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);
3212 * Abort the call if the server is over the busy threshold. This
3213 * can be used without requiring a call structure be initialised,
3214 * or connected to a particular channel
3217 rxi_AbortIfServerBusy(osi_socket socket, afs_uint32 host,
3218 u_short port, struct rx_packet *np)
3220 if ((rx_BusyThreshold > 0) &&
3221 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3222 rxi_SendRawAbort(socket, host, port, rx_BusyError, np, 0);
3223 if (rx_stats_active)
3224 rx_atomic_inc(&rx_stats.nBusies);
3231 /* There are two packet tracing routines available for testing and monitoring
3232 * Rx. One is called just after every packet is received and the other is
3233 * called just before every packet is sent. Received packets, have had their
3234 * headers decoded, and packets to be sent have not yet had their headers
3235 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3236 * containing the network address. Both can be modified. The return value, if
3237 * non-zero, indicates that the packet should be dropped. */
3239 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3240 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3242 /* A packet has been received off the interface. Np is the packet, socket is
3243 * the socket number it was received from (useful in determining which service
3244 * this packet corresponds to), and (host, port) reflect the host,port of the
3245 * sender. This call returns the packet to the caller if it is finished with
3246 * it, rather than de-allocating it, just as a small performance hack */
3249 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3250 afs_uint32 host, u_short port, int *tnop,
3251 struct rx_call **newcallp)
3253 struct rx_call *call;
3254 struct rx_connection *conn;
3256 afs_uint32 currentCallNumber;
3261 struct rx_packet *tnp;
3264 /* We don't print out the packet until now because (1) the time may not be
3265 * accurate enough until now in the lwp implementation (rx_Listener only gets
3266 * the time after the packet is read) and (2) from a protocol point of view,
3267 * this is the first time the packet has been seen */
3268 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3269 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3270 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3271 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3272 np->header.epoch, np->header.cid, np->header.callNumber,
3273 np->header.seq, np->header.flags, np));
3276 /* Account for connectionless packets */
3277 if (rx_stats_active &&
3278 ((np->header.type == RX_PACKET_TYPE_VERSION) ||
3279 (np->header.type == RX_PACKET_TYPE_DEBUG))) {
3280 struct rx_peer *peer;
3282 /* Try to look up the peer structure, but don't create one */
3283 peer = rxi_FindPeer(host, port, 0, 0);
3285 /* Since this may not be associated with a connection, it may have
3286 * no refCount, meaning we could race with ReapConnections
3289 if (peer && (peer->refCount > 0)) {
3290 #ifdef AFS_RXERRQ_ENV
3291 if (rx_atomic_read(&peer->neterrs)) {
3292 rx_atomic_set(&peer->neterrs, 0);
3295 MUTEX_ENTER(&peer->peer_lock);
3296 peer->bytesReceived += np->length;
3297 MUTEX_EXIT(&peer->peer_lock);
3301 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3302 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3305 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3306 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3309 /* If an input tracer function is defined, call it with the packet and
3310 * network address. Note this function may modify its arguments. */
3311 if (rx_justReceived) {
3312 struct sockaddr_in addr;
3314 addr.sin_family = AF_INET;
3315 addr.sin_port = port;
3316 addr.sin_addr.s_addr = host;
3317 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3318 addr.sin_len = sizeof(addr);
3319 #endif /* AFS_OSF_ENV */
3320 drop = (*rx_justReceived) (np, &addr);
3321 /* drop packet if return value is non-zero */
3324 port = addr.sin_port; /* in case fcn changed addr */
3325 host = addr.sin_addr.s_addr;
3329 /* If packet was not sent by the client, then *we* must be the client */
3330 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3331 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3333 /* Find the connection (or fabricate one, if we're the server & if
3334 * necessary) associated with this packet */
3336 rxi_FindConnection(socket, host, port, np->header.serviceId,
3337 np->header.cid, np->header.epoch, type,
3338 np->header.securityIndex);
3340 /* To avoid having 2 connections just abort at each other,
3341 don't abort an abort. */
3343 if (np->header.type != RX_PACKET_TYPE_ABORT)
3344 rxi_SendRawAbort(socket, host, port, RX_INVALID_OPERATION,
3349 #ifdef AFS_RXERRQ_ENV
3350 if (rx_atomic_read(&conn->peer->neterrs)) {
3351 rx_atomic_set(&conn->peer->neterrs, 0);
3355 /* If we're doing statistics, then account for the incoming packet */
3356 if (rx_stats_active) {
3357 MUTEX_ENTER(&conn->peer->peer_lock);
3358 conn->peer->bytesReceived += np->length;
3359 MUTEX_EXIT(&conn->peer->peer_lock);
3362 /* If the connection is in an error state, send an abort packet and ignore
3363 * the incoming packet */
3365 /* Don't respond to an abort packet--we don't want loops! */
3366 MUTEX_ENTER(&conn->conn_data_lock);
3367 if (np->header.type != RX_PACKET_TYPE_ABORT)
3368 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3369 putConnection(conn);
3370 MUTEX_EXIT(&conn->conn_data_lock);
3374 /* Check for connection-only requests (i.e. not call specific). */
3375 if (np->header.callNumber == 0) {
3376 switch (np->header.type) {
3377 case RX_PACKET_TYPE_ABORT: {
3378 /* What if the supplied error is zero? */
3379 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3380 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3381 rxi_ConnectionError(conn, errcode);
3382 putConnection(conn);
3385 case RX_PACKET_TYPE_CHALLENGE:
3386 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3387 putConnection(conn);
3389 case RX_PACKET_TYPE_RESPONSE:
3390 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3391 putConnection(conn);
3393 case RX_PACKET_TYPE_PARAMS:
3394 case RX_PACKET_TYPE_PARAMS + 1:
3395 case RX_PACKET_TYPE_PARAMS + 2:
3396 /* ignore these packet types for now */
3397 putConnection(conn);
3401 /* Should not reach here, unless the peer is broken: send an
3403 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3404 MUTEX_ENTER(&conn->conn_data_lock);
3405 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3406 putConnection(conn);
3407 MUTEX_EXIT(&conn->conn_data_lock);
3412 channel = np->header.cid & RX_CHANNELMASK;
3413 MUTEX_ENTER(&conn->conn_call_lock);
3414 call = conn->call[channel];
3417 MUTEX_ENTER(&call->lock);
3418 currentCallNumber = conn->callNumber[channel];
3419 MUTEX_EXIT(&conn->conn_call_lock);
3420 } else if (type == RX_SERVER_CONNECTION) { /* No call allocated */
3422 if (rxi_AbortIfServerBusy(socket, host, port, np)) {
3423 putConnection(conn);
3427 call = rxi_NewCall(conn, channel); /* returns locked call */
3428 *call->callNumber = currentCallNumber = np->header.callNumber;
3429 MUTEX_EXIT(&conn->conn_call_lock);
3430 call->state = RX_STATE_PRECALL;
3431 clock_GetTime(&call->queueTime);
3432 call->app.bytesSent = 0;
3433 call->app.bytesRcvd = 0;
3434 rxi_KeepAliveOn(call);
3436 } else { /* RX_CLIENT_CONNECTION and No call allocated */
3437 /* This packet can't be for this call. If the new call address is
3438 * 0 then no call is running on this channel. If there is a call
3439 * then, since this is a client connection we're getting data for
3440 * it must be for the previous call.
3442 MUTEX_EXIT(&conn->conn_call_lock);
3443 if (rx_stats_active)
3444 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3445 putConnection(conn);
3449 /* There is a non-NULL locked call at this point */
3450 if (type == RX_SERVER_CONNECTION) { /* We're the server */
3451 if (np->header.callNumber < currentCallNumber) {
3452 MUTEX_EXIT(&call->lock);
3453 if (rx_stats_active)
3454 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3455 putConnection(conn);
3457 } else if (np->header.callNumber != currentCallNumber) {
3458 /* Wait until the transmit queue is idle before deciding
3459 * whether to reset the current call. Chances are that the
3460 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3463 #ifdef RX_ENABLE_LOCKS
3464 if (call->state == RX_STATE_ACTIVE) {
3465 rxi_WaitforTQBusy(call);
3467 * If we entered error state while waiting,
3468 * must call rxi_CallError to permit rxi_ResetCall
3469 * to processed when the tqWaiter count hits zero.
3472 rxi_CallError(call, call->error);
3473 MUTEX_EXIT(&call->lock);
3474 putConnection(conn);
3478 #endif /* RX_ENABLE_LOCKS */
3479 /* If the new call cannot be taken right now send a busy and set
3480 * the error condition in this call, so that it terminates as
3481 * quickly as possible */
3482 if (call->state == RX_STATE_ACTIVE) {
3483 struct rx_packet *tp;
3485 rxi_CallError(call, RX_CALL_DEAD);
3486 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3488 MUTEX_EXIT(&call->lock);
3489 putConnection(conn);
3493 if (rxi_AbortIfServerBusy(socket, host, port, np)) {
3494 MUTEX_EXIT(&call->lock);
3495 putConnection(conn);
3499 rxi_ResetCall(call, 0);
3501 * The conn_call_lock is not held but no one else should be
3502 * using this call channel while we are processing this incoming
3503 * packet. This assignment should be safe.
3505 *call->callNumber = np->header.callNumber;
3506 call->state = RX_STATE_PRECALL;
3507 clock_GetTime(&call->queueTime);
3508 call->app.bytesSent = 0;
3509 call->app.bytesRcvd = 0;
3510 rxi_KeepAliveOn(call);
3512 /* Continuing call; do nothing here. */
3514 } else { /* we're the client */
3515 /* Ignore all incoming acknowledgements for calls in DALLY state */
3516 if ((call->state == RX_STATE_DALLY)
3517 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3518 if (rx_stats_active)
3519 rx_atomic_inc(&rx_stats.ignorePacketDally);
3520 MUTEX_EXIT(&call->lock);
3521 putConnection(conn);
3525 /* Ignore anything that's not relevant to the current call. If there
3526 * isn't a current call, then no packet is relevant. */
3527 if (np->header.callNumber != currentCallNumber) {
3528 if (rx_stats_active)
3529 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3530 MUTEX_EXIT(&call->lock);
3531 putConnection(conn);
3534 /* If the service security object index stamped in the packet does not
3535 * match the connection's security index, ignore the packet */
3536 if (np->header.securityIndex != conn->securityIndex) {
3537 MUTEX_EXIT(&call->lock);
3538 putConnection(conn);
3542 /* If we're receiving the response, then all transmit packets are
3543 * implicitly acknowledged. Get rid of them. */
3544 if (np->header.type == RX_PACKET_TYPE_DATA) {
3545 #ifdef RX_ENABLE_LOCKS
3546 /* XXX Hack. Because we must release the call lock when
3547 * sending packets (osi_NetSend) we drop all acks while we're
3548 * traversing the tq in rxi_Start sending packets out because
3549 * packets may move to the freePacketQueue as result of being here!
3550 * So we drop these packets until we're safely out of the
3551 * traversing. Really ugly!
3552 * For fine grain RX locking, we set the acked field in the
3553 * packets and let rxi_Start remove them from the transmit queue.
3555 if (call->flags & RX_CALL_TQ_BUSY) {
3556 rxi_SetAcksInTransmitQueue(call);
3558 rxi_ClearTransmitQueue(call, 0);
3560 #else /* RX_ENABLE_LOCKS */
3561 rxi_ClearTransmitQueue(call, 0);
3562 #endif /* RX_ENABLE_LOCKS */
3564 if (np->header.type == RX_PACKET_TYPE_ACK) {
3565 /* now check to see if this is an ack packet acknowledging that the
3566 * server actually *lost* some hard-acked data. If this happens we
3567 * ignore this packet, as it may indicate that the server restarted in
3568 * the middle of a call. It is also possible that this is an old ack
3569 * packet. We don't abort the connection in this case, because this
3570 * *might* just be an old ack packet. The right way to detect a server
3571 * restart in the midst of a call is to notice that the server epoch
3573 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3574 * XXX unacknowledged. I think that this is off-by-one, but
3575 * XXX I don't dare change it just yet, since it will
3576 * XXX interact badly with the server-restart detection
3577 * XXX code in receiveackpacket. */
3578 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3579 if (rx_stats_active)
3580 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3581 MUTEX_EXIT(&call->lock);
3582 putConnection(conn);
3586 } /* else not a data packet */
3589 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3590 /* Set remote user defined status from packet */
3591 call->remoteStatus = np->header.userStatus;
3593 /* Now do packet type-specific processing */
3594 switch (np->header.type) {
3595 case RX_PACKET_TYPE_DATA:
3596 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3599 case RX_PACKET_TYPE_ACK:
3600 /* Respond immediately to ack packets requesting acknowledgement
3602 if (np->header.flags & RX_REQUEST_ACK) {
3604 (void)rxi_SendCallAbort(call, 0, 1, 0);
3606 (void)rxi_SendAck(call, 0, np->header.serial,
3607 RX_ACK_PING_RESPONSE, 1);
3609 np = rxi_ReceiveAckPacket(call, np, 1);
3611 case RX_PACKET_TYPE_ABORT: {
3612 /* An abort packet: reset the call, passing the error up to the user. */
3613 /* What if error is zero? */
3614 /* What if the error is -1? the application will treat it as a timeout. */
3615 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3616 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3617 rxi_CallError(call, errdata);
3618 MUTEX_EXIT(&call->lock);
3619 putConnection(conn);
3620 return np; /* xmitting; drop packet */
3622 case RX_PACKET_TYPE_BUSY: {
3623 struct clock busyTime;
3625 clock_GetTime(&busyTime);
3627 MUTEX_EXIT(&call->lock);
3629 MUTEX_ENTER(&conn->conn_call_lock);
3630 MUTEX_ENTER(&call->lock);
3631 conn->lastBusy[call->channel] = busyTime.sec;
3632 call->flags |= RX_CALL_PEER_BUSY;
3633 MUTEX_EXIT(&call->lock);
3634 MUTEX_EXIT(&conn->conn_call_lock);
3636 putConnection(conn);
3640 case RX_PACKET_TYPE_ACKALL:
3641 /* All packets acknowledged, so we can drop all packets previously
3642 * readied for sending */
3643 #ifdef RX_ENABLE_LOCKS
3644 /* XXX Hack. We because we can't release the call lock when
3645 * sending packets (osi_NetSend) we drop all ack pkts while we're
3646 * traversing the tq in rxi_Start sending packets out because
3647 * packets may move to the freePacketQueue as result of being
3648 * here! So we drop these packets until we're safely out of the
3649 * traversing. Really ugly!
3650 * For fine grain RX locking, we set the acked field in the packets
3651 * and let rxi_Start remove the packets from the transmit queue.
3653 if (call->flags & RX_CALL_TQ_BUSY) {
3654 rxi_SetAcksInTransmitQueue(call);
3657 #endif /* RX_ENABLE_LOCKS */
3658 rxi_ClearTransmitQueue(call, 0);
3661 /* Should not reach here, unless the peer is broken: send an abort
3663 rxi_CallError(call, RX_PROTOCOL_ERROR);
3664 np = rxi_SendCallAbort(call, np, 1, 0);
3667 /* Note when this last legitimate packet was received, for keep-alive
3668 * processing. Note, we delay getting the time until now in the hope that
3669 * the packet will be delivered to the user before any get time is required
3670 * (if not, then the time won't actually be re-evaluated here). */
3671 call->lastReceiveTime = clock_Sec();
3672 /* we've received a legit packet, so the channel is not busy */
3673 call->flags &= ~RX_CALL_PEER_BUSY;
3674 MUTEX_EXIT(&call->lock);
3675 putConnection(conn);
3679 /* return true if this is an "interesting" connection from the point of view
3680 of someone trying to debug the system */
3682 rxi_IsConnInteresting(struct rx_connection *aconn)
3685 struct rx_call *tcall;
3687 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3690 for (i = 0; i < RX_MAXCALLS; i++) {
3691 tcall = aconn->call[i];
3693 if ((tcall->state == RX_STATE_PRECALL)
3694 || (tcall->state == RX_STATE_ACTIVE))
3696 if ((tcall->app.mode == RX_MODE_SENDING)
3697 || (tcall->app.mode == RX_MODE_RECEIVING))
3705 /* if this is one of the last few packets AND it wouldn't be used by the
3706 receiving call to immediately satisfy a read request, then drop it on
3707 the floor, since accepting it might prevent a lock-holding thread from
3708 making progress in its reading. If a call has been cleared while in
3709 the precall state then ignore all subsequent packets until the call
3710 is assigned to a thread. */
3713 TooLow(struct rx_packet *ap, struct rx_call *acall)
3717 MUTEX_ENTER(&rx_quota_mutex);
3718 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3719 && (acall->state == RX_STATE_PRECALL))
3720 || ((rx_nFreePackets < rxi_dataQuota + 2)
3721 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3722 && (acall->flags & RX_CALL_READER_WAIT)))) {
3725 MUTEX_EXIT(&rx_quota_mutex);
3731 * Clear the attach wait flag on a connection and proceed.
3733 * Any processing waiting for a connection to be attached should be
3734 * unblocked. We clear the flag and do any other needed tasks.
3737 * the conn to unmark waiting for attach
3739 * @pre conn's conn_data_lock must be locked before calling this function
3743 rxi_ConnClearAttachWait(struct rx_connection *conn)
3745 /* Indicate that rxi_CheckReachEvent is no longer running by
3746 * clearing the flag. Must be atomic under conn_data_lock to
3747 * avoid a new call slipping by: rxi_CheckConnReach holds
3748 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3750 conn->flags &= ~RX_CONN_ATTACHWAIT;
3751 if (conn->flags & RX_CONN_NAT_PING) {
3752 conn->flags &= ~RX_CONN_NAT_PING;
3753 rxi_ScheduleNatKeepAliveEvent(conn);
3758 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3760 struct rx_connection *conn = arg1;
3761 struct rx_call *acall = arg2;
3762 struct rx_call *call = acall;
3763 struct clock when, now;
3766 MUTEX_ENTER(&conn->conn_data_lock);
3769 rxevent_Put(conn->checkReachEvent);
3770 conn->checkReachEvent = NULL;
3773 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3775 putConnection(conn);
3777 MUTEX_EXIT(&conn->conn_data_lock);
3781 MUTEX_ENTER(&conn->conn_call_lock);
3782 MUTEX_ENTER(&conn->conn_data_lock);
3783 for (i = 0; i < RX_MAXCALLS; i++) {
3784 struct rx_call *tc = conn->call[i];
3785 if (tc && tc->state == RX_STATE_PRECALL) {
3791 rxi_ConnClearAttachWait(conn);
3792 MUTEX_EXIT(&conn->conn_data_lock);
3793 MUTEX_EXIT(&conn->conn_call_lock);
3798 MUTEX_ENTER(&call->lock);
3799 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3801 MUTEX_EXIT(&call->lock);
3803 clock_GetTime(&now);
3805 when.sec += RX_CHECKREACH_TIMEOUT;
3806 MUTEX_ENTER(&conn->conn_data_lock);
3807 if (!conn->checkReachEvent) {
3808 MUTEX_ENTER(&rx_refcnt_mutex);
3810 MUTEX_EXIT(&rx_refcnt_mutex);
3811 conn->checkReachEvent = rxevent_Post(&when, &now,
3812 rxi_CheckReachEvent, conn,
3815 MUTEX_EXIT(&conn->conn_data_lock);
3821 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3823 struct rx_service *service = conn->service;
3824 struct rx_peer *peer = conn->peer;
3825 afs_uint32 now, lastReach;
3827 if (service->checkReach == 0)
3831 MUTEX_ENTER(&peer->peer_lock);
3832 lastReach = peer->lastReachTime;
3833 MUTEX_EXIT(&peer->peer_lock);
3834 if (now - lastReach < RX_CHECKREACH_TTL)
3837 MUTEX_ENTER(&conn->conn_data_lock);
3838 if (conn->flags & RX_CONN_ATTACHWAIT) {
3839 MUTEX_EXIT(&conn->conn_data_lock);
3842 conn->flags |= RX_CONN_ATTACHWAIT;
3843 MUTEX_EXIT(&conn->conn_data_lock);
3844 if (!conn->checkReachEvent)
3845 rxi_CheckReachEvent(NULL, conn, call, 0);
3850 /* try to attach call, if authentication is complete */
3852 TryAttach(struct rx_call *acall, osi_socket socket,
3853 int *tnop, struct rx_call **newcallp,
3856 struct rx_connection *conn = acall->conn;
3858 if (conn->type == RX_SERVER_CONNECTION
3859 && acall->state == RX_STATE_PRECALL) {
3860 /* Don't attach until we have any req'd. authentication. */
3861 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3862 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3863 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3864 /* Note: this does not necessarily succeed; there
3865 * may not any proc available
3868 rxi_ChallengeOn(acall->conn);
3873 /* A data packet has been received off the interface. This packet is
3874 * appropriate to the call (the call is in the right state, etc.). This
3875 * routine can return a packet to the caller, for re-use */
3877 static struct rx_packet *
3878 rxi_ReceiveDataPacket(struct rx_call *call,
3879 struct rx_packet *np, int istack,
3880 osi_socket socket, afs_uint32 host, u_short port,
3881 int *tnop, struct rx_call **newcallp)
3883 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3888 afs_uint32 serial=0, flags=0;
3890 struct rx_packet *tnp;
3891 if (rx_stats_active)
3892 rx_atomic_inc(&rx_stats.dataPacketsRead);
3895 /* If there are no packet buffers, drop this new packet, unless we can find
3896 * packet buffers from inactive calls */
3898 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3899 MUTEX_ENTER(&rx_freePktQ_lock);
3900 rxi_NeedMorePackets = TRUE;
3901 MUTEX_EXIT(&rx_freePktQ_lock);
3902 if (rx_stats_active)
3903 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3904 rxi_calltrace(RX_TRACE_DROP, call);
3905 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3906 /* We used to clear the receive queue here, in an attempt to free
3907 * packets. However this is unsafe if the queue has received a
3908 * soft ACK for the final packet */
3909 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3915 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3916 * packet is one of several packets transmitted as a single
3917 * datagram. Do not send any soft or hard acks until all packets
3918 * in a jumbogram have been processed. Send negative acks right away.
