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;
207 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
208 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
210 /* Incoming calls wait on this queue when there are no available
211 * server processes */
212 struct opr_queue rx_incomingCallQueue;
214 /* Server processes wait on this queue when there are no appropriate
215 * calls to process */
216 struct opr_queue rx_idleServerQueue;
218 #if !defined(offsetof)
219 #include <stddef.h> /* for definition of offsetof() */
222 #ifdef RX_ENABLE_LOCKS
223 afs_kmutex_t rx_atomic_mutex;
226 /* Forward prototypes */
227 static struct rx_call * rxi_NewCall(struct rx_connection *, int);
230 putConnection (struct rx_connection *conn) {
231 MUTEX_ENTER(&rx_refcnt_mutex);
233 MUTEX_EXIT(&rx_refcnt_mutex);
236 #ifdef AFS_PTHREAD_ENV
239 * Use procedural initialization of mutexes/condition variables
243 extern afs_kmutex_t rx_quota_mutex;
244 extern afs_kmutex_t rx_pthread_mutex;
245 extern afs_kmutex_t rx_packets_mutex;
246 extern afs_kmutex_t rx_refcnt_mutex;
247 extern afs_kmutex_t des_init_mutex;
248 extern afs_kmutex_t des_random_mutex;
250 extern afs_kmutex_t rx_clock_mutex;
251 extern afs_kmutex_t rxi_connCacheMutex;
252 extern afs_kmutex_t event_handler_mutex;
253 extern afs_kmutex_t listener_mutex;
254 extern afs_kmutex_t rx_if_init_mutex;
255 extern afs_kmutex_t rx_if_mutex;
257 extern afs_kcondvar_t rx_event_handler_cond;
258 extern afs_kcondvar_t rx_listener_cond;
261 static afs_kmutex_t epoch_mutex;
262 static afs_kmutex_t rx_init_mutex;
263 static afs_kmutex_t rx_debug_mutex;
264 static afs_kmutex_t rx_rpc_stats;
267 rxi_InitPthread(void)
269 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
270 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
271 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
272 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
274 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
275 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
276 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
277 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
278 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
279 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
281 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
282 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
283 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
284 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
285 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
288 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
289 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
292 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
293 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
295 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
296 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
297 MUTEX_INIT(&rx_mallocedPktQ_lock, "rx_mallocedPktQ_lock", MUTEX_DEFAULT,
300 #ifdef RX_ENABLE_LOCKS
303 #endif /* RX_LOCKS_DB */
304 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
305 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
307 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
309 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
311 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
313 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
315 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
317 #endif /* RX_ENABLE_LOCKS */
320 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
321 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
323 * The rx_stats_mutex mutex protects the following global variables:
324 * rxi_lowConnRefCount
325 * rxi_lowPeerRefCount
334 * The rx_quota_mutex mutex protects the following global variables:
342 * The rx_freePktQ_lock protects the following global variables:
347 * The rx_packets_mutex mutex protects the following global variables:
355 * The rx_pthread_mutex mutex protects the following global variables:
356 * rxi_fcfs_thread_num
359 #define INIT_PTHREAD_LOCKS
363 /* Variables for handling the minProcs implementation. availProcs gives the
364 * number of threads available in the pool at this moment (not counting dudes
365 * executing right now). totalMin gives the total number of procs required
366 * for handling all minProcs requests. minDeficit is a dynamic variable
367 * tracking the # of procs required to satisfy all of the remaining minProcs
369 * For fine grain locking to work, the quota check and the reservation of
370 * a server thread has to come while rxi_availProcs and rxi_minDeficit
371 * are locked. To this end, the code has been modified under #ifdef
372 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
373 * same time. A new function, ReturnToServerPool() returns the allocation.
375 * A call can be on several queue's (but only one at a time). When
376 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
377 * that no one else is touching the queue. To this end, we store the address
378 * of the queue lock in the call structure (under the call lock) when we
379 * put the call on a queue, and we clear the call_queue_lock when the
380 * call is removed from a queue (once the call lock has been obtained).
381 * This allows rxi_ResetCall to safely synchronize with others wishing
382 * to manipulate the queue.
385 #if defined(RX_ENABLE_LOCKS)
386 static afs_kmutex_t rx_rpc_stats;
389 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
390 ** pretty good that the next packet coming in is from the same connection
391 ** as the last packet, since we're send multiple packets in a transmit window.
393 struct rx_connection *rxLastConn = 0;
395 #ifdef RX_ENABLE_LOCKS
396 /* The locking hierarchy for rx fine grain locking is composed of these
399 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
400 * also protects updates to rx_nextCid
401 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
402 * call->lock - locks call data fields.
403 * These are independent of each other:
404 * rx_freeCallQueue_lock
409 * serverQueueEntry->lock
410 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
412 * peer->lock - locks peer data fields.
413 * conn_data_lock - that more than one thread is not updating a conn data
414 * field at the same time.
425 * Do we need a lock to protect the peer field in the conn structure?
426 * conn->peer was previously a constant for all intents and so has no
427 * lock protecting this field. The multihomed client delta introduced
428 * a RX code change : change the peer field in the connection structure
429 * to that remote interface from which the last packet for this
430 * connection was sent out. This may become an issue if further changes
433 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
434 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
436 /* rxdb_fileID is used to identify the lock location, along with line#. */
437 static int rxdb_fileID = RXDB_FILE_RX;
438 #endif /* RX_LOCKS_DB */
439 #else /* RX_ENABLE_LOCKS */
440 #define SET_CALL_QUEUE_LOCK(C, L)
441 #define CLEAR_CALL_QUEUE_LOCK(C)
442 #endif /* RX_ENABLE_LOCKS */
443 struct rx_serverQueueEntry *rx_waitForPacket = 0;
445 /* ------------Exported Interfaces------------- */
447 /* Initialize rx. A port number may be mentioned, in which case this
448 * becomes the default port number for any service installed later.
449 * If 0 is provided for the port number, a random port will be chosen
450 * by the kernel. Whether this will ever overlap anything in
451 * /etc/services is anybody's guess... Returns 0 on success, -1 on
453 #if !(defined(AFS_NT40_ENV) || defined(RXK_UPCALL_ENV))
456 rx_atomic_t rxinit_status = RX_ATOMIC_INIT(1);
459 rx_InitHost(u_int host, u_int port)
466 char *htable, *ptable;
471 if (!rx_atomic_test_and_clear_bit(&rxinit_status, 0))
472 return 0; /* already started */
478 if (afs_winsockInit() < 0)
484 * Initialize anything necessary to provide a non-premptive threading
487 rxi_InitializeThreadSupport();
490 /* Allocate and initialize a socket for client and perhaps server
493 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
494 if (rx_socket == OSI_NULLSOCKET) {
497 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
500 #endif /* RX_LOCKS_DB */
501 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
502 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
503 MUTEX_INIT(&rx_atomic_mutex, "rx_atomic_mutex", MUTEX_DEFAULT, 0);
504 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
505 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
506 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
507 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
508 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
509 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
510 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
512 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
514 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
516 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
518 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
519 MUTEX_INIT(&rx_mallocedPktQ_lock, "rx_mallocedPktQ_lock", MUTEX_DEFAULT,
522 #if defined(AFS_HPUX110_ENV)
524 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
525 #endif /* AFS_HPUX110_ENV */
526 #endif /* RX_ENABLE_LOCKS && KERNEL */
529 rx_connDeadTime = 12;
530 rx_tranquil = 0; /* reset flag */
531 rxi_ResetStatistics();
532 htable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
533 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
534 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
535 ptable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
536 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
537 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
539 /* Malloc up a bunch of packets & buffers */
541 opr_queue_Init(&rx_freePacketQueue);
542 rxi_NeedMorePackets = FALSE;
543 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
544 opr_queue_Init(&rx_mallocedPacketQueue);
546 /* enforce a minimum number of allocated packets */
547 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
548 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
550 /* allocate the initial free packet pool */
551 #ifdef RX_ENABLE_TSFPQ
552 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
553 #else /* RX_ENABLE_TSFPQ */
554 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
555 #endif /* RX_ENABLE_TSFPQ */
562 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
563 tv.tv_sec = clock_now.sec;
564 tv.tv_usec = clock_now.usec;
565 srand((unsigned int)tv.tv_usec);
572 #if defined(KERNEL) && !defined(UKERNEL)
573 /* Really, this should never happen in a real kernel */
576 struct sockaddr_in addr;
578 int addrlen = sizeof(addr);
580 socklen_t addrlen = sizeof(addr);
582 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
584 osi_Free(htable, rx_hashTableSize * sizeof(struct rx_connection *));
587 rx_port = addr.sin_port;
590 rx_stats.minRtt.sec = 9999999;
591 if (RAND_bytes(&rx_epoch, sizeof(rx_epoch)) != 1)
593 rx_epoch = (rx_epoch & ~0x40000000) | 0x80000000;
594 if (RAND_bytes(&rx_nextCid, sizeof(rx_nextCid)) != 1)
596 rx_nextCid &= RX_CIDMASK;
597 MUTEX_ENTER(&rx_quota_mutex);
598 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
599 MUTEX_EXIT(&rx_quota_mutex);
600 /* *Slightly* random start time for the cid. This is just to help
601 * out with the hashing function at the peer */
602 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
603 rx_connHashTable = (struct rx_connection **)htable;
604 rx_peerHashTable = (struct rx_peer **)ptable;
606 rx_hardAckDelay.sec = 0;
607 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
609 rxevent_Init(20, rxi_ReScheduleEvents);
611 /* Initialize various global queues */
612 opr_queue_Init(&rx_idleServerQueue);
613 opr_queue_Init(&rx_incomingCallQueue);
614 opr_queue_Init(&rx_freeCallQueue);
616 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
617 /* Initialize our list of usable IP addresses. */
621 /* Start listener process (exact function is dependent on the
622 * implementation environment--kernel or user space) */
626 rx_atomic_clear_bit(&rxinit_status, 0);
633 return rx_InitHost(htonl(INADDR_ANY), port);
639 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
640 * maintaing the round trip timer.
645 * Start a new RTT timer for a given call and packet.
647 * There must be no resendEvent already listed for this call, otherwise this
648 * will leak events - intended for internal use within the RTO code only
651 * the RX call to start the timer for
652 * @param[in] lastPacket
653 * a flag indicating whether the last packet has been sent or not
655 * @pre call must be locked before calling this function
659 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
661 struct clock now, retryTime;
663 MUTEX_ASSERT(&call->lock);
667 clock_Add(&retryTime, &call->rto);
669 /* If we're sending the last packet, and we're the client, then the server
670 * may wait for an additional 400ms before returning the ACK, wait for it
671 * rather than hitting a timeout */
672 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
673 clock_Addmsec(&retryTime, 400);
675 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
676 call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
681 * Cancel an RTT timer for a given call.
685 * the RX call to cancel the timer for
687 * @pre call must be locked before calling this function
692 rxi_rto_cancel(struct rx_call *call)
694 MUTEX_ASSERT(&call->lock);
695 if (rxevent_Cancel(&call->resendEvent))
696 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
700 * Tell the RTO timer that we have sent a packet.
702 * If the timer isn't already running, then start it. If the timer is running,
706 * the RX call that the packet has been sent on
707 * @param[in] lastPacket
708 * A flag which is true if this is the last packet for the call
710 * @pre The call must be locked before calling this function
715 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
717 if (call->resendEvent)
720 rxi_rto_startTimer(call, lastPacket, istack);
724 * Tell the RTO timer that we have received an new ACK message
726 * This function should be called whenever a call receives an ACK that
727 * acknowledges new packets. Whatever happens, we stop the current timer.
728 * If there are unacked packets in the queue which have been sent, then
729 * we restart the timer from now. Otherwise, we leave it stopped.
732 * the RX call that the ACK has been received on
736 rxi_rto_packet_acked(struct rx_call *call, int istack)
738 struct opr_queue *cursor;
740 rxi_rto_cancel(call);
742 if (opr_queue_IsEmpty(&call->tq))
745 for (opr_queue_Scan(&call->tq, cursor)) {
746 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
747 if (p->header.seq > call->tfirst + call->twind)
750 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
751 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
759 * Set an initial round trip timeout for a peer connection
761 * @param[in] secs The timeout to set in seconds
765 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
766 peer->rtt = secs * 8000;
770 * Set a delayed ack event on the specified call for the given time
772 * @param[in] call - the call on which to set the event
773 * @param[in] offset - the delay from now after which the event fires
776 rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
778 struct clock now, when;
780 MUTEX_ASSERT(&call->lock);
783 clock_Add(&when, offset);
785 if (clock_Gt(&call->delayedAckTime, &when) &&
786 rxevent_Cancel(&call->delayedAckEvent)) {
787 /* We successfully cancelled an event too far in the future to install
788 * our new one; we can reuse the reference on the call. */
789 call->delayedAckEvent = rxevent_Post(&when, &now, rxi_SendDelayedAck,
792 call->delayedAckTime = when;
793 } else if (call->delayedAckEvent == NULL) {
794 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
795 call->delayedAckEvent = rxevent_Post(&when, &now,
798 call->delayedAckTime = when;
803 rxi_CancelDelayedAckEvent(struct rx_call *call)
805 MUTEX_ASSERT(&call->lock);
806 /* Only drop the ref if we cancelled it before it could run. */
807 if (rxevent_Cancel(&call->delayedAckEvent))
808 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
811 /* called with unincremented nRequestsRunning to see if it is OK to start
812 * a new thread in this service. Could be "no" for two reasons: over the
813 * max quota, or would prevent others from reaching their min quota.
815 #ifdef RX_ENABLE_LOCKS
816 /* This verion of QuotaOK reserves quota if it's ok while the
817 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
820 QuotaOK(struct rx_service *aservice)
822 /* check if over max quota */
823 if (aservice->nRequestsRunning >= aservice->maxProcs) {
827 /* under min quota, we're OK */
828 /* otherwise, can use only if there are enough to allow everyone
829 * to go to their min quota after this guy starts.
832 MUTEX_ENTER(&rx_quota_mutex);
833 if ((aservice->nRequestsRunning < aservice->minProcs)
834 || (rxi_availProcs > rxi_minDeficit)) {
835 aservice->nRequestsRunning++;
836 /* just started call in minProcs pool, need fewer to maintain
838 if (aservice->nRequestsRunning <= aservice->minProcs)
841 MUTEX_EXIT(&rx_quota_mutex);
844 MUTEX_EXIT(&rx_quota_mutex);
850 ReturnToServerPool(struct rx_service *aservice)
852 aservice->nRequestsRunning--;
853 MUTEX_ENTER(&rx_quota_mutex);
854 if (aservice->nRequestsRunning < aservice->minProcs)
857 MUTEX_EXIT(&rx_quota_mutex);
860 #else /* RX_ENABLE_LOCKS */
862 QuotaOK(struct rx_service *aservice)
865 /* under min quota, we're OK */
866 if (aservice->nRequestsRunning < aservice->minProcs)
869 /* check if over max quota */
870 if (aservice->nRequestsRunning >= aservice->maxProcs)
873 /* otherwise, can use only if there are enough to allow everyone
874 * to go to their min quota after this guy starts.
876 MUTEX_ENTER(&rx_quota_mutex);
877 if (rxi_availProcs > rxi_minDeficit)
879 MUTEX_EXIT(&rx_quota_mutex);
882 #endif /* RX_ENABLE_LOCKS */
885 /* Called by rx_StartServer to start up lwp's to service calls.
886 NExistingProcs gives the number of procs already existing, and which
887 therefore needn't be created. */
889 rxi_StartServerProcs(int nExistingProcs)
891 struct rx_service *service;
896 /* For each service, reserve N processes, where N is the "minimum"
897 * number of processes that MUST be able to execute a request in parallel,
898 * at any time, for that process. Also compute the maximum difference
899 * between any service's maximum number of processes that can run
900 * (i.e. the maximum number that ever will be run, and a guarantee
901 * that this number will run if other services aren't running), and its
902 * minimum number. The result is the extra number of processes that
903 * we need in order to provide the latter guarantee */
904 for (i = 0; i < RX_MAX_SERVICES; i++) {
906 service = rx_services[i];
907 if (service == (struct rx_service *)0)
909 nProcs += service->minProcs;
910 diff = service->maxProcs - service->minProcs;
914 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
915 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
916 for (i = 0; i < nProcs; i++) {
917 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
923 /* This routine is only required on Windows */
925 rx_StartClientThread(void)
927 #ifdef AFS_PTHREAD_ENV
929 pid = pthread_self();
930 #endif /* AFS_PTHREAD_ENV */
932 #endif /* AFS_NT40_ENV */
934 /* This routine must be called if any services are exported. If the
935 * donateMe flag is set, the calling process is donated to the server
938 rx_StartServer(int donateMe)
940 struct rx_service *service;
946 /* Start server processes, if necessary (exact function is dependent
947 * on the implementation environment--kernel or user space). DonateMe
948 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
949 * case, one less new proc will be created rx_StartServerProcs.
951 rxi_StartServerProcs(donateMe);
953 /* count up the # of threads in minProcs, and add set the min deficit to
954 * be that value, too.
956 for (i = 0; i < RX_MAX_SERVICES; i++) {
957 service = rx_services[i];
958 if (service == (struct rx_service *)0)
960 MUTEX_ENTER(&rx_quota_mutex);
961 rxi_totalMin += service->minProcs;
962 /* below works even if a thread is running, since minDeficit would
963 * still have been decremented and later re-incremented.
965 rxi_minDeficit += service->minProcs;
966 MUTEX_EXIT(&rx_quota_mutex);
969 /* Turn on reaping of idle server connections */
970 rxi_ReapConnections(NULL, NULL, NULL, 0);
979 #ifdef AFS_PTHREAD_ENV
981 pid = afs_pointer_to_int(pthread_self());
982 #else /* AFS_PTHREAD_ENV */
984 LWP_CurrentProcess(&pid);
985 #endif /* AFS_PTHREAD_ENV */
987 sprintf(name, "srv_%d", ++nProcs);
989 (*registerProgram) (pid, name);
991 #endif /* AFS_NT40_ENV */
992 rx_ServerProc(NULL); /* Never returns */
994 #ifdef RX_ENABLE_TSFPQ
995 /* no use leaving packets around in this thread's local queue if
996 * it isn't getting donated to the server thread pool.
998 rxi_FlushLocalPacketsTSFPQ();
999 #endif /* RX_ENABLE_TSFPQ */
1003 /* Create a new client connection to the specified service, using the
1004 * specified security object to implement the security model for this
1006 struct rx_connection *
1007 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
1008 struct rx_securityClass *securityObject,
1009 int serviceSecurityIndex)
1012 struct rx_connection *conn;
1017 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
1018 "serviceSecurityIndex %d)\n",
1019 ntohl(shost), ntohs(sport), sservice, securityObject,
1020 serviceSecurityIndex));
1022 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
1023 * the case of kmem_alloc? */
1024 conn = rxi_AllocConnection();
1025 #ifdef RX_ENABLE_LOCKS
1026 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
1027 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
1028 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
1031 MUTEX_ENTER(&rx_connHashTable_lock);
1032 conn->type = RX_CLIENT_CONNECTION;
1033 conn->epoch = rx_epoch;
1034 conn->cid = rx_nextCid;
1036 conn->peer = rxi_FindPeer(shost, sport, 1);
1037 conn->serviceId = sservice;
1038 conn->securityObject = securityObject;
1039 conn->securityData = (void *) 0;
1040 conn->securityIndex = serviceSecurityIndex;
1041 rx_SetConnDeadTime(conn, rx_connDeadTime);
1042 rx_SetConnSecondsUntilNatPing(conn, 0);
1043 conn->ackRate = RX_FAST_ACK_RATE;
1044 conn->nSpecific = 0;
1045 conn->specific = NULL;
1046 conn->challengeEvent = NULL;
1047 conn->delayedAbortEvent = NULL;
1048 conn->abortCount = 0;
1050 for (i = 0; i < RX_MAXCALLS; i++) {
1051 conn->twind[i] = rx_initSendWindow;
1052 conn->rwind[i] = rx_initReceiveWindow;
1053 conn->lastBusy[i] = 0;
1056 RXS_NewConnection(securityObject, conn);
1058 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1060 conn->refCount++; /* no lock required since only this thread knows... */
1061 conn->next = rx_connHashTable[hashindex];
1062 rx_connHashTable[hashindex] = conn;
1063 if (rx_stats_active)
1064 rx_atomic_inc(&rx_stats.nClientConns);
1065 MUTEX_EXIT(&rx_connHashTable_lock);
1071 * Ensure a connection's timeout values are valid.
1073 * @param[in] conn The connection to check
1075 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1076 * unless idleDeadTime and/or hardDeadTime are not set
1080 rxi_CheckConnTimeouts(struct rx_connection *conn)
1082 /* a connection's timeouts must have the relationship
1083 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1084 * total loss of network to a peer may cause an idle timeout instead of a
1085 * dead timeout, simply because the idle timeout gets hit first. Also set
1086 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1087 /* this logic is slightly complicated by the fact that
1088 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1090 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1091 if (conn->idleDeadTime) {
1092 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1094 if (conn->hardDeadTime) {
1095 if (conn->idleDeadTime) {
1096 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1098 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1104 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1106 /* The idea is to set the dead time to a value that allows several
1107 * keepalives to be dropped without timing out the connection. */
1108 conn->secondsUntilDead = seconds;
1109 rxi_CheckConnTimeouts(conn);
1110 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1114 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1116 conn->hardDeadTime = seconds;
1117 rxi_CheckConnTimeouts(conn);
1121 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1123 conn->idleDeadTime = seconds;
1124 rxi_CheckConnTimeouts(conn);
1127 int rxi_lowPeerRefCount = 0;
1128 int rxi_lowConnRefCount = 0;
1131 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1132 * NOTE: must not be called with rx_connHashTable_lock held.
1135 rxi_CleanupConnection(struct rx_connection *conn)
1137 /* Notify the service exporter, if requested, that this connection
1138 * is being destroyed */
1139 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1140 (*conn->service->destroyConnProc) (conn);
1142 /* Notify the security module that this connection is being destroyed */
1143 RXS_DestroyConnection(conn->securityObject, conn);
1145 /* If this is the last connection using the rx_peer struct, set its
1146 * idle time to now. rxi_ReapConnections will reap it if it's still
1147 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1149 MUTEX_ENTER(&rx_peerHashTable_lock);
1150 if (conn->peer->refCount < 2) {
1151 conn->peer->idleWhen = clock_Sec();
1152 if (conn->peer->refCount < 1) {
1153 conn->peer->refCount = 1;
1154 if (rx_stats_active) {
1155 MUTEX_ENTER(&rx_stats_mutex);
1156 rxi_lowPeerRefCount++;
1157 MUTEX_EXIT(&rx_stats_mutex);
1161 conn->peer->refCount--;
1162 MUTEX_EXIT(&rx_peerHashTable_lock);
1164 if (rx_stats_active)
1166 if (conn->type == RX_SERVER_CONNECTION)
1167 rx_atomic_dec(&rx_stats.nServerConns);
1169 rx_atomic_dec(&rx_stats.nClientConns);
1172 if (conn->specific) {
1174 for (i = 0; i < conn->nSpecific; i++) {
1175 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1176 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1177 conn->specific[i] = NULL;
1179 free(conn->specific);
1181 conn->specific = NULL;
1182 conn->nSpecific = 0;
1183 #endif /* !KERNEL */
1185 MUTEX_DESTROY(&conn->conn_call_lock);
1186 MUTEX_DESTROY(&conn->conn_data_lock);
1187 CV_DESTROY(&conn->conn_call_cv);
1189 rxi_FreeConnection(conn);
1192 /* Destroy the specified connection */
1194 rxi_DestroyConnection(struct rx_connection *conn)
1196 MUTEX_ENTER(&rx_connHashTable_lock);
1197 rxi_DestroyConnectionNoLock(conn);
1198 /* conn should be at the head of the cleanup list */
1199 if (conn == rx_connCleanup_list) {
1200 rx_connCleanup_list = rx_connCleanup_list->next;
1201 MUTEX_EXIT(&rx_connHashTable_lock);
1202 rxi_CleanupConnection(conn);
1204 #ifdef RX_ENABLE_LOCKS
1206 MUTEX_EXIT(&rx_connHashTable_lock);
1208 #endif /* RX_ENABLE_LOCKS */
1212 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1214 struct rx_connection **conn_ptr;
1222 MUTEX_ENTER(&conn->conn_data_lock);
1223 MUTEX_ENTER(&rx_refcnt_mutex);
1224 if (conn->refCount > 0)
1227 #ifdef RX_REFCOUNT_CHECK
1228 osi_Assert(conn->refCount == 0);
1230 if (rx_stats_active) {
1231 MUTEX_ENTER(&rx_stats_mutex);
1232 rxi_lowConnRefCount++;
1233 MUTEX_EXIT(&rx_stats_mutex);
1237 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1238 /* Busy; wait till the last guy before proceeding */
1239 MUTEX_EXIT(&rx_refcnt_mutex);
1240 MUTEX_EXIT(&conn->conn_data_lock);
1245 /* If the client previously called rx_NewCall, but it is still
1246 * waiting, treat this as a running call, and wait to destroy the
1247 * connection later when the call completes. */
1248 if ((conn->type == RX_CLIENT_CONNECTION)
1249 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1250 conn->flags |= RX_CONN_DESTROY_ME;
1251 MUTEX_EXIT(&rx_refcnt_mutex);
1252 MUTEX_EXIT(&conn->conn_data_lock);
1256 MUTEX_EXIT(&rx_refcnt_mutex);
1257 MUTEX_EXIT(&conn->conn_data_lock);
1259 /* Check for extant references to this connection */
1260 MUTEX_ENTER(&conn->conn_call_lock);
1261 for (i = 0; i < RX_MAXCALLS; i++) {
1262 struct rx_call *call = conn->call[i];
1265 if (conn->type == RX_CLIENT_CONNECTION) {
1266 MUTEX_ENTER(&call->lock);
1267 if (call->delayedAckEvent) {
1268 /* Push the final acknowledgment out now--there
1269 * won't be a subsequent call to acknowledge the
1270 * last reply packets */
1271 rxi_CancelDelayedAckEvent(call);
1272 if (call->state == RX_STATE_PRECALL
1273 || call->state == RX_STATE_ACTIVE) {
1274 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1279 MUTEX_EXIT(&call->lock);
1283 MUTEX_EXIT(&conn->conn_call_lock);
1285 #ifdef RX_ENABLE_LOCKS
1287 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1288 MUTEX_EXIT(&conn->conn_data_lock);
1290 /* Someone is accessing a packet right now. */
1294 #endif /* RX_ENABLE_LOCKS */
1297 /* Don't destroy the connection if there are any call
1298 * structures still in use */
1299 MUTEX_ENTER(&conn->conn_data_lock);
1300 conn->flags |= RX_CONN_DESTROY_ME;
1301 MUTEX_EXIT(&conn->conn_data_lock);
1306 /* Remove from connection hash table before proceeding */
1308 &rx_connHashTable[CONN_HASH
1309 (peer->host, peer->port, conn->cid, conn->epoch,
1311 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1312 if (*conn_ptr == conn) {
1313 *conn_ptr = conn->next;
1317 /* if the conn that we are destroying was the last connection, then we
1318 * clear rxLastConn as well */
1319 if (rxLastConn == conn)
1322 /* Make sure the connection is completely reset before deleting it. */
1324 * Pending events hold a refcount, so we can't get here if they are
1326 osi_Assert(conn->challengeEvent == NULL);
1327 osi_Assert(conn->delayedAbortEvent == NULL);
1328 osi_Assert(conn->natKeepAliveEvent == NULL);
1329 osi_Assert(conn->checkReachEvent == NULL);
1331 /* Add the connection to the list of destroyed connections that
1332 * need to be cleaned up. This is necessary to avoid deadlocks
1333 * in the routines we call to inform others that this connection is
1334 * being destroyed. */
1335 conn->next = rx_connCleanup_list;
1336 rx_connCleanup_list = conn;
1339 /* Externally available version */
1341 rx_DestroyConnection(struct rx_connection *conn)
1346 rxi_DestroyConnection(conn);
1351 rx_GetConnection(struct rx_connection *conn)
1356 MUTEX_ENTER(&rx_refcnt_mutex);
1358 MUTEX_EXIT(&rx_refcnt_mutex);
1362 #ifdef RX_ENABLE_LOCKS
1363 /* Wait for the transmit queue to no longer be busy.
1364 * requires the call->lock to be held */
1366 rxi_WaitforTQBusy(struct rx_call *call) {
1367 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1368 call->flags |= RX_CALL_TQ_WAIT;
1370 MUTEX_ASSERT(&call->lock);
1371 CV_WAIT(&call->cv_tq, &call->lock);
1373 if (call->tqWaiters == 0) {
1374 call->flags &= ~RX_CALL_TQ_WAIT;
1381 rxi_WakeUpTransmitQueue(struct rx_call *call)
1383 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1384 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1385 call, call->tqWaiters, call->flags));
1386 #ifdef RX_ENABLE_LOCKS
1387 MUTEX_ASSERT(&call->lock);
1388 CV_BROADCAST(&call->cv_tq);
1389 #else /* RX_ENABLE_LOCKS */
1390 osi_rxWakeup(&call->tq);
1391 #endif /* RX_ENABLE_LOCKS */
1395 /* Start a new rx remote procedure call, on the specified connection.
1396 * If wait is set to 1, wait for a free call channel; otherwise return
1397 * 0. Maxtime gives the maximum number of seconds this call may take,
1398 * after rx_NewCall returns. After this time interval, a call to any
1399 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1400 * For fine grain locking, we hold the conn_call_lock in order to
1401 * to ensure that we don't get signalle after we found a call in an active
1402 * state and before we go to sleep.
1405 rx_NewCall(struct rx_connection *conn)
1407 int i, wait, ignoreBusy = 1;
1408 struct rx_call *call;
1409 struct clock queueTime;
1410 afs_uint32 leastBusy = 0;
1414 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1417 clock_GetTime(&queueTime);
1419 * Check if there are others waiting for a new call.
1420 * If so, let them go first to avoid starving them.
1421 * This is a fairly simple scheme, and might not be
1422 * a complete solution for large numbers of waiters.
