * by a window size change in the ack. These have reached the
4309 * application layer, and may be discarded. These are packets
4310 * with sequence numbers < ap->firstPacket.
4312 * The second section is packets which have sequence numbers in
4313 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4314 * contents of the packet's ack array determines whether these
4315 * packets are acknowledged or not.
4317 * The third section is packets which fall above the range
4318 * addressed in the ack packet. These have not yet been received
4321 * The four section is packets which have not yet been transmitted.
4322 * These packets will have a header.serial of 0.
4325 /* First section - implicitly acknowledged packets that can be
4329 tp = queue_First(&call->tq, rx_packet);
4330 while(!queue_IsEnd(&call->tq, tp) && tp->header.seq < first) {
4331 struct rx_packet *next;
4333 next = queue_Next(tp, rx_packet);
4334 call->tfirst = tp->header.seq + 1;
4336 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4338 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4341 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4342 /* XXX Hack. Because we have to release the global rx lock when sending
4343 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4344 * in rxi_Start sending packets out because packets may move to the
4345 * freePacketQueue as result of being here! So we drop these packets until
4346 * we're safely out of the traversing. Really ugly!
4347 * To make it even uglier, if we're using fine grain locking, we can
4348 * set the ack bits in the packets and have rxi_Start remove the packets
4349 * when it's done transmitting.
4351 if (call->flags & RX_CALL_TQ_BUSY) {
4352 #ifdef RX_ENABLE_LOCKS
4353 tp->flags |= RX_PKTFLAG_ACKED;
4354 call->flags |= RX_CALL_TQ_SOME_ACKED;
4355 #else /* RX_ENABLE_LOCKS */
4357 #endif /* RX_ENABLE_LOCKS */
4359 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4362 #ifdef RX_TRACK_PACKETS
4363 tp->flags &= ~RX_PKTFLAG_TQ;
4365 #ifdef RXDEBUG_PACKET
4367 #endif /* RXDEBUG_PACKET */
4368 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4373 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4375 /* Second section of the queue - packets for which we are receiving
4378 * Go through the explicit acks/nacks and record the results in
4379 * the waiting packets. These are packets that can't be released
4380 * yet, even with a positive acknowledge. This positive
4381 * acknowledge only means the packet has been received by the
4382 * peer, not that it will be retained long enough to be sent to
4383 * the peer's upper level. In addition, reset the transmit timers
4384 * of any missing packets (those packets that must be missing
4385 * because this packet was out of sequence) */
4387 call->nSoftAcked = 0;
4389 while (!queue_IsEnd(&call->tq, tp) && tp->header.seq < first + nAcks) {
4390 /* Set the acknowledge flag per packet based on the
4391 * information in the ack packet. An acknowlegded packet can
4392 * be downgraded when the server has discarded a packet it
4393 * soacked previously, or when an ack packet is received
4394 * out of sequence. */
4395 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4396 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4398 tp->flags |= RX_PKTFLAG_ACKED;
4399 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4406 } else /* RX_ACK_TYPE_NACK */ {
4407 tp->flags &= ~RX_PKTFLAG_ACKED;
4411 tp = queue_Next(tp, rx_packet);
4414 /* We don't need to take any action with the 3rd or 4th section in the
4415 * queue - they're not addressed by the contents of this ACK packet.
4418 /* If the window has been extended by this acknowledge packet,
4419 * then wakeup a sender waiting in alloc for window space, or try
4420 * sending packets now, if he's been sitting on packets due to
4421 * lack of window space */
4422 if (call->tnext < (call->tfirst + call->twind)) {
4423 #ifdef RX_ENABLE_LOCKS
4424 CV_SIGNAL(&call->cv_twind);
4426 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4427 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4428 osi_rxWakeup(&call->twind);
4431 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4432 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4436 /* if the ack packet has a receivelen field hanging off it,
4437 * update our state */
4438 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4441 /* If the ack packet has a "recommended" size that is less than
4442 * what I am using now, reduce my size to match */
4443 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4444 (int)sizeof(afs_int32), &tSize);
4445 tSize = (afs_uint32) ntohl(tSize);
4446 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4448 /* Get the maximum packet size to send to this peer */
4449 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4451 tSize = (afs_uint32) ntohl(tSize);
4452 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4453 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4455 /* sanity check - peer might have restarted with different params.
4456 * If peer says "send less", dammit, send less... Peer should never
4457 * be unable to accept packets of the size that prior AFS versions would
4458 * send without asking. */
4459 if (peer->maxMTU != tSize) {
4460 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4462 peer->maxMTU = tSize;
4463 peer->MTU = MIN(tSize, peer->MTU);
4464 call->MTU = MIN(call->MTU, tSize);
4467 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4470 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4471 (int)sizeof(afs_int32), &tSize);
4472 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4473 if (tSize < call->twind) { /* smaller than our send */
4474 call->twind = tSize; /* window, we must send less... */
4475 call->ssthresh = MIN(call->twind, call->ssthresh);
4476 call->conn->twind[call->channel] = call->twind;
4479 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4480 * network MTU confused with the loopback MTU. Calculate the
4481 * maximum MTU here for use in the slow start code below.
4483 /* Did peer restart with older RX version? */
4484 if (peer->maxDgramPackets > 1) {
4485 peer->maxDgramPackets = 1;
4487 } else if (np->length >=
4488 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4491 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4492 sizeof(afs_int32), &tSize);
4493 tSize = (afs_uint32) ntohl(tSize);
4495 * As of AFS 3.5 we set the send window to match the receive window.
4497 if (tSize < call->twind) {
4498 call->twind = tSize;
4499 call->conn->twind[call->channel] = call->twind;
4500 call->ssthresh = MIN(call->twind, call->ssthresh);
4501 } else if (tSize > call->twind) {
4502 call->twind = tSize;
4503 call->conn->twind[call->channel] = call->twind;
4507 * As of AFS 3.5, a jumbogram is more than one fixed size
4508 * packet transmitted in a single UDP datagram. If the remote
4509 * MTU is smaller than our local MTU then never send a datagram
4510 * larger than the natural MTU.
4513 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4514 (int)sizeof(afs_int32), &tSize);
4515 maxDgramPackets = (afs_uint32) ntohl(tSize);
4516 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4518 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4519 if (maxDgramPackets > 1) {
4520 peer->maxDgramPackets = maxDgramPackets;
4521 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4523 peer->maxDgramPackets = 1;
4524 call->MTU = peer->natMTU;
4526 } else if (peer->maxDgramPackets > 1) {
4527 /* Restarted with lower version of RX */
4528 peer->maxDgramPackets = 1;
4530 } else if (peer->maxDgramPackets > 1
4531 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4532 /* Restarted with lower version of RX */
4533 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4534 peer->natMTU = OLD_MAX_PACKET_SIZE;
4535 peer->MTU = OLD_MAX_PACKET_SIZE;
4536 peer->maxDgramPackets = 1;
4537 peer->nDgramPackets = 1;
4539 call->MTU = OLD_MAX_PACKET_SIZE;
4544 * Calculate how many datagrams were successfully received after
4545 * the first missing packet and adjust the negative ack counter
4550 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4551 if (call->nNacks < nNacked) {
4552 call->nNacks = nNacked;
4555 call->nAcks += newAckCount;
4559 /* If the packet contained new acknowledgements, rather than just
4560 * being a duplicate of one we have previously seen, then we can restart
4563 if (newAckCount > 0)
4564 rxi_rto_packet_acked(call, istack);
4566 if (call->flags & RX_CALL_FAST_RECOVER) {
4567 if (newAckCount == 0) {
4568 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4570 call->flags &= ~RX_CALL_FAST_RECOVER;
4571 call->cwind = call->nextCwind;
4572 call->nextCwind = 0;
4575 call->nCwindAcks = 0;
4576 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4577 /* Three negative acks in a row trigger congestion recovery */
4578 call->flags |= RX_CALL_FAST_RECOVER;
4579 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4581 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4582 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4583 call->nextCwind = call->ssthresh;
4586 peer->MTU = call->MTU;
4587 peer->cwind = call->nextCwind;
4588 peer->nDgramPackets = call->nDgramPackets;
4590 call->congestSeq = peer->congestSeq;
4592 /* Reset the resend times on the packets that were nacked
4593 * so we will retransmit as soon as the window permits
4596 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4598 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4599 tp->flags &= ~RX_PKTFLAG_SENT;
4601 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4606 /* If cwind is smaller than ssthresh, then increase
4607 * the window one packet for each ack we receive (exponential
4609 * If cwind is greater than or equal to ssthresh then increase
4610 * the congestion window by one packet for each cwind acks we
4611 * receive (linear growth). */
4612 if (call->cwind < call->ssthresh) {
4614 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4615 call->nCwindAcks = 0;
4617 call->nCwindAcks += newAckCount;
4618 if (call->nCwindAcks >= call->cwind) {
4619 call->nCwindAcks = 0;
4620 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4624 * If we have received several acknowledgements in a row then
4625 * it is time to increase the size of our datagrams
4627 if ((int)call->nAcks > rx_nDgramThreshold) {
4628 if (peer->maxDgramPackets > 1) {
4629 if (call->nDgramPackets < peer->maxDgramPackets) {
4630 call->nDgramPackets++;
4632 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4633 } else if (call->MTU < peer->maxMTU) {
4634 /* don't upgrade if we can't handle it */
4635 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4636 call->MTU = peer->ifMTU;
4638 call->MTU += peer->natMTU;
4639 call->MTU = MIN(call->MTU, peer->maxMTU);
4646 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4648 /* Servers need to hold the call until all response packets have
4649 * been acknowledged. Soft acks are good enough since clients
4650 * are not allowed to clear their receive queues. */
4651 if (call->state == RX_STATE_HOLD
4652 && call->tfirst + call->nSoftAcked >= call->tnext) {
4653 call->state = RX_STATE_DALLY;
4654 rxi_ClearTransmitQueue(call, 0);
4655 rxevent_Cancel(&call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4656 } else if (!queue_IsEmpty(&call->tq)) {
4657 rxi_Start(call, istack);
4662 /* Received a response to a challenge packet */
4664 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4665 struct rx_packet *np, int istack)
4669 /* Ignore the packet if we're the client */
4670 if (conn->type == RX_CLIENT_CONNECTION)
4673 /* If already authenticated, ignore the packet (it's probably a retry) */
4674 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4677 /* Otherwise, have the security object evaluate the response packet */
4678 error = RXS_CheckResponse(conn->securityObject, conn, np);
4680 /* If the response is invalid, reset the connection, sending
4681 * an abort to the peer */
4685 rxi_ConnectionError(conn, error);
4686 MUTEX_ENTER(&conn->conn_data_lock);
4687 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4688 MUTEX_EXIT(&conn->conn_data_lock);
4691 /* If the response is valid, any calls waiting to attach
4692 * servers can now do so */
4695 for (i = 0; i < RX_MAXCALLS; i++) {
4696 struct rx_call *call = conn->call[i];
4698 MUTEX_ENTER(&call->lock);
4699 if (call->state == RX_STATE_PRECALL)
4700 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4701 /* tnop can be null if newcallp is null */
4702 MUTEX_EXIT(&call->lock);
4706 /* Update the peer reachability information, just in case
4707 * some calls went into attach-wait while we were waiting
4708 * for authentication..
4710 rxi_UpdatePeerReach(conn, NULL);
4715 /* A client has received an authentication challenge: the security
4716 * object is asked to cough up a respectable response packet to send
4717 * back to the server. The server is responsible for retrying the
4718 * challenge if it fails to get a response. */
4721 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4722 struct rx_packet *np, int istack)
4726 /* Ignore the challenge if we're the server */
4727 if (conn->type == RX_SERVER_CONNECTION)
4730 /* Ignore the challenge if the connection is otherwise idle; someone's
4731 * trying to use us as an oracle. */
4732 if (!rxi_HasActiveCalls(conn))
4735 /* Send the security object the challenge packet. It is expected to fill
4736 * in the response. */
4737 error = RXS_GetResponse(conn->securityObject, conn, np);
4739 /* If the security object is unable to return a valid response, reset the
4740 * connection and send an abort to the peer. Otherwise send the response
4741 * packet to the peer connection. */
4743 rxi_ConnectionError(conn, error);
4744 MUTEX_ENTER(&conn->conn_data_lock);
4745 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4746 MUTEX_EXIT(&conn->conn_data_lock);
4748 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4749 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4755 /* Find an available server process to service the current request in
4756 * the given call structure. If one isn't available, queue up this
4757 * call so it eventually gets one */
4759 rxi_AttachServerProc(struct rx_call *call,
4760 osi_socket socket, int *tnop,
4761 struct rx_call **newcallp)
4763 struct rx_serverQueueEntry *sq;
4764 struct rx_service *service = call->conn->service;
4767 /* May already be attached */
4768 if (call->state == RX_STATE_ACTIVE)
4771 MUTEX_ENTER(&rx_serverPool_lock);
4773 haveQuota = QuotaOK(service);
4774 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4775 /* If there are no processes available to service this call,
4776 * put the call on the incoming call queue (unless it's
4777 * already on the queue).
4779 #ifdef RX_ENABLE_LOCKS
4781 ReturnToServerPool(service);
4782 #endif /* RX_ENABLE_LOCKS */
4784 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4785 call->flags |= RX_CALL_WAIT_PROC;
4786 rx_atomic_inc(&rx_nWaiting);
4787 rx_atomic_inc(&rx_nWaited);
4788 rxi_calltrace(RX_CALL_ARRIVAL, call);
4789 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4790 queue_Append(&rx_incomingCallQueue, call);
4793 sq = queue_Last(&rx_idleServerQueue, rx_serverQueueEntry);
4795 /* If hot threads are enabled, and both newcallp and sq->socketp
4796 * are non-null, then this thread will process the call, and the
4797 * idle server thread will start listening on this threads socket.
