2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
13 #include <afs/param.h>
16 # include "afs/sysincludes.h"
17 # include "afsincludes.h"
22 # ifdef AFS_LINUX20_ENV
23 # include "h/socket.h"
25 # include "netinet/in.h"
27 # include "netinet/ip6.h"
28 # include "inet/common.h"
30 # include "inet/ip_ire.h"
32 # include "afs/afs_args.h"
33 # include "afs/afs_osi.h"
34 # ifdef RX_KERNEL_TRACE
35 # include "rx_kcommon.h"
37 # if defined(AFS_AIX_ENV)
41 # undef RXDEBUG /* turn off debugging */
43 # if defined(AFS_SGI_ENV)
44 # include "sys/debug.h"
47 # include "afs/sysincludes.h"
48 # include "afsincludes.h"
49 # endif /* !UKERNEL */
50 # include "afs/lock.h"
51 # include "rx_kmutex.h"
52 # include "rx_kernel.h"
53 # define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
54 # define AFSOP_STOP_AFS 211 /* Stop AFS process */
55 # define AFSOP_STOP_BKG 212 /* Stop BKG process */
56 extern afs_int32 afs_termState;
58 # include "sys/lockl.h"
59 # include "sys/lock_def.h"
60 # endif /* AFS_AIX41_ENV */
61 # include "afs/rxgen_consts.h"
66 # include <afs/afsutil.h>
67 # include <WINNT\afsreg.h>
76 #include "rx_atomic.h"
77 #include "rx_globals.h"
79 #include "rx_internal.h"
82 #include <afs/rxgen_consts.h>
85 #ifdef AFS_PTHREAD_ENV
87 int (*registerProgram) (pid_t, char *) = 0;
88 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
91 int (*registerProgram) (PROCESS, char *) = 0;
92 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
96 /* Local static routines */
97 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
98 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
99 struct rx_call *, struct rx_peer *,
101 static void rxi_Resend(struct rxevent *event, void *arg0, void *arg1,
104 #ifdef RX_ENABLE_LOCKS
105 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
108 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
110 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
111 rx_atomic_t rxi_start_in_error;
113 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
115 /* Constant delay time before sending an acknowledge of the last packet
116 * received. This is to avoid sending an extra acknowledge when the
117 * client is about to make another call, anyway, or the server is
120 * The lastAckDelay may not exceeed 400ms without causing peers to
121 * unecessarily timeout.
123 struct clock rx_lastAckDelay = {0, 400000};
125 /* Constant delay time before sending a soft ack when none was requested.
126 * This is to make sure we send soft acks before the sender times out,
127 * Normally we wait and send a hard ack when the receiver consumes the packet
129 * This value has been 100ms in all shipping versions of OpenAFS. Changing it
130 * will require changes to the peer's RTT calculations.
132 struct clock rx_softAckDelay = {0, 100000};
135 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
136 * currently allocated within rx. This number is used to allocate the
137 * memory required to return the statistics when queried.
138 * Protected by the rx_rpc_stats mutex.
141 static unsigned int rxi_rpc_peer_stat_cnt;
144 * rxi_rpc_process_stat_cnt counts the total number of local process stat
145 * structures currently allocated within rx. The number is used to allocate
146 * the memory required to return the statistics when queried.
147 * Protected by the rx_rpc_stats mutex.
150 static unsigned int rxi_rpc_process_stat_cnt;
153 * rxi_busyChannelError is the error to return to the application when a call
154 * channel appears busy (inferred from the receipt of RX_PACKET_TYPE_BUSY
155 * packets on the channel), and there are other call channels in the
156 * connection that are not busy. If 0, we do not return errors upon receiving
157 * busy packets; we just keep trying on the same call channel until we hit a
160 static afs_int32 rxi_busyChannelError = 0;
162 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
163 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
165 #if !defined(offsetof)
166 #include <stddef.h> /* for definition of offsetof() */
169 #ifdef RX_ENABLE_LOCKS
170 afs_kmutex_t rx_atomic_mutex;
173 /* Forward prototypes */
174 static struct rx_call * rxi_NewCall(struct rx_connection *, int);
176 #ifdef AFS_PTHREAD_ENV
179 * Use procedural initialization of mutexes/condition variables
183 extern afs_kmutex_t rx_quota_mutex;
184 extern afs_kmutex_t rx_pthread_mutex;
185 extern afs_kmutex_t rx_packets_mutex;
186 extern afs_kmutex_t rx_refcnt_mutex;
187 extern afs_kmutex_t des_init_mutex;
188 extern afs_kmutex_t des_random_mutex;
189 extern afs_kmutex_t rx_clock_mutex;
190 extern afs_kmutex_t rxi_connCacheMutex;
191 extern afs_kmutex_t rx_event_mutex;
192 extern afs_kmutex_t event_handler_mutex;
193 extern afs_kmutex_t listener_mutex;
194 extern afs_kmutex_t rx_if_init_mutex;
195 extern afs_kmutex_t rx_if_mutex;
197 extern afs_kcondvar_t rx_event_handler_cond;
198 extern afs_kcondvar_t rx_listener_cond;
200 static afs_kmutex_t epoch_mutex;
201 static afs_kmutex_t rx_init_mutex;
202 static afs_kmutex_t rx_debug_mutex;
203 static afs_kmutex_t rx_rpc_stats;
206 rxi_InitPthread(void)
208 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
209 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
210 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
211 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
212 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
213 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
214 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
215 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
216 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
217 MUTEX_INIT(&rx_event_mutex, "event", MUTEX_DEFAULT, 0);
218 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
219 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
220 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
221 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
222 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
223 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
225 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
226 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
228 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
229 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
231 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
232 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
233 #ifdef RX_ENABLE_LOCKS
236 #endif /* RX_LOCKS_DB */
237 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
238 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
240 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
242 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
244 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
246 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
247 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
248 #endif /* RX_ENABLE_LOCKS */
251 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
252 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
254 * The rx_stats_mutex mutex protects the following global variables:
255 * rxi_lowConnRefCount
256 * rxi_lowPeerRefCount
265 * The rx_quota_mutex mutex protects the following global variables:
273 * The rx_freePktQ_lock protects the following global variables:
278 * The rx_packets_mutex mutex protects the following global variables:
286 * The rx_pthread_mutex mutex protects the following global variables:
287 * rxi_fcfs_thread_num
290 #define INIT_PTHREAD_LOCKS
294 /* Variables for handling the minProcs implementation. availProcs gives the
295 * number of threads available in the pool at this moment (not counting dudes
296 * executing right now). totalMin gives the total number of procs required
297 * for handling all minProcs requests. minDeficit is a dynamic variable
298 * tracking the # of procs required to satisfy all of the remaining minProcs
300 * For fine grain locking to work, the quota check and the reservation of
301 * a server thread has to come while rxi_availProcs and rxi_minDeficit
302 * are locked. To this end, the code has been modified under #ifdef
303 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
304 * same time. A new function, ReturnToServerPool() returns the allocation.
306 * A call can be on several queue's (but only one at a time). When
307 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
308 * that no one else is touching the queue. To this end, we store the address
309 * of the queue lock in the call structure (under the call lock) when we
310 * put the call on a queue, and we clear the call_queue_lock when the
311 * call is removed from a queue (once the call lock has been obtained).
312 * This allows rxi_ResetCall to safely synchronize with others wishing
313 * to manipulate the queue.
316 #if defined(RX_ENABLE_LOCKS)
317 static afs_kmutex_t rx_rpc_stats;
320 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
321 ** pretty good that the next packet coming in is from the same connection
322 ** as the last packet, since we're send multiple packets in a transmit window.
324 struct rx_connection *rxLastConn = 0;
326 #ifdef RX_ENABLE_LOCKS
327 /* The locking hierarchy for rx fine grain locking is composed of these
330 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
331 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
332 * call->lock - locks call data fields.
333 * These are independent of each other:
334 * rx_freeCallQueue_lock
339 * serverQueueEntry->lock
340 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
342 * peer->lock - locks peer data fields.
343 * conn_data_lock - that more than one thread is not updating a conn data
344 * field at the same time.
355 * Do we need a lock to protect the peer field in the conn structure?
356 * conn->peer was previously a constant for all intents and so has no
357 * lock protecting this field. The multihomed client delta introduced
358 * a RX code change : change the peer field in the connection structure
359 * to that remote interface from which the last packet for this
360 * connection was sent out. This may become an issue if further changes
363 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
364 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
366 /* rxdb_fileID is used to identify the lock location, along with line#. */
367 static int rxdb_fileID = RXDB_FILE_RX;
368 #endif /* RX_LOCKS_DB */
369 #else /* RX_ENABLE_LOCKS */
370 #define SET_CALL_QUEUE_LOCK(C, L)
371 #define CLEAR_CALL_QUEUE_LOCK(C)
372 #endif /* RX_ENABLE_LOCKS */
373 struct rx_serverQueueEntry *rx_waitForPacket = 0;
374 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
376 /* ------------Exported Interfaces------------- */
378 /* This function allows rxkad to set the epoch to a suitably random number
379 * which rx_NewConnection will use in the future. The principle purpose is to
380 * get rxnull connections to use the same epoch as the rxkad connections do, at
381 * least once the first rxkad connection is established. This is important now
382 * that the host/port addresses aren't used in FindConnection: the uniqueness
383 * of epoch/cid matters and the start time won't do. */
385 #ifdef AFS_PTHREAD_ENV
387 * This mutex protects the following global variables:
391 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
392 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
396 #endif /* AFS_PTHREAD_ENV */
399 rx_SetEpoch(afs_uint32 epoch)
406 /* Initialize rx. A port number may be mentioned, in which case this
407 * becomes the default port number for any service installed later.
408 * If 0 is provided for the port number, a random port will be chosen
409 * by the kernel. Whether this will ever overlap anything in
410 * /etc/services is anybody's guess... Returns 0 on success, -1 on
415 int rxinit_status = 1;
416 #ifdef AFS_PTHREAD_ENV
418 * This mutex protects the following global variables:
422 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
423 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
426 #define UNLOCK_RX_INIT
430 rx_InitHost(u_int host, u_int port)
437 char *htable, *ptable;
444 if (rxinit_status == 0) {
445 tmp_status = rxinit_status;
447 return tmp_status; /* Already started; return previous error code. */
453 if (afs_winsockInit() < 0)
459 * Initialize anything necessary to provide a non-premptive threading
462 rxi_InitializeThreadSupport();
465 /* Allocate and initialize a socket for client and perhaps server
468 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
469 if (rx_socket == OSI_NULLSOCKET) {
473 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
476 #endif /* RX_LOCKS_DB */
477 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
478 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
479 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
480 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
481 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
482 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
483 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
484 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
485 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
487 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
489 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
491 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
493 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
494 #if defined(AFS_HPUX110_ENV)
496 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
497 #endif /* AFS_HPUX110_ENV */
498 #endif /* RX_ENABLE_LOCKS && KERNEL */
501 rx_connDeadTime = 12;
502 rx_tranquil = 0; /* reset flag */
503 rxi_ResetStatistics();
505 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
506 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
507 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
508 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
509 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
510 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
512 /* Malloc up a bunch of packets & buffers */
514 queue_Init(&rx_freePacketQueue);
515 rxi_NeedMorePackets = FALSE;
516 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
518 /* enforce a minimum number of allocated packets */
519 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
520 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
522 /* allocate the initial free packet pool */
523 #ifdef RX_ENABLE_TSFPQ
524 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
525 #else /* RX_ENABLE_TSFPQ */
526 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
527 #endif /* RX_ENABLE_TSFPQ */
534 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
535 tv.tv_sec = clock_now.sec;
536 tv.tv_usec = clock_now.usec;
537 srand((unsigned int)tv.tv_usec);
544 #if defined(KERNEL) && !defined(UKERNEL)
545 /* Really, this should never happen in a real kernel */
548 struct sockaddr_in addr;
550 int addrlen = sizeof(addr);
552 socklen_t addrlen = sizeof(addr);
554 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
558 rx_port = addr.sin_port;
561 rx_stats.minRtt.sec = 9999999;
563 rx_SetEpoch(tv.tv_sec | 0x80000000);
565 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
566 * will provide a randomer value. */
568 MUTEX_ENTER(&rx_quota_mutex);
569 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
570 MUTEX_EXIT(&rx_quota_mutex);
571 /* *Slightly* random start time for the cid. This is just to help
572 * out with the hashing function at the peer */
573 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
574 rx_connHashTable = (struct rx_connection **)htable;
575 rx_peerHashTable = (struct rx_peer **)ptable;
577 rx_hardAckDelay.sec = 0;
578 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
580 rxevent_Init(20, rxi_ReScheduleEvents);
582 /* Initialize various global queues */
583 queue_Init(&rx_idleServerQueue);
584 queue_Init(&rx_incomingCallQueue);
585 queue_Init(&rx_freeCallQueue);
587 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
588 /* Initialize our list of usable IP addresses. */
592 #if defined(RXK_LISTENER_ENV) || !defined(KERNEL)
593 /* Start listener process (exact function is dependent on the
594 * implementation environment--kernel or user space) */
599 tmp_status = rxinit_status = 0;
607 return rx_InitHost(htonl(INADDR_ANY), port);
613 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
614 * maintaing the round trip timer.
619 * Start a new RTT timer for a given call and packet.
621 * There must be no resendEvent already listed for this call, otherwise this
622 * will leak events - intended for internal use within the RTO code only
625 * the RX call to start the timer for
626 * @param[in] lastPacket
627 * a flag indicating whether the last packet has been sent or not
629 * @pre call must be locked before calling this function
633 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
635 struct clock now, retryTime;
640 clock_Add(&retryTime, &call->rto);
642 /* If we're sending the last packet, and we're the client, then the server
643 * may wait for an additional 400ms before returning the ACK, wait for it
644 * rather than hitting a timeout */
645 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
646 clock_Addmsec(&retryTime, 400);
648 MUTEX_ENTER(&rx_refcnt_mutex);
649 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
650 MUTEX_EXIT(&rx_refcnt_mutex);
651 call->resendEvent = rxevent_PostNow2(&retryTime, &now, rxi_Resend,
656 * Cancel an RTT timer for a given call.
660 * the RX call to cancel the timer for
662 * @pre call must be locked before calling this function
667 rxi_rto_cancel(struct rx_call *call)
669 if (!call->resendEvent)
672 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
676 * Tell the RTO timer that we have sent a packet.
678 * If the timer isn't already running, then start it. If the timer is running,
682 * the RX call that the packet has been sent on
683 * @param[in] lastPacket
684 * A flag which is true if this is the last packet for the call
686 * @pre The call must be locked before calling this function
691 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
693 if (call->resendEvent)
696 rxi_rto_startTimer(call, lastPacket, istack);
700 * Tell the RTO timer that we have received an new ACK message
702 * This function should be called whenever a call receives an ACK that
703 * acknowledges new packets. Whatever happens, we stop the current timer.
704 * If there are unacked packets in the queue which have been sent, then
705 * we restart the timer from now. Otherwise, we leave it stopped.
708 * the RX call that the ACK has been received on
712 rxi_rto_packet_acked(struct rx_call *call, int istack)
714 struct rx_packet *p, *nxp;
716 rxi_rto_cancel(call);
718 if (queue_IsEmpty(&call->tq))
721 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
722 if (p->header.seq > call->tfirst + call->twind)
725 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
726 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
734 * Set an initial round trip timeout for a peer connection
736 * @param[in] secs The timeout to set in seconds
740 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
741 peer->rtt = secs * 8000;
745 * Sets the error generated when a busy call channel is detected.
747 * @param[in] error The error to return for a call on a busy channel.
749 * @pre Neither rx_Init nor rx_InitHost have been called yet
752 rx_SetBusyChannelError(afs_int32 error)
754 osi_Assert(rxinit_status != 0);
755 rxi_busyChannelError = error;
758 /* called with unincremented nRequestsRunning to see if it is OK to start
759 * a new thread in this service. Could be "no" for two reasons: over the
760 * max quota, or would prevent others from reaching their min quota.
762 #ifdef RX_ENABLE_LOCKS
763 /* This verion of QuotaOK reserves quota if it's ok while the
764 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
767 QuotaOK(struct rx_service *aservice)
769 /* check if over max quota */
770 if (aservice->nRequestsRunning >= aservice->maxProcs) {
774 /* under min quota, we're OK */
775 /* otherwise, can use only if there are enough to allow everyone
776 * to go to their min quota after this guy starts.
779 MUTEX_ENTER(&rx_quota_mutex);
780 if ((aservice->nRequestsRunning < aservice->minProcs)
781 || (rxi_availProcs > rxi_minDeficit)) {
782 aservice->nRequestsRunning++;
783 /* just started call in minProcs pool, need fewer to maintain
785 if (aservice->nRequestsRunning <= aservice->minProcs)
788 MUTEX_EXIT(&rx_quota_mutex);
791 MUTEX_EXIT(&rx_quota_mutex);
797 ReturnToServerPool(struct rx_service *aservice)
799 aservice->nRequestsRunning--;
800 MUTEX_ENTER(&rx_quota_mutex);
801 if (aservice->nRequestsRunning < aservice->minProcs)
804 MUTEX_EXIT(&rx_quota_mutex);
807 #else /* RX_ENABLE_LOCKS */
809 QuotaOK(struct rx_service *aservice)
812 /* under min quota, we're OK */
813 if (aservice->nRequestsRunning < aservice->minProcs)
816 /* check if over max quota */
817 if (aservice->nRequestsRunning >= aservice->maxProcs)
820 /* otherwise, can use only if there are enough to allow everyone
821 * to go to their min quota after this guy starts.
823 MUTEX_ENTER(&rx_quota_mutex);
824 if (rxi_availProcs > rxi_minDeficit)
826 MUTEX_EXIT(&rx_quota_mutex);
829 #endif /* RX_ENABLE_LOCKS */
832 /* Called by rx_StartServer to start up lwp's to service calls.
833 NExistingProcs gives the number of procs already existing, and which
834 therefore needn't be created. */
836 rxi_StartServerProcs(int nExistingProcs)
838 struct rx_service *service;
843 /* For each service, reserve N processes, where N is the "minimum"
844 * number of processes that MUST be able to execute a request in parallel,
845 * at any time, for that process. Also compute the maximum difference
846 * between any service's maximum number of processes that can run
847 * (i.e. the maximum number that ever will be run, and a guarantee
848 * that this number will run if other services aren't running), and its
849 * minimum number. The result is the extra number of processes that
850 * we need in order to provide the latter guarantee */
851 for (i = 0; i < RX_MAX_SERVICES; i++) {
853 service = rx_services[i];
854 if (service == (struct rx_service *)0)
856 nProcs += service->minProcs;
857 diff = service->maxProcs - service->minProcs;
861 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
862 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
863 for (i = 0; i < nProcs; i++) {
864 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
870 /* This routine is only required on Windows */
872 rx_StartClientThread(void)
874 #ifdef AFS_PTHREAD_ENV
876 pid = pthread_self();
877 #endif /* AFS_PTHREAD_ENV */
879 #endif /* AFS_NT40_ENV */
881 /* This routine must be called if any services are exported. If the
882 * donateMe flag is set, the calling process is donated to the server
885 rx_StartServer(int donateMe)
887 struct rx_service *service;
893 /* Start server processes, if necessary (exact function is dependent
894 * on the implementation environment--kernel or user space). DonateMe
895 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
896 * case, one less new proc will be created rx_StartServerProcs.
898 rxi_StartServerProcs(donateMe);
900 /* count up the # of threads in minProcs, and add set the min deficit to
901 * be that value, too.
903 for (i = 0; i < RX_MAX_SERVICES; i++) {
904 service = rx_services[i];
905 if (service == (struct rx_service *)0)
907 MUTEX_ENTER(&rx_quota_mutex);
908 rxi_totalMin += service->minProcs;
909 /* below works even if a thread is running, since minDeficit would
910 * still have been decremented and later re-incremented.
912 rxi_minDeficit += service->minProcs;
913 MUTEX_EXIT(&rx_quota_mutex);
916 /* Turn on reaping of idle server connections */
917 rxi_ReapConnections(NULL, NULL, NULL);
926 #ifdef AFS_PTHREAD_ENV
928 pid = afs_pointer_to_int(pthread_self());
929 #else /* AFS_PTHREAD_ENV */
931 LWP_CurrentProcess(&pid);
932 #endif /* AFS_PTHREAD_ENV */
934 sprintf(name, "srv_%d", ++nProcs);
936 (*registerProgram) (pid, name);
938 #endif /* AFS_NT40_ENV */
939 rx_ServerProc(NULL); /* Never returns */
941 #ifdef RX_ENABLE_TSFPQ
942 /* no use leaving packets around in this thread's local queue if
943 * it isn't getting donated to the server thread pool.
945 rxi_FlushLocalPacketsTSFPQ();
946 #endif /* RX_ENABLE_TSFPQ */
950 /* Create a new client connection to the specified service, using the
951 * specified security object to implement the security model for this
953 struct rx_connection *
954 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
955 struct rx_securityClass *securityObject,
956 int serviceSecurityIndex)
960 struct rx_connection *conn;
965 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
966 "serviceSecurityIndex %d)\n",
967 ntohl(shost), ntohs(sport), sservice, securityObject,
968 serviceSecurityIndex));
970 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
971 * the case of kmem_alloc? */
972 conn = rxi_AllocConnection();
973 #ifdef RX_ENABLE_LOCKS
974 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
975 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
976 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
979 MUTEX_ENTER(&rx_connHashTable_lock);
980 cid = (rx_nextCid += RX_MAXCALLS);
981 conn->type = RX_CLIENT_CONNECTION;
983 conn->epoch = rx_epoch;
984 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
985 conn->serviceId = sservice;
986 conn->securityObject = securityObject;
987 conn->securityData = (void *) 0;
988 conn->securityIndex = serviceSecurityIndex;
989 rx_SetConnDeadTime(conn, rx_connDeadTime);
990 rx_SetConnSecondsUntilNatPing(conn, 0);
991 conn->ackRate = RX_FAST_ACK_RATE;
993 conn->specific = NULL;
994 conn->challengeEvent = NULL;
995 conn->delayedAbortEvent = NULL;
996 conn->abortCount = 0;
998 for (i = 0; i < RX_MAXCALLS; i++) {
999 conn->twind[i] = rx_initSendWindow;
1000 conn->rwind[i] = rx_initReceiveWindow;
1001 conn->lastBusy[i] = 0;
1004 RXS_NewConnection(securityObject, conn);
1006 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1008 conn->refCount++; /* no lock required since only this thread knows... */
1009 conn->next = rx_connHashTable[hashindex];
1010 rx_connHashTable[hashindex] = conn;
1011 if (rx_stats_active)
1012 rx_atomic_inc(&rx_stats.nClientConns);
1013 MUTEX_EXIT(&rx_connHashTable_lock);
1019 * Ensure a connection's timeout values are valid.
1021 * @param[in] conn The connection to check
1023 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1024 * unless idleDeadTime and/or hardDeadTime are not set
1028 rxi_CheckConnTimeouts(struct rx_connection *conn)
1030 /* a connection's timeouts must have the relationship
1031 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1032 * total loss of network to a peer may cause an idle timeout instead of a
1033 * dead timeout, simply because the idle timeout gets hit first. Also set
1034 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1035 /* this logic is slightly complicated by the fact that
1036 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1038 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1039 if (conn->idleDeadTime) {
1040 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1042 if (conn->hardDeadTime) {
1043 if (conn->idleDeadTime) {
1044 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1046 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1052 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1054 /* The idea is to set the dead time to a value that allows several
1055 * keepalives to be dropped without timing out the connection. */
1056 conn->secondsUntilDead = seconds;
1057 rxi_CheckConnTimeouts(conn);
1058 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1062 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1064 conn->hardDeadTime = seconds;
1065 rxi_CheckConnTimeouts(conn);
1069 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1071 conn->idleDeadTime = seconds;
1072 rxi_CheckConnTimeouts(conn);
1075 int rxi_lowPeerRefCount = 0;
1076 int rxi_lowConnRefCount = 0;
1079 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1080 * NOTE: must not be called with rx_connHashTable_lock held.
1083 rxi_CleanupConnection(struct rx_connection *conn)
1085 /* Notify the service exporter, if requested, that this connection
1086 * is being destroyed */
1087 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1088 (*conn->service->destroyConnProc) (conn);
1090 /* Notify the security module that this connection is being destroyed */
1091 RXS_DestroyConnection(conn->securityObject, conn);
1093 /* If this is the last connection using the rx_peer struct, set its
1094 * idle time to now. rxi_ReapConnections will reap it if it's still
1095 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1097 MUTEX_ENTER(&rx_peerHashTable_lock);
1098 if (conn->peer->refCount < 2) {
1099 conn->peer->idleWhen = clock_Sec();
1100 if (conn->peer->refCount < 1) {
1101 conn->peer->refCount = 1;
1102 if (rx_stats_active) {
1103 MUTEX_ENTER(&rx_stats_mutex);
1104 rxi_lowPeerRefCount++;
1105 MUTEX_EXIT(&rx_stats_mutex);
1109 conn->peer->refCount--;
1110 MUTEX_EXIT(&rx_peerHashTable_lock);
1112 if (rx_stats_active)
1114 if (conn->type == RX_SERVER_CONNECTION)
1115 rx_atomic_dec(&rx_stats.nServerConns);
1117 rx_atomic_dec(&rx_stats.nClientConns);
1120 if (conn->specific) {
1122 for (i = 0; i < conn->nSpecific; i++) {
1123 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1124 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1125 conn->specific[i] = NULL;
1127 free(conn->specific);
1129 conn->specific = NULL;
1130 conn->nSpecific = 0;
1131 #endif /* !KERNEL */
1133 MUTEX_DESTROY(&conn->conn_call_lock);
1134 MUTEX_DESTROY(&conn->conn_data_lock);
1135 CV_DESTROY(&conn->conn_call_cv);
1137 rxi_FreeConnection(conn);
1140 /* Destroy the specified connection */
1142 rxi_DestroyConnection(struct rx_connection *conn)
1144 MUTEX_ENTER(&rx_connHashTable_lock);
1145 rxi_DestroyConnectionNoLock(conn);
1146 /* conn should be at the head of the cleanup list */
1147 if (conn == rx_connCleanup_list) {
1148 rx_connCleanup_list = rx_connCleanup_list->next;
1149 MUTEX_EXIT(&rx_connHashTable_lock);
1150 rxi_CleanupConnection(conn);
1152 #ifdef RX_ENABLE_LOCKS
1154 MUTEX_EXIT(&rx_connHashTable_lock);
1156 #endif /* RX_ENABLE_LOCKS */
1160 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1162 struct rx_connection **conn_ptr;
1164 struct rx_packet *packet;
1171 MUTEX_ENTER(&conn->conn_data_lock);
1172 MUTEX_ENTER(&rx_refcnt_mutex);
1173 if (conn->refCount > 0)
1176 if (rx_stats_active) {
1177 MUTEX_ENTER(&rx_stats_mutex);
1178 rxi_lowConnRefCount++;
1179 MUTEX_EXIT(&rx_stats_mutex);
1183 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1184 /* Busy; wait till the last guy before proceeding */
1185 MUTEX_EXIT(&rx_refcnt_mutex);
1186 MUTEX_EXIT(&conn->conn_data_lock);
1191 /* If the client previously called rx_NewCall, but it is still
1192 * waiting, treat this as a running call, and wait to destroy the
1193 * connection later when the call completes. */
1194 if ((conn->type == RX_CLIENT_CONNECTION)
1195 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1196 conn->flags |= RX_CONN_DESTROY_ME;
1197 MUTEX_EXIT(&conn->conn_data_lock);
1201 MUTEX_EXIT(&rx_refcnt_mutex);
1202 MUTEX_EXIT(&conn->conn_data_lock);
1204 /* Check for extant references to this connection */
1205 MUTEX_ENTER(&conn->conn_call_lock);
1206 for (i = 0; i < RX_MAXCALLS; i++) {
1207 struct rx_call *call = conn->call[i];
1210 if (conn->type == RX_CLIENT_CONNECTION) {
1211 MUTEX_ENTER(&call->lock);
1212 if (call->delayedAckEvent) {
1213 /* Push the final acknowledgment out now--there
1214 * won't be a subsequent call to acknowledge the
1215 * last reply packets */
1216 rxevent_Cancel(call->delayedAckEvent, call,
1217 RX_CALL_REFCOUNT_DELAY);
1218 if (call->state == RX_STATE_PRECALL
1219 || call->state == RX_STATE_ACTIVE) {
1220 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1222 rxi_AckAll(NULL, call, 0);
1225 MUTEX_EXIT(&call->lock);
1229 MUTEX_EXIT(&conn->conn_call_lock);
1231 #ifdef RX_ENABLE_LOCKS
1233 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1234 MUTEX_EXIT(&conn->conn_data_lock);
1236 /* Someone is accessing a packet right now. */
1240 #endif /* RX_ENABLE_LOCKS */
1243 /* Don't destroy the connection if there are any call
1244 * structures still in use */
1245 MUTEX_ENTER(&conn->conn_data_lock);
1246 conn->flags |= RX_CONN_DESTROY_ME;
1247 MUTEX_EXIT(&conn->conn_data_lock);
1252 if (conn->natKeepAliveEvent) {
1253 rxi_NatKeepAliveOff(conn);
1256 if (conn->delayedAbortEvent) {
1257 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1258 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1260 MUTEX_ENTER(&conn->conn_data_lock);
1261 rxi_SendConnectionAbort(conn, packet, 0, 1);
1262 MUTEX_EXIT(&conn->conn_data_lock);
1263 rxi_FreePacket(packet);
1267 /* Remove from connection hash table before proceeding */
1269 &rx_connHashTable[CONN_HASH
1270 (peer->host, peer->port, conn->cid, conn->epoch,
1272 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1273 if (*conn_ptr == conn) {
1274 *conn_ptr = conn->next;
1278 /* if the conn that we are destroying was the last connection, then we
1279 * clear rxLastConn as well */
1280 if (rxLastConn == conn)
1283 /* Make sure the connection is completely reset before deleting it. */
1284 /* get rid of pending events that could zap us later */
1285 if (conn->challengeEvent)
1286 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1287 if (conn->checkReachEvent)
1288 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1289 if (conn->natKeepAliveEvent)
1290 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
1292 /* Add the connection to the list of destroyed connections that
1293 * need to be cleaned up. This is necessary to avoid deadlocks
1294 * in the routines we call to inform others that this connection is
1295 * being destroyed. */
1296 conn->next = rx_connCleanup_list;
1297 rx_connCleanup_list = conn;
1300 /* Externally available version */
1302 rx_DestroyConnection(struct rx_connection *conn)
1307 rxi_DestroyConnection(conn);
1312 rx_GetConnection(struct rx_connection *conn)
1317 MUTEX_ENTER(&rx_refcnt_mutex);
1319 MUTEX_EXIT(&rx_refcnt_mutex);
1323 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1324 /* Wait for the transmit queue to no longer be busy.