1424 * makeCallWaiters keeps track of the number of
1425 * threads waiting to make calls and the
1426 * RX_CONN_MAKECALL_WAITING flag bit is used to
1427 * indicate that there are indeed calls waiting.
1428 * The flag is set when the waiter is incremented.
1429 * It is only cleared when makeCallWaiters is 0.
1430 * This prevents us from accidently destroying the
1431 * connection while it is potentially about to be used.
1433 MUTEX_ENTER(&conn->conn_call_lock);
1434 MUTEX_ENTER(&conn->conn_data_lock);
1435 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1436 conn->flags |= RX_CONN_MAKECALL_WAITING;
1437 conn->makeCallWaiters++;
1438 MUTEX_EXIT(&conn->conn_data_lock);
1440 #ifdef RX_ENABLE_LOCKS
1441 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1445 MUTEX_ENTER(&conn->conn_data_lock);
1446 conn->makeCallWaiters--;
1447 if (conn->makeCallWaiters == 0)
1448 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1451 /* We are now the active thread in rx_NewCall */
1452 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1453 MUTEX_EXIT(&conn->conn_data_lock);
1458 for (i = 0; i < RX_MAXCALLS; i++) {
1459 call = conn->call[i];
1461 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1462 /* we're not ignoring busy call slots; only look at the
1463 * call slot that is the "least" busy */
1467 if (call->state == RX_STATE_DALLY) {
1468 MUTEX_ENTER(&call->lock);
1469 if (call->state == RX_STATE_DALLY) {
1470 if (ignoreBusy && conn->lastBusy[i]) {
1471 /* if we're ignoring busy call slots, skip any ones that
1472 * have lastBusy set */
1473 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1474 leastBusy = conn->lastBusy[i];
1476 MUTEX_EXIT(&call->lock);
1481 * We are setting the state to RX_STATE_RESET to
1482 * ensure that no one else will attempt to use this
1483 * call once we drop the conn->conn_call_lock and
1484 * call->lock. We must drop the conn->conn_call_lock
1485 * before calling rxi_ResetCall because the process
1486 * of clearing the transmit queue can block for an
1487 * extended period of time. If we block while holding
1488 * the conn->conn_call_lock, then all rx_EndCall
1489 * processing will block as well. This has a detrimental
1490 * effect on overall system performance.
1492 call->state = RX_STATE_RESET;
1493 (*call->callNumber)++;
1494 MUTEX_EXIT(&conn->conn_call_lock);
1495 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1496 rxi_ResetCall(call, 0);
1497 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1501 * If we failed to be able to safely obtain the
1502 * conn->conn_call_lock we will have to drop the
1503 * call->lock to avoid a deadlock. When the call->lock
1504 * is released the state of the call can change. If it
1505 * is no longer RX_STATE_RESET then some other thread is
1508 MUTEX_EXIT(&call->lock);
1509 MUTEX_ENTER(&conn->conn_call_lock);
1510 MUTEX_ENTER(&call->lock);
1512 if (call->state == RX_STATE_RESET)
1516 * If we get here it means that after dropping
1517 * the conn->conn_call_lock and call->lock that
1518 * the call is no longer ours. If we can't find
1519 * a free call in the remaining slots we should
1520 * not go immediately to RX_CONN_MAKECALL_WAITING
1521 * because by dropping the conn->conn_call_lock
1522 * we have given up synchronization with rx_EndCall.
1523 * Instead, cycle through one more time to see if
1524 * we can find a call that can call our own.
1526 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1529 MUTEX_EXIT(&call->lock);
1532 if (ignoreBusy && conn->lastBusy[i]) {
1533 /* if we're ignoring busy call slots, skip any ones that
1534 * have lastBusy set */
1535 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1536 leastBusy = conn->lastBusy[i];
1541 /* rxi_NewCall returns with mutex locked */
1542 call = rxi_NewCall(conn, i);
1543 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1547 if (i < RX_MAXCALLS) {
1548 conn->lastBusy[i] = 0;
1553 if (leastBusy && ignoreBusy) {
1554 /* we didn't find a useable call slot, but we did see at least one
1555 * 'busy' slot; look again and only use a slot with the 'least
1561 MUTEX_ENTER(&conn->conn_data_lock);
1562 conn->flags |= RX_CONN_MAKECALL_WAITING;
1563 conn->makeCallWaiters++;
1564 MUTEX_EXIT(&conn->conn_data_lock);
1566 #ifdef RX_ENABLE_LOCKS
1567 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1571 MUTEX_ENTER(&conn->conn_data_lock);
1572 conn->makeCallWaiters--;
1573 if (conn->makeCallWaiters == 0)
1574 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1575 MUTEX_EXIT(&conn->conn_data_lock);
1577 /* Client is initially in send mode */
1578 call->state = RX_STATE_ACTIVE;
1579 call->error = conn->error;
1581 call->app.mode = RX_MODE_ERROR;
1583 call->app.mode = RX_MODE_SENDING;
1585 #ifdef AFS_RXERRQ_ENV
1586 /* remember how many network errors the peer has when we started, so if
1587 * more errors are encountered after the call starts, we know the other endpoint won't be
1588 * responding to us */
1589 call->neterr_gen = rx_atomic_read(&conn->peer->neterrs);
1592 /* remember start time for call in case we have hard dead time limit */
1593 call->queueTime = queueTime;
1594 clock_GetTime(&call->startTime);
1595 call->app.bytesSent = 0;
1596 call->app.bytesRcvd = 0;
1598 /* Turn on busy protocol. */
1599 rxi_KeepAliveOn(call);
1601 /* Attempt MTU discovery */
1602 rxi_GrowMTUOn(call);
1605 * We are no longer the active thread in rx_NewCall
1607 MUTEX_ENTER(&conn->conn_data_lock);
1608 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1609 MUTEX_EXIT(&conn->conn_data_lock);
1612 * Wake up anyone else who might be giving us a chance to
1613 * run (see code above that avoids resource starvation).
1615 #ifdef RX_ENABLE_LOCKS
1616 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1617 osi_Panic("rx_NewCall call about to be used without an empty tq");
1620 CV_BROADCAST(&conn->conn_call_cv);
1624 MUTEX_EXIT(&conn->conn_call_lock);
1625 MUTEX_EXIT(&call->lock);
1628 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1633 rxi_HasActiveCalls(struct rx_connection *aconn)
1636 struct rx_call *tcall;
1640 for (i = 0; i < RX_MAXCALLS; i++) {
1641 if ((tcall = aconn->call[i])) {
1642 if ((tcall->state == RX_STATE_ACTIVE)
1643 || (tcall->state == RX_STATE_PRECALL)) {
1654 rxi_GetCallNumberVector(struct rx_connection *aconn,
1655 afs_int32 * aint32s)
1658 struct rx_call *tcall;
1662 MUTEX_ENTER(&aconn->conn_call_lock);
1663 for (i = 0; i < RX_MAXCALLS; i++) {
1664 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1665 aint32s[i] = aconn->callNumber[i] + 1;
1667 aint32s[i] = aconn->callNumber[i];
1669 MUTEX_EXIT(&aconn->conn_call_lock);
1675 rxi_SetCallNumberVector(struct rx_connection *aconn,
1676 afs_int32 * aint32s)
1679 struct rx_call *tcall;
1683 MUTEX_ENTER(&aconn->conn_call_lock);
1684 for (i = 0; i < RX_MAXCALLS; i++) {
1685 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1686 aconn->callNumber[i] = aint32s[i] - 1;
1688 aconn->callNumber[i] = aint32s[i];
1690 MUTEX_EXIT(&aconn->conn_call_lock);
1695 /* Advertise a new service. A service is named locally by a UDP port
1696 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1699 char *serviceName; Name for identification purposes (e.g. the
1700 service name might be used for probing for
1703 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1704 char *serviceName, struct rx_securityClass **securityObjects,
1705 int nSecurityObjects,
1706 afs_int32(*serviceProc) (struct rx_call * acall))
1708 osi_socket socket = OSI_NULLSOCKET;
1709 struct rx_service *tservice;
1715 if (serviceId == 0) {
1717 "rx_NewService: service id for service %s is not non-zero.\n",
1724 "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",
1732 tservice = rxi_AllocService();
1735 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1737 for (i = 0; i < RX_MAX_SERVICES; i++) {
1738 struct rx_service *service = rx_services[i];
1740 if (port == service->servicePort && host == service->serviceHost) {
1741 if (service->serviceId == serviceId) {
1742 /* The identical service has already been
1743 * installed; if the caller was intending to
1744 * change the security classes used by this
1745 * service, he/she loses. */
1747 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1748 serviceName, serviceId, service->serviceName);
1750 rxi_FreeService(tservice);
1753 /* Different service, same port: re-use the socket
1754 * which is bound to the same port */
1755 socket = service->socket;
1758 if (socket == OSI_NULLSOCKET) {
1759 /* If we don't already have a socket (from another
1760 * service on same port) get a new one */
1761 socket = rxi_GetHostUDPSocket(host, port);
1762 if (socket == OSI_NULLSOCKET) {
1764 rxi_FreeService(tservice);
1769 service->socket = socket;
1770 service->serviceHost = host;
1771 service->servicePort = port;
1772 service->serviceId = serviceId;
1773 service->serviceName = serviceName;
1774 service->nSecurityObjects = nSecurityObjects;
1775 service->securityObjects = securityObjects;
1776 service->minProcs = 0;
1777 service->maxProcs = 1;
1778 service->idleDeadTime = 60;
1779 service->connDeadTime = rx_connDeadTime;
1780 service->executeRequestProc = serviceProc;
1781 service->checkReach = 0;
1782 service->nSpecific = 0;
1783 service->specific = NULL;
1784 rx_services[i] = service; /* not visible until now */
1790 rxi_FreeService(tservice);
1791 (osi_Msg "rx_NewService: cannot support > %d services\n",
1796 /* Set configuration options for all of a service's security objects */
1799 rx_SetSecurityConfiguration(struct rx_service *service,
1800 rx_securityConfigVariables type,
1804 for (i = 0; i<service->nSecurityObjects; i++) {
1805 if (service->securityObjects[i]) {
1806 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1814 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1815 struct rx_securityClass **securityObjects, int nSecurityObjects,
1816 afs_int32(*serviceProc) (struct rx_call * acall))
1818 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1821 /* Generic request processing loop. This routine should be called
1822 * by the implementation dependent rx_ServerProc. If socketp is
1823 * non-null, it will be set to the file descriptor that this thread
1824 * is now listening on. If socketp is null, this routine will never
1827 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1829 struct rx_call *call;
1831 struct rx_service *tservice = NULL;
1838 call = rx_GetCall(threadID, tservice, socketp);
1839 if (socketp && *socketp != OSI_NULLSOCKET) {
1840 /* We are now a listener thread */
1846 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1847 #ifdef RX_ENABLE_LOCKS
1849 #endif /* RX_ENABLE_LOCKS */
1850 afs_termState = AFSOP_STOP_AFS;
1851 afs_osi_Wakeup(&afs_termState);
1852 #ifdef RX_ENABLE_LOCKS
1854 #endif /* RX_ENABLE_LOCKS */
1859 /* if server is restarting( typically smooth shutdown) then do not
1860 * allow any new calls.
1863 if (rx_tranquil && (call != NULL)) {
1867 MUTEX_ENTER(&call->lock);
1869 rxi_CallError(call, RX_RESTARTING);
1870 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1872 MUTEX_EXIT(&call->lock);
1877 tservice = call->conn->service;
1879 if (tservice->beforeProc)
1880 (*tservice->beforeProc) (call);
1882 code = tservice->executeRequestProc(call);
1884 if (tservice->afterProc)
1885 (*tservice->afterProc) (call, code);
1887 rx_EndCall(call, code);
1889 if (tservice->postProc)
1890 (*tservice->postProc) (code);
1892 if (rx_stats_active) {
1893 MUTEX_ENTER(&rx_stats_mutex);
1895 MUTEX_EXIT(&rx_stats_mutex);
1902 rx_WakeupServerProcs(void)
1904 struct rx_serverQueueEntry *np, *tqp;
1905 struct opr_queue *cursor;
1909 MUTEX_ENTER(&rx_serverPool_lock);
1911 #ifdef RX_ENABLE_LOCKS
1912 if (rx_waitForPacket)
1913 CV_BROADCAST(&rx_waitForPacket->cv);
1914 #else /* RX_ENABLE_LOCKS */
1915 if (rx_waitForPacket)
1916 osi_rxWakeup(rx_waitForPacket);
1917 #endif /* RX_ENABLE_LOCKS */
1918 MUTEX_ENTER(&freeSQEList_lock);
1919 for (np = rx_FreeSQEList; np; np = tqp) {
1920 tqp = *(struct rx_serverQueueEntry **)np;
1921 #ifdef RX_ENABLE_LOCKS
1922 CV_BROADCAST(&np->cv);
1923 #else /* RX_ENABLE_LOCKS */
1925 #endif /* RX_ENABLE_LOCKS */
1927 MUTEX_EXIT(&freeSQEList_lock);
1928 for (opr_queue_Scan(&rx_idleServerQueue, cursor)) {
1929 np = opr_queue_Entry(cursor, struct rx_serverQueueEntry, entry);
1930 #ifdef RX_ENABLE_LOCKS
1931 CV_BROADCAST(&np->cv);
1932 #else /* RX_ENABLE_LOCKS */
1934 #endif /* RX_ENABLE_LOCKS */
1936 MUTEX_EXIT(&rx_serverPool_lock);
1941 * One thing that seems to happen is that all the server threads get
1942 * tied up on some empty or slow call, and then a whole bunch of calls
1943 * arrive at once, using up the packet pool, so now there are more
1944 * empty calls. The most critical resources here are server threads
1945 * and the free packet pool. The "doreclaim" code seems to help in
1946 * general. I think that eventually we arrive in this state: there
1947 * are lots of pending calls which do have all their packets present,
1948 * so they won't be reclaimed, are multi-packet calls, so they won't
1949 * be scheduled until later, and thus are tying up most of the free
1950 * packet pool for a very long time.
1952 * 1. schedule multi-packet calls if all the packets are present.
1953 * Probably CPU-bound operation, useful to return packets to pool.
1954 * Do what if there is a full window, but the last packet isn't here?
1955 * 3. preserve one thread which *only* runs "best" calls, otherwise
1956 * it sleeps and waits for that type of call.
1957 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1958 * the current dataquota business is badly broken. The quota isn't adjusted
1959 * to reflect how many packets are presently queued for a running call.
1960 * So, when we schedule a queued call with a full window of packets queued
1961 * up for it, that *should* free up a window full of packets for other 2d-class
1962 * calls to be able to use from the packet pool. But it doesn't.
1964 * NB. Most of the time, this code doesn't run -- since idle server threads
1965 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1966 * as a new call arrives.
1968 /* Sleep until a call arrives. Returns a pointer to the call, ready
1969 * for an rx_Read. */
1970 #ifdef RX_ENABLE_LOCKS
1972 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1974 struct rx_serverQueueEntry *sq;
1975 struct rx_call *call = (struct rx_call *)0;
1976 struct rx_service *service = NULL;
1978 MUTEX_ENTER(&freeSQEList_lock);
1980 if ((sq = rx_FreeSQEList)) {
1981 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1982 MUTEX_EXIT(&freeSQEList_lock);
1983 } else { /* otherwise allocate a new one and return that */
1984 MUTEX_EXIT(&freeSQEList_lock);
1985 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1986 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1987 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1990 MUTEX_ENTER(&rx_serverPool_lock);
1991 if (cur_service != NULL) {
1992 ReturnToServerPool(cur_service);
1995 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
1996 struct rx_call *tcall, *choice2 = NULL;
1997 struct opr_queue *cursor;
1999 /* Scan for eligible incoming calls. A call is not eligible
2000 * if the maximum number of calls for its service type are
2001 * already executing */
2002 /* One thread will process calls FCFS (to prevent starvation),
2003 * while the other threads may run ahead looking for calls which
2004 * have all their input data available immediately. This helps
2005 * keep threads from blocking, waiting for data from the client. */
2006 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2007 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2009 service = tcall->conn->service;
2010 if (!QuotaOK(service)) {
2013 MUTEX_ENTER(&rx_pthread_mutex);
2014 if (tno == rxi_fcfs_thread_num
2015 || opr_queue_IsEnd(&rx_incomingCallQueue, cursor)) {
2016 MUTEX_EXIT(&rx_pthread_mutex);
2017 /* If we're the fcfs thread , then we'll just use
2018 * this call. If we haven't been able to find an optimal
2019 * choice, and we're at the end of the list, then use a
2020 * 2d choice if one has been identified. Otherwise... */
2021 call = (choice2 ? choice2 : tcall);
2022 service = call->conn->service;
2024 MUTEX_EXIT(&rx_pthread_mutex);
2025 if (!opr_queue_IsEmpty(&tcall->rq)) {
2026 struct rx_packet *rp;
2027 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2029 if (rp->header.seq == 1) {
2031 || (rp->header.flags & RX_LAST_PACKET)) {
2033 } else if (rxi_2dchoice && !choice2
2034 && !(tcall->flags & RX_CALL_CLEARED)
2035 && (tcall->rprev > rxi_HardAckRate)) {
2045 ReturnToServerPool(service);
2051 opr_queue_Remove(&call->entry);
2052 MUTEX_EXIT(&rx_serverPool_lock);
2053 MUTEX_ENTER(&call->lock);
2055 if (call->flags & RX_CALL_WAIT_PROC) {
2056 call->flags &= ~RX_CALL_WAIT_PROC;
2057 rx_atomic_dec(&rx_nWaiting);
2060 if (call->state != RX_STATE_PRECALL || call->error) {
2061 MUTEX_EXIT(&call->lock);
2062 MUTEX_ENTER(&rx_serverPool_lock);
2063 ReturnToServerPool(service);
2068 if (opr_queue_IsEmpty(&call->rq)
2069 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1)
2070 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2072 CLEAR_CALL_QUEUE_LOCK(call);
2075 /* If there are no eligible incoming calls, add this process
2076 * to the idle server queue, to wait for one */
2080 *socketp = OSI_NULLSOCKET;
2082 sq->socketp = socketp;
2083 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2084 #ifndef AFS_AIX41_ENV
2085 rx_waitForPacket = sq;
2086 #endif /* AFS_AIX41_ENV */
2088 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2090 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2091 MUTEX_EXIT(&rx_serverPool_lock);
2092 return (struct rx_call *)0;
2095 } while (!(call = sq->newcall)
2096 && !(socketp && *socketp != OSI_NULLSOCKET));
2097 MUTEX_EXIT(&rx_serverPool_lock);
2099 MUTEX_ENTER(&call->lock);
2105 MUTEX_ENTER(&freeSQEList_lock);
2106 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2107 rx_FreeSQEList = sq;
2108 MUTEX_EXIT(&freeSQEList_lock);
2111 clock_GetTime(&call->startTime);
2112 call->state = RX_STATE_ACTIVE;
2113 call->app.mode = RX_MODE_RECEIVING;
2114 #ifdef RX_KERNEL_TRACE
2115 if (ICL_SETACTIVE(afs_iclSetp)) {
2116 int glockOwner = ISAFS_GLOCK();
2119 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2120 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2127 rxi_calltrace(RX_CALL_START, call);
2128 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2129 call->conn->service->servicePort, call->conn->service->serviceId,
2132 MUTEX_EXIT(&call->lock);
2133 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2135 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2140 #else /* RX_ENABLE_LOCKS */
2142 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2144 struct rx_serverQueueEntry *sq;
2145 struct rx_call *call = (struct rx_call *)0, *choice2;
2146 struct rx_service *service = NULL;
2150 MUTEX_ENTER(&freeSQEList_lock);
2152 if ((sq = rx_FreeSQEList)) {
2153 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2154 MUTEX_EXIT(&freeSQEList_lock);
2155 } else { /* otherwise allocate a new one and return that */
2156 MUTEX_EXIT(&freeSQEList_lock);
2157 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2158 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2159 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2161 MUTEX_ENTER(&sq->lock);
2163 if (cur_service != NULL) {
2164 cur_service->nRequestsRunning--;
2165 MUTEX_ENTER(&rx_quota_mutex);
2166 if (cur_service->nRequestsRunning < cur_service->minProcs)
2169 MUTEX_EXIT(&rx_quota_mutex);
2171 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2172 struct rx_call *tcall;
2173 struct opr_queue *cursor;
2174 /* Scan for eligible incoming calls. A call is not eligible
2175 * if the maximum number of calls for its service type are
2176 * already executing */
2177 /* One thread will process calls FCFS (to prevent starvation),
2178 * while the other threads may run ahead looking for calls which
2179 * have all their input data available immediately. This helps
2180 * keep threads from blocking, waiting for data from the client. */
2181 choice2 = (struct rx_call *)0;
2182 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2183 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2184 service = tcall->conn->service;
2185 if (QuotaOK(service)) {
2186 MUTEX_ENTER(&rx_pthread_mutex);
2187 /* XXX - If tcall->entry.next is NULL, then we're no longer
2188 * on a queue at all. This shouldn't happen. */
2189 if (tno == rxi_fcfs_thread_num || !tcall->entry.next) {
2190 MUTEX_EXIT(&rx_pthread_mutex);
2191 /* If we're the fcfs thread, then we'll just use
2192 * this call. If we haven't been able to find an optimal
2193 * choice, and we're at the end of the list, then use a
2194 * 2d choice if one has been identified. Otherwise... */
2195 call = (choice2 ? choice2 : tcall);
2196 service = call->conn->service;
2198 MUTEX_EXIT(&rx_pthread_mutex);
2199 if (!opr_queue_IsEmpty(&tcall->rq)) {
2200 struct rx_packet *rp;
2201 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2203 if (rp->header.seq == 1
2205 || (rp->header.flags & RX_LAST_PACKET))) {
2207 } else if (rxi_2dchoice && !choice2
2208 && !(tcall->flags & RX_CALL_CLEARED)
2209 && (tcall->rprev > rxi_HardAckRate)) {
2222 opr_queue_Remove(&call->entry);
2223 /* we can't schedule a call if there's no data!!! */
2224 /* send an ack if there's no data, if we're missing the
2225 * first packet, or we're missing something between first
2226 * and last -- there's a "hole" in the incoming data. */
2227 if (opr_queue_IsEmpty(&call->rq)
2228 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1
2229 || call->rprev != opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq)
2230 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2232 call->flags &= (~RX_CALL_WAIT_PROC);
2233 service->nRequestsRunning++;
2234 /* just started call in minProcs pool, need fewer to maintain
2236 MUTEX_ENTER(&rx_quota_mutex);
2237 if (service->nRequestsRunning <= service->minProcs)
2240 MUTEX_EXIT(&rx_quota_mutex);
2241 rx_atomic_dec(&rx_nWaiting);
2242 /* MUTEX_EXIT(&call->lock); */
2244 /* If there are no eligible incoming calls, add this process
2245 * to the idle server queue, to wait for one */
2248 *socketp = OSI_NULLSOCKET;
2250 sq->socketp = socketp;
2251 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2255 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2257 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2258 return (struct rx_call *)0;
2261 } while (!(call = sq->newcall)
2262 && !(socketp && *socketp != OSI_NULLSOCKET));
2264 MUTEX_EXIT(&sq->lock);
2266 MUTEX_ENTER(&freeSQEList_lock);
2267 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2268 rx_FreeSQEList = sq;
2269 MUTEX_EXIT(&freeSQEList_lock);
2272 clock_GetTime(&call->startTime);
2273 call->state = RX_STATE_ACTIVE;
2274 call->app.mode = RX_MODE_RECEIVING;
2275 #ifdef RX_KERNEL_TRACE
2276 if (ICL_SETACTIVE(afs_iclSetp)) {
2277 int glockOwner = ISAFS_GLOCK();
2280 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2281 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2288 rxi_calltrace(RX_CALL_START, call);
2289 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2290 call->conn->service->servicePort, call->conn->service->serviceId,
2293 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2300 #endif /* RX_ENABLE_LOCKS */
2304 /* Establish a procedure to be called when a packet arrives for a
2305 * call. This routine will be called at most once after each call,
2306 * and will also be called if there is an error condition on the or
2307 * the call is complete. Used by multi rx to build a selection
2308 * function which determines which of several calls is likely to be a
2309 * good one to read from.
2310 * NOTE: the way this is currently implemented it is probably only a
2311 * good idea to (1) use it immediately after a newcall (clients only)
2312 * and (2) only use it once. Other uses currently void your warranty
2315 rx_SetArrivalProc(struct rx_call *call,
2316 void (*proc) (struct rx_call * call,
2319 void * handle, int arg)
2321 call->arrivalProc = proc;
2322 call->arrivalProcHandle = handle;
2323 call->arrivalProcArg = arg;
2326 /* Call is finished (possibly prematurely). Return rc to the peer, if
2327 * appropriate, and return the final error code from the conversation
2331 rx_EndCall(struct rx_call *call, afs_int32 rc)
2333 struct rx_connection *conn = call->conn;
2337 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2338 call, rc, call->error, call->abortCode));
2341 MUTEX_ENTER(&call->lock);
2343 if (rc == 0 && call->error == 0) {
2344 call->abortCode = 0;
2345 call->abortCount = 0;
2348 call->arrivalProc = (void (*)())0;
2349 if (rc && call->error == 0) {
2350 rxi_CallError(call, rc);
2351 call->app.mode = RX_MODE_ERROR;
2352 /* Send an abort message to the peer if this error code has
2353 * only just been set. If it was set previously, assume the
2354 * peer has already been sent the error code or will request it
2356 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2358 if (conn->type == RX_SERVER_CONNECTION) {
2359 /* Make sure reply or at least dummy reply is sent */
2360 if (call->app.mode == RX_MODE_RECEIVING) {
2361 MUTEX_EXIT(&call->lock);
2362 rxi_WriteProc(call, 0, 0);
2363 MUTEX_ENTER(&call->lock);
2365 if (call->app.mode == RX_MODE_SENDING) {
2366 rxi_FlushWriteLocked(call);
2368 rxi_calltrace(RX_CALL_END, call);
2369 /* Call goes to hold state until reply packets are acknowledged */
2370 if (call->tfirst + call->nSoftAcked < call->tnext) {
2371 call->state = RX_STATE_HOLD;
2373 call->state = RX_STATE_DALLY;
2374 rxi_ClearTransmitQueue(call, 0);
2375 rxi_rto_cancel(call);
2376 rxi_CancelKeepAliveEvent(call);
2378 } else { /* Client connection */
2380 /* Make sure server receives input packets, in the case where
2381 * no reply arguments are expected */
2383 if ((call->app.mode == RX_MODE_SENDING)
2384 || (call->app.mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2385 MUTEX_EXIT(&call->lock);
2386 (void)rxi_ReadProc(call, &dummy, 1);
2387 MUTEX_ENTER(&call->lock);
2390 /* If we had an outstanding delayed ack, be nice to the server
2391 * and force-send it now.
2393 if (call->delayedAckEvent) {
2394 rxi_CancelDelayedAckEvent(call);
2395 rxi_SendDelayedAck(NULL, call, NULL, 0);
2398 /* We need to release the call lock since it's lower than the
2399 * conn_call_lock and we don't want to hold the conn_call_lock
2400 * over the rx_ReadProc call. The conn_call_lock needs to be held
2401 * here for the case where rx_NewCall is perusing the calls on
2402 * the connection structure. We don't want to signal until
2403 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2404 * have checked this call, found it active and by the time it
2405 * goes to sleep, will have missed the signal.
2407 MUTEX_EXIT(&call->lock);
2408 MUTEX_ENTER(&conn->conn_call_lock);
2409 MUTEX_ENTER(&call->lock);
2412 /* While there are some circumstances where a call with an error is
2413 * obviously not on a "busy" channel, be conservative (clearing
2414 * lastBusy is just best-effort to possibly speed up rx_NewCall).
2415 * The call channel is definitely not busy if we just successfully
2416 * completed a call on it. */
2417 conn->lastBusy[call->channel] = 0;
2419 } else if (call->error == RX_CALL_TIMEOUT) {
2420 /* The call is still probably running on the server side, so try to
2421 * avoid this call channel in the future. */
2422 conn->lastBusy[call->channel] = clock_Sec();
2425 MUTEX_ENTER(&conn->conn_data_lock);
2426 conn->flags |= RX_CONN_BUSY;
2427 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2428 MUTEX_EXIT(&conn->conn_data_lock);
2429 #ifdef RX_ENABLE_LOCKS
2430 CV_BROADCAST(&conn->conn_call_cv);
2435 #ifdef RX_ENABLE_LOCKS
2437 MUTEX_EXIT(&conn->conn_data_lock);
2439 #endif /* RX_ENABLE_LOCKS */
2440 call->state = RX_STATE_DALLY;
2442 error = call->error;
2444 /* currentPacket, nLeft, and NFree must be zeroed here, because
2445 * ResetCall cannot: ResetCall may be called at splnet(), in the
2446 * kernel version, and may interrupt the macros rx_Read or
2447 * rx_Write, which run at normal priority for efficiency. */
2448 if (call->app.currentPacket) {
2449 #ifdef RX_TRACK_PACKETS
2450 call->app.currentPacket->flags &= ~RX_PKTFLAG_CP;
2452 rxi_FreePacket(call->app.currentPacket);
2453 call->app.currentPacket = (struct rx_packet *)0;
2456 call->app.nLeft = call->app.nFree = call->app.curlen = 0;
2458 /* Free any packets from the last call to ReadvProc/WritevProc */
2459 #ifdef RXDEBUG_PACKET
2461 #endif /* RXDEBUG_PACKET */
2462 rxi_FreePackets(0, &call->app.iovq);
2463 MUTEX_EXIT(&call->lock);
2465 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2466 if (conn->type == RX_CLIENT_CONNECTION) {
2467 MUTEX_ENTER(&conn->conn_data_lock);
2468 conn->flags &= ~RX_CONN_BUSY;
2469 MUTEX_EXIT(&conn->conn_data_lock);
2470 MUTEX_EXIT(&conn->conn_call_lock);
2474 * Map errors to the local host's errno.h format.
2476 error = ntoh_syserr_conv(error);
2478 /* If the caller said the call failed with some error, we had better
2479 * return an error code. */
2480 osi_Assert(!rc || error);
2484 #if !defined(KERNEL)
2486 /* Call this routine when shutting down a server or client (especially
2487 * clients). This will allow Rx to gracefully garbage collect server
2488 * connections, and reduce the number of retries that a server might
2489 * make to a dead client.
