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(&conn->conn_data_lock);
1255 MUTEX_EXIT(&rx_refcnt_mutex);
1256 MUTEX_EXIT(&conn->conn_data_lock);
1258 /* Check for extant references to this connection */
1259 MUTEX_ENTER(&conn->conn_call_lock);
1260 for (i = 0; i < RX_MAXCALLS; i++) {
1261 struct rx_call *call = conn->call[i];
1264 if (conn->type == RX_CLIENT_CONNECTION) {
1265 MUTEX_ENTER(&call->lock);
1266 if (call->delayedAckEvent) {
1267 /* Push the final acknowledgment out now--there
1268 * won't be a subsequent call to acknowledge the
1269 * last reply packets */
1270 rxi_CancelDelayedAckEvent(call);
1271 if (call->state == RX_STATE_PRECALL
1272 || call->state == RX_STATE_ACTIVE) {
1273 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1278 MUTEX_EXIT(&call->lock);
1282 MUTEX_EXIT(&conn->conn_call_lock);
1284 #ifdef RX_ENABLE_LOCKS
1286 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1287 MUTEX_EXIT(&conn->conn_data_lock);
1289 /* Someone is accessing a packet right now. */
1293 #endif /* RX_ENABLE_LOCKS */
1296 /* Don't destroy the connection if there are any call
1297 * structures still in use */
1298 MUTEX_ENTER(&conn->conn_data_lock);
1299 conn->flags |= RX_CONN_DESTROY_ME;
1300 MUTEX_EXIT(&conn->conn_data_lock);
1305 /* Remove from connection hash table before proceeding */
1307 &rx_connHashTable[CONN_HASH
1308 (peer->host, peer->port, conn->cid, conn->epoch,
1310 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1311 if (*conn_ptr == conn) {
1312 *conn_ptr = conn->next;
1316 /* if the conn that we are destroying was the last connection, then we
1317 * clear rxLastConn as well */
1318 if (rxLastConn == conn)
1321 /* Make sure the connection is completely reset before deleting it. */
1323 * Pending events hold a refcount, so we can't get here if they are
1325 osi_Assert(conn->challengeEvent == NULL);
1326 osi_Assert(conn->delayedAbortEvent == NULL);
1327 osi_Assert(conn->natKeepAliveEvent == NULL);
1328 osi_Assert(conn->checkReachEvent == NULL);
1330 /* Add the connection to the list of destroyed connections that
1331 * need to be cleaned up. This is necessary to avoid deadlocks
1332 * in the routines we call to inform others that this connection is
1333 * being destroyed. */
1334 conn->next = rx_connCleanup_list;
1335 rx_connCleanup_list = conn;
1338 /* Externally available version */
1340 rx_DestroyConnection(struct rx_connection *conn)
1345 rxi_DestroyConnection(conn);
1350 rx_GetConnection(struct rx_connection *conn)
1355 MUTEX_ENTER(&rx_refcnt_mutex);
1357 MUTEX_EXIT(&rx_refcnt_mutex);
1361 #ifdef RX_ENABLE_LOCKS
1362 /* Wait for the transmit queue to no longer be busy.
1363 * requires the call->lock to be held */
1365 rxi_WaitforTQBusy(struct rx_call *call) {
1366 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1367 call->flags |= RX_CALL_TQ_WAIT;
1369 MUTEX_ASSERT(&call->lock);
1370 CV_WAIT(&call->cv_tq, &call->lock);
1372 if (call->tqWaiters == 0) {
1373 call->flags &= ~RX_CALL_TQ_WAIT;
1380 rxi_WakeUpTransmitQueue(struct rx_call *call)
1382 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1383 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1384 call, call->tqWaiters, call->flags));
1385 #ifdef RX_ENABLE_LOCKS
1386 MUTEX_ASSERT(&call->lock);
1387 CV_BROADCAST(&call->cv_tq);
1388 #else /* RX_ENABLE_LOCKS */
1389 osi_rxWakeup(&call->tq);
1390 #endif /* RX_ENABLE_LOCKS */
1394 /* Start a new rx remote procedure call, on the specified connection.
1395 * If wait is set to 1, wait for a free call channel; otherwise return
1396 * 0. Maxtime gives the maximum number of seconds this call may take,
1397 * after rx_NewCall returns. After this time interval, a call to any
1398 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1399 * For fine grain locking, we hold the conn_call_lock in order to
1400 * to ensure that we don't get signalle after we found a call in an active
1401 * state and before we go to sleep.
1404 rx_NewCall(struct rx_connection *conn)
1406 int i, wait, ignoreBusy = 1;
1407 struct rx_call *call;
1408 struct clock queueTime;
1409 afs_uint32 leastBusy = 0;
1413 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1416 clock_GetTime(&queueTime);
1418 * Check if there are others waiting for a new call.
1419 * If so, let them go first to avoid starving them.
1420 * This is a fairly simple scheme, and might not be
1421 * a complete solution for large numbers of waiters.
1423 * makeCallWaiters keeps track of the number of
1424 * threads waiting to make calls and the
1425 * RX_CONN_MAKECALL_WAITING flag bit is used to
1426 * indicate that there are indeed calls waiting.
1427 * The flag is set when the waiter is incremented.
1428 * It is only cleared when makeCallWaiters is 0.
1429 * This prevents us from accidently destroying the
1430 * connection while it is potentially about to be used.
1432 MUTEX_ENTER(&conn->conn_call_lock);
1433 MUTEX_ENTER(&conn->conn_data_lock);
1434 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1435 conn->flags |= RX_CONN_MAKECALL_WAITING;
1436 conn->makeCallWaiters++;
1437 MUTEX_EXIT(&conn->conn_data_lock);
1439 #ifdef RX_ENABLE_LOCKS
1440 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1444 MUTEX_ENTER(&conn->conn_data_lock);
1445 conn->makeCallWaiters--;
1446 if (conn->makeCallWaiters == 0)
1447 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1450 /* We are now the active thread in rx_NewCall */
1451 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1452 MUTEX_EXIT(&conn->conn_data_lock);
1457 for (i = 0; i < RX_MAXCALLS; i++) {
1458 call = conn->call[i];
1460 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1461 /* we're not ignoring busy call slots; only look at the
1462 * call slot that is the "least" busy */
1466 if (call->state == RX_STATE_DALLY) {
1467 MUTEX_ENTER(&call->lock);
1468 if (call->state == RX_STATE_DALLY) {
1469 if (ignoreBusy && conn->lastBusy[i]) {
1470 /* if we're ignoring busy call slots, skip any ones that
1471 * have lastBusy set */
1472 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1473 leastBusy = conn->lastBusy[i];
1475 MUTEX_EXIT(&call->lock);
1480 * We are setting the state to RX_STATE_RESET to
1481 * ensure that no one else will attempt to use this
1482 * call once we drop the conn->conn_call_lock and
1483 * call->lock. We must drop the conn->conn_call_lock
1484 * before calling rxi_ResetCall because the process
1485 * of clearing the transmit queue can block for an
1486 * extended period of time. If we block while holding
1487 * the conn->conn_call_lock, then all rx_EndCall
1488 * processing will block as well. This has a detrimental
1489 * effect on overall system performance.
1491 call->state = RX_STATE_RESET;
1492 (*call->callNumber)++;
1493 MUTEX_EXIT(&conn->conn_call_lock);
1494 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1495 rxi_ResetCall(call, 0);
1496 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1500 * If we failed to be able to safely obtain the
1501 * conn->conn_call_lock we will have to drop the
1502 * call->lock to avoid a deadlock. When the call->lock
1503 * is released the state of the call can change. If it
1504 * is no longer RX_STATE_RESET then some other thread is
1507 MUTEX_EXIT(&call->lock);
1508 MUTEX_ENTER(&conn->conn_call_lock);
1509 MUTEX_ENTER(&call->lock);
1511 if (call->state == RX_STATE_RESET)
1515 * If we get here it means that after dropping
1516 * the conn->conn_call_lock and call->lock that
1517 * the call is no longer ours. If we can't find
1518 * a free call in the remaining slots we should
1519 * not go immediately to RX_CONN_MAKECALL_WAITING
1520 * because by dropping the conn->conn_call_lock
1521 * we have given up synchronization with rx_EndCall.
1522 * Instead, cycle through one more time to see if
1523 * we can find a call that can call our own.
1525 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1528 MUTEX_EXIT(&call->lock);
1531 if (ignoreBusy && conn->lastBusy[i]) {
1532 /* if we're ignoring busy call slots, skip any ones that
1533 * have lastBusy set */
1534 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1535 leastBusy = conn->lastBusy[i];
1540 /* rxi_NewCall returns with mutex locked */
1541 call = rxi_NewCall(conn, i);
1542 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1546 if (i < RX_MAXCALLS) {
1547 conn->lastBusy[i] = 0;
1552 if (leastBusy && ignoreBusy) {
1553 /* we didn't find a useable call slot, but we did see at least one
1554 * 'busy' slot; look again and only use a slot with the 'least
1560 MUTEX_ENTER(&conn->conn_data_lock);
1561 conn->flags |= RX_CONN_MAKECALL_WAITING;
1562 conn->makeCallWaiters++;
1563 MUTEX_EXIT(&conn->conn_data_lock);
1565 #ifdef RX_ENABLE_LOCKS
1566 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1570 MUTEX_ENTER(&conn->conn_data_lock);
1571 conn->makeCallWaiters--;
1572 if (conn->makeCallWaiters == 0)
1573 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1574 MUTEX_EXIT(&conn->conn_data_lock);
1576 /* Client is initially in send mode */
1577 call->state = RX_STATE_ACTIVE;
1578 call->error = conn->error;
1580 call->app.mode = RX_MODE_ERROR;
1582 call->app.mode = RX_MODE_SENDING;
1584 #ifdef AFS_RXERRQ_ENV
1585 /* remember how many network errors the peer has when we started, so if
1586 * more errors are encountered after the call starts, we know the other endpoint won't be
1587 * responding to us */
1588 call->neterr_gen = rx_atomic_read(&conn->peer->neterrs);
1591 /* remember start time for call in case we have hard dead time limit */
1592 call->queueTime = queueTime;
1593 clock_GetTime(&call->startTime);
1594 call->app.bytesSent = 0;
1595 call->app.bytesRcvd = 0;
1597 /* Turn on busy protocol. */
1598 rxi_KeepAliveOn(call);
1600 /* Attempt MTU discovery */
1601 rxi_GrowMTUOn(call);
1604 * We are no longer the active thread in rx_NewCall
1606 MUTEX_ENTER(&conn->conn_data_lock);
1607 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1608 MUTEX_EXIT(&conn->conn_data_lock);
1611 * Wake up anyone else who might be giving us a chance to
1612 * run (see code above that avoids resource starvation).
1614 #ifdef RX_ENABLE_LOCKS
1615 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1616 osi_Panic("rx_NewCall call about to be used without an empty tq");
1619 CV_BROADCAST(&conn->conn_call_cv);
1623 MUTEX_EXIT(&conn->conn_call_lock);
1624 MUTEX_EXIT(&call->lock);
1627 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1632 rxi_HasActiveCalls(struct rx_connection *aconn)
1635 struct rx_call *tcall;
1639 for (i = 0; i < RX_MAXCALLS; i++) {
1640 if ((tcall = aconn->call[i])) {
1641 if ((tcall->state == RX_STATE_ACTIVE)
1642 || (tcall->state == RX_STATE_PRECALL)) {
1653 rxi_GetCallNumberVector(struct rx_connection *aconn,
1654 afs_int32 * aint32s)
1657 struct rx_call *tcall;
1661 MUTEX_ENTER(&aconn->conn_call_lock);
1662 for (i = 0; i < RX_MAXCALLS; i++) {
1663 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1664 aint32s[i] = aconn->callNumber[i] + 1;
1666 aint32s[i] = aconn->callNumber[i];
1668 MUTEX_EXIT(&aconn->conn_call_lock);
1674 rxi_SetCallNumberVector(struct rx_connection *aconn,
1675 afs_int32 * aint32s)
1678 struct rx_call *tcall;
1682 MUTEX_ENTER(&aconn->conn_call_lock);
1683 for (i = 0; i < RX_MAXCALLS; i++) {
1684 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1685 aconn->callNumber[i] = aint32s[i] - 1;
1687 aconn->callNumber[i] = aint32s[i];
1689 MUTEX_EXIT(&aconn->conn_call_lock);
1694 /* Advertise a new service. A service is named locally by a UDP port
1695 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1698 char *serviceName; Name for identification purposes (e.g. the
1699 service name might be used for probing for
1702 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1703 char *serviceName, struct rx_securityClass **securityObjects,
1704 int nSecurityObjects,
1705 afs_int32(*serviceProc) (struct rx_call * acall))
1707 osi_socket socket = OSI_NULLSOCKET;
1708 struct rx_service *tservice;
1714 if (serviceId == 0) {
1716 "rx_NewService: service id for service %s is not non-zero.\n",
1723 "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",
1731 tservice = rxi_AllocService();
1734 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1736 for (i = 0; i < RX_MAX_SERVICES; i++) {
1737 struct rx_service *service = rx_services[i];
1739 if (port == service->servicePort && host == service->serviceHost) {
1740 if (service->serviceId == serviceId) {
1741 /* The identical service has already been
1742 * installed; if the caller was intending to
1743 * change the security classes used by this
1744 * service, he/she loses. */
1746 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1747 serviceName, serviceId, service->serviceName);
1749 rxi_FreeService(tservice);
1752 /* Different service, same port: re-use the socket
1753 * which is bound to the same port */
1754 socket = service->socket;
1757 if (socket == OSI_NULLSOCKET) {
1758 /* If we don't already have a socket (from another
1759 * service on same port) get a new one */
1760 socket = rxi_GetHostUDPSocket(host, port);
1761 if (socket == OSI_NULLSOCKET) {
1763 rxi_FreeService(tservice);
1768 service->socket = socket;
1769 service->serviceHost = host;
1770 service->servicePort = port;
1771 service->serviceId = serviceId;
1772 service->serviceName = serviceName;
1773 service->nSecurityObjects = nSecurityObjects;
1774 service->securityObjects = securityObjects;
1775 service->minProcs = 0;
1776 service->maxProcs = 1;
1777 service->idleDeadTime = 60;
1778 service->connDeadTime = rx_connDeadTime;
1779 service->executeRequestProc = serviceProc;
1780 service->checkReach = 0;
1781 service->nSpecific = 0;
1782 service->specific = NULL;
1783 rx_services[i] = service; /* not visible until now */
1789 rxi_FreeService(tservice);
1790 (osi_Msg "rx_NewService: cannot support > %d services\n",
1795 /* Set configuration options for all of a service's security objects */
1798 rx_SetSecurityConfiguration(struct rx_service *service,
1799 rx_securityConfigVariables type,
1803 for (i = 0; i<service->nSecurityObjects; i++) {
1804 if (service->securityObjects[i]) {
1805 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1813 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1814 struct rx_securityClass **securityObjects, int nSecurityObjects,
1815 afs_int32(*serviceProc) (struct rx_call * acall))
1817 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1820 /* Generic request processing loop. This routine should be called
1821 * by the implementation dependent rx_ServerProc. If socketp is
1822 * non-null, it will be set to the file descriptor that this thread
1823 * is now listening on. If socketp is null, this routine will never
1826 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1828 struct rx_call *call;
1830 struct rx_service *tservice = NULL;
1837 call = rx_GetCall(threadID, tservice, socketp);
1838 if (socketp && *socketp != OSI_NULLSOCKET) {
1839 /* We are now a listener thread */
1845 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1846 #ifdef RX_ENABLE_LOCKS
1848 #endif /* RX_ENABLE_LOCKS */
1849 afs_termState = AFSOP_STOP_AFS;
1850 afs_osi_Wakeup(&afs_termState);
1851 #ifdef RX_ENABLE_LOCKS
1853 #endif /* RX_ENABLE_LOCKS */
1858 /* if server is restarting( typically smooth shutdown) then do not
1859 * allow any new calls.
1862 if (rx_tranquil && (call != NULL)) {
1866 MUTEX_ENTER(&call->lock);
1868 rxi_CallError(call, RX_RESTARTING);
1869 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1871 MUTEX_EXIT(&call->lock);
1876 tservice = call->conn->service;
1878 if (tservice->beforeProc)
1879 (*tservice->beforeProc) (call);
1881 code = tservice->executeRequestProc(call);
1883 if (tservice->afterProc)
1884 (*tservice->afterProc) (call, code);
1886 rx_EndCall(call, code);
1888 if (tservice->postProc)
1889 (*tservice->postProc) (code);
1891 if (rx_stats_active) {
1892 MUTEX_ENTER(&rx_stats_mutex);
1894 MUTEX_EXIT(&rx_stats_mutex);
1901 rx_WakeupServerProcs(void)
1903 struct rx_serverQueueEntry *np, *tqp;
1904 struct opr_queue *cursor;
1908 MUTEX_ENTER(&rx_serverPool_lock);
1910 #ifdef RX_ENABLE_LOCKS
1911 if (rx_waitForPacket)
1912 CV_BROADCAST(&rx_waitForPacket->cv);
1913 #else /* RX_ENABLE_LOCKS */
1914 if (rx_waitForPacket)
1915 osi_rxWakeup(rx_waitForPacket);
1916 #endif /* RX_ENABLE_LOCKS */
1917 MUTEX_ENTER(&freeSQEList_lock);
1918 for (np = rx_FreeSQEList; np; np = tqp) {
1919 tqp = *(struct rx_serverQueueEntry **)np;
1920 #ifdef RX_ENABLE_LOCKS
1921 CV_BROADCAST(&np->cv);
1922 #else /* RX_ENABLE_LOCKS */
1924 #endif /* RX_ENABLE_LOCKS */
1926 MUTEX_EXIT(&freeSQEList_lock);
1927 for (opr_queue_Scan(&rx_idleServerQueue, cursor)) {
1928 np = opr_queue_Entry(cursor, struct rx_serverQueueEntry, entry);
1929 #ifdef RX_ENABLE_LOCKS
1930 CV_BROADCAST(&np->cv);
1931 #else /* RX_ENABLE_LOCKS */
1933 #endif /* RX_ENABLE_LOCKS */
1935 MUTEX_EXIT(&rx_serverPool_lock);
1940 * One thing that seems to happen is that all the server threads get
1941 * tied up on some empty or slow call, and then a whole bunch of calls
1942 * arrive at once, using up the packet pool, so now there are more
1943 * empty calls. The most critical resources here are server threads
1944 * and the free packet pool. The "doreclaim" code seems to help in
1945 * general. I think that eventually we arrive in this state: there
1946 * are lots of pending calls which do have all their packets present,
1947 * so they won't be reclaimed, are multi-packet calls, so they won't
1948 * be scheduled until later, and thus are tying up most of the free
1949 * packet pool for a very long time.
1951 * 1. schedule multi-packet calls if all the packets are present.
1952 * Probably CPU-bound operation, useful to return packets to pool.
1953 * Do what if there is a full window, but the last packet isn't here?
1954 * 3. preserve one thread which *only* runs "best" calls, otherwise
1955 * it sleeps and waits for that type of call.
1956 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1957 * the current dataquota business is badly broken. The quota isn't adjusted
1958 * to reflect how many packets are presently queued for a running call.
1959 * So, when we schedule a queued call with a full window of packets queued
1960 * up for it, that *should* free up a window full of packets for other 2d-class
1961 * calls to be able to use from the packet pool. But it doesn't.
1963 * NB. Most of the time, this code doesn't run -- since idle server threads
1964 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1965 * as a new call arrives.