3920 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3921 /* tnp is non-null when there are more packets in the
3922 * current jumbo gram */
3929 seq = np->header.seq;
3930 serial = np->header.serial;
3931 flags = np->header.flags;
3933 /* If the call is in an error state, send an abort message */
3935 return rxi_SendCallAbort(call, np, istack, 0);
3937 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3938 * AFS 3.5 jumbogram. */
3939 if (flags & RX_JUMBO_PACKET) {
3940 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3945 if (np->header.spare != 0) {
3946 MUTEX_ENTER(&call->conn->conn_data_lock);
3947 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3948 MUTEX_EXIT(&call->conn->conn_data_lock);
3951 /* The usual case is that this is the expected next packet */
3952 if (seq == call->rnext) {
3954 /* Check to make sure it is not a duplicate of one already queued */
3955 if (!opr_queue_IsEmpty(&call->rq)
3956 && opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq == seq) {
3957 if (rx_stats_active)
3958 rx_atomic_inc(&rx_stats.dupPacketsRead);
3959 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3960 rxevent_Cancel(&call->delayedAckEvent, call,
3961 RX_CALL_REFCOUNT_DELAY);
3962 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3968 /* It's the next packet. Stick it on the receive queue
3969 * for this call. Set newPackets to make sure we wake
3970 * the reader once all packets have been processed */
3971 #ifdef RX_TRACK_PACKETS
3972 np->flags |= RX_PKTFLAG_RQ;
3974 opr_queue_Prepend(&call->rq, &np->entry);
3975 #ifdef RXDEBUG_PACKET
3977 #endif /* RXDEBUG_PACKET */
3979 np = NULL; /* We can't use this anymore */
3982 /* If an ack is requested then set a flag to make sure we
3983 * send an acknowledgement for this packet */
3984 if (flags & RX_REQUEST_ACK) {
3985 ackNeeded = RX_ACK_REQUESTED;
3988 /* Keep track of whether we have received the last packet */
3989 if (flags & RX_LAST_PACKET) {
3990 call->flags |= RX_CALL_HAVE_LAST;
3994 /* Check whether we have all of the packets for this call */
3995 if (call->flags & RX_CALL_HAVE_LAST) {
3996 afs_uint32 tseq; /* temporary sequence number */
3997 struct opr_queue *cursor;
3999 for (tseq = seq, opr_queue_Scan(&call->rq, cursor)) {
4000 struct rx_packet *tp;
4002 tp = opr_queue_Entry(cursor, struct rx_packet, entry);
4003 if (tseq != tp->header.seq)
4005 if (tp->header.flags & RX_LAST_PACKET) {
4006 call->flags |= RX_CALL_RECEIVE_DONE;
4013 /* Provide asynchronous notification for those who want it
4014 * (e.g. multi rx) */
4015 if (call->arrivalProc) {
4016 (*call->arrivalProc) (call, call->arrivalProcHandle,
4017 call->arrivalProcArg);
4018 call->arrivalProc = (void (*)())0;
4021 /* Update last packet received */
4024 /* If there is no server process serving this call, grab
4025 * one, if available. We only need to do this once. If a
4026 * server thread is available, this thread becomes a server
4027 * thread and the server thread becomes a listener thread. */
4029 TryAttach(call, socket, tnop, newcallp, 0);
4032 /* This is not the expected next packet. */
4034 /* Determine whether this is a new or old packet, and if it's
4035 * a new one, whether it fits into the current receive window.
4036 * Also figure out whether the packet was delivered in sequence.
4037 * We use the prev variable to determine whether the new packet
4038 * is the successor of its immediate predecessor in the
4039 * receive queue, and the missing flag to determine whether
4040 * any of this packets predecessors are missing. */
4042 afs_uint32 prev; /* "Previous packet" sequence number */
4043 struct opr_queue *cursor;
4044 int missing; /* Are any predecessors missing? */
4046 /* If the new packet's sequence number has been sent to the
4047 * application already, then this is a duplicate */
4048 if (seq < call->rnext) {
4049 if (rx_stats_active)
4050 rx_atomic_inc(&rx_stats.dupPacketsRead);
4051 rxevent_Cancel(&call->delayedAckEvent, call,
4052 RX_CALL_REFCOUNT_DELAY);
4053 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
4059 /* If the sequence number is greater than what can be
4060 * accomodated by the current window, then send a negative
4061 * acknowledge and drop the packet */
4062 if ((call->rnext + call->rwind) <= seq) {
4063 rxevent_Cancel(&call->delayedAckEvent, call,
4064 RX_CALL_REFCOUNT_DELAY);
4065 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
4072 /* Look for the packet in the queue of old received packets */
4073 prev = call->rnext - 1;
4075 for (opr_queue_Scan(&call->rq, cursor)) {
4076 struct rx_packet *tp
4077 = opr_queue_Entry(cursor, struct rx_packet, entry);
4079 /*Check for duplicate packet */
4080 if (seq == tp->header.seq) {
4081 if (rx_stats_active)
4082 rx_atomic_inc(&rx_stats.dupPacketsRead);
4083 rxevent_Cancel(&call->delayedAckEvent, call,
4084 RX_CALL_REFCOUNT_DELAY);
4085 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
4091 /* If we find a higher sequence packet, break out and
4092 * insert the new packet here. */
4093 if (seq < tp->header.seq)
4095 /* Check for missing packet */
4096 if (tp->header.seq != prev + 1) {
4100 prev = tp->header.seq;
4103 /* Keep track of whether we have received the last packet. */
4104 if (flags & RX_LAST_PACKET) {
4105 call->flags |= RX_CALL_HAVE_LAST;
4108 /* It's within the window: add it to the the receive queue.
4109 * tp is left by the previous loop either pointing at the
4110 * packet before which to insert the new packet, or at the
4111 * queue head if the queue is empty or the packet should be
4113 #ifdef RX_TRACK_PACKETS
4114 np->flags |= RX_PKTFLAG_RQ;
4116 #ifdef RXDEBUG_PACKET
4118 #endif /* RXDEBUG_PACKET */
4119 opr_queue_InsertBefore(cursor, &np->entry);
4123 /* Check whether we have all of the packets for this call */
4124 if ((call->flags & RX_CALL_HAVE_LAST)
4125 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4126 afs_uint32 tseq; /* temporary sequence number */
4129 for (opr_queue_Scan(&call->rq, cursor)) {
4130 struct rx_packet *tp
4131 = opr_queue_Entry(cursor, struct rx_packet, entry);
4132 if (tseq != tp->header.seq)
4134 if (tp->header.flags & RX_LAST_PACKET) {
4135 call->flags |= RX_CALL_RECEIVE_DONE;
4142 /* We need to send an ack of the packet is out of sequence,
4143 * or if an ack was requested by the peer. */
4144 if (seq != prev + 1 || missing) {
4145 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4146 } else if (flags & RX_REQUEST_ACK) {
4147 ackNeeded = RX_ACK_REQUESTED;
4150 /* Acknowledge the last packet for each call */
4151 if (flags & RX_LAST_PACKET) {
4162 * If the receiver is waiting for an iovec, fill the iovec
4163 * using the data from the receive queue */
4164 if (call->flags & RX_CALL_IOVEC_WAIT) {
4165 didHardAck = rxi_FillReadVec(call, serial);
4166 /* the call may have been aborted */
4175 /* Wakeup the reader if any */
4176 if ((call->flags & RX_CALL_READER_WAIT)
4177 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4178 || (call->iovNext >= call->iovMax)
4179 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4180 call->flags &= ~RX_CALL_READER_WAIT;
4181 #ifdef RX_ENABLE_LOCKS
4182 CV_BROADCAST(&call->cv_rq);
4184 osi_rxWakeup(&call->rq);
4190 * Send an ack when requested by the peer, or once every
4191 * rxi_SoftAckRate packets until the last packet has been
4192 * received. Always send a soft ack for the last packet in
4193 * the server's reply. */
4195 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4196 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4197 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4198 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4199 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4200 } else if (call->nSoftAcks) {
4201 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4202 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4204 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4205 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4206 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4213 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4215 struct rx_peer *peer = conn->peer;
4217 MUTEX_ENTER(&peer->peer_lock);
4218 peer->lastReachTime = clock_Sec();
4219 MUTEX_EXIT(&peer->peer_lock);
4221 MUTEX_ENTER(&conn->conn_data_lock);
4222 if (conn->flags & RX_CONN_ATTACHWAIT) {
4225 rxi_ConnClearAttachWait(conn);
4226 MUTEX_EXIT(&conn->conn_data_lock);
4228 for (i = 0; i < RX_MAXCALLS; i++) {
4229 struct rx_call *call = conn->call[i];
4232 MUTEX_ENTER(&call->lock);
4233 /* tnop can be null if newcallp is null */
4234 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4236 MUTEX_EXIT(&call->lock);
4240 MUTEX_EXIT(&conn->conn_data_lock);
4243 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4245 rx_ack_reason(int reason)
4248 case RX_ACK_REQUESTED:
4250 case RX_ACK_DUPLICATE:
4252 case RX_ACK_OUT_OF_SEQUENCE:
4254 case RX_ACK_EXCEEDS_WINDOW:
4256 case RX_ACK_NOSPACE:
4260 case RX_ACK_PING_RESPONSE:
4273 /* The real smarts of the whole thing. */
4274 static struct rx_packet *
4275 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4278 struct rx_ackPacket *ap;
4280 struct rx_packet *tp;
4281 struct rx_connection *conn = call->conn;
4282 struct rx_peer *peer = conn->peer;
4283 struct opr_queue *cursor;
4284 struct clock now; /* Current time, for RTT calculations */
4292 int newAckCount = 0;
4293 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4294 int pktsize = 0; /* Set if we need to update the peer mtu */
4295 int conn_data_locked = 0;
4297 if (rx_stats_active)
4298 rx_atomic_inc(&rx_stats.ackPacketsRead);
4299 ap = (struct rx_ackPacket *)rx_DataOf(np);
4300 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4302 return np; /* truncated ack packet */
4304 /* depends on ack packet struct */
4305 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4306 first = ntohl(ap->firstPacket);
4307 prev = ntohl(ap->previousPacket);
4308 serial = ntohl(ap->serial);
4311 * Ignore ack packets received out of order while protecting
4312 * against peers that set the previousPacket field to a packet
4313 * serial number instead of a sequence number.
4315 if (first < call->tfirst ||
4316 (first == call->tfirst && prev < call->tprev && prev < call->tfirst
4323 if (np->header.flags & RX_SLOW_START_OK) {
4324 call->flags |= RX_CALL_SLOW_START_OK;
4327 if (ap->reason == RX_ACK_PING_RESPONSE)
4328 rxi_UpdatePeerReach(conn, call);
4330 if (conn->lastPacketSizeSeq) {
4331 MUTEX_ENTER(&conn->conn_data_lock);
4332 conn_data_locked = 1;
4333 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4334 pktsize = conn->lastPacketSize;
4335 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4338 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4339 if (!conn_data_locked) {
4340 MUTEX_ENTER(&conn->conn_data_lock);
4341 conn_data_locked = 1;
4343 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4344 /* process mtu ping ack */
4345 pktsize = conn->lastPingSize;
4346 conn->lastPingSizeSer = conn->lastPingSize = 0;
4350 if (conn_data_locked) {
4351 MUTEX_EXIT(&conn->conn_data_lock);
4352 conn_data_locked = 0;
4356 if (rxdebug_active) {
4360 len = _snprintf(msg, sizeof(msg),
4361 "tid[%d] RACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4362 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4363 ntohl(ap->serial), ntohl(ap->previousPacket),
4364 (unsigned int)np->header.seq, ntohl(ap->firstPacket),
4365 ap->nAcks, ntohs(ap->bufferSpace) );
4369 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4370 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4374 OutputDebugString(msg);
4376 #else /* AFS_NT40_ENV */
4379 "RACK: reason %x previous %u seq %u serial %u first %u",
4380 ap->reason, ntohl(ap->previousPacket),
4381 (unsigned int)np->header.seq, (unsigned int)serial,
4382 ntohl(ap->firstPacket));
4385 for (offset = 0; offset < nAcks; offset++)
4386 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4391 #endif /* AFS_NT40_ENV */
4394 MUTEX_ENTER(&peer->peer_lock);
4397 * Start somewhere. Can't assume we can send what we can receive,
4398 * but we are clearly receiving.
4400 if (!peer->maxPacketSize)
4401 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4403 if (pktsize > peer->maxPacketSize) {
4404 peer->maxPacketSize = pktsize;
4405 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4406 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4407 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4408 rxi_ScheduleGrowMTUEvent(call, 1);
4413 clock_GetTime(&now);
4415 /* The transmit queue splits into 4 sections.
4417 * The first section is packets which have now been acknowledged
4418 * by a window size change in the ack. These have reached the
4419 * application layer, and may be discarded. These are packets
4420 * with sequence numbers < ap->firstPacket.
4422 * The second section is packets which have sequence numbers in
4423 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4424 * contents of the packet's ack array determines whether these
4425 * packets are acknowledged or not.
4427 * The third section is packets which fall above the range
4428 * addressed in the ack packet. These have not yet been received
4431 * The four section is packets which have not yet been transmitted.
4432 * These packets will have a header.serial of 0.
4435 /* First section - implicitly acknowledged packets that can be
4439 tp = opr_queue_First(&call->tq, struct rx_packet, entry);
4440 while(!opr_queue_IsEnd(&call->tq, &tp->entry) && tp->header.seq < first) {
4441 struct rx_packet *next;
4443 next = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4444 call->tfirst = tp->header.seq + 1;
4446 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4448 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4451 #ifdef RX_ENABLE_LOCKS
4452 /* XXX Hack. Because we have to release the global call lock when sending
4453 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4454 * in rxi_Start sending packets out because packets may move to the
4455 * freePacketQueue as result of being here! So we drop these packets until
4456 * we're safely out of the traversing. Really ugly!
4457 * To make it even uglier, if we're using fine grain locking, we can
4458 * set the ack bits in the packets and have rxi_Start remove the packets
4459 * when it's done transmitting.
4461 if (call->flags & RX_CALL_TQ_BUSY) {
4462 tp->flags |= RX_PKTFLAG_ACKED;
4463 call->flags |= RX_CALL_TQ_SOME_ACKED;
4465 #endif /* RX_ENABLE_LOCKS */
4467 opr_queue_Remove(&tp->entry);
4468 #ifdef RX_TRACK_PACKETS
4469 tp->flags &= ~RX_PKTFLAG_TQ;
4471 #ifdef RXDEBUG_PACKET
4473 #endif /* RXDEBUG_PACKET */
4474 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4479 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4481 /* Second section of the queue - packets for which we are receiving
4484 * Go through the explicit acks/nacks and record the results in
4485 * the waiting packets. These are packets that can't be released
4486 * yet, even with a positive acknowledge. This positive
4487 * acknowledge only means the packet has been received by the
4488 * peer, not that it will be retained long enough to be sent to
4489 * the peer's upper level. In addition, reset the transmit timers
4490 * of any missing packets (those packets that must be missing
4491 * because this packet was out of sequence) */
4493 call->nSoftAcked = 0;
4495 while (!opr_queue_IsEnd(&call->tq, &tp->entry)
4496 && tp->header.seq < first + nAcks) {
4497 /* Set the acknowledge flag per packet based on the
4498 * information in the ack packet. An acknowlegded packet can
4499 * be downgraded when the server has discarded a packet it
4500 * soacked previously, or when an ack packet is received
4501 * out of sequence. */
4502 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4503 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4505 tp->flags |= RX_PKTFLAG_ACKED;
4506 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4513 } else /* RX_ACK_TYPE_NACK */ {
4514 tp->flags &= ~RX_PKTFLAG_ACKED;
4518 tp = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4521 /* We don't need to take any action with the 3rd or 4th section in the
4522 * queue - they're not addressed by the contents of this ACK packet.
4525 /* If the window has been extended by this acknowledge packet,
4526 * then wakeup a sender waiting in alloc for window space, or try
4527 * sending packets now, if he's been sitting on packets due to
4528 * lack of window space */
4529 if (call->tnext < (call->tfirst + call->twind)) {
4530 #ifdef RX_ENABLE_LOCKS
4531 CV_SIGNAL(&call->cv_twind);
4533 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4534 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4535 osi_rxWakeup(&call->twind);
4538 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4539 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4543 /* if the ack packet has a receivelen field hanging off it,
4544 * update our state */
4545 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4548 /* If the ack packet has a "recommended" size that is less than
4549 * what I am using now, reduce my size to match */
4550 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4551 (int)sizeof(afs_int32), &tSize);
4552 tSize = (afs_uint32) ntohl(tSize);
4553 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4555 /* Get the maximum packet size to send to this peer */
4556 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4558 tSize = (afs_uint32) ntohl(tSize);
4559 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4560 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4562 /* sanity check - peer might have restarted with different params.
4563 * If peer says "send less", dammit, send less... Peer should never
4564 * be unable to accept packets of the size that prior AFS versions would
4565 * send without asking. */
4566 if (peer->maxMTU != tSize) {
4567 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4569 peer->maxMTU = tSize;
4570 peer->MTU = MIN(tSize, peer->MTU);
4571 call->MTU = MIN(call->MTU, tSize);
4574 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4577 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4578 (int)sizeof(afs_int32), &tSize);
4579 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4580 if (tSize < call->twind) { /* smaller than our send */
4581 call->twind = tSize; /* window, we must send less... */
4582 call->ssthresh = MIN(call->twind, call->ssthresh);
4583 call->conn->twind[call->channel] = call->twind;
4586 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4587 * network MTU confused with the loopback MTU. Calculate the
4588 * maximum MTU here for use in the slow start code below.
4590 /* Did peer restart with older RX version? */
4591 if (peer->maxDgramPackets > 1) {
4592 peer->maxDgramPackets = 1;
4594 } else if (np->length >=
4595 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4598 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4599 sizeof(afs_int32), &tSize);
4600 tSize = (afs_uint32) ntohl(tSize);
4602 * As of AFS 3.5 we set the send window to match the receive window.