2490 * This is not quite right, since some calls may still be ongoing and
2491 * we can't lock them to destroy them. */
2495 struct rx_connection **conn_ptr, **conn_end;
2498 if (rx_atomic_test_and_set_bit(&rxinit_status, 0))
2499 return; /* Already shutdown. */
2501 rxi_DeleteCachedConnections();
2502 if (rx_connHashTable) {
2503 MUTEX_ENTER(&rx_connHashTable_lock);
2504 for (conn_ptr = &rx_connHashTable[0], conn_end =
2505 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2507 struct rx_connection *conn, *next;
2508 for (conn = *conn_ptr; conn; conn = next) {
2510 if (conn->type == RX_CLIENT_CONNECTION) {
2511 rx_GetConnection(conn);
2512 #ifdef RX_ENABLE_LOCKS
2513 rxi_DestroyConnectionNoLock(conn);
2514 #else /* RX_ENABLE_LOCKS */
2515 rxi_DestroyConnection(conn);
2516 #endif /* RX_ENABLE_LOCKS */
2520 #ifdef RX_ENABLE_LOCKS
2521 while (rx_connCleanup_list) {
2522 struct rx_connection *conn;
2523 conn = rx_connCleanup_list;
2524 rx_connCleanup_list = rx_connCleanup_list->next;
2525 MUTEX_EXIT(&rx_connHashTable_lock);
2526 rxi_CleanupConnection(conn);
2527 MUTEX_ENTER(&rx_connHashTable_lock);
2529 MUTEX_EXIT(&rx_connHashTable_lock);
2530 #endif /* RX_ENABLE_LOCKS */
2535 afs_winsockCleanup();
2541 /* if we wakeup packet waiter too often, can get in loop with two
2542 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2544 rxi_PacketsUnWait(void)
2546 if (!rx_waitingForPackets) {
2550 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2551 return; /* still over quota */
2554 rx_waitingForPackets = 0;
2555 #ifdef RX_ENABLE_LOCKS
2556 CV_BROADCAST(&rx_waitingForPackets_cv);
2558 osi_rxWakeup(&rx_waitingForPackets);
2564 /* ------------------Internal interfaces------------------------- */
2566 /* Return this process's service structure for the
2567 * specified socket and service */
2568 static struct rx_service *
2569 rxi_FindService(osi_socket socket, u_short serviceId)
2571 struct rx_service **sp;
2572 for (sp = &rx_services[0]; *sp; sp++) {
2573 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2579 #ifdef RXDEBUG_PACKET
2580 #ifdef KDUMP_RX_LOCK
2581 static struct rx_call_rx_lock *rx_allCallsp = 0;
2583 static struct rx_call *rx_allCallsp = 0;
2585 #endif /* RXDEBUG_PACKET */
2587 /* Allocate a call structure, for the indicated channel of the
2588 * supplied connection. The mode and state of the call must be set by
2589 * the caller. Returns the call with mutex locked. */
2590 static struct rx_call *
2591 rxi_NewCall(struct rx_connection *conn, int channel)
2593 struct rx_call *call;
2594 #ifdef RX_ENABLE_LOCKS
2595 struct rx_call *cp; /* Call pointer temp */
2596 struct opr_queue *cursor;
2599 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2601 /* Grab an existing call structure, or allocate a new one.
2602 * Existing call structures are assumed to have been left reset by
2604 MUTEX_ENTER(&rx_freeCallQueue_lock);
2606 #ifdef RX_ENABLE_LOCKS
2608 * EXCEPT that the TQ might not yet be cleared out.
2609 * Skip over those with in-use TQs.
2612 for (opr_queue_Scan(&rx_freeCallQueue, cursor)) {
2613 cp = opr_queue_Entry(cursor, struct rx_call, entry);
2614 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2620 #else /* RX_ENABLE_LOCKS */
2621 if (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
2622 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
2623 #endif /* RX_ENABLE_LOCKS */
2624 opr_queue_Remove(&call->entry);
2625 if (rx_stats_active)
2626 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2627 MUTEX_EXIT(&rx_freeCallQueue_lock);
2628 MUTEX_ENTER(&call->lock);
2629 CLEAR_CALL_QUEUE_LOCK(call);
2630 #ifdef RX_ENABLE_LOCKS
2631 /* Now, if TQ wasn't cleared earlier, do it now. */
2632 rxi_WaitforTQBusy(call);
2633 if (call->flags & RX_CALL_TQ_CLEARME) {
2634 rxi_ClearTransmitQueue(call, 1);
2635 /*queue_Init(&call->tq);*/
2637 #endif /* RX_ENABLE_LOCKS */
2638 /* Bind the call to its connection structure */
2640 rxi_ResetCall(call, 1);
2643 call = rxi_Alloc(sizeof(struct rx_call));
2644 #ifdef RXDEBUG_PACKET
2645 call->allNextp = rx_allCallsp;
2646 rx_allCallsp = call;
2648 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2649 #else /* RXDEBUG_PACKET */
2650 rx_atomic_inc(&rx_stats.nCallStructs);
2651 #endif /* RXDEBUG_PACKET */
2653 MUTEX_EXIT(&rx_freeCallQueue_lock);
2654 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2655 MUTEX_ENTER(&call->lock);
2656 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2657 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2658 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2660 /* Initialize once-only items */
2661 opr_queue_Init(&call->tq);
2662 opr_queue_Init(&call->rq);
2663 opr_queue_Init(&call->app.iovq);
2664 #ifdef RXDEBUG_PACKET
2665 call->rqc = call->tqc = call->iovqc = 0;
2666 #endif /* RXDEBUG_PACKET */
2667 /* Bind the call to its connection structure (prereq for reset) */
2669 rxi_ResetCall(call, 1);
2671 call->channel = channel;
2672 call->callNumber = &conn->callNumber[channel];
2673 call->rwind = conn->rwind[channel];
2674 call->twind = conn->twind[channel];
2675 /* Note that the next expected call number is retained (in
2676 * conn->callNumber[i]), even if we reallocate the call structure
2678 conn->call[channel] = call;
2679 /* if the channel's never been used (== 0), we should start at 1, otherwise
2680 * the call number is valid from the last time this channel was used */
2681 if (*call->callNumber == 0)
2682 *call->callNumber = 1;
2687 /* A call has been inactive long enough that so we can throw away
2688 * state, including the call structure, which is placed on the call
2691 * call->lock amd rx_refcnt_mutex are held upon entry.
2692 * haveCTLock is set when called from rxi_ReapConnections.
2694 * return 1 if the call is freed, 0 if not.
2697 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2699 int channel = call->channel;
2700 struct rx_connection *conn = call->conn;
2701 u_char state = call->state;
2704 * We are setting the state to RX_STATE_RESET to
2705 * ensure that no one else will attempt to use this
2706 * call once we drop the refcnt lock. We must drop
2707 * the refcnt lock before calling rxi_ResetCall
2708 * because it cannot be held across acquiring the
2709 * freepktQ lock. NewCall does the same.
2711 call->state = RX_STATE_RESET;
2712 MUTEX_EXIT(&rx_refcnt_mutex);
2713 rxi_ResetCall(call, 0);
2715 if (MUTEX_TRYENTER(&conn->conn_call_lock))
2717 if (state == RX_STATE_DALLY || state == RX_STATE_HOLD)
2718 (*call->callNumber)++;
2720 if (call->conn->call[channel] == call)
2721 call->conn->call[channel] = 0;
2722 MUTEX_EXIT(&conn->conn_call_lock);
2725 * We couldn't obtain the conn_call_lock so we can't
2726 * disconnect the call from the connection. Set the
2727 * call state to dally so that the call can be reused.
2729 MUTEX_ENTER(&rx_refcnt_mutex);
2730 call->state = RX_STATE_DALLY;
2734 MUTEX_ENTER(&rx_freeCallQueue_lock);
2735 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2736 #ifdef RX_ENABLE_LOCKS
2737 /* A call may be free even though its transmit queue is still in use.
2738 * Since we search the call list from head to tail, put busy calls at
2739 * the head of the list, and idle calls at the tail.
2741 if (call->flags & RX_CALL_TQ_BUSY)
2742 opr_queue_Prepend(&rx_freeCallQueue, &call->entry);
2744 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2745 #else /* RX_ENABLE_LOCKS */
2746 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2747 #endif /* RX_ENABLE_LOCKS */
2748 if (rx_stats_active)
2749 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2750 MUTEX_EXIT(&rx_freeCallQueue_lock);
2752 /* Destroy the connection if it was previously slated for
2753 * destruction, i.e. the Rx client code previously called
2754 * rx_DestroyConnection (client connections), or
2755 * rxi_ReapConnections called the same routine (server
2756 * connections). Only do this, however, if there are no
2757 * outstanding calls. Note that for fine grain locking, there appears
2758 * to be a deadlock in that rxi_FreeCall has a call locked and
2759 * DestroyConnectionNoLock locks each call in the conn. But note a
2760 * few lines up where we have removed this call from the conn.
2761 * If someone else destroys a connection, they either have no
2762 * call lock held or are going through this section of code.
2764 MUTEX_ENTER(&conn->conn_data_lock);
2765 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2766 rx_GetConnection(conn);
2767 MUTEX_EXIT(&conn->conn_data_lock);
2768 #ifdef RX_ENABLE_LOCKS
2770 rxi_DestroyConnectionNoLock(conn);
2772 rxi_DestroyConnection(conn);
2773 #else /* RX_ENABLE_LOCKS */
2774 rxi_DestroyConnection(conn);
2775 #endif /* RX_ENABLE_LOCKS */
2777 MUTEX_EXIT(&conn->conn_data_lock);
2779 MUTEX_ENTER(&rx_refcnt_mutex);
2783 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2784 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2787 rxi_Alloc(size_t size)
2791 if (rx_stats_active) {
2792 rx_atomic_add(&rxi_Allocsize, (int) size);
2793 rx_atomic_inc(&rxi_Alloccnt);
2797 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2798 afs_osi_Alloc_NoSleep(size);
2803 osi_Panic("rxi_Alloc error");
2809 rxi_Free(void *addr, size_t size)
2811 if (rx_stats_active) {
2812 rx_atomic_sub(&rxi_Allocsize, (int) size);
2813 rx_atomic_dec(&rxi_Alloccnt);
2815 osi_Free(addr, size);
2819 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2821 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2822 struct rx_peer *next = NULL;
2826 MUTEX_ENTER(&rx_peerHashTable_lock);
2828 peer_ptr = &rx_peerHashTable[0];
2829 peer_end = &rx_peerHashTable[rx_hashTableSize];
2832 for ( ; peer_ptr < peer_end; peer_ptr++) {
2835 for ( ; peer; peer = next) {
2837 if (host == peer->host)
2842 hashIndex = PEER_HASH(host, port);
2843 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2844 if ((peer->host == host) && (peer->port == port))
2849 MUTEX_ENTER(&rx_peerHashTable_lock);
2854 MUTEX_EXIT(&rx_peerHashTable_lock);
2856 MUTEX_ENTER(&peer->peer_lock);
2857 /* We don't handle dropping below min, so don't */
2858 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2859 peer->ifMTU=MIN(mtu, peer->ifMTU);
2860 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2861 /* if we tweaked this down, need to tune our peer MTU too */
2862 peer->MTU = MIN(peer->MTU, peer->natMTU);
2863 /* if we discovered a sub-1500 mtu, degrade */
2864 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2865 peer->maxDgramPackets = 1;
2866 /* We no longer have valid peer packet information */
2867 if (peer->maxPacketSize + RX_HEADER_SIZE > peer->ifMTU)
2868 peer->maxPacketSize = 0;
2869 MUTEX_EXIT(&peer->peer_lock);
2871 MUTEX_ENTER(&rx_peerHashTable_lock);
2873 if (host && !port) {
2875 /* pick up where we left off */
2879 MUTEX_EXIT(&rx_peerHashTable_lock);
2882 #ifdef AFS_RXERRQ_ENV
2884 rxi_SetPeerDead(struct sock_extended_err *err, afs_uint32 host, afs_uint16 port)
2886 int hashIndex = PEER_HASH(host, port);
2887 struct rx_peer *peer;
2889 MUTEX_ENTER(&rx_peerHashTable_lock);
2891 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2892 if (peer->host == host && peer->port == port) {
2898 MUTEX_EXIT(&rx_peerHashTable_lock);
2901 rx_atomic_inc(&peer->neterrs);
2902 MUTEX_ENTER(&peer->peer_lock);
2903 peer->last_err_origin = RX_NETWORK_ERROR_ORIGIN_ICMP;
2904 peer->last_err_type = err->ee_type;
2905 peer->last_err_code = err->ee_code;
2906 MUTEX_EXIT(&peer->peer_lock);
2908 MUTEX_ENTER(&rx_peerHashTable_lock);
2910 MUTEX_EXIT(&rx_peerHashTable_lock);
2915 rxi_ProcessNetError(struct sock_extended_err *err, afs_uint32 addr, afs_uint16 port)
2917 # ifdef AFS_ADAPT_PMTU
2918 if (err->ee_errno == EMSGSIZE && err->ee_info >= 68) {
2919 rxi_SetPeerMtu(NULL, addr, port, err->ee_info - RX_IPUDP_SIZE);
2923 if (err->ee_origin == SO_EE_ORIGIN_ICMP && err->ee_type == ICMP_DEST_UNREACH) {
2924 switch (err->ee_code) {
2925 case ICMP_NET_UNREACH:
2926 case ICMP_HOST_UNREACH:
2927 case ICMP_PORT_UNREACH:
2930 rxi_SetPeerDead(err, addr, port);
2937 rxi_TranslateICMP(int type, int code)
2940 case ICMP_DEST_UNREACH:
2942 case ICMP_NET_UNREACH:
2943 return "Destination Net Unreachable";
2944 case ICMP_HOST_UNREACH:
2945 return "Destination Host Unreachable";
2946 case ICMP_PROT_UNREACH:
2947 return "Destination Protocol Unreachable";
2948 case ICMP_PORT_UNREACH:
2949 return "Destination Port Unreachable";
2951 return "Destination Net Prohibited";
2953 return "Destination Host Prohibited";
2959 #endif /* AFS_RXERRQ_ENV */
2962 * Get the last network error for a connection
2964 * A "network error" here means an error retrieved from ICMP, or some other
2965 * mechanism outside of Rx that informs us of errors in network reachability.
2967 * If a peer associated with the given Rx connection has received a network
2968 * error recently, this function allows the caller to know what error
2969 * specifically occurred. This can be useful to know, since e.g. ICMP errors
2970 * can cause calls to that peer to be quickly aborted. So, this function can
2971 * help see why a call was aborted due to network errors.
2973 * If we have received traffic from a peer since the last network error, we
2974 * treat that peer as if we had not received an network error for it.
2976 * @param[in] conn The Rx connection to examine
2977 * @param[out] err_origin The origin of the last network error (e.g. ICMP);
2978 * one of the RX_NETWORK_ERROR_ORIGIN_* constants
2979 * @param[out] err_type The type of the last error
2980 * @param[out] err_code The code of the last error
2981 * @param[out] msg Human-readable error message, if applicable; NULL otherwise
2983 * @return If we have an error
2984 * @retval -1 No error to get; 'out' params are undefined
2985 * @retval 0 We have an error; 'out' params contain the last error
2988 rx_GetNetworkError(struct rx_connection *conn, int *err_origin, int *err_type,
2989 int *err_code, const char **msg)
2991 #ifdef AFS_RXERRQ_ENV
2992 struct rx_peer *peer = conn->peer;
2993 if (rx_atomic_read(&peer->neterrs)) {
2994 MUTEX_ENTER(&peer->peer_lock);
2995 *err_origin = peer->last_err_origin;
2996 *err_type = peer->last_err_type;
2997 *err_code = peer->last_err_code;
2998 MUTEX_EXIT(&peer->peer_lock);
3001 if (*err_origin == RX_NETWORK_ERROR_ORIGIN_ICMP) {
3002 *msg = rxi_TranslateICMP(*err_type, *err_code);
3011 /* Find the peer process represented by the supplied (host,port)
3012 * combination. If there is no appropriate active peer structure, a
3013 * new one will be allocated and initialized
3016 rxi_FindPeer(afs_uint32 host, u_short port, int create)
3020 hashIndex = PEER_HASH(host, port);
3021 MUTEX_ENTER(&rx_peerHashTable_lock);
3022 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
3023 if ((pp->host == host) && (pp->port == port))
3028 pp = rxi_AllocPeer(); /* This bzero's *pp */
3029 pp->host = host; /* set here or in InitPeerParams is zero */
3031 #ifdef AFS_RXERRQ_ENV
3032 rx_atomic_set(&pp->neterrs, 0);
3034 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
3035 opr_queue_Init(&pp->rpcStats);
3036 pp->next = rx_peerHashTable[hashIndex];
3037 rx_peerHashTable[hashIndex] = pp;
3038 rxi_InitPeerParams(pp);
3039 if (rx_stats_active)
3040 rx_atomic_inc(&rx_stats.nPeerStructs);
3046 MUTEX_EXIT(&rx_peerHashTable_lock);
3051 /* Find the connection at (host, port) started at epoch, and with the
3052 * given connection id. Creates the server connection if necessary.
3053 * The type specifies whether a client connection or a server
3054 * connection is desired. In both cases, (host, port) specify the
3055 * peer's (host, pair) pair. Client connections are not made
3056 * automatically by this routine. The parameter socket gives the
3057 * socket descriptor on which the packet was received. This is used,
3058 * in the case of server connections, to check that *new* connections
3059 * come via a valid (port, serviceId). Finally, the securityIndex
3060 * parameter must match the existing index for the connection. If a
3061 * server connection is created, it will be created using the supplied
3062 * index, if the index is valid for this service */
3063 static struct rx_connection *
3064 rxi_FindConnection(osi_socket socket, afs_uint32 host,
3065 u_short port, u_short serviceId, afs_uint32 cid,
3066 afs_uint32 epoch, int type, u_int securityIndex,
3067 int *unknownService)
3069 int hashindex, flag, i;
3070 struct rx_connection *conn;
3071 *unknownService = 0;
3072 hashindex = CONN_HASH(host, port, cid, epoch, type);
3073 MUTEX_ENTER(&rx_connHashTable_lock);
3074 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
3075 rx_connHashTable[hashindex],
3078 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
3079 && (epoch == conn->epoch)) {
3080 struct rx_peer *pp = conn->peer;
3081 if (securityIndex != conn->securityIndex) {
3082 /* this isn't supposed to happen, but someone could forge a packet
3083 * like this, and there seems to be some CM bug that makes this
3084 * happen from time to time -- in which case, the fileserver
3086 MUTEX_EXIT(&rx_connHashTable_lock);
3087 return (struct rx_connection *)0;
3089 if (pp->host == host && pp->port == port)
3091 if (type == RX_CLIENT_CONNECTION && pp->port == port)
3093 /* So what happens when it's a callback connection? */
3094 if ( /*type == RX_CLIENT_CONNECTION && */
3095 (conn->epoch & 0x80000000))
3099 /* the connection rxLastConn that was used the last time is not the
3100 ** one we are looking for now. Hence, start searching in the hash */
3102 conn = rx_connHashTable[hashindex];
3107 struct rx_service *service;
3108 if (type == RX_CLIENT_CONNECTION) {
3109 MUTEX_EXIT(&rx_connHashTable_lock);
3110 return (struct rx_connection *)0;
3112 service = rxi_FindService(socket, serviceId);
3113 if (!service || (securityIndex >= service->nSecurityObjects)
3114 || (service->securityObjects[securityIndex] == 0)) {
3115 MUTEX_EXIT(&rx_connHashTable_lock);
3116 *unknownService = 1;
3117 return (struct rx_connection *)0;
3119 conn = rxi_AllocConnection(); /* This bzero's the connection */
3120 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
3121 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
3122 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
3123 conn->next = rx_connHashTable[hashindex];
3124 rx_connHashTable[hashindex] = conn;
3125 conn->peer = rxi_FindPeer(host, port, 1);
3126 conn->type = RX_SERVER_CONNECTION;
3127 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
3128 conn->epoch = epoch;
3129 conn->cid = cid & RX_CIDMASK;
3130 conn->ackRate = RX_FAST_ACK_RATE;
3131 conn->service = service;
3132 conn->serviceId = serviceId;
3133 conn->securityIndex = securityIndex;
3134 conn->securityObject = service->securityObjects[securityIndex];
3135 conn->nSpecific = 0;
3136 conn->specific = NULL;
3137 rx_SetConnDeadTime(conn, service->connDeadTime);
3138 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
3139 for (i = 0; i < RX_MAXCALLS; i++) {
3140 conn->twind[i] = rx_initSendWindow;
3141 conn->rwind[i] = rx_initReceiveWindow;
3143 /* Notify security object of the new connection */
3144 RXS_NewConnection(conn->securityObject, conn);
3145 /* XXXX Connection timeout? */
3146 if (service->newConnProc)
3147 (*service->newConnProc) (conn);
3148 if (rx_stats_active)
3149 rx_atomic_inc(&rx_stats.nServerConns);
3152 rx_GetConnection(conn);
3154 rxLastConn = conn; /* store this connection as the last conn used */
3155 MUTEX_EXIT(&rx_connHashTable_lock);
3160 * Abort the call if the server is over the busy threshold. This
3161 * can be used without requiring a call structure be initialised,
3162 * or connected to a particular channel
3165 rxi_AbortIfServerBusy(osi_socket socket, struct rx_connection *conn,
3166 struct rx_packet *np)
3170 if ((rx_BusyThreshold > 0) &&
3171 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3172 MUTEX_ENTER(&conn->conn_data_lock);
3173 serial = ++conn->serial;
3174 MUTEX_EXIT(&conn->conn_data_lock);
3175 rxi_SendRawAbort(socket, conn->peer->host, conn->peer->port,
3176 serial, rx_BusyError, np, 0);
3177 if (rx_stats_active)
3178 rx_atomic_inc(&rx_stats.nBusies);
3185 static_inline struct rx_call *
3186 rxi_ReceiveClientCall(struct rx_packet *np, struct rx_connection *conn)
3189 struct rx_call *call;
3191 channel = np->header.cid & RX_CHANNELMASK;
3192 MUTEX_ENTER(&conn->conn_call_lock);
3193 call = conn->call[channel];
3194 if (np->header.type == RX_PACKET_TYPE_BUSY) {
3195 conn->lastBusy[channel] = clock_Sec();
3197 if (!call || conn->callNumber[channel] != np->header.callNumber) {
3198 MUTEX_EXIT(&conn->conn_call_lock);
3199 if (rx_stats_active)
3200 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3204 MUTEX_ENTER(&call->lock);
3205 MUTEX_EXIT(&conn->conn_call_lock);
3207 if ((call->state == RX_STATE_DALLY)
3208 && np->header.type == RX_PACKET_TYPE_ACK) {
3209 if (rx_stats_active)
3210 rx_atomic_inc(&rx_stats.ignorePacketDally);
3211 MUTEX_EXIT(&call->lock);
3218 static_inline struct rx_call *
3219 rxi_ReceiveServerCall(osi_socket socket, struct rx_packet *np,
3220 struct rx_connection *conn)
3223 struct rx_call *call;
3225 channel = np->header.cid & RX_CHANNELMASK;
3226 MUTEX_ENTER(&conn->conn_call_lock);
3227 call = conn->call[channel];
3230 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3231 MUTEX_EXIT(&conn->conn_call_lock);
3235 call = rxi_NewCall(conn, channel); /* returns locked call */
3236 *call->callNumber = np->header.callNumber;
3237 MUTEX_EXIT(&conn->conn_call_lock);
3239 call->state = RX_STATE_PRECALL;
3240 clock_GetTime(&call->queueTime);
3241 call->app.bytesSent = 0;
3242 call->app.bytesRcvd = 0;
3243 rxi_KeepAliveOn(call);
3248 if (np->header.callNumber == conn->callNumber[channel]) {
3249 MUTEX_ENTER(&call->lock);
3250 MUTEX_EXIT(&conn->conn_call_lock);
3254 if (np->header.callNumber < conn->callNumber[channel]) {
3255 MUTEX_EXIT(&conn->conn_call_lock);
3256 if (rx_stats_active)
3257 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3261 MUTEX_ENTER(&call->lock);
3262 MUTEX_EXIT(&conn->conn_call_lock);
3264 /* Wait until the transmit queue is idle before deciding
3265 * whether to reset the current call. Chances are that the
3266 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3269 #ifdef RX_ENABLE_LOCKS
3270 if (call->state == RX_STATE_ACTIVE && !call->error) {
3271 rxi_WaitforTQBusy(call);
3272 /* If we entered error state while waiting,
3273 * must call rxi_CallError to permit rxi_ResetCall
3274 * to processed when the tqWaiter count hits zero.
3277 rxi_CallError(call, call->error);
3278 MUTEX_EXIT(&call->lock);
3282 #endif /* RX_ENABLE_LOCKS */
3283 /* If the new call cannot be taken right now send a busy and set
3284 * the error condition in this call, so that it terminates as
3285 * quickly as possible */
3286 if (call->state == RX_STATE_ACTIVE) {
3287 rxi_CallError(call, RX_CALL_DEAD);
3288 rxi_SendSpecial(call, conn, NULL, RX_PACKET_TYPE_BUSY,
3290 MUTEX_EXIT(&call->lock);
3294 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3295 MUTEX_EXIT(&call->lock);
3299 rxi_ResetCall(call, 0);
3300 /* The conn_call_lock is not held but no one else should be
3301 * using this call channel while we are processing this incoming
3302 * packet. This assignment should be safe.
