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 (*call->callNumber)++;
1499 MUTEX_EXIT(&conn->conn_call_lock);
1500 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1501 rxi_ResetCall(call, 0);
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 MUTEX_ENTER(&aconn->conn_call_lock);
1665 for (i = 0; i < RX_MAXCALLS; i++) {
1666 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1667 aint32s[i] = aconn->callNumber[i] + 1;
1669 aint32s[i] = aconn->callNumber[i];
1671 MUTEX_EXIT(&aconn->conn_call_lock);
1677 rxi_SetCallNumberVector(struct rx_connection *aconn,
1678 afs_int32 * aint32s)
1681 struct rx_call *tcall;
1685 MUTEX_ENTER(&aconn->conn_call_lock);
1686 for (i = 0; i < RX_MAXCALLS; i++) {
1687 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1688 aconn->callNumber[i] = aint32s[i] - 1;
1690 aconn->callNumber[i] = aint32s[i];
1692 MUTEX_EXIT(&aconn->conn_call_lock);
1697 /* Advertise a new service. A service is named locally by a UDP port
1698 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1701 char *serviceName; Name for identification purposes (e.g. the
1702 service name might be used for probing for
1705 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1706 char *serviceName, struct rx_securityClass **securityObjects,
1707 int nSecurityObjects,
1708 afs_int32(*serviceProc) (struct rx_call * acall))
1710 osi_socket socket = OSI_NULLSOCKET;
1711 struct rx_service *tservice;
1717 if (serviceId == 0) {
1719 "rx_NewService: service id for service %s is not non-zero.\n",
1726 "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",
1734 tservice = rxi_AllocService();
1737 #ifdef RX_ENABLE_LOCKS
1738 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1741 for (i = 0; i < RX_MAX_SERVICES; i++) {
1742 struct rx_service *service = rx_services[i];
1744 if (port == service->servicePort && host == service->serviceHost) {
1745 if (service->serviceId == serviceId) {
1746 /* The identical service has already been
1747 * installed; if the caller was intending to
1748 * change the security classes used by this
1749 * service, he/she loses. */
1751 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1752 serviceName, serviceId, service->serviceName);
1754 rxi_FreeService(tservice);
1757 /* Different service, same port: re-use the socket
1758 * which is bound to the same port */
1759 socket = service->socket;
1762 if (socket == OSI_NULLSOCKET) {
1763 /* If we don't already have a socket (from another
1764 * service on same port) get a new one */
1765 socket = rxi_GetHostUDPSocket(host, port);
1766 if (socket == OSI_NULLSOCKET) {
1768 rxi_FreeService(tservice);
1773 service->socket = socket;
1774 service->serviceHost = host;
1775 service->servicePort = port;
1776 service->serviceId = serviceId;
1777 service->serviceName = serviceName;
1778 service->nSecurityObjects = nSecurityObjects;
1779 service->securityObjects = securityObjects;
1780 service->minProcs = 0;
1781 service->maxProcs = 1;
1782 service->idleDeadTime = 60;
1783 service->idleDeadErr = 0;
1784 service->connDeadTime = rx_connDeadTime;
1785 service->executeRequestProc = serviceProc;
1786 service->checkReach = 0;
1787 service->nSpecific = 0;
1788 service->specific = NULL;
1789 rx_services[i] = service; /* not visible until now */
1795 rxi_FreeService(tservice);
1796 (osi_Msg "rx_NewService: cannot support > %d services\n",
1801 /* Set configuration options for all of a service's security objects */
1804 rx_SetSecurityConfiguration(struct rx_service *service,
1805 rx_securityConfigVariables type,
1809 for (i = 0; i<service->nSecurityObjects; i++) {
1810 if (service->securityObjects[i]) {
1811 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1819 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1820 struct rx_securityClass **securityObjects, int nSecurityObjects,
1821 afs_int32(*serviceProc) (struct rx_call * acall))
1823 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1826 /* Generic request processing loop. This routine should be called
1827 * by the implementation dependent rx_ServerProc. If socketp is
1828 * non-null, it will be set to the file descriptor that this thread
1829 * is now listening on. If socketp is null, this routine will never
1832 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1834 struct rx_call *call;
1836 struct rx_service *tservice = NULL;
1843 call = rx_GetCall(threadID, tservice, socketp);
1844 if (socketp && *socketp != OSI_NULLSOCKET) {
1845 /* We are now a listener thread */
1851 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1852 #ifdef RX_ENABLE_LOCKS
1854 #endif /* RX_ENABLE_LOCKS */
1855 afs_termState = AFSOP_STOP_AFS;
1856 afs_osi_Wakeup(&afs_termState);
1857 #ifdef RX_ENABLE_LOCKS
1859 #endif /* RX_ENABLE_LOCKS */
1864 /* if server is restarting( typically smooth shutdown) then do not
1865 * allow any new calls.
1868 if (rx_tranquil && (call != NULL)) {
1872 MUTEX_ENTER(&call->lock);
1874 rxi_CallError(call, RX_RESTARTING);
1875 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1877 MUTEX_EXIT(&call->lock);
1882 tservice = call->conn->service;
1884 if (tservice->beforeProc)
1885 (*tservice->beforeProc) (call);
1887 code = tservice->executeRequestProc(call);
1889 if (tservice->afterProc)
1890 (*tservice->afterProc) (call, code);
1892 rx_EndCall(call, code);
1894 if (tservice->postProc)
1895 (*tservice->postProc) (code);
1897 if (rx_stats_active) {
1898 MUTEX_ENTER(&rx_stats_mutex);
1900 MUTEX_EXIT(&rx_stats_mutex);
1907 rx_WakeupServerProcs(void)
1909 struct rx_serverQueueEntry *np, *tqp;
1913 MUTEX_ENTER(&rx_serverPool_lock);
1915 #ifdef RX_ENABLE_LOCKS
1916 if (rx_waitForPacket)
1917 CV_BROADCAST(&rx_waitForPacket->cv);
1918 #else /* RX_ENABLE_LOCKS */
1919 if (rx_waitForPacket)
1920 osi_rxWakeup(rx_waitForPacket);
1921 #endif /* RX_ENABLE_LOCKS */
1922 MUTEX_ENTER(&freeSQEList_lock);
1923 for (np = rx_FreeSQEList; np; np = tqp) {
1924 tqp = *(struct rx_serverQueueEntry **)np;
1925 #ifdef RX_ENABLE_LOCKS
1926 CV_BROADCAST(&np->cv);
1927 #else /* RX_ENABLE_LOCKS */
1929 #endif /* RX_ENABLE_LOCKS */
1931 MUTEX_EXIT(&freeSQEList_lock);
1932 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1933 #ifdef RX_ENABLE_LOCKS
1934 CV_BROADCAST(&np->cv);
1935 #else /* RX_ENABLE_LOCKS */
1937 #endif /* RX_ENABLE_LOCKS */
1939 MUTEX_EXIT(&rx_serverPool_lock);
1944 * One thing that seems to happen is that all the server threads get
1945 * tied up on some empty or slow call, and then a whole bunch of calls
1946 * arrive at once, using up the packet pool, so now there are more
1947 * empty calls. The most critical resources here are server threads
1948 * and the free packet pool. The "doreclaim" code seems to help in
1949 * general. I think that eventually we arrive in this state: there
1950 * are lots of pending calls which do have all their packets present,
1951 * so they won't be reclaimed, are multi-packet calls, so they won't
1952 * be scheduled until later, and thus are tying up most of the free
1953 * packet pool for a very long time.
1955 * 1. schedule multi-packet calls if all the packets are present.
1956 * Probably CPU-bound operation, useful to return packets to pool.
1957 * Do what if there is a full window, but the last packet isn't here?
1958 * 3. preserve one thread which *only* runs "best" calls, otherwise
1959 * it sleeps and waits for that type of call.
1960 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1961 * the current dataquota business is badly broken. The quota isn't adjusted
1962 * to reflect how many packets are presently queued for a running call.
1963 * So, when we schedule a queued call with a full window of packets queued
1964 * up for it, that *should* free up a window full of packets for other 2d-class
1965 * calls to be able to use from the packet pool. But it doesn't.
1967 * NB. Most of the time, this code doesn't run -- since idle server threads
1968 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1969 * as a new call arrives.
1971 /* Sleep until a call arrives. Returns a pointer to the call, ready
1972 * for an rx_Read. */
1973 #ifdef RX_ENABLE_LOCKS
1975 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1977 struct rx_serverQueueEntry *sq;
1978 struct rx_call *call = (struct rx_call *)0;
1979 struct rx_service *service = NULL;
1981 MUTEX_ENTER(&freeSQEList_lock);
1983 if ((sq = rx_FreeSQEList)) {
1984 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1985 MUTEX_EXIT(&freeSQEList_lock);
1986 } else { /* otherwise allocate a new one and return that */
1987 MUTEX_EXIT(&freeSQEList_lock);
1988 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1989 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1990 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1993 MUTEX_ENTER(&rx_serverPool_lock);
1994 if (cur_service != NULL) {
1995 ReturnToServerPool(cur_service);
1998 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1999 struct rx_call *tcall, *ncall, *choice2 = NULL;
2001 /* Scan for eligible incoming calls. A call is not eligible
2002 * if the maximum number of calls for its service type are
2003 * already executing */
2004 /* One thread will process calls FCFS (to prevent starvation),
2005 * while the other threads may run ahead looking for calls which
2006 * have all their input data available immediately. This helps
2007 * keep threads from blocking, waiting for data from the client. */
2008 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
2009 service = tcall->conn->service;
2010 if (!QuotaOK(service)) {
2013 MUTEX_ENTER(&rx_pthread_mutex);
2014 if (tno == rxi_fcfs_thread_num
2015 || queue_IsLast(&rx_incomingCallQueue, tcall)) {
2016 MUTEX_EXIT(&rx_pthread_mutex);
2017 /* If we're the fcfs thread , then we'll just use
2018 * this call. If we haven't been able to find an optimal
2019 * choice, and we're at the end of the list, then use a
2020 * 2d choice if one has been identified. Otherwise... */
2021 call = (choice2 ? choice2 : tcall);
2022 service = call->conn->service;
2024 MUTEX_EXIT(&rx_pthread_mutex);
2025 if (!queue_IsEmpty(&tcall->rq)) {
2026 struct rx_packet *rp;
2027 rp = queue_First(&tcall->rq, rx_packet);
2028 if (rp->header.seq == 1) {
2030 || (rp->header.flags & RX_LAST_PACKET)) {
2032 } else if (rxi_2dchoice && !choice2
2033 && !(tcall->flags & RX_CALL_CLEARED)
2034 && (tcall->rprev > rxi_HardAckRate)) {
2044 ReturnToServerPool(service);
2051 MUTEX_EXIT(&rx_serverPool_lock);
2052 MUTEX_ENTER(&call->lock);
2054 if (call->flags & RX_CALL_WAIT_PROC) {
2055 call->flags &= ~RX_CALL_WAIT_PROC;
2056 rx_atomic_dec(&rx_nWaiting);
2059 if (call->state != RX_STATE_PRECALL || call->error) {
2060 MUTEX_EXIT(&call->lock);
2061 MUTEX_ENTER(&rx_serverPool_lock);
2062 ReturnToServerPool(service);
2067 if (queue_IsEmpty(&call->rq)
2068 || queue_First(&call->rq, rx_packet)->header.seq != 1)
2069 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2071 CLEAR_CALL_QUEUE_LOCK(call);
2074 /* If there are no eligible incoming calls, add this process
2075 * to the idle server queue, to wait for one */
2079 *socketp = OSI_NULLSOCKET;
2081 sq->socketp = socketp;
2082 queue_Append(&rx_idleServerQueue, sq);
2083 #ifndef AFS_AIX41_ENV
2084 rx_waitForPacket = sq;
2086 rx_waitingForPacket = sq;
2087 #endif /* AFS_AIX41_ENV */
2089 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2091 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2092 MUTEX_EXIT(&rx_serverPool_lock);
2093 return (struct rx_call *)0;
2096 } while (!(call = sq->newcall)
2097 && !(socketp && *socketp != OSI_NULLSOCKET));
2098 MUTEX_EXIT(&rx_serverPool_lock);
2100 MUTEX_ENTER(&call->lock);
2106 MUTEX_ENTER(&freeSQEList_lock);
2107 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2108 rx_FreeSQEList = sq;
2109 MUTEX_EXIT(&freeSQEList_lock);
2112 clock_GetTime(&call->startTime);
2113 call->state = RX_STATE_ACTIVE;
2114 call->mode = RX_MODE_RECEIVING;
2115 #ifdef RX_KERNEL_TRACE
2116 if (ICL_SETACTIVE(afs_iclSetp)) {
2117 int glockOwner = ISAFS_GLOCK();
2120 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2121 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2128 rxi_calltrace(RX_CALL_START, call);
2129 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2130 call->conn->service->servicePort, call->conn->service->serviceId,
2133 MUTEX_EXIT(&call->lock);
2134 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2136 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2141 #else /* RX_ENABLE_LOCKS */
2143 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2145 struct rx_serverQueueEntry *sq;
2146 struct rx_call *call = (struct rx_call *)0, *choice2;
2147 struct rx_service *service = NULL;
2151 MUTEX_ENTER(&freeSQEList_lock);
2153 if ((sq = rx_FreeSQEList)) {
2154 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2155 MUTEX_EXIT(&freeSQEList_lock);
2156 } else { /* otherwise allocate a new one and return that */
2157 MUTEX_EXIT(&freeSQEList_lock);
2158 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2159 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2160 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2162 MUTEX_ENTER(&sq->lock);
2164 if (cur_service != NULL) {
2165 cur_service->nRequestsRunning--;
2166 MUTEX_ENTER(&rx_quota_mutex);
2167 if (cur_service->nRequestsRunning < cur_service->minProcs)
2170 MUTEX_EXIT(&rx_quota_mutex);
2172 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
2173 struct rx_call *tcall, *ncall;
2174 /* Scan for eligible incoming calls. A call is not eligible
2175 * if the maximum number of calls for its service type are
2176 * already executing */
2177 /* One thread will process calls FCFS (to prevent starvation),
2178 * while the other threads may run ahead looking for calls which
2179 * have all their input data available immediately. This helps
2180 * keep threads from blocking, waiting for data from the client. */
2181 choice2 = (struct rx_call *)0;
2182 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
2183 service = tcall->conn->service;
2184 if (QuotaOK(service)) {
2185 MUTEX_ENTER(&rx_pthread_mutex);
2186 if (tno == rxi_fcfs_thread_num
2187 || !tcall->queue_item_header.next) {
2188 MUTEX_EXIT(&rx_pthread_mutex);
2189 /* If we're the fcfs thread, then we'll just use
2190 * this call. If we haven't been able to find an optimal
2191 * choice, and we're at the end of the list, then use a
2192 * 2d choice if one has been identified. Otherwise... */
2193 call = (choice2 ? choice2 : tcall);
2194 service = call->conn->service;
2196 MUTEX_EXIT(&rx_pthread_mutex);
2197 if (!queue_IsEmpty(&tcall->rq)) {
2198 struct rx_packet *rp;
2199 rp = queue_First(&tcall->rq, rx_packet);
2200 if (rp->header.seq == 1
2202 || (rp->header.flags & RX_LAST_PACKET))) {
2204 } else if (rxi_2dchoice && !choice2
2205 && !(tcall->flags & RX_CALL_CLEARED)
2206 && (tcall->rprev > rxi_HardAckRate)) {
2220 /* we can't schedule a call if there's no data!!! */
2221 /* send an ack if there's no data, if we're missing the
2222 * first packet, or we're missing something between first
2223 * and last -- there's a "hole" in the incoming data. */
2224 if (queue_IsEmpty(&call->rq)
2225 || queue_First(&call->rq, rx_packet)->header.seq != 1
2226 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
2227 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2229 call->flags &= (~RX_CALL_WAIT_PROC);
2230 service->nRequestsRunning++;
2231 /* just started call in minProcs pool, need fewer to maintain
2233 MUTEX_ENTER(&rx_quota_mutex);
2234 if (service->nRequestsRunning <= service->minProcs)
2237 MUTEX_EXIT(&rx_quota_mutex);
2238 rx_atomic_dec(&rx_nWaiting);
2239 /* MUTEX_EXIT(&call->lock); */
2241 /* If there are no eligible incoming calls, add this process
2242 * to the idle server queue, to wait for one */
2245 *socketp = OSI_NULLSOCKET;
2247 sq->socketp = socketp;
2248 queue_Append(&rx_idleServerQueue, sq);
2252 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2254 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2255 return (struct rx_call *)0;
2258 } while (!(call = sq->newcall)
2259 && !(socketp && *socketp != OSI_NULLSOCKET));
2261 MUTEX_EXIT(&sq->lock);
2263 MUTEX_ENTER(&freeSQEList_lock);
2264 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2265 rx_FreeSQEList = sq;
2266 MUTEX_EXIT(&freeSQEList_lock);
2269 clock_GetTime(&call->startTime);
2270 call->state = RX_STATE_ACTIVE;
2271 call->mode = RX_MODE_RECEIVING;
2272 #ifdef RX_KERNEL_TRACE
2273 if (ICL_SETACTIVE(afs_iclSetp)) {
2274 int glockOwner = ISAFS_GLOCK();
2277 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2278 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2285 rxi_calltrace(RX_CALL_START, call);
2286 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2287 call->conn->service->servicePort, call->conn->service->serviceId,
2290 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2297 #endif /* RX_ENABLE_LOCKS */
2301 /* Establish a procedure to be called when a packet arrives for a
2302 * call. This routine will be called at most once after each call,
2303 * and will also be called if there is an error condition on the or
2304 * the call is complete. Used by multi rx to build a selection
2305 * function which determines which of several calls is likely to be a
2306 * good one to read from.
2307 * NOTE: the way this is currently implemented it is probably only a
2308 * good idea to (1) use it immediately after a newcall (clients only)
2309 * and (2) only use it once. Other uses currently void your warranty
2312 rx_SetArrivalProc(struct rx_call *call,
2313 void (*proc) (struct rx_call * call,
2316 void * handle, int arg)
2318 call->arrivalProc = proc;
2319 call->arrivalProcHandle = handle;
2320 call->arrivalProcArg = arg;
2323 /* Call is finished (possibly prematurely). Return rc to the peer, if
2324 * appropriate, and return the final error code from the conversation
2328 rx_EndCall(struct rx_call *call, afs_int32 rc)
2330 struct rx_connection *conn = call->conn;
2334 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2335 call, rc, call->error, call->abortCode));
2338 MUTEX_ENTER(&call->lock);
2340 if (rc == 0 && call->error == 0) {
2341 call->abortCode = 0;
2342 call->abortCount = 0;
2345 call->arrivalProc = (void (*)())0;
2346 if (rc && call->error == 0) {
2347 rxi_CallError(call, rc);
2348 call->mode = RX_MODE_ERROR;
2349 /* Send an abort message to the peer if this error code has
2350 * only just been set. If it was set previously, assume the
2351 * peer has already been sent the error code or will request it
2353 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2355 if (conn->type == RX_SERVER_CONNECTION) {
2356 /* Make sure reply or at least dummy reply is sent */
2357 if (call->mode == RX_MODE_RECEIVING) {
2358 MUTEX_EXIT(&call->lock);
2359 rxi_WriteProc(call, 0, 0);
2360 MUTEX_ENTER(&call->lock);
2362 if (call->mode == RX_MODE_SENDING) {
2363 MUTEX_EXIT(&call->lock);
2364 rxi_FlushWrite(call);
2365 MUTEX_ENTER(&call->lock);
2367 rxi_calltrace(RX_CALL_END, call);
2368 /* Call goes to hold state until reply packets are acknowledged */
2369 if (call->tfirst + call->nSoftAcked < call->tnext) {
2370 call->state = RX_STATE_HOLD;
2372 call->state = RX_STATE_DALLY;
2373 rxi_ClearTransmitQueue(call, 0);
2374 rxi_rto_cancel(call);
2375 rxevent_Cancel(&call->keepAliveEvent, call,
2376 RX_CALL_REFCOUNT_ALIVE);
2378 } else { /* Client connection */
2380 /* Make sure server receives input packets, in the case where
2381 * no reply arguments are expected */
2382 if ((call->mode == RX_MODE_SENDING)
2383 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2384 MUTEX_EXIT(&call->lock);
2385 (void)rxi_ReadProc(call, &dummy, 1);
2386 MUTEX_ENTER(&call->lock);
2389 /* If we had an outstanding delayed ack, be nice to the server
2390 * and force-send it now.
2392 if (call->delayedAckEvent) {
2393 rxevent_Cancel(&call->delayedAckEvent, call,
2394 RX_CALL_REFCOUNT_DELAY);
2395 rxi_SendDelayedAck(NULL, call, NULL, 0);
2398 /* We need to release the call lock since it's lower than the
2399 * conn_call_lock and we don't want to hold the conn_call_lock
2400 * over the rx_ReadProc call. The conn_call_lock needs to be held
2401 * here for the case where rx_NewCall is perusing the calls on
2402 * the connection structure. We don't want to signal until
2403 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2404 * have checked this call, found it active and by the time it
2405 * goes to sleep, will have missed the signal.
2407 MUTEX_EXIT(&call->lock);
2408 MUTEX_ENTER(&conn->conn_call_lock);
2409 MUTEX_ENTER(&call->lock);
2411 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2412 conn->lastBusy[call->channel] = 0;
2415 MUTEX_ENTER(&conn->conn_data_lock);
2416 conn->flags |= RX_CONN_BUSY;
2417 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2418 MUTEX_EXIT(&conn->conn_data_lock);
2419 #ifdef RX_ENABLE_LOCKS
2420 CV_BROADCAST(&conn->conn_call_cv);
2425 #ifdef RX_ENABLE_LOCKS
2427 MUTEX_EXIT(&conn->conn_data_lock);
2429 #endif /* RX_ENABLE_LOCKS */
2430 call->state = RX_STATE_DALLY;
2432 error = call->error;
2434 /* currentPacket, nLeft, and NFree must be zeroed here, because
2435 * ResetCall cannot: ResetCall may be called at splnet(), in the
2436 * kernel version, and may interrupt the macros rx_Read or
2437 * rx_Write, which run at normal priority for efficiency. */
2438 if (call->currentPacket) {
2439 #ifdef RX_TRACK_PACKETS
2440 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2442 rxi_FreePacket(call->currentPacket);
2443 call->currentPacket = (struct rx_packet *)0;
2446 call->nLeft = call->nFree = call->curlen = 0;
2448 /* Free any packets from the last call to ReadvProc/WritevProc */
2449 #ifdef RXDEBUG_PACKET
2451 #endif /* RXDEBUG_PACKET */
2452 rxi_FreePackets(0, &call->iovq);
2453 MUTEX_EXIT(&call->lock);
2455 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2456 if (conn->type == RX_CLIENT_CONNECTION) {
2457 MUTEX_ENTER(&conn->conn_data_lock);
2458 conn->flags &= ~RX_CONN_BUSY;
2459 MUTEX_EXIT(&conn->conn_data_lock);
2460 MUTEX_EXIT(&conn->conn_call_lock);
2464 * Map errors to the local host's errno.h format.
2466 error = ntoh_syserr_conv(error);
2470 #if !defined(KERNEL)
2472 /* Call this routine when shutting down a server or client (especially
2473 * clients). This will allow Rx to gracefully garbage collect server
2474 * connections, and reduce the number of retries that a server might
2475 * make to a dead client.
