2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
13 #include <afs/param.h>
16 # include "afs/sysincludes.h"
17 # include "afsincludes.h"
22 # ifdef AFS_LINUX20_ENV
23 # include "h/socket.h"
25 # include "netinet/in.h"
27 # include "netinet/ip6.h"
28 # include "inet/common.h"
30 # include "inet/ip_ire.h"
32 # include "afs/afs_args.h"
33 # include "afs/afs_osi.h"
34 # ifdef RX_KERNEL_TRACE
35 # include "rx_kcommon.h"
37 # if defined(AFS_AIX_ENV)
41 # undef RXDEBUG /* turn off debugging */
43 # if defined(AFS_SGI_ENV)
44 # include "sys/debug.h"
47 # include "afs/sysincludes.h"
48 # include "afsincludes.h"
49 # endif /* !UKERNEL */
50 # include "afs/lock.h"
51 # include "rx_kmutex.h"
52 # include "rx_kernel.h"
53 # define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
54 # define AFSOP_STOP_AFS 211 /* Stop AFS process */
55 # define AFSOP_STOP_BKG 212 /* Stop BKG process */
56 extern afs_int32 afs_termState;
58 # include "sys/lockl.h"
59 # include "sys/lock_def.h"
60 # endif /* AFS_AIX41_ENV */
61 # include "afs/rxgen_consts.h"
66 # include <afs/afsutil.h>
67 # include <WINNT\afsreg.h>
75 #include <opr/queue.h>
76 #include <hcrypto/rand.h>
80 #include "rx_atomic.h"
81 #include "rx_globals.h"
83 #include "rx_internal.h"
90 #include "rx_packet.h"
91 #include "rx_server.h"
93 #include <afs/rxgen_consts.h>
96 #ifdef AFS_PTHREAD_ENV
98 int (*registerProgram) (pid_t, char *) = 0;
99 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
102 int (*registerProgram) (PROCESS, char *) = 0;
103 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
107 /* Local static routines */
108 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
109 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
110 struct rx_call *, struct rx_peer *,
112 static void rxi_Resend(struct rxevent *event, void *arg0, void *arg1,
114 static void rxi_SendDelayedAck(struct rxevent *event, void *call,
115 void *dummy, int dummy2);
116 static void rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1,
117 void *dummy, int dummy2);
118 static void rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1,
119 void *unused, int unused2);
120 static void rxi_ReapConnections(struct rxevent *unused, void *unused1,
121 void *unused2, int unused3);
122 static struct rx_packet *rxi_SendCallAbort(struct rx_call *call,
123 struct rx_packet *packet,
124 int istack, int force);
125 static void rxi_AckAll(struct rx_call *call);
126 static struct rx_connection
127 *rxi_FindConnection(osi_socket socket, afs_uint32 host, u_short port,
128 u_short serviceId, afs_uint32 cid,
129 afs_uint32 epoch, int type, u_int securityIndex,
130 int *unknownService);
131 static struct rx_packet
132 *rxi_ReceiveDataPacket(struct rx_call *call, struct rx_packet *np,
133 int istack, osi_socket socket,
134 afs_uint32 host, u_short port, int *tnop,
135 struct rx_call **newcallp);
136 static struct rx_packet
137 *rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
139 static struct rx_packet
140 *rxi_ReceiveResponsePacket(struct rx_connection *conn,
141 struct rx_packet *np, int istack);
142 static struct rx_packet
143 *rxi_ReceiveChallengePacket(struct rx_connection *conn,
144 struct rx_packet *np, int istack);
145 static void rxi_AttachServerProc(struct rx_call *call, osi_socket socket,
146 int *tnop, struct rx_call **newcallp);
147 static void rxi_ClearTransmitQueue(struct rx_call *call, int force);
148 static void rxi_ClearReceiveQueue(struct rx_call *call);
149 static void rxi_ResetCall(struct rx_call *call, int newcall);
150 static void rxi_ScheduleKeepAliveEvent(struct rx_call *call);
151 static void rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn);
152 static void rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs);
153 static void rxi_KeepAliveOn(struct rx_call *call);
154 static void rxi_GrowMTUOn(struct rx_call *call);
155 static void rxi_ChallengeOn(struct rx_connection *conn);
156 static int rxi_CheckCall(struct rx_call *call, int haveCTLock);
157 static void rxi_AckAllInTransmitQueue(struct rx_call *call);
158 static void rxi_CancelKeepAliveEvent(struct rx_call *call);
159 static void rxi_CancelDelayedAbortEvent(struct rx_call *call);
160 static void rxi_CancelGrowMTUEvent(struct rx_call *call);
161 static void update_nextCid(void);
163 #ifdef RX_ENABLE_LOCKS
165 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
166 rx_atomic_t rxi_start_in_error;
168 #endif /* RX_ENABLE_LOCKS */
170 /* Constant delay time before sending an acknowledge of the last packet
171 * received. This is to avoid sending an extra acknowledge when the
172 * client is about to make another call, anyway, or the server is
175 * The lastAckDelay may not exceeed 400ms without causing peers to
176 * unecessarily timeout.
178 struct clock rx_lastAckDelay = {0, 400000};
180 /* Constant delay time before sending a soft ack when none was requested.
181 * This is to make sure we send soft acks before the sender times out,
182 * Normally we wait and send a hard ack when the receiver consumes the packet
184 * This value has been 100ms in all shipping versions of OpenAFS. Changing it
185 * will require changes to the peer's RTT calculations.
187 struct clock rx_softAckDelay = {0, 100000};
190 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
191 * currently allocated within rx. This number is used to allocate the
192 * memory required to return the statistics when queried.
193 * Protected by the rx_rpc_stats mutex.
196 static unsigned int rxi_rpc_peer_stat_cnt;
199 * rxi_rpc_process_stat_cnt counts the total number of local process stat
200 * structures currently allocated within rx. The number is used to allocate
201 * the memory required to return the statistics when queried.
202 * Protected by the rx_rpc_stats mutex.
205 static unsigned int rxi_rpc_process_stat_cnt;
207 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
208 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
210 /* Incoming calls wait on this queue when there are no available
211 * server processes */
212 struct opr_queue rx_incomingCallQueue;
214 /* Server processes wait on this queue when there are no appropriate
215 * calls to process */
216 struct opr_queue rx_idleServerQueue;
218 #if !defined(offsetof)
219 #include <stddef.h> /* for definition of offsetof() */
222 #ifdef RX_ENABLE_LOCKS
223 afs_kmutex_t rx_atomic_mutex;
226 /* Forward prototypes */
227 static struct rx_call * rxi_NewCall(struct rx_connection *, int);
230 putConnection (struct rx_connection *conn) {
231 MUTEX_ENTER(&rx_refcnt_mutex);
233 MUTEX_EXIT(&rx_refcnt_mutex);
236 #ifdef AFS_PTHREAD_ENV
239 * Use procedural initialization of mutexes/condition variables
243 extern afs_kmutex_t rx_quota_mutex;
244 extern afs_kmutex_t rx_pthread_mutex;
245 extern afs_kmutex_t rx_packets_mutex;
246 extern afs_kmutex_t rx_refcnt_mutex;
247 extern afs_kmutex_t des_init_mutex;
248 extern afs_kmutex_t des_random_mutex;
250 extern afs_kmutex_t rx_clock_mutex;
251 extern afs_kmutex_t rxi_connCacheMutex;
252 extern afs_kmutex_t event_handler_mutex;
253 extern afs_kmutex_t listener_mutex;
254 extern afs_kmutex_t rx_if_init_mutex;
255 extern afs_kmutex_t rx_if_mutex;
257 extern afs_kcondvar_t rx_event_handler_cond;
258 extern afs_kcondvar_t rx_listener_cond;
261 static afs_kmutex_t epoch_mutex;
262 static afs_kmutex_t rx_init_mutex;
263 static afs_kmutex_t rx_debug_mutex;
264 static afs_kmutex_t rx_rpc_stats;
267 rxi_InitPthread(void)
269 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
270 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
271 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
272 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
274 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
275 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
276 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
277 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
278 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
279 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
281 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
282 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
283 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
284 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
285 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
288 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
289 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
292 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
293 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
295 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
296 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
297 MUTEX_INIT(&rx_mallocedPktQ_lock, "rx_mallocedPktQ_lock", MUTEX_DEFAULT,
300 #ifdef RX_ENABLE_LOCKS
303 #endif /* RX_LOCKS_DB */
304 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
305 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
307 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
309 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
311 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
313 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
315 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
317 #endif /* RX_ENABLE_LOCKS */
320 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
321 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
323 * The rx_stats_mutex mutex protects the following global variables:
324 * rxi_lowConnRefCount
325 * rxi_lowPeerRefCount
334 * The rx_quota_mutex mutex protects the following global variables:
342 * The rx_freePktQ_lock protects the following global variables:
347 * The rx_packets_mutex mutex protects the following global variables:
355 * The rx_pthread_mutex mutex protects the following global variables:
356 * rxi_fcfs_thread_num
359 #define INIT_PTHREAD_LOCKS
363 /* Variables for handling the minProcs implementation. availProcs gives the
364 * number of threads available in the pool at this moment (not counting dudes
365 * executing right now). totalMin gives the total number of procs required
366 * for handling all minProcs requests. minDeficit is a dynamic variable
367 * tracking the # of procs required to satisfy all of the remaining minProcs
369 * For fine grain locking to work, the quota check and the reservation of
370 * a server thread has to come while rxi_availProcs and rxi_minDeficit
371 * are locked. To this end, the code has been modified under #ifdef
372 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
373 * same time. A new function, ReturnToServerPool() returns the allocation.
375 * A call can be on several queue's (but only one at a time). When
376 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
377 * that no one else is touching the queue. To this end, we store the address
378 * of the queue lock in the call structure (under the call lock) when we
379 * put the call on a queue, and we clear the call_queue_lock when the
380 * call is removed from a queue (once the call lock has been obtained).
381 * This allows rxi_ResetCall to safely synchronize with others wishing
382 * to manipulate the queue.
385 #if defined(RX_ENABLE_LOCKS)
386 static afs_kmutex_t rx_rpc_stats;
389 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
390 ** pretty good that the next packet coming in is from the same connection
391 ** as the last packet, since we're send multiple packets in a transmit window.
393 struct rx_connection *rxLastConn = 0;
395 #ifdef RX_ENABLE_LOCKS
396 /* The locking hierarchy for rx fine grain locking is composed of these
399 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
400 * also protects updates to rx_nextCid
401 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
402 * call->lock - locks call data fields.
403 * These are independent of each other:
404 * rx_freeCallQueue_lock
409 * serverQueueEntry->lock
410 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
412 * peer->lock - locks peer data fields.
413 * conn_data_lock - that more than one thread is not updating a conn data
414 * field at the same time.
425 * Do we need a lock to protect the peer field in the conn structure?
426 * conn->peer was previously a constant for all intents and so has no
427 * lock protecting this field. The multihomed client delta introduced
428 * a RX code change : change the peer field in the connection structure
429 * to that remote interface from which the last packet for this
430 * connection was sent out. This may become an issue if further changes
433 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
434 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
436 /* rxdb_fileID is used to identify the lock location, along with line#. */
437 static int rxdb_fileID = RXDB_FILE_RX;
438 #endif /* RX_LOCKS_DB */
439 #else /* RX_ENABLE_LOCKS */
440 #define SET_CALL_QUEUE_LOCK(C, L)
441 #define CLEAR_CALL_QUEUE_LOCK(C)
442 #endif /* RX_ENABLE_LOCKS */
443 struct rx_serverQueueEntry *rx_waitForPacket = 0;
445 /* ------------Exported Interfaces------------- */
447 /* Initialize rx. A port number may be mentioned, in which case this
448 * becomes the default port number for any service installed later.
449 * If 0 is provided for the port number, a random port will be chosen
450 * by the kernel. Whether this will ever overlap anything in
451 * /etc/services is anybody's guess... Returns 0 on success, -1 on
453 #if !(defined(AFS_NT40_ENV) || defined(RXK_UPCALL_ENV))
456 rx_atomic_t rxinit_status = RX_ATOMIC_INIT(1);
459 rx_InitHost(u_int host, u_int port)
466 char *htable, *ptable;
471 if (!rx_atomic_test_and_clear_bit(&rxinit_status, 0))
472 return 0; /* already started */
478 if (afs_winsockInit() < 0)
484 * Initialize anything necessary to provide a non-premptive threading
487 rxi_InitializeThreadSupport();
490 /* Allocate and initialize a socket for client and perhaps server
493 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
494 if (rx_socket == OSI_NULLSOCKET) {
497 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
500 #endif /* RX_LOCKS_DB */
501 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
502 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
503 MUTEX_INIT(&rx_atomic_mutex, "rx_atomic_mutex", MUTEX_DEFAULT, 0);
504 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
505 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
506 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
507 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
508 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
509 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
510 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
512 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
514 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
516 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
518 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
519 MUTEX_INIT(&rx_mallocedPktQ_lock, "rx_mallocedPktQ_lock", MUTEX_DEFAULT,
522 #if defined(AFS_HPUX110_ENV)
524 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
525 #endif /* AFS_HPUX110_ENV */
526 #endif /* RX_ENABLE_LOCKS && KERNEL */
529 rx_connDeadTime = 12;
530 rx_tranquil = 0; /* reset flag */
531 rxi_ResetStatistics();
532 htable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
533 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
534 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
535 ptable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
536 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
537 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
539 /* Malloc up a bunch of packets & buffers */
541 opr_queue_Init(&rx_freePacketQueue);
542 rxi_NeedMorePackets = FALSE;
543 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
544 opr_queue_Init(&rx_mallocedPacketQueue);
546 /* enforce a minimum number of allocated packets */
547 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
548 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
550 /* allocate the initial free packet pool */
551 #ifdef RX_ENABLE_TSFPQ
552 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
553 #else /* RX_ENABLE_TSFPQ */
554 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
555 #endif /* RX_ENABLE_TSFPQ */
562 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
563 tv.tv_sec = clock_now.sec;
564 tv.tv_usec = clock_now.usec;
565 srand((unsigned int)tv.tv_usec);
572 #if defined(KERNEL) && !defined(UKERNEL)
573 /* Really, this should never happen in a real kernel */
576 struct sockaddr_in addr;
578 int addrlen = sizeof(addr);
580 socklen_t addrlen = sizeof(addr);
582 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
584 osi_Free(htable, rx_hashTableSize * sizeof(struct rx_connection *));
587 rx_port = addr.sin_port;
590 rx_stats.minRtt.sec = 9999999;
591 if (RAND_bytes(&rx_epoch, sizeof(rx_epoch)) != 1)
593 rx_epoch = (rx_epoch & ~0x40000000) | 0x80000000;
594 if (RAND_bytes(&rx_nextCid, sizeof(rx_nextCid)) != 1)
596 rx_nextCid &= RX_CIDMASK;
597 MUTEX_ENTER(&rx_quota_mutex);
598 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
599 MUTEX_EXIT(&rx_quota_mutex);
600 /* *Slightly* random start time for the cid. This is just to help
601 * out with the hashing function at the peer */
602 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
603 rx_connHashTable = (struct rx_connection **)htable;
604 rx_peerHashTable = (struct rx_peer **)ptable;
606 rx_hardAckDelay.sec = 0;
607 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
609 rxevent_Init(20, rxi_ReScheduleEvents);
611 /* Initialize various global queues */
612 opr_queue_Init(&rx_idleServerQueue);
613 opr_queue_Init(&rx_incomingCallQueue);
614 opr_queue_Init(&rx_freeCallQueue);
616 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
617 /* Initialize our list of usable IP addresses. */
621 /* Start listener process (exact function is dependent on the
622 * implementation environment--kernel or user space) */
626 rx_atomic_clear_bit(&rxinit_status, 0);
633 return rx_InitHost(htonl(INADDR_ANY), port);
639 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
640 * maintaing the round trip timer.
645 * Start a new RTT timer for a given call and packet.
647 * There must be no resendEvent already listed for this call, otherwise this
648 * will leak events - intended for internal use within the RTO code only
651 * the RX call to start the timer for
652 * @param[in] lastPacket
653 * a flag indicating whether the last packet has been sent or not
655 * @pre call must be locked before calling this function
659 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
661 struct clock now, retryTime;
663 MUTEX_ASSERT(&call->lock);
667 clock_Add(&retryTime, &call->rto);
669 /* If we're sending the last packet, and we're the client, then the server
670 * may wait for an additional 400ms before returning the ACK, wait for it
671 * rather than hitting a timeout */
672 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
673 clock_Addmsec(&retryTime, 400);
675 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
676 call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
681 * Cancel an RTT timer for a given call.
685 * the RX call to cancel the timer for
687 * @pre call must be locked before calling this function
692 rxi_rto_cancel(struct rx_call *call)
694 MUTEX_ASSERT(&call->lock);
695 if (rxevent_Cancel(&call->resendEvent))
696 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
700 * Tell the RTO timer that we have sent a packet.
702 * If the timer isn't already running, then start it. If the timer is running,
706 * the RX call that the packet has been sent on
707 * @param[in] lastPacket
708 * A flag which is true if this is the last packet for the call
710 * @pre The call must be locked before calling this function
715 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
717 if (call->resendEvent)
720 rxi_rto_startTimer(call, lastPacket, istack);
724 * Tell the RTO timer that we have received an new ACK message
726 * This function should be called whenever a call receives an ACK that
727 * acknowledges new packets. Whatever happens, we stop the current timer.
728 * If there are unacked packets in the queue which have been sent, then
729 * we restart the timer from now. Otherwise, we leave it stopped.
732 * the RX call that the ACK has been received on
736 rxi_rto_packet_acked(struct rx_call *call, int istack)
738 struct opr_queue *cursor;
740 rxi_rto_cancel(call);
742 if (opr_queue_IsEmpty(&call->tq))
745 for (opr_queue_Scan(&call->tq, cursor)) {
746 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
747 if (p->header.seq > call->tfirst + call->twind)
750 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
751 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
759 * Set an initial round trip timeout for a peer connection
761 * @param[in] secs The timeout to set in seconds
765 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
766 peer->rtt = secs * 8000;
770 * Set a delayed ack event on the specified call for the given time
772 * @param[in] call - the call on which to set the event
773 * @param[in] offset - the delay from now after which the event fires
776 rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
778 struct clock now, when;
780 MUTEX_ASSERT(&call->lock);
783 clock_Add(&when, offset);
785 if (clock_Gt(&call->delayedAckTime, &when) &&
786 rxevent_Cancel(&call->delayedAckEvent)) {
787 /* We successfully cancelled an event too far in the future to install
788 * our new one; we can reuse the reference on the call. */
789 call->delayedAckEvent = rxevent_Post(&when, &now, rxi_SendDelayedAck,
792 call->delayedAckTime = when;
793 } else if (call->delayedAckEvent == NULL) {
794 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
795 call->delayedAckEvent = rxevent_Post(&when, &now,
798 call->delayedAckTime = when;
803 rxi_CancelDelayedAckEvent(struct rx_call *call)
805 MUTEX_ASSERT(&call->lock);
806 /* Only drop the ref if we cancelled it before it could run. */
807 if (rxevent_Cancel(&call->delayedAckEvent))
808 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
811 /* called with unincremented nRequestsRunning to see if it is OK to start
812 * a new thread in this service. Could be "no" for two reasons: over the
813 * max quota, or would prevent others from reaching their min quota.
815 #ifdef RX_ENABLE_LOCKS
816 /* This verion of QuotaOK reserves quota if it's ok while the
817 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
820 QuotaOK(struct rx_service *aservice)
822 /* check if over max quota */
823 if (aservice->nRequestsRunning >= aservice->maxProcs) {
827 /* under min quota, we're OK */
828 /* otherwise, can use only if there are enough to allow everyone
829 * to go to their min quota after this guy starts.
832 MUTEX_ENTER(&rx_quota_mutex);
833 if ((aservice->nRequestsRunning < aservice->minProcs)
834 || (rxi_availProcs > rxi_minDeficit)) {
835 aservice->nRequestsRunning++;
836 /* just started call in minProcs pool, need fewer to maintain
838 if (aservice->nRequestsRunning <= aservice->minProcs)
841 MUTEX_EXIT(&rx_quota_mutex);
844 MUTEX_EXIT(&rx_quota_mutex);
850 ReturnToServerPool(struct rx_service *aservice)
852 aservice->nRequestsRunning--;
853 MUTEX_ENTER(&rx_quota_mutex);
854 if (aservice->nRequestsRunning < aservice->minProcs)
857 MUTEX_EXIT(&rx_quota_mutex);
860 #else /* RX_ENABLE_LOCKS */
862 QuotaOK(struct rx_service *aservice)
865 /* under min quota, we're OK */
866 if (aservice->nRequestsRunning < aservice->minProcs)
869 /* check if over max quota */
870 if (aservice->nRequestsRunning >= aservice->maxProcs)
873 /* otherwise, can use only if there are enough to allow everyone
874 * to go to their min quota after this guy starts.
876 MUTEX_ENTER(&rx_quota_mutex);
877 if (rxi_availProcs > rxi_minDeficit)
879 MUTEX_EXIT(&rx_quota_mutex);
882 #endif /* RX_ENABLE_LOCKS */
885 /* Called by rx_StartServer to start up lwp's to service calls.
886 NExistingProcs gives the number of procs already existing, and which
887 therefore needn't be created. */
889 rxi_StartServerProcs(int nExistingProcs)
891 struct rx_service *service;
896 /* For each service, reserve N processes, where N is the "minimum"
897 * number of processes that MUST be able to execute a request in parallel,
898 * at any time, for that process. Also compute the maximum difference
899 * between any service's maximum number of processes that can run
900 * (i.e. the maximum number that ever will be run, and a guarantee
901 * that this number will run if other services aren't running), and its
902 * minimum number. The result is the extra number of processes that
903 * we need in order to provide the latter guarantee */
904 for (i = 0; i < RX_MAX_SERVICES; i++) {
906 service = rx_services[i];
907 if (service == (struct rx_service *)0)
909 nProcs += service->minProcs;
910 diff = service->maxProcs - service->minProcs;
914 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
915 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
916 for (i = 0; i < nProcs; i++) {
917 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
923 /* This routine is only required on Windows */
925 rx_StartClientThread(void)
927 #ifdef AFS_PTHREAD_ENV
929 pid = pthread_self();
930 #endif /* AFS_PTHREAD_ENV */
932 #endif /* AFS_NT40_ENV */
934 /* This routine must be called if any services are exported. If the
935 * donateMe flag is set, the calling process is donated to the server
938 rx_StartServer(int donateMe)
940 struct rx_service *service;
946 /* Start server processes, if necessary (exact function is dependent
947 * on the implementation environment--kernel or user space). DonateMe
948 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
949 * case, one less new proc will be created rx_StartServerProcs.
951 rxi_StartServerProcs(donateMe);
953 /* count up the # of threads in minProcs, and add set the min deficit to
954 * be that value, too.
956 for (i = 0; i < RX_MAX_SERVICES; i++) {
957 service = rx_services[i];
958 if (service == (struct rx_service *)0)
960 MUTEX_ENTER(&rx_quota_mutex);
961 rxi_totalMin += service->minProcs;
962 /* below works even if a thread is running, since minDeficit would
963 * still have been decremented and later re-incremented.
965 rxi_minDeficit += service->minProcs;
966 MUTEX_EXIT(&rx_quota_mutex);
969 /* Turn on reaping of idle server connections */
970 rxi_ReapConnections(NULL, NULL, NULL, 0);
979 #ifdef AFS_PTHREAD_ENV
981 pid = afs_pointer_to_int(pthread_self());
982 #else /* AFS_PTHREAD_ENV */
984 LWP_CurrentProcess(&pid);
985 #endif /* AFS_PTHREAD_ENV */
987 sprintf(name, "srv_%d", ++nProcs);
989 (*registerProgram) (pid, name);
991 #endif /* AFS_NT40_ENV */
992 rx_ServerProc(NULL); /* Never returns */
994 #ifdef RX_ENABLE_TSFPQ
995 /* no use leaving packets around in this thread's local queue if
996 * it isn't getting donated to the server thread pool.
998 rxi_FlushLocalPacketsTSFPQ();
999 #endif /* RX_ENABLE_TSFPQ */
1003 /* Create a new client connection to the specified service, using the
1004 * specified security object to implement the security model for this
1006 struct rx_connection *
1007 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
1008 struct rx_securityClass *securityObject,
1009 int serviceSecurityIndex)
1012 struct rx_connection *conn;
1017 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
1018 "serviceSecurityIndex %d)\n",
1019 ntohl(shost), ntohs(sport), sservice, securityObject,
1020 serviceSecurityIndex));
1022 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
1023 * the case of kmem_alloc? */
1024 conn = rxi_AllocConnection();
1025 #ifdef RX_ENABLE_LOCKS
1026 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
1027 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
1028 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
1031 MUTEX_ENTER(&rx_connHashTable_lock);
1032 conn->type = RX_CLIENT_CONNECTION;
1033 conn->epoch = rx_epoch;
1034 conn->cid = rx_nextCid;
1036 conn->peer = rxi_FindPeer(shost, sport, 1);
1037 conn->serviceId = sservice;
1038 conn->securityObject = securityObject;
1039 conn->securityData = (void *) 0;
1040 conn->securityIndex = serviceSecurityIndex;
1041 rx_SetConnDeadTime(conn, rx_connDeadTime);
1042 rx_SetConnSecondsUntilNatPing(conn, 0);
1043 conn->ackRate = RX_FAST_ACK_RATE;
1044 conn->nSpecific = 0;
1045 conn->specific = NULL;
1046 conn->challengeEvent = NULL;
1047 conn->delayedAbortEvent = NULL;
1048 conn->abortCount = 0;
1050 for (i = 0; i < RX_MAXCALLS; i++) {
1051 conn->twind[i] = rx_initSendWindow;
1052 conn->rwind[i] = rx_initReceiveWindow;
1053 conn->lastBusy[i] = 0;
1056 RXS_NewConnection(securityObject, conn);
1058 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1060 conn->refCount++; /* no lock required since only this thread knows... */
1061 conn->next = rx_connHashTable[hashindex];
1062 rx_connHashTable[hashindex] = conn;
1063 if (rx_stats_active)
1064 rx_atomic_inc(&rx_stats.nClientConns);
1065 MUTEX_EXIT(&rx_connHashTable_lock);
1071 * Ensure a connection's timeout values are valid.
1073 * @param[in] conn The connection to check
1075 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1076 * unless idleDeadTime and/or hardDeadTime are not set
1080 rxi_CheckConnTimeouts(struct rx_connection *conn)
1082 /* a connection's timeouts must have the relationship
1083 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1084 * total loss of network to a peer may cause an idle timeout instead of a
1085 * dead timeout, simply because the idle timeout gets hit first. Also set
1086 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1087 /* this logic is slightly complicated by the fact that
1088 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1090 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1091 if (conn->idleDeadTime) {
1092 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1094 if (conn->hardDeadTime) {
1095 if (conn->idleDeadTime) {
1096 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1098 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1104 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1106 /* The idea is to set the dead time to a value that allows several
1107 * keepalives to be dropped without timing out the connection. */
1108 conn->secondsUntilDead = seconds;
1109 rxi_CheckConnTimeouts(conn);
1110 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1114 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1116 conn->hardDeadTime = seconds;
1117 rxi_CheckConnTimeouts(conn);
1121 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1123 conn->idleDeadTime = seconds;
1124 rxi_CheckConnTimeouts(conn);
1127 int rxi_lowPeerRefCount = 0;
1128 int rxi_lowConnRefCount = 0;
1131 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1132 * NOTE: must not be called with rx_connHashTable_lock held.