1967 /* Sleep until a call arrives. Returns a pointer to the call, ready
1968 * for an rx_Read. */
1969 #ifdef RX_ENABLE_LOCKS
1971 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1973 struct rx_serverQueueEntry *sq;
1974 struct rx_call *call = (struct rx_call *)0;
1975 struct rx_service *service = NULL;
1977 MUTEX_ENTER(&freeSQEList_lock);
1979 if ((sq = rx_FreeSQEList)) {
1980 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1981 MUTEX_EXIT(&freeSQEList_lock);
1982 } else { /* otherwise allocate a new one and return that */
1983 MUTEX_EXIT(&freeSQEList_lock);
1984 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1985 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1986 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1989 MUTEX_ENTER(&rx_serverPool_lock);
1990 if (cur_service != NULL) {
1991 ReturnToServerPool(cur_service);
1994 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
1995 struct rx_call *tcall, *choice2 = NULL;
1996 struct opr_queue *cursor;
1998 /* Scan for eligible incoming calls. A call is not eligible
1999 * if the maximum number of calls for its service type are
2000 * already executing */
2001 /* One thread will process calls FCFS (to prevent starvation),
2002 * while the other threads may run ahead looking for calls which
2003 * have all their input data available immediately. This helps
2004 * keep threads from blocking, waiting for data from the client. */
2005 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2006 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2008 service = tcall->conn->service;
2009 if (!QuotaOK(service)) {
2012 MUTEX_ENTER(&rx_pthread_mutex);
2013 if (tno == rxi_fcfs_thread_num
2014 || opr_queue_IsEnd(&rx_incomingCallQueue, cursor)) {
2015 MUTEX_EXIT(&rx_pthread_mutex);
2016 /* If we're the fcfs thread , then we'll just use
2017 * this call. If we haven't been able to find an optimal
2018 * choice, and we're at the end of the list, then use a
2019 * 2d choice if one has been identified. Otherwise... */
2020 call = (choice2 ? choice2 : tcall);
2021 service = call->conn->service;
2023 MUTEX_EXIT(&rx_pthread_mutex);
2024 if (!opr_queue_IsEmpty(&tcall->rq)) {
2025 struct rx_packet *rp;
2026 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2028 if (rp->header.seq == 1) {
2030 || (rp->header.flags & RX_LAST_PACKET)) {
2032 } else if (rxi_2dchoice && !choice2
2033 && !(tcall->flags & RX_CALL_CLEARED)
2034 && (tcall->rprev > rxi_HardAckRate)) {
2044 ReturnToServerPool(service);
2050 opr_queue_Remove(&call->entry);
2051 MUTEX_EXIT(&rx_serverPool_lock);
2052 MUTEX_ENTER(&call->lock);
2054 if (call->flags & RX_CALL_WAIT_PROC) {
2055 call->flags &= ~RX_CALL_WAIT_PROC;
2056 rx_atomic_dec(&rx_nWaiting);
2059 if (call->state != RX_STATE_PRECALL || call->error) {
2060 MUTEX_EXIT(&call->lock);
2061 MUTEX_ENTER(&rx_serverPool_lock);
2062 ReturnToServerPool(service);
2067 if (opr_queue_IsEmpty(&call->rq)
2068 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1)
2069 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2071 CLEAR_CALL_QUEUE_LOCK(call);
2074 /* If there are no eligible incoming calls, add this process
2075 * to the idle server queue, to wait for one */
2079 *socketp = OSI_NULLSOCKET;
2081 sq->socketp = socketp;
2082 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2083 #ifndef AFS_AIX41_ENV
2084 rx_waitForPacket = sq;
2085 #endif /* AFS_AIX41_ENV */
2087 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2089 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2090 MUTEX_EXIT(&rx_serverPool_lock);
2091 return (struct rx_call *)0;
2094 } while (!(call = sq->newcall)
2095 && !(socketp && *socketp != OSI_NULLSOCKET));
2096 MUTEX_EXIT(&rx_serverPool_lock);
2098 MUTEX_ENTER(&call->lock);
2104 MUTEX_ENTER(&freeSQEList_lock);
2105 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2106 rx_FreeSQEList = sq;
2107 MUTEX_EXIT(&freeSQEList_lock);
2110 clock_GetTime(&call->startTime);
2111 call->state = RX_STATE_ACTIVE;
2112 call->app.mode = RX_MODE_RECEIVING;
2113 #ifdef RX_KERNEL_TRACE
2114 if (ICL_SETACTIVE(afs_iclSetp)) {
2115 int glockOwner = ISAFS_GLOCK();
2118 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2119 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2126 rxi_calltrace(RX_CALL_START, call);
2127 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2128 call->conn->service->servicePort, call->conn->service->serviceId,
2131 MUTEX_EXIT(&call->lock);
2132 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2134 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2139 #else /* RX_ENABLE_LOCKS */
2141 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2143 struct rx_serverQueueEntry *sq;
2144 struct rx_call *call = (struct rx_call *)0, *choice2;
2145 struct rx_service *service = NULL;
2149 MUTEX_ENTER(&freeSQEList_lock);
2151 if ((sq = rx_FreeSQEList)) {
2152 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2153 MUTEX_EXIT(&freeSQEList_lock);
2154 } else { /* otherwise allocate a new one and return that */
2155 MUTEX_EXIT(&freeSQEList_lock);
2156 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2157 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2158 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2160 MUTEX_ENTER(&sq->lock);
2162 if (cur_service != NULL) {
2163 cur_service->nRequestsRunning--;
2164 MUTEX_ENTER(&rx_quota_mutex);
2165 if (cur_service->nRequestsRunning < cur_service->minProcs)
2168 MUTEX_EXIT(&rx_quota_mutex);
2170 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2171 struct rx_call *tcall;
2172 struct opr_queue *cursor;
2173 /* Scan for eligible incoming calls. A call is not eligible
2174 * if the maximum number of calls for its service type are
2175 * already executing */
2176 /* One thread will process calls FCFS (to prevent starvation),
2177 * while the other threads may run ahead looking for calls which
2178 * have all their input data available immediately. This helps
2179 * keep threads from blocking, waiting for data from the client. */
2180 choice2 = (struct rx_call *)0;
2181 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2182 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2183 service = tcall->conn->service;
2184 if (QuotaOK(service)) {
2185 MUTEX_ENTER(&rx_pthread_mutex);
2186 /* XXX - If tcall->entry.next is NULL, then we're no longer
2187 * on a queue at all. This shouldn't happen. */
2188 if (tno == rxi_fcfs_thread_num || !tcall->entry.next) {
2189 MUTEX_EXIT(&rx_pthread_mutex);
2190 /* If we're the fcfs thread, then we'll just use
2191 * this call. If we haven't been able to find an optimal
2192 * choice, and we're at the end of the list, then use a
2193 * 2d choice if one has been identified. Otherwise... */
2194 call = (choice2 ? choice2 : tcall);
2195 service = call->conn->service;
2197 MUTEX_EXIT(&rx_pthread_mutex);
2198 if (!opr_queue_IsEmpty(&tcall->rq)) {
2199 struct rx_packet *rp;
2200 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2202 if (rp->header.seq == 1
2204 || (rp->header.flags & RX_LAST_PACKET))) {
2206 } else if (rxi_2dchoice && !choice2
2207 && !(tcall->flags & RX_CALL_CLEARED)
2208 && (tcall->rprev > rxi_HardAckRate)) {
2221 opr_queue_Remove(&call->entry);
2222 /* we can't schedule a call if there's no data!!! */
2223 /* send an ack if there's no data, if we're missing the
2224 * first packet, or we're missing something between first
2225 * and last -- there's a "hole" in the incoming data. */
2226 if (opr_queue_IsEmpty(&call->rq)
2227 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1
2228 || call->rprev != opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq)
2229 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2231 call->flags &= (~RX_CALL_WAIT_PROC);
2232 service->nRequestsRunning++;
2233 /* just started call in minProcs pool, need fewer to maintain
2235 MUTEX_ENTER(&rx_quota_mutex);
2236 if (service->nRequestsRunning <= service->minProcs)
2239 MUTEX_EXIT(&rx_quota_mutex);
2240 rx_atomic_dec(&rx_nWaiting);
2241 /* MUTEX_EXIT(&call->lock); */
2243 /* If there are no eligible incoming calls, add this process
2244 * to the idle server queue, to wait for one */
2247 *socketp = OSI_NULLSOCKET;
2249 sq->socketp = socketp;
2250 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2254 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2256 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2257 return (struct rx_call *)0;
2260 } while (!(call = sq->newcall)
2261 && !(socketp && *socketp != OSI_NULLSOCKET));
2263 MUTEX_EXIT(&sq->lock);
2265 MUTEX_ENTER(&freeSQEList_lock);
2266 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2267 rx_FreeSQEList = sq;
2268 MUTEX_EXIT(&freeSQEList_lock);
2271 clock_GetTime(&call->startTime);
2272 call->state = RX_STATE_ACTIVE;
2273 call->app.mode = RX_MODE_RECEIVING;
2274 #ifdef RX_KERNEL_TRACE
2275 if (ICL_SETACTIVE(afs_iclSetp)) {
2276 int glockOwner = ISAFS_GLOCK();
2279 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2280 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2287 rxi_calltrace(RX_CALL_START, call);
2288 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2289 call->conn->service->servicePort, call->conn->service->serviceId,
2292 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2299 #endif /* RX_ENABLE_LOCKS */
2303 /* Establish a procedure to be called when a packet arrives for a
2304 * call. This routine will be called at most once after each call,
2305 * and will also be called if there is an error condition on the or
2306 * the call is complete. Used by multi rx to build a selection
2307 * function which determines which of several calls is likely to be a
2308 * good one to read from.
2309 * NOTE: the way this is currently implemented it is probably only a
2310 * good idea to (1) use it immediately after a newcall (clients only)
2311 * and (2) only use it once. Other uses currently void your warranty
2314 rx_SetArrivalProc(struct rx_call *call,
2315 void (*proc) (struct rx_call * call,
2318 void * handle, int arg)
2320 call->arrivalProc = proc;
2321 call->arrivalProcHandle = handle;
2322 call->arrivalProcArg = arg;
2325 /* Call is finished (possibly prematurely). Return rc to the peer, if
2326 * appropriate, and return the final error code from the conversation
2330 rx_EndCall(struct rx_call *call, afs_int32 rc)
2332 struct rx_connection *conn = call->conn;
2336 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2337 call, rc, call->error, call->abortCode));
2340 MUTEX_ENTER(&call->lock);
2342 if (rc == 0 && call->error == 0) {
2343 call->abortCode = 0;
2344 call->abortCount = 0;
2347 call->arrivalProc = (void (*)())0;
2348 if (rc && call->error == 0) {
2349 rxi_CallError(call, rc);
2350 call->app.mode = RX_MODE_ERROR;
2351 /* Send an abort message to the peer if this error code has
2352 * only just been set. If it was set previously, assume the
2353 * peer has already been sent the error code or will request it
2355 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2357 if (conn->type == RX_SERVER_CONNECTION) {
2358 /* Make sure reply or at least dummy reply is sent */
2359 if (call->app.mode == RX_MODE_RECEIVING) {
2360 MUTEX_EXIT(&call->lock);
2361 rxi_WriteProc(call, 0, 0);
2362 MUTEX_ENTER(&call->lock);
2364 if (call->app.mode == RX_MODE_SENDING) {
2365 rxi_FlushWriteLocked(call);
2367 rxi_calltrace(RX_CALL_END, call);
2368 /* Call goes to hold state until reply packets are acknowledged */
2369 if (call->tfirst + call->nSoftAcked < call->tnext) {
2370 call->state = RX_STATE_HOLD;
2372 call->state = RX_STATE_DALLY;
2373 rxi_ClearTransmitQueue(call, 0);
2374 rxi_rto_cancel(call);
2375 rxi_CancelKeepAliveEvent(call);
2377 } else { /* Client connection */
2379 /* Make sure server receives input packets, in the case where
2380 * no reply arguments are expected */
2382 if ((call->app.mode == RX_MODE_SENDING)
2383 || (call->app.mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2384 MUTEX_EXIT(&call->lock);
2385 (void)rxi_ReadProc(call, &dummy, 1);
2386 MUTEX_ENTER(&call->lock);
2389 /* If we had an outstanding delayed ack, be nice to the server
2390 * and force-send it now.
2392 if (call->delayedAckEvent) {
2393 rxi_CancelDelayedAckEvent(call);
2394 rxi_SendDelayedAck(NULL, call, NULL, 0);
2397 /* We need to release the call lock since it's lower than the
2398 * conn_call_lock and we don't want to hold the conn_call_lock
2399 * over the rx_ReadProc call. The conn_call_lock needs to be held
2400 * here for the case where rx_NewCall is perusing the calls on
2401 * the connection structure. We don't want to signal until
2402 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2403 * have checked this call, found it active and by the time it
2404 * goes to sleep, will have missed the signal.
2406 MUTEX_EXIT(&call->lock);
2407 MUTEX_ENTER(&conn->conn_call_lock);
2408 MUTEX_ENTER(&call->lock);
2411 /* While there are some circumstances where a call with an error is
2412 * obviously not on a "busy" channel, be conservative (clearing
2413 * lastBusy is just best-effort to possibly speed up rx_NewCall).
2414 * The call channel is definitely not busy if we just successfully
2415 * completed a call on it. */
2416 conn->lastBusy[call->channel] = 0;
2418 } else if (call->error == RX_CALL_TIMEOUT) {
2419 /* The call is still probably running on the server side, so try to
2420 * avoid this call channel in the future. */
2421 conn->lastBusy[call->channel] = clock_Sec();
2424 MUTEX_ENTER(&conn->conn_data_lock);
2425 conn->flags |= RX_CONN_BUSY;
2426 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2427 MUTEX_EXIT(&conn->conn_data_lock);
2428 #ifdef RX_ENABLE_LOCKS
2429 CV_BROADCAST(&conn->conn_call_cv);
2434 #ifdef RX_ENABLE_LOCKS
2436 MUTEX_EXIT(&conn->conn_data_lock);
2438 #endif /* RX_ENABLE_LOCKS */
2439 call->state = RX_STATE_DALLY;
2441 error = call->error;
2443 /* currentPacket, nLeft, and NFree must be zeroed here, because
2444 * ResetCall cannot: ResetCall may be called at splnet(), in the
2445 * kernel version, and may interrupt the macros rx_Read or
2446 * rx_Write, which run at normal priority for efficiency. */
2447 if (call->app.currentPacket) {
2448 #ifdef RX_TRACK_PACKETS
2449 call->app.currentPacket->flags &= ~RX_PKTFLAG_CP;
2451 rxi_FreePacket(call->app.currentPacket);
2452 call->app.currentPacket = (struct rx_packet *)0;
2455 call->app.nLeft = call->app.nFree = call->app.curlen = 0;
2457 /* Free any packets from the last call to ReadvProc/WritevProc */
2458 #ifdef RXDEBUG_PACKET
2460 #endif /* RXDEBUG_PACKET */
2461 rxi_FreePackets(0, &call->app.iovq);
2462 MUTEX_EXIT(&call->lock);
2464 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2465 if (conn->type == RX_CLIENT_CONNECTION) {
2466 MUTEX_ENTER(&conn->conn_data_lock);
2467 conn->flags &= ~RX_CONN_BUSY;
2468 MUTEX_EXIT(&conn->conn_data_lock);
2469 MUTEX_EXIT(&conn->conn_call_lock);
2473 * Map errors to the local host's errno.h format.
2475 error = ntoh_syserr_conv(error);
2477 /* If the caller said the call failed with some error, we had better
2478 * return an error code. */
2479 osi_Assert(!rc || error);
2483 #if !defined(KERNEL)
2485 /* Call this routine when shutting down a server or client (especially
2486 * clients). This will allow Rx to gracefully garbage collect server
2487 * connections, and reduce the number of retries that a server might
2488 * make to a dead client.
2489 * This is not quite right, since some calls may still be ongoing and
2490 * we can't lock them to destroy them. */
2494 struct rx_connection **conn_ptr, **conn_end;
2497 if (rx_atomic_test_and_set_bit(&rxinit_status, 0))
2498 return; /* Already shutdown. */
2500 rxi_DeleteCachedConnections();
2501 if (rx_connHashTable) {
2502 MUTEX_ENTER(&rx_connHashTable_lock);
2503 for (conn_ptr = &rx_connHashTable[0], conn_end =
2504 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2506 struct rx_connection *conn, *next;
2507 for (conn = *conn_ptr; conn; conn = next) {
2509 if (conn->type == RX_CLIENT_CONNECTION) {
2510 MUTEX_ENTER(&rx_refcnt_mutex);
2512 MUTEX_EXIT(&rx_refcnt_mutex);
2513 #ifdef RX_ENABLE_LOCKS
2514 rxi_DestroyConnectionNoLock(conn);
2515 #else /* RX_ENABLE_LOCKS */
2516 rxi_DestroyConnection(conn);
2517 #endif /* RX_ENABLE_LOCKS */
2521 #ifdef RX_ENABLE_LOCKS
2522 while (rx_connCleanup_list) {
2523 struct rx_connection *conn;
2524 conn = rx_connCleanup_list;
2525 rx_connCleanup_list = rx_connCleanup_list->next;
2526 MUTEX_EXIT(&rx_connHashTable_lock);
2527 rxi_CleanupConnection(conn);
2528 MUTEX_ENTER(&rx_connHashTable_lock);
2530 MUTEX_EXIT(&rx_connHashTable_lock);
2531 #endif /* RX_ENABLE_LOCKS */
2536 afs_winsockCleanup();
2542 /* if we wakeup packet waiter too often, can get in loop with two
2543 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2545 rxi_PacketsUnWait(void)
2547 if (!rx_waitingForPackets) {
2551 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2552 return; /* still over quota */
2555 rx_waitingForPackets = 0;
2556 #ifdef RX_ENABLE_LOCKS
2557 CV_BROADCAST(&rx_waitingForPackets_cv);
2559 osi_rxWakeup(&rx_waitingForPackets);
2565 /* ------------------Internal interfaces------------------------- */
2567 /* Return this process's service structure for the
2568 * specified socket and service */
2569 static struct rx_service *
2570 rxi_FindService(osi_socket socket, u_short serviceId)
2572 struct rx_service **sp;
2573 for (sp = &rx_services[0]; *sp; sp++) {
2574 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2580 #ifdef RXDEBUG_PACKET
2581 #ifdef KDUMP_RX_LOCK
2582 static struct rx_call_rx_lock *rx_allCallsp = 0;
2584 static struct rx_call *rx_allCallsp = 0;
2586 #endif /* RXDEBUG_PACKET */
2588 /* Allocate a call structure, for the indicated channel of the
2589 * supplied connection. The mode and state of the call must be set by
2590 * the caller. Returns the call with mutex locked. */
2591 static struct rx_call *
2592 rxi_NewCall(struct rx_connection *conn, int channel)
2594 struct rx_call *call;
2595 #ifdef RX_ENABLE_LOCKS
2596 struct rx_call *cp; /* Call pointer temp */
2597 struct opr_queue *cursor;
2600 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2602 /* Grab an existing call structure, or allocate a new one.
2603 * Existing call structures are assumed to have been left reset by
2605 MUTEX_ENTER(&rx_freeCallQueue_lock);
2607 #ifdef RX_ENABLE_LOCKS
2609 * EXCEPT that the TQ might not yet be cleared out.
2610 * Skip over those with in-use TQs.
2613 for (opr_queue_Scan(&rx_freeCallQueue, cursor)) {
2614 cp = opr_queue_Entry(cursor, struct rx_call, entry);
2615 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2621 #else /* RX_ENABLE_LOCKS */
2622 if (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
2623 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
2624 #endif /* RX_ENABLE_LOCKS */
2625 opr_queue_Remove(&call->entry);
2626 if (rx_stats_active)
2627 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2628 MUTEX_EXIT(&rx_freeCallQueue_lock);
2629 MUTEX_ENTER(&call->lock);
2630 CLEAR_CALL_QUEUE_LOCK(call);
2631 #ifdef RX_ENABLE_LOCKS
2632 /* Now, if TQ wasn't cleared earlier, do it now. */
2633 rxi_WaitforTQBusy(call);
2634 if (call->flags & RX_CALL_TQ_CLEARME) {
2635 rxi_ClearTransmitQueue(call, 1);
2636 /*queue_Init(&call->tq);*/
2638 #endif /* RX_ENABLE_LOCKS */
2639 /* Bind the call to its connection structure */
2641 rxi_ResetCall(call, 1);
2644 call = rxi_Alloc(sizeof(struct rx_call));
2645 #ifdef RXDEBUG_PACKET
2646 call->allNextp = rx_allCallsp;
2647 rx_allCallsp = call;
2649 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2650 #else /* RXDEBUG_PACKET */
2651 rx_atomic_inc(&rx_stats.nCallStructs);
2652 #endif /* RXDEBUG_PACKET */
2654 MUTEX_EXIT(&rx_freeCallQueue_lock);
2655 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2656 MUTEX_ENTER(&call->lock);
2657 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2658 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2659 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2661 /* Initialize once-only items */
2662 opr_queue_Init(&call->tq);
2663 opr_queue_Init(&call->rq);
2664 opr_queue_Init(&call->app.iovq);
2665 #ifdef RXDEBUG_PACKET
2666 call->rqc = call->tqc = call->iovqc = 0;
2667 #endif /* RXDEBUG_PACKET */
2668 /* Bind the call to its connection structure (prereq for reset) */
2670 rxi_ResetCall(call, 1);
2672 call->channel = channel;
2673 call->callNumber = &conn->callNumber[channel];
2674 call->rwind = conn->rwind[channel];
2675 call->twind = conn->twind[channel];
2676 /* Note that the next expected call number is retained (in
2677 * conn->callNumber[i]), even if we reallocate the call structure
2679 conn->call[channel] = call;
2680 /* if the channel's never been used (== 0), we should start at 1, otherwise
2681 * the call number is valid from the last time this channel was used */
2682 if (*call->callNumber == 0)
2683 *call->callNumber = 1;
2688 /* A call has been inactive long enough that so we can throw away
2689 * state, including the call structure, which is placed on the call
2692 * call->lock amd rx_refcnt_mutex are held upon entry.
2693 * haveCTLock is set when called from rxi_ReapConnections.
2695 * return 1 if the call is freed, 0 if not.
2698 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2700 int channel = call->channel;
2701 struct rx_connection *conn = call->conn;
2702 u_char state = call->state;
2705 * We are setting the state to RX_STATE_RESET to
2706 * ensure that no one else will attempt to use this
2707 * call once we drop the refcnt lock. We must drop
2708 * the refcnt lock before calling rxi_ResetCall
2709 * because it cannot be held across acquiring the
2710 * freepktQ lock. NewCall does the same.
2712 call->state = RX_STATE_RESET;
2713 MUTEX_EXIT(&rx_refcnt_mutex);
2714 rxi_ResetCall(call, 0);
2716 if (MUTEX_TRYENTER(&conn->conn_call_lock))
2718 if (state == RX_STATE_DALLY || state == RX_STATE_HOLD)
2719 (*call->callNumber)++;
2721 if (call->conn->call[channel] == call)
2722 call->conn->call[channel] = 0;
2723 MUTEX_EXIT(&conn->conn_call_lock);
2726 * We couldn't obtain the conn_call_lock so we can't
2727 * disconnect the call from the connection. Set the
2728 * call state to dally so that the call can be reused.
2730 MUTEX_ENTER(&rx_refcnt_mutex);
2731 call->state = RX_STATE_DALLY;
2735 MUTEX_ENTER(&rx_freeCallQueue_lock);
2736 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2737 #ifdef RX_ENABLE_LOCKS
2738 /* A call may be free even though its transmit queue is still in use.
2739 * Since we search the call list from head to tail, put busy calls at
2740 * the head of the list, and idle calls at the tail.
2742 if (call->flags & RX_CALL_TQ_BUSY)
2743 opr_queue_Prepend(&rx_freeCallQueue, &call->entry);
2745 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2746 #else /* RX_ENABLE_LOCKS */
2747 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2748 #endif /* RX_ENABLE_LOCKS */
2749 if (rx_stats_active)
2750 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2751 MUTEX_EXIT(&rx_freeCallQueue_lock);
2753 /* Destroy the connection if it was previously slated for
2754 * destruction, i.e. the Rx client code previously called
2755 * rx_DestroyConnection (client connections), or
2756 * rxi_ReapConnections called the same routine (server
2757 * connections). Only do this, however, if there are no
2758 * outstanding calls. Note that for fine grain locking, there appears
2759 * to be a deadlock in that rxi_FreeCall has a call locked and
2760 * DestroyConnectionNoLock locks each call in the conn. But note a
2761 * few lines up where we have removed this call from the conn.
2762 * If someone else destroys a connection, they either have no
2763 * call lock held or are going through this section of code.