4800 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4803 *sq->socketp = socket;
4804 clock_GetTime(&call->startTime);
4805 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4809 if (call->flags & RX_CALL_WAIT_PROC) {
4810 /* Conservative: I don't think this should happen */
4811 call->flags &= ~RX_CALL_WAIT_PROC;
4812 if (queue_IsOnQueue(call)) {
4815 rx_atomic_dec(&rx_nWaiting);
4818 call->state = RX_STATE_ACTIVE;
4819 call->mode = RX_MODE_RECEIVING;
4820 #ifdef RX_KERNEL_TRACE
4822 int glockOwner = ISAFS_GLOCK();
4825 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4826 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4832 if (call->flags & RX_CALL_CLEARED) {
4833 /* send an ack now to start the packet flow up again */
4834 call->flags &= ~RX_CALL_CLEARED;
4835 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4837 #ifdef RX_ENABLE_LOCKS
4840 service->nRequestsRunning++;
4841 MUTEX_ENTER(&rx_quota_mutex);
4842 if (service->nRequestsRunning <= service->minProcs)
4845 MUTEX_EXIT(&rx_quota_mutex);
4849 MUTEX_EXIT(&rx_serverPool_lock);
4852 /* Delay the sending of an acknowledge event for a short while, while
4853 * a new call is being prepared (in the case of a client) or a reply
4854 * is being prepared (in the case of a server). Rather than sending
4855 * an ack packet, an ACKALL packet is sent. */
4857 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4859 #ifdef RX_ENABLE_LOCKS
4861 MUTEX_ENTER(&call->lock);
4862 rxevent_Put(call->delayedAckEvent);
4863 call->delayedAckEvent = NULL;
4864 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4866 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4867 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4868 call->flags |= RX_CALL_ACKALL_SENT;
4870 MUTEX_EXIT(&call->lock);
4871 #else /* RX_ENABLE_LOCKS */
4873 rxevent_Put(call->delayedAckEvent);
4874 call->delayedAckEvent = NULL;
4876 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4877 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4878 call->flags |= RX_CALL_ACKALL_SENT;
4879 #endif /* RX_ENABLE_LOCKS */
4883 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4886 struct rx_call *call = arg1;
4887 #ifdef RX_ENABLE_LOCKS
4889 MUTEX_ENTER(&call->lock);
4890 if (event == call->delayedAckEvent) {
4891 rxevent_Put(call->delayedAckEvent);
4892 call->delayedAckEvent = NULL;
4894 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4896 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4898 MUTEX_EXIT(&call->lock);
4899 #else /* RX_ENABLE_LOCKS */
4901 rxevent_Put(call->delayedAckEvent);
4902 call->delayedAckEvent = NULL;
4904 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4905 #endif /* RX_ENABLE_LOCKS */
4909 #ifdef RX_ENABLE_LOCKS
4910 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4911 * clearing them out.
4914 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4916 struct rx_packet *p, *tp;
4919 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4920 p->flags |= RX_PKTFLAG_ACKED;
4924 call->flags |= RX_CALL_TQ_CLEARME;
4925 call->flags |= RX_CALL_TQ_SOME_ACKED;
4928 rxi_rto_cancel(call);
4930 call->tfirst = call->tnext;
4931 call->nSoftAcked = 0;
4933 if (call->flags & RX_CALL_FAST_RECOVER) {
4934 call->flags &= ~RX_CALL_FAST_RECOVER;
4935 call->cwind = call->nextCwind;
4936 call->nextCwind = 0;
4939 CV_SIGNAL(&call->cv_twind);
4941 #endif /* RX_ENABLE_LOCKS */
4943 /* Clear out the transmit queue for the current call (all packets have
4944 * been received by peer) */
4946 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4948 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4949 struct rx_packet *p, *tp;
4951 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4953 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4954 p->flags |= RX_PKTFLAG_ACKED;
4958 call->flags |= RX_CALL_TQ_CLEARME;
4959 call->flags |= RX_CALL_TQ_SOME_ACKED;
4962 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4963 #ifdef RXDEBUG_PACKET
4965 #endif /* RXDEBUG_PACKET */
4966 rxi_FreePackets(0, &call->tq);
4967 rxi_WakeUpTransmitQueue(call);
4968 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4969 call->flags &= ~RX_CALL_TQ_CLEARME;
4971 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4973 rxi_rto_cancel(call);
4974 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4975 call->nSoftAcked = 0;
4977 if (call->flags & RX_CALL_FAST_RECOVER) {
4978 call->flags &= ~RX_CALL_FAST_RECOVER;
4979 call->cwind = call->nextCwind;
4981 #ifdef RX_ENABLE_LOCKS
4982 CV_SIGNAL(&call->cv_twind);
4984 osi_rxWakeup(&call->twind);
4989 rxi_ClearReceiveQueue(struct rx_call *call)
4991 if (queue_IsNotEmpty(&call->rq)) {
4994 count = rxi_FreePackets(0, &call->rq);
4995 rx_packetReclaims += count;
4996 #ifdef RXDEBUG_PACKET
4998 if ( call->rqc != 0 )
4999 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5001 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5003 if (call->state == RX_STATE_PRECALL) {
5004 call->flags |= RX_CALL_CLEARED;
5008 /* Send an abort packet for the specified call */
5010 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5011 int istack, int force)
5013 afs_int32 error, cerror;
5014 struct clock when, now;
5019 switch (call->error) {
5022 cerror = RX_CALL_TIMEOUT;
5025 cerror = call->error;
5028 /* Clients should never delay abort messages */
5029 if (rx_IsClientConn(call->conn))
5032 if (call->abortCode != cerror) {
5033 call->abortCode = cerror;
5034 call->abortCount = 0;
5037 if (force || rxi_callAbortThreshhold == 0
5038 || call->abortCount < rxi_callAbortThreshhold) {
5039 if (call->delayedAbortEvent) {
5040 rxevent_Cancel(&call->delayedAbortEvent, call,
5041 RX_CALL_REFCOUNT_ABORT);
5043 error = htonl(cerror);
5046 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5047 (char *)&error, sizeof(error), istack);
5048 } else if (!call->delayedAbortEvent) {
5049 clock_GetTime(&now);
5051 clock_Addmsec(&when, rxi_callAbortDelay);
5052 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5053 call->delayedAbortEvent =
5054 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5059 /* Send an abort packet for the specified connection. Packet is an
5060 * optional pointer to a packet that can be used to send the abort.
5061 * Once the number of abort messages reaches the threshhold, an
5062 * event is scheduled to send the abort. Setting the force flag
5063 * overrides sending delayed abort messages.
5065 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5066 * to send the abort packet.
5069 rxi_SendConnectionAbort(struct rx_connection *conn,
5070 struct rx_packet *packet, int istack, int force)
5073 struct clock when, now;
5078 /* Clients should never delay abort messages */
5079 if (rx_IsClientConn(conn))
5082 if (force || rxi_connAbortThreshhold == 0
5083 || conn->abortCount < rxi_connAbortThreshhold) {
5085 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
5086 error = htonl(conn->error);
5088 MUTEX_EXIT(&conn->conn_data_lock);
5090 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5091 RX_PACKET_TYPE_ABORT, (char *)&error,
5092 sizeof(error), istack);
5093 MUTEX_ENTER(&conn->conn_data_lock);
5094 } else if (!conn->delayedAbortEvent) {
5095 clock_GetTime(&now);
5097 clock_Addmsec(&when, rxi_connAbortDelay);
5098 conn->delayedAbortEvent =
5099 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
5104 /* Associate an error all of the calls owned by a connection. Called
5105 * with error non-zero. This is only for really fatal things, like
5106 * bad authentication responses. The connection itself is set in
5107 * error at this point, so that future packets received will be
5110 rxi_ConnectionError(struct rx_connection *conn,
5116 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5118 MUTEX_ENTER(&conn->conn_data_lock);
5119 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
5120 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
5121 if (conn->checkReachEvent) {
5122 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
5123 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5124 putConnection(conn);
5126 MUTEX_EXIT(&conn->conn_data_lock);
5127 for (i = 0; i < RX_MAXCALLS; i++) {
5128 struct rx_call *call = conn->call[i];
5130 MUTEX_ENTER(&call->lock);
5131 rxi_CallError(call, error);
5132 MUTEX_EXIT(&call->lock);
5135 conn->error = error;
5136 if (rx_stats_active)
5137 rx_atomic_inc(&rx_stats.fatalErrors);
5142 * Interrupt an in-progress call with the specified error and wakeup waiters.
5144 * @param[in] call The call to interrupt
5145 * @param[in] error The error code to send to the peer
5148 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5150 MUTEX_ENTER(&call->lock);
5151 rxi_CallError(call, error);
5152 rxi_SendCallAbort(call, NULL, 0, 1);
5153 MUTEX_EXIT(&call->lock);
5157 rxi_CallError(struct rx_call *call, afs_int32 error)
5160 osirx_AssertMine(&call->lock, "rxi_CallError");
5162 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5164 error = call->error;
5166 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5167 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5168 rxi_ResetCall(call, 0);
5171 rxi_ResetCall(call, 0);
5173 call->error = error;
5176 /* Reset various fields in a call structure, and wakeup waiting
5177 * processes. Some fields aren't changed: state & mode are not
5178 * touched (these must be set by the caller), and bufptr, nLeft, and
5179 * nFree are not reset, since these fields are manipulated by
5180 * unprotected macros, and may only be reset by non-interrupting code.
5184 rxi_ResetCall(struct rx_call *call, int newcall)
5187 struct rx_peer *peer;
5188 struct rx_packet *packet;
5190 osirx_AssertMine(&call->lock, "rxi_ResetCall");
5192 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5194 /* Notify anyone who is waiting for asynchronous packet arrival */
5195 if (call->arrivalProc) {
5196 (*call->arrivalProc) (call, call->arrivalProcHandle,
5197 call->arrivalProcArg);
5198 call->arrivalProc = (void (*)())0;
5202 rxevent_Cancel(&call->growMTUEvent, call, RX_CALL_REFCOUNT_ALIVE);
5204 if (call->delayedAbortEvent) {
5205 rxevent_Cancel(&call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
5206 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5208 rxi_SendCallAbort(call, packet, 0, 1);
5209 rxi_FreePacket(packet);
5214 * Update the peer with the congestion information in this call
5215 * so other calls on this connection can pick up where this call
5216 * left off. If the congestion sequence numbers don't match then
5217 * another call experienced a retransmission.
5219 peer = call->conn->peer;
5220 MUTEX_ENTER(&peer->peer_lock);
5222 if (call->congestSeq == peer->congestSeq) {
5223 peer->cwind = MAX(peer->cwind, call->cwind);
5224 peer->MTU = MAX(peer->MTU, call->MTU);
5225 peer->nDgramPackets =
5226 MAX(peer->nDgramPackets, call->nDgramPackets);
5229 call->abortCode = 0;
5230 call->abortCount = 0;
5232 if (peer->maxDgramPackets > 1) {
5233 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5235 call->MTU = peer->MTU;
5237 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5238 call->ssthresh = rx_maxSendWindow;
5239 call->nDgramPackets = peer->nDgramPackets;
5240 call->congestSeq = peer->congestSeq;
5241 call->rtt = peer->rtt;
5242 call->rtt_dev = peer->rtt_dev;
5243 clock_Zero(&call->rto);
5244 clock_Addmsec(&call->rto,
5245 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5246 MUTEX_EXIT(&peer->peer_lock);
5248 flags = call->flags;
5249 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5250 rxi_WaitforTQBusy(call);
5251 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5253 rxi_ClearTransmitQueue(call, 1);
5254 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5255 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5259 if (!newcall && (flags & RX_CALL_PEER_BUSY)) {
5260 /* The call channel is still busy; resetting the call doesn't change
5261 * that. However, if 'newcall' is set, we are processing a call
5262 * structure that has either been recycled from the free list, or has
5263 * been newly allocated. So, RX_CALL_PEER_BUSY is not relevant if
5264 * 'newcall' is set, since it describes a completely different call
5265 * channel which we do not care about. */
5266 call->flags |= RX_CALL_PEER_BUSY;
5269 rxi_ClearReceiveQueue(call);
5270 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5274 call->twind = call->conn->twind[call->channel];
5275 call->rwind = call->conn->rwind[call->channel];
5276 call->nSoftAcked = 0;
5277 call->nextCwind = 0;
5280 call->nCwindAcks = 0;
5281 call->nSoftAcks = 0;
5282 call->nHardAcks = 0;
5284 call->tfirst = call->rnext = call->tnext = 1;
5287 call->lastAcked = 0;
5288 call->localStatus = call->remoteStatus = 0;
5290 if (flags & RX_CALL_READER_WAIT) {
5291 #ifdef RX_ENABLE_LOCKS
5292 CV_BROADCAST(&call->cv_rq);
5294 osi_rxWakeup(&call->rq);
5297 if (flags & RX_CALL_WAIT_PACKETS) {
5298 MUTEX_ENTER(&rx_freePktQ_lock);
5299 rxi_PacketsUnWait(); /* XXX */
5300 MUTEX_EXIT(&rx_freePktQ_lock);
5302 #ifdef RX_ENABLE_LOCKS
5303 CV_SIGNAL(&call->cv_twind);
5305 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5306 osi_rxWakeup(&call->twind);
5309 #ifdef RX_ENABLE_LOCKS
5310 /* The following ensures that we don't mess with any queue while some
5311 * other thread might also be doing so. The call_queue_lock field is
5312 * is only modified under the call lock. If the call is in the process
5313 * of being removed from a queue, the call is not locked until the
5314 * the queue lock is dropped and only then is the call_queue_lock field
5315 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5316 * Note that any other routine which removes a call from a queue has to
5317 * obtain the queue lock before examing the queue and removing the call.
5319 if (call->call_queue_lock) {
5320 MUTEX_ENTER(call->call_queue_lock);
5321 if (queue_IsOnQueue(call)) {
5323 if (flags & RX_CALL_WAIT_PROC) {
5324 rx_atomic_dec(&rx_nWaiting);
5327 MUTEX_EXIT(call->call_queue_lock);
5328 CLEAR_CALL_QUEUE_LOCK(call);
5330 #else /* RX_ENABLE_LOCKS */
5331 if (queue_IsOnQueue(call)) {
5333 if (flags & RX_CALL_WAIT_PROC)
5334 rx_atomic_dec(&rx_nWaiting);
5336 #endif /* RX_ENABLE_LOCKS */
5338 rxi_KeepAliveOff(call);
5339 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5342 /* Send an acknowledge for the indicated packet (seq,serial) of the
5343 * indicated call, for the indicated reason (reason). This
5344 * acknowledge will specifically acknowledge receiving the packet, and
5345 * will also specify which other packets for this call have been
5346 * received. This routine returns the packet that was used to the
5347 * caller. The caller is responsible for freeing it or re-using it.
5348 * This acknowledgement also returns the highest sequence number
5349 * actually read out by the higher level to the sender; the sender
5350 * promises to keep around packets that have not been read by the
5351 * higher level yet (unless, of course, the sender decides to abort
5352 * the call altogether). Any of p, seq, serial, pflags, or reason may
5353 * be set to zero without ill effect. That is, if they are zero, they
5354 * will not convey any information.