1135 rxi_CleanupConnection(struct rx_connection *conn)
1137 /* Notify the service exporter, if requested, that this connection
1138 * is being destroyed */
1139 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1140 (*conn->service->destroyConnProc) (conn);
1142 /* Notify the security module that this connection is being destroyed */
1143 RXS_DestroyConnection(conn->securityObject, conn);
1145 /* If this is the last connection using the rx_peer struct, set its
1146 * idle time to now. rxi_ReapConnections will reap it if it's still
1147 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1149 MUTEX_ENTER(&rx_peerHashTable_lock);
1150 if (conn->peer->refCount < 2) {
1151 conn->peer->idleWhen = clock_Sec();
1152 if (conn->peer->refCount < 1) {
1153 conn->peer->refCount = 1;
1154 if (rx_stats_active) {
1155 MUTEX_ENTER(&rx_stats_mutex);
1156 rxi_lowPeerRefCount++;
1157 MUTEX_EXIT(&rx_stats_mutex);
1161 conn->peer->refCount--;
1162 MUTEX_EXIT(&rx_peerHashTable_lock);
1164 if (rx_stats_active)
1166 if (conn->type == RX_SERVER_CONNECTION)
1167 rx_atomic_dec(&rx_stats.nServerConns);
1169 rx_atomic_dec(&rx_stats.nClientConns);
1172 if (conn->specific) {
1174 for (i = 0; i < conn->nSpecific; i++) {
1175 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1176 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1177 conn->specific[i] = NULL;
1179 free(conn->specific);
1181 conn->specific = NULL;
1182 conn->nSpecific = 0;
1183 #endif /* !KERNEL */
1185 MUTEX_DESTROY(&conn->conn_call_lock);
1186 MUTEX_DESTROY(&conn->conn_data_lock);
1187 CV_DESTROY(&conn->conn_call_cv);
1189 rxi_FreeConnection(conn);
1192 /* Destroy the specified connection */
1194 rxi_DestroyConnection(struct rx_connection *conn)
1196 MUTEX_ENTER(&rx_connHashTable_lock);
1197 rxi_DestroyConnectionNoLock(conn);
1198 /* conn should be at the head of the cleanup list */
1199 if (conn == rx_connCleanup_list) {
1200 rx_connCleanup_list = rx_connCleanup_list->next;
1201 MUTEX_EXIT(&rx_connHashTable_lock);
1202 rxi_CleanupConnection(conn);
1204 #ifdef RX_ENABLE_LOCKS
1206 MUTEX_EXIT(&rx_connHashTable_lock);
1208 #endif /* RX_ENABLE_LOCKS */
1212 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1214 struct rx_connection **conn_ptr;
1222 MUTEX_ENTER(&conn->conn_data_lock);
1223 MUTEX_ENTER(&rx_refcnt_mutex);
1224 if (conn->refCount > 0)
1227 #ifdef RX_REFCOUNT_CHECK
1228 osi_Assert(conn->refCount == 0);
1230 if (rx_stats_active) {
1231 MUTEX_ENTER(&rx_stats_mutex);
1232 rxi_lowConnRefCount++;
1233 MUTEX_EXIT(&rx_stats_mutex);
1237 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1238 /* Busy; wait till the last guy before proceeding */
1239 MUTEX_EXIT(&rx_refcnt_mutex);
1240 MUTEX_EXIT(&conn->conn_data_lock);
1245 /* If the client previously called rx_NewCall, but it is still
1246 * waiting, treat this as a running call, and wait to destroy the
1247 * connection later when the call completes. */
1248 if ((conn->type == RX_CLIENT_CONNECTION)
1249 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1250 conn->flags |= RX_CONN_DESTROY_ME;
1251 MUTEX_EXIT(&rx_refcnt_mutex);
1252 MUTEX_EXIT(&conn->conn_data_lock);
1256 MUTEX_EXIT(&rx_refcnt_mutex);
1257 MUTEX_EXIT(&conn->conn_data_lock);
1259 /* Check for extant references to this connection */
1260 MUTEX_ENTER(&conn->conn_call_lock);
1261 for (i = 0; i < RX_MAXCALLS; i++) {
1262 struct rx_call *call = conn->call[i];
1265 if (conn->type == RX_CLIENT_CONNECTION) {
1266 MUTEX_ENTER(&call->lock);
1267 if (call->delayedAckEvent) {
1268 /* Push the final acknowledgment out now--there
1269 * won't be a subsequent call to acknowledge the
1270 * last reply packets */
1271 rxi_CancelDelayedAckEvent(call);
1272 if (call->state == RX_STATE_PRECALL
1273 || call->state == RX_STATE_ACTIVE) {
1274 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1279 MUTEX_EXIT(&call->lock);
1283 MUTEX_EXIT(&conn->conn_call_lock);
1285 #ifdef RX_ENABLE_LOCKS
1287 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1288 MUTEX_EXIT(&conn->conn_data_lock);
1290 /* Someone is accessing a packet right now. */
1294 #endif /* RX_ENABLE_LOCKS */
1297 /* Don't destroy the connection if there are any call
1298 * structures still in use */
1299 MUTEX_ENTER(&conn->conn_data_lock);
1300 conn->flags |= RX_CONN_DESTROY_ME;
1301 MUTEX_EXIT(&conn->conn_data_lock);
1306 /* Remove from connection hash table before proceeding */
1308 &rx_connHashTable[CONN_HASH
1309 (peer->host, peer->port, conn->cid, conn->epoch,
1311 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1312 if (*conn_ptr == conn) {
1313 *conn_ptr = conn->next;
1317 /* if the conn that we are destroying was the last connection, then we
1318 * clear rxLastConn as well */
1319 if (rxLastConn == conn)
1322 /* Make sure the connection is completely reset before deleting it. */
1324 * Pending events hold a refcount, so we can't get here if they are
1326 osi_Assert(conn->challengeEvent == NULL);
1327 osi_Assert(conn->delayedAbortEvent == NULL);
1328 osi_Assert(conn->natKeepAliveEvent == NULL);
1329 osi_Assert(conn->checkReachEvent == NULL);
1331 /* Add the connection to the list of destroyed connections that
1332 * need to be cleaned up. This is necessary to avoid deadlocks
1333 * in the routines we call to inform others that this connection is
1334 * being destroyed. */
1335 conn->next = rx_connCleanup_list;
1336 rx_connCleanup_list = conn;
1339 /* Externally available version */
1341 rx_DestroyConnection(struct rx_connection *conn)
1346 rxi_DestroyConnection(conn);
1351 rx_GetConnection(struct rx_connection *conn)
1356 MUTEX_ENTER(&rx_refcnt_mutex);
1358 MUTEX_EXIT(&rx_refcnt_mutex);
1362 #ifdef RX_ENABLE_LOCKS
1363 /* Wait for the transmit queue to no longer be busy.
1364 * requires the call->lock to be held */
1366 rxi_WaitforTQBusy(struct rx_call *call) {
1367 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1368 call->flags |= RX_CALL_TQ_WAIT;
1370 MUTEX_ASSERT(&call->lock);
1371 CV_WAIT(&call->cv_tq, &call->lock);
1373 if (call->tqWaiters == 0) {
1374 call->flags &= ~RX_CALL_TQ_WAIT;
1381 rxi_WakeUpTransmitQueue(struct rx_call *call)
1383 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1384 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1385 call, call->tqWaiters, call->flags));
1386 #ifdef RX_ENABLE_LOCKS
1387 MUTEX_ASSERT(&call->lock);
1388 CV_BROADCAST(&call->cv_tq);
1389 #else /* RX_ENABLE_LOCKS */
1390 osi_rxWakeup(&call->tq);
1391 #endif /* RX_ENABLE_LOCKS */
1395 /* Start a new rx remote procedure call, on the specified connection.
1396 * If wait is set to 1, wait for a free call channel; otherwise return
1397 * 0. Maxtime gives the maximum number of seconds this call may take,
1398 * after rx_NewCall returns. After this time interval, a call to any
1399 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1400 * For fine grain locking, we hold the conn_call_lock in order to
1401 * to ensure that we don't get signalle after we found a call in an active
1402 * state and before we go to sleep.
1405 rx_NewCall(struct rx_connection *conn)
1407 int i, wait, ignoreBusy = 1;
1408 struct rx_call *call;
1409 struct clock queueTime;
1410 afs_uint32 leastBusy = 0;
1414 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1417 clock_GetTime(&queueTime);
1419 * Check if there are others waiting for a new call.
1420 * If so, let them go first to avoid starving them.
1421 * This is a fairly simple scheme, and might not be
1422 * a complete solution for large numbers of waiters.
1424 * makeCallWaiters keeps track of the number of
1425 * threads waiting to make calls and the
1426 * RX_CONN_MAKECALL_WAITING flag bit is used to
1427 * indicate that there are indeed calls waiting.
1428 * The flag is set when the waiter is incremented.
1429 * It is only cleared when makeCallWaiters is 0.
1430 * This prevents us from accidently destroying the
1431 * connection while it is potentially about to be used.
1433 MUTEX_ENTER(&conn->conn_call_lock);
1434 MUTEX_ENTER(&conn->conn_data_lock);
1435 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1436 conn->flags |= RX_CONN_MAKECALL_WAITING;
1437 conn->makeCallWaiters++;
1438 MUTEX_EXIT(&conn->conn_data_lock);
1440 #ifdef RX_ENABLE_LOCKS
1441 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1445 MUTEX_ENTER(&conn->conn_data_lock);
1446 conn->makeCallWaiters--;
1447 if (conn->makeCallWaiters == 0)
1448 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1451 /* We are now the active thread in rx_NewCall */
1452 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1453 MUTEX_EXIT(&conn->conn_data_lock);
1458 for (i = 0; i < RX_MAXCALLS; i++) {
1459 call = conn->call[i];
1461 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1462 /* we're not ignoring busy call slots; only look at the
1463 * call slot that is the "least" busy */
1467 if (call->state == RX_STATE_DALLY) {
1468 MUTEX_ENTER(&call->lock);
1469 if (call->state == RX_STATE_DALLY) {
1470 if (ignoreBusy && conn->lastBusy[i]) {
1471 /* if we're ignoring busy call slots, skip any ones that
1472 * have lastBusy set */
1473 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1474 leastBusy = conn->lastBusy[i];
1476 MUTEX_EXIT(&call->lock);
1481 * We are setting the state to RX_STATE_RESET to
1482 * ensure that no one else will attempt to use this
1483 * call once we drop the conn->conn_call_lock and
1484 * call->lock. We must drop the conn->conn_call_lock
1485 * before calling rxi_ResetCall because the process
1486 * of clearing the transmit queue can block for an
1487 * extended period of time. If we block while holding
1488 * the conn->conn_call_lock, then all rx_EndCall
1489 * processing will block as well. This has a detrimental
1490 * effect on overall system performance.
1492 call->state = RX_STATE_RESET;
1493 (*call->callNumber)++;
1494 MUTEX_EXIT(&conn->conn_call_lock);
1495 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1496 rxi_ResetCall(call, 0);
1497 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1501 * If we failed to be able to safely obtain the
1502 * conn->conn_call_lock we will have to drop the
1503 * call->lock to avoid a deadlock. When the call->lock
1504 * is released the state of the call can change. If it
1505 * is no longer RX_STATE_RESET then some other thread is
1508 MUTEX_EXIT(&call->lock);
1509 MUTEX_ENTER(&conn->conn_call_lock);
1510 MUTEX_ENTER(&call->lock);
1512 if (call->state == RX_STATE_RESET)
1516 * If we get here it means that after dropping
1517 * the conn->conn_call_lock and call->lock that
1518 * the call is no longer ours. If we can't find
1519 * a free call in the remaining slots we should
1520 * not go immediately to RX_CONN_MAKECALL_WAITING
1521 * because by dropping the conn->conn_call_lock
1522 * we have given up synchronization with rx_EndCall.
1523 * Instead, cycle through one more time to see if
1524 * we can find a call that can call our own.
1526 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1529 MUTEX_EXIT(&call->lock);
1532 if (ignoreBusy && conn->lastBusy[i]) {
1533 /* if we're ignoring busy call slots, skip any ones that
1534 * have lastBusy set */
1535 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1536 leastBusy = conn->lastBusy[i];
1541 /* rxi_NewCall returns with mutex locked */
1542 call = rxi_NewCall(conn, i);
1543 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1547 if (i < RX_MAXCALLS) {
1548 conn->lastBusy[i] = 0;
1553 if (leastBusy && ignoreBusy) {
1554 /* we didn't find a useable call slot, but we did see at least one
1555 * 'busy' slot; look again and only use a slot with the 'least
1561 MUTEX_ENTER(&conn->conn_data_lock);
1562 conn->flags |= RX_CONN_MAKECALL_WAITING;
1563 conn->makeCallWaiters++;
1564 MUTEX_EXIT(&conn->conn_data_lock);
1566 #ifdef RX_ENABLE_LOCKS
1567 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1571 MUTEX_ENTER(&conn->conn_data_lock);
1572 conn->makeCallWaiters--;
1573 if (conn->makeCallWaiters == 0)
1574 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1575 MUTEX_EXIT(&conn->conn_data_lock);
1577 /* Client is initially in send mode */
1578 call->state = RX_STATE_ACTIVE;
1579 call->error = conn->error;
1581 call->app.mode = RX_MODE_ERROR;
1583 call->app.mode = RX_MODE_SENDING;
1585 #ifdef AFS_RXERRQ_ENV
1586 /* remember how many network errors the peer has when we started, so if
1587 * more errors are encountered after the call starts, we know the other endpoint won't be
1588 * responding to us */
1589 call->neterr_gen = rx_atomic_read(&conn->peer->neterrs);
1592 /* remember start time for call in case we have hard dead time limit */
1593 call->queueTime = queueTime;
1594 clock_GetTime(&call->startTime);
1595 call->app.bytesSent = 0;
1596 call->app.bytesRcvd = 0;
1598 /* Turn on busy protocol. */
1599 rxi_KeepAliveOn(call);
1601 /* Attempt MTU discovery */
1602 rxi_GrowMTUOn(call);
1605 * We are no longer the active thread in rx_NewCall
1607 MUTEX_ENTER(&conn->conn_data_lock);
1608 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1609 MUTEX_EXIT(&conn->conn_data_lock);
1612 * Wake up anyone else who might be giving us a chance to
1613 * run (see code above that avoids resource starvation).
1615 #ifdef RX_ENABLE_LOCKS
1616 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1617 osi_Panic("rx_NewCall call about to be used without an empty tq");
1620 CV_BROADCAST(&conn->conn_call_cv);
1624 MUTEX_EXIT(&conn->conn_call_lock);
1625 MUTEX_EXIT(&call->lock);
1628 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1633 rxi_HasActiveCalls(struct rx_connection *aconn)
1636 struct rx_call *tcall;
1640 for (i = 0; i < RX_MAXCALLS; i++) {
1641 if ((tcall = aconn->call[i])) {
1642 if ((tcall->state == RX_STATE_ACTIVE)
1643 || (tcall->state == RX_STATE_PRECALL)) {
1654 rxi_GetCallNumberVector(struct rx_connection *aconn,
1655 afs_int32 * aint32s)
1658 struct rx_call *tcall;
1662 MUTEX_ENTER(&aconn->conn_call_lock);
1663 for (i = 0; i < RX_MAXCALLS; i++) {
1664 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1665 aint32s[i] = aconn->callNumber[i] + 1;
1667 aint32s[i] = aconn->callNumber[i];
1669 MUTEX_EXIT(&aconn->conn_call_lock);
1675 rxi_SetCallNumberVector(struct rx_connection *aconn,
1676 afs_int32 * aint32s)
1679 struct rx_call *tcall;
1683 MUTEX_ENTER(&aconn->conn_call_lock);
1684 for (i = 0; i < RX_MAXCALLS; i++) {
1685 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1686 aconn->callNumber[i] = aint32s[i] - 1;
1688 aconn->callNumber[i] = aint32s[i];
1690 MUTEX_EXIT(&aconn->conn_call_lock);
1695 /* Advertise a new service. A service is named locally by a UDP port
1696 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1699 char *serviceName; Name for identification purposes (e.g. the
1700 service name might be used for probing for
1703 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1704 char *serviceName, struct rx_securityClass **securityObjects,
1705 int nSecurityObjects,
1706 afs_int32(*serviceProc) (struct rx_call * acall))
1708 osi_socket socket = OSI_NULLSOCKET;
1709 struct rx_service *tservice;
1715 if (serviceId == 0) {
1717 "rx_NewService: service id for service %s is not non-zero.\n",
1724 "rx_NewService: A non-zero port must be specified on this call if a non-zero port was not provided at Rx initialization (service %s).\n",
1732 tservice = rxi_AllocService();
1735 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1737 for (i = 0; i < RX_MAX_SERVICES; i++) {
1738 struct rx_service *service = rx_services[i];
1740 if (port == service->servicePort && host == service->serviceHost) {
1741 if (service->serviceId == serviceId) {
1742 /* The identical service has already been
1743 * installed; if the caller was intending to
1744 * change the security classes used by this
1745 * service, he/she loses. */
1747 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1748 serviceName, serviceId, service->serviceName);
1750 rxi_FreeService(tservice);
1753 /* Different service, same port: re-use the socket
1754 * which is bound to the same port */
1755 socket = service->socket;
1758 if (socket == OSI_NULLSOCKET) {
1759 /* If we don't already have a socket (from another
1760 * service on same port) get a new one */
1761 socket = rxi_GetHostUDPSocket(host, port);
1762 if (socket == OSI_NULLSOCKET) {
1764 rxi_FreeService(tservice);
1769 service->socket = socket;
1770 service->serviceHost = host;
1771 service->servicePort = port;
1772 service->serviceId = serviceId;
1773 service->serviceName = serviceName;
1774 service->nSecurityObjects = nSecurityObjects;
1775 service->securityObjects = securityObjects;
1776 service->minProcs = 0;
1777 service->maxProcs = 1;
1778 service->idleDeadTime = 60;
1779 service->connDeadTime = rx_connDeadTime;
1780 service->executeRequestProc = serviceProc;
1781 service->checkReach = 0;
1782 service->nSpecific = 0;
1783 service->specific = NULL;
1784 rx_services[i] = service; /* not visible until now */
1790 rxi_FreeService(tservice);
1791 (osi_Msg "rx_NewService: cannot support > %d services\n",
1796 /* Set configuration options for all of a service's security objects */
1799 rx_SetSecurityConfiguration(struct rx_service *service,
1800 rx_securityConfigVariables type,
1804 for (i = 0; i<service->nSecurityObjects; i++) {
1805 if (service->securityObjects[i]) {
1806 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1814 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1815 struct rx_securityClass **securityObjects, int nSecurityObjects,
1816 afs_int32(*serviceProc) (struct rx_call * acall))
1818 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1821 /* Generic request processing loop. This routine should be called
1822 * by the implementation dependent rx_ServerProc. If socketp is
1823 * non-null, it will be set to the file descriptor that this thread
1824 * is now listening on. If socketp is null, this routine will never
1827 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1829 struct rx_call *call;
1831 struct rx_service *tservice = NULL;
1838 call = rx_GetCall(threadID, tservice, socketp);
1839 if (socketp && *socketp != OSI_NULLSOCKET) {
1840 /* We are now a listener thread */
1846 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1847 #ifdef RX_ENABLE_LOCKS
1849 #endif /* RX_ENABLE_LOCKS */
1850 afs_termState = AFSOP_STOP_AFS;
1851 afs_osi_Wakeup(&afs_termState);
1852 #ifdef RX_ENABLE_LOCKS
1854 #endif /* RX_ENABLE_LOCKS */
1859 /* if server is restarting( typically smooth shutdown) then do not
1860 * allow any new calls.
1863 if (rx_tranquil && (call != NULL)) {
1867 MUTEX_ENTER(&call->lock);
1869 rxi_CallError(call, RX_RESTARTING);
1870 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1872 MUTEX_EXIT(&call->lock);
1877 tservice = call->conn->service;
1879 if (tservice->beforeProc)
1880 (*tservice->beforeProc) (call);
1882 code = tservice->executeRequestProc(call);
1884 if (tservice->afterProc)
1885 (*tservice->afterProc) (call, code);
1887 rx_EndCall(call, code);
1889 if (tservice->postProc)
1890 (*tservice->postProc) (code);
1892 if (rx_stats_active) {
1893 MUTEX_ENTER(&rx_stats_mutex);
1895 MUTEX_EXIT(&rx_stats_mutex);
1902 rx_WakeupServerProcs(void)
1904 struct rx_serverQueueEntry *np, *tqp;
1905 struct opr_queue *cursor;
1909 MUTEX_ENTER(&rx_serverPool_lock);
1911 #ifdef RX_ENABLE_LOCKS
1912 if (rx_waitForPacket)
1913 CV_BROADCAST(&rx_waitForPacket->cv);
1914 #else /* RX_ENABLE_LOCKS */
1915 if (rx_waitForPacket)
1916 osi_rxWakeup(rx_waitForPacket);
1917 #endif /* RX_ENABLE_LOCKS */
1918 MUTEX_ENTER(&freeSQEList_lock);
1919 for (np = rx_FreeSQEList; np; np = tqp) {
1920 tqp = *(struct rx_serverQueueEntry **)np;
1921 #ifdef RX_ENABLE_LOCKS
1922 CV_BROADCAST(&np->cv);
1923 #else /* RX_ENABLE_LOCKS */
1925 #endif /* RX_ENABLE_LOCKS */
1927 MUTEX_EXIT(&freeSQEList_lock);
1928 for (opr_queue_Scan(&rx_idleServerQueue, cursor)) {
1929 np = opr_queue_Entry(cursor, struct rx_serverQueueEntry, entry);
1930 #ifdef RX_ENABLE_LOCKS
1931 CV_BROADCAST(&np->cv);
1932 #else /* RX_ENABLE_LOCKS */
1934 #endif /* RX_ENABLE_LOCKS */
1936 MUTEX_EXIT(&rx_serverPool_lock);
1941 * One thing that seems to happen is that all the server threads get
1942 * tied up on some empty or slow call, and then a whole bunch of calls
1943 * arrive at once, using up the packet pool, so now there are more
1944 * empty calls. The most critical resources here are server threads
1945 * and the free packet pool. The "doreclaim" code seems to help in
1946 * general. I think that eventually we arrive in this state: there
1947 * are lots of pending calls which do have all their packets present,
1948 * so they won't be reclaimed, are multi-packet calls, so they won't
1949 * be scheduled until later, and thus are tying up most of the free
1950 * packet pool for a very long time.
1952 * 1. schedule multi-packet calls if all the packets are present.
1953 * Probably CPU-bound operation, useful to return packets to pool.
1954 * Do what if there is a full window, but the last packet isn't here?
1955 * 3. preserve one thread which *only* runs "best" calls, otherwise
1956 * it sleeps and waits for that type of call.
1957 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1958 * the current dataquota business is badly broken. The quota isn't adjusted
1959 * to reflect how many packets are presently queued for a running call.
1960 * So, when we schedule a queued call with a full window of packets queued
1961 * up for it, that *should* free up a window full of packets for other 2d-class
1962 * calls to be able to use from the packet pool. But it doesn't.
1964 * NB. Most of the time, this code doesn't run -- since idle server threads
1965 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1966 * as a new call arrives.
1968 /* Sleep until a call arrives. Returns a pointer to the call, ready
1969 * for an rx_Read. */
1970 #ifdef RX_ENABLE_LOCKS
1972 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1974 struct rx_serverQueueEntry *sq;
1975 struct rx_call *call = (struct rx_call *)0;
1976 struct rx_service *service = NULL;
1978 MUTEX_ENTER(&freeSQEList_lock);
1980 if ((sq = rx_FreeSQEList)) {
1981 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1982 MUTEX_EXIT(&freeSQEList_lock);
1983 } else { /* otherwise allocate a new one and return that */
1984 MUTEX_EXIT(&freeSQEList_lock);
1985 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1986 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1987 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1990 MUTEX_ENTER(&rx_serverPool_lock);
1991 if (cur_service != NULL) {
1992 ReturnToServerPool(cur_service);
1995 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
1996 struct rx_call *tcall, *choice2 = NULL;
1997 struct opr_queue *cursor;
1999 /* Scan for eligible incoming calls. A call is not eligible
2000 * if the maximum number of calls for its service type are
2001 * already executing */
2002 /* One thread will process calls FCFS (to prevent starvation),
2003 * while the other threads may run ahead looking for calls which
2004 * have all their input data available immediately. This helps
2005 * keep threads from blocking, waiting for data from the client. */
2006 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2007 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2009 service = tcall->conn->service;
2010 if (!QuotaOK(service)) {
2013 MUTEX_ENTER(&rx_pthread_mutex);
2014 if (tno == rxi_fcfs_thread_num
2015 || opr_queue_IsEnd(&rx_incomingCallQueue, cursor)) {
2016 MUTEX_EXIT(&rx_pthread_mutex);
2017 /* If we're the fcfs thread , then we'll just use
2018 * this call. If we haven't been able to find an optimal
2019 * choice, and we're at the end of the list, then use a
2020 * 2d choice if one has been identified. Otherwise... */
2021 call = (choice2 ? choice2 : tcall);
2022 service = call->conn->service;
2024 MUTEX_EXIT(&rx_pthread_mutex);
2025 if (!opr_queue_IsEmpty(&tcall->rq)) {
2026 struct rx_packet *rp;
2027 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2029 if (rp->header.seq == 1) {
2031 || (rp->header.flags & RX_LAST_PACKET)) {
2033 } else if (rxi_2dchoice && !choice2
2034 && !(tcall->flags & RX_CALL_CLEARED)
2035 && (tcall->rprev > rxi_HardAckRate)) {
2045 ReturnToServerPool(service);
2051 opr_queue_Remove(&call->entry);
2052 MUTEX_EXIT(&rx_serverPool_lock);
2053 MUTEX_ENTER(&call->lock);
2055 if (call->flags & RX_CALL_WAIT_PROC) {
2056 call->flags &= ~RX_CALL_WAIT_PROC;
2057 rx_atomic_dec(&rx_nWaiting);
2060 if (call->state != RX_STATE_PRECALL || call->error) {
2061 MUTEX_EXIT(&call->lock);
2062 MUTEX_ENTER(&rx_serverPool_lock);
2063 ReturnToServerPool(service);
2068 if (opr_queue_IsEmpty(&call->rq)
2069 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1)
2070 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2072 CLEAR_CALL_QUEUE_LOCK(call);
2075 /* If there are no eligible incoming calls, add this process
2076 * to the idle server queue, to wait for one */
2080 *socketp = OSI_NULLSOCKET;
2082 sq->socketp = socketp;
2083 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2084 #ifndef AFS_AIX41_ENV
2085 rx_waitForPacket = sq;
2086 #endif /* AFS_AIX41_ENV */
2088 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2090 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2091 MUTEX_EXIT(&rx_serverPool_lock);
2092 return (struct rx_call *)0;
2095 } while (!(call = sq->newcall)
2096 && !(socketp && *socketp != OSI_NULLSOCKET));
2097 MUTEX_EXIT(&rx_serverPool_lock);
2099 MUTEX_ENTER(&call->lock);
2105 MUTEX_ENTER(&freeSQEList_lock);
2106 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2107 rx_FreeSQEList = sq;
2108 MUTEX_EXIT(&freeSQEList_lock);
2111 clock_GetTime(&call->startTime);
2112 call->state = RX_STATE_ACTIVE;
2113 call->app.mode = RX_MODE_RECEIVING;
2114 #ifdef RX_KERNEL_TRACE
2115 if (ICL_SETACTIVE(afs_iclSetp)) {
2116 int glockOwner = ISAFS_GLOCK();
2119 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2120 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2127 rxi_calltrace(RX_CALL_START, call);
2128 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2129 call->conn->service->servicePort, call->conn->service->serviceId,
2132 MUTEX_EXIT(&call->lock);
2133 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2135 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2140 #else /* RX_ENABLE_LOCKS */
2142 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2144 struct rx_serverQueueEntry *sq;
2145 struct rx_call *call = (struct rx_call *)0, *choice2;
2146 struct rx_service *service = NULL;
2150 MUTEX_ENTER(&freeSQEList_lock);
2152 if ((sq = rx_FreeSQEList)) {
2153 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2154 MUTEX_EXIT(&freeSQEList_lock);
2155 } else { /* otherwise allocate a new one and return that */
2156 MUTEX_EXIT(&freeSQEList_lock);
2157 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2158 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2159 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2161 MUTEX_ENTER(&sq->lock);
2163 if (cur_service != NULL) {
2164 cur_service->nRequestsRunning--;
2165 MUTEX_ENTER(&rx_quota_mutex);
2166 if (cur_service->nRequestsRunning < cur_service->minProcs)
2169 MUTEX_EXIT(&rx_quota_mutex);
2171 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2172 struct rx_call *tcall;
2173 struct opr_queue *cursor;
2174 /* Scan for eligible incoming calls. A call is not eligible
2175 * if the maximum number of calls for its service type are
2176 * already executing */
2177 /* One thread will process calls FCFS (to prevent starvation),
2178 * while the other threads may run ahead looking for calls which
2179 * have all their input data available immediately. This helps
2180 * keep threads from blocking, waiting for data from the client. */
2181 choice2 = (struct rx_call *)0;
2182 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2183 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2184 service = tcall->conn->service;
2185 if (QuotaOK(service)) {
2186 MUTEX_ENTER(&rx_pthread_mutex);
2187 /* XXX - If tcall->entry.next is NULL, then we're no longer
2188 * on a queue at all. This shouldn't happen. */
2189 if (tno == rxi_fcfs_thread_num || !tcall->entry.next) {
2190 MUTEX_EXIT(&rx_pthread_mutex);
2191 /* If we're the fcfs thread, then we'll just use
2192 * this call. If we haven't been able to find an optimal
2193 * choice, and we're at the end of the list, then use a
2194 * 2d choice if one has been identified. Otherwise... */
2195 call = (choice2 ? choice2 : tcall);
2196 service = call->conn->service;
2198 MUTEX_EXIT(&rx_pthread_mutex);
2199 if (!opr_queue_IsEmpty(&tcall->rq)) {
2200 struct rx_packet *rp;
2201 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2203 if (rp->header.seq == 1
2205 || (rp->header.flags & RX_LAST_PACKET))) {
2207 } else if (rxi_2dchoice && !choice2
2208 && !(tcall->flags & RX_CALL_CLEARED)
2209 && (tcall->rprev > rxi_HardAckRate)) {
2222 opr_queue_Remove(&call->entry);
2223 /* we can't schedule a call if there's no data!!! */
2224 /* send an ack if there's no data, if we're missing the
2225 * first packet, or we're missing something between first
2226 * and last -- there's a "hole" in the incoming data. */
2227 if (opr_queue_IsEmpty(&call->rq)
2228 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1
2229 || call->rprev != opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq)
2230 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2232 call->flags &= (~RX_CALL_WAIT_PROC);
2233 service->nRequestsRunning++;
2234 /* just started call in minProcs pool, need fewer to maintain
2236 MUTEX_ENTER(&rx_quota_mutex);
2237 if (service->nRequestsRunning <= service->minProcs)
2240 MUTEX_EXIT(&rx_quota_mutex);
2241 rx_atomic_dec(&rx_nWaiting);
2242 /* MUTEX_EXIT(&call->lock); */
2244 /* If there are no eligible incoming calls, add this process
2245 * to the idle server queue, to wait for one */
2248 *socketp = OSI_NULLSOCKET;
2250 sq->socketp = socketp;
2251 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2255 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2257 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2258 return (struct rx_call *)0;
2261 } while (!(call = sq->newcall)
2262 && !(socketp && *socketp != OSI_NULLSOCKET));
2264 MUTEX_EXIT(&sq->lock);
2266 MUTEX_ENTER(&freeSQEList_lock);
2267 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2268 rx_FreeSQEList = sq;
2269 MUTEX_EXIT(&freeSQEList_lock);
2272 clock_GetTime(&call->startTime);
2273 call->state = RX_STATE_ACTIVE;
2274 call->app.mode = RX_MODE_RECEIVING;
2275 #ifdef RX_KERNEL_TRACE
2276 if (ICL_SETACTIVE(afs_iclSetp)) {
2277 int glockOwner = ISAFS_GLOCK();
2280 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2281 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2288 rxi_calltrace(RX_CALL_START, call);
2289 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2290 call->conn->service->servicePort, call->conn->service->serviceId,
2293 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2300 #endif /* RX_ENABLE_LOCKS */
2304 /* Establish a procedure to be called when a packet arrives for a
2305 * call. This routine will be called at most once after each call,
2306 * and will also be called if there is an error condition on the or
2307 * the call is complete. Used by multi rx to build a selection
2308 * function which determines which of several calls is likely to be a
2309 * good one to read from.
2310 * NOTE: the way this is currently implemented it is probably only a
2311 * good idea to (1) use it immediately after a newcall (clients only)
2312 * and (2) only use it once. Other uses currently void your warranty
2315 rx_SetArrivalProc(struct rx_call *call,
2316 void (*proc) (struct rx_call * call,
2319 void * handle, int arg)
2321 call->arrivalProc = proc;
2322 call->arrivalProcHandle = handle;
2323 call->arrivalProcArg = arg;
2326 /* Call is finished (possibly prematurely). Return rc to the peer, if
2327 * appropriate, and return the final error code from the conversation
2331 rx_EndCall(struct rx_call *call, afs_int32 rc)
2333 struct rx_connection *conn = call->conn;
2337 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2338 call, rc, call->error, call->abortCode));
2341 MUTEX_ENTER(&call->lock);
2343 if (rc == 0 && call->error == 0) {
2344 call->abortCode = 0;
2345 call->abortCount = 0;
2348 call->arrivalProc = (void (*)())0;
2349 if (rc && call->error == 0) {
2350 rxi_CallError(call, rc);
2351 call->app.mode = RX_MODE_ERROR;
2352 /* Send an abort message to the peer if this error code has
2353 * only just been set. If it was set previously, assume the
2354 * peer has already been sent the error code or will request it
2356 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2358 if (conn->type == RX_SERVER_CONNECTION) {
2359 /* Make sure reply or at least dummy reply is sent */
2360 if (call->app.mode == RX_MODE_RECEIVING) {
2361 MUTEX_EXIT(&call->lock);
2362 rxi_WriteProc(call, 0, 0);
2363 MUTEX_ENTER(&call->lock);
2365 if (call->app.mode == RX_MODE_SENDING) {
2366 rxi_FlushWriteLocked(call);
2368 rxi_calltrace(RX_CALL_END, call);
2369 /* Call goes to hold state until reply packets are acknowledged */
2370 if (call->tfirst + call->nSoftAcked < call->tnext) {
2371 call->state = RX_STATE_HOLD;
2373 call->state = RX_STATE_DALLY;
2374 rxi_ClearTransmitQueue(call, 0);
2375 rxi_rto_cancel(call);
2376 rxi_CancelKeepAliveEvent(call);
2378 } else { /* Client connection */
2380 /* Make sure server receives input packets, in the case where
2381 * no reply arguments are expected */
2383 if ((call->app.mode == RX_MODE_SENDING)
2384 || (call->app.mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2385 MUTEX_EXIT(&call->lock);
2386 (void)rxi_ReadProc(call, &dummy, 1);
2387 MUTEX_ENTER(&call->lock);
2390 /* If we had an outstanding delayed ack, be nice to the server
2391 * and force-send it now.