2476 * This is not quite right, since some calls may still be ongoing and
2477 * we can't lock them to destroy them. */
2481 struct rx_connection **conn_ptr, **conn_end;
2485 if (rxinit_status == 1) {
2487 return; /* Already shutdown. */
2489 rxi_DeleteCachedConnections();
2490 if (rx_connHashTable) {
2491 MUTEX_ENTER(&rx_connHashTable_lock);
2492 for (conn_ptr = &rx_connHashTable[0], conn_end =
2493 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2495 struct rx_connection *conn, *next;
2496 for (conn = *conn_ptr; conn; conn = next) {
2498 if (conn->type == RX_CLIENT_CONNECTION) {
2499 MUTEX_ENTER(&rx_refcnt_mutex);
2501 MUTEX_EXIT(&rx_refcnt_mutex);
2502 #ifdef RX_ENABLE_LOCKS
2503 rxi_DestroyConnectionNoLock(conn);
2504 #else /* RX_ENABLE_LOCKS */
2505 rxi_DestroyConnection(conn);
2506 #endif /* RX_ENABLE_LOCKS */
2510 #ifdef RX_ENABLE_LOCKS
2511 while (rx_connCleanup_list) {
2512 struct rx_connection *conn;
2513 conn = rx_connCleanup_list;
2514 rx_connCleanup_list = rx_connCleanup_list->next;
2515 MUTEX_EXIT(&rx_connHashTable_lock);
2516 rxi_CleanupConnection(conn);
2517 MUTEX_ENTER(&rx_connHashTable_lock);
2519 MUTEX_EXIT(&rx_connHashTable_lock);
2520 #endif /* RX_ENABLE_LOCKS */
2525 afs_winsockCleanup();
2533 /* if we wakeup packet waiter too often, can get in loop with two
2534 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2536 rxi_PacketsUnWait(void)
2538 if (!rx_waitingForPackets) {
2542 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2543 return; /* still over quota */
2546 rx_waitingForPackets = 0;
2547 #ifdef RX_ENABLE_LOCKS
2548 CV_BROADCAST(&rx_waitingForPackets_cv);
2550 osi_rxWakeup(&rx_waitingForPackets);
2556 /* ------------------Internal interfaces------------------------- */
2558 /* Return this process's service structure for the
2559 * specified socket and service */
2560 static struct rx_service *
2561 rxi_FindService(osi_socket socket, u_short serviceId)
2563 struct rx_service **sp;
2564 for (sp = &rx_services[0]; *sp; sp++) {
2565 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2571 #ifdef RXDEBUG_PACKET
2572 #ifdef KDUMP_RX_LOCK
2573 static struct rx_call_rx_lock *rx_allCallsp = 0;
2575 static struct rx_call *rx_allCallsp = 0;
2577 #endif /* RXDEBUG_PACKET */
2579 /* Allocate a call structure, for the indicated channel of the
2580 * supplied connection. The mode and state of the call must be set by
2581 * the caller. Returns the call with mutex locked. */
2582 static struct rx_call *
2583 rxi_NewCall(struct rx_connection *conn, int channel)
2585 struct rx_call *call;
2586 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2587 struct rx_call *cp; /* Call pointer temp */
2588 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2589 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2591 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2593 /* Grab an existing call structure, or allocate a new one.
2594 * Existing call structures are assumed to have been left reset by
2596 MUTEX_ENTER(&rx_freeCallQueue_lock);
2598 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2600 * EXCEPT that the TQ might not yet be cleared out.
2601 * Skip over those with in-use TQs.
2604 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2605 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2611 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2612 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2613 call = queue_First(&rx_freeCallQueue, rx_call);
2614 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2616 if (rx_stats_active)
2617 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2618 MUTEX_EXIT(&rx_freeCallQueue_lock);
2619 MUTEX_ENTER(&call->lock);
2620 CLEAR_CALL_QUEUE_LOCK(call);
2621 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2622 /* Now, if TQ wasn't cleared earlier, do it now. */
2623 rxi_WaitforTQBusy(call);
2624 if (call->flags & RX_CALL_TQ_CLEARME) {
2625 rxi_ClearTransmitQueue(call, 1);
2626 /*queue_Init(&call->tq);*/
2628 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2629 /* Bind the call to its connection structure */
2631 rxi_ResetCall(call, 1);
2634 call = rxi_Alloc(sizeof(struct rx_call));
2635 #ifdef RXDEBUG_PACKET
2636 call->allNextp = rx_allCallsp;
2637 rx_allCallsp = call;
2639 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2640 #else /* RXDEBUG_PACKET */
2641 rx_atomic_inc(&rx_stats.nCallStructs);
2642 #endif /* RXDEBUG_PACKET */
2644 MUTEX_EXIT(&rx_freeCallQueue_lock);
2645 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2646 MUTEX_ENTER(&call->lock);
2647 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2648 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2649 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2651 /* Initialize once-only items */
2652 queue_Init(&call->tq);
2653 queue_Init(&call->rq);
2654 queue_Init(&call->iovq);
2655 #ifdef RXDEBUG_PACKET
2656 call->rqc = call->tqc = call->iovqc = 0;
2657 #endif /* RXDEBUG_PACKET */
2658 /* Bind the call to its connection structure (prereq for reset) */
2660 rxi_ResetCall(call, 1);
2662 call->channel = channel;
2663 call->callNumber = &conn->callNumber[channel];
2664 call->rwind = conn->rwind[channel];
2665 call->twind = conn->twind[channel];
2666 /* Note that the next expected call number is retained (in
2667 * conn->callNumber[i]), even if we reallocate the call structure
2669 conn->call[channel] = call;
2670 /* if the channel's never been used (== 0), we should start at 1, otherwise
2671 * the call number is valid from the last time this channel was used */
2672 if (*call->callNumber == 0)
2673 *call->callNumber = 1;
2678 /* A call has been inactive long enough that so we can throw away
2679 * state, including the call structure, which is placed on the call
2682 * call->lock amd rx_refcnt_mutex are held upon entry.
2683 * haveCTLock is set when called from rxi_ReapConnections.
2685 * return 1 if the call is freed, 0 if not.
2688 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2690 int channel = call->channel;
2691 struct rx_connection *conn = call->conn;
2692 u_char state = call->state;
2695 * We are setting the state to RX_STATE_RESET to
2696 * ensure that no one else will attempt to use this
2697 * call once we drop the refcnt lock. We must drop
2698 * the refcnt lock before calling rxi_ResetCall
2699 * because it cannot be held across acquiring the
2700 * freepktQ lock. NewCall does the same.
2702 call->state = RX_STATE_RESET;
2703 MUTEX_EXIT(&rx_refcnt_mutex);
2704 rxi_ResetCall(call, 0);
2706 if (MUTEX_TRYENTER(&conn->conn_call_lock))
2708 if (state == RX_STATE_DALLY || state == RX_STATE_HOLD)
2709 (*call->callNumber)++;
2711 if (call->conn->call[channel] == call)
2712 call->conn->call[channel] = 0;
2713 MUTEX_EXIT(&conn->conn_call_lock);
2716 * We couldn't obtain the conn_call_lock so we can't
2717 * disconnect the call from the connection. Set the
2718 * call state to dally so that the call can be reused.
2720 MUTEX_ENTER(&rx_refcnt_mutex);
2721 call->state = RX_STATE_DALLY;
2725 MUTEX_ENTER(&rx_freeCallQueue_lock);
2726 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2727 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2728 /* A call may be free even though its transmit queue is still in use.
2729 * Since we search the call list from head to tail, put busy calls at
2730 * the head of the list, and idle calls at the tail.
2732 if (call->flags & RX_CALL_TQ_BUSY)
2733 queue_Prepend(&rx_freeCallQueue, call);
2735 queue_Append(&rx_freeCallQueue, call);
2736 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2737 queue_Append(&rx_freeCallQueue, call);
2738 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2739 if (rx_stats_active)
2740 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2741 MUTEX_EXIT(&rx_freeCallQueue_lock);
2743 /* Destroy the connection if it was previously slated for
2744 * destruction, i.e. the Rx client code previously called
2745 * rx_DestroyConnection (client connections), or
2746 * rxi_ReapConnections called the same routine (server
2747 * connections). Only do this, however, if there are no
2748 * outstanding calls. Note that for fine grain locking, there appears
2749 * to be a deadlock in that rxi_FreeCall has a call locked and
2750 * DestroyConnectionNoLock locks each call in the conn. But note a
2751 * few lines up where we have removed this call from the conn.
2752 * If someone else destroys a connection, they either have no
2753 * call lock held or are going through this section of code.
2755 MUTEX_ENTER(&conn->conn_data_lock);
2756 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2757 MUTEX_ENTER(&rx_refcnt_mutex);
2759 MUTEX_EXIT(&rx_refcnt_mutex);
2760 MUTEX_EXIT(&conn->conn_data_lock);
2761 #ifdef RX_ENABLE_LOCKS
2763 rxi_DestroyConnectionNoLock(conn);
2765 rxi_DestroyConnection(conn);
2766 #else /* RX_ENABLE_LOCKS */
2767 rxi_DestroyConnection(conn);
2768 #endif /* RX_ENABLE_LOCKS */
2770 MUTEX_EXIT(&conn->conn_data_lock);
2772 MUTEX_ENTER(&rx_refcnt_mutex);
2776 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2777 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2780 rxi_Alloc(size_t size)
2784 if (rx_stats_active) {
2785 rx_atomic_add(&rxi_Allocsize, (int) size);
2786 rx_atomic_inc(&rxi_Alloccnt);
2790 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2791 afs_osi_Alloc_NoSleep(size);
2796 osi_Panic("rxi_Alloc error");
2802 rxi_Free(void *addr, size_t size)
2804 if (rx_stats_active) {
2805 rx_atomic_sub(&rxi_Allocsize, (int) size);
2806 rx_atomic_dec(&rxi_Alloccnt);
2808 osi_Free(addr, size);
2812 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2814 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2815 struct rx_peer *next = NULL;
2819 MUTEX_ENTER(&rx_peerHashTable_lock);
2821 peer_ptr = &rx_peerHashTable[0];
2822 peer_end = &rx_peerHashTable[rx_hashTableSize];
2825 for ( ; peer_ptr < peer_end; peer_ptr++) {
2828 for ( ; peer; peer = next) {
2830 if (host == peer->host)
2835 hashIndex = PEER_HASH(host, port);
2836 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2837 if ((peer->host == host) && (peer->port == port))
2842 MUTEX_ENTER(&rx_peerHashTable_lock);
2847 MUTEX_EXIT(&rx_peerHashTable_lock);
2849 MUTEX_ENTER(&peer->peer_lock);
2850 /* We don't handle dropping below min, so don't */
2851 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2852 peer->ifMTU=MIN(mtu, peer->ifMTU);
2853 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2854 /* if we tweaked this down, need to tune our peer MTU too */
2855 peer->MTU = MIN(peer->MTU, peer->natMTU);
2856 /* if we discovered a sub-1500 mtu, degrade */
2857 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2858 peer->maxDgramPackets = 1;
2859 /* We no longer have valid peer packet information */
2860 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2861 peer->maxPacketSize = 0;
2862 MUTEX_EXIT(&peer->peer_lock);
2864 MUTEX_ENTER(&rx_peerHashTable_lock);
2866 if (host && !port) {
2868 /* pick up where we left off */
2872 MUTEX_EXIT(&rx_peerHashTable_lock);
2875 /* Find the peer process represented by the supplied (host,port)
2876 * combination. If there is no appropriate active peer structure, a
2877 * new one will be allocated and initialized
2878 * The origPeer, if set, is a pointer to a peer structure on which the
2879 * refcount will be be decremented. This is used to replace the peer
2880 * structure hanging off a connection structure */
2882 rxi_FindPeer(afs_uint32 host, u_short port,
2883 struct rx_peer *origPeer, int create)
2887 hashIndex = PEER_HASH(host, port);
2888 MUTEX_ENTER(&rx_peerHashTable_lock);
2889 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2890 if ((pp->host == host) && (pp->port == port))
2895 pp = rxi_AllocPeer(); /* This bzero's *pp */
2896 pp->host = host; /* set here or in InitPeerParams is zero */
2898 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2899 queue_Init(&pp->congestionQueue);
2900 queue_Init(&pp->rpcStats);
2901 pp->next = rx_peerHashTable[hashIndex];
2902 rx_peerHashTable[hashIndex] = pp;
2903 rxi_InitPeerParams(pp);
2904 if (rx_stats_active)
2905 rx_atomic_inc(&rx_stats.nPeerStructs);
2912 origPeer->refCount--;
2913 MUTEX_EXIT(&rx_peerHashTable_lock);
2918 /* Find the connection at (host, port) started at epoch, and with the
2919 * given connection id. Creates the server connection if necessary.
2920 * The type specifies whether a client connection or a server
2921 * connection is desired. In both cases, (host, port) specify the
2922 * peer's (host, pair) pair. Client connections are not made
2923 * automatically by this routine. The parameter socket gives the
2924 * socket descriptor on which the packet was received. This is used,
2925 * in the case of server connections, to check that *new* connections
2926 * come via a valid (port, serviceId). Finally, the securityIndex
2927 * parameter must match the existing index for the connection. If a
2928 * server connection is created, it will be created using the supplied
2929 * index, if the index is valid for this service */
2930 struct rx_connection *
2931 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2932 u_short port, u_short serviceId, afs_uint32 cid,
2933 afs_uint32 epoch, int type, u_int securityIndex)
2935 int hashindex, flag, i;
2936 struct rx_connection *conn;
2937 hashindex = CONN_HASH(host, port, cid, epoch, type);
2938 MUTEX_ENTER(&rx_connHashTable_lock);
2939 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2940 rx_connHashTable[hashindex],
2943 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2944 && (epoch == conn->epoch)) {
2945 struct rx_peer *pp = conn->peer;
2946 if (securityIndex != conn->securityIndex) {
2947 /* this isn't supposed to happen, but someone could forge a packet
2948 * like this, and there seems to be some CM bug that makes this
2949 * happen from time to time -- in which case, the fileserver
2951 MUTEX_EXIT(&rx_connHashTable_lock);
2952 return (struct rx_connection *)0;
2954 if (pp->host == host && pp->port == port)
2956 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2958 /* So what happens when it's a callback connection? */
2959 if ( /*type == RX_CLIENT_CONNECTION && */
2960 (conn->epoch & 0x80000000))
2964 /* the connection rxLastConn that was used the last time is not the
2965 ** one we are looking for now. Hence, start searching in the hash */
2967 conn = rx_connHashTable[hashindex];
2972 struct rx_service *service;
2973 if (type == RX_CLIENT_CONNECTION) {
2974 MUTEX_EXIT(&rx_connHashTable_lock);
2975 return (struct rx_connection *)0;
2977 service = rxi_FindService(socket, serviceId);
2978 if (!service || (securityIndex >= service->nSecurityObjects)
2979 || (service->securityObjects[securityIndex] == 0)) {
2980 MUTEX_EXIT(&rx_connHashTable_lock);
2981 return (struct rx_connection *)0;
2983 conn = rxi_AllocConnection(); /* This bzero's the connection */
2984 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2985 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2986 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2987 conn->next = rx_connHashTable[hashindex];
2988 rx_connHashTable[hashindex] = conn;
2989 conn->peer = rxi_FindPeer(host, port, 0, 1);
2990 conn->type = RX_SERVER_CONNECTION;
2991 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2992 conn->epoch = epoch;
2993 conn->cid = cid & RX_CIDMASK;
2994 /* conn->serial = conn->lastSerial = 0; */
2995 /* conn->timeout = 0; */
2996 conn->ackRate = RX_FAST_ACK_RATE;
2997 conn->service = service;
2998 conn->serviceId = serviceId;
2999 conn->securityIndex = securityIndex;
3000 conn->securityObject = service->securityObjects[securityIndex];
3001 conn->nSpecific = 0;
3002 conn->specific = NULL;
3003 rx_SetConnDeadTime(conn, service->connDeadTime);
3004 conn->idleDeadTime = service->idleDeadTime;
3005 conn->idleDeadDetection = service->idleDeadErr ? 1 : 0;
3006 for (i = 0; i < RX_MAXCALLS; i++) {
3007 conn->twind[i] = rx_initSendWindow;
3008 conn->rwind[i] = rx_initReceiveWindow;
3010 /* Notify security object of the new connection */
3011 RXS_NewConnection(conn->securityObject, conn);
3012 /* XXXX Connection timeout? */
3013 if (service->newConnProc)
3014 (*service->newConnProc) (conn);
3015 if (rx_stats_active)
3016 rx_atomic_inc(&rx_stats.nServerConns);
3019 MUTEX_ENTER(&rx_refcnt_mutex);
3021 MUTEX_EXIT(&rx_refcnt_mutex);
3023 rxLastConn = conn; /* store this connection as the last conn used */
3024 MUTEX_EXIT(&rx_connHashTable_lock);
3029 * Timeout a call on a busy call channel if appropriate.
3031 * @param[in] call The busy call.
3033 * @pre 'call' is marked as busy (namely,
3034 * call->conn->lastBusy[call->channel] != 0)
3036 * @pre call->lock is held
3037 * @pre rxi_busyChannelError is nonzero
3039 * @note call->lock is dropped and reacquired
3042 rxi_CheckBusy(struct rx_call *call)
3044 struct rx_connection *conn = call->conn;
3045 int channel = call->channel;
3046 int freechannel = 0;
3048 afs_uint32 callNumber;
3050 MUTEX_EXIT(&call->lock);
3052 MUTEX_ENTER(&conn->conn_call_lock);
3053 callNumber = *call->callNumber;
3055 /* Are there any other call slots on this conn that we should try? Look for
3056 * slots that are empty and are either non-busy, or were marked as busy
3057 * longer than conn->secondsUntilDead seconds before this call started. */
3059 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3061 /* only look at channels that aren't us */
3065 if (conn->lastBusy[i]) {
3066 /* if this channel looked busy too recently, don't look at it */
3067 if (conn->lastBusy[i] >= call->startTime.sec) {
3070 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3075 if (conn->call[i]) {
3076 struct rx_call *tcall = conn->call[i];
3077 MUTEX_ENTER(&tcall->lock);
3078 if (tcall->state == RX_STATE_DALLY) {
3081 MUTEX_EXIT(&tcall->lock);
3087 MUTEX_ENTER(&call->lock);
3089 /* Since the call->lock and conn->conn_call_lock have been released it is
3090 * possible that (1) the call may no longer be busy and/or (2) the call may
3091 * have been reused by another waiting thread. Therefore, we must confirm
3092 * that the call state has not changed when deciding whether or not to
3093 * force this application thread to retry by forcing a Timeout error. */
3095 if (freechannel && *call->callNumber == callNumber &&
3096 (call->flags & RX_CALL_PEER_BUSY)) {
3097 /* Since 'freechannel' is set, there exists another channel in this
3098 * rx_conn that the application thread might be able to use. We know
3099 * that we have the correct call since callNumber is unchanged, and we
3100 * know that the call is still busy. So, set the call error state to
3101 * rxi_busyChannelError so the application can retry the request,
3102 * presumably on a less-busy call channel. */
3104 rxi_CallError(call, RX_CALL_BUSY);
3106 MUTEX_EXIT(&conn->conn_call_lock);
3109 /* There are two packet tracing routines available for testing and monitoring
3110 * Rx. One is called just after every packet is received and the other is
3111 * called just before every packet is sent. Received packets, have had their
3112 * headers decoded, and packets to be sent have not yet had their headers
3113 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3114 * containing the network address. Both can be modified. The return value, if
3115 * non-zero, indicates that the packet should be dropped. */
3117 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3118 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3120 /* A packet has been received off the interface. Np is the packet, socket is
3121 * the socket number it was received from (useful in determining which service
3122 * this packet corresponds to), and (host, port) reflect the host,port of the
3123 * sender. This call returns the packet to the caller if it is finished with
3124 * it, rather than de-allocating it, just as a small performance hack */
3127 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3128 afs_uint32 host, u_short port, int *tnop,
3129 struct rx_call **newcallp)
3131 struct rx_call *call;
3132 struct rx_connection *conn;
3134 afs_uint32 currentCallNumber;
3140 struct rx_packet *tnp;
3143 /* We don't print out the packet until now because (1) the time may not be
3144 * accurate enough until now in the lwp implementation (rx_Listener only gets
3145 * the time after the packet is read) and (2) from a protocol point of view,
3146 * this is the first time the packet has been seen */
3147 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3148 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3149 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3150 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3151 np->header.epoch, np->header.cid, np->header.callNumber,
3152 np->header.seq, np->header.flags, np));
3155 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3156 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3159 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3160 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3163 /* If an input tracer function is defined, call it with the packet and
3164 * network address. Note this function may modify its arguments. */
3165 if (rx_justReceived) {
3166 struct sockaddr_in addr;
3168 addr.sin_family = AF_INET;
3169 addr.sin_port = port;
3170 addr.sin_addr.s_addr = host;
3171 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3172 addr.sin_len = sizeof(addr);
3173 #endif /* AFS_OSF_ENV */
3174 drop = (*rx_justReceived) (np, &addr);
3175 /* drop packet if return value is non-zero */
3178 port = addr.sin_port; /* in case fcn changed addr */
3179 host = addr.sin_addr.s_addr;
3183 /* If packet was not sent by the client, then *we* must be the client */
3184 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3185 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3187 /* Find the connection (or fabricate one, if we're the server & if
3188 * necessary) associated with this packet */
3190 rxi_FindConnection(socket, host, port, np->header.serviceId,
3191 np->header.cid, np->header.epoch, type,
3192 np->header.securityIndex);
3195 /* If no connection found or fabricated, just ignore the packet.
3196 * (An argument could be made for sending an abort packet for
3201 /* If the connection is in an error state, send an abort packet and ignore
3202 * the incoming packet */
3204 /* Don't respond to an abort packet--we don't want loops! */
3205 MUTEX_ENTER(&conn->conn_data_lock);
3206 if (np->header.type != RX_PACKET_TYPE_ABORT)
3207 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3208 putConnection(conn);
3209 MUTEX_EXIT(&conn->conn_data_lock);
3213 /* Check for connection-only requests (i.e. not call specific). */
3214 if (np->header.callNumber == 0) {
3215 switch (np->header.type) {
3216 case RX_PACKET_TYPE_ABORT: {
3217 /* What if the supplied error is zero? */
3218 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3219 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3220 rxi_ConnectionError(conn, errcode);
3221 putConnection(conn);
3224 case RX_PACKET_TYPE_CHALLENGE:
3225 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3226 putConnection(conn);
3228 case RX_PACKET_TYPE_RESPONSE:
3229 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3230 putConnection(conn);
3232 case RX_PACKET_TYPE_PARAMS:
3233 case RX_PACKET_TYPE_PARAMS + 1:
3234 case RX_PACKET_TYPE_PARAMS + 2:
3235 /* ignore these packet types for now */
3236 putConnection(conn);
3240 /* Should not reach here, unless the peer is broken: send an
3242 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3243 MUTEX_ENTER(&conn->conn_data_lock);
3244 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3245 putConnection(conn);
3246 MUTEX_EXIT(&conn->conn_data_lock);
3251 channel = np->header.cid & RX_CHANNELMASK;
3252 MUTEX_ENTER(&conn->conn_call_lock);
3253 call = conn->call[channel];
3256 MUTEX_ENTER(&call->lock);
3257 currentCallNumber = conn->callNumber[channel];
3258 MUTEX_EXIT(&conn->conn_call_lock);
3259 } else if (type == RX_SERVER_CONNECTION) { /* No call allocated */
3260 call = conn->call[channel];
3262 MUTEX_ENTER(&call->lock);
3263 currentCallNumber = conn->callNumber[channel];
3264 MUTEX_EXIT(&conn->conn_call_lock);
3266 call = rxi_NewCall(conn, channel); /* returns locked call */
3267 *call->callNumber = currentCallNumber = np->header.callNumber;
3268 MUTEX_EXIT(&conn->conn_call_lock);
3270 if (np->header.callNumber == 0)
3271 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3272 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3273 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3274 np->header.flags, np, np->length));
3276 call->state = RX_STATE_PRECALL;
3277 clock_GetTime(&call->queueTime);
3278 hzero(call->bytesSent);
3279 hzero(call->bytesRcvd);
3281 * If the number of queued calls exceeds the overload
3282 * threshold then abort this call.
3284 if ((rx_BusyThreshold > 0) &&
3285 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3286 struct rx_packet *tp;
3288 rxi_CallError(call, rx_BusyError);
3289 tp = rxi_SendCallAbort(call, np, 1, 0);
3290 MUTEX_EXIT(&call->lock);
3291 putConnection(conn);
3292 if (rx_stats_active)
3293 rx_atomic_inc(&rx_stats.nBusies);
3296 rxi_KeepAliveOn(call);
3298 } else { /* RX_CLIENT_CONNECTION and No call allocated */
3299 /* This packet can't be for this call. If the new call address is
3300 * 0 then no call is running on this channel. If there is a call
3301 * then, since this is a client connection we're getting data for
3302 * it must be for the previous call.
