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>
76 #include <hcrypto/rand.h>
80 #include "rx_atomic.h"
81 #include "rx_globals.h"
83 #include "rx_internal.h"
90 #include "rx_packet.h"
91 #include "rx_server.h"
93 #include <afs/rxgen_consts.h>
96 #ifdef AFS_PTHREAD_ENV
98 int (*registerProgram) (pid_t, char *) = 0;
99 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
102 int (*registerProgram) (PROCESS, char *) = 0;
103 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
107 /* Local static routines */
108 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
109 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
110 struct rx_call *, struct rx_peer *,
112 static void rxi_Resend(struct rxevent *event, void *arg0, void *arg1,
114 static void rxi_SendDelayedAck(struct rxevent *event, void *call,
115 void *dummy, int dummy2);
116 static void rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1,
117 void *dummy, int dummy2);
118 static void rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1,
119 void *unused, int unused2);
120 static void rxi_ReapConnections(struct rxevent *unused, void *unused1,
121 void *unused2, int unused3);
122 static struct rx_packet *rxi_SendCallAbort(struct rx_call *call,
123 struct rx_packet *packet,
124 int istack, int force);
125 static void rxi_AckAll(struct rx_call *call);
126 static struct rx_connection
127 *rxi_FindConnection(osi_socket socket, afs_uint32 host, u_short port,
128 u_short serviceId, afs_uint32 cid,
129 afs_uint32 epoch, int type, u_int securityIndex,
130 int *unknownService);
131 static struct rx_packet
132 *rxi_ReceiveDataPacket(struct rx_call *call, struct rx_packet *np,
133 int istack, osi_socket socket,
134 afs_uint32 host, u_short port, int *tnop,
135 struct rx_call **newcallp);
136 static struct rx_packet
137 *rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
139 static struct rx_packet
140 *rxi_ReceiveResponsePacket(struct rx_connection *conn,
141 struct rx_packet *np, int istack);
142 static struct rx_packet
143 *rxi_ReceiveChallengePacket(struct rx_connection *conn,
144 struct rx_packet *np, int istack);
145 static void rxi_AttachServerProc(struct rx_call *call, osi_socket socket,
146 int *tnop, struct rx_call **newcallp);
147 static void rxi_ClearTransmitQueue(struct rx_call *call, int force);
148 static void rxi_ClearReceiveQueue(struct rx_call *call);
149 static void rxi_ResetCall(struct rx_call *call, int newcall);
150 static void rxi_ScheduleKeepAliveEvent(struct rx_call *call);
151 static void rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn);
152 static void rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs);
153 static void rxi_KeepAliveOn(struct rx_call *call);
154 static void rxi_GrowMTUOn(struct rx_call *call);
155 static void rxi_ChallengeOn(struct rx_connection *conn);
156 static int rxi_CheckCall(struct rx_call *call, int haveCTLock);
157 static void rxi_AckAllInTransmitQueue(struct rx_call *call);
158 static void rxi_CancelKeepAliveEvent(struct rx_call *call);
159 static void rxi_CancelDelayedAbortEvent(struct rx_call *call);
160 static void rxi_CancelGrowMTUEvent(struct rx_call *call);
161 static void update_nextCid(void);
163 #ifdef RX_ENABLE_LOCKS
165 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
166 rx_atomic_t rxi_start_in_error;
168 #endif /* RX_ENABLE_LOCKS */
170 /* Constant delay time before sending an acknowledge of the last packet
171 * received. This is to avoid sending an extra acknowledge when the
172 * client is about to make another call, anyway, or the server is
175 * The lastAckDelay may not exceeed 400ms without causing peers to
176 * unecessarily timeout.
178 struct clock rx_lastAckDelay = {0, 400000};
180 /* Constant delay time before sending a soft ack when none was requested.
181 * This is to make sure we send soft acks before the sender times out,
182 * Normally we wait and send a hard ack when the receiver consumes the packet
184 * This value has been 100ms in all shipping versions of OpenAFS. Changing it
185 * will require changes to the peer's RTT calculations.
187 struct clock rx_softAckDelay = {0, 100000};
190 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
191 * currently allocated within rx. This number is used to allocate the
192 * memory required to return the statistics when queried.
193 * Protected by the rx_rpc_stats mutex.
196 static unsigned int rxi_rpc_peer_stat_cnt;
199 * rxi_rpc_process_stat_cnt counts the total number of local process stat
200 * structures currently allocated within rx. The number is used to allocate
201 * the memory required to return the statistics when queried.
202 * Protected by the rx_rpc_stats mutex.
205 static unsigned int rxi_rpc_process_stat_cnt;
208 * rxi_busyChannelError is a boolean. It indicates whether or not RX_CALL_BUSY
209 * errors should be reported to the application when a call channel appears busy
210 * (inferred from the receipt of RX_PACKET_TYPE_BUSY packets on the channel),
211 * and there are other call channels in the connection that are not busy.
212 * If 0, we do not return errors upon receiving busy packets; we just keep
213 * trying on the same call channel until we hit a timeout.
215 static afs_int32 rxi_busyChannelError = 0;
217 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
218 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
220 /* Incoming calls wait on this queue when there are no available
221 * server processes */
222 struct opr_queue rx_incomingCallQueue;
224 /* Server processes wait on this queue when there are no appropriate
225 * calls to process */
226 struct opr_queue rx_idleServerQueue;
228 #if !defined(offsetof)
229 #include <stddef.h> /* for definition of offsetof() */
232 #ifdef RX_ENABLE_LOCKS
233 afs_kmutex_t rx_atomic_mutex;
236 /* Forward prototypes */
237 static struct rx_call * rxi_NewCall(struct rx_connection *, int);
240 putConnection (struct rx_connection *conn) {
241 MUTEX_ENTER(&rx_refcnt_mutex);
243 MUTEX_EXIT(&rx_refcnt_mutex);
246 #ifdef AFS_PTHREAD_ENV
249 * Use procedural initialization of mutexes/condition variables
253 extern afs_kmutex_t rx_quota_mutex;
254 extern afs_kmutex_t rx_pthread_mutex;
255 extern afs_kmutex_t rx_packets_mutex;
256 extern afs_kmutex_t rx_refcnt_mutex;
257 extern afs_kmutex_t des_init_mutex;
258 extern afs_kmutex_t des_random_mutex;
260 extern afs_kmutex_t rx_clock_mutex;
261 extern afs_kmutex_t rxi_connCacheMutex;
262 extern afs_kmutex_t event_handler_mutex;
263 extern afs_kmutex_t listener_mutex;
264 extern afs_kmutex_t rx_if_init_mutex;
265 extern afs_kmutex_t rx_if_mutex;
267 extern afs_kcondvar_t rx_event_handler_cond;
268 extern afs_kcondvar_t rx_listener_cond;
271 static afs_kmutex_t epoch_mutex;
272 static afs_kmutex_t rx_init_mutex;
273 static afs_kmutex_t rx_debug_mutex;
274 static afs_kmutex_t rx_rpc_stats;
277 rxi_InitPthread(void)
279 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
280 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
281 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
282 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
284 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
285 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
286 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
287 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
288 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
289 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
291 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
292 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
293 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
294 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
295 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
298 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
299 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
302 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
303 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
305 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
306 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
307 #ifdef RX_ENABLE_LOCKS
310 #endif /* RX_LOCKS_DB */
311 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
312 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
314 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
316 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
318 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
320 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
322 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
324 #endif /* RX_ENABLE_LOCKS */
327 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
328 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
330 * The rx_stats_mutex mutex protects the following global variables:
331 * rxi_lowConnRefCount
332 * rxi_lowPeerRefCount
341 * The rx_quota_mutex mutex protects the following global variables:
349 * The rx_freePktQ_lock protects the following global variables:
354 * The rx_packets_mutex mutex protects the following global variables:
362 * The rx_pthread_mutex mutex protects the following global variables:
363 * rxi_fcfs_thread_num
366 #define INIT_PTHREAD_LOCKS
370 /* Variables for handling the minProcs implementation. availProcs gives the
371 * number of threads available in the pool at this moment (not counting dudes
372 * executing right now). totalMin gives the total number of procs required
373 * for handling all minProcs requests. minDeficit is a dynamic variable
374 * tracking the # of procs required to satisfy all of the remaining minProcs
376 * For fine grain locking to work, the quota check and the reservation of
377 * a server thread has to come while rxi_availProcs and rxi_minDeficit
378 * are locked. To this end, the code has been modified under #ifdef
379 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
380 * same time. A new function, ReturnToServerPool() returns the allocation.
382 * A call can be on several queue's (but only one at a time). When
383 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
384 * that no one else is touching the queue. To this end, we store the address
385 * of the queue lock in the call structure (under the call lock) when we
386 * put the call on a queue, and we clear the call_queue_lock when the
387 * call is removed from a queue (once the call lock has been obtained).
388 * This allows rxi_ResetCall to safely synchronize with others wishing
389 * to manipulate the queue.
392 #if defined(RX_ENABLE_LOCKS)
393 static afs_kmutex_t rx_rpc_stats;
396 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
397 ** pretty good that the next packet coming in is from the same connection
398 ** as the last packet, since we're send multiple packets in a transmit window.
400 struct rx_connection *rxLastConn = 0;
402 #ifdef RX_ENABLE_LOCKS
403 /* The locking hierarchy for rx fine grain locking is composed of these
406 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
407 * also protects updates to rx_nextCid
408 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
409 * call->lock - locks call data fields.
410 * These are independent of each other:
411 * rx_freeCallQueue_lock
416 * serverQueueEntry->lock
417 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
419 * peer->lock - locks peer data fields.
420 * conn_data_lock - that more than one thread is not updating a conn data
421 * field at the same time.
432 * Do we need a lock to protect the peer field in the conn structure?
433 * conn->peer was previously a constant for all intents and so has no
434 * lock protecting this field. The multihomed client delta introduced
435 * a RX code change : change the peer field in the connection structure
436 * to that remote interface from which the last packet for this
437 * connection was sent out. This may become an issue if further changes
440 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
441 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
443 /* rxdb_fileID is used to identify the lock location, along with line#. */
444 static int rxdb_fileID = RXDB_FILE_RX;
445 #endif /* RX_LOCKS_DB */
446 #else /* RX_ENABLE_LOCKS */
447 #define SET_CALL_QUEUE_LOCK(C, L)
448 #define CLEAR_CALL_QUEUE_LOCK(C)
449 #endif /* RX_ENABLE_LOCKS */
450 struct rx_serverQueueEntry *rx_waitForPacket = 0;
452 /* ------------Exported Interfaces------------- */
454 /* Initialize rx. A port number may be mentioned, in which case this
455 * becomes the default port number for any service installed later.
456 * If 0 is provided for the port number, a random port will be chosen
457 * by the kernel. Whether this will ever overlap anything in
458 * /etc/services is anybody's guess... Returns 0 on success, -1 on
463 rx_atomic_t rxinit_status = RX_ATOMIC_INIT(1);
466 rx_InitHost(u_int host, u_int port)
473 char *htable, *ptable;
478 if (!rx_atomic_test_and_clear_bit(&rxinit_status, 0))
479 return 0; /* already started */
485 if (afs_winsockInit() < 0)
491 * Initialize anything necessary to provide a non-premptive threading
494 rxi_InitializeThreadSupport();
497 /* Allocate and initialize a socket for client and perhaps server
500 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
501 if (rx_socket == OSI_NULLSOCKET) {
504 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
507 #endif /* RX_LOCKS_DB */
508 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
509 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
510 MUTEX_INIT(&rx_atomic_mutex, "rx_atomic_mutex", MUTEX_DEFAULT, 0);
511 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
512 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
513 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
514 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
515 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
516 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
517 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
519 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
521 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
523 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
525 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
526 #if defined(AFS_HPUX110_ENV)
528 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
529 #endif /* AFS_HPUX110_ENV */
530 #endif /* RX_ENABLE_LOCKS && KERNEL */
533 rx_connDeadTime = 12;
534 rx_tranquil = 0; /* reset flag */
535 rxi_ResetStatistics();
536 htable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
537 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
538 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
539 ptable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
540 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
541 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
543 /* Malloc up a bunch of packets & buffers */
545 opr_queue_Init(&rx_freePacketQueue);
546 rxi_NeedMorePackets = FALSE;
547 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
549 /* enforce a minimum number of allocated packets */
550 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
551 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
553 /* allocate the initial free packet pool */
554 #ifdef RX_ENABLE_TSFPQ
555 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
556 #else /* RX_ENABLE_TSFPQ */
557 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
558 #endif /* RX_ENABLE_TSFPQ */
565 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
566 tv.tv_sec = clock_now.sec;
567 tv.tv_usec = clock_now.usec;
568 srand((unsigned int)tv.tv_usec);
575 #if defined(KERNEL) && !defined(UKERNEL)
576 /* Really, this should never happen in a real kernel */
579 struct sockaddr_in addr;
581 int addrlen = sizeof(addr);
583 socklen_t addrlen = sizeof(addr);
585 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
587 osi_Free(htable, rx_hashTableSize * sizeof(struct rx_connection *));
590 rx_port = addr.sin_port;
593 rx_stats.minRtt.sec = 9999999;
594 if (RAND_bytes(&rx_epoch, sizeof(rx_epoch)) != 1)
596 rx_epoch = (rx_epoch & ~0x40000000) | 0x80000000;
597 if (RAND_bytes(&rx_nextCid, sizeof(rx_nextCid)) != 1)
599 rx_nextCid &= RX_CIDMASK;
600 MUTEX_ENTER(&rx_quota_mutex);
601 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
602 MUTEX_EXIT(&rx_quota_mutex);
603 /* *Slightly* random start time for the cid. This is just to help
604 * out with the hashing function at the peer */
605 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
606 rx_connHashTable = (struct rx_connection **)htable;
607 rx_peerHashTable = (struct rx_peer **)ptable;
609 rx_hardAckDelay.sec = 0;
610 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
612 rxevent_Init(20, rxi_ReScheduleEvents);
614 /* Initialize various global queues */
615 opr_queue_Init(&rx_idleServerQueue);
616 opr_queue_Init(&rx_incomingCallQueue);
617 opr_queue_Init(&rx_freeCallQueue);
619 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
620 /* Initialize our list of usable IP addresses. */
624 /* Start listener process (exact function is dependent on the
625 * implementation environment--kernel or user space) */
629 rx_atomic_clear_bit(&rxinit_status, 0);
636 return rx_InitHost(htonl(INADDR_ANY), port);
642 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
643 * maintaing the round trip timer.
648 * Start a new RTT timer for a given call and packet.
650 * There must be no resendEvent already listed for this call, otherwise this
651 * will leak events - intended for internal use within the RTO code only
654 * the RX call to start the timer for
655 * @param[in] lastPacket
656 * a flag indicating whether the last packet has been sent or not
658 * @pre call must be locked before calling this function
662 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
664 struct clock now, retryTime;
669 clock_Add(&retryTime, &call->rto);
671 /* If we're sending the last packet, and we're the client, then the server
672 * may wait for an additional 400ms before returning the ACK, wait for it
673 * rather than hitting a timeout */
674 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
675 clock_Addmsec(&retryTime, 400);
677 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
678 call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
683 * Cancel an RTT timer for a given call.
687 * the RX call to cancel the timer for
689 * @pre call must be locked before calling this function
694 rxi_rto_cancel(struct rx_call *call)
696 if (call->resendEvent != NULL) {
697 rxevent_Cancel(&call->resendEvent);
698 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
703 * Tell the RTO timer that we have sent a packet.
705 * If the timer isn't already running, then start it. If the timer is running,
709 * the RX call that the packet has been sent on
710 * @param[in] lastPacket
711 * A flag which is true if this is the last packet for the call
713 * @pre The call must be locked before calling this function
718 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
720 if (call->resendEvent)
723 rxi_rto_startTimer(call, lastPacket, istack);
727 * Tell the RTO timer that we have received an new ACK message
729 * This function should be called whenever a call receives an ACK that
730 * acknowledges new packets. Whatever happens, we stop the current timer.
731 * If there are unacked packets in the queue which have been sent, then
732 * we restart the timer from now. Otherwise, we leave it stopped.
735 * the RX call that the ACK has been received on
739 rxi_rto_packet_acked(struct rx_call *call, int istack)
741 struct opr_queue *cursor;
743 rxi_rto_cancel(call);
745 if (opr_queue_IsEmpty(&call->tq))
748 for (opr_queue_Scan(&call->tq, cursor)) {
749 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
750 if (p->header.seq > call->tfirst + call->twind)
753 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
754 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
762 * Set an initial round trip timeout for a peer connection
764 * @param[in] secs The timeout to set in seconds
768 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
769 peer->rtt = secs * 8000;
773 * Enables or disables the busy call channel error (RX_CALL_BUSY).
775 * @param[in] onoff Non-zero to enable busy call channel errors.
777 * @pre Neither rx_Init nor rx_InitHost have been called yet
780 rx_SetBusyChannelError(afs_int32 onoff)
782 osi_Assert(rx_atomic_test_bit(&rxinit_status, 0));
783 rxi_busyChannelError = onoff ? 1 : 0;
787 * Set a delayed ack event on the specified call for the given time
789 * @param[in] call - the call on which to set the event
790 * @param[in] offset - the delay from now after which the event fires
793 rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
795 struct clock now, when;
799 clock_Add(&when, offset);
801 if (call->delayedAckEvent && clock_Gt(&call->delayedAckTime, &when)) {
802 /* The event we're cancelling already has a reference, so we don't
804 rxevent_Cancel(&call->delayedAckEvent);
805 call->delayedAckEvent = rxevent_Post(&when, &now, rxi_SendDelayedAck,
808 call->delayedAckTime = when;
809 } else if (!call->delayedAckEvent) {
810 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
811 call->delayedAckEvent = rxevent_Post(&when, &now,
814 call->delayedAckTime = when;
819 rxi_CancelDelayedAckEvent(struct rx_call *call)
821 if (call->delayedAckEvent) {
822 rxevent_Cancel(&call->delayedAckEvent);
823 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
827 /* called with unincremented nRequestsRunning to see if it is OK to start
828 * a new thread in this service. Could be "no" for two reasons: over the
829 * max quota, or would prevent others from reaching their min quota.
831 #ifdef RX_ENABLE_LOCKS
832 /* This verion of QuotaOK reserves quota if it's ok while the
833 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
836 QuotaOK(struct rx_service *aservice)
838 /* check if over max quota */
839 if (aservice->nRequestsRunning >= aservice->maxProcs) {
843 /* under min quota, we're OK */
844 /* otherwise, can use only if there are enough to allow everyone
845 * to go to their min quota after this guy starts.
848 MUTEX_ENTER(&rx_quota_mutex);
849 if ((aservice->nRequestsRunning < aservice->minProcs)
850 || (rxi_availProcs > rxi_minDeficit)) {
851 aservice->nRequestsRunning++;
852 /* just started call in minProcs pool, need fewer to maintain
854 if (aservice->nRequestsRunning <= aservice->minProcs)
857 MUTEX_EXIT(&rx_quota_mutex);
860 MUTEX_EXIT(&rx_quota_mutex);
866 ReturnToServerPool(struct rx_service *aservice)
868 aservice->nRequestsRunning--;
869 MUTEX_ENTER(&rx_quota_mutex);
870 if (aservice->nRequestsRunning < aservice->minProcs)
873 MUTEX_EXIT(&rx_quota_mutex);
876 #else /* RX_ENABLE_LOCKS */
878 QuotaOK(struct rx_service *aservice)
881 /* under min quota, we're OK */
882 if (aservice->nRequestsRunning < aservice->minProcs)
885 /* check if over max quota */
886 if (aservice->nRequestsRunning >= aservice->maxProcs)
889 /* otherwise, can use only if there are enough to allow everyone
890 * to go to their min quota after this guy starts.
892 MUTEX_ENTER(&rx_quota_mutex);
893 if (rxi_availProcs > rxi_minDeficit)
895 MUTEX_EXIT(&rx_quota_mutex);
898 #endif /* RX_ENABLE_LOCKS */
901 /* Called by rx_StartServer to start up lwp's to service calls.
902 NExistingProcs gives the number of procs already existing, and which
903 therefore needn't be created. */
905 rxi_StartServerProcs(int nExistingProcs)
907 struct rx_service *service;
912 /* For each service, reserve N processes, where N is the "minimum"
913 * number of processes that MUST be able to execute a request in parallel,
914 * at any time, for that process. Also compute the maximum difference
915 * between any service's maximum number of processes that can run
916 * (i.e. the maximum number that ever will be run, and a guarantee
917 * that this number will run if other services aren't running), and its
918 * minimum number. The result is the extra number of processes that
919 * we need in order to provide the latter guarantee */
920 for (i = 0; i < RX_MAX_SERVICES; i++) {
922 service = rx_services[i];
923 if (service == (struct rx_service *)0)
925 nProcs += service->minProcs;
926 diff = service->maxProcs - service->minProcs;
930 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
931 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
932 for (i = 0; i < nProcs; i++) {
933 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
939 /* This routine is only required on Windows */
941 rx_StartClientThread(void)
943 #ifdef AFS_PTHREAD_ENV
945 pid = pthread_self();
946 #endif /* AFS_PTHREAD_ENV */
948 #endif /* AFS_NT40_ENV */
950 /* This routine must be called if any services are exported. If the
951 * donateMe flag is set, the calling process is donated to the server
954 rx_StartServer(int donateMe)
956 struct rx_service *service;
962 /* Start server processes, if necessary (exact function is dependent
963 * on the implementation environment--kernel or user space). DonateMe
964 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
965 * case, one less new proc will be created rx_StartServerProcs.
967 rxi_StartServerProcs(donateMe);
969 /* count up the # of threads in minProcs, and add set the min deficit to
970 * be that value, too.
972 for (i = 0; i < RX_MAX_SERVICES; i++) {
973 service = rx_services[i];
974 if (service == (struct rx_service *)0)
976 MUTEX_ENTER(&rx_quota_mutex);
977 rxi_totalMin += service->minProcs;
978 /* below works even if a thread is running, since minDeficit would
979 * still have been decremented and later re-incremented.
981 rxi_minDeficit += service->minProcs;
982 MUTEX_EXIT(&rx_quota_mutex);
985 /* Turn on reaping of idle server connections */
986 rxi_ReapConnections(NULL, NULL, NULL, 0);
995 #ifdef AFS_PTHREAD_ENV
997 pid = afs_pointer_to_int(pthread_self());
998 #else /* AFS_PTHREAD_ENV */
1000 LWP_CurrentProcess(&pid);
1001 #endif /* AFS_PTHREAD_ENV */
1003 sprintf(name, "srv_%d", ++nProcs);
1004 if (registerProgram)
1005 (*registerProgram) (pid, name);
1007 #endif /* AFS_NT40_ENV */
1008 rx_ServerProc(NULL); /* Never returns */
1010 #ifdef RX_ENABLE_TSFPQ
1011 /* no use leaving packets around in this thread's local queue if
1012 * it isn't getting donated to the server thread pool.
1014 rxi_FlushLocalPacketsTSFPQ();
1015 #endif /* RX_ENABLE_TSFPQ */
1019 /* Create a new client connection to the specified service, using the
1020 * specified security object to implement the security model for this
1022 struct rx_connection *
1023 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
1024 struct rx_securityClass *securityObject,
1025 int serviceSecurityIndex)
1028 struct rx_connection *conn;
1033 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
1034 "serviceSecurityIndex %d)\n",
1035 ntohl(shost), ntohs(sport), sservice, securityObject,
1036 serviceSecurityIndex));
1038 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
1039 * the case of kmem_alloc? */
1040 conn = rxi_AllocConnection();
1041 #ifdef RX_ENABLE_LOCKS
1042 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
1043 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
1044 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
1047 MUTEX_ENTER(&rx_connHashTable_lock);
1048 conn->type = RX_CLIENT_CONNECTION;
1049 conn->epoch = rx_epoch;
1050 conn->cid = rx_nextCid;
1052 conn->peer = rxi_FindPeer(shost, sport, 1);
1053 conn->serviceId = sservice;
1054 conn->securityObject = securityObject;
1055 conn->securityData = (void *) 0;
1056 conn->securityIndex = serviceSecurityIndex;
1057 rx_SetConnDeadTime(conn, rx_connDeadTime);
1058 rx_SetConnSecondsUntilNatPing(conn, 0);
1059 conn->ackRate = RX_FAST_ACK_RATE;
1060 conn->nSpecific = 0;
1061 conn->specific = NULL;
1062 conn->challengeEvent = NULL;
1063 conn->delayedAbortEvent = NULL;
1064 conn->abortCount = 0;
1066 for (i = 0; i < RX_MAXCALLS; i++) {
1067 conn->twind[i] = rx_initSendWindow;
1068 conn->rwind[i] = rx_initReceiveWindow;
1069 conn->lastBusy[i] = 0;
1072 RXS_NewConnection(securityObject, conn);
1074 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1076 conn->refCount++; /* no lock required since only this thread knows... */
1077 conn->next = rx_connHashTable[hashindex];
1078 rx_connHashTable[hashindex] = conn;
1079 if (rx_stats_active)
1080 rx_atomic_inc(&rx_stats.nClientConns);
1081 MUTEX_EXIT(&rx_connHashTable_lock);
1087 * Ensure a connection's timeout values are valid.
1089 * @param[in] conn The connection to check
1091 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1092 * unless idleDeadTime and/or hardDeadTime are not set
1096 rxi_CheckConnTimeouts(struct rx_connection *conn)
1098 /* a connection's timeouts must have the relationship
1099 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1100 * total loss of network to a peer may cause an idle timeout instead of a
1101 * dead timeout, simply because the idle timeout gets hit first. Also set
1102 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1103 /* this logic is slightly complicated by the fact that
1104 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1106 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1107 if (conn->idleDeadTime) {
1108 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1110 if (conn->hardDeadTime) {
1111 if (conn->idleDeadTime) {
1112 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1114 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1120 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1122 /* The idea is to set the dead time to a value that allows several
1123 * keepalives to be dropped without timing out the connection. */
1124 conn->secondsUntilDead = seconds;
1125 rxi_CheckConnTimeouts(conn);
1126 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1130 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1132 conn->hardDeadTime = seconds;
1133 rxi_CheckConnTimeouts(conn);
1137 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1139 conn->idleDeadTime = seconds;
1140 rxi_CheckConnTimeouts(conn);
1143 int rxi_lowPeerRefCount = 0;
1144 int rxi_lowConnRefCount = 0;
1147 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1148 * NOTE: must not be called with rx_connHashTable_lock held.
1151 rxi_CleanupConnection(struct rx_connection *conn)
1153 /* Notify the service exporter, if requested, that this connection
1154 * is being destroyed */
1155 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1156 (*conn->service->destroyConnProc) (conn);
1158 /* Notify the security module that this connection is being destroyed */
1159 RXS_DestroyConnection(conn->securityObject, conn);
1161 /* If this is the last connection using the rx_peer struct, set its
1162 * idle time to now. rxi_ReapConnections will reap it if it's still
1163 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1165 MUTEX_ENTER(&rx_peerHashTable_lock);
1166 if (conn->peer->refCount < 2) {
1167 conn->peer->idleWhen = clock_Sec();
1168 if (conn->peer->refCount < 1) {
1169 conn->peer->refCount = 1;
1170 if (rx_stats_active) {
1171 MUTEX_ENTER(&rx_stats_mutex);
1172 rxi_lowPeerRefCount++;
1173 MUTEX_EXIT(&rx_stats_mutex);
1177 conn->peer->refCount--;
1178 MUTEX_EXIT(&rx_peerHashTable_lock);
1180 if (rx_stats_active)
1182 if (conn->type == RX_SERVER_CONNECTION)
1183 rx_atomic_dec(&rx_stats.nServerConns);
1185 rx_atomic_dec(&rx_stats.nClientConns);
1188 if (conn->specific) {
1190 for (i = 0; i < conn->nSpecific; i++) {
1191 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1192 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1193 conn->specific[i] = NULL;
1195 free(conn->specific);
1197 conn->specific = NULL;
1198 conn->nSpecific = 0;
1199 #endif /* !KERNEL */
1201 MUTEX_DESTROY(&conn->conn_call_lock);
1202 MUTEX_DESTROY(&conn->conn_data_lock);
1203 CV_DESTROY(&conn->conn_call_cv);
1205 rxi_FreeConnection(conn);
1208 /* Destroy the specified connection */
1210 rxi_DestroyConnection(struct rx_connection *conn)
1212 MUTEX_ENTER(&rx_connHashTable_lock);
1213 rxi_DestroyConnectionNoLock(conn);
1214 /* conn should be at the head of the cleanup list */
1215 if (conn == rx_connCleanup_list) {
1216 rx_connCleanup_list = rx_connCleanup_list->next;
1217 MUTEX_EXIT(&rx_connHashTable_lock);
1218 rxi_CleanupConnection(conn);
1220 #ifdef RX_ENABLE_LOCKS
1222 MUTEX_EXIT(&rx_connHashTable_lock);
1224 #endif /* RX_ENABLE_LOCKS */
1228 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1230 struct rx_connection **conn_ptr;
1232 struct rx_packet *packet;
1239 MUTEX_ENTER(&conn->conn_data_lock);
1240 MUTEX_ENTER(&rx_refcnt_mutex);
1241 if (conn->refCount > 0)
1244 if (rx_stats_active) {
1245 MUTEX_ENTER(&rx_stats_mutex);
1246 rxi_lowConnRefCount++;
1247 MUTEX_EXIT(&rx_stats_mutex);
1251 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1252 /* Busy; wait till the last guy before proceeding */
1253 MUTEX_EXIT(&rx_refcnt_mutex);
1254 MUTEX_EXIT(&conn->conn_data_lock);
1259 /* If the client previously called rx_NewCall, but it is still
1260 * waiting, treat this as a running call, and wait to destroy the
1261 * connection later when the call completes. */
1262 if ((conn->type == RX_CLIENT_CONNECTION)
1263 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1264 conn->flags |= RX_CONN_DESTROY_ME;
1265 MUTEX_EXIT(&conn->conn_data_lock);
1269 MUTEX_EXIT(&rx_refcnt_mutex);
1270 MUTEX_EXIT(&conn->conn_data_lock);
1272 /* Check for extant references to this connection */
1273 MUTEX_ENTER(&conn->conn_call_lock);
1274 for (i = 0; i < RX_MAXCALLS; i++) {
1275 struct rx_call *call = conn->call[i];
1278 if (conn->type == RX_CLIENT_CONNECTION) {
1279 MUTEX_ENTER(&call->lock);
1280 if (call->delayedAckEvent) {
1281 /* Push the final acknowledgment out now--there
1282 * won't be a subsequent call to acknowledge the
1283 * last reply packets */
1284 rxi_CancelDelayedAckEvent(call);
1285 if (call->state == RX_STATE_PRECALL
1286 || call->state == RX_STATE_ACTIVE) {
1287 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1292 MUTEX_EXIT(&call->lock);
1296 MUTEX_EXIT(&conn->conn_call_lock);
1298 #ifdef RX_ENABLE_LOCKS
1300 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1301 MUTEX_EXIT(&conn->conn_data_lock);
1303 /* Someone is accessing a packet right now. */
1307 #endif /* RX_ENABLE_LOCKS */
1310 /* Don't destroy the connection if there are any call
1311 * structures still in use */
1312 MUTEX_ENTER(&conn->conn_data_lock);
1313 conn->flags |= RX_CONN_DESTROY_ME;
1314 MUTEX_EXIT(&conn->conn_data_lock);
1319 if (conn->natKeepAliveEvent) {
1320 rxi_NatKeepAliveOff(conn);
1323 if (conn->delayedAbortEvent) {
1324 rxevent_Cancel(&conn->delayedAbortEvent);
1325 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1327 MUTEX_ENTER(&conn->conn_data_lock);
1328 rxi_SendConnectionAbort(conn, packet, 0, 1);
1329 MUTEX_EXIT(&conn->conn_data_lock);
1330 rxi_FreePacket(packet);
1334 /* Remove from connection hash table before proceeding */
1336 &rx_connHashTable[CONN_HASH
1337 (peer->host, peer->port, conn->cid, conn->epoch,
1339 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1340 if (*conn_ptr == conn) {
1341 *conn_ptr = conn->next;
1345 /* if the conn that we are destroying was the last connection, then we
1346 * clear rxLastConn as well */
1347 if (rxLastConn == conn)
1350 /* Make sure the connection is completely reset before deleting it. */
1351 /* get rid of pending events that could zap us later */
1352 rxevent_Cancel(&conn->challengeEvent);
1353 rxevent_Cancel(&conn->checkReachEvent);
1354 rxevent_Cancel(&conn->natKeepAliveEvent);
1356 /* Add the connection to the list of destroyed connections that
1357 * need to be cleaned up. This is necessary to avoid deadlocks
1358 * in the routines we call to inform others that this connection is
1359 * being destroyed. */
1360 conn->next = rx_connCleanup_list;
1361 rx_connCleanup_list = conn;
1364 /* Externally available version */
1366 rx_DestroyConnection(struct rx_connection *conn)
1371 rxi_DestroyConnection(conn);
1376 rx_GetConnection(struct rx_connection *conn)
1381 MUTEX_ENTER(&rx_refcnt_mutex);
1383 MUTEX_EXIT(&rx_refcnt_mutex);
1387 #ifdef RX_ENABLE_LOCKS
1388 /* Wait for the transmit queue to no longer be busy.