2393 if (call->delayedAckEvent) {
2394 rxi_CancelDelayedAckEvent(call);
2395 rxi_SendDelayedAck(NULL, call, NULL, 0);
2398 /* We need to release the call lock since it's lower than the
2399 * conn_call_lock and we don't want to hold the conn_call_lock
2400 * over the rx_ReadProc call. The conn_call_lock needs to be held
2401 * here for the case where rx_NewCall is perusing the calls on
2402 * the connection structure. We don't want to signal until
2403 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2404 * have checked this call, found it active and by the time it
2405 * goes to sleep, will have missed the signal.
2407 MUTEX_EXIT(&call->lock);
2408 MUTEX_ENTER(&conn->conn_call_lock);
2409 MUTEX_ENTER(&call->lock);
2412 /* While there are some circumstances where a call with an error is
2413 * obviously not on a "busy" channel, be conservative (clearing
2414 * lastBusy is just best-effort to possibly speed up rx_NewCall).
2415 * The call channel is definitely not busy if we just successfully
2416 * completed a call on it. */
2417 conn->lastBusy[call->channel] = 0;
2419 } else if (call->error == RX_CALL_TIMEOUT) {
2420 /* The call is still probably running on the server side, so try to
2421 * avoid this call channel in the future. */
2422 conn->lastBusy[call->channel] = clock_Sec();
2425 MUTEX_ENTER(&conn->conn_data_lock);
2426 conn->flags |= RX_CONN_BUSY;
2427 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2428 MUTEX_EXIT(&conn->conn_data_lock);
2429 #ifdef RX_ENABLE_LOCKS
2430 CV_BROADCAST(&conn->conn_call_cv);
2435 #ifdef RX_ENABLE_LOCKS
2437 MUTEX_EXIT(&conn->conn_data_lock);
2439 #endif /* RX_ENABLE_LOCKS */
2440 call->state = RX_STATE_DALLY;
2442 error = call->error;
2444 /* currentPacket, nLeft, and NFree must be zeroed here, because
2445 * ResetCall cannot: ResetCall may be called at splnet(), in the
2446 * kernel version, and may interrupt the macros rx_Read or
2447 * rx_Write, which run at normal priority for efficiency. */
2448 if (call->app.currentPacket) {
2449 #ifdef RX_TRACK_PACKETS
2450 call->app.currentPacket->flags &= ~RX_PKTFLAG_CP;
2452 rxi_FreePacket(call->app.currentPacket);
2453 call->app.currentPacket = (struct rx_packet *)0;
2456 call->app.nLeft = call->app.nFree = call->app.curlen = 0;
2458 /* Free any packets from the last call to ReadvProc/WritevProc */
2459 #ifdef RXDEBUG_PACKET
2461 #endif /* RXDEBUG_PACKET */
2462 rxi_FreePackets(0, &call->app.iovq);
2463 MUTEX_EXIT(&call->lock);
2465 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2466 if (conn->type == RX_CLIENT_CONNECTION) {
2467 MUTEX_ENTER(&conn->conn_data_lock);
2468 conn->flags &= ~RX_CONN_BUSY;
2469 MUTEX_EXIT(&conn->conn_data_lock);
2470 MUTEX_EXIT(&conn->conn_call_lock);
2474 * Map errors to the local host's errno.h format.
2476 error = ntoh_syserr_conv(error);
2478 /* If the caller said the call failed with some error, we had better
2479 * return an error code. */
2480 osi_Assert(!rc || error);
2484 #if !defined(KERNEL)
2486 /* Call this routine when shutting down a server or client (especially
2487 * clients). This will allow Rx to gracefully garbage collect server
2488 * connections, and reduce the number of retries that a server might
2489 * make to a dead client.
2490 * This is not quite right, since some calls may still be ongoing and
2491 * we can't lock them to destroy them. */
2495 struct rx_connection **conn_ptr, **conn_end;
2498 if (rx_atomic_test_and_set_bit(&rxinit_status, 0))
2499 return; /* Already shutdown. */
2501 rxi_DeleteCachedConnections();
2502 if (rx_connHashTable) {
2503 MUTEX_ENTER(&rx_connHashTable_lock);
2504 for (conn_ptr = &rx_connHashTable[0], conn_end =
2505 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2507 struct rx_connection *conn, *next;
2508 for (conn = *conn_ptr; conn; conn = next) {
2510 if (conn->type == RX_CLIENT_CONNECTION) {
2511 MUTEX_ENTER(&rx_refcnt_mutex);
2513 MUTEX_EXIT(&rx_refcnt_mutex);
2514 #ifdef RX_ENABLE_LOCKS
2515 rxi_DestroyConnectionNoLock(conn);
2516 #else /* RX_ENABLE_LOCKS */
2517 rxi_DestroyConnection(conn);
2518 #endif /* RX_ENABLE_LOCKS */
2522 #ifdef RX_ENABLE_LOCKS
2523 while (rx_connCleanup_list) {
2524 struct rx_connection *conn;
2525 conn = rx_connCleanup_list;
2526 rx_connCleanup_list = rx_connCleanup_list->next;
2527 MUTEX_EXIT(&rx_connHashTable_lock);
2528 rxi_CleanupConnection(conn);
2529 MUTEX_ENTER(&rx_connHashTable_lock);
2531 MUTEX_EXIT(&rx_connHashTable_lock);
2532 #endif /* RX_ENABLE_LOCKS */
2537 afs_winsockCleanup();
2543 /* if we wakeup packet waiter too often, can get in loop with two
2544 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2546 rxi_PacketsUnWait(void)
2548 if (!rx_waitingForPackets) {
2552 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2553 return; /* still over quota */
2556 rx_waitingForPackets = 0;
2557 #ifdef RX_ENABLE_LOCKS
2558 CV_BROADCAST(&rx_waitingForPackets_cv);
2560 osi_rxWakeup(&rx_waitingForPackets);
2566 /* ------------------Internal interfaces------------------------- */
2568 /* Return this process's service structure for the
2569 * specified socket and service */
2570 static struct rx_service *
2571 rxi_FindService(osi_socket socket, u_short serviceId)
2573 struct rx_service **sp;
2574 for (sp = &rx_services[0]; *sp; sp++) {
2575 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2581 #ifdef RXDEBUG_PACKET
2582 #ifdef KDUMP_RX_LOCK
2583 static struct rx_call_rx_lock *rx_allCallsp = 0;
2585 static struct rx_call *rx_allCallsp = 0;
2587 #endif /* RXDEBUG_PACKET */
2589 /* Allocate a call structure, for the indicated channel of the
2590 * supplied connection. The mode and state of the call must be set by
2591 * the caller. Returns the call with mutex locked. */
2592 static struct rx_call *
2593 rxi_NewCall(struct rx_connection *conn, int channel)
2595 struct rx_call *call;
2596 #ifdef RX_ENABLE_LOCKS
2597 struct rx_call *cp; /* Call pointer temp */
2598 struct opr_queue *cursor;
2601 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2603 /* Grab an existing call structure, or allocate a new one.
2604 * Existing call structures are assumed to have been left reset by
2606 MUTEX_ENTER(&rx_freeCallQueue_lock);
2608 #ifdef RX_ENABLE_LOCKS
2610 * EXCEPT that the TQ might not yet be cleared out.
2611 * Skip over those with in-use TQs.
2614 for (opr_queue_Scan(&rx_freeCallQueue, cursor)) {
2615 cp = opr_queue_Entry(cursor, struct rx_call, entry);
2616 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2622 #else /* RX_ENABLE_LOCKS */
2623 if (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
2624 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
2625 #endif /* RX_ENABLE_LOCKS */
2626 opr_queue_Remove(&call->entry);
2627 if (rx_stats_active)
2628 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2629 MUTEX_EXIT(&rx_freeCallQueue_lock);
2630 MUTEX_ENTER(&call->lock);
2631 CLEAR_CALL_QUEUE_LOCK(call);
2632 #ifdef RX_ENABLE_LOCKS
2633 /* Now, if TQ wasn't cleared earlier, do it now. */
2634 rxi_WaitforTQBusy(call);
2635 if (call->flags & RX_CALL_TQ_CLEARME) {
2636 rxi_ClearTransmitQueue(call, 1);
2637 /*queue_Init(&call->tq);*/
2639 #endif /* RX_ENABLE_LOCKS */
2640 /* Bind the call to its connection structure */
2642 rxi_ResetCall(call, 1);
2645 call = rxi_Alloc(sizeof(struct rx_call));
2646 #ifdef RXDEBUG_PACKET
2647 call->allNextp = rx_allCallsp;
2648 rx_allCallsp = call;
2650 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2651 #else /* RXDEBUG_PACKET */
2652 rx_atomic_inc(&rx_stats.nCallStructs);
2653 #endif /* RXDEBUG_PACKET */
2655 MUTEX_EXIT(&rx_freeCallQueue_lock);
2656 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2657 MUTEX_ENTER(&call->lock);
2658 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2659 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2660 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2662 /* Initialize once-only items */
2663 opr_queue_Init(&call->tq);
2664 opr_queue_Init(&call->rq);
2665 opr_queue_Init(&call->app.iovq);
2666 #ifdef RXDEBUG_PACKET
2667 call->rqc = call->tqc = call->iovqc = 0;
2668 #endif /* RXDEBUG_PACKET */
2669 /* Bind the call to its connection structure (prereq for reset) */
2671 rxi_ResetCall(call, 1);
2673 call->channel = channel;
2674 call->callNumber = &conn->callNumber[channel];
2675 call->rwind = conn->rwind[channel];
2676 call->twind = conn->twind[channel];
2677 /* Note that the next expected call number is retained (in
2678 * conn->callNumber[i]), even if we reallocate the call structure
2680 conn->call[channel] = call;
2681 /* if the channel's never been used (== 0), we should start at 1, otherwise
2682 * the call number is valid from the last time this channel was used */
2683 if (*call->callNumber == 0)
2684 *call->callNumber = 1;
2689 /* A call has been inactive long enough that so we can throw away
2690 * state, including the call structure, which is placed on the call
2693 * call->lock amd rx_refcnt_mutex are held upon entry.
2694 * haveCTLock is set when called from rxi_ReapConnections.
2696 * return 1 if the call is freed, 0 if not.
2699 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2701 int channel = call->channel;
2702 struct rx_connection *conn = call->conn;
2703 u_char state = call->state;
2706 * We are setting the state to RX_STATE_RESET to
2707 * ensure that no one else will attempt to use this
2708 * call once we drop the refcnt lock. We must drop
2709 * the refcnt lock before calling rxi_ResetCall
2710 * because it cannot be held across acquiring the
2711 * freepktQ lock. NewCall does the same.
2713 call->state = RX_STATE_RESET;
2714 MUTEX_EXIT(&rx_refcnt_mutex);
2715 rxi_ResetCall(call, 0);
2717 if (MUTEX_TRYENTER(&conn->conn_call_lock))
2719 if (state == RX_STATE_DALLY || state == RX_STATE_HOLD)
2720 (*call->callNumber)++;
2722 if (call->conn->call[channel] == call)
2723 call->conn->call[channel] = 0;
2724 MUTEX_EXIT(&conn->conn_call_lock);
2727 * We couldn't obtain the conn_call_lock so we can't
2728 * disconnect the call from the connection. Set the
2729 * call state to dally so that the call can be reused.
2731 MUTEX_ENTER(&rx_refcnt_mutex);
2732 call->state = RX_STATE_DALLY;
2736 MUTEX_ENTER(&rx_freeCallQueue_lock);
2737 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2738 #ifdef RX_ENABLE_LOCKS
2739 /* A call may be free even though its transmit queue is still in use.
2740 * Since we search the call list from head to tail, put busy calls at
2741 * the head of the list, and idle calls at the tail.
2743 if (call->flags & RX_CALL_TQ_BUSY)
2744 opr_queue_Prepend(&rx_freeCallQueue, &call->entry);
2746 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2747 #else /* RX_ENABLE_LOCKS */
2748 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2749 #endif /* RX_ENABLE_LOCKS */
2750 if (rx_stats_active)
2751 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2752 MUTEX_EXIT(&rx_freeCallQueue_lock);
2754 /* Destroy the connection if it was previously slated for
2755 * destruction, i.e. the Rx client code previously called
2756 * rx_DestroyConnection (client connections), or
2757 * rxi_ReapConnections called the same routine (server
2758 * connections). Only do this, however, if there are no
2759 * outstanding calls. Note that for fine grain locking, there appears
2760 * to be a deadlock in that rxi_FreeCall has a call locked and
2761 * DestroyConnectionNoLock locks each call in the conn. But note a
2762 * few lines up where we have removed this call from the conn.
2763 * If someone else destroys a connection, they either have no
2764 * call lock held or are going through this section of code.
2766 MUTEX_ENTER(&conn->conn_data_lock);
2767 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2768 MUTEX_ENTER(&rx_refcnt_mutex);
2770 MUTEX_EXIT(&rx_refcnt_mutex);
2771 MUTEX_EXIT(&conn->conn_data_lock);
2772 #ifdef RX_ENABLE_LOCKS
2774 rxi_DestroyConnectionNoLock(conn);
2776 rxi_DestroyConnection(conn);
2777 #else /* RX_ENABLE_LOCKS */
2778 rxi_DestroyConnection(conn);
2779 #endif /* RX_ENABLE_LOCKS */
2781 MUTEX_EXIT(&conn->conn_data_lock);
2783 MUTEX_ENTER(&rx_refcnt_mutex);
2787 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2788 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2791 rxi_Alloc(size_t size)
2795 if (rx_stats_active) {
2796 rx_atomic_add(&rxi_Allocsize, (int) size);
2797 rx_atomic_inc(&rxi_Alloccnt);
2801 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2802 afs_osi_Alloc_NoSleep(size);
2807 osi_Panic("rxi_Alloc error");
2813 rxi_Free(void *addr, size_t size)
2815 if (rx_stats_active) {
2816 rx_atomic_sub(&rxi_Allocsize, (int) size);
2817 rx_atomic_dec(&rxi_Alloccnt);
2819 osi_Free(addr, size);
2823 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2825 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2826 struct rx_peer *next = NULL;
2830 MUTEX_ENTER(&rx_peerHashTable_lock);
2832 peer_ptr = &rx_peerHashTable[0];
2833 peer_end = &rx_peerHashTable[rx_hashTableSize];
2836 for ( ; peer_ptr < peer_end; peer_ptr++) {
2839 for ( ; peer; peer = next) {
2841 if (host == peer->host)
2846 hashIndex = PEER_HASH(host, port);
2847 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2848 if ((peer->host == host) && (peer->port == port))
2853 MUTEX_ENTER(&rx_peerHashTable_lock);
2858 MUTEX_EXIT(&rx_peerHashTable_lock);
2860 MUTEX_ENTER(&peer->peer_lock);
2861 /* We don't handle dropping below min, so don't */
2862 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2863 peer->ifMTU=MIN(mtu, peer->ifMTU);
2864 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2865 /* if we tweaked this down, need to tune our peer MTU too */
2866 peer->MTU = MIN(peer->MTU, peer->natMTU);
2867 /* if we discovered a sub-1500 mtu, degrade */
2868 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2869 peer->maxDgramPackets = 1;
2870 /* We no longer have valid peer packet information */
2871 if (peer->maxPacketSize + RX_HEADER_SIZE > peer->ifMTU)
2872 peer->maxPacketSize = 0;
2873 MUTEX_EXIT(&peer->peer_lock);
2875 MUTEX_ENTER(&rx_peerHashTable_lock);
2877 if (host && !port) {
2879 /* pick up where we left off */
2883 MUTEX_EXIT(&rx_peerHashTable_lock);
2886 #ifdef AFS_RXERRQ_ENV
2888 rxi_SetPeerDead(struct sock_extended_err *err, afs_uint32 host, afs_uint16 port)
2890 int hashIndex = PEER_HASH(host, port);
2891 struct rx_peer *peer;
2893 MUTEX_ENTER(&rx_peerHashTable_lock);
2895 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2896 if (peer->host == host && peer->port == port) {
2902 MUTEX_EXIT(&rx_peerHashTable_lock);
2905 rx_atomic_inc(&peer->neterrs);
2906 MUTEX_ENTER(&peer->peer_lock);
2907 peer->last_err_origin = RX_NETWORK_ERROR_ORIGIN_ICMP;
2908 peer->last_err_type = err->ee_type;
2909 peer->last_err_code = err->ee_code;
2910 MUTEX_EXIT(&peer->peer_lock);
2912 MUTEX_ENTER(&rx_peerHashTable_lock);
2914 MUTEX_EXIT(&rx_peerHashTable_lock);
2919 rxi_ProcessNetError(struct sock_extended_err *err, afs_uint32 addr, afs_uint16 port)
2921 # ifdef AFS_ADAPT_PMTU
2922 if (err->ee_errno == EMSGSIZE && err->ee_info >= 68) {
2923 rxi_SetPeerMtu(NULL, addr, port, err->ee_info - RX_IPUDP_SIZE);
2927 if (err->ee_origin == SO_EE_ORIGIN_ICMP && err->ee_type == ICMP_DEST_UNREACH) {
2928 switch (err->ee_code) {
2929 case ICMP_NET_UNREACH:
2930 case ICMP_HOST_UNREACH:
2931 case ICMP_PORT_UNREACH:
2934 rxi_SetPeerDead(err, addr, port);
2941 rxi_TranslateICMP(int type, int code)
2944 case ICMP_DEST_UNREACH:
2946 case ICMP_NET_UNREACH:
2947 return "Destination Net Unreachable";
2948 case ICMP_HOST_UNREACH:
2949 return "Destination Host Unreachable";
2950 case ICMP_PROT_UNREACH:
2951 return "Destination Protocol Unreachable";
2952 case ICMP_PORT_UNREACH:
2953 return "Destination Port Unreachable";
2955 return "Destination Net Prohibited";
2957 return "Destination Host Prohibited";
2963 #endif /* AFS_RXERRQ_ENV */
2966 * Get the last network error for a connection
2968 * A "network error" here means an error retrieved from ICMP, or some other
2969 * mechanism outside of Rx that informs us of errors in network reachability.
2971 * If a peer associated with the given Rx connection has received a network
2972 * error recently, this function allows the caller to know what error
2973 * specifically occurred. This can be useful to know, since e.g. ICMP errors
2974 * can cause calls to that peer to be quickly aborted. So, this function can
2975 * help see why a call was aborted due to network errors.
2977 * If we have received traffic from a peer since the last network error, we
2978 * treat that peer as if we had not received an network error for it.
2980 * @param[in] conn The Rx connection to examine
2981 * @param[out] err_origin The origin of the last network error (e.g. ICMP);
2982 * one of the RX_NETWORK_ERROR_ORIGIN_* constants
2983 * @param[out] err_type The type of the last error
2984 * @param[out] err_code The code of the last error
2985 * @param[out] msg Human-readable error message, if applicable; NULL otherwise
2987 * @return If we have an error
2988 * @retval -1 No error to get; 'out' params are undefined
2989 * @retval 0 We have an error; 'out' params contain the last error
2992 rx_GetNetworkError(struct rx_connection *conn, int *err_origin, int *err_type,
2993 int *err_code, const char **msg)
2995 #ifdef AFS_RXERRQ_ENV
2996 struct rx_peer *peer = conn->peer;
2997 if (rx_atomic_read(&peer->neterrs)) {
2998 MUTEX_ENTER(&peer->peer_lock);
2999 *err_origin = peer->last_err_origin;
3000 *err_type = peer->last_err_type;
3001 *err_code = peer->last_err_code;
3002 MUTEX_EXIT(&peer->peer_lock);
3005 if (*err_origin == RX_NETWORK_ERROR_ORIGIN_ICMP) {
3006 *msg = rxi_TranslateICMP(*err_type, *err_code);
3015 /* Find the peer process represented by the supplied (host,port)
3016 * combination. If there is no appropriate active peer structure, a
3017 * new one will be allocated and initialized
3020 rxi_FindPeer(afs_uint32 host, u_short port, int create)
3024 hashIndex = PEER_HASH(host, port);
3025 MUTEX_ENTER(&rx_peerHashTable_lock);
3026 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
3027 if ((pp->host == host) && (pp->port == port))
3032 pp = rxi_AllocPeer(); /* This bzero's *pp */
3033 pp->host = host; /* set here or in InitPeerParams is zero */
3035 #ifdef AFS_RXERRQ_ENV
3036 rx_atomic_set(&pp->neterrs, 0);
3038 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
3039 opr_queue_Init(&pp->rpcStats);
3040 pp->next = rx_peerHashTable[hashIndex];
3041 rx_peerHashTable[hashIndex] = pp;
3042 rxi_InitPeerParams(pp);
3043 if (rx_stats_active)
3044 rx_atomic_inc(&rx_stats.nPeerStructs);
3050 MUTEX_EXIT(&rx_peerHashTable_lock);
3055 /* Find the connection at (host, port) started at epoch, and with the
3056 * given connection id. Creates the server connection if necessary.
3057 * The type specifies whether a client connection or a server
3058 * connection is desired. In both cases, (host, port) specify the
3059 * peer's (host, pair) pair. Client connections are not made
3060 * automatically by this routine. The parameter socket gives the
3061 * socket descriptor on which the packet was received. This is used,
3062 * in the case of server connections, to check that *new* connections
3063 * come via a valid (port, serviceId). Finally, the securityIndex
3064 * parameter must match the existing index for the connection. If a
3065 * server connection is created, it will be created using the supplied
3066 * index, if the index is valid for this service */
3067 static struct rx_connection *
3068 rxi_FindConnection(osi_socket socket, afs_uint32 host,
3069 u_short port, u_short serviceId, afs_uint32 cid,
3070 afs_uint32 epoch, int type, u_int securityIndex,
3071 int *unknownService)
3073 int hashindex, flag, i;
3074 struct rx_connection *conn;
3075 *unknownService = 0;
3076 hashindex = CONN_HASH(host, port, cid, epoch, type);
3077 MUTEX_ENTER(&rx_connHashTable_lock);
3078 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
3079 rx_connHashTable[hashindex],
3082 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
3083 && (epoch == conn->epoch)) {
3084 struct rx_peer *pp = conn->peer;
3085 if (securityIndex != conn->securityIndex) {
3086 /* this isn't supposed to happen, but someone could forge a packet
3087 * like this, and there seems to be some CM bug that makes this
3088 * happen from time to time -- in which case, the fileserver
3090 MUTEX_EXIT(&rx_connHashTable_lock);
3091 return (struct rx_connection *)0;
3093 if (pp->host == host && pp->port == port)
3095 if (type == RX_CLIENT_CONNECTION && pp->port == port)
3097 /* So what happens when it's a callback connection? */
3098 if ( /*type == RX_CLIENT_CONNECTION && */
3099 (conn->epoch & 0x80000000))
3103 /* the connection rxLastConn that was used the last time is not the
3104 ** one we are looking for now. Hence, start searching in the hash */
3106 conn = rx_connHashTable[hashindex];
3111 struct rx_service *service;
3112 if (type == RX_CLIENT_CONNECTION) {
3113 MUTEX_EXIT(&rx_connHashTable_lock);
3114 return (struct rx_connection *)0;
3116 service = rxi_FindService(socket, serviceId);
3117 if (!service || (securityIndex >= service->nSecurityObjects)
3118 || (service->securityObjects[securityIndex] == 0)) {
3119 MUTEX_EXIT(&rx_connHashTable_lock);
3120 *unknownService = 1;
3121 return (struct rx_connection *)0;
3123 conn = rxi_AllocConnection(); /* This bzero's the connection */
3124 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
3125 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
3126 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
3127 conn->next = rx_connHashTable[hashindex];
3128 rx_connHashTable[hashindex] = conn;
3129 conn->peer = rxi_FindPeer(host, port, 1);
3130 conn->type = RX_SERVER_CONNECTION;
3131 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
3132 conn->epoch = epoch;
3133 conn->cid = cid & RX_CIDMASK;
3134 conn->ackRate = RX_FAST_ACK_RATE;
3135 conn->service = service;
3136 conn->serviceId = serviceId;
3137 conn->securityIndex = securityIndex;
3138 conn->securityObject = service->securityObjects[securityIndex];
3139 conn->nSpecific = 0;
3140 conn->specific = NULL;
3141 rx_SetConnDeadTime(conn, service->connDeadTime);
3142 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
3143 for (i = 0; i < RX_MAXCALLS; i++) {
3144 conn->twind[i] = rx_initSendWindow;
3145 conn->rwind[i] = rx_initReceiveWindow;
3147 /* Notify security object of the new connection */
3148 RXS_NewConnection(conn->securityObject, conn);
3149 /* XXXX Connection timeout? */
3150 if (service->newConnProc)
3151 (*service->newConnProc) (conn);
3152 if (rx_stats_active)
3153 rx_atomic_inc(&rx_stats.nServerConns);
3156 MUTEX_ENTER(&rx_refcnt_mutex);
3158 MUTEX_EXIT(&rx_refcnt_mutex);
3160 rxLastConn = conn; /* store this connection as the last conn used */
3161 MUTEX_EXIT(&rx_connHashTable_lock);
3166 * Abort the call if the server is over the busy threshold. This
3167 * can be used without requiring a call structure be initialised,
3168 * or connected to a particular channel
3171 rxi_AbortIfServerBusy(osi_socket socket, struct rx_connection *conn,
3172 struct rx_packet *np)
3174 if ((rx_BusyThreshold > 0) &&
3175 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3176 rxi_SendRawAbort(socket, conn->peer->host, conn->peer->port,
3177 rx_BusyError, np, 0);
3178 if (rx_stats_active)
3179 rx_atomic_inc(&rx_stats.nBusies);
3186 static_inline struct rx_call *
3187 rxi_ReceiveClientCall(struct rx_packet *np, struct rx_connection *conn)
3190 struct rx_call *call;
3192 channel = np->header.cid & RX_CHANNELMASK;
3193 MUTEX_ENTER(&conn->conn_call_lock);
3194 call = conn->call[channel];
3195 if (np->header.type == RX_PACKET_TYPE_BUSY) {
3196 conn->lastBusy[channel] = clock_Sec();
3198 if (!call || conn->callNumber[channel] != np->header.callNumber) {
3199 MUTEX_EXIT(&conn->conn_call_lock);
3200 if (rx_stats_active)
3201 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3205 MUTEX_ENTER(&call->lock);
3206 MUTEX_EXIT(&conn->conn_call_lock);
3208 if ((call->state == RX_STATE_DALLY)
3209 && np->header.type == RX_PACKET_TYPE_ACK) {
3210 if (rx_stats_active)
3211 rx_atomic_inc(&rx_stats.ignorePacketDally);
3212 MUTEX_EXIT(&call->lock);
3219 static_inline struct rx_call *
3220 rxi_ReceiveServerCall(osi_socket socket, struct rx_packet *np,
3221 struct rx_connection *conn)
3224 struct rx_call *call;
3226 channel = np->header.cid & RX_CHANNELMASK;
3227 MUTEX_ENTER(&conn->conn_call_lock);
3228 call = conn->call[channel];
3231 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3232 MUTEX_EXIT(&conn->conn_call_lock);
3236 call = rxi_NewCall(conn, channel); /* returns locked call */
3237 *call->callNumber = np->header.callNumber;
3238 MUTEX_EXIT(&conn->conn_call_lock);
3240 call->state = RX_STATE_PRECALL;
3241 clock_GetTime(&call->queueTime);
3242 call->app.bytesSent = 0;
3243 call->app.bytesRcvd = 0;
3244 rxi_KeepAliveOn(call);
3249 if (np->header.callNumber == conn->callNumber[channel]) {
3250 MUTEX_ENTER(&call->lock);
3251 MUTEX_EXIT(&conn->conn_call_lock);
3255 if (np->header.callNumber < conn->callNumber[channel]) {
3256 MUTEX_EXIT(&conn->conn_call_lock);
3257 if (rx_stats_active)
3258 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3262 MUTEX_ENTER(&call->lock);
3263 MUTEX_EXIT(&conn->conn_call_lock);
3265 /* Wait until the transmit queue is idle before deciding
3266 * whether to reset the current call. Chances are that the
3267 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3270 #ifdef RX_ENABLE_LOCKS
3271 if (call->state == RX_STATE_ACTIVE && !call->error) {
3272 rxi_WaitforTQBusy(call);
3273 /* If we entered error state while waiting,
3274 * must call rxi_CallError to permit rxi_ResetCall
3275 * to processed when the tqWaiter count hits zero.
3278 rxi_CallError(call, call->error);
3279 MUTEX_EXIT(&call->lock);
3283 #endif /* RX_ENABLE_LOCKS */
3284 /* If the new call cannot be taken right now send a busy and set
3285 * the error condition in this call, so that it terminates as
3286 * quickly as possible */
3287 if (call->state == RX_STATE_ACTIVE) {
3288 rxi_CallError(call, RX_CALL_DEAD);
3289 rxi_SendSpecial(call, conn, NULL, RX_PACKET_TYPE_BUSY,
3291 MUTEX_EXIT(&call->lock);
3295 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3296 MUTEX_EXIT(&call->lock);
3300 rxi_ResetCall(call, 0);
3301 /* The conn_call_lock is not held but no one else should be
3302 * using this call channel while we are processing this incoming
3303 * packet. This assignment should be safe.
