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"
82 #include <afs/rxgen_consts.h>
85 #ifdef AFS_PTHREAD_ENV
87 int (*registerProgram) (pid_t, char *) = 0;
88 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
91 int (*registerProgram) (PROCESS, char *) = 0;
92 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
96 /* Local static routines */
97 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
98 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
99 struct rx_call *, struct rx_peer *,
101 static void rxi_Resend(struct rxevent *event, void *arg0, void *arg1,
104 #ifdef RX_ENABLE_LOCKS
105 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
108 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
110 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
111 rx_atomic_t rxi_start_in_error;
113 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
115 /* Constant delay time before sending an acknowledge of the last packet
116 * received. This is to avoid sending an extra acknowledge when the
117 * client is about to make another call, anyway, or the server is
120 * The lastAckDelay may not exceeed 400ms without causing peers to
121 * unecessarily timeout.
123 struct clock rx_lastAckDelay = {0, 400000};
125 /* Constant delay time before sending a soft ack when none was requested.
126 * This is to make sure we send soft acks before the sender times out,
127 * Normally we wait and send a hard ack when the receiver consumes the packet
129 * This value has been 100ms in all shipping versions of OpenAFS. Changing it
130 * will require changes to the peer's RTT calculations.
132 struct clock rx_softAckDelay = {0, 100000};
135 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
136 * currently allocated within rx. This number is used to allocate the
137 * memory required to return the statistics when queried.
138 * Protected by the rx_rpc_stats mutex.
141 static unsigned int rxi_rpc_peer_stat_cnt;
144 * rxi_rpc_process_stat_cnt counts the total number of local process stat
145 * structures currently allocated within rx. The number is used to allocate
146 * the memory required to return the statistics when queried.
147 * Protected by the rx_rpc_stats mutex.
150 static unsigned int rxi_rpc_process_stat_cnt;
153 * rxi_busyChannelError is the error to return to the application when a call
154 * channel appears busy (inferred from the receipt of RX_PACKET_TYPE_BUSY
155 * packets on the channel), and there are other call channels in the
156 * connection that are not busy. If 0, we do not return errors upon receiving
157 * busy packets; we just keep trying on the same call channel until we hit a
160 static afs_int32 rxi_busyChannelError = 0;
162 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
163 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
165 #if !defined(offsetof)
166 #include <stddef.h> /* for definition of offsetof() */
169 #ifdef RX_ENABLE_LOCKS
170 afs_kmutex_t rx_atomic_mutex;
173 /* Forward prototypes */
174 static struct rx_call * rxi_NewCall(struct rx_connection *, int);
176 #ifdef AFS_PTHREAD_ENV
179 * Use procedural initialization of mutexes/condition variables
183 extern afs_kmutex_t rx_quota_mutex;
184 extern afs_kmutex_t rx_pthread_mutex;
185 extern afs_kmutex_t rx_packets_mutex;
186 extern afs_kmutex_t rx_refcnt_mutex;
187 extern afs_kmutex_t des_init_mutex;
188 extern afs_kmutex_t des_random_mutex;
189 extern afs_kmutex_t rx_clock_mutex;
190 extern afs_kmutex_t rxi_connCacheMutex;
191 extern afs_kmutex_t rx_event_mutex;
192 extern afs_kmutex_t event_handler_mutex;
193 extern afs_kmutex_t listener_mutex;
194 extern afs_kmutex_t rx_if_init_mutex;
195 extern afs_kmutex_t rx_if_mutex;
197 extern afs_kcondvar_t rx_event_handler_cond;
198 extern afs_kcondvar_t rx_listener_cond;
200 static afs_kmutex_t epoch_mutex;
201 static afs_kmutex_t rx_init_mutex;
202 static afs_kmutex_t rx_debug_mutex;
203 static afs_kmutex_t rx_rpc_stats;
206 rxi_InitPthread(void)
208 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
209 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
210 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
211 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
212 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
213 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
214 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
215 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
216 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
217 MUTEX_INIT(&rx_event_mutex, "event", MUTEX_DEFAULT, 0);
218 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
219 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
220 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
221 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
222 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
223 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
225 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
226 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
228 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
229 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
231 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
232 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
233 #ifdef RX_ENABLE_LOCKS
236 #endif /* RX_LOCKS_DB */
237 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
238 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
240 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
242 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
244 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
246 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
247 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
248 #endif /* RX_ENABLE_LOCKS */
251 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
252 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
254 * The rx_stats_mutex mutex protects the following global variables:
255 * rxi_lowConnRefCount
256 * rxi_lowPeerRefCount
265 * The rx_quota_mutex mutex protects the following global variables:
273 * The rx_freePktQ_lock protects the following global variables:
278 * The rx_packets_mutex mutex protects the following global variables:
286 * The rx_pthread_mutex mutex protects the following global variables:
287 * rxi_fcfs_thread_num
290 #define INIT_PTHREAD_LOCKS
294 /* Variables for handling the minProcs implementation. availProcs gives the
295 * number of threads available in the pool at this moment (not counting dudes
296 * executing right now). totalMin gives the total number of procs required
297 * for handling all minProcs requests. minDeficit is a dynamic variable
298 * tracking the # of procs required to satisfy all of the remaining minProcs
300 * For fine grain locking to work, the quota check and the reservation of
301 * a server thread has to come while rxi_availProcs and rxi_minDeficit
302 * are locked. To this end, the code has been modified under #ifdef
303 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
304 * same time. A new function, ReturnToServerPool() returns the allocation.
306 * A call can be on several queue's (but only one at a time). When
307 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
308 * that no one else is touching the queue. To this end, we store the address
309 * of the queue lock in the call structure (under the call lock) when we
310 * put the call on a queue, and we clear the call_queue_lock when the
311 * call is removed from a queue (once the call lock has been obtained).
312 * This allows rxi_ResetCall to safely synchronize with others wishing
313 * to manipulate the queue.
316 #if defined(RX_ENABLE_LOCKS)
317 static afs_kmutex_t rx_rpc_stats;
320 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
321 ** pretty good that the next packet coming in is from the same connection
322 ** as the last packet, since we're send multiple packets in a transmit window.
324 struct rx_connection *rxLastConn = 0;
326 #ifdef RX_ENABLE_LOCKS
327 /* The locking hierarchy for rx fine grain locking is composed of these
330 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
331 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
332 * call->lock - locks call data fields.
333 * These are independent of each other:
334 * rx_freeCallQueue_lock
339 * serverQueueEntry->lock
340 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
342 * peer->lock - locks peer data fields.
343 * conn_data_lock - that more than one thread is not updating a conn data
344 * field at the same time.
355 * Do we need a lock to protect the peer field in the conn structure?
356 * conn->peer was previously a constant for all intents and so has no
357 * lock protecting this field. The multihomed client delta introduced
358 * a RX code change : change the peer field in the connection structure
359 * to that remote interface from which the last packet for this
360 * connection was sent out. This may become an issue if further changes
363 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
364 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
366 /* rxdb_fileID is used to identify the lock location, along with line#. */
367 static int rxdb_fileID = RXDB_FILE_RX;
368 #endif /* RX_LOCKS_DB */
369 #else /* RX_ENABLE_LOCKS */
370 #define SET_CALL_QUEUE_LOCK(C, L)
371 #define CLEAR_CALL_QUEUE_LOCK(C)
372 #endif /* RX_ENABLE_LOCKS */
373 struct rx_serverQueueEntry *rx_waitForPacket = 0;
374 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
376 /* ------------Exported Interfaces------------- */
378 /* This function allows rxkad to set the epoch to a suitably random number
379 * which rx_NewConnection will use in the future. The principle purpose is to
380 * get rxnull connections to use the same epoch as the rxkad connections do, at
381 * least once the first rxkad connection is established. This is important now
382 * that the host/port addresses aren't used in FindConnection: the uniqueness
383 * of epoch/cid matters and the start time won't do. */
385 #ifdef AFS_PTHREAD_ENV
387 * This mutex protects the following global variables:
391 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
392 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
396 #endif /* AFS_PTHREAD_ENV */
399 rx_SetEpoch(afs_uint32 epoch)
406 /* Initialize rx. A port number may be mentioned, in which case this
407 * becomes the default port number for any service installed later.
408 * If 0 is provided for the port number, a random port will be chosen
409 * by the kernel. Whether this will ever overlap anything in
410 * /etc/services is anybody's guess... Returns 0 on success, -1 on
415 int rxinit_status = 1;
416 #ifdef AFS_PTHREAD_ENV
418 * This mutex protects the following global variables:
422 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
423 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
426 #define UNLOCK_RX_INIT
430 rx_InitHost(u_int host, u_int port)
437 char *htable, *ptable;
444 if (rxinit_status == 0) {
445 tmp_status = rxinit_status;
447 return tmp_status; /* Already started; return previous error code. */
453 if (afs_winsockInit() < 0)
459 * Initialize anything necessary to provide a non-premptive threading
462 rxi_InitializeThreadSupport();
465 /* Allocate and initialize a socket for client and perhaps server
468 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
469 if (rx_socket == OSI_NULLSOCKET) {
473 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
476 #endif /* RX_LOCKS_DB */
477 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
478 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
479 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
480 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
481 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
482 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
483 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
484 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
485 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
487 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
489 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
491 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
493 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
494 #if defined(AFS_HPUX110_ENV)
496 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
497 #endif /* AFS_HPUX110_ENV */
498 #endif /* RX_ENABLE_LOCKS && KERNEL */
501 rx_connDeadTime = 12;
502 rx_tranquil = 0; /* reset flag */
503 rxi_ResetStatistics();
505 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
506 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
507 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
508 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
509 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
510 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
512 /* Malloc up a bunch of packets & buffers */
514 queue_Init(&rx_freePacketQueue);
515 rxi_NeedMorePackets = FALSE;
516 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
518 /* enforce a minimum number of allocated packets */
519 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
520 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
522 /* allocate the initial free packet pool */
523 #ifdef RX_ENABLE_TSFPQ
524 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
525 #else /* RX_ENABLE_TSFPQ */
526 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
527 #endif /* RX_ENABLE_TSFPQ */
534 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
535 tv.tv_sec = clock_now.sec;
536 tv.tv_usec = clock_now.usec;
537 srand((unsigned int)tv.tv_usec);
544 #if defined(KERNEL) && !defined(UKERNEL)
545 /* Really, this should never happen in a real kernel */
548 struct sockaddr_in addr;
550 int addrlen = sizeof(addr);
552 socklen_t addrlen = sizeof(addr);
554 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
558 rx_port = addr.sin_port;
561 rx_stats.minRtt.sec = 9999999;
563 rx_SetEpoch(tv.tv_sec | 0x80000000);
565 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
566 * will provide a randomer value. */
568 MUTEX_ENTER(&rx_quota_mutex);
569 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
570 MUTEX_EXIT(&rx_quota_mutex);
571 /* *Slightly* random start time for the cid. This is just to help
572 * out with the hashing function at the peer */
573 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
574 rx_connHashTable = (struct rx_connection **)htable;
575 rx_peerHashTable = (struct rx_peer **)ptable;
577 rx_hardAckDelay.sec = 0;
578 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
580 rxevent_Init(20, rxi_ReScheduleEvents);
582 /* Initialize various global queues */
583 queue_Init(&rx_idleServerQueue);
584 queue_Init(&rx_incomingCallQueue);
585 queue_Init(&rx_freeCallQueue);
587 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
588 /* Initialize our list of usable IP addresses. */
592 #if defined(RXK_LISTENER_ENV) || !defined(KERNEL)
593 /* Start listener process (exact function is dependent on the
594 * implementation environment--kernel or user space) */
599 tmp_status = rxinit_status = 0;
607 return rx_InitHost(htonl(INADDR_ANY), port);
613 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
614 * maintaing the round trip timer.
619 * Start a new RTT timer for a given call and packet.
621 * There must be no resendEvent already listed for this call, otherwise this
622 * will leak events - intended for internal use within the RTO code only
625 * the RX call to start the timer for
626 * @param[in] lastPacket
627 * a flag indicating whether the last packet has been sent or not
629 * @pre call must be locked before calling this function
633 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
635 struct clock now, retryTime;
640 clock_Add(&retryTime, &call->rto);
642 /* If we're sending the last packet, and we're the client, then the server
643 * may wait for an additional 400ms before returning the ACK, wait for it
644 * rather than hitting a timeout */
645 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
646 clock_Addmsec(&retryTime, 400);
648 MUTEX_ENTER(&rx_refcnt_mutex);
649 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
650 MUTEX_EXIT(&rx_refcnt_mutex);
651 call->resendEvent = rxevent_PostNow2(&retryTime, &now, rxi_Resend,
656 * Cancel an RTT timer for a given call.
660 * the RX call to cancel the timer for
662 * @pre call must be locked before calling this function
667 rxi_rto_cancel(struct rx_call *call)
669 if (!call->resendEvent)
672 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
676 * Tell the RTO timer that we have sent a packet.
678 * If the timer isn't already running, then start it. If the timer is running,
682 * the RX call that the packet has been sent on
683 * @param[in] lastPacket
684 * A flag which is true if this is the last packet for the call
686 * @pre The call must be locked before calling this function
691 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
693 if (call->resendEvent)
696 rxi_rto_startTimer(call, lastPacket, istack);
700 * Tell the RTO timer that we have received an new ACK message
702 * This function should be called whenever a call receives an ACK that
703 * acknowledges new packets. Whatever happens, we stop the current timer.
704 * If there are unacked packets in the queue which have been sent, then
705 * we restart the timer from now. Otherwise, we leave it stopped.
708 * the RX call that the ACK has been received on
712 rxi_rto_packet_acked(struct rx_call *call, int istack)
714 struct rx_packet *p, *nxp;
716 rxi_rto_cancel(call);
718 if (queue_IsEmpty(&call->tq))
721 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
722 if (p->header.seq > call->tfirst + call->twind)
725 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
726 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
734 * Set an initial round trip timeout for a peer connection
736 * @param[in] secs The timeout to set in seconds
740 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
741 peer->rtt = secs * 8000;
745 * Sets the error generated when a busy call channel is detected.
747 * @param[in] error The error to return for a call on a busy channel.
749 * @pre Neither rx_Init nor rx_InitHost have been called yet
752 rx_SetBusyChannelError(afs_int32 error)
754 osi_Assert(rxinit_status != 0);
755 rxi_busyChannelError = error;
758 /* called with unincremented nRequestsRunning to see if it is OK to start
759 * a new thread in this service. Could be "no" for two reasons: over the
760 * max quota, or would prevent others from reaching their min quota.
762 #ifdef RX_ENABLE_LOCKS
763 /* This verion of QuotaOK reserves quota if it's ok while the
764 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
767 QuotaOK(struct rx_service *aservice)
769 /* check if over max quota */
770 if (aservice->nRequestsRunning >= aservice->maxProcs) {
774 /* under min quota, we're OK */
775 /* otherwise, can use only if there are enough to allow everyone
776 * to go to their min quota after this guy starts.
779 MUTEX_ENTER(&rx_quota_mutex);
780 if ((aservice->nRequestsRunning < aservice->minProcs)
781 || (rxi_availProcs > rxi_minDeficit)) {
782 aservice->nRequestsRunning++;
783 /* just started call in minProcs pool, need fewer to maintain
785 if (aservice->nRequestsRunning <= aservice->minProcs)
788 MUTEX_EXIT(&rx_quota_mutex);
791 MUTEX_EXIT(&rx_quota_mutex);
797 ReturnToServerPool(struct rx_service *aservice)
799 aservice->nRequestsRunning--;
800 MUTEX_ENTER(&rx_quota_mutex);
801 if (aservice->nRequestsRunning < aservice->minProcs)
804 MUTEX_EXIT(&rx_quota_mutex);
807 #else /* RX_ENABLE_LOCKS */
809 QuotaOK(struct rx_service *aservice)
812 /* under min quota, we're OK */
813 if (aservice->nRequestsRunning < aservice->minProcs)
816 /* check if over max quota */
817 if (aservice->nRequestsRunning >= aservice->maxProcs)
820 /* otherwise, can use only if there are enough to allow everyone
821 * to go to their min quota after this guy starts.
823 MUTEX_ENTER(&rx_quota_mutex);
824 if (rxi_availProcs > rxi_minDeficit)
826 MUTEX_EXIT(&rx_quota_mutex);
829 #endif /* RX_ENABLE_LOCKS */
832 /* Called by rx_StartServer to start up lwp's to service calls.
833 NExistingProcs gives the number of procs already existing, and which
834 therefore needn't be created. */
836 rxi_StartServerProcs(int nExistingProcs)
838 struct rx_service *service;
843 /* For each service, reserve N processes, where N is the "minimum"
844 * number of processes that MUST be able to execute a request in parallel,
845 * at any time, for that process. Also compute the maximum difference
846 * between any service's maximum number of processes that can run
847 * (i.e. the maximum number that ever will be run, and a guarantee
848 * that this number will run if other services aren't running), and its
849 * minimum number. The result is the extra number of processes that
850 * we need in order to provide the latter guarantee */
851 for (i = 0; i < RX_MAX_SERVICES; i++) {
853 service = rx_services[i];
854 if (service == (struct rx_service *)0)
856 nProcs += service->minProcs;
857 diff = service->maxProcs - service->minProcs;
861 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
862 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
863 for (i = 0; i < nProcs; i++) {
864 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
870 /* This routine is only required on Windows */
872 rx_StartClientThread(void)
874 #ifdef AFS_PTHREAD_ENV
876 pid = pthread_self();
877 #endif /* AFS_PTHREAD_ENV */
879 #endif /* AFS_NT40_ENV */
881 /* This routine must be called if any services are exported. If the
882 * donateMe flag is set, the calling process is donated to the server
885 rx_StartServer(int donateMe)
887 struct rx_service *service;
893 /* Start server processes, if necessary (exact function is dependent
894 * on the implementation environment--kernel or user space). DonateMe
895 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
896 * case, one less new proc will be created rx_StartServerProcs.
898 rxi_StartServerProcs(donateMe);
900 /* count up the # of threads in minProcs, and add set the min deficit to
901 * be that value, too.
903 for (i = 0; i < RX_MAX_SERVICES; i++) {
904 service = rx_services[i];
905 if (service == (struct rx_service *)0)
907 MUTEX_ENTER(&rx_quota_mutex);
908 rxi_totalMin += service->minProcs;
909 /* below works even if a thread is running, since minDeficit would
910 * still have been decremented and later re-incremented.
912 rxi_minDeficit += service->minProcs;
913 MUTEX_EXIT(&rx_quota_mutex);
916 /* Turn on reaping of idle server connections */
917 rxi_ReapConnections(NULL, NULL, NULL);
926 #ifdef AFS_PTHREAD_ENV
928 pid = afs_pointer_to_int(pthread_self());
929 #else /* AFS_PTHREAD_ENV */
931 LWP_CurrentProcess(&pid);
932 #endif /* AFS_PTHREAD_ENV */
934 sprintf(name, "srv_%d", ++nProcs);
936 (*registerProgram) (pid, name);
938 #endif /* AFS_NT40_ENV */
939 rx_ServerProc(NULL); /* Never returns */
941 #ifdef RX_ENABLE_TSFPQ
942 /* no use leaving packets around in this thread's local queue if
943 * it isn't getting donated to the server thread pool.
945 rxi_FlushLocalPacketsTSFPQ();
946 #endif /* RX_ENABLE_TSFPQ */
950 /* Create a new client connection to the specified service, using the
951 * specified security object to implement the security model for this
953 struct rx_connection *
954 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
955 struct rx_securityClass *securityObject,
956 int serviceSecurityIndex)
960 struct rx_connection *conn;
965 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
966 "serviceSecurityIndex %d)\n",
967 ntohl(shost), ntohs(sport), sservice, securityObject,
968 serviceSecurityIndex));
970 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
971 * the case of kmem_alloc? */
972 conn = rxi_AllocConnection();
973 #ifdef RX_ENABLE_LOCKS
974 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
975 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
976 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
979 MUTEX_ENTER(&rx_connHashTable_lock);
980 cid = (rx_nextCid += RX_MAXCALLS);
981 conn->type = RX_CLIENT_CONNECTION;
983 conn->epoch = rx_epoch;
984 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
985 conn->serviceId = sservice;
986 conn->securityObject = securityObject;
987 conn->securityData = (void *) 0;
988 conn->securityIndex = serviceSecurityIndex;
989 rx_SetConnDeadTime(conn, rx_connDeadTime);
990 rx_SetConnSecondsUntilNatPing(conn, 0);
991 conn->ackRate = RX_FAST_ACK_RATE;
993 conn->specific = NULL;
994 conn->challengeEvent = NULL;
995 conn->delayedAbortEvent = NULL;
996 conn->abortCount = 0;
998 for (i = 0; i < RX_MAXCALLS; i++) {
999 conn->twind[i] = rx_initSendWindow;
1000 conn->rwind[i] = rx_initReceiveWindow;
1001 conn->lastBusy[i] = 0;
1004 RXS_NewConnection(securityObject, conn);
1006 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1008 conn->refCount++; /* no lock required since only this thread knows... */
1009 conn->next = rx_connHashTable[hashindex];
1010 rx_connHashTable[hashindex] = conn;
1011 if (rx_stats_active)
1012 rx_atomic_inc(&rx_stats.nClientConns);
1013 MUTEX_EXIT(&rx_connHashTable_lock);
1019 * Ensure a connection's timeout values are valid.
1021 * @param[in] conn The connection to check
1023 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1024 * unless idleDeadTime and/or hardDeadTime are not set
1028 rxi_CheckConnTimeouts(struct rx_connection *conn)
1030 /* a connection's timeouts must have the relationship
1031 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1032 * total loss of network to a peer may cause an idle timeout instead of a
1033 * dead timeout, simply because the idle timeout gets hit first. Also set
1034 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1035 /* this logic is slightly complicated by the fact that
1036 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1038 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1039 if (conn->idleDeadTime) {
1040 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1042 if (conn->hardDeadTime) {
1043 if (conn->idleDeadTime) {
1044 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1046 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1052 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1054 /* The idea is to set the dead time to a value that allows several
1055 * keepalives to be dropped without timing out the connection. */
1056 conn->secondsUntilDead = seconds;
1057 rxi_CheckConnTimeouts(conn);
1058 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1062 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1064 conn->hardDeadTime = seconds;
1065 rxi_CheckConnTimeouts(conn);
1069 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1071 conn->idleDeadTime = seconds;
1072 rxi_CheckConnTimeouts(conn);
1075 int rxi_lowPeerRefCount = 0;
1076 int rxi_lowConnRefCount = 0;
1079 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1080 * NOTE: must not be called with rx_connHashTable_lock held.
1083 rxi_CleanupConnection(struct rx_connection *conn)
1085 /* Notify the service exporter, if requested, that this connection
1086 * is being destroyed */
1087 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1088 (*conn->service->destroyConnProc) (conn);
1090 /* Notify the security module that this connection is being destroyed */
1091 RXS_DestroyConnection(conn->securityObject, conn);
1093 /* If this is the last connection using the rx_peer struct, set its
1094 * idle time to now. rxi_ReapConnections will reap it if it's still
1095 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1097 MUTEX_ENTER(&rx_peerHashTable_lock);
1098 if (conn->peer->refCount < 2) {
1099 conn->peer->idleWhen = clock_Sec();
1100 if (conn->peer->refCount < 1) {
1101 conn->peer->refCount = 1;
1102 if (rx_stats_active) {
1103 MUTEX_ENTER(&rx_stats_mutex);
1104 rxi_lowPeerRefCount++;
1105 MUTEX_EXIT(&rx_stats_mutex);
1109 conn->peer->refCount--;
1110 MUTEX_EXIT(&rx_peerHashTable_lock);
1112 if (rx_stats_active)
1114 if (conn->type == RX_SERVER_CONNECTION)
1115 rx_atomic_dec(&rx_stats.nServerConns);
1117 rx_atomic_dec(&rx_stats.nClientConns);
1120 if (conn->specific) {
1122 for (i = 0; i < conn->nSpecific; i++) {
1123 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1124 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1125 conn->specific[i] = NULL;
1127 free(conn->specific);
1129 conn->specific = NULL;
1130 conn->nSpecific = 0;
1131 #endif /* !KERNEL */
1133 MUTEX_DESTROY(&conn->conn_call_lock);
1134 MUTEX_DESTROY(&conn->conn_data_lock);
1135 CV_DESTROY(&conn->conn_call_cv);
1137 rxi_FreeConnection(conn);
1140 /* Destroy the specified connection */
1142 rxi_DestroyConnection(struct rx_connection *conn)
1144 MUTEX_ENTER(&rx_connHashTable_lock);
1145 rxi_DestroyConnectionNoLock(conn);
1146 /* conn should be at the head of the cleanup list */
1147 if (conn == rx_connCleanup_list) {
1148 rx_connCleanup_list = rx_connCleanup_list->next;
1149 MUTEX_EXIT(&rx_connHashTable_lock);
1150 rxi_CleanupConnection(conn);
1152 #ifdef RX_ENABLE_LOCKS
1154 MUTEX_EXIT(&rx_connHashTable_lock);
1156 #endif /* RX_ENABLE_LOCKS */
1160 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1162 struct rx_connection **conn_ptr;
1164 struct rx_packet *packet;
1171 MUTEX_ENTER(&conn->conn_data_lock);
1172 MUTEX_ENTER(&rx_refcnt_mutex);
1173 if (conn->refCount > 0)
1176 if (rx_stats_active) {
1177 MUTEX_ENTER(&rx_stats_mutex);
1178 rxi_lowConnRefCount++;
1179 MUTEX_EXIT(&rx_stats_mutex);
1183 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1184 /* Busy; wait till the last guy before proceeding */
1185 MUTEX_EXIT(&rx_refcnt_mutex);
1186 MUTEX_EXIT(&conn->conn_data_lock);
1191 /* If the client previously called rx_NewCall, but it is still
1192 * waiting, treat this as a running call, and wait to destroy the
1193 * connection later when the call completes. */
1194 if ((conn->type == RX_CLIENT_CONNECTION)
1195 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1196 conn->flags |= RX_CONN_DESTROY_ME;
1197 MUTEX_EXIT(&conn->conn_data_lock);
1201 MUTEX_EXIT(&rx_refcnt_mutex);
1202 MUTEX_EXIT(&conn->conn_data_lock);
1204 /* Check for extant references to this connection */
1205 MUTEX_ENTER(&conn->conn_call_lock);
1206 for (i = 0; i < RX_MAXCALLS; i++) {
1207 struct rx_call *call = conn->call[i];
1210 if (conn->type == RX_CLIENT_CONNECTION) {
1211 MUTEX_ENTER(&call->lock);
1212 if (call->delayedAckEvent) {
1213 /* Push the final acknowledgment out now--there
1214 * won't be a subsequent call to acknowledge the
1215 * last reply packets */
1216 rxevent_Cancel(call->delayedAckEvent, call,
1217 RX_CALL_REFCOUNT_DELAY);
1218 if (call->state == RX_STATE_PRECALL
1219 || call->state == RX_STATE_ACTIVE) {
1220 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1222 rxi_AckAll(NULL, call, 0);
1225 MUTEX_EXIT(&call->lock);
1229 MUTEX_EXIT(&conn->conn_call_lock);
1231 #ifdef RX_ENABLE_LOCKS
1233 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1234 MUTEX_EXIT(&conn->conn_data_lock);
1236 /* Someone is accessing a packet right now. */
1240 #endif /* RX_ENABLE_LOCKS */
1243 /* Don't destroy the connection if there are any call
1244 * structures still in use */
1245 MUTEX_ENTER(&conn->conn_data_lock);
1246 conn->flags |= RX_CONN_DESTROY_ME;
1247 MUTEX_EXIT(&conn->conn_data_lock);
1252 if (conn->natKeepAliveEvent) {
1253 rxi_NatKeepAliveOff(conn);
1256 if (conn->delayedAbortEvent) {
1257 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1258 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1260 MUTEX_ENTER(&conn->conn_data_lock);
1261 rxi_SendConnectionAbort(conn, packet, 0, 1);
1262 MUTEX_EXIT(&conn->conn_data_lock);
1263 rxi_FreePacket(packet);
1267 /* Remove from connection hash table before proceeding */
1269 &rx_connHashTable[CONN_HASH
1270 (peer->host, peer->port, conn->cid, conn->epoch,
1272 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1273 if (*conn_ptr == conn) {
1274 *conn_ptr = conn->next;
1278 /* if the conn that we are destroying was the last connection, then we
1279 * clear rxLastConn as well */
1280 if (rxLastConn == conn)
1283 /* Make sure the connection is completely reset before deleting it. */
1284 /* get rid of pending events that could zap us later */
1285 if (conn->challengeEvent)
1286 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1287 if (conn->checkReachEvent)
1288 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1289 if (conn->natKeepAliveEvent)
1290 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
1292 /* Add the connection to the list of destroyed connections that
1293 * need to be cleaned up. This is necessary to avoid deadlocks
1294 * in the routines we call to inform others that this connection is
1295 * being destroyed. */
1296 conn->next = rx_connCleanup_list;
1297 rx_connCleanup_list = conn;
1300 /* Externally available version */
1302 rx_DestroyConnection(struct rx_connection *conn)
1307 rxi_DestroyConnection(conn);
1312 rx_GetConnection(struct rx_connection *conn)
1317 MUTEX_ENTER(&rx_refcnt_mutex);
1319 MUTEX_EXIT(&rx_refcnt_mutex);
1323 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1324 /* Wait for the transmit queue to no longer be busy.
