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>
75 #include <opr/queue.h>
79 #include "rx_atomic.h"
80 #include "rx_globals.h"
82 #include "rx_internal.h"
89 #include "rx_packet.h"
90 #include "rx_server.h"
92 #include <afs/rxgen_consts.h>
95 #ifdef AFS_PTHREAD_ENV
97 int (*registerProgram) (pid_t, char *) = 0;
98 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
101 int (*registerProgram) (PROCESS, char *) = 0;
102 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
106 /* Local static routines */
107 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
108 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
109 struct rx_call *, struct rx_peer *,
111 static void rxi_Resend(struct rxevent *event, void *arg0, void *arg1,
113 static void rxi_SendDelayedAck(struct rxevent *event, void *call,
114 void *dummy, int dummy2);
115 static void rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1,
116 void *dummy, int dummy2);
117 static void rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1,
118 void *unused, int unused2);
119 static void rxi_ReapConnections(struct rxevent *unused, void *unused1,
120 void *unused2, int unused3);
121 static struct rx_packet *rxi_SendCallAbort(struct rx_call *call,
122 struct rx_packet *packet,
123 int istack, int force);
124 static void rxi_AckAll(struct rx_call *call);
125 static struct rx_connection
126 *rxi_FindConnection(osi_socket socket, afs_uint32 host, u_short port,
127 u_short serviceId, afs_uint32 cid,
128 afs_uint32 epoch, int type, u_int securityIndex);
129 static struct rx_packet
130 *rxi_ReceiveDataPacket(struct rx_call *call, struct rx_packet *np,
131 int istack, osi_socket socket,
132 afs_uint32 host, u_short port, int *tnop,
133 struct rx_call **newcallp);
134 static struct rx_packet
135 *rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
137 static struct rx_packet
138 *rxi_ReceiveResponsePacket(struct rx_connection *conn,
139 struct rx_packet *np, int istack);
140 static struct rx_packet
141 *rxi_ReceiveChallengePacket(struct rx_connection *conn,
142 struct rx_packet *np, int istack);
143 static void rxi_AttachServerProc(struct rx_call *call, osi_socket socket,
144 int *tnop, struct rx_call **newcallp);
145 static void rxi_ClearTransmitQueue(struct rx_call *call, int force);
146 static void rxi_ClearReceiveQueue(struct rx_call *call);
147 static void rxi_ResetCall(struct rx_call *call, int newcall);
148 static void rxi_ScheduleKeepAliveEvent(struct rx_call *call);
149 static void rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn);
150 static void rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs);
151 static void rxi_KeepAliveOn(struct rx_call *call);
152 static void rxi_GrowMTUOn(struct rx_call *call);
153 static void rxi_ChallengeOn(struct rx_connection *conn);
154 static int rxi_CheckCall(struct rx_call *call, int haveCTLock);
156 #ifdef RX_ENABLE_LOCKS
157 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
160 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
161 rx_atomic_t rxi_start_in_error;
163 #endif /* RX_ENABLE_LOCKS */
165 /* Constant delay time before sending an acknowledge of the last packet
166 * received. This is to avoid sending an extra acknowledge when the
167 * client is about to make another call, anyway, or the server is
170 * The lastAckDelay may not exceeed 400ms without causing peers to
171 * unecessarily timeout.
173 struct clock rx_lastAckDelay = {0, 400000};
175 /* Constant delay time before sending a soft ack when none was requested.
176 * This is to make sure we send soft acks before the sender times out,
177 * Normally we wait and send a hard ack when the receiver consumes the packet
179 * This value has been 100ms in all shipping versions of OpenAFS. Changing it
180 * will require changes to the peer's RTT calculations.
182 struct clock rx_softAckDelay = {0, 100000};
185 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
186 * currently allocated within rx. This number is used to allocate the
187 * memory required to return the statistics when queried.
188 * Protected by the rx_rpc_stats mutex.
191 static unsigned int rxi_rpc_peer_stat_cnt;
194 * rxi_rpc_process_stat_cnt counts the total number of local process stat
195 * structures currently allocated within rx. The number is used to allocate
196 * the memory required to return the statistics when queried.
197 * Protected by the rx_rpc_stats mutex.
200 static unsigned int rxi_rpc_process_stat_cnt;
203 * rxi_busyChannelError is a boolean. It indicates whether or not RX_CALL_BUSY
204 * errors should be reported to the application when a call channel appears busy
205 * (inferred from the receipt of RX_PACKET_TYPE_BUSY packets on the channel),
206 * and there are other call channels in the connection that are not busy.
207 * If 0, we do not return errors upon receiving busy packets; we just keep
208 * trying on the same call channel until we hit a timeout.
210 static afs_int32 rxi_busyChannelError = 0;
212 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
213 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
215 /* Incoming calls wait on this queue when there are no available
216 * server processes */
217 struct opr_queue rx_incomingCallQueue;
219 /* Server processes wait on this queue when there are no appropriate
220 * calls to process */
221 struct opr_queue rx_idleServerQueue;
223 #if !defined(offsetof)
224 #include <stddef.h> /* for definition of offsetof() */
227 #ifdef RX_ENABLE_LOCKS
228 afs_kmutex_t rx_atomic_mutex;
231 /* Forward prototypes */
232 static struct rx_call * rxi_NewCall(struct rx_connection *, int);
235 putConnection (struct rx_connection *conn) {
236 MUTEX_ENTER(&rx_refcnt_mutex);
238 MUTEX_EXIT(&rx_refcnt_mutex);
241 #ifdef AFS_PTHREAD_ENV
244 * Use procedural initialization of mutexes/condition variables
248 extern afs_kmutex_t rx_quota_mutex;
249 extern afs_kmutex_t rx_pthread_mutex;
250 extern afs_kmutex_t rx_packets_mutex;
251 extern afs_kmutex_t rx_refcnt_mutex;
252 extern afs_kmutex_t des_init_mutex;
253 extern afs_kmutex_t des_random_mutex;
254 extern afs_kmutex_t rx_clock_mutex;
255 extern afs_kmutex_t rxi_connCacheMutex;
256 extern afs_kmutex_t event_handler_mutex;
257 extern afs_kmutex_t listener_mutex;
258 extern afs_kmutex_t rx_if_init_mutex;
259 extern afs_kmutex_t rx_if_mutex;
261 extern afs_kcondvar_t rx_event_handler_cond;
262 extern afs_kcondvar_t rx_listener_cond;
264 static afs_kmutex_t epoch_mutex;
265 static afs_kmutex_t rx_init_mutex;
266 static afs_kmutex_t rx_debug_mutex;
267 static afs_kmutex_t rx_rpc_stats;
270 rxi_InitPthread(void)
272 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
273 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
274 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
275 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
276 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
277 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
278 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
279 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
280 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
281 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
282 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
283 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
284 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
285 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
286 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
288 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
289 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
291 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
292 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
294 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
295 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
296 #ifdef RX_ENABLE_LOCKS
299 #endif /* RX_LOCKS_DB */
300 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
301 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
303 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
305 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
307 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
309 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
310 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
311 #endif /* RX_ENABLE_LOCKS */
314 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
315 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
317 * The rx_stats_mutex mutex protects the following global variables:
318 * rxi_lowConnRefCount
319 * rxi_lowPeerRefCount
328 * The rx_quota_mutex mutex protects the following global variables:
336 * The rx_freePktQ_lock protects the following global variables:
341 * The rx_packets_mutex mutex protects the following global variables:
349 * The rx_pthread_mutex mutex protects the following global variables:
350 * rxi_fcfs_thread_num
353 #define INIT_PTHREAD_LOCKS
357 /* Variables for handling the minProcs implementation. availProcs gives the
358 * number of threads available in the pool at this moment (not counting dudes
359 * executing right now). totalMin gives the total number of procs required
360 * for handling all minProcs requests. minDeficit is a dynamic variable
361 * tracking the # of procs required to satisfy all of the remaining minProcs
363 * For fine grain locking to work, the quota check and the reservation of
364 * a server thread has to come while rxi_availProcs and rxi_minDeficit
365 * are locked. To this end, the code has been modified under #ifdef
366 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
367 * same time. A new function, ReturnToServerPool() returns the allocation.
369 * A call can be on several queue's (but only one at a time). When
370 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
371 * that no one else is touching the queue. To this end, we store the address
372 * of the queue lock in the call structure (under the call lock) when we
373 * put the call on a queue, and we clear the call_queue_lock when the
374 * call is removed from a queue (once the call lock has been obtained).
375 * This allows rxi_ResetCall to safely synchronize with others wishing
376 * to manipulate the queue.
379 #if defined(RX_ENABLE_LOCKS)
380 static afs_kmutex_t rx_rpc_stats;
383 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
384 ** pretty good that the next packet coming in is from the same connection
385 ** as the last packet, since we're send multiple packets in a transmit window.
387 struct rx_connection *rxLastConn = 0;
389 #ifdef RX_ENABLE_LOCKS
390 /* The locking hierarchy for rx fine grain locking is composed of these
393 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
394 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
395 * call->lock - locks call data fields.
396 * These are independent of each other:
397 * rx_freeCallQueue_lock
402 * serverQueueEntry->lock
403 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
405 * peer->lock - locks peer data fields.
406 * conn_data_lock - that more than one thread is not updating a conn data
407 * field at the same time.
418 * Do we need a lock to protect the peer field in the conn structure?
419 * conn->peer was previously a constant for all intents and so has no
420 * lock protecting this field. The multihomed client delta introduced
421 * a RX code change : change the peer field in the connection structure
422 * to that remote interface from which the last packet for this
423 * connection was sent out. This may become an issue if further changes
426 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
427 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
429 /* rxdb_fileID is used to identify the lock location, along with line#. */
430 static int rxdb_fileID = RXDB_FILE_RX;
431 #endif /* RX_LOCKS_DB */
432 #else /* RX_ENABLE_LOCKS */
433 #define SET_CALL_QUEUE_LOCK(C, L)
434 #define CLEAR_CALL_QUEUE_LOCK(C)
435 #endif /* RX_ENABLE_LOCKS */
436 struct rx_serverQueueEntry *rx_waitForPacket = 0;
437 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
439 /* ------------Exported Interfaces------------- */
441 /* This function allows rxkad to set the epoch to a suitably random number
442 * which rx_NewConnection will use in the future. The principle purpose is to
443 * get rxnull connections to use the same epoch as the rxkad connections do, at
444 * least once the first rxkad connection is established. This is important now
445 * that the host/port addresses aren't used in FindConnection: the uniqueness
446 * of epoch/cid matters and the start time won't do. */
448 #ifdef AFS_PTHREAD_ENV
450 * This mutex protects the following global variables:
454 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
455 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
459 #endif /* AFS_PTHREAD_ENV */
462 rx_SetEpoch(afs_uint32 epoch)
469 /* Initialize rx. A port number may be mentioned, in which case this
470 * becomes the default port number for any service installed later.
471 * If 0 is provided for the port number, a random port will be chosen
472 * by the kernel. Whether this will ever overlap anything in
473 * /etc/services is anybody's guess... Returns 0 on success, -1 on
478 int rxinit_status = 1;
479 #ifdef AFS_PTHREAD_ENV
481 * This mutex protects the following global variables:
485 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
486 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
489 #define UNLOCK_RX_INIT
493 rx_InitHost(u_int host, u_int port)
500 char *htable, *ptable;
507 if (rxinit_status == 0) {
508 tmp_status = rxinit_status;
510 return tmp_status; /* Already started; return previous error code. */
516 if (afs_winsockInit() < 0)
522 * Initialize anything necessary to provide a non-premptive threading
525 rxi_InitializeThreadSupport();
528 /* Allocate and initialize a socket for client and perhaps server
531 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
532 if (rx_socket == OSI_NULLSOCKET) {
536 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
539 #endif /* RX_LOCKS_DB */
540 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
541 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
542 MUTEX_INIT(&rx_atomic_mutex, "rx_atomic_mutex", MUTEX_DEFAULT, 0);
543 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
544 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
545 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
546 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
547 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
548 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
549 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
551 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
553 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
555 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
557 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
558 #if defined(AFS_HPUX110_ENV)
560 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
561 #endif /* AFS_HPUX110_ENV */
562 #endif /* RX_ENABLE_LOCKS && KERNEL */
565 rx_connDeadTime = 12;
566 rx_tranquil = 0; /* reset flag */
567 rxi_ResetStatistics();
568 htable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
569 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
570 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
571 ptable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
572 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
573 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
575 /* Malloc up a bunch of packets & buffers */
577 opr_queue_Init(&rx_freePacketQueue);
578 rxi_NeedMorePackets = FALSE;
579 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
581 /* enforce a minimum number of allocated packets */
582 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
583 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
585 /* allocate the initial free packet pool */
586 #ifdef RX_ENABLE_TSFPQ
587 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
588 #else /* RX_ENABLE_TSFPQ */
589 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
590 #endif /* RX_ENABLE_TSFPQ */
597 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
598 tv.tv_sec = clock_now.sec;
599 tv.tv_usec = clock_now.usec;
600 srand((unsigned int)tv.tv_usec);
607 #if defined(KERNEL) && !defined(UKERNEL)
608 /* Really, this should never happen in a real kernel */
611 struct sockaddr_in addr;
613 int addrlen = sizeof(addr);
615 socklen_t addrlen = sizeof(addr);
617 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
619 osi_Free(htable, rx_hashTableSize * sizeof(struct rx_connection *));
622 rx_port = addr.sin_port;
625 rx_stats.minRtt.sec = 9999999;
627 rx_SetEpoch(tv.tv_sec | 0x80000000);
629 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
630 * will provide a randomer value. */
632 MUTEX_ENTER(&rx_quota_mutex);
633 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
634 MUTEX_EXIT(&rx_quota_mutex);
635 /* *Slightly* random start time for the cid. This is just to help
636 * out with the hashing function at the peer */
637 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
638 rx_connHashTable = (struct rx_connection **)htable;
639 rx_peerHashTable = (struct rx_peer **)ptable;
641 rx_hardAckDelay.sec = 0;
642 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
644 rxevent_Init(20, rxi_ReScheduleEvents);
646 /* Initialize various global queues */
647 opr_queue_Init(&rx_idleServerQueue);
648 opr_queue_Init(&rx_incomingCallQueue);
649 opr_queue_Init(&rx_freeCallQueue);
651 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
652 /* Initialize our list of usable IP addresses. */
656 #if defined(RXK_LISTENER_ENV) || !defined(KERNEL)
657 /* Start listener process (exact function is dependent on the
658 * implementation environment--kernel or user space) */
663 tmp_status = rxinit_status = 0;
671 return rx_InitHost(htonl(INADDR_ANY), port);
677 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
678 * maintaing the round trip timer.
683 * Start a new RTT timer for a given call and packet.
685 * There must be no resendEvent already listed for this call, otherwise this
686 * will leak events - intended for internal use within the RTO code only
689 * the RX call to start the timer for
690 * @param[in] lastPacket
691 * a flag indicating whether the last packet has been sent or not
693 * @pre call must be locked before calling this function
697 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
699 struct clock now, retryTime;
704 clock_Add(&retryTime, &call->rto);
706 /* If we're sending the last packet, and we're the client, then the server
707 * may wait for an additional 400ms before returning the ACK, wait for it
708 * rather than hitting a timeout */
709 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
710 clock_Addmsec(&retryTime, 400);
712 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
713 call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
718 * Cancel an RTT timer for a given call.
722 * the RX call to cancel the timer for
724 * @pre call must be locked before calling this function
729 rxi_rto_cancel(struct rx_call *call)
731 rxevent_Cancel(&call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
735 * Tell the RTO timer that we have sent a packet.
737 * If the timer isn't already running, then start it. If the timer is running,
741 * the RX call that the packet has been sent on
742 * @param[in] lastPacket
743 * A flag which is true if this is the last packet for the call
745 * @pre The call must be locked before calling this function
750 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
752 if (call->resendEvent)
755 rxi_rto_startTimer(call, lastPacket, istack);
759 * Tell the RTO timer that we have received an new ACK message
761 * This function should be called whenever a call receives an ACK that
762 * acknowledges new packets. Whatever happens, we stop the current timer.
763 * If there are unacked packets in the queue which have been sent, then
764 * we restart the timer from now. Otherwise, we leave it stopped.
767 * the RX call that the ACK has been received on
771 rxi_rto_packet_acked(struct rx_call *call, int istack)
773 struct opr_queue *cursor;
775 rxi_rto_cancel(call);
777 if (opr_queue_IsEmpty(&call->tq))
780 for (opr_queue_Scan(&call->tq, cursor)) {
781 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
782 if (p->header.seq > call->tfirst + call->twind)
785 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
786 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
794 * Set an initial round trip timeout for a peer connection
796 * @param[in] secs The timeout to set in seconds
800 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
801 peer->rtt = secs * 8000;
805 * Enables or disables the busy call channel error (RX_CALL_BUSY).
807 * @param[in] onoff Non-zero to enable busy call channel errors.
809 * @pre Neither rx_Init nor rx_InitHost have been called yet
812 rx_SetBusyChannelError(afs_int32 onoff)
814 osi_Assert(rxinit_status != 0);
815 rxi_busyChannelError = onoff ? 1 : 0;
819 * Set a delayed ack event on the specified call for the given time
821 * @param[in] call - the call on which to set the event
822 * @param[in] offset - the delay from now after which the event fires
825 rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
827 struct clock now, when;
831 clock_Add(&when, offset);
833 if (!call->delayedAckEvent
834 || clock_Gt(&call->delayedAckTime, &when)) {
836 rxevent_Cancel(&call->delayedAckEvent, call,
837 RX_CALL_REFCOUNT_DELAY);
838 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
840 call->delayedAckEvent = rxevent_Post(&when, &now,
843 call->delayedAckTime = when;
847 /* called with unincremented nRequestsRunning to see if it is OK to start
848 * a new thread in this service. Could be "no" for two reasons: over the
849 * max quota, or would prevent others from reaching their min quota.
851 #ifdef RX_ENABLE_LOCKS
852 /* This verion of QuotaOK reserves quota if it's ok while the
853 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
856 QuotaOK(struct rx_service *aservice)
858 /* check if over max quota */
859 if (aservice->nRequestsRunning >= aservice->maxProcs) {
863 /* under min quota, we're OK */
864 /* otherwise, can use only if there are enough to allow everyone
865 * to go to their min quota after this guy starts.
868 MUTEX_ENTER(&rx_quota_mutex);
869 if ((aservice->nRequestsRunning < aservice->minProcs)
870 || (rxi_availProcs > rxi_minDeficit)) {
871 aservice->nRequestsRunning++;
872 /* just started call in minProcs pool, need fewer to maintain
874 if (aservice->nRequestsRunning <= aservice->minProcs)
877 MUTEX_EXIT(&rx_quota_mutex);
880 MUTEX_EXIT(&rx_quota_mutex);
886 ReturnToServerPool(struct rx_service *aservice)
888 aservice->nRequestsRunning--;
889 MUTEX_ENTER(&rx_quota_mutex);
890 if (aservice->nRequestsRunning < aservice->minProcs)
893 MUTEX_EXIT(&rx_quota_mutex);
896 #else /* RX_ENABLE_LOCKS */
898 QuotaOK(struct rx_service *aservice)
901 /* under min quota, we're OK */
902 if (aservice->nRequestsRunning < aservice->minProcs)
905 /* check if over max quota */
906 if (aservice->nRequestsRunning >= aservice->maxProcs)
909 /* otherwise, can use only if there are enough to allow everyone
910 * to go to their min quota after this guy starts.
912 MUTEX_ENTER(&rx_quota_mutex);
913 if (rxi_availProcs > rxi_minDeficit)
915 MUTEX_EXIT(&rx_quota_mutex);
918 #endif /* RX_ENABLE_LOCKS */
921 /* Called by rx_StartServer to start up lwp's to service calls.
922 NExistingProcs gives the number of procs already existing, and which
923 therefore needn't be created. */
925 rxi_StartServerProcs(int nExistingProcs)
927 struct rx_service *service;
932 /* For each service, reserve N processes, where N is the "minimum"
933 * number of processes that MUST be able to execute a request in parallel,
934 * at any time, for that process. Also compute the maximum difference
935 * between any service's maximum number of processes that can run
936 * (i.e. the maximum number that ever will be run, and a guarantee
937 * that this number will run if other services aren't running), and its
938 * minimum number. The result is the extra number of processes that
939 * we need in order to provide the latter guarantee */
940 for (i = 0; i < RX_MAX_SERVICES; i++) {
942 service = rx_services[i];
943 if (service == (struct rx_service *)0)
945 nProcs += service->minProcs;
946 diff = service->maxProcs - service->minProcs;
950 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
951 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
952 for (i = 0; i < nProcs; i++) {
953 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
959 /* This routine is only required on Windows */
961 rx_StartClientThread(void)
963 #ifdef AFS_PTHREAD_ENV
965 pid = pthread_self();
966 #endif /* AFS_PTHREAD_ENV */
968 #endif /* AFS_NT40_ENV */
970 /* This routine must be called if any services are exported. If the
971 * donateMe flag is set, the calling process is donated to the server
974 rx_StartServer(int donateMe)
976 struct rx_service *service;
982 /* Start server processes, if necessary (exact function is dependent
983 * on the implementation environment--kernel or user space). DonateMe
984 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
985 * case, one less new proc will be created rx_StartServerProcs.
987 rxi_StartServerProcs(donateMe);
989 /* count up the # of threads in minProcs, and add set the min deficit to
990 * be that value, too.
992 for (i = 0; i < RX_MAX_SERVICES; i++) {
993 service = rx_services[i];
994 if (service == (struct rx_service *)0)
996 MUTEX_ENTER(&rx_quota_mutex);
997 rxi_totalMin += service->minProcs;
998 /* below works even if a thread is running, since minDeficit would
999 * still have been decremented and later re-incremented.
1001 rxi_minDeficit += service->minProcs;
1002 MUTEX_EXIT(&rx_quota_mutex);
1005 /* Turn on reaping of idle server connections */
1006 rxi_ReapConnections(NULL, NULL, NULL, 0);
1011 #ifndef AFS_NT40_ENV
1015 #ifdef AFS_PTHREAD_ENV
1017 pid = afs_pointer_to_int(pthread_self());
1018 #else /* AFS_PTHREAD_ENV */
1020 LWP_CurrentProcess(&pid);
1021 #endif /* AFS_PTHREAD_ENV */
1023 sprintf(name, "srv_%d", ++nProcs);
1024 if (registerProgram)
1025 (*registerProgram) (pid, name);
1027 #endif /* AFS_NT40_ENV */
1028 rx_ServerProc(NULL); /* Never returns */
1030 #ifdef RX_ENABLE_TSFPQ
1031 /* no use leaving packets around in this thread's local queue if
1032 * it isn't getting donated to the server thread pool.
1034 rxi_FlushLocalPacketsTSFPQ();
1035 #endif /* RX_ENABLE_TSFPQ */
1039 /* Create a new client connection to the specified service, using the
1040 * specified security object to implement the security model for this
1042 struct rx_connection *
1043 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
1044 struct rx_securityClass *securityObject,
1045 int serviceSecurityIndex)
1049 struct rx_connection *conn;
1054 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
1055 "serviceSecurityIndex %d)\n",
1056 ntohl(shost), ntohs(sport), sservice, securityObject,
1057 serviceSecurityIndex));
1059 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
1060 * the case of kmem_alloc? */
1061 conn = rxi_AllocConnection();
1062 #ifdef RX_ENABLE_LOCKS
1063 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
1064 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
1065 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
1068 MUTEX_ENTER(&rx_connHashTable_lock);
1069 cid = (rx_nextCid += RX_MAXCALLS);
1070 conn->type = RX_CLIENT_CONNECTION;
1072 conn->epoch = rx_epoch;
1073 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
1074 conn->serviceId = sservice;
1075 conn->securityObject = securityObject;
1076 conn->securityData = (void *) 0;
1077 conn->securityIndex = serviceSecurityIndex;
1078 rx_SetConnDeadTime(conn, rx_connDeadTime);
1079 rx_SetConnSecondsUntilNatPing(conn, 0);
1080 conn->ackRate = RX_FAST_ACK_RATE;
1081 conn->nSpecific = 0;
1082 conn->specific = NULL;
1083 conn->challengeEvent = NULL;
1084 conn->delayedAbortEvent = NULL;
1085 conn->abortCount = 0;
1087 for (i = 0; i < RX_MAXCALLS; i++) {
1088 conn->twind[i] = rx_initSendWindow;
1089 conn->rwind[i] = rx_initReceiveWindow;
1090 conn->lastBusy[i] = 0;
1093 RXS_NewConnection(securityObject, conn);
1095 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1097 conn->refCount++; /* no lock required since only this thread knows... */
1098 conn->next = rx_connHashTable[hashindex];
1099 rx_connHashTable[hashindex] = conn;
1100 if (rx_stats_active)
1101 rx_atomic_inc(&rx_stats.nClientConns);
1102 MUTEX_EXIT(&rx_connHashTable_lock);
1108 * Ensure a connection's timeout values are valid.
1110 * @param[in] conn The connection to check
1112 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1113 * unless idleDeadTime and/or hardDeadTime are not set
1117 rxi_CheckConnTimeouts(struct rx_connection *conn)
1119 /* a connection's timeouts must have the relationship
1120 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1121 * total loss of network to a peer may cause an idle timeout instead of a
1122 * dead timeout, simply because the idle timeout gets hit first. Also set
1123 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1124 /* this logic is slightly complicated by the fact that
1125 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1127 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1128 if (conn->idleDeadTime) {
1129 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1131 if (conn->hardDeadTime) {
1132 if (conn->idleDeadTime) {
1133 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1135 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1141 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1143 /* The idea is to set the dead time to a value that allows several
1144 * keepalives to be dropped without timing out the connection. */
1145 conn->secondsUntilDead = seconds;
1146 rxi_CheckConnTimeouts(conn);
1147 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1151 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1153 conn->hardDeadTime = seconds;
1154 rxi_CheckConnTimeouts(conn);
1158 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1160 conn->idleDeadTime = seconds;
1161 conn->idleDeadDetection = (seconds ? 1 : 0);
1162 rxi_CheckConnTimeouts(conn);
1165 int rxi_lowPeerRefCount = 0;
1166 int rxi_lowConnRefCount = 0;
1169 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1170 * NOTE: must not be called with rx_connHashTable_lock held.
1173 rxi_CleanupConnection(struct rx_connection *conn)
1175 /* Notify the service exporter, if requested, that this connection
1176 * is being destroyed */
1177 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1178 (*conn->service->destroyConnProc) (conn);
1180 /* Notify the security module that this connection is being destroyed */
1181 RXS_DestroyConnection(conn->securityObject, conn);
1183 /* If this is the last connection using the rx_peer struct, set its
1184 * idle time to now. rxi_ReapConnections will reap it if it's still
1185 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1187 MUTEX_ENTER(&rx_peerHashTable_lock);
1188 if (conn->peer->refCount < 2) {
1189 conn->peer->idleWhen = clock_Sec();
1190 if (conn->peer->refCount < 1) {
1191 conn->peer->refCount = 1;
1192 if (rx_stats_active) {
1193 MUTEX_ENTER(&rx_stats_mutex);
1194 rxi_lowPeerRefCount++;
1195 MUTEX_EXIT(&rx_stats_mutex);
1199 conn->peer->refCount--;
1200 MUTEX_EXIT(&rx_peerHashTable_lock);
1202 if (rx_stats_active)
1204 if (conn->type == RX_SERVER_CONNECTION)
1205 rx_atomic_dec(&rx_stats.nServerConns);
1207 rx_atomic_dec(&rx_stats.nClientConns);
1210 if (conn->specific) {
1212 for (i = 0; i < conn->nSpecific; i++) {
1213 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1214 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1215 conn->specific[i] = NULL;
1217 free(conn->specific);
1219 conn->specific = NULL;
1220 conn->nSpecific = 0;
1221 #endif /* !KERNEL */
1223 MUTEX_DESTROY(&conn->conn_call_lock);
1224 MUTEX_DESTROY(&conn->conn_data_lock);
1225 CV_DESTROY(&conn->conn_call_cv);
1227 rxi_FreeConnection(conn);
1230 /* Destroy the specified connection */
1232 rxi_DestroyConnection(struct rx_connection *conn)
1234 MUTEX_ENTER(&rx_connHashTable_lock);
1235 rxi_DestroyConnectionNoLock(conn);
1236 /* conn should be at the head of the cleanup list */
1237 if (conn == rx_connCleanup_list) {
1238 rx_connCleanup_list = rx_connCleanup_list->next;
1239 MUTEX_EXIT(&rx_connHashTable_lock);
1240 rxi_CleanupConnection(conn);
1242 #ifdef RX_ENABLE_LOCKS
1244 MUTEX_EXIT(&rx_connHashTable_lock);
1246 #endif /* RX_ENABLE_LOCKS */
1250 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1252 struct rx_connection **conn_ptr;
1254 struct rx_packet *packet;
1261 MUTEX_ENTER(&conn->conn_data_lock);
1262 MUTEX_ENTER(&rx_refcnt_mutex);
1263 if (conn->refCount > 0)
1266 if (rx_stats_active) {
1267 MUTEX_ENTER(&rx_stats_mutex);
1268 rxi_lowConnRefCount++;
1269 MUTEX_EXIT(&rx_stats_mutex);
1273 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1274 /* Busy; wait till the last guy before proceeding */
1275 MUTEX_EXIT(&rx_refcnt_mutex);
1276 MUTEX_EXIT(&conn->conn_data_lock);
1281 /* If the client previously called rx_NewCall, but it is still
1282 * waiting, treat this as a running call, and wait to destroy the
1283 * connection later when the call completes. */
1284 if ((conn->type == RX_CLIENT_CONNECTION)
1285 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1286 conn->flags |= RX_CONN_DESTROY_ME;
1287 MUTEX_EXIT(&conn->conn_data_lock);
1291 MUTEX_EXIT(&rx_refcnt_mutex);
1292 MUTEX_EXIT(&conn->conn_data_lock);
1294 /* Check for extant references to this connection */
1295 MUTEX_ENTER(&conn->conn_call_lock);
1296 for (i = 0; i < RX_MAXCALLS; i++) {
1297 struct rx_call *call = conn->call[i];
1300 if (conn->type == RX_CLIENT_CONNECTION) {
1301 MUTEX_ENTER(&call->lock);
1302 if (call->delayedAckEvent) {
1303 /* Push the final acknowledgment out now--there
1304 * won't be a subsequent call to acknowledge the
1305 * last reply packets */
1306 rxevent_Cancel(&call->delayedAckEvent, call,
1307 RX_CALL_REFCOUNT_DELAY);
1308 if (call->state == RX_STATE_PRECALL
1309 || call->state == RX_STATE_ACTIVE) {
1310 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1315 MUTEX_EXIT(&call->lock);
1319 MUTEX_EXIT(&conn->conn_call_lock);
1321 #ifdef RX_ENABLE_LOCKS
1323 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1324 MUTEX_EXIT(&conn->conn_data_lock);
1326 /* Someone is accessing a packet right now. */
1330 #endif /* RX_ENABLE_LOCKS */
1333 /* Don't destroy the connection if there are any call
1334 * structures still in use */
1335 MUTEX_ENTER(&conn->conn_data_lock);
1336 conn->flags |= RX_CONN_DESTROY_ME;
1337 MUTEX_EXIT(&conn->conn_data_lock);
1342 if (conn->natKeepAliveEvent) {
1343 rxi_NatKeepAliveOff(conn);
1346 if (conn->delayedAbortEvent) {
1347 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
1348 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1350 MUTEX_ENTER(&conn->conn_data_lock);
1351 rxi_SendConnectionAbort(conn, packet, 0, 1);
1352 MUTEX_EXIT(&conn->conn_data_lock);
1353 rxi_FreePacket(packet);
1357 /* Remove from connection hash table before proceeding */
1359 &rx_connHashTable[CONN_HASH
1360 (peer->host, peer->port, conn->cid, conn->epoch,
1362 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1363 if (*conn_ptr == conn) {
1364 *conn_ptr = conn->next;
1368 /* if the conn that we are destroying was the last connection, then we
1369 * clear rxLastConn as well */
1370 if (rxLastConn == conn)
1373 /* Make sure the connection is completely reset before deleting it. */
1374 /* get rid of pending events that could zap us later */
1375 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
1376 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
1377 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
1379 /* Add the connection to the list of destroyed connections that
1380 * need to be cleaned up. This is necessary to avoid deadlocks
1381 * in the routines we call to inform others that this connection is
1382 * being destroyed. */
1383 conn->next = rx_connCleanup_list;
1384 rx_connCleanup_list = conn;
1387 /* Externally available version */
1389 rx_DestroyConnection(struct rx_connection *conn)
1394 rxi_DestroyConnection(conn);
1399 rx_GetConnection(struct rx_connection *conn)
1404 MUTEX_ENTER(&rx_refcnt_mutex);
1406 MUTEX_EXIT(&rx_refcnt_mutex);
1410 #ifdef RX_ENABLE_LOCKS
1411 /* Wait for the transmit queue to no longer be busy.