3304 if (rx_stats_active)
3305 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3306 putConnection(conn);
3310 /* There is a non-NULL locked call at this point */
3311 if (type == RX_SERVER_CONNECTION) { /* We're the server */
3312 if (np->header.callNumber < currentCallNumber) {
3313 MUTEX_EXIT(&call->lock);
3314 if (rx_stats_active)
3315 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3316 putConnection(conn);
3318 } else if (np->header.callNumber != currentCallNumber) {
3319 /* Wait until the transmit queue is idle before deciding
3320 * whether to reset the current call. Chances are that the
3321 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3324 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3325 if (call->state == RX_STATE_ACTIVE) {
3326 rxi_WaitforTQBusy(call);
3328 * If we entered error state while waiting,
3329 * must call rxi_CallError to permit rxi_ResetCall
3330 * to processed when the tqWaiter count hits zero.
3333 rxi_CallError(call, call->error);
3334 MUTEX_EXIT(&call->lock);
3335 putConnection(conn);
3339 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3340 /* If the new call cannot be taken right now send a busy and set
3341 * the error condition in this call, so that it terminates as
3342 * quickly as possible */
3343 if (call->state == RX_STATE_ACTIVE) {
3344 struct rx_packet *tp;
3346 rxi_CallError(call, RX_CALL_DEAD);
3347 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3349 MUTEX_EXIT(&call->lock);
3350 putConnection(conn);
3353 rxi_ResetCall(call, 0);
3355 * The conn_call_lock is not held but no one else should be
3356 * using this call channel while we are processing this incoming
3357 * packet. This assignment should be safe.
3359 *call->callNumber = np->header.callNumber;
3361 if (np->header.callNumber == 0)
3362 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3363 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3364 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3365 np->header.flags, np, np->length));
3367 call->state = RX_STATE_PRECALL;
3368 clock_GetTime(&call->queueTime);
3369 hzero(call->bytesSent);
3370 hzero(call->bytesRcvd);
3372 * If the number of queued calls exceeds the overload
3373 * threshold then abort this call.
3375 if ((rx_BusyThreshold > 0) &&
3376 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3377 struct rx_packet *tp;
3379 rxi_CallError(call, rx_BusyError);
3380 tp = rxi_SendCallAbort(call, np, 1, 0);
3381 MUTEX_EXIT(&call->lock);
3382 putConnection(conn);
3383 if (rx_stats_active)
3384 rx_atomic_inc(&rx_stats.nBusies);
3387 rxi_KeepAliveOn(call);
3389 /* Continuing call; do nothing here. */
3391 } else { /* we're the client */
3392 /* Ignore all incoming acknowledgements for calls in DALLY state */
3393 if ((call->state == RX_STATE_DALLY)
3394 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3395 if (rx_stats_active)
3396 rx_atomic_inc(&rx_stats.ignorePacketDally);
3397 MUTEX_EXIT(&call->lock);
3398 putConnection(conn);
3402 /* Ignore anything that's not relevant to the current call. If there
3403 * isn't a current call, then no packet is relevant. */
3404 if (np->header.callNumber != currentCallNumber) {
3405 if (rx_stats_active)
3406 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3407 MUTEX_EXIT(&call->lock);
3408 putConnection(conn);
3411 /* If the service security object index stamped in the packet does not
3412 * match the connection's security index, ignore the packet */
3413 if (np->header.securityIndex != conn->securityIndex) {
3414 MUTEX_EXIT(&call->lock);
3415 putConnection(conn);
3419 /* If we're receiving the response, then all transmit packets are
3420 * implicitly acknowledged. Get rid of them. */
3421 if (np->header.type == RX_PACKET_TYPE_DATA) {
3422 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3423 /* XXX Hack. Because we must release the global rx lock when
3424 * sending packets (osi_NetSend) we drop all acks while we're
3425 * traversing the tq in rxi_Start sending packets out because
3426 * packets may move to the freePacketQueue as result of being here!
3427 * So we drop these packets until we're safely out of the
3428 * traversing. Really ugly!
3429 * For fine grain RX locking, we set the acked field in the
3430 * packets and let rxi_Start remove them from the transmit queue.
3432 if (call->flags & RX_CALL_TQ_BUSY) {
3433 #ifdef RX_ENABLE_LOCKS
3434 rxi_SetAcksInTransmitQueue(call);
3436 putConnection(conn);
3437 return np; /* xmitting; drop packet */
3440 rxi_ClearTransmitQueue(call, 0);
3442 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3443 rxi_ClearTransmitQueue(call, 0);
3444 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3446 if (np->header.type == RX_PACKET_TYPE_ACK) {
3447 /* now check to see if this is an ack packet acknowledging that the
3448 * server actually *lost* some hard-acked data. If this happens we
3449 * ignore this packet, as it may indicate that the server restarted in
3450 * the middle of a call. It is also possible that this is an old ack
3451 * packet. We don't abort the connection in this case, because this
3452 * *might* just be an old ack packet. The right way to detect a server
3453 * restart in the midst of a call is to notice that the server epoch
3455 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3456 * XXX unacknowledged. I think that this is off-by-one, but
3457 * XXX I don't dare change it just yet, since it will
3458 * XXX interact badly with the server-restart detection
3459 * XXX code in receiveackpacket. */
3460 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3461 if (rx_stats_active)
3462 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3463 MUTEX_EXIT(&call->lock);
3464 putConnection(conn);
3468 } /* else not a data packet */
3471 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3472 /* Set remote user defined status from packet */
3473 call->remoteStatus = np->header.userStatus;
3475 /* Note the gap between the expected next packet and the actual
3476 * packet that arrived, when the new packet has a smaller serial number
3477 * than expected. Rioses frequently reorder packets all by themselves,
3478 * so this will be quite important with very large window sizes.
3479 * Skew is checked against 0 here to avoid any dependence on the type of
3480 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3482 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3483 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3484 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3486 MUTEX_ENTER(&conn->conn_data_lock);
3487 skew = conn->lastSerial - np->header.serial;
3488 conn->lastSerial = np->header.serial;
3489 MUTEX_EXIT(&conn->conn_data_lock);
3491 struct rx_peer *peer;
3493 if (skew > peer->inPacketSkew) {
3494 dpf(("*** In skew changed from %d to %d\n",
3495 peer->inPacketSkew, skew));
3496 peer->inPacketSkew = skew;
3500 /* Now do packet type-specific processing */
3501 switch (np->header.type) {
3502 case RX_PACKET_TYPE_DATA:
3503 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3506 case RX_PACKET_TYPE_ACK:
3507 /* Respond immediately to ack packets requesting acknowledgement
3509 if (np->header.flags & RX_REQUEST_ACK) {
3511 (void)rxi_SendCallAbort(call, 0, 1, 0);
3513 (void)rxi_SendAck(call, 0, np->header.serial,
3514 RX_ACK_PING_RESPONSE, 1);
3516 np = rxi_ReceiveAckPacket(call, np, 1);
3518 case RX_PACKET_TYPE_ABORT: {
3519 /* An abort packet: reset the call, passing the error up to the user. */
3520 /* What if error is zero? */
3521 /* What if the error is -1? the application will treat it as a timeout. */
3522 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3523 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3524 rxi_CallError(call, errdata);
3525 MUTEX_EXIT(&call->lock);
3526 putConnection(conn);
3527 return np; /* xmitting; drop packet */
3529 case RX_PACKET_TYPE_BUSY: {
3530 struct clock busyTime;
3532 clock_GetTime(&busyTime);
3534 MUTEX_EXIT(&call->lock);
3536 MUTEX_ENTER(&conn->conn_call_lock);
3537 MUTEX_ENTER(&call->lock);
3538 conn->lastBusy[call->channel] = busyTime.sec;
3539 call->flags |= RX_CALL_PEER_BUSY;
3540 MUTEX_EXIT(&call->lock);
3541 MUTEX_EXIT(&conn->conn_call_lock);
3543 putConnection(conn);
3547 case RX_PACKET_TYPE_ACKALL:
3548 /* All packets acknowledged, so we can drop all packets previously
3549 * readied for sending */
3550 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3551 /* XXX Hack. We because we can't release the global rx lock when
3552 * sending packets (osi_NetSend) we drop all ack pkts while we're
3553 * traversing the tq in rxi_Start sending packets out because
3554 * packets may move to the freePacketQueue as result of being
3555 * here! So we drop these packets until we're safely out of the
3556 * traversing. Really ugly!
3557 * For fine grain RX locking, we set the acked field in the packets
3558 * and let rxi_Start remove the packets from the transmit queue.
3560 if (call->flags & RX_CALL_TQ_BUSY) {
3561 #ifdef RX_ENABLE_LOCKS
3562 rxi_SetAcksInTransmitQueue(call);
3564 #else /* RX_ENABLE_LOCKS */
3565 MUTEX_EXIT(&call->lock);
3566 putConnection(conn);
3567 return np; /* xmitting; drop packet */
3568 #endif /* RX_ENABLE_LOCKS */
3570 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3571 rxi_ClearTransmitQueue(call, 0);
3574 /* Should not reach here, unless the peer is broken: send an abort
3576 rxi_CallError(call, RX_PROTOCOL_ERROR);
3577 np = rxi_SendCallAbort(call, np, 1, 0);
3580 /* Note when this last legitimate packet was received, for keep-alive
3581 * processing. Note, we delay getting the time until now in the hope that
3582 * the packet will be delivered to the user before any get time is required
3583 * (if not, then the time won't actually be re-evaluated here). */
3584 call->lastReceiveTime = clock_Sec();
3585 /* we've received a legit packet, so the channel is not busy */
3586 call->flags &= ~RX_CALL_PEER_BUSY;
3587 MUTEX_EXIT(&call->lock);
3588 putConnection(conn);
3592 /* return true if this is an "interesting" connection from the point of view
3593 of someone trying to debug the system */
3595 rxi_IsConnInteresting(struct rx_connection *aconn)
3598 struct rx_call *tcall;
3600 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3603 for (i = 0; i < RX_MAXCALLS; i++) {
3604 tcall = aconn->call[i];
3606 if ((tcall->state == RX_STATE_PRECALL)
3607 || (tcall->state == RX_STATE_ACTIVE))
3609 if ((tcall->mode == RX_MODE_SENDING)
3610 || (tcall->mode == RX_MODE_RECEIVING))
3618 /* if this is one of the last few packets AND it wouldn't be used by the
3619 receiving call to immediately satisfy a read request, then drop it on
3620 the floor, since accepting it might prevent a lock-holding thread from
3621 making progress in its reading. If a call has been cleared while in
3622 the precall state then ignore all subsequent packets until the call
3623 is assigned to a thread. */
3626 TooLow(struct rx_packet *ap, struct rx_call *acall)
3630 MUTEX_ENTER(&rx_quota_mutex);
3631 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3632 && (acall->state == RX_STATE_PRECALL))
3633 || ((rx_nFreePackets < rxi_dataQuota + 2)
3634 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3635 && (acall->flags & RX_CALL_READER_WAIT)))) {
3638 MUTEX_EXIT(&rx_quota_mutex);
3644 * Clear the attach wait flag on a connection and proceed.
3646 * Any processing waiting for a connection to be attached should be
3647 * unblocked. We clear the flag and do any other needed tasks.
3650 * the conn to unmark waiting for attach
3652 * @pre conn's conn_data_lock must be locked before calling this function
3656 rxi_ConnClearAttachWait(struct rx_connection *conn)
3658 /* Indicate that rxi_CheckReachEvent is no longer running by
3659 * clearing the flag. Must be atomic under conn_data_lock to
3660 * avoid a new call slipping by: rxi_CheckConnReach holds
3661 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3663 conn->flags &= ~RX_CONN_ATTACHWAIT;
3664 if (conn->flags & RX_CONN_NAT_PING) {
3665 conn->flags &= ~RX_CONN_NAT_PING;
3666 rxi_ScheduleNatKeepAliveEvent(conn);
3671 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3673 struct rx_connection *conn = arg1;
3674 struct rx_call *acall = arg2;
3675 struct rx_call *call = acall;
3676 struct clock when, now;
3679 MUTEX_ENTER(&conn->conn_data_lock);
3682 rxevent_Put(conn->checkReachEvent);
3683 conn->checkReachEvent = NULL;
3686 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3688 putConnection(conn);
3690 MUTEX_EXIT(&conn->conn_data_lock);
3694 MUTEX_ENTER(&conn->conn_call_lock);
3695 MUTEX_ENTER(&conn->conn_data_lock);
3696 for (i = 0; i < RX_MAXCALLS; i++) {
3697 struct rx_call *tc = conn->call[i];
3698 if (tc && tc->state == RX_STATE_PRECALL) {
3704 rxi_ConnClearAttachWait(conn);
3705 MUTEX_EXIT(&conn->conn_data_lock);
3706 MUTEX_EXIT(&conn->conn_call_lock);
3711 MUTEX_ENTER(&call->lock);
3712 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3714 MUTEX_EXIT(&call->lock);
3716 clock_GetTime(&now);
3718 when.sec += RX_CHECKREACH_TIMEOUT;
3719 MUTEX_ENTER(&conn->conn_data_lock);
3720 if (!conn->checkReachEvent) {
3721 MUTEX_ENTER(&rx_refcnt_mutex);
3723 MUTEX_EXIT(&rx_refcnt_mutex);
3724 conn->checkReachEvent = rxevent_Post(&when, &now,
3725 rxi_CheckReachEvent, conn,
3728 MUTEX_EXIT(&conn->conn_data_lock);
3734 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3736 struct rx_service *service = conn->service;
3737 struct rx_peer *peer = conn->peer;
3738 afs_uint32 now, lastReach;
3740 if (service->checkReach == 0)
3744 MUTEX_ENTER(&peer->peer_lock);
3745 lastReach = peer->lastReachTime;
3746 MUTEX_EXIT(&peer->peer_lock);
3747 if (now - lastReach < RX_CHECKREACH_TTL)
3750 MUTEX_ENTER(&conn->conn_data_lock);
3751 if (conn->flags & RX_CONN_ATTACHWAIT) {
3752 MUTEX_EXIT(&conn->conn_data_lock);
3755 conn->flags |= RX_CONN_ATTACHWAIT;
3756 MUTEX_EXIT(&conn->conn_data_lock);
3757 if (!conn->checkReachEvent)
3758 rxi_CheckReachEvent(NULL, conn, call, 0);
3763 /* try to attach call, if authentication is complete */
3765 TryAttach(struct rx_call *acall, osi_socket socket,
3766 int *tnop, struct rx_call **newcallp,
3769 struct rx_connection *conn = acall->conn;
3771 if (conn->type == RX_SERVER_CONNECTION
3772 && acall->state == RX_STATE_PRECALL) {
3773 /* Don't attach until we have any req'd. authentication. */
3774 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3775 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3776 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3777 /* Note: this does not necessarily succeed; there
3778 * may not any proc available
3781 rxi_ChallengeOn(acall->conn);
3786 /* A data packet has been received off the interface. This packet is
3787 * appropriate to the call (the call is in the right state, etc.). This
3788 * routine can return a packet to the caller, for re-use */
3791 rxi_ReceiveDataPacket(struct rx_call *call,
3792 struct rx_packet *np, int istack,
3793 osi_socket socket, afs_uint32 host, u_short port,
3794 int *tnop, struct rx_call **newcallp)
3796 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3801 afs_uint32 serial=0, flags=0;
3803 struct rx_packet *tnp;
3804 if (rx_stats_active)
3805 rx_atomic_inc(&rx_stats.dataPacketsRead);
3808 /* If there are no packet buffers, drop this new packet, unless we can find
3809 * packet buffers from inactive calls */
3811 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3812 MUTEX_ENTER(&rx_freePktQ_lock);
3813 rxi_NeedMorePackets = TRUE;
3814 MUTEX_EXIT(&rx_freePktQ_lock);
3815 if (rx_stats_active)
3816 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3817 call->rprev = np->header.serial;
3818 rxi_calltrace(RX_TRACE_DROP, call);
3819 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3820 /* We used to clear the receive queue here, in an attempt to free
3821 * packets. However this is unsafe if the queue has received a
3822 * soft ACK for the final packet */
3823 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3825 /* we've damaged this call already, might as well do it in. */
3831 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3832 * packet is one of several packets transmitted as a single
3833 * datagram. Do not send any soft or hard acks until all packets
3834 * in a jumbogram have been processed. Send negative acks right away.
3836 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3837 /* tnp is non-null when there are more packets in the
3838 * current jumbo gram */
3845 seq = np->header.seq;
3846 serial = np->header.serial;
3847 flags = np->header.flags;
3849 /* If the call is in an error state, send an abort message */
3851 return rxi_SendCallAbort(call, np, istack, 0);
3853 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3854 * AFS 3.5 jumbogram. */
3855 if (flags & RX_JUMBO_PACKET) {
3856 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3861 if (np->header.spare != 0) {
3862 MUTEX_ENTER(&call->conn->conn_data_lock);
3863 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3864 MUTEX_EXIT(&call->conn->conn_data_lock);
3867 /* The usual case is that this is the expected next packet */
3868 if (seq == call->rnext) {
3870 /* Check to make sure it is not a duplicate of one already queued */
3871 if (queue_IsNotEmpty(&call->rq)
3872 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3873 if (rx_stats_active)
3874 rx_atomic_inc(&rx_stats.dupPacketsRead);
3875 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3876 rxevent_Cancel(&call->delayedAckEvent, call,
3877 RX_CALL_REFCOUNT_DELAY);
3878 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3884 /* It's the next packet. Stick it on the receive queue
3885 * for this call. Set newPackets to make sure we wake
3886 * the reader once all packets have been processed */
3887 #ifdef RX_TRACK_PACKETS
3888 np->flags |= RX_PKTFLAG_RQ;
3890 queue_Prepend(&call->rq, np);
3891 #ifdef RXDEBUG_PACKET
3893 #endif /* RXDEBUG_PACKET */
3895 np = NULL; /* We can't use this anymore */
3898 /* If an ack is requested then set a flag to make sure we
3899 * send an acknowledgement for this packet */
3900 if (flags & RX_REQUEST_ACK) {
3901 ackNeeded = RX_ACK_REQUESTED;
3904 /* Keep track of whether we have received the last packet */
3905 if (flags & RX_LAST_PACKET) {
3906 call->flags |= RX_CALL_HAVE_LAST;
3910 /* Check whether we have all of the packets for this call */
3911 if (call->flags & RX_CALL_HAVE_LAST) {
3912 afs_uint32 tseq; /* temporary sequence number */
3913 struct rx_packet *tp; /* Temporary packet pointer */
3914 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3916 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3917 if (tseq != tp->header.seq)
3919 if (tp->header.flags & RX_LAST_PACKET) {
3920 call->flags |= RX_CALL_RECEIVE_DONE;
3927 /* Provide asynchronous notification for those who want it
3928 * (e.g. multi rx) */
3929 if (call->arrivalProc) {
3930 (*call->arrivalProc) (call, call->arrivalProcHandle,
3931 call->arrivalProcArg);
3932 call->arrivalProc = (void (*)())0;
3935 /* Update last packet received */
3938 /* If there is no server process serving this call, grab
3939 * one, if available. We only need to do this once. If a
3940 * server thread is available, this thread becomes a server
3941 * thread and the server thread becomes a listener thread. */
3943 TryAttach(call, socket, tnop, newcallp, 0);
3946 /* This is not the expected next packet. */
3948 /* Determine whether this is a new or old packet, and if it's
3949 * a new one, whether it fits into the current receive window.
3950 * Also figure out whether the packet was delivered in sequence.
3951 * We use the prev variable to determine whether the new packet
3952 * is the successor of its immediate predecessor in the
3953 * receive queue, and the missing flag to determine whether
3954 * any of this packets predecessors are missing. */
3956 afs_uint32 prev; /* "Previous packet" sequence number */
3957 struct rx_packet *tp; /* Temporary packet pointer */
3958 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3959 int missing; /* Are any predecessors missing? */
3961 /* If the new packet's sequence number has been sent to the
3962 * application already, then this is a duplicate */
3963 if (seq < call->rnext) {
3964 if (rx_stats_active)
3965 rx_atomic_inc(&rx_stats.dupPacketsRead);
3966 rxevent_Cancel(&call->delayedAckEvent, call,
3967 RX_CALL_REFCOUNT_DELAY);
3968 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3974 /* If the sequence number is greater than what can be
3975 * accomodated by the current window, then send a negative
3976 * acknowledge and drop the packet */
3977 if ((call->rnext + call->rwind) <= seq) {
3978 rxevent_Cancel(&call->delayedAckEvent, call,
3979 RX_CALL_REFCOUNT_DELAY);
3980 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3987 /* Look for the packet in the queue of old received packets */
3988 for (prev = call->rnext - 1, missing =
3989 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3990 /*Check for duplicate packet */
3991 if (seq == tp->header.seq) {
3992 if (rx_stats_active)
3993 rx_atomic_inc(&rx_stats.dupPacketsRead);
3994 rxevent_Cancel(&call->delayedAckEvent, call,
3995 RX_CALL_REFCOUNT_DELAY);
3996 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
4002 /* If we find a higher sequence packet, break out and
4003 * insert the new packet here. */
4004 if (seq < tp->header.seq)
4006 /* Check for missing packet */
4007 if (tp->header.seq != prev + 1) {
4011 prev = tp->header.seq;
4014 /* Keep track of whether we have received the last packet. */
4015 if (flags & RX_LAST_PACKET) {
4016 call->flags |= RX_CALL_HAVE_LAST;
4019 /* It's within the window: add it to the the receive queue.