2765 MUTEX_ENTER(&conn->conn_data_lock);
2766 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2767 MUTEX_ENTER(&rx_refcnt_mutex);
2769 MUTEX_EXIT(&rx_refcnt_mutex);
2770 MUTEX_EXIT(&conn->conn_data_lock);
2771 #ifdef RX_ENABLE_LOCKS
2773 rxi_DestroyConnectionNoLock(conn);
2775 rxi_DestroyConnection(conn);
2776 #else /* RX_ENABLE_LOCKS */
2777 rxi_DestroyConnection(conn);
2778 #endif /* RX_ENABLE_LOCKS */
2780 MUTEX_EXIT(&conn->conn_data_lock);
2782 MUTEX_ENTER(&rx_refcnt_mutex);
2786 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2787 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2790 rxi_Alloc(size_t size)
2794 if (rx_stats_active) {
2795 rx_atomic_add(&rxi_Allocsize, (int) size);
2796 rx_atomic_inc(&rxi_Alloccnt);
2800 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2801 afs_osi_Alloc_NoSleep(size);
2806 osi_Panic("rxi_Alloc error");
2812 rxi_Free(void *addr, size_t size)
2814 if (rx_stats_active) {
2815 rx_atomic_sub(&rxi_Allocsize, (int) size);
2816 rx_atomic_dec(&rxi_Alloccnt);
2818 osi_Free(addr, size);
2822 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2824 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2825 struct rx_peer *next = NULL;
2829 MUTEX_ENTER(&rx_peerHashTable_lock);
2831 peer_ptr = &rx_peerHashTable[0];
2832 peer_end = &rx_peerHashTable[rx_hashTableSize];
2835 for ( ; peer_ptr < peer_end; peer_ptr++) {
2838 for ( ; peer; peer = next) {
2840 if (host == peer->host)
2845 hashIndex = PEER_HASH(host, port);
2846 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2847 if ((peer->host == host) && (peer->port == port))
2852 MUTEX_ENTER(&rx_peerHashTable_lock);
2857 MUTEX_EXIT(&rx_peerHashTable_lock);
2859 MUTEX_ENTER(&peer->peer_lock);
2860 /* We don't handle dropping below min, so don't */
2861 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2862 peer->ifMTU=MIN(mtu, peer->ifMTU);
2863 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2864 /* if we tweaked this down, need to tune our peer MTU too */
2865 peer->MTU = MIN(peer->MTU, peer->natMTU);
2866 /* if we discovered a sub-1500 mtu, degrade */
2867 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2868 peer->maxDgramPackets = 1;
2869 /* We no longer have valid peer packet information */
2870 if (peer->maxPacketSize + RX_HEADER_SIZE > peer->ifMTU)
2871 peer->maxPacketSize = 0;
2872 MUTEX_EXIT(&peer->peer_lock);
2874 MUTEX_ENTER(&rx_peerHashTable_lock);
2876 if (host && !port) {
2878 /* pick up where we left off */
2882 MUTEX_EXIT(&rx_peerHashTable_lock);
2885 #ifdef AFS_RXERRQ_ENV
2887 rxi_SetPeerDead(struct sock_extended_err *err, afs_uint32 host, afs_uint16 port)
2889 int hashIndex = PEER_HASH(host, port);
2890 struct rx_peer *peer;
2892 MUTEX_ENTER(&rx_peerHashTable_lock);
2894 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2895 if (peer->host == host && peer->port == port) {
2901 MUTEX_EXIT(&rx_peerHashTable_lock);
2904 rx_atomic_inc(&peer->neterrs);
2905 MUTEX_ENTER(&peer->peer_lock);
2906 peer->last_err_origin = RX_NETWORK_ERROR_ORIGIN_ICMP;
2907 peer->last_err_type = err->ee_type;
2908 peer->last_err_code = err->ee_code;
2909 MUTEX_EXIT(&peer->peer_lock);
2911 MUTEX_ENTER(&rx_peerHashTable_lock);
2913 MUTEX_EXIT(&rx_peerHashTable_lock);
2918 rxi_ProcessNetError(struct sock_extended_err *err, afs_uint32 addr, afs_uint16 port)
2920 # ifdef AFS_ADAPT_PMTU
2921 if (err->ee_errno == EMSGSIZE && err->ee_info >= 68) {
2922 rxi_SetPeerMtu(NULL, addr, port, err->ee_info - RX_IPUDP_SIZE);
2926 if (err->ee_origin == SO_EE_ORIGIN_ICMP && err->ee_type == ICMP_DEST_UNREACH) {
2927 switch (err->ee_code) {
2928 case ICMP_NET_UNREACH:
2929 case ICMP_HOST_UNREACH:
2930 case ICMP_PORT_UNREACH:
2933 rxi_SetPeerDead(err, addr, port);
2940 rxi_TranslateICMP(int type, int code)
2943 case ICMP_DEST_UNREACH:
2945 case ICMP_NET_UNREACH:
2946 return "Destination Net Unreachable";
2947 case ICMP_HOST_UNREACH:
2948 return "Destination Host Unreachable";
2949 case ICMP_PROT_UNREACH:
2950 return "Destination Protocol Unreachable";
2951 case ICMP_PORT_UNREACH:
2952 return "Destination Port Unreachable";
2954 return "Destination Net Prohibited";
2956 return "Destination Host Prohibited";
2962 #endif /* AFS_RXERRQ_ENV */
2965 * Get the last network error for a connection
2967 * A "network error" here means an error retrieved from ICMP, or some other
2968 * mechanism outside of Rx that informs us of errors in network reachability.
2970 * If a peer associated with the given Rx connection has received a network
2971 * error recently, this function allows the caller to know what error
2972 * specifically occurred. This can be useful to know, since e.g. ICMP errors
2973 * can cause calls to that peer to be quickly aborted. So, this function can
2974 * help see why a call was aborted due to network errors.
2976 * If we have received traffic from a peer since the last network error, we
2977 * treat that peer as if we had not received an network error for it.
2979 * @param[in] conn The Rx connection to examine
2980 * @param[out] err_origin The origin of the last network error (e.g. ICMP);
2981 * one of the RX_NETWORK_ERROR_ORIGIN_* constants
2982 * @param[out] err_type The type of the last error
2983 * @param[out] err_code The code of the last error
2984 * @param[out] msg Human-readable error message, if applicable; NULL otherwise
2986 * @return If we have an error
2987 * @retval -1 No error to get; 'out' params are undefined
2988 * @retval 0 We have an error; 'out' params contain the last error
2991 rx_GetNetworkError(struct rx_connection *conn, int *err_origin, int *err_type,
2992 int *err_code, const char **msg)
2994 #ifdef AFS_RXERRQ_ENV
2995 struct rx_peer *peer = conn->peer;
2996 if (rx_atomic_read(&peer->neterrs)) {
2997 MUTEX_ENTER(&peer->peer_lock);
2998 *err_origin = peer->last_err_origin;
2999 *err_type = peer->last_err_type;
3000 *err_code = peer->last_err_code;
3001 MUTEX_EXIT(&peer->peer_lock);
3004 if (*err_origin == RX_NETWORK_ERROR_ORIGIN_ICMP) {
3005 *msg = rxi_TranslateICMP(*err_type, *err_code);
3014 /* Find the peer process represented by the supplied (host,port)
3015 * combination. If there is no appropriate active peer structure, a
3016 * new one will be allocated and initialized
3019 rxi_FindPeer(afs_uint32 host, u_short port, int create)
3023 hashIndex = PEER_HASH(host, port);
3024 MUTEX_ENTER(&rx_peerHashTable_lock);
3025 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
3026 if ((pp->host == host) && (pp->port == port))
3031 pp = rxi_AllocPeer(); /* This bzero's *pp */
3032 pp->host = host; /* set here or in InitPeerParams is zero */
3034 #ifdef AFS_RXERRQ_ENV
3035 rx_atomic_set(&pp->neterrs, 0);
3037 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
3038 opr_queue_Init(&pp->rpcStats);
3039 pp->next = rx_peerHashTable[hashIndex];
3040 rx_peerHashTable[hashIndex] = pp;
3041 rxi_InitPeerParams(pp);
3042 if (rx_stats_active)
3043 rx_atomic_inc(&rx_stats.nPeerStructs);
3049 MUTEX_EXIT(&rx_peerHashTable_lock);
3054 /* Find the connection at (host, port) started at epoch, and with the
3055 * given connection id. Creates the server connection if necessary.
3056 * The type specifies whether a client connection or a server
3057 * connection is desired. In both cases, (host, port) specify the
3058 * peer's (host, pair) pair. Client connections are not made
3059 * automatically by this routine. The parameter socket gives the
3060 * socket descriptor on which the packet was received. This is used,
3061 * in the case of server connections, to check that *new* connections
3062 * come via a valid (port, serviceId). Finally, the securityIndex
3063 * parameter must match the existing index for the connection. If a
3064 * server connection is created, it will be created using the supplied
3065 * index, if the index is valid for this service */
3066 static struct rx_connection *
3067 rxi_FindConnection(osi_socket socket, afs_uint32 host,
3068 u_short port, u_short serviceId, afs_uint32 cid,
3069 afs_uint32 epoch, int type, u_int securityIndex,
3070 int *unknownService)
3072 int hashindex, flag, i;
3073 struct rx_connection *conn;
3074 *unknownService = 0;
3075 hashindex = CONN_HASH(host, port, cid, epoch, type);
3076 MUTEX_ENTER(&rx_connHashTable_lock);
3077 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
3078 rx_connHashTable[hashindex],
3081 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
3082 && (epoch == conn->epoch)) {
3083 struct rx_peer *pp = conn->peer;
3084 if (securityIndex != conn->securityIndex) {
3085 /* this isn't supposed to happen, but someone could forge a packet
3086 * like this, and there seems to be some CM bug that makes this
3087 * happen from time to time -- in which case, the fileserver
3089 MUTEX_EXIT(&rx_connHashTable_lock);
3090 return (struct rx_connection *)0;
3092 if (pp->host == host && pp->port == port)
3094 if (type == RX_CLIENT_CONNECTION && pp->port == port)
3096 /* So what happens when it's a callback connection? */
3097 if ( /*type == RX_CLIENT_CONNECTION && */
3098 (conn->epoch & 0x80000000))
3102 /* the connection rxLastConn that was used the last time is not the
3103 ** one we are looking for now. Hence, start searching in the hash */
3105 conn = rx_connHashTable[hashindex];
3110 struct rx_service *service;
3111 if (type == RX_CLIENT_CONNECTION) {
3112 MUTEX_EXIT(&rx_connHashTable_lock);
3113 return (struct rx_connection *)0;
3115 service = rxi_FindService(socket, serviceId);
3116 if (!service || (securityIndex >= service->nSecurityObjects)
3117 || (service->securityObjects[securityIndex] == 0)) {
3118 MUTEX_EXIT(&rx_connHashTable_lock);
3119 *unknownService = 1;
3120 return (struct rx_connection *)0;
3122 conn = rxi_AllocConnection(); /* This bzero's the connection */
3123 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
3124 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
3125 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
3126 conn->next = rx_connHashTable[hashindex];
3127 rx_connHashTable[hashindex] = conn;
3128 conn->peer = rxi_FindPeer(host, port, 1);
3129 conn->type = RX_SERVER_CONNECTION;
3130 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
3131 conn->epoch = epoch;
3132 conn->cid = cid & RX_CIDMASK;
3133 conn->ackRate = RX_FAST_ACK_RATE;
3134 conn->service = service;
3135 conn->serviceId = serviceId;
3136 conn->securityIndex = securityIndex;
3137 conn->securityObject = service->securityObjects[securityIndex];
3138 conn->nSpecific = 0;
3139 conn->specific = NULL;
3140 rx_SetConnDeadTime(conn, service->connDeadTime);
3141 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
3142 for (i = 0; i < RX_MAXCALLS; i++) {
3143 conn->twind[i] = rx_initSendWindow;
3144 conn->rwind[i] = rx_initReceiveWindow;
3146 /* Notify security object of the new connection */
3147 RXS_NewConnection(conn->securityObject, conn);
3148 /* XXXX Connection timeout? */
3149 if (service->newConnProc)
3150 (*service->newConnProc) (conn);
3151 if (rx_stats_active)
3152 rx_atomic_inc(&rx_stats.nServerConns);
3155 MUTEX_ENTER(&rx_refcnt_mutex);
3157 MUTEX_EXIT(&rx_refcnt_mutex);
3159 rxLastConn = conn; /* store this connection as the last conn used */
3160 MUTEX_EXIT(&rx_connHashTable_lock);
3165 * Abort the call if the server is over the busy threshold. This
3166 * can be used without requiring a call structure be initialised,
3167 * or connected to a particular channel
3170 rxi_AbortIfServerBusy(osi_socket socket, struct rx_connection *conn,
3171 struct rx_packet *np)
3173 if ((rx_BusyThreshold > 0) &&
3174 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3175 rxi_SendRawAbort(socket, conn->peer->host, conn->peer->port,
3176 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, 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 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4437 /* Get the maximum packet size to send to this peer */
4438 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4440 tSize = (afs_uint32) ntohl(tSize);
4441 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4442 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4444 /* sanity check - peer might have restarted with different params.
4445 * If peer says "send less", dammit, send less... Peer should never
4446 * be unable to accept packets of the size that prior AFS versions would
4447 * send without asking. */
4448 if (peer->maxMTU != tSize) {
4449 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4451 peer->maxMTU = tSize;
4452 peer->MTU = MIN(tSize, peer->MTU);
4453 call->MTU = MIN(call->MTU, tSize);
4456 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4459 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4460 (int)sizeof(afs_int32), &tSize);
4461 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4462 if (tSize < call->twind) { /* smaller than our send */
4463 call->twind = tSize; /* window, we must send less... */
4464 call->ssthresh = MIN(call->twind, call->ssthresh);
4465 call->conn->twind[call->channel] = call->twind;
4468 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4469 * network MTU confused with the loopback MTU. Calculate the
4470 * maximum MTU here for use in the slow start code below.
4472 /* Did peer restart with older RX version? */
4473 if (peer->maxDgramPackets > 1) {
4474 peer->maxDgramPackets = 1;
4476 } else if (np->length >=
4477 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4480 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4481 sizeof(afs_int32), &tSize);
4482 tSize = (afs_uint32) ntohl(tSize);
4484 * As of AFS 3.5 we set the send window to match the receive window.
4486 if (tSize < call->twind) {
4487 call->twind = tSize;
4488 call->conn->twind[call->channel] = call->twind;
4489 call->ssthresh = MIN(call->twind, call->ssthresh);
4490 } else if (tSize > call->twind) {
4491 call->twind = tSize;
4492 call->conn->twind[call->channel] = call->twind;
4496 * As of AFS 3.5, a jumbogram is more than one fixed size
4497 * packet transmitted in a single UDP datagram. If the remote
4498 * MTU is smaller than our local MTU then never send a datagram
4499 * larger than the natural MTU.
4502 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4503 (int)sizeof(afs_int32), &tSize);
4504 maxDgramPackets = (afs_uint32) ntohl(tSize);
4505 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4507 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4508 if (maxDgramPackets > 1) {
4509 peer->maxDgramPackets = maxDgramPackets;
4510 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4512 peer->maxDgramPackets = 1;
4513 call->MTU = peer->natMTU;
4515 } else if (peer->maxDgramPackets > 1) {
4516 /* Restarted with lower version of RX */
4517 peer->maxDgramPackets = 1;
4519 } else if (peer->maxDgramPackets > 1
4520 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4521 /* Restarted with lower version of RX */
4522 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4523 peer->natMTU = OLD_MAX_PACKET_SIZE;
4524 peer->MTU = OLD_MAX_PACKET_SIZE;
4525 peer->maxDgramPackets = 1;
4526 peer->nDgramPackets = 1;
4528 call->MTU = OLD_MAX_PACKET_SIZE;
4531 /* If the window has been extended by this acknowledge packet,
4532 * then wakeup a sender waiting in alloc for window space, or try
4533 * sending packets now, if he's been sitting on packets due to
4534 * lack of window space */
4535 if (call->tnext < (call->tfirst + call->twind)) {
4536 #ifdef RX_ENABLE_LOCKS
4537 CV_SIGNAL(&call->cv_twind);
4539 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4540 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4541 osi_rxWakeup(&call->twind);
4544 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4545 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4551 * Calculate how many datagrams were successfully received after
4552 * the first missing packet and adjust the negative ack counter
4557 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4558 if (call->nNacks < nNacked) {
4559 call->nNacks = nNacked;
4562 call->nAcks += newAckCount;
4566 /* If the packet contained new acknowledgements, rather than just
4567 * being a duplicate of one we have previously seen, then we can restart
4570 if (newAckCount > 0)
4571 rxi_rto_packet_acked(call, istack);
4573 if (call->flags & RX_CALL_FAST_RECOVER) {
4574 if (newAckCount == 0) {
4575 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4577 call->flags &= ~RX_CALL_FAST_RECOVER;
4578 call->cwind = call->nextCwind;
4579 call->nextCwind = 0;
4582 call->nCwindAcks = 0;
4583 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4584 /* Three negative acks in a row trigger congestion recovery */
4585 call->flags |= RX_CALL_FAST_RECOVER;
4586 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4588 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4589 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4590 call->nextCwind = call->ssthresh;
4593 peer->MTU = call->MTU;
4594 peer->cwind = call->nextCwind;
4595 peer->nDgramPackets = call->nDgramPackets;
4597 call->congestSeq = peer->congestSeq;
4599 /* Reset the resend times on the packets that were nacked
4600 * so we will retransmit as soon as the window permits
4604 for (opr_queue_ScanBackwards(&call->tq, cursor)) {
4605 struct rx_packet *tp =
4606 opr_queue_Entry(cursor, struct rx_packet, entry);
4608 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4609 tp->flags &= ~RX_PKTFLAG_SENT;
4611 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4616 /* If cwind is smaller than ssthresh, then increase
4617 * the window one packet for each ack we receive (exponential
4619 * If cwind is greater than or equal to ssthresh then increase
4620 * the congestion window by one packet for each cwind acks we
4621 * receive (linear growth). */
4622 if (call->cwind < call->ssthresh) {
4624 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4625 call->nCwindAcks = 0;
4627 call->nCwindAcks += newAckCount;
4628 if (call->nCwindAcks >= call->cwind) {
4629 call->nCwindAcks = 0;
4630 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4634 * If we have received several acknowledgements in a row then
4635 * it is time to increase the size of our datagrams
4637 if ((int)call->nAcks > rx_nDgramThreshold) {
4638 if (peer->maxDgramPackets > 1) {
4639 if (call->nDgramPackets < peer->maxDgramPackets) {
4640 call->nDgramPackets++;
4642 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4643 } else if (call->MTU < peer->maxMTU) {
4644 /* don't upgrade if we can't handle it */
4645 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4646 call->MTU = peer->ifMTU;
4648 call->MTU += peer->natMTU;
4649 call->MTU = MIN(call->MTU, peer->maxMTU);
4656 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4658 /* Servers need to hold the call until all response packets have
4659 * been acknowledged. Soft acks are good enough since clients
4660 * are not allowed to clear their receive queues. */
4661 if (call->state == RX_STATE_HOLD
4662 && call->tfirst + call->nSoftAcked >= call->tnext) {
4663 call->state = RX_STATE_DALLY;
4664 rxi_ClearTransmitQueue(call, 0);
4665 rxi_CancelKeepAliveEvent(call);
4666 } else if (!opr_queue_IsEmpty(&call->tq)) {
4667 rxi_Start(call, istack);
4673 * Schedule a connection abort to be sent after some delay.
4675 * @param[in] conn The connection to send the abort on.
4676 * @param[in] msec The number of milliseconds to wait before sending.
4678 * @pre conn_data_lock must be held
4681 rxi_SendConnectionAbortLater(struct rx_connection *conn, int msec)
4683 struct clock when, now;
4685 MUTEX_ASSERT(&conn->conn_data_lock);
4689 if (!conn->delayedAbortEvent) {
4690 clock_GetTime(&now);
4692 clock_Addmsec(&when, msec);
4693 rx_GetConnection(conn);
4694 conn->delayedAbortEvent =
4695 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
4699 /* Received a response to a challenge packet */
4700 static struct rx_packet *
4701 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4702 struct rx_packet *np, int istack)
4706 /* Ignore the packet if we're the client */
4707 if (conn->type == RX_CLIENT_CONNECTION)
4710 /* If already authenticated, ignore the packet (it's probably a retry) */
4711 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4714 if (!conn->securityChallengeSent) {
4715 /* We've never sent out a challenge for this connection, so this
4716 * response cannot possibly be correct; ignore it. This can happen
4717 * if we sent a challenge to the client, then we were restarted, and
4718 * then the client sent us a response. If we ignore the response, the
4719 * client will eventually resend a data packet, causing us to send a
4720 * new challenge and the client to send a new response. */
4724 /* Otherwise, have the security object evaluate the response packet */
4725 error = RXS_CheckResponse(conn->securityObject, conn, np);
4727 /* If the response is invalid, reset the connection, sending
4728 * an abort to the peer. Send the abort with a 1 second delay,
4729 * to avoid a peer hammering us by constantly recreating a
4730 * connection with bad credentials. */
4731 rxi_ConnectionError(conn, error);
4732 MUTEX_ENTER(&conn->conn_data_lock);
4733 rxi_SendConnectionAbortLater(conn, 1000);
4734 MUTEX_EXIT(&conn->conn_data_lock);
4737 /* If the response is valid, any calls waiting to attach
4738 * servers can now do so */
4741 for (i = 0; i < RX_MAXCALLS; i++) {
4742 struct rx_call *call = conn->call[i];
4744 MUTEX_ENTER(&call->lock);
4745 if (call->state == RX_STATE_PRECALL)
4746 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4747 /* tnop can be null if newcallp is null */
4748 MUTEX_EXIT(&call->lock);
4752 /* Update the peer reachability information, just in case
4753 * some calls went into attach-wait while we were waiting
4754 * for authentication..
4756 rxi_UpdatePeerReach(conn, NULL);
4761 /* A client has received an authentication challenge: the security
4762 * object is asked to cough up a respectable response packet to send
4763 * back to the server. The server is responsible for retrying the
4764 * challenge if it fails to get a response. */
4766 static struct rx_packet *
4767 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4768 struct rx_packet *np, int istack)
4772 /* Ignore the challenge if we're the server */
4773 if (conn->type == RX_SERVER_CONNECTION)
4776 /* Ignore the challenge if the connection is otherwise idle; someone's
4777 * trying to use us as an oracle. */
4778 if (!rxi_HasActiveCalls(conn))
4781 /* Send the security object the challenge packet. It is expected to fill
4782 * in the response. */
4783 error = RXS_GetResponse(conn->securityObject, conn, np);
4785 /* If the security object is unable to return a valid response, reset the
4786 * connection and send an abort to the peer. Otherwise send the response
4787 * packet to the peer connection. */
4789 rxi_ConnectionError(conn, error);
4790 MUTEX_ENTER(&conn->conn_data_lock);
4791 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4792 MUTEX_EXIT(&conn->conn_data_lock);
4794 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4795 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4801 /* Find an available server process to service the current request in
4802 * the given call structure. If one isn't available, queue up this
4803 * call so it eventually gets one */
4805 rxi_AttachServerProc(struct rx_call *call,
4806 osi_socket socket, int *tnop,
4807 struct rx_call **newcallp)
4809 struct rx_serverQueueEntry *sq;
4810 struct rx_service *service = call->conn->service;
4813 /* May already be attached */
4814 if (call->state == RX_STATE_ACTIVE)
4817 MUTEX_ENTER(&rx_serverPool_lock);
4819 haveQuota = QuotaOK(service);
4820 if ((!haveQuota) || opr_queue_IsEmpty(&rx_idleServerQueue)) {
4821 /* If there are no processes available to service this call,
4822 * put the call on the incoming call queue (unless it's
4823 * already on the queue).
