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 */
1811 /* if server is restarting( typically smooth shutdown) then do not
1812 * allow any new calls.
1815 if (rx_tranquil && (call != NULL)) {
1819 MUTEX_ENTER(&call->lock);
1821 rxi_CallError(call, RX_RESTARTING);
1822 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1824 MUTEX_EXIT(&call->lock);
1828 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1829 #ifdef RX_ENABLE_LOCKS
1831 #endif /* RX_ENABLE_LOCKS */
1832 afs_termState = AFSOP_STOP_AFS;
1833 afs_osi_Wakeup(&afs_termState);
1834 #ifdef RX_ENABLE_LOCKS
1836 #endif /* RX_ENABLE_LOCKS */
1841 tservice = call->conn->service;
1843 if (tservice->beforeProc)
1844 (*tservice->beforeProc) (call);
1846 code = tservice->executeRequestProc(call);
1848 if (tservice->afterProc)
1849 (*tservice->afterProc) (call, code);
1851 rx_EndCall(call, code);
1852 if (rx_stats_active) {
1853 MUTEX_ENTER(&rx_stats_mutex);
1855 MUTEX_EXIT(&rx_stats_mutex);
1862 rx_WakeupServerProcs(void)
1864 struct rx_serverQueueEntry *np, *tqp;
1868 MUTEX_ENTER(&rx_serverPool_lock);
1870 #ifdef RX_ENABLE_LOCKS
1871 if (rx_waitForPacket)
1872 CV_BROADCAST(&rx_waitForPacket->cv);
1873 #else /* RX_ENABLE_LOCKS */
1874 if (rx_waitForPacket)
1875 osi_rxWakeup(rx_waitForPacket);
1876 #endif /* RX_ENABLE_LOCKS */
1877 MUTEX_ENTER(&freeSQEList_lock);
1878 for (np = rx_FreeSQEList; np; np = tqp) {
1879 tqp = *(struct rx_serverQueueEntry **)np;
1880 #ifdef RX_ENABLE_LOCKS
1881 CV_BROADCAST(&np->cv);
1882 #else /* RX_ENABLE_LOCKS */
1884 #endif /* RX_ENABLE_LOCKS */
1886 MUTEX_EXIT(&freeSQEList_lock);
1887 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1888 #ifdef RX_ENABLE_LOCKS
1889 CV_BROADCAST(&np->cv);
1890 #else /* RX_ENABLE_LOCKS */
1892 #endif /* RX_ENABLE_LOCKS */
1894 MUTEX_EXIT(&rx_serverPool_lock);
1899 * One thing that seems to happen is that all the server threads get
1900 * tied up on some empty or slow call, and then a whole bunch of calls
1901 * arrive at once, using up the packet pool, so now there are more
1902 * empty calls. The most critical resources here are server threads
1903 * and the free packet pool. The "doreclaim" code seems to help in
1904 * general. I think that eventually we arrive in this state: there
1905 * are lots of pending calls which do have all their packets present,
1906 * so they won't be reclaimed, are multi-packet calls, so they won't
1907 * be scheduled until later, and thus are tying up most of the free
1908 * packet pool for a very long time.
1910 * 1. schedule multi-packet calls if all the packets are present.
1911 * Probably CPU-bound operation, useful to return packets to pool.
1912 * Do what if there is a full window, but the last packet isn't here?
1913 * 3. preserve one thread which *only* runs "best" calls, otherwise
1914 * it sleeps and waits for that type of call.
1915 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1916 * the current dataquota business is badly broken. The quota isn't adjusted
1917 * to reflect how many packets are presently queued for a running call.
1918 * So, when we schedule a queued call with a full window of packets queued
1919 * up for it, that *should* free up a window full of packets for other 2d-class
1920 * calls to be able to use from the packet pool. But it doesn't.
1922 * NB. Most of the time, this code doesn't run -- since idle server threads
1923 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1924 * as a new call arrives.
1926 /* Sleep until a call arrives. Returns a pointer to the call, ready
1927 * for an rx_Read. */
1928 #ifdef RX_ENABLE_LOCKS
1930 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1932 struct rx_serverQueueEntry *sq;
1933 struct rx_call *call = (struct rx_call *)0;
1934 struct rx_service *service = NULL;
1936 MUTEX_ENTER(&freeSQEList_lock);
1938 if ((sq = rx_FreeSQEList)) {
1939 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1940 MUTEX_EXIT(&freeSQEList_lock);
1941 } else { /* otherwise allocate a new one and return that */
1942 MUTEX_EXIT(&freeSQEList_lock);
1943 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1944 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1945 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1948 MUTEX_ENTER(&rx_serverPool_lock);
1949 if (cur_service != NULL) {
1950 ReturnToServerPool(cur_service);
1953 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1954 struct rx_call *tcall, *ncall, *choice2 = NULL;
1956 /* Scan for eligible incoming calls. A call is not eligible
1957 * if the maximum number of calls for its service type are
1958 * already executing */
1959 /* One thread will process calls FCFS (to prevent starvation),
1960 * while the other threads may run ahead looking for calls which
1961 * have all their input data available immediately. This helps
1962 * keep threads from blocking, waiting for data from the client. */
1963 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1964 service = tcall->conn->service;
1965 if (!QuotaOK(service)) {
1968 MUTEX_ENTER(&rx_pthread_mutex);
1969 if (tno == rxi_fcfs_thread_num
1970 || !tcall->queue_item_header.next) {
1971 MUTEX_EXIT(&rx_pthread_mutex);
1972 /* If we're the fcfs thread , then we'll just use
1973 * this call. If we haven't been able to find an optimal
1974 * choice, and we're at the end of the list, then use a
1975 * 2d choice if one has been identified. Otherwise... */
1976 call = (choice2 ? choice2 : tcall);
1977 service = call->conn->service;
1979 MUTEX_EXIT(&rx_pthread_mutex);
1980 if (!queue_IsEmpty(&tcall->rq)) {
1981 struct rx_packet *rp;
1982 rp = queue_First(&tcall->rq, rx_packet);
1983 if (rp->header.seq == 1) {
1985 || (rp->header.flags & RX_LAST_PACKET)) {
1987 } else if (rxi_2dchoice && !choice2
1988 && !(tcall->flags & RX_CALL_CLEARED)
1989 && (tcall->rprev > rxi_HardAckRate)) {
1999 ReturnToServerPool(service);
2006 MUTEX_EXIT(&rx_serverPool_lock);
2007 MUTEX_ENTER(&call->lock);
2009 if (call->flags & RX_CALL_WAIT_PROC) {
2010 call->flags &= ~RX_CALL_WAIT_PROC;
2011 rx_atomic_dec(&rx_nWaiting);
2014 if (call->state != RX_STATE_PRECALL || call->error) {
2015 MUTEX_EXIT(&call->lock);
2016 MUTEX_ENTER(&rx_serverPool_lock);
2017 ReturnToServerPool(service);
2022 if (queue_IsEmpty(&call->rq)
2023 || queue_First(&call->rq, rx_packet)->header.seq != 1)
2024 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2026 CLEAR_CALL_QUEUE_LOCK(call);
2029 /* If there are no eligible incoming calls, add this process
2030 * to the idle server queue, to wait for one */
2034 *socketp = OSI_NULLSOCKET;
2036 sq->socketp = socketp;
2037 queue_Append(&rx_idleServerQueue, sq);
2038 #ifndef AFS_AIX41_ENV
2039 rx_waitForPacket = sq;
2041 rx_waitingForPacket = sq;
2042 #endif /* AFS_AIX41_ENV */
2044 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2046 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2047 MUTEX_EXIT(&rx_serverPool_lock);
2048 return (struct rx_call *)0;
2051 } while (!(call = sq->newcall)
2052 && !(socketp && *socketp != OSI_NULLSOCKET));
2053 MUTEX_EXIT(&rx_serverPool_lock);
2055 MUTEX_ENTER(&call->lock);
2061 MUTEX_ENTER(&freeSQEList_lock);
2062 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2063 rx_FreeSQEList = sq;
2064 MUTEX_EXIT(&freeSQEList_lock);
2067 clock_GetTime(&call->startTime);
2068 call->state = RX_STATE_ACTIVE;
2069 call->mode = RX_MODE_RECEIVING;
2070 #ifdef RX_KERNEL_TRACE
2071 if (ICL_SETACTIVE(afs_iclSetp)) {
2072 int glockOwner = ISAFS_GLOCK();
2075 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2076 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2083 rxi_calltrace(RX_CALL_START, call);
2084 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2085 call->conn->service->servicePort, call->conn->service->serviceId,
2088 MUTEX_EXIT(&call->lock);
2089 MUTEX_ENTER(&rx_refcnt_mutex);
2090 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2091 MUTEX_EXIT(&rx_refcnt_mutex);
2093 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2098 #else /* RX_ENABLE_LOCKS */
2100 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2102 struct rx_serverQueueEntry *sq;
2103 struct rx_call *call = (struct rx_call *)0, *choice2;
2104 struct rx_service *service = NULL;
2108 MUTEX_ENTER(&freeSQEList_lock);
2110 if ((sq = rx_FreeSQEList)) {
2111 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2112 MUTEX_EXIT(&freeSQEList_lock);
2113 } else { /* otherwise allocate a new one and return that */
2114 MUTEX_EXIT(&freeSQEList_lock);
2115 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2116 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2117 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2119 MUTEX_ENTER(&sq->lock);
2121 if (cur_service != NULL) {
2122 cur_service->nRequestsRunning--;
2123 MUTEX_ENTER(&rx_quota_mutex);
2124 if (cur_service->nRequestsRunning < cur_service->minProcs)
2127 MUTEX_EXIT(&rx_quota_mutex);
2129 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
2130 struct rx_call *tcall, *ncall;
2131 /* Scan for eligible incoming calls. A call is not eligible
2132 * if the maximum number of calls for its service type are
2133 * already executing */
2134 /* One thread will process calls FCFS (to prevent starvation),
2135 * while the other threads may run ahead looking for calls which
2136 * have all their input data available immediately. This helps
2137 * keep threads from blocking, waiting for data from the client. */
2138 choice2 = (struct rx_call *)0;
2139 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
2140 service = tcall->conn->service;
2141 if (QuotaOK(service)) {
2142 MUTEX_ENTER(&rx_pthread_mutex);
2143 if (tno == rxi_fcfs_thread_num
2144 || !tcall->queue_item_header.next) {
2145 MUTEX_EXIT(&rx_pthread_mutex);
2146 /* If we're the fcfs thread, then we'll just use
2147 * this call. If we haven't been able to find an optimal
2148 * choice, and we're at the end of the list, then use a
2149 * 2d choice if one has been identified. Otherwise... */
2150 call = (choice2 ? choice2 : tcall);
2151 service = call->conn->service;
2153 MUTEX_EXIT(&rx_pthread_mutex);
2154 if (!queue_IsEmpty(&tcall->rq)) {
2155 struct rx_packet *rp;
2156 rp = queue_First(&tcall->rq, rx_packet);
2157 if (rp->header.seq == 1
2159 || (rp->header.flags & RX_LAST_PACKET))) {
2161 } else if (rxi_2dchoice && !choice2
2162 && !(tcall->flags & RX_CALL_CLEARED)
2163 && (tcall->rprev > rxi_HardAckRate)) {
2177 /* we can't schedule a call if there's no data!!! */
2178 /* send an ack if there's no data, if we're missing the
2179 * first packet, or we're missing something between first
2180 * and last -- there's a "hole" in the incoming data. */
2181 if (queue_IsEmpty(&call->rq)
2182 || queue_First(&call->rq, rx_packet)->header.seq != 1
2183 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
2184 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2186 call->flags &= (~RX_CALL_WAIT_PROC);
2187 service->nRequestsRunning++;
2188 /* just started call in minProcs pool, need fewer to maintain
2190 MUTEX_ENTER(&rx_quota_mutex);
2191 if (service->nRequestsRunning <= service->minProcs)
2194 MUTEX_EXIT(&rx_quota_mutex);
2195 rx_atomic_dec(&rx_nWaiting);
2196 /* MUTEX_EXIT(&call->lock); */
2198 /* If there are no eligible incoming calls, add this process
2199 * to the idle server queue, to wait for one */
2202 *socketp = OSI_NULLSOCKET;
2204 sq->socketp = socketp;
2205 queue_Append(&rx_idleServerQueue, sq);
2209 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2211 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2212 return (struct rx_call *)0;
2215 } while (!(call = sq->newcall)
2216 && !(socketp && *socketp != OSI_NULLSOCKET));
2218 MUTEX_EXIT(&sq->lock);
2220 MUTEX_ENTER(&freeSQEList_lock);
2221 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2222 rx_FreeSQEList = sq;
2223 MUTEX_EXIT(&freeSQEList_lock);
2226 clock_GetTime(&call->startTime);
2227 call->state = RX_STATE_ACTIVE;
2228 call->mode = RX_MODE_RECEIVING;
2229 #ifdef RX_KERNEL_TRACE
2230 if (ICL_SETACTIVE(afs_iclSetp)) {
2231 int glockOwner = ISAFS_GLOCK();
2234 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2235 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2242 rxi_calltrace(RX_CALL_START, call);
2243 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2244 call->conn->service->servicePort, call->conn->service->serviceId,
2247 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2254 #endif /* RX_ENABLE_LOCKS */
2258 /* Establish a procedure to be called when a packet arrives for a
2259 * call. This routine will be called at most once after each call,
2260 * and will also be called if there is an error condition on the or
2261 * the call is complete. Used by multi rx to build a selection
2262 * function which determines which of several calls is likely to be a
2263 * good one to read from.
2264 * NOTE: the way this is currently implemented it is probably only a
2265 * good idea to (1) use it immediately after a newcall (clients only)
2266 * and (2) only use it once. Other uses currently void your warranty
2269 rx_SetArrivalProc(struct rx_call *call,
2270 void (*proc) (struct rx_call * call,
2273 void * handle, int arg)
2275 call->arrivalProc = proc;
2276 call->arrivalProcHandle = handle;
2277 call->arrivalProcArg = arg;
2280 /* Call is finished (possibly prematurely). Return rc to the peer, if
2281 * appropriate, and return the final error code from the conversation
2285 rx_EndCall(struct rx_call *call, afs_int32 rc)
2287 struct rx_connection *conn = call->conn;
2291 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2292 call, rc, call->error, call->abortCode));
2295 MUTEX_ENTER(&call->lock);
2297 if (rc == 0 && call->error == 0) {
2298 call->abortCode = 0;
2299 call->abortCount = 0;
2302 call->arrivalProc = (void (*)())0;
2303 if (rc && call->error == 0) {
2304 rxi_CallError(call, rc);
2305 call->mode = RX_MODE_ERROR;
2306 /* Send an abort message to the peer if this error code has
2307 * only just been set. If it was set previously, assume the
2308 * peer has already been sent the error code or will request it
2310 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2312 if (conn->type == RX_SERVER_CONNECTION) {
2313 /* Make sure reply or at least dummy reply is sent */
2314 if (call->mode == RX_MODE_RECEIVING) {
2315 MUTEX_EXIT(&call->lock);
2316 rxi_WriteProc(call, 0, 0);
2317 MUTEX_ENTER(&call->lock);
2319 if (call->mode == RX_MODE_SENDING) {
2320 MUTEX_EXIT(&call->lock);
2321 rxi_FlushWrite(call);
2322 MUTEX_ENTER(&call->lock);
2324 rxi_calltrace(RX_CALL_END, call);
2325 /* Call goes to hold state until reply packets are acknowledged */
2326 if (call->tfirst + call->nSoftAcked < call->tnext) {
2327 call->state = RX_STATE_HOLD;
2329 call->state = RX_STATE_DALLY;
2330 rxi_ClearTransmitQueue(call, 0);
2331 rxi_rto_cancel(call);
2332 rxevent_Cancel(call->keepAliveEvent, call,
2333 RX_CALL_REFCOUNT_ALIVE);
2335 } else { /* Client connection */
2337 /* Make sure server receives input packets, in the case where
2338 * no reply arguments are expected */
2339 if ((call->mode == RX_MODE_SENDING)
2340 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2341 MUTEX_EXIT(&call->lock);
2342 (void)rxi_ReadProc(call, &dummy, 1);
2343 MUTEX_ENTER(&call->lock);
2346 /* If we had an outstanding delayed ack, be nice to the server
2347 * and force-send it now.
2349 if (call->delayedAckEvent) {
2350 rxevent_Cancel(call->delayedAckEvent, call,
2351 RX_CALL_REFCOUNT_DELAY);
2352 call->delayedAckEvent = NULL;
2353 rxi_SendDelayedAck(NULL, call, NULL);
2356 /* We need to release the call lock since it's lower than the
2357 * conn_call_lock and we don't want to hold the conn_call_lock
2358 * over the rx_ReadProc call. The conn_call_lock needs to be held
2359 * here for the case where rx_NewCall is perusing the calls on
2360 * the connection structure. We don't want to signal until
2361 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2362 * have checked this call, found it active and by the time it
2363 * goes to sleep, will have missed the signal.
2365 MUTEX_EXIT(&call->lock);
2366 MUTEX_ENTER(&conn->conn_call_lock);
2367 MUTEX_ENTER(&call->lock);
2369 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2370 conn->lastBusy[call->channel] = 0;
2373 MUTEX_ENTER(&conn->conn_data_lock);
2374 conn->flags |= RX_CONN_BUSY;
2375 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2376 MUTEX_EXIT(&conn->conn_data_lock);
2377 #ifdef RX_ENABLE_LOCKS
2378 CV_BROADCAST(&conn->conn_call_cv);
2383 #ifdef RX_ENABLE_LOCKS
2385 MUTEX_EXIT(&conn->conn_data_lock);
2387 #endif /* RX_ENABLE_LOCKS */
2388 call->state = RX_STATE_DALLY;
2390 error = call->error;
2392 /* currentPacket, nLeft, and NFree must be zeroed here, because
2393 * ResetCall cannot: ResetCall may be called at splnet(), in the
2394 * kernel version, and may interrupt the macros rx_Read or
2395 * rx_Write, which run at normal priority for efficiency. */
2396 if (call->currentPacket) {
2397 #ifdef RX_TRACK_PACKETS
2398 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2400 rxi_FreePacket(call->currentPacket);
2401 call->currentPacket = (struct rx_packet *)0;
2404 call->nLeft = call->nFree = call->curlen = 0;
2406 /* Free any packets from the last call to ReadvProc/WritevProc */
2407 #ifdef RXDEBUG_PACKET
2409 #endif /* RXDEBUG_PACKET */
2410 rxi_FreePackets(0, &call->iovq);
2411 MUTEX_EXIT(&call->lock);
2413 MUTEX_ENTER(&rx_refcnt_mutex);
2414 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2415 MUTEX_EXIT(&rx_refcnt_mutex);
2416 if (conn->type == RX_CLIENT_CONNECTION) {
2417 MUTEX_ENTER(&conn->conn_data_lock);
2418 conn->flags &= ~RX_CONN_BUSY;
2419 MUTEX_EXIT(&conn->conn_data_lock);
2420 MUTEX_EXIT(&conn->conn_call_lock);
2424 * Map errors to the local host's errno.h format.
2426 error = ntoh_syserr_conv(error);
2430 #if !defined(KERNEL)
2432 /* Call this routine when shutting down a server or client (especially
2433 * clients). This will allow Rx to gracefully garbage collect server
2434 * connections, and reduce the number of retries that a server might
2435 * make to a dead client.
2436 * This is not quite right, since some calls may still be ongoing and
2437 * we can't lock them to destroy them. */
2441 struct rx_connection **conn_ptr, **conn_end;
2445 if (rxinit_status == 1) {
2447 return; /* Already shutdown. */
2449 rxi_DeleteCachedConnections();
2450 if (rx_connHashTable) {
2451 MUTEX_ENTER(&rx_connHashTable_lock);
2452 for (conn_ptr = &rx_connHashTable[0], conn_end =
2453 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2455 struct rx_connection *conn, *next;
2456 for (conn = *conn_ptr; conn; conn = next) {
2458 if (conn->type == RX_CLIENT_CONNECTION) {
2459 MUTEX_ENTER(&rx_refcnt_mutex);
2461 MUTEX_EXIT(&rx_refcnt_mutex);
2462 #ifdef RX_ENABLE_LOCKS
2463 rxi_DestroyConnectionNoLock(conn);
2464 #else /* RX_ENABLE_LOCKS */
2465 rxi_DestroyConnection(conn);
2466 #endif /* RX_ENABLE_LOCKS */
2470 #ifdef RX_ENABLE_LOCKS
2471 while (rx_connCleanup_list) {
2472 struct rx_connection *conn;
2473 conn = rx_connCleanup_list;
2474 rx_connCleanup_list = rx_connCleanup_list->next;
2475 MUTEX_EXIT(&rx_connHashTable_lock);
2476 rxi_CleanupConnection(conn);
2477 MUTEX_ENTER(&rx_connHashTable_lock);
2479 MUTEX_EXIT(&rx_connHashTable_lock);
2480 #endif /* RX_ENABLE_LOCKS */
2485 afs_winsockCleanup();
2493 /* if we wakeup packet waiter too often, can get in loop with two
2494 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2496 rxi_PacketsUnWait(void)
2498 if (!rx_waitingForPackets) {
2502 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2503 return; /* still over quota */
2506 rx_waitingForPackets = 0;
2507 #ifdef RX_ENABLE_LOCKS
2508 CV_BROADCAST(&rx_waitingForPackets_cv);
2510 osi_rxWakeup(&rx_waitingForPackets);
2516 /* ------------------Internal interfaces------------------------- */
2518 /* Return this process's service structure for the
2519 * specified socket and service */
2520 static struct rx_service *
2521 rxi_FindService(osi_socket socket, u_short serviceId)
2523 struct rx_service **sp;
2524 for (sp = &rx_services[0]; *sp; sp++) {
2525 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2531 #ifdef RXDEBUG_PACKET
2532 #ifdef KDUMP_RX_LOCK
2533 static struct rx_call_rx_lock *rx_allCallsp = 0;
2535 static struct rx_call *rx_allCallsp = 0;
2537 #endif /* RXDEBUG_PACKET */
2539 /* Allocate a call structure, for the indicated channel of the
2540 * supplied connection. The mode and state of the call must be set by
2541 * the caller. Returns the call with mutex locked. */
2542 static struct rx_call *
2543 rxi_NewCall(struct rx_connection *conn, int channel)
2545 struct rx_call *call;
2546 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2547 struct rx_call *cp; /* Call pointer temp */
2548 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2549 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2551 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2553 /* Grab an existing call structure, or allocate a new one.
2554 * Existing call structures are assumed to have been left reset by
2556 MUTEX_ENTER(&rx_freeCallQueue_lock);
2558 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2560 * EXCEPT that the TQ might not yet be cleared out.
2561 * Skip over those with in-use TQs.
2564 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2565 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2571 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2572 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2573 call = queue_First(&rx_freeCallQueue, rx_call);
2574 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2576 if (rx_stats_active)
2577 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2578 MUTEX_EXIT(&rx_freeCallQueue_lock);
2579 MUTEX_ENTER(&call->lock);
2580 CLEAR_CALL_QUEUE_LOCK(call);
2581 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2582 /* Now, if TQ wasn't cleared earlier, do it now. */
2583 rxi_WaitforTQBusy(call);
2584 if (call->flags & RX_CALL_TQ_CLEARME) {
2585 rxi_ClearTransmitQueue(call, 1);
2586 /*queue_Init(&call->tq);*/
2588 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2589 /* Bind the call to its connection structure */
2591 rxi_ResetCall(call, 1);
2594 call = rxi_Alloc(sizeof(struct rx_call));
2595 #ifdef RXDEBUG_PACKET
2596 call->allNextp = rx_allCallsp;
2597 rx_allCallsp = call;
2599 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2600 #else /* RXDEBUG_PACKET */
2601 rx_atomic_inc(&rx_stats.nCallStructs);
2602 #endif /* RXDEBUG_PACKET */
2604 MUTEX_EXIT(&rx_freeCallQueue_lock);
2605 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2606 MUTEX_ENTER(&call->lock);
2607 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2608 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2609 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2611 /* Initialize once-only items */
2612 queue_Init(&call->tq);
2613 queue_Init(&call->rq);
2614 queue_Init(&call->iovq);
2615 #ifdef RXDEBUG_PACKET
2616 call->rqc = call->tqc = call->iovqc = 0;
2617 #endif /* RXDEBUG_PACKET */
2618 /* Bind the call to its connection structure (prereq for reset) */
2620 rxi_ResetCall(call, 1);
2622 call->channel = channel;
2623 call->callNumber = &conn->callNumber[channel];
2624 call->rwind = conn->rwind[channel];
2625 call->twind = conn->twind[channel];
2626 /* Note that the next expected call number is retained (in
2627 * conn->callNumber[i]), even if we reallocate the call structure
2629 conn->call[channel] = call;
2630 /* if the channel's never been used (== 0), we should start at 1, otherwise
2631 * the call number is valid from the last time this channel was used */
2632 if (*call->callNumber == 0)
2633 *call->callNumber = 1;
2638 /* A call has been inactive long enough that so we can throw away
2639 * state, including the call structure, which is placed on the call
2642 * call->lock amd rx_refcnt_mutex are held upon entry.
