2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
13 #include <afs/param.h>
16 # include "afs/sysincludes.h"
17 # include "afsincludes.h"
22 # ifdef AFS_LINUX20_ENV
23 # include "h/socket.h"
25 # include "netinet/in.h"
27 # include "netinet/ip6.h"
28 # include "inet/common.h"
30 # include "inet/ip_ire.h"
32 # include "afs/afs_args.h"
33 # include "afs/afs_osi.h"
34 # ifdef RX_KERNEL_TRACE
35 # include "rx_kcommon.h"
37 # if defined(AFS_AIX_ENV)
41 # undef RXDEBUG /* turn off debugging */
43 # if defined(AFS_SGI_ENV)
44 # include "sys/debug.h"
47 # include "afs/sysincludes.h"
48 # include "afsincludes.h"
49 # endif /* !UKERNEL */
50 # include "afs/lock.h"
51 # include "rx_kmutex.h"
52 # include "rx_kernel.h"
53 # define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
54 # define AFSOP_STOP_AFS 211 /* Stop AFS process */
55 # define AFSOP_STOP_BKG 212 /* Stop BKG process */
56 extern afs_int32 afs_termState;
58 # include "sys/lockl.h"
59 # include "sys/lock_def.h"
60 # endif /* AFS_AIX41_ENV */
61 # include "afs/rxgen_consts.h"
66 # include <afs/afsutil.h>
67 # include <WINNT\afsreg.h>
75 #include <opr/queue.h>
79 #include "rx_atomic.h"
80 #include "rx_globals.h"
82 #include "rx_internal.h"
89 #include "rx_packet.h"
90 #include "rx_server.h"
92 #include <afs/rxgen_consts.h>
95 #ifdef AFS_PTHREAD_ENV
97 int (*registerProgram) (pid_t, char *) = 0;
98 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
101 int (*registerProgram) (PROCESS, char *) = 0;
102 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
106 /* Local static routines */
107 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
108 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
109 struct rx_call *, struct rx_peer *,
111 static void rxi_Resend(struct rxevent *event, void *arg0, void *arg1,
113 static void rxi_SendDelayedAck(struct rxevent *event, void *call,
114 void *dummy, int dummy2);
115 static void rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1,
116 void *dummy, int dummy2);
117 static void rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1,
118 void *unused, int unused2);
119 static void rxi_ReapConnections(struct rxevent *unused, void *unused1,
120 void *unused2, int unused3);
121 static struct rx_packet *rxi_SendCallAbort(struct rx_call *call,
122 struct rx_packet *packet,
123 int istack, int force);
124 static void rxi_AckAll(struct rx_call *call);
125 static struct rx_connection
126 *rxi_FindConnection(osi_socket socket, afs_uint32 host, u_short port,
127 u_short serviceId, afs_uint32 cid,
128 afs_uint32 epoch, int type, u_int securityIndex,
129 int *unknownService);
130 static struct rx_packet
131 *rxi_ReceiveDataPacket(struct rx_call *call, struct rx_packet *np,
132 int istack, osi_socket socket,
133 afs_uint32 host, u_short port, int *tnop,
134 struct rx_call **newcallp);
135 static struct rx_packet
136 *rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
138 static struct rx_packet
139 *rxi_ReceiveResponsePacket(struct rx_connection *conn,
140 struct rx_packet *np, int istack);
141 static struct rx_packet
142 *rxi_ReceiveChallengePacket(struct rx_connection *conn,
143 struct rx_packet *np, int istack);
144 static void rxi_AttachServerProc(struct rx_call *call, osi_socket socket,
145 int *tnop, struct rx_call **newcallp);
146 static void rxi_ClearTransmitQueue(struct rx_call *call, int force);
147 static void rxi_ClearReceiveQueue(struct rx_call *call);
148 static void rxi_ResetCall(struct rx_call *call, int newcall);
149 static void rxi_ScheduleKeepAliveEvent(struct rx_call *call);
150 static void rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn);
151 static void rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs);
152 static void rxi_KeepAliveOn(struct rx_call *call);
153 static void rxi_GrowMTUOn(struct rx_call *call);
154 static void rxi_ChallengeOn(struct rx_connection *conn);
155 static int rxi_CheckCall(struct rx_call *call, int haveCTLock);
156 static void rxi_AckAllInTransmitQueue(struct rx_call *call);
157 static void rxi_CancelKeepAliveEvent(struct rx_call *call);
158 static void rxi_CancelDelayedAbortEvent(struct rx_call *call);
159 static void rxi_CancelGrowMTUEvent(struct rx_call *call);
161 #ifdef RX_ENABLE_LOCKS
163 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
164 rx_atomic_t rxi_start_in_error;
166 #endif /* RX_ENABLE_LOCKS */
168 /* Constant delay time before sending an acknowledge of the last packet
169 * received. This is to avoid sending an extra acknowledge when the
170 * client is about to make another call, anyway, or the server is
173 * The lastAckDelay may not exceeed 400ms without causing peers to
174 * unecessarily timeout.
176 struct clock rx_lastAckDelay = {0, 400000};
178 /* Constant delay time before sending a soft ack when none was requested.
179 * This is to make sure we send soft acks before the sender times out,
180 * Normally we wait and send a hard ack when the receiver consumes the packet
182 * This value has been 100ms in all shipping versions of OpenAFS. Changing it
183 * will require changes to the peer's RTT calculations.
185 struct clock rx_softAckDelay = {0, 100000};
188 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
189 * currently allocated within rx. This number is used to allocate the
190 * memory required to return the statistics when queried.
191 * Protected by the rx_rpc_stats mutex.
194 static unsigned int rxi_rpc_peer_stat_cnt;
197 * rxi_rpc_process_stat_cnt counts the total number of local process stat
198 * structures currently allocated within rx. The number is used to allocate
199 * the memory required to return the statistics when queried.
200 * Protected by the rx_rpc_stats mutex.
203 static unsigned int rxi_rpc_process_stat_cnt;
206 * rxi_busyChannelError is a boolean. It indicates whether or not RX_CALL_BUSY
207 * errors should be reported to the application when a call channel appears busy
208 * (inferred from the receipt of RX_PACKET_TYPE_BUSY packets on the channel),
209 * and there are other call channels in the connection that are not busy.
210 * If 0, we do not return errors upon receiving busy packets; we just keep
211 * trying on the same call channel until we hit a timeout.
213 static afs_int32 rxi_busyChannelError = 0;
215 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
216 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
218 /* Incoming calls wait on this queue when there are no available
219 * server processes */
220 struct opr_queue rx_incomingCallQueue;
222 /* Server processes wait on this queue when there are no appropriate
223 * calls to process */
224 struct opr_queue rx_idleServerQueue;
226 #if !defined(offsetof)
227 #include <stddef.h> /* for definition of offsetof() */
230 #ifdef RX_ENABLE_LOCKS
231 afs_kmutex_t rx_atomic_mutex;
234 /* Forward prototypes */
235 static struct rx_call * rxi_NewCall(struct rx_connection *, int);
238 putConnection (struct rx_connection *conn) {
239 MUTEX_ENTER(&rx_refcnt_mutex);
241 MUTEX_EXIT(&rx_refcnt_mutex);
244 #ifdef AFS_PTHREAD_ENV
247 * Use procedural initialization of mutexes/condition variables
251 extern afs_kmutex_t rx_quota_mutex;
252 extern afs_kmutex_t rx_pthread_mutex;
253 extern afs_kmutex_t rx_packets_mutex;
254 extern afs_kmutex_t rx_refcnt_mutex;
255 extern afs_kmutex_t des_init_mutex;
256 extern afs_kmutex_t des_random_mutex;
257 extern afs_kmutex_t rx_clock_mutex;
258 extern afs_kmutex_t rxi_connCacheMutex;
259 extern afs_kmutex_t event_handler_mutex;
260 extern afs_kmutex_t listener_mutex;
261 extern afs_kmutex_t rx_if_init_mutex;
262 extern afs_kmutex_t rx_if_mutex;
264 extern afs_kcondvar_t rx_event_handler_cond;
265 extern afs_kcondvar_t rx_listener_cond;
267 static afs_kmutex_t epoch_mutex;
268 static afs_kmutex_t rx_init_mutex;
269 static afs_kmutex_t rx_debug_mutex;
270 static afs_kmutex_t rx_rpc_stats;
273 rxi_InitPthread(void)
275 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
276 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
277 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
278 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
279 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
280 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
281 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
282 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
283 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
284 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
285 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
286 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
287 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
288 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
289 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
291 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
292 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
294 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
295 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
297 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
298 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
299 #ifdef RX_ENABLE_LOCKS
302 #endif /* RX_LOCKS_DB */
303 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
304 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
306 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
308 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
310 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
312 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
313 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
314 #endif /* RX_ENABLE_LOCKS */
317 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
318 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
320 * The rx_stats_mutex mutex protects the following global variables:
321 * rxi_lowConnRefCount
322 * rxi_lowPeerRefCount
331 * The rx_quota_mutex mutex protects the following global variables:
339 * The rx_freePktQ_lock protects the following global variables:
344 * The rx_packets_mutex mutex protects the following global variables:
352 * The rx_pthread_mutex mutex protects the following global variables:
353 * rxi_fcfs_thread_num
356 #define INIT_PTHREAD_LOCKS
360 /* Variables for handling the minProcs implementation. availProcs gives the
361 * number of threads available in the pool at this moment (not counting dudes
362 * executing right now). totalMin gives the total number of procs required
363 * for handling all minProcs requests. minDeficit is a dynamic variable
364 * tracking the # of procs required to satisfy all of the remaining minProcs
366 * For fine grain locking to work, the quota check and the reservation of
367 * a server thread has to come while rxi_availProcs and rxi_minDeficit
368 * are locked. To this end, the code has been modified under #ifdef
369 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
370 * same time. A new function, ReturnToServerPool() returns the allocation.
372 * A call can be on several queue's (but only one at a time). When
373 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
374 * that no one else is touching the queue. To this end, we store the address
375 * of the queue lock in the call structure (under the call lock) when we
376 * put the call on a queue, and we clear the call_queue_lock when the
377 * call is removed from a queue (once the call lock has been obtained).
378 * This allows rxi_ResetCall to safely synchronize with others wishing
379 * to manipulate the queue.
382 #if defined(RX_ENABLE_LOCKS)
383 static afs_kmutex_t rx_rpc_stats;
386 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
387 ** pretty good that the next packet coming in is from the same connection
388 ** as the last packet, since we're send multiple packets in a transmit window.
390 struct rx_connection *rxLastConn = 0;
392 #ifdef RX_ENABLE_LOCKS
393 /* The locking hierarchy for rx fine grain locking is composed of these
396 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
397 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
398 * call->lock - locks call data fields.
399 * These are independent of each other:
400 * rx_freeCallQueue_lock
405 * serverQueueEntry->lock
406 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
408 * peer->lock - locks peer data fields.
409 * conn_data_lock - that more than one thread is not updating a conn data
410 * field at the same time.
421 * Do we need a lock to protect the peer field in the conn structure?
422 * conn->peer was previously a constant for all intents and so has no
423 * lock protecting this field. The multihomed client delta introduced
424 * a RX code change : change the peer field in the connection structure
425 * to that remote interface from which the last packet for this
426 * connection was sent out. This may become an issue if further changes
429 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
430 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
432 /* rxdb_fileID is used to identify the lock location, along with line#. */
433 static int rxdb_fileID = RXDB_FILE_RX;
434 #endif /* RX_LOCKS_DB */
435 #else /* RX_ENABLE_LOCKS */
436 #define SET_CALL_QUEUE_LOCK(C, L)
437 #define CLEAR_CALL_QUEUE_LOCK(C)
438 #endif /* RX_ENABLE_LOCKS */
439 struct rx_serverQueueEntry *rx_waitForPacket = 0;
440 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
442 /* ------------Exported Interfaces------------- */
444 /* This function allows rxkad to set the epoch to a suitably random number
445 * which rx_NewConnection will use in the future. The principle purpose is to
446 * get rxnull connections to use the same epoch as the rxkad connections do, at
447 * least once the first rxkad connection is established. This is important now
448 * that the host/port addresses aren't used in FindConnection: the uniqueness
449 * of epoch/cid matters and the start time won't do. */
451 #ifdef AFS_PTHREAD_ENV
453 * This mutex protects the following global variables:
457 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
458 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
462 #endif /* AFS_PTHREAD_ENV */
465 rx_SetEpoch(afs_uint32 epoch)
472 /* Initialize rx. A port number may be mentioned, in which case this
473 * becomes the default port number for any service installed later.
474 * If 0 is provided for the port number, a random port will be chosen
475 * by the kernel. Whether this will ever overlap anything in
476 * /etc/services is anybody's guess... Returns 0 on success, -1 on
481 rx_atomic_t rxinit_status = RX_ATOMIC_INIT(1);
484 rx_InitHost(u_int host, u_int port)
491 char *htable, *ptable;
496 if (!rx_atomic_test_and_clear_bit(&rxinit_status, 0))
497 return 0; /* already started */
503 if (afs_winsockInit() < 0)
509 * Initialize anything necessary to provide a non-premptive threading
512 rxi_InitializeThreadSupport();
515 /* Allocate and initialize a socket for client and perhaps server
518 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
519 if (rx_socket == OSI_NULLSOCKET) {
522 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
525 #endif /* RX_LOCKS_DB */
526 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
527 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
528 MUTEX_INIT(&rx_atomic_mutex, "rx_atomic_mutex", MUTEX_DEFAULT, 0);
529 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
530 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
531 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
532 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
533 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
534 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
535 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
537 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
539 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
541 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
543 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
544 #if defined(AFS_HPUX110_ENV)
546 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
547 #endif /* AFS_HPUX110_ENV */
548 #endif /* RX_ENABLE_LOCKS && KERNEL */
551 rx_connDeadTime = 12;
552 rx_tranquil = 0; /* reset flag */
553 rxi_ResetStatistics();
554 htable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
555 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
556 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
557 ptable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
558 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
559 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
561 /* Malloc up a bunch of packets & buffers */
563 opr_queue_Init(&rx_freePacketQueue);
564 rxi_NeedMorePackets = FALSE;
565 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
567 /* enforce a minimum number of allocated packets */
568 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
569 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
571 /* allocate the initial free packet pool */
572 #ifdef RX_ENABLE_TSFPQ
573 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
574 #else /* RX_ENABLE_TSFPQ */
575 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
576 #endif /* RX_ENABLE_TSFPQ */
583 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
584 tv.tv_sec = clock_now.sec;
585 tv.tv_usec = clock_now.usec;
586 srand((unsigned int)tv.tv_usec);
593 #if defined(KERNEL) && !defined(UKERNEL)
594 /* Really, this should never happen in a real kernel */
597 struct sockaddr_in addr;
599 int addrlen = sizeof(addr);
601 socklen_t addrlen = sizeof(addr);
603 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
605 osi_Free(htable, rx_hashTableSize * sizeof(struct rx_connection *));
608 rx_port = addr.sin_port;
611 rx_stats.minRtt.sec = 9999999;
613 rx_SetEpoch(tv.tv_sec | 0x80000000);
615 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
616 * will provide a randomer value. */
618 MUTEX_ENTER(&rx_quota_mutex);
619 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
620 MUTEX_EXIT(&rx_quota_mutex);
621 /* *Slightly* random start time for the cid. This is just to help
622 * out with the hashing function at the peer */
623 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
624 rx_connHashTable = (struct rx_connection **)htable;
625 rx_peerHashTable = (struct rx_peer **)ptable;
627 rx_hardAckDelay.sec = 0;
628 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
630 rxevent_Init(20, rxi_ReScheduleEvents);
632 /* Initialize various global queues */
633 opr_queue_Init(&rx_idleServerQueue);
634 opr_queue_Init(&rx_incomingCallQueue);
635 opr_queue_Init(&rx_freeCallQueue);
637 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
638 /* Initialize our list of usable IP addresses. */
642 /* Start listener process (exact function is dependent on the
643 * implementation environment--kernel or user space) */
647 rx_atomic_clear_bit(&rxinit_status, 0);
654 return rx_InitHost(htonl(INADDR_ANY), port);
660 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
661 * maintaing the round trip timer.
666 * Start a new RTT timer for a given call and packet.
668 * There must be no resendEvent already listed for this call, otherwise this
669 * will leak events - intended for internal use within the RTO code only
672 * the RX call to start the timer for
673 * @param[in] lastPacket
674 * a flag indicating whether the last packet has been sent or not
676 * @pre call must be locked before calling this function
680 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
682 struct clock now, retryTime;
687 clock_Add(&retryTime, &call->rto);
689 /* If we're sending the last packet, and we're the client, then the server
690 * may wait for an additional 400ms before returning the ACK, wait for it
691 * rather than hitting a timeout */
692 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
693 clock_Addmsec(&retryTime, 400);
695 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
696 call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
701 * Cancel an RTT timer for a given call.
705 * the RX call to cancel the timer for
707 * @pre call must be locked before calling this function
712 rxi_rto_cancel(struct rx_call *call)
714 rxevent_Cancel(&call->resendEvent);
715 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
719 * Tell the RTO timer that we have sent a packet.
721 * If the timer isn't already running, then start it. If the timer is running,
725 * the RX call that the packet has been sent on
726 * @param[in] lastPacket
727 * A flag which is true if this is the last packet for the call
729 * @pre The call must be locked before calling this function
734 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
736 if (call->resendEvent)
739 rxi_rto_startTimer(call, lastPacket, istack);
743 * Tell the RTO timer that we have received an new ACK message
745 * This function should be called whenever a call receives an ACK that
746 * acknowledges new packets. Whatever happens, we stop the current timer.
747 * If there are unacked packets in the queue which have been sent, then
748 * we restart the timer from now. Otherwise, we leave it stopped.
751 * the RX call that the ACK has been received on
755 rxi_rto_packet_acked(struct rx_call *call, int istack)
757 struct opr_queue *cursor;
759 rxi_rto_cancel(call);
761 if (opr_queue_IsEmpty(&call->tq))
764 for (opr_queue_Scan(&call->tq, cursor)) {
765 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
766 if (p->header.seq > call->tfirst + call->twind)
769 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
770 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
778 * Set an initial round trip timeout for a peer connection
780 * @param[in] secs The timeout to set in seconds
784 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
785 peer->rtt = secs * 8000;
789 * Enables or disables the busy call channel error (RX_CALL_BUSY).
791 * @param[in] onoff Non-zero to enable busy call channel errors.
793 * @pre Neither rx_Init nor rx_InitHost have been called yet
796 rx_SetBusyChannelError(afs_int32 onoff)
798 osi_Assert(rx_atomic_test_bit(&rxinit_status, 0));
799 rxi_busyChannelError = onoff ? 1 : 0;
803 * Set a delayed ack event on the specified call for the given time
805 * @param[in] call - the call on which to set the event
806 * @param[in] offset - the delay from now after which the event fires
809 rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
811 struct clock now, when;
815 clock_Add(&when, offset);
817 if (call->delayedAckEvent && clock_Gt(&call->delayedAckTime, &when)) {
818 /* The event we're cancelling already has a reference, so we don't
820 rxevent_Cancel(&call->delayedAckEvent);
821 call->delayedAckEvent = rxevent_Post(&when, &now, rxi_SendDelayedAck,
824 call->delayedAckTime = when;
825 } else if (!call->delayedAckEvent) {
826 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
827 call->delayedAckEvent = rxevent_Post(&when, &now,
830 call->delayedAckTime = when;
835 rxi_CancelDelayedAckEvent(struct rx_call *call)
837 if (call->delayedAckEvent) {
838 rxevent_Cancel(&call->delayedAckEvent);
839 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
843 /* called with unincremented nRequestsRunning to see if it is OK to start
844 * a new thread in this service. Could be "no" for two reasons: over the
845 * max quota, or would prevent others from reaching their min quota.
847 #ifdef RX_ENABLE_LOCKS
848 /* This verion of QuotaOK reserves quota if it's ok while the
849 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
852 QuotaOK(struct rx_service *aservice)
854 /* check if over max quota */
855 if (aservice->nRequestsRunning >= aservice->maxProcs) {
859 /* under min quota, we're OK */
860 /* otherwise, can use only if there are enough to allow everyone
861 * to go to their min quota after this guy starts.
864 MUTEX_ENTER(&rx_quota_mutex);
865 if ((aservice->nRequestsRunning < aservice->minProcs)
866 || (rxi_availProcs > rxi_minDeficit)) {
867 aservice->nRequestsRunning++;
868 /* just started call in minProcs pool, need fewer to maintain
870 if (aservice->nRequestsRunning <= aservice->minProcs)
873 MUTEX_EXIT(&rx_quota_mutex);
876 MUTEX_EXIT(&rx_quota_mutex);
882 ReturnToServerPool(struct rx_service *aservice)
884 aservice->nRequestsRunning--;
885 MUTEX_ENTER(&rx_quota_mutex);
886 if (aservice->nRequestsRunning < aservice->minProcs)
889 MUTEX_EXIT(&rx_quota_mutex);
892 #else /* RX_ENABLE_LOCKS */
894 QuotaOK(struct rx_service *aservice)
897 /* under min quota, we're OK */
898 if (aservice->nRequestsRunning < aservice->minProcs)
901 /* check if over max quota */
902 if (aservice->nRequestsRunning >= aservice->maxProcs)
905 /* otherwise, can use only if there are enough to allow everyone
906 * to go to their min quota after this guy starts.
908 MUTEX_ENTER(&rx_quota_mutex);
909 if (rxi_availProcs > rxi_minDeficit)
911 MUTEX_EXIT(&rx_quota_mutex);
914 #endif /* RX_ENABLE_LOCKS */
917 /* Called by rx_StartServer to start up lwp's to service calls.
918 NExistingProcs gives the number of procs already existing, and which
919 therefore needn't be created. */
921 rxi_StartServerProcs(int nExistingProcs)
923 struct rx_service *service;
928 /* For each service, reserve N processes, where N is the "minimum"
929 * number of processes that MUST be able to execute a request in parallel,
930 * at any time, for that process. Also compute the maximum difference
931 * between any service's maximum number of processes that can run
932 * (i.e. the maximum number that ever will be run, and a guarantee
933 * that this number will run if other services aren't running), and its
934 * minimum number. The result is the extra number of processes that
935 * we need in order to provide the latter guarantee */
936 for (i = 0; i < RX_MAX_SERVICES; i++) {
938 service = rx_services[i];
939 if (service == (struct rx_service *)0)
941 nProcs += service->minProcs;
942 diff = service->maxProcs - service->minProcs;
946 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
947 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
948 for (i = 0; i < nProcs; i++) {
949 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
955 /* This routine is only required on Windows */
957 rx_StartClientThread(void)
959 #ifdef AFS_PTHREAD_ENV
961 pid = pthread_self();
962 #endif /* AFS_PTHREAD_ENV */
964 #endif /* AFS_NT40_ENV */
966 /* This routine must be called if any services are exported. If the
967 * donateMe flag is set, the calling process is donated to the server
970 rx_StartServer(int donateMe)
972 struct rx_service *service;
978 /* Start server processes, if necessary (exact function is dependent
979 * on the implementation environment--kernel or user space). DonateMe
980 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
981 * case, one less new proc will be created rx_StartServerProcs.
983 rxi_StartServerProcs(donateMe);
985 /* count up the # of threads in minProcs, and add set the min deficit to
986 * be that value, too.
988 for (i = 0; i < RX_MAX_SERVICES; i++) {
989 service = rx_services[i];
990 if (service == (struct rx_service *)0)
992 MUTEX_ENTER(&rx_quota_mutex);
993 rxi_totalMin += service->minProcs;
994 /* below works even if a thread is running, since minDeficit would
995 * still have been decremented and later re-incremented.
997 rxi_minDeficit += service->minProcs;
998 MUTEX_EXIT(&rx_quota_mutex);
1001 /* Turn on reaping of idle server connections */
1002 rxi_ReapConnections(NULL, NULL, NULL, 0);
1007 #ifndef AFS_NT40_ENV
1011 #ifdef AFS_PTHREAD_ENV
1013 pid = afs_pointer_to_int(pthread_self());
1014 #else /* AFS_PTHREAD_ENV */
1016 LWP_CurrentProcess(&pid);
1017 #endif /* AFS_PTHREAD_ENV */
1019 sprintf(name, "srv_%d", ++nProcs);
1020 if (registerProgram)
1021 (*registerProgram) (pid, name);
1023 #endif /* AFS_NT40_ENV */
1024 rx_ServerProc(NULL); /* Never returns */
1026 #ifdef RX_ENABLE_TSFPQ
1027 /* no use leaving packets around in this thread's local queue if
1028 * it isn't getting donated to the server thread pool.
1030 rxi_FlushLocalPacketsTSFPQ();
1031 #endif /* RX_ENABLE_TSFPQ */
1035 /* Create a new client connection to the specified service, using the
1036 * specified security object to implement the security model for this
1038 struct rx_connection *
1039 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
1040 struct rx_securityClass *securityObject,
1041 int serviceSecurityIndex)
1045 struct rx_connection *conn;
1050 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
1051 "serviceSecurityIndex %d)\n",
1052 ntohl(shost), ntohs(sport), sservice, securityObject,
1053 serviceSecurityIndex));
1055 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
1056 * the case of kmem_alloc? */
1057 conn = rxi_AllocConnection();
1058 #ifdef RX_ENABLE_LOCKS
1059 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
1060 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
1061 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
1064 MUTEX_ENTER(&rx_connHashTable_lock);
1065 cid = (rx_nextCid += RX_MAXCALLS);
1066 conn->type = RX_CLIENT_CONNECTION;
1068 conn->epoch = rx_epoch;
1069 conn->peer = rxi_FindPeer(shost, sport, 1);
1070 conn->serviceId = sservice;
1071 conn->securityObject = securityObject;
1072 conn->securityData = (void *) 0;
1073 conn->securityIndex = serviceSecurityIndex;
1074 rx_SetConnDeadTime(conn, rx_connDeadTime);
1075 rx_SetConnSecondsUntilNatPing(conn, 0);
1076 conn->ackRate = RX_FAST_ACK_RATE;
1077 conn->nSpecific = 0;
1078 conn->specific = NULL;
1079 conn->challengeEvent = NULL;
1080 conn->delayedAbortEvent = NULL;
1081 conn->abortCount = 0;
1083 for (i = 0; i < RX_MAXCALLS; i++) {
1084 conn->twind[i] = rx_initSendWindow;
1085 conn->rwind[i] = rx_initReceiveWindow;
1086 conn->lastBusy[i] = 0;
1089 RXS_NewConnection(securityObject, conn);
1091 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1093 conn->refCount++; /* no lock required since only this thread knows... */
1094 conn->next = rx_connHashTable[hashindex];
1095 rx_connHashTable[hashindex] = conn;
1096 if (rx_stats_active)
1097 rx_atomic_inc(&rx_stats.nClientConns);
1098 MUTEX_EXIT(&rx_connHashTable_lock);
1104 * Ensure a connection's timeout values are valid.
1106 * @param[in] conn The connection to check
1108 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1109 * unless idleDeadTime and/or hardDeadTime are not set
1113 rxi_CheckConnTimeouts(struct rx_connection *conn)
1115 /* a connection's timeouts must have the relationship
1116 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1117 * total loss of network to a peer may cause an idle timeout instead of a
1118 * dead timeout, simply because the idle timeout gets hit first. Also set
1119 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1120 /* this logic is slightly complicated by the fact that
1121 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1123 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1124 if (conn->idleDeadTime) {
1125 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1127 if (conn->hardDeadTime) {
1128 if (conn->idleDeadTime) {
1129 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1131 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1137 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1139 /* The idea is to set the dead time to a value that allows several
1140 * keepalives to be dropped without timing out the connection. */
1141 conn->secondsUntilDead = seconds;
1142 rxi_CheckConnTimeouts(conn);
1143 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1147 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1149 conn->hardDeadTime = seconds;
1150 rxi_CheckConnTimeouts(conn);
1154 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1156 conn->idleDeadTime = seconds;
1157 conn->idleDeadDetection = (seconds ? 1 : 0);
1158 rxi_CheckConnTimeouts(conn);
1161 int rxi_lowPeerRefCount = 0;
1162 int rxi_lowConnRefCount = 0;
1165 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1166 * NOTE: must not be called with rx_connHashTable_lock held.