1389 * requires the call->lock to be held */
1391 rxi_WaitforTQBusy(struct rx_call *call) {
1392 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1393 call->flags |= RX_CALL_TQ_WAIT;
1395 MUTEX_ASSERT(&call->lock);
1396 CV_WAIT(&call->cv_tq, &call->lock);
1398 if (call->tqWaiters == 0) {
1399 call->flags &= ~RX_CALL_TQ_WAIT;
1406 rxi_WakeUpTransmitQueue(struct rx_call *call)
1408 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1409 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1410 call, call->tqWaiters, call->flags));
1411 #ifdef RX_ENABLE_LOCKS
1412 MUTEX_ASSERT(&call->lock);
1413 CV_BROADCAST(&call->cv_tq);
1414 #else /* RX_ENABLE_LOCKS */
1415 osi_rxWakeup(&call->tq);
1416 #endif /* RX_ENABLE_LOCKS */
1420 /* Start a new rx remote procedure call, on the specified connection.
1421 * If wait is set to 1, wait for a free call channel; otherwise return
1422 * 0. Maxtime gives the maximum number of seconds this call may take,
1423 * after rx_NewCall returns. After this time interval, a call to any
1424 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1425 * For fine grain locking, we hold the conn_call_lock in order to
1426 * to ensure that we don't get signalle after we found a call in an active
1427 * state and before we go to sleep.
1430 rx_NewCall(struct rx_connection *conn)
1432 int i, wait, ignoreBusy = 1;
1433 struct rx_call *call;
1434 struct clock queueTime;
1435 afs_uint32 leastBusy = 0;
1439 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1442 clock_GetTime(&queueTime);
1444 * Check if there are others waiting for a new call.
1445 * If so, let them go first to avoid starving them.
1446 * This is a fairly simple scheme, and might not be
1447 * a complete solution for large numbers of waiters.
1449 * makeCallWaiters keeps track of the number of
1450 * threads waiting to make calls and the
1451 * RX_CONN_MAKECALL_WAITING flag bit is used to
1452 * indicate that there are indeed calls waiting.
1453 * The flag is set when the waiter is incremented.
1454 * It is only cleared when makeCallWaiters is 0.
1455 * This prevents us from accidently destroying the
1456 * connection while it is potentially about to be used.
1458 MUTEX_ENTER(&conn->conn_call_lock);
1459 MUTEX_ENTER(&conn->conn_data_lock);
1460 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1461 conn->flags |= RX_CONN_MAKECALL_WAITING;
1462 conn->makeCallWaiters++;
1463 MUTEX_EXIT(&conn->conn_data_lock);
1465 #ifdef RX_ENABLE_LOCKS
1466 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1470 MUTEX_ENTER(&conn->conn_data_lock);
1471 conn->makeCallWaiters--;
1472 if (conn->makeCallWaiters == 0)
1473 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1476 /* We are now the active thread in rx_NewCall */
1477 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1478 MUTEX_EXIT(&conn->conn_data_lock);
1483 for (i = 0; i < RX_MAXCALLS; i++) {
1484 call = conn->call[i];
1486 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1487 /* we're not ignoring busy call slots; only look at the
1488 * call slot that is the "least" busy */
1492 if (call->state == RX_STATE_DALLY) {
1493 MUTEX_ENTER(&call->lock);
1494 if (call->state == RX_STATE_DALLY) {
1495 if (ignoreBusy && conn->lastBusy[i]) {
1496 /* if we're ignoring busy call slots, skip any ones that
1497 * have lastBusy set */
1498 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1499 leastBusy = conn->lastBusy[i];
1501 MUTEX_EXIT(&call->lock);
1506 * We are setting the state to RX_STATE_RESET to
1507 * ensure that no one else will attempt to use this
1508 * call once we drop the conn->conn_call_lock and
1509 * call->lock. We must drop the conn->conn_call_lock
1510 * before calling rxi_ResetCall because the process
1511 * of clearing the transmit queue can block for an
1512 * extended period of time. If we block while holding
1513 * the conn->conn_call_lock, then all rx_EndCall
1514 * processing will block as well. This has a detrimental
1515 * effect on overall system performance.
1517 call->state = RX_STATE_RESET;
1518 (*call->callNumber)++;
1519 MUTEX_EXIT(&conn->conn_call_lock);
1520 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1521 rxi_ResetCall(call, 0);
1522 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1526 * If we failed to be able to safely obtain the
1527 * conn->conn_call_lock we will have to drop the
1528 * call->lock to avoid a deadlock. When the call->lock
1529 * is released the state of the call can change. If it
1530 * is no longer RX_STATE_RESET then some other thread is
1533 MUTEX_EXIT(&call->lock);
1534 MUTEX_ENTER(&conn->conn_call_lock);
1535 MUTEX_ENTER(&call->lock);
1537 if (call->state == RX_STATE_RESET)
1541 * If we get here it means that after dropping
1542 * the conn->conn_call_lock and call->lock that
1543 * the call is no longer ours. If we can't find
1544 * a free call in the remaining slots we should
1545 * not go immediately to RX_CONN_MAKECALL_WAITING
1546 * because by dropping the conn->conn_call_lock
1547 * we have given up synchronization with rx_EndCall.
1548 * Instead, cycle through one more time to see if
1549 * we can find a call that can call our own.
1551 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1554 MUTEX_EXIT(&call->lock);
1557 if (ignoreBusy && conn->lastBusy[i]) {
1558 /* if we're ignoring busy call slots, skip any ones that
1559 * have lastBusy set */
1560 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1561 leastBusy = conn->lastBusy[i];
1566 /* rxi_NewCall returns with mutex locked */
1567 call = rxi_NewCall(conn, i);
1568 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1572 if (i < RX_MAXCALLS) {
1573 conn->lastBusy[i] = 0;
1574 call->flags &= ~RX_CALL_PEER_BUSY;
1579 if (leastBusy && ignoreBusy) {
1580 /* we didn't find a useable call slot, but we did see at least one
1581 * 'busy' slot; look again and only use a slot with the 'least
1587 MUTEX_ENTER(&conn->conn_data_lock);
1588 conn->flags |= RX_CONN_MAKECALL_WAITING;
1589 conn->makeCallWaiters++;
1590 MUTEX_EXIT(&conn->conn_data_lock);
1592 #ifdef RX_ENABLE_LOCKS
1593 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1597 MUTEX_ENTER(&conn->conn_data_lock);
1598 conn->makeCallWaiters--;
1599 if (conn->makeCallWaiters == 0)
1600 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1601 MUTEX_EXIT(&conn->conn_data_lock);
1603 /* Client is initially in send mode */
1604 call->state = RX_STATE_ACTIVE;
1605 call->error = conn->error;
1607 call->app.mode = RX_MODE_ERROR;
1609 call->app.mode = RX_MODE_SENDING;
1611 #ifdef AFS_RXERRQ_ENV
1612 /* remember how many network errors the peer has when we started, so if
1613 * more errors are encountered after the call starts, we know the other endpoint won't be
1614 * responding to us */
1615 call->neterr_gen = rx_atomic_read(&conn->peer->neterrs);
1618 /* remember start time for call in case we have hard dead time limit */
1619 call->queueTime = queueTime;
1620 clock_GetTime(&call->startTime);
1621 call->app.bytesSent = 0;
1622 call->app.bytesRcvd = 0;
1624 /* Turn on busy protocol. */
1625 rxi_KeepAliveOn(call);
1627 /* Attempt MTU discovery */
1628 rxi_GrowMTUOn(call);
1631 * We are no longer the active thread in rx_NewCall
1633 MUTEX_ENTER(&conn->conn_data_lock);
1634 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1635 MUTEX_EXIT(&conn->conn_data_lock);
1638 * Wake up anyone else who might be giving us a chance to
1639 * run (see code above that avoids resource starvation).
1641 #ifdef RX_ENABLE_LOCKS
1642 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1643 osi_Panic("rx_NewCall call about to be used without an empty tq");
1646 CV_BROADCAST(&conn->conn_call_cv);
1650 MUTEX_EXIT(&conn->conn_call_lock);
1651 MUTEX_EXIT(&call->lock);
1654 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1659 rxi_HasActiveCalls(struct rx_connection *aconn)
1662 struct rx_call *tcall;
1666 for (i = 0; i < RX_MAXCALLS; i++) {
1667 if ((tcall = aconn->call[i])) {
1668 if ((tcall->state == RX_STATE_ACTIVE)
1669 || (tcall->state == RX_STATE_PRECALL)) {
1680 rxi_GetCallNumberVector(struct rx_connection *aconn,
1681 afs_int32 * aint32s)
1684 struct rx_call *tcall;
1688 MUTEX_ENTER(&aconn->conn_call_lock);
1689 for (i = 0; i < RX_MAXCALLS; i++) {
1690 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1691 aint32s[i] = aconn->callNumber[i] + 1;
1693 aint32s[i] = aconn->callNumber[i];
1695 MUTEX_EXIT(&aconn->conn_call_lock);
1701 rxi_SetCallNumberVector(struct rx_connection *aconn,
1702 afs_int32 * aint32s)
1705 struct rx_call *tcall;
1709 MUTEX_ENTER(&aconn->conn_call_lock);
1710 for (i = 0; i < RX_MAXCALLS; i++) {
1711 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1712 aconn->callNumber[i] = aint32s[i] - 1;
1714 aconn->callNumber[i] = aint32s[i];
1716 MUTEX_EXIT(&aconn->conn_call_lock);
1721 /* Advertise a new service. A service is named locally by a UDP port
1722 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1725 char *serviceName; Name for identification purposes (e.g. the
1726 service name might be used for probing for
1729 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1730 char *serviceName, struct rx_securityClass **securityObjects,
1731 int nSecurityObjects,
1732 afs_int32(*serviceProc) (struct rx_call * acall))
1734 osi_socket socket = OSI_NULLSOCKET;
1735 struct rx_service *tservice;
1741 if (serviceId == 0) {
1743 "rx_NewService: service id for service %s is not non-zero.\n",
1750 "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",
1758 tservice = rxi_AllocService();
1761 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1763 for (i = 0; i < RX_MAX_SERVICES; i++) {
1764 struct rx_service *service = rx_services[i];
1766 if (port == service->servicePort && host == service->serviceHost) {
1767 if (service->serviceId == serviceId) {
1768 /* The identical service has already been
1769 * installed; if the caller was intending to
1770 * change the security classes used by this
1771 * service, he/she loses. */
1773 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1774 serviceName, serviceId, service->serviceName);
1776 rxi_FreeService(tservice);
1779 /* Different service, same port: re-use the socket
1780 * which is bound to the same port */
1781 socket = service->socket;
1784 if (socket == OSI_NULLSOCKET) {
1785 /* If we don't already have a socket (from another
1786 * service on same port) get a new one */
1787 socket = rxi_GetHostUDPSocket(host, port);
1788 if (socket == OSI_NULLSOCKET) {
1790 rxi_FreeService(tservice);
1795 service->socket = socket;
1796 service->serviceHost = host;
1797 service->servicePort = port;
1798 service->serviceId = serviceId;
1799 service->serviceName = serviceName;
1800 service->nSecurityObjects = nSecurityObjects;
1801 service->securityObjects = securityObjects;
1802 service->minProcs = 0;
1803 service->maxProcs = 1;
1804 service->idleDeadTime = 60;
1805 service->connDeadTime = rx_connDeadTime;
1806 service->executeRequestProc = serviceProc;
1807 service->checkReach = 0;
1808 service->nSpecific = 0;
1809 service->specific = NULL;
1810 rx_services[i] = service; /* not visible until now */
1816 rxi_FreeService(tservice);
1817 (osi_Msg "rx_NewService: cannot support > %d services\n",
1822 /* Set configuration options for all of a service's security objects */
1825 rx_SetSecurityConfiguration(struct rx_service *service,
1826 rx_securityConfigVariables type,
1830 for (i = 0; i<service->nSecurityObjects; i++) {
1831 if (service->securityObjects[i]) {
1832 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1840 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1841 struct rx_securityClass **securityObjects, int nSecurityObjects,
1842 afs_int32(*serviceProc) (struct rx_call * acall))
1844 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1847 /* Generic request processing loop. This routine should be called
1848 * by the implementation dependent rx_ServerProc. If socketp is
1849 * non-null, it will be set to the file descriptor that this thread
1850 * is now listening on. If socketp is null, this routine will never
1853 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1855 struct rx_call *call;
1857 struct rx_service *tservice = NULL;
1864 call = rx_GetCall(threadID, tservice, socketp);
1865 if (socketp && *socketp != OSI_NULLSOCKET) {
1866 /* We are now a listener thread */
1872 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1873 #ifdef RX_ENABLE_LOCKS
1875 #endif /* RX_ENABLE_LOCKS */
1876 afs_termState = AFSOP_STOP_AFS;
1877 afs_osi_Wakeup(&afs_termState);
1878 #ifdef RX_ENABLE_LOCKS
1880 #endif /* RX_ENABLE_LOCKS */
1885 /* if server is restarting( typically smooth shutdown) then do not
1886 * allow any new calls.
1889 if (rx_tranquil && (call != NULL)) {
1893 MUTEX_ENTER(&call->lock);
1895 rxi_CallError(call, RX_RESTARTING);
1896 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1898 MUTEX_EXIT(&call->lock);
1903 tservice = call->conn->service;
1905 if (tservice->beforeProc)
1906 (*tservice->beforeProc) (call);
1908 code = tservice->executeRequestProc(call);
1910 if (tservice->afterProc)
1911 (*tservice->afterProc) (call, code);
1913 rx_EndCall(call, code);
1915 if (tservice->postProc)
1916 (*tservice->postProc) (code);
1918 if (rx_stats_active) {
1919 MUTEX_ENTER(&rx_stats_mutex);
1921 MUTEX_EXIT(&rx_stats_mutex);
1928 rx_WakeupServerProcs(void)
1930 struct rx_serverQueueEntry *np, *tqp;
1931 struct opr_queue *cursor;
1935 MUTEX_ENTER(&rx_serverPool_lock);
1937 #ifdef RX_ENABLE_LOCKS
1938 if (rx_waitForPacket)
1939 CV_BROADCAST(&rx_waitForPacket->cv);
1940 #else /* RX_ENABLE_LOCKS */
1941 if (rx_waitForPacket)
1942 osi_rxWakeup(rx_waitForPacket);
1943 #endif /* RX_ENABLE_LOCKS */
1944 MUTEX_ENTER(&freeSQEList_lock);
1945 for (np = rx_FreeSQEList; np; np = tqp) {
1946 tqp = *(struct rx_serverQueueEntry **)np;
1947 #ifdef RX_ENABLE_LOCKS
1948 CV_BROADCAST(&np->cv);
1949 #else /* RX_ENABLE_LOCKS */
1951 #endif /* RX_ENABLE_LOCKS */
1953 MUTEX_EXIT(&freeSQEList_lock);
1954 for (opr_queue_Scan(&rx_idleServerQueue, cursor)) {
1955 np = opr_queue_Entry(cursor, struct rx_serverQueueEntry, entry);
1956 #ifdef RX_ENABLE_LOCKS
1957 CV_BROADCAST(&np->cv);
1958 #else /* RX_ENABLE_LOCKS */
1960 #endif /* RX_ENABLE_LOCKS */
1962 MUTEX_EXIT(&rx_serverPool_lock);
1967 * One thing that seems to happen is that all the server threads get
1968 * tied up on some empty or slow call, and then a whole bunch of calls
1969 * arrive at once, using up the packet pool, so now there are more
1970 * empty calls. The most critical resources here are server threads
1971 * and the free packet pool. The "doreclaim" code seems to help in
1972 * general. I think that eventually we arrive in this state: there
1973 * are lots of pending calls which do have all their packets present,
1974 * so they won't be reclaimed, are multi-packet calls, so they won't
1975 * be scheduled until later, and thus are tying up most of the free
1976 * packet pool for a very long time.
1978 * 1. schedule multi-packet calls if all the packets are present.
1979 * Probably CPU-bound operation, useful to return packets to pool.
1980 * Do what if there is a full window, but the last packet isn't here?
1981 * 3. preserve one thread which *only* runs "best" calls, otherwise
1982 * it sleeps and waits for that type of call.
1983 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1984 * the current dataquota business is badly broken. The quota isn't adjusted
1985 * to reflect how many packets are presently queued for a running call.
1986 * So, when we schedule a queued call with a full window of packets queued
1987 * up for it, that *should* free up a window full of packets for other 2d-class
1988 * calls to be able to use from the packet pool. But it doesn't.
1990 * NB. Most of the time, this code doesn't run -- since idle server threads
1991 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1992 * as a new call arrives.
1994 /* Sleep until a call arrives. Returns a pointer to the call, ready
1995 * for an rx_Read. */
1996 #ifdef RX_ENABLE_LOCKS
1998 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2000 struct rx_serverQueueEntry *sq;
2001 struct rx_call *call = (struct rx_call *)0;
2002 struct rx_service *service = NULL;
2004 MUTEX_ENTER(&freeSQEList_lock);
2006 if ((sq = rx_FreeSQEList)) {
2007 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2008 MUTEX_EXIT(&freeSQEList_lock);
2009 } else { /* otherwise allocate a new one and return that */
2010 MUTEX_EXIT(&freeSQEList_lock);
2011 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2012 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2013 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2016 MUTEX_ENTER(&rx_serverPool_lock);
2017 if (cur_service != NULL) {
2018 ReturnToServerPool(cur_service);
2021 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2022 struct rx_call *tcall, *choice2 = NULL;
2023 struct opr_queue *cursor;
2025 /* Scan for eligible incoming calls. A call is not eligible
2026 * if the maximum number of calls for its service type are
2027 * already executing */
2028 /* One thread will process calls FCFS (to prevent starvation),
2029 * while the other threads may run ahead looking for calls which
2030 * have all their input data available immediately. This helps
2031 * keep threads from blocking, waiting for data from the client. */
2032 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2033 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2035 service = tcall->conn->service;
2036 if (!QuotaOK(service)) {
2039 MUTEX_ENTER(&rx_pthread_mutex);
2040 if (tno == rxi_fcfs_thread_num
2041 || opr_queue_IsEnd(&rx_incomingCallQueue, cursor)) {
2042 MUTEX_EXIT(&rx_pthread_mutex);
2043 /* If we're the fcfs thread , then we'll just use
2044 * this call. If we haven't been able to find an optimal
2045 * choice, and we're at the end of the list, then use a
2046 * 2d choice if one has been identified. Otherwise... */
2047 call = (choice2 ? choice2 : tcall);
2048 service = call->conn->service;
2050 MUTEX_EXIT(&rx_pthread_mutex);
2051 if (!opr_queue_IsEmpty(&tcall->rq)) {
2052 struct rx_packet *rp;
2053 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2055 if (rp->header.seq == 1) {
2057 || (rp->header.flags & RX_LAST_PACKET)) {
2059 } else if (rxi_2dchoice && !choice2
2060 && !(tcall->flags & RX_CALL_CLEARED)
2061 && (tcall->rprev > rxi_HardAckRate)) {
2071 ReturnToServerPool(service);
2077 opr_queue_Remove(&call->entry);
2078 MUTEX_EXIT(&rx_serverPool_lock);
2079 MUTEX_ENTER(&call->lock);
2081 if (call->flags & RX_CALL_WAIT_PROC) {
2082 call->flags &= ~RX_CALL_WAIT_PROC;
2083 rx_atomic_dec(&rx_nWaiting);
2086 if (call->state != RX_STATE_PRECALL || call->error) {
2087 MUTEX_EXIT(&call->lock);
2088 MUTEX_ENTER(&rx_serverPool_lock);
2089 ReturnToServerPool(service);
2094 if (opr_queue_IsEmpty(&call->rq)
2095 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1)
2096 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2098 CLEAR_CALL_QUEUE_LOCK(call);
2101 /* If there are no eligible incoming calls, add this process
2102 * to the idle server queue, to wait for one */
2106 *socketp = OSI_NULLSOCKET;
2108 sq->socketp = socketp;
2109 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2110 #ifndef AFS_AIX41_ENV
2111 rx_waitForPacket = sq;
2112 #endif /* AFS_AIX41_ENV */
2114 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2116 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2117 MUTEX_EXIT(&rx_serverPool_lock);
2118 return (struct rx_call *)0;
2121 } while (!(call = sq->newcall)
2122 && !(socketp && *socketp != OSI_NULLSOCKET));
2123 MUTEX_EXIT(&rx_serverPool_lock);
2125 MUTEX_ENTER(&call->lock);
2131 MUTEX_ENTER(&freeSQEList_lock);
2132 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2133 rx_FreeSQEList = sq;
2134 MUTEX_EXIT(&freeSQEList_lock);
2137 clock_GetTime(&call->startTime);
2138 call->state = RX_STATE_ACTIVE;
2139 call->app.mode = RX_MODE_RECEIVING;
2140 #ifdef RX_KERNEL_TRACE
2141 if (ICL_SETACTIVE(afs_iclSetp)) {
2142 int glockOwner = ISAFS_GLOCK();
2145 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2146 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2153 rxi_calltrace(RX_CALL_START, call);
2154 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2155 call->conn->service->servicePort, call->conn->service->serviceId,
2158 MUTEX_EXIT(&call->lock);
2159 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2161 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2166 #else /* RX_ENABLE_LOCKS */
2168 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2170 struct rx_serverQueueEntry *sq;
2171 struct rx_call *call = (struct rx_call *)0, *choice2;
2172 struct rx_service *service = NULL;
2176 MUTEX_ENTER(&freeSQEList_lock);
2178 if ((sq = rx_FreeSQEList)) {
2179 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2180 MUTEX_EXIT(&freeSQEList_lock);
2181 } else { /* otherwise allocate a new one and return that */
2182 MUTEX_EXIT(&freeSQEList_lock);
2183 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2184 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2185 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2187 MUTEX_ENTER(&sq->lock);
2189 if (cur_service != NULL) {
2190 cur_service->nRequestsRunning--;
2191 MUTEX_ENTER(&rx_quota_mutex);
2192 if (cur_service->nRequestsRunning < cur_service->minProcs)
2195 MUTEX_EXIT(&rx_quota_mutex);
2197 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2198 struct rx_call *tcall;
2199 struct opr_queue *cursor;
2200 /* Scan for eligible incoming calls. A call is not eligible
2201 * if the maximum number of calls for its service type are
2202 * already executing */
2203 /* One thread will process calls FCFS (to prevent starvation),
2204 * while the other threads may run ahead looking for calls which
2205 * have all their input data available immediately. This helps
2206 * keep threads from blocking, waiting for data from the client. */
2207 choice2 = (struct rx_call *)0;
2208 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2209 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2210 service = tcall->conn->service;
2211 if (QuotaOK(service)) {
2212 MUTEX_ENTER(&rx_pthread_mutex);
2213 /* XXX - If tcall->entry.next is NULL, then we're no longer
2214 * on a queue at all. This shouldn't happen. */
2215 if (tno == rxi_fcfs_thread_num || !tcall->entry.next) {
2216 MUTEX_EXIT(&rx_pthread_mutex);
2217 /* If we're the fcfs thread, then we'll just use
2218 * this call. If we haven't been able to find an optimal
2219 * choice, and we're at the end of the list, then use a
2220 * 2d choice if one has been identified. Otherwise... */
2221 call = (choice2 ? choice2 : tcall);
2222 service = call->conn->service;
2224 MUTEX_EXIT(&rx_pthread_mutex);
2225 if (!opr_queue_IsEmpty(&tcall->rq)) {
2226 struct rx_packet *rp;
2227 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2229 if (rp->header.seq == 1
2231 || (rp->header.flags & RX_LAST_PACKET))) {
2233 } else if (rxi_2dchoice && !choice2
2234 && !(tcall->flags & RX_CALL_CLEARED)
2235 && (tcall->rprev > rxi_HardAckRate)) {
2248 opr_queue_Remove(&call->entry);
2249 /* we can't schedule a call if there's no data!!! */
2250 /* send an ack if there's no data, if we're missing the
2251 * first packet, or we're missing something between first
2252 * and last -- there's a "hole" in the incoming data. */
2253 if (opr_queue_IsEmpty(&call->rq)
2254 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1
2255 || call->rprev != opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq)
2256 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2258 call->flags &= (~RX_CALL_WAIT_PROC);
2259 service->nRequestsRunning++;
2260 /* just started call in minProcs pool, need fewer to maintain
2262 MUTEX_ENTER(&rx_quota_mutex);
2263 if (service->nRequestsRunning <= service->minProcs)
2266 MUTEX_EXIT(&rx_quota_mutex);
2267 rx_atomic_dec(&rx_nWaiting);
2268 /* MUTEX_EXIT(&call->lock); */
2270 /* If there are no eligible incoming calls, add this process
2271 * to the idle server queue, to wait for one */
2274 *socketp = OSI_NULLSOCKET;
2276 sq->socketp = socketp;
2277 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2281 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2283 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2284 return (struct rx_call *)0;
2287 } while (!(call = sq->newcall)
2288 && !(socketp && *socketp != OSI_NULLSOCKET));
2290 MUTEX_EXIT(&sq->lock);
2292 MUTEX_ENTER(&freeSQEList_lock);
2293 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2294 rx_FreeSQEList = sq;
2295 MUTEX_EXIT(&freeSQEList_lock);
2298 clock_GetTime(&call->startTime);
2299 call->state = RX_STATE_ACTIVE;
2300 call->app.mode = RX_MODE_RECEIVING;
2301 #ifdef RX_KERNEL_TRACE
2302 if (ICL_SETACTIVE(afs_iclSetp)) {
2303 int glockOwner = ISAFS_GLOCK();
2306 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2307 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2314 rxi_calltrace(RX_CALL_START, call);
2315 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2316 call->conn->service->servicePort, call->conn->service->serviceId,
2319 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2326 #endif /* RX_ENABLE_LOCKS */
2330 /* Establish a procedure to be called when a packet arrives for a
2331 * call. This routine will be called at most once after each call,
2332 * and will also be called if there is an error condition on the or
2333 * the call is complete. Used by multi rx to build a selection
2334 * function which determines which of several calls is likely to be a
2335 * good one to read from.
2336 * NOTE: the way this is currently implemented it is probably only a
2337 * good idea to (1) use it immediately after a newcall (clients only)
2338 * and (2) only use it once. Other uses currently void your warranty
2341 rx_SetArrivalProc(struct rx_call *call,
2342 void (*proc) (struct rx_call * call,
2345 void * handle, int arg)
2347 call->arrivalProc = proc;
2348 call->arrivalProcHandle = handle;
2349 call->arrivalProcArg = arg;
2352 /* Call is finished (possibly prematurely). Return rc to the peer, if
2353 * appropriate, and return the final error code from the conversation
2357 rx_EndCall(struct rx_call *call, afs_int32 rc)
2359 struct rx_connection *conn = call->conn;
2363 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2364 call, rc, call->error, call->abortCode));
2367 MUTEX_ENTER(&call->lock);
2369 if (rc == 0 && call->error == 0) {
2370 call->abortCode = 0;
2371 call->abortCount = 0;
2374 call->arrivalProc = (void (*)())0;
2375 if (rc && call->error == 0) {
2376 rxi_CallError(call, rc);
2377 call->app.mode = RX_MODE_ERROR;
2378 /* Send an abort message to the peer if this error code has
2379 * only just been set. If it was set previously, assume the
2380 * peer has already been sent the error code or will request it
2382 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2384 if (conn->type == RX_SERVER_CONNECTION) {
2385 /* Make sure reply or at least dummy reply is sent */
2386 if (call->app.mode == RX_MODE_RECEIVING) {
2387 MUTEX_EXIT(&call->lock);
2388 rxi_WriteProc(call, 0, 0);
2389 MUTEX_ENTER(&call->lock);
2391 if (call->app.mode == RX_MODE_SENDING) {
2392 MUTEX_EXIT(&call->lock);
2393 rxi_FlushWrite(call);
2394 MUTEX_ENTER(&call->lock);
2396 rxi_calltrace(RX_CALL_END, call);
2397 /* Call goes to hold state until reply packets are acknowledged */
2398 if (call->tfirst + call->nSoftAcked < call->tnext) {
2399 call->state = RX_STATE_HOLD;
2401 call->state = RX_STATE_DALLY;
2402 rxi_ClearTransmitQueue(call, 0);
2403 rxi_rto_cancel(call);
2404 rxi_CancelKeepAliveEvent(call);
2406 } else { /* Client connection */
2408 /* Make sure server receives input packets, in the case where
2409 * no reply arguments are expected */
2411 if ((call->app.mode == RX_MODE_SENDING)
2412 || (call->app.mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2413 MUTEX_EXIT(&call->lock);
2414 (void)rxi_ReadProc(call, &dummy, 1);
2415 MUTEX_ENTER(&call->lock);
2418 /* If we had an outstanding delayed ack, be nice to the server
2419 * and force-send it now.