3305 *call->callNumber = np->header.callNumber;
3306 call->state = RX_STATE_PRECALL;
3307 clock_GetTime(&call->queueTime);
3308 call->app.bytesSent = 0;
3309 call->app.bytesRcvd = 0;
3310 rxi_KeepAliveOn(call);
3316 /* There are two packet tracing routines available for testing and monitoring
3317 * Rx. One is called just after every packet is received and the other is
3318 * called just before every packet is sent. Received packets, have had their
3319 * headers decoded, and packets to be sent have not yet had their headers
3320 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3321 * containing the network address. Both can be modified. The return value, if
3322 * non-zero, indicates that the packet should be dropped. */
3324 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3325 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3327 /* A packet has been received off the interface. Np is the packet, socket is
3328 * the socket number it was received from (useful in determining which service
3329 * this packet corresponds to), and (host, port) reflect the host,port of the
3330 * sender. This call returns the packet to the caller if it is finished with
3331 * it, rather than de-allocating it, just as a small performance hack */
3334 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3335 afs_uint32 host, u_short port, int *tnop,
3336 struct rx_call **newcallp)
3338 struct rx_call *call;
3339 struct rx_connection *conn;
3341 int unknownService = 0;
3345 struct rx_packet *tnp;
3348 /* We don't print out the packet until now because (1) the time may not be
3349 * accurate enough until now in the lwp implementation (rx_Listener only gets
3350 * the time after the packet is read) and (2) from a protocol point of view,
3351 * this is the first time the packet has been seen */
3352 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3353 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3354 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3355 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3356 np->header.epoch, np->header.cid, np->header.callNumber,
3357 np->header.seq, np->header.flags, np));
3360 /* Account for connectionless packets */
3361 if (rx_stats_active &&
3362 ((np->header.type == RX_PACKET_TYPE_VERSION) ||
3363 (np->header.type == RX_PACKET_TYPE_DEBUG))) {
3364 struct rx_peer *peer;
3366 /* Try to look up the peer structure, but don't create one */
3367 peer = rxi_FindPeer(host, port, 0);
3369 /* Since this may not be associated with a connection, it may have
3370 * no refCount, meaning we could race with ReapConnections
3373 if (peer && (peer->refCount > 0)) {
3374 #ifdef AFS_RXERRQ_ENV
3375 if (rx_atomic_read(&peer->neterrs)) {
3376 rx_atomic_set(&peer->neterrs, 0);
3379 MUTEX_ENTER(&peer->peer_lock);
3380 peer->bytesReceived += np->length;
3381 MUTEX_EXIT(&peer->peer_lock);
3385 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3386 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3389 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3390 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3393 /* If an input tracer function is defined, call it with the packet and
3394 * network address. Note this function may modify its arguments. */
3395 if (rx_justReceived) {
3396 struct sockaddr_in addr;
3398 addr.sin_family = AF_INET;
3399 addr.sin_port = port;
3400 addr.sin_addr.s_addr = host;
3401 memset(&addr.sin_zero, 0, sizeof(addr.sin_zero));
3402 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3403 addr.sin_len = sizeof(addr);
3404 #endif /* AFS_OSF_ENV */
3405 drop = (*rx_justReceived) (np, &addr);
3406 /* drop packet if return value is non-zero */
3409 port = addr.sin_port; /* in case fcn changed addr */
3410 host = addr.sin_addr.s_addr;
3414 /* If packet was not sent by the client, then *we* must be the client */
3415 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3416 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3418 /* Find the connection (or fabricate one, if we're the server & if
3419 * necessary) associated with this packet */
3421 rxi_FindConnection(socket, host, port, np->header.serviceId,
3422 np->header.cid, np->header.epoch, type,
3423 np->header.securityIndex, &unknownService);
3425 /* To avoid having 2 connections just abort at each other,
3426 don't abort an abort. */
3428 if (unknownService && (np->header.type != RX_PACKET_TYPE_ABORT))
3429 rxi_SendRawAbort(socket, host, port, RX_INVALID_OPERATION,
3434 #ifdef AFS_RXERRQ_ENV
3435 if (rx_atomic_read(&conn->peer->neterrs)) {
3436 rx_atomic_set(&conn->peer->neterrs, 0);
3440 /* If we're doing statistics, then account for the incoming packet */
3441 if (rx_stats_active) {
3442 MUTEX_ENTER(&conn->peer->peer_lock);
3443 conn->peer->bytesReceived += np->length;
3444 MUTEX_EXIT(&conn->peer->peer_lock);
3447 /* If the connection is in an error state, send an abort packet and ignore
3448 * the incoming packet */
3450 /* Don't respond to an abort packet--we don't want loops! */
3451 MUTEX_ENTER(&conn->conn_data_lock);
3452 if (np->header.type != RX_PACKET_TYPE_ABORT)
3453 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3454 putConnection(conn);
3455 MUTEX_EXIT(&conn->conn_data_lock);
3459 /* Check for connection-only requests (i.e. not call specific). */
3460 if (np->header.callNumber == 0) {
3461 switch (np->header.type) {
3462 case RX_PACKET_TYPE_ABORT: {
3463 /* What if the supplied error is zero? */
3464 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3465 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3466 rxi_ConnectionError(conn, errcode);
3467 putConnection(conn);
3470 case RX_PACKET_TYPE_CHALLENGE:
3471 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3472 putConnection(conn);
3474 case RX_PACKET_TYPE_RESPONSE:
3475 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3476 putConnection(conn);
3478 case RX_PACKET_TYPE_PARAMS:
3479 case RX_PACKET_TYPE_PARAMS + 1:
3480 case RX_PACKET_TYPE_PARAMS + 2:
3481 /* ignore these packet types for now */
3482 putConnection(conn);
3486 /* Should not reach here, unless the peer is broken: send an
3488 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3489 MUTEX_ENTER(&conn->conn_data_lock);
3490 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3491 putConnection(conn);
3492 MUTEX_EXIT(&conn->conn_data_lock);
3497 if (type == RX_SERVER_CONNECTION)
3498 call = rxi_ReceiveServerCall(socket, np, conn);
3500 call = rxi_ReceiveClientCall(np, conn);
3503 putConnection(conn);
3507 MUTEX_ASSERT(&call->lock);
3508 /* Set remote user defined status from packet */
3509 call->remoteStatus = np->header.userStatus;
3511 /* Now do packet type-specific processing */
3512 switch (np->header.type) {
3513 case RX_PACKET_TYPE_DATA:
3514 /* If we're a client, and receiving a response, then all the packets
3515 * we transmitted packets are implicitly acknowledged. */
3516 if (type == RX_CLIENT_CONNECTION && !opr_queue_IsEmpty(&call->tq))
3517 rxi_AckAllInTransmitQueue(call);
3519 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3522 case RX_PACKET_TYPE_ACK:
3523 /* Respond immediately to ack packets requesting acknowledgement
3525 if (np->header.flags & RX_REQUEST_ACK) {
3527 (void)rxi_SendCallAbort(call, 0, 1, 0);
3529 (void)rxi_SendAck(call, 0, np->header.serial,
3530 RX_ACK_PING_RESPONSE, 1);
3532 np = rxi_ReceiveAckPacket(call, np, 1);
3534 case RX_PACKET_TYPE_ABORT: {
3535 /* An abort packet: reset the call, passing the error up to the user. */
3536 /* What if error is zero? */
3537 /* What if the error is -1? the application will treat it as a timeout. */
3538 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3539 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3540 rxi_CallError(call, errdata);
3541 MUTEX_EXIT(&call->lock);
3542 putConnection(conn);
3543 return np; /* xmitting; drop packet */
3545 case RX_PACKET_TYPE_BUSY:
3546 /* Mostly ignore BUSY packets. We will update lastReceiveTime below,
3547 * so we don't think the endpoint is completely dead, but otherwise
3548 * just act as if we never saw anything. If all we get are BUSY packets
3549 * back, then we will eventually error out with RX_CALL_TIMEOUT if the
3550 * connection is configured with idle/hard timeouts. */
3553 case RX_PACKET_TYPE_ACKALL:
3554 /* All packets acknowledged, so we can drop all packets previously
3555 * readied for sending */
3556 rxi_AckAllInTransmitQueue(call);
3559 /* Should not reach here, unless the peer is broken: send an abort
3561 rxi_CallError(call, RX_PROTOCOL_ERROR);
3562 np = rxi_SendCallAbort(call, np, 1, 0);
3565 /* Note when this last legitimate packet was received, for keep-alive
3566 * processing. Note, we delay getting the time until now in the hope that
3567 * the packet will be delivered to the user before any get time is required
3568 * (if not, then the time won't actually be re-evaluated here). */
3569 call->lastReceiveTime = clock_Sec();
3570 MUTEX_EXIT(&call->lock);
3571 putConnection(conn);
3575 /* return true if this is an "interesting" connection from the point of view
3576 of someone trying to debug the system */
3578 rxi_IsConnInteresting(struct rx_connection *aconn)
3581 struct rx_call *tcall;
3583 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3586 for (i = 0; i < RX_MAXCALLS; i++) {
3587 tcall = aconn->call[i];
3589 if ((tcall->state == RX_STATE_PRECALL)
3590 || (tcall->state == RX_STATE_ACTIVE))
3592 if ((tcall->app.mode == RX_MODE_SENDING)
3593 || (tcall->app.mode == RX_MODE_RECEIVING))
3601 /* if this is one of the last few packets AND it wouldn't be used by the
3602 receiving call to immediately satisfy a read request, then drop it on
3603 the floor, since accepting it might prevent a lock-holding thread from
3604 making progress in its reading. If a call has been cleared while in
3605 the precall state then ignore all subsequent packets until the call
3606 is assigned to a thread. */
3609 TooLow(struct rx_packet *ap, struct rx_call *acall)
3613 MUTEX_ENTER(&rx_quota_mutex);
3614 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3615 && (acall->state == RX_STATE_PRECALL))
3616 || ((rx_nFreePackets < rxi_dataQuota + 2)
3617 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3618 && (acall->flags & RX_CALL_READER_WAIT)))) {
3621 MUTEX_EXIT(&rx_quota_mutex);
3627 * Clear the attach wait flag on a connection and proceed.
3629 * Any processing waiting for a connection to be attached should be
3630 * unblocked. We clear the flag and do any other needed tasks.
3633 * the conn to unmark waiting for attach
3635 * @pre conn's conn_data_lock must be locked before calling this function
3639 rxi_ConnClearAttachWait(struct rx_connection *conn)
3641 /* Indicate that rxi_CheckReachEvent is no longer running by
3642 * clearing the flag. Must be atomic under conn_data_lock to
3643 * avoid a new call slipping by: rxi_CheckConnReach holds
3644 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3646 conn->flags &= ~RX_CONN_ATTACHWAIT;
3647 if (conn->flags & RX_CONN_NAT_PING) {
3648 conn->flags &= ~RX_CONN_NAT_PING;
3649 rxi_ScheduleNatKeepAliveEvent(conn);
3654 * Event handler function for connection-specific events for checking
3655 * reachability. Also called directly from main code with |event| == NULL
3656 * in order to trigger the initial reachability check.
3658 * When |event| == NULL, must be called with the connection data lock held,
3659 * but returns with the lock unlocked.
3662 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3664 struct rx_connection *conn = arg1;
3665 struct rx_call *acall = arg2;
3666 struct rx_call *call = acall;
3667 struct clock when, now;
3671 MUTEX_ENTER(&conn->conn_data_lock);
3673 MUTEX_ASSERT(&conn->conn_data_lock);
3675 if (event != NULL && event == conn->checkReachEvent)
3676 rxevent_Put(&conn->checkReachEvent);
3677 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3678 MUTEX_EXIT(&conn->conn_data_lock);
3682 MUTEX_ENTER(&conn->conn_call_lock);
3683 MUTEX_ENTER(&conn->conn_data_lock);
3684 for (i = 0; i < RX_MAXCALLS; i++) {
3685 struct rx_call *tc = conn->call[i];
3686 if (tc && tc->state == RX_STATE_PRECALL) {
3692 rxi_ConnClearAttachWait(conn);
3693 MUTEX_EXIT(&conn->conn_data_lock);
3694 MUTEX_EXIT(&conn->conn_call_lock);
3699 MUTEX_ENTER(&call->lock);
3700 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3702 MUTEX_EXIT(&call->lock);
3704 clock_GetTime(&now);
3706 when.sec += RX_CHECKREACH_TIMEOUT;
3707 MUTEX_ENTER(&conn->conn_data_lock);
3708 if (!conn->checkReachEvent) {
3709 rx_GetConnection(conn);
3710 conn->checkReachEvent = rxevent_Post(&when, &now,
3711 rxi_CheckReachEvent, conn,
3714 MUTEX_EXIT(&conn->conn_data_lock);
3717 /* If fired as an event handler, drop our refcount on the connection. */
3719 putConnection(conn);
3723 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3725 struct rx_service *service = conn->service;
3726 struct rx_peer *peer = conn->peer;
3727 afs_uint32 now, lastReach;
3729 if (service->checkReach == 0)
3733 MUTEX_ENTER(&peer->peer_lock);
3734 lastReach = peer->lastReachTime;
3735 MUTEX_EXIT(&peer->peer_lock);
3736 if (now - lastReach < RX_CHECKREACH_TTL)
3739 MUTEX_ENTER(&conn->conn_data_lock);
3740 if (conn->flags & RX_CONN_ATTACHWAIT) {
3741 MUTEX_EXIT(&conn->conn_data_lock);
3744 conn->flags |= RX_CONN_ATTACHWAIT;
3745 if (conn->checkReachEvent == NULL) {
3746 /* rxi_CheckReachEvent(NULL, ...) will drop the lock. */
3747 rxi_CheckReachEvent(NULL, conn, call, 0);
3749 MUTEX_EXIT(&conn->conn_data_lock);
3755 /* try to attach call, if authentication is complete */
3757 TryAttach(struct rx_call *acall, osi_socket socket,
3758 int *tnop, struct rx_call **newcallp,
3761 struct rx_connection *conn = acall->conn;
3763 if (conn->type == RX_SERVER_CONNECTION
3764 && acall->state == RX_STATE_PRECALL) {
3765 /* Don't attach until we have any req'd. authentication. */
3766 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3767 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3768 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3769 /* Note: this does not necessarily succeed; there
3770 * may not any proc available
3773 rxi_ChallengeOn(acall->conn);
3778 /* A data packet has been received off the interface. This packet is
3779 * appropriate to the call (the call is in the right state, etc.). This
3780 * routine can return a packet to the caller, for re-use */
3782 static struct rx_packet *
3783 rxi_ReceiveDataPacket(struct rx_call *call,
3784 struct rx_packet *np, int istack,
3785 osi_socket socket, afs_uint32 host, u_short port,
3786 int *tnop, struct rx_call **newcallp)
3788 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3793 afs_uint32 serial=0, flags=0;
3795 struct rx_packet *tnp;
3796 if (rx_stats_active)
3797 rx_atomic_inc(&rx_stats.dataPacketsRead);
3800 /* If there are no packet buffers, drop this new packet, unless we can find
3801 * packet buffers from inactive calls */
3803 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3804 MUTEX_ENTER(&rx_freePktQ_lock);
3805 rxi_NeedMorePackets = TRUE;
3806 MUTEX_EXIT(&rx_freePktQ_lock);
3807 if (rx_stats_active)
3808 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3809 rxi_calltrace(RX_TRACE_DROP, call);
3810 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3811 /* We used to clear the receive queue here, in an attempt to free
3812 * packets. However this is unsafe if the queue has received a
3813 * soft ACK for the final packet */
3814 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3820 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3821 * packet is one of several packets transmitted as a single
3822 * datagram. Do not send any soft or hard acks until all packets
3823 * in a jumbogram have been processed. Send negative acks right away.
3825 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3826 /* tnp is non-null when there are more packets in the
3827 * current jumbo gram */
3834 seq = np->header.seq;
3835 serial = np->header.serial;
3836 flags = np->header.flags;
3838 /* If the call is in an error state, send an abort message */
3840 return rxi_SendCallAbort(call, np, istack, 0);
3842 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3843 * AFS 3.5 jumbogram. */
3844 if (flags & RX_JUMBO_PACKET) {
3845 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3850 if (np->header.spare != 0) {
3851 MUTEX_ENTER(&call->conn->conn_data_lock);
3852 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3853 MUTEX_EXIT(&call->conn->conn_data_lock);
3856 /* The usual case is that this is the expected next packet */
3857 if (seq == call->rnext) {
3859 /* Check to make sure it is not a duplicate of one already queued */
3860 if (!opr_queue_IsEmpty(&call->rq)
3861 && opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq == seq) {
3862 if (rx_stats_active)
3863 rx_atomic_inc(&rx_stats.dupPacketsRead);
3864 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3865 rxi_CancelDelayedAckEvent(call);
3866 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3872 /* It's the next packet. Stick it on the receive queue
3873 * for this call. Set newPackets to make sure we wake
3874 * the reader once all packets have been processed */
3875 #ifdef RX_TRACK_PACKETS
3876 np->flags |= RX_PKTFLAG_RQ;
3878 opr_queue_Prepend(&call->rq, &np->entry);
3879 #ifdef RXDEBUG_PACKET
3881 #endif /* RXDEBUG_PACKET */
3883 np = NULL; /* We can't use this anymore */
3886 /* If an ack is requested then set a flag to make sure we
3887 * send an acknowledgement for this packet */
3888 if (flags & RX_REQUEST_ACK) {
3889 ackNeeded = RX_ACK_REQUESTED;
3892 /* Keep track of whether we have received the last packet */
3893 if (flags & RX_LAST_PACKET) {
3894 call->flags |= RX_CALL_HAVE_LAST;
3898 /* Check whether we have all of the packets for this call */
3899 if (call->flags & RX_CALL_HAVE_LAST) {
3900 afs_uint32 tseq; /* temporary sequence number */
3901 struct opr_queue *cursor;
3903 for (tseq = seq, opr_queue_Scan(&call->rq, cursor)) {
3904 struct rx_packet *tp;
3906 tp = opr_queue_Entry(cursor, struct rx_packet, entry);
3907 if (tseq != tp->header.seq)
3909 if (tp->header.flags & RX_LAST_PACKET) {
3910 call->flags |= RX_CALL_RECEIVE_DONE;
3917 /* Provide asynchronous notification for those who want it
3918 * (e.g. multi rx) */
3919 if (call->arrivalProc) {
3920 (*call->arrivalProc) (call, call->arrivalProcHandle,
3921 call->arrivalProcArg);
3922 call->arrivalProc = (void (*)())0;
3925 /* Update last packet received */
3928 /* If there is no server process serving this call, grab
3929 * one, if available. We only need to do this once. If a
3930 * server thread is available, this thread becomes a server
3931 * thread and the server thread becomes a listener thread. */
3933 TryAttach(call, socket, tnop, newcallp, 0);
3936 /* This is not the expected next packet. */
3938 /* Determine whether this is a new or old packet, and if it's
3939 * a new one, whether it fits into the current receive window.
3940 * Also figure out whether the packet was delivered in sequence.
3941 * We use the prev variable to determine whether the new packet
3942 * is the successor of its immediate predecessor in the
3943 * receive queue, and the missing flag to determine whether
3944 * any of this packets predecessors are missing. */
3946 afs_uint32 prev; /* "Previous packet" sequence number */
3947 struct opr_queue *cursor;
3948 int missing; /* Are any predecessors missing? */
3950 /* If the new packet's sequence number has been sent to the
3951 * application already, then this is a duplicate */
3952 if (seq < call->rnext) {
3953 if (rx_stats_active)
3954 rx_atomic_inc(&rx_stats.dupPacketsRead);
3955 rxi_CancelDelayedAckEvent(call);
3956 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3962 /* If the sequence number is greater than what can be
3963 * accomodated by the current window, then send a negative
3964 * acknowledge and drop the packet */
3965 if ((call->rnext + call->rwind) <= seq) {
3966 rxi_CancelDelayedAckEvent(call);
3967 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3974 /* Look for the packet in the queue of old received packets */
3975 prev = call->rnext - 1;
3977 for (opr_queue_Scan(&call->rq, cursor)) {
3978 struct rx_packet *tp
3979 = opr_queue_Entry(cursor, struct rx_packet, entry);
3981 /*Check for duplicate packet */
3982 if (seq == tp->header.seq) {
3983 if (rx_stats_active)
3984 rx_atomic_inc(&rx_stats.dupPacketsRead);
3985 rxi_CancelDelayedAckEvent(call);
3986 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3992 /* If we find a higher sequence packet, break out and
3993 * insert the new packet here. */
3994 if (seq < tp->header.seq)
3996 /* Check for missing packet */
3997 if (tp->header.seq != prev + 1) {
4001 prev = tp->header.seq;
4004 /* Keep track of whether we have received the last packet. */
4005 if (flags & RX_LAST_PACKET) {
4006 call->flags |= RX_CALL_HAVE_LAST;
4009 /* It's within the window: add it to the the receive queue.
4010 * tp is left by the previous loop either pointing at the
4011 * packet before which to insert the new packet, or at the
4012 * queue head if the queue is empty or the packet should be
4014 #ifdef RX_TRACK_PACKETS
4015 np->flags |= RX_PKTFLAG_RQ;
4017 #ifdef RXDEBUG_PACKET
4019 #endif /* RXDEBUG_PACKET */
4020 opr_queue_InsertBefore(cursor, &np->entry);
4024 /* Check whether we have all of the packets for this call */
4025 if ((call->flags & RX_CALL_HAVE_LAST)
4026 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4027 afs_uint32 tseq; /* temporary sequence number */
4030 for (opr_queue_Scan(&call->rq, cursor)) {
4031 struct rx_packet *tp
4032 = opr_queue_Entry(cursor, struct rx_packet, entry);
4033 if (tseq != tp->header.seq)
4035 if (tp->header.flags & RX_LAST_PACKET) {
4036 call->flags |= RX_CALL_RECEIVE_DONE;
4043 /* We need to send an ack of the packet is out of sequence,
4044 * or if an ack was requested by the peer. */
4045 if (seq != prev + 1 || missing) {
4046 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4047 } else if (flags & RX_REQUEST_ACK) {
4048 ackNeeded = RX_ACK_REQUESTED;
4051 /* Acknowledge the last packet for each call */
4052 if (flags & RX_LAST_PACKET) {
4063 * If the receiver is waiting for an iovec, fill the iovec
4064 * using the data from the receive queue */
4065 if (call->flags & RX_CALL_IOVEC_WAIT) {
4066 didHardAck = rxi_FillReadVec(call, serial);
4067 /* the call may have been aborted */
4076 /* Wakeup the reader if any */
4077 if ((call->flags & RX_CALL_READER_WAIT)
4078 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4079 || (call->iovNext >= call->iovMax)
4080 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4081 call->flags &= ~RX_CALL_READER_WAIT;
4082 #ifdef RX_ENABLE_LOCKS
4083 CV_BROADCAST(&call->cv_rq);
4085 osi_rxWakeup(&call->rq);
4091 * Send an ack when requested by the peer, or once every
4092 * rxi_SoftAckRate packets until the last packet has been
4093 * received. Always send a soft ack for the last packet in
4094 * the server's reply. */
4096 rxi_CancelDelayedAckEvent(call);
4097 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4098 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4099 rxi_CancelDelayedAckEvent(call);
4100 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4101 } else if (call->nSoftAcks) {
4102 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4103 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4105 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4106 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4107 rxi_CancelDelayedAckEvent(call);
4114 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4116 struct rx_peer *peer = conn->peer;
4118 MUTEX_ENTER(&peer->peer_lock);
4119 peer->lastReachTime = clock_Sec();
4120 MUTEX_EXIT(&peer->peer_lock);
4122 MUTEX_ENTER(&conn->conn_data_lock);
4123 if (conn->flags & RX_CONN_ATTACHWAIT) {
4126 rxi_ConnClearAttachWait(conn);
4127 MUTEX_EXIT(&conn->conn_data_lock);
4129 for (i = 0; i < RX_MAXCALLS; i++) {
4130 struct rx_call *call = conn->call[i];
4133 MUTEX_ENTER(&call->lock);
4134 /* tnop can be null if newcallp is null */
4135 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4137 MUTEX_EXIT(&call->lock);
4141 MUTEX_EXIT(&conn->conn_data_lock);
4144 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4146 rx_ack_reason(int reason)
4149 case RX_ACK_REQUESTED:
4151 case RX_ACK_DUPLICATE:
4153 case RX_ACK_OUT_OF_SEQUENCE:
4155 case RX_ACK_EXCEEDS_WINDOW:
4157 case RX_ACK_NOSPACE:
4161 case RX_ACK_PING_RESPONSE:
4174 /* The real smarts of the whole thing. */
4175 static struct rx_packet *
4176 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4179 struct rx_ackPacket *ap;
4181 struct rx_packet *tp;
4182 struct rx_connection *conn = call->conn;
4183 struct rx_peer *peer = conn->peer;
4184 struct opr_queue *cursor;
4185 struct clock now; /* Current time, for RTT calculations */
4193 int newAckCount = 0;
4194 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4195 int pktsize = 0; /* Set if we need to update the peer mtu */
4196 int conn_data_locked = 0;
4198 if (rx_stats_active)
4199 rx_atomic_inc(&rx_stats.ackPacketsRead);
4200 ap = (struct rx_ackPacket *)rx_DataOf(np);
4201 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4203 return np; /* truncated ack packet */
4205 /* depends on ack packet struct */
4206 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4207 first = ntohl(ap->firstPacket);
4208 prev = ntohl(ap->previousPacket);
4209 serial = ntohl(ap->serial);
4212 * Ignore ack packets received out of order while protecting
4213 * against peers that set the previousPacket field to a packet
4214 * serial number instead of a sequence number.
4216 if (first < call->tfirst ||
4217 (first == call->tfirst && prev < call->tprev && prev < call->tfirst
4224 if (np->header.flags & RX_SLOW_START_OK) {
4225 call->flags |= RX_CALL_SLOW_START_OK;
4228 if (ap->reason == RX_ACK_PING_RESPONSE)
4229 rxi_UpdatePeerReach(conn, call);
4231 if (conn->lastPacketSizeSeq) {
4232 MUTEX_ENTER(&conn->conn_data_lock);
4233 conn_data_locked = 1;
4234 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4235 pktsize = conn->lastPacketSize;
4236 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4239 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4240 if (!conn_data_locked) {
4241 MUTEX_ENTER(&conn->conn_data_lock);
4242 conn_data_locked = 1;
4244 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4245 /* process mtu ping ack */
4246 pktsize = conn->lastPingSize;
4247 conn->lastPingSizeSer = conn->lastPingSize = 0;
4251 if (conn_data_locked) {
4252 MUTEX_EXIT(&conn->conn_data_lock);
4253 conn_data_locked = 0;
4257 if (rxdebug_active) {
4261 len = _snprintf(msg, sizeof(msg),
4262 "tid[%d] RACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4263 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4264 ntohl(ap->serial), ntohl(ap->previousPacket),
4265 (unsigned int)np->header.seq, ntohl(ap->firstPacket),
4266 ap->nAcks, ntohs(ap->bufferSpace) );
4270 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4271 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4275 OutputDebugString(msg);
4277 #else /* AFS_NT40_ENV */
4280 "RACK: reason %x previous %u seq %u serial %u first %u",
4281 ap->reason, ntohl(ap->previousPacket),
4282 (unsigned int)np->header.seq, (unsigned int)serial,
4283 ntohl(ap->firstPacket));
4286 for (offset = 0; offset < nAcks; offset++)
4287 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4292 #endif /* AFS_NT40_ENV */
4295 MUTEX_ENTER(&peer->peer_lock);
4298 * Start somewhere. Can't assume we can send what we can receive,
4299 * but we are clearly receiving.
4301 if (!peer->maxPacketSize)
4302 peer->maxPacketSize = RX_MIN_PACKET_SIZE - RX_HEADER_SIZE;
4304 if (pktsize > peer->maxPacketSize) {
4305 peer->maxPacketSize = pktsize;
4306 if ((pktsize + RX_HEADER_SIZE > peer->ifMTU)) {
4307 peer->ifMTU = pktsize + RX_HEADER_SIZE;
4308 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4309 rxi_ScheduleGrowMTUEvent(call, 1);
4314 clock_GetTime(&now);
4316 /* The transmit queue splits into 4 sections.
4318 * The first section is packets which have now been acknowledged
4319 * by a window size change in the ack. These have reached the
4320 * application layer, and may be discarded. These are packets
4321 * with sequence numbers < ap->firstPacket.
4323 * The second section is packets which have sequence numbers in
4324 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4325 * contents of the packet's ack array determines whether these
4326 * packets are acknowledged or not.
4328 * The third section is packets which fall above the range
4329 * addressed in the ack packet. These have not yet been received
4332 * The four section is packets which have not yet been transmitted.
4333 * These packets will have a header.serial of 0.
4336 /* First section - implicitly acknowledged packets that can be
4340 tp = opr_queue_First(&call->tq, struct rx_packet, entry);
4341 while(!opr_queue_IsEnd(&call->tq, &tp->entry) && tp->header.seq < first) {
4342 struct rx_packet *next;
4344 next = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4345 call->tfirst = tp->header.seq + 1;
4347 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4349 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4352 #ifdef RX_ENABLE_LOCKS
4353 /* XXX Hack. Because we have to release the global call lock when sending
4354 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4355 * in rxi_Start sending packets out because packets may move to the
4356 * freePacketQueue as result of being here! So we drop these packets until
4357 * we're safely out of the traversing. Really ugly!
4358 * To make it even uglier, if we're using fine grain locking, we can
4359 * set the ack bits in the packets and have rxi_Start remove the packets
4360 * when it's done transmitting.
4362 if (call->flags & RX_CALL_TQ_BUSY) {
4363 tp->flags |= RX_PKTFLAG_ACKED;
4364 call->flags |= RX_CALL_TQ_SOME_ACKED;
4366 #endif /* RX_ENABLE_LOCKS */
4368 opr_queue_Remove(&tp->entry);
4369 #ifdef RX_TRACK_PACKETS
4370 tp->flags &= ~RX_PKTFLAG_TQ;
4372 #ifdef RXDEBUG_PACKET
4374 #endif /* RXDEBUG_PACKET */
4375 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4380 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4382 /* Second section of the queue - packets for which we are receiving
4385 * Go through the explicit acks/nacks and record the results in
4386 * the waiting packets. These are packets that can't be released
4387 * yet, even with a positive acknowledge. This positive
4388 * acknowledge only means the packet has been received by the
4389 * peer, not that it will be retained long enough to be sent to
4390 * the peer's upper level. In addition, reset the transmit timers
4391 * of any missing packets (those packets that must be missing
4392 * because this packet was out of sequence) */
4394 call->nSoftAcked = 0;
4396 while (!opr_queue_IsEnd(&call->tq, &tp->entry)
4397 && tp->header.seq < first + nAcks) {
4398 /* Set the acknowledge flag per packet based on the
4399 * information in the ack packet. An acknowlegded packet can
4400 * be downgraded when the server has discarded a packet it
4401 * soacked previously, or when an ack packet is received
4402 * out of sequence. */
4403 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4404 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4406 tp->flags |= RX_PKTFLAG_ACKED;
4407 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4414 } else /* RX_ACK_TYPE_NACK */ {
4415 tp->flags &= ~RX_PKTFLAG_ACKED;
4419 tp = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4422 /* We don't need to take any action with the 3rd or 4th section in the
4423 * queue - they're not addressed by the contents of this ACK packet.