2643 * haveCTLock is set when called from rxi_ReapConnections.
2646 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2648 int channel = call->channel;
2649 struct rx_connection *conn = call->conn;
2652 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2653 (*call->callNumber)++;
2655 * We are setting the state to RX_STATE_RESET to
2656 * ensure that no one else will attempt to use this
2657 * call once we drop the refcnt lock. We must drop
2658 * the refcnt lock before calling rxi_ResetCall
2659 * because it cannot be held across acquiring the
2660 * freepktQ lock. NewCall does the same.
2662 call->state = RX_STATE_RESET;
2663 MUTEX_EXIT(&rx_refcnt_mutex);
2664 rxi_ResetCall(call, 0);
2666 MUTEX_ENTER(&conn->conn_call_lock);
2667 if (call->conn->call[channel] == call)
2668 call->conn->call[channel] = 0;
2669 MUTEX_EXIT(&conn->conn_call_lock);
2671 MUTEX_ENTER(&rx_freeCallQueue_lock);
2672 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2673 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2674 /* A call may be free even though its transmit queue is still in use.
2675 * Since we search the call list from head to tail, put busy calls at
2676 * the head of the list, and idle calls at the tail.
2678 if (call->flags & RX_CALL_TQ_BUSY)
2679 queue_Prepend(&rx_freeCallQueue, call);
2681 queue_Append(&rx_freeCallQueue, call);
2682 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2683 queue_Append(&rx_freeCallQueue, call);
2684 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2685 if (rx_stats_active)
2686 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2687 MUTEX_EXIT(&rx_freeCallQueue_lock);
2689 /* Destroy the connection if it was previously slated for
2690 * destruction, i.e. the Rx client code previously called
2691 * rx_DestroyConnection (client connections), or
2692 * rxi_ReapConnections called the same routine (server
2693 * connections). Only do this, however, if there are no
2694 * outstanding calls. Note that for fine grain locking, there appears
2695 * to be a deadlock in that rxi_FreeCall has a call locked and
2696 * DestroyConnectionNoLock locks each call in the conn. But note a
2697 * few lines up where we have removed this call from the conn.
2698 * If someone else destroys a connection, they either have no
2699 * call lock held or are going through this section of code.
2701 MUTEX_ENTER(&conn->conn_data_lock);
2702 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2703 MUTEX_ENTER(&rx_refcnt_mutex);
2705 MUTEX_EXIT(&rx_refcnt_mutex);
2706 MUTEX_EXIT(&conn->conn_data_lock);
2707 #ifdef RX_ENABLE_LOCKS
2709 rxi_DestroyConnectionNoLock(conn);
2711 rxi_DestroyConnection(conn);
2712 #else /* RX_ENABLE_LOCKS */
2713 rxi_DestroyConnection(conn);
2714 #endif /* RX_ENABLE_LOCKS */
2716 MUTEX_EXIT(&conn->conn_data_lock);
2718 MUTEX_ENTER(&rx_refcnt_mutex);
2721 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2722 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2725 rxi_Alloc(size_t size)
2729 if (rx_stats_active) {
2730 rx_atomic_add(&rxi_Allocsize, (int) size);
2731 rx_atomic_inc(&rxi_Alloccnt);
2735 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2736 afs_osi_Alloc_NoSleep(size);
2741 osi_Panic("rxi_Alloc error");
2747 rxi_Free(void *addr, size_t size)
2749 if (rx_stats_active) {
2750 rx_atomic_sub(&rxi_Allocsize, (int) size);
2751 rx_atomic_dec(&rxi_Alloccnt);
2753 osi_Free(addr, size);
2757 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2759 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2760 struct rx_peer *next = NULL;
2764 MUTEX_ENTER(&rx_peerHashTable_lock);
2766 peer_ptr = &rx_peerHashTable[0];
2767 peer_end = &rx_peerHashTable[rx_hashTableSize];
2770 for ( ; peer_ptr < peer_end; peer_ptr++) {
2773 for ( ; peer; peer = next) {
2775 if (host == peer->host)
2780 hashIndex = PEER_HASH(host, port);
2781 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2782 if ((peer->host == host) && (peer->port == port))
2787 MUTEX_ENTER(&rx_peerHashTable_lock);
2792 MUTEX_EXIT(&rx_peerHashTable_lock);
2794 MUTEX_ENTER(&peer->peer_lock);
2795 /* We don't handle dropping below min, so don't */
2796 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2797 peer->ifMTU=MIN(mtu, peer->ifMTU);
2798 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2799 /* if we tweaked this down, need to tune our peer MTU too */
2800 peer->MTU = MIN(peer->MTU, peer->natMTU);
2801 /* if we discovered a sub-1500 mtu, degrade */
2802 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2803 peer->maxDgramPackets = 1;
2804 /* We no longer have valid peer packet information */
2805 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2806 peer->maxPacketSize = 0;
2807 MUTEX_EXIT(&peer->peer_lock);
2809 MUTEX_ENTER(&rx_peerHashTable_lock);
2811 if (host && !port) {
2813 /* pick up where we left off */
2817 MUTEX_EXIT(&rx_peerHashTable_lock);
2820 /* Find the peer process represented by the supplied (host,port)
2821 * combination. If there is no appropriate active peer structure, a
2822 * new one will be allocated and initialized
2823 * The origPeer, if set, is a pointer to a peer structure on which the
2824 * refcount will be be decremented. This is used to replace the peer
2825 * structure hanging off a connection structure */
2827 rxi_FindPeer(afs_uint32 host, u_short port,
2828 struct rx_peer *origPeer, int create)
2832 hashIndex = PEER_HASH(host, port);
2833 MUTEX_ENTER(&rx_peerHashTable_lock);
2834 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2835 if ((pp->host == host) && (pp->port == port))
2840 pp = rxi_AllocPeer(); /* This bzero's *pp */
2841 pp->host = host; /* set here or in InitPeerParams is zero */
2843 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2844 queue_Init(&pp->congestionQueue);
2845 queue_Init(&pp->rpcStats);
2846 pp->next = rx_peerHashTable[hashIndex];
2847 rx_peerHashTable[hashIndex] = pp;
2848 rxi_InitPeerParams(pp);
2849 if (rx_stats_active)
2850 rx_atomic_inc(&rx_stats.nPeerStructs);
2857 origPeer->refCount--;
2858 MUTEX_EXIT(&rx_peerHashTable_lock);
2863 /* Find the connection at (host, port) started at epoch, and with the
2864 * given connection id. Creates the server connection if necessary.
2865 * The type specifies whether a client connection or a server
2866 * connection is desired. In both cases, (host, port) specify the
2867 * peer's (host, pair) pair. Client connections are not made
2868 * automatically by this routine. The parameter socket gives the
2869 * socket descriptor on which the packet was received. This is used,
2870 * in the case of server connections, to check that *new* connections
2871 * come via a valid (port, serviceId). Finally, the securityIndex
2872 * parameter must match the existing index for the connection. If a
2873 * server connection is created, it will be created using the supplied
2874 * index, if the index is valid for this service */
2875 struct rx_connection *
2876 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2877 u_short port, u_short serviceId, afs_uint32 cid,
2878 afs_uint32 epoch, int type, u_int securityIndex)
2880 int hashindex, flag, i;
2881 struct rx_connection *conn;
2882 hashindex = CONN_HASH(host, port, cid, epoch, type);
2883 MUTEX_ENTER(&rx_connHashTable_lock);
2884 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2885 rx_connHashTable[hashindex],
2888 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2889 && (epoch == conn->epoch)) {
2890 struct rx_peer *pp = conn->peer;
2891 if (securityIndex != conn->securityIndex) {
2892 /* this isn't supposed to happen, but someone could forge a packet
2893 * like this, and there seems to be some CM bug that makes this
2894 * happen from time to time -- in which case, the fileserver
2896 MUTEX_EXIT(&rx_connHashTable_lock);
2897 return (struct rx_connection *)0;
2899 if (pp->host == host && pp->port == port)
2901 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2903 /* So what happens when it's a callback connection? */
2904 if ( /*type == RX_CLIENT_CONNECTION && */
2905 (conn->epoch & 0x80000000))
2909 /* the connection rxLastConn that was used the last time is not the
2910 ** one we are looking for now. Hence, start searching in the hash */
2912 conn = rx_connHashTable[hashindex];
2917 struct rx_service *service;
2918 if (type == RX_CLIENT_CONNECTION) {
2919 MUTEX_EXIT(&rx_connHashTable_lock);
2920 return (struct rx_connection *)0;
2922 service = rxi_FindService(socket, serviceId);
2923 if (!service || (securityIndex >= service->nSecurityObjects)
2924 || (service->securityObjects[securityIndex] == 0)) {
2925 MUTEX_EXIT(&rx_connHashTable_lock);
2926 return (struct rx_connection *)0;
2928 conn = rxi_AllocConnection(); /* This bzero's the connection */
2929 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2930 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2931 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2932 conn->next = rx_connHashTable[hashindex];
2933 rx_connHashTable[hashindex] = conn;
2934 conn->peer = rxi_FindPeer(host, port, 0, 1);
2935 conn->type = RX_SERVER_CONNECTION;
2936 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2937 conn->epoch = epoch;
2938 conn->cid = cid & RX_CIDMASK;
2939 /* conn->serial = conn->lastSerial = 0; */
2940 /* conn->timeout = 0; */
2941 conn->ackRate = RX_FAST_ACK_RATE;
2942 conn->service = service;
2943 conn->serviceId = serviceId;
2944 conn->securityIndex = securityIndex;
2945 conn->securityObject = service->securityObjects[securityIndex];
2946 conn->nSpecific = 0;
2947 conn->specific = NULL;
2948 rx_SetConnDeadTime(conn, service->connDeadTime);
2949 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2950 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2951 for (i = 0; i < RX_MAXCALLS; i++) {
2952 conn->twind[i] = rx_initSendWindow;
2953 conn->rwind[i] = rx_initReceiveWindow;
2955 /* Notify security object of the new connection */
2956 RXS_NewConnection(conn->securityObject, conn);
2957 /* XXXX Connection timeout? */
2958 if (service->newConnProc)
2959 (*service->newConnProc) (conn);
2960 if (rx_stats_active)
2961 rx_atomic_inc(&rx_stats.nServerConns);
2964 MUTEX_ENTER(&rx_refcnt_mutex);
2966 MUTEX_EXIT(&rx_refcnt_mutex);
2968 rxLastConn = conn; /* store this connection as the last conn used */
2969 MUTEX_EXIT(&rx_connHashTable_lock);
2974 * Timeout a call on a busy call channel if appropriate.
2976 * @param[in] call The busy call.
2978 * @pre 'call' is marked as busy (namely,
2979 * call->conn->lastBusy[call->channel] != 0)
2981 * @pre call->lock is held
2982 * @pre rxi_busyChannelError is nonzero
2984 * @note call->lock is dropped and reacquired
2987 rxi_CheckBusy(struct rx_call *call)
2989 struct rx_connection *conn = call->conn;
2990 int channel = call->channel;
2991 int freechannel = 0;
2993 afs_uint32 callNumber = *call->callNumber;
2995 MUTEX_EXIT(&call->lock);
2997 MUTEX_ENTER(&conn->conn_call_lock);
2999 /* Are there any other call slots on this conn that we should try? Look for
3000 * slots that are empty and are either non-busy, or were marked as busy
3001 * longer than conn->secondsUntilDead seconds before this call started. */
3003 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3005 /* only look at channels that aren't us */
3009 if (conn->lastBusy[i]) {
3010 /* if this channel looked busy too recently, don't look at it */
3011 if (conn->lastBusy[i] >= call->startTime.sec) {
3014 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3019 if (conn->call[i]) {
3020 struct rx_call *tcall = conn->call[i];
3021 MUTEX_ENTER(&tcall->lock);
3022 if (tcall->state == RX_STATE_DALLY) {
3025 MUTEX_EXIT(&tcall->lock);
3031 MUTEX_EXIT(&conn->conn_call_lock);
3033 MUTEX_ENTER(&call->lock);
3035 /* Since the call->lock and conn->conn_call_lock have been released it is
3036 * possible that (1) the call may no longer be busy and/or (2) the call may
3037 * have been reused by another waiting thread. Therefore, we must confirm
3038 * that the call state has not changed when deciding whether or not to
3039 * force this application thread to retry by forcing a Timeout error. */
3041 if (freechannel && *call->callNumber == callNumber &&
3042 (call->flags & RX_CALL_PEER_BUSY)) {
3043 /* Since 'freechannel' is set, there exists another channel in this
3044 * rx_conn that the application thread might be able to use. We know
3045 * that we have the correct call since callNumber is unchanged, and we
3046 * know that the call is still busy. So, set the call error state to
3047 * rxi_busyChannelError so the application can retry the request,
3048 * presumably on a less-busy call channel. */
3050 rxi_CallError(call, rxi_busyChannelError);
3054 /* There are two packet tracing routines available for testing and monitoring
3055 * Rx. One is called just after every packet is received and the other is
3056 * called just before every packet is sent. Received packets, have had their
3057 * headers decoded, and packets to be sent have not yet had their headers
3058 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3059 * containing the network address. Both can be modified. The return value, if
3060 * non-zero, indicates that the packet should be dropped. */
3062 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3063 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3065 /* A packet has been received off the interface. Np is the packet, socket is
3066 * the socket number it was received from (useful in determining which service
3067 * this packet corresponds to), and (host, port) reflect the host,port of the
3068 * sender. This call returns the packet to the caller if it is finished with
3069 * it, rather than de-allocating it, just as a small performance hack */
3072 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3073 afs_uint32 host, u_short port, int *tnop,
3074 struct rx_call **newcallp)
3076 struct rx_call *call;
3077 struct rx_connection *conn;
3079 afs_uint32 currentCallNumber;
3085 struct rx_packet *tnp;
3088 /* We don't print out the packet until now because (1) the time may not be
3089 * accurate enough until now in the lwp implementation (rx_Listener only gets
3090 * the time after the packet is read) and (2) from a protocol point of view,
3091 * this is the first time the packet has been seen */
3092 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3093 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3094 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3095 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3096 np->header.epoch, np->header.cid, np->header.callNumber,
3097 np->header.seq, np->header.flags, np));
3100 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3101 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3104 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3105 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3108 /* If an input tracer function is defined, call it with the packet and
3109 * network address. Note this function may modify its arguments. */
3110 if (rx_justReceived) {
3111 struct sockaddr_in addr;
3113 addr.sin_family = AF_INET;
3114 addr.sin_port = port;
3115 addr.sin_addr.s_addr = host;
3116 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3117 addr.sin_len = sizeof(addr);
3118 #endif /* AFS_OSF_ENV */
3119 drop = (*rx_justReceived) (np, &addr);
3120 /* drop packet if return value is non-zero */
3123 port = addr.sin_port; /* in case fcn changed addr */
3124 host = addr.sin_addr.s_addr;
3128 /* If packet was not sent by the client, then *we* must be the client */
3129 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3130 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3132 /* Find the connection (or fabricate one, if we're the server & if
3133 * necessary) associated with this packet */
3135 rxi_FindConnection(socket, host, port, np->header.serviceId,
3136 np->header.cid, np->header.epoch, type,
3137 np->header.securityIndex);
3140 /* If no connection found or fabricated, just ignore the packet.
3141 * (An argument could be made for sending an abort packet for
3146 /* If the connection is in an error state, send an abort packet and ignore
3147 * the incoming packet */
3149 /* Don't respond to an abort packet--we don't want loops! */
3150 MUTEX_ENTER(&conn->conn_data_lock);
3151 if (np->header.type != RX_PACKET_TYPE_ABORT)
3152 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3153 MUTEX_ENTER(&rx_refcnt_mutex);
3155 MUTEX_EXIT(&rx_refcnt_mutex);
3156 MUTEX_EXIT(&conn->conn_data_lock);
3160 /* Check for connection-only requests (i.e. not call specific). */
3161 if (np->header.callNumber == 0) {
3162 switch (np->header.type) {
3163 case RX_PACKET_TYPE_ABORT: {
3164 /* What if the supplied error is zero? */
3165 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3166 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3167 rxi_ConnectionError(conn, errcode);
3168 MUTEX_ENTER(&rx_refcnt_mutex);
3170 MUTEX_EXIT(&rx_refcnt_mutex);
3173 case RX_PACKET_TYPE_CHALLENGE:
3174 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3175 MUTEX_ENTER(&rx_refcnt_mutex);
3177 MUTEX_EXIT(&rx_refcnt_mutex);
3179 case RX_PACKET_TYPE_RESPONSE:
3180 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3181 MUTEX_ENTER(&rx_refcnt_mutex);
3183 MUTEX_EXIT(&rx_refcnt_mutex);
3185 case RX_PACKET_TYPE_PARAMS:
3186 case RX_PACKET_TYPE_PARAMS + 1:
3187 case RX_PACKET_TYPE_PARAMS + 2:
3188 /* ignore these packet types for now */
3189 MUTEX_ENTER(&rx_refcnt_mutex);
3191 MUTEX_EXIT(&rx_refcnt_mutex);
3196 /* Should not reach here, unless the peer is broken: send an
3198 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3199 MUTEX_ENTER(&conn->conn_data_lock);
3200 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3201 MUTEX_ENTER(&rx_refcnt_mutex);
3203 MUTEX_EXIT(&rx_refcnt_mutex);
3204 MUTEX_EXIT(&conn->conn_data_lock);
3209 channel = np->header.cid & RX_CHANNELMASK;
3210 call = conn->call[channel];
3213 MUTEX_ENTER(&call->lock);
3214 currentCallNumber = conn->callNumber[channel];
3215 } else if (type == RX_SERVER_CONNECTION) { /* No call allocated */
3216 MUTEX_ENTER(&conn->conn_call_lock);
3217 call = conn->call[channel];
3219 MUTEX_ENTER(&call->lock);
3220 MUTEX_EXIT(&conn->conn_call_lock);
3221 currentCallNumber = conn->callNumber[channel];
3223 call = rxi_NewCall(conn, channel); /* returns locked call */
3224 MUTEX_EXIT(&conn->conn_call_lock);
3225 *call->callNumber = currentCallNumber = np->header.callNumber;
3227 if (np->header.callNumber == 0)
3228 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3229 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3230 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3231 np->header.flags, np, np->length));
3233 call->state = RX_STATE_PRECALL;
3234 clock_GetTime(&call->queueTime);
3235 hzero(call->bytesSent);
3236 hzero(call->bytesRcvd);
3238 * If the number of queued calls exceeds the overload
3239 * threshold then abort this call.
3241 if ((rx_BusyThreshold > 0) &&
3242 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3243 struct rx_packet *tp;
3245 rxi_CallError(call, rx_BusyError);
3246 tp = rxi_SendCallAbort(call, np, 1, 0);
3247 MUTEX_EXIT(&call->lock);
3248 MUTEX_ENTER(&rx_refcnt_mutex);
3250 MUTEX_EXIT(&rx_refcnt_mutex);
3251 if (rx_stats_active)
3252 rx_atomic_inc(&rx_stats.nBusies);
3255 rxi_KeepAliveOn(call);
3257 } else { /* RX_CLIENT_CONNECTION and No call allocated */
3258 /* This packet can't be for this call. If the new call address is
3259 * 0 then no call is running on this channel. If there is a call
3260 * then, since this is a client connection we're getting data for
3261 * it must be for the previous call.
3263 if (rx_stats_active)
3264 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3265 MUTEX_ENTER(&rx_refcnt_mutex);
3267 MUTEX_EXIT(&rx_refcnt_mutex);
3271 /* There is a non-NULL locked call at this point */
3272 if (type == RX_SERVER_CONNECTION) { /* We're the server */
3273 if (np->header.callNumber < currentCallNumber) {
3274 MUTEX_EXIT(&call->lock);
3275 if (rx_stats_active)
3276 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3277 MUTEX_ENTER(&rx_refcnt_mutex);
3279 MUTEX_EXIT(&rx_refcnt_mutex);
3281 } else if (np->header.callNumber != currentCallNumber) {
3282 /* Wait until the transmit queue is idle before deciding
3283 * whether to reset the current call. Chances are that the
3284 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3287 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3288 if (call->state == RX_STATE_ACTIVE) {
3289 rxi_WaitforTQBusy(call);
3291 * If we entered error state while waiting,
3292 * must call rxi_CallError to permit rxi_ResetCall
3293 * to processed when the tqWaiter count hits zero.
3296 rxi_CallError(call, call->error);
3297 MUTEX_EXIT(&call->lock);
3298 MUTEX_ENTER(&rx_refcnt_mutex);
3300 MUTEX_EXIT(&rx_refcnt_mutex);
3304 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3305 /* If the new call cannot be taken right now send a busy and set
3306 * the error condition in this call, so that it terminates as
3307 * quickly as possible */
3308 if (call->state == RX_STATE_ACTIVE) {
3309 struct rx_packet *tp;
3311 rxi_CallError(call, RX_CALL_DEAD);
3312 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3314 MUTEX_EXIT(&call->lock);
3315 MUTEX_ENTER(&rx_refcnt_mutex);
3317 MUTEX_EXIT(&rx_refcnt_mutex);
3320 rxi_ResetCall(call, 0);
3321 *call->callNumber = np->header.callNumber;
3323 if (np->header.callNumber == 0)
3324 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3325 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3326 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3327 np->header.flags, np, np->length));
3329 call->state = RX_STATE_PRECALL;
3330 clock_GetTime(&call->queueTime);
3331 hzero(call->bytesSent);
3332 hzero(call->bytesRcvd);
3334 * If the number of queued calls exceeds the overload
3335 * threshold then abort this call.
3337 if ((rx_BusyThreshold > 0) &&
3338 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3339 struct rx_packet *tp;
3341 rxi_CallError(call, rx_BusyError);
3342 tp = rxi_SendCallAbort(call, np, 1, 0);
3343 MUTEX_EXIT(&call->lock);
3344 MUTEX_ENTER(&rx_refcnt_mutex);
3346 MUTEX_EXIT(&rx_refcnt_mutex);
3347 if (rx_stats_active)
3348 rx_atomic_inc(&rx_stats.nBusies);
3351 rxi_KeepAliveOn(call);
3353 /* Continuing call; do nothing here. */
3355 } else { /* we're the client */
3356 /* Ignore all incoming acknowledgements for calls in DALLY state */
3357 if ((call->state == RX_STATE_DALLY)
3358 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3359 if (rx_stats_active)
3360 rx_atomic_inc(&rx_stats.ignorePacketDally);
3361 MUTEX_EXIT(&call->lock);
3362 MUTEX_ENTER(&rx_refcnt_mutex);
3364 MUTEX_EXIT(&rx_refcnt_mutex);
3368 /* Ignore anything that's not relevant to the current call. If there
3369 * isn't a current call, then no packet is relevant. */
3370 if (np->header.callNumber != currentCallNumber) {
3371 if (rx_stats_active)
3372 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3373 MUTEX_EXIT(&call->lock);
3374 MUTEX_ENTER(&rx_refcnt_mutex);
3376 MUTEX_EXIT(&rx_refcnt_mutex);
3379 /* If the service security object index stamped in the packet does not
3380 * match the connection's security index, ignore the packet */
3381 if (np->header.securityIndex != conn->securityIndex) {
3382 MUTEX_EXIT(&call->lock);
3383 MUTEX_ENTER(&rx_refcnt_mutex);
3385 MUTEX_EXIT(&rx_refcnt_mutex);
3389 /* If we're receiving the response, then all transmit packets are
3390 * implicitly acknowledged. Get rid of them. */
3391 if (np->header.type == RX_PACKET_TYPE_DATA) {
3392 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3393 /* XXX Hack. Because we must release the global rx lock when
3394 * sending packets (osi_NetSend) we drop all acks while we're
3395 * traversing the tq in rxi_Start sending packets out because
3396 * packets may move to the freePacketQueue as result of being here!
3397 * So we drop these packets until we're safely out of the
3398 * traversing. Really ugly!
3399 * For fine grain RX locking, we set the acked field in the
3400 * packets and let rxi_Start remove them from the transmit queue.