1325 * requires the call->lock to be held */
1327 rxi_WaitforTQBusy(struct rx_call *call) {
1328 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1329 call->flags |= RX_CALL_TQ_WAIT;
1331 #ifdef RX_ENABLE_LOCKS
1332 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1333 CV_WAIT(&call->cv_tq, &call->lock);
1334 #else /* RX_ENABLE_LOCKS */
1335 osi_rxSleep(&call->tq);
1336 #endif /* RX_ENABLE_LOCKS */
1338 if (call->tqWaiters == 0) {
1339 call->flags &= ~RX_CALL_TQ_WAIT;
1346 rxi_WakeUpTransmitQueue(struct rx_call *call)
1348 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1349 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1350 call, call->tqWaiters, call->flags));
1351 #ifdef RX_ENABLE_LOCKS
1352 osirx_AssertMine(&call->lock, "rxi_Start start");
1353 CV_BROADCAST(&call->cv_tq);
1354 #else /* RX_ENABLE_LOCKS */
1355 osi_rxWakeup(&call->tq);
1356 #endif /* RX_ENABLE_LOCKS */
1360 /* Start a new rx remote procedure call, on the specified connection.
1361 * If wait is set to 1, wait for a free call channel; otherwise return
1362 * 0. Maxtime gives the maximum number of seconds this call may take,
1363 * after rx_NewCall returns. After this time interval, a call to any
1364 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1365 * For fine grain locking, we hold the conn_call_lock in order to
1366 * to ensure that we don't get signalle after we found a call in an active
1367 * state and before we go to sleep.
1370 rx_NewCall(struct rx_connection *conn)
1372 int i, wait, ignoreBusy = 1;
1373 struct rx_call *call;
1374 struct clock queueTime;
1375 afs_uint32 leastBusy = 0;
1379 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1382 clock_GetTime(&queueTime);
1384 * Check if there are others waiting for a new call.
1385 * If so, let them go first to avoid starving them.
1386 * This is a fairly simple scheme, and might not be
1387 * a complete solution for large numbers of waiters.
1389 * makeCallWaiters keeps track of the number of
1390 * threads waiting to make calls and the
1391 * RX_CONN_MAKECALL_WAITING flag bit is used to
1392 * indicate that there are indeed calls waiting.
1393 * The flag is set when the waiter is incremented.
1394 * It is only cleared when makeCallWaiters is 0.
1395 * This prevents us from accidently destroying the
1396 * connection while it is potentially about to be used.
1398 MUTEX_ENTER(&conn->conn_call_lock);
1399 MUTEX_ENTER(&conn->conn_data_lock);
1400 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1401 conn->flags |= RX_CONN_MAKECALL_WAITING;
1402 conn->makeCallWaiters++;
1403 MUTEX_EXIT(&conn->conn_data_lock);
1405 #ifdef RX_ENABLE_LOCKS
1406 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1410 MUTEX_ENTER(&conn->conn_data_lock);
1411 conn->makeCallWaiters--;
1412 if (conn->makeCallWaiters == 0)
1413 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1416 /* We are now the active thread in rx_NewCall */
1417 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1418 MUTEX_EXIT(&conn->conn_data_lock);
1423 for (i = 0; i < RX_MAXCALLS; i++) {
1424 call = conn->call[i];
1426 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1427 /* we're not ignoring busy call slots; only look at the
1428 * call slot that is the "least" busy */
1432 if (call->state == RX_STATE_DALLY) {
1433 MUTEX_ENTER(&call->lock);
1434 if (call->state == RX_STATE_DALLY) {
1435 if (ignoreBusy && conn->lastBusy[i]) {
1436 /* if we're ignoring busy call slots, skip any ones that
1437 * have lastBusy set */
1438 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1439 leastBusy = conn->lastBusy[i];
1441 MUTEX_EXIT(&call->lock);
1446 * We are setting the state to RX_STATE_RESET to
1447 * ensure that no one else will attempt to use this
1448 * call once we drop the conn->conn_call_lock and
1449 * call->lock. We must drop the conn->conn_call_lock
1450 * before calling rxi_ResetCall because the process
1451 * of clearing the transmit queue can block for an
1452 * extended period of time. If we block while holding
1453 * the conn->conn_call_lock, then all rx_EndCall
1454 * processing will block as well. This has a detrimental
1455 * effect on overall system performance.
1457 call->state = RX_STATE_RESET;
1458 MUTEX_EXIT(&conn->conn_call_lock);
1459 MUTEX_ENTER(&rx_refcnt_mutex);
1460 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1461 MUTEX_EXIT(&rx_refcnt_mutex);
1462 rxi_ResetCall(call, 0);
1463 (*call->callNumber)++;
1464 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1468 * If we failed to be able to safely obtain the
1469 * conn->conn_call_lock we will have to drop the
1470 * call->lock to avoid a deadlock. When the call->lock
1471 * is released the state of the call can change. If it
1472 * is no longer RX_STATE_RESET then some other thread is
1475 MUTEX_EXIT(&call->lock);
1476 MUTEX_ENTER(&conn->conn_call_lock);
1477 MUTEX_ENTER(&call->lock);
1479 if (call->state == RX_STATE_RESET)
1483 * If we get here it means that after dropping
1484 * the conn->conn_call_lock and call->lock that
1485 * the call is no longer ours. If we can't find
1486 * a free call in the remaining slots we should
1487 * not go immediately to RX_CONN_MAKECALL_WAITING
1488 * because by dropping the conn->conn_call_lock
1489 * we have given up synchronization with rx_EndCall.
1490 * Instead, cycle through one more time to see if
1491 * we can find a call that can call our own.
1493 MUTEX_ENTER(&rx_refcnt_mutex);
1494 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1495 MUTEX_EXIT(&rx_refcnt_mutex);
1498 MUTEX_EXIT(&call->lock);
1501 if (ignoreBusy && conn->lastBusy[i]) {
1502 /* if we're ignoring busy call slots, skip any ones that
1503 * have lastBusy set */
1504 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1505 leastBusy = conn->lastBusy[i];
1510 /* rxi_NewCall returns with mutex locked */
1511 call = rxi_NewCall(conn, i);
1512 MUTEX_ENTER(&rx_refcnt_mutex);
1513 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1514 MUTEX_EXIT(&rx_refcnt_mutex);
1518 if (i < RX_MAXCALLS) {
1519 conn->lastBusy[i] = 0;
1524 if (leastBusy && ignoreBusy) {
1525 /* we didn't find a useable call slot, but we did see at least one
1526 * 'busy' slot; look again and only use a slot with the 'least
1532 MUTEX_ENTER(&conn->conn_data_lock);
1533 conn->flags |= RX_CONN_MAKECALL_WAITING;
1534 conn->makeCallWaiters++;
1535 MUTEX_EXIT(&conn->conn_data_lock);
1537 #ifdef RX_ENABLE_LOCKS
1538 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1542 MUTEX_ENTER(&conn->conn_data_lock);
1543 conn->makeCallWaiters--;
1544 if (conn->makeCallWaiters == 0)
1545 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1546 MUTEX_EXIT(&conn->conn_data_lock);
1548 /* Client is initially in send mode */
1549 call->state = RX_STATE_ACTIVE;
1550 call->error = conn->error;
1552 call->mode = RX_MODE_ERROR;
1554 call->mode = RX_MODE_SENDING;
1556 /* remember start time for call in case we have hard dead time limit */
1557 call->queueTime = queueTime;
1558 clock_GetTime(&call->startTime);
1559 hzero(call->bytesSent);
1560 hzero(call->bytesRcvd);
1562 /* Turn on busy protocol. */
1563 rxi_KeepAliveOn(call);
1565 /* Attempt MTU discovery */
1566 rxi_GrowMTUOn(call);
1569 * We are no longer the active thread in rx_NewCall
1571 MUTEX_ENTER(&conn->conn_data_lock);
1572 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1573 MUTEX_EXIT(&conn->conn_data_lock);
1576 * Wake up anyone else who might be giving us a chance to
1577 * run (see code above that avoids resource starvation).
1579 #ifdef RX_ENABLE_LOCKS
1580 CV_BROADCAST(&conn->conn_call_cv);
1584 MUTEX_EXIT(&conn->conn_call_lock);
1586 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1587 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1588 osi_Panic("rx_NewCall call about to be used without an empty tq");
1590 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1592 MUTEX_EXIT(&call->lock);
1595 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1600 rxi_HasActiveCalls(struct rx_connection *aconn)
1603 struct rx_call *tcall;
1607 for (i = 0; i < RX_MAXCALLS; i++) {
1608 if ((tcall = aconn->call[i])) {
1609 if ((tcall->state == RX_STATE_ACTIVE)
1610 || (tcall->state == RX_STATE_PRECALL)) {
1621 rxi_GetCallNumberVector(struct rx_connection *aconn,
1622 afs_int32 * aint32s)
1625 struct rx_call *tcall;
1629 for (i = 0; i < RX_MAXCALLS; i++) {
1630 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1631 aint32s[i] = aconn->callNumber[i] + 1;
1633 aint32s[i] = aconn->callNumber[i];
1640 rxi_SetCallNumberVector(struct rx_connection *aconn,
1641 afs_int32 * aint32s)
1644 struct rx_call *tcall;
1648 for (i = 0; i < RX_MAXCALLS; i++) {
1649 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1650 aconn->callNumber[i] = aint32s[i] - 1;
1652 aconn->callNumber[i] = aint32s[i];
1658 /* Advertise a new service. A service is named locally by a UDP port
1659 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1662 char *serviceName; Name for identification purposes (e.g. the
1663 service name might be used for probing for
1666 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1667 char *serviceName, struct rx_securityClass **securityObjects,
1668 int nSecurityObjects,
1669 afs_int32(*serviceProc) (struct rx_call * acall))
1671 osi_socket socket = OSI_NULLSOCKET;
1672 struct rx_service *tservice;
1678 if (serviceId == 0) {
1680 "rx_NewService: service id for service %s is not non-zero.\n",
1687 "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",
1695 tservice = rxi_AllocService();
1698 #ifdef RX_ENABLE_LOCKS
1699 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1702 for (i = 0; i < RX_MAX_SERVICES; i++) {
1703 struct rx_service *service = rx_services[i];
1705 if (port == service->servicePort && host == service->serviceHost) {
1706 if (service->serviceId == serviceId) {
1707 /* The identical service has already been
1708 * installed; if the caller was intending to
1709 * change the security classes used by this
1710 * service, he/she loses. */
1712 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1713 serviceName, serviceId, service->serviceName);
1715 rxi_FreeService(tservice);
1718 /* Different service, same port: re-use the socket
1719 * which is bound to the same port */
1720 socket = service->socket;
1723 if (socket == OSI_NULLSOCKET) {
1724 /* If we don't already have a socket (from another
1725 * service on same port) get a new one */
1726 socket = rxi_GetHostUDPSocket(host, port);
1727 if (socket == OSI_NULLSOCKET) {
1729 rxi_FreeService(tservice);
1734 service->socket = socket;
1735 service->serviceHost = host;
1736 service->servicePort = port;
1737 service->serviceId = serviceId;
1738 service->serviceName = serviceName;
1739 service->nSecurityObjects = nSecurityObjects;
1740 service->securityObjects = securityObjects;
1741 service->minProcs = 0;
1742 service->maxProcs = 1;
1743 service->idleDeadTime = 60;
1744 service->idleDeadErr = 0;
1745 service->connDeadTime = rx_connDeadTime;
1746 service->executeRequestProc = serviceProc;
1747 service->checkReach = 0;
1748 service->nSpecific = 0;
1749 service->specific = NULL;
1750 rx_services[i] = service; /* not visible until now */
1756 rxi_FreeService(tservice);
1757 (osi_Msg "rx_NewService: cannot support > %d services\n",
1762 /* Set configuration options for all of a service's security objects */
1765 rx_SetSecurityConfiguration(struct rx_service *service,
1766 rx_securityConfigVariables type,
1770 for (i = 0; i<service->nSecurityObjects; i++) {
1771 if (service->securityObjects[i]) {
1772 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1780 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1781 struct rx_securityClass **securityObjects, int nSecurityObjects,
1782 afs_int32(*serviceProc) (struct rx_call * acall))
1784 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1787 /* Generic request processing loop. This routine should be called
1788 * by the implementation dependent rx_ServerProc. If socketp is
1789 * non-null, it will be set to the file descriptor that this thread
1790 * is now listening on. If socketp is null, this routine will never
1793 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1795 struct rx_call *call;
1797 struct rx_service *tservice = NULL;
1804 call = rx_GetCall(threadID, tservice, socketp);
1805 if (socketp && *socketp != OSI_NULLSOCKET) {
1806 /* We are now a listener thread */
1812 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1813 #ifdef RX_ENABLE_LOCKS
1815 #endif /* RX_ENABLE_LOCKS */
1816 afs_termState = AFSOP_STOP_AFS;
1817 afs_osi_Wakeup(&afs_termState);
1818 #ifdef RX_ENABLE_LOCKS
1820 #endif /* RX_ENABLE_LOCKS */
1825 /* if server is restarting( typically smooth shutdown) then do not
1826 * allow any new calls.
1829 if (rx_tranquil && (call != NULL)) {
1833 MUTEX_ENTER(&call->lock);
1835 rxi_CallError(call, RX_RESTARTING);
1836 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1838 MUTEX_EXIT(&call->lock);
1843 tservice = call->conn->service;
1845 if (tservice->beforeProc)
1846 (*tservice->beforeProc) (call);
1848 code = tservice->executeRequestProc(call);
1850 if (tservice->afterProc)
1851 (*tservice->afterProc) (call, code);
1853 rx_EndCall(call, code);
1854 if (rx_stats_active) {
1855 MUTEX_ENTER(&rx_stats_mutex);
1857 MUTEX_EXIT(&rx_stats_mutex);
1864 rx_WakeupServerProcs(void)
1866 struct rx_serverQueueEntry *np, *tqp;
1870 MUTEX_ENTER(&rx_serverPool_lock);
1872 #ifdef RX_ENABLE_LOCKS
1873 if (rx_waitForPacket)
1874 CV_BROADCAST(&rx_waitForPacket->cv);
1875 #else /* RX_ENABLE_LOCKS */
1876 if (rx_waitForPacket)
1877 osi_rxWakeup(rx_waitForPacket);
1878 #endif /* RX_ENABLE_LOCKS */
1879 MUTEX_ENTER(&freeSQEList_lock);
1880 for (np = rx_FreeSQEList; np; np = tqp) {
1881 tqp = *(struct rx_serverQueueEntry **)np;
1882 #ifdef RX_ENABLE_LOCKS
1883 CV_BROADCAST(&np->cv);
1884 #else /* RX_ENABLE_LOCKS */
1886 #endif /* RX_ENABLE_LOCKS */
1888 MUTEX_EXIT(&freeSQEList_lock);
1889 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1890 #ifdef RX_ENABLE_LOCKS
1891 CV_BROADCAST(&np->cv);
1892 #else /* RX_ENABLE_LOCKS */
1894 #endif /* RX_ENABLE_LOCKS */
1896 MUTEX_EXIT(&rx_serverPool_lock);
1901 * One thing that seems to happen is that all the server threads get
1902 * tied up on some empty or slow call, and then a whole bunch of calls
1903 * arrive at once, using up the packet pool, so now there are more
1904 * empty calls. The most critical resources here are server threads
1905 * and the free packet pool. The "doreclaim" code seems to help in
1906 * general. I think that eventually we arrive in this state: there
1907 * are lots of pending calls which do have all their packets present,
1908 * so they won't be reclaimed, are multi-packet calls, so they won't
1909 * be scheduled until later, and thus are tying up most of the free
1910 * packet pool for a very long time.
1912 * 1. schedule multi-packet calls if all the packets are present.
1913 * Probably CPU-bound operation, useful to return packets to pool.
1914 * Do what if there is a full window, but the last packet isn't here?
1915 * 3. preserve one thread which *only* runs "best" calls, otherwise
1916 * it sleeps and waits for that type of call.
1917 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1918 * the current dataquota business is badly broken. The quota isn't adjusted
1919 * to reflect how many packets are presently queued for a running call.
1920 * So, when we schedule a queued call with a full window of packets queued
1921 * up for it, that *should* free up a window full of packets for other 2d-class
1922 * calls to be able to use from the packet pool. But it doesn't.
1924 * NB. Most of the time, this code doesn't run -- since idle server threads
1925 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1926 * as a new call arrives.
1928 /* Sleep until a call arrives. Returns a pointer to the call, ready
1929 * for an rx_Read. */
1930 #ifdef RX_ENABLE_LOCKS
1932 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1934 struct rx_serverQueueEntry *sq;
1935 struct rx_call *call = (struct rx_call *)0;
1936 struct rx_service *service = NULL;
1938 MUTEX_ENTER(&freeSQEList_lock);
1940 if ((sq = rx_FreeSQEList)) {
1941 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1942 MUTEX_EXIT(&freeSQEList_lock);
1943 } else { /* otherwise allocate a new one and return that */
1944 MUTEX_EXIT(&freeSQEList_lock);
1945 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1946 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1947 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1950 MUTEX_ENTER(&rx_serverPool_lock);
1951 if (cur_service != NULL) {
1952 ReturnToServerPool(cur_service);
1955 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1956 struct rx_call *tcall, *ncall, *choice2 = NULL;
1958 /* Scan for eligible incoming calls. A call is not eligible
1959 * if the maximum number of calls for its service type are
1960 * already executing */
1961 /* One thread will process calls FCFS (to prevent starvation),
1962 * while the other threads may run ahead looking for calls which
1963 * have all their input data available immediately. This helps
1964 * keep threads from blocking, waiting for data from the client. */
1965 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1966 service = tcall->conn->service;
1967 if (!QuotaOK(service)) {
1970 MUTEX_ENTER(&rx_pthread_mutex);
1971 if (tno == rxi_fcfs_thread_num
1972 || !tcall->queue_item_header.next) {
1973 MUTEX_EXIT(&rx_pthread_mutex);
1974 /* If we're the fcfs thread , then we'll just use
1975 * this call. If we haven't been able to find an optimal
1976 * choice, and we're at the end of the list, then use a
1977 * 2d choice if one has been identified. Otherwise... */
1978 call = (choice2 ? choice2 : tcall);
1979 service = call->conn->service;
1981 MUTEX_EXIT(&rx_pthread_mutex);
1982 if (!queue_IsEmpty(&tcall->rq)) {
1983 struct rx_packet *rp;
1984 rp = queue_First(&tcall->rq, rx_packet);
1985 if (rp->header.seq == 1) {
1987 || (rp->header.flags & RX_LAST_PACKET)) {
1989 } else if (rxi_2dchoice && !choice2
1990 && !(tcall->flags & RX_CALL_CLEARED)
1991 && (tcall->rprev > rxi_HardAckRate)) {
2001 ReturnToServerPool(service);
2008 MUTEX_EXIT(&rx_serverPool_lock);
2009 MUTEX_ENTER(&call->lock);
2011 if (call->flags & RX_CALL_WAIT_PROC) {
2012 call->flags &= ~RX_CALL_WAIT_PROC;
2013 rx_atomic_dec(&rx_nWaiting);
2016 if (call->state != RX_STATE_PRECALL || call->error) {
2017 MUTEX_EXIT(&call->lock);
2018 MUTEX_ENTER(&rx_serverPool_lock);
2019 ReturnToServerPool(service);
2024 if (queue_IsEmpty(&call->rq)
2025 || queue_First(&call->rq, rx_packet)->header.seq != 1)
2026 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2028 CLEAR_CALL_QUEUE_LOCK(call);
2031 /* If there are no eligible incoming calls, add this process
2032 * to the idle server queue, to wait for one */
2036 *socketp = OSI_NULLSOCKET;
2038 sq->socketp = socketp;
2039 queue_Append(&rx_idleServerQueue, sq);
2040 #ifndef AFS_AIX41_ENV
2041 rx_waitForPacket = sq;
2043 rx_waitingForPacket = sq;
2044 #endif /* AFS_AIX41_ENV */
2046 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2048 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2049 MUTEX_EXIT(&rx_serverPool_lock);
2050 return (struct rx_call *)0;
2053 } while (!(call = sq->newcall)
2054 && !(socketp && *socketp != OSI_NULLSOCKET));
2055 MUTEX_EXIT(&rx_serverPool_lock);
2057 MUTEX_ENTER(&call->lock);
2063 MUTEX_ENTER(&freeSQEList_lock);
2064 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2065 rx_FreeSQEList = sq;
2066 MUTEX_EXIT(&freeSQEList_lock);
2069 clock_GetTime(&call->startTime);
2070 call->state = RX_STATE_ACTIVE;
2071 call->mode = RX_MODE_RECEIVING;
2072 #ifdef RX_KERNEL_TRACE
2073 if (ICL_SETACTIVE(afs_iclSetp)) {
2074 int glockOwner = ISAFS_GLOCK();
2077 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2078 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2085 rxi_calltrace(RX_CALL_START, call);
2086 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2087 call->conn->service->servicePort, call->conn->service->serviceId,
2090 MUTEX_EXIT(&call->lock);
2091 MUTEX_ENTER(&rx_refcnt_mutex);
2092 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2093 MUTEX_EXIT(&rx_refcnt_mutex);
2095 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2100 #else /* RX_ENABLE_LOCKS */
2102 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2104 struct rx_serverQueueEntry *sq;
2105 struct rx_call *call = (struct rx_call *)0, *choice2;
2106 struct rx_service *service = NULL;
2110 MUTEX_ENTER(&freeSQEList_lock);
2112 if ((sq = rx_FreeSQEList)) {
2113 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2114 MUTEX_EXIT(&freeSQEList_lock);
2115 } else { /* otherwise allocate a new one and return that */
2116 MUTEX_EXIT(&freeSQEList_lock);
2117 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2118 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2119 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2121 MUTEX_ENTER(&sq->lock);
2123 if (cur_service != NULL) {
2124 cur_service->nRequestsRunning--;
2125 MUTEX_ENTER(&rx_quota_mutex);
2126 if (cur_service->nRequestsRunning < cur_service->minProcs)
2129 MUTEX_EXIT(&rx_quota_mutex);
2131 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
2132 struct rx_call *tcall, *ncall;
2133 /* Scan for eligible incoming calls. A call is not eligible
2134 * if the maximum number of calls for its service type are
2135 * already executing */
2136 /* One thread will process calls FCFS (to prevent starvation),
2137 * while the other threads may run ahead looking for calls which
2138 * have all their input data available immediately. This helps
2139 * keep threads from blocking, waiting for data from the client. */
2140 choice2 = (struct rx_call *)0;
2141 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
2142 service = tcall->conn->service;
2143 if (QuotaOK(service)) {
2144 MUTEX_ENTER(&rx_pthread_mutex);
2145 if (tno == rxi_fcfs_thread_num
2146 || !tcall->queue_item_header.next) {
2147 MUTEX_EXIT(&rx_pthread_mutex);
2148 /* If we're the fcfs thread, then we'll just use
2149 * this call. If we haven't been able to find an optimal
2150 * choice, and we're at the end of the list, then use a
2151 * 2d choice if one has been identified. Otherwise... */
2152 call = (choice2 ? choice2 : tcall);
2153 service = call->conn->service;
2155 MUTEX_EXIT(&rx_pthread_mutex);
2156 if (!queue_IsEmpty(&tcall->rq)) {
2157 struct rx_packet *rp;
2158 rp = queue_First(&tcall->rq, rx_packet);
2159 if (rp->header.seq == 1
2161 || (rp->header.flags & RX_LAST_PACKET))) {
2163 } else if (rxi_2dchoice && !choice2
2164 && !(tcall->flags & RX_CALL_CLEARED)
2165 && (tcall->rprev > rxi_HardAckRate)) {
2179 /* we can't schedule a call if there's no data!!! */
2180 /* send an ack if there's no data, if we're missing the
2181 * first packet, or we're missing something between first
2182 * and last -- there's a "hole" in the incoming data. */
2183 if (queue_IsEmpty(&call->rq)
2184 || queue_First(&call->rq, rx_packet)->header.seq != 1
2185 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
2186 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2188 call->flags &= (~RX_CALL_WAIT_PROC);
2189 service->nRequestsRunning++;
2190 /* just started call in minProcs pool, need fewer to maintain
2192 MUTEX_ENTER(&rx_quota_mutex);
2193 if (service->nRequestsRunning <= service->minProcs)
2196 MUTEX_EXIT(&rx_quota_mutex);
2197 rx_atomic_dec(&rx_nWaiting);
2198 /* MUTEX_EXIT(&call->lock); */
2200 /* If there are no eligible incoming calls, add this process
2201 * to the idle server queue, to wait for one */
2204 *socketp = OSI_NULLSOCKET;
2206 sq->socketp = socketp;
2207 queue_Append(&rx_idleServerQueue, sq);
2211 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2213 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2214 return (struct rx_call *)0;
2217 } while (!(call = sq->newcall)
2218 && !(socketp && *socketp != OSI_NULLSOCKET));
2220 MUTEX_EXIT(&sq->lock);
2222 MUTEX_ENTER(&freeSQEList_lock);
2223 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2224 rx_FreeSQEList = sq;
2225 MUTEX_EXIT(&freeSQEList_lock);
2228 clock_GetTime(&call->startTime);
2229 call->state = RX_STATE_ACTIVE;
2230 call->mode = RX_MODE_RECEIVING;
2231 #ifdef RX_KERNEL_TRACE
2232 if (ICL_SETACTIVE(afs_iclSetp)) {
2233 int glockOwner = ISAFS_GLOCK();
2236 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2237 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2244 rxi_calltrace(RX_CALL_START, call);
2245 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2246 call->conn->service->servicePort, call->conn->service->serviceId,
2249 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2256 #endif /* RX_ENABLE_LOCKS */
2260 /* Establish a procedure to be called when a packet arrives for a
2261 * call. This routine will be called at most once after each call,
2262 * and will also be called if there is an error condition on the or
2263 * the call is complete. Used by multi rx to build a selection
2264 * function which determines which of several calls is likely to be a
2265 * good one to read from.
2266 * NOTE: the way this is currently implemented it is probably only a
2267 * good idea to (1) use it immediately after a newcall (clients only)
2268 * and (2) only use it once. Other uses currently void your warranty
2271 rx_SetArrivalProc(struct rx_call *call,
2272 void (*proc) (struct rx_call * call,
2275 void * handle, int arg)
2277 call->arrivalProc = proc;
2278 call->arrivalProcHandle = handle;
2279 call->arrivalProcArg = arg;
2282 /* Call is finished (possibly prematurely). Return rc to the peer, if
2283 * appropriate, and return the final error code from the conversation
2287 rx_EndCall(struct rx_call *call, afs_int32 rc)
2289 struct rx_connection *conn = call->conn;
2293 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2294 call, rc, call->error, call->abortCode));
2297 MUTEX_ENTER(&call->lock);
2299 if (rc == 0 && call->error == 0) {
2300 call->abortCode = 0;
2301 call->abortCount = 0;
2304 call->arrivalProc = (void (*)())0;
2305 if (rc && call->error == 0) {
2306 rxi_CallError(call, rc);
2307 call->mode = RX_MODE_ERROR;
2308 /* Send an abort message to the peer if this error code has
2309 * only just been set. If it was set previously, assume the
2310 * peer has already been sent the error code or will request it
2312 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2314 if (conn->type == RX_SERVER_CONNECTION) {
2315 /* Make sure reply or at least dummy reply is sent */
2316 if (call->mode == RX_MODE_RECEIVING) {
2317 MUTEX_EXIT(&call->lock);
2318 rxi_WriteProc(call, 0, 0);
2319 MUTEX_ENTER(&call->lock);
2321 if (call->mode == RX_MODE_SENDING) {
2322 MUTEX_EXIT(&call->lock);
2323 rxi_FlushWrite(call);
2324 MUTEX_ENTER(&call->lock);
2326 rxi_calltrace(RX_CALL_END, call);
2327 /* Call goes to hold state until reply packets are acknowledged */
2328 if (call->tfirst + call->nSoftAcked < call->tnext) {
2329 call->state = RX_STATE_HOLD;
2331 call->state = RX_STATE_DALLY;
2332 rxi_ClearTransmitQueue(call, 0);
2333 rxi_rto_cancel(call);
2334 rxevent_Cancel(call->keepAliveEvent, call,
2335 RX_CALL_REFCOUNT_ALIVE);
2337 } else { /* Client connection */
2339 /* Make sure server receives input packets, in the case where
2340 * no reply arguments are expected */
2341 if ((call->mode == RX_MODE_SENDING)
2342 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2343 MUTEX_EXIT(&call->lock);
2344 (void)rxi_ReadProc(call, &dummy, 1);
2345 MUTEX_ENTER(&call->lock);
2348 /* If we had an outstanding delayed ack, be nice to the server
2349 * and force-send it now.
2351 if (call->delayedAckEvent) {
2352 rxevent_Cancel(call->delayedAckEvent, call,
2353 RX_CALL_REFCOUNT_DELAY);
2354 call->delayedAckEvent = NULL;
2355 rxi_SendDelayedAck(NULL, call, NULL);
2358 /* We need to release the call lock since it's lower than the
2359 * conn_call_lock and we don't want to hold the conn_call_lock
2360 * over the rx_ReadProc call. The conn_call_lock needs to be held
2361 * here for the case where rx_NewCall is perusing the calls on
2362 * the connection structure. We don't want to signal until
2363 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2364 * have checked this call, found it active and by the time it
2365 * goes to sleep, will have missed the signal.