5355 * NOW there is a trailer field, after the ack where it will safely be
5356 * ignored by mundanes, which indicates the maximum size packet this
5357 * host can swallow. */
5359 struct rx_packet *optionalPacket; use to send ack (or null)
5360 int seq; Sequence number of the packet we are acking
5361 int serial; Serial number of the packet
5362 int pflags; Flags field from packet header
5363 int reason; Reason an acknowledge was prompted
5367 rxi_SendAck(struct rx_call *call,
5368 struct rx_packet *optionalPacket, int serial, int reason,
5371 struct rx_ackPacket *ap;
5372 struct rx_packet *rqp;
5373 struct rx_packet *nxp; /* For queue_Scan */
5374 struct rx_packet *p;
5377 afs_uint32 padbytes = 0;
5378 #ifdef RX_ENABLE_TSFPQ
5379 struct rx_ts_info_t * rx_ts_info;
5383 * Open the receive window once a thread starts reading packets
5385 if (call->rnext > 1) {
5386 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5389 /* Don't attempt to grow MTU if this is a critical ping */
5390 if (reason == RX_ACK_MTU) {
5391 /* keep track of per-call attempts, if we're over max, do in small
5392 * otherwise in larger? set a size to increment by, decrease
5395 if (call->conn->peer->maxPacketSize &&
5396 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5398 padbytes = call->conn->peer->maxPacketSize+16;
5400 padbytes = call->conn->peer->maxMTU + 128;
5402 /* do always try a minimum size ping */
5403 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5405 /* subtract the ack payload */
5406 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5407 reason = RX_ACK_PING;
5410 call->nHardAcks = 0;
5411 call->nSoftAcks = 0;
5412 if (call->rnext > call->lastAcked)
5413 call->lastAcked = call->rnext;
5417 rx_computelen(p, p->length); /* reset length, you never know */
5418 } /* where that's been... */
5419 #ifdef RX_ENABLE_TSFPQ
5421 RX_TS_INFO_GET(rx_ts_info);
5422 if ((p = rx_ts_info->local_special_packet)) {
5423 rx_computelen(p, p->length);
5424 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5425 rx_ts_info->local_special_packet = p;
5426 } else { /* We won't send the ack, but don't panic. */
5427 return optionalPacket;
5431 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5432 /* We won't send the ack, but don't panic. */
5433 return optionalPacket;
5438 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5441 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5442 #ifndef RX_ENABLE_TSFPQ
5443 if (!optionalPacket)
5446 return optionalPacket;
5448 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5449 if (rx_Contiguous(p) < templ) {
5450 #ifndef RX_ENABLE_TSFPQ
5451 if (!optionalPacket)
5454 return optionalPacket;
5459 /* MTUXXX failing to send an ack is very serious. We should */
5460 /* try as hard as possible to send even a partial ack; it's */
5461 /* better than nothing. */
5462 ap = (struct rx_ackPacket *)rx_DataOf(p);
5463 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5464 ap->reason = reason;
5466 /* The skew computation used to be bogus, I think it's better now. */
5467 /* We should start paying attention to skew. XXX */
5468 ap->serial = htonl(serial);
5469 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5472 * First packet not yet forwarded to reader. When ACKALL has been
5473 * sent the peer has been told that all received packets will be
5474 * delivered to the reader. The value 'rnext' is used internally
5475 * to refer to the next packet in the receive queue that must be
5476 * delivered to the reader. From the perspective of the peer it
5477 * already has so report the last sequence number plus one if there
5478 * are packets in the receive queue awaiting processing.
5480 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5481 !queue_IsEmpty(&call->rq)) {
5482 ap->firstPacket = htonl(queue_Last(&call->rq, rx_packet)->header.seq + 1);
5484 ap->firstPacket = htonl(call->rnext);
5486 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5488 /* No fear of running out of ack packet here because there can only be at most
5489 * one window full of unacknowledged packets. The window size must be constrained
5490 * to be less than the maximum ack size, of course. Also, an ack should always
5491 * fit into a single packet -- it should not ever be fragmented. */
5492 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5493 if (!rqp || !call->rq.next
5494 || (rqp->header.seq > (call->rnext + call->rwind))) {
5495 #ifndef RX_ENABLE_TSFPQ
5496 if (!optionalPacket)
5499 rxi_CallError(call, RX_CALL_DEAD);
5500 return optionalPacket;
5503 while (rqp->header.seq > call->rnext + offset)
5504 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5505 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5507 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5508 #ifndef RX_ENABLE_TSFPQ
5509 if (!optionalPacket)
5512 rxi_CallError(call, RX_CALL_DEAD);
5513 return optionalPacket;
5519 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5521 /* these are new for AFS 3.3 */
5522 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5523 templ = htonl(templ);
5524 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5525 templ = htonl(call->conn->peer->ifMTU);
5526 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5527 sizeof(afs_int32), &templ);
5529 /* new for AFS 3.4 */
5530 templ = htonl(call->rwind);
5531 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5532 sizeof(afs_int32), &templ);
5534 /* new for AFS 3.5 */
5535 templ = htonl(call->conn->peer->ifDgramPackets);
5536 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5537 sizeof(afs_int32), &templ);
5539 p->header.serviceId = call->conn->serviceId;
5540 p->header.cid = (call->conn->cid | call->channel);
5541 p->header.callNumber = *call->callNumber;
5543 p->header.securityIndex = call->conn->securityIndex;
5544 p->header.epoch = call->conn->epoch;
5545 p->header.type = RX_PACKET_TYPE_ACK;
5546 p->header.flags = RX_SLOW_START_OK;
5547 if (reason == RX_ACK_PING) {
5548 p->header.flags |= RX_REQUEST_ACK;
5550 p->length = padbytes +
5551 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5554 /* not fast but we can potentially use this if truncated
5555 * fragments are delivered to figure out the mtu.
5557 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5558 sizeof(afs_int32), sizeof(afs_int32),
5562 if (call->conn->type == RX_CLIENT_CONNECTION)
5563 p->header.flags |= RX_CLIENT_INITIATED;
5567 if (rxdebug_active) {
5571 len = _snprintf(msg, sizeof(msg),
5572 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5573 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5574 ntohl(ap->serial), ntohl(ap->previousPacket),
5575 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5576 ap->nAcks, ntohs(ap->bufferSpace) );
5580 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5581 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5585 OutputDebugString(msg);
5587 #else /* AFS_NT40_ENV */
5589 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5590 ap->reason, ntohl(ap->previousPacket),
5591 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5593 for (offset = 0; offset < ap->nAcks; offset++)
5594 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5599 #endif /* AFS_NT40_ENV */
5602 int i, nbytes = p->length;
5604 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5605 if (nbytes <= p->wirevec[i].iov_len) {
5608 savelen = p->wirevec[i].iov_len;
5610 p->wirevec[i].iov_len = nbytes;
5612 rxi_Send(call, p, istack);
5613 p->wirevec[i].iov_len = savelen;
5617 nbytes -= p->wirevec[i].iov_len;
5620 if (rx_stats_active)
5621 rx_atomic_inc(&rx_stats.ackPacketsSent);
5622 #ifndef RX_ENABLE_TSFPQ
5623 if (!optionalPacket)
5626 return optionalPacket; /* Return packet for re-use by caller */
5630 struct rx_packet **list;
5635 /* Send all of the packets in the list in single datagram */
5637 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5638 int istack, int moreFlag)
5644 struct rx_connection *conn = call->conn;
5645 struct rx_peer *peer = conn->peer;
5647 MUTEX_ENTER(&peer->peer_lock);
5648 peer->nSent += xmit->len;
5649 if (xmit->resending)
5650 peer->reSends += xmit->len;
5651 MUTEX_EXIT(&peer->peer_lock);
5653 if (rx_stats_active) {
5654 if (xmit->resending)
5655 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5657 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5660 clock_GetTime(&now);
5662 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5666 /* Set the packet flags and schedule the resend events */
5667 /* Only request an ack for the last packet in the list */
5668 for (i = 0; i < xmit->len; i++) {
5669 struct rx_packet *packet = xmit->list[i];
5671 /* Record the time sent */
5672 packet->timeSent = now;
5673 packet->flags |= RX_PKTFLAG_SENT;
5675 /* Ask for an ack on retransmitted packets, on every other packet
5676 * if the peer doesn't support slow start. Ask for an ack on every
5677 * packet until the congestion window reaches the ack rate. */
5678 if (packet->header.serial) {
5681 packet->firstSent = now;
5682 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5683 || (!(call->flags & RX_CALL_SLOW_START_OK)
5684 && (packet->header.seq & 1)))) {
5689 /* Tag this packet as not being the last in this group,
5690 * for the receiver's benefit */
5691 if (i < xmit->len - 1 || moreFlag) {
5692 packet->header.flags |= RX_MORE_PACKETS;
5697 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5700 /* Since we're about to send a data packet to the peer, it's
5701 * safe to nuke any scheduled end-of-packets ack */
5702 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5704 MUTEX_EXIT(&call->lock);
5705 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5706 if (xmit->len > 1) {
5707 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5709 rxi_SendPacket(call, conn, xmit->list[0], istack);
5711 MUTEX_ENTER(&call->lock);
5712 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5714 /* Tell the RTO calculation engine that we have sent a packet, and
5715 * if it was the last one */
5716 rxi_rto_packet_sent(call, lastPacket, istack);
5718 /* Update last send time for this call (for keep-alive
5719 * processing), and for the connection (so that we can discover
5720 * idle connections) */
5721 conn->lastSendTime = call->lastSendTime = clock_Sec();
5722 /* Let a set of retransmits trigger an idle timeout */
5723 if (!xmit->resending)
5724 call->lastSendData = call->lastSendTime;
5727 /* When sending packets we need to follow these rules:
5728 * 1. Never send more than maxDgramPackets in a jumbogram.
5729 * 2. Never send a packet with more than two iovecs in a jumbogram.
5730 * 3. Never send a retransmitted packet in a jumbogram.
5731 * 4. Never send more than cwind/4 packets in a jumbogram
5732 * We always keep the last list we should have sent so we
5733 * can set the RX_MORE_PACKETS flags correctly.
5737 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5742 struct xmitlist working;
5743 struct xmitlist last;
5745 struct rx_peer *peer = call->conn->peer;
5746 int morePackets = 0;
5748 memset(&last, 0, sizeof(struct xmitlist));
5749 working.list = &list[0];
5751 working.resending = 0;
5753 recovery = call->flags & RX_CALL_FAST_RECOVER;
5755 for (i = 0; i < len; i++) {
5756 /* Does the current packet force us to flush the current list? */
5758 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5759 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5761 /* This sends the 'last' list and then rolls the current working
5762 * set into the 'last' one, and resets the working set */
5765 rxi_SendList(call, &last, istack, 1);
5766 /* If the call enters an error state stop sending, or if
5767 * we entered congestion recovery mode, stop sending */
5769 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5774 working.resending = 0;
5775 working.list = &list[i];
5777 /* Add the current packet to the list if it hasn't been acked.
5778 * Otherwise adjust the list pointer to skip the current packet. */
5779 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5782 if (list[i]->header.serial)
5783 working.resending = 1;
5785 /* Do we need to flush the list? */
5786 if (working.len >= (int)peer->maxDgramPackets
5787 || working.len >= (int)call->nDgramPackets
5788 || working.len >= (int)call->cwind
5789 || list[i]->header.serial
5790 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5792 rxi_SendList(call, &last, istack, 1);
5793 /* If the call enters an error state stop sending, or if
5794 * we entered congestion recovery mode, stop sending */
5796 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5801 working.resending = 0;
5802 working.list = &list[i + 1];
5805 if (working.len != 0) {
5806 osi_Panic("rxi_SendList error");
5808 working.list = &list[i + 1];
5812 /* Send the whole list when the call is in receive mode, when
5813 * the call is in eof mode, when we are in fast recovery mode,
5814 * and when we have the last packet */
5815 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5816 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5817 || (call->flags & RX_CALL_FAST_RECOVER)) {
5818 /* Check for the case where the current list contains
5819 * an acked packet. Since we always send retransmissions
5820 * in a separate packet, we only need to check the first
5821 * packet in the list */
5822 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5826 rxi_SendList(call, &last, istack, morePackets);
5827 /* If the call enters an error state stop sending, or if
5828 * we entered congestion recovery mode, stop sending */
5830 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5834 rxi_SendList(call, &working, istack, 0);
5836 } else if (last.len > 0) {
5837 rxi_SendList(call, &last, istack, 0);
5838 /* Packets which are in 'working' are not sent by this call */
5843 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
5845 struct rx_call *call = arg0;
5846 struct rx_peer *peer;
5847 struct rx_packet *p, *nxp;
5848 struct clock maxTimeout = { 60, 0 };
5850 MUTEX_ENTER(&call->lock);
5852 peer = call->conn->peer;
5854 /* Make sure that the event pointer is removed from the call
5855 * structure, since there is no longer a per-call retransmission
5857 if (event == call->resendEvent) {
5858 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5859 rxevent_Put(call->resendEvent);
5860 call->resendEvent = NULL;
5863 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
5864 rxi_CheckBusy(call);
5867 if (queue_IsEmpty(&call->tq)) {
5868 /* Nothing to do. This means that we've been raced, and that an
5869 * ACK has come in between when we were triggered, and when we
5870 * actually got to run. */
5874 /* We're in loss recovery */
5875 call->flags |= RX_CALL_FAST_RECOVER;
5877 /* Mark all of the pending packets in the queue as being lost */
5878 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5879 if (!(p->flags & RX_PKTFLAG_ACKED))
5880 p->flags &= ~RX_PKTFLAG_SENT;
5883 /* We're resending, so we double the timeout of the call. This will be
5884 * dropped back down by the first successful ACK that we receive.