1325 * requires the call->lock to be held */
1327 rxi_WaitforTQBusy(struct rx_call *call) {
1328 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1329 call->flags |= RX_CALL_TQ_WAIT;
1331 #ifdef RX_ENABLE_LOCKS
1332 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1333 CV_WAIT(&call->cv_tq, &call->lock);
1334 #else /* RX_ENABLE_LOCKS */
1335 osi_rxSleep(&call->tq);
1336 #endif /* RX_ENABLE_LOCKS */
1338 if (call->tqWaiters == 0) {
1339 call->flags &= ~RX_CALL_TQ_WAIT;
1346 rxi_WakeUpTransmitQueue(struct rx_call *call)
1348 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1349 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1350 call, call->tqWaiters, call->flags));
1351 #ifdef RX_ENABLE_LOCKS
1352 osirx_AssertMine(&call->lock, "rxi_Start start");
1353 CV_BROADCAST(&call->cv_tq);
1354 #else /* RX_ENABLE_LOCKS */
1355 osi_rxWakeup(&call->tq);
1356 #endif /* RX_ENABLE_LOCKS */
1360 /* Start a new rx remote procedure call, on the specified connection.
1361 * If wait is set to 1, wait for a free call channel; otherwise return
1362 * 0. Maxtime gives the maximum number of seconds this call may take,
1363 * after rx_NewCall returns. After this time interval, a call to any
1364 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1365 * For fine grain locking, we hold the conn_call_lock in order to
1366 * to ensure that we don't get signalle after we found a call in an active
1367 * state and before we go to sleep.
1370 rx_NewCall(struct rx_connection *conn)
1372 int i, wait, ignoreBusy = 1;
1373 struct rx_call *call;
1374 struct clock queueTime;
1375 afs_uint32 leastBusy = 0;
1379 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1382 clock_GetTime(&queueTime);
1384 * Check if there are others waiting for a new call.
1385 * If so, let them go first to avoid starving them.
1386 * This is a fairly simple scheme, and might not be
1387 * a complete solution for large numbers of waiters.
1389 * makeCallWaiters keeps track of the number of
1390 * threads waiting to make calls and the
1391 * RX_CONN_MAKECALL_WAITING flag bit is used to
1392 * indicate that there are indeed calls waiting.
1393 * The flag is set when the waiter is incremented.
1394 * It is only cleared when makeCallWaiters is 0.
1395 * This prevents us from accidently destroying the
1396 * connection while it is potentially about to be used.
1398 MUTEX_ENTER(&conn->conn_call_lock);
1399 MUTEX_ENTER(&conn->conn_data_lock);
1400 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1401 conn->flags |= RX_CONN_MAKECALL_WAITING;
1402 conn->makeCallWaiters++;
1403 MUTEX_EXIT(&conn->conn_data_lock);
1405 #ifdef RX_ENABLE_LOCKS
1406 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1410 MUTEX_ENTER(&conn->conn_data_lock);
1411 conn->makeCallWaiters--;
1412 if (conn->makeCallWaiters == 0)
1413 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1416 /* We are now the active thread in rx_NewCall */
1417 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1418 MUTEX_EXIT(&conn->conn_data_lock);
1423 for (i = 0; i < RX_MAXCALLS; i++) {
1424 call = conn->call[i];
1426 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1427 /* we're not ignoring busy call slots; only look at the
1428 * call slot that is the "least" busy */
1432 if (call->state == RX_STATE_DALLY) {
1433 MUTEX_ENTER(&call->lock);
1434 if (call->state == RX_STATE_DALLY) {
1435 if (ignoreBusy && conn->lastBusy[i]) {
1436 /* if we're ignoring busy call slots, skip any ones that
1437 * have lastBusy set */
1438 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1439 leastBusy = conn->lastBusy[i];
1441 MUTEX_EXIT(&call->lock);
1446 * We are setting the state to RX_STATE_RESET to
1447 * ensure that no one else will attempt to use this
1448 * call once we drop the conn->conn_call_lock and
1449 * call->lock. We must drop the conn->conn_call_lock
1450 * before calling rxi_ResetCall because the process
1451 * of clearing the transmit queue can block for an
1452 * extended period of time. If we block while holding
1453 * the conn->conn_call_lock, then all rx_EndCall
1454 * processing will block as well. This has a detrimental
1455 * effect on overall system performance.
1457 call->state = RX_STATE_RESET;
1458 MUTEX_EXIT(&conn->conn_call_lock);
1459 MUTEX_ENTER(&rx_refcnt_mutex);
1460 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1461 MUTEX_EXIT(&rx_refcnt_mutex);
1462 rxi_ResetCall(call, 0);
1463 (*call->callNumber)++;
1464 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1468 * If we failed to be able to safely obtain the
1469 * conn->conn_call_lock we will have to drop the
1470 * call->lock to avoid a deadlock. When the call->lock
1471 * is released the state of the call can change. If it
1472 * is no longer RX_STATE_RESET then some other thread is
1475 MUTEX_EXIT(&call->lock);
1476 MUTEX_ENTER(&conn->conn_call_lock);
1477 MUTEX_ENTER(&call->lock);
1479 if (call->state == RX_STATE_RESET)
1483 * If we get here it means that after dropping
1484 * the conn->conn_call_lock and call->lock that
1485 * the call is no longer ours. If we can't find
1486 * a free call in the remaining slots we should
1487 * not go immediately to RX_CONN_MAKECALL_WAITING
1488 * because by dropping the conn->conn_call_lock
1489 * we have given up synchronization with rx_EndCall.
1490 * Instead, cycle through one more time to see if
1491 * we can find a call that can call our own.
1493 MUTEX_ENTER(&rx_refcnt_mutex);
1494 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1495 MUTEX_EXIT(&rx_refcnt_mutex);
1498 MUTEX_EXIT(&call->lock);
1501 if (ignoreBusy && conn->lastBusy[i]) {
1502 /* if we're ignoring busy call slots, skip any ones that
1503 * have lastBusy set */
1504 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1505 leastBusy = conn->lastBusy[i];
1510 /* rxi_NewCall returns with mutex locked */
1511 call = rxi_NewCall(conn, i);
1512 MUTEX_ENTER(&rx_refcnt_mutex);
1513 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1514 MUTEX_EXIT(&rx_refcnt_mutex);
1518 if (i < RX_MAXCALLS) {
1519 conn->lastBusy[i] = 0;
1524 if (leastBusy && ignoreBusy) {
1525 /* we didn't find a useable call slot, but we did see at least one
1526 * 'busy' slot; look again and only use a slot with the 'least
1532 MUTEX_ENTER(&conn->conn_data_lock);
1533 conn->flags |= RX_CONN_MAKECALL_WAITING;
1534 conn->makeCallWaiters++;
1535 MUTEX_EXIT(&conn->conn_data_lock);
1537 #ifdef RX_ENABLE_LOCKS
1538 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1542 MUTEX_ENTER(&conn->conn_data_lock);
1543 conn->makeCallWaiters--;
1544 if (conn->makeCallWaiters == 0)
1545 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1546 MUTEX_EXIT(&conn->conn_data_lock);
1548 /* Client is initially in send mode */
1549 call->state = RX_STATE_ACTIVE;
1550 call->error = conn->error;
1552 call->mode = RX_MODE_ERROR;
1554 call->mode = RX_MODE_SENDING;
1556 /* remember start time for call in case we have hard dead time limit */
1557 call->queueTime = queueTime;
1558 clock_GetTime(&call->startTime);
1559 hzero(call->bytesSent);
1560 hzero(call->bytesRcvd);
1562 /* Turn on busy protocol. */
1563 rxi_KeepAliveOn(call);
1565 /* Attempt MTU discovery */
1566 rxi_GrowMTUOn(call);
1569 * We are no longer the active thread in rx_NewCall
1571 MUTEX_ENTER(&conn->conn_data_lock);
1572 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1573 MUTEX_EXIT(&conn->conn_data_lock);
1576 * Wake up anyone else who might be giving us a chance to
1577 * run (see code above that avoids resource starvation).
1579 #ifdef RX_ENABLE_LOCKS
1580 CV_BROADCAST(&conn->conn_call_cv);
1584 MUTEX_EXIT(&conn->conn_call_lock);
1586 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1587 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1588 osi_Panic("rx_NewCall call about to be used without an empty tq");
1590 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1592 MUTEX_EXIT(&call->lock);
1595 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1600 rxi_HasActiveCalls(struct rx_connection *aconn)
1603 struct rx_call *tcall;
1607 for (i = 0; i < RX_MAXCALLS; i++) {
1608 if ((tcall = aconn->call[i])) {
1609 if ((tcall->state == RX_STATE_ACTIVE)
1610 || (tcall->state == RX_STATE_PRECALL)) {
1621 rxi_GetCallNumberVector(struct rx_connection *aconn,
1622 afs_int32 * aint32s)
1625 struct rx_call *tcall;
1629 for (i = 0; i < RX_MAXCALLS; i++) {
1630 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1631 aint32s[i] = aconn->callNumber[i] + 1;
1633 aint32s[i] = aconn->callNumber[i];
1640 rxi_SetCallNumberVector(struct rx_connection *aconn,
1641 afs_int32 * aint32s)
1644 struct rx_call *tcall;
1648 for (i = 0; i < RX_MAXCALLS; i++) {
1649 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1650 aconn->callNumber[i] = aint32s[i] - 1;
1652 aconn->callNumber[i] = aint32s[i];
1658 /* Advertise a new service. A service is named locally by a UDP port
1659 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1662 char *serviceName; Name for identification purposes (e.g. the
1663 service name might be used for probing for
1666 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1667 char *serviceName, struct rx_securityClass **securityObjects,
1668 int nSecurityObjects,
1669 afs_int32(*serviceProc) (struct rx_call * acall))
1671 osi_socket socket = OSI_NULLSOCKET;
1672 struct rx_service *tservice;
1678 if (serviceId == 0) {
1680 "rx_NewService: service id for service %s is not non-zero.\n",
1687 "rx_NewService: A non-zero port must be specified on this call if a non-zero port was not provided at Rx initialization (service %s).\n",
1695 tservice = rxi_AllocService();
1698 #ifdef RX_ENABLE_LOCKS
1699 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1702 for (i = 0; i < RX_MAX_SERVICES; i++) {
1703 struct rx_service *service = rx_services[i];
1705 if (port == service->servicePort && host == service->serviceHost) {
1706 if (service->serviceId == serviceId) {
1707 /* The identical service has already been
1708 * installed; if the caller was intending to
1709 * change the security classes used by this
1710 * service, he/she loses. */
1712 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1713 serviceName, serviceId, service->serviceName);
1715 rxi_FreeService(tservice);
1718 /* Different service, same port: re-use the socket
1719 * which is bound to the same port */
1720 socket = service->socket;
1723 if (socket == OSI_NULLSOCKET) {
1724 /* If we don't already have a socket (from another
1725 * service on same port) get a new one */
1726 socket = rxi_GetHostUDPSocket(host, port);
1727 if (socket == OSI_NULLSOCKET) {
1729 rxi_FreeService(tservice);
1734 service->socket = socket;
1735 service->serviceHost = host;
1736 service->servicePort = port;
1737 service->serviceId = serviceId;
1738 service->serviceName = serviceName;
1739 service->nSecurityObjects = nSecurityObjects;
1740 service->securityObjects = securityObjects;
1741 service->minProcs = 0;
1742 service->maxProcs = 1;
1743 service->idleDeadTime = 60;
1744 service->idleDeadErr = 0;
1745 service->connDeadTime = rx_connDeadTime;
1746 service->executeRequestProc = serviceProc;
1747 service->checkReach = 0;
1748 service->nSpecific = 0;
1749 service->specific = NULL;
1750 rx_services[i] = service; /* not visible until now */
1756 rxi_FreeService(tservice);
1757 (osi_Msg "rx_NewService: cannot support > %d services\n",
1762 /* Set configuration options for all of a service's security objects */
1765 rx_SetSecurityConfiguration(struct rx_service *service,
1766 rx_securityConfigVariables type,
1770 for (i = 0; i<service->nSecurityObjects; i++) {
1771 if (service->securityObjects[i]) {
1772 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1780 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1781 struct rx_securityClass **securityObjects, int nSecurityObjects,
1782 afs_int32(*serviceProc) (struct rx_call * acall))
1784 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1787 /* Generic request processing loop. This routine should be called
1788 * by the implementation dependent rx_ServerProc. If socketp is
1789 * non-null, it will be set to the file descriptor that this thread
1790 * is now listening on. If socketp is null, this routine will never
1793 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1795 struct rx_call *call;
1797 struct rx_service *tservice = NULL;
1804 call = rx_GetCall(threadID, tservice, socketp);
1805 if (socketp && *socketp != OSI_NULLSOCKET) {
1806 /* We are now a listener thread */
1812 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1813 #ifdef RX_ENABLE_LOCKS
1815 #endif /* RX_ENABLE_LOCKS */
1816 afs_termState = AFSOP_STOP_AFS;
1817 afs_osi_Wakeup(&afs_termState);
1818 #ifdef RX_ENABLE_LOCKS
1820 #endif /* RX_ENABLE_LOCKS */
1825 /* if server is restarting( typically smooth shutdown) then do not
1826 * allow any new calls.
1829 if (rx_tranquil && (call != NULL)) {
1833 MUTEX_ENTER(&call->lock);
1835 rxi_CallError(call, RX_RESTARTING);
1836 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1838 MUTEX_EXIT(&call->lock);
1843 tservice = call->conn->service;
1845 if (tservice->beforeProc)
1846 (*tservice->beforeProc) (call);
1848 code = tservice->executeRequestProc(call);
1850 if (tservice->afterProc)
1851 (*tservice->afterProc) (call, code);
1853 rx_EndCall(call, code);
1855 if (tservice->postProc)
1856 (*tservice->postProc) (code);
1858 if (rx_stats_active) {
1859 MUTEX_ENTER(&rx_stats_mutex);
1861 MUTEX_EXIT(&rx_stats_mutex);
1868 rx_WakeupServerProcs(void)
1870 struct rx_serverQueueEntry *np, *tqp;
1874 MUTEX_ENTER(&rx_serverPool_lock);
1876 #ifdef RX_ENABLE_LOCKS
1877 if (rx_waitForPacket)
1878 CV_BROADCAST(&rx_waitForPacket->cv);
1879 #else /* RX_ENABLE_LOCKS */
1880 if (rx_waitForPacket)
1881 osi_rxWakeup(rx_waitForPacket);
1882 #endif /* RX_ENABLE_LOCKS */
1883 MUTEX_ENTER(&freeSQEList_lock);
1884 for (np = rx_FreeSQEList; np; np = tqp) {
1885 tqp = *(struct rx_serverQueueEntry **)np;
1886 #ifdef RX_ENABLE_LOCKS
1887 CV_BROADCAST(&np->cv);
1888 #else /* RX_ENABLE_LOCKS */
1890 #endif /* RX_ENABLE_LOCKS */
1892 MUTEX_EXIT(&freeSQEList_lock);
1893 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1894 #ifdef RX_ENABLE_LOCKS
1895 CV_BROADCAST(&np->cv);
1896 #else /* RX_ENABLE_LOCKS */
1898 #endif /* RX_ENABLE_LOCKS */
1900 MUTEX_EXIT(&rx_serverPool_lock);
1905 * One thing that seems to happen is that all the server threads get
1906 * tied up on some empty or slow call, and then a whole bunch of calls
1907 * arrive at once, using up the packet pool, so now there are more
1908 * empty calls. The most critical resources here are server threads
1909 * and the free packet pool. The "doreclaim" code seems to help in
1910 * general. I think that eventually we arrive in this state: there
1911 * are lots of pending calls which do have all their packets present,
1912 * so they won't be reclaimed, are multi-packet calls, so they won't
1913 * be scheduled until later, and thus are tying up most of the free
1914 * packet pool for a very long time.
1916 * 1. schedule multi-packet calls if all the packets are present.
1917 * Probably CPU-bound operation, useful to return packets to pool.
1918 * Do what if there is a full window, but the last packet isn't here?
1919 * 3. preserve one thread which *only* runs "best" calls, otherwise
1920 * it sleeps and waits for that type of call.
1921 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1922 * the current dataquota business is badly broken. The quota isn't adjusted
1923 * to reflect how many packets are presently queued for a running call.
1924 * So, when we schedule a queued call with a full window of packets queued
1925 * up for it, that *should* free up a window full of packets for other 2d-class
1926 * calls to be able to use from the packet pool. But it doesn't.
1928 * NB. Most of the time, this code doesn't run -- since idle server threads
1929 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1930 * as a new call arrives.
1932 /* Sleep until a call arrives. Returns a pointer to the call, ready
1933 * for an rx_Read. */
1934 #ifdef RX_ENABLE_LOCKS
1936 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1938 struct rx_serverQueueEntry *sq;
1939 struct rx_call *call = (struct rx_call *)0;
1940 struct rx_service *service = NULL;
1942 MUTEX_ENTER(&freeSQEList_lock);
1944 if ((sq = rx_FreeSQEList)) {
1945 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1946 MUTEX_EXIT(&freeSQEList_lock);
1947 } else { /* otherwise allocate a new one and return that */
1948 MUTEX_EXIT(&freeSQEList_lock);
1949 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1950 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1951 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1954 MUTEX_ENTER(&rx_serverPool_lock);
1955 if (cur_service != NULL) {
1956 ReturnToServerPool(cur_service);
1959 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1960 struct rx_call *tcall, *ncall, *choice2 = NULL;
1962 /* Scan for eligible incoming calls. A call is not eligible
1963 * if the maximum number of calls for its service type are
1964 * already executing */
1965 /* One thread will process calls FCFS (to prevent starvation),
1966 * while the other threads may run ahead looking for calls which
1967 * have all their input data available immediately. This helps
1968 * keep threads from blocking, waiting for data from the client. */
1969 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1970 service = tcall->conn->service;
1971 if (!QuotaOK(service)) {
1974 MUTEX_ENTER(&rx_pthread_mutex);
1975 if (tno == rxi_fcfs_thread_num
1976 || !tcall->queue_item_header.next) {
1977 MUTEX_EXIT(&rx_pthread_mutex);
1978 /* If we're the fcfs thread , then we'll just use
1979 * this call. If we haven't been able to find an optimal
1980 * choice, and we're at the end of the list, then use a
1981 * 2d choice if one has been identified. Otherwise... */
1982 call = (choice2 ? choice2 : tcall);
1983 service = call->conn->service;
1985 MUTEX_EXIT(&rx_pthread_mutex);
1986 if (!queue_IsEmpty(&tcall->rq)) {
1987 struct rx_packet *rp;
1988 rp = queue_First(&tcall->rq, rx_packet);
1989 if (rp->header.seq == 1) {
1991 || (rp->header.flags & RX_LAST_PACKET)) {
1993 } else if (rxi_2dchoice && !choice2
1994 && !(tcall->flags & RX_CALL_CLEARED)
1995 && (tcall->rprev > rxi_HardAckRate)) {
2005 ReturnToServerPool(service);
2012 MUTEX_EXIT(&rx_serverPool_lock);
2013 MUTEX_ENTER(&call->lock);
2015 if (call->flags & RX_CALL_WAIT_PROC) {
2016 call->flags &= ~RX_CALL_WAIT_PROC;
2017 rx_atomic_dec(&rx_nWaiting);
2020 if (call->state != RX_STATE_PRECALL || call->error) {
2021 MUTEX_EXIT(&call->lock);
2022 MUTEX_ENTER(&rx_serverPool_lock);
2023 ReturnToServerPool(service);
2028 if (queue_IsEmpty(&call->rq)
2029 || queue_First(&call->rq, rx_packet)->header.seq != 1)
2030 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2032 CLEAR_CALL_QUEUE_LOCK(call);
2035 /* If there are no eligible incoming calls, add this process
2036 * to the idle server queue, to wait for one */
2040 *socketp = OSI_NULLSOCKET;
2042 sq->socketp = socketp;
2043 queue_Append(&rx_idleServerQueue, sq);
2044 #ifndef AFS_AIX41_ENV
2045 rx_waitForPacket = sq;
2047 rx_waitingForPacket = sq;
2048 #endif /* AFS_AIX41_ENV */
2050 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2052 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2053 MUTEX_EXIT(&rx_serverPool_lock);
2054 return (struct rx_call *)0;
2057 } while (!(call = sq->newcall)
2058 && !(socketp && *socketp != OSI_NULLSOCKET));
2059 MUTEX_EXIT(&rx_serverPool_lock);
2061 MUTEX_ENTER(&call->lock);
2067 MUTEX_ENTER(&freeSQEList_lock);
2068 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2069 rx_FreeSQEList = sq;
2070 MUTEX_EXIT(&freeSQEList_lock);
2073 clock_GetTime(&call->startTime);
2074 call->state = RX_STATE_ACTIVE;
2075 call->mode = RX_MODE_RECEIVING;
2076 #ifdef RX_KERNEL_TRACE
2077 if (ICL_SETACTIVE(afs_iclSetp)) {
2078 int glockOwner = ISAFS_GLOCK();
2081 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2082 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2089 rxi_calltrace(RX_CALL_START, call);
2090 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2091 call->conn->service->servicePort, call->conn->service->serviceId,
2094 MUTEX_EXIT(&call->lock);
2095 MUTEX_ENTER(&rx_refcnt_mutex);
2096 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2097 MUTEX_EXIT(&rx_refcnt_mutex);
2099 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2104 #else /* RX_ENABLE_LOCKS */
2106 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2108 struct rx_serverQueueEntry *sq;
2109 struct rx_call *call = (struct rx_call *)0, *choice2;
2110 struct rx_service *service = NULL;
2114 MUTEX_ENTER(&freeSQEList_lock);
2116 if ((sq = rx_FreeSQEList)) {
2117 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2118 MUTEX_EXIT(&freeSQEList_lock);
2119 } else { /* otherwise allocate a new one and return that */
2120 MUTEX_EXIT(&freeSQEList_lock);
2121 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2122 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2123 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2125 MUTEX_ENTER(&sq->lock);
2127 if (cur_service != NULL) {
2128 cur_service->nRequestsRunning--;
2129 MUTEX_ENTER(&rx_quota_mutex);
2130 if (cur_service->nRequestsRunning < cur_service->minProcs)
2133 MUTEX_EXIT(&rx_quota_mutex);
2135 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
2136 struct rx_call *tcall, *ncall;
2137 /* Scan for eligible incoming calls. A call is not eligible
2138 * if the maximum number of calls for its service type are
2139 * already executing */
2140 /* One thread will process calls FCFS (to prevent starvation),
2141 * while the other threads may run ahead looking for calls which
2142 * have all their input data available immediately. This helps
2143 * keep threads from blocking, waiting for data from the client. */
2144 choice2 = (struct rx_call *)0;
2145 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
2146 service = tcall->conn->service;
2147 if (QuotaOK(service)) {
2148 MUTEX_ENTER(&rx_pthread_mutex);
2149 if (tno == rxi_fcfs_thread_num
2150 || !tcall->queue_item_header.next) {
2151 MUTEX_EXIT(&rx_pthread_mutex);
2152 /* If we're the fcfs thread, then we'll just use
2153 * this call. If we haven't been able to find an optimal
2154 * choice, and we're at the end of the list, then use a
2155 * 2d choice if one has been identified. Otherwise... */
2156 call = (choice2 ? choice2 : tcall);
2157 service = call->conn->service;
2159 MUTEX_EXIT(&rx_pthread_mutex);
2160 if (!queue_IsEmpty(&tcall->rq)) {
2161 struct rx_packet *rp;
2162 rp = queue_First(&tcall->rq, rx_packet);
2163 if (rp->header.seq == 1
2165 || (rp->header.flags & RX_LAST_PACKET))) {
2167 } else if (rxi_2dchoice && !choice2
2168 && !(tcall->flags & RX_CALL_CLEARED)
2169 && (tcall->rprev > rxi_HardAckRate)) {
2183 /* we can't schedule a call if there's no data!!! */
2184 /* send an ack if there's no data, if we're missing the
2185 * first packet, or we're missing something between first
2186 * and last -- there's a "hole" in the incoming data. */
2187 if (queue_IsEmpty(&call->rq)
2188 || queue_First(&call->rq, rx_packet)->header.seq != 1
2189 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
2190 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2192 call->flags &= (~RX_CALL_WAIT_PROC);
2193 service->nRequestsRunning++;
2194 /* just started call in minProcs pool, need fewer to maintain
2196 MUTEX_ENTER(&rx_quota_mutex);
2197 if (service->nRequestsRunning <= service->minProcs)
2200 MUTEX_EXIT(&rx_quota_mutex);
2201 rx_atomic_dec(&rx_nWaiting);
2202 /* MUTEX_EXIT(&call->lock); */
2204 /* If there are no eligible incoming calls, add this process
2205 * to the idle server queue, to wait for one */
2208 *socketp = OSI_NULLSOCKET;
2210 sq->socketp = socketp;
2211 queue_Append(&rx_idleServerQueue, sq);
2215 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2217 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2218 return (struct rx_call *)0;
2221 } while (!(call = sq->newcall)
2222 && !(socketp && *socketp != OSI_NULLSOCKET));
2224 MUTEX_EXIT(&sq->lock);
2226 MUTEX_ENTER(&freeSQEList_lock);
2227 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2228 rx_FreeSQEList = sq;
2229 MUTEX_EXIT(&freeSQEList_lock);
2232 clock_GetTime(&call->startTime);
2233 call->state = RX_STATE_ACTIVE;
2234 call->mode = RX_MODE_RECEIVING;
2235 #ifdef RX_KERNEL_TRACE
2236 if (ICL_SETACTIVE(afs_iclSetp)) {
2237 int glockOwner = ISAFS_GLOCK();
2240 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2241 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2248 rxi_calltrace(RX_CALL_START, call);
2249 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2250 call->conn->service->servicePort, call->conn->service->serviceId,
2253 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2260 #endif /* RX_ENABLE_LOCKS */
2264 /* Establish a procedure to be called when a packet arrives for a
2265 * call. This routine will be called at most once after each call,
2266 * and will also be called if there is an error condition on the or
2267 * the call is complete. Used by multi rx to build a selection
2268 * function which determines which of several calls is likely to be a
2269 * good one to read from.
2270 * NOTE: the way this is currently implemented it is probably only a
2271 * good idea to (1) use it immediately after a newcall (clients only)
2272 * and (2) only use it once. Other uses currently void your warranty
2275 rx_SetArrivalProc(struct rx_call *call,
2276 void (*proc) (struct rx_call * call,
2279 void * handle, int arg)
2281 call->arrivalProc = proc;
2282 call->arrivalProcHandle = handle;
2283 call->arrivalProcArg = arg;
2286 /* Call is finished (possibly prematurely). Return rc to the peer, if
2287 * appropriate, and return the final error code from the conversation
2291 rx_EndCall(struct rx_call *call, afs_int32 rc)
2293 struct rx_connection *conn = call->conn;
2297 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2298 call, rc, call->error, call->abortCode));
2301 MUTEX_ENTER(&call->lock);
2303 if (rc == 0 && call->error == 0) {
2304 call->abortCode = 0;
2305 call->abortCount = 0;
2308 call->arrivalProc = (void (*)())0;
2309 if (rc && call->error == 0) {
2310 rxi_CallError(call, rc);
2311 call->mode = RX_MODE_ERROR;
2312 /* Send an abort message to the peer if this error code has
2313 * only just been set. If it was set previously, assume the
2314 * peer has already been sent the error code or will request it
2316 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2318 if (conn->type == RX_SERVER_CONNECTION) {
2319 /* Make sure reply or at least dummy reply is sent */
2320 if (call->mode == RX_MODE_RECEIVING) {
2321 MUTEX_EXIT(&call->lock);
2322 rxi_WriteProc(call, 0, 0);
2323 MUTEX_ENTER(&call->lock);
2325 if (call->mode == RX_MODE_SENDING) {
2326 MUTEX_EXIT(&call->lock);
2327 rxi_FlushWrite(call);
2328 MUTEX_ENTER(&call->lock);
2330 rxi_calltrace(RX_CALL_END, call);
2331 /* Call goes to hold state until reply packets are acknowledged */
2332 if (call->tfirst + call->nSoftAcked < call->tnext) {
2333 call->state = RX_STATE_HOLD;
2335 call->state = RX_STATE_DALLY;
2336 rxi_ClearTransmitQueue(call, 0);
2337 rxi_rto_cancel(call);
2338 rxevent_Cancel(call->keepAliveEvent, call,
2339 RX_CALL_REFCOUNT_ALIVE);
2341 } else { /* Client connection */
2343 /* Make sure server receives input packets, in the case where
2344 * no reply arguments are expected */
2345 if ((call->mode == RX_MODE_SENDING)
2346 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2347 MUTEX_EXIT(&call->lock);
2348 (void)rxi_ReadProc(call, &dummy, 1);
2349 MUTEX_ENTER(&call->lock);
2352 /* If we had an outstanding delayed ack, be nice to the server
2353 * and force-send it now.
2355 if (call->delayedAckEvent) {
2356 rxevent_Cancel(call->delayedAckEvent, call,
2357 RX_CALL_REFCOUNT_DELAY);
2358 call->delayedAckEvent = NULL;
2359 rxi_SendDelayedAck(NULL, call, NULL);
2362 /* We need to release the call lock since it's lower than the
2363 * conn_call_lock and we don't want to hold the conn_call_lock
2364 * over the rx_ReadProc call. The conn_call_lock needs to be held
2365 * here for the case where rx_NewCall is perusing the calls on
2366 * the connection structure. We don't want to signal until
2367 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2368 * have checked this call, found it active and by the time it
2369 * goes to sleep, will have missed the signal.