1169 rxi_CleanupConnection(struct rx_connection *conn)
1171 /* Notify the service exporter, if requested, that this connection
1172 * is being destroyed */
1173 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1174 (*conn->service->destroyConnProc) (conn);
1176 /* Notify the security module that this connection is being destroyed */
1177 RXS_DestroyConnection(conn->securityObject, conn);
1179 /* If this is the last connection using the rx_peer struct, set its
1180 * idle time to now. rxi_ReapConnections will reap it if it's still
1181 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1183 MUTEX_ENTER(&rx_peerHashTable_lock);
1184 if (conn->peer->refCount < 2) {
1185 conn->peer->idleWhen = clock_Sec();
1186 if (conn->peer->refCount < 1) {
1187 conn->peer->refCount = 1;
1188 if (rx_stats_active) {
1189 MUTEX_ENTER(&rx_stats_mutex);
1190 rxi_lowPeerRefCount++;
1191 MUTEX_EXIT(&rx_stats_mutex);
1195 conn->peer->refCount--;
1196 MUTEX_EXIT(&rx_peerHashTable_lock);
1198 if (rx_stats_active)
1200 if (conn->type == RX_SERVER_CONNECTION)
1201 rx_atomic_dec(&rx_stats.nServerConns);
1203 rx_atomic_dec(&rx_stats.nClientConns);
1206 if (conn->specific) {
1208 for (i = 0; i < conn->nSpecific; i++) {
1209 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1210 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1211 conn->specific[i] = NULL;
1213 free(conn->specific);
1215 conn->specific = NULL;
1216 conn->nSpecific = 0;
1217 #endif /* !KERNEL */
1219 MUTEX_DESTROY(&conn->conn_call_lock);
1220 MUTEX_DESTROY(&conn->conn_data_lock);
1221 CV_DESTROY(&conn->conn_call_cv);
1223 rxi_FreeConnection(conn);
1226 /* Destroy the specified connection */
1228 rxi_DestroyConnection(struct rx_connection *conn)
1230 MUTEX_ENTER(&rx_connHashTable_lock);
1231 rxi_DestroyConnectionNoLock(conn);
1232 /* conn should be at the head of the cleanup list */
1233 if (conn == rx_connCleanup_list) {
1234 rx_connCleanup_list = rx_connCleanup_list->next;
1235 MUTEX_EXIT(&rx_connHashTable_lock);
1236 rxi_CleanupConnection(conn);
1238 #ifdef RX_ENABLE_LOCKS
1240 MUTEX_EXIT(&rx_connHashTable_lock);
1242 #endif /* RX_ENABLE_LOCKS */
1246 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1248 struct rx_connection **conn_ptr;
1250 struct rx_packet *packet;
1257 MUTEX_ENTER(&conn->conn_data_lock);
1258 MUTEX_ENTER(&rx_refcnt_mutex);
1259 if (conn->refCount > 0)
1262 if (rx_stats_active) {
1263 MUTEX_ENTER(&rx_stats_mutex);
1264 rxi_lowConnRefCount++;
1265 MUTEX_EXIT(&rx_stats_mutex);
1269 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1270 /* Busy; wait till the last guy before proceeding */
1271 MUTEX_EXIT(&rx_refcnt_mutex);
1272 MUTEX_EXIT(&conn->conn_data_lock);
1277 /* If the client previously called rx_NewCall, but it is still
1278 * waiting, treat this as a running call, and wait to destroy the
1279 * connection later when the call completes. */
1280 if ((conn->type == RX_CLIENT_CONNECTION)
1281 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1282 conn->flags |= RX_CONN_DESTROY_ME;
1283 MUTEX_EXIT(&conn->conn_data_lock);
1287 MUTEX_EXIT(&rx_refcnt_mutex);
1288 MUTEX_EXIT(&conn->conn_data_lock);
1290 /* Check for extant references to this connection */
1291 MUTEX_ENTER(&conn->conn_call_lock);
1292 for (i = 0; i < RX_MAXCALLS; i++) {
1293 struct rx_call *call = conn->call[i];
1296 if (conn->type == RX_CLIENT_CONNECTION) {
1297 MUTEX_ENTER(&call->lock);
1298 if (call->delayedAckEvent) {
1299 /* Push the final acknowledgment out now--there
1300 * won't be a subsequent call to acknowledge the
1301 * last reply packets */
1302 rxi_CancelDelayedAckEvent(call);
1303 if (call->state == RX_STATE_PRECALL
1304 || call->state == RX_STATE_ACTIVE) {
1305 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1310 MUTEX_EXIT(&call->lock);
1314 MUTEX_EXIT(&conn->conn_call_lock);
1316 #ifdef RX_ENABLE_LOCKS
1318 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1319 MUTEX_EXIT(&conn->conn_data_lock);
1321 /* Someone is accessing a packet right now. */
1325 #endif /* RX_ENABLE_LOCKS */
1328 /* Don't destroy the connection if there are any call
1329 * structures still in use */
1330 MUTEX_ENTER(&conn->conn_data_lock);
1331 conn->flags |= RX_CONN_DESTROY_ME;
1332 MUTEX_EXIT(&conn->conn_data_lock);
1337 if (conn->natKeepAliveEvent) {
1338 rxi_NatKeepAliveOff(conn);
1341 if (conn->delayedAbortEvent) {
1342 rxevent_Cancel(&conn->delayedAbortEvent);
1343 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1345 MUTEX_ENTER(&conn->conn_data_lock);
1346 rxi_SendConnectionAbort(conn, packet, 0, 1);
1347 MUTEX_EXIT(&conn->conn_data_lock);
1348 rxi_FreePacket(packet);
1352 /* Remove from connection hash table before proceeding */
1354 &rx_connHashTable[CONN_HASH
1355 (peer->host, peer->port, conn->cid, conn->epoch,
1357 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1358 if (*conn_ptr == conn) {
1359 *conn_ptr = conn->next;
1363 /* if the conn that we are destroying was the last connection, then we
1364 * clear rxLastConn as well */
1365 if (rxLastConn == conn)
1368 /* Make sure the connection is completely reset before deleting it. */
1369 /* get rid of pending events that could zap us later */
1370 rxevent_Cancel(&conn->challengeEvent);
1371 rxevent_Cancel(&conn->checkReachEvent);
1372 rxevent_Cancel(&conn->natKeepAliveEvent);
1374 /* Add the connection to the list of destroyed connections that
1375 * need to be cleaned up. This is necessary to avoid deadlocks
1376 * in the routines we call to inform others that this connection is
1377 * being destroyed. */
1378 conn->next = rx_connCleanup_list;
1379 rx_connCleanup_list = conn;
1382 /* Externally available version */
1384 rx_DestroyConnection(struct rx_connection *conn)
1389 rxi_DestroyConnection(conn);
1394 rx_GetConnection(struct rx_connection *conn)
1399 MUTEX_ENTER(&rx_refcnt_mutex);
1401 MUTEX_EXIT(&rx_refcnt_mutex);
1405 #ifdef RX_ENABLE_LOCKS
1406 /* Wait for the transmit queue to no longer be busy.
1407 * requires the call->lock to be held */
1409 rxi_WaitforTQBusy(struct rx_call *call) {
1410 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1411 call->flags |= RX_CALL_TQ_WAIT;
1413 MUTEX_ASSERT(&call->lock);
1414 CV_WAIT(&call->cv_tq, &call->lock);
1416 if (call->tqWaiters == 0) {
1417 call->flags &= ~RX_CALL_TQ_WAIT;
1424 rxi_WakeUpTransmitQueue(struct rx_call *call)
1426 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1427 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1428 call, call->tqWaiters, call->flags));
1429 #ifdef RX_ENABLE_LOCKS
1430 MUTEX_ASSERT(&call->lock);
1431 CV_BROADCAST(&call->cv_tq);
1432 #else /* RX_ENABLE_LOCKS */
1433 osi_rxWakeup(&call->tq);
1434 #endif /* RX_ENABLE_LOCKS */
1438 /* Start a new rx remote procedure call, on the specified connection.
1439 * If wait is set to 1, wait for a free call channel; otherwise return
1440 * 0. Maxtime gives the maximum number of seconds this call may take,
1441 * after rx_NewCall returns. After this time interval, a call to any
1442 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1443 * For fine grain locking, we hold the conn_call_lock in order to
1444 * to ensure that we don't get signalle after we found a call in an active
1445 * state and before we go to sleep.
1448 rx_NewCall(struct rx_connection *conn)
1450 int i, wait, ignoreBusy = 1;
1451 struct rx_call *call;
1452 struct clock queueTime;
1453 afs_uint32 leastBusy = 0;
1457 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1460 clock_GetTime(&queueTime);
1462 * Check if there are others waiting for a new call.
1463 * If so, let them go first to avoid starving them.
1464 * This is a fairly simple scheme, and might not be
1465 * a complete solution for large numbers of waiters.
1467 * makeCallWaiters keeps track of the number of
1468 * threads waiting to make calls and the
1469 * RX_CONN_MAKECALL_WAITING flag bit is used to
1470 * indicate that there are indeed calls waiting.
1471 * The flag is set when the waiter is incremented.
1472 * It is only cleared when makeCallWaiters is 0.
1473 * This prevents us from accidently destroying the
1474 * connection while it is potentially about to be used.
1476 MUTEX_ENTER(&conn->conn_call_lock);
1477 MUTEX_ENTER(&conn->conn_data_lock);
1478 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1479 conn->flags |= RX_CONN_MAKECALL_WAITING;
1480 conn->makeCallWaiters++;
1481 MUTEX_EXIT(&conn->conn_data_lock);
1483 #ifdef RX_ENABLE_LOCKS
1484 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1488 MUTEX_ENTER(&conn->conn_data_lock);
1489 conn->makeCallWaiters--;
1490 if (conn->makeCallWaiters == 0)
1491 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1494 /* We are now the active thread in rx_NewCall */
1495 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1496 MUTEX_EXIT(&conn->conn_data_lock);
1501 for (i = 0; i < RX_MAXCALLS; i++) {
1502 call = conn->call[i];
1504 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1505 /* we're not ignoring busy call slots; only look at the
1506 * call slot that is the "least" busy */
1510 if (call->state == RX_STATE_DALLY) {
1511 MUTEX_ENTER(&call->lock);
1512 if (call->state == RX_STATE_DALLY) {
1513 if (ignoreBusy && conn->lastBusy[i]) {
1514 /* if we're ignoring busy call slots, skip any ones that
1515 * have lastBusy set */
1516 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1517 leastBusy = conn->lastBusy[i];
1519 MUTEX_EXIT(&call->lock);
1524 * We are setting the state to RX_STATE_RESET to
1525 * ensure that no one else will attempt to use this
1526 * call once we drop the conn->conn_call_lock and
1527 * call->lock. We must drop the conn->conn_call_lock
1528 * before calling rxi_ResetCall because the process
1529 * of clearing the transmit queue can block for an
1530 * extended period of time. If we block while holding
1531 * the conn->conn_call_lock, then all rx_EndCall
1532 * processing will block as well. This has a detrimental
1533 * effect on overall system performance.
1535 call->state = RX_STATE_RESET;
1536 (*call->callNumber)++;
1537 MUTEX_EXIT(&conn->conn_call_lock);
1538 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1539 rxi_ResetCall(call, 0);
1540 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1544 * If we failed to be able to safely obtain the
1545 * conn->conn_call_lock we will have to drop the
1546 * call->lock to avoid a deadlock. When the call->lock
1547 * is released the state of the call can change. If it
1548 * is no longer RX_STATE_RESET then some other thread is
1551 MUTEX_EXIT(&call->lock);
1552 MUTEX_ENTER(&conn->conn_call_lock);
1553 MUTEX_ENTER(&call->lock);
1555 if (call->state == RX_STATE_RESET)
1559 * If we get here it means that after dropping
1560 * the conn->conn_call_lock and call->lock that
1561 * the call is no longer ours. If we can't find
1562 * a free call in the remaining slots we should
1563 * not go immediately to RX_CONN_MAKECALL_WAITING
1564 * because by dropping the conn->conn_call_lock
1565 * we have given up synchronization with rx_EndCall.
1566 * Instead, cycle through one more time to see if
1567 * we can find a call that can call our own.
1569 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1572 MUTEX_EXIT(&call->lock);
1575 if (ignoreBusy && conn->lastBusy[i]) {
1576 /* if we're ignoring busy call slots, skip any ones that
1577 * have lastBusy set */
1578 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1579 leastBusy = conn->lastBusy[i];
1584 /* rxi_NewCall returns with mutex locked */
1585 call = rxi_NewCall(conn, i);
1586 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1590 if (i < RX_MAXCALLS) {
1591 conn->lastBusy[i] = 0;
1592 call->flags &= ~RX_CALL_PEER_BUSY;
1597 if (leastBusy && ignoreBusy) {
1598 /* we didn't find a useable call slot, but we did see at least one
1599 * 'busy' slot; look again and only use a slot with the 'least
1605 MUTEX_ENTER(&conn->conn_data_lock);
1606 conn->flags |= RX_CONN_MAKECALL_WAITING;
1607 conn->makeCallWaiters++;
1608 MUTEX_EXIT(&conn->conn_data_lock);
1610 #ifdef RX_ENABLE_LOCKS
1611 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1615 MUTEX_ENTER(&conn->conn_data_lock);
1616 conn->makeCallWaiters--;
1617 if (conn->makeCallWaiters == 0)
1618 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1619 MUTEX_EXIT(&conn->conn_data_lock);
1621 /* Client is initially in send mode */
1622 call->state = RX_STATE_ACTIVE;
1623 call->error = conn->error;
1625 call->app.mode = RX_MODE_ERROR;
1627 call->app.mode = RX_MODE_SENDING;
1629 #ifdef AFS_RXERRQ_ENV
1630 /* remember how many network errors the peer has when we started, so if
1631 * more errors are encountered after the call starts, we know the other endpoint won't be
1632 * responding to us */
1633 call->neterr_gen = rx_atomic_read(&conn->peer->neterrs);
1636 /* remember start time for call in case we have hard dead time limit */
1637 call->queueTime = queueTime;
1638 clock_GetTime(&call->startTime);
1639 call->app.bytesSent = 0;
1640 call->app.bytesRcvd = 0;
1642 /* Turn on busy protocol. */
1643 rxi_KeepAliveOn(call);
1645 /* Attempt MTU discovery */
1646 rxi_GrowMTUOn(call);
1649 * We are no longer the active thread in rx_NewCall
1651 MUTEX_ENTER(&conn->conn_data_lock);
1652 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1653 MUTEX_EXIT(&conn->conn_data_lock);
1656 * Wake up anyone else who might be giving us a chance to
1657 * run (see code above that avoids resource starvation).
1659 #ifdef RX_ENABLE_LOCKS
1660 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1661 osi_Panic("rx_NewCall call about to be used without an empty tq");
1664 CV_BROADCAST(&conn->conn_call_cv);
1668 MUTEX_EXIT(&conn->conn_call_lock);
1669 MUTEX_EXIT(&call->lock);
1672 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1677 rxi_HasActiveCalls(struct rx_connection *aconn)
1680 struct rx_call *tcall;
1684 for (i = 0; i < RX_MAXCALLS; i++) {
1685 if ((tcall = aconn->call[i])) {
1686 if ((tcall->state == RX_STATE_ACTIVE)
1687 || (tcall->state == RX_STATE_PRECALL)) {
1698 rxi_GetCallNumberVector(struct rx_connection *aconn,
1699 afs_int32 * aint32s)
1702 struct rx_call *tcall;
1706 MUTEX_ENTER(&aconn->conn_call_lock);
1707 for (i = 0; i < RX_MAXCALLS; i++) {
1708 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1709 aint32s[i] = aconn->callNumber[i] + 1;
1711 aint32s[i] = aconn->callNumber[i];
1713 MUTEX_EXIT(&aconn->conn_call_lock);
1719 rxi_SetCallNumberVector(struct rx_connection *aconn,
1720 afs_int32 * aint32s)
1723 struct rx_call *tcall;
1727 MUTEX_ENTER(&aconn->conn_call_lock);
1728 for (i = 0; i < RX_MAXCALLS; i++) {
1729 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1730 aconn->callNumber[i] = aint32s[i] - 1;
1732 aconn->callNumber[i] = aint32s[i];
1734 MUTEX_EXIT(&aconn->conn_call_lock);
1739 /* Advertise a new service. A service is named locally by a UDP port
1740 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1743 char *serviceName; Name for identification purposes (e.g. the
1744 service name might be used for probing for
1747 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1748 char *serviceName, struct rx_securityClass **securityObjects,
1749 int nSecurityObjects,
1750 afs_int32(*serviceProc) (struct rx_call * acall))
1752 osi_socket socket = OSI_NULLSOCKET;
1753 struct rx_service *tservice;
1759 if (serviceId == 0) {
1761 "rx_NewService: service id for service %s is not non-zero.\n",
1768 "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",
1776 tservice = rxi_AllocService();
1779 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1781 for (i = 0; i < RX_MAX_SERVICES; i++) {
1782 struct rx_service *service = rx_services[i];
1784 if (port == service->servicePort && host == service->serviceHost) {
1785 if (service->serviceId == serviceId) {
1786 /* The identical service has already been
1787 * installed; if the caller was intending to
1788 * change the security classes used by this
1789 * service, he/she loses. */
1791 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1792 serviceName, serviceId, service->serviceName);
1794 rxi_FreeService(tservice);
1797 /* Different service, same port: re-use the socket
1798 * which is bound to the same port */
1799 socket = service->socket;
1802 if (socket == OSI_NULLSOCKET) {
1803 /* If we don't already have a socket (from another
1804 * service on same port) get a new one */
1805 socket = rxi_GetHostUDPSocket(host, port);
1806 if (socket == OSI_NULLSOCKET) {
1808 rxi_FreeService(tservice);
1813 service->socket = socket;
1814 service->serviceHost = host;
1815 service->servicePort = port;
1816 service->serviceId = serviceId;
1817 service->serviceName = serviceName;
1818 service->nSecurityObjects = nSecurityObjects;
1819 service->securityObjects = securityObjects;
1820 service->minProcs = 0;
1821 service->maxProcs = 1;
1822 service->idleDeadTime = 60;
1823 service->idleDeadErr = 0;
1824 service->connDeadTime = rx_connDeadTime;
1825 service->executeRequestProc = serviceProc;
1826 service->checkReach = 0;
1827 service->nSpecific = 0;
1828 service->specific = NULL;
1829 rx_services[i] = service; /* not visible until now */
1835 rxi_FreeService(tservice);
1836 (osi_Msg "rx_NewService: cannot support > %d services\n",
1841 /* Set configuration options for all of a service's security objects */
1844 rx_SetSecurityConfiguration(struct rx_service *service,
1845 rx_securityConfigVariables type,
1849 for (i = 0; i<service->nSecurityObjects; i++) {
1850 if (service->securityObjects[i]) {
1851 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1859 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1860 struct rx_securityClass **securityObjects, int nSecurityObjects,
1861 afs_int32(*serviceProc) (struct rx_call * acall))
1863 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1866 /* Generic request processing loop. This routine should be called
1867 * by the implementation dependent rx_ServerProc. If socketp is
1868 * non-null, it will be set to the file descriptor that this thread
1869 * is now listening on. If socketp is null, this routine will never
1872 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1874 struct rx_call *call;
1876 struct rx_service *tservice = NULL;
1883 call = rx_GetCall(threadID, tservice, socketp);
1884 if (socketp && *socketp != OSI_NULLSOCKET) {
1885 /* We are now a listener thread */
1891 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1892 #ifdef RX_ENABLE_LOCKS
1894 #endif /* RX_ENABLE_LOCKS */
1895 afs_termState = AFSOP_STOP_AFS;
1896 afs_osi_Wakeup(&afs_termState);
1897 #ifdef RX_ENABLE_LOCKS
1899 #endif /* RX_ENABLE_LOCKS */
1904 /* if server is restarting( typically smooth shutdown) then do not
1905 * allow any new calls.
1908 if (rx_tranquil && (call != NULL)) {
1912 MUTEX_ENTER(&call->lock);
1914 rxi_CallError(call, RX_RESTARTING);
1915 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1917 MUTEX_EXIT(&call->lock);
1922 tservice = call->conn->service;
1924 if (tservice->beforeProc)
1925 (*tservice->beforeProc) (call);
1927 code = tservice->executeRequestProc(call);
1929 if (tservice->afterProc)
1930 (*tservice->afterProc) (call, code);
1932 rx_EndCall(call, code);
1934 if (tservice->postProc)
1935 (*tservice->postProc) (code);
1937 if (rx_stats_active) {
1938 MUTEX_ENTER(&rx_stats_mutex);
1940 MUTEX_EXIT(&rx_stats_mutex);
1947 rx_WakeupServerProcs(void)
1949 struct rx_serverQueueEntry *np, *tqp;
1950 struct opr_queue *cursor;
1954 MUTEX_ENTER(&rx_serverPool_lock);
1956 #ifdef RX_ENABLE_LOCKS
1957 if (rx_waitForPacket)
1958 CV_BROADCAST(&rx_waitForPacket->cv);
1959 #else /* RX_ENABLE_LOCKS */
1960 if (rx_waitForPacket)
1961 osi_rxWakeup(rx_waitForPacket);
1962 #endif /* RX_ENABLE_LOCKS */
1963 MUTEX_ENTER(&freeSQEList_lock);
1964 for (np = rx_FreeSQEList; np; np = tqp) {
1965 tqp = *(struct rx_serverQueueEntry **)np;
1966 #ifdef RX_ENABLE_LOCKS
1967 CV_BROADCAST(&np->cv);
1968 #else /* RX_ENABLE_LOCKS */
1970 #endif /* RX_ENABLE_LOCKS */
1972 MUTEX_EXIT(&freeSQEList_lock);
1973 for (opr_queue_Scan(&rx_idleServerQueue, cursor)) {
1974 np = opr_queue_Entry(cursor, struct rx_serverQueueEntry, entry);
1975 #ifdef RX_ENABLE_LOCKS
1976 CV_BROADCAST(&np->cv);
1977 #else /* RX_ENABLE_LOCKS */
1979 #endif /* RX_ENABLE_LOCKS */
1981 MUTEX_EXIT(&rx_serverPool_lock);
1986 * One thing that seems to happen is that all the server threads get
1987 * tied up on some empty or slow call, and then a whole bunch of calls
1988 * arrive at once, using up the packet pool, so now there are more
1989 * empty calls. The most critical resources here are server threads
1990 * and the free packet pool. The "doreclaim" code seems to help in
1991 * general. I think that eventually we arrive in this state: there
1992 * are lots of pending calls which do have all their packets present,
1993 * so they won't be reclaimed, are multi-packet calls, so they won't
1994 * be scheduled until later, and thus are tying up most of the free
1995 * packet pool for a very long time.
1997 * 1. schedule multi-packet calls if all the packets are present.
1998 * Probably CPU-bound operation, useful to return packets to pool.
1999 * Do what if there is a full window, but the last packet isn't here?
2000 * 3. preserve one thread which *only* runs "best" calls, otherwise
2001 * it sleeps and waits for that type of call.
2002 * 4. Don't necessarily reserve a whole window for each thread. In fact,
2003 * the current dataquota business is badly broken. The quota isn't adjusted
2004 * to reflect how many packets are presently queued for a running call.
2005 * So, when we schedule a queued call with a full window of packets queued
2006 * up for it, that *should* free up a window full of packets for other 2d-class
2007 * calls to be able to use from the packet pool. But it doesn't.
2009 * NB. Most of the time, this code doesn't run -- since idle server threads
2010 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
2011 * as a new call arrives.
2013 /* Sleep until a call arrives. Returns a pointer to the call, ready
2014 * for an rx_Read. */
2015 #ifdef RX_ENABLE_LOCKS
2017 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2019 struct rx_serverQueueEntry *sq;
2020 struct rx_call *call = (struct rx_call *)0;
2021 struct rx_service *service = NULL;
2023 MUTEX_ENTER(&freeSQEList_lock);
2025 if ((sq = rx_FreeSQEList)) {
2026 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2027 MUTEX_EXIT(&freeSQEList_lock);
2028 } else { /* otherwise allocate a new one and return that */
2029 MUTEX_EXIT(&freeSQEList_lock);
2030 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2031 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2032 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2035 MUTEX_ENTER(&rx_serverPool_lock);
2036 if (cur_service != NULL) {
2037 ReturnToServerPool(cur_service);
2040 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2041 struct rx_call *tcall, *choice2 = NULL;
2042 struct opr_queue *cursor;
2044 /* Scan for eligible incoming calls. A call is not eligible
2045 * if the maximum number of calls for its service type are
2046 * already executing */
2047 /* One thread will process calls FCFS (to prevent starvation),
2048 * while the other threads may run ahead looking for calls which
2049 * have all their input data available immediately. This helps
2050 * keep threads from blocking, waiting for data from the client. */
2051 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2052 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2054 service = tcall->conn->service;
2055 if (!QuotaOK(service)) {
2058 MUTEX_ENTER(&rx_pthread_mutex);
2059 if (tno == rxi_fcfs_thread_num
2060 || opr_queue_IsEnd(&rx_incomingCallQueue, cursor)) {
2061 MUTEX_EXIT(&rx_pthread_mutex);
2062 /* If we're the fcfs thread , then we'll just use
2063 * this call. If we haven't been able to find an optimal
2064 * choice, and we're at the end of the list, then use a
2065 * 2d choice if one has been identified. Otherwise... */
2066 call = (choice2 ? choice2 : tcall);
2067 service = call->conn->service;
2069 MUTEX_EXIT(&rx_pthread_mutex);
2070 if (!opr_queue_IsEmpty(&tcall->rq)) {
2071 struct rx_packet *rp;
2072 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2074 if (rp->header.seq == 1) {
2076 || (rp->header.flags & RX_LAST_PACKET)) {
2078 } else if (rxi_2dchoice && !choice2
2079 && !(tcall->flags & RX_CALL_CLEARED)
2080 && (tcall->rprev > rxi_HardAckRate)) {
2090 ReturnToServerPool(service);
2096 opr_queue_Remove(&call->entry);
2097 MUTEX_EXIT(&rx_serverPool_lock);
2098 MUTEX_ENTER(&call->lock);
2100 if (call->flags & RX_CALL_WAIT_PROC) {
2101 call->flags &= ~RX_CALL_WAIT_PROC;
2102 rx_atomic_dec(&rx_nWaiting);
2105 if (call->state != RX_STATE_PRECALL || call->error) {
2106 MUTEX_EXIT(&call->lock);
2107 MUTEX_ENTER(&rx_serverPool_lock);
2108 ReturnToServerPool(service);
2113 if (opr_queue_IsEmpty(&call->rq)
2114 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1)
2115 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2117 CLEAR_CALL_QUEUE_LOCK(call);
2120 /* If there are no eligible incoming calls, add this process
2121 * to the idle server queue, to wait for one */
2125 *socketp = OSI_NULLSOCKET;
2127 sq->socketp = socketp;
2128 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2129 #ifndef AFS_AIX41_ENV
2130 rx_waitForPacket = sq;
2132 rx_waitingForPacket = sq;
2133 #endif /* AFS_AIX41_ENV */
2135 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2137 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2138 MUTEX_EXIT(&rx_serverPool_lock);
2139 return (struct rx_call *)0;
2142 } while (!(call = sq->newcall)
2143 && !(socketp && *socketp != OSI_NULLSOCKET));
2144 MUTEX_EXIT(&rx_serverPool_lock);
2146 MUTEX_ENTER(&call->lock);
2152 MUTEX_ENTER(&freeSQEList_lock);
2153 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2154 rx_FreeSQEList = sq;
2155 MUTEX_EXIT(&freeSQEList_lock);
2158 clock_GetTime(&call->startTime);
2159 call->state = RX_STATE_ACTIVE;
2160 call->app.mode = RX_MODE_RECEIVING;
2161 #ifdef RX_KERNEL_TRACE
2162 if (ICL_SETACTIVE(afs_iclSetp)) {
2163 int glockOwner = ISAFS_GLOCK();
2166 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2167 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2174 rxi_calltrace(RX_CALL_START, call);
2175 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2176 call->conn->service->servicePort, call->conn->service->serviceId,
2179 MUTEX_EXIT(&call->lock);
2180 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2182 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2187 #else /* RX_ENABLE_LOCKS */
2189 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2191 struct rx_serverQueueEntry *sq;
2192 struct rx_call *call = (struct rx_call *)0, *choice2;
2193 struct rx_service *service = NULL;
2197 MUTEX_ENTER(&freeSQEList_lock);
2199 if ((sq = rx_FreeSQEList)) {
2200 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2201 MUTEX_EXIT(&freeSQEList_lock);
2202 } else { /* otherwise allocate a new one and return that */
2203 MUTEX_EXIT(&freeSQEList_lock);
2204 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2205 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2206 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2208 MUTEX_ENTER(&sq->lock);
2210 if (cur_service != NULL) {
2211 cur_service->nRequestsRunning--;
2212 MUTEX_ENTER(&rx_quota_mutex);
2213 if (cur_service->nRequestsRunning < cur_service->minProcs)
2216 MUTEX_EXIT(&rx_quota_mutex);
2218 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2219 struct rx_call *tcall;
2220 struct opr_queue *cursor;
2221 /* Scan for eligible incoming calls. A call is not eligible
2222 * if the maximum number of calls for its service type are
2223 * already executing */
2224 /* One thread will process calls FCFS (to prevent starvation),
2225 * while the other threads may run ahead looking for calls which
2226 * have all their input data available immediately. This helps
2227 * keep threads from blocking, waiting for data from the client. */
2228 choice2 = (struct rx_call *)0;
2229 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2230 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2231 service = tcall->conn->service;
2232 if (QuotaOK(service)) {
2233 MUTEX_ENTER(&rx_pthread_mutex);
2234 /* XXX - If tcall->entry.next is NULL, then we're no longer
2235 * on a queue at all. This shouldn't happen. */
2236 if (tno == rxi_fcfs_thread_num || !tcall->entry.next) {
2237 MUTEX_EXIT(&rx_pthread_mutex);
2238 /* If we're the fcfs thread, then we'll just use
2239 * this call. If we haven't been able to find an optimal
2240 * choice, and we're at the end of the list, then use a
2241 * 2d choice if one has been identified. Otherwise... */
2242 call = (choice2 ? choice2 : tcall);
2243 service = call->conn->service;
2245 MUTEX_EXIT(&rx_pthread_mutex);
2246 if (!opr_queue_IsEmpty(&tcall->rq)) {
2247 struct rx_packet *rp;
2248 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2250 if (rp->header.seq == 1
2252 || (rp->header.flags & RX_LAST_PACKET))) {
2254 } else if (rxi_2dchoice && !choice2
2255 && !(tcall->flags & RX_CALL_CLEARED)
2256 && (tcall->rprev > rxi_HardAckRate)) {
2269 opr_queue_Remove(&call->entry);
2270 /* we can't schedule a call if there's no data!!! */
2271 /* send an ack if there's no data, if we're missing the
2272 * first packet, or we're missing something between first
2273 * and last -- there's a "hole" in the incoming data. */
2274 if (opr_queue_IsEmpty(&call->rq)
2275 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1
2276 || call->rprev != opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq)
2277 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2279 call->flags &= (~RX_CALL_WAIT_PROC);
2280 service->nRequestsRunning++;
2281 /* just started call in minProcs pool, need fewer to maintain
2283 MUTEX_ENTER(&rx_quota_mutex);
2284 if (service->nRequestsRunning <= service->minProcs)
2287 MUTEX_EXIT(&rx_quota_mutex);
2288 rx_atomic_dec(&rx_nWaiting);
2289 /* MUTEX_EXIT(&call->lock); */
2291 /* If there are no eligible incoming calls, add this process
2292 * to the idle server queue, to wait for one */
2295 *socketp = OSI_NULLSOCKET;
2297 sq->socketp = socketp;
2298 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2302 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2304 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2305 return (struct rx_call *)0;
2308 } while (!(call = sq->newcall)
2309 && !(socketp && *socketp != OSI_NULLSOCKET));
2311 MUTEX_EXIT(&sq->lock);
2313 MUTEX_ENTER(&freeSQEList_lock);
2314 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2315 rx_FreeSQEList = sq;
2316 MUTEX_EXIT(&freeSQEList_lock);
2319 clock_GetTime(&call->startTime);
2320 call->state = RX_STATE_ACTIVE;
2321 call->app.mode = RX_MODE_RECEIVING;
2322 #ifdef RX_KERNEL_TRACE
2323 if (ICL_SETACTIVE(afs_iclSetp)) {
2324 int glockOwner = ISAFS_GLOCK();
2327 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2328 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2335 rxi_calltrace(RX_CALL_START, call);
2336 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2337 call->conn->service->servicePort, call->conn->service->serviceId,
2340 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2347 #endif /* RX_ENABLE_LOCKS */
2351 /* Establish a procedure to be called when a packet arrives for a
2352 * call. This routine will be called at most once after each call,
2353 * and will also be called if there is an error condition on the or
2354 * the call is complete. Used by multi rx to build a selection
2355 * function which determines which of several calls is likely to be a
2356 * good one to read from.