2421 if (call->delayedAckEvent) {
2422 rxi_CancelDelayedAckEvent(call);
2423 rxi_SendDelayedAck(NULL, call, NULL, 0);
2426 /* We need to release the call lock since it's lower than the
2427 * conn_call_lock and we don't want to hold the conn_call_lock
2428 * over the rx_ReadProc call. The conn_call_lock needs to be held
2429 * here for the case where rx_NewCall is perusing the calls on
2430 * the connection structure. We don't want to signal until
2431 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2432 * have checked this call, found it active and by the time it
2433 * goes to sleep, will have missed the signal.
2435 MUTEX_EXIT(&call->lock);
2436 MUTEX_ENTER(&conn->conn_call_lock);
2437 MUTEX_ENTER(&call->lock);
2439 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2440 conn->lastBusy[call->channel] = 0;
2443 MUTEX_ENTER(&conn->conn_data_lock);
2444 conn->flags |= RX_CONN_BUSY;
2445 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2446 MUTEX_EXIT(&conn->conn_data_lock);
2447 #ifdef RX_ENABLE_LOCKS
2448 CV_BROADCAST(&conn->conn_call_cv);
2453 #ifdef RX_ENABLE_LOCKS
2455 MUTEX_EXIT(&conn->conn_data_lock);
2457 #endif /* RX_ENABLE_LOCKS */
2458 call->state = RX_STATE_DALLY;
2460 error = call->error;
2462 /* currentPacket, nLeft, and NFree must be zeroed here, because
2463 * ResetCall cannot: ResetCall may be called at splnet(), in the
2464 * kernel version, and may interrupt the macros rx_Read or
2465 * rx_Write, which run at normal priority for efficiency. */
2466 if (call->app.currentPacket) {
2467 #ifdef RX_TRACK_PACKETS
2468 call->app.currentPacket->flags &= ~RX_PKTFLAG_CP;
2470 rxi_FreePacket(call->app.currentPacket);
2471 call->app.currentPacket = (struct rx_packet *)0;
2474 call->app.nLeft = call->app.nFree = call->app.curlen = 0;
2476 /* Free any packets from the last call to ReadvProc/WritevProc */
2477 #ifdef RXDEBUG_PACKET
2479 #endif /* RXDEBUG_PACKET */
2480 rxi_FreePackets(0, &call->app.iovq);
2481 MUTEX_EXIT(&call->lock);
2483 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2484 if (conn->type == RX_CLIENT_CONNECTION) {
2485 MUTEX_ENTER(&conn->conn_data_lock);
2486 conn->flags &= ~RX_CONN_BUSY;
2487 MUTEX_EXIT(&conn->conn_data_lock);
2488 MUTEX_EXIT(&conn->conn_call_lock);
2492 * Map errors to the local host's errno.h format.
2494 error = ntoh_syserr_conv(error);
2496 /* If the caller said the call failed with some error, we had better
2497 * return an error code. */
2498 osi_Assert(!rc || error);
2502 #if !defined(KERNEL)
2504 /* Call this routine when shutting down a server or client (especially
2505 * clients). This will allow Rx to gracefully garbage collect server
2506 * connections, and reduce the number of retries that a server might
2507 * make to a dead client.
2508 * This is not quite right, since some calls may still be ongoing and
2509 * we can't lock them to destroy them. */
2513 struct rx_connection **conn_ptr, **conn_end;
2516 if (rx_atomic_test_and_set_bit(&rxinit_status, 0))
2517 return; /* Already shutdown. */
2519 rxi_DeleteCachedConnections();
2520 if (rx_connHashTable) {
2521 MUTEX_ENTER(&rx_connHashTable_lock);
2522 for (conn_ptr = &rx_connHashTable[0], conn_end =
2523 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2525 struct rx_connection *conn, *next;
2526 for (conn = *conn_ptr; conn; conn = next) {
2528 if (conn->type == RX_CLIENT_CONNECTION) {
2529 MUTEX_ENTER(&rx_refcnt_mutex);
2531 MUTEX_EXIT(&rx_refcnt_mutex);
2532 #ifdef RX_ENABLE_LOCKS
2533 rxi_DestroyConnectionNoLock(conn);
2534 #else /* RX_ENABLE_LOCKS */
2535 rxi_DestroyConnection(conn);
2536 #endif /* RX_ENABLE_LOCKS */
2540 #ifdef RX_ENABLE_LOCKS
2541 while (rx_connCleanup_list) {
2542 struct rx_connection *conn;
2543 conn = rx_connCleanup_list;
2544 rx_connCleanup_list = rx_connCleanup_list->next;
2545 MUTEX_EXIT(&rx_connHashTable_lock);
2546 rxi_CleanupConnection(conn);
2547 MUTEX_ENTER(&rx_connHashTable_lock);
2549 MUTEX_EXIT(&rx_connHashTable_lock);
2550 #endif /* RX_ENABLE_LOCKS */
2555 afs_winsockCleanup();
2561 /* if we wakeup packet waiter too often, can get in loop with two
2562 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2564 rxi_PacketsUnWait(void)
2566 if (!rx_waitingForPackets) {
2570 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2571 return; /* still over quota */
2574 rx_waitingForPackets = 0;
2575 #ifdef RX_ENABLE_LOCKS
2576 CV_BROADCAST(&rx_waitingForPackets_cv);
2578 osi_rxWakeup(&rx_waitingForPackets);
2584 /* ------------------Internal interfaces------------------------- */
2586 /* Return this process's service structure for the
2587 * specified socket and service */
2588 static struct rx_service *
2589 rxi_FindService(osi_socket socket, u_short serviceId)
2591 struct rx_service **sp;
2592 for (sp = &rx_services[0]; *sp; sp++) {
2593 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2599 #ifdef RXDEBUG_PACKET
2600 #ifdef KDUMP_RX_LOCK
2601 static struct rx_call_rx_lock *rx_allCallsp = 0;
2603 static struct rx_call *rx_allCallsp = 0;
2605 #endif /* RXDEBUG_PACKET */
2607 /* Allocate a call structure, for the indicated channel of the
2608 * supplied connection. The mode and state of the call must be set by
2609 * the caller. Returns the call with mutex locked. */
2610 static struct rx_call *
2611 rxi_NewCall(struct rx_connection *conn, int channel)
2613 struct rx_call *call;
2614 #ifdef RX_ENABLE_LOCKS
2615 struct rx_call *cp; /* Call pointer temp */
2616 struct opr_queue *cursor;
2619 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2621 /* Grab an existing call structure, or allocate a new one.
2622 * Existing call structures are assumed to have been left reset by
2624 MUTEX_ENTER(&rx_freeCallQueue_lock);
2626 #ifdef RX_ENABLE_LOCKS
2628 * EXCEPT that the TQ might not yet be cleared out.
2629 * Skip over those with in-use TQs.
2632 for (opr_queue_Scan(&rx_freeCallQueue, cursor)) {
2633 cp = opr_queue_Entry(cursor, struct rx_call, entry);
2634 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2640 #else /* RX_ENABLE_LOCKS */
2641 if (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
2642 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
2643 #endif /* RX_ENABLE_LOCKS */
2644 opr_queue_Remove(&call->entry);
2645 if (rx_stats_active)
2646 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2647 MUTEX_EXIT(&rx_freeCallQueue_lock);
2648 MUTEX_ENTER(&call->lock);
2649 CLEAR_CALL_QUEUE_LOCK(call);
2650 #ifdef RX_ENABLE_LOCKS
2651 /* Now, if TQ wasn't cleared earlier, do it now. */
2652 rxi_WaitforTQBusy(call);
2653 if (call->flags & RX_CALL_TQ_CLEARME) {
2654 rxi_ClearTransmitQueue(call, 1);
2655 /*queue_Init(&call->tq);*/
2657 #endif /* RX_ENABLE_LOCKS */
2658 /* Bind the call to its connection structure */
2660 rxi_ResetCall(call, 1);
2663 call = rxi_Alloc(sizeof(struct rx_call));
2664 #ifdef RXDEBUG_PACKET
2665 call->allNextp = rx_allCallsp;
2666 rx_allCallsp = call;
2668 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2669 #else /* RXDEBUG_PACKET */
2670 rx_atomic_inc(&rx_stats.nCallStructs);
2671 #endif /* RXDEBUG_PACKET */
2673 MUTEX_EXIT(&rx_freeCallQueue_lock);
2674 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2675 MUTEX_ENTER(&call->lock);
2676 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2677 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2678 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2680 /* Initialize once-only items */
2681 opr_queue_Init(&call->tq);
2682 opr_queue_Init(&call->rq);
2683 opr_queue_Init(&call->app.iovq);
2684 #ifdef RXDEBUG_PACKET
2685 call->rqc = call->tqc = call->iovqc = 0;
2686 #endif /* RXDEBUG_PACKET */
2687 /* Bind the call to its connection structure (prereq for reset) */
2689 rxi_ResetCall(call, 1);
2691 call->channel = channel;
2692 call->callNumber = &conn->callNumber[channel];
2693 call->rwind = conn->rwind[channel];
2694 call->twind = conn->twind[channel];
2695 /* Note that the next expected call number is retained (in
2696 * conn->callNumber[i]), even if we reallocate the call structure
2698 conn->call[channel] = call;
2699 /* if the channel's never been used (== 0), we should start at 1, otherwise
2700 * the call number is valid from the last time this channel was used */
2701 if (*call->callNumber == 0)
2702 *call->callNumber = 1;
2707 /* A call has been inactive long enough that so we can throw away
2708 * state, including the call structure, which is placed on the call
2711 * call->lock amd rx_refcnt_mutex are held upon entry.
2712 * haveCTLock is set when called from rxi_ReapConnections.
2714 * return 1 if the call is freed, 0 if not.
2717 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2719 int channel = call->channel;
2720 struct rx_connection *conn = call->conn;
2721 u_char state = call->state;
2724 * We are setting the state to RX_STATE_RESET to
2725 * ensure that no one else will attempt to use this
2726 * call once we drop the refcnt lock. We must drop
2727 * the refcnt lock before calling rxi_ResetCall
2728 * because it cannot be held across acquiring the
2729 * freepktQ lock. NewCall does the same.
2731 call->state = RX_STATE_RESET;
2732 MUTEX_EXIT(&rx_refcnt_mutex);
2733 rxi_ResetCall(call, 0);
2735 if (MUTEX_TRYENTER(&conn->conn_call_lock))
2737 if (state == RX_STATE_DALLY || state == RX_STATE_HOLD)
2738 (*call->callNumber)++;
2740 if (call->conn->call[channel] == call)
2741 call->conn->call[channel] = 0;
2742 MUTEX_EXIT(&conn->conn_call_lock);
2745 * We couldn't obtain the conn_call_lock so we can't
2746 * disconnect the call from the connection. Set the
2747 * call state to dally so that the call can be reused.
2749 MUTEX_ENTER(&rx_refcnt_mutex);
2750 call->state = RX_STATE_DALLY;
2754 MUTEX_ENTER(&rx_freeCallQueue_lock);
2755 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2756 #ifdef RX_ENABLE_LOCKS
2757 /* A call may be free even though its transmit queue is still in use.
2758 * Since we search the call list from head to tail, put busy calls at
2759 * the head of the list, and idle calls at the tail.
2761 if (call->flags & RX_CALL_TQ_BUSY)
2762 opr_queue_Prepend(&rx_freeCallQueue, &call->entry);
2764 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2765 #else /* RX_ENABLE_LOCKS */
2766 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2767 #endif /* RX_ENABLE_LOCKS */
2768 if (rx_stats_active)
2769 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2770 MUTEX_EXIT(&rx_freeCallQueue_lock);
2772 /* Destroy the connection if it was previously slated for
2773 * destruction, i.e. the Rx client code previously called
2774 * rx_DestroyConnection (client connections), or
2775 * rxi_ReapConnections called the same routine (server
2776 * connections). Only do this, however, if there are no
2777 * outstanding calls. Note that for fine grain locking, there appears
2778 * to be a deadlock in that rxi_FreeCall has a call locked and
2779 * DestroyConnectionNoLock locks each call in the conn. But note a
2780 * few lines up where we have removed this call from the conn.
2781 * If someone else destroys a connection, they either have no
2782 * call lock held or are going through this section of code.
2784 MUTEX_ENTER(&conn->conn_data_lock);
2785 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2786 MUTEX_ENTER(&rx_refcnt_mutex);
2788 MUTEX_EXIT(&rx_refcnt_mutex);
2789 MUTEX_EXIT(&conn->conn_data_lock);
2790 #ifdef RX_ENABLE_LOCKS
2792 rxi_DestroyConnectionNoLock(conn);
2794 rxi_DestroyConnection(conn);
2795 #else /* RX_ENABLE_LOCKS */
2796 rxi_DestroyConnection(conn);
2797 #endif /* RX_ENABLE_LOCKS */
2799 MUTEX_EXIT(&conn->conn_data_lock);
2801 MUTEX_ENTER(&rx_refcnt_mutex);
2805 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2806 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2809 rxi_Alloc(size_t size)
2813 if (rx_stats_active) {
2814 rx_atomic_add(&rxi_Allocsize, (int) size);
2815 rx_atomic_inc(&rxi_Alloccnt);
2819 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2820 afs_osi_Alloc_NoSleep(size);
2825 osi_Panic("rxi_Alloc error");
2831 rxi_Free(void *addr, size_t size)
2833 if (rx_stats_active) {
2834 rx_atomic_sub(&rxi_Allocsize, (int) size);
2835 rx_atomic_dec(&rxi_Alloccnt);
2837 osi_Free(addr, size);
2841 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2843 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2844 struct rx_peer *next = NULL;
2848 MUTEX_ENTER(&rx_peerHashTable_lock);
2850 peer_ptr = &rx_peerHashTable[0];
2851 peer_end = &rx_peerHashTable[rx_hashTableSize];
2854 for ( ; peer_ptr < peer_end; peer_ptr++) {
2857 for ( ; peer; peer = next) {
2859 if (host == peer->host)
2864 hashIndex = PEER_HASH(host, port);
2865 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2866 if ((peer->host == host) && (peer->port == port))
2871 MUTEX_ENTER(&rx_peerHashTable_lock);
2876 MUTEX_EXIT(&rx_peerHashTable_lock);
2878 MUTEX_ENTER(&peer->peer_lock);
2879 /* We don't handle dropping below min, so don't */
2880 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2881 peer->ifMTU=MIN(mtu, peer->ifMTU);
2882 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2883 /* if we tweaked this down, need to tune our peer MTU too */
2884 peer->MTU = MIN(peer->MTU, peer->natMTU);
2885 /* if we discovered a sub-1500 mtu, degrade */
2886 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2887 peer->maxDgramPackets = 1;
2888 /* We no longer have valid peer packet information */
2889 if (peer->maxPacketSize + RX_HEADER_SIZE > peer->ifMTU)
2890 peer->maxPacketSize = 0;
2891 MUTEX_EXIT(&peer->peer_lock);
2893 MUTEX_ENTER(&rx_peerHashTable_lock);
2895 if (host && !port) {
2897 /* pick up where we left off */
2901 MUTEX_EXIT(&rx_peerHashTable_lock);
2904 #ifdef AFS_RXERRQ_ENV
2906 rxi_SetPeerDead(struct sock_extended_err *err, afs_uint32 host, afs_uint16 port)
2908 int hashIndex = PEER_HASH(host, port);
2909 struct rx_peer *peer;
2911 MUTEX_ENTER(&rx_peerHashTable_lock);
2913 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2914 if (peer->host == host && peer->port == port) {
2920 MUTEX_EXIT(&rx_peerHashTable_lock);
2923 rx_atomic_inc(&peer->neterrs);
2924 MUTEX_ENTER(&peer->peer_lock);
2925 peer->last_err_origin = RX_NETWORK_ERROR_ORIGIN_ICMP;
2926 peer->last_err_type = err->ee_type;
2927 peer->last_err_code = err->ee_code;
2928 MUTEX_EXIT(&peer->peer_lock);
2930 MUTEX_ENTER(&rx_peerHashTable_lock);
2932 MUTEX_EXIT(&rx_peerHashTable_lock);
2937 rxi_ProcessNetError(struct sock_extended_err *err, afs_uint32 addr, afs_uint16 port)
2939 # ifdef AFS_ADAPT_PMTU
2940 if (err->ee_errno == EMSGSIZE && err->ee_info >= 68) {
2941 rxi_SetPeerMtu(NULL, addr, port, err->ee_info - RX_IPUDP_SIZE);
2945 if (err->ee_origin == SO_EE_ORIGIN_ICMP && err->ee_type == ICMP_DEST_UNREACH) {
2946 switch (err->ee_code) {
2947 case ICMP_NET_UNREACH:
2948 case ICMP_HOST_UNREACH:
2949 case ICMP_PORT_UNREACH:
2952 rxi_SetPeerDead(err, addr, port);
2959 rxi_TranslateICMP(int type, int code)
2962 case ICMP_DEST_UNREACH:
2964 case ICMP_NET_UNREACH:
2965 return "Destination Net Unreachable";
2966 case ICMP_HOST_UNREACH:
2967 return "Destination Host Unreachable";
2968 case ICMP_PROT_UNREACH:
2969 return "Destination Protocol Unreachable";
2970 case ICMP_PORT_UNREACH:
2971 return "Destination Port Unreachable";
2973 return "Destination Net Prohibited";
2975 return "Destination Host Prohibited";
2981 #endif /* AFS_RXERRQ_ENV */
2984 * Get the last network error for a connection
2986 * A "network error" here means an error retrieved from ICMP, or some other
2987 * mechanism outside of Rx that informs us of errors in network reachability.
2989 * If a peer associated with the given Rx connection has received a network
2990 * error recently, this function allows the caller to know what error
2991 * specifically occurred. This can be useful to know, since e.g. ICMP errors
2992 * can cause calls to that peer to be quickly aborted. So, this function can
2993 * help see why a call was aborted due to network errors.
2995 * If we have received traffic from a peer since the last network error, we
2996 * treat that peer as if we had not received an network error for it.
2998 * @param[in] conn The Rx connection to examine
2999 * @param[out] err_origin The origin of the last network error (e.g. ICMP);
3000 * one of the RX_NETWORK_ERROR_ORIGIN_* constants
3001 * @param[out] err_type The type of the last error
3002 * @param[out] err_code The code of the last error
3003 * @param[out] msg Human-readable error message, if applicable; NULL otherwise
3005 * @return If we have an error
3006 * @retval -1 No error to get; 'out' params are undefined
3007 * @retval 0 We have an error; 'out' params contain the last error
3010 rx_GetNetworkError(struct rx_connection *conn, int *err_origin, int *err_type,
3011 int *err_code, const char **msg)
3013 #ifdef AFS_RXERRQ_ENV
3014 struct rx_peer *peer = conn->peer;
3015 if (rx_atomic_read(&peer->neterrs)) {
3016 MUTEX_ENTER(&peer->peer_lock);
3017 *err_origin = peer->last_err_origin;
3018 *err_type = peer->last_err_type;
3019 *err_code = peer->last_err_code;
3020 MUTEX_EXIT(&peer->peer_lock);
3023 if (*err_origin == RX_NETWORK_ERROR_ORIGIN_ICMP) {
3024 *msg = rxi_TranslateICMP(*err_type, *err_code);
3033 /* Find the peer process represented by the supplied (host,port)
3034 * combination. If there is no appropriate active peer structure, a
3035 * new one will be allocated and initialized
3038 rxi_FindPeer(afs_uint32 host, u_short port, int create)
3042 hashIndex = PEER_HASH(host, port);
3043 MUTEX_ENTER(&rx_peerHashTable_lock);
3044 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
3045 if ((pp->host == host) && (pp->port == port))
3050 pp = rxi_AllocPeer(); /* This bzero's *pp */
3051 pp->host = host; /* set here or in InitPeerParams is zero */
3053 #ifdef AFS_RXERRQ_ENV
3054 rx_atomic_set(&pp->neterrs, 0);
3056 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
3057 opr_queue_Init(&pp->rpcStats);
3058 pp->next = rx_peerHashTable[hashIndex];
3059 rx_peerHashTable[hashIndex] = pp;
3060 rxi_InitPeerParams(pp);
3061 if (rx_stats_active)
3062 rx_atomic_inc(&rx_stats.nPeerStructs);
3068 MUTEX_EXIT(&rx_peerHashTable_lock);
3073 /* Find the connection at (host, port) started at epoch, and with the
3074 * given connection id. Creates the server connection if necessary.
3075 * The type specifies whether a client connection or a server
3076 * connection is desired. In both cases, (host, port) specify the
3077 * peer's (host, pair) pair. Client connections are not made
3078 * automatically by this routine. The parameter socket gives the
3079 * socket descriptor on which the packet was received. This is used,
3080 * in the case of server connections, to check that *new* connections
3081 * come via a valid (port, serviceId). Finally, the securityIndex
3082 * parameter must match the existing index for the connection. If a
3083 * server connection is created, it will be created using the supplied
3084 * index, if the index is valid for this service */
3085 static struct rx_connection *
3086 rxi_FindConnection(osi_socket socket, afs_uint32 host,
3087 u_short port, u_short serviceId, afs_uint32 cid,
3088 afs_uint32 epoch, int type, u_int securityIndex,
3089 int *unknownService)
3091 int hashindex, flag, i;
3092 struct rx_connection *conn;
3093 *unknownService = 0;
3094 hashindex = CONN_HASH(host, port, cid, epoch, type);
3095 MUTEX_ENTER(&rx_connHashTable_lock);
3096 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
3097 rx_connHashTable[hashindex],
3100 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
3101 && (epoch == conn->epoch)) {
3102 struct rx_peer *pp = conn->peer;
3103 if (securityIndex != conn->securityIndex) {
3104 /* this isn't supposed to happen, but someone could forge a packet
3105 * like this, and there seems to be some CM bug that makes this
3106 * happen from time to time -- in which case, the fileserver
3108 MUTEX_EXIT(&rx_connHashTable_lock);
3109 return (struct rx_connection *)0;
3111 if (pp->host == host && pp->port == port)
3113 if (type == RX_CLIENT_CONNECTION && pp->port == port)
3115 /* So what happens when it's a callback connection? */
3116 if ( /*type == RX_CLIENT_CONNECTION && */
3117 (conn->epoch & 0x80000000))
3121 /* the connection rxLastConn that was used the last time is not the
3122 ** one we are looking for now. Hence, start searching in the hash */
3124 conn = rx_connHashTable[hashindex];
3129 struct rx_service *service;
3130 if (type == RX_CLIENT_CONNECTION) {
3131 MUTEX_EXIT(&rx_connHashTable_lock);
3132 return (struct rx_connection *)0;
3134 service = rxi_FindService(socket, serviceId);
3135 if (!service || (securityIndex >= service->nSecurityObjects)
3136 || (service->securityObjects[securityIndex] == 0)) {
3137 MUTEX_EXIT(&rx_connHashTable_lock);
3138 *unknownService = 1;
3139 return (struct rx_connection *)0;
3141 conn = rxi_AllocConnection(); /* This bzero's the connection */
3142 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
3143 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
3144 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
3145 conn->next = rx_connHashTable[hashindex];
3146 rx_connHashTable[hashindex] = conn;
3147 conn->peer = rxi_FindPeer(host, port, 1);
3148 conn->type = RX_SERVER_CONNECTION;
3149 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
3150 conn->epoch = epoch;
3151 conn->cid = cid & RX_CIDMASK;
3152 conn->ackRate = RX_FAST_ACK_RATE;
3153 conn->service = service;
3154 conn->serviceId = serviceId;
3155 conn->securityIndex = securityIndex;
3156 conn->securityObject = service->securityObjects[securityIndex];
3157 conn->nSpecific = 0;
3158 conn->specific = NULL;
3159 rx_SetConnDeadTime(conn, service->connDeadTime);
3160 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
3161 for (i = 0; i < RX_MAXCALLS; i++) {
3162 conn->twind[i] = rx_initSendWindow;
3163 conn->rwind[i] = rx_initReceiveWindow;
3165 /* Notify security object of the new connection */
3166 RXS_NewConnection(conn->securityObject, conn);
3167 /* XXXX Connection timeout? */
3168 if (service->newConnProc)
3169 (*service->newConnProc) (conn);
3170 if (rx_stats_active)
3171 rx_atomic_inc(&rx_stats.nServerConns);
3174 MUTEX_ENTER(&rx_refcnt_mutex);
3176 MUTEX_EXIT(&rx_refcnt_mutex);
3178 rxLastConn = conn; /* store this connection as the last conn used */
3179 MUTEX_EXIT(&rx_connHashTable_lock);
3184 * Timeout a call on a busy call channel if appropriate.
3186 * @param[in] call The busy call.
3188 * @pre 'call' is marked as busy (namely,
3189 * call->conn->lastBusy[call->channel] != 0)
3191 * @pre call->lock is held
3192 * @pre rxi_busyChannelError is nonzero
3194 * @note call->lock is dropped and reacquired
3197 rxi_CheckBusy(struct rx_call *call)
3199 struct rx_connection *conn = call->conn;
3200 int channel = call->channel;
3201 int freechannel = 0;
3204 MUTEX_EXIT(&call->lock);
3206 MUTEX_ENTER(&conn->conn_call_lock);
3208 /* Are there any other call slots on this conn that we should try? Look for
3209 * slots that are empty and are either non-busy, or were marked as busy
3210 * longer than conn->secondsUntilDead seconds before this call started. */
3212 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3214 /* only look at channels that aren't us */
3218 if (conn->lastBusy[i]) {
3219 /* if this channel looked busy too recently, don't look at it */
3220 if (conn->lastBusy[i] >= call->startTime.sec) {
3223 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3228 if (conn->call[i]) {
3229 struct rx_call *tcall = conn->call[i];
3230 MUTEX_ENTER(&tcall->lock);
3231 if (tcall->state == RX_STATE_DALLY) {
3234 MUTEX_EXIT(&tcall->lock);
3240 MUTEX_ENTER(&call->lock);
3242 /* Since the call->lock has been released it is possible that the call may
3243 * no longer be busy (the call channel cannot have been reallocated as we
3244 * haven't dropped the conn_call_lock) Therefore, we must confirm
3245 * that the call state has not changed when deciding whether or not to
3246 * force this application thread to retry by forcing a Timeout error. */
3248 if (freechannel && (call->flags & RX_CALL_PEER_BUSY)) {
3249 /* Since 'freechannel' is set, there exists another channel in this
3250 * rx_conn that the application thread might be able to use. We know
3251 * that we have the correct call since callNumber is unchanged, and we
3252 * know that the call is still busy. So, set the call error state to
3253 * rxi_busyChannelError so the application can retry the request,
3254 * presumably on a less-busy call channel. */
3256 rxi_CallError(call, RX_CALL_BUSY);
3258 MUTEX_EXIT(&conn->conn_call_lock);
3262 * Abort the call if the server is over the busy threshold. This
3263 * can be used without requiring a call structure be initialised,
3264 * or connected to a particular channel
3267 rxi_AbortIfServerBusy(osi_socket socket, struct rx_connection *conn,
3268 struct rx_packet *np)
3270 if ((rx_BusyThreshold > 0) &&
3271 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3272 rxi_SendRawAbort(socket, conn->peer->host, conn->peer->port,
3273 rx_BusyError, np, 0);
3274 if (rx_stats_active)
3275 rx_atomic_inc(&rx_stats.nBusies);
3282 static_inline struct rx_call *
3283 rxi_ReceiveClientCall(struct rx_packet *np, struct rx_connection *conn)
3286 struct rx_call *call;
3288 channel = np->header.cid & RX_CHANNELMASK;
3289 MUTEX_ENTER(&conn->conn_call_lock);
3290 call = conn->call[channel];
3291 if (!call || conn->callNumber[channel] != np->header.callNumber) {
3292 MUTEX_EXIT(&conn->conn_call_lock);
3293 if (rx_stats_active)
3294 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3298 MUTEX_ENTER(&call->lock);
3299 MUTEX_EXIT(&conn->conn_call_lock);
3301 if ((call->state == RX_STATE_DALLY)
3302 && np->header.type == RX_PACKET_TYPE_ACK) {
3303 if (rx_stats_active)
3304 rx_atomic_inc(&rx_stats.ignorePacketDally);
3305 MUTEX_EXIT(&call->lock);
3312 static_inline struct rx_call *
3313 rxi_ReceiveServerCall(osi_socket socket, struct rx_packet *np,
3314 struct rx_connection *conn)
3317 struct rx_call *call;
3319 channel = np->header.cid & RX_CHANNELMASK;
3320 MUTEX_ENTER(&conn->conn_call_lock);
3321 call = conn->call[channel];
3324 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3325 MUTEX_EXIT(&conn->conn_call_lock);
3329 call = rxi_NewCall(conn, channel); /* returns locked call */
3330 *call->callNumber = np->header.callNumber;
3331 MUTEX_EXIT(&conn->conn_call_lock);
3333 call->state = RX_STATE_PRECALL;
3334 clock_GetTime(&call->queueTime);
3335 call->app.bytesSent = 0;
3336 call->app.bytesRcvd = 0;
3337 rxi_KeepAliveOn(call);
3342 if (np->header.callNumber == conn->callNumber[channel]) {
3343 MUTEX_ENTER(&call->lock);
3344 MUTEX_EXIT(&conn->conn_call_lock);
3348 if (np->header.callNumber < conn->callNumber[channel]) {
3349 MUTEX_EXIT(&conn->conn_call_lock);
3350 if (rx_stats_active)
3351 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3355 MUTEX_ENTER(&call->lock);
3356 MUTEX_EXIT(&conn->conn_call_lock);
3358 /* Wait until the transmit queue is idle before deciding
3359 * whether to reset the current call. Chances are that the
3360 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3363 #ifdef RX_ENABLE_LOCKS
3364 if (call->state == RX_STATE_ACTIVE && !call->error) {
3365 rxi_WaitforTQBusy(call);
3366 /* If we entered error state while waiting,
3367 * must call rxi_CallError to permit rxi_ResetCall
3368 * to processed when the tqWaiter count hits zero.
3371 rxi_CallError(call, call->error);
3372 MUTEX_EXIT(&call->lock);
3376 #endif /* RX_ENABLE_LOCKS */
3377 /* If the new call cannot be taken right now send a busy and set
3378 * the error condition in this call, so that it terminates as
3379 * quickly as possible */
3380 if (call->state == RX_STATE_ACTIVE) {
3381 rxi_CallError(call, RX_CALL_DEAD);
3382 rxi_SendSpecial(call, conn, NULL, RX_PACKET_TYPE_BUSY,
3384 MUTEX_EXIT(&call->lock);
3388 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3389 MUTEX_EXIT(&call->lock);
3393 rxi_ResetCall(call, 0);
3394 /* The conn_call_lock is not held but no one else should be
3395 * using this call channel while we are processing this incoming
3396 * packet. This assignment should be safe.