5886 * We apply a maximum value here of 60 seconds
5888 clock_Add(&call->rto, &call->rto);
5889 if (clock_Gt(&call->rto, &maxTimeout))
5890 call->rto = maxTimeout;
5892 /* Packet loss is most likely due to congestion, so drop our window size
5893 * and start again from the beginning */
5894 if (peer->maxDgramPackets >1) {
5895 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5896 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5898 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5899 call->nDgramPackets = 1;
5901 call->nextCwind = 1;
5904 MUTEX_ENTER(&peer->peer_lock);
5905 peer->MTU = call->MTU;
5906 peer->cwind = call->cwind;
5907 peer->nDgramPackets = 1;
5909 call->congestSeq = peer->congestSeq;
5910 MUTEX_EXIT(&peer->peer_lock);
5912 rxi_Start(call, istack);
5915 MUTEX_EXIT(&call->lock);
5918 /* This routine is called when new packets are readied for
5919 * transmission and when retransmission may be necessary, or when the
5920 * transmission window or burst count are favourable. This should be
5921 * better optimized for new packets, the usual case, now that we've
5922 * got rid of queues of send packets. XXXXXXXXXXX */
5924 rxi_Start(struct rx_call *call, int istack)
5927 struct rx_packet *p;
5928 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5933 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5934 if (rx_stats_active)
5935 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5940 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5942 /* Send (or resend) any packets that need it, subject to
5943 * window restrictions and congestion burst control
5944 * restrictions. Ask for an ack on the last packet sent in
5945 * this burst. For now, we're relying upon the window being
5946 * considerably bigger than the largest number of packets that
5947 * are typically sent at once by one initial call to
5948 * rxi_Start. This is probably bogus (perhaps we should ask
5949 * for an ack when we're half way through the current
5950 * window?). Also, for non file transfer applications, this
5951 * may end up asking for an ack for every packet. Bogus. XXXX
5954 * But check whether we're here recursively, and let the other guy
5957 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5958 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5959 call->flags |= RX_CALL_TQ_BUSY;
5961 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5963 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5964 call->flags &= ~RX_CALL_NEED_START;
5965 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5967 maxXmitPackets = MIN(call->twind, call->cwind);
5968 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5969 #ifdef RX_TRACK_PACKETS
5970 if ((p->flags & RX_PKTFLAG_FREE)
5971 || (!queue_IsEnd(&call->tq, nxp)
5972 && (nxp->flags & RX_PKTFLAG_FREE))
5973 || (p == (struct rx_packet *)&rx_freePacketQueue)
5974 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5975 osi_Panic("rxi_Start: xmit queue clobbered");
5978 if (p->flags & RX_PKTFLAG_ACKED) {
5979 /* Since we may block, don't trust this */
5980 if (rx_stats_active)
5981 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
5982 continue; /* Ignore this packet if it has been acknowledged */
5985 /* Turn off all flags except these ones, which are the same
5986 * on each transmission */
5987 p->header.flags &= RX_PRESET_FLAGS;
5989 if (p->header.seq >=
5990 call->tfirst + MIN((int)call->twind,
5991 (int)(call->nSoftAcked +
5993 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5994 /* Note: if we're waiting for more window space, we can
5995 * still send retransmits; hence we don't return here, but
5996 * break out to schedule a retransmit event */
5997 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
5998 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6003 /* Transmit the packet if it needs to be sent. */
6004 if (!(p->flags & RX_PKTFLAG_SENT)) {
6005 if (nXmitPackets == maxXmitPackets) {
6006 rxi_SendXmitList(call, call->xmitList,
6007 nXmitPackets, istack);
6010 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6011 *(call->callNumber), p));
6012 call->xmitList[nXmitPackets++] = p;
6016 /* xmitList now hold pointers to all of the packets that are
6017 * ready to send. Now we loop to send the packets */
6018 if (nXmitPackets > 0) {
6019 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6023 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6025 /* We went into the error state while sending packets. Now is
6026 * the time to reset the call. This will also inform the using
6027 * process that the call is in an error state.
6029 if (rx_stats_active)
6030 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6031 call->flags &= ~RX_CALL_TQ_BUSY;
6032 rxi_WakeUpTransmitQueue(call);
6033 rxi_CallError(call, call->error);
6036 #ifdef RX_ENABLE_LOCKS
6037 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6039 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6040 /* Some packets have received acks. If they all have, we can clear
6041 * the transmit queue.
6044 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
6045 if (p->header.seq < call->tfirst
6046 && (p->flags & RX_PKTFLAG_ACKED)) {
6048 #ifdef RX_TRACK_PACKETS
6049 p->flags &= ~RX_PKTFLAG_TQ;
6051 #ifdef RXDEBUG_PACKET
6059 call->flags |= RX_CALL_TQ_CLEARME;
6061 #endif /* RX_ENABLE_LOCKS */
6062 if (call->flags & RX_CALL_TQ_CLEARME)
6063 rxi_ClearTransmitQueue(call, 1);
6064 } while (call->flags & RX_CALL_NEED_START);
6066 * TQ references no longer protected by this flag; they must remain
6067 * protected by the global lock.
6069 call->flags &= ~RX_CALL_TQ_BUSY;
6070 rxi_WakeUpTransmitQueue(call);
6072 call->flags |= RX_CALL_NEED_START;
6074 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6076 rxi_rto_cancel(call);
6080 /* Also adjusts the keep alive parameters for the call, to reflect
6081 * that we have just sent a packet (so keep alives aren't sent
6084 rxi_Send(struct rx_call *call, struct rx_packet *p,
6087 struct rx_connection *conn = call->conn;
6089 /* Stamp each packet with the user supplied status */
6090 p->header.userStatus = call->localStatus;
6092 /* Allow the security object controlling this call's security to
6093 * make any last-minute changes to the packet */
6094 RXS_SendPacket(conn->securityObject, call, p);
6096 /* Since we're about to send SOME sort of packet to the peer, it's
6097 * safe to nuke any scheduled end-of-packets ack */
6098 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
6100 /* Actually send the packet, filling in more connection-specific fields */
6101 MUTEX_EXIT(&call->lock);
6102 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6103 rxi_SendPacket(call, conn, p, istack);
6104 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6105 MUTEX_ENTER(&call->lock);
6107 /* Update last send time for this call (for keep-alive
6108 * processing), and for the connection (so that we can discover
6109 * idle connections) */
6110 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6111 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6112 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6114 conn->lastSendTime = call->lastSendTime = clock_Sec();
6115 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6116 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6117 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6118 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6119 RX_ACK_PING_RESPONSE)))
6120 call->lastSendData = call->lastSendTime;
6124 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6125 * that things are fine. Also called periodically to guarantee that nothing
6126 * falls through the cracks (e.g. (error + dally) connections have keepalive
6127 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6129 * haveCTLock Set if calling from rxi_ReapConnections
6131 #ifdef RX_ENABLE_LOCKS
6133 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6134 #else /* RX_ENABLE_LOCKS */
6136 rxi_CheckCall(struct rx_call *call)
6137 #endif /* RX_ENABLE_LOCKS */
6139 struct rx_connection *conn = call->conn;
6141 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6142 afs_uint32 fudgeFactor;
6145 int idle_timeout = 0;
6146 afs_int32 clock_diff = 0;
6150 /* Large swings in the clock can have a significant impact on
6151 * the performance of RX call processing. Forward clock shifts
6152 * will result in premature event triggering or timeouts.
6153 * Backward shifts can result in calls not completing until
6154 * the clock catches up with the original start clock value.
6156 * If a backward clock shift of more than five minutes is noticed,
6157 * just fail the call.
6159 if (now < call->lastSendTime)
6160 clock_diff = call->lastSendTime - now;
6161 if (now < call->startWait)
6162 clock_diff = MAX(clock_diff, call->startWait - now);
6163 if (now < call->lastReceiveTime)
6164 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6165 if (clock_diff > 5 * 60)
6167 if (call->state == RX_STATE_ACTIVE)
6168 rxi_CallError(call, RX_CALL_TIMEOUT);
6172 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6173 if (call->flags & RX_CALL_TQ_BUSY) {
6174 /* Call is active and will be reset by rxi_Start if it's
6175 * in an error state.
6180 /* RTT + 8*MDEV, rounded up to the next second. */
6181 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6182 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6184 deadTime = conn->secondsUntilDead + fudgeFactor;
6185 /* These are computed to the second (+- 1 second). But that's
6186 * good enough for these values, which should be a significant
6187 * number of seconds. */
6188 if (now > (call->lastReceiveTime + deadTime)) {
6189 if (call->state == RX_STATE_ACTIVE) {
6191 #if defined(KERNEL) && defined(AFS_SUN5_ENV)
6193 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6194 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6195 ip_stack_t *ipst = ns->netstack_ip;
6197 ire = ire_cache_lookup(conn->peer->host
6198 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6200 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
6202 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6209 if (ire && ire->ire_max_frag > 0)
6210 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6212 #if defined(GLOBAL_NETSTACKID)
6216 #endif /* ADAPT_PMTU */
6217 cerror = RX_CALL_DEAD;
6220 #ifdef RX_ENABLE_LOCKS
6221 /* Cancel pending events */
6222 rxevent_Cancel(&call->delayedAckEvent, call,
6223 RX_CALL_REFCOUNT_DELAY);
6224 rxi_rto_cancel(call);
6225 rxevent_Cancel(&call->keepAliveEvent, call,
6226 RX_CALL_REFCOUNT_ALIVE);
6227 rxevent_Cancel(&call->growMTUEvent, call,
6228 RX_CALL_REFCOUNT_ALIVE);
6229 MUTEX_ENTER(&rx_refcnt_mutex);
6230 if (call->refCount == 0) {
6231 rxi_FreeCall(call, haveCTLock);
6232 MUTEX_EXIT(&rx_refcnt_mutex);
6235 MUTEX_EXIT(&rx_refcnt_mutex);
6237 #else /* RX_ENABLE_LOCKS */
6238 rxi_FreeCall(call, 0);
6240 #endif /* RX_ENABLE_LOCKS */
6242 /* Non-active calls are destroyed if they are not responding
6243 * to pings; active calls are simply flagged in error, so the
6244 * attached process can die reasonably gracefully. */
6247 if (conn->idleDeadDetection) {
6248 if (conn->idleDeadTime) {
6249 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6253 /* see if we have a non-activity timeout */
6254 if (call->startWait && ((call->startWait + idleDeadTime) < now) &&
6255 (call->flags & RX_CALL_READER_WAIT)) {
6256 if (call->state == RX_STATE_ACTIVE) {
6257 cerror = RX_CALL_TIMEOUT;
6262 if (call->lastSendData && ((call->lastSendData + idleDeadTime) < now)) {
6263 if (call->state == RX_STATE_ACTIVE) {
6264 cerror = conn->service ? conn->service->idleDeadErr : RX_CALL_IDLE;
6272 if (conn->hardDeadTime) {
6273 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6276 /* see if we have a hard timeout */
6278 && (now > (hardDeadTime + call->startTime.sec))) {
6279 if (call->state == RX_STATE_ACTIVE)
6280 rxi_CallError(call, RX_CALL_TIMEOUT);
6285 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6286 call->lastReceiveTime) {
6287 int oldMTU = conn->peer->ifMTU;
6289 /* if we thought we could send more, perhaps things got worse */
6290 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6291 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6292 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6293 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6295 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6297 /* minimum capped in SetPeerMtu */
6298 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6301 conn->lastPacketSize = 0;
6303 /* needed so ResetCall doesn't clobber us. */
6304 call->MTU = conn->peer->ifMTU;
6306 /* if we never succeeded, let the error pass out as-is */
6307 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6308 cerror = conn->msgsizeRetryErr;
6311 rxi_CallError(call, cerror);
6316 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6317 void *dummy, int dummy2)
6319 struct rx_connection *conn = arg1;
6320 struct rx_header theader;
6321 char tbuffer[1 + sizeof(struct rx_header)];
6322 struct sockaddr_in taddr;
6325 struct iovec tmpiov[2];
6328 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6331 tp = &tbuffer[sizeof(struct rx_header)];
6332 taddr.sin_family = AF_INET;
6333 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6334 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6335 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6336 taddr.sin_len = sizeof(struct sockaddr_in);
6338 memset(&theader, 0, sizeof(theader));
6339 theader.epoch = htonl(999);
6341 theader.callNumber = 0;
6344 theader.type = RX_PACKET_TYPE_VERSION;
6345 theader.flags = RX_LAST_PACKET;
6346 theader.serviceId = 0;
6348 memcpy(tbuffer, &theader, sizeof(theader));
6349 memcpy(tp, &a, sizeof(a));
6350 tmpiov[0].iov_base = tbuffer;
6351 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6353 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6355 MUTEX_ENTER(&conn->conn_data_lock);
6356 MUTEX_ENTER(&rx_refcnt_mutex);
6357 /* Only reschedule ourselves if the connection would not be destroyed */
6358 if (conn->refCount <= 1) {
6359 rxevent_Put(conn->natKeepAliveEvent);
6360 conn->natKeepAliveEvent = NULL;
6361 MUTEX_EXIT(&rx_refcnt_mutex);
6362 MUTEX_EXIT(&conn->conn_data_lock);
6363 rx_DestroyConnection(conn); /* drop the reference for this */
6365 conn->refCount--; /* drop the reference for this */
6366 MUTEX_EXIT(&rx_refcnt_mutex);
6367 rxevent_Put(conn->natKeepAliveEvent);
6368 conn->natKeepAliveEvent = NULL;
6369 rxi_ScheduleNatKeepAliveEvent(conn);
6370 MUTEX_EXIT(&conn->conn_data_lock);
6375 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6377 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6378 struct clock when, now;
6379 clock_GetTime(&now);
6381 when.sec += conn->secondsUntilNatPing;
6382 MUTEX_ENTER(&rx_refcnt_mutex);
6383 conn->refCount++; /* hold a reference for this */
6384 MUTEX_EXIT(&rx_refcnt_mutex);
6385 conn->natKeepAliveEvent =
6386 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6391 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6393 MUTEX_ENTER(&conn->conn_data_lock);
6394 conn->secondsUntilNatPing = seconds;
6396 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6397 rxi_ScheduleNatKeepAliveEvent(conn);
6399 conn->flags |= RX_CONN_NAT_PING;
6401 MUTEX_EXIT(&conn->conn_data_lock);
6405 rxi_NatKeepAliveOn(struct rx_connection *conn)
6407 MUTEX_ENTER(&conn->conn_data_lock);
6408 /* if it's already attached */
6409 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6410 rxi_ScheduleNatKeepAliveEvent(conn);
6412 conn->flags |= RX_CONN_NAT_PING;
6413 MUTEX_EXIT(&conn->conn_data_lock);
6416 /* When a call is in progress, this routine is called occasionally to
6417 * make sure that some traffic has arrived (or been sent to) the peer.