1412 * requires the call->lock to be held */
1414 rxi_WaitforTQBusy(struct rx_call *call) {
1415 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1416 call->flags |= RX_CALL_TQ_WAIT;
1418 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1419 CV_WAIT(&call->cv_tq, &call->lock);
1421 if (call->tqWaiters == 0) {
1422 call->flags &= ~RX_CALL_TQ_WAIT;
1429 rxi_WakeUpTransmitQueue(struct rx_call *call)
1431 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1432 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1433 call, call->tqWaiters, call->flags));
1434 #ifdef RX_ENABLE_LOCKS
1435 osirx_AssertMine(&call->lock, "rxi_Start start");
1436 CV_BROADCAST(&call->cv_tq);
1437 #else /* RX_ENABLE_LOCKS */
1438 osi_rxWakeup(&call->tq);
1439 #endif /* RX_ENABLE_LOCKS */
1443 /* Start a new rx remote procedure call, on the specified connection.
1444 * If wait is set to 1, wait for a free call channel; otherwise return
1445 * 0. Maxtime gives the maximum number of seconds this call may take,
1446 * after rx_NewCall returns. After this time interval, a call to any
1447 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1448 * For fine grain locking, we hold the conn_call_lock in order to
1449 * to ensure that we don't get signalle after we found a call in an active
1450 * state and before we go to sleep.
1453 rx_NewCall(struct rx_connection *conn)
1455 int i, wait, ignoreBusy = 1;
1456 struct rx_call *call;
1457 struct clock queueTime;
1458 afs_uint32 leastBusy = 0;
1462 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1465 clock_GetTime(&queueTime);
1467 * Check if there are others waiting for a new call.
1468 * If so, let them go first to avoid starving them.
1469 * This is a fairly simple scheme, and might not be
1470 * a complete solution for large numbers of waiters.
1472 * makeCallWaiters keeps track of the number of
1473 * threads waiting to make calls and the
1474 * RX_CONN_MAKECALL_WAITING flag bit is used to
1475 * indicate that there are indeed calls waiting.
1476 * The flag is set when the waiter is incremented.
1477 * It is only cleared when makeCallWaiters is 0.
1478 * This prevents us from accidently destroying the
1479 * connection while it is potentially about to be used.
1481 MUTEX_ENTER(&conn->conn_call_lock);
1482 MUTEX_ENTER(&conn->conn_data_lock);
1483 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1484 conn->flags |= RX_CONN_MAKECALL_WAITING;
1485 conn->makeCallWaiters++;
1486 MUTEX_EXIT(&conn->conn_data_lock);
1488 #ifdef RX_ENABLE_LOCKS
1489 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1493 MUTEX_ENTER(&conn->conn_data_lock);
1494 conn->makeCallWaiters--;
1495 if (conn->makeCallWaiters == 0)
1496 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1499 /* We are now the active thread in rx_NewCall */
1500 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1501 MUTEX_EXIT(&conn->conn_data_lock);
1506 for (i = 0; i < RX_MAXCALLS; i++) {
1507 call = conn->call[i];
1509 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1510 /* we're not ignoring busy call slots; only look at the
1511 * call slot that is the "least" busy */
1515 if (call->state == RX_STATE_DALLY) {
1516 MUTEX_ENTER(&call->lock);
1517 if (call->state == RX_STATE_DALLY) {
1518 if (ignoreBusy && conn->lastBusy[i]) {
1519 /* if we're ignoring busy call slots, skip any ones that
1520 * have lastBusy set */
1521 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1522 leastBusy = conn->lastBusy[i];
1524 MUTEX_EXIT(&call->lock);
1529 * We are setting the state to RX_STATE_RESET to
1530 * ensure that no one else will attempt to use this
1531 * call once we drop the conn->conn_call_lock and
1532 * call->lock. We must drop the conn->conn_call_lock
1533 * before calling rxi_ResetCall because the process
1534 * of clearing the transmit queue can block for an
1535 * extended period of time. If we block while holding
1536 * the conn->conn_call_lock, then all rx_EndCall
1537 * processing will block as well. This has a detrimental
1538 * effect on overall system performance.
1540 call->state = RX_STATE_RESET;
1541 (*call->callNumber)++;
1542 MUTEX_EXIT(&conn->conn_call_lock);
1543 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1544 rxi_ResetCall(call, 0);
1545 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1549 * If we failed to be able to safely obtain the
1550 * conn->conn_call_lock we will have to drop the
1551 * call->lock to avoid a deadlock. When the call->lock
1552 * is released the state of the call can change. If it
1553 * is no longer RX_STATE_RESET then some other thread is
1556 MUTEX_EXIT(&call->lock);
1557 MUTEX_ENTER(&conn->conn_call_lock);
1558 MUTEX_ENTER(&call->lock);
1560 if (call->state == RX_STATE_RESET)
1564 * If we get here it means that after dropping
1565 * the conn->conn_call_lock and call->lock that
1566 * the call is no longer ours. If we can't find
1567 * a free call in the remaining slots we should
1568 * not go immediately to RX_CONN_MAKECALL_WAITING
1569 * because by dropping the conn->conn_call_lock
1570 * we have given up synchronization with rx_EndCall.
1571 * Instead, cycle through one more time to see if
1572 * we can find a call that can call our own.
1574 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1577 MUTEX_EXIT(&call->lock);
1580 if (ignoreBusy && conn->lastBusy[i]) {
1581 /* if we're ignoring busy call slots, skip any ones that
1582 * have lastBusy set */
1583 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1584 leastBusy = conn->lastBusy[i];
1589 /* rxi_NewCall returns with mutex locked */
1590 call = rxi_NewCall(conn, i);
1591 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1595 if (i < RX_MAXCALLS) {
1596 conn->lastBusy[i] = 0;
1597 call->flags &= ~RX_CALL_PEER_BUSY;
1602 if (leastBusy && ignoreBusy) {
1603 /* we didn't find a useable call slot, but we did see at least one
1604 * 'busy' slot; look again and only use a slot with the 'least
1610 MUTEX_ENTER(&conn->conn_data_lock);
1611 conn->flags |= RX_CONN_MAKECALL_WAITING;
1612 conn->makeCallWaiters++;
1613 MUTEX_EXIT(&conn->conn_data_lock);
1615 #ifdef RX_ENABLE_LOCKS
1616 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1620 MUTEX_ENTER(&conn->conn_data_lock);
1621 conn->makeCallWaiters--;
1622 if (conn->makeCallWaiters == 0)
1623 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1624 MUTEX_EXIT(&conn->conn_data_lock);
1626 /* Client is initially in send mode */
1627 call->state = RX_STATE_ACTIVE;
1628 call->error = conn->error;
1630 call->app.mode = RX_MODE_ERROR;
1632 call->app.mode = RX_MODE_SENDING;
1634 #ifdef AFS_RXERRQ_ENV
1635 /* remember how many network errors the peer has when we started, so if
1636 * more errors are encountered after the call starts, we know the other endpoint won't be
1637 * responding to us */
1638 call->neterr_gen = rx_atomic_read(&conn->peer->neterrs);
1641 /* remember start time for call in case we have hard dead time limit */
1642 call->queueTime = queueTime;
1643 clock_GetTime(&call->startTime);
1644 call->app.bytesSent = 0;
1645 call->app.bytesRcvd = 0;
1647 /* Turn on busy protocol. */
1648 rxi_KeepAliveOn(call);
1650 /* Attempt MTU discovery */
1651 rxi_GrowMTUOn(call);
1654 * We are no longer the active thread in rx_NewCall
1656 MUTEX_ENTER(&conn->conn_data_lock);
1657 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1658 MUTEX_EXIT(&conn->conn_data_lock);
1661 * Wake up anyone else who might be giving us a chance to
1662 * run (see code above that avoids resource starvation).
1664 #ifdef RX_ENABLE_LOCKS
1665 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1666 osi_Panic("rx_NewCall call about to be used without an empty tq");
1669 CV_BROADCAST(&conn->conn_call_cv);
1673 MUTEX_EXIT(&conn->conn_call_lock);
1674 MUTEX_EXIT(&call->lock);
1677 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1682 rxi_HasActiveCalls(struct rx_connection *aconn)
1685 struct rx_call *tcall;
1689 for (i = 0; i < RX_MAXCALLS; i++) {
1690 if ((tcall = aconn->call[i])) {
1691 if ((tcall->state == RX_STATE_ACTIVE)
1692 || (tcall->state == RX_STATE_PRECALL)) {
1703 rxi_GetCallNumberVector(struct rx_connection *aconn,
1704 afs_int32 * aint32s)
1707 struct rx_call *tcall;
1711 MUTEX_ENTER(&aconn->conn_call_lock);
1712 for (i = 0; i < RX_MAXCALLS; i++) {
1713 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1714 aint32s[i] = aconn->callNumber[i] + 1;
1716 aint32s[i] = aconn->callNumber[i];
1718 MUTEX_EXIT(&aconn->conn_call_lock);
1724 rxi_SetCallNumberVector(struct rx_connection *aconn,
1725 afs_int32 * aint32s)
1728 struct rx_call *tcall;
1732 MUTEX_ENTER(&aconn->conn_call_lock);
1733 for (i = 0; i < RX_MAXCALLS; i++) {
1734 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1735 aconn->callNumber[i] = aint32s[i] - 1;
1737 aconn->callNumber[i] = aint32s[i];
1739 MUTEX_EXIT(&aconn->conn_call_lock);
1744 /* Advertise a new service. A service is named locally by a UDP port
1745 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1748 char *serviceName; Name for identification purposes (e.g. the
1749 service name might be used for probing for
1752 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1753 char *serviceName, struct rx_securityClass **securityObjects,
1754 int nSecurityObjects,
1755 afs_int32(*serviceProc) (struct rx_call * acall))
1757 osi_socket socket = OSI_NULLSOCKET;
1758 struct rx_service *tservice;
1764 if (serviceId == 0) {
1766 "rx_NewService: service id for service %s is not non-zero.\n",
1773 "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",
1781 tservice = rxi_AllocService();
1784 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1786 for (i = 0; i < RX_MAX_SERVICES; i++) {
1787 struct rx_service *service = rx_services[i];
1789 if (port == service->servicePort && host == service->serviceHost) {
1790 if (service->serviceId == serviceId) {
1791 /* The identical service has already been
1792 * installed; if the caller was intending to
1793 * change the security classes used by this
1794 * service, he/she loses. */
1796 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1797 serviceName, serviceId, service->serviceName);
1799 rxi_FreeService(tservice);
1802 /* Different service, same port: re-use the socket
1803 * which is bound to the same port */
1804 socket = service->socket;
1807 if (socket == OSI_NULLSOCKET) {
1808 /* If we don't already have a socket (from another
1809 * service on same port) get a new one */
1810 socket = rxi_GetHostUDPSocket(host, port);
1811 if (socket == OSI_NULLSOCKET) {
1813 rxi_FreeService(tservice);
1818 service->socket = socket;
1819 service->serviceHost = host;
1820 service->servicePort = port;
1821 service->serviceId = serviceId;
1822 service->serviceName = serviceName;
1823 service->nSecurityObjects = nSecurityObjects;
1824 service->securityObjects = securityObjects;
1825 service->minProcs = 0;
1826 service->maxProcs = 1;
1827 service->idleDeadTime = 60;
1828 service->idleDeadErr = 0;
1829 service->connDeadTime = rx_connDeadTime;
1830 service->executeRequestProc = serviceProc;
1831 service->checkReach = 0;
1832 service->nSpecific = 0;
1833 service->specific = NULL;
1834 rx_services[i] = service; /* not visible until now */
1840 rxi_FreeService(tservice);
1841 (osi_Msg "rx_NewService: cannot support > %d services\n",
1846 /* Set configuration options for all of a service's security objects */
1849 rx_SetSecurityConfiguration(struct rx_service *service,
1850 rx_securityConfigVariables type,
1854 for (i = 0; i<service->nSecurityObjects; i++) {
1855 if (service->securityObjects[i]) {
1856 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1864 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1865 struct rx_securityClass **securityObjects, int nSecurityObjects,
1866 afs_int32(*serviceProc) (struct rx_call * acall))
1868 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1871 /* Generic request processing loop. This routine should be called
1872 * by the implementation dependent rx_ServerProc. If socketp is
1873 * non-null, it will be set to the file descriptor that this thread
1874 * is now listening on. If socketp is null, this routine will never
1877 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1879 struct rx_call *call;
1881 struct rx_service *tservice = NULL;
1888 call = rx_GetCall(threadID, tservice, socketp);
1889 if (socketp && *socketp != OSI_NULLSOCKET) {
1890 /* We are now a listener thread */
1896 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1897 #ifdef RX_ENABLE_LOCKS
1899 #endif /* RX_ENABLE_LOCKS */
1900 afs_termState = AFSOP_STOP_AFS;
1901 afs_osi_Wakeup(&afs_termState);
1902 #ifdef RX_ENABLE_LOCKS
1904 #endif /* RX_ENABLE_LOCKS */
1909 /* if server is restarting( typically smooth shutdown) then do not
1910 * allow any new calls.
1913 if (rx_tranquil && (call != NULL)) {
1917 MUTEX_ENTER(&call->lock);
1919 rxi_CallError(call, RX_RESTARTING);
1920 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1922 MUTEX_EXIT(&call->lock);
1927 tservice = call->conn->service;
1929 if (tservice->beforeProc)
1930 (*tservice->beforeProc) (call);
1932 code = tservice->executeRequestProc(call);
1934 if (tservice->afterProc)
1935 (*tservice->afterProc) (call, code);
1937 rx_EndCall(call, code);
1939 if (tservice->postProc)
1940 (*tservice->postProc) (code);
1942 if (rx_stats_active) {
1943 MUTEX_ENTER(&rx_stats_mutex);
1945 MUTEX_EXIT(&rx_stats_mutex);
1952 rx_WakeupServerProcs(void)
1954 struct rx_serverQueueEntry *np, *tqp;
1955 struct opr_queue *cursor;
1959 MUTEX_ENTER(&rx_serverPool_lock);
1961 #ifdef RX_ENABLE_LOCKS
1962 if (rx_waitForPacket)
1963 CV_BROADCAST(&rx_waitForPacket->cv);
1964 #else /* RX_ENABLE_LOCKS */
1965 if (rx_waitForPacket)
1966 osi_rxWakeup(rx_waitForPacket);
1967 #endif /* RX_ENABLE_LOCKS */
1968 MUTEX_ENTER(&freeSQEList_lock);
1969 for (np = rx_FreeSQEList; np; np = tqp) {
1970 tqp = *(struct rx_serverQueueEntry **)np;
1971 #ifdef RX_ENABLE_LOCKS
1972 CV_BROADCAST(&np->cv);
1973 #else /* RX_ENABLE_LOCKS */
1975 #endif /* RX_ENABLE_LOCKS */
1977 MUTEX_EXIT(&freeSQEList_lock);
1978 for (opr_queue_Scan(&rx_idleServerQueue, cursor)) {
1979 np = opr_queue_Entry(cursor, struct rx_serverQueueEntry, entry);
1980 #ifdef RX_ENABLE_LOCKS
1981 CV_BROADCAST(&np->cv);
1982 #else /* RX_ENABLE_LOCKS */
1984 #endif /* RX_ENABLE_LOCKS */
1986 MUTEX_EXIT(&rx_serverPool_lock);
1991 * One thing that seems to happen is that all the server threads get
1992 * tied up on some empty or slow call, and then a whole bunch of calls
1993 * arrive at once, using up the packet pool, so now there are more
1994 * empty calls. The most critical resources here are server threads
1995 * and the free packet pool. The "doreclaim" code seems to help in
1996 * general. I think that eventually we arrive in this state: there
1997 * are lots of pending calls which do have all their packets present,
1998 * so they won't be reclaimed, are multi-packet calls, so they won't
1999 * be scheduled until later, and thus are tying up most of the free
2000 * packet pool for a very long time.
2002 * 1. schedule multi-packet calls if all the packets are present.
2003 * Probably CPU-bound operation, useful to return packets to pool.
2004 * Do what if there is a full window, but the last packet isn't here?
2005 * 3. preserve one thread which *only* runs "best" calls, otherwise
2006 * it sleeps and waits for that type of call.
2007 * 4. Don't necessarily reserve a whole window for each thread. In fact,
2008 * the current dataquota business is badly broken. The quota isn't adjusted
2009 * to reflect how many packets are presently queued for a running call.
2010 * So, when we schedule a queued call with a full window of packets queued
2011 * up for it, that *should* free up a window full of packets for other 2d-class
2012 * calls to be able to use from the packet pool. But it doesn't.
2014 * NB. Most of the time, this code doesn't run -- since idle server threads
2015 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
2016 * as a new call arrives.
2018 /* Sleep until a call arrives. Returns a pointer to the call, ready
2019 * for an rx_Read. */
2020 #ifdef RX_ENABLE_LOCKS
2022 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2024 struct rx_serverQueueEntry *sq;
2025 struct rx_call *call = (struct rx_call *)0;
2026 struct rx_service *service = NULL;
2028 MUTEX_ENTER(&freeSQEList_lock);
2030 if ((sq = rx_FreeSQEList)) {
2031 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2032 MUTEX_EXIT(&freeSQEList_lock);
2033 } else { /* otherwise allocate a new one and return that */
2034 MUTEX_EXIT(&freeSQEList_lock);
2035 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2036 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2037 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2040 MUTEX_ENTER(&rx_serverPool_lock);
2041 if (cur_service != NULL) {
2042 ReturnToServerPool(cur_service);
2045 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2046 struct rx_call *tcall, *choice2 = NULL;
2047 struct opr_queue *cursor;
2049 /* Scan for eligible incoming calls. A call is not eligible
2050 * if the maximum number of calls for its service type are
2051 * already executing */
2052 /* One thread will process calls FCFS (to prevent starvation),
2053 * while the other threads may run ahead looking for calls which
2054 * have all their input data available immediately. This helps
2055 * keep threads from blocking, waiting for data from the client. */
2056 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2057 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2059 service = tcall->conn->service;
2060 if (!QuotaOK(service)) {
2063 MUTEX_ENTER(&rx_pthread_mutex);
2064 if (tno == rxi_fcfs_thread_num
2065 || opr_queue_IsEnd(&rx_incomingCallQueue, cursor)) {
2066 MUTEX_EXIT(&rx_pthread_mutex);
2067 /* If we're the fcfs thread , then we'll just use
2068 * this call. If we haven't been able to find an optimal
2069 * choice, and we're at the end of the list, then use a
2070 * 2d choice if one has been identified. Otherwise... */
2071 call = (choice2 ? choice2 : tcall);
2072 service = call->conn->service;
2074 MUTEX_EXIT(&rx_pthread_mutex);
2075 if (!opr_queue_IsEmpty(&tcall->rq)) {
2076 struct rx_packet *rp;
2077 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2079 if (rp->header.seq == 1) {
2081 || (rp->header.flags & RX_LAST_PACKET)) {
2083 } else if (rxi_2dchoice && !choice2
2084 && !(tcall->flags & RX_CALL_CLEARED)
2085 && (tcall->rprev > rxi_HardAckRate)) {
2095 ReturnToServerPool(service);
2101 opr_queue_Remove(&call->entry);
2102 MUTEX_EXIT(&rx_serverPool_lock);
2103 MUTEX_ENTER(&call->lock);
2105 if (call->flags & RX_CALL_WAIT_PROC) {
2106 call->flags &= ~RX_CALL_WAIT_PROC;
2107 rx_atomic_dec(&rx_nWaiting);
2110 if (call->state != RX_STATE_PRECALL || call->error) {
2111 MUTEX_EXIT(&call->lock);
2112 MUTEX_ENTER(&rx_serverPool_lock);
2113 ReturnToServerPool(service);
2118 if (opr_queue_IsEmpty(&call->rq)
2119 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1)
2120 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2122 CLEAR_CALL_QUEUE_LOCK(call);
2125 /* If there are no eligible incoming calls, add this process
2126 * to the idle server queue, to wait for one */
2130 *socketp = OSI_NULLSOCKET;
2132 sq->socketp = socketp;
2133 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2134 #ifndef AFS_AIX41_ENV
2135 rx_waitForPacket = sq;
2137 rx_waitingForPacket = sq;
2138 #endif /* AFS_AIX41_ENV */
2140 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2142 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2143 MUTEX_EXIT(&rx_serverPool_lock);
2144 return (struct rx_call *)0;
2147 } while (!(call = sq->newcall)
2148 && !(socketp && *socketp != OSI_NULLSOCKET));
2149 MUTEX_EXIT(&rx_serverPool_lock);
2151 MUTEX_ENTER(&call->lock);
2157 MUTEX_ENTER(&freeSQEList_lock);
2158 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2159 rx_FreeSQEList = sq;
2160 MUTEX_EXIT(&freeSQEList_lock);
2163 clock_GetTime(&call->startTime);
2164 call->state = RX_STATE_ACTIVE;
2165 call->app.mode = RX_MODE_RECEIVING;
2166 #ifdef RX_KERNEL_TRACE
2167 if (ICL_SETACTIVE(afs_iclSetp)) {
2168 int glockOwner = ISAFS_GLOCK();
2171 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2172 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2179 rxi_calltrace(RX_CALL_START, call);
2180 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2181 call->conn->service->servicePort, call->conn->service->serviceId,
2184 MUTEX_EXIT(&call->lock);
2185 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2187 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2192 #else /* RX_ENABLE_LOCKS */
2194 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2196 struct rx_serverQueueEntry *sq;
2197 struct rx_call *call = (struct rx_call *)0, *choice2;
2198 struct rx_service *service = NULL;
2202 MUTEX_ENTER(&freeSQEList_lock);
2204 if ((sq = rx_FreeSQEList)) {
2205 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2206 MUTEX_EXIT(&freeSQEList_lock);
2207 } else { /* otherwise allocate a new one and return that */
2208 MUTEX_EXIT(&freeSQEList_lock);
2209 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2210 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2211 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2213 MUTEX_ENTER(&sq->lock);
2215 if (cur_service != NULL) {
2216 cur_service->nRequestsRunning--;
2217 MUTEX_ENTER(&rx_quota_mutex);
2218 if (cur_service->nRequestsRunning < cur_service->minProcs)
2221 MUTEX_EXIT(&rx_quota_mutex);
2223 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2224 struct rx_call *tcall;
2225 struct opr_queue *cursor;
2226 /* Scan for eligible incoming calls. A call is not eligible
2227 * if the maximum number of calls for its service type are
2228 * already executing */
2229 /* One thread will process calls FCFS (to prevent starvation),
2230 * while the other threads may run ahead looking for calls which
2231 * have all their input data available immediately. This helps
2232 * keep threads from blocking, waiting for data from the client. */
2233 choice2 = (struct rx_call *)0;
2234 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2235 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2236 service = tcall->conn->service;
2237 if (QuotaOK(service)) {
2238 MUTEX_ENTER(&rx_pthread_mutex);
2239 /* XXX - If tcall->entry.next is NULL, then we're no longer
2240 * on a queue at all. This shouldn't happen. */
2241 if (tno == rxi_fcfs_thread_num || !tcall->entry.next) {
2242 MUTEX_EXIT(&rx_pthread_mutex);
2243 /* If we're the fcfs thread, then we'll just use
2244 * this call. If we haven't been able to find an optimal
2245 * choice, and we're at the end of the list, then use a
2246 * 2d choice if one has been identified. Otherwise... */
2247 call = (choice2 ? choice2 : tcall);
2248 service = call->conn->service;
2250 MUTEX_EXIT(&rx_pthread_mutex);
2251 if (!opr_queue_IsEmpty(&tcall->rq)) {
2252 struct rx_packet *rp;
2253 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2255 if (rp->header.seq == 1
2257 || (rp->header.flags & RX_LAST_PACKET))) {
2259 } else if (rxi_2dchoice && !choice2
2260 && !(tcall->flags & RX_CALL_CLEARED)
2261 && (tcall->rprev > rxi_HardAckRate)) {
2274 opr_queue_Remove(&call->entry);
2275 /* we can't schedule a call if there's no data!!! */
2276 /* send an ack if there's no data, if we're missing the
2277 * first packet, or we're missing something between first
2278 * and last -- there's a "hole" in the incoming data. */
2279 if (opr_queue_IsEmpty(&call->rq)
2280 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1
2281 || call->rprev != opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq)
2282 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2284 call->flags &= (~RX_CALL_WAIT_PROC);
2285 service->nRequestsRunning++;
2286 /* just started call in minProcs pool, need fewer to maintain
2288 MUTEX_ENTER(&rx_quota_mutex);
2289 if (service->nRequestsRunning <= service->minProcs)
2292 MUTEX_EXIT(&rx_quota_mutex);
2293 rx_atomic_dec(&rx_nWaiting);
2294 /* MUTEX_EXIT(&call->lock); */
2296 /* If there are no eligible incoming calls, add this process
2297 * to the idle server queue, to wait for one */
2300 *socketp = OSI_NULLSOCKET;
2302 sq->socketp = socketp;
2303 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2307 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2309 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2310 return (struct rx_call *)0;
2313 } while (!(call = sq->newcall)
2314 && !(socketp && *socketp != OSI_NULLSOCKET));
2316 MUTEX_EXIT(&sq->lock);
2318 MUTEX_ENTER(&freeSQEList_lock);
2319 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2320 rx_FreeSQEList = sq;
2321 MUTEX_EXIT(&freeSQEList_lock);
2324 clock_GetTime(&call->startTime);
2325 call->state = RX_STATE_ACTIVE;
2326 call->app.mode = RX_MODE_RECEIVING;
2327 #ifdef RX_KERNEL_TRACE
2328 if (ICL_SETACTIVE(afs_iclSetp)) {
2329 int glockOwner = ISAFS_GLOCK();
2332 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2333 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2340 rxi_calltrace(RX_CALL_START, call);
2341 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2342 call->conn->service->servicePort, call->conn->service->serviceId,
2345 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2352 #endif /* RX_ENABLE_LOCKS */
2356 /* Establish a procedure to be called when a packet arrives for a
2357 * call. This routine will be called at most once after each call,
2358 * and will also be called if there is an error condition on the or
2359 * the call is complete. Used by multi rx to build a selection
2360 * function which determines which of several calls is likely to be a
2361 * good one to read from.
2362 * NOTE: the way this is currently implemented it is probably only a
2363 * good idea to (1) use it immediately after a newcall (clients only)
2364 * and (2) only use it once. Other uses currently void your warranty
2367 rx_SetArrivalProc(struct rx_call *call,
2368 void (*proc) (struct rx_call * call,
2371 void * handle, int arg)
2373 call->arrivalProc = proc;
2374 call->arrivalProcHandle = handle;
2375 call->arrivalProcArg = arg;
2378 /* Call is finished (possibly prematurely). Return rc to the peer, if
2379 * appropriate, and return the final error code from the conversation
2383 rx_EndCall(struct rx_call *call, afs_int32 rc)
2385 struct rx_connection *conn = call->conn;
2389 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2390 call, rc, call->error, call->abortCode));
2393 MUTEX_ENTER(&call->lock);
2395 if (rc == 0 && call->error == 0) {
2396 call->abortCode = 0;
2397 call->abortCount = 0;
2400 call->arrivalProc = (void (*)())0;
2401 if (rc && call->error == 0) {
2402 rxi_CallError(call, rc);
2403 call->app.mode = RX_MODE_ERROR;
2404 /* Send an abort message to the peer if this error code has
2405 * only just been set. If it was set previously, assume the
2406 * peer has already been sent the error code or will request it
2408 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2410 if (conn->type == RX_SERVER_CONNECTION) {
2411 /* Make sure reply or at least dummy reply is sent */
2412 if (call->app.mode == RX_MODE_RECEIVING) {
2413 MUTEX_EXIT(&call->lock);
2414 rxi_WriteProc(call, 0, 0);
2415 MUTEX_ENTER(&call->lock);
2417 if (call->app.mode == RX_MODE_SENDING) {
2418 MUTEX_EXIT(&call->lock);
2419 rxi_FlushWrite(call);
2420 MUTEX_ENTER(&call->lock);
2422 rxi_calltrace(RX_CALL_END, call);
2423 /* Call goes to hold state until reply packets are acknowledged */
2424 if (call->tfirst + call->nSoftAcked < call->tnext) {
2425 call->state = RX_STATE_HOLD;
2427 call->state = RX_STATE_DALLY;
2428 rxi_ClearTransmitQueue(call, 0);
2429 rxi_rto_cancel(call);
2430 rxevent_Cancel(&call->keepAliveEvent, call,
2431 RX_CALL_REFCOUNT_ALIVE);
2433 } else { /* Client connection */
2435 /* Make sure server receives input packets, in the case where
2436 * no reply arguments are expected */
2438 if ((call->app.mode == RX_MODE_SENDING)
2439 || (call->app.mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2440 MUTEX_EXIT(&call->lock);
2441 (void)rxi_ReadProc(call, &dummy, 1);
2442 MUTEX_ENTER(&call->lock);
2445 /* If we had an outstanding delayed ack, be nice to the server
2446 * and force-send it now.