2371 MUTEX_EXIT(&call->lock);
2372 MUTEX_ENTER(&conn->conn_call_lock);
2373 MUTEX_ENTER(&call->lock);
2375 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2376 conn->lastBusy[call->channel] = 0;
2379 MUTEX_ENTER(&conn->conn_data_lock);
2380 conn->flags |= RX_CONN_BUSY;
2381 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2382 MUTEX_EXIT(&conn->conn_data_lock);
2383 #ifdef RX_ENABLE_LOCKS
2384 CV_BROADCAST(&conn->conn_call_cv);
2389 #ifdef RX_ENABLE_LOCKS
2391 MUTEX_EXIT(&conn->conn_data_lock);
2393 #endif /* RX_ENABLE_LOCKS */
2394 call->state = RX_STATE_DALLY;
2396 error = call->error;
2398 /* currentPacket, nLeft, and NFree must be zeroed here, because
2399 * ResetCall cannot: ResetCall may be called at splnet(), in the
2400 * kernel version, and may interrupt the macros rx_Read or
2401 * rx_Write, which run at normal priority for efficiency. */
2402 if (call->currentPacket) {
2403 #ifdef RX_TRACK_PACKETS
2404 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2406 rxi_FreePacket(call->currentPacket);
2407 call->currentPacket = (struct rx_packet *)0;
2410 call->nLeft = call->nFree = call->curlen = 0;
2412 /* Free any packets from the last call to ReadvProc/WritevProc */
2413 #ifdef RXDEBUG_PACKET
2415 #endif /* RXDEBUG_PACKET */
2416 rxi_FreePackets(0, &call->iovq);
2417 MUTEX_EXIT(&call->lock);
2419 MUTEX_ENTER(&rx_refcnt_mutex);
2420 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2421 MUTEX_EXIT(&rx_refcnt_mutex);
2422 if (conn->type == RX_CLIENT_CONNECTION) {
2423 MUTEX_ENTER(&conn->conn_data_lock);
2424 conn->flags &= ~RX_CONN_BUSY;
2425 MUTEX_EXIT(&conn->conn_data_lock);
2426 MUTEX_EXIT(&conn->conn_call_lock);
2430 * Map errors to the local host's errno.h format.
2432 error = ntoh_syserr_conv(error);
2436 #if !defined(KERNEL)
2438 /* Call this routine when shutting down a server or client (especially
2439 * clients). This will allow Rx to gracefully garbage collect server
2440 * connections, and reduce the number of retries that a server might
2441 * make to a dead client.
2442 * This is not quite right, since some calls may still be ongoing and
2443 * we can't lock them to destroy them. */
2447 struct rx_connection **conn_ptr, **conn_end;
2451 if (rxinit_status == 1) {
2453 return; /* Already shutdown. */
2455 rxi_DeleteCachedConnections();
2456 if (rx_connHashTable) {
2457 MUTEX_ENTER(&rx_connHashTable_lock);
2458 for (conn_ptr = &rx_connHashTable[0], conn_end =
2459 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2461 struct rx_connection *conn, *next;
2462 for (conn = *conn_ptr; conn; conn = next) {
2464 if (conn->type == RX_CLIENT_CONNECTION) {
2465 MUTEX_ENTER(&rx_refcnt_mutex);
2467 MUTEX_EXIT(&rx_refcnt_mutex);
2468 #ifdef RX_ENABLE_LOCKS
2469 rxi_DestroyConnectionNoLock(conn);
2470 #else /* RX_ENABLE_LOCKS */
2471 rxi_DestroyConnection(conn);
2472 #endif /* RX_ENABLE_LOCKS */
2476 #ifdef RX_ENABLE_LOCKS
2477 while (rx_connCleanup_list) {
2478 struct rx_connection *conn;
2479 conn = rx_connCleanup_list;
2480 rx_connCleanup_list = rx_connCleanup_list->next;
2481 MUTEX_EXIT(&rx_connHashTable_lock);
2482 rxi_CleanupConnection(conn);
2483 MUTEX_ENTER(&rx_connHashTable_lock);
2485 MUTEX_EXIT(&rx_connHashTable_lock);
2486 #endif /* RX_ENABLE_LOCKS */
2491 afs_winsockCleanup();
2499 /* if we wakeup packet waiter too often, can get in loop with two
2500 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2502 rxi_PacketsUnWait(void)
2504 if (!rx_waitingForPackets) {
2508 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2509 return; /* still over quota */
2512 rx_waitingForPackets = 0;
2513 #ifdef RX_ENABLE_LOCKS
2514 CV_BROADCAST(&rx_waitingForPackets_cv);
2516 osi_rxWakeup(&rx_waitingForPackets);
2522 /* ------------------Internal interfaces------------------------- */
2524 /* Return this process's service structure for the
2525 * specified socket and service */
2526 static struct rx_service *
2527 rxi_FindService(osi_socket socket, u_short serviceId)
2529 struct rx_service **sp;
2530 for (sp = &rx_services[0]; *sp; sp++) {
2531 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2537 #ifdef RXDEBUG_PACKET
2538 #ifdef KDUMP_RX_LOCK
2539 static struct rx_call_rx_lock *rx_allCallsp = 0;
2541 static struct rx_call *rx_allCallsp = 0;
2543 #endif /* RXDEBUG_PACKET */
2545 /* Allocate a call structure, for the indicated channel of the
2546 * supplied connection. The mode and state of the call must be set by
2547 * the caller. Returns the call with mutex locked. */
2548 static struct rx_call *
2549 rxi_NewCall(struct rx_connection *conn, int channel)
2551 struct rx_call *call;
2552 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2553 struct rx_call *cp; /* Call pointer temp */
2554 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2555 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2557 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2559 /* Grab an existing call structure, or allocate a new one.
2560 * Existing call structures are assumed to have been left reset by
2562 MUTEX_ENTER(&rx_freeCallQueue_lock);
2564 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2566 * EXCEPT that the TQ might not yet be cleared out.
2567 * Skip over those with in-use TQs.
2570 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2571 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2577 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2578 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2579 call = queue_First(&rx_freeCallQueue, rx_call);
2580 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2582 if (rx_stats_active)
2583 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2584 MUTEX_EXIT(&rx_freeCallQueue_lock);
2585 MUTEX_ENTER(&call->lock);
2586 CLEAR_CALL_QUEUE_LOCK(call);
2587 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2588 /* Now, if TQ wasn't cleared earlier, do it now. */
2589 rxi_WaitforTQBusy(call);
2590 if (call->flags & RX_CALL_TQ_CLEARME) {
2591 rxi_ClearTransmitQueue(call, 1);
2592 /*queue_Init(&call->tq);*/
2594 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2595 /* Bind the call to its connection structure */
2597 rxi_ResetCall(call, 1);
2600 call = rxi_Alloc(sizeof(struct rx_call));
2601 #ifdef RXDEBUG_PACKET
2602 call->allNextp = rx_allCallsp;
2603 rx_allCallsp = call;
2605 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2606 #else /* RXDEBUG_PACKET */
2607 rx_atomic_inc(&rx_stats.nCallStructs);
2608 #endif /* RXDEBUG_PACKET */
2610 MUTEX_EXIT(&rx_freeCallQueue_lock);
2611 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2612 MUTEX_ENTER(&call->lock);
2613 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2614 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2615 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2617 /* Initialize once-only items */
2618 queue_Init(&call->tq);
2619 queue_Init(&call->rq);
2620 queue_Init(&call->iovq);
2621 #ifdef RXDEBUG_PACKET
2622 call->rqc = call->tqc = call->iovqc = 0;
2623 #endif /* RXDEBUG_PACKET */
2624 /* Bind the call to its connection structure (prereq for reset) */
2626 rxi_ResetCall(call, 1);
2628 call->channel = channel;
2629 call->callNumber = &conn->callNumber[channel];
2630 call->rwind = conn->rwind[channel];
2631 call->twind = conn->twind[channel];
2632 /* Note that the next expected call number is retained (in
2633 * conn->callNumber[i]), even if we reallocate the call structure
2635 conn->call[channel] = call;
2636 /* if the channel's never been used (== 0), we should start at 1, otherwise
2637 * the call number is valid from the last time this channel was used */
2638 if (*call->callNumber == 0)
2639 *call->callNumber = 1;
2644 /* A call has been inactive long enough that so we can throw away
2645 * state, including the call structure, which is placed on the call
2648 * call->lock amd rx_refcnt_mutex are held upon entry.
2649 * haveCTLock is set when called from rxi_ReapConnections.
2652 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2654 int channel = call->channel;
2655 struct rx_connection *conn = call->conn;
2658 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2659 (*call->callNumber)++;
2661 * We are setting the state to RX_STATE_RESET to
2662 * ensure that no one else will attempt to use this
2663 * call once we drop the refcnt lock. We must drop
2664 * the refcnt lock before calling rxi_ResetCall
2665 * because it cannot be held across acquiring the
2666 * freepktQ lock. NewCall does the same.
2668 call->state = RX_STATE_RESET;
2669 MUTEX_EXIT(&rx_refcnt_mutex);
2670 rxi_ResetCall(call, 0);
2672 MUTEX_ENTER(&conn->conn_call_lock);
2673 if (call->conn->call[channel] == call)
2674 call->conn->call[channel] = 0;
2675 MUTEX_EXIT(&conn->conn_call_lock);
2677 MUTEX_ENTER(&rx_freeCallQueue_lock);
2678 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2679 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2680 /* A call may be free even though its transmit queue is still in use.
2681 * Since we search the call list from head to tail, put busy calls at
2682 * the head of the list, and idle calls at the tail.
2684 if (call->flags & RX_CALL_TQ_BUSY)
2685 queue_Prepend(&rx_freeCallQueue, call);
2687 queue_Append(&rx_freeCallQueue, call);
2688 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2689 queue_Append(&rx_freeCallQueue, call);
2690 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2691 if (rx_stats_active)
2692 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2693 MUTEX_EXIT(&rx_freeCallQueue_lock);
2695 /* Destroy the connection if it was previously slated for
2696 * destruction, i.e. the Rx client code previously called
2697 * rx_DestroyConnection (client connections), or
2698 * rxi_ReapConnections called the same routine (server
2699 * connections). Only do this, however, if there are no
2700 * outstanding calls. Note that for fine grain locking, there appears
2701 * to be a deadlock in that rxi_FreeCall has a call locked and
2702 * DestroyConnectionNoLock locks each call in the conn. But note a
2703 * few lines up where we have removed this call from the conn.
2704 * If someone else destroys a connection, they either have no
2705 * call lock held or are going through this section of code.
2707 MUTEX_ENTER(&conn->conn_data_lock);
2708 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2709 MUTEX_ENTER(&rx_refcnt_mutex);
2711 MUTEX_EXIT(&rx_refcnt_mutex);
2712 MUTEX_EXIT(&conn->conn_data_lock);
2713 #ifdef RX_ENABLE_LOCKS
2715 rxi_DestroyConnectionNoLock(conn);
2717 rxi_DestroyConnection(conn);
2718 #else /* RX_ENABLE_LOCKS */
2719 rxi_DestroyConnection(conn);
2720 #endif /* RX_ENABLE_LOCKS */
2722 MUTEX_EXIT(&conn->conn_data_lock);
2724 MUTEX_ENTER(&rx_refcnt_mutex);
2727 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2728 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2731 rxi_Alloc(size_t size)
2735 if (rx_stats_active) {
2736 rx_atomic_add(&rxi_Allocsize, (int) size);
2737 rx_atomic_inc(&rxi_Alloccnt);
2741 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2742 afs_osi_Alloc_NoSleep(size);
2747 osi_Panic("rxi_Alloc error");
2753 rxi_Free(void *addr, size_t size)
2755 if (rx_stats_active) {
2756 rx_atomic_sub(&rxi_Allocsize, (int) size);
2757 rx_atomic_dec(&rxi_Alloccnt);
2759 osi_Free(addr, size);
2763 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2765 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2766 struct rx_peer *next = NULL;
2770 MUTEX_ENTER(&rx_peerHashTable_lock);
2772 peer_ptr = &rx_peerHashTable[0];
2773 peer_end = &rx_peerHashTable[rx_hashTableSize];
2776 for ( ; peer_ptr < peer_end; peer_ptr++) {
2779 for ( ; peer; peer = next) {
2781 if (host == peer->host)
2786 hashIndex = PEER_HASH(host, port);
2787 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2788 if ((peer->host == host) && (peer->port == port))
2793 MUTEX_ENTER(&rx_peerHashTable_lock);
2798 MUTEX_EXIT(&rx_peerHashTable_lock);
2800 MUTEX_ENTER(&peer->peer_lock);
2801 /* We don't handle dropping below min, so don't */
2802 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2803 peer->ifMTU=MIN(mtu, peer->ifMTU);
2804 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2805 /* if we tweaked this down, need to tune our peer MTU too */
2806 peer->MTU = MIN(peer->MTU, peer->natMTU);
2807 /* if we discovered a sub-1500 mtu, degrade */
2808 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2809 peer->maxDgramPackets = 1;
2810 /* We no longer have valid peer packet information */
2811 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2812 peer->maxPacketSize = 0;
2813 MUTEX_EXIT(&peer->peer_lock);
2815 MUTEX_ENTER(&rx_peerHashTable_lock);
2817 if (host && !port) {
2819 /* pick up where we left off */
2823 MUTEX_EXIT(&rx_peerHashTable_lock);
2826 /* Find the peer process represented by the supplied (host,port)
2827 * combination. If there is no appropriate active peer structure, a
2828 * new one will be allocated and initialized
2829 * The origPeer, if set, is a pointer to a peer structure on which the
2830 * refcount will be be decremented. This is used to replace the peer
2831 * structure hanging off a connection structure */
2833 rxi_FindPeer(afs_uint32 host, u_short port,
2834 struct rx_peer *origPeer, int create)
2838 hashIndex = PEER_HASH(host, port);
2839 MUTEX_ENTER(&rx_peerHashTable_lock);
2840 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2841 if ((pp->host == host) && (pp->port == port))
2846 pp = rxi_AllocPeer(); /* This bzero's *pp */
2847 pp->host = host; /* set here or in InitPeerParams is zero */
2849 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2850 queue_Init(&pp->congestionQueue);
2851 queue_Init(&pp->rpcStats);
2852 pp->next = rx_peerHashTable[hashIndex];
2853 rx_peerHashTable[hashIndex] = pp;
2854 rxi_InitPeerParams(pp);
2855 if (rx_stats_active)
2856 rx_atomic_inc(&rx_stats.nPeerStructs);
2863 origPeer->refCount--;
2864 MUTEX_EXIT(&rx_peerHashTable_lock);
2869 /* Find the connection at (host, port) started at epoch, and with the
2870 * given connection id. Creates the server connection if necessary.
2871 * The type specifies whether a client connection or a server
2872 * connection is desired. In both cases, (host, port) specify the
2873 * peer's (host, pair) pair. Client connections are not made
2874 * automatically by this routine. The parameter socket gives the
2875 * socket descriptor on which the packet was received. This is used,
2876 * in the case of server connections, to check that *new* connections
2877 * come via a valid (port, serviceId). Finally, the securityIndex
2878 * parameter must match the existing index for the connection. If a
2879 * server connection is created, it will be created using the supplied
2880 * index, if the index is valid for this service */
2881 struct rx_connection *
2882 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2883 u_short port, u_short serviceId, afs_uint32 cid,
2884 afs_uint32 epoch, int type, u_int securityIndex)
2886 int hashindex, flag, i;
2887 struct rx_connection *conn;
2888 hashindex = CONN_HASH(host, port, cid, epoch, type);
2889 MUTEX_ENTER(&rx_connHashTable_lock);
2890 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2891 rx_connHashTable[hashindex],
2894 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2895 && (epoch == conn->epoch)) {
2896 struct rx_peer *pp = conn->peer;
2897 if (securityIndex != conn->securityIndex) {
2898 /* this isn't supposed to happen, but someone could forge a packet
2899 * like this, and there seems to be some CM bug that makes this
2900 * happen from time to time -- in which case, the fileserver
2902 MUTEX_EXIT(&rx_connHashTable_lock);
2903 return (struct rx_connection *)0;
2905 if (pp->host == host && pp->port == port)
2907 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2909 /* So what happens when it's a callback connection? */
2910 if ( /*type == RX_CLIENT_CONNECTION && */
2911 (conn->epoch & 0x80000000))
2915 /* the connection rxLastConn that was used the last time is not the
2916 ** one we are looking for now. Hence, start searching in the hash */
2918 conn = rx_connHashTable[hashindex];
2923 struct rx_service *service;
2924 if (type == RX_CLIENT_CONNECTION) {
2925 MUTEX_EXIT(&rx_connHashTable_lock);
2926 return (struct rx_connection *)0;
2928 service = rxi_FindService(socket, serviceId);
2929 if (!service || (securityIndex >= service->nSecurityObjects)
2930 || (service->securityObjects[securityIndex] == 0)) {
2931 MUTEX_EXIT(&rx_connHashTable_lock);
2932 return (struct rx_connection *)0;
2934 conn = rxi_AllocConnection(); /* This bzero's the connection */
2935 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2936 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2937 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2938 conn->next = rx_connHashTable[hashindex];
2939 rx_connHashTable[hashindex] = conn;
2940 conn->peer = rxi_FindPeer(host, port, 0, 1);
2941 conn->type = RX_SERVER_CONNECTION;
2942 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2943 conn->epoch = epoch;
2944 conn->cid = cid & RX_CIDMASK;
2945 /* conn->serial = conn->lastSerial = 0; */
2946 /* conn->timeout = 0; */
2947 conn->ackRate = RX_FAST_ACK_RATE;
2948 conn->service = service;
2949 conn->serviceId = serviceId;
2950 conn->securityIndex = securityIndex;
2951 conn->securityObject = service->securityObjects[securityIndex];
2952 conn->nSpecific = 0;
2953 conn->specific = NULL;
2954 rx_SetConnDeadTime(conn, service->connDeadTime);
2955 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2956 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2957 for (i = 0; i < RX_MAXCALLS; i++) {
2958 conn->twind[i] = rx_initSendWindow;
2959 conn->rwind[i] = rx_initReceiveWindow;
2961 /* Notify security object of the new connection */
2962 RXS_NewConnection(conn->securityObject, conn);
2963 /* XXXX Connection timeout? */
2964 if (service->newConnProc)
2965 (*service->newConnProc) (conn);
2966 if (rx_stats_active)
2967 rx_atomic_inc(&rx_stats.nServerConns);
2970 MUTEX_ENTER(&rx_refcnt_mutex);
2972 MUTEX_EXIT(&rx_refcnt_mutex);
2974 rxLastConn = conn; /* store this connection as the last conn used */
2975 MUTEX_EXIT(&rx_connHashTable_lock);
2980 * Timeout a call on a busy call channel if appropriate.
2982 * @param[in] call The busy call.
2984 * @pre 'call' is marked as busy (namely,
2985 * call->conn->lastBusy[call->channel] != 0)
2987 * @pre call->lock is held
2988 * @pre rxi_busyChannelError is nonzero
2990 * @note call->lock is dropped and reacquired
2993 rxi_CheckBusy(struct rx_call *call)
2995 struct rx_connection *conn = call->conn;
2996 int channel = call->channel;
2997 int freechannel = 0;
2999 afs_uint32 callNumber = *call->callNumber;
3001 MUTEX_EXIT(&call->lock);
3003 MUTEX_ENTER(&conn->conn_call_lock);
3005 /* Are there any other call slots on this conn that we should try? Look for
3006 * slots that are empty and are either non-busy, or were marked as busy
3007 * longer than conn->secondsUntilDead seconds before this call started. */
3009 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3011 /* only look at channels that aren't us */
3015 if (conn->lastBusy[i]) {
3016 /* if this channel looked busy too recently, don't look at it */
3017 if (conn->lastBusy[i] >= call->startTime.sec) {
3020 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3025 if (conn->call[i]) {
3026 struct rx_call *tcall = conn->call[i];
3027 MUTEX_ENTER(&tcall->lock);
3028 if (tcall->state == RX_STATE_DALLY) {
3031 MUTEX_EXIT(&tcall->lock);
3037 MUTEX_EXIT(&conn->conn_call_lock);
3039 MUTEX_ENTER(&call->lock);
3041 /* Since the call->lock and conn->conn_call_lock have been released it is
3042 * possible that (1) the call may no longer be busy and/or (2) the call may
3043 * have been reused by another waiting thread. Therefore, we must confirm
3044 * that the call state has not changed when deciding whether or not to
3045 * force this application thread to retry by forcing a Timeout error. */
3047 if (freechannel && *call->callNumber == callNumber &&
3048 (call->flags & RX_CALL_PEER_BUSY)) {
3049 /* Since 'freechannel' is set, there exists another channel in this
3050 * rx_conn that the application thread might be able to use. We know
3051 * that we have the correct call since callNumber is unchanged, and we
3052 * know that the call is still busy. So, set the call error state to
3053 * rxi_busyChannelError so the application can retry the request,
3054 * presumably on a less-busy call channel. */
3056 rxi_CallError(call, rxi_busyChannelError);
3060 /* There are two packet tracing routines available for testing and monitoring
3061 * Rx. One is called just after every packet is received and the other is
3062 * called just before every packet is sent. Received packets, have had their
3063 * headers decoded, and packets to be sent have not yet had their headers
3064 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3065 * containing the network address. Both can be modified. The return value, if
3066 * non-zero, indicates that the packet should be dropped. */
3068 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3069 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3071 /* A packet has been received off the interface. Np is the packet, socket is
3072 * the socket number it was received from (useful in determining which service
3073 * this packet corresponds to), and (host, port) reflect the host,port of the
3074 * sender. This call returns the packet to the caller if it is finished with
3075 * it, rather than de-allocating it, just as a small performance hack */
3078 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3079 afs_uint32 host, u_short port, int *tnop,
3080 struct rx_call **newcallp)
3082 struct rx_call *call;
3083 struct rx_connection *conn;
3085 afs_uint32 currentCallNumber;
3091 struct rx_packet *tnp;
3094 /* We don't print out the packet until now because (1) the time may not be
3095 * accurate enough until now in the lwp implementation (rx_Listener only gets
3096 * the time after the packet is read) and (2) from a protocol point of view,
3097 * this is the first time the packet has been seen */
3098 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3099 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3100 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3101 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3102 np->header.epoch, np->header.cid, np->header.callNumber,
3103 np->header.seq, np->header.flags, np));
3106 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3107 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3110 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3111 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3114 /* If an input tracer function is defined, call it with the packet and
3115 * network address. Note this function may modify its arguments. */
3116 if (rx_justReceived) {
3117 struct sockaddr_in addr;
3119 addr.sin_family = AF_INET;
3120 addr.sin_port = port;
3121 addr.sin_addr.s_addr = host;
3122 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3123 addr.sin_len = sizeof(addr);
3124 #endif /* AFS_OSF_ENV */
3125 drop = (*rx_justReceived) (np, &addr);
3126 /* drop packet if return value is non-zero */
3129 port = addr.sin_port; /* in case fcn changed addr */
3130 host = addr.sin_addr.s_addr;
3134 /* If packet was not sent by the client, then *we* must be the client */
3135 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3136 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3138 /* Find the connection (or fabricate one, if we're the server & if
3139 * necessary) associated with this packet */
3141 rxi_FindConnection(socket, host, port, np->header.serviceId,
3142 np->header.cid, np->header.epoch, type,
3143 np->header.securityIndex);
3146 /* If no connection found or fabricated, just ignore the packet.
3147 * (An argument could be made for sending an abort packet for
3152 /* If the connection is in an error state, send an abort packet and ignore
3153 * the incoming packet */
3155 /* Don't respond to an abort packet--we don't want loops! */
3156 MUTEX_ENTER(&conn->conn_data_lock);
3157 if (np->header.type != RX_PACKET_TYPE_ABORT)
3158 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3159 MUTEX_ENTER(&rx_refcnt_mutex);
3161 MUTEX_EXIT(&rx_refcnt_mutex);
3162 MUTEX_EXIT(&conn->conn_data_lock);
3166 /* Check for connection-only requests (i.e. not call specific). */
3167 if (np->header.callNumber == 0) {
3168 switch (np->header.type) {
3169 case RX_PACKET_TYPE_ABORT: {
3170 /* What if the supplied error is zero? */
3171 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3172 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3173 rxi_ConnectionError(conn, errcode);
3174 MUTEX_ENTER(&rx_refcnt_mutex);
3176 MUTEX_EXIT(&rx_refcnt_mutex);
3179 case RX_PACKET_TYPE_CHALLENGE:
3180 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3181 MUTEX_ENTER(&rx_refcnt_mutex);
3183 MUTEX_EXIT(&rx_refcnt_mutex);
3185 case RX_PACKET_TYPE_RESPONSE:
3186 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3187 MUTEX_ENTER(&rx_refcnt_mutex);
3189 MUTEX_EXIT(&rx_refcnt_mutex);
3191 case RX_PACKET_TYPE_PARAMS:
3192 case RX_PACKET_TYPE_PARAMS + 1:
3193 case RX_PACKET_TYPE_PARAMS + 2:
3194 /* ignore these packet types for now */
3195 MUTEX_ENTER(&rx_refcnt_mutex);
3197 MUTEX_EXIT(&rx_refcnt_mutex);
3202 /* Should not reach here, unless the peer is broken: send an
3204 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3205 MUTEX_ENTER(&conn->conn_data_lock);
3206 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3207 MUTEX_ENTER(&rx_refcnt_mutex);
3209 MUTEX_EXIT(&rx_refcnt_mutex);
3210 MUTEX_EXIT(&conn->conn_data_lock);
3215 channel = np->header.cid & RX_CHANNELMASK;
3216 call = conn->call[channel];
3219 MUTEX_ENTER(&call->lock);
3220 currentCallNumber = conn->callNumber[channel];
3221 } else if (type == RX_SERVER_CONNECTION) { /* No call allocated */
3222 MUTEX_ENTER(&conn->conn_call_lock);
3223 call = conn->call[channel];
3225 MUTEX_ENTER(&call->lock);
3226 MUTEX_EXIT(&conn->conn_call_lock);
3227 currentCallNumber = conn->callNumber[channel];
3229 call = rxi_NewCall(conn, channel); /* returns locked call */
3230 MUTEX_EXIT(&conn->conn_call_lock);
3231 *call->callNumber = currentCallNumber = np->header.callNumber;
3233 if (np->header.callNumber == 0)
3234 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3235 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3236 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3237 np->header.flags, np, np->length));
3239 call->state = RX_STATE_PRECALL;
3240 clock_GetTime(&call->queueTime);
3241 hzero(call->bytesSent);
3242 hzero(call->bytesRcvd);
3244 * If the number of queued calls exceeds the overload
3245 * threshold then abort this call.
3247 if ((rx_BusyThreshold > 0) &&
3248 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3249 struct rx_packet *tp;
3251 rxi_CallError(call, rx_BusyError);
3252 tp = rxi_SendCallAbort(call, np, 1, 0);
3253 MUTEX_EXIT(&call->lock);
3254 MUTEX_ENTER(&rx_refcnt_mutex);
3256 MUTEX_EXIT(&rx_refcnt_mutex);
3257 if (rx_stats_active)
3258 rx_atomic_inc(&rx_stats.nBusies);
3261 rxi_KeepAliveOn(call);
3263 } else { /* RX_CLIENT_CONNECTION and No call allocated */
3264 /* This packet can't be for this call. If the new call address is
3265 * 0 then no call is running on this channel. If there is a call
3266 * then, since this is a client connection we're getting data for
3267 * it must be for the previous call.
3269 if (rx_stats_active)
3270 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3271 MUTEX_ENTER(&rx_refcnt_mutex);
3273 MUTEX_EXIT(&rx_refcnt_mutex);
3277 /* There is a non-NULL locked call at this point */
3278 if (type == RX_SERVER_CONNECTION) { /* We're the server */
3279 if (np->header.callNumber < currentCallNumber) {
3280 MUTEX_EXIT(&call->lock);
3281 if (rx_stats_active)
3282 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3283 MUTEX_ENTER(&rx_refcnt_mutex);
3285 MUTEX_EXIT(&rx_refcnt_mutex);
3287 } else if (np->header.callNumber != currentCallNumber) {
3288 /* Wait until the transmit queue is idle before deciding
3289 * whether to reset the current call. Chances are that the
3290 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3293 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3294 if (call->state == RX_STATE_ACTIVE) {
3295 rxi_WaitforTQBusy(call);
3297 * If we entered error state while waiting,
3298 * must call rxi_CallError to permit rxi_ResetCall
3299 * to processed when the tqWaiter count hits zero.
3302 rxi_CallError(call, call->error);
3303 MUTEX_EXIT(&call->lock);
3304 MUTEX_ENTER(&rx_refcnt_mutex);
3306 MUTEX_EXIT(&rx_refcnt_mutex);
3310 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3311 /* If the new call cannot be taken right now send a busy and set
3312 * the error condition in this call, so that it terminates as
3313 * quickly as possible */
3314 if (call->state == RX_STATE_ACTIVE) {
3315 struct rx_packet *tp;
3317 rxi_CallError(call, RX_CALL_DEAD);
3318 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3320 MUTEX_EXIT(&call->lock);
3321 MUTEX_ENTER(&rx_refcnt_mutex);
3323 MUTEX_EXIT(&rx_refcnt_mutex);
3326 rxi_ResetCall(call, 0);
3327 *call->callNumber = np->header.callNumber;
3329 if (np->header.callNumber == 0)
3330 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3331 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3332 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3333 np->header.flags, np, np->length));
3335 call->state = RX_STATE_PRECALL;
3336 clock_GetTime(&call->queueTime);
3337 hzero(call->bytesSent);
3338 hzero(call->bytesRcvd);
3340 * If the number of queued calls exceeds the overload
3341 * threshold then abort this call.