4020 * tp is left by the previous loop either pointing at the
4021 * packet before which to insert the new packet, or at the
4022 * queue head if the queue is empty or the packet should be
4024 #ifdef RX_TRACK_PACKETS
4025 np->flags |= RX_PKTFLAG_RQ;
4027 #ifdef RXDEBUG_PACKET
4029 #endif /* RXDEBUG_PACKET */
4030 queue_InsertBefore(tp, np);
4034 /* Check whether we have all of the packets for this call */
4035 if ((call->flags & RX_CALL_HAVE_LAST)
4036 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4037 afs_uint32 tseq; /* temporary sequence number */
4040 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
4041 if (tseq != tp->header.seq)
4043 if (tp->header.flags & RX_LAST_PACKET) {
4044 call->flags |= RX_CALL_RECEIVE_DONE;
4051 /* We need to send an ack of the packet is out of sequence,
4052 * or if an ack was requested by the peer. */
4053 if (seq != prev + 1 || missing) {
4054 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4055 } else if (flags & RX_REQUEST_ACK) {
4056 ackNeeded = RX_ACK_REQUESTED;
4059 /* Acknowledge the last packet for each call */
4060 if (flags & RX_LAST_PACKET) {
4071 * If the receiver is waiting for an iovec, fill the iovec
4072 * using the data from the receive queue */
4073 if (call->flags & RX_CALL_IOVEC_WAIT) {
4074 didHardAck = rxi_FillReadVec(call, serial);
4075 /* the call may have been aborted */
4084 /* Wakeup the reader if any */
4085 if ((call->flags & RX_CALL_READER_WAIT)
4086 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4087 || (call->iovNext >= call->iovMax)
4088 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4089 call->flags &= ~RX_CALL_READER_WAIT;
4090 #ifdef RX_ENABLE_LOCKS
4091 CV_BROADCAST(&call->cv_rq);
4093 osi_rxWakeup(&call->rq);
4099 * Send an ack when requested by the peer, or once every
4100 * rxi_SoftAckRate packets until the last packet has been
4101 * received. Always send a soft ack for the last packet in
4102 * the server's reply. */
4104 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4105 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4106 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4107 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4108 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4109 } else if (call->nSoftAcks) {
4110 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4111 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4113 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4114 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4115 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4122 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4124 struct rx_peer *peer = conn->peer;
4126 MUTEX_ENTER(&peer->peer_lock);
4127 peer->lastReachTime = clock_Sec();
4128 MUTEX_EXIT(&peer->peer_lock);
4130 MUTEX_ENTER(&conn->conn_data_lock);
4131 if (conn->flags & RX_CONN_ATTACHWAIT) {
4134 rxi_ConnClearAttachWait(conn);
4135 MUTEX_EXIT(&conn->conn_data_lock);
4137 for (i = 0; i < RX_MAXCALLS; i++) {
4138 struct rx_call *call = conn->call[i];
4141 MUTEX_ENTER(&call->lock);
4142 /* tnop can be null if newcallp is null */
4143 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4145 MUTEX_EXIT(&call->lock);
4149 MUTEX_EXIT(&conn->conn_data_lock);
4152 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4154 rx_ack_reason(int reason)
4157 case RX_ACK_REQUESTED:
4159 case RX_ACK_DUPLICATE:
4161 case RX_ACK_OUT_OF_SEQUENCE:
4163 case RX_ACK_EXCEEDS_WINDOW:
4165 case RX_ACK_NOSPACE:
4169 case RX_ACK_PING_RESPONSE:
4182 /* The real smarts of the whole thing. */
4184 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4187 struct rx_ackPacket *ap;
4189 struct rx_packet *tp;
4190 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
4191 struct rx_connection *conn = call->conn;
4192 struct rx_peer *peer = conn->peer;
4193 struct clock now; /* Current time, for RTT calculations */
4197 /* because there are CM's that are bogus, sending weird values for this. */
4198 afs_uint32 skew = 0;
4203 int newAckCount = 0;
4204 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4205 int pktsize = 0; /* Set if we need to update the peer mtu */
4206 int conn_data_locked = 0;
4208 if (rx_stats_active)
4209 rx_atomic_inc(&rx_stats.ackPacketsRead);
4210 ap = (struct rx_ackPacket *)rx_DataOf(np);
4211 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4213 return np; /* truncated ack packet */
4215 /* depends on ack packet struct */
4216 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4217 first = ntohl(ap->firstPacket);
4218 prev = ntohl(ap->previousPacket);
4219 serial = ntohl(ap->serial);
4220 /* temporarily disabled -- needs to degrade over time
4221 * skew = ntohs(ap->maxSkew); */
4223 /* Ignore ack packets received out of order */
4224 if (first < call->tfirst ||
4225 (first == call->tfirst && prev < call->tprev)) {
4231 if (np->header.flags & RX_SLOW_START_OK) {
4232 call->flags |= RX_CALL_SLOW_START_OK;
4235 if (ap->reason == RX_ACK_PING_RESPONSE)
4236 rxi_UpdatePeerReach(conn, call);
4238 if (conn->lastPacketSizeSeq) {
4239 MUTEX_ENTER(&conn->conn_data_lock);
4240 conn_data_locked = 1;
4241 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4242 pktsize = conn->lastPacketSize;
4243 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4246 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4247 if (!conn_data_locked) {
4248 MUTEX_ENTER(&conn->conn_data_lock);
4249 conn_data_locked = 1;
4251 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4252 /* process mtu ping ack */
4253 pktsize = conn->lastPingSize;
4254 conn->lastPingSizeSer = conn->lastPingSize = 0;
4258 if (conn_data_locked) {
4259 MUTEX_EXIT(&conn->conn_data_lock);
4260 conn_data_locked = 0;
4264 if (rxdebug_active) {
4268 len = _snprintf(msg, sizeof(msg),
4269 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
4270 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4271 ntohl(ap->serial), ntohl(ap->previousPacket),
4272 (unsigned int)np->header.seq, (unsigned int)skew,
4273 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
4277 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4278 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4282 OutputDebugString(msg);
4284 #else /* AFS_NT40_ENV */
4287 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
4288 ap->reason, ntohl(ap->previousPacket),
4289 (unsigned int)np->header.seq, (unsigned int)serial,
4290 (unsigned int)skew, ntohl(ap->firstPacket));
4293 for (offset = 0; offset < nAcks; offset++)
4294 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4299 #endif /* AFS_NT40_ENV */
4302 MUTEX_ENTER(&peer->peer_lock);
4305 * Start somewhere. Can't assume we can send what we can receive,
4306 * but we are clearly receiving.
4308 if (!peer->maxPacketSize)
4309 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4311 if (pktsize > peer->maxPacketSize) {
4312 peer->maxPacketSize = pktsize;
4313 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4314 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4315 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4316 rxi_ScheduleGrowMTUEvent(call, 1);
4321 /* Update the outgoing packet skew value to the latest value of
4322 * the peer's incoming packet skew value. The ack packet, of
4323 * course, could arrive out of order, but that won't affect things
4325 peer->outPacketSkew = skew;
4328 clock_GetTime(&now);
4330 /* The transmit queue splits into 4 sections.
4332 * The first section is packets which have now been acknowledged
4333 * by a window size change in the ack. These have reached the
4334 * application layer, and may be discarded. These are packets
4335 * with sequence numbers < ap->firstPacket.
4337 * The second section is packets which have sequence numbers in
4338 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4339 * contents of the packet's ack array determines whether these
4340 * packets are acknowledged or not.
4342 * The third section is packets which fall above the range
4343 * addressed in the ack packet. These have not yet been received
4346 * The four section is packets which have not yet been transmitted.
4347 * These packets will have a header.serial of 0.
4350 /* First section - implicitly acknowledged packets that can be
4354 tp = queue_First(&call->tq, rx_packet);
4355 while(!queue_IsEnd(&call->tq, tp) && tp->header.seq < first) {
4356 struct rx_packet *next;
4358 next = queue_Next(tp, rx_packet);
4359 call->tfirst = tp->header.seq + 1;
4361 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4363 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4366 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4367 /* XXX Hack. Because we have to release the global rx lock when sending
4368 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4369 * in rxi_Start sending packets out because packets may move to the
4370 * freePacketQueue as result of being here! So we drop these packets until
4371 * we're safely out of the traversing. Really ugly!
4372 * To make it even uglier, if we're using fine grain locking, we can
4373 * set the ack bits in the packets and have rxi_Start remove the packets
4374 * when it's done transmitting.
4376 if (call->flags & RX_CALL_TQ_BUSY) {
4377 #ifdef RX_ENABLE_LOCKS
4378 tp->flags |= RX_PKTFLAG_ACKED;
4379 call->flags |= RX_CALL_TQ_SOME_ACKED;
4380 #else /* RX_ENABLE_LOCKS */
4382 #endif /* RX_ENABLE_LOCKS */
4384 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4387 #ifdef RX_TRACK_PACKETS
4388 tp->flags &= ~RX_PKTFLAG_TQ;
4390 #ifdef RXDEBUG_PACKET
4392 #endif /* RXDEBUG_PACKET */
4393 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4398 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4400 /* Second section of the queue - packets for which we are receiving
4403 * Go through the explicit acks/nacks and record the results in
4404 * the waiting packets. These are packets that can't be released
4405 * yet, even with a positive acknowledge. This positive
4406 * acknowledge only means the packet has been received by the
4407 * peer, not that it will be retained long enough to be sent to
4408 * the peer's upper level. In addition, reset the transmit timers
4409 * of any missing packets (those packets that must be missing
4410 * because this packet was out of sequence) */
4412 call->nSoftAcked = 0;
4414 while (!queue_IsEnd(&call->tq, tp) && tp->header.seq < first + nAcks) {
4415 /* Set the acknowledge flag per packet based on the
4416 * information in the ack packet. An acknowlegded packet can
4417 * be downgraded when the server has discarded a packet it
4418 * soacked previously, or when an ack packet is received
4419 * out of sequence. */
4420 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4421 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4423 tp->flags |= RX_PKTFLAG_ACKED;
4424 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4431 } else /* RX_ACK_TYPE_NACK */ {
4432 tp->flags &= ~RX_PKTFLAG_ACKED;
4436 tp = queue_Next(tp, rx_packet);
4439 /* We don't need to take any action with the 3rd or 4th section in the
4440 * queue - they're not addressed by the contents of this ACK packet.
4443 /* If the window has been extended by this acknowledge packet,
4444 * then wakeup a sender waiting in alloc for window space, or try
4445 * sending packets now, if he's been sitting on packets due to
4446 * lack of window space */
4447 if (call->tnext < (call->tfirst + call->twind)) {
4448 #ifdef RX_ENABLE_LOCKS
4449 CV_SIGNAL(&call->cv_twind);
4451 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4452 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4453 osi_rxWakeup(&call->twind);
4456 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4457 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4461 /* if the ack packet has a receivelen field hanging off it,
4462 * update our state */
4463 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4466 /* If the ack packet has a "recommended" size that is less than
4467 * what I am using now, reduce my size to match */
4468 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4469 (int)sizeof(afs_int32), &tSize);
4470 tSize = (afs_uint32) ntohl(tSize);
4471 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4473 /* Get the maximum packet size to send to this peer */
4474 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4476 tSize = (afs_uint32) ntohl(tSize);
4477 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4478 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4480 /* sanity check - peer might have restarted with different params.
4481 * If peer says "send less", dammit, send less... Peer should never
4482 * be unable to accept packets of the size that prior AFS versions would
4483 * send without asking. */
4484 if (peer->maxMTU != tSize) {
4485 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4487 peer->maxMTU = tSize;
4488 peer->MTU = MIN(tSize, peer->MTU);
4489 call->MTU = MIN(call->MTU, tSize);
4492 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4495 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4496 (int)sizeof(afs_int32), &tSize);
4497 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4498 if (tSize < call->twind) { /* smaller than our send */
4499 call->twind = tSize; /* window, we must send less... */
4500 call->ssthresh = MIN(call->twind, call->ssthresh);
4501 call->conn->twind[call->channel] = call->twind;
4504 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4505 * network MTU confused with the loopback MTU. Calculate the
4506 * maximum MTU here for use in the slow start code below.
4508 /* Did peer restart with older RX version? */
4509 if (peer->maxDgramPackets > 1) {
4510 peer->maxDgramPackets = 1;
4512 } else if (np->length >=
4513 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4516 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4517 sizeof(afs_int32), &tSize);
4518 tSize = (afs_uint32) ntohl(tSize);
4520 * As of AFS 3.5 we set the send window to match the receive window.
4522 if (tSize < call->twind) {
4523 call->twind = tSize;
4524 call->conn->twind[call->channel] = call->twind;
4525 call->ssthresh = MIN(call->twind, call->ssthresh);
4526 } else if (tSize > call->twind) {
4527 call->twind = tSize;
4528 call->conn->twind[call->channel] = call->twind;
4532 * As of AFS 3.5, a jumbogram is more than one fixed size
4533 * packet transmitted in a single UDP datagram. If the remote
4534 * MTU is smaller than our local MTU then never send a datagram
4535 * larger than the natural MTU.
4538 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4539 (int)sizeof(afs_int32), &tSize);
4540 maxDgramPackets = (afs_uint32) ntohl(tSize);
4541 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4543 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4544 if (maxDgramPackets > 1) {
4545 peer->maxDgramPackets = maxDgramPackets;
4546 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4548 peer->maxDgramPackets = 1;
4549 call->MTU = peer->natMTU;
4551 } else if (peer->maxDgramPackets > 1) {
4552 /* Restarted with lower version of RX */
4553 peer->maxDgramPackets = 1;
4555 } else if (peer->maxDgramPackets > 1
4556 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4557 /* Restarted with lower version of RX */
4558 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4559 peer->natMTU = OLD_MAX_PACKET_SIZE;
4560 peer->MTU = OLD_MAX_PACKET_SIZE;
4561 peer->maxDgramPackets = 1;
4562 peer->nDgramPackets = 1;
4564 call->MTU = OLD_MAX_PACKET_SIZE;
4569 * Calculate how many datagrams were successfully received after
4570 * the first missing packet and adjust the negative ack counter
4575 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4576 if (call->nNacks < nNacked) {
4577 call->nNacks = nNacked;
4580 call->nAcks += newAckCount;
4584 /* If the packet contained new acknowledgements, rather than just
4585 * being a duplicate of one we have previously seen, then we can restart
4588 if (newAckCount > 0)
4589 rxi_rto_packet_acked(call, istack);
4591 if (call->flags & RX_CALL_FAST_RECOVER) {
4592 if (newAckCount == 0) {
4593 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4595 call->flags &= ~RX_CALL_FAST_RECOVER;
4596 call->cwind = call->nextCwind;
4597 call->nextCwind = 0;
4600 call->nCwindAcks = 0;
4601 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4602 /* Three negative acks in a row trigger congestion recovery */
4603 call->flags |= RX_CALL_FAST_RECOVER;
4604 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4606 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4607 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4608 call->nextCwind = call->ssthresh;
4611 peer->MTU = call->MTU;
4612 peer->cwind = call->nextCwind;
4613 peer->nDgramPackets = call->nDgramPackets;
4615 call->congestSeq = peer->congestSeq;
4617 /* Reset the resend times on the packets that were nacked
4618 * so we will retransmit as soon as the window permits
4621 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4623 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4624 tp->flags &= ~RX_PKTFLAG_SENT;
4626 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4631 /* If cwind is smaller than ssthresh, then increase
4632 * the window one packet for each ack we receive (exponential
4634 * If cwind is greater than or equal to ssthresh then increase
4635 * the congestion window by one packet for each cwind acks we
4636 * receive (linear growth). */
4637 if (call->cwind < call->ssthresh) {
4639 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4640 call->nCwindAcks = 0;
4642 call->nCwindAcks += newAckCount;
4643 if (call->nCwindAcks >= call->cwind) {
4644 call->nCwindAcks = 0;
4645 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4649 * If we have received several acknowledgements in a row then
4650 * it is time to increase the size of our datagrams
4652 if ((int)call->nAcks > rx_nDgramThreshold) {
4653 if (peer->maxDgramPackets > 1) {
4654 if (call->nDgramPackets < peer->maxDgramPackets) {
4655 call->nDgramPackets++;
4657 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4658 } else if (call->MTU < peer->maxMTU) {
4659 /* don't upgrade if we can't handle it */
4660 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4661 call->MTU = peer->ifMTU;
4663 call->MTU += peer->natMTU;
4664 call->MTU = MIN(call->MTU, peer->maxMTU);
4671 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4673 /* Servers need to hold the call until all response packets have
4674 * been acknowledged. Soft acks are good enough since clients
4675 * are not allowed to clear their receive queues. */
4676 if (call->state == RX_STATE_HOLD
4677 && call->tfirst + call->nSoftAcked >= call->tnext) {
4678 call->state = RX_STATE_DALLY;
4679 rxi_ClearTransmitQueue(call, 0);
4680 rxevent_Cancel(&call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4681 } else if (!queue_IsEmpty(&call->tq)) {
4682 rxi_Start(call, istack);
4687 /* Received a response to a challenge packet */
4689 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4690 struct rx_packet *np, int istack)
4694 /* Ignore the packet if we're the client */
4695 if (conn->type == RX_CLIENT_CONNECTION)
4698 /* If already authenticated, ignore the packet (it's probably a retry) */
4699 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4702 /* Otherwise, have the security object evaluate the response packet */
4703 error = RXS_CheckResponse(conn->securityObject, conn, np);
4705 /* If the response is invalid, reset the connection, sending
4706 * an abort to the peer */
4710 rxi_ConnectionError(conn, error);
4711 MUTEX_ENTER(&conn->conn_data_lock);
4712 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4713 MUTEX_EXIT(&conn->conn_data_lock);
4716 /* If the response is valid, any calls waiting to attach
4717 * servers can now do so */
4720 for (i = 0; i < RX_MAXCALLS; i++) {
4721 struct rx_call *call = conn->call[i];
4723 MUTEX_ENTER(&call->lock);
4724 if (call->state == RX_STATE_PRECALL)
4725 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4726 /* tnop can be null if newcallp is null */
4727 MUTEX_EXIT(&call->lock);
4731 /* Update the peer reachability information, just in case
4732 * some calls went into attach-wait while we were waiting
4733 * for authentication..
4735 rxi_UpdatePeerReach(conn, NULL);
4740 /* A client has received an authentication challenge: the security
4741 * object is asked to cough up a respectable response packet to send
4742 * back to the server. The server is responsible for retrying the
4743 * challenge if it fails to get a response. */
4746 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4747 struct rx_packet *np, int istack)
4751 /* Ignore the challenge if we're the server */
4752 if (conn->type == RX_SERVER_CONNECTION)
4755 /* Ignore the challenge if the connection is otherwise idle; someone's
4756 * trying to use us as an oracle. */
4757 if (!rxi_HasActiveCalls(conn))
4760 /* Send the security object the challenge packet. It is expected to fill
4761 * in the response. */
4762 error = RXS_GetResponse(conn->securityObject, conn, np);
4764 /* If the security object is unable to return a valid response, reset the
4765 * connection and send an abort to the peer. Otherwise send the response
4766 * packet to the peer connection. */
4768 rxi_ConnectionError(conn, error);
4769 MUTEX_ENTER(&conn->conn_data_lock);
4770 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4771 MUTEX_EXIT(&conn->conn_data_lock);
4773 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4774 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4780 /* Find an available server process to service the current request in
4781 * the given call structure. If one isn't available, queue up this
4782 * call so it eventually gets one */
4784 rxi_AttachServerProc(struct rx_call *call,
4785 osi_socket socket, int *tnop,
4786 struct rx_call **newcallp)
4788 struct rx_serverQueueEntry *sq;
4789 struct rx_service *service = call->conn->service;
4792 /* May already be attached */
4793 if (call->state == RX_STATE_ACTIVE)
4796 MUTEX_ENTER(&rx_serverPool_lock);
4798 haveQuota = QuotaOK(service);
4799 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4800 /* If there are no processes available to service this call,
4801 * put the call on the incoming call queue (unless it's
4802 * already on the queue).
4804 #ifdef RX_ENABLE_LOCKS
4806 ReturnToServerPool(service);
4807 #endif /* RX_ENABLE_LOCKS */
4809 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4810 call->flags |= RX_CALL_WAIT_PROC;
4811 rx_atomic_inc(&rx_nWaiting);
4812 rx_atomic_inc(&rx_nWaited);
4813 rxi_calltrace(RX_CALL_ARRIVAL, call);
4814 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4815 queue_Append(&rx_incomingCallQueue, call);
4818 sq = queue_Last(&rx_idleServerQueue, rx_serverQueueEntry);
4820 /* If hot threads are enabled, and both newcallp and sq->socketp
4821 * are non-null, then this thread will process the call, and the
4822 * idle server thread will start listening on this threads socket.
4825 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4828 *sq->socketp = socket;
4829 clock_GetTime(&call->startTime);
4830 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4834 if (call->flags & RX_CALL_WAIT_PROC) {
4835 /* Conservative: I don't think this should happen */
4836 call->flags &= ~RX_CALL_WAIT_PROC;
4837 if (queue_IsOnQueue(call)) {
4840 rx_atomic_dec(&rx_nWaiting);
4843 call->state = RX_STATE_ACTIVE;
4844 call->mode = RX_MODE_RECEIVING;
4845 #ifdef RX_KERNEL_TRACE
4847 int glockOwner = ISAFS_GLOCK();
4850 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4851 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4857 if (call->flags & RX_CALL_CLEARED) {
4858 /* send an ack now to start the packet flow up again */
4859 call->flags &= ~RX_CALL_CLEARED;
4860 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4862 #ifdef RX_ENABLE_LOCKS
4865 service->nRequestsRunning++;
4866 MUTEX_ENTER(&rx_quota_mutex);
4867 if (service->nRequestsRunning <= service->minProcs)
4870 MUTEX_EXIT(&rx_quota_mutex);
4874 MUTEX_EXIT(&rx_serverPool_lock);
4877 /* Delay the sending of an acknowledge event for a short while, while
4878 * a new call is being prepared (in the case of a client) or a reply
4879 * is being prepared (in the case of a server). Rather than sending
4880 * an ack packet, an ACKALL packet is sent. */
4882 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4884 #ifdef RX_ENABLE_LOCKS
4886 MUTEX_ENTER(&call->lock);
4887 rxevent_Put(call->delayedAckEvent);
4888 call->delayedAckEvent = NULL;
4889 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4891 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4892 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4893 call->flags |= RX_CALL_ACKALL_SENT;
4895 MUTEX_EXIT(&call->lock);
4896 #else /* RX_ENABLE_LOCKS */
4898 rxevent_Put(call->delayedAckEvent);
4899 call->delayedAckEvent = NULL;
4901 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4902 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4903 call->flags |= RX_CALL_ACKALL_SENT;
4904 #endif /* RX_ENABLE_LOCKS */
4908 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4911 struct rx_call *call = arg1;
4912 #ifdef RX_ENABLE_LOCKS
4914 MUTEX_ENTER(&call->lock);
4915 if (event == call->delayedAckEvent) {
4916 rxevent_Put(call->delayedAckEvent);
4917 call->delayedAckEvent = NULL;
4919 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4921 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4923 MUTEX_EXIT(&call->lock);
4924 #else /* RX_ENABLE_LOCKS */
4926 rxevent_Put(call->delayedAckEvent);
4927 call->delayedAckEvent = NULL;
4929 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4930 #endif /* RX_ENABLE_LOCKS */
4934 #ifdef RX_ENABLE_LOCKS
4935 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4936 * clearing them out.
4939 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4941 struct rx_packet *p, *tp;
4944 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4945 p->flags |= RX_PKTFLAG_ACKED;
4949 call->flags |= RX_CALL_TQ_CLEARME;
4950 call->flags |= RX_CALL_TQ_SOME_ACKED;
4953 rxi_rto_cancel(call);
4955 call->tfirst = call->tnext;
4956 call->nSoftAcked = 0;
4958 if (call->flags & RX_CALL_FAST_RECOVER) {
4959 call->flags &= ~RX_CALL_FAST_RECOVER;
4960 call->cwind = call->nextCwind;
4961 call->nextCwind = 0;
4964 CV_SIGNAL(&call->cv_twind);
4966 #endif /* RX_ENABLE_LOCKS */
4968 /* Clear out the transmit queue for the current call (all packets have
4969 * been received by peer) */
4971 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4973 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4974 struct rx_packet *p, *tp;
4976 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4978 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4979 p->flags |= RX_PKTFLAG_ACKED;
4983 call->flags |= RX_CALL_TQ_CLEARME;
4984 call->flags |= RX_CALL_TQ_SOME_ACKED;
4987 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4988 #ifdef RXDEBUG_PACKET
4990 #endif /* RXDEBUG_PACKET */
4991 rxi_FreePackets(0, &call->tq);
4992 rxi_WakeUpTransmitQueue(call);
4993 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4994 call->flags &= ~RX_CALL_TQ_CLEARME;
4996 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4998 rxi_rto_cancel(call);
4999 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5000 call->nSoftAcked = 0;
5002 if (call->flags & RX_CALL_FAST_RECOVER) {
5003 call->flags &= ~RX_CALL_FAST_RECOVER;
5004 call->cwind = call->nextCwind;
5006 #ifdef RX_ENABLE_LOCKS
5007 CV_SIGNAL(&call->cv_twind);
5009 osi_rxWakeup(&call->twind);
5014 rxi_ClearReceiveQueue(struct rx_call *call)
5016 if (queue_IsNotEmpty(&call->rq)) {
5019 count = rxi_FreePackets(0, &call->rq);
5020 rx_packetReclaims += count;
5021 #ifdef RXDEBUG_PACKET
5023 if ( call->rqc != 0 )
5024 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5026 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5028 if (call->state == RX_STATE_PRECALL) {
5029 call->flags |= RX_CALL_CLEARED;
5033 /* Send an abort packet for the specified call */
5035 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5036 int istack, int force)
5038 afs_int32 error, cerror;
5039 struct clock when, now;
5044 switch (call->error) {
5047 cerror = RX_CALL_TIMEOUT;
5050 cerror = call->error;
5053 /* Clients should never delay abort messages */
5054 if (rx_IsClientConn(call->conn))
5057 if (call->abortCode != cerror) {
5058 call->abortCode = cerror;
5059 call->abortCount = 0;
5062 if (force || rxi_callAbortThreshhold == 0
5063 || call->abortCount < rxi_callAbortThreshhold) {
5064 if (call->delayedAbortEvent) {
5065 rxevent_Cancel(&call->delayedAbortEvent, call,
5066 RX_CALL_REFCOUNT_ABORT);
5068 error = htonl(cerror);
5071 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5072 (char *)&error, sizeof(error), istack);
5073 } else if (!call->delayedAbortEvent) {
5074 clock_GetTime(&now);
5076 clock_Addmsec(&when, rxi_callAbortDelay);
5077 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5078 call->delayedAbortEvent =
5079 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5084 /* Send an abort packet for the specified connection. Packet is an
5085 * optional pointer to a packet that can be used to send the abort.
5086 * Once the number of abort messages reaches the threshhold, an
5087 * event is scheduled to send the abort. Setting the force flag
5088 * overrides sending delayed abort messages.
5090 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5091 * to send the abort packet.