2448 if (call->delayedAckEvent) {
2449 rxevent_Cancel(&call->delayedAckEvent, call,
2450 RX_CALL_REFCOUNT_DELAY);
2451 rxi_SendDelayedAck(NULL, call, NULL, 0);
2454 /* We need to release the call lock since it's lower than the
2455 * conn_call_lock and we don't want to hold the conn_call_lock
2456 * over the rx_ReadProc call. The conn_call_lock needs to be held
2457 * here for the case where rx_NewCall is perusing the calls on
2458 * the connection structure. We don't want to signal until
2459 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2460 * have checked this call, found it active and by the time it
2461 * goes to sleep, will have missed the signal.
2463 MUTEX_EXIT(&call->lock);
2464 MUTEX_ENTER(&conn->conn_call_lock);
2465 MUTEX_ENTER(&call->lock);
2467 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2468 conn->lastBusy[call->channel] = 0;
2471 MUTEX_ENTER(&conn->conn_data_lock);
2472 conn->flags |= RX_CONN_BUSY;
2473 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2474 MUTEX_EXIT(&conn->conn_data_lock);
2475 #ifdef RX_ENABLE_LOCKS
2476 CV_BROADCAST(&conn->conn_call_cv);
2481 #ifdef RX_ENABLE_LOCKS
2483 MUTEX_EXIT(&conn->conn_data_lock);
2485 #endif /* RX_ENABLE_LOCKS */
2486 call->state = RX_STATE_DALLY;
2488 error = call->error;
2490 /* currentPacket, nLeft, and NFree must be zeroed here, because
2491 * ResetCall cannot: ResetCall may be called at splnet(), in the
2492 * kernel version, and may interrupt the macros rx_Read or
2493 * rx_Write, which run at normal priority for efficiency. */
2494 if (call->app.currentPacket) {
2495 #ifdef RX_TRACK_PACKETS
2496 call->app.currentPacket->flags &= ~RX_PKTFLAG_CP;
2498 rxi_FreePacket(call->app.currentPacket);
2499 call->app.currentPacket = (struct rx_packet *)0;
2502 call->app.nLeft = call->app.nFree = call->app.curlen = 0;
2504 /* Free any packets from the last call to ReadvProc/WritevProc */
2505 #ifdef RXDEBUG_PACKET
2507 #endif /* RXDEBUG_PACKET */
2508 rxi_FreePackets(0, &call->app.iovq);
2509 MUTEX_EXIT(&call->lock);
2511 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2512 if (conn->type == RX_CLIENT_CONNECTION) {
2513 MUTEX_ENTER(&conn->conn_data_lock);
2514 conn->flags &= ~RX_CONN_BUSY;
2515 MUTEX_EXIT(&conn->conn_data_lock);
2516 MUTEX_EXIT(&conn->conn_call_lock);
2520 * Map errors to the local host's errno.h format.
2522 error = ntoh_syserr_conv(error);
2526 #if !defined(KERNEL)
2528 /* Call this routine when shutting down a server or client (especially
2529 * clients). This will allow Rx to gracefully garbage collect server
2530 * connections, and reduce the number of retries that a server might
2531 * make to a dead client.
2532 * This is not quite right, since some calls may still be ongoing and
2533 * we can't lock them to destroy them. */
2537 struct rx_connection **conn_ptr, **conn_end;
2541 if (rxinit_status == 1) {
2543 return; /* Already shutdown. */
2545 rxi_DeleteCachedConnections();
2546 if (rx_connHashTable) {
2547 MUTEX_ENTER(&rx_connHashTable_lock);
2548 for (conn_ptr = &rx_connHashTable[0], conn_end =
2549 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2551 struct rx_connection *conn, *next;
2552 for (conn = *conn_ptr; conn; conn = next) {
2554 if (conn->type == RX_CLIENT_CONNECTION) {
2555 MUTEX_ENTER(&rx_refcnt_mutex);
2557 MUTEX_EXIT(&rx_refcnt_mutex);
2558 #ifdef RX_ENABLE_LOCKS
2559 rxi_DestroyConnectionNoLock(conn);
2560 #else /* RX_ENABLE_LOCKS */
2561 rxi_DestroyConnection(conn);
2562 #endif /* RX_ENABLE_LOCKS */
2566 #ifdef RX_ENABLE_LOCKS
2567 while (rx_connCleanup_list) {
2568 struct rx_connection *conn;
2569 conn = rx_connCleanup_list;
2570 rx_connCleanup_list = rx_connCleanup_list->next;
2571 MUTEX_EXIT(&rx_connHashTable_lock);
2572 rxi_CleanupConnection(conn);
2573 MUTEX_ENTER(&rx_connHashTable_lock);
2575 MUTEX_EXIT(&rx_connHashTable_lock);
2576 #endif /* RX_ENABLE_LOCKS */
2581 afs_winsockCleanup();
2589 /* if we wakeup packet waiter too often, can get in loop with two
2590 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2592 rxi_PacketsUnWait(void)
2594 if (!rx_waitingForPackets) {
2598 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2599 return; /* still over quota */
2602 rx_waitingForPackets = 0;
2603 #ifdef RX_ENABLE_LOCKS
2604 CV_BROADCAST(&rx_waitingForPackets_cv);
2606 osi_rxWakeup(&rx_waitingForPackets);
2612 /* ------------------Internal interfaces------------------------- */
2614 /* Return this process's service structure for the
2615 * specified socket and service */
2616 static struct rx_service *
2617 rxi_FindService(osi_socket socket, u_short serviceId)
2619 struct rx_service **sp;
2620 for (sp = &rx_services[0]; *sp; sp++) {
2621 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2627 #ifdef RXDEBUG_PACKET
2628 #ifdef KDUMP_RX_LOCK
2629 static struct rx_call_rx_lock *rx_allCallsp = 0;
2631 static struct rx_call *rx_allCallsp = 0;
2633 #endif /* RXDEBUG_PACKET */
2635 /* Allocate a call structure, for the indicated channel of the
2636 * supplied connection. The mode and state of the call must be set by
2637 * the caller. Returns the call with mutex locked. */
2638 static struct rx_call *
2639 rxi_NewCall(struct rx_connection *conn, int channel)
2641 struct rx_call *call;
2642 #ifdef RX_ENABLE_LOCKS
2643 struct rx_call *cp; /* Call pointer temp */
2644 struct opr_queue *cursor;
2647 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2649 /* Grab an existing call structure, or allocate a new one.
2650 * Existing call structures are assumed to have been left reset by
2652 MUTEX_ENTER(&rx_freeCallQueue_lock);
2654 #ifdef RX_ENABLE_LOCKS
2656 * EXCEPT that the TQ might not yet be cleared out.
2657 * Skip over those with in-use TQs.
2660 for (opr_queue_Scan(&rx_freeCallQueue, cursor)) {
2661 cp = opr_queue_Entry(cursor, struct rx_call, entry);
2662 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2668 #else /* RX_ENABLE_LOCKS */
2669 if (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
2670 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
2671 #endif /* RX_ENABLE_LOCKS */
2672 opr_queue_Remove(&call->entry);
2673 if (rx_stats_active)
2674 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2675 MUTEX_EXIT(&rx_freeCallQueue_lock);
2676 MUTEX_ENTER(&call->lock);
2677 CLEAR_CALL_QUEUE_LOCK(call);
2678 #ifdef RX_ENABLE_LOCKS
2679 /* Now, if TQ wasn't cleared earlier, do it now. */
2680 rxi_WaitforTQBusy(call);
2681 if (call->flags & RX_CALL_TQ_CLEARME) {
2682 rxi_ClearTransmitQueue(call, 1);
2683 /*queue_Init(&call->tq);*/
2685 #endif /* RX_ENABLE_LOCKS */
2686 /* Bind the call to its connection structure */
2688 rxi_ResetCall(call, 1);
2691 call = rxi_Alloc(sizeof(struct rx_call));
2692 #ifdef RXDEBUG_PACKET
2693 call->allNextp = rx_allCallsp;
2694 rx_allCallsp = call;
2696 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2697 #else /* RXDEBUG_PACKET */
2698 rx_atomic_inc(&rx_stats.nCallStructs);
2699 #endif /* RXDEBUG_PACKET */
2701 MUTEX_EXIT(&rx_freeCallQueue_lock);
2702 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2703 MUTEX_ENTER(&call->lock);
2704 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2705 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2706 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2708 /* Initialize once-only items */
2709 opr_queue_Init(&call->tq);
2710 opr_queue_Init(&call->rq);
2711 opr_queue_Init(&call->app.iovq);
2712 #ifdef RXDEBUG_PACKET
2713 call->rqc = call->tqc = call->iovqc = 0;
2714 #endif /* RXDEBUG_PACKET */
2715 /* Bind the call to its connection structure (prereq for reset) */
2717 rxi_ResetCall(call, 1);
2719 call->channel = channel;
2720 call->callNumber = &conn->callNumber[channel];
2721 call->rwind = conn->rwind[channel];
2722 call->twind = conn->twind[channel];
2723 /* Note that the next expected call number is retained (in
2724 * conn->callNumber[i]), even if we reallocate the call structure
2726 conn->call[channel] = call;
2727 /* if the channel's never been used (== 0), we should start at 1, otherwise
2728 * the call number is valid from the last time this channel was used */
2729 if (*call->callNumber == 0)
2730 *call->callNumber = 1;
2735 /* A call has been inactive long enough that so we can throw away
2736 * state, including the call structure, which is placed on the call
2739 * call->lock amd rx_refcnt_mutex are held upon entry.
2740 * haveCTLock is set when called from rxi_ReapConnections.
2742 * return 1 if the call is freed, 0 if not.
2745 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2747 int channel = call->channel;
2748 struct rx_connection *conn = call->conn;
2749 u_char state = call->state;
2752 * We are setting the state to RX_STATE_RESET to
2753 * ensure that no one else will attempt to use this
2754 * call once we drop the refcnt lock. We must drop
2755 * the refcnt lock before calling rxi_ResetCall
2756 * because it cannot be held across acquiring the
2757 * freepktQ lock. NewCall does the same.
2759 call->state = RX_STATE_RESET;
2760 MUTEX_EXIT(&rx_refcnt_mutex);
2761 rxi_ResetCall(call, 0);
2763 if (MUTEX_TRYENTER(&conn->conn_call_lock))
2765 if (state == RX_STATE_DALLY || state == RX_STATE_HOLD)
2766 (*call->callNumber)++;
2768 if (call->conn->call[channel] == call)
2769 call->conn->call[channel] = 0;
2770 MUTEX_EXIT(&conn->conn_call_lock);
2773 * We couldn't obtain the conn_call_lock so we can't
2774 * disconnect the call from the connection. Set the
2775 * call state to dally so that the call can be reused.
2777 MUTEX_ENTER(&rx_refcnt_mutex);
2778 call->state = RX_STATE_DALLY;
2782 MUTEX_ENTER(&rx_freeCallQueue_lock);
2783 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2784 #ifdef RX_ENABLE_LOCKS
2785 /* A call may be free even though its transmit queue is still in use.
2786 * Since we search the call list from head to tail, put busy calls at
2787 * the head of the list, and idle calls at the tail.
2789 if (call->flags & RX_CALL_TQ_BUSY)
2790 opr_queue_Prepend(&rx_freeCallQueue, &call->entry);
2792 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2793 #else /* RX_ENABLE_LOCKS */
2794 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2795 #endif /* RX_ENABLE_LOCKS */
2796 if (rx_stats_active)
2797 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2798 MUTEX_EXIT(&rx_freeCallQueue_lock);
2800 /* Destroy the connection if it was previously slated for
2801 * destruction, i.e. the Rx client code previously called
2802 * rx_DestroyConnection (client connections), or
2803 * rxi_ReapConnections called the same routine (server
2804 * connections). Only do this, however, if there are no
2805 * outstanding calls. Note that for fine grain locking, there appears
2806 * to be a deadlock in that rxi_FreeCall has a call locked and
2807 * DestroyConnectionNoLock locks each call in the conn. But note a
2808 * few lines up where we have removed this call from the conn.
2809 * If someone else destroys a connection, they either have no
2810 * call lock held or are going through this section of code.
2812 MUTEX_ENTER(&conn->conn_data_lock);
2813 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2814 MUTEX_ENTER(&rx_refcnt_mutex);
2816 MUTEX_EXIT(&rx_refcnt_mutex);
2817 MUTEX_EXIT(&conn->conn_data_lock);
2818 #ifdef RX_ENABLE_LOCKS
2820 rxi_DestroyConnectionNoLock(conn);
2822 rxi_DestroyConnection(conn);
2823 #else /* RX_ENABLE_LOCKS */
2824 rxi_DestroyConnection(conn);
2825 #endif /* RX_ENABLE_LOCKS */
2827 MUTEX_EXIT(&conn->conn_data_lock);
2829 MUTEX_ENTER(&rx_refcnt_mutex);
2833 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2834 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2837 rxi_Alloc(size_t size)
2841 if (rx_stats_active) {
2842 rx_atomic_add(&rxi_Allocsize, (int) size);
2843 rx_atomic_inc(&rxi_Alloccnt);
2847 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2848 afs_osi_Alloc_NoSleep(size);
2853 osi_Panic("rxi_Alloc error");
2859 rxi_Free(void *addr, size_t size)
2861 if (rx_stats_active) {
2862 rx_atomic_sub(&rxi_Allocsize, (int) size);
2863 rx_atomic_dec(&rxi_Alloccnt);
2865 osi_Free(addr, size);
2869 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2871 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2872 struct rx_peer *next = NULL;
2876 MUTEX_ENTER(&rx_peerHashTable_lock);
2878 peer_ptr = &rx_peerHashTable[0];
2879 peer_end = &rx_peerHashTable[rx_hashTableSize];
2882 for ( ; peer_ptr < peer_end; peer_ptr++) {
2885 for ( ; peer; peer = next) {
2887 if (host == peer->host)
2892 hashIndex = PEER_HASH(host, port);
2893 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2894 if ((peer->host == host) && (peer->port == port))
2899 MUTEX_ENTER(&rx_peerHashTable_lock);
2904 MUTEX_EXIT(&rx_peerHashTable_lock);
2906 MUTEX_ENTER(&peer->peer_lock);
2907 /* We don't handle dropping below min, so don't */
2908 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2909 peer->ifMTU=MIN(mtu, peer->ifMTU);
2910 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2911 /* if we tweaked this down, need to tune our peer MTU too */
2912 peer->MTU = MIN(peer->MTU, peer->natMTU);
2913 /* if we discovered a sub-1500 mtu, degrade */
2914 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2915 peer->maxDgramPackets = 1;
2916 /* We no longer have valid peer packet information */
2917 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2918 peer->maxPacketSize = 0;
2919 MUTEX_EXIT(&peer->peer_lock);
2921 MUTEX_ENTER(&rx_peerHashTable_lock);
2923 if (host && !port) {
2925 /* pick up where we left off */
2929 MUTEX_EXIT(&rx_peerHashTable_lock);
2932 #ifdef AFS_RXERRQ_ENV
2934 rxi_SetPeerDead(afs_uint32 host, afs_uint16 port)
2936 int hashIndex = PEER_HASH(host, port);
2937 struct rx_peer *peer;
2939 MUTEX_ENTER(&rx_peerHashTable_lock);
2941 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2942 if (peer->host == host && peer->port == port) {
2948 rx_atomic_inc(&peer->neterrs);
2951 MUTEX_EXIT(&rx_peerHashTable_lock);
2955 rxi_ProcessNetError(struct sock_extended_err *err, afs_uint32 addr, afs_uint16 port)
2957 # ifdef AFS_ADAPT_PMTU
2958 if (err->ee_errno == EMSGSIZE && err->ee_info >= 68) {
2959 rxi_SetPeerMtu(NULL, addr, port, err->ee_info - RX_IPUDP_SIZE);
2963 if (err->ee_origin == SO_EE_ORIGIN_ICMP && err->ee_type == ICMP_DEST_UNREACH) {
2964 switch (err->ee_code) {
2965 case ICMP_NET_UNREACH:
2966 case ICMP_HOST_UNREACH:
2967 case ICMP_PORT_UNREACH:
2970 rxi_SetPeerDead(addr, port);
2975 #endif /* AFS_RXERRQ_ENV */
2977 /* Find the peer process represented by the supplied (host,port)
2978 * combination. If there is no appropriate active peer structure, a
2979 * new one will be allocated and initialized
2980 * The origPeer, if set, is a pointer to a peer structure on which the
2981 * refcount will be be decremented. This is used to replace the peer
2982 * structure hanging off a connection structure */
2984 rxi_FindPeer(afs_uint32 host, u_short port,
2985 struct rx_peer *origPeer, int create)
2989 hashIndex = PEER_HASH(host, port);
2990 MUTEX_ENTER(&rx_peerHashTable_lock);
2991 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2992 if ((pp->host == host) && (pp->port == port))
2997 pp = rxi_AllocPeer(); /* This bzero's *pp */
2998 pp->host = host; /* set here or in InitPeerParams is zero */
3000 #ifdef AFS_RXERRQ_ENV
3001 rx_atomic_set(&pp->neterrs, 0);
3003 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
3004 opr_queue_Init(&pp->rpcStats);
3005 pp->next = rx_peerHashTable[hashIndex];
3006 rx_peerHashTable[hashIndex] = pp;
3007 rxi_InitPeerParams(pp);
3008 if (rx_stats_active)
3009 rx_atomic_inc(&rx_stats.nPeerStructs);
3016 origPeer->refCount--;
3017 MUTEX_EXIT(&rx_peerHashTable_lock);
3022 /* Find the connection at (host, port) started at epoch, and with the
3023 * given connection id. Creates the server connection if necessary.
3024 * The type specifies whether a client connection or a server
3025 * connection is desired. In both cases, (host, port) specify the
3026 * peer's (host, pair) pair. Client connections are not made
3027 * automatically by this routine. The parameter socket gives the
3028 * socket descriptor on which the packet was received. This is used,
3029 * in the case of server connections, to check that *new* connections
3030 * come via a valid (port, serviceId). Finally, the securityIndex
3031 * parameter must match the existing index for the connection. If a
3032 * server connection is created, it will be created using the supplied
3033 * index, if the index is valid for this service */
3034 static struct rx_connection *
3035 rxi_FindConnection(osi_socket socket, afs_uint32 host,
3036 u_short port, u_short serviceId, afs_uint32 cid,
3037 afs_uint32 epoch, int type, u_int securityIndex)
3039 int hashindex, flag, i;
3040 struct rx_connection *conn;
3041 hashindex = CONN_HASH(host, port, cid, epoch, type);
3042 MUTEX_ENTER(&rx_connHashTable_lock);
3043 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
3044 rx_connHashTable[hashindex],
3047 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
3048 && (epoch == conn->epoch)) {
3049 struct rx_peer *pp = conn->peer;
3050 if (securityIndex != conn->securityIndex) {
3051 /* this isn't supposed to happen, but someone could forge a packet
3052 * like this, and there seems to be some CM bug that makes this
3053 * happen from time to time -- in which case, the fileserver
3055 MUTEX_EXIT(&rx_connHashTable_lock);
3056 return (struct rx_connection *)0;
3058 if (pp->host == host && pp->port == port)
3060 if (type == RX_CLIENT_CONNECTION && pp->port == port)
3062 /* So what happens when it's a callback connection? */
3063 if ( /*type == RX_CLIENT_CONNECTION && */
3064 (conn->epoch & 0x80000000))
3068 /* the connection rxLastConn that was used the last time is not the
3069 ** one we are looking for now. Hence, start searching in the hash */
3071 conn = rx_connHashTable[hashindex];
3076 struct rx_service *service;
3077 if (type == RX_CLIENT_CONNECTION) {
3078 MUTEX_EXIT(&rx_connHashTable_lock);
3079 return (struct rx_connection *)0;
3081 service = rxi_FindService(socket, serviceId);
3082 if (!service || (securityIndex >= service->nSecurityObjects)
3083 || (service->securityObjects[securityIndex] == 0)) {
3084 MUTEX_EXIT(&rx_connHashTable_lock);
3085 return (struct rx_connection *)0;
3087 conn = rxi_AllocConnection(); /* This bzero's the connection */
3088 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
3089 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
3090 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
3091 conn->next = rx_connHashTable[hashindex];
3092 rx_connHashTable[hashindex] = conn;
3093 conn->peer = rxi_FindPeer(host, port, 0, 1);
3094 conn->type = RX_SERVER_CONNECTION;
3095 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
3096 conn->epoch = epoch;
3097 conn->cid = cid & RX_CIDMASK;
3098 conn->ackRate = RX_FAST_ACK_RATE;
3099 conn->service = service;
3100 conn->serviceId = serviceId;
3101 conn->securityIndex = securityIndex;
3102 conn->securityObject = service->securityObjects[securityIndex];
3103 conn->nSpecific = 0;
3104 conn->specific = NULL;
3105 rx_SetConnDeadTime(conn, service->connDeadTime);
3106 conn->idleDeadTime = service->idleDeadTime;
3107 conn->idleDeadDetection = service->idleDeadErr ? 1 : 0;
3108 for (i = 0; i < RX_MAXCALLS; i++) {
3109 conn->twind[i] = rx_initSendWindow;
3110 conn->rwind[i] = rx_initReceiveWindow;
3112 /* Notify security object of the new connection */
3113 RXS_NewConnection(conn->securityObject, conn);
3114 /* XXXX Connection timeout? */
3115 if (service->newConnProc)
3116 (*service->newConnProc) (conn);
3117 if (rx_stats_active)
3118 rx_atomic_inc(&rx_stats.nServerConns);
3121 MUTEX_ENTER(&rx_refcnt_mutex);
3123 MUTEX_EXIT(&rx_refcnt_mutex);
3125 rxLastConn = conn; /* store this connection as the last conn used */
3126 MUTEX_EXIT(&rx_connHashTable_lock);
3131 * Timeout a call on a busy call channel if appropriate.
3133 * @param[in] call The busy call.
3135 * @pre 'call' is marked as busy (namely,
3136 * call->conn->lastBusy[call->channel] != 0)
3138 * @pre call->lock is held
3139 * @pre rxi_busyChannelError is nonzero
3141 * @note call->lock is dropped and reacquired
3144 rxi_CheckBusy(struct rx_call *call)
3146 struct rx_connection *conn = call->conn;
3147 int channel = call->channel;
3148 int freechannel = 0;
3150 afs_uint32 callNumber;
3152 MUTEX_EXIT(&call->lock);
3154 MUTEX_ENTER(&conn->conn_call_lock);
3155 callNumber = *call->callNumber;
3157 /* Are there any other call slots on this conn that we should try? Look for
3158 * slots that are empty and are either non-busy, or were marked as busy
3159 * longer than conn->secondsUntilDead seconds before this call started. */
3161 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3163 /* only look at channels that aren't us */
3167 if (conn->lastBusy[i]) {
3168 /* if this channel looked busy too recently, don't look at it */
3169 if (conn->lastBusy[i] >= call->startTime.sec) {
3172 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3177 if (conn->call[i]) {
3178 struct rx_call *tcall = conn->call[i];
3179 MUTEX_ENTER(&tcall->lock);
3180 if (tcall->state == RX_STATE_DALLY) {
3183 MUTEX_EXIT(&tcall->lock);
3189 MUTEX_ENTER(&call->lock);
3191 /* Since the call->lock and conn->conn_call_lock have been released it is
3192 * possible that (1) the call may no longer be busy and/or (2) the call may
3193 * have been reused by another waiting thread. Therefore, we must confirm
3194 * that the call state has not changed when deciding whether or not to
3195 * force this application thread to retry by forcing a Timeout error. */
3197 if (freechannel && *call->callNumber == callNumber &&
3198 (call->flags & RX_CALL_PEER_BUSY)) {
3199 /* Since 'freechannel' is set, there exists another channel in this
3200 * rx_conn that the application thread might be able to use. We know
3201 * that we have the correct call since callNumber is unchanged, and we
3202 * know that the call is still busy. So, set the call error state to
3203 * rxi_busyChannelError so the application can retry the request,
3204 * presumably on a less-busy call channel. */
3206 rxi_CallError(call, RX_CALL_BUSY);
3208 MUTEX_EXIT(&conn->conn_call_lock);
3211 /* There are two packet tracing routines available for testing and monitoring
3212 * Rx. One is called just after every packet is received and the other is
3213 * called just before every packet is sent. Received packets, have had their
3214 * headers decoded, and packets to be sent have not yet had their headers
3215 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3216 * containing the network address. Both can be modified. The return value, if
3217 * non-zero, indicates that the packet should be dropped. */
3219 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3220 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3222 /* A packet has been received off the interface. Np is the packet, socket is
3223 * the socket number it was received from (useful in determining which service
3224 * this packet corresponds to), and (host, port) reflect the host,port of the
3225 * sender. This call returns the packet to the caller if it is finished with
3226 * it, rather than de-allocating it, just as a small performance hack */
3229 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3230 afs_uint32 host, u_short port, int *tnop,
3231 struct rx_call **newcallp)
3233 struct rx_call *call;
3234 struct rx_connection *conn;
3236 afs_uint32 currentCallNumber;
3241 struct rx_packet *tnp;
3244 /* We don't print out the packet until now because (1) the time may not be
3245 * accurate enough until now in the lwp implementation (rx_Listener only gets
3246 * the time after the packet is read) and (2) from a protocol point of view,
3247 * this is the first time the packet has been seen */
3248 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3249 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3250 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3251 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3252 np->header.epoch, np->header.cid, np->header.callNumber,
3253 np->header.seq, np->header.flags, np));
3256 /* Account for connectionless packets */
3257 if (rx_stats_active &&
3258 ((np->header.type == RX_PACKET_TYPE_VERSION) ||
3259 (np->header.type == RX_PACKET_TYPE_DEBUG))) {
3260 struct rx_peer *peer;
3262 /* Try to look up the peer structure, but don't create one */
3263 peer = rxi_FindPeer(host, port, 0, 0);
3265 /* Since this may not be associated with a connection, it may have
3266 * no refCount, meaning we could race with ReapConnections
3269 if (peer && (peer->refCount > 0)) {
3270 #ifdef AFS_RXERRQ_ENV
3271 if (rx_atomic_read(&peer->neterrs)) {
3272 rx_atomic_set(&peer->neterrs, 0);
3275 MUTEX_ENTER(&peer->peer_lock);
3276 peer->bytesReceived += np->length;
3277 MUTEX_EXIT(&peer->peer_lock);
3281 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3282 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3285 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3286 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3289 /* If an input tracer function is defined, call it with the packet and
3290 * network address. Note this function may modify its arguments. */
3291 if (rx_justReceived) {
3292 struct sockaddr_in addr;
3294 addr.sin_family = AF_INET;
3295 addr.sin_port = port;
3296 addr.sin_addr.s_addr = host;
3297 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3298 addr.sin_len = sizeof(addr);
3299 #endif /* AFS_OSF_ENV */
3300 drop = (*rx_justReceived) (np, &addr);
3301 /* drop packet if return value is non-zero */
3304 port = addr.sin_port; /* in case fcn changed addr */
3305 host = addr.sin_addr.s_addr;
3309 /* If packet was not sent by the client, then *we* must be the client */
3310 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3311 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3313 /* Find the connection (or fabricate one, if we're the server & if
3314 * necessary) associated with this packet */
3316 rxi_FindConnection(socket, host, port, np->header.serviceId,
3317 np->header.cid, np->header.epoch, type,
3318 np->header.securityIndex);
3320 /* To avoid having 2 connections just abort at each other,
3321 don't abort an abort. */
3323 if (np->header.type != RX_PACKET_TYPE_ABORT)
3324 rxi_SendRawAbort(socket, host, port, RX_INVALID_OPERATION,
3329 #ifdef AFS_RXERRQ_ENV
3330 if (rx_atomic_read(&conn->peer->neterrs)) {
3331 rx_atomic_set(&conn->peer->neterrs, 0);
3335 /* If we're doing statistics, then account for the incoming packet */
3336 if (rx_stats_active) {
3337 MUTEX_ENTER(&conn->peer->peer_lock);
3338 conn->peer->bytesReceived += np->length;
3339 MUTEX_EXIT(&conn->peer->peer_lock);
3342 /* If the connection is in an error state, send an abort packet and ignore
3343 * the incoming packet */
3345 /* Don't respond to an abort packet--we don't want loops! */
3346 MUTEX_ENTER(&conn->conn_data_lock);
3347 if (np->header.type != RX_PACKET_TYPE_ABORT)
3348 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3349 putConnection(conn);
3350 MUTEX_EXIT(&conn->conn_data_lock);
3354 /* Check for connection-only requests (i.e. not call specific). */
3355 if (np->header.callNumber == 0) {
3356 switch (np->header.type) {
3357 case RX_PACKET_TYPE_ABORT: {
3358 /* What if the supplied error is zero? */
3359 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3360 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3361 rxi_ConnectionError(conn, errcode);
3362 putConnection(conn);
3365 case RX_PACKET_TYPE_CHALLENGE:
3366 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3367 putConnection(conn);
3369 case RX_PACKET_TYPE_RESPONSE:
3370 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3371 putConnection(conn);
3373 case RX_PACKET_TYPE_PARAMS:
3374 case RX_PACKET_TYPE_PARAMS + 1:
3375 case RX_PACKET_TYPE_PARAMS + 2:
3376 /* ignore these packet types for now */
3377 putConnection(conn);
3381 /* Should not reach here, unless the peer is broken: send an
3383 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3384 MUTEX_ENTER(&conn->conn_data_lock);
3385 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3386 putConnection(conn);
3387 MUTEX_EXIT(&conn->conn_data_lock);
3392 channel = np->header.cid & RX_CHANNELMASK;
3393 MUTEX_ENTER(&conn->conn_call_lock);
3394 call = conn->call[channel];
3397 MUTEX_ENTER(&call->lock);
3398 currentCallNumber = conn->callNumber[channel];
3399 MUTEX_EXIT(&conn->conn_call_lock);
3400 } else if (type == RX_SERVER_CONNECTION) { /* No call allocated */
3401 call = rxi_NewCall(conn, channel); /* returns locked call */
3402 *call->callNumber = currentCallNumber = np->header.callNumber;
3403 MUTEX_EXIT(&conn->conn_call_lock);
3405 if (np->header.callNumber == 0)
3406 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, "
3407 "packet %"AFS_PTR_FMT" len %d\n",
3408 np->header.serial, rx_packetTypes[np->header.type - 1],
3409 ntohl(conn->peer->host), ntohs(conn->peer->port),
3410 np->header.serial, np->header.epoch, np->header.cid,
3411 np->header.callNumber, np->header.seq,
3412 np->header.flags, np, np->length));
3414 call->state = RX_STATE_PRECALL;
3415 clock_GetTime(&call->queueTime);
3416 call->app.bytesSent = 0;
3417 call->app.bytesRcvd = 0;
3419 * If the number of queued calls exceeds the overload
3420 * threshold then abort this call.