4604 if (tSize < call->twind) {
4605 call->twind = tSize;
4606 call->conn->twind[call->channel] = call->twind;
4607 call->ssthresh = MIN(call->twind, call->ssthresh);
4608 } else if (tSize > call->twind) {
4609 call->twind = tSize;
4610 call->conn->twind[call->channel] = call->twind;
4614 * As of AFS 3.5, a jumbogram is more than one fixed size
4615 * packet transmitted in a single UDP datagram. If the remote
4616 * MTU is smaller than our local MTU then never send a datagram
4617 * larger than the natural MTU.
4620 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4621 (int)sizeof(afs_int32), &tSize);
4622 maxDgramPackets = (afs_uint32) ntohl(tSize);
4623 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4625 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4626 if (maxDgramPackets > 1) {
4627 peer->maxDgramPackets = maxDgramPackets;
4628 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4630 peer->maxDgramPackets = 1;
4631 call->MTU = peer->natMTU;
4633 } else if (peer->maxDgramPackets > 1) {
4634 /* Restarted with lower version of RX */
4635 peer->maxDgramPackets = 1;
4637 } else if (peer->maxDgramPackets > 1
4638 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4639 /* Restarted with lower version of RX */
4640 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4641 peer->natMTU = OLD_MAX_PACKET_SIZE;
4642 peer->MTU = OLD_MAX_PACKET_SIZE;
4643 peer->maxDgramPackets = 1;
4644 peer->nDgramPackets = 1;
4646 call->MTU = OLD_MAX_PACKET_SIZE;
4651 * Calculate how many datagrams were successfully received after
4652 * the first missing packet and adjust the negative ack counter
4657 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4658 if (call->nNacks < nNacked) {
4659 call->nNacks = nNacked;
4662 call->nAcks += newAckCount;
4666 /* If the packet contained new acknowledgements, rather than just
4667 * being a duplicate of one we have previously seen, then we can restart
4670 if (newAckCount > 0)
4671 rxi_rto_packet_acked(call, istack);
4673 if (call->flags & RX_CALL_FAST_RECOVER) {
4674 if (newAckCount == 0) {
4675 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4677 call->flags &= ~RX_CALL_FAST_RECOVER;
4678 call->cwind = call->nextCwind;
4679 call->nextCwind = 0;
4682 call->nCwindAcks = 0;
4683 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4684 /* Three negative acks in a row trigger congestion recovery */
4685 call->flags |= RX_CALL_FAST_RECOVER;
4686 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4688 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4689 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4690 call->nextCwind = call->ssthresh;
4693 peer->MTU = call->MTU;
4694 peer->cwind = call->nextCwind;
4695 peer->nDgramPackets = call->nDgramPackets;
4697 call->congestSeq = peer->congestSeq;
4699 /* Reset the resend times on the packets that were nacked
4700 * so we will retransmit as soon as the window permits
4704 for (opr_queue_ScanBackwards(&call->tq, cursor)) {
4705 struct rx_packet *tp =
4706 opr_queue_Entry(cursor, struct rx_packet, entry);
4708 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4709 tp->flags &= ~RX_PKTFLAG_SENT;
4711 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4716 /* If cwind is smaller than ssthresh, then increase
4717 * the window one packet for each ack we receive (exponential
4719 * If cwind is greater than or equal to ssthresh then increase
4720 * the congestion window by one packet for each cwind acks we
4721 * receive (linear growth). */
4722 if (call->cwind < call->ssthresh) {
4724 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4725 call->nCwindAcks = 0;
4727 call->nCwindAcks += newAckCount;
4728 if (call->nCwindAcks >= call->cwind) {
4729 call->nCwindAcks = 0;
4730 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4734 * If we have received several acknowledgements in a row then
4735 * it is time to increase the size of our datagrams
4737 if ((int)call->nAcks > rx_nDgramThreshold) {
4738 if (peer->maxDgramPackets > 1) {
4739 if (call->nDgramPackets < peer->maxDgramPackets) {
4740 call->nDgramPackets++;
4742 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4743 } else if (call->MTU < peer->maxMTU) {
4744 /* don't upgrade if we can't handle it */
4745 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4746 call->MTU = peer->ifMTU;
4748 call->MTU += peer->natMTU;
4749 call->MTU = MIN(call->MTU, peer->maxMTU);
4756 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4758 /* Servers need to hold the call until all response packets have
4759 * been acknowledged. Soft acks are good enough since clients
4760 * are not allowed to clear their receive queues. */
4761 if (call->state == RX_STATE_HOLD
4762 && call->tfirst + call->nSoftAcked >= call->tnext) {
4763 call->state = RX_STATE_DALLY;
4764 rxi_ClearTransmitQueue(call, 0);
4765 rxevent_Cancel(&call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4766 } else if (!opr_queue_IsEmpty(&call->tq)) {
4767 rxi_Start(call, istack);
4772 /* Received a response to a challenge packet */
4773 static struct rx_packet *
4774 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4775 struct rx_packet *np, int istack)
4779 /* Ignore the packet if we're the client */
4780 if (conn->type == RX_CLIENT_CONNECTION)
4783 /* If already authenticated, ignore the packet (it's probably a retry) */
4784 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4787 /* Otherwise, have the security object evaluate the response packet */
4788 error = RXS_CheckResponse(conn->securityObject, conn, np);
4790 /* If the response is invalid, reset the connection, sending
4791 * an abort to the peer */
4795 rxi_ConnectionError(conn, error);
4796 MUTEX_ENTER(&conn->conn_data_lock);
4797 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4798 MUTEX_EXIT(&conn->conn_data_lock);
4801 /* If the response is valid, any calls waiting to attach
4802 * servers can now do so */
4805 for (i = 0; i < RX_MAXCALLS; i++) {
4806 struct rx_call *call = conn->call[i];
4808 MUTEX_ENTER(&call->lock);
4809 if (call->state == RX_STATE_PRECALL)
4810 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4811 /* tnop can be null if newcallp is null */
4812 MUTEX_EXIT(&call->lock);
4816 /* Update the peer reachability information, just in case
4817 * some calls went into attach-wait while we were waiting
4818 * for authentication..
4820 rxi_UpdatePeerReach(conn, NULL);
4825 /* A client has received an authentication challenge: the security
4826 * object is asked to cough up a respectable response packet to send
4827 * back to the server. The server is responsible for retrying the
4828 * challenge if it fails to get a response. */
4830 static struct rx_packet *
4831 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4832 struct rx_packet *np, int istack)
4836 /* Ignore the challenge if we're the server */
4837 if (conn->type == RX_SERVER_CONNECTION)
4840 /* Ignore the challenge if the connection is otherwise idle; someone's
4841 * trying to use us as an oracle. */
4842 if (!rxi_HasActiveCalls(conn))
4845 /* Send the security object the challenge packet. It is expected to fill
4846 * in the response. */
4847 error = RXS_GetResponse(conn->securityObject, conn, np);
4849 /* If the security object is unable to return a valid response, reset the
4850 * connection and send an abort to the peer. Otherwise send the response
4851 * packet to the peer connection. */
4853 rxi_ConnectionError(conn, error);
4854 MUTEX_ENTER(&conn->conn_data_lock);
4855 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4856 MUTEX_EXIT(&conn->conn_data_lock);
4858 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4859 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4865 /* Find an available server process to service the current request in
4866 * the given call structure. If one isn't available, queue up this
4867 * call so it eventually gets one */
4869 rxi_AttachServerProc(struct rx_call *call,
4870 osi_socket socket, int *tnop,
4871 struct rx_call **newcallp)
4873 struct rx_serverQueueEntry *sq;
4874 struct rx_service *service = call->conn->service;
4877 /* May already be attached */
4878 if (call->state == RX_STATE_ACTIVE)
4881 MUTEX_ENTER(&rx_serverPool_lock);
4883 haveQuota = QuotaOK(service);
4884 if ((!haveQuota) || opr_queue_IsEmpty(&rx_idleServerQueue)) {
4885 /* If there are no processes available to service this call,
4886 * put the call on the incoming call queue (unless it's
4887 * already on the queue).
4889 #ifdef RX_ENABLE_LOCKS
4891 ReturnToServerPool(service);
4892 #endif /* RX_ENABLE_LOCKS */
4894 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4895 call->flags |= RX_CALL_WAIT_PROC;
4896 rx_atomic_inc(&rx_nWaiting);
4897 rx_atomic_inc(&rx_nWaited);
4898 rxi_calltrace(RX_CALL_ARRIVAL, call);
4899 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4900 opr_queue_Append(&rx_incomingCallQueue, &call->entry);
4903 sq = opr_queue_Last(&rx_idleServerQueue,
4904 struct rx_serverQueueEntry, entry);
4906 /* If hot threads are enabled, and both newcallp and sq->socketp
4907 * are non-null, then this thread will process the call, and the
4908 * idle server thread will start listening on this threads socket.
4910 opr_queue_Remove(&sq->entry);
4912 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4915 *sq->socketp = socket;
4916 clock_GetTime(&call->startTime);
4917 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4921 if (call->flags & RX_CALL_WAIT_PROC) {
4922 /* Conservative: I don't think this should happen */
4923 call->flags &= ~RX_CALL_WAIT_PROC;
4924 rx_atomic_dec(&rx_nWaiting);
4925 if (opr_queue_IsOnQueue(&call->entry)) {
4926 opr_queue_Remove(&call->entry);
4929 call->state = RX_STATE_ACTIVE;
4930 call->app.mode = RX_MODE_RECEIVING;
4931 #ifdef RX_KERNEL_TRACE
4933 int glockOwner = ISAFS_GLOCK();
4936 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4937 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4943 if (call->flags & RX_CALL_CLEARED) {
4944 /* send an ack now to start the packet flow up again */
4945 call->flags &= ~RX_CALL_CLEARED;
4946 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4948 #ifdef RX_ENABLE_LOCKS
4951 service->nRequestsRunning++;
4952 MUTEX_ENTER(&rx_quota_mutex);
4953 if (service->nRequestsRunning <= service->minProcs)
4956 MUTEX_EXIT(&rx_quota_mutex);
4960 MUTEX_EXIT(&rx_serverPool_lock);
4963 /* Delay the sending of an acknowledge event for a short while, while
4964 * a new call is being prepared (in the case of a client) or a reply
4965 * is being prepared (in the case of a server). Rather than sending
4966 * an ack packet, an ACKALL packet is sent. */
4968 rxi_AckAll(struct rx_call *call)
4970 rxi_SendSpecial(call, call->conn, NULL, RX_PACKET_TYPE_ACKALL,
4972 call->flags |= RX_CALL_ACKALL_SENT;
4976 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4979 struct rx_call *call = arg1;
4980 #ifdef RX_ENABLE_LOCKS
4982 MUTEX_ENTER(&call->lock);
4983 if (event == call->delayedAckEvent) {
4984 rxevent_Put(call->delayedAckEvent);
4985 call->delayedAckEvent = NULL;
4987 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4989 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4991 MUTEX_EXIT(&call->lock);
4992 #else /* RX_ENABLE_LOCKS */
4994 rxevent_Put(call->delayedAckEvent);
4995 call->delayedAckEvent = NULL;
4997 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4998 #endif /* RX_ENABLE_LOCKS */
5002 #ifdef RX_ENABLE_LOCKS
5003 /* Set ack in all packets in transmit queue. rxi_Start will deal with
5004 * clearing them out.
5007 rxi_SetAcksInTransmitQueue(struct rx_call *call)
5009 struct opr_queue *cursor;
5012 for (opr_queue_Scan(&call->tq, cursor)) {
5014 = opr_queue_Entry(cursor, struct rx_packet, entry);
5016 p->flags |= RX_PKTFLAG_ACKED;
5021 call->flags |= RX_CALL_TQ_CLEARME;
5022 call->flags |= RX_CALL_TQ_SOME_ACKED;
5025 rxi_rto_cancel(call);
5027 call->tfirst = call->tnext;
5028 call->nSoftAcked = 0;
5030 if (call->flags & RX_CALL_FAST_RECOVER) {
5031 call->flags &= ~RX_CALL_FAST_RECOVER;
5032 call->cwind = call->nextCwind;
5033 call->nextCwind = 0;
5036 CV_SIGNAL(&call->cv_twind);
5038 #endif /* RX_ENABLE_LOCKS */
5040 /* Clear out the transmit queue for the current call (all packets have
5041 * been received by peer) */
5043 rxi_ClearTransmitQueue(struct rx_call *call, int force)
5045 #ifdef RX_ENABLE_LOCKS
5046 struct opr_queue *cursor;
5047 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
5049 for (opr_queue_Scan(&call->tq, cursor)) {
5051 = opr_queue_Entry(cursor, struct rx_packet, entry);
5053 p->flags |= RX_PKTFLAG_ACKED;
5057 call->flags |= RX_CALL_TQ_CLEARME;
5058 call->flags |= RX_CALL_TQ_SOME_ACKED;
5061 #endif /* RX_ENABLE_LOCKS */
5062 #ifdef RXDEBUG_PACKET
5064 #endif /* RXDEBUG_PACKET */
5065 rxi_FreePackets(0, &call->tq);
5066 rxi_WakeUpTransmitQueue(call);
5067 #ifdef RX_ENABLE_LOCKS
5068 call->flags &= ~RX_CALL_TQ_CLEARME;
5072 rxi_rto_cancel(call);
5073 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5074 call->nSoftAcked = 0;
5076 if (call->flags & RX_CALL_FAST_RECOVER) {
5077 call->flags &= ~RX_CALL_FAST_RECOVER;
5078 call->cwind = call->nextCwind;
5080 #ifdef RX_ENABLE_LOCKS
5081 CV_SIGNAL(&call->cv_twind);
5083 osi_rxWakeup(&call->twind);
5088 rxi_ClearReceiveQueue(struct rx_call *call)
5090 if (!opr_queue_IsEmpty(&call->rq)) {
5093 count = rxi_FreePackets(0, &call->rq);
5094 rx_packetReclaims += count;
5095 #ifdef RXDEBUG_PACKET
5097 if ( call->rqc != 0 )
5098 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5100 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5102 if (call->state == RX_STATE_PRECALL) {
5103 call->flags |= RX_CALL_CLEARED;
5107 /* Send an abort packet for the specified call */
5108 static struct rx_packet *
5109 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5110 int istack, int force)
5112 afs_int32 error, cerror;
5113 struct clock when, now;
5118 switch (call->error) {
5121 cerror = RX_CALL_TIMEOUT;
5124 cerror = call->error;
5127 /* Clients should never delay abort messages */
5128 if (rx_IsClientConn(call->conn))
5131 if (call->abortCode != cerror) {
5132 call->abortCode = cerror;
5133 call->abortCount = 0;
5136 if (force || rxi_callAbortThreshhold == 0
5137 || call->abortCount < rxi_callAbortThreshhold) {
5138 if (call->delayedAbortEvent) {
5139 rxevent_Cancel(&call->delayedAbortEvent, call,
5140 RX_CALL_REFCOUNT_ABORT);
5142 error = htonl(cerror);
5145 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5146 (char *)&error, sizeof(error), istack);
5147 } else if (!call->delayedAbortEvent) {
5148 clock_GetTime(&now);
5150 clock_Addmsec(&when, rxi_callAbortDelay);
5151 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5152 call->delayedAbortEvent =
5153 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5158 /* Send an abort packet for the specified connection. Packet is an
5159 * optional pointer to a packet that can be used to send the abort.
5160 * Once the number of abort messages reaches the threshhold, an
5161 * event is scheduled to send the abort. Setting the force flag
5162 * overrides sending delayed abort messages.
5164 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5165 * to send the abort packet.
5168 rxi_SendConnectionAbort(struct rx_connection *conn,
5169 struct rx_packet *packet, int istack, int force)
5172 struct clock when, now;
5177 /* Clients should never delay abort messages */
5178 if (rx_IsClientConn(conn))
5181 if (force || rxi_connAbortThreshhold == 0
5182 || conn->abortCount < rxi_connAbortThreshhold) {
5184 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
5185 error = htonl(conn->error);
5187 MUTEX_EXIT(&conn->conn_data_lock);
5189 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5190 RX_PACKET_TYPE_ABORT, (char *)&error,
5191 sizeof(error), istack);
5192 MUTEX_ENTER(&conn->conn_data_lock);
5193 } else if (!conn->delayedAbortEvent) {
5194 clock_GetTime(&now);
5196 clock_Addmsec(&when, rxi_connAbortDelay);
5197 conn->delayedAbortEvent =
5198 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
5203 /* Associate an error all of the calls owned by a connection. Called
5204 * with error non-zero. This is only for really fatal things, like
5205 * bad authentication responses. The connection itself is set in
5206 * error at this point, so that future packets received will be
5209 rxi_ConnectionError(struct rx_connection *conn,
5215 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5217 MUTEX_ENTER(&conn->conn_data_lock);
5218 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
5219 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
5220 if (conn->checkReachEvent) {
5221 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
5222 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5223 putConnection(conn);
5225 MUTEX_EXIT(&conn->conn_data_lock);
5226 for (i = 0; i < RX_MAXCALLS; i++) {
5227 struct rx_call *call = conn->call[i];
5229 MUTEX_ENTER(&call->lock);
5230 rxi_CallError(call, error);
5231 MUTEX_EXIT(&call->lock);
5234 conn->error = error;
5235 if (rx_stats_active)
5236 rx_atomic_inc(&rx_stats.fatalErrors);
5241 * Interrupt an in-progress call with the specified error and wakeup waiters.
5243 * @param[in] call The call to interrupt
5244 * @param[in] error The error code to send to the peer
5247 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5249 MUTEX_ENTER(&call->lock);
5250 rxi_CallError(call, error);
5251 rxi_SendCallAbort(call, NULL, 0, 1);
5252 MUTEX_EXIT(&call->lock);
5256 rxi_CallError(struct rx_call *call, afs_int32 error)
5259 osirx_AssertMine(&call->lock, "rxi_CallError");
5261 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5263 error = call->error;
5265 #ifdef RX_ENABLE_LOCKS
5266 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5267 rxi_ResetCall(call, 0);
5270 rxi_ResetCall(call, 0);
5272 call->error = error;
5275 /* Reset various fields in a call structure, and wakeup waiting
5276 * processes. Some fields aren't changed: state & mode are not
5277 * touched (these must be set by the caller), and bufptr, nLeft, and
5278 * nFree are not reset, since these fields are manipulated by
5279 * unprotected macros, and may only be reset by non-interrupting code.
5283 rxi_ResetCall(struct rx_call *call, int newcall)
5286 struct rx_peer *peer;
5287 struct rx_packet *packet;
5289 osirx_AssertMine(&call->lock, "rxi_ResetCall");
5291 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5293 /* Notify anyone who is waiting for asynchronous packet arrival */
5294 if (call->arrivalProc) {
5295 (*call->arrivalProc) (call, call->arrivalProcHandle,
5296 call->arrivalProcArg);
5297 call->arrivalProc = (void (*)())0;
5301 rxevent_Cancel(&call->growMTUEvent, call, RX_CALL_REFCOUNT_MTU);
5303 if (call->delayedAbortEvent) {
5304 rxevent_Cancel(&call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
5305 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5307 rxi_SendCallAbort(call, packet, 0, 1);
5308 rxi_FreePacket(packet);
5313 * Update the peer with the congestion information in this call
5314 * so other calls on this connection can pick up where this call
5315 * left off. If the congestion sequence numbers don't match then
5316 * another call experienced a retransmission.