3402 if (call->flags & RX_CALL_TQ_BUSY) {
3403 #ifdef RX_ENABLE_LOCKS
3404 rxi_SetAcksInTransmitQueue(call);
3406 MUTEX_ENTER(&rx_refcnt_mutex);
3408 MUTEX_EXIT(&rx_refcnt_mutex);
3409 return np; /* xmitting; drop packet */
3412 rxi_ClearTransmitQueue(call, 0);
3414 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3415 rxi_ClearTransmitQueue(call, 0);
3416 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3418 if (np->header.type == RX_PACKET_TYPE_ACK) {
3419 /* now check to see if this is an ack packet acknowledging that the
3420 * server actually *lost* some hard-acked data. If this happens we
3421 * ignore this packet, as it may indicate that the server restarted in
3422 * the middle of a call. It is also possible that this is an old ack
3423 * packet. We don't abort the connection in this case, because this
3424 * *might* just be an old ack packet. The right way to detect a server
3425 * restart in the midst of a call is to notice that the server epoch
3427 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3428 * XXX unacknowledged. I think that this is off-by-one, but
3429 * XXX I don't dare change it just yet, since it will
3430 * XXX interact badly with the server-restart detection
3431 * XXX code in receiveackpacket. */
3432 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3433 if (rx_stats_active)
3434 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3435 MUTEX_EXIT(&call->lock);
3436 MUTEX_ENTER(&rx_refcnt_mutex);
3438 MUTEX_EXIT(&rx_refcnt_mutex);
3442 } /* else not a data packet */
3445 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3446 /* Set remote user defined status from packet */
3447 call->remoteStatus = np->header.userStatus;
3449 /* Note the gap between the expected next packet and the actual
3450 * packet that arrived, when the new packet has a smaller serial number
3451 * than expected. Rioses frequently reorder packets all by themselves,
3452 * so this will be quite important with very large window sizes.
3453 * Skew is checked against 0 here to avoid any dependence on the type of
3454 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3456 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3457 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3458 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3460 MUTEX_ENTER(&conn->conn_data_lock);
3461 skew = conn->lastSerial - np->header.serial;
3462 conn->lastSerial = np->header.serial;
3463 MUTEX_EXIT(&conn->conn_data_lock);
3465 struct rx_peer *peer;
3467 if (skew > peer->inPacketSkew) {
3468 dpf(("*** In skew changed from %d to %d\n",
3469 peer->inPacketSkew, skew));
3470 peer->inPacketSkew = skew;
3474 /* Now do packet type-specific processing */
3475 switch (np->header.type) {
3476 case RX_PACKET_TYPE_DATA:
3477 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3480 case RX_PACKET_TYPE_ACK:
3481 /* Respond immediately to ack packets requesting acknowledgement
3483 if (np->header.flags & RX_REQUEST_ACK) {
3485 (void)rxi_SendCallAbort(call, 0, 1, 0);
3487 (void)rxi_SendAck(call, 0, np->header.serial,
3488 RX_ACK_PING_RESPONSE, 1);
3490 np = rxi_ReceiveAckPacket(call, np, 1);
3492 case RX_PACKET_TYPE_ABORT: {
3493 /* An abort packet: reset the call, passing the error up to the user. */
3494 /* What if error is zero? */
3495 /* What if the error is -1? the application will treat it as a timeout. */
3496 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3497 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3498 rxi_CallError(call, errdata);
3499 MUTEX_EXIT(&call->lock);
3500 MUTEX_ENTER(&rx_refcnt_mutex);
3502 MUTEX_EXIT(&rx_refcnt_mutex);
3503 return np; /* xmitting; drop packet */
3505 case RX_PACKET_TYPE_BUSY: {
3506 struct clock busyTime;
3508 clock_GetTime(&busyTime);
3510 MUTEX_EXIT(&call->lock);
3512 MUTEX_ENTER(&conn->conn_call_lock);
3513 MUTEX_ENTER(&call->lock);
3514 conn->lastBusy[call->channel] = busyTime.sec;
3515 call->flags |= RX_CALL_PEER_BUSY;
3516 MUTEX_EXIT(&call->lock);
3517 MUTEX_EXIT(&conn->conn_call_lock);
3519 MUTEX_ENTER(&rx_refcnt_mutex);
3521 MUTEX_EXIT(&rx_refcnt_mutex);
3525 case RX_PACKET_TYPE_ACKALL:
3526 /* All packets acknowledged, so we can drop all packets previously
3527 * readied for sending */
3528 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3529 /* XXX Hack. We because we can't release the global rx lock when
3530 * sending packets (osi_NetSend) we drop all ack pkts while we're
3531 * traversing the tq in rxi_Start sending packets out because
3532 * packets may move to the freePacketQueue as result of being
3533 * here! So we drop these packets until we're safely out of the
3534 * traversing. Really ugly!
3535 * For fine grain RX locking, we set the acked field in the packets
3536 * and let rxi_Start remove the packets from the transmit queue.
3538 if (call->flags & RX_CALL_TQ_BUSY) {
3539 #ifdef RX_ENABLE_LOCKS
3540 rxi_SetAcksInTransmitQueue(call);
3542 #else /* RX_ENABLE_LOCKS */
3543 MUTEX_EXIT(&call->lock);
3544 MUTEX_ENTER(&rx_refcnt_mutex);
3546 MUTEX_EXIT(&rx_refcnt_mutex);
3547 return np; /* xmitting; drop packet */
3548 #endif /* RX_ENABLE_LOCKS */
3550 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3551 rxi_ClearTransmitQueue(call, 0);
3554 /* Should not reach here, unless the peer is broken: send an abort
3556 rxi_CallError(call, RX_PROTOCOL_ERROR);
3557 np = rxi_SendCallAbort(call, np, 1, 0);
3560 /* Note when this last legitimate packet was received, for keep-alive
3561 * processing. Note, we delay getting the time until now in the hope that
3562 * the packet will be delivered to the user before any get time is required
3563 * (if not, then the time won't actually be re-evaluated here). */
3564 call->lastReceiveTime = clock_Sec();
3565 /* we've received a legit packet, so the channel is not busy */
3566 call->flags &= ~RX_CALL_PEER_BUSY;
3567 MUTEX_EXIT(&call->lock);
3568 MUTEX_ENTER(&rx_refcnt_mutex);
3570 MUTEX_EXIT(&rx_refcnt_mutex);
3574 /* return true if this is an "interesting" connection from the point of view
3575 of someone trying to debug the system */
3577 rxi_IsConnInteresting(struct rx_connection *aconn)
3580 struct rx_call *tcall;
3582 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3585 for (i = 0; i < RX_MAXCALLS; i++) {
3586 tcall = aconn->call[i];
3588 if ((tcall->state == RX_STATE_PRECALL)
3589 || (tcall->state == RX_STATE_ACTIVE))
3591 if ((tcall->mode == RX_MODE_SENDING)
3592 || (tcall->mode == RX_MODE_RECEIVING))
3600 /* if this is one of the last few packets AND it wouldn't be used by the
3601 receiving call to immediately satisfy a read request, then drop it on
3602 the floor, since accepting it might prevent a lock-holding thread from
3603 making progress in its reading. If a call has been cleared while in
3604 the precall state then ignore all subsequent packets until the call
3605 is assigned to a thread. */
3608 TooLow(struct rx_packet *ap, struct rx_call *acall)
3612 MUTEX_ENTER(&rx_quota_mutex);
3613 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3614 && (acall->state == RX_STATE_PRECALL))
3615 || ((rx_nFreePackets < rxi_dataQuota + 2)
3616 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3617 && (acall->flags & RX_CALL_READER_WAIT)))) {
3620 MUTEX_EXIT(&rx_quota_mutex);
3626 * Clear the attach wait flag on a connection and proceed.
3628 * Any processing waiting for a connection to be attached should be
3629 * unblocked. We clear the flag and do any other needed tasks.
3632 * the conn to unmark waiting for attach
3634 * @pre conn's conn_data_lock must be locked before calling this function
3638 rxi_ConnClearAttachWait(struct rx_connection *conn)
3640 /* Indicate that rxi_CheckReachEvent is no longer running by
3641 * clearing the flag. Must be atomic under conn_data_lock to
3642 * avoid a new call slipping by: rxi_CheckConnReach holds
3643 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3645 conn->flags &= ~RX_CONN_ATTACHWAIT;
3646 if (conn->flags & RX_CONN_NAT_PING) {
3647 conn->flags &= ~RX_CONN_NAT_PING;
3648 rxi_ScheduleNatKeepAliveEvent(conn);
3653 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3655 struct rx_connection *conn = arg1;
3656 struct rx_call *acall = arg2;
3657 struct rx_call *call = acall;
3658 struct clock when, now;
3661 MUTEX_ENTER(&conn->conn_data_lock);
3662 conn->checkReachEvent = NULL;
3663 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3665 MUTEX_ENTER(&rx_refcnt_mutex);
3667 MUTEX_EXIT(&rx_refcnt_mutex);
3669 MUTEX_EXIT(&conn->conn_data_lock);
3673 MUTEX_ENTER(&conn->conn_call_lock);
3674 MUTEX_ENTER(&conn->conn_data_lock);
3675 for (i = 0; i < RX_MAXCALLS; i++) {
3676 struct rx_call *tc = conn->call[i];
3677 if (tc && tc->state == RX_STATE_PRECALL) {
3683 rxi_ConnClearAttachWait(conn);
3684 MUTEX_EXIT(&conn->conn_data_lock);
3685 MUTEX_EXIT(&conn->conn_call_lock);
3690 MUTEX_ENTER(&call->lock);
3691 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3693 MUTEX_EXIT(&call->lock);
3695 clock_GetTime(&now);
3697 when.sec += RX_CHECKREACH_TIMEOUT;
3698 MUTEX_ENTER(&conn->conn_data_lock);
3699 if (!conn->checkReachEvent) {
3700 MUTEX_ENTER(&rx_refcnt_mutex);
3702 MUTEX_EXIT(&rx_refcnt_mutex);
3703 conn->checkReachEvent =
3704 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3707 MUTEX_EXIT(&conn->conn_data_lock);
3713 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3715 struct rx_service *service = conn->service;
3716 struct rx_peer *peer = conn->peer;
3717 afs_uint32 now, lastReach;
3719 if (service->checkReach == 0)
3723 MUTEX_ENTER(&peer->peer_lock);
3724 lastReach = peer->lastReachTime;
3725 MUTEX_EXIT(&peer->peer_lock);
3726 if (now - lastReach < RX_CHECKREACH_TTL)
3729 MUTEX_ENTER(&conn->conn_data_lock);
3730 if (conn->flags & RX_CONN_ATTACHWAIT) {
3731 MUTEX_EXIT(&conn->conn_data_lock);
3734 conn->flags |= RX_CONN_ATTACHWAIT;
3735 MUTEX_EXIT(&conn->conn_data_lock);
3736 if (!conn->checkReachEvent)
3737 rxi_CheckReachEvent(NULL, conn, call);
3742 /* try to attach call, if authentication is complete */
3744 TryAttach(struct rx_call *acall, osi_socket socket,
3745 int *tnop, struct rx_call **newcallp,
3748 struct rx_connection *conn = acall->conn;
3750 if (conn->type == RX_SERVER_CONNECTION
3751 && acall->state == RX_STATE_PRECALL) {
3752 /* Don't attach until we have any req'd. authentication. */
3753 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3754 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3755 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3756 /* Note: this does not necessarily succeed; there
3757 * may not any proc available
3760 rxi_ChallengeOn(acall->conn);
3765 /* A data packet has been received off the interface. This packet is
3766 * appropriate to the call (the call is in the right state, etc.). This
3767 * routine can return a packet to the caller, for re-use */
3770 rxi_ReceiveDataPacket(struct rx_call *call,
3771 struct rx_packet *np, int istack,
3772 osi_socket socket, afs_uint32 host, u_short port,
3773 int *tnop, struct rx_call **newcallp)
3775 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3780 afs_uint32 serial=0, flags=0;
3782 struct rx_packet *tnp;
3783 struct clock when, now;
3784 if (rx_stats_active)
3785 rx_atomic_inc(&rx_stats.dataPacketsRead);
3788 /* If there are no packet buffers, drop this new packet, unless we can find
3789 * packet buffers from inactive calls */
3791 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3792 MUTEX_ENTER(&rx_freePktQ_lock);
3793 rxi_NeedMorePackets = TRUE;
3794 MUTEX_EXIT(&rx_freePktQ_lock);
3795 if (rx_stats_active)
3796 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3797 call->rprev = np->header.serial;
3798 rxi_calltrace(RX_TRACE_DROP, call);
3799 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3801 rxi_ClearReceiveQueue(call);
3802 clock_GetTime(&now);
3804 clock_Add(&when, &rx_softAckDelay);
3805 if (!call->delayedAckEvent
3806 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3807 rxevent_Cancel(call->delayedAckEvent, call,
3808 RX_CALL_REFCOUNT_DELAY);
3809 MUTEX_ENTER(&rx_refcnt_mutex);
3810 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3811 MUTEX_EXIT(&rx_refcnt_mutex);
3813 call->delayedAckEvent =
3814 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3816 /* we've damaged this call already, might as well do it in. */
3822 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3823 * packet is one of several packets transmitted as a single
3824 * datagram. Do not send any soft or hard acks until all packets
3825 * in a jumbogram have been processed. Send negative acks right away.
3827 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3828 /* tnp is non-null when there are more packets in the
3829 * current jumbo gram */
3836 seq = np->header.seq;
3837 serial = np->header.serial;
3838 flags = np->header.flags;
3840 /* If the call is in an error state, send an abort message */
3842 return rxi_SendCallAbort(call, np, istack, 0);
3844 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3845 * AFS 3.5 jumbogram. */
3846 if (flags & RX_JUMBO_PACKET) {
3847 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3852 if (np->header.spare != 0) {
3853 MUTEX_ENTER(&call->conn->conn_data_lock);
3854 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3855 MUTEX_EXIT(&call->conn->conn_data_lock);
3858 /* The usual case is that this is the expected next packet */
3859 if (seq == call->rnext) {
3861 /* Check to make sure it is not a duplicate of one already queued */
3862 if (queue_IsNotEmpty(&call->rq)
3863 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3864 if (rx_stats_active)
3865 rx_atomic_inc(&rx_stats.dupPacketsRead);
3866 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3867 rxevent_Cancel(call->delayedAckEvent, call,
3868 RX_CALL_REFCOUNT_DELAY);
3869 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3875 /* It's the next packet. Stick it on the receive queue
3876 * for this call. Set newPackets to make sure we wake
3877 * the reader once all packets have been processed */
3878 #ifdef RX_TRACK_PACKETS
3879 np->flags |= RX_PKTFLAG_RQ;
3881 queue_Prepend(&call->rq, np);
3882 #ifdef RXDEBUG_PACKET
3884 #endif /* RXDEBUG_PACKET */
3886 np = NULL; /* We can't use this anymore */
3889 /* If an ack is requested then set a flag to make sure we
3890 * send an acknowledgement for this packet */
3891 if (flags & RX_REQUEST_ACK) {
3892 ackNeeded = RX_ACK_REQUESTED;
3895 /* Keep track of whether we have received the last packet */
3896 if (flags & RX_LAST_PACKET) {
3897 call->flags |= RX_CALL_HAVE_LAST;
3901 /* Check whether we have all of the packets for this call */
3902 if (call->flags & RX_CALL_HAVE_LAST) {
3903 afs_uint32 tseq; /* temporary sequence number */
3904 struct rx_packet *tp; /* Temporary packet pointer */
3905 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3907 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3908 if (tseq != tp->header.seq)
3910 if (tp->header.flags & RX_LAST_PACKET) {
3911 call->flags |= RX_CALL_RECEIVE_DONE;
3918 /* Provide asynchronous notification for those who want it
3919 * (e.g. multi rx) */
3920 if (call->arrivalProc) {
3921 (*call->arrivalProc) (call, call->arrivalProcHandle,
3922 call->arrivalProcArg);
3923 call->arrivalProc = (void (*)())0;
3926 /* Update last packet received */
3929 /* If there is no server process serving this call, grab
3930 * one, if available. We only need to do this once. If a
3931 * server thread is available, this thread becomes a server
3932 * thread and the server thread becomes a listener thread. */
3934 TryAttach(call, socket, tnop, newcallp, 0);
3937 /* This is not the expected next packet. */
3939 /* Determine whether this is a new or old packet, and if it's
3940 * a new one, whether it fits into the current receive window.
3941 * Also figure out whether the packet was delivered in sequence.
3942 * We use the prev variable to determine whether the new packet
3943 * is the successor of its immediate predecessor in the
3944 * receive queue, and the missing flag to determine whether
3945 * any of this packets predecessors are missing. */
3947 afs_uint32 prev; /* "Previous packet" sequence number */
3948 struct rx_packet *tp; /* Temporary packet pointer */
3949 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3950 int missing; /* Are any predecessors missing? */
3952 /* If the new packet's sequence number has been sent to the
3953 * application already, then this is a duplicate */
3954 if (seq < call->rnext) {
3955 if (rx_stats_active)
3956 rx_atomic_inc(&rx_stats.dupPacketsRead);
3957 rxevent_Cancel(call->delayedAckEvent, call,
3958 RX_CALL_REFCOUNT_DELAY);
3959 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3965 /* If the sequence number is greater than what can be
3966 * accomodated by the current window, then send a negative
3967 * acknowledge and drop the packet */
3968 if ((call->rnext + call->rwind) <= seq) {
3969 rxevent_Cancel(call->delayedAckEvent, call,
3970 RX_CALL_REFCOUNT_DELAY);
3971 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3978 /* Look for the packet in the queue of old received packets */
3979 for (prev = call->rnext - 1, missing =
3980 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3981 /*Check for duplicate packet */
3982 if (seq == tp->header.seq) {
3983 if (rx_stats_active)
3984 rx_atomic_inc(&rx_stats.dupPacketsRead);
3985 rxevent_Cancel(call->delayedAckEvent, call,
3986 RX_CALL_REFCOUNT_DELAY);
3987 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3993 /* If we find a higher sequence packet, break out and
3994 * insert the new packet here. */
3995 if (seq < tp->header.seq)
3997 /* Check for missing packet */
3998 if (tp->header.seq != prev + 1) {
4002 prev = tp->header.seq;
4005 /* Keep track of whether we have received the last packet. */
4006 if (flags & RX_LAST_PACKET) {
4007 call->flags |= RX_CALL_HAVE_LAST;
4010 /* It's within the window: add it to the the receive queue.
4011 * tp is left by the previous loop either pointing at the
4012 * packet before which to insert the new packet, or at the
4013 * queue head if the queue is empty or the packet should be
4015 #ifdef RX_TRACK_PACKETS
4016 np->flags |= RX_PKTFLAG_RQ;
4018 #ifdef RXDEBUG_PACKET
4020 #endif /* RXDEBUG_PACKET */
4021 queue_InsertBefore(tp, np);
4025 /* Check whether we have all of the packets for this call */
4026 if ((call->flags & RX_CALL_HAVE_LAST)
4027 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4028 afs_uint32 tseq; /* temporary sequence number */
4031 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
4032 if (tseq != tp->header.seq)
4034 if (tp->header.flags & RX_LAST_PACKET) {
4035 call->flags |= RX_CALL_RECEIVE_DONE;
4042 /* We need to send an ack of the packet is out of sequence,
4043 * or if an ack was requested by the peer. */
4044 if (seq != prev + 1 || missing) {
4045 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4046 } else if (flags & RX_REQUEST_ACK) {
4047 ackNeeded = RX_ACK_REQUESTED;
4050 /* Acknowledge the last packet for each call */
4051 if (flags & RX_LAST_PACKET) {
4062 * If the receiver is waiting for an iovec, fill the iovec
4063 * using the data from the receive queue */
4064 if (call->flags & RX_CALL_IOVEC_WAIT) {
4065 didHardAck = rxi_FillReadVec(call, serial);
4066 /* the call may have been aborted */
4075 /* Wakeup the reader if any */
4076 if ((call->flags & RX_CALL_READER_WAIT)
4077 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4078 || (call->iovNext >= call->iovMax)
4079 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4080 call->flags &= ~RX_CALL_READER_WAIT;
4081 #ifdef RX_ENABLE_LOCKS
4082 CV_BROADCAST(&call->cv_rq);
4084 osi_rxWakeup(&call->rq);
4090 * Send an ack when requested by the peer, or once every
4091 * rxi_SoftAckRate packets until the last packet has been
4092 * received. Always send a soft ack for the last packet in
4093 * the server's reply. */
4095 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4096 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4097 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4098 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4099 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4100 } else if (call->nSoftAcks) {
4101 clock_GetTime(&now);
4103 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
4104 clock_Add(&when, &rx_lastAckDelay);
4106 clock_Add(&when, &rx_softAckDelay);
4108 if (!call->delayedAckEvent
4109 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
4110 rxevent_Cancel(call->delayedAckEvent, call,
4111 RX_CALL_REFCOUNT_DELAY);
4112 MUTEX_ENTER(&rx_refcnt_mutex);
4113 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
4114 MUTEX_EXIT(&rx_refcnt_mutex);
4115 call->delayedAckEvent =
4116 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
4118 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4119 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4126 static void rxi_ComputeRate();
4130 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4132 struct rx_peer *peer = conn->peer;
4134 MUTEX_ENTER(&peer->peer_lock);
4135 peer->lastReachTime = clock_Sec();
4136 MUTEX_EXIT(&peer->peer_lock);
4138 MUTEX_ENTER(&conn->conn_data_lock);
4139 if (conn->flags & RX_CONN_ATTACHWAIT) {
4142 rxi_ConnClearAttachWait(conn);
4143 MUTEX_EXIT(&conn->conn_data_lock);
4145 for (i = 0; i < RX_MAXCALLS; i++) {
4146 struct rx_call *call = conn->call[i];
4149 MUTEX_ENTER(&call->lock);
4150 /* tnop can be null if newcallp is null */
4151 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4153 MUTEX_EXIT(&call->lock);
4157 MUTEX_EXIT(&conn->conn_data_lock);
4160 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4162 rx_ack_reason(int reason)
4165 case RX_ACK_REQUESTED:
4167 case RX_ACK_DUPLICATE:
4169 case RX_ACK_OUT_OF_SEQUENCE:
4171 case RX_ACK_EXCEEDS_WINDOW:
4173 case RX_ACK_NOSPACE:
4177 case RX_ACK_PING_RESPONSE:
4190 /* The real smarts of the whole thing. */
4192 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4195 struct rx_ackPacket *ap;
4197 struct rx_packet *tp;
4198 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
4199 struct rx_connection *conn = call->conn;
4200 struct rx_peer *peer = conn->peer;
4201 struct clock now; /* Current time, for RTT calculations */
4205 /* because there are CM's that are bogus, sending weird values for this. */
4206 afs_uint32 skew = 0;
4211 int newAckCount = 0;
4212 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4213 int pktsize = 0; /* Set if we need to update the peer mtu */
4214 int conn_data_locked = 0;
4216 if (rx_stats_active)
4217 rx_atomic_inc(&rx_stats.ackPacketsRead);
4218 ap = (struct rx_ackPacket *)rx_DataOf(np);
4219 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4221 return np; /* truncated ack packet */
4223 /* depends on ack packet struct */
4224 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4225 first = ntohl(ap->firstPacket);
4226 prev = ntohl(ap->previousPacket);
4227 serial = ntohl(ap->serial);
4228 /* temporarily disabled -- needs to degrade over time
4229 * skew = ntohs(ap->maxSkew); */
4231 /* Ignore ack packets received out of order */
4232 if (first < call->tfirst ||
4233 (first == call->tfirst && prev < call->tprev)) {
4239 if (np->header.flags & RX_SLOW_START_OK) {
4240 call->flags |= RX_CALL_SLOW_START_OK;
4243 if (ap->reason == RX_ACK_PING_RESPONSE)
4244 rxi_UpdatePeerReach(conn, call);
4246 if (conn->lastPacketSizeSeq) {
4247 MUTEX_ENTER(&conn->conn_data_lock);
4248 conn_data_locked = 1;
4249 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4250 pktsize = conn->lastPacketSize;
4251 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4254 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4255 if (!conn_data_locked) {
4256 MUTEX_ENTER(&conn->conn_data_lock);
4257 conn_data_locked = 1;
4259 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4260 /* process mtu ping ack */
4261 pktsize = conn->lastPingSize;
4262 conn->lastPingSizeSer = conn->lastPingSize = 0;
4266 if (conn_data_locked) {
4267 MUTEX_EXIT(&conn->conn_data_lock);
4268 conn_data_locked = 0;
4272 if (rxdebug_active) {
4276 len = _snprintf(msg, sizeof(msg),
4277 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
4278 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4279 ntohl(ap->serial), ntohl(ap->previousPacket),
4280 (unsigned int)np->header.seq, (unsigned int)skew,
4281 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
4285 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4286 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4290 OutputDebugString(msg);
4292 #else /* AFS_NT40_ENV */
4295 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
4296 ap->reason, ntohl(ap->previousPacket),
4297 (unsigned int)np->header.seq, (unsigned int)serial,
4298 (unsigned int)skew, ntohl(ap->firstPacket));
4301 for (offset = 0; offset < nAcks; offset++)
4302 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4307 #endif /* AFS_NT40_ENV */
4310 MUTEX_ENTER(&peer->peer_lock);
4313 * Start somewhere. Can't assume we can send what we can receive,
4314 * but we are clearly receiving.