4426 /* if the ack packet has a receivelen field hanging off it,
4427 * update our state */
4428 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4431 /* If the ack packet has a "recommended" size that is less than
4432 * what I am using now, reduce my size to match */
4433 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4434 (int)sizeof(afs_int32), &tSize);
4435 tSize = (afs_uint32) ntohl(tSize);
4436 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4438 /* Get the maximum packet size to send to this peer */
4439 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4441 tSize = (afs_uint32) ntohl(tSize);
4442 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4443 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4445 /* sanity check - peer might have restarted with different params.
4446 * If peer says "send less", dammit, send less... Peer should never
4447 * be unable to accept packets of the size that prior AFS versions would
4448 * send without asking. */
4449 if (peer->maxMTU != tSize) {
4450 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4452 peer->maxMTU = tSize;
4453 peer->MTU = MIN(tSize, peer->MTU);
4454 call->MTU = MIN(call->MTU, tSize);
4457 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4460 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4461 (int)sizeof(afs_int32), &tSize);
4462 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4463 if (tSize < call->twind) { /* smaller than our send */
4464 call->twind = tSize; /* window, we must send less... */
4465 call->ssthresh = MIN(call->twind, call->ssthresh);
4466 call->conn->twind[call->channel] = call->twind;
4469 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4470 * network MTU confused with the loopback MTU. Calculate the
4471 * maximum MTU here for use in the slow start code below.
4473 /* Did peer restart with older RX version? */
4474 if (peer->maxDgramPackets > 1) {
4475 peer->maxDgramPackets = 1;
4477 } else if (np->length >=
4478 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4481 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4482 sizeof(afs_int32), &tSize);
4483 tSize = (afs_uint32) ntohl(tSize);
4485 * As of AFS 3.5 we set the send window to match the receive window.
4487 if (tSize < call->twind) {
4488 call->twind = tSize;
4489 call->conn->twind[call->channel] = call->twind;
4490 call->ssthresh = MIN(call->twind, call->ssthresh);
4491 } else if (tSize > call->twind) {
4492 call->twind = tSize;
4493 call->conn->twind[call->channel] = call->twind;
4497 * As of AFS 3.5, a jumbogram is more than one fixed size
4498 * packet transmitted in a single UDP datagram. If the remote
4499 * MTU is smaller than our local MTU then never send a datagram
4500 * larger than the natural MTU.
4503 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4504 (int)sizeof(afs_int32), &tSize);
4505 maxDgramPackets = (afs_uint32) ntohl(tSize);
4506 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4508 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4509 if (maxDgramPackets > 1) {
4510 peer->maxDgramPackets = maxDgramPackets;
4511 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4513 peer->maxDgramPackets = 1;
4514 call->MTU = peer->natMTU;
4516 } else if (peer->maxDgramPackets > 1) {
4517 /* Restarted with lower version of RX */
4518 peer->maxDgramPackets = 1;
4520 } else if (peer->maxDgramPackets > 1
4521 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4522 /* Restarted with lower version of RX */
4523 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4524 peer->natMTU = OLD_MAX_PACKET_SIZE;
4525 peer->MTU = OLD_MAX_PACKET_SIZE;
4526 peer->maxDgramPackets = 1;
4527 peer->nDgramPackets = 1;
4529 call->MTU = OLD_MAX_PACKET_SIZE;
4532 /* If the window has been extended by this acknowledge packet,
4533 * then wakeup a sender waiting in alloc for window space, or try
4534 * sending packets now, if he's been sitting on packets due to
4535 * lack of window space */
4536 if (call->tnext < (call->tfirst + call->twind)) {
4537 #ifdef RX_ENABLE_LOCKS
4538 CV_SIGNAL(&call->cv_twind);
4540 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4541 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4542 osi_rxWakeup(&call->twind);
4545 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4546 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4552 * Calculate how many datagrams were successfully received after
4553 * the first missing packet and adjust the negative ack counter
4558 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4559 if (call->nNacks < nNacked) {
4560 call->nNacks = nNacked;
4563 call->nAcks += newAckCount;
4567 /* If the packet contained new acknowledgements, rather than just
4568 * being a duplicate of one we have previously seen, then we can restart
4571 if (newAckCount > 0)
4572 rxi_rto_packet_acked(call, istack);
4574 if (call->flags & RX_CALL_FAST_RECOVER) {
4575 if (newAckCount == 0) {
4576 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4578 call->flags &= ~RX_CALL_FAST_RECOVER;
4579 call->cwind = call->nextCwind;
4580 call->nextCwind = 0;
4583 call->nCwindAcks = 0;
4584 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4585 /* Three negative acks in a row trigger congestion recovery */
4586 call->flags |= RX_CALL_FAST_RECOVER;
4587 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4589 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4590 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4591 call->nextCwind = call->ssthresh;
4594 peer->MTU = call->MTU;
4595 peer->cwind = call->nextCwind;
4596 peer->nDgramPackets = call->nDgramPackets;
4598 call->congestSeq = peer->congestSeq;
4600 /* Reset the resend times on the packets that were nacked
4601 * so we will retransmit as soon as the window permits
4605 for (opr_queue_ScanBackwards(&call->tq, cursor)) {
4606 struct rx_packet *tp =
4607 opr_queue_Entry(cursor, struct rx_packet, entry);
4609 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4610 tp->flags &= ~RX_PKTFLAG_SENT;
4612 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4617 /* If cwind is smaller than ssthresh, then increase
4618 * the window one packet for each ack we receive (exponential
4620 * If cwind is greater than or equal to ssthresh then increase
4621 * the congestion window by one packet for each cwind acks we
4622 * receive (linear growth). */
4623 if (call->cwind < call->ssthresh) {
4625 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4626 call->nCwindAcks = 0;
4628 call->nCwindAcks += newAckCount;
4629 if (call->nCwindAcks >= call->cwind) {
4630 call->nCwindAcks = 0;
4631 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4635 * If we have received several acknowledgements in a row then
4636 * it is time to increase the size of our datagrams
4638 if ((int)call->nAcks > rx_nDgramThreshold) {
4639 if (peer->maxDgramPackets > 1) {
4640 if (call->nDgramPackets < peer->maxDgramPackets) {
4641 call->nDgramPackets++;
4643 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4644 } else if (call->MTU < peer->maxMTU) {
4645 /* don't upgrade if we can't handle it */
4646 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4647 call->MTU = peer->ifMTU;
4649 call->MTU += peer->natMTU;
4650 call->MTU = MIN(call->MTU, peer->maxMTU);
4657 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4659 /* Servers need to hold the call until all response packets have
4660 * been acknowledged. Soft acks are good enough since clients
4661 * are not allowed to clear their receive queues. */
4662 if (call->state == RX_STATE_HOLD
4663 && call->tfirst + call->nSoftAcked >= call->tnext) {
4664 call->state = RX_STATE_DALLY;
4665 rxi_ClearTransmitQueue(call, 0);
4666 rxi_CancelKeepAliveEvent(call);
4667 } else if (!opr_queue_IsEmpty(&call->tq)) {
4668 rxi_Start(call, istack);
4674 * Schedule a connection abort to be sent after some delay.
4676 * @param[in] conn The connection to send the abort on.
4677 * @param[in] msec The number of milliseconds to wait before sending.
4679 * @pre conn_data_lock must be held
4682 rxi_SendConnectionAbortLater(struct rx_connection *conn, int msec)
4684 struct clock when, now;
4686 MUTEX_ASSERT(&conn->conn_data_lock);
4690 if (!conn->delayedAbortEvent) {
4691 clock_GetTime(&now);
4693 clock_Addmsec(&when, msec);
4694 rx_GetConnection(conn);
4695 conn->delayedAbortEvent =
4696 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
4700 /* Received a response to a challenge packet */
4701 static struct rx_packet *
4702 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4703 struct rx_packet *np, int istack)
4707 /* Ignore the packet if we're the client */
4708 if (conn->type == RX_CLIENT_CONNECTION)
4711 /* If already authenticated, ignore the packet (it's probably a retry) */
4712 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4715 if (!conn->securityChallengeSent) {
4716 /* We've never sent out a challenge for this connection, so this
4717 * response cannot possibly be correct; ignore it. This can happen
4718 * if we sent a challenge to the client, then we were restarted, and
4719 * then the client sent us a response. If we ignore the response, the
4720 * client will eventually resend a data packet, causing us to send a
4721 * new challenge and the client to send a new response. */
4725 /* Otherwise, have the security object evaluate the response packet */
4726 error = RXS_CheckResponse(conn->securityObject, conn, np);
4728 /* If the response is invalid, reset the connection, sending
4729 * an abort to the peer. Send the abort with a 1 second delay,
4730 * to avoid a peer hammering us by constantly recreating a
4731 * connection with bad credentials. */
4732 rxi_ConnectionError(conn, error);
4733 MUTEX_ENTER(&conn->conn_data_lock);
4734 rxi_SendConnectionAbortLater(conn, 1000);
4735 MUTEX_EXIT(&conn->conn_data_lock);
4738 /* If the response is valid, any calls waiting to attach
4739 * servers can now do so */
4742 for (i = 0; i < RX_MAXCALLS; i++) {
4743 struct rx_call *call = conn->call[i];
4745 MUTEX_ENTER(&call->lock);
4746 if (call->state == RX_STATE_PRECALL)
4747 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4748 /* tnop can be null if newcallp is null */
4749 MUTEX_EXIT(&call->lock);
4753 /* Update the peer reachability information, just in case
4754 * some calls went into attach-wait while we were waiting
4755 * for authentication..
4757 rxi_UpdatePeerReach(conn, NULL);
4762 /* A client has received an authentication challenge: the security
4763 * object is asked to cough up a respectable response packet to send
4764 * back to the server. The server is responsible for retrying the
4765 * challenge if it fails to get a response. */
4767 static struct rx_packet *
4768 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4769 struct rx_packet *np, int istack)
4773 /* Ignore the challenge if we're the server */
4774 if (conn->type == RX_SERVER_CONNECTION)
4777 /* Ignore the challenge if the connection is otherwise idle; someone's
4778 * trying to use us as an oracle. */
4779 if (!rxi_HasActiveCalls(conn))
4782 /* Send the security object the challenge packet. It is expected to fill
4783 * in the response. */
4784 error = RXS_GetResponse(conn->securityObject, conn, np);
4786 /* If the security object is unable to return a valid response, reset the
4787 * connection and send an abort to the peer. Otherwise send the response
4788 * packet to the peer connection. */
4790 rxi_ConnectionError(conn, error);
4791 MUTEX_ENTER(&conn->conn_data_lock);
4792 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4793 MUTEX_EXIT(&conn->conn_data_lock);
4795 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4796 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4802 /* Find an available server process to service the current request in
4803 * the given call structure. If one isn't available, queue up this
4804 * call so it eventually gets one */
4806 rxi_AttachServerProc(struct rx_call *call,
4807 osi_socket socket, int *tnop,
4808 struct rx_call **newcallp)
4810 struct rx_serverQueueEntry *sq;
4811 struct rx_service *service = call->conn->service;
4814 /* May already be attached */
4815 if (call->state == RX_STATE_ACTIVE)
4818 MUTEX_ENTER(&rx_serverPool_lock);
4820 haveQuota = QuotaOK(service);
4821 if ((!haveQuota) || opr_queue_IsEmpty(&rx_idleServerQueue)) {
4822 /* If there are no processes available to service this call,
4823 * put the call on the incoming call queue (unless it's
4824 * already on the queue).
4826 #ifdef RX_ENABLE_LOCKS
4828 ReturnToServerPool(service);
4829 #endif /* RX_ENABLE_LOCKS */
4831 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4832 call->flags |= RX_CALL_WAIT_PROC;
4833 rx_atomic_inc(&rx_nWaiting);
4834 rx_atomic_inc(&rx_nWaited);
4835 rxi_calltrace(RX_CALL_ARRIVAL, call);
4836 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4837 opr_queue_Append(&rx_incomingCallQueue, &call->entry);
4840 sq = opr_queue_Last(&rx_idleServerQueue,
4841 struct rx_serverQueueEntry, entry);
4843 /* If hot threads are enabled, and both newcallp and sq->socketp
4844 * are non-null, then this thread will process the call, and the
4845 * idle server thread will start listening on this threads socket.
4847 opr_queue_Remove(&sq->entry);
4849 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4852 *sq->socketp = socket;
4853 clock_GetTime(&call->startTime);
4854 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4858 if (call->flags & RX_CALL_WAIT_PROC) {
4859 /* Conservative: I don't think this should happen */
4860 call->flags &= ~RX_CALL_WAIT_PROC;
4861 rx_atomic_dec(&rx_nWaiting);
4862 if (opr_queue_IsOnQueue(&call->entry)) {
4863 opr_queue_Remove(&call->entry);
4866 call->state = RX_STATE_ACTIVE;
4867 call->app.mode = RX_MODE_RECEIVING;
4868 #ifdef RX_KERNEL_TRACE
4870 int glockOwner = ISAFS_GLOCK();
4873 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4874 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4880 if (call->flags & RX_CALL_CLEARED) {
4881 /* send an ack now to start the packet flow up again */
4882 call->flags &= ~RX_CALL_CLEARED;
4883 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4885 #ifdef RX_ENABLE_LOCKS
4888 service->nRequestsRunning++;
4889 MUTEX_ENTER(&rx_quota_mutex);
4890 if (service->nRequestsRunning <= service->minProcs)
4893 MUTEX_EXIT(&rx_quota_mutex);
4897 MUTEX_EXIT(&rx_serverPool_lock);
4900 /* Delay the sending of an acknowledge event for a short while, while
4901 * a new call is being prepared (in the case of a client) or a reply
4902 * is being prepared (in the case of a server). Rather than sending
4903 * an ack packet, an ACKALL packet is sent. */
4905 rxi_AckAll(struct rx_call *call)
4907 rxi_SendSpecial(call, call->conn, NULL, RX_PACKET_TYPE_ACKALL,
4909 call->flags |= RX_CALL_ACKALL_SENT;
4913 * Event handler for per-call delayed acks.
4914 * Also called synchronously, with |event| == NULL, to send a "delayed" ack
4918 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4921 struct rx_call *call = arg1;
4922 #ifdef RX_ENABLE_LOCKS
4924 MUTEX_ENTER(&call->lock);
4925 if (event == call->delayedAckEvent)
4926 rxevent_Put(&call->delayedAckEvent);
4928 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4930 MUTEX_EXIT(&call->lock);
4931 #else /* RX_ENABLE_LOCKS */
4933 rxevent_Put(&call->delayedAckEvent);
4934 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4935 #endif /* RX_ENABLE_LOCKS */
4936 /* Release the call reference for the event that fired. */
4938 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4941 #ifdef RX_ENABLE_LOCKS
4942 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4943 * clearing them out.
4946 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4948 struct opr_queue *cursor;
4951 for (opr_queue_Scan(&call->tq, cursor)) {
4953 = opr_queue_Entry(cursor, struct rx_packet, entry);
4955 p->flags |= RX_PKTFLAG_ACKED;
4960 call->flags |= RX_CALL_TQ_CLEARME;
4961 call->flags |= RX_CALL_TQ_SOME_ACKED;
4964 rxi_rto_cancel(call);
4966 call->tfirst = call->tnext;
4967 call->nSoftAcked = 0;
4969 if (call->flags & RX_CALL_FAST_RECOVER) {
4970 call->flags &= ~RX_CALL_FAST_RECOVER;
4971 call->cwind = call->nextCwind;
4972 call->nextCwind = 0;
4975 CV_SIGNAL(&call->cv_twind);
4977 #endif /* RX_ENABLE_LOCKS */
4980 * Acknowledge the whole transmit queue.
4982 * If we're running without locks, or the transmit queue isn't busy, then
4983 * we can just clear the queue now. Otherwise, we have to mark all of the
4984 * packets as acknowledged, and let rxi_Start clear it later on
4987 rxi_AckAllInTransmitQueue(struct rx_call *call)
4989 #ifdef RX_ENABLE_LOCKS
4990 if (call->flags & RX_CALL_TQ_BUSY) {
4991 rxi_SetAcksInTransmitQueue(call);
4995 rxi_ClearTransmitQueue(call, 0);
4997 /* Clear out the transmit queue for the current call (all packets have
4998 * been received by peer) */
5000 rxi_ClearTransmitQueue(struct rx_call *call, int force)
5002 #ifdef RX_ENABLE_LOCKS
5003 struct opr_queue *cursor;
5004 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
5006 for (opr_queue_Scan(&call->tq, cursor)) {
5008 = opr_queue_Entry(cursor, struct rx_packet, entry);
5010 p->flags |= RX_PKTFLAG_ACKED;
5014 call->flags |= RX_CALL_TQ_CLEARME;
5015 call->flags |= RX_CALL_TQ_SOME_ACKED;
5018 #endif /* RX_ENABLE_LOCKS */
5019 #ifdef RXDEBUG_PACKET
5021 #endif /* RXDEBUG_PACKET */
5022 rxi_FreePackets(0, &call->tq);
5023 rxi_WakeUpTransmitQueue(call);
5024 #ifdef RX_ENABLE_LOCKS
5025 call->flags &= ~RX_CALL_TQ_CLEARME;
5029 rxi_rto_cancel(call);
5030 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5031 call->nSoftAcked = 0;
5033 if (call->flags & RX_CALL_FAST_RECOVER) {
5034 call->flags &= ~RX_CALL_FAST_RECOVER;
5035 call->cwind = call->nextCwind;
5037 #ifdef RX_ENABLE_LOCKS
5038 CV_SIGNAL(&call->cv_twind);
5040 osi_rxWakeup(&call->twind);
5045 rxi_ClearReceiveQueue(struct rx_call *call)
5047 if (!opr_queue_IsEmpty(&call->rq)) {
5050 count = rxi_FreePackets(0, &call->rq);
5051 rx_packetReclaims += count;
5052 #ifdef RXDEBUG_PACKET
5054 if ( call->rqc != 0 )
5055 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5057 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5059 if (call->state == RX_STATE_PRECALL) {
5060 call->flags |= RX_CALL_CLEARED;
5064 /* Send an abort packet for the specified call */
5065 static struct rx_packet *
5066 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5067 int istack, int force)
5070 struct clock when, now;
5075 /* Clients should never delay abort messages */
5076 if (rx_IsClientConn(call->conn))
5079 if (call->abortCode != call->error) {
5080 call->abortCode = call->error;
5081 call->abortCount = 0;
5084 if (force || rxi_callAbortThreshhold == 0
5085 || call->abortCount < rxi_callAbortThreshhold) {
5086 rxi_CancelDelayedAbortEvent(call);
5087 error = htonl(call->error);
5090 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5091 (char *)&error, sizeof(error), istack);
5092 } else if (!call->delayedAbortEvent) {
5093 clock_GetTime(&now);
5095 clock_Addmsec(&when, rxi_callAbortDelay);
5096 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5097 call->delayedAbortEvent =
5098 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5104 rxi_CancelDelayedAbortEvent(struct rx_call *call)
5106 MUTEX_ASSERT(&call->lock);
5107 if (rxevent_Cancel(&call->delayedAbortEvent))
5108 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5111 /* Send an abort packet for the specified connection. Packet is an
5112 * optional pointer to a packet that can be used to send the abort.
5113 * Once the number of abort messages reaches the threshhold, an
5114 * event is scheduled to send the abort. Setting the force flag
5115 * overrides sending delayed abort messages.
5117 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5118 * to send the abort packet.
5121 rxi_SendConnectionAbort(struct rx_connection *conn,
5122 struct rx_packet *packet, int istack, int force)
5129 /* Clients should never delay abort messages */
5130 if (rx_IsClientConn(conn))
5133 if (force || rxi_connAbortThreshhold == 0
5134 || conn->abortCount < rxi_connAbortThreshhold) {
5136 if (rxevent_Cancel(&conn->delayedAbortEvent))
5137 putConnection(conn);
5138 error = htonl(conn->error);
5140 MUTEX_EXIT(&conn->conn_data_lock);
5142 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5143 RX_PACKET_TYPE_ABORT, (char *)&error,
5144 sizeof(error), istack);
5145 MUTEX_ENTER(&conn->conn_data_lock);
5147 rxi_SendConnectionAbortLater(conn, rxi_connAbortDelay);
5152 /* Associate an error all of the calls owned by a connection. Called
5153 * with error non-zero. This is only for really fatal things, like
5154 * bad authentication responses. The connection itself is set in
5155 * error at this point, so that future packets received will be
5158 rxi_ConnectionError(struct rx_connection *conn,
5164 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5166 MUTEX_ENTER(&conn->conn_data_lock);
5167 if (rxevent_Cancel(&conn->challengeEvent))
5168 putConnection(conn);
5169 if (rxevent_Cancel(&conn->natKeepAliveEvent))
5170 putConnection(conn);
5171 if (rxevent_Cancel(&conn->checkReachEvent)) {
5172 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5173 putConnection(conn);
5175 MUTEX_EXIT(&conn->conn_data_lock);
5176 for (i = 0; i < RX_MAXCALLS; i++) {
5177 struct rx_call *call = conn->call[i];
5179 MUTEX_ENTER(&call->lock);
5180 rxi_CallError(call, error);
5181 MUTEX_EXIT(&call->lock);
5184 conn->error = error;
5185 if (rx_stats_active)
5186 rx_atomic_inc(&rx_stats.fatalErrors);
5191 * Interrupt an in-progress call with the specified error and wakeup waiters.
5193 * @param[in] call The call to interrupt
5194 * @param[in] error The error code to send to the peer
5197 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5199 MUTEX_ENTER(&call->lock);
5200 rxi_CallError(call, error);
5201 rxi_SendCallAbort(call, NULL, 0, 1);
5202 MUTEX_EXIT(&call->lock);
5206 rxi_CallError(struct rx_call *call, afs_int32 error)
5208 MUTEX_ASSERT(&call->lock);
5209 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5211 error = call->error;
5213 #ifdef RX_ENABLE_LOCKS
5214 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5215 rxi_ResetCall(call, 0);
5218 rxi_ResetCall(call, 0);
5220 call->error = error;
5223 /* Reset various fields in a call structure, and wakeup waiting
5224 * processes. Some fields aren't changed: state & mode are not
5225 * touched (these must be set by the caller), and bufptr, nLeft, and
5226 * nFree are not reset, since these fields are manipulated by
5227 * unprotected macros, and may only be reset by non-interrupting code.
5231 rxi_ResetCall(struct rx_call *call, int newcall)
5234 struct rx_peer *peer;
5235 struct rx_packet *packet;
5237 MUTEX_ASSERT(&call->lock);
5238 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5240 /* Notify anyone who is waiting for asynchronous packet arrival */
5241 if (call->arrivalProc) {
5242 (*call->arrivalProc) (call, call->arrivalProcHandle,
5243 call->arrivalProcArg);
5244 call->arrivalProc = (void (*)())0;
5248 rxi_CancelGrowMTUEvent(call);
5250 if (call->delayedAbortEvent) {
5251 rxi_CancelDelayedAbortEvent(call);
5252 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5254 rxi_SendCallAbort(call, packet, 0, 1);
5255 rxi_FreePacket(packet);
5260 * Update the peer with the congestion information in this call
5261 * so other calls on this connection can pick up where this call
5262 * left off. If the congestion sequence numbers don't match then
5263 * another call experienced a retransmission.
5265 peer = call->conn->peer;
5266 MUTEX_ENTER(&peer->peer_lock);
5268 if (call->congestSeq == peer->congestSeq) {
5269 peer->cwind = MAX(peer->cwind, call->cwind);
5270 peer->MTU = MAX(peer->MTU, call->MTU);
5271 peer->nDgramPackets =
5272 MAX(peer->nDgramPackets, call->nDgramPackets);
5275 call->abortCode = 0;
5276 call->abortCount = 0;
5278 if (peer->maxDgramPackets > 1) {
5279 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5281 call->MTU = peer->MTU;
5283 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5284 call->ssthresh = rx_maxSendWindow;
5285 call->nDgramPackets = peer->nDgramPackets;
5286 call->congestSeq = peer->congestSeq;
5287 call->rtt = peer->rtt;
5288 call->rtt_dev = peer->rtt_dev;
5289 clock_Zero(&call->rto);
5290 clock_Addmsec(&call->rto,
5291 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5292 MUTEX_EXIT(&peer->peer_lock);
5294 flags = call->flags;
5295 rxi_WaitforTQBusy(call);
5297 rxi_ClearTransmitQueue(call, 1);
5298 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5299 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5303 rxi_ClearReceiveQueue(call);
5304 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5308 call->twind = call->conn->twind[call->channel];
5309 call->rwind = call->conn->rwind[call->channel];
5310 call->nSoftAcked = 0;
5311 call->nextCwind = 0;
5314 call->nCwindAcks = 0;
5315 call->nSoftAcks = 0;
5316 call->nHardAcks = 0;
5318 call->tfirst = call->rnext = call->tnext = 1;
5321 call->lastAcked = 0;
5322 call->localStatus = call->remoteStatus = 0;
5324 if (flags & RX_CALL_READER_WAIT) {
5325 #ifdef RX_ENABLE_LOCKS
5326 CV_BROADCAST(&call->cv_rq);
5328 osi_rxWakeup(&call->rq);
5331 if (flags & RX_CALL_WAIT_PACKETS) {
5332 MUTEX_ENTER(&rx_freePktQ_lock);
5333 rxi_PacketsUnWait(); /* XXX */
5334 MUTEX_EXIT(&rx_freePktQ_lock);
5336 #ifdef RX_ENABLE_LOCKS
5337 CV_SIGNAL(&call->cv_twind);
5339 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5340 osi_rxWakeup(&call->twind);
5343 if (flags & RX_CALL_WAIT_PROC) {
5344 rx_atomic_dec(&rx_nWaiting);
5346 #ifdef RX_ENABLE_LOCKS
5347 /* The following ensures that we don't mess with any queue while some
5348 * other thread might also be doing so. The call_queue_lock field is
5349 * is only modified under the call lock. If the call is in the process
5350 * of being removed from a queue, the call is not locked until the
5351 * the queue lock is dropped and only then is the call_queue_lock field
5352 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5353 * Note that any other routine which removes a call from a queue has to
5354 * obtain the queue lock before examing the queue and removing the call.
5356 if (call->call_queue_lock) {
5357 MUTEX_ENTER(call->call_queue_lock);
5358 if (opr_queue_IsOnQueue(&call->entry)) {
5359 opr_queue_Remove(&call->entry);
5361 MUTEX_EXIT(call->call_queue_lock);
5362 CLEAR_CALL_QUEUE_LOCK(call);
5364 #else /* RX_ENABLE_LOCKS */
5365 if (opr_queue_IsOnQueue(&call->entry)) {
5366 opr_queue_Remove(&call->entry);
5368 #endif /* RX_ENABLE_LOCKS */
5370 rxi_CancelKeepAliveEvent(call);
5371 rxi_CancelDelayedAckEvent(call);
5374 /* Send an acknowledge for the indicated packet (seq,serial) of the
5375 * indicated call, for the indicated reason (reason). This
5376 * acknowledge will specifically acknowledge receiving the packet, and
5377 * will also specify which other packets for this call have been
5378 * received. This routine returns the packet that was used to the
5379 * caller. The caller is responsible for freeing it or re-using it.
5380 * This acknowledgement also returns the highest sequence number
5381 * actually read out by the higher level to the sender; the sender
5382 * promises to keep around packets that have not been read by the
5383 * higher level yet (unless, of course, the sender decides to abort
5384 * the call altogether). Any of p, seq, serial, pflags, or reason may
5385 * be set to zero without ill effect. That is, if they are zero, they
5386 * will not convey any information.
5387 * NOW there is a trailer field, after the ack where it will safely be
5388 * ignored by mundanes, which indicates the maximum size packet this
5389 * host can swallow. */
5391 struct rx_packet *optionalPacket; use to send ack (or null)
5392 int seq; Sequence number of the packet we are acking
5393 int serial; Serial number of the packet
5394 int pflags; Flags field from packet header
5395 int reason; Reason an acknowledge was prompted
5398 #define RX_ZEROS 1024
5399 static char rx_zeros[RX_ZEROS];
5402 rxi_SendAck(struct rx_call *call,
5403 struct rx_packet *optionalPacket, int serial, int reason,
5406 struct rx_ackPacket *ap;
5407 struct rx_packet *p;
5408 struct opr_queue *cursor;
5411 afs_uint32 padbytes = 0;
5412 #ifdef RX_ENABLE_TSFPQ
5413 struct rx_ts_info_t * rx_ts_info;
5417 * Open the receive window once a thread starts reading packets
5419 if (call->rnext > 1) {
5420 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5423 /* Don't attempt to grow MTU if this is a critical ping */
5424 if (reason == RX_ACK_MTU) {
5425 /* keep track of per-call attempts, if we're over max, do in small
5426 * otherwise in larger? set a size to increment by, decrease
5429 if (call->conn->peer->maxPacketSize &&
5430 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5432 padbytes = call->conn->peer->maxPacketSize+16;
5434 padbytes = call->conn->peer->maxMTU + 128;
5436 /* do always try a minimum size ping */
5437 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5439 /* subtract the ack payload */
5440 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5441 reason = RX_ACK_PING;
5444 call->nHardAcks = 0;
5445 call->nSoftAcks = 0;
5446 if (call->rnext > call->lastAcked)
5447 call->lastAcked = call->rnext;
5451 rx_computelen(p, p->length); /* reset length, you never know */
5452 } /* where that's been... */
5453 #ifdef RX_ENABLE_TSFPQ
5455 RX_TS_INFO_GET(rx_ts_info);
5456 if ((p = rx_ts_info->local_special_packet)) {
5457 rx_computelen(p, p->length);
5458 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5459 rx_ts_info->local_special_packet = p;
5460 } else { /* We won't send the ack, but don't panic. */
5461 return optionalPacket;
5465 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5466 /* We won't send the ack, but don't panic. */
5467 return optionalPacket;
5472 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5475 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5476 #ifndef RX_ENABLE_TSFPQ
5477 if (!optionalPacket)
5480 return optionalPacket;
5482 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5483 if (rx_Contiguous(p) < templ) {
5484 #ifndef RX_ENABLE_TSFPQ
5485 if (!optionalPacket)
5488 return optionalPacket;
5493 /* MTUXXX failing to send an ack is very serious. We should */
5494 /* try as hard as possible to send even a partial ack; it's */
5495 /* better than nothing. */
5496 ap = (struct rx_ackPacket *)rx_DataOf(p);
5497 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5498 ap->reason = reason;
5500 /* The skew computation used to be bogus, I think it's better now. */
5501 /* We should start paying attention to skew. XXX */
5502 ap->serial = htonl(serial);
5503 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5506 * First packet not yet forwarded to reader. When ACKALL has been
5507 * sent the peer has been told that all received packets will be
5508 * delivered to the reader. The value 'rnext' is used internally
5509 * to refer to the next packet in the receive queue that must be
5510 * delivered to the reader. From the perspective of the peer it
5511 * already has so report the last sequence number plus one if there
5512 * are packets in the receive queue awaiting processing.