2357 * NOTE: the way this is currently implemented it is probably only a
2358 * good idea to (1) use it immediately after a newcall (clients only)
2359 * and (2) only use it once. Other uses currently void your warranty
2362 rx_SetArrivalProc(struct rx_call *call,
2363 void (*proc) (struct rx_call * call,
2366 void * handle, int arg)
2368 call->arrivalProc = proc;
2369 call->arrivalProcHandle = handle;
2370 call->arrivalProcArg = arg;
2373 /* Call is finished (possibly prematurely). Return rc to the peer, if
2374 * appropriate, and return the final error code from the conversation
2378 rx_EndCall(struct rx_call *call, afs_int32 rc)
2380 struct rx_connection *conn = call->conn;
2384 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2385 call, rc, call->error, call->abortCode));
2388 MUTEX_ENTER(&call->lock);
2390 if (rc == 0 && call->error == 0) {
2391 call->abortCode = 0;
2392 call->abortCount = 0;
2395 call->arrivalProc = (void (*)())0;
2396 if (rc && call->error == 0) {
2397 rxi_CallError(call, rc);
2398 call->app.mode = RX_MODE_ERROR;
2399 /* Send an abort message to the peer if this error code has
2400 * only just been set. If it was set previously, assume the
2401 * peer has already been sent the error code or will request it
2403 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2405 if (conn->type == RX_SERVER_CONNECTION) {
2406 /* Make sure reply or at least dummy reply is sent */
2407 if (call->app.mode == RX_MODE_RECEIVING) {
2408 MUTEX_EXIT(&call->lock);
2409 rxi_WriteProc(call, 0, 0);
2410 MUTEX_ENTER(&call->lock);
2412 if (call->app.mode == RX_MODE_SENDING) {
2413 MUTEX_EXIT(&call->lock);
2414 rxi_FlushWrite(call);
2415 MUTEX_ENTER(&call->lock);
2417 rxi_calltrace(RX_CALL_END, call);
2418 /* Call goes to hold state until reply packets are acknowledged */
2419 if (call->tfirst + call->nSoftAcked < call->tnext) {
2420 call->state = RX_STATE_HOLD;
2422 call->state = RX_STATE_DALLY;
2423 rxi_ClearTransmitQueue(call, 0);
2424 rxi_rto_cancel(call);
2425 rxi_CancelKeepAliveEvent(call);
2427 } else { /* Client connection */
2429 /* Make sure server receives input packets, in the case where
2430 * no reply arguments are expected */
2432 if ((call->app.mode == RX_MODE_SENDING)
2433 || (call->app.mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2434 MUTEX_EXIT(&call->lock);
2435 (void)rxi_ReadProc(call, &dummy, 1);
2436 MUTEX_ENTER(&call->lock);
2439 /* If we had an outstanding delayed ack, be nice to the server
2440 * and force-send it now.
2442 if (call->delayedAckEvent) {
2443 rxi_CancelDelayedAckEvent(call);
2444 rxi_SendDelayedAck(NULL, call, NULL, 0);
2447 /* We need to release the call lock since it's lower than the
2448 * conn_call_lock and we don't want to hold the conn_call_lock
2449 * over the rx_ReadProc call. The conn_call_lock needs to be held
2450 * here for the case where rx_NewCall is perusing the calls on
2451 * the connection structure. We don't want to signal until
2452 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2453 * have checked this call, found it active and by the time it
2454 * goes to sleep, will have missed the signal.
2456 MUTEX_EXIT(&call->lock);
2457 MUTEX_ENTER(&conn->conn_call_lock);
2458 MUTEX_ENTER(&call->lock);
2460 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2461 conn->lastBusy[call->channel] = 0;
2464 MUTEX_ENTER(&conn->conn_data_lock);
2465 conn->flags |= RX_CONN_BUSY;
2466 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2467 MUTEX_EXIT(&conn->conn_data_lock);
2468 #ifdef RX_ENABLE_LOCKS
2469 CV_BROADCAST(&conn->conn_call_cv);
2474 #ifdef RX_ENABLE_LOCKS
2476 MUTEX_EXIT(&conn->conn_data_lock);
2478 #endif /* RX_ENABLE_LOCKS */
2479 call->state = RX_STATE_DALLY;
2481 error = call->error;
2483 /* currentPacket, nLeft, and NFree must be zeroed here, because
2484 * ResetCall cannot: ResetCall may be called at splnet(), in the
2485 * kernel version, and may interrupt the macros rx_Read or
2486 * rx_Write, which run at normal priority for efficiency. */
2487 if (call->app.currentPacket) {
2488 #ifdef RX_TRACK_PACKETS
2489 call->app.currentPacket->flags &= ~RX_PKTFLAG_CP;
2491 rxi_FreePacket(call->app.currentPacket);
2492 call->app.currentPacket = (struct rx_packet *)0;
2495 call->app.nLeft = call->app.nFree = call->app.curlen = 0;
2497 /* Free any packets from the last call to ReadvProc/WritevProc */
2498 #ifdef RXDEBUG_PACKET
2500 #endif /* RXDEBUG_PACKET */
2501 rxi_FreePackets(0, &call->app.iovq);
2502 MUTEX_EXIT(&call->lock);
2504 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2505 if (conn->type == RX_CLIENT_CONNECTION) {
2506 MUTEX_ENTER(&conn->conn_data_lock);
2507 conn->flags &= ~RX_CONN_BUSY;
2508 MUTEX_EXIT(&conn->conn_data_lock);
2509 MUTEX_EXIT(&conn->conn_call_lock);
2513 * Map errors to the local host's errno.h format.
2515 error = ntoh_syserr_conv(error);
2519 #if !defined(KERNEL)
2521 /* Call this routine when shutting down a server or client (especially
2522 * clients). This will allow Rx to gracefully garbage collect server
2523 * connections, and reduce the number of retries that a server might
2524 * make to a dead client.
2525 * This is not quite right, since some calls may still be ongoing and
2526 * we can't lock them to destroy them. */
2530 struct rx_connection **conn_ptr, **conn_end;
2533 if (rx_atomic_test_and_set_bit(&rxinit_status, 0))
2534 return; /* Already shutdown. */
2536 rxi_DeleteCachedConnections();
2537 if (rx_connHashTable) {
2538 MUTEX_ENTER(&rx_connHashTable_lock);
2539 for (conn_ptr = &rx_connHashTable[0], conn_end =
2540 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2542 struct rx_connection *conn, *next;
2543 for (conn = *conn_ptr; conn; conn = next) {
2545 if (conn->type == RX_CLIENT_CONNECTION) {
2546 MUTEX_ENTER(&rx_refcnt_mutex);
2548 MUTEX_EXIT(&rx_refcnt_mutex);
2549 #ifdef RX_ENABLE_LOCKS
2550 rxi_DestroyConnectionNoLock(conn);
2551 #else /* RX_ENABLE_LOCKS */
2552 rxi_DestroyConnection(conn);
2553 #endif /* RX_ENABLE_LOCKS */
2557 #ifdef RX_ENABLE_LOCKS
2558 while (rx_connCleanup_list) {
2559 struct rx_connection *conn;
2560 conn = rx_connCleanup_list;
2561 rx_connCleanup_list = rx_connCleanup_list->next;
2562 MUTEX_EXIT(&rx_connHashTable_lock);
2563 rxi_CleanupConnection(conn);
2564 MUTEX_ENTER(&rx_connHashTable_lock);
2566 MUTEX_EXIT(&rx_connHashTable_lock);
2567 #endif /* RX_ENABLE_LOCKS */
2572 afs_winsockCleanup();
2578 /* if we wakeup packet waiter too often, can get in loop with two
2579 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2581 rxi_PacketsUnWait(void)
2583 if (!rx_waitingForPackets) {
2587 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2588 return; /* still over quota */
2591 rx_waitingForPackets = 0;
2592 #ifdef RX_ENABLE_LOCKS
2593 CV_BROADCAST(&rx_waitingForPackets_cv);
2595 osi_rxWakeup(&rx_waitingForPackets);
2601 /* ------------------Internal interfaces------------------------- */
2603 /* Return this process's service structure for the
2604 * specified socket and service */
2605 static struct rx_service *
2606 rxi_FindService(osi_socket socket, u_short serviceId)
2608 struct rx_service **sp;
2609 for (sp = &rx_services[0]; *sp; sp++) {
2610 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2616 #ifdef RXDEBUG_PACKET
2617 #ifdef KDUMP_RX_LOCK
2618 static struct rx_call_rx_lock *rx_allCallsp = 0;
2620 static struct rx_call *rx_allCallsp = 0;
2622 #endif /* RXDEBUG_PACKET */
2624 /* Allocate a call structure, for the indicated channel of the
2625 * supplied connection. The mode and state of the call must be set by
2626 * the caller. Returns the call with mutex locked. */
2627 static struct rx_call *
2628 rxi_NewCall(struct rx_connection *conn, int channel)
2630 struct rx_call *call;
2631 #ifdef RX_ENABLE_LOCKS
2632 struct rx_call *cp; /* Call pointer temp */
2633 struct opr_queue *cursor;
2636 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2638 /* Grab an existing call structure, or allocate a new one.
2639 * Existing call structures are assumed to have been left reset by
2641 MUTEX_ENTER(&rx_freeCallQueue_lock);
2643 #ifdef RX_ENABLE_LOCKS
2645 * EXCEPT that the TQ might not yet be cleared out.
2646 * Skip over those with in-use TQs.
2649 for (opr_queue_Scan(&rx_freeCallQueue, cursor)) {
2650 cp = opr_queue_Entry(cursor, struct rx_call, entry);
2651 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2657 #else /* RX_ENABLE_LOCKS */
2658 if (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
2659 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
2660 #endif /* RX_ENABLE_LOCKS */
2661 opr_queue_Remove(&call->entry);
2662 if (rx_stats_active)
2663 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2664 MUTEX_EXIT(&rx_freeCallQueue_lock);
2665 MUTEX_ENTER(&call->lock);
2666 CLEAR_CALL_QUEUE_LOCK(call);
2667 #ifdef RX_ENABLE_LOCKS
2668 /* Now, if TQ wasn't cleared earlier, do it now. */
2669 rxi_WaitforTQBusy(call);
2670 if (call->flags & RX_CALL_TQ_CLEARME) {
2671 rxi_ClearTransmitQueue(call, 1);
2672 /*queue_Init(&call->tq);*/
2674 #endif /* RX_ENABLE_LOCKS */
2675 /* Bind the call to its connection structure */
2677 rxi_ResetCall(call, 1);
2680 call = rxi_Alloc(sizeof(struct rx_call));
2681 #ifdef RXDEBUG_PACKET
2682 call->allNextp = rx_allCallsp;
2683 rx_allCallsp = call;
2685 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2686 #else /* RXDEBUG_PACKET */
2687 rx_atomic_inc(&rx_stats.nCallStructs);
2688 #endif /* RXDEBUG_PACKET */
2690 MUTEX_EXIT(&rx_freeCallQueue_lock);
2691 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2692 MUTEX_ENTER(&call->lock);
2693 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2694 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2695 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2697 /* Initialize once-only items */
2698 opr_queue_Init(&call->tq);
2699 opr_queue_Init(&call->rq);
2700 opr_queue_Init(&call->app.iovq);
2701 #ifdef RXDEBUG_PACKET
2702 call->rqc = call->tqc = call->iovqc = 0;
2703 #endif /* RXDEBUG_PACKET */
2704 /* Bind the call to its connection structure (prereq for reset) */
2706 rxi_ResetCall(call, 1);
2708 call->channel = channel;
2709 call->callNumber = &conn->callNumber[channel];
2710 call->rwind = conn->rwind[channel];
2711 call->twind = conn->twind[channel];
2712 /* Note that the next expected call number is retained (in
2713 * conn->callNumber[i]), even if we reallocate the call structure
2715 conn->call[channel] = call;
2716 /* if the channel's never been used (== 0), we should start at 1, otherwise
2717 * the call number is valid from the last time this channel was used */
2718 if (*call->callNumber == 0)
2719 *call->callNumber = 1;
2724 /* A call has been inactive long enough that so we can throw away
2725 * state, including the call structure, which is placed on the call
2728 * call->lock amd rx_refcnt_mutex are held upon entry.
2729 * haveCTLock is set when called from rxi_ReapConnections.
2731 * return 1 if the call is freed, 0 if not.
2734 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2736 int channel = call->channel;
2737 struct rx_connection *conn = call->conn;
2738 u_char state = call->state;
2741 * We are setting the state to RX_STATE_RESET to
2742 * ensure that no one else will attempt to use this
2743 * call once we drop the refcnt lock. We must drop
2744 * the refcnt lock before calling rxi_ResetCall
2745 * because it cannot be held across acquiring the
2746 * freepktQ lock. NewCall does the same.
2748 call->state = RX_STATE_RESET;
2749 MUTEX_EXIT(&rx_refcnt_mutex);
2750 rxi_ResetCall(call, 0);
2752 if (MUTEX_TRYENTER(&conn->conn_call_lock))
2754 if (state == RX_STATE_DALLY || state == RX_STATE_HOLD)
2755 (*call->callNumber)++;
2757 if (call->conn->call[channel] == call)
2758 call->conn->call[channel] = 0;
2759 MUTEX_EXIT(&conn->conn_call_lock);
2762 * We couldn't obtain the conn_call_lock so we can't
2763 * disconnect the call from the connection. Set the
2764 * call state to dally so that the call can be reused.
2766 MUTEX_ENTER(&rx_refcnt_mutex);
2767 call->state = RX_STATE_DALLY;
2771 MUTEX_ENTER(&rx_freeCallQueue_lock);
2772 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2773 #ifdef RX_ENABLE_LOCKS
2774 /* A call may be free even though its transmit queue is still in use.
2775 * Since we search the call list from head to tail, put busy calls at
2776 * the head of the list, and idle calls at the tail.
2778 if (call->flags & RX_CALL_TQ_BUSY)
2779 opr_queue_Prepend(&rx_freeCallQueue, &call->entry);
2781 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2782 #else /* RX_ENABLE_LOCKS */
2783 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2784 #endif /* RX_ENABLE_LOCKS */
2785 if (rx_stats_active)
2786 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2787 MUTEX_EXIT(&rx_freeCallQueue_lock);
2789 /* Destroy the connection if it was previously slated for
2790 * destruction, i.e. the Rx client code previously called
2791 * rx_DestroyConnection (client connections), or
2792 * rxi_ReapConnections called the same routine (server
2793 * connections). Only do this, however, if there are no
2794 * outstanding calls. Note that for fine grain locking, there appears
2795 * to be a deadlock in that rxi_FreeCall has a call locked and
2796 * DestroyConnectionNoLock locks each call in the conn. But note a
2797 * few lines up where we have removed this call from the conn.
2798 * If someone else destroys a connection, they either have no
2799 * call lock held or are going through this section of code.
2801 MUTEX_ENTER(&conn->conn_data_lock);
2802 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2803 MUTEX_ENTER(&rx_refcnt_mutex);
2805 MUTEX_EXIT(&rx_refcnt_mutex);
2806 MUTEX_EXIT(&conn->conn_data_lock);
2807 #ifdef RX_ENABLE_LOCKS
2809 rxi_DestroyConnectionNoLock(conn);
2811 rxi_DestroyConnection(conn);
2812 #else /* RX_ENABLE_LOCKS */
2813 rxi_DestroyConnection(conn);
2814 #endif /* RX_ENABLE_LOCKS */
2816 MUTEX_EXIT(&conn->conn_data_lock);
2818 MUTEX_ENTER(&rx_refcnt_mutex);
2822 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2823 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2826 rxi_Alloc(size_t size)
2830 if (rx_stats_active) {
2831 rx_atomic_add(&rxi_Allocsize, (int) size);
2832 rx_atomic_inc(&rxi_Alloccnt);
2836 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2837 afs_osi_Alloc_NoSleep(size);
2842 osi_Panic("rxi_Alloc error");
2848 rxi_Free(void *addr, size_t size)
2850 if (rx_stats_active) {
2851 rx_atomic_sub(&rxi_Allocsize, (int) size);
2852 rx_atomic_dec(&rxi_Alloccnt);
2854 osi_Free(addr, size);
2858 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2860 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2861 struct rx_peer *next = NULL;
2865 MUTEX_ENTER(&rx_peerHashTable_lock);
2867 peer_ptr = &rx_peerHashTable[0];
2868 peer_end = &rx_peerHashTable[rx_hashTableSize];
2871 for ( ; peer_ptr < peer_end; peer_ptr++) {
2874 for ( ; peer; peer = next) {
2876 if (host == peer->host)
2881 hashIndex = PEER_HASH(host, port);
2882 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2883 if ((peer->host == host) && (peer->port == port))
2888 MUTEX_ENTER(&rx_peerHashTable_lock);
2893 MUTEX_EXIT(&rx_peerHashTable_lock);
2895 MUTEX_ENTER(&peer->peer_lock);
2896 /* We don't handle dropping below min, so don't */
2897 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2898 peer->ifMTU=MIN(mtu, peer->ifMTU);
2899 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2900 /* if we tweaked this down, need to tune our peer MTU too */
2901 peer->MTU = MIN(peer->MTU, peer->natMTU);
2902 /* if we discovered a sub-1500 mtu, degrade */
2903 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2904 peer->maxDgramPackets = 1;
2905 /* We no longer have valid peer packet information */
2906 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2907 peer->maxPacketSize = 0;
2908 MUTEX_EXIT(&peer->peer_lock);
2910 MUTEX_ENTER(&rx_peerHashTable_lock);
2912 if (host && !port) {
2914 /* pick up where we left off */
2918 MUTEX_EXIT(&rx_peerHashTable_lock);
2921 #ifdef AFS_RXERRQ_ENV
2923 rxi_SetPeerDead(struct sock_extended_err *err, afs_uint32 host, afs_uint16 port)
2925 int hashIndex = PEER_HASH(host, port);
2926 struct rx_peer *peer;
2928 MUTEX_ENTER(&rx_peerHashTable_lock);
2930 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2931 if (peer->host == host && peer->port == port) {
2937 MUTEX_EXIT(&rx_peerHashTable_lock);
2940 rx_atomic_inc(&peer->neterrs);
2941 MUTEX_ENTER(&peer->peer_lock);
2942 peer->last_err_origin = RX_NETWORK_ERROR_ORIGIN_ICMP;
2943 peer->last_err_type = err->ee_type;
2944 peer->last_err_code = err->ee_code;
2945 MUTEX_EXIT(&peer->peer_lock);
2947 MUTEX_ENTER(&rx_peerHashTable_lock);
2949 MUTEX_EXIT(&rx_peerHashTable_lock);
2954 rxi_ProcessNetError(struct sock_extended_err *err, afs_uint32 addr, afs_uint16 port)
2956 # ifdef AFS_ADAPT_PMTU
2957 if (err->ee_errno == EMSGSIZE && err->ee_info >= 68) {
2958 rxi_SetPeerMtu(NULL, addr, port, err->ee_info - RX_IPUDP_SIZE);
2962 if (err->ee_origin == SO_EE_ORIGIN_ICMP && err->ee_type == ICMP_DEST_UNREACH) {
2963 switch (err->ee_code) {
2964 case ICMP_NET_UNREACH:
2965 case ICMP_HOST_UNREACH:
2966 case ICMP_PORT_UNREACH:
2969 rxi_SetPeerDead(err, addr, port);
2976 rxi_TranslateICMP(int type, int code)
2979 case ICMP_DEST_UNREACH:
2981 case ICMP_NET_UNREACH:
2982 return "Destination Net Unreachable";
2983 case ICMP_HOST_UNREACH:
2984 return "Destination Host Unreachable";
2985 case ICMP_PROT_UNREACH:
2986 return "Destination Protocol Unreachable";
2987 case ICMP_PORT_UNREACH:
2988 return "Destination Port Unreachable";
2990 return "Destination Net Prohibited";
2992 return "Destination Host Prohibited";
2998 #endif /* AFS_RXERRQ_ENV */
3001 * Get the last network error for a connection
3003 * A "network error" here means an error retrieved from ICMP, or some other
3004 * mechanism outside of Rx that informs us of errors in network reachability.
3006 * If a peer associated with the given Rx connection has received a network
3007 * error recently, this function allows the caller to know what error
3008 * specifically occurred. This can be useful to know, since e.g. ICMP errors
3009 * can cause calls to that peer to be quickly aborted. So, this function can
3010 * help see why a call was aborted due to network errors.
3012 * If we have received traffic from a peer since the last network error, we
3013 * treat that peer as if we had not received an network error for it.
3015 * @param[in] conn The Rx connection to examine
3016 * @param[out] err_origin The origin of the last network error (e.g. ICMP);
3017 * one of the RX_NETWORK_ERROR_ORIGIN_* constants
3018 * @param[out] err_type The type of the last error
3019 * @param[out] err_code The code of the last error
3020 * @param[out] msg Human-readable error message, if applicable; NULL otherwise
3022 * @return If we have an error
3023 * @retval -1 No error to get; 'out' params are undefined
3024 * @retval 0 We have an error; 'out' params contain the last error
3027 rx_GetNetworkError(struct rx_connection *conn, int *err_origin, int *err_type,
3028 int *err_code, const char **msg)
3030 #ifdef AFS_RXERRQ_ENV
3031 struct rx_peer *peer = conn->peer;
3032 if (rx_atomic_read(&peer->neterrs)) {
3033 MUTEX_ENTER(&peer->peer_lock);
3034 *err_origin = peer->last_err_origin;
3035 *err_type = peer->last_err_type;
3036 *err_code = peer->last_err_code;
3037 MUTEX_EXIT(&peer->peer_lock);
3040 if (*err_origin == RX_NETWORK_ERROR_ORIGIN_ICMP) {
3041 *msg = rxi_TranslateICMP(*err_type, *err_code);
3050 /* Find the peer process represented by the supplied (host,port)
3051 * combination. If there is no appropriate active peer structure, a
3052 * new one will be allocated and initialized
3055 rxi_FindPeer(afs_uint32 host, u_short port, int create)
3059 hashIndex = PEER_HASH(host, port);
3060 MUTEX_ENTER(&rx_peerHashTable_lock);
3061 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
3062 if ((pp->host == host) && (pp->port == port))
3067 pp = rxi_AllocPeer(); /* This bzero's *pp */
3068 pp->host = host; /* set here or in InitPeerParams is zero */
3070 #ifdef AFS_RXERRQ_ENV
3071 rx_atomic_set(&pp->neterrs, 0);
3073 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
3074 opr_queue_Init(&pp->rpcStats);
3075 pp->next = rx_peerHashTable[hashIndex];
3076 rx_peerHashTable[hashIndex] = pp;
3077 rxi_InitPeerParams(pp);
3078 if (rx_stats_active)
3079 rx_atomic_inc(&rx_stats.nPeerStructs);
3085 MUTEX_EXIT(&rx_peerHashTable_lock);
3090 /* Find the connection at (host, port) started at epoch, and with the
3091 * given connection id. Creates the server connection if necessary.
3092 * The type specifies whether a client connection or a server
3093 * connection is desired. In both cases, (host, port) specify the
3094 * peer's (host, pair) pair. Client connections are not made
3095 * automatically by this routine. The parameter socket gives the
3096 * socket descriptor on which the packet was received. This is used,
3097 * in the case of server connections, to check that *new* connections
3098 * come via a valid (port, serviceId). Finally, the securityIndex
3099 * parameter must match the existing index for the connection. If a
3100 * server connection is created, it will be created using the supplied
3101 * index, if the index is valid for this service */
3102 static struct rx_connection *
3103 rxi_FindConnection(osi_socket socket, afs_uint32 host,
3104 u_short port, u_short serviceId, afs_uint32 cid,
3105 afs_uint32 epoch, int type, u_int securityIndex,
3106 int *unknownService)
3108 int hashindex, flag, i;
3109 struct rx_connection *conn;
3110 *unknownService = 0;
3111 hashindex = CONN_HASH(host, port, cid, epoch, type);
3112 MUTEX_ENTER(&rx_connHashTable_lock);
3113 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
3114 rx_connHashTable[hashindex],
3117 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
3118 && (epoch == conn->epoch)) {
3119 struct rx_peer *pp = conn->peer;
3120 if (securityIndex != conn->securityIndex) {
3121 /* this isn't supposed to happen, but someone could forge a packet
3122 * like this, and there seems to be some CM bug that makes this
3123 * happen from time to time -- in which case, the fileserver
3125 MUTEX_EXIT(&rx_connHashTable_lock);
3126 return (struct rx_connection *)0;
3128 if (pp->host == host && pp->port == port)
3130 if (type == RX_CLIENT_CONNECTION && pp->port == port)
3132 /* So what happens when it's a callback connection? */
3133 if ( /*type == RX_CLIENT_CONNECTION && */
3134 (conn->epoch & 0x80000000))
3138 /* the connection rxLastConn that was used the last time is not the
3139 ** one we are looking for now. Hence, start searching in the hash */
3141 conn = rx_connHashTable[hashindex];
3146 struct rx_service *service;
3147 if (type == RX_CLIENT_CONNECTION) {
3148 MUTEX_EXIT(&rx_connHashTable_lock);
3149 return (struct rx_connection *)0;
3151 service = rxi_FindService(socket, serviceId);
3152 if (!service || (securityIndex >= service->nSecurityObjects)
3153 || (service->securityObjects[securityIndex] == 0)) {
3154 MUTEX_EXIT(&rx_connHashTable_lock);
3155 *unknownService = 1;
3156 return (struct rx_connection *)0;
3158 conn = rxi_AllocConnection(); /* This bzero's the connection */
3159 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
3160 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
3161 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
3162 conn->next = rx_connHashTable[hashindex];
3163 rx_connHashTable[hashindex] = conn;
3164 conn->peer = rxi_FindPeer(host, port, 1);
3165 conn->type = RX_SERVER_CONNECTION;
3166 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
3167 conn->epoch = epoch;
3168 conn->cid = cid & RX_CIDMASK;
3169 conn->ackRate = RX_FAST_ACK_RATE;
3170 conn->service = service;
3171 conn->serviceId = serviceId;
3172 conn->securityIndex = securityIndex;
3173 conn->securityObject = service->securityObjects[securityIndex];
3174 conn->nSpecific = 0;
3175 conn->specific = NULL;
3176 rx_SetConnDeadTime(conn, service->connDeadTime);
3177 conn->idleDeadTime = service->idleDeadTime;
3178 conn->idleDeadDetection = service->idleDeadErr ? 1 : 0;
3179 for (i = 0; i < RX_MAXCALLS; i++) {
3180 conn->twind[i] = rx_initSendWindow;
3181 conn->rwind[i] = rx_initReceiveWindow;
3183 /* Notify security object of the new connection */
3184 RXS_NewConnection(conn->securityObject, conn);
3185 /* XXXX Connection timeout? */
3186 if (service->newConnProc)
3187 (*service->newConnProc) (conn);
3188 if (rx_stats_active)
3189 rx_atomic_inc(&rx_stats.nServerConns);
3192 MUTEX_ENTER(&rx_refcnt_mutex);
3194 MUTEX_EXIT(&rx_refcnt_mutex);
3196 rxLastConn = conn; /* store this connection as the last conn used */
3197 MUTEX_EXIT(&rx_connHashTable_lock);
3202 * Timeout a call on a busy call channel if appropriate.