4825 #ifdef RX_ENABLE_LOCKS
4827 ReturnToServerPool(service);
4828 #endif /* RX_ENABLE_LOCKS */
4830 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4831 call->flags |= RX_CALL_WAIT_PROC;
4832 rx_atomic_inc(&rx_nWaiting);
4833 rx_atomic_inc(&rx_nWaited);
4834 rxi_calltrace(RX_CALL_ARRIVAL, call);
4835 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4836 opr_queue_Append(&rx_incomingCallQueue, &call->entry);
4839 sq = opr_queue_Last(&rx_idleServerQueue,
4840 struct rx_serverQueueEntry, entry);
4842 /* If hot threads are enabled, and both newcallp and sq->socketp
4843 * are non-null, then this thread will process the call, and the
4844 * idle server thread will start listening on this threads socket.
4846 opr_queue_Remove(&sq->entry);
4848 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4851 *sq->socketp = socket;
4852 clock_GetTime(&call->startTime);
4853 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4857 if (call->flags & RX_CALL_WAIT_PROC) {
4858 /* Conservative: I don't think this should happen */
4859 call->flags &= ~RX_CALL_WAIT_PROC;
4860 rx_atomic_dec(&rx_nWaiting);
4861 if (opr_queue_IsOnQueue(&call->entry)) {
4862 opr_queue_Remove(&call->entry);
4865 call->state = RX_STATE_ACTIVE;
4866 call->app.mode = RX_MODE_RECEIVING;
4867 #ifdef RX_KERNEL_TRACE
4869 int glockOwner = ISAFS_GLOCK();
4872 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4873 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4879 if (call->flags & RX_CALL_CLEARED) {
4880 /* send an ack now to start the packet flow up again */
4881 call->flags &= ~RX_CALL_CLEARED;
4882 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4884 #ifdef RX_ENABLE_LOCKS
4887 service->nRequestsRunning++;
4888 MUTEX_ENTER(&rx_quota_mutex);
4889 if (service->nRequestsRunning <= service->minProcs)
4892 MUTEX_EXIT(&rx_quota_mutex);
4896 MUTEX_EXIT(&rx_serverPool_lock);
4899 /* Delay the sending of an acknowledge event for a short while, while
4900 * a new call is being prepared (in the case of a client) or a reply
4901 * is being prepared (in the case of a server). Rather than sending
4902 * an ack packet, an ACKALL packet is sent. */
4904 rxi_AckAll(struct rx_call *call)
4906 rxi_SendSpecial(call, call->conn, NULL, RX_PACKET_TYPE_ACKALL,
4908 call->flags |= RX_CALL_ACKALL_SENT;
4912 * Event handler for per-call delayed acks.
4913 * Also called synchronously, with |event| == NULL, to send a "delayed" ack
4917 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4920 struct rx_call *call = arg1;
4921 #ifdef RX_ENABLE_LOCKS
4923 MUTEX_ENTER(&call->lock);
4924 if (event == call->delayedAckEvent)
4925 rxevent_Put(&call->delayedAckEvent);
4927 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4929 MUTEX_EXIT(&call->lock);
4930 #else /* RX_ENABLE_LOCKS */
4932 rxevent_Put(&call->delayedAckEvent);
4933 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4934 #endif /* RX_ENABLE_LOCKS */
4935 /* Release the call reference for the event that fired. */
4937 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4940 #ifdef RX_ENABLE_LOCKS
4941 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4942 * clearing them out.
4945 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4947 struct opr_queue *cursor;
4950 for (opr_queue_Scan(&call->tq, cursor)) {
4952 = opr_queue_Entry(cursor, struct rx_packet, entry);
4954 p->flags |= RX_PKTFLAG_ACKED;
4959 call->flags |= RX_CALL_TQ_CLEARME;
4960 call->flags |= RX_CALL_TQ_SOME_ACKED;
4963 rxi_rto_cancel(call);
4965 call->tfirst = call->tnext;
4966 call->nSoftAcked = 0;
4968 if (call->flags & RX_CALL_FAST_RECOVER) {
4969 call->flags &= ~RX_CALL_FAST_RECOVER;
4970 call->cwind = call->nextCwind;
4971 call->nextCwind = 0;
4974 CV_SIGNAL(&call->cv_twind);
4976 #endif /* RX_ENABLE_LOCKS */
4979 * Acknowledge the whole transmit queue.
4981 * If we're running without locks, or the transmit queue isn't busy, then
4982 * we can just clear the queue now. Otherwise, we have to mark all of the
4983 * packets as acknowledged, and let rxi_Start clear it later on
4986 rxi_AckAllInTransmitQueue(struct rx_call *call)
4988 #ifdef RX_ENABLE_LOCKS
4989 if (call->flags & RX_CALL_TQ_BUSY) {
4990 rxi_SetAcksInTransmitQueue(call);
4994 rxi_ClearTransmitQueue(call, 0);
4996 /* Clear out the transmit queue for the current call (all packets have
4997 * been received by peer) */
4999 rxi_ClearTransmitQueue(struct rx_call *call, int force)
5001 #ifdef RX_ENABLE_LOCKS
5002 struct opr_queue *cursor;
5003 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
5005 for (opr_queue_Scan(&call->tq, cursor)) {
5007 = opr_queue_Entry(cursor, struct rx_packet, entry);
5009 p->flags |= RX_PKTFLAG_ACKED;
5013 call->flags |= RX_CALL_TQ_CLEARME;
5014 call->flags |= RX_CALL_TQ_SOME_ACKED;
5017 #endif /* RX_ENABLE_LOCKS */
5018 #ifdef RXDEBUG_PACKET
5020 #endif /* RXDEBUG_PACKET */
5021 rxi_FreePackets(0, &call->tq);
5022 rxi_WakeUpTransmitQueue(call);
5023 #ifdef RX_ENABLE_LOCKS
5024 call->flags &= ~RX_CALL_TQ_CLEARME;
5028 rxi_rto_cancel(call);
5029 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5030 call->nSoftAcked = 0;
5032 if (call->flags & RX_CALL_FAST_RECOVER) {
5033 call->flags &= ~RX_CALL_FAST_RECOVER;
5034 call->cwind = call->nextCwind;
5036 #ifdef RX_ENABLE_LOCKS
5037 CV_SIGNAL(&call->cv_twind);
5039 osi_rxWakeup(&call->twind);
5044 rxi_ClearReceiveQueue(struct rx_call *call)
5046 if (!opr_queue_IsEmpty(&call->rq)) {
5049 count = rxi_FreePackets(0, &call->rq);
5050 rx_packetReclaims += count;
5051 #ifdef RXDEBUG_PACKET
5053 if ( call->rqc != 0 )
5054 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5056 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5058 if (call->state == RX_STATE_PRECALL) {
5059 call->flags |= RX_CALL_CLEARED;
5063 /* Send an abort packet for the specified call */
5064 static struct rx_packet *
5065 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5066 int istack, int force)
5069 struct clock when, now;
5074 /* Clients should never delay abort messages */
5075 if (rx_IsClientConn(call->conn))
5078 if (call->abortCode != call->error) {
5079 call->abortCode = call->error;
5080 call->abortCount = 0;
5083 if (force || rxi_callAbortThreshhold == 0
5084 || call->abortCount < rxi_callAbortThreshhold) {
5085 rxi_CancelDelayedAbortEvent(call);
5086 error = htonl(call->error);
5089 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5090 (char *)&error, sizeof(error), istack);
5091 } else if (!call->delayedAbortEvent) {
5092 clock_GetTime(&now);
5094 clock_Addmsec(&when, rxi_callAbortDelay);
5095 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5096 call->delayedAbortEvent =
5097 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5103 rxi_CancelDelayedAbortEvent(struct rx_call *call)
5105 MUTEX_ASSERT(&call->lock);
5106 if (rxevent_Cancel(&call->delayedAbortEvent))
5107 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5110 /* Send an abort packet for the specified connection. Packet is an
5111 * optional pointer to a packet that can be used to send the abort.
5112 * Once the number of abort messages reaches the threshhold, an
5113 * event is scheduled to send the abort. Setting the force flag
5114 * overrides sending delayed abort messages.
5116 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5117 * to send the abort packet.
5120 rxi_SendConnectionAbort(struct rx_connection *conn,
5121 struct rx_packet *packet, int istack, int force)
5128 /* Clients should never delay abort messages */
5129 if (rx_IsClientConn(conn))
5132 if (force || rxi_connAbortThreshhold == 0
5133 || conn->abortCount < rxi_connAbortThreshhold) {
5135 if (rxevent_Cancel(&conn->delayedAbortEvent))
5136 putConnection(conn);
5137 error = htonl(conn->error);
5139 MUTEX_EXIT(&conn->conn_data_lock);
5141 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5142 RX_PACKET_TYPE_ABORT, (char *)&error,
5143 sizeof(error), istack);
5144 MUTEX_ENTER(&conn->conn_data_lock);
5146 rxi_SendConnectionAbortLater(conn, rxi_connAbortDelay);
5151 /* Associate an error all of the calls owned by a connection. Called
5152 * with error non-zero. This is only for really fatal things, like
5153 * bad authentication responses. The connection itself is set in
5154 * error at this point, so that future packets received will be
5157 rxi_ConnectionError(struct rx_connection *conn,
5163 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5165 MUTEX_ENTER(&conn->conn_data_lock);
5166 if (rxevent_Cancel(&conn->challengeEvent))
5167 putConnection(conn);
5168 if (rxevent_Cancel(&conn->natKeepAliveEvent))
5169 putConnection(conn);
5170 if (rxevent_Cancel(&conn->checkReachEvent)) {
5171 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5172 putConnection(conn);
5174 MUTEX_EXIT(&conn->conn_data_lock);
5175 for (i = 0; i < RX_MAXCALLS; i++) {
5176 struct rx_call *call = conn->call[i];
5178 MUTEX_ENTER(&call->lock);
5179 rxi_CallError(call, error);
5180 MUTEX_EXIT(&call->lock);
5183 conn->error = error;
5184 if (rx_stats_active)
5185 rx_atomic_inc(&rx_stats.fatalErrors);
5190 * Interrupt an in-progress call with the specified error and wakeup waiters.
5192 * @param[in] call The call to interrupt
5193 * @param[in] error The error code to send to the peer
5196 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5198 MUTEX_ENTER(&call->lock);
5199 rxi_CallError(call, error);
5200 rxi_SendCallAbort(call, NULL, 0, 1);
5201 MUTEX_EXIT(&call->lock);
5205 rxi_CallError(struct rx_call *call, afs_int32 error)
5207 MUTEX_ASSERT(&call->lock);
5208 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5210 error = call->error;
5212 #ifdef RX_ENABLE_LOCKS
5213 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5214 rxi_ResetCall(call, 0);
5217 rxi_ResetCall(call, 0);
5219 call->error = error;
5222 /* Reset various fields in a call structure, and wakeup waiting
5223 * processes. Some fields aren't changed: state & mode are not
5224 * touched (these must be set by the caller), and bufptr, nLeft, and
5225 * nFree are not reset, since these fields are manipulated by
5226 * unprotected macros, and may only be reset by non-interrupting code.
5230 rxi_ResetCall(struct rx_call *call, int newcall)
5233 struct rx_peer *peer;
5234 struct rx_packet *packet;
5236 MUTEX_ASSERT(&call->lock);
5237 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5239 /* Notify anyone who is waiting for asynchronous packet arrival */
5240 if (call->arrivalProc) {
5241 (*call->arrivalProc) (call, call->arrivalProcHandle,
5242 call->arrivalProcArg);
5243 call->arrivalProc = (void (*)())0;
5247 rxi_CancelGrowMTUEvent(call);
5249 if (call->delayedAbortEvent) {
5250 rxi_CancelDelayedAbortEvent(call);
5251 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5253 rxi_SendCallAbort(call, packet, 0, 1);
5254 rxi_FreePacket(packet);
5259 * Update the peer with the congestion information in this call
5260 * so other calls on this connection can pick up where this call
5261 * left off. If the congestion sequence numbers don't match then
5262 * another call experienced a retransmission.
5264 peer = call->conn->peer;
5265 MUTEX_ENTER(&peer->peer_lock);
5267 if (call->congestSeq == peer->congestSeq) {
5268 peer->cwind = MAX(peer->cwind, call->cwind);
5269 peer->MTU = MAX(peer->MTU, call->MTU);
5270 peer->nDgramPackets =
5271 MAX(peer->nDgramPackets, call->nDgramPackets);
5274 call->abortCode = 0;
5275 call->abortCount = 0;
5277 if (peer->maxDgramPackets > 1) {
5278 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5280 call->MTU = peer->MTU;
5282 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5283 call->ssthresh = rx_maxSendWindow;
5284 call->nDgramPackets = peer->nDgramPackets;
5285 call->congestSeq = peer->congestSeq;
5286 call->rtt = peer->rtt;
5287 call->rtt_dev = peer->rtt_dev;
5288 clock_Zero(&call->rto);
5289 clock_Addmsec(&call->rto,
5290 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5291 MUTEX_EXIT(&peer->peer_lock);
5293 flags = call->flags;
5294 rxi_WaitforTQBusy(call);
5296 rxi_ClearTransmitQueue(call, 1);
5297 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5298 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5302 rxi_ClearReceiveQueue(call);
5303 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5307 call->twind = call->conn->twind[call->channel];
5308 call->rwind = call->conn->rwind[call->channel];
5309 call->nSoftAcked = 0;
5310 call->nextCwind = 0;
5313 call->nCwindAcks = 0;
5314 call->nSoftAcks = 0;
5315 call->nHardAcks = 0;
5317 call->tfirst = call->rnext = call->tnext = 1;
5320 call->lastAcked = 0;
5321 call->localStatus = call->remoteStatus = 0;
5323 if (flags & RX_CALL_READER_WAIT) {
5324 #ifdef RX_ENABLE_LOCKS
5325 CV_BROADCAST(&call->cv_rq);
5327 osi_rxWakeup(&call->rq);
5330 if (flags & RX_CALL_WAIT_PACKETS) {
5331 MUTEX_ENTER(&rx_freePktQ_lock);
5332 rxi_PacketsUnWait(); /* XXX */
5333 MUTEX_EXIT(&rx_freePktQ_lock);
5335 #ifdef RX_ENABLE_LOCKS
5336 CV_SIGNAL(&call->cv_twind);
5338 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5339 osi_rxWakeup(&call->twind);
5342 if (flags & RX_CALL_WAIT_PROC) {
5343 rx_atomic_dec(&rx_nWaiting);
5345 #ifdef RX_ENABLE_LOCKS
5346 /* The following ensures that we don't mess with any queue while some
5347 * other thread might also be doing so. The call_queue_lock field is
5348 * is only modified under the call lock. If the call is in the process
5349 * of being removed from a queue, the call is not locked until the
5350 * the queue lock is dropped and only then is the call_queue_lock field
5351 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5352 * Note that any other routine which removes a call from a queue has to
5353 * obtain the queue lock before examing the queue and removing the call.
5355 if (call->call_queue_lock) {
5356 MUTEX_ENTER(call->call_queue_lock);
5357 if (opr_queue_IsOnQueue(&call->entry)) {
5358 opr_queue_Remove(&call->entry);
5360 MUTEX_EXIT(call->call_queue_lock);
5361 CLEAR_CALL_QUEUE_LOCK(call);
5363 #else /* RX_ENABLE_LOCKS */
5364 if (opr_queue_IsOnQueue(&call->entry)) {
5365 opr_queue_Remove(&call->entry);
5367 #endif /* RX_ENABLE_LOCKS */
5369 rxi_CancelKeepAliveEvent(call);
5370 rxi_CancelDelayedAckEvent(call);
5373 /* Send an acknowledge for the indicated packet (seq,serial) of the
5374 * indicated call, for the indicated reason (reason). This
5375 * acknowledge will specifically acknowledge receiving the packet, and
5376 * will also specify which other packets for this call have been
5377 * received. This routine returns the packet that was used to the
5378 * caller. The caller is responsible for freeing it or re-using it.
5379 * This acknowledgement also returns the highest sequence number
5380 * actually read out by the higher level to the sender; the sender
5381 * promises to keep around packets that have not been read by the
5382 * higher level yet (unless, of course, the sender decides to abort
5383 * the call altogether). Any of p, seq, serial, pflags, or reason may
5384 * be set to zero without ill effect. That is, if they are zero, they
5385 * will not convey any information.
5386 * NOW there is a trailer field, after the ack where it will safely be
5387 * ignored by mundanes, which indicates the maximum size packet this
5388 * host can swallow. */
5390 struct rx_packet *optionalPacket; use to send ack (or null)
5391 int seq; Sequence number of the packet we are acking
5392 int serial; Serial number of the packet
5393 int pflags; Flags field from packet header
5394 int reason; Reason an acknowledge was prompted
5397 #define RX_ZEROS 1024
5398 static char rx_zeros[RX_ZEROS];
5401 rxi_SendAck(struct rx_call *call,
5402 struct rx_packet *optionalPacket, int serial, int reason,
5405 struct rx_ackPacket *ap;
5406 struct rx_packet *p;
5407 struct opr_queue *cursor;
5410 afs_uint32 padbytes = 0;
5411 #ifdef RX_ENABLE_TSFPQ
5412 struct rx_ts_info_t * rx_ts_info;
5416 * Open the receive window once a thread starts reading packets
5418 if (call->rnext > 1) {
5419 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5422 /* Don't attempt to grow MTU if this is a critical ping */
5423 if (reason == RX_ACK_MTU) {
5424 /* keep track of per-call attempts, if we're over max, do in small
5425 * otherwise in larger? set a size to increment by, decrease
5428 if (call->conn->peer->maxPacketSize &&
5429 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5431 padbytes = call->conn->peer->maxPacketSize+16;
5433 padbytes = call->conn->peer->maxMTU + 128;
5435 /* do always try a minimum size ping */
5436 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5438 /* subtract the ack payload */
5439 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5440 reason = RX_ACK_PING;
5443 call->nHardAcks = 0;
5444 call->nSoftAcks = 0;
5445 if (call->rnext > call->lastAcked)
5446 call->lastAcked = call->rnext;
5450 rx_computelen(p, p->length); /* reset length, you never know */
5451 } /* where that's been... */
5452 #ifdef RX_ENABLE_TSFPQ
5454 RX_TS_INFO_GET(rx_ts_info);
5455 if ((p = rx_ts_info->local_special_packet)) {
5456 rx_computelen(p, p->length);
5457 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5458 rx_ts_info->local_special_packet = p;
5459 } else { /* We won't send the ack, but don't panic. */
5460 return optionalPacket;
5464 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5465 /* We won't send the ack, but don't panic. */
5466 return optionalPacket;
5471 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5474 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5475 #ifndef RX_ENABLE_TSFPQ
5476 if (!optionalPacket)
5479 return optionalPacket;
5481 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5482 if (rx_Contiguous(p) < templ) {
5483 #ifndef RX_ENABLE_TSFPQ
5484 if (!optionalPacket)
5487 return optionalPacket;
5492 /* MTUXXX failing to send an ack is very serious. We should */
5493 /* try as hard as possible to send even a partial ack; it's */
5494 /* better than nothing. */
5495 ap = (struct rx_ackPacket *)rx_DataOf(p);
5496 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5497 ap->reason = reason;
5499 /* The skew computation used to be bogus, I think it's better now. */
5500 /* We should start paying attention to skew. XXX */
5501 ap->serial = htonl(serial);
5502 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5505 * First packet not yet forwarded to reader. When ACKALL has been
5506 * sent the peer has been told that all received packets will be
5507 * delivered to the reader. The value 'rnext' is used internally
5508 * to refer to the next packet in the receive queue that must be
5509 * delivered to the reader. From the perspective of the peer it
5510 * already has so report the last sequence number plus one if there
5511 * are packets in the receive queue awaiting processing.