2367 MUTEX_EXIT(&call->lock);
2368 MUTEX_ENTER(&conn->conn_call_lock);
2369 MUTEX_ENTER(&call->lock);
2371 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2372 conn->lastBusy[call->channel] = 0;
2375 MUTEX_ENTER(&conn->conn_data_lock);
2376 conn->flags |= RX_CONN_BUSY;
2377 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2378 MUTEX_EXIT(&conn->conn_data_lock);
2379 #ifdef RX_ENABLE_LOCKS
2380 CV_BROADCAST(&conn->conn_call_cv);
2385 #ifdef RX_ENABLE_LOCKS
2387 MUTEX_EXIT(&conn->conn_data_lock);
2389 #endif /* RX_ENABLE_LOCKS */
2390 call->state = RX_STATE_DALLY;
2392 error = call->error;
2394 /* currentPacket, nLeft, and NFree must be zeroed here, because
2395 * ResetCall cannot: ResetCall may be called at splnet(), in the
2396 * kernel version, and may interrupt the macros rx_Read or
2397 * rx_Write, which run at normal priority for efficiency. */
2398 if (call->currentPacket) {
2399 #ifdef RX_TRACK_PACKETS
2400 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2402 rxi_FreePacket(call->currentPacket);
2403 call->currentPacket = (struct rx_packet *)0;
2406 call->nLeft = call->nFree = call->curlen = 0;
2408 /* Free any packets from the last call to ReadvProc/WritevProc */
2409 #ifdef RXDEBUG_PACKET
2411 #endif /* RXDEBUG_PACKET */
2412 rxi_FreePackets(0, &call->iovq);
2413 MUTEX_EXIT(&call->lock);
2415 MUTEX_ENTER(&rx_refcnt_mutex);
2416 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2417 MUTEX_EXIT(&rx_refcnt_mutex);
2418 if (conn->type == RX_CLIENT_CONNECTION) {
2419 MUTEX_ENTER(&conn->conn_data_lock);
2420 conn->flags &= ~RX_CONN_BUSY;
2421 MUTEX_EXIT(&conn->conn_data_lock);
2422 MUTEX_EXIT(&conn->conn_call_lock);
2426 * Map errors to the local host's errno.h format.
2428 error = ntoh_syserr_conv(error);
2432 #if !defined(KERNEL)
2434 /* Call this routine when shutting down a server or client (especially
2435 * clients). This will allow Rx to gracefully garbage collect server
2436 * connections, and reduce the number of retries that a server might
2437 * make to a dead client.
2438 * This is not quite right, since some calls may still be ongoing and
2439 * we can't lock them to destroy them. */
2443 struct rx_connection **conn_ptr, **conn_end;
2447 if (rxinit_status == 1) {
2449 return; /* Already shutdown. */
2451 rxi_DeleteCachedConnections();
2452 if (rx_connHashTable) {
2453 MUTEX_ENTER(&rx_connHashTable_lock);
2454 for (conn_ptr = &rx_connHashTable[0], conn_end =
2455 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2457 struct rx_connection *conn, *next;
2458 for (conn = *conn_ptr; conn; conn = next) {
2460 if (conn->type == RX_CLIENT_CONNECTION) {
2461 MUTEX_ENTER(&rx_refcnt_mutex);
2463 MUTEX_EXIT(&rx_refcnt_mutex);
2464 #ifdef RX_ENABLE_LOCKS
2465 rxi_DestroyConnectionNoLock(conn);
2466 #else /* RX_ENABLE_LOCKS */
2467 rxi_DestroyConnection(conn);
2468 #endif /* RX_ENABLE_LOCKS */
2472 #ifdef RX_ENABLE_LOCKS
2473 while (rx_connCleanup_list) {
2474 struct rx_connection *conn;
2475 conn = rx_connCleanup_list;
2476 rx_connCleanup_list = rx_connCleanup_list->next;
2477 MUTEX_EXIT(&rx_connHashTable_lock);
2478 rxi_CleanupConnection(conn);
2479 MUTEX_ENTER(&rx_connHashTable_lock);
2481 MUTEX_EXIT(&rx_connHashTable_lock);
2482 #endif /* RX_ENABLE_LOCKS */
2487 afs_winsockCleanup();
2495 /* if we wakeup packet waiter too often, can get in loop with two
2496 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2498 rxi_PacketsUnWait(void)
2500 if (!rx_waitingForPackets) {
2504 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2505 return; /* still over quota */
2508 rx_waitingForPackets = 0;
2509 #ifdef RX_ENABLE_LOCKS
2510 CV_BROADCAST(&rx_waitingForPackets_cv);
2512 osi_rxWakeup(&rx_waitingForPackets);
2518 /* ------------------Internal interfaces------------------------- */
2520 /* Return this process's service structure for the
2521 * specified socket and service */
2522 static struct rx_service *
2523 rxi_FindService(osi_socket socket, u_short serviceId)
2525 struct rx_service **sp;
2526 for (sp = &rx_services[0]; *sp; sp++) {
2527 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2533 #ifdef RXDEBUG_PACKET
2534 #ifdef KDUMP_RX_LOCK
2535 static struct rx_call_rx_lock *rx_allCallsp = 0;
2537 static struct rx_call *rx_allCallsp = 0;
2539 #endif /* RXDEBUG_PACKET */
2541 /* Allocate a call structure, for the indicated channel of the
2542 * supplied connection. The mode and state of the call must be set by
2543 * the caller. Returns the call with mutex locked. */
2544 static struct rx_call *
2545 rxi_NewCall(struct rx_connection *conn, int channel)
2547 struct rx_call *call;
2548 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2549 struct rx_call *cp; /* Call pointer temp */
2550 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2551 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2553 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2555 /* Grab an existing call structure, or allocate a new one.
2556 * Existing call structures are assumed to have been left reset by
2558 MUTEX_ENTER(&rx_freeCallQueue_lock);
2560 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2562 * EXCEPT that the TQ might not yet be cleared out.
2563 * Skip over those with in-use TQs.
2566 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2567 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2573 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2574 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2575 call = queue_First(&rx_freeCallQueue, rx_call);
2576 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2578 if (rx_stats_active)
2579 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2580 MUTEX_EXIT(&rx_freeCallQueue_lock);
2581 MUTEX_ENTER(&call->lock);
2582 CLEAR_CALL_QUEUE_LOCK(call);
2583 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2584 /* Now, if TQ wasn't cleared earlier, do it now. */
2585 rxi_WaitforTQBusy(call);
2586 if (call->flags & RX_CALL_TQ_CLEARME) {
2587 rxi_ClearTransmitQueue(call, 1);
2588 /*queue_Init(&call->tq);*/
2590 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2591 /* Bind the call to its connection structure */
2593 rxi_ResetCall(call, 1);
2596 call = rxi_Alloc(sizeof(struct rx_call));
2597 #ifdef RXDEBUG_PACKET
2598 call->allNextp = rx_allCallsp;
2599 rx_allCallsp = call;
2601 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2602 #else /* RXDEBUG_PACKET */
2603 rx_atomic_inc(&rx_stats.nCallStructs);
2604 #endif /* RXDEBUG_PACKET */
2606 MUTEX_EXIT(&rx_freeCallQueue_lock);
2607 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2608 MUTEX_ENTER(&call->lock);
2609 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2610 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2611 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2613 /* Initialize once-only items */
2614 queue_Init(&call->tq);
2615 queue_Init(&call->rq);
2616 queue_Init(&call->iovq);
2617 #ifdef RXDEBUG_PACKET
2618 call->rqc = call->tqc = call->iovqc = 0;
2619 #endif /* RXDEBUG_PACKET */
2620 /* Bind the call to its connection structure (prereq for reset) */
2622 rxi_ResetCall(call, 1);
2624 call->channel = channel;
2625 call->callNumber = &conn->callNumber[channel];
2626 call->rwind = conn->rwind[channel];
2627 call->twind = conn->twind[channel];
2628 /* Note that the next expected call number is retained (in
2629 * conn->callNumber[i]), even if we reallocate the call structure
2631 conn->call[channel] = call;
2632 /* if the channel's never been used (== 0), we should start at 1, otherwise
2633 * the call number is valid from the last time this channel was used */
2634 if (*call->callNumber == 0)
2635 *call->callNumber = 1;
2640 /* A call has been inactive long enough that so we can throw away
2641 * state, including the call structure, which is placed on the call
2644 * call->lock amd rx_refcnt_mutex are held upon entry.
2645 * haveCTLock is set when called from rxi_ReapConnections.
2648 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2650 int channel = call->channel;
2651 struct rx_connection *conn = call->conn;
2654 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2655 (*call->callNumber)++;
2657 * We are setting the state to RX_STATE_RESET to
2658 * ensure that no one else will attempt to use this
2659 * call once we drop the refcnt lock. We must drop
2660 * the refcnt lock before calling rxi_ResetCall
2661 * because it cannot be held across acquiring the
2662 * freepktQ lock. NewCall does the same.
2664 call->state = RX_STATE_RESET;
2665 MUTEX_EXIT(&rx_refcnt_mutex);
2666 rxi_ResetCall(call, 0);
2668 MUTEX_ENTER(&conn->conn_call_lock);
2669 if (call->conn->call[channel] == call)
2670 call->conn->call[channel] = 0;
2671 MUTEX_EXIT(&conn->conn_call_lock);
2673 MUTEX_ENTER(&rx_freeCallQueue_lock);
2674 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2675 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2676 /* A call may be free even though its transmit queue is still in use.
2677 * Since we search the call list from head to tail, put busy calls at
2678 * the head of the list, and idle calls at the tail.
2680 if (call->flags & RX_CALL_TQ_BUSY)
2681 queue_Prepend(&rx_freeCallQueue, call);
2683 queue_Append(&rx_freeCallQueue, call);
2684 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2685 queue_Append(&rx_freeCallQueue, call);
2686 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2687 if (rx_stats_active)
2688 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2689 MUTEX_EXIT(&rx_freeCallQueue_lock);
2691 /* Destroy the connection if it was previously slated for
2692 * destruction, i.e. the Rx client code previously called
2693 * rx_DestroyConnection (client connections), or
2694 * rxi_ReapConnections called the same routine (server
2695 * connections). Only do this, however, if there are no
2696 * outstanding calls. Note that for fine grain locking, there appears
2697 * to be a deadlock in that rxi_FreeCall has a call locked and
2698 * DestroyConnectionNoLock locks each call in the conn. But note a
2699 * few lines up where we have removed this call from the conn.
2700 * If someone else destroys a connection, they either have no
2701 * call lock held or are going through this section of code.
2703 MUTEX_ENTER(&conn->conn_data_lock);
2704 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2705 MUTEX_ENTER(&rx_refcnt_mutex);
2707 MUTEX_EXIT(&rx_refcnt_mutex);
2708 MUTEX_EXIT(&conn->conn_data_lock);
2709 #ifdef RX_ENABLE_LOCKS
2711 rxi_DestroyConnectionNoLock(conn);
2713 rxi_DestroyConnection(conn);
2714 #else /* RX_ENABLE_LOCKS */
2715 rxi_DestroyConnection(conn);
2716 #endif /* RX_ENABLE_LOCKS */
2718 MUTEX_EXIT(&conn->conn_data_lock);
2720 MUTEX_ENTER(&rx_refcnt_mutex);
2723 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2724 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2727 rxi_Alloc(size_t size)
2731 if (rx_stats_active) {
2732 rx_atomic_add(&rxi_Allocsize, (int) size);
2733 rx_atomic_inc(&rxi_Alloccnt);
2737 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2738 afs_osi_Alloc_NoSleep(size);
2743 osi_Panic("rxi_Alloc error");
2749 rxi_Free(void *addr, size_t size)
2751 if (rx_stats_active) {
2752 rx_atomic_sub(&rxi_Allocsize, (int) size);
2753 rx_atomic_dec(&rxi_Alloccnt);
2755 osi_Free(addr, size);
2759 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2761 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2762 struct rx_peer *next = NULL;
2766 MUTEX_ENTER(&rx_peerHashTable_lock);
2768 peer_ptr = &rx_peerHashTable[0];
2769 peer_end = &rx_peerHashTable[rx_hashTableSize];
2772 for ( ; peer_ptr < peer_end; peer_ptr++) {
2775 for ( ; peer; peer = next) {
2777 if (host == peer->host)
2782 hashIndex = PEER_HASH(host, port);
2783 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2784 if ((peer->host == host) && (peer->port == port))
2789 MUTEX_ENTER(&rx_peerHashTable_lock);
2794 MUTEX_EXIT(&rx_peerHashTable_lock);
2796 MUTEX_ENTER(&peer->peer_lock);
2797 /* We don't handle dropping below min, so don't */
2798 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2799 peer->ifMTU=MIN(mtu, peer->ifMTU);
2800 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2801 /* if we tweaked this down, need to tune our peer MTU too */
2802 peer->MTU = MIN(peer->MTU, peer->natMTU);
2803 /* if we discovered a sub-1500 mtu, degrade */
2804 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2805 peer->maxDgramPackets = 1;
2806 /* We no longer have valid peer packet information */
2807 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2808 peer->maxPacketSize = 0;
2809 MUTEX_EXIT(&peer->peer_lock);
2811 MUTEX_ENTER(&rx_peerHashTable_lock);
2813 if (host && !port) {
2815 /* pick up where we left off */
2819 MUTEX_EXIT(&rx_peerHashTable_lock);
2822 /* Find the peer process represented by the supplied (host,port)
2823 * combination. If there is no appropriate active peer structure, a
2824 * new one will be allocated and initialized
2825 * The origPeer, if set, is a pointer to a peer structure on which the
2826 * refcount will be be decremented. This is used to replace the peer
2827 * structure hanging off a connection structure */
2829 rxi_FindPeer(afs_uint32 host, u_short port,
2830 struct rx_peer *origPeer, int create)
2834 hashIndex = PEER_HASH(host, port);
2835 MUTEX_ENTER(&rx_peerHashTable_lock);
2836 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2837 if ((pp->host == host) && (pp->port == port))
2842 pp = rxi_AllocPeer(); /* This bzero's *pp */
2843 pp->host = host; /* set here or in InitPeerParams is zero */
2845 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2846 queue_Init(&pp->congestionQueue);
2847 queue_Init(&pp->rpcStats);
2848 pp->next = rx_peerHashTable[hashIndex];
2849 rx_peerHashTable[hashIndex] = pp;
2850 rxi_InitPeerParams(pp);
2851 if (rx_stats_active)
2852 rx_atomic_inc(&rx_stats.nPeerStructs);
2859 origPeer->refCount--;
2860 MUTEX_EXIT(&rx_peerHashTable_lock);
2865 /* Find the connection at (host, port) started at epoch, and with the
2866 * given connection id. Creates the server connection if necessary.
2867 * The type specifies whether a client connection or a server
2868 * connection is desired. In both cases, (host, port) specify the
2869 * peer's (host, pair) pair. Client connections are not made
2870 * automatically by this routine. The parameter socket gives the
2871 * socket descriptor on which the packet was received. This is used,
2872 * in the case of server connections, to check that *new* connections
2873 * come via a valid (port, serviceId). Finally, the securityIndex
2874 * parameter must match the existing index for the connection. If a
2875 * server connection is created, it will be created using the supplied
2876 * index, if the index is valid for this service */
2877 struct rx_connection *
2878 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2879 u_short port, u_short serviceId, afs_uint32 cid,
2880 afs_uint32 epoch, int type, u_int securityIndex)
2882 int hashindex, flag, i;
2883 struct rx_connection *conn;
2884 hashindex = CONN_HASH(host, port, cid, epoch, type);
2885 MUTEX_ENTER(&rx_connHashTable_lock);
2886 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2887 rx_connHashTable[hashindex],
2890 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2891 && (epoch == conn->epoch)) {
2892 struct rx_peer *pp = conn->peer;
2893 if (securityIndex != conn->securityIndex) {
2894 /* this isn't supposed to happen, but someone could forge a packet
2895 * like this, and there seems to be some CM bug that makes this
2896 * happen from time to time -- in which case, the fileserver
2898 MUTEX_EXIT(&rx_connHashTable_lock);
2899 return (struct rx_connection *)0;
2901 if (pp->host == host && pp->port == port)
2903 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2905 /* So what happens when it's a callback connection? */
2906 if ( /*type == RX_CLIENT_CONNECTION && */
2907 (conn->epoch & 0x80000000))
2911 /* the connection rxLastConn that was used the last time is not the
2912 ** one we are looking for now. Hence, start searching in the hash */
2914 conn = rx_connHashTable[hashindex];
2919 struct rx_service *service;
2920 if (type == RX_CLIENT_CONNECTION) {
2921 MUTEX_EXIT(&rx_connHashTable_lock);
2922 return (struct rx_connection *)0;
2924 service = rxi_FindService(socket, serviceId);
2925 if (!service || (securityIndex >= service->nSecurityObjects)
2926 || (service->securityObjects[securityIndex] == 0)) {
2927 MUTEX_EXIT(&rx_connHashTable_lock);
2928 return (struct rx_connection *)0;
2930 conn = rxi_AllocConnection(); /* This bzero's the connection */
2931 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2932 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2933 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2934 conn->next = rx_connHashTable[hashindex];
2935 rx_connHashTable[hashindex] = conn;
2936 conn->peer = rxi_FindPeer(host, port, 0, 1);
2937 conn->type = RX_SERVER_CONNECTION;
2938 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2939 conn->epoch = epoch;
2940 conn->cid = cid & RX_CIDMASK;
2941 /* conn->serial = conn->lastSerial = 0; */
2942 /* conn->timeout = 0; */
2943 conn->ackRate = RX_FAST_ACK_RATE;
2944 conn->service = service;
2945 conn->serviceId = serviceId;
2946 conn->securityIndex = securityIndex;
2947 conn->securityObject = service->securityObjects[securityIndex];
2948 conn->nSpecific = 0;
2949 conn->specific = NULL;
2950 rx_SetConnDeadTime(conn, service->connDeadTime);
2951 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2952 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2953 for (i = 0; i < RX_MAXCALLS; i++) {
2954 conn->twind[i] = rx_initSendWindow;
2955 conn->rwind[i] = rx_initReceiveWindow;
2957 /* Notify security object of the new connection */
2958 RXS_NewConnection(conn->securityObject, conn);
2959 /* XXXX Connection timeout? */
2960 if (service->newConnProc)
2961 (*service->newConnProc) (conn);
2962 if (rx_stats_active)
2963 rx_atomic_inc(&rx_stats.nServerConns);
2966 MUTEX_ENTER(&rx_refcnt_mutex);
2968 MUTEX_EXIT(&rx_refcnt_mutex);
2970 rxLastConn = conn; /* store this connection as the last conn used */
2971 MUTEX_EXIT(&rx_connHashTable_lock);
2976 * Timeout a call on a busy call channel if appropriate.
2978 * @param[in] call The busy call.
2980 * @pre 'call' is marked as busy (namely,
2981 * call->conn->lastBusy[call->channel] != 0)
2983 * @pre call->lock is held
2984 * @pre rxi_busyChannelError is nonzero
2986 * @note call->lock is dropped and reacquired
2989 rxi_CheckBusy(struct rx_call *call)
2991 struct rx_connection *conn = call->conn;
2992 int channel = call->channel;
2993 int freechannel = 0;
2995 afs_uint32 callNumber = *call->callNumber;
2997 MUTEX_EXIT(&call->lock);
2999 MUTEX_ENTER(&conn->conn_call_lock);
3001 /* Are there any other call slots on this conn that we should try? Look for
3002 * slots that are empty and are either non-busy, or were marked as busy
3003 * longer than conn->secondsUntilDead seconds before this call started. */
3005 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3007 /* only look at channels that aren't us */
3011 if (conn->lastBusy[i]) {
3012 /* if this channel looked busy too recently, don't look at it */
3013 if (conn->lastBusy[i] >= call->startTime.sec) {
3016 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3021 if (conn->call[i]) {
3022 struct rx_call *tcall = conn->call[i];
3023 MUTEX_ENTER(&tcall->lock);
3024 if (tcall->state == RX_STATE_DALLY) {
3027 MUTEX_EXIT(&tcall->lock);
3033 MUTEX_EXIT(&conn->conn_call_lock);
3035 MUTEX_ENTER(&call->lock);
3037 /* Since the call->lock and conn->conn_call_lock have been released it is
3038 * possible that (1) the call may no longer be busy and/or (2) the call may
3039 * have been reused by another waiting thread. Therefore, we must confirm
3040 * that the call state has not changed when deciding whether or not to
3041 * force this application thread to retry by forcing a Timeout error. */
3043 if (freechannel && *call->callNumber == callNumber &&
3044 (call->flags & RX_CALL_PEER_BUSY)) {
3045 /* Since 'freechannel' is set, there exists another channel in this
3046 * rx_conn that the application thread might be able to use. We know
3047 * that we have the correct call since callNumber is unchanged, and we
3048 * know that the call is still busy. So, set the call error state to
3049 * rxi_busyChannelError so the application can retry the request,
3050 * presumably on a less-busy call channel. */
3052 rxi_CallError(call, rxi_busyChannelError);
3056 /* There are two packet tracing routines available for testing and monitoring
3057 * Rx. One is called just after every packet is received and the other is
3058 * called just before every packet is sent. Received packets, have had their
3059 * headers decoded, and packets to be sent have not yet had their headers
3060 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3061 * containing the network address. Both can be modified. The return value, if
3062 * non-zero, indicates that the packet should be dropped. */
3064 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3065 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3067 /* A packet has been received off the interface. Np is the packet, socket is
3068 * the socket number it was received from (useful in determining which service
3069 * this packet corresponds to), and (host, port) reflect the host,port of the
3070 * sender. This call returns the packet to the caller if it is finished with
3071 * it, rather than de-allocating it, just as a small performance hack */
3074 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3075 afs_uint32 host, u_short port, int *tnop,
3076 struct rx_call **newcallp)
3078 struct rx_call *call;
3079 struct rx_connection *conn;
3081 afs_uint32 currentCallNumber;
3087 struct rx_packet *tnp;
3090 /* We don't print out the packet until now because (1) the time may not be
3091 * accurate enough until now in the lwp implementation (rx_Listener only gets
3092 * the time after the packet is read) and (2) from a protocol point of view,
3093 * this is the first time the packet has been seen */
3094 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3095 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3096 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3097 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3098 np->header.epoch, np->header.cid, np->header.callNumber,
3099 np->header.seq, np->header.flags, np));
3102 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3103 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3106 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3107 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3110 /* If an input tracer function is defined, call it with the packet and
3111 * network address. Note this function may modify its arguments. */
3112 if (rx_justReceived) {
3113 struct sockaddr_in addr;
3115 addr.sin_family = AF_INET;
3116 addr.sin_port = port;
3117 addr.sin_addr.s_addr = host;
3118 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3119 addr.sin_len = sizeof(addr);
3120 #endif /* AFS_OSF_ENV */
3121 drop = (*rx_justReceived) (np, &addr);
3122 /* drop packet if return value is non-zero */
3125 port = addr.sin_port; /* in case fcn changed addr */
3126 host = addr.sin_addr.s_addr;
3130 /* If packet was not sent by the client, then *we* must be the client */
3131 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3132 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3134 /* Find the connection (or fabricate one, if we're the server & if
3135 * necessary) associated with this packet */
3137 rxi_FindConnection(socket, host, port, np->header.serviceId,
3138 np->header.cid, np->header.epoch, type,
3139 np->header.securityIndex);
3142 /* If no connection found or fabricated, just ignore the packet.
3143 * (An argument could be made for sending an abort packet for
3148 /* If the connection is in an error state, send an abort packet and ignore
3149 * the incoming packet */
3151 /* Don't respond to an abort packet--we don't want loops! */
3152 MUTEX_ENTER(&conn->conn_data_lock);
3153 if (np->header.type != RX_PACKET_TYPE_ABORT)
3154 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3155 MUTEX_ENTER(&rx_refcnt_mutex);
3157 MUTEX_EXIT(&rx_refcnt_mutex);
3158 MUTEX_EXIT(&conn->conn_data_lock);
3162 /* Check for connection-only requests (i.e. not call specific). */
3163 if (np->header.callNumber == 0) {
3164 switch (np->header.type) {
3165 case RX_PACKET_TYPE_ABORT: {
3166 /* What if the supplied error is zero? */
3167 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3168 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3169 rxi_ConnectionError(conn, errcode);
3170 MUTEX_ENTER(&rx_refcnt_mutex);
3172 MUTEX_EXIT(&rx_refcnt_mutex);
3175 case RX_PACKET_TYPE_CHALLENGE:
3176 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3177 MUTEX_ENTER(&rx_refcnt_mutex);
3179 MUTEX_EXIT(&rx_refcnt_mutex);
3181 case RX_PACKET_TYPE_RESPONSE:
3182 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3183 MUTEX_ENTER(&rx_refcnt_mutex);
3185 MUTEX_EXIT(&rx_refcnt_mutex);
3187 case RX_PACKET_TYPE_PARAMS:
3188 case RX_PACKET_TYPE_PARAMS + 1:
3189 case RX_PACKET_TYPE_PARAMS + 2:
3190 /* ignore these packet types for now */
3191 MUTEX_ENTER(&rx_refcnt_mutex);
3193 MUTEX_EXIT(&rx_refcnt_mutex);
3198 /* Should not reach here, unless the peer is broken: send an
3200 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3201 MUTEX_ENTER(&conn->conn_data_lock);
3202 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3203 MUTEX_ENTER(&rx_refcnt_mutex);
3205 MUTEX_EXIT(&rx_refcnt_mutex);
3206 MUTEX_EXIT(&conn->conn_data_lock);
3211 channel = np->header.cid & RX_CHANNELMASK;
3212 call = conn->call[channel];
3215 MUTEX_ENTER(&call->lock);
3216 currentCallNumber = conn->callNumber[channel];
3217 } else if (type == RX_SERVER_CONNECTION) { /* No call allocated */
3218 MUTEX_ENTER(&conn->conn_call_lock);
3219 call = conn->call[channel];
3221 MUTEX_ENTER(&call->lock);
3222 MUTEX_EXIT(&conn->conn_call_lock);
3223 currentCallNumber = conn->callNumber[channel];
3225 call = rxi_NewCall(conn, channel); /* returns locked call */
3226 MUTEX_EXIT(&conn->conn_call_lock);
3227 *call->callNumber = currentCallNumber = np->header.callNumber;
3229 if (np->header.callNumber == 0)
3230 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3231 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3232 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3233 np->header.flags, np, np->length));
3235 call->state = RX_STATE_PRECALL;
3236 clock_GetTime(&call->queueTime);
3237 hzero(call->bytesSent);
3238 hzero(call->bytesRcvd);
3240 * If the number of queued calls exceeds the overload
3241 * threshold then abort this call.
3243 if ((rx_BusyThreshold > 0) &&
3244 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3245 struct rx_packet *tp;
3247 rxi_CallError(call, rx_BusyError);
3248 tp = rxi_SendCallAbort(call, np, 1, 0);
3249 MUTEX_EXIT(&call->lock);
3250 MUTEX_ENTER(&rx_refcnt_mutex);
3252 MUTEX_EXIT(&rx_refcnt_mutex);
3253 if (rx_stats_active)
3254 rx_atomic_inc(&rx_stats.nBusies);
3257 rxi_KeepAliveOn(call);
3259 } else { /* RX_CLIENT_CONNECTION and No call allocated */
3260 /* This packet can't be for this call. If the new call address is
3261 * 0 then no call is running on this channel. If there is a call
3262 * then, since this is a client connection we're getting data for
3263 * it must be for the previous call.
3265 if (rx_stats_active)
3266 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3267 MUTEX_ENTER(&rx_refcnt_mutex);
3269 MUTEX_EXIT(&rx_refcnt_mutex);
3273 /* There is a non-NULL locked call at this point */
3274 if (type == RX_SERVER_CONNECTION) { /* We're the server */
3275 if (np->header.callNumber < currentCallNumber) {
3276 MUTEX_EXIT(&call->lock);
3277 if (rx_stats_active)
3278 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3279 MUTEX_ENTER(&rx_refcnt_mutex);
3281 MUTEX_EXIT(&rx_refcnt_mutex);
3283 } else if (np->header.callNumber != currentCallNumber) {
3284 /* Wait until the transmit queue is idle before deciding
3285 * whether to reset the current call. Chances are that the
3286 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3289 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3290 if (call->state == RX_STATE_ACTIVE) {
3291 rxi_WaitforTQBusy(call);
3293 * If we entered error state while waiting,
3294 * must call rxi_CallError to permit rxi_ResetCall
3295 * to processed when the tqWaiter count hits zero.
3298 rxi_CallError(call, call->error);
3299 MUTEX_EXIT(&call->lock);
3300 MUTEX_ENTER(&rx_refcnt_mutex);
3302 MUTEX_EXIT(&rx_refcnt_mutex);
3306 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3307 /* If the new call cannot be taken right now send a busy and set
3308 * the error condition in this call, so that it terminates as
3309 * quickly as possible */
3310 if (call->state == RX_STATE_ACTIVE) {
3311 struct rx_packet *tp;
3313 rxi_CallError(call, RX_CALL_DEAD);
3314 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3316 MUTEX_EXIT(&call->lock);
3317 MUTEX_ENTER(&rx_refcnt_mutex);
3319 MUTEX_EXIT(&rx_refcnt_mutex);
3322 rxi_ResetCall(call, 0);
3323 *call->callNumber = np->header.callNumber;
3325 if (np->header.callNumber == 0)
3326 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3327 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3328 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3329 np->header.flags, np, np->length));
3331 call->state = RX_STATE_PRECALL;
3332 clock_GetTime(&call->queueTime);
3333 hzero(call->bytesSent);
3334 hzero(call->bytesRcvd);
3336 * If the number of queued calls exceeds the overload
3337 * threshold then abort this call.