5514 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5515 !opr_queue_IsEmpty(&call->rq)) {
5516 ap->firstPacket = htonl(opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq + 1);
5518 ap->firstPacket = htonl(call->rnext);
5520 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5522 /* No fear of running out of ack packet here because there can only
5523 * be at most one window full of unacknowledged packets. The window
5524 * size must be constrained to be less than the maximum ack size,
5525 * of course. Also, an ack should always fit into a single packet
5526 * -- it should not ever be fragmented. */
5528 for (opr_queue_Scan(&call->rq, cursor)) {
5529 struct rx_packet *rqp
5530 = opr_queue_Entry(cursor, struct rx_packet, entry);
5532 if (!rqp || !call->rq.next
5533 || (rqp->header.seq > (call->rnext + call->rwind))) {
5534 #ifndef RX_ENABLE_TSFPQ
5535 if (!optionalPacket)
5538 rxi_CallError(call, RX_CALL_DEAD);
5539 return optionalPacket;
5542 while (rqp->header.seq > call->rnext + offset)
5543 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5544 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5546 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5547 #ifndef RX_ENABLE_TSFPQ
5548 if (!optionalPacket)
5551 rxi_CallError(call, RX_CALL_DEAD);
5552 return optionalPacket;
5558 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5560 /* Must zero the 3 octets that rx_AckDataSize skips at the end of the
5563 rx_packetwrite(p, rx_AckDataSize(offset) - 3, 3, rx_zeros);
5565 /* these are new for AFS 3.3 */
5566 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5567 templ = htonl(templ);
5568 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5569 templ = htonl(call->conn->peer->ifMTU);
5570 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5571 sizeof(afs_int32), &templ);
5573 /* new for AFS 3.4 */
5574 templ = htonl(call->rwind);
5575 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5576 sizeof(afs_int32), &templ);
5578 /* new for AFS 3.5 */
5579 templ = htonl(call->conn->peer->ifDgramPackets);
5580 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5581 sizeof(afs_int32), &templ);
5583 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5585 p->header.serviceId = call->conn->serviceId;
5586 p->header.cid = (call->conn->cid | call->channel);
5587 p->header.callNumber = *call->callNumber;
5589 p->header.securityIndex = call->conn->securityIndex;
5590 p->header.epoch = call->conn->epoch;
5591 p->header.type = RX_PACKET_TYPE_ACK;
5592 p->header.flags = RX_SLOW_START_OK;
5593 if (reason == RX_ACK_PING)
5594 p->header.flags |= RX_REQUEST_ACK;
5596 while (padbytes > 0) {
5597 if (padbytes > RX_ZEROS) {
5598 rx_packetwrite(p, p->length, RX_ZEROS, rx_zeros);
5599 p->length += RX_ZEROS;
5600 padbytes -= RX_ZEROS;
5602 rx_packetwrite(p, p->length, padbytes, rx_zeros);
5603 p->length += padbytes;
5608 if (call->conn->type == RX_CLIENT_CONNECTION)
5609 p->header.flags |= RX_CLIENT_INITIATED;
5613 if (rxdebug_active) {
5617 len = _snprintf(msg, sizeof(msg),
5618 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5619 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5620 ntohl(ap->serial), ntohl(ap->previousPacket),
5621 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5622 ap->nAcks, ntohs(ap->bufferSpace) );
5626 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5627 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5631 OutputDebugString(msg);
5633 #else /* AFS_NT40_ENV */
5635 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5636 ap->reason, ntohl(ap->previousPacket),
5637 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5639 for (offset = 0; offset < ap->nAcks; offset++)
5640 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5645 #endif /* AFS_NT40_ENV */
5648 int i, nbytes = p->length;
5650 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5651 if (nbytes <= p->wirevec[i].iov_len) {
5654 savelen = p->wirevec[i].iov_len;
5656 p->wirevec[i].iov_len = nbytes;
5658 rxi_Send(call, p, istack);
5659 p->wirevec[i].iov_len = savelen;
5663 nbytes -= p->wirevec[i].iov_len;
5666 if (rx_stats_active)
5667 rx_atomic_inc(&rx_stats.ackPacketsSent);
5668 #ifndef RX_ENABLE_TSFPQ
5669 if (!optionalPacket)
5672 return optionalPacket; /* Return packet for re-use by caller */
5676 struct rx_packet **list;
5681 /* Send all of the packets in the list in single datagram */
5683 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5684 int istack, int moreFlag)
5690 struct rx_connection *conn = call->conn;
5691 struct rx_peer *peer = conn->peer;
5693 MUTEX_ENTER(&peer->peer_lock);
5694 peer->nSent += xmit->len;
5695 if (xmit->resending)
5696 peer->reSends += xmit->len;
5697 MUTEX_EXIT(&peer->peer_lock);
5699 if (rx_stats_active) {
5700 if (xmit->resending)
5701 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5703 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5706 clock_GetTime(&now);
5708 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5712 /* Set the packet flags and schedule the resend events */
5713 /* Only request an ack for the last packet in the list */
5714 for (i = 0; i < xmit->len; i++) {
5715 struct rx_packet *packet = xmit->list[i];
5717 /* Record the time sent */
5718 packet->timeSent = now;
5719 packet->flags |= RX_PKTFLAG_SENT;
5721 /* Ask for an ack on retransmitted packets, on every other packet
5722 * if the peer doesn't support slow start. Ask for an ack on every
5723 * packet until the congestion window reaches the ack rate. */
5724 if (packet->header.serial) {
5727 packet->firstSent = now;
5728 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5729 || (!(call->flags & RX_CALL_SLOW_START_OK)
5730 && (packet->header.seq & 1)))) {
5735 /* Tag this packet as not being the last in this group,
5736 * for the receiver's benefit */
5737 if (i < xmit->len - 1 || moreFlag) {
5738 packet->header.flags |= RX_MORE_PACKETS;
5743 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5746 /* Since we're about to send a data packet to the peer, it's
5747 * safe to nuke any scheduled end-of-packets ack */
5748 rxi_CancelDelayedAckEvent(call);
5750 MUTEX_EXIT(&call->lock);
5751 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5752 if (xmit->len > 1) {
5753 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5755 rxi_SendPacket(call, conn, xmit->list[0], istack);
5757 MUTEX_ENTER(&call->lock);
5758 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5760 /* Tell the RTO calculation engine that we have sent a packet, and
5761 * if it was the last one */
5762 rxi_rto_packet_sent(call, lastPacket, istack);
5764 /* Update last send time for this call (for keep-alive
5765 * processing), and for the connection (so that we can discover
5766 * idle connections) */
5767 conn->lastSendTime = call->lastSendTime = clock_Sec();
5770 /* When sending packets we need to follow these rules:
5771 * 1. Never send more than maxDgramPackets in a jumbogram.
5772 * 2. Never send a packet with more than two iovecs in a jumbogram.
5773 * 3. Never send a retransmitted packet in a jumbogram.
5774 * 4. Never send more than cwind/4 packets in a jumbogram
5775 * We always keep the last list we should have sent so we
5776 * can set the RX_MORE_PACKETS flags correctly.
5780 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5785 struct xmitlist working;
5786 struct xmitlist last;
5788 struct rx_peer *peer = call->conn->peer;
5789 int morePackets = 0;
5791 memset(&last, 0, sizeof(struct xmitlist));
5792 working.list = &list[0];
5794 working.resending = 0;
5796 recovery = call->flags & RX_CALL_FAST_RECOVER;
5798 for (i = 0; i < len; i++) {
5799 /* Does the current packet force us to flush the current list? */
5801 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5802 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5804 /* This sends the 'last' list and then rolls the current working
5805 * set into the 'last' one, and resets the working set */
5808 rxi_SendList(call, &last, istack, 1);
5809 /* If the call enters an error state stop sending, or if
5810 * we entered congestion recovery mode, stop sending */
5812 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5817 working.resending = 0;
5818 working.list = &list[i];
5820 /* Add the current packet to the list if it hasn't been acked.
5821 * Otherwise adjust the list pointer to skip the current packet. */
5822 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5825 if (list[i]->header.serial)
5826 working.resending = 1;
5828 /* Do we need to flush the list? */
5829 if (working.len >= (int)peer->maxDgramPackets
5830 || working.len >= (int)call->nDgramPackets
5831 || working.len >= (int)call->cwind
5832 || list[i]->header.serial
5833 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5835 rxi_SendList(call, &last, istack, 1);
5836 /* If the call enters an error state stop sending, or if
5837 * we entered congestion recovery mode, stop sending */
5839 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5844 working.resending = 0;
5845 working.list = &list[i + 1];
5848 if (working.len != 0) {
5849 osi_Panic("rxi_SendList error");
5851 working.list = &list[i + 1];
5855 /* Send the whole list when the call is in receive mode, when
5856 * the call is in eof mode, when we are in fast recovery mode,
5857 * and when we have the last packet */
5858 /* XXX - The accesses to app.mode aren't safe, as this may be called by
5859 * the listener or event threads
5861 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5862 || (call->flags & RX_CALL_FLUSH)
5863 || (call->flags & RX_CALL_FAST_RECOVER)) {
5864 /* Check for the case where the current list contains
5865 * an acked packet. Since we always send retransmissions
5866 * in a separate packet, we only need to check the first
5867 * packet in the list */
5868 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5872 rxi_SendList(call, &last, istack, morePackets);
5873 /* If the call enters an error state stop sending, or if
5874 * we entered congestion recovery mode, stop sending */
5876 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5880 rxi_SendList(call, &working, istack, 0);
5882 } else if (last.len > 0) {
5883 rxi_SendList(call, &last, istack, 0);
5884 /* Packets which are in 'working' are not sent by this call */
5889 * Check if the peer for the given call is known to be dead
5891 * If the call's peer appears dead (it has encountered fatal network errors
5892 * since the call started) the call is killed with RX_CALL_DEAD if the call
5893 * is active. Otherwise, we do nothing.
5895 * @param[in] call The call to check
5898 * @retval 0 The call is fine, and we haven't done anything to the call
5899 * @retval nonzero The call's peer appears dead, and the call has been
5900 * terminated if it was active
5902 * @pre call->lock must be locked
5905 rxi_CheckPeerDead(struct rx_call *call)
5907 #ifdef AFS_RXERRQ_ENV
5910 if (call->state == RX_STATE_DALLY) {
5914 peererrs = rx_atomic_read(&call->conn->peer->neterrs);
5915 if (call->neterr_gen < peererrs) {
5916 /* we have received network errors since this call started; kill
5918 if (call->state == RX_STATE_ACTIVE) {
5919 rxi_CallError(call, RX_CALL_DEAD);
5923 if (call->neterr_gen > peererrs) {
5924 /* someone has reset the number of peer errors; set the call error gen
5925 * so we can detect if more errors are encountered */
5926 call->neterr_gen = peererrs;
5933 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
5935 struct rx_call *call = arg0;
5936 struct rx_peer *peer;
5937 struct opr_queue *cursor;
5938 struct clock maxTimeout = { 60, 0 };
5940 MUTEX_ENTER(&call->lock);
5942 peer = call->conn->peer;
5944 /* Make sure that the event pointer is removed from the call
5945 * structure, since there is no longer a per-call retransmission
5947 if (event == call->resendEvent)
5948 rxevent_Put(&call->resendEvent);
5950 rxi_CheckPeerDead(call);
5952 if (opr_queue_IsEmpty(&call->tq)) {
5953 /* Nothing to do. This means that we've been raced, and that an
5954 * ACK has come in between when we were triggered, and when we
5955 * actually got to run. */
5959 /* We're in loss recovery */
5960 call->flags |= RX_CALL_FAST_RECOVER;
5962 /* Mark all of the pending packets in the queue as being lost */
5963 for (opr_queue_Scan(&call->tq, cursor)) {
5964 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
5965 if (!(p->flags & RX_PKTFLAG_ACKED))
5966 p->flags &= ~RX_PKTFLAG_SENT;
5969 /* We're resending, so we double the timeout of the call. This will be
5970 * dropped back down by the first successful ACK that we receive.
5972 * We apply a maximum value here of 60 seconds
5974 clock_Add(&call->rto, &call->rto);
5975 if (clock_Gt(&call->rto, &maxTimeout))
5976 call->rto = maxTimeout;
5978 /* Packet loss is most likely due to congestion, so drop our window size
5979 * and start again from the beginning */
5980 if (peer->maxDgramPackets >1) {
5981 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5982 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5984 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5985 call->nDgramPackets = 1;
5987 call->nextCwind = 1;
5990 MUTEX_ENTER(&peer->peer_lock);
5991 peer->MTU = call->MTU;
5992 peer->cwind = call->cwind;
5993 peer->nDgramPackets = 1;
5995 call->congestSeq = peer->congestSeq;
5996 MUTEX_EXIT(&peer->peer_lock);
5998 rxi_Start(call, istack);
6001 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
6002 MUTEX_EXIT(&call->lock);
6005 /* This routine is called when new packets are readied for
6006 * transmission and when retransmission may be necessary, or when the
6007 * transmission window or burst count are favourable. This should be
6008 * better optimized for new packets, the usual case, now that we've
6009 * got rid of queues of send packets. XXXXXXXXXXX */
6011 rxi_Start(struct rx_call *call, int istack)
6013 struct opr_queue *cursor;
6014 #ifdef RX_ENABLE_LOCKS
6015 struct opr_queue *store;
6021 #ifdef RX_ENABLE_LOCKS
6022 if (rx_stats_active)
6023 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
6028 if (!opr_queue_IsEmpty(&call->tq)) { /* If we have anything to send */
6029 /* Send (or resend) any packets that need it, subject to
6030 * window restrictions and congestion burst control
6031 * restrictions. Ask for an ack on the last packet sent in
6032 * this burst. For now, we're relying upon the window being
6033 * considerably bigger than the largest number of packets that
6034 * are typically sent at once by one initial call to
6035 * rxi_Start. This is probably bogus (perhaps we should ask
6036 * for an ack when we're half way through the current
6037 * window?). Also, for non file transfer applications, this
6038 * may end up asking for an ack for every packet. Bogus. XXXX
6041 * But check whether we're here recursively, and let the other guy
6044 #ifdef RX_ENABLE_LOCKS
6045 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6046 call->flags |= RX_CALL_TQ_BUSY;
6048 #endif /* RX_ENABLE_LOCKS */
6050 #ifdef RX_ENABLE_LOCKS
6051 call->flags &= ~RX_CALL_NEED_START;
6052 #endif /* RX_ENABLE_LOCKS */
6054 maxXmitPackets = MIN(call->twind, call->cwind);
6055 for (opr_queue_Scan(&call->tq, cursor)) {
6057 = opr_queue_Entry(cursor, struct rx_packet, entry);
6059 if (p->flags & RX_PKTFLAG_ACKED) {
6060 /* Since we may block, don't trust this */
6061 if (rx_stats_active)
6062 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6063 continue; /* Ignore this packet if it has been acknowledged */
6066 /* Turn off all flags except these ones, which are the same
6067 * on each transmission */
6068 p->header.flags &= RX_PRESET_FLAGS;
6070 if (p->header.seq >=
6071 call->tfirst + MIN((int)call->twind,
6072 (int)(call->nSoftAcked +
6074 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6075 /* Note: if we're waiting for more window space, we can
6076 * still send retransmits; hence we don't return here, but
6077 * break out to schedule a retransmit event */
6078 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6079 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6084 /* Transmit the packet if it needs to be sent. */
6085 if (!(p->flags & RX_PKTFLAG_SENT)) {
6086 if (nXmitPackets == maxXmitPackets) {
6087 rxi_SendXmitList(call, call->xmitList,
6088 nXmitPackets, istack);
6091 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6092 *(call->callNumber), p));
6093 call->xmitList[nXmitPackets++] = p;
6095 } /* end of the queue_Scan */
6097 /* xmitList now hold pointers to all of the packets that are
6098 * ready to send. Now we loop to send the packets */
6099 if (nXmitPackets > 0) {
6100 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6104 #ifdef RX_ENABLE_LOCKS
6106 /* We went into the error state while sending packets. Now is
6107 * the time to reset the call. This will also inform the using
6108 * process that the call is in an error state.
6110 if (rx_stats_active)
6111 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6112 call->flags &= ~RX_CALL_TQ_BUSY;
6113 rxi_WakeUpTransmitQueue(call);
6114 rxi_CallError(call, call->error);
6118 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6120 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6121 /* Some packets have received acks. If they all have, we can clear
6122 * the transmit queue.
6125 for (opr_queue_ScanSafe(&call->tq, cursor, store)) {
6127 = opr_queue_Entry(cursor, struct rx_packet, entry);
6129 if (p->header.seq < call->tfirst
6130 && (p->flags & RX_PKTFLAG_ACKED)) {
6131 opr_queue_Remove(&p->entry);
6132 #ifdef RX_TRACK_PACKETS
6133 p->flags &= ~RX_PKTFLAG_TQ;
6135 #ifdef RXDEBUG_PACKET
6143 call->flags |= RX_CALL_TQ_CLEARME;
6145 if (call->flags & RX_CALL_TQ_CLEARME)
6146 rxi_ClearTransmitQueue(call, 1);
6147 } while (call->flags & RX_CALL_NEED_START);
6149 * TQ references no longer protected by this flag; they must remain
6150 * protected by the call lock.
6152 call->flags &= ~RX_CALL_TQ_BUSY;
6153 rxi_WakeUpTransmitQueue(call);
6155 call->flags |= RX_CALL_NEED_START;
6157 #endif /* RX_ENABLE_LOCKS */
6159 rxi_rto_cancel(call);
6163 /* Also adjusts the keep alive parameters for the call, to reflect
6164 * that we have just sent a packet (so keep alives aren't sent
6167 rxi_Send(struct rx_call *call, struct rx_packet *p,
6170 struct rx_connection *conn = call->conn;
6172 /* Stamp each packet with the user supplied status */
6173 p->header.userStatus = call->localStatus;
6175 /* Allow the security object controlling this call's security to
6176 * make any last-minute changes to the packet */
6177 RXS_SendPacket(conn->securityObject, call, p);
6179 /* Since we're about to send SOME sort of packet to the peer, it's
6180 * safe to nuke any scheduled end-of-packets ack */
6181 rxi_CancelDelayedAckEvent(call);
6183 /* Actually send the packet, filling in more connection-specific fields */
6184 MUTEX_EXIT(&call->lock);
6185 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6186 rxi_SendPacket(call, conn, p, istack);
6187 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6188 MUTEX_ENTER(&call->lock);
6190 /* Update last send time for this call (for keep-alive
6191 * processing), and for the connection (so that we can discover
6192 * idle connections) */
6193 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6194 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6195 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6197 conn->lastSendTime = call->lastSendTime = clock_Sec();
6201 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6202 * that things are fine. Also called periodically to guarantee that nothing
6203 * falls through the cracks (e.g. (error + dally) connections have keepalive
6204 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6206 * haveCTLock Set if calling from rxi_ReapConnections
6209 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6211 struct rx_connection *conn = call->conn;
6213 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6214 afs_uint32 fudgeFactor;
6217 int idle_timeout = 0;
6218 afs_int32 clock_diff = 0;
6220 if (rxi_CheckPeerDead(call)) {
6226 /* Large swings in the clock can have a significant impact on
6227 * the performance of RX call processing. Forward clock shifts
6228 * will result in premature event triggering or timeouts.
6229 * Backward shifts can result in calls not completing until
6230 * the clock catches up with the original start clock value.
6232 * If a backward clock shift of more than five minutes is noticed,
6233 * just fail the call.
6235 if (now < call->lastSendTime)
6236 clock_diff = call->lastSendTime - now;
6237 if (now < call->startWait)
6238 clock_diff = MAX(clock_diff, call->startWait - now);
6239 if (now < call->lastReceiveTime)
6240 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6241 if (clock_diff > 5 * 60)
6243 if (call->state == RX_STATE_ACTIVE)
6244 rxi_CallError(call, RX_CALL_TIMEOUT);
6248 #ifdef RX_ENABLE_LOCKS
6249 if (call->flags & RX_CALL_TQ_BUSY) {
6250 /* Call is active and will be reset by rxi_Start if it's
6251 * in an error state.
6256 /* RTT + 8*MDEV, rounded up to the next second. */
6257 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6258 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6260 deadTime = conn->secondsUntilDead + fudgeFactor;
6261 /* These are computed to the second (+- 1 second). But that's
6262 * good enough for these values, which should be a significant
6263 * number of seconds. */
6264 if (now > (call->lastReceiveTime + deadTime)) {
6265 if (call->state == RX_STATE_ACTIVE) {
6266 cerror = RX_CALL_DEAD;
6269 #ifdef RX_ENABLE_LOCKS
6270 /* Cancel pending events */
6271 rxi_CancelDelayedAckEvent(call);
6272 rxi_rto_cancel(call);
6273 rxi_CancelKeepAliveEvent(call);
6274 rxi_CancelGrowMTUEvent(call);
6275 MUTEX_ENTER(&rx_refcnt_mutex);
6276 /* if rxi_FreeCall returns 1 it has freed the call */
6277 if (call->refCount == 0 &&
6278 rxi_FreeCall(call, haveCTLock))
6280 MUTEX_EXIT(&rx_refcnt_mutex);
6283 MUTEX_EXIT(&rx_refcnt_mutex);
6285 #else /* RX_ENABLE_LOCKS */
6286 rxi_FreeCall(call, 0);
6288 #endif /* RX_ENABLE_LOCKS */
6290 /* Non-active calls are destroyed if they are not responding
6291 * to pings; active calls are simply flagged in error, so the
6292 * attached process can die reasonably gracefully. */
6295 if (conn->idleDeadTime) {
6296 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6300 /* see if we have a non-activity timeout */
6301 if (call->startWait && ((call->startWait + idleDeadTime) < now)) {
6302 if (call->state == RX_STATE_ACTIVE) {
6303 cerror = RX_CALL_TIMEOUT;
6309 if (conn->hardDeadTime) {
6310 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6313 /* see if we have a hard timeout */
6315 && (now > (hardDeadTime + call->startTime.sec))) {
6316 if (call->state == RX_STATE_ACTIVE)
6317 rxi_CallError(call, RX_CALL_TIMEOUT);
6322 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6323 call->lastReceiveTime) {
6324 int oldMTU = conn->peer->ifMTU;
6326 /* If we thought we could send more, perhaps things got worse.
6327 * Shrink by 128 bytes and try again. */
6328 if (conn->peer->maxPacketSize < conn->lastPacketSize)
6329 /* maxPacketSize will be cleared in rxi_SetPeerMtu */
6330 newmtu = MAX(conn->peer->maxPacketSize + RX_HEADER_SIZE,
6331 conn->lastPacketSize - 128 + RX_HEADER_SIZE);
6333 newmtu = conn->lastPacketSize - 128 + RX_HEADER_SIZE;
6335 /* minimum capped in SetPeerMtu */
6336 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6339 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
6341 /* needed so ResetCall doesn't clobber us. */
6342 call->MTU = conn->peer->ifMTU;
6344 /* if we never succeeded, let the error pass out as-is */
6345 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6346 cerror = conn->msgsizeRetryErr;
6349 rxi_CallError(call, cerror);
6354 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6355 void *dummy, int dummy2)
6357 struct rx_connection *conn = arg1;
6358 struct rx_header theader;
6359 char tbuffer[1 + sizeof(struct rx_header)];
6360 struct sockaddr_in taddr;
6364 struct iovec tmpiov[2];
6367 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6370 tp = &tbuffer[sizeof(struct rx_header)];
6371 taddr.sin_family = AF_INET;
6372 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6373 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6374 memset(&taddr.sin_zero, 0, sizeof(taddr.sin_zero));
6375 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6376 taddr.sin_len = sizeof(struct sockaddr_in);
6378 memset(&theader, 0, sizeof(theader));
6379 theader.epoch = htonl(999);
6381 theader.callNumber = 0;
6384 theader.type = RX_PACKET_TYPE_VERSION;
6385 theader.flags = RX_LAST_PACKET;
6386 theader.serviceId = 0;
6388 memcpy(tbuffer, &theader, sizeof(theader));
6389 memcpy(tp, &a, sizeof(a));
6390 tmpiov[0].iov_base = tbuffer;
6391 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6393 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6395 MUTEX_ENTER(&conn->conn_data_lock);
6396 /* We ran, so the handle is no longer needed to try to cancel ourselves. */
6397 if (event == conn->natKeepAliveEvent)
6398 rxevent_Put(&conn->natKeepAliveEvent);
6399 MUTEX_ENTER(&rx_refcnt_mutex);
6400 /* Only reschedule ourselves if the connection would not be destroyed */
6401 if (conn->refCount > 1)
6403 if (conn->refCount <= 0) {
6404 #ifdef RX_REFCOUNT_CHECK
6405 osi_Assert(conn->refCount == 0);
6407 if (rx_stats_active) {
6408 MUTEX_ENTER(&rx_stats_mutex);
6409 rxi_lowConnRefCount++;
6410 MUTEX_EXIT(&rx_stats_mutex);
6413 MUTEX_EXIT(&rx_refcnt_mutex);
6415 rxi_ScheduleNatKeepAliveEvent(conn);
6416 MUTEX_EXIT(&conn->conn_data_lock);
6417 putConnection(conn);
6421 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6423 MUTEX_ASSERT(&conn->conn_data_lock);
6424 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6425 struct clock when, now;
6426 clock_GetTime(&now);
6428 when.sec += conn->secondsUntilNatPing;
6429 rx_GetConnection(conn);
6430 conn->natKeepAliveEvent =
6431 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6436 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6438 MUTEX_ENTER(&conn->conn_data_lock);
6439 conn->secondsUntilNatPing = seconds;
6441 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6442 rxi_ScheduleNatKeepAliveEvent(conn);
6444 conn->flags |= RX_CONN_NAT_PING;
6446 MUTEX_EXIT(&conn->conn_data_lock);
6449 /* When a call is in progress, this routine is called occasionally to
6450 * make sure that some traffic has arrived (or been sent to) the peer.
6451 * If nothing has arrived in a reasonable amount of time, the call is
6452 * declared dead; if nothing has been sent for a while, we send a
6453 * keep-alive packet (if we're actually trying to keep the call alive)
6456 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6459 struct rx_call *call = arg1;
6460 struct rx_connection *conn;
6463 MUTEX_ENTER(&call->lock);
6465 if (event == call->keepAliveEvent)
6466 rxevent_Put(&call->keepAliveEvent);
6470 if (rxi_CheckCall(call, 0)) {
6471 MUTEX_EXIT(&call->lock);
6475 /* Don't try to keep alive dallying calls */
6476 if (call->state == RX_STATE_DALLY) {
6477 MUTEX_EXIT(&call->lock);
6478 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6483 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6484 /* Don't try to send keepalives if there is unacknowledged data */
6485 /* the rexmit code should be good enough, this little hack
6486 * doesn't quite work XXX */
6487 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6489 rxi_ScheduleKeepAliveEvent(call);
6490 MUTEX_EXIT(&call->lock);
6491 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6494 /* Does what's on the nameplate. */
6496 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6498 struct rx_call *call = arg1;
6499 struct rx_connection *conn;
6501 MUTEX_ENTER(&call->lock);
6503 if (event == call->growMTUEvent)
6504 rxevent_Put(&call->growMTUEvent);
6506 if (rxi_CheckCall(call, 0))
6509 /* Don't bother with dallying calls */
6510 if (call->state == RX_STATE_DALLY)
6516 * keep being scheduled, just don't do anything if we're at peak,
6517 * or we're not set up to be properly handled (idle timeout required)
6519 if ((conn->peer->maxPacketSize != 0) &&
6520 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6522 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6523 rxi_ScheduleGrowMTUEvent(call, 0);
6525 MUTEX_EXIT(&call->lock);
6526 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6530 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6532 MUTEX_ASSERT(&call->lock);
6533 if (!call->keepAliveEvent) {
6534 struct clock when, now;
6535 clock_GetTime(&now);
6537 when.sec += call->conn->secondsUntilPing;
6538 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6539 call->keepAliveEvent =
6540 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6545 rxi_CancelKeepAliveEvent(struct rx_call *call) {
6546 MUTEX_ASSERT(&call->lock);
6547 if (rxevent_Cancel(&call->keepAliveEvent))
6548 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6552 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6554 MUTEX_ASSERT(&call->lock);
6555 if (!call->growMTUEvent) {
6556 struct clock when, now;
6558 clock_GetTime(&now);
6561 if (call->conn->secondsUntilPing)
6562 secs = (6*call->conn->secondsUntilPing)-1;
6564 if (call->conn->secondsUntilDead)
6565 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6569 CALL_HOLD(call, RX_CALL_REFCOUNT_MTU);
6570 call->growMTUEvent =
6571 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6576 rxi_CancelGrowMTUEvent(struct rx_call *call)
6578 MUTEX_ASSERT(&call->lock);
6579 if (rxevent_Cancel(&call->growMTUEvent))
6580 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6584 * Increment the counter for the next connection ID, handling overflow.
6587 update_nextCid(void)
6589 /* Overflow is technically undefined behavior; avoid it. */
6590 if (rx_nextCid > MAX_AFS_INT32 - (1 << RX_CIDSHIFT))
6591 rx_nextCid = -1 * ((MAX_AFS_INT32 / RX_CIDSHIFT) * RX_CIDSHIFT);
6593 rx_nextCid += 1 << RX_CIDSHIFT;
6597 rxi_KeepAliveOn(struct rx_call *call)
6599 /* Pretend last packet received was received now--i.e. if another
6600 * packet isn't received within the keep alive time, then the call
6601 * will die; Initialize last send time to the current time--even
6602 * if a packet hasn't been sent yet. This will guarantee that a
6603 * keep-alive is sent within the ping time */
6604 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6605 rxi_ScheduleKeepAliveEvent(call);
6609 rxi_GrowMTUOn(struct rx_call *call)
6611 struct rx_connection *conn = call->conn;
6612 MUTEX_ENTER(&conn->conn_data_lock);
6613 conn->lastPingSizeSer = conn->lastPingSize = 0;
6614 MUTEX_EXIT(&conn->conn_data_lock);
6615 rxi_ScheduleGrowMTUEvent(call, 1);
6618 /* This routine is called to send connection abort messages
6619 * that have been delayed to throttle looping clients. */
6621 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6624 struct rx_connection *conn = arg1;
6627 struct rx_packet *packet;
6629 MUTEX_ENTER(&conn->conn_data_lock);
6630 if (event == conn->delayedAbortEvent)
6631 rxevent_Put(&conn->delayedAbortEvent);
6632 error = htonl(conn->error);
6634 MUTEX_EXIT(&conn->conn_data_lock);
6635 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6638 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6639 RX_PACKET_TYPE_ABORT, (char *)&error,
6641 rxi_FreePacket(packet);
6643 putConnection(conn);
6646 /* This routine is called to send call abort messages
6647 * that have been delayed to throttle looping clients. */
6649 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6652 struct rx_call *call = arg1;
6655 struct rx_packet *packet;
6657 MUTEX_ENTER(&call->lock);
6658 if (event == call->delayedAbortEvent)
6659 rxevent_Put(&call->delayedAbortEvent);
6660 error = htonl(call->error);
6662 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6665 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6666 (char *)&error, sizeof(error), 0);
6667 rxi_FreePacket(packet);
6669 MUTEX_EXIT(&call->lock);
6670 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6674 * This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6675 * seconds) to ask the client to authenticate itself. The routine
6676 * issues a challenge to the client, which is obtained from the
6677 * security object associated with the connection
6679 * This routine is both an event handler and a function called directly;
6680 * when called directly the passed |event| is NULL and the
6681 * conn->conn->data>lock must must not be held.