3422 if ((rx_BusyThreshold > 0) &&
3423 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3424 struct rx_packet *tp;
3426 rxi_CallError(call, rx_BusyError);
3427 tp = rxi_SendCallAbort(call, np, 1, 0);
3428 MUTEX_EXIT(&call->lock);
3429 putConnection(conn);
3430 if (rx_stats_active)
3431 rx_atomic_inc(&rx_stats.nBusies);
3434 rxi_KeepAliveOn(call);
3435 } else { /* RX_CLIENT_CONNECTION and No call allocated */
3436 /* This packet can't be for this call. If the new call address is
3437 * 0 then no call is running on this channel. If there is a call
3438 * then, since this is a client connection we're getting data for
3439 * it must be for the previous call.
3441 MUTEX_EXIT(&conn->conn_call_lock);
3442 if (rx_stats_active)
3443 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3444 putConnection(conn);
3448 /* There is a non-NULL locked call at this point */
3449 if (type == RX_SERVER_CONNECTION) { /* We're the server */
3450 if (np->header.callNumber < currentCallNumber) {
3451 MUTEX_EXIT(&call->lock);
3452 if (rx_stats_active)
3453 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3454 putConnection(conn);
3456 } else if (np->header.callNumber != currentCallNumber) {
3457 /* Wait until the transmit queue is idle before deciding
3458 * whether to reset the current call. Chances are that the
3459 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3462 #ifdef RX_ENABLE_LOCKS
3463 if (call->state == RX_STATE_ACTIVE) {
3464 rxi_WaitforTQBusy(call);
3466 * If we entered error state while waiting,
3467 * must call rxi_CallError to permit rxi_ResetCall
3468 * to processed when the tqWaiter count hits zero.
3471 rxi_CallError(call, call->error);
3472 MUTEX_EXIT(&call->lock);
3473 putConnection(conn);
3477 #endif /* RX_ENABLE_LOCKS */
3478 /* If the new call cannot be taken right now send a busy and set
3479 * the error condition in this call, so that it terminates as
3480 * quickly as possible */
3481 if (call->state == RX_STATE_ACTIVE) {
3482 struct rx_packet *tp;
3484 rxi_CallError(call, RX_CALL_DEAD);
3485 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3487 MUTEX_EXIT(&call->lock);
3488 putConnection(conn);
3491 rxi_ResetCall(call, 0);
3493 * The conn_call_lock is not held but no one else should be
3494 * using this call channel while we are processing this incoming
3495 * packet. This assignment should be safe.
3497 *call->callNumber = np->header.callNumber;
3499 if (np->header.callNumber == 0)
3500 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3501 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3502 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3503 np->header.flags, np, np->length));
3505 call->state = RX_STATE_PRECALL;
3506 clock_GetTime(&call->queueTime);
3507 call->app.bytesSent = 0;
3508 call->app.bytesRcvd = 0;
3510 * If the number of queued calls exceeds the overload
3511 * threshold then abort this call.
3513 if ((rx_BusyThreshold > 0) &&
3514 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3515 struct rx_packet *tp;
3517 rxi_CallError(call, rx_BusyError);
3518 tp = rxi_SendCallAbort(call, np, 1, 0);
3519 MUTEX_EXIT(&call->lock);
3520 putConnection(conn);
3521 if (rx_stats_active)
3522 rx_atomic_inc(&rx_stats.nBusies);
3525 rxi_KeepAliveOn(call);
3527 /* Continuing call; do nothing here. */
3529 } else { /* we're the client */
3530 /* Ignore all incoming acknowledgements for calls in DALLY state */
3531 if ((call->state == RX_STATE_DALLY)
3532 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3533 if (rx_stats_active)
3534 rx_atomic_inc(&rx_stats.ignorePacketDally);
3535 MUTEX_EXIT(&call->lock);
3536 putConnection(conn);
3540 /* Ignore anything that's not relevant to the current call. If there
3541 * isn't a current call, then no packet is relevant. */
3542 if (np->header.callNumber != currentCallNumber) {
3543 if (rx_stats_active)
3544 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3545 MUTEX_EXIT(&call->lock);
3546 putConnection(conn);
3549 /* If the service security object index stamped in the packet does not
3550 * match the connection's security index, ignore the packet */
3551 if (np->header.securityIndex != conn->securityIndex) {
3552 MUTEX_EXIT(&call->lock);
3553 putConnection(conn);
3557 /* If we're receiving the response, then all transmit packets are
3558 * implicitly acknowledged. Get rid of them. */
3559 if (np->header.type == RX_PACKET_TYPE_DATA) {
3560 #ifdef RX_ENABLE_LOCKS
3561 /* XXX Hack. Because we must release the call lock when
3562 * sending packets (osi_NetSend) we drop all acks while we're
3563 * traversing the tq in rxi_Start sending packets out because
3564 * packets may move to the freePacketQueue as result of being here!
3565 * So we drop these packets until we're safely out of the
3566 * traversing. Really ugly!
3567 * For fine grain RX locking, we set the acked field in the
3568 * packets and let rxi_Start remove them from the transmit queue.
3570 if (call->flags & RX_CALL_TQ_BUSY) {
3571 rxi_SetAcksInTransmitQueue(call);
3573 rxi_ClearTransmitQueue(call, 0);
3575 #else /* RX_ENABLE_LOCKS */
3576 rxi_ClearTransmitQueue(call, 0);
3577 #endif /* RX_ENABLE_LOCKS */
3579 if (np->header.type == RX_PACKET_TYPE_ACK) {
3580 /* now check to see if this is an ack packet acknowledging that the
3581 * server actually *lost* some hard-acked data. If this happens we
3582 * ignore this packet, as it may indicate that the server restarted in
3583 * the middle of a call. It is also possible that this is an old ack
3584 * packet. We don't abort the connection in this case, because this
3585 * *might* just be an old ack packet. The right way to detect a server
3586 * restart in the midst of a call is to notice that the server epoch
3588 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3589 * XXX unacknowledged. I think that this is off-by-one, but
3590 * XXX I don't dare change it just yet, since it will
3591 * XXX interact badly with the server-restart detection
3592 * XXX code in receiveackpacket. */
3593 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3594 if (rx_stats_active)
3595 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3596 MUTEX_EXIT(&call->lock);
3597 putConnection(conn);
3601 } /* else not a data packet */
3604 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3605 /* Set remote user defined status from packet */
3606 call->remoteStatus = np->header.userStatus;
3608 /* Now do packet type-specific processing */
3609 switch (np->header.type) {
3610 case RX_PACKET_TYPE_DATA:
3611 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3614 case RX_PACKET_TYPE_ACK:
3615 /* Respond immediately to ack packets requesting acknowledgement
3617 if (np->header.flags & RX_REQUEST_ACK) {
3619 (void)rxi_SendCallAbort(call, 0, 1, 0);
3621 (void)rxi_SendAck(call, 0, np->header.serial,
3622 RX_ACK_PING_RESPONSE, 1);
3624 np = rxi_ReceiveAckPacket(call, np, 1);
3626 case RX_PACKET_TYPE_ABORT: {
3627 /* An abort packet: reset the call, passing the error up to the user. */
3628 /* What if error is zero? */
3629 /* What if the error is -1? the application will treat it as a timeout. */
3630 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3631 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3632 rxi_CallError(call, errdata);
3633 MUTEX_EXIT(&call->lock);
3634 putConnection(conn);
3635 return np; /* xmitting; drop packet */
3637 case RX_PACKET_TYPE_BUSY: {
3638 struct clock busyTime;
3640 clock_GetTime(&busyTime);
3642 MUTEX_EXIT(&call->lock);
3644 MUTEX_ENTER(&conn->conn_call_lock);
3645 MUTEX_ENTER(&call->lock);
3646 conn->lastBusy[call->channel] = busyTime.sec;
3647 call->flags |= RX_CALL_PEER_BUSY;
3648 MUTEX_EXIT(&call->lock);
3649 MUTEX_EXIT(&conn->conn_call_lock);
3651 putConnection(conn);
3655 case RX_PACKET_TYPE_ACKALL:
3656 /* All packets acknowledged, so we can drop all packets previously
3657 * readied for sending */
3658 #ifdef RX_ENABLE_LOCKS
3659 /* XXX Hack. We because we can't release the call lock when
3660 * sending packets (osi_NetSend) we drop all ack pkts while we're
3661 * traversing the tq in rxi_Start sending packets out because
3662 * packets may move to the freePacketQueue as result of being
3663 * here! So we drop these packets until we're safely out of the
3664 * traversing. Really ugly!
3665 * For fine grain RX locking, we set the acked field in the packets
3666 * and let rxi_Start remove the packets from the transmit queue.
3668 if (call->flags & RX_CALL_TQ_BUSY) {
3669 rxi_SetAcksInTransmitQueue(call);
3672 #endif /* RX_ENABLE_LOCKS */
3673 rxi_ClearTransmitQueue(call, 0);
3676 /* Should not reach here, unless the peer is broken: send an abort
3678 rxi_CallError(call, RX_PROTOCOL_ERROR);
3679 np = rxi_SendCallAbort(call, np, 1, 0);
3682 /* Note when this last legitimate packet was received, for keep-alive
3683 * processing. Note, we delay getting the time until now in the hope that
3684 * the packet will be delivered to the user before any get time is required
3685 * (if not, then the time won't actually be re-evaluated here). */
3686 call->lastReceiveTime = clock_Sec();
3687 /* we've received a legit packet, so the channel is not busy */
3688 call->flags &= ~RX_CALL_PEER_BUSY;
3689 MUTEX_EXIT(&call->lock);
3690 putConnection(conn);
3694 /* return true if this is an "interesting" connection from the point of view
3695 of someone trying to debug the system */
3697 rxi_IsConnInteresting(struct rx_connection *aconn)
3700 struct rx_call *tcall;
3702 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3705 for (i = 0; i < RX_MAXCALLS; i++) {
3706 tcall = aconn->call[i];
3708 if ((tcall->state == RX_STATE_PRECALL)
3709 || (tcall->state == RX_STATE_ACTIVE))
3711 if ((tcall->app.mode == RX_MODE_SENDING)
3712 || (tcall->app.mode == RX_MODE_RECEIVING))
3720 /* if this is one of the last few packets AND it wouldn't be used by the
3721 receiving call to immediately satisfy a read request, then drop it on
3722 the floor, since accepting it might prevent a lock-holding thread from
3723 making progress in its reading. If a call has been cleared while in
3724 the precall state then ignore all subsequent packets until the call
3725 is assigned to a thread. */
3728 TooLow(struct rx_packet *ap, struct rx_call *acall)
3732 MUTEX_ENTER(&rx_quota_mutex);
3733 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3734 && (acall->state == RX_STATE_PRECALL))
3735 || ((rx_nFreePackets < rxi_dataQuota + 2)
3736 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3737 && (acall->flags & RX_CALL_READER_WAIT)))) {
3740 MUTEX_EXIT(&rx_quota_mutex);
3746 * Clear the attach wait flag on a connection and proceed.
3748 * Any processing waiting for a connection to be attached should be
3749 * unblocked. We clear the flag and do any other needed tasks.
3752 * the conn to unmark waiting for attach
3754 * @pre conn's conn_data_lock must be locked before calling this function
3758 rxi_ConnClearAttachWait(struct rx_connection *conn)
3760 /* Indicate that rxi_CheckReachEvent is no longer running by
3761 * clearing the flag. Must be atomic under conn_data_lock to
3762 * avoid a new call slipping by: rxi_CheckConnReach holds
3763 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3765 conn->flags &= ~RX_CONN_ATTACHWAIT;
3766 if (conn->flags & RX_CONN_NAT_PING) {
3767 conn->flags &= ~RX_CONN_NAT_PING;
3768 rxi_ScheduleNatKeepAliveEvent(conn);
3773 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3775 struct rx_connection *conn = arg1;
3776 struct rx_call *acall = arg2;
3777 struct rx_call *call = acall;
3778 struct clock when, now;
3781 MUTEX_ENTER(&conn->conn_data_lock);
3784 rxevent_Put(conn->checkReachEvent);
3785 conn->checkReachEvent = NULL;
3788 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3790 putConnection(conn);
3792 MUTEX_EXIT(&conn->conn_data_lock);
3796 MUTEX_ENTER(&conn->conn_call_lock);
3797 MUTEX_ENTER(&conn->conn_data_lock);
3798 for (i = 0; i < RX_MAXCALLS; i++) {
3799 struct rx_call *tc = conn->call[i];
3800 if (tc && tc->state == RX_STATE_PRECALL) {
3806 rxi_ConnClearAttachWait(conn);
3807 MUTEX_EXIT(&conn->conn_data_lock);
3808 MUTEX_EXIT(&conn->conn_call_lock);
3813 MUTEX_ENTER(&call->lock);
3814 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3816 MUTEX_EXIT(&call->lock);
3818 clock_GetTime(&now);
3820 when.sec += RX_CHECKREACH_TIMEOUT;
3821 MUTEX_ENTER(&conn->conn_data_lock);
3822 if (!conn->checkReachEvent) {
3823 MUTEX_ENTER(&rx_refcnt_mutex);
3825 MUTEX_EXIT(&rx_refcnt_mutex);
3826 conn->checkReachEvent = rxevent_Post(&when, &now,
3827 rxi_CheckReachEvent, conn,
3830 MUTEX_EXIT(&conn->conn_data_lock);
3836 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3838 struct rx_service *service = conn->service;
3839 struct rx_peer *peer = conn->peer;
3840 afs_uint32 now, lastReach;
3842 if (service->checkReach == 0)
3846 MUTEX_ENTER(&peer->peer_lock);
3847 lastReach = peer->lastReachTime;
3848 MUTEX_EXIT(&peer->peer_lock);
3849 if (now - lastReach < RX_CHECKREACH_TTL)
3852 MUTEX_ENTER(&conn->conn_data_lock);
3853 if (conn->flags & RX_CONN_ATTACHWAIT) {
3854 MUTEX_EXIT(&conn->conn_data_lock);
3857 conn->flags |= RX_CONN_ATTACHWAIT;
3858 MUTEX_EXIT(&conn->conn_data_lock);
3859 if (!conn->checkReachEvent)
3860 rxi_CheckReachEvent(NULL, conn, call, 0);
3865 /* try to attach call, if authentication is complete */
3867 TryAttach(struct rx_call *acall, osi_socket socket,
3868 int *tnop, struct rx_call **newcallp,
3871 struct rx_connection *conn = acall->conn;
3873 if (conn->type == RX_SERVER_CONNECTION
3874 && acall->state == RX_STATE_PRECALL) {
3875 /* Don't attach until we have any req'd. authentication. */
3876 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3877 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3878 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3879 /* Note: this does not necessarily succeed; there
3880 * may not any proc available
3883 rxi_ChallengeOn(acall->conn);
3888 /* A data packet has been received off the interface. This packet is
3889 * appropriate to the call (the call is in the right state, etc.). This
3890 * routine can return a packet to the caller, for re-use */
3892 static struct rx_packet *
3893 rxi_ReceiveDataPacket(struct rx_call *call,
3894 struct rx_packet *np, int istack,
3895 osi_socket socket, afs_uint32 host, u_short port,
3896 int *tnop, struct rx_call **newcallp)
3898 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3903 afs_uint32 serial=0, flags=0;
3905 struct rx_packet *tnp;
3906 if (rx_stats_active)
3907 rx_atomic_inc(&rx_stats.dataPacketsRead);
3910 /* If there are no packet buffers, drop this new packet, unless we can find
3911 * packet buffers from inactive calls */
3913 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3914 MUTEX_ENTER(&rx_freePktQ_lock);
3915 rxi_NeedMorePackets = TRUE;
3916 MUTEX_EXIT(&rx_freePktQ_lock);
3917 if (rx_stats_active)
3918 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3919 rxi_calltrace(RX_TRACE_DROP, call);
3920 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3921 /* We used to clear the receive queue here, in an attempt to free
3922 * packets. However this is unsafe if the queue has received a
3923 * soft ACK for the final packet */
3924 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3930 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3931 * packet is one of several packets transmitted as a single
3932 * datagram. Do not send any soft or hard acks until all packets
3933 * in a jumbogram have been processed. Send negative acks right away.
3935 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3936 /* tnp is non-null when there are more packets in the
3937 * current jumbo gram */
3944 seq = np->header.seq;
3945 serial = np->header.serial;
3946 flags = np->header.flags;
3948 /* If the call is in an error state, send an abort message */
3950 return rxi_SendCallAbort(call, np, istack, 0);
3952 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3953 * AFS 3.5 jumbogram. */
3954 if (flags & RX_JUMBO_PACKET) {
3955 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3960 if (np->header.spare != 0) {
3961 MUTEX_ENTER(&call->conn->conn_data_lock);
3962 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3963 MUTEX_EXIT(&call->conn->conn_data_lock);
3966 /* The usual case is that this is the expected next packet */
3967 if (seq == call->rnext) {
3969 /* Check to make sure it is not a duplicate of one already queued */
3970 if (!opr_queue_IsEmpty(&call->rq)
3971 && opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq == seq) {
3972 if (rx_stats_active)
3973 rx_atomic_inc(&rx_stats.dupPacketsRead);
3974 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3975 rxevent_Cancel(&call->delayedAckEvent, call,
3976 RX_CALL_REFCOUNT_DELAY);
3977 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3983 /* It's the next packet. Stick it on the receive queue
3984 * for this call. Set newPackets to make sure we wake
3985 * the reader once all packets have been processed */
3986 #ifdef RX_TRACK_PACKETS
3987 np->flags |= RX_PKTFLAG_RQ;
3989 opr_queue_Prepend(&call->rq, &np->entry);
3990 #ifdef RXDEBUG_PACKET
3992 #endif /* RXDEBUG_PACKET */
3994 np = NULL; /* We can't use this anymore */
3997 /* If an ack is requested then set a flag to make sure we
3998 * send an acknowledgement for this packet */
3999 if (flags & RX_REQUEST_ACK) {
4000 ackNeeded = RX_ACK_REQUESTED;
4003 /* Keep track of whether we have received the last packet */
4004 if (flags & RX_LAST_PACKET) {
4005 call->flags |= RX_CALL_HAVE_LAST;
4009 /* Check whether we have all of the packets for this call */
4010 if (call->flags & RX_CALL_HAVE_LAST) {
4011 afs_uint32 tseq; /* temporary sequence number */
4012 struct opr_queue *cursor;
4014 for (tseq = seq, opr_queue_Scan(&call->rq, cursor)) {
4015 struct rx_packet *tp;
4017 tp = opr_queue_Entry(cursor, struct rx_packet, entry);
4018 if (tseq != tp->header.seq)
4020 if (tp->header.flags & RX_LAST_PACKET) {
4021 call->flags |= RX_CALL_RECEIVE_DONE;
4028 /* Provide asynchronous notification for those who want it
4029 * (e.g. multi rx) */
4030 if (call->arrivalProc) {
4031 (*call->arrivalProc) (call, call->arrivalProcHandle,
4032 call->arrivalProcArg);
4033 call->arrivalProc = (void (*)())0;
4036 /* Update last packet received */
4039 /* If there is no server process serving this call, grab
4040 * one, if available. We only need to do this once. If a
4041 * server thread is available, this thread becomes a server
4042 * thread and the server thread becomes a listener thread. */
4044 TryAttach(call, socket, tnop, newcallp, 0);
4047 /* This is not the expected next packet. */
4049 /* Determine whether this is a new or old packet, and if it's
4050 * a new one, whether it fits into the current receive window.
4051 * Also figure out whether the packet was delivered in sequence.
4052 * We use the prev variable to determine whether the new packet
4053 * is the successor of its immediate predecessor in the
4054 * receive queue, and the missing flag to determine whether
4055 * any of this packets predecessors are missing. */
4057 afs_uint32 prev; /* "Previous packet" sequence number */
4058 struct opr_queue *cursor;
4059 int missing; /* Are any predecessors missing? */
4061 /* If the new packet's sequence number has been sent to the
4062 * application already, then this is a duplicate */
4063 if (seq < call->rnext) {
4064 if (rx_stats_active)
4065 rx_atomic_inc(&rx_stats.dupPacketsRead);
4066 rxevent_Cancel(&call->delayedAckEvent, call,
4067 RX_CALL_REFCOUNT_DELAY);
4068 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
4074 /* If the sequence number is greater than what can be
4075 * accomodated by the current window, then send a negative
4076 * acknowledge and drop the packet */
4077 if ((call->rnext + call->rwind) <= seq) {
4078 rxevent_Cancel(&call->delayedAckEvent, call,
4079 RX_CALL_REFCOUNT_DELAY);
4080 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
4087 /* Look for the packet in the queue of old received packets */
4088 prev = call->rnext - 1;
4090 for (opr_queue_Scan(&call->rq, cursor)) {
4091 struct rx_packet *tp
4092 = opr_queue_Entry(cursor, struct rx_packet, entry);
4094 /*Check for duplicate packet */
4095 if (seq == tp->header.seq) {
4096 if (rx_stats_active)
4097 rx_atomic_inc(&rx_stats.dupPacketsRead);
4098 rxevent_Cancel(&call->delayedAckEvent, call,
4099 RX_CALL_REFCOUNT_DELAY);
4100 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
4106 /* If we find a higher sequence packet, break out and
4107 * insert the new packet here. */
4108 if (seq < tp->header.seq)
4110 /* Check for missing packet */
4111 if (tp->header.seq != prev + 1) {
4115 prev = tp->header.seq;
4118 /* Keep track of whether we have received the last packet. */
4119 if (flags & RX_LAST_PACKET) {
4120 call->flags |= RX_CALL_HAVE_LAST;
4123 /* It's within the window: add it to the the receive queue.
4124 * tp is left by the previous loop either pointing at the
4125 * packet before which to insert the new packet, or at the
4126 * queue head if the queue is empty or the packet should be
4128 #ifdef RX_TRACK_PACKETS
4129 np->flags |= RX_PKTFLAG_RQ;
4131 #ifdef RXDEBUG_PACKET
4133 #endif /* RXDEBUG_PACKET */
4134 opr_queue_InsertBefore(cursor, &np->entry);
4138 /* Check whether we have all of the packets for this call */
4139 if ((call->flags & RX_CALL_HAVE_LAST)
4140 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4141 afs_uint32 tseq; /* temporary sequence number */
4144 for (opr_queue_Scan(&call->rq, cursor)) {
4145 struct rx_packet *tp
4146 = opr_queue_Entry(cursor, struct rx_packet, entry);
4147 if (tseq != tp->header.seq)
4149 if (tp->header.flags & RX_LAST_PACKET) {
4150 call->flags |= RX_CALL_RECEIVE_DONE;
4157 /* We need to send an ack of the packet is out of sequence,
4158 * or if an ack was requested by the peer. */
4159 if (seq != prev + 1 || missing) {
4160 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4161 } else if (flags & RX_REQUEST_ACK) {
4162 ackNeeded = RX_ACK_REQUESTED;
4165 /* Acknowledge the last packet for each call */
4166 if (flags & RX_LAST_PACKET) {
4177 * If the receiver is waiting for an iovec, fill the iovec
4178 * using the data from the receive queue */
4179 if (call->flags & RX_CALL_IOVEC_WAIT) {
4180 didHardAck = rxi_FillReadVec(call, serial);
4181 /* the call may have been aborted */
4190 /* Wakeup the reader if any */
4191 if ((call->flags & RX_CALL_READER_WAIT)
4192 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4193 || (call->iovNext >= call->iovMax)
4194 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4195 call->flags &= ~RX_CALL_READER_WAIT;
4196 #ifdef RX_ENABLE_LOCKS
4197 CV_BROADCAST(&call->cv_rq);
4199 osi_rxWakeup(&call->rq);
4205 * Send an ack when requested by the peer, or once every
4206 * rxi_SoftAckRate packets until the last packet has been
4207 * received. Always send a soft ack for the last packet in
4208 * the server's reply. */
4210 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4211 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4212 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4213 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4214 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4215 } else if (call->nSoftAcks) {
4216 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4217 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4219 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4220 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4221 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4228 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4230 struct rx_peer *peer = conn->peer;
4232 MUTEX_ENTER(&peer->peer_lock);
4233 peer->lastReachTime = clock_Sec();
4234 MUTEX_EXIT(&peer->peer_lock);
4236 MUTEX_ENTER(&conn->conn_data_lock);
4237 if (conn->flags & RX_CONN_ATTACHWAIT) {
4240 rxi_ConnClearAttachWait(conn);
4241 MUTEX_EXIT(&conn->conn_data_lock);
4243 for (i = 0; i < RX_MAXCALLS; i++) {
4244 struct rx_call *call = conn->call[i];
4247 MUTEX_ENTER(&call->lock);
4248 /* tnop can be null if newcallp is null */
4249 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4251 MUTEX_EXIT(&call->lock);
4255 MUTEX_EXIT(&conn->conn_data_lock);
4258 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4260 rx_ack_reason(int reason)
4263 case RX_ACK_REQUESTED:
4265 case RX_ACK_DUPLICATE:
4267 case RX_ACK_OUT_OF_SEQUENCE:
4269 case RX_ACK_EXCEEDS_WINDOW:
4271 case RX_ACK_NOSPACE:
4275 case RX_ACK_PING_RESPONSE:
4288 /* The real smarts of the whole thing. */
4289 static struct rx_packet *
4290 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4293 struct rx_ackPacket *ap;
4295 struct rx_packet *tp;
4296 struct rx_connection *conn = call->conn;
4297 struct rx_peer *peer = conn->peer;
4298 struct opr_queue *cursor;
4299 struct clock now; /* Current time, for RTT calculations */
4307 int newAckCount = 0;
4308 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4309 int pktsize = 0; /* Set if we need to update the peer mtu */
4310 int conn_data_locked = 0;
4312 if (rx_stats_active)
4313 rx_atomic_inc(&rx_stats.ackPacketsRead);
4314 ap = (struct rx_ackPacket *)rx_DataOf(np);
4315 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4317 return np; /* truncated ack packet */
4319 /* depends on ack packet struct */
4320 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4321 first = ntohl(ap->firstPacket);
4322 prev = ntohl(ap->previousPacket);
4323 serial = ntohl(ap->serial);
4326 * Ignore ack packets received out of order while protecting
4327 * against peers that set the previousPacket field to a packet
4328 * serial number instead of a sequence number.
4330 if (first < call->tfirst ||
4331 (first == call->tfirst && prev < call->tprev && prev < call->tfirst
4338 if (np->header.flags & RX_SLOW_START_OK) {
4339 call->flags |= RX_CALL_SLOW_START_OK;
4342 if (ap->reason == RX_ACK_PING_RESPONSE)
4343 rxi_UpdatePeerReach(conn, call);
4345 if (conn->lastPacketSizeSeq) {
4346 MUTEX_ENTER(&conn->conn_data_lock);
4347 conn_data_locked = 1;
4348 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4349 pktsize = conn->lastPacketSize;
4350 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4353 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4354 if (!conn_data_locked) {
4355 MUTEX_ENTER(&conn->conn_data_lock);
4356 conn_data_locked = 1;
4358 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4359 /* process mtu ping ack */
4360 pktsize = conn->lastPingSize;
4361 conn->lastPingSizeSer = conn->lastPingSize = 0;
4365 if (conn_data_locked) {
4366 MUTEX_EXIT(&conn->conn_data_lock);
4367 conn_data_locked = 0;
4371 if (rxdebug_active) {
4375 len = _snprintf(msg, sizeof(msg),
4376 "tid[%d] RACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4377 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4378 ntohl(ap->serial), ntohl(ap->previousPacket),
4379 (unsigned int)np->header.seq, ntohl(ap->firstPacket),
4380 ap->nAcks, ntohs(ap->bufferSpace) );
4384 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4385 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4389 OutputDebugString(msg);
4391 #else /* AFS_NT40_ENV */
4394 "RACK: reason %x previous %u seq %u serial %u first %u",
4395 ap->reason, ntohl(ap->previousPacket),
4396 (unsigned int)np->header.seq, (unsigned int)serial,
4397 ntohl(ap->firstPacket));
4400 for (offset = 0; offset < nAcks; offset++)
4401 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4406 #endif /* AFS_NT40_ENV */
4409 MUTEX_ENTER(&peer->peer_lock);
4412 * Start somewhere. Can't assume we can send what we can receive,
4413 * but we are clearly receiving.
4415 if (!peer->maxPacketSize)
4416 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4418 if (pktsize > peer->maxPacketSize) {
4419 peer->maxPacketSize = pktsize;
4420 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4421 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4422 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4423 rxi_ScheduleGrowMTUEvent(call, 1);
4428 clock_GetTime(&now);
4430 /* The transmit queue splits into 4 sections.
4432 * The first section is packets which have now been acknowledged
4433 * by a window size change in the ack. These have reached the
4434 * application layer, and may be discarded. These are packets
4435 * with sequence numbers < ap->firstPacket.
4437 * The second section is packets which have sequence numbers in
4438 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4439 * contents of the packet's ack array determines whether these
4440 * packets are acknowledged or not.
4442 * The third section is packets which fall above the range
4443 * addressed in the ack packet. These have not yet been received
4446 * The four section is packets which have not yet been transmitted.
4447 * These packets will have a header.serial of 0.
4450 /* First section - implicitly acknowledged packets that can be
4454 tp = opr_queue_First(&call->tq, struct rx_packet, entry);
4455 while(!opr_queue_IsEnd(&call->tq, &tp->entry) && tp->header.seq < first) {
4456 struct rx_packet *next;
4458 next = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4459 call->tfirst = tp->header.seq + 1;
4461 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4463 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4466 #ifdef RX_ENABLE_LOCKS
4467 /* XXX Hack. Because we have to release the global call lock when sending
4468 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4469 * in rxi_Start sending packets out because packets may move to the
4470 * freePacketQueue as result of being here! So we drop these packets until
4471 * we're safely out of the traversing. Really ugly!
4472 * To make it even uglier, if we're using fine grain locking, we can
4473 * set the ack bits in the packets and have rxi_Start remove the packets
4474 * when it's done transmitting.