3304 *call->callNumber = np->header.callNumber;
3305 call->state = RX_STATE_PRECALL;
3306 clock_GetTime(&call->queueTime);
3307 call->app.bytesSent = 0;
3308 call->app.bytesRcvd = 0;
3309 rxi_KeepAliveOn(call);
3315 /* There are two packet tracing routines available for testing and monitoring
3316 * Rx. One is called just after every packet is received and the other is
3317 * called just before every packet is sent. Received packets, have had their
3318 * headers decoded, and packets to be sent have not yet had their headers
3319 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3320 * containing the network address. Both can be modified. The return value, if
3321 * non-zero, indicates that the packet should be dropped. */
3323 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3324 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3326 /* A packet has been received off the interface. Np is the packet, socket is
3327 * the socket number it was received from (useful in determining which service
3328 * this packet corresponds to), and (host, port) reflect the host,port of the
3329 * sender. This call returns the packet to the caller if it is finished with
3330 * it, rather than de-allocating it, just as a small performance hack */
3333 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3334 afs_uint32 host, u_short port, int *tnop,
3335 struct rx_call **newcallp)
3337 struct rx_call *call;
3338 struct rx_connection *conn;
3340 int unknownService = 0;
3344 struct rx_packet *tnp;
3347 /* We don't print out the packet until now because (1) the time may not be
3348 * accurate enough until now in the lwp implementation (rx_Listener only gets
3349 * the time after the packet is read) and (2) from a protocol point of view,
3350 * this is the first time the packet has been seen */
3351 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3352 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3353 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3354 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3355 np->header.epoch, np->header.cid, np->header.callNumber,
3356 np->header.seq, np->header.flags, np));
3359 /* Account for connectionless packets */
3360 if (rx_stats_active &&
3361 ((np->header.type == RX_PACKET_TYPE_VERSION) ||
3362 (np->header.type == RX_PACKET_TYPE_DEBUG))) {
3363 struct rx_peer *peer;
3365 /* Try to look up the peer structure, but don't create one */
3366 peer = rxi_FindPeer(host, port, 0);
3368 /* Since this may not be associated with a connection, it may have
3369 * no refCount, meaning we could race with ReapConnections
3372 if (peer && (peer->refCount > 0)) {
3373 #ifdef AFS_RXERRQ_ENV
3374 if (rx_atomic_read(&peer->neterrs)) {
3375 rx_atomic_set(&peer->neterrs, 0);
3378 MUTEX_ENTER(&peer->peer_lock);
3379 peer->bytesReceived += np->length;
3380 MUTEX_EXIT(&peer->peer_lock);
3384 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3385 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3388 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3389 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3392 /* If an input tracer function is defined, call it with the packet and
3393 * network address. Note this function may modify its arguments. */
3394 if (rx_justReceived) {
3395 struct sockaddr_in addr;
3397 addr.sin_family = AF_INET;
3398 addr.sin_port = port;
3399 addr.sin_addr.s_addr = host;
3400 memset(&addr.sin_zero, 0, sizeof(addr.sin_zero));
3401 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3402 addr.sin_len = sizeof(addr);
3403 #endif /* AFS_OSF_ENV */
3404 drop = (*rx_justReceived) (np, &addr);
3405 /* drop packet if return value is non-zero */
3408 port = addr.sin_port; /* in case fcn changed addr */
3409 host = addr.sin_addr.s_addr;
3413 /* If packet was not sent by the client, then *we* must be the client */
3414 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3415 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3417 /* Find the connection (or fabricate one, if we're the server & if
3418 * necessary) associated with this packet */
3420 rxi_FindConnection(socket, host, port, np->header.serviceId,
3421 np->header.cid, np->header.epoch, type,
3422 np->header.securityIndex, &unknownService);
3424 /* To avoid having 2 connections just abort at each other,
3425 don't abort an abort. */
3427 if (unknownService && (np->header.type != RX_PACKET_TYPE_ABORT))
3428 rxi_SendRawAbort(socket, host, port, 0, RX_INVALID_OPERATION,
3433 #ifdef AFS_RXERRQ_ENV
3434 if (rx_atomic_read(&conn->peer->neterrs)) {
3435 rx_atomic_set(&conn->peer->neterrs, 0);
3439 /* If we're doing statistics, then account for the incoming packet */
3440 if (rx_stats_active) {
3441 MUTEX_ENTER(&conn->peer->peer_lock);
3442 conn->peer->bytesReceived += np->length;
3443 MUTEX_EXIT(&conn->peer->peer_lock);
3446 /* If the connection is in an error state, send an abort packet and ignore
3447 * the incoming packet */
3449 /* Don't respond to an abort packet--we don't want loops! */
3450 MUTEX_ENTER(&conn->conn_data_lock);
3451 if (np->header.type != RX_PACKET_TYPE_ABORT)
3452 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3453 putConnection(conn);
3454 MUTEX_EXIT(&conn->conn_data_lock);
3458 /* Check for connection-only requests (i.e. not call specific). */
3459 if (np->header.callNumber == 0) {
3460 switch (np->header.type) {
3461 case RX_PACKET_TYPE_ABORT: {
3462 /* What if the supplied error is zero? */
3463 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3464 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3465 rxi_ConnectionError(conn, errcode);
3466 putConnection(conn);
3469 case RX_PACKET_TYPE_CHALLENGE:
3470 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3471 putConnection(conn);
3473 case RX_PACKET_TYPE_RESPONSE:
3474 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3475 putConnection(conn);
3477 case RX_PACKET_TYPE_PARAMS:
3478 case RX_PACKET_TYPE_PARAMS + 1:
3479 case RX_PACKET_TYPE_PARAMS + 2:
3480 /* ignore these packet types for now */
3481 putConnection(conn);
3485 /* Should not reach here, unless the peer is broken: send an
3487 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3488 MUTEX_ENTER(&conn->conn_data_lock);
3489 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3490 putConnection(conn);
3491 MUTEX_EXIT(&conn->conn_data_lock);
3496 if (type == RX_SERVER_CONNECTION)
3497 call = rxi_ReceiveServerCall(socket, np, conn);
3499 call = rxi_ReceiveClientCall(np, conn);
3502 putConnection(conn);
3506 MUTEX_ASSERT(&call->lock);
3507 /* Set remote user defined status from packet */
3508 call->remoteStatus = np->header.userStatus;
3510 /* Now do packet type-specific processing */
3511 switch (np->header.type) {
3512 case RX_PACKET_TYPE_DATA:
3513 /* If we're a client, and receiving a response, then all the packets
3514 * we transmitted packets are implicitly acknowledged. */
3515 if (type == RX_CLIENT_CONNECTION && !opr_queue_IsEmpty(&call->tq))
3516 rxi_AckAllInTransmitQueue(call);
3518 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3521 case RX_PACKET_TYPE_ACK:
3522 /* Respond immediately to ack packets requesting acknowledgement
3524 if (np->header.flags & RX_REQUEST_ACK) {
3526 (void)rxi_SendCallAbort(call, 0, 1, 0);
3528 (void)rxi_SendAck(call, 0, np->header.serial,
3529 RX_ACK_PING_RESPONSE, 1);
3531 np = rxi_ReceiveAckPacket(call, np, 1);
3533 case RX_PACKET_TYPE_ABORT: {
3534 /* An abort packet: reset the call, passing the error up to the user. */
3535 /* What if error is zero? */
3536 /* What if the error is -1? the application will treat it as a timeout. */
3537 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3538 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3539 rxi_CallError(call, errdata);
3540 MUTEX_EXIT(&call->lock);
3541 putConnection(conn);
3542 return np; /* xmitting; drop packet */
3544 case RX_PACKET_TYPE_BUSY:
3545 /* Mostly ignore BUSY packets. We will update lastReceiveTime below,
3546 * so we don't think the endpoint is completely dead, but otherwise
3547 * just act as if we never saw anything. If all we get are BUSY packets
3548 * back, then we will eventually error out with RX_CALL_TIMEOUT if the
3549 * connection is configured with idle/hard timeouts. */
3552 case RX_PACKET_TYPE_ACKALL:
3553 /* All packets acknowledged, so we can drop all packets previously
3554 * readied for sending */
3555 rxi_AckAllInTransmitQueue(call);
3558 /* Should not reach here, unless the peer is broken: send an abort
3560 rxi_CallError(call, RX_PROTOCOL_ERROR);
3561 np = rxi_SendCallAbort(call, np, 1, 0);
3564 /* Note when this last legitimate packet was received, for keep-alive
3565 * processing. Note, we delay getting the time until now in the hope that
3566 * the packet will be delivered to the user before any get time is required
3567 * (if not, then the time won't actually be re-evaluated here). */
3568 call->lastReceiveTime = clock_Sec();
3569 MUTEX_EXIT(&call->lock);
3570 putConnection(conn);
3574 /* return true if this is an "interesting" connection from the point of view
3575 of someone trying to debug the system */
3577 rxi_IsConnInteresting(struct rx_connection *aconn)
3580 struct rx_call *tcall;
3582 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3585 for (i = 0; i < RX_MAXCALLS; i++) {
3586 tcall = aconn->call[i];
3588 if ((tcall->state == RX_STATE_PRECALL)
3589 || (tcall->state == RX_STATE_ACTIVE))
3591 if ((tcall->app.mode == RX_MODE_SENDING)
3592 || (tcall->app.mode == RX_MODE_RECEIVING))
3600 /* if this is one of the last few packets AND it wouldn't be used by the
3601 receiving call to immediately satisfy a read request, then drop it on
3602 the floor, since accepting it might prevent a lock-holding thread from
3603 making progress in its reading. If a call has been cleared while in
3604 the precall state then ignore all subsequent packets until the call
3605 is assigned to a thread. */
3608 TooLow(struct rx_packet *ap, struct rx_call *acall)
3612 MUTEX_ENTER(&rx_quota_mutex);
3613 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3614 && (acall->state == RX_STATE_PRECALL))
3615 || ((rx_nFreePackets < rxi_dataQuota + 2)
3616 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3617 && (acall->flags & RX_CALL_READER_WAIT)))) {
3620 MUTEX_EXIT(&rx_quota_mutex);
3626 * Clear the attach wait flag on a connection and proceed.
3628 * Any processing waiting for a connection to be attached should be
3629 * unblocked. We clear the flag and do any other needed tasks.
3632 * the conn to unmark waiting for attach
3634 * @pre conn's conn_data_lock must be locked before calling this function
3638 rxi_ConnClearAttachWait(struct rx_connection *conn)
3640 /* Indicate that rxi_CheckReachEvent is no longer running by
3641 * clearing the flag. Must be atomic under conn_data_lock to
3642 * avoid a new call slipping by: rxi_CheckConnReach holds
3643 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3645 conn->flags &= ~RX_CONN_ATTACHWAIT;
3646 if (conn->flags & RX_CONN_NAT_PING) {
3647 conn->flags &= ~RX_CONN_NAT_PING;
3648 rxi_ScheduleNatKeepAliveEvent(conn);
3653 * Event handler function for connection-specific events for checking
3654 * reachability. Also called directly from main code with |event| == NULL
3655 * in order to trigger the initial reachability check.
3657 * When |event| == NULL, must be called with the connection data lock held,
3658 * but returns with the lock unlocked.
3661 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3663 struct rx_connection *conn = arg1;
3664 struct rx_call *acall = arg2;
3665 struct rx_call *call = acall;
3666 struct clock when, now;
3670 MUTEX_ENTER(&conn->conn_data_lock);
3672 MUTEX_ASSERT(&conn->conn_data_lock);
3674 if (event != NULL && event == conn->checkReachEvent)
3675 rxevent_Put(&conn->checkReachEvent);
3676 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3677 MUTEX_EXIT(&conn->conn_data_lock);
3681 MUTEX_ENTER(&conn->conn_call_lock);
3682 MUTEX_ENTER(&conn->conn_data_lock);
3683 for (i = 0; i < RX_MAXCALLS; i++) {
3684 struct rx_call *tc = conn->call[i];
3685 if (tc && tc->state == RX_STATE_PRECALL) {
3691 rxi_ConnClearAttachWait(conn);
3692 MUTEX_EXIT(&conn->conn_data_lock);
3693 MUTEX_EXIT(&conn->conn_call_lock);
3698 MUTEX_ENTER(&call->lock);
3699 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3701 MUTEX_EXIT(&call->lock);
3703 clock_GetTime(&now);
3705 when.sec += RX_CHECKREACH_TIMEOUT;
3706 MUTEX_ENTER(&conn->conn_data_lock);
3707 if (!conn->checkReachEvent) {
3708 rx_GetConnection(conn);
3709 conn->checkReachEvent = rxevent_Post(&when, &now,
3710 rxi_CheckReachEvent, conn,
3713 MUTEX_EXIT(&conn->conn_data_lock);
3716 /* If fired as an event handler, drop our refcount on the connection. */
3718 putConnection(conn);
3722 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3724 struct rx_service *service = conn->service;
3725 struct rx_peer *peer = conn->peer;
3726 afs_uint32 now, lastReach;
3728 if (service->checkReach == 0)
3732 MUTEX_ENTER(&peer->peer_lock);
3733 lastReach = peer->lastReachTime;
3734 MUTEX_EXIT(&peer->peer_lock);
3735 if (now - lastReach < RX_CHECKREACH_TTL)
3738 MUTEX_ENTER(&conn->conn_data_lock);
3739 if (conn->flags & RX_CONN_ATTACHWAIT) {
3740 MUTEX_EXIT(&conn->conn_data_lock);
3743 conn->flags |= RX_CONN_ATTACHWAIT;
3744 if (conn->checkReachEvent == NULL) {
3745 /* rxi_CheckReachEvent(NULL, ...) will drop the lock. */
3746 rxi_CheckReachEvent(NULL, conn, call, 0);
3748 MUTEX_EXIT(&conn->conn_data_lock);
3754 /* try to attach call, if authentication is complete */
3756 TryAttach(struct rx_call *acall, osi_socket socket,
3757 int *tnop, struct rx_call **newcallp,
3760 struct rx_connection *conn = acall->conn;
3762 if (conn->type == RX_SERVER_CONNECTION
3763 && acall->state == RX_STATE_PRECALL) {
3764 /* Don't attach until we have any req'd. authentication. */
3765 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3766 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3767 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3768 /* Note: this does not necessarily succeed; there
3769 * may not any proc available
3772 rxi_ChallengeOn(acall->conn);
3777 /* A data packet has been received off the interface. This packet is
3778 * appropriate to the call (the call is in the right state, etc.). This
3779 * routine can return a packet to the caller, for re-use */
3781 static struct rx_packet *
3782 rxi_ReceiveDataPacket(struct rx_call *call,
3783 struct rx_packet *np, int istack,
3784 osi_socket socket, afs_uint32 host, u_short port,
3785 int *tnop, struct rx_call **newcallp)
3787 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3792 afs_uint32 serial=0, flags=0;
3794 struct rx_packet *tnp;
3795 if (rx_stats_active)
3796 rx_atomic_inc(&rx_stats.dataPacketsRead);
3799 /* If there are no packet buffers, drop this new packet, unless we can find
3800 * packet buffers from inactive calls */
3802 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3803 MUTEX_ENTER(&rx_freePktQ_lock);
3804 rxi_NeedMorePackets = TRUE;
3805 MUTEX_EXIT(&rx_freePktQ_lock);
3806 if (rx_stats_active)
3807 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3808 rxi_calltrace(RX_TRACE_DROP, call);
3809 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3810 /* We used to clear the receive queue here, in an attempt to free
3811 * packets. However this is unsafe if the queue has received a
3812 * soft ACK for the final packet */
3813 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3819 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3820 * packet is one of several packets transmitted as a single
3821 * datagram. Do not send any soft or hard acks until all packets
3822 * in a jumbogram have been processed. Send negative acks right away.
3824 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3825 /* tnp is non-null when there are more packets in the
3826 * current jumbo gram */
3833 seq = np->header.seq;
3834 serial = np->header.serial;
3835 flags = np->header.flags;
3837 /* If the call is in an error state, send an abort message */
3839 return rxi_SendCallAbort(call, np, istack, 0);
3841 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3842 * AFS 3.5 jumbogram. */
3843 if (flags & RX_JUMBO_PACKET) {
3844 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3849 if (np->header.spare != 0) {
3850 MUTEX_ENTER(&call->conn->conn_data_lock);
3851 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3852 MUTEX_EXIT(&call->conn->conn_data_lock);
3855 /* The usual case is that this is the expected next packet */
3856 if (seq == call->rnext) {
3858 /* Check to make sure it is not a duplicate of one already queued */
3859 if (!opr_queue_IsEmpty(&call->rq)
3860 && opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq == seq) {
3861 if (rx_stats_active)
3862 rx_atomic_inc(&rx_stats.dupPacketsRead);
3863 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3864 rxi_CancelDelayedAckEvent(call);
3865 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3871 /* It's the next packet. Stick it on the receive queue
3872 * for this call. Set newPackets to make sure we wake
3873 * the reader once all packets have been processed */
3874 #ifdef RX_TRACK_PACKETS
3875 np->flags |= RX_PKTFLAG_RQ;
3877 opr_queue_Prepend(&call->rq, &np->entry);
3878 #ifdef RXDEBUG_PACKET
3880 #endif /* RXDEBUG_PACKET */
3882 np = NULL; /* We can't use this anymore */
3885 /* If an ack is requested then set a flag to make sure we
3886 * send an acknowledgement for this packet */
3887 if (flags & RX_REQUEST_ACK) {
3888 ackNeeded = RX_ACK_REQUESTED;
3891 /* Keep track of whether we have received the last packet */
3892 if (flags & RX_LAST_PACKET) {
3893 call->flags |= RX_CALL_HAVE_LAST;
3897 /* Check whether we have all of the packets for this call */
3898 if (call->flags & RX_CALL_HAVE_LAST) {
3899 afs_uint32 tseq; /* temporary sequence number */
3900 struct opr_queue *cursor;
3902 for (tseq = seq, opr_queue_Scan(&call->rq, cursor)) {
3903 struct rx_packet *tp;
3905 tp = opr_queue_Entry(cursor, struct rx_packet, entry);
3906 if (tseq != tp->header.seq)
3908 if (tp->header.flags & RX_LAST_PACKET) {
3909 call->flags |= RX_CALL_RECEIVE_DONE;
3916 /* Provide asynchronous notification for those who want it
3917 * (e.g. multi rx) */
3918 if (call->arrivalProc) {
3919 (*call->arrivalProc) (call, call->arrivalProcHandle,
3920 call->arrivalProcArg);
3921 call->arrivalProc = (void (*)())0;
3924 /* Update last packet received */
3927 /* If there is no server process serving this call, grab
3928 * one, if available. We only need to do this once. If a
3929 * server thread is available, this thread becomes a server
3930 * thread and the server thread becomes a listener thread. */
3932 TryAttach(call, socket, tnop, newcallp, 0);
3935 /* This is not the expected next packet. */
3937 /* Determine whether this is a new or old packet, and if it's
3938 * a new one, whether it fits into the current receive window.
3939 * Also figure out whether the packet was delivered in sequence.
3940 * We use the prev variable to determine whether the new packet
3941 * is the successor of its immediate predecessor in the
3942 * receive queue, and the missing flag to determine whether
3943 * any of this packets predecessors are missing. */
3945 afs_uint32 prev; /* "Previous packet" sequence number */
3946 struct opr_queue *cursor;
3947 int missing; /* Are any predecessors missing? */
3949 /* If the new packet's sequence number has been sent to the
3950 * application already, then this is a duplicate */
3951 if (seq < call->rnext) {
3952 if (rx_stats_active)
3953 rx_atomic_inc(&rx_stats.dupPacketsRead);
3954 rxi_CancelDelayedAckEvent(call);
3955 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3961 /* If the sequence number is greater than what can be
3962 * accomodated by the current window, then send a negative
3963 * acknowledge and drop the packet */
3964 if ((call->rnext + call->rwind) <= seq) {
3965 rxi_CancelDelayedAckEvent(call);
3966 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3973 /* Look for the packet in the queue of old received packets */
3974 prev = call->rnext - 1;
3976 for (opr_queue_Scan(&call->rq, cursor)) {
3977 struct rx_packet *tp
3978 = opr_queue_Entry(cursor, struct rx_packet, entry);
3980 /*Check for duplicate packet */
3981 if (seq == tp->header.seq) {
3982 if (rx_stats_active)
3983 rx_atomic_inc(&rx_stats.dupPacketsRead);
3984 rxi_CancelDelayedAckEvent(call);
3985 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3991 /* If we find a higher sequence packet, break out and
3992 * insert the new packet here. */
3993 if (seq < tp->header.seq)
3995 /* Check for missing packet */
3996 if (tp->header.seq != prev + 1) {
4000 prev = tp->header.seq;
4003 /* Keep track of whether we have received the last packet. */
4004 if (flags & RX_LAST_PACKET) {
4005 call->flags |= RX_CALL_HAVE_LAST;
4008 /* It's within the window: add it to the the receive queue.
4009 * tp is left by the previous loop either pointing at the
4010 * packet before which to insert the new packet, or at the
4011 * queue head if the queue is empty or the packet should be
4013 #ifdef RX_TRACK_PACKETS
4014 np->flags |= RX_PKTFLAG_RQ;
4016 #ifdef RXDEBUG_PACKET
4018 #endif /* RXDEBUG_PACKET */
4019 opr_queue_InsertBefore(cursor, &np->entry);
4023 /* Check whether we have all of the packets for this call */
4024 if ((call->flags & RX_CALL_HAVE_LAST)
4025 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4026 afs_uint32 tseq; /* temporary sequence number */
4029 for (opr_queue_Scan(&call->rq, cursor)) {
4030 struct rx_packet *tp
4031 = opr_queue_Entry(cursor, struct rx_packet, entry);
4032 if (tseq != tp->header.seq)
4034 if (tp->header.flags & RX_LAST_PACKET) {
4035 call->flags |= RX_CALL_RECEIVE_DONE;
4042 /* We need to send an ack of the packet is out of sequence,
4043 * or if an ack was requested by the peer. */
4044 if (seq != prev + 1 || missing) {
4045 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4046 } else if (flags & RX_REQUEST_ACK) {
4047 ackNeeded = RX_ACK_REQUESTED;
4050 /* Acknowledge the last packet for each call */
4051 if (flags & RX_LAST_PACKET) {
4062 * If the receiver is waiting for an iovec, fill the iovec
4063 * using the data from the receive queue */
4064 if (call->flags & RX_CALL_IOVEC_WAIT) {
4065 didHardAck = rxi_FillReadVec(call, serial);
4066 /* the call may have been aborted */
4075 /* Wakeup the reader if any */
4076 if ((call->flags & RX_CALL_READER_WAIT)
4077 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4078 || (call->iovNext >= call->iovMax)
4079 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4080 call->flags &= ~RX_CALL_READER_WAIT;
4081 #ifdef RX_ENABLE_LOCKS
4082 CV_BROADCAST(&call->cv_rq);
4084 osi_rxWakeup(&call->rq);
4090 * Send an ack when requested by the peer, or once every
4091 * rxi_SoftAckRate packets until the last packet has been
4092 * received. Always send a soft ack for the last packet in
4093 * the server's reply. */
4095 rxi_CancelDelayedAckEvent(call);
4096 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4097 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4098 rxi_CancelDelayedAckEvent(call);
4099 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4100 } else if (call->nSoftAcks) {
4101 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4102 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4104 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4105 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4106 rxi_CancelDelayedAckEvent(call);
4113 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4115 struct rx_peer *peer = conn->peer;
4117 MUTEX_ENTER(&peer->peer_lock);
4118 peer->lastReachTime = clock_Sec();
4119 MUTEX_EXIT(&peer->peer_lock);
4121 MUTEX_ENTER(&conn->conn_data_lock);
4122 if (conn->flags & RX_CONN_ATTACHWAIT) {
4125 rxi_ConnClearAttachWait(conn);
4126 MUTEX_EXIT(&conn->conn_data_lock);
4128 for (i = 0; i < RX_MAXCALLS; i++) {
4129 struct rx_call *call = conn->call[i];
4132 MUTEX_ENTER(&call->lock);
4133 /* tnop can be null if newcallp is null */
4134 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4136 MUTEX_EXIT(&call->lock);
4140 MUTEX_EXIT(&conn->conn_data_lock);
4143 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4145 rx_ack_reason(int reason)
4148 case RX_ACK_REQUESTED:
4150 case RX_ACK_DUPLICATE:
4152 case RX_ACK_OUT_OF_SEQUENCE:
4154 case RX_ACK_EXCEEDS_WINDOW:
4156 case RX_ACK_NOSPACE:
4160 case RX_ACK_PING_RESPONSE:
4173 /* The real smarts of the whole thing. */
4174 static struct rx_packet *
4175 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4178 struct rx_ackPacket *ap;
4180 struct rx_packet *tp;
4181 struct rx_connection *conn = call->conn;
4182 struct rx_peer *peer = conn->peer;
4183 struct opr_queue *cursor;
4184 struct clock now; /* Current time, for RTT calculations */
4192 int newAckCount = 0;
4193 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4194 int pktsize = 0; /* Set if we need to update the peer mtu */
4195 int conn_data_locked = 0;
4197 if (rx_stats_active)
4198 rx_atomic_inc(&rx_stats.ackPacketsRead);
4199 ap = (struct rx_ackPacket *)rx_DataOf(np);
4200 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4202 return np; /* truncated ack packet */
4204 /* depends on ack packet struct */
4205 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4206 first = ntohl(ap->firstPacket);
4207 prev = ntohl(ap->previousPacket);
4208 serial = ntohl(ap->serial);
4211 * Ignore ack packets received out of order while protecting
4212 * against peers that set the previousPacket field to a packet
4213 * serial number instead of a sequence number.
4215 if (first < call->tfirst ||
4216 (first == call->tfirst && prev < call->tprev && prev < call->tfirst
4223 if (np->header.flags & RX_SLOW_START_OK) {
4224 call->flags |= RX_CALL_SLOW_START_OK;
4227 if (ap->reason == RX_ACK_PING_RESPONSE)
4228 rxi_UpdatePeerReach(conn, call);
4230 if (conn->lastPacketSizeSeq) {
4231 MUTEX_ENTER(&conn->conn_data_lock);
4232 conn_data_locked = 1;
4233 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4234 pktsize = conn->lastPacketSize;
4235 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4238 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4239 if (!conn_data_locked) {
4240 MUTEX_ENTER(&conn->conn_data_lock);
4241 conn_data_locked = 1;
4243 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4244 /* process mtu ping ack */
4245 pktsize = conn->lastPingSize;
4246 conn->lastPingSizeSer = conn->lastPingSize = 0;
4250 if (conn_data_locked) {
4251 MUTEX_EXIT(&conn->conn_data_lock);
4252 conn_data_locked = 0;
4256 if (rxdebug_active) {
4260 len = _snprintf(msg, sizeof(msg),
4261 "tid[%d] RACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4262 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4263 ntohl(ap->serial), ntohl(ap->previousPacket),
4264 (unsigned int)np->header.seq, ntohl(ap->firstPacket),
4265 ap->nAcks, ntohs(ap->bufferSpace) );
4269 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4270 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4274 OutputDebugString(msg);
4276 #else /* AFS_NT40_ENV */
4279 "RACK: reason %x previous %u seq %u serial %u first %u",
4280 ap->reason, ntohl(ap->previousPacket),
4281 (unsigned int)np->header.seq, (unsigned int)serial,
4282 ntohl(ap->firstPacket));
4285 for (offset = 0; offset < nAcks; offset++)
4286 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4291 #endif /* AFS_NT40_ENV */
4294 MUTEX_ENTER(&peer->peer_lock);
4297 * Start somewhere. Can't assume we can send what we can receive,
4298 * but we are clearly receiving.
4300 if (!peer->maxPacketSize)
4301 peer->maxPacketSize = RX_MIN_PACKET_SIZE - RX_HEADER_SIZE;
4303 if (pktsize > peer->maxPacketSize) {
4304 peer->maxPacketSize = pktsize;
4305 if ((pktsize + RX_HEADER_SIZE > peer->ifMTU)) {
4306 peer->ifMTU = pktsize + RX_HEADER_SIZE;
4307 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4308 rxi_ScheduleGrowMTUEvent(call, 1);
4313 clock_GetTime(&now);
4315 /* The transmit queue splits into 4 sections.
4317 * The first section is packets which have now been acknowledged
4318 * by a window size change in the ack. These have reached the
4319 * application layer, and may be discarded. These are packets
4320 * with sequence numbers < ap->firstPacket.
4322 * The second section is packets which have sequence numbers in
4323 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4324 * contents of the packet's ack array determines whether these
4325 * packets are acknowledged or not.
4327 * The third section is packets which fall above the range
4328 * addressed in the ack packet. These have not yet been received
4331 * The four section is packets which have not yet been transmitted.
4332 * These packets will have a header.serial of 0.
4335 /* First section - implicitly acknowledged packets that can be
4339 tp = opr_queue_First(&call->tq, struct rx_packet, entry);
4340 while(!opr_queue_IsEnd(&call->tq, &tp->entry) && tp->header.seq < first) {
4341 struct rx_packet *next;
4343 next = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4344 call->tfirst = tp->header.seq + 1;
4346 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4348 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4351 #ifdef RX_ENABLE_LOCKS
4352 /* XXX Hack. Because we have to release the global call lock when sending
4353 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4354 * in rxi_Start sending packets out because packets may move to the
4355 * freePacketQueue as result of being here! So we drop these packets until
4356 * we're safely out of the traversing. Really ugly!
4357 * To make it even uglier, if we're using fine grain locking, we can
4358 * set the ack bits in the packets and have rxi_Start remove the packets
4359 * when it's done transmitting.
4361 if (call->flags & RX_CALL_TQ_BUSY) {
4362 tp->flags |= RX_PKTFLAG_ACKED;
4363 call->flags |= RX_CALL_TQ_SOME_ACKED;
4365 #endif /* RX_ENABLE_LOCKS */
4367 opr_queue_Remove(&tp->entry);
4368 #ifdef RX_TRACK_PACKETS
4369 tp->flags &= ~RX_PKTFLAG_TQ;
4371 #ifdef RXDEBUG_PACKET
4373 #endif /* RXDEBUG_PACKET */
4374 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4379 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4381 /* Second section of the queue - packets for which we are receiving
4384 * Go through the explicit acks/nacks and record the results in
4385 * the waiting packets. These are packets that can't be released
4386 * yet, even with a positive acknowledge. This positive
4387 * acknowledge only means the packet has been received by the
4388 * peer, not that it will be retained long enough to be sent to
4389 * the peer's upper level. In addition, reset the transmit timers
4390 * of any missing packets (those packets that must be missing
4391 * because this packet was out of sequence) */
4393 call->nSoftAcked = 0;
4395 while (!opr_queue_IsEnd(&call->tq, &tp->entry)
4396 && tp->header.seq < first + nAcks) {
4397 /* Set the acknowledge flag per packet based on the
4398 * information in the ack packet. An acknowlegded packet can
4399 * be downgraded when the server has discarded a packet it
4400 * soacked previously, or when an ack packet is received
4401 * out of sequence. */
4402 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4403 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4405 tp->flags |= RX_PKTFLAG_ACKED;
4406 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4413 } else /* RX_ACK_TYPE_NACK */ {
4414 tp->flags &= ~RX_PKTFLAG_ACKED;
4418 tp = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4421 /* We don't need to take any action with the 3rd or 4th section in the
4422 * queue - they're not addressed by the contents of this ACK packet.
4425 /* if the ack packet has a receivelen field hanging off it,
4426 * update our state */
4427 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4430 /* If the ack packet has a "recommended" size that is less than
4431 * what I am using now, reduce my size to match */
4432 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4433 (int)sizeof(afs_int32), &tSize);
4434 tSize = (afs_uint32) ntohl(tSize);
4435 if (tSize > RX_MAX_PACKET_SIZE)
4436 tSize = RX_MAX_PACKET_SIZE;
4437 if (tSize < RX_MIN_PACKET_SIZE)
4438 tSize = RX_MIN_PACKET_SIZE;
4439 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4441 /* Get the maximum packet size to send to this peer */
4442 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4444 tSize = (afs_uint32) ntohl(tSize);
4445 if (tSize > RX_MAX_PACKET_SIZE)
4446 tSize = RX_MAX_PACKET_SIZE;
4447 if (tSize < RX_MIN_PACKET_SIZE)
4448 tSize = RX_MIN_PACKET_SIZE;
4449 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4450 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4452 /* sanity check - peer might have restarted with different params.
4453 * If peer says "send less", dammit, send less... Peer should never
4454 * be unable to accept packets of the size that prior AFS versions would
4455 * send without asking. */
4456 if (peer->maxMTU != tSize) {
4457 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4459 peer->maxMTU = tSize;
4460 peer->MTU = MIN(tSize, peer->MTU);
4461 call->MTU = MIN(call->MTU, tSize);
4464 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4467 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4468 (int)sizeof(afs_int32), &tSize);
4469 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4472 if (tSize >= rx_maxSendWindow)
4473 tSize = rx_maxSendWindow;
4474 if (tSize < call->twind) { /* smaller than our send */
4475 call->twind = tSize; /* window, we must send less... */
4476 call->ssthresh = MIN(call->twind, call->ssthresh);
4477 call->conn->twind[call->channel] = call->twind;
4480 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4481 * network MTU confused with the loopback MTU. Calculate the
4482 * maximum MTU here for use in the slow start code below.
4484 /* Did peer restart with older RX version? */
4485 if (peer->maxDgramPackets > 1) {
4486 peer->maxDgramPackets = 1;
4488 } else if (np->length >=
4489 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4492 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4493 sizeof(afs_int32), &tSize);
4494 tSize = (afs_uint32) ntohl(tSize);
4497 if (tSize >= rx_maxSendWindow)
4498 tSize = rx_maxSendWindow;
4500 * As of AFS 3.5 we set the send window to match the receive window.
4502 if (tSize < call->twind) {
4503 call->twind = tSize;
4504 call->conn->twind[call->channel] = call->twind;
4505 call->ssthresh = MIN(call->twind, call->ssthresh);
4506 } else if (tSize > call->twind) {
4507 call->twind = tSize;
4508 call->conn->twind[call->channel] = call->twind;
4512 * As of AFS 3.5, a jumbogram is more than one fixed size
4513 * packet transmitted in a single UDP datagram. If the remote
4514 * MTU is smaller than our local MTU then never send a datagram
4515 * larger than the natural MTU.