6418 * If nothing has arrived in a reasonable amount of time, the call is
6419 * declared dead; if nothing has been sent for a while, we send a
6420 * keep-alive packet (if we're actually trying to keep the call alive)
6423 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6426 struct rx_call *call = arg1;
6427 struct rx_connection *conn;
6430 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6431 MUTEX_ENTER(&call->lock);
6433 if (event == call->keepAliveEvent) {
6434 rxevent_Put(call->keepAliveEvent);
6435 call->keepAliveEvent = NULL;
6440 #ifdef RX_ENABLE_LOCKS
6441 if (rxi_CheckCall(call, 0)) {
6442 MUTEX_EXIT(&call->lock);
6445 #else /* RX_ENABLE_LOCKS */
6446 if (rxi_CheckCall(call))
6448 #endif /* RX_ENABLE_LOCKS */
6450 /* Don't try to keep alive dallying calls */
6451 if (call->state == RX_STATE_DALLY) {
6452 MUTEX_EXIT(&call->lock);
6457 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6458 /* Don't try to send keepalives if there is unacknowledged data */
6459 /* the rexmit code should be good enough, this little hack
6460 * doesn't quite work XXX */
6461 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6463 rxi_ScheduleKeepAliveEvent(call);
6464 MUTEX_EXIT(&call->lock);
6467 /* Does what's on the nameplate. */
6469 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6471 struct rx_call *call = arg1;
6472 struct rx_connection *conn;
6474 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6475 MUTEX_ENTER(&call->lock);
6477 if (event == call->growMTUEvent) {
6478 rxevent_Put(call->growMTUEvent);
6479 call->growMTUEvent = NULL;
6482 #ifdef RX_ENABLE_LOCKS
6483 if (rxi_CheckCall(call, 0)) {
6484 MUTEX_EXIT(&call->lock);
6487 #else /* RX_ENABLE_LOCKS */
6488 if (rxi_CheckCall(call))
6490 #endif /* RX_ENABLE_LOCKS */
6492 /* Don't bother with dallying calls */
6493 if (call->state == RX_STATE_DALLY) {
6494 MUTEX_EXIT(&call->lock);
6501 * keep being scheduled, just don't do anything if we're at peak,
6502 * or we're not set up to be properly handled (idle timeout required)
6504 if ((conn->peer->maxPacketSize != 0) &&
6505 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6506 conn->idleDeadDetection)
6507 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6508 rxi_ScheduleGrowMTUEvent(call, 0);
6509 MUTEX_EXIT(&call->lock);
6513 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6515 if (!call->keepAliveEvent) {
6516 struct clock when, now;
6517 clock_GetTime(&now);
6519 when.sec += call->conn->secondsUntilPing;
6520 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6521 call->keepAliveEvent =
6522 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6527 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6529 if (!call->growMTUEvent) {
6530 struct clock when, now;
6532 clock_GetTime(&now);
6535 if (call->conn->secondsUntilPing)
6536 secs = (6*call->conn->secondsUntilPing)-1;
6538 if (call->conn->secondsUntilDead)
6539 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6543 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6544 call->growMTUEvent =
6545 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6549 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6551 rxi_KeepAliveOn(struct rx_call *call)
6553 /* Pretend last packet received was received now--i.e. if another
6554 * packet isn't received within the keep alive time, then the call
6555 * will die; Initialize last send time to the current time--even
6556 * if a packet hasn't been sent yet. This will guarantee that a
6557 * keep-alive is sent within the ping time */
6558 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6559 rxi_ScheduleKeepAliveEvent(call);
6563 * Solely in order that callers not need to include rx_call.h
6566 rx_KeepAliveOff(struct rx_call *call)
6568 rxi_KeepAliveOff(call);
6571 rx_KeepAliveOn(struct rx_call *call)
6573 rxi_KeepAliveOn(call);
6577 rxi_GrowMTUOn(struct rx_call *call)
6579 struct rx_connection *conn = call->conn;
6580 MUTEX_ENTER(&conn->conn_data_lock);
6581 conn->lastPingSizeSer = conn->lastPingSize = 0;
6582 MUTEX_EXIT(&conn->conn_data_lock);
6583 rxi_ScheduleGrowMTUEvent(call, 1);
6586 /* This routine is called to send connection abort messages
6587 * that have been delayed to throttle looping clients. */
6589 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6592 struct rx_connection *conn = arg1;
6595 struct rx_packet *packet;
6597 MUTEX_ENTER(&conn->conn_data_lock);
6598 rxevent_Put(conn->delayedAbortEvent);
6599 conn->delayedAbortEvent = NULL;
6600 error = htonl(conn->error);
6602 MUTEX_EXIT(&conn->conn_data_lock);
6603 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6606 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6607 RX_PACKET_TYPE_ABORT, (char *)&error,
6609 rxi_FreePacket(packet);
6613 /* This routine is called to send call abort messages
6614 * that have been delayed to throttle looping clients. */
6616 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6619 struct rx_call *call = arg1;
6622 struct rx_packet *packet;
6624 MUTEX_ENTER(&call->lock);
6625 rxevent_Put(call->delayedAbortEvent);
6626 call->delayedAbortEvent = NULL;
6627 error = htonl(call->error);
6629 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6632 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6633 (char *)&error, sizeof(error), 0);
6634 rxi_FreePacket(packet);
6636 MUTEX_EXIT(&call->lock);
6637 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6640 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6641 * seconds) to ask the client to authenticate itself. The routine
6642 * issues a challenge to the client, which is obtained from the
6643 * security object associated with the connection */
6645 rxi_ChallengeEvent(struct rxevent *event,
6646 void *arg0, void *arg1, int tries)
6648 struct rx_connection *conn = arg0;
6651 rxevent_Put(conn->challengeEvent);
6652 conn->challengeEvent = NULL;
6655 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6656 struct rx_packet *packet;
6657 struct clock when, now;
6660 /* We've failed to authenticate for too long.
6661 * Reset any calls waiting for authentication;
6662 * they are all in RX_STATE_PRECALL.
6666 MUTEX_ENTER(&conn->conn_call_lock);
6667 for (i = 0; i < RX_MAXCALLS; i++) {
6668 struct rx_call *call = conn->call[i];
6670 MUTEX_ENTER(&call->lock);
6671 if (call->state == RX_STATE_PRECALL) {
6672 rxi_CallError(call, RX_CALL_DEAD);
6673 rxi_SendCallAbort(call, NULL, 0, 0);
6675 MUTEX_EXIT(&call->lock);
6678 MUTEX_EXIT(&conn->conn_call_lock);
6682 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6684 /* If there's no packet available, do this later. */
6685 RXS_GetChallenge(conn->securityObject, conn, packet);
6686 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6687 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6688 rxi_FreePacket(packet);
6690 clock_GetTime(&now);
6692 when.sec += RX_CHALLENGE_TIMEOUT;
6693 conn->challengeEvent =
6694 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6699 /* Call this routine to start requesting the client to authenticate
6700 * itself. This will continue until authentication is established,
6701 * the call times out, or an invalid response is returned. The
6702 * security object associated with the connection is asked to create
6703 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6704 * defined earlier. */
6706 rxi_ChallengeOn(struct rx_connection *conn)
6708 if (!conn->challengeEvent) {
6709 RXS_CreateChallenge(conn->securityObject, conn);
6710 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6715 /* rxi_ComputeRoundTripTime is called with peer locked. */
6716 /* peer may be null */
6718 rxi_ComputeRoundTripTime(struct rx_packet *p,
6719 struct rx_ackPacket *ack,
6720 struct rx_call *call,
6721 struct rx_peer *peer,
6724 struct clock thisRtt, *sentp;
6728 /* If the ACK is delayed, then do nothing */
6729 if (ack->reason == RX_ACK_DELAY)
6732 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6733 * their RTT multiple times, so only include the RTT of the last packet
6735 if (p->flags & RX_JUMBO_PACKET)
6738 /* Use the serial number to determine which transmission the ACK is for,
6739 * and set the sent time to match this. If we have no serial number, then
6740 * only use the ACK for RTT calculations if the packet has not been
6744 serial = ntohl(ack->serial);
6746 if (serial == p->header.serial) {
6747 sentp = &p->timeSent;
6748 } else if (serial == p->firstSerial) {
6749 sentp = &p->firstSent;
6750 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6751 sentp = &p->firstSent;
6755 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6756 sentp = &p->firstSent;
6763 if (clock_Lt(&thisRtt, sentp))
6764 return; /* somebody set the clock back, don't count this time. */
6766 clock_Sub(&thisRtt, sentp);
6767 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6768 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6770 if (clock_IsZero(&thisRtt)) {
6772 * The actual round trip time is shorter than the
6773 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6774 * Since we can't tell which at the moment we will assume 1ms.
6776 thisRtt.usec = 1000;
6779 if (rx_stats_active) {
6780 MUTEX_ENTER(&rx_stats_mutex);
6781 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6782 rx_stats.minRtt = thisRtt;
6783 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6784 if (thisRtt.sec > 60) {
6785 MUTEX_EXIT(&rx_stats_mutex);
6786 return; /* somebody set the clock ahead */
6788 rx_stats.maxRtt = thisRtt;
6790 clock_Add(&rx_stats.totalRtt, &thisRtt);
6791 rx_atomic_inc(&rx_stats.nRttSamples);
6792 MUTEX_EXIT(&rx_stats_mutex);
6795 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6797 /* Apply VanJacobson round-trip estimations */
6802 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6803 * srtt is stored as fixed point with 3 bits after the binary
6804 * point (i.e., scaled by 8). The following magic is
6805 * equivalent to the smoothing algorithm in rfc793 with an
6806 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6807 * srtt'*8 = rtt + srtt*7
6808 * srtt'*8 = srtt*8 + rtt - srtt
6809 * srtt' = srtt + rtt/8 - srtt/8
6810 * srtt' = srtt + (rtt - srtt)/8
6813 delta = _8THMSEC(&thisRtt) - call->rtt;
6814 call->rtt += (delta >> 3);
6817 * We accumulate a smoothed rtt variance (actually, a smoothed
6818 * mean difference), then set the retransmit timer to smoothed
6819 * rtt + 4 times the smoothed variance (was 2x in van's original
6820 * paper, but 4x works better for me, and apparently for him as
6822 * rttvar is stored as
6823 * fixed point with 2 bits after the binary point (scaled by
6824 * 4). The following is equivalent to rfc793 smoothing with
6825 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6826 * rttvar'*4 = rttvar*3 + |delta|
6827 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6828 * rttvar' = rttvar + |delta|/4 - rttvar/4
6829 * rttvar' = rttvar + (|delta| - rttvar)/4
6830 * This replaces rfc793's wired-in beta.
6831 * dev*4 = dev*4 + (|actual - expected| - dev)
6837 delta -= (call->rtt_dev << 1);
6838 call->rtt_dev += (delta >> 3);
6840 /* I don't have a stored RTT so I start with this value. Since I'm
6841 * probably just starting a call, and will be pushing more data down
6842 * this, I expect congestion to increase rapidly. So I fudge a
6843 * little, and I set deviance to half the rtt. In practice,
6844 * deviance tends to approach something a little less than
6845 * half the smoothed rtt. */
6846 call->rtt = _8THMSEC(&thisRtt) + 8;
6847 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6849 /* the smoothed RTT time is RTT + 4*MDEV
6851 * We allow a user specified minimum to be set for this, to allow clamping
6852 * at a minimum value in the same way as TCP. In addition, we have to allow
6853 * for the possibility that this packet is answered by a delayed ACK, so we
6854 * add on a fixed 200ms to account for that timer expiring.
6857 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6858 rx_minPeerTimeout) + 200;
6859 clock_Zero(&call->rto);
6860 clock_Addmsec(&call->rto, rtt_timeout);
6862 /* Update the peer, so any new calls start with our values */
6863 peer->rtt_dev = call->rtt_dev;
6864 peer->rtt = call->rtt;
6866 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6867 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6871 /* Find all server connections that have not been active for a long time, and
6874 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
6877 struct clock now, when;
6878 clock_GetTime(&now);
6880 /* Find server connection structures that haven't been used for
6881 * greater than rx_idleConnectionTime */
6883 struct rx_connection **conn_ptr, **conn_end;
6884 int i, havecalls = 0;
6885 MUTEX_ENTER(&rx_connHashTable_lock);
6886 for (conn_ptr = &rx_connHashTable[0], conn_end =
6887 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6889 struct rx_connection *conn, *next;
6890 struct rx_call *call;
6894 for (conn = *conn_ptr; conn; conn = next) {
6895 /* XXX -- Shouldn't the connection be locked? */
6898 for (i = 0; i < RX_MAXCALLS; i++) {
6899 call = conn->call[i];
6903 code = MUTEX_TRYENTER(&call->lock);
6906 #ifdef RX_ENABLE_LOCKS
6907 result = rxi_CheckCall(call, 1);
6908 #else /* RX_ENABLE_LOCKS */
6909 result = rxi_CheckCall(call);
6910 #endif /* RX_ENABLE_LOCKS */
6911 MUTEX_EXIT(&call->lock);
6913 /* If CheckCall freed the call, it might
6914 * have destroyed the connection as well,
6915 * which screws up the linked lists.
6921 if (conn->type == RX_SERVER_CONNECTION) {
6922 /* This only actually destroys the connection if
6923 * there are no outstanding calls */
6924 MUTEX_ENTER(&conn->conn_data_lock);
6925 MUTEX_ENTER(&rx_refcnt_mutex);
6926 if (!havecalls && !conn->refCount
6927 && ((conn->lastSendTime + rx_idleConnectionTime) <
6929 conn->refCount++; /* it will be decr in rx_DestroyConn */
6930 MUTEX_EXIT(&rx_refcnt_mutex);
6931 MUTEX_EXIT(&conn->conn_data_lock);
6932 #ifdef RX_ENABLE_LOCKS
6933 rxi_DestroyConnectionNoLock(conn);
6934 #else /* RX_ENABLE_LOCKS */
6935 rxi_DestroyConnection(conn);
6936 #endif /* RX_ENABLE_LOCKS */
6938 #ifdef RX_ENABLE_LOCKS
6940 MUTEX_EXIT(&rx_refcnt_mutex);
6941 MUTEX_EXIT(&conn->conn_data_lock);
6943 #endif /* RX_ENABLE_LOCKS */
6947 #ifdef RX_ENABLE_LOCKS
6948 while (rx_connCleanup_list) {
6949 struct rx_connection *conn;
6950 conn = rx_connCleanup_list;
6951 rx_connCleanup_list = rx_connCleanup_list->next;
6952 MUTEX_EXIT(&rx_connHashTable_lock);
6953 rxi_CleanupConnection(conn);
6954 MUTEX_ENTER(&rx_connHashTable_lock);
6956 MUTEX_EXIT(&rx_connHashTable_lock);
6957 #endif /* RX_ENABLE_LOCKS */
6960 /* Find any peer structures that haven't been used (haven't had an
6961 * associated connection) for greater than rx_idlePeerTime */
6963 struct rx_peer **peer_ptr, **peer_end;
6967 * Why do we need to hold the rx_peerHashTable_lock across
6968 * the incrementing of peer_ptr since the rx_peerHashTable
6969 * array is not changing? We don't.
6971 * By dropping the lock periodically we can permit other
6972 * activities to be performed while a rxi_ReapConnections
6973 * call is in progress. The goal of reap connections
6974 * is to clean up quickly without causing large amounts
6975 * of contention. Therefore, it is important that global
6976 * mutexes not be held for extended periods of time.
6978 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6979 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6981 struct rx_peer *peer, *next, *prev;
6983 MUTEX_ENTER(&rx_peerHashTable_lock);
6984 for (prev = peer = *peer_ptr; peer; peer = next) {
6986 code = MUTEX_TRYENTER(&peer->peer_lock);
6987 if ((code) && (peer->refCount == 0)
6988 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6989 rx_interface_stat_p rpc_stat, nrpc_stat;
6993 * now know that this peer object is one to be
6994 * removed from the hash table. Once it is removed
6995 * it can't be referenced by other threads.
6996 * Lets remove it first and decrement the struct
6997 * nPeerStructs count.
6999 if (peer == *peer_ptr) {
7005 if (rx_stats_active)
7006 rx_atomic_dec(&rx_stats.nPeerStructs);
7009 * Now if we hold references on 'prev' and 'next'
7010 * we can safely drop the rx_peerHashTable_lock
7011 * while we destroy this 'peer' object.