6684 rxi_ChallengeEvent(struct rxevent *event,
6685 void *arg0, void *arg1, int tries)
6687 struct rx_connection *conn = arg0;
6689 MUTEX_ENTER(&conn->conn_data_lock);
6690 if (event != NULL && event == conn->challengeEvent)
6691 rxevent_Put(&conn->challengeEvent);
6692 MUTEX_EXIT(&conn->conn_data_lock);
6694 /* If there are no active calls it is not worth re-issuing the
6695 * challenge. If the client issues another call on this connection
6696 * the challenge can be requested at that time.
6698 if (!rxi_HasActiveCalls(conn)) {
6699 putConnection(conn);
6703 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6704 struct rx_packet *packet;
6705 struct clock when, now;
6708 /* We've failed to authenticate for too long.
6709 * Reset any calls waiting for authentication;
6710 * they are all in RX_STATE_PRECALL.
6714 MUTEX_ENTER(&conn->conn_call_lock);
6715 for (i = 0; i < RX_MAXCALLS; i++) {
6716 struct rx_call *call = conn->call[i];
6718 MUTEX_ENTER(&call->lock);
6719 if (call->state == RX_STATE_PRECALL) {
6720 rxi_CallError(call, RX_CALL_DEAD);
6721 rxi_SendCallAbort(call, NULL, 0, 0);
6723 MUTEX_EXIT(&call->lock);
6726 MUTEX_EXIT(&conn->conn_call_lock);
6727 putConnection(conn);
6731 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6733 /* If there's no packet available, do this later. */
6734 RXS_GetChallenge(conn->securityObject, conn, packet);
6735 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6736 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6737 rxi_FreePacket(packet);
6738 conn->securityChallengeSent = 1;
6740 clock_GetTime(&now);
6742 when.sec += RX_CHALLENGE_TIMEOUT;
6743 MUTEX_ENTER(&conn->conn_data_lock);
6744 /* Only reschedule ourselves if not already pending. */
6745 if (conn->challengeEvent == NULL) {
6746 rx_GetConnection(conn);
6747 conn->challengeEvent =
6748 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6751 MUTEX_EXIT(&conn->conn_data_lock);
6753 putConnection(conn);
6756 /* Call this routine to start requesting the client to authenticate
6757 * itself. This will continue until authentication is established,
6758 * the call times out, or an invalid response is returned. The
6759 * security object associated with the connection is asked to create
6760 * the challenge at this time. */
6762 rxi_ChallengeOn(struct rx_connection *conn)
6765 MUTEX_ENTER(&conn->conn_data_lock);
6766 if (!conn->challengeEvent)
6768 MUTEX_EXIT(&conn->conn_data_lock);
6770 RXS_CreateChallenge(conn->securityObject, conn);
6771 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6776 /* rxi_ComputeRoundTripTime is called with peer locked. */
6777 /* peer may be null */
6779 rxi_ComputeRoundTripTime(struct rx_packet *p,
6780 struct rx_ackPacket *ack,
6781 struct rx_call *call,
6782 struct rx_peer *peer,
6785 struct clock thisRtt, *sentp;
6789 /* If the ACK is delayed, then do nothing */
6790 if (ack->reason == RX_ACK_DELAY)
6793 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6794 * their RTT multiple times, so only include the RTT of the last packet
6796 if (p->flags & RX_JUMBO_PACKET)
6799 /* Use the serial number to determine which transmission the ACK is for,
6800 * and set the sent time to match this. If we have no serial number, then
6801 * only use the ACK for RTT calculations if the packet has not been
6805 serial = ntohl(ack->serial);
6807 if (serial == p->header.serial) {
6808 sentp = &p->timeSent;
6809 } else if (serial == p->firstSerial) {
6810 sentp = &p->firstSent;
6811 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6812 sentp = &p->firstSent;
6816 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6817 sentp = &p->firstSent;
6824 if (clock_Lt(&thisRtt, sentp))
6825 return; /* somebody set the clock back, don't count this time. */
6827 clock_Sub(&thisRtt, sentp);
6828 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6829 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6831 if (clock_IsZero(&thisRtt)) {
6833 * The actual round trip time is shorter than the
6834 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6835 * Since we can't tell which at the moment we will assume 1ms.
6837 thisRtt.usec = 1000;
6840 if (rx_stats_active) {
6841 MUTEX_ENTER(&rx_stats_mutex);
6842 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6843 rx_stats.minRtt = thisRtt;
6844 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6845 if (thisRtt.sec > 60) {
6846 MUTEX_EXIT(&rx_stats_mutex);
6847 return; /* somebody set the clock ahead */
6849 rx_stats.maxRtt = thisRtt;
6851 clock_Add(&rx_stats.totalRtt, &thisRtt);
6852 rx_atomic_inc(&rx_stats.nRttSamples);
6853 MUTEX_EXIT(&rx_stats_mutex);
6856 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6858 /* Apply VanJacobson round-trip estimations */
6863 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6864 * srtt is stored as fixed point with 3 bits after the binary
6865 * point (i.e., scaled by 8). The following magic is
6866 * equivalent to the smoothing algorithm in rfc793 with an
6867 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6868 * srtt'*8 = rtt + srtt*7
6869 * srtt'*8 = srtt*8 + rtt - srtt
6870 * srtt' = srtt + rtt/8 - srtt/8
6871 * srtt' = srtt + (rtt - srtt)/8
6874 delta = _8THMSEC(&thisRtt) - call->rtt;
6875 call->rtt += (delta >> 3);
6878 * We accumulate a smoothed rtt variance (actually, a smoothed
6879 * mean difference), then set the retransmit timer to smoothed
6880 * rtt + 4 times the smoothed variance (was 2x in van's original
6881 * paper, but 4x works better for me, and apparently for him as
6883 * rttvar is stored as
6884 * fixed point with 2 bits after the binary point (scaled by
6885 * 4). The following is equivalent to rfc793 smoothing with
6886 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6887 * rttvar'*4 = rttvar*3 + |delta|
6888 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6889 * rttvar' = rttvar + |delta|/4 - rttvar/4
6890 * rttvar' = rttvar + (|delta| - rttvar)/4
6891 * This replaces rfc793's wired-in beta.
6892 * dev*4 = dev*4 + (|actual - expected| - dev)
6898 delta -= (call->rtt_dev << 1);
6899 call->rtt_dev += (delta >> 3);
6901 /* I don't have a stored RTT so I start with this value. Since I'm
6902 * probably just starting a call, and will be pushing more data down
6903 * this, I expect congestion to increase rapidly. So I fudge a
6904 * little, and I set deviance to half the rtt. In practice,
6905 * deviance tends to approach something a little less than
6906 * half the smoothed rtt. */
6907 call->rtt = _8THMSEC(&thisRtt) + 8;
6908 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6910 /* the smoothed RTT time is RTT + 4*MDEV
6912 * We allow a user specified minimum to be set for this, to allow clamping
6913 * at a minimum value in the same way as TCP. In addition, we have to allow
6914 * for the possibility that this packet is answered by a delayed ACK, so we
6915 * add on a fixed 200ms to account for that timer expiring.
6918 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6919 rx_minPeerTimeout) + 200;
6920 clock_Zero(&call->rto);
6921 clock_Addmsec(&call->rto, rtt_timeout);
6923 /* Update the peer, so any new calls start with our values */
6924 peer->rtt_dev = call->rtt_dev;
6925 peer->rtt = call->rtt;
6927 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6928 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6932 /* Find all server connections that have not been active for a long time, and
6935 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
6938 struct clock now, when;
6939 struct rxevent *event;
6940 clock_GetTime(&now);
6942 /* Find server connection structures that haven't been used for
6943 * greater than rx_idleConnectionTime */
6945 struct rx_connection **conn_ptr, **conn_end;
6946 int i, havecalls = 0;
6947 MUTEX_ENTER(&rx_connHashTable_lock);
6948 for (conn_ptr = &rx_connHashTable[0], conn_end =
6949 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6951 struct rx_connection *conn, *next;
6952 struct rx_call *call;
6956 for (conn = *conn_ptr; conn; conn = next) {
6957 /* XXX -- Shouldn't the connection be locked? */
6960 for (i = 0; i < RX_MAXCALLS; i++) {
6961 call = conn->call[i];
6965 code = MUTEX_TRYENTER(&call->lock);
6968 result = rxi_CheckCall(call, 1);
6969 MUTEX_EXIT(&call->lock);
6971 /* If CheckCall freed the call, it might
6972 * have destroyed the connection as well,
6973 * which screws up the linked lists.
6979 if (conn->type == RX_SERVER_CONNECTION) {
6980 /* This only actually destroys the connection if
6981 * there are no outstanding calls */
6982 MUTEX_ENTER(&conn->conn_data_lock);
6983 MUTEX_ENTER(&rx_refcnt_mutex);
6984 if (!havecalls && !conn->refCount
6985 && ((conn->lastSendTime + rx_idleConnectionTime) <
6987 conn->refCount++; /* it will be decr in rx_DestroyConn */
6988 MUTEX_EXIT(&rx_refcnt_mutex);
6989 MUTEX_EXIT(&conn->conn_data_lock);
6990 #ifdef RX_ENABLE_LOCKS
6991 rxi_DestroyConnectionNoLock(conn);
6992 #else /* RX_ENABLE_LOCKS */
6993 rxi_DestroyConnection(conn);
6994 #endif /* RX_ENABLE_LOCKS */
6996 #ifdef RX_ENABLE_LOCKS
6998 MUTEX_EXIT(&rx_refcnt_mutex);
6999 MUTEX_EXIT(&conn->conn_data_lock);
7001 #endif /* RX_ENABLE_LOCKS */
7005 #ifdef RX_ENABLE_LOCKS
7006 while (rx_connCleanup_list) {
7007 struct rx_connection *conn;
7008 conn = rx_connCleanup_list;
7009 rx_connCleanup_list = rx_connCleanup_list->next;
7010 MUTEX_EXIT(&rx_connHashTable_lock);
7011 rxi_CleanupConnection(conn);
7012 MUTEX_ENTER(&rx_connHashTable_lock);
7014 MUTEX_EXIT(&rx_connHashTable_lock);
7015 #endif /* RX_ENABLE_LOCKS */
7018 /* Find any peer structures that haven't been used (haven't had an
7019 * associated connection) for greater than rx_idlePeerTime */
7021 struct rx_peer **peer_ptr, **peer_end;
7025 * Why do we need to hold the rx_peerHashTable_lock across
7026 * the incrementing of peer_ptr since the rx_peerHashTable
7027 * array is not changing? We don't.
7029 * By dropping the lock periodically we can permit other
7030 * activities to be performed while a rxi_ReapConnections
7031 * call is in progress. The goal of reap connections
7032 * is to clean up quickly without causing large amounts
7033 * of contention. Therefore, it is important that global
7034 * mutexes not be held for extended periods of time.
7036 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7037 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7039 struct rx_peer *peer, *next, *prev;
7041 MUTEX_ENTER(&rx_peerHashTable_lock);
7042 for (prev = peer = *peer_ptr; peer; peer = next) {
7044 code = MUTEX_TRYENTER(&peer->peer_lock);
7045 if ((code) && (peer->refCount == 0)
7046 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7047 struct opr_queue *cursor, *store;
7051 * now know that this peer object is one to be
7052 * removed from the hash table. Once it is removed
7053 * it can't be referenced by other threads.
7054 * Lets remove it first and decrement the struct
7055 * nPeerStructs count.
7057 if (peer == *peer_ptr) {
7063 if (rx_stats_active)
7064 rx_atomic_dec(&rx_stats.nPeerStructs);
7067 * Now if we hold references on 'prev' and 'next'
7068 * we can safely drop the rx_peerHashTable_lock
7069 * while we destroy this 'peer' object.
7075 MUTEX_EXIT(&rx_peerHashTable_lock);
7077 MUTEX_EXIT(&peer->peer_lock);
7078 MUTEX_DESTROY(&peer->peer_lock);
7080 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7081 unsigned int num_funcs;
7082 struct rx_interface_stat *rpc_stat
7083 = opr_queue_Entry(cursor, struct rx_interface_stat,
7088 opr_queue_Remove(&rpc_stat->entry);
7089 opr_queue_Remove(&rpc_stat->entryPeers);
7091 num_funcs = rpc_stat->stats[0].func_total;
7093 sizeof(rx_interface_stat_t) +
7094 rpc_stat->stats[0].func_total *
7095 sizeof(rx_function_entry_v1_t);
7097 rxi_Free(rpc_stat, space);
7099 MUTEX_ENTER(&rx_rpc_stats);
7100 rxi_rpc_peer_stat_cnt -= num_funcs;
7101 MUTEX_EXIT(&rx_rpc_stats);
7106 * Regain the rx_peerHashTable_lock and
7107 * decrement the reference count on 'prev'
7110 MUTEX_ENTER(&rx_peerHashTable_lock);
7117 MUTEX_EXIT(&peer->peer_lock);
7122 MUTEX_EXIT(&rx_peerHashTable_lock);
7126 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7127 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7128 * GC, just below. Really, we shouldn't have to keep moving packets from
7129 * one place to another, but instead ought to always know if we can
7130 * afford to hold onto a packet in its particular use. */
7131 MUTEX_ENTER(&rx_freePktQ_lock);
7132 if (rx_waitingForPackets) {
7133 rx_waitingForPackets = 0;
7134 #ifdef RX_ENABLE_LOCKS
7135 CV_BROADCAST(&rx_waitingForPackets_cv);
7137 osi_rxWakeup(&rx_waitingForPackets);
7140 MUTEX_EXIT(&rx_freePktQ_lock);
7143 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7144 event = rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0);
7145 rxevent_Put(&event);
7149 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7150 * rx.h is sort of strange this is better. This is called with a security
7151 * object before it is discarded. Each connection using a security object has
7152 * its own refcount to the object so it won't actually be freed until the last
7153 * connection is destroyed.
7155 * This is the only rxs module call. A hold could also be written but no one
7159 rxs_Release(struct rx_securityClass *aobj)
7161 return RXS_Close(aobj);
7169 #define TRACE_OPTION_RX_DEBUG 16
7177 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7178 0, KEY_QUERY_VALUE, &parmKey);
7179 if (code != ERROR_SUCCESS)
7182 dummyLen = sizeof(TraceOption);
7183 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7184 (BYTE *) &TraceOption, &dummyLen);
7185 if (code == ERROR_SUCCESS) {
7186 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7188 RegCloseKey (parmKey);
7189 #endif /* AFS_NT40_ENV */
7194 rx_DebugOnOff(int on)
7198 rxdebug_active = on;
7204 rx_StatsOnOff(int on)
7206 rx_stats_active = on;
7210 /* Don't call this debugging routine directly; use dpf */
7212 rxi_DebugPrint(char *format, ...)
7221 va_start(ap, format);
7223 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7226 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7228 OutputDebugString(msg);
7234 va_start(ap, format);
7236 clock_GetTime(&now);
7237 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7238 (unsigned int)now.usec);
7239 vfprintf(rx_Log, format, ap);
7247 * This function is used to process the rx_stats structure that is local
7248 * to a process as well as an rx_stats structure received from a remote
7249 * process (via rxdebug). Therefore, it needs to do minimal version
7253 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7254 afs_int32 freePackets, char version)
7258 if (size != sizeof(struct rx_statistics)) {
7260 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7261 size, sizeof(struct rx_statistics));
7264 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7267 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7268 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7269 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7270 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7271 s->specialPktAllocFailures);
7273 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7274 s->receivePktAllocFailures, s->sendPktAllocFailures,
7275 s->specialPktAllocFailures);
7279 " greedy %u, " "bogusReads %u (last from host %x), "
7280 "noPackets %u, " "noBuffers %u, " "selects %u, "
7281 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7282 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7283 s->selects, s->sendSelects);
7285 fprintf(file, " packets read: ");
7286 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7287 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7289 fprintf(file, "\n");
7292 " other read counters: data %u, " "ack %u, " "dup %u "
7293 "spurious %u " "dally %u\n", s->dataPacketsRead,
7294 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7295 s->ignorePacketDally);
7297 fprintf(file, " packets sent: ");
7298 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7299 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7301 fprintf(file, "\n");
7304 " other send counters: ack %u, " "data %u (not resends), "
7305 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7306 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7307 s->dataPacketsPushed, s->ignoreAckedPacket);
7310 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7311 s->netSendFailures, (int)s->fatalErrors);
7313 if (s->nRttSamples) {
7314 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7315 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7317 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7318 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7322 " %d server connections, " "%d client connections, "
7323 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7324 s->nServerConns, s->nClientConns, s->nPeerStructs,
7325 s->nCallStructs, s->nFreeCallStructs);
7327 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7328 fprintf(file, " %d clock updates\n", clock_nUpdates);
7332 /* for backward compatibility */
7334 rx_PrintStats(FILE * file)
7336 MUTEX_ENTER(&rx_stats_mutex);
7337 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7338 sizeof(rx_stats), rx_nFreePackets,
7340 MUTEX_EXIT(&rx_stats_mutex);
7344 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7346 fprintf(file, "Peer %x.%d.\n",
7347 ntohl(peer->host), (int)ntohs(peer->port));
7350 " Rtt %d, " "total sent %d, " "resent %d\n",
7351 peer->rtt, peer->nSent, peer->reSends);
7353 fprintf(file, " Packet size %d\n", peer->ifMTU);
7357 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7359 * This mutex protects the following static variables:
7363 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7364 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7366 #define LOCK_RX_DEBUG
7367 #define UNLOCK_RX_DEBUG
7368 #endif /* AFS_PTHREAD_ENV */
7370 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7372 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7373 u_char type, void *inputData, size_t inputLength,
7374 void *outputData, size_t outputLength)
7376 static afs_int32 counter = 100;
7377 time_t waitTime, waitCount;
7378 struct rx_header theader;
7381 struct timeval tv_now, tv_wake, tv_delta;
7382 struct sockaddr_in taddr, faddr;
7396 tp = &tbuffer[sizeof(struct rx_header)];
7397 taddr.sin_family = AF_INET;
7398 taddr.sin_port = remotePort;
7399 taddr.sin_addr.s_addr = remoteAddr;
7400 memset(&taddr.sin_zero, 0, sizeof(taddr.sin_zero));
7401 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7402 taddr.sin_len = sizeof(struct sockaddr_in);
7405 memset(&theader, 0, sizeof(theader));
7406 theader.epoch = htonl(999);
7408 theader.callNumber = htonl(counter);
7411 theader.type = type;
7412 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7413 theader.serviceId = 0;
7415 memcpy(tbuffer, &theader, sizeof(theader));
7416 memcpy(tp, inputData, inputLength);
7418 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7419 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7421 /* see if there's a packet available */
7422 gettimeofday(&tv_wake, NULL);
7423 tv_wake.tv_sec += waitTime;
7426 FD_SET(socket, &imask);
7427 tv_delta.tv_sec = tv_wake.tv_sec;
7428 tv_delta.tv_usec = tv_wake.tv_usec;
7429 gettimeofday(&tv_now, NULL);
7431 if (tv_delta.tv_usec < tv_now.tv_usec) {
7433 tv_delta.tv_usec += 1000000;
7436 tv_delta.tv_usec -= tv_now.tv_usec;
7438 if (tv_delta.tv_sec < tv_now.tv_sec) {
7442 tv_delta.tv_sec -= tv_now.tv_sec;
7445 code = select(0, &imask, 0, 0, &tv_delta);
7446 #else /* AFS_NT40_ENV */
7447 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7448 #endif /* AFS_NT40_ENV */
7449 if (code == 1 && FD_ISSET(socket, &imask)) {
7450 /* now receive a packet */
7451 faddrLen = sizeof(struct sockaddr_in);
7453 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7454 (struct sockaddr *)&faddr, &faddrLen);
7457 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7458 if (counter == ntohl(theader.callNumber))
7466 /* see if we've timed out */
7474 code -= sizeof(struct rx_header);
7475 if (code > outputLength)
7476 code = outputLength;
7477 memcpy(outputData, tp, code);
7480 #endif /* RXDEBUG */
7483 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7484 afs_uint16 remotePort, struct rx_debugStats * stat,
7485 afs_uint32 * supportedValues)
7487 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7489 struct rx_debugIn in;
7491 *supportedValues = 0;
7492 in.type = htonl(RX_DEBUGI_GETSTATS);
7495 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7496 &in, sizeof(in), stat, sizeof(*stat));
7499 * If the call was successful, fixup the version and indicate
7500 * what contents of the stat structure are valid.
7501 * Also do net to host conversion of fields here.
7505 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7506 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7508 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7509 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7511 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7512 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7514 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7515 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7517 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7518 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7520 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7521 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7523 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7524 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7526 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7527 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7529 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7530 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7532 stat->nFreePackets = ntohl(stat->nFreePackets);
7533 stat->packetReclaims = ntohl(stat->packetReclaims);
7534 stat->callsExecuted = ntohl(stat->callsExecuted);
7535 stat->nWaiting = ntohl(stat->nWaiting);
7536 stat->idleThreads = ntohl(stat->idleThreads);
7537 stat->nWaited = ntohl(stat->nWaited);
7538 stat->nPackets = ntohl(stat->nPackets);
7547 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7548 afs_uint16 remotePort, struct rx_statistics * stat,
7549 afs_uint32 * supportedValues)
7551 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7553 struct rx_debugIn in;
7554 afs_int32 *lp = (afs_int32 *) stat;
7558 * supportedValues is currently unused, but added to allow future
7559 * versioning of this function.
7562 *supportedValues = 0;
7563 in.type = htonl(RX_DEBUGI_RXSTATS);
7565 memset(stat, 0, sizeof(*stat));
7567 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7568 &in, sizeof(in), stat, sizeof(*stat));
7573 * Do net to host conversion here
7576 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7587 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7588 afs_uint16 remotePort, size_t version_length,
7591 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7593 return MakeDebugCall(socket, remoteAddr, remotePort,
7594 RX_PACKET_TYPE_VERSION, a, 1, version,
7602 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7603 afs_uint16 remotePort, afs_int32 * nextConnection,
7604 int allConnections, afs_uint32 debugSupportedValues,
7605 struct rx_debugConn * conn,
7606 afs_uint32 * supportedValues)
7608 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7610 struct rx_debugIn in;
7614 * supportedValues is currently unused, but added to allow future
7615 * versioning of this function.
7618 *supportedValues = 0;
7619 if (allConnections) {
7620 in.type = htonl(RX_DEBUGI_GETALLCONN);
7622 in.type = htonl(RX_DEBUGI_GETCONN);
7624 in.index = htonl(*nextConnection);
7625 memset(conn, 0, sizeof(*conn));
7627 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7628 &in, sizeof(in), conn, sizeof(*conn));
7631 *nextConnection += 1;
7634 * Convert old connection format to new structure.
7637 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7638 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7639 #define MOVEvL(a) (conn->a = vL->a)
7641 /* any old or unrecognized version... */
7642 for (i = 0; i < RX_MAXCALLS; i++) {
7643 MOVEvL(callState[i]);
7644 MOVEvL(callMode[i]);
7645 MOVEvL(callFlags[i]);
7646 MOVEvL(callOther[i]);
7648 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7649 MOVEvL(secStats.type);
7650 MOVEvL(secStats.level);
7651 MOVEvL(secStats.flags);
7652 MOVEvL(secStats.expires);
7653 MOVEvL(secStats.packetsReceived);
7654 MOVEvL(secStats.packetsSent);
7655 MOVEvL(secStats.bytesReceived);
7656 MOVEvL(secStats.bytesSent);
7661 * Do net to host conversion here
7663 * I don't convert host or port since we are most likely
7664 * going to want these in NBO.
7666 conn->cid = ntohl(conn->cid);
7667 conn->serial = ntohl(conn->serial);
7668 for (i = 0; i < RX_MAXCALLS; i++) {
7669 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7671 conn->error = ntohl(conn->error);
7672 conn->secStats.flags = ntohl(conn->secStats.flags);
7673 conn->secStats.expires = ntohl(conn->secStats.expires);
7674 conn->secStats.packetsReceived =
7675 ntohl(conn->secStats.packetsReceived);
7676 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7677 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7678 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7679 conn->epoch = ntohl(conn->epoch);
7680 conn->natMTU = ntohl(conn->natMTU);
7689 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7690 afs_uint16 remotePort, afs_int32 * nextPeer,
7691 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7692 afs_uint32 * supportedValues)
7694 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7696 struct rx_debugIn in;
7699 * supportedValues is currently unused, but added to allow future
7700 * versioning of this function.
7703 *supportedValues = 0;
7704 in.type = htonl(RX_DEBUGI_GETPEER);
7705 in.index = htonl(*nextPeer);
7706 memset(peer, 0, sizeof(*peer));
7708 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7709 &in, sizeof(in), peer, sizeof(*peer));
7715 * Do net to host conversion here
7717 * I don't convert host or port since we are most likely
7718 * going to want these in NBO.
7720 peer->ifMTU = ntohs(peer->ifMTU);
7721 peer->idleWhen = ntohl(peer->idleWhen);
7722 peer->refCount = ntohs(peer->refCount);
7723 peer->rtt = ntohl(peer->rtt);
7724 peer->rtt_dev = ntohl(peer->rtt_dev);
7725 peer->timeout.sec = 0;
7726 peer->timeout.usec = 0;
7727 peer->nSent = ntohl(peer->nSent);
7728 peer->reSends = ntohl(peer->reSends);
7729 peer->natMTU = ntohs(peer->natMTU);
7730 peer->maxMTU = ntohs(peer->maxMTU);
7731 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7732 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7733 peer->MTU = ntohs(peer->MTU);
7734 peer->cwind = ntohs(peer->cwind);
7735 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7736 peer->congestSeq = ntohs(peer->congestSeq);
7737 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7738 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7739 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7740 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7749 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7750 struct rx_debugPeer * peerStats)
7753 afs_int32 error = 1; /* default to "did not succeed" */
7754 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7756 MUTEX_ENTER(&rx_peerHashTable_lock);
7757 for(tp = rx_peerHashTable[hashValue];
7758 tp != NULL; tp = tp->next) {
7759 if (tp->host == peerHost)
7765 MUTEX_EXIT(&rx_peerHashTable_lock);
7769 MUTEX_ENTER(&tp->peer_lock);
7770 peerStats->host = tp->host;
7771 peerStats->port = tp->port;
7772 peerStats->ifMTU = tp->ifMTU;
7773 peerStats->idleWhen = tp->idleWhen;
7774 peerStats->refCount = tp->refCount;
7775 peerStats->burstSize = 0;
7776 peerStats->burst = 0;
7777 peerStats->burstWait.sec = 0;
7778 peerStats->burstWait.usec = 0;
7779 peerStats->rtt = tp->rtt;
7780 peerStats->rtt_dev = tp->rtt_dev;
7781 peerStats->timeout.sec = 0;
7782 peerStats->timeout.usec = 0;
7783 peerStats->nSent = tp->nSent;
7784 peerStats->reSends = tp->reSends;
7785 peerStats->natMTU = tp->natMTU;
7786 peerStats->maxMTU = tp->maxMTU;
7787 peerStats->maxDgramPackets = tp->maxDgramPackets;
7788 peerStats->ifDgramPackets = tp->ifDgramPackets;
7789 peerStats->MTU = tp->MTU;
7790 peerStats->cwind = tp->cwind;
7791 peerStats->nDgramPackets = tp->nDgramPackets;
7792 peerStats->congestSeq = tp->congestSeq;
7793 peerStats->bytesSent.high = tp->bytesSent >> 32;
7794 peerStats->bytesSent.low = tp->bytesSent & MAX_AFS_UINT32;
7795 peerStats->bytesReceived.high = tp->bytesReceived >> 32;
7796 peerStats->bytesReceived.low
7797 = tp->bytesReceived & MAX_AFS_UINT32;
7798 MUTEX_EXIT(&tp->peer_lock);
7800 MUTEX_ENTER(&rx_peerHashTable_lock);
7803 MUTEX_EXIT(&rx_peerHashTable_lock);
7811 struct rx_serverQueueEntry *np;
7814 struct rx_call *call;
7815 struct rx_serverQueueEntry *sq;
7818 if (rx_atomic_test_and_set_bit(&rxinit_status, 0))
7819 return; /* Already shutdown. */
7823 #ifndef AFS_PTHREAD_ENV
7824 FD_ZERO(&rx_selectMask);
7825 #endif /* AFS_PTHREAD_ENV */
7826 rxi_dataQuota = RX_MAX_QUOTA;
7827 #ifndef AFS_PTHREAD_ENV
7829 #endif /* AFS_PTHREAD_ENV */
7832 #ifndef AFS_PTHREAD_ENV
7833 #ifndef AFS_USE_GETTIMEOFDAY
7835 #endif /* AFS_USE_GETTIMEOFDAY */
7836 #endif /* AFS_PTHREAD_ENV */
7838 while (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
7839 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
7840 opr_queue_Remove(&call->entry);
7841 rxi_Free(call, sizeof(struct rx_call));
7844 while (!opr_queue_IsEmpty(&rx_idleServerQueue)) {
7845 sq = opr_queue_First(&rx_idleServerQueue, struct rx_serverQueueEntry,
7847 opr_queue_Remove(&sq->entry);
7852 struct rx_peer **peer_ptr, **peer_end;
7853 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7854 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7856 struct rx_peer *peer, *next;
7858 MUTEX_ENTER(&rx_peerHashTable_lock);
7859 for (peer = *peer_ptr; peer; peer = next) {
7860 struct opr_queue *cursor, *store;
7863 MUTEX_ENTER(&rx_rpc_stats);
7864 MUTEX_ENTER(&peer->peer_lock);
7865 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7866 unsigned int num_funcs;
7867 struct rx_interface_stat *rpc_stat
7868 = opr_queue_Entry(cursor, struct rx_interface_stat,
7872 opr_queue_Remove(&rpc_stat->entry);
7873 opr_queue_Remove(&rpc_stat->entryPeers);
7874 num_funcs = rpc_stat->stats[0].func_total;
7876 sizeof(rx_interface_stat_t) +
7877 rpc_stat->stats[0].func_total *
7878 sizeof(rx_function_entry_v1_t);
7880 rxi_Free(rpc_stat, space);
7882 /* rx_rpc_stats must be held */
7883 rxi_rpc_peer_stat_cnt -= num_funcs;
7885 MUTEX_EXIT(&peer->peer_lock);
7886 MUTEX_EXIT(&rx_rpc_stats);
7890 if (rx_stats_active)
7891 rx_atomic_dec(&rx_stats.nPeerStructs);
7893 MUTEX_EXIT(&rx_peerHashTable_lock);
7896 for (i = 0; i < RX_MAX_SERVICES; i++) {
7898 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7900 for (i = 0; i < rx_hashTableSize; i++) {
7901 struct rx_connection *tc, *ntc;
7902 MUTEX_ENTER(&rx_connHashTable_lock);
7903 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7905 for (j = 0; j < RX_MAXCALLS; j++) {
7907 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7910 rxi_Free(tc, sizeof(*tc));
7912 MUTEX_EXIT(&rx_connHashTable_lock);
7915 MUTEX_ENTER(&freeSQEList_lock);
7917 while ((np = rx_FreeSQEList)) {
7918 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7919 MUTEX_DESTROY(&np->lock);
7920 rxi_Free(np, sizeof(*np));
7923 MUTEX_EXIT(&freeSQEList_lock);
7924 MUTEX_DESTROY(&freeSQEList_lock);
7925 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7926 MUTEX_DESTROY(&rx_connHashTable_lock);
7927 MUTEX_DESTROY(&rx_peerHashTable_lock);
7928 MUTEX_DESTROY(&rx_serverPool_lock);
7930 osi_Free(rx_connHashTable,
7931 rx_hashTableSize * sizeof(struct rx_connection *));
7932 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7934 UNPIN(rx_connHashTable,
7935 rx_hashTableSize * sizeof(struct rx_connection *));
7936 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7938 MUTEX_ENTER(&rx_quota_mutex);
7939 rxi_dataQuota = RX_MAX_QUOTA;
7940 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7941 MUTEX_EXIT(&rx_quota_mutex);
7947 * Routines to implement connection specific data.