5513 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5514 !opr_queue_IsEmpty(&call->rq)) {
5515 ap->firstPacket = htonl(opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq + 1);
5517 ap->firstPacket = htonl(call->rnext);
5519 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5521 /* No fear of running out of ack packet here because there can only
5522 * be at most one window full of unacknowledged packets. The window
5523 * size must be constrained to be less than the maximum ack size,
5524 * of course. Also, an ack should always fit into a single packet
5525 * -- it should not ever be fragmented. */
5527 for (opr_queue_Scan(&call->rq, cursor)) {
5528 struct rx_packet *rqp
5529 = opr_queue_Entry(cursor, struct rx_packet, entry);
5531 if (!rqp || !call->rq.next
5532 || (rqp->header.seq > (call->rnext + call->rwind))) {
5533 #ifndef RX_ENABLE_TSFPQ
5534 if (!optionalPacket)
5537 rxi_CallError(call, RX_CALL_DEAD);
5538 return optionalPacket;
5541 while (rqp->header.seq > call->rnext + offset)
5542 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5543 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5545 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5546 #ifndef RX_ENABLE_TSFPQ
5547 if (!optionalPacket)
5550 rxi_CallError(call, RX_CALL_DEAD);
5551 return optionalPacket;
5557 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5559 /* Must zero the 3 octets that rx_AckDataSize skips at the end of the
5562 rx_packetwrite(p, rx_AckDataSize(offset) - 3, 3, rx_zeros);
5564 /* these are new for AFS 3.3 */
5565 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5566 templ = htonl(templ);
5567 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5568 templ = htonl(call->conn->peer->ifMTU);
5569 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5570 sizeof(afs_int32), &templ);
5572 /* new for AFS 3.4 */
5573 templ = htonl(call->rwind);
5574 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5575 sizeof(afs_int32), &templ);
5577 /* new for AFS 3.5 */
5578 templ = htonl(call->conn->peer->ifDgramPackets);
5579 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5580 sizeof(afs_int32), &templ);
5582 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5584 p->header.serviceId = call->conn->serviceId;
5585 p->header.cid = (call->conn->cid | call->channel);
5586 p->header.callNumber = *call->callNumber;
5588 p->header.securityIndex = call->conn->securityIndex;
5589 p->header.epoch = call->conn->epoch;
5590 p->header.type = RX_PACKET_TYPE_ACK;
5591 p->header.flags = RX_SLOW_START_OK;
5592 if (reason == RX_ACK_PING)
5593 p->header.flags |= RX_REQUEST_ACK;
5595 while (padbytes > 0) {
5596 if (padbytes > RX_ZEROS) {
5597 rx_packetwrite(p, p->length, RX_ZEROS, rx_zeros);
5598 p->length += RX_ZEROS;
5599 padbytes -= RX_ZEROS;
5601 rx_packetwrite(p, p->length, padbytes, rx_zeros);
5602 p->length += padbytes;
5607 if (call->conn->type == RX_CLIENT_CONNECTION)
5608 p->header.flags |= RX_CLIENT_INITIATED;
5612 if (rxdebug_active) {
5616 len = _snprintf(msg, sizeof(msg),
5617 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5618 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5619 ntohl(ap->serial), ntohl(ap->previousPacket),
5620 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5621 ap->nAcks, ntohs(ap->bufferSpace) );
5625 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5626 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5630 OutputDebugString(msg);
5632 #else /* AFS_NT40_ENV */
5634 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5635 ap->reason, ntohl(ap->previousPacket),
5636 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5638 for (offset = 0; offset < ap->nAcks; offset++)
5639 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5644 #endif /* AFS_NT40_ENV */
5647 int i, nbytes = p->length;
5649 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5650 if (nbytes <= p->wirevec[i].iov_len) {
5653 savelen = p->wirevec[i].iov_len;
5655 p->wirevec[i].iov_len = nbytes;
5657 rxi_Send(call, p, istack);
5658 p->wirevec[i].iov_len = savelen;
5662 nbytes -= p->wirevec[i].iov_len;
5665 if (rx_stats_active)
5666 rx_atomic_inc(&rx_stats.ackPacketsSent);
5667 #ifndef RX_ENABLE_TSFPQ
5668 if (!optionalPacket)
5671 return optionalPacket; /* Return packet for re-use by caller */
5675 struct rx_packet **list;
5680 /* Send all of the packets in the list in single datagram */
5682 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5683 int istack, int moreFlag)
5689 struct rx_connection *conn = call->conn;
5690 struct rx_peer *peer = conn->peer;
5692 MUTEX_ENTER(&peer->peer_lock);
5693 peer->nSent += xmit->len;
5694 if (xmit->resending)
5695 peer->reSends += xmit->len;
5696 MUTEX_EXIT(&peer->peer_lock);
5698 if (rx_stats_active) {
5699 if (xmit->resending)
5700 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5702 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5705 clock_GetTime(&now);
5707 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5711 /* Set the packet flags and schedule the resend events */
5712 /* Only request an ack for the last packet in the list */
5713 for (i = 0; i < xmit->len; i++) {
5714 struct rx_packet *packet = xmit->list[i];
5716 /* Record the time sent */
5717 packet->timeSent = now;
5718 packet->flags |= RX_PKTFLAG_SENT;
5720 /* Ask for an ack on retransmitted packets, on every other packet
5721 * if the peer doesn't support slow start. Ask for an ack on every
5722 * packet until the congestion window reaches the ack rate. */
5723 if (packet->header.serial) {
5726 packet->firstSent = now;
5727 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5728 || (!(call->flags & RX_CALL_SLOW_START_OK)
5729 && (packet->header.seq & 1)))) {
5734 /* Tag this packet as not being the last in this group,
5735 * for the receiver's benefit */
5736 if (i < xmit->len - 1 || moreFlag) {
5737 packet->header.flags |= RX_MORE_PACKETS;
5742 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5745 /* Since we're about to send a data packet to the peer, it's
5746 * safe to nuke any scheduled end-of-packets ack */
5747 rxi_CancelDelayedAckEvent(call);
5749 MUTEX_EXIT(&call->lock);
5750 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5751 if (xmit->len > 1) {
5752 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5754 rxi_SendPacket(call, conn, xmit->list[0], istack);
5756 MUTEX_ENTER(&call->lock);
5757 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5759 /* Tell the RTO calculation engine that we have sent a packet, and
5760 * if it was the last one */
5761 rxi_rto_packet_sent(call, lastPacket, istack);
5763 /* Update last send time for this call (for keep-alive
5764 * processing), and for the connection (so that we can discover
5765 * idle connections) */
5766 conn->lastSendTime = call->lastSendTime = clock_Sec();
5769 /* When sending packets we need to follow these rules:
5770 * 1. Never send more than maxDgramPackets in a jumbogram.
5771 * 2. Never send a packet with more than two iovecs in a jumbogram.
5772 * 3. Never send a retransmitted packet in a jumbogram.
5773 * 4. Never send more than cwind/4 packets in a jumbogram
5774 * We always keep the last list we should have sent so we
5775 * can set the RX_MORE_PACKETS flags correctly.
5779 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5784 struct xmitlist working;
5785 struct xmitlist last;
5787 struct rx_peer *peer = call->conn->peer;
5788 int morePackets = 0;
5790 memset(&last, 0, sizeof(struct xmitlist));
5791 working.list = &list[0];
5793 working.resending = 0;
5795 recovery = call->flags & RX_CALL_FAST_RECOVER;
5797 for (i = 0; i < len; i++) {
5798 /* Does the current packet force us to flush the current list? */
5800 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5801 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5803 /* This sends the 'last' list and then rolls the current working
5804 * set into the 'last' one, and resets the working set */
5807 rxi_SendList(call, &last, istack, 1);
5808 /* If the call enters an error state stop sending, or if
5809 * we entered congestion recovery mode, stop sending */
5811 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5816 working.resending = 0;
5817 working.list = &list[i];
5819 /* Add the current packet to the list if it hasn't been acked.
5820 * Otherwise adjust the list pointer to skip the current packet. */
5821 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5824 if (list[i]->header.serial)
5825 working.resending = 1;
5827 /* Do we need to flush the list? */
5828 if (working.len >= (int)peer->maxDgramPackets
5829 || working.len >= (int)call->nDgramPackets
5830 || working.len >= (int)call->cwind
5831 || list[i]->header.serial
5832 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5834 rxi_SendList(call, &last, istack, 1);
5835 /* If the call enters an error state stop sending, or if
5836 * we entered congestion recovery mode, stop sending */
5838 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5843 working.resending = 0;
5844 working.list = &list[i + 1];
5847 if (working.len != 0) {
5848 osi_Panic("rxi_SendList error");
5850 working.list = &list[i + 1];
5854 /* Send the whole list when the call is in receive mode, when
5855 * the call is in eof mode, when we are in fast recovery mode,
5856 * and when we have the last packet */
5857 /* XXX - The accesses to app.mode aren't safe, as this may be called by
5858 * the listener or event threads
5860 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5861 || (call->flags & RX_CALL_FLUSH)
5862 || (call->flags & RX_CALL_FAST_RECOVER)) {
5863 /* Check for the case where the current list contains
5864 * an acked packet. Since we always send retransmissions
5865 * in a separate packet, we only need to check the first
5866 * packet in the list */
5867 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5871 rxi_SendList(call, &last, istack, morePackets);
5872 /* If the call enters an error state stop sending, or if
5873 * we entered congestion recovery mode, stop sending */
5875 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5879 rxi_SendList(call, &working, istack, 0);
5881 } else if (last.len > 0) {
5882 rxi_SendList(call, &last, istack, 0);
5883 /* Packets which are in 'working' are not sent by this call */
5888 * Check if the peer for the given call is known to be dead
5890 * If the call's peer appears dead (it has encountered fatal network errors
5891 * since the call started) the call is killed with RX_CALL_DEAD if the call
5892 * is active. Otherwise, we do nothing.
5894 * @param[in] call The call to check
5897 * @retval 0 The call is fine, and we haven't done anything to the call
5898 * @retval nonzero The call's peer appears dead, and the call has been
5899 * terminated if it was active
5901 * @pre call->lock must be locked
5904 rxi_CheckPeerDead(struct rx_call *call)
5906 #ifdef AFS_RXERRQ_ENV
5909 if (call->state == RX_STATE_DALLY) {
5913 peererrs = rx_atomic_read(&call->conn->peer->neterrs);
5914 if (call->neterr_gen < peererrs) {
5915 /* we have received network errors since this call started; kill
5917 if (call->state == RX_STATE_ACTIVE) {
5918 rxi_CallError(call, RX_CALL_DEAD);
5922 if (call->neterr_gen > peererrs) {
5923 /* someone has reset the number of peer errors; set the call error gen
5924 * so we can detect if more errors are encountered */
5925 call->neterr_gen = peererrs;
5932 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
5934 struct rx_call *call = arg0;
5935 struct rx_peer *peer;
5936 struct opr_queue *cursor;
5937 struct clock maxTimeout = { 60, 0 };
5939 MUTEX_ENTER(&call->lock);
5941 peer = call->conn->peer;
5943 /* Make sure that the event pointer is removed from the call
5944 * structure, since there is no longer a per-call retransmission
5946 if (event == call->resendEvent)
5947 rxevent_Put(&call->resendEvent);
5949 rxi_CheckPeerDead(call);
5951 if (opr_queue_IsEmpty(&call->tq)) {
5952 /* Nothing to do. This means that we've been raced, and that an
5953 * ACK has come in between when we were triggered, and when we
5954 * actually got to run. */
5958 /* We're in loss recovery */
5959 call->flags |= RX_CALL_FAST_RECOVER;
5961 /* Mark all of the pending packets in the queue as being lost */
5962 for (opr_queue_Scan(&call->tq, cursor)) {
5963 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
5964 if (!(p->flags & RX_PKTFLAG_ACKED))
5965 p->flags &= ~RX_PKTFLAG_SENT;
5968 /* We're resending, so we double the timeout of the call. This will be
5969 * dropped back down by the first successful ACK that we receive.
5971 * We apply a maximum value here of 60 seconds
5973 clock_Add(&call->rto, &call->rto);
5974 if (clock_Gt(&call->rto, &maxTimeout))
5975 call->rto = maxTimeout;
5977 /* Packet loss is most likely due to congestion, so drop our window size
5978 * and start again from the beginning */
5979 if (peer->maxDgramPackets >1) {
5980 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5981 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5983 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5984 call->nDgramPackets = 1;
5986 call->nextCwind = 1;
5989 MUTEX_ENTER(&peer->peer_lock);
5990 peer->MTU = call->MTU;
5991 peer->cwind = call->cwind;
5992 peer->nDgramPackets = 1;
5994 call->congestSeq = peer->congestSeq;
5995 MUTEX_EXIT(&peer->peer_lock);
5997 rxi_Start(call, istack);
6000 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
6001 MUTEX_EXIT(&call->lock);
6004 /* This routine is called when new packets are readied for
6005 * transmission and when retransmission may be necessary, or when the
6006 * transmission window or burst count are favourable. This should be
6007 * better optimized for new packets, the usual case, now that we've
6008 * got rid of queues of send packets. XXXXXXXXXXX */
6010 rxi_Start(struct rx_call *call, int istack)
6012 struct opr_queue *cursor;
6013 #ifdef RX_ENABLE_LOCKS
6014 struct opr_queue *store;
6020 #ifdef RX_ENABLE_LOCKS
6021 if (rx_stats_active)
6022 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
6027 if (!opr_queue_IsEmpty(&call->tq)) { /* If we have anything to send */
6028 /* Send (or resend) any packets that need it, subject to
6029 * window restrictions and congestion burst control
6030 * restrictions. Ask for an ack on the last packet sent in
6031 * this burst. For now, we're relying upon the window being
6032 * considerably bigger than the largest number of packets that
6033 * are typically sent at once by one initial call to
6034 * rxi_Start. This is probably bogus (perhaps we should ask
6035 * for an ack when we're half way through the current
6036 * window?). Also, for non file transfer applications, this
6037 * may end up asking for an ack for every packet. Bogus. XXXX
6040 * But check whether we're here recursively, and let the other guy
6043 #ifdef RX_ENABLE_LOCKS
6044 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6045 call->flags |= RX_CALL_TQ_BUSY;
6047 #endif /* RX_ENABLE_LOCKS */
6049 #ifdef RX_ENABLE_LOCKS
6050 call->flags &= ~RX_CALL_NEED_START;
6051 #endif /* RX_ENABLE_LOCKS */
6053 maxXmitPackets = MIN(call->twind, call->cwind);
6054 for (opr_queue_Scan(&call->tq, cursor)) {
6056 = opr_queue_Entry(cursor, struct rx_packet, entry);
6058 if (p->flags & RX_PKTFLAG_ACKED) {
6059 /* Since we may block, don't trust this */
6060 if (rx_stats_active)
6061 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6062 continue; /* Ignore this packet if it has been acknowledged */
6065 /* Turn off all flags except these ones, which are the same
6066 * on each transmission */
6067 p->header.flags &= RX_PRESET_FLAGS;
6069 if (p->header.seq >=
6070 call->tfirst + MIN((int)call->twind,
6071 (int)(call->nSoftAcked +
6073 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6074 /* Note: if we're waiting for more window space, we can
6075 * still send retransmits; hence we don't return here, but
6076 * break out to schedule a retransmit event */
6077 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6078 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6083 /* Transmit the packet if it needs to be sent. */
6084 if (!(p->flags & RX_PKTFLAG_SENT)) {
6085 if (nXmitPackets == maxXmitPackets) {
6086 rxi_SendXmitList(call, call->xmitList,
6087 nXmitPackets, istack);
6090 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6091 *(call->callNumber), p));
6092 call->xmitList[nXmitPackets++] = p;
6094 } /* end of the queue_Scan */
6096 /* xmitList now hold pointers to all of the packets that are
6097 * ready to send. Now we loop to send the packets */
6098 if (nXmitPackets > 0) {
6099 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6103 #ifdef RX_ENABLE_LOCKS
6105 /* We went into the error state while sending packets. Now is
6106 * the time to reset the call. This will also inform the using
6107 * process that the call is in an error state.
6109 if (rx_stats_active)
6110 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6111 call->flags &= ~RX_CALL_TQ_BUSY;
6112 rxi_WakeUpTransmitQueue(call);
6113 rxi_CallError(call, call->error);
6117 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6119 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6120 /* Some packets have received acks. If they all have, we can clear
6121 * the transmit queue.
6124 for (opr_queue_ScanSafe(&call->tq, cursor, store)) {
6126 = opr_queue_Entry(cursor, struct rx_packet, entry);
6128 if (p->header.seq < call->tfirst
6129 && (p->flags & RX_PKTFLAG_ACKED)) {
6130 opr_queue_Remove(&p->entry);
6131 #ifdef RX_TRACK_PACKETS
6132 p->flags &= ~RX_PKTFLAG_TQ;
6134 #ifdef RXDEBUG_PACKET
6142 call->flags |= RX_CALL_TQ_CLEARME;
6144 if (call->flags & RX_CALL_TQ_CLEARME)
6145 rxi_ClearTransmitQueue(call, 1);
6146 } while (call->flags & RX_CALL_NEED_START);
6148 * TQ references no longer protected by this flag; they must remain
6149 * protected by the call lock.
6151 call->flags &= ~RX_CALL_TQ_BUSY;
6152 rxi_WakeUpTransmitQueue(call);
6154 call->flags |= RX_CALL_NEED_START;
6156 #endif /* RX_ENABLE_LOCKS */
6158 rxi_rto_cancel(call);
6162 /* Also adjusts the keep alive parameters for the call, to reflect
6163 * that we have just sent a packet (so keep alives aren't sent
6166 rxi_Send(struct rx_call *call, struct rx_packet *p,
6169 struct rx_connection *conn = call->conn;
6171 /* Stamp each packet with the user supplied status */
6172 p->header.userStatus = call->localStatus;
6174 /* Allow the security object controlling this call's security to
6175 * make any last-minute changes to the packet */
6176 RXS_SendPacket(conn->securityObject, call, p);
6178 /* Since we're about to send SOME sort of packet to the peer, it's
6179 * safe to nuke any scheduled end-of-packets ack */
6180 rxi_CancelDelayedAckEvent(call);
6182 /* Actually send the packet, filling in more connection-specific fields */
6183 MUTEX_EXIT(&call->lock);
6184 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6185 rxi_SendPacket(call, conn, p, istack);
6186 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6187 MUTEX_ENTER(&call->lock);
6189 /* Update last send time for this call (for keep-alive
6190 * processing), and for the connection (so that we can discover
6191 * idle connections) */
6192 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6193 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6194 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6196 conn->lastSendTime = call->lastSendTime = clock_Sec();
6200 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6201 * that things are fine. Also called periodically to guarantee that nothing
6202 * falls through the cracks (e.g. (error + dally) connections have keepalive
6203 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6205 * haveCTLock Set if calling from rxi_ReapConnections
6208 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6210 struct rx_connection *conn = call->conn;
6212 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6213 afs_uint32 fudgeFactor;
6216 int idle_timeout = 0;
6217 afs_int32 clock_diff = 0;
6219 if (rxi_CheckPeerDead(call)) {
6225 /* Large swings in the clock can have a significant impact on
6226 * the performance of RX call processing. Forward clock shifts
6227 * will result in premature event triggering or timeouts.
6228 * Backward shifts can result in calls not completing until
6229 * the clock catches up with the original start clock value.
6231 * If a backward clock shift of more than five minutes is noticed,
6232 * just fail the call.
6234 if (now < call->lastSendTime)
6235 clock_diff = call->lastSendTime - now;
6236 if (now < call->startWait)
6237 clock_diff = MAX(clock_diff, call->startWait - now);
6238 if (now < call->lastReceiveTime)
6239 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6240 if (clock_diff > 5 * 60)
6242 if (call->state == RX_STATE_ACTIVE)
6243 rxi_CallError(call, RX_CALL_TIMEOUT);
6247 #ifdef RX_ENABLE_LOCKS
6248 if (call->flags & RX_CALL_TQ_BUSY) {
6249 /* Call is active and will be reset by rxi_Start if it's
6250 * in an error state.
6255 /* RTT + 8*MDEV, rounded up to the next second. */
6256 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6257 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6259 deadTime = conn->secondsUntilDead + fudgeFactor;
6260 /* These are computed to the second (+- 1 second). But that's
6261 * good enough for these values, which should be a significant
6262 * number of seconds. */
6263 if (now > (call->lastReceiveTime + deadTime)) {
6264 if (call->state == RX_STATE_ACTIVE) {
6265 cerror = RX_CALL_DEAD;
6268 #ifdef RX_ENABLE_LOCKS
6269 /* Cancel pending events */
6270 rxi_CancelDelayedAckEvent(call);
6271 rxi_rto_cancel(call);
6272 rxi_CancelKeepAliveEvent(call);
6273 rxi_CancelGrowMTUEvent(call);
6274 MUTEX_ENTER(&rx_refcnt_mutex);
6275 /* if rxi_FreeCall returns 1 it has freed the call */
6276 if (call->refCount == 0 &&
6277 rxi_FreeCall(call, haveCTLock))
6279 MUTEX_EXIT(&rx_refcnt_mutex);
6282 MUTEX_EXIT(&rx_refcnt_mutex);
6284 #else /* RX_ENABLE_LOCKS */
6285 rxi_FreeCall(call, 0);
6287 #endif /* RX_ENABLE_LOCKS */
6289 /* Non-active calls are destroyed if they are not responding
6290 * to pings; active calls are simply flagged in error, so the
6291 * attached process can die reasonably gracefully. */
6294 if (conn->idleDeadTime) {
6295 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6299 /* see if we have a non-activity timeout */
6300 if (call->startWait && ((call->startWait + idleDeadTime) < now)) {
6301 if (call->state == RX_STATE_ACTIVE) {
6302 cerror = RX_CALL_TIMEOUT;
6308 if (conn->hardDeadTime) {
6309 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6312 /* see if we have a hard timeout */
6314 && (now > (hardDeadTime + call->startTime.sec))) {
6315 if (call->state == RX_STATE_ACTIVE)
6316 rxi_CallError(call, RX_CALL_TIMEOUT);
6321 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6322 call->lastReceiveTime) {
6323 int oldMTU = conn->peer->ifMTU;
6325 /* If we thought we could send more, perhaps things got worse.
6326 * Shrink by 128 bytes and try again. */
6327 if (conn->peer->maxPacketSize < conn->lastPacketSize)
6328 /* maxPacketSize will be cleared in rxi_SetPeerMtu */
6329 newmtu = MAX(conn->peer->maxPacketSize + RX_HEADER_SIZE,
6330 conn->lastPacketSize - 128 + RX_HEADER_SIZE);
6332 newmtu = conn->lastPacketSize - 128 + RX_HEADER_SIZE;
6334 /* minimum capped in SetPeerMtu */
6335 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6338 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
6340 /* needed so ResetCall doesn't clobber us. */
6341 call->MTU = conn->peer->ifMTU;
6343 /* if we never succeeded, let the error pass out as-is */
6344 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6345 cerror = conn->msgsizeRetryErr;
6348 rxi_CallError(call, cerror);
6353 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6354 void *dummy, int dummy2)
6356 struct rx_connection *conn = arg1;
6357 struct rx_header theader;
6358 char tbuffer[1 + sizeof(struct rx_header)];
6359 struct sockaddr_in taddr;
6363 struct iovec tmpiov[2];
6366 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6369 tp = &tbuffer[sizeof(struct rx_header)];
6370 taddr.sin_family = AF_INET;
6371 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6372 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6373 memset(&taddr.sin_zero, 0, sizeof(taddr.sin_zero));
6374 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6375 taddr.sin_len = sizeof(struct sockaddr_in);
6377 memset(&theader, 0, sizeof(theader));
6378 theader.epoch = htonl(999);
6380 theader.callNumber = 0;
6383 theader.type = RX_PACKET_TYPE_VERSION;
6384 theader.flags = RX_LAST_PACKET;
6385 theader.serviceId = 0;
6387 memcpy(tbuffer, &theader, sizeof(theader));
6388 memcpy(tp, &a, sizeof(a));
6389 tmpiov[0].iov_base = tbuffer;
6390 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6392 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6394 MUTEX_ENTER(&conn->conn_data_lock);
6395 /* We ran, so the handle is no longer needed to try to cancel ourselves. */
6396 if (event == conn->natKeepAliveEvent)
6397 rxevent_Put(&conn->natKeepAliveEvent);
6398 MUTEX_ENTER(&rx_refcnt_mutex);
6399 /* Only reschedule ourselves if the connection would not be destroyed */
6400 if (conn->refCount > 1)
6402 if (conn->refCount <= 0) {
6403 #ifdef RX_REFCOUNT_CHECK
6404 osi_Assert(conn->refCount == 0);
6406 if (rx_stats_active) {
6407 MUTEX_ENTER(&rx_stats_mutex);
6408 rxi_lowConnRefCount++;
6409 MUTEX_EXIT(&rx_stats_mutex);
6412 MUTEX_EXIT(&rx_refcnt_mutex);
6414 rxi_ScheduleNatKeepAliveEvent(conn);
6415 MUTEX_EXIT(&conn->conn_data_lock);
6416 putConnection(conn);
6420 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6422 MUTEX_ASSERT(&conn->conn_data_lock);
6423 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6424 struct clock when, now;
6425 clock_GetTime(&now);
6427 when.sec += conn->secondsUntilNatPing;
6428 rx_GetConnection(conn);
6429 conn->natKeepAliveEvent =
6430 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6435 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6437 MUTEX_ENTER(&conn->conn_data_lock);
6438 conn->secondsUntilNatPing = seconds;
6440 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6441 rxi_ScheduleNatKeepAliveEvent(conn);
6443 conn->flags |= RX_CONN_NAT_PING;
6445 MUTEX_EXIT(&conn->conn_data_lock);
6448 /* When a call is in progress, this routine is called occasionally to
6449 * make sure that some traffic has arrived (or been sent to) the peer.