3398 *call->callNumber = np->header.callNumber;
3399 call->state = RX_STATE_PRECALL;
3400 clock_GetTime(&call->queueTime);
3401 call->app.bytesSent = 0;
3402 call->app.bytesRcvd = 0;
3403 rxi_KeepAliveOn(call);
3409 /* There are two packet tracing routines available for testing and monitoring
3410 * Rx. One is called just after every packet is received and the other is
3411 * called just before every packet is sent. Received packets, have had their
3412 * headers decoded, and packets to be sent have not yet had their headers
3413 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3414 * containing the network address. Both can be modified. The return value, if
3415 * non-zero, indicates that the packet should be dropped. */
3417 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3418 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3420 /* A packet has been received off the interface. Np is the packet, socket is
3421 * the socket number it was received from (useful in determining which service
3422 * this packet corresponds to), and (host, port) reflect the host,port of the
3423 * sender. This call returns the packet to the caller if it is finished with
3424 * it, rather than de-allocating it, just as a small performance hack */
3427 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3428 afs_uint32 host, u_short port, int *tnop,
3429 struct rx_call **newcallp)
3431 struct rx_call *call;
3432 struct rx_connection *conn;
3434 int unknownService = 0;
3438 struct rx_packet *tnp;
3441 /* We don't print out the packet until now because (1) the time may not be
3442 * accurate enough until now in the lwp implementation (rx_Listener only gets
3443 * the time after the packet is read) and (2) from a protocol point of view,
3444 * this is the first time the packet has been seen */
3445 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3446 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3447 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3448 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3449 np->header.epoch, np->header.cid, np->header.callNumber,
3450 np->header.seq, np->header.flags, np));
3453 /* Account for connectionless packets */
3454 if (rx_stats_active &&
3455 ((np->header.type == RX_PACKET_TYPE_VERSION) ||
3456 (np->header.type == RX_PACKET_TYPE_DEBUG))) {
3457 struct rx_peer *peer;
3459 /* Try to look up the peer structure, but don't create one */
3460 peer = rxi_FindPeer(host, port, 0);
3462 /* Since this may not be associated with a connection, it may have
3463 * no refCount, meaning we could race with ReapConnections
3466 if (peer && (peer->refCount > 0)) {
3467 #ifdef AFS_RXERRQ_ENV
3468 if (rx_atomic_read(&peer->neterrs)) {
3469 rx_atomic_set(&peer->neterrs, 0);
3472 MUTEX_ENTER(&peer->peer_lock);
3473 peer->bytesReceived += np->length;
3474 MUTEX_EXIT(&peer->peer_lock);
3478 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3479 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3482 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3483 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3486 /* If an input tracer function is defined, call it with the packet and
3487 * network address. Note this function may modify its arguments. */
3488 if (rx_justReceived) {
3489 struct sockaddr_in addr;
3491 addr.sin_family = AF_INET;
3492 addr.sin_port = port;
3493 addr.sin_addr.s_addr = host;
3494 memset(&addr.sin_zero, 0, sizeof(addr.sin_zero));
3495 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3496 addr.sin_len = sizeof(addr);
3497 #endif /* AFS_OSF_ENV */
3498 drop = (*rx_justReceived) (np, &addr);
3499 /* drop packet if return value is non-zero */
3502 port = addr.sin_port; /* in case fcn changed addr */
3503 host = addr.sin_addr.s_addr;
3507 /* If packet was not sent by the client, then *we* must be the client */
3508 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3509 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3511 /* Find the connection (or fabricate one, if we're the server & if
3512 * necessary) associated with this packet */
3514 rxi_FindConnection(socket, host, port, np->header.serviceId,
3515 np->header.cid, np->header.epoch, type,
3516 np->header.securityIndex, &unknownService);
3518 /* To avoid having 2 connections just abort at each other,
3519 don't abort an abort. */
3521 if (unknownService && (np->header.type != RX_PACKET_TYPE_ABORT))
3522 rxi_SendRawAbort(socket, host, port, RX_INVALID_OPERATION,
3527 #ifdef AFS_RXERRQ_ENV
3528 if (rx_atomic_read(&conn->peer->neterrs)) {
3529 rx_atomic_set(&conn->peer->neterrs, 0);
3533 /* If we're doing statistics, then account for the incoming packet */
3534 if (rx_stats_active) {
3535 MUTEX_ENTER(&conn->peer->peer_lock);
3536 conn->peer->bytesReceived += np->length;
3537 MUTEX_EXIT(&conn->peer->peer_lock);
3540 /* If the connection is in an error state, send an abort packet and ignore
3541 * the incoming packet */
3543 /* Don't respond to an abort packet--we don't want loops! */
3544 MUTEX_ENTER(&conn->conn_data_lock);
3545 if (np->header.type != RX_PACKET_TYPE_ABORT)
3546 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3547 putConnection(conn);
3548 MUTEX_EXIT(&conn->conn_data_lock);
3552 /* Check for connection-only requests (i.e. not call specific). */
3553 if (np->header.callNumber == 0) {
3554 switch (np->header.type) {
3555 case RX_PACKET_TYPE_ABORT: {
3556 /* What if the supplied error is zero? */
3557 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3558 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3559 rxi_ConnectionError(conn, errcode);
3560 putConnection(conn);
3563 case RX_PACKET_TYPE_CHALLENGE:
3564 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3565 putConnection(conn);
3567 case RX_PACKET_TYPE_RESPONSE:
3568 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3569 putConnection(conn);
3571 case RX_PACKET_TYPE_PARAMS:
3572 case RX_PACKET_TYPE_PARAMS + 1:
3573 case RX_PACKET_TYPE_PARAMS + 2:
3574 /* ignore these packet types for now */
3575 putConnection(conn);
3579 /* Should not reach here, unless the peer is broken: send an
3581 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3582 MUTEX_ENTER(&conn->conn_data_lock);
3583 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3584 putConnection(conn);
3585 MUTEX_EXIT(&conn->conn_data_lock);
3590 if (type == RX_SERVER_CONNECTION)
3591 call = rxi_ReceiveServerCall(socket, np, conn);
3593 call = rxi_ReceiveClientCall(np, conn);
3596 putConnection(conn);
3600 MUTEX_ASSERT(&call->lock);
3601 /* Set remote user defined status from packet */
3602 call->remoteStatus = np->header.userStatus;
3604 /* Now do packet type-specific processing */
3605 switch (np->header.type) {
3606 case RX_PACKET_TYPE_DATA:
3607 /* If we're a client, and receiving a response, then all the packets
3608 * we transmitted packets are implicitly acknowledged. */
3609 if (type == RX_CLIENT_CONNECTION && !opr_queue_IsEmpty(&call->tq))
3610 rxi_AckAllInTransmitQueue(call);
3612 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3615 case RX_PACKET_TYPE_ACK:
3616 /* Respond immediately to ack packets requesting acknowledgement
3618 if (np->header.flags & RX_REQUEST_ACK) {
3620 (void)rxi_SendCallAbort(call, 0, 1, 0);
3622 (void)rxi_SendAck(call, 0, np->header.serial,
3623 RX_ACK_PING_RESPONSE, 1);
3625 np = rxi_ReceiveAckPacket(call, np, 1);
3627 case RX_PACKET_TYPE_ABORT: {
3628 /* An abort packet: reset the call, passing the error up to the user. */
3629 /* What if error is zero? */
3630 /* What if the error is -1? the application will treat it as a timeout. */
3631 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3632 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3633 rxi_CallError(call, errdata);
3634 MUTEX_EXIT(&call->lock);
3635 putConnection(conn);
3636 return np; /* xmitting; drop packet */
3638 case RX_PACKET_TYPE_BUSY: {
3639 struct clock busyTime;
3641 clock_GetTime(&busyTime);
3643 MUTEX_EXIT(&call->lock);
3645 MUTEX_ENTER(&conn->conn_call_lock);
3646 MUTEX_ENTER(&call->lock);
3647 conn->lastBusy[call->channel] = busyTime.sec;
3648 call->flags |= RX_CALL_PEER_BUSY;
3649 MUTEX_EXIT(&call->lock);
3650 MUTEX_EXIT(&conn->conn_call_lock);
3652 putConnection(conn);
3656 case RX_PACKET_TYPE_ACKALL:
3657 /* All packets acknowledged, so we can drop all packets previously
3658 * readied for sending */
3659 rxi_AckAllInTransmitQueue(call);
3662 /* Should not reach here, unless the peer is broken: send an abort
3664 rxi_CallError(call, RX_PROTOCOL_ERROR);
3665 np = rxi_SendCallAbort(call, np, 1, 0);
3668 /* Note when this last legitimate packet was received, for keep-alive
3669 * processing. Note, we delay getting the time until now in the hope that
3670 * the packet will be delivered to the user before any get time is required
3671 * (if not, then the time won't actually be re-evaluated here). */
3672 call->lastReceiveTime = clock_Sec();
3673 /* we've received a legit packet, so the channel is not busy */
3674 call->flags &= ~RX_CALL_PEER_BUSY;
3675 MUTEX_EXIT(&call->lock);
3676 putConnection(conn);
3680 /* return true if this is an "interesting" connection from the point of view
3681 of someone trying to debug the system */
3683 rxi_IsConnInteresting(struct rx_connection *aconn)
3686 struct rx_call *tcall;
3688 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3691 for (i = 0; i < RX_MAXCALLS; i++) {
3692 tcall = aconn->call[i];
3694 if ((tcall->state == RX_STATE_PRECALL)
3695 || (tcall->state == RX_STATE_ACTIVE))
3697 if ((tcall->app.mode == RX_MODE_SENDING)
3698 || (tcall->app.mode == RX_MODE_RECEIVING))
3706 /* if this is one of the last few packets AND it wouldn't be used by the
3707 receiving call to immediately satisfy a read request, then drop it on
3708 the floor, since accepting it might prevent a lock-holding thread from
3709 making progress in its reading. If a call has been cleared while in
3710 the precall state then ignore all subsequent packets until the call
3711 is assigned to a thread. */
3714 TooLow(struct rx_packet *ap, struct rx_call *acall)
3718 MUTEX_ENTER(&rx_quota_mutex);
3719 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3720 && (acall->state == RX_STATE_PRECALL))
3721 || ((rx_nFreePackets < rxi_dataQuota + 2)
3722 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3723 && (acall->flags & RX_CALL_READER_WAIT)))) {
3726 MUTEX_EXIT(&rx_quota_mutex);
3732 * Clear the attach wait flag on a connection and proceed.
3734 * Any processing waiting for a connection to be attached should be
3735 * unblocked. We clear the flag and do any other needed tasks.
3738 * the conn to unmark waiting for attach
3740 * @pre conn's conn_data_lock must be locked before calling this function
3744 rxi_ConnClearAttachWait(struct rx_connection *conn)
3746 /* Indicate that rxi_CheckReachEvent is no longer running by
3747 * clearing the flag. Must be atomic under conn_data_lock to
3748 * avoid a new call slipping by: rxi_CheckConnReach holds
3749 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3751 conn->flags &= ~RX_CONN_ATTACHWAIT;
3752 if (conn->flags & RX_CONN_NAT_PING) {
3753 conn->flags &= ~RX_CONN_NAT_PING;
3754 rxi_ScheduleNatKeepAliveEvent(conn);
3759 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3761 struct rx_connection *conn = arg1;
3762 struct rx_call *acall = arg2;
3763 struct rx_call *call = acall;
3764 struct clock when, now;
3767 MUTEX_ENTER(&conn->conn_data_lock);
3770 rxevent_Put(&conn->checkReachEvent);
3772 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3774 putConnection(conn);
3776 MUTEX_EXIT(&conn->conn_data_lock);
3780 MUTEX_ENTER(&conn->conn_call_lock);
3781 MUTEX_ENTER(&conn->conn_data_lock);
3782 for (i = 0; i < RX_MAXCALLS; i++) {
3783 struct rx_call *tc = conn->call[i];
3784 if (tc && tc->state == RX_STATE_PRECALL) {
3790 rxi_ConnClearAttachWait(conn);
3791 MUTEX_EXIT(&conn->conn_data_lock);
3792 MUTEX_EXIT(&conn->conn_call_lock);
3797 MUTEX_ENTER(&call->lock);
3798 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3800 MUTEX_EXIT(&call->lock);
3802 clock_GetTime(&now);
3804 when.sec += RX_CHECKREACH_TIMEOUT;
3805 MUTEX_ENTER(&conn->conn_data_lock);
3806 if (!conn->checkReachEvent) {
3807 MUTEX_ENTER(&rx_refcnt_mutex);
3809 MUTEX_EXIT(&rx_refcnt_mutex);
3810 conn->checkReachEvent = rxevent_Post(&when, &now,
3811 rxi_CheckReachEvent, conn,
3814 MUTEX_EXIT(&conn->conn_data_lock);
3820 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3822 struct rx_service *service = conn->service;
3823 struct rx_peer *peer = conn->peer;
3824 afs_uint32 now, lastReach;
3826 if (service->checkReach == 0)
3830 MUTEX_ENTER(&peer->peer_lock);
3831 lastReach = peer->lastReachTime;
3832 MUTEX_EXIT(&peer->peer_lock);
3833 if (now - lastReach < RX_CHECKREACH_TTL)
3836 MUTEX_ENTER(&conn->conn_data_lock);
3837 if (conn->flags & RX_CONN_ATTACHWAIT) {
3838 MUTEX_EXIT(&conn->conn_data_lock);
3841 conn->flags |= RX_CONN_ATTACHWAIT;
3842 MUTEX_EXIT(&conn->conn_data_lock);
3843 if (!conn->checkReachEvent)
3844 rxi_CheckReachEvent(NULL, conn, call, 0);
3849 /* try to attach call, if authentication is complete */
3851 TryAttach(struct rx_call *acall, osi_socket socket,
3852 int *tnop, struct rx_call **newcallp,
3855 struct rx_connection *conn = acall->conn;
3857 if (conn->type == RX_SERVER_CONNECTION
3858 && acall->state == RX_STATE_PRECALL) {
3859 /* Don't attach until we have any req'd. authentication. */
3860 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3861 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3862 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3863 /* Note: this does not necessarily succeed; there
3864 * may not any proc available
3867 rxi_ChallengeOn(acall->conn);
3872 /* A data packet has been received off the interface. This packet is
3873 * appropriate to the call (the call is in the right state, etc.). This
3874 * routine can return a packet to the caller, for re-use */
3876 static struct rx_packet *
3877 rxi_ReceiveDataPacket(struct rx_call *call,
3878 struct rx_packet *np, int istack,
3879 osi_socket socket, afs_uint32 host, u_short port,
3880 int *tnop, struct rx_call **newcallp)
3882 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3887 afs_uint32 serial=0, flags=0;
3889 struct rx_packet *tnp;
3890 if (rx_stats_active)
3891 rx_atomic_inc(&rx_stats.dataPacketsRead);
3894 /* If there are no packet buffers, drop this new packet, unless we can find
3895 * packet buffers from inactive calls */
3897 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3898 MUTEX_ENTER(&rx_freePktQ_lock);
3899 rxi_NeedMorePackets = TRUE;
3900 MUTEX_EXIT(&rx_freePktQ_lock);
3901 if (rx_stats_active)
3902 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3903 rxi_calltrace(RX_TRACE_DROP, call);
3904 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3905 /* We used to clear the receive queue here, in an attempt to free
3906 * packets. However this is unsafe if the queue has received a
3907 * soft ACK for the final packet */
3908 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3914 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3915 * packet is one of several packets transmitted as a single
3916 * datagram. Do not send any soft or hard acks until all packets
3917 * in a jumbogram have been processed. Send negative acks right away.
3919 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3920 /* tnp is non-null when there are more packets in the
3921 * current jumbo gram */
3928 seq = np->header.seq;
3929 serial = np->header.serial;
3930 flags = np->header.flags;
3932 /* If the call is in an error state, send an abort message */
3934 return rxi_SendCallAbort(call, np, istack, 0);
3936 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3937 * AFS 3.5 jumbogram. */
3938 if (flags & RX_JUMBO_PACKET) {
3939 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3944 if (np->header.spare != 0) {
3945 MUTEX_ENTER(&call->conn->conn_data_lock);
3946 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3947 MUTEX_EXIT(&call->conn->conn_data_lock);
3950 /* The usual case is that this is the expected next packet */
3951 if (seq == call->rnext) {
3953 /* Check to make sure it is not a duplicate of one already queued */
3954 if (!opr_queue_IsEmpty(&call->rq)
3955 && opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq == seq) {
3956 if (rx_stats_active)
3957 rx_atomic_inc(&rx_stats.dupPacketsRead);
3958 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3959 rxi_CancelDelayedAckEvent(call);
3960 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3966 /* It's the next packet. Stick it on the receive queue
3967 * for this call. Set newPackets to make sure we wake
3968 * the reader once all packets have been processed */
3969 #ifdef RX_TRACK_PACKETS
3970 np->flags |= RX_PKTFLAG_RQ;
3972 opr_queue_Prepend(&call->rq, &np->entry);
3973 #ifdef RXDEBUG_PACKET
3975 #endif /* RXDEBUG_PACKET */
3977 np = NULL; /* We can't use this anymore */
3980 /* If an ack is requested then set a flag to make sure we
3981 * send an acknowledgement for this packet */
3982 if (flags & RX_REQUEST_ACK) {
3983 ackNeeded = RX_ACK_REQUESTED;
3986 /* Keep track of whether we have received the last packet */
3987 if (flags & RX_LAST_PACKET) {
3988 call->flags |= RX_CALL_HAVE_LAST;
3992 /* Check whether we have all of the packets for this call */
3993 if (call->flags & RX_CALL_HAVE_LAST) {
3994 afs_uint32 tseq; /* temporary sequence number */
3995 struct opr_queue *cursor;
3997 for (tseq = seq, opr_queue_Scan(&call->rq, cursor)) {
3998 struct rx_packet *tp;
4000 tp = opr_queue_Entry(cursor, struct rx_packet, entry);
4001 if (tseq != tp->header.seq)
4003 if (tp->header.flags & RX_LAST_PACKET) {
4004 call->flags |= RX_CALL_RECEIVE_DONE;
4011 /* Provide asynchronous notification for those who want it
4012 * (e.g. multi rx) */
4013 if (call->arrivalProc) {
4014 (*call->arrivalProc) (call, call->arrivalProcHandle,
4015 call->arrivalProcArg);
4016 call->arrivalProc = (void (*)())0;
4019 /* Update last packet received */
4022 /* If there is no server process serving this call, grab
4023 * one, if available. We only need to do this once. If a
4024 * server thread is available, this thread becomes a server
4025 * thread and the server thread becomes a listener thread. */
4027 TryAttach(call, socket, tnop, newcallp, 0);
4030 /* This is not the expected next packet. */
4032 /* Determine whether this is a new or old packet, and if it's
4033 * a new one, whether it fits into the current receive window.
4034 * Also figure out whether the packet was delivered in sequence.
4035 * We use the prev variable to determine whether the new packet
4036 * is the successor of its immediate predecessor in the
4037 * receive queue, and the missing flag to determine whether
4038 * any of this packets predecessors are missing. */
4040 afs_uint32 prev; /* "Previous packet" sequence number */
4041 struct opr_queue *cursor;
4042 int missing; /* Are any predecessors missing? */
4044 /* If the new packet's sequence number has been sent to the
4045 * application already, then this is a duplicate */
4046 if (seq < call->rnext) {
4047 if (rx_stats_active)
4048 rx_atomic_inc(&rx_stats.dupPacketsRead);
4049 rxi_CancelDelayedAckEvent(call);
4050 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
4056 /* If the sequence number is greater than what can be
4057 * accomodated by the current window, then send a negative
4058 * acknowledge and drop the packet */
4059 if ((call->rnext + call->rwind) <= seq) {
4060 rxi_CancelDelayedAckEvent(call);
4061 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
4068 /* Look for the packet in the queue of old received packets */
4069 prev = call->rnext - 1;
4071 for (opr_queue_Scan(&call->rq, cursor)) {
4072 struct rx_packet *tp
4073 = opr_queue_Entry(cursor, struct rx_packet, entry);
4075 /*Check for duplicate packet */
4076 if (seq == tp->header.seq) {
4077 if (rx_stats_active)
4078 rx_atomic_inc(&rx_stats.dupPacketsRead);
4079 rxi_CancelDelayedAckEvent(call);
4080 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
4086 /* If we find a higher sequence packet, break out and
4087 * insert the new packet here. */
4088 if (seq < tp->header.seq)
4090 /* Check for missing packet */
4091 if (tp->header.seq != prev + 1) {
4095 prev = tp->header.seq;
4098 /* Keep track of whether we have received the last packet. */
4099 if (flags & RX_LAST_PACKET) {
4100 call->flags |= RX_CALL_HAVE_LAST;
4103 /* It's within the window: add it to the the receive queue.
4104 * tp is left by the previous loop either pointing at the
4105 * packet before which to insert the new packet, or at the
4106 * queue head if the queue is empty or the packet should be
4108 #ifdef RX_TRACK_PACKETS
4109 np->flags |= RX_PKTFLAG_RQ;
4111 #ifdef RXDEBUG_PACKET
4113 #endif /* RXDEBUG_PACKET */
4114 opr_queue_InsertBefore(cursor, &np->entry);
4118 /* Check whether we have all of the packets for this call */
4119 if ((call->flags & RX_CALL_HAVE_LAST)
4120 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4121 afs_uint32 tseq; /* temporary sequence number */
4124 for (opr_queue_Scan(&call->rq, cursor)) {
4125 struct rx_packet *tp
4126 = opr_queue_Entry(cursor, struct rx_packet, entry);
4127 if (tseq != tp->header.seq)
4129 if (tp->header.flags & RX_LAST_PACKET) {
4130 call->flags |= RX_CALL_RECEIVE_DONE;
4137 /* We need to send an ack of the packet is out of sequence,
4138 * or if an ack was requested by the peer. */
4139 if (seq != prev + 1 || missing) {
4140 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4141 } else if (flags & RX_REQUEST_ACK) {
4142 ackNeeded = RX_ACK_REQUESTED;
4145 /* Acknowledge the last packet for each call */
4146 if (flags & RX_LAST_PACKET) {
4157 * If the receiver is waiting for an iovec, fill the iovec
4158 * using the data from the receive queue */
4159 if (call->flags & RX_CALL_IOVEC_WAIT) {
4160 didHardAck = rxi_FillReadVec(call, serial);
4161 /* the call may have been aborted */
4170 /* Wakeup the reader if any */
4171 if ((call->flags & RX_CALL_READER_WAIT)
4172 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4173 || (call->iovNext >= call->iovMax)
4174 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4175 call->flags &= ~RX_CALL_READER_WAIT;
4176 #ifdef RX_ENABLE_LOCKS
4177 CV_BROADCAST(&call->cv_rq);
4179 osi_rxWakeup(&call->rq);
4185 * Send an ack when requested by the peer, or once every
4186 * rxi_SoftAckRate packets until the last packet has been
4187 * received. Always send a soft ack for the last packet in
4188 * the server's reply. */
4190 rxi_CancelDelayedAckEvent(call);
4191 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4192 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4193 rxi_CancelDelayedAckEvent(call);
4194 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4195 } else if (call->nSoftAcks) {
4196 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4197 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4199 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4200 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4201 rxi_CancelDelayedAckEvent(call);
4208 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4210 struct rx_peer *peer = conn->peer;
4212 MUTEX_ENTER(&peer->peer_lock);
4213 peer->lastReachTime = clock_Sec();
4214 MUTEX_EXIT(&peer->peer_lock);
4216 MUTEX_ENTER(&conn->conn_data_lock);
4217 if (conn->flags & RX_CONN_ATTACHWAIT) {
4220 rxi_ConnClearAttachWait(conn);
4221 MUTEX_EXIT(&conn->conn_data_lock);
4223 for (i = 0; i < RX_MAXCALLS; i++) {
4224 struct rx_call *call = conn->call[i];
4227 MUTEX_ENTER(&call->lock);
4228 /* tnop can be null if newcallp is null */
4229 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4231 MUTEX_EXIT(&call->lock);
4235 MUTEX_EXIT(&conn->conn_data_lock);
4238 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4240 rx_ack_reason(int reason)
4243 case RX_ACK_REQUESTED:
4245 case RX_ACK_DUPLICATE:
4247 case RX_ACK_OUT_OF_SEQUENCE:
4249 case RX_ACK_EXCEEDS_WINDOW:
4251 case RX_ACK_NOSPACE:
4255 case RX_ACK_PING_RESPONSE:
4268 /* The real smarts of the whole thing. */
4269 static struct rx_packet *
4270 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4273 struct rx_ackPacket *ap;
4275 struct rx_packet *tp;
4276 struct rx_connection *conn = call->conn;
4277 struct rx_peer *peer = conn->peer;
4278 struct opr_queue *cursor;
4279 struct clock now; /* Current time, for RTT calculations */
4287 int newAckCount = 0;
4288 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4289 int pktsize = 0; /* Set if we need to update the peer mtu */
4290 int conn_data_locked = 0;
4292 if (rx_stats_active)
4293 rx_atomic_inc(&rx_stats.ackPacketsRead);
4294 ap = (struct rx_ackPacket *)rx_DataOf(np);
4295 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4297 return np; /* truncated ack packet */
4299 /* depends on ack packet struct */
4300 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4301 first = ntohl(ap->firstPacket);
4302 prev = ntohl(ap->previousPacket);
4303 serial = ntohl(ap->serial);
4306 * Ignore ack packets received out of order while protecting
4307 * against peers that set the previousPacket field to a packet
4308 * serial number instead of a sequence number.
4310 if (first < call->tfirst ||
4311 (first == call->tfirst && prev < call->tprev && prev < call->tfirst
4318 if (np->header.flags & RX_SLOW_START_OK) {
4319 call->flags |= RX_CALL_SLOW_START_OK;
4322 if (ap->reason == RX_ACK_PING_RESPONSE)
4323 rxi_UpdatePeerReach(conn, call);
4325 if (conn->lastPacketSizeSeq) {
4326 MUTEX_ENTER(&conn->conn_data_lock);
4327 conn_data_locked = 1;
4328 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4329 pktsize = conn->lastPacketSize;
4330 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4333 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4334 if (!conn_data_locked) {
4335 MUTEX_ENTER(&conn->conn_data_lock);
4336 conn_data_locked = 1;
4338 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4339 /* process mtu ping ack */
4340 pktsize = conn->lastPingSize;
4341 conn->lastPingSizeSer = conn->lastPingSize = 0;
4345 if (conn_data_locked) {
4346 MUTEX_EXIT(&conn->conn_data_lock);
4347 conn_data_locked = 0;
4351 if (rxdebug_active) {
4355 len = _snprintf(msg, sizeof(msg),
4356 "tid[%d] RACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4357 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4358 ntohl(ap->serial), ntohl(ap->previousPacket),
4359 (unsigned int)np->header.seq, ntohl(ap->firstPacket),
4360 ap->nAcks, ntohs(ap->bufferSpace) );
4364 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4365 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4369 OutputDebugString(msg);
4371 #else /* AFS_NT40_ENV */
4374 "RACK: reason %x previous %u seq %u serial %u first %u",
4375 ap->reason, ntohl(ap->previousPacket),
4376 (unsigned int)np->header.seq, (unsigned int)serial,
4377 ntohl(ap->firstPacket));
4380 for (offset = 0; offset < nAcks; offset++)
4381 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4386 #endif /* AFS_NT40_ENV */
4389 MUTEX_ENTER(&peer->peer_lock);
4392 * Start somewhere. Can't assume we can send what we can receive,
4393 * but we are clearly receiving.
4395 if (!peer->maxPacketSize)
4396 peer->maxPacketSize = RX_MIN_PACKET_SIZE - RX_HEADER_SIZE;
4398 if (pktsize > peer->maxPacketSize) {
4399 peer->maxPacketSize = pktsize;
4400 if ((pktsize + RX_HEADER_SIZE > peer->ifMTU)) {
4401 peer->ifMTU = pktsize + RX_HEADER_SIZE;
4402 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4403 rxi_ScheduleGrowMTUEvent(call, 1);
4408 clock_GetTime(&now);
4410 /* The transmit queue splits into 4 sections.
4412 * The first section is packets which have now been acknowledged
4413 * by a window size change in the ack. These have reached the
4414 * application layer, and may be discarded. These are packets
4415 * with sequence numbers < ap->firstPacket.
4417 * The second section is packets which have sequence numbers in
4418 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4419 * contents of the packet's ack array determines whether these
4420 * packets are acknowledged or not.
4422 * The third section is packets which fall above the range
4423 * addressed in the ack packet. These have not yet been received
4426 * The four section is packets which have not yet been transmitted.
4427 * These packets will have a header.serial of 0.
4430 /* First section - implicitly acknowledged packets that can be
4434 tp = opr_queue_First(&call->tq, struct rx_packet, entry);
4435 while(!opr_queue_IsEnd(&call->tq, &tp->entry) && tp->header.seq < first) {
4436 struct rx_packet *next;
4438 next = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4439 call->tfirst = tp->header.seq + 1;
4441 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4443 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4446 #ifdef RX_ENABLE_LOCKS
4447 /* XXX Hack. Because we have to release the global call lock when sending
4448 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4449 * in rxi_Start sending packets out because packets may move to the
4450 * freePacketQueue as result of being here! So we drop these packets until
4451 * we're safely out of the traversing. Really ugly!
4452 * To make it even uglier, if we're using fine grain locking, we can
4453 * set the ack bits in the packets and have rxi_Start remove the packets
4454 * when it's done transmitting.
4456 if (call->flags & RX_CALL_TQ_BUSY) {
4457 tp->flags |= RX_PKTFLAG_ACKED;
4458 call->flags |= RX_CALL_TQ_SOME_ACKED;
4460 #endif /* RX_ENABLE_LOCKS */
4462 opr_queue_Remove(&tp->entry);
4463 #ifdef RX_TRACK_PACKETS
4464 tp->flags &= ~RX_PKTFLAG_TQ;
4466 #ifdef RXDEBUG_PACKET
4468 #endif /* RXDEBUG_PACKET */
4469 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4474 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4476 /* Second section of the queue - packets for which we are receiving
4479 * Go through the explicit acks/nacks and record the results in
4480 * the waiting packets. These are packets that can't be released
4481 * yet, even with a positive acknowledge. This positive
4482 * acknowledge only means the packet has been received by the
4483 * peer, not that it will be retained long enough to be sent to
4484 * the peer's upper level. In addition, reset the transmit timers
4485 * of any missing packets (those packets that must be missing
4486 * because this packet was out of sequence) */
4488 call->nSoftAcked = 0;
4490 while (!opr_queue_IsEnd(&call->tq, &tp->entry)
4491 && tp->header.seq < first + nAcks) {
4492 /* Set the acknowledge flag per packet based on the
4493 * information in the ack packet. An acknowlegded packet can
4494 * be downgraded when the server has discarded a packet it
4495 * soacked previously, or when an ack packet is received
4496 * out of sequence. */
4497 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4498 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4500 tp->flags |= RX_PKTFLAG_ACKED;
4501 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4508 } else /* RX_ACK_TYPE_NACK */ {
4509 tp->flags &= ~RX_PKTFLAG_ACKED;
4513 tp = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4516 /* We don't need to take any action with the 3rd or 4th section in the
4517 * queue - they're not addressed by the contents of this ACK packet.