3339 if ((rx_BusyThreshold > 0) &&
3340 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3341 struct rx_packet *tp;
3343 rxi_CallError(call, rx_BusyError);
3344 tp = rxi_SendCallAbort(call, np, 1, 0);
3345 MUTEX_EXIT(&call->lock);
3346 MUTEX_ENTER(&rx_refcnt_mutex);
3348 MUTEX_EXIT(&rx_refcnt_mutex);
3349 if (rx_stats_active)
3350 rx_atomic_inc(&rx_stats.nBusies);
3353 rxi_KeepAliveOn(call);
3355 /* Continuing call; do nothing here. */
3357 } else { /* we're the client */
3358 /* Ignore all incoming acknowledgements for calls in DALLY state */
3359 if ((call->state == RX_STATE_DALLY)
3360 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3361 if (rx_stats_active)
3362 rx_atomic_inc(&rx_stats.ignorePacketDally);
3363 MUTEX_EXIT(&call->lock);
3364 MUTEX_ENTER(&rx_refcnt_mutex);
3366 MUTEX_EXIT(&rx_refcnt_mutex);
3370 /* Ignore anything that's not relevant to the current call. If there
3371 * isn't a current call, then no packet is relevant. */
3372 if (np->header.callNumber != currentCallNumber) {
3373 if (rx_stats_active)
3374 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3375 MUTEX_EXIT(&call->lock);
3376 MUTEX_ENTER(&rx_refcnt_mutex);
3378 MUTEX_EXIT(&rx_refcnt_mutex);
3381 /* If the service security object index stamped in the packet does not
3382 * match the connection's security index, ignore the packet */
3383 if (np->header.securityIndex != conn->securityIndex) {
3384 MUTEX_EXIT(&call->lock);
3385 MUTEX_ENTER(&rx_refcnt_mutex);
3387 MUTEX_EXIT(&rx_refcnt_mutex);
3391 /* If we're receiving the response, then all transmit packets are
3392 * implicitly acknowledged. Get rid of them. */
3393 if (np->header.type == RX_PACKET_TYPE_DATA) {
3394 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3395 /* XXX Hack. Because we must release the global rx lock when
3396 * sending packets (osi_NetSend) we drop all acks while we're
3397 * traversing the tq in rxi_Start sending packets out because
3398 * packets may move to the freePacketQueue as result of being here!
3399 * So we drop these packets until we're safely out of the
3400 * traversing. Really ugly!
3401 * For fine grain RX locking, we set the acked field in the
3402 * packets and let rxi_Start remove them from the transmit queue.
3404 if (call->flags & RX_CALL_TQ_BUSY) {
3405 #ifdef RX_ENABLE_LOCKS
3406 rxi_SetAcksInTransmitQueue(call);
3408 MUTEX_ENTER(&rx_refcnt_mutex);
3410 MUTEX_EXIT(&rx_refcnt_mutex);
3411 return np; /* xmitting; drop packet */
3414 rxi_ClearTransmitQueue(call, 0);
3416 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3417 rxi_ClearTransmitQueue(call, 0);
3418 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3420 if (np->header.type == RX_PACKET_TYPE_ACK) {
3421 /* now check to see if this is an ack packet acknowledging that the
3422 * server actually *lost* some hard-acked data. If this happens we
3423 * ignore this packet, as it may indicate that the server restarted in
3424 * the middle of a call. It is also possible that this is an old ack
3425 * packet. We don't abort the connection in this case, because this
3426 * *might* just be an old ack packet. The right way to detect a server
3427 * restart in the midst of a call is to notice that the server epoch
3429 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3430 * XXX unacknowledged. I think that this is off-by-one, but
3431 * XXX I don't dare change it just yet, since it will
3432 * XXX interact badly with the server-restart detection
3433 * XXX code in receiveackpacket. */
3434 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3435 if (rx_stats_active)
3436 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3437 MUTEX_EXIT(&call->lock);
3438 MUTEX_ENTER(&rx_refcnt_mutex);
3440 MUTEX_EXIT(&rx_refcnt_mutex);
3444 } /* else not a data packet */
3447 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3448 /* Set remote user defined status from packet */
3449 call->remoteStatus = np->header.userStatus;
3451 /* Note the gap between the expected next packet and the actual
3452 * packet that arrived, when the new packet has a smaller serial number
3453 * than expected. Rioses frequently reorder packets all by themselves,
3454 * so this will be quite important with very large window sizes.
3455 * Skew is checked against 0 here to avoid any dependence on the type of
3456 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3458 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3459 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3460 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3462 MUTEX_ENTER(&conn->conn_data_lock);
3463 skew = conn->lastSerial - np->header.serial;
3464 conn->lastSerial = np->header.serial;
3465 MUTEX_EXIT(&conn->conn_data_lock);
3467 struct rx_peer *peer;
3469 if (skew > peer->inPacketSkew) {
3470 dpf(("*** In skew changed from %d to %d\n",
3471 peer->inPacketSkew, skew));
3472 peer->inPacketSkew = skew;
3476 /* Now do packet type-specific processing */
3477 switch (np->header.type) {
3478 case RX_PACKET_TYPE_DATA:
3479 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3482 case RX_PACKET_TYPE_ACK:
3483 /* Respond immediately to ack packets requesting acknowledgement
3485 if (np->header.flags & RX_REQUEST_ACK) {
3487 (void)rxi_SendCallAbort(call, 0, 1, 0);
3489 (void)rxi_SendAck(call, 0, np->header.serial,
3490 RX_ACK_PING_RESPONSE, 1);
3492 np = rxi_ReceiveAckPacket(call, np, 1);
3494 case RX_PACKET_TYPE_ABORT: {
3495 /* An abort packet: reset the call, passing the error up to the user. */
3496 /* What if error is zero? */
3497 /* What if the error is -1? the application will treat it as a timeout. */
3498 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3499 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3500 rxi_CallError(call, errdata);
3501 MUTEX_EXIT(&call->lock);
3502 MUTEX_ENTER(&rx_refcnt_mutex);
3504 MUTEX_EXIT(&rx_refcnt_mutex);
3505 return np; /* xmitting; drop packet */
3507 case RX_PACKET_TYPE_BUSY: {
3508 struct clock busyTime;
3510 clock_GetTime(&busyTime);
3512 MUTEX_EXIT(&call->lock);
3514 MUTEX_ENTER(&conn->conn_call_lock);
3515 MUTEX_ENTER(&call->lock);
3516 conn->lastBusy[call->channel] = busyTime.sec;
3517 call->flags |= RX_CALL_PEER_BUSY;
3518 MUTEX_EXIT(&call->lock);
3519 MUTEX_EXIT(&conn->conn_call_lock);
3521 MUTEX_ENTER(&rx_refcnt_mutex);
3523 MUTEX_EXIT(&rx_refcnt_mutex);
3527 case RX_PACKET_TYPE_ACKALL:
3528 /* All packets acknowledged, so we can drop all packets previously
3529 * readied for sending */
3530 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3531 /* XXX Hack. We because we can't release the global rx lock when
3532 * sending packets (osi_NetSend) we drop all ack pkts while we're
3533 * traversing the tq in rxi_Start sending packets out because
3534 * packets may move to the freePacketQueue as result of being
3535 * here! So we drop these packets until we're safely out of the
3536 * traversing. Really ugly!
3537 * For fine grain RX locking, we set the acked field in the packets
3538 * and let rxi_Start remove the packets from the transmit queue.
3540 if (call->flags & RX_CALL_TQ_BUSY) {
3541 #ifdef RX_ENABLE_LOCKS
3542 rxi_SetAcksInTransmitQueue(call);
3544 #else /* RX_ENABLE_LOCKS */
3545 MUTEX_EXIT(&call->lock);
3546 MUTEX_ENTER(&rx_refcnt_mutex);
3548 MUTEX_EXIT(&rx_refcnt_mutex);
3549 return np; /* xmitting; drop packet */
3550 #endif /* RX_ENABLE_LOCKS */
3552 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3553 rxi_ClearTransmitQueue(call, 0);
3556 /* Should not reach here, unless the peer is broken: send an abort
3558 rxi_CallError(call, RX_PROTOCOL_ERROR);
3559 np = rxi_SendCallAbort(call, np, 1, 0);
3562 /* Note when this last legitimate packet was received, for keep-alive
3563 * processing. Note, we delay getting the time until now in the hope that
3564 * the packet will be delivered to the user before any get time is required
3565 * (if not, then the time won't actually be re-evaluated here). */
3566 call->lastReceiveTime = clock_Sec();
3567 /* we've received a legit packet, so the channel is not busy */
3568 call->flags &= ~RX_CALL_PEER_BUSY;
3569 MUTEX_EXIT(&call->lock);
3570 MUTEX_ENTER(&rx_refcnt_mutex);
3572 MUTEX_EXIT(&rx_refcnt_mutex);
3576 /* return true if this is an "interesting" connection from the point of view
3577 of someone trying to debug the system */
3579 rxi_IsConnInteresting(struct rx_connection *aconn)
3582 struct rx_call *tcall;
3584 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3587 for (i = 0; i < RX_MAXCALLS; i++) {
3588 tcall = aconn->call[i];
3590 if ((tcall->state == RX_STATE_PRECALL)
3591 || (tcall->state == RX_STATE_ACTIVE))
3593 if ((tcall->mode == RX_MODE_SENDING)
3594 || (tcall->mode == RX_MODE_RECEIVING))
3602 /* if this is one of the last few packets AND it wouldn't be used by the
3603 receiving call to immediately satisfy a read request, then drop it on
3604 the floor, since accepting it might prevent a lock-holding thread from
3605 making progress in its reading. If a call has been cleared while in
3606 the precall state then ignore all subsequent packets until the call
3607 is assigned to a thread. */
3610 TooLow(struct rx_packet *ap, struct rx_call *acall)
3614 MUTEX_ENTER(&rx_quota_mutex);
3615 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3616 && (acall->state == RX_STATE_PRECALL))
3617 || ((rx_nFreePackets < rxi_dataQuota + 2)
3618 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3619 && (acall->flags & RX_CALL_READER_WAIT)))) {
3622 MUTEX_EXIT(&rx_quota_mutex);
3628 * Clear the attach wait flag on a connection and proceed.
3630 * Any processing waiting for a connection to be attached should be
3631 * unblocked. We clear the flag and do any other needed tasks.
3634 * the conn to unmark waiting for attach
3636 * @pre conn's conn_data_lock must be locked before calling this function
3640 rxi_ConnClearAttachWait(struct rx_connection *conn)
3642 /* Indicate that rxi_CheckReachEvent is no longer running by
3643 * clearing the flag. Must be atomic under conn_data_lock to
3644 * avoid a new call slipping by: rxi_CheckConnReach holds
3645 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3647 conn->flags &= ~RX_CONN_ATTACHWAIT;
3648 if (conn->flags & RX_CONN_NAT_PING) {
3649 conn->flags &= ~RX_CONN_NAT_PING;
3650 rxi_ScheduleNatKeepAliveEvent(conn);
3655 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3657 struct rx_connection *conn = arg1;
3658 struct rx_call *acall = arg2;
3659 struct rx_call *call = acall;
3660 struct clock when, now;
3663 MUTEX_ENTER(&conn->conn_data_lock);
3664 conn->checkReachEvent = NULL;
3665 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3667 MUTEX_ENTER(&rx_refcnt_mutex);
3669 MUTEX_EXIT(&rx_refcnt_mutex);
3671 MUTEX_EXIT(&conn->conn_data_lock);
3675 MUTEX_ENTER(&conn->conn_call_lock);
3676 MUTEX_ENTER(&conn->conn_data_lock);
3677 for (i = 0; i < RX_MAXCALLS; i++) {
3678 struct rx_call *tc = conn->call[i];
3679 if (tc && tc->state == RX_STATE_PRECALL) {
3685 rxi_ConnClearAttachWait(conn);
3686 MUTEX_EXIT(&conn->conn_data_lock);
3687 MUTEX_EXIT(&conn->conn_call_lock);
3692 MUTEX_ENTER(&call->lock);
3693 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3695 MUTEX_EXIT(&call->lock);
3697 clock_GetTime(&now);
3699 when.sec += RX_CHECKREACH_TIMEOUT;
3700 MUTEX_ENTER(&conn->conn_data_lock);
3701 if (!conn->checkReachEvent) {
3702 MUTEX_ENTER(&rx_refcnt_mutex);
3704 MUTEX_EXIT(&rx_refcnt_mutex);
3705 conn->checkReachEvent =
3706 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3709 MUTEX_EXIT(&conn->conn_data_lock);
3715 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3717 struct rx_service *service = conn->service;
3718 struct rx_peer *peer = conn->peer;
3719 afs_uint32 now, lastReach;
3721 if (service->checkReach == 0)
3725 MUTEX_ENTER(&peer->peer_lock);
3726 lastReach = peer->lastReachTime;
3727 MUTEX_EXIT(&peer->peer_lock);
3728 if (now - lastReach < RX_CHECKREACH_TTL)
3731 MUTEX_ENTER(&conn->conn_data_lock);
3732 if (conn->flags & RX_CONN_ATTACHWAIT) {
3733 MUTEX_EXIT(&conn->conn_data_lock);
3736 conn->flags |= RX_CONN_ATTACHWAIT;
3737 MUTEX_EXIT(&conn->conn_data_lock);
3738 if (!conn->checkReachEvent)
3739 rxi_CheckReachEvent(NULL, conn, call);
3744 /* try to attach call, if authentication is complete */
3746 TryAttach(struct rx_call *acall, osi_socket socket,
3747 int *tnop, struct rx_call **newcallp,
3750 struct rx_connection *conn = acall->conn;
3752 if (conn->type == RX_SERVER_CONNECTION
3753 && acall->state == RX_STATE_PRECALL) {
3754 /* Don't attach until we have any req'd. authentication. */
3755 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3756 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3757 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3758 /* Note: this does not necessarily succeed; there
3759 * may not any proc available
3762 rxi_ChallengeOn(acall->conn);
3767 /* A data packet has been received off the interface. This packet is
3768 * appropriate to the call (the call is in the right state, etc.). This
3769 * routine can return a packet to the caller, for re-use */
3772 rxi_ReceiveDataPacket(struct rx_call *call,
3773 struct rx_packet *np, int istack,
3774 osi_socket socket, afs_uint32 host, u_short port,
3775 int *tnop, struct rx_call **newcallp)
3777 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3782 afs_uint32 serial=0, flags=0;
3784 struct rx_packet *tnp;
3785 struct clock when, now;
3786 if (rx_stats_active)
3787 rx_atomic_inc(&rx_stats.dataPacketsRead);
3790 /* If there are no packet buffers, drop this new packet, unless we can find
3791 * packet buffers from inactive calls */
3793 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3794 MUTEX_ENTER(&rx_freePktQ_lock);
3795 rxi_NeedMorePackets = TRUE;
3796 MUTEX_EXIT(&rx_freePktQ_lock);
3797 if (rx_stats_active)
3798 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3799 call->rprev = np->header.serial;
3800 rxi_calltrace(RX_TRACE_DROP, call);
3801 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3803 rxi_ClearReceiveQueue(call);
3804 clock_GetTime(&now);
3806 clock_Add(&when, &rx_softAckDelay);
3807 if (!call->delayedAckEvent
3808 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3809 rxevent_Cancel(call->delayedAckEvent, call,
3810 RX_CALL_REFCOUNT_DELAY);
3811 MUTEX_ENTER(&rx_refcnt_mutex);
3812 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3813 MUTEX_EXIT(&rx_refcnt_mutex);
3815 call->delayedAckEvent =
3816 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3818 /* we've damaged this call already, might as well do it in. */
3824 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3825 * packet is one of several packets transmitted as a single
3826 * datagram. Do not send any soft or hard acks until all packets
3827 * in a jumbogram have been processed. Send negative acks right away.
3829 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3830 /* tnp is non-null when there are more packets in the
3831 * current jumbo gram */
3838 seq = np->header.seq;
3839 serial = np->header.serial;
3840 flags = np->header.flags;
3842 /* If the call is in an error state, send an abort message */
3844 return rxi_SendCallAbort(call, np, istack, 0);
3846 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3847 * AFS 3.5 jumbogram. */
3848 if (flags & RX_JUMBO_PACKET) {
3849 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3854 if (np->header.spare != 0) {
3855 MUTEX_ENTER(&call->conn->conn_data_lock);
3856 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3857 MUTEX_EXIT(&call->conn->conn_data_lock);
3860 /* The usual case is that this is the expected next packet */
3861 if (seq == call->rnext) {
3863 /* Check to make sure it is not a duplicate of one already queued */
3864 if (queue_IsNotEmpty(&call->rq)
3865 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3866 if (rx_stats_active)
3867 rx_atomic_inc(&rx_stats.dupPacketsRead);
3868 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3869 rxevent_Cancel(call->delayedAckEvent, call,
3870 RX_CALL_REFCOUNT_DELAY);
3871 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3877 /* It's the next packet. Stick it on the receive queue
3878 * for this call. Set newPackets to make sure we wake
3879 * the reader once all packets have been processed */
3880 #ifdef RX_TRACK_PACKETS
3881 np->flags |= RX_PKTFLAG_RQ;
3883 queue_Prepend(&call->rq, np);
3884 #ifdef RXDEBUG_PACKET
3886 #endif /* RXDEBUG_PACKET */
3888 np = NULL; /* We can't use this anymore */
3891 /* If an ack is requested then set a flag to make sure we
3892 * send an acknowledgement for this packet */
3893 if (flags & RX_REQUEST_ACK) {
3894 ackNeeded = RX_ACK_REQUESTED;
3897 /* Keep track of whether we have received the last packet */
3898 if (flags & RX_LAST_PACKET) {
3899 call->flags |= RX_CALL_HAVE_LAST;
3903 /* Check whether we have all of the packets for this call */
3904 if (call->flags & RX_CALL_HAVE_LAST) {
3905 afs_uint32 tseq; /* temporary sequence number */
3906 struct rx_packet *tp; /* Temporary packet pointer */
3907 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3909 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3910 if (tseq != tp->header.seq)
3912 if (tp->header.flags & RX_LAST_PACKET) {
3913 call->flags |= RX_CALL_RECEIVE_DONE;
3920 /* Provide asynchronous notification for those who want it
3921 * (e.g. multi rx) */
3922 if (call->arrivalProc) {
3923 (*call->arrivalProc) (call, call->arrivalProcHandle,
3924 call->arrivalProcArg);
3925 call->arrivalProc = (void (*)())0;
3928 /* Update last packet received */
3931 /* If there is no server process serving this call, grab
3932 * one, if available. We only need to do this once. If a
3933 * server thread is available, this thread becomes a server
3934 * thread and the server thread becomes a listener thread. */
3936 TryAttach(call, socket, tnop, newcallp, 0);
3939 /* This is not the expected next packet. */
3941 /* Determine whether this is a new or old packet, and if it's
3942 * a new one, whether it fits into the current receive window.
3943 * Also figure out whether the packet was delivered in sequence.
3944 * We use the prev variable to determine whether the new packet
3945 * is the successor of its immediate predecessor in the
3946 * receive queue, and the missing flag to determine whether
3947 * any of this packets predecessors are missing. */
3949 afs_uint32 prev; /* "Previous packet" sequence number */
3950 struct rx_packet *tp; /* Temporary packet pointer */
3951 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3952 int missing; /* Are any predecessors missing? */
3954 /* If the new packet's sequence number has been sent to the
3955 * application already, then this is a duplicate */
3956 if (seq < call->rnext) {
3957 if (rx_stats_active)
3958 rx_atomic_inc(&rx_stats.dupPacketsRead);
3959 rxevent_Cancel(call->delayedAckEvent, call,
3960 RX_CALL_REFCOUNT_DELAY);
3961 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3967 /* If the sequence number is greater than what can be
3968 * accomodated by the current window, then send a negative
3969 * acknowledge and drop the packet */
3970 if ((call->rnext + call->rwind) <= seq) {
3971 rxevent_Cancel(call->delayedAckEvent, call,
3972 RX_CALL_REFCOUNT_DELAY);
3973 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3980 /* Look for the packet in the queue of old received packets */
3981 for (prev = call->rnext - 1, missing =
3982 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3983 /*Check for duplicate packet */
3984 if (seq == tp->header.seq) {
3985 if (rx_stats_active)
3986 rx_atomic_inc(&rx_stats.dupPacketsRead);
3987 rxevent_Cancel(call->delayedAckEvent, call,
3988 RX_CALL_REFCOUNT_DELAY);
3989 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3995 /* If we find a higher sequence packet, break out and
3996 * insert the new packet here. */
3997 if (seq < tp->header.seq)
3999 /* Check for missing packet */
4000 if (tp->header.seq != prev + 1) {
4004 prev = tp->header.seq;
4007 /* Keep track of whether we have received the last packet. */
4008 if (flags & RX_LAST_PACKET) {
4009 call->flags |= RX_CALL_HAVE_LAST;
4012 /* It's within the window: add it to the the receive queue.
4013 * tp is left by the previous loop either pointing at the
4014 * packet before which to insert the new packet, or at the
4015 * queue head if the queue is empty or the packet should be
4017 #ifdef RX_TRACK_PACKETS
4018 np->flags |= RX_PKTFLAG_RQ;
4020 #ifdef RXDEBUG_PACKET
4022 #endif /* RXDEBUG_PACKET */
4023 queue_InsertBefore(tp, np);
4027 /* Check whether we have all of the packets for this call */
4028 if ((call->flags & RX_CALL_HAVE_LAST)
4029 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4030 afs_uint32 tseq; /* temporary sequence number */
4033 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
4034 if (tseq != tp->header.seq)
4036 if (tp->header.flags & RX_LAST_PACKET) {
4037 call->flags |= RX_CALL_RECEIVE_DONE;
4044 /* We need to send an ack of the packet is out of sequence,
4045 * or if an ack was requested by the peer. */
4046 if (seq != prev + 1 || missing) {
4047 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4048 } else if (flags & RX_REQUEST_ACK) {
4049 ackNeeded = RX_ACK_REQUESTED;
4052 /* Acknowledge the last packet for each call */
4053 if (flags & RX_LAST_PACKET) {
4064 * If the receiver is waiting for an iovec, fill the iovec
4065 * using the data from the receive queue */
4066 if (call->flags & RX_CALL_IOVEC_WAIT) {
4067 didHardAck = rxi_FillReadVec(call, serial);
4068 /* the call may have been aborted */
4077 /* Wakeup the reader if any */
4078 if ((call->flags & RX_CALL_READER_WAIT)
4079 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4080 || (call->iovNext >= call->iovMax)
4081 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4082 call->flags &= ~RX_CALL_READER_WAIT;
4083 #ifdef RX_ENABLE_LOCKS
4084 CV_BROADCAST(&call->cv_rq);
4086 osi_rxWakeup(&call->rq);
4092 * Send an ack when requested by the peer, or once every
4093 * rxi_SoftAckRate packets until the last packet has been
4094 * received. Always send a soft ack for the last packet in
4095 * the server's reply. */
4097 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4098 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4099 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4100 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4101 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4102 } else if (call->nSoftAcks) {
4103 clock_GetTime(&now);
4105 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
4106 clock_Add(&when, &rx_lastAckDelay);
4108 clock_Add(&when, &rx_softAckDelay);
4110 if (!call->delayedAckEvent
4111 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
4112 rxevent_Cancel(call->delayedAckEvent, call,
4113 RX_CALL_REFCOUNT_DELAY);
4114 MUTEX_ENTER(&rx_refcnt_mutex);
4115 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
4116 MUTEX_EXIT(&rx_refcnt_mutex);
4117 call->delayedAckEvent =
4118 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
4120 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4121 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4128 static void rxi_ComputeRate();
4132 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4134 struct rx_peer *peer = conn->peer;
4136 MUTEX_ENTER(&peer->peer_lock);
4137 peer->lastReachTime = clock_Sec();
4138 MUTEX_EXIT(&peer->peer_lock);
4140 MUTEX_ENTER(&conn->conn_data_lock);
4141 if (conn->flags & RX_CONN_ATTACHWAIT) {
4144 rxi_ConnClearAttachWait(conn);
4145 MUTEX_EXIT(&conn->conn_data_lock);
4147 for (i = 0; i < RX_MAXCALLS; i++) {
4148 struct rx_call *call = conn->call[i];
4151 MUTEX_ENTER(&call->lock);
4152 /* tnop can be null if newcallp is null */
4153 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4155 MUTEX_EXIT(&call->lock);
4159 MUTEX_EXIT(&conn->conn_data_lock);
4162 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4164 rx_ack_reason(int reason)
4167 case RX_ACK_REQUESTED:
4169 case RX_ACK_DUPLICATE:
4171 case RX_ACK_OUT_OF_SEQUENCE:
4173 case RX_ACK_EXCEEDS_WINDOW:
4175 case RX_ACK_NOSPACE:
4179 case RX_ACK_PING_RESPONSE:
4192 /* The real smarts of the whole thing. */
4194 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4197 struct rx_ackPacket *ap;
4199 struct rx_packet *tp;
4200 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
4201 struct rx_connection *conn = call->conn;
4202 struct rx_peer *peer = conn->peer;
4203 struct clock now; /* Current time, for RTT calculations */
4207 /* because there are CM's that are bogus, sending weird values for this. */
4208 afs_uint32 skew = 0;
4213 int newAckCount = 0;
4214 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4215 int pktsize = 0; /* Set if we need to update the peer mtu */
4216 int conn_data_locked = 0;
4218 if (rx_stats_active)
4219 rx_atomic_inc(&rx_stats.ackPacketsRead);
4220 ap = (struct rx_ackPacket *)rx_DataOf(np);
4221 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4223 return np; /* truncated ack packet */
4225 /* depends on ack packet struct */
4226 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4227 first = ntohl(ap->firstPacket);
4228 prev = ntohl(ap->previousPacket);
4229 serial = ntohl(ap->serial);
4230 /* temporarily disabled -- needs to degrade over time
4231 * skew = ntohs(ap->maxSkew); */
4233 /* Ignore ack packets received out of order */
4234 if (first < call->tfirst ||
4235 (first == call->tfirst && prev < call->tprev)) {
4241 if (np->header.flags & RX_SLOW_START_OK) {
4242 call->flags |= RX_CALL_SLOW_START_OK;
4245 if (ap->reason == RX_ACK_PING_RESPONSE)
4246 rxi_UpdatePeerReach(conn, call);
4248 if (conn->lastPacketSizeSeq) {
4249 MUTEX_ENTER(&conn->conn_data_lock);
4250 conn_data_locked = 1;
4251 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4252 pktsize = conn->lastPacketSize;
4253 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4256 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4257 if (!conn_data_locked) {
4258 MUTEX_ENTER(&conn->conn_data_lock);
4259 conn_data_locked = 1;
4261 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4262 /* process mtu ping ack */
4263 pktsize = conn->lastPingSize;
4264 conn->lastPingSizeSer = conn->lastPingSize = 0;
4268 if (conn_data_locked) {
4269 MUTEX_EXIT(&conn->conn_data_lock);
4270 conn_data_locked = 0;
4274 if (rxdebug_active) {
4278 len = _snprintf(msg, sizeof(msg),
4279 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
4280 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4281 ntohl(ap->serial), ntohl(ap->previousPacket),
4282 (unsigned int)np->header.seq, (unsigned int)skew,
4283 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
4287 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4288 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4292 OutputDebugString(msg);
4294 #else /* AFS_NT40_ENV */
4297 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
4298 ap->reason, ntohl(ap->previousPacket),
4299 (unsigned int)np->header.seq, (unsigned int)serial,
4300 (unsigned int)skew, ntohl(ap->firstPacket));
4303 for (offset = 0; offset < nAcks; offset++)
4304 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4309 #endif /* AFS_NT40_ENV */
4312 MUTEX_ENTER(&peer->peer_lock);
4315 * Start somewhere. Can't assume we can send what we can receive,
4316 * but we are clearly receiving.
4318 if (!peer->maxPacketSize)
4319 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4321 if (pktsize > peer->maxPacketSize) {
4322 peer->maxPacketSize = pktsize;
4323 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4324 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4325 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4326 rxi_ScheduleGrowMTUEvent(call, 1);
4331 /* Update the outgoing packet skew value to the latest value of
4332 * the peer's incoming packet skew value. The ack packet, of
4333 * course, could arrive out of order, but that won't affect things
4335 peer->outPacketSkew = skew;
4338 clock_GetTime(&now);
4340 /* The transmit queue splits into 4 sections.
4342 * The first section is packets which have now been acknowledged
4343 * by a window size change in the ack. These have reached the
4344 * application layer, and may be discarded. These are packets
4345 * with sequence numbers < ap->firstPacket.
4347 * The second section is packets which have sequence numbers in
4348 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4349 * contents of the packet's ack array determines whether these
4350 * packets are acknowledged or not.
4352 * The third section is packets which fall above the range
4353 * addressed in the ack packet. These have not yet been received
4356 * The four section is packets which have not yet been transmitted.
4357 * These packets will have a header.serial of 0.
4360 /* First section - implicitly acknowledged packets that can be
4364 tp = queue_First(&call->tq, rx_packet);
4365 while(!queue_IsEnd(&call->tq, tp) && tp->header.seq < first) {
4366 struct rx_packet *next;
4368 next = queue_Next(tp, rx_packet);
4369 call->tfirst = tp->header.seq + 1;
4371 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4373 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4377 rxi_ComputeRate(call->conn->peer, call, p, np, ap->reason);
4380 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4381 /* XXX Hack. Because we have to release the global rx lock when sending
4382 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4383 * in rxi_Start sending packets out because packets may move to the
4384 * freePacketQueue as result of being here! So we drop these packets until
4385 * we're safely out of the traversing. Really ugly!
4386 * To make it even uglier, if we're using fine grain locking, we can
4387 * set the ack bits in the packets and have rxi_Start remove the packets
4388 * when it's done transmitting.