5094 rxi_SendConnectionAbort(struct rx_connection *conn,
5095 struct rx_packet *packet, int istack, int force)
5098 struct clock when, now;
5103 /* Clients should never delay abort messages */
5104 if (rx_IsClientConn(conn))
5107 if (force || rxi_connAbortThreshhold == 0
5108 || conn->abortCount < rxi_connAbortThreshhold) {
5110 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
5111 error = htonl(conn->error);
5113 MUTEX_EXIT(&conn->conn_data_lock);
5115 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5116 RX_PACKET_TYPE_ABORT, (char *)&error,
5117 sizeof(error), istack);
5118 MUTEX_ENTER(&conn->conn_data_lock);
5119 } else if (!conn->delayedAbortEvent) {
5120 clock_GetTime(&now);
5122 clock_Addmsec(&when, rxi_connAbortDelay);
5123 conn->delayedAbortEvent =
5124 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
5129 /* Associate an error all of the calls owned by a connection. Called
5130 * with error non-zero. This is only for really fatal things, like
5131 * bad authentication responses. The connection itself is set in
5132 * error at this point, so that future packets received will be
5135 rxi_ConnectionError(struct rx_connection *conn,
5141 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5143 MUTEX_ENTER(&conn->conn_data_lock);
5144 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
5145 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
5146 if (conn->checkReachEvent) {
5147 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
5148 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5149 putConnection(conn);
5151 MUTEX_EXIT(&conn->conn_data_lock);
5152 for (i = 0; i < RX_MAXCALLS; i++) {
5153 struct rx_call *call = conn->call[i];
5155 MUTEX_ENTER(&call->lock);
5156 rxi_CallError(call, error);
5157 MUTEX_EXIT(&call->lock);
5160 conn->error = error;
5161 if (rx_stats_active)
5162 rx_atomic_inc(&rx_stats.fatalErrors);
5167 * Interrupt an in-progress call with the specified error and wakeup waiters.
5169 * @param[in] call The call to interrupt
5170 * @param[in] error The error code to send to the peer
5173 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5175 MUTEX_ENTER(&call->lock);
5176 rxi_CallError(call, error);
5177 rxi_SendCallAbort(call, NULL, 0, 1);
5178 MUTEX_EXIT(&call->lock);
5182 rxi_CallError(struct rx_call *call, afs_int32 error)
5185 osirx_AssertMine(&call->lock, "rxi_CallError");
5187 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5189 error = call->error;
5191 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5192 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5193 rxi_ResetCall(call, 0);
5196 rxi_ResetCall(call, 0);
5198 call->error = error;
5201 /* Reset various fields in a call structure, and wakeup waiting
5202 * processes. Some fields aren't changed: state & mode are not
5203 * touched (these must be set by the caller), and bufptr, nLeft, and
5204 * nFree are not reset, since these fields are manipulated by
5205 * unprotected macros, and may only be reset by non-interrupting code.
5209 rxi_ResetCall(struct rx_call *call, int newcall)
5212 struct rx_peer *peer;
5213 struct rx_packet *packet;
5215 osirx_AssertMine(&call->lock, "rxi_ResetCall");
5217 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5219 /* Notify anyone who is waiting for asynchronous packet arrival */
5220 if (call->arrivalProc) {
5221 (*call->arrivalProc) (call, call->arrivalProcHandle,
5222 call->arrivalProcArg);
5223 call->arrivalProc = (void (*)())0;
5227 rxevent_Cancel(&call->growMTUEvent, call, RX_CALL_REFCOUNT_MTU);
5229 if (call->delayedAbortEvent) {
5230 rxevent_Cancel(&call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
5231 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5233 rxi_SendCallAbort(call, packet, 0, 1);
5234 rxi_FreePacket(packet);
5239 * Update the peer with the congestion information in this call
5240 * so other calls on this connection can pick up where this call
5241 * left off. If the congestion sequence numbers don't match then
5242 * another call experienced a retransmission.
5244 peer = call->conn->peer;
5245 MUTEX_ENTER(&peer->peer_lock);
5247 if (call->congestSeq == peer->congestSeq) {
5248 peer->cwind = MAX(peer->cwind, call->cwind);
5249 peer->MTU = MAX(peer->MTU, call->MTU);
5250 peer->nDgramPackets =
5251 MAX(peer->nDgramPackets, call->nDgramPackets);
5254 call->abortCode = 0;
5255 call->abortCount = 0;
5257 if (peer->maxDgramPackets > 1) {
5258 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5260 call->MTU = peer->MTU;
5262 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5263 call->ssthresh = rx_maxSendWindow;
5264 call->nDgramPackets = peer->nDgramPackets;
5265 call->congestSeq = peer->congestSeq;
5266 call->rtt = peer->rtt;
5267 call->rtt_dev = peer->rtt_dev;
5268 clock_Zero(&call->rto);
5269 clock_Addmsec(&call->rto,
5270 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5271 MUTEX_EXIT(&peer->peer_lock);
5273 flags = call->flags;
5274 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5275 rxi_WaitforTQBusy(call);
5276 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5278 rxi_ClearTransmitQueue(call, 1);
5279 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5280 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5284 if (!newcall && (flags & RX_CALL_PEER_BUSY)) {
5285 /* The call channel is still busy; resetting the call doesn't change
5286 * that. However, if 'newcall' is set, we are processing a call
5287 * structure that has either been recycled from the free list, or has
5288 * been newly allocated. So, RX_CALL_PEER_BUSY is not relevant if
5289 * 'newcall' is set, since it describes a completely different call
5290 * channel which we do not care about. */
5291 call->flags |= RX_CALL_PEER_BUSY;
5294 rxi_ClearReceiveQueue(call);
5295 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5299 call->twind = call->conn->twind[call->channel];
5300 call->rwind = call->conn->rwind[call->channel];
5301 call->nSoftAcked = 0;
5302 call->nextCwind = 0;
5305 call->nCwindAcks = 0;
5306 call->nSoftAcks = 0;
5307 call->nHardAcks = 0;
5309 call->tfirst = call->rnext = call->tnext = 1;
5312 call->lastAcked = 0;
5313 call->localStatus = call->remoteStatus = 0;
5315 if (flags & RX_CALL_READER_WAIT) {
5316 #ifdef RX_ENABLE_LOCKS
5317 CV_BROADCAST(&call->cv_rq);
5319 osi_rxWakeup(&call->rq);
5322 if (flags & RX_CALL_WAIT_PACKETS) {
5323 MUTEX_ENTER(&rx_freePktQ_lock);
5324 rxi_PacketsUnWait(); /* XXX */
5325 MUTEX_EXIT(&rx_freePktQ_lock);
5327 #ifdef RX_ENABLE_LOCKS
5328 CV_SIGNAL(&call->cv_twind);
5330 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5331 osi_rxWakeup(&call->twind);
5334 #ifdef RX_ENABLE_LOCKS
5335 /* The following ensures that we don't mess with any queue while some
5336 * other thread might also be doing so. The call_queue_lock field is
5337 * is only modified under the call lock. If the call is in the process
5338 * of being removed from a queue, the call is not locked until the
5339 * the queue lock is dropped and only then is the call_queue_lock field
5340 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5341 * Note that any other routine which removes a call from a queue has to
5342 * obtain the queue lock before examing the queue and removing the call.
5344 if (call->call_queue_lock) {
5345 MUTEX_ENTER(call->call_queue_lock);
5346 if (queue_IsOnQueue(call)) {
5348 if (flags & RX_CALL_WAIT_PROC) {
5349 rx_atomic_dec(&rx_nWaiting);
5352 MUTEX_EXIT(call->call_queue_lock);
5353 CLEAR_CALL_QUEUE_LOCK(call);
5355 #else /* RX_ENABLE_LOCKS */
5356 if (queue_IsOnQueue(call)) {
5358 if (flags & RX_CALL_WAIT_PROC)
5359 rx_atomic_dec(&rx_nWaiting);
5361 #endif /* RX_ENABLE_LOCKS */
5363 rxi_KeepAliveOff(call);
5364 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5367 /* Send an acknowledge for the indicated packet (seq,serial) of the
5368 * indicated call, for the indicated reason (reason). This
5369 * acknowledge will specifically acknowledge receiving the packet, and
5370 * will also specify which other packets for this call have been
5371 * received. This routine returns the packet that was used to the
5372 * caller. The caller is responsible for freeing it or re-using it.
5373 * This acknowledgement also returns the highest sequence number
5374 * actually read out by the higher level to the sender; the sender
5375 * promises to keep around packets that have not been read by the
5376 * higher level yet (unless, of course, the sender decides to abort
5377 * the call altogether). Any of p, seq, serial, pflags, or reason may
5378 * be set to zero without ill effect. That is, if they are zero, they
5379 * will not convey any information.
5380 * NOW there is a trailer field, after the ack where it will safely be
5381 * ignored by mundanes, which indicates the maximum size packet this
5382 * host can swallow. */
5384 struct rx_packet *optionalPacket; use to send ack (or null)
5385 int seq; Sequence number of the packet we are acking
5386 int serial; Serial number of the packet
5387 int pflags; Flags field from packet header
5388 int reason; Reason an acknowledge was prompted
5392 rxi_SendAck(struct rx_call *call,
5393 struct rx_packet *optionalPacket, int serial, int reason,
5396 struct rx_ackPacket *ap;
5397 struct rx_packet *rqp;
5398 struct rx_packet *nxp; /* For queue_Scan */
5399 struct rx_packet *p;
5402 afs_uint32 padbytes = 0;
5403 #ifdef RX_ENABLE_TSFPQ
5404 struct rx_ts_info_t * rx_ts_info;
5408 * Open the receive window once a thread starts reading packets
5410 if (call->rnext > 1) {
5411 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5414 /* Don't attempt to grow MTU if this is a critical ping */
5415 if (reason == RX_ACK_MTU) {
5416 /* keep track of per-call attempts, if we're over max, do in small
5417 * otherwise in larger? set a size to increment by, decrease
5420 if (call->conn->peer->maxPacketSize &&
5421 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5423 padbytes = call->conn->peer->maxPacketSize+16;
5425 padbytes = call->conn->peer->maxMTU + 128;
5427 /* do always try a minimum size ping */
5428 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5430 /* subtract the ack payload */
5431 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5432 reason = RX_ACK_PING;
5435 call->nHardAcks = 0;
5436 call->nSoftAcks = 0;
5437 if (call->rnext > call->lastAcked)
5438 call->lastAcked = call->rnext;
5442 rx_computelen(p, p->length); /* reset length, you never know */
5443 } /* where that's been... */
5444 #ifdef RX_ENABLE_TSFPQ
5446 RX_TS_INFO_GET(rx_ts_info);
5447 if ((p = rx_ts_info->local_special_packet)) {
5448 rx_computelen(p, p->length);
5449 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5450 rx_ts_info->local_special_packet = p;
5451 } else { /* We won't send the ack, but don't panic. */
5452 return optionalPacket;
5456 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5457 /* We won't send the ack, but don't panic. */
5458 return optionalPacket;
5463 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5466 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5467 #ifndef RX_ENABLE_TSFPQ
5468 if (!optionalPacket)
5471 return optionalPacket;
5473 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5474 if (rx_Contiguous(p) < templ) {
5475 #ifndef RX_ENABLE_TSFPQ
5476 if (!optionalPacket)
5479 return optionalPacket;
5484 /* MTUXXX failing to send an ack is very serious. We should */
5485 /* try as hard as possible to send even a partial ack; it's */
5486 /* better than nothing. */
5487 ap = (struct rx_ackPacket *)rx_DataOf(p);
5488 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5489 ap->reason = reason;
5491 /* The skew computation used to be bogus, I think it's better now. */
5492 /* We should start paying attention to skew. XXX */
5493 ap->serial = htonl(serial);
5494 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5497 * First packet not yet forwarded to reader. When ACKALL has been
5498 * sent the peer has been told that all received packets will be
5499 * delivered to the reader. The value 'rnext' is used internally
5500 * to refer to the next packet in the receive queue that must be
5501 * delivered to the reader. From the perspective of the peer it
5502 * already has so report the last sequence number plus one if there
5503 * are packets in the receive queue awaiting processing.
5505 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5506 !queue_IsEmpty(&call->rq)) {
5507 ap->firstPacket = htonl(queue_Last(&call->rq, rx_packet)->header.seq + 1);
5509 ap->firstPacket = htonl(call->rnext);
5511 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5513 /* No fear of running out of ack packet here because there can only be at most
5514 * one window full of unacknowledged packets. The window size must be constrained
5515 * to be less than the maximum ack size, of course. Also, an ack should always
5516 * fit into a single packet -- it should not ever be fragmented. */
5517 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5518 if (!rqp || !call->rq.next
5519 || (rqp->header.seq > (call->rnext + call->rwind))) {
5520 #ifndef RX_ENABLE_TSFPQ
5521 if (!optionalPacket)
5524 rxi_CallError(call, RX_CALL_DEAD);
5525 return optionalPacket;
5528 while (rqp->header.seq > call->rnext + offset)
5529 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5530 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5532 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5533 #ifndef RX_ENABLE_TSFPQ
5534 if (!optionalPacket)
5537 rxi_CallError(call, RX_CALL_DEAD);
5538 return optionalPacket;
5544 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5546 /* these are new for AFS 3.3 */
5547 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5548 templ = htonl(templ);
5549 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5550 templ = htonl(call->conn->peer->ifMTU);
5551 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5552 sizeof(afs_int32), &templ);
5554 /* new for AFS 3.4 */
5555 templ = htonl(call->rwind);
5556 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5557 sizeof(afs_int32), &templ);
5559 /* new for AFS 3.5 */
5560 templ = htonl(call->conn->peer->ifDgramPackets);
5561 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5562 sizeof(afs_int32), &templ);
5564 p->header.serviceId = call->conn->serviceId;
5565 p->header.cid = (call->conn->cid | call->channel);
5566 p->header.callNumber = *call->callNumber;
5568 p->header.securityIndex = call->conn->securityIndex;
5569 p->header.epoch = call->conn->epoch;
5570 p->header.type = RX_PACKET_TYPE_ACK;
5571 p->header.flags = RX_SLOW_START_OK;
5572 if (reason == RX_ACK_PING) {
5573 p->header.flags |= RX_REQUEST_ACK;
5575 p->length = padbytes +
5576 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5579 /* not fast but we can potentially use this if truncated
5580 * fragments are delivered to figure out the mtu.
5582 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5583 sizeof(afs_int32), sizeof(afs_int32),
5587 if (call->conn->type == RX_CLIENT_CONNECTION)
5588 p->header.flags |= RX_CLIENT_INITIATED;
5592 if (rxdebug_active) {
5596 len = _snprintf(msg, sizeof(msg),
5597 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5598 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5599 ntohl(ap->serial), ntohl(ap->previousPacket),
5600 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5601 ap->nAcks, ntohs(ap->bufferSpace) );
5605 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5606 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5610 OutputDebugString(msg);
5612 #else /* AFS_NT40_ENV */
5614 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5615 ap->reason, ntohl(ap->previousPacket),
5616 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5618 for (offset = 0; offset < ap->nAcks; offset++)
5619 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5624 #endif /* AFS_NT40_ENV */
5627 int i, nbytes = p->length;
5629 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5630 if (nbytes <= p->wirevec[i].iov_len) {
5633 savelen = p->wirevec[i].iov_len;
5635 p->wirevec[i].iov_len = nbytes;
5637 rxi_Send(call, p, istack);
5638 p->wirevec[i].iov_len = savelen;
5642 nbytes -= p->wirevec[i].iov_len;
5645 if (rx_stats_active)
5646 rx_atomic_inc(&rx_stats.ackPacketsSent);
5647 #ifndef RX_ENABLE_TSFPQ
5648 if (!optionalPacket)
5651 return optionalPacket; /* Return packet for re-use by caller */
5655 struct rx_packet **list;
5660 /* Send all of the packets in the list in single datagram */
5662 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5663 int istack, int moreFlag)
5669 struct rx_connection *conn = call->conn;
5670 struct rx_peer *peer = conn->peer;
5672 MUTEX_ENTER(&peer->peer_lock);
5673 peer->nSent += xmit->len;
5674 if (xmit->resending)
5675 peer->reSends += xmit->len;
5676 MUTEX_EXIT(&peer->peer_lock);
5678 if (rx_stats_active) {
5679 if (xmit->resending)
5680 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5682 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5685 clock_GetTime(&now);
5687 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5691 /* Set the packet flags and schedule the resend events */
5692 /* Only request an ack for the last packet in the list */
5693 for (i = 0; i < xmit->len; i++) {
5694 struct rx_packet *packet = xmit->list[i];
5696 /* Record the time sent */
5697 packet->timeSent = now;
5698 packet->flags |= RX_PKTFLAG_SENT;
5700 /* Ask for an ack on retransmitted packets, on every other packet
5701 * if the peer doesn't support slow start. Ask for an ack on every
5702 * packet until the congestion window reaches the ack rate. */
5703 if (packet->header.serial) {
5706 packet->firstSent = now;
5707 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5708 || (!(call->flags & RX_CALL_SLOW_START_OK)
5709 && (packet->header.seq & 1)))) {
5714 /* Tag this packet as not being the last in this group,
5715 * for the receiver's benefit */
5716 if (i < xmit->len - 1 || moreFlag) {
5717 packet->header.flags |= RX_MORE_PACKETS;
5722 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5725 /* Since we're about to send a data packet to the peer, it's
5726 * safe to nuke any scheduled end-of-packets ack */
5727 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5729 MUTEX_EXIT(&call->lock);
5730 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5731 if (xmit->len > 1) {
5732 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5734 rxi_SendPacket(call, conn, xmit->list[0], istack);
5736 MUTEX_ENTER(&call->lock);
5737 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5739 /* Tell the RTO calculation engine that we have sent a packet, and
5740 * if it was the last one */
5741 rxi_rto_packet_sent(call, lastPacket, istack);
5743 /* Update last send time for this call (for keep-alive
5744 * processing), and for the connection (so that we can discover
5745 * idle connections) */
5746 conn->lastSendTime = call->lastSendTime = clock_Sec();
5747 /* Let a set of retransmits trigger an idle timeout */
5748 if (!xmit->resending)
5749 call->lastSendData = call->lastSendTime;
5752 /* When sending packets we need to follow these rules:
5753 * 1. Never send more than maxDgramPackets in a jumbogram.
5754 * 2. Never send a packet with more than two iovecs in a jumbogram.
5755 * 3. Never send a retransmitted packet in a jumbogram.
5756 * 4. Never send more than cwind/4 packets in a jumbogram
5757 * We always keep the last list we should have sent so we
5758 * can set the RX_MORE_PACKETS flags correctly.
5762 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5767 struct xmitlist working;
5768 struct xmitlist last;
5770 struct rx_peer *peer = call->conn->peer;
5771 int morePackets = 0;
5773 memset(&last, 0, sizeof(struct xmitlist));
5774 working.list = &list[0];
5776 working.resending = 0;
5778 recovery = call->flags & RX_CALL_FAST_RECOVER;
5780 for (i = 0; i < len; i++) {
5781 /* Does the current packet force us to flush the current list? */
5783 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5784 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5786 /* This sends the 'last' list and then rolls the current working
5787 * set into the 'last' one, and resets the working set */
5790 rxi_SendList(call, &last, istack, 1);
5791 /* If the call enters an error state stop sending, or if
5792 * we entered congestion recovery mode, stop sending */
5794 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5799 working.resending = 0;
5800 working.list = &list[i];
5802 /* Add the current packet to the list if it hasn't been acked.
5803 * Otherwise adjust the list pointer to skip the current packet. */
5804 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5807 if (list[i]->header.serial)
5808 working.resending = 1;
5810 /* Do we need to flush the list? */
5811 if (working.len >= (int)peer->maxDgramPackets
5812 || working.len >= (int)call->nDgramPackets
5813 || working.len >= (int)call->cwind
5814 || list[i]->header.serial
5815 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5817 rxi_SendList(call, &last, istack, 1);
5818 /* If the call enters an error state stop sending, or if
5819 * we entered congestion recovery mode, stop sending */
5821 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5826 working.resending = 0;
5827 working.list = &list[i + 1];
5830 if (working.len != 0) {
5831 osi_Panic("rxi_SendList error");
5833 working.list = &list[i + 1];
5837 /* Send the whole list when the call is in receive mode, when
5838 * the call is in eof mode, when we are in fast recovery mode,
5839 * and when we have the last packet */
5840 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5841 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5842 || (call->flags & RX_CALL_FAST_RECOVER)) {
5843 /* Check for the case where the current list contains
5844 * an acked packet. Since we always send retransmissions
5845 * in a separate packet, we only need to check the first
5846 * packet in the list */
5847 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5851 rxi_SendList(call, &last, istack, morePackets);
5852 /* If the call enters an error state stop sending, or if
5853 * we entered congestion recovery mode, stop sending */
5855 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5859 rxi_SendList(call, &working, istack, 0);
5861 } else if (last.len > 0) {
5862 rxi_SendList(call, &last, istack, 0);
5863 /* Packets which are in 'working' are not sent by this call */
5868 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
5870 struct rx_call *call = arg0;
5871 struct rx_peer *peer;
5872 struct rx_packet *p, *nxp;
5873 struct clock maxTimeout = { 60, 0 };
5875 MUTEX_ENTER(&call->lock);
5877 peer = call->conn->peer;
5879 /* Make sure that the event pointer is removed from the call
5880 * structure, since there is no longer a per-call retransmission
5882 if (event == call->resendEvent) {
5883 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5884 rxevent_Put(call->resendEvent);
5885 call->resendEvent = NULL;
5888 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
5889 rxi_CheckBusy(call);
5892 if (queue_IsEmpty(&call->tq)) {
5893 /* Nothing to do. This means that we've been raced, and that an
5894 * ACK has come in between when we were triggered, and when we
5895 * actually got to run. */
5899 /* We're in loss recovery */
5900 call->flags |= RX_CALL_FAST_RECOVER;
5902 /* Mark all of the pending packets in the queue as being lost */
5903 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5904 if (!(p->flags & RX_PKTFLAG_ACKED))
5905 p->flags &= ~RX_PKTFLAG_SENT;
5908 /* We're resending, so we double the timeout of the call. This will be
5909 * dropped back down by the first successful ACK that we receive.
5911 * We apply a maximum value here of 60 seconds
5913 clock_Add(&call->rto, &call->rto);
5914 if (clock_Gt(&call->rto, &maxTimeout))
5915 call->rto = maxTimeout;
5917 /* Packet loss is most likely due to congestion, so drop our window size
5918 * and start again from the beginning */
5919 if (peer->maxDgramPackets >1) {
5920 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5921 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5923 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5924 call->nDgramPackets = 1;
5926 call->nextCwind = 1;
5929 MUTEX_ENTER(&peer->peer_lock);
5930 peer->MTU = call->MTU;
5931 peer->cwind = call->cwind;
5932 peer->nDgramPackets = 1;
5934 call->congestSeq = peer->congestSeq;
5935 MUTEX_EXIT(&peer->peer_lock);
5937 rxi_Start(call, istack);
5940 MUTEX_EXIT(&call->lock);
5943 /* This routine is called when new packets are readied for
5944 * transmission and when retransmission may be necessary, or when the
5945 * transmission window or burst count are favourable. This should be
5946 * better optimized for new packets, the usual case, now that we've
5947 * got rid of queues of send packets. XXXXXXXXXXX */
5949 rxi_Start(struct rx_call *call, int istack)
5952 struct rx_packet *p;
5953 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5958 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5959 if (rx_stats_active)
5960 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5965 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5967 /* Send (or resend) any packets that need it, subject to
5968 * window restrictions and congestion burst control
5969 * restrictions. Ask for an ack on the last packet sent in
5970 * this burst. For now, we're relying upon the window being
5971 * considerably bigger than the largest number of packets that
5972 * are typically sent at once by one initial call to
5973 * rxi_Start. This is probably bogus (perhaps we should ask
5974 * for an ack when we're half way through the current
5975 * window?). Also, for non file transfer applications, this
5976 * may end up asking for an ack for every packet. Bogus. XXXX
5979 * But check whether we're here recursively, and let the other guy
5982 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5983 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5984 call->flags |= RX_CALL_TQ_BUSY;
5986 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5988 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5989 call->flags &= ~RX_CALL_NEED_START;
5990 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5992 maxXmitPackets = MIN(call->twind, call->cwind);
5993 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5994 #ifdef RX_TRACK_PACKETS
5995 if ((p->flags & RX_PKTFLAG_FREE)
5996 || (!queue_IsEnd(&call->tq, nxp)
5997 && (nxp->flags & RX_PKTFLAG_FREE))
5998 || (p == (struct rx_packet *)&rx_freePacketQueue)
5999 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
6000 osi_Panic("rxi_Start: xmit queue clobbered");
6003 if (p->flags & RX_PKTFLAG_ACKED) {
6004 /* Since we may block, don't trust this */
6005 if (rx_stats_active)
6006 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6007 continue; /* Ignore this packet if it has been acknowledged */
6010 /* Turn off all flags except these ones, which are the same
6011 * on each transmission */
6012 p->header.flags &= RX_PRESET_FLAGS;
6014 if (p->header.seq >=
6015 call->tfirst + MIN((int)call->twind,
6016 (int)(call->nSoftAcked +
6018 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6019 /* Note: if we're waiting for more window space, we can
6020 * still send retransmits; hence we don't return here, but
6021 * break out to schedule a retransmit event */
6022 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6023 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6028 /* Transmit the packet if it needs to be sent. */
6029 if (!(p->flags & RX_PKTFLAG_SENT)) {
6030 if (nXmitPackets == maxXmitPackets) {
6031 rxi_SendXmitList(call, call->xmitList,
6032 nXmitPackets, istack);
6035 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6036 *(call->callNumber), p));
6037 call->xmitList[nXmitPackets++] = p;
6041 /* xmitList now hold pointers to all of the packets that are
6042 * ready to send. Now we loop to send the packets */
6043 if (nXmitPackets > 0) {
6044 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6048 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6050 /* We went into the error state while sending packets. Now is
6051 * the time to reset the call. This will also inform the using
6052 * process that the call is in an error state.