3343 if ((rx_BusyThreshold > 0) &&
3344 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3345 struct rx_packet *tp;
3347 rxi_CallError(call, rx_BusyError);
3348 tp = rxi_SendCallAbort(call, np, 1, 0);
3349 MUTEX_EXIT(&call->lock);
3350 MUTEX_ENTER(&rx_refcnt_mutex);
3352 MUTEX_EXIT(&rx_refcnt_mutex);
3353 if (rx_stats_active)
3354 rx_atomic_inc(&rx_stats.nBusies);
3357 rxi_KeepAliveOn(call);
3359 /* Continuing call; do nothing here. */
3361 } else { /* we're the client */
3362 /* Ignore all incoming acknowledgements for calls in DALLY state */
3363 if ((call->state == RX_STATE_DALLY)
3364 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3365 if (rx_stats_active)
3366 rx_atomic_inc(&rx_stats.ignorePacketDally);
3367 MUTEX_EXIT(&call->lock);
3368 MUTEX_ENTER(&rx_refcnt_mutex);
3370 MUTEX_EXIT(&rx_refcnt_mutex);
3374 /* Ignore anything that's not relevant to the current call. If there
3375 * isn't a current call, then no packet is relevant. */
3376 if (np->header.callNumber != currentCallNumber) {
3377 if (rx_stats_active)
3378 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3379 MUTEX_EXIT(&call->lock);
3380 MUTEX_ENTER(&rx_refcnt_mutex);
3382 MUTEX_EXIT(&rx_refcnt_mutex);
3385 /* If the service security object index stamped in the packet does not
3386 * match the connection's security index, ignore the packet */
3387 if (np->header.securityIndex != conn->securityIndex) {
3388 MUTEX_EXIT(&call->lock);
3389 MUTEX_ENTER(&rx_refcnt_mutex);
3391 MUTEX_EXIT(&rx_refcnt_mutex);
3395 /* If we're receiving the response, then all transmit packets are
3396 * implicitly acknowledged. Get rid of them. */
3397 if (np->header.type == RX_PACKET_TYPE_DATA) {
3398 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3399 /* XXX Hack. Because we must release the global rx lock when
3400 * sending packets (osi_NetSend) we drop all acks while we're
3401 * traversing the tq in rxi_Start sending packets out because
3402 * packets may move to the freePacketQueue as result of being here!
3403 * So we drop these packets until we're safely out of the
3404 * traversing. Really ugly!
3405 * For fine grain RX locking, we set the acked field in the
3406 * packets and let rxi_Start remove them from the transmit queue.
3408 if (call->flags & RX_CALL_TQ_BUSY) {
3409 #ifdef RX_ENABLE_LOCKS
3410 rxi_SetAcksInTransmitQueue(call);
3412 MUTEX_ENTER(&rx_refcnt_mutex);
3414 MUTEX_EXIT(&rx_refcnt_mutex);
3415 return np; /* xmitting; drop packet */
3418 rxi_ClearTransmitQueue(call, 0);
3420 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3421 rxi_ClearTransmitQueue(call, 0);
3422 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3424 if (np->header.type == RX_PACKET_TYPE_ACK) {
3425 /* now check to see if this is an ack packet acknowledging that the
3426 * server actually *lost* some hard-acked data. If this happens we
3427 * ignore this packet, as it may indicate that the server restarted in
3428 * the middle of a call. It is also possible that this is an old ack
3429 * packet. We don't abort the connection in this case, because this
3430 * *might* just be an old ack packet. The right way to detect a server
3431 * restart in the midst of a call is to notice that the server epoch
3433 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3434 * XXX unacknowledged. I think that this is off-by-one, but
3435 * XXX I don't dare change it just yet, since it will
3436 * XXX interact badly with the server-restart detection
3437 * XXX code in receiveackpacket. */
3438 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3439 if (rx_stats_active)
3440 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3441 MUTEX_EXIT(&call->lock);
3442 MUTEX_ENTER(&rx_refcnt_mutex);
3444 MUTEX_EXIT(&rx_refcnt_mutex);
3448 } /* else not a data packet */
3451 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3452 /* Set remote user defined status from packet */
3453 call->remoteStatus = np->header.userStatus;
3455 /* Note the gap between the expected next packet and the actual
3456 * packet that arrived, when the new packet has a smaller serial number
3457 * than expected. Rioses frequently reorder packets all by themselves,
3458 * so this will be quite important with very large window sizes.
3459 * Skew is checked against 0 here to avoid any dependence on the type of
3460 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
3462 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
3463 * see CalculateRoundTripTime for an example of how to keep smoothed values.
3464 * I think using a beta of 1/8 is probably appropriate. 93.04.21
3466 MUTEX_ENTER(&conn->conn_data_lock);
3467 skew = conn->lastSerial - np->header.serial;
3468 conn->lastSerial = np->header.serial;
3469 MUTEX_EXIT(&conn->conn_data_lock);
3471 struct rx_peer *peer;
3473 if (skew > peer->inPacketSkew) {
3474 dpf(("*** In skew changed from %d to %d\n",
3475 peer->inPacketSkew, skew));
3476 peer->inPacketSkew = skew;
3480 /* Now do packet type-specific processing */
3481 switch (np->header.type) {
3482 case RX_PACKET_TYPE_DATA:
3483 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3486 case RX_PACKET_TYPE_ACK:
3487 /* Respond immediately to ack packets requesting acknowledgement
3489 if (np->header.flags & RX_REQUEST_ACK) {
3491 (void)rxi_SendCallAbort(call, 0, 1, 0);
3493 (void)rxi_SendAck(call, 0, np->header.serial,
3494 RX_ACK_PING_RESPONSE, 1);
3496 np = rxi_ReceiveAckPacket(call, np, 1);
3498 case RX_PACKET_TYPE_ABORT: {
3499 /* An abort packet: reset the call, passing the error up to the user. */
3500 /* What if error is zero? */
3501 /* What if the error is -1? the application will treat it as a timeout. */
3502 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3503 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3504 rxi_CallError(call, errdata);
3505 MUTEX_EXIT(&call->lock);
3506 MUTEX_ENTER(&rx_refcnt_mutex);
3508 MUTEX_EXIT(&rx_refcnt_mutex);
3509 return np; /* xmitting; drop packet */
3511 case RX_PACKET_TYPE_BUSY: {
3512 struct clock busyTime;
3514 clock_GetTime(&busyTime);
3516 MUTEX_EXIT(&call->lock);
3518 MUTEX_ENTER(&conn->conn_call_lock);
3519 MUTEX_ENTER(&call->lock);
3520 conn->lastBusy[call->channel] = busyTime.sec;
3521 call->flags |= RX_CALL_PEER_BUSY;
3522 MUTEX_EXIT(&call->lock);
3523 MUTEX_EXIT(&conn->conn_call_lock);
3525 MUTEX_ENTER(&rx_refcnt_mutex);
3527 MUTEX_EXIT(&rx_refcnt_mutex);
3531 case RX_PACKET_TYPE_ACKALL:
3532 /* All packets acknowledged, so we can drop all packets previously
3533 * readied for sending */
3534 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3535 /* XXX Hack. We because we can't release the global rx lock when
3536 * sending packets (osi_NetSend) we drop all ack pkts while we're
3537 * traversing the tq in rxi_Start sending packets out because
3538 * packets may move to the freePacketQueue as result of being
3539 * here! So we drop these packets until we're safely out of the
3540 * traversing. Really ugly!
3541 * For fine grain RX locking, we set the acked field in the packets
3542 * and let rxi_Start remove the packets from the transmit queue.
3544 if (call->flags & RX_CALL_TQ_BUSY) {
3545 #ifdef RX_ENABLE_LOCKS
3546 rxi_SetAcksInTransmitQueue(call);
3548 #else /* RX_ENABLE_LOCKS */
3549 MUTEX_EXIT(&call->lock);
3550 MUTEX_ENTER(&rx_refcnt_mutex);
3552 MUTEX_EXIT(&rx_refcnt_mutex);
3553 return np; /* xmitting; drop packet */
3554 #endif /* RX_ENABLE_LOCKS */
3556 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3557 rxi_ClearTransmitQueue(call, 0);
3560 /* Should not reach here, unless the peer is broken: send an abort
3562 rxi_CallError(call, RX_PROTOCOL_ERROR);
3563 np = rxi_SendCallAbort(call, np, 1, 0);
3566 /* Note when this last legitimate packet was received, for keep-alive
3567 * processing. Note, we delay getting the time until now in the hope that
3568 * the packet will be delivered to the user before any get time is required
3569 * (if not, then the time won't actually be re-evaluated here). */
3570 call->lastReceiveTime = clock_Sec();
3571 /* we've received a legit packet, so the channel is not busy */
3572 call->flags &= ~RX_CALL_PEER_BUSY;
3573 MUTEX_EXIT(&call->lock);
3574 MUTEX_ENTER(&rx_refcnt_mutex);
3576 MUTEX_EXIT(&rx_refcnt_mutex);
3580 /* return true if this is an "interesting" connection from the point of view
3581 of someone trying to debug the system */
3583 rxi_IsConnInteresting(struct rx_connection *aconn)
3586 struct rx_call *tcall;
3588 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3591 for (i = 0; i < RX_MAXCALLS; i++) {
3592 tcall = aconn->call[i];
3594 if ((tcall->state == RX_STATE_PRECALL)
3595 || (tcall->state == RX_STATE_ACTIVE))
3597 if ((tcall->mode == RX_MODE_SENDING)
3598 || (tcall->mode == RX_MODE_RECEIVING))
3606 /* if this is one of the last few packets AND it wouldn't be used by the
3607 receiving call to immediately satisfy a read request, then drop it on
3608 the floor, since accepting it might prevent a lock-holding thread from
3609 making progress in its reading. If a call has been cleared while in
3610 the precall state then ignore all subsequent packets until the call
3611 is assigned to a thread. */
3614 TooLow(struct rx_packet *ap, struct rx_call *acall)
3618 MUTEX_ENTER(&rx_quota_mutex);
3619 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3620 && (acall->state == RX_STATE_PRECALL))
3621 || ((rx_nFreePackets < rxi_dataQuota + 2)
3622 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3623 && (acall->flags & RX_CALL_READER_WAIT)))) {
3626 MUTEX_EXIT(&rx_quota_mutex);
3632 * Clear the attach wait flag on a connection and proceed.
3634 * Any processing waiting for a connection to be attached should be
3635 * unblocked. We clear the flag and do any other needed tasks.
3638 * the conn to unmark waiting for attach
3640 * @pre conn's conn_data_lock must be locked before calling this function
3644 rxi_ConnClearAttachWait(struct rx_connection *conn)
3646 /* Indicate that rxi_CheckReachEvent is no longer running by
3647 * clearing the flag. Must be atomic under conn_data_lock to
3648 * avoid a new call slipping by: rxi_CheckConnReach holds
3649 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3651 conn->flags &= ~RX_CONN_ATTACHWAIT;
3652 if (conn->flags & RX_CONN_NAT_PING) {
3653 conn->flags &= ~RX_CONN_NAT_PING;
3654 rxi_ScheduleNatKeepAliveEvent(conn);
3659 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3661 struct rx_connection *conn = arg1;
3662 struct rx_call *acall = arg2;
3663 struct rx_call *call = acall;
3664 struct clock when, now;
3667 MUTEX_ENTER(&conn->conn_data_lock);
3668 conn->checkReachEvent = NULL;
3669 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3671 MUTEX_ENTER(&rx_refcnt_mutex);
3673 MUTEX_EXIT(&rx_refcnt_mutex);
3675 MUTEX_EXIT(&conn->conn_data_lock);
3679 MUTEX_ENTER(&conn->conn_call_lock);
3680 MUTEX_ENTER(&conn->conn_data_lock);
3681 for (i = 0; i < RX_MAXCALLS; i++) {
3682 struct rx_call *tc = conn->call[i];
3683 if (tc && tc->state == RX_STATE_PRECALL) {
3689 rxi_ConnClearAttachWait(conn);
3690 MUTEX_EXIT(&conn->conn_data_lock);
3691 MUTEX_EXIT(&conn->conn_call_lock);
3696 MUTEX_ENTER(&call->lock);
3697 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3699 MUTEX_EXIT(&call->lock);
3701 clock_GetTime(&now);
3703 when.sec += RX_CHECKREACH_TIMEOUT;
3704 MUTEX_ENTER(&conn->conn_data_lock);
3705 if (!conn->checkReachEvent) {
3706 MUTEX_ENTER(&rx_refcnt_mutex);
3708 MUTEX_EXIT(&rx_refcnt_mutex);
3709 conn->checkReachEvent =
3710 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3713 MUTEX_EXIT(&conn->conn_data_lock);
3719 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3721 struct rx_service *service = conn->service;
3722 struct rx_peer *peer = conn->peer;
3723 afs_uint32 now, lastReach;
3725 if (service->checkReach == 0)
3729 MUTEX_ENTER(&peer->peer_lock);
3730 lastReach = peer->lastReachTime;
3731 MUTEX_EXIT(&peer->peer_lock);
3732 if (now - lastReach < RX_CHECKREACH_TTL)
3735 MUTEX_ENTER(&conn->conn_data_lock);
3736 if (conn->flags & RX_CONN_ATTACHWAIT) {
3737 MUTEX_EXIT(&conn->conn_data_lock);
3740 conn->flags |= RX_CONN_ATTACHWAIT;
3741 MUTEX_EXIT(&conn->conn_data_lock);
3742 if (!conn->checkReachEvent)
3743 rxi_CheckReachEvent(NULL, conn, call);
3748 /* try to attach call, if authentication is complete */
3750 TryAttach(struct rx_call *acall, osi_socket socket,
3751 int *tnop, struct rx_call **newcallp,
3754 struct rx_connection *conn = acall->conn;
3756 if (conn->type == RX_SERVER_CONNECTION
3757 && acall->state == RX_STATE_PRECALL) {
3758 /* Don't attach until we have any req'd. authentication. */
3759 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3760 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3761 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3762 /* Note: this does not necessarily succeed; there
3763 * may not any proc available
3766 rxi_ChallengeOn(acall->conn);
3771 /* A data packet has been received off the interface. This packet is
3772 * appropriate to the call (the call is in the right state, etc.). This
3773 * routine can return a packet to the caller, for re-use */
3776 rxi_ReceiveDataPacket(struct rx_call *call,
3777 struct rx_packet *np, int istack,
3778 osi_socket socket, afs_uint32 host, u_short port,
3779 int *tnop, struct rx_call **newcallp)
3781 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3786 afs_uint32 serial=0, flags=0;
3788 struct rx_packet *tnp;
3789 struct clock when, now;
3790 if (rx_stats_active)
3791 rx_atomic_inc(&rx_stats.dataPacketsRead);
3794 /* If there are no packet buffers, drop this new packet, unless we can find
3795 * packet buffers from inactive calls */
3797 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3798 MUTEX_ENTER(&rx_freePktQ_lock);
3799 rxi_NeedMorePackets = TRUE;
3800 MUTEX_EXIT(&rx_freePktQ_lock);
3801 if (rx_stats_active)
3802 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3803 call->rprev = np->header.serial;
3804 rxi_calltrace(RX_TRACE_DROP, call);
3805 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3807 rxi_ClearReceiveQueue(call);
3808 clock_GetTime(&now);
3810 clock_Add(&when, &rx_softAckDelay);
3811 if (!call->delayedAckEvent
3812 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3813 rxevent_Cancel(call->delayedAckEvent, call,
3814 RX_CALL_REFCOUNT_DELAY);
3815 MUTEX_ENTER(&rx_refcnt_mutex);
3816 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3817 MUTEX_EXIT(&rx_refcnt_mutex);
3819 call->delayedAckEvent =
3820 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3822 /* we've damaged this call already, might as well do it in. */
3828 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3829 * packet is one of several packets transmitted as a single
3830 * datagram. Do not send any soft or hard acks until all packets
3831 * in a jumbogram have been processed. Send negative acks right away.
3833 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3834 /* tnp is non-null when there are more packets in the
3835 * current jumbo gram */
3842 seq = np->header.seq;
3843 serial = np->header.serial;
3844 flags = np->header.flags;
3846 /* If the call is in an error state, send an abort message */
3848 return rxi_SendCallAbort(call, np, istack, 0);
3850 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3851 * AFS 3.5 jumbogram. */
3852 if (flags & RX_JUMBO_PACKET) {
3853 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3858 if (np->header.spare != 0) {
3859 MUTEX_ENTER(&call->conn->conn_data_lock);
3860 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3861 MUTEX_EXIT(&call->conn->conn_data_lock);
3864 /* The usual case is that this is the expected next packet */
3865 if (seq == call->rnext) {
3867 /* Check to make sure it is not a duplicate of one already queued */
3868 if (queue_IsNotEmpty(&call->rq)
3869 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3870 if (rx_stats_active)
3871 rx_atomic_inc(&rx_stats.dupPacketsRead);
3872 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3873 rxevent_Cancel(call->delayedAckEvent, call,
3874 RX_CALL_REFCOUNT_DELAY);
3875 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3881 /* It's the next packet. Stick it on the receive queue
3882 * for this call. Set newPackets to make sure we wake
3883 * the reader once all packets have been processed */
3884 #ifdef RX_TRACK_PACKETS
3885 np->flags |= RX_PKTFLAG_RQ;
3887 queue_Prepend(&call->rq, np);
3888 #ifdef RXDEBUG_PACKET
3890 #endif /* RXDEBUG_PACKET */
3892 np = NULL; /* We can't use this anymore */
3895 /* If an ack is requested then set a flag to make sure we
3896 * send an acknowledgement for this packet */
3897 if (flags & RX_REQUEST_ACK) {
3898 ackNeeded = RX_ACK_REQUESTED;
3901 /* Keep track of whether we have received the last packet */
3902 if (flags & RX_LAST_PACKET) {
3903 call->flags |= RX_CALL_HAVE_LAST;
3907 /* Check whether we have all of the packets for this call */
3908 if (call->flags & RX_CALL_HAVE_LAST) {
3909 afs_uint32 tseq; /* temporary sequence number */
3910 struct rx_packet *tp; /* Temporary packet pointer */
3911 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3913 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3914 if (tseq != tp->header.seq)
3916 if (tp->header.flags & RX_LAST_PACKET) {
3917 call->flags |= RX_CALL_RECEIVE_DONE;
3924 /* Provide asynchronous notification for those who want it
3925 * (e.g. multi rx) */
3926 if (call->arrivalProc) {
3927 (*call->arrivalProc) (call, call->arrivalProcHandle,
3928 call->arrivalProcArg);
3929 call->arrivalProc = (void (*)())0;
3932 /* Update last packet received */
3935 /* If there is no server process serving this call, grab
3936 * one, if available. We only need to do this once. If a
3937 * server thread is available, this thread becomes a server
3938 * thread and the server thread becomes a listener thread. */
3940 TryAttach(call, socket, tnop, newcallp, 0);
3943 /* This is not the expected next packet. */
3945 /* Determine whether this is a new or old packet, and if it's
3946 * a new one, whether it fits into the current receive window.
3947 * Also figure out whether the packet was delivered in sequence.
3948 * We use the prev variable to determine whether the new packet
3949 * is the successor of its immediate predecessor in the
3950 * receive queue, and the missing flag to determine whether
3951 * any of this packets predecessors are missing. */
3953 afs_uint32 prev; /* "Previous packet" sequence number */
3954 struct rx_packet *tp; /* Temporary packet pointer */
3955 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3956 int missing; /* Are any predecessors missing? */
3958 /* If the new packet's sequence number has been sent to the
3959 * application already, then this is a duplicate */
3960 if (seq < call->rnext) {
3961 if (rx_stats_active)
3962 rx_atomic_inc(&rx_stats.dupPacketsRead);
3963 rxevent_Cancel(call->delayedAckEvent, call,
3964 RX_CALL_REFCOUNT_DELAY);
3965 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3971 /* If the sequence number is greater than what can be
3972 * accomodated by the current window, then send a negative
3973 * acknowledge and drop the packet */
3974 if ((call->rnext + call->rwind) <= seq) {
3975 rxevent_Cancel(call->delayedAckEvent, call,
3976 RX_CALL_REFCOUNT_DELAY);
3977 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3984 /* Look for the packet in the queue of old received packets */
3985 for (prev = call->rnext - 1, missing =
3986 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3987 /*Check for duplicate packet */
3988 if (seq == tp->header.seq) {
3989 if (rx_stats_active)
3990 rx_atomic_inc(&rx_stats.dupPacketsRead);
3991 rxevent_Cancel(call->delayedAckEvent, call,
3992 RX_CALL_REFCOUNT_DELAY);
3993 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3999 /* If we find a higher sequence packet, break out and
4000 * insert the new packet here. */
4001 if (seq < tp->header.seq)
4003 /* Check for missing packet */
4004 if (tp->header.seq != prev + 1) {
4008 prev = tp->header.seq;
4011 /* Keep track of whether we have received the last packet. */
4012 if (flags & RX_LAST_PACKET) {
4013 call->flags |= RX_CALL_HAVE_LAST;
4016 /* It's within the window: add it to the the receive queue.
4017 * tp is left by the previous loop either pointing at the
4018 * packet before which to insert the new packet, or at the
4019 * queue head if the queue is empty or the packet should be
4021 #ifdef RX_TRACK_PACKETS
4022 np->flags |= RX_PKTFLAG_RQ;
4024 #ifdef RXDEBUG_PACKET
4026 #endif /* RXDEBUG_PACKET */
4027 queue_InsertBefore(tp, np);
4031 /* Check whether we have all of the packets for this call */
4032 if ((call->flags & RX_CALL_HAVE_LAST)
4033 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4034 afs_uint32 tseq; /* temporary sequence number */
4037 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
4038 if (tseq != tp->header.seq)
4040 if (tp->header.flags & RX_LAST_PACKET) {
4041 call->flags |= RX_CALL_RECEIVE_DONE;
4048 /* We need to send an ack of the packet is out of sequence,
4049 * or if an ack was requested by the peer. */
4050 if (seq != prev + 1 || missing) {
4051 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4052 } else if (flags & RX_REQUEST_ACK) {
4053 ackNeeded = RX_ACK_REQUESTED;
4056 /* Acknowledge the last packet for each call */
4057 if (flags & RX_LAST_PACKET) {
4068 * If the receiver is waiting for an iovec, fill the iovec
4069 * using the data from the receive queue */
4070 if (call->flags & RX_CALL_IOVEC_WAIT) {
4071 didHardAck = rxi_FillReadVec(call, serial);
4072 /* the call may have been aborted */
4081 /* Wakeup the reader if any */
4082 if ((call->flags & RX_CALL_READER_WAIT)
4083 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4084 || (call->iovNext >= call->iovMax)
4085 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4086 call->flags &= ~RX_CALL_READER_WAIT;
4087 #ifdef RX_ENABLE_LOCKS
4088 CV_BROADCAST(&call->cv_rq);
4090 osi_rxWakeup(&call->rq);
4096 * Send an ack when requested by the peer, or once every
4097 * rxi_SoftAckRate packets until the last packet has been
4098 * received. Always send a soft ack for the last packet in
4099 * the server's reply. */
4101 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4102 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4103 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4104 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4105 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4106 } else if (call->nSoftAcks) {
4107 clock_GetTime(&now);
4109 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
4110 clock_Add(&when, &rx_lastAckDelay);
4112 clock_Add(&when, &rx_softAckDelay);
4114 if (!call->delayedAckEvent
4115 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
4116 rxevent_Cancel(call->delayedAckEvent, call,
4117 RX_CALL_REFCOUNT_DELAY);
4118 MUTEX_ENTER(&rx_refcnt_mutex);
4119 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
4120 MUTEX_EXIT(&rx_refcnt_mutex);
4121 call->delayedAckEvent =
4122 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
4124 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4125 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4132 static void rxi_ComputeRate();
4136 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4138 struct rx_peer *peer = conn->peer;
4140 MUTEX_ENTER(&peer->peer_lock);
4141 peer->lastReachTime = clock_Sec();
4142 MUTEX_EXIT(&peer->peer_lock);
4144 MUTEX_ENTER(&conn->conn_data_lock);
4145 if (conn->flags & RX_CONN_ATTACHWAIT) {
4148 rxi_ConnClearAttachWait(conn);
4149 MUTEX_EXIT(&conn->conn_data_lock);
4151 for (i = 0; i < RX_MAXCALLS; i++) {
4152 struct rx_call *call = conn->call[i];
4155 MUTEX_ENTER(&call->lock);
4156 /* tnop can be null if newcallp is null */
4157 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4159 MUTEX_EXIT(&call->lock);
4163 MUTEX_EXIT(&conn->conn_data_lock);
4166 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4168 rx_ack_reason(int reason)
4171 case RX_ACK_REQUESTED:
4173 case RX_ACK_DUPLICATE:
4175 case RX_ACK_OUT_OF_SEQUENCE:
4177 case RX_ACK_EXCEEDS_WINDOW:
4179 case RX_ACK_NOSPACE:
4183 case RX_ACK_PING_RESPONSE:
4196 /* The real smarts of the whole thing. */
4198 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4201 struct rx_ackPacket *ap;
4203 struct rx_packet *tp;
4204 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
4205 struct rx_connection *conn = call->conn;
4206 struct rx_peer *peer = conn->peer;
4207 struct clock now; /* Current time, for RTT calculations */
4211 /* because there are CM's that are bogus, sending weird values for this. */
4212 afs_uint32 skew = 0;
4217 int newAckCount = 0;
4218 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4219 int pktsize = 0; /* Set if we need to update the peer mtu */
4220 int conn_data_locked = 0;
4222 if (rx_stats_active)
4223 rx_atomic_inc(&rx_stats.ackPacketsRead);
4224 ap = (struct rx_ackPacket *)rx_DataOf(np);
4225 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4227 return np; /* truncated ack packet */
4229 /* depends on ack packet struct */
4230 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4231 first = ntohl(ap->firstPacket);
4232 prev = ntohl(ap->previousPacket);
4233 serial = ntohl(ap->serial);
4234 /* temporarily disabled -- needs to degrade over time
4235 * skew = ntohs(ap->maxSkew); */
4237 /* Ignore ack packets received out of order */
4238 if (first < call->tfirst ||
4239 (first == call->tfirst && prev < call->tprev)) {
4245 if (np->header.flags & RX_SLOW_START_OK) {
4246 call->flags |= RX_CALL_SLOW_START_OK;
4249 if (ap->reason == RX_ACK_PING_RESPONSE)
4250 rxi_UpdatePeerReach(conn, call);
4252 if (conn->lastPacketSizeSeq) {
4253 MUTEX_ENTER(&conn->conn_data_lock);
4254 conn_data_locked = 1;
4255 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4256 pktsize = conn->lastPacketSize;
4257 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4260 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4261 if (!conn_data_locked) {
4262 MUTEX_ENTER(&conn->conn_data_lock);
4263 conn_data_locked = 1;
4265 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4266 /* process mtu ping ack */
4267 pktsize = conn->lastPingSize;
4268 conn->lastPingSizeSer = conn->lastPingSize = 0;
4272 if (conn_data_locked) {
4273 MUTEX_EXIT(&conn->conn_data_lock);
4274 conn_data_locked = 0;
4278 if (rxdebug_active) {
4282 len = _snprintf(msg, sizeof(msg),
4283 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
4284 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4285 ntohl(ap->serial), ntohl(ap->previousPacket),
4286 (unsigned int)np->header.seq, (unsigned int)skew,
4287 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
4291 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4292 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4296 OutputDebugString(msg);
4298 #else /* AFS_NT40_ENV */
4301 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
4302 ap->reason, ntohl(ap->previousPacket),
4303 (unsigned int)np->header.seq, (unsigned int)serial,
4304 (unsigned int)skew, ntohl(ap->firstPacket));
4307 for (offset = 0; offset < nAcks; offset++)
4308 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4313 #endif /* AFS_NT40_ENV */
4316 MUTEX_ENTER(&peer->peer_lock);
4319 * Start somewhere. Can't assume we can send what we can receive,
4320 * but we are clearly receiving.
4322 if (!peer->maxPacketSize)
4323 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4325 if (pktsize > peer->maxPacketSize) {
4326 peer->maxPacketSize = pktsize;
4327 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4328 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4329 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4330 rxi_ScheduleGrowMTUEvent(call, 1);
4335 /* Update the outgoing packet skew value to the latest value of
4336 * the peer's incoming packet skew value. The ack packet, of
4337 * course, could arrive out of order, but that won't affect things
4339 peer->outPacketSkew = skew;
4342 clock_GetTime(&now);
4344 /* The transmit queue splits into 4 sections.
4346 * The first section is packets which have now been acknowledged
4347 * by a window size change in the ack. These have reached the
4348 * application layer, and may be discarded. These are packets
4349 * with sequence numbers < ap->firstPacket.
4351 * The second section is packets which have sequence numbers in
4352 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4353 * contents of the packet's ack array determines whether these
4354 * packets are acknowledged or not.
4356 * The third section is packets which fall above the range
4357 * addressed in the ack packet. These have not yet been received
4360 * The four section is packets which have not yet been transmitted.
4361 * These packets will have a header.serial of 0.
4364 /* First section - implicitly acknowledged packets that can be
4368 tp = queue_First(&call->tq, rx_packet);
4369 while(!queue_IsEnd(&call->tq, tp) && tp->header.seq < first) {
4370 struct rx_packet *next;
4372 next = queue_Next(tp, rx_packet);
4373 call->tfirst = tp->header.seq + 1;
4375 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4377 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4381 rxi_ComputeRate(call->conn->peer, call, p, np, ap->reason);
4384 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4385 /* XXX Hack. Because we have to release the global rx lock when sending
4386 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4387 * in rxi_Start sending packets out because packets may move to the
4388 * freePacketQueue as result of being here! So we drop these packets until
4389 * we're safely out of the traversing. Really ugly!
4390 * To make it even uglier, if we're using fine grain locking, we can
4391 * set the ack bits in the packets and have rxi_Start remove the packets
4392 * when it's done transmitting.