4316 if (!peer->maxPacketSize)
4317 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4319 if (pktsize > peer->maxPacketSize) {
4320 peer->maxPacketSize = pktsize;
4321 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4322 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4323 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4324 rxi_ScheduleGrowMTUEvent(call, 1);
4329 /* Update the outgoing packet skew value to the latest value of
4330 * the peer's incoming packet skew value. The ack packet, of
4331 * course, could arrive out of order, but that won't affect things
4333 peer->outPacketSkew = skew;
4336 clock_GetTime(&now);
4338 /* The transmit queue splits into 4 sections.
4340 * The first section is packets which have now been acknowledged
4341 * by a window size change in the ack. These have reached the
4342 * application layer, and may be discarded. These are packets
4343 * with sequence numbers < ap->firstPacket.
4345 * The second section is packets which have sequence numbers in
4346 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4347 * contents of the packet's ack array determines whether these
4348 * packets are acknowledged or not.
4350 * The third section is packets which fall above the range
4351 * addressed in the ack packet. These have not yet been received
4354 * The four section is packets which have not yet been transmitted.
4355 * These packets will have a header.serial of 0.
4358 /* First section - implicitly acknowledged packets that can be
4362 tp = queue_First(&call->tq, rx_packet);
4363 while(!queue_IsEnd(&call->tq, tp) && tp->header.seq < first) {
4364 struct rx_packet *next;
4366 next = queue_Next(tp, rx_packet);
4367 call->tfirst = tp->header.seq + 1;
4369 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4371 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4375 rxi_ComputeRate(call->conn->peer, call, p, np, ap->reason);
4378 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4379 /* XXX Hack. Because we have to release the global rx lock when sending
4380 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4381 * in rxi_Start sending packets out because packets may move to the
4382 * freePacketQueue as result of being here! So we drop these packets until
4383 * we're safely out of the traversing. Really ugly!
4384 * To make it even uglier, if we're using fine grain locking, we can
4385 * set the ack bits in the packets and have rxi_Start remove the packets
4386 * when it's done transmitting.
4388 if (call->flags & RX_CALL_TQ_BUSY) {
4389 #ifdef RX_ENABLE_LOCKS
4390 tp->flags |= RX_PKTFLAG_ACKED;
4391 call->flags |= RX_CALL_TQ_SOME_ACKED;
4392 #else /* RX_ENABLE_LOCKS */
4394 #endif /* RX_ENABLE_LOCKS */
4396 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4399 #ifdef RX_TRACK_PACKETS
4400 tp->flags &= ~RX_PKTFLAG_TQ;
4402 #ifdef RXDEBUG_PACKET
4404 #endif /* RXDEBUG_PACKET */
4405 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4411 /* Give rate detector a chance to respond to ping requests */
4412 if (ap->reason == RX_ACK_PING_RESPONSE) {
4413 rxi_ComputeRate(peer, call, 0, np, ap->reason);
4417 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4419 /* Second section of the queue - packets for which we are receiving
4422 * Go through the explicit acks/nacks and record the results in
4423 * the waiting packets. These are packets that can't be released
4424 * yet, even with a positive acknowledge. This positive
4425 * acknowledge only means the packet has been received by the
4426 * peer, not that it will be retained long enough to be sent to
4427 * the peer's upper level. In addition, reset the transmit timers
4428 * of any missing packets (those packets that must be missing
4429 * because this packet was out of sequence) */
4431 call->nSoftAcked = 0;
4433 while (!queue_IsEnd(&call->tq, tp) && tp->header.seq < first + nAcks) {
4434 /* Set the acknowledge flag per packet based on the
4435 * information in the ack packet. An acknowlegded packet can
4436 * be downgraded when the server has discarded a packet it
4437 * soacked previously, or when an ack packet is received
4438 * out of sequence. */
4439 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4440 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4442 tp->flags |= RX_PKTFLAG_ACKED;
4443 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4445 rxi_ComputeRate(call->conn->peer, call, tp, np, ap->reason);
4453 } else /* RX_ACK_TYPE_NACK */ {
4454 tp->flags &= ~RX_PKTFLAG_ACKED;
4458 tp = queue_Next(tp, rx_packet);
4461 /* We don't need to take any action with the 3rd or 4th section in the
4462 * queue - they're not addressed by the contents of this ACK packet.
4465 /* If the window has been extended by this acknowledge packet,
4466 * then wakeup a sender waiting in alloc for window space, or try
4467 * sending packets now, if he's been sitting on packets due to
4468 * lack of window space */
4469 if (call->tnext < (call->tfirst + call->twind)) {
4470 #ifdef RX_ENABLE_LOCKS
4471 CV_SIGNAL(&call->cv_twind);
4473 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4474 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4475 osi_rxWakeup(&call->twind);
4478 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4479 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4483 /* if the ack packet has a receivelen field hanging off it,
4484 * update our state */
4485 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4488 /* If the ack packet has a "recommended" size that is less than
4489 * what I am using now, reduce my size to match */
4490 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4491 (int)sizeof(afs_int32), &tSize);
4492 tSize = (afs_uint32) ntohl(tSize);
4493 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4495 /* Get the maximum packet size to send to this peer */
4496 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4498 tSize = (afs_uint32) ntohl(tSize);
4499 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4500 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4502 /* sanity check - peer might have restarted with different params.
4503 * If peer says "send less", dammit, send less... Peer should never
4504 * be unable to accept packets of the size that prior AFS versions would
4505 * send without asking. */
4506 if (peer->maxMTU != tSize) {
4507 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4509 peer->maxMTU = tSize;
4510 peer->MTU = MIN(tSize, peer->MTU);
4511 call->MTU = MIN(call->MTU, tSize);
4514 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4517 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4518 (int)sizeof(afs_int32), &tSize);
4519 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4520 if (tSize < call->twind) { /* smaller than our send */
4521 call->twind = tSize; /* window, we must send less... */
4522 call->ssthresh = MIN(call->twind, call->ssthresh);
4523 call->conn->twind[call->channel] = call->twind;
4526 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4527 * network MTU confused with the loopback MTU. Calculate the
4528 * maximum MTU here for use in the slow start code below.
4530 /* Did peer restart with older RX version? */
4531 if (peer->maxDgramPackets > 1) {
4532 peer->maxDgramPackets = 1;
4534 } else if (np->length >=
4535 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4538 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4539 sizeof(afs_int32), &tSize);
4540 tSize = (afs_uint32) ntohl(tSize);
4542 * As of AFS 3.5 we set the send window to match the receive window.
4544 if (tSize < call->twind) {
4545 call->twind = tSize;
4546 call->conn->twind[call->channel] = call->twind;
4547 call->ssthresh = MIN(call->twind, call->ssthresh);
4548 } else if (tSize > call->twind) {
4549 call->twind = tSize;
4550 call->conn->twind[call->channel] = call->twind;
4554 * As of AFS 3.5, a jumbogram is more than one fixed size
4555 * packet transmitted in a single UDP datagram. If the remote
4556 * MTU is smaller than our local MTU then never send a datagram
4557 * larger than the natural MTU.
4560 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4561 (int)sizeof(afs_int32), &tSize);
4562 maxDgramPackets = (afs_uint32) ntohl(tSize);
4563 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4565 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4566 if (maxDgramPackets > 1) {
4567 peer->maxDgramPackets = maxDgramPackets;
4568 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4570 peer->maxDgramPackets = 1;
4571 call->MTU = peer->natMTU;
4573 } else if (peer->maxDgramPackets > 1) {
4574 /* Restarted with lower version of RX */
4575 peer->maxDgramPackets = 1;
4577 } else if (peer->maxDgramPackets > 1
4578 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4579 /* Restarted with lower version of RX */
4580 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4581 peer->natMTU = OLD_MAX_PACKET_SIZE;
4582 peer->MTU = OLD_MAX_PACKET_SIZE;
4583 peer->maxDgramPackets = 1;
4584 peer->nDgramPackets = 1;
4586 call->MTU = OLD_MAX_PACKET_SIZE;
4591 * Calculate how many datagrams were successfully received after
4592 * the first missing packet and adjust the negative ack counter
4597 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4598 if (call->nNacks < nNacked) {
4599 call->nNacks = nNacked;
4602 call->nAcks += newAckCount;
4606 /* If the packet contained new acknowledgements, rather than just
4607 * being a duplicate of one we have previously seen, then we can restart
4610 if (newAckCount > 0)
4611 rxi_rto_packet_acked(call, istack);
4613 if (call->flags & RX_CALL_FAST_RECOVER) {
4614 if (newAckCount == 0) {
4615 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4617 call->flags &= ~RX_CALL_FAST_RECOVER;
4618 call->cwind = call->nextCwind;
4619 call->nextCwind = 0;
4622 call->nCwindAcks = 0;
4623 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4624 /* Three negative acks in a row trigger congestion recovery */
4625 call->flags |= RX_CALL_FAST_RECOVER;
4626 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4628 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4629 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4630 call->nextCwind = call->ssthresh;
4633 peer->MTU = call->MTU;
4634 peer->cwind = call->nextCwind;
4635 peer->nDgramPackets = call->nDgramPackets;
4637 call->congestSeq = peer->congestSeq;
4639 /* Reset the resend times on the packets that were nacked
4640 * so we will retransmit as soon as the window permits
4643 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4645 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4646 tp->flags &= ~RX_PKTFLAG_SENT;
4648 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4653 /* If cwind is smaller than ssthresh, then increase
4654 * the window one packet for each ack we receive (exponential
4656 * If cwind is greater than or equal to ssthresh then increase
4657 * the congestion window by one packet for each cwind acks we
4658 * receive (linear growth). */
4659 if (call->cwind < call->ssthresh) {
4661 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4662 call->nCwindAcks = 0;
4664 call->nCwindAcks += newAckCount;
4665 if (call->nCwindAcks >= call->cwind) {
4666 call->nCwindAcks = 0;
4667 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4671 * If we have received several acknowledgements in a row then
4672 * it is time to increase the size of our datagrams
4674 if ((int)call->nAcks > rx_nDgramThreshold) {
4675 if (peer->maxDgramPackets > 1) {
4676 if (call->nDgramPackets < peer->maxDgramPackets) {
4677 call->nDgramPackets++;
4679 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4680 } else if (call->MTU < peer->maxMTU) {
4681 /* don't upgrade if we can't handle it */
4682 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4683 call->MTU = peer->ifMTU;
4685 call->MTU += peer->natMTU;
4686 call->MTU = MIN(call->MTU, peer->maxMTU);
4693 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4695 /* Servers need to hold the call until all response packets have
4696 * been acknowledged. Soft acks are good enough since clients
4697 * are not allowed to clear their receive queues. */
4698 if (call->state == RX_STATE_HOLD
4699 && call->tfirst + call->nSoftAcked >= call->tnext) {
4700 call->state = RX_STATE_DALLY;
4701 rxi_ClearTransmitQueue(call, 0);
4702 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4703 } else if (!queue_IsEmpty(&call->tq)) {
4704 rxi_Start(call, istack);
4709 /* Received a response to a challenge packet */
4711 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4712 struct rx_packet *np, int istack)
4716 /* Ignore the packet if we're the client */
4717 if (conn->type == RX_CLIENT_CONNECTION)
4720 /* If already authenticated, ignore the packet (it's probably a retry) */
4721 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4724 /* Otherwise, have the security object evaluate the response packet */
4725 error = RXS_CheckResponse(conn->securityObject, conn, np);
4727 /* If the response is invalid, reset the connection, sending
4728 * an abort to the peer */
4732 rxi_ConnectionError(conn, error);
4733 MUTEX_ENTER(&conn->conn_data_lock);
4734 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4735 MUTEX_EXIT(&conn->conn_data_lock);
4738 /* If the response is valid, any calls waiting to attach
4739 * servers can now do so */
4742 for (i = 0; i < RX_MAXCALLS; i++) {
4743 struct rx_call *call = conn->call[i];
4745 MUTEX_ENTER(&call->lock);
4746 if (call->state == RX_STATE_PRECALL)
4747 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4748 /* tnop can be null if newcallp is null */
4749 MUTEX_EXIT(&call->lock);
4753 /* Update the peer reachability information, just in case
4754 * some calls went into attach-wait while we were waiting
4755 * for authentication..
4757 rxi_UpdatePeerReach(conn, NULL);
4762 /* A client has received an authentication challenge: the security
4763 * object is asked to cough up a respectable response packet to send
4764 * back to the server. The server is responsible for retrying the
4765 * challenge if it fails to get a response. */
4768 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4769 struct rx_packet *np, int istack)
4773 /* Ignore the challenge if we're the server */
4774 if (conn->type == RX_SERVER_CONNECTION)
4777 /* Ignore the challenge if the connection is otherwise idle; someone's
4778 * trying to use us as an oracle. */
4779 if (!rxi_HasActiveCalls(conn))
4782 /* Send the security object the challenge packet. It is expected to fill
4783 * in the response. */
4784 error = RXS_GetResponse(conn->securityObject, conn, np);
4786 /* If the security object is unable to return a valid response, reset the
4787 * connection and send an abort to the peer. Otherwise send the response
4788 * packet to the peer connection. */
4790 rxi_ConnectionError(conn, error);
4791 MUTEX_ENTER(&conn->conn_data_lock);
4792 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4793 MUTEX_EXIT(&conn->conn_data_lock);
4795 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4796 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4802 /* Find an available server process to service the current request in
4803 * the given call structure. If one isn't available, queue up this
4804 * call so it eventually gets one */
4806 rxi_AttachServerProc(struct rx_call *call,
4807 osi_socket socket, int *tnop,
4808 struct rx_call **newcallp)
4810 struct rx_serverQueueEntry *sq;
4811 struct rx_service *service = call->conn->service;
4814 /* May already be attached */
4815 if (call->state == RX_STATE_ACTIVE)
4818 MUTEX_ENTER(&rx_serverPool_lock);
4820 haveQuota = QuotaOK(service);
4821 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4822 /* If there are no processes available to service this call,
4823 * put the call on the incoming call queue (unless it's
4824 * already on the queue).
4826 #ifdef RX_ENABLE_LOCKS
4828 ReturnToServerPool(service);
4829 #endif /* RX_ENABLE_LOCKS */
4831 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4832 call->flags |= RX_CALL_WAIT_PROC;
4833 rx_atomic_inc(&rx_nWaiting);
4834 rx_atomic_inc(&rx_nWaited);
4835 rxi_calltrace(RX_CALL_ARRIVAL, call);
4836 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4837 queue_Append(&rx_incomingCallQueue, call);
4840 sq = queue_Last(&rx_idleServerQueue, rx_serverQueueEntry);
4842 /* If hot threads are enabled, and both newcallp and sq->socketp
4843 * are non-null, then this thread will process the call, and the
4844 * idle server thread will start listening on this threads socket.
4847 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4850 *sq->socketp = socket;
4851 clock_GetTime(&call->startTime);
4852 MUTEX_ENTER(&rx_refcnt_mutex);
4853 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4854 MUTEX_EXIT(&rx_refcnt_mutex);
4858 if (call->flags & RX_CALL_WAIT_PROC) {
4859 /* Conservative: I don't think this should happen */
4860 call->flags &= ~RX_CALL_WAIT_PROC;
4861 if (queue_IsOnQueue(call)) {
4864 rx_atomic_dec(&rx_nWaiting);
4867 call->state = RX_STATE_ACTIVE;
4868 call->mode = RX_MODE_RECEIVING;
4869 #ifdef RX_KERNEL_TRACE
4871 int glockOwner = ISAFS_GLOCK();
4874 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4875 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4881 if (call->flags & RX_CALL_CLEARED) {
4882 /* send an ack now to start the packet flow up again */
4883 call->flags &= ~RX_CALL_CLEARED;
4884 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4886 #ifdef RX_ENABLE_LOCKS
4889 service->nRequestsRunning++;
4890 MUTEX_ENTER(&rx_quota_mutex);
4891 if (service->nRequestsRunning <= service->minProcs)
4894 MUTEX_EXIT(&rx_quota_mutex);
4898 MUTEX_EXIT(&rx_serverPool_lock);
4901 /* Delay the sending of an acknowledge event for a short while, while
4902 * a new call is being prepared (in the case of a client) or a reply
4903 * is being prepared (in the case of a server). Rather than sending
4904 * an ack packet, an ACKALL packet is sent. */
4906 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4908 #ifdef RX_ENABLE_LOCKS
4910 MUTEX_ENTER(&call->lock);
4911 call->delayedAckEvent = NULL;
4912 MUTEX_ENTER(&rx_refcnt_mutex);
4913 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4914 MUTEX_EXIT(&rx_refcnt_mutex);
4916 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4917 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4918 call->flags |= RX_CALL_ACKALL_SENT;
4920 MUTEX_EXIT(&call->lock);
4921 #else /* RX_ENABLE_LOCKS */
4923 call->delayedAckEvent = NULL;
4924 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4925 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4926 call->flags |= RX_CALL_ACKALL_SENT;
4927 #endif /* RX_ENABLE_LOCKS */
4931 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4933 struct rx_call *call = arg1;
4934 #ifdef RX_ENABLE_LOCKS
4936 MUTEX_ENTER(&call->lock);
4937 if (event == call->delayedAckEvent)
4938 call->delayedAckEvent = NULL;
4939 MUTEX_ENTER(&rx_refcnt_mutex);
4940 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4941 MUTEX_EXIT(&rx_refcnt_mutex);
4943 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4945 MUTEX_EXIT(&call->lock);
4946 #else /* RX_ENABLE_LOCKS */
4948 call->delayedAckEvent = NULL;
4949 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4950 #endif /* RX_ENABLE_LOCKS */
4954 #ifdef RX_ENABLE_LOCKS
4955 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4956 * clearing them out.
4959 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4961 struct rx_packet *p, *tp;
4964 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4965 p->flags |= RX_PKTFLAG_ACKED;
4969 call->flags |= RX_CALL_TQ_CLEARME;
4970 call->flags |= RX_CALL_TQ_SOME_ACKED;
4973 rxi_rto_cancel(call);
4975 call->tfirst = call->tnext;
4976 call->nSoftAcked = 0;
4978 if (call->flags & RX_CALL_FAST_RECOVER) {
4979 call->flags &= ~RX_CALL_FAST_RECOVER;
4980 call->cwind = call->nextCwind;
4981 call->nextCwind = 0;
4984 CV_SIGNAL(&call->cv_twind);
4986 #endif /* RX_ENABLE_LOCKS */
4988 /* Clear out the transmit queue for the current call (all packets have
4989 * been received by peer) */
4991 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4993 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4994 struct rx_packet *p, *tp;
4996 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4998 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4999 p->flags |= RX_PKTFLAG_ACKED;
5003 call->flags |= RX_CALL_TQ_CLEARME;
5004 call->flags |= RX_CALL_TQ_SOME_ACKED;
5007 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5008 #ifdef RXDEBUG_PACKET
5010 #endif /* RXDEBUG_PACKET */
5011 rxi_FreePackets(0, &call->tq);
5012 rxi_WakeUpTransmitQueue(call);
5013 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5014 call->flags &= ~RX_CALL_TQ_CLEARME;
5016 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5018 rxi_rto_cancel(call);
5019 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5020 call->nSoftAcked = 0;
5022 if (call->flags & RX_CALL_FAST_RECOVER) {
5023 call->flags &= ~RX_CALL_FAST_RECOVER;
5024 call->cwind = call->nextCwind;
5026 #ifdef RX_ENABLE_LOCKS
5027 CV_SIGNAL(&call->cv_twind);
5029 osi_rxWakeup(&call->twind);
5034 rxi_ClearReceiveQueue(struct rx_call *call)
5036 if (queue_IsNotEmpty(&call->rq)) {
5039 count = rxi_FreePackets(0, &call->rq);
5040 rx_packetReclaims += count;
5041 #ifdef RXDEBUG_PACKET
5043 if ( call->rqc != 0 )
5044 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5046 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5048 if (call->state == RX_STATE_PRECALL) {
5049 call->flags |= RX_CALL_CLEARED;
5053 /* Send an abort packet for the specified call */
5055 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5056 int istack, int force)
5059 struct clock when, now;
5064 /* Clients should never delay abort messages */
5065 if (rx_IsClientConn(call->conn))
5068 if (call->abortCode != call->error) {
5069 call->abortCode = call->error;
5070 call->abortCount = 0;
5073 if (force || rxi_callAbortThreshhold == 0
5074 || call->abortCount < rxi_callAbortThreshhold) {
5075 if (call->delayedAbortEvent) {
5076 rxevent_Cancel(call->delayedAbortEvent, call,
5077 RX_CALL_REFCOUNT_ABORT);
5079 error = htonl(call->error);
5082 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5083 (char *)&error, sizeof(error), istack);
5084 } else if (!call->delayedAbortEvent) {
5085 clock_GetTime(&now);
5087 clock_Addmsec(&when, rxi_callAbortDelay);
5088 MUTEX_ENTER(&rx_refcnt_mutex);
5089 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5090 MUTEX_EXIT(&rx_refcnt_mutex);
5091 call->delayedAbortEvent =
5092 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
5097 /* Send an abort packet for the specified connection. Packet is an
5098 * optional pointer to a packet that can be used to send the abort.
5099 * Once the number of abort messages reaches the threshhold, an
5100 * event is scheduled to send the abort. Setting the force flag
5101 * overrides sending delayed abort messages.
5103 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5104 * to send the abort packet.
5107 rxi_SendConnectionAbort(struct rx_connection *conn,
5108 struct rx_packet *packet, int istack, int force)
5111 struct clock when, now;
5116 /* Clients should never delay abort messages */
5117 if (rx_IsClientConn(conn))
5120 if (force || rxi_connAbortThreshhold == 0
5121 || conn->abortCount < rxi_connAbortThreshhold) {
5122 if (conn->delayedAbortEvent) {
5123 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
5125 error = htonl(conn->error);
5127 MUTEX_EXIT(&conn->conn_data_lock);
5129 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5130 RX_PACKET_TYPE_ABORT, (char *)&error,
5131 sizeof(error), istack);
5132 MUTEX_ENTER(&conn->conn_data_lock);
5133 } else if (!conn->delayedAbortEvent) {
5134 clock_GetTime(&now);
5136 clock_Addmsec(&when, rxi_connAbortDelay);
5137 conn->delayedAbortEvent =
5138 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
5143 /* Associate an error all of the calls owned by a connection. Called
5144 * with error non-zero. This is only for really fatal things, like
5145 * bad authentication responses. The connection itself is set in
5146 * error at this point, so that future packets received will be
5149 rxi_ConnectionError(struct rx_connection *conn,
5155 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5157 MUTEX_ENTER(&conn->conn_data_lock);
5158 if (conn->challengeEvent)
5159 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
5160 if (conn->natKeepAliveEvent)
5161 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
5162 if (conn->checkReachEvent) {
5163 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
5164 conn->checkReachEvent = 0;
5165 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5166 MUTEX_ENTER(&rx_refcnt_mutex);
5168 MUTEX_EXIT(&rx_refcnt_mutex);
5170 MUTEX_EXIT(&conn->conn_data_lock);
5171 for (i = 0; i < RX_MAXCALLS; i++) {
5172 struct rx_call *call = conn->call[i];
5174 MUTEX_ENTER(&call->lock);
5175 rxi_CallError(call, error);
5176 MUTEX_EXIT(&call->lock);
5179 conn->error = error;
5180 if (rx_stats_active)
5181 rx_atomic_inc(&rx_stats.fatalErrors);
5186 * Interrupt an in-progress call with the specified error and wakeup waiters.
5188 * @param[in] call The call to interrupt
5189 * @param[in] error The error code to send to the peer
5192 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5194 MUTEX_ENTER(&call->lock);
5195 rxi_CallError(call, error);
5196 rxi_SendCallAbort(call, NULL, 0, 1);
5197 MUTEX_EXIT(&call->lock);
5201 rxi_CallError(struct rx_call *call, afs_int32 error)
5204 osirx_AssertMine(&call->lock, "rxi_CallError");
5206 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5208 error = call->error;
5210 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5211 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5212 rxi_ResetCall(call, 0);
5215 rxi_ResetCall(call, 0);
5217 call->error = error;
5220 /* Reset various fields in a call structure, and wakeup waiting
5221 * processes. Some fields aren't changed: state & mode are not
5222 * touched (these must be set by the caller), and bufptr, nLeft, and
5223 * nFree are not reset, since these fields are manipulated by
5224 * unprotected macros, and may only be reset by non-interrupting code.