6054 if (rx_stats_active)
6055 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6056 call->flags &= ~RX_CALL_TQ_BUSY;
6057 rxi_WakeUpTransmitQueue(call);
6058 rxi_CallError(call, call->error);
6061 #ifdef RX_ENABLE_LOCKS
6062 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6064 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6065 /* Some packets have received acks. If they all have, we can clear
6066 * the transmit queue.
6069 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
6070 if (p->header.seq < call->tfirst
6071 && (p->flags & RX_PKTFLAG_ACKED)) {
6073 #ifdef RX_TRACK_PACKETS
6074 p->flags &= ~RX_PKTFLAG_TQ;
6076 #ifdef RXDEBUG_PACKET
6084 call->flags |= RX_CALL_TQ_CLEARME;
6086 #endif /* RX_ENABLE_LOCKS */
6087 if (call->flags & RX_CALL_TQ_CLEARME)
6088 rxi_ClearTransmitQueue(call, 1);
6089 } while (call->flags & RX_CALL_NEED_START);
6091 * TQ references no longer protected by this flag; they must remain
6092 * protected by the global lock.
6094 call->flags &= ~RX_CALL_TQ_BUSY;
6095 rxi_WakeUpTransmitQueue(call);
6097 call->flags |= RX_CALL_NEED_START;
6099 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6101 rxi_rto_cancel(call);
6105 /* Also adjusts the keep alive parameters for the call, to reflect
6106 * that we have just sent a packet (so keep alives aren't sent
6109 rxi_Send(struct rx_call *call, struct rx_packet *p,
6112 struct rx_connection *conn = call->conn;
6114 /* Stamp each packet with the user supplied status */
6115 p->header.userStatus = call->localStatus;
6117 /* Allow the security object controlling this call's security to
6118 * make any last-minute changes to the packet */
6119 RXS_SendPacket(conn->securityObject, call, p);
6121 /* Since we're about to send SOME sort of packet to the peer, it's
6122 * safe to nuke any scheduled end-of-packets ack */
6123 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
6125 /* Actually send the packet, filling in more connection-specific fields */
6126 MUTEX_EXIT(&call->lock);
6127 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6128 rxi_SendPacket(call, conn, p, istack);
6129 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6130 MUTEX_ENTER(&call->lock);
6132 /* Update last send time for this call (for keep-alive
6133 * processing), and for the connection (so that we can discover
6134 * idle connections) */
6135 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6136 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6137 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6139 conn->lastSendTime = call->lastSendTime = clock_Sec();
6140 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6141 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6142 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6143 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6144 RX_ACK_PING_RESPONSE)))
6145 call->lastSendData = call->lastSendTime;
6149 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6150 * that things are fine. Also called periodically to guarantee that nothing
6151 * falls through the cracks (e.g. (error + dally) connections have keepalive
6152 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6154 * haveCTLock Set if calling from rxi_ReapConnections
6156 #ifdef RX_ENABLE_LOCKS
6158 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6159 #else /* RX_ENABLE_LOCKS */
6161 rxi_CheckCall(struct rx_call *call)
6162 #endif /* RX_ENABLE_LOCKS */
6164 struct rx_connection *conn = call->conn;
6166 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6167 afs_uint32 fudgeFactor;
6170 int idle_timeout = 0;
6171 afs_int32 clock_diff = 0;
6175 /* Large swings in the clock can have a significant impact on
6176 * the performance of RX call processing. Forward clock shifts
6177 * will result in premature event triggering or timeouts.
6178 * Backward shifts can result in calls not completing until
6179 * the clock catches up with the original start clock value.
6181 * If a backward clock shift of more than five minutes is noticed,
6182 * just fail the call.
6184 if (now < call->lastSendTime)
6185 clock_diff = call->lastSendTime - now;
6186 if (now < call->startWait)
6187 clock_diff = MAX(clock_diff, call->startWait - now);
6188 if (now < call->lastReceiveTime)
6189 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6190 if (clock_diff > 5 * 60)
6192 if (call->state == RX_STATE_ACTIVE)
6193 rxi_CallError(call, RX_CALL_TIMEOUT);
6197 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6198 if (call->flags & RX_CALL_TQ_BUSY) {
6199 /* Call is active and will be reset by rxi_Start if it's
6200 * in an error state.
6205 /* RTT + 8*MDEV, rounded up to the next second. */
6206 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6207 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6209 deadTime = conn->secondsUntilDead + fudgeFactor;
6210 /* These are computed to the second (+- 1 second). But that's
6211 * good enough for these values, which should be a significant
6212 * number of seconds. */
6213 if (now > (call->lastReceiveTime + deadTime)) {
6214 if (call->state == RX_STATE_ACTIVE) {
6216 #if defined(KERNEL) && defined(AFS_SUN5_ENV)
6218 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6219 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6220 ip_stack_t *ipst = ns->netstack_ip;
6222 ire = ire_cache_lookup(conn->peer->host
6223 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6225 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
6227 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6234 if (ire && ire->ire_max_frag > 0)
6235 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6237 #if defined(GLOBAL_NETSTACKID)
6241 #endif /* ADAPT_PMTU */
6242 cerror = RX_CALL_DEAD;
6245 #ifdef RX_ENABLE_LOCKS
6246 /* Cancel pending events */
6247 rxevent_Cancel(&call->delayedAckEvent, call,
6248 RX_CALL_REFCOUNT_DELAY);
6249 rxi_rto_cancel(call);
6250 rxevent_Cancel(&call->keepAliveEvent, call,
6251 RX_CALL_REFCOUNT_ALIVE);
6252 rxevent_Cancel(&call->growMTUEvent, call,
6253 RX_CALL_REFCOUNT_MTU);
6254 MUTEX_ENTER(&rx_refcnt_mutex);
6255 /* if rxi_FreeCall returns 1 it has freed the call */
6256 if (call->refCount == 0 &&
6257 rxi_FreeCall(call, haveCTLock))
6259 MUTEX_EXIT(&rx_refcnt_mutex);
6262 MUTEX_EXIT(&rx_refcnt_mutex);
6264 #else /* RX_ENABLE_LOCKS */
6265 rxi_FreeCall(call, 0);
6267 #endif /* RX_ENABLE_LOCKS */
6269 /* Non-active calls are destroyed if they are not responding
6270 * to pings; active calls are simply flagged in error, so the
6271 * attached process can die reasonably gracefully. */
6274 if (conn->idleDeadDetection) {
6275 if (conn->idleDeadTime) {
6276 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6280 /* see if we have a non-activity timeout */
6281 if (call->startWait && ((call->startWait + idleDeadTime) < now) &&
6282 (call->flags & RX_CALL_READER_WAIT)) {
6283 if (call->state == RX_STATE_ACTIVE) {
6284 cerror = RX_CALL_TIMEOUT;
6289 if (call->lastSendData && ((call->lastSendData + idleDeadTime) < now)) {
6290 if (call->state == RX_STATE_ACTIVE) {
6291 cerror = conn->service ? conn->service->idleDeadErr : RX_CALL_IDLE;
6299 if (conn->hardDeadTime) {
6300 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6303 /* see if we have a hard timeout */
6305 && (now > (hardDeadTime + call->startTime.sec))) {
6306 if (call->state == RX_STATE_ACTIVE)
6307 rxi_CallError(call, RX_CALL_TIMEOUT);
6312 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6313 call->lastReceiveTime) {
6314 int oldMTU = conn->peer->ifMTU;
6316 /* if we thought we could send more, perhaps things got worse */
6317 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6318 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6319 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6320 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6322 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6324 /* minimum capped in SetPeerMtu */
6325 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6328 conn->lastPacketSize = 0;
6330 /* needed so ResetCall doesn't clobber us. */
6331 call->MTU = conn->peer->ifMTU;
6333 /* if we never succeeded, let the error pass out as-is */
6334 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6335 cerror = conn->msgsizeRetryErr;
6338 rxi_CallError(call, cerror);
6343 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6344 void *dummy, int dummy2)
6346 struct rx_connection *conn = arg1;
6347 struct rx_header theader;
6348 char tbuffer[1 + sizeof(struct rx_header)];
6349 struct sockaddr_in taddr;
6352 struct iovec tmpiov[2];
6355 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6358 tp = &tbuffer[sizeof(struct rx_header)];
6359 taddr.sin_family = AF_INET;
6360 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6361 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6362 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6363 taddr.sin_len = sizeof(struct sockaddr_in);
6365 memset(&theader, 0, sizeof(theader));
6366 theader.epoch = htonl(999);
6368 theader.callNumber = 0;
6371 theader.type = RX_PACKET_TYPE_VERSION;
6372 theader.flags = RX_LAST_PACKET;
6373 theader.serviceId = 0;
6375 memcpy(tbuffer, &theader, sizeof(theader));
6376 memcpy(tp, &a, sizeof(a));
6377 tmpiov[0].iov_base = tbuffer;
6378 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6380 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6382 MUTEX_ENTER(&conn->conn_data_lock);
6383 MUTEX_ENTER(&rx_refcnt_mutex);
6384 /* Only reschedule ourselves if the connection would not be destroyed */
6385 if (conn->refCount <= 1) {
6386 rxevent_Put(conn->natKeepAliveEvent);
6387 conn->natKeepAliveEvent = NULL;
6388 MUTEX_EXIT(&rx_refcnt_mutex);
6389 MUTEX_EXIT(&conn->conn_data_lock);
6390 rx_DestroyConnection(conn); /* drop the reference for this */
6392 conn->refCount--; /* drop the reference for this */
6393 MUTEX_EXIT(&rx_refcnt_mutex);
6394 rxevent_Put(conn->natKeepAliveEvent);
6395 conn->natKeepAliveEvent = NULL;
6396 rxi_ScheduleNatKeepAliveEvent(conn);
6397 MUTEX_EXIT(&conn->conn_data_lock);
6402 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6404 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6405 struct clock when, now;
6406 clock_GetTime(&now);
6408 when.sec += conn->secondsUntilNatPing;
6409 MUTEX_ENTER(&rx_refcnt_mutex);
6410 conn->refCount++; /* hold a reference for this */
6411 MUTEX_EXIT(&rx_refcnt_mutex);
6412 conn->natKeepAliveEvent =
6413 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6418 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6420 MUTEX_ENTER(&conn->conn_data_lock);
6421 conn->secondsUntilNatPing = seconds;
6423 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6424 rxi_ScheduleNatKeepAliveEvent(conn);
6426 conn->flags |= RX_CONN_NAT_PING;
6428 MUTEX_EXIT(&conn->conn_data_lock);
6432 rxi_NatKeepAliveOn(struct rx_connection *conn)
6434 MUTEX_ENTER(&conn->conn_data_lock);
6435 /* if it's already attached */
6436 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6437 rxi_ScheduleNatKeepAliveEvent(conn);
6439 conn->flags |= RX_CONN_NAT_PING;
6440 MUTEX_EXIT(&conn->conn_data_lock);
6443 /* When a call is in progress, this routine is called occasionally to
6444 * make sure that some traffic has arrived (or been sent to) the peer.
6445 * If nothing has arrived in a reasonable amount of time, the call is
6446 * declared dead; if nothing has been sent for a while, we send a
6447 * keep-alive packet (if we're actually trying to keep the call alive)
6450 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6453 struct rx_call *call = arg1;
6454 struct rx_connection *conn;
6457 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6458 MUTEX_ENTER(&call->lock);
6460 if (event == call->keepAliveEvent) {
6461 rxevent_Put(call->keepAliveEvent);
6462 call->keepAliveEvent = NULL;
6467 #ifdef RX_ENABLE_LOCKS
6468 if (rxi_CheckCall(call, 0)) {
6469 MUTEX_EXIT(&call->lock);
6472 #else /* RX_ENABLE_LOCKS */
6473 if (rxi_CheckCall(call))
6475 #endif /* RX_ENABLE_LOCKS */
6477 /* Don't try to keep alive dallying calls */
6478 if (call->state == RX_STATE_DALLY) {
6479 MUTEX_EXIT(&call->lock);
6484 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6485 /* Don't try to send keepalives if there is unacknowledged data */
6486 /* the rexmit code should be good enough, this little hack
6487 * doesn't quite work XXX */
6488 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6490 rxi_ScheduleKeepAliveEvent(call);
6491 MUTEX_EXIT(&call->lock);
6494 /* Does what's on the nameplate. */
6496 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6498 struct rx_call *call = arg1;
6499 struct rx_connection *conn;
6501 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6502 MUTEX_ENTER(&call->lock);
6504 if (event == call->growMTUEvent) {
6505 rxevent_Put(call->growMTUEvent);
6506 call->growMTUEvent = NULL;
6509 #ifdef RX_ENABLE_LOCKS
6510 if (rxi_CheckCall(call, 0)) {
6511 MUTEX_EXIT(&call->lock);
6514 #else /* RX_ENABLE_LOCKS */
6515 if (rxi_CheckCall(call))
6517 #endif /* RX_ENABLE_LOCKS */
6519 /* Don't bother with dallying calls */
6520 if (call->state == RX_STATE_DALLY) {
6521 MUTEX_EXIT(&call->lock);
6528 * keep being scheduled, just don't do anything if we're at peak,
6529 * or we're not set up to be properly handled (idle timeout required)
6531 if ((conn->peer->maxPacketSize != 0) &&
6532 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6533 conn->idleDeadDetection)
6534 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6535 rxi_ScheduleGrowMTUEvent(call, 0);
6536 MUTEX_EXIT(&call->lock);
6540 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6542 if (!call->keepAliveEvent) {
6543 struct clock when, now;
6544 clock_GetTime(&now);
6546 when.sec += call->conn->secondsUntilPing;
6547 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6548 call->keepAliveEvent =
6549 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6554 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6556 if (!call->growMTUEvent) {
6557 struct clock when, now;
6559 clock_GetTime(&now);
6562 if (call->conn->secondsUntilPing)
6563 secs = (6*call->conn->secondsUntilPing)-1;
6565 if (call->conn->secondsUntilDead)
6566 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6570 CALL_HOLD(call, RX_CALL_REFCOUNT_MTU);
6571 call->growMTUEvent =
6572 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6576 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6578 rxi_KeepAliveOn(struct rx_call *call)
6580 /* Pretend last packet received was received now--i.e. if another
6581 * packet isn't received within the keep alive time, then the call
6582 * will die; Initialize last send time to the current time--even
6583 * if a packet hasn't been sent yet. This will guarantee that a
6584 * keep-alive is sent within the ping time */
6585 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6586 rxi_ScheduleKeepAliveEvent(call);
6590 * Solely in order that callers not need to include rx_call.h
6593 rx_KeepAliveOff(struct rx_call *call)
6595 rxi_KeepAliveOff(call);
6598 rx_KeepAliveOn(struct rx_call *call)
6600 rxi_KeepAliveOn(call);
6604 rxi_GrowMTUOn(struct rx_call *call)
6606 struct rx_connection *conn = call->conn;
6607 MUTEX_ENTER(&conn->conn_data_lock);
6608 conn->lastPingSizeSer = conn->lastPingSize = 0;
6609 MUTEX_EXIT(&conn->conn_data_lock);
6610 rxi_ScheduleGrowMTUEvent(call, 1);
6613 /* This routine is called to send connection abort messages
6614 * that have been delayed to throttle looping clients. */
6616 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6619 struct rx_connection *conn = arg1;
6622 struct rx_packet *packet;
6624 MUTEX_ENTER(&conn->conn_data_lock);
6625 rxevent_Put(conn->delayedAbortEvent);
6626 conn->delayedAbortEvent = NULL;
6627 error = htonl(conn->error);
6629 MUTEX_EXIT(&conn->conn_data_lock);
6630 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6633 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6634 RX_PACKET_TYPE_ABORT, (char *)&error,
6636 rxi_FreePacket(packet);
6640 /* This routine is called to send call abort messages
6641 * that have been delayed to throttle looping clients. */
6643 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6646 struct rx_call *call = arg1;
6649 struct rx_packet *packet;
6651 MUTEX_ENTER(&call->lock);
6652 rxevent_Put(call->delayedAbortEvent);
6653 call->delayedAbortEvent = NULL;
6654 error = htonl(call->error);
6656 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6659 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6660 (char *)&error, sizeof(error), 0);
6661 rxi_FreePacket(packet);
6663 MUTEX_EXIT(&call->lock);
6664 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6667 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6668 * seconds) to ask the client to authenticate itself. The routine
6669 * issues a challenge to the client, which is obtained from the
6670 * security object associated with the connection */
6672 rxi_ChallengeEvent(struct rxevent *event,
6673 void *arg0, void *arg1, int tries)
6675 struct rx_connection *conn = arg0;
6678 rxevent_Put(conn->challengeEvent);
6679 conn->challengeEvent = NULL;
6682 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6683 struct rx_packet *packet;
6684 struct clock when, now;
6687 /* We've failed to authenticate for too long.
6688 * Reset any calls waiting for authentication;
6689 * they are all in RX_STATE_PRECALL.
6693 MUTEX_ENTER(&conn->conn_call_lock);
6694 for (i = 0; i < RX_MAXCALLS; i++) {
6695 struct rx_call *call = conn->call[i];
6697 MUTEX_ENTER(&call->lock);
6698 if (call->state == RX_STATE_PRECALL) {
6699 rxi_CallError(call, RX_CALL_DEAD);
6700 rxi_SendCallAbort(call, NULL, 0, 0);
6702 MUTEX_EXIT(&call->lock);
6705 MUTEX_EXIT(&conn->conn_call_lock);
6709 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6711 /* If there's no packet available, do this later. */
6712 RXS_GetChallenge(conn->securityObject, conn, packet);
6713 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6714 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6715 rxi_FreePacket(packet);
6717 clock_GetTime(&now);
6719 when.sec += RX_CHALLENGE_TIMEOUT;
6720 conn->challengeEvent =
6721 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6726 /* Call this routine to start requesting the client to authenticate
6727 * itself. This will continue until authentication is established,
6728 * the call times out, or an invalid response is returned. The
6729 * security object associated with the connection is asked to create
6730 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6731 * defined earlier. */
6733 rxi_ChallengeOn(struct rx_connection *conn)
6735 if (!conn->challengeEvent) {
6736 RXS_CreateChallenge(conn->securityObject, conn);
6737 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6742 /* rxi_ComputeRoundTripTime is called with peer locked. */
6743 /* peer may be null */
6745 rxi_ComputeRoundTripTime(struct rx_packet *p,
6746 struct rx_ackPacket *ack,
6747 struct rx_call *call,
6748 struct rx_peer *peer,
6751 struct clock thisRtt, *sentp;
6755 /* If the ACK is delayed, then do nothing */
6756 if (ack->reason == RX_ACK_DELAY)
6759 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6760 * their RTT multiple times, so only include the RTT of the last packet
6762 if (p->flags & RX_JUMBO_PACKET)
6765 /* Use the serial number to determine which transmission the ACK is for,
6766 * and set the sent time to match this. If we have no serial number, then
6767 * only use the ACK for RTT calculations if the packet has not been
6771 serial = ntohl(ack->serial);
6773 if (serial == p->header.serial) {
6774 sentp = &p->timeSent;
6775 } else if (serial == p->firstSerial) {
6776 sentp = &p->firstSent;
6777 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6778 sentp = &p->firstSent;
6782 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6783 sentp = &p->firstSent;
6790 if (clock_Lt(&thisRtt, sentp))
6791 return; /* somebody set the clock back, don't count this time. */
6793 clock_Sub(&thisRtt, sentp);
6794 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6795 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6797 if (clock_IsZero(&thisRtt)) {
6799 * The actual round trip time is shorter than the
6800 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6801 * Since we can't tell which at the moment we will assume 1ms.
6803 thisRtt.usec = 1000;
6806 if (rx_stats_active) {
6807 MUTEX_ENTER(&rx_stats_mutex);
6808 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6809 rx_stats.minRtt = thisRtt;
6810 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6811 if (thisRtt.sec > 60) {
6812 MUTEX_EXIT(&rx_stats_mutex);
6813 return; /* somebody set the clock ahead */
6815 rx_stats.maxRtt = thisRtt;
6817 clock_Add(&rx_stats.totalRtt, &thisRtt);
6818 rx_atomic_inc(&rx_stats.nRttSamples);
6819 MUTEX_EXIT(&rx_stats_mutex);
6822 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6824 /* Apply VanJacobson round-trip estimations */
6829 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6830 * srtt is stored as fixed point with 3 bits after the binary
6831 * point (i.e., scaled by 8). The following magic is
6832 * equivalent to the smoothing algorithm in rfc793 with an
6833 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6834 * srtt'*8 = rtt + srtt*7
6835 * srtt'*8 = srtt*8 + rtt - srtt
6836 * srtt' = srtt + rtt/8 - srtt/8
6837 * srtt' = srtt + (rtt - srtt)/8
6840 delta = _8THMSEC(&thisRtt) - call->rtt;
6841 call->rtt += (delta >> 3);
6844 * We accumulate a smoothed rtt variance (actually, a smoothed
6845 * mean difference), then set the retransmit timer to smoothed
6846 * rtt + 4 times the smoothed variance (was 2x in van's original
6847 * paper, but 4x works better for me, and apparently for him as
6849 * rttvar is stored as
6850 * fixed point with 2 bits after the binary point (scaled by
6851 * 4). The following is equivalent to rfc793 smoothing with
6852 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6853 * rttvar'*4 = rttvar*3 + |delta|
6854 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6855 * rttvar' = rttvar + |delta|/4 - rttvar/4
6856 * rttvar' = rttvar + (|delta| - rttvar)/4
6857 * This replaces rfc793's wired-in beta.
6858 * dev*4 = dev*4 + (|actual - expected| - dev)
6864 delta -= (call->rtt_dev << 1);
6865 call->rtt_dev += (delta >> 3);
6867 /* I don't have a stored RTT so I start with this value. Since I'm
6868 * probably just starting a call, and will be pushing more data down
6869 * this, I expect congestion to increase rapidly. So I fudge a
6870 * little, and I set deviance to half the rtt. In practice,
6871 * deviance tends to approach something a little less than
6872 * half the smoothed rtt. */
6873 call->rtt = _8THMSEC(&thisRtt) + 8;
6874 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6876 /* the smoothed RTT time is RTT + 4*MDEV
6878 * We allow a user specified minimum to be set for this, to allow clamping
6879 * at a minimum value in the same way as TCP. In addition, we have to allow
6880 * for the possibility that this packet is answered by a delayed ACK, so we
6881 * add on a fixed 200ms to account for that timer expiring.