4476 if (call->flags & RX_CALL_TQ_BUSY) {
4477 tp->flags |= RX_PKTFLAG_ACKED;
4478 call->flags |= RX_CALL_TQ_SOME_ACKED;
4480 #endif /* RX_ENABLE_LOCKS */
4482 opr_queue_Remove(&tp->entry);
4483 #ifdef RX_TRACK_PACKETS
4484 tp->flags &= ~RX_PKTFLAG_TQ;
4486 #ifdef RXDEBUG_PACKET
4488 #endif /* RXDEBUG_PACKET */
4489 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4494 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4496 /* Second section of the queue - packets for which we are receiving
4499 * Go through the explicit acks/nacks and record the results in
4500 * the waiting packets. These are packets that can't be released
4501 * yet, even with a positive acknowledge. This positive
4502 * acknowledge only means the packet has been received by the
4503 * peer, not that it will be retained long enough to be sent to
4504 * the peer's upper level. In addition, reset the transmit timers
4505 * of any missing packets (those packets that must be missing
4506 * because this packet was out of sequence) */
4508 call->nSoftAcked = 0;
4510 while (!opr_queue_IsEnd(&call->tq, &tp->entry)
4511 && tp->header.seq < first + nAcks) {
4512 /* Set the acknowledge flag per packet based on the
4513 * information in the ack packet. An acknowlegded packet can
4514 * be downgraded when the server has discarded a packet it
4515 * soacked previously, or when an ack packet is received
4516 * out of sequence. */
4517 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4518 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4520 tp->flags |= RX_PKTFLAG_ACKED;
4521 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4528 } else /* RX_ACK_TYPE_NACK */ {
4529 tp->flags &= ~RX_PKTFLAG_ACKED;
4533 tp = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4536 /* We don't need to take any action with the 3rd or 4th section in the
4537 * queue - they're not addressed by the contents of this ACK packet.
4540 /* If the window has been extended by this acknowledge packet,
4541 * then wakeup a sender waiting in alloc for window space, or try
4542 * sending packets now, if he's been sitting on packets due to
4543 * lack of window space */
4544 if (call->tnext < (call->tfirst + call->twind)) {
4545 #ifdef RX_ENABLE_LOCKS
4546 CV_SIGNAL(&call->cv_twind);
4548 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4549 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4550 osi_rxWakeup(&call->twind);
4553 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4554 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4558 /* if the ack packet has a receivelen field hanging off it,
4559 * update our state */
4560 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4563 /* If the ack packet has a "recommended" size that is less than
4564 * what I am using now, reduce my size to match */
4565 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4566 (int)sizeof(afs_int32), &tSize);
4567 tSize = (afs_uint32) ntohl(tSize);
4568 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4570 /* Get the maximum packet size to send to this peer */
4571 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4573 tSize = (afs_uint32) ntohl(tSize);
4574 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4575 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4577 /* sanity check - peer might have restarted with different params.
4578 * If peer says "send less", dammit, send less... Peer should never
4579 * be unable to accept packets of the size that prior AFS versions would
4580 * send without asking. */
4581 if (peer->maxMTU != tSize) {
4582 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4584 peer->maxMTU = tSize;
4585 peer->MTU = MIN(tSize, peer->MTU);
4586 call->MTU = MIN(call->MTU, tSize);
4589 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4592 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4593 (int)sizeof(afs_int32), &tSize);
4594 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4595 if (tSize < call->twind) { /* smaller than our send */
4596 call->twind = tSize; /* window, we must send less... */
4597 call->ssthresh = MIN(call->twind, call->ssthresh);
4598 call->conn->twind[call->channel] = call->twind;
4601 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4602 * network MTU confused with the loopback MTU. Calculate the
4603 * maximum MTU here for use in the slow start code below.
4605 /* Did peer restart with older RX version? */
4606 if (peer->maxDgramPackets > 1) {
4607 peer->maxDgramPackets = 1;
4609 } else if (np->length >=
4610 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4613 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4614 sizeof(afs_int32), &tSize);
4615 tSize = (afs_uint32) ntohl(tSize);
4617 * As of AFS 3.5 we set the send window to match the receive window.
4619 if (tSize < call->twind) {
4620 call->twind = tSize;
4621 call->conn->twind[call->channel] = call->twind;
4622 call->ssthresh = MIN(call->twind, call->ssthresh);
4623 } else if (tSize > call->twind) {
4624 call->twind = tSize;
4625 call->conn->twind[call->channel] = call->twind;
4629 * As of AFS 3.5, a jumbogram is more than one fixed size
4630 * packet transmitted in a single UDP datagram. If the remote
4631 * MTU is smaller than our local MTU then never send a datagram
4632 * larger than the natural MTU.
4635 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4636 (int)sizeof(afs_int32), &tSize);
4637 maxDgramPackets = (afs_uint32) ntohl(tSize);
4638 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4640 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4641 if (maxDgramPackets > 1) {
4642 peer->maxDgramPackets = maxDgramPackets;
4643 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4645 peer->maxDgramPackets = 1;
4646 call->MTU = peer->natMTU;
4648 } else if (peer->maxDgramPackets > 1) {
4649 /* Restarted with lower version of RX */
4650 peer->maxDgramPackets = 1;
4652 } else if (peer->maxDgramPackets > 1
4653 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4654 /* Restarted with lower version of RX */
4655 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4656 peer->natMTU = OLD_MAX_PACKET_SIZE;
4657 peer->MTU = OLD_MAX_PACKET_SIZE;
4658 peer->maxDgramPackets = 1;
4659 peer->nDgramPackets = 1;
4661 call->MTU = OLD_MAX_PACKET_SIZE;
4666 * Calculate how many datagrams were successfully received after
4667 * the first missing packet and adjust the negative ack counter
4672 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4673 if (call->nNacks < nNacked) {
4674 call->nNacks = nNacked;
4677 call->nAcks += newAckCount;
4681 /* If the packet contained new acknowledgements, rather than just
4682 * being a duplicate of one we have previously seen, then we can restart
4685 if (newAckCount > 0)
4686 rxi_rto_packet_acked(call, istack);
4688 if (call->flags & RX_CALL_FAST_RECOVER) {
4689 if (newAckCount == 0) {
4690 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4692 call->flags &= ~RX_CALL_FAST_RECOVER;
4693 call->cwind = call->nextCwind;
4694 call->nextCwind = 0;
4697 call->nCwindAcks = 0;
4698 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4699 /* Three negative acks in a row trigger congestion recovery */
4700 call->flags |= RX_CALL_FAST_RECOVER;
4701 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4703 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4704 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4705 call->nextCwind = call->ssthresh;
4708 peer->MTU = call->MTU;
4709 peer->cwind = call->nextCwind;
4710 peer->nDgramPackets = call->nDgramPackets;
4712 call->congestSeq = peer->congestSeq;
4714 /* Reset the resend times on the packets that were nacked
4715 * so we will retransmit as soon as the window permits
4719 for (opr_queue_ScanBackwards(&call->tq, cursor)) {
4720 struct rx_packet *tp =
4721 opr_queue_Entry(cursor, struct rx_packet, entry);
4723 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4724 tp->flags &= ~RX_PKTFLAG_SENT;
4726 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4731 /* If cwind is smaller than ssthresh, then increase
4732 * the window one packet for each ack we receive (exponential
4734 * If cwind is greater than or equal to ssthresh then increase
4735 * the congestion window by one packet for each cwind acks we
4736 * receive (linear growth). */
4737 if (call->cwind < call->ssthresh) {
4739 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4740 call->nCwindAcks = 0;
4742 call->nCwindAcks += newAckCount;
4743 if (call->nCwindAcks >= call->cwind) {
4744 call->nCwindAcks = 0;
4745 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4749 * If we have received several acknowledgements in a row then
4750 * it is time to increase the size of our datagrams
4752 if ((int)call->nAcks > rx_nDgramThreshold) {
4753 if (peer->maxDgramPackets > 1) {
4754 if (call->nDgramPackets < peer->maxDgramPackets) {
4755 call->nDgramPackets++;
4757 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4758 } else if (call->MTU < peer->maxMTU) {
4759 /* don't upgrade if we can't handle it */
4760 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4761 call->MTU = peer->ifMTU;
4763 call->MTU += peer->natMTU;
4764 call->MTU = MIN(call->MTU, peer->maxMTU);
4771 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4773 /* Servers need to hold the call until all response packets have
4774 * been acknowledged. Soft acks are good enough since clients
4775 * are not allowed to clear their receive queues. */
4776 if (call->state == RX_STATE_HOLD
4777 && call->tfirst + call->nSoftAcked >= call->tnext) {
4778 call->state = RX_STATE_DALLY;
4779 rxi_ClearTransmitQueue(call, 0);
4780 rxevent_Cancel(&call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4781 } else if (!opr_queue_IsEmpty(&call->tq)) {
4782 rxi_Start(call, istack);
4787 /* Received a response to a challenge packet */
4788 static struct rx_packet *
4789 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4790 struct rx_packet *np, int istack)
4794 /* Ignore the packet if we're the client */
4795 if (conn->type == RX_CLIENT_CONNECTION)
4798 /* If already authenticated, ignore the packet (it's probably a retry) */
4799 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4802 /* Otherwise, have the security object evaluate the response packet */
4803 error = RXS_CheckResponse(conn->securityObject, conn, np);
4805 /* If the response is invalid, reset the connection, sending
4806 * an abort to the peer */
4810 rxi_ConnectionError(conn, error);
4811 MUTEX_ENTER(&conn->conn_data_lock);
4812 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4813 MUTEX_EXIT(&conn->conn_data_lock);
4816 /* If the response is valid, any calls waiting to attach
4817 * servers can now do so */
4820 for (i = 0; i < RX_MAXCALLS; i++) {
4821 struct rx_call *call = conn->call[i];
4823 MUTEX_ENTER(&call->lock);
4824 if (call->state == RX_STATE_PRECALL)
4825 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4826 /* tnop can be null if newcallp is null */
4827 MUTEX_EXIT(&call->lock);
4831 /* Update the peer reachability information, just in case
4832 * some calls went into attach-wait while we were waiting
4833 * for authentication..
4835 rxi_UpdatePeerReach(conn, NULL);
4840 /* A client has received an authentication challenge: the security
4841 * object is asked to cough up a respectable response packet to send
4842 * back to the server. The server is responsible for retrying the
4843 * challenge if it fails to get a response. */
4845 static struct rx_packet *
4846 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4847 struct rx_packet *np, int istack)
4851 /* Ignore the challenge if we're the server */
4852 if (conn->type == RX_SERVER_CONNECTION)
4855 /* Ignore the challenge if the connection is otherwise idle; someone's
4856 * trying to use us as an oracle. */
4857 if (!rxi_HasActiveCalls(conn))
4860 /* Send the security object the challenge packet. It is expected to fill
4861 * in the response. */
4862 error = RXS_GetResponse(conn->securityObject, conn, np);
4864 /* If the security object is unable to return a valid response, reset the
4865 * connection and send an abort to the peer. Otherwise send the response
4866 * packet to the peer connection. */
4868 rxi_ConnectionError(conn, error);
4869 MUTEX_ENTER(&conn->conn_data_lock);
4870 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4871 MUTEX_EXIT(&conn->conn_data_lock);
4873 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4874 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4880 /* Find an available server process to service the current request in
4881 * the given call structure. If one isn't available, queue up this
4882 * call so it eventually gets one */
4884 rxi_AttachServerProc(struct rx_call *call,
4885 osi_socket socket, int *tnop,
4886 struct rx_call **newcallp)
4888 struct rx_serverQueueEntry *sq;
4889 struct rx_service *service = call->conn->service;
4892 /* May already be attached */
4893 if (call->state == RX_STATE_ACTIVE)
4896 MUTEX_ENTER(&rx_serverPool_lock);
4898 haveQuota = QuotaOK(service);
4899 if ((!haveQuota) || opr_queue_IsEmpty(&rx_idleServerQueue)) {
4900 /* If there are no processes available to service this call,
4901 * put the call on the incoming call queue (unless it's
4902 * already on the queue).
4904 #ifdef RX_ENABLE_LOCKS
4906 ReturnToServerPool(service);
4907 #endif /* RX_ENABLE_LOCKS */
4909 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4910 call->flags |= RX_CALL_WAIT_PROC;
4911 rx_atomic_inc(&rx_nWaiting);
4912 rx_atomic_inc(&rx_nWaited);
4913 rxi_calltrace(RX_CALL_ARRIVAL, call);
4914 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4915 opr_queue_Append(&rx_incomingCallQueue, &call->entry);
4918 sq = opr_queue_Last(&rx_idleServerQueue,
4919 struct rx_serverQueueEntry, entry);
4921 /* If hot threads are enabled, and both newcallp and sq->socketp
4922 * are non-null, then this thread will process the call, and the
4923 * idle server thread will start listening on this threads socket.
4925 opr_queue_Remove(&sq->entry);
4927 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4930 *sq->socketp = socket;
4931 clock_GetTime(&call->startTime);
4932 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4936 if (call->flags & RX_CALL_WAIT_PROC) {
4937 /* Conservative: I don't think this should happen */
4938 call->flags &= ~RX_CALL_WAIT_PROC;
4939 rx_atomic_dec(&rx_nWaiting);
4940 if (opr_queue_IsOnQueue(&call->entry)) {
4941 opr_queue_Remove(&call->entry);
4944 call->state = RX_STATE_ACTIVE;
4945 call->app.mode = RX_MODE_RECEIVING;
4946 #ifdef RX_KERNEL_TRACE
4948 int glockOwner = ISAFS_GLOCK();
4951 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4952 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4958 if (call->flags & RX_CALL_CLEARED) {
4959 /* send an ack now to start the packet flow up again */
4960 call->flags &= ~RX_CALL_CLEARED;
4961 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4963 #ifdef RX_ENABLE_LOCKS
4966 service->nRequestsRunning++;
4967 MUTEX_ENTER(&rx_quota_mutex);
4968 if (service->nRequestsRunning <= service->minProcs)
4971 MUTEX_EXIT(&rx_quota_mutex);
4975 MUTEX_EXIT(&rx_serverPool_lock);
4978 /* Delay the sending of an acknowledge event for a short while, while
4979 * a new call is being prepared (in the case of a client) or a reply
4980 * is being prepared (in the case of a server). Rather than sending
4981 * an ack packet, an ACKALL packet is sent. */
4983 rxi_AckAll(struct rx_call *call)
4985 rxi_SendSpecial(call, call->conn, NULL, RX_PACKET_TYPE_ACKALL,
4987 call->flags |= RX_CALL_ACKALL_SENT;
4991 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4994 struct rx_call *call = arg1;
4995 #ifdef RX_ENABLE_LOCKS
4997 MUTEX_ENTER(&call->lock);
4998 if (event == call->delayedAckEvent) {
4999 rxevent_Put(call->delayedAckEvent);
5000 call->delayedAckEvent = NULL;
5002 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
5004 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
5006 MUTEX_EXIT(&call->lock);
5007 #else /* RX_ENABLE_LOCKS */
5009 rxevent_Put(call->delayedAckEvent);
5010 call->delayedAckEvent = NULL;
5012 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
5013 #endif /* RX_ENABLE_LOCKS */
5017 #ifdef RX_ENABLE_LOCKS
5018 /* Set ack in all packets in transmit queue. rxi_Start will deal with
5019 * clearing them out.
5022 rxi_SetAcksInTransmitQueue(struct rx_call *call)
5024 struct opr_queue *cursor;
5027 for (opr_queue_Scan(&call->tq, cursor)) {
5029 = opr_queue_Entry(cursor, struct rx_packet, entry);
5031 p->flags |= RX_PKTFLAG_ACKED;
5036 call->flags |= RX_CALL_TQ_CLEARME;
5037 call->flags |= RX_CALL_TQ_SOME_ACKED;
5040 rxi_rto_cancel(call);
5042 call->tfirst = call->tnext;
5043 call->nSoftAcked = 0;
5045 if (call->flags & RX_CALL_FAST_RECOVER) {
5046 call->flags &= ~RX_CALL_FAST_RECOVER;
5047 call->cwind = call->nextCwind;
5048 call->nextCwind = 0;
5051 CV_SIGNAL(&call->cv_twind);
5053 #endif /* RX_ENABLE_LOCKS */
5055 /* Clear out the transmit queue for the current call (all packets have
5056 * been received by peer) */
5058 rxi_ClearTransmitQueue(struct rx_call *call, int force)
5060 #ifdef RX_ENABLE_LOCKS
5061 struct opr_queue *cursor;
5062 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
5064 for (opr_queue_Scan(&call->tq, cursor)) {
5066 = opr_queue_Entry(cursor, struct rx_packet, entry);
5068 p->flags |= RX_PKTFLAG_ACKED;
5072 call->flags |= RX_CALL_TQ_CLEARME;
5073 call->flags |= RX_CALL_TQ_SOME_ACKED;
5076 #endif /* RX_ENABLE_LOCKS */
5077 #ifdef RXDEBUG_PACKET
5079 #endif /* RXDEBUG_PACKET */
5080 rxi_FreePackets(0, &call->tq);
5081 rxi_WakeUpTransmitQueue(call);
5082 #ifdef RX_ENABLE_LOCKS
5083 call->flags &= ~RX_CALL_TQ_CLEARME;
5087 rxi_rto_cancel(call);
5088 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5089 call->nSoftAcked = 0;
5091 if (call->flags & RX_CALL_FAST_RECOVER) {
5092 call->flags &= ~RX_CALL_FAST_RECOVER;
5093 call->cwind = call->nextCwind;
5095 #ifdef RX_ENABLE_LOCKS
5096 CV_SIGNAL(&call->cv_twind);
5098 osi_rxWakeup(&call->twind);
5103 rxi_ClearReceiveQueue(struct rx_call *call)
5105 if (!opr_queue_IsEmpty(&call->rq)) {
5108 count = rxi_FreePackets(0, &call->rq);
5109 rx_packetReclaims += count;
5110 #ifdef RXDEBUG_PACKET
5112 if ( call->rqc != 0 )
5113 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5115 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5117 if (call->state == RX_STATE_PRECALL) {
5118 call->flags |= RX_CALL_CLEARED;
5122 /* Send an abort packet for the specified call */
5123 static struct rx_packet *
5124 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5125 int istack, int force)
5127 afs_int32 error, cerror;
5128 struct clock when, now;
5133 switch (call->error) {
5136 cerror = RX_CALL_TIMEOUT;
5139 cerror = call->error;
5142 /* Clients should never delay abort messages */
5143 if (rx_IsClientConn(call->conn))
5146 if (call->abortCode != cerror) {
5147 call->abortCode = cerror;
5148 call->abortCount = 0;
5151 if (force || rxi_callAbortThreshhold == 0
5152 || call->abortCount < rxi_callAbortThreshhold) {
5153 if (call->delayedAbortEvent) {
5154 rxevent_Cancel(&call->delayedAbortEvent, call,
5155 RX_CALL_REFCOUNT_ABORT);
5157 error = htonl(cerror);
5160 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5161 (char *)&error, sizeof(error), istack);
5162 } else if (!call->delayedAbortEvent) {
5163 clock_GetTime(&now);
5165 clock_Addmsec(&when, rxi_callAbortDelay);
5166 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5167 call->delayedAbortEvent =
5168 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5173 /* Send an abort packet for the specified connection. Packet is an
5174 * optional pointer to a packet that can be used to send the abort.
5175 * Once the number of abort messages reaches the threshhold, an
5176 * event is scheduled to send the abort. Setting the force flag
5177 * overrides sending delayed abort messages.
5179 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5180 * to send the abort packet.
5183 rxi_SendConnectionAbort(struct rx_connection *conn,
5184 struct rx_packet *packet, int istack, int force)
5187 struct clock when, now;
5192 /* Clients should never delay abort messages */
5193 if (rx_IsClientConn(conn))
5196 if (force || rxi_connAbortThreshhold == 0
5197 || conn->abortCount < rxi_connAbortThreshhold) {
5199 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
5200 error = htonl(conn->error);
5202 MUTEX_EXIT(&conn->conn_data_lock);
5204 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5205 RX_PACKET_TYPE_ABORT, (char *)&error,
5206 sizeof(error), istack);
5207 MUTEX_ENTER(&conn->conn_data_lock);
5208 } else if (!conn->delayedAbortEvent) {
5209 clock_GetTime(&now);
5211 clock_Addmsec(&when, rxi_connAbortDelay);
5212 conn->delayedAbortEvent =
5213 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
5218 /* Associate an error all of the calls owned by a connection. Called
5219 * with error non-zero. This is only for really fatal things, like
5220 * bad authentication responses. The connection itself is set in
5221 * error at this point, so that future packets received will be
5224 rxi_ConnectionError(struct rx_connection *conn,
5230 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5232 MUTEX_ENTER(&conn->conn_data_lock);
5233 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
5234 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
5235 if (conn->checkReachEvent) {
5236 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
5237 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5238 putConnection(conn);
5240 MUTEX_EXIT(&conn->conn_data_lock);
5241 for (i = 0; i < RX_MAXCALLS; i++) {
5242 struct rx_call *call = conn->call[i];
5244 MUTEX_ENTER(&call->lock);
5245 rxi_CallError(call, error);
5246 MUTEX_EXIT(&call->lock);
5249 conn->error = error;
5250 if (rx_stats_active)
5251 rx_atomic_inc(&rx_stats.fatalErrors);
5256 * Interrupt an in-progress call with the specified error and wakeup waiters.
5258 * @param[in] call The call to interrupt
5259 * @param[in] error The error code to send to the peer
5262 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5264 MUTEX_ENTER(&call->lock);
5265 rxi_CallError(call, error);
5266 rxi_SendCallAbort(call, NULL, 0, 1);
5267 MUTEX_EXIT(&call->lock);
5271 rxi_CallError(struct rx_call *call, afs_int32 error)
5274 osirx_AssertMine(&call->lock, "rxi_CallError");
5276 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5278 error = call->error;
5280 #ifdef RX_ENABLE_LOCKS
5281 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5282 rxi_ResetCall(call, 0);
5285 rxi_ResetCall(call, 0);
5287 call->error = error;
5290 /* Reset various fields in a call structure, and wakeup waiting
5291 * processes. Some fields aren't changed: state & mode are not
5292 * touched (these must be set by the caller), and bufptr, nLeft, and
5293 * nFree are not reset, since these fields are manipulated by
5294 * unprotected macros, and may only be reset by non-interrupting code.
5298 rxi_ResetCall(struct rx_call *call, int newcall)
5301 struct rx_peer *peer;
5302 struct rx_packet *packet;
5304 osirx_AssertMine(&call->lock, "rxi_ResetCall");
5306 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5308 /* Notify anyone who is waiting for asynchronous packet arrival */
5309 if (call->arrivalProc) {
5310 (*call->arrivalProc) (call, call->arrivalProcHandle,
5311 call->arrivalProcArg);
5312 call->arrivalProc = (void (*)())0;
5316 rxevent_Cancel(&call->growMTUEvent, call, RX_CALL_REFCOUNT_MTU);
5318 if (call->delayedAbortEvent) {
5319 rxevent_Cancel(&call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
5320 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5322 rxi_SendCallAbort(call, packet, 0, 1);
5323 rxi_FreePacket(packet);
5328 * Update the peer with the congestion information in this call
5329 * so other calls on this connection can pick up where this call
5330 * left off. If the congestion sequence numbers don't match then
5331 * another call experienced a retransmission.
5333 peer = call->conn->peer;
5334 MUTEX_ENTER(&peer->peer_lock);
5336 if (call->congestSeq == peer->congestSeq) {
5337 peer->cwind = MAX(peer->cwind, call->cwind);
5338 peer->MTU = MAX(peer->MTU, call->MTU);
5339 peer->nDgramPackets =
5340 MAX(peer->nDgramPackets, call->nDgramPackets);
5343 call->abortCode = 0;
5344 call->abortCount = 0;
5346 if (peer->maxDgramPackets > 1) {
5347 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5349 call->MTU = peer->MTU;
5351 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5352 call->ssthresh = rx_maxSendWindow;
5353 call->nDgramPackets = peer->nDgramPackets;
5354 call->congestSeq = peer->congestSeq;
5355 call->rtt = peer->rtt;
5356 call->rtt_dev = peer->rtt_dev;
5357 clock_Zero(&call->rto);
5358 clock_Addmsec(&call->rto,
5359 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5360 MUTEX_EXIT(&peer->peer_lock);
5362 flags = call->flags;
5363 rxi_WaitforTQBusy(call);
5365 rxi_ClearTransmitQueue(call, 1);
5366 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5367 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5371 if (!newcall && (flags & RX_CALL_PEER_BUSY)) {
5372 /* The call channel is still busy; resetting the call doesn't change
5373 * that. However, if 'newcall' is set, we are processing a call
5374 * structure that has either been recycled from the free list, or has
5375 * been newly allocated. So, RX_CALL_PEER_BUSY is not relevant if
5376 * 'newcall' is set, since it describes a completely different call
5377 * channel which we do not care about. */
5378 call->flags |= RX_CALL_PEER_BUSY;
5381 rxi_ClearReceiveQueue(call);
5382 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5386 call->twind = call->conn->twind[call->channel];
5387 call->rwind = call->conn->rwind[call->channel];
5388 call->nSoftAcked = 0;
5389 call->nextCwind = 0;
5392 call->nCwindAcks = 0;
5393 call->nSoftAcks = 0;
5394 call->nHardAcks = 0;
5396 call->tfirst = call->rnext = call->tnext = 1;
5399 call->lastAcked = 0;
5400 call->localStatus = call->remoteStatus = 0;
5402 if (flags & RX_CALL_READER_WAIT) {
5403 #ifdef RX_ENABLE_LOCKS
5404 CV_BROADCAST(&call->cv_rq);
5406 osi_rxWakeup(&call->rq);
5409 if (flags & RX_CALL_WAIT_PACKETS) {
5410 MUTEX_ENTER(&rx_freePktQ_lock);
5411 rxi_PacketsUnWait(); /* XXX */
5412 MUTEX_EXIT(&rx_freePktQ_lock);
5414 #ifdef RX_ENABLE_LOCKS
5415 CV_SIGNAL(&call->cv_twind);
5417 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5418 osi_rxWakeup(&call->twind);
5421 if (flags & RX_CALL_WAIT_PROC) {
5422 rx_atomic_dec(&rx_nWaiting);
5424 #ifdef RX_ENABLE_LOCKS
5425 /* The following ensures that we don't mess with any queue while some
5426 * other thread might also be doing so. The call_queue_lock field is
5427 * is only modified under the call lock. If the call is in the process
5428 * of being removed from a queue, the call is not locked until the
5429 * the queue lock is dropped and only then is the call_queue_lock field
5430 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5431 * Note that any other routine which removes a call from a queue has to
5432 * obtain the queue lock before examing the queue and removing the call.
5434 if (call->call_queue_lock) {
5435 MUTEX_ENTER(call->call_queue_lock);
5436 if (opr_queue_IsOnQueue(&call->entry)) {
5437 opr_queue_Remove(&call->entry);
5439 MUTEX_EXIT(call->call_queue_lock);
5440 CLEAR_CALL_QUEUE_LOCK(call);
5442 #else /* RX_ENABLE_LOCKS */
5443 if (opr_queue_IsOnQueue(&call->entry)) {
5444 opr_queue_Remove(&call->entry);
5446 #endif /* RX_ENABLE_LOCKS */
5448 rxi_KeepAliveOff(call);
5449 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5452 /* Send an acknowledge for the indicated packet (seq,serial) of the
5453 * indicated call, for the indicated reason (reason). This
5454 * acknowledge will specifically acknowledge receiving the packet, and
5455 * will also specify which other packets for this call have been
5456 * received. This routine returns the packet that was used to the
5457 * caller. The caller is responsible for freeing it or re-using it.
5458 * This acknowledgement also returns the highest sequence number
5459 * actually read out by the higher level to the sender; the sender
5460 * promises to keep around packets that have not been read by the
5461 * higher level yet (unless, of course, the sender decides to abort
5462 * the call altogether). Any of p, seq, serial, pflags, or reason may
5463 * be set to zero without ill effect. That is, if they are zero, they
5464 * will not convey any information.
5465 * NOW there is a trailer field, after the ack where it will safely be
5466 * ignored by mundanes, which indicates the maximum size packet this
5467 * host can swallow. */
5469 struct rx_packet *optionalPacket; use to send ack (or null)
5470 int seq; Sequence number of the packet we are acking
5471 int serial; Serial number of the packet
5472 int pflags; Flags field from packet header
5473 int reason; Reason an acknowledge was prompted
5477 rxi_SendAck(struct rx_call *call,
5478 struct rx_packet *optionalPacket, int serial, int reason,
5481 struct rx_ackPacket *ap;
5482 struct rx_packet *p;
5483 struct opr_queue *cursor;
5486 afs_uint32 padbytes = 0;
5487 #ifdef RX_ENABLE_TSFPQ
5488 struct rx_ts_info_t * rx_ts_info;
5492 * Open the receive window once a thread starts reading packets
5494 if (call->rnext > 1) {
5495 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5498 /* Don't attempt to grow MTU if this is a critical ping */
5499 if (reason == RX_ACK_MTU) {
5500 /* keep track of per-call attempts, if we're over max, do in small
5501 * otherwise in larger? set a size to increment by, decrease
5504 if (call->conn->peer->maxPacketSize &&
5505 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5507 padbytes = call->conn->peer->maxPacketSize+16;
5509 padbytes = call->conn->peer->maxMTU + 128;
5511 /* do always try a minimum size ping */
5512 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5514 /* subtract the ack payload */
5515 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5516 reason = RX_ACK_PING;
5519 call->nHardAcks = 0;
5520 call->nSoftAcks = 0;
5521 if (call->rnext > call->lastAcked)
5522 call->lastAcked = call->rnext;
5526 rx_computelen(p, p->length); /* reset length, you never know */
5527 } /* where that's been... */
5528 #ifdef RX_ENABLE_TSFPQ
5530 RX_TS_INFO_GET(rx_ts_info);
5531 if ((p = rx_ts_info->local_special_packet)) {
5532 rx_computelen(p, p->length);
5533 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5534 rx_ts_info->local_special_packet = p;
5535 } else { /* We won't send the ack, but don't panic. */
5536 return optionalPacket;
5540 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5541 /* We won't send the ack, but don't panic. */
5542 return optionalPacket;
5547 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5550 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5551 #ifndef RX_ENABLE_TSFPQ
5552 if (!optionalPacket)
5555 return optionalPacket;
5557 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5558 if (rx_Contiguous(p) < templ) {
5559 #ifndef RX_ENABLE_TSFPQ
5560 if (!optionalPacket)
5563 return optionalPacket;
5568 /* MTUXXX failing to send an ack is very serious. We should */
5569 /* try as hard as possible to send even a partial ack; it's */
5570 /* better than nothing. */
5571 ap = (struct rx_ackPacket *)rx_DataOf(p);
5572 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5573 ap->reason = reason;
5575 /* The skew computation used to be bogus, I think it's better now. */
5576 /* We should start paying attention to skew. XXX */
5577 ap->serial = htonl(serial);
5578 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5581 * First packet not yet forwarded to reader. When ACKALL has been
5582 * sent the peer has been told that all received packets will be
5583 * delivered to the reader. The value 'rnext' is used internally
5584 * to refer to the next packet in the receive queue that must be
5585 * delivered to the reader. From the perspective of the peer it
5586 * already has so report the last sequence number plus one if there
5587 * are packets in the receive queue awaiting processing.