5227 /* this code requires that call->conn be set properly as a pre-condition. */
5228 #endif /* ADAPT_WINDOW */
5231 rxi_ResetCall(struct rx_call *call, int newcall)
5234 struct rx_peer *peer;
5235 struct rx_packet *packet;
5237 osirx_AssertMine(&call->lock, "rxi_ResetCall");
5239 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5241 /* Notify anyone who is waiting for asynchronous packet arrival */
5242 if (call->arrivalProc) {
5243 (*call->arrivalProc) (call, call->arrivalProcHandle,
5244 call->arrivalProcArg);
5245 call->arrivalProc = (void (*)())0;
5248 if (call->growMTUEvent)
5249 rxevent_Cancel(call->growMTUEvent, call,
5250 RX_CALL_REFCOUNT_ALIVE);
5252 if (call->delayedAbortEvent) {
5253 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
5254 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5256 rxi_SendCallAbort(call, packet, 0, 1);
5257 rxi_FreePacket(packet);
5262 * Update the peer with the congestion information in this call
5263 * so other calls on this connection can pick up where this call
5264 * left off. If the congestion sequence numbers don't match then
5265 * another call experienced a retransmission.
5267 peer = call->conn->peer;
5268 MUTEX_ENTER(&peer->peer_lock);
5270 if (call->congestSeq == peer->congestSeq) {
5271 peer->cwind = MAX(peer->cwind, call->cwind);
5272 peer->MTU = MAX(peer->MTU, call->MTU);
5273 peer->nDgramPackets =
5274 MAX(peer->nDgramPackets, call->nDgramPackets);
5277 call->abortCode = 0;
5278 call->abortCount = 0;
5280 if (peer->maxDgramPackets > 1) {
5281 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5283 call->MTU = peer->MTU;
5285 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5286 call->ssthresh = rx_maxSendWindow;
5287 call->nDgramPackets = peer->nDgramPackets;
5288 call->congestSeq = peer->congestSeq;
5289 call->rtt = peer->rtt;
5290 call->rtt_dev = peer->rtt_dev;
5291 clock_Zero(&call->rto);
5292 clock_Addmsec(&call->rto,
5293 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5294 MUTEX_EXIT(&peer->peer_lock);
5296 flags = call->flags;
5297 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5298 rxi_WaitforTQBusy(call);
5299 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5301 rxi_ClearTransmitQueue(call, 1);
5302 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5303 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5307 if ((flags & RX_CALL_PEER_BUSY)) {
5308 /* The call channel is still busy; resetting the call doesn't change
5310 call->flags |= RX_CALL_PEER_BUSY;
5313 rxi_ClearReceiveQueue(call);
5314 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5318 call->twind = call->conn->twind[call->channel];
5319 call->rwind = call->conn->rwind[call->channel];
5320 call->nSoftAcked = 0;
5321 call->nextCwind = 0;
5324 call->nCwindAcks = 0;
5325 call->nSoftAcks = 0;
5326 call->nHardAcks = 0;
5328 call->tfirst = call->rnext = call->tnext = 1;
5331 call->lastAcked = 0;
5332 call->localStatus = call->remoteStatus = 0;
5334 if (flags & RX_CALL_READER_WAIT) {
5335 #ifdef RX_ENABLE_LOCKS
5336 CV_BROADCAST(&call->cv_rq);
5338 osi_rxWakeup(&call->rq);
5341 if (flags & RX_CALL_WAIT_PACKETS) {
5342 MUTEX_ENTER(&rx_freePktQ_lock);
5343 rxi_PacketsUnWait(); /* XXX */
5344 MUTEX_EXIT(&rx_freePktQ_lock);
5346 #ifdef RX_ENABLE_LOCKS
5347 CV_SIGNAL(&call->cv_twind);
5349 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5350 osi_rxWakeup(&call->twind);
5353 #ifdef RX_ENABLE_LOCKS
5354 /* The following ensures that we don't mess with any queue while some
5355 * other thread might also be doing so. The call_queue_lock field is
5356 * is only modified under the call lock. If the call is in the process
5357 * of being removed from a queue, the call is not locked until the
5358 * the queue lock is dropped and only then is the call_queue_lock field
5359 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5360 * Note that any other routine which removes a call from a queue has to
5361 * obtain the queue lock before examing the queue and removing the call.
5363 if (call->call_queue_lock) {
5364 MUTEX_ENTER(call->call_queue_lock);
5365 if (queue_IsOnQueue(call)) {
5367 if (flags & RX_CALL_WAIT_PROC) {
5368 rx_atomic_dec(&rx_nWaiting);
5371 MUTEX_EXIT(call->call_queue_lock);
5372 CLEAR_CALL_QUEUE_LOCK(call);
5374 #else /* RX_ENABLE_LOCKS */
5375 if (queue_IsOnQueue(call)) {
5377 if (flags & RX_CALL_WAIT_PROC)
5378 rx_atomic_dec(&rx_nWaiting);
5380 #endif /* RX_ENABLE_LOCKS */
5382 rxi_KeepAliveOff(call);
5383 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5386 /* Send an acknowledge for the indicated packet (seq,serial) of the
5387 * indicated call, for the indicated reason (reason). This
5388 * acknowledge will specifically acknowledge receiving the packet, and
5389 * will also specify which other packets for this call have been
5390 * received. This routine returns the packet that was used to the
5391 * caller. The caller is responsible for freeing it or re-using it.
5392 * This acknowledgement also returns the highest sequence number
5393 * actually read out by the higher level to the sender; the sender
5394 * promises to keep around packets that have not been read by the
5395 * higher level yet (unless, of course, the sender decides to abort
5396 * the call altogether). Any of p, seq, serial, pflags, or reason may
5397 * be set to zero without ill effect. That is, if they are zero, they
5398 * will not convey any information.
5399 * NOW there is a trailer field, after the ack where it will safely be
5400 * ignored by mundanes, which indicates the maximum size packet this
5401 * host can swallow. */
5403 struct rx_packet *optionalPacket; use to send ack (or null)
5404 int seq; Sequence number of the packet we are acking
5405 int serial; Serial number of the packet
5406 int pflags; Flags field from packet header
5407 int reason; Reason an acknowledge was prompted
5411 rxi_SendAck(struct rx_call *call,
5412 struct rx_packet *optionalPacket, int serial, int reason,
5415 struct rx_ackPacket *ap;
5416 struct rx_packet *rqp;
5417 struct rx_packet *nxp; /* For queue_Scan */
5418 struct rx_packet *p;
5421 afs_uint32 padbytes = 0;
5422 #ifdef RX_ENABLE_TSFPQ
5423 struct rx_ts_info_t * rx_ts_info;
5427 * Open the receive window once a thread starts reading packets
5429 if (call->rnext > 1) {
5430 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5433 /* Don't attempt to grow MTU if this is a critical ping */
5434 if (reason == RX_ACK_MTU) {
5435 /* keep track of per-call attempts, if we're over max, do in small
5436 * otherwise in larger? set a size to increment by, decrease
5439 if (call->conn->peer->maxPacketSize &&
5440 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5442 padbytes = call->conn->peer->maxPacketSize+16;
5444 padbytes = call->conn->peer->maxMTU + 128;
5446 /* do always try a minimum size ping */
5447 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5449 /* subtract the ack payload */
5450 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5451 reason = RX_ACK_PING;
5454 call->nHardAcks = 0;
5455 call->nSoftAcks = 0;
5456 if (call->rnext > call->lastAcked)
5457 call->lastAcked = call->rnext;
5461 rx_computelen(p, p->length); /* reset length, you never know */
5462 } /* where that's been... */
5463 #ifdef RX_ENABLE_TSFPQ
5465 RX_TS_INFO_GET(rx_ts_info);
5466 if ((p = rx_ts_info->local_special_packet)) {
5467 rx_computelen(p, p->length);
5468 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5469 rx_ts_info->local_special_packet = p;
5470 } else { /* We won't send the ack, but don't panic. */
5471 return optionalPacket;
5475 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5476 /* We won't send the ack, but don't panic. */
5477 return optionalPacket;
5482 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5485 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5486 #ifndef RX_ENABLE_TSFPQ
5487 if (!optionalPacket)
5490 return optionalPacket;
5492 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5493 if (rx_Contiguous(p) < templ) {
5494 #ifndef RX_ENABLE_TSFPQ
5495 if (!optionalPacket)
5498 return optionalPacket;
5503 /* MTUXXX failing to send an ack is very serious. We should */
5504 /* try as hard as possible to send even a partial ack; it's */
5505 /* better than nothing. */
5506 ap = (struct rx_ackPacket *)rx_DataOf(p);
5507 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5508 ap->reason = reason;
5510 /* The skew computation used to be bogus, I think it's better now. */
5511 /* We should start paying attention to skew. XXX */
5512 ap->serial = htonl(serial);
5513 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5516 * First packet not yet forwarded to reader. When ACKALL has been
5517 * sent the peer has been told that all received packets will be
5518 * delivered to the reader. The value 'rnext' is used internally
5519 * to refer to the next packet in the receive queue that must be
5520 * delivered to the reader. From the perspective of the peer it
5521 * already has so report the last sequence number plus one if there
5522 * are packets in the receive queue awaiting processing.
5524 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5525 !queue_IsEmpty(&call->rq)) {
5526 ap->firstPacket = htonl(queue_Last(&call->rq, rx_packet)->header.seq + 1);
5528 ap->firstPacket = htonl(call->rnext);
5530 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5532 /* No fear of running out of ack packet here because there can only be at most
5533 * one window full of unacknowledged packets. The window size must be constrained
5534 * to be less than the maximum ack size, of course. Also, an ack should always
5535 * fit into a single packet -- it should not ever be fragmented. */
5536 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5537 if (!rqp || !call->rq.next
5538 || (rqp->header.seq > (call->rnext + call->rwind))) {
5539 #ifndef RX_ENABLE_TSFPQ
5540 if (!optionalPacket)
5543 rxi_CallError(call, RX_CALL_DEAD);
5544 return optionalPacket;
5547 while (rqp->header.seq > call->rnext + offset)
5548 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5549 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5551 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5552 #ifndef RX_ENABLE_TSFPQ
5553 if (!optionalPacket)
5556 rxi_CallError(call, RX_CALL_DEAD);
5557 return optionalPacket;
5562 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5564 /* these are new for AFS 3.3 */
5565 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5566 templ = htonl(templ);
5567 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5568 templ = htonl(call->conn->peer->ifMTU);
5569 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5570 sizeof(afs_int32), &templ);
5572 /* new for AFS 3.4 */
5573 templ = htonl(call->rwind);
5574 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5575 sizeof(afs_int32), &templ);
5577 /* new for AFS 3.5 */
5578 templ = htonl(call->conn->peer->ifDgramPackets);
5579 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5580 sizeof(afs_int32), &templ);
5582 p->header.serviceId = call->conn->serviceId;
5583 p->header.cid = (call->conn->cid | call->channel);
5584 p->header.callNumber = *call->callNumber;
5586 p->header.securityIndex = call->conn->securityIndex;
5587 p->header.epoch = call->conn->epoch;
5588 p->header.type = RX_PACKET_TYPE_ACK;
5589 p->header.flags = RX_SLOW_START_OK;
5590 if (reason == RX_ACK_PING) {
5591 p->header.flags |= RX_REQUEST_ACK;
5593 clock_GetTime(&call->pingRequestTime);
5596 p->length = padbytes +
5597 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5600 /* not fast but we can potentially use this if truncated
5601 * fragments are delivered to figure out the mtu.
5603 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5604 sizeof(afs_int32), sizeof(afs_int32),
5608 if (call->conn->type == RX_CLIENT_CONNECTION)
5609 p->header.flags |= RX_CLIENT_INITIATED;
5613 if (rxdebug_active) {
5617 len = _snprintf(msg, sizeof(msg),
5618 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5619 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5620 ntohl(ap->serial), ntohl(ap->previousPacket),
5621 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5622 ap->nAcks, ntohs(ap->bufferSpace) );
5626 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5627 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5631 OutputDebugString(msg);
5633 #else /* AFS_NT40_ENV */
5635 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5636 ap->reason, ntohl(ap->previousPacket),
5637 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5639 for (offset = 0; offset < ap->nAcks; offset++)
5640 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5645 #endif /* AFS_NT40_ENV */
5648 int i, nbytes = p->length;
5650 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5651 if (nbytes <= p->wirevec[i].iov_len) {
5654 savelen = p->wirevec[i].iov_len;
5656 p->wirevec[i].iov_len = nbytes;
5658 rxi_Send(call, p, istack);
5659 p->wirevec[i].iov_len = savelen;
5663 nbytes -= p->wirevec[i].iov_len;
5666 if (rx_stats_active)
5667 rx_atomic_inc(&rx_stats.ackPacketsSent);
5668 #ifndef RX_ENABLE_TSFPQ
5669 if (!optionalPacket)
5672 return optionalPacket; /* Return packet for re-use by caller */
5676 struct rx_packet **list;
5681 /* Send all of the packets in the list in single datagram */
5683 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5684 int istack, int moreFlag)
5690 struct rx_connection *conn = call->conn;
5691 struct rx_peer *peer = conn->peer;
5693 MUTEX_ENTER(&peer->peer_lock);
5694 peer->nSent += xmit->len;
5695 if (xmit->resending)
5696 peer->reSends += xmit->len;
5697 MUTEX_EXIT(&peer->peer_lock);
5699 if (rx_stats_active) {
5700 if (xmit->resending)
5701 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5703 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5706 clock_GetTime(&now);
5708 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5712 /* Set the packet flags and schedule the resend events */
5713 /* Only request an ack for the last packet in the list */
5714 for (i = 0; i < xmit->len; i++) {
5715 struct rx_packet *packet = xmit->list[i];
5717 /* Record the time sent */
5718 packet->timeSent = now;
5719 packet->flags |= RX_PKTFLAG_SENT;
5721 /* Ask for an ack on retransmitted packets, on every other packet
5722 * if the peer doesn't support slow start. Ask for an ack on every
5723 * packet until the congestion window reaches the ack rate. */
5724 if (packet->header.serial) {
5727 packet->firstSent = now;
5728 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5729 || (!(call->flags & RX_CALL_SLOW_START_OK)
5730 && (packet->header.seq & 1)))) {
5735 /* Tag this packet as not being the last in this group,
5736 * for the receiver's benefit */
5737 if (i < xmit->len - 1 || moreFlag) {
5738 packet->header.flags |= RX_MORE_PACKETS;
5743 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5746 /* Since we're about to send a data packet to the peer, it's
5747 * safe to nuke any scheduled end-of-packets ack */
5748 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5750 MUTEX_EXIT(&call->lock);
5751 MUTEX_ENTER(&rx_refcnt_mutex);
5752 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5753 MUTEX_EXIT(&rx_refcnt_mutex);
5754 if (xmit->len > 1) {
5755 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5757 rxi_SendPacket(call, conn, xmit->list[0], istack);
5759 MUTEX_ENTER(&call->lock);
5760 MUTEX_ENTER(&rx_refcnt_mutex);
5761 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5762 MUTEX_EXIT(&rx_refcnt_mutex);
5764 /* Tell the RTO calculation engine that we have sent a packet, and
5765 * if it was the last one */
5766 rxi_rto_packet_sent(call, lastPacket, istack);
5768 /* Update last send time for this call (for keep-alive
5769 * processing), and for the connection (so that we can discover
5770 * idle connections) */
5771 conn->lastSendTime = call->lastSendTime = clock_Sec();
5772 /* Let a set of retransmits trigger an idle timeout */
5773 if (!xmit->resending)
5774 call->lastSendData = call->lastSendTime;
5777 /* When sending packets we need to follow these rules:
5778 * 1. Never send more than maxDgramPackets in a jumbogram.
5779 * 2. Never send a packet with more than two iovecs in a jumbogram.
5780 * 3. Never send a retransmitted packet in a jumbogram.
5781 * 4. Never send more than cwind/4 packets in a jumbogram
5782 * We always keep the last list we should have sent so we
5783 * can set the RX_MORE_PACKETS flags correctly.
5787 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5792 struct xmitlist working;
5793 struct xmitlist last;
5795 struct rx_peer *peer = call->conn->peer;
5796 int morePackets = 0;
5798 memset(&last, 0, sizeof(struct xmitlist));
5799 working.list = &list[0];
5801 working.resending = 0;
5803 recovery = call->flags & RX_CALL_FAST_RECOVER;
5805 for (i = 0; i < len; i++) {
5806 /* Does the current packet force us to flush the current list? */
5808 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5809 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5811 /* This sends the 'last' list and then rolls the current working
5812 * set into the 'last' one, and resets the working set */
5815 rxi_SendList(call, &last, istack, 1);
5816 /* If the call enters an error state stop sending, or if
5817 * we entered congestion recovery mode, stop sending */
5819 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5824 working.resending = 0;
5825 working.list = &list[i];
5827 /* Add the current packet to the list if it hasn't been acked.
5828 * Otherwise adjust the list pointer to skip the current packet. */
5829 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5832 if (list[i]->header.serial)
5833 working.resending = 1;
5835 /* Do we need to flush the list? */
5836 if (working.len >= (int)peer->maxDgramPackets
5837 || working.len >= (int)call->nDgramPackets
5838 || working.len >= (int)call->cwind
5839 || list[i]->header.serial
5840 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5842 rxi_SendList(call, &last, istack, 1);
5843 /* If the call enters an error state stop sending, or if
5844 * we entered congestion recovery mode, stop sending */
5846 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5851 working.resending = 0;
5852 working.list = &list[i + 1];
5855 if (working.len != 0) {
5856 osi_Panic("rxi_SendList error");
5858 working.list = &list[i + 1];
5862 /* Send the whole list when the call is in receive mode, when
5863 * the call is in eof mode, when we are in fast recovery mode,
5864 * and when we have the last packet */
5865 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5866 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5867 || (call->flags & RX_CALL_FAST_RECOVER)) {
5868 /* Check for the case where the current list contains
5869 * an acked packet. Since we always send retransmissions
5870 * in a separate packet, we only need to check the first
5871 * packet in the list */
5872 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5876 rxi_SendList(call, &last, istack, morePackets);
5877 /* If the call enters an error state stop sending, or if
5878 * we entered congestion recovery mode, stop sending */
5880 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5884 rxi_SendList(call, &working, istack, 0);
5886 } else if (last.len > 0) {
5887 rxi_SendList(call, &last, istack, 0);
5888 /* Packets which are in 'working' are not sent by this call */
5893 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
5895 struct rx_call *call = arg0;
5896 struct rx_peer *peer;
5897 struct rx_packet *p, *nxp;
5898 struct clock maxTimeout = { 60, 0 };
5900 MUTEX_ENTER(&call->lock);
5902 peer = call->conn->peer;
5904 /* Make sure that the event pointer is removed from the call
5905 * structure, since there is no longer a per-call retransmission
5907 if (event == call->resendEvent) {
5908 MUTEX_ENTER(&rx_refcnt_mutex);
5909 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5910 MUTEX_EXIT(&rx_refcnt_mutex);
5911 call->resendEvent = NULL;
5914 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
5915 rxi_CheckBusy(call);
5918 if (queue_IsEmpty(&call->tq)) {
5919 /* Nothing to do. This means that we've been raced, and that an
5920 * ACK has come in between when we were triggered, and when we
5921 * actually got to run. */
5925 /* We're in loss recovery */
5926 call->flags |= RX_CALL_FAST_RECOVER;
5928 /* Mark all of the pending packets in the queue as being lost */
5929 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5930 if (!(p->flags & RX_PKTFLAG_ACKED))
5931 p->flags &= ~RX_PKTFLAG_SENT;
5934 /* We're resending, so we double the timeout of the call. This will be
5935 * dropped back down by the first successful ACK that we receive.
5937 * We apply a maximum value here of 60 seconds
5939 clock_Add(&call->rto, &call->rto);
5940 if (clock_Gt(&call->rto, &maxTimeout))
5941 call->rto = maxTimeout;
5943 /* Packet loss is most likely due to congestion, so drop our window size
5944 * and start again from the beginning */
5945 if (peer->maxDgramPackets >1) {
5946 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5947 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5949 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5950 call->nDgramPackets = 1;
5952 call->nextCwind = 1;
5955 MUTEX_ENTER(&peer->peer_lock);
5956 peer->MTU = call->MTU;
5957 peer->cwind = call->cwind;
5958 peer->nDgramPackets = 1;
5960 call->congestSeq = peer->congestSeq;
5961 MUTEX_EXIT(&peer->peer_lock);
5963 rxi_Start(call, istack);
5966 MUTEX_EXIT(&call->lock);
5969 /* This routine is called when new packets are readied for
5970 * transmission and when retransmission may be necessary, or when the
5971 * transmission window or burst count are favourable. This should be
5972 * better optimized for new packets, the usual case, now that we've
5973 * got rid of queues of send packets. XXXXXXXXXXX */
5975 rxi_Start(struct rx_call *call, int istack)
5978 struct rx_packet *p;
5979 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5984 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5985 if (rx_stats_active)
5986 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5991 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5993 /* Send (or resend) any packets that need it, subject to
5994 * window restrictions and congestion burst control
5995 * restrictions. Ask for an ack on the last packet sent in
5996 * this burst. For now, we're relying upon the window being
5997 * considerably bigger than the largest number of packets that
5998 * are typically sent at once by one initial call to
5999 * rxi_Start. This is probably bogus (perhaps we should ask
6000 * for an ack when we're half way through the current
6001 * window?). Also, for non file transfer applications, this
6002 * may end up asking for an ack for every packet. Bogus. XXXX
6005 * But check whether we're here recursively, and let the other guy
6008 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6009 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6010 call->flags |= RX_CALL_TQ_BUSY;
6012 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6014 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6015 call->flags &= ~RX_CALL_NEED_START;
6016 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6018 maxXmitPackets = MIN(call->twind, call->cwind);
6019 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
6020 #ifdef RX_TRACK_PACKETS
6021 if ((p->flags & RX_PKTFLAG_FREE)
6022 || (!queue_IsEnd(&call->tq, nxp)
6023 && (nxp->flags & RX_PKTFLAG_FREE))
6024 || (p == (struct rx_packet *)&rx_freePacketQueue)
6025 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
6026 osi_Panic("rxi_Start: xmit queue clobbered");
6029 if (p->flags & RX_PKTFLAG_ACKED) {
6030 /* Since we may block, don't trust this */
6031 if (rx_stats_active)
6032 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6033 continue; /* Ignore this packet if it has been acknowledged */
6036 /* Turn off all flags except these ones, which are the same
6037 * on each transmission */
6038 p->header.flags &= RX_PRESET_FLAGS;
6040 if (p->header.seq >=
6041 call->tfirst + MIN((int)call->twind,
6042 (int)(call->nSoftAcked +
6044 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6045 /* Note: if we're waiting for more window space, we can
6046 * still send retransmits; hence we don't return here, but
6047 * break out to schedule a retransmit event */
6048 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6049 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6054 /* Transmit the packet if it needs to be sent. */
6055 if (!(p->flags & RX_PKTFLAG_SENT)) {
6056 if (nXmitPackets == maxXmitPackets) {
6057 rxi_SendXmitList(call, call->xmitList,
6058 nXmitPackets, istack);
6061 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6062 *(call->callNumber), p));
6063 call->xmitList[nXmitPackets++] = p;
6067 /* xmitList now hold pointers to all of the packets that are
6068 * ready to send. Now we loop to send the packets */
6069 if (nXmitPackets > 0) {
6070 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6074 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6076 /* We went into the error state while sending packets. Now is
6077 * the time to reset the call. This will also inform the using
6078 * process that the call is in an error state.
6080 if (rx_stats_active)
6081 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6082 call->flags &= ~RX_CALL_TQ_BUSY;
6083 rxi_WakeUpTransmitQueue(call);
6084 rxi_CallError(call, call->error);
6087 #ifdef RX_ENABLE_LOCKS
6088 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6090 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6091 /* Some packets have received acks. If they all have, we can clear
6092 * the transmit queue.
6095 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
6096 if (p->header.seq < call->tfirst
6097 && (p->flags & RX_PKTFLAG_ACKED)) {
6099 #ifdef RX_TRACK_PACKETS
6100 p->flags &= ~RX_PKTFLAG_TQ;
6102 #ifdef RXDEBUG_PACKET
6110 call->flags |= RX_CALL_TQ_CLEARME;
6112 #endif /* RX_ENABLE_LOCKS */
6113 if (call->flags & RX_CALL_TQ_CLEARME)
6114 rxi_ClearTransmitQueue(call, 1);
6115 } while (call->flags & RX_CALL_NEED_START);
6117 * TQ references no longer protected by this flag; they must remain
6118 * protected by the global lock.