4394 if (call->flags & RX_CALL_TQ_BUSY) {
4395 #ifdef RX_ENABLE_LOCKS
4396 tp->flags |= RX_PKTFLAG_ACKED;
4397 call->flags |= RX_CALL_TQ_SOME_ACKED;
4398 #else /* RX_ENABLE_LOCKS */
4400 #endif /* RX_ENABLE_LOCKS */
4402 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4405 #ifdef RX_TRACK_PACKETS
4406 tp->flags &= ~RX_PKTFLAG_TQ;
4408 #ifdef RXDEBUG_PACKET
4410 #endif /* RXDEBUG_PACKET */
4411 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4417 /* Give rate detector a chance to respond to ping requests */
4418 if (ap->reason == RX_ACK_PING_RESPONSE) {
4419 rxi_ComputeRate(peer, call, 0, np, ap->reason);
4423 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4425 /* Second section of the queue - packets for which we are receiving
4428 * Go through the explicit acks/nacks and record the results in
4429 * the waiting packets. These are packets that can't be released
4430 * yet, even with a positive acknowledge. This positive
4431 * acknowledge only means the packet has been received by the
4432 * peer, not that it will be retained long enough to be sent to
4433 * the peer's upper level. In addition, reset the transmit timers
4434 * of any missing packets (those packets that must be missing
4435 * because this packet was out of sequence) */
4437 call->nSoftAcked = 0;
4439 while (!queue_IsEnd(&call->tq, tp) && tp->header.seq < first + nAcks) {
4440 /* Set the acknowledge flag per packet based on the
4441 * information in the ack packet. An acknowlegded packet can
4442 * be downgraded when the server has discarded a packet it
4443 * soacked previously, or when an ack packet is received
4444 * out of sequence. */
4445 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4446 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4448 tp->flags |= RX_PKTFLAG_ACKED;
4449 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4451 rxi_ComputeRate(call->conn->peer, call, tp, np, ap->reason);
4459 } else /* RX_ACK_TYPE_NACK */ {
4460 tp->flags &= ~RX_PKTFLAG_ACKED;
4464 tp = queue_Next(tp, rx_packet);
4467 /* We don't need to take any action with the 3rd or 4th section in the
4468 * queue - they're not addressed by the contents of this ACK packet.
4471 /* If the window has been extended by this acknowledge packet,
4472 * then wakeup a sender waiting in alloc for window space, or try
4473 * sending packets now, if he's been sitting on packets due to
4474 * lack of window space */
4475 if (call->tnext < (call->tfirst + call->twind)) {
4476 #ifdef RX_ENABLE_LOCKS
4477 CV_SIGNAL(&call->cv_twind);
4479 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4480 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4481 osi_rxWakeup(&call->twind);
4484 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4485 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4489 /* if the ack packet has a receivelen field hanging off it,
4490 * update our state */
4491 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4494 /* If the ack packet has a "recommended" size that is less than
4495 * what I am using now, reduce my size to match */
4496 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4497 (int)sizeof(afs_int32), &tSize);
4498 tSize = (afs_uint32) ntohl(tSize);
4499 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4501 /* Get the maximum packet size to send to this peer */
4502 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4504 tSize = (afs_uint32) ntohl(tSize);
4505 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4506 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4508 /* sanity check - peer might have restarted with different params.
4509 * If peer says "send less", dammit, send less... Peer should never
4510 * be unable to accept packets of the size that prior AFS versions would
4511 * send without asking. */
4512 if (peer->maxMTU != tSize) {
4513 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4515 peer->maxMTU = tSize;
4516 peer->MTU = MIN(tSize, peer->MTU);
4517 call->MTU = MIN(call->MTU, tSize);
4520 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4523 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4524 (int)sizeof(afs_int32), &tSize);
4525 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4526 if (tSize < call->twind) { /* smaller than our send */
4527 call->twind = tSize; /* window, we must send less... */
4528 call->ssthresh = MIN(call->twind, call->ssthresh);
4529 call->conn->twind[call->channel] = call->twind;
4532 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4533 * network MTU confused with the loopback MTU. Calculate the
4534 * maximum MTU here for use in the slow start code below.
4536 /* Did peer restart with older RX version? */
4537 if (peer->maxDgramPackets > 1) {
4538 peer->maxDgramPackets = 1;
4540 } else if (np->length >=
4541 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4544 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4545 sizeof(afs_int32), &tSize);
4546 tSize = (afs_uint32) ntohl(tSize);
4548 * As of AFS 3.5 we set the send window to match the receive window.
4550 if (tSize < call->twind) {
4551 call->twind = tSize;
4552 call->conn->twind[call->channel] = call->twind;
4553 call->ssthresh = MIN(call->twind, call->ssthresh);
4554 } else if (tSize > call->twind) {
4555 call->twind = tSize;
4556 call->conn->twind[call->channel] = call->twind;
4560 * As of AFS 3.5, a jumbogram is more than one fixed size
4561 * packet transmitted in a single UDP datagram. If the remote
4562 * MTU is smaller than our local MTU then never send a datagram
4563 * larger than the natural MTU.
4566 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4567 (int)sizeof(afs_int32), &tSize);
4568 maxDgramPackets = (afs_uint32) ntohl(tSize);
4569 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4571 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4572 if (maxDgramPackets > 1) {
4573 peer->maxDgramPackets = maxDgramPackets;
4574 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4576 peer->maxDgramPackets = 1;
4577 call->MTU = peer->natMTU;
4579 } else if (peer->maxDgramPackets > 1) {
4580 /* Restarted with lower version of RX */
4581 peer->maxDgramPackets = 1;
4583 } else if (peer->maxDgramPackets > 1
4584 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4585 /* Restarted with lower version of RX */
4586 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4587 peer->natMTU = OLD_MAX_PACKET_SIZE;
4588 peer->MTU = OLD_MAX_PACKET_SIZE;
4589 peer->maxDgramPackets = 1;
4590 peer->nDgramPackets = 1;
4592 call->MTU = OLD_MAX_PACKET_SIZE;
4597 * Calculate how many datagrams were successfully received after
4598 * the first missing packet and adjust the negative ack counter
4603 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4604 if (call->nNacks < nNacked) {
4605 call->nNacks = nNacked;
4608 call->nAcks += newAckCount;
4612 /* If the packet contained new acknowledgements, rather than just
4613 * being a duplicate of one we have previously seen, then we can restart
4616 if (newAckCount > 0)
4617 rxi_rto_packet_acked(call, istack);
4619 if (call->flags & RX_CALL_FAST_RECOVER) {
4620 if (newAckCount == 0) {
4621 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4623 call->flags &= ~RX_CALL_FAST_RECOVER;
4624 call->cwind = call->nextCwind;
4625 call->nextCwind = 0;
4628 call->nCwindAcks = 0;
4629 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4630 /* Three negative acks in a row trigger congestion recovery */
4631 call->flags |= RX_CALL_FAST_RECOVER;
4632 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4634 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4635 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4636 call->nextCwind = call->ssthresh;
4639 peer->MTU = call->MTU;
4640 peer->cwind = call->nextCwind;
4641 peer->nDgramPackets = call->nDgramPackets;
4643 call->congestSeq = peer->congestSeq;
4645 /* Reset the resend times on the packets that were nacked
4646 * so we will retransmit as soon as the window permits
4649 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4651 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4652 tp->flags &= ~RX_PKTFLAG_SENT;
4654 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4659 /* If cwind is smaller than ssthresh, then increase
4660 * the window one packet for each ack we receive (exponential
4662 * If cwind is greater than or equal to ssthresh then increase
4663 * the congestion window by one packet for each cwind acks we
4664 * receive (linear growth). */
4665 if (call->cwind < call->ssthresh) {
4667 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4668 call->nCwindAcks = 0;
4670 call->nCwindAcks += newAckCount;
4671 if (call->nCwindAcks >= call->cwind) {
4672 call->nCwindAcks = 0;
4673 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4677 * If we have received several acknowledgements in a row then
4678 * it is time to increase the size of our datagrams
4680 if ((int)call->nAcks > rx_nDgramThreshold) {
4681 if (peer->maxDgramPackets > 1) {
4682 if (call->nDgramPackets < peer->maxDgramPackets) {
4683 call->nDgramPackets++;
4685 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4686 } else if (call->MTU < peer->maxMTU) {
4687 /* don't upgrade if we can't handle it */
4688 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4689 call->MTU = peer->ifMTU;
4691 call->MTU += peer->natMTU;
4692 call->MTU = MIN(call->MTU, peer->maxMTU);
4699 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4701 /* Servers need to hold the call until all response packets have
4702 * been acknowledged. Soft acks are good enough since clients
4703 * are not allowed to clear their receive queues. */
4704 if (call->state == RX_STATE_HOLD
4705 && call->tfirst + call->nSoftAcked >= call->tnext) {
4706 call->state = RX_STATE_DALLY;
4707 rxi_ClearTransmitQueue(call, 0);
4708 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4709 } else if (!queue_IsEmpty(&call->tq)) {
4710 rxi_Start(call, istack);
4715 /* Received a response to a challenge packet */
4717 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4718 struct rx_packet *np, int istack)
4722 /* Ignore the packet if we're the client */
4723 if (conn->type == RX_CLIENT_CONNECTION)
4726 /* If already authenticated, ignore the packet (it's probably a retry) */
4727 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4730 /* Otherwise, have the security object evaluate the response packet */
4731 error = RXS_CheckResponse(conn->securityObject, conn, np);
4733 /* If the response is invalid, reset the connection, sending
4734 * an abort to the peer */
4738 rxi_ConnectionError(conn, error);
4739 MUTEX_ENTER(&conn->conn_data_lock);
4740 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4741 MUTEX_EXIT(&conn->conn_data_lock);
4744 /* If the response is valid, any calls waiting to attach
4745 * servers can now do so */
4748 for (i = 0; i < RX_MAXCALLS; i++) {
4749 struct rx_call *call = conn->call[i];
4751 MUTEX_ENTER(&call->lock);
4752 if (call->state == RX_STATE_PRECALL)
4753 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4754 /* tnop can be null if newcallp is null */
4755 MUTEX_EXIT(&call->lock);
4759 /* Update the peer reachability information, just in case
4760 * some calls went into attach-wait while we were waiting
4761 * for authentication..
4763 rxi_UpdatePeerReach(conn, NULL);
4768 /* A client has received an authentication challenge: the security
4769 * object is asked to cough up a respectable response packet to send
4770 * back to the server. The server is responsible for retrying the
4771 * challenge if it fails to get a response. */
4774 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4775 struct rx_packet *np, int istack)
4779 /* Ignore the challenge if we're the server */
4780 if (conn->type == RX_SERVER_CONNECTION)
4783 /* Ignore the challenge if the connection is otherwise idle; someone's
4784 * trying to use us as an oracle. */
4785 if (!rxi_HasActiveCalls(conn))
4788 /* Send the security object the challenge packet. It is expected to fill
4789 * in the response. */
4790 error = RXS_GetResponse(conn->securityObject, conn, np);
4792 /* If the security object is unable to return a valid response, reset the
4793 * connection and send an abort to the peer. Otherwise send the response
4794 * packet to the peer connection. */
4796 rxi_ConnectionError(conn, error);
4797 MUTEX_ENTER(&conn->conn_data_lock);
4798 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4799 MUTEX_EXIT(&conn->conn_data_lock);
4801 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4802 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4808 /* Find an available server process to service the current request in
4809 * the given call structure. If one isn't available, queue up this
4810 * call so it eventually gets one */
4812 rxi_AttachServerProc(struct rx_call *call,
4813 osi_socket socket, int *tnop,
4814 struct rx_call **newcallp)
4816 struct rx_serverQueueEntry *sq;
4817 struct rx_service *service = call->conn->service;
4820 /* May already be attached */
4821 if (call->state == RX_STATE_ACTIVE)
4824 MUTEX_ENTER(&rx_serverPool_lock);
4826 haveQuota = QuotaOK(service);
4827 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4828 /* If there are no processes available to service this call,
4829 * put the call on the incoming call queue (unless it's
4830 * already on the queue).
4832 #ifdef RX_ENABLE_LOCKS
4834 ReturnToServerPool(service);
4835 #endif /* RX_ENABLE_LOCKS */
4837 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4838 call->flags |= RX_CALL_WAIT_PROC;
4839 rx_atomic_inc(&rx_nWaiting);
4840 rx_atomic_inc(&rx_nWaited);
4841 rxi_calltrace(RX_CALL_ARRIVAL, call);
4842 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4843 queue_Append(&rx_incomingCallQueue, call);
4846 sq = queue_Last(&rx_idleServerQueue, rx_serverQueueEntry);
4848 /* If hot threads are enabled, and both newcallp and sq->socketp
4849 * are non-null, then this thread will process the call, and the
4850 * idle server thread will start listening on this threads socket.
4853 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4856 *sq->socketp = socket;
4857 clock_GetTime(&call->startTime);
4858 MUTEX_ENTER(&rx_refcnt_mutex);
4859 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4860 MUTEX_EXIT(&rx_refcnt_mutex);
4864 if (call->flags & RX_CALL_WAIT_PROC) {
4865 /* Conservative: I don't think this should happen */
4866 call->flags &= ~RX_CALL_WAIT_PROC;
4867 if (queue_IsOnQueue(call)) {
4870 rx_atomic_dec(&rx_nWaiting);
4873 call->state = RX_STATE_ACTIVE;
4874 call->mode = RX_MODE_RECEIVING;
4875 #ifdef RX_KERNEL_TRACE
4877 int glockOwner = ISAFS_GLOCK();
4880 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4881 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4887 if (call->flags & RX_CALL_CLEARED) {
4888 /* send an ack now to start the packet flow up again */
4889 call->flags &= ~RX_CALL_CLEARED;
4890 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4892 #ifdef RX_ENABLE_LOCKS
4895 service->nRequestsRunning++;
4896 MUTEX_ENTER(&rx_quota_mutex);
4897 if (service->nRequestsRunning <= service->minProcs)
4900 MUTEX_EXIT(&rx_quota_mutex);
4904 MUTEX_EXIT(&rx_serverPool_lock);
4907 /* Delay the sending of an acknowledge event for a short while, while
4908 * a new call is being prepared (in the case of a client) or a reply
4909 * is being prepared (in the case of a server). Rather than sending
4910 * an ack packet, an ACKALL packet is sent. */
4912 rxi_AckAll(struct rxevent *event, struct rx_call *call, char *dummy)
4914 #ifdef RX_ENABLE_LOCKS
4916 MUTEX_ENTER(&call->lock);
4917 call->delayedAckEvent = NULL;
4918 MUTEX_ENTER(&rx_refcnt_mutex);
4919 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4920 MUTEX_EXIT(&rx_refcnt_mutex);
4922 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4923 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4924 call->flags |= RX_CALL_ACKALL_SENT;
4926 MUTEX_EXIT(&call->lock);
4927 #else /* RX_ENABLE_LOCKS */
4929 call->delayedAckEvent = NULL;
4930 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4931 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4932 call->flags |= RX_CALL_ACKALL_SENT;
4933 #endif /* RX_ENABLE_LOCKS */
4937 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4939 struct rx_call *call = arg1;
4940 #ifdef RX_ENABLE_LOCKS
4942 MUTEX_ENTER(&call->lock);
4943 if (event == call->delayedAckEvent)
4944 call->delayedAckEvent = NULL;
4945 MUTEX_ENTER(&rx_refcnt_mutex);
4946 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4947 MUTEX_EXIT(&rx_refcnt_mutex);
4949 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4951 MUTEX_EXIT(&call->lock);
4952 #else /* RX_ENABLE_LOCKS */
4954 call->delayedAckEvent = NULL;
4955 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4956 #endif /* RX_ENABLE_LOCKS */
4960 #ifdef RX_ENABLE_LOCKS
4961 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4962 * clearing them out.
4965 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4967 struct rx_packet *p, *tp;
4970 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4971 p->flags |= RX_PKTFLAG_ACKED;
4975 call->flags |= RX_CALL_TQ_CLEARME;
4976 call->flags |= RX_CALL_TQ_SOME_ACKED;
4979 rxi_rto_cancel(call);
4981 call->tfirst = call->tnext;
4982 call->nSoftAcked = 0;
4984 if (call->flags & RX_CALL_FAST_RECOVER) {
4985 call->flags &= ~RX_CALL_FAST_RECOVER;
4986 call->cwind = call->nextCwind;
4987 call->nextCwind = 0;
4990 CV_SIGNAL(&call->cv_twind);
4992 #endif /* RX_ENABLE_LOCKS */
4994 /* Clear out the transmit queue for the current call (all packets have
4995 * been received by peer) */
4997 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4999 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5000 struct rx_packet *p, *tp;
5002 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
5004 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
5005 p->flags |= RX_PKTFLAG_ACKED;
5009 call->flags |= RX_CALL_TQ_CLEARME;
5010 call->flags |= RX_CALL_TQ_SOME_ACKED;
5013 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5014 #ifdef RXDEBUG_PACKET
5016 #endif /* RXDEBUG_PACKET */
5017 rxi_FreePackets(0, &call->tq);
5018 rxi_WakeUpTransmitQueue(call);
5019 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5020 call->flags &= ~RX_CALL_TQ_CLEARME;
5022 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5024 rxi_rto_cancel(call);
5025 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5026 call->nSoftAcked = 0;
5028 if (call->flags & RX_CALL_FAST_RECOVER) {
5029 call->flags &= ~RX_CALL_FAST_RECOVER;
5030 call->cwind = call->nextCwind;
5032 #ifdef RX_ENABLE_LOCKS
5033 CV_SIGNAL(&call->cv_twind);
5035 osi_rxWakeup(&call->twind);
5040 rxi_ClearReceiveQueue(struct rx_call *call)
5042 if (queue_IsNotEmpty(&call->rq)) {
5045 count = rxi_FreePackets(0, &call->rq);
5046 rx_packetReclaims += count;
5047 #ifdef RXDEBUG_PACKET
5049 if ( call->rqc != 0 )
5050 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5052 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5054 if (call->state == RX_STATE_PRECALL) {
5055 call->flags |= RX_CALL_CLEARED;
5059 /* Send an abort packet for the specified call */
5061 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5062 int istack, int force)
5065 struct clock when, now;
5070 /* Clients should never delay abort messages */
5071 if (rx_IsClientConn(call->conn))
5074 if (call->abortCode != call->error) {
5075 call->abortCode = call->error;
5076 call->abortCount = 0;
5079 if (force || rxi_callAbortThreshhold == 0
5080 || call->abortCount < rxi_callAbortThreshhold) {
5081 if (call->delayedAbortEvent) {
5082 rxevent_Cancel(call->delayedAbortEvent, call,
5083 RX_CALL_REFCOUNT_ABORT);
5085 error = htonl(call->error);
5088 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5089 (char *)&error, sizeof(error), istack);
5090 } else if (!call->delayedAbortEvent) {
5091 clock_GetTime(&now);
5093 clock_Addmsec(&when, rxi_callAbortDelay);
5094 MUTEX_ENTER(&rx_refcnt_mutex);
5095 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5096 MUTEX_EXIT(&rx_refcnt_mutex);
5097 call->delayedAbortEvent =
5098 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
5103 /* Send an abort packet for the specified connection. Packet is an
5104 * optional pointer to a packet that can be used to send the abort.
5105 * Once the number of abort messages reaches the threshhold, an
5106 * event is scheduled to send the abort. Setting the force flag
5107 * overrides sending delayed abort messages.
5109 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5110 * to send the abort packet.
5113 rxi_SendConnectionAbort(struct rx_connection *conn,
5114 struct rx_packet *packet, int istack, int force)
5117 struct clock when, now;
5122 /* Clients should never delay abort messages */
5123 if (rx_IsClientConn(conn))
5126 if (force || rxi_connAbortThreshhold == 0
5127 || conn->abortCount < rxi_connAbortThreshhold) {
5128 if (conn->delayedAbortEvent) {
5129 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
5131 error = htonl(conn->error);
5133 MUTEX_EXIT(&conn->conn_data_lock);
5135 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5136 RX_PACKET_TYPE_ABORT, (char *)&error,
5137 sizeof(error), istack);
5138 MUTEX_ENTER(&conn->conn_data_lock);
5139 } else if (!conn->delayedAbortEvent) {
5140 clock_GetTime(&now);
5142 clock_Addmsec(&when, rxi_connAbortDelay);
5143 conn->delayedAbortEvent =
5144 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
5149 /* Associate an error all of the calls owned by a connection. Called
5150 * with error non-zero. This is only for really fatal things, like
5151 * bad authentication responses. The connection itself is set in
5152 * error at this point, so that future packets received will be
5155 rxi_ConnectionError(struct rx_connection *conn,
5161 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5163 MUTEX_ENTER(&conn->conn_data_lock);
5164 if (conn->challengeEvent)
5165 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
5166 if (conn->natKeepAliveEvent)
5167 rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);
5168 if (conn->checkReachEvent) {
5169 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
5170 conn->checkReachEvent = 0;
5171 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5172 MUTEX_ENTER(&rx_refcnt_mutex);
5174 MUTEX_EXIT(&rx_refcnt_mutex);
5176 MUTEX_EXIT(&conn->conn_data_lock);
5177 for (i = 0; i < RX_MAXCALLS; i++) {
5178 struct rx_call *call = conn->call[i];
5180 MUTEX_ENTER(&call->lock);
5181 rxi_CallError(call, error);
5182 MUTEX_EXIT(&call->lock);
5185 conn->error = error;
5186 if (rx_stats_active)
5187 rx_atomic_inc(&rx_stats.fatalErrors);
5192 * Interrupt an in-progress call with the specified error and wakeup waiters.
5194 * @param[in] call The call to interrupt
5195 * @param[in] error The error code to send to the peer
5198 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5200 MUTEX_ENTER(&call->lock);
5201 rxi_CallError(call, error);
5202 rxi_SendCallAbort(call, NULL, 0, 1);
5203 MUTEX_EXIT(&call->lock);
5207 rxi_CallError(struct rx_call *call, afs_int32 error)
5210 osirx_AssertMine(&call->lock, "rxi_CallError");
5212 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5214 error = call->error;
5216 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5217 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5218 rxi_ResetCall(call, 0);
5221 rxi_ResetCall(call, 0);
5223 call->error = error;
5226 /* Reset various fields in a call structure, and wakeup waiting
5227 * processes. Some fields aren't changed: state & mode are not
5228 * touched (these must be set by the caller), and bufptr, nLeft, and
5229 * nFree are not reset, since these fields are manipulated by
5230 * unprotected macros, and may only be reset by non-interrupting code.
5233 /* this code requires that call->conn be set properly as a pre-condition. */
5234 #endif /* ADAPT_WINDOW */
5237 rxi_ResetCall(struct rx_call *call, int newcall)
5240 struct rx_peer *peer;
5241 struct rx_packet *packet;
5243 osirx_AssertMine(&call->lock, "rxi_ResetCall");
5245 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5247 /* Notify anyone who is waiting for asynchronous packet arrival */
5248 if (call->arrivalProc) {
5249 (*call->arrivalProc) (call, call->arrivalProcHandle,
5250 call->arrivalProcArg);
5251 call->arrivalProc = (void (*)())0;
5254 if (call->growMTUEvent)
5255 rxevent_Cancel(call->growMTUEvent, call,
5256 RX_CALL_REFCOUNT_ALIVE);
5258 if (call->delayedAbortEvent) {
5259 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
5260 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5262 rxi_SendCallAbort(call, packet, 0, 1);
5263 rxi_FreePacket(packet);
5268 * Update the peer with the congestion information in this call
5269 * so other calls on this connection can pick up where this call
5270 * left off. If the congestion sequence numbers don't match then
5271 * another call experienced a retransmission.
5273 peer = call->conn->peer;
5274 MUTEX_ENTER(&peer->peer_lock);
5276 if (call->congestSeq == peer->congestSeq) {
5277 peer->cwind = MAX(peer->cwind, call->cwind);
5278 peer->MTU = MAX(peer->MTU, call->MTU);
5279 peer->nDgramPackets =
5280 MAX(peer->nDgramPackets, call->nDgramPackets);
5283 call->abortCode = 0;
5284 call->abortCount = 0;
5286 if (peer->maxDgramPackets > 1) {
5287 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5289 call->MTU = peer->MTU;
5291 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5292 call->ssthresh = rx_maxSendWindow;
5293 call->nDgramPackets = peer->nDgramPackets;
5294 call->congestSeq = peer->congestSeq;
5295 call->rtt = peer->rtt;
5296 call->rtt_dev = peer->rtt_dev;
5297 clock_Zero(&call->rto);
5298 clock_Addmsec(&call->rto,
5299 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5300 MUTEX_EXIT(&peer->peer_lock);
5302 flags = call->flags;
5303 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5304 rxi_WaitforTQBusy(call);
5305 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5307 rxi_ClearTransmitQueue(call, 1);
5308 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5309 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5313 if ((flags & RX_CALL_PEER_BUSY)) {
5314 /* The call channel is still busy; resetting the call doesn't change
5316 call->flags |= RX_CALL_PEER_BUSY;
5319 rxi_ClearReceiveQueue(call);
5320 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5324 call->twind = call->conn->twind[call->channel];
5325 call->rwind = call->conn->rwind[call->channel];
5326 call->nSoftAcked = 0;
5327 call->nextCwind = 0;
5330 call->nCwindAcks = 0;
5331 call->nSoftAcks = 0;
5332 call->nHardAcks = 0;
5334 call->tfirst = call->rnext = call->tnext = 1;
5337 call->lastAcked = 0;
5338 call->localStatus = call->remoteStatus = 0;
5340 if (flags & RX_CALL_READER_WAIT) {
5341 #ifdef RX_ENABLE_LOCKS
5342 CV_BROADCAST(&call->cv_rq);
5344 osi_rxWakeup(&call->rq);
5347 if (flags & RX_CALL_WAIT_PACKETS) {
5348 MUTEX_ENTER(&rx_freePktQ_lock);
5349 rxi_PacketsUnWait(); /* XXX */
5350 MUTEX_EXIT(&rx_freePktQ_lock);
5352 #ifdef RX_ENABLE_LOCKS
5353 CV_SIGNAL(&call->cv_twind);
5355 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5356 osi_rxWakeup(&call->twind);
5359 #ifdef RX_ENABLE_LOCKS
5360 /* The following ensures that we don't mess with any queue while some
5361 * other thread might also be doing so. The call_queue_lock field is
5362 * is only modified under the call lock. If the call is in the process
5363 * of being removed from a queue, the call is not locked until the
5364 * the queue lock is dropped and only then is the call_queue_lock field
5365 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5366 * Note that any other routine which removes a call from a queue has to
5367 * obtain the queue lock before examing the queue and removing the call.
5369 if (call->call_queue_lock) {
5370 MUTEX_ENTER(call->call_queue_lock);
5371 if (queue_IsOnQueue(call)) {
5373 if (flags & RX_CALL_WAIT_PROC) {
5374 rx_atomic_dec(&rx_nWaiting);
5377 MUTEX_EXIT(call->call_queue_lock);
5378 CLEAR_CALL_QUEUE_LOCK(call);
5380 #else /* RX_ENABLE_LOCKS */
5381 if (queue_IsOnQueue(call)) {
5383 if (flags & RX_CALL_WAIT_PROC)
5384 rx_atomic_dec(&rx_nWaiting);
5386 #endif /* RX_ENABLE_LOCKS */
5388 rxi_KeepAliveOff(call);
5389 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5392 /* Send an acknowledge for the indicated packet (seq,serial) of the
5393 * indicated call, for the indicated reason (reason). This
5394 * acknowledge will specifically acknowledge receiving the packet, and
5395 * will also specify which other packets for this call have been
5396 * received. This routine returns the packet that was used to the
5397 * caller. The caller is responsible for freeing it or re-using it.
5398 * This acknowledgement also returns the highest sequence number
5399 * actually read out by the higher level to the sender; the sender
5400 * promises to keep around packets that have not been read by the
5401 * higher level yet (unless, of course, the sender decides to abort
5402 * the call altogether). Any of p, seq, serial, pflags, or reason may
5403 * be set to zero without ill effect. That is, if they are zero, they
5404 * will not convey any information.
5405 * NOW there is a trailer field, after the ack where it will safely be
5406 * ignored by mundanes, which indicates the maximum size packet this
5407 * host can swallow. */
5409 struct rx_packet *optionalPacket; use to send ack (or null)
5410 int seq; Sequence number of the packet we are acking
5411 int serial; Serial number of the packet
5412 int pflags; Flags field from packet header
5413 int reason; Reason an acknowledge was prompted
5417 rxi_SendAck(struct rx_call *call,
5418 struct rx_packet *optionalPacket, int serial, int reason,
5421 struct rx_ackPacket *ap;
5422 struct rx_packet *rqp;
5423 struct rx_packet *nxp; /* For queue_Scan */
5424 struct rx_packet *p;
5427 afs_uint32 padbytes = 0;
5428 #ifdef RX_ENABLE_TSFPQ
5429 struct rx_ts_info_t * rx_ts_info;
5433 * Open the receive window once a thread starts reading packets
5435 if (call->rnext > 1) {
5436 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5439 /* Don't attempt to grow MTU if this is a critical ping */
5440 if (reason == RX_ACK_MTU) {
5441 /* keep track of per-call attempts, if we're over max, do in small
5442 * otherwise in larger? set a size to increment by, decrease
5445 if (call->conn->peer->maxPacketSize &&
5446 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5448 padbytes = call->conn->peer->maxPacketSize+16;
5450 padbytes = call->conn->peer->maxMTU + 128;
5452 /* do always try a minimum size ping */
5453 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5455 /* subtract the ack payload */
5456 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5457 reason = RX_ACK_PING;
5460 call->nHardAcks = 0;
5461 call->nSoftAcks = 0;
5462 if (call->rnext > call->lastAcked)
5463 call->lastAcked = call->rnext;
5467 rx_computelen(p, p->length); /* reset length, you never know */
5468 } /* where that's been... */
5469 #ifdef RX_ENABLE_TSFPQ
5471 RX_TS_INFO_GET(rx_ts_info);
5472 if ((p = rx_ts_info->local_special_packet)) {
5473 rx_computelen(p, p->length);
5474 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5475 rx_ts_info->local_special_packet = p;
5476 } else { /* We won't send the ack, but don't panic. */
5477 return optionalPacket;
5481 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5482 /* We won't send the ack, but don't panic. */
5483 return optionalPacket;
5488 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5491 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5492 #ifndef RX_ENABLE_TSFPQ
5493 if (!optionalPacket)
5496 return optionalPacket;
5498 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5499 if (rx_Contiguous(p) < templ) {
5500 #ifndef RX_ENABLE_TSFPQ
5501 if (!optionalPacket)
5504 return optionalPacket;
5509 /* MTUXXX failing to send an ack is very serious. We should */
5510 /* try as hard as possible to send even a partial ack; it's */
5511 /* better than nothing. */
5512 ap = (struct rx_ackPacket *)rx_DataOf(p);
5513 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5514 ap->reason = reason;
5516 /* The skew computation used to be bogus, I think it's better now. */
5517 /* We should start paying attention to skew. XXX */
5518 ap->serial = htonl(serial);
5519 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5522 * First packet not yet forwarded to reader. When ACKALL has been
5523 * sent the peer has been told that all received packets will be
5524 * delivered to the reader. The value 'rnext' is used internally
5525 * to refer to the next packet in the receive queue that must be
5526 * delivered to the reader. From the perspective of the peer it
5527 * already has so report the last sequence number plus one if there
5528 * are packets in the receive queue awaiting processing.