6884 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6885 rx_minPeerTimeout) + 200;
6886 clock_Zero(&call->rto);
6887 clock_Addmsec(&call->rto, rtt_timeout);
6889 /* Update the peer, so any new calls start with our values */
6890 peer->rtt_dev = call->rtt_dev;
6891 peer->rtt = call->rtt;
6893 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6894 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6898 /* Find all server connections that have not been active for a long time, and
6901 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
6904 struct clock now, when;
6905 clock_GetTime(&now);
6907 /* Find server connection structures that haven't been used for
6908 * greater than rx_idleConnectionTime */
6910 struct rx_connection **conn_ptr, **conn_end;
6911 int i, havecalls = 0;
6912 MUTEX_ENTER(&rx_connHashTable_lock);
6913 for (conn_ptr = &rx_connHashTable[0], conn_end =
6914 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6916 struct rx_connection *conn, *next;
6917 struct rx_call *call;
6921 for (conn = *conn_ptr; conn; conn = next) {
6922 /* XXX -- Shouldn't the connection be locked? */
6925 for (i = 0; i < RX_MAXCALLS; i++) {
6926 call = conn->call[i];
6930 code = MUTEX_TRYENTER(&call->lock);
6933 #ifdef RX_ENABLE_LOCKS
6934 result = rxi_CheckCall(call, 1);
6935 #else /* RX_ENABLE_LOCKS */
6936 result = rxi_CheckCall(call);
6937 #endif /* RX_ENABLE_LOCKS */
6938 MUTEX_EXIT(&call->lock);
6940 /* If CheckCall freed the call, it might
6941 * have destroyed the connection as well,
6942 * which screws up the linked lists.
6948 if (conn->type == RX_SERVER_CONNECTION) {
6949 /* This only actually destroys the connection if
6950 * there are no outstanding calls */
6951 MUTEX_ENTER(&conn->conn_data_lock);
6952 MUTEX_ENTER(&rx_refcnt_mutex);
6953 if (!havecalls && !conn->refCount
6954 && ((conn->lastSendTime + rx_idleConnectionTime) <
6956 conn->refCount++; /* it will be decr in rx_DestroyConn */
6957 MUTEX_EXIT(&rx_refcnt_mutex);
6958 MUTEX_EXIT(&conn->conn_data_lock);
6959 #ifdef RX_ENABLE_LOCKS
6960 rxi_DestroyConnectionNoLock(conn);
6961 #else /* RX_ENABLE_LOCKS */
6962 rxi_DestroyConnection(conn);
6963 #endif /* RX_ENABLE_LOCKS */
6965 #ifdef RX_ENABLE_LOCKS
6967 MUTEX_EXIT(&rx_refcnt_mutex);
6968 MUTEX_EXIT(&conn->conn_data_lock);
6970 #endif /* RX_ENABLE_LOCKS */
6974 #ifdef RX_ENABLE_LOCKS
6975 while (rx_connCleanup_list) {
6976 struct rx_connection *conn;
6977 conn = rx_connCleanup_list;
6978 rx_connCleanup_list = rx_connCleanup_list->next;
6979 MUTEX_EXIT(&rx_connHashTable_lock);
6980 rxi_CleanupConnection(conn);
6981 MUTEX_ENTER(&rx_connHashTable_lock);
6983 MUTEX_EXIT(&rx_connHashTable_lock);
6984 #endif /* RX_ENABLE_LOCKS */
6987 /* Find any peer structures that haven't been used (haven't had an
6988 * associated connection) for greater than rx_idlePeerTime */
6990 struct rx_peer **peer_ptr, **peer_end;
6994 * Why do we need to hold the rx_peerHashTable_lock across
6995 * the incrementing of peer_ptr since the rx_peerHashTable
6996 * array is not changing? We don't.
6998 * By dropping the lock periodically we can permit other
6999 * activities to be performed while a rxi_ReapConnections
7000 * call is in progress. The goal of reap connections
7001 * is to clean up quickly without causing large amounts
7002 * of contention. Therefore, it is important that global
7003 * mutexes not be held for extended periods of time.
7005 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7006 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7008 struct rx_peer *peer, *next, *prev;
7010 MUTEX_ENTER(&rx_peerHashTable_lock);
7011 for (prev = peer = *peer_ptr; peer; peer = next) {
7013 code = MUTEX_TRYENTER(&peer->peer_lock);
7014 if ((code) && (peer->refCount == 0)
7015 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7016 rx_interface_stat_p rpc_stat, nrpc_stat;
7020 * now know that this peer object is one to be
7021 * removed from the hash table. Once it is removed
7022 * it can't be referenced by other threads.
7023 * Lets remove it first and decrement the struct
7024 * nPeerStructs count.
7026 if (peer == *peer_ptr) {
7032 if (rx_stats_active)
7033 rx_atomic_dec(&rx_stats.nPeerStructs);
7036 * Now if we hold references on 'prev' and 'next'
7037 * we can safely drop the rx_peerHashTable_lock
7038 * while we destroy this 'peer' object.
7044 MUTEX_EXIT(&rx_peerHashTable_lock);
7046 MUTEX_EXIT(&peer->peer_lock);
7047 MUTEX_DESTROY(&peer->peer_lock);
7049 (&peer->rpcStats, rpc_stat, nrpc_stat,
7050 rx_interface_stat)) {
7051 unsigned int num_funcs;
7054 queue_Remove(&rpc_stat->queue_header);
7055 queue_Remove(&rpc_stat->all_peers);
7056 num_funcs = rpc_stat->stats[0].func_total;
7058 sizeof(rx_interface_stat_t) +
7059 rpc_stat->stats[0].func_total *
7060 sizeof(rx_function_entry_v1_t);
7062 rxi_Free(rpc_stat, space);
7064 MUTEX_ENTER(&rx_rpc_stats);
7065 rxi_rpc_peer_stat_cnt -= num_funcs;
7066 MUTEX_EXIT(&rx_rpc_stats);
7071 * Regain the rx_peerHashTable_lock and
7072 * decrement the reference count on 'prev'
7075 MUTEX_ENTER(&rx_peerHashTable_lock);
7082 MUTEX_EXIT(&peer->peer_lock);
7087 MUTEX_EXIT(&rx_peerHashTable_lock);
7091 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7092 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7093 * GC, just below. Really, we shouldn't have to keep moving packets from
7094 * one place to another, but instead ought to always know if we can
7095 * afford to hold onto a packet in its particular use. */
7096 MUTEX_ENTER(&rx_freePktQ_lock);
7097 if (rx_waitingForPackets) {
7098 rx_waitingForPackets = 0;
7099 #ifdef RX_ENABLE_LOCKS
7100 CV_BROADCAST(&rx_waitingForPackets_cv);
7102 osi_rxWakeup(&rx_waitingForPackets);
7105 MUTEX_EXIT(&rx_freePktQ_lock);
7108 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7109 rxevent_Put(rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0));
7113 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7114 * rx.h is sort of strange this is better. This is called with a security
7115 * object before it is discarded. Each connection using a security object has
7116 * its own refcount to the object so it won't actually be freed until the last
7117 * connection is destroyed.
7119 * This is the only rxs module call. A hold could also be written but no one
7123 rxs_Release(struct rx_securityClass *aobj)
7125 return RXS_Close(aobj);
7133 #define TRACE_OPTION_RX_DEBUG 16
7141 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7142 0, KEY_QUERY_VALUE, &parmKey);
7143 if (code != ERROR_SUCCESS)
7146 dummyLen = sizeof(TraceOption);
7147 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7148 (BYTE *) &TraceOption, &dummyLen);
7149 if (code == ERROR_SUCCESS) {
7150 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7152 RegCloseKey (parmKey);
7153 #endif /* AFS_NT40_ENV */
7158 rx_DebugOnOff(int on)
7162 rxdebug_active = on;
7168 rx_StatsOnOff(int on)
7170 rx_stats_active = on;
7174 /* Don't call this debugging routine directly; use dpf */
7176 rxi_DebugPrint(char *format, ...)
7185 va_start(ap, format);
7187 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7190 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7192 OutputDebugString(msg);
7198 va_start(ap, format);
7200 clock_GetTime(&now);
7201 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7202 (unsigned int)now.usec);
7203 vfprintf(rx_Log, format, ap);
7211 * This function is used to process the rx_stats structure that is local
7212 * to a process as well as an rx_stats structure received from a remote
7213 * process (via rxdebug). Therefore, it needs to do minimal version
7217 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7218 afs_int32 freePackets, char version)
7222 if (size != sizeof(struct rx_statistics)) {
7224 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7225 size, sizeof(struct rx_statistics));
7228 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7231 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7232 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7233 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7234 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7235 s->specialPktAllocFailures);
7237 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7238 s->receivePktAllocFailures, s->sendPktAllocFailures,
7239 s->specialPktAllocFailures);
7243 " greedy %u, " "bogusReads %u (last from host %x), "
7244 "noPackets %u, " "noBuffers %u, " "selects %u, "
7245 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7246 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7247 s->selects, s->sendSelects);
7249 fprintf(file, " packets read: ");
7250 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7251 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7253 fprintf(file, "\n");
7256 " other read counters: data %u, " "ack %u, " "dup %u "
7257 "spurious %u " "dally %u\n", s->dataPacketsRead,
7258 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7259 s->ignorePacketDally);
7261 fprintf(file, " packets sent: ");
7262 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7263 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7265 fprintf(file, "\n");
7268 " other send counters: ack %u, " "data %u (not resends), "
7269 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7270 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7271 s->dataPacketsPushed, s->ignoreAckedPacket);
7274 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7275 s->netSendFailures, (int)s->fatalErrors);
7277 if (s->nRttSamples) {
7278 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7279 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7281 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7282 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7286 " %d server connections, " "%d client connections, "
7287 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7288 s->nServerConns, s->nClientConns, s->nPeerStructs,
7289 s->nCallStructs, s->nFreeCallStructs);
7291 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7292 fprintf(file, " %d clock updates\n", clock_nUpdates);
7296 /* for backward compatibility */
7298 rx_PrintStats(FILE * file)
7300 MUTEX_ENTER(&rx_stats_mutex);
7301 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7302 sizeof(rx_stats), rx_nFreePackets,
7304 MUTEX_EXIT(&rx_stats_mutex);
7308 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7310 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
7311 ntohl(peer->host), (int)ntohs(peer->port), (int)peer->burstSize,
7312 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
7315 " Rtt %d, " "total sent %d, " "resent %d\n",
7316 peer->rtt, peer->nSent, peer->reSends);
7319 " Packet size %d, " "max in packet skew %d, "
7320 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
7321 (int)peer->outPacketSkew);
7325 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7327 * This mutex protects the following static variables:
7331 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7332 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7334 #define LOCK_RX_DEBUG
7335 #define UNLOCK_RX_DEBUG
7336 #endif /* AFS_PTHREAD_ENV */
7338 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7340 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7341 u_char type, void *inputData, size_t inputLength,
7342 void *outputData, size_t outputLength)
7344 static afs_int32 counter = 100;
7345 time_t waitTime, waitCount;
7346 struct rx_header theader;
7349 struct timeval tv_now, tv_wake, tv_delta;
7350 struct sockaddr_in taddr, faddr;
7364 tp = &tbuffer[sizeof(struct rx_header)];
7365 taddr.sin_family = AF_INET;
7366 taddr.sin_port = remotePort;
7367 taddr.sin_addr.s_addr = remoteAddr;
7368 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7369 taddr.sin_len = sizeof(struct sockaddr_in);
7372 memset(&theader, 0, sizeof(theader));
7373 theader.epoch = htonl(999);
7375 theader.callNumber = htonl(counter);
7378 theader.type = type;
7379 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7380 theader.serviceId = 0;
7382 memcpy(tbuffer, &theader, sizeof(theader));
7383 memcpy(tp, inputData, inputLength);
7385 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7386 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7388 /* see if there's a packet available */
7389 gettimeofday(&tv_wake, NULL);
7390 tv_wake.tv_sec += waitTime;
7393 FD_SET(socket, &imask);
7394 tv_delta.tv_sec = tv_wake.tv_sec;
7395 tv_delta.tv_usec = tv_wake.tv_usec;
7396 gettimeofday(&tv_now, NULL);
7398 if (tv_delta.tv_usec < tv_now.tv_usec) {
7400 tv_delta.tv_usec += 1000000;
7403 tv_delta.tv_usec -= tv_now.tv_usec;
7405 if (tv_delta.tv_sec < tv_now.tv_sec) {
7409 tv_delta.tv_sec -= tv_now.tv_sec;
7412 code = select(0, &imask, 0, 0, &tv_delta);
7413 #else /* AFS_NT40_ENV */
7414 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7415 #endif /* AFS_NT40_ENV */
7416 if (code == 1 && FD_ISSET(socket, &imask)) {
7417 /* now receive a packet */
7418 faddrLen = sizeof(struct sockaddr_in);
7420 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7421 (struct sockaddr *)&faddr, &faddrLen);
7424 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7425 if (counter == ntohl(theader.callNumber))
7433 /* see if we've timed out */
7441 code -= sizeof(struct rx_header);
7442 if (code > outputLength)
7443 code = outputLength;
7444 memcpy(outputData, tp, code);
7447 #endif /* RXDEBUG */
7450 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7451 afs_uint16 remotePort, struct rx_debugStats * stat,
7452 afs_uint32 * supportedValues)
7454 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7456 struct rx_debugIn in;
7458 *supportedValues = 0;
7459 in.type = htonl(RX_DEBUGI_GETSTATS);
7462 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7463 &in, sizeof(in), stat, sizeof(*stat));
7466 * If the call was successful, fixup the version and indicate
7467 * what contents of the stat structure are valid.
7468 * Also do net to host conversion of fields here.
7472 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7473 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7475 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7476 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7478 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7479 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7481 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7482 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7484 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7485 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7487 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7488 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7490 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7491 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7493 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7494 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7496 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7497 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7499 stat->nFreePackets = ntohl(stat->nFreePackets);
7500 stat->packetReclaims = ntohl(stat->packetReclaims);
7501 stat->callsExecuted = ntohl(stat->callsExecuted);
7502 stat->nWaiting = ntohl(stat->nWaiting);
7503 stat->idleThreads = ntohl(stat->idleThreads);
7504 stat->nWaited = ntohl(stat->nWaited);
7505 stat->nPackets = ntohl(stat->nPackets);
7514 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7515 afs_uint16 remotePort, struct rx_statistics * stat,
7516 afs_uint32 * supportedValues)
7518 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7520 struct rx_debugIn in;
7521 afs_int32 *lp = (afs_int32 *) stat;
7525 * supportedValues is currently unused, but added to allow future
7526 * versioning of this function.
7529 *supportedValues = 0;
7530 in.type = htonl(RX_DEBUGI_RXSTATS);
7532 memset(stat, 0, sizeof(*stat));
7534 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7535 &in, sizeof(in), stat, sizeof(*stat));
7540 * Do net to host conversion here
7543 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7554 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7555 afs_uint16 remotePort, size_t version_length,
7558 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7560 return MakeDebugCall(socket, remoteAddr, remotePort,
7561 RX_PACKET_TYPE_VERSION, a, 1, version,
7569 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7570 afs_uint16 remotePort, afs_int32 * nextConnection,
7571 int allConnections, afs_uint32 debugSupportedValues,
7572 struct rx_debugConn * conn,
7573 afs_uint32 * supportedValues)
7575 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7577 struct rx_debugIn in;
7581 * supportedValues is currently unused, but added to allow future
7582 * versioning of this function.
7585 *supportedValues = 0;
7586 if (allConnections) {
7587 in.type = htonl(RX_DEBUGI_GETALLCONN);
7589 in.type = htonl(RX_DEBUGI_GETCONN);
7591 in.index = htonl(*nextConnection);
7592 memset(conn, 0, sizeof(*conn));
7594 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7595 &in, sizeof(in), conn, sizeof(*conn));
7598 *nextConnection += 1;
7601 * Convert old connection format to new structure.
7604 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7605 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7606 #define MOVEvL(a) (conn->a = vL->a)
7608 /* any old or unrecognized version... */
7609 for (i = 0; i < RX_MAXCALLS; i++) {
7610 MOVEvL(callState[i]);
7611 MOVEvL(callMode[i]);
7612 MOVEvL(callFlags[i]);
7613 MOVEvL(callOther[i]);
7615 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7616 MOVEvL(secStats.type);
7617 MOVEvL(secStats.level);
7618 MOVEvL(secStats.flags);
7619 MOVEvL(secStats.expires);
7620 MOVEvL(secStats.packetsReceived);
7621 MOVEvL(secStats.packetsSent);
7622 MOVEvL(secStats.bytesReceived);
7623 MOVEvL(secStats.bytesSent);
7628 * Do net to host conversion here
7630 * I don't convert host or port since we are most likely
7631 * going to want these in NBO.
7633 conn->cid = ntohl(conn->cid);
7634 conn->serial = ntohl(conn->serial);
7635 for (i = 0; i < RX_MAXCALLS; i++) {
7636 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7638 conn->error = ntohl(conn->error);
7639 conn->secStats.flags = ntohl(conn->secStats.flags);
7640 conn->secStats.expires = ntohl(conn->secStats.expires);
7641 conn->secStats.packetsReceived =
7642 ntohl(conn->secStats.packetsReceived);
7643 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7644 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7645 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7646 conn->epoch = ntohl(conn->epoch);
7647 conn->natMTU = ntohl(conn->natMTU);
7656 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7657 afs_uint16 remotePort, afs_int32 * nextPeer,
7658 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7659 afs_uint32 * supportedValues)
7661 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7663 struct rx_debugIn in;
7666 * supportedValues is currently unused, but added to allow future
7667 * versioning of this function.
7670 *supportedValues = 0;
7671 in.type = htonl(RX_DEBUGI_GETPEER);
7672 in.index = htonl(*nextPeer);
7673 memset(peer, 0, sizeof(*peer));
7675 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7676 &in, sizeof(in), peer, sizeof(*peer));
7682 * Do net to host conversion here
7684 * I don't convert host or port since we are most likely
7685 * going to want these in NBO.
7687 peer->ifMTU = ntohs(peer->ifMTU);
7688 peer->idleWhen = ntohl(peer->idleWhen);
7689 peer->refCount = ntohs(peer->refCount);
7690 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7691 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7692 peer->rtt = ntohl(peer->rtt);
7693 peer->rtt_dev = ntohl(peer->rtt_dev);
7694 peer->timeout.sec = 0;
7695 peer->timeout.usec = 0;
7696 peer->nSent = ntohl(peer->nSent);
7697 peer->reSends = ntohl(peer->reSends);
7698 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7699 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7700 peer->natMTU = ntohs(peer->natMTU);
7701 peer->maxMTU = ntohs(peer->maxMTU);
7702 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7703 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7704 peer->MTU = ntohs(peer->MTU);
7705 peer->cwind = ntohs(peer->cwind);
7706 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7707 peer->congestSeq = ntohs(peer->congestSeq);
7708 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7709 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7710 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7711 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7720 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7721 struct rx_debugPeer * peerStats)
7724 afs_int32 error = 1; /* default to "did not succeed" */
7725 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7727 MUTEX_ENTER(&rx_peerHashTable_lock);
7728 for(tp = rx_peerHashTable[hashValue];
7729 tp != NULL; tp = tp->next) {
7730 if (tp->host == peerHost)
7736 MUTEX_EXIT(&rx_peerHashTable_lock);
7740 MUTEX_ENTER(&tp->peer_lock);
7741 peerStats->host = tp->host;
7742 peerStats->port = tp->port;
7743 peerStats->ifMTU = tp->ifMTU;
7744 peerStats->idleWhen = tp->idleWhen;
7745 peerStats->refCount = tp->refCount;
7746 peerStats->burstSize = tp->burstSize;
7747 peerStats->burst = tp->burst;
7748 peerStats->burstWait.sec = tp->burstWait.sec;
7749 peerStats->burstWait.usec = tp->burstWait.usec;
7750 peerStats->rtt = tp->rtt;
7751 peerStats->rtt_dev = tp->rtt_dev;
7752 peerStats->timeout.sec = 0;
7753 peerStats->timeout.usec = 0;
7754 peerStats->nSent = tp->nSent;
7755 peerStats->reSends = tp->reSends;
7756 peerStats->inPacketSkew = tp->inPacketSkew;
7757 peerStats->outPacketSkew = tp->outPacketSkew;
7758 peerStats->natMTU = tp->natMTU;
7759 peerStats->maxMTU = tp->maxMTU;
7760 peerStats->maxDgramPackets = tp->maxDgramPackets;
7761 peerStats->ifDgramPackets = tp->ifDgramPackets;
7762 peerStats->MTU = tp->MTU;
7763 peerStats->cwind = tp->cwind;
7764 peerStats->nDgramPackets = tp->nDgramPackets;
7765 peerStats->congestSeq = tp->congestSeq;
7766 peerStats->bytesSent.high = tp->bytesSent.high;
7767 peerStats->bytesSent.low = tp->bytesSent.low;
7768 peerStats->bytesReceived.high = tp->bytesReceived.high;
7769 peerStats->bytesReceived.low = tp->bytesReceived.low;
7770 MUTEX_EXIT(&tp->peer_lock);
7772 MUTEX_ENTER(&rx_peerHashTable_lock);
7775 MUTEX_EXIT(&rx_peerHashTable_lock);
7783 struct rx_serverQueueEntry *np;
7786 struct rx_call *call;
7787 struct rx_serverQueueEntry *sq;
7791 if (rxinit_status == 1) {
7793 return; /* Already shutdown. */
7797 #ifndef AFS_PTHREAD_ENV
7798 FD_ZERO(&rx_selectMask);
7799 #endif /* AFS_PTHREAD_ENV */
7800 rxi_dataQuota = RX_MAX_QUOTA;
7801 #ifndef AFS_PTHREAD_ENV
7803 #endif /* AFS_PTHREAD_ENV */
7806 #ifndef AFS_PTHREAD_ENV
7807 #ifndef AFS_USE_GETTIMEOFDAY
7809 #endif /* AFS_USE_GETTIMEOFDAY */
7810 #endif /* AFS_PTHREAD_ENV */
7812 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7813 call = queue_First(&rx_freeCallQueue, rx_call);
7815 rxi_Free(call, sizeof(struct rx_call));
7818 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7819 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7825 struct rx_peer **peer_ptr, **peer_end;
7826 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7827 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7829 struct rx_peer *peer, *next;
7831 MUTEX_ENTER(&rx_peerHashTable_lock);
7832 for (peer = *peer_ptr; peer; peer = next) {
7833 rx_interface_stat_p rpc_stat, nrpc_stat;
7836 MUTEX_ENTER(&rx_rpc_stats);
7837 MUTEX_ENTER(&peer->peer_lock);
7839 (&peer->rpcStats, rpc_stat, nrpc_stat,
7840 rx_interface_stat)) {
7841 unsigned int num_funcs;
7844 queue_Remove(&rpc_stat->queue_header);
7845 queue_Remove(&rpc_stat->all_peers);
7846 num_funcs = rpc_stat->stats[0].func_total;
7848 sizeof(rx_interface_stat_t) +
7849 rpc_stat->stats[0].func_total *
7850 sizeof(rx_function_entry_v1_t);
7852 rxi_Free(rpc_stat, space);
7854 /* rx_rpc_stats must be held */
7855 rxi_rpc_peer_stat_cnt -= num_funcs;
7857 MUTEX_EXIT(&peer->peer_lock);
7858 MUTEX_EXIT(&rx_rpc_stats);
7862 if (rx_stats_active)
7863 rx_atomic_dec(&rx_stats.nPeerStructs);
7865 MUTEX_EXIT(&rx_peerHashTable_lock);
7868 for (i = 0; i < RX_MAX_SERVICES; i++) {
7870 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7872 for (i = 0; i < rx_hashTableSize; i++) {
7873 struct rx_connection *tc, *ntc;
7874 MUTEX_ENTER(&rx_connHashTable_lock);
7875 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7877 for (j = 0; j < RX_MAXCALLS; j++) {
7879 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7882 rxi_Free(tc, sizeof(*tc));
7884 MUTEX_EXIT(&rx_connHashTable_lock);
7887 MUTEX_ENTER(&freeSQEList_lock);
7889 while ((np = rx_FreeSQEList)) {
7890 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7891 MUTEX_DESTROY(&np->lock);
7892 rxi_Free(np, sizeof(*np));
7895 MUTEX_EXIT(&freeSQEList_lock);
7896 MUTEX_DESTROY(&freeSQEList_lock);
7897 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7898 MUTEX_DESTROY(&rx_connHashTable_lock);
7899 MUTEX_DESTROY(&rx_peerHashTable_lock);
7900 MUTEX_DESTROY(&rx_serverPool_lock);
7902 osi_Free(rx_connHashTable,
7903 rx_hashTableSize * sizeof(struct rx_connection *));
7904 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7906 UNPIN(rx_connHashTable,
7907 rx_hashTableSize * sizeof(struct rx_connection *));
7908 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7910 rxi_FreeAllPackets();
7912 MUTEX_ENTER(&rx_quota_mutex);
7913 rxi_dataQuota = RX_MAX_QUOTA;
7914 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7915 MUTEX_EXIT(&rx_quota_mutex);
7920 #ifdef RX_ENABLE_LOCKS
7922 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7924 if (!MUTEX_ISMINE(lockaddr))
7925 osi_Panic("Lock not held: %s", msg);
7927 #endif /* RX_ENABLE_LOCKS */
7932 * Routines to implement connection specific data.