3204 * @param[in] call The busy call.
3206 * @pre 'call' is marked as busy (namely,
3207 * call->conn->lastBusy[call->channel] != 0)
3209 * @pre call->lock is held
3210 * @pre rxi_busyChannelError is nonzero
3212 * @note call->lock is dropped and reacquired
3215 rxi_CheckBusy(struct rx_call *call)
3217 struct rx_connection *conn = call->conn;
3218 int channel = call->channel;
3219 int freechannel = 0;
3222 MUTEX_EXIT(&call->lock);
3224 MUTEX_ENTER(&conn->conn_call_lock);
3226 /* Are there any other call slots on this conn that we should try? Look for
3227 * slots that are empty and are either non-busy, or were marked as busy
3228 * longer than conn->secondsUntilDead seconds before this call started. */
3230 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3232 /* only look at channels that aren't us */
3236 if (conn->lastBusy[i]) {
3237 /* if this channel looked busy too recently, don't look at it */
3238 if (conn->lastBusy[i] >= call->startTime.sec) {
3241 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3246 if (conn->call[i]) {
3247 struct rx_call *tcall = conn->call[i];
3248 MUTEX_ENTER(&tcall->lock);
3249 if (tcall->state == RX_STATE_DALLY) {
3252 MUTEX_EXIT(&tcall->lock);
3258 MUTEX_ENTER(&call->lock);
3260 /* Since the call->lock has been released it is possible that the call may
3261 * no longer be busy (the call channel cannot have been reallocated as we
3262 * haven't dropped the conn_call_lock) Therefore, we must confirm
3263 * that the call state has not changed when deciding whether or not to
3264 * force this application thread to retry by forcing a Timeout error. */
3266 if (freechannel && (call->flags & RX_CALL_PEER_BUSY)) {
3267 /* Since 'freechannel' is set, there exists another channel in this
3268 * rx_conn that the application thread might be able to use. We know
3269 * that we have the correct call since callNumber is unchanged, and we
3270 * know that the call is still busy. So, set the call error state to
3271 * rxi_busyChannelError so the application can retry the request,
3272 * presumably on a less-busy call channel. */
3274 rxi_CallError(call, RX_CALL_BUSY);
3276 MUTEX_EXIT(&conn->conn_call_lock);
3280 * Abort the call if the server is over the busy threshold. This
3281 * can be used without requiring a call structure be initialised,
3282 * or connected to a particular channel
3285 rxi_AbortIfServerBusy(osi_socket socket, struct rx_connection *conn,
3286 struct rx_packet *np)
3288 if ((rx_BusyThreshold > 0) &&
3289 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3290 rxi_SendRawAbort(socket, conn->peer->host, conn->peer->port,
3291 rx_BusyError, np, 0);
3292 if (rx_stats_active)
3293 rx_atomic_inc(&rx_stats.nBusies);
3300 static_inline struct rx_call *
3301 rxi_ReceiveClientCall(struct rx_packet *np, struct rx_connection *conn)
3304 struct rx_call *call;
3306 channel = np->header.cid & RX_CHANNELMASK;
3307 MUTEX_ENTER(&conn->conn_call_lock);
3308 call = conn->call[channel];
3309 if (!call || conn->callNumber[channel] != np->header.callNumber) {
3310 MUTEX_EXIT(&conn->conn_call_lock);
3311 if (rx_stats_active)
3312 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3316 MUTEX_ENTER(&call->lock);
3317 MUTEX_EXIT(&conn->conn_call_lock);
3319 if ((call->state == RX_STATE_DALLY)
3320 && np->header.type == RX_PACKET_TYPE_ACK) {
3321 if (rx_stats_active)
3322 rx_atomic_inc(&rx_stats.ignorePacketDally);
3323 MUTEX_EXIT(&call->lock);
3330 static_inline struct rx_call *
3331 rxi_ReceiveServerCall(osi_socket socket, struct rx_packet *np,
3332 struct rx_connection *conn)
3335 struct rx_call *call;
3337 channel = np->header.cid & RX_CHANNELMASK;
3338 MUTEX_ENTER(&conn->conn_call_lock);
3339 call = conn->call[channel];
3342 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3343 MUTEX_EXIT(&conn->conn_call_lock);
3347 call = rxi_NewCall(conn, channel); /* returns locked call */
3348 *call->callNumber = np->header.callNumber;
3349 MUTEX_EXIT(&conn->conn_call_lock);
3351 call->state = RX_STATE_PRECALL;
3352 clock_GetTime(&call->queueTime);
3353 call->app.bytesSent = 0;
3354 call->app.bytesRcvd = 0;
3355 rxi_KeepAliveOn(call);
3360 if (np->header.callNumber == conn->callNumber[channel]) {
3361 MUTEX_ENTER(&call->lock);
3362 MUTEX_EXIT(&conn->conn_call_lock);
3366 if (np->header.callNumber < conn->callNumber[channel]) {
3367 MUTEX_EXIT(&conn->conn_call_lock);
3368 if (rx_stats_active)
3369 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3373 MUTEX_ENTER(&call->lock);
3374 MUTEX_EXIT(&conn->conn_call_lock);
3376 /* Wait until the transmit queue is idle before deciding
3377 * whether to reset the current call. Chances are that the
3378 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3381 #ifdef RX_ENABLE_LOCKS
3382 if (call->state == RX_STATE_ACTIVE && !call->error) {
3383 rxi_WaitforTQBusy(call);
3384 /* If we entered error state while waiting,
3385 * must call rxi_CallError to permit rxi_ResetCall
3386 * to processed when the tqWaiter count hits zero.
3389 rxi_CallError(call, call->error);
3390 MUTEX_EXIT(&call->lock);
3394 #endif /* RX_ENABLE_LOCKS */
3395 /* If the new call cannot be taken right now send a busy and set
3396 * the error condition in this call, so that it terminates as
3397 * quickly as possible */
3398 if (call->state == RX_STATE_ACTIVE) {
3399 rxi_CallError(call, RX_CALL_DEAD);
3400 rxi_SendSpecial(call, conn, NULL, RX_PACKET_TYPE_BUSY,
3402 MUTEX_EXIT(&call->lock);
3406 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3407 MUTEX_EXIT(&call->lock);
3411 rxi_ResetCall(call, 0);
3412 /* The conn_call_lock is not held but no one else should be
3413 * using this call channel while we are processing this incoming
3414 * packet. This assignment should be safe.
3416 *call->callNumber = np->header.callNumber;
3417 call->state = RX_STATE_PRECALL;
3418 clock_GetTime(&call->queueTime);
3419 call->app.bytesSent = 0;
3420 call->app.bytesRcvd = 0;
3421 rxi_KeepAliveOn(call);
3427 /* There are two packet tracing routines available for testing and monitoring
3428 * Rx. One is called just after every packet is received and the other is
3429 * called just before every packet is sent. Received packets, have had their
3430 * headers decoded, and packets to be sent have not yet had their headers
3431 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3432 * containing the network address. Both can be modified. The return value, if
3433 * non-zero, indicates that the packet should be dropped. */
3435 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3436 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3438 /* A packet has been received off the interface. Np is the packet, socket is
3439 * the socket number it was received from (useful in determining which service
3440 * this packet corresponds to), and (host, port) reflect the host,port of the
3441 * sender. This call returns the packet to the caller if it is finished with
3442 * it, rather than de-allocating it, just as a small performance hack */
3445 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3446 afs_uint32 host, u_short port, int *tnop,
3447 struct rx_call **newcallp)
3449 struct rx_call *call;
3450 struct rx_connection *conn;
3452 int unknownService = 0;
3456 struct rx_packet *tnp;
3459 /* We don't print out the packet until now because (1) the time may not be
3460 * accurate enough until now in the lwp implementation (rx_Listener only gets
3461 * the time after the packet is read) and (2) from a protocol point of view,
3462 * this is the first time the packet has been seen */
3463 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3464 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3465 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3466 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3467 np->header.epoch, np->header.cid, np->header.callNumber,
3468 np->header.seq, np->header.flags, np));
3471 /* Account for connectionless packets */
3472 if (rx_stats_active &&
3473 ((np->header.type == RX_PACKET_TYPE_VERSION) ||
3474 (np->header.type == RX_PACKET_TYPE_DEBUG))) {
3475 struct rx_peer *peer;
3477 /* Try to look up the peer structure, but don't create one */
3478 peer = rxi_FindPeer(host, port, 0);
3480 /* Since this may not be associated with a connection, it may have
3481 * no refCount, meaning we could race with ReapConnections
3484 if (peer && (peer->refCount > 0)) {
3485 #ifdef AFS_RXERRQ_ENV
3486 if (rx_atomic_read(&peer->neterrs)) {
3487 rx_atomic_set(&peer->neterrs, 0);
3490 MUTEX_ENTER(&peer->peer_lock);
3491 peer->bytesReceived += np->length;
3492 MUTEX_EXIT(&peer->peer_lock);
3496 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3497 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3500 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3501 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3504 /* If an input tracer function is defined, call it with the packet and
3505 * network address. Note this function may modify its arguments. */
3506 if (rx_justReceived) {
3507 struct sockaddr_in addr;
3509 addr.sin_family = AF_INET;
3510 addr.sin_port = port;
3511 addr.sin_addr.s_addr = host;
3512 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3513 addr.sin_len = sizeof(addr);
3514 #endif /* AFS_OSF_ENV */
3515 drop = (*rx_justReceived) (np, &addr);
3516 /* drop packet if return value is non-zero */
3519 port = addr.sin_port; /* in case fcn changed addr */
3520 host = addr.sin_addr.s_addr;
3524 /* If packet was not sent by the client, then *we* must be the client */
3525 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3526 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3528 /* Find the connection (or fabricate one, if we're the server & if
3529 * necessary) associated with this packet */
3531 rxi_FindConnection(socket, host, port, np->header.serviceId,
3532 np->header.cid, np->header.epoch, type,
3533 np->header.securityIndex, &unknownService);
3535 /* To avoid having 2 connections just abort at each other,
3536 don't abort an abort. */
3538 if (unknownService && (np->header.type != RX_PACKET_TYPE_ABORT))
3539 rxi_SendRawAbort(socket, host, port, RX_INVALID_OPERATION,
3544 #ifdef AFS_RXERRQ_ENV
3545 if (rx_atomic_read(&conn->peer->neterrs)) {
3546 rx_atomic_set(&conn->peer->neterrs, 0);
3550 /* If we're doing statistics, then account for the incoming packet */
3551 if (rx_stats_active) {
3552 MUTEX_ENTER(&conn->peer->peer_lock);
3553 conn->peer->bytesReceived += np->length;
3554 MUTEX_EXIT(&conn->peer->peer_lock);
3557 /* If the connection is in an error state, send an abort packet and ignore
3558 * the incoming packet */
3560 /* Don't respond to an abort packet--we don't want loops! */
3561 MUTEX_ENTER(&conn->conn_data_lock);
3562 if (np->header.type != RX_PACKET_TYPE_ABORT)
3563 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3564 putConnection(conn);
3565 MUTEX_EXIT(&conn->conn_data_lock);
3569 /* Check for connection-only requests (i.e. not call specific). */
3570 if (np->header.callNumber == 0) {
3571 switch (np->header.type) {
3572 case RX_PACKET_TYPE_ABORT: {
3573 /* What if the supplied error is zero? */
3574 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3575 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3576 rxi_ConnectionError(conn, errcode);
3577 putConnection(conn);
3580 case RX_PACKET_TYPE_CHALLENGE:
3581 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3582 putConnection(conn);
3584 case RX_PACKET_TYPE_RESPONSE:
3585 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3586 putConnection(conn);
3588 case RX_PACKET_TYPE_PARAMS:
3589 case RX_PACKET_TYPE_PARAMS + 1:
3590 case RX_PACKET_TYPE_PARAMS + 2:
3591 /* ignore these packet types for now */
3592 putConnection(conn);
3596 /* Should not reach here, unless the peer is broken: send an
3598 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3599 MUTEX_ENTER(&conn->conn_data_lock);
3600 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3601 putConnection(conn);
3602 MUTEX_EXIT(&conn->conn_data_lock);
3607 if (type == RX_SERVER_CONNECTION)
3608 call = rxi_ReceiveServerCall(socket, np, conn);
3610 call = rxi_ReceiveClientCall(np, conn);
3613 putConnection(conn);
3617 MUTEX_ASSERT(&call->lock);
3618 /* Set remote user defined status from packet */
3619 call->remoteStatus = np->header.userStatus;
3621 /* Now do packet type-specific processing */
3622 switch (np->header.type) {
3623 case RX_PACKET_TYPE_DATA:
3624 /* If we're a client, and receiving a response, then all the packets
3625 * we transmitted packets are implicitly acknowledged. */
3626 if (type == RX_CLIENT_CONNECTION && !opr_queue_IsEmpty(&call->tq))
3627 rxi_AckAllInTransmitQueue(call);
3629 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3632 case RX_PACKET_TYPE_ACK:
3633 /* Respond immediately to ack packets requesting acknowledgement
3635 if (np->header.flags & RX_REQUEST_ACK) {
3637 (void)rxi_SendCallAbort(call, 0, 1, 0);
3639 (void)rxi_SendAck(call, 0, np->header.serial,
3640 RX_ACK_PING_RESPONSE, 1);
3642 np = rxi_ReceiveAckPacket(call, np, 1);
3644 case RX_PACKET_TYPE_ABORT: {
3645 /* An abort packet: reset the call, passing the error up to the user. */
3646 /* What if error is zero? */
3647 /* What if the error is -1? the application will treat it as a timeout. */
3648 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3649 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3650 rxi_CallError(call, errdata);
3651 MUTEX_EXIT(&call->lock);
3652 putConnection(conn);
3653 return np; /* xmitting; drop packet */
3655 case RX_PACKET_TYPE_BUSY: {
3656 struct clock busyTime;
3658 clock_GetTime(&busyTime);
3660 MUTEX_EXIT(&call->lock);
3662 MUTEX_ENTER(&conn->conn_call_lock);
3663 MUTEX_ENTER(&call->lock);
3664 conn->lastBusy[call->channel] = busyTime.sec;
3665 call->flags |= RX_CALL_PEER_BUSY;
3666 MUTEX_EXIT(&call->lock);
3667 MUTEX_EXIT(&conn->conn_call_lock);
3669 putConnection(conn);
3673 case RX_PACKET_TYPE_ACKALL:
3674 /* All packets acknowledged, so we can drop all packets previously
3675 * readied for sending */
3676 rxi_AckAllInTransmitQueue(call);
3679 /* Should not reach here, unless the peer is broken: send an abort
3681 rxi_CallError(call, RX_PROTOCOL_ERROR);
3682 np = rxi_SendCallAbort(call, np, 1, 0);
3685 /* Note when this last legitimate packet was received, for keep-alive
3686 * processing. Note, we delay getting the time until now in the hope that
3687 * the packet will be delivered to the user before any get time is required
3688 * (if not, then the time won't actually be re-evaluated here). */
3689 call->lastReceiveTime = clock_Sec();
3690 /* we've received a legit packet, so the channel is not busy */
3691 call->flags &= ~RX_CALL_PEER_BUSY;
3692 MUTEX_EXIT(&call->lock);
3693 putConnection(conn);
3697 /* return true if this is an "interesting" connection from the point of view
3698 of someone trying to debug the system */
3700 rxi_IsConnInteresting(struct rx_connection *aconn)
3703 struct rx_call *tcall;
3705 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3708 for (i = 0; i < RX_MAXCALLS; i++) {
3709 tcall = aconn->call[i];
3711 if ((tcall->state == RX_STATE_PRECALL)
3712 || (tcall->state == RX_STATE_ACTIVE))
3714 if ((tcall->app.mode == RX_MODE_SENDING)
3715 || (tcall->app.mode == RX_MODE_RECEIVING))
3723 /* if this is one of the last few packets AND it wouldn't be used by the
3724 receiving call to immediately satisfy a read request, then drop it on
3725 the floor, since accepting it might prevent a lock-holding thread from
3726 making progress in its reading. If a call has been cleared while in
3727 the precall state then ignore all subsequent packets until the call
3728 is assigned to a thread. */
3731 TooLow(struct rx_packet *ap, struct rx_call *acall)
3735 MUTEX_ENTER(&rx_quota_mutex);
3736 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3737 && (acall->state == RX_STATE_PRECALL))
3738 || ((rx_nFreePackets < rxi_dataQuota + 2)
3739 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3740 && (acall->flags & RX_CALL_READER_WAIT)))) {
3743 MUTEX_EXIT(&rx_quota_mutex);
3749 * Clear the attach wait flag on a connection and proceed.
3751 * Any processing waiting for a connection to be attached should be
3752 * unblocked. We clear the flag and do any other needed tasks.
3755 * the conn to unmark waiting for attach
3757 * @pre conn's conn_data_lock must be locked before calling this function
3761 rxi_ConnClearAttachWait(struct rx_connection *conn)
3763 /* Indicate that rxi_CheckReachEvent is no longer running by
3764 * clearing the flag. Must be atomic under conn_data_lock to
3765 * avoid a new call slipping by: rxi_CheckConnReach holds
3766 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3768 conn->flags &= ~RX_CONN_ATTACHWAIT;
3769 if (conn->flags & RX_CONN_NAT_PING) {
3770 conn->flags &= ~RX_CONN_NAT_PING;
3771 rxi_ScheduleNatKeepAliveEvent(conn);
3776 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3778 struct rx_connection *conn = arg1;
3779 struct rx_call *acall = arg2;
3780 struct rx_call *call = acall;
3781 struct clock when, now;
3784 MUTEX_ENTER(&conn->conn_data_lock);
3787 rxevent_Put(&conn->checkReachEvent);
3789 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3791 putConnection(conn);
3793 MUTEX_EXIT(&conn->conn_data_lock);
3797 MUTEX_ENTER(&conn->conn_call_lock);
3798 MUTEX_ENTER(&conn->conn_data_lock);
3799 for (i = 0; i < RX_MAXCALLS; i++) {
3800 struct rx_call *tc = conn->call[i];
3801 if (tc && tc->state == RX_STATE_PRECALL) {
3807 rxi_ConnClearAttachWait(conn);
3808 MUTEX_EXIT(&conn->conn_data_lock);
3809 MUTEX_EXIT(&conn->conn_call_lock);
3814 MUTEX_ENTER(&call->lock);
3815 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3817 MUTEX_EXIT(&call->lock);
3819 clock_GetTime(&now);
3821 when.sec += RX_CHECKREACH_TIMEOUT;
3822 MUTEX_ENTER(&conn->conn_data_lock);
3823 if (!conn->checkReachEvent) {
3824 MUTEX_ENTER(&rx_refcnt_mutex);
3826 MUTEX_EXIT(&rx_refcnt_mutex);
3827 conn->checkReachEvent = rxevent_Post(&when, &now,
3828 rxi_CheckReachEvent, conn,
3831 MUTEX_EXIT(&conn->conn_data_lock);
3837 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3839 struct rx_service *service = conn->service;
3840 struct rx_peer *peer = conn->peer;
3841 afs_uint32 now, lastReach;
3843 if (service->checkReach == 0)
3847 MUTEX_ENTER(&peer->peer_lock);
3848 lastReach = peer->lastReachTime;
3849 MUTEX_EXIT(&peer->peer_lock);
3850 if (now - lastReach < RX_CHECKREACH_TTL)
3853 MUTEX_ENTER(&conn->conn_data_lock);
3854 if (conn->flags & RX_CONN_ATTACHWAIT) {
3855 MUTEX_EXIT(&conn->conn_data_lock);
3858 conn->flags |= RX_CONN_ATTACHWAIT;
3859 MUTEX_EXIT(&conn->conn_data_lock);
3860 if (!conn->checkReachEvent)
3861 rxi_CheckReachEvent(NULL, conn, call, 0);
3866 /* try to attach call, if authentication is complete */
3868 TryAttach(struct rx_call *acall, osi_socket socket,
3869 int *tnop, struct rx_call **newcallp,
3872 struct rx_connection *conn = acall->conn;
3874 if (conn->type == RX_SERVER_CONNECTION
3875 && acall->state == RX_STATE_PRECALL) {
3876 /* Don't attach until we have any req'd. authentication. */
3877 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3878 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3879 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3880 /* Note: this does not necessarily succeed; there
3881 * may not any proc available
3884 rxi_ChallengeOn(acall->conn);
3889 /* A data packet has been received off the interface. This packet is
3890 * appropriate to the call (the call is in the right state, etc.). This
3891 * routine can return a packet to the caller, for re-use */
3893 static struct rx_packet *
3894 rxi_ReceiveDataPacket(struct rx_call *call,
3895 struct rx_packet *np, int istack,
3896 osi_socket socket, afs_uint32 host, u_short port,
3897 int *tnop, struct rx_call **newcallp)
3899 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3904 afs_uint32 serial=0, flags=0;
3906 struct rx_packet *tnp;
3907 if (rx_stats_active)
3908 rx_atomic_inc(&rx_stats.dataPacketsRead);
3911 /* If there are no packet buffers, drop this new packet, unless we can find
3912 * packet buffers from inactive calls */
3914 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3915 MUTEX_ENTER(&rx_freePktQ_lock);
3916 rxi_NeedMorePackets = TRUE;
3917 MUTEX_EXIT(&rx_freePktQ_lock);
3918 if (rx_stats_active)
3919 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3920 rxi_calltrace(RX_TRACE_DROP, call);
3921 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3922 /* We used to clear the receive queue here, in an attempt to free
3923 * packets. However this is unsafe if the queue has received a
3924 * soft ACK for the final packet */
3925 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3931 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3932 * packet is one of several packets transmitted as a single
3933 * datagram. Do not send any soft or hard acks until all packets
3934 * in a jumbogram have been processed. Send negative acks right away.
3936 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3937 /* tnp is non-null when there are more packets in the
3938 * current jumbo gram */
3945 seq = np->header.seq;
3946 serial = np->header.serial;
3947 flags = np->header.flags;
3949 /* If the call is in an error state, send an abort message */
3951 return rxi_SendCallAbort(call, np, istack, 0);
3953 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3954 * AFS 3.5 jumbogram. */
3955 if (flags & RX_JUMBO_PACKET) {
3956 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3961 if (np->header.spare != 0) {
3962 MUTEX_ENTER(&call->conn->conn_data_lock);
3963 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3964 MUTEX_EXIT(&call->conn->conn_data_lock);
3967 /* The usual case is that this is the expected next packet */
3968 if (seq == call->rnext) {
3970 /* Check to make sure it is not a duplicate of one already queued */
3971 if (!opr_queue_IsEmpty(&call->rq)
3972 && opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq == seq) {
3973 if (rx_stats_active)
3974 rx_atomic_inc(&rx_stats.dupPacketsRead);
3975 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3976 rxi_CancelDelayedAckEvent(call);
3977 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3983 /* It's the next packet. Stick it on the receive queue
3984 * for this call. Set newPackets to make sure we wake
3985 * the reader once all packets have been processed */
3986 #ifdef RX_TRACK_PACKETS
3987 np->flags |= RX_PKTFLAG_RQ;
3989 opr_queue_Prepend(&call->rq, &np->entry);
3990 #ifdef RXDEBUG_PACKET
3992 #endif /* RXDEBUG_PACKET */
3994 np = NULL; /* We can't use this anymore */
3997 /* If an ack is requested then set a flag to make sure we
3998 * send an acknowledgement for this packet */
3999 if (flags & RX_REQUEST_ACK) {
4000 ackNeeded = RX_ACK_REQUESTED;
4003 /* Keep track of whether we have received the last packet */
4004 if (flags & RX_LAST_PACKET) {
4005 call->flags |= RX_CALL_HAVE_LAST;
4009 /* Check whether we have all of the packets for this call */
4010 if (call->flags & RX_CALL_HAVE_LAST) {
4011 afs_uint32 tseq; /* temporary sequence number */
4012 struct opr_queue *cursor;
4014 for (tseq = seq, opr_queue_Scan(&call->rq, cursor)) {
4015 struct rx_packet *tp;
4017 tp = opr_queue_Entry(cursor, struct rx_packet, entry);
4018 if (tseq != tp->header.seq)
4020 if (tp->header.flags & RX_LAST_PACKET) {
4021 call->flags |= RX_CALL_RECEIVE_DONE;
4028 /* Provide asynchronous notification for those who want it
4029 * (e.g. multi rx) */
4030 if (call->arrivalProc) {
4031 (*call->arrivalProc) (call, call->arrivalProcHandle,
4032 call->arrivalProcArg);
4033 call->arrivalProc = (void (*)())0;
4036 /* Update last packet received */
4039 /* If there is no server process serving this call, grab
4040 * one, if available. We only need to do this once. If a
4041 * server thread is available, this thread becomes a server
4042 * thread and the server thread becomes a listener thread. */
4044 TryAttach(call, socket, tnop, newcallp, 0);
4047 /* This is not the expected next packet. */
4049 /* Determine whether this is a new or old packet, and if it's
4050 * a new one, whether it fits into the current receive window.
4051 * Also figure out whether the packet was delivered in sequence.
4052 * We use the prev variable to determine whether the new packet
4053 * is the successor of its immediate predecessor in the
4054 * receive queue, and the missing flag to determine whether
4055 * any of this packets predecessors are missing. */
4057 afs_uint32 prev; /* "Previous packet" sequence number */
4058 struct opr_queue *cursor;
4059 int missing; /* Are any predecessors missing? */
4061 /* If the new packet's sequence number has been sent to the
4062 * application already, then this is a duplicate */
4063 if (seq < call->rnext) {
4064 if (rx_stats_active)
4065 rx_atomic_inc(&rx_stats.dupPacketsRead);
4066 rxi_CancelDelayedAckEvent(call);
4067 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
4073 /* If the sequence number is greater than what can be
4074 * accomodated by the current window, then send a negative
4075 * acknowledge and drop the packet */
4076 if ((call->rnext + call->rwind) <= seq) {
4077 rxi_CancelDelayedAckEvent(call);
4078 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
4085 /* Look for the packet in the queue of old received packets */
4086 prev = call->rnext - 1;
4088 for (opr_queue_Scan(&call->rq, cursor)) {
4089 struct rx_packet *tp
4090 = opr_queue_Entry(cursor, struct rx_packet, entry);
4092 /*Check for duplicate packet */
4093 if (seq == tp->header.seq) {
4094 if (rx_stats_active)
4095 rx_atomic_inc(&rx_stats.dupPacketsRead);
4096 rxi_CancelDelayedAckEvent(call);
4097 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
4103 /* If we find a higher sequence packet, break out and
4104 * insert the new packet here. */
4105 if (seq < tp->header.seq)
4107 /* Check for missing packet */
4108 if (tp->header.seq != prev + 1) {
4112 prev = tp->header.seq;
4115 /* Keep track of whether we have received the last packet. */
4116 if (flags & RX_LAST_PACKET) {
4117 call->flags |= RX_CALL_HAVE_LAST;
4120 /* It's within the window: add it to the the receive queue.