7017 MUTEX_EXIT(&rx_peerHashTable_lock);
7019 MUTEX_EXIT(&peer->peer_lock);
7020 MUTEX_DESTROY(&peer->peer_lock);
7022 (&peer->rpcStats, rpc_stat, nrpc_stat,
7023 rx_interface_stat)) {
7024 unsigned int num_funcs;
7027 queue_Remove(&rpc_stat->queue_header);
7028 queue_Remove(&rpc_stat->all_peers);
7029 num_funcs = rpc_stat->stats[0].func_total;
7031 sizeof(rx_interface_stat_t) +
7032 rpc_stat->stats[0].func_total *
7033 sizeof(rx_function_entry_v1_t);
7035 rxi_Free(rpc_stat, space);
7037 MUTEX_ENTER(&rx_rpc_stats);
7038 rxi_rpc_peer_stat_cnt -= num_funcs;
7039 MUTEX_EXIT(&rx_rpc_stats);
7044 * Regain the rx_peerHashTable_lock and
7045 * decrement the reference count on 'prev'
7048 MUTEX_ENTER(&rx_peerHashTable_lock);
7055 MUTEX_EXIT(&peer->peer_lock);
7060 MUTEX_EXIT(&rx_peerHashTable_lock);
7064 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7065 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7066 * GC, just below. Really, we shouldn't have to keep moving packets from
7067 * one place to another, but instead ought to always know if we can
7068 * afford to hold onto a packet in its particular use. */
7069 MUTEX_ENTER(&rx_freePktQ_lock);
7070 if (rx_waitingForPackets) {
7071 rx_waitingForPackets = 0;
7072 #ifdef RX_ENABLE_LOCKS
7073 CV_BROADCAST(&rx_waitingForPackets_cv);
7075 osi_rxWakeup(&rx_waitingForPackets);
7078 MUTEX_EXIT(&rx_freePktQ_lock);
7081 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7082 rxevent_Put(rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0));
7086 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7087 * rx.h is sort of strange this is better. This is called with a security
7088 * object before it is discarded. Each connection using a security object has
7089 * its own refcount to the object so it won't actually be freed until the last
7090 * connection is destroyed.
7092 * This is the only rxs module call. A hold could also be written but no one
7096 rxs_Release(struct rx_securityClass *aobj)
7098 return RXS_Close(aobj);
7106 #define TRACE_OPTION_RX_DEBUG 16
7114 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7115 0, KEY_QUERY_VALUE, &parmKey);
7116 if (code != ERROR_SUCCESS)
7119 dummyLen = sizeof(TraceOption);
7120 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7121 (BYTE *) &TraceOption, &dummyLen);
7122 if (code == ERROR_SUCCESS) {
7123 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7125 RegCloseKey (parmKey);
7126 #endif /* AFS_NT40_ENV */
7131 rx_DebugOnOff(int on)
7135 rxdebug_active = on;
7141 rx_StatsOnOff(int on)
7143 rx_stats_active = on;
7147 /* Don't call this debugging routine directly; use dpf */
7149 rxi_DebugPrint(char *format, ...)
7158 va_start(ap, format);
7160 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7163 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7165 OutputDebugString(msg);
7171 va_start(ap, format);
7173 clock_GetTime(&now);
7174 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7175 (unsigned int)now.usec);
7176 vfprintf(rx_Log, format, ap);
7184 * This function is used to process the rx_stats structure that is local
7185 * to a process as well as an rx_stats structure received from a remote
7186 * process (via rxdebug). Therefore, it needs to do minimal version
7190 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7191 afs_int32 freePackets, char version)
7195 if (size != sizeof(struct rx_statistics)) {
7197 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7198 size, sizeof(struct rx_statistics));
7201 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7204 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7205 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7206 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7207 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7208 s->specialPktAllocFailures);
7210 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7211 s->receivePktAllocFailures, s->sendPktAllocFailures,
7212 s->specialPktAllocFailures);
7216 " greedy %u, " "bogusReads %u (last from host %x), "
7217 "noPackets %u, " "noBuffers %u, " "selects %u, "
7218 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7219 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7220 s->selects, s->sendSelects);
7222 fprintf(file, " packets read: ");
7223 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7224 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7226 fprintf(file, "\n");
7229 " other read counters: data %u, " "ack %u, " "dup %u "
7230 "spurious %u " "dally %u\n", s->dataPacketsRead,
7231 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7232 s->ignorePacketDally);
7234 fprintf(file, " packets sent: ");
7235 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7236 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7238 fprintf(file, "\n");
7241 " other send counters: ack %u, " "data %u (not resends), "
7242 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7243 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7244 s->dataPacketsPushed, s->ignoreAckedPacket);
7247 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7248 s->netSendFailures, (int)s->fatalErrors);
7250 if (s->nRttSamples) {
7251 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7252 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7254 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7255 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7259 " %d server connections, " "%d client connections, "
7260 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7261 s->nServerConns, s->nClientConns, s->nPeerStructs,
7262 s->nCallStructs, s->nFreeCallStructs);
7264 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7265 fprintf(file, " %d clock updates\n", clock_nUpdates);
7269 /* for backward compatibility */
7271 rx_PrintStats(FILE * file)
7273 MUTEX_ENTER(&rx_stats_mutex);
7274 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7275 sizeof(rx_stats), rx_nFreePackets,
7277 MUTEX_EXIT(&rx_stats_mutex);
7281 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7283 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
7284 ntohl(peer->host), (int)ntohs(peer->port), (int)peer->burstSize,
7285 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
7288 " Rtt %d, " "total sent %d, " "resent %d\n",
7289 peer->rtt, peer->nSent, peer->reSends);
7292 " Packet size %d, " "max in packet skew %d, "
7293 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
7294 (int)peer->outPacketSkew);
7298 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7300 * This mutex protects the following static variables:
7304 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7305 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7307 #define LOCK_RX_DEBUG
7308 #define UNLOCK_RX_DEBUG
7309 #endif /* AFS_PTHREAD_ENV */
7311 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7313 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7314 u_char type, void *inputData, size_t inputLength,
7315 void *outputData, size_t outputLength)
7317 static afs_int32 counter = 100;
7318 time_t waitTime, waitCount;
7319 struct rx_header theader;
7322 struct timeval tv_now, tv_wake, tv_delta;
7323 struct sockaddr_in taddr, faddr;
7337 tp = &tbuffer[sizeof(struct rx_header)];
7338 taddr.sin_family = AF_INET;
7339 taddr.sin_port = remotePort;
7340 taddr.sin_addr.s_addr = remoteAddr;
7341 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7342 taddr.sin_len = sizeof(struct sockaddr_in);
7345 memset(&theader, 0, sizeof(theader));
7346 theader.epoch = htonl(999);
7348 theader.callNumber = htonl(counter);
7351 theader.type = type;
7352 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7353 theader.serviceId = 0;
7355 memcpy(tbuffer, &theader, sizeof(theader));
7356 memcpy(tp, inputData, inputLength);
7358 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7359 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7361 /* see if there's a packet available */
7362 gettimeofday(&tv_wake, NULL);
7363 tv_wake.tv_sec += waitTime;
7366 FD_SET(socket, &imask);
7367 tv_delta.tv_sec = tv_wake.tv_sec;
7368 tv_delta.tv_usec = tv_wake.tv_usec;
7369 gettimeofday(&tv_now, NULL);
7371 if (tv_delta.tv_usec < tv_now.tv_usec) {
7373 tv_delta.tv_usec += 1000000;
7376 tv_delta.tv_usec -= tv_now.tv_usec;
7378 if (tv_delta.tv_sec < tv_now.tv_sec) {
7382 tv_delta.tv_sec -= tv_now.tv_sec;
7385 code = select(0, &imask, 0, 0, &tv_delta);
7386 #else /* AFS_NT40_ENV */
7387 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7388 #endif /* AFS_NT40_ENV */
7389 if (code == 1 && FD_ISSET(socket, &imask)) {
7390 /* now receive a packet */
7391 faddrLen = sizeof(struct sockaddr_in);
7393 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7394 (struct sockaddr *)&faddr, &faddrLen);
7397 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7398 if (counter == ntohl(theader.callNumber))
7406 /* see if we've timed out */
7414 code -= sizeof(struct rx_header);
7415 if (code > outputLength)
7416 code = outputLength;
7417 memcpy(outputData, tp, code);
7420 #endif /* RXDEBUG */
7423 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7424 afs_uint16 remotePort, struct rx_debugStats * stat,
7425 afs_uint32 * supportedValues)
7427 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7429 struct rx_debugIn in;
7431 *supportedValues = 0;
7432 in.type = htonl(RX_DEBUGI_GETSTATS);
7435 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7436 &in, sizeof(in), stat, sizeof(*stat));
7439 * If the call was successful, fixup the version and indicate
7440 * what contents of the stat structure are valid.
7441 * Also do net to host conversion of fields here.
7445 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7446 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7448 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7449 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7451 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7452 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7454 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7455 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7457 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7458 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7460 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7461 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7463 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7464 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7466 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7467 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7469 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7470 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7472 stat->nFreePackets = ntohl(stat->nFreePackets);
7473 stat->packetReclaims = ntohl(stat->packetReclaims);
7474 stat->callsExecuted = ntohl(stat->callsExecuted);
7475 stat->nWaiting = ntohl(stat->nWaiting);
7476 stat->idleThreads = ntohl(stat->idleThreads);
7477 stat->nWaited = ntohl(stat->nWaited);
7478 stat->nPackets = ntohl(stat->nPackets);
7487 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7488 afs_uint16 remotePort, struct rx_statistics * stat,
7489 afs_uint32 * supportedValues)
7491 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7493 struct rx_debugIn in;
7494 afs_int32 *lp = (afs_int32 *) stat;
7498 * supportedValues is currently unused, but added to allow future
7499 * versioning of this function.
7502 *supportedValues = 0;
7503 in.type = htonl(RX_DEBUGI_RXSTATS);
7505 memset(stat, 0, sizeof(*stat));
7507 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7508 &in, sizeof(in), stat, sizeof(*stat));
7513 * Do net to host conversion here
7516 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7527 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7528 afs_uint16 remotePort, size_t version_length,
7531 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7533 return MakeDebugCall(socket, remoteAddr, remotePort,
7534 RX_PACKET_TYPE_VERSION, a, 1, version,
7542 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7543 afs_uint16 remotePort, afs_int32 * nextConnection,
7544 int allConnections, afs_uint32 debugSupportedValues,
7545 struct rx_debugConn * conn,
7546 afs_uint32 * supportedValues)
7548 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7550 struct rx_debugIn in;
7554 * supportedValues is currently unused, but added to allow future
7555 * versioning of this function.
7558 *supportedValues = 0;
7559 if (allConnections) {
7560 in.type = htonl(RX_DEBUGI_GETALLCONN);
7562 in.type = htonl(RX_DEBUGI_GETCONN);
7564 in.index = htonl(*nextConnection);
7565 memset(conn, 0, sizeof(*conn));
7567 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7568 &in, sizeof(in), conn, sizeof(*conn));
7571 *nextConnection += 1;
7574 * Convert old connection format to new structure.
7577 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7578 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7579 #define MOVEvL(a) (conn->a = vL->a)
7581 /* any old or unrecognized version... */
7582 for (i = 0; i < RX_MAXCALLS; i++) {
7583 MOVEvL(callState[i]);
7584 MOVEvL(callMode[i]);
7585 MOVEvL(callFlags[i]);
7586 MOVEvL(callOther[i]);
7588 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7589 MOVEvL(secStats.type);
7590 MOVEvL(secStats.level);
7591 MOVEvL(secStats.flags);
7592 MOVEvL(secStats.expires);
7593 MOVEvL(secStats.packetsReceived);
7594 MOVEvL(secStats.packetsSent);
7595 MOVEvL(secStats.bytesReceived);
7596 MOVEvL(secStats.bytesSent);
7601 * Do net to host conversion here
7603 * I don't convert host or port since we are most likely
7604 * going to want these in NBO.
7606 conn->cid = ntohl(conn->cid);
7607 conn->serial = ntohl(conn->serial);
7608 for (i = 0; i < RX_MAXCALLS; i++) {
7609 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7611 conn->error = ntohl(conn->error);
7612 conn->secStats.flags = ntohl(conn->secStats.flags);
7613 conn->secStats.expires = ntohl(conn->secStats.expires);
7614 conn->secStats.packetsReceived =
7615 ntohl(conn->secStats.packetsReceived);
7616 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7617 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7618 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7619 conn->epoch = ntohl(conn->epoch);
7620 conn->natMTU = ntohl(conn->natMTU);
7629 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7630 afs_uint16 remotePort, afs_int32 * nextPeer,
7631 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7632 afs_uint32 * supportedValues)
7634 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7636 struct rx_debugIn in;
7639 * supportedValues is currently unused, but added to allow future
7640 * versioning of this function.
7643 *supportedValues = 0;
7644 in.type = htonl(RX_DEBUGI_GETPEER);
7645 in.index = htonl(*nextPeer);
7646 memset(peer, 0, sizeof(*peer));
7648 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7649 &in, sizeof(in), peer, sizeof(*peer));
7655 * Do net to host conversion here
7657 * I don't convert host or port since we are most likely
7658 * going to want these in NBO.