4520 /* If the window has been extended by this acknowledge packet,
4521 * then wakeup a sender waiting in alloc for window space, or try
4522 * sending packets now, if he's been sitting on packets due to
4523 * lack of window space */
4524 if (call->tnext < (call->tfirst + call->twind)) {
4525 #ifdef RX_ENABLE_LOCKS
4526 CV_SIGNAL(&call->cv_twind);
4528 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4529 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4530 osi_rxWakeup(&call->twind);
4533 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4534 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4538 /* if the ack packet has a receivelen field hanging off it,
4539 * update our state */
4540 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4543 /* If the ack packet has a "recommended" size that is less than
4544 * what I am using now, reduce my size to match */
4545 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4546 (int)sizeof(afs_int32), &tSize);
4547 tSize = (afs_uint32) ntohl(tSize);
4548 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4550 /* Get the maximum packet size to send to this peer */
4551 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4553 tSize = (afs_uint32) ntohl(tSize);
4554 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4555 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4557 /* sanity check - peer might have restarted with different params.
4558 * If peer says "send less", dammit, send less... Peer should never
4559 * be unable to accept packets of the size that prior AFS versions would
4560 * send without asking. */
4561 if (peer->maxMTU != tSize) {
4562 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4564 peer->maxMTU = tSize;
4565 peer->MTU = MIN(tSize, peer->MTU);
4566 call->MTU = MIN(call->MTU, tSize);
4569 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4572 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4573 (int)sizeof(afs_int32), &tSize);
4574 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4575 if (tSize < call->twind) { /* smaller than our send */
4576 call->twind = tSize; /* window, we must send less... */
4577 call->ssthresh = MIN(call->twind, call->ssthresh);
4578 call->conn->twind[call->channel] = call->twind;
4581 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4582 * network MTU confused with the loopback MTU. Calculate the
4583 * maximum MTU here for use in the slow start code below.
4585 /* Did peer restart with older RX version? */
4586 if (peer->maxDgramPackets > 1) {
4587 peer->maxDgramPackets = 1;
4589 } else if (np->length >=
4590 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4593 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4594 sizeof(afs_int32), &tSize);
4595 tSize = (afs_uint32) ntohl(tSize);
4597 * As of AFS 3.5 we set the send window to match the receive window.
4599 if (tSize < call->twind) {
4600 call->twind = tSize;
4601 call->conn->twind[call->channel] = call->twind;
4602 call->ssthresh = MIN(call->twind, call->ssthresh);
4603 } else if (tSize > call->twind) {
4604 call->twind = tSize;
4605 call->conn->twind[call->channel] = call->twind;
4609 * As of AFS 3.5, a jumbogram is more than one fixed size
4610 * packet transmitted in a single UDP datagram. If the remote
4611 * MTU is smaller than our local MTU then never send a datagram
4612 * larger than the natural MTU.
4615 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4616 (int)sizeof(afs_int32), &tSize);
4617 maxDgramPackets = (afs_uint32) ntohl(tSize);
4618 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4620 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4621 if (maxDgramPackets > 1) {
4622 peer->maxDgramPackets = maxDgramPackets;
4623 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4625 peer->maxDgramPackets = 1;
4626 call->MTU = peer->natMTU;
4628 } else if (peer->maxDgramPackets > 1) {
4629 /* Restarted with lower version of RX */
4630 peer->maxDgramPackets = 1;
4632 } else if (peer->maxDgramPackets > 1
4633 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4634 /* Restarted with lower version of RX */
4635 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4636 peer->natMTU = OLD_MAX_PACKET_SIZE;
4637 peer->MTU = OLD_MAX_PACKET_SIZE;
4638 peer->maxDgramPackets = 1;
4639 peer->nDgramPackets = 1;
4641 call->MTU = OLD_MAX_PACKET_SIZE;
4646 * Calculate how many datagrams were successfully received after
4647 * the first missing packet and adjust the negative ack counter
4652 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4653 if (call->nNacks < nNacked) {
4654 call->nNacks = nNacked;
4657 call->nAcks += newAckCount;
4661 /* If the packet contained new acknowledgements, rather than just
4662 * being a duplicate of one we have previously seen, then we can restart
4665 if (newAckCount > 0)
4666 rxi_rto_packet_acked(call, istack);
4668 if (call->flags & RX_CALL_FAST_RECOVER) {
4669 if (newAckCount == 0) {
4670 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4672 call->flags &= ~RX_CALL_FAST_RECOVER;
4673 call->cwind = call->nextCwind;
4674 call->nextCwind = 0;
4677 call->nCwindAcks = 0;
4678 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4679 /* Three negative acks in a row trigger congestion recovery */
4680 call->flags |= RX_CALL_FAST_RECOVER;
4681 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4683 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4684 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4685 call->nextCwind = call->ssthresh;
4688 peer->MTU = call->MTU;
4689 peer->cwind = call->nextCwind;
4690 peer->nDgramPackets = call->nDgramPackets;
4692 call->congestSeq = peer->congestSeq;
4694 /* Reset the resend times on the packets that were nacked
4695 * so we will retransmit as soon as the window permits
4699 for (opr_queue_ScanBackwards(&call->tq, cursor)) {
4700 struct rx_packet *tp =
4701 opr_queue_Entry(cursor, struct rx_packet, entry);
4703 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4704 tp->flags &= ~RX_PKTFLAG_SENT;
4706 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4711 /* If cwind is smaller than ssthresh, then increase
4712 * the window one packet for each ack we receive (exponential
4714 * If cwind is greater than or equal to ssthresh then increase
4715 * the congestion window by one packet for each cwind acks we
4716 * receive (linear growth). */
4717 if (call->cwind < call->ssthresh) {
4719 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4720 call->nCwindAcks = 0;
4722 call->nCwindAcks += newAckCount;
4723 if (call->nCwindAcks >= call->cwind) {
4724 call->nCwindAcks = 0;
4725 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4729 * If we have received several acknowledgements in a row then
4730 * it is time to increase the size of our datagrams
4732 if ((int)call->nAcks > rx_nDgramThreshold) {
4733 if (peer->maxDgramPackets > 1) {
4734 if (call->nDgramPackets < peer->maxDgramPackets) {
4735 call->nDgramPackets++;
4737 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4738 } else if (call->MTU < peer->maxMTU) {
4739 /* don't upgrade if we can't handle it */
4740 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4741 call->MTU = peer->ifMTU;
4743 call->MTU += peer->natMTU;
4744 call->MTU = MIN(call->MTU, peer->maxMTU);
4751 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4753 /* Servers need to hold the call until all response packets have
4754 * been acknowledged. Soft acks are good enough since clients
4755 * are not allowed to clear their receive queues. */
4756 if (call->state == RX_STATE_HOLD
4757 && call->tfirst + call->nSoftAcked >= call->tnext) {
4758 call->state = RX_STATE_DALLY;
4759 rxi_ClearTransmitQueue(call, 0);
4760 rxi_CancelKeepAliveEvent(call);
4761 } else if (!opr_queue_IsEmpty(&call->tq)) {
4762 rxi_Start(call, istack);
4768 * Schedule a connection abort to be sent after some delay.
4770 * @param[in] conn The connection to send the abort on.
4771 * @param[in] msec The number of milliseconds to wait before sending.
4773 * @pre conn_data_lock must be held
4776 rxi_SendConnectionAbortLater(struct rx_connection *conn, int msec)
4778 struct clock when, now;
4782 if (!conn->delayedAbortEvent) {
4783 clock_GetTime(&now);
4785 clock_Addmsec(&when, msec);
4786 conn->delayedAbortEvent =
4787 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
4791 /* Received a response to a challenge packet */
4792 static struct rx_packet *
4793 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4794 struct rx_packet *np, int istack)
4798 /* Ignore the packet if we're the client */
4799 if (conn->type == RX_CLIENT_CONNECTION)
4802 /* If already authenticated, ignore the packet (it's probably a retry) */
4803 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4806 if (!conn->securityChallengeSent) {
4807 /* We've never sent out a challenge for this connection, so this
4808 * response cannot possibly be correct; ignore it. This can happen
4809 * if we sent a challenge to the client, then we were restarted, and
4810 * then the client sent us a response. If we ignore the response, the
4811 * client will eventually resend a data packet, causing us to send a
4812 * new challenge and the client to send a new response. */
4816 /* Otherwise, have the security object evaluate the response packet */
4817 error = RXS_CheckResponse(conn->securityObject, conn, np);
4819 /* If the response is invalid, reset the connection, sending
4820 * an abort to the peer. Send the abort with a 1 second delay,
4821 * to avoid a peer hammering us by constantly recreating a
4822 * connection with bad credentials. */
4823 rxi_ConnectionError(conn, error);
4824 MUTEX_ENTER(&conn->conn_data_lock);
4825 rxi_SendConnectionAbortLater(conn, 1000);
4826 MUTEX_EXIT(&conn->conn_data_lock);
4829 /* If the response is valid, any calls waiting to attach
4830 * servers can now do so */
4833 for (i = 0; i < RX_MAXCALLS; i++) {
4834 struct rx_call *call = conn->call[i];
4836 MUTEX_ENTER(&call->lock);
4837 if (call->state == RX_STATE_PRECALL)
4838 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4839 /* tnop can be null if newcallp is null */
4840 MUTEX_EXIT(&call->lock);
4844 /* Update the peer reachability information, just in case
4845 * some calls went into attach-wait while we were waiting
4846 * for authentication..
4848 rxi_UpdatePeerReach(conn, NULL);
4853 /* A client has received an authentication challenge: the security
4854 * object is asked to cough up a respectable response packet to send
4855 * back to the server. The server is responsible for retrying the
4856 * challenge if it fails to get a response. */
4858 static struct rx_packet *
4859 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4860 struct rx_packet *np, int istack)
4864 /* Ignore the challenge if we're the server */
4865 if (conn->type == RX_SERVER_CONNECTION)
4868 /* Ignore the challenge if the connection is otherwise idle; someone's
4869 * trying to use us as an oracle. */
4870 if (!rxi_HasActiveCalls(conn))
4873 /* Send the security object the challenge packet. It is expected to fill
4874 * in the response. */
4875 error = RXS_GetResponse(conn->securityObject, conn, np);
4877 /* If the security object is unable to return a valid response, reset the
4878 * connection and send an abort to the peer. Otherwise send the response
4879 * packet to the peer connection. */
4881 rxi_ConnectionError(conn, error);
4882 MUTEX_ENTER(&conn->conn_data_lock);
4883 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4884 MUTEX_EXIT(&conn->conn_data_lock);
4886 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4887 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4893 /* Find an available server process to service the current request in
4894 * the given call structure. If one isn't available, queue up this
4895 * call so it eventually gets one */
4897 rxi_AttachServerProc(struct rx_call *call,
4898 osi_socket socket, int *tnop,
4899 struct rx_call **newcallp)
4901 struct rx_serverQueueEntry *sq;
4902 struct rx_service *service = call->conn->service;
4905 /* May already be attached */
4906 if (call->state == RX_STATE_ACTIVE)
4909 MUTEX_ENTER(&rx_serverPool_lock);
4911 haveQuota = QuotaOK(service);
4912 if ((!haveQuota) || opr_queue_IsEmpty(&rx_idleServerQueue)) {
4913 /* If there are no processes available to service this call,
4914 * put the call on the incoming call queue (unless it's
4915 * already on the queue).
4917 #ifdef RX_ENABLE_LOCKS
4919 ReturnToServerPool(service);
4920 #endif /* RX_ENABLE_LOCKS */
4922 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4923 call->flags |= RX_CALL_WAIT_PROC;
4924 rx_atomic_inc(&rx_nWaiting);
4925 rx_atomic_inc(&rx_nWaited);
4926 rxi_calltrace(RX_CALL_ARRIVAL, call);
4927 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4928 opr_queue_Append(&rx_incomingCallQueue, &call->entry);
4931 sq = opr_queue_Last(&rx_idleServerQueue,
4932 struct rx_serverQueueEntry, entry);
4934 /* If hot threads are enabled, and both newcallp and sq->socketp
4935 * are non-null, then this thread will process the call, and the
4936 * idle server thread will start listening on this threads socket.
4938 opr_queue_Remove(&sq->entry);
4940 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4943 *sq->socketp = socket;
4944 clock_GetTime(&call->startTime);
4945 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4949 if (call->flags & RX_CALL_WAIT_PROC) {
4950 /* Conservative: I don't think this should happen */
4951 call->flags &= ~RX_CALL_WAIT_PROC;
4952 rx_atomic_dec(&rx_nWaiting);
4953 if (opr_queue_IsOnQueue(&call->entry)) {
4954 opr_queue_Remove(&call->entry);
4957 call->state = RX_STATE_ACTIVE;
4958 call->app.mode = RX_MODE_RECEIVING;
4959 #ifdef RX_KERNEL_TRACE
4961 int glockOwner = ISAFS_GLOCK();
4964 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4965 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4971 if (call->flags & RX_CALL_CLEARED) {
4972 /* send an ack now to start the packet flow up again */
4973 call->flags &= ~RX_CALL_CLEARED;
4974 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4976 #ifdef RX_ENABLE_LOCKS
4979 service->nRequestsRunning++;
4980 MUTEX_ENTER(&rx_quota_mutex);
4981 if (service->nRequestsRunning <= service->minProcs)
4984 MUTEX_EXIT(&rx_quota_mutex);
4988 MUTEX_EXIT(&rx_serverPool_lock);
4991 /* Delay the sending of an acknowledge event for a short while, while
4992 * a new call is being prepared (in the case of a client) or a reply
4993 * is being prepared (in the case of a server). Rather than sending
4994 * an ack packet, an ACKALL packet is sent. */
4996 rxi_AckAll(struct rx_call *call)
4998 rxi_SendSpecial(call, call->conn, NULL, RX_PACKET_TYPE_ACKALL,
5000 call->flags |= RX_CALL_ACKALL_SENT;
5004 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
5007 struct rx_call *call = arg1;
5008 #ifdef RX_ENABLE_LOCKS
5010 MUTEX_ENTER(&call->lock);
5011 if (event == call->delayedAckEvent)
5012 rxevent_Put(&call->delayedAckEvent);
5013 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
5015 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
5017 MUTEX_EXIT(&call->lock);
5018 #else /* RX_ENABLE_LOCKS */
5020 rxevent_Put(&call->delayedAckEvent);
5021 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
5022 #endif /* RX_ENABLE_LOCKS */
5025 #ifdef RX_ENABLE_LOCKS
5026 /* Set ack in all packets in transmit queue. rxi_Start will deal with
5027 * clearing them out.
5030 rxi_SetAcksInTransmitQueue(struct rx_call *call)
5032 struct opr_queue *cursor;
5035 for (opr_queue_Scan(&call->tq, cursor)) {
5037 = opr_queue_Entry(cursor, struct rx_packet, entry);
5039 p->flags |= RX_PKTFLAG_ACKED;
5044 call->flags |= RX_CALL_TQ_CLEARME;
5045 call->flags |= RX_CALL_TQ_SOME_ACKED;
5048 rxi_rto_cancel(call);
5050 call->tfirst = call->tnext;
5051 call->nSoftAcked = 0;
5053 if (call->flags & RX_CALL_FAST_RECOVER) {
5054 call->flags &= ~RX_CALL_FAST_RECOVER;
5055 call->cwind = call->nextCwind;
5056 call->nextCwind = 0;
5059 CV_SIGNAL(&call->cv_twind);
5061 #endif /* RX_ENABLE_LOCKS */
5064 * Acknowledge the whole transmit queue.
5066 * If we're running without locks, or the transmit queue isn't busy, then
5067 * we can just clear the queue now. Otherwise, we have to mark all of the
5068 * packets as acknowledged, and let rxi_Start clear it later on
5071 rxi_AckAllInTransmitQueue(struct rx_call *call)
5073 #ifdef RX_ENABLE_LOCKS
5074 if (call->flags & RX_CALL_TQ_BUSY) {
5075 rxi_SetAcksInTransmitQueue(call);
5079 rxi_ClearTransmitQueue(call, 0);
5081 /* Clear out the transmit queue for the current call (all packets have
5082 * been received by peer) */
5084 rxi_ClearTransmitQueue(struct rx_call *call, int force)
5086 #ifdef RX_ENABLE_LOCKS
5087 struct opr_queue *cursor;
5088 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
5090 for (opr_queue_Scan(&call->tq, cursor)) {
5092 = opr_queue_Entry(cursor, struct rx_packet, entry);
5094 p->flags |= RX_PKTFLAG_ACKED;
5098 call->flags |= RX_CALL_TQ_CLEARME;
5099 call->flags |= RX_CALL_TQ_SOME_ACKED;
5102 #endif /* RX_ENABLE_LOCKS */
5103 #ifdef RXDEBUG_PACKET
5105 #endif /* RXDEBUG_PACKET */
5106 rxi_FreePackets(0, &call->tq);
5107 rxi_WakeUpTransmitQueue(call);
5108 #ifdef RX_ENABLE_LOCKS
5109 call->flags &= ~RX_CALL_TQ_CLEARME;
5113 rxi_rto_cancel(call);
5114 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5115 call->nSoftAcked = 0;
5117 if (call->flags & RX_CALL_FAST_RECOVER) {
5118 call->flags &= ~RX_CALL_FAST_RECOVER;
5119 call->cwind = call->nextCwind;
5121 #ifdef RX_ENABLE_LOCKS
5122 CV_SIGNAL(&call->cv_twind);
5124 osi_rxWakeup(&call->twind);
5129 rxi_ClearReceiveQueue(struct rx_call *call)
5131 if (!opr_queue_IsEmpty(&call->rq)) {
5134 count = rxi_FreePackets(0, &call->rq);
5135 rx_packetReclaims += count;
5136 #ifdef RXDEBUG_PACKET
5138 if ( call->rqc != 0 )
5139 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5141 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5143 if (call->state == RX_STATE_PRECALL) {
5144 call->flags |= RX_CALL_CLEARED;
5148 /* Send an abort packet for the specified call */
5149 static struct rx_packet *
5150 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5151 int istack, int force)
5153 afs_int32 error, cerror;
5154 struct clock when, now;
5159 switch (call->error) {
5162 cerror = RX_CALL_TIMEOUT;
5165 cerror = call->error;
5168 /* Clients should never delay abort messages */
5169 if (rx_IsClientConn(call->conn))
5172 if (call->abortCode != cerror) {
5173 call->abortCode = cerror;
5174 call->abortCount = 0;
5177 if (force || rxi_callAbortThreshhold == 0
5178 || call->abortCount < rxi_callAbortThreshhold) {
5179 rxi_CancelDelayedAbortEvent(call);
5180 error = htonl(cerror);
5183 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5184 (char *)&error, sizeof(error), istack);
5185 } else if (!call->delayedAbortEvent) {
5186 clock_GetTime(&now);
5188 clock_Addmsec(&when, rxi_callAbortDelay);
5189 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5190 call->delayedAbortEvent =
5191 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5197 rxi_CancelDelayedAbortEvent(struct rx_call *call)
5199 if (call->delayedAbortEvent) {
5200 rxevent_Cancel(&call->delayedAbortEvent);
5201 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5205 /* Send an abort packet for the specified connection. Packet is an
5206 * optional pointer to a packet that can be used to send the abort.
5207 * Once the number of abort messages reaches the threshhold, an
5208 * event is scheduled to send the abort. Setting the force flag
5209 * overrides sending delayed abort messages.
5211 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5212 * to send the abort packet.
5215 rxi_SendConnectionAbort(struct rx_connection *conn,
5216 struct rx_packet *packet, int istack, int force)
5223 /* Clients should never delay abort messages */
5224 if (rx_IsClientConn(conn))
5227 if (force || rxi_connAbortThreshhold == 0
5228 || conn->abortCount < rxi_connAbortThreshhold) {
5230 rxevent_Cancel(&conn->delayedAbortEvent);
5231 error = htonl(conn->error);
5233 MUTEX_EXIT(&conn->conn_data_lock);
5235 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5236 RX_PACKET_TYPE_ABORT, (char *)&error,
5237 sizeof(error), istack);
5238 MUTEX_ENTER(&conn->conn_data_lock);
5240 rxi_SendConnectionAbortLater(conn, rxi_connAbortDelay);
5245 /* Associate an error all of the calls owned by a connection. Called
5246 * with error non-zero. This is only for really fatal things, like
5247 * bad authentication responses. The connection itself is set in
5248 * error at this point, so that future packets received will be
5251 rxi_ConnectionError(struct rx_connection *conn,
5257 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5259 MUTEX_ENTER(&conn->conn_data_lock);
5260 rxevent_Cancel(&conn->challengeEvent);
5261 rxevent_Cancel(&conn->natKeepAliveEvent);
5262 if (conn->checkReachEvent) {
5263 rxevent_Cancel(&conn->checkReachEvent);
5264 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5265 putConnection(conn);
5267 MUTEX_EXIT(&conn->conn_data_lock);
5268 for (i = 0; i < RX_MAXCALLS; i++) {
5269 struct rx_call *call = conn->call[i];
5271 MUTEX_ENTER(&call->lock);
5272 rxi_CallError(call, error);
5273 MUTEX_EXIT(&call->lock);
5276 conn->error = error;
5277 if (rx_stats_active)
5278 rx_atomic_inc(&rx_stats.fatalErrors);
5283 * Interrupt an in-progress call with the specified error and wakeup waiters.
5285 * @param[in] call The call to interrupt
5286 * @param[in] error The error code to send to the peer
5289 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5291 MUTEX_ENTER(&call->lock);
5292 rxi_CallError(call, error);
5293 rxi_SendCallAbort(call, NULL, 0, 1);
5294 MUTEX_EXIT(&call->lock);
5298 rxi_CallError(struct rx_call *call, afs_int32 error)
5300 MUTEX_ASSERT(&call->lock);
5301 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5303 error = call->error;
5305 #ifdef RX_ENABLE_LOCKS
5306 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5307 rxi_ResetCall(call, 0);
5310 rxi_ResetCall(call, 0);
5312 call->error = error;
5315 /* Reset various fields in a call structure, and wakeup waiting
5316 * processes. Some fields aren't changed: state & mode are not
5317 * touched (these must be set by the caller), and bufptr, nLeft, and
5318 * nFree are not reset, since these fields are manipulated by
5319 * unprotected macros, and may only be reset by non-interrupting code.
5323 rxi_ResetCall(struct rx_call *call, int newcall)
5326 struct rx_peer *peer;
5327 struct rx_packet *packet;
5329 MUTEX_ASSERT(&call->lock);
5330 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5332 /* Notify anyone who is waiting for asynchronous packet arrival */
5333 if (call->arrivalProc) {
5334 (*call->arrivalProc) (call, call->arrivalProcHandle,
5335 call->arrivalProcArg);
5336 call->arrivalProc = (void (*)())0;
5340 rxi_CancelGrowMTUEvent(call);
5342 if (call->delayedAbortEvent) {
5343 rxi_CancelDelayedAbortEvent(call);
5344 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5346 rxi_SendCallAbort(call, packet, 0, 1);
5347 rxi_FreePacket(packet);
5352 * Update the peer with the congestion information in this call
5353 * so other calls on this connection can pick up where this call
5354 * left off. If the congestion sequence numbers don't match then
5355 * another call experienced a retransmission.
5357 peer = call->conn->peer;
5358 MUTEX_ENTER(&peer->peer_lock);
5360 if (call->congestSeq == peer->congestSeq) {
5361 peer->cwind = MAX(peer->cwind, call->cwind);
5362 peer->MTU = MAX(peer->MTU, call->MTU);
5363 peer->nDgramPackets =
5364 MAX(peer->nDgramPackets, call->nDgramPackets);
5367 call->abortCode = 0;
5368 call->abortCount = 0;
5370 if (peer->maxDgramPackets > 1) {
5371 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5373 call->MTU = peer->MTU;
5375 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5376 call->ssthresh = rx_maxSendWindow;
5377 call->nDgramPackets = peer->nDgramPackets;
5378 call->congestSeq = peer->congestSeq;
5379 call->rtt = peer->rtt;
5380 call->rtt_dev = peer->rtt_dev;
5381 clock_Zero(&call->rto);
5382 clock_Addmsec(&call->rto,
5383 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5384 MUTEX_EXIT(&peer->peer_lock);
5386 flags = call->flags;
5387 rxi_WaitforTQBusy(call);
5389 rxi_ClearTransmitQueue(call, 1);
5390 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5391 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5395 if (!newcall && (flags & RX_CALL_PEER_BUSY)) {
5396 /* The call channel is still busy; resetting the call doesn't change
5397 * that. However, if 'newcall' is set, we are processing a call
5398 * structure that has either been recycled from the free list, or has
5399 * been newly allocated. So, RX_CALL_PEER_BUSY is not relevant if
5400 * 'newcall' is set, since it describes a completely different call
5401 * channel which we do not care about. */
5402 call->flags |= RX_CALL_PEER_BUSY;
5405 rxi_ClearReceiveQueue(call);
5406 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5410 call->twind = call->conn->twind[call->channel];
5411 call->rwind = call->conn->rwind[call->channel];
5412 call->nSoftAcked = 0;
5413 call->nextCwind = 0;
5416 call->nCwindAcks = 0;
5417 call->nSoftAcks = 0;
5418 call->nHardAcks = 0;
5420 call->tfirst = call->rnext = call->tnext = 1;
5423 call->lastAcked = 0;
5424 call->localStatus = call->remoteStatus = 0;
5426 if (flags & RX_CALL_READER_WAIT) {
5427 #ifdef RX_ENABLE_LOCKS
5428 CV_BROADCAST(&call->cv_rq);
5430 osi_rxWakeup(&call->rq);
5433 if (flags & RX_CALL_WAIT_PACKETS) {
5434 MUTEX_ENTER(&rx_freePktQ_lock);
5435 rxi_PacketsUnWait(); /* XXX */
5436 MUTEX_EXIT(&rx_freePktQ_lock);
5438 #ifdef RX_ENABLE_LOCKS
5439 CV_SIGNAL(&call->cv_twind);
5441 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5442 osi_rxWakeup(&call->twind);
5445 if (flags & RX_CALL_WAIT_PROC) {
5446 rx_atomic_dec(&rx_nWaiting);
5448 #ifdef RX_ENABLE_LOCKS
5449 /* The following ensures that we don't mess with any queue while some
5450 * other thread might also be doing so. The call_queue_lock field is
5451 * is only modified under the call lock. If the call is in the process
5452 * of being removed from a queue, the call is not locked until the
5453 * the queue lock is dropped and only then is the call_queue_lock field
5454 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5455 * Note that any other routine which removes a call from a queue has to
5456 * obtain the queue lock before examing the queue and removing the call.
5458 if (call->call_queue_lock) {
5459 MUTEX_ENTER(call->call_queue_lock);
5460 if (opr_queue_IsOnQueue(&call->entry)) {
5461 opr_queue_Remove(&call->entry);
5463 MUTEX_EXIT(call->call_queue_lock);
5464 CLEAR_CALL_QUEUE_LOCK(call);
5466 #else /* RX_ENABLE_LOCKS */
5467 if (opr_queue_IsOnQueue(&call->entry)) {
5468 opr_queue_Remove(&call->entry);
5470 #endif /* RX_ENABLE_LOCKS */
5472 rxi_CancelKeepAliveEvent(call);
5473 rxi_CancelDelayedAckEvent(call);
5476 /* Send an acknowledge for the indicated packet (seq,serial) of the
5477 * indicated call, for the indicated reason (reason). This
5478 * acknowledge will specifically acknowledge receiving the packet, and
5479 * will also specify which other packets for this call have been
5480 * received. This routine returns the packet that was used to the
5481 * caller. The caller is responsible for freeing it or re-using it.
5482 * This acknowledgement also returns the highest sequence number
5483 * actually read out by the higher level to the sender; the sender
5484 * promises to keep around packets that have not been read by the
5485 * higher level yet (unless, of course, the sender decides to abort
5486 * the call altogether). Any of p, seq, serial, pflags, or reason may
5487 * be set to zero without ill effect. That is, if they are zero, they
5488 * will not convey any information.
5489 * NOW there is a trailer field, after the ack where it will safely be
5490 * ignored by mundanes, which indicates the maximum size packet this
5491 * host can swallow. */
5493 struct rx_packet *optionalPacket; use to send ack (or null)
5494 int seq; Sequence number of the packet we are acking
5495 int serial; Serial number of the packet
5496 int pflags; Flags field from packet header
5497 int reason; Reason an acknowledge was prompted
5501 rxi_SendAck(struct rx_call *call,
5502 struct rx_packet *optionalPacket, int serial, int reason,
5505 struct rx_ackPacket *ap;
5506 struct rx_packet *p;
5507 struct opr_queue *cursor;
5510 afs_uint32 padbytes = 0;
5511 #ifdef RX_ENABLE_TSFPQ
5512 struct rx_ts_info_t * rx_ts_info;
5516 * Open the receive window once a thread starts reading packets
5518 if (call->rnext > 1) {
5519 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5522 /* Don't attempt to grow MTU if this is a critical ping */
5523 if (reason == RX_ACK_MTU) {
5524 /* keep track of per-call attempts, if we're over max, do in small
5525 * otherwise in larger? set a size to increment by, decrease
5528 if (call->conn->peer->maxPacketSize &&
5529 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5531 padbytes = call->conn->peer->maxPacketSize+16;
5533 padbytes = call->conn->peer->maxMTU + 128;
5535 /* do always try a minimum size ping */
5536 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5538 /* subtract the ack payload */
5539 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5540 reason = RX_ACK_PING;
5543 call->nHardAcks = 0;
5544 call->nSoftAcks = 0;
5545 if (call->rnext > call->lastAcked)
5546 call->lastAcked = call->rnext;
5550 rx_computelen(p, p->length); /* reset length, you never know */
5551 } /* where that's been... */
5552 #ifdef RX_ENABLE_TSFPQ
5554 RX_TS_INFO_GET(rx_ts_info);
5555 if ((p = rx_ts_info->local_special_packet)) {
5556 rx_computelen(p, p->length);
5557 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5558 rx_ts_info->local_special_packet = p;
5559 } else { /* We won't send the ack, but don't panic. */
5560 return optionalPacket;
5564 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5565 /* We won't send the ack, but don't panic. */
5566 return optionalPacket;
5571 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5574 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5575 #ifndef RX_ENABLE_TSFPQ
5576 if (!optionalPacket)
5579 return optionalPacket;
5581 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5582 if (rx_Contiguous(p) < templ) {
5583 #ifndef RX_ENABLE_TSFPQ
5584 if (!optionalPacket)
5587 return optionalPacket;
5592 /* MTUXXX failing to send an ack is very serious. We should */
5593 /* try as hard as possible to send even a partial ack; it's */
5594 /* better than nothing. */
5595 ap = (struct rx_ackPacket *)rx_DataOf(p);
5596 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5597 ap->reason = reason;
5599 /* The skew computation used to be bogus, I think it's better now. */
5600 /* We should start paying attention to skew. XXX */
5601 ap->serial = htonl(serial);
5602 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5605 * First packet not yet forwarded to reader. When ACKALL has been
5606 * sent the peer has been told that all received packets will be
5607 * delivered to the reader. The value 'rnext' is used internally
5608 * to refer to the next packet in the receive queue that must be
5609 * delivered to the reader. From the perspective of the peer it
5610 * already has so report the last sequence number plus one if there
5611 * are packets in the receive queue awaiting processing.