4121 * tp is left by the previous loop either pointing at the
4122 * packet before which to insert the new packet, or at the
4123 * queue head if the queue is empty or the packet should be
4125 #ifdef RX_TRACK_PACKETS
4126 np->flags |= RX_PKTFLAG_RQ;
4128 #ifdef RXDEBUG_PACKET
4130 #endif /* RXDEBUG_PACKET */
4131 opr_queue_InsertBefore(cursor, &np->entry);
4135 /* Check whether we have all of the packets for this call */
4136 if ((call->flags & RX_CALL_HAVE_LAST)
4137 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4138 afs_uint32 tseq; /* temporary sequence number */
4141 for (opr_queue_Scan(&call->rq, cursor)) {
4142 struct rx_packet *tp
4143 = opr_queue_Entry(cursor, struct rx_packet, entry);
4144 if (tseq != tp->header.seq)
4146 if (tp->header.flags & RX_LAST_PACKET) {
4147 call->flags |= RX_CALL_RECEIVE_DONE;
4154 /* We need to send an ack of the packet is out of sequence,
4155 * or if an ack was requested by the peer. */
4156 if (seq != prev + 1 || missing) {
4157 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4158 } else if (flags & RX_REQUEST_ACK) {
4159 ackNeeded = RX_ACK_REQUESTED;
4162 /* Acknowledge the last packet for each call */
4163 if (flags & RX_LAST_PACKET) {
4174 * If the receiver is waiting for an iovec, fill the iovec
4175 * using the data from the receive queue */
4176 if (call->flags & RX_CALL_IOVEC_WAIT) {
4177 didHardAck = rxi_FillReadVec(call, serial);
4178 /* the call may have been aborted */
4187 /* Wakeup the reader if any */
4188 if ((call->flags & RX_CALL_READER_WAIT)
4189 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4190 || (call->iovNext >= call->iovMax)
4191 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4192 call->flags &= ~RX_CALL_READER_WAIT;
4193 #ifdef RX_ENABLE_LOCKS
4194 CV_BROADCAST(&call->cv_rq);
4196 osi_rxWakeup(&call->rq);
4202 * Send an ack when requested by the peer, or once every
4203 * rxi_SoftAckRate packets until the last packet has been
4204 * received. Always send a soft ack for the last packet in
4205 * the server's reply. */
4207 rxi_CancelDelayedAckEvent(call);
4208 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4209 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4210 rxi_CancelDelayedAckEvent(call);
4211 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4212 } else if (call->nSoftAcks) {
4213 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4214 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4216 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4217 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4218 rxi_CancelDelayedAckEvent(call);
4225 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4227 struct rx_peer *peer = conn->peer;
4229 MUTEX_ENTER(&peer->peer_lock);
4230 peer->lastReachTime = clock_Sec();
4231 MUTEX_EXIT(&peer->peer_lock);
4233 MUTEX_ENTER(&conn->conn_data_lock);
4234 if (conn->flags & RX_CONN_ATTACHWAIT) {
4237 rxi_ConnClearAttachWait(conn);
4238 MUTEX_EXIT(&conn->conn_data_lock);
4240 for (i = 0; i < RX_MAXCALLS; i++) {
4241 struct rx_call *call = conn->call[i];
4244 MUTEX_ENTER(&call->lock);
4245 /* tnop can be null if newcallp is null */
4246 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4248 MUTEX_EXIT(&call->lock);
4252 MUTEX_EXIT(&conn->conn_data_lock);
4255 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4257 rx_ack_reason(int reason)
4260 case RX_ACK_REQUESTED:
4262 case RX_ACK_DUPLICATE:
4264 case RX_ACK_OUT_OF_SEQUENCE:
4266 case RX_ACK_EXCEEDS_WINDOW:
4268 case RX_ACK_NOSPACE:
4272 case RX_ACK_PING_RESPONSE:
4285 /* The real smarts of the whole thing. */
4286 static struct rx_packet *
4287 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4290 struct rx_ackPacket *ap;
4292 struct rx_packet *tp;
4293 struct rx_connection *conn = call->conn;
4294 struct rx_peer *peer = conn->peer;
4295 struct opr_queue *cursor;
4296 struct clock now; /* Current time, for RTT calculations */
4304 int newAckCount = 0;
4305 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4306 int pktsize = 0; /* Set if we need to update the peer mtu */
4307 int conn_data_locked = 0;
4309 if (rx_stats_active)
4310 rx_atomic_inc(&rx_stats.ackPacketsRead);
4311 ap = (struct rx_ackPacket *)rx_DataOf(np);
4312 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4314 return np; /* truncated ack packet */
4316 /* depends on ack packet struct */
4317 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4318 first = ntohl(ap->firstPacket);
4319 prev = ntohl(ap->previousPacket);
4320 serial = ntohl(ap->serial);
4323 * Ignore ack packets received out of order while protecting
4324 * against peers that set the previousPacket field to a packet
4325 * serial number instead of a sequence number.
4327 if (first < call->tfirst ||
4328 (first == call->tfirst && prev < call->tprev && prev < call->tfirst
4335 if (np->header.flags & RX_SLOW_START_OK) {
4336 call->flags |= RX_CALL_SLOW_START_OK;
4339 if (ap->reason == RX_ACK_PING_RESPONSE)
4340 rxi_UpdatePeerReach(conn, call);
4342 if (conn->lastPacketSizeSeq) {
4343 MUTEX_ENTER(&conn->conn_data_lock);
4344 conn_data_locked = 1;
4345 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4346 pktsize = conn->lastPacketSize;
4347 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4350 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4351 if (!conn_data_locked) {
4352 MUTEX_ENTER(&conn->conn_data_lock);
4353 conn_data_locked = 1;
4355 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4356 /* process mtu ping ack */
4357 pktsize = conn->lastPingSize;
4358 conn->lastPingSizeSer = conn->lastPingSize = 0;
4362 if (conn_data_locked) {
4363 MUTEX_EXIT(&conn->conn_data_lock);
4364 conn_data_locked = 0;
4368 if (rxdebug_active) {
4372 len = _snprintf(msg, sizeof(msg),
4373 "tid[%d] RACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4374 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4375 ntohl(ap->serial), ntohl(ap->previousPacket),
4376 (unsigned int)np->header.seq, ntohl(ap->firstPacket),
4377 ap->nAcks, ntohs(ap->bufferSpace) );
4381 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4382 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4386 OutputDebugString(msg);
4388 #else /* AFS_NT40_ENV */
4391 "RACK: reason %x previous %u seq %u serial %u first %u",
4392 ap->reason, ntohl(ap->previousPacket),
4393 (unsigned int)np->header.seq, (unsigned int)serial,
4394 ntohl(ap->firstPacket));
4397 for (offset = 0; offset < nAcks; offset++)
4398 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4403 #endif /* AFS_NT40_ENV */
4406 MUTEX_ENTER(&peer->peer_lock);
4409 * Start somewhere. Can't assume we can send what we can receive,
4410 * but we are clearly receiving.
4412 if (!peer->maxPacketSize)
4413 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4415 if (pktsize > peer->maxPacketSize) {
4416 peer->maxPacketSize = pktsize;
4417 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4418 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4419 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4420 rxi_ScheduleGrowMTUEvent(call, 1);
4425 clock_GetTime(&now);
4427 /* The transmit queue splits into 4 sections.
4429 * The first section is packets which have now been acknowledged
4430 * by a window size change in the ack. These have reached the
4431 * application layer, and may be discarded. These are packets
4432 * with sequence numbers < ap->firstPacket.
4434 * The second section is packets which have sequence numbers in
4435 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4436 * contents of the packet's ack array determines whether these
4437 * packets are acknowledged or not.
4439 * The third section is packets which fall above the range
4440 * addressed in the ack packet. These have not yet been received
4443 * The four section is packets which have not yet been transmitted.
4444 * These packets will have a header.serial of 0.
4447 /* First section - implicitly acknowledged packets that can be
4451 tp = opr_queue_First(&call->tq, struct rx_packet, entry);
4452 while(!opr_queue_IsEnd(&call->tq, &tp->entry) && tp->header.seq < first) {
4453 struct rx_packet *next;
4455 next = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4456 call->tfirst = tp->header.seq + 1;
4458 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4460 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4463 #ifdef RX_ENABLE_LOCKS
4464 /* XXX Hack. Because we have to release the global call lock when sending
4465 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4466 * in rxi_Start sending packets out because packets may move to the
4467 * freePacketQueue as result of being here! So we drop these packets until
4468 * we're safely out of the traversing. Really ugly!
4469 * To make it even uglier, if we're using fine grain locking, we can
4470 * set the ack bits in the packets and have rxi_Start remove the packets
4471 * when it's done transmitting.
4473 if (call->flags & RX_CALL_TQ_BUSY) {
4474 tp->flags |= RX_PKTFLAG_ACKED;
4475 call->flags |= RX_CALL_TQ_SOME_ACKED;
4477 #endif /* RX_ENABLE_LOCKS */
4479 opr_queue_Remove(&tp->entry);
4480 #ifdef RX_TRACK_PACKETS
4481 tp->flags &= ~RX_PKTFLAG_TQ;
4483 #ifdef RXDEBUG_PACKET
4485 #endif /* RXDEBUG_PACKET */
4486 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4491 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4493 /* Second section of the queue - packets for which we are receiving
4496 * Go through the explicit acks/nacks and record the results in
4497 * the waiting packets. These are packets that can't be released
4498 * yet, even with a positive acknowledge. This positive
4499 * acknowledge only means the packet has been received by the
4500 * peer, not that it will be retained long enough to be sent to
4501 * the peer's upper level. In addition, reset the transmit timers
4502 * of any missing packets (those packets that must be missing
4503 * because this packet was out of sequence) */
4505 call->nSoftAcked = 0;
4507 while (!opr_queue_IsEnd(&call->tq, &tp->entry)
4508 && tp->header.seq < first + nAcks) {
4509 /* Set the acknowledge flag per packet based on the
4510 * information in the ack packet. An acknowlegded packet can
4511 * be downgraded when the server has discarded a packet it
4512 * soacked previously, or when an ack packet is received
4513 * out of sequence. */
4514 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4515 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4517 tp->flags |= RX_PKTFLAG_ACKED;
4518 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4525 } else /* RX_ACK_TYPE_NACK */ {
4526 tp->flags &= ~RX_PKTFLAG_ACKED;
4530 tp = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4533 /* We don't need to take any action with the 3rd or 4th section in the
4534 * queue - they're not addressed by the contents of this ACK packet.
4537 /* If the window has been extended by this acknowledge packet,
4538 * then wakeup a sender waiting in alloc for window space, or try
4539 * sending packets now, if he's been sitting on packets due to
4540 * lack of window space */
4541 if (call->tnext < (call->tfirst + call->twind)) {
4542 #ifdef RX_ENABLE_LOCKS
4543 CV_SIGNAL(&call->cv_twind);
4545 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4546 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4547 osi_rxWakeup(&call->twind);
4550 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4551 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4555 /* if the ack packet has a receivelen field hanging off it,
4556 * update our state */
4557 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4560 /* If the ack packet has a "recommended" size that is less than
4561 * what I am using now, reduce my size to match */
4562 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4563 (int)sizeof(afs_int32), &tSize);
4564 tSize = (afs_uint32) ntohl(tSize);
4565 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4567 /* Get the maximum packet size to send to this peer */
4568 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4570 tSize = (afs_uint32) ntohl(tSize);
4571 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4572 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4574 /* sanity check - peer might have restarted with different params.
4575 * If peer says "send less", dammit, send less... Peer should never
4576 * be unable to accept packets of the size that prior AFS versions would
4577 * send without asking. */
4578 if (peer->maxMTU != tSize) {
4579 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4581 peer->maxMTU = tSize;
4582 peer->MTU = MIN(tSize, peer->MTU);
4583 call->MTU = MIN(call->MTU, tSize);
4586 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4589 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4590 (int)sizeof(afs_int32), &tSize);
4591 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4592 if (tSize < call->twind) { /* smaller than our send */
4593 call->twind = tSize; /* window, we must send less... */
4594 call->ssthresh = MIN(call->twind, call->ssthresh);
4595 call->conn->twind[call->channel] = call->twind;
4598 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4599 * network MTU confused with the loopback MTU. Calculate the
4600 * maximum MTU here for use in the slow start code below.
4602 /* Did peer restart with older RX version? */
4603 if (peer->maxDgramPackets > 1) {
4604 peer->maxDgramPackets = 1;
4606 } else if (np->length >=
4607 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4610 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4611 sizeof(afs_int32), &tSize);
4612 tSize = (afs_uint32) ntohl(tSize);
4614 * As of AFS 3.5 we set the send window to match the receive window.
4616 if (tSize < call->twind) {
4617 call->twind = tSize;
4618 call->conn->twind[call->channel] = call->twind;
4619 call->ssthresh = MIN(call->twind, call->ssthresh);
4620 } else if (tSize > call->twind) {
4621 call->twind = tSize;
4622 call->conn->twind[call->channel] = call->twind;
4626 * As of AFS 3.5, a jumbogram is more than one fixed size
4627 * packet transmitted in a single UDP datagram. If the remote
4628 * MTU is smaller than our local MTU then never send a datagram
4629 * larger than the natural MTU.
4632 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4633 (int)sizeof(afs_int32), &tSize);
4634 maxDgramPackets = (afs_uint32) ntohl(tSize);
4635 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4637 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4638 if (maxDgramPackets > 1) {
4639 peer->maxDgramPackets = maxDgramPackets;
4640 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4642 peer->maxDgramPackets = 1;
4643 call->MTU = peer->natMTU;
4645 } else if (peer->maxDgramPackets > 1) {
4646 /* Restarted with lower version of RX */
4647 peer->maxDgramPackets = 1;
4649 } else if (peer->maxDgramPackets > 1
4650 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4651 /* Restarted with lower version of RX */
4652 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4653 peer->natMTU = OLD_MAX_PACKET_SIZE;
4654 peer->MTU = OLD_MAX_PACKET_SIZE;
4655 peer->maxDgramPackets = 1;
4656 peer->nDgramPackets = 1;
4658 call->MTU = OLD_MAX_PACKET_SIZE;
4663 * Calculate how many datagrams were successfully received after
4664 * the first missing packet and adjust the negative ack counter
4669 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4670 if (call->nNacks < nNacked) {
4671 call->nNacks = nNacked;
4674 call->nAcks += newAckCount;
4678 /* If the packet contained new acknowledgements, rather than just
4679 * being a duplicate of one we have previously seen, then we can restart
4682 if (newAckCount > 0)
4683 rxi_rto_packet_acked(call, istack);
4685 if (call->flags & RX_CALL_FAST_RECOVER) {
4686 if (newAckCount == 0) {
4687 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4689 call->flags &= ~RX_CALL_FAST_RECOVER;
4690 call->cwind = call->nextCwind;
4691 call->nextCwind = 0;
4694 call->nCwindAcks = 0;
4695 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4696 /* Three negative acks in a row trigger congestion recovery */
4697 call->flags |= RX_CALL_FAST_RECOVER;
4698 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4700 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4701 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4702 call->nextCwind = call->ssthresh;
4705 peer->MTU = call->MTU;
4706 peer->cwind = call->nextCwind;
4707 peer->nDgramPackets = call->nDgramPackets;
4709 call->congestSeq = peer->congestSeq;
4711 /* Reset the resend times on the packets that were nacked
4712 * so we will retransmit as soon as the window permits
4716 for (opr_queue_ScanBackwards(&call->tq, cursor)) {
4717 struct rx_packet *tp =
4718 opr_queue_Entry(cursor, struct rx_packet, entry);
4720 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4721 tp->flags &= ~RX_PKTFLAG_SENT;
4723 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4728 /* If cwind is smaller than ssthresh, then increase
4729 * the window one packet for each ack we receive (exponential
4731 * If cwind is greater than or equal to ssthresh then increase
4732 * the congestion window by one packet for each cwind acks we
4733 * receive (linear growth). */
4734 if (call->cwind < call->ssthresh) {
4736 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4737 call->nCwindAcks = 0;
4739 call->nCwindAcks += newAckCount;
4740 if (call->nCwindAcks >= call->cwind) {
4741 call->nCwindAcks = 0;
4742 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4746 * If we have received several acknowledgements in a row then
4747 * it is time to increase the size of our datagrams
4749 if ((int)call->nAcks > rx_nDgramThreshold) {
4750 if (peer->maxDgramPackets > 1) {
4751 if (call->nDgramPackets < peer->maxDgramPackets) {
4752 call->nDgramPackets++;
4754 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4755 } else if (call->MTU < peer->maxMTU) {
4756 /* don't upgrade if we can't handle it */
4757 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4758 call->MTU = peer->ifMTU;
4760 call->MTU += peer->natMTU;
4761 call->MTU = MIN(call->MTU, peer->maxMTU);
4768 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4770 /* Servers need to hold the call until all response packets have
4771 * been acknowledged. Soft acks are good enough since clients
4772 * are not allowed to clear their receive queues. */
4773 if (call->state == RX_STATE_HOLD
4774 && call->tfirst + call->nSoftAcked >= call->tnext) {
4775 call->state = RX_STATE_DALLY;
4776 rxi_ClearTransmitQueue(call, 0);
4777 rxi_CancelKeepAliveEvent(call);
4778 } else if (!opr_queue_IsEmpty(&call->tq)) {
4779 rxi_Start(call, istack);
4784 /* Received a response to a challenge packet */
4785 static struct rx_packet *
4786 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4787 struct rx_packet *np, int istack)
4791 /* Ignore the packet if we're the client */
4792 if (conn->type == RX_CLIENT_CONNECTION)
4795 /* If already authenticated, ignore the packet (it's probably a retry) */
4796 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4799 /* Otherwise, have the security object evaluate the response packet */
4800 error = RXS_CheckResponse(conn->securityObject, conn, np);
4802 /* If the response is invalid, reset the connection, sending
4803 * an abort to the peer */
4807 rxi_ConnectionError(conn, error);
4808 MUTEX_ENTER(&conn->conn_data_lock);
4809 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4810 MUTEX_EXIT(&conn->conn_data_lock);
4813 /* If the response is valid, any calls waiting to attach
4814 * servers can now do so */
4817 for (i = 0; i < RX_MAXCALLS; i++) {
4818 struct rx_call *call = conn->call[i];
4820 MUTEX_ENTER(&call->lock);
4821 if (call->state == RX_STATE_PRECALL)
4822 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4823 /* tnop can be null if newcallp is null */
4824 MUTEX_EXIT(&call->lock);
4828 /* Update the peer reachability information, just in case
4829 * some calls went into attach-wait while we were waiting
4830 * for authentication..
4832 rxi_UpdatePeerReach(conn, NULL);
4837 /* A client has received an authentication challenge: the security
4838 * object is asked to cough up a respectable response packet to send
4839 * back to the server. The server is responsible for retrying the
4840 * challenge if it fails to get a response. */
4842 static struct rx_packet *
4843 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4844 struct rx_packet *np, int istack)
4848 /* Ignore the challenge if we're the server */
4849 if (conn->type == RX_SERVER_CONNECTION)
4852 /* Ignore the challenge if the connection is otherwise idle; someone's
4853 * trying to use us as an oracle. */
4854 if (!rxi_HasActiveCalls(conn))
4857 /* Send the security object the challenge packet. It is expected to fill
4858 * in the response. */
4859 error = RXS_GetResponse(conn->securityObject, conn, np);
4861 /* If the security object is unable to return a valid response, reset the
4862 * connection and send an abort to the peer. Otherwise send the response
4863 * packet to the peer connection. */
4865 rxi_ConnectionError(conn, error);
4866 MUTEX_ENTER(&conn->conn_data_lock);
4867 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4868 MUTEX_EXIT(&conn->conn_data_lock);
4870 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4871 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4877 /* Find an available server process to service the current request in
4878 * the given call structure. If one isn't available, queue up this
4879 * call so it eventually gets one */
4881 rxi_AttachServerProc(struct rx_call *call,
4882 osi_socket socket, int *tnop,
4883 struct rx_call **newcallp)
4885 struct rx_serverQueueEntry *sq;
4886 struct rx_service *service = call->conn->service;
4889 /* May already be attached */
4890 if (call->state == RX_STATE_ACTIVE)
4893 MUTEX_ENTER(&rx_serverPool_lock);
4895 haveQuota = QuotaOK(service);
4896 if ((!haveQuota) || opr_queue_IsEmpty(&rx_idleServerQueue)) {
4897 /* If there are no processes available to service this call,
4898 * put the call on the incoming call queue (unless it's
4899 * already on the queue).
4901 #ifdef RX_ENABLE_LOCKS
4903 ReturnToServerPool(service);
4904 #endif /* RX_ENABLE_LOCKS */
4906 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4907 call->flags |= RX_CALL_WAIT_PROC;
4908 rx_atomic_inc(&rx_nWaiting);
4909 rx_atomic_inc(&rx_nWaited);
4910 rxi_calltrace(RX_CALL_ARRIVAL, call);
4911 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4912 opr_queue_Append(&rx_incomingCallQueue, &call->entry);
4915 sq = opr_queue_Last(&rx_idleServerQueue,
4916 struct rx_serverQueueEntry, entry);
4918 /* If hot threads are enabled, and both newcallp and sq->socketp
4919 * are non-null, then this thread will process the call, and the
4920 * idle server thread will start listening on this threads socket.
4922 opr_queue_Remove(&sq->entry);
4924 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4927 *sq->socketp = socket;
4928 clock_GetTime(&call->startTime);
4929 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4933 if (call->flags & RX_CALL_WAIT_PROC) {
4934 /* Conservative: I don't think this should happen */
4935 call->flags &= ~RX_CALL_WAIT_PROC;
4936 rx_atomic_dec(&rx_nWaiting);
4937 if (opr_queue_IsOnQueue(&call->entry)) {
4938 opr_queue_Remove(&call->entry);
4941 call->state = RX_STATE_ACTIVE;
4942 call->app.mode = RX_MODE_RECEIVING;
4943 #ifdef RX_KERNEL_TRACE
4945 int glockOwner = ISAFS_GLOCK();
4948 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4949 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4955 if (call->flags & RX_CALL_CLEARED) {
4956 /* send an ack now to start the packet flow up again */
4957 call->flags &= ~RX_CALL_CLEARED;
4958 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4960 #ifdef RX_ENABLE_LOCKS
4963 service->nRequestsRunning++;
4964 MUTEX_ENTER(&rx_quota_mutex);
4965 if (service->nRequestsRunning <= service->minProcs)
4968 MUTEX_EXIT(&rx_quota_mutex);
4972 MUTEX_EXIT(&rx_serverPool_lock);
4975 /* Delay the sending of an acknowledge event for a short while, while
4976 * a new call is being prepared (in the case of a client) or a reply
4977 * is being prepared (in the case of a server). Rather than sending
4978 * an ack packet, an ACKALL packet is sent. */
4980 rxi_AckAll(struct rx_call *call)
4982 rxi_SendSpecial(call, call->conn, NULL, RX_PACKET_TYPE_ACKALL,
4984 call->flags |= RX_CALL_ACKALL_SENT;
4988 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4991 struct rx_call *call = arg1;
4992 #ifdef RX_ENABLE_LOCKS
4994 MUTEX_ENTER(&call->lock);
4995 if (event == call->delayedAckEvent)
4996 rxevent_Put(&call->delayedAckEvent);
4997 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4999 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
5001 MUTEX_EXIT(&call->lock);
5002 #else /* RX_ENABLE_LOCKS */
5004 rxevent_Put(&call->delayedAckEvent);
5005 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
5006 #endif /* RX_ENABLE_LOCKS */
5009 #ifdef RX_ENABLE_LOCKS
5010 /* Set ack in all packets in transmit queue. rxi_Start will deal with
5011 * clearing them out.
5014 rxi_SetAcksInTransmitQueue(struct rx_call *call)
5016 struct opr_queue *cursor;
5019 for (opr_queue_Scan(&call->tq, cursor)) {
5021 = opr_queue_Entry(cursor, struct rx_packet, entry);
5023 p->flags |= RX_PKTFLAG_ACKED;
5028 call->flags |= RX_CALL_TQ_CLEARME;
5029 call->flags |= RX_CALL_TQ_SOME_ACKED;
5032 rxi_rto_cancel(call);
5034 call->tfirst = call->tnext;
5035 call->nSoftAcked = 0;
5037 if (call->flags & RX_CALL_FAST_RECOVER) {
5038 call->flags &= ~RX_CALL_FAST_RECOVER;
5039 call->cwind = call->nextCwind;
5040 call->nextCwind = 0;
5043 CV_SIGNAL(&call->cv_twind);
5045 #endif /* RX_ENABLE_LOCKS */
5048 * Acknowledge the whole transmit queue.
5050 * If we're running without locks, or the transmit queue isn't busy, then
5051 * we can just clear the queue now. Otherwise, we have to mark all of the
5052 * packets as acknowledged, and let rxi_Start clear it later on
5055 rxi_AckAllInTransmitQueue(struct rx_call *call)
5057 #ifdef RX_ENABLE_LOCKS
5058 if (call->flags & RX_CALL_TQ_BUSY) {
5059 rxi_SetAcksInTransmitQueue(call);
5063 rxi_ClearTransmitQueue(call, 0);
5065 /* Clear out the transmit queue for the current call (all packets have
5066 * been received by peer) */
5068 rxi_ClearTransmitQueue(struct rx_call *call, int force)
5070 #ifdef RX_ENABLE_LOCKS
5071 struct opr_queue *cursor;
5072 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
5074 for (opr_queue_Scan(&call->tq, cursor)) {
5076 = opr_queue_Entry(cursor, struct rx_packet, entry);
5078 p->flags |= RX_PKTFLAG_ACKED;
5082 call->flags |= RX_CALL_TQ_CLEARME;
5083 call->flags |= RX_CALL_TQ_SOME_ACKED;
5086 #endif /* RX_ENABLE_LOCKS */
5087 #ifdef RXDEBUG_PACKET
5089 #endif /* RXDEBUG_PACKET */
5090 rxi_FreePackets(0, &call->tq);
5091 rxi_WakeUpTransmitQueue(call);
5092 #ifdef RX_ENABLE_LOCKS
5093 call->flags &= ~RX_CALL_TQ_CLEARME;
5097 rxi_rto_cancel(call);
5098 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5099 call->nSoftAcked = 0;
5101 if (call->flags & RX_CALL_FAST_RECOVER) {
5102 call->flags &= ~RX_CALL_FAST_RECOVER;
5103 call->cwind = call->nextCwind;
5105 #ifdef RX_ENABLE_LOCKS
5106 CV_SIGNAL(&call->cv_twind);
5108 osi_rxWakeup(&call->twind);
5113 rxi_ClearReceiveQueue(struct rx_call *call)
5115 if (!opr_queue_IsEmpty(&call->rq)) {
5118 count = rxi_FreePackets(0, &call->rq);
5119 rx_packetReclaims += count;
5120 #ifdef RXDEBUG_PACKET
5122 if ( call->rqc != 0 )
5123 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5125 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5127 if (call->state == RX_STATE_PRECALL) {
5128 call->flags |= RX_CALL_CLEARED;
5132 /* Send an abort packet for the specified call */
5133 static struct rx_packet *
5134 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5135 int istack, int force)
5137 afs_int32 error, cerror;
5138 struct clock when, now;
5143 switch (call->error) {
5146 cerror = RX_CALL_TIMEOUT;
5149 cerror = call->error;
5152 /* Clients should never delay abort messages */
5153 if (rx_IsClientConn(call->conn))
5156 if (call->abortCode != cerror) {
5157 call->abortCode = cerror;
5158 call->abortCount = 0;
5161 if (force || rxi_callAbortThreshhold == 0
5162 || call->abortCount < rxi_callAbortThreshhold) {
5163 rxi_CancelDelayedAbortEvent(call);
5164 error = htonl(cerror);
5167 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5168 (char *)&error, sizeof(error), istack);
5169 } else if (!call->delayedAbortEvent) {
5170 clock_GetTime(&now);
5172 clock_Addmsec(&when, rxi_callAbortDelay);
5173 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5174 call->delayedAbortEvent =
5175 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5181 rxi_CancelDelayedAbortEvent(struct rx_call *call)
5183 if (call->delayedAbortEvent) {
5184 rxevent_Cancel(&call->delayedAbortEvent);
5185 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5189 /* Send an abort packet for the specified connection. Packet is an
5190 * optional pointer to a packet that can be used to send the abort.
5191 * Once the number of abort messages reaches the threshhold, an
5192 * event is scheduled to send the abort. Setting the force flag
5193 * overrides sending delayed abort messages.
5195 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5196 * to send the abort packet.