5318 peer = call->conn->peer;
5319 MUTEX_ENTER(&peer->peer_lock);
5321 if (call->congestSeq == peer->congestSeq) {
5322 peer->cwind = MAX(peer->cwind, call->cwind);
5323 peer->MTU = MAX(peer->MTU, call->MTU);
5324 peer->nDgramPackets =
5325 MAX(peer->nDgramPackets, call->nDgramPackets);
5328 call->abortCode = 0;
5329 call->abortCount = 0;
5331 if (peer->maxDgramPackets > 1) {
5332 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5334 call->MTU = peer->MTU;
5336 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5337 call->ssthresh = rx_maxSendWindow;
5338 call->nDgramPackets = peer->nDgramPackets;
5339 call->congestSeq = peer->congestSeq;
5340 call->rtt = peer->rtt;
5341 call->rtt_dev = peer->rtt_dev;
5342 clock_Zero(&call->rto);
5343 clock_Addmsec(&call->rto,
5344 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5345 MUTEX_EXIT(&peer->peer_lock);
5347 flags = call->flags;
5348 rxi_WaitforTQBusy(call);
5350 rxi_ClearTransmitQueue(call, 1);
5351 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5352 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5356 if (!newcall && (flags & RX_CALL_PEER_BUSY)) {
5357 /* The call channel is still busy; resetting the call doesn't change
5358 * that. However, if 'newcall' is set, we are processing a call
5359 * structure that has either been recycled from the free list, or has
5360 * been newly allocated. So, RX_CALL_PEER_BUSY is not relevant if
5361 * 'newcall' is set, since it describes a completely different call
5362 * channel which we do not care about. */
5363 call->flags |= RX_CALL_PEER_BUSY;
5366 rxi_ClearReceiveQueue(call);
5367 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5371 call->twind = call->conn->twind[call->channel];
5372 call->rwind = call->conn->rwind[call->channel];
5373 call->nSoftAcked = 0;
5374 call->nextCwind = 0;
5377 call->nCwindAcks = 0;
5378 call->nSoftAcks = 0;
5379 call->nHardAcks = 0;
5381 call->tfirst = call->rnext = call->tnext = 1;
5384 call->lastAcked = 0;
5385 call->localStatus = call->remoteStatus = 0;
5387 if (flags & RX_CALL_READER_WAIT) {
5388 #ifdef RX_ENABLE_LOCKS
5389 CV_BROADCAST(&call->cv_rq);
5391 osi_rxWakeup(&call->rq);
5394 if (flags & RX_CALL_WAIT_PACKETS) {
5395 MUTEX_ENTER(&rx_freePktQ_lock);
5396 rxi_PacketsUnWait(); /* XXX */
5397 MUTEX_EXIT(&rx_freePktQ_lock);
5399 #ifdef RX_ENABLE_LOCKS
5400 CV_SIGNAL(&call->cv_twind);
5402 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5403 osi_rxWakeup(&call->twind);
5406 if (flags & RX_CALL_WAIT_PROC) {
5407 rx_atomic_dec(&rx_nWaiting);
5409 #ifdef RX_ENABLE_LOCKS
5410 /* The following ensures that we don't mess with any queue while some
5411 * other thread might also be doing so. The call_queue_lock field is
5412 * is only modified under the call lock. If the call is in the process
5413 * of being removed from a queue, the call is not locked until the
5414 * the queue lock is dropped and only then is the call_queue_lock field
5415 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5416 * Note that any other routine which removes a call from a queue has to
5417 * obtain the queue lock before examing the queue and removing the call.
5419 if (call->call_queue_lock) {
5420 MUTEX_ENTER(call->call_queue_lock);
5421 if (opr_queue_IsOnQueue(&call->entry)) {
5422 opr_queue_Remove(&call->entry);
5424 MUTEX_EXIT(call->call_queue_lock);
5425 CLEAR_CALL_QUEUE_LOCK(call);
5427 #else /* RX_ENABLE_LOCKS */
5428 if (opr_queue_IsOnQueue(&call->entry)) {
5429 opr_queue_Remove(&call->entry);
5431 #endif /* RX_ENABLE_LOCKS */
5433 rxi_KeepAliveOff(call);
5434 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5437 /* Send an acknowledge for the indicated packet (seq,serial) of the
5438 * indicated call, for the indicated reason (reason). This
5439 * acknowledge will specifically acknowledge receiving the packet, and
5440 * will also specify which other packets for this call have been
5441 * received. This routine returns the packet that was used to the
5442 * caller. The caller is responsible for freeing it or re-using it.
5443 * This acknowledgement also returns the highest sequence number
5444 * actually read out by the higher level to the sender; the sender
5445 * promises to keep around packets that have not been read by the
5446 * higher level yet (unless, of course, the sender decides to abort
5447 * the call altogether). Any of p, seq, serial, pflags, or reason may
5448 * be set to zero without ill effect. That is, if they are zero, they
5449 * will not convey any information.
5450 * NOW there is a trailer field, after the ack where it will safely be
5451 * ignored by mundanes, which indicates the maximum size packet this
5452 * host can swallow. */
5454 struct rx_packet *optionalPacket; use to send ack (or null)
5455 int seq; Sequence number of the packet we are acking
5456 int serial; Serial number of the packet
5457 int pflags; Flags field from packet header
5458 int reason; Reason an acknowledge was prompted
5462 rxi_SendAck(struct rx_call *call,
5463 struct rx_packet *optionalPacket, int serial, int reason,
5466 struct rx_ackPacket *ap;
5467 struct rx_packet *p;
5468 struct opr_queue *cursor;
5471 afs_uint32 padbytes = 0;
5472 #ifdef RX_ENABLE_TSFPQ
5473 struct rx_ts_info_t * rx_ts_info;
5477 * Open the receive window once a thread starts reading packets
5479 if (call->rnext > 1) {
5480 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5483 /* Don't attempt to grow MTU if this is a critical ping */
5484 if (reason == RX_ACK_MTU) {
5485 /* keep track of per-call attempts, if we're over max, do in small
5486 * otherwise in larger? set a size to increment by, decrease
5489 if (call->conn->peer->maxPacketSize &&
5490 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5492 padbytes = call->conn->peer->maxPacketSize+16;
5494 padbytes = call->conn->peer->maxMTU + 128;
5496 /* do always try a minimum size ping */
5497 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5499 /* subtract the ack payload */
5500 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5501 reason = RX_ACK_PING;
5504 call->nHardAcks = 0;
5505 call->nSoftAcks = 0;
5506 if (call->rnext > call->lastAcked)
5507 call->lastAcked = call->rnext;
5511 rx_computelen(p, p->length); /* reset length, you never know */
5512 } /* where that's been... */
5513 #ifdef RX_ENABLE_TSFPQ
5515 RX_TS_INFO_GET(rx_ts_info);
5516 if ((p = rx_ts_info->local_special_packet)) {
5517 rx_computelen(p, p->length);
5518 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5519 rx_ts_info->local_special_packet = p;
5520 } else { /* We won't send the ack, but don't panic. */
5521 return optionalPacket;
5525 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5526 /* We won't send the ack, but don't panic. */
5527 return optionalPacket;
5532 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5535 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5536 #ifndef RX_ENABLE_TSFPQ
5537 if (!optionalPacket)
5540 return optionalPacket;
5542 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5543 if (rx_Contiguous(p) < templ) {
5544 #ifndef RX_ENABLE_TSFPQ
5545 if (!optionalPacket)
5548 return optionalPacket;
5553 /* MTUXXX failing to send an ack is very serious. We should */
5554 /* try as hard as possible to send even a partial ack; it's */
5555 /* better than nothing. */
5556 ap = (struct rx_ackPacket *)rx_DataOf(p);
5557 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5558 ap->reason = reason;
5560 /* The skew computation used to be bogus, I think it's better now. */
5561 /* We should start paying attention to skew. XXX */
5562 ap->serial = htonl(serial);
5563 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5566 * First packet not yet forwarded to reader. When ACKALL has been
5567 * sent the peer has been told that all received packets will be
5568 * delivered to the reader. The value 'rnext' is used internally
5569 * to refer to the next packet in the receive queue that must be
5570 * delivered to the reader. From the perspective of the peer it
5571 * already has so report the last sequence number plus one if there
5572 * are packets in the receive queue awaiting processing.
5574 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5575 !opr_queue_IsEmpty(&call->rq)) {
5576 ap->firstPacket = htonl(opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq + 1);
5578 ap->firstPacket = htonl(call->rnext);
5580 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5582 /* No fear of running out of ack packet here because there can only
5583 * be at most one window full of unacknowledged packets. The window
5584 * size must be constrained to be less than the maximum ack size,
5585 * of course. Also, an ack should always fit into a single packet
5586 * -- it should not ever be fragmented. */
5588 for (opr_queue_Scan(&call->rq, cursor)) {
5589 struct rx_packet *rqp
5590 = opr_queue_Entry(cursor, struct rx_packet, entry);
5592 if (!rqp || !call->rq.next
5593 || (rqp->header.seq > (call->rnext + call->rwind))) {
5594 #ifndef RX_ENABLE_TSFPQ
5595 if (!optionalPacket)
5598 rxi_CallError(call, RX_CALL_DEAD);
5599 return optionalPacket;
5602 while (rqp->header.seq > call->rnext + offset)
5603 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5604 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5606 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5607 #ifndef RX_ENABLE_TSFPQ
5608 if (!optionalPacket)
5611 rxi_CallError(call, RX_CALL_DEAD);
5612 return optionalPacket;
5618 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5620 /* these are new for AFS 3.3 */
5621 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5622 templ = htonl(templ);
5623 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5624 templ = htonl(call->conn->peer->ifMTU);
5625 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5626 sizeof(afs_int32), &templ);
5628 /* new for AFS 3.4 */
5629 templ = htonl(call->rwind);
5630 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5631 sizeof(afs_int32), &templ);
5633 /* new for AFS 3.5 */
5634 templ = htonl(call->conn->peer->ifDgramPackets);
5635 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5636 sizeof(afs_int32), &templ);
5638 p->header.serviceId = call->conn->serviceId;
5639 p->header.cid = (call->conn->cid | call->channel);
5640 p->header.callNumber = *call->callNumber;
5642 p->header.securityIndex = call->conn->securityIndex;
5643 p->header.epoch = call->conn->epoch;
5644 p->header.type = RX_PACKET_TYPE_ACK;
5645 p->header.flags = RX_SLOW_START_OK;
5646 if (reason == RX_ACK_PING) {
5647 p->header.flags |= RX_REQUEST_ACK;
5649 p->length = padbytes +
5650 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5653 /* not fast but we can potentially use this if truncated
5654 * fragments are delivered to figure out the mtu.
5656 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5657 sizeof(afs_int32), sizeof(afs_int32),
5661 if (call->conn->type == RX_CLIENT_CONNECTION)
5662 p->header.flags |= RX_CLIENT_INITIATED;
5666 if (rxdebug_active) {
5670 len = _snprintf(msg, sizeof(msg),
5671 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5672 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5673 ntohl(ap->serial), ntohl(ap->previousPacket),
5674 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5675 ap->nAcks, ntohs(ap->bufferSpace) );
5679 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5680 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5684 OutputDebugString(msg);
5686 #else /* AFS_NT40_ENV */
5688 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5689 ap->reason, ntohl(ap->previousPacket),
5690 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5692 for (offset = 0; offset < ap->nAcks; offset++)
5693 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5698 #endif /* AFS_NT40_ENV */
5701 int i, nbytes = p->length;
5703 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5704 if (nbytes <= p->wirevec[i].iov_len) {
5707 savelen = p->wirevec[i].iov_len;
5709 p->wirevec[i].iov_len = nbytes;
5711 rxi_Send(call, p, istack);
5712 p->wirevec[i].iov_len = savelen;
5716 nbytes -= p->wirevec[i].iov_len;
5719 if (rx_stats_active)
5720 rx_atomic_inc(&rx_stats.ackPacketsSent);
5721 #ifndef RX_ENABLE_TSFPQ
5722 if (!optionalPacket)
5725 return optionalPacket; /* Return packet for re-use by caller */
5729 struct rx_packet **list;
5734 /* Send all of the packets in the list in single datagram */
5736 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5737 int istack, int moreFlag)
5743 struct rx_connection *conn = call->conn;
5744 struct rx_peer *peer = conn->peer;
5746 MUTEX_ENTER(&peer->peer_lock);
5747 peer->nSent += xmit->len;
5748 if (xmit->resending)
5749 peer->reSends += xmit->len;
5750 MUTEX_EXIT(&peer->peer_lock);
5752 if (rx_stats_active) {
5753 if (xmit->resending)
5754 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5756 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5759 clock_GetTime(&now);
5761 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5765 /* Set the packet flags and schedule the resend events */
5766 /* Only request an ack for the last packet in the list */
5767 for (i = 0; i < xmit->len; i++) {
5768 struct rx_packet *packet = xmit->list[i];
5770 /* Record the time sent */
5771 packet->timeSent = now;
5772 packet->flags |= RX_PKTFLAG_SENT;
5774 /* Ask for an ack on retransmitted packets, on every other packet
5775 * if the peer doesn't support slow start. Ask for an ack on every
5776 * packet until the congestion window reaches the ack rate. */
5777 if (packet->header.serial) {
5780 packet->firstSent = now;
5781 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5782 || (!(call->flags & RX_CALL_SLOW_START_OK)
5783 && (packet->header.seq & 1)))) {
5788 /* Tag this packet as not being the last in this group,
5789 * for the receiver's benefit */
5790 if (i < xmit->len - 1 || moreFlag) {
5791 packet->header.flags |= RX_MORE_PACKETS;
5796 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5799 /* Since we're about to send a data packet to the peer, it's
5800 * safe to nuke any scheduled end-of-packets ack */
5801 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5803 MUTEX_EXIT(&call->lock);
5804 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5805 if (xmit->len > 1) {
5806 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5808 rxi_SendPacket(call, conn, xmit->list[0], istack);
5810 MUTEX_ENTER(&call->lock);
5811 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5813 /* Tell the RTO calculation engine that we have sent a packet, and
5814 * if it was the last one */
5815 rxi_rto_packet_sent(call, lastPacket, istack);
5817 /* Update last send time for this call (for keep-alive
5818 * processing), and for the connection (so that we can discover
5819 * idle connections) */
5820 conn->lastSendTime = call->lastSendTime = clock_Sec();
5821 /* Let a set of retransmits trigger an idle timeout */
5822 if (!xmit->resending)
5823 call->lastSendData = call->lastSendTime;
5826 /* When sending packets we need to follow these rules:
5827 * 1. Never send more than maxDgramPackets in a jumbogram.
5828 * 2. Never send a packet with more than two iovecs in a jumbogram.
5829 * 3. Never send a retransmitted packet in a jumbogram.
5830 * 4. Never send more than cwind/4 packets in a jumbogram
5831 * We always keep the last list we should have sent so we
5832 * can set the RX_MORE_PACKETS flags correctly.
5836 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5841 struct xmitlist working;
5842 struct xmitlist last;
5844 struct rx_peer *peer = call->conn->peer;
5845 int morePackets = 0;
5847 memset(&last, 0, sizeof(struct xmitlist));
5848 working.list = &list[0];
5850 working.resending = 0;
5852 recovery = call->flags & RX_CALL_FAST_RECOVER;
5854 for (i = 0; i < len; i++) {
5855 /* Does the current packet force us to flush the current list? */
5857 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5858 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5860 /* This sends the 'last' list and then rolls the current working
5861 * set into the 'last' one, and resets the working set */
5864 rxi_SendList(call, &last, istack, 1);
5865 /* If the call enters an error state stop sending, or if
5866 * we entered congestion recovery mode, stop sending */
5868 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5873 working.resending = 0;
5874 working.list = &list[i];
5876 /* Add the current packet to the list if it hasn't been acked.
5877 * Otherwise adjust the list pointer to skip the current packet. */
5878 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5881 if (list[i]->header.serial)
5882 working.resending = 1;
5884 /* Do we need to flush the list? */
5885 if (working.len >= (int)peer->maxDgramPackets
5886 || working.len >= (int)call->nDgramPackets
5887 || working.len >= (int)call->cwind
5888 || list[i]->header.serial
5889 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5891 rxi_SendList(call, &last, istack, 1);
5892 /* If the call enters an error state stop sending, or if
5893 * we entered congestion recovery mode, stop sending */
5895 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5900 working.resending = 0;
5901 working.list = &list[i + 1];
5904 if (working.len != 0) {
5905 osi_Panic("rxi_SendList error");
5907 working.list = &list[i + 1];
5911 /* Send the whole list when the call is in receive mode, when
5912 * the call is in eof mode, when we are in fast recovery mode,
5913 * and when we have the last packet */
5914 /* XXX - The accesses to app.mode aren't safe, as this may be called by
5915 * the listener or event threads
5917 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5918 || (call->flags & RX_CALL_FLUSH)
5919 || (call->flags & RX_CALL_FAST_RECOVER)) {
5920 /* Check for the case where the current list contains
5921 * an acked packet. Since we always send retransmissions
5922 * in a separate packet, we only need to check the first
5923 * packet in the list */
5924 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5928 rxi_SendList(call, &last, istack, morePackets);
5929 /* If the call enters an error state stop sending, or if
5930 * we entered congestion recovery mode, stop sending */
5932 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5936 rxi_SendList(call, &working, istack, 0);
5938 } else if (last.len > 0) {
5939 rxi_SendList(call, &last, istack, 0);
5940 /* Packets which are in 'working' are not sent by this call */
5945 * Check if the peer for the given call is known to be dead
5947 * If the call's peer appears dead (it has encountered fatal network errors
5948 * since the call started) the call is killed with RX_CALL_DEAD if the call
5949 * is active. Otherwise, we do nothing.
5951 * @param[in] call The call to check
5954 * @retval 0 The call is fine, and we haven't done anything to the call
5955 * @retval nonzero The call's peer appears dead, and the call has been
5956 * terminated if it was active
5958 * @pre call->lock must be locked
5961 rxi_CheckPeerDead(struct rx_call *call)
5963 #ifdef AFS_RXERRQ_ENV
5966 if (call->state == RX_STATE_DALLY) {
5970 peererrs = rx_atomic_read(&call->conn->peer->neterrs);
5971 if (call->neterr_gen < peererrs) {
5972 /* we have received network errors since this call started; kill
5974 if (call->state == RX_STATE_ACTIVE) {
5975 rxi_CallError(call, RX_CALL_DEAD);
5979 if (call->neterr_gen > peererrs) {
5980 /* someone has reset the number of peer errors; set the call error gen
5981 * so we can detect if more errors are encountered */
5982 call->neterr_gen = peererrs;
5989 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
5991 struct rx_call *call = arg0;
5992 struct rx_peer *peer;
5993 struct opr_queue *cursor;
5994 struct clock maxTimeout = { 60, 0 };
5996 MUTEX_ENTER(&call->lock);
5998 peer = call->conn->peer;
6000 /* Make sure that the event pointer is removed from the call
6001 * structure, since there is no longer a per-call retransmission
6003 if (event == call->resendEvent) {
6004 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
6005 rxevent_Put(call->resendEvent);
6006 call->resendEvent = NULL;
6009 rxi_CheckPeerDead(call);
6011 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
6012 rxi_CheckBusy(call);
6015 if (opr_queue_IsEmpty(&call->tq)) {
6016 /* Nothing to do. This means that we've been raced, and that an
6017 * ACK has come in between when we were triggered, and when we
6018 * actually got to run. */
6022 /* We're in loss recovery */
6023 call->flags |= RX_CALL_FAST_RECOVER;
6025 /* Mark all of the pending packets in the queue as being lost */
6026 for (opr_queue_Scan(&call->tq, cursor)) {
6027 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
6028 if (!(p->flags & RX_PKTFLAG_ACKED))
6029 p->flags &= ~RX_PKTFLAG_SENT;
6032 /* We're resending, so we double the timeout of the call. This will be
6033 * dropped back down by the first successful ACK that we receive.