7951 rx_KeyCreate(rx_destructor_t rtn)
7954 MUTEX_ENTER(&rxi_keyCreate_lock);
7955 key = rxi_keyCreate_counter++;
7956 rxi_keyCreate_destructor = (rx_destructor_t *)
7957 realloc((void *)rxi_keyCreate_destructor,
7958 (key + 1) * sizeof(rx_destructor_t));
7959 rxi_keyCreate_destructor[key] = rtn;
7960 MUTEX_EXIT(&rxi_keyCreate_lock);
7965 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7968 MUTEX_ENTER(&conn->conn_data_lock);
7969 if (!conn->specific) {
7970 conn->specific = malloc((key + 1) * sizeof(void *));
7971 for (i = 0; i < key; i++)
7972 conn->specific[i] = NULL;
7973 conn->nSpecific = key + 1;
7974 conn->specific[key] = ptr;
7975 } else if (key >= conn->nSpecific) {
7976 conn->specific = (void **)
7977 realloc(conn->specific, (key + 1) * sizeof(void *));
7978 for (i = conn->nSpecific; i < key; i++)
7979 conn->specific[i] = NULL;
7980 conn->nSpecific = key + 1;
7981 conn->specific[key] = ptr;
7983 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7984 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7985 conn->specific[key] = ptr;
7987 MUTEX_EXIT(&conn->conn_data_lock);
7991 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7994 MUTEX_ENTER(&svc->svc_data_lock);
7995 if (!svc->specific) {
7996 svc->specific = malloc((key + 1) * sizeof(void *));
7997 for (i = 0; i < key; i++)
7998 svc->specific[i] = NULL;
7999 svc->nSpecific = key + 1;
8000 svc->specific[key] = ptr;
8001 } else if (key >= svc->nSpecific) {
8002 svc->specific = (void **)
8003 realloc(svc->specific, (key + 1) * sizeof(void *));
8004 for (i = svc->nSpecific; i < key; i++)
8005 svc->specific[i] = NULL;
8006 svc->nSpecific = key + 1;
8007 svc->specific[key] = ptr;
8009 if (svc->specific[key] && rxi_keyCreate_destructor[key])
8010 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
8011 svc->specific[key] = ptr;
8013 MUTEX_EXIT(&svc->svc_data_lock);
8017 rx_GetSpecific(struct rx_connection *conn, int key)
8020 MUTEX_ENTER(&conn->conn_data_lock);
8021 if (key >= conn->nSpecific)
8024 ptr = conn->specific[key];
8025 MUTEX_EXIT(&conn->conn_data_lock);
8030 rx_GetServiceSpecific(struct rx_service *svc, int key)
8033 MUTEX_ENTER(&svc->svc_data_lock);
8034 if (key >= svc->nSpecific)
8037 ptr = svc->specific[key];
8038 MUTEX_EXIT(&svc->svc_data_lock);
8043 #endif /* !KERNEL */
8046 * processStats is a queue used to store the statistics for the local
8047 * process. Its contents are similar to the contents of the rpcStats
8048 * queue on a rx_peer structure, but the actual data stored within
8049 * this queue contains totals across the lifetime of the process (assuming
8050 * the stats have not been reset) - unlike the per peer structures
8051 * which can come and go based upon the peer lifetime.
8054 static struct opr_queue processStats = { &processStats, &processStats };
8057 * peerStats is a queue used to store the statistics for all peer structs.
8058 * Its contents are the union of all the peer rpcStats queues.
8061 static struct opr_queue peerStats = { &peerStats, &peerStats };
8064 * rxi_monitor_processStats is used to turn process wide stat collection
8068 static int rxi_monitor_processStats = 0;
8071 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8074 static int rxi_monitor_peerStats = 0;
8078 rxi_ClearRPCOpStat(rx_function_entry_v1_p rpc_stat)
8080 rpc_stat->invocations = 0;
8081 rpc_stat->bytes_sent = 0;
8082 rpc_stat->bytes_rcvd = 0;
8083 rpc_stat->queue_time_sum.sec = 0;
8084 rpc_stat->queue_time_sum.usec = 0;
8085 rpc_stat->queue_time_sum_sqr.sec = 0;
8086 rpc_stat->queue_time_sum_sqr.usec = 0;
8087 rpc_stat->queue_time_min.sec = 9999999;
8088 rpc_stat->queue_time_min.usec = 9999999;
8089 rpc_stat->queue_time_max.sec = 0;
8090 rpc_stat->queue_time_max.usec = 0;
8091 rpc_stat->execution_time_sum.sec = 0;
8092 rpc_stat->execution_time_sum.usec = 0;
8093 rpc_stat->execution_time_sum_sqr.sec = 0;
8094 rpc_stat->execution_time_sum_sqr.usec = 0;
8095 rpc_stat->execution_time_min.sec = 9999999;
8096 rpc_stat->execution_time_min.usec = 9999999;
8097 rpc_stat->execution_time_max.sec = 0;
8098 rpc_stat->execution_time_max.usec = 0;
8102 * Given all of the information for a particular rpc
8103 * call, find or create (if requested) the stat structure for the rpc.
8106 * the queue of stats that will be updated with the new value
8108 * @param rxInterface
8109 * a unique number that identifies the rpc interface
8112 * the total number of functions in this interface. this is only
8113 * required if create is true
8116 * if true, this invocation was made to a server
8119 * the ip address of the remote host. this is only required if create
8120 * and addToPeerList are true
8123 * the port of the remote host. this is only required if create
8124 * and addToPeerList are true
8126 * @param addToPeerList
8127 * if != 0, add newly created stat to the global peer list
8130 * if a new stats structure is allocated, the counter will
8131 * be updated with the new number of allocated stat structures.
8132 * only required if create is true
8135 * if no stats structure exists, allocate one
8139 static rx_interface_stat_p
8140 rxi_FindRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8141 afs_uint32 totalFunc, int isServer, afs_uint32 remoteHost,
8142 afs_uint32 remotePort, int addToPeerList,
8143 unsigned int *counter, int create)
8145 rx_interface_stat_p rpc_stat = NULL;
8146 struct opr_queue *cursor;
8149 * See if there's already a structure for this interface
8152 for (opr_queue_Scan(stats, cursor)) {
8153 rpc_stat = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8155 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8156 && (rpc_stat->stats[0].remote_is_server == isServer))
8160 /* if they didn't ask us to create, we're done */
8162 if (opr_queue_IsEnd(stats, cursor))
8168 /* can't proceed without these */
8169 if (!totalFunc || !counter)
8173 * Didn't find a match so allocate a new structure and add it to the
8177 if (opr_queue_IsEnd(stats, cursor) || (rpc_stat == NULL)
8178 || (rpc_stat->stats[0].interfaceId != rxInterface)
8179 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8184 sizeof(rx_interface_stat_t) +
8185 totalFunc * sizeof(rx_function_entry_v1_t);
8187 rpc_stat = rxi_Alloc(space);
8188 if (rpc_stat == NULL)
8191 *counter += totalFunc;
8192 for (i = 0; i < totalFunc; i++) {
8193 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8194 rpc_stat->stats[i].remote_peer = remoteHost;
8195 rpc_stat->stats[i].remote_port = remotePort;
8196 rpc_stat->stats[i].remote_is_server = isServer;
8197 rpc_stat->stats[i].interfaceId = rxInterface;
8198 rpc_stat->stats[i].func_total = totalFunc;
8199 rpc_stat->stats[i].func_index = i;
8201 opr_queue_Prepend(stats, &rpc_stat->entry);
8202 if (addToPeerList) {
8203 opr_queue_Prepend(&peerStats, &rpc_stat->entryPeers);
8210 rx_ClearProcessRPCStats(afs_int32 rxInterface)
8212 rx_interface_stat_p rpc_stat;
8215 if (rxInterface == -1)
8218 MUTEX_ENTER(&rx_rpc_stats);
8219 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8222 totalFunc = rpc_stat->stats[0].func_total;
8223 for (i = 0; i < totalFunc; i++)
8224 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8226 MUTEX_EXIT(&rx_rpc_stats);
8231 rx_ClearPeerRPCStats(afs_int32 rxInterface, afs_uint32 peerHost, afs_uint16 peerPort)
8233 rx_interface_stat_p rpc_stat;
8235 struct rx_peer * peer;
8237 if (rxInterface == -1)
8240 peer = rxi_FindPeer(peerHost, peerPort, 0);
8244 MUTEX_ENTER(&rx_rpc_stats);
8245 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8248 totalFunc = rpc_stat->stats[0].func_total;
8249 for (i = 0; i < totalFunc; i++)
8250 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8252 MUTEX_EXIT(&rx_rpc_stats);
8257 rx_CopyProcessRPCStats(afs_uint64 op)
8259 rx_interface_stat_p rpc_stat;
8260 rx_function_entry_v1_p rpcop_stat =
8261 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8262 int currentFunc = (op & MAX_AFS_UINT32);
8263 afs_int32 rxInterface = (op >> 32);
8265 if (!rxi_monitor_processStats)
8268 if (rxInterface == -1)
8271 if (rpcop_stat == NULL)
8274 MUTEX_ENTER(&rx_rpc_stats);
8275 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8278 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8279 sizeof(rx_function_entry_v1_t));
8280 MUTEX_EXIT(&rx_rpc_stats);
8282 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8289 rx_CopyPeerRPCStats(afs_uint64 op, afs_uint32 peerHost, afs_uint16 peerPort)
8291 rx_interface_stat_p rpc_stat;
8292 rx_function_entry_v1_p rpcop_stat =
8293 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8294 int currentFunc = (op & MAX_AFS_UINT32);
8295 afs_int32 rxInterface = (op >> 32);
8296 struct rx_peer *peer;
8298 if (!rxi_monitor_peerStats)
8301 if (rxInterface == -1)
8304 if (rpcop_stat == NULL)
8307 peer = rxi_FindPeer(peerHost, peerPort, 0);
8311 MUTEX_ENTER(&rx_rpc_stats);
8312 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8315 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8316 sizeof(rx_function_entry_v1_t));
8317 MUTEX_EXIT(&rx_rpc_stats);
8319 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8326 rx_ReleaseRPCStats(void *stats)
8329 rxi_Free(stats, sizeof(rx_function_entry_v1_t));
8333 * Given all of the information for a particular rpc
8334 * call, create (if needed) and update the stat totals for the rpc.
8337 * the queue of stats that will be updated with the new value
8339 * @param rxInterface
8340 * a unique number that identifies the rpc interface
8342 * @param currentFunc
8343 * the index of the function being invoked
8346 * the total number of functions in this interface
8349 * the amount of time this function waited for a thread
8352 * the amount of time this function invocation took to execute
8355 * the number bytes sent by this invocation
8358 * the number bytes received by this invocation
8361 * if true, this invocation was made to a server
8364 * the ip address of the remote host
8367 * the port of the remote host
8369 * @param addToPeerList
8370 * if != 0, add newly created stat to the global peer list
8373 * if a new stats structure is allocated, the counter will
8374 * be updated with the new number of allocated stat structures
8379 rxi_AddRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8380 afs_uint32 currentFunc, afs_uint32 totalFunc,
8381 struct clock *queueTime, struct clock *execTime,
8382 afs_uint64 bytesSent, afs_uint64 bytesRcvd, int isServer,
8383 afs_uint32 remoteHost, afs_uint32 remotePort,
8384 int addToPeerList, unsigned int *counter)
8387 rx_interface_stat_p rpc_stat;
8389 rpc_stat = rxi_FindRpcStat(stats, rxInterface, totalFunc, isServer,
8390 remoteHost, remotePort, addToPeerList, counter,
8398 * Increment the stats for this function
8401 rpc_stat->stats[currentFunc].invocations++;
8402 rpc_stat->stats[currentFunc].bytes_sent += bytesSent;
8403 rpc_stat->stats[currentFunc].bytes_rcvd += bytesRcvd;
8404 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8405 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8406 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8407 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8409 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8410 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8412 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8413 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8415 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8416 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8418 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8419 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8427 rxi_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8428 afs_uint32 currentFunc, afs_uint32 totalFunc,
8429 struct clock *queueTime, struct clock *execTime,
8430 afs_uint64 bytesSent, afs_uint64 bytesRcvd,
8434 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8437 MUTEX_ENTER(&rx_rpc_stats);
8439 if (rxi_monitor_peerStats) {
8440 MUTEX_ENTER(&peer->peer_lock);
8441 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8442 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8443 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8444 MUTEX_EXIT(&peer->peer_lock);
8447 if (rxi_monitor_processStats) {
8448 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8449 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8450 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8453 MUTEX_EXIT(&rx_rpc_stats);
8457 * Increment the times and count for a particular rpc function.
8459 * Traditionally this call was invoked from rxgen stubs. Modern stubs
8460 * call rx_RecordCallStatistics instead, so the public version of this
8461 * function is left purely for legacy callers.
8464 * The peer who invoked the rpc
8466 * @param rxInterface
8467 * A unique number that identifies the rpc interface
8469 * @param currentFunc
8470 * The index of the function being invoked
8473 * The total number of functions in this interface
8476 * The amount of time this function waited for a thread
8479 * The amount of time this function invocation took to execute
8482 * The number bytes sent by this invocation
8485 * The number bytes received by this invocation
8488 * If true, this invocation was made to a server
8492 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8493 afs_uint32 currentFunc, afs_uint32 totalFunc,
8494 struct clock *queueTime, struct clock *execTime,
8495 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8501 sent64 = ((afs_uint64)bytesSent->high << 32) + bytesSent->low;
8502 rcvd64 = ((afs_uint64)bytesRcvd->high << 32) + bytesRcvd->low;
8504 rxi_IncrementTimeAndCount(peer, rxInterface, currentFunc, totalFunc,
8505 queueTime, execTime, sent64, rcvd64,
8512 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8516 * IN callerVersion - the rpc stat version of the caller.
8518 * IN count - the number of entries to marshall.
8520 * IN stats - pointer to stats to be marshalled.
8522 * OUT ptr - Where to store the marshalled data.
8529 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8530 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8536 * We only support the first version
8538 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8539 *(ptr++) = stats->remote_peer;
8540 *(ptr++) = stats->remote_port;
8541 *(ptr++) = stats->remote_is_server;
8542 *(ptr++) = stats->interfaceId;
8543 *(ptr++) = stats->func_total;
8544 *(ptr++) = stats->func_index;
8545 *(ptr++) = stats->invocations >> 32;
8546 *(ptr++) = stats->invocations & MAX_AFS_UINT32;
8547 *(ptr++) = stats->bytes_sent >> 32;
8548 *(ptr++) = stats->bytes_sent & MAX_AFS_UINT32;
8549 *(ptr++) = stats->bytes_rcvd >> 32;
8550 *(ptr++) = stats->bytes_rcvd & MAX_AFS_UINT32;
8551 *(ptr++) = stats->queue_time_sum.sec;
8552 *(ptr++) = stats->queue_time_sum.usec;
8553 *(ptr++) = stats->queue_time_sum_sqr.sec;
8554 *(ptr++) = stats->queue_time_sum_sqr.usec;
8555 *(ptr++) = stats->queue_time_min.sec;
8556 *(ptr++) = stats->queue_time_min.usec;
8557 *(ptr++) = stats->queue_time_max.sec;
8558 *(ptr++) = stats->queue_time_max.usec;
8559 *(ptr++) = stats->execution_time_sum.sec;
8560 *(ptr++) = stats->execution_time_sum.usec;
8561 *(ptr++) = stats->execution_time_sum_sqr.sec;
8562 *(ptr++) = stats->execution_time_sum_sqr.usec;
8563 *(ptr++) = stats->execution_time_min.sec;
8564 *(ptr++) = stats->execution_time_min.usec;
8565 *(ptr++) = stats->execution_time_max.sec;
8566 *(ptr++) = stats->execution_time_max.usec;
8572 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8577 * IN callerVersion - the rpc stat version of the caller
8579 * OUT myVersion - the rpc stat version of this function
8581 * OUT clock_sec - local time seconds
8583 * OUT clock_usec - local time microseconds
8585 * OUT allocSize - the number of bytes allocated to contain stats
8587 * OUT statCount - the number stats retrieved from this process.
8589 * OUT stats - the actual stats retrieved from this process.
8593 * Returns void. If successful, stats will != NULL.
8597 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8598 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8599 size_t * allocSize, afs_uint32 * statCount,
8600 afs_uint32 ** stats)
8610 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8613 * Check to see if stats are enabled
8616 MUTEX_ENTER(&rx_rpc_stats);
8617 if (!rxi_monitor_processStats) {
8618 MUTEX_EXIT(&rx_rpc_stats);
8622 clock_GetTime(&now);
8623 *clock_sec = now.sec;
8624 *clock_usec = now.usec;
8627 * Allocate the space based upon the caller version
8629 * If the client is at an older version than we are,
8630 * we return the statistic data in the older data format, but
8631 * we still return our version number so the client knows we
8632 * are maintaining more data than it can retrieve.
8635 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8636 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8637 *statCount = rxi_rpc_process_stat_cnt;
8640 * This can't happen yet, but in the future version changes
8641 * can be handled by adding additional code here
8645 if (space > (size_t) 0) {
8647 ptr = *stats = rxi_Alloc(space);
8650 struct opr_queue *cursor;
8652 for (opr_queue_Scan(&processStats, cursor)) {
8653 struct rx_interface_stat *rpc_stat =
8654 opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8656 * Copy the data based upon the caller version
8658 rx_MarshallProcessRPCStats(callerVersion,
8659 rpc_stat->stats[0].func_total,
8660 rpc_stat->stats, &ptr);
8666 MUTEX_EXIT(&rx_rpc_stats);
8671 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8675 * IN callerVersion - the rpc stat version of the caller
8677 * OUT myVersion - the rpc stat version of this function
8679 * OUT clock_sec - local time seconds
8681 * OUT clock_usec - local time microseconds
8683 * OUT allocSize - the number of bytes allocated to contain stats
8685 * OUT statCount - the number of stats retrieved from the individual
8688 * OUT stats - the actual stats retrieved from the individual peer structures.
8692 * Returns void. If successful, stats will != NULL.
8696 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8697 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8698 size_t * allocSize, afs_uint32 * statCount,
8699 afs_uint32 ** stats)
8709 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8712 * Check to see if stats are enabled
8715 MUTEX_ENTER(&rx_rpc_stats);
8716 if (!rxi_monitor_peerStats) {
8717 MUTEX_EXIT(&rx_rpc_stats);
8721 clock_GetTime(&now);
8722 *clock_sec = now.sec;
8723 *clock_usec = now.usec;
8726 * Allocate the space based upon the caller version
8728 * If the client is at an older version than we are,
8729 * we return the statistic data in the older data format, but
8730 * we still return our version number so the client knows we
8731 * are maintaining more data than it can retrieve.
8734 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8735 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8736 *statCount = rxi_rpc_peer_stat_cnt;
8739 * This can't happen yet, but in the future version changes
8740 * can be handled by adding additional code here
8744 if (space > (size_t) 0) {
8746 ptr = *stats = rxi_Alloc(space);
8749 struct opr_queue *cursor;
8751 for (opr_queue_Scan(&peerStats, cursor)) {
8752 struct rx_interface_stat *rpc_stat
8753 = opr_queue_Entry(cursor, struct rx_interface_stat,
8757 * Copy the data based upon the caller version
8759 rx_MarshallProcessRPCStats(callerVersion,
8760 rpc_stat->stats[0].func_total,
8761 rpc_stat->stats, &ptr);
8767 MUTEX_EXIT(&rx_rpc_stats);
8772 * rx_FreeRPCStats - free memory allocated by
8773 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8777 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8778 * rx_RetrievePeerRPCStats
8780 * IN allocSize - the number of bytes in stats.
8788 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8790 rxi_Free(stats, allocSize);
8794 * rx_queryProcessRPCStats - see if process rpc stat collection is
8795 * currently enabled.
8801 * Returns 0 if stats are not enabled != 0 otherwise
8805 rx_queryProcessRPCStats(void)
8808 MUTEX_ENTER(&rx_rpc_stats);
8809 rc = rxi_monitor_processStats;
8810 MUTEX_EXIT(&rx_rpc_stats);
8815 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8821 * Returns 0 if stats are not enabled != 0 otherwise
8825 rx_queryPeerRPCStats(void)
8828 MUTEX_ENTER(&rx_rpc_stats);
8829 rc = rxi_monitor_peerStats;
8830 MUTEX_EXIT(&rx_rpc_stats);
8835 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8845 rx_enableProcessRPCStats(void)
8847 MUTEX_ENTER(&rx_rpc_stats);
8848 rx_enable_stats = 1;
8849 rxi_monitor_processStats = 1;
8850 MUTEX_EXIT(&rx_rpc_stats);
8854 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8864 rx_enablePeerRPCStats(void)
8866 MUTEX_ENTER(&rx_rpc_stats);
8867 rx_enable_stats = 1;
8868 rxi_monitor_peerStats = 1;
8869 MUTEX_EXIT(&rx_rpc_stats);
8873 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8883 rx_disableProcessRPCStats(void)
8885 struct opr_queue *cursor, *store;
8888 MUTEX_ENTER(&rx_rpc_stats);
8891 * Turn off process statistics and if peer stats is also off, turn
8895 rxi_monitor_processStats = 0;
8896 if (rxi_monitor_peerStats == 0) {
8897 rx_enable_stats = 0;
8900 for (opr_queue_ScanSafe(&processStats, cursor, store)) {
8901 unsigned int num_funcs = 0;
8902 struct rx_interface_stat *rpc_stat
8903 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8905 opr_queue_Remove(&rpc_stat->entry);
8907 num_funcs = rpc_stat->stats[0].func_total;
8909 sizeof(rx_interface_stat_t) +
8910 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8912 rxi_Free(rpc_stat, space);
8913 rxi_rpc_process_stat_cnt -= num_funcs;
8915 MUTEX_EXIT(&rx_rpc_stats);
8919 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8929 rx_disablePeerRPCStats(void)
8931 struct rx_peer **peer_ptr, **peer_end;
8935 * Turn off peer statistics and if process stats is also off, turn
8939 rxi_monitor_peerStats = 0;
8940 if (rxi_monitor_processStats == 0) {
8941 rx_enable_stats = 0;
8944 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8945 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8947 struct rx_peer *peer, *next, *prev;
8949 MUTEX_ENTER(&rx_peerHashTable_lock);
8950 MUTEX_ENTER(&rx_rpc_stats);
8951 for (prev = peer = *peer_ptr; peer; peer = next) {
8953 code = MUTEX_TRYENTER(&peer->peer_lock);
8956 struct opr_queue *cursor, *store;
8958 if (prev == *peer_ptr) {
8969 MUTEX_EXIT(&rx_peerHashTable_lock);
8971 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
8972 unsigned int num_funcs = 0;
8973 struct rx_interface_stat *rpc_stat
8974 = opr_queue_Entry(cursor, struct rx_interface_stat,
8977 opr_queue_Remove(&rpc_stat->entry);
8978 opr_queue_Remove(&rpc_stat->entryPeers);
8979 num_funcs = rpc_stat->stats[0].func_total;
8981 sizeof(rx_interface_stat_t) +
8982 rpc_stat->stats[0].func_total *
8983 sizeof(rx_function_entry_v1_t);
8985 rxi_Free(rpc_stat, space);
8986 rxi_rpc_peer_stat_cnt -= num_funcs;
8988 MUTEX_EXIT(&peer->peer_lock);
8990 MUTEX_ENTER(&rx_peerHashTable_lock);
9000 MUTEX_EXIT(&rx_rpc_stats);
9001 MUTEX_EXIT(&rx_peerHashTable_lock);
9006 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
9011 * IN clearFlag - flag indicating which stats to clear
9019 rx_clearProcessRPCStats(afs_uint32 clearFlag)
9021 struct opr_queue *cursor;
9023 MUTEX_ENTER(&rx_rpc_stats);
9025 for (opr_queue_Scan(&processStats, cursor)) {
9026 unsigned int num_funcs = 0, i;
9027 struct rx_interface_stat *rpc_stat
9028 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
9030 num_funcs = rpc_stat->stats[0].func_total;
9031 for (i = 0; i < num_funcs; i++) {
9032 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9033 rpc_stat->stats[i].invocations = 0;
9035 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9036 rpc_stat->stats[i].bytes_sent = 0;
9038 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9039 rpc_stat->stats[i].bytes_rcvd = 0;
9041 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9042 rpc_stat->stats[i].queue_time_sum.sec = 0;
9043 rpc_stat->stats[i].queue_time_sum.usec = 0;
9045 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9046 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9047 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9049 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9050 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9051 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9053 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9054 rpc_stat->stats[i].queue_time_max.sec = 0;
9055 rpc_stat->stats[i].queue_time_max.usec = 0;
9057 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9058 rpc_stat->stats[i].execution_time_sum.sec = 0;
9059 rpc_stat->stats[i].execution_time_sum.usec = 0;
9061 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9062 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9063 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9065 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9066 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9067 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9069 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9070 rpc_stat->stats[i].execution_time_max.sec = 0;
9071 rpc_stat->stats[i].execution_time_max.usec = 0;
9076 MUTEX_EXIT(&rx_rpc_stats);
9080 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
9085 * IN clearFlag - flag indicating which stats to clear
9093 rx_clearPeerRPCStats(afs_uint32 clearFlag)
9095 struct opr_queue *cursor;
9097 MUTEX_ENTER(&rx_rpc_stats);
9099 for (opr_queue_Scan(&peerStats, cursor)) {
9100 unsigned int num_funcs, i;
9101 struct rx_interface_stat *rpc_stat
9102 = opr_queue_Entry(cursor, struct rx_interface_stat, entryPeers);
9104 num_funcs = rpc_stat->stats[0].func_total;
9105 for (i = 0; i < num_funcs; i++) {
9106 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9107 rpc_stat->stats[i].invocations = 0;
9109 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9110 rpc_stat->stats[i].bytes_sent = 0;
9112 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9113 rpc_stat->stats[i].bytes_rcvd = 0;
9115 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9116 rpc_stat->stats[i].queue_time_sum.sec = 0;
9117 rpc_stat->stats[i].queue_time_sum.usec = 0;
9119 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9120 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9121 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9123 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9124 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9125 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9127 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9128 rpc_stat->stats[i].queue_time_max.sec = 0;
9129 rpc_stat->stats[i].queue_time_max.usec = 0;
9131 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9132 rpc_stat->stats[i].execution_time_sum.sec = 0;
9133 rpc_stat->stats[i].execution_time_sum.usec = 0;
9135 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9136 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9137 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9139 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9140 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9141 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9143 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9144 rpc_stat->stats[i].execution_time_max.sec = 0;
9145 rpc_stat->stats[i].execution_time_max.usec = 0;
9150 MUTEX_EXIT(&rx_rpc_stats);
9154 * rxi_rxstat_userok points to a routine that returns 1 if the caller
9155 * is authorized to enable/disable/clear RX statistics.
9157 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
9160 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
9162 rxi_rxstat_userok = proc;
9166 rx_RxStatUserOk(struct rx_call *call)
9168 if (!rxi_rxstat_userok)
9170 return rxi_rxstat_userok(call);
9175 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
9176 * function in the MSVC runtime DLL (msvcrt.dll).
9178 * Note: the system serializes calls to this function.
9181 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
9182 DWORD reason, /* reason function is being called */
9183 LPVOID reserved) /* reserved for future use */
9186 case DLL_PROCESS_ATTACH:
9187 /* library is being attached to a process */
9191 case DLL_PROCESS_DETACH:
9198 #endif /* AFS_NT40_ENV */
9201 int rx_DumpCalls(FILE *outputFile, char *cookie)
9203 #ifdef RXDEBUG_PACKET
9204 #ifdef KDUMP_RX_LOCK
9205 struct rx_call_rx_lock *c;
9212 #define RXDPRINTF sprintf
9213 #define RXDPRINTOUT output
9215 #define RXDPRINTF fprintf
9216 #define RXDPRINTOUT outputFile
9219 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9221 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9224 for (c = rx_allCallsp; c; c = c->allNextp) {
9225 u_short rqc, tqc, iovqc;
9227 MUTEX_ENTER(&c->lock);
9228 rqc = opr_queue_Count(&c->rq);
9229 tqc = opr_queue_Count(&c->tq);
9230 iovqc = opr_queue_Count(&c->app.iovq);
9232 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, "
9233 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9234 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9235 "resendEvent=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9236 "lastSendTime=%u, lastRecvTime=%u"
9237 #ifdef RX_ENABLE_LOCKS
9240 #ifdef RX_REFCOUNT_CHECK
9241 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9242 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9245 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,
9246 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9247 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9248 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9249 c->resendEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9250 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime
9251 #ifdef RX_ENABLE_LOCKS
9252 , (afs_uint32)c->refCount
9254 #ifdef RX_REFCOUNT_CHECK
9255 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9258 MUTEX_EXIT(&c->lock);
9261 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9264 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9266 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9268 #endif /* RXDEBUG_PACKET */