5199 rxi_SendConnectionAbort(struct rx_connection *conn,
5200 struct rx_packet *packet, int istack, int force)
5203 struct clock when, now;
5208 /* Clients should never delay abort messages */
5209 if (rx_IsClientConn(conn))
5212 if (force || rxi_connAbortThreshhold == 0
5213 || conn->abortCount < rxi_connAbortThreshhold) {
5215 rxevent_Cancel(&conn->delayedAbortEvent);
5216 error = htonl(conn->error);
5218 MUTEX_EXIT(&conn->conn_data_lock);
5220 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5221 RX_PACKET_TYPE_ABORT, (char *)&error,
5222 sizeof(error), istack);
5223 MUTEX_ENTER(&conn->conn_data_lock);
5224 } else if (!conn->delayedAbortEvent) {
5225 clock_GetTime(&now);
5227 clock_Addmsec(&when, rxi_connAbortDelay);
5228 conn->delayedAbortEvent =
5229 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
5234 /* Associate an error all of the calls owned by a connection. Called
5235 * with error non-zero. This is only for really fatal things, like
5236 * bad authentication responses. The connection itself is set in
5237 * error at this point, so that future packets received will be
5240 rxi_ConnectionError(struct rx_connection *conn,
5246 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5248 MUTEX_ENTER(&conn->conn_data_lock);
5249 rxevent_Cancel(&conn->challengeEvent);
5250 rxevent_Cancel(&conn->natKeepAliveEvent);
5251 if (conn->checkReachEvent) {
5252 rxevent_Cancel(&conn->checkReachEvent);
5253 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5254 putConnection(conn);
5256 MUTEX_EXIT(&conn->conn_data_lock);
5257 for (i = 0; i < RX_MAXCALLS; i++) {
5258 struct rx_call *call = conn->call[i];
5260 MUTEX_ENTER(&call->lock);
5261 rxi_CallError(call, error);
5262 MUTEX_EXIT(&call->lock);
5265 conn->error = error;
5266 if (rx_stats_active)
5267 rx_atomic_inc(&rx_stats.fatalErrors);
5272 * Interrupt an in-progress call with the specified error and wakeup waiters.
5274 * @param[in] call The call to interrupt
5275 * @param[in] error The error code to send to the peer
5278 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5280 MUTEX_ENTER(&call->lock);
5281 rxi_CallError(call, error);
5282 rxi_SendCallAbort(call, NULL, 0, 1);
5283 MUTEX_EXIT(&call->lock);
5287 rxi_CallError(struct rx_call *call, afs_int32 error)
5289 MUTEX_ASSERT(&call->lock);
5290 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5292 error = call->error;
5294 #ifdef RX_ENABLE_LOCKS
5295 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5296 rxi_ResetCall(call, 0);
5299 rxi_ResetCall(call, 0);
5301 call->error = error;
5304 /* Reset various fields in a call structure, and wakeup waiting
5305 * processes. Some fields aren't changed: state & mode are not
5306 * touched (these must be set by the caller), and bufptr, nLeft, and
5307 * nFree are not reset, since these fields are manipulated by
5308 * unprotected macros, and may only be reset by non-interrupting code.
5312 rxi_ResetCall(struct rx_call *call, int newcall)
5315 struct rx_peer *peer;
5316 struct rx_packet *packet;
5318 MUTEX_ASSERT(&call->lock);
5319 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5321 /* Notify anyone who is waiting for asynchronous packet arrival */
5322 if (call->arrivalProc) {
5323 (*call->arrivalProc) (call, call->arrivalProcHandle,
5324 call->arrivalProcArg);
5325 call->arrivalProc = (void (*)())0;
5329 rxi_CancelGrowMTUEvent(call);
5331 if (call->delayedAbortEvent) {
5332 rxi_CancelDelayedAbortEvent(call);
5333 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5335 rxi_SendCallAbort(call, packet, 0, 1);
5336 rxi_FreePacket(packet);
5341 * Update the peer with the congestion information in this call
5342 * so other calls on this connection can pick up where this call
5343 * left off. If the congestion sequence numbers don't match then
5344 * another call experienced a retransmission.
5346 peer = call->conn->peer;
5347 MUTEX_ENTER(&peer->peer_lock);
5349 if (call->congestSeq == peer->congestSeq) {
5350 peer->cwind = MAX(peer->cwind, call->cwind);
5351 peer->MTU = MAX(peer->MTU, call->MTU);
5352 peer->nDgramPackets =
5353 MAX(peer->nDgramPackets, call->nDgramPackets);
5356 call->abortCode = 0;
5357 call->abortCount = 0;
5359 if (peer->maxDgramPackets > 1) {
5360 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5362 call->MTU = peer->MTU;
5364 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5365 call->ssthresh = rx_maxSendWindow;
5366 call->nDgramPackets = peer->nDgramPackets;
5367 call->congestSeq = peer->congestSeq;
5368 call->rtt = peer->rtt;
5369 call->rtt_dev = peer->rtt_dev;
5370 clock_Zero(&call->rto);
5371 clock_Addmsec(&call->rto,
5372 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5373 MUTEX_EXIT(&peer->peer_lock);
5375 flags = call->flags;
5376 rxi_WaitforTQBusy(call);
5378 rxi_ClearTransmitQueue(call, 1);
5379 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5380 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5384 if (!newcall && (flags & RX_CALL_PEER_BUSY)) {
5385 /* The call channel is still busy; resetting the call doesn't change
5386 * that. However, if 'newcall' is set, we are processing a call
5387 * structure that has either been recycled from the free list, or has
5388 * been newly allocated. So, RX_CALL_PEER_BUSY is not relevant if
5389 * 'newcall' is set, since it describes a completely different call
5390 * channel which we do not care about. */
5391 call->flags |= RX_CALL_PEER_BUSY;
5394 rxi_ClearReceiveQueue(call);
5395 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5399 call->twind = call->conn->twind[call->channel];
5400 call->rwind = call->conn->rwind[call->channel];
5401 call->nSoftAcked = 0;
5402 call->nextCwind = 0;
5405 call->nCwindAcks = 0;
5406 call->nSoftAcks = 0;
5407 call->nHardAcks = 0;
5409 call->tfirst = call->rnext = call->tnext = 1;
5412 call->lastAcked = 0;
5413 call->localStatus = call->remoteStatus = 0;
5415 if (flags & RX_CALL_READER_WAIT) {
5416 #ifdef RX_ENABLE_LOCKS
5417 CV_BROADCAST(&call->cv_rq);
5419 osi_rxWakeup(&call->rq);
5422 if (flags & RX_CALL_WAIT_PACKETS) {
5423 MUTEX_ENTER(&rx_freePktQ_lock);
5424 rxi_PacketsUnWait(); /* XXX */
5425 MUTEX_EXIT(&rx_freePktQ_lock);
5427 #ifdef RX_ENABLE_LOCKS
5428 CV_SIGNAL(&call->cv_twind);
5430 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5431 osi_rxWakeup(&call->twind);
5434 if (flags & RX_CALL_WAIT_PROC) {
5435 rx_atomic_dec(&rx_nWaiting);
5437 #ifdef RX_ENABLE_LOCKS
5438 /* The following ensures that we don't mess with any queue while some
5439 * other thread might also be doing so. The call_queue_lock field is
5440 * is only modified under the call lock. If the call is in the process
5441 * of being removed from a queue, the call is not locked until the
5442 * the queue lock is dropped and only then is the call_queue_lock field
5443 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5444 * Note that any other routine which removes a call from a queue has to
5445 * obtain the queue lock before examing the queue and removing the call.
5447 if (call->call_queue_lock) {
5448 MUTEX_ENTER(call->call_queue_lock);
5449 if (opr_queue_IsOnQueue(&call->entry)) {
5450 opr_queue_Remove(&call->entry);
5452 MUTEX_EXIT(call->call_queue_lock);
5453 CLEAR_CALL_QUEUE_LOCK(call);
5455 #else /* RX_ENABLE_LOCKS */
5456 if (opr_queue_IsOnQueue(&call->entry)) {
5457 opr_queue_Remove(&call->entry);
5459 #endif /* RX_ENABLE_LOCKS */
5461 rxi_CancelKeepAliveEvent(call);
5462 rxi_CancelDelayedAckEvent(call);
5465 /* Send an acknowledge for the indicated packet (seq,serial) of the
5466 * indicated call, for the indicated reason (reason). This
5467 * acknowledge will specifically acknowledge receiving the packet, and
5468 * will also specify which other packets for this call have been
5469 * received. This routine returns the packet that was used to the
5470 * caller. The caller is responsible for freeing it or re-using it.
5471 * This acknowledgement also returns the highest sequence number
5472 * actually read out by the higher level to the sender; the sender
5473 * promises to keep around packets that have not been read by the
5474 * higher level yet (unless, of course, the sender decides to abort
5475 * the call altogether). Any of p, seq, serial, pflags, or reason may
5476 * be set to zero without ill effect. That is, if they are zero, they
5477 * will not convey any information.
5478 * NOW there is a trailer field, after the ack where it will safely be
5479 * ignored by mundanes, which indicates the maximum size packet this
5480 * host can swallow. */
5482 struct rx_packet *optionalPacket; use to send ack (or null)
5483 int seq; Sequence number of the packet we are acking
5484 int serial; Serial number of the packet
5485 int pflags; Flags field from packet header
5486 int reason; Reason an acknowledge was prompted
5490 rxi_SendAck(struct rx_call *call,
5491 struct rx_packet *optionalPacket, int serial, int reason,
5494 struct rx_ackPacket *ap;
5495 struct rx_packet *p;
5496 struct opr_queue *cursor;
5499 afs_uint32 padbytes = 0;
5500 #ifdef RX_ENABLE_TSFPQ
5501 struct rx_ts_info_t * rx_ts_info;
5505 * Open the receive window once a thread starts reading packets
5507 if (call->rnext > 1) {
5508 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5511 /* Don't attempt to grow MTU if this is a critical ping */
5512 if (reason == RX_ACK_MTU) {
5513 /* keep track of per-call attempts, if we're over max, do in small
5514 * otherwise in larger? set a size to increment by, decrease
5517 if (call->conn->peer->maxPacketSize &&
5518 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5520 padbytes = call->conn->peer->maxPacketSize+16;
5522 padbytes = call->conn->peer->maxMTU + 128;
5524 /* do always try a minimum size ping */
5525 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5527 /* subtract the ack payload */
5528 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5529 reason = RX_ACK_PING;
5532 call->nHardAcks = 0;
5533 call->nSoftAcks = 0;
5534 if (call->rnext > call->lastAcked)
5535 call->lastAcked = call->rnext;
5539 rx_computelen(p, p->length); /* reset length, you never know */
5540 } /* where that's been... */
5541 #ifdef RX_ENABLE_TSFPQ
5543 RX_TS_INFO_GET(rx_ts_info);
5544 if ((p = rx_ts_info->local_special_packet)) {
5545 rx_computelen(p, p->length);
5546 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5547 rx_ts_info->local_special_packet = p;
5548 } else { /* We won't send the ack, but don't panic. */
5549 return optionalPacket;
5553 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5554 /* We won't send the ack, but don't panic. */
5555 return optionalPacket;
5560 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5563 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5564 #ifndef RX_ENABLE_TSFPQ
5565 if (!optionalPacket)
5568 return optionalPacket;
5570 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5571 if (rx_Contiguous(p) < templ) {
5572 #ifndef RX_ENABLE_TSFPQ
5573 if (!optionalPacket)
5576 return optionalPacket;
5581 /* MTUXXX failing to send an ack is very serious. We should */
5582 /* try as hard as possible to send even a partial ack; it's */
5583 /* better than nothing. */
5584 ap = (struct rx_ackPacket *)rx_DataOf(p);
5585 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5586 ap->reason = reason;
5588 /* The skew computation used to be bogus, I think it's better now. */
5589 /* We should start paying attention to skew. XXX */
5590 ap->serial = htonl(serial);
5591 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5594 * First packet not yet forwarded to reader. When ACKALL has been
5595 * sent the peer has been told that all received packets will be
5596 * delivered to the reader. The value 'rnext' is used internally
5597 * to refer to the next packet in the receive queue that must be
5598 * delivered to the reader. From the perspective of the peer it
5599 * already has so report the last sequence number plus one if there
5600 * are packets in the receive queue awaiting processing.
5602 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5603 !opr_queue_IsEmpty(&call->rq)) {
5604 ap->firstPacket = htonl(opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq + 1);
5606 ap->firstPacket = htonl(call->rnext);
5608 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5610 /* No fear of running out of ack packet here because there can only
5611 * be at most one window full of unacknowledged packets. The window
5612 * size must be constrained to be less than the maximum ack size,
5613 * of course. Also, an ack should always fit into a single packet
5614 * -- it should not ever be fragmented. */
5616 for (opr_queue_Scan(&call->rq, cursor)) {
5617 struct rx_packet *rqp
5618 = opr_queue_Entry(cursor, struct rx_packet, entry);
5620 if (!rqp || !call->rq.next
5621 || (rqp->header.seq > (call->rnext + call->rwind))) {
5622 #ifndef RX_ENABLE_TSFPQ
5623 if (!optionalPacket)
5626 rxi_CallError(call, RX_CALL_DEAD);
5627 return optionalPacket;
5630 while (rqp->header.seq > call->rnext + offset)
5631 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5632 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5634 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5635 #ifndef RX_ENABLE_TSFPQ
5636 if (!optionalPacket)
5639 rxi_CallError(call, RX_CALL_DEAD);
5640 return optionalPacket;
5646 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5648 /* these are new for AFS 3.3 */
5649 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5650 templ = htonl(templ);
5651 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5652 templ = htonl(call->conn->peer->ifMTU);
5653 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5654 sizeof(afs_int32), &templ);
5656 /* new for AFS 3.4 */
5657 templ = htonl(call->rwind);
5658 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5659 sizeof(afs_int32), &templ);
5661 /* new for AFS 3.5 */
5662 templ = htonl(call->conn->peer->ifDgramPackets);
5663 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5664 sizeof(afs_int32), &templ);
5666 p->header.serviceId = call->conn->serviceId;
5667 p->header.cid = (call->conn->cid | call->channel);
5668 p->header.callNumber = *call->callNumber;
5670 p->header.securityIndex = call->conn->securityIndex;
5671 p->header.epoch = call->conn->epoch;
5672 p->header.type = RX_PACKET_TYPE_ACK;
5673 p->header.flags = RX_SLOW_START_OK;
5674 if (reason == RX_ACK_PING) {
5675 p->header.flags |= RX_REQUEST_ACK;
5677 p->length = padbytes +
5678 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5681 /* not fast but we can potentially use this if truncated
5682 * fragments are delivered to figure out the mtu.
5684 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5685 sizeof(afs_int32), sizeof(afs_int32),
5689 if (call->conn->type == RX_CLIENT_CONNECTION)
5690 p->header.flags |= RX_CLIENT_INITIATED;
5694 if (rxdebug_active) {
5698 len = _snprintf(msg, sizeof(msg),
5699 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5700 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5701 ntohl(ap->serial), ntohl(ap->previousPacket),
5702 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5703 ap->nAcks, ntohs(ap->bufferSpace) );
5707 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5708 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5712 OutputDebugString(msg);
5714 #else /* AFS_NT40_ENV */
5716 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5717 ap->reason, ntohl(ap->previousPacket),
5718 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5720 for (offset = 0; offset < ap->nAcks; offset++)
5721 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5726 #endif /* AFS_NT40_ENV */
5729 int i, nbytes = p->length;
5731 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5732 if (nbytes <= p->wirevec[i].iov_len) {
5735 savelen = p->wirevec[i].iov_len;
5737 p->wirevec[i].iov_len = nbytes;
5739 rxi_Send(call, p, istack);
5740 p->wirevec[i].iov_len = savelen;
5744 nbytes -= p->wirevec[i].iov_len;
5747 if (rx_stats_active)
5748 rx_atomic_inc(&rx_stats.ackPacketsSent);
5749 #ifndef RX_ENABLE_TSFPQ
5750 if (!optionalPacket)
5753 return optionalPacket; /* Return packet for re-use by caller */
5757 struct rx_packet **list;
5762 /* Send all of the packets in the list in single datagram */
5764 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5765 int istack, int moreFlag)
5771 struct rx_connection *conn = call->conn;
5772 struct rx_peer *peer = conn->peer;
5774 MUTEX_ENTER(&peer->peer_lock);
5775 peer->nSent += xmit->len;
5776 if (xmit->resending)
5777 peer->reSends += xmit->len;
5778 MUTEX_EXIT(&peer->peer_lock);
5780 if (rx_stats_active) {
5781 if (xmit->resending)
5782 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5784 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5787 clock_GetTime(&now);
5789 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5793 /* Set the packet flags and schedule the resend events */
5794 /* Only request an ack for the last packet in the list */
5795 for (i = 0; i < xmit->len; i++) {
5796 struct rx_packet *packet = xmit->list[i];
5798 /* Record the time sent */
5799 packet->timeSent = now;
5800 packet->flags |= RX_PKTFLAG_SENT;
5802 /* Ask for an ack on retransmitted packets, on every other packet
5803 * if the peer doesn't support slow start. Ask for an ack on every
5804 * packet until the congestion window reaches the ack rate. */
5805 if (packet->header.serial) {
5808 packet->firstSent = now;
5809 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5810 || (!(call->flags & RX_CALL_SLOW_START_OK)
5811 && (packet->header.seq & 1)))) {
5816 /* Tag this packet as not being the last in this group,
5817 * for the receiver's benefit */
5818 if (i < xmit->len - 1 || moreFlag) {
5819 packet->header.flags |= RX_MORE_PACKETS;
5824 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5827 /* Since we're about to send a data packet to the peer, it's
5828 * safe to nuke any scheduled end-of-packets ack */
5829 rxi_CancelDelayedAckEvent(call);
5831 MUTEX_EXIT(&call->lock);
5832 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5833 if (xmit->len > 1) {
5834 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5836 rxi_SendPacket(call, conn, xmit->list[0], istack);
5838 MUTEX_ENTER(&call->lock);
5839 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5841 /* Tell the RTO calculation engine that we have sent a packet, and
5842 * if it was the last one */
5843 rxi_rto_packet_sent(call, lastPacket, istack);
5845 /* Update last send time for this call (for keep-alive
5846 * processing), and for the connection (so that we can discover
5847 * idle connections) */
5848 conn->lastSendTime = call->lastSendTime = clock_Sec();
5849 /* Let a set of retransmits trigger an idle timeout */
5850 if (!xmit->resending)
5851 call->lastSendData = call->lastSendTime;
5854 /* When sending packets we need to follow these rules:
5855 * 1. Never send more than maxDgramPackets in a jumbogram.
5856 * 2. Never send a packet with more than two iovecs in a jumbogram.
5857 * 3. Never send a retransmitted packet in a jumbogram.
5858 * 4. Never send more than cwind/4 packets in a jumbogram
5859 * We always keep the last list we should have sent so we
5860 * can set the RX_MORE_PACKETS flags correctly.
5864 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5869 struct xmitlist working;
5870 struct xmitlist last;
5872 struct rx_peer *peer = call->conn->peer;
5873 int morePackets = 0;
5875 memset(&last, 0, sizeof(struct xmitlist));
5876 working.list = &list[0];
5878 working.resending = 0;
5880 recovery = call->flags & RX_CALL_FAST_RECOVER;
5882 for (i = 0; i < len; i++) {
5883 /* Does the current packet force us to flush the current list? */
5885 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5886 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5888 /* This sends the 'last' list and then rolls the current working
5889 * set into the 'last' one, and resets the working set */
5892 rxi_SendList(call, &last, istack, 1);
5893 /* If the call enters an error state stop sending, or if
5894 * we entered congestion recovery mode, stop sending */
5896 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5901 working.resending = 0;
5902 working.list = &list[i];
5904 /* Add the current packet to the list if it hasn't been acked.
5905 * Otherwise adjust the list pointer to skip the current packet. */
5906 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5909 if (list[i]->header.serial)
5910 working.resending = 1;
5912 /* Do we need to flush the list? */
5913 if (working.len >= (int)peer->maxDgramPackets
5914 || working.len >= (int)call->nDgramPackets
5915 || working.len >= (int)call->cwind
5916 || list[i]->header.serial
5917 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5919 rxi_SendList(call, &last, istack, 1);
5920 /* If the call enters an error state stop sending, or if
5921 * we entered congestion recovery mode, stop sending */
5923 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5928 working.resending = 0;
5929 working.list = &list[i + 1];
5932 if (working.len != 0) {
5933 osi_Panic("rxi_SendList error");
5935 working.list = &list[i + 1];
5939 /* Send the whole list when the call is in receive mode, when
5940 * the call is in eof mode, when we are in fast recovery mode,
5941 * and when we have the last packet */
5942 /* XXX - The accesses to app.mode aren't safe, as this may be called by
5943 * the listener or event threads
5945 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5946 || (call->flags & RX_CALL_FLUSH)
5947 || (call->flags & RX_CALL_FAST_RECOVER)) {
5948 /* Check for the case where the current list contains
5949 * an acked packet. Since we always send retransmissions
5950 * in a separate packet, we only need to check the first
5951 * packet in the list */
5952 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5956 rxi_SendList(call, &last, istack, morePackets);
5957 /* If the call enters an error state stop sending, or if
5958 * we entered congestion recovery mode, stop sending */
5960 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5964 rxi_SendList(call, &working, istack, 0);
5966 } else if (last.len > 0) {
5967 rxi_SendList(call, &last, istack, 0);
5968 /* Packets which are in 'working' are not sent by this call */
5973 * Check if the peer for the given call is known to be dead
5975 * If the call's peer appears dead (it has encountered fatal network errors
5976 * since the call started) the call is killed with RX_CALL_DEAD if the call
5977 * is active. Otherwise, we do nothing.
5979 * @param[in] call The call to check
5982 * @retval 0 The call is fine, and we haven't done anything to the call
5983 * @retval nonzero The call's peer appears dead, and the call has been
5984 * terminated if it was active
5986 * @pre call->lock must be locked
5989 rxi_CheckPeerDead(struct rx_call *call)
5991 #ifdef AFS_RXERRQ_ENV
5994 if (call->state == RX_STATE_DALLY) {
5998 peererrs = rx_atomic_read(&call->conn->peer->neterrs);
5999 if (call->neterr_gen < peererrs) {
6000 /* we have received network errors since this call started; kill
6002 if (call->state == RX_STATE_ACTIVE) {
6003 rxi_CallError(call, RX_CALL_DEAD);
6007 if (call->neterr_gen > peererrs) {
6008 /* someone has reset the number of peer errors; set the call error gen
6009 * so we can detect if more errors are encountered */
6010 call->neterr_gen = peererrs;
6017 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
6019 struct rx_call *call = arg0;
6020 struct rx_peer *peer;
6021 struct opr_queue *cursor;
6022 struct clock maxTimeout = { 60, 0 };
6024 MUTEX_ENTER(&call->lock);
6026 peer = call->conn->peer;
6028 /* Make sure that the event pointer is removed from the call
6029 * structure, since there is no longer a per-call retransmission
6031 if (event == call->resendEvent) {
6032 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
6033 rxevent_Put(&call->resendEvent);
6036 rxi_CheckPeerDead(call);
6038 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
6039 rxi_CheckBusy(call);
6042 if (opr_queue_IsEmpty(&call->tq)) {
6043 /* Nothing to do. This means that we've been raced, and that an
6044 * ACK has come in between when we were triggered, and when we
6045 * actually got to run. */
6049 /* We're in loss recovery */
6050 call->flags |= RX_CALL_FAST_RECOVER;
6052 /* Mark all of the pending packets in the queue as being lost */
6053 for (opr_queue_Scan(&call->tq, cursor)) {
6054 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
6055 if (!(p->flags & RX_PKTFLAG_ACKED))
6056 p->flags &= ~RX_PKTFLAG_SENT;
6059 /* We're resending, so we double the timeout of the call. This will be
6060 * dropped back down by the first successful ACK that we receive.
6062 * We apply a maximum value here of 60 seconds
6064 clock_Add(&call->rto, &call->rto);
6065 if (clock_Gt(&call->rto, &maxTimeout))
6066 call->rto = maxTimeout;
6068 /* Packet loss is most likely due to congestion, so drop our window size
6069 * and start again from the beginning */
6070 if (peer->maxDgramPackets >1) {
6071 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
6072 call->MTU = MIN(peer->natMTU, peer->maxMTU);
6074 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
6075 call->nDgramPackets = 1;
6077 call->nextCwind = 1;
6080 MUTEX_ENTER(&peer->peer_lock);
6081 peer->MTU = call->MTU;
6082 peer->cwind = call->cwind;
6083 peer->nDgramPackets = 1;
6085 call->congestSeq = peer->congestSeq;
6086 MUTEX_EXIT(&peer->peer_lock);
6088 rxi_Start(call, istack);
6091 MUTEX_EXIT(&call->lock);
6094 /* This routine is called when new packets are readied for
6095 * transmission and when retransmission may be necessary, or when the
6096 * transmission window or burst count are favourable. This should be
6097 * better optimized for new packets, the usual case, now that we've
6098 * got rid of queues of send packets. XXXXXXXXXXX */
6100 rxi_Start(struct rx_call *call, int istack)
6102 struct opr_queue *cursor;
6103 #ifdef RX_ENABLE_LOCKS
6104 struct opr_queue *store;
6110 #ifdef RX_ENABLE_LOCKS
6111 if (rx_stats_active)
6112 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
6117 if (!opr_queue_IsEmpty(&call->tq)) { /* If we have anything to send */
6118 /* Send (or resend) any packets that need it, subject to
6119 * window restrictions and congestion burst control
6120 * restrictions. Ask for an ack on the last packet sent in
6121 * this burst. For now, we're relying upon the window being
6122 * considerably bigger than the largest number of packets that
6123 * are typically sent at once by one initial call to
6124 * rxi_Start. This is probably bogus (perhaps we should ask
6125 * for an ack when we're half way through the current
6126 * window?). Also, for non file transfer applications, this
6127 * may end up asking for an ack for every packet. Bogus. XXXX
6130 * But check whether we're here recursively, and let the other guy
6133 #ifdef RX_ENABLE_LOCKS
6134 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6135 call->flags |= RX_CALL_TQ_BUSY;
6137 #endif /* RX_ENABLE_LOCKS */
6139 #ifdef RX_ENABLE_LOCKS
6140 call->flags &= ~RX_CALL_NEED_START;
6141 #endif /* RX_ENABLE_LOCKS */
6143 maxXmitPackets = MIN(call->twind, call->cwind);
6144 for (opr_queue_Scan(&call->tq, cursor)) {
6146 = opr_queue_Entry(cursor, struct rx_packet, entry);
6148 if (p->flags & RX_PKTFLAG_ACKED) {
6149 /* Since we may block, don't trust this */
6150 if (rx_stats_active)
6151 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6152 continue; /* Ignore this packet if it has been acknowledged */
6155 /* Turn off all flags except these ones, which are the same
6156 * on each transmission */
6157 p->header.flags &= RX_PRESET_FLAGS;
6159 if (p->header.seq >=
6160 call->tfirst + MIN((int)call->twind,
6161 (int)(call->nSoftAcked +
6163 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6164 /* Note: if we're waiting for more window space, we can
6165 * still send retransmits; hence we don't return here, but
6166 * break out to schedule a retransmit event */
6167 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6168 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6173 /* Transmit the packet if it needs to be sent. */
6174 if (!(p->flags & RX_PKTFLAG_SENT)) {
6175 if (nXmitPackets == maxXmitPackets) {
6176 rxi_SendXmitList(call, call->xmitList,
6177 nXmitPackets, istack);
6180 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6181 *(call->callNumber), p));
6182 call->xmitList[nXmitPackets++] = p;
6184 } /* end of the queue_Scan */
6186 /* xmitList now hold pointers to all of the packets that are
6187 * ready to send. Now we loop to send the packets */
6188 if (nXmitPackets > 0) {
6189 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6193 #ifdef RX_ENABLE_LOCKS
6195 /* We went into the error state while sending packets. Now is
6196 * the time to reset the call. This will also inform the using
6197 * process that the call is in an error state.
6199 if (rx_stats_active)
6200 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6201 call->flags &= ~RX_CALL_TQ_BUSY;
6202 rxi_WakeUpTransmitQueue(call);
6203 rxi_CallError(call, call->error);
6207 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6209 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6210 /* Some packets have received acks. If they all have, we can clear
6211 * the transmit queue.
6214 for (opr_queue_ScanSafe(&call->tq, cursor, store)) {
6216 = opr_queue_Entry(cursor, struct rx_packet, entry);
6218 if (p->header.seq < call->tfirst
6219 && (p->flags & RX_PKTFLAG_ACKED)) {
6220 opr_queue_Remove(&p->entry);
6221 #ifdef RX_TRACK_PACKETS
6222 p->flags &= ~RX_PKTFLAG_TQ;
6224 #ifdef RXDEBUG_PACKET
6232 call->flags |= RX_CALL_TQ_CLEARME;
6234 if (call->flags & RX_CALL_TQ_CLEARME)
6235 rxi_ClearTransmitQueue(call, 1);
6236 } while (call->flags & RX_CALL_NEED_START);
6238 * TQ references no longer protected by this flag; they must remain
6239 * protected by the call lock.