6450 * If nothing has arrived in a reasonable amount of time, the call is
6451 * declared dead; if nothing has been sent for a while, we send a
6452 * keep-alive packet (if we're actually trying to keep the call alive)
6455 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6458 struct rx_call *call = arg1;
6459 struct rx_connection *conn;
6462 MUTEX_ENTER(&call->lock);
6464 if (event == call->keepAliveEvent)
6465 rxevent_Put(&call->keepAliveEvent);
6469 if (rxi_CheckCall(call, 0)) {
6470 MUTEX_EXIT(&call->lock);
6474 /* Don't try to keep alive dallying calls */
6475 if (call->state == RX_STATE_DALLY) {
6476 MUTEX_EXIT(&call->lock);
6477 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6482 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6483 /* Don't try to send keepalives if there is unacknowledged data */
6484 /* the rexmit code should be good enough, this little hack
6485 * doesn't quite work XXX */
6486 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6488 rxi_ScheduleKeepAliveEvent(call);
6489 MUTEX_EXIT(&call->lock);
6490 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6493 /* Does what's on the nameplate. */
6495 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6497 struct rx_call *call = arg1;
6498 struct rx_connection *conn;
6500 MUTEX_ENTER(&call->lock);
6502 if (event == call->growMTUEvent)
6503 rxevent_Put(&call->growMTUEvent);
6505 if (rxi_CheckCall(call, 0))
6508 /* Don't bother with dallying calls */
6509 if (call->state == RX_STATE_DALLY)
6515 * keep being scheduled, just don't do anything if we're at peak,
6516 * or we're not set up to be properly handled (idle timeout required)
6518 if ((conn->peer->maxPacketSize != 0) &&
6519 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6521 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6522 rxi_ScheduleGrowMTUEvent(call, 0);
6524 MUTEX_EXIT(&call->lock);
6525 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6529 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6531 MUTEX_ASSERT(&call->lock);
6532 if (!call->keepAliveEvent) {
6533 struct clock when, now;
6534 clock_GetTime(&now);
6536 when.sec += call->conn->secondsUntilPing;
6537 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6538 call->keepAliveEvent =
6539 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6544 rxi_CancelKeepAliveEvent(struct rx_call *call) {
6545 MUTEX_ASSERT(&call->lock);
6546 if (rxevent_Cancel(&call->keepAliveEvent))
6547 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6551 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6553 MUTEX_ASSERT(&call->lock);
6554 if (!call->growMTUEvent) {
6555 struct clock when, now;
6557 clock_GetTime(&now);
6560 if (call->conn->secondsUntilPing)
6561 secs = (6*call->conn->secondsUntilPing)-1;
6563 if (call->conn->secondsUntilDead)
6564 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6568 CALL_HOLD(call, RX_CALL_REFCOUNT_MTU);
6569 call->growMTUEvent =
6570 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6575 rxi_CancelGrowMTUEvent(struct rx_call *call)
6577 MUTEX_ASSERT(&call->lock);
6578 if (rxevent_Cancel(&call->growMTUEvent))
6579 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6583 * Increment the counter for the next connection ID, handling overflow.
6586 update_nextCid(void)
6588 /* Overflow is technically undefined behavior; avoid it. */
6589 if (rx_nextCid > MAX_AFS_INT32 - (1 << RX_CIDSHIFT))
6590 rx_nextCid = -1 * ((MAX_AFS_INT32 / RX_CIDSHIFT) * RX_CIDSHIFT);
6592 rx_nextCid += 1 << RX_CIDSHIFT;
6596 rxi_KeepAliveOn(struct rx_call *call)
6598 /* Pretend last packet received was received now--i.e. if another
6599 * packet isn't received within the keep alive time, then the call
6600 * will die; Initialize last send time to the current time--even
6601 * if a packet hasn't been sent yet. This will guarantee that a
6602 * keep-alive is sent within the ping time */
6603 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6604 rxi_ScheduleKeepAliveEvent(call);
6608 rxi_GrowMTUOn(struct rx_call *call)
6610 struct rx_connection *conn = call->conn;
6611 MUTEX_ENTER(&conn->conn_data_lock);
6612 conn->lastPingSizeSer = conn->lastPingSize = 0;
6613 MUTEX_EXIT(&conn->conn_data_lock);
6614 rxi_ScheduleGrowMTUEvent(call, 1);
6617 /* This routine is called to send connection abort messages
6618 * that have been delayed to throttle looping clients. */
6620 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6623 struct rx_connection *conn = arg1;
6626 struct rx_packet *packet;
6628 MUTEX_ENTER(&conn->conn_data_lock);
6629 if (event == conn->delayedAbortEvent)
6630 rxevent_Put(&conn->delayedAbortEvent);
6631 error = htonl(conn->error);
6633 MUTEX_EXIT(&conn->conn_data_lock);
6634 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6637 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6638 RX_PACKET_TYPE_ABORT, (char *)&error,
6640 rxi_FreePacket(packet);
6642 putConnection(conn);
6645 /* This routine is called to send call abort messages
6646 * that have been delayed to throttle looping clients. */
6648 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6651 struct rx_call *call = arg1;
6654 struct rx_packet *packet;
6656 MUTEX_ENTER(&call->lock);
6657 if (event == call->delayedAbortEvent)
6658 rxevent_Put(&call->delayedAbortEvent);
6659 error = htonl(call->error);
6661 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6664 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6665 (char *)&error, sizeof(error), 0);
6666 rxi_FreePacket(packet);
6668 MUTEX_EXIT(&call->lock);
6669 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6673 * This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6674 * seconds) to ask the client to authenticate itself. The routine
6675 * issues a challenge to the client, which is obtained from the
6676 * security object associated with the connection
6678 * This routine is both an event handler and a function called directly;
6679 * when called directly the passed |event| is NULL and the
6680 * conn->conn->data>lock must must not be held.
6683 rxi_ChallengeEvent(struct rxevent *event,
6684 void *arg0, void *arg1, int tries)
6686 struct rx_connection *conn = arg0;
6688 MUTEX_ENTER(&conn->conn_data_lock);
6689 if (event != NULL && event == conn->challengeEvent)
6690 rxevent_Put(&conn->challengeEvent);
6691 MUTEX_EXIT(&conn->conn_data_lock);
6693 /* If there are no active calls it is not worth re-issuing the
6694 * challenge. If the client issues another call on this connection
6695 * the challenge can be requested at that time.
6697 if (!rxi_HasActiveCalls(conn)) {
6698 putConnection(conn);
6702 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6703 struct rx_packet *packet;
6704 struct clock when, now;
6707 /* We've failed to authenticate for too long.
6708 * Reset any calls waiting for authentication;
6709 * they are all in RX_STATE_PRECALL.
6713 MUTEX_ENTER(&conn->conn_call_lock);
6714 for (i = 0; i < RX_MAXCALLS; i++) {
6715 struct rx_call *call = conn->call[i];
6717 MUTEX_ENTER(&call->lock);
6718 if (call->state == RX_STATE_PRECALL) {
6719 rxi_CallError(call, RX_CALL_DEAD);
6720 rxi_SendCallAbort(call, NULL, 0, 0);
6722 MUTEX_EXIT(&call->lock);
6725 MUTEX_EXIT(&conn->conn_call_lock);
6726 putConnection(conn);
6730 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6732 /* If there's no packet available, do this later. */
6733 RXS_GetChallenge(conn->securityObject, conn, packet);
6734 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6735 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6736 rxi_FreePacket(packet);
6737 conn->securityChallengeSent = 1;
6739 clock_GetTime(&now);
6741 when.sec += RX_CHALLENGE_TIMEOUT;
6742 MUTEX_ENTER(&conn->conn_data_lock);
6743 /* Only reschedule ourselves if not already pending. */
6744 if (conn->challengeEvent == NULL) {
6745 rx_GetConnection(conn);
6746 conn->challengeEvent =
6747 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6750 MUTEX_EXIT(&conn->conn_data_lock);
6752 putConnection(conn);
6755 /* Call this routine to start requesting the client to authenticate
6756 * itself. This will continue until authentication is established,
6757 * the call times out, or an invalid response is returned. The
6758 * security object associated with the connection is asked to create
6759 * the challenge at this time. */
6761 rxi_ChallengeOn(struct rx_connection *conn)
6764 MUTEX_ENTER(&conn->conn_data_lock);
6765 if (!conn->challengeEvent)
6767 MUTEX_EXIT(&conn->conn_data_lock);
6769 RXS_CreateChallenge(conn->securityObject, conn);
6770 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6775 /* rxi_ComputeRoundTripTime is called with peer locked. */
6776 /* peer may be null */
6778 rxi_ComputeRoundTripTime(struct rx_packet *p,
6779 struct rx_ackPacket *ack,
6780 struct rx_call *call,
6781 struct rx_peer *peer,
6784 struct clock thisRtt, *sentp;
6788 /* If the ACK is delayed, then do nothing */
6789 if (ack->reason == RX_ACK_DELAY)
6792 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6793 * their RTT multiple times, so only include the RTT of the last packet
6795 if (p->flags & RX_JUMBO_PACKET)
6798 /* Use the serial number to determine which transmission the ACK is for,
6799 * and set the sent time to match this. If we have no serial number, then
6800 * only use the ACK for RTT calculations if the packet has not been
6804 serial = ntohl(ack->serial);
6806 if (serial == p->header.serial) {
6807 sentp = &p->timeSent;
6808 } else if (serial == p->firstSerial) {
6809 sentp = &p->firstSent;
6810 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6811 sentp = &p->firstSent;
6815 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6816 sentp = &p->firstSent;
6823 if (clock_Lt(&thisRtt, sentp))
6824 return; /* somebody set the clock back, don't count this time. */
6826 clock_Sub(&thisRtt, sentp);
6827 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6828 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6830 if (clock_IsZero(&thisRtt)) {
6832 * The actual round trip time is shorter than the
6833 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6834 * Since we can't tell which at the moment we will assume 1ms.
6836 thisRtt.usec = 1000;
6839 if (rx_stats_active) {
6840 MUTEX_ENTER(&rx_stats_mutex);
6841 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6842 rx_stats.minRtt = thisRtt;
6843 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6844 if (thisRtt.sec > 60) {
6845 MUTEX_EXIT(&rx_stats_mutex);
6846 return; /* somebody set the clock ahead */
6848 rx_stats.maxRtt = thisRtt;
6850 clock_Add(&rx_stats.totalRtt, &thisRtt);
6851 rx_atomic_inc(&rx_stats.nRttSamples);
6852 MUTEX_EXIT(&rx_stats_mutex);
6855 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6857 /* Apply VanJacobson round-trip estimations */
6862 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6863 * srtt is stored as fixed point with 3 bits after the binary
6864 * point (i.e., scaled by 8). The following magic is
6865 * equivalent to the smoothing algorithm in rfc793 with an
6866 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6867 * srtt'*8 = rtt + srtt*7
6868 * srtt'*8 = srtt*8 + rtt - srtt
6869 * srtt' = srtt + rtt/8 - srtt/8
6870 * srtt' = srtt + (rtt - srtt)/8
6873 delta = _8THMSEC(&thisRtt) - call->rtt;
6874 call->rtt += (delta >> 3);
6877 * We accumulate a smoothed rtt variance (actually, a smoothed
6878 * mean difference), then set the retransmit timer to smoothed
6879 * rtt + 4 times the smoothed variance (was 2x in van's original
6880 * paper, but 4x works better for me, and apparently for him as
6882 * rttvar is stored as
6883 * fixed point with 2 bits after the binary point (scaled by
6884 * 4). The following is equivalent to rfc793 smoothing with
6885 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6886 * rttvar'*4 = rttvar*3 + |delta|
6887 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6888 * rttvar' = rttvar + |delta|/4 - rttvar/4
6889 * rttvar' = rttvar + (|delta| - rttvar)/4
6890 * This replaces rfc793's wired-in beta.
6891 * dev*4 = dev*4 + (|actual - expected| - dev)
6897 delta -= (call->rtt_dev << 1);
6898 call->rtt_dev += (delta >> 3);
6900 /* I don't have a stored RTT so I start with this value. Since I'm
6901 * probably just starting a call, and will be pushing more data down
6902 * this, I expect congestion to increase rapidly. So I fudge a
6903 * little, and I set deviance to half the rtt. In practice,
6904 * deviance tends to approach something a little less than
6905 * half the smoothed rtt. */
6906 call->rtt = _8THMSEC(&thisRtt) + 8;
6907 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6909 /* the smoothed RTT time is RTT + 4*MDEV
6911 * We allow a user specified minimum to be set for this, to allow clamping
6912 * at a minimum value in the same way as TCP. In addition, we have to allow
6913 * for the possibility that this packet is answered by a delayed ACK, so we
6914 * add on a fixed 200ms to account for that timer expiring.
6917 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6918 rx_minPeerTimeout) + 200;
6919 clock_Zero(&call->rto);
6920 clock_Addmsec(&call->rto, rtt_timeout);
6922 /* Update the peer, so any new calls start with our values */
6923 peer->rtt_dev = call->rtt_dev;
6924 peer->rtt = call->rtt;
6926 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6927 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6931 /* Find all server connections that have not been active for a long time, and
6934 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
6937 struct clock now, when;
6938 struct rxevent *event;
6939 clock_GetTime(&now);
6941 /* Find server connection structures that haven't been used for
6942 * greater than rx_idleConnectionTime */
6944 struct rx_connection **conn_ptr, **conn_end;
6945 int i, havecalls = 0;
6946 MUTEX_ENTER(&rx_connHashTable_lock);
6947 for (conn_ptr = &rx_connHashTable[0], conn_end =
6948 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6950 struct rx_connection *conn, *next;
6951 struct rx_call *call;
6955 for (conn = *conn_ptr; conn; conn = next) {
6956 /* XXX -- Shouldn't the connection be locked? */
6959 for (i = 0; i < RX_MAXCALLS; i++) {
6960 call = conn->call[i];
6964 code = MUTEX_TRYENTER(&call->lock);
6967 result = rxi_CheckCall(call, 1);
6968 MUTEX_EXIT(&call->lock);
6970 /* If CheckCall freed the call, it might
6971 * have destroyed the connection as well,
6972 * which screws up the linked lists.
6978 if (conn->type == RX_SERVER_CONNECTION) {
6979 /* This only actually destroys the connection if
6980 * there are no outstanding calls */
6981 MUTEX_ENTER(&conn->conn_data_lock);
6982 MUTEX_ENTER(&rx_refcnt_mutex);
6983 if (!havecalls && !conn->refCount
6984 && ((conn->lastSendTime + rx_idleConnectionTime) <
6986 conn->refCount++; /* it will be decr in rx_DestroyConn */
6987 MUTEX_EXIT(&rx_refcnt_mutex);
6988 MUTEX_EXIT(&conn->conn_data_lock);
6989 #ifdef RX_ENABLE_LOCKS
6990 rxi_DestroyConnectionNoLock(conn);
6991 #else /* RX_ENABLE_LOCKS */
6992 rxi_DestroyConnection(conn);
6993 #endif /* RX_ENABLE_LOCKS */
6995 #ifdef RX_ENABLE_LOCKS
6997 MUTEX_EXIT(&rx_refcnt_mutex);
6998 MUTEX_EXIT(&conn->conn_data_lock);
7000 #endif /* RX_ENABLE_LOCKS */
7004 #ifdef RX_ENABLE_LOCKS
7005 while (rx_connCleanup_list) {
7006 struct rx_connection *conn;
7007 conn = rx_connCleanup_list;
7008 rx_connCleanup_list = rx_connCleanup_list->next;
7009 MUTEX_EXIT(&rx_connHashTable_lock);
7010 rxi_CleanupConnection(conn);
7011 MUTEX_ENTER(&rx_connHashTable_lock);
7013 MUTEX_EXIT(&rx_connHashTable_lock);
7014 #endif /* RX_ENABLE_LOCKS */
7017 /* Find any peer structures that haven't been used (haven't had an
7018 * associated connection) for greater than rx_idlePeerTime */
7020 struct rx_peer **peer_ptr, **peer_end;
7024 * Why do we need to hold the rx_peerHashTable_lock across
7025 * the incrementing of peer_ptr since the rx_peerHashTable
7026 * array is not changing? We don't.
7028 * By dropping the lock periodically we can permit other
7029 * activities to be performed while a rxi_ReapConnections
7030 * call is in progress. The goal of reap connections
7031 * is to clean up quickly without causing large amounts
7032 * of contention. Therefore, it is important that global
7033 * mutexes not be held for extended periods of time.
7035 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7036 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7038 struct rx_peer *peer, *next, *prev;
7040 MUTEX_ENTER(&rx_peerHashTable_lock);
7041 for (prev = peer = *peer_ptr; peer; peer = next) {
7043 code = MUTEX_TRYENTER(&peer->peer_lock);
7044 if ((code) && (peer->refCount == 0)
7045 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7046 struct opr_queue *cursor, *store;
7050 * now know that this peer object is one to be
7051 * removed from the hash table. Once it is removed
7052 * it can't be referenced by other threads.
7053 * Lets remove it first and decrement the struct
7054 * nPeerStructs count.
7056 if (peer == *peer_ptr) {
7062 if (rx_stats_active)
7063 rx_atomic_dec(&rx_stats.nPeerStructs);
7066 * Now if we hold references on 'prev' and 'next'
7067 * we can safely drop the rx_peerHashTable_lock
7068 * while we destroy this 'peer' object.
7074 MUTEX_EXIT(&rx_peerHashTable_lock);
7076 MUTEX_EXIT(&peer->peer_lock);
7077 MUTEX_DESTROY(&peer->peer_lock);
7079 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7080 unsigned int num_funcs;
7081 struct rx_interface_stat *rpc_stat
7082 = opr_queue_Entry(cursor, struct rx_interface_stat,
7087 opr_queue_Remove(&rpc_stat->entry);
7088 opr_queue_Remove(&rpc_stat->entryPeers);
7090 num_funcs = rpc_stat->stats[0].func_total;
7092 sizeof(rx_interface_stat_t) +
7093 rpc_stat->stats[0].func_total *
7094 sizeof(rx_function_entry_v1_t);
7096 rxi_Free(rpc_stat, space);
7098 MUTEX_ENTER(&rx_rpc_stats);
7099 rxi_rpc_peer_stat_cnt -= num_funcs;
7100 MUTEX_EXIT(&rx_rpc_stats);
7105 * Regain the rx_peerHashTable_lock and
7106 * decrement the reference count on 'prev'
7109 MUTEX_ENTER(&rx_peerHashTable_lock);
7116 MUTEX_EXIT(&peer->peer_lock);
7121 MUTEX_EXIT(&rx_peerHashTable_lock);
7125 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7126 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7127 * GC, just below. Really, we shouldn't have to keep moving packets from
7128 * one place to another, but instead ought to always know if we can
7129 * afford to hold onto a packet in its particular use. */
7130 MUTEX_ENTER(&rx_freePktQ_lock);
7131 if (rx_waitingForPackets) {
7132 rx_waitingForPackets = 0;
7133 #ifdef RX_ENABLE_LOCKS
7134 CV_BROADCAST(&rx_waitingForPackets_cv);
7136 osi_rxWakeup(&rx_waitingForPackets);
7139 MUTEX_EXIT(&rx_freePktQ_lock);
7142 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7143 event = rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0);
7144 rxevent_Put(&event);
7148 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7149 * rx.h is sort of strange this is better. This is called with a security
7150 * object before it is discarded. Each connection using a security object has
7151 * its own refcount to the object so it won't actually be freed until the last
7152 * connection is destroyed.
7154 * This is the only rxs module call. A hold could also be written but no one
7158 rxs_Release(struct rx_securityClass *aobj)
7160 return RXS_Close(aobj);
7168 #define TRACE_OPTION_RX_DEBUG 16
7176 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7177 0, KEY_QUERY_VALUE, &parmKey);
7178 if (code != ERROR_SUCCESS)
7181 dummyLen = sizeof(TraceOption);
7182 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7183 (BYTE *) &TraceOption, &dummyLen);
7184 if (code == ERROR_SUCCESS) {
7185 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7187 RegCloseKey (parmKey);
7188 #endif /* AFS_NT40_ENV */
7193 rx_DebugOnOff(int on)
7197 rxdebug_active = on;
7203 rx_StatsOnOff(int on)
7205 rx_stats_active = on;
7209 /* Don't call this debugging routine directly; use dpf */
7211 rxi_DebugPrint(char *format, ...)