6035 * We apply a maximum value here of 60 seconds
6037 clock_Add(&call->rto, &call->rto);
6038 if (clock_Gt(&call->rto, &maxTimeout))
6039 call->rto = maxTimeout;
6041 /* Packet loss is most likely due to congestion, so drop our window size
6042 * and start again from the beginning */
6043 if (peer->maxDgramPackets >1) {
6044 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
6045 call->MTU = MIN(peer->natMTU, peer->maxMTU);
6047 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
6048 call->nDgramPackets = 1;
6050 call->nextCwind = 1;
6053 MUTEX_ENTER(&peer->peer_lock);
6054 peer->MTU = call->MTU;
6055 peer->cwind = call->cwind;
6056 peer->nDgramPackets = 1;
6058 call->congestSeq = peer->congestSeq;
6059 MUTEX_EXIT(&peer->peer_lock);
6061 rxi_Start(call, istack);
6064 MUTEX_EXIT(&call->lock);
6067 /* This routine is called when new packets are readied for
6068 * transmission and when retransmission may be necessary, or when the
6069 * transmission window or burst count are favourable. This should be
6070 * better optimized for new packets, the usual case, now that we've
6071 * got rid of queues of send packets. XXXXXXXXXXX */
6073 rxi_Start(struct rx_call *call, int istack)
6075 struct opr_queue *cursor;
6076 #ifdef RX_ENABLE_LOCKS
6077 struct opr_queue *store;
6083 #ifdef RX_ENABLE_LOCKS
6084 if (rx_stats_active)
6085 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
6090 if (!opr_queue_IsEmpty(&call->tq)) { /* If we have anything to send */
6091 /* Send (or resend) any packets that need it, subject to
6092 * window restrictions and congestion burst control
6093 * restrictions. Ask for an ack on the last packet sent in
6094 * this burst. For now, we're relying upon the window being
6095 * considerably bigger than the largest number of packets that
6096 * are typically sent at once by one initial call to
6097 * rxi_Start. This is probably bogus (perhaps we should ask
6098 * for an ack when we're half way through the current
6099 * window?). Also, for non file transfer applications, this
6100 * may end up asking for an ack for every packet. Bogus. XXXX
6103 * But check whether we're here recursively, and let the other guy
6106 #ifdef RX_ENABLE_LOCKS
6107 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6108 call->flags |= RX_CALL_TQ_BUSY;
6110 #endif /* RX_ENABLE_LOCKS */
6112 #ifdef RX_ENABLE_LOCKS
6113 call->flags &= ~RX_CALL_NEED_START;
6114 #endif /* RX_ENABLE_LOCKS */
6116 maxXmitPackets = MIN(call->twind, call->cwind);
6117 for (opr_queue_Scan(&call->tq, cursor)) {
6119 = opr_queue_Entry(cursor, struct rx_packet, entry);
6121 if (p->flags & RX_PKTFLAG_ACKED) {
6122 /* Since we may block, don't trust this */
6123 if (rx_stats_active)
6124 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6125 continue; /* Ignore this packet if it has been acknowledged */
6128 /* Turn off all flags except these ones, which are the same
6129 * on each transmission */
6130 p->header.flags &= RX_PRESET_FLAGS;
6132 if (p->header.seq >=
6133 call->tfirst + MIN((int)call->twind,
6134 (int)(call->nSoftAcked +
6136 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6137 /* Note: if we're waiting for more window space, we can
6138 * still send retransmits; hence we don't return here, but
6139 * break out to schedule a retransmit event */
6140 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6141 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6146 /* Transmit the packet if it needs to be sent. */
6147 if (!(p->flags & RX_PKTFLAG_SENT)) {
6148 if (nXmitPackets == maxXmitPackets) {
6149 rxi_SendXmitList(call, call->xmitList,
6150 nXmitPackets, istack);
6153 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6154 *(call->callNumber), p));
6155 call->xmitList[nXmitPackets++] = p;
6157 } /* end of the queue_Scan */
6159 /* xmitList now hold pointers to all of the packets that are
6160 * ready to send. Now we loop to send the packets */
6161 if (nXmitPackets > 0) {
6162 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6166 #ifdef RX_ENABLE_LOCKS
6168 /* We went into the error state while sending packets. Now is
6169 * the time to reset the call. This will also inform the using
6170 * process that the call is in an error state.
6172 if (rx_stats_active)
6173 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6174 call->flags &= ~RX_CALL_TQ_BUSY;
6175 rxi_WakeUpTransmitQueue(call);
6176 rxi_CallError(call, call->error);
6180 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6182 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6183 /* Some packets have received acks. If they all have, we can clear
6184 * the transmit queue.
6187 for (opr_queue_ScanSafe(&call->tq, cursor, store)) {
6189 = opr_queue_Entry(cursor, struct rx_packet, entry);
6191 if (p->header.seq < call->tfirst
6192 && (p->flags & RX_PKTFLAG_ACKED)) {
6193 opr_queue_Remove(&p->entry);
6194 #ifdef RX_TRACK_PACKETS
6195 p->flags &= ~RX_PKTFLAG_TQ;
6197 #ifdef RXDEBUG_PACKET
6205 call->flags |= RX_CALL_TQ_CLEARME;
6207 if (call->flags & RX_CALL_TQ_CLEARME)
6208 rxi_ClearTransmitQueue(call, 1);
6209 } while (call->flags & RX_CALL_NEED_START);
6211 * TQ references no longer protected by this flag; they must remain
6212 * protected by the call lock.
6214 call->flags &= ~RX_CALL_TQ_BUSY;
6215 rxi_WakeUpTransmitQueue(call);
6217 call->flags |= RX_CALL_NEED_START;
6219 #endif /* RX_ENABLE_LOCKS */
6221 rxi_rto_cancel(call);
6225 /* Also adjusts the keep alive parameters for the call, to reflect
6226 * that we have just sent a packet (so keep alives aren't sent
6229 rxi_Send(struct rx_call *call, struct rx_packet *p,
6232 struct rx_connection *conn = call->conn;
6234 /* Stamp each packet with the user supplied status */
6235 p->header.userStatus = call->localStatus;
6237 /* Allow the security object controlling this call's security to
6238 * make any last-minute changes to the packet */
6239 RXS_SendPacket(conn->securityObject, call, p);
6241 /* Since we're about to send SOME sort of packet to the peer, it's
6242 * safe to nuke any scheduled end-of-packets ack */
6243 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
6245 /* Actually send the packet, filling in more connection-specific fields */
6246 MUTEX_EXIT(&call->lock);
6247 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6248 rxi_SendPacket(call, conn, p, istack);
6249 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6250 MUTEX_ENTER(&call->lock);
6252 /* Update last send time for this call (for keep-alive
6253 * processing), and for the connection (so that we can discover
6254 * idle connections) */
6255 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6256 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6257 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6259 conn->lastSendTime = call->lastSendTime = clock_Sec();
6260 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6261 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6262 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6263 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6264 RX_ACK_PING_RESPONSE)))
6265 call->lastSendData = call->lastSendTime;
6269 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6270 * that things are fine. Also called periodically to guarantee that nothing
6271 * falls through the cracks (e.g. (error + dally) connections have keepalive
6272 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6274 * haveCTLock Set if calling from rxi_ReapConnections
6277 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6279 struct rx_connection *conn = call->conn;
6281 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6282 afs_uint32 fudgeFactor;
6285 int idle_timeout = 0;
6286 afs_int32 clock_diff = 0;
6288 if (rxi_CheckPeerDead(call)) {
6294 /* Large swings in the clock can have a significant impact on
6295 * the performance of RX call processing. Forward clock shifts
6296 * will result in premature event triggering or timeouts.
6297 * Backward shifts can result in calls not completing until
6298 * the clock catches up with the original start clock value.
6300 * If a backward clock shift of more than five minutes is noticed,
6301 * just fail the call.
6303 if (now < call->lastSendTime)
6304 clock_diff = call->lastSendTime - now;
6305 if (now < call->startWait)
6306 clock_diff = MAX(clock_diff, call->startWait - now);
6307 if (now < call->lastReceiveTime)
6308 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6309 if (clock_diff > 5 * 60)
6311 if (call->state == RX_STATE_ACTIVE)
6312 rxi_CallError(call, RX_CALL_TIMEOUT);
6316 #ifdef RX_ENABLE_LOCKS
6317 if (call->flags & RX_CALL_TQ_BUSY) {
6318 /* Call is active and will be reset by rxi_Start if it's
6319 * in an error state.
6324 /* RTT + 8*MDEV, rounded up to the next second. */
6325 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6326 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6328 deadTime = conn->secondsUntilDead + fudgeFactor;
6329 /* These are computed to the second (+- 1 second). But that's
6330 * good enough for these values, which should be a significant
6331 * number of seconds. */
6332 if (now > (call->lastReceiveTime + deadTime)) {
6333 if (call->state == RX_STATE_ACTIVE) {
6334 #ifdef AFS_ADAPT_PMTU
6335 # if defined(KERNEL) && defined(AFS_SUN5_ENV)
6337 # if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6338 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6339 ip_stack_t *ipst = ns->netstack_ip;
6341 ire = ire_cache_lookup(conn->peer->host
6342 # if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6344 # if defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID)
6346 # if defined(GLOBAL_NETSTACKID)
6353 if (ire && ire->ire_max_frag > 0)
6354 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6356 # if defined(GLOBAL_NETSTACKID)
6360 #endif /* AFS_ADAPT_PMTU */
6361 cerror = RX_CALL_DEAD;
6364 #ifdef RX_ENABLE_LOCKS
6365 /* Cancel pending events */
6366 rxevent_Cancel(&call->delayedAckEvent, call,
6367 RX_CALL_REFCOUNT_DELAY);
6368 rxi_rto_cancel(call);
6369 rxevent_Cancel(&call->keepAliveEvent, call,
6370 RX_CALL_REFCOUNT_ALIVE);
6371 rxevent_Cancel(&call->growMTUEvent, call,
6372 RX_CALL_REFCOUNT_MTU);
6373 MUTEX_ENTER(&rx_refcnt_mutex);
6374 /* if rxi_FreeCall returns 1 it has freed the call */
6375 if (call->refCount == 0 &&
6376 rxi_FreeCall(call, haveCTLock))
6378 MUTEX_EXIT(&rx_refcnt_mutex);
6381 MUTEX_EXIT(&rx_refcnt_mutex);
6383 #else /* RX_ENABLE_LOCKS */
6384 rxi_FreeCall(call, 0);
6386 #endif /* RX_ENABLE_LOCKS */
6388 /* Non-active calls are destroyed if they are not responding
6389 * to pings; active calls are simply flagged in error, so the
6390 * attached process can die reasonably gracefully. */
6393 if (conn->idleDeadDetection) {
6394 if (conn->idleDeadTime) {
6395 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6399 /* see if we have a non-activity timeout */
6400 if (call->startWait && ((call->startWait + idleDeadTime) < now) &&
6401 (call->flags & RX_CALL_READER_WAIT)) {
6402 if (call->state == RX_STATE_ACTIVE) {
6403 cerror = RX_CALL_TIMEOUT;
6408 if (call->lastSendData && ((call->lastSendData + idleDeadTime) < now)) {
6409 if (call->state == RX_STATE_ACTIVE) {
6410 cerror = conn->service ? conn->service->idleDeadErr : RX_CALL_IDLE;
6418 if (conn->hardDeadTime) {
6419 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6422 /* see if we have a hard timeout */
6424 && (now > (hardDeadTime + call->startTime.sec))) {
6425 if (call->state == RX_STATE_ACTIVE)
6426 rxi_CallError(call, RX_CALL_TIMEOUT);
6431 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6432 call->lastReceiveTime) {
6433 int oldMTU = conn->peer->ifMTU;
6435 /* if we thought we could send more, perhaps things got worse */
6436 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6437 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6438 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6439 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6441 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6443 /* minimum capped in SetPeerMtu */
6444 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6447 conn->lastPacketSize = 0;
6449 /* needed so ResetCall doesn't clobber us. */
6450 call->MTU = conn->peer->ifMTU;
6452 /* if we never succeeded, let the error pass out as-is */
6453 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6454 cerror = conn->msgsizeRetryErr;
6457 rxi_CallError(call, cerror);
6462 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6463 void *dummy, int dummy2)
6465 struct rx_connection *conn = arg1;
6466 struct rx_header theader;
6467 char tbuffer[1 + sizeof(struct rx_header)];
6468 struct sockaddr_in taddr;
6471 struct iovec tmpiov[2];
6474 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6477 tp = &tbuffer[sizeof(struct rx_header)];
6478 taddr.sin_family = AF_INET;
6479 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6480 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6481 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6482 taddr.sin_len = sizeof(struct sockaddr_in);
6484 memset(&theader, 0, sizeof(theader));
6485 theader.epoch = htonl(999);
6487 theader.callNumber = 0;
6490 theader.type = RX_PACKET_TYPE_VERSION;
6491 theader.flags = RX_LAST_PACKET;
6492 theader.serviceId = 0;
6494 memcpy(tbuffer, &theader, sizeof(theader));
6495 memcpy(tp, &a, sizeof(a));
6496 tmpiov[0].iov_base = tbuffer;
6497 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6499 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6501 MUTEX_ENTER(&conn->conn_data_lock);
6502 MUTEX_ENTER(&rx_refcnt_mutex);
6503 /* Only reschedule ourselves if the connection would not be destroyed */
6504 if (conn->refCount <= 1) {
6505 rxevent_Put(conn->natKeepAliveEvent);
6506 conn->natKeepAliveEvent = NULL;
6507 MUTEX_EXIT(&rx_refcnt_mutex);
6508 MUTEX_EXIT(&conn->conn_data_lock);
6509 rx_DestroyConnection(conn); /* drop the reference for this */
6511 conn->refCount--; /* drop the reference for this */
6512 MUTEX_EXIT(&rx_refcnt_mutex);
6513 rxevent_Put(conn->natKeepAliveEvent);
6514 conn->natKeepAliveEvent = NULL;
6515 rxi_ScheduleNatKeepAliveEvent(conn);
6516 MUTEX_EXIT(&conn->conn_data_lock);
6521 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6523 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6524 struct clock when, now;
6525 clock_GetTime(&now);
6527 when.sec += conn->secondsUntilNatPing;
6528 MUTEX_ENTER(&rx_refcnt_mutex);
6529 conn->refCount++; /* hold a reference for this */
6530 MUTEX_EXIT(&rx_refcnt_mutex);
6531 conn->natKeepAliveEvent =
6532 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6537 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6539 MUTEX_ENTER(&conn->conn_data_lock);
6540 conn->secondsUntilNatPing = seconds;
6542 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6543 rxi_ScheduleNatKeepAliveEvent(conn);
6545 conn->flags |= RX_CONN_NAT_PING;
6547 MUTEX_EXIT(&conn->conn_data_lock);
6550 /* When a call is in progress, this routine is called occasionally to
6551 * make sure that some traffic has arrived (or been sent to) the peer.
6552 * If nothing has arrived in a reasonable amount of time, the call is
6553 * declared dead; if nothing has been sent for a while, we send a
6554 * keep-alive packet (if we're actually trying to keep the call alive)
6557 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6560 struct rx_call *call = arg1;
6561 struct rx_connection *conn;
6564 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6565 MUTEX_ENTER(&call->lock);
6567 if (event == call->keepAliveEvent) {
6568 rxevent_Put(call->keepAliveEvent);
6569 call->keepAliveEvent = NULL;
6574 if (rxi_CheckCall(call, 0)) {
6575 MUTEX_EXIT(&call->lock);
6579 /* Don't try to keep alive dallying calls */
6580 if (call->state == RX_STATE_DALLY) {
6581 MUTEX_EXIT(&call->lock);
6586 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6587 /* Don't try to send keepalives if there is unacknowledged data */
6588 /* the rexmit code should be good enough, this little hack
6589 * doesn't quite work XXX */
6590 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6592 rxi_ScheduleKeepAliveEvent(call);
6593 MUTEX_EXIT(&call->lock);
6596 /* Does what's on the nameplate. */
6598 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6600 struct rx_call *call = arg1;
6601 struct rx_connection *conn;
6603 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6604 MUTEX_ENTER(&call->lock);
6606 if (event == call->growMTUEvent) {
6607 rxevent_Put(call->growMTUEvent);
6608 call->growMTUEvent = NULL;
6611 if (rxi_CheckCall(call, 0)) {
6612 MUTEX_EXIT(&call->lock);
6616 /* Don't bother with dallying calls */
6617 if (call->state == RX_STATE_DALLY) {
6618 MUTEX_EXIT(&call->lock);
6625 * keep being scheduled, just don't do anything if we're at peak,
6626 * or we're not set up to be properly handled (idle timeout required)
6628 if ((conn->peer->maxPacketSize != 0) &&
6629 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6630 conn->idleDeadDetection)
6631 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6632 rxi_ScheduleGrowMTUEvent(call, 0);
6633 MUTEX_EXIT(&call->lock);
6637 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6639 if (!call->keepAliveEvent) {
6640 struct clock when, now;
6641 clock_GetTime(&now);
6643 when.sec += call->conn->secondsUntilPing;
6644 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6645 call->keepAliveEvent =
6646 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6651 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6653 if (!call->growMTUEvent) {
6654 struct clock when, now;
6656 clock_GetTime(&now);
6659 if (call->conn->secondsUntilPing)
6660 secs = (6*call->conn->secondsUntilPing)-1;
6662 if (call->conn->secondsUntilDead)
6663 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6667 CALL_HOLD(call, RX_CALL_REFCOUNT_MTU);
6668 call->growMTUEvent =
6669 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6673 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6675 rxi_KeepAliveOn(struct rx_call *call)
6677 /* Pretend last packet received was received now--i.e. if another
6678 * packet isn't received within the keep alive time, then the call
6679 * will die; Initialize last send time to the current time--even
6680 * if a packet hasn't been sent yet. This will guarantee that a
6681 * keep-alive is sent within the ping time */
6682 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6683 rxi_ScheduleKeepAliveEvent(call);
6687 * Solely in order that callers not need to include rx_call.h
6690 rx_KeepAliveOff(struct rx_call *call)
6692 rxi_KeepAliveOff(call);
6695 rx_KeepAliveOn(struct rx_call *call)
6697 rxi_KeepAliveOn(call);
6701 rxi_GrowMTUOn(struct rx_call *call)
6703 struct rx_connection *conn = call->conn;
6704 MUTEX_ENTER(&conn->conn_data_lock);
6705 conn->lastPingSizeSer = conn->lastPingSize = 0;
6706 MUTEX_EXIT(&conn->conn_data_lock);
6707 rxi_ScheduleGrowMTUEvent(call, 1);
6710 /* This routine is called to send connection abort messages
6711 * that have been delayed to throttle looping clients. */
6713 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6716 struct rx_connection *conn = arg1;
6719 struct rx_packet *packet;
6721 MUTEX_ENTER(&conn->conn_data_lock);
6722 rxevent_Put(conn->delayedAbortEvent);
6723 conn->delayedAbortEvent = NULL;
6724 error = htonl(conn->error);
6726 MUTEX_EXIT(&conn->conn_data_lock);
6727 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6730 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6731 RX_PACKET_TYPE_ABORT, (char *)&error,
6733 rxi_FreePacket(packet);
6737 /* This routine is called to send call abort messages
6738 * that have been delayed to throttle looping clients. */
6740 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6743 struct rx_call *call = arg1;
6746 struct rx_packet *packet;
6748 MUTEX_ENTER(&call->lock);
6749 rxevent_Put(call->delayedAbortEvent);
6750 call->delayedAbortEvent = NULL;
6751 error = htonl(call->error);
6753 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6756 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6757 (char *)&error, sizeof(error), 0);
6758 rxi_FreePacket(packet);
6760 MUTEX_EXIT(&call->lock);
6761 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6764 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6765 * seconds) to ask the client to authenticate itself. The routine
6766 * issues a challenge to the client, which is obtained from the
6767 * security object associated with the connection */
6769 rxi_ChallengeEvent(struct rxevent *event,
6770 void *arg0, void *arg1, int tries)
6772 struct rx_connection *conn = arg0;
6775 rxevent_Put(conn->challengeEvent);
6776 conn->challengeEvent = NULL;
6779 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6780 struct rx_packet *packet;
6781 struct clock when, now;
6784 /* We've failed to authenticate for too long.
6785 * Reset any calls waiting for authentication;
6786 * they are all in RX_STATE_PRECALL.