4518 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4519 (int)sizeof(afs_int32), &tSize);
4520 maxDgramPackets = (afs_uint32) ntohl(tSize);
4521 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4523 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4524 if (maxDgramPackets > 1) {
4525 peer->maxDgramPackets = maxDgramPackets;
4526 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4528 peer->maxDgramPackets = 1;
4529 call->MTU = peer->natMTU;
4531 } else if (peer->maxDgramPackets > 1) {
4532 /* Restarted with lower version of RX */
4533 peer->maxDgramPackets = 1;
4535 } else if (peer->maxDgramPackets > 1
4536 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4537 /* Restarted with lower version of RX */
4538 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4539 peer->natMTU = OLD_MAX_PACKET_SIZE;
4540 peer->MTU = OLD_MAX_PACKET_SIZE;
4541 peer->maxDgramPackets = 1;
4542 peer->nDgramPackets = 1;
4544 call->MTU = OLD_MAX_PACKET_SIZE;
4547 /* If the window has been extended by this acknowledge packet,
4548 * then wakeup a sender waiting in alloc for window space, or try
4549 * sending packets now, if he's been sitting on packets due to
4550 * lack of window space */
4551 if (call->tnext < (call->tfirst + call->twind)) {
4552 #ifdef RX_ENABLE_LOCKS
4553 CV_SIGNAL(&call->cv_twind);
4555 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4556 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4557 osi_rxWakeup(&call->twind);
4560 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4561 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4567 * Calculate how many datagrams were successfully received after
4568 * the first missing packet and adjust the negative ack counter
4573 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4574 if (call->nNacks < nNacked) {
4575 call->nNacks = nNacked;
4578 call->nAcks += newAckCount;
4582 /* If the packet contained new acknowledgements, rather than just
4583 * being a duplicate of one we have previously seen, then we can restart
4586 if (newAckCount > 0)
4587 rxi_rto_packet_acked(call, istack);
4589 if (call->flags & RX_CALL_FAST_RECOVER) {
4590 if (newAckCount == 0) {
4591 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4593 call->flags &= ~RX_CALL_FAST_RECOVER;
4594 call->cwind = call->nextCwind;
4595 call->nextCwind = 0;
4598 call->nCwindAcks = 0;
4599 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4600 /* Three negative acks in a row trigger congestion recovery */
4601 call->flags |= RX_CALL_FAST_RECOVER;
4602 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4604 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4605 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4606 call->nextCwind = call->ssthresh;
4609 peer->MTU = call->MTU;
4610 peer->cwind = call->nextCwind;
4611 peer->nDgramPackets = call->nDgramPackets;
4613 call->congestSeq = peer->congestSeq;
4615 /* Reset the resend times on the packets that were nacked
4616 * so we will retransmit as soon as the window permits
4620 for (opr_queue_ScanBackwards(&call->tq, cursor)) {
4621 struct rx_packet *tp =
4622 opr_queue_Entry(cursor, struct rx_packet, entry);
4624 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4625 tp->flags &= ~RX_PKTFLAG_SENT;
4627 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4632 /* If cwind is smaller than ssthresh, then increase
4633 * the window one packet for each ack we receive (exponential
4635 * If cwind is greater than or equal to ssthresh then increase
4636 * the congestion window by one packet for each cwind acks we
4637 * receive (linear growth). */
4638 if (call->cwind < call->ssthresh) {
4640 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4641 call->nCwindAcks = 0;
4643 call->nCwindAcks += newAckCount;
4644 if (call->nCwindAcks >= call->cwind) {
4645 call->nCwindAcks = 0;
4646 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4650 * If we have received several acknowledgements in a row then
4651 * it is time to increase the size of our datagrams
4653 if ((int)call->nAcks > rx_nDgramThreshold) {
4654 if (peer->maxDgramPackets > 1) {
4655 if (call->nDgramPackets < peer->maxDgramPackets) {
4656 call->nDgramPackets++;
4658 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4659 } else if (call->MTU < peer->maxMTU) {
4660 /* don't upgrade if we can't handle it */
4661 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4662 call->MTU = peer->ifMTU;
4664 call->MTU += peer->natMTU;
4665 call->MTU = MIN(call->MTU, peer->maxMTU);
4672 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4674 /* Servers need to hold the call until all response packets have
4675 * been acknowledged. Soft acks are good enough since clients
4676 * are not allowed to clear their receive queues. */
4677 if (call->state == RX_STATE_HOLD
4678 && call->tfirst + call->nSoftAcked >= call->tnext) {
4679 call->state = RX_STATE_DALLY;
4680 rxi_ClearTransmitQueue(call, 0);
4681 rxi_CancelKeepAliveEvent(call);
4682 } else if (!opr_queue_IsEmpty(&call->tq)) {
4683 rxi_Start(call, istack);
4689 * Schedule a connection abort to be sent after some delay.
4691 * @param[in] conn The connection to send the abort on.
4692 * @param[in] msec The number of milliseconds to wait before sending.
4694 * @pre conn_data_lock must be held
4697 rxi_SendConnectionAbortLater(struct rx_connection *conn, int msec)
4699 struct clock when, now;
4701 MUTEX_ASSERT(&conn->conn_data_lock);
4705 if (!conn->delayedAbortEvent) {
4706 clock_GetTime(&now);
4708 clock_Addmsec(&when, msec);
4709 rx_GetConnection(conn);
4710 conn->delayedAbortEvent =
4711 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
4715 /* Received a response to a challenge packet */
4716 static struct rx_packet *
4717 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4718 struct rx_packet *np, int istack)
4722 /* Ignore the packet if we're the client */
4723 if (conn->type == RX_CLIENT_CONNECTION)
4726 /* If already authenticated, ignore the packet (it's probably a retry) */
4727 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4730 if (!conn->securityChallengeSent) {
4731 /* We've never sent out a challenge for this connection, so this
4732 * response cannot possibly be correct; ignore it. This can happen
4733 * if we sent a challenge to the client, then we were restarted, and
4734 * then the client sent us a response. If we ignore the response, the
4735 * client will eventually resend a data packet, causing us to send a
4736 * new challenge and the client to send a new response. */
4740 /* Otherwise, have the security object evaluate the response packet */
4741 error = RXS_CheckResponse(conn->securityObject, conn, np);
4743 /* If the response is invalid, reset the connection, sending
4744 * an abort to the peer. Send the abort with a 1 second delay,
4745 * to avoid a peer hammering us by constantly recreating a
4746 * connection with bad credentials. */
4747 rxi_ConnectionError(conn, error);
4748 MUTEX_ENTER(&conn->conn_data_lock);
4749 rxi_SendConnectionAbortLater(conn, 1000);
4750 MUTEX_EXIT(&conn->conn_data_lock);
4753 /* If the response is valid, any calls waiting to attach
4754 * servers can now do so */
4757 for (i = 0; i < RX_MAXCALLS; i++) {
4758 struct rx_call *call = conn->call[i];
4760 MUTEX_ENTER(&call->lock);
4761 if (call->state == RX_STATE_PRECALL)
4762 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4763 /* tnop can be null if newcallp is null */
4764 MUTEX_EXIT(&call->lock);
4768 /* Update the peer reachability information, just in case
4769 * some calls went into attach-wait while we were waiting
4770 * for authentication..
4772 rxi_UpdatePeerReach(conn, NULL);
4777 /* A client has received an authentication challenge: the security
4778 * object is asked to cough up a respectable response packet to send
4779 * back to the server. The server is responsible for retrying the
4780 * challenge if it fails to get a response. */
4782 static struct rx_packet *
4783 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4784 struct rx_packet *np, int istack)
4788 /* Ignore the challenge if we're the server */
4789 if (conn->type == RX_SERVER_CONNECTION)
4792 /* Ignore the challenge if the connection is otherwise idle; someone's
4793 * trying to use us as an oracle. */
4794 if (!rxi_HasActiveCalls(conn))
4797 /* Send the security object the challenge packet. It is expected to fill
4798 * in the response. */
4799 error = RXS_GetResponse(conn->securityObject, conn, np);
4801 /* If the security object is unable to return a valid response, reset the
4802 * connection and send an abort to the peer. Otherwise send the response
4803 * packet to the peer connection. */
4805 rxi_ConnectionError(conn, error);
4806 MUTEX_ENTER(&conn->conn_data_lock);
4807 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4808 MUTEX_EXIT(&conn->conn_data_lock);
4810 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4811 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4817 /* Find an available server process to service the current request in
4818 * the given call structure. If one isn't available, queue up this
4819 * call so it eventually gets one */
4821 rxi_AttachServerProc(struct rx_call *call,
4822 osi_socket socket, int *tnop,
4823 struct rx_call **newcallp)
4825 struct rx_serverQueueEntry *sq;
4826 struct rx_service *service = call->conn->service;
4829 /* May already be attached */
4830 if (call->state == RX_STATE_ACTIVE)
4833 MUTEX_ENTER(&rx_serverPool_lock);
4835 haveQuota = QuotaOK(service);
4836 if ((!haveQuota) || opr_queue_IsEmpty(&rx_idleServerQueue)) {
4837 /* If there are no processes available to service this call,
4838 * put the call on the incoming call queue (unless it's
4839 * already on the queue).
4841 #ifdef RX_ENABLE_LOCKS
4843 ReturnToServerPool(service);
4844 #endif /* RX_ENABLE_LOCKS */
4846 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4847 call->flags |= RX_CALL_WAIT_PROC;
4848 rx_atomic_inc(&rx_nWaiting);
4849 rx_atomic_inc(&rx_nWaited);
4850 rxi_calltrace(RX_CALL_ARRIVAL, call);
4851 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4852 opr_queue_Append(&rx_incomingCallQueue, &call->entry);
4855 sq = opr_queue_Last(&rx_idleServerQueue,
4856 struct rx_serverQueueEntry, entry);
4858 /* If hot threads are enabled, and both newcallp and sq->socketp
4859 * are non-null, then this thread will process the call, and the
4860 * idle server thread will start listening on this threads socket.
4862 opr_queue_Remove(&sq->entry);
4864 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4867 *sq->socketp = socket;
4868 clock_GetTime(&call->startTime);
4869 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4873 if (call->flags & RX_CALL_WAIT_PROC) {
4874 /* Conservative: I don't think this should happen */
4875 call->flags &= ~RX_CALL_WAIT_PROC;
4876 rx_atomic_dec(&rx_nWaiting);
4877 if (opr_queue_IsOnQueue(&call->entry)) {
4878 opr_queue_Remove(&call->entry);
4881 call->state = RX_STATE_ACTIVE;
4882 call->app.mode = RX_MODE_RECEIVING;
4883 #ifdef RX_KERNEL_TRACE
4885 int glockOwner = ISAFS_GLOCK();
4888 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4889 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4895 if (call->flags & RX_CALL_CLEARED) {
4896 /* send an ack now to start the packet flow up again */
4897 call->flags &= ~RX_CALL_CLEARED;
4898 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4900 #ifdef RX_ENABLE_LOCKS
4903 service->nRequestsRunning++;
4904 MUTEX_ENTER(&rx_quota_mutex);
4905 if (service->nRequestsRunning <= service->minProcs)
4908 MUTEX_EXIT(&rx_quota_mutex);
4912 MUTEX_EXIT(&rx_serverPool_lock);
4915 /* Delay the sending of an acknowledge event for a short while, while
4916 * a new call is being prepared (in the case of a client) or a reply
4917 * is being prepared (in the case of a server). Rather than sending
4918 * an ack packet, an ACKALL packet is sent. */
4920 rxi_AckAll(struct rx_call *call)
4922 rxi_SendSpecial(call, call->conn, NULL, RX_PACKET_TYPE_ACKALL,
4924 call->flags |= RX_CALL_ACKALL_SENT;
4928 * Event handler for per-call delayed acks.
4929 * Also called synchronously, with |event| == NULL, to send a "delayed" ack
4933 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4936 struct rx_call *call = arg1;
4937 #ifdef RX_ENABLE_LOCKS
4939 MUTEX_ENTER(&call->lock);
4940 if (event == call->delayedAckEvent)
4941 rxevent_Put(&call->delayedAckEvent);
4943 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4945 MUTEX_EXIT(&call->lock);
4946 #else /* RX_ENABLE_LOCKS */
4948 rxevent_Put(&call->delayedAckEvent);
4949 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4950 #endif /* RX_ENABLE_LOCKS */
4951 /* Release the call reference for the event that fired. */
4953 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4956 #ifdef RX_ENABLE_LOCKS
4957 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4958 * clearing them out.
4961 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4963 struct opr_queue *cursor;
4966 for (opr_queue_Scan(&call->tq, cursor)) {
4968 = opr_queue_Entry(cursor, struct rx_packet, entry);
4970 p->flags |= RX_PKTFLAG_ACKED;
4975 call->flags |= RX_CALL_TQ_CLEARME;
4976 call->flags |= RX_CALL_TQ_SOME_ACKED;
4979 rxi_rto_cancel(call);
4981 call->tfirst = call->tnext;
4982 call->nSoftAcked = 0;
4984 if (call->flags & RX_CALL_FAST_RECOVER) {
4985 call->flags &= ~RX_CALL_FAST_RECOVER;
4986 call->cwind = call->nextCwind;
4987 call->nextCwind = 0;
4990 CV_SIGNAL(&call->cv_twind);
4992 #endif /* RX_ENABLE_LOCKS */
4995 * Acknowledge the whole transmit queue.
4997 * If we're running without locks, or the transmit queue isn't busy, then
4998 * we can just clear the queue now. Otherwise, we have to mark all of the
4999 * packets as acknowledged, and let rxi_Start clear it later on
5002 rxi_AckAllInTransmitQueue(struct rx_call *call)
5004 #ifdef RX_ENABLE_LOCKS
5005 if (call->flags & RX_CALL_TQ_BUSY) {
5006 rxi_SetAcksInTransmitQueue(call);
5010 rxi_ClearTransmitQueue(call, 0);
5012 /* Clear out the transmit queue for the current call (all packets have
5013 * been received by peer) */
5015 rxi_ClearTransmitQueue(struct rx_call *call, int force)
5017 #ifdef RX_ENABLE_LOCKS
5018 struct opr_queue *cursor;
5019 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
5021 for (opr_queue_Scan(&call->tq, cursor)) {
5023 = opr_queue_Entry(cursor, struct rx_packet, entry);
5025 p->flags |= RX_PKTFLAG_ACKED;
5029 call->flags |= RX_CALL_TQ_CLEARME;
5030 call->flags |= RX_CALL_TQ_SOME_ACKED;
5033 #endif /* RX_ENABLE_LOCKS */
5034 #ifdef RXDEBUG_PACKET
5036 #endif /* RXDEBUG_PACKET */
5037 rxi_FreePackets(0, &call->tq);
5038 rxi_WakeUpTransmitQueue(call);
5039 #ifdef RX_ENABLE_LOCKS
5040 call->flags &= ~RX_CALL_TQ_CLEARME;
5044 rxi_rto_cancel(call);
5045 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5046 call->nSoftAcked = 0;
5048 if (call->flags & RX_CALL_FAST_RECOVER) {
5049 call->flags &= ~RX_CALL_FAST_RECOVER;
5050 call->cwind = call->nextCwind;
5052 #ifdef RX_ENABLE_LOCKS
5053 CV_SIGNAL(&call->cv_twind);
5055 osi_rxWakeup(&call->twind);
5060 rxi_ClearReceiveQueue(struct rx_call *call)
5062 if (!opr_queue_IsEmpty(&call->rq)) {
5065 count = rxi_FreePackets(0, &call->rq);
5066 rx_packetReclaims += count;
5067 #ifdef RXDEBUG_PACKET
5069 if ( call->rqc != 0 )
5070 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5072 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5074 if (call->state == RX_STATE_PRECALL) {
5075 call->flags |= RX_CALL_CLEARED;
5079 /* Send an abort packet for the specified call */
5080 static struct rx_packet *
5081 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5082 int istack, int force)
5085 struct clock when, now;
5090 /* Clients should never delay abort messages */
5091 if (rx_IsClientConn(call->conn))
5095 * An opcode that has been deprecated or has yet to be implemented is not
5096 * a misbehavior of the client. Do not punish the client by introducing
5099 if (call->error == RXGEN_OPCODE) {
5101 } else if (call->abortCode != call->error) {
5102 call->abortCode = call->error;
5103 call->abortCount = 0;
5106 if (force || rxi_callAbortThreshhold == 0
5107 || call->abortCount < rxi_callAbortThreshhold) {
5108 rxi_CancelDelayedAbortEvent(call);
5109 error = htonl(call->error);
5113 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5114 (char *)&error, sizeof(error), istack);
5115 } else if (!call->delayedAbortEvent) {
5116 clock_GetTime(&now);
5118 clock_Addmsec(&when, rxi_callAbortDelay);
5119 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5120 call->delayedAbortEvent =
5121 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5127 rxi_CancelDelayedAbortEvent(struct rx_call *call)
5129 MUTEX_ASSERT(&call->lock);
5130 if (rxevent_Cancel(&call->delayedAbortEvent))
5131 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5134 /* Send an abort packet for the specified connection. Packet is an
5135 * optional pointer to a packet that can be used to send the abort.
5136 * Once the number of abort messages reaches the threshhold, an
5137 * event is scheduled to send the abort. Setting the force flag
5138 * overrides sending delayed abort messages.
5140 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5141 * to send the abort packet.
5144 rxi_SendConnectionAbort(struct rx_connection *conn,
5145 struct rx_packet *packet, int istack, int force)
5152 /* Clients should never delay abort messages */
5153 if (rx_IsClientConn(conn))
5156 if (force || rxi_connAbortThreshhold == 0
5157 || conn->abortCount < rxi_connAbortThreshhold) {
5159 if (rxevent_Cancel(&conn->delayedAbortEvent))
5160 putConnection(conn);
5161 error = htonl(conn->error);
5163 MUTEX_EXIT(&conn->conn_data_lock);
5165 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5166 RX_PACKET_TYPE_ABORT, (char *)&error,
5167 sizeof(error), istack);
5168 MUTEX_ENTER(&conn->conn_data_lock);
5170 rxi_SendConnectionAbortLater(conn, rxi_connAbortDelay);
5175 /* Associate an error all of the calls owned by a connection. Called
5176 * with error non-zero. This is only for really fatal things, like
5177 * bad authentication responses. The connection itself is set in
5178 * error at this point, so that future packets received will be
5181 rxi_ConnectionError(struct rx_connection *conn,
5187 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5189 MUTEX_ENTER(&conn->conn_data_lock);
5190 if (rxevent_Cancel(&conn->challengeEvent))
5191 putConnection(conn);
5192 if (rxevent_Cancel(&conn->natKeepAliveEvent))
5193 putConnection(conn);
5194 if (rxevent_Cancel(&conn->checkReachEvent)) {
5195 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5196 putConnection(conn);
5198 MUTEX_EXIT(&conn->conn_data_lock);
5199 for (i = 0; i < RX_MAXCALLS; i++) {
5200 struct rx_call *call = conn->call[i];
5202 MUTEX_ENTER(&call->lock);
5203 rxi_CallError(call, error);
5204 MUTEX_EXIT(&call->lock);
5207 conn->error = error;
5208 if (rx_stats_active)
5209 rx_atomic_inc(&rx_stats.fatalErrors);
5214 * Interrupt an in-progress call with the specified error and wakeup waiters.
5216 * @param[in] call The call to interrupt
5217 * @param[in] error The error code to send to the peer
5220 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5222 MUTEX_ENTER(&call->lock);
5223 rxi_CallError(call, error);
5224 rxi_SendCallAbort(call, NULL, 0, 1);
5225 MUTEX_EXIT(&call->lock);
5229 rxi_CallError(struct rx_call *call, afs_int32 error)
5231 MUTEX_ASSERT(&call->lock);
5232 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5234 error = call->error;
5236 #ifdef RX_ENABLE_LOCKS
5237 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5238 rxi_ResetCall(call, 0);
5241 rxi_ResetCall(call, 0);
5243 call->error = error;
5246 /* Reset various fields in a call structure, and wakeup waiting
5247 * processes. Some fields aren't changed: state & mode are not
5248 * touched (these must be set by the caller), and bufptr, nLeft, and
5249 * nFree are not reset, since these fields are manipulated by
5250 * unprotected macros, and may only be reset by non-interrupting code.
5254 rxi_ResetCall(struct rx_call *call, int newcall)
5257 struct rx_peer *peer;
5258 struct rx_packet *packet;
5260 MUTEX_ASSERT(&call->lock);
5261 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5263 /* Notify anyone who is waiting for asynchronous packet arrival */
5264 if (call->arrivalProc) {
5265 (*call->arrivalProc) (call, call->arrivalProcHandle,
5266 call->arrivalProcArg);
5267 call->arrivalProc = (void (*)())0;
5271 rxi_CancelGrowMTUEvent(call);
5273 if (call->delayedAbortEvent) {
5274 rxi_CancelDelayedAbortEvent(call);
5275 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5277 rxi_SendCallAbort(call, packet, 0, 1);
5278 rxi_FreePacket(packet);
5283 * Update the peer with the congestion information in this call
5284 * so other calls on this connection can pick up where this call
5285 * left off. If the congestion sequence numbers don't match then
5286 * another call experienced a retransmission.
5288 peer = call->conn->peer;
5289 MUTEX_ENTER(&peer->peer_lock);
5291 if (call->congestSeq == peer->congestSeq) {
5292 peer->cwind = MAX(peer->cwind, call->cwind);
5293 peer->MTU = MAX(peer->MTU, call->MTU);
5294 peer->nDgramPackets =
5295 MAX(peer->nDgramPackets, call->nDgramPackets);
5298 call->abortCode = 0;
5299 call->abortCount = 0;
5301 if (peer->maxDgramPackets > 1) {
5302 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5304 call->MTU = peer->MTU;
5306 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5307 call->ssthresh = rx_maxSendWindow;
5308 call->nDgramPackets = peer->nDgramPackets;
5309 call->congestSeq = peer->congestSeq;
5310 call->rtt = peer->rtt;
5311 call->rtt_dev = peer->rtt_dev;
5312 clock_Zero(&call->rto);
5313 clock_Addmsec(&call->rto,
5314 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5315 MUTEX_EXIT(&peer->peer_lock);
5317 flags = call->flags;
5318 rxi_WaitforTQBusy(call);
5320 rxi_ClearTransmitQueue(call, 1);
5321 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5322 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5326 rxi_ClearReceiveQueue(call);
5327 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5331 call->twind = call->conn->twind[call->channel];
5332 call->rwind = call->conn->rwind[call->channel];
5333 call->nSoftAcked = 0;
5334 call->nextCwind = 0;
5337 call->nCwindAcks = 0;
5338 call->nSoftAcks = 0;
5339 call->nHardAcks = 0;
5341 call->tfirst = call->rnext = call->tnext = 1;
5344 call->lastAcked = 0;
5345 call->localStatus = call->remoteStatus = 0;
5347 if (flags & RX_CALL_READER_WAIT) {
5348 #ifdef RX_ENABLE_LOCKS
5349 CV_BROADCAST(&call->cv_rq);
5351 osi_rxWakeup(&call->rq);
5354 if (flags & RX_CALL_WAIT_PACKETS) {
5355 MUTEX_ENTER(&rx_freePktQ_lock);
5356 rxi_PacketsUnWait(); /* XXX */
5357 MUTEX_EXIT(&rx_freePktQ_lock);
5359 #ifdef RX_ENABLE_LOCKS
5360 CV_SIGNAL(&call->cv_twind);
5362 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5363 osi_rxWakeup(&call->twind);
5366 if (flags & RX_CALL_WAIT_PROC) {
5367 rx_atomic_dec(&rx_nWaiting);
5369 #ifdef RX_ENABLE_LOCKS
5370 /* The following ensures that we don't mess with any queue while some
5371 * other thread might also be doing so. The call_queue_lock field is
5372 * is only modified under the call lock. If the call is in the process
5373 * of being removed from a queue, the call is not locked until the
5374 * the queue lock is dropped and only then is the call_queue_lock field
5375 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5376 * Note that any other routine which removes a call from a queue has to
5377 * obtain the queue lock before examing the queue and removing the call.
5379 if (call->call_queue_lock) {
5380 MUTEX_ENTER(call->call_queue_lock);
5381 if (opr_queue_IsOnQueue(&call->entry)) {
5382 opr_queue_Remove(&call->entry);
5384 MUTEX_EXIT(call->call_queue_lock);
5385 CLEAR_CALL_QUEUE_LOCK(call);
5387 #else /* RX_ENABLE_LOCKS */
5388 if (opr_queue_IsOnQueue(&call->entry)) {
5389 opr_queue_Remove(&call->entry);
5391 #endif /* RX_ENABLE_LOCKS */
5393 rxi_CancelKeepAliveEvent(call);
5394 rxi_CancelDelayedAckEvent(call);
5397 /* Send an acknowledge for the indicated packet (seq,serial) of the
5398 * indicated call, for the indicated reason (reason). This
5399 * acknowledge will specifically acknowledge receiving the packet, and
5400 * will also specify which other packets for this call have been
5401 * received. This routine returns the packet that was used to the
5402 * caller. The caller is responsible for freeing it or re-using it.
5403 * This acknowledgement also returns the highest sequence number
5404 * actually read out by the higher level to the sender; the sender
5405 * promises to keep around packets that have not been read by the
5406 * higher level yet (unless, of course, the sender decides to abort
5407 * the call altogether). Any of p, seq, serial, pflags, or reason may
5408 * be set to zero without ill effect. That is, if they are zero, they
5409 * will not convey any information.
5410 * NOW there is a trailer field, after the ack where it will safely be
5411 * ignored by mundanes, which indicates the maximum size packet this
5412 * host can swallow. */
5414 struct rx_packet *optionalPacket; use to send ack (or null)
5415 int seq; Sequence number of the packet we are acking
5416 int serial; Serial number of the packet
5417 int pflags; Flags field from packet header
5418 int reason; Reason an acknowledge was prompted
5421 #define RX_ZEROS 1024
5422 static char rx_zeros[RX_ZEROS];
5425 rxi_SendAck(struct rx_call *call,
5426 struct rx_packet *optionalPacket, int serial, int reason,
5429 struct rx_ackPacket *ap;
5430 struct rx_packet *p;
5431 struct opr_queue *cursor;
5434 afs_uint32 padbytes = 0;
5435 #ifdef RX_ENABLE_TSFPQ
5436 struct rx_ts_info_t * rx_ts_info;
5440 * Open the receive window once a thread starts reading packets
5442 if (call->rnext > 1) {
5443 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5446 /* Don't attempt to grow MTU if this is a critical ping */
5447 if (reason == RX_ACK_MTU) {
5448 /* keep track of per-call attempts, if we're over max, do in small
5449 * otherwise in larger? set a size to increment by, decrease
5452 if (call->conn->peer->maxPacketSize &&
5453 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5455 padbytes = call->conn->peer->maxPacketSize+16;
5457 padbytes = call->conn->peer->maxMTU + 128;
5459 /* do always try a minimum size ping */
5460 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5462 /* subtract the ack payload */
5463 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5464 reason = RX_ACK_PING;
5467 call->nHardAcks = 0;
5468 call->nSoftAcks = 0;
5469 if (call->rnext > call->lastAcked)
5470 call->lastAcked = call->rnext;
5474 rx_computelen(p, p->length); /* reset length, you never know */
5475 } /* where that's been... */
5476 #ifdef RX_ENABLE_TSFPQ
5478 RX_TS_INFO_GET(rx_ts_info);
5479 if ((p = rx_ts_info->local_special_packet)) {
5480 rx_computelen(p, p->length);
5481 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5482 rx_ts_info->local_special_packet = p;
5483 } else { /* We won't send the ack, but don't panic. */
5484 return optionalPacket;
5488 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5489 /* We won't send the ack, but don't panic. */
5490 return optionalPacket;
5495 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5498 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5499 #ifndef RX_ENABLE_TSFPQ
5500 if (!optionalPacket)
5503 return optionalPacket;
5505 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5506 if (rx_Contiguous(p) < templ) {
5507 #ifndef RX_ENABLE_TSFPQ
5508 if (!optionalPacket)
5511 return optionalPacket;
5516 /* MTUXXX failing to send an ack is very serious. We should */
5517 /* try as hard as possible to send even a partial ack; it's */
5518 /* better than nothing. */
5519 ap = (struct rx_ackPacket *)rx_DataOf(p);
5520 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5521 ap->reason = reason;
5523 /* The skew computation used to be bogus, I think it's better now. */
5524 /* We should start paying attention to skew. XXX */
5525 ap->serial = htonl(serial);
5526 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5529 * First packet not yet forwarded to reader. When ACKALL has been
5530 * sent the peer has been told that all received packets will be
5531 * delivered to the reader. The value 'rnext' is used internally
5532 * to refer to the next packet in the receive queue that must be
5533 * delivered to the reader. From the perspective of the peer it
5534 * already has so report the last sequence number plus one if there
5535 * are packets in the receive queue awaiting processing.