5530 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5531 !queue_IsEmpty(&call->rq)) {
5532 ap->firstPacket = htonl(queue_Last(&call->rq, rx_packet)->header.seq + 1);
5534 ap->firstPacket = htonl(call->rnext);
5536 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5538 /* No fear of running out of ack packet here because there can only be at most
5539 * one window full of unacknowledged packets. The window size must be constrained
5540 * to be less than the maximum ack size, of course. Also, an ack should always
5541 * fit into a single packet -- it should not ever be fragmented. */
5542 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5543 if (!rqp || !call->rq.next
5544 || (rqp->header.seq > (call->rnext + call->rwind))) {
5545 #ifndef RX_ENABLE_TSFPQ
5546 if (!optionalPacket)
5549 rxi_CallError(call, RX_CALL_DEAD);
5550 return optionalPacket;
5553 while (rqp->header.seq > call->rnext + offset)
5554 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5555 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5557 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5558 #ifndef RX_ENABLE_TSFPQ
5559 if (!optionalPacket)
5562 rxi_CallError(call, RX_CALL_DEAD);
5563 return optionalPacket;
5568 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5570 /* these are new for AFS 3.3 */
5571 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5572 templ = htonl(templ);
5573 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5574 templ = htonl(call->conn->peer->ifMTU);
5575 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5576 sizeof(afs_int32), &templ);
5578 /* new for AFS 3.4 */
5579 templ = htonl(call->rwind);
5580 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5581 sizeof(afs_int32), &templ);
5583 /* new for AFS 3.5 */
5584 templ = htonl(call->conn->peer->ifDgramPackets);
5585 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5586 sizeof(afs_int32), &templ);
5588 p->header.serviceId = call->conn->serviceId;
5589 p->header.cid = (call->conn->cid | call->channel);
5590 p->header.callNumber = *call->callNumber;
5592 p->header.securityIndex = call->conn->securityIndex;
5593 p->header.epoch = call->conn->epoch;
5594 p->header.type = RX_PACKET_TYPE_ACK;
5595 p->header.flags = RX_SLOW_START_OK;
5596 if (reason == RX_ACK_PING) {
5597 p->header.flags |= RX_REQUEST_ACK;
5599 clock_GetTime(&call->pingRequestTime);
5602 p->length = padbytes +
5603 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5606 /* not fast but we can potentially use this if truncated
5607 * fragments are delivered to figure out the mtu.
5609 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5610 sizeof(afs_int32), sizeof(afs_int32),
5614 if (call->conn->type == RX_CLIENT_CONNECTION)
5615 p->header.flags |= RX_CLIENT_INITIATED;
5619 if (rxdebug_active) {
5623 len = _snprintf(msg, sizeof(msg),
5624 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5625 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5626 ntohl(ap->serial), ntohl(ap->previousPacket),
5627 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5628 ap->nAcks, ntohs(ap->bufferSpace) );
5632 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5633 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5637 OutputDebugString(msg);
5639 #else /* AFS_NT40_ENV */
5641 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5642 ap->reason, ntohl(ap->previousPacket),
5643 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5645 for (offset = 0; offset < ap->nAcks; offset++)
5646 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5651 #endif /* AFS_NT40_ENV */
5654 int i, nbytes = p->length;
5656 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5657 if (nbytes <= p->wirevec[i].iov_len) {
5660 savelen = p->wirevec[i].iov_len;
5662 p->wirevec[i].iov_len = nbytes;
5664 rxi_Send(call, p, istack);
5665 p->wirevec[i].iov_len = savelen;
5669 nbytes -= p->wirevec[i].iov_len;
5672 if (rx_stats_active)
5673 rx_atomic_inc(&rx_stats.ackPacketsSent);
5674 #ifndef RX_ENABLE_TSFPQ
5675 if (!optionalPacket)
5678 return optionalPacket; /* Return packet for re-use by caller */
5682 struct rx_packet **list;
5687 /* Send all of the packets in the list in single datagram */
5689 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5690 int istack, int moreFlag)
5696 struct rx_connection *conn = call->conn;
5697 struct rx_peer *peer = conn->peer;
5699 MUTEX_ENTER(&peer->peer_lock);
5700 peer->nSent += xmit->len;
5701 if (xmit->resending)
5702 peer->reSends += xmit->len;
5703 MUTEX_EXIT(&peer->peer_lock);
5705 if (rx_stats_active) {
5706 if (xmit->resending)
5707 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5709 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5712 clock_GetTime(&now);
5714 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5718 /* Set the packet flags and schedule the resend events */
5719 /* Only request an ack for the last packet in the list */
5720 for (i = 0; i < xmit->len; i++) {
5721 struct rx_packet *packet = xmit->list[i];
5723 /* Record the time sent */
5724 packet->timeSent = now;
5725 packet->flags |= RX_PKTFLAG_SENT;
5727 /* Ask for an ack on retransmitted packets, on every other packet
5728 * if the peer doesn't support slow start. Ask for an ack on every
5729 * packet until the congestion window reaches the ack rate. */
5730 if (packet->header.serial) {
5733 packet->firstSent = now;
5734 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5735 || (!(call->flags & RX_CALL_SLOW_START_OK)
5736 && (packet->header.seq & 1)))) {
5741 /* Tag this packet as not being the last in this group,
5742 * for the receiver's benefit */
5743 if (i < xmit->len - 1 || moreFlag) {
5744 packet->header.flags |= RX_MORE_PACKETS;
5749 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5752 /* Since we're about to send a data packet to the peer, it's
5753 * safe to nuke any scheduled end-of-packets ack */
5754 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5756 MUTEX_EXIT(&call->lock);
5757 MUTEX_ENTER(&rx_refcnt_mutex);
5758 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5759 MUTEX_EXIT(&rx_refcnt_mutex);
5760 if (xmit->len > 1) {
5761 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5763 rxi_SendPacket(call, conn, xmit->list[0], istack);
5765 MUTEX_ENTER(&call->lock);
5766 MUTEX_ENTER(&rx_refcnt_mutex);
5767 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5768 MUTEX_EXIT(&rx_refcnt_mutex);
5770 /* Tell the RTO calculation engine that we have sent a packet, and
5771 * if it was the last one */
5772 rxi_rto_packet_sent(call, lastPacket, istack);
5774 /* Update last send time for this call (for keep-alive
5775 * processing), and for the connection (so that we can discover
5776 * idle connections) */
5777 conn->lastSendTime = call->lastSendTime = clock_Sec();
5778 /* Let a set of retransmits trigger an idle timeout */
5779 if (!xmit->resending)
5780 call->lastSendData = call->lastSendTime;
5783 /* When sending packets we need to follow these rules:
5784 * 1. Never send more than maxDgramPackets in a jumbogram.
5785 * 2. Never send a packet with more than two iovecs in a jumbogram.
5786 * 3. Never send a retransmitted packet in a jumbogram.
5787 * 4. Never send more than cwind/4 packets in a jumbogram
5788 * We always keep the last list we should have sent so we
5789 * can set the RX_MORE_PACKETS flags correctly.
5793 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5798 struct xmitlist working;
5799 struct xmitlist last;
5801 struct rx_peer *peer = call->conn->peer;
5802 int morePackets = 0;
5804 memset(&last, 0, sizeof(struct xmitlist));
5805 working.list = &list[0];
5807 working.resending = 0;
5809 recovery = call->flags & RX_CALL_FAST_RECOVER;
5811 for (i = 0; i < len; i++) {
5812 /* Does the current packet force us to flush the current list? */
5814 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5815 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5817 /* This sends the 'last' list and then rolls the current working
5818 * set into the 'last' one, and resets the working set */
5821 rxi_SendList(call, &last, istack, 1);
5822 /* If the call enters an error state stop sending, or if
5823 * we entered congestion recovery mode, stop sending */
5825 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5830 working.resending = 0;
5831 working.list = &list[i];
5833 /* Add the current packet to the list if it hasn't been acked.
5834 * Otherwise adjust the list pointer to skip the current packet. */
5835 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5838 if (list[i]->header.serial)
5839 working.resending = 1;
5841 /* Do we need to flush the list? */
5842 if (working.len >= (int)peer->maxDgramPackets
5843 || working.len >= (int)call->nDgramPackets
5844 || working.len >= (int)call->cwind
5845 || list[i]->header.serial
5846 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5848 rxi_SendList(call, &last, istack, 1);
5849 /* If the call enters an error state stop sending, or if
5850 * we entered congestion recovery mode, stop sending */
5852 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5857 working.resending = 0;
5858 working.list = &list[i + 1];
5861 if (working.len != 0) {
5862 osi_Panic("rxi_SendList error");
5864 working.list = &list[i + 1];
5868 /* Send the whole list when the call is in receive mode, when
5869 * the call is in eof mode, when we are in fast recovery mode,
5870 * and when we have the last packet */
5871 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5872 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5873 || (call->flags & RX_CALL_FAST_RECOVER)) {
5874 /* Check for the case where the current list contains
5875 * an acked packet. Since we always send retransmissions
5876 * in a separate packet, we only need to check the first
5877 * packet in the list */
5878 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5882 rxi_SendList(call, &last, istack, morePackets);
5883 /* If the call enters an error state stop sending, or if
5884 * we entered congestion recovery mode, stop sending */
5886 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5890 rxi_SendList(call, &working, istack, 0);
5892 } else if (last.len > 0) {
5893 rxi_SendList(call, &last, istack, 0);
5894 /* Packets which are in 'working' are not sent by this call */
5899 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
5901 struct rx_call *call = arg0;
5902 struct rx_peer *peer;
5903 struct rx_packet *p, *nxp;
5904 struct clock maxTimeout = { 60, 0 };
5906 MUTEX_ENTER(&call->lock);
5908 peer = call->conn->peer;
5910 /* Make sure that the event pointer is removed from the call
5911 * structure, since there is no longer a per-call retransmission
5913 if (event == call->resendEvent) {
5914 MUTEX_ENTER(&rx_refcnt_mutex);
5915 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5916 MUTEX_EXIT(&rx_refcnt_mutex);
5917 call->resendEvent = NULL;
5920 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
5921 rxi_CheckBusy(call);
5924 if (queue_IsEmpty(&call->tq)) {
5925 /* Nothing to do. This means that we've been raced, and that an
5926 * ACK has come in between when we were triggered, and when we
5927 * actually got to run. */
5931 /* We're in loss recovery */
5932 call->flags |= RX_CALL_FAST_RECOVER;
5934 /* Mark all of the pending packets in the queue as being lost */
5935 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5936 if (!(p->flags & RX_PKTFLAG_ACKED))
5937 p->flags &= ~RX_PKTFLAG_SENT;
5940 /* We're resending, so we double the timeout of the call. This will be
5941 * dropped back down by the first successful ACK that we receive.
5943 * We apply a maximum value here of 60 seconds
5945 clock_Add(&call->rto, &call->rto);
5946 if (clock_Gt(&call->rto, &maxTimeout))
5947 call->rto = maxTimeout;
5949 /* Packet loss is most likely due to congestion, so drop our window size
5950 * and start again from the beginning */
5951 if (peer->maxDgramPackets >1) {
5952 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5953 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5955 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5956 call->nDgramPackets = 1;
5958 call->nextCwind = 1;
5961 MUTEX_ENTER(&peer->peer_lock);
5962 peer->MTU = call->MTU;
5963 peer->cwind = call->cwind;
5964 peer->nDgramPackets = 1;
5966 call->congestSeq = peer->congestSeq;
5967 MUTEX_EXIT(&peer->peer_lock);
5969 rxi_Start(call, istack);
5972 MUTEX_EXIT(&call->lock);
5975 /* This routine is called when new packets are readied for
5976 * transmission and when retransmission may be necessary, or when the
5977 * transmission window or burst count are favourable. This should be
5978 * better optimized for new packets, the usual case, now that we've
5979 * got rid of queues of send packets. XXXXXXXXXXX */
5981 rxi_Start(struct rx_call *call, int istack)
5984 struct rx_packet *p;
5985 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5990 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5991 if (rx_stats_active)
5992 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5997 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5999 /* Send (or resend) any packets that need it, subject to
6000 * window restrictions and congestion burst control
6001 * restrictions. Ask for an ack on the last packet sent in
6002 * this burst. For now, we're relying upon the window being
6003 * considerably bigger than the largest number of packets that
6004 * are typically sent at once by one initial call to
6005 * rxi_Start. This is probably bogus (perhaps we should ask
6006 * for an ack when we're half way through the current
6007 * window?). Also, for non file transfer applications, this
6008 * may end up asking for an ack for every packet. Bogus. XXXX
6011 * But check whether we're here recursively, and let the other guy
6014 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6015 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6016 call->flags |= RX_CALL_TQ_BUSY;
6018 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6020 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6021 call->flags &= ~RX_CALL_NEED_START;
6022 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6024 maxXmitPackets = MIN(call->twind, call->cwind);
6025 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
6026 #ifdef RX_TRACK_PACKETS
6027 if ((p->flags & RX_PKTFLAG_FREE)
6028 || (!queue_IsEnd(&call->tq, nxp)
6029 && (nxp->flags & RX_PKTFLAG_FREE))
6030 || (p == (struct rx_packet *)&rx_freePacketQueue)
6031 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
6032 osi_Panic("rxi_Start: xmit queue clobbered");
6035 if (p->flags & RX_PKTFLAG_ACKED) {
6036 /* Since we may block, don't trust this */
6037 if (rx_stats_active)
6038 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6039 continue; /* Ignore this packet if it has been acknowledged */
6042 /* Turn off all flags except these ones, which are the same
6043 * on each transmission */
6044 p->header.flags &= RX_PRESET_FLAGS;
6046 if (p->header.seq >=
6047 call->tfirst + MIN((int)call->twind,
6048 (int)(call->nSoftAcked +
6050 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6051 /* Note: if we're waiting for more window space, we can
6052 * still send retransmits; hence we don't return here, but
6053 * break out to schedule a retransmit event */
6054 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6055 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6060 /* Transmit the packet if it needs to be sent. */
6061 if (!(p->flags & RX_PKTFLAG_SENT)) {
6062 if (nXmitPackets == maxXmitPackets) {
6063 rxi_SendXmitList(call, call->xmitList,
6064 nXmitPackets, istack);
6067 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6068 *(call->callNumber), p));
6069 call->xmitList[nXmitPackets++] = p;
6073 /* xmitList now hold pointers to all of the packets that are
6074 * ready to send. Now we loop to send the packets */
6075 if (nXmitPackets > 0) {
6076 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6080 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6082 /* We went into the error state while sending packets. Now is
6083 * the time to reset the call. This will also inform the using
6084 * process that the call is in an error state.
6086 if (rx_stats_active)
6087 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6088 call->flags &= ~RX_CALL_TQ_BUSY;
6089 rxi_WakeUpTransmitQueue(call);
6090 rxi_CallError(call, call->error);
6093 #ifdef RX_ENABLE_LOCKS
6094 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6096 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6097 /* Some packets have received acks. If they all have, we can clear
6098 * the transmit queue.
6101 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
6102 if (p->header.seq < call->tfirst
6103 && (p->flags & RX_PKTFLAG_ACKED)) {
6105 #ifdef RX_TRACK_PACKETS
6106 p->flags &= ~RX_PKTFLAG_TQ;
6108 #ifdef RXDEBUG_PACKET
6116 call->flags |= RX_CALL_TQ_CLEARME;
6118 #endif /* RX_ENABLE_LOCKS */
6119 if (call->flags & RX_CALL_TQ_CLEARME)
6120 rxi_ClearTransmitQueue(call, 1);
6121 } while (call->flags & RX_CALL_NEED_START);
6123 * TQ references no longer protected by this flag; they must remain
6124 * protected by the global lock.
6126 call->flags &= ~RX_CALL_TQ_BUSY;
6127 rxi_WakeUpTransmitQueue(call);
6129 call->flags |= RX_CALL_NEED_START;
6131 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6133 rxi_rto_cancel(call);
6137 /* Also adjusts the keep alive parameters for the call, to reflect
6138 * that we have just sent a packet (so keep alives aren't sent
6141 rxi_Send(struct rx_call *call, struct rx_packet *p,
6144 struct rx_connection *conn = call->conn;
6146 /* Stamp each packet with the user supplied status */
6147 p->header.userStatus = call->localStatus;
6149 /* Allow the security object controlling this call's security to
6150 * make any last-minute changes to the packet */
6151 RXS_SendPacket(conn->securityObject, call, p);
6153 /* Since we're about to send SOME sort of packet to the peer, it's
6154 * safe to nuke any scheduled end-of-packets ack */
6155 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
6157 /* Actually send the packet, filling in more connection-specific fields */
6158 MUTEX_EXIT(&call->lock);
6159 MUTEX_ENTER(&rx_refcnt_mutex);
6160 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6161 MUTEX_EXIT(&rx_refcnt_mutex);
6162 rxi_SendPacket(call, conn, p, istack);
6163 MUTEX_ENTER(&rx_refcnt_mutex);
6164 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6165 MUTEX_EXIT(&rx_refcnt_mutex);
6166 MUTEX_ENTER(&call->lock);
6168 /* Update last send time for this call (for keep-alive
6169 * processing), and for the connection (so that we can discover
6170 * idle connections) */
6171 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6172 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6173 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6175 conn->lastSendTime = call->lastSendTime = clock_Sec();
6176 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6177 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6178 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6179 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6180 RX_ACK_PING_RESPONSE)))
6181 call->lastSendData = call->lastSendTime;
6185 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6186 * that things are fine. Also called periodically to guarantee that nothing
6187 * falls through the cracks (e.g. (error + dally) connections have keepalive
6188 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6190 * haveCTLock Set if calling from rxi_ReapConnections
6192 #ifdef RX_ENABLE_LOCKS
6194 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6195 #else /* RX_ENABLE_LOCKS */
6197 rxi_CheckCall(struct rx_call *call)
6198 #endif /* RX_ENABLE_LOCKS */
6200 struct rx_connection *conn = call->conn;
6202 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6203 afs_uint32 fudgeFactor;
6207 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6208 if (call->flags & RX_CALL_TQ_BUSY) {
6209 /* Call is active and will be reset by rxi_Start if it's
6210 * in an error state.
6215 /* RTT + 8*MDEV, rounded up to the next second. */
6216 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6217 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6219 deadTime = conn->secondsUntilDead + fudgeFactor;
6221 /* These are computed to the second (+- 1 second). But that's
6222 * good enough for these values, which should be a significant
6223 * number of seconds. */
6224 if (now > (call->lastReceiveTime + deadTime)) {
6225 if (call->state == RX_STATE_ACTIVE) {
6227 #if defined(KERNEL) && defined(AFS_SUN5_ENV)
6229 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6230 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6231 ip_stack_t *ipst = ns->netstack_ip;
6233 ire = ire_cache_lookup(conn->peer->host
6234 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6236 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
6238 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6245 if (ire && ire->ire_max_frag > 0)
6246 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6248 #if defined(GLOBAL_NETSTACKID)
6252 #endif /* ADAPT_PMTU */
6253 cerror = RX_CALL_DEAD;
6256 #ifdef RX_ENABLE_LOCKS
6257 /* Cancel pending events */
6258 rxevent_Cancel(call->delayedAckEvent, call,
6259 RX_CALL_REFCOUNT_DELAY);
6260 rxi_rto_cancel(call);
6261 rxevent_Cancel(call->keepAliveEvent, call,
6262 RX_CALL_REFCOUNT_ALIVE);
6263 if (call->growMTUEvent)
6264 rxevent_Cancel(call->growMTUEvent, call,
6265 RX_CALL_REFCOUNT_ALIVE);
6266 MUTEX_ENTER(&rx_refcnt_mutex);
6267 if (call->refCount == 0) {
6268 rxi_FreeCall(call, haveCTLock);
6269 MUTEX_EXIT(&rx_refcnt_mutex);
6272 MUTEX_EXIT(&rx_refcnt_mutex);
6274 #else /* RX_ENABLE_LOCKS */
6275 rxi_FreeCall(call, 0);
6277 #endif /* RX_ENABLE_LOCKS */
6279 /* Non-active calls are destroyed if they are not responding
6280 * to pings; active calls are simply flagged in error, so the
6281 * attached process can die reasonably gracefully. */
6284 if (conn->idleDeadTime) {
6285 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6288 /* see if we have a non-activity timeout */
6289 if (call->startWait && idleDeadTime
6290 && ((call->startWait + idleDeadTime) < now) &&
6291 (call->flags & RX_CALL_READER_WAIT)) {
6292 if (call->state == RX_STATE_ACTIVE) {
6293 cerror = RX_CALL_TIMEOUT;
6297 if (call->lastSendData && idleDeadTime && (conn->idleDeadErr != 0)
6298 && ((call->lastSendData + idleDeadTime) < now)) {
6299 if (call->state == RX_STATE_ACTIVE) {
6300 cerror = conn->idleDeadErr;
6305 if (conn->hardDeadTime) {
6306 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6309 /* see if we have a hard timeout */
6311 && (now > (hardDeadTime + call->startTime.sec))) {
6312 if (call->state == RX_STATE_ACTIVE)
6313 rxi_CallError(call, RX_CALL_TIMEOUT);
6318 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT
6319 && call->lastReceiveTime) {
6320 int oldMTU = conn->peer->ifMTU;
6322 /* if we thought we could send more, perhaps things got worse */
6323 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6324 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6325 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6326 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6328 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6330 /* minimum capped in SetPeerMtu */
6331 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6334 conn->lastPacketSize = 0;
6336 /* needed so ResetCall doesn't clobber us. */
6337 call->MTU = conn->peer->ifMTU;
6339 /* if we never succeeded, let the error pass out as-is */
6340 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6341 cerror = conn->msgsizeRetryErr;
6344 rxi_CallError(call, cerror);
6349 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6351 struct rx_connection *conn = arg1;
6352 struct rx_header theader;
6353 char tbuffer[1 + sizeof(struct rx_header)];
6354 struct sockaddr_in taddr;
6357 struct iovec tmpiov[2];
6360 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6363 tp = &tbuffer[sizeof(struct rx_header)];
6364 taddr.sin_family = AF_INET;
6365 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6366 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6367 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6368 taddr.sin_len = sizeof(struct sockaddr_in);
6370 memset(&theader, 0, sizeof(theader));
6371 theader.epoch = htonl(999);
6373 theader.callNumber = 0;
6376 theader.type = RX_PACKET_TYPE_VERSION;
6377 theader.flags = RX_LAST_PACKET;
6378 theader.serviceId = 0;
6380 memcpy(tbuffer, &theader, sizeof(theader));
6381 memcpy(tp, &a, sizeof(a));
6382 tmpiov[0].iov_base = tbuffer;
6383 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6385 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6387 MUTEX_ENTER(&conn->conn_data_lock);
6388 MUTEX_ENTER(&rx_refcnt_mutex);
6389 /* Only reschedule ourselves if the connection would not be destroyed */
6390 if (conn->refCount <= 1) {
6391 conn->natKeepAliveEvent = NULL;
6392 MUTEX_EXIT(&rx_refcnt_mutex);
6393 MUTEX_EXIT(&conn->conn_data_lock);
6394 rx_DestroyConnection(conn); /* drop the reference for this */
6396 conn->refCount--; /* drop the reference for this */
6397 MUTEX_EXIT(&rx_refcnt_mutex);
6398 conn->natKeepAliveEvent = NULL;
6399 rxi_ScheduleNatKeepAliveEvent(conn);
6400 MUTEX_EXIT(&conn->conn_data_lock);
6405 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6407 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6408 struct clock when, now;
6409 clock_GetTime(&now);
6411 when.sec += conn->secondsUntilNatPing;
6412 MUTEX_ENTER(&rx_refcnt_mutex);
6413 conn->refCount++; /* hold a reference for this */
6414 MUTEX_EXIT(&rx_refcnt_mutex);
6415 conn->natKeepAliveEvent =
6416 rxevent_PostNow(&when, &now, rxi_NatKeepAliveEvent, conn, 0);
6421 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6423 MUTEX_ENTER(&conn->conn_data_lock);
6424 conn->secondsUntilNatPing = seconds;
6426 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6427 rxi_ScheduleNatKeepAliveEvent(conn);
6429 conn->flags |= RX_CONN_NAT_PING;
6431 MUTEX_EXIT(&conn->conn_data_lock);
6435 rxi_NatKeepAliveOn(struct rx_connection *conn)
6437 MUTEX_ENTER(&conn->conn_data_lock);
6438 /* if it's already attached */
6439 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6440 rxi_ScheduleNatKeepAliveEvent(conn);
6442 conn->flags |= RX_CONN_NAT_PING;
6443 MUTEX_EXIT(&conn->conn_data_lock);
6446 /* When a call is in progress, this routine is called occasionally to
6447 * make sure that some traffic has arrived (or been sent to) the peer.
6448 * If nothing has arrived in a reasonable amount of time, the call is
6449 * declared dead; if nothing has been sent for a while, we send a
6450 * keep-alive packet (if we're actually trying to keep the call alive)
6453 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
6455 struct rx_call *call = arg1;
6456 struct rx_connection *conn;
6459 MUTEX_ENTER(&rx_refcnt_mutex);
6460 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6461 MUTEX_EXIT(&rx_refcnt_mutex);
6462 MUTEX_ENTER(&call->lock);
6463 if (event == call->keepAliveEvent)
6464 call->keepAliveEvent = NULL;
6467 #ifdef RX_ENABLE_LOCKS
6468 if (rxi_CheckCall(call, 0)) {
6469 MUTEX_EXIT(&call->lock);
6472 #else /* RX_ENABLE_LOCKS */
6473 if (rxi_CheckCall(call))
6475 #endif /* RX_ENABLE_LOCKS */
6477 /* Don't try to keep alive dallying calls */
6478 if (call->state == RX_STATE_DALLY) {
6479 MUTEX_EXIT(&call->lock);
6484 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6485 /* Don't try to send keepalives if there is unacknowledged data */
6486 /* the rexmit code should be good enough, this little hack
6487 * doesn't quite work XXX */
6488 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6490 rxi_ScheduleKeepAliveEvent(call);
6491 MUTEX_EXIT(&call->lock);
6494 /* Does what's on the nameplate. */
6496 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy)
6498 struct rx_call *call = arg1;
6499 struct rx_connection *conn;
6501 MUTEX_ENTER(&rx_refcnt_mutex);
6502 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6503 MUTEX_EXIT(&rx_refcnt_mutex);
6504 MUTEX_ENTER(&call->lock);
6506 if (event == call->growMTUEvent)
6507 call->growMTUEvent = NULL;
6509 #ifdef RX_ENABLE_LOCKS
6510 if (rxi_CheckCall(call, 0)) {
6511 MUTEX_EXIT(&call->lock);
6514 #else /* RX_ENABLE_LOCKS */
6515 if (rxi_CheckCall(call))
6517 #endif /* RX_ENABLE_LOCKS */
6519 /* Don't bother with dallying calls */
6520 if (call->state == RX_STATE_DALLY) {
6521 MUTEX_EXIT(&call->lock);
6528 * keep being scheduled, just don't do anything if we're at peak,
6529 * or we're not set up to be properly handled (idle timeout required)
6531 if ((conn->peer->maxPacketSize != 0) &&
6532 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6533 (conn->idleDeadErr))
6534 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6535 rxi_ScheduleGrowMTUEvent(call, 0);
6536 MUTEX_EXIT(&call->lock);
6540 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6542 if (!call->keepAliveEvent) {
6543 struct clock when, now;
6544 clock_GetTime(&now);
6546 when.sec += call->conn->secondsUntilPing;
6547 MUTEX_ENTER(&rx_refcnt_mutex);
6548 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6549 MUTEX_EXIT(&rx_refcnt_mutex);
6550 call->keepAliveEvent =
6551 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
6556 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6558 if (!call->growMTUEvent) {
6559 struct clock when, now;
6561 clock_GetTime(&now);
6564 if (call->conn->secondsUntilPing)
6565 secs = (6*call->conn->secondsUntilPing)-1;
6567 if (call->conn->secondsUntilDead)
6568 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6572 MUTEX_ENTER(&rx_refcnt_mutex);
6573 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6574 MUTEX_EXIT(&rx_refcnt_mutex);
6575 call->growMTUEvent =
6576 rxevent_PostNow(&when, &now, rxi_GrowMTUEvent, call, 0);
6580 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6582 rxi_KeepAliveOn(struct rx_call *call)
6584 /* Pretend last packet received was received now--i.e. if another
6585 * packet isn't received within the keep alive time, then the call
6586 * will die; Initialize last send time to the current time--even
6587 * if a packet hasn't been sent yet. This will guarantee that a
6588 * keep-alive is sent within the ping time */
6589 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6590 rxi_ScheduleKeepAliveEvent(call);
6594 rxi_GrowMTUOn(struct rx_call *call)
6596 struct rx_connection *conn = call->conn;
6597 MUTEX_ENTER(&conn->conn_data_lock);
6598 conn->lastPingSizeSer = conn->lastPingSize = 0;
6599 MUTEX_EXIT(&conn->conn_data_lock);
6600 rxi_ScheduleGrowMTUEvent(call, 1);
6603 /* This routine is called to send connection abort messages
6604 * that have been delayed to throttle looping clients. */
6606 rxi_SendDelayedConnAbort(struct rxevent *event,
6607 void *arg1, void *unused)
6609 struct rx_connection *conn = arg1;
6612 struct rx_packet *packet;
6614 MUTEX_ENTER(&conn->conn_data_lock);
6615 conn->delayedAbortEvent = NULL;
6616 error = htonl(conn->error);
6618 MUTEX_EXIT(&conn->conn_data_lock);
6619 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6622 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6623 RX_PACKET_TYPE_ABORT, (char *)&error,
6625 rxi_FreePacket(packet);
6629 /* This routine is called to send call abort messages
6630 * that have been delayed to throttle looping clients. */
6632 rxi_SendDelayedCallAbort(struct rxevent *event,
6633 void *arg1, void *dummy)
6635 struct rx_call *call = arg1;
6638 struct rx_packet *packet;
6640 MUTEX_ENTER(&call->lock);
6641 call->delayedAbortEvent = NULL;
6642 error = htonl(call->error);
6644 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6647 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6648 (char *)&error, sizeof(error), 0);
6649 rxi_FreePacket(packet);
6651 MUTEX_EXIT(&call->lock);
6652 MUTEX_ENTER(&rx_refcnt_mutex);
6653 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6654 MUTEX_EXIT(&rx_refcnt_mutex);
6657 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6658 * seconds) to ask the client to authenticate itself. The routine
6659 * issues a challenge to the client, which is obtained from the
6660 * security object associated with the connection */
6662 rxi_ChallengeEvent(struct rxevent *event,
6663 void *arg0, void *arg1, int tries)
6665 struct rx_connection *conn = arg0;
6667 conn->challengeEvent = NULL;
6668 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6669 struct rx_packet *packet;
6670 struct clock when, now;
6673 /* We've failed to authenticate for too long.