6790 MUTEX_ENTER(&conn->conn_call_lock);
6791 for (i = 0; i < RX_MAXCALLS; i++) {
6792 struct rx_call *call = conn->call[i];
6794 MUTEX_ENTER(&call->lock);
6795 if (call->state == RX_STATE_PRECALL) {
6796 rxi_CallError(call, RX_CALL_DEAD);
6797 rxi_SendCallAbort(call, NULL, 0, 0);
6799 MUTEX_EXIT(&call->lock);
6802 MUTEX_EXIT(&conn->conn_call_lock);
6806 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6808 /* If there's no packet available, do this later. */
6809 RXS_GetChallenge(conn->securityObject, conn, packet);
6810 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6811 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6812 rxi_FreePacket(packet);
6814 clock_GetTime(&now);
6816 when.sec += RX_CHALLENGE_TIMEOUT;
6817 conn->challengeEvent =
6818 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6823 /* Call this routine to start requesting the client to authenticate
6824 * itself. This will continue until authentication is established,
6825 * the call times out, or an invalid response is returned. The
6826 * security object associated with the connection is asked to create
6827 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6828 * defined earlier. */
6830 rxi_ChallengeOn(struct rx_connection *conn)
6832 if (!conn->challengeEvent) {
6833 RXS_CreateChallenge(conn->securityObject, conn);
6834 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6839 /* rxi_ComputeRoundTripTime is called with peer locked. */
6840 /* peer may be null */
6842 rxi_ComputeRoundTripTime(struct rx_packet *p,
6843 struct rx_ackPacket *ack,
6844 struct rx_call *call,
6845 struct rx_peer *peer,
6848 struct clock thisRtt, *sentp;
6852 /* If the ACK is delayed, then do nothing */
6853 if (ack->reason == RX_ACK_DELAY)
6856 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6857 * their RTT multiple times, so only include the RTT of the last packet
6859 if (p->flags & RX_JUMBO_PACKET)
6862 /* Use the serial number to determine which transmission the ACK is for,
6863 * and set the sent time to match this. If we have no serial number, then
6864 * only use the ACK for RTT calculations if the packet has not been
6868 serial = ntohl(ack->serial);
6870 if (serial == p->header.serial) {
6871 sentp = &p->timeSent;
6872 } else if (serial == p->firstSerial) {
6873 sentp = &p->firstSent;
6874 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6875 sentp = &p->firstSent;
6879 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6880 sentp = &p->firstSent;
6887 if (clock_Lt(&thisRtt, sentp))
6888 return; /* somebody set the clock back, don't count this time. */
6890 clock_Sub(&thisRtt, sentp);
6891 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6892 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6894 if (clock_IsZero(&thisRtt)) {
6896 * The actual round trip time is shorter than the
6897 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6898 * Since we can't tell which at the moment we will assume 1ms.
6900 thisRtt.usec = 1000;
6903 if (rx_stats_active) {
6904 MUTEX_ENTER(&rx_stats_mutex);
6905 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6906 rx_stats.minRtt = thisRtt;
6907 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6908 if (thisRtt.sec > 60) {
6909 MUTEX_EXIT(&rx_stats_mutex);
6910 return; /* somebody set the clock ahead */
6912 rx_stats.maxRtt = thisRtt;
6914 clock_Add(&rx_stats.totalRtt, &thisRtt);
6915 rx_atomic_inc(&rx_stats.nRttSamples);
6916 MUTEX_EXIT(&rx_stats_mutex);
6919 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6921 /* Apply VanJacobson round-trip estimations */
6926 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6927 * srtt is stored as fixed point with 3 bits after the binary
6928 * point (i.e., scaled by 8). The following magic is
6929 * equivalent to the smoothing algorithm in rfc793 with an
6930 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6931 * srtt'*8 = rtt + srtt*7
6932 * srtt'*8 = srtt*8 + rtt - srtt
6933 * srtt' = srtt + rtt/8 - srtt/8
6934 * srtt' = srtt + (rtt - srtt)/8
6937 delta = _8THMSEC(&thisRtt) - call->rtt;
6938 call->rtt += (delta >> 3);
6941 * We accumulate a smoothed rtt variance (actually, a smoothed
6942 * mean difference), then set the retransmit timer to smoothed
6943 * rtt + 4 times the smoothed variance (was 2x in van's original
6944 * paper, but 4x works better for me, and apparently for him as
6946 * rttvar is stored as
6947 * fixed point with 2 bits after the binary point (scaled by
6948 * 4). The following is equivalent to rfc793 smoothing with
6949 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6950 * rttvar'*4 = rttvar*3 + |delta|
6951 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6952 * rttvar' = rttvar + |delta|/4 - rttvar/4
6953 * rttvar' = rttvar + (|delta| - rttvar)/4
6954 * This replaces rfc793's wired-in beta.
6955 * dev*4 = dev*4 + (|actual - expected| - dev)
6961 delta -= (call->rtt_dev << 1);
6962 call->rtt_dev += (delta >> 3);
6964 /* I don't have a stored RTT so I start with this value. Since I'm
6965 * probably just starting a call, and will be pushing more data down
6966 * this, I expect congestion to increase rapidly. So I fudge a
6967 * little, and I set deviance to half the rtt. In practice,
6968 * deviance tends to approach something a little less than
6969 * half the smoothed rtt. */
6970 call->rtt = _8THMSEC(&thisRtt) + 8;
6971 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6973 /* the smoothed RTT time is RTT + 4*MDEV
6975 * We allow a user specified minimum to be set for this, to allow clamping
6976 * at a minimum value in the same way as TCP. In addition, we have to allow
6977 * for the possibility that this packet is answered by a delayed ACK, so we
6978 * add on a fixed 200ms to account for that timer expiring.
6981 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6982 rx_minPeerTimeout) + 200;
6983 clock_Zero(&call->rto);
6984 clock_Addmsec(&call->rto, rtt_timeout);
6986 /* Update the peer, so any new calls start with our values */
6987 peer->rtt_dev = call->rtt_dev;
6988 peer->rtt = call->rtt;
6990 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6991 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6995 /* Find all server connections that have not been active for a long time, and
6998 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
7001 struct clock now, when;
7002 clock_GetTime(&now);
7004 /* Find server connection structures that haven't been used for
7005 * greater than rx_idleConnectionTime */
7007 struct rx_connection **conn_ptr, **conn_end;
7008 int i, havecalls = 0;
7009 MUTEX_ENTER(&rx_connHashTable_lock);
7010 for (conn_ptr = &rx_connHashTable[0], conn_end =
7011 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
7013 struct rx_connection *conn, *next;
7014 struct rx_call *call;
7018 for (conn = *conn_ptr; conn; conn = next) {
7019 /* XXX -- Shouldn't the connection be locked? */
7022 for (i = 0; i < RX_MAXCALLS; i++) {
7023 call = conn->call[i];
7027 code = MUTEX_TRYENTER(&call->lock);
7030 result = rxi_CheckCall(call, 1);
7031 MUTEX_EXIT(&call->lock);
7033 /* If CheckCall freed the call, it might
7034 * have destroyed the connection as well,
7035 * which screws up the linked lists.
7041 if (conn->type == RX_SERVER_CONNECTION) {
7042 /* This only actually destroys the connection if
7043 * there are no outstanding calls */
7044 MUTEX_ENTER(&conn->conn_data_lock);
7045 MUTEX_ENTER(&rx_refcnt_mutex);
7046 if (!havecalls && !conn->refCount
7047 && ((conn->lastSendTime + rx_idleConnectionTime) <
7049 conn->refCount++; /* it will be decr in rx_DestroyConn */
7050 MUTEX_EXIT(&rx_refcnt_mutex);
7051 MUTEX_EXIT(&conn->conn_data_lock);
7052 #ifdef RX_ENABLE_LOCKS
7053 rxi_DestroyConnectionNoLock(conn);
7054 #else /* RX_ENABLE_LOCKS */
7055 rxi_DestroyConnection(conn);
7056 #endif /* RX_ENABLE_LOCKS */
7058 #ifdef RX_ENABLE_LOCKS
7060 MUTEX_EXIT(&rx_refcnt_mutex);
7061 MUTEX_EXIT(&conn->conn_data_lock);
7063 #endif /* RX_ENABLE_LOCKS */
7067 #ifdef RX_ENABLE_LOCKS
7068 while (rx_connCleanup_list) {
7069 struct rx_connection *conn;
7070 conn = rx_connCleanup_list;
7071 rx_connCleanup_list = rx_connCleanup_list->next;
7072 MUTEX_EXIT(&rx_connHashTable_lock);
7073 rxi_CleanupConnection(conn);
7074 MUTEX_ENTER(&rx_connHashTable_lock);
7076 MUTEX_EXIT(&rx_connHashTable_lock);
7077 #endif /* RX_ENABLE_LOCKS */
7080 /* Find any peer structures that haven't been used (haven't had an
7081 * associated connection) for greater than rx_idlePeerTime */
7083 struct rx_peer **peer_ptr, **peer_end;
7087 * Why do we need to hold the rx_peerHashTable_lock across
7088 * the incrementing of peer_ptr since the rx_peerHashTable
7089 * array is not changing? We don't.
7091 * By dropping the lock periodically we can permit other
7092 * activities to be performed while a rxi_ReapConnections
7093 * call is in progress. The goal of reap connections
7094 * is to clean up quickly without causing large amounts
7095 * of contention. Therefore, it is important that global
7096 * mutexes not be held for extended periods of time.
7098 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7099 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7101 struct rx_peer *peer, *next, *prev;
7103 MUTEX_ENTER(&rx_peerHashTable_lock);
7104 for (prev = peer = *peer_ptr; peer; peer = next) {
7106 code = MUTEX_TRYENTER(&peer->peer_lock);
7107 if ((code) && (peer->refCount == 0)
7108 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7109 struct opr_queue *cursor, *store;
7113 * now know that this peer object is one to be
7114 * removed from the hash table. Once it is removed
7115 * it can't be referenced by other threads.
7116 * Lets remove it first and decrement the struct
7117 * nPeerStructs count.
7119 if (peer == *peer_ptr) {
7125 if (rx_stats_active)
7126 rx_atomic_dec(&rx_stats.nPeerStructs);
7129 * Now if we hold references on 'prev' and 'next'
7130 * we can safely drop the rx_peerHashTable_lock
7131 * while we destroy this 'peer' object.
7137 MUTEX_EXIT(&rx_peerHashTable_lock);
7139 MUTEX_EXIT(&peer->peer_lock);
7140 MUTEX_DESTROY(&peer->peer_lock);
7142 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7143 unsigned int num_funcs;
7144 struct rx_interface_stat *rpc_stat
7145 = opr_queue_Entry(cursor, struct rx_interface_stat,
7150 opr_queue_Remove(&rpc_stat->entry);
7151 opr_queue_Remove(&rpc_stat->entryPeers);
7153 num_funcs = rpc_stat->stats[0].func_total;
7155 sizeof(rx_interface_stat_t) +
7156 rpc_stat->stats[0].func_total *
7157 sizeof(rx_function_entry_v1_t);
7159 rxi_Free(rpc_stat, space);
7161 MUTEX_ENTER(&rx_rpc_stats);
7162 rxi_rpc_peer_stat_cnt -= num_funcs;
7163 MUTEX_EXIT(&rx_rpc_stats);
7168 * Regain the rx_peerHashTable_lock and
7169 * decrement the reference count on 'prev'
7172 MUTEX_ENTER(&rx_peerHashTable_lock);
7179 MUTEX_EXIT(&peer->peer_lock);
7184 MUTEX_EXIT(&rx_peerHashTable_lock);
7188 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7189 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7190 * GC, just below. Really, we shouldn't have to keep moving packets from
7191 * one place to another, but instead ought to always know if we can
7192 * afford to hold onto a packet in its particular use. */
7193 MUTEX_ENTER(&rx_freePktQ_lock);
7194 if (rx_waitingForPackets) {
7195 rx_waitingForPackets = 0;
7196 #ifdef RX_ENABLE_LOCKS
7197 CV_BROADCAST(&rx_waitingForPackets_cv);
7199 osi_rxWakeup(&rx_waitingForPackets);
7202 MUTEX_EXIT(&rx_freePktQ_lock);
7205 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7206 rxevent_Put(rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0));
7210 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7211 * rx.h is sort of strange this is better. This is called with a security
7212 * object before it is discarded. Each connection using a security object has
7213 * its own refcount to the object so it won't actually be freed until the last
7214 * connection is destroyed.
7216 * This is the only rxs module call. A hold could also be written but no one
7220 rxs_Release(struct rx_securityClass *aobj)
7222 return RXS_Close(aobj);
7230 #define TRACE_OPTION_RX_DEBUG 16
7238 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7239 0, KEY_QUERY_VALUE, &parmKey);
7240 if (code != ERROR_SUCCESS)
7243 dummyLen = sizeof(TraceOption);
7244 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7245 (BYTE *) &TraceOption, &dummyLen);
7246 if (code == ERROR_SUCCESS) {
7247 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7249 RegCloseKey (parmKey);
7250 #endif /* AFS_NT40_ENV */
7255 rx_DebugOnOff(int on)
7259 rxdebug_active = on;
7265 rx_StatsOnOff(int on)
7267 rx_stats_active = on;
7271 /* Don't call this debugging routine directly; use dpf */
7273 rxi_DebugPrint(char *format, ...)
7282 va_start(ap, format);
7284 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7287 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7289 OutputDebugString(msg);
7295 va_start(ap, format);
7297 clock_GetTime(&now);
7298 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7299 (unsigned int)now.usec);
7300 vfprintf(rx_Log, format, ap);
7308 * This function is used to process the rx_stats structure that is local
7309 * to a process as well as an rx_stats structure received from a remote
7310 * process (via rxdebug). Therefore, it needs to do minimal version
7314 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7315 afs_int32 freePackets, char version)
7319 if (size != sizeof(struct rx_statistics)) {
7321 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7322 size, sizeof(struct rx_statistics));
7325 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7328 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7329 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7330 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7331 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7332 s->specialPktAllocFailures);
7334 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7335 s->receivePktAllocFailures, s->sendPktAllocFailures,
7336 s->specialPktAllocFailures);
7340 " greedy %u, " "bogusReads %u (last from host %x), "
7341 "noPackets %u, " "noBuffers %u, " "selects %u, "
7342 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7343 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7344 s->selects, s->sendSelects);
7346 fprintf(file, " packets read: ");
7347 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7348 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7350 fprintf(file, "\n");
7353 " other read counters: data %u, " "ack %u, " "dup %u "
7354 "spurious %u " "dally %u\n", s->dataPacketsRead,
7355 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7356 s->ignorePacketDally);
7358 fprintf(file, " packets sent: ");
7359 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7360 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7362 fprintf(file, "\n");
7365 " other send counters: ack %u, " "data %u (not resends), "
7366 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7367 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7368 s->dataPacketsPushed, s->ignoreAckedPacket);
7371 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7372 s->netSendFailures, (int)s->fatalErrors);
7374 if (s->nRttSamples) {
7375 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7376 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7378 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7379 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7383 " %d server connections, " "%d client connections, "
7384 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7385 s->nServerConns, s->nClientConns, s->nPeerStructs,
7386 s->nCallStructs, s->nFreeCallStructs);
7388 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7389 fprintf(file, " %d clock updates\n", clock_nUpdates);
7393 /* for backward compatibility */
7395 rx_PrintStats(FILE * file)
7397 MUTEX_ENTER(&rx_stats_mutex);
7398 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7399 sizeof(rx_stats), rx_nFreePackets,
7401 MUTEX_EXIT(&rx_stats_mutex);
7405 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7407 fprintf(file, "Peer %x.%d.\n",
7408 ntohl(peer->host), (int)ntohs(peer->port));
7411 " Rtt %d, " "total sent %d, " "resent %d\n",
7412 peer->rtt, peer->nSent, peer->reSends);
7414 fprintf(file, " Packet size %d\n", peer->ifMTU);
7418 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7420 * This mutex protects the following static variables:
7424 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7425 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7427 #define LOCK_RX_DEBUG
7428 #define UNLOCK_RX_DEBUG
7429 #endif /* AFS_PTHREAD_ENV */
7431 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7433 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7434 u_char type, void *inputData, size_t inputLength,
7435 void *outputData, size_t outputLength)
7437 static afs_int32 counter = 100;
7438 time_t waitTime, waitCount;
7439 struct rx_header theader;
7442 struct timeval tv_now, tv_wake, tv_delta;
7443 struct sockaddr_in taddr, faddr;
7457 tp = &tbuffer[sizeof(struct rx_header)];
7458 taddr.sin_family = AF_INET;
7459 taddr.sin_port = remotePort;
7460 taddr.sin_addr.s_addr = remoteAddr;
7461 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7462 taddr.sin_len = sizeof(struct sockaddr_in);
7465 memset(&theader, 0, sizeof(theader));
7466 theader.epoch = htonl(999);
7468 theader.callNumber = htonl(counter);
7471 theader.type = type;
7472 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7473 theader.serviceId = 0;
7475 memcpy(tbuffer, &theader, sizeof(theader));
7476 memcpy(tp, inputData, inputLength);
7478 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7479 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7481 /* see if there's a packet available */
7482 gettimeofday(&tv_wake, NULL);
7483 tv_wake.tv_sec += waitTime;
7486 FD_SET(socket, &imask);
7487 tv_delta.tv_sec = tv_wake.tv_sec;
7488 tv_delta.tv_usec = tv_wake.tv_usec;
7489 gettimeofday(&tv_now, NULL);
7491 if (tv_delta.tv_usec < tv_now.tv_usec) {
7493 tv_delta.tv_usec += 1000000;
7496 tv_delta.tv_usec -= tv_now.tv_usec;
7498 if (tv_delta.tv_sec < tv_now.tv_sec) {
7502 tv_delta.tv_sec -= tv_now.tv_sec;
7505 code = select(0, &imask, 0, 0, &tv_delta);
7506 #else /* AFS_NT40_ENV */
7507 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7508 #endif /* AFS_NT40_ENV */
7509 if (code == 1 && FD_ISSET(socket, &imask)) {
7510 /* now receive a packet */
7511 faddrLen = sizeof(struct sockaddr_in);
7513 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7514 (struct sockaddr *)&faddr, &faddrLen);
7517 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7518 if (counter == ntohl(theader.callNumber))
7526 /* see if we've timed out */
7534 code -= sizeof(struct rx_header);
7535 if (code > outputLength)
7536 code = outputLength;
7537 memcpy(outputData, tp, code);
7540 #endif /* RXDEBUG */
7543 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7544 afs_uint16 remotePort, struct rx_debugStats * stat,
7545 afs_uint32 * supportedValues)
7547 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7549 struct rx_debugIn in;
7551 *supportedValues = 0;
7552 in.type = htonl(RX_DEBUGI_GETSTATS);
7555 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7556 &in, sizeof(in), stat, sizeof(*stat));
7559 * If the call was successful, fixup the version and indicate
7560 * what contents of the stat structure are valid.
7561 * Also do net to host conversion of fields here.
7565 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7566 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7568 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7569 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7571 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7572 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7574 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7575 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7577 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7578 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7580 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7581 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7583 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7584 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7586 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7587 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7589 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7590 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7592 stat->nFreePackets = ntohl(stat->nFreePackets);
7593 stat->packetReclaims = ntohl(stat->packetReclaims);
7594 stat->callsExecuted = ntohl(stat->callsExecuted);
7595 stat->nWaiting = ntohl(stat->nWaiting);
7596 stat->idleThreads = ntohl(stat->idleThreads);
7597 stat->nWaited = ntohl(stat->nWaited);
7598 stat->nPackets = ntohl(stat->nPackets);
7607 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7608 afs_uint16 remotePort, struct rx_statistics * stat,
7609 afs_uint32 * supportedValues)
7611 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7613 struct rx_debugIn in;
7614 afs_int32 *lp = (afs_int32 *) stat;
7618 * supportedValues is currently unused, but added to allow future
7619 * versioning of this function.