5589 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5590 !opr_queue_IsEmpty(&call->rq)) {
5591 ap->firstPacket = htonl(opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq + 1);
5593 ap->firstPacket = htonl(call->rnext);
5595 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5597 /* No fear of running out of ack packet here because there can only
5598 * be at most one window full of unacknowledged packets. The window
5599 * size must be constrained to be less than the maximum ack size,
5600 * of course. Also, an ack should always fit into a single packet
5601 * -- it should not ever be fragmented. */
5603 for (opr_queue_Scan(&call->rq, cursor)) {
5604 struct rx_packet *rqp
5605 = opr_queue_Entry(cursor, struct rx_packet, entry);
5607 if (!rqp || !call->rq.next
5608 || (rqp->header.seq > (call->rnext + call->rwind))) {
5609 #ifndef RX_ENABLE_TSFPQ
5610 if (!optionalPacket)
5613 rxi_CallError(call, RX_CALL_DEAD);
5614 return optionalPacket;
5617 while (rqp->header.seq > call->rnext + offset)
5618 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5619 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5621 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5622 #ifndef RX_ENABLE_TSFPQ
5623 if (!optionalPacket)
5626 rxi_CallError(call, RX_CALL_DEAD);
5627 return optionalPacket;
5633 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5635 /* these are new for AFS 3.3 */
5636 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5637 templ = htonl(templ);
5638 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5639 templ = htonl(call->conn->peer->ifMTU);
5640 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5641 sizeof(afs_int32), &templ);
5643 /* new for AFS 3.4 */
5644 templ = htonl(call->rwind);
5645 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5646 sizeof(afs_int32), &templ);
5648 /* new for AFS 3.5 */
5649 templ = htonl(call->conn->peer->ifDgramPackets);
5650 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5651 sizeof(afs_int32), &templ);
5653 p->header.serviceId = call->conn->serviceId;
5654 p->header.cid = (call->conn->cid | call->channel);
5655 p->header.callNumber = *call->callNumber;
5657 p->header.securityIndex = call->conn->securityIndex;
5658 p->header.epoch = call->conn->epoch;
5659 p->header.type = RX_PACKET_TYPE_ACK;
5660 p->header.flags = RX_SLOW_START_OK;
5661 if (reason == RX_ACK_PING) {
5662 p->header.flags |= RX_REQUEST_ACK;
5664 p->length = padbytes +
5665 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5668 /* not fast but we can potentially use this if truncated
5669 * fragments are delivered to figure out the mtu.
5671 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5672 sizeof(afs_int32), sizeof(afs_int32),
5676 if (call->conn->type == RX_CLIENT_CONNECTION)
5677 p->header.flags |= RX_CLIENT_INITIATED;
5681 if (rxdebug_active) {
5685 len = _snprintf(msg, sizeof(msg),
5686 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5687 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5688 ntohl(ap->serial), ntohl(ap->previousPacket),
5689 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5690 ap->nAcks, ntohs(ap->bufferSpace) );
5694 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5695 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5699 OutputDebugString(msg);
5701 #else /* AFS_NT40_ENV */
5703 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5704 ap->reason, ntohl(ap->previousPacket),
5705 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5707 for (offset = 0; offset < ap->nAcks; offset++)
5708 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5713 #endif /* AFS_NT40_ENV */
5716 int i, nbytes = p->length;
5718 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5719 if (nbytes <= p->wirevec[i].iov_len) {
5722 savelen = p->wirevec[i].iov_len;
5724 p->wirevec[i].iov_len = nbytes;
5726 rxi_Send(call, p, istack);
5727 p->wirevec[i].iov_len = savelen;
5731 nbytes -= p->wirevec[i].iov_len;
5734 if (rx_stats_active)
5735 rx_atomic_inc(&rx_stats.ackPacketsSent);
5736 #ifndef RX_ENABLE_TSFPQ
5737 if (!optionalPacket)
5740 return optionalPacket; /* Return packet for re-use by caller */
5744 struct rx_packet **list;
5749 /* Send all of the packets in the list in single datagram */
5751 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5752 int istack, int moreFlag)
5758 struct rx_connection *conn = call->conn;
5759 struct rx_peer *peer = conn->peer;
5761 MUTEX_ENTER(&peer->peer_lock);
5762 peer->nSent += xmit->len;
5763 if (xmit->resending)
5764 peer->reSends += xmit->len;
5765 MUTEX_EXIT(&peer->peer_lock);
5767 if (rx_stats_active) {
5768 if (xmit->resending)
5769 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5771 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5774 clock_GetTime(&now);
5776 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5780 /* Set the packet flags and schedule the resend events */
5781 /* Only request an ack for the last packet in the list */
5782 for (i = 0; i < xmit->len; i++) {
5783 struct rx_packet *packet = xmit->list[i];
5785 /* Record the time sent */
5786 packet->timeSent = now;
5787 packet->flags |= RX_PKTFLAG_SENT;
5789 /* Ask for an ack on retransmitted packets, on every other packet
5790 * if the peer doesn't support slow start. Ask for an ack on every
5791 * packet until the congestion window reaches the ack rate. */
5792 if (packet->header.serial) {
5795 packet->firstSent = now;
5796 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5797 || (!(call->flags & RX_CALL_SLOW_START_OK)
5798 && (packet->header.seq & 1)))) {
5803 /* Tag this packet as not being the last in this group,
5804 * for the receiver's benefit */
5805 if (i < xmit->len - 1 || moreFlag) {
5806 packet->header.flags |= RX_MORE_PACKETS;
5811 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5814 /* Since we're about to send a data packet to the peer, it's
5815 * safe to nuke any scheduled end-of-packets ack */
5816 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5818 MUTEX_EXIT(&call->lock);
5819 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5820 if (xmit->len > 1) {
5821 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5823 rxi_SendPacket(call, conn, xmit->list[0], istack);
5825 MUTEX_ENTER(&call->lock);
5826 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5828 /* Tell the RTO calculation engine that we have sent a packet, and
5829 * if it was the last one */
5830 rxi_rto_packet_sent(call, lastPacket, istack);
5832 /* Update last send time for this call (for keep-alive
5833 * processing), and for the connection (so that we can discover
5834 * idle connections) */
5835 conn->lastSendTime = call->lastSendTime = clock_Sec();
5836 /* Let a set of retransmits trigger an idle timeout */
5837 if (!xmit->resending)
5838 call->lastSendData = call->lastSendTime;
5841 /* When sending packets we need to follow these rules:
5842 * 1. Never send more than maxDgramPackets in a jumbogram.
5843 * 2. Never send a packet with more than two iovecs in a jumbogram.
5844 * 3. Never send a retransmitted packet in a jumbogram.
5845 * 4. Never send more than cwind/4 packets in a jumbogram
5846 * We always keep the last list we should have sent so we
5847 * can set the RX_MORE_PACKETS flags correctly.
5851 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5856 struct xmitlist working;
5857 struct xmitlist last;
5859 struct rx_peer *peer = call->conn->peer;
5860 int morePackets = 0;
5862 memset(&last, 0, sizeof(struct xmitlist));
5863 working.list = &list[0];
5865 working.resending = 0;
5867 recovery = call->flags & RX_CALL_FAST_RECOVER;
5869 for (i = 0; i < len; i++) {
5870 /* Does the current packet force us to flush the current list? */
5872 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5873 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5875 /* This sends the 'last' list and then rolls the current working
5876 * set into the 'last' one, and resets the working set */
5879 rxi_SendList(call, &last, istack, 1);
5880 /* If the call enters an error state stop sending, or if
5881 * we entered congestion recovery mode, stop sending */
5883 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5888 working.resending = 0;
5889 working.list = &list[i];
5891 /* Add the current packet to the list if it hasn't been acked.
5892 * Otherwise adjust the list pointer to skip the current packet. */
5893 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5896 if (list[i]->header.serial)
5897 working.resending = 1;
5899 /* Do we need to flush the list? */
5900 if (working.len >= (int)peer->maxDgramPackets
5901 || working.len >= (int)call->nDgramPackets
5902 || working.len >= (int)call->cwind
5903 || list[i]->header.serial
5904 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5906 rxi_SendList(call, &last, istack, 1);
5907 /* If the call enters an error state stop sending, or if
5908 * we entered congestion recovery mode, stop sending */
5910 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5915 working.resending = 0;
5916 working.list = &list[i + 1];
5919 if (working.len != 0) {
5920 osi_Panic("rxi_SendList error");
5922 working.list = &list[i + 1];
5926 /* Send the whole list when the call is in receive mode, when
5927 * the call is in eof mode, when we are in fast recovery mode,
5928 * and when we have the last packet */
5929 /* XXX - The accesses to app.mode aren't safe, as this may be called by
5930 * the listener or event threads
5932 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5933 || (call->flags & RX_CALL_FLUSH)
5934 || (call->flags & RX_CALL_FAST_RECOVER)) {
5935 /* Check for the case where the current list contains
5936 * an acked packet. Since we always send retransmissions
5937 * in a separate packet, we only need to check the first
5938 * packet in the list */
5939 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5943 rxi_SendList(call, &last, istack, morePackets);
5944 /* If the call enters an error state stop sending, or if
5945 * we entered congestion recovery mode, stop sending */
5947 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5951 rxi_SendList(call, &working, istack, 0);
5953 } else if (last.len > 0) {
5954 rxi_SendList(call, &last, istack, 0);
5955 /* Packets which are in 'working' are not sent by this call */
5960 * Check if the peer for the given call is known to be dead
5962 * If the call's peer appears dead (it has encountered fatal network errors
5963 * since the call started) the call is killed with RX_CALL_DEAD if the call
5964 * is active. Otherwise, we do nothing.
5966 * @param[in] call The call to check
5969 * @retval 0 The call is fine, and we haven't done anything to the call
5970 * @retval nonzero The call's peer appears dead, and the call has been
5971 * terminated if it was active
5973 * @pre call->lock must be locked
5976 rxi_CheckPeerDead(struct rx_call *call)
5978 #ifdef AFS_RXERRQ_ENV
5981 if (call->state == RX_STATE_DALLY) {
5985 peererrs = rx_atomic_read(&call->conn->peer->neterrs);
5986 if (call->neterr_gen < peererrs) {
5987 /* we have received network errors since this call started; kill
5989 if (call->state == RX_STATE_ACTIVE) {
5990 rxi_CallError(call, RX_CALL_DEAD);
5994 if (call->neterr_gen > peererrs) {
5995 /* someone has reset the number of peer errors; set the call error gen
5996 * so we can detect if more errors are encountered */
5997 call->neterr_gen = peererrs;
6004 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
6006 struct rx_call *call = arg0;
6007 struct rx_peer *peer;
6008 struct opr_queue *cursor;
6009 struct clock maxTimeout = { 60, 0 };
6011 MUTEX_ENTER(&call->lock);
6013 peer = call->conn->peer;
6015 /* Make sure that the event pointer is removed from the call
6016 * structure, since there is no longer a per-call retransmission
6018 if (event == call->resendEvent) {
6019 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
6020 rxevent_Put(call->resendEvent);
6021 call->resendEvent = NULL;
6024 rxi_CheckPeerDead(call);
6026 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
6027 rxi_CheckBusy(call);
6030 if (opr_queue_IsEmpty(&call->tq)) {
6031 /* Nothing to do. This means that we've been raced, and that an
6032 * ACK has come in between when we were triggered, and when we
6033 * actually got to run. */
6037 /* We're in loss recovery */
6038 call->flags |= RX_CALL_FAST_RECOVER;
6040 /* Mark all of the pending packets in the queue as being lost */
6041 for (opr_queue_Scan(&call->tq, cursor)) {
6042 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
6043 if (!(p->flags & RX_PKTFLAG_ACKED))
6044 p->flags &= ~RX_PKTFLAG_SENT;
6047 /* We're resending, so we double the timeout of the call. This will be
6048 * dropped back down by the first successful ACK that we receive.
6050 * We apply a maximum value here of 60 seconds
6052 clock_Add(&call->rto, &call->rto);
6053 if (clock_Gt(&call->rto, &maxTimeout))
6054 call->rto = maxTimeout;
6056 /* Packet loss is most likely due to congestion, so drop our window size
6057 * and start again from the beginning */
6058 if (peer->maxDgramPackets >1) {
6059 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
6060 call->MTU = MIN(peer->natMTU, peer->maxMTU);
6062 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
6063 call->nDgramPackets = 1;
6065 call->nextCwind = 1;
6068 MUTEX_ENTER(&peer->peer_lock);
6069 peer->MTU = call->MTU;
6070 peer->cwind = call->cwind;
6071 peer->nDgramPackets = 1;
6073 call->congestSeq = peer->congestSeq;
6074 MUTEX_EXIT(&peer->peer_lock);
6076 rxi_Start(call, istack);
6079 MUTEX_EXIT(&call->lock);
6082 /* This routine is called when new packets are readied for
6083 * transmission and when retransmission may be necessary, or when the
6084 * transmission window or burst count are favourable. This should be
6085 * better optimized for new packets, the usual case, now that we've
6086 * got rid of queues of send packets. XXXXXXXXXXX */
6088 rxi_Start(struct rx_call *call, int istack)
6090 struct opr_queue *cursor;
6091 #ifdef RX_ENABLE_LOCKS
6092 struct opr_queue *store;
6098 #ifdef RX_ENABLE_LOCKS
6099 if (rx_stats_active)
6100 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
6105 if (!opr_queue_IsEmpty(&call->tq)) { /* If we have anything to send */
6106 /* Send (or resend) any packets that need it, subject to
6107 * window restrictions and congestion burst control
6108 * restrictions. Ask for an ack on the last packet sent in
6109 * this burst. For now, we're relying upon the window being
6110 * considerably bigger than the largest number of packets that
6111 * are typically sent at once by one initial call to
6112 * rxi_Start. This is probably bogus (perhaps we should ask
6113 * for an ack when we're half way through the current
6114 * window?). Also, for non file transfer applications, this
6115 * may end up asking for an ack for every packet. Bogus. XXXX
6118 * But check whether we're here recursively, and let the other guy
6121 #ifdef RX_ENABLE_LOCKS
6122 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6123 call->flags |= RX_CALL_TQ_BUSY;
6125 #endif /* RX_ENABLE_LOCKS */
6127 #ifdef RX_ENABLE_LOCKS
6128 call->flags &= ~RX_CALL_NEED_START;
6129 #endif /* RX_ENABLE_LOCKS */
6131 maxXmitPackets = MIN(call->twind, call->cwind);
6132 for (opr_queue_Scan(&call->tq, cursor)) {
6134 = opr_queue_Entry(cursor, struct rx_packet, entry);
6136 if (p->flags & RX_PKTFLAG_ACKED) {
6137 /* Since we may block, don't trust this */
6138 if (rx_stats_active)
6139 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6140 continue; /* Ignore this packet if it has been acknowledged */
6143 /* Turn off all flags except these ones, which are the same
6144 * on each transmission */
6145 p->header.flags &= RX_PRESET_FLAGS;
6147 if (p->header.seq >=
6148 call->tfirst + MIN((int)call->twind,
6149 (int)(call->nSoftAcked +
6151 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6152 /* Note: if we're waiting for more window space, we can
6153 * still send retransmits; hence we don't return here, but
6154 * break out to schedule a retransmit event */
6155 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6156 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6161 /* Transmit the packet if it needs to be sent. */
6162 if (!(p->flags & RX_PKTFLAG_SENT)) {
6163 if (nXmitPackets == maxXmitPackets) {
6164 rxi_SendXmitList(call, call->xmitList,
6165 nXmitPackets, istack);
6168 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6169 *(call->callNumber), p));
6170 call->xmitList[nXmitPackets++] = p;
6172 } /* end of the queue_Scan */
6174 /* xmitList now hold pointers to all of the packets that are
6175 * ready to send. Now we loop to send the packets */
6176 if (nXmitPackets > 0) {
6177 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6181 #ifdef RX_ENABLE_LOCKS
6183 /* We went into the error state while sending packets. Now is
6184 * the time to reset the call. This will also inform the using
6185 * process that the call is in an error state.
6187 if (rx_stats_active)
6188 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6189 call->flags &= ~RX_CALL_TQ_BUSY;
6190 rxi_WakeUpTransmitQueue(call);
6191 rxi_CallError(call, call->error);
6195 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6197 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6198 /* Some packets have received acks. If they all have, we can clear
6199 * the transmit queue.
6202 for (opr_queue_ScanSafe(&call->tq, cursor, store)) {
6204 = opr_queue_Entry(cursor, struct rx_packet, entry);
6206 if (p->header.seq < call->tfirst
6207 && (p->flags & RX_PKTFLAG_ACKED)) {
6208 opr_queue_Remove(&p->entry);
6209 #ifdef RX_TRACK_PACKETS
6210 p->flags &= ~RX_PKTFLAG_TQ;
6212 #ifdef RXDEBUG_PACKET
6220 call->flags |= RX_CALL_TQ_CLEARME;
6222 if (call->flags & RX_CALL_TQ_CLEARME)
6223 rxi_ClearTransmitQueue(call, 1);
6224 } while (call->flags & RX_CALL_NEED_START);
6226 * TQ references no longer protected by this flag; they must remain
6227 * protected by the call lock.
6229 call->flags &= ~RX_CALL_TQ_BUSY;
6230 rxi_WakeUpTransmitQueue(call);
6232 call->flags |= RX_CALL_NEED_START;
6234 #endif /* RX_ENABLE_LOCKS */
6236 rxi_rto_cancel(call);
6240 /* Also adjusts the keep alive parameters for the call, to reflect
6241 * that we have just sent a packet (so keep alives aren't sent
6244 rxi_Send(struct rx_call *call, struct rx_packet *p,
6247 struct rx_connection *conn = call->conn;
6249 /* Stamp each packet with the user supplied status */
6250 p->header.userStatus = call->localStatus;
6252 /* Allow the security object controlling this call's security to
6253 * make any last-minute changes to the packet */
6254 RXS_SendPacket(conn->securityObject, call, p);
6256 /* Since we're about to send SOME sort of packet to the peer, it's
6257 * safe to nuke any scheduled end-of-packets ack */
6258 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
6260 /* Actually send the packet, filling in more connection-specific fields */
6261 MUTEX_EXIT(&call->lock);
6262 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6263 rxi_SendPacket(call, conn, p, istack);
6264 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6265 MUTEX_ENTER(&call->lock);
6267 /* Update last send time for this call (for keep-alive
6268 * processing), and for the connection (so that we can discover
6269 * idle connections) */
6270 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6271 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6272 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6274 conn->lastSendTime = call->lastSendTime = clock_Sec();
6275 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6276 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6277 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6278 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6279 RX_ACK_PING_RESPONSE)))
6280 call->lastSendData = call->lastSendTime;
6284 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6285 * that things are fine. Also called periodically to guarantee that nothing
6286 * falls through the cracks (e.g. (error + dally) connections have keepalive
6287 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6289 * haveCTLock Set if calling from rxi_ReapConnections
6292 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6294 struct rx_connection *conn = call->conn;
6296 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6297 afs_uint32 fudgeFactor;
6300 int idle_timeout = 0;
6301 afs_int32 clock_diff = 0;
6303 if (rxi_CheckPeerDead(call)) {
6309 /* Large swings in the clock can have a significant impact on
6310 * the performance of RX call processing. Forward clock shifts
6311 * will result in premature event triggering or timeouts.
6312 * Backward shifts can result in calls not completing until
6313 * the clock catches up with the original start clock value.
6315 * If a backward clock shift of more than five minutes is noticed,
6316 * just fail the call.
6318 if (now < call->lastSendTime)
6319 clock_diff = call->lastSendTime - now;
6320 if (now < call->startWait)
6321 clock_diff = MAX(clock_diff, call->startWait - now);
6322 if (now < call->lastReceiveTime)
6323 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6324 if (clock_diff > 5 * 60)
6326 if (call->state == RX_STATE_ACTIVE)
6327 rxi_CallError(call, RX_CALL_TIMEOUT);
6331 #ifdef RX_ENABLE_LOCKS
6332 if (call->flags & RX_CALL_TQ_BUSY) {
6333 /* Call is active and will be reset by rxi_Start if it's
6334 * in an error state.
6339 /* RTT + 8*MDEV, rounded up to the next second. */
6340 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6341 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6343 deadTime = conn->secondsUntilDead + fudgeFactor;
6344 /* These are computed to the second (+- 1 second). But that's
6345 * good enough for these values, which should be a significant
6346 * number of seconds. */
6347 if (now > (call->lastReceiveTime + deadTime)) {
6348 if (call->state == RX_STATE_ACTIVE) {
6349 #ifdef AFS_ADAPT_PMTU
6350 # if defined(KERNEL) && defined(AFS_SUN5_ENV)
6352 # if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6353 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6354 ip_stack_t *ipst = ns->netstack_ip;
6356 ire = ire_cache_lookup(conn->peer->host
6357 # if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6359 # if defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID)
6361 # if defined(GLOBAL_NETSTACKID)
6368 if (ire && ire->ire_max_frag > 0)
6369 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6371 # if defined(GLOBAL_NETSTACKID)
6375 #endif /* AFS_ADAPT_PMTU */
6376 cerror = RX_CALL_DEAD;
6379 #ifdef RX_ENABLE_LOCKS
6380 /* Cancel pending events */
6381 rxevent_Cancel(&call->delayedAckEvent, call,
6382 RX_CALL_REFCOUNT_DELAY);
6383 rxi_rto_cancel(call);
6384 rxevent_Cancel(&call->keepAliveEvent, call,
6385 RX_CALL_REFCOUNT_ALIVE);
6386 rxevent_Cancel(&call->growMTUEvent, call,
6387 RX_CALL_REFCOUNT_MTU);
6388 MUTEX_ENTER(&rx_refcnt_mutex);
6389 /* if rxi_FreeCall returns 1 it has freed the call */
6390 if (call->refCount == 0 &&
6391 rxi_FreeCall(call, haveCTLock))
6393 MUTEX_EXIT(&rx_refcnt_mutex);
6396 MUTEX_EXIT(&rx_refcnt_mutex);
6398 #else /* RX_ENABLE_LOCKS */
6399 rxi_FreeCall(call, 0);
6401 #endif /* RX_ENABLE_LOCKS */
6403 /* Non-active calls are destroyed if they are not responding
6404 * to pings; active calls are simply flagged in error, so the
6405 * attached process can die reasonably gracefully. */
6408 if (conn->idleDeadDetection) {
6409 if (conn->idleDeadTime) {
6410 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6414 /* see if we have a non-activity timeout */
6415 if (call->startWait && ((call->startWait + idleDeadTime) < now) &&
6416 (call->flags & RX_CALL_READER_WAIT)) {
6417 if (call->state == RX_STATE_ACTIVE) {
6418 cerror = RX_CALL_TIMEOUT;
6423 if (call->lastSendData && ((call->lastSendData + idleDeadTime) < now)) {
6424 if (call->state == RX_STATE_ACTIVE) {
6425 cerror = conn->service ? conn->service->idleDeadErr : RX_CALL_IDLE;
6433 if (conn->hardDeadTime) {
6434 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6437 /* see if we have a hard timeout */
6439 && (now > (hardDeadTime + call->startTime.sec))) {
6440 if (call->state == RX_STATE_ACTIVE)
6441 rxi_CallError(call, RX_CALL_TIMEOUT);
6446 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6447 call->lastReceiveTime) {
6448 int oldMTU = conn->peer->ifMTU;
6450 /* if we thought we could send more, perhaps things got worse */
6451 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6452 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6453 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6454 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6456 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6458 /* minimum capped in SetPeerMtu */
6459 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6462 conn->lastPacketSize = 0;
6464 /* needed so ResetCall doesn't clobber us. */
6465 call->MTU = conn->peer->ifMTU;
6467 /* if we never succeeded, let the error pass out as-is */
6468 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6469 cerror = conn->msgsizeRetryErr;
6472 rxi_CallError(call, cerror);
6477 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6478 void *dummy, int dummy2)
6480 struct rx_connection *conn = arg1;
6481 struct rx_header theader;
6482 char tbuffer[1 + sizeof(struct rx_header)];
6483 struct sockaddr_in taddr;
6486 struct iovec tmpiov[2];
6489 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6492 tp = &tbuffer[sizeof(struct rx_header)];
6493 taddr.sin_family = AF_INET;
6494 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6495 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6496 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6497 taddr.sin_len = sizeof(struct sockaddr_in);
6499 memset(&theader, 0, sizeof(theader));
6500 theader.epoch = htonl(999);
6502 theader.callNumber = 0;
6505 theader.type = RX_PACKET_TYPE_VERSION;
6506 theader.flags = RX_LAST_PACKET;
6507 theader.serviceId = 0;
6509 memcpy(tbuffer, &theader, sizeof(theader));
6510 memcpy(tp, &a, sizeof(a));
6511 tmpiov[0].iov_base = tbuffer;
6512 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6514 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6516 MUTEX_ENTER(&conn->conn_data_lock);
6517 MUTEX_ENTER(&rx_refcnt_mutex);
6518 /* Only reschedule ourselves if the connection would not be destroyed */
6519 if (conn->refCount <= 1) {
6520 rxevent_Put(conn->natKeepAliveEvent);
6521 conn->natKeepAliveEvent = NULL;
6522 MUTEX_EXIT(&rx_refcnt_mutex);
6523 MUTEX_EXIT(&conn->conn_data_lock);
6524 rx_DestroyConnection(conn); /* drop the reference for this */
6526 conn->refCount--; /* drop the reference for this */
6527 MUTEX_EXIT(&rx_refcnt_mutex);
6528 rxevent_Put(conn->natKeepAliveEvent);
6529 conn->natKeepAliveEvent = NULL;
6530 rxi_ScheduleNatKeepAliveEvent(conn);
6531 MUTEX_EXIT(&conn->conn_data_lock);
6536 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6538 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6539 struct clock when, now;
6540 clock_GetTime(&now);
6542 when.sec += conn->secondsUntilNatPing;
6543 MUTEX_ENTER(&rx_refcnt_mutex);
6544 conn->refCount++; /* hold a reference for this */
6545 MUTEX_EXIT(&rx_refcnt_mutex);
6546 conn->natKeepAliveEvent =
6547 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6552 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6554 MUTEX_ENTER(&conn->conn_data_lock);
6555 conn->secondsUntilNatPing = seconds;
6557 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6558 rxi_ScheduleNatKeepAliveEvent(conn);
6560 conn->flags |= RX_CONN_NAT_PING;
6562 MUTEX_EXIT(&conn->conn_data_lock);
6565 /* When a call is in progress, this routine is called occasionally to
6566 * make sure that some traffic has arrived (or been sent to) the peer.
6567 * If nothing has arrived in a reasonable amount of time, the call is
6568 * declared dead; if nothing has been sent for a while, we send a
6569 * keep-alive packet (if we're actually trying to keep the call alive)
6572 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6575 struct rx_call *call = arg1;
6576 struct rx_connection *conn;
6579 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6580 MUTEX_ENTER(&call->lock);
6582 if (event == call->keepAliveEvent) {
6583 rxevent_Put(call->keepAliveEvent);
6584 call->keepAliveEvent = NULL;
6589 if (rxi_CheckCall(call, 0)) {
6590 MUTEX_EXIT(&call->lock);
6594 /* Don't try to keep alive dallying calls */
6595 if (call->state == RX_STATE_DALLY) {
6596 MUTEX_EXIT(&call->lock);
6601 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6602 /* Don't try to send keepalives if there is unacknowledged data */
6603 /* the rexmit code should be good enough, this little hack
6604 * doesn't quite work XXX */
6605 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6607 rxi_ScheduleKeepAliveEvent(call);
6608 MUTEX_EXIT(&call->lock);
6611 /* Does what's on the nameplate. */
6613 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6615 struct rx_call *call = arg1;
6616 struct rx_connection *conn;
6618 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6619 MUTEX_ENTER(&call->lock);
6621 if (event == call->growMTUEvent) {
6622 rxevent_Put(call->growMTUEvent);
6623 call->growMTUEvent = NULL;
6626 if (rxi_CheckCall(call, 0)) {
6627 MUTEX_EXIT(&call->lock);
6631 /* Don't bother with dallying calls */
6632 if (call->state == RX_STATE_DALLY) {
6633 MUTEX_EXIT(&call->lock);
6640 * keep being scheduled, just don't do anything if we're at peak,
6641 * or we're not set up to be properly handled (idle timeout required)
6643 if ((conn->peer->maxPacketSize != 0) &&
6644 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6645 conn->idleDeadDetection)
6646 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6647 rxi_ScheduleGrowMTUEvent(call, 0);
6648 MUTEX_EXIT(&call->lock);
6652 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6654 if (!call->keepAliveEvent) {
6655 struct clock when, now;
6656 clock_GetTime(&now);
6658 when.sec += call->conn->secondsUntilPing;
6659 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6660 call->keepAliveEvent =
6661 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6666 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6668 if (!call->growMTUEvent) {
6669 struct clock when, now;
6671 clock_GetTime(&now);
6674 if (call->conn->secondsUntilPing)
6675 secs = (6*call->conn->secondsUntilPing)-1;
6677 if (call->conn->secondsUntilDead)
6678 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6682 CALL_HOLD(call, RX_CALL_REFCOUNT_MTU);
6683 call->growMTUEvent =
6684 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6688 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6690 rxi_KeepAliveOn(struct rx_call *call)
6692 /* Pretend last packet received was received now--i.e. if another
6693 * packet isn't received within the keep alive time, then the call
6694 * will die; Initialize last send time to the current time--even
6695 * if a packet hasn't been sent yet. This will guarantee that a
6696 * keep-alive is sent within the ping time */
6697 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6698 rxi_ScheduleKeepAliveEvent(call);
6702 * Solely in order that callers not need to include rx_call.h
6705 rx_KeepAliveOff(struct rx_call *call)
6707 rxi_KeepAliveOff(call);
6710 rx_KeepAliveOn(struct rx_call *call)
6712 rxi_KeepAliveOn(call);
6716 rxi_GrowMTUOn(struct rx_call *call)
6718 struct rx_connection *conn = call->conn;
6719 MUTEX_ENTER(&conn->conn_data_lock);
6720 conn->lastPingSizeSer = conn->lastPingSize = 0;
6721 MUTEX_EXIT(&conn->conn_data_lock);
6722 rxi_ScheduleGrowMTUEvent(call, 1);
6725 /* This routine is called to send connection abort messages
6726 * that have been delayed to throttle looping clients. */
6728 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6731 struct rx_connection *conn = arg1;
6734 struct rx_packet *packet;
6736 MUTEX_ENTER(&conn->conn_data_lock);
6737 rxevent_Put(conn->delayedAbortEvent);
6738 conn->delayedAbortEvent = NULL;
6739 error = htonl(conn->error);
6741 MUTEX_EXIT(&conn->conn_data_lock);
6742 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6745 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6746 RX_PACKET_TYPE_ABORT, (char *)&error,
6748 rxi_FreePacket(packet);
6752 /* This routine is called to send call abort messages
6753 * that have been delayed to throttle looping clients. */
6755 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6758 struct rx_call *call = arg1;
6761 struct rx_packet *packet;
6763 MUTEX_ENTER(&call->lock);
6764 rxevent_Put(call->delayedAbortEvent);
6765 call->delayedAbortEvent = NULL;
6766 error = htonl(call->error);
6768 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6771 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6772 (char *)&error, sizeof(error), 0);
6773 rxi_FreePacket(packet);
6775 MUTEX_EXIT(&call->lock);
6776 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6779 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6780 * seconds) to ask the client to authenticate itself. The routine
6781 * issues a challenge to the client, which is obtained from the
6782 * security object associated with the connection */
6784 rxi_ChallengeEvent(struct rxevent *event,
6785 void *arg0, void *arg1, int tries)
6787 struct rx_connection *conn = arg0;
6790 rxevent_Put(conn->challengeEvent);
6791 conn->challengeEvent = NULL;
6794 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6795 struct rx_packet *packet;
6796 struct clock when, now;
6799 /* We've failed to authenticate for too long.