7660 peer->ifMTU = ntohs(peer->ifMTU);
7661 peer->idleWhen = ntohl(peer->idleWhen);
7662 peer->refCount = ntohs(peer->refCount);
7663 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7664 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7665 peer->rtt = ntohl(peer->rtt);
7666 peer->rtt_dev = ntohl(peer->rtt_dev);
7667 peer->timeout.sec = 0;
7668 peer->timeout.usec = 0;
7669 peer->nSent = ntohl(peer->nSent);
7670 peer->reSends = ntohl(peer->reSends);
7671 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7672 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7673 peer->natMTU = ntohs(peer->natMTU);
7674 peer->maxMTU = ntohs(peer->maxMTU);
7675 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7676 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7677 peer->MTU = ntohs(peer->MTU);
7678 peer->cwind = ntohs(peer->cwind);
7679 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7680 peer->congestSeq = ntohs(peer->congestSeq);
7681 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7682 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7683 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7684 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7693 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7694 struct rx_debugPeer * peerStats)
7697 afs_int32 error = 1; /* default to "did not succeed" */
7698 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7700 MUTEX_ENTER(&rx_peerHashTable_lock);
7701 for(tp = rx_peerHashTable[hashValue];
7702 tp != NULL; tp = tp->next) {
7703 if (tp->host == peerHost)
7709 MUTEX_EXIT(&rx_peerHashTable_lock);
7713 MUTEX_ENTER(&tp->peer_lock);
7714 peerStats->host = tp->host;
7715 peerStats->port = tp->port;
7716 peerStats->ifMTU = tp->ifMTU;
7717 peerStats->idleWhen = tp->idleWhen;
7718 peerStats->refCount = tp->refCount;
7719 peerStats->burstSize = tp->burstSize;
7720 peerStats->burst = tp->burst;
7721 peerStats->burstWait.sec = tp->burstWait.sec;
7722 peerStats->burstWait.usec = tp->burstWait.usec;
7723 peerStats->rtt = tp->rtt;
7724 peerStats->rtt_dev = tp->rtt_dev;
7725 peerStats->timeout.sec = 0;
7726 peerStats->timeout.usec = 0;
7727 peerStats->nSent = tp->nSent;
7728 peerStats->reSends = tp->reSends;
7729 peerStats->inPacketSkew = tp->inPacketSkew;
7730 peerStats->outPacketSkew = tp->outPacketSkew;
7731 peerStats->natMTU = tp->natMTU;
7732 peerStats->maxMTU = tp->maxMTU;
7733 peerStats->maxDgramPackets = tp->maxDgramPackets;
7734 peerStats->ifDgramPackets = tp->ifDgramPackets;
7735 peerStats->MTU = tp->MTU;
7736 peerStats->cwind = tp->cwind;
7737 peerStats->nDgramPackets = tp->nDgramPackets;
7738 peerStats->congestSeq = tp->congestSeq;
7739 peerStats->bytesSent.high = tp->bytesSent.high;
7740 peerStats->bytesSent.low = tp->bytesSent.low;
7741 peerStats->bytesReceived.high = tp->bytesReceived.high;
7742 peerStats->bytesReceived.low = tp->bytesReceived.low;
7743 MUTEX_EXIT(&tp->peer_lock);
7745 MUTEX_ENTER(&rx_peerHashTable_lock);
7748 MUTEX_EXIT(&rx_peerHashTable_lock);
7756 struct rx_serverQueueEntry *np;
7759 struct rx_call *call;
7760 struct rx_serverQueueEntry *sq;
7764 if (rxinit_status == 1) {
7766 return; /* Already shutdown. */
7770 #ifndef AFS_PTHREAD_ENV
7771 FD_ZERO(&rx_selectMask);
7772 #endif /* AFS_PTHREAD_ENV */
7773 rxi_dataQuota = RX_MAX_QUOTA;
7774 #ifndef AFS_PTHREAD_ENV
7776 #endif /* AFS_PTHREAD_ENV */
7779 #ifndef AFS_PTHREAD_ENV
7780 #ifndef AFS_USE_GETTIMEOFDAY
7782 #endif /* AFS_USE_GETTIMEOFDAY */
7783 #endif /* AFS_PTHREAD_ENV */
7785 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7786 call = queue_First(&rx_freeCallQueue, rx_call);
7788 rxi_Free(call, sizeof(struct rx_call));
7791 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7792 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7798 struct rx_peer **peer_ptr, **peer_end;
7799 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7800 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7802 struct rx_peer *peer, *next;
7804 MUTEX_ENTER(&rx_peerHashTable_lock);
7805 for (peer = *peer_ptr; peer; peer = next) {
7806 rx_interface_stat_p rpc_stat, nrpc_stat;
7809 MUTEX_ENTER(&rx_rpc_stats);
7810 MUTEX_ENTER(&peer->peer_lock);
7812 (&peer->rpcStats, rpc_stat, nrpc_stat,
7813 rx_interface_stat)) {
7814 unsigned int num_funcs;
7817 queue_Remove(&rpc_stat->queue_header);
7818 queue_Remove(&rpc_stat->all_peers);
7819 num_funcs = rpc_stat->stats[0].func_total;
7821 sizeof(rx_interface_stat_t) +
7822 rpc_stat->stats[0].func_total *
7823 sizeof(rx_function_entry_v1_t);
7825 rxi_Free(rpc_stat, space);
7827 /* rx_rpc_stats must be held */
7828 rxi_rpc_peer_stat_cnt -= num_funcs;
7830 MUTEX_EXIT(&peer->peer_lock);
7831 MUTEX_EXIT(&rx_rpc_stats);
7835 if (rx_stats_active)
7836 rx_atomic_dec(&rx_stats.nPeerStructs);
7838 MUTEX_EXIT(&rx_peerHashTable_lock);
7841 for (i = 0; i < RX_MAX_SERVICES; i++) {
7843 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7845 for (i = 0; i < rx_hashTableSize; i++) {
7846 struct rx_connection *tc, *ntc;
7847 MUTEX_ENTER(&rx_connHashTable_lock);
7848 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7850 for (j = 0; j < RX_MAXCALLS; j++) {
7852 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7855 rxi_Free(tc, sizeof(*tc));
7857 MUTEX_EXIT(&rx_connHashTable_lock);
7860 MUTEX_ENTER(&freeSQEList_lock);
7862 while ((np = rx_FreeSQEList)) {
7863 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7864 MUTEX_DESTROY(&np->lock);
7865 rxi_Free(np, sizeof(*np));
7868 MUTEX_EXIT(&freeSQEList_lock);
7869 MUTEX_DESTROY(&freeSQEList_lock);
7870 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7871 MUTEX_DESTROY(&rx_connHashTable_lock);
7872 MUTEX_DESTROY(&rx_peerHashTable_lock);
7873 MUTEX_DESTROY(&rx_serverPool_lock);
7875 osi_Free(rx_connHashTable,
7876 rx_hashTableSize * sizeof(struct rx_connection *));
7877 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7879 UNPIN(rx_connHashTable,
7880 rx_hashTableSize * sizeof(struct rx_connection *));
7881 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7883 rxi_FreeAllPackets();
7885 MUTEX_ENTER(&rx_quota_mutex);
7886 rxi_dataQuota = RX_MAX_QUOTA;
7887 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7888 MUTEX_EXIT(&rx_quota_mutex);
7893 #ifdef RX_ENABLE_LOCKS
7895 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7897 if (!MUTEX_ISMINE(lockaddr))
7898 osi_Panic("Lock not held: %s", msg);
7900 #endif /* RX_ENABLE_LOCKS */
7905 * Routines to implement connection specific data.
7909 rx_KeyCreate(rx_destructor_t rtn)
7912 MUTEX_ENTER(&rxi_keyCreate_lock);
7913 key = rxi_keyCreate_counter++;
7914 rxi_keyCreate_destructor = (rx_destructor_t *)
7915 realloc((void *)rxi_keyCreate_destructor,
7916 (key + 1) * sizeof(rx_destructor_t));
7917 rxi_keyCreate_destructor[key] = rtn;
7918 MUTEX_EXIT(&rxi_keyCreate_lock);
7923 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7926 MUTEX_ENTER(&conn->conn_data_lock);
7927 if (!conn->specific) {
7928 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7929 for (i = 0; i < key; i++)
7930 conn->specific[i] = NULL;
7931 conn->nSpecific = key + 1;
7932 conn->specific[key] = ptr;
7933 } else if (key >= conn->nSpecific) {
7934 conn->specific = (void **)
7935 realloc(conn->specific, (key + 1) * sizeof(void *));
7936 for (i = conn->nSpecific; i < key; i++)
7937 conn->specific[i] = NULL;
7938 conn->nSpecific = key + 1;
7939 conn->specific[key] = ptr;
7941 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7942 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7943 conn->specific[key] = ptr;
7945 MUTEX_EXIT(&conn->conn_data_lock);
7949 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7952 MUTEX_ENTER(&svc->svc_data_lock);
7953 if (!svc->specific) {
7954 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
7955 for (i = 0; i < key; i++)
7956 svc->specific[i] = NULL;
7957 svc->nSpecific = key + 1;
7958 svc->specific[key] = ptr;
7959 } else if (key >= svc->nSpecific) {
7960 svc->specific = (void **)
7961 realloc(svc->specific, (key + 1) * sizeof(void *));
7962 for (i = svc->nSpecific; i < key; i++)
7963 svc->specific[i] = NULL;
7964 svc->nSpecific = key + 1;
7965 svc->specific[key] = ptr;
7967 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7968 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7969 svc->specific[key] = ptr;
7971 MUTEX_EXIT(&svc->svc_data_lock);
7975 rx_GetSpecific(struct rx_connection *conn, int key)
7978 MUTEX_ENTER(&conn->conn_data_lock);
7979 if (key >= conn->nSpecific)
7982 ptr = conn->specific[key];
7983 MUTEX_EXIT(&conn->conn_data_lock);
7988 rx_GetServiceSpecific(struct rx_service *svc, int key)
7991 MUTEX_ENTER(&svc->svc_data_lock);
7992 if (key >= svc->nSpecific)
7995 ptr = svc->specific[key];
7996 MUTEX_EXIT(&svc->svc_data_lock);
8001 #endif /* !KERNEL */
8004 * processStats is a queue used to store the statistics for the local
8005 * process. Its contents are similar to the contents of the rpcStats
8006 * queue on a rx_peer structure, but the actual data stored within
8007 * this queue contains totals across the lifetime of the process (assuming
8008 * the stats have not been reset) - unlike the per peer structures
8009 * which can come and go based upon the peer lifetime.
8012 static struct rx_queue processStats = { &processStats, &processStats };
8015 * peerStats is a queue used to store the statistics for all peer structs.
8016 * Its contents are the union of all the peer rpcStats queues.
8019 static struct rx_queue peerStats = { &peerStats, &peerStats };
8022 * rxi_monitor_processStats is used to turn process wide stat collection
8026 static int rxi_monitor_processStats = 0;
8029 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8032 static int rxi_monitor_peerStats = 0;
8035 * rxi_AddRpcStat - given all of the information for a particular rpc
8036 * call, create (if needed) and update the stat totals for the rpc.
8040 * IN stats - the queue of stats that will be updated with the new value
8042 * IN rxInterface - a unique number that identifies the rpc interface
8044 * IN currentFunc - the index of the function being invoked
8046 * IN totalFunc - the total number of functions in this interface
8048 * IN queueTime - the amount of time this function waited for a thread
8050 * IN execTime - the amount of time this function invocation took to execute
8052 * IN bytesSent - the number bytes sent by this invocation
8054 * IN bytesRcvd - the number bytes received by this invocation
8056 * IN isServer - if true, this invocation was made to a server
8058 * IN remoteHost - the ip address of the remote host
8060 * IN remotePort - the port of the remote host
8062 * IN addToPeerList - if != 0, add newly created stat to the global peer list
8064 * INOUT counter - if a new stats structure is allocated, the counter will
8065 * be updated with the new number of allocated stat structures
8073 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
8074 afs_uint32 currentFunc, afs_uint32 totalFunc,
8075 struct clock *queueTime, struct clock *execTime,
8076 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
8077 afs_uint32 remoteHost, afs_uint32 remotePort,
8078 int addToPeerList, unsigned int *counter)
8081 rx_interface_stat_p rpc_stat, nrpc_stat;
8084 * See if there's already a structure for this interface
8087 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8088 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8089 && (rpc_stat->stats[0].remote_is_server == isServer))
8094 * Didn't find a match so allocate a new structure and add it to the
8098 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
8099 || (rpc_stat->stats[0].interfaceId != rxInterface)
8100 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8105 sizeof(rx_interface_stat_t) +
8106 totalFunc * sizeof(rx_function_entry_v1_t);
8108 rpc_stat = rxi_Alloc(space);
8109 if (rpc_stat == NULL) {
8113 *counter += totalFunc;
8114 for (i = 0; i < totalFunc; i++) {
8115 rpc_stat->stats[i].remote_peer = remoteHost;
8116 rpc_stat->stats[i].remote_port = remotePort;
8117 rpc_stat->stats[i].remote_is_server = isServer;
8118 rpc_stat->stats[i].interfaceId = rxInterface;
8119 rpc_stat->stats[i].func_total = totalFunc;
8120 rpc_stat->stats[i].func_index = i;
8121 hzero(rpc_stat->stats[i].invocations);
8122 hzero(rpc_stat->stats[i].bytes_sent);
8123 hzero(rpc_stat->stats[i].bytes_rcvd);
8124 rpc_stat->stats[i].queue_time_sum.sec = 0;
8125 rpc_stat->stats[i].queue_time_sum.usec = 0;
8126 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8127 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8128 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8129 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8130 rpc_stat->stats[i].queue_time_max.sec = 0;
8131 rpc_stat->stats[i].queue_time_max.usec = 0;
8132 rpc_stat->stats[i].execution_time_sum.sec = 0;
8133 rpc_stat->stats[i].execution_time_sum.usec = 0;
8134 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8135 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8136 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8137 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8138 rpc_stat->stats[i].execution_time_max.sec = 0;
8139 rpc_stat->stats[i].execution_time_max.usec = 0;
8141 queue_Prepend(stats, rpc_stat);
8142 if (addToPeerList) {
8143 queue_Prepend(&peerStats, &rpc_stat->all_peers);
8148 * Increment the stats for this function
8151 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
8152 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
8153 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
8154 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8155 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8156 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8157 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8159 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8160 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8162 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8163 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8165 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8166 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8168 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8169 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8177 * rx_IncrementTimeAndCount - increment the times and count for a particular
8182 * IN peer - the peer who invoked the rpc
8184 * IN rxInterface - a unique number that identifies the rpc interface
8186 * IN currentFunc - the index of the function being invoked
8188 * IN totalFunc - the total number of functions in this interface
8190 * IN queueTime - the amount of time this function waited for a thread
8192 * IN execTime - the amount of time this function invocation took to execute
8194 * IN bytesSent - the number bytes sent by this invocation
8196 * IN bytesRcvd - the number bytes received by this invocation
8198 * IN isServer - if true, this invocation was made to a server
8206 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8207 afs_uint32 currentFunc, afs_uint32 totalFunc,
8208 struct clock *queueTime, struct clock *execTime,
8209 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8213 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8216 MUTEX_ENTER(&rx_rpc_stats);
8218 if (rxi_monitor_peerStats) {
8219 MUTEX_ENTER(&peer->peer_lock);
8220 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8221 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8222 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8223 MUTEX_EXIT(&peer->peer_lock);
8226 if (rxi_monitor_processStats) {
8227 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8228 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8229 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8232 MUTEX_EXIT(&rx_rpc_stats);
8237 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8241 * IN callerVersion - the rpc stat version of the caller.
8243 * IN count - the number of entries to marshall.
8245 * IN stats - pointer to stats to be marshalled.
8247 * OUT ptr - Where to store the marshalled data.