6120 call->flags &= ~RX_CALL_TQ_BUSY;
6121 rxi_WakeUpTransmitQueue(call);
6123 call->flags |= RX_CALL_NEED_START;
6125 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6127 rxi_rto_cancel(call);
6131 /* Also adjusts the keep alive parameters for the call, to reflect
6132 * that we have just sent a packet (so keep alives aren't sent
6135 rxi_Send(struct rx_call *call, struct rx_packet *p,
6138 struct rx_connection *conn = call->conn;
6140 /* Stamp each packet with the user supplied status */
6141 p->header.userStatus = call->localStatus;
6143 /* Allow the security object controlling this call's security to
6144 * make any last-minute changes to the packet */
6145 RXS_SendPacket(conn->securityObject, call, p);
6147 /* Since we're about to send SOME sort of packet to the peer, it's
6148 * safe to nuke any scheduled end-of-packets ack */
6149 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
6151 /* Actually send the packet, filling in more connection-specific fields */
6152 MUTEX_EXIT(&call->lock);
6153 MUTEX_ENTER(&rx_refcnt_mutex);
6154 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6155 MUTEX_EXIT(&rx_refcnt_mutex);
6156 rxi_SendPacket(call, conn, p, istack);
6157 MUTEX_ENTER(&rx_refcnt_mutex);
6158 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6159 MUTEX_EXIT(&rx_refcnt_mutex);
6160 MUTEX_ENTER(&call->lock);
6162 /* Update last send time for this call (for keep-alive
6163 * processing), and for the connection (so that we can discover
6164 * idle connections) */
6165 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6166 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6167 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6169 conn->lastSendTime = call->lastSendTime = clock_Sec();
6170 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6171 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6172 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6173 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6174 RX_ACK_PING_RESPONSE)))
6175 call->lastSendData = call->lastSendTime;
6179 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6180 * that things are fine. Also called periodically to guarantee that nothing
6181 * falls through the cracks (e.g. (error + dally) connections have keepalive
6182 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6184 * haveCTLock Set if calling from rxi_ReapConnections
6186 #ifdef RX_ENABLE_LOCKS
6188 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6189 #else /* RX_ENABLE_LOCKS */
6191 rxi_CheckCall(struct rx_call *call)
6192 #endif /* RX_ENABLE_LOCKS */
6194 struct rx_connection *conn = call->conn;
6196 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6197 afs_uint32 fudgeFactor;
6201 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6202 if (call->flags & RX_CALL_TQ_BUSY) {
6203 /* Call is active and will be reset by rxi_Start if it's
6204 * in an error state.
6209 /* RTT + 8*MDEV, rounded up to the next second. */
6210 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6211 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6213 deadTime = conn->secondsUntilDead + fudgeFactor;
6215 /* These are computed to the second (+- 1 second). But that's
6216 * good enough for these values, which should be a significant
6217 * number of seconds. */
6218 if (now > (call->lastReceiveTime + deadTime)) {
6219 if (call->state == RX_STATE_ACTIVE) {
6221 #if defined(KERNEL) && defined(AFS_SUN5_ENV)
6223 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6224 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6225 ip_stack_t *ipst = ns->netstack_ip;
6227 ire = ire_cache_lookup(conn->peer->host
6228 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6230 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
6232 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6239 if (ire && ire->ire_max_frag > 0)
6240 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6242 #if defined(GLOBAL_NETSTACKID)
6246 #endif /* ADAPT_PMTU */
6247 cerror = RX_CALL_DEAD;
6250 #ifdef RX_ENABLE_LOCKS
6251 /* Cancel pending events */
6252 rxevent_Cancel(call->delayedAckEvent, call,
6253 RX_CALL_REFCOUNT_DELAY);
6254 rxi_rto_cancel(call);
6255 rxevent_Cancel(call->keepAliveEvent, call,
6256 RX_CALL_REFCOUNT_ALIVE);
6257 if (call->growMTUEvent)
6258 rxevent_Cancel(call->growMTUEvent, call,
6259 RX_CALL_REFCOUNT_ALIVE);
6260 MUTEX_ENTER(&rx_refcnt_mutex);
6261 if (call->refCount == 0) {
6262 rxi_FreeCall(call, haveCTLock);
6263 MUTEX_EXIT(&rx_refcnt_mutex);
6266 MUTEX_EXIT(&rx_refcnt_mutex);
6268 #else /* RX_ENABLE_LOCKS */
6269 rxi_FreeCall(call, 0);
6271 #endif /* RX_ENABLE_LOCKS */
6273 /* Non-active calls are destroyed if they are not responding
6274 * to pings; active calls are simply flagged in error, so the
6275 * attached process can die reasonably gracefully. */
6278 if (conn->idleDeadTime) {
6279 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6282 /* see if we have a non-activity timeout */
6283 if (call->startWait && idleDeadTime
6284 && ((call->startWait + idleDeadTime) < now) &&
6285 (call->flags & RX_CALL_READER_WAIT)) {
6286 if (call->state == RX_STATE_ACTIVE) {
6287 cerror = RX_CALL_TIMEOUT;
6291 if (call->lastSendData && idleDeadTime && (conn->idleDeadErr != 0)
6292 && ((call->lastSendData + idleDeadTime) < now)) {
6293 if (call->state == RX_STATE_ACTIVE) {
6294 cerror = conn->idleDeadErr;
6299 if (conn->hardDeadTime) {
6300 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6303 /* see if we have a hard timeout */
6305 && (now > (hardDeadTime + call->startTime.sec))) {
6306 if (call->state == RX_STATE_ACTIVE)
6307 rxi_CallError(call, RX_CALL_TIMEOUT);
6312 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT
6313 && call->lastReceiveTime) {
6314 int oldMTU = conn->peer->ifMTU;
6316 /* if we thought we could send more, perhaps things got worse */
6317 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6318 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6319 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6320 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6322 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6324 /* minimum capped in SetPeerMtu */
6325 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6328 conn->lastPacketSize = 0;
6330 /* needed so ResetCall doesn't clobber us. */
6331 call->MTU = conn->peer->ifMTU;
6333 /* if we never succeeded, let the error pass out as-is */
6334 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6335 cerror = conn->msgsizeRetryErr;
6338 rxi_CallError(call, cerror);
6343 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6345 struct rx_connection *conn = arg1;
6346 struct rx_header theader;
6347 char tbuffer[1 + sizeof(struct rx_header)];
6348 struct sockaddr_in taddr;
6351 struct iovec tmpiov[2];
6354 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6357 tp = &tbuffer[sizeof(struct rx_header)];
6358 taddr.sin_family = AF_INET;
6359 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6360 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6361 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6362 taddr.sin_len = sizeof(struct sockaddr_in);
6364 memset(&theader, 0, sizeof(theader));
6365 theader.epoch = htonl(999);
6367 theader.callNumber = 0;
6370 theader.type = RX_PACKET_TYPE_VERSION;
6371 theader.flags = RX_LAST_PACKET;
6372 theader.serviceId = 0;
6374 memcpy(tbuffer, &theader, sizeof(theader));
6375 memcpy(tp, &a, sizeof(a));
6376 tmpiov[0].iov_base = tbuffer;
6377 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6379 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6381 MUTEX_ENTER(&conn->conn_data_lock);
6382 MUTEX_ENTER(&rx_refcnt_mutex);
6383 /* Only reschedule ourselves if the connection would not be destroyed */
6384 if (conn->refCount <= 1) {
6385 conn->natKeepAliveEvent = NULL;
6386 MUTEX_EXIT(&rx_refcnt_mutex);
6387 MUTEX_EXIT(&conn->conn_data_lock);
6388 rx_DestroyConnection(conn); /* drop the reference for this */
6390 conn->refCount--; /* drop the reference for this */
6391 MUTEX_EXIT(&rx_refcnt_mutex);
6392 conn->natKeepAliveEvent = NULL;
6393 rxi_ScheduleNatKeepAliveEvent(conn);
6394 MUTEX_EXIT(&conn->conn_data_lock);
6399 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6401 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6402 struct clock when, now;
6403 clock_GetTime(&now);
6405 when.sec += conn->secondsUntilNatPing;
6406 MUTEX_ENTER(&rx_refcnt_mutex);
6407 conn->refCount++; /* hold a reference for this */
6408 MUTEX_EXIT(&rx_refcnt_mutex);
6409 conn->natKeepAliveEvent =
6410 rxevent_PostNow(&when, &now, rxi_NatKeepAliveEvent, conn, 0);
6415 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6417 MUTEX_ENTER(&conn->conn_data_lock);
6418 conn->secondsUntilNatPing = seconds;
6420 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6421 rxi_ScheduleNatKeepAliveEvent(conn);
6423 conn->flags |= RX_CONN_NAT_PING;
6425 MUTEX_EXIT(&conn->conn_data_lock);
6429 rxi_NatKeepAliveOn(struct rx_connection *conn)
6431 MUTEX_ENTER(&conn->conn_data_lock);
6432 /* if it's already attached */
6433 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6434 rxi_ScheduleNatKeepAliveEvent(conn);
6436 conn->flags |= RX_CONN_NAT_PING;
6437 MUTEX_EXIT(&conn->conn_data_lock);
6440 /* When a call is in progress, this routine is called occasionally to
6441 * make sure that some traffic has arrived (or been sent to) the peer.
6442 * If nothing has arrived in a reasonable amount of time, the call is
6443 * declared dead; if nothing has been sent for a while, we send a
6444 * keep-alive packet (if we're actually trying to keep the call alive)
6447 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6449 struct rx_call *call = arg1;
6450 struct rx_connection *conn;
6453 MUTEX_ENTER(&rx_refcnt_mutex);
6454 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6455 MUTEX_EXIT(&rx_refcnt_mutex);
6456 MUTEX_ENTER(&call->lock);
6457 if (event == call->keepAliveEvent)
6458 call->keepAliveEvent = NULL;
6461 #ifdef RX_ENABLE_LOCKS
6462 if (rxi_CheckCall(call, 0)) {
6463 MUTEX_EXIT(&call->lock);
6466 #else /* RX_ENABLE_LOCKS */
6467 if (rxi_CheckCall(call))
6469 #endif /* RX_ENABLE_LOCKS */
6471 /* Don't try to keep alive dallying calls */
6472 if (call->state == RX_STATE_DALLY) {
6473 MUTEX_EXIT(&call->lock);
6478 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6479 /* Don't try to send keepalives if there is unacknowledged data */
6480 /* the rexmit code should be good enough, this little hack
6481 * doesn't quite work XXX */
6482 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6484 rxi_ScheduleKeepAliveEvent(call);
6485 MUTEX_EXIT(&call->lock);
6488 /* Does what's on the nameplate. */
6490 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy)
6492 struct rx_call *call = arg1;
6493 struct rx_connection *conn;
6495 MUTEX_ENTER(&rx_refcnt_mutex);
6496 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6497 MUTEX_EXIT(&rx_refcnt_mutex);
6498 MUTEX_ENTER(&call->lock);
6500 if (event == call->growMTUEvent)
6501 call->growMTUEvent = NULL;
6503 #ifdef RX_ENABLE_LOCKS
6504 if (rxi_CheckCall(call, 0)) {
6505 MUTEX_EXIT(&call->lock);
6508 #else /* RX_ENABLE_LOCKS */
6509 if (rxi_CheckCall(call))
6511 #endif /* RX_ENABLE_LOCKS */
6513 /* Don't bother with dallying calls */
6514 if (call->state == RX_STATE_DALLY) {
6515 MUTEX_EXIT(&call->lock);
6522 * keep being scheduled, just don't do anything if we're at peak,
6523 * or we're not set up to be properly handled (idle timeout required)
6525 if ((conn->peer->maxPacketSize != 0) &&
6526 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6527 (conn->idleDeadErr))
6528 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6529 rxi_ScheduleGrowMTUEvent(call, 0);
6530 MUTEX_EXIT(&call->lock);
6534 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6536 if (!call->keepAliveEvent) {
6537 struct clock when, now;
6538 clock_GetTime(&now);
6540 when.sec += call->conn->secondsUntilPing;
6541 MUTEX_ENTER(&rx_refcnt_mutex);
6542 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6543 MUTEX_EXIT(&rx_refcnt_mutex);
6544 call->keepAliveEvent =
6545 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
6550 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6552 if (!call->growMTUEvent) {
6553 struct clock when, now;
6555 clock_GetTime(&now);
6558 if (call->conn->secondsUntilPing)
6559 secs = (6*call->conn->secondsUntilPing)-1;
6561 if (call->conn->secondsUntilDead)
6562 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6566 MUTEX_ENTER(&rx_refcnt_mutex);
6567 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6568 MUTEX_EXIT(&rx_refcnt_mutex);
6569 call->growMTUEvent =
6570 rxevent_PostNow(&when, &now, rxi_GrowMTUEvent, call, 0);
6574 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6576 rxi_KeepAliveOn(struct rx_call *call)
6578 /* Pretend last packet received was received now--i.e. if another
6579 * packet isn't received within the keep alive time, then the call
6580 * will die; Initialize last send time to the current time--even
6581 * if a packet hasn't been sent yet. This will guarantee that a
6582 * keep-alive is sent within the ping time */
6583 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6584 rxi_ScheduleKeepAliveEvent(call);
6588 rxi_GrowMTUOn(struct rx_call *call)
6590 struct rx_connection *conn = call->conn;
6591 MUTEX_ENTER(&conn->conn_data_lock);
6592 conn->lastPingSizeSer = conn->lastPingSize = 0;
6593 MUTEX_EXIT(&conn->conn_data_lock);
6594 rxi_ScheduleGrowMTUEvent(call, 1);
6597 /* This routine is called to send connection abort messages
6598 * that have been delayed to throttle looping clients. */
6600 rxi_SendDelayedConnAbort(struct rxevent *event,
6601 void *arg1, void *unused)
6603 struct rx_connection *conn = arg1;
6606 struct rx_packet *packet;
6608 MUTEX_ENTER(&conn->conn_data_lock);
6609 conn->delayedAbortEvent = NULL;
6610 error = htonl(conn->error);
6612 MUTEX_EXIT(&conn->conn_data_lock);
6613 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6616 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6617 RX_PACKET_TYPE_ABORT, (char *)&error,
6619 rxi_FreePacket(packet);
6623 /* This routine is called to send call abort messages
6624 * that have been delayed to throttle looping clients. */
6626 rxi_SendDelayedCallAbort(struct rxevent *event,
6627 void *arg1, void *dummy)
6629 struct rx_call *call = arg1;
6632 struct rx_packet *packet;
6634 MUTEX_ENTER(&call->lock);
6635 call->delayedAbortEvent = NULL;
6636 error = htonl(call->error);
6638 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6641 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6642 (char *)&error, sizeof(error), 0);
6643 rxi_FreePacket(packet);
6645 MUTEX_EXIT(&call->lock);
6646 MUTEX_ENTER(&rx_refcnt_mutex);
6647 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6648 MUTEX_EXIT(&rx_refcnt_mutex);
6651 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6652 * seconds) to ask the client to authenticate itself. The routine
6653 * issues a challenge to the client, which is obtained from the
6654 * security object associated with the connection */
6656 rxi_ChallengeEvent(struct rxevent *event,
6657 void *arg0, void *arg1, int tries)
6659 struct rx_connection *conn = arg0;
6661 conn->challengeEvent = NULL;
6662 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6663 struct rx_packet *packet;
6664 struct clock when, now;
6667 /* We've failed to authenticate for too long.
6668 * Reset any calls waiting for authentication;
6669 * they are all in RX_STATE_PRECALL.
6673 MUTEX_ENTER(&conn->conn_call_lock);
6674 for (i = 0; i < RX_MAXCALLS; i++) {
6675 struct rx_call *call = conn->call[i];
6677 MUTEX_ENTER(&call->lock);
6678 if (call->state == RX_STATE_PRECALL) {
6679 rxi_CallError(call, RX_CALL_DEAD);
6680 rxi_SendCallAbort(call, NULL, 0, 0);
6682 MUTEX_EXIT(&call->lock);
6685 MUTEX_EXIT(&conn->conn_call_lock);
6689 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6691 /* If there's no packet available, do this later. */
6692 RXS_GetChallenge(conn->securityObject, conn, packet);
6693 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6694 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6695 rxi_FreePacket(packet);
6697 clock_GetTime(&now);
6699 when.sec += RX_CHALLENGE_TIMEOUT;
6700 conn->challengeEvent =
6701 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
6706 /* Call this routine to start requesting the client to authenticate
6707 * itself. This will continue until authentication is established,
6708 * the call times out, or an invalid response is returned. The
6709 * security object associated with the connection is asked to create
6710 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6711 * defined earlier. */
6713 rxi_ChallengeOn(struct rx_connection *conn)
6715 if (!conn->challengeEvent) {
6716 RXS_CreateChallenge(conn->securityObject, conn);
6717 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6722 /* rxi_ComputeRoundTripTime is called with peer locked. */
6723 /* peer may be null */
6725 rxi_ComputeRoundTripTime(struct rx_packet *p,
6726 struct rx_ackPacket *ack,
6727 struct rx_call *call,
6728 struct rx_peer *peer,
6731 struct clock thisRtt, *sentp;
6735 /* If the ACK is delayed, then do nothing */
6736 if (ack->reason == RX_ACK_DELAY)
6739 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6740 * their RTT multiple times, so only include the RTT of the last packet
6742 if (p->flags & RX_JUMBO_PACKET)
6745 /* Use the serial number to determine which transmission the ACK is for,
6746 * and set the sent time to match this. If we have no serial number, then
6747 * only use the ACK for RTT calculations if the packet has not been
6751 serial = ntohl(ack->serial);
6753 if (serial == p->header.serial) {
6754 sentp = &p->timeSent;
6755 } else if (serial == p->firstSerial) {
6756 sentp = &p->firstSent;
6757 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6758 sentp = &p->firstSent;
6762 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6763 sentp = &p->firstSent;
6770 if (clock_Lt(&thisRtt, sentp))
6771 return; /* somebody set the clock back, don't count this time. */
6773 clock_Sub(&thisRtt, sentp);
6774 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6775 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6777 if (clock_IsZero(&thisRtt)) {
6779 * The actual round trip time is shorter than the
6780 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6781 * Since we can't tell which at the moment we will assume 1ms.
6783 thisRtt.usec = 1000;
6786 if (rx_stats_active) {
6787 MUTEX_ENTER(&rx_stats_mutex);
6788 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6789 rx_stats.minRtt = thisRtt;
6790 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6791 if (thisRtt.sec > 60) {
6792 MUTEX_EXIT(&rx_stats_mutex);
6793 return; /* somebody set the clock ahead */
6795 rx_stats.maxRtt = thisRtt;
6797 clock_Add(&rx_stats.totalRtt, &thisRtt);
6798 rx_atomic_inc(&rx_stats.nRttSamples);
6799 MUTEX_EXIT(&rx_stats_mutex);
6802 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6804 /* Apply VanJacobson round-trip estimations */
6809 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6810 * srtt is stored as fixed point with 3 bits after the binary
6811 * point (i.e., scaled by 8). The following magic is
6812 * equivalent to the smoothing algorithm in rfc793 with an
6813 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6814 * srtt'*8 = rtt + srtt*7
6815 * srtt'*8 = srtt*8 + rtt - srtt
6816 * srtt' = srtt + rtt/8 - srtt/8
6817 * srtt' = srtt + (rtt - srtt)/8
6820 delta = _8THMSEC(&thisRtt) - call->rtt;
6821 call->rtt += (delta >> 3);
6824 * We accumulate a smoothed rtt variance (actually, a smoothed
6825 * mean difference), then set the retransmit timer to smoothed
6826 * rtt + 4 times the smoothed variance (was 2x in van's original
6827 * paper, but 4x works better for me, and apparently for him as
6829 * rttvar is stored as
6830 * fixed point with 2 bits after the binary point (scaled by
6831 * 4). The following is equivalent to rfc793 smoothing with
6832 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6833 * rttvar'*4 = rttvar*3 + |delta|
6834 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6835 * rttvar' = rttvar + |delta|/4 - rttvar/4
6836 * rttvar' = rttvar + (|delta| - rttvar)/4
6837 * This replaces rfc793's wired-in beta.
6838 * dev*4 = dev*4 + (|actual - expected| - dev)
6844 delta -= (call->rtt_dev << 1);
6845 call->rtt_dev += (delta >> 3);
6847 /* I don't have a stored RTT so I start with this value. Since I'm
6848 * probably just starting a call, and will be pushing more data down
6849 * this, I expect congestion to increase rapidly. So I fudge a
6850 * little, and I set deviance to half the rtt. In practice,
6851 * deviance tends to approach something a little less than
6852 * half the smoothed rtt. */
6853 call->rtt = _8THMSEC(&thisRtt) + 8;
6854 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6856 /* the smoothed RTT time is RTT + 4*MDEV
6858 * We allow a user specified minimum to be set for this, to allow clamping
6859 * at a minimum value in the same way as TCP. In addition, we have to allow
6860 * for the possibility that this packet is answered by a delayed ACK, so we
6861 * add on a fixed 200ms to account for that timer expiring.
6864 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6865 rx_minPeerTimeout) + 200;
6866 clock_Zero(&call->rto);
6867 clock_Addmsec(&call->rto, rtt_timeout);
6869 /* Update the peer, so any new calls start with our values */
6870 peer->rtt_dev = call->rtt_dev;
6871 peer->rtt = call->rtt;
6873 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6874 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6878 /* Find all server connections that have not been active for a long time, and
6881 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6883 struct clock now, when;
6884 clock_GetTime(&now);
6886 /* Find server connection structures that haven't been used for
6887 * greater than rx_idleConnectionTime */
6889 struct rx_connection **conn_ptr, **conn_end;
6890 int i, havecalls = 0;
6891 MUTEX_ENTER(&rx_connHashTable_lock);
6892 for (conn_ptr = &rx_connHashTable[0], conn_end =
6893 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6895 struct rx_connection *conn, *next;
6896 struct rx_call *call;
6900 for (conn = *conn_ptr; conn; conn = next) {
6901 /* XXX -- Shouldn't the connection be locked? */
6904 for (i = 0; i < RX_MAXCALLS; i++) {
6905 call = conn->call[i];
6909 code = MUTEX_TRYENTER(&call->lock);
6912 #ifdef RX_ENABLE_LOCKS
6913 result = rxi_CheckCall(call, 1);
6914 #else /* RX_ENABLE_LOCKS */
6915 result = rxi_CheckCall(call);
6916 #endif /* RX_ENABLE_LOCKS */
6917 MUTEX_EXIT(&call->lock);
6919 /* If CheckCall freed the call, it might
6920 * have destroyed the connection as well,
6921 * which screws up the linked lists.
6927 if (conn->type == RX_SERVER_CONNECTION) {
6928 /* This only actually destroys the connection if
6929 * there are no outstanding calls */
6930 MUTEX_ENTER(&conn->conn_data_lock);
6931 MUTEX_ENTER(&rx_refcnt_mutex);
6932 if (!havecalls && !conn->refCount
6933 && ((conn->lastSendTime + rx_idleConnectionTime) <
6935 conn->refCount++; /* it will be decr in rx_DestroyConn */
6936 MUTEX_EXIT(&rx_refcnt_mutex);
6937 MUTEX_EXIT(&conn->conn_data_lock);
6938 #ifdef RX_ENABLE_LOCKS
6939 rxi_DestroyConnectionNoLock(conn);
6940 #else /* RX_ENABLE_LOCKS */
6941 rxi_DestroyConnection(conn);
6942 #endif /* RX_ENABLE_LOCKS */
6944 #ifdef RX_ENABLE_LOCKS
6946 MUTEX_EXIT(&rx_refcnt_mutex);
6947 MUTEX_EXIT(&conn->conn_data_lock);
6949 #endif /* RX_ENABLE_LOCKS */
6953 #ifdef RX_ENABLE_LOCKS
6954 while (rx_connCleanup_list) {
6955 struct rx_connection *conn;
6956 conn = rx_connCleanup_list;
6957 rx_connCleanup_list = rx_connCleanup_list->next;
6958 MUTEX_EXIT(&rx_connHashTable_lock);
6959 rxi_CleanupConnection(conn);
6960 MUTEX_ENTER(&rx_connHashTable_lock);
6962 MUTEX_EXIT(&rx_connHashTable_lock);
6963 #endif /* RX_ENABLE_LOCKS */
6966 /* Find any peer structures that haven't been used (haven't had an
6967 * associated connection) for greater than rx_idlePeerTime */
6969 struct rx_peer **peer_ptr, **peer_end;
6973 * Why do we need to hold the rx_peerHashTable_lock across
6974 * the incrementing of peer_ptr since the rx_peerHashTable
6975 * array is not changing? We don't.
6977 * By dropping the lock periodically we can permit other
6978 * activities to be performed while a rxi_ReapConnections
6979 * call is in progress. The goal of reap connections
6980 * is to clean up quickly without causing large amounts
6981 * of contention. Therefore, it is important that global
6982 * mutexes not be held for extended periods of time.