7220 va_start(ap, format);
7222 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7225 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7227 OutputDebugString(msg);
7233 va_start(ap, format);
7235 clock_GetTime(&now);
7236 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7237 (unsigned int)now.usec);
7238 vfprintf(rx_Log, format, ap);
7246 * This function is used to process the rx_stats structure that is local
7247 * to a process as well as an rx_stats structure received from a remote
7248 * process (via rxdebug). Therefore, it needs to do minimal version
7252 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7253 afs_int32 freePackets, char version)
7257 if (size != sizeof(struct rx_statistics)) {
7259 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7260 size, sizeof(struct rx_statistics));
7263 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7266 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7267 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7268 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7269 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7270 s->specialPktAllocFailures);
7272 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7273 s->receivePktAllocFailures, s->sendPktAllocFailures,
7274 s->specialPktAllocFailures);
7278 " greedy %u, " "bogusReads %u (last from host %x), "
7279 "noPackets %u, " "noBuffers %u, " "selects %u, "
7280 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7281 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7282 s->selects, s->sendSelects);
7284 fprintf(file, " packets read: ");
7285 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7286 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7288 fprintf(file, "\n");
7291 " other read counters: data %u, " "ack %u, " "dup %u "
7292 "spurious %u " "dally %u\n", s->dataPacketsRead,
7293 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7294 s->ignorePacketDally);
7296 fprintf(file, " packets sent: ");
7297 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7298 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7300 fprintf(file, "\n");
7303 " other send counters: ack %u, " "data %u (not resends), "
7304 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7305 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7306 s->dataPacketsPushed, s->ignoreAckedPacket);
7309 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7310 s->netSendFailures, (int)s->fatalErrors);
7312 if (s->nRttSamples) {
7313 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7314 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7316 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7317 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7321 " %d server connections, " "%d client connections, "
7322 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7323 s->nServerConns, s->nClientConns, s->nPeerStructs,
7324 s->nCallStructs, s->nFreeCallStructs);
7326 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7327 fprintf(file, " %d clock updates\n", clock_nUpdates);
7331 /* for backward compatibility */
7333 rx_PrintStats(FILE * file)
7335 MUTEX_ENTER(&rx_stats_mutex);
7336 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7337 sizeof(rx_stats), rx_nFreePackets,
7339 MUTEX_EXIT(&rx_stats_mutex);
7343 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7345 fprintf(file, "Peer %x.%d.\n",
7346 ntohl(peer->host), (int)ntohs(peer->port));
7349 " Rtt %d, " "total sent %d, " "resent %d\n",
7350 peer->rtt, peer->nSent, peer->reSends);
7352 fprintf(file, " Packet size %d\n", peer->ifMTU);
7356 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7358 * This mutex protects the following static variables:
7362 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7363 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7365 #define LOCK_RX_DEBUG
7366 #define UNLOCK_RX_DEBUG
7367 #endif /* AFS_PTHREAD_ENV */
7369 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7371 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7372 u_char type, void *inputData, size_t inputLength,
7373 void *outputData, size_t outputLength)
7375 static afs_int32 counter = 100;
7376 time_t waitTime, waitCount;
7377 struct rx_header theader;
7380 struct timeval tv_now, tv_wake, tv_delta;
7381 struct sockaddr_in taddr, faddr;
7395 tp = &tbuffer[sizeof(struct rx_header)];
7396 taddr.sin_family = AF_INET;
7397 taddr.sin_port = remotePort;
7398 taddr.sin_addr.s_addr = remoteAddr;
7399 memset(&taddr.sin_zero, 0, sizeof(taddr.sin_zero));
7400 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7401 taddr.sin_len = sizeof(struct sockaddr_in);
7404 memset(&theader, 0, sizeof(theader));
7405 theader.epoch = htonl(999);
7407 theader.callNumber = htonl(counter);
7410 theader.type = type;
7411 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7412 theader.serviceId = 0;
7414 memcpy(tbuffer, &theader, sizeof(theader));
7415 memcpy(tp, inputData, inputLength);
7417 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7418 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7420 /* see if there's a packet available */
7421 gettimeofday(&tv_wake, NULL);
7422 tv_wake.tv_sec += waitTime;
7425 FD_SET(socket, &imask);
7426 tv_delta.tv_sec = tv_wake.tv_sec;
7427 tv_delta.tv_usec = tv_wake.tv_usec;
7428 gettimeofday(&tv_now, NULL);
7430 if (tv_delta.tv_usec < tv_now.tv_usec) {
7432 tv_delta.tv_usec += 1000000;
7435 tv_delta.tv_usec -= tv_now.tv_usec;
7437 if (tv_delta.tv_sec < tv_now.tv_sec) {
7441 tv_delta.tv_sec -= tv_now.tv_sec;
7444 code = select(0, &imask, 0, 0, &tv_delta);
7445 #else /* AFS_NT40_ENV */
7446 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7447 #endif /* AFS_NT40_ENV */
7448 if (code == 1 && FD_ISSET(socket, &imask)) {
7449 /* now receive a packet */
7450 faddrLen = sizeof(struct sockaddr_in);
7452 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7453 (struct sockaddr *)&faddr, &faddrLen);
7456 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7457 if (counter == ntohl(theader.callNumber))
7465 /* see if we've timed out */
7473 code -= sizeof(struct rx_header);
7474 if (code > outputLength)
7475 code = outputLength;
7476 memcpy(outputData, tp, code);
7479 #endif /* RXDEBUG */
7482 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7483 afs_uint16 remotePort, struct rx_debugStats * stat,
7484 afs_uint32 * supportedValues)
7486 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7488 struct rx_debugIn in;
7490 *supportedValues = 0;
7491 in.type = htonl(RX_DEBUGI_GETSTATS);
7494 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7495 &in, sizeof(in), stat, sizeof(*stat));
7498 * If the call was successful, fixup the version and indicate
7499 * what contents of the stat structure are valid.
7500 * Also do net to host conversion of fields here.
7504 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7505 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7507 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7508 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7510 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7511 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7513 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7514 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7516 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7517 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7519 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7520 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7522 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7523 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7525 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7526 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7528 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7529 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7531 stat->nFreePackets = ntohl(stat->nFreePackets);
7532 stat->packetReclaims = ntohl(stat->packetReclaims);
7533 stat->callsExecuted = ntohl(stat->callsExecuted);
7534 stat->nWaiting = ntohl(stat->nWaiting);
7535 stat->idleThreads = ntohl(stat->idleThreads);
7536 stat->nWaited = ntohl(stat->nWaited);
7537 stat->nPackets = ntohl(stat->nPackets);
7546 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7547 afs_uint16 remotePort, struct rx_statistics * stat,
7548 afs_uint32 * supportedValues)
7550 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7552 struct rx_debugIn in;
7553 afs_int32 *lp = (afs_int32 *) stat;
7557 * supportedValues is currently unused, but added to allow future
7558 * versioning of this function.
7561 *supportedValues = 0;
7562 in.type = htonl(RX_DEBUGI_RXSTATS);
7564 memset(stat, 0, sizeof(*stat));
7566 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7567 &in, sizeof(in), stat, sizeof(*stat));
7572 * Do net to host conversion here
7575 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7586 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7587 afs_uint16 remotePort, size_t version_length,
7590 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7592 return MakeDebugCall(socket, remoteAddr, remotePort,
7593 RX_PACKET_TYPE_VERSION, a, 1, version,
7601 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7602 afs_uint16 remotePort, afs_int32 * nextConnection,
7603 int allConnections, afs_uint32 debugSupportedValues,
7604 struct rx_debugConn * conn,
7605 afs_uint32 * supportedValues)
7607 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7609 struct rx_debugIn in;
7613 * supportedValues is currently unused, but added to allow future
7614 * versioning of this function.
7617 *supportedValues = 0;
7618 if (allConnections) {
7619 in.type = htonl(RX_DEBUGI_GETALLCONN);
7621 in.type = htonl(RX_DEBUGI_GETCONN);
7623 in.index = htonl(*nextConnection);
7624 memset(conn, 0, sizeof(*conn));
7626 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7627 &in, sizeof(in), conn, sizeof(*conn));
7630 *nextConnection += 1;
7633 * Convert old connection format to new structure.
7636 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7637 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7638 #define MOVEvL(a) (conn->a = vL->a)
7640 /* any old or unrecognized version... */
7641 for (i = 0; i < RX_MAXCALLS; i++) {
7642 MOVEvL(callState[i]);
7643 MOVEvL(callMode[i]);
7644 MOVEvL(callFlags[i]);
7645 MOVEvL(callOther[i]);
7647 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7648 MOVEvL(secStats.type);
7649 MOVEvL(secStats.level);
7650 MOVEvL(secStats.flags);
7651 MOVEvL(secStats.expires);
7652 MOVEvL(secStats.packetsReceived);
7653 MOVEvL(secStats.packetsSent);
7654 MOVEvL(secStats.bytesReceived);
7655 MOVEvL(secStats.bytesSent);
7660 * Do net to host conversion here
7662 * I don't convert host or port since we are most likely
7663 * going to want these in NBO.
7665 conn->cid = ntohl(conn->cid);
7666 conn->serial = ntohl(conn->serial);
7667 for (i = 0; i < RX_MAXCALLS; i++) {
7668 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7670 conn->error = ntohl(conn->error);
7671 conn->secStats.flags = ntohl(conn->secStats.flags);
7672 conn->secStats.expires = ntohl(conn->secStats.expires);
7673 conn->secStats.packetsReceived =
7674 ntohl(conn->secStats.packetsReceived);
7675 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7676 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7677 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7678 conn->epoch = ntohl(conn->epoch);
7679 conn->natMTU = ntohl(conn->natMTU);
7688 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7689 afs_uint16 remotePort, afs_int32 * nextPeer,
7690 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7691 afs_uint32 * supportedValues)
7693 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7695 struct rx_debugIn in;
7698 * supportedValues is currently unused, but added to allow future
7699 * versioning of this function.
7702 *supportedValues = 0;
7703 in.type = htonl(RX_DEBUGI_GETPEER);
7704 in.index = htonl(*nextPeer);
7705 memset(peer, 0, sizeof(*peer));
7707 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7708 &in, sizeof(in), peer, sizeof(*peer));
7714 * Do net to host conversion here
7716 * I don't convert host or port since we are most likely
7717 * going to want these in NBO.
7719 peer->ifMTU = ntohs(peer->ifMTU);
7720 peer->idleWhen = ntohl(peer->idleWhen);
7721 peer->refCount = ntohs(peer->refCount);
7722 peer->rtt = ntohl(peer->rtt);
7723 peer->rtt_dev = ntohl(peer->rtt_dev);
7724 peer->timeout.sec = 0;
7725 peer->timeout.usec = 0;
7726 peer->nSent = ntohl(peer->nSent);
7727 peer->reSends = ntohl(peer->reSends);
7728 peer->natMTU = ntohs(peer->natMTU);
7729 peer->maxMTU = ntohs(peer->maxMTU);
7730 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7731 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7732 peer->MTU = ntohs(peer->MTU);
7733 peer->cwind = ntohs(peer->cwind);
7734 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7735 peer->congestSeq = ntohs(peer->congestSeq);
7736 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7737 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7738 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7739 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7748 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7749 struct rx_debugPeer * peerStats)
7752 afs_int32 error = 1; /* default to "did not succeed" */
7753 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7755 MUTEX_ENTER(&rx_peerHashTable_lock);
7756 for(tp = rx_peerHashTable[hashValue];
7757 tp != NULL; tp = tp->next) {
7758 if (tp->host == peerHost)
7764 MUTEX_EXIT(&rx_peerHashTable_lock);
7768 MUTEX_ENTER(&tp->peer_lock);
7769 peerStats->host = tp->host;
7770 peerStats->port = tp->port;
7771 peerStats->ifMTU = tp->ifMTU;
7772 peerStats->idleWhen = tp->idleWhen;
7773 peerStats->refCount = tp->refCount;
7774 peerStats->burstSize = 0;
7775 peerStats->burst = 0;
7776 peerStats->burstWait.sec = 0;
7777 peerStats->burstWait.usec = 0;
7778 peerStats->rtt = tp->rtt;
7779 peerStats->rtt_dev = tp->rtt_dev;
7780 peerStats->timeout.sec = 0;
7781 peerStats->timeout.usec = 0;
7782 peerStats->nSent = tp->nSent;
7783 peerStats->reSends = tp->reSends;
7784 peerStats->natMTU = tp->natMTU;
7785 peerStats->maxMTU = tp->maxMTU;
7786 peerStats->maxDgramPackets = tp->maxDgramPackets;
7787 peerStats->ifDgramPackets = tp->ifDgramPackets;
7788 peerStats->MTU = tp->MTU;
7789 peerStats->cwind = tp->cwind;
7790 peerStats->nDgramPackets = tp->nDgramPackets;
7791 peerStats->congestSeq = tp->congestSeq;
7792 peerStats->bytesSent.high = tp->bytesSent >> 32;
7793 peerStats->bytesSent.low = tp->bytesSent & MAX_AFS_UINT32;
7794 peerStats->bytesReceived.high = tp->bytesReceived >> 32;
7795 peerStats->bytesReceived.low
7796 = tp->bytesReceived & MAX_AFS_UINT32;
7797 MUTEX_EXIT(&tp->peer_lock);
7799 MUTEX_ENTER(&rx_peerHashTable_lock);
7802 MUTEX_EXIT(&rx_peerHashTable_lock);
7810 struct rx_serverQueueEntry *np;
7813 struct rx_call *call;
7814 struct rx_serverQueueEntry *sq;
7817 if (rx_atomic_test_and_set_bit(&rxinit_status, 0))
7818 return; /* Already shutdown. */
7822 #ifndef AFS_PTHREAD_ENV
7823 FD_ZERO(&rx_selectMask);
7824 #endif /* AFS_PTHREAD_ENV */
7825 rxi_dataQuota = RX_MAX_QUOTA;
7826 #ifndef AFS_PTHREAD_ENV
7828 #endif /* AFS_PTHREAD_ENV */
7831 #ifndef AFS_PTHREAD_ENV
7832 #ifndef AFS_USE_GETTIMEOFDAY
7834 #endif /* AFS_USE_GETTIMEOFDAY */
7835 #endif /* AFS_PTHREAD_ENV */
7837 while (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
7838 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
7839 opr_queue_Remove(&call->entry);
7840 rxi_Free(call, sizeof(struct rx_call));
7843 while (!opr_queue_IsEmpty(&rx_idleServerQueue)) {
7844 sq = opr_queue_First(&rx_idleServerQueue, struct rx_serverQueueEntry,
7846 opr_queue_Remove(&sq->entry);
7851 struct rx_peer **peer_ptr, **peer_end;
7852 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7853 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7855 struct rx_peer *peer, *next;
7857 MUTEX_ENTER(&rx_peerHashTable_lock);
7858 for (peer = *peer_ptr; peer; peer = next) {
7859 struct opr_queue *cursor, *store;
7862 MUTEX_ENTER(&rx_rpc_stats);
7863 MUTEX_ENTER(&peer->peer_lock);
7864 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7865 unsigned int num_funcs;
7866 struct rx_interface_stat *rpc_stat
7867 = opr_queue_Entry(cursor, struct rx_interface_stat,
7871 opr_queue_Remove(&rpc_stat->entry);
7872 opr_queue_Remove(&rpc_stat->entryPeers);
7873 num_funcs = rpc_stat->stats[0].func_total;
7875 sizeof(rx_interface_stat_t) +
7876 rpc_stat->stats[0].func_total *
7877 sizeof(rx_function_entry_v1_t);
7879 rxi_Free(rpc_stat, space);
7881 /* rx_rpc_stats must be held */
7882 rxi_rpc_peer_stat_cnt -= num_funcs;
7884 MUTEX_EXIT(&peer->peer_lock);
7885 MUTEX_EXIT(&rx_rpc_stats);
7889 if (rx_stats_active)
7890 rx_atomic_dec(&rx_stats.nPeerStructs);
7892 MUTEX_EXIT(&rx_peerHashTable_lock);
7895 for (i = 0; i < RX_MAX_SERVICES; i++) {
7897 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7899 for (i = 0; i < rx_hashTableSize; i++) {
7900 struct rx_connection *tc, *ntc;
7901 MUTEX_ENTER(&rx_connHashTable_lock);
7902 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7904 for (j = 0; j < RX_MAXCALLS; j++) {
7906 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7909 rxi_Free(tc, sizeof(*tc));
7911 MUTEX_EXIT(&rx_connHashTable_lock);
7914 MUTEX_ENTER(&freeSQEList_lock);
7916 while ((np = rx_FreeSQEList)) {
7917 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7918 MUTEX_DESTROY(&np->lock);
7919 rxi_Free(np, sizeof(*np));
7922 MUTEX_EXIT(&freeSQEList_lock);
7923 MUTEX_DESTROY(&freeSQEList_lock);
7924 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7925 MUTEX_DESTROY(&rx_connHashTable_lock);
7926 MUTEX_DESTROY(&rx_peerHashTable_lock);
7927 MUTEX_DESTROY(&rx_serverPool_lock);
7929 osi_Free(rx_connHashTable,
7930 rx_hashTableSize * sizeof(struct rx_connection *));
7931 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7933 UNPIN(rx_connHashTable,
7934 rx_hashTableSize * sizeof(struct rx_connection *));
7935 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7937 MUTEX_ENTER(&rx_quota_mutex);
7938 rxi_dataQuota = RX_MAX_QUOTA;
7939 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7940 MUTEX_EXIT(&rx_quota_mutex);
7946 * Routines to implement connection specific data.
7950 rx_KeyCreate(rx_destructor_t rtn)
7953 MUTEX_ENTER(&rxi_keyCreate_lock);
7954 key = rxi_keyCreate_counter++;
7955 rxi_keyCreate_destructor = (rx_destructor_t *)
7956 realloc((void *)rxi_keyCreate_destructor,
7957 (key + 1) * sizeof(rx_destructor_t));
7958 rxi_keyCreate_destructor[key] = rtn;
7959 MUTEX_EXIT(&rxi_keyCreate_lock);
7964 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7967 MUTEX_ENTER(&conn->conn_data_lock);
7968 if (!conn->specific) {
7969 conn->specific = malloc((key + 1) * sizeof(void *));
7970 for (i = 0; i < key; i++)
7971 conn->specific[i] = NULL;
7972 conn->nSpecific = key + 1;
7973 conn->specific[key] = ptr;
7974 } else if (key >= conn->nSpecific) {
7975 conn->specific = (void **)
7976 realloc(conn->specific, (key + 1) * sizeof(void *));
7977 for (i = conn->nSpecific; i < key; i++)
7978 conn->specific[i] = NULL;
7979 conn->nSpecific = key + 1;
7980 conn->specific[key] = ptr;
7982 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7983 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7984 conn->specific[key] = ptr;
7986 MUTEX_EXIT(&conn->conn_data_lock);
7990 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7993 MUTEX_ENTER(&svc->svc_data_lock);
7994 if (!svc->specific) {
7995 svc->specific = malloc((key + 1) * sizeof(void *));
7996 for (i = 0; i < key; i++)
7997 svc->specific[i] = NULL;
7998 svc->nSpecific = key + 1;
7999 svc->specific[key] = ptr;
8000 } else if (key >= svc->nSpecific) {
8001 svc->specific = (void **)
8002 realloc(svc->specific, (key + 1) * sizeof(void *));
8003 for (i = svc->nSpecific; i < key; i++)
8004 svc->specific[i] = NULL;
8005 svc->nSpecific = key + 1;
8006 svc->specific[key] = ptr;
8008 if (svc->specific[key] && rxi_keyCreate_destructor[key])
8009 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
8010 svc->specific[key] = ptr;
8012 MUTEX_EXIT(&svc->svc_data_lock);
8016 rx_GetSpecific(struct rx_connection *conn, int key)
8019 MUTEX_ENTER(&conn->conn_data_lock);
8020 if (key >= conn->nSpecific)
8023 ptr = conn->specific[key];
8024 MUTEX_EXIT(&conn->conn_data_lock);
8029 rx_GetServiceSpecific(struct rx_service *svc, int key)
8032 MUTEX_ENTER(&svc->svc_data_lock);
8033 if (key >= svc->nSpecific)
8036 ptr = svc->specific[key];
8037 MUTEX_EXIT(&svc->svc_data_lock);
8042 #endif /* !KERNEL */
8045 * processStats is a queue used to store the statistics for the local
8046 * process. Its contents are similar to the contents of the rpcStats
8047 * queue on a rx_peer structure, but the actual data stored within
8048 * this queue contains totals across the lifetime of the process (assuming
8049 * the stats have not been reset) - unlike the per peer structures
8050 * which can come and go based upon the peer lifetime.
8053 static struct opr_queue processStats = { &processStats, &processStats };
8056 * peerStats is a queue used to store the statistics for all peer structs.
8057 * Its contents are the union of all the peer rpcStats queues.
8060 static struct opr_queue peerStats = { &peerStats, &peerStats };
8063 * rxi_monitor_processStats is used to turn process wide stat collection
8067 static int rxi_monitor_processStats = 0;
8070 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8073 static int rxi_monitor_peerStats = 0;
8077 rxi_ClearRPCOpStat(rx_function_entry_v1_p rpc_stat)
8079 rpc_stat->invocations = 0;
8080 rpc_stat->bytes_sent = 0;
8081 rpc_stat->bytes_rcvd = 0;
8082 rpc_stat->queue_time_sum.sec = 0;
8083 rpc_stat->queue_time_sum.usec = 0;
8084 rpc_stat->queue_time_sum_sqr.sec = 0;
8085 rpc_stat->queue_time_sum_sqr.usec = 0;
8086 rpc_stat->queue_time_min.sec = 9999999;
8087 rpc_stat->queue_time_min.usec = 9999999;
8088 rpc_stat->queue_time_max.sec = 0;
8089 rpc_stat->queue_time_max.usec = 0;
8090 rpc_stat->execution_time_sum.sec = 0;
8091 rpc_stat->execution_time_sum.usec = 0;
8092 rpc_stat->execution_time_sum_sqr.sec = 0;
8093 rpc_stat->execution_time_sum_sqr.usec = 0;
8094 rpc_stat->execution_time_min.sec = 9999999;
8095 rpc_stat->execution_time_min.usec = 9999999;
8096 rpc_stat->execution_time_max.sec = 0;
8097 rpc_stat->execution_time_max.usec = 0;
8101 * Given all of the information for a particular rpc
8102 * call, find or create (if requested) the stat structure for the rpc.
8105 * the queue of stats that will be updated with the new value
8107 * @param rxInterface
8108 * a unique number that identifies the rpc interface
8111 * the total number of functions in this interface. this is only
8112 * required if create is true
8115 * if true, this invocation was made to a server
8118 * the ip address of the remote host. this is only required if create
8119 * and addToPeerList are true
8122 * the port of the remote host. this is only required if create
8123 * and addToPeerList are true
8125 * @param addToPeerList
8126 * if != 0, add newly created stat to the global peer list
8129 * if a new stats structure is allocated, the counter will
8130 * be updated with the new number of allocated stat structures.