5537 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5538 !opr_queue_IsEmpty(&call->rq)) {
5539 ap->firstPacket = htonl(opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq + 1);
5541 ap->firstPacket = htonl(call->rnext);
5543 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5545 /* No fear of running out of ack packet here because there can only
5546 * be at most one window full of unacknowledged packets. The window
5547 * size must be constrained to be less than the maximum ack size,
5548 * of course. Also, an ack should always fit into a single packet
5549 * -- it should not ever be fragmented. */
5551 for (opr_queue_Scan(&call->rq, cursor)) {
5552 struct rx_packet *rqp
5553 = opr_queue_Entry(cursor, struct rx_packet, entry);
5555 if (!rqp || !call->rq.next
5556 || (rqp->header.seq > (call->rnext + call->rwind))) {
5557 #ifndef RX_ENABLE_TSFPQ
5558 if (!optionalPacket)
5561 rxi_CallError(call, RX_CALL_DEAD);
5562 return optionalPacket;
5565 while (rqp->header.seq > call->rnext + offset)
5566 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5567 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5569 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5570 #ifndef RX_ENABLE_TSFPQ
5571 if (!optionalPacket)
5574 rxi_CallError(call, RX_CALL_DEAD);
5575 return optionalPacket;
5581 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5583 /* Must zero the 3 octets that rx_AckDataSize skips at the end of the
5586 rx_packetwrite(p, rx_AckDataSize(offset) - 3, 3, rx_zeros);
5588 /* these are new for AFS 3.3 */
5589 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5590 templ = htonl(templ);
5591 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5592 templ = htonl(call->conn->peer->ifMTU);
5593 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5594 sizeof(afs_int32), &templ);
5596 /* new for AFS 3.4 */
5597 templ = htonl(call->rwind);
5598 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5599 sizeof(afs_int32), &templ);
5601 /* new for AFS 3.5 */
5602 templ = htonl(call->conn->peer->ifDgramPackets);
5603 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5604 sizeof(afs_int32), &templ);
5606 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5608 p->header.serviceId = call->conn->serviceId;
5609 p->header.cid = (call->conn->cid | call->channel);
5610 p->header.callNumber = *call->callNumber;
5612 p->header.securityIndex = call->conn->securityIndex;
5613 p->header.epoch = call->conn->epoch;
5614 p->header.type = RX_PACKET_TYPE_ACK;
5615 p->header.flags = RX_SLOW_START_OK;
5616 if (reason == RX_ACK_PING)
5617 p->header.flags |= RX_REQUEST_ACK;
5619 while (padbytes > 0) {
5620 if (padbytes > RX_ZEROS) {
5621 rx_packetwrite(p, p->length, RX_ZEROS, rx_zeros);
5622 p->length += RX_ZEROS;
5623 padbytes -= RX_ZEROS;
5625 rx_packetwrite(p, p->length, padbytes, rx_zeros);
5626 p->length += padbytes;
5631 if (call->conn->type == RX_CLIENT_CONNECTION)
5632 p->header.flags |= RX_CLIENT_INITIATED;
5636 if (rxdebug_active) {
5640 len = _snprintf(msg, sizeof(msg),
5641 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5642 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5643 ntohl(ap->serial), ntohl(ap->previousPacket),
5644 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5645 ap->nAcks, ntohs(ap->bufferSpace) );
5649 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5650 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5654 OutputDebugString(msg);
5656 #else /* AFS_NT40_ENV */
5658 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5659 ap->reason, ntohl(ap->previousPacket),
5660 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5662 for (offset = 0; offset < ap->nAcks; offset++)
5663 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5668 #endif /* AFS_NT40_ENV */
5671 int i, nbytes = p->length;
5673 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5674 if (nbytes <= p->wirevec[i].iov_len) {
5677 savelen = p->wirevec[i].iov_len;
5679 p->wirevec[i].iov_len = nbytes;
5681 rxi_Send(call, p, istack);
5682 p->wirevec[i].iov_len = savelen;
5686 nbytes -= p->wirevec[i].iov_len;
5689 if (rx_stats_active)
5690 rx_atomic_inc(&rx_stats.ackPacketsSent);
5691 #ifndef RX_ENABLE_TSFPQ
5692 if (!optionalPacket)
5695 return optionalPacket; /* Return packet for re-use by caller */
5699 struct rx_packet **list;
5704 /* Send all of the packets in the list in single datagram */
5706 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5707 int istack, int moreFlag)
5713 struct rx_connection *conn = call->conn;
5714 struct rx_peer *peer = conn->peer;
5716 MUTEX_ENTER(&peer->peer_lock);
5717 peer->nSent += xmit->len;
5718 if (xmit->resending)
5719 peer->reSends += xmit->len;
5720 MUTEX_EXIT(&peer->peer_lock);
5722 if (rx_stats_active) {
5723 if (xmit->resending)
5724 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5726 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5729 clock_GetTime(&now);
5731 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5735 /* Set the packet flags and schedule the resend events */
5736 /* Only request an ack for the last packet in the list */
5737 for (i = 0; i < xmit->len; i++) {
5738 struct rx_packet *packet = xmit->list[i];
5740 /* Record the time sent */
5741 packet->timeSent = now;
5742 packet->flags |= RX_PKTFLAG_SENT;
5744 /* Ask for an ack on retransmitted packets, on every other packet
5745 * if the peer doesn't support slow start. Ask for an ack on every
5746 * packet until the congestion window reaches the ack rate. */
5747 if (packet->header.serial) {
5750 packet->firstSent = now;
5751 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5752 || (!(call->flags & RX_CALL_SLOW_START_OK)
5753 && (packet->header.seq & 1)))) {
5758 /* Tag this packet as not being the last in this group,
5759 * for the receiver's benefit */
5760 if (i < xmit->len - 1 || moreFlag) {
5761 packet->header.flags |= RX_MORE_PACKETS;
5766 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5769 /* Since we're about to send a data packet to the peer, it's
5770 * safe to nuke any scheduled end-of-packets ack */
5771 rxi_CancelDelayedAckEvent(call);
5773 MUTEX_EXIT(&call->lock);
5774 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5775 if (xmit->len > 1) {
5776 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5778 rxi_SendPacket(call, conn, xmit->list[0], istack);
5780 MUTEX_ENTER(&call->lock);
5781 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5783 /* Tell the RTO calculation engine that we have sent a packet, and
5784 * if it was the last one */
5785 rxi_rto_packet_sent(call, lastPacket, istack);
5787 /* Update last send time for this call (for keep-alive
5788 * processing), and for the connection (so that we can discover
5789 * idle connections) */
5790 conn->lastSendTime = call->lastSendTime = clock_Sec();
5793 /* When sending packets we need to follow these rules:
5794 * 1. Never send more than maxDgramPackets in a jumbogram.
5795 * 2. Never send a packet with more than two iovecs in a jumbogram.
5796 * 3. Never send a retransmitted packet in a jumbogram.
5797 * 4. Never send more than cwind/4 packets in a jumbogram
5798 * We always keep the last list we should have sent so we
5799 * can set the RX_MORE_PACKETS flags correctly.
5803 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5808 struct xmitlist working;
5809 struct xmitlist last;
5811 struct rx_peer *peer = call->conn->peer;
5812 int morePackets = 0;
5814 memset(&last, 0, sizeof(struct xmitlist));
5815 working.list = &list[0];
5817 working.resending = 0;
5819 recovery = call->flags & RX_CALL_FAST_RECOVER;
5821 for (i = 0; i < len; i++) {
5822 /* Does the current packet force us to flush the current list? */
5824 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5825 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5827 /* This sends the 'last' list and then rolls the current working
5828 * set into the 'last' one, and resets the working set */
5831 rxi_SendList(call, &last, istack, 1);
5832 /* If the call enters an error state stop sending, or if
5833 * we entered congestion recovery mode, stop sending */
5835 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5840 working.resending = 0;
5841 working.list = &list[i];
5843 /* Add the current packet to the list if it hasn't been acked.
5844 * Otherwise adjust the list pointer to skip the current packet. */
5845 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5848 if (list[i]->header.serial)
5849 working.resending = 1;
5851 /* Do we need to flush the list? */
5852 if (working.len >= (int)peer->maxDgramPackets
5853 || working.len >= (int)call->nDgramPackets
5854 || working.len >= (int)call->cwind
5855 || list[i]->header.serial
5856 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5858 rxi_SendList(call, &last, istack, 1);
5859 /* If the call enters an error state stop sending, or if
5860 * we entered congestion recovery mode, stop sending */
5862 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5867 working.resending = 0;
5868 working.list = &list[i + 1];
5871 if (working.len != 0) {
5872 osi_Panic("rxi_SendList error");
5874 working.list = &list[i + 1];
5878 /* Send the whole list when the call is in receive mode, when
5879 * the call is in eof mode, when we are in fast recovery mode,
5880 * and when we have the last packet */
5881 /* XXX - The accesses to app.mode aren't safe, as this may be called by
5882 * the listener or event threads
5884 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5885 || (call->flags & RX_CALL_FLUSH)
5886 || (call->flags & RX_CALL_FAST_RECOVER)) {
5887 /* Check for the case where the current list contains
5888 * an acked packet. Since we always send retransmissions
5889 * in a separate packet, we only need to check the first
5890 * packet in the list */
5891 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5895 rxi_SendList(call, &last, istack, morePackets);
5896 /* If the call enters an error state stop sending, or if
5897 * we entered congestion recovery mode, stop sending */
5899 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5903 rxi_SendList(call, &working, istack, 0);
5905 } else if (last.len > 0) {
5906 rxi_SendList(call, &last, istack, 0);
5907 /* Packets which are in 'working' are not sent by this call */
5912 * Check if the peer for the given call is known to be dead
5914 * If the call's peer appears dead (it has encountered fatal network errors
5915 * since the call started) the call is killed with RX_CALL_DEAD if the call
5916 * is active. Otherwise, we do nothing.
5918 * @param[in] call The call to check
5921 * @retval 0 The call is fine, and we haven't done anything to the call
5922 * @retval nonzero The call's peer appears dead, and the call has been
5923 * terminated if it was active
5925 * @pre call->lock must be locked
5928 rxi_CheckPeerDead(struct rx_call *call)
5930 #ifdef AFS_RXERRQ_ENV
5933 if (call->state == RX_STATE_DALLY) {
5937 peererrs = rx_atomic_read(&call->conn->peer->neterrs);
5938 if (call->neterr_gen < peererrs) {
5939 /* we have received network errors since this call started; kill
5941 if (call->state == RX_STATE_ACTIVE) {
5942 rxi_CallError(call, RX_CALL_DEAD);
5946 if (call->neterr_gen > peererrs) {
5947 /* someone has reset the number of peer errors; set the call error gen
5948 * so we can detect if more errors are encountered */
5949 call->neterr_gen = peererrs;
5956 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
5958 struct rx_call *call = arg0;
5959 struct rx_peer *peer;
5960 struct opr_queue *cursor;
5961 struct clock maxTimeout = { 60, 0 };
5963 MUTEX_ENTER(&call->lock);
5965 peer = call->conn->peer;
5967 /* Make sure that the event pointer is removed from the call
5968 * structure, since there is no longer a per-call retransmission
5970 if (event == call->resendEvent)
5971 rxevent_Put(&call->resendEvent);
5973 rxi_CheckPeerDead(call);
5975 if (opr_queue_IsEmpty(&call->tq)) {
5976 /* Nothing to do. This means that we've been raced, and that an
5977 * ACK has come in between when we were triggered, and when we
5978 * actually got to run. */
5982 /* We're in loss recovery */
5983 call->flags |= RX_CALL_FAST_RECOVER;
5985 /* Mark all of the pending packets in the queue as being lost */
5986 for (opr_queue_Scan(&call->tq, cursor)) {
5987 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
5988 if (!(p->flags & RX_PKTFLAG_ACKED))
5989 p->flags &= ~RX_PKTFLAG_SENT;
5992 /* We're resending, so we double the timeout of the call. This will be
5993 * dropped back down by the first successful ACK that we receive.
5995 * We apply a maximum value here of 60 seconds
5997 clock_Add(&call->rto, &call->rto);
5998 if (clock_Gt(&call->rto, &maxTimeout))
5999 call->rto = maxTimeout;
6001 /* Packet loss is most likely due to congestion, so drop our window size
6002 * and start again from the beginning */
6003 if (peer->maxDgramPackets >1) {
6004 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
6005 call->MTU = MIN(peer->natMTU, peer->maxMTU);
6007 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
6008 call->nDgramPackets = 1;
6010 call->nextCwind = 1;
6013 MUTEX_ENTER(&peer->peer_lock);
6014 peer->MTU = call->MTU;
6015 peer->cwind = call->cwind;
6016 peer->nDgramPackets = 1;
6018 call->congestSeq = peer->congestSeq;
6019 MUTEX_EXIT(&peer->peer_lock);
6021 rxi_Start(call, istack);
6024 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
6025 MUTEX_EXIT(&call->lock);
6028 /* This routine is called when new packets are readied for
6029 * transmission and when retransmission may be necessary, or when the
6030 * transmission window or burst count are favourable. This should be
6031 * better optimized for new packets, the usual case, now that we've
6032 * got rid of queues of send packets. XXXXXXXXXXX */
6034 rxi_Start(struct rx_call *call, int istack)
6036 struct opr_queue *cursor;
6037 #ifdef RX_ENABLE_LOCKS
6038 struct opr_queue *store;
6044 #ifdef RX_ENABLE_LOCKS
6045 if (rx_stats_active)
6046 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
6051 if (!opr_queue_IsEmpty(&call->tq)) { /* If we have anything to send */
6052 /* Send (or resend) any packets that need it, subject to
6053 * window restrictions and congestion burst control
6054 * restrictions. Ask for an ack on the last packet sent in
6055 * this burst. For now, we're relying upon the window being
6056 * considerably bigger than the largest number of packets that
6057 * are typically sent at once by one initial call to
6058 * rxi_Start. This is probably bogus (perhaps we should ask
6059 * for an ack when we're half way through the current
6060 * window?). Also, for non file transfer applications, this
6061 * may end up asking for an ack for every packet. Bogus. XXXX
6064 * But check whether we're here recursively, and let the other guy
6067 #ifdef RX_ENABLE_LOCKS
6068 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6069 call->flags |= RX_CALL_TQ_BUSY;
6071 #endif /* RX_ENABLE_LOCKS */
6073 #ifdef RX_ENABLE_LOCKS
6074 call->flags &= ~RX_CALL_NEED_START;
6075 #endif /* RX_ENABLE_LOCKS */
6077 maxXmitPackets = MIN(call->twind, call->cwind);
6078 for (opr_queue_Scan(&call->tq, cursor)) {
6080 = opr_queue_Entry(cursor, struct rx_packet, entry);
6082 if (p->flags & RX_PKTFLAG_ACKED) {
6083 /* Since we may block, don't trust this */
6084 if (rx_stats_active)
6085 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6086 continue; /* Ignore this packet if it has been acknowledged */
6089 /* Turn off all flags except these ones, which are the same
6090 * on each transmission */
6091 p->header.flags &= RX_PRESET_FLAGS;
6093 if (p->header.seq >=
6094 call->tfirst + MIN((int)call->twind,
6095 (int)(call->nSoftAcked +
6097 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6098 /* Note: if we're waiting for more window space, we can
6099 * still send retransmits; hence we don't return here, but
6100 * break out to schedule a retransmit event */
6101 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6102 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6107 /* Transmit the packet if it needs to be sent. */
6108 if (!(p->flags & RX_PKTFLAG_SENT)) {
6109 if (nXmitPackets == maxXmitPackets) {
6110 rxi_SendXmitList(call, call->xmitList,
6111 nXmitPackets, istack);
6114 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6115 *(call->callNumber), p));
6116 call->xmitList[nXmitPackets++] = p;
6118 } /* end of the queue_Scan */
6120 /* xmitList now hold pointers to all of the packets that are
6121 * ready to send. Now we loop to send the packets */
6122 if (nXmitPackets > 0) {
6123 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6127 #ifdef RX_ENABLE_LOCKS
6129 /* We went into the error state while sending packets. Now is
6130 * the time to reset the call. This will also inform the using
6131 * process that the call is in an error state.
6133 if (rx_stats_active)
6134 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6135 call->flags &= ~RX_CALL_TQ_BUSY;
6136 rxi_WakeUpTransmitQueue(call);
6137 rxi_CallError(call, call->error);
6141 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6143 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6144 /* Some packets have received acks. If they all have, we can clear
6145 * the transmit queue.
6148 for (opr_queue_ScanSafe(&call->tq, cursor, store)) {
6150 = opr_queue_Entry(cursor, struct rx_packet, entry);
6152 if (p->header.seq < call->tfirst
6153 && (p->flags & RX_PKTFLAG_ACKED)) {
6154 opr_queue_Remove(&p->entry);
6155 #ifdef RX_TRACK_PACKETS
6156 p->flags &= ~RX_PKTFLAG_TQ;
6158 #ifdef RXDEBUG_PACKET
6166 call->flags |= RX_CALL_TQ_CLEARME;
6168 if (call->flags & RX_CALL_TQ_CLEARME)
6169 rxi_ClearTransmitQueue(call, 1);
6170 } while (call->flags & RX_CALL_NEED_START);
6172 * TQ references no longer protected by this flag; they must remain
6173 * protected by the call lock.
6175 call->flags &= ~RX_CALL_TQ_BUSY;
6176 rxi_WakeUpTransmitQueue(call);
6178 call->flags |= RX_CALL_NEED_START;
6180 #endif /* RX_ENABLE_LOCKS */
6182 rxi_rto_cancel(call);
6186 /* Also adjusts the keep alive parameters for the call, to reflect
6187 * that we have just sent a packet (so keep alives aren't sent
6190 rxi_Send(struct rx_call *call, struct rx_packet *p,
6193 struct rx_connection *conn = call->conn;
6195 /* Stamp each packet with the user supplied status */
6196 p->header.userStatus = call->localStatus;
6198 /* Allow the security object controlling this call's security to
6199 * make any last-minute changes to the packet */
6200 RXS_SendPacket(conn->securityObject, call, p);
6202 /* Since we're about to send SOME sort of packet to the peer, it's
6203 * safe to nuke any scheduled end-of-packets ack */
6204 rxi_CancelDelayedAckEvent(call);
6206 /* Actually send the packet, filling in more connection-specific fields */
6207 MUTEX_EXIT(&call->lock);
6208 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6209 rxi_SendPacket(call, conn, p, istack);
6210 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6211 MUTEX_ENTER(&call->lock);
6213 /* Update last send time for this call (for keep-alive
6214 * processing), and for the connection (so that we can discover
6215 * idle connections) */
6216 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6217 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6218 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6220 conn->lastSendTime = call->lastSendTime = clock_Sec();
6224 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6225 * that things are fine. Also called periodically to guarantee that nothing
6226 * falls through the cracks (e.g. (error + dally) connections have keepalive
6227 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6229 * haveCTLock Set if calling from rxi_ReapConnections
6232 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6234 struct rx_connection *conn = call->conn;
6236 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6237 afs_uint32 fudgeFactor;
6240 int idle_timeout = 0;
6241 afs_int32 clock_diff = 0;
6243 if (rxi_CheckPeerDead(call)) {
6249 /* Large swings in the clock can have a significant impact on
6250 * the performance of RX call processing. Forward clock shifts
6251 * will result in premature event triggering or timeouts.
6252 * Backward shifts can result in calls not completing until
6253 * the clock catches up with the original start clock value.
6255 * If a backward clock shift of more than five minutes is noticed,
6256 * just fail the call.
6258 if (now < call->lastSendTime)
6259 clock_diff = call->lastSendTime - now;
6260 if (now < call->startWait)
6261 clock_diff = MAX(clock_diff, call->startWait - now);
6262 if (now < call->lastReceiveTime)
6263 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6264 if (clock_diff > 5 * 60)
6266 if (call->state == RX_STATE_ACTIVE)
6267 rxi_CallError(call, RX_CALL_TIMEOUT);
6271 #ifdef RX_ENABLE_LOCKS
6272 if (call->flags & RX_CALL_TQ_BUSY) {
6273 /* Call is active and will be reset by rxi_Start if it's
6274 * in an error state.
6279 /* RTT + 8*MDEV, rounded up to the next second. */
6280 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6281 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6283 deadTime = conn->secondsUntilDead + fudgeFactor;
6284 /* These are computed to the second (+- 1 second). But that's
6285 * good enough for these values, which should be a significant
6286 * number of seconds. */
6287 if (now > (call->lastReceiveTime + deadTime)) {
6288 if (call->state == RX_STATE_ACTIVE) {
6289 cerror = RX_CALL_DEAD;
6292 #ifdef RX_ENABLE_LOCKS
6293 /* Cancel pending events */
6294 rxi_CancelDelayedAckEvent(call);
6295 rxi_rto_cancel(call);
6296 rxi_CancelKeepAliveEvent(call);
6297 rxi_CancelGrowMTUEvent(call);
6298 MUTEX_ENTER(&rx_refcnt_mutex);
6299 /* if rxi_FreeCall returns 1 it has freed the call */
6300 if (call->refCount == 0 &&
6301 rxi_FreeCall(call, haveCTLock))
6303 MUTEX_EXIT(&rx_refcnt_mutex);
6306 MUTEX_EXIT(&rx_refcnt_mutex);
6308 #else /* RX_ENABLE_LOCKS */
6309 rxi_FreeCall(call, 0);
6311 #endif /* RX_ENABLE_LOCKS */
6313 /* Non-active calls are destroyed if they are not responding
6314 * to pings; active calls are simply flagged in error, so the
6315 * attached process can die reasonably gracefully. */
6318 if (conn->idleDeadTime) {
6319 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6323 /* see if we have a non-activity timeout */
6324 if (call->startWait && ((call->startWait + idleDeadTime) < now)) {
6325 if (call->state == RX_STATE_ACTIVE) {
6326 cerror = RX_CALL_TIMEOUT;
6332 if (conn->hardDeadTime) {
6333 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6336 /* see if we have a hard timeout */
6338 && (now > (hardDeadTime + call->startTime.sec))) {
6339 if (call->state == RX_STATE_ACTIVE)
6340 rxi_CallError(call, RX_CALL_TIMEOUT);
6345 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6346 call->lastReceiveTime) {
6347 int oldMTU = conn->peer->ifMTU;
6349 /* If we thought we could send more, perhaps things got worse.
6350 * Shrink by 128 bytes and try again. */
6351 if (conn->peer->maxPacketSize < conn->lastPacketSize)
6352 /* maxPacketSize will be cleared in rxi_SetPeerMtu */
6353 newmtu = MAX(conn->peer->maxPacketSize + RX_HEADER_SIZE,
6354 conn->lastPacketSize - 128 + RX_HEADER_SIZE);
6356 newmtu = conn->lastPacketSize - 128 + RX_HEADER_SIZE;
6358 /* minimum capped in SetPeerMtu */
6359 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6362 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
6364 /* needed so ResetCall doesn't clobber us. */
6365 call->MTU = conn->peer->ifMTU;
6367 /* if we never succeeded, let the error pass out as-is */
6368 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6369 cerror = conn->msgsizeRetryErr;
6372 rxi_CallError(call, cerror);
6377 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6378 void *dummy, int dummy2)
6380 struct rx_connection *conn = arg1;
6381 struct rx_header theader;
6382 char tbuffer[1 + sizeof(struct rx_header)];
6383 struct sockaddr_in taddr;
6387 struct iovec tmpiov[2];
6390 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6393 tp = &tbuffer[sizeof(struct rx_header)];
6394 taddr.sin_family = AF_INET;
6395 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6396 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6397 memset(&taddr.sin_zero, 0, sizeof(taddr.sin_zero));
6398 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6399 taddr.sin_len = sizeof(struct sockaddr_in);
6401 memset(&theader, 0, sizeof(theader));
6402 theader.epoch = htonl(999);
6404 theader.callNumber = 0;
6407 theader.type = RX_PACKET_TYPE_VERSION;
6408 theader.flags = RX_LAST_PACKET;
6409 theader.serviceId = 0;
6411 memcpy(tbuffer, &theader, sizeof(theader));
6412 memcpy(tp, &a, sizeof(a));
6413 tmpiov[0].iov_base = tbuffer;
6414 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6416 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6418 MUTEX_ENTER(&conn->conn_data_lock);
6419 /* We ran, so the handle is no longer needed to try to cancel ourselves. */
6420 if (event == conn->natKeepAliveEvent)
6421 rxevent_Put(&conn->natKeepAliveEvent);
6422 MUTEX_ENTER(&rx_refcnt_mutex);
6423 /* Only reschedule ourselves if the connection would not be destroyed */
6424 if (conn->refCount > 1)
6426 if (conn->refCount <= 0) {
6427 #ifdef RX_REFCOUNT_CHECK
6428 osi_Assert(conn->refCount == 0);
6430 if (rx_stats_active) {
6431 MUTEX_ENTER(&rx_stats_mutex);
6432 rxi_lowConnRefCount++;
6433 MUTEX_EXIT(&rx_stats_mutex);
6436 MUTEX_EXIT(&rx_refcnt_mutex);
6438 rxi_ScheduleNatKeepAliveEvent(conn);
6439 MUTEX_EXIT(&conn->conn_data_lock);
6440 putConnection(conn);
6444 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6446 MUTEX_ASSERT(&conn->conn_data_lock);
6447 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6448 struct clock when, now;
6449 clock_GetTime(&now);
6451 when.sec += conn->secondsUntilNatPing;
6452 rx_GetConnection(conn);
6453 conn->natKeepAliveEvent =
6454 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6459 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6461 MUTEX_ENTER(&conn->conn_data_lock);
6462 conn->secondsUntilNatPing = seconds;
6464 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6465 rxi_ScheduleNatKeepAliveEvent(conn);
6467 conn->flags |= RX_CONN_NAT_PING;
6469 MUTEX_EXIT(&conn->conn_data_lock);
6472 /* When a call is in progress, this routine is called occasionally to
6473 * make sure that some traffic has arrived (or been sent to) the peer.
6474 * If nothing has arrived in a reasonable amount of time, the call is
6475 * declared dead; if nothing has been sent for a while, we send a
6476 * keep-alive packet (if we're actually trying to keep the call alive)
6479 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6482 struct rx_call *call = arg1;
6483 struct rx_connection *conn;
6486 MUTEX_ENTER(&call->lock);
6488 if (event == call->keepAliveEvent)
6489 rxevent_Put(&call->keepAliveEvent);
6493 if (rxi_CheckCall(call, 0)) {
6494 MUTEX_EXIT(&call->lock);
6495 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6499 /* Don't try to keep alive dallying calls */
6500 if (call->state == RX_STATE_DALLY) {
6501 MUTEX_EXIT(&call->lock);
6502 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6507 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6508 /* Don't try to send keepalives if there is unacknowledged data */
6509 /* the rexmit code should be good enough, this little hack
6510 * doesn't quite work XXX */
6511 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6513 rxi_ScheduleKeepAliveEvent(call);
6514 MUTEX_EXIT(&call->lock);
6515 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6518 /* Does what's on the nameplate. */
6520 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6522 struct rx_call *call = arg1;
6523 struct rx_connection *conn;
6525 MUTEX_ENTER(&call->lock);
6527 if (event == call->growMTUEvent)
6528 rxevent_Put(&call->growMTUEvent);
6530 if (rxi_CheckCall(call, 0))
6533 /* Don't bother with dallying calls */
6534 if (call->state == RX_STATE_DALLY)
6540 * keep being scheduled, just don't do anything if we're at peak,
6541 * or we're not set up to be properly handled (idle timeout required)
6543 if ((conn->peer->maxPacketSize != 0) &&
6544 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6546 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6547 rxi_ScheduleGrowMTUEvent(call, 0);
6549 MUTEX_EXIT(&call->lock);
6550 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6554 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6556 MUTEX_ASSERT(&call->lock);
6557 if (!call->keepAliveEvent) {
6558 struct clock when, now;
6559 clock_GetTime(&now);
6561 when.sec += call->conn->secondsUntilPing;
6562 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6563 call->keepAliveEvent =
6564 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6569 rxi_CancelKeepAliveEvent(struct rx_call *call) {
6570 MUTEX_ASSERT(&call->lock);
6571 if (rxevent_Cancel(&call->keepAliveEvent))
6572 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6576 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6578 MUTEX_ASSERT(&call->lock);
6579 if (!call->growMTUEvent) {
6580 struct clock when, now;
6582 clock_GetTime(&now);
6585 if (call->conn->secondsUntilPing)
6586 secs = (6*call->conn->secondsUntilPing)-1;
6588 if (call->conn->secondsUntilDead)
6589 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6593 CALL_HOLD(call, RX_CALL_REFCOUNT_MTU);
6594 call->growMTUEvent =
6595 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6600 rxi_CancelGrowMTUEvent(struct rx_call *call)
6602 MUTEX_ASSERT(&call->lock);
6603 if (rxevent_Cancel(&call->growMTUEvent))
6604 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6608 * Increment the counter for the next connection ID, handling overflow.
6611 update_nextCid(void)
6613 /* Overflow is technically undefined behavior; avoid it. */
6614 if (rx_nextCid > MAX_AFS_INT32 - (1 << RX_CIDSHIFT))
6615 rx_nextCid = -1 * ((MAX_AFS_INT32 / RX_CIDSHIFT) * RX_CIDSHIFT);
6617 rx_nextCid += 1 << RX_CIDSHIFT;
6621 rxi_KeepAliveOn(struct rx_call *call)
6623 /* Pretend last packet received was received now--i.e. if another
6624 * packet isn't received within the keep alive time, then the call
6625 * will die; Initialize last send time to the current time--even
6626 * if a packet hasn't been sent yet. This will guarantee that a
6627 * keep-alive is sent within the ping time */
6628 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6629 rxi_ScheduleKeepAliveEvent(call);
6633 rxi_GrowMTUOn(struct rx_call *call)
6635 struct rx_connection *conn = call->conn;
6636 MUTEX_ENTER(&conn->conn_data_lock);
6637 conn->lastPingSizeSer = conn->lastPingSize = 0;
6638 MUTEX_EXIT(&conn->conn_data_lock);
6639 rxi_ScheduleGrowMTUEvent(call, 1);
6642 /* This routine is called to send connection abort messages
6643 * that have been delayed to throttle looping clients. */
6645 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6648 struct rx_connection *conn = arg1;
6651 struct rx_packet *packet;
6653 MUTEX_ENTER(&conn->conn_data_lock);
6654 if (event == conn->delayedAbortEvent)
6655 rxevent_Put(&conn->delayedAbortEvent);
6656 error = htonl(conn->error);
6658 MUTEX_EXIT(&conn->conn_data_lock);
6659 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6662 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6663 RX_PACKET_TYPE_ABORT, (char *)&error,
6665 rxi_FreePacket(packet);
6667 putConnection(conn);
6670 /* This routine is called to send call abort messages
6671 * that have been delayed to throttle looping clients. */
6673 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6676 struct rx_call *call = arg1;
6679 struct rx_packet *packet;
6681 MUTEX_ENTER(&call->lock);
6682 if (event == call->delayedAbortEvent)
6683 rxevent_Put(&call->delayedAbortEvent);
6684 error = htonl(call->error);
6686 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6689 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6690 (char *)&error, sizeof(error), 0);
6691 rxi_FreePacket(packet);
6693 MUTEX_EXIT(&call->lock);
6694 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6698 * This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6699 * seconds) to ask the client to authenticate itself. The routine
6700 * issues a challenge to the client, which is obtained from the
6701 * security object associated with the connection
6703 * This routine is both an event handler and a function called directly;
6704 * when called directly the passed |event| is NULL and the
6705 * conn->conn->data>lock must must not be held.
6708 rxi_ChallengeEvent(struct rxevent *event,
6709 void *arg0, void *arg1, int tries)
6711 struct rx_connection *conn = arg0;
6713 MUTEX_ENTER(&conn->conn_data_lock);
6714 if (event != NULL && event == conn->challengeEvent)
6715 rxevent_Put(&conn->challengeEvent);
6716 MUTEX_EXIT(&conn->conn_data_lock);
6718 /* If there are no active calls it is not worth re-issuing the
6719 * challenge. If the client issues another call on this connection
6720 * the challenge can be requested at that time.