7936 rx_KeyCreate(rx_destructor_t rtn)
7939 MUTEX_ENTER(&rxi_keyCreate_lock);
7940 key = rxi_keyCreate_counter++;
7941 rxi_keyCreate_destructor = (rx_destructor_t *)
7942 realloc((void *)rxi_keyCreate_destructor,
7943 (key + 1) * sizeof(rx_destructor_t));
7944 rxi_keyCreate_destructor[key] = rtn;
7945 MUTEX_EXIT(&rxi_keyCreate_lock);
7950 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7953 MUTEX_ENTER(&conn->conn_data_lock);
7954 if (!conn->specific) {
7955 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7956 for (i = 0; i < key; i++)
7957 conn->specific[i] = NULL;
7958 conn->nSpecific = key + 1;
7959 conn->specific[key] = ptr;
7960 } else if (key >= conn->nSpecific) {
7961 conn->specific = (void **)
7962 realloc(conn->specific, (key + 1) * sizeof(void *));
7963 for (i = conn->nSpecific; i < key; i++)
7964 conn->specific[i] = NULL;
7965 conn->nSpecific = key + 1;
7966 conn->specific[key] = ptr;
7968 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7969 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7970 conn->specific[key] = ptr;
7972 MUTEX_EXIT(&conn->conn_data_lock);
7976 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7979 MUTEX_ENTER(&svc->svc_data_lock);
7980 if (!svc->specific) {
7981 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
7982 for (i = 0; i < key; i++)
7983 svc->specific[i] = NULL;
7984 svc->nSpecific = key + 1;
7985 svc->specific[key] = ptr;
7986 } else if (key >= svc->nSpecific) {
7987 svc->specific = (void **)
7988 realloc(svc->specific, (key + 1) * sizeof(void *));
7989 for (i = svc->nSpecific; i < key; i++)
7990 svc->specific[i] = NULL;
7991 svc->nSpecific = key + 1;
7992 svc->specific[key] = ptr;
7994 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7995 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7996 svc->specific[key] = ptr;
7998 MUTEX_EXIT(&svc->svc_data_lock);
8002 rx_GetSpecific(struct rx_connection *conn, int key)
8005 MUTEX_ENTER(&conn->conn_data_lock);
8006 if (key >= conn->nSpecific)
8009 ptr = conn->specific[key];
8010 MUTEX_EXIT(&conn->conn_data_lock);
8015 rx_GetServiceSpecific(struct rx_service *svc, int key)
8018 MUTEX_ENTER(&svc->svc_data_lock);
8019 if (key >= svc->nSpecific)
8022 ptr = svc->specific[key];
8023 MUTEX_EXIT(&svc->svc_data_lock);
8028 #endif /* !KERNEL */
8031 * processStats is a queue used to store the statistics for the local
8032 * process. Its contents are similar to the contents of the rpcStats
8033 * queue on a rx_peer structure, but the actual data stored within
8034 * this queue contains totals across the lifetime of the process (assuming
8035 * the stats have not been reset) - unlike the per peer structures
8036 * which can come and go based upon the peer lifetime.
8039 static struct rx_queue processStats = { &processStats, &processStats };
8042 * peerStats is a queue used to store the statistics for all peer structs.
8043 * Its contents are the union of all the peer rpcStats queues.
8046 static struct rx_queue peerStats = { &peerStats, &peerStats };
8049 * rxi_monitor_processStats is used to turn process wide stat collection
8053 static int rxi_monitor_processStats = 0;
8056 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8059 static int rxi_monitor_peerStats = 0;
8062 * rxi_AddRpcStat - given all of the information for a particular rpc
8063 * call, create (if needed) and update the stat totals for the rpc.
8067 * IN stats - the queue of stats that will be updated with the new value
8069 * IN rxInterface - a unique number that identifies the rpc interface
8071 * IN currentFunc - the index of the function being invoked
8073 * IN totalFunc - the total number of functions in this interface
8075 * IN queueTime - the amount of time this function waited for a thread
8077 * IN execTime - the amount of time this function invocation took to execute
8079 * IN bytesSent - the number bytes sent by this invocation
8081 * IN bytesRcvd - the number bytes received by this invocation
8083 * IN isServer - if true, this invocation was made to a server
8085 * IN remoteHost - the ip address of the remote host
8087 * IN remotePort - the port of the remote host
8089 * IN addToPeerList - if != 0, add newly created stat to the global peer list
8091 * INOUT counter - if a new stats structure is allocated, the counter will
8092 * be updated with the new number of allocated stat structures
8100 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
8101 afs_uint32 currentFunc, afs_uint32 totalFunc,
8102 struct clock *queueTime, struct clock *execTime,
8103 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
8104 afs_uint32 remoteHost, afs_uint32 remotePort,
8105 int addToPeerList, unsigned int *counter)
8108 rx_interface_stat_p rpc_stat, nrpc_stat;
8111 * See if there's already a structure for this interface
8114 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8115 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8116 && (rpc_stat->stats[0].remote_is_server == isServer))
8121 * Didn't find a match so allocate a new structure and add it to the
8125 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
8126 || (rpc_stat->stats[0].interfaceId != rxInterface)
8127 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8132 sizeof(rx_interface_stat_t) +
8133 totalFunc * sizeof(rx_function_entry_v1_t);
8135 rpc_stat = rxi_Alloc(space);
8136 if (rpc_stat == NULL) {
8140 *counter += totalFunc;
8141 for (i = 0; i < totalFunc; i++) {
8142 rpc_stat->stats[i].remote_peer = remoteHost;
8143 rpc_stat->stats[i].remote_port = remotePort;
8144 rpc_stat->stats[i].remote_is_server = isServer;
8145 rpc_stat->stats[i].interfaceId = rxInterface;
8146 rpc_stat->stats[i].func_total = totalFunc;
8147 rpc_stat->stats[i].func_index = i;
8148 hzero(rpc_stat->stats[i].invocations);
8149 hzero(rpc_stat->stats[i].bytes_sent);
8150 hzero(rpc_stat->stats[i].bytes_rcvd);
8151 rpc_stat->stats[i].queue_time_sum.sec = 0;
8152 rpc_stat->stats[i].queue_time_sum.usec = 0;
8153 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8154 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8155 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8156 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8157 rpc_stat->stats[i].queue_time_max.sec = 0;
8158 rpc_stat->stats[i].queue_time_max.usec = 0;
8159 rpc_stat->stats[i].execution_time_sum.sec = 0;
8160 rpc_stat->stats[i].execution_time_sum.usec = 0;
8161 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8162 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8163 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8164 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8165 rpc_stat->stats[i].execution_time_max.sec = 0;
8166 rpc_stat->stats[i].execution_time_max.usec = 0;
8168 queue_Prepend(stats, rpc_stat);
8169 if (addToPeerList) {
8170 queue_Prepend(&peerStats, &rpc_stat->all_peers);
8175 * Increment the stats for this function
8178 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
8179 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
8180 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
8181 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8182 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8183 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8184 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8186 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8187 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8189 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8190 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8192 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8193 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8195 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8196 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8204 * rx_IncrementTimeAndCount - increment the times and count for a particular
8209 * IN peer - the peer who invoked the rpc
8211 * IN rxInterface - a unique number that identifies the rpc interface
8213 * IN currentFunc - the index of the function being invoked
8215 * IN totalFunc - the total number of functions in this interface
8217 * IN queueTime - the amount of time this function waited for a thread
8219 * IN execTime - the amount of time this function invocation took to execute
8221 * IN bytesSent - the number bytes sent by this invocation
8223 * IN bytesRcvd - the number bytes received by this invocation
8225 * IN isServer - if true, this invocation was made to a server
8233 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8234 afs_uint32 currentFunc, afs_uint32 totalFunc,
8235 struct clock *queueTime, struct clock *execTime,
8236 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8240 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8243 MUTEX_ENTER(&rx_rpc_stats);
8245 if (rxi_monitor_peerStats) {
8246 MUTEX_ENTER(&peer->peer_lock);
8247 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8248 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8249 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8250 MUTEX_EXIT(&peer->peer_lock);
8253 if (rxi_monitor_processStats) {
8254 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8255 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8256 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8259 MUTEX_EXIT(&rx_rpc_stats);
8264 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8268 * IN callerVersion - the rpc stat version of the caller.
8270 * IN count - the number of entries to marshall.
8272 * IN stats - pointer to stats to be marshalled.
8274 * OUT ptr - Where to store the marshalled data.
8281 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8282 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8288 * We only support the first version
8290 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8291 *(ptr++) = stats->remote_peer;
8292 *(ptr++) = stats->remote_port;
8293 *(ptr++) = stats->remote_is_server;
8294 *(ptr++) = stats->interfaceId;
8295 *(ptr++) = stats->func_total;
8296 *(ptr++) = stats->func_index;
8297 *(ptr++) = hgethi(stats->invocations);
8298 *(ptr++) = hgetlo(stats->invocations);
8299 *(ptr++) = hgethi(stats->bytes_sent);
8300 *(ptr++) = hgetlo(stats->bytes_sent);
8301 *(ptr++) = hgethi(stats->bytes_rcvd);
8302 *(ptr++) = hgetlo(stats->bytes_rcvd);
8303 *(ptr++) = stats->queue_time_sum.sec;
8304 *(ptr++) = stats->queue_time_sum.usec;
8305 *(ptr++) = stats->queue_time_sum_sqr.sec;
8306 *(ptr++) = stats->queue_time_sum_sqr.usec;
8307 *(ptr++) = stats->queue_time_min.sec;
8308 *(ptr++) = stats->queue_time_min.usec;
8309 *(ptr++) = stats->queue_time_max.sec;
8310 *(ptr++) = stats->queue_time_max.usec;
8311 *(ptr++) = stats->execution_time_sum.sec;
8312 *(ptr++) = stats->execution_time_sum.usec;
8313 *(ptr++) = stats->execution_time_sum_sqr.sec;
8314 *(ptr++) = stats->execution_time_sum_sqr.usec;
8315 *(ptr++) = stats->execution_time_min.sec;
8316 *(ptr++) = stats->execution_time_min.usec;
8317 *(ptr++) = stats->execution_time_max.sec;
8318 *(ptr++) = stats->execution_time_max.usec;
8324 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8329 * IN callerVersion - the rpc stat version of the caller
8331 * OUT myVersion - the rpc stat version of this function
8333 * OUT clock_sec - local time seconds
8335 * OUT clock_usec - local time microseconds
8337 * OUT allocSize - the number of bytes allocated to contain stats
8339 * OUT statCount - the number stats retrieved from this process.
8341 * OUT stats - the actual stats retrieved from this process.
8345 * Returns void. If successful, stats will != NULL.
8349 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8350 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8351 size_t * allocSize, afs_uint32 * statCount,
8352 afs_uint32 ** stats)
8362 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8365 * Check to see if stats are enabled
8368 MUTEX_ENTER(&rx_rpc_stats);
8369 if (!rxi_monitor_processStats) {
8370 MUTEX_EXIT(&rx_rpc_stats);
8374 clock_GetTime(&now);
8375 *clock_sec = now.sec;
8376 *clock_usec = now.usec;
8379 * Allocate the space based upon the caller version
8381 * If the client is at an older version than we are,
8382 * we return the statistic data in the older data format, but
8383 * we still return our version number so the client knows we
8384 * are maintaining more data than it can retrieve.
8387 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8388 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8389 *statCount = rxi_rpc_process_stat_cnt;
8392 * This can't happen yet, but in the future version changes
8393 * can be handled by adding additional code here
8397 if (space > (size_t) 0) {
8399 ptr = *stats = rxi_Alloc(space);
8402 rx_interface_stat_p rpc_stat, nrpc_stat;
8406 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8408 * Copy the data based upon the caller version
8410 rx_MarshallProcessRPCStats(callerVersion,
8411 rpc_stat->stats[0].func_total,
8412 rpc_stat->stats, &ptr);
8418 MUTEX_EXIT(&rx_rpc_stats);
8423 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8427 * IN callerVersion - the rpc stat version of the caller
8429 * OUT myVersion - the rpc stat version of this function
8431 * OUT clock_sec - local time seconds
8433 * OUT clock_usec - local time microseconds
8435 * OUT allocSize - the number of bytes allocated to contain stats
8437 * OUT statCount - the number of stats retrieved from the individual
8440 * OUT stats - the actual stats retrieved from the individual peer structures.
8444 * Returns void. If successful, stats will != NULL.
8448 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8449 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8450 size_t * allocSize, afs_uint32 * statCount,
8451 afs_uint32 ** stats)
8461 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8464 * Check to see if stats are enabled
8467 MUTEX_ENTER(&rx_rpc_stats);
8468 if (!rxi_monitor_peerStats) {
8469 MUTEX_EXIT(&rx_rpc_stats);
8473 clock_GetTime(&now);
8474 *clock_sec = now.sec;
8475 *clock_usec = now.usec;
8478 * Allocate the space based upon the caller version
8480 * If the client is at an older version than we are,
8481 * we return the statistic data in the older data format, but
8482 * we still return our version number so the client knows we
8483 * are maintaining more data than it can retrieve.
8486 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8487 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8488 *statCount = rxi_rpc_peer_stat_cnt;
8491 * This can't happen yet, but in the future version changes
8492 * can be handled by adding additional code here
8496 if (space > (size_t) 0) {
8498 ptr = *stats = rxi_Alloc(space);
8501 rx_interface_stat_p rpc_stat, nrpc_stat;
8505 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8507 * We have to fix the offset of rpc_stat since we are
8508 * keeping this structure on two rx_queues. The rx_queue
8509 * package assumes that the rx_queue member is the first
8510 * member of the structure. That is, rx_queue assumes that
8511 * any one item is only on one queue at a time. We are
8512 * breaking that assumption and so we have to do a little
8513 * math to fix our pointers.
8516 fix_offset = (char *)rpc_stat;
8517 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8518 rpc_stat = (rx_interface_stat_p) fix_offset;
8521 * Copy the data based upon the caller version
8523 rx_MarshallProcessRPCStats(callerVersion,
8524 rpc_stat->stats[0].func_total,
8525 rpc_stat->stats, &ptr);
8531 MUTEX_EXIT(&rx_rpc_stats);
8536 * rx_FreeRPCStats - free memory allocated by
8537 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8541 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8542 * rx_RetrievePeerRPCStats
8544 * IN allocSize - the number of bytes in stats.
8552 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8554 rxi_Free(stats, allocSize);
8558 * rx_queryProcessRPCStats - see if process rpc stat collection is
8559 * currently enabled.
8565 * Returns 0 if stats are not enabled != 0 otherwise
8569 rx_queryProcessRPCStats(void)
8572 MUTEX_ENTER(&rx_rpc_stats);
8573 rc = rxi_monitor_processStats;
8574 MUTEX_EXIT(&rx_rpc_stats);
8579 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8585 * Returns 0 if stats are not enabled != 0 otherwise
8589 rx_queryPeerRPCStats(void)
8592 MUTEX_ENTER(&rx_rpc_stats);
8593 rc = rxi_monitor_peerStats;
8594 MUTEX_EXIT(&rx_rpc_stats);
8599 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8609 rx_enableProcessRPCStats(void)
8611 MUTEX_ENTER(&rx_rpc_stats);
8612 rx_enable_stats = 1;
8613 rxi_monitor_processStats = 1;
8614 MUTEX_EXIT(&rx_rpc_stats);
8618 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8628 rx_enablePeerRPCStats(void)
8630 MUTEX_ENTER(&rx_rpc_stats);
8631 rx_enable_stats = 1;
8632 rxi_monitor_peerStats = 1;
8633 MUTEX_EXIT(&rx_rpc_stats);
8637 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8647 rx_disableProcessRPCStats(void)
8649 rx_interface_stat_p rpc_stat, nrpc_stat;
8652 MUTEX_ENTER(&rx_rpc_stats);
8655 * Turn off process statistics and if peer stats is also off, turn
8659 rxi_monitor_processStats = 0;
8660 if (rxi_monitor_peerStats == 0) {
8661 rx_enable_stats = 0;
8664 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8665 unsigned int num_funcs = 0;
8668 queue_Remove(rpc_stat);
8669 num_funcs = rpc_stat->stats[0].func_total;
8671 sizeof(rx_interface_stat_t) +
8672 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8674 rxi_Free(rpc_stat, space);
8675 rxi_rpc_process_stat_cnt -= num_funcs;
8677 MUTEX_EXIT(&rx_rpc_stats);
8681 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8691 rx_disablePeerRPCStats(void)
8693 struct rx_peer **peer_ptr, **peer_end;
8697 * Turn off peer statistics and if process stats is also off, turn
8701 rxi_monitor_peerStats = 0;
8702 if (rxi_monitor_processStats == 0) {
8703 rx_enable_stats = 0;
8706 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8707 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8709 struct rx_peer *peer, *next, *prev;
8711 MUTEX_ENTER(&rx_peerHashTable_lock);
8712 MUTEX_ENTER(&rx_rpc_stats);
8713 for (prev = peer = *peer_ptr; peer; peer = next) {
8715 code = MUTEX_TRYENTER(&peer->peer_lock);
8717 rx_interface_stat_p rpc_stat, nrpc_stat;
8720 if (prev == *peer_ptr) {
8731 MUTEX_EXIT(&rx_peerHashTable_lock);
8734 (&peer->rpcStats, rpc_stat, nrpc_stat,
8735 rx_interface_stat)) {
8736 unsigned int num_funcs = 0;
8739 queue_Remove(&rpc_stat->queue_header);
8740 queue_Remove(&rpc_stat->all_peers);
8741 num_funcs = rpc_stat->stats[0].func_total;
8743 sizeof(rx_interface_stat_t) +
8744 rpc_stat->stats[0].func_total *
8745 sizeof(rx_function_entry_v1_t);
8747 rxi_Free(rpc_stat, space);
8748 rxi_rpc_peer_stat_cnt -= num_funcs;
8750 MUTEX_EXIT(&peer->peer_lock);
8752 MUTEX_ENTER(&rx_peerHashTable_lock);
8762 MUTEX_EXIT(&rx_rpc_stats);
8763 MUTEX_EXIT(&rx_peerHashTable_lock);
8768 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8773 * IN clearFlag - flag indicating which stats to clear
8781 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8783 rx_interface_stat_p rpc_stat, nrpc_stat;
8785 MUTEX_ENTER(&rx_rpc_stats);
8787 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8788 unsigned int num_funcs = 0, i;
8789 num_funcs = rpc_stat->stats[0].func_total;
8790 for (i = 0; i < num_funcs; i++) {
8791 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8792 hzero(rpc_stat->stats[i].invocations);
8794 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8795 hzero(rpc_stat->stats[i].bytes_sent);
8797 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8798 hzero(rpc_stat->stats[i].bytes_rcvd);
8800 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8801 rpc_stat->stats[i].queue_time_sum.sec = 0;
8802 rpc_stat->stats[i].queue_time_sum.usec = 0;
8804 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8805 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8806 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8808 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8809 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8810 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8812 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8813 rpc_stat->stats[i].queue_time_max.sec = 0;
8814 rpc_stat->stats[i].queue_time_max.usec = 0;
8816 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8817 rpc_stat->stats[i].execution_time_sum.sec = 0;
8818 rpc_stat->stats[i].execution_time_sum.usec = 0;
8820 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8821 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8822 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8824 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8825 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8826 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8828 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8829 rpc_stat->stats[i].execution_time_max.sec = 0;
8830 rpc_stat->stats[i].execution_time_max.usec = 0;
8835 MUTEX_EXIT(&rx_rpc_stats);
8839 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8844 * IN clearFlag - flag indicating which stats to clear
8852 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8854 rx_interface_stat_p rpc_stat, nrpc_stat;
8856 MUTEX_ENTER(&rx_rpc_stats);
8858 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8859 unsigned int num_funcs = 0, i;
8862 * We have to fix the offset of rpc_stat since we are
8863 * keeping this structure on two rx_queues. The rx_queue
8864 * package assumes that the rx_queue member is the first
8865 * member of the structure. That is, rx_queue assumes that
8866 * any one item is only on one queue at a time. We are
8867 * breaking that assumption and so we have to do a little
8868 * math to fix our pointers.
8871 fix_offset = (char *)rpc_stat;
8872 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8873 rpc_stat = (rx_interface_stat_p) fix_offset;
8875 num_funcs = rpc_stat->stats[0].func_total;
8876 for (i = 0; i < num_funcs; i++) {
8877 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8878 hzero(rpc_stat->stats[i].invocations);
8880 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8881 hzero(rpc_stat->stats[i].bytes_sent);
8883 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8884 hzero(rpc_stat->stats[i].bytes_rcvd);
8886 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8887 rpc_stat->stats[i].queue_time_sum.sec = 0;
8888 rpc_stat->stats[i].queue_time_sum.usec = 0;
8890 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8891 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8892 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8894 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8895 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8896 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8898 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8899 rpc_stat->stats[i].queue_time_max.sec = 0;
8900 rpc_stat->stats[i].queue_time_max.usec = 0;
8902 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8903 rpc_stat->stats[i].execution_time_sum.sec = 0;
8904 rpc_stat->stats[i].execution_time_sum.usec = 0;
8906 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8907 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8908 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8910 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8911 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8912 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8914 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8915 rpc_stat->stats[i].execution_time_max.sec = 0;
8916 rpc_stat->stats[i].execution_time_max.usec = 0;
8921 MUTEX_EXIT(&rx_rpc_stats);
8925 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8926 * is authorized to enable/disable/clear RX statistics.
8928 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8931 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8933 rxi_rxstat_userok = proc;
8937 rx_RxStatUserOk(struct rx_call *call)
8939 if (!rxi_rxstat_userok)
8941 return rxi_rxstat_userok(call);
8946 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8947 * function in the MSVC runtime DLL (msvcrt.dll).
8949 * Note: the system serializes calls to this function.
8952 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8953 DWORD reason, /* reason function is being called */
8954 LPVOID reserved) /* reserved for future use */
8957 case DLL_PROCESS_ATTACH:
8958 /* library is being attached to a process */
8962 case DLL_PROCESS_DETACH:
8969 #endif /* AFS_NT40_ENV */
8972 int rx_DumpCalls(FILE *outputFile, char *cookie)
8974 #ifdef RXDEBUG_PACKET
8975 #ifdef KDUMP_RX_LOCK
8976 struct rx_call_rx_lock *c;
8983 #define RXDPRINTF sprintf
8984 #define RXDPRINTOUT output
8986 #define RXDPRINTF fprintf
8987 #define RXDPRINTOUT outputFile
8990 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8992 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8995 for (c = rx_allCallsp; c; c = c->allNextp) {
8996 u_short rqc, tqc, iovqc;
8997 struct rx_packet *p, *np;
8999 MUTEX_ENTER(&c->lock);
9000 queue_Count(&c->rq, p, np, rx_packet, rqc);
9001 queue_Count(&c->tq, p, np, rx_packet, tqc);
9002 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
9004 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, "
9005 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9006 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9007 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9008 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
9009 #ifdef RX_ENABLE_LOCKS
9012 #ifdef RX_REFCOUNT_CHECK
9013 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9014 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9017 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,
9018 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9019 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9020 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9021 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9022 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
9023 #ifdef RX_ENABLE_LOCKS
9024 , (afs_uint32)c->refCount
9026 #ifdef RX_REFCOUNT_CHECK
9027 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9030 MUTEX_EXIT(&c->lock);
9033 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9036 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9038 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9040 #endif /* RXDEBUG_PACKET */