5613 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5614 !opr_queue_IsEmpty(&call->rq)) {
5615 ap->firstPacket = htonl(opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq + 1);
5617 ap->firstPacket = htonl(call->rnext);
5619 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5621 /* No fear of running out of ack packet here because there can only
5622 * be at most one window full of unacknowledged packets. The window
5623 * size must be constrained to be less than the maximum ack size,
5624 * of course. Also, an ack should always fit into a single packet
5625 * -- it should not ever be fragmented. */
5627 for (opr_queue_Scan(&call->rq, cursor)) {
5628 struct rx_packet *rqp
5629 = opr_queue_Entry(cursor, struct rx_packet, entry);
5631 if (!rqp || !call->rq.next
5632 || (rqp->header.seq > (call->rnext + call->rwind))) {
5633 #ifndef RX_ENABLE_TSFPQ
5634 if (!optionalPacket)
5637 rxi_CallError(call, RX_CALL_DEAD);
5638 return optionalPacket;
5641 while (rqp->header.seq > call->rnext + offset)
5642 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5643 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5645 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5646 #ifndef RX_ENABLE_TSFPQ
5647 if (!optionalPacket)
5650 rxi_CallError(call, RX_CALL_DEAD);
5651 return optionalPacket;
5657 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5659 /* these are new for AFS 3.3 */
5660 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5661 templ = htonl(templ);
5662 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5663 templ = htonl(call->conn->peer->ifMTU);
5664 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5665 sizeof(afs_int32), &templ);
5667 /* new for AFS 3.4 */
5668 templ = htonl(call->rwind);
5669 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5670 sizeof(afs_int32), &templ);
5672 /* new for AFS 3.5 */
5673 templ = htonl(call->conn->peer->ifDgramPackets);
5674 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5675 sizeof(afs_int32), &templ);
5677 p->header.serviceId = call->conn->serviceId;
5678 p->header.cid = (call->conn->cid | call->channel);
5679 p->header.callNumber = *call->callNumber;
5681 p->header.securityIndex = call->conn->securityIndex;
5682 p->header.epoch = call->conn->epoch;
5683 p->header.type = RX_PACKET_TYPE_ACK;
5684 p->header.flags = RX_SLOW_START_OK;
5685 if (reason == RX_ACK_PING) {
5686 p->header.flags |= RX_REQUEST_ACK;
5688 p->length = padbytes +
5689 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5692 /* not fast but we can potentially use this if truncated
5693 * fragments are delivered to figure out the mtu.
5695 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5696 sizeof(afs_int32), sizeof(afs_int32),
5700 if (call->conn->type == RX_CLIENT_CONNECTION)
5701 p->header.flags |= RX_CLIENT_INITIATED;
5705 if (rxdebug_active) {
5709 len = _snprintf(msg, sizeof(msg),
5710 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5711 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5712 ntohl(ap->serial), ntohl(ap->previousPacket),
5713 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5714 ap->nAcks, ntohs(ap->bufferSpace) );
5718 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5719 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5723 OutputDebugString(msg);
5725 #else /* AFS_NT40_ENV */
5727 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5728 ap->reason, ntohl(ap->previousPacket),
5729 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5731 for (offset = 0; offset < ap->nAcks; offset++)
5732 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5737 #endif /* AFS_NT40_ENV */
5740 int i, nbytes = p->length;
5742 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5743 if (nbytes <= p->wirevec[i].iov_len) {
5746 savelen = p->wirevec[i].iov_len;
5748 p->wirevec[i].iov_len = nbytes;
5750 rxi_Send(call, p, istack);
5751 p->wirevec[i].iov_len = savelen;
5755 nbytes -= p->wirevec[i].iov_len;
5758 if (rx_stats_active)
5759 rx_atomic_inc(&rx_stats.ackPacketsSent);
5760 #ifndef RX_ENABLE_TSFPQ
5761 if (!optionalPacket)
5764 return optionalPacket; /* Return packet for re-use by caller */
5768 struct rx_packet **list;
5773 /* Send all of the packets in the list in single datagram */
5775 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5776 int istack, int moreFlag)
5782 struct rx_connection *conn = call->conn;
5783 struct rx_peer *peer = conn->peer;
5785 MUTEX_ENTER(&peer->peer_lock);
5786 peer->nSent += xmit->len;
5787 if (xmit->resending)
5788 peer->reSends += xmit->len;
5789 MUTEX_EXIT(&peer->peer_lock);
5791 if (rx_stats_active) {
5792 if (xmit->resending)
5793 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5795 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5798 clock_GetTime(&now);
5800 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5804 /* Set the packet flags and schedule the resend events */
5805 /* Only request an ack for the last packet in the list */
5806 for (i = 0; i < xmit->len; i++) {
5807 struct rx_packet *packet = xmit->list[i];
5809 /* Record the time sent */
5810 packet->timeSent = now;
5811 packet->flags |= RX_PKTFLAG_SENT;
5813 /* Ask for an ack on retransmitted packets, on every other packet
5814 * if the peer doesn't support slow start. Ask for an ack on every
5815 * packet until the congestion window reaches the ack rate. */
5816 if (packet->header.serial) {
5819 packet->firstSent = now;
5820 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5821 || (!(call->flags & RX_CALL_SLOW_START_OK)
5822 && (packet->header.seq & 1)))) {
5827 /* Tag this packet as not being the last in this group,
5828 * for the receiver's benefit */
5829 if (i < xmit->len - 1 || moreFlag) {
5830 packet->header.flags |= RX_MORE_PACKETS;
5835 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5838 /* Since we're about to send a data packet to the peer, it's
5839 * safe to nuke any scheduled end-of-packets ack */
5840 rxi_CancelDelayedAckEvent(call);
5842 MUTEX_EXIT(&call->lock);
5843 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5844 if (xmit->len > 1) {
5845 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5847 rxi_SendPacket(call, conn, xmit->list[0], istack);
5849 MUTEX_ENTER(&call->lock);
5850 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5852 /* Tell the RTO calculation engine that we have sent a packet, and
5853 * if it was the last one */
5854 rxi_rto_packet_sent(call, lastPacket, istack);
5856 /* Update last send time for this call (for keep-alive
5857 * processing), and for the connection (so that we can discover
5858 * idle connections) */
5859 conn->lastSendTime = call->lastSendTime = clock_Sec();
5862 /* When sending packets we need to follow these rules:
5863 * 1. Never send more than maxDgramPackets in a jumbogram.
5864 * 2. Never send a packet with more than two iovecs in a jumbogram.
5865 * 3. Never send a retransmitted packet in a jumbogram.
5866 * 4. Never send more than cwind/4 packets in a jumbogram
5867 * We always keep the last list we should have sent so we
5868 * can set the RX_MORE_PACKETS flags correctly.
5872 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5877 struct xmitlist working;
5878 struct xmitlist last;
5880 struct rx_peer *peer = call->conn->peer;
5881 int morePackets = 0;
5883 memset(&last, 0, sizeof(struct xmitlist));
5884 working.list = &list[0];
5886 working.resending = 0;
5888 recovery = call->flags & RX_CALL_FAST_RECOVER;
5890 for (i = 0; i < len; i++) {
5891 /* Does the current packet force us to flush the current list? */
5893 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5894 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5896 /* This sends the 'last' list and then rolls the current working
5897 * set into the 'last' one, and resets the working set */
5900 rxi_SendList(call, &last, istack, 1);
5901 /* If the call enters an error state stop sending, or if
5902 * we entered congestion recovery mode, stop sending */
5904 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5909 working.resending = 0;
5910 working.list = &list[i];
5912 /* Add the current packet to the list if it hasn't been acked.
5913 * Otherwise adjust the list pointer to skip the current packet. */
5914 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5917 if (list[i]->header.serial)
5918 working.resending = 1;
5920 /* Do we need to flush the list? */
5921 if (working.len >= (int)peer->maxDgramPackets
5922 || working.len >= (int)call->nDgramPackets
5923 || working.len >= (int)call->cwind
5924 || list[i]->header.serial
5925 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5927 rxi_SendList(call, &last, istack, 1);
5928 /* If the call enters an error state stop sending, or if
5929 * we entered congestion recovery mode, stop sending */
5931 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5936 working.resending = 0;
5937 working.list = &list[i + 1];
5940 if (working.len != 0) {
5941 osi_Panic("rxi_SendList error");
5943 working.list = &list[i + 1];
5947 /* Send the whole list when the call is in receive mode, when
5948 * the call is in eof mode, when we are in fast recovery mode,
5949 * and when we have the last packet */
5950 /* XXX - The accesses to app.mode aren't safe, as this may be called by
5951 * the listener or event threads
5953 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5954 || (call->flags & RX_CALL_FLUSH)
5955 || (call->flags & RX_CALL_FAST_RECOVER)) {
5956 /* Check for the case where the current list contains
5957 * an acked packet. Since we always send retransmissions
5958 * in a separate packet, we only need to check the first
5959 * packet in the list */
5960 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5964 rxi_SendList(call, &last, istack, morePackets);
5965 /* If the call enters an error state stop sending, or if
5966 * we entered congestion recovery mode, stop sending */
5968 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5972 rxi_SendList(call, &working, istack, 0);
5974 } else if (last.len > 0) {
5975 rxi_SendList(call, &last, istack, 0);
5976 /* Packets which are in 'working' are not sent by this call */
5981 * Check if the peer for the given call is known to be dead
5983 * If the call's peer appears dead (it has encountered fatal network errors
5984 * since the call started) the call is killed with RX_CALL_DEAD if the call
5985 * is active. Otherwise, we do nothing.
5987 * @param[in] call The call to check
5990 * @retval 0 The call is fine, and we haven't done anything to the call
5991 * @retval nonzero The call's peer appears dead, and the call has been
5992 * terminated if it was active
5994 * @pre call->lock must be locked
5997 rxi_CheckPeerDead(struct rx_call *call)
5999 #ifdef AFS_RXERRQ_ENV
6002 if (call->state == RX_STATE_DALLY) {
6006 peererrs = rx_atomic_read(&call->conn->peer->neterrs);
6007 if (call->neterr_gen < peererrs) {
6008 /* we have received network errors since this call started; kill
6010 if (call->state == RX_STATE_ACTIVE) {
6011 rxi_CallError(call, RX_CALL_DEAD);
6015 if (call->neterr_gen > peererrs) {
6016 /* someone has reset the number of peer errors; set the call error gen
6017 * so we can detect if more errors are encountered */
6018 call->neterr_gen = peererrs;
6025 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
6027 struct rx_call *call = arg0;
6028 struct rx_peer *peer;
6029 struct opr_queue *cursor;
6030 struct clock maxTimeout = { 60, 0 };
6032 MUTEX_ENTER(&call->lock);
6034 peer = call->conn->peer;
6036 /* Make sure that the event pointer is removed from the call
6037 * structure, since there is no longer a per-call retransmission
6039 if (event == call->resendEvent) {
6040 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
6041 rxevent_Put(&call->resendEvent);
6044 rxi_CheckPeerDead(call);
6046 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
6047 rxi_CheckBusy(call);
6050 if (opr_queue_IsEmpty(&call->tq)) {
6051 /* Nothing to do. This means that we've been raced, and that an
6052 * ACK has come in between when we were triggered, and when we
6053 * actually got to run. */
6057 /* We're in loss recovery */
6058 call->flags |= RX_CALL_FAST_RECOVER;
6060 /* Mark all of the pending packets in the queue as being lost */
6061 for (opr_queue_Scan(&call->tq, cursor)) {
6062 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
6063 if (!(p->flags & RX_PKTFLAG_ACKED))
6064 p->flags &= ~RX_PKTFLAG_SENT;
6067 /* We're resending, so we double the timeout of the call. This will be
6068 * dropped back down by the first successful ACK that we receive.
6070 * We apply a maximum value here of 60 seconds
6072 clock_Add(&call->rto, &call->rto);
6073 if (clock_Gt(&call->rto, &maxTimeout))
6074 call->rto = maxTimeout;
6076 /* Packet loss is most likely due to congestion, so drop our window size
6077 * and start again from the beginning */
6078 if (peer->maxDgramPackets >1) {
6079 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
6080 call->MTU = MIN(peer->natMTU, peer->maxMTU);
6082 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
6083 call->nDgramPackets = 1;
6085 call->nextCwind = 1;
6088 MUTEX_ENTER(&peer->peer_lock);
6089 peer->MTU = call->MTU;
6090 peer->cwind = call->cwind;
6091 peer->nDgramPackets = 1;
6093 call->congestSeq = peer->congestSeq;
6094 MUTEX_EXIT(&peer->peer_lock);
6096 rxi_Start(call, istack);
6099 MUTEX_EXIT(&call->lock);
6102 /* This routine is called when new packets are readied for
6103 * transmission and when retransmission may be necessary, or when the
6104 * transmission window or burst count are favourable. This should be
6105 * better optimized for new packets, the usual case, now that we've
6106 * got rid of queues of send packets. XXXXXXXXXXX */
6108 rxi_Start(struct rx_call *call, int istack)
6110 struct opr_queue *cursor;
6111 #ifdef RX_ENABLE_LOCKS
6112 struct opr_queue *store;
6118 #ifdef RX_ENABLE_LOCKS
6119 if (rx_stats_active)
6120 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
6125 if (!opr_queue_IsEmpty(&call->tq)) { /* If we have anything to send */
6126 /* Send (or resend) any packets that need it, subject to
6127 * window restrictions and congestion burst control
6128 * restrictions. Ask for an ack on the last packet sent in
6129 * this burst. For now, we're relying upon the window being
6130 * considerably bigger than the largest number of packets that
6131 * are typically sent at once by one initial call to
6132 * rxi_Start. This is probably bogus (perhaps we should ask
6133 * for an ack when we're half way through the current
6134 * window?). Also, for non file transfer applications, this
6135 * may end up asking for an ack for every packet. Bogus. XXXX
6138 * But check whether we're here recursively, and let the other guy
6141 #ifdef RX_ENABLE_LOCKS
6142 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6143 call->flags |= RX_CALL_TQ_BUSY;
6145 #endif /* RX_ENABLE_LOCKS */
6147 #ifdef RX_ENABLE_LOCKS
6148 call->flags &= ~RX_CALL_NEED_START;
6149 #endif /* RX_ENABLE_LOCKS */
6151 maxXmitPackets = MIN(call->twind, call->cwind);
6152 for (opr_queue_Scan(&call->tq, cursor)) {
6154 = opr_queue_Entry(cursor, struct rx_packet, entry);
6156 if (p->flags & RX_PKTFLAG_ACKED) {
6157 /* Since we may block, don't trust this */
6158 if (rx_stats_active)
6159 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6160 continue; /* Ignore this packet if it has been acknowledged */
6163 /* Turn off all flags except these ones, which are the same
6164 * on each transmission */
6165 p->header.flags &= RX_PRESET_FLAGS;
6167 if (p->header.seq >=
6168 call->tfirst + MIN((int)call->twind,
6169 (int)(call->nSoftAcked +
6171 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6172 /* Note: if we're waiting for more window space, we can
6173 * still send retransmits; hence we don't return here, but
6174 * break out to schedule a retransmit event */
6175 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6176 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6181 /* Transmit the packet if it needs to be sent. */
6182 if (!(p->flags & RX_PKTFLAG_SENT)) {
6183 if (nXmitPackets == maxXmitPackets) {
6184 rxi_SendXmitList(call, call->xmitList,
6185 nXmitPackets, istack);
6188 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6189 *(call->callNumber), p));
6190 call->xmitList[nXmitPackets++] = p;
6192 } /* end of the queue_Scan */
6194 /* xmitList now hold pointers to all of the packets that are
6195 * ready to send. Now we loop to send the packets */
6196 if (nXmitPackets > 0) {
6197 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6201 #ifdef RX_ENABLE_LOCKS
6203 /* We went into the error state while sending packets. Now is
6204 * the time to reset the call. This will also inform the using
6205 * process that the call is in an error state.
6207 if (rx_stats_active)
6208 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6209 call->flags &= ~RX_CALL_TQ_BUSY;
6210 rxi_WakeUpTransmitQueue(call);
6211 rxi_CallError(call, call->error);
6215 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6217 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6218 /* Some packets have received acks. If they all have, we can clear
6219 * the transmit queue.
6222 for (opr_queue_ScanSafe(&call->tq, cursor, store)) {
6224 = opr_queue_Entry(cursor, struct rx_packet, entry);
6226 if (p->header.seq < call->tfirst
6227 && (p->flags & RX_PKTFLAG_ACKED)) {
6228 opr_queue_Remove(&p->entry);
6229 #ifdef RX_TRACK_PACKETS
6230 p->flags &= ~RX_PKTFLAG_TQ;
6232 #ifdef RXDEBUG_PACKET
6240 call->flags |= RX_CALL_TQ_CLEARME;
6242 if (call->flags & RX_CALL_TQ_CLEARME)
6243 rxi_ClearTransmitQueue(call, 1);
6244 } while (call->flags & RX_CALL_NEED_START);
6246 * TQ references no longer protected by this flag; they must remain
6247 * protected by the call lock.
6249 call->flags &= ~RX_CALL_TQ_BUSY;
6250 rxi_WakeUpTransmitQueue(call);
6252 call->flags |= RX_CALL_NEED_START;
6254 #endif /* RX_ENABLE_LOCKS */
6256 rxi_rto_cancel(call);
6260 /* Also adjusts the keep alive parameters for the call, to reflect
6261 * that we have just sent a packet (so keep alives aren't sent
6264 rxi_Send(struct rx_call *call, struct rx_packet *p,
6267 struct rx_connection *conn = call->conn;
6269 /* Stamp each packet with the user supplied status */
6270 p->header.userStatus = call->localStatus;
6272 /* Allow the security object controlling this call's security to
6273 * make any last-minute changes to the packet */
6274 RXS_SendPacket(conn->securityObject, call, p);
6276 /* Since we're about to send SOME sort of packet to the peer, it's
6277 * safe to nuke any scheduled end-of-packets ack */
6278 rxi_CancelDelayedAckEvent(call);
6280 /* Actually send the packet, filling in more connection-specific fields */
6281 MUTEX_EXIT(&call->lock);
6282 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6283 rxi_SendPacket(call, conn, p, istack);
6284 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6285 MUTEX_ENTER(&call->lock);
6287 /* Update last send time for this call (for keep-alive
6288 * processing), and for the connection (so that we can discover
6289 * idle connections) */
6290 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6291 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6292 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6294 conn->lastSendTime = call->lastSendTime = clock_Sec();
6298 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6299 * that things are fine. Also called periodically to guarantee that nothing
6300 * falls through the cracks (e.g. (error + dally) connections have keepalive
6301 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6303 * haveCTLock Set if calling from rxi_ReapConnections
6306 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6308 struct rx_connection *conn = call->conn;
6310 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6311 afs_uint32 fudgeFactor;
6314 int idle_timeout = 0;
6315 afs_int32 clock_diff = 0;
6317 if (rxi_CheckPeerDead(call)) {
6323 /* Large swings in the clock can have a significant impact on
6324 * the performance of RX call processing. Forward clock shifts
6325 * will result in premature event triggering or timeouts.
6326 * Backward shifts can result in calls not completing until
6327 * the clock catches up with the original start clock value.
6329 * If a backward clock shift of more than five minutes is noticed,
6330 * just fail the call.
6332 if (now < call->lastSendTime)
6333 clock_diff = call->lastSendTime - now;
6334 if (now < call->startWait)
6335 clock_diff = MAX(clock_diff, call->startWait - now);
6336 if (now < call->lastReceiveTime)
6337 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6338 if (clock_diff > 5 * 60)
6340 if (call->state == RX_STATE_ACTIVE)
6341 rxi_CallError(call, RX_CALL_TIMEOUT);
6345 #ifdef RX_ENABLE_LOCKS
6346 if (call->flags & RX_CALL_TQ_BUSY) {
6347 /* Call is active and will be reset by rxi_Start if it's
6348 * in an error state.
6353 /* RTT + 8*MDEV, rounded up to the next second. */
6354 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6355 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6357 deadTime = conn->secondsUntilDead + fudgeFactor;
6358 /* These are computed to the second (+- 1 second). But that's
6359 * good enough for these values, which should be a significant
6360 * number of seconds. */
6361 if (now > (call->lastReceiveTime + deadTime)) {
6362 if (call->state == RX_STATE_ACTIVE) {
6363 cerror = RX_CALL_DEAD;
6366 #ifdef RX_ENABLE_LOCKS
6367 /* Cancel pending events */
6368 rxi_CancelDelayedAckEvent(call);
6369 rxi_rto_cancel(call);
6370 rxi_CancelKeepAliveEvent(call);
6371 rxi_CancelGrowMTUEvent(call);
6372 MUTEX_ENTER(&rx_refcnt_mutex);
6373 /* if rxi_FreeCall returns 1 it has freed the call */
6374 if (call->refCount == 0 &&
6375 rxi_FreeCall(call, haveCTLock))
6377 MUTEX_EXIT(&rx_refcnt_mutex);
6380 MUTEX_EXIT(&rx_refcnt_mutex);
6382 #else /* RX_ENABLE_LOCKS */
6383 rxi_FreeCall(call, 0);
6385 #endif /* RX_ENABLE_LOCKS */
6387 /* Non-active calls are destroyed if they are not responding
6388 * to pings; active calls are simply flagged in error, so the
6389 * attached process can die reasonably gracefully. */
6392 if (conn->idleDeadTime) {
6393 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6397 /* see if we have a non-activity timeout */
6398 if (call->startWait && ((call->startWait + idleDeadTime) < now)) {
6399 if (call->state == RX_STATE_ACTIVE) {
6400 cerror = RX_CALL_TIMEOUT;
6406 if (conn->hardDeadTime) {
6407 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6410 /* see if we have a hard timeout */
6412 && (now > (hardDeadTime + call->startTime.sec))) {
6413 if (call->state == RX_STATE_ACTIVE)
6414 rxi_CallError(call, RX_CALL_TIMEOUT);
6419 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6420 call->lastReceiveTime) {
6421 int oldMTU = conn->peer->ifMTU;
6423 /* If we thought we could send more, perhaps things got worse.
6424 * Shrink by 128 bytes and try again. */
6425 if (conn->peer->maxPacketSize < conn->lastPacketSize)
6426 /* maxPacketSize will be cleared in rxi_SetPeerMtu */
6427 newmtu = MAX(conn->peer->maxPacketSize + RX_HEADER_SIZE,
6428 conn->lastPacketSize - 128 + RX_HEADER_SIZE);
6430 newmtu = conn->lastPacketSize - 128 + RX_HEADER_SIZE;
6432 /* minimum capped in SetPeerMtu */
6433 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6436 conn->lastPacketSize = 0;
6438 /* needed so ResetCall doesn't clobber us. */
6439 call->MTU = conn->peer->ifMTU;
6441 /* if we never succeeded, let the error pass out as-is */
6442 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6443 cerror = conn->msgsizeRetryErr;
6446 rxi_CallError(call, cerror);
6451 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6452 void *dummy, int dummy2)
6454 struct rx_connection *conn = arg1;
6455 struct rx_header theader;
6456 char tbuffer[1 + sizeof(struct rx_header)];
6457 struct sockaddr_in taddr;
6460 struct iovec tmpiov[2];
6463 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6466 tp = &tbuffer[sizeof(struct rx_header)];
6467 taddr.sin_family = AF_INET;
6468 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6469 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6470 memset(&taddr.sin_zero, 0, sizeof(taddr.sin_zero));
6471 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6472 taddr.sin_len = sizeof(struct sockaddr_in);
6474 memset(&theader, 0, sizeof(theader));
6475 theader.epoch = htonl(999);
6477 theader.callNumber = 0;
6480 theader.type = RX_PACKET_TYPE_VERSION;
6481 theader.flags = RX_LAST_PACKET;
6482 theader.serviceId = 0;
6484 memcpy(tbuffer, &theader, sizeof(theader));
6485 memcpy(tp, &a, sizeof(a));
6486 tmpiov[0].iov_base = tbuffer;
6487 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6489 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6491 MUTEX_ENTER(&conn->conn_data_lock);
6492 MUTEX_ENTER(&rx_refcnt_mutex);
6493 /* Only reschedule ourselves if the connection would not be destroyed */
6494 if (conn->refCount <= 1) {
6495 rxevent_Put(&conn->natKeepAliveEvent);
6496 MUTEX_EXIT(&rx_refcnt_mutex);
6497 MUTEX_EXIT(&conn->conn_data_lock);
6498 rx_DestroyConnection(conn); /* drop the reference for this */
6500 conn->refCount--; /* drop the reference for this */
6501 MUTEX_EXIT(&rx_refcnt_mutex);
6502 rxevent_Put(&conn->natKeepAliveEvent);
6503 rxi_ScheduleNatKeepAliveEvent(conn);
6504 MUTEX_EXIT(&conn->conn_data_lock);
6509 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6511 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6512 struct clock when, now;
6513 clock_GetTime(&now);
6515 when.sec += conn->secondsUntilNatPing;
6516 MUTEX_ENTER(&rx_refcnt_mutex);
6517 conn->refCount++; /* hold a reference for this */
6518 MUTEX_EXIT(&rx_refcnt_mutex);
6519 conn->natKeepAliveEvent =
6520 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6525 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6527 MUTEX_ENTER(&conn->conn_data_lock);
6528 conn->secondsUntilNatPing = seconds;
6530 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6531 rxi_ScheduleNatKeepAliveEvent(conn);
6533 conn->flags |= RX_CONN_NAT_PING;
6535 MUTEX_EXIT(&conn->conn_data_lock);
6538 /* When a call is in progress, this routine is called occasionally to
6539 * make sure that some traffic has arrived (or been sent to) the peer.
6540 * If nothing has arrived in a reasonable amount of time, the call is
6541 * declared dead; if nothing has been sent for a while, we send a
6542 * keep-alive packet (if we're actually trying to keep the call alive)
6545 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6548 struct rx_call *call = arg1;
6549 struct rx_connection *conn;
6552 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6553 MUTEX_ENTER(&call->lock);
6555 if (event == call->keepAliveEvent)
6556 rxevent_Put(&call->keepAliveEvent);
6560 if (rxi_CheckCall(call, 0)) {
6561 MUTEX_EXIT(&call->lock);
6565 /* Don't try to keep alive dallying calls */
6566 if (call->state == RX_STATE_DALLY) {
6567 MUTEX_EXIT(&call->lock);
6572 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6573 /* Don't try to send keepalives if there is unacknowledged data */
6574 /* the rexmit code should be good enough, this little hack
6575 * doesn't quite work XXX */
6576 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6578 rxi_ScheduleKeepAliveEvent(call);
6579 MUTEX_EXIT(&call->lock);
6582 /* Does what's on the nameplate. */
6584 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6586 struct rx_call *call = arg1;
6587 struct rx_connection *conn;
6589 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6590 MUTEX_ENTER(&call->lock);
6592 if (event == call->growMTUEvent)
6593 rxevent_Put(&call->growMTUEvent);
6595 if (rxi_CheckCall(call, 0)) {
6596 MUTEX_EXIT(&call->lock);
6600 /* Don't bother with dallying calls */
6601 if (call->state == RX_STATE_DALLY) {
6602 MUTEX_EXIT(&call->lock);
6609 * keep being scheduled, just don't do anything if we're at peak,
6610 * or we're not set up to be properly handled (idle timeout required)
6612 if ((conn->peer->maxPacketSize != 0) &&
6613 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6615 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6616 rxi_ScheduleGrowMTUEvent(call, 0);
6617 MUTEX_EXIT(&call->lock);
6621 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6623 if (!call->keepAliveEvent) {
6624 struct clock when, now;
6625 clock_GetTime(&now);
6627 when.sec += call->conn->secondsUntilPing;
6628 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6629 call->keepAliveEvent =
6630 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6635 rxi_CancelKeepAliveEvent(struct rx_call *call) {
6636 if (call->keepAliveEvent) {
6637 rxevent_Cancel(&call->keepAliveEvent);
6638 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6643 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6645 if (!call->growMTUEvent) {
6646 struct clock when, now;
6648 clock_GetTime(&now);
6651 if (call->conn->secondsUntilPing)
6652 secs = (6*call->conn->secondsUntilPing)-1;
6654 if (call->conn->secondsUntilDead)
6655 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6659 CALL_HOLD(call, RX_CALL_REFCOUNT_MTU);
6660 call->growMTUEvent =
6661 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6666 rxi_CancelGrowMTUEvent(struct rx_call *call)
6668 if (call->growMTUEvent) {
6669 rxevent_Cancel(&call->growMTUEvent);
6670 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6675 * Increment the counter for the next connection ID, handling overflow.