4390 if (call->flags & RX_CALL_TQ_BUSY) {
4391 #ifdef RX_ENABLE_LOCKS
4392 tp->flags |= RX_PKTFLAG_ACKED;
4393 call->flags |= RX_CALL_TQ_SOME_ACKED;
4394 #else /* RX_ENABLE_LOCKS */
4396 #endif /* RX_ENABLE_LOCKS */
4398 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4401 #ifdef RX_TRACK_PACKETS
4402 tp->flags &= ~RX_PKTFLAG_TQ;
4404 #ifdef RXDEBUG_PACKET
4406 #endif /* RXDEBUG_PACKET */
4407 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4413 /* Give rate detector a chance to respond to ping requests */
4414 if (ap->reason == RX_ACK_PING_RESPONSE) {
4415 rxi_ComputeRate(peer, call, 0, np, ap->reason);
4419 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4421 /* Second section of the queue - packets for which we are receiving
4424 * Go through the explicit acks/nacks and record the results in
4425 * the waiting packets. These are packets that can't be released
4426 * yet, even with a positive acknowledge. This positive
4427 * acknowledge only means the packet has been received by the
4428 * peer, not that it will be retained long enough to be sent to
4429 * the peer's upper level. In addition, reset the transmit timers
4430 * of any missing packets (those packets that must be missing
4431 * because this packet was out of sequence) */
4433 call->nSoftAcked = 0;
4435 while (!queue_IsEnd(&call->tq, tp) && tp->header.seq < first + nAcks) {
4436 /* Set the acknowledge flag per packet based on the
4437 * information in the ack packet. An acknowlegded packet can
4438 * be downgraded when the server has discarded a packet it
4439 * soacked previously, or when an ack packet is received
4440 * out of sequence. */
4441 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4442 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4444 tp->flags |= RX_PKTFLAG_ACKED;
4445 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4447 rxi_ComputeRate(call->conn->peer, call, tp, np, ap->reason);
4455 } else /* RX_ACK_TYPE_NACK */ {
4456 tp->flags &= ~RX_PKTFLAG_ACKED;
4460 tp = queue_Next(tp, rx_packet);
4463 /* We don't need to take any action with the 3rd or 4th section in the
4464 * queue - they're not addressed by the contents of this ACK packet.
4467 /* If the window has been extended by this acknowledge packet,
4468 * then wakeup a sender waiting in alloc for window space, or try
4469 * sending packets now, if he's been sitting on packets due to
4470 * lack of window space */
4471 if (call->tnext < (call->tfirst + call->twind)) {
4472 #ifdef RX_ENABLE_LOCKS
4473 CV_SIGNAL(&call->cv_twind);
4475 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4476 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4477 osi_rxWakeup(&call->twind);
4480 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4481 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4485 /* if the ack packet has a receivelen field hanging off it,
4486 * update our state */
4487 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4490 /* If the ack packet has a "recommended" size that is less than
4491 * what I am using now, reduce my size to match */
4492 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4493 (int)sizeof(afs_int32), &tSize);
4494 tSize = (afs_uint32) ntohl(tSize);
4495 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4497 /* Get the maximum packet size to send to this peer */
4498 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4500 tSize = (afs_uint32) ntohl(tSize);
4501 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4502 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4504 /* sanity check - peer might have restarted with different params.
4505 * If peer says "send less", dammit, send less... Peer should never
4506 * be unable to accept packets of the size that prior AFS versions would
4507 * send without asking. */
4508 if (peer->maxMTU != tSize) {
4509 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4511 peer->maxMTU = tSize;
4512 peer->MTU = MIN(tSize, peer->MTU);
4513 call->MTU = MIN(call->MTU, tSize);
4516 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4519 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4520 (int)sizeof(afs_int32), &tSize);
4521 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4522 if (tSize < call->twind) { /* smaller than our send */
4523 call->twind = tSize; /* window, we must send less... */
4524 call->ssthresh = MIN(call->twind, call->ssthresh);
4525 call->conn->twind[call->channel] = call->twind;
4528 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4529 * network MTU confused with the loopback MTU. Calculate the
4530 * maximum MTU here for use in the slow start code below.
4532 /* Did peer restart with older RX version? */
4533 if (peer->maxDgramPackets > 1) {
4534 peer->maxDgramPackets = 1;
4536 } else if (np->length >=
4537 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4540 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4541 sizeof(afs_int32), &tSize);
4542 tSize = (afs_uint32) ntohl(tSize);
4544 * As of AFS 3.5 we set the send window to match the receive window.
4546 if (tSize < call->twind) {
4547 call->twind = tSize;
4548 call->conn->twind[call->channel] = call->twind;
4549 call->ssthresh = MIN(call->twind, call->ssthresh);
4550 } else if (tSize > call->twind) {
4551 call->twind = tSize;
4552 call->conn->twind[call->channel] = call->twind;
4556 * As of AFS 3.5, a jumbogram is more than one fixed size
4557 * packet transmitted in a single UDP datagram. If the remote
4558 * MTU is smaller than our local MTU then never send a datagram
4559 * larger than the natural MTU.
4562 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4563 (int)sizeof(afs_int32), &tSize);
4564 maxDgramPackets = (afs_uint32) ntohl(tSize);
4565 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4567 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4568 if (maxDgramPackets > 1) {
4569 peer->maxDgramPackets = maxDgramPackets;
4570 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4572 peer->maxDgramPackets = 1;
4573 call->MTU = peer->natMTU;
4575 } else if (peer->maxDgramPackets > 1) {
4576 /* Restarted with lower version of RX */
4577 peer->maxDgramPackets = 1;
4579 } else if (peer->maxDgramPackets > 1
4580 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4581 /* Restarted with lower version of RX */
4582 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4583 peer->natMTU = OLD_MAX_PACKET_SIZE;
4584 peer->MTU = OLD_MAX_PACKET_SIZE;
4585 peer->maxDgramPackets = 1;
4586 peer->nDgramPackets = 1;
4588 call->MTU = OLD_MAX_PACKET_SIZE;
4593 * Calculate how many datagrams were successfully received after
4594 * the first missing packet and adjust the negative ack counter
4599 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4600 if (call->nNacks < nNacked) {
4601 call->nNacks = nNacked;
4604 call->nAcks += newAckCount;
4608 /* If the packet contained new acknowledgements, rather than just
4609 * being a duplicate of one we have previously seen, then we can restart
4612 if (newAckCount > 0)
4613 rxi_rto_packet_acked(call, istack);
4615 if (call->flags & RX_CALL_FAST_RECOVER) {
4616 if (newAckCount == 0) {
4617 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4619 call->flags &= ~RX_CALL_FAST_RECOVER;
4620 call->cwind = call->nextCwind;
4621 call->nextCwind = 0;
4624 call->nCwindAcks = 0;
4625 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4626 /* Three negative acks in a row trigger congestion recovery */
4627 call->flags |= RX_CALL_FAST_RECOVER;
4628 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4630 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4631 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4632 call->nextCwind = call->ssthresh;
4635 peer->MTU = call->MTU;
4636 peer->cwind = call->nextCwind;
4637 peer->nDgramPackets = call->nDgramPackets;
4639 call->congestSeq = peer->congestSeq;
4641 /* Reset the resend times on the packets that were nacked
4642 * so we will retransmit as soon as the window permits
4645 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4647 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4648 tp->flags &= ~RX_PKTFLAG_SENT;
4650 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4655 /* If cwind is smaller than ssthresh, then increase
4656 * the window one packet for each ack we receive (exponential
4658 * If cwind is greater than or equal to ssthresh then increase
4659 * the congestion window by one packet for each cwind acks we
4660 * receive (linear growth). */
4661 if (call->cwind < call->ssthresh) {
4663 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4664 call->nCwindAcks = 0;
4666 call->nCwindAcks += newAckCount;
4667 if (call->nCwindAcks >= call->cwind) {
4668 call->nCwindAcks = 0;
4669 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4673 * If we have received several acknowledgements in a row then
4674 * it is time to increase the size of our datagrams
4676 if ((int)call->nAcks > rx_nDgramThreshold) {
4677 if (peer->maxDgramPackets > 1) {
4678 if (call->nDgramPackets < peer->maxDgramPackets) {
4679 call->nDgramPackets++;
4681 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4682 } else if (call->MTU < peer->maxMTU) {
4683 /* don't upgrade if we can't handle it */
4684 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4685 call->MTU = peer->ifMTU;
4687 call->MTU += peer->natMTU;
4688 call->MTU = MIN(call->MTU, peer->maxMTU);
4695 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4697 /* Servers need to hold the call until all response packets have
4698 * been acknowledged. Soft acks are good enough since clients
4699 * are not allowed to clear their receive queues. */
4700 if (call->state == RX_STATE_HOLD
4701 && call->tfirst + call->nSoftAcked >= call->tnext) {
4702 call->state = RX_STATE_DALLY;
4703 rxi_ClearTransmitQueue(call, 0);
4704 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4705 } else if (!queue_IsEmpty(&call->tq)) {
4706 rxi_Start(call, istack);
4711 /* Received a response to a challenge packet */
4713 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4714 struct rx_packet *np, int istack)
4718 /* Ignore the packet if we're the client */
4719 if (conn->type == RX_CLIENT_CONNECTION)
4722 /* If already authenticated, ignore the packet (it's probably a retry) */
4723 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4726 /* Otherwise, have the security object evaluate the response packet */
4727 error = RXS_CheckResponse(conn->securityObject, conn, np);
4729 /* If the response is invalid, reset the connection, sending
4730 * an abort to the peer */
4734 rxi_ConnectionError(conn, error);
4735 MUTEX_ENTER(&conn->conn_data_lock);
4736 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4737 MUTEX_EXIT(&conn->conn_data_lock);
4740 /* If the response is valid, any calls waiting to attach
4741 * servers can now do so */
4744 for (i = 0; i < RX_MAXCALLS; i++) {
4745 struct rx_call *call = conn->call[i];
4747 MUTEX_ENTER(&call->lock);
4748 if (call->state == RX_STATE_PRECALL)
4749 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4750 /* tnop can be null if newcallp is null */
4751 MUTEX_EXIT(&call->lock);
4755 /* Update the peer reachability information, just in case
4756 * some calls went into attach-wait while we were waiting
4757 * for authentication..
4759 rxi_UpdatePeerReach(conn, NULL);
4764 /* A client has received an authentication challenge: the security
4765 * object is asked to cough up a respectable response packet to send
4766 * back to the server. The server is responsible for retrying the
4767 * challenge if it fails to get a response. */
4770 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4771 struct rx_packet *np, int istack)
4775 /* Ignore the challenge if we're the server */
4776 if (conn->type == RX_SERVER_CONNECTION)
4779 /* Ignore the challenge if the connection is otherwise idle; someone's
4780 * trying to use us as an oracle. */
4781 if (!rxi_HasActiveCalls(conn))
4784 /* Send the security object the challenge packet. It is expected to fill
4785 * in the response. */
4786 error = RXS_GetResponse(conn->securityObject, conn, np);
4788 /* If the security object is unable to return a valid response, reset the
4789 * connection and send an abort to the peer. Otherwise send the response
4790 * packet to the peer connection. */
4792 rxi_ConnectionError(conn, error);
4793 MUTEX_ENTER(&conn->conn_data_lock);
4794 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4795 MUTEX_EXIT(&conn->conn_data_lock);
4797 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4798 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4804 /* Find an available server process to service the current request in
4805 * the given call structure. If one isn't available, queue up this
4806 * call so it eventually gets one */
4808 rxi_AttachServerProc(struct rx_call *call,
4809 osi_socket socket, int *tnop,
4810 struct rx_call **newcallp)
4812 struct rx_serverQueueEntry *sq;
4813 struct rx_service *service = call->conn->service;
4816 /* May already be attached */
4817 if (call->state == RX_STATE_ACTIVE)
4820 MUTEX_ENTER(&rx_serverPool_lock);
4822 haveQuota = QuotaOK(service);
4823 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4824 /* If there are no processes available to service this call,
4825 * put the call on the incoming call queue (unless it's
4826 * already on the queue).
4828 #ifdef RX_ENABLE_LOCKS
4830 ReturnToServerPool(service);
4831 #endif /* RX_ENABLE_LOCKS */
4833 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4834 call->flags |= RX_CALL_WAIT_PROC;
4835 rx_atomic_inc(&rx_nWaiting);
4836 rx_atomic_inc(&rx_nWaited);
4837 rxi_calltrace(RX_CALL_ARRIVAL, call);
4838 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4839 queue_Append(&rx_incomingCallQueue, call);
4842 sq = queue_Last(&rx_idleServerQueue, rx_serverQueueEntry);
4844 /* If hot threads are enabled, and both newcallp and sq->socketp
4845 * are non-null, then this thread will process the call, and the
4846 * idle server thread will start listening on this threads socket.
4849 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4852 *sq->socketp = socket;
4853 clock_GetTime(&call->startTime);
4854 MUTEX_ENTER(&rx_refcnt_mutex);
4855 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4856 MUTEX_EXIT(&rx_refcnt_mutex);
4860 if (call->flags & RX_CALL_WAIT_PROC) {
4861 /* Conservative: I don't think this should happen */
4862 call->flags &= ~RX_CALL_WAIT_PROC;
4863 if (queue_IsOnQueue(call)) {
4866 rx_atomic_dec(&rx_nWaiting);
4869 call->state = RX_STATE_ACTIVE;
4870 call->mode = RX_MODE_RECEIVING;
4871 #ifdef RX_KERNEL_TRACE
4873 int glockOwner = ISAFS_GLOCK();
4876 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4877 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4883 if (call->flags & RX_CALL_CLEARED) {
4884 /* send an ack now to start the packet flow up again */
4885 call->flags &= ~RX_CALL_CLEARED;
4886 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4888 #ifdef RX_ENABLE_LOCKS
4891 service->nRequestsRunning++;
4892 MUTEX_ENTER(&rx_quota_mutex);
4893 if (service->nRequestsRunning <= service->minProcs)
4896 MUTEX_EXIT(&rx_quota_mutex);
4900 MUTEX_EXIT(&rx_serverPool_lock);
4903 /* Delay the sending of an acknowledge event for a short while, while
4904 * a new call is being prepared (in the case of a client) or a reply
4905 * is being prepared (in the case of a server). Rather than sending
4906 * an ack packet, an ACKALL packet is sent. */
4908 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4910 #ifdef RX_ENABLE_LOCKS
4912 MUTEX_ENTER(&call->lock);
4913 call->delayedAckEvent = NULL;
4914 MUTEX_ENTER(&rx_refcnt_mutex);
4915 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4916 MUTEX_EXIT(&rx_refcnt_mutex);
4918 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4919 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4920 call->flags |= RX_CALL_ACKALL_SENT;
4922 MUTEX_EXIT(&call->lock);
4923 #else /* RX_ENABLE_LOCKS */
4925 call->delayedAckEvent = NULL;
4926 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4927 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4928 call->flags |= RX_CALL_ACKALL_SENT;
4929 #endif /* RX_ENABLE_LOCKS */
4933 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4935 struct rx_call *call = arg1;
4936 #ifdef RX_ENABLE_LOCKS
4938 MUTEX_ENTER(&call->lock);
4939 if (event == call->delayedAckEvent)
4940 call->delayedAckEvent = NULL;
4941 MUTEX_ENTER(&rx_refcnt_mutex);
4942 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4943 MUTEX_EXIT(&rx_refcnt_mutex);
4945 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4947 MUTEX_EXIT(&call->lock);
4948 #else /* RX_ENABLE_LOCKS */
4950 call->delayedAckEvent = NULL;
4951 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4952 #endif /* RX_ENABLE_LOCKS */
4956 #ifdef RX_ENABLE_LOCKS
4957 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4958 * clearing them out.
4961 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4963 struct rx_packet *p, *tp;
4966 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4967 p->flags |= RX_PKTFLAG_ACKED;
4971 call->flags |= RX_CALL_TQ_CLEARME;
4972 call->flags |= RX_CALL_TQ_SOME_ACKED;
4975 rxi_rto_cancel(call);
4977 call->tfirst = call->tnext;
4978 call->nSoftAcked = 0;
4980 if (call->flags & RX_CALL_FAST_RECOVER) {
4981 call->flags &= ~RX_CALL_FAST_RECOVER;
4982 call->cwind = call->nextCwind;
4983 call->nextCwind = 0;
4986 CV_SIGNAL(&call->cv_twind);
4988 #endif /* RX_ENABLE_LOCKS */
4990 /* Clear out the transmit queue for the current call (all packets have
4991 * been received by peer) */
4993 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4995 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4996 struct rx_packet *p, *tp;
4998 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
5000 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
5001 p->flags |= RX_PKTFLAG_ACKED;
5005 call->flags |= RX_CALL_TQ_CLEARME;
5006 call->flags |= RX_CALL_TQ_SOME_ACKED;
5009 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5010 #ifdef RXDEBUG_PACKET
5012 #endif /* RXDEBUG_PACKET */
5013 rxi_FreePackets(0, &call->tq);
5014 rxi_WakeUpTransmitQueue(call);
5015 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5016 call->flags &= ~RX_CALL_TQ_CLEARME;
5018 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5020 rxi_rto_cancel(call);
5021 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5022 call->nSoftAcked = 0;
5024 if (call->flags & RX_CALL_FAST_RECOVER) {
5025 call->flags &= ~RX_CALL_FAST_RECOVER;
5026 call->cwind = call->nextCwind;
5028 #ifdef RX_ENABLE_LOCKS
5029 CV_SIGNAL(&call->cv_twind);
5031 osi_rxWakeup(&call->twind);
5036 rxi_ClearReceiveQueue(struct rx_call *call)
5038 if (queue_IsNotEmpty(&call->rq)) {
5041 count = rxi_FreePackets(0, &call->rq);
5042 rx_packetReclaims += count;
5043 #ifdef RXDEBUG_PACKET
5045 if ( call->rqc != 0 )
5046 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5048 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5050 if (call->state == RX_STATE_PRECALL) {
5051 call->flags |= RX_CALL_CLEARED;
5055 /* Send an abort packet for the specified call */
5057 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5058 int istack, int force)
5061 struct clock when, now;
5066 /* Clients should never delay abort messages */
5067 if (rx_IsClientConn(call->conn))
5070 if (call->abortCode != call->error) {
5071 call->abortCode = call->error;
5072 call->abortCount = 0;
5075 if (force || rxi_callAbortThreshhold == 0
5076 || call->abortCount < rxi_callAbortThreshhold) {
5077 if (call->delayedAbortEvent) {
5078 rxevent_Cancel(call->delayedAbortEvent, call,
5079 RX_CALL_REFCOUNT_ABORT);
5081 error = htonl(call->error);
5084 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5085 (char *)&error, sizeof(error), istack);
5086 } else if (!call->delayedAbortEvent) {
5087 clock_GetTime(&now);
5089 clock_Addmsec(&when, rxi_callAbortDelay);
5090 MUTEX_ENTER(&rx_refcnt_mutex);
5091 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5092 MUTEX_EXIT(&rx_refcnt_mutex);
5093 call->delayedAbortEvent =
5094 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
5099 /* Send an abort packet for the specified connection. Packet is an
5100 * optional pointer to a packet that can be used to send the abort.
5101 * Once the number of abort messages reaches the threshhold, an
5102 * event is scheduled to send the abort. Setting the force flag
5103 * overrides sending delayed abort messages.
5105 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5106 * to send the abort packet.
5109 rxi_SendConnectionAbort(struct rx_connection *conn,
5110 struct rx_packet *packet, int istack, int force)
5113 struct clock when, now;
5118 /* Clients should never delay abort messages */
5119 if (rx_IsClientConn(conn))
5122 if (force || rxi_connAbortThreshhold == 0
5123 || conn->abortCount < rxi_connAbortThreshhold) {
5124 if (conn->delayedAbortEvent) {
5125 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
5127 error = htonl(conn->error);
5129 MUTEX_EXIT(&conn->conn_data_lock);
5131 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5132 RX_PACKET_TYPE_ABORT, (char *)&error,
5133 sizeof(error), istack);
5134 MUTEX_ENTER(&conn->conn_data_lock);
5135 } else if (!conn->delayedAbortEvent) {
5136 clock_GetTime(&now);
5138 clock_Addmsec(&when, rxi_connAbortDelay);
5139 conn->delayedAbortEvent =
5140 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
5145 /* Associate an error all of the calls owned by a connection. Called
5146 * with error non-zero. This is only for really fatal things, like
5147 * bad authentication responses. The connection itself is set in
5148 * error at this point, so that future packets received will be
5151 rxi_ConnectionError(struct rx_connection *conn,
5157 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5159 MUTEX_ENTER(&conn->conn_data_lock);
5160 if (conn->challengeEvent)
5161 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
5162 if (conn->natKeepAliveEvent)
5163 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
5164 if (conn->checkReachEvent) {
5165 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
5166 conn->checkReachEvent = 0;
5167 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5168 MUTEX_ENTER(&rx_refcnt_mutex);
5170 MUTEX_EXIT(&rx_refcnt_mutex);
5172 MUTEX_EXIT(&conn->conn_data_lock);
5173 for (i = 0; i < RX_MAXCALLS; i++) {
5174 struct rx_call *call = conn->call[i];
5176 MUTEX_ENTER(&call->lock);
5177 rxi_CallError(call, error);
5178 MUTEX_EXIT(&call->lock);
5181 conn->error = error;
5182 if (rx_stats_active)
5183 rx_atomic_inc(&rx_stats.fatalErrors);
5188 * Interrupt an in-progress call with the specified error and wakeup waiters.
5190 * @param[in] call The call to interrupt
5191 * @param[in] error The error code to send to the peer
5194 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5196 MUTEX_ENTER(&call->lock);
5197 rxi_CallError(call, error);
5198 rxi_SendCallAbort(call, NULL, 0, 1);
5199 MUTEX_EXIT(&call->lock);
5203 rxi_CallError(struct rx_call *call, afs_int32 error)
5206 osirx_AssertMine(&call->lock, "rxi_CallError");
5208 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5210 error = call->error;
5212 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5213 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5214 rxi_ResetCall(call, 0);
5217 rxi_ResetCall(call, 0);
5219 call->error = error;
5222 /* Reset various fields in a call structure, and wakeup waiting
5223 * processes. Some fields aren't changed: state & mode are not
5224 * touched (these must be set by the caller), and bufptr, nLeft, and
5225 * nFree are not reset, since these fields are manipulated by
5226 * unprotected macros, and may only be reset by non-interrupting code.
5229 /* this code requires that call->conn be set properly as a pre-condition. */
5230 #endif /* ADAPT_WINDOW */
5233 rxi_ResetCall(struct rx_call *call, int newcall)
5236 struct rx_peer *peer;
5237 struct rx_packet *packet;
5239 osirx_AssertMine(&call->lock, "rxi_ResetCall");
5241 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5243 /* Notify anyone who is waiting for asynchronous packet arrival */
5244 if (call->arrivalProc) {
5245 (*call->arrivalProc) (call, call->arrivalProcHandle,
5246 call->arrivalProcArg);
5247 call->arrivalProc = (void (*)())0;
5250 if (call->growMTUEvent)
5251 rxevent_Cancel(call->growMTUEvent, call,
5252 RX_CALL_REFCOUNT_ALIVE);
5254 if (call->delayedAbortEvent) {
5255 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
5256 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5258 rxi_SendCallAbort(call, packet, 0, 1);
5259 rxi_FreePacket(packet);
5264 * Update the peer with the congestion information in this call
5265 * so other calls on this connection can pick up where this call
5266 * left off. If the congestion sequence numbers don't match then
5267 * another call experienced a retransmission.
5269 peer = call->conn->peer;
5270 MUTEX_ENTER(&peer->peer_lock);
5272 if (call->congestSeq == peer->congestSeq) {
5273 peer->cwind = MAX(peer->cwind, call->cwind);
5274 peer->MTU = MAX(peer->MTU, call->MTU);
5275 peer->nDgramPackets =
5276 MAX(peer->nDgramPackets, call->nDgramPackets);
5279 call->abortCode = 0;
5280 call->abortCount = 0;
5282 if (peer->maxDgramPackets > 1) {
5283 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5285 call->MTU = peer->MTU;
5287 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5288 call->ssthresh = rx_maxSendWindow;
5289 call->nDgramPackets = peer->nDgramPackets;
5290 call->congestSeq = peer->congestSeq;
5291 call->rtt = peer->rtt;
5292 call->rtt_dev = peer->rtt_dev;
5293 clock_Zero(&call->rto);
5294 clock_Addmsec(&call->rto,
5295 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5296 MUTEX_EXIT(&peer->peer_lock);
5298 flags = call->flags;
5299 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5300 rxi_WaitforTQBusy(call);
5301 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5303 rxi_ClearTransmitQueue(call, 1);
5304 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5305 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5309 if ((flags & RX_CALL_PEER_BUSY)) {
5310 /* The call channel is still busy; resetting the call doesn't change
5312 call->flags |= RX_CALL_PEER_BUSY;
5315 rxi_ClearReceiveQueue(call);
5316 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5320 call->twind = call->conn->twind[call->channel];
5321 call->rwind = call->conn->rwind[call->channel];
5322 call->nSoftAcked = 0;
5323 call->nextCwind = 0;
5326 call->nCwindAcks = 0;
5327 call->nSoftAcks = 0;
5328 call->nHardAcks = 0;
5330 call->tfirst = call->rnext = call->tnext = 1;
5333 call->lastAcked = 0;
5334 call->localStatus = call->remoteStatus = 0;
5336 if (flags & RX_CALL_READER_WAIT) {
5337 #ifdef RX_ENABLE_LOCKS
5338 CV_BROADCAST(&call->cv_rq);
5340 osi_rxWakeup(&call->rq);
5343 if (flags & RX_CALL_WAIT_PACKETS) {
5344 MUTEX_ENTER(&rx_freePktQ_lock);
5345 rxi_PacketsUnWait(); /* XXX */
5346 MUTEX_EXIT(&rx_freePktQ_lock);
5348 #ifdef RX_ENABLE_LOCKS
5349 CV_SIGNAL(&call->cv_twind);
5351 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5352 osi_rxWakeup(&call->twind);
5355 #ifdef RX_ENABLE_LOCKS
5356 /* The following ensures that we don't mess with any queue while some
5357 * other thread might also be doing so. The call_queue_lock field is
5358 * is only modified under the call lock. If the call is in the process
5359 * of being removed from a queue, the call is not locked until the
5360 * the queue lock is dropped and only then is the call_queue_lock field
5361 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5362 * Note that any other routine which removes a call from a queue has to
5363 * obtain the queue lock before examing the queue and removing the call.
5365 if (call->call_queue_lock) {
5366 MUTEX_ENTER(call->call_queue_lock);
5367 if (queue_IsOnQueue(call)) {
5369 if (flags & RX_CALL_WAIT_PROC) {
5370 rx_atomic_dec(&rx_nWaiting);
5373 MUTEX_EXIT(call->call_queue_lock);
5374 CLEAR_CALL_QUEUE_LOCK(call);
5376 #else /* RX_ENABLE_LOCKS */
5377 if (queue_IsOnQueue(call)) {
5379 if (flags & RX_CALL_WAIT_PROC)
5380 rx_atomic_dec(&rx_nWaiting);
5382 #endif /* RX_ENABLE_LOCKS */
5384 rxi_KeepAliveOff(call);
5385 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5388 /* Send an acknowledge for the indicated packet (seq,serial) of the
5389 * indicated call, for the indicated reason (reason). This
5390 * acknowledge will specifically acknowledge receiving the packet, and
5391 * will also specify which other packets for this call have been
5392 * received. This routine returns the packet that was used to the
5393 * caller. The caller is responsible for freeing it or re-using it.
5394 * This acknowledgement also returns the highest sequence number
5395 * actually read out by the higher level to the sender; the sender
5396 * promises to keep around packets that have not been read by the
5397 * higher level yet (unless, of course, the sender decides to abort
5398 * the call altogether). Any of p, seq, serial, pflags, or reason may
5399 * be set to zero without ill effect. That is, if they are zero, they
5400 * will not convey any information.
5401 * NOW there is a trailer field, after the ack where it will safely be
5402 * ignored by mundanes, which indicates the maximum size packet this
5403 * host can swallow. */
5405 struct rx_packet *optionalPacket; use to send ack (or null)
5406 int seq; Sequence number of the packet we are acking
5407 int serial; Serial number of the packet
5408 int pflags; Flags field from packet header
5409 int reason; Reason an acknowledge was prompted
5413 rxi_SendAck(struct rx_call *call,
5414 struct rx_packet *optionalPacket, int serial, int reason,
5417 struct rx_ackPacket *ap;
5418 struct rx_packet *rqp;
5419 struct rx_packet *nxp; /* For queue_Scan */
5420 struct rx_packet *p;
5423 afs_uint32 padbytes = 0;
5424 #ifdef RX_ENABLE_TSFPQ
5425 struct rx_ts_info_t * rx_ts_info;
5429 * Open the receive window once a thread starts reading packets
5431 if (call->rnext > 1) {
5432 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5435 /* Don't attempt to grow MTU if this is a critical ping */
5436 if (reason == RX_ACK_MTU) {
5437 /* keep track of per-call attempts, if we're over max, do in small
5438 * otherwise in larger? set a size to increment by, decrease
5441 if (call->conn->peer->maxPacketSize &&
5442 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5444 padbytes = call->conn->peer->maxPacketSize+16;
5446 padbytes = call->conn->peer->maxMTU + 128;
5448 /* do always try a minimum size ping */
5449 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5451 /* subtract the ack payload */
5452 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5453 reason = RX_ACK_PING;
5456 call->nHardAcks = 0;
5457 call->nSoftAcks = 0;
5458 if (call->rnext > call->lastAcked)
5459 call->lastAcked = call->rnext;
5463 rx_computelen(p, p->length); /* reset length, you never know */
5464 } /* where that's been... */
5465 #ifdef RX_ENABLE_TSFPQ
5467 RX_TS_INFO_GET(rx_ts_info);
5468 if ((p = rx_ts_info->local_special_packet)) {
5469 rx_computelen(p, p->length);
5470 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5471 rx_ts_info->local_special_packet = p;
5472 } else { /* We won't send the ack, but don't panic. */
5473 return optionalPacket;
5477 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5478 /* We won't send the ack, but don't panic. */
5479 return optionalPacket;
5484 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5487 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5488 #ifndef RX_ENABLE_TSFPQ
5489 if (!optionalPacket)
5492 return optionalPacket;
5494 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5495 if (rx_Contiguous(p) < templ) {
5496 #ifndef RX_ENABLE_TSFPQ
5497 if (!optionalPacket)
5500 return optionalPacket;
5505 /* MTUXXX failing to send an ack is very serious. We should */
5506 /* try as hard as possible to send even a partial ack; it's */
5507 /* better than nothing. */
5508 ap = (struct rx_ackPacket *)rx_DataOf(p);
5509 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5510 ap->reason = reason;
5512 /* The skew computation used to be bogus, I think it's better now. */
5513 /* We should start paying attention to skew. XXX */
5514 ap->serial = htonl(serial);
5515 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5518 * First packet not yet forwarded to reader. When ACKALL has been
5519 * sent the peer has been told that all received packets will be
5520 * delivered to the reader. The value 'rnext' is used internally
5521 * to refer to the next packet in the receive queue that must be
5522 * delivered to the reader. From the perspective of the peer it
5523 * already has so report the last sequence number plus one if there
5524 * are packets in the receive queue awaiting processing.