6722 if (!rxi_HasActiveCalls(conn)) {
6723 putConnection(conn);
6727 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6728 struct rx_packet *packet;
6729 struct clock when, now;
6732 /* We've failed to authenticate for too long.
6733 * Reset any calls waiting for authentication;
6734 * they are all in RX_STATE_PRECALL.
6738 MUTEX_ENTER(&conn->conn_call_lock);
6739 for (i = 0; i < RX_MAXCALLS; i++) {
6740 struct rx_call *call = conn->call[i];
6742 MUTEX_ENTER(&call->lock);
6743 if (call->state == RX_STATE_PRECALL) {
6744 rxi_CallError(call, RX_CALL_DEAD);
6745 rxi_SendCallAbort(call, NULL, 0, 0);
6747 MUTEX_EXIT(&call->lock);
6750 MUTEX_EXIT(&conn->conn_call_lock);
6751 putConnection(conn);
6755 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6757 /* If there's no packet available, do this later. */
6758 RXS_GetChallenge(conn->securityObject, conn, packet);
6759 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6760 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6761 rxi_FreePacket(packet);
6762 conn->securityChallengeSent = 1;
6764 clock_GetTime(&now);
6766 when.sec += RX_CHALLENGE_TIMEOUT;
6767 MUTEX_ENTER(&conn->conn_data_lock);
6768 /* Only reschedule ourselves if not already pending. */
6769 if (conn->challengeEvent == NULL) {
6770 rx_GetConnection(conn);
6771 conn->challengeEvent =
6772 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6775 MUTEX_EXIT(&conn->conn_data_lock);
6777 putConnection(conn);
6780 /* Call this routine to start requesting the client to authenticate
6781 * itself. This will continue until authentication is established,
6782 * the call times out, or an invalid response is returned. The
6783 * security object associated with the connection is asked to create
6784 * the challenge at this time. */
6786 rxi_ChallengeOn(struct rx_connection *conn)
6789 MUTEX_ENTER(&conn->conn_data_lock);
6790 if (!conn->challengeEvent)
6792 MUTEX_EXIT(&conn->conn_data_lock);
6794 RXS_CreateChallenge(conn->securityObject, conn);
6795 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6800 /* rxi_ComputeRoundTripTime is called with peer locked. */
6801 /* peer may be null */
6803 rxi_ComputeRoundTripTime(struct rx_packet *p,
6804 struct rx_ackPacket *ack,
6805 struct rx_call *call,
6806 struct rx_peer *peer,
6809 struct clock thisRtt, *sentp;
6813 /* If the ACK is delayed, then do nothing */
6814 if (ack->reason == RX_ACK_DELAY)
6817 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6818 * their RTT multiple times, so only include the RTT of the last packet
6820 if (p->flags & RX_JUMBO_PACKET)
6823 /* Use the serial number to determine which transmission the ACK is for,
6824 * and set the sent time to match this. If we have no serial number, then
6825 * only use the ACK for RTT calculations if the packet has not been
6829 serial = ntohl(ack->serial);
6831 if (serial == p->header.serial) {
6832 sentp = &p->timeSent;
6833 } else if (serial == p->firstSerial) {
6834 sentp = &p->firstSent;
6835 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6836 sentp = &p->firstSent;
6840 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6841 sentp = &p->firstSent;
6848 if (clock_Lt(&thisRtt, sentp))
6849 return; /* somebody set the clock back, don't count this time. */
6851 clock_Sub(&thisRtt, sentp);
6852 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6853 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6855 if (clock_IsZero(&thisRtt)) {
6857 * The actual round trip time is shorter than the
6858 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6859 * Since we can't tell which at the moment we will assume 1ms.
6861 thisRtt.usec = 1000;
6864 if (rx_stats_active) {
6865 MUTEX_ENTER(&rx_stats_mutex);
6866 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6867 rx_stats.minRtt = thisRtt;
6868 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6869 if (thisRtt.sec > 60) {
6870 MUTEX_EXIT(&rx_stats_mutex);
6871 return; /* somebody set the clock ahead */
6873 rx_stats.maxRtt = thisRtt;
6875 clock_Add(&rx_stats.totalRtt, &thisRtt);
6876 rx_atomic_inc(&rx_stats.nRttSamples);
6877 MUTEX_EXIT(&rx_stats_mutex);
6880 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6882 /* Apply VanJacobson round-trip estimations */
6887 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6888 * srtt is stored as fixed point with 3 bits after the binary
6889 * point (i.e., scaled by 8). The following magic is
6890 * equivalent to the smoothing algorithm in rfc793 with an
6891 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6892 * srtt'*8 = rtt + srtt*7
6893 * srtt'*8 = srtt*8 + rtt - srtt
6894 * srtt' = srtt + rtt/8 - srtt/8
6895 * srtt' = srtt + (rtt - srtt)/8
6898 delta = _8THMSEC(&thisRtt) - call->rtt;
6899 call->rtt += (delta >> 3);
6902 * We accumulate a smoothed rtt variance (actually, a smoothed
6903 * mean difference), then set the retransmit timer to smoothed
6904 * rtt + 4 times the smoothed variance (was 2x in van's original
6905 * paper, but 4x works better for me, and apparently for him as
6907 * rttvar is stored as
6908 * fixed point with 2 bits after the binary point (scaled by
6909 * 4). The following is equivalent to rfc793 smoothing with
6910 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6911 * rttvar'*4 = rttvar*3 + |delta|
6912 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6913 * rttvar' = rttvar + |delta|/4 - rttvar/4
6914 * rttvar' = rttvar + (|delta| - rttvar)/4
6915 * This replaces rfc793's wired-in beta.
6916 * dev*4 = dev*4 + (|actual - expected| - dev)
6922 delta -= (call->rtt_dev << 1);
6923 call->rtt_dev += (delta >> 3);
6925 /* I don't have a stored RTT so I start with this value. Since I'm
6926 * probably just starting a call, and will be pushing more data down
6927 * this, I expect congestion to increase rapidly. So I fudge a
6928 * little, and I set deviance to half the rtt. In practice,
6929 * deviance tends to approach something a little less than
6930 * half the smoothed rtt. */
6931 call->rtt = _8THMSEC(&thisRtt) + 8;
6932 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6934 /* the smoothed RTT time is RTT + 4*MDEV
6936 * We allow a user specified minimum to be set for this, to allow clamping
6937 * at a minimum value in the same way as TCP. In addition, we have to allow
6938 * for the possibility that this packet is answered by a delayed ACK, so we
6939 * add on a fixed 200ms to account for that timer expiring.
6942 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6943 rx_minPeerTimeout) + 200;
6944 clock_Zero(&call->rto);
6945 clock_Addmsec(&call->rto, rtt_timeout);
6947 /* Update the peer, so any new calls start with our values */
6948 peer->rtt_dev = call->rtt_dev;
6949 peer->rtt = call->rtt;
6951 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6952 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6956 /* Find all server connections that have not been active for a long time, and
6959 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
6962 struct clock now, when;
6963 struct rxevent *event;
6964 clock_GetTime(&now);
6966 /* Find server connection structures that haven't been used for
6967 * greater than rx_idleConnectionTime */
6969 struct rx_connection **conn_ptr, **conn_end;
6970 int i, havecalls = 0;
6971 MUTEX_ENTER(&rx_connHashTable_lock);
6972 for (conn_ptr = &rx_connHashTable[0], conn_end =
6973 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6975 struct rx_connection *conn, *next;
6976 struct rx_call *call;
6980 for (conn = *conn_ptr; conn; conn = next) {
6981 /* XXX -- Shouldn't the connection be locked? */
6984 for (i = 0; i < RX_MAXCALLS; i++) {
6985 call = conn->call[i];
6989 code = MUTEX_TRYENTER(&call->lock);
6992 result = rxi_CheckCall(call, 1);
6993 MUTEX_EXIT(&call->lock);
6995 /* If CheckCall freed the call, it might
6996 * have destroyed the connection as well,
6997 * which screws up the linked lists.
7003 if (conn->type == RX_SERVER_CONNECTION) {
7004 /* This only actually destroys the connection if
7005 * there are no outstanding calls */
7006 MUTEX_ENTER(&conn->conn_data_lock);
7007 MUTEX_ENTER(&rx_refcnt_mutex);
7008 if (!havecalls && !conn->refCount
7009 && ((conn->lastSendTime + rx_idleConnectionTime) <
7011 conn->refCount++; /* it will be decr in rx_DestroyConn */
7012 MUTEX_EXIT(&rx_refcnt_mutex);
7013 MUTEX_EXIT(&conn->conn_data_lock);
7014 #ifdef RX_ENABLE_LOCKS
7015 rxi_DestroyConnectionNoLock(conn);
7016 #else /* RX_ENABLE_LOCKS */
7017 rxi_DestroyConnection(conn);
7018 #endif /* RX_ENABLE_LOCKS */
7020 #ifdef RX_ENABLE_LOCKS
7022 MUTEX_EXIT(&rx_refcnt_mutex);
7023 MUTEX_EXIT(&conn->conn_data_lock);
7025 #endif /* RX_ENABLE_LOCKS */
7029 #ifdef RX_ENABLE_LOCKS
7030 while (rx_connCleanup_list) {
7031 struct rx_connection *conn;
7032 conn = rx_connCleanup_list;
7033 rx_connCleanup_list = rx_connCleanup_list->next;
7034 MUTEX_EXIT(&rx_connHashTable_lock);
7035 rxi_CleanupConnection(conn);
7036 MUTEX_ENTER(&rx_connHashTable_lock);
7038 MUTEX_EXIT(&rx_connHashTable_lock);
7039 #endif /* RX_ENABLE_LOCKS */
7042 /* Find any peer structures that haven't been used (haven't had an
7043 * associated connection) for greater than rx_idlePeerTime */
7045 struct rx_peer **peer_ptr, **peer_end;
7049 * Why do we need to hold the rx_peerHashTable_lock across
7050 * the incrementing of peer_ptr since the rx_peerHashTable
7051 * array is not changing? We don't.
7053 * By dropping the lock periodically we can permit other
7054 * activities to be performed while a rxi_ReapConnections
7055 * call is in progress. The goal of reap connections
7056 * is to clean up quickly without causing large amounts
7057 * of contention. Therefore, it is important that global
7058 * mutexes not be held for extended periods of time.
7060 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7061 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7063 struct rx_peer *peer, *next, *prev;
7065 MUTEX_ENTER(&rx_peerHashTable_lock);
7066 for (prev = peer = *peer_ptr; peer; peer = next) {
7068 code = MUTEX_TRYENTER(&peer->peer_lock);
7069 if ((code) && (peer->refCount == 0)
7070 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7071 struct opr_queue *cursor, *store;
7075 * now know that this peer object is one to be
7076 * removed from the hash table. Once it is removed
7077 * it can't be referenced by other threads.
7078 * Lets remove it first and decrement the struct
7079 * nPeerStructs count.
7081 if (peer == *peer_ptr) {
7087 if (rx_stats_active)
7088 rx_atomic_dec(&rx_stats.nPeerStructs);
7091 * Now if we hold references on 'prev' and 'next'
7092 * we can safely drop the rx_peerHashTable_lock
7093 * while we destroy this 'peer' object.
7099 MUTEX_EXIT(&rx_peerHashTable_lock);
7101 MUTEX_EXIT(&peer->peer_lock);
7102 MUTEX_DESTROY(&peer->peer_lock);
7104 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7105 unsigned int num_funcs;
7106 struct rx_interface_stat *rpc_stat
7107 = opr_queue_Entry(cursor, struct rx_interface_stat,
7112 opr_queue_Remove(&rpc_stat->entry);
7113 opr_queue_Remove(&rpc_stat->entryPeers);
7115 num_funcs = rpc_stat->stats[0].func_total;
7117 sizeof(rx_interface_stat_t) +
7118 rpc_stat->stats[0].func_total *
7119 sizeof(rx_function_entry_v1_t);
7121 rxi_Free(rpc_stat, space);
7123 MUTEX_ENTER(&rx_rpc_stats);
7124 rxi_rpc_peer_stat_cnt -= num_funcs;
7125 MUTEX_EXIT(&rx_rpc_stats);
7130 * Regain the rx_peerHashTable_lock and
7131 * decrement the reference count on 'prev'
7134 MUTEX_ENTER(&rx_peerHashTable_lock);
7141 MUTEX_EXIT(&peer->peer_lock);
7146 MUTEX_EXIT(&rx_peerHashTable_lock);
7150 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7151 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7152 * GC, just below. Really, we shouldn't have to keep moving packets from
7153 * one place to another, but instead ought to always know if we can
7154 * afford to hold onto a packet in its particular use. */
7155 MUTEX_ENTER(&rx_freePktQ_lock);
7156 if (rx_waitingForPackets) {
7157 rx_waitingForPackets = 0;
7158 #ifdef RX_ENABLE_LOCKS
7159 CV_BROADCAST(&rx_waitingForPackets_cv);
7161 osi_rxWakeup(&rx_waitingForPackets);
7164 MUTEX_EXIT(&rx_freePktQ_lock);
7167 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7168 event = rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0);
7169 rxevent_Put(&event);
7173 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7174 * rx.h is sort of strange this is better. This is called with a security
7175 * object before it is discarded. Each connection using a security object has
7176 * its own refcount to the object so it won't actually be freed until the last
7177 * connection is destroyed.
7179 * This is the only rxs module call. A hold could also be written but no one
7183 rxs_Release(struct rx_securityClass *aobj)
7185 return RXS_Close(aobj);
7193 #define TRACE_OPTION_RX_DEBUG 16
7201 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7202 0, KEY_QUERY_VALUE, &parmKey);
7203 if (code != ERROR_SUCCESS)
7206 dummyLen = sizeof(TraceOption);
7207 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7208 (BYTE *) &TraceOption, &dummyLen);
7209 if (code == ERROR_SUCCESS) {
7210 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7212 RegCloseKey (parmKey);
7213 #endif /* AFS_NT40_ENV */
7218 rx_DebugOnOff(int on)
7222 rxdebug_active = on;
7228 rx_StatsOnOff(int on)
7230 rx_stats_active = on;
7234 /* Don't call this debugging routine directly; use dpf */
7236 rxi_DebugPrint(char *format, ...)
7245 va_start(ap, format);
7247 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7250 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7252 OutputDebugString(msg);
7258 va_start(ap, format);
7260 clock_GetTime(&now);
7261 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7262 (unsigned int)now.usec);
7263 vfprintf(rx_Log, format, ap);
7271 * This function is used to process the rx_stats structure that is local
7272 * to a process as well as an rx_stats structure received from a remote
7273 * process (via rxdebug). Therefore, it needs to do minimal version
7277 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7278 afs_int32 freePackets, char version)
7282 if (size != sizeof(struct rx_statistics)) {
7284 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7285 size, sizeof(struct rx_statistics));
7288 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7291 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7292 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7293 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7294 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7295 s->specialPktAllocFailures);
7297 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7298 s->receivePktAllocFailures, s->sendPktAllocFailures,
7299 s->specialPktAllocFailures);
7303 " greedy %u, " "bogusReads %u (last from host %x), "
7304 "noPackets %u, " "noBuffers %u, " "selects %u, "
7305 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7306 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7307 s->selects, s->sendSelects);
7309 fprintf(file, " packets read: ");
7310 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7311 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7313 fprintf(file, "\n");
7316 " other read counters: data %u, " "ack %u, " "dup %u "
7317 "spurious %u " "dally %u\n", s->dataPacketsRead,
7318 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7319 s->ignorePacketDally);
7321 fprintf(file, " packets sent: ");
7322 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7323 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7325 fprintf(file, "\n");
7328 " other send counters: ack %u, " "data %u (not resends), "
7329 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7330 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7331 s->dataPacketsPushed, s->ignoreAckedPacket);
7334 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7335 s->netSendFailures, (int)s->fatalErrors);
7337 if (s->nRttSamples) {
7338 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7339 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7341 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7342 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7346 " %d server connections, " "%d client connections, "
7347 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7348 s->nServerConns, s->nClientConns, s->nPeerStructs,
7349 s->nCallStructs, s->nFreeCallStructs);
7351 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7352 fprintf(file, " %d clock updates\n", clock_nUpdates);
7356 /* for backward compatibility */
7358 rx_PrintStats(FILE * file)
7360 MUTEX_ENTER(&rx_stats_mutex);
7361 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7362 sizeof(rx_stats), rx_nFreePackets,
7364 MUTEX_EXIT(&rx_stats_mutex);
7368 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7370 fprintf(file, "Peer %x.%d.\n",
7371 ntohl(peer->host), (int)ntohs(peer->port));
7374 " Rtt %d, " "total sent %d, " "resent %d\n",
7375 peer->rtt, peer->nSent, peer->reSends);
7377 fprintf(file, " Packet size %d\n", peer->ifMTU);
7381 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7383 * This mutex protects the following static variables:
7387 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7388 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7390 #define LOCK_RX_DEBUG
7391 #define UNLOCK_RX_DEBUG
7392 #endif /* AFS_PTHREAD_ENV */
7394 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7396 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7397 u_char type, void *inputData, size_t inputLength,
7398 void *outputData, size_t outputLength)
7400 static afs_int32 counter = 100;
7401 time_t waitTime, waitCount;
7402 struct rx_header theader;
7405 struct timeval tv_now, tv_wake, tv_delta;
7406 struct sockaddr_in taddr, faddr;
7420 tp = &tbuffer[sizeof(struct rx_header)];
7421 taddr.sin_family = AF_INET;
7422 taddr.sin_port = remotePort;
7423 taddr.sin_addr.s_addr = remoteAddr;
7424 memset(&taddr.sin_zero, 0, sizeof(taddr.sin_zero));
7425 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7426 taddr.sin_len = sizeof(struct sockaddr_in);
7429 memset(&theader, 0, sizeof(theader));
7430 theader.epoch = htonl(999);
7432 theader.callNumber = htonl(counter);
7435 theader.type = type;
7436 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7437 theader.serviceId = 0;
7439 memcpy(tbuffer, &theader, sizeof(theader));
7440 memcpy(tp, inputData, inputLength);
7442 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7443 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7445 /* see if there's a packet available */
7446 gettimeofday(&tv_wake, NULL);
7447 tv_wake.tv_sec += waitTime;
7450 FD_SET(socket, &imask);
7451 tv_delta.tv_sec = tv_wake.tv_sec;
7452 tv_delta.tv_usec = tv_wake.tv_usec;
7453 gettimeofday(&tv_now, NULL);
7455 if (tv_delta.tv_usec < tv_now.tv_usec) {
7457 tv_delta.tv_usec += 1000000;
7460 tv_delta.tv_usec -= tv_now.tv_usec;
7462 if (tv_delta.tv_sec < tv_now.tv_sec) {
7466 tv_delta.tv_sec -= tv_now.tv_sec;
7469 code = select(0, &imask, 0, 0, &tv_delta);
7470 #else /* AFS_NT40_ENV */
7471 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7472 #endif /* AFS_NT40_ENV */
7473 if (code == 1 && FD_ISSET(socket, &imask)) {
7474 /* now receive a packet */
7475 faddrLen = sizeof(struct sockaddr_in);
7477 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7478 (struct sockaddr *)&faddr, &faddrLen);
7481 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7482 if (counter == ntohl(theader.callNumber))
7490 /* see if we've timed out */
7498 code -= sizeof(struct rx_header);
7499 if (code > outputLength)
7500 code = outputLength;
7501 memcpy(outputData, tp, code);
7504 #endif /* RXDEBUG */
7507 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7508 afs_uint16 remotePort, struct rx_debugStats * stat,
7509 afs_uint32 * supportedValues)
7511 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7513 struct rx_debugIn in;
7515 *supportedValues = 0;
7516 in.type = htonl(RX_DEBUGI_GETSTATS);
7519 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7520 &in, sizeof(in), stat, sizeof(*stat));
7523 * If the call was successful, fixup the version and indicate
7524 * what contents of the stat structure are valid.
7525 * Also do net to host conversion of fields here.
7529 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7530 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7532 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7533 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7535 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7536 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7538 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7539 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7541 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7542 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7544 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7545 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7547 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7548 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7550 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7551 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7553 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7554 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7556 stat->nFreePackets = ntohl(stat->nFreePackets);
7557 stat->packetReclaims = ntohl(stat->packetReclaims);
7558 stat->callsExecuted = ntohl(stat->callsExecuted);
7559 stat->nWaiting = ntohl(stat->nWaiting);
7560 stat->idleThreads = ntohl(stat->idleThreads);
7561 stat->nWaited = ntohl(stat->nWaited);
7562 stat->nPackets = ntohl(stat->nPackets);
7571 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7572 afs_uint16 remotePort, struct rx_statistics * stat,
7573 afs_uint32 * supportedValues)
7575 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7577 struct rx_debugIn in;
7578 afs_int32 *lp = (afs_int32 *) stat;
7582 * supportedValues is currently unused, but added to allow future
7583 * versioning of this function.
7586 *supportedValues = 0;
7587 in.type = htonl(RX_DEBUGI_RXSTATS);
7589 memset(stat, 0, sizeof(*stat));
7591 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7592 &in, sizeof(in), stat, sizeof(*stat));
7597 * Do net to host conversion here
7600 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7611 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7612 afs_uint16 remotePort, size_t version_length,
7615 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7617 return MakeDebugCall(socket, remoteAddr, remotePort,
7618 RX_PACKET_TYPE_VERSION, a, 1, version,
7626 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7627 afs_uint16 remotePort, afs_int32 * nextConnection,
7628 int allConnections, afs_uint32 debugSupportedValues,
7629 struct rx_debugConn * conn,
7630 afs_uint32 * supportedValues)
7632 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7634 struct rx_debugIn in;
7638 * supportedValues is currently unused, but added to allow future
7639 * versioning of this function.
7642 *supportedValues = 0;
7643 if (allConnections) {
7644 in.type = htonl(RX_DEBUGI_GETALLCONN);
7646 in.type = htonl(RX_DEBUGI_GETCONN);
7648 in.index = htonl(*nextConnection);
7649 memset(conn, 0, sizeof(*conn));
7651 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7652 &in, sizeof(in), conn, sizeof(*conn));
7655 *nextConnection += 1;
7658 * Convert old connection format to new structure.
7661 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7662 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7663 #define MOVEvL(a) (conn->a = vL->a)
7665 /* any old or unrecognized version... */
7666 for (i = 0; i < RX_MAXCALLS; i++) {
7667 MOVEvL(callState[i]);
7668 MOVEvL(callMode[i]);
7669 MOVEvL(callFlags[i]);
7670 MOVEvL(callOther[i]);
7672 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7673 MOVEvL(secStats.type);
7674 MOVEvL(secStats.level);
7675 MOVEvL(secStats.flags);
7676 MOVEvL(secStats.expires);
7677 MOVEvL(secStats.packetsReceived);
7678 MOVEvL(secStats.packetsSent);
7679 MOVEvL(secStats.bytesReceived);
7680 MOVEvL(secStats.bytesSent);
7685 * Do net to host conversion here
7687 * I don't convert host or port since we are most likely
7688 * going to want these in NBO.
7690 conn->cid = ntohl(conn->cid);
7691 conn->serial = ntohl(conn->serial);
7692 for (i = 0; i < RX_MAXCALLS; i++) {
7693 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7695 conn->error = ntohl(conn->error);
7696 conn->secStats.flags = ntohl(conn->secStats.flags);
7697 conn->secStats.expires = ntohl(conn->secStats.expires);
7698 conn->secStats.packetsReceived =
7699 ntohl(conn->secStats.packetsReceived);
7700 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7701 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7702 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7703 conn->epoch = ntohl(conn->epoch);
7704 conn->natMTU = ntohl(conn->natMTU);
7713 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7714 afs_uint16 remotePort, afs_int32 * nextPeer,
7715 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7716 afs_uint32 * supportedValues)
7718 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7720 struct rx_debugIn in;
7723 * supportedValues is currently unused, but added to allow future
7724 * versioning of this function.
7727 *supportedValues = 0;
7728 in.type = htonl(RX_DEBUGI_GETPEER);
7729 in.index = htonl(*nextPeer);
7730 memset(peer, 0, sizeof(*peer));
7732 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7733 &in, sizeof(in), peer, sizeof(*peer));
7739 * Do net to host conversion here
7741 * I don't convert host or port since we are most likely
7742 * going to want these in NBO.
7744 peer->ifMTU = ntohs(peer->ifMTU);
7745 peer->idleWhen = ntohl(peer->idleWhen);
7746 peer->refCount = ntohs(peer->refCount);
7747 peer->rtt = ntohl(peer->rtt);
7748 peer->rtt_dev = ntohl(peer->rtt_dev);
7749 peer->timeout.sec = 0;
7750 peer->timeout.usec = 0;
7751 peer->nSent = ntohl(peer->nSent);
7752 peer->reSends = ntohl(peer->reSends);
7753 peer->natMTU = ntohs(peer->natMTU);
7754 peer->maxMTU = ntohs(peer->maxMTU);
7755 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7756 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7757 peer->MTU = ntohs(peer->MTU);
7758 peer->cwind = ntohs(peer->cwind);
7759 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7760 peer->congestSeq = ntohs(peer->congestSeq);
7761 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7762 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7763 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7764 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7773 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7774 struct rx_debugPeer * peerStats)
7777 afs_int32 error = 1; /* default to "did not succeed" */
7778 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7780 MUTEX_ENTER(&rx_peerHashTable_lock);
7781 for(tp = rx_peerHashTable[hashValue];
7782 tp != NULL; tp = tp->next) {
7783 if (tp->host == peerHost)
7789 MUTEX_EXIT(&rx_peerHashTable_lock);
7793 MUTEX_ENTER(&tp->peer_lock);
7794 peerStats->host = tp->host;
7795 peerStats->port = tp->port;
7796 peerStats->ifMTU = tp->ifMTU;
7797 peerStats->idleWhen = tp->idleWhen;
7798 peerStats->refCount = tp->refCount;
7799 peerStats->burstSize = 0;
7800 peerStats->burst = 0;
7801 peerStats->burstWait.sec = 0;
7802 peerStats->burstWait.usec = 0;
7803 peerStats->rtt = tp->rtt;
7804 peerStats->rtt_dev = tp->rtt_dev;
7805 peerStats->timeout.sec = 0;
7806 peerStats->timeout.usec = 0;
7807 peerStats->nSent = tp->nSent;
7808 peerStats->reSends = tp->reSends;
7809 peerStats->natMTU = tp->natMTU;
7810 peerStats->maxMTU = tp->maxMTU;
7811 peerStats->maxDgramPackets = tp->maxDgramPackets;
7812 peerStats->ifDgramPackets = tp->ifDgramPackets;
7813 peerStats->MTU = tp->MTU;
7814 peerStats->cwind = tp->cwind;
7815 peerStats->nDgramPackets = tp->nDgramPackets;
7816 peerStats->congestSeq = tp->congestSeq;
7817 peerStats->bytesSent.high = tp->bytesSent >> 32;
7818 peerStats->bytesSent.low = tp->bytesSent & MAX_AFS_UINT32;
7819 peerStats->bytesReceived.high = tp->bytesReceived >> 32;
7820 peerStats->bytesReceived.low
7821 = tp->bytesReceived & MAX_AFS_UINT32;
7822 MUTEX_EXIT(&tp->peer_lock);
7824 MUTEX_ENTER(&rx_peerHashTable_lock);
7827 MUTEX_EXIT(&rx_peerHashTable_lock);
7835 struct rx_serverQueueEntry *np;
7838 struct rx_call *call;
7839 struct rx_serverQueueEntry *sq;
7842 if (rx_atomic_test_and_set_bit(&rxinit_status, 0))
7843 return; /* Already shutdown. */
7847 #ifndef AFS_PTHREAD_ENV
7848 FD_ZERO(&rx_selectMask);
7849 #endif /* AFS_PTHREAD_ENV */
7850 rxi_dataQuota = RX_MAX_QUOTA;
7851 #ifndef AFS_PTHREAD_ENV
7853 #endif /* AFS_PTHREAD_ENV */
7856 #ifndef AFS_PTHREAD_ENV
7857 #ifndef AFS_USE_GETTIMEOFDAY
7859 #endif /* AFS_USE_GETTIMEOFDAY */
7860 #endif /* AFS_PTHREAD_ENV */
7862 while (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
7863 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
7864 opr_queue_Remove(&call->entry);
7865 rxi_Free(call, sizeof(struct rx_call));
7868 while (!opr_queue_IsEmpty(&rx_idleServerQueue)) {
7869 sq = opr_queue_First(&rx_idleServerQueue, struct rx_serverQueueEntry,
7871 opr_queue_Remove(&sq->entry);
7876 struct rx_peer **peer_ptr, **peer_end;
7877 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7878 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7880 struct rx_peer *peer, *next;
7882 MUTEX_ENTER(&rx_peerHashTable_lock);
7883 for (peer = *peer_ptr; peer; peer = next) {
7884 struct opr_queue *cursor, *store;
7887 MUTEX_ENTER(&rx_rpc_stats);
7888 MUTEX_ENTER(&peer->peer_lock);
7889 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7890 unsigned int num_funcs;
7891 struct rx_interface_stat *rpc_stat
7892 = opr_queue_Entry(cursor, struct rx_interface_stat,
7896 opr_queue_Remove(&rpc_stat->entry);
7897 opr_queue_Remove(&rpc_stat->entryPeers);
7898 num_funcs = rpc_stat->stats[0].func_total;
7900 sizeof(rx_interface_stat_t) +
7901 rpc_stat->stats[0].func_total *
7902 sizeof(rx_function_entry_v1_t);
7904 rxi_Free(rpc_stat, space);
7906 /* rx_rpc_stats must be held */
7907 rxi_rpc_peer_stat_cnt -= num_funcs;
7909 MUTEX_EXIT(&peer->peer_lock);
7910 MUTEX_EXIT(&rx_rpc_stats);
7914 if (rx_stats_active)
7915 rx_atomic_dec(&rx_stats.nPeerStructs);
7917 MUTEX_EXIT(&rx_peerHashTable_lock);
7920 for (i = 0; i < RX_MAX_SERVICES; i++) {
7922 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7924 for (i = 0; i < rx_hashTableSize; i++) {
7925 struct rx_connection *tc, *ntc;
7926 MUTEX_ENTER(&rx_connHashTable_lock);
7927 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7929 for (j = 0; j < RX_MAXCALLS; j++) {
7931 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7934 rxi_Free(tc, sizeof(*tc));
7936 MUTEX_EXIT(&rx_connHashTable_lock);
7939 MUTEX_ENTER(&freeSQEList_lock);
7941 while ((np = rx_FreeSQEList)) {
7942 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7943 MUTEX_DESTROY(&np->lock);
7944 rxi_Free(np, sizeof(*np));
7947 MUTEX_EXIT(&freeSQEList_lock);
7948 MUTEX_DESTROY(&freeSQEList_lock);
7949 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7950 MUTEX_DESTROY(&rx_connHashTable_lock);
7951 MUTEX_DESTROY(&rx_peerHashTable_lock);
7952 MUTEX_DESTROY(&rx_serverPool_lock);
7954 osi_Free(rx_connHashTable,
7955 rx_hashTableSize * sizeof(struct rx_connection *));
7956 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7958 UNPIN(rx_connHashTable,
7959 rx_hashTableSize * sizeof(struct rx_connection *));
7960 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7962 MUTEX_ENTER(&rx_quota_mutex);
7963 rxi_dataQuota = RX_MAX_QUOTA;
7964 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7965 MUTEX_EXIT(&rx_quota_mutex);
7971 * Routines to implement connection specific data.