6984 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6985 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6987 struct rx_peer *peer, *next, *prev;
6989 MUTEX_ENTER(&rx_peerHashTable_lock);
6990 for (prev = peer = *peer_ptr; peer; peer = next) {
6992 code = MUTEX_TRYENTER(&peer->peer_lock);
6993 if ((code) && (peer->refCount == 0)
6994 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6995 rx_interface_stat_p rpc_stat, nrpc_stat;
6999 * now know that this peer object is one to be
7000 * removed from the hash table. Once it is removed
7001 * it can't be referenced by other threads.
7002 * Lets remove it first and decrement the struct
7003 * nPeerStructs count.
7005 if (peer == *peer_ptr) {
7011 if (rx_stats_active)
7012 rx_atomic_dec(&rx_stats.nPeerStructs);
7015 * Now if we hold references on 'prev' and 'next'
7016 * we can safely drop the rx_peerHashTable_lock
7017 * while we destroy this 'peer' object.
7023 MUTEX_EXIT(&rx_peerHashTable_lock);
7025 MUTEX_EXIT(&peer->peer_lock);
7026 MUTEX_DESTROY(&peer->peer_lock);
7028 (&peer->rpcStats, rpc_stat, nrpc_stat,
7029 rx_interface_stat)) {
7030 unsigned int num_funcs;
7033 queue_Remove(&rpc_stat->queue_header);
7034 queue_Remove(&rpc_stat->all_peers);
7035 num_funcs = rpc_stat->stats[0].func_total;
7037 sizeof(rx_interface_stat_t) +
7038 rpc_stat->stats[0].func_total *
7039 sizeof(rx_function_entry_v1_t);
7041 rxi_Free(rpc_stat, space);
7043 MUTEX_ENTER(&rx_rpc_stats);
7044 rxi_rpc_peer_stat_cnt -= num_funcs;
7045 MUTEX_EXIT(&rx_rpc_stats);
7050 * Regain the rx_peerHashTable_lock and
7051 * decrement the reference count on 'prev'
7054 MUTEX_ENTER(&rx_peerHashTable_lock);
7061 MUTEX_EXIT(&peer->peer_lock);
7066 MUTEX_EXIT(&rx_peerHashTable_lock);
7070 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7071 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7072 * GC, just below. Really, we shouldn't have to keep moving packets from
7073 * one place to another, but instead ought to always know if we can
7074 * afford to hold onto a packet in its particular use. */
7075 MUTEX_ENTER(&rx_freePktQ_lock);
7076 if (rx_waitingForPackets) {
7077 rx_waitingForPackets = 0;
7078 #ifdef RX_ENABLE_LOCKS
7079 CV_BROADCAST(&rx_waitingForPackets_cv);
7081 osi_rxWakeup(&rx_waitingForPackets);
7084 MUTEX_EXIT(&rx_freePktQ_lock);
7087 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7088 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
7092 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7093 * rx.h is sort of strange this is better. This is called with a security
7094 * object before it is discarded. Each connection using a security object has
7095 * its own refcount to the object so it won't actually be freed until the last
7096 * connection is destroyed.
7098 * This is the only rxs module call. A hold could also be written but no one
7102 rxs_Release(struct rx_securityClass *aobj)
7104 return RXS_Close(aobj);
7108 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
7109 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
7110 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
7111 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
7113 /* Adjust our estimate of the transmission rate to this peer, given
7114 * that the packet p was just acked. We can adjust peer->timeout and
7115 * call->twind. Pragmatically, this is called
7116 * only with packets of maximal length.
7117 * Called with peer and call locked.
7121 rxi_ComputeRate(struct rx_peer *peer, struct rx_call *call,
7122 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
7124 afs_int32 xferSize, xferMs;
7128 /* Count down packets */
7129 if (peer->rateFlag > 0)
7131 /* Do nothing until we're enabled */
7132 if (peer->rateFlag != 0)
7137 /* Count only when the ack seems legitimate */
7138 switch (ackReason) {
7139 case RX_ACK_REQUESTED:
7141 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
7145 case RX_ACK_PING_RESPONSE:
7146 if (p) /* want the response to ping-request, not data send */
7148 clock_GetTime(&newTO);
7149 if (clock_Gt(&newTO, &call->pingRequestTime)) {
7150 clock_Sub(&newTO, &call->pingRequestTime);
7151 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
7155 xferSize = rx_AckDataSize(rx_maxSendWindow) + RX_HEADER_SIZE;
7162 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %d.%06d, rtt %u, ps %u)\n",
7163 ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"),
7164 xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
7166 /* Track only packets that are big enough. */
7167 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
7171 /* absorb RTT data (in milliseconds) for these big packets */
7172 if (peer->smRtt == 0) {
7173 peer->smRtt = xferMs;
7175 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
7180 if (peer->countDown) {
7184 peer->countDown = 10; /* recalculate only every so often */
7186 /* In practice, we can measure only the RTT for full packets,
7187 * because of the way Rx acks the data that it receives. (If it's
7188 * smaller than a full packet, it often gets implicitly acked
7189 * either by the call response (from a server) or by the next call
7190 * (from a client), and either case confuses transmission times
7191 * with processing times.) Therefore, replace the above
7192 * more-sophisticated processing with a simpler version, where the
7193 * smoothed RTT is kept for full-size packets, and the time to
7194 * transmit a windowful of full-size packets is simply RTT *
7195 * windowSize. Again, we take two steps:
7196 - ensure the timeout is large enough for a single packet's RTT;
7197 - ensure that the window is small enough to fit in the desired timeout.*/
7199 /* First, the timeout check. */
7200 minTime = peer->smRtt;
7201 /* Get a reasonable estimate for a timeout period */
7203 newTO.sec = minTime / 1000;
7204 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
7206 /* Increase the timeout period so that we can always do at least
7207 * one packet exchange */
7208 if (clock_Gt(&newTO, &peer->timeout)) {
7210 dpf(("CONG peer %lx/%u: timeout %d.%06d ==> %ld.%06d (rtt %u)\n",
7211 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec,
7212 newTO.sec, newTO.usec, peer->smRtt));
7214 peer->timeout = newTO;
7217 /* Now, get an estimate for the transmit window size. */
7218 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
7219 /* Now, convert to the number of full packets that could fit in a
7220 * reasonable fraction of that interval */
7221 minTime /= (peer->smRtt << 1);
7222 minTime = MAX(minTime, rx_minPeerTimeout);
7223 xferSize = minTime; /* (make a copy) */
7225 /* Now clamp the size to reasonable bounds. */
7228 else if (minTime > rx_maxSendWindow)
7229 minTime = rx_maxSendWindow;
7230 /* if (minTime != peer->maxWindow) {
7231 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u)\n",
7232 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
7233 peer->timeout.sec, peer->timeout.usec, peer->smRtt));
7234 peer->maxWindow = minTime;
7235 elide... call->twind = minTime;
7239 /* Cut back on the peer timeout if it had earlier grown unreasonably.
7240 * Discern this by calculating the timeout necessary for rx_Window
7242 if ((xferSize > rx_maxSendWindow) && (peer->timeout.sec >= 3)) {
7243 /* calculate estimate for transmission interval in milliseconds */
7244 minTime = rx_maxSendWindow * peer->smRtt;
7245 if (minTime < 1000) {
7246 dpf(("CONG peer %lx/%u: cut TO %d.%06d by 0.5 (rtt %u)\n",
7247 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
7248 peer->timeout.usec, peer->smRtt));
7250 newTO.sec = 0; /* cut back on timeout by half a second */
7251 newTO.usec = 500000;
7252 clock_Sub(&peer->timeout, &newTO);
7257 } /* end of rxi_ComputeRate */
7258 #endif /* ADAPT_WINDOW */
7266 #define TRACE_OPTION_RX_DEBUG 16
7274 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7275 0, KEY_QUERY_VALUE, &parmKey);
7276 if (code != ERROR_SUCCESS)
7279 dummyLen = sizeof(TraceOption);
7280 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7281 (BYTE *) &TraceOption, &dummyLen);
7282 if (code == ERROR_SUCCESS) {
7283 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7285 RegCloseKey (parmKey);
7286 #endif /* AFS_NT40_ENV */
7291 rx_DebugOnOff(int on)
7295 rxdebug_active = on;
7301 rx_StatsOnOff(int on)
7303 rx_stats_active = on;
7307 /* Don't call this debugging routine directly; use dpf */
7309 rxi_DebugPrint(char *format, ...)
7318 va_start(ap, format);
7320 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7323 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7325 OutputDebugString(msg);
7331 va_start(ap, format);
7333 clock_GetTime(&now);
7334 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7335 (unsigned int)now.usec);
7336 vfprintf(rx_Log, format, ap);
7344 * This function is used to process the rx_stats structure that is local
7345 * to a process as well as an rx_stats structure received from a remote
7346 * process (via rxdebug). Therefore, it needs to do minimal version
7350 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7351 afs_int32 freePackets, char version)
7355 if (size != sizeof(struct rx_statistics)) {
7357 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7358 size, sizeof(struct rx_statistics));
7361 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7364 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7365 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7366 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7367 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7368 s->specialPktAllocFailures);
7370 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7371 s->receivePktAllocFailures, s->sendPktAllocFailures,
7372 s->specialPktAllocFailures);
7376 " greedy %u, " "bogusReads %u (last from host %x), "
7377 "noPackets %u, " "noBuffers %u, " "selects %u, "
7378 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7379 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7380 s->selects, s->sendSelects);
7382 fprintf(file, " packets read: ");
7383 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7384 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7386 fprintf(file, "\n");
7389 " other read counters: data %u, " "ack %u, " "dup %u "
7390 "spurious %u " "dally %u\n", s->dataPacketsRead,
7391 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7392 s->ignorePacketDally);
7394 fprintf(file, " packets sent: ");
7395 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7396 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7398 fprintf(file, "\n");
7401 " other send counters: ack %u, " "data %u (not resends), "
7402 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7403 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7404 s->dataPacketsPushed, s->ignoreAckedPacket);
7407 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7408 s->netSendFailures, (int)s->fatalErrors);
7410 if (s->nRttSamples) {
7411 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7412 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7414 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7415 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7419 " %d server connections, " "%d client connections, "
7420 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7421 s->nServerConns, s->nClientConns, s->nPeerStructs,
7422 s->nCallStructs, s->nFreeCallStructs);
7424 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7425 fprintf(file, " %d clock updates\n", clock_nUpdates);
7429 /* for backward compatibility */
7431 rx_PrintStats(FILE * file)
7433 MUTEX_ENTER(&rx_stats_mutex);
7434 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7435 sizeof(rx_stats), rx_nFreePackets,
7437 MUTEX_EXIT(&rx_stats_mutex);
7441 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7443 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
7444 ntohl(peer->host), (int)ntohs(peer->port), (int)peer->burstSize,
7445 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
7448 " Rtt %d, " "total sent %d, " "resent %d\n",
7449 peer->rtt, peer->nSent, peer->reSends);
7452 " Packet size %d, " "max in packet skew %d, "
7453 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
7454 (int)peer->outPacketSkew);
7458 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7460 * This mutex protects the following static variables:
7464 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7465 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7467 #define LOCK_RX_DEBUG
7468 #define UNLOCK_RX_DEBUG
7469 #endif /* AFS_PTHREAD_ENV */
7471 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7473 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7474 u_char type, void *inputData, size_t inputLength,
7475 void *outputData, size_t outputLength)
7477 static afs_int32 counter = 100;
7478 time_t waitTime, waitCount;
7479 struct rx_header theader;
7482 struct timeval tv_now, tv_wake, tv_delta;
7483 struct sockaddr_in taddr, faddr;
7497 tp = &tbuffer[sizeof(struct rx_header)];
7498 taddr.sin_family = AF_INET;
7499 taddr.sin_port = remotePort;
7500 taddr.sin_addr.s_addr = remoteAddr;
7501 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7502 taddr.sin_len = sizeof(struct sockaddr_in);
7505 memset(&theader, 0, sizeof(theader));
7506 theader.epoch = htonl(999);
7508 theader.callNumber = htonl(counter);
7511 theader.type = type;
7512 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7513 theader.serviceId = 0;
7515 memcpy(tbuffer, &theader, sizeof(theader));
7516 memcpy(tp, inputData, inputLength);
7518 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7519 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7521 /* see if there's a packet available */
7522 gettimeofday(&tv_wake, NULL);
7523 tv_wake.tv_sec += waitTime;
7526 FD_SET(socket, &imask);
7527 tv_delta.tv_sec = tv_wake.tv_sec;
7528 tv_delta.tv_usec = tv_wake.tv_usec;
7529 gettimeofday(&tv_now, NULL);
7531 if (tv_delta.tv_usec < tv_now.tv_usec) {
7533 tv_delta.tv_usec += 1000000;
7536 tv_delta.tv_usec -= tv_now.tv_usec;
7538 if (tv_delta.tv_sec < tv_now.tv_sec) {
7542 tv_delta.tv_sec -= tv_now.tv_sec;
7545 code = select(0, &imask, 0, 0, &tv_delta);
7546 #else /* AFS_NT40_ENV */
7547 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7548 #endif /* AFS_NT40_ENV */
7549 if (code == 1 && FD_ISSET(socket, &imask)) {
7550 /* now receive a packet */
7551 faddrLen = sizeof(struct sockaddr_in);
7553 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7554 (struct sockaddr *)&faddr, &faddrLen);
7557 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7558 if (counter == ntohl(theader.callNumber))
7566 /* see if we've timed out */
7574 code -= sizeof(struct rx_header);
7575 if (code > outputLength)
7576 code = outputLength;
7577 memcpy(outputData, tp, code);
7580 #endif /* RXDEBUG */
7583 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7584 afs_uint16 remotePort, struct rx_debugStats * stat,
7585 afs_uint32 * supportedValues)
7587 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7589 struct rx_debugIn in;
7591 *supportedValues = 0;
7592 in.type = htonl(RX_DEBUGI_GETSTATS);
7595 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7596 &in, sizeof(in), stat, sizeof(*stat));
7599 * If the call was successful, fixup the version and indicate
7600 * what contents of the stat structure are valid.
7601 * Also do net to host conversion of fields here.
7605 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7606 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7608 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7609 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7611 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7612 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7614 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7615 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7617 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7618 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7620 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7621 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7623 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7624 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7626 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7627 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7629 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7630 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7632 stat->nFreePackets = ntohl(stat->nFreePackets);
7633 stat->packetReclaims = ntohl(stat->packetReclaims);
7634 stat->callsExecuted = ntohl(stat->callsExecuted);
7635 stat->nWaiting = ntohl(stat->nWaiting);
7636 stat->idleThreads = ntohl(stat->idleThreads);
7637 stat->nWaited = ntohl(stat->nWaited);
7638 stat->nPackets = ntohl(stat->nPackets);
7647 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7648 afs_uint16 remotePort, struct rx_statistics * stat,
7649 afs_uint32 * supportedValues)
7651 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7653 struct rx_debugIn in;
7654 afs_int32 *lp = (afs_int32 *) stat;
7658 * supportedValues is currently unused, but added to allow future
7659 * versioning of this function.
7662 *supportedValues = 0;
7663 in.type = htonl(RX_DEBUGI_RXSTATS);
7665 memset(stat, 0, sizeof(*stat));
7667 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7668 &in, sizeof(in), stat, sizeof(*stat));
7673 * Do net to host conversion here
7676 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7687 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7688 afs_uint16 remotePort, size_t version_length,
7691 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7693 return MakeDebugCall(socket, remoteAddr, remotePort,
7694 RX_PACKET_TYPE_VERSION, a, 1, version,
7702 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7703 afs_uint16 remotePort, afs_int32 * nextConnection,
7704 int allConnections, afs_uint32 debugSupportedValues,
7705 struct rx_debugConn * conn,
7706 afs_uint32 * supportedValues)
7708 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7710 struct rx_debugIn in;
7714 * supportedValues is currently unused, but added to allow future
7715 * versioning of this function.
7718 *supportedValues = 0;
7719 if (allConnections) {
7720 in.type = htonl(RX_DEBUGI_GETALLCONN);
7722 in.type = htonl(RX_DEBUGI_GETCONN);
7724 in.index = htonl(*nextConnection);
7725 memset(conn, 0, sizeof(*conn));
7727 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7728 &in, sizeof(in), conn, sizeof(*conn));
7731 *nextConnection += 1;
7734 * Convert old connection format to new structure.
7737 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7738 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7739 #define MOVEvL(a) (conn->a = vL->a)
7741 /* any old or unrecognized version... */
7742 for (i = 0; i < RX_MAXCALLS; i++) {
7743 MOVEvL(callState[i]);
7744 MOVEvL(callMode[i]);
7745 MOVEvL(callFlags[i]);
7746 MOVEvL(callOther[i]);
7748 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7749 MOVEvL(secStats.type);
7750 MOVEvL(secStats.level);
7751 MOVEvL(secStats.flags);
7752 MOVEvL(secStats.expires);
7753 MOVEvL(secStats.packetsReceived);
7754 MOVEvL(secStats.packetsSent);
7755 MOVEvL(secStats.bytesReceived);
7756 MOVEvL(secStats.bytesSent);
7761 * Do net to host conversion here
7763 * I don't convert host or port since we are most likely
7764 * going to want these in NBO.
7766 conn->cid = ntohl(conn->cid);
7767 conn->serial = ntohl(conn->serial);
7768 for (i = 0; i < RX_MAXCALLS; i++) {
7769 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7771 conn->error = ntohl(conn->error);
7772 conn->secStats.flags = ntohl(conn->secStats.flags);
7773 conn->secStats.expires = ntohl(conn->secStats.expires);
7774 conn->secStats.packetsReceived =
7775 ntohl(conn->secStats.packetsReceived);
7776 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7777 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7778 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7779 conn->epoch = ntohl(conn->epoch);
7780 conn->natMTU = ntohl(conn->natMTU);
7789 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7790 afs_uint16 remotePort, afs_int32 * nextPeer,
7791 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7792 afs_uint32 * supportedValues)
7794 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7796 struct rx_debugIn in;
7799 * supportedValues is currently unused, but added to allow future
7800 * versioning of this function.
7803 *supportedValues = 0;
7804 in.type = htonl(RX_DEBUGI_GETPEER);
7805 in.index = htonl(*nextPeer);
7806 memset(peer, 0, sizeof(*peer));
7808 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7809 &in, sizeof(in), peer, sizeof(*peer));
7815 * Do net to host conversion here
7817 * I don't convert host or port since we are most likely
7818 * going to want these in NBO.
7820 peer->ifMTU = ntohs(peer->ifMTU);
7821 peer->idleWhen = ntohl(peer->idleWhen);
7822 peer->refCount = ntohs(peer->refCount);
7823 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7824 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7825 peer->rtt = ntohl(peer->rtt);
7826 peer->rtt_dev = ntohl(peer->rtt_dev);
7827 peer->timeout.sec = 0;
7828 peer->timeout.usec = 0;
7829 peer->nSent = ntohl(peer->nSent);
7830 peer->reSends = ntohl(peer->reSends);
7831 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7832 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7833 peer->rateFlag = ntohl(peer->rateFlag);
7834 peer->natMTU = ntohs(peer->natMTU);
7835 peer->maxMTU = ntohs(peer->maxMTU);
7836 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7837 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7838 peer->MTU = ntohs(peer->MTU);
7839 peer->cwind = ntohs(peer->cwind);
7840 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7841 peer->congestSeq = ntohs(peer->congestSeq);
7842 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7843 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7844 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7845 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7854 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7855 struct rx_debugPeer * peerStats)
7858 afs_int32 error = 1; /* default to "did not succeed" */
7859 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7861 MUTEX_ENTER(&rx_peerHashTable_lock);
7862 for(tp = rx_peerHashTable[hashValue];
7863 tp != NULL; tp = tp->next) {
7864 if (tp->host == peerHost)
7870 MUTEX_EXIT(&rx_peerHashTable_lock);
7874 MUTEX_ENTER(&tp->peer_lock);
7875 peerStats->host = tp->host;
7876 peerStats->port = tp->port;
7877 peerStats->ifMTU = tp->ifMTU;
7878 peerStats->idleWhen = tp->idleWhen;
7879 peerStats->refCount = tp->refCount;
7880 peerStats->burstSize = tp->burstSize;
7881 peerStats->burst = tp->burst;
7882 peerStats->burstWait.sec = tp->burstWait.sec;
7883 peerStats->burstWait.usec = tp->burstWait.usec;
7884 peerStats->rtt = tp->rtt;
7885 peerStats->rtt_dev = tp->rtt_dev;
7886 peerStats->timeout.sec = 0;
7887 peerStats->timeout.usec = 0;
7888 peerStats->nSent = tp->nSent;
7889 peerStats->reSends = tp->reSends;
7890 peerStats->inPacketSkew = tp->inPacketSkew;
7891 peerStats->outPacketSkew = tp->outPacketSkew;
7892 peerStats->rateFlag = tp->rateFlag;
7893 peerStats->natMTU = tp->natMTU;
7894 peerStats->maxMTU = tp->maxMTU;
7895 peerStats->maxDgramPackets = tp->maxDgramPackets;
7896 peerStats->ifDgramPackets = tp->ifDgramPackets;
7897 peerStats->MTU = tp->MTU;
7898 peerStats->cwind = tp->cwind;
7899 peerStats->nDgramPackets = tp->nDgramPackets;
7900 peerStats->congestSeq = tp->congestSeq;
7901 peerStats->bytesSent.high = tp->bytesSent.high;
7902 peerStats->bytesSent.low = tp->bytesSent.low;
7903 peerStats->bytesReceived.high = tp->bytesReceived.high;
7904 peerStats->bytesReceived.low = tp->bytesReceived.low;
7905 MUTEX_EXIT(&tp->peer_lock);
7907 MUTEX_ENTER(&rx_peerHashTable_lock);
7910 MUTEX_EXIT(&rx_peerHashTable_lock);
7918 struct rx_serverQueueEntry *np;
7921 struct rx_call *call;
7922 struct rx_serverQueueEntry *sq;
7926 if (rxinit_status == 1) {
7928 return; /* Already shutdown. */
7932 #ifndef AFS_PTHREAD_ENV
7933 FD_ZERO(&rx_selectMask);
7934 #endif /* AFS_PTHREAD_ENV */
7935 rxi_dataQuota = RX_MAX_QUOTA;
7936 #ifndef AFS_PTHREAD_ENV
7938 #endif /* AFS_PTHREAD_ENV */
7941 #ifndef AFS_PTHREAD_ENV
7942 #ifndef AFS_USE_GETTIMEOFDAY
7944 #endif /* AFS_USE_GETTIMEOFDAY */
7945 #endif /* AFS_PTHREAD_ENV */
7947 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7948 call = queue_First(&rx_freeCallQueue, rx_call);
7950 rxi_Free(call, sizeof(struct rx_call));
7953 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7954 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7960 struct rx_peer **peer_ptr, **peer_end;
7961 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7962 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7964 struct rx_peer *peer, *next;
7966 MUTEX_ENTER(&rx_peerHashTable_lock);
7967 for (peer = *peer_ptr; peer; peer = next) {
7968 rx_interface_stat_p rpc_stat, nrpc_stat;
7971 MUTEX_ENTER(&rx_rpc_stats);
7972 MUTEX_ENTER(&peer->peer_lock);
7974 (&peer->rpcStats, rpc_stat, nrpc_stat,
7975 rx_interface_stat)) {
7976 unsigned int num_funcs;
7979 queue_Remove(&rpc_stat->queue_header);
7980 queue_Remove(&rpc_stat->all_peers);
7981 num_funcs = rpc_stat->stats[0].func_total;
7983 sizeof(rx_interface_stat_t) +
7984 rpc_stat->stats[0].func_total *
7985 sizeof(rx_function_entry_v1_t);
7987 rxi_Free(rpc_stat, space);
7989 /* rx_rpc_stats must be held */
7990 rxi_rpc_peer_stat_cnt -= num_funcs;
7992 MUTEX_EXIT(&peer->peer_lock);
7993 MUTEX_EXIT(&rx_rpc_stats);
7997 if (rx_stats_active)
7998 rx_atomic_dec(&rx_stats.nPeerStructs);
8000 MUTEX_EXIT(&rx_peerHashTable_lock);
8003 for (i = 0; i < RX_MAX_SERVICES; i++) {
8005 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
8007 for (i = 0; i < rx_hashTableSize; i++) {
8008 struct rx_connection *tc, *ntc;
8009 MUTEX_ENTER(&rx_connHashTable_lock);
8010 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
8012 for (j = 0; j < RX_MAXCALLS; j++) {
8014 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
8017 rxi_Free(tc, sizeof(*tc));
8019 MUTEX_EXIT(&rx_connHashTable_lock);
8022 MUTEX_ENTER(&freeSQEList_lock);
8024 while ((np = rx_FreeSQEList)) {
8025 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
8026 MUTEX_DESTROY(&np->lock);
8027 rxi_Free(np, sizeof(*np));
8030 MUTEX_EXIT(&freeSQEList_lock);
8031 MUTEX_DESTROY(&freeSQEList_lock);
8032 MUTEX_DESTROY(&rx_freeCallQueue_lock);
8033 MUTEX_DESTROY(&rx_connHashTable_lock);
8034 MUTEX_DESTROY(&rx_peerHashTable_lock);
8035 MUTEX_DESTROY(&rx_serverPool_lock);
8037 osi_Free(rx_connHashTable,
8038 rx_hashTableSize * sizeof(struct rx_connection *));
8039 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8041 UNPIN(rx_connHashTable,
8042 rx_hashTableSize * sizeof(struct rx_connection *));
8043 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8045 rxi_FreeAllPackets();
8047 MUTEX_ENTER(&rx_quota_mutex);
8048 rxi_dataQuota = RX_MAX_QUOTA;
8049 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
8050 MUTEX_EXIT(&rx_quota_mutex);
8055 #ifdef RX_ENABLE_LOCKS
8057 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
8059 if (!MUTEX_ISMINE(lockaddr))
8060 osi_Panic("Lock not held: %s", msg);
8062 #endif /* RX_ENABLE_LOCKS */
8067 * Routines to implement connection specific data.