5526 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5527 !queue_IsEmpty(&call->rq)) {
5528 ap->firstPacket = htonl(queue_Last(&call->rq, rx_packet)->header.seq + 1);
5530 ap->firstPacket = htonl(call->rnext);
5532 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5534 /* No fear of running out of ack packet here because there can only be at most
5535 * one window full of unacknowledged packets. The window size must be constrained
5536 * to be less than the maximum ack size, of course. Also, an ack should always
5537 * fit into a single packet -- it should not ever be fragmented. */
5538 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5539 if (!rqp || !call->rq.next
5540 || (rqp->header.seq > (call->rnext + call->rwind))) {
5541 #ifndef RX_ENABLE_TSFPQ
5542 if (!optionalPacket)
5545 rxi_CallError(call, RX_CALL_DEAD);
5546 return optionalPacket;
5549 while (rqp->header.seq > call->rnext + offset)
5550 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5551 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5553 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5554 #ifndef RX_ENABLE_TSFPQ
5555 if (!optionalPacket)
5558 rxi_CallError(call, RX_CALL_DEAD);
5559 return optionalPacket;
5564 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5566 /* these are new for AFS 3.3 */
5567 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5568 templ = htonl(templ);
5569 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5570 templ = htonl(call->conn->peer->ifMTU);
5571 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5572 sizeof(afs_int32), &templ);
5574 /* new for AFS 3.4 */
5575 templ = htonl(call->rwind);
5576 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5577 sizeof(afs_int32), &templ);
5579 /* new for AFS 3.5 */
5580 templ = htonl(call->conn->peer->ifDgramPackets);
5581 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5582 sizeof(afs_int32), &templ);
5584 p->header.serviceId = call->conn->serviceId;
5585 p->header.cid = (call->conn->cid | call->channel);
5586 p->header.callNumber = *call->callNumber;
5588 p->header.securityIndex = call->conn->securityIndex;
5589 p->header.epoch = call->conn->epoch;
5590 p->header.type = RX_PACKET_TYPE_ACK;
5591 p->header.flags = RX_SLOW_START_OK;
5592 if (reason == RX_ACK_PING) {
5593 p->header.flags |= RX_REQUEST_ACK;
5595 clock_GetTime(&call->pingRequestTime);
5598 p->length = padbytes +
5599 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5602 /* not fast but we can potentially use this if truncated
5603 * fragments are delivered to figure out the mtu.
5605 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5606 sizeof(afs_int32), sizeof(afs_int32),
5610 if (call->conn->type == RX_CLIENT_CONNECTION)
5611 p->header.flags |= RX_CLIENT_INITIATED;
5615 if (rxdebug_active) {
5619 len = _snprintf(msg, sizeof(msg),
5620 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5621 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5622 ntohl(ap->serial), ntohl(ap->previousPacket),
5623 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5624 ap->nAcks, ntohs(ap->bufferSpace) );
5628 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5629 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5633 OutputDebugString(msg);
5635 #else /* AFS_NT40_ENV */
5637 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5638 ap->reason, ntohl(ap->previousPacket),
5639 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5641 for (offset = 0; offset < ap->nAcks; offset++)
5642 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5647 #endif /* AFS_NT40_ENV */
5650 int i, nbytes = p->length;
5652 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5653 if (nbytes <= p->wirevec[i].iov_len) {
5656 savelen = p->wirevec[i].iov_len;
5658 p->wirevec[i].iov_len = nbytes;
5660 rxi_Send(call, p, istack);
5661 p->wirevec[i].iov_len = savelen;
5665 nbytes -= p->wirevec[i].iov_len;
5668 if (rx_stats_active)
5669 rx_atomic_inc(&rx_stats.ackPacketsSent);
5670 #ifndef RX_ENABLE_TSFPQ
5671 if (!optionalPacket)
5674 return optionalPacket; /* Return packet for re-use by caller */
5678 struct rx_packet **list;
5683 /* Send all of the packets in the list in single datagram */
5685 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5686 int istack, int moreFlag)
5692 struct rx_connection *conn = call->conn;
5693 struct rx_peer *peer = conn->peer;
5695 MUTEX_ENTER(&peer->peer_lock);
5696 peer->nSent += xmit->len;
5697 if (xmit->resending)
5698 peer->reSends += xmit->len;
5699 MUTEX_EXIT(&peer->peer_lock);
5701 if (rx_stats_active) {
5702 if (xmit->resending)
5703 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5705 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5708 clock_GetTime(&now);
5710 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5714 /* Set the packet flags and schedule the resend events */
5715 /* Only request an ack for the last packet in the list */
5716 for (i = 0; i < xmit->len; i++) {
5717 struct rx_packet *packet = xmit->list[i];
5719 /* Record the time sent */
5720 packet->timeSent = now;
5721 packet->flags |= RX_PKTFLAG_SENT;
5723 /* Ask for an ack on retransmitted packets, on every other packet
5724 * if the peer doesn't support slow start. Ask for an ack on every
5725 * packet until the congestion window reaches the ack rate. */
5726 if (packet->header.serial) {
5729 packet->firstSent = now;
5730 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5731 || (!(call->flags & RX_CALL_SLOW_START_OK)
5732 && (packet->header.seq & 1)))) {
5737 /* Tag this packet as not being the last in this group,
5738 * for the receiver's benefit */
5739 if (i < xmit->len - 1 || moreFlag) {
5740 packet->header.flags |= RX_MORE_PACKETS;
5745 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5748 /* Since we're about to send a data packet to the peer, it's
5749 * safe to nuke any scheduled end-of-packets ack */
5750 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5752 MUTEX_EXIT(&call->lock);
5753 MUTEX_ENTER(&rx_refcnt_mutex);
5754 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5755 MUTEX_EXIT(&rx_refcnt_mutex);
5756 if (xmit->len > 1) {
5757 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5759 rxi_SendPacket(call, conn, xmit->list[0], istack);
5761 MUTEX_ENTER(&call->lock);
5762 MUTEX_ENTER(&rx_refcnt_mutex);
5763 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5764 MUTEX_EXIT(&rx_refcnt_mutex);
5766 /* Tell the RTO calculation engine that we have sent a packet, and
5767 * if it was the last one */
5768 rxi_rto_packet_sent(call, lastPacket, istack);
5770 /* Update last send time for this call (for keep-alive
5771 * processing), and for the connection (so that we can discover
5772 * idle connections) */
5773 conn->lastSendTime = call->lastSendTime = clock_Sec();
5774 /* Let a set of retransmits trigger an idle timeout */
5775 if (!xmit->resending)
5776 call->lastSendData = call->lastSendTime;
5779 /* When sending packets we need to follow these rules:
5780 * 1. Never send more than maxDgramPackets in a jumbogram.
5781 * 2. Never send a packet with more than two iovecs in a jumbogram.
5782 * 3. Never send a retransmitted packet in a jumbogram.
5783 * 4. Never send more than cwind/4 packets in a jumbogram
5784 * We always keep the last list we should have sent so we
5785 * can set the RX_MORE_PACKETS flags correctly.
5789 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5794 struct xmitlist working;
5795 struct xmitlist last;
5797 struct rx_peer *peer = call->conn->peer;
5798 int morePackets = 0;
5800 memset(&last, 0, sizeof(struct xmitlist));
5801 working.list = &list[0];
5803 working.resending = 0;
5805 recovery = call->flags & RX_CALL_FAST_RECOVER;
5807 for (i = 0; i < len; i++) {
5808 /* Does the current packet force us to flush the current list? */
5810 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5811 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5813 /* This sends the 'last' list and then rolls the current working
5814 * set into the 'last' one, and resets the working set */
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];
5829 /* Add the current packet to the list if it hasn't been acked.
5830 * Otherwise adjust the list pointer to skip the current packet. */
5831 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5834 if (list[i]->header.serial)
5835 working.resending = 1;
5837 /* Do we need to flush the list? */
5838 if (working.len >= (int)peer->maxDgramPackets
5839 || working.len >= (int)call->nDgramPackets
5840 || working.len >= (int)call->cwind
5841 || list[i]->header.serial
5842 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5844 rxi_SendList(call, &last, istack, 1);
5845 /* If the call enters an error state stop sending, or if
5846 * we entered congestion recovery mode, stop sending */
5848 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5853 working.resending = 0;
5854 working.list = &list[i + 1];
5857 if (working.len != 0) {
5858 osi_Panic("rxi_SendList error");
5860 working.list = &list[i + 1];
5864 /* Send the whole list when the call is in receive mode, when
5865 * the call is in eof mode, when we are in fast recovery mode,
5866 * and when we have the last packet */
5867 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5868 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5869 || (call->flags & RX_CALL_FAST_RECOVER)) {
5870 /* Check for the case where the current list contains
5871 * an acked packet. Since we always send retransmissions
5872 * in a separate packet, we only need to check the first
5873 * packet in the list */
5874 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5878 rxi_SendList(call, &last, istack, morePackets);
5879 /* If the call enters an error state stop sending, or if
5880 * we entered congestion recovery mode, stop sending */
5882 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5886 rxi_SendList(call, &working, istack, 0);
5888 } else if (last.len > 0) {
5889 rxi_SendList(call, &last, istack, 0);
5890 /* Packets which are in 'working' are not sent by this call */
5895 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
5897 struct rx_call *call = arg0;
5898 struct rx_peer *peer;
5899 struct rx_packet *p, *nxp;
5900 struct clock maxTimeout = { 60, 0 };
5902 MUTEX_ENTER(&call->lock);
5904 peer = call->conn->peer;
5906 /* Make sure that the event pointer is removed from the call
5907 * structure, since there is no longer a per-call retransmission
5909 if (event == call->resendEvent) {
5910 MUTEX_ENTER(&rx_refcnt_mutex);
5911 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5912 MUTEX_EXIT(&rx_refcnt_mutex);
5913 call->resendEvent = NULL;
5916 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
5917 rxi_CheckBusy(call);
5920 if (queue_IsEmpty(&call->tq)) {
5921 /* Nothing to do. This means that we've been raced, and that an
5922 * ACK has come in between when we were triggered, and when we
5923 * actually got to run. */
5927 /* We're in loss recovery */
5928 call->flags |= RX_CALL_FAST_RECOVER;
5930 /* Mark all of the pending packets in the queue as being lost */
5931 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5932 if (!(p->flags & RX_PKTFLAG_ACKED))
5933 p->flags &= ~RX_PKTFLAG_SENT;
5936 /* We're resending, so we double the timeout of the call. This will be
5937 * dropped back down by the first successful ACK that we receive.
5939 * We apply a maximum value here of 60 seconds
5941 clock_Add(&call->rto, &call->rto);
5942 if (clock_Gt(&call->rto, &maxTimeout))
5943 call->rto = maxTimeout;
5945 /* Packet loss is most likely due to congestion, so drop our window size
5946 * and start again from the beginning */
5947 if (peer->maxDgramPackets >1) {
5948 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5949 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5951 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5952 call->nDgramPackets = 1;
5954 call->nextCwind = 1;
5957 MUTEX_ENTER(&peer->peer_lock);
5958 peer->MTU = call->MTU;
5959 peer->cwind = call->cwind;
5960 peer->nDgramPackets = 1;
5962 call->congestSeq = peer->congestSeq;
5963 MUTEX_EXIT(&peer->peer_lock);
5965 rxi_Start(call, istack);
5968 MUTEX_EXIT(&call->lock);
5971 /* This routine is called when new packets are readied for
5972 * transmission and when retransmission may be necessary, or when the
5973 * transmission window or burst count are favourable. This should be
5974 * better optimized for new packets, the usual case, now that we've
5975 * got rid of queues of send packets. XXXXXXXXXXX */
5977 rxi_Start(struct rx_call *call, int istack)
5980 struct rx_packet *p;
5981 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5986 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5987 if (rx_stats_active)
5988 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5993 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5995 /* Send (or resend) any packets that need it, subject to
5996 * window restrictions and congestion burst control
5997 * restrictions. Ask for an ack on the last packet sent in
5998 * this burst. For now, we're relying upon the window being
5999 * considerably bigger than the largest number of packets that
6000 * are typically sent at once by one initial call to
6001 * rxi_Start. This is probably bogus (perhaps we should ask
6002 * for an ack when we're half way through the current
6003 * window?). Also, for non file transfer applications, this
6004 * may end up asking for an ack for every packet. Bogus. XXXX
6007 * But check whether we're here recursively, and let the other guy
6010 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6011 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6012 call->flags |= RX_CALL_TQ_BUSY;
6014 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6016 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6017 call->flags &= ~RX_CALL_NEED_START;
6018 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6020 maxXmitPackets = MIN(call->twind, call->cwind);
6021 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
6022 #ifdef RX_TRACK_PACKETS
6023 if ((p->flags & RX_PKTFLAG_FREE)
6024 || (!queue_IsEnd(&call->tq, nxp)
6025 && (nxp->flags & RX_PKTFLAG_FREE))
6026 || (p == (struct rx_packet *)&rx_freePacketQueue)
6027 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
6028 osi_Panic("rxi_Start: xmit queue clobbered");
6031 if (p->flags & RX_PKTFLAG_ACKED) {
6032 /* Since we may block, don't trust this */
6033 if (rx_stats_active)
6034 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6035 continue; /* Ignore this packet if it has been acknowledged */
6038 /* Turn off all flags except these ones, which are the same
6039 * on each transmission */
6040 p->header.flags &= RX_PRESET_FLAGS;
6042 if (p->header.seq >=
6043 call->tfirst + MIN((int)call->twind,
6044 (int)(call->nSoftAcked +
6046 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6047 /* Note: if we're waiting for more window space, we can
6048 * still send retransmits; hence we don't return here, but
6049 * break out to schedule a retransmit event */
6050 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6051 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6056 /* Transmit the packet if it needs to be sent. */
6057 if (!(p->flags & RX_PKTFLAG_SENT)) {
6058 if (nXmitPackets == maxXmitPackets) {
6059 rxi_SendXmitList(call, call->xmitList,
6060 nXmitPackets, istack);
6063 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6064 *(call->callNumber), p));
6065 call->xmitList[nXmitPackets++] = p;
6069 /* xmitList now hold pointers to all of the packets that are
6070 * ready to send. Now we loop to send the packets */
6071 if (nXmitPackets > 0) {
6072 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6076 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6078 /* We went into the error state while sending packets. Now is
6079 * the time to reset the call. This will also inform the using
6080 * process that the call is in an error state.
6082 if (rx_stats_active)
6083 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6084 call->flags &= ~RX_CALL_TQ_BUSY;
6085 rxi_WakeUpTransmitQueue(call);
6086 rxi_CallError(call, call->error);
6089 #ifdef RX_ENABLE_LOCKS
6090 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6092 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6093 /* Some packets have received acks. If they all have, we can clear
6094 * the transmit queue.
6097 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
6098 if (p->header.seq < call->tfirst
6099 && (p->flags & RX_PKTFLAG_ACKED)) {
6101 #ifdef RX_TRACK_PACKETS
6102 p->flags &= ~RX_PKTFLAG_TQ;
6104 #ifdef RXDEBUG_PACKET
6112 call->flags |= RX_CALL_TQ_CLEARME;
6114 #endif /* RX_ENABLE_LOCKS */
6115 if (call->flags & RX_CALL_TQ_CLEARME)
6116 rxi_ClearTransmitQueue(call, 1);
6117 } while (call->flags & RX_CALL_NEED_START);
6119 * TQ references no longer protected by this flag; they must remain
6120 * protected by the global lock.
6122 call->flags &= ~RX_CALL_TQ_BUSY;
6123 rxi_WakeUpTransmitQueue(call);
6125 call->flags |= RX_CALL_NEED_START;
6127 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6129 rxi_rto_cancel(call);
6133 /* Also adjusts the keep alive parameters for the call, to reflect
6134 * that we have just sent a packet (so keep alives aren't sent
6137 rxi_Send(struct rx_call *call, struct rx_packet *p,
6140 struct rx_connection *conn = call->conn;
6142 /* Stamp each packet with the user supplied status */
6143 p->header.userStatus = call->localStatus;
6145 /* Allow the security object controlling this call's security to
6146 * make any last-minute changes to the packet */
6147 RXS_SendPacket(conn->securityObject, call, p);
6149 /* Since we're about to send SOME sort of packet to the peer, it's
6150 * safe to nuke any scheduled end-of-packets ack */
6151 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
6153 /* Actually send the packet, filling in more connection-specific fields */
6154 MUTEX_EXIT(&call->lock);
6155 MUTEX_ENTER(&rx_refcnt_mutex);
6156 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6157 MUTEX_EXIT(&rx_refcnt_mutex);
6158 rxi_SendPacket(call, conn, p, istack);
6159 MUTEX_ENTER(&rx_refcnt_mutex);
6160 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6161 MUTEX_EXIT(&rx_refcnt_mutex);
6162 MUTEX_ENTER(&call->lock);
6164 /* Update last send time for this call (for keep-alive
6165 * processing), and for the connection (so that we can discover
6166 * idle connections) */
6167 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6168 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6169 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6171 conn->lastSendTime = call->lastSendTime = clock_Sec();
6172 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6173 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6174 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6175 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6176 RX_ACK_PING_RESPONSE)))
6177 call->lastSendData = call->lastSendTime;
6181 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6182 * that things are fine. Also called periodically to guarantee that nothing
6183 * falls through the cracks (e.g. (error + dally) connections have keepalive
6184 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6186 * haveCTLock Set if calling from rxi_ReapConnections
6188 #ifdef RX_ENABLE_LOCKS
6190 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6191 #else /* RX_ENABLE_LOCKS */
6193 rxi_CheckCall(struct rx_call *call)
6194 #endif /* RX_ENABLE_LOCKS */
6196 struct rx_connection *conn = call->conn;
6198 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6199 afs_uint32 fudgeFactor;
6203 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6204 if (call->flags & RX_CALL_TQ_BUSY) {
6205 /* Call is active and will be reset by rxi_Start if it's
6206 * in an error state.
6211 /* RTT + 8*MDEV, rounded up to the next second. */
6212 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6213 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6215 deadTime = conn->secondsUntilDead + fudgeFactor;
6217 /* These are computed to the second (+- 1 second). But that's
6218 * good enough for these values, which should be a significant
6219 * number of seconds. */
6220 if (now > (call->lastReceiveTime + deadTime)) {
6221 if (call->state == RX_STATE_ACTIVE) {
6223 #if defined(KERNEL) && defined(AFS_SUN5_ENV)
6225 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6226 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6227 ip_stack_t *ipst = ns->netstack_ip;
6229 ire = ire_cache_lookup(conn->peer->host
6230 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6232 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
6234 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6241 if (ire && ire->ire_max_frag > 0)
6242 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6244 #if defined(GLOBAL_NETSTACKID)
6248 #endif /* ADAPT_PMTU */
6249 cerror = RX_CALL_DEAD;
6252 #ifdef RX_ENABLE_LOCKS
6253 /* Cancel pending events */
6254 rxevent_Cancel(call->delayedAckEvent, call,
6255 RX_CALL_REFCOUNT_DELAY);
6256 rxi_rto_cancel(call);
6257 rxevent_Cancel(call->keepAliveEvent, call,
6258 RX_CALL_REFCOUNT_ALIVE);
6259 if (call->growMTUEvent)
6260 rxevent_Cancel(call->growMTUEvent, call,
6261 RX_CALL_REFCOUNT_ALIVE);
6262 MUTEX_ENTER(&rx_refcnt_mutex);
6263 if (call->refCount == 0) {
6264 rxi_FreeCall(call, haveCTLock);
6265 MUTEX_EXIT(&rx_refcnt_mutex);
6268 MUTEX_EXIT(&rx_refcnt_mutex);
6270 #else /* RX_ENABLE_LOCKS */
6271 rxi_FreeCall(call, 0);
6273 #endif /* RX_ENABLE_LOCKS */
6275 /* Non-active calls are destroyed if they are not responding
6276 * to pings; active calls are simply flagged in error, so the
6277 * attached process can die reasonably gracefully. */
6280 if (conn->idleDeadTime) {
6281 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6284 /* see if we have a non-activity timeout */
6285 if (call->startWait && idleDeadTime
6286 && ((call->startWait + idleDeadTime) < now) &&
6287 (call->flags & RX_CALL_READER_WAIT)) {
6288 if (call->state == RX_STATE_ACTIVE) {
6289 cerror = RX_CALL_TIMEOUT;
6293 if (call->lastSendData && idleDeadTime && (conn->idleDeadErr != 0)
6294 && ((call->lastSendData + idleDeadTime) < now)) {
6295 if (call->state == RX_STATE_ACTIVE) {
6296 cerror = conn->idleDeadErr;
6301 if (conn->hardDeadTime) {
6302 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6305 /* see if we have a hard timeout */
6307 && (now > (hardDeadTime + call->startTime.sec))) {
6308 if (call->state == RX_STATE_ACTIVE)
6309 rxi_CallError(call, RX_CALL_TIMEOUT);
6314 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT
6315 && call->lastReceiveTime) {
6316 int oldMTU = conn->peer->ifMTU;
6318 /* if we thought we could send more, perhaps things got worse */
6319 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6320 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6321 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6322 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6324 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6326 /* minimum capped in SetPeerMtu */
6327 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6330 conn->lastPacketSize = 0;
6332 /* needed so ResetCall doesn't clobber us. */
6333 call->MTU = conn->peer->ifMTU;
6335 /* if we never succeeded, let the error pass out as-is */
6336 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6337 cerror = conn->msgsizeRetryErr;
6340 rxi_CallError(call, cerror);
6345 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6347 struct rx_connection *conn = arg1;
6348 struct rx_header theader;
6349 char tbuffer[1 + sizeof(struct rx_header)];
6350 struct sockaddr_in taddr;
6353 struct iovec tmpiov[2];
6356 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6359 tp = &tbuffer[sizeof(struct rx_header)];
6360 taddr.sin_family = AF_INET;
6361 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6362 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6363 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6364 taddr.sin_len = sizeof(struct sockaddr_in);
6366 memset(&theader, 0, sizeof(theader));
6367 theader.epoch = htonl(999);
6369 theader.callNumber = 0;
6372 theader.type = RX_PACKET_TYPE_VERSION;
6373 theader.flags = RX_LAST_PACKET;
6374 theader.serviceId = 0;
6376 memcpy(tbuffer, &theader, sizeof(theader));
6377 memcpy(tp, &a, sizeof(a));
6378 tmpiov[0].iov_base = tbuffer;
6379 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6381 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6383 MUTEX_ENTER(&conn->conn_data_lock);
6384 MUTEX_ENTER(&rx_refcnt_mutex);
6385 /* Only reschedule ourselves if the connection would not be destroyed */
6386 if (conn->refCount <= 1) {
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 conn->natKeepAliveEvent = NULL;
6395 rxi_ScheduleNatKeepAliveEvent(conn);
6396 MUTEX_EXIT(&conn->conn_data_lock);
6401 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6403 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6404 struct clock when, now;
6405 clock_GetTime(&now);
6407 when.sec += conn->secondsUntilNatPing;
6408 MUTEX_ENTER(&rx_refcnt_mutex);
6409 conn->refCount++; /* hold a reference for this */
6410 MUTEX_EXIT(&rx_refcnt_mutex);
6411 conn->natKeepAliveEvent =
6412 rxevent_PostNow(&when, &now, rxi_NatKeepAliveEvent, conn, 0);
6417 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6419 MUTEX_ENTER(&conn->conn_data_lock);
6420 conn->secondsUntilNatPing = seconds;
6422 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6423 rxi_ScheduleNatKeepAliveEvent(conn);
6425 conn->flags |= RX_CONN_NAT_PING;
6427 MUTEX_EXIT(&conn->conn_data_lock);
6431 rxi_NatKeepAliveOn(struct rx_connection *conn)
6433 MUTEX_ENTER(&conn->conn_data_lock);
6434 /* if it's already attached */
6435 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6436 rxi_ScheduleNatKeepAliveEvent(conn);
6438 conn->flags |= RX_CONN_NAT_PING;
6439 MUTEX_EXIT(&conn->conn_data_lock);
6442 /* When a call is in progress, this routine is called occasionally to
6443 * make sure that some traffic has arrived (or been sent to) the peer.
6444 * If nothing has arrived in a reasonable amount of time, the call is
6445 * declared dead; if nothing has been sent for a while, we send a
6446 * keep-alive packet (if we're actually trying to keep the call alive)
6449 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6451 struct rx_call *call = arg1;
6452 struct rx_connection *conn;
6455 MUTEX_ENTER(&rx_refcnt_mutex);
6456 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6457 MUTEX_EXIT(&rx_refcnt_mutex);
6458 MUTEX_ENTER(&call->lock);
6459 if (event == call->keepAliveEvent)
6460 call->keepAliveEvent = NULL;
6463 #ifdef RX_ENABLE_LOCKS
6464 if (rxi_CheckCall(call, 0)) {
6465 MUTEX_EXIT(&call->lock);
6468 #else /* RX_ENABLE_LOCKS */
6469 if (rxi_CheckCall(call))
6471 #endif /* RX_ENABLE_LOCKS */
6473 /* Don't try to keep alive dallying calls */
6474 if (call->state == RX_STATE_DALLY) {
6475 MUTEX_EXIT(&call->lock);
6480 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6481 /* Don't try to send keepalives if there is unacknowledged data */
6482 /* the rexmit code should be good enough, this little hack
6483 * doesn't quite work XXX */
6484 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6486 rxi_ScheduleKeepAliveEvent(call);
6487 MUTEX_EXIT(&call->lock);
6490 /* Does what's on the nameplate. */
6492 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy)
6494 struct rx_call *call = arg1;
6495 struct rx_connection *conn;
6497 MUTEX_ENTER(&rx_refcnt_mutex);
6498 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6499 MUTEX_EXIT(&rx_refcnt_mutex);
6500 MUTEX_ENTER(&call->lock);
6502 if (event == call->growMTUEvent)
6503 call->growMTUEvent = NULL;
6505 #ifdef RX_ENABLE_LOCKS
6506 if (rxi_CheckCall(call, 0)) {
6507 MUTEX_EXIT(&call->lock);
6510 #else /* RX_ENABLE_LOCKS */
6511 if (rxi_CheckCall(call))
6513 #endif /* RX_ENABLE_LOCKS */
6515 /* Don't bother with dallying calls */
6516 if (call->state == RX_STATE_DALLY) {
6517 MUTEX_EXIT(&call->lock);
6524 * keep being scheduled, just don't do anything if we're at peak,
6525 * or we're not set up to be properly handled (idle timeout required)
6527 if ((conn->peer->maxPacketSize != 0) &&
6528 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6529 (conn->idleDeadErr))
6530 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6531 rxi_ScheduleGrowMTUEvent(call, 0);
6532 MUTEX_EXIT(&call->lock);
6536 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6538 if (!call->keepAliveEvent) {
6539 struct clock when, now;
6540 clock_GetTime(&now);
6542 when.sec += call->conn->secondsUntilPing;
6543 MUTEX_ENTER(&rx_refcnt_mutex);
6544 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6545 MUTEX_EXIT(&rx_refcnt_mutex);
6546 call->keepAliveEvent =
6547 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
6552 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6554 if (!call->growMTUEvent) {
6555 struct clock when, now;
6557 clock_GetTime(&now);
6560 if (call->conn->secondsUntilPing)
6561 secs = (6*call->conn->secondsUntilPing)-1;
6563 if (call->conn->secondsUntilDead)
6564 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6568 MUTEX_ENTER(&rx_refcnt_mutex);
6569 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6570 MUTEX_EXIT(&rx_refcnt_mutex);
6571 call->growMTUEvent =
6572 rxevent_PostNow(&when, &now, rxi_GrowMTUEvent, call, 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 rxi_GrowMTUOn(struct rx_call *call)
6592 struct rx_connection *conn = call->conn;
6593 MUTEX_ENTER(&conn->conn_data_lock);
6594 conn->lastPingSizeSer = conn->lastPingSize = 0;
6595 MUTEX_EXIT(&conn->conn_data_lock);
6596 rxi_ScheduleGrowMTUEvent(call, 1);
6599 /* This routine is called to send connection abort messages
6600 * that have been delayed to throttle looping clients. */
6602 rxi_SendDelayedConnAbort(struct rxevent *event,
6603 void *arg1, void *unused)
6605 struct rx_connection *conn = arg1;
6608 struct rx_packet *packet;
6610 MUTEX_ENTER(&conn->conn_data_lock);
6611 conn->delayedAbortEvent = NULL;
6612 error = htonl(conn->error);
6614 MUTEX_EXIT(&conn->conn_data_lock);
6615 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6618 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6619 RX_PACKET_TYPE_ABORT, (char *)&error,
6621 rxi_FreePacket(packet);
6625 /* This routine is called to send call abort messages
6626 * that have been delayed to throttle looping clients. */
6628 rxi_SendDelayedCallAbort(struct rxevent *event,
6629 void *arg1, void *dummy)
6631 struct rx_call *call = arg1;
6634 struct rx_packet *packet;
6636 MUTEX_ENTER(&call->lock);
6637 call->delayedAbortEvent = NULL;
6638 error = htonl(call->error);
6640 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6643 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6644 (char *)&error, sizeof(error), 0);
6645 rxi_FreePacket(packet);
6647 MUTEX_EXIT(&call->lock);
6648 MUTEX_ENTER(&rx_refcnt_mutex);
6649 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6650 MUTEX_EXIT(&rx_refcnt_mutex);
6653 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6654 * seconds) to ask the client to authenticate itself. The routine
6655 * issues a challenge to the client, which is obtained from the
6656 * security object associated with the connection */
6658 rxi_ChallengeEvent(struct rxevent *event,
6659 void *arg0, void *arg1, int tries)
6661 struct rx_connection *conn = arg0;
6663 conn->challengeEvent = NULL;
6664 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6665 struct rx_packet *packet;
6666 struct clock when, now;
6669 /* We've failed to authenticate for too long.