6241 call->flags &= ~RX_CALL_TQ_BUSY;
6242 rxi_WakeUpTransmitQueue(call);
6244 call->flags |= RX_CALL_NEED_START;
6246 #endif /* RX_ENABLE_LOCKS */
6248 rxi_rto_cancel(call);
6252 /* Also adjusts the keep alive parameters for the call, to reflect
6253 * that we have just sent a packet (so keep alives aren't sent
6256 rxi_Send(struct rx_call *call, struct rx_packet *p,
6259 struct rx_connection *conn = call->conn;
6261 /* Stamp each packet with the user supplied status */
6262 p->header.userStatus = call->localStatus;
6264 /* Allow the security object controlling this call's security to
6265 * make any last-minute changes to the packet */
6266 RXS_SendPacket(conn->securityObject, call, p);
6268 /* Since we're about to send SOME sort of packet to the peer, it's
6269 * safe to nuke any scheduled end-of-packets ack */
6270 rxi_CancelDelayedAckEvent(call);
6272 /* Actually send the packet, filling in more connection-specific fields */
6273 MUTEX_EXIT(&call->lock);
6274 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6275 rxi_SendPacket(call, conn, p, istack);
6276 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6277 MUTEX_ENTER(&call->lock);
6279 /* Update last send time for this call (for keep-alive
6280 * processing), and for the connection (so that we can discover
6281 * idle connections) */
6282 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6283 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6284 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6286 conn->lastSendTime = call->lastSendTime = clock_Sec();
6287 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6288 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6289 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6290 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6291 RX_ACK_PING_RESPONSE)))
6292 call->lastSendData = call->lastSendTime;
6296 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6297 * that things are fine. Also called periodically to guarantee that nothing
6298 * falls through the cracks (e.g. (error + dally) connections have keepalive
6299 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6301 * haveCTLock Set if calling from rxi_ReapConnections
6304 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6306 struct rx_connection *conn = call->conn;
6308 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6309 afs_uint32 fudgeFactor;
6312 int idle_timeout = 0;
6313 afs_int32 clock_diff = 0;
6315 if (rxi_CheckPeerDead(call)) {
6321 /* Large swings in the clock can have a significant impact on
6322 * the performance of RX call processing. Forward clock shifts
6323 * will result in premature event triggering or timeouts.
6324 * Backward shifts can result in calls not completing until
6325 * the clock catches up with the original start clock value.
6327 * If a backward clock shift of more than five minutes is noticed,
6328 * just fail the call.
6330 if (now < call->lastSendTime)
6331 clock_diff = call->lastSendTime - now;
6332 if (now < call->startWait)
6333 clock_diff = MAX(clock_diff, call->startWait - now);
6334 if (now < call->lastReceiveTime)
6335 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6336 if (clock_diff > 5 * 60)
6338 if (call->state == RX_STATE_ACTIVE)
6339 rxi_CallError(call, RX_CALL_TIMEOUT);
6343 #ifdef RX_ENABLE_LOCKS
6344 if (call->flags & RX_CALL_TQ_BUSY) {
6345 /* Call is active and will be reset by rxi_Start if it's
6346 * in an error state.
6351 /* RTT + 8*MDEV, rounded up to the next second. */
6352 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6353 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6355 deadTime = conn->secondsUntilDead + fudgeFactor;
6356 /* These are computed to the second (+- 1 second). But that's
6357 * good enough for these values, which should be a significant
6358 * number of seconds. */
6359 if (now > (call->lastReceiveTime + deadTime)) {
6360 if (call->state == RX_STATE_ACTIVE) {
6361 #ifdef AFS_ADAPT_PMTU
6362 # if defined(KERNEL) && defined(AFS_SUN5_ENV)
6364 # if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6365 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6366 ip_stack_t *ipst = ns->netstack_ip;
6368 ire = ire_cache_lookup(conn->peer->host
6369 # if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6371 # if defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID)
6373 # if defined(GLOBAL_NETSTACKID)
6380 if (ire && ire->ire_max_frag > 0)
6381 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6383 # if defined(GLOBAL_NETSTACKID)
6387 #endif /* AFS_ADAPT_PMTU */
6388 cerror = RX_CALL_DEAD;
6391 #ifdef RX_ENABLE_LOCKS
6392 /* Cancel pending events */
6393 rxi_CancelDelayedAckEvent(call);
6394 rxi_rto_cancel(call);
6395 rxi_CancelKeepAliveEvent(call);
6396 rxi_CancelGrowMTUEvent(call);
6397 MUTEX_ENTER(&rx_refcnt_mutex);
6398 /* if rxi_FreeCall returns 1 it has freed the call */
6399 if (call->refCount == 0 &&
6400 rxi_FreeCall(call, haveCTLock))
6402 MUTEX_EXIT(&rx_refcnt_mutex);
6405 MUTEX_EXIT(&rx_refcnt_mutex);
6407 #else /* RX_ENABLE_LOCKS */
6408 rxi_FreeCall(call, 0);
6410 #endif /* RX_ENABLE_LOCKS */
6412 /* Non-active calls are destroyed if they are not responding
6413 * to pings; active calls are simply flagged in error, so the
6414 * attached process can die reasonably gracefully. */
6417 if (conn->idleDeadDetection) {
6418 if (conn->idleDeadTime) {
6419 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6423 /* see if we have a non-activity timeout */
6424 if (call->startWait && ((call->startWait + idleDeadTime) < now) &&
6425 (call->flags & RX_CALL_READER_WAIT)) {
6426 if (call->state == RX_STATE_ACTIVE) {
6427 cerror = RX_CALL_TIMEOUT;
6432 if (call->lastSendData && ((call->lastSendData + idleDeadTime) < now)) {
6433 if (call->state == RX_STATE_ACTIVE) {
6434 cerror = conn->service ? conn->service->idleDeadErr : RX_CALL_IDLE;
6442 if (conn->hardDeadTime) {
6443 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6446 /* see if we have a hard timeout */
6448 && (now > (hardDeadTime + call->startTime.sec))) {
6449 if (call->state == RX_STATE_ACTIVE)
6450 rxi_CallError(call, RX_CALL_TIMEOUT);
6455 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6456 call->lastReceiveTime) {
6457 int oldMTU = conn->peer->ifMTU;
6459 /* if we thought we could send more, perhaps things got worse */
6460 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6461 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6462 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6463 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6465 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6467 /* minimum capped in SetPeerMtu */
6468 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6471 conn->lastPacketSize = 0;
6473 /* needed so ResetCall doesn't clobber us. */
6474 call->MTU = conn->peer->ifMTU;
6476 /* if we never succeeded, let the error pass out as-is */
6477 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6478 cerror = conn->msgsizeRetryErr;
6481 rxi_CallError(call, cerror);
6486 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6487 void *dummy, int dummy2)
6489 struct rx_connection *conn = arg1;
6490 struct rx_header theader;
6491 char tbuffer[1 + sizeof(struct rx_header)];
6492 struct sockaddr_in taddr;
6495 struct iovec tmpiov[2];
6498 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6501 tp = &tbuffer[sizeof(struct rx_header)];
6502 taddr.sin_family = AF_INET;
6503 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6504 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6505 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6506 taddr.sin_len = sizeof(struct sockaddr_in);
6508 memset(&theader, 0, sizeof(theader));
6509 theader.epoch = htonl(999);
6511 theader.callNumber = 0;
6514 theader.type = RX_PACKET_TYPE_VERSION;
6515 theader.flags = RX_LAST_PACKET;
6516 theader.serviceId = 0;
6518 memcpy(tbuffer, &theader, sizeof(theader));
6519 memcpy(tp, &a, sizeof(a));
6520 tmpiov[0].iov_base = tbuffer;
6521 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6523 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6525 MUTEX_ENTER(&conn->conn_data_lock);
6526 MUTEX_ENTER(&rx_refcnt_mutex);
6527 /* Only reschedule ourselves if the connection would not be destroyed */
6528 if (conn->refCount <= 1) {
6529 rxevent_Put(&conn->natKeepAliveEvent);
6530 MUTEX_EXIT(&rx_refcnt_mutex);
6531 MUTEX_EXIT(&conn->conn_data_lock);
6532 rx_DestroyConnection(conn); /* drop the reference for this */
6534 conn->refCount--; /* drop the reference for this */
6535 MUTEX_EXIT(&rx_refcnt_mutex);
6536 rxevent_Put(&conn->natKeepAliveEvent);
6537 rxi_ScheduleNatKeepAliveEvent(conn);
6538 MUTEX_EXIT(&conn->conn_data_lock);
6543 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6545 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6546 struct clock when, now;
6547 clock_GetTime(&now);
6549 when.sec += conn->secondsUntilNatPing;
6550 MUTEX_ENTER(&rx_refcnt_mutex);
6551 conn->refCount++; /* hold a reference for this */
6552 MUTEX_EXIT(&rx_refcnt_mutex);
6553 conn->natKeepAliveEvent =
6554 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6559 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6561 MUTEX_ENTER(&conn->conn_data_lock);
6562 conn->secondsUntilNatPing = seconds;
6564 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6565 rxi_ScheduleNatKeepAliveEvent(conn);
6567 conn->flags |= RX_CONN_NAT_PING;
6569 MUTEX_EXIT(&conn->conn_data_lock);
6572 /* When a call is in progress, this routine is called occasionally to
6573 * make sure that some traffic has arrived (or been sent to) the peer.
6574 * If nothing has arrived in a reasonable amount of time, the call is
6575 * declared dead; if nothing has been sent for a while, we send a
6576 * keep-alive packet (if we're actually trying to keep the call alive)
6579 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6582 struct rx_call *call = arg1;
6583 struct rx_connection *conn;
6586 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6587 MUTEX_ENTER(&call->lock);
6589 if (event == call->keepAliveEvent)
6590 rxevent_Put(&call->keepAliveEvent);
6594 if (rxi_CheckCall(call, 0)) {
6595 MUTEX_EXIT(&call->lock);
6599 /* Don't try to keep alive dallying calls */
6600 if (call->state == RX_STATE_DALLY) {
6601 MUTEX_EXIT(&call->lock);
6606 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6607 /* Don't try to send keepalives if there is unacknowledged data */
6608 /* the rexmit code should be good enough, this little hack
6609 * doesn't quite work XXX */
6610 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6612 rxi_ScheduleKeepAliveEvent(call);
6613 MUTEX_EXIT(&call->lock);
6616 /* Does what's on the nameplate. */
6618 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6620 struct rx_call *call = arg1;
6621 struct rx_connection *conn;
6623 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6624 MUTEX_ENTER(&call->lock);
6626 if (event == call->growMTUEvent)
6627 rxevent_Put(&call->growMTUEvent);
6629 if (rxi_CheckCall(call, 0)) {
6630 MUTEX_EXIT(&call->lock);
6634 /* Don't bother with dallying calls */
6635 if (call->state == RX_STATE_DALLY) {
6636 MUTEX_EXIT(&call->lock);
6643 * keep being scheduled, just don't do anything if we're at peak,
6644 * or we're not set up to be properly handled (idle timeout required)
6646 if ((conn->peer->maxPacketSize != 0) &&
6647 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6648 conn->idleDeadDetection)
6649 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6650 rxi_ScheduleGrowMTUEvent(call, 0);
6651 MUTEX_EXIT(&call->lock);
6655 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6657 if (!call->keepAliveEvent) {
6658 struct clock when, now;
6659 clock_GetTime(&now);
6661 when.sec += call->conn->secondsUntilPing;
6662 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6663 call->keepAliveEvent =
6664 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6669 rxi_CancelKeepAliveEvent(struct rx_call *call) {
6670 if (call->keepAliveEvent) {
6671 rxevent_Cancel(&call->keepAliveEvent);
6672 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6677 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6679 if (!call->growMTUEvent) {
6680 struct clock when, now;
6682 clock_GetTime(&now);
6685 if (call->conn->secondsUntilPing)
6686 secs = (6*call->conn->secondsUntilPing)-1;
6688 if (call->conn->secondsUntilDead)
6689 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6693 CALL_HOLD(call, RX_CALL_REFCOUNT_MTU);
6694 call->growMTUEvent =
6695 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6700 rxi_CancelGrowMTUEvent(struct rx_call *call)
6702 if (call->growMTUEvent) {
6703 rxevent_Cancel(&call->growMTUEvent);
6704 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6709 rxi_KeepAliveOn(struct rx_call *call)
6711 /* Pretend last packet received was received now--i.e. if another
6712 * packet isn't received within the keep alive time, then the call
6713 * will die; Initialize last send time to the current time--even
6714 * if a packet hasn't been sent yet. This will guarantee that a
6715 * keep-alive is sent within the ping time */
6716 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6717 rxi_ScheduleKeepAliveEvent(call);
6721 rx_KeepAliveOff(struct rx_call *call)
6723 MUTEX_ENTER(&call->lock);
6724 rxi_CancelKeepAliveEvent(call);
6725 MUTEX_EXIT(&call->lock);
6729 rx_KeepAliveOn(struct rx_call *call)
6731 MUTEX_ENTER(&call->lock);
6732 rxi_KeepAliveOn(call);
6733 MUTEX_EXIT(&call->lock);
6737 rxi_GrowMTUOn(struct rx_call *call)
6739 struct rx_connection *conn = call->conn;
6740 MUTEX_ENTER(&conn->conn_data_lock);
6741 conn->lastPingSizeSer = conn->lastPingSize = 0;
6742 MUTEX_EXIT(&conn->conn_data_lock);
6743 rxi_ScheduleGrowMTUEvent(call, 1);
6746 /* This routine is called to send connection abort messages
6747 * that have been delayed to throttle looping clients. */
6749 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6752 struct rx_connection *conn = arg1;
6755 struct rx_packet *packet;
6757 MUTEX_ENTER(&conn->conn_data_lock);
6758 rxevent_Put(&conn->delayedAbortEvent);
6759 error = htonl(conn->error);
6761 MUTEX_EXIT(&conn->conn_data_lock);
6762 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6765 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6766 RX_PACKET_TYPE_ABORT, (char *)&error,
6768 rxi_FreePacket(packet);
6772 /* This routine is called to send call abort messages
6773 * that have been delayed to throttle looping clients. */
6775 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6778 struct rx_call *call = arg1;
6781 struct rx_packet *packet;
6783 MUTEX_ENTER(&call->lock);
6784 rxevent_Put(&call->delayedAbortEvent);
6785 error = htonl(call->error);
6787 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6790 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6791 (char *)&error, sizeof(error), 0);
6792 rxi_FreePacket(packet);
6794 MUTEX_EXIT(&call->lock);
6795 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6798 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6799 * seconds) to ask the client to authenticate itself. The routine
6800 * issues a challenge to the client, which is obtained from the
6801 * security object associated with the connection */
6803 rxi_ChallengeEvent(struct rxevent *event,
6804 void *arg0, void *arg1, int tries)
6806 struct rx_connection *conn = arg0;
6809 rxevent_Put(&conn->challengeEvent);
6811 /* If there are no active calls it is not worth re-issuing the
6812 * challenge. If the client issues another call on this connection
6813 * the challenge can be requested at that time.
6815 if (!rxi_HasActiveCalls(conn))
6818 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6819 struct rx_packet *packet;
6820 struct clock when, now;
6823 /* We've failed to authenticate for too long.
6824 * Reset any calls waiting for authentication;
6825 * they are all in RX_STATE_PRECALL.
6829 MUTEX_ENTER(&conn->conn_call_lock);
6830 for (i = 0; i < RX_MAXCALLS; i++) {
6831 struct rx_call *call = conn->call[i];
6833 MUTEX_ENTER(&call->lock);
6834 if (call->state == RX_STATE_PRECALL) {
6835 rxi_CallError(call, RX_CALL_DEAD);
6836 rxi_SendCallAbort(call, NULL, 0, 0);
6838 MUTEX_EXIT(&call->lock);
6841 MUTEX_EXIT(&conn->conn_call_lock);
6845 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6847 /* If there's no packet available, do this later. */
6848 RXS_GetChallenge(conn->securityObject, conn, packet);
6849 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6850 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6851 rxi_FreePacket(packet);
6853 clock_GetTime(&now);
6855 when.sec += RX_CHALLENGE_TIMEOUT;
6856 conn->challengeEvent =
6857 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6862 /* Call this routine to start requesting the client to authenticate
6863 * itself. This will continue until authentication is established,
6864 * the call times out, or an invalid response is returned. The
6865 * security object associated with the connection is asked to create
6866 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6867 * defined earlier. */
6869 rxi_ChallengeOn(struct rx_connection *conn)
6871 if (!conn->challengeEvent) {
6872 RXS_CreateChallenge(conn->securityObject, conn);
6873 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6878 /* rxi_ComputeRoundTripTime is called with peer locked. */
6879 /* peer may be null */
6881 rxi_ComputeRoundTripTime(struct rx_packet *p,
6882 struct rx_ackPacket *ack,
6883 struct rx_call *call,
6884 struct rx_peer *peer,
6887 struct clock thisRtt, *sentp;
6891 /* If the ACK is delayed, then do nothing */
6892 if (ack->reason == RX_ACK_DELAY)
6895 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6896 * their RTT multiple times, so only include the RTT of the last packet
6898 if (p->flags & RX_JUMBO_PACKET)
6901 /* Use the serial number to determine which transmission the ACK is for,
6902 * and set the sent time to match this. If we have no serial number, then
6903 * only use the ACK for RTT calculations if the packet has not been
6907 serial = ntohl(ack->serial);
6909 if (serial == p->header.serial) {
6910 sentp = &p->timeSent;
6911 } else if (serial == p->firstSerial) {
6912 sentp = &p->firstSent;
6913 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6914 sentp = &p->firstSent;
6918 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6919 sentp = &p->firstSent;
6926 if (clock_Lt(&thisRtt, sentp))
6927 return; /* somebody set the clock back, don't count this time. */
6929 clock_Sub(&thisRtt, sentp);
6930 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6931 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6933 if (clock_IsZero(&thisRtt)) {
6935 * The actual round trip time is shorter than the
6936 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6937 * Since we can't tell which at the moment we will assume 1ms.
6939 thisRtt.usec = 1000;
6942 if (rx_stats_active) {
6943 MUTEX_ENTER(&rx_stats_mutex);
6944 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6945 rx_stats.minRtt = thisRtt;
6946 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6947 if (thisRtt.sec > 60) {
6948 MUTEX_EXIT(&rx_stats_mutex);
6949 return; /* somebody set the clock ahead */
6951 rx_stats.maxRtt = thisRtt;
6953 clock_Add(&rx_stats.totalRtt, &thisRtt);
6954 rx_atomic_inc(&rx_stats.nRttSamples);
6955 MUTEX_EXIT(&rx_stats_mutex);
6958 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6960 /* Apply VanJacobson round-trip estimations */
6965 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6966 * srtt is stored as fixed point with 3 bits after the binary
6967 * point (i.e., scaled by 8). The following magic is
6968 * equivalent to the smoothing algorithm in rfc793 with an
6969 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6970 * srtt'*8 = rtt + srtt*7
6971 * srtt'*8 = srtt*8 + rtt - srtt
6972 * srtt' = srtt + rtt/8 - srtt/8
6973 * srtt' = srtt + (rtt - srtt)/8
6976 delta = _8THMSEC(&thisRtt) - call->rtt;
6977 call->rtt += (delta >> 3);
6980 * We accumulate a smoothed rtt variance (actually, a smoothed
6981 * mean difference), then set the retransmit timer to smoothed
6982 * rtt + 4 times the smoothed variance (was 2x in van's original
6983 * paper, but 4x works better for me, and apparently for him as
6985 * rttvar is stored as
6986 * fixed point with 2 bits after the binary point (scaled by
6987 * 4). The following is equivalent to rfc793 smoothing with
6988 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6989 * rttvar'*4 = rttvar*3 + |delta|
6990 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6991 * rttvar' = rttvar + |delta|/4 - rttvar/4
6992 * rttvar' = rttvar + (|delta| - rttvar)/4
6993 * This replaces rfc793's wired-in beta.
6994 * dev*4 = dev*4 + (|actual - expected| - dev)
7000 delta -= (call->rtt_dev << 1);
7001 call->rtt_dev += (delta >> 3);
7003 /* I don't have a stored RTT so I start with this value. Since I'm
7004 * probably just starting a call, and will be pushing more data down
7005 * this, I expect congestion to increase rapidly. So I fudge a
7006 * little, and I set deviance to half the rtt. In practice,
7007 * deviance tends to approach something a little less than
7008 * half the smoothed rtt. */
7009 call->rtt = _8THMSEC(&thisRtt) + 8;
7010 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
7012 /* the smoothed RTT time is RTT + 4*MDEV
7014 * We allow a user specified minimum to be set for this, to allow clamping
7015 * at a minimum value in the same way as TCP. In addition, we have to allow
7016 * for the possibility that this packet is answered by a delayed ACK, so we
7017 * add on a fixed 200ms to account for that timer expiring.
7020 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
7021 rx_minPeerTimeout) + 200;
7022 clock_Zero(&call->rto);
7023 clock_Addmsec(&call->rto, rtt_timeout);
7025 /* Update the peer, so any new calls start with our values */
7026 peer->rtt_dev = call->rtt_dev;
7027 peer->rtt = call->rtt;
7029 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
7030 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
7034 /* Find all server connections that have not been active for a long time, and
7037 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
7040 struct clock now, when;
7041 struct rxevent *event;
7042 clock_GetTime(&now);
7044 /* Find server connection structures that haven't been used for
7045 * greater than rx_idleConnectionTime */
7047 struct rx_connection **conn_ptr, **conn_end;
7048 int i, havecalls = 0;
7049 MUTEX_ENTER(&rx_connHashTable_lock);
7050 for (conn_ptr = &rx_connHashTable[0], conn_end =
7051 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
7053 struct rx_connection *conn, *next;
7054 struct rx_call *call;
7058 for (conn = *conn_ptr; conn; conn = next) {
7059 /* XXX -- Shouldn't the connection be locked? */
7062 for (i = 0; i < RX_MAXCALLS; i++) {
7063 call = conn->call[i];
7067 code = MUTEX_TRYENTER(&call->lock);
7070 result = rxi_CheckCall(call, 1);
7071 MUTEX_EXIT(&call->lock);
7073 /* If CheckCall freed the call, it might
7074 * have destroyed the connection as well,
7075 * which screws up the linked lists.
7081 if (conn->type == RX_SERVER_CONNECTION) {
7082 /* This only actually destroys the connection if
7083 * there are no outstanding calls */
7084 MUTEX_ENTER(&conn->conn_data_lock);
7085 MUTEX_ENTER(&rx_refcnt_mutex);
7086 if (!havecalls && !conn->refCount
7087 && ((conn->lastSendTime + rx_idleConnectionTime) <
7089 conn->refCount++; /* it will be decr in rx_DestroyConn */
7090 MUTEX_EXIT(&rx_refcnt_mutex);
7091 MUTEX_EXIT(&conn->conn_data_lock);
7092 #ifdef RX_ENABLE_LOCKS
7093 rxi_DestroyConnectionNoLock(conn);
7094 #else /* RX_ENABLE_LOCKS */
7095 rxi_DestroyConnection(conn);
7096 #endif /* RX_ENABLE_LOCKS */
7098 #ifdef RX_ENABLE_LOCKS
7100 MUTEX_EXIT(&rx_refcnt_mutex);
7101 MUTEX_EXIT(&conn->conn_data_lock);
7103 #endif /* RX_ENABLE_LOCKS */
7107 #ifdef RX_ENABLE_LOCKS
7108 while (rx_connCleanup_list) {
7109 struct rx_connection *conn;
7110 conn = rx_connCleanup_list;
7111 rx_connCleanup_list = rx_connCleanup_list->next;
7112 MUTEX_EXIT(&rx_connHashTable_lock);
7113 rxi_CleanupConnection(conn);
7114 MUTEX_ENTER(&rx_connHashTable_lock);
7116 MUTEX_EXIT(&rx_connHashTable_lock);
7117 #endif /* RX_ENABLE_LOCKS */
7120 /* Find any peer structures that haven't been used (haven't had an
7121 * associated connection) for greater than rx_idlePeerTime */
7123 struct rx_peer **peer_ptr, **peer_end;
7127 * Why do we need to hold the rx_peerHashTable_lock across
7128 * the incrementing of peer_ptr since the rx_peerHashTable
7129 * array is not changing? We don't.
7131 * By dropping the lock periodically we can permit other
7132 * activities to be performed while a rxi_ReapConnections
7133 * call is in progress. The goal of reap connections
7134 * is to clean up quickly without causing large amounts
7135 * of contention. Therefore, it is important that global
7136 * mutexes not be held for extended periods of time.
7138 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7139 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7141 struct rx_peer *peer, *next, *prev;
7143 MUTEX_ENTER(&rx_peerHashTable_lock);
7144 for (prev = peer = *peer_ptr; peer; peer = next) {
7146 code = MUTEX_TRYENTER(&peer->peer_lock);
7147 if ((code) && (peer->refCount == 0)
7148 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7149 struct opr_queue *cursor, *store;
7153 * now know that this peer object is one to be
7154 * removed from the hash table. Once it is removed
7155 * it can't be referenced by other threads.
7156 * Lets remove it first and decrement the struct
7157 * nPeerStructs count.
7159 if (peer == *peer_ptr) {
7165 if (rx_stats_active)
7166 rx_atomic_dec(&rx_stats.nPeerStructs);
7169 * Now if we hold references on 'prev' and 'next'
7170 * we can safely drop the rx_peerHashTable_lock
7171 * while we destroy this 'peer' object.
7177 MUTEX_EXIT(&rx_peerHashTable_lock);
7179 MUTEX_EXIT(&peer->peer_lock);
7180 MUTEX_DESTROY(&peer->peer_lock);
7182 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7183 unsigned int num_funcs;
7184 struct rx_interface_stat *rpc_stat
7185 = opr_queue_Entry(cursor, struct rx_interface_stat,
7190 opr_queue_Remove(&rpc_stat->entry);
7191 opr_queue_Remove(&rpc_stat->entryPeers);
7193 num_funcs = rpc_stat->stats[0].func_total;
7195 sizeof(rx_interface_stat_t) +
7196 rpc_stat->stats[0].func_total *
7197 sizeof(rx_function_entry_v1_t);
7199 rxi_Free(rpc_stat, space);
7201 MUTEX_ENTER(&rx_rpc_stats);
7202 rxi_rpc_peer_stat_cnt -= num_funcs;
7203 MUTEX_EXIT(&rx_rpc_stats);
7208 * Regain the rx_peerHashTable_lock and
7209 * decrement the reference count on 'prev'
7212 MUTEX_ENTER(&rx_peerHashTable_lock);
7219 MUTEX_EXIT(&peer->peer_lock);
7224 MUTEX_EXIT(&rx_peerHashTable_lock);
7228 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7229 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7230 * GC, just below. Really, we shouldn't have to keep moving packets from
7231 * one place to another, but instead ought to always know if we can
7232 * afford to hold onto a packet in its particular use. */
7233 MUTEX_ENTER(&rx_freePktQ_lock);
7234 if (rx_waitingForPackets) {
7235 rx_waitingForPackets = 0;
7236 #ifdef RX_ENABLE_LOCKS
7237 CV_BROADCAST(&rx_waitingForPackets_cv);
7239 osi_rxWakeup(&rx_waitingForPackets);
7242 MUTEX_EXIT(&rx_freePktQ_lock);
7245 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7246 event = rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0);
7247 rxevent_Put(&event);
7251 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7252 * rx.h is sort of strange this is better. This is called with a security
7253 * object before it is discarded. Each connection using a security object has
7254 * its own refcount to the object so it won't actually be freed until the last
7255 * connection is destroyed.
7257 * This is the only rxs module call. A hold could also be written but no one
7261 rxs_Release(struct rx_securityClass *aobj)
7263 return RXS_Close(aobj);
7271 #define TRACE_OPTION_RX_DEBUG 16
7279 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7280 0, KEY_QUERY_VALUE, &parmKey);
7281 if (code != ERROR_SUCCESS)
7284 dummyLen = sizeof(TraceOption);
7285 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7286 (BYTE *) &TraceOption, &dummyLen);
7287 if (code == ERROR_SUCCESS) {
7288 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7290 RegCloseKey (parmKey);
7291 #endif /* AFS_NT40_ENV */
7296 rx_DebugOnOff(int on)
7300 rxdebug_active = on;
7306 rx_StatsOnOff(int on)
7308 rx_stats_active = on;
7312 /* Don't call this debugging routine directly; use dpf */
7314 rxi_DebugPrint(char *format, ...)