6678 update_nextCid(void)
6680 /* Overflow is technically undefined behavior; avoid it. */
6681 if (rx_nextCid > MAX_AFS_INT32 - (1 << RX_CIDSHIFT))
6682 rx_nextCid = -1 * ((MAX_AFS_INT32 / RX_CIDSHIFT) * RX_CIDSHIFT);
6684 rx_nextCid += 1 << RX_CIDSHIFT;
6688 rxi_KeepAliveOn(struct rx_call *call)
6690 /* Pretend last packet received was received now--i.e. if another
6691 * packet isn't received within the keep alive time, then the call
6692 * will die; Initialize last send time to the current time--even
6693 * if a packet hasn't been sent yet. This will guarantee that a
6694 * keep-alive is sent within the ping time */
6695 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6696 rxi_ScheduleKeepAliveEvent(call);
6700 rxi_GrowMTUOn(struct rx_call *call)
6702 struct rx_connection *conn = call->conn;
6703 MUTEX_ENTER(&conn->conn_data_lock);
6704 conn->lastPingSizeSer = conn->lastPingSize = 0;
6705 MUTEX_EXIT(&conn->conn_data_lock);
6706 rxi_ScheduleGrowMTUEvent(call, 1);
6709 /* This routine is called to send connection abort messages
6710 * that have been delayed to throttle looping clients. */
6712 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6715 struct rx_connection *conn = arg1;
6718 struct rx_packet *packet;
6720 MUTEX_ENTER(&conn->conn_data_lock);
6721 rxevent_Put(&conn->delayedAbortEvent);
6722 error = htonl(conn->error);
6724 MUTEX_EXIT(&conn->conn_data_lock);
6725 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6728 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6729 RX_PACKET_TYPE_ABORT, (char *)&error,
6731 rxi_FreePacket(packet);
6735 /* This routine is called to send call abort messages
6736 * that have been delayed to throttle looping clients. */
6738 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6741 struct rx_call *call = arg1;
6744 struct rx_packet *packet;
6746 MUTEX_ENTER(&call->lock);
6747 rxevent_Put(&call->delayedAbortEvent);
6748 error = htonl(call->error);
6750 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6753 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6754 (char *)&error, sizeof(error), 0);
6755 rxi_FreePacket(packet);
6757 MUTEX_EXIT(&call->lock);
6758 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6761 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6762 * seconds) to ask the client to authenticate itself. The routine
6763 * issues a challenge to the client, which is obtained from the
6764 * security object associated with the connection */
6766 rxi_ChallengeEvent(struct rxevent *event,
6767 void *arg0, void *arg1, int tries)
6769 struct rx_connection *conn = arg0;
6772 rxevent_Put(&conn->challengeEvent);
6774 /* If there are no active calls it is not worth re-issuing the
6775 * challenge. If the client issues another call on this connection
6776 * the challenge can be requested at that time.
6778 if (!rxi_HasActiveCalls(conn))
6781 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6782 struct rx_packet *packet;
6783 struct clock when, now;
6786 /* We've failed to authenticate for too long.
6787 * Reset any calls waiting for authentication;
6788 * they are all in RX_STATE_PRECALL.
6792 MUTEX_ENTER(&conn->conn_call_lock);
6793 for (i = 0; i < RX_MAXCALLS; i++) {
6794 struct rx_call *call = conn->call[i];
6796 MUTEX_ENTER(&call->lock);
6797 if (call->state == RX_STATE_PRECALL) {
6798 rxi_CallError(call, RX_CALL_DEAD);
6799 rxi_SendCallAbort(call, NULL, 0, 0);
6801 MUTEX_EXIT(&call->lock);
6804 MUTEX_EXIT(&conn->conn_call_lock);
6808 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6810 /* If there's no packet available, do this later. */
6811 RXS_GetChallenge(conn->securityObject, conn, packet);
6812 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6813 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6814 rxi_FreePacket(packet);
6815 conn->securityChallengeSent = 1;
6817 clock_GetTime(&now);
6819 when.sec += RX_CHALLENGE_TIMEOUT;
6820 conn->challengeEvent =
6821 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6826 /* Call this routine to start requesting the client to authenticate
6827 * itself. This will continue until authentication is established,
6828 * the call times out, or an invalid response is returned. The
6829 * security object associated with the connection is asked to create
6830 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6831 * defined earlier. */
6833 rxi_ChallengeOn(struct rx_connection *conn)
6835 if (!conn->challengeEvent) {
6836 RXS_CreateChallenge(conn->securityObject, conn);
6837 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6842 /* rxi_ComputeRoundTripTime is called with peer locked. */
6843 /* peer may be null */
6845 rxi_ComputeRoundTripTime(struct rx_packet *p,
6846 struct rx_ackPacket *ack,
6847 struct rx_call *call,
6848 struct rx_peer *peer,
6851 struct clock thisRtt, *sentp;
6855 /* If the ACK is delayed, then do nothing */
6856 if (ack->reason == RX_ACK_DELAY)
6859 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6860 * their RTT multiple times, so only include the RTT of the last packet
6862 if (p->flags & RX_JUMBO_PACKET)
6865 /* Use the serial number to determine which transmission the ACK is for,
6866 * and set the sent time to match this. If we have no serial number, then
6867 * only use the ACK for RTT calculations if the packet has not been
6871 serial = ntohl(ack->serial);
6873 if (serial == p->header.serial) {
6874 sentp = &p->timeSent;
6875 } else if (serial == p->firstSerial) {
6876 sentp = &p->firstSent;
6877 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6878 sentp = &p->firstSent;
6882 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6883 sentp = &p->firstSent;
6890 if (clock_Lt(&thisRtt, sentp))
6891 return; /* somebody set the clock back, don't count this time. */
6893 clock_Sub(&thisRtt, sentp);
6894 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6895 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6897 if (clock_IsZero(&thisRtt)) {
6899 * The actual round trip time is shorter than the
6900 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6901 * Since we can't tell which at the moment we will assume 1ms.
6903 thisRtt.usec = 1000;
6906 if (rx_stats_active) {
6907 MUTEX_ENTER(&rx_stats_mutex);
6908 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6909 rx_stats.minRtt = thisRtt;
6910 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6911 if (thisRtt.sec > 60) {
6912 MUTEX_EXIT(&rx_stats_mutex);
6913 return; /* somebody set the clock ahead */
6915 rx_stats.maxRtt = thisRtt;
6917 clock_Add(&rx_stats.totalRtt, &thisRtt);
6918 rx_atomic_inc(&rx_stats.nRttSamples);
6919 MUTEX_EXIT(&rx_stats_mutex);
6922 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6924 /* Apply VanJacobson round-trip estimations */
6929 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6930 * srtt is stored as fixed point with 3 bits after the binary
6931 * point (i.e., scaled by 8). The following magic is
6932 * equivalent to the smoothing algorithm in rfc793 with an
6933 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6934 * srtt'*8 = rtt + srtt*7
6935 * srtt'*8 = srtt*8 + rtt - srtt
6936 * srtt' = srtt + rtt/8 - srtt/8
6937 * srtt' = srtt + (rtt - srtt)/8
6940 delta = _8THMSEC(&thisRtt) - call->rtt;
6941 call->rtt += (delta >> 3);
6944 * We accumulate a smoothed rtt variance (actually, a smoothed
6945 * mean difference), then set the retransmit timer to smoothed
6946 * rtt + 4 times the smoothed variance (was 2x in van's original
6947 * paper, but 4x works better for me, and apparently for him as
6949 * rttvar is stored as
6950 * fixed point with 2 bits after the binary point (scaled by
6951 * 4). The following is equivalent to rfc793 smoothing with
6952 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6953 * rttvar'*4 = rttvar*3 + |delta|
6954 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6955 * rttvar' = rttvar + |delta|/4 - rttvar/4
6956 * rttvar' = rttvar + (|delta| - rttvar)/4
6957 * This replaces rfc793's wired-in beta.
6958 * dev*4 = dev*4 + (|actual - expected| - dev)
6964 delta -= (call->rtt_dev << 1);
6965 call->rtt_dev += (delta >> 3);
6967 /* I don't have a stored RTT so I start with this value. Since I'm
6968 * probably just starting a call, and will be pushing more data down
6969 * this, I expect congestion to increase rapidly. So I fudge a
6970 * little, and I set deviance to half the rtt. In practice,
6971 * deviance tends to approach something a little less than
6972 * half the smoothed rtt. */
6973 call->rtt = _8THMSEC(&thisRtt) + 8;
6974 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6976 /* the smoothed RTT time is RTT + 4*MDEV
6978 * We allow a user specified minimum to be set for this, to allow clamping
6979 * at a minimum value in the same way as TCP. In addition, we have to allow
6980 * for the possibility that this packet is answered by a delayed ACK, so we
6981 * add on a fixed 200ms to account for that timer expiring.
6984 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6985 rx_minPeerTimeout) + 200;
6986 clock_Zero(&call->rto);
6987 clock_Addmsec(&call->rto, rtt_timeout);
6989 /* Update the peer, so any new calls start with our values */
6990 peer->rtt_dev = call->rtt_dev;
6991 peer->rtt = call->rtt;
6993 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6994 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6998 /* Find all server connections that have not been active for a long time, and
7001 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
7004 struct clock now, when;
7005 struct rxevent *event;
7006 clock_GetTime(&now);
7008 /* Find server connection structures that haven't been used for
7009 * greater than rx_idleConnectionTime */
7011 struct rx_connection **conn_ptr, **conn_end;
7012 int i, havecalls = 0;
7013 MUTEX_ENTER(&rx_connHashTable_lock);
7014 for (conn_ptr = &rx_connHashTable[0], conn_end =
7015 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
7017 struct rx_connection *conn, *next;
7018 struct rx_call *call;
7022 for (conn = *conn_ptr; conn; conn = next) {
7023 /* XXX -- Shouldn't the connection be locked? */
7026 for (i = 0; i < RX_MAXCALLS; i++) {
7027 call = conn->call[i];
7031 code = MUTEX_TRYENTER(&call->lock);
7034 result = rxi_CheckCall(call, 1);
7035 MUTEX_EXIT(&call->lock);
7037 /* If CheckCall freed the call, it might
7038 * have destroyed the connection as well,
7039 * which screws up the linked lists.
7045 if (conn->type == RX_SERVER_CONNECTION) {
7046 /* This only actually destroys the connection if
7047 * there are no outstanding calls */
7048 MUTEX_ENTER(&conn->conn_data_lock);
7049 MUTEX_ENTER(&rx_refcnt_mutex);
7050 if (!havecalls && !conn->refCount
7051 && ((conn->lastSendTime + rx_idleConnectionTime) <
7053 conn->refCount++; /* it will be decr in rx_DestroyConn */
7054 MUTEX_EXIT(&rx_refcnt_mutex);
7055 MUTEX_EXIT(&conn->conn_data_lock);
7056 #ifdef RX_ENABLE_LOCKS
7057 rxi_DestroyConnectionNoLock(conn);
7058 #else /* RX_ENABLE_LOCKS */
7059 rxi_DestroyConnection(conn);
7060 #endif /* RX_ENABLE_LOCKS */
7062 #ifdef RX_ENABLE_LOCKS
7064 MUTEX_EXIT(&rx_refcnt_mutex);
7065 MUTEX_EXIT(&conn->conn_data_lock);
7067 #endif /* RX_ENABLE_LOCKS */
7071 #ifdef RX_ENABLE_LOCKS
7072 while (rx_connCleanup_list) {
7073 struct rx_connection *conn;
7074 conn = rx_connCleanup_list;
7075 rx_connCleanup_list = rx_connCleanup_list->next;
7076 MUTEX_EXIT(&rx_connHashTable_lock);
7077 rxi_CleanupConnection(conn);
7078 MUTEX_ENTER(&rx_connHashTable_lock);
7080 MUTEX_EXIT(&rx_connHashTable_lock);
7081 #endif /* RX_ENABLE_LOCKS */
7084 /* Find any peer structures that haven't been used (haven't had an
7085 * associated connection) for greater than rx_idlePeerTime */
7087 struct rx_peer **peer_ptr, **peer_end;
7091 * Why do we need to hold the rx_peerHashTable_lock across
7092 * the incrementing of peer_ptr since the rx_peerHashTable
7093 * array is not changing? We don't.
7095 * By dropping the lock periodically we can permit other
7096 * activities to be performed while a rxi_ReapConnections
7097 * call is in progress. The goal of reap connections
7098 * is to clean up quickly without causing large amounts
7099 * of contention. Therefore, it is important that global
7100 * mutexes not be held for extended periods of time.
7102 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7103 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7105 struct rx_peer *peer, *next, *prev;
7107 MUTEX_ENTER(&rx_peerHashTable_lock);
7108 for (prev = peer = *peer_ptr; peer; peer = next) {
7110 code = MUTEX_TRYENTER(&peer->peer_lock);
7111 if ((code) && (peer->refCount == 0)
7112 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7113 struct opr_queue *cursor, *store;
7117 * now know that this peer object is one to be
7118 * removed from the hash table. Once it is removed
7119 * it can't be referenced by other threads.
7120 * Lets remove it first and decrement the struct
7121 * nPeerStructs count.
7123 if (peer == *peer_ptr) {
7129 if (rx_stats_active)
7130 rx_atomic_dec(&rx_stats.nPeerStructs);
7133 * Now if we hold references on 'prev' and 'next'
7134 * we can safely drop the rx_peerHashTable_lock
7135 * while we destroy this 'peer' object.
7141 MUTEX_EXIT(&rx_peerHashTable_lock);
7143 MUTEX_EXIT(&peer->peer_lock);
7144 MUTEX_DESTROY(&peer->peer_lock);
7146 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7147 unsigned int num_funcs;
7148 struct rx_interface_stat *rpc_stat
7149 = opr_queue_Entry(cursor, struct rx_interface_stat,
7154 opr_queue_Remove(&rpc_stat->entry);
7155 opr_queue_Remove(&rpc_stat->entryPeers);
7157 num_funcs = rpc_stat->stats[0].func_total;
7159 sizeof(rx_interface_stat_t) +
7160 rpc_stat->stats[0].func_total *
7161 sizeof(rx_function_entry_v1_t);
7163 rxi_Free(rpc_stat, space);
7165 MUTEX_ENTER(&rx_rpc_stats);
7166 rxi_rpc_peer_stat_cnt -= num_funcs;
7167 MUTEX_EXIT(&rx_rpc_stats);
7172 * Regain the rx_peerHashTable_lock and
7173 * decrement the reference count on 'prev'
7176 MUTEX_ENTER(&rx_peerHashTable_lock);
7183 MUTEX_EXIT(&peer->peer_lock);
7188 MUTEX_EXIT(&rx_peerHashTable_lock);
7192 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7193 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7194 * GC, just below. Really, we shouldn't have to keep moving packets from
7195 * one place to another, but instead ought to always know if we can
7196 * afford to hold onto a packet in its particular use. */
7197 MUTEX_ENTER(&rx_freePktQ_lock);
7198 if (rx_waitingForPackets) {
7199 rx_waitingForPackets = 0;
7200 #ifdef RX_ENABLE_LOCKS
7201 CV_BROADCAST(&rx_waitingForPackets_cv);
7203 osi_rxWakeup(&rx_waitingForPackets);
7206 MUTEX_EXIT(&rx_freePktQ_lock);
7209 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7210 event = rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0);
7211 rxevent_Put(&event);
7215 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7216 * rx.h is sort of strange this is better. This is called with a security
7217 * object before it is discarded. Each connection using a security object has
7218 * its own refcount to the object so it won't actually be freed until the last
7219 * connection is destroyed.
7221 * This is the only rxs module call. A hold could also be written but no one
7225 rxs_Release(struct rx_securityClass *aobj)
7227 return RXS_Close(aobj);
7235 #define TRACE_OPTION_RX_DEBUG 16
7243 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7244 0, KEY_QUERY_VALUE, &parmKey);
7245 if (code != ERROR_SUCCESS)
7248 dummyLen = sizeof(TraceOption);
7249 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7250 (BYTE *) &TraceOption, &dummyLen);
7251 if (code == ERROR_SUCCESS) {
7252 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7254 RegCloseKey (parmKey);
7255 #endif /* AFS_NT40_ENV */
7260 rx_DebugOnOff(int on)
7264 rxdebug_active = on;
7270 rx_StatsOnOff(int on)
7272 rx_stats_active = on;
7276 /* Don't call this debugging routine directly; use dpf */
7278 rxi_DebugPrint(char *format, ...)
7287 va_start(ap, format);
7289 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7292 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7294 OutputDebugString(msg);
7300 va_start(ap, format);
7302 clock_GetTime(&now);
7303 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7304 (unsigned int)now.usec);
7305 vfprintf(rx_Log, format, ap);
7313 * This function is used to process the rx_stats structure that is local
7314 * to a process as well as an rx_stats structure received from a remote
7315 * process (via rxdebug). Therefore, it needs to do minimal version
7319 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7320 afs_int32 freePackets, char version)
7324 if (size != sizeof(struct rx_statistics)) {
7326 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7327 size, sizeof(struct rx_statistics));
7330 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7333 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7334 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7335 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7336 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7337 s->specialPktAllocFailures);
7339 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7340 s->receivePktAllocFailures, s->sendPktAllocFailures,
7341 s->specialPktAllocFailures);
7345 " greedy %u, " "bogusReads %u (last from host %x), "
7346 "noPackets %u, " "noBuffers %u, " "selects %u, "
7347 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7348 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7349 s->selects, s->sendSelects);
7351 fprintf(file, " packets read: ");
7352 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7353 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7355 fprintf(file, "\n");
7358 " other read counters: data %u, " "ack %u, " "dup %u "
7359 "spurious %u " "dally %u\n", s->dataPacketsRead,
7360 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7361 s->ignorePacketDally);
7363 fprintf(file, " packets sent: ");
7364 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7365 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7367 fprintf(file, "\n");
7370 " other send counters: ack %u, " "data %u (not resends), "
7371 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7372 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7373 s->dataPacketsPushed, s->ignoreAckedPacket);
7376 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7377 s->netSendFailures, (int)s->fatalErrors);
7379 if (s->nRttSamples) {
7380 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7381 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7383 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7384 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7388 " %d server connections, " "%d client connections, "
7389 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7390 s->nServerConns, s->nClientConns, s->nPeerStructs,
7391 s->nCallStructs, s->nFreeCallStructs);
7393 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7394 fprintf(file, " %d clock updates\n", clock_nUpdates);
7398 /* for backward compatibility */
7400 rx_PrintStats(FILE * file)
7402 MUTEX_ENTER(&rx_stats_mutex);
7403 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7404 sizeof(rx_stats), rx_nFreePackets,
7406 MUTEX_EXIT(&rx_stats_mutex);
7410 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7412 fprintf(file, "Peer %x.%d.\n",
7413 ntohl(peer->host), (int)ntohs(peer->port));
7416 " Rtt %d, " "total sent %d, " "resent %d\n",
7417 peer->rtt, peer->nSent, peer->reSends);
7419 fprintf(file, " Packet size %d\n", peer->ifMTU);
7423 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7425 * This mutex protects the following static variables:
7429 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7430 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7432 #define LOCK_RX_DEBUG
7433 #define UNLOCK_RX_DEBUG
7434 #endif /* AFS_PTHREAD_ENV */
7436 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7438 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7439 u_char type, void *inputData, size_t inputLength,
7440 void *outputData, size_t outputLength)
7442 static afs_int32 counter = 100;
7443 time_t waitTime, waitCount;
7444 struct rx_header theader;
7447 struct timeval tv_now, tv_wake, tv_delta;
7448 struct sockaddr_in taddr, faddr;
7462 tp = &tbuffer[sizeof(struct rx_header)];
7463 taddr.sin_family = AF_INET;
7464 taddr.sin_port = remotePort;
7465 taddr.sin_addr.s_addr = remoteAddr;
7466 memset(&taddr.sin_zero, 0, sizeof(taddr.sin_zero));
7467 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7468 taddr.sin_len = sizeof(struct sockaddr_in);
7471 memset(&theader, 0, sizeof(theader));
7472 theader.epoch = htonl(999);
7474 theader.callNumber = htonl(counter);
7477 theader.type = type;
7478 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7479 theader.serviceId = 0;
7481 memcpy(tbuffer, &theader, sizeof(theader));
7482 memcpy(tp, inputData, inputLength);
7484 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7485 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7487 /* see if there's a packet available */
7488 gettimeofday(&tv_wake, NULL);
7489 tv_wake.tv_sec += waitTime;
7492 FD_SET(socket, &imask);
7493 tv_delta.tv_sec = tv_wake.tv_sec;
7494 tv_delta.tv_usec = tv_wake.tv_usec;
7495 gettimeofday(&tv_now, NULL);
7497 if (tv_delta.tv_usec < tv_now.tv_usec) {
7499 tv_delta.tv_usec += 1000000;
7502 tv_delta.tv_usec -= tv_now.tv_usec;
7504 if (tv_delta.tv_sec < tv_now.tv_sec) {
7508 tv_delta.tv_sec -= tv_now.tv_sec;
7511 code = select(0, &imask, 0, 0, &tv_delta);
7512 #else /* AFS_NT40_ENV */
7513 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7514 #endif /* AFS_NT40_ENV */
7515 if (code == 1 && FD_ISSET(socket, &imask)) {
7516 /* now receive a packet */
7517 faddrLen = sizeof(struct sockaddr_in);
7519 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7520 (struct sockaddr *)&faddr, &faddrLen);
7523 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7524 if (counter == ntohl(theader.callNumber))
7532 /* see if we've timed out */
7540 code -= sizeof(struct rx_header);
7541 if (code > outputLength)
7542 code = outputLength;
7543 memcpy(outputData, tp, code);
7546 #endif /* RXDEBUG */
7549 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7550 afs_uint16 remotePort, struct rx_debugStats * stat,
7551 afs_uint32 * supportedValues)
7553 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7555 struct rx_debugIn in;
7557 *supportedValues = 0;
7558 in.type = htonl(RX_DEBUGI_GETSTATS);
7561 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7562 &in, sizeof(in), stat, sizeof(*stat));
7565 * If the call was successful, fixup the version and indicate
7566 * what contents of the stat structure are valid.
7567 * Also do net to host conversion of fields here.
7571 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7572 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7574 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7575 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7577 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7578 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7580 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7581 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7583 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7584 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7586 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7587 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7589 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7590 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7592 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7593 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7595 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7596 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7598 stat->nFreePackets = ntohl(stat->nFreePackets);
7599 stat->packetReclaims = ntohl(stat->packetReclaims);
7600 stat->callsExecuted = ntohl(stat->callsExecuted);
7601 stat->nWaiting = ntohl(stat->nWaiting);
7602 stat->idleThreads = ntohl(stat->idleThreads);
7603 stat->nWaited = ntohl(stat->nWaited);
7604 stat->nPackets = ntohl(stat->nPackets);
7613 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7614 afs_uint16 remotePort, struct rx_statistics * stat,
7615 afs_uint32 * supportedValues)
7617 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7619 struct rx_debugIn in;
7620 afs_int32 *lp = (afs_int32 *) stat;
7624 * supportedValues is currently unused, but added to allow future
7625 * versioning of this function.
7628 *supportedValues = 0;
7629 in.type = htonl(RX_DEBUGI_RXSTATS);
7631 memset(stat, 0, sizeof(*stat));
7633 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7634 &in, sizeof(in), stat, sizeof(*stat));
7639 * Do net to host conversion here
7642 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7653 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7654 afs_uint16 remotePort, size_t version_length,
7657 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7659 return MakeDebugCall(socket, remoteAddr, remotePort,
7660 RX_PACKET_TYPE_VERSION, a, 1, version,
7668 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7669 afs_uint16 remotePort, afs_int32 * nextConnection,
7670 int allConnections, afs_uint32 debugSupportedValues,
7671 struct rx_debugConn * conn,
7672 afs_uint32 * supportedValues)
7674 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7676 struct rx_debugIn in;
7680 * supportedValues is currently unused, but added to allow future
7681 * versioning of this function.
7684 *supportedValues = 0;
7685 if (allConnections) {
7686 in.type = htonl(RX_DEBUGI_GETALLCONN);
7688 in.type = htonl(RX_DEBUGI_GETCONN);
7690 in.index = htonl(*nextConnection);
7691 memset(conn, 0, sizeof(*conn));
7693 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7694 &in, sizeof(in), conn, sizeof(*conn));
7697 *nextConnection += 1;
7700 * Convert old connection format to new structure.
7703 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7704 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7705 #define MOVEvL(a) (conn->a = vL->a)
7707 /* any old or unrecognized version... */
7708 for (i = 0; i < RX_MAXCALLS; i++) {
7709 MOVEvL(callState[i]);
7710 MOVEvL(callMode[i]);
7711 MOVEvL(callFlags[i]);
7712 MOVEvL(callOther[i]);
7714 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7715 MOVEvL(secStats.type);
7716 MOVEvL(secStats.level);
7717 MOVEvL(secStats.flags);
7718 MOVEvL(secStats.expires);
7719 MOVEvL(secStats.packetsReceived);
7720 MOVEvL(secStats.packetsSent);
7721 MOVEvL(secStats.bytesReceived);
7722 MOVEvL(secStats.bytesSent);
7727 * Do net to host conversion here
7729 * I don't convert host or port since we are most likely
7730 * going to want these in NBO.
7732 conn->cid = ntohl(conn->cid);
7733 conn->serial = ntohl(conn->serial);
7734 for (i = 0; i < RX_MAXCALLS; i++) {
7735 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7737 conn->error = ntohl(conn->error);
7738 conn->secStats.flags = ntohl(conn->secStats.flags);
7739 conn->secStats.expires = ntohl(conn->secStats.expires);
7740 conn->secStats.packetsReceived =
7741 ntohl(conn->secStats.packetsReceived);
7742 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7743 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7744 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7745 conn->epoch = ntohl(conn->epoch);
7746 conn->natMTU = ntohl(conn->natMTU);
7755 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7756 afs_uint16 remotePort, afs_int32 * nextPeer,
7757 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7758 afs_uint32 * supportedValues)
7760 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7762 struct rx_debugIn in;
7765 * supportedValues is currently unused, but added to allow future
7766 * versioning of this function.
7769 *supportedValues = 0;
7770 in.type = htonl(RX_DEBUGI_GETPEER);
7771 in.index = htonl(*nextPeer);
7772 memset(peer, 0, sizeof(*peer));
7774 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7775 &in, sizeof(in), peer, sizeof(*peer));
7781 * Do net to host conversion here
7783 * I don't convert host or port since we are most likely
7784 * going to want these in NBO.
7786 peer->ifMTU = ntohs(peer->ifMTU);
7787 peer->idleWhen = ntohl(peer->idleWhen);
7788 peer->refCount = ntohs(peer->refCount);
7789 peer->rtt = ntohl(peer->rtt);
7790 peer->rtt_dev = ntohl(peer->rtt_dev);
7791 peer->timeout.sec = 0;
7792 peer->timeout.usec = 0;
7793 peer->nSent = ntohl(peer->nSent);
7794 peer->reSends = ntohl(peer->reSends);
7795 peer->natMTU = ntohs(peer->natMTU);
7796 peer->maxMTU = ntohs(peer->maxMTU);
7797 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7798 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7799 peer->MTU = ntohs(peer->MTU);
7800 peer->cwind = ntohs(peer->cwind);
7801 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7802 peer->congestSeq = ntohs(peer->congestSeq);
7803 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7804 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7805 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7806 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7815 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7816 struct rx_debugPeer * peerStats)
7819 afs_int32 error = 1; /* default to "did not succeed" */
7820 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7822 MUTEX_ENTER(&rx_peerHashTable_lock);
7823 for(tp = rx_peerHashTable[hashValue];
7824 tp != NULL; tp = tp->next) {
7825 if (tp->host == peerHost)
7831 MUTEX_EXIT(&rx_peerHashTable_lock);
7835 MUTEX_ENTER(&tp->peer_lock);
7836 peerStats->host = tp->host;
7837 peerStats->port = tp->port;
7838 peerStats->ifMTU = tp->ifMTU;
7839 peerStats->idleWhen = tp->idleWhen;
7840 peerStats->refCount = tp->refCount;
7841 peerStats->burstSize = 0;
7842 peerStats->burst = 0;
7843 peerStats->burstWait.sec = 0;
7844 peerStats->burstWait.usec = 0;
7845 peerStats->rtt = tp->rtt;
7846 peerStats->rtt_dev = tp->rtt_dev;
7847 peerStats->timeout.sec = 0;
7848 peerStats->timeout.usec = 0;
7849 peerStats->nSent = tp->nSent;
7850 peerStats->reSends = tp->reSends;
7851 peerStats->natMTU = tp->natMTU;
7852 peerStats->maxMTU = tp->maxMTU;
7853 peerStats->maxDgramPackets = tp->maxDgramPackets;
7854 peerStats->ifDgramPackets = tp->ifDgramPackets;
7855 peerStats->MTU = tp->MTU;
7856 peerStats->cwind = tp->cwind;
7857 peerStats->nDgramPackets = tp->nDgramPackets;
7858 peerStats->congestSeq = tp->congestSeq;
7859 peerStats->bytesSent.high = tp->bytesSent >> 32;
7860 peerStats->bytesSent.low = tp->bytesSent & MAX_AFS_UINT32;
7861 peerStats->bytesReceived.high = tp->bytesReceived >> 32;
7862 peerStats->bytesReceived.low
7863 = tp->bytesReceived & MAX_AFS_UINT32;
7864 MUTEX_EXIT(&tp->peer_lock);
7866 MUTEX_ENTER(&rx_peerHashTable_lock);
7869 MUTEX_EXIT(&rx_peerHashTable_lock);
7877 struct rx_serverQueueEntry *np;
7880 struct rx_call *call;
7881 struct rx_serverQueueEntry *sq;
7884 if (rx_atomic_test_and_set_bit(&rxinit_status, 0))
7885 return; /* Already shutdown. */
7889 #ifndef AFS_PTHREAD_ENV
7890 FD_ZERO(&rx_selectMask);
7891 #endif /* AFS_PTHREAD_ENV */
7892 rxi_dataQuota = RX_MAX_QUOTA;
7893 #ifndef AFS_PTHREAD_ENV
7895 #endif /* AFS_PTHREAD_ENV */
7898 #ifndef AFS_PTHREAD_ENV
7899 #ifndef AFS_USE_GETTIMEOFDAY
7901 #endif /* AFS_USE_GETTIMEOFDAY */
7902 #endif /* AFS_PTHREAD_ENV */
7904 while (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
7905 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
7906 opr_queue_Remove(&call->entry);
7907 rxi_Free(call, sizeof(struct rx_call));
7910 while (!opr_queue_IsEmpty(&rx_idleServerQueue)) {
7911 sq = opr_queue_First(&rx_idleServerQueue, struct rx_serverQueueEntry,
7913 opr_queue_Remove(&sq->entry);
7918 struct rx_peer **peer_ptr, **peer_end;
7919 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7920 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7922 struct rx_peer *peer, *next;
7924 MUTEX_ENTER(&rx_peerHashTable_lock);
7925 for (peer = *peer_ptr; peer; peer = next) {
7926 struct opr_queue *cursor, *store;
7929 MUTEX_ENTER(&rx_rpc_stats);
7930 MUTEX_ENTER(&peer->peer_lock);
7931 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7932 unsigned int num_funcs;
7933 struct rx_interface_stat *rpc_stat
7934 = opr_queue_Entry(cursor, struct rx_interface_stat,
7938 opr_queue_Remove(&rpc_stat->entry);
7939 opr_queue_Remove(&rpc_stat->entryPeers);
7940 num_funcs = rpc_stat->stats[0].func_total;
7942 sizeof(rx_interface_stat_t) +
7943 rpc_stat->stats[0].func_total *
7944 sizeof(rx_function_entry_v1_t);
7946 rxi_Free(rpc_stat, space);
7948 /* rx_rpc_stats must be held */
7949 rxi_rpc_peer_stat_cnt -= num_funcs;
7951 MUTEX_EXIT(&peer->peer_lock);
7952 MUTEX_EXIT(&rx_rpc_stats);
7956 if (rx_stats_active)
7957 rx_atomic_dec(&rx_stats.nPeerStructs);
7959 MUTEX_EXIT(&rx_peerHashTable_lock);
7962 for (i = 0; i < RX_MAX_SERVICES; i++) {
7964 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7966 for (i = 0; i < rx_hashTableSize; i++) {
7967 struct rx_connection *tc, *ntc;
7968 MUTEX_ENTER(&rx_connHashTable_lock);
7969 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7971 for (j = 0; j < RX_MAXCALLS; j++) {
7973 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7976 rxi_Free(tc, sizeof(*tc));
7978 MUTEX_EXIT(&rx_connHashTable_lock);
7981 MUTEX_ENTER(&freeSQEList_lock);
7983 while ((np = rx_FreeSQEList)) {
7984 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7985 MUTEX_DESTROY(&np->lock);
7986 rxi_Free(np, sizeof(*np));
7989 MUTEX_EXIT(&freeSQEList_lock);
7990 MUTEX_DESTROY(&freeSQEList_lock);
7991 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7992 MUTEX_DESTROY(&rx_connHashTable_lock);
7993 MUTEX_DESTROY(&rx_peerHashTable_lock);
7994 MUTEX_DESTROY(&rx_serverPool_lock);
7996 osi_Free(rx_connHashTable,
7997 rx_hashTableSize * sizeof(struct rx_connection *));
7998 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8000 UNPIN(rx_connHashTable,
8001 rx_hashTableSize * sizeof(struct rx_connection *));
8002 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8004 rxi_FreeAllPackets();
8006 MUTEX_ENTER(&rx_quota_mutex);
8007 rxi_dataQuota = RX_MAX_QUOTA;
8008 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
8009 MUTEX_EXIT(&rx_quota_mutex);
8015 * Routines to implement connection specific data.