8131 * only required if create is true
8134 * if no stats structure exists, allocate one
8138 static rx_interface_stat_p
8139 rxi_FindRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8140 afs_uint32 totalFunc, int isServer, afs_uint32 remoteHost,
8141 afs_uint32 remotePort, int addToPeerList,
8142 unsigned int *counter, int create)
8144 rx_interface_stat_p rpc_stat = NULL;
8145 struct opr_queue *cursor;
8148 * See if there's already a structure for this interface
8151 for (opr_queue_Scan(stats, cursor)) {
8152 rpc_stat = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8154 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8155 && (rpc_stat->stats[0].remote_is_server == isServer))
8159 /* if they didn't ask us to create, we're done */
8161 if (opr_queue_IsEnd(stats, cursor))
8167 /* can't proceed without these */
8168 if (!totalFunc || !counter)
8172 * Didn't find a match so allocate a new structure and add it to the
8176 if (opr_queue_IsEnd(stats, cursor) || (rpc_stat == NULL)
8177 || (rpc_stat->stats[0].interfaceId != rxInterface)
8178 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8183 sizeof(rx_interface_stat_t) +
8184 totalFunc * sizeof(rx_function_entry_v1_t);
8186 rpc_stat = rxi_Alloc(space);
8187 if (rpc_stat == NULL)
8190 *counter += totalFunc;
8191 for (i = 0; i < totalFunc; i++) {
8192 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8193 rpc_stat->stats[i].remote_peer = remoteHost;
8194 rpc_stat->stats[i].remote_port = remotePort;
8195 rpc_stat->stats[i].remote_is_server = isServer;
8196 rpc_stat->stats[i].interfaceId = rxInterface;
8197 rpc_stat->stats[i].func_total = totalFunc;
8198 rpc_stat->stats[i].func_index = i;
8200 opr_queue_Prepend(stats, &rpc_stat->entry);
8201 if (addToPeerList) {
8202 opr_queue_Prepend(&peerStats, &rpc_stat->entryPeers);
8209 rx_ClearProcessRPCStats(afs_int32 rxInterface)
8211 rx_interface_stat_p rpc_stat;
8214 if (rxInterface == -1)
8217 MUTEX_ENTER(&rx_rpc_stats);
8218 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8221 totalFunc = rpc_stat->stats[0].func_total;
8222 for (i = 0; i < totalFunc; i++)
8223 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8225 MUTEX_EXIT(&rx_rpc_stats);
8230 rx_ClearPeerRPCStats(afs_int32 rxInterface, afs_uint32 peerHost, afs_uint16 peerPort)
8232 rx_interface_stat_p rpc_stat;
8234 struct rx_peer * peer;
8236 if (rxInterface == -1)
8239 peer = rxi_FindPeer(peerHost, peerPort, 0);
8243 MUTEX_ENTER(&rx_rpc_stats);
8244 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8247 totalFunc = rpc_stat->stats[0].func_total;
8248 for (i = 0; i < totalFunc; i++)
8249 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8251 MUTEX_EXIT(&rx_rpc_stats);
8256 rx_CopyProcessRPCStats(afs_uint64 op)
8258 rx_interface_stat_p rpc_stat;
8259 rx_function_entry_v1_p rpcop_stat =
8260 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8261 int currentFunc = (op & MAX_AFS_UINT32);
8262 afs_int32 rxInterface = (op >> 32);
8264 if (!rxi_monitor_processStats)
8267 if (rxInterface == -1)
8270 if (rpcop_stat == NULL)
8273 MUTEX_ENTER(&rx_rpc_stats);
8274 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8277 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8278 sizeof(rx_function_entry_v1_t));
8279 MUTEX_EXIT(&rx_rpc_stats);
8281 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8288 rx_CopyPeerRPCStats(afs_uint64 op, afs_uint32 peerHost, afs_uint16 peerPort)
8290 rx_interface_stat_p rpc_stat;
8291 rx_function_entry_v1_p rpcop_stat =
8292 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8293 int currentFunc = (op & MAX_AFS_UINT32);
8294 afs_int32 rxInterface = (op >> 32);
8295 struct rx_peer *peer;
8297 if (!rxi_monitor_peerStats)
8300 if (rxInterface == -1)
8303 if (rpcop_stat == NULL)
8306 peer = rxi_FindPeer(peerHost, peerPort, 0);
8310 MUTEX_ENTER(&rx_rpc_stats);
8311 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8314 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8315 sizeof(rx_function_entry_v1_t));
8316 MUTEX_EXIT(&rx_rpc_stats);
8318 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8325 rx_ReleaseRPCStats(void *stats)
8328 rxi_Free(stats, sizeof(rx_function_entry_v1_t));
8332 * Given all of the information for a particular rpc
8333 * call, create (if needed) and update the stat totals for the rpc.
8336 * the queue of stats that will be updated with the new value
8338 * @param rxInterface
8339 * a unique number that identifies the rpc interface
8341 * @param currentFunc
8342 * the index of the function being invoked
8345 * the total number of functions in this interface
8348 * the amount of time this function waited for a thread
8351 * the amount of time this function invocation took to execute
8354 * the number bytes sent by this invocation
8357 * the number bytes received by this invocation
8360 * if true, this invocation was made to a server
8363 * the ip address of the remote host
8366 * the port of the remote host
8368 * @param addToPeerList
8369 * if != 0, add newly created stat to the global peer list
8372 * if a new stats structure is allocated, the counter will
8373 * be updated with the new number of allocated stat structures
8378 rxi_AddRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8379 afs_uint32 currentFunc, afs_uint32 totalFunc,
8380 struct clock *queueTime, struct clock *execTime,
8381 afs_uint64 bytesSent, afs_uint64 bytesRcvd, int isServer,
8382 afs_uint32 remoteHost, afs_uint32 remotePort,
8383 int addToPeerList, unsigned int *counter)
8386 rx_interface_stat_p rpc_stat;
8388 rpc_stat = rxi_FindRpcStat(stats, rxInterface, totalFunc, isServer,
8389 remoteHost, remotePort, addToPeerList, counter,
8397 * Increment the stats for this function
8400 rpc_stat->stats[currentFunc].invocations++;
8401 rpc_stat->stats[currentFunc].bytes_sent += bytesSent;
8402 rpc_stat->stats[currentFunc].bytes_rcvd += bytesRcvd;
8403 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8404 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8405 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8406 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8408 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8409 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8411 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8412 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8414 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8415 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8417 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8418 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8426 rxi_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8427 afs_uint32 currentFunc, afs_uint32 totalFunc,
8428 struct clock *queueTime, struct clock *execTime,
8429 afs_uint64 bytesSent, afs_uint64 bytesRcvd,
8433 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8436 MUTEX_ENTER(&rx_rpc_stats);
8438 if (rxi_monitor_peerStats) {
8439 MUTEX_ENTER(&peer->peer_lock);
8440 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8441 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8442 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8443 MUTEX_EXIT(&peer->peer_lock);
8446 if (rxi_monitor_processStats) {
8447 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8448 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8449 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8452 MUTEX_EXIT(&rx_rpc_stats);
8456 * Increment the times and count for a particular rpc function.
8458 * Traditionally this call was invoked from rxgen stubs. Modern stubs
8459 * call rx_RecordCallStatistics instead, so the public version of this
8460 * function is left purely for legacy callers.
8463 * The peer who invoked the rpc
8465 * @param rxInterface
8466 * A unique number that identifies the rpc interface
8468 * @param currentFunc
8469 * The index of the function being invoked
8472 * The total number of functions in this interface
8475 * The amount of time this function waited for a thread
8478 * The amount of time this function invocation took to execute
8481 * The number bytes sent by this invocation
8484 * The number bytes received by this invocation
8487 * If true, this invocation was made to a server
8491 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8492 afs_uint32 currentFunc, afs_uint32 totalFunc,
8493 struct clock *queueTime, struct clock *execTime,
8494 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8500 sent64 = ((afs_uint64)bytesSent->high << 32) + bytesSent->low;
8501 rcvd64 = ((afs_uint64)bytesRcvd->high << 32) + bytesRcvd->low;
8503 rxi_IncrementTimeAndCount(peer, rxInterface, currentFunc, totalFunc,
8504 queueTime, execTime, sent64, rcvd64,
8511 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8515 * IN callerVersion - the rpc stat version of the caller.
8517 * IN count - the number of entries to marshall.
8519 * IN stats - pointer to stats to be marshalled.
8521 * OUT ptr - Where to store the marshalled data.
8528 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8529 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8535 * We only support the first version
8537 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8538 *(ptr++) = stats->remote_peer;
8539 *(ptr++) = stats->remote_port;
8540 *(ptr++) = stats->remote_is_server;
8541 *(ptr++) = stats->interfaceId;
8542 *(ptr++) = stats->func_total;
8543 *(ptr++) = stats->func_index;
8544 *(ptr++) = stats->invocations >> 32;
8545 *(ptr++) = stats->invocations & MAX_AFS_UINT32;
8546 *(ptr++) = stats->bytes_sent >> 32;
8547 *(ptr++) = stats->bytes_sent & MAX_AFS_UINT32;
8548 *(ptr++) = stats->bytes_rcvd >> 32;
8549 *(ptr++) = stats->bytes_rcvd & MAX_AFS_UINT32;
8550 *(ptr++) = stats->queue_time_sum.sec;
8551 *(ptr++) = stats->queue_time_sum.usec;
8552 *(ptr++) = stats->queue_time_sum_sqr.sec;
8553 *(ptr++) = stats->queue_time_sum_sqr.usec;
8554 *(ptr++) = stats->queue_time_min.sec;
8555 *(ptr++) = stats->queue_time_min.usec;
8556 *(ptr++) = stats->queue_time_max.sec;
8557 *(ptr++) = stats->queue_time_max.usec;
8558 *(ptr++) = stats->execution_time_sum.sec;
8559 *(ptr++) = stats->execution_time_sum.usec;
8560 *(ptr++) = stats->execution_time_sum_sqr.sec;
8561 *(ptr++) = stats->execution_time_sum_sqr.usec;
8562 *(ptr++) = stats->execution_time_min.sec;
8563 *(ptr++) = stats->execution_time_min.usec;
8564 *(ptr++) = stats->execution_time_max.sec;
8565 *(ptr++) = stats->execution_time_max.usec;
8571 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8576 * IN callerVersion - the rpc stat version of the caller
8578 * OUT myVersion - the rpc stat version of this function
8580 * OUT clock_sec - local time seconds
8582 * OUT clock_usec - local time microseconds
8584 * OUT allocSize - the number of bytes allocated to contain stats
8586 * OUT statCount - the number stats retrieved from this process.
8588 * OUT stats - the actual stats retrieved from this process.
8592 * Returns void. If successful, stats will != NULL.
8596 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8597 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8598 size_t * allocSize, afs_uint32 * statCount,
8599 afs_uint32 ** stats)
8609 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8612 * Check to see if stats are enabled
8615 MUTEX_ENTER(&rx_rpc_stats);
8616 if (!rxi_monitor_processStats) {
8617 MUTEX_EXIT(&rx_rpc_stats);
8621 clock_GetTime(&now);
8622 *clock_sec = now.sec;
8623 *clock_usec = now.usec;
8626 * Allocate the space based upon the caller version
8628 * If the client is at an older version than we are,
8629 * we return the statistic data in the older data format, but
8630 * we still return our version number so the client knows we
8631 * are maintaining more data than it can retrieve.
8634 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8635 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8636 *statCount = rxi_rpc_process_stat_cnt;
8639 * This can't happen yet, but in the future version changes
8640 * can be handled by adding additional code here
8644 if (space > (size_t) 0) {
8646 ptr = *stats = rxi_Alloc(space);
8649 struct opr_queue *cursor;
8651 for (opr_queue_Scan(&processStats, cursor)) {
8652 struct rx_interface_stat *rpc_stat =
8653 opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8655 * Copy the data based upon the caller version
8657 rx_MarshallProcessRPCStats(callerVersion,
8658 rpc_stat->stats[0].func_total,
8659 rpc_stat->stats, &ptr);
8665 MUTEX_EXIT(&rx_rpc_stats);
8670 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8674 * IN callerVersion - the rpc stat version of the caller
8676 * OUT myVersion - the rpc stat version of this function
8678 * OUT clock_sec - local time seconds
8680 * OUT clock_usec - local time microseconds
8682 * OUT allocSize - the number of bytes allocated to contain stats
8684 * OUT statCount - the number of stats retrieved from the individual
8687 * OUT stats - the actual stats retrieved from the individual peer structures.
8691 * Returns void. If successful, stats will != NULL.
8695 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8696 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8697 size_t * allocSize, afs_uint32 * statCount,
8698 afs_uint32 ** stats)
8708 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8711 * Check to see if stats are enabled
8714 MUTEX_ENTER(&rx_rpc_stats);
8715 if (!rxi_monitor_peerStats) {
8716 MUTEX_EXIT(&rx_rpc_stats);
8720 clock_GetTime(&now);
8721 *clock_sec = now.sec;
8722 *clock_usec = now.usec;
8725 * Allocate the space based upon the caller version
8727 * If the client is at an older version than we are,
8728 * we return the statistic data in the older data format, but
8729 * we still return our version number so the client knows we
8730 * are maintaining more data than it can retrieve.
8733 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8734 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8735 *statCount = rxi_rpc_peer_stat_cnt;
8738 * This can't happen yet, but in the future version changes
8739 * can be handled by adding additional code here
8743 if (space > (size_t) 0) {
8745 ptr = *stats = rxi_Alloc(space);
8748 struct opr_queue *cursor;
8750 for (opr_queue_Scan(&peerStats, cursor)) {
8751 struct rx_interface_stat *rpc_stat
8752 = opr_queue_Entry(cursor, struct rx_interface_stat,
8756 * Copy the data based upon the caller version
8758 rx_MarshallProcessRPCStats(callerVersion,
8759 rpc_stat->stats[0].func_total,
8760 rpc_stat->stats, &ptr);
8766 MUTEX_EXIT(&rx_rpc_stats);
8771 * rx_FreeRPCStats - free memory allocated by
8772 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8776 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8777 * rx_RetrievePeerRPCStats
8779 * IN allocSize - the number of bytes in stats.
8787 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8789 rxi_Free(stats, allocSize);
8793 * rx_queryProcessRPCStats - see if process rpc stat collection is
8794 * currently enabled.
8800 * Returns 0 if stats are not enabled != 0 otherwise
8804 rx_queryProcessRPCStats(void)
8807 MUTEX_ENTER(&rx_rpc_stats);
8808 rc = rxi_monitor_processStats;
8809 MUTEX_EXIT(&rx_rpc_stats);
8814 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8820 * Returns 0 if stats are not enabled != 0 otherwise
8824 rx_queryPeerRPCStats(void)
8827 MUTEX_ENTER(&rx_rpc_stats);
8828 rc = rxi_monitor_peerStats;
8829 MUTEX_EXIT(&rx_rpc_stats);
8834 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8844 rx_enableProcessRPCStats(void)
8846 MUTEX_ENTER(&rx_rpc_stats);
8847 rx_enable_stats = 1;
8848 rxi_monitor_processStats = 1;
8849 MUTEX_EXIT(&rx_rpc_stats);
8853 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8863 rx_enablePeerRPCStats(void)
8865 MUTEX_ENTER(&rx_rpc_stats);
8866 rx_enable_stats = 1;
8867 rxi_monitor_peerStats = 1;
8868 MUTEX_EXIT(&rx_rpc_stats);
8872 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8882 rx_disableProcessRPCStats(void)
8884 struct opr_queue *cursor, *store;
8887 MUTEX_ENTER(&rx_rpc_stats);
8890 * Turn off process statistics and if peer stats is also off, turn
8894 rxi_monitor_processStats = 0;
8895 if (rxi_monitor_peerStats == 0) {
8896 rx_enable_stats = 0;
8899 for (opr_queue_ScanSafe(&processStats, cursor, store)) {
8900 unsigned int num_funcs = 0;
8901 struct rx_interface_stat *rpc_stat
8902 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8904 opr_queue_Remove(&rpc_stat->entry);
8906 num_funcs = rpc_stat->stats[0].func_total;
8908 sizeof(rx_interface_stat_t) +
8909 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8911 rxi_Free(rpc_stat, space);
8912 rxi_rpc_process_stat_cnt -= num_funcs;
8914 MUTEX_EXIT(&rx_rpc_stats);
8918 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8928 rx_disablePeerRPCStats(void)
8930 struct rx_peer **peer_ptr, **peer_end;
8934 * Turn off peer statistics and if process stats is also off, turn
8938 rxi_monitor_peerStats = 0;
8939 if (rxi_monitor_processStats == 0) {
8940 rx_enable_stats = 0;
8943 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8944 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8946 struct rx_peer *peer, *next, *prev;
8948 MUTEX_ENTER(&rx_peerHashTable_lock);
8949 MUTEX_ENTER(&rx_rpc_stats);
8950 for (prev = peer = *peer_ptr; peer; peer = next) {
8952 code = MUTEX_TRYENTER(&peer->peer_lock);
8955 struct opr_queue *cursor, *store;
8957 if (prev == *peer_ptr) {
8968 MUTEX_EXIT(&rx_peerHashTable_lock);
8970 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
8971 unsigned int num_funcs = 0;
8972 struct rx_interface_stat *rpc_stat
8973 = opr_queue_Entry(cursor, struct rx_interface_stat,
8976 opr_queue_Remove(&rpc_stat->entry);
8977 opr_queue_Remove(&rpc_stat->entryPeers);
8978 num_funcs = rpc_stat->stats[0].func_total;
8980 sizeof(rx_interface_stat_t) +
8981 rpc_stat->stats[0].func_total *
8982 sizeof(rx_function_entry_v1_t);
8984 rxi_Free(rpc_stat, space);
8985 rxi_rpc_peer_stat_cnt -= num_funcs;
8987 MUTEX_EXIT(&peer->peer_lock);
8989 MUTEX_ENTER(&rx_peerHashTable_lock);
8999 MUTEX_EXIT(&rx_rpc_stats);
9000 MUTEX_EXIT(&rx_peerHashTable_lock);
9005 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
9010 * IN clearFlag - flag indicating which stats to clear
9018 rx_clearProcessRPCStats(afs_uint32 clearFlag)
9020 struct opr_queue *cursor;
9022 MUTEX_ENTER(&rx_rpc_stats);
9024 for (opr_queue_Scan(&processStats, cursor)) {
9025 unsigned int num_funcs = 0, i;
9026 struct rx_interface_stat *rpc_stat
9027 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
9029 num_funcs = rpc_stat->stats[0].func_total;
9030 for (i = 0; i < num_funcs; i++) {
9031 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9032 rpc_stat->stats[i].invocations = 0;
9034 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9035 rpc_stat->stats[i].bytes_sent = 0;
9037 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9038 rpc_stat->stats[i].bytes_rcvd = 0;
9040 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9041 rpc_stat->stats[i].queue_time_sum.sec = 0;
9042 rpc_stat->stats[i].queue_time_sum.usec = 0;
9044 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9045 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9046 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9048 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9049 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9050 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9052 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9053 rpc_stat->stats[i].queue_time_max.sec = 0;
9054 rpc_stat->stats[i].queue_time_max.usec = 0;
9056 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9057 rpc_stat->stats[i].execution_time_sum.sec = 0;
9058 rpc_stat->stats[i].execution_time_sum.usec = 0;
9060 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9061 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9062 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9064 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9065 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9066 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9068 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9069 rpc_stat->stats[i].execution_time_max.sec = 0;
9070 rpc_stat->stats[i].execution_time_max.usec = 0;
9075 MUTEX_EXIT(&rx_rpc_stats);
9079 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
9084 * IN clearFlag - flag indicating which stats to clear
9092 rx_clearPeerRPCStats(afs_uint32 clearFlag)
9094 struct opr_queue *cursor;
9096 MUTEX_ENTER(&rx_rpc_stats);
9098 for (opr_queue_Scan(&peerStats, cursor)) {
9099 unsigned int num_funcs, i;
9100 struct rx_interface_stat *rpc_stat
9101 = opr_queue_Entry(cursor, struct rx_interface_stat, entryPeers);
9103 num_funcs = rpc_stat->stats[0].func_total;
9104 for (i = 0; i < num_funcs; i++) {
9105 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9106 rpc_stat->stats[i].invocations = 0;
9108 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9109 rpc_stat->stats[i].bytes_sent = 0;
9111 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9112 rpc_stat->stats[i].bytes_rcvd = 0;
9114 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9115 rpc_stat->stats[i].queue_time_sum.sec = 0;
9116 rpc_stat->stats[i].queue_time_sum.usec = 0;
9118 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9119 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9120 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9122 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9123 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9124 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9126 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9127 rpc_stat->stats[i].queue_time_max.sec = 0;
9128 rpc_stat->stats[i].queue_time_max.usec = 0;
9130 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9131 rpc_stat->stats[i].execution_time_sum.sec = 0;
9132 rpc_stat->stats[i].execution_time_sum.usec = 0;
9134 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9135 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9136 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9138 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9139 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9140 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9142 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9143 rpc_stat->stats[i].execution_time_max.sec = 0;
9144 rpc_stat->stats[i].execution_time_max.usec = 0;
9149 MUTEX_EXIT(&rx_rpc_stats);
9153 * rxi_rxstat_userok points to a routine that returns 1 if the caller
9154 * is authorized to enable/disable/clear RX statistics.
9156 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
9159 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
9161 rxi_rxstat_userok = proc;
9165 rx_RxStatUserOk(struct rx_call *call)
9167 if (!rxi_rxstat_userok)
9169 return rxi_rxstat_userok(call);
9174 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
9175 * function in the MSVC runtime DLL (msvcrt.dll).
9177 * Note: the system serializes calls to this function.
9180 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
9181 DWORD reason, /* reason function is being called */
9182 LPVOID reserved) /* reserved for future use */
9185 case DLL_PROCESS_ATTACH:
9186 /* library is being attached to a process */
9190 case DLL_PROCESS_DETACH:
9197 #endif /* AFS_NT40_ENV */
9200 int rx_DumpCalls(FILE *outputFile, char *cookie)
9202 #ifdef RXDEBUG_PACKET
9203 #ifdef KDUMP_RX_LOCK
9204 struct rx_call_rx_lock *c;
9211 #define RXDPRINTF sprintf
9212 #define RXDPRINTOUT output
9214 #define RXDPRINTF fprintf
9215 #define RXDPRINTOUT outputFile
9218 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9220 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9223 for (c = rx_allCallsp; c; c = c->allNextp) {
9224 u_short rqc, tqc, iovqc;
9226 MUTEX_ENTER(&c->lock);
9227 rqc = opr_queue_Count(&c->rq);
9228 tqc = opr_queue_Count(&c->tq);
9229 iovqc = opr_queue_Count(&c->app.iovq);
9231 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, "
9232 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9233 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9234 "resendEvent=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9235 "lastSendTime=%u, lastRecvTime=%u"
9236 #ifdef RX_ENABLE_LOCKS
9239 #ifdef RX_REFCOUNT_CHECK
9240 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9241 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9244 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,
9245 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9246 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9247 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9248 c->resendEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9249 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime
9250 #ifdef RX_ENABLE_LOCKS
9251 , (afs_uint32)c->refCount
9253 #ifdef RX_REFCOUNT_CHECK
9254 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9257 MUTEX_EXIT(&c->lock);
9260 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9263 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9265 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9267 #endif /* RXDEBUG_PACKET */