6670 * Reset any calls waiting for authentication;
6671 * they are all in RX_STATE_PRECALL.
6675 MUTEX_ENTER(&conn->conn_call_lock);
6676 for (i = 0; i < RX_MAXCALLS; i++) {
6677 struct rx_call *call = conn->call[i];
6679 MUTEX_ENTER(&call->lock);
6680 if (call->state == RX_STATE_PRECALL) {
6681 rxi_CallError(call, RX_CALL_DEAD);
6682 rxi_SendCallAbort(call, NULL, 0, 0);
6684 MUTEX_EXIT(&call->lock);
6687 MUTEX_EXIT(&conn->conn_call_lock);
6691 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6693 /* If there's no packet available, do this later. */
6694 RXS_GetChallenge(conn->securityObject, conn, packet);
6695 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6696 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6697 rxi_FreePacket(packet);
6699 clock_GetTime(&now);
6701 when.sec += RX_CHALLENGE_TIMEOUT;
6702 conn->challengeEvent =
6703 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
6708 /* Call this routine to start requesting the client to authenticate
6709 * itself. This will continue until authentication is established,
6710 * the call times out, or an invalid response is returned. The
6711 * security object associated with the connection is asked to create
6712 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6713 * defined earlier. */
6715 rxi_ChallengeOn(struct rx_connection *conn)
6717 if (!conn->challengeEvent) {
6718 RXS_CreateChallenge(conn->securityObject, conn);
6719 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6724 /* rxi_ComputeRoundTripTime is called with peer locked. */
6725 /* peer may be null */
6727 rxi_ComputeRoundTripTime(struct rx_packet *p,
6728 struct rx_ackPacket *ack,
6729 struct rx_call *call,
6730 struct rx_peer *peer,
6733 struct clock thisRtt, *sentp;
6737 /* If the ACK is delayed, then do nothing */
6738 if (ack->reason == RX_ACK_DELAY)
6741 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6742 * their RTT multiple times, so only include the RTT of the last packet
6744 if (p->flags & RX_JUMBO_PACKET)
6747 /* Use the serial number to determine which transmission the ACK is for,
6748 * and set the sent time to match this. If we have no serial number, then
6749 * only use the ACK for RTT calculations if the packet has not been
6753 serial = ntohl(ack->serial);
6755 if (serial == p->header.serial) {
6756 sentp = &p->timeSent;
6757 } else if (serial == p->firstSerial) {
6758 sentp = &p->firstSent;
6759 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6760 sentp = &p->firstSent;
6764 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6765 sentp = &p->firstSent;
6772 if (clock_Lt(&thisRtt, sentp))
6773 return; /* somebody set the clock back, don't count this time. */
6775 clock_Sub(&thisRtt, sentp);
6776 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6777 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6779 if (clock_IsZero(&thisRtt)) {
6781 * The actual round trip time is shorter than the
6782 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6783 * Since we can't tell which at the moment we will assume 1ms.
6785 thisRtt.usec = 1000;
6788 if (rx_stats_active) {
6789 MUTEX_ENTER(&rx_stats_mutex);
6790 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6791 rx_stats.minRtt = thisRtt;
6792 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6793 if (thisRtt.sec > 60) {
6794 MUTEX_EXIT(&rx_stats_mutex);
6795 return; /* somebody set the clock ahead */
6797 rx_stats.maxRtt = thisRtt;
6799 clock_Add(&rx_stats.totalRtt, &thisRtt);
6800 rx_atomic_inc(&rx_stats.nRttSamples);
6801 MUTEX_EXIT(&rx_stats_mutex);
6804 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6806 /* Apply VanJacobson round-trip estimations */
6811 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6812 * srtt is stored as fixed point with 3 bits after the binary
6813 * point (i.e., scaled by 8). The following magic is
6814 * equivalent to the smoothing algorithm in rfc793 with an
6815 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6816 * srtt'*8 = rtt + srtt*7
6817 * srtt'*8 = srtt*8 + rtt - srtt
6818 * srtt' = srtt + rtt/8 - srtt/8
6819 * srtt' = srtt + (rtt - srtt)/8
6822 delta = _8THMSEC(&thisRtt) - call->rtt;
6823 call->rtt += (delta >> 3);
6826 * We accumulate a smoothed rtt variance (actually, a smoothed
6827 * mean difference), then set the retransmit timer to smoothed
6828 * rtt + 4 times the smoothed variance (was 2x in van's original
6829 * paper, but 4x works better for me, and apparently for him as
6831 * rttvar is stored as
6832 * fixed point with 2 bits after the binary point (scaled by
6833 * 4). The following is equivalent to rfc793 smoothing with
6834 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6835 * rttvar'*4 = rttvar*3 + |delta|
6836 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6837 * rttvar' = rttvar + |delta|/4 - rttvar/4
6838 * rttvar' = rttvar + (|delta| - rttvar)/4
6839 * This replaces rfc793's wired-in beta.
6840 * dev*4 = dev*4 + (|actual - expected| - dev)
6846 delta -= (call->rtt_dev << 1);
6847 call->rtt_dev += (delta >> 3);
6849 /* I don't have a stored RTT so I start with this value. Since I'm
6850 * probably just starting a call, and will be pushing more data down
6851 * this, I expect congestion to increase rapidly. So I fudge a
6852 * little, and I set deviance to half the rtt. In practice,
6853 * deviance tends to approach something a little less than
6854 * half the smoothed rtt. */
6855 call->rtt = _8THMSEC(&thisRtt) + 8;
6856 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6858 /* the smoothed RTT time is RTT + 4*MDEV
6860 * We allow a user specified minimum to be set for this, to allow clamping
6861 * at a minimum value in the same way as TCP. In addition, we have to allow
6862 * for the possibility that this packet is answered by a delayed ACK, so we
6863 * add on a fixed 200ms to account for that timer expiring.
6866 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6867 rx_minPeerTimeout) + 200;
6868 clock_Zero(&call->rto);
6869 clock_Addmsec(&call->rto, rtt_timeout);
6871 /* Update the peer, so any new calls start with our values */
6872 peer->rtt_dev = call->rtt_dev;
6873 peer->rtt = call->rtt;
6875 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6876 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6880 /* Find all server connections that have not been active for a long time, and
6883 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6885 struct clock now, when;
6886 clock_GetTime(&now);
6888 /* Find server connection structures that haven't been used for
6889 * greater than rx_idleConnectionTime */
6891 struct rx_connection **conn_ptr, **conn_end;
6892 int i, havecalls = 0;
6893 MUTEX_ENTER(&rx_connHashTable_lock);
6894 for (conn_ptr = &rx_connHashTable[0], conn_end =
6895 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6897 struct rx_connection *conn, *next;
6898 struct rx_call *call;
6902 for (conn = *conn_ptr; conn; conn = next) {
6903 /* XXX -- Shouldn't the connection be locked? */
6906 for (i = 0; i < RX_MAXCALLS; i++) {
6907 call = conn->call[i];
6911 code = MUTEX_TRYENTER(&call->lock);
6914 #ifdef RX_ENABLE_LOCKS
6915 result = rxi_CheckCall(call, 1);
6916 #else /* RX_ENABLE_LOCKS */
6917 result = rxi_CheckCall(call);
6918 #endif /* RX_ENABLE_LOCKS */
6919 MUTEX_EXIT(&call->lock);
6921 /* If CheckCall freed the call, it might
6922 * have destroyed the connection as well,
6923 * which screws up the linked lists.
6929 if (conn->type == RX_SERVER_CONNECTION) {
6930 /* This only actually destroys the connection if
6931 * there are no outstanding calls */
6932 MUTEX_ENTER(&conn->conn_data_lock);
6933 MUTEX_ENTER(&rx_refcnt_mutex);
6934 if (!havecalls && !conn->refCount
6935 && ((conn->lastSendTime + rx_idleConnectionTime) <
6937 conn->refCount++; /* it will be decr in rx_DestroyConn */
6938 MUTEX_EXIT(&rx_refcnt_mutex);
6939 MUTEX_EXIT(&conn->conn_data_lock);
6940 #ifdef RX_ENABLE_LOCKS
6941 rxi_DestroyConnectionNoLock(conn);
6942 #else /* RX_ENABLE_LOCKS */
6943 rxi_DestroyConnection(conn);
6944 #endif /* RX_ENABLE_LOCKS */
6946 #ifdef RX_ENABLE_LOCKS
6948 MUTEX_EXIT(&rx_refcnt_mutex);
6949 MUTEX_EXIT(&conn->conn_data_lock);
6951 #endif /* RX_ENABLE_LOCKS */
6955 #ifdef RX_ENABLE_LOCKS
6956 while (rx_connCleanup_list) {
6957 struct rx_connection *conn;
6958 conn = rx_connCleanup_list;
6959 rx_connCleanup_list = rx_connCleanup_list->next;
6960 MUTEX_EXIT(&rx_connHashTable_lock);
6961 rxi_CleanupConnection(conn);
6962 MUTEX_ENTER(&rx_connHashTable_lock);
6964 MUTEX_EXIT(&rx_connHashTable_lock);
6965 #endif /* RX_ENABLE_LOCKS */
6968 /* Find any peer structures that haven't been used (haven't had an
6969 * associated connection) for greater than rx_idlePeerTime */
6971 struct rx_peer **peer_ptr, **peer_end;
6975 * Why do we need to hold the rx_peerHashTable_lock across
6976 * the incrementing of peer_ptr since the rx_peerHashTable
6977 * array is not changing? We don't.
6979 * By dropping the lock periodically we can permit other
6980 * activities to be performed while a rxi_ReapConnections
6981 * call is in progress. The goal of reap connections
6982 * is to clean up quickly without causing large amounts
6983 * of contention. Therefore, it is important that global
6984 * mutexes not be held for extended periods of time.
6986 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6987 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6989 struct rx_peer *peer, *next, *prev;
6991 MUTEX_ENTER(&rx_peerHashTable_lock);
6992 for (prev = peer = *peer_ptr; peer; peer = next) {
6994 code = MUTEX_TRYENTER(&peer->peer_lock);
6995 if ((code) && (peer->refCount == 0)
6996 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6997 rx_interface_stat_p rpc_stat, nrpc_stat;
7001 * now know that this peer object is one to be
7002 * removed from the hash table. Once it is removed
7003 * it can't be referenced by other threads.
7004 * Lets remove it first and decrement the struct
7005 * nPeerStructs count.
7007 if (peer == *peer_ptr) {
7013 if (rx_stats_active)
7014 rx_atomic_dec(&rx_stats.nPeerStructs);
7017 * Now if we hold references on 'prev' and 'next'
7018 * we can safely drop the rx_peerHashTable_lock
7019 * while we destroy this 'peer' object.
7025 MUTEX_EXIT(&rx_peerHashTable_lock);
7027 MUTEX_EXIT(&peer->peer_lock);
7028 MUTEX_DESTROY(&peer->peer_lock);
7030 (&peer->rpcStats, rpc_stat, nrpc_stat,
7031 rx_interface_stat)) {
7032 unsigned int num_funcs;
7035 queue_Remove(&rpc_stat->queue_header);
7036 queue_Remove(&rpc_stat->all_peers);
7037 num_funcs = rpc_stat->stats[0].func_total;
7039 sizeof(rx_interface_stat_t) +
7040 rpc_stat->stats[0].func_total *
7041 sizeof(rx_function_entry_v1_t);
7043 rxi_Free(rpc_stat, space);
7045 MUTEX_ENTER(&rx_rpc_stats);
7046 rxi_rpc_peer_stat_cnt -= num_funcs;
7047 MUTEX_EXIT(&rx_rpc_stats);
7052 * Regain the rx_peerHashTable_lock and
7053 * decrement the reference count on 'prev'
7056 MUTEX_ENTER(&rx_peerHashTable_lock);
7063 MUTEX_EXIT(&peer->peer_lock);
7068 MUTEX_EXIT(&rx_peerHashTable_lock);
7072 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7073 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7074 * GC, just below. Really, we shouldn't have to keep moving packets from
7075 * one place to another, but instead ought to always know if we can
7076 * afford to hold onto a packet in its particular use. */
7077 MUTEX_ENTER(&rx_freePktQ_lock);
7078 if (rx_waitingForPackets) {
7079 rx_waitingForPackets = 0;
7080 #ifdef RX_ENABLE_LOCKS
7081 CV_BROADCAST(&rx_waitingForPackets_cv);
7083 osi_rxWakeup(&rx_waitingForPackets);
7086 MUTEX_EXIT(&rx_freePktQ_lock);
7089 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7090 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
7094 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7095 * rx.h is sort of strange this is better. This is called with a security
7096 * object before it is discarded. Each connection using a security object has
7097 * its own refcount to the object so it won't actually be freed until the last
7098 * connection is destroyed.
7100 * This is the only rxs module call. A hold could also be written but no one
7104 rxs_Release(struct rx_securityClass *aobj)
7106 return RXS_Close(aobj);
7110 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
7111 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
7112 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
7113 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
7115 /* Adjust our estimate of the transmission rate to this peer, given
7116 * that the packet p was just acked. We can adjust peer->timeout and
7117 * call->twind. Pragmatically, this is called
7118 * only with packets of maximal length.
7119 * Called with peer and call locked.
7123 rxi_ComputeRate(struct rx_peer *peer, struct rx_call *call,
7124 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
7126 afs_int32 xferSize, xferMs;
7130 /* Count down packets */
7131 if (peer->rateFlag > 0)
7133 /* Do nothing until we're enabled */
7134 if (peer->rateFlag != 0)
7139 /* Count only when the ack seems legitimate */
7140 switch (ackReason) {
7141 case RX_ACK_REQUESTED:
7143 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
7147 case RX_ACK_PING_RESPONSE:
7148 if (p) /* want the response to ping-request, not data send */
7150 clock_GetTime(&newTO);
7151 if (clock_Gt(&newTO, &call->pingRequestTime)) {
7152 clock_Sub(&newTO, &call->pingRequestTime);
7153 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
7157 xferSize = rx_AckDataSize(rx_maxSendWindow) + RX_HEADER_SIZE;
7164 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %d.%06d, rtt %u, ps %u)\n",
7165 ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"),
7166 xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
7168 /* Track only packets that are big enough. */
7169 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
7173 /* absorb RTT data (in milliseconds) for these big packets */
7174 if (peer->smRtt == 0) {
7175 peer->smRtt = xferMs;
7177 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
7182 if (peer->countDown) {
7186 peer->countDown = 10; /* recalculate only every so often */
7188 /* In practice, we can measure only the RTT for full packets,
7189 * because of the way Rx acks the data that it receives. (If it's
7190 * smaller than a full packet, it often gets implicitly acked
7191 * either by the call response (from a server) or by the next call
7192 * (from a client), and either case confuses transmission times
7193 * with processing times.) Therefore, replace the above
7194 * more-sophisticated processing with a simpler version, where the
7195 * smoothed RTT is kept for full-size packets, and the time to
7196 * transmit a windowful of full-size packets is simply RTT *
7197 * windowSize. Again, we take two steps:
7198 - ensure the timeout is large enough for a single packet's RTT;
7199 - ensure that the window is small enough to fit in the desired timeout.*/
7201 /* First, the timeout check. */
7202 minTime = peer->smRtt;
7203 /* Get a reasonable estimate for a timeout period */
7205 newTO.sec = minTime / 1000;
7206 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
7208 /* Increase the timeout period so that we can always do at least
7209 * one packet exchange */
7210 if (clock_Gt(&newTO, &peer->timeout)) {
7212 dpf(("CONG peer %lx/%u: timeout %d.%06d ==> %ld.%06d (rtt %u)\n",
7213 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec,
7214 newTO.sec, newTO.usec, peer->smRtt));
7216 peer->timeout = newTO;
7219 /* Now, get an estimate for the transmit window size. */
7220 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
7221 /* Now, convert to the number of full packets that could fit in a
7222 * reasonable fraction of that interval */
7223 minTime /= (peer->smRtt << 1);
7224 minTime = MAX(minTime, rx_minPeerTimeout);
7225 xferSize = minTime; /* (make a copy) */
7227 /* Now clamp the size to reasonable bounds. */
7230 else if (minTime > rx_maxSendWindow)
7231 minTime = rx_maxSendWindow;
7232 /* if (minTime != peer->maxWindow) {
7233 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u)\n",
7234 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
7235 peer->timeout.sec, peer->timeout.usec, peer->smRtt));
7236 peer->maxWindow = minTime;
7237 elide... call->twind = minTime;
7241 /* Cut back on the peer timeout if it had earlier grown unreasonably.
7242 * Discern this by calculating the timeout necessary for rx_Window
7244 if ((xferSize > rx_maxSendWindow) && (peer->timeout.sec >= 3)) {
7245 /* calculate estimate for transmission interval in milliseconds */
7246 minTime = rx_maxSendWindow * peer->smRtt;
7247 if (minTime < 1000) {
7248 dpf(("CONG peer %lx/%u: cut TO %d.%06d by 0.5 (rtt %u)\n",
7249 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
7250 peer->timeout.usec, peer->smRtt));
7252 newTO.sec = 0; /* cut back on timeout by half a second */
7253 newTO.usec = 500000;
7254 clock_Sub(&peer->timeout, &newTO);
7259 } /* end of rxi_ComputeRate */
7260 #endif /* ADAPT_WINDOW */
7268 #define TRACE_OPTION_RX_DEBUG 16
7276 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7277 0, KEY_QUERY_VALUE, &parmKey);
7278 if (code != ERROR_SUCCESS)
7281 dummyLen = sizeof(TraceOption);
7282 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7283 (BYTE *) &TraceOption, &dummyLen);
7284 if (code == ERROR_SUCCESS) {
7285 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7287 RegCloseKey (parmKey);
7288 #endif /* AFS_NT40_ENV */
7293 rx_DebugOnOff(int on)
7297 rxdebug_active = on;
7303 rx_StatsOnOff(int on)
7305 rx_stats_active = on;
7309 /* Don't call this debugging routine directly; use dpf */
7311 rxi_DebugPrint(char *format, ...)
7320 va_start(ap, format);
7322 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7325 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7327 OutputDebugString(msg);
7333 va_start(ap, format);
7335 clock_GetTime(&now);
7336 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7337 (unsigned int)now.usec);
7338 vfprintf(rx_Log, format, ap);
7346 * This function is used to process the rx_stats structure that is local
7347 * to a process as well as an rx_stats structure received from a remote
7348 * process (via rxdebug). Therefore, it needs to do minimal version
7352 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7353 afs_int32 freePackets, char version)
7357 if (size != sizeof(struct rx_statistics)) {
7359 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7360 size, sizeof(struct rx_statistics));
7363 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7366 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7367 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7368 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7369 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7370 s->specialPktAllocFailures);
7372 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7373 s->receivePktAllocFailures, s->sendPktAllocFailures,
7374 s->specialPktAllocFailures);
7378 " greedy %u, " "bogusReads %u (last from host %x), "
7379 "noPackets %u, " "noBuffers %u, " "selects %u, "
7380 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7381 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7382 s->selects, s->sendSelects);
7384 fprintf(file, " packets read: ");
7385 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7386 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7388 fprintf(file, "\n");
7391 " other read counters: data %u, " "ack %u, " "dup %u "
7392 "spurious %u " "dally %u\n", s->dataPacketsRead,
7393 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7394 s->ignorePacketDally);
7396 fprintf(file, " packets sent: ");
7397 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7398 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7400 fprintf(file, "\n");
7403 " other send counters: ack %u, " "data %u (not resends), "
7404 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7405 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7406 s->dataPacketsPushed, s->ignoreAckedPacket);
7409 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7410 s->netSendFailures, (int)s->fatalErrors);
7412 if (s->nRttSamples) {
7413 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7414 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7416 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7417 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7421 " %d server connections, " "%d client connections, "
7422 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7423 s->nServerConns, s->nClientConns, s->nPeerStructs,
7424 s->nCallStructs, s->nFreeCallStructs);
7426 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7427 fprintf(file, " %d clock updates\n", clock_nUpdates);
7431 /* for backward compatibility */
7433 rx_PrintStats(FILE * file)
7435 MUTEX_ENTER(&rx_stats_mutex);
7436 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7437 sizeof(rx_stats), rx_nFreePackets,
7439 MUTEX_EXIT(&rx_stats_mutex);
7443 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7445 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
7446 ntohl(peer->host), (int)ntohs(peer->port), (int)peer->burstSize,
7447 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
7450 " Rtt %d, " "total sent %d, " "resent %d\n",
7451 peer->rtt, peer->nSent, peer->reSends);
7454 " Packet size %d, " "max in packet skew %d, "
7455 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
7456 (int)peer->outPacketSkew);
7460 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7462 * This mutex protects the following static variables:
7466 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7467 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7469 #define LOCK_RX_DEBUG
7470 #define UNLOCK_RX_DEBUG
7471 #endif /* AFS_PTHREAD_ENV */
7473 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7475 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7476 u_char type, void *inputData, size_t inputLength,
7477 void *outputData, size_t outputLength)
7479 static afs_int32 counter = 100;
7480 time_t waitTime, waitCount;
7481 struct rx_header theader;
7484 struct timeval tv_now, tv_wake, tv_delta;
7485 struct sockaddr_in taddr, faddr;
7499 tp = &tbuffer[sizeof(struct rx_header)];
7500 taddr.sin_family = AF_INET;
7501 taddr.sin_port = remotePort;
7502 taddr.sin_addr.s_addr = remoteAddr;
7503 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7504 taddr.sin_len = sizeof(struct sockaddr_in);
7507 memset(&theader, 0, sizeof(theader));
7508 theader.epoch = htonl(999);
7510 theader.callNumber = htonl(counter);
7513 theader.type = type;
7514 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7515 theader.serviceId = 0;
7517 memcpy(tbuffer, &theader, sizeof(theader));
7518 memcpy(tp, inputData, inputLength);
7520 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7521 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7523 /* see if there's a packet available */
7524 gettimeofday(&tv_wake, NULL);
7525 tv_wake.tv_sec += waitTime;
7528 FD_SET(socket, &imask);
7529 tv_delta.tv_sec = tv_wake.tv_sec;
7530 tv_delta.tv_usec = tv_wake.tv_usec;
7531 gettimeofday(&tv_now, NULL);
7533 if (tv_delta.tv_usec < tv_now.tv_usec) {
7535 tv_delta.tv_usec += 1000000;
7538 tv_delta.tv_usec -= tv_now.tv_usec;
7540 if (tv_delta.tv_sec < tv_now.tv_sec) {
7544 tv_delta.tv_sec -= tv_now.tv_sec;
7547 code = select(0, &imask, 0, 0, &tv_delta);
7548 #else /* AFS_NT40_ENV */
7549 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7550 #endif /* AFS_NT40_ENV */
7551 if (code == 1 && FD_ISSET(socket, &imask)) {
7552 /* now receive a packet */
7553 faddrLen = sizeof(struct sockaddr_in);
7555 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7556 (struct sockaddr *)&faddr, &faddrLen);
7559 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7560 if (counter == ntohl(theader.callNumber))
7568 /* see if we've timed out */
7576 code -= sizeof(struct rx_header);
7577 if (code > outputLength)
7578 code = outputLength;
7579 memcpy(outputData, tp, code);
7582 #endif /* RXDEBUG */
7585 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7586 afs_uint16 remotePort, struct rx_debugStats * stat,
7587 afs_uint32 * supportedValues)
7589 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7591 struct rx_debugIn in;
7593 *supportedValues = 0;
7594 in.type = htonl(RX_DEBUGI_GETSTATS);
7597 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7598 &in, sizeof(in), stat, sizeof(*stat));
7601 * If the call was successful, fixup the version and indicate
7602 * what contents of the stat structure are valid.
7603 * Also do net to host conversion of fields here.
7607 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7608 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7610 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7611 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7613 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7614 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7616 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7617 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7619 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7620 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7622 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7623 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7625 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7626 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7628 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7629 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7631 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7632 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7634 stat->nFreePackets = ntohl(stat->nFreePackets);
7635 stat->packetReclaims = ntohl(stat->packetReclaims);
7636 stat->callsExecuted = ntohl(stat->callsExecuted);
7637 stat->nWaiting = ntohl(stat->nWaiting);
7638 stat->idleThreads = ntohl(stat->idleThreads);
7639 stat->nWaited = ntohl(stat->nWaited);
7640 stat->nPackets = ntohl(stat->nPackets);
7649 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7650 afs_uint16 remotePort, struct rx_statistics * stat,
7651 afs_uint32 * supportedValues)
7653 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7655 struct rx_debugIn in;
7656 afs_int32 *lp = (afs_int32 *) stat;
7660 * supportedValues is currently unused, but added to allow future
7661 * versioning of this function.
7664 *supportedValues = 0;
7665 in.type = htonl(RX_DEBUGI_RXSTATS);
7667 memset(stat, 0, sizeof(*stat));
7669 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7670 &in, sizeof(in), stat, sizeof(*stat));
7675 * Do net to host conversion here
7678 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7689 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7690 afs_uint16 remotePort, size_t version_length,
7693 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7695 return MakeDebugCall(socket, remoteAddr, remotePort,
7696 RX_PACKET_TYPE_VERSION, a, 1, version,
7704 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7705 afs_uint16 remotePort, afs_int32 * nextConnection,
7706 int allConnections, afs_uint32 debugSupportedValues,
7707 struct rx_debugConn * conn,
7708 afs_uint32 * supportedValues)
7710 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7712 struct rx_debugIn in;
7716 * supportedValues is currently unused, but added to allow future
7717 * versioning of this function.
7720 *supportedValues = 0;
7721 if (allConnections) {
7722 in.type = htonl(RX_DEBUGI_GETALLCONN);
7724 in.type = htonl(RX_DEBUGI_GETCONN);
7726 in.index = htonl(*nextConnection);
7727 memset(conn, 0, sizeof(*conn));
7729 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7730 &in, sizeof(in), conn, sizeof(*conn));
7733 *nextConnection += 1;
7736 * Convert old connection format to new structure.
7739 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7740 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7741 #define MOVEvL(a) (conn->a = vL->a)
7743 /* any old or unrecognized version... */
7744 for (i = 0; i < RX_MAXCALLS; i++) {
7745 MOVEvL(callState[i]);
7746 MOVEvL(callMode[i]);
7747 MOVEvL(callFlags[i]);
7748 MOVEvL(callOther[i]);
7750 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7751 MOVEvL(secStats.type);
7752 MOVEvL(secStats.level);
7753 MOVEvL(secStats.flags);
7754 MOVEvL(secStats.expires);
7755 MOVEvL(secStats.packetsReceived);
7756 MOVEvL(secStats.packetsSent);
7757 MOVEvL(secStats.bytesReceived);
7758 MOVEvL(secStats.bytesSent);
7763 * Do net to host conversion here
7765 * I don't convert host or port since we are most likely
7766 * going to want these in NBO.
7768 conn->cid = ntohl(conn->cid);
7769 conn->serial = ntohl(conn->serial);
7770 for (i = 0; i < RX_MAXCALLS; i++) {
7771 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7773 conn->error = ntohl(conn->error);
7774 conn->secStats.flags = ntohl(conn->secStats.flags);
7775 conn->secStats.expires = ntohl(conn->secStats.expires);
7776 conn->secStats.packetsReceived =
7777 ntohl(conn->secStats.packetsReceived);
7778 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7779 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7780 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7781 conn->epoch = ntohl(conn->epoch);
7782 conn->natMTU = ntohl(conn->natMTU);
7791 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7792 afs_uint16 remotePort, afs_int32 * nextPeer,
7793 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7794 afs_uint32 * supportedValues)
7796 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7798 struct rx_debugIn in;
7801 * supportedValues is currently unused, but added to allow future
7802 * versioning of this function.
7805 *supportedValues = 0;
7806 in.type = htonl(RX_DEBUGI_GETPEER);
7807 in.index = htonl(*nextPeer);
7808 memset(peer, 0, sizeof(*peer));
7810 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7811 &in, sizeof(in), peer, sizeof(*peer));
7817 * Do net to host conversion here
7819 * I don't convert host or port since we are most likely
7820 * going to want these in NBO.