6800 * Reset any calls waiting for authentication;
6801 * they are all in RX_STATE_PRECALL.
6805 MUTEX_ENTER(&conn->conn_call_lock);
6806 for (i = 0; i < RX_MAXCALLS; i++) {
6807 struct rx_call *call = conn->call[i];
6809 MUTEX_ENTER(&call->lock);
6810 if (call->state == RX_STATE_PRECALL) {
6811 rxi_CallError(call, RX_CALL_DEAD);
6812 rxi_SendCallAbort(call, NULL, 0, 0);
6814 MUTEX_EXIT(&call->lock);
6817 MUTEX_EXIT(&conn->conn_call_lock);
6821 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6823 /* If there's no packet available, do this later. */
6824 RXS_GetChallenge(conn->securityObject, conn, packet);
6825 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6826 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6827 rxi_FreePacket(packet);
6829 clock_GetTime(&now);
6831 when.sec += RX_CHALLENGE_TIMEOUT;
6832 conn->challengeEvent =
6833 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6838 /* Call this routine to start requesting the client to authenticate
6839 * itself. This will continue until authentication is established,
6840 * the call times out, or an invalid response is returned. The
6841 * security object associated with the connection is asked to create
6842 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6843 * defined earlier. */
6845 rxi_ChallengeOn(struct rx_connection *conn)
6847 if (!conn->challengeEvent) {
6848 RXS_CreateChallenge(conn->securityObject, conn);
6849 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6854 /* rxi_ComputeRoundTripTime is called with peer locked. */
6855 /* peer may be null */
6857 rxi_ComputeRoundTripTime(struct rx_packet *p,
6858 struct rx_ackPacket *ack,
6859 struct rx_call *call,
6860 struct rx_peer *peer,
6863 struct clock thisRtt, *sentp;
6867 /* If the ACK is delayed, then do nothing */
6868 if (ack->reason == RX_ACK_DELAY)
6871 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6872 * their RTT multiple times, so only include the RTT of the last packet
6874 if (p->flags & RX_JUMBO_PACKET)
6877 /* Use the serial number to determine which transmission the ACK is for,
6878 * and set the sent time to match this. If we have no serial number, then
6879 * only use the ACK for RTT calculations if the packet has not been
6883 serial = ntohl(ack->serial);
6885 if (serial == p->header.serial) {
6886 sentp = &p->timeSent;
6887 } else if (serial == p->firstSerial) {
6888 sentp = &p->firstSent;
6889 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6890 sentp = &p->firstSent;
6894 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6895 sentp = &p->firstSent;
6902 if (clock_Lt(&thisRtt, sentp))
6903 return; /* somebody set the clock back, don't count this time. */
6905 clock_Sub(&thisRtt, sentp);
6906 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6907 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6909 if (clock_IsZero(&thisRtt)) {
6911 * The actual round trip time is shorter than the
6912 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6913 * Since we can't tell which at the moment we will assume 1ms.
6915 thisRtt.usec = 1000;
6918 if (rx_stats_active) {
6919 MUTEX_ENTER(&rx_stats_mutex);
6920 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6921 rx_stats.minRtt = thisRtt;
6922 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6923 if (thisRtt.sec > 60) {
6924 MUTEX_EXIT(&rx_stats_mutex);
6925 return; /* somebody set the clock ahead */
6927 rx_stats.maxRtt = thisRtt;
6929 clock_Add(&rx_stats.totalRtt, &thisRtt);
6930 rx_atomic_inc(&rx_stats.nRttSamples);
6931 MUTEX_EXIT(&rx_stats_mutex);
6934 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6936 /* Apply VanJacobson round-trip estimations */
6941 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6942 * srtt is stored as fixed point with 3 bits after the binary
6943 * point (i.e., scaled by 8). The following magic is
6944 * equivalent to the smoothing algorithm in rfc793 with an
6945 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6946 * srtt'*8 = rtt + srtt*7
6947 * srtt'*8 = srtt*8 + rtt - srtt
6948 * srtt' = srtt + rtt/8 - srtt/8
6949 * srtt' = srtt + (rtt - srtt)/8
6952 delta = _8THMSEC(&thisRtt) - call->rtt;
6953 call->rtt += (delta >> 3);
6956 * We accumulate a smoothed rtt variance (actually, a smoothed
6957 * mean difference), then set the retransmit timer to smoothed
6958 * rtt + 4 times the smoothed variance (was 2x in van's original
6959 * paper, but 4x works better for me, and apparently for him as
6961 * rttvar is stored as
6962 * fixed point with 2 bits after the binary point (scaled by
6963 * 4). The following is equivalent to rfc793 smoothing with
6964 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6965 * rttvar'*4 = rttvar*3 + |delta|
6966 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6967 * rttvar' = rttvar + |delta|/4 - rttvar/4
6968 * rttvar' = rttvar + (|delta| - rttvar)/4
6969 * This replaces rfc793's wired-in beta.
6970 * dev*4 = dev*4 + (|actual - expected| - dev)
6976 delta -= (call->rtt_dev << 1);
6977 call->rtt_dev += (delta >> 3);
6979 /* I don't have a stored RTT so I start with this value. Since I'm
6980 * probably just starting a call, and will be pushing more data down
6981 * this, I expect congestion to increase rapidly. So I fudge a
6982 * little, and I set deviance to half the rtt. In practice,
6983 * deviance tends to approach something a little less than
6984 * half the smoothed rtt. */
6985 call->rtt = _8THMSEC(&thisRtt) + 8;
6986 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6988 /* the smoothed RTT time is RTT + 4*MDEV
6990 * We allow a user specified minimum to be set for this, to allow clamping
6991 * at a minimum value in the same way as TCP. In addition, we have to allow
6992 * for the possibility that this packet is answered by a delayed ACK, so we
6993 * add on a fixed 200ms to account for that timer expiring.
6996 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6997 rx_minPeerTimeout) + 200;
6998 clock_Zero(&call->rto);
6999 clock_Addmsec(&call->rto, rtt_timeout);
7001 /* Update the peer, so any new calls start with our values */
7002 peer->rtt_dev = call->rtt_dev;
7003 peer->rtt = call->rtt;
7005 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
7006 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
7010 /* Find all server connections that have not been active for a long time, and
7013 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
7016 struct clock now, when;
7017 clock_GetTime(&now);
7019 /* Find server connection structures that haven't been used for
7020 * greater than rx_idleConnectionTime */
7022 struct rx_connection **conn_ptr, **conn_end;
7023 int i, havecalls = 0;
7024 MUTEX_ENTER(&rx_connHashTable_lock);
7025 for (conn_ptr = &rx_connHashTable[0], conn_end =
7026 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
7028 struct rx_connection *conn, *next;
7029 struct rx_call *call;
7033 for (conn = *conn_ptr; conn; conn = next) {
7034 /* XXX -- Shouldn't the connection be locked? */
7037 for (i = 0; i < RX_MAXCALLS; i++) {
7038 call = conn->call[i];
7042 code = MUTEX_TRYENTER(&call->lock);
7045 result = rxi_CheckCall(call, 1);
7046 MUTEX_EXIT(&call->lock);
7048 /* If CheckCall freed the call, it might
7049 * have destroyed the connection as well,
7050 * which screws up the linked lists.
7056 if (conn->type == RX_SERVER_CONNECTION) {
7057 /* This only actually destroys the connection if
7058 * there are no outstanding calls */
7059 MUTEX_ENTER(&conn->conn_data_lock);
7060 MUTEX_ENTER(&rx_refcnt_mutex);
7061 if (!havecalls && !conn->refCount
7062 && ((conn->lastSendTime + rx_idleConnectionTime) <
7064 conn->refCount++; /* it will be decr in rx_DestroyConn */
7065 MUTEX_EXIT(&rx_refcnt_mutex);
7066 MUTEX_EXIT(&conn->conn_data_lock);
7067 #ifdef RX_ENABLE_LOCKS
7068 rxi_DestroyConnectionNoLock(conn);
7069 #else /* RX_ENABLE_LOCKS */
7070 rxi_DestroyConnection(conn);
7071 #endif /* RX_ENABLE_LOCKS */
7073 #ifdef RX_ENABLE_LOCKS
7075 MUTEX_EXIT(&rx_refcnt_mutex);
7076 MUTEX_EXIT(&conn->conn_data_lock);
7078 #endif /* RX_ENABLE_LOCKS */
7082 #ifdef RX_ENABLE_LOCKS
7083 while (rx_connCleanup_list) {
7084 struct rx_connection *conn;
7085 conn = rx_connCleanup_list;
7086 rx_connCleanup_list = rx_connCleanup_list->next;
7087 MUTEX_EXIT(&rx_connHashTable_lock);
7088 rxi_CleanupConnection(conn);
7089 MUTEX_ENTER(&rx_connHashTable_lock);
7091 MUTEX_EXIT(&rx_connHashTable_lock);
7092 #endif /* RX_ENABLE_LOCKS */
7095 /* Find any peer structures that haven't been used (haven't had an
7096 * associated connection) for greater than rx_idlePeerTime */
7098 struct rx_peer **peer_ptr, **peer_end;
7102 * Why do we need to hold the rx_peerHashTable_lock across
7103 * the incrementing of peer_ptr since the rx_peerHashTable
7104 * array is not changing? We don't.
7106 * By dropping the lock periodically we can permit other
7107 * activities to be performed while a rxi_ReapConnections
7108 * call is in progress. The goal of reap connections
7109 * is to clean up quickly without causing large amounts
7110 * of contention. Therefore, it is important that global
7111 * mutexes not be held for extended periods of time.
7113 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7114 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7116 struct rx_peer *peer, *next, *prev;
7118 MUTEX_ENTER(&rx_peerHashTable_lock);
7119 for (prev = peer = *peer_ptr; peer; peer = next) {
7121 code = MUTEX_TRYENTER(&peer->peer_lock);
7122 if ((code) && (peer->refCount == 0)
7123 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7124 struct opr_queue *cursor, *store;
7128 * now know that this peer object is one to be
7129 * removed from the hash table. Once it is removed
7130 * it can't be referenced by other threads.
7131 * Lets remove it first and decrement the struct
7132 * nPeerStructs count.
7134 if (peer == *peer_ptr) {
7140 if (rx_stats_active)
7141 rx_atomic_dec(&rx_stats.nPeerStructs);
7144 * Now if we hold references on 'prev' and 'next'
7145 * we can safely drop the rx_peerHashTable_lock
7146 * while we destroy this 'peer' object.
7152 MUTEX_EXIT(&rx_peerHashTable_lock);
7154 MUTEX_EXIT(&peer->peer_lock);
7155 MUTEX_DESTROY(&peer->peer_lock);
7157 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7158 unsigned int num_funcs;
7159 struct rx_interface_stat *rpc_stat
7160 = opr_queue_Entry(cursor, struct rx_interface_stat,
7165 opr_queue_Remove(&rpc_stat->entry);
7166 opr_queue_Remove(&rpc_stat->entryPeers);
7168 num_funcs = rpc_stat->stats[0].func_total;
7170 sizeof(rx_interface_stat_t) +
7171 rpc_stat->stats[0].func_total *
7172 sizeof(rx_function_entry_v1_t);
7174 rxi_Free(rpc_stat, space);
7176 MUTEX_ENTER(&rx_rpc_stats);
7177 rxi_rpc_peer_stat_cnt -= num_funcs;
7178 MUTEX_EXIT(&rx_rpc_stats);
7183 * Regain the rx_peerHashTable_lock and
7184 * decrement the reference count on 'prev'
7187 MUTEX_ENTER(&rx_peerHashTable_lock);
7194 MUTEX_EXIT(&peer->peer_lock);
7199 MUTEX_EXIT(&rx_peerHashTable_lock);
7203 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7204 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7205 * GC, just below. Really, we shouldn't have to keep moving packets from
7206 * one place to another, but instead ought to always know if we can
7207 * afford to hold onto a packet in its particular use. */
7208 MUTEX_ENTER(&rx_freePktQ_lock);
7209 if (rx_waitingForPackets) {
7210 rx_waitingForPackets = 0;
7211 #ifdef RX_ENABLE_LOCKS
7212 CV_BROADCAST(&rx_waitingForPackets_cv);
7214 osi_rxWakeup(&rx_waitingForPackets);
7217 MUTEX_EXIT(&rx_freePktQ_lock);
7220 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7221 rxevent_Put(rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0));
7225 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7226 * rx.h is sort of strange this is better. This is called with a security
7227 * object before it is discarded. Each connection using a security object has
7228 * its own refcount to the object so it won't actually be freed until the last
7229 * connection is destroyed.
7231 * This is the only rxs module call. A hold could also be written but no one
7235 rxs_Release(struct rx_securityClass *aobj)
7237 return RXS_Close(aobj);
7245 #define TRACE_OPTION_RX_DEBUG 16
7253 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7254 0, KEY_QUERY_VALUE, &parmKey);
7255 if (code != ERROR_SUCCESS)
7258 dummyLen = sizeof(TraceOption);
7259 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7260 (BYTE *) &TraceOption, &dummyLen);
7261 if (code == ERROR_SUCCESS) {
7262 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7264 RegCloseKey (parmKey);
7265 #endif /* AFS_NT40_ENV */
7270 rx_DebugOnOff(int on)
7274 rxdebug_active = on;
7280 rx_StatsOnOff(int on)
7282 rx_stats_active = on;
7286 /* Don't call this debugging routine directly; use dpf */
7288 rxi_DebugPrint(char *format, ...)
7297 va_start(ap, format);
7299 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7302 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7304 OutputDebugString(msg);
7310 va_start(ap, format);
7312 clock_GetTime(&now);
7313 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7314 (unsigned int)now.usec);
7315 vfprintf(rx_Log, format, ap);
7323 * This function is used to process the rx_stats structure that is local
7324 * to a process as well as an rx_stats structure received from a remote
7325 * process (via rxdebug). Therefore, it needs to do minimal version
7329 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7330 afs_int32 freePackets, char version)
7334 if (size != sizeof(struct rx_statistics)) {
7336 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7337 size, sizeof(struct rx_statistics));
7340 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7343 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7344 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7345 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7346 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7347 s->specialPktAllocFailures);
7349 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7350 s->receivePktAllocFailures, s->sendPktAllocFailures,
7351 s->specialPktAllocFailures);
7355 " greedy %u, " "bogusReads %u (last from host %x), "
7356 "noPackets %u, " "noBuffers %u, " "selects %u, "
7357 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7358 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7359 s->selects, s->sendSelects);
7361 fprintf(file, " packets read: ");
7362 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7363 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7365 fprintf(file, "\n");
7368 " other read counters: data %u, " "ack %u, " "dup %u "
7369 "spurious %u " "dally %u\n", s->dataPacketsRead,
7370 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7371 s->ignorePacketDally);
7373 fprintf(file, " packets sent: ");
7374 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7375 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7377 fprintf(file, "\n");
7380 " other send counters: ack %u, " "data %u (not resends), "
7381 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7382 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7383 s->dataPacketsPushed, s->ignoreAckedPacket);
7386 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7387 s->netSendFailures, (int)s->fatalErrors);
7389 if (s->nRttSamples) {
7390 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7391 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7393 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7394 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7398 " %d server connections, " "%d client connections, "
7399 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7400 s->nServerConns, s->nClientConns, s->nPeerStructs,
7401 s->nCallStructs, s->nFreeCallStructs);
7403 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7404 fprintf(file, " %d clock updates\n", clock_nUpdates);
7408 /* for backward compatibility */
7410 rx_PrintStats(FILE * file)
7412 MUTEX_ENTER(&rx_stats_mutex);
7413 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7414 sizeof(rx_stats), rx_nFreePackets,
7416 MUTEX_EXIT(&rx_stats_mutex);
7420 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7422 fprintf(file, "Peer %x.%d.\n",
7423 ntohl(peer->host), (int)ntohs(peer->port));
7426 " Rtt %d, " "total sent %d, " "resent %d\n",
7427 peer->rtt, peer->nSent, peer->reSends);
7429 fprintf(file, " Packet size %d\n", peer->ifMTU);
7433 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7435 * This mutex protects the following static variables:
7439 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7440 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7442 #define LOCK_RX_DEBUG
7443 #define UNLOCK_RX_DEBUG
7444 #endif /* AFS_PTHREAD_ENV */
7446 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7448 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7449 u_char type, void *inputData, size_t inputLength,
7450 void *outputData, size_t outputLength)
7452 static afs_int32 counter = 100;
7453 time_t waitTime, waitCount;
7454 struct rx_header theader;
7457 struct timeval tv_now, tv_wake, tv_delta;
7458 struct sockaddr_in taddr, faddr;
7472 tp = &tbuffer[sizeof(struct rx_header)];
7473 taddr.sin_family = AF_INET;
7474 taddr.sin_port = remotePort;
7475 taddr.sin_addr.s_addr = remoteAddr;
7476 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7477 taddr.sin_len = sizeof(struct sockaddr_in);
7480 memset(&theader, 0, sizeof(theader));
7481 theader.epoch = htonl(999);
7483 theader.callNumber = htonl(counter);
7486 theader.type = type;
7487 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7488 theader.serviceId = 0;
7490 memcpy(tbuffer, &theader, sizeof(theader));
7491 memcpy(tp, inputData, inputLength);
7493 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7494 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7496 /* see if there's a packet available */
7497 gettimeofday(&tv_wake, NULL);
7498 tv_wake.tv_sec += waitTime;
7501 FD_SET(socket, &imask);
7502 tv_delta.tv_sec = tv_wake.tv_sec;
7503 tv_delta.tv_usec = tv_wake.tv_usec;
7504 gettimeofday(&tv_now, NULL);
7506 if (tv_delta.tv_usec < tv_now.tv_usec) {
7508 tv_delta.tv_usec += 1000000;
7511 tv_delta.tv_usec -= tv_now.tv_usec;
7513 if (tv_delta.tv_sec < tv_now.tv_sec) {
7517 tv_delta.tv_sec -= tv_now.tv_sec;
7520 code = select(0, &imask, 0, 0, &tv_delta);
7521 #else /* AFS_NT40_ENV */
7522 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7523 #endif /* AFS_NT40_ENV */
7524 if (code == 1 && FD_ISSET(socket, &imask)) {
7525 /* now receive a packet */
7526 faddrLen = sizeof(struct sockaddr_in);
7528 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7529 (struct sockaddr *)&faddr, &faddrLen);
7532 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7533 if (counter == ntohl(theader.callNumber))
7541 /* see if we've timed out */
7549 code -= sizeof(struct rx_header);
7550 if (code > outputLength)
7551 code = outputLength;
7552 memcpy(outputData, tp, code);
7555 #endif /* RXDEBUG */
7558 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7559 afs_uint16 remotePort, struct rx_debugStats * stat,
7560 afs_uint32 * supportedValues)
7562 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7564 struct rx_debugIn in;
7566 *supportedValues = 0;
7567 in.type = htonl(RX_DEBUGI_GETSTATS);
7570 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7571 &in, sizeof(in), stat, sizeof(*stat));
7574 * If the call was successful, fixup the version and indicate
7575 * what contents of the stat structure are valid.
7576 * Also do net to host conversion of fields here.
7580 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7581 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7583 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7584 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7586 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7587 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7589 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7590 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7592 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7593 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7595 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7596 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7598 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7599 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7601 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7602 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7604 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7605 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7607 stat->nFreePackets = ntohl(stat->nFreePackets);
7608 stat->packetReclaims = ntohl(stat->packetReclaims);
7609 stat->callsExecuted = ntohl(stat->callsExecuted);
7610 stat->nWaiting = ntohl(stat->nWaiting);
7611 stat->idleThreads = ntohl(stat->idleThreads);
7612 stat->nWaited = ntohl(stat->nWaited);
7613 stat->nPackets = ntohl(stat->nPackets);
7622 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7623 afs_uint16 remotePort, struct rx_statistics * stat,
7624 afs_uint32 * supportedValues)
7626 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7628 struct rx_debugIn in;
7629 afs_int32 *lp = (afs_int32 *) stat;
7633 * supportedValues is currently unused, but added to allow future
7634 * versioning of this function.
7637 *supportedValues = 0;
7638 in.type = htonl(RX_DEBUGI_RXSTATS);
7640 memset(stat, 0, sizeof(*stat));
7642 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7643 &in, sizeof(in), stat, sizeof(*stat));
7648 * Do net to host conversion here
7651 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7662 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7663 afs_uint16 remotePort, size_t version_length,
7666 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7668 return MakeDebugCall(socket, remoteAddr, remotePort,
7669 RX_PACKET_TYPE_VERSION, a, 1, version,
7677 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7678 afs_uint16 remotePort, afs_int32 * nextConnection,
7679 int allConnections, afs_uint32 debugSupportedValues,
7680 struct rx_debugConn * conn,
7681 afs_uint32 * supportedValues)
7683 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7685 struct rx_debugIn in;
7689 * supportedValues is currently unused, but added to allow future
7690 * versioning of this function.
7693 *supportedValues = 0;
7694 if (allConnections) {
7695 in.type = htonl(RX_DEBUGI_GETALLCONN);
7697 in.type = htonl(RX_DEBUGI_GETCONN);
7699 in.index = htonl(*nextConnection);
7700 memset(conn, 0, sizeof(*conn));
7702 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7703 &in, sizeof(in), conn, sizeof(*conn));
7706 *nextConnection += 1;
7709 * Convert old connection format to new structure.
7712 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7713 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7714 #define MOVEvL(a) (conn->a = vL->a)
7716 /* any old or unrecognized version... */
7717 for (i = 0; i < RX_MAXCALLS; i++) {
7718 MOVEvL(callState[i]);
7719 MOVEvL(callMode[i]);
7720 MOVEvL(callFlags[i]);
7721 MOVEvL(callOther[i]);
7723 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7724 MOVEvL(secStats.type);
7725 MOVEvL(secStats.level);
7726 MOVEvL(secStats.flags);
7727 MOVEvL(secStats.expires);
7728 MOVEvL(secStats.packetsReceived);
7729 MOVEvL(secStats.packetsSent);
7730 MOVEvL(secStats.bytesReceived);
7731 MOVEvL(secStats.bytesSent);
7736 * Do net to host conversion here
7738 * I don't convert host or port since we are most likely
7739 * going to want these in NBO.
7741 conn->cid = ntohl(conn->cid);
7742 conn->serial = ntohl(conn->serial);
7743 for (i = 0; i < RX_MAXCALLS; i++) {
7744 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7746 conn->error = ntohl(conn->error);
7747 conn->secStats.flags = ntohl(conn->secStats.flags);
7748 conn->secStats.expires = ntohl(conn->secStats.expires);
7749 conn->secStats.packetsReceived =
7750 ntohl(conn->secStats.packetsReceived);
7751 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7752 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7753 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7754 conn->epoch = ntohl(conn->epoch);
7755 conn->natMTU = ntohl(conn->natMTU);
7764 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7765 afs_uint16 remotePort, afs_int32 * nextPeer,
7766 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7767 afs_uint32 * supportedValues)
7769 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7771 struct rx_debugIn in;
7774 * supportedValues is currently unused, but added to allow future
7775 * versioning of this function.
7778 *supportedValues = 0;
7779 in.type = htonl(RX_DEBUGI_GETPEER);
7780 in.index = htonl(*nextPeer);
7781 memset(peer, 0, sizeof(*peer));
7783 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7784 &in, sizeof(in), peer, sizeof(*peer));
7790 * Do net to host conversion here
7792 * I don't convert host or port since we are most likely
7793 * going to want these in NBO.
7795 peer->ifMTU = ntohs(peer->ifMTU);
7796 peer->idleWhen = ntohl(peer->idleWhen);
7797 peer->refCount = ntohs(peer->refCount);
7798 peer->rtt = ntohl(peer->rtt);
7799 peer->rtt_dev = ntohl(peer->rtt_dev);
7800 peer->timeout.sec = 0;
7801 peer->timeout.usec = 0;
7802 peer->nSent = ntohl(peer->nSent);
7803 peer->reSends = ntohl(peer->reSends);
7804 peer->natMTU = ntohs(peer->natMTU);
7805 peer->maxMTU = ntohs(peer->maxMTU);
7806 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7807 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7808 peer->MTU = ntohs(peer->MTU);
7809 peer->cwind = ntohs(peer->cwind);
7810 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7811 peer->congestSeq = ntohs(peer->congestSeq);
7812 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7813 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7814 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7815 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7824 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7825 struct rx_debugPeer * peerStats)
7828 afs_int32 error = 1; /* default to "did not succeed" */
7829 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7831 MUTEX_ENTER(&rx_peerHashTable_lock);
7832 for(tp = rx_peerHashTable[hashValue];
7833 tp != NULL; tp = tp->next) {
7834 if (tp->host == peerHost)
7840 MUTEX_EXIT(&rx_peerHashTable_lock);
7844 MUTEX_ENTER(&tp->peer_lock);
7845 peerStats->host = tp->host;
7846 peerStats->port = tp->port;
7847 peerStats->ifMTU = tp->ifMTU;
7848 peerStats->idleWhen = tp->idleWhen;
7849 peerStats->refCount = tp->refCount;
7850 peerStats->burstSize = 0;
7851 peerStats->burst = 0;
7852 peerStats->burstWait.sec = 0;
7853 peerStats->burstWait.usec = 0;
7854 peerStats->rtt = tp->rtt;
7855 peerStats->rtt_dev = tp->rtt_dev;
7856 peerStats->timeout.sec = 0;
7857 peerStats->timeout.usec = 0;
7858 peerStats->nSent = tp->nSent;
7859 peerStats->reSends = tp->reSends;
7860 peerStats->natMTU = tp->natMTU;
7861 peerStats->maxMTU = tp->maxMTU;
7862 peerStats->maxDgramPackets = tp->maxDgramPackets;
7863 peerStats->ifDgramPackets = tp->ifDgramPackets;
7864 peerStats->MTU = tp->MTU;
7865 peerStats->cwind = tp->cwind;
7866 peerStats->nDgramPackets = tp->nDgramPackets;
7867 peerStats->congestSeq = tp->congestSeq;
7868 peerStats->bytesSent.high = tp->bytesSent >> 32;
7869 peerStats->bytesSent.low = tp->bytesSent & MAX_AFS_UINT32;
7870 peerStats->bytesReceived.high = tp->bytesReceived >> 32;
7871 peerStats->bytesReceived.low
7872 = tp->bytesReceived & MAX_AFS_UINT32;
7873 MUTEX_EXIT(&tp->peer_lock);
7875 MUTEX_ENTER(&rx_peerHashTable_lock);
7878 MUTEX_EXIT(&rx_peerHashTable_lock);
7886 struct rx_serverQueueEntry *np;
7889 struct rx_call *call;
7890 struct rx_serverQueueEntry *sq;
7894 if (rxinit_status == 1) {
7896 return; /* Already shutdown. */
7900 #ifndef AFS_PTHREAD_ENV
7901 FD_ZERO(&rx_selectMask);
7902 #endif /* AFS_PTHREAD_ENV */
7903 rxi_dataQuota = RX_MAX_QUOTA;
7904 #ifndef AFS_PTHREAD_ENV
7906 #endif /* AFS_PTHREAD_ENV */
7909 #ifndef AFS_PTHREAD_ENV
7910 #ifndef AFS_USE_GETTIMEOFDAY
7912 #endif /* AFS_USE_GETTIMEOFDAY */
7913 #endif /* AFS_PTHREAD_ENV */
7915 while (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
7916 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
7917 opr_queue_Remove(&call->entry);
7918 rxi_Free(call, sizeof(struct rx_call));
7921 while (!opr_queue_IsEmpty(&rx_idleServerQueue)) {
7922 sq = opr_queue_First(&rx_idleServerQueue, struct rx_serverQueueEntry,
7924 opr_queue_Remove(&sq->entry);
7929 struct rx_peer **peer_ptr, **peer_end;
7930 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7931 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7933 struct rx_peer *peer, *next;
7935 MUTEX_ENTER(&rx_peerHashTable_lock);
7936 for (peer = *peer_ptr; peer; peer = next) {
7937 struct opr_queue *cursor, *store;
7940 MUTEX_ENTER(&rx_rpc_stats);
7941 MUTEX_ENTER(&peer->peer_lock);
7942 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7943 unsigned int num_funcs;
7944 struct rx_interface_stat *rpc_stat
7945 = opr_queue_Entry(cursor, struct rx_interface_stat,
7949 opr_queue_Remove(&rpc_stat->entry);
7950 opr_queue_Remove(&rpc_stat->entryPeers);
7951 num_funcs = rpc_stat->stats[0].func_total;
7953 sizeof(rx_interface_stat_t) +
7954 rpc_stat->stats[0].func_total *
7955 sizeof(rx_function_entry_v1_t);
7957 rxi_Free(rpc_stat, space);
7959 /* rx_rpc_stats must be held */
7960 rxi_rpc_peer_stat_cnt -= num_funcs;
7962 MUTEX_EXIT(&peer->peer_lock);
7963 MUTEX_EXIT(&rx_rpc_stats);
7967 if (rx_stats_active)
7968 rx_atomic_dec(&rx_stats.nPeerStructs);
7970 MUTEX_EXIT(&rx_peerHashTable_lock);
7973 for (i = 0; i < RX_MAX_SERVICES; i++) {
7975 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7977 for (i = 0; i < rx_hashTableSize; i++) {
7978 struct rx_connection *tc, *ntc;
7979 MUTEX_ENTER(&rx_connHashTable_lock);
7980 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7982 for (j = 0; j < RX_MAXCALLS; j++) {
7984 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7987 rxi_Free(tc, sizeof(*tc));
7989 MUTEX_EXIT(&rx_connHashTable_lock);
7992 MUTEX_ENTER(&freeSQEList_lock);
7994 while ((np = rx_FreeSQEList)) {
7995 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7996 MUTEX_DESTROY(&np->lock);
7997 rxi_Free(np, sizeof(*np));
8000 MUTEX_EXIT(&freeSQEList_lock);
8001 MUTEX_DESTROY(&freeSQEList_lock);
8002 MUTEX_DESTROY(&rx_freeCallQueue_lock);
8003 MUTEX_DESTROY(&rx_connHashTable_lock);
8004 MUTEX_DESTROY(&rx_peerHashTable_lock);
8005 MUTEX_DESTROY(&rx_serverPool_lock);
8007 osi_Free(rx_connHashTable,
8008 rx_hashTableSize * sizeof(struct rx_connection *));
8009 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8011 UNPIN(rx_connHashTable,
8012 rx_hashTableSize * sizeof(struct rx_connection *));
8013 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8015 rxi_FreeAllPackets();
8017 MUTEX_ENTER(&rx_quota_mutex);
8018 rxi_dataQuota = RX_MAX_QUOTA;
8019 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
8020 MUTEX_EXIT(&rx_quota_mutex);
8025 #ifdef RX_ENABLE_LOCKS
8027 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
8029 if (!MUTEX_ISMINE(lockaddr))
8030 osi_Panic("Lock not held: %s", msg);
8032 #endif /* RX_ENABLE_LOCKS */
8037 * Routines to implement connection specific data.