8019 rx_KeyCreate(rx_destructor_t rtn)
8022 MUTEX_ENTER(&rxi_keyCreate_lock);
8023 key = rxi_keyCreate_counter++;
8024 rxi_keyCreate_destructor = (rx_destructor_t *)
8025 realloc((void *)rxi_keyCreate_destructor,
8026 (key + 1) * sizeof(rx_destructor_t));
8027 rxi_keyCreate_destructor[key] = rtn;
8028 MUTEX_EXIT(&rxi_keyCreate_lock);
8033 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
8036 MUTEX_ENTER(&conn->conn_data_lock);
8037 if (!conn->specific) {
8038 conn->specific = malloc((key + 1) * sizeof(void *));
8039 for (i = 0; i < key; i++)
8040 conn->specific[i] = NULL;
8041 conn->nSpecific = key + 1;
8042 conn->specific[key] = ptr;
8043 } else if (key >= conn->nSpecific) {
8044 conn->specific = (void **)
8045 realloc(conn->specific, (key + 1) * sizeof(void *));
8046 for (i = conn->nSpecific; i < key; i++)
8047 conn->specific[i] = NULL;
8048 conn->nSpecific = key + 1;
8049 conn->specific[key] = ptr;
8051 if (conn->specific[key] && rxi_keyCreate_destructor[key])
8052 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
8053 conn->specific[key] = ptr;
8055 MUTEX_EXIT(&conn->conn_data_lock);
8059 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
8062 MUTEX_ENTER(&svc->svc_data_lock);
8063 if (!svc->specific) {
8064 svc->specific = malloc((key + 1) * sizeof(void *));
8065 for (i = 0; i < key; i++)
8066 svc->specific[i] = NULL;
8067 svc->nSpecific = key + 1;
8068 svc->specific[key] = ptr;
8069 } else if (key >= svc->nSpecific) {
8070 svc->specific = (void **)
8071 realloc(svc->specific, (key + 1) * sizeof(void *));
8072 for (i = svc->nSpecific; i < key; i++)
8073 svc->specific[i] = NULL;
8074 svc->nSpecific = key + 1;
8075 svc->specific[key] = ptr;
8077 if (svc->specific[key] && rxi_keyCreate_destructor[key])
8078 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
8079 svc->specific[key] = ptr;
8081 MUTEX_EXIT(&svc->svc_data_lock);
8085 rx_GetSpecific(struct rx_connection *conn, int key)
8088 MUTEX_ENTER(&conn->conn_data_lock);
8089 if (key >= conn->nSpecific)
8092 ptr = conn->specific[key];
8093 MUTEX_EXIT(&conn->conn_data_lock);
8098 rx_GetServiceSpecific(struct rx_service *svc, int key)
8101 MUTEX_ENTER(&svc->svc_data_lock);
8102 if (key >= svc->nSpecific)
8105 ptr = svc->specific[key];
8106 MUTEX_EXIT(&svc->svc_data_lock);
8111 #endif /* !KERNEL */
8114 * processStats is a queue used to store the statistics for the local
8115 * process. Its contents are similar to the contents of the rpcStats
8116 * queue on a rx_peer structure, but the actual data stored within
8117 * this queue contains totals across the lifetime of the process (assuming
8118 * the stats have not been reset) - unlike the per peer structures
8119 * which can come and go based upon the peer lifetime.
8122 static struct opr_queue processStats = { &processStats, &processStats };
8125 * peerStats is a queue used to store the statistics for all peer structs.
8126 * Its contents are the union of all the peer rpcStats queues.
8129 static struct opr_queue peerStats = { &peerStats, &peerStats };
8132 * rxi_monitor_processStats is used to turn process wide stat collection
8136 static int rxi_monitor_processStats = 0;
8139 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8142 static int rxi_monitor_peerStats = 0;
8146 rxi_ClearRPCOpStat(rx_function_entry_v1_p rpc_stat)
8148 rpc_stat->invocations = 0;
8149 rpc_stat->bytes_sent = 0;
8150 rpc_stat->bytes_rcvd = 0;
8151 rpc_stat->queue_time_sum.sec = 0;
8152 rpc_stat->queue_time_sum.usec = 0;
8153 rpc_stat->queue_time_sum_sqr.sec = 0;
8154 rpc_stat->queue_time_sum_sqr.usec = 0;
8155 rpc_stat->queue_time_min.sec = 9999999;
8156 rpc_stat->queue_time_min.usec = 9999999;
8157 rpc_stat->queue_time_max.sec = 0;
8158 rpc_stat->queue_time_max.usec = 0;
8159 rpc_stat->execution_time_sum.sec = 0;
8160 rpc_stat->execution_time_sum.usec = 0;
8161 rpc_stat->execution_time_sum_sqr.sec = 0;
8162 rpc_stat->execution_time_sum_sqr.usec = 0;
8163 rpc_stat->execution_time_min.sec = 9999999;
8164 rpc_stat->execution_time_min.usec = 9999999;
8165 rpc_stat->execution_time_max.sec = 0;
8166 rpc_stat->execution_time_max.usec = 0;
8170 * Given all of the information for a particular rpc
8171 * call, find or create (if requested) the stat structure for the rpc.
8174 * the queue of stats that will be updated with the new value
8176 * @param rxInterface
8177 * a unique number that identifies the rpc interface
8180 * the total number of functions in this interface. this is only
8181 * required if create is true
8184 * if true, this invocation was made to a server
8187 * the ip address of the remote host. this is only required if create
8188 * and addToPeerList are true
8191 * the port of the remote host. this is only required if create
8192 * and addToPeerList are true
8194 * @param addToPeerList
8195 * if != 0, add newly created stat to the global peer list
8198 * if a new stats structure is allocated, the counter will
8199 * be updated with the new number of allocated stat structures.
8200 * only required if create is true
8203 * if no stats structure exists, allocate one
8207 static rx_interface_stat_p
8208 rxi_FindRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8209 afs_uint32 totalFunc, int isServer, afs_uint32 remoteHost,
8210 afs_uint32 remotePort, int addToPeerList,
8211 unsigned int *counter, int create)
8213 rx_interface_stat_p rpc_stat = NULL;
8214 struct opr_queue *cursor;
8217 * See if there's already a structure for this interface
8220 for (opr_queue_Scan(stats, cursor)) {
8221 rpc_stat = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8223 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8224 && (rpc_stat->stats[0].remote_is_server == isServer))
8228 /* if they didn't ask us to create, we're done */
8230 if (opr_queue_IsEnd(stats, cursor))
8236 /* can't proceed without these */
8237 if (!totalFunc || !counter)
8241 * Didn't find a match so allocate a new structure and add it to the
8245 if (opr_queue_IsEnd(stats, cursor) || (rpc_stat == NULL)
8246 || (rpc_stat->stats[0].interfaceId != rxInterface)
8247 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8252 sizeof(rx_interface_stat_t) +
8253 totalFunc * sizeof(rx_function_entry_v1_t);
8255 rpc_stat = rxi_Alloc(space);
8256 if (rpc_stat == NULL)
8259 *counter += totalFunc;
8260 for (i = 0; i < totalFunc; i++) {
8261 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8262 rpc_stat->stats[i].remote_peer = remoteHost;
8263 rpc_stat->stats[i].remote_port = remotePort;
8264 rpc_stat->stats[i].remote_is_server = isServer;
8265 rpc_stat->stats[i].interfaceId = rxInterface;
8266 rpc_stat->stats[i].func_total = totalFunc;
8267 rpc_stat->stats[i].func_index = i;
8269 opr_queue_Prepend(stats, &rpc_stat->entry);
8270 if (addToPeerList) {
8271 opr_queue_Prepend(&peerStats, &rpc_stat->entryPeers);
8278 rx_ClearProcessRPCStats(afs_int32 rxInterface)
8280 rx_interface_stat_p rpc_stat;
8283 if (rxInterface == -1)
8286 MUTEX_ENTER(&rx_rpc_stats);
8287 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8290 totalFunc = rpc_stat->stats[0].func_total;
8291 for (i = 0; i < totalFunc; i++)
8292 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8294 MUTEX_EXIT(&rx_rpc_stats);
8299 rx_ClearPeerRPCStats(afs_int32 rxInterface, afs_uint32 peerHost, afs_uint16 peerPort)
8301 rx_interface_stat_p rpc_stat;
8303 struct rx_peer * peer;
8305 if (rxInterface == -1)
8308 peer = rxi_FindPeer(peerHost, peerPort, 0);
8312 MUTEX_ENTER(&rx_rpc_stats);
8313 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8316 totalFunc = rpc_stat->stats[0].func_total;
8317 for (i = 0; i < totalFunc; i++)
8318 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8320 MUTEX_EXIT(&rx_rpc_stats);
8325 rx_CopyProcessRPCStats(afs_uint64 op)
8327 rx_interface_stat_p rpc_stat;
8328 rx_function_entry_v1_p rpcop_stat =
8329 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8330 int currentFunc = (op & MAX_AFS_UINT32);
8331 afs_int32 rxInterface = (op >> 32);
8333 if (!rxi_monitor_processStats)
8336 if (rxInterface == -1)
8339 if (rpcop_stat == NULL)
8342 MUTEX_ENTER(&rx_rpc_stats);
8343 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8346 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8347 sizeof(rx_function_entry_v1_t));
8348 MUTEX_EXIT(&rx_rpc_stats);
8350 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8357 rx_CopyPeerRPCStats(afs_uint64 op, afs_uint32 peerHost, afs_uint16 peerPort)
8359 rx_interface_stat_p rpc_stat;
8360 rx_function_entry_v1_p rpcop_stat =
8361 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8362 int currentFunc = (op & MAX_AFS_UINT32);
8363 afs_int32 rxInterface = (op >> 32);
8364 struct rx_peer *peer;
8366 if (!rxi_monitor_peerStats)
8369 if (rxInterface == -1)
8372 if (rpcop_stat == NULL)
8375 peer = rxi_FindPeer(peerHost, peerPort, 0);
8379 MUTEX_ENTER(&rx_rpc_stats);
8380 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8383 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8384 sizeof(rx_function_entry_v1_t));
8385 MUTEX_EXIT(&rx_rpc_stats);
8387 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8394 rx_ReleaseRPCStats(void *stats)
8397 rxi_Free(stats, sizeof(rx_function_entry_v1_t));
8401 * Given all of the information for a particular rpc
8402 * call, create (if needed) and update the stat totals for the rpc.
8405 * the queue of stats that will be updated with the new value
8407 * @param rxInterface
8408 * a unique number that identifies the rpc interface
8410 * @param currentFunc
8411 * the index of the function being invoked
8414 * the total number of functions in this interface
8417 * the amount of time this function waited for a thread
8420 * the amount of time this function invocation took to execute
8423 * the number bytes sent by this invocation
8426 * the number bytes received by this invocation
8429 * if true, this invocation was made to a server
8432 * the ip address of the remote host
8435 * the port of the remote host
8437 * @param addToPeerList
8438 * if != 0, add newly created stat to the global peer list
8441 * if a new stats structure is allocated, the counter will
8442 * be updated with the new number of allocated stat structures
8447 rxi_AddRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8448 afs_uint32 currentFunc, afs_uint32 totalFunc,
8449 struct clock *queueTime, struct clock *execTime,
8450 afs_uint64 bytesSent, afs_uint64 bytesRcvd, int isServer,
8451 afs_uint32 remoteHost, afs_uint32 remotePort,
8452 int addToPeerList, unsigned int *counter)
8455 rx_interface_stat_p rpc_stat;
8457 rpc_stat = rxi_FindRpcStat(stats, rxInterface, totalFunc, isServer,
8458 remoteHost, remotePort, addToPeerList, counter,
8466 * Increment the stats for this function
8469 rpc_stat->stats[currentFunc].invocations++;
8470 rpc_stat->stats[currentFunc].bytes_sent += bytesSent;
8471 rpc_stat->stats[currentFunc].bytes_rcvd += bytesRcvd;
8472 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8473 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8474 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8475 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8477 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8478 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8480 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8481 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8483 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8484 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8486 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8487 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8495 rxi_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8496 afs_uint32 currentFunc, afs_uint32 totalFunc,
8497 struct clock *queueTime, struct clock *execTime,
8498 afs_uint64 bytesSent, afs_uint64 bytesRcvd,
8502 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8505 MUTEX_ENTER(&rx_rpc_stats);
8507 if (rxi_monitor_peerStats) {
8508 MUTEX_ENTER(&peer->peer_lock);
8509 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8510 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8511 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8512 MUTEX_EXIT(&peer->peer_lock);
8515 if (rxi_monitor_processStats) {
8516 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8517 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8518 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8521 MUTEX_EXIT(&rx_rpc_stats);
8525 * Increment the times and count for a particular rpc function.
8527 * Traditionally this call was invoked from rxgen stubs. Modern stubs
8528 * call rx_RecordCallStatistics instead, so the public version of this
8529 * function is left purely for legacy callers.
8532 * The peer who invoked the rpc
8534 * @param rxInterface
8535 * A unique number that identifies the rpc interface
8537 * @param currentFunc
8538 * The index of the function being invoked
8541 * The total number of functions in this interface
8544 * The amount of time this function waited for a thread
8547 * The amount of time this function invocation took to execute
8550 * The number bytes sent by this invocation
8553 * The number bytes received by this invocation
8556 * If true, this invocation was made to a server
8560 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8561 afs_uint32 currentFunc, afs_uint32 totalFunc,
8562 struct clock *queueTime, struct clock *execTime,
8563 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8569 sent64 = ((afs_uint64)bytesSent->high << 32) + bytesSent->low;
8570 rcvd64 = ((afs_uint64)bytesRcvd->high << 32) + bytesRcvd->low;
8572 rxi_IncrementTimeAndCount(peer, rxInterface, currentFunc, totalFunc,
8573 queueTime, execTime, sent64, rcvd64,
8580 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8584 * IN callerVersion - the rpc stat version of the caller.
8586 * IN count - the number of entries to marshall.
8588 * IN stats - pointer to stats to be marshalled.
8590 * OUT ptr - Where to store the marshalled data.
8597 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8598 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8604 * We only support the first version
8606 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8607 *(ptr++) = stats->remote_peer;
8608 *(ptr++) = stats->remote_port;
8609 *(ptr++) = stats->remote_is_server;
8610 *(ptr++) = stats->interfaceId;
8611 *(ptr++) = stats->func_total;
8612 *(ptr++) = stats->func_index;
8613 *(ptr++) = stats->invocations >> 32;
8614 *(ptr++) = stats->invocations & MAX_AFS_UINT32;
8615 *(ptr++) = stats->bytes_sent >> 32;
8616 *(ptr++) = stats->bytes_sent & MAX_AFS_UINT32;
8617 *(ptr++) = stats->bytes_rcvd >> 32;
8618 *(ptr++) = stats->bytes_rcvd & MAX_AFS_UINT32;
8619 *(ptr++) = stats->queue_time_sum.sec;
8620 *(ptr++) = stats->queue_time_sum.usec;
8621 *(ptr++) = stats->queue_time_sum_sqr.sec;
8622 *(ptr++) = stats->queue_time_sum_sqr.usec;
8623 *(ptr++) = stats->queue_time_min.sec;
8624 *(ptr++) = stats->queue_time_min.usec;
8625 *(ptr++) = stats->queue_time_max.sec;
8626 *(ptr++) = stats->queue_time_max.usec;
8627 *(ptr++) = stats->execution_time_sum.sec;
8628 *(ptr++) = stats->execution_time_sum.usec;
8629 *(ptr++) = stats->execution_time_sum_sqr.sec;
8630 *(ptr++) = stats->execution_time_sum_sqr.usec;
8631 *(ptr++) = stats->execution_time_min.sec;
8632 *(ptr++) = stats->execution_time_min.usec;
8633 *(ptr++) = stats->execution_time_max.sec;
8634 *(ptr++) = stats->execution_time_max.usec;
8640 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8645 * IN callerVersion - the rpc stat version of the caller
8647 * OUT myVersion - the rpc stat version of this function
8649 * OUT clock_sec - local time seconds
8651 * OUT clock_usec - local time microseconds
8653 * OUT allocSize - the number of bytes allocated to contain stats
8655 * OUT statCount - the number stats retrieved from this process.
8657 * OUT stats - the actual stats retrieved from this process.
8661 * Returns void. If successful, stats will != NULL.
8665 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8666 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8667 size_t * allocSize, afs_uint32 * statCount,
8668 afs_uint32 ** stats)
8678 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8681 * Check to see if stats are enabled
8684 MUTEX_ENTER(&rx_rpc_stats);
8685 if (!rxi_monitor_processStats) {
8686 MUTEX_EXIT(&rx_rpc_stats);
8690 clock_GetTime(&now);
8691 *clock_sec = now.sec;
8692 *clock_usec = now.usec;
8695 * Allocate the space based upon the caller version
8697 * If the client is at an older version than we are,
8698 * we return the statistic data in the older data format, but
8699 * we still return our version number so the client knows we
8700 * are maintaining more data than it can retrieve.
8703 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8704 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8705 *statCount = rxi_rpc_process_stat_cnt;
8708 * This can't happen yet, but in the future version changes
8709 * can be handled by adding additional code here
8713 if (space > (size_t) 0) {
8715 ptr = *stats = rxi_Alloc(space);
8718 struct opr_queue *cursor;
8720 for (opr_queue_Scan(&processStats, cursor)) {
8721 struct rx_interface_stat *rpc_stat =
8722 opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8724 * Copy the data based upon the caller version
8726 rx_MarshallProcessRPCStats(callerVersion,
8727 rpc_stat->stats[0].func_total,
8728 rpc_stat->stats, &ptr);
8734 MUTEX_EXIT(&rx_rpc_stats);
8739 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8743 * IN callerVersion - the rpc stat version of the caller
8745 * OUT myVersion - the rpc stat version of this function
8747 * OUT clock_sec - local time seconds
8749 * OUT clock_usec - local time microseconds
8751 * OUT allocSize - the number of bytes allocated to contain stats
8753 * OUT statCount - the number of stats retrieved from the individual
8756 * OUT stats - the actual stats retrieved from the individual peer structures.
8760 * Returns void. If successful, stats will != NULL.
8764 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8765 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8766 size_t * allocSize, afs_uint32 * statCount,
8767 afs_uint32 ** stats)
8777 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8780 * Check to see if stats are enabled
8783 MUTEX_ENTER(&rx_rpc_stats);
8784 if (!rxi_monitor_peerStats) {
8785 MUTEX_EXIT(&rx_rpc_stats);
8789 clock_GetTime(&now);
8790 *clock_sec = now.sec;
8791 *clock_usec = now.usec;
8794 * Allocate the space based upon the caller version
8796 * If the client is at an older version than we are,
8797 * we return the statistic data in the older data format, but
8798 * we still return our version number so the client knows we
8799 * are maintaining more data than it can retrieve.
8802 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8803 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8804 *statCount = rxi_rpc_peer_stat_cnt;
8807 * This can't happen yet, but in the future version changes
8808 * can be handled by adding additional code here
8812 if (space > (size_t) 0) {
8814 ptr = *stats = rxi_Alloc(space);
8817 struct opr_queue *cursor;
8819 for (opr_queue_Scan(&peerStats, cursor)) {
8820 struct rx_interface_stat *rpc_stat
8821 = opr_queue_Entry(cursor, struct rx_interface_stat,
8825 * Copy the data based upon the caller version
8827 rx_MarshallProcessRPCStats(callerVersion,
8828 rpc_stat->stats[0].func_total,
8829 rpc_stat->stats, &ptr);
8835 MUTEX_EXIT(&rx_rpc_stats);
8840 * rx_FreeRPCStats - free memory allocated by
8841 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8845 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8846 * rx_RetrievePeerRPCStats
8848 * IN allocSize - the number of bytes in stats.
8856 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8858 rxi_Free(stats, allocSize);
8862 * rx_queryProcessRPCStats - see if process rpc stat collection is
8863 * currently enabled.
8869 * Returns 0 if stats are not enabled != 0 otherwise
8873 rx_queryProcessRPCStats(void)
8876 MUTEX_ENTER(&rx_rpc_stats);
8877 rc = rxi_monitor_processStats;
8878 MUTEX_EXIT(&rx_rpc_stats);
8883 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8889 * Returns 0 if stats are not enabled != 0 otherwise
8893 rx_queryPeerRPCStats(void)
8896 MUTEX_ENTER(&rx_rpc_stats);
8897 rc = rxi_monitor_peerStats;
8898 MUTEX_EXIT(&rx_rpc_stats);
8903 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8913 rx_enableProcessRPCStats(void)
8915 MUTEX_ENTER(&rx_rpc_stats);
8916 rx_enable_stats = 1;
8917 rxi_monitor_processStats = 1;
8918 MUTEX_EXIT(&rx_rpc_stats);
8922 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8932 rx_enablePeerRPCStats(void)
8934 MUTEX_ENTER(&rx_rpc_stats);
8935 rx_enable_stats = 1;
8936 rxi_monitor_peerStats = 1;
8937 MUTEX_EXIT(&rx_rpc_stats);
8941 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8951 rx_disableProcessRPCStats(void)
8953 struct opr_queue *cursor, *store;
8956 MUTEX_ENTER(&rx_rpc_stats);
8959 * Turn off process statistics and if peer stats is also off, turn
8963 rxi_monitor_processStats = 0;
8964 if (rxi_monitor_peerStats == 0) {
8965 rx_enable_stats = 0;
8968 for (opr_queue_ScanSafe(&processStats, cursor, store)) {
8969 unsigned int num_funcs = 0;
8970 struct rx_interface_stat *rpc_stat
8971 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8973 opr_queue_Remove(&rpc_stat->entry);
8975 num_funcs = rpc_stat->stats[0].func_total;
8977 sizeof(rx_interface_stat_t) +
8978 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8980 rxi_Free(rpc_stat, space);
8981 rxi_rpc_process_stat_cnt -= num_funcs;
8983 MUTEX_EXIT(&rx_rpc_stats);
8987 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8997 rx_disablePeerRPCStats(void)
8999 struct rx_peer **peer_ptr, **peer_end;
9003 * Turn off peer statistics and if process stats is also off, turn
9007 rxi_monitor_peerStats = 0;
9008 if (rxi_monitor_processStats == 0) {
9009 rx_enable_stats = 0;
9012 for (peer_ptr = &rx_peerHashTable[0], peer_end =
9013 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
9015 struct rx_peer *peer, *next, *prev;
9017 MUTEX_ENTER(&rx_peerHashTable_lock);
9018 MUTEX_ENTER(&rx_rpc_stats);
9019 for (prev = peer = *peer_ptr; peer; peer = next) {
9021 code = MUTEX_TRYENTER(&peer->peer_lock);
9024 struct opr_queue *cursor, *store;
9026 if (prev == *peer_ptr) {
9037 MUTEX_EXIT(&rx_peerHashTable_lock);
9039 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
9040 unsigned int num_funcs = 0;
9041 struct rx_interface_stat *rpc_stat
9042 = opr_queue_Entry(cursor, struct rx_interface_stat,
9045 opr_queue_Remove(&rpc_stat->entry);
9046 opr_queue_Remove(&rpc_stat->entryPeers);
9047 num_funcs = rpc_stat->stats[0].func_total;
9049 sizeof(rx_interface_stat_t) +
9050 rpc_stat->stats[0].func_total *
9051 sizeof(rx_function_entry_v1_t);
9053 rxi_Free(rpc_stat, space);
9054 rxi_rpc_peer_stat_cnt -= num_funcs;
9056 MUTEX_EXIT(&peer->peer_lock);
9058 MUTEX_ENTER(&rx_peerHashTable_lock);
9068 MUTEX_EXIT(&rx_rpc_stats);
9069 MUTEX_EXIT(&rx_peerHashTable_lock);
9074 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
9079 * IN clearFlag - flag indicating which stats to clear
9087 rx_clearProcessRPCStats(afs_uint32 clearFlag)
9089 struct opr_queue *cursor;
9091 MUTEX_ENTER(&rx_rpc_stats);
9093 for (opr_queue_Scan(&processStats, cursor)) {
9094 unsigned int num_funcs = 0, i;
9095 struct rx_interface_stat *rpc_stat
9096 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
9098 num_funcs = rpc_stat->stats[0].func_total;
9099 for (i = 0; i < num_funcs; i++) {
9100 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9101 rpc_stat->stats[i].invocations = 0;
9103 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9104 rpc_stat->stats[i].bytes_sent = 0;
9106 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9107 rpc_stat->stats[i].bytes_rcvd = 0;
9109 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9110 rpc_stat->stats[i].queue_time_sum.sec = 0;
9111 rpc_stat->stats[i].queue_time_sum.usec = 0;
9113 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9114 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9115 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9117 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9118 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9119 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9121 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9122 rpc_stat->stats[i].queue_time_max.sec = 0;
9123 rpc_stat->stats[i].queue_time_max.usec = 0;
9125 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9126 rpc_stat->stats[i].execution_time_sum.sec = 0;
9127 rpc_stat->stats[i].execution_time_sum.usec = 0;
9129 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9130 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9131 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9133 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9134 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9135 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9137 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9138 rpc_stat->stats[i].execution_time_max.sec = 0;
9139 rpc_stat->stats[i].execution_time_max.usec = 0;
9144 MUTEX_EXIT(&rx_rpc_stats);
9148 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
9153 * IN clearFlag - flag indicating which stats to clear
9161 rx_clearPeerRPCStats(afs_uint32 clearFlag)
9163 struct opr_queue *cursor;
9165 MUTEX_ENTER(&rx_rpc_stats);
9167 for (opr_queue_Scan(&peerStats, cursor)) {
9168 unsigned int num_funcs, i;
9169 struct rx_interface_stat *rpc_stat
9170 = opr_queue_Entry(cursor, struct rx_interface_stat, entryPeers);
9172 num_funcs = rpc_stat->stats[0].func_total;
9173 for (i = 0; i < num_funcs; i++) {
9174 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9175 rpc_stat->stats[i].invocations = 0;
9177 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9178 rpc_stat->stats[i].bytes_sent = 0;
9180 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9181 rpc_stat->stats[i].bytes_rcvd = 0;
9183 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9184 rpc_stat->stats[i].queue_time_sum.sec = 0;
9185 rpc_stat->stats[i].queue_time_sum.usec = 0;
9187 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9188 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9189 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9191 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9192 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9193 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9195 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9196 rpc_stat->stats[i].queue_time_max.sec = 0;
9197 rpc_stat->stats[i].queue_time_max.usec = 0;
9199 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9200 rpc_stat->stats[i].execution_time_sum.sec = 0;
9201 rpc_stat->stats[i].execution_time_sum.usec = 0;
9203 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9204 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9205 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9207 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9208 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9209 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9211 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9212 rpc_stat->stats[i].execution_time_max.sec = 0;
9213 rpc_stat->stats[i].execution_time_max.usec = 0;
9218 MUTEX_EXIT(&rx_rpc_stats);
9222 * rxi_rxstat_userok points to a routine that returns 1 if the caller
9223 * is authorized to enable/disable/clear RX statistics.
9225 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
9228 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
9230 rxi_rxstat_userok = proc;
9234 rx_RxStatUserOk(struct rx_call *call)
9236 if (!rxi_rxstat_userok)
9238 return rxi_rxstat_userok(call);
9243 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
9244 * function in the MSVC runtime DLL (msvcrt.dll).
9246 * Note: the system serializes calls to this function.
9249 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
9250 DWORD reason, /* reason function is being called */
9251 LPVOID reserved) /* reserved for future use */
9254 case DLL_PROCESS_ATTACH:
9255 /* library is being attached to a process */
9259 case DLL_PROCESS_DETACH:
9266 #endif /* AFS_NT40_ENV */
9269 int rx_DumpCalls(FILE *outputFile, char *cookie)
9271 #ifdef RXDEBUG_PACKET
9272 #ifdef KDUMP_RX_LOCK
9273 struct rx_call_rx_lock *c;
9280 #define RXDPRINTF sprintf
9281 #define RXDPRINTOUT output
9283 #define RXDPRINTF fprintf
9284 #define RXDPRINTOUT outputFile
9287 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9289 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9292 for (c = rx_allCallsp; c; c = c->allNextp) {
9293 u_short rqc, tqc, iovqc;
9295 MUTEX_ENTER(&c->lock);
9296 rqc = opr_queue_Count(&c->rq);
9297 tqc = opr_queue_Count(&c->tq);
9298 iovqc = opr_queue_Count(&c->app.iovq);
9300 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, "
9301 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9302 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9303 "resendEvent=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9304 "lastSendTime=%u, lastRecvTime=%u"
9305 #ifdef RX_ENABLE_LOCKS
9308 #ifdef RX_REFCOUNT_CHECK
9309 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9310 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9313 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,
9314 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9315 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9316 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9317 c->resendEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9318 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime
9319 #ifdef RX_ENABLE_LOCKS
9320 , (afs_uint32)c->refCount
9322 #ifdef RX_REFCOUNT_CHECK
9323 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9326 MUTEX_EXIT(&c->lock);
9329 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9332 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9334 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9336 #endif /* RXDEBUG_PACKET */