7822 peer->ifMTU = ntohs(peer->ifMTU);
7823 peer->idleWhen = ntohl(peer->idleWhen);
7824 peer->refCount = ntohs(peer->refCount);
7825 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7826 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7827 peer->rtt = ntohl(peer->rtt);
7828 peer->rtt_dev = ntohl(peer->rtt_dev);
7829 peer->timeout.sec = 0;
7830 peer->timeout.usec = 0;
7831 peer->nSent = ntohl(peer->nSent);
7832 peer->reSends = ntohl(peer->reSends);
7833 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7834 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7835 peer->rateFlag = ntohl(peer->rateFlag);
7836 peer->natMTU = ntohs(peer->natMTU);
7837 peer->maxMTU = ntohs(peer->maxMTU);
7838 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7839 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7840 peer->MTU = ntohs(peer->MTU);
7841 peer->cwind = ntohs(peer->cwind);
7842 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7843 peer->congestSeq = ntohs(peer->congestSeq);
7844 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7845 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7846 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7847 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7856 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7857 struct rx_debugPeer * peerStats)
7860 afs_int32 error = 1; /* default to "did not succeed" */
7861 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7863 MUTEX_ENTER(&rx_peerHashTable_lock);
7864 for(tp = rx_peerHashTable[hashValue];
7865 tp != NULL; tp = tp->next) {
7866 if (tp->host == peerHost)
7872 MUTEX_EXIT(&rx_peerHashTable_lock);
7876 MUTEX_ENTER(&tp->peer_lock);
7877 peerStats->host = tp->host;
7878 peerStats->port = tp->port;
7879 peerStats->ifMTU = tp->ifMTU;
7880 peerStats->idleWhen = tp->idleWhen;
7881 peerStats->refCount = tp->refCount;
7882 peerStats->burstSize = tp->burstSize;
7883 peerStats->burst = tp->burst;
7884 peerStats->burstWait.sec = tp->burstWait.sec;
7885 peerStats->burstWait.usec = tp->burstWait.usec;
7886 peerStats->rtt = tp->rtt;
7887 peerStats->rtt_dev = tp->rtt_dev;
7888 peerStats->timeout.sec = 0;
7889 peerStats->timeout.usec = 0;
7890 peerStats->nSent = tp->nSent;
7891 peerStats->reSends = tp->reSends;
7892 peerStats->inPacketSkew = tp->inPacketSkew;
7893 peerStats->outPacketSkew = tp->outPacketSkew;
7894 peerStats->rateFlag = tp->rateFlag;
7895 peerStats->natMTU = tp->natMTU;
7896 peerStats->maxMTU = tp->maxMTU;
7897 peerStats->maxDgramPackets = tp->maxDgramPackets;
7898 peerStats->ifDgramPackets = tp->ifDgramPackets;
7899 peerStats->MTU = tp->MTU;
7900 peerStats->cwind = tp->cwind;
7901 peerStats->nDgramPackets = tp->nDgramPackets;
7902 peerStats->congestSeq = tp->congestSeq;
7903 peerStats->bytesSent.high = tp->bytesSent.high;
7904 peerStats->bytesSent.low = tp->bytesSent.low;
7905 peerStats->bytesReceived.high = tp->bytesReceived.high;
7906 peerStats->bytesReceived.low = tp->bytesReceived.low;
7907 MUTEX_EXIT(&tp->peer_lock);
7909 MUTEX_ENTER(&rx_peerHashTable_lock);
7912 MUTEX_EXIT(&rx_peerHashTable_lock);
7920 struct rx_serverQueueEntry *np;
7923 struct rx_call *call;
7924 struct rx_serverQueueEntry *sq;
7928 if (rxinit_status == 1) {
7930 return; /* Already shutdown. */
7934 #ifndef AFS_PTHREAD_ENV
7935 FD_ZERO(&rx_selectMask);
7936 #endif /* AFS_PTHREAD_ENV */
7937 rxi_dataQuota = RX_MAX_QUOTA;
7938 #ifndef AFS_PTHREAD_ENV
7940 #endif /* AFS_PTHREAD_ENV */
7943 #ifndef AFS_PTHREAD_ENV
7944 #ifndef AFS_USE_GETTIMEOFDAY
7946 #endif /* AFS_USE_GETTIMEOFDAY */
7947 #endif /* AFS_PTHREAD_ENV */
7949 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7950 call = queue_First(&rx_freeCallQueue, rx_call);
7952 rxi_Free(call, sizeof(struct rx_call));
7955 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7956 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7962 struct rx_peer **peer_ptr, **peer_end;
7963 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7964 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7966 struct rx_peer *peer, *next;
7968 MUTEX_ENTER(&rx_peerHashTable_lock);
7969 for (peer = *peer_ptr; peer; peer = next) {
7970 rx_interface_stat_p rpc_stat, nrpc_stat;
7973 MUTEX_ENTER(&rx_rpc_stats);
7974 MUTEX_ENTER(&peer->peer_lock);
7976 (&peer->rpcStats, rpc_stat, nrpc_stat,
7977 rx_interface_stat)) {
7978 unsigned int num_funcs;
7981 queue_Remove(&rpc_stat->queue_header);
7982 queue_Remove(&rpc_stat->all_peers);
7983 num_funcs = rpc_stat->stats[0].func_total;
7985 sizeof(rx_interface_stat_t) +
7986 rpc_stat->stats[0].func_total *
7987 sizeof(rx_function_entry_v1_t);
7989 rxi_Free(rpc_stat, space);
7991 /* rx_rpc_stats must be held */
7992 rxi_rpc_peer_stat_cnt -= num_funcs;
7994 MUTEX_EXIT(&peer->peer_lock);
7995 MUTEX_EXIT(&rx_rpc_stats);
7999 if (rx_stats_active)
8000 rx_atomic_dec(&rx_stats.nPeerStructs);
8002 MUTEX_EXIT(&rx_peerHashTable_lock);
8005 for (i = 0; i < RX_MAX_SERVICES; i++) {
8007 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
8009 for (i = 0; i < rx_hashTableSize; i++) {
8010 struct rx_connection *tc, *ntc;
8011 MUTEX_ENTER(&rx_connHashTable_lock);
8012 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
8014 for (j = 0; j < RX_MAXCALLS; j++) {
8016 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
8019 rxi_Free(tc, sizeof(*tc));
8021 MUTEX_EXIT(&rx_connHashTable_lock);
8024 MUTEX_ENTER(&freeSQEList_lock);
8026 while ((np = rx_FreeSQEList)) {
8027 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
8028 MUTEX_DESTROY(&np->lock);
8029 rxi_Free(np, sizeof(*np));
8032 MUTEX_EXIT(&freeSQEList_lock);
8033 MUTEX_DESTROY(&freeSQEList_lock);
8034 MUTEX_DESTROY(&rx_freeCallQueue_lock);
8035 MUTEX_DESTROY(&rx_connHashTable_lock);
8036 MUTEX_DESTROY(&rx_peerHashTable_lock);
8037 MUTEX_DESTROY(&rx_serverPool_lock);
8039 osi_Free(rx_connHashTable,
8040 rx_hashTableSize * sizeof(struct rx_connection *));
8041 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8043 UNPIN(rx_connHashTable,
8044 rx_hashTableSize * sizeof(struct rx_connection *));
8045 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8047 rxi_FreeAllPackets();
8049 MUTEX_ENTER(&rx_quota_mutex);
8050 rxi_dataQuota = RX_MAX_QUOTA;
8051 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
8052 MUTEX_EXIT(&rx_quota_mutex);
8057 #ifdef RX_ENABLE_LOCKS
8059 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
8061 if (!MUTEX_ISMINE(lockaddr))
8062 osi_Panic("Lock not held: %s", msg);
8064 #endif /* RX_ENABLE_LOCKS */
8069 * Routines to implement connection specific data.
8073 rx_KeyCreate(rx_destructor_t rtn)
8076 MUTEX_ENTER(&rxi_keyCreate_lock);
8077 key = rxi_keyCreate_counter++;
8078 rxi_keyCreate_destructor = (rx_destructor_t *)
8079 realloc((void *)rxi_keyCreate_destructor,
8080 (key + 1) * sizeof(rx_destructor_t));
8081 rxi_keyCreate_destructor[key] = rtn;
8082 MUTEX_EXIT(&rxi_keyCreate_lock);
8087 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
8090 MUTEX_ENTER(&conn->conn_data_lock);
8091 if (!conn->specific) {
8092 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
8093 for (i = 0; i < key; i++)
8094 conn->specific[i] = NULL;
8095 conn->nSpecific = key + 1;
8096 conn->specific[key] = ptr;
8097 } else if (key >= conn->nSpecific) {
8098 conn->specific = (void **)
8099 realloc(conn->specific, (key + 1) * sizeof(void *));
8100 for (i = conn->nSpecific; i < key; i++)
8101 conn->specific[i] = NULL;
8102 conn->nSpecific = key + 1;
8103 conn->specific[key] = ptr;
8105 if (conn->specific[key] && rxi_keyCreate_destructor[key])
8106 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
8107 conn->specific[key] = ptr;
8109 MUTEX_EXIT(&conn->conn_data_lock);
8113 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
8116 MUTEX_ENTER(&svc->svc_data_lock);
8117 if (!svc->specific) {
8118 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
8119 for (i = 0; i < key; i++)
8120 svc->specific[i] = NULL;
8121 svc->nSpecific = key + 1;
8122 svc->specific[key] = ptr;
8123 } else if (key >= svc->nSpecific) {
8124 svc->specific = (void **)
8125 realloc(svc->specific, (key + 1) * sizeof(void *));
8126 for (i = svc->nSpecific; i < key; i++)
8127 svc->specific[i] = NULL;
8128 svc->nSpecific = key + 1;
8129 svc->specific[key] = ptr;
8131 if (svc->specific[key] && rxi_keyCreate_destructor[key])
8132 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
8133 svc->specific[key] = ptr;
8135 MUTEX_EXIT(&svc->svc_data_lock);
8139 rx_GetSpecific(struct rx_connection *conn, int key)
8142 MUTEX_ENTER(&conn->conn_data_lock);
8143 if (key >= conn->nSpecific)
8146 ptr = conn->specific[key];
8147 MUTEX_EXIT(&conn->conn_data_lock);
8152 rx_GetServiceSpecific(struct rx_service *svc, int key)
8155 MUTEX_ENTER(&svc->svc_data_lock);
8156 if (key >= svc->nSpecific)
8159 ptr = svc->specific[key];
8160 MUTEX_EXIT(&svc->svc_data_lock);
8165 #endif /* !KERNEL */
8168 * processStats is a queue used to store the statistics for the local
8169 * process. Its contents are similar to the contents of the rpcStats
8170 * queue on a rx_peer structure, but the actual data stored within
8171 * this queue contains totals across the lifetime of the process (assuming
8172 * the stats have not been reset) - unlike the per peer structures
8173 * which can come and go based upon the peer lifetime.
8176 static struct rx_queue processStats = { &processStats, &processStats };
8179 * peerStats is a queue used to store the statistics for all peer structs.
8180 * Its contents are the union of all the peer rpcStats queues.
8183 static struct rx_queue peerStats = { &peerStats, &peerStats };
8186 * rxi_monitor_processStats is used to turn process wide stat collection
8190 static int rxi_monitor_processStats = 0;
8193 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8196 static int rxi_monitor_peerStats = 0;
8199 * rxi_AddRpcStat - given all of the information for a particular rpc
8200 * call, create (if needed) and update the stat totals for the rpc.
8204 * IN stats - the queue of stats that will be updated with the new value
8206 * IN rxInterface - a unique number that identifies the rpc interface
8208 * IN currentFunc - the index of the function being invoked
8210 * IN totalFunc - the total number of functions in this interface
8212 * IN queueTime - the amount of time this function waited for a thread
8214 * IN execTime - the amount of time this function invocation took to execute
8216 * IN bytesSent - the number bytes sent by this invocation
8218 * IN bytesRcvd - the number bytes received by this invocation
8220 * IN isServer - if true, this invocation was made to a server
8222 * IN remoteHost - the ip address of the remote host
8224 * IN remotePort - the port of the remote host
8226 * IN addToPeerList - if != 0, add newly created stat to the global peer list
8228 * INOUT counter - if a new stats structure is allocated, the counter will
8229 * be updated with the new number of allocated stat structures
8237 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
8238 afs_uint32 currentFunc, afs_uint32 totalFunc,
8239 struct clock *queueTime, struct clock *execTime,
8240 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
8241 afs_uint32 remoteHost, afs_uint32 remotePort,
8242 int addToPeerList, unsigned int *counter)
8245 rx_interface_stat_p rpc_stat, nrpc_stat;
8248 * See if there's already a structure for this interface
8251 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8252 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8253 && (rpc_stat->stats[0].remote_is_server == isServer))
8258 * Didn't find a match so allocate a new structure and add it to the
8262 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
8263 || (rpc_stat->stats[0].interfaceId != rxInterface)
8264 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8269 sizeof(rx_interface_stat_t) +
8270 totalFunc * sizeof(rx_function_entry_v1_t);
8272 rpc_stat = rxi_Alloc(space);
8273 if (rpc_stat == NULL) {
8277 *counter += totalFunc;
8278 for (i = 0; i < totalFunc; i++) {
8279 rpc_stat->stats[i].remote_peer = remoteHost;
8280 rpc_stat->stats[i].remote_port = remotePort;
8281 rpc_stat->stats[i].remote_is_server = isServer;
8282 rpc_stat->stats[i].interfaceId = rxInterface;
8283 rpc_stat->stats[i].func_total = totalFunc;
8284 rpc_stat->stats[i].func_index = i;
8285 hzero(rpc_stat->stats[i].invocations);
8286 hzero(rpc_stat->stats[i].bytes_sent);
8287 hzero(rpc_stat->stats[i].bytes_rcvd);
8288 rpc_stat->stats[i].queue_time_sum.sec = 0;
8289 rpc_stat->stats[i].queue_time_sum.usec = 0;
8290 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8291 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8292 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8293 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8294 rpc_stat->stats[i].queue_time_max.sec = 0;
8295 rpc_stat->stats[i].queue_time_max.usec = 0;
8296 rpc_stat->stats[i].execution_time_sum.sec = 0;
8297 rpc_stat->stats[i].execution_time_sum.usec = 0;
8298 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8299 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8300 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8301 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8302 rpc_stat->stats[i].execution_time_max.sec = 0;
8303 rpc_stat->stats[i].execution_time_max.usec = 0;
8305 queue_Prepend(stats, rpc_stat);
8306 if (addToPeerList) {
8307 queue_Prepend(&peerStats, &rpc_stat->all_peers);
8312 * Increment the stats for this function
8315 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
8316 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
8317 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
8318 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8319 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8320 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8321 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8323 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8324 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8326 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8327 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8329 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8330 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8332 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8333 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8341 * rx_IncrementTimeAndCount - increment the times and count for a particular
8346 * IN peer - the peer who invoked the rpc
8348 * IN rxInterface - a unique number that identifies the rpc interface
8350 * IN currentFunc - the index of the function being invoked
8352 * IN totalFunc - the total number of functions in this interface
8354 * IN queueTime - the amount of time this function waited for a thread
8356 * IN execTime - the amount of time this function invocation took to execute
8358 * IN bytesSent - the number bytes sent by this invocation
8360 * IN bytesRcvd - the number bytes received by this invocation
8362 * IN isServer - if true, this invocation was made to a server
8370 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8371 afs_uint32 currentFunc, afs_uint32 totalFunc,
8372 struct clock *queueTime, struct clock *execTime,
8373 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8377 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8380 MUTEX_ENTER(&rx_rpc_stats);
8382 if (rxi_monitor_peerStats) {
8383 MUTEX_ENTER(&peer->peer_lock);
8384 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8385 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8386 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8387 MUTEX_EXIT(&peer->peer_lock);
8390 if (rxi_monitor_processStats) {
8391 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8392 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8393 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8396 MUTEX_EXIT(&rx_rpc_stats);
8401 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8405 * IN callerVersion - the rpc stat version of the caller.
8407 * IN count - the number of entries to marshall.
8409 * IN stats - pointer to stats to be marshalled.
8411 * OUT ptr - Where to store the marshalled data.
8418 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8419 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8425 * We only support the first version
8427 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8428 *(ptr++) = stats->remote_peer;
8429 *(ptr++) = stats->remote_port;
8430 *(ptr++) = stats->remote_is_server;
8431 *(ptr++) = stats->interfaceId;
8432 *(ptr++) = stats->func_total;
8433 *(ptr++) = stats->func_index;
8434 *(ptr++) = hgethi(stats->invocations);
8435 *(ptr++) = hgetlo(stats->invocations);
8436 *(ptr++) = hgethi(stats->bytes_sent);
8437 *(ptr++) = hgetlo(stats->bytes_sent);
8438 *(ptr++) = hgethi(stats->bytes_rcvd);
8439 *(ptr++) = hgetlo(stats->bytes_rcvd);
8440 *(ptr++) = stats->queue_time_sum.sec;
8441 *(ptr++) = stats->queue_time_sum.usec;
8442 *(ptr++) = stats->queue_time_sum_sqr.sec;
8443 *(ptr++) = stats->queue_time_sum_sqr.usec;
8444 *(ptr++) = stats->queue_time_min.sec;
8445 *(ptr++) = stats->queue_time_min.usec;
8446 *(ptr++) = stats->queue_time_max.sec;
8447 *(ptr++) = stats->queue_time_max.usec;
8448 *(ptr++) = stats->execution_time_sum.sec;
8449 *(ptr++) = stats->execution_time_sum.usec;
8450 *(ptr++) = stats->execution_time_sum_sqr.sec;
8451 *(ptr++) = stats->execution_time_sum_sqr.usec;
8452 *(ptr++) = stats->execution_time_min.sec;
8453 *(ptr++) = stats->execution_time_min.usec;
8454 *(ptr++) = stats->execution_time_max.sec;
8455 *(ptr++) = stats->execution_time_max.usec;
8461 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8466 * IN callerVersion - the rpc stat version of the caller
8468 * OUT myVersion - the rpc stat version of this function
8470 * OUT clock_sec - local time seconds
8472 * OUT clock_usec - local time microseconds
8474 * OUT allocSize - the number of bytes allocated to contain stats
8476 * OUT statCount - the number stats retrieved from this process.
8478 * OUT stats - the actual stats retrieved from this process.
8482 * Returns void. If successful, stats will != NULL.
8486 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8487 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8488 size_t * allocSize, afs_uint32 * statCount,
8489 afs_uint32 ** stats)
8499 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8502 * Check to see if stats are enabled
8505 MUTEX_ENTER(&rx_rpc_stats);
8506 if (!rxi_monitor_processStats) {
8507 MUTEX_EXIT(&rx_rpc_stats);
8511 clock_GetTime(&now);
8512 *clock_sec = now.sec;
8513 *clock_usec = now.usec;
8516 * Allocate the space based upon the caller version
8518 * If the client is at an older version than we are,
8519 * we return the statistic data in the older data format, but
8520 * we still return our version number so the client knows we
8521 * are maintaining more data than it can retrieve.
8524 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8525 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8526 *statCount = rxi_rpc_process_stat_cnt;
8529 * This can't happen yet, but in the future version changes
8530 * can be handled by adding additional code here
8534 if (space > (size_t) 0) {
8536 ptr = *stats = rxi_Alloc(space);
8539 rx_interface_stat_p rpc_stat, nrpc_stat;
8543 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8545 * Copy the data based upon the caller version
8547 rx_MarshallProcessRPCStats(callerVersion,
8548 rpc_stat->stats[0].func_total,
8549 rpc_stat->stats, &ptr);
8555 MUTEX_EXIT(&rx_rpc_stats);
8560 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8564 * IN callerVersion - the rpc stat version of the caller
8566 * OUT myVersion - the rpc stat version of this function
8568 * OUT clock_sec - local time seconds
8570 * OUT clock_usec - local time microseconds
8572 * OUT allocSize - the number of bytes allocated to contain stats
8574 * OUT statCount - the number of stats retrieved from the individual
8577 * OUT stats - the actual stats retrieved from the individual peer structures.
8581 * Returns void. If successful, stats will != NULL.
8585 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8586 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8587 size_t * allocSize, afs_uint32 * statCount,
8588 afs_uint32 ** stats)
8598 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8601 * Check to see if stats are enabled
8604 MUTEX_ENTER(&rx_rpc_stats);
8605 if (!rxi_monitor_peerStats) {
8606 MUTEX_EXIT(&rx_rpc_stats);
8610 clock_GetTime(&now);
8611 *clock_sec = now.sec;
8612 *clock_usec = now.usec;
8615 * Allocate the space based upon the caller version
8617 * If the client is at an older version than we are,
8618 * we return the statistic data in the older data format, but
8619 * we still return our version number so the client knows we
8620 * are maintaining more data than it can retrieve.
8623 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8624 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8625 *statCount = rxi_rpc_peer_stat_cnt;
8628 * This can't happen yet, but in the future version changes
8629 * can be handled by adding additional code here
8633 if (space > (size_t) 0) {
8635 ptr = *stats = rxi_Alloc(space);
8638 rx_interface_stat_p rpc_stat, nrpc_stat;
8642 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8644 * We have to fix the offset of rpc_stat since we are
8645 * keeping this structure on two rx_queues. The rx_queue
8646 * package assumes that the rx_queue member is the first
8647 * member of the structure. That is, rx_queue assumes that
8648 * any one item is only on one queue at a time. We are
8649 * breaking that assumption and so we have to do a little
8650 * math to fix our pointers.
8653 fix_offset = (char *)rpc_stat;
8654 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8655 rpc_stat = (rx_interface_stat_p) fix_offset;
8658 * Copy the data based upon the caller version
8660 rx_MarshallProcessRPCStats(callerVersion,
8661 rpc_stat->stats[0].func_total,
8662 rpc_stat->stats, &ptr);
8668 MUTEX_EXIT(&rx_rpc_stats);
8673 * rx_FreeRPCStats - free memory allocated by
8674 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8678 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8679 * rx_RetrievePeerRPCStats
8681 * IN allocSize - the number of bytes in stats.
8689 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8691 rxi_Free(stats, allocSize);
8695 * rx_queryProcessRPCStats - see if process rpc stat collection is
8696 * currently enabled.
8702 * Returns 0 if stats are not enabled != 0 otherwise
8706 rx_queryProcessRPCStats(void)
8709 MUTEX_ENTER(&rx_rpc_stats);
8710 rc = rxi_monitor_processStats;
8711 MUTEX_EXIT(&rx_rpc_stats);
8716 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8722 * Returns 0 if stats are not enabled != 0 otherwise
8726 rx_queryPeerRPCStats(void)
8729 MUTEX_ENTER(&rx_rpc_stats);
8730 rc = rxi_monitor_peerStats;
8731 MUTEX_EXIT(&rx_rpc_stats);
8736 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8746 rx_enableProcessRPCStats(void)
8748 MUTEX_ENTER(&rx_rpc_stats);
8749 rx_enable_stats = 1;
8750 rxi_monitor_processStats = 1;
8751 MUTEX_EXIT(&rx_rpc_stats);
8755 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8765 rx_enablePeerRPCStats(void)
8767 MUTEX_ENTER(&rx_rpc_stats);
8768 rx_enable_stats = 1;
8769 rxi_monitor_peerStats = 1;
8770 MUTEX_EXIT(&rx_rpc_stats);
8774 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8784 rx_disableProcessRPCStats(void)
8786 rx_interface_stat_p rpc_stat, nrpc_stat;
8789 MUTEX_ENTER(&rx_rpc_stats);
8792 * Turn off process statistics and if peer stats is also off, turn
8796 rxi_monitor_processStats = 0;
8797 if (rxi_monitor_peerStats == 0) {
8798 rx_enable_stats = 0;
8801 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8802 unsigned int num_funcs = 0;
8805 queue_Remove(rpc_stat);
8806 num_funcs = rpc_stat->stats[0].func_total;
8808 sizeof(rx_interface_stat_t) +
8809 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8811 rxi_Free(rpc_stat, space);
8812 rxi_rpc_process_stat_cnt -= num_funcs;
8814 MUTEX_EXIT(&rx_rpc_stats);
8818 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8828 rx_disablePeerRPCStats(void)
8830 struct rx_peer **peer_ptr, **peer_end;
8834 * Turn off peer statistics and if process stats is also off, turn
8838 rxi_monitor_peerStats = 0;
8839 if (rxi_monitor_processStats == 0) {
8840 rx_enable_stats = 0;
8843 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8844 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8846 struct rx_peer *peer, *next, *prev;
8848 MUTEX_ENTER(&rx_peerHashTable_lock);
8849 MUTEX_ENTER(&rx_rpc_stats);
8850 for (prev = peer = *peer_ptr; peer; peer = next) {
8852 code = MUTEX_TRYENTER(&peer->peer_lock);
8854 rx_interface_stat_p rpc_stat, nrpc_stat;
8857 if (prev == *peer_ptr) {
8868 MUTEX_EXIT(&rx_peerHashTable_lock);
8871 (&peer->rpcStats, rpc_stat, nrpc_stat,
8872 rx_interface_stat)) {
8873 unsigned int num_funcs = 0;
8876 queue_Remove(&rpc_stat->queue_header);
8877 queue_Remove(&rpc_stat->all_peers);
8878 num_funcs = rpc_stat->stats[0].func_total;
8880 sizeof(rx_interface_stat_t) +
8881 rpc_stat->stats[0].func_total *
8882 sizeof(rx_function_entry_v1_t);
8884 rxi_Free(rpc_stat, space);
8885 rxi_rpc_peer_stat_cnt -= num_funcs;
8887 MUTEX_EXIT(&peer->peer_lock);
8889 MUTEX_ENTER(&rx_peerHashTable_lock);
8899 MUTEX_EXIT(&rx_rpc_stats);
8900 MUTEX_EXIT(&rx_peerHashTable_lock);
8905 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8910 * IN clearFlag - flag indicating which stats to clear
8918 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8920 rx_interface_stat_p rpc_stat, nrpc_stat;
8922 MUTEX_ENTER(&rx_rpc_stats);
8924 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8925 unsigned int num_funcs = 0, i;
8926 num_funcs = rpc_stat->stats[0].func_total;
8927 for (i = 0; i < num_funcs; i++) {
8928 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8929 hzero(rpc_stat->stats[i].invocations);
8931 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8932 hzero(rpc_stat->stats[i].bytes_sent);
8934 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8935 hzero(rpc_stat->stats[i].bytes_rcvd);
8937 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8938 rpc_stat->stats[i].queue_time_sum.sec = 0;
8939 rpc_stat->stats[i].queue_time_sum.usec = 0;
8941 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8942 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8943 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8945 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8946 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8947 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8949 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8950 rpc_stat->stats[i].queue_time_max.sec = 0;
8951 rpc_stat->stats[i].queue_time_max.usec = 0;
8953 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8954 rpc_stat->stats[i].execution_time_sum.sec = 0;
8955 rpc_stat->stats[i].execution_time_sum.usec = 0;
8957 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8958 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8959 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8961 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8962 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8963 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8965 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8966 rpc_stat->stats[i].execution_time_max.sec = 0;
8967 rpc_stat->stats[i].execution_time_max.usec = 0;
8972 MUTEX_EXIT(&rx_rpc_stats);
8976 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8981 * IN clearFlag - flag indicating which stats to clear
8989 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8991 rx_interface_stat_p rpc_stat, nrpc_stat;
8993 MUTEX_ENTER(&rx_rpc_stats);
8995 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8996 unsigned int num_funcs = 0, i;
8999 * We have to fix the offset of rpc_stat since we are
9000 * keeping this structure on two rx_queues. The rx_queue
9001 * package assumes that the rx_queue member is the first
9002 * member of the structure. That is, rx_queue assumes that
9003 * any one item is only on one queue at a time. We are
9004 * breaking that assumption and so we have to do a little
9005 * math to fix our pointers.
9008 fix_offset = (char *)rpc_stat;
9009 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
9010 rpc_stat = (rx_interface_stat_p) fix_offset;
9012 num_funcs = rpc_stat->stats[0].func_total;
9013 for (i = 0; i < num_funcs; i++) {
9014 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9015 hzero(rpc_stat->stats[i].invocations);
9017 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9018 hzero(rpc_stat->stats[i].bytes_sent);
9020 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9021 hzero(rpc_stat->stats[i].bytes_rcvd);
9023 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9024 rpc_stat->stats[i].queue_time_sum.sec = 0;
9025 rpc_stat->stats[i].queue_time_sum.usec = 0;
9027 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9028 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9029 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9031 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9032 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9033 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9035 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9036 rpc_stat->stats[i].queue_time_max.sec = 0;
9037 rpc_stat->stats[i].queue_time_max.usec = 0;
9039 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9040 rpc_stat->stats[i].execution_time_sum.sec = 0;
9041 rpc_stat->stats[i].execution_time_sum.usec = 0;
9043 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9044 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9045 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9047 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9048 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9049 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9051 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9052 rpc_stat->stats[i].execution_time_max.sec = 0;
9053 rpc_stat->stats[i].execution_time_max.usec = 0;
9058 MUTEX_EXIT(&rx_rpc_stats);
9062 * rxi_rxstat_userok points to a routine that returns 1 if the caller
9063 * is authorized to enable/disable/clear RX statistics.
9065 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
9068 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
9070 rxi_rxstat_userok = proc;
9074 rx_RxStatUserOk(struct rx_call *call)
9076 if (!rxi_rxstat_userok)
9078 return rxi_rxstat_userok(call);
9083 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
9084 * function in the MSVC runtime DLL (msvcrt.dll).
9086 * Note: the system serializes calls to this function.
9089 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
9090 DWORD reason, /* reason function is being called */
9091 LPVOID reserved) /* reserved for future use */
9094 case DLL_PROCESS_ATTACH:
9095 /* library is being attached to a process */
9099 case DLL_PROCESS_DETACH:
9106 #endif /* AFS_NT40_ENV */
9109 int rx_DumpCalls(FILE *outputFile, char *cookie)
9111 #ifdef RXDEBUG_PACKET
9112 #ifdef KDUMP_RX_LOCK
9113 struct rx_call_rx_lock *c;
9120 #define RXDPRINTF sprintf
9121 #define RXDPRINTOUT output
9123 #define RXDPRINTF fprintf
9124 #define RXDPRINTOUT outputFile
9127 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9129 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9132 for (c = rx_allCallsp; c; c = c->allNextp) {
9133 u_short rqc, tqc, iovqc;
9134 struct rx_packet *p, *np;
9136 MUTEX_ENTER(&c->lock);
9137 queue_Count(&c->rq, p, np, rx_packet, rqc);
9138 queue_Count(&c->tq, p, np, rx_packet, tqc);
9139 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
9141 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, "
9142 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9143 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9144 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9145 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
9146 #ifdef RX_ENABLE_LOCKS
9149 #ifdef RX_REFCOUNT_CHECK
9150 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9151 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9154 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,
9155 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9156 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9157 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9158 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9159 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
9160 #ifdef RX_ENABLE_LOCKS
9161 , (afs_uint32)c->refCount
9163 #ifdef RX_REFCOUNT_CHECK
9164 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9167 MUTEX_EXIT(&c->lock);
9170 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9173 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9175 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9177 #endif /* RXDEBUG_PACKET */