7622 *supportedValues = 0;
7623 in.type = htonl(RX_DEBUGI_RXSTATS);
7625 memset(stat, 0, sizeof(*stat));
7627 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7628 &in, sizeof(in), stat, sizeof(*stat));
7633 * Do net to host conversion here
7636 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7647 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7648 afs_uint16 remotePort, size_t version_length,
7651 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7653 return MakeDebugCall(socket, remoteAddr, remotePort,
7654 RX_PACKET_TYPE_VERSION, a, 1, version,
7662 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7663 afs_uint16 remotePort, afs_int32 * nextConnection,
7664 int allConnections, afs_uint32 debugSupportedValues,
7665 struct rx_debugConn * conn,
7666 afs_uint32 * supportedValues)
7668 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7670 struct rx_debugIn in;
7674 * supportedValues is currently unused, but added to allow future
7675 * versioning of this function.
7678 *supportedValues = 0;
7679 if (allConnections) {
7680 in.type = htonl(RX_DEBUGI_GETALLCONN);
7682 in.type = htonl(RX_DEBUGI_GETCONN);
7684 in.index = htonl(*nextConnection);
7685 memset(conn, 0, sizeof(*conn));
7687 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7688 &in, sizeof(in), conn, sizeof(*conn));
7691 *nextConnection += 1;
7694 * Convert old connection format to new structure.
7697 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7698 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7699 #define MOVEvL(a) (conn->a = vL->a)
7701 /* any old or unrecognized version... */
7702 for (i = 0; i < RX_MAXCALLS; i++) {
7703 MOVEvL(callState[i]);
7704 MOVEvL(callMode[i]);
7705 MOVEvL(callFlags[i]);
7706 MOVEvL(callOther[i]);
7708 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7709 MOVEvL(secStats.type);
7710 MOVEvL(secStats.level);
7711 MOVEvL(secStats.flags);
7712 MOVEvL(secStats.expires);
7713 MOVEvL(secStats.packetsReceived);
7714 MOVEvL(secStats.packetsSent);
7715 MOVEvL(secStats.bytesReceived);
7716 MOVEvL(secStats.bytesSent);
7721 * Do net to host conversion here
7723 * I don't convert host or port since we are most likely
7724 * going to want these in NBO.
7726 conn->cid = ntohl(conn->cid);
7727 conn->serial = ntohl(conn->serial);
7728 for (i = 0; i < RX_MAXCALLS; i++) {
7729 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7731 conn->error = ntohl(conn->error);
7732 conn->secStats.flags = ntohl(conn->secStats.flags);
7733 conn->secStats.expires = ntohl(conn->secStats.expires);
7734 conn->secStats.packetsReceived =
7735 ntohl(conn->secStats.packetsReceived);
7736 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7737 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7738 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7739 conn->epoch = ntohl(conn->epoch);
7740 conn->natMTU = ntohl(conn->natMTU);
7749 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7750 afs_uint16 remotePort, afs_int32 * nextPeer,
7751 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7752 afs_uint32 * supportedValues)
7754 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7756 struct rx_debugIn in;
7759 * supportedValues is currently unused, but added to allow future
7760 * versioning of this function.
7763 *supportedValues = 0;
7764 in.type = htonl(RX_DEBUGI_GETPEER);
7765 in.index = htonl(*nextPeer);
7766 memset(peer, 0, sizeof(*peer));
7768 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7769 &in, sizeof(in), peer, sizeof(*peer));
7775 * Do net to host conversion here
7777 * I don't convert host or port since we are most likely
7778 * going to want these in NBO.
7780 peer->ifMTU = ntohs(peer->ifMTU);
7781 peer->idleWhen = ntohl(peer->idleWhen);
7782 peer->refCount = ntohs(peer->refCount);
7783 peer->rtt = ntohl(peer->rtt);
7784 peer->rtt_dev = ntohl(peer->rtt_dev);
7785 peer->timeout.sec = 0;
7786 peer->timeout.usec = 0;
7787 peer->nSent = ntohl(peer->nSent);
7788 peer->reSends = ntohl(peer->reSends);
7789 peer->natMTU = ntohs(peer->natMTU);
7790 peer->maxMTU = ntohs(peer->maxMTU);
7791 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7792 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7793 peer->MTU = ntohs(peer->MTU);
7794 peer->cwind = ntohs(peer->cwind);
7795 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7796 peer->congestSeq = ntohs(peer->congestSeq);
7797 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7798 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7799 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7800 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7809 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7810 struct rx_debugPeer * peerStats)
7813 afs_int32 error = 1; /* default to "did not succeed" */
7814 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7816 MUTEX_ENTER(&rx_peerHashTable_lock);
7817 for(tp = rx_peerHashTable[hashValue];
7818 tp != NULL; tp = tp->next) {
7819 if (tp->host == peerHost)
7825 MUTEX_EXIT(&rx_peerHashTable_lock);
7829 MUTEX_ENTER(&tp->peer_lock);
7830 peerStats->host = tp->host;
7831 peerStats->port = tp->port;
7832 peerStats->ifMTU = tp->ifMTU;
7833 peerStats->idleWhen = tp->idleWhen;
7834 peerStats->refCount = tp->refCount;
7835 peerStats->burstSize = 0;
7836 peerStats->burst = 0;
7837 peerStats->burstWait.sec = 0;
7838 peerStats->burstWait.usec = 0;
7839 peerStats->rtt = tp->rtt;
7840 peerStats->rtt_dev = tp->rtt_dev;
7841 peerStats->timeout.sec = 0;
7842 peerStats->timeout.usec = 0;
7843 peerStats->nSent = tp->nSent;
7844 peerStats->reSends = tp->reSends;
7845 peerStats->natMTU = tp->natMTU;
7846 peerStats->maxMTU = tp->maxMTU;
7847 peerStats->maxDgramPackets = tp->maxDgramPackets;
7848 peerStats->ifDgramPackets = tp->ifDgramPackets;
7849 peerStats->MTU = tp->MTU;
7850 peerStats->cwind = tp->cwind;
7851 peerStats->nDgramPackets = tp->nDgramPackets;
7852 peerStats->congestSeq = tp->congestSeq;
7853 peerStats->bytesSent.high = tp->bytesSent >> 32;
7854 peerStats->bytesSent.low = tp->bytesSent & MAX_AFS_UINT32;
7855 peerStats->bytesReceived.high = tp->bytesReceived >> 32;
7856 peerStats->bytesReceived.low
7857 = tp->bytesReceived & MAX_AFS_UINT32;
7858 MUTEX_EXIT(&tp->peer_lock);
7860 MUTEX_ENTER(&rx_peerHashTable_lock);
7863 MUTEX_EXIT(&rx_peerHashTable_lock);
7871 struct rx_serverQueueEntry *np;
7874 struct rx_call *call;
7875 struct rx_serverQueueEntry *sq;
7879 if (rxinit_status == 1) {
7881 return; /* Already shutdown. */
7885 #ifndef AFS_PTHREAD_ENV
7886 FD_ZERO(&rx_selectMask);
7887 #endif /* AFS_PTHREAD_ENV */
7888 rxi_dataQuota = RX_MAX_QUOTA;
7889 #ifndef AFS_PTHREAD_ENV
7891 #endif /* AFS_PTHREAD_ENV */
7894 #ifndef AFS_PTHREAD_ENV
7895 #ifndef AFS_USE_GETTIMEOFDAY
7897 #endif /* AFS_USE_GETTIMEOFDAY */
7898 #endif /* AFS_PTHREAD_ENV */
7900 while (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
7901 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
7902 opr_queue_Remove(&call->entry);
7903 rxi_Free(call, sizeof(struct rx_call));
7906 while (!opr_queue_IsEmpty(&rx_idleServerQueue)) {
7907 sq = opr_queue_First(&rx_idleServerQueue, struct rx_serverQueueEntry,
7909 opr_queue_Remove(&sq->entry);
7914 struct rx_peer **peer_ptr, **peer_end;
7915 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7916 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7918 struct rx_peer *peer, *next;
7920 MUTEX_ENTER(&rx_peerHashTable_lock);
7921 for (peer = *peer_ptr; peer; peer = next) {
7922 struct opr_queue *cursor, *store;
7925 MUTEX_ENTER(&rx_rpc_stats);
7926 MUTEX_ENTER(&peer->peer_lock);
7927 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7928 unsigned int num_funcs;
7929 struct rx_interface_stat *rpc_stat
7930 = opr_queue_Entry(cursor, struct rx_interface_stat,
7934 opr_queue_Remove(&rpc_stat->entry);
7935 opr_queue_Remove(&rpc_stat->entryPeers);
7936 num_funcs = rpc_stat->stats[0].func_total;
7938 sizeof(rx_interface_stat_t) +
7939 rpc_stat->stats[0].func_total *
7940 sizeof(rx_function_entry_v1_t);
7942 rxi_Free(rpc_stat, space);
7944 /* rx_rpc_stats must be held */
7945 rxi_rpc_peer_stat_cnt -= num_funcs;
7947 MUTEX_EXIT(&peer->peer_lock);
7948 MUTEX_EXIT(&rx_rpc_stats);
7952 if (rx_stats_active)
7953 rx_atomic_dec(&rx_stats.nPeerStructs);
7955 MUTEX_EXIT(&rx_peerHashTable_lock);
7958 for (i = 0; i < RX_MAX_SERVICES; i++) {
7960 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7962 for (i = 0; i < rx_hashTableSize; i++) {
7963 struct rx_connection *tc, *ntc;
7964 MUTEX_ENTER(&rx_connHashTable_lock);
7965 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7967 for (j = 0; j < RX_MAXCALLS; j++) {
7969 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7972 rxi_Free(tc, sizeof(*tc));
7974 MUTEX_EXIT(&rx_connHashTable_lock);
7977 MUTEX_ENTER(&freeSQEList_lock);
7979 while ((np = rx_FreeSQEList)) {
7980 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7981 MUTEX_DESTROY(&np->lock);
7982 rxi_Free(np, sizeof(*np));
7985 MUTEX_EXIT(&freeSQEList_lock);
7986 MUTEX_DESTROY(&freeSQEList_lock);
7987 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7988 MUTEX_DESTROY(&rx_connHashTable_lock);
7989 MUTEX_DESTROY(&rx_peerHashTable_lock);
7990 MUTEX_DESTROY(&rx_serverPool_lock);
7992 osi_Free(rx_connHashTable,
7993 rx_hashTableSize * sizeof(struct rx_connection *));
7994 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7996 UNPIN(rx_connHashTable,
7997 rx_hashTableSize * sizeof(struct rx_connection *));
7998 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8000 rxi_FreeAllPackets();
8002 MUTEX_ENTER(&rx_quota_mutex);
8003 rxi_dataQuota = RX_MAX_QUOTA;
8004 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
8005 MUTEX_EXIT(&rx_quota_mutex);
8010 #ifdef RX_ENABLE_LOCKS
8012 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
8014 if (!MUTEX_ISMINE(lockaddr))
8015 osi_Panic("Lock not held: %s", msg);
8017 #endif /* RX_ENABLE_LOCKS */
8022 * Routines to implement connection specific data.
8026 rx_KeyCreate(rx_destructor_t rtn)
8029 MUTEX_ENTER(&rxi_keyCreate_lock);
8030 key = rxi_keyCreate_counter++;
8031 rxi_keyCreate_destructor = (rx_destructor_t *)
8032 realloc((void *)rxi_keyCreate_destructor,
8033 (key + 1) * sizeof(rx_destructor_t));
8034 rxi_keyCreate_destructor[key] = rtn;
8035 MUTEX_EXIT(&rxi_keyCreate_lock);
8040 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
8043 MUTEX_ENTER(&conn->conn_data_lock);
8044 if (!conn->specific) {
8045 conn->specific = malloc((key + 1) * sizeof(void *));
8046 for (i = 0; i < key; i++)
8047 conn->specific[i] = NULL;
8048 conn->nSpecific = key + 1;
8049 conn->specific[key] = ptr;
8050 } else if (key >= conn->nSpecific) {
8051 conn->specific = (void **)
8052 realloc(conn->specific, (key + 1) * sizeof(void *));
8053 for (i = conn->nSpecific; i < key; i++)
8054 conn->specific[i] = NULL;
8055 conn->nSpecific = key + 1;
8056 conn->specific[key] = ptr;
8058 if (conn->specific[key] && rxi_keyCreate_destructor[key])
8059 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
8060 conn->specific[key] = ptr;
8062 MUTEX_EXIT(&conn->conn_data_lock);
8066 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
8069 MUTEX_ENTER(&svc->svc_data_lock);
8070 if (!svc->specific) {
8071 svc->specific = malloc((key + 1) * sizeof(void *));
8072 for (i = 0; i < key; i++)
8073 svc->specific[i] = NULL;
8074 svc->nSpecific = key + 1;
8075 svc->specific[key] = ptr;
8076 } else if (key >= svc->nSpecific) {
8077 svc->specific = (void **)
8078 realloc(svc->specific, (key + 1) * sizeof(void *));
8079 for (i = svc->nSpecific; i < key; i++)
8080 svc->specific[i] = NULL;
8081 svc->nSpecific = key + 1;
8082 svc->specific[key] = ptr;
8084 if (svc->specific[key] && rxi_keyCreate_destructor[key])
8085 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
8086 svc->specific[key] = ptr;
8088 MUTEX_EXIT(&svc->svc_data_lock);
8092 rx_GetSpecific(struct rx_connection *conn, int key)
8095 MUTEX_ENTER(&conn->conn_data_lock);
8096 if (key >= conn->nSpecific)
8099 ptr = conn->specific[key];
8100 MUTEX_EXIT(&conn->conn_data_lock);
8105 rx_GetServiceSpecific(struct rx_service *svc, int key)
8108 MUTEX_ENTER(&svc->svc_data_lock);
8109 if (key >= svc->nSpecific)
8112 ptr = svc->specific[key];
8113 MUTEX_EXIT(&svc->svc_data_lock);
8118 #endif /* !KERNEL */
8121 * processStats is a queue used to store the statistics for the local
8122 * process. Its contents are similar to the contents of the rpcStats
8123 * queue on a rx_peer structure, but the actual data stored within
8124 * this queue contains totals across the lifetime of the process (assuming
8125 * the stats have not been reset) - unlike the per peer structures
8126 * which can come and go based upon the peer lifetime.
8129 static struct opr_queue processStats = { &processStats, &processStats };
8132 * peerStats is a queue used to store the statistics for all peer structs.
8133 * Its contents are the union of all the peer rpcStats queues.
8136 static struct opr_queue peerStats = { &peerStats, &peerStats };
8139 * rxi_monitor_processStats is used to turn process wide stat collection
8143 static int rxi_monitor_processStats = 0;
8146 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8149 static int rxi_monitor_peerStats = 0;
8153 rxi_ClearRPCOpStat(rx_function_entry_v1_p rpc_stat)
8155 rpc_stat->invocations = 0;
8156 rpc_stat->bytes_sent = 0;
8157 rpc_stat->bytes_rcvd = 0;
8158 rpc_stat->queue_time_sum.sec = 0;
8159 rpc_stat->queue_time_sum.usec = 0;
8160 rpc_stat->queue_time_sum_sqr.sec = 0;
8161 rpc_stat->queue_time_sum_sqr.usec = 0;
8162 rpc_stat->queue_time_min.sec = 9999999;
8163 rpc_stat->queue_time_min.usec = 9999999;
8164 rpc_stat->queue_time_max.sec = 0;
8165 rpc_stat->queue_time_max.usec = 0;
8166 rpc_stat->execution_time_sum.sec = 0;
8167 rpc_stat->execution_time_sum.usec = 0;
8168 rpc_stat->execution_time_sum_sqr.sec = 0;
8169 rpc_stat->execution_time_sum_sqr.usec = 0;
8170 rpc_stat->execution_time_min.sec = 9999999;
8171 rpc_stat->execution_time_min.usec = 9999999;
8172 rpc_stat->execution_time_max.sec = 0;
8173 rpc_stat->execution_time_max.usec = 0;
8177 * Given all of the information for a particular rpc
8178 * call, find or create (if requested) the stat structure for the rpc.
8181 * the queue of stats that will be updated with the new value
8183 * @param rxInterface
8184 * a unique number that identifies the rpc interface
8187 * the total number of functions in this interface. this is only
8188 * required if create is true
8191 * if true, this invocation was made to a server
8194 * the ip address of the remote host. this is only required if create
8195 * and addToPeerList are true
8198 * the port of the remote host. this is only required if create
8199 * and addToPeerList are true
8201 * @param addToPeerList
8202 * if != 0, add newly created stat to the global peer list
8205 * if a new stats structure is allocated, the counter will
8206 * be updated with the new number of allocated stat structures.
8207 * only required if create is true
8210 * if no stats structure exists, allocate one
8214 static rx_interface_stat_p
8215 rxi_FindRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8216 afs_uint32 totalFunc, int isServer, afs_uint32 remoteHost,
8217 afs_uint32 remotePort, int addToPeerList,
8218 unsigned int *counter, int create)
8220 rx_interface_stat_p rpc_stat = NULL;
8221 struct opr_queue *cursor;
8224 * See if there's already a structure for this interface
8227 for (opr_queue_Scan(stats, cursor)) {
8228 rpc_stat = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8230 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8231 && (rpc_stat->stats[0].remote_is_server == isServer))
8235 /* if they didn't ask us to create, we're done */
8237 if (opr_queue_IsEnd(stats, cursor))
8243 /* can't proceed without these */
8244 if (!totalFunc || !counter)
8248 * Didn't find a match so allocate a new structure and add it to the
8252 if (opr_queue_IsEnd(stats, cursor) || (rpc_stat == NULL)
8253 || (rpc_stat->stats[0].interfaceId != rxInterface)
8254 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8259 sizeof(rx_interface_stat_t) +
8260 totalFunc * sizeof(rx_function_entry_v1_t);
8262 rpc_stat = rxi_Alloc(space);
8263 if (rpc_stat == NULL)
8266 *counter += totalFunc;
8267 for (i = 0; i < totalFunc; i++) {
8268 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8269 rpc_stat->stats[i].remote_peer = remoteHost;
8270 rpc_stat->stats[i].remote_port = remotePort;
8271 rpc_stat->stats[i].remote_is_server = isServer;
8272 rpc_stat->stats[i].interfaceId = rxInterface;
8273 rpc_stat->stats[i].func_total = totalFunc;
8274 rpc_stat->stats[i].func_index = i;
8276 opr_queue_Prepend(stats, &rpc_stat->entry);
8277 if (addToPeerList) {
8278 opr_queue_Prepend(&peerStats, &rpc_stat->entryPeers);
8285 rx_ClearProcessRPCStats(afs_int32 rxInterface)
8287 rx_interface_stat_p rpc_stat;
8290 if (rxInterface == -1)
8293 MUTEX_ENTER(&rx_rpc_stats);
8294 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8297 totalFunc = rpc_stat->stats[0].func_total;
8298 for (i = 0; i < totalFunc; i++)
8299 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8301 MUTEX_EXIT(&rx_rpc_stats);
8306 rx_ClearPeerRPCStats(afs_int32 rxInterface, afs_uint32 peerHost, afs_uint16 peerPort)
8308 rx_interface_stat_p rpc_stat;
8310 struct rx_peer * peer;
8312 if (rxInterface == -1)
8315 peer = rxi_FindPeer(peerHost, peerPort, 0, 0);
8319 MUTEX_ENTER(&rx_rpc_stats);
8320 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8323 totalFunc = rpc_stat->stats[0].func_total;
8324 for (i = 0; i < totalFunc; i++)
8325 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8327 MUTEX_EXIT(&rx_rpc_stats);
8332 rx_CopyProcessRPCStats(afs_uint64 op)
8334 rx_interface_stat_p rpc_stat;
8335 rx_function_entry_v1_p rpcop_stat =
8336 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8337 int currentFunc = (op & MAX_AFS_UINT32);
8338 afs_int32 rxInterface = (op >> 32);
8340 if (!rxi_monitor_processStats)
8343 if (rxInterface == -1)
8346 if (rpcop_stat == NULL)
8349 MUTEX_ENTER(&rx_rpc_stats);
8350 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8353 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8354 sizeof(rx_function_entry_v1_t));
8355 MUTEX_EXIT(&rx_rpc_stats);
8357 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8364 rx_CopyPeerRPCStats(afs_uint64 op, afs_uint32 peerHost, afs_uint16 peerPort)
8366 rx_interface_stat_p rpc_stat;
8367 rx_function_entry_v1_p rpcop_stat =
8368 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8369 int currentFunc = (op & MAX_AFS_UINT32);
8370 afs_int32 rxInterface = (op >> 32);
8371 struct rx_peer *peer;
8373 if (!rxi_monitor_peerStats)
8376 if (rxInterface == -1)
8379 if (rpcop_stat == NULL)
8382 peer = rxi_FindPeer(peerHost, peerPort, 0, 0);
8386 MUTEX_ENTER(&rx_rpc_stats);
8387 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8390 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8391 sizeof(rx_function_entry_v1_t));
8392 MUTEX_EXIT(&rx_rpc_stats);
8394 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8401 rx_ReleaseRPCStats(void *stats)
8404 rxi_Free(stats, sizeof(rx_function_entry_v1_t));
8408 * Given all of the information for a particular rpc
8409 * call, create (if needed) and update the stat totals for the rpc.