6674 * Reset any calls waiting for authentication;
6675 * they are all in RX_STATE_PRECALL.
6679 MUTEX_ENTER(&conn->conn_call_lock);
6680 for (i = 0; i < RX_MAXCALLS; i++) {
6681 struct rx_call *call = conn->call[i];
6683 MUTEX_ENTER(&call->lock);
6684 if (call->state == RX_STATE_PRECALL) {
6685 rxi_CallError(call, RX_CALL_DEAD);
6686 rxi_SendCallAbort(call, NULL, 0, 0);
6688 MUTEX_EXIT(&call->lock);
6691 MUTEX_EXIT(&conn->conn_call_lock);
6695 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6697 /* If there's no packet available, do this later. */
6698 RXS_GetChallenge(conn->securityObject, conn, packet);
6699 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6700 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6701 rxi_FreePacket(packet);
6703 clock_GetTime(&now);
6705 when.sec += RX_CHALLENGE_TIMEOUT;
6706 conn->challengeEvent =
6707 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
6712 /* Call this routine to start requesting the client to authenticate
6713 * itself. This will continue until authentication is established,
6714 * the call times out, or an invalid response is returned. The
6715 * security object associated with the connection is asked to create
6716 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6717 * defined earlier. */
6719 rxi_ChallengeOn(struct rx_connection *conn)
6721 if (!conn->challengeEvent) {
6722 RXS_CreateChallenge(conn->securityObject, conn);
6723 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6728 /* rxi_ComputeRoundTripTime is called with peer locked. */
6729 /* peer may be null */
6731 rxi_ComputeRoundTripTime(struct rx_packet *p,
6732 struct rx_ackPacket *ack,
6733 struct rx_call *call,
6734 struct rx_peer *peer,
6737 struct clock thisRtt, *sentp;
6741 /* If the ACK is delayed, then do nothing */
6742 if (ack->reason == RX_ACK_DELAY)
6745 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6746 * their RTT multiple times, so only include the RTT of the last packet
6748 if (p->flags & RX_JUMBO_PACKET)
6751 /* Use the serial number to determine which transmission the ACK is for,
6752 * and set the sent time to match this. If we have no serial number, then
6753 * only use the ACK for RTT calculations if the packet has not been
6757 serial = ntohl(ack->serial);
6759 if (serial == p->header.serial) {
6760 sentp = &p->timeSent;
6761 } else if (serial == p->firstSerial) {
6762 sentp = &p->firstSent;
6763 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6764 sentp = &p->firstSent;
6768 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6769 sentp = &p->firstSent;
6776 if (clock_Lt(&thisRtt, sentp))
6777 return; /* somebody set the clock back, don't count this time. */
6779 clock_Sub(&thisRtt, sentp);
6780 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6781 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6783 if (clock_IsZero(&thisRtt)) {
6785 * The actual round trip time is shorter than the
6786 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6787 * Since we can't tell which at the moment we will assume 1ms.
6789 thisRtt.usec = 1000;
6792 if (rx_stats_active) {
6793 MUTEX_ENTER(&rx_stats_mutex);
6794 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6795 rx_stats.minRtt = thisRtt;
6796 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6797 if (thisRtt.sec > 60) {
6798 MUTEX_EXIT(&rx_stats_mutex);
6799 return; /* somebody set the clock ahead */
6801 rx_stats.maxRtt = thisRtt;
6803 clock_Add(&rx_stats.totalRtt, &thisRtt);
6804 rx_atomic_inc(&rx_stats.nRttSamples);
6805 MUTEX_EXIT(&rx_stats_mutex);
6808 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6810 /* Apply VanJacobson round-trip estimations */
6815 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6816 * srtt is stored as fixed point with 3 bits after the binary
6817 * point (i.e., scaled by 8). The following magic is
6818 * equivalent to the smoothing algorithm in rfc793 with an
6819 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6820 * srtt'*8 = rtt + srtt*7
6821 * srtt'*8 = srtt*8 + rtt - srtt
6822 * srtt' = srtt + rtt/8 - srtt/8
6823 * srtt' = srtt + (rtt - srtt)/8
6826 delta = _8THMSEC(&thisRtt) - call->rtt;
6827 call->rtt += (delta >> 3);
6830 * We accumulate a smoothed rtt variance (actually, a smoothed
6831 * mean difference), then set the retransmit timer to smoothed
6832 * rtt + 4 times the smoothed variance (was 2x in van's original
6833 * paper, but 4x works better for me, and apparently for him as
6835 * rttvar is stored as
6836 * fixed point with 2 bits after the binary point (scaled by
6837 * 4). The following is equivalent to rfc793 smoothing with
6838 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6839 * rttvar'*4 = rttvar*3 + |delta|
6840 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6841 * rttvar' = rttvar + |delta|/4 - rttvar/4
6842 * rttvar' = rttvar + (|delta| - rttvar)/4
6843 * This replaces rfc793's wired-in beta.
6844 * dev*4 = dev*4 + (|actual - expected| - dev)
6850 delta -= (call->rtt_dev << 1);
6851 call->rtt_dev += (delta >> 3);
6853 /* I don't have a stored RTT so I start with this value. Since I'm
6854 * probably just starting a call, and will be pushing more data down
6855 * this, I expect congestion to increase rapidly. So I fudge a
6856 * little, and I set deviance to half the rtt. In practice,
6857 * deviance tends to approach something a little less than
6858 * half the smoothed rtt. */
6859 call->rtt = _8THMSEC(&thisRtt) + 8;
6860 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6862 /* the smoothed RTT time is RTT + 4*MDEV
6864 * We allow a user specified minimum to be set for this, to allow clamping
6865 * at a minimum value in the same way as TCP. In addition, we have to allow
6866 * for the possibility that this packet is answered by a delayed ACK, so we
6867 * add on a fixed 200ms to account for that timer expiring.
6870 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6871 rx_minPeerTimeout) + 200;
6872 clock_Zero(&call->rto);
6873 clock_Addmsec(&call->rto, rtt_timeout);
6875 /* Update the peer, so any new calls start with our values */
6876 peer->rtt_dev = call->rtt_dev;
6877 peer->rtt = call->rtt;
6879 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6880 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6884 /* Find all server connections that have not been active for a long time, and
6887 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6889 struct clock now, when;
6890 clock_GetTime(&now);
6892 /* Find server connection structures that haven't been used for
6893 * greater than rx_idleConnectionTime */
6895 struct rx_connection **conn_ptr, **conn_end;
6896 int i, havecalls = 0;
6897 MUTEX_ENTER(&rx_connHashTable_lock);
6898 for (conn_ptr = &rx_connHashTable[0], conn_end =
6899 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6901 struct rx_connection *conn, *next;
6902 struct rx_call *call;
6906 for (conn = *conn_ptr; conn; conn = next) {
6907 /* XXX -- Shouldn't the connection be locked? */
6910 for (i = 0; i < RX_MAXCALLS; i++) {
6911 call = conn->call[i];
6915 code = MUTEX_TRYENTER(&call->lock);
6918 #ifdef RX_ENABLE_LOCKS
6919 result = rxi_CheckCall(call, 1);
6920 #else /* RX_ENABLE_LOCKS */
6921 result = rxi_CheckCall(call);
6922 #endif /* RX_ENABLE_LOCKS */
6923 MUTEX_EXIT(&call->lock);
6925 /* If CheckCall freed the call, it might
6926 * have destroyed the connection as well,
6927 * which screws up the linked lists.
6933 if (conn->type == RX_SERVER_CONNECTION) {
6934 /* This only actually destroys the connection if
6935 * there are no outstanding calls */
6936 MUTEX_ENTER(&conn->conn_data_lock);
6937 MUTEX_ENTER(&rx_refcnt_mutex);
6938 if (!havecalls && !conn->refCount
6939 && ((conn->lastSendTime + rx_idleConnectionTime) <
6941 conn->refCount++; /* it will be decr in rx_DestroyConn */
6942 MUTEX_EXIT(&rx_refcnt_mutex);
6943 MUTEX_EXIT(&conn->conn_data_lock);
6944 #ifdef RX_ENABLE_LOCKS
6945 rxi_DestroyConnectionNoLock(conn);
6946 #else /* RX_ENABLE_LOCKS */
6947 rxi_DestroyConnection(conn);
6948 #endif /* RX_ENABLE_LOCKS */
6950 #ifdef RX_ENABLE_LOCKS
6952 MUTEX_EXIT(&rx_refcnt_mutex);
6953 MUTEX_EXIT(&conn->conn_data_lock);
6955 #endif /* RX_ENABLE_LOCKS */
6959 #ifdef RX_ENABLE_LOCKS
6960 while (rx_connCleanup_list) {
6961 struct rx_connection *conn;
6962 conn = rx_connCleanup_list;
6963 rx_connCleanup_list = rx_connCleanup_list->next;
6964 MUTEX_EXIT(&rx_connHashTable_lock);
6965 rxi_CleanupConnection(conn);
6966 MUTEX_ENTER(&rx_connHashTable_lock);
6968 MUTEX_EXIT(&rx_connHashTable_lock);
6969 #endif /* RX_ENABLE_LOCKS */
6972 /* Find any peer structures that haven't been used (haven't had an
6973 * associated connection) for greater than rx_idlePeerTime */
6975 struct rx_peer **peer_ptr, **peer_end;
6979 * Why do we need to hold the rx_peerHashTable_lock across
6980 * the incrementing of peer_ptr since the rx_peerHashTable
6981 * array is not changing? We don't.
6983 * By dropping the lock periodically we can permit other
6984 * activities to be performed while a rxi_ReapConnections
6985 * call is in progress. The goal of reap connections
6986 * is to clean up quickly without causing large amounts
6987 * of contention. Therefore, it is important that global
6988 * mutexes not be held for extended periods of time.
6990 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6991 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6993 struct rx_peer *peer, *next, *prev;
6995 MUTEX_ENTER(&rx_peerHashTable_lock);
6996 for (prev = peer = *peer_ptr; peer; peer = next) {
6998 code = MUTEX_TRYENTER(&peer->peer_lock);
6999 if ((code) && (peer->refCount == 0)
7000 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7001 rx_interface_stat_p rpc_stat, nrpc_stat;
7005 * now know that this peer object is one to be
7006 * removed from the hash table. Once it is removed
7007 * it can't be referenced by other threads.
7008 * Lets remove it first and decrement the struct
7009 * nPeerStructs count.
7011 if (peer == *peer_ptr) {
7017 if (rx_stats_active)
7018 rx_atomic_dec(&rx_stats.nPeerStructs);
7021 * Now if we hold references on 'prev' and 'next'
7022 * we can safely drop the rx_peerHashTable_lock
7023 * while we destroy this 'peer' object.
7029 MUTEX_EXIT(&rx_peerHashTable_lock);
7031 MUTEX_EXIT(&peer->peer_lock);
7032 MUTEX_DESTROY(&peer->peer_lock);
7034 (&peer->rpcStats, rpc_stat, nrpc_stat,
7035 rx_interface_stat)) {
7036 unsigned int num_funcs;
7039 queue_Remove(&rpc_stat->queue_header);
7040 queue_Remove(&rpc_stat->all_peers);
7041 num_funcs = rpc_stat->stats[0].func_total;
7043 sizeof(rx_interface_stat_t) +
7044 rpc_stat->stats[0].func_total *
7045 sizeof(rx_function_entry_v1_t);
7047 rxi_Free(rpc_stat, space);
7049 MUTEX_ENTER(&rx_rpc_stats);
7050 rxi_rpc_peer_stat_cnt -= num_funcs;
7051 MUTEX_EXIT(&rx_rpc_stats);
7056 * Regain the rx_peerHashTable_lock and
7057 * decrement the reference count on 'prev'
7060 MUTEX_ENTER(&rx_peerHashTable_lock);
7067 MUTEX_EXIT(&peer->peer_lock);
7072 MUTEX_EXIT(&rx_peerHashTable_lock);
7076 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7077 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7078 * GC, just below. Really, we shouldn't have to keep moving packets from
7079 * one place to another, but instead ought to always know if we can
7080 * afford to hold onto a packet in its particular use. */
7081 MUTEX_ENTER(&rx_freePktQ_lock);
7082 if (rx_waitingForPackets) {
7083 rx_waitingForPackets = 0;
7084 #ifdef RX_ENABLE_LOCKS
7085 CV_BROADCAST(&rx_waitingForPackets_cv);
7087 osi_rxWakeup(&rx_waitingForPackets);
7090 MUTEX_EXIT(&rx_freePktQ_lock);
7093 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7094 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
7098 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7099 * rx.h is sort of strange this is better. This is called with a security
7100 * object before it is discarded. Each connection using a security object has
7101 * its own refcount to the object so it won't actually be freed until the last
7102 * connection is destroyed.
7104 * This is the only rxs module call. A hold could also be written but no one
7108 rxs_Release(struct rx_securityClass *aobj)
7110 return RXS_Close(aobj);
7114 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
7115 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
7116 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
7117 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
7119 /* Adjust our estimate of the transmission rate to this peer, given
7120 * that the packet p was just acked. We can adjust peer->timeout and
7121 * call->twind. Pragmatically, this is called
7122 * only with packets of maximal length.
7123 * Called with peer and call locked.
7127 rxi_ComputeRate(struct rx_peer *peer, struct rx_call *call,
7128 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
7130 afs_int32 xferSize, xferMs;
7134 /* Count down packets */
7135 if (peer->rateFlag > 0)
7137 /* Do nothing until we're enabled */
7138 if (peer->rateFlag != 0)
7143 /* Count only when the ack seems legitimate */
7144 switch (ackReason) {
7145 case RX_ACK_REQUESTED:
7147 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
7151 case RX_ACK_PING_RESPONSE:
7152 if (p) /* want the response to ping-request, not data send */
7154 clock_GetTime(&newTO);
7155 if (clock_Gt(&newTO, &call->pingRequestTime)) {
7156 clock_Sub(&newTO, &call->pingRequestTime);
7157 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
7161 xferSize = rx_AckDataSize(rx_maxSendWindow) + RX_HEADER_SIZE;
7168 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %d.%06d, rtt %u, ps %u)\n",
7169 ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"),
7170 xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
7172 /* Track only packets that are big enough. */
7173 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
7177 /* absorb RTT data (in milliseconds) for these big packets */
7178 if (peer->smRtt == 0) {
7179 peer->smRtt = xferMs;
7181 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
7186 if (peer->countDown) {
7190 peer->countDown = 10; /* recalculate only every so often */
7192 /* In practice, we can measure only the RTT for full packets,
7193 * because of the way Rx acks the data that it receives. (If it's
7194 * smaller than a full packet, it often gets implicitly acked
7195 * either by the call response (from a server) or by the next call
7196 * (from a client), and either case confuses transmission times
7197 * with processing times.) Therefore, replace the above
7198 * more-sophisticated processing with a simpler version, where the
7199 * smoothed RTT is kept for full-size packets, and the time to
7200 * transmit a windowful of full-size packets is simply RTT *
7201 * windowSize. Again, we take two steps:
7202 - ensure the timeout is large enough for a single packet's RTT;
7203 - ensure that the window is small enough to fit in the desired timeout.*/
7205 /* First, the timeout check. */
7206 minTime = peer->smRtt;
7207 /* Get a reasonable estimate for a timeout period */
7209 newTO.sec = minTime / 1000;
7210 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
7212 /* Increase the timeout period so that we can always do at least
7213 * one packet exchange */
7214 if (clock_Gt(&newTO, &peer->timeout)) {
7216 dpf(("CONG peer %lx/%u: timeout %d.%06d ==> %ld.%06d (rtt %u)\n",
7217 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec,
7218 newTO.sec, newTO.usec, peer->smRtt));
7220 peer->timeout = newTO;
7223 /* Now, get an estimate for the transmit window size. */
7224 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
7225 /* Now, convert to the number of full packets that could fit in a
7226 * reasonable fraction of that interval */
7227 minTime /= (peer->smRtt << 1);
7228 minTime = MAX(minTime, rx_minPeerTimeout);
7229 xferSize = minTime; /* (make a copy) */
7231 /* Now clamp the size to reasonable bounds. */
7234 else if (minTime > rx_maxSendWindow)
7235 minTime = rx_maxSendWindow;
7236 /* if (minTime != peer->maxWindow) {
7237 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u)\n",
7238 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
7239 peer->timeout.sec, peer->timeout.usec, peer->smRtt));
7240 peer->maxWindow = minTime;
7241 elide... call->twind = minTime;
7245 /* Cut back on the peer timeout if it had earlier grown unreasonably.
7246 * Discern this by calculating the timeout necessary for rx_Window
7248 if ((xferSize > rx_maxSendWindow) && (peer->timeout.sec >= 3)) {
7249 /* calculate estimate for transmission interval in milliseconds */
7250 minTime = rx_maxSendWindow * peer->smRtt;
7251 if (minTime < 1000) {
7252 dpf(("CONG peer %lx/%u: cut TO %d.%06d by 0.5 (rtt %u)\n",
7253 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
7254 peer->timeout.usec, peer->smRtt));
7256 newTO.sec = 0; /* cut back on timeout by half a second */
7257 newTO.usec = 500000;
7258 clock_Sub(&peer->timeout, &newTO);
7263 } /* end of rxi_ComputeRate */
7264 #endif /* ADAPT_WINDOW */
7272 #define TRACE_OPTION_RX_DEBUG 16
7280 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7281 0, KEY_QUERY_VALUE, &parmKey);
7282 if (code != ERROR_SUCCESS)
7285 dummyLen = sizeof(TraceOption);
7286 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7287 (BYTE *) &TraceOption, &dummyLen);
7288 if (code == ERROR_SUCCESS) {
7289 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7291 RegCloseKey (parmKey);
7292 #endif /* AFS_NT40_ENV */
7297 rx_DebugOnOff(int on)
7301 rxdebug_active = on;
7307 rx_StatsOnOff(int on)
7309 rx_stats_active = on;
7313 /* Don't call this debugging routine directly; use dpf */
7315 rxi_DebugPrint(char *format, ...)
7324 va_start(ap, format);
7326 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7329 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7331 OutputDebugString(msg);
7337 va_start(ap, format);
7339 clock_GetTime(&now);
7340 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7341 (unsigned int)now.usec);
7342 vfprintf(rx_Log, format, ap);
7350 * This function is used to process the rx_stats structure that is local
7351 * to a process as well as an rx_stats structure received from a remote
7352 * process (via rxdebug). Therefore, it needs to do minimal version
7356 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7357 afs_int32 freePackets, char version)
7361 if (size != sizeof(struct rx_statistics)) {
7363 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7364 size, sizeof(struct rx_statistics));
7367 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7370 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7371 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7372 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7373 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7374 s->specialPktAllocFailures);
7376 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7377 s->receivePktAllocFailures, s->sendPktAllocFailures,
7378 s->specialPktAllocFailures);
7382 " greedy %u, " "bogusReads %u (last from host %x), "
7383 "noPackets %u, " "noBuffers %u, " "selects %u, "
7384 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7385 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7386 s->selects, s->sendSelects);
7388 fprintf(file, " packets read: ");
7389 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7390 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7392 fprintf(file, "\n");
7395 " other read counters: data %u, " "ack %u, " "dup %u "
7396 "spurious %u " "dally %u\n", s->dataPacketsRead,
7397 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7398 s->ignorePacketDally);
7400 fprintf(file, " packets sent: ");
7401 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7402 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7404 fprintf(file, "\n");
7407 " other send counters: ack %u, " "data %u (not resends), "
7408 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7409 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7410 s->dataPacketsPushed, s->ignoreAckedPacket);
7413 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7414 s->netSendFailures, (int)s->fatalErrors);
7416 if (s->nRttSamples) {
7417 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7418 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7420 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7421 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7425 " %d server connections, " "%d client connections, "
7426 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7427 s->nServerConns, s->nClientConns, s->nPeerStructs,
7428 s->nCallStructs, s->nFreeCallStructs);
7430 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7431 fprintf(file, " %d clock updates\n", clock_nUpdates);
7435 /* for backward compatibility */
7437 rx_PrintStats(FILE * file)
7439 MUTEX_ENTER(&rx_stats_mutex);
7440 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7441 sizeof(rx_stats), rx_nFreePackets,
7443 MUTEX_EXIT(&rx_stats_mutex);
7447 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7449 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %d.%06d.\n",
7450 ntohl(peer->host), (int)ntohs(peer->port), (int)peer->burstSize,
7451 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
7454 " Rtt %d, " "total sent %d, " "resent %d\n",
7455 peer->rtt, peer->nSent, peer->reSends);
7458 " Packet size %d, " "max in packet skew %d, "
7459 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
7460 (int)peer->outPacketSkew);
7464 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7466 * This mutex protects the following static variables:
7470 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7471 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7473 #define LOCK_RX_DEBUG
7474 #define UNLOCK_RX_DEBUG
7475 #endif /* AFS_PTHREAD_ENV */
7477 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7479 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7480 u_char type, void *inputData, size_t inputLength,
7481 void *outputData, size_t outputLength)
7483 static afs_int32 counter = 100;
7484 time_t waitTime, waitCount;
7485 struct rx_header theader;
7488 struct timeval tv_now, tv_wake, tv_delta;
7489 struct sockaddr_in taddr, faddr;
7503 tp = &tbuffer[sizeof(struct rx_header)];
7504 taddr.sin_family = AF_INET;
7505 taddr.sin_port = remotePort;
7506 taddr.sin_addr.s_addr = remoteAddr;
7507 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7508 taddr.sin_len = sizeof(struct sockaddr_in);
7511 memset(&theader, 0, sizeof(theader));
7512 theader.epoch = htonl(999);
7514 theader.callNumber = htonl(counter);
7517 theader.type = type;
7518 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7519 theader.serviceId = 0;
7521 memcpy(tbuffer, &theader, sizeof(theader));
7522 memcpy(tp, inputData, inputLength);
7524 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7525 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7527 /* see if there's a packet available */
7528 gettimeofday(&tv_wake, NULL);
7529 tv_wake.tv_sec += waitTime;
7532 FD_SET(socket, &imask);
7533 tv_delta.tv_sec = tv_wake.tv_sec;
7534 tv_delta.tv_usec = tv_wake.tv_usec;
7535 gettimeofday(&tv_now, NULL);
7537 if (tv_delta.tv_usec < tv_now.tv_usec) {
7539 tv_delta.tv_usec += 1000000;
7542 tv_delta.tv_usec -= tv_now.tv_usec;
7544 if (tv_delta.tv_sec < tv_now.tv_sec) {
7548 tv_delta.tv_sec -= tv_now.tv_sec;
7551 code = select(0, &imask, 0, 0, &tv_delta);
7552 #else /* AFS_NT40_ENV */
7553 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7554 #endif /* AFS_NT40_ENV */
7555 if (code == 1 && FD_ISSET(socket, &imask)) {
7556 /* now receive a packet */
7557 faddrLen = sizeof(struct sockaddr_in);
7559 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7560 (struct sockaddr *)&faddr, &faddrLen);
7563 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7564 if (counter == ntohl(theader.callNumber))
7572 /* see if we've timed out */
7580 code -= sizeof(struct rx_header);
7581 if (code > outputLength)
7582 code = outputLength;
7583 memcpy(outputData, tp, code);
7586 #endif /* RXDEBUG */
7589 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7590 afs_uint16 remotePort, struct rx_debugStats * stat,
7591 afs_uint32 * supportedValues)
7593 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7595 struct rx_debugIn in;
7597 *supportedValues = 0;
7598 in.type = htonl(RX_DEBUGI_GETSTATS);
7601 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7602 &in, sizeof(in), stat, sizeof(*stat));
7605 * If the call was successful, fixup the version and indicate
7606 * what contents of the stat structure are valid.
7607 * Also do net to host conversion of fields here.
7611 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7612 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7614 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7615 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7617 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7618 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7620 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7621 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7623 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7624 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7626 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7627 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7629 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7630 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7632 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7633 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7635 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7636 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7638 stat->nFreePackets = ntohl(stat->nFreePackets);
7639 stat->packetReclaims = ntohl(stat->packetReclaims);
7640 stat->callsExecuted = ntohl(stat->callsExecuted);
7641 stat->nWaiting = ntohl(stat->nWaiting);
7642 stat->idleThreads = ntohl(stat->idleThreads);
7643 stat->nWaited = ntohl(stat->nWaited);
7644 stat->nPackets = ntohl(stat->nPackets);
7653 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7654 afs_uint16 remotePort, struct rx_statistics * stat,
7655 afs_uint32 * supportedValues)
7657 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7659 struct rx_debugIn in;
7660 afs_int32 *lp = (afs_int32 *) stat;
7664 * supportedValues is currently unused, but added to allow future
7665 * versioning of this function.
7668 *supportedValues = 0;
7669 in.type = htonl(RX_DEBUGI_RXSTATS);
7671 memset(stat, 0, sizeof(*stat));
7673 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7674 &in, sizeof(in), stat, sizeof(*stat));
7679 * Do net to host conversion here
7682 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7693 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7694 afs_uint16 remotePort, size_t version_length,
7697 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7699 return MakeDebugCall(socket, remoteAddr, remotePort,
7700 RX_PACKET_TYPE_VERSION, a, 1, version,
7708 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7709 afs_uint16 remotePort, afs_int32 * nextConnection,
7710 int allConnections, afs_uint32 debugSupportedValues,
7711 struct rx_debugConn * conn,
7712 afs_uint32 * supportedValues)
7714 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7716 struct rx_debugIn in;
7720 * supportedValues is currently unused, but added to allow future
7721 * versioning of this function.
7724 *supportedValues = 0;
7725 if (allConnections) {
7726 in.type = htonl(RX_DEBUGI_GETALLCONN);
7728 in.type = htonl(RX_DEBUGI_GETCONN);
7730 in.index = htonl(*nextConnection);
7731 memset(conn, 0, sizeof(*conn));
7733 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7734 &in, sizeof(in), conn, sizeof(*conn));
7737 *nextConnection += 1;
7740 * Convert old connection format to new structure.
7743 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7744 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7745 #define MOVEvL(a) (conn->a = vL->a)
7747 /* any old or unrecognized version... */
7748 for (i = 0; i < RX_MAXCALLS; i++) {
7749 MOVEvL(callState[i]);
7750 MOVEvL(callMode[i]);
7751 MOVEvL(callFlags[i]);
7752 MOVEvL(callOther[i]);
7754 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7755 MOVEvL(secStats.type);
7756 MOVEvL(secStats.level);
7757 MOVEvL(secStats.flags);
7758 MOVEvL(secStats.expires);
7759 MOVEvL(secStats.packetsReceived);
7760 MOVEvL(secStats.packetsSent);
7761 MOVEvL(secStats.bytesReceived);
7762 MOVEvL(secStats.bytesSent);
7767 * Do net to host conversion here
7769 * I don't convert host or port since we are most likely
7770 * going to want these in NBO.
7772 conn->cid = ntohl(conn->cid);
7773 conn->serial = ntohl(conn->serial);
7774 for (i = 0; i < RX_MAXCALLS; i++) {
7775 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7777 conn->error = ntohl(conn->error);
7778 conn->secStats.flags = ntohl(conn->secStats.flags);
7779 conn->secStats.expires = ntohl(conn->secStats.expires);
7780 conn->secStats.packetsReceived =
7781 ntohl(conn->secStats.packetsReceived);
7782 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7783 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7784 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7785 conn->epoch = ntohl(conn->epoch);
7786 conn->natMTU = ntohl(conn->natMTU);
7795 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7796 afs_uint16 remotePort, afs_int32 * nextPeer,
7797 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7798 afs_uint32 * supportedValues)
7800 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7802 struct rx_debugIn in;
7805 * supportedValues is currently unused, but added to allow future
7806 * versioning of this function.
7809 *supportedValues = 0;
7810 in.type = htonl(RX_DEBUGI_GETPEER);
7811 in.index = htonl(*nextPeer);
7812 memset(peer, 0, sizeof(*peer));
7814 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7815 &in, sizeof(in), peer, sizeof(*peer));
7821 * Do net to host conversion here
7823 * I don't convert host or port since we are most likely
7824 * going to want these in NBO.