8071 rx_KeyCreate(rx_destructor_t rtn)
8074 MUTEX_ENTER(&rxi_keyCreate_lock);
8075 key = rxi_keyCreate_counter++;
8076 rxi_keyCreate_destructor = (rx_destructor_t *)
8077 realloc((void *)rxi_keyCreate_destructor,
8078 (key + 1) * sizeof(rx_destructor_t));
8079 rxi_keyCreate_destructor[key] = rtn;
8080 MUTEX_EXIT(&rxi_keyCreate_lock);
8085 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
8088 MUTEX_ENTER(&conn->conn_data_lock);
8089 if (!conn->specific) {
8090 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
8091 for (i = 0; i < key; i++)
8092 conn->specific[i] = NULL;
8093 conn->nSpecific = key + 1;
8094 conn->specific[key] = ptr;
8095 } else if (key >= conn->nSpecific) {
8096 conn->specific = (void **)
8097 realloc(conn->specific, (key + 1) * sizeof(void *));
8098 for (i = conn->nSpecific; i < key; i++)
8099 conn->specific[i] = NULL;
8100 conn->nSpecific = key + 1;
8101 conn->specific[key] = ptr;
8103 if (conn->specific[key] && rxi_keyCreate_destructor[key])
8104 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
8105 conn->specific[key] = ptr;
8107 MUTEX_EXIT(&conn->conn_data_lock);
8111 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
8114 MUTEX_ENTER(&svc->svc_data_lock);
8115 if (!svc->specific) {
8116 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
8117 for (i = 0; i < key; i++)
8118 svc->specific[i] = NULL;
8119 svc->nSpecific = key + 1;
8120 svc->specific[key] = ptr;
8121 } else if (key >= svc->nSpecific) {
8122 svc->specific = (void **)
8123 realloc(svc->specific, (key + 1) * sizeof(void *));
8124 for (i = svc->nSpecific; i < key; i++)
8125 svc->specific[i] = NULL;
8126 svc->nSpecific = key + 1;
8127 svc->specific[key] = ptr;
8129 if (svc->specific[key] && rxi_keyCreate_destructor[key])
8130 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
8131 svc->specific[key] = ptr;
8133 MUTEX_EXIT(&svc->svc_data_lock);
8137 rx_GetSpecific(struct rx_connection *conn, int key)
8140 MUTEX_ENTER(&conn->conn_data_lock);
8141 if (key >= conn->nSpecific)
8144 ptr = conn->specific[key];
8145 MUTEX_EXIT(&conn->conn_data_lock);
8150 rx_GetServiceSpecific(struct rx_service *svc, int key)
8153 MUTEX_ENTER(&svc->svc_data_lock);
8154 if (key >= svc->nSpecific)
8157 ptr = svc->specific[key];
8158 MUTEX_EXIT(&svc->svc_data_lock);
8163 #endif /* !KERNEL */
8166 * processStats is a queue used to store the statistics for the local
8167 * process. Its contents are similar to the contents of the rpcStats
8168 * queue on a rx_peer structure, but the actual data stored within
8169 * this queue contains totals across the lifetime of the process (assuming
8170 * the stats have not been reset) - unlike the per peer structures
8171 * which can come and go based upon the peer lifetime.
8174 static struct rx_queue processStats = { &processStats, &processStats };
8177 * peerStats is a queue used to store the statistics for all peer structs.
8178 * Its contents are the union of all the peer rpcStats queues.
8181 static struct rx_queue peerStats = { &peerStats, &peerStats };
8184 * rxi_monitor_processStats is used to turn process wide stat collection
8188 static int rxi_monitor_processStats = 0;
8191 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8194 static int rxi_monitor_peerStats = 0;
8197 * rxi_AddRpcStat - given all of the information for a particular rpc
8198 * call, create (if needed) and update the stat totals for the rpc.
8202 * IN stats - the queue of stats that will be updated with the new value
8204 * IN rxInterface - a unique number that identifies the rpc interface
8206 * IN currentFunc - the index of the function being invoked
8208 * IN totalFunc - the total number of functions in this interface
8210 * IN queueTime - the amount of time this function waited for a thread
8212 * IN execTime - the amount of time this function invocation took to execute
8214 * IN bytesSent - the number bytes sent by this invocation
8216 * IN bytesRcvd - the number bytes received by this invocation
8218 * IN isServer - if true, this invocation was made to a server
8220 * IN remoteHost - the ip address of the remote host
8222 * IN remotePort - the port of the remote host
8224 * IN addToPeerList - if != 0, add newly created stat to the global peer list
8226 * INOUT counter - if a new stats structure is allocated, the counter will
8227 * be updated with the new number of allocated stat structures
8235 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
8236 afs_uint32 currentFunc, afs_uint32 totalFunc,
8237 struct clock *queueTime, struct clock *execTime,
8238 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
8239 afs_uint32 remoteHost, afs_uint32 remotePort,
8240 int addToPeerList, unsigned int *counter)
8243 rx_interface_stat_p rpc_stat, nrpc_stat;
8246 * See if there's already a structure for this interface
8249 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8250 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8251 && (rpc_stat->stats[0].remote_is_server == isServer))
8256 * Didn't find a match so allocate a new structure and add it to the
8260 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
8261 || (rpc_stat->stats[0].interfaceId != rxInterface)
8262 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8267 sizeof(rx_interface_stat_t) +
8268 totalFunc * sizeof(rx_function_entry_v1_t);
8270 rpc_stat = rxi_Alloc(space);
8271 if (rpc_stat == NULL) {
8275 *counter += totalFunc;
8276 for (i = 0; i < totalFunc; i++) {
8277 rpc_stat->stats[i].remote_peer = remoteHost;
8278 rpc_stat->stats[i].remote_port = remotePort;
8279 rpc_stat->stats[i].remote_is_server = isServer;
8280 rpc_stat->stats[i].interfaceId = rxInterface;
8281 rpc_stat->stats[i].func_total = totalFunc;
8282 rpc_stat->stats[i].func_index = i;
8283 hzero(rpc_stat->stats[i].invocations);
8284 hzero(rpc_stat->stats[i].bytes_sent);
8285 hzero(rpc_stat->stats[i].bytes_rcvd);
8286 rpc_stat->stats[i].queue_time_sum.sec = 0;
8287 rpc_stat->stats[i].queue_time_sum.usec = 0;
8288 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8289 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8290 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8291 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8292 rpc_stat->stats[i].queue_time_max.sec = 0;
8293 rpc_stat->stats[i].queue_time_max.usec = 0;
8294 rpc_stat->stats[i].execution_time_sum.sec = 0;
8295 rpc_stat->stats[i].execution_time_sum.usec = 0;
8296 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8297 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8298 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8299 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8300 rpc_stat->stats[i].execution_time_max.sec = 0;
8301 rpc_stat->stats[i].execution_time_max.usec = 0;
8303 queue_Prepend(stats, rpc_stat);
8304 if (addToPeerList) {
8305 queue_Prepend(&peerStats, &rpc_stat->all_peers);
8310 * Increment the stats for this function
8313 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
8314 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
8315 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
8316 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8317 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8318 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8319 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8321 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8322 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8324 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8325 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8327 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8328 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8330 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8331 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8339 * rx_IncrementTimeAndCount - increment the times and count for a particular
8344 * IN peer - the peer who invoked the rpc
8346 * IN rxInterface - a unique number that identifies the rpc interface
8348 * IN currentFunc - the index of the function being invoked
8350 * IN totalFunc - the total number of functions in this interface
8352 * IN queueTime - the amount of time this function waited for a thread
8354 * IN execTime - the amount of time this function invocation took to execute
8356 * IN bytesSent - the number bytes sent by this invocation
8358 * IN bytesRcvd - the number bytes received by this invocation
8360 * IN isServer - if true, this invocation was made to a server
8368 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8369 afs_uint32 currentFunc, afs_uint32 totalFunc,
8370 struct clock *queueTime, struct clock *execTime,
8371 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8375 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8378 MUTEX_ENTER(&rx_rpc_stats);
8380 if (rxi_monitor_peerStats) {
8381 MUTEX_ENTER(&peer->peer_lock);
8382 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8383 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8384 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8385 MUTEX_EXIT(&peer->peer_lock);
8388 if (rxi_monitor_processStats) {
8389 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8390 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8391 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8394 MUTEX_EXIT(&rx_rpc_stats);
8399 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8403 * IN callerVersion - the rpc stat version of the caller.
8405 * IN count - the number of entries to marshall.
8407 * IN stats - pointer to stats to be marshalled.
8409 * OUT ptr - Where to store the marshalled data.
8416 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8417 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8423 * We only support the first version
8425 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8426 *(ptr++) = stats->remote_peer;
8427 *(ptr++) = stats->remote_port;
8428 *(ptr++) = stats->remote_is_server;
8429 *(ptr++) = stats->interfaceId;
8430 *(ptr++) = stats->func_total;
8431 *(ptr++) = stats->func_index;
8432 *(ptr++) = hgethi(stats->invocations);
8433 *(ptr++) = hgetlo(stats->invocations);
8434 *(ptr++) = hgethi(stats->bytes_sent);
8435 *(ptr++) = hgetlo(stats->bytes_sent);
8436 *(ptr++) = hgethi(stats->bytes_rcvd);
8437 *(ptr++) = hgetlo(stats->bytes_rcvd);
8438 *(ptr++) = stats->queue_time_sum.sec;
8439 *(ptr++) = stats->queue_time_sum.usec;
8440 *(ptr++) = stats->queue_time_sum_sqr.sec;
8441 *(ptr++) = stats->queue_time_sum_sqr.usec;
8442 *(ptr++) = stats->queue_time_min.sec;
8443 *(ptr++) = stats->queue_time_min.usec;
8444 *(ptr++) = stats->queue_time_max.sec;
8445 *(ptr++) = stats->queue_time_max.usec;
8446 *(ptr++) = stats->execution_time_sum.sec;
8447 *(ptr++) = stats->execution_time_sum.usec;
8448 *(ptr++) = stats->execution_time_sum_sqr.sec;
8449 *(ptr++) = stats->execution_time_sum_sqr.usec;
8450 *(ptr++) = stats->execution_time_min.sec;
8451 *(ptr++) = stats->execution_time_min.usec;
8452 *(ptr++) = stats->execution_time_max.sec;
8453 *(ptr++) = stats->execution_time_max.usec;
8459 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8464 * IN callerVersion - the rpc stat version of the caller
8466 * OUT myVersion - the rpc stat version of this function
8468 * OUT clock_sec - local time seconds
8470 * OUT clock_usec - local time microseconds
8472 * OUT allocSize - the number of bytes allocated to contain stats
8474 * OUT statCount - the number stats retrieved from this process.
8476 * OUT stats - the actual stats retrieved from this process.
8480 * Returns void. If successful, stats will != NULL.
8484 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8485 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8486 size_t * allocSize, afs_uint32 * statCount,
8487 afs_uint32 ** stats)
8497 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8500 * Check to see if stats are enabled
8503 MUTEX_ENTER(&rx_rpc_stats);
8504 if (!rxi_monitor_processStats) {
8505 MUTEX_EXIT(&rx_rpc_stats);
8509 clock_GetTime(&now);
8510 *clock_sec = now.sec;
8511 *clock_usec = now.usec;
8514 * Allocate the space based upon the caller version
8516 * If the client is at an older version than we are,
8517 * we return the statistic data in the older data format, but
8518 * we still return our version number so the client knows we
8519 * are maintaining more data than it can retrieve.
8522 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8523 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8524 *statCount = rxi_rpc_process_stat_cnt;
8527 * This can't happen yet, but in the future version changes
8528 * can be handled by adding additional code here
8532 if (space > (size_t) 0) {
8534 ptr = *stats = rxi_Alloc(space);
8537 rx_interface_stat_p rpc_stat, nrpc_stat;
8541 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8543 * Copy the data based upon the caller version
8545 rx_MarshallProcessRPCStats(callerVersion,
8546 rpc_stat->stats[0].func_total,
8547 rpc_stat->stats, &ptr);
8553 MUTEX_EXIT(&rx_rpc_stats);
8558 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8562 * IN callerVersion - the rpc stat version of the caller
8564 * OUT myVersion - the rpc stat version of this function
8566 * OUT clock_sec - local time seconds
8568 * OUT clock_usec - local time microseconds
8570 * OUT allocSize - the number of bytes allocated to contain stats
8572 * OUT statCount - the number of stats retrieved from the individual
8575 * OUT stats - the actual stats retrieved from the individual peer structures.
8579 * Returns void. If successful, stats will != NULL.
8583 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8584 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8585 size_t * allocSize, afs_uint32 * statCount,
8586 afs_uint32 ** stats)
8596 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8599 * Check to see if stats are enabled
8602 MUTEX_ENTER(&rx_rpc_stats);
8603 if (!rxi_monitor_peerStats) {
8604 MUTEX_EXIT(&rx_rpc_stats);
8608 clock_GetTime(&now);
8609 *clock_sec = now.sec;
8610 *clock_usec = now.usec;
8613 * Allocate the space based upon the caller version
8615 * If the client is at an older version than we are,
8616 * we return the statistic data in the older data format, but
8617 * we still return our version number so the client knows we
8618 * are maintaining more data than it can retrieve.
8621 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8622 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8623 *statCount = rxi_rpc_peer_stat_cnt;
8626 * This can't happen yet, but in the future version changes
8627 * can be handled by adding additional code here
8631 if (space > (size_t) 0) {
8633 ptr = *stats = rxi_Alloc(space);
8636 rx_interface_stat_p rpc_stat, nrpc_stat;
8640 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8642 * We have to fix the offset of rpc_stat since we are
8643 * keeping this structure on two rx_queues. The rx_queue
8644 * package assumes that the rx_queue member is the first
8645 * member of the structure. That is, rx_queue assumes that
8646 * any one item is only on one queue at a time. We are
8647 * breaking that assumption and so we have to do a little
8648 * math to fix our pointers.
8651 fix_offset = (char *)rpc_stat;
8652 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8653 rpc_stat = (rx_interface_stat_p) fix_offset;
8656 * Copy the data based upon the caller version
8658 rx_MarshallProcessRPCStats(callerVersion,
8659 rpc_stat->stats[0].func_total,
8660 rpc_stat->stats, &ptr);
8666 MUTEX_EXIT(&rx_rpc_stats);
8671 * rx_FreeRPCStats - free memory allocated by
8672 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8676 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8677 * rx_RetrievePeerRPCStats
8679 * IN allocSize - the number of bytes in stats.
8687 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8689 rxi_Free(stats, allocSize);
8693 * rx_queryProcessRPCStats - see if process rpc stat collection is
8694 * currently enabled.
8700 * Returns 0 if stats are not enabled != 0 otherwise
8704 rx_queryProcessRPCStats(void)
8707 MUTEX_ENTER(&rx_rpc_stats);
8708 rc = rxi_monitor_processStats;
8709 MUTEX_EXIT(&rx_rpc_stats);
8714 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8720 * Returns 0 if stats are not enabled != 0 otherwise
8724 rx_queryPeerRPCStats(void)
8727 MUTEX_ENTER(&rx_rpc_stats);
8728 rc = rxi_monitor_peerStats;
8729 MUTEX_EXIT(&rx_rpc_stats);
8734 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8744 rx_enableProcessRPCStats(void)
8746 MUTEX_ENTER(&rx_rpc_stats);
8747 rx_enable_stats = 1;
8748 rxi_monitor_processStats = 1;
8749 MUTEX_EXIT(&rx_rpc_stats);
8753 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8763 rx_enablePeerRPCStats(void)
8765 MUTEX_ENTER(&rx_rpc_stats);
8766 rx_enable_stats = 1;
8767 rxi_monitor_peerStats = 1;
8768 MUTEX_EXIT(&rx_rpc_stats);
8772 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8782 rx_disableProcessRPCStats(void)
8784 rx_interface_stat_p rpc_stat, nrpc_stat;
8787 MUTEX_ENTER(&rx_rpc_stats);
8790 * Turn off process statistics and if peer stats is also off, turn
8794 rxi_monitor_processStats = 0;
8795 if (rxi_monitor_peerStats == 0) {
8796 rx_enable_stats = 0;
8799 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8800 unsigned int num_funcs = 0;
8803 queue_Remove(rpc_stat);
8804 num_funcs = rpc_stat->stats[0].func_total;
8806 sizeof(rx_interface_stat_t) +
8807 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8809 rxi_Free(rpc_stat, space);
8810 rxi_rpc_process_stat_cnt -= num_funcs;
8812 MUTEX_EXIT(&rx_rpc_stats);
8816 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8826 rx_disablePeerRPCStats(void)
8828 struct rx_peer **peer_ptr, **peer_end;
8832 * Turn off peer statistics and if process stats is also off, turn
8836 rxi_monitor_peerStats = 0;
8837 if (rxi_monitor_processStats == 0) {
8838 rx_enable_stats = 0;
8841 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8842 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8844 struct rx_peer *peer, *next, *prev;
8846 MUTEX_ENTER(&rx_peerHashTable_lock);
8847 MUTEX_ENTER(&rx_rpc_stats);
8848 for (prev = peer = *peer_ptr; peer; peer = next) {
8850 code = MUTEX_TRYENTER(&peer->peer_lock);
8852 rx_interface_stat_p rpc_stat, nrpc_stat;
8855 if (prev == *peer_ptr) {
8866 MUTEX_EXIT(&rx_peerHashTable_lock);
8869 (&peer->rpcStats, rpc_stat, nrpc_stat,
8870 rx_interface_stat)) {
8871 unsigned int num_funcs = 0;
8874 queue_Remove(&rpc_stat->queue_header);
8875 queue_Remove(&rpc_stat->all_peers);
8876 num_funcs = rpc_stat->stats[0].func_total;
8878 sizeof(rx_interface_stat_t) +
8879 rpc_stat->stats[0].func_total *
8880 sizeof(rx_function_entry_v1_t);
8882 rxi_Free(rpc_stat, space);
8883 rxi_rpc_peer_stat_cnt -= num_funcs;
8885 MUTEX_EXIT(&peer->peer_lock);
8887 MUTEX_ENTER(&rx_peerHashTable_lock);
8897 MUTEX_EXIT(&rx_rpc_stats);
8898 MUTEX_EXIT(&rx_peerHashTable_lock);
8903 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8908 * IN clearFlag - flag indicating which stats to clear
8916 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8918 rx_interface_stat_p rpc_stat, nrpc_stat;
8920 MUTEX_ENTER(&rx_rpc_stats);
8922 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8923 unsigned int num_funcs = 0, i;
8924 num_funcs = rpc_stat->stats[0].func_total;
8925 for (i = 0; i < num_funcs; i++) {
8926 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8927 hzero(rpc_stat->stats[i].invocations);
8929 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8930 hzero(rpc_stat->stats[i].bytes_sent);
8932 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8933 hzero(rpc_stat->stats[i].bytes_rcvd);
8935 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8936 rpc_stat->stats[i].queue_time_sum.sec = 0;
8937 rpc_stat->stats[i].queue_time_sum.usec = 0;
8939 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8940 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8941 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8943 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8944 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8945 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8947 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8948 rpc_stat->stats[i].queue_time_max.sec = 0;
8949 rpc_stat->stats[i].queue_time_max.usec = 0;
8951 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8952 rpc_stat->stats[i].execution_time_sum.sec = 0;
8953 rpc_stat->stats[i].execution_time_sum.usec = 0;
8955 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8956 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8957 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8959 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8960 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8961 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8963 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8964 rpc_stat->stats[i].execution_time_max.sec = 0;
8965 rpc_stat->stats[i].execution_time_max.usec = 0;
8970 MUTEX_EXIT(&rx_rpc_stats);
8974 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8979 * IN clearFlag - flag indicating which stats to clear
8987 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8989 rx_interface_stat_p rpc_stat, nrpc_stat;
8991 MUTEX_ENTER(&rx_rpc_stats);
8993 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8994 unsigned int num_funcs = 0, i;
8997 * We have to fix the offset of rpc_stat since we are
8998 * keeping this structure on two rx_queues. The rx_queue
8999 * package assumes that the rx_queue member is the first
9000 * member of the structure. That is, rx_queue assumes that
9001 * any one item is only on one queue at a time. We are
9002 * breaking that assumption and so we have to do a little
9003 * math to fix our pointers.
9006 fix_offset = (char *)rpc_stat;
9007 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
9008 rpc_stat = (rx_interface_stat_p) fix_offset;
9010 num_funcs = rpc_stat->stats[0].func_total;
9011 for (i = 0; i < num_funcs; i++) {
9012 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9013 hzero(rpc_stat->stats[i].invocations);
9015 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9016 hzero(rpc_stat->stats[i].bytes_sent);
9018 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9019 hzero(rpc_stat->stats[i].bytes_rcvd);
9021 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9022 rpc_stat->stats[i].queue_time_sum.sec = 0;
9023 rpc_stat->stats[i].queue_time_sum.usec = 0;
9025 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9026 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9027 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9029 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9030 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9031 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9033 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9034 rpc_stat->stats[i].queue_time_max.sec = 0;
9035 rpc_stat->stats[i].queue_time_max.usec = 0;
9037 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9038 rpc_stat->stats[i].execution_time_sum.sec = 0;
9039 rpc_stat->stats[i].execution_time_sum.usec = 0;
9041 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9042 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9043 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9045 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9046 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9047 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9049 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9050 rpc_stat->stats[i].execution_time_max.sec = 0;
9051 rpc_stat->stats[i].execution_time_max.usec = 0;
9056 MUTEX_EXIT(&rx_rpc_stats);
9060 * rxi_rxstat_userok points to a routine that returns 1 if the caller
9061 * is authorized to enable/disable/clear RX statistics.
9063 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
9066 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
9068 rxi_rxstat_userok = proc;
9072 rx_RxStatUserOk(struct rx_call *call)
9074 if (!rxi_rxstat_userok)
9076 return rxi_rxstat_userok(call);
9081 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
9082 * function in the MSVC runtime DLL (msvcrt.dll).
9084 * Note: the system serializes calls to this function.
9087 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
9088 DWORD reason, /* reason function is being called */
9089 LPVOID reserved) /* reserved for future use */
9092 case DLL_PROCESS_ATTACH:
9093 /* library is being attached to a process */
9097 case DLL_PROCESS_DETACH:
9104 #endif /* AFS_NT40_ENV */
9107 int rx_DumpCalls(FILE *outputFile, char *cookie)
9109 #ifdef RXDEBUG_PACKET
9110 #ifdef KDUMP_RX_LOCK
9111 struct rx_call_rx_lock *c;
9118 #define RXDPRINTF sprintf
9119 #define RXDPRINTOUT output
9121 #define RXDPRINTF fprintf
9122 #define RXDPRINTOUT outputFile
9125 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9127 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9130 for (c = rx_allCallsp; c; c = c->allNextp) {
9131 u_short rqc, tqc, iovqc;
9132 struct rx_packet *p, *np;
9134 MUTEX_ENTER(&c->lock);
9135 queue_Count(&c->rq, p, np, rx_packet, rqc);
9136 queue_Count(&c->tq, p, np, rx_packet, tqc);
9137 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
9139 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, "
9140 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9141 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9142 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9143 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
9144 #ifdef RX_ENABLE_LOCKS
9147 #ifdef RX_REFCOUNT_CHECK
9148 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9149 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9152 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,
9153 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9154 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9155 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9156 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9157 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
9158 #ifdef RX_ENABLE_LOCKS
9159 , (afs_uint32)c->refCount
9161 #ifdef RX_REFCOUNT_CHECK
9162 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9165 MUTEX_EXIT(&c->lock);
9168 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9171 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9173 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9175 #endif /* RXDEBUG_PACKET */