8412 * the queue of stats that will be updated with the new value
8414 * @param rxInterface
8415 * a unique number that identifies the rpc interface
8417 * @param currentFunc
8418 * the index of the function being invoked
8421 * the total number of functions in this interface
8424 * the amount of time this function waited for a thread
8427 * the amount of time this function invocation took to execute
8430 * the number bytes sent by this invocation
8433 * the number bytes received by this invocation
8436 * if true, this invocation was made to a server
8439 * the ip address of the remote host
8442 * the port of the remote host
8444 * @param addToPeerList
8445 * if != 0, add newly created stat to the global peer list
8448 * if a new stats structure is allocated, the counter will
8449 * be updated with the new number of allocated stat structures
8454 rxi_AddRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8455 afs_uint32 currentFunc, afs_uint32 totalFunc,
8456 struct clock *queueTime, struct clock *execTime,
8457 afs_uint64 bytesSent, afs_uint64 bytesRcvd, int isServer,
8458 afs_uint32 remoteHost, afs_uint32 remotePort,
8459 int addToPeerList, unsigned int *counter)
8462 rx_interface_stat_p rpc_stat;
8464 rpc_stat = rxi_FindRpcStat(stats, rxInterface, totalFunc, isServer,
8465 remoteHost, remotePort, addToPeerList, counter,
8473 * Increment the stats for this function
8476 rpc_stat->stats[currentFunc].invocations++;
8477 rpc_stat->stats[currentFunc].bytes_sent += bytesSent;
8478 rpc_stat->stats[currentFunc].bytes_rcvd += bytesRcvd;
8479 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8480 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8481 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8482 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8484 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8485 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8487 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8488 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8490 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8491 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8493 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8494 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8502 rxi_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8503 afs_uint32 currentFunc, afs_uint32 totalFunc,
8504 struct clock *queueTime, struct clock *execTime,
8505 afs_uint64 bytesSent, afs_uint64 bytesRcvd,
8509 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8512 MUTEX_ENTER(&rx_rpc_stats);
8514 if (rxi_monitor_peerStats) {
8515 MUTEX_ENTER(&peer->peer_lock);
8516 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8517 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8518 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8519 MUTEX_EXIT(&peer->peer_lock);
8522 if (rxi_monitor_processStats) {
8523 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8524 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8525 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8528 MUTEX_EXIT(&rx_rpc_stats);
8532 * Increment the times and count for a particular rpc function.
8534 * Traditionally this call was invoked from rxgen stubs. Modern stubs
8535 * call rx_RecordCallStatistics instead, so the public version of this
8536 * function is left purely for legacy callers.
8539 * The peer who invoked the rpc
8541 * @param rxInterface
8542 * A unique number that identifies the rpc interface
8544 * @param currentFunc
8545 * The index of the function being invoked
8548 * The total number of functions in this interface
8551 * The amount of time this function waited for a thread
8554 * The amount of time this function invocation took to execute
8557 * The number bytes sent by this invocation
8560 * The number bytes received by this invocation
8563 * If true, this invocation was made to a server
8567 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8568 afs_uint32 currentFunc, afs_uint32 totalFunc,
8569 struct clock *queueTime, struct clock *execTime,
8570 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8576 sent64 = ((afs_uint64)bytesSent->high << 32) + bytesSent->low;
8577 rcvd64 = ((afs_uint64)bytesRcvd->high << 32) + bytesRcvd->low;
8579 rxi_IncrementTimeAndCount(peer, rxInterface, currentFunc, totalFunc,
8580 queueTime, execTime, sent64, rcvd64,
8587 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8591 * IN callerVersion - the rpc stat version of the caller.
8593 * IN count - the number of entries to marshall.
8595 * IN stats - pointer to stats to be marshalled.
8597 * OUT ptr - Where to store the marshalled data.
8604 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8605 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8611 * We only support the first version
8613 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8614 *(ptr++) = stats->remote_peer;
8615 *(ptr++) = stats->remote_port;
8616 *(ptr++) = stats->remote_is_server;
8617 *(ptr++) = stats->interfaceId;
8618 *(ptr++) = stats->func_total;
8619 *(ptr++) = stats->func_index;
8620 *(ptr++) = stats->invocations >> 32;
8621 *(ptr++) = stats->invocations & MAX_AFS_UINT32;
8622 *(ptr++) = stats->bytes_sent >> 32;
8623 *(ptr++) = stats->bytes_sent & MAX_AFS_UINT32;
8624 *(ptr++) = stats->bytes_rcvd >> 32;
8625 *(ptr++) = stats->bytes_rcvd & MAX_AFS_UINT32;
8626 *(ptr++) = stats->queue_time_sum.sec;
8627 *(ptr++) = stats->queue_time_sum.usec;
8628 *(ptr++) = stats->queue_time_sum_sqr.sec;
8629 *(ptr++) = stats->queue_time_sum_sqr.usec;
8630 *(ptr++) = stats->queue_time_min.sec;
8631 *(ptr++) = stats->queue_time_min.usec;
8632 *(ptr++) = stats->queue_time_max.sec;
8633 *(ptr++) = stats->queue_time_max.usec;
8634 *(ptr++) = stats->execution_time_sum.sec;
8635 *(ptr++) = stats->execution_time_sum.usec;
8636 *(ptr++) = stats->execution_time_sum_sqr.sec;
8637 *(ptr++) = stats->execution_time_sum_sqr.usec;
8638 *(ptr++) = stats->execution_time_min.sec;
8639 *(ptr++) = stats->execution_time_min.usec;
8640 *(ptr++) = stats->execution_time_max.sec;
8641 *(ptr++) = stats->execution_time_max.usec;
8647 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8652 * IN callerVersion - the rpc stat version of the caller
8654 * OUT myVersion - the rpc stat version of this function
8656 * OUT clock_sec - local time seconds
8658 * OUT clock_usec - local time microseconds
8660 * OUT allocSize - the number of bytes allocated to contain stats
8662 * OUT statCount - the number stats retrieved from this process.
8664 * OUT stats - the actual stats retrieved from this process.
8668 * Returns void. If successful, stats will != NULL.
8672 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8673 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8674 size_t * allocSize, afs_uint32 * statCount,
8675 afs_uint32 ** stats)
8685 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8688 * Check to see if stats are enabled
8691 MUTEX_ENTER(&rx_rpc_stats);
8692 if (!rxi_monitor_processStats) {
8693 MUTEX_EXIT(&rx_rpc_stats);
8697 clock_GetTime(&now);
8698 *clock_sec = now.sec;
8699 *clock_usec = now.usec;
8702 * Allocate the space based upon the caller version
8704 * If the client is at an older version than we are,
8705 * we return the statistic data in the older data format, but
8706 * we still return our version number so the client knows we
8707 * are maintaining more data than it can retrieve.
8710 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8711 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8712 *statCount = rxi_rpc_process_stat_cnt;
8715 * This can't happen yet, but in the future version changes
8716 * can be handled by adding additional code here
8720 if (space > (size_t) 0) {
8722 ptr = *stats = rxi_Alloc(space);
8725 struct opr_queue *cursor;
8727 for (opr_queue_Scan(&processStats, cursor)) {
8728 struct rx_interface_stat *rpc_stat =
8729 opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8731 * Copy the data based upon the caller version
8733 rx_MarshallProcessRPCStats(callerVersion,
8734 rpc_stat->stats[0].func_total,
8735 rpc_stat->stats, &ptr);
8741 MUTEX_EXIT(&rx_rpc_stats);
8746 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8750 * IN callerVersion - the rpc stat version of the caller
8752 * OUT myVersion - the rpc stat version of this function
8754 * OUT clock_sec - local time seconds
8756 * OUT clock_usec - local time microseconds
8758 * OUT allocSize - the number of bytes allocated to contain stats
8760 * OUT statCount - the number of stats retrieved from the individual
8763 * OUT stats - the actual stats retrieved from the individual peer structures.
8767 * Returns void. If successful, stats will != NULL.
8771 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8772 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8773 size_t * allocSize, afs_uint32 * statCount,
8774 afs_uint32 ** stats)
8784 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8787 * Check to see if stats are enabled
8790 MUTEX_ENTER(&rx_rpc_stats);
8791 if (!rxi_monitor_peerStats) {
8792 MUTEX_EXIT(&rx_rpc_stats);
8796 clock_GetTime(&now);
8797 *clock_sec = now.sec;
8798 *clock_usec = now.usec;
8801 * Allocate the space based upon the caller version
8803 * If the client is at an older version than we are,
8804 * we return the statistic data in the older data format, but
8805 * we still return our version number so the client knows we
8806 * are maintaining more data than it can retrieve.
8809 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8810 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8811 *statCount = rxi_rpc_peer_stat_cnt;
8814 * This can't happen yet, but in the future version changes
8815 * can be handled by adding additional code here
8819 if (space > (size_t) 0) {
8821 ptr = *stats = rxi_Alloc(space);
8824 struct opr_queue *cursor;
8826 for (opr_queue_Scan(&peerStats, cursor)) {
8827 struct rx_interface_stat *rpc_stat
8828 = opr_queue_Entry(cursor, struct rx_interface_stat,
8832 * Copy the data based upon the caller version
8834 rx_MarshallProcessRPCStats(callerVersion,
8835 rpc_stat->stats[0].func_total,
8836 rpc_stat->stats, &ptr);
8842 MUTEX_EXIT(&rx_rpc_stats);
8847 * rx_FreeRPCStats - free memory allocated by
8848 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8852 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8853 * rx_RetrievePeerRPCStats
8855 * IN allocSize - the number of bytes in stats.
8863 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8865 rxi_Free(stats, allocSize);
8869 * rx_queryProcessRPCStats - see if process rpc stat collection is
8870 * currently enabled.
8876 * Returns 0 if stats are not enabled != 0 otherwise
8880 rx_queryProcessRPCStats(void)
8883 MUTEX_ENTER(&rx_rpc_stats);
8884 rc = rxi_monitor_processStats;
8885 MUTEX_EXIT(&rx_rpc_stats);
8890 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8896 * Returns 0 if stats are not enabled != 0 otherwise
8900 rx_queryPeerRPCStats(void)
8903 MUTEX_ENTER(&rx_rpc_stats);
8904 rc = rxi_monitor_peerStats;
8905 MUTEX_EXIT(&rx_rpc_stats);
8910 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8920 rx_enableProcessRPCStats(void)
8922 MUTEX_ENTER(&rx_rpc_stats);
8923 rx_enable_stats = 1;
8924 rxi_monitor_processStats = 1;
8925 MUTEX_EXIT(&rx_rpc_stats);
8929 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8939 rx_enablePeerRPCStats(void)
8941 MUTEX_ENTER(&rx_rpc_stats);
8942 rx_enable_stats = 1;
8943 rxi_monitor_peerStats = 1;
8944 MUTEX_EXIT(&rx_rpc_stats);
8948 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8958 rx_disableProcessRPCStats(void)
8960 struct opr_queue *cursor, *store;
8963 MUTEX_ENTER(&rx_rpc_stats);
8966 * Turn off process statistics and if peer stats is also off, turn
8970 rxi_monitor_processStats = 0;
8971 if (rxi_monitor_peerStats == 0) {
8972 rx_enable_stats = 0;
8975 for (opr_queue_ScanSafe(&processStats, cursor, store)) {
8976 unsigned int num_funcs = 0;
8977 struct rx_interface_stat *rpc_stat
8978 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8980 opr_queue_Remove(&rpc_stat->entry);
8982 num_funcs = rpc_stat->stats[0].func_total;
8984 sizeof(rx_interface_stat_t) +
8985 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8987 rxi_Free(rpc_stat, space);
8988 rxi_rpc_process_stat_cnt -= num_funcs;
8990 MUTEX_EXIT(&rx_rpc_stats);
8994 * rx_disablePeerRPCStats - stop rpc stat collection for peers
9004 rx_disablePeerRPCStats(void)
9006 struct rx_peer **peer_ptr, **peer_end;
9010 * Turn off peer statistics and if process stats is also off, turn
9014 rxi_monitor_peerStats = 0;
9015 if (rxi_monitor_processStats == 0) {
9016 rx_enable_stats = 0;
9019 for (peer_ptr = &rx_peerHashTable[0], peer_end =
9020 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
9022 struct rx_peer *peer, *next, *prev;
9024 MUTEX_ENTER(&rx_peerHashTable_lock);
9025 MUTEX_ENTER(&rx_rpc_stats);
9026 for (prev = peer = *peer_ptr; peer; peer = next) {
9028 code = MUTEX_TRYENTER(&peer->peer_lock);
9031 struct opr_queue *cursor, *store;
9033 if (prev == *peer_ptr) {
9044 MUTEX_EXIT(&rx_peerHashTable_lock);
9046 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
9047 unsigned int num_funcs = 0;
9048 struct rx_interface_stat *rpc_stat
9049 = opr_queue_Entry(cursor, struct rx_interface_stat,
9052 opr_queue_Remove(&rpc_stat->entry);
9053 opr_queue_Remove(&rpc_stat->entryPeers);
9054 num_funcs = rpc_stat->stats[0].func_total;
9056 sizeof(rx_interface_stat_t) +
9057 rpc_stat->stats[0].func_total *
9058 sizeof(rx_function_entry_v1_t);
9060 rxi_Free(rpc_stat, space);
9061 rxi_rpc_peer_stat_cnt -= num_funcs;
9063 MUTEX_EXIT(&peer->peer_lock);
9065 MUTEX_ENTER(&rx_peerHashTable_lock);
9075 MUTEX_EXIT(&rx_rpc_stats);
9076 MUTEX_EXIT(&rx_peerHashTable_lock);
9081 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
9086 * IN clearFlag - flag indicating which stats to clear
9094 rx_clearProcessRPCStats(afs_uint32 clearFlag)
9096 struct opr_queue *cursor;
9098 MUTEX_ENTER(&rx_rpc_stats);
9100 for (opr_queue_Scan(&processStats, cursor)) {
9101 unsigned int num_funcs = 0, i;
9102 struct rx_interface_stat *rpc_stat
9103 = opr_queue_Entry(rpc_stat, struct rx_interface_stat, entry);
9105 num_funcs = rpc_stat->stats[0].func_total;
9106 for (i = 0; i < num_funcs; i++) {
9107 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9108 rpc_stat->stats[i].invocations = 0;
9110 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9111 rpc_stat->stats[i].bytes_sent = 0;
9113 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9114 rpc_stat->stats[i].bytes_rcvd = 0;
9116 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9117 rpc_stat->stats[i].queue_time_sum.sec = 0;
9118 rpc_stat->stats[i].queue_time_sum.usec = 0;
9120 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9121 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9122 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9124 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9125 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9126 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9128 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9129 rpc_stat->stats[i].queue_time_max.sec = 0;
9130 rpc_stat->stats[i].queue_time_max.usec = 0;
9132 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9133 rpc_stat->stats[i].execution_time_sum.sec = 0;
9134 rpc_stat->stats[i].execution_time_sum.usec = 0;
9136 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9137 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9138 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9140 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9141 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9142 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9144 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9145 rpc_stat->stats[i].execution_time_max.sec = 0;
9146 rpc_stat->stats[i].execution_time_max.usec = 0;
9151 MUTEX_EXIT(&rx_rpc_stats);
9155 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
9160 * IN clearFlag - flag indicating which stats to clear
9168 rx_clearPeerRPCStats(afs_uint32 clearFlag)
9170 struct opr_queue *cursor;
9172 MUTEX_ENTER(&rx_rpc_stats);
9174 for (opr_queue_Scan(&peerStats, cursor)) {
9175 unsigned int num_funcs, i;
9176 struct rx_interface_stat *rpc_stat
9177 = opr_queue_Entry(cursor, struct rx_interface_stat, entryPeers);
9179 num_funcs = rpc_stat->stats[0].func_total;
9180 for (i = 0; i < num_funcs; i++) {
9181 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9182 rpc_stat->stats[i].invocations = 0;
9184 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9185 rpc_stat->stats[i].bytes_sent = 0;
9187 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9188 rpc_stat->stats[i].bytes_rcvd = 0;
9190 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9191 rpc_stat->stats[i].queue_time_sum.sec = 0;
9192 rpc_stat->stats[i].queue_time_sum.usec = 0;
9194 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9195 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9196 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9198 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9199 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9200 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9202 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9203 rpc_stat->stats[i].queue_time_max.sec = 0;
9204 rpc_stat->stats[i].queue_time_max.usec = 0;
9206 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9207 rpc_stat->stats[i].execution_time_sum.sec = 0;
9208 rpc_stat->stats[i].execution_time_sum.usec = 0;
9210 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9211 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9212 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9214 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9215 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9216 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9218 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9219 rpc_stat->stats[i].execution_time_max.sec = 0;
9220 rpc_stat->stats[i].execution_time_max.usec = 0;
9225 MUTEX_EXIT(&rx_rpc_stats);
9229 * rxi_rxstat_userok points to a routine that returns 1 if the caller
9230 * is authorized to enable/disable/clear RX statistics.
9232 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
9235 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
9237 rxi_rxstat_userok = proc;
9241 rx_RxStatUserOk(struct rx_call *call)
9243 if (!rxi_rxstat_userok)
9245 return rxi_rxstat_userok(call);
9250 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
9251 * function in the MSVC runtime DLL (msvcrt.dll).
9253 * Note: the system serializes calls to this function.
9256 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
9257 DWORD reason, /* reason function is being called */
9258 LPVOID reserved) /* reserved for future use */
9261 case DLL_PROCESS_ATTACH:
9262 /* library is being attached to a process */
9266 case DLL_PROCESS_DETACH:
9273 #endif /* AFS_NT40_ENV */
9276 int rx_DumpCalls(FILE *outputFile, char *cookie)
9278 #ifdef RXDEBUG_PACKET
9279 #ifdef KDUMP_RX_LOCK
9280 struct rx_call_rx_lock *c;
9287 #define RXDPRINTF sprintf
9288 #define RXDPRINTOUT output
9290 #define RXDPRINTF fprintf
9291 #define RXDPRINTOUT outputFile
9294 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9296 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9299 for (c = rx_allCallsp; c; c = c->allNextp) {
9300 u_short rqc, tqc, iovqc;
9302 MUTEX_ENTER(&c->lock);
9303 rqc = opr_queue_Count(&c->rq);
9304 tqc = opr_queue_Count(&c->tq);
9305 iovqc = opr_queue_Count(&c->app.iovq);
9307 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, "
9308 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9309 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9310 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9311 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
9312 #ifdef RX_ENABLE_LOCKS
9315 #ifdef RX_REFCOUNT_CHECK
9316 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9317 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9320 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,
9321 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9322 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9323 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9324 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9325 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
9326 #ifdef RX_ENABLE_LOCKS
9327 , (afs_uint32)c->refCount
9329 #ifdef RX_REFCOUNT_CHECK
9330 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9333 MUTEX_EXIT(&c->lock);
9336 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9339 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9341 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9343 #endif /* RXDEBUG_PACKET */