7826 peer->ifMTU = ntohs(peer->ifMTU);
7827 peer->idleWhen = ntohl(peer->idleWhen);
7828 peer->refCount = ntohs(peer->refCount);
7829 peer->burstWait.sec = ntohl(peer->burstWait.sec);
7830 peer->burstWait.usec = ntohl(peer->burstWait.usec);
7831 peer->rtt = ntohl(peer->rtt);
7832 peer->rtt_dev = ntohl(peer->rtt_dev);
7833 peer->timeout.sec = 0;
7834 peer->timeout.usec = 0;
7835 peer->nSent = ntohl(peer->nSent);
7836 peer->reSends = ntohl(peer->reSends);
7837 peer->inPacketSkew = ntohl(peer->inPacketSkew);
7838 peer->outPacketSkew = ntohl(peer->outPacketSkew);
7839 peer->rateFlag = ntohl(peer->rateFlag);
7840 peer->natMTU = ntohs(peer->natMTU);
7841 peer->maxMTU = ntohs(peer->maxMTU);
7842 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7843 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7844 peer->MTU = ntohs(peer->MTU);
7845 peer->cwind = ntohs(peer->cwind);
7846 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7847 peer->congestSeq = ntohs(peer->congestSeq);
7848 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7849 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7850 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7851 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7860 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7861 struct rx_debugPeer * peerStats)
7864 afs_int32 error = 1; /* default to "did not succeed" */
7865 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7867 MUTEX_ENTER(&rx_peerHashTable_lock);
7868 for(tp = rx_peerHashTable[hashValue];
7869 tp != NULL; tp = tp->next) {
7870 if (tp->host == peerHost)
7876 MUTEX_EXIT(&rx_peerHashTable_lock);
7880 MUTEX_ENTER(&tp->peer_lock);
7881 peerStats->host = tp->host;
7882 peerStats->port = tp->port;
7883 peerStats->ifMTU = tp->ifMTU;
7884 peerStats->idleWhen = tp->idleWhen;
7885 peerStats->refCount = tp->refCount;
7886 peerStats->burstSize = tp->burstSize;
7887 peerStats->burst = tp->burst;
7888 peerStats->burstWait.sec = tp->burstWait.sec;
7889 peerStats->burstWait.usec = tp->burstWait.usec;
7890 peerStats->rtt = tp->rtt;
7891 peerStats->rtt_dev = tp->rtt_dev;
7892 peerStats->timeout.sec = 0;
7893 peerStats->timeout.usec = 0;
7894 peerStats->nSent = tp->nSent;
7895 peerStats->reSends = tp->reSends;
7896 peerStats->inPacketSkew = tp->inPacketSkew;
7897 peerStats->outPacketSkew = tp->outPacketSkew;
7898 peerStats->rateFlag = tp->rateFlag;
7899 peerStats->natMTU = tp->natMTU;
7900 peerStats->maxMTU = tp->maxMTU;
7901 peerStats->maxDgramPackets = tp->maxDgramPackets;
7902 peerStats->ifDgramPackets = tp->ifDgramPackets;
7903 peerStats->MTU = tp->MTU;
7904 peerStats->cwind = tp->cwind;
7905 peerStats->nDgramPackets = tp->nDgramPackets;
7906 peerStats->congestSeq = tp->congestSeq;
7907 peerStats->bytesSent.high = tp->bytesSent.high;
7908 peerStats->bytesSent.low = tp->bytesSent.low;
7909 peerStats->bytesReceived.high = tp->bytesReceived.high;
7910 peerStats->bytesReceived.low = tp->bytesReceived.low;
7911 MUTEX_EXIT(&tp->peer_lock);
7913 MUTEX_ENTER(&rx_peerHashTable_lock);
7916 MUTEX_EXIT(&rx_peerHashTable_lock);
7924 struct rx_serverQueueEntry *np;
7927 struct rx_call *call;
7928 struct rx_serverQueueEntry *sq;
7932 if (rxinit_status == 1) {
7934 return; /* Already shutdown. */
7938 #ifndef AFS_PTHREAD_ENV
7939 FD_ZERO(&rx_selectMask);
7940 #endif /* AFS_PTHREAD_ENV */
7941 rxi_dataQuota = RX_MAX_QUOTA;
7942 #ifndef AFS_PTHREAD_ENV
7944 #endif /* AFS_PTHREAD_ENV */
7947 #ifndef AFS_PTHREAD_ENV
7948 #ifndef AFS_USE_GETTIMEOFDAY
7950 #endif /* AFS_USE_GETTIMEOFDAY */
7951 #endif /* AFS_PTHREAD_ENV */
7953 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7954 call = queue_First(&rx_freeCallQueue, rx_call);
7956 rxi_Free(call, sizeof(struct rx_call));
7959 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7960 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7966 struct rx_peer **peer_ptr, **peer_end;
7967 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7968 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7970 struct rx_peer *peer, *next;
7972 MUTEX_ENTER(&rx_peerHashTable_lock);
7973 for (peer = *peer_ptr; peer; peer = next) {
7974 rx_interface_stat_p rpc_stat, nrpc_stat;
7977 MUTEX_ENTER(&rx_rpc_stats);
7978 MUTEX_ENTER(&peer->peer_lock);
7980 (&peer->rpcStats, rpc_stat, nrpc_stat,
7981 rx_interface_stat)) {
7982 unsigned int num_funcs;
7985 queue_Remove(&rpc_stat->queue_header);
7986 queue_Remove(&rpc_stat->all_peers);
7987 num_funcs = rpc_stat->stats[0].func_total;
7989 sizeof(rx_interface_stat_t) +
7990 rpc_stat->stats[0].func_total *
7991 sizeof(rx_function_entry_v1_t);
7993 rxi_Free(rpc_stat, space);
7995 /* rx_rpc_stats must be held */
7996 rxi_rpc_peer_stat_cnt -= num_funcs;
7998 MUTEX_EXIT(&peer->peer_lock);
7999 MUTEX_EXIT(&rx_rpc_stats);
8003 if (rx_stats_active)
8004 rx_atomic_dec(&rx_stats.nPeerStructs);
8006 MUTEX_EXIT(&rx_peerHashTable_lock);
8009 for (i = 0; i < RX_MAX_SERVICES; i++) {
8011 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
8013 for (i = 0; i < rx_hashTableSize; i++) {
8014 struct rx_connection *tc, *ntc;
8015 MUTEX_ENTER(&rx_connHashTable_lock);
8016 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
8018 for (j = 0; j < RX_MAXCALLS; j++) {
8020 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
8023 rxi_Free(tc, sizeof(*tc));
8025 MUTEX_EXIT(&rx_connHashTable_lock);
8028 MUTEX_ENTER(&freeSQEList_lock);
8030 while ((np = rx_FreeSQEList)) {
8031 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
8032 MUTEX_DESTROY(&np->lock);
8033 rxi_Free(np, sizeof(*np));
8036 MUTEX_EXIT(&freeSQEList_lock);
8037 MUTEX_DESTROY(&freeSQEList_lock);
8038 MUTEX_DESTROY(&rx_freeCallQueue_lock);
8039 MUTEX_DESTROY(&rx_connHashTable_lock);
8040 MUTEX_DESTROY(&rx_peerHashTable_lock);
8041 MUTEX_DESTROY(&rx_serverPool_lock);
8043 osi_Free(rx_connHashTable,
8044 rx_hashTableSize * sizeof(struct rx_connection *));
8045 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8047 UNPIN(rx_connHashTable,
8048 rx_hashTableSize * sizeof(struct rx_connection *));
8049 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8051 rxi_FreeAllPackets();
8053 MUTEX_ENTER(&rx_quota_mutex);
8054 rxi_dataQuota = RX_MAX_QUOTA;
8055 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
8056 MUTEX_EXIT(&rx_quota_mutex);
8061 #ifdef RX_ENABLE_LOCKS
8063 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
8065 if (!MUTEX_ISMINE(lockaddr))
8066 osi_Panic("Lock not held: %s", msg);
8068 #endif /* RX_ENABLE_LOCKS */
8073 * Routines to implement connection specific data.
8077 rx_KeyCreate(rx_destructor_t rtn)
8080 MUTEX_ENTER(&rxi_keyCreate_lock);
8081 key = rxi_keyCreate_counter++;
8082 rxi_keyCreate_destructor = (rx_destructor_t *)
8083 realloc((void *)rxi_keyCreate_destructor,
8084 (key + 1) * sizeof(rx_destructor_t));
8085 rxi_keyCreate_destructor[key] = rtn;
8086 MUTEX_EXIT(&rxi_keyCreate_lock);
8091 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
8094 MUTEX_ENTER(&conn->conn_data_lock);
8095 if (!conn->specific) {
8096 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
8097 for (i = 0; i < key; i++)
8098 conn->specific[i] = NULL;
8099 conn->nSpecific = key + 1;
8100 conn->specific[key] = ptr;
8101 } else if (key >= conn->nSpecific) {
8102 conn->specific = (void **)
8103 realloc(conn->specific, (key + 1) * sizeof(void *));
8104 for (i = conn->nSpecific; i < key; i++)
8105 conn->specific[i] = NULL;
8106 conn->nSpecific = key + 1;
8107 conn->specific[key] = ptr;
8109 if (conn->specific[key] && rxi_keyCreate_destructor[key])
8110 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
8111 conn->specific[key] = ptr;
8113 MUTEX_EXIT(&conn->conn_data_lock);
8117 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
8120 MUTEX_ENTER(&svc->svc_data_lock);
8121 if (!svc->specific) {
8122 svc->specific = (void **)malloc((key + 1) * sizeof(void *));
8123 for (i = 0; i < key; i++)
8124 svc->specific[i] = NULL;
8125 svc->nSpecific = key + 1;
8126 svc->specific[key] = ptr;
8127 } else if (key >= svc->nSpecific) {
8128 svc->specific = (void **)
8129 realloc(svc->specific, (key + 1) * sizeof(void *));
8130 for (i = svc->nSpecific; i < key; i++)
8131 svc->specific[i] = NULL;
8132 svc->nSpecific = key + 1;
8133 svc->specific[key] = ptr;
8135 if (svc->specific[key] && rxi_keyCreate_destructor[key])
8136 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
8137 svc->specific[key] = ptr;
8139 MUTEX_EXIT(&svc->svc_data_lock);
8143 rx_GetSpecific(struct rx_connection *conn, int key)
8146 MUTEX_ENTER(&conn->conn_data_lock);
8147 if (key >= conn->nSpecific)
8150 ptr = conn->specific[key];
8151 MUTEX_EXIT(&conn->conn_data_lock);
8156 rx_GetServiceSpecific(struct rx_service *svc, int key)
8159 MUTEX_ENTER(&svc->svc_data_lock);
8160 if (key >= svc->nSpecific)
8163 ptr = svc->specific[key];
8164 MUTEX_EXIT(&svc->svc_data_lock);
8169 #endif /* !KERNEL */
8172 * processStats is a queue used to store the statistics for the local
8173 * process. Its contents are similar to the contents of the rpcStats
8174 * queue on a rx_peer structure, but the actual data stored within
8175 * this queue contains totals across the lifetime of the process (assuming
8176 * the stats have not been reset) - unlike the per peer structures
8177 * which can come and go based upon the peer lifetime.
8180 static struct rx_queue processStats = { &processStats, &processStats };
8183 * peerStats is a queue used to store the statistics for all peer structs.
8184 * Its contents are the union of all the peer rpcStats queues.
8187 static struct rx_queue peerStats = { &peerStats, &peerStats };
8190 * rxi_monitor_processStats is used to turn process wide stat collection
8194 static int rxi_monitor_processStats = 0;
8197 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8200 static int rxi_monitor_peerStats = 0;
8203 * rxi_AddRpcStat - given all of the information for a particular rpc
8204 * call, create (if needed) and update the stat totals for the rpc.
8208 * IN stats - the queue of stats that will be updated with the new value
8210 * IN rxInterface - a unique number that identifies the rpc interface
8212 * IN currentFunc - the index of the function being invoked
8214 * IN totalFunc - the total number of functions in this interface
8216 * IN queueTime - the amount of time this function waited for a thread
8218 * IN execTime - the amount of time this function invocation took to execute
8220 * IN bytesSent - the number bytes sent by this invocation
8222 * IN bytesRcvd - the number bytes received by this invocation
8224 * IN isServer - if true, this invocation was made to a server
8226 * IN remoteHost - the ip address of the remote host
8228 * IN remotePort - the port of the remote host
8230 * IN addToPeerList - if != 0, add newly created stat to the global peer list
8232 * INOUT counter - if a new stats structure is allocated, the counter will
8233 * be updated with the new number of allocated stat structures
8241 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
8242 afs_uint32 currentFunc, afs_uint32 totalFunc,
8243 struct clock *queueTime, struct clock *execTime,
8244 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
8245 afs_uint32 remoteHost, afs_uint32 remotePort,
8246 int addToPeerList, unsigned int *counter)
8249 rx_interface_stat_p rpc_stat, nrpc_stat;
8252 * See if there's already a structure for this interface
8255 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8256 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8257 && (rpc_stat->stats[0].remote_is_server == isServer))
8262 * Didn't find a match so allocate a new structure and add it to the
8266 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
8267 || (rpc_stat->stats[0].interfaceId != rxInterface)
8268 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8273 sizeof(rx_interface_stat_t) +
8274 totalFunc * sizeof(rx_function_entry_v1_t);
8276 rpc_stat = rxi_Alloc(space);
8277 if (rpc_stat == NULL) {
8281 *counter += totalFunc;
8282 for (i = 0; i < totalFunc; i++) {
8283 rpc_stat->stats[i].remote_peer = remoteHost;
8284 rpc_stat->stats[i].remote_port = remotePort;
8285 rpc_stat->stats[i].remote_is_server = isServer;
8286 rpc_stat->stats[i].interfaceId = rxInterface;
8287 rpc_stat->stats[i].func_total = totalFunc;
8288 rpc_stat->stats[i].func_index = i;
8289 hzero(rpc_stat->stats[i].invocations);
8290 hzero(rpc_stat->stats[i].bytes_sent);
8291 hzero(rpc_stat->stats[i].bytes_rcvd);
8292 rpc_stat->stats[i].queue_time_sum.sec = 0;
8293 rpc_stat->stats[i].queue_time_sum.usec = 0;
8294 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8295 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8296 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8297 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8298 rpc_stat->stats[i].queue_time_max.sec = 0;
8299 rpc_stat->stats[i].queue_time_max.usec = 0;
8300 rpc_stat->stats[i].execution_time_sum.sec = 0;
8301 rpc_stat->stats[i].execution_time_sum.usec = 0;
8302 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8303 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8304 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8305 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8306 rpc_stat->stats[i].execution_time_max.sec = 0;
8307 rpc_stat->stats[i].execution_time_max.usec = 0;
8309 queue_Prepend(stats, rpc_stat);
8310 if (addToPeerList) {
8311 queue_Prepend(&peerStats, &rpc_stat->all_peers);
8316 * Increment the stats for this function
8319 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
8320 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
8321 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
8322 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8323 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8324 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8325 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8327 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8328 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8330 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8331 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8333 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8334 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8336 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8337 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8345 * rx_IncrementTimeAndCount - increment the times and count for a particular
8350 * IN peer - the peer who invoked the rpc
8352 * IN rxInterface - a unique number that identifies the rpc interface
8354 * IN currentFunc - the index of the function being invoked
8356 * IN totalFunc - the total number of functions in this interface
8358 * IN queueTime - the amount of time this function waited for a thread
8360 * IN execTime - the amount of time this function invocation took to execute
8362 * IN bytesSent - the number bytes sent by this invocation
8364 * IN bytesRcvd - the number bytes received by this invocation
8366 * IN isServer - if true, this invocation was made to a server
8374 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8375 afs_uint32 currentFunc, afs_uint32 totalFunc,
8376 struct clock *queueTime, struct clock *execTime,
8377 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8381 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8384 MUTEX_ENTER(&rx_rpc_stats);
8386 if (rxi_monitor_peerStats) {
8387 MUTEX_ENTER(&peer->peer_lock);
8388 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8389 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8390 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8391 MUTEX_EXIT(&peer->peer_lock);
8394 if (rxi_monitor_processStats) {
8395 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8396 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8397 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8400 MUTEX_EXIT(&rx_rpc_stats);
8405 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8409 * IN callerVersion - the rpc stat version of the caller.
8411 * IN count - the number of entries to marshall.
8413 * IN stats - pointer to stats to be marshalled.
8415 * OUT ptr - Where to store the marshalled data.
8422 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8423 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8429 * We only support the first version
8431 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8432 *(ptr++) = stats->remote_peer;
8433 *(ptr++) = stats->remote_port;
8434 *(ptr++) = stats->remote_is_server;
8435 *(ptr++) = stats->interfaceId;
8436 *(ptr++) = stats->func_total;
8437 *(ptr++) = stats->func_index;
8438 *(ptr++) = hgethi(stats->invocations);
8439 *(ptr++) = hgetlo(stats->invocations);
8440 *(ptr++) = hgethi(stats->bytes_sent);
8441 *(ptr++) = hgetlo(stats->bytes_sent);
8442 *(ptr++) = hgethi(stats->bytes_rcvd);
8443 *(ptr++) = hgetlo(stats->bytes_rcvd);
8444 *(ptr++) = stats->queue_time_sum.sec;
8445 *(ptr++) = stats->queue_time_sum.usec;
8446 *(ptr++) = stats->queue_time_sum_sqr.sec;
8447 *(ptr++) = stats->queue_time_sum_sqr.usec;
8448 *(ptr++) = stats->queue_time_min.sec;
8449 *(ptr++) = stats->queue_time_min.usec;
8450 *(ptr++) = stats->queue_time_max.sec;
8451 *(ptr++) = stats->queue_time_max.usec;
8452 *(ptr++) = stats->execution_time_sum.sec;
8453 *(ptr++) = stats->execution_time_sum.usec;
8454 *(ptr++) = stats->execution_time_sum_sqr.sec;
8455 *(ptr++) = stats->execution_time_sum_sqr.usec;
8456 *(ptr++) = stats->execution_time_min.sec;
8457 *(ptr++) = stats->execution_time_min.usec;
8458 *(ptr++) = stats->execution_time_max.sec;
8459 *(ptr++) = stats->execution_time_max.usec;
8465 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8470 * IN callerVersion - the rpc stat version of the caller
8472 * OUT myVersion - the rpc stat version of this function
8474 * OUT clock_sec - local time seconds
8476 * OUT clock_usec - local time microseconds
8478 * OUT allocSize - the number of bytes allocated to contain stats
8480 * OUT statCount - the number stats retrieved from this process.
8482 * OUT stats - the actual stats retrieved from this process.
8486 * Returns void. If successful, stats will != NULL.
8490 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8491 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8492 size_t * allocSize, afs_uint32 * statCount,
8493 afs_uint32 ** stats)
8503 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8506 * Check to see if stats are enabled
8509 MUTEX_ENTER(&rx_rpc_stats);
8510 if (!rxi_monitor_processStats) {
8511 MUTEX_EXIT(&rx_rpc_stats);
8515 clock_GetTime(&now);
8516 *clock_sec = now.sec;
8517 *clock_usec = now.usec;
8520 * Allocate the space based upon the caller version
8522 * If the client is at an older version than we are,
8523 * we return the statistic data in the older data format, but
8524 * we still return our version number so the client knows we
8525 * are maintaining more data than it can retrieve.
8528 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8529 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8530 *statCount = rxi_rpc_process_stat_cnt;
8533 * This can't happen yet, but in the future version changes
8534 * can be handled by adding additional code here
8538 if (space > (size_t) 0) {
8540 ptr = *stats = rxi_Alloc(space);
8543 rx_interface_stat_p rpc_stat, nrpc_stat;
8547 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8549 * Copy the data based upon the caller version
8551 rx_MarshallProcessRPCStats(callerVersion,
8552 rpc_stat->stats[0].func_total,
8553 rpc_stat->stats, &ptr);
8559 MUTEX_EXIT(&rx_rpc_stats);
8564 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8568 * IN callerVersion - the rpc stat version of the caller
8570 * OUT myVersion - the rpc stat version of this function
8572 * OUT clock_sec - local time seconds
8574 * OUT clock_usec - local time microseconds
8576 * OUT allocSize - the number of bytes allocated to contain stats
8578 * OUT statCount - the number of stats retrieved from the individual
8581 * OUT stats - the actual stats retrieved from the individual peer structures.
8585 * Returns void. If successful, stats will != NULL.
8589 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8590 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8591 size_t * allocSize, afs_uint32 * statCount,
8592 afs_uint32 ** stats)
8602 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8605 * Check to see if stats are enabled
8608 MUTEX_ENTER(&rx_rpc_stats);
8609 if (!rxi_monitor_peerStats) {
8610 MUTEX_EXIT(&rx_rpc_stats);
8614 clock_GetTime(&now);
8615 *clock_sec = now.sec;
8616 *clock_usec = now.usec;
8619 * Allocate the space based upon the caller version
8621 * If the client is at an older version than we are,
8622 * we return the statistic data in the older data format, but
8623 * we still return our version number so the client knows we
8624 * are maintaining more data than it can retrieve.
8627 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8628 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8629 *statCount = rxi_rpc_peer_stat_cnt;
8632 * This can't happen yet, but in the future version changes
8633 * can be handled by adding additional code here
8637 if (space > (size_t) 0) {
8639 ptr = *stats = rxi_Alloc(space);
8642 rx_interface_stat_p rpc_stat, nrpc_stat;
8646 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8648 * We have to fix the offset of rpc_stat since we are
8649 * keeping this structure on two rx_queues. The rx_queue
8650 * package assumes that the rx_queue member is the first
8651 * member of the structure. That is, rx_queue assumes that
8652 * any one item is only on one queue at a time. We are
8653 * breaking that assumption and so we have to do a little
8654 * math to fix our pointers.
8657 fix_offset = (char *)rpc_stat;
8658 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8659 rpc_stat = (rx_interface_stat_p) fix_offset;
8662 * Copy the data based upon the caller version
8664 rx_MarshallProcessRPCStats(callerVersion,
8665 rpc_stat->stats[0].func_total,
8666 rpc_stat->stats, &ptr);
8672 MUTEX_EXIT(&rx_rpc_stats);
8677 * rx_FreeRPCStats - free memory allocated by
8678 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8682 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8683 * rx_RetrievePeerRPCStats
8685 * IN allocSize - the number of bytes in stats.
8693 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8695 rxi_Free(stats, allocSize);
8699 * rx_queryProcessRPCStats - see if process rpc stat collection is
8700 * currently enabled.
8706 * Returns 0 if stats are not enabled != 0 otherwise
8710 rx_queryProcessRPCStats(void)
8713 MUTEX_ENTER(&rx_rpc_stats);
8714 rc = rxi_monitor_processStats;
8715 MUTEX_EXIT(&rx_rpc_stats);
8720 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8726 * Returns 0 if stats are not enabled != 0 otherwise
8730 rx_queryPeerRPCStats(void)
8733 MUTEX_ENTER(&rx_rpc_stats);
8734 rc = rxi_monitor_peerStats;
8735 MUTEX_EXIT(&rx_rpc_stats);
8740 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8750 rx_enableProcessRPCStats(void)
8752 MUTEX_ENTER(&rx_rpc_stats);
8753 rx_enable_stats = 1;
8754 rxi_monitor_processStats = 1;
8755 MUTEX_EXIT(&rx_rpc_stats);
8759 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8769 rx_enablePeerRPCStats(void)
8771 MUTEX_ENTER(&rx_rpc_stats);
8772 rx_enable_stats = 1;
8773 rxi_monitor_peerStats = 1;
8774 MUTEX_EXIT(&rx_rpc_stats);
8778 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8788 rx_disableProcessRPCStats(void)
8790 rx_interface_stat_p rpc_stat, nrpc_stat;
8793 MUTEX_ENTER(&rx_rpc_stats);
8796 * Turn off process statistics and if peer stats is also off, turn
8800 rxi_monitor_processStats = 0;
8801 if (rxi_monitor_peerStats == 0) {
8802 rx_enable_stats = 0;
8805 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8806 unsigned int num_funcs = 0;
8809 queue_Remove(rpc_stat);
8810 num_funcs = rpc_stat->stats[0].func_total;
8812 sizeof(rx_interface_stat_t) +
8813 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8815 rxi_Free(rpc_stat, space);
8816 rxi_rpc_process_stat_cnt -= num_funcs;
8818 MUTEX_EXIT(&rx_rpc_stats);
8822 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8832 rx_disablePeerRPCStats(void)
8834 struct rx_peer **peer_ptr, **peer_end;
8838 * Turn off peer statistics and if process stats is also off, turn
8842 rxi_monitor_peerStats = 0;
8843 if (rxi_monitor_processStats == 0) {
8844 rx_enable_stats = 0;
8847 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8848 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8850 struct rx_peer *peer, *next, *prev;
8852 MUTEX_ENTER(&rx_peerHashTable_lock);
8853 MUTEX_ENTER(&rx_rpc_stats);
8854 for (prev = peer = *peer_ptr; peer; peer = next) {
8856 code = MUTEX_TRYENTER(&peer->peer_lock);
8858 rx_interface_stat_p rpc_stat, nrpc_stat;
8861 if (prev == *peer_ptr) {
8872 MUTEX_EXIT(&rx_peerHashTable_lock);
8875 (&peer->rpcStats, rpc_stat, nrpc_stat,
8876 rx_interface_stat)) {
8877 unsigned int num_funcs = 0;
8880 queue_Remove(&rpc_stat->queue_header);
8881 queue_Remove(&rpc_stat->all_peers);
8882 num_funcs = rpc_stat->stats[0].func_total;
8884 sizeof(rx_interface_stat_t) +
8885 rpc_stat->stats[0].func_total *
8886 sizeof(rx_function_entry_v1_t);
8888 rxi_Free(rpc_stat, space);
8889 rxi_rpc_peer_stat_cnt -= num_funcs;
8891 MUTEX_EXIT(&peer->peer_lock);
8893 MUTEX_ENTER(&rx_peerHashTable_lock);
8903 MUTEX_EXIT(&rx_rpc_stats);
8904 MUTEX_EXIT(&rx_peerHashTable_lock);
8909 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8914 * IN clearFlag - flag indicating which stats to clear
8922 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8924 rx_interface_stat_p rpc_stat, nrpc_stat;
8926 MUTEX_ENTER(&rx_rpc_stats);
8928 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8929 unsigned int num_funcs = 0, i;
8930 num_funcs = rpc_stat->stats[0].func_total;
8931 for (i = 0; i < num_funcs; i++) {
8932 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8933 hzero(rpc_stat->stats[i].invocations);
8935 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8936 hzero(rpc_stat->stats[i].bytes_sent);
8938 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8939 hzero(rpc_stat->stats[i].bytes_rcvd);
8941 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8942 rpc_stat->stats[i].queue_time_sum.sec = 0;
8943 rpc_stat->stats[i].queue_time_sum.usec = 0;
8945 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8946 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8947 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8949 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8950 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8951 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8953 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8954 rpc_stat->stats[i].queue_time_max.sec = 0;
8955 rpc_stat->stats[i].queue_time_max.usec = 0;
8957 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8958 rpc_stat->stats[i].execution_time_sum.sec = 0;
8959 rpc_stat->stats[i].execution_time_sum.usec = 0;
8961 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8962 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8963 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8965 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8966 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8967 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8969 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8970 rpc_stat->stats[i].execution_time_max.sec = 0;
8971 rpc_stat->stats[i].execution_time_max.usec = 0;
8976 MUTEX_EXIT(&rx_rpc_stats);
8980 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8985 * IN clearFlag - flag indicating which stats to clear
8993 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8995 rx_interface_stat_p rpc_stat, nrpc_stat;
8997 MUTEX_ENTER(&rx_rpc_stats);
8999 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
9000 unsigned int num_funcs = 0, i;
9003 * We have to fix the offset of rpc_stat since we are
9004 * keeping this structure on two rx_queues. The rx_queue
9005 * package assumes that the rx_queue member is the first
9006 * member of the structure. That is, rx_queue assumes that
9007 * any one item is only on one queue at a time. We are
9008 * breaking that assumption and so we have to do a little
9009 * math to fix our pointers.
9012 fix_offset = (char *)rpc_stat;
9013 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
9014 rpc_stat = (rx_interface_stat_p) fix_offset;
9016 num_funcs = rpc_stat->stats[0].func_total;
9017 for (i = 0; i < num_funcs; i++) {
9018 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9019 hzero(rpc_stat->stats[i].invocations);
9021 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9022 hzero(rpc_stat->stats[i].bytes_sent);
9024 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9025 hzero(rpc_stat->stats[i].bytes_rcvd);
9027 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9028 rpc_stat->stats[i].queue_time_sum.sec = 0;
9029 rpc_stat->stats[i].queue_time_sum.usec = 0;
9031 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9032 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9033 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9035 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9036 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9037 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9039 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9040 rpc_stat->stats[i].queue_time_max.sec = 0;
9041 rpc_stat->stats[i].queue_time_max.usec = 0;
9043 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9044 rpc_stat->stats[i].execution_time_sum.sec = 0;
9045 rpc_stat->stats[i].execution_time_sum.usec = 0;
9047 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9048 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9049 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9051 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9052 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9053 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9055 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9056 rpc_stat->stats[i].execution_time_max.sec = 0;
9057 rpc_stat->stats[i].execution_time_max.usec = 0;
9062 MUTEX_EXIT(&rx_rpc_stats);
9066 * rxi_rxstat_userok points to a routine that returns 1 if the caller
9067 * is authorized to enable/disable/clear RX statistics.
9069 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
9072 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
9074 rxi_rxstat_userok = proc;
9078 rx_RxStatUserOk(struct rx_call *call)
9080 if (!rxi_rxstat_userok)
9082 return rxi_rxstat_userok(call);
9087 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
9088 * function in the MSVC runtime DLL (msvcrt.dll).
9090 * Note: the system serializes calls to this function.
9093 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
9094 DWORD reason, /* reason function is being called */
9095 LPVOID reserved) /* reserved for future use */
9098 case DLL_PROCESS_ATTACH:
9099 /* library is being attached to a process */
9103 case DLL_PROCESS_DETACH:
9110 #endif /* AFS_NT40_ENV */
9113 int rx_DumpCalls(FILE *outputFile, char *cookie)
9115 #ifdef RXDEBUG_PACKET
9116 #ifdef KDUMP_RX_LOCK
9117 struct rx_call_rx_lock *c;
9124 #define RXDPRINTF sprintf
9125 #define RXDPRINTOUT output
9127 #define RXDPRINTF fprintf
9128 #define RXDPRINTOUT outputFile
9131 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9133 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9136 for (c = rx_allCallsp; c; c = c->allNextp) {
9137 u_short rqc, tqc, iovqc;
9138 struct rx_packet *p, *np;
9140 MUTEX_ENTER(&c->lock);
9141 queue_Count(&c->rq, p, np, rx_packet, rqc);
9142 queue_Count(&c->tq, p, np, rx_packet, tqc);
9143 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
9145 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, "
9146 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9147 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9148 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9149 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
9150 #ifdef RX_ENABLE_LOCKS
9153 #ifdef RX_REFCOUNT_CHECK
9154 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9155 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9158 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,
9159 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9160 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9161 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9162 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9163 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
9164 #ifdef RX_ENABLE_LOCKS
9165 , (afs_uint32)c->refCount
9167 #ifdef RX_REFCOUNT_CHECK
9168 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9171 MUTEX_EXIT(&c->lock);
9174 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9177 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9179 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9181 #endif /* RXDEBUG_PACKET */