8254 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8255 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8261 * We only support the first version
8263 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8264 *(ptr++) = stats->remote_peer;
8265 *(ptr++) = stats->remote_port;
8266 *(ptr++) = stats->remote_is_server;
8267 *(ptr++) = stats->interfaceId;
8268 *(ptr++) = stats->func_total;
8269 *(ptr++) = stats->func_index;
8270 *(ptr++) = hgethi(stats->invocations);
8271 *(ptr++) = hgetlo(stats->invocations);
8272 *(ptr++) = hgethi(stats->bytes_sent);
8273 *(ptr++) = hgetlo(stats->bytes_sent);
8274 *(ptr++) = hgethi(stats->bytes_rcvd);
8275 *(ptr++) = hgetlo(stats->bytes_rcvd);
8276 *(ptr++) = stats->queue_time_sum.sec;
8277 *(ptr++) = stats->queue_time_sum.usec;
8278 *(ptr++) = stats->queue_time_sum_sqr.sec;
8279 *(ptr++) = stats->queue_time_sum_sqr.usec;
8280 *(ptr++) = stats->queue_time_min.sec;
8281 *(ptr++) = stats->queue_time_min.usec;
8282 *(ptr++) = stats->queue_time_max.sec;
8283 *(ptr++) = stats->queue_time_max.usec;
8284 *(ptr++) = stats->execution_time_sum.sec;
8285 *(ptr++) = stats->execution_time_sum.usec;
8286 *(ptr++) = stats->execution_time_sum_sqr.sec;
8287 *(ptr++) = stats->execution_time_sum_sqr.usec;
8288 *(ptr++) = stats->execution_time_min.sec;
8289 *(ptr++) = stats->execution_time_min.usec;
8290 *(ptr++) = stats->execution_time_max.sec;
8291 *(ptr++) = stats->execution_time_max.usec;
8297 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8302 * IN callerVersion - the rpc stat version of the caller
8304 * OUT myVersion - the rpc stat version of this function
8306 * OUT clock_sec - local time seconds
8308 * OUT clock_usec - local time microseconds
8310 * OUT allocSize - the number of bytes allocated to contain stats
8312 * OUT statCount - the number stats retrieved from this process.
8314 * OUT stats - the actual stats retrieved from this process.
8318 * Returns void. If successful, stats will != NULL.
8322 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8323 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8324 size_t * allocSize, afs_uint32 * statCount,
8325 afs_uint32 ** stats)
8335 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8338 * Check to see if stats are enabled
8341 MUTEX_ENTER(&rx_rpc_stats);
8342 if (!rxi_monitor_processStats) {
8343 MUTEX_EXIT(&rx_rpc_stats);
8347 clock_GetTime(&now);
8348 *clock_sec = now.sec;
8349 *clock_usec = now.usec;
8352 * Allocate the space based upon the caller version
8354 * If the client is at an older version than we are,
8355 * we return the statistic data in the older data format, but
8356 * we still return our version number so the client knows we
8357 * are maintaining more data than it can retrieve.
8360 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8361 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8362 *statCount = rxi_rpc_process_stat_cnt;
8365 * This can't happen yet, but in the future version changes
8366 * can be handled by adding additional code here
8370 if (space > (size_t) 0) {
8372 ptr = *stats = rxi_Alloc(space);
8375 rx_interface_stat_p rpc_stat, nrpc_stat;
8379 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8381 * Copy the data based upon the caller version
8383 rx_MarshallProcessRPCStats(callerVersion,
8384 rpc_stat->stats[0].func_total,
8385 rpc_stat->stats, &ptr);
8391 MUTEX_EXIT(&rx_rpc_stats);
8396 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8400 * IN callerVersion - the rpc stat version of the caller
8402 * OUT myVersion - the rpc stat version of this function
8404 * OUT clock_sec - local time seconds
8406 * OUT clock_usec - local time microseconds
8408 * OUT allocSize - the number of bytes allocated to contain stats
8410 * OUT statCount - the number of stats retrieved from the individual
8413 * OUT stats - the actual stats retrieved from the individual peer structures.
8417 * Returns void. If successful, stats will != NULL.
8421 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8422 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8423 size_t * allocSize, afs_uint32 * statCount,
8424 afs_uint32 ** stats)
8434 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8437 * Check to see if stats are enabled
8440 MUTEX_ENTER(&rx_rpc_stats);
8441 if (!rxi_monitor_peerStats) {
8442 MUTEX_EXIT(&rx_rpc_stats);
8446 clock_GetTime(&now);
8447 *clock_sec = now.sec;
8448 *clock_usec = now.usec;
8451 * Allocate the space based upon the caller version
8453 * If the client is at an older version than we are,
8454 * we return the statistic data in the older data format, but
8455 * we still return our version number so the client knows we
8456 * are maintaining more data than it can retrieve.
8459 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8460 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8461 *statCount = rxi_rpc_peer_stat_cnt;
8464 * This can't happen yet, but in the future version changes
8465 * can be handled by adding additional code here
8469 if (space > (size_t) 0) {
8471 ptr = *stats = rxi_Alloc(space);
8474 rx_interface_stat_p rpc_stat, nrpc_stat;
8478 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8480 * We have to fix the offset of rpc_stat since we are
8481 * keeping this structure on two rx_queues. The rx_queue
8482 * package assumes that the rx_queue member is the first
8483 * member of the structure. That is, rx_queue assumes that
8484 * any one item is only on one queue at a time. We are
8485 * breaking that assumption and so we have to do a little
8486 * math to fix our pointers.
8489 fix_offset = (char *)rpc_stat;
8490 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8491 rpc_stat = (rx_interface_stat_p) fix_offset;
8494 * Copy the data based upon the caller version
8496 rx_MarshallProcessRPCStats(callerVersion,
8497 rpc_stat->stats[0].func_total,
8498 rpc_stat->stats, &ptr);
8504 MUTEX_EXIT(&rx_rpc_stats);
8509 * rx_FreeRPCStats - free memory allocated by
8510 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8514 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8515 * rx_RetrievePeerRPCStats
8517 * IN allocSize - the number of bytes in stats.
8525 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8527 rxi_Free(stats, allocSize);
8531 * rx_queryProcessRPCStats - see if process rpc stat collection is
8532 * currently enabled.
8538 * Returns 0 if stats are not enabled != 0 otherwise
8542 rx_queryProcessRPCStats(void)
8545 MUTEX_ENTER(&rx_rpc_stats);
8546 rc = rxi_monitor_processStats;
8547 MUTEX_EXIT(&rx_rpc_stats);
8552 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8558 * Returns 0 if stats are not enabled != 0 otherwise
8562 rx_queryPeerRPCStats(void)
8565 MUTEX_ENTER(&rx_rpc_stats);
8566 rc = rxi_monitor_peerStats;
8567 MUTEX_EXIT(&rx_rpc_stats);
8572 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8582 rx_enableProcessRPCStats(void)
8584 MUTEX_ENTER(&rx_rpc_stats);
8585 rx_enable_stats = 1;
8586 rxi_monitor_processStats = 1;
8587 MUTEX_EXIT(&rx_rpc_stats);
8591 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8601 rx_enablePeerRPCStats(void)
8603 MUTEX_ENTER(&rx_rpc_stats);
8604 rx_enable_stats = 1;
8605 rxi_monitor_peerStats = 1;
8606 MUTEX_EXIT(&rx_rpc_stats);
8610 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8620 rx_disableProcessRPCStats(void)
8622 rx_interface_stat_p rpc_stat, nrpc_stat;
8625 MUTEX_ENTER(&rx_rpc_stats);
8628 * Turn off process statistics and if peer stats is also off, turn
8632 rxi_monitor_processStats = 0;
8633 if (rxi_monitor_peerStats == 0) {
8634 rx_enable_stats = 0;
8637 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8638 unsigned int num_funcs = 0;
8641 queue_Remove(rpc_stat);
8642 num_funcs = rpc_stat->stats[0].func_total;
8644 sizeof(rx_interface_stat_t) +
8645 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8647 rxi_Free(rpc_stat, space);
8648 rxi_rpc_process_stat_cnt -= num_funcs;
8650 MUTEX_EXIT(&rx_rpc_stats);
8654 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8664 rx_disablePeerRPCStats(void)
8666 struct rx_peer **peer_ptr, **peer_end;
8670 * Turn off peer statistics and if process stats is also off, turn
8674 rxi_monitor_peerStats = 0;
8675 if (rxi_monitor_processStats == 0) {
8676 rx_enable_stats = 0;
8679 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8680 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8682 struct rx_peer *peer, *next, *prev;
8684 MUTEX_ENTER(&rx_peerHashTable_lock);
8685 MUTEX_ENTER(&rx_rpc_stats);
8686 for (prev = peer = *peer_ptr; peer; peer = next) {
8688 code = MUTEX_TRYENTER(&peer->peer_lock);
8690 rx_interface_stat_p rpc_stat, nrpc_stat;
8693 if (prev == *peer_ptr) {
8704 MUTEX_EXIT(&rx_peerHashTable_lock);
8707 (&peer->rpcStats, rpc_stat, nrpc_stat,
8708 rx_interface_stat)) {
8709 unsigned int num_funcs = 0;
8712 queue_Remove(&rpc_stat->queue_header);
8713 queue_Remove(&rpc_stat->all_peers);
8714 num_funcs = rpc_stat->stats[0].func_total;
8716 sizeof(rx_interface_stat_t) +
8717 rpc_stat->stats[0].func_total *
8718 sizeof(rx_function_entry_v1_t);
8720 rxi_Free(rpc_stat, space);
8721 rxi_rpc_peer_stat_cnt -= num_funcs;
8723 MUTEX_EXIT(&peer->peer_lock);
8725 MUTEX_ENTER(&rx_peerHashTable_lock);
8735 MUTEX_EXIT(&rx_rpc_stats);
8736 MUTEX_EXIT(&rx_peerHashTable_lock);
8741 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8746 * IN clearFlag - flag indicating which stats to clear
8754 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8756 rx_interface_stat_p rpc_stat, nrpc_stat;
8758 MUTEX_ENTER(&rx_rpc_stats);
8760 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8761 unsigned int num_funcs = 0, i;
8762 num_funcs = rpc_stat->stats[0].func_total;
8763 for (i = 0; i < num_funcs; i++) {
8764 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8765 hzero(rpc_stat->stats[i].invocations);
8767 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8768 hzero(rpc_stat->stats[i].bytes_sent);
8770 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8771 hzero(rpc_stat->stats[i].bytes_rcvd);
8773 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8774 rpc_stat->stats[i].queue_time_sum.sec = 0;
8775 rpc_stat->stats[i].queue_time_sum.usec = 0;
8777 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8778 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8779 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8781 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8782 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8783 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8785 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8786 rpc_stat->stats[i].queue_time_max.sec = 0;
8787 rpc_stat->stats[i].queue_time_max.usec = 0;
8789 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8790 rpc_stat->stats[i].execution_time_sum.sec = 0;
8791 rpc_stat->stats[i].execution_time_sum.usec = 0;
8793 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8794 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8795 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8797 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8798 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8799 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8801 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8802 rpc_stat->stats[i].execution_time_max.sec = 0;
8803 rpc_stat->stats[i].execution_time_max.usec = 0;
8808 MUTEX_EXIT(&rx_rpc_stats);
8812 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8817 * IN clearFlag - flag indicating which stats to clear
8825 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8827 rx_interface_stat_p rpc_stat, nrpc_stat;
8829 MUTEX_ENTER(&rx_rpc_stats);
8831 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8832 unsigned int num_funcs = 0, i;
8835 * We have to fix the offset of rpc_stat since we are
8836 * keeping this structure on two rx_queues. The rx_queue
8837 * package assumes that the rx_queue member is the first
8838 * member of the structure. That is, rx_queue assumes that
8839 * any one item is only on one queue at a time. We are
8840 * breaking that assumption and so we have to do a little
8841 * math to fix our pointers.
8844 fix_offset = (char *)rpc_stat;
8845 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8846 rpc_stat = (rx_interface_stat_p) fix_offset;
8848 num_funcs = rpc_stat->stats[0].func_total;
8849 for (i = 0; i < num_funcs; i++) {
8850 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8851 hzero(rpc_stat->stats[i].invocations);
8853 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8854 hzero(rpc_stat->stats[i].bytes_sent);
8856 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8857 hzero(rpc_stat->stats[i].bytes_rcvd);
8859 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8860 rpc_stat->stats[i].queue_time_sum.sec = 0;
8861 rpc_stat->stats[i].queue_time_sum.usec = 0;
8863 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8864 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8865 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8867 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8868 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8869 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8871 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8872 rpc_stat->stats[i].queue_time_max.sec = 0;
8873 rpc_stat->stats[i].queue_time_max.usec = 0;
8875 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8876 rpc_stat->stats[i].execution_time_sum.sec = 0;
8877 rpc_stat->stats[i].execution_time_sum.usec = 0;
8879 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8880 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8881 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8883 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8884 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8885 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8887 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8888 rpc_stat->stats[i].execution_time_max.sec = 0;
8889 rpc_stat->stats[i].execution_time_max.usec = 0;
8894 MUTEX_EXIT(&rx_rpc_stats);
8898 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8899 * is authorized to enable/disable/clear RX statistics.
8901 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8904 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8906 rxi_rxstat_userok = proc;
8910 rx_RxStatUserOk(struct rx_call *call)
8912 if (!rxi_rxstat_userok)
8914 return rxi_rxstat_userok(call);
8919 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8920 * function in the MSVC runtime DLL (msvcrt.dll).
8922 * Note: the system serializes calls to this function.
8925 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8926 DWORD reason, /* reason function is being called */
8927 LPVOID reserved) /* reserved for future use */
8930 case DLL_PROCESS_ATTACH:
8931 /* library is being attached to a process */
8935 case DLL_PROCESS_DETACH:
8942 #endif /* AFS_NT40_ENV */
8945 int rx_DumpCalls(FILE *outputFile, char *cookie)
8947 #ifdef RXDEBUG_PACKET
8948 #ifdef KDUMP_RX_LOCK
8949 struct rx_call_rx_lock *c;
8956 #define RXDPRINTF sprintf
8957 #define RXDPRINTOUT output
8959 #define RXDPRINTF fprintf
8960 #define RXDPRINTOUT outputFile
8963 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8965 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8968 for (c = rx_allCallsp; c; c = c->allNextp) {
8969 u_short rqc, tqc, iovqc;
8970 struct rx_packet *p, *np;
8972 MUTEX_ENTER(&c->lock);
8973 queue_Count(&c->rq, p, np, rx_packet, rqc);
8974 queue_Count(&c->tq, p, np, rx_packet, tqc);
8975 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8977 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, "
8978 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8979 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8980 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8981 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8982 #ifdef RX_ENABLE_LOCKS
8985 #ifdef RX_REFCOUNT_CHECK
8986 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
8987 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
8990 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,
8991 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
8992 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
8993 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
8994 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
8995 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
8996 #ifdef RX_ENABLE_LOCKS
8997 , (afs_uint32)c->refCount
8999 #ifdef RX_REFCOUNT_CHECK
9000 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9003 MUTEX_EXIT(&c->lock);
9006 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9009 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9011 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9013 #endif /* RXDEBUG_PACKET */