7975 rx_KeyCreate(rx_destructor_t rtn)
7978 MUTEX_ENTER(&rxi_keyCreate_lock);
7979 key = rxi_keyCreate_counter++;
7980 rxi_keyCreate_destructor = (rx_destructor_t *)
7981 realloc((void *)rxi_keyCreate_destructor,
7982 (key + 1) * sizeof(rx_destructor_t));
7983 rxi_keyCreate_destructor[key] = rtn;
7984 MUTEX_EXIT(&rxi_keyCreate_lock);
7989 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7992 MUTEX_ENTER(&conn->conn_data_lock);
7993 if (!conn->specific) {
7994 conn->specific = malloc((key + 1) * sizeof(void *));
7995 for (i = 0; i < key; i++)
7996 conn->specific[i] = NULL;
7997 conn->nSpecific = key + 1;
7998 conn->specific[key] = ptr;
7999 } else if (key >= conn->nSpecific) {
8000 conn->specific = (void **)
8001 realloc(conn->specific, (key + 1) * sizeof(void *));
8002 for (i = conn->nSpecific; i < key; i++)
8003 conn->specific[i] = NULL;
8004 conn->nSpecific = key + 1;
8005 conn->specific[key] = ptr;
8007 if (conn->specific[key] && rxi_keyCreate_destructor[key])
8008 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
8009 conn->specific[key] = ptr;
8011 MUTEX_EXIT(&conn->conn_data_lock);
8015 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
8018 MUTEX_ENTER(&svc->svc_data_lock);
8019 if (!svc->specific) {
8020 svc->specific = malloc((key + 1) * sizeof(void *));
8021 for (i = 0; i < key; i++)
8022 svc->specific[i] = NULL;
8023 svc->nSpecific = key + 1;
8024 svc->specific[key] = ptr;
8025 } else if (key >= svc->nSpecific) {
8026 svc->specific = (void **)
8027 realloc(svc->specific, (key + 1) * sizeof(void *));
8028 for (i = svc->nSpecific; i < key; i++)
8029 svc->specific[i] = NULL;
8030 svc->nSpecific = key + 1;
8031 svc->specific[key] = ptr;
8033 if (svc->specific[key] && rxi_keyCreate_destructor[key])
8034 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
8035 svc->specific[key] = ptr;
8037 MUTEX_EXIT(&svc->svc_data_lock);
8041 rx_GetSpecific(struct rx_connection *conn, int key)
8044 MUTEX_ENTER(&conn->conn_data_lock);
8045 if (key >= conn->nSpecific)
8048 ptr = conn->specific[key];
8049 MUTEX_EXIT(&conn->conn_data_lock);
8054 rx_GetServiceSpecific(struct rx_service *svc, int key)
8057 MUTEX_ENTER(&svc->svc_data_lock);
8058 if (key >= svc->nSpecific)
8061 ptr = svc->specific[key];
8062 MUTEX_EXIT(&svc->svc_data_lock);
8067 #endif /* !KERNEL */
8070 * processStats is a queue used to store the statistics for the local
8071 * process. Its contents are similar to the contents of the rpcStats
8072 * queue on a rx_peer structure, but the actual data stored within
8073 * this queue contains totals across the lifetime of the process (assuming
8074 * the stats have not been reset) - unlike the per peer structures
8075 * which can come and go based upon the peer lifetime.
8078 static struct opr_queue processStats = { &processStats, &processStats };
8081 * peerStats is a queue used to store the statistics for all peer structs.
8082 * Its contents are the union of all the peer rpcStats queues.
8085 static struct opr_queue peerStats = { &peerStats, &peerStats };
8088 * rxi_monitor_processStats is used to turn process wide stat collection
8092 static int rxi_monitor_processStats = 0;
8095 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8098 static int rxi_monitor_peerStats = 0;
8102 rxi_ClearRPCOpStat(rx_function_entry_v1_p rpc_stat)
8104 rpc_stat->invocations = 0;
8105 rpc_stat->bytes_sent = 0;
8106 rpc_stat->bytes_rcvd = 0;
8107 rpc_stat->queue_time_sum.sec = 0;
8108 rpc_stat->queue_time_sum.usec = 0;
8109 rpc_stat->queue_time_sum_sqr.sec = 0;
8110 rpc_stat->queue_time_sum_sqr.usec = 0;
8111 rpc_stat->queue_time_min.sec = 9999999;
8112 rpc_stat->queue_time_min.usec = 9999999;
8113 rpc_stat->queue_time_max.sec = 0;
8114 rpc_stat->queue_time_max.usec = 0;
8115 rpc_stat->execution_time_sum.sec = 0;
8116 rpc_stat->execution_time_sum.usec = 0;
8117 rpc_stat->execution_time_sum_sqr.sec = 0;
8118 rpc_stat->execution_time_sum_sqr.usec = 0;
8119 rpc_stat->execution_time_min.sec = 9999999;
8120 rpc_stat->execution_time_min.usec = 9999999;
8121 rpc_stat->execution_time_max.sec = 0;
8122 rpc_stat->execution_time_max.usec = 0;
8126 * Given all of the information for a particular rpc
8127 * call, find or create (if requested) the stat structure for the rpc.
8130 * the queue of stats that will be updated with the new value
8132 * @param rxInterface
8133 * a unique number that identifies the rpc interface
8136 * the total number of functions in this interface. this is only
8137 * required if create is true
8140 * if true, this invocation was made to a server
8143 * the ip address of the remote host. this is only required if create
8144 * and addToPeerList are true
8147 * the port of the remote host. this is only required if create
8148 * and addToPeerList are true
8150 * @param addToPeerList
8151 * if != 0, add newly created stat to the global peer list
8154 * if a new stats structure is allocated, the counter will
8155 * be updated with the new number of allocated stat structures.
8156 * only required if create is true
8159 * if no stats structure exists, allocate one
8163 static rx_interface_stat_p
8164 rxi_FindRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8165 afs_uint32 totalFunc, int isServer, afs_uint32 remoteHost,
8166 afs_uint32 remotePort, int addToPeerList,
8167 unsigned int *counter, int create)
8169 rx_interface_stat_p rpc_stat = NULL;
8170 struct opr_queue *cursor;
8173 * See if there's already a structure for this interface
8176 for (opr_queue_Scan(stats, cursor)) {
8177 rpc_stat = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8179 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8180 && (rpc_stat->stats[0].remote_is_server == isServer))
8184 /* if they didn't ask us to create, we're done */
8186 if (opr_queue_IsEnd(stats, cursor))
8192 /* can't proceed without these */
8193 if (!totalFunc || !counter)
8197 * Didn't find a match so allocate a new structure and add it to the
8201 if (opr_queue_IsEnd(stats, cursor) || (rpc_stat == NULL)
8202 || (rpc_stat->stats[0].interfaceId != rxInterface)
8203 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8208 sizeof(rx_interface_stat_t) +
8209 totalFunc * sizeof(rx_function_entry_v1_t);
8211 rpc_stat = rxi_Alloc(space);
8212 if (rpc_stat == NULL)
8215 *counter += totalFunc;
8216 for (i = 0; i < totalFunc; i++) {
8217 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8218 rpc_stat->stats[i].remote_peer = remoteHost;
8219 rpc_stat->stats[i].remote_port = remotePort;
8220 rpc_stat->stats[i].remote_is_server = isServer;
8221 rpc_stat->stats[i].interfaceId = rxInterface;
8222 rpc_stat->stats[i].func_total = totalFunc;
8223 rpc_stat->stats[i].func_index = i;
8225 opr_queue_Prepend(stats, &rpc_stat->entry);
8226 if (addToPeerList) {
8227 opr_queue_Prepend(&peerStats, &rpc_stat->entryPeers);
8234 rx_ClearProcessRPCStats(afs_int32 rxInterface)
8236 rx_interface_stat_p rpc_stat;
8239 if (rxInterface == -1)
8242 MUTEX_ENTER(&rx_rpc_stats);
8243 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8246 totalFunc = rpc_stat->stats[0].func_total;
8247 for (i = 0; i < totalFunc; i++)
8248 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8250 MUTEX_EXIT(&rx_rpc_stats);
8255 rx_ClearPeerRPCStats(afs_int32 rxInterface, afs_uint32 peerHost, afs_uint16 peerPort)
8257 rx_interface_stat_p rpc_stat;
8259 struct rx_peer * peer;
8261 if (rxInterface == -1)
8264 peer = rxi_FindPeer(peerHost, peerPort, 0);
8268 MUTEX_ENTER(&rx_rpc_stats);
8269 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8272 totalFunc = rpc_stat->stats[0].func_total;
8273 for (i = 0; i < totalFunc; i++)
8274 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8276 MUTEX_EXIT(&rx_rpc_stats);
8281 rx_CopyProcessRPCStats(afs_uint64 op)
8283 rx_interface_stat_p rpc_stat;
8284 rx_function_entry_v1_p rpcop_stat =
8285 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8286 int currentFunc = (op & MAX_AFS_UINT32);
8287 afs_int32 rxInterface = (op >> 32);
8289 if (!rxi_monitor_processStats)
8292 if (rxInterface == -1)
8295 if (rpcop_stat == NULL)
8298 MUTEX_ENTER(&rx_rpc_stats);
8299 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8302 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8303 sizeof(rx_function_entry_v1_t));
8304 MUTEX_EXIT(&rx_rpc_stats);
8306 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8313 rx_CopyPeerRPCStats(afs_uint64 op, afs_uint32 peerHost, afs_uint16 peerPort)
8315 rx_interface_stat_p rpc_stat;
8316 rx_function_entry_v1_p rpcop_stat =
8317 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8318 int currentFunc = (op & MAX_AFS_UINT32);
8319 afs_int32 rxInterface = (op >> 32);
8320 struct rx_peer *peer;
8322 if (!rxi_monitor_peerStats)
8325 if (rxInterface == -1)
8328 if (rpcop_stat == NULL)
8331 peer = rxi_FindPeer(peerHost, peerPort, 0);
8335 MUTEX_ENTER(&rx_rpc_stats);
8336 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8339 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8340 sizeof(rx_function_entry_v1_t));
8341 MUTEX_EXIT(&rx_rpc_stats);
8343 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8350 rx_ReleaseRPCStats(void *stats)
8353 rxi_Free(stats, sizeof(rx_function_entry_v1_t));
8357 * Given all of the information for a particular rpc
8358 * call, create (if needed) and update the stat totals for the rpc.
8361 * the queue of stats that will be updated with the new value
8363 * @param rxInterface
8364 * a unique number that identifies the rpc interface
8366 * @param currentFunc
8367 * the index of the function being invoked
8370 * the total number of functions in this interface
8373 * the amount of time this function waited for a thread
8376 * the amount of time this function invocation took to execute
8379 * the number bytes sent by this invocation
8382 * the number bytes received by this invocation
8385 * if true, this invocation was made to a server
8388 * the ip address of the remote host
8391 * the port of the remote host
8393 * @param addToPeerList
8394 * if != 0, add newly created stat to the global peer list
8397 * if a new stats structure is allocated, the counter will
8398 * be updated with the new number of allocated stat structures
8403 rxi_AddRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8404 afs_uint32 currentFunc, afs_uint32 totalFunc,
8405 struct clock *queueTime, struct clock *execTime,
8406 afs_uint64 bytesSent, afs_uint64 bytesRcvd, int isServer,
8407 afs_uint32 remoteHost, afs_uint32 remotePort,
8408 int addToPeerList, unsigned int *counter)
8411 rx_interface_stat_p rpc_stat;
8413 rpc_stat = rxi_FindRpcStat(stats, rxInterface, totalFunc, isServer,
8414 remoteHost, remotePort, addToPeerList, counter,
8422 * Increment the stats for this function
8425 rpc_stat->stats[currentFunc].invocations++;
8426 rpc_stat->stats[currentFunc].bytes_sent += bytesSent;
8427 rpc_stat->stats[currentFunc].bytes_rcvd += bytesRcvd;
8428 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8429 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8430 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8431 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8433 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8434 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8436 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8437 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8439 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8440 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8442 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8443 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8451 rxi_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8452 afs_uint32 currentFunc, afs_uint32 totalFunc,
8453 struct clock *queueTime, struct clock *execTime,
8454 afs_uint64 bytesSent, afs_uint64 bytesRcvd,
8458 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8461 MUTEX_ENTER(&rx_rpc_stats);
8463 if (rxi_monitor_peerStats) {
8464 MUTEX_ENTER(&peer->peer_lock);
8465 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8466 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8467 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8468 MUTEX_EXIT(&peer->peer_lock);
8471 if (rxi_monitor_processStats) {
8472 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8473 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8474 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8477 MUTEX_EXIT(&rx_rpc_stats);
8481 * Increment the times and count for a particular rpc function.
8483 * Traditionally this call was invoked from rxgen stubs. Modern stubs
8484 * call rx_RecordCallStatistics instead, so the public version of this
8485 * function is left purely for legacy callers.
8488 * The peer who invoked the rpc
8490 * @param rxInterface
8491 * A unique number that identifies the rpc interface
8493 * @param currentFunc
8494 * The index of the function being invoked
8497 * The total number of functions in this interface
8500 * The amount of time this function waited for a thread
8503 * The amount of time this function invocation took to execute
8506 * The number bytes sent by this invocation
8509 * The number bytes received by this invocation
8512 * If true, this invocation was made to a server
8516 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8517 afs_uint32 currentFunc, afs_uint32 totalFunc,
8518 struct clock *queueTime, struct clock *execTime,
8519 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8525 sent64 = ((afs_uint64)bytesSent->high << 32) + bytesSent->low;
8526 rcvd64 = ((afs_uint64)bytesRcvd->high << 32) + bytesRcvd->low;
8528 rxi_IncrementTimeAndCount(peer, rxInterface, currentFunc, totalFunc,
8529 queueTime, execTime, sent64, rcvd64,
8536 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8540 * IN callerVersion - the rpc stat version of the caller.
8542 * IN count - the number of entries to marshall.
8544 * IN stats - pointer to stats to be marshalled.
8546 * OUT ptr - Where to store the marshalled data.
8553 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8554 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8560 * We only support the first version
8562 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8563 *(ptr++) = stats->remote_peer;
8564 *(ptr++) = stats->remote_port;
8565 *(ptr++) = stats->remote_is_server;
8566 *(ptr++) = stats->interfaceId;
8567 *(ptr++) = stats->func_total;
8568 *(ptr++) = stats->func_index;
8569 *(ptr++) = stats->invocations >> 32;
8570 *(ptr++) = stats->invocations & MAX_AFS_UINT32;
8571 *(ptr++) = stats->bytes_sent >> 32;
8572 *(ptr++) = stats->bytes_sent & MAX_AFS_UINT32;
8573 *(ptr++) = stats->bytes_rcvd >> 32;
8574 *(ptr++) = stats->bytes_rcvd & MAX_AFS_UINT32;
8575 *(ptr++) = stats->queue_time_sum.sec;
8576 *(ptr++) = stats->queue_time_sum.usec;
8577 *(ptr++) = stats->queue_time_sum_sqr.sec;
8578 *(ptr++) = stats->queue_time_sum_sqr.usec;
8579 *(ptr++) = stats->queue_time_min.sec;
8580 *(ptr++) = stats->queue_time_min.usec;
8581 *(ptr++) = stats->queue_time_max.sec;
8582 *(ptr++) = stats->queue_time_max.usec;
8583 *(ptr++) = stats->execution_time_sum.sec;
8584 *(ptr++) = stats->execution_time_sum.usec;
8585 *(ptr++) = stats->execution_time_sum_sqr.sec;
8586 *(ptr++) = stats->execution_time_sum_sqr.usec;
8587 *(ptr++) = stats->execution_time_min.sec;
8588 *(ptr++) = stats->execution_time_min.usec;
8589 *(ptr++) = stats->execution_time_max.sec;
8590 *(ptr++) = stats->execution_time_max.usec;
8596 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8601 * IN callerVersion - the rpc stat version of the caller
8603 * OUT myVersion - the rpc stat version of this function
8605 * OUT clock_sec - local time seconds
8607 * OUT clock_usec - local time microseconds
8609 * OUT allocSize - the number of bytes allocated to contain stats
8611 * OUT statCount - the number stats retrieved from this process.
8613 * OUT stats - the actual stats retrieved from this process.
8617 * Returns void. If successful, stats will != NULL.
8621 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8622 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8623 size_t * allocSize, afs_uint32 * statCount,
8624 afs_uint32 ** stats)
8634 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8637 * Check to see if stats are enabled
8640 MUTEX_ENTER(&rx_rpc_stats);
8641 if (!rxi_monitor_processStats) {
8642 MUTEX_EXIT(&rx_rpc_stats);
8646 clock_GetTime(&now);
8647 *clock_sec = now.sec;
8648 *clock_usec = now.usec;
8651 * Allocate the space based upon the caller version
8653 * If the client is at an older version than we are,
8654 * we return the statistic data in the older data format, but
8655 * we still return our version number so the client knows we
8656 * are maintaining more data than it can retrieve.
8659 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8660 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8661 *statCount = rxi_rpc_process_stat_cnt;
8664 * This can't happen yet, but in the future version changes
8665 * can be handled by adding additional code here
8669 if (space > (size_t) 0) {
8671 ptr = *stats = rxi_Alloc(space);
8674 struct opr_queue *cursor;
8676 for (opr_queue_Scan(&processStats, cursor)) {
8677 struct rx_interface_stat *rpc_stat =
8678 opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8680 * Copy the data based upon the caller version
8682 rx_MarshallProcessRPCStats(callerVersion,
8683 rpc_stat->stats[0].func_total,
8684 rpc_stat->stats, &ptr);
8690 MUTEX_EXIT(&rx_rpc_stats);
8695 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8699 * IN callerVersion - the rpc stat version of the caller
8701 * OUT myVersion - the rpc stat version of this function
8703 * OUT clock_sec - local time seconds
8705 * OUT clock_usec - local time microseconds
8707 * OUT allocSize - the number of bytes allocated to contain stats
8709 * OUT statCount - the number of stats retrieved from the individual
8712 * OUT stats - the actual stats retrieved from the individual peer structures.
8716 * Returns void. If successful, stats will != NULL.
8720 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8721 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8722 size_t * allocSize, afs_uint32 * statCount,
8723 afs_uint32 ** stats)
8733 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8736 * Check to see if stats are enabled
8739 MUTEX_ENTER(&rx_rpc_stats);
8740 if (!rxi_monitor_peerStats) {
8741 MUTEX_EXIT(&rx_rpc_stats);
8745 clock_GetTime(&now);
8746 *clock_sec = now.sec;
8747 *clock_usec = now.usec;
8750 * Allocate the space based upon the caller version
8752 * If the client is at an older version than we are,
8753 * we return the statistic data in the older data format, but
8754 * we still return our version number so the client knows we
8755 * are maintaining more data than it can retrieve.
8758 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8759 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8760 *statCount = rxi_rpc_peer_stat_cnt;
8763 * This can't happen yet, but in the future version changes
8764 * can be handled by adding additional code here
8768 if (space > (size_t) 0) {
8770 ptr = *stats = rxi_Alloc(space);
8773 struct opr_queue *cursor;
8775 for (opr_queue_Scan(&peerStats, cursor)) {
8776 struct rx_interface_stat *rpc_stat
8777 = opr_queue_Entry(cursor, struct rx_interface_stat,
8781 * Copy the data based upon the caller version
8783 rx_MarshallProcessRPCStats(callerVersion,
8784 rpc_stat->stats[0].func_total,
8785 rpc_stat->stats, &ptr);
8791 MUTEX_EXIT(&rx_rpc_stats);
8796 * rx_FreeRPCStats - free memory allocated by
8797 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8801 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8802 * rx_RetrievePeerRPCStats
8804 * IN allocSize - the number of bytes in stats.
8812 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8814 rxi_Free(stats, allocSize);
8818 * rx_queryProcessRPCStats - see if process rpc stat collection is
8819 * currently enabled.
8825 * Returns 0 if stats are not enabled != 0 otherwise
8829 rx_queryProcessRPCStats(void)
8832 MUTEX_ENTER(&rx_rpc_stats);
8833 rc = rxi_monitor_processStats;
8834 MUTEX_EXIT(&rx_rpc_stats);
8839 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8845 * Returns 0 if stats are not enabled != 0 otherwise
8849 rx_queryPeerRPCStats(void)
8852 MUTEX_ENTER(&rx_rpc_stats);
8853 rc = rxi_monitor_peerStats;
8854 MUTEX_EXIT(&rx_rpc_stats);
8859 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8869 rx_enableProcessRPCStats(void)
8871 MUTEX_ENTER(&rx_rpc_stats);
8872 rx_enable_stats = 1;
8873 rxi_monitor_processStats = 1;
8874 MUTEX_EXIT(&rx_rpc_stats);
8878 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8888 rx_enablePeerRPCStats(void)
8890 MUTEX_ENTER(&rx_rpc_stats);
8891 rx_enable_stats = 1;
8892 rxi_monitor_peerStats = 1;
8893 MUTEX_EXIT(&rx_rpc_stats);
8897 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8907 rx_disableProcessRPCStats(void)
8909 struct opr_queue *cursor, *store;
8912 MUTEX_ENTER(&rx_rpc_stats);
8915 * Turn off process statistics and if peer stats is also off, turn
8919 rxi_monitor_processStats = 0;
8920 if (rxi_monitor_peerStats == 0) {
8921 rx_enable_stats = 0;
8924 for (opr_queue_ScanSafe(&processStats, cursor, store)) {
8925 unsigned int num_funcs = 0;
8926 struct rx_interface_stat *rpc_stat
8927 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8929 opr_queue_Remove(&rpc_stat->entry);
8931 num_funcs = rpc_stat->stats[0].func_total;
8933 sizeof(rx_interface_stat_t) +
8934 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8936 rxi_Free(rpc_stat, space);
8937 rxi_rpc_process_stat_cnt -= num_funcs;
8939 MUTEX_EXIT(&rx_rpc_stats);
8943 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8953 rx_disablePeerRPCStats(void)
8955 struct rx_peer **peer_ptr, **peer_end;
8959 * Turn off peer statistics and if process stats is also off, turn
8963 rxi_monitor_peerStats = 0;
8964 if (rxi_monitor_processStats == 0) {
8965 rx_enable_stats = 0;
8968 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8969 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8971 struct rx_peer *peer, *next, *prev;
8973 MUTEX_ENTER(&rx_peerHashTable_lock);
8974 MUTEX_ENTER(&rx_rpc_stats);
8975 for (prev = peer = *peer_ptr; peer; peer = next) {
8977 code = MUTEX_TRYENTER(&peer->peer_lock);
8980 struct opr_queue *cursor, *store;
8982 if (prev == *peer_ptr) {
8993 MUTEX_EXIT(&rx_peerHashTable_lock);
8995 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
8996 unsigned int num_funcs = 0;
8997 struct rx_interface_stat *rpc_stat
8998 = opr_queue_Entry(cursor, struct rx_interface_stat,
9001 opr_queue_Remove(&rpc_stat->entry);
9002 opr_queue_Remove(&rpc_stat->entryPeers);
9003 num_funcs = rpc_stat->stats[0].func_total;
9005 sizeof(rx_interface_stat_t) +
9006 rpc_stat->stats[0].func_total *
9007 sizeof(rx_function_entry_v1_t);
9009 rxi_Free(rpc_stat, space);
9010 rxi_rpc_peer_stat_cnt -= num_funcs;
9012 MUTEX_EXIT(&peer->peer_lock);
9014 MUTEX_ENTER(&rx_peerHashTable_lock);
9024 MUTEX_EXIT(&rx_rpc_stats);
9025 MUTEX_EXIT(&rx_peerHashTable_lock);
9030 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
9035 * IN clearFlag - flag indicating which stats to clear
9043 rx_clearProcessRPCStats(afs_uint32 clearFlag)
9045 struct opr_queue *cursor;
9047 MUTEX_ENTER(&rx_rpc_stats);
9049 for (opr_queue_Scan(&processStats, cursor)) {
9050 unsigned int num_funcs = 0, i;
9051 struct rx_interface_stat *rpc_stat
9052 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
9054 num_funcs = rpc_stat->stats[0].func_total;
9055 for (i = 0; i < num_funcs; i++) {
9056 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9057 rpc_stat->stats[i].invocations = 0;
9059 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9060 rpc_stat->stats[i].bytes_sent = 0;
9062 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9063 rpc_stat->stats[i].bytes_rcvd = 0;
9065 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9066 rpc_stat->stats[i].queue_time_sum.sec = 0;
9067 rpc_stat->stats[i].queue_time_sum.usec = 0;
9069 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9070 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9071 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9073 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9074 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9075 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9077 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9078 rpc_stat->stats[i].queue_time_max.sec = 0;
9079 rpc_stat->stats[i].queue_time_max.usec = 0;
9081 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9082 rpc_stat->stats[i].execution_time_sum.sec = 0;
9083 rpc_stat->stats[i].execution_time_sum.usec = 0;
9085 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9086 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9087 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9089 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9090 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9091 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9093 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9094 rpc_stat->stats[i].execution_time_max.sec = 0;
9095 rpc_stat->stats[i].execution_time_max.usec = 0;
9100 MUTEX_EXIT(&rx_rpc_stats);
9104 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
9109 * IN clearFlag - flag indicating which stats to clear
9117 rx_clearPeerRPCStats(afs_uint32 clearFlag)
9119 struct opr_queue *cursor;
9121 MUTEX_ENTER(&rx_rpc_stats);
9123 for (opr_queue_Scan(&peerStats, cursor)) {
9124 unsigned int num_funcs, i;
9125 struct rx_interface_stat *rpc_stat
9126 = opr_queue_Entry(cursor, struct rx_interface_stat, entryPeers);
9128 num_funcs = rpc_stat->stats[0].func_total;
9129 for (i = 0; i < num_funcs; i++) {
9130 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9131 rpc_stat->stats[i].invocations = 0;
9133 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9134 rpc_stat->stats[i].bytes_sent = 0;
9136 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9137 rpc_stat->stats[i].bytes_rcvd = 0;
9139 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9140 rpc_stat->stats[i].queue_time_sum.sec = 0;
9141 rpc_stat->stats[i].queue_time_sum.usec = 0;
9143 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9144 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9145 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9147 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9148 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9149 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9151 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9152 rpc_stat->stats[i].queue_time_max.sec = 0;
9153 rpc_stat->stats[i].queue_time_max.usec = 0;
9155 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9156 rpc_stat->stats[i].execution_time_sum.sec = 0;
9157 rpc_stat->stats[i].execution_time_sum.usec = 0;
9159 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9160 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9161 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9163 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9164 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9165 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9167 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9168 rpc_stat->stats[i].execution_time_max.sec = 0;
9169 rpc_stat->stats[i].execution_time_max.usec = 0;
9174 MUTEX_EXIT(&rx_rpc_stats);
9178 * rxi_rxstat_userok points to a routine that returns 1 if the caller
9179 * is authorized to enable/disable/clear RX statistics.
9181 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
9184 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
9186 rxi_rxstat_userok = proc;
9190 rx_RxStatUserOk(struct rx_call *call)
9192 if (!rxi_rxstat_userok)
9194 return rxi_rxstat_userok(call);
9199 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
9200 * function in the MSVC runtime DLL (msvcrt.dll).
9202 * Note: the system serializes calls to this function.
9205 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
9206 DWORD reason, /* reason function is being called */
9207 LPVOID reserved) /* reserved for future use */
9210 case DLL_PROCESS_ATTACH:
9211 /* library is being attached to a process */
9215 case DLL_PROCESS_DETACH:
9222 #endif /* AFS_NT40_ENV */
9225 int rx_DumpCalls(FILE *outputFile, char *cookie)
9227 #ifdef RXDEBUG_PACKET
9228 #ifdef KDUMP_RX_LOCK
9229 struct rx_call_rx_lock *c;
9236 #define RXDPRINTF sprintf
9237 #define RXDPRINTOUT output
9239 #define RXDPRINTF fprintf
9240 #define RXDPRINTOUT outputFile
9243 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9245 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9248 for (c = rx_allCallsp; c; c = c->allNextp) {
9249 u_short rqc, tqc, iovqc;
9251 MUTEX_ENTER(&c->lock);
9252 rqc = opr_queue_Count(&c->rq);
9253 tqc = opr_queue_Count(&c->tq);
9254 iovqc = opr_queue_Count(&c->app.iovq);
9256 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, "
9257 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9258 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9259 "resendEvent=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9260 "lastSendTime=%u, lastRecvTime=%u"
9261 #ifdef RX_ENABLE_LOCKS
9264 #ifdef RX_REFCOUNT_CHECK
9265 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9266 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9269 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,
9270 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9271 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9272 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9273 c->resendEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9274 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime
9275 #ifdef RX_ENABLE_LOCKS
9276 , (afs_uint32)c->refCount
9278 #ifdef RX_REFCOUNT_CHECK
9279 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9282 MUTEX_EXIT(&c->lock);
9285 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9288 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9290 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9292 #endif /* RXDEBUG_PACKET */