8041 rx_KeyCreate(rx_destructor_t rtn)
8044 MUTEX_ENTER(&rxi_keyCreate_lock);
8045 key = rxi_keyCreate_counter++;
8046 rxi_keyCreate_destructor = (rx_destructor_t *)
8047 realloc((void *)rxi_keyCreate_destructor,
8048 (key + 1) * sizeof(rx_destructor_t));
8049 rxi_keyCreate_destructor[key] = rtn;
8050 MUTEX_EXIT(&rxi_keyCreate_lock);
8055 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
8058 MUTEX_ENTER(&conn->conn_data_lock);
8059 if (!conn->specific) {
8060 conn->specific = malloc((key + 1) * sizeof(void *));
8061 for (i = 0; i < key; i++)
8062 conn->specific[i] = NULL;
8063 conn->nSpecific = key + 1;
8064 conn->specific[key] = ptr;
8065 } else if (key >= conn->nSpecific) {
8066 conn->specific = (void **)
8067 realloc(conn->specific, (key + 1) * sizeof(void *));
8068 for (i = conn->nSpecific; i < key; i++)
8069 conn->specific[i] = NULL;
8070 conn->nSpecific = key + 1;
8071 conn->specific[key] = ptr;
8073 if (conn->specific[key] && rxi_keyCreate_destructor[key])
8074 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
8075 conn->specific[key] = ptr;
8077 MUTEX_EXIT(&conn->conn_data_lock);
8081 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
8084 MUTEX_ENTER(&svc->svc_data_lock);
8085 if (!svc->specific) {
8086 svc->specific = malloc((key + 1) * sizeof(void *));
8087 for (i = 0; i < key; i++)
8088 svc->specific[i] = NULL;
8089 svc->nSpecific = key + 1;
8090 svc->specific[key] = ptr;
8091 } else if (key >= svc->nSpecific) {
8092 svc->specific = (void **)
8093 realloc(svc->specific, (key + 1) * sizeof(void *));
8094 for (i = svc->nSpecific; i < key; i++)
8095 svc->specific[i] = NULL;
8096 svc->nSpecific = key + 1;
8097 svc->specific[key] = ptr;
8099 if (svc->specific[key] && rxi_keyCreate_destructor[key])
8100 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
8101 svc->specific[key] = ptr;
8103 MUTEX_EXIT(&svc->svc_data_lock);
8107 rx_GetSpecific(struct rx_connection *conn, int key)
8110 MUTEX_ENTER(&conn->conn_data_lock);
8111 if (key >= conn->nSpecific)
8114 ptr = conn->specific[key];
8115 MUTEX_EXIT(&conn->conn_data_lock);
8120 rx_GetServiceSpecific(struct rx_service *svc, int key)
8123 MUTEX_ENTER(&svc->svc_data_lock);
8124 if (key >= svc->nSpecific)
8127 ptr = svc->specific[key];
8128 MUTEX_EXIT(&svc->svc_data_lock);
8133 #endif /* !KERNEL */
8136 * processStats is a queue used to store the statistics for the local
8137 * process. Its contents are similar to the contents of the rpcStats
8138 * queue on a rx_peer structure, but the actual data stored within
8139 * this queue contains totals across the lifetime of the process (assuming
8140 * the stats have not been reset) - unlike the per peer structures
8141 * which can come and go based upon the peer lifetime.
8144 static struct opr_queue processStats = { &processStats, &processStats };
8147 * peerStats is a queue used to store the statistics for all peer structs.
8148 * Its contents are the union of all the peer rpcStats queues.
8151 static struct opr_queue peerStats = { &peerStats, &peerStats };
8154 * rxi_monitor_processStats is used to turn process wide stat collection
8158 static int rxi_monitor_processStats = 0;
8161 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8164 static int rxi_monitor_peerStats = 0;
8168 rxi_ClearRPCOpStat(rx_function_entry_v1_p rpc_stat)
8170 rpc_stat->invocations = 0;
8171 rpc_stat->bytes_sent = 0;
8172 rpc_stat->bytes_rcvd = 0;
8173 rpc_stat->queue_time_sum.sec = 0;
8174 rpc_stat->queue_time_sum.usec = 0;
8175 rpc_stat->queue_time_sum_sqr.sec = 0;
8176 rpc_stat->queue_time_sum_sqr.usec = 0;
8177 rpc_stat->queue_time_min.sec = 9999999;
8178 rpc_stat->queue_time_min.usec = 9999999;
8179 rpc_stat->queue_time_max.sec = 0;
8180 rpc_stat->queue_time_max.usec = 0;
8181 rpc_stat->execution_time_sum.sec = 0;
8182 rpc_stat->execution_time_sum.usec = 0;
8183 rpc_stat->execution_time_sum_sqr.sec = 0;
8184 rpc_stat->execution_time_sum_sqr.usec = 0;
8185 rpc_stat->execution_time_min.sec = 9999999;
8186 rpc_stat->execution_time_min.usec = 9999999;
8187 rpc_stat->execution_time_max.sec = 0;
8188 rpc_stat->execution_time_max.usec = 0;
8192 * Given all of the information for a particular rpc
8193 * call, find or create (if requested) the stat structure for the rpc.
8196 * the queue of stats that will be updated with the new value
8198 * @param rxInterface
8199 * a unique number that identifies the rpc interface
8202 * the total number of functions in this interface. this is only
8203 * required if create is true
8206 * if true, this invocation was made to a server
8209 * the ip address of the remote host. this is only required if create
8210 * and addToPeerList are true
8213 * the port of the remote host. this is only required if create
8214 * and addToPeerList are true
8216 * @param addToPeerList
8217 * if != 0, add newly created stat to the global peer list
8220 * if a new stats structure is allocated, the counter will
8221 * be updated with the new number of allocated stat structures.
8222 * only required if create is true
8225 * if no stats structure exists, allocate one
8229 static rx_interface_stat_p
8230 rxi_FindRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8231 afs_uint32 totalFunc, int isServer, afs_uint32 remoteHost,
8232 afs_uint32 remotePort, int addToPeerList,
8233 unsigned int *counter, int create)
8235 rx_interface_stat_p rpc_stat = NULL;
8236 struct opr_queue *cursor;
8239 * See if there's already a structure for this interface
8242 for (opr_queue_Scan(stats, cursor)) {
8243 rpc_stat = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8245 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8246 && (rpc_stat->stats[0].remote_is_server == isServer))
8250 /* if they didn't ask us to create, we're done */
8252 if (opr_queue_IsEnd(stats, cursor))
8258 /* can't proceed without these */
8259 if (!totalFunc || !counter)
8263 * Didn't find a match so allocate a new structure and add it to the
8267 if (opr_queue_IsEnd(stats, cursor) || (rpc_stat == NULL)
8268 || (rpc_stat->stats[0].interfaceId != rxInterface)
8269 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8274 sizeof(rx_interface_stat_t) +
8275 totalFunc * sizeof(rx_function_entry_v1_t);
8277 rpc_stat = rxi_Alloc(space);
8278 if (rpc_stat == NULL)
8281 *counter += totalFunc;
8282 for (i = 0; i < totalFunc; i++) {
8283 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8284 rpc_stat->stats[i].remote_peer = remoteHost;
8285 rpc_stat->stats[i].remote_port = remotePort;
8286 rpc_stat->stats[i].remote_is_server = isServer;
8287 rpc_stat->stats[i].interfaceId = rxInterface;
8288 rpc_stat->stats[i].func_total = totalFunc;
8289 rpc_stat->stats[i].func_index = i;
8291 opr_queue_Prepend(stats, &rpc_stat->entry);
8292 if (addToPeerList) {
8293 opr_queue_Prepend(&peerStats, &rpc_stat->entryPeers);
8300 rx_ClearProcessRPCStats(afs_int32 rxInterface)
8302 rx_interface_stat_p rpc_stat;
8305 if (rxInterface == -1)
8308 MUTEX_ENTER(&rx_rpc_stats);
8309 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8312 totalFunc = rpc_stat->stats[0].func_total;
8313 for (i = 0; i < totalFunc; i++)
8314 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8316 MUTEX_EXIT(&rx_rpc_stats);
8321 rx_ClearPeerRPCStats(afs_int32 rxInterface, afs_uint32 peerHost, afs_uint16 peerPort)
8323 rx_interface_stat_p rpc_stat;
8325 struct rx_peer * peer;
8327 if (rxInterface == -1)
8330 peer = rxi_FindPeer(peerHost, peerPort, 0, 0);
8334 MUTEX_ENTER(&rx_rpc_stats);
8335 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8338 totalFunc = rpc_stat->stats[0].func_total;
8339 for (i = 0; i < totalFunc; i++)
8340 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8342 MUTEX_EXIT(&rx_rpc_stats);
8347 rx_CopyProcessRPCStats(afs_uint64 op)
8349 rx_interface_stat_p rpc_stat;
8350 rx_function_entry_v1_p rpcop_stat =
8351 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8352 int currentFunc = (op & MAX_AFS_UINT32);
8353 afs_int32 rxInterface = (op >> 32);
8355 if (!rxi_monitor_processStats)
8358 if (rxInterface == -1)
8361 if (rpcop_stat == NULL)
8364 MUTEX_ENTER(&rx_rpc_stats);
8365 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8368 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8369 sizeof(rx_function_entry_v1_t));
8370 MUTEX_EXIT(&rx_rpc_stats);
8372 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8379 rx_CopyPeerRPCStats(afs_uint64 op, afs_uint32 peerHost, afs_uint16 peerPort)
8381 rx_interface_stat_p rpc_stat;
8382 rx_function_entry_v1_p rpcop_stat =
8383 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8384 int currentFunc = (op & MAX_AFS_UINT32);
8385 afs_int32 rxInterface = (op >> 32);
8386 struct rx_peer *peer;
8388 if (!rxi_monitor_peerStats)
8391 if (rxInterface == -1)
8394 if (rpcop_stat == NULL)
8397 peer = rxi_FindPeer(peerHost, peerPort, 0, 0);
8401 MUTEX_ENTER(&rx_rpc_stats);
8402 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8405 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8406 sizeof(rx_function_entry_v1_t));
8407 MUTEX_EXIT(&rx_rpc_stats);
8409 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8416 rx_ReleaseRPCStats(void *stats)
8419 rxi_Free(stats, sizeof(rx_function_entry_v1_t));
8423 * Given all of the information for a particular rpc
8424 * call, create (if needed) and update the stat totals for the rpc.
8427 * the queue of stats that will be updated with the new value
8429 * @param rxInterface
8430 * a unique number that identifies the rpc interface
8432 * @param currentFunc
8433 * the index of the function being invoked
8436 * the total number of functions in this interface
8439 * the amount of time this function waited for a thread
8442 * the amount of time this function invocation took to execute
8445 * the number bytes sent by this invocation
8448 * the number bytes received by this invocation
8451 * if true, this invocation was made to a server
8454 * the ip address of the remote host
8457 * the port of the remote host
8459 * @param addToPeerList
8460 * if != 0, add newly created stat to the global peer list
8463 * if a new stats structure is allocated, the counter will
8464 * be updated with the new number of allocated stat structures
8469 rxi_AddRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8470 afs_uint32 currentFunc, afs_uint32 totalFunc,
8471 struct clock *queueTime, struct clock *execTime,
8472 afs_uint64 bytesSent, afs_uint64 bytesRcvd, int isServer,
8473 afs_uint32 remoteHost, afs_uint32 remotePort,
8474 int addToPeerList, unsigned int *counter)
8477 rx_interface_stat_p rpc_stat;
8479 rpc_stat = rxi_FindRpcStat(stats, rxInterface, totalFunc, isServer,
8480 remoteHost, remotePort, addToPeerList, counter,
8488 * Increment the stats for this function
8491 rpc_stat->stats[currentFunc].invocations++;
8492 rpc_stat->stats[currentFunc].bytes_sent += bytesSent;
8493 rpc_stat->stats[currentFunc].bytes_rcvd += bytesRcvd;
8494 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8495 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8496 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8497 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8499 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8500 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8502 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8503 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8505 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8506 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8508 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8509 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8517 rxi_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8518 afs_uint32 currentFunc, afs_uint32 totalFunc,
8519 struct clock *queueTime, struct clock *execTime,
8520 afs_uint64 bytesSent, afs_uint64 bytesRcvd,
8524 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8527 MUTEX_ENTER(&rx_rpc_stats);
8529 if (rxi_monitor_peerStats) {
8530 MUTEX_ENTER(&peer->peer_lock);
8531 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8532 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8533 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8534 MUTEX_EXIT(&peer->peer_lock);
8537 if (rxi_monitor_processStats) {
8538 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8539 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8540 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8543 MUTEX_EXIT(&rx_rpc_stats);
8547 * Increment the times and count for a particular rpc function.
8549 * Traditionally this call was invoked from rxgen stubs. Modern stubs
8550 * call rx_RecordCallStatistics instead, so the public version of this
8551 * function is left purely for legacy callers.
8554 * The peer who invoked the rpc
8556 * @param rxInterface
8557 * A unique number that identifies the rpc interface
8559 * @param currentFunc
8560 * The index of the function being invoked
8563 * The total number of functions in this interface
8566 * The amount of time this function waited for a thread
8569 * The amount of time this function invocation took to execute
8572 * The number bytes sent by this invocation
8575 * The number bytes received by this invocation
8578 * If true, this invocation was made to a server
8582 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8583 afs_uint32 currentFunc, afs_uint32 totalFunc,
8584 struct clock *queueTime, struct clock *execTime,
8585 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8591 sent64 = ((afs_uint64)bytesSent->high << 32) + bytesSent->low;
8592 rcvd64 = ((afs_uint64)bytesRcvd->high << 32) + bytesRcvd->low;
8594 rxi_IncrementTimeAndCount(peer, rxInterface, currentFunc, totalFunc,
8595 queueTime, execTime, sent64, rcvd64,
8602 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8606 * IN callerVersion - the rpc stat version of the caller.
8608 * IN count - the number of entries to marshall.
8610 * IN stats - pointer to stats to be marshalled.
8612 * OUT ptr - Where to store the marshalled data.
8619 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8620 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8626 * We only support the first version
8628 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8629 *(ptr++) = stats->remote_peer;
8630 *(ptr++) = stats->remote_port;
8631 *(ptr++) = stats->remote_is_server;
8632 *(ptr++) = stats->interfaceId;
8633 *(ptr++) = stats->func_total;
8634 *(ptr++) = stats->func_index;
8635 *(ptr++) = stats->invocations >> 32;
8636 *(ptr++) = stats->invocations & MAX_AFS_UINT32;
8637 *(ptr++) = stats->bytes_sent >> 32;
8638 *(ptr++) = stats->bytes_sent & MAX_AFS_UINT32;
8639 *(ptr++) = stats->bytes_rcvd >> 32;
8640 *(ptr++) = stats->bytes_rcvd & MAX_AFS_UINT32;
8641 *(ptr++) = stats->queue_time_sum.sec;
8642 *(ptr++) = stats->queue_time_sum.usec;
8643 *(ptr++) = stats->queue_time_sum_sqr.sec;
8644 *(ptr++) = stats->queue_time_sum_sqr.usec;
8645 *(ptr++) = stats->queue_time_min.sec;
8646 *(ptr++) = stats->queue_time_min.usec;
8647 *(ptr++) = stats->queue_time_max.sec;
8648 *(ptr++) = stats->queue_time_max.usec;
8649 *(ptr++) = stats->execution_time_sum.sec;
8650 *(ptr++) = stats->execution_time_sum.usec;
8651 *(ptr++) = stats->execution_time_sum_sqr.sec;
8652 *(ptr++) = stats->execution_time_sum_sqr.usec;
8653 *(ptr++) = stats->execution_time_min.sec;
8654 *(ptr++) = stats->execution_time_min.usec;
8655 *(ptr++) = stats->execution_time_max.sec;
8656 *(ptr++) = stats->execution_time_max.usec;
8662 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8667 * IN callerVersion - the rpc stat version of the caller
8669 * OUT myVersion - the rpc stat version of this function
8671 * OUT clock_sec - local time seconds
8673 * OUT clock_usec - local time microseconds
8675 * OUT allocSize - the number of bytes allocated to contain stats
8677 * OUT statCount - the number stats retrieved from this process.
8679 * OUT stats - the actual stats retrieved from this process.
8683 * Returns void. If successful, stats will != NULL.
8687 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8688 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8689 size_t * allocSize, afs_uint32 * statCount,
8690 afs_uint32 ** stats)
8700 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8703 * Check to see if stats are enabled
8706 MUTEX_ENTER(&rx_rpc_stats);
8707 if (!rxi_monitor_processStats) {
8708 MUTEX_EXIT(&rx_rpc_stats);
8712 clock_GetTime(&now);
8713 *clock_sec = now.sec;
8714 *clock_usec = now.usec;
8717 * Allocate the space based upon the caller version
8719 * If the client is at an older version than we are,
8720 * we return the statistic data in the older data format, but
8721 * we still return our version number so the client knows we
8722 * are maintaining more data than it can retrieve.
8725 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8726 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8727 *statCount = rxi_rpc_process_stat_cnt;
8730 * This can't happen yet, but in the future version changes
8731 * can be handled by adding additional code here
8735 if (space > (size_t) 0) {
8737 ptr = *stats = rxi_Alloc(space);
8740 struct opr_queue *cursor;
8742 for (opr_queue_Scan(&processStats, cursor)) {
8743 struct rx_interface_stat *rpc_stat =
8744 opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8746 * Copy the data based upon the caller version
8748 rx_MarshallProcessRPCStats(callerVersion,
8749 rpc_stat->stats[0].func_total,
8750 rpc_stat->stats, &ptr);
8756 MUTEX_EXIT(&rx_rpc_stats);
8761 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8765 * IN callerVersion - the rpc stat version of the caller
8767 * OUT myVersion - the rpc stat version of this function
8769 * OUT clock_sec - local time seconds
8771 * OUT clock_usec - local time microseconds
8773 * OUT allocSize - the number of bytes allocated to contain stats
8775 * OUT statCount - the number of stats retrieved from the individual
8778 * OUT stats - the actual stats retrieved from the individual peer structures.
8782 * Returns void. If successful, stats will != NULL.
8786 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8787 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8788 size_t * allocSize, afs_uint32 * statCount,
8789 afs_uint32 ** stats)
8799 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8802 * Check to see if stats are enabled
8805 MUTEX_ENTER(&rx_rpc_stats);
8806 if (!rxi_monitor_peerStats) {
8807 MUTEX_EXIT(&rx_rpc_stats);
8811 clock_GetTime(&now);
8812 *clock_sec = now.sec;
8813 *clock_usec = now.usec;
8816 * Allocate the space based upon the caller version
8818 * If the client is at an older version than we are,
8819 * we return the statistic data in the older data format, but
8820 * we still return our version number so the client knows we
8821 * are maintaining more data than it can retrieve.
8824 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8825 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8826 *statCount = rxi_rpc_peer_stat_cnt;
8829 * This can't happen yet, but in the future version changes
8830 * can be handled by adding additional code here
8834 if (space > (size_t) 0) {
8836 ptr = *stats = rxi_Alloc(space);
8839 struct opr_queue *cursor;
8841 for (opr_queue_Scan(&peerStats, cursor)) {
8842 struct rx_interface_stat *rpc_stat
8843 = opr_queue_Entry(cursor, struct rx_interface_stat,
8847 * Copy the data based upon the caller version
8849 rx_MarshallProcessRPCStats(callerVersion,
8850 rpc_stat->stats[0].func_total,
8851 rpc_stat->stats, &ptr);
8857 MUTEX_EXIT(&rx_rpc_stats);
8862 * rx_FreeRPCStats - free memory allocated by
8863 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8867 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8868 * rx_RetrievePeerRPCStats
8870 * IN allocSize - the number of bytes in stats.
8878 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8880 rxi_Free(stats, allocSize);
8884 * rx_queryProcessRPCStats - see if process rpc stat collection is
8885 * currently enabled.
8891 * Returns 0 if stats are not enabled != 0 otherwise
8895 rx_queryProcessRPCStats(void)
8898 MUTEX_ENTER(&rx_rpc_stats);
8899 rc = rxi_monitor_processStats;
8900 MUTEX_EXIT(&rx_rpc_stats);
8905 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8911 * Returns 0 if stats are not enabled != 0 otherwise
8915 rx_queryPeerRPCStats(void)
8918 MUTEX_ENTER(&rx_rpc_stats);
8919 rc = rxi_monitor_peerStats;
8920 MUTEX_EXIT(&rx_rpc_stats);
8925 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8935 rx_enableProcessRPCStats(void)
8937 MUTEX_ENTER(&rx_rpc_stats);
8938 rx_enable_stats = 1;
8939 rxi_monitor_processStats = 1;
8940 MUTEX_EXIT(&rx_rpc_stats);
8944 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8954 rx_enablePeerRPCStats(void)
8956 MUTEX_ENTER(&rx_rpc_stats);
8957 rx_enable_stats = 1;
8958 rxi_monitor_peerStats = 1;
8959 MUTEX_EXIT(&rx_rpc_stats);
8963 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8973 rx_disableProcessRPCStats(void)
8975 struct opr_queue *cursor, *store;
8978 MUTEX_ENTER(&rx_rpc_stats);
8981 * Turn off process statistics and if peer stats is also off, turn
8985 rxi_monitor_processStats = 0;
8986 if (rxi_monitor_peerStats == 0) {
8987 rx_enable_stats = 0;
8990 for (opr_queue_ScanSafe(&processStats, cursor, store)) {
8991 unsigned int num_funcs = 0;
8992 struct rx_interface_stat *rpc_stat
8993 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8995 opr_queue_Remove(&rpc_stat->entry);
8997 num_funcs = rpc_stat->stats[0].func_total;
8999 sizeof(rx_interface_stat_t) +
9000 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
9002 rxi_Free(rpc_stat, space);
9003 rxi_rpc_process_stat_cnt -= num_funcs;
9005 MUTEX_EXIT(&rx_rpc_stats);
9009 * rx_disablePeerRPCStats - stop rpc stat collection for peers
9019 rx_disablePeerRPCStats(void)
9021 struct rx_peer **peer_ptr, **peer_end;
9025 * Turn off peer statistics and if process stats is also off, turn
9029 rxi_monitor_peerStats = 0;
9030 if (rxi_monitor_processStats == 0) {
9031 rx_enable_stats = 0;
9034 for (peer_ptr = &rx_peerHashTable[0], peer_end =
9035 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
9037 struct rx_peer *peer, *next, *prev;
9039 MUTEX_ENTER(&rx_peerHashTable_lock);
9040 MUTEX_ENTER(&rx_rpc_stats);
9041 for (prev = peer = *peer_ptr; peer; peer = next) {
9043 code = MUTEX_TRYENTER(&peer->peer_lock);
9046 struct opr_queue *cursor, *store;
9048 if (prev == *peer_ptr) {
9059 MUTEX_EXIT(&rx_peerHashTable_lock);
9061 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
9062 unsigned int num_funcs = 0;
9063 struct rx_interface_stat *rpc_stat
9064 = opr_queue_Entry(cursor, struct rx_interface_stat,
9067 opr_queue_Remove(&rpc_stat->entry);
9068 opr_queue_Remove(&rpc_stat->entryPeers);
9069 num_funcs = rpc_stat->stats[0].func_total;
9071 sizeof(rx_interface_stat_t) +
9072 rpc_stat->stats[0].func_total *
9073 sizeof(rx_function_entry_v1_t);
9075 rxi_Free(rpc_stat, space);
9076 rxi_rpc_peer_stat_cnt -= num_funcs;
9078 MUTEX_EXIT(&peer->peer_lock);
9080 MUTEX_ENTER(&rx_peerHashTable_lock);
9090 MUTEX_EXIT(&rx_rpc_stats);
9091 MUTEX_EXIT(&rx_peerHashTable_lock);
9096 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
9101 * IN clearFlag - flag indicating which stats to clear
9109 rx_clearProcessRPCStats(afs_uint32 clearFlag)
9111 struct opr_queue *cursor;
9113 MUTEX_ENTER(&rx_rpc_stats);
9115 for (opr_queue_Scan(&processStats, cursor)) {
9116 unsigned int num_funcs = 0, i;
9117 struct rx_interface_stat *rpc_stat
9118 = opr_queue_Entry(rpc_stat, struct rx_interface_stat, entry);
9120 num_funcs = rpc_stat->stats[0].func_total;
9121 for (i = 0; i < num_funcs; i++) {
9122 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9123 rpc_stat->stats[i].invocations = 0;
9125 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9126 rpc_stat->stats[i].bytes_sent = 0;
9128 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9129 rpc_stat->stats[i].bytes_rcvd = 0;
9131 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9132 rpc_stat->stats[i].queue_time_sum.sec = 0;
9133 rpc_stat->stats[i].queue_time_sum.usec = 0;
9135 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9136 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9137 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9139 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9140 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9141 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9143 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9144 rpc_stat->stats[i].queue_time_max.sec = 0;
9145 rpc_stat->stats[i].queue_time_max.usec = 0;
9147 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9148 rpc_stat->stats[i].execution_time_sum.sec = 0;
9149 rpc_stat->stats[i].execution_time_sum.usec = 0;
9151 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9152 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9153 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9155 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9156 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9157 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9159 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9160 rpc_stat->stats[i].execution_time_max.sec = 0;
9161 rpc_stat->stats[i].execution_time_max.usec = 0;
9166 MUTEX_EXIT(&rx_rpc_stats);
9170 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
9175 * IN clearFlag - flag indicating which stats to clear
9183 rx_clearPeerRPCStats(afs_uint32 clearFlag)
9185 struct opr_queue *cursor;
9187 MUTEX_ENTER(&rx_rpc_stats);
9189 for (opr_queue_Scan(&peerStats, cursor)) {
9190 unsigned int num_funcs, i;
9191 struct rx_interface_stat *rpc_stat
9192 = opr_queue_Entry(cursor, struct rx_interface_stat, entryPeers);
9194 num_funcs = rpc_stat->stats[0].func_total;
9195 for (i = 0; i < num_funcs; i++) {
9196 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9197 rpc_stat->stats[i].invocations = 0;
9199 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9200 rpc_stat->stats[i].bytes_sent = 0;
9202 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9203 rpc_stat->stats[i].bytes_rcvd = 0;
9205 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9206 rpc_stat->stats[i].queue_time_sum.sec = 0;
9207 rpc_stat->stats[i].queue_time_sum.usec = 0;
9209 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9210 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9211 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9213 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9214 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9215 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9217 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9218 rpc_stat->stats[i].queue_time_max.sec = 0;
9219 rpc_stat->stats[i].queue_time_max.usec = 0;
9221 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9222 rpc_stat->stats[i].execution_time_sum.sec = 0;
9223 rpc_stat->stats[i].execution_time_sum.usec = 0;
9225 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9226 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9227 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9229 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9230 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9231 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9233 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9234 rpc_stat->stats[i].execution_time_max.sec = 0;
9235 rpc_stat->stats[i].execution_time_max.usec = 0;
9240 MUTEX_EXIT(&rx_rpc_stats);
9244 * rxi_rxstat_userok points to a routine that returns 1 if the caller
9245 * is authorized to enable/disable/clear RX statistics.
9247 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
9250 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
9252 rxi_rxstat_userok = proc;
9256 rx_RxStatUserOk(struct rx_call *call)
9258 if (!rxi_rxstat_userok)
9260 return rxi_rxstat_userok(call);
9265 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
9266 * function in the MSVC runtime DLL (msvcrt.dll).
9268 * Note: the system serializes calls to this function.
9271 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
9272 DWORD reason, /* reason function is being called */
9273 LPVOID reserved) /* reserved for future use */
9276 case DLL_PROCESS_ATTACH:
9277 /* library is being attached to a process */
9281 case DLL_PROCESS_DETACH:
9288 #endif /* AFS_NT40_ENV */
9291 int rx_DumpCalls(FILE *outputFile, char *cookie)
9293 #ifdef RXDEBUG_PACKET
9294 #ifdef KDUMP_RX_LOCK
9295 struct rx_call_rx_lock *c;
9302 #define RXDPRINTF sprintf
9303 #define RXDPRINTOUT output
9305 #define RXDPRINTF fprintf
9306 #define RXDPRINTOUT outputFile
9309 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9311 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9314 for (c = rx_allCallsp; c; c = c->allNextp) {
9315 u_short rqc, tqc, iovqc;
9317 MUTEX_ENTER(&c->lock);
9318 rqc = opr_queue_Count(&c->rq);
9319 tqc = opr_queue_Count(&c->tq);
9320 iovqc = opr_queue_Count(&c->app.iovq);
9322 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, "
9323 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9324 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9325 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9326 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
9327 #ifdef RX_ENABLE_LOCKS
9330 #ifdef RX_REFCOUNT_CHECK
9331 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9332 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9335 cookie, c, c->call_id, (afs_uint32)c->state, (afs_uint32)c->app.mode, c->conn, c->conn?c->conn->epoch:0, c->conn?c->conn->cid:0,
9336 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9337 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9338 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9339 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9340 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
9341 #ifdef RX_ENABLE_LOCKS
9342 , (afs_uint32)c->refCount
9344 #ifdef RX_REFCOUNT_CHECK
9345 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9348 MUTEX_EXIT(&c->lock);
9351 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9354 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9356 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9358 #endif /* RXDEBUG_PACKET */