7323 va_start(ap, format);
7325 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7328 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7330 OutputDebugString(msg);
7336 va_start(ap, format);
7338 clock_GetTime(&now);
7339 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7340 (unsigned int)now.usec);
7341 vfprintf(rx_Log, format, ap);
7349 * This function is used to process the rx_stats structure that is local
7350 * to a process as well as an rx_stats structure received from a remote
7351 * process (via rxdebug). Therefore, it needs to do minimal version
7355 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7356 afs_int32 freePackets, char version)
7360 if (size != sizeof(struct rx_statistics)) {
7362 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7363 size, sizeof(struct rx_statistics));
7366 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7369 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7370 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7371 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7372 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7373 s->specialPktAllocFailures);
7375 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7376 s->receivePktAllocFailures, s->sendPktAllocFailures,
7377 s->specialPktAllocFailures);
7381 " greedy %u, " "bogusReads %u (last from host %x), "
7382 "noPackets %u, " "noBuffers %u, " "selects %u, "
7383 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7384 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7385 s->selects, s->sendSelects);
7387 fprintf(file, " packets read: ");
7388 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7389 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7391 fprintf(file, "\n");
7394 " other read counters: data %u, " "ack %u, " "dup %u "
7395 "spurious %u " "dally %u\n", s->dataPacketsRead,
7396 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7397 s->ignorePacketDally);
7399 fprintf(file, " packets sent: ");
7400 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7401 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7403 fprintf(file, "\n");
7406 " other send counters: ack %u, " "data %u (not resends), "
7407 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7408 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7409 s->dataPacketsPushed, s->ignoreAckedPacket);
7412 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7413 s->netSendFailures, (int)s->fatalErrors);
7415 if (s->nRttSamples) {
7416 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7417 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7419 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7420 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7424 " %d server connections, " "%d client connections, "
7425 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7426 s->nServerConns, s->nClientConns, s->nPeerStructs,
7427 s->nCallStructs, s->nFreeCallStructs);
7429 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7430 fprintf(file, " %d clock updates\n", clock_nUpdates);
7434 /* for backward compatibility */
7436 rx_PrintStats(FILE * file)
7438 MUTEX_ENTER(&rx_stats_mutex);
7439 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7440 sizeof(rx_stats), rx_nFreePackets,
7442 MUTEX_EXIT(&rx_stats_mutex);
7446 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7448 fprintf(file, "Peer %x.%d.\n",
7449 ntohl(peer->host), (int)ntohs(peer->port));
7452 " Rtt %d, " "total sent %d, " "resent %d\n",
7453 peer->rtt, peer->nSent, peer->reSends);
7455 fprintf(file, " Packet size %d\n", peer->ifMTU);
7459 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7461 * This mutex protects the following static variables:
7465 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7466 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7468 #define LOCK_RX_DEBUG
7469 #define UNLOCK_RX_DEBUG
7470 #endif /* AFS_PTHREAD_ENV */
7472 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7474 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7475 u_char type, void *inputData, size_t inputLength,
7476 void *outputData, size_t outputLength)
7478 static afs_int32 counter = 100;
7479 time_t waitTime, waitCount;
7480 struct rx_header theader;
7483 struct timeval tv_now, tv_wake, tv_delta;
7484 struct sockaddr_in taddr, faddr;
7498 tp = &tbuffer[sizeof(struct rx_header)];
7499 taddr.sin_family = AF_INET;
7500 taddr.sin_port = remotePort;
7501 taddr.sin_addr.s_addr = remoteAddr;
7502 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7503 taddr.sin_len = sizeof(struct sockaddr_in);
7506 memset(&theader, 0, sizeof(theader));
7507 theader.epoch = htonl(999);
7509 theader.callNumber = htonl(counter);
7512 theader.type = type;
7513 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7514 theader.serviceId = 0;
7516 memcpy(tbuffer, &theader, sizeof(theader));
7517 memcpy(tp, inputData, inputLength);
7519 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7520 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7522 /* see if there's a packet available */
7523 gettimeofday(&tv_wake, NULL);
7524 tv_wake.tv_sec += waitTime;
7527 FD_SET(socket, &imask);
7528 tv_delta.tv_sec = tv_wake.tv_sec;
7529 tv_delta.tv_usec = tv_wake.tv_usec;
7530 gettimeofday(&tv_now, NULL);
7532 if (tv_delta.tv_usec < tv_now.tv_usec) {
7534 tv_delta.tv_usec += 1000000;
7537 tv_delta.tv_usec -= tv_now.tv_usec;
7539 if (tv_delta.tv_sec < tv_now.tv_sec) {
7543 tv_delta.tv_sec -= tv_now.tv_sec;
7546 code = select(0, &imask, 0, 0, &tv_delta);
7547 #else /* AFS_NT40_ENV */
7548 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7549 #endif /* AFS_NT40_ENV */
7550 if (code == 1 && FD_ISSET(socket, &imask)) {
7551 /* now receive a packet */
7552 faddrLen = sizeof(struct sockaddr_in);
7554 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7555 (struct sockaddr *)&faddr, &faddrLen);
7558 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7559 if (counter == ntohl(theader.callNumber))
7567 /* see if we've timed out */
7575 code -= sizeof(struct rx_header);
7576 if (code > outputLength)
7577 code = outputLength;
7578 memcpy(outputData, tp, code);
7581 #endif /* RXDEBUG */
7584 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7585 afs_uint16 remotePort, struct rx_debugStats * stat,
7586 afs_uint32 * supportedValues)
7588 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7590 struct rx_debugIn in;
7592 *supportedValues = 0;
7593 in.type = htonl(RX_DEBUGI_GETSTATS);
7596 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7597 &in, sizeof(in), stat, sizeof(*stat));
7600 * If the call was successful, fixup the version and indicate
7601 * what contents of the stat structure are valid.
7602 * Also do net to host conversion of fields here.
7606 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7607 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7609 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7610 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7612 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7613 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7615 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7616 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7618 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7619 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7621 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7622 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7624 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7625 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7627 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7628 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7630 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7631 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7633 stat->nFreePackets = ntohl(stat->nFreePackets);
7634 stat->packetReclaims = ntohl(stat->packetReclaims);
7635 stat->callsExecuted = ntohl(stat->callsExecuted);
7636 stat->nWaiting = ntohl(stat->nWaiting);
7637 stat->idleThreads = ntohl(stat->idleThreads);
7638 stat->nWaited = ntohl(stat->nWaited);
7639 stat->nPackets = ntohl(stat->nPackets);
7648 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7649 afs_uint16 remotePort, struct rx_statistics * stat,
7650 afs_uint32 * supportedValues)
7652 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7654 struct rx_debugIn in;
7655 afs_int32 *lp = (afs_int32 *) stat;
7659 * supportedValues is currently unused, but added to allow future
7660 * versioning of this function.
7663 *supportedValues = 0;
7664 in.type = htonl(RX_DEBUGI_RXSTATS);
7666 memset(stat, 0, sizeof(*stat));
7668 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7669 &in, sizeof(in), stat, sizeof(*stat));
7674 * Do net to host conversion here
7677 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7688 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7689 afs_uint16 remotePort, size_t version_length,
7692 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7694 return MakeDebugCall(socket, remoteAddr, remotePort,
7695 RX_PACKET_TYPE_VERSION, a, 1, version,
7703 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7704 afs_uint16 remotePort, afs_int32 * nextConnection,
7705 int allConnections, afs_uint32 debugSupportedValues,
7706 struct rx_debugConn * conn,
7707 afs_uint32 * supportedValues)
7709 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7711 struct rx_debugIn in;
7715 * supportedValues is currently unused, but added to allow future
7716 * versioning of this function.
7719 *supportedValues = 0;
7720 if (allConnections) {
7721 in.type = htonl(RX_DEBUGI_GETALLCONN);
7723 in.type = htonl(RX_DEBUGI_GETCONN);
7725 in.index = htonl(*nextConnection);
7726 memset(conn, 0, sizeof(*conn));
7728 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7729 &in, sizeof(in), conn, sizeof(*conn));
7732 *nextConnection += 1;
7735 * Convert old connection format to new structure.
7738 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7739 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7740 #define MOVEvL(a) (conn->a = vL->a)
7742 /* any old or unrecognized version... */
7743 for (i = 0; i < RX_MAXCALLS; i++) {
7744 MOVEvL(callState[i]);
7745 MOVEvL(callMode[i]);
7746 MOVEvL(callFlags[i]);
7747 MOVEvL(callOther[i]);
7749 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7750 MOVEvL(secStats.type);
7751 MOVEvL(secStats.level);
7752 MOVEvL(secStats.flags);
7753 MOVEvL(secStats.expires);
7754 MOVEvL(secStats.packetsReceived);
7755 MOVEvL(secStats.packetsSent);
7756 MOVEvL(secStats.bytesReceived);
7757 MOVEvL(secStats.bytesSent);
7762 * Do net to host conversion here
7764 * I don't convert host or port since we are most likely
7765 * going to want these in NBO.
7767 conn->cid = ntohl(conn->cid);
7768 conn->serial = ntohl(conn->serial);
7769 for (i = 0; i < RX_MAXCALLS; i++) {
7770 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7772 conn->error = ntohl(conn->error);
7773 conn->secStats.flags = ntohl(conn->secStats.flags);
7774 conn->secStats.expires = ntohl(conn->secStats.expires);
7775 conn->secStats.packetsReceived =
7776 ntohl(conn->secStats.packetsReceived);
7777 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7778 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7779 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7780 conn->epoch = ntohl(conn->epoch);
7781 conn->natMTU = ntohl(conn->natMTU);
7790 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7791 afs_uint16 remotePort, afs_int32 * nextPeer,
7792 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7793 afs_uint32 * supportedValues)
7795 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7797 struct rx_debugIn in;
7800 * supportedValues is currently unused, but added to allow future
7801 * versioning of this function.
7804 *supportedValues = 0;
7805 in.type = htonl(RX_DEBUGI_GETPEER);
7806 in.index = htonl(*nextPeer);
7807 memset(peer, 0, sizeof(*peer));
7809 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7810 &in, sizeof(in), peer, sizeof(*peer));
7816 * Do net to host conversion here
7818 * I don't convert host or port since we are most likely
7819 * going to want these in NBO.
7821 peer->ifMTU = ntohs(peer->ifMTU);
7822 peer->idleWhen = ntohl(peer->idleWhen);
7823 peer->refCount = ntohs(peer->refCount);
7824 peer->rtt = ntohl(peer->rtt);
7825 peer->rtt_dev = ntohl(peer->rtt_dev);
7826 peer->timeout.sec = 0;
7827 peer->timeout.usec = 0;
7828 peer->nSent = ntohl(peer->nSent);
7829 peer->reSends = ntohl(peer->reSends);
7830 peer->natMTU = ntohs(peer->natMTU);
7831 peer->maxMTU = ntohs(peer->maxMTU);
7832 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7833 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7834 peer->MTU = ntohs(peer->MTU);
7835 peer->cwind = ntohs(peer->cwind);
7836 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7837 peer->congestSeq = ntohs(peer->congestSeq);
7838 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7839 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7840 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7841 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7850 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7851 struct rx_debugPeer * peerStats)
7854 afs_int32 error = 1; /* default to "did not succeed" */
7855 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7857 MUTEX_ENTER(&rx_peerHashTable_lock);
7858 for(tp = rx_peerHashTable[hashValue];
7859 tp != NULL; tp = tp->next) {
7860 if (tp->host == peerHost)
7866 MUTEX_EXIT(&rx_peerHashTable_lock);
7870 MUTEX_ENTER(&tp->peer_lock);
7871 peerStats->host = tp->host;
7872 peerStats->port = tp->port;
7873 peerStats->ifMTU = tp->ifMTU;
7874 peerStats->idleWhen = tp->idleWhen;
7875 peerStats->refCount = tp->refCount;
7876 peerStats->burstSize = 0;
7877 peerStats->burst = 0;
7878 peerStats->burstWait.sec = 0;
7879 peerStats->burstWait.usec = 0;
7880 peerStats->rtt = tp->rtt;
7881 peerStats->rtt_dev = tp->rtt_dev;
7882 peerStats->timeout.sec = 0;
7883 peerStats->timeout.usec = 0;
7884 peerStats->nSent = tp->nSent;
7885 peerStats->reSends = tp->reSends;
7886 peerStats->natMTU = tp->natMTU;
7887 peerStats->maxMTU = tp->maxMTU;
7888 peerStats->maxDgramPackets = tp->maxDgramPackets;
7889 peerStats->ifDgramPackets = tp->ifDgramPackets;
7890 peerStats->MTU = tp->MTU;
7891 peerStats->cwind = tp->cwind;
7892 peerStats->nDgramPackets = tp->nDgramPackets;
7893 peerStats->congestSeq = tp->congestSeq;
7894 peerStats->bytesSent.high = tp->bytesSent >> 32;
7895 peerStats->bytesSent.low = tp->bytesSent & MAX_AFS_UINT32;
7896 peerStats->bytesReceived.high = tp->bytesReceived >> 32;
7897 peerStats->bytesReceived.low
7898 = tp->bytesReceived & MAX_AFS_UINT32;
7899 MUTEX_EXIT(&tp->peer_lock);
7901 MUTEX_ENTER(&rx_peerHashTable_lock);
7904 MUTEX_EXIT(&rx_peerHashTable_lock);
7912 struct rx_serverQueueEntry *np;
7915 struct rx_call *call;
7916 struct rx_serverQueueEntry *sq;
7919 if (rx_atomic_test_and_set_bit(&rxinit_status, 0))
7920 return; /* Already shutdown. */
7924 #ifndef AFS_PTHREAD_ENV
7925 FD_ZERO(&rx_selectMask);
7926 #endif /* AFS_PTHREAD_ENV */
7927 rxi_dataQuota = RX_MAX_QUOTA;
7928 #ifndef AFS_PTHREAD_ENV
7930 #endif /* AFS_PTHREAD_ENV */
7933 #ifndef AFS_PTHREAD_ENV
7934 #ifndef AFS_USE_GETTIMEOFDAY
7936 #endif /* AFS_USE_GETTIMEOFDAY */
7937 #endif /* AFS_PTHREAD_ENV */
7939 while (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
7940 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
7941 opr_queue_Remove(&call->entry);
7942 rxi_Free(call, sizeof(struct rx_call));
7945 while (!opr_queue_IsEmpty(&rx_idleServerQueue)) {
7946 sq = opr_queue_First(&rx_idleServerQueue, struct rx_serverQueueEntry,
7948 opr_queue_Remove(&sq->entry);
7953 struct rx_peer **peer_ptr, **peer_end;
7954 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7955 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7957 struct rx_peer *peer, *next;
7959 MUTEX_ENTER(&rx_peerHashTable_lock);
7960 for (peer = *peer_ptr; peer; peer = next) {
7961 struct opr_queue *cursor, *store;
7964 MUTEX_ENTER(&rx_rpc_stats);
7965 MUTEX_ENTER(&peer->peer_lock);
7966 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7967 unsigned int num_funcs;
7968 struct rx_interface_stat *rpc_stat
7969 = opr_queue_Entry(cursor, struct rx_interface_stat,
7973 opr_queue_Remove(&rpc_stat->entry);
7974 opr_queue_Remove(&rpc_stat->entryPeers);
7975 num_funcs = rpc_stat->stats[0].func_total;
7977 sizeof(rx_interface_stat_t) +
7978 rpc_stat->stats[0].func_total *
7979 sizeof(rx_function_entry_v1_t);
7981 rxi_Free(rpc_stat, space);
7983 /* rx_rpc_stats must be held */
7984 rxi_rpc_peer_stat_cnt -= num_funcs;
7986 MUTEX_EXIT(&peer->peer_lock);
7987 MUTEX_EXIT(&rx_rpc_stats);
7991 if (rx_stats_active)
7992 rx_atomic_dec(&rx_stats.nPeerStructs);
7994 MUTEX_EXIT(&rx_peerHashTable_lock);
7997 for (i = 0; i < RX_MAX_SERVICES; i++) {
7999 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
8001 for (i = 0; i < rx_hashTableSize; i++) {
8002 struct rx_connection *tc, *ntc;
8003 MUTEX_ENTER(&rx_connHashTable_lock);
8004 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
8006 for (j = 0; j < RX_MAXCALLS; j++) {
8008 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
8011 rxi_Free(tc, sizeof(*tc));
8013 MUTEX_EXIT(&rx_connHashTable_lock);
8016 MUTEX_ENTER(&freeSQEList_lock);
8018 while ((np = rx_FreeSQEList)) {
8019 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
8020 MUTEX_DESTROY(&np->lock);
8021 rxi_Free(np, sizeof(*np));
8024 MUTEX_EXIT(&freeSQEList_lock);
8025 MUTEX_DESTROY(&freeSQEList_lock);
8026 MUTEX_DESTROY(&rx_freeCallQueue_lock);
8027 MUTEX_DESTROY(&rx_connHashTable_lock);
8028 MUTEX_DESTROY(&rx_peerHashTable_lock);
8029 MUTEX_DESTROY(&rx_serverPool_lock);
8031 osi_Free(rx_connHashTable,
8032 rx_hashTableSize * sizeof(struct rx_connection *));
8033 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8035 UNPIN(rx_connHashTable,
8036 rx_hashTableSize * sizeof(struct rx_connection *));
8037 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8039 rxi_FreeAllPackets();
8041 MUTEX_ENTER(&rx_quota_mutex);
8042 rxi_dataQuota = RX_MAX_QUOTA;
8043 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
8044 MUTEX_EXIT(&rx_quota_mutex);
8050 * Routines to implement connection specific data.
8054 rx_KeyCreate(rx_destructor_t rtn)
8057 MUTEX_ENTER(&rxi_keyCreate_lock);
8058 key = rxi_keyCreate_counter++;
8059 rxi_keyCreate_destructor = (rx_destructor_t *)
8060 realloc((void *)rxi_keyCreate_destructor,
8061 (key + 1) * sizeof(rx_destructor_t));
8062 rxi_keyCreate_destructor[key] = rtn;
8063 MUTEX_EXIT(&rxi_keyCreate_lock);
8068 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
8071 MUTEX_ENTER(&conn->conn_data_lock);
8072 if (!conn->specific) {
8073 conn->specific = malloc((key + 1) * sizeof(void *));
8074 for (i = 0; i < key; i++)
8075 conn->specific[i] = NULL;
8076 conn->nSpecific = key + 1;
8077 conn->specific[key] = ptr;
8078 } else if (key >= conn->nSpecific) {
8079 conn->specific = (void **)
8080 realloc(conn->specific, (key + 1) * sizeof(void *));
8081 for (i = conn->nSpecific; i < key; i++)
8082 conn->specific[i] = NULL;
8083 conn->nSpecific = key + 1;
8084 conn->specific[key] = ptr;
8086 if (conn->specific[key] && rxi_keyCreate_destructor[key])
8087 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
8088 conn->specific[key] = ptr;
8090 MUTEX_EXIT(&conn->conn_data_lock);
8094 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
8097 MUTEX_ENTER(&svc->svc_data_lock);
8098 if (!svc->specific) {
8099 svc->specific = malloc((key + 1) * sizeof(void *));
8100 for (i = 0; i < key; i++)
8101 svc->specific[i] = NULL;
8102 svc->nSpecific = key + 1;
8103 svc->specific[key] = ptr;
8104 } else if (key >= svc->nSpecific) {
8105 svc->specific = (void **)
8106 realloc(svc->specific, (key + 1) * sizeof(void *));
8107 for (i = svc->nSpecific; i < key; i++)
8108 svc->specific[i] = NULL;
8109 svc->nSpecific = key + 1;
8110 svc->specific[key] = ptr;
8112 if (svc->specific[key] && rxi_keyCreate_destructor[key])
8113 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
8114 svc->specific[key] = ptr;
8116 MUTEX_EXIT(&svc->svc_data_lock);
8120 rx_GetSpecific(struct rx_connection *conn, int key)
8123 MUTEX_ENTER(&conn->conn_data_lock);
8124 if (key >= conn->nSpecific)
8127 ptr = conn->specific[key];
8128 MUTEX_EXIT(&conn->conn_data_lock);
8133 rx_GetServiceSpecific(struct rx_service *svc, int key)
8136 MUTEX_ENTER(&svc->svc_data_lock);
8137 if (key >= svc->nSpecific)
8140 ptr = svc->specific[key];
8141 MUTEX_EXIT(&svc->svc_data_lock);
8146 #endif /* !KERNEL */
8149 * processStats is a queue used to store the statistics for the local
8150 * process. Its contents are similar to the contents of the rpcStats
8151 * queue on a rx_peer structure, but the actual data stored within
8152 * this queue contains totals across the lifetime of the process (assuming
8153 * the stats have not been reset) - unlike the per peer structures
8154 * which can come and go based upon the peer lifetime.
8157 static struct opr_queue processStats = { &processStats, &processStats };
8160 * peerStats is a queue used to store the statistics for all peer structs.
8161 * Its contents are the union of all the peer rpcStats queues.
8164 static struct opr_queue peerStats = { &peerStats, &peerStats };
8167 * rxi_monitor_processStats is used to turn process wide stat collection
8171 static int rxi_monitor_processStats = 0;
8174 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8177 static int rxi_monitor_peerStats = 0;
8181 rxi_ClearRPCOpStat(rx_function_entry_v1_p rpc_stat)
8183 rpc_stat->invocations = 0;
8184 rpc_stat->bytes_sent = 0;
8185 rpc_stat->bytes_rcvd = 0;
8186 rpc_stat->queue_time_sum.sec = 0;
8187 rpc_stat->queue_time_sum.usec = 0;
8188 rpc_stat->queue_time_sum_sqr.sec = 0;
8189 rpc_stat->queue_time_sum_sqr.usec = 0;
8190 rpc_stat->queue_time_min.sec = 9999999;
8191 rpc_stat->queue_time_min.usec = 9999999;
8192 rpc_stat->queue_time_max.sec = 0;
8193 rpc_stat->queue_time_max.usec = 0;
8194 rpc_stat->execution_time_sum.sec = 0;
8195 rpc_stat->execution_time_sum.usec = 0;
8196 rpc_stat->execution_time_sum_sqr.sec = 0;
8197 rpc_stat->execution_time_sum_sqr.usec = 0;
8198 rpc_stat->execution_time_min.sec = 9999999;
8199 rpc_stat->execution_time_min.usec = 9999999;
8200 rpc_stat->execution_time_max.sec = 0;
8201 rpc_stat->execution_time_max.usec = 0;
8205 * Given all of the information for a particular rpc
8206 * call, find or create (if requested) the stat structure for the rpc.
8209 * the queue of stats that will be updated with the new value
8211 * @param rxInterface
8212 * a unique number that identifies the rpc interface
8215 * the total number of functions in this interface. this is only
8216 * required if create is true
8219 * if true, this invocation was made to a server
8222 * the ip address of the remote host. this is only required if create
8223 * and addToPeerList are true
8226 * the port of the remote host. this is only required if create
8227 * and addToPeerList are true
8229 * @param addToPeerList
8230 * if != 0, add newly created stat to the global peer list
8233 * if a new stats structure is allocated, the counter will
8234 * be updated with the new number of allocated stat structures.
8235 * only required if create is true
8238 * if no stats structure exists, allocate one
8242 static rx_interface_stat_p
8243 rxi_FindRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8244 afs_uint32 totalFunc, int isServer, afs_uint32 remoteHost,
8245 afs_uint32 remotePort, int addToPeerList,
8246 unsigned int *counter, int create)
8248 rx_interface_stat_p rpc_stat = NULL;
8249 struct opr_queue *cursor;
8252 * See if there's already a structure for this interface
8255 for (opr_queue_Scan(stats, cursor)) {
8256 rpc_stat = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8258 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8259 && (rpc_stat->stats[0].remote_is_server == isServer))
8263 /* if they didn't ask us to create, we're done */
8265 if (opr_queue_IsEnd(stats, cursor))
8271 /* can't proceed without these */
8272 if (!totalFunc || !counter)
8276 * Didn't find a match so allocate a new structure and add it to the
8280 if (opr_queue_IsEnd(stats, cursor) || (rpc_stat == NULL)
8281 || (rpc_stat->stats[0].interfaceId != rxInterface)
8282 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8287 sizeof(rx_interface_stat_t) +
8288 totalFunc * sizeof(rx_function_entry_v1_t);
8290 rpc_stat = rxi_Alloc(space);
8291 if (rpc_stat == NULL)
8294 *counter += totalFunc;
8295 for (i = 0; i < totalFunc; i++) {
8296 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8297 rpc_stat->stats[i].remote_peer = remoteHost;
8298 rpc_stat->stats[i].remote_port = remotePort;
8299 rpc_stat->stats[i].remote_is_server = isServer;
8300 rpc_stat->stats[i].interfaceId = rxInterface;
8301 rpc_stat->stats[i].func_total = totalFunc;
8302 rpc_stat->stats[i].func_index = i;
8304 opr_queue_Prepend(stats, &rpc_stat->entry);
8305 if (addToPeerList) {
8306 opr_queue_Prepend(&peerStats, &rpc_stat->entryPeers);
8313 rx_ClearProcessRPCStats(afs_int32 rxInterface)
8315 rx_interface_stat_p rpc_stat;
8318 if (rxInterface == -1)
8321 MUTEX_ENTER(&rx_rpc_stats);
8322 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8325 totalFunc = rpc_stat->stats[0].func_total;
8326 for (i = 0; i < totalFunc; i++)
8327 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8329 MUTEX_EXIT(&rx_rpc_stats);
8334 rx_ClearPeerRPCStats(afs_int32 rxInterface, afs_uint32 peerHost, afs_uint16 peerPort)
8336 rx_interface_stat_p rpc_stat;
8338 struct rx_peer * peer;
8340 if (rxInterface == -1)
8343 peer = rxi_FindPeer(peerHost, peerPort, 0);
8347 MUTEX_ENTER(&rx_rpc_stats);
8348 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8351 totalFunc = rpc_stat->stats[0].func_total;
8352 for (i = 0; i < totalFunc; i++)
8353 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8355 MUTEX_EXIT(&rx_rpc_stats);
8360 rx_CopyProcessRPCStats(afs_uint64 op)
8362 rx_interface_stat_p rpc_stat;
8363 rx_function_entry_v1_p rpcop_stat =
8364 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8365 int currentFunc = (op & MAX_AFS_UINT32);
8366 afs_int32 rxInterface = (op >> 32);
8368 if (!rxi_monitor_processStats)
8371 if (rxInterface == -1)
8374 if (rpcop_stat == NULL)
8377 MUTEX_ENTER(&rx_rpc_stats);
8378 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8381 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8382 sizeof(rx_function_entry_v1_t));
8383 MUTEX_EXIT(&rx_rpc_stats);
8385 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8392 rx_CopyPeerRPCStats(afs_uint64 op, afs_uint32 peerHost, afs_uint16 peerPort)
8394 rx_interface_stat_p rpc_stat;
8395 rx_function_entry_v1_p rpcop_stat =
8396 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8397 int currentFunc = (op & MAX_AFS_UINT32);
8398 afs_int32 rxInterface = (op >> 32);
8399 struct rx_peer *peer;
8401 if (!rxi_monitor_peerStats)
8404 if (rxInterface == -1)
8407 if (rpcop_stat == NULL)
8410 peer = rxi_FindPeer(peerHost, peerPort, 0);
8414 MUTEX_ENTER(&rx_rpc_stats);
8415 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8418 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8419 sizeof(rx_function_entry_v1_t));
8420 MUTEX_EXIT(&rx_rpc_stats);
8422 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8429 rx_ReleaseRPCStats(void *stats)
8432 rxi_Free(stats, sizeof(rx_function_entry_v1_t));
8436 * Given all of the information for a particular rpc
8437 * call, create (if needed) and update the stat totals for the rpc.
8440 * the queue of stats that will be updated with the new value
8442 * @param rxInterface
8443 * a unique number that identifies the rpc interface
8445 * @param currentFunc
8446 * the index of the function being invoked
8449 * the total number of functions in this interface
8452 * the amount of time this function waited for a thread
8455 * the amount of time this function invocation took to execute
8458 * the number bytes sent by this invocation
8461 * the number bytes received by this invocation
8464 * if true, this invocation was made to a server
8467 * the ip address of the remote host
8470 * the port of the remote host
8472 * @param addToPeerList
8473 * if != 0, add newly created stat to the global peer list
8476 * if a new stats structure is allocated, the counter will
8477 * be updated with the new number of allocated stat structures
8482 rxi_AddRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8483 afs_uint32 currentFunc, afs_uint32 totalFunc,
8484 struct clock *queueTime, struct clock *execTime,
8485 afs_uint64 bytesSent, afs_uint64 bytesRcvd, int isServer,
8486 afs_uint32 remoteHost, afs_uint32 remotePort,
8487 int addToPeerList, unsigned int *counter)
8490 rx_interface_stat_p rpc_stat;
8492 rpc_stat = rxi_FindRpcStat(stats, rxInterface, totalFunc, isServer,
8493 remoteHost, remotePort, addToPeerList, counter,
8501 * Increment the stats for this function
8504 rpc_stat->stats[currentFunc].invocations++;
8505 rpc_stat->stats[currentFunc].bytes_sent += bytesSent;
8506 rpc_stat->stats[currentFunc].bytes_rcvd += bytesRcvd;
8507 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8508 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8509 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8510 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8512 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8513 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8515 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8516 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8518 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8519 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8521 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8522 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8530 rxi_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8531 afs_uint32 currentFunc, afs_uint32 totalFunc,
8532 struct clock *queueTime, struct clock *execTime,
8533 afs_uint64 bytesSent, afs_uint64 bytesRcvd,
8537 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8540 MUTEX_ENTER(&rx_rpc_stats);
8542 if (rxi_monitor_peerStats) {
8543 MUTEX_ENTER(&peer->peer_lock);
8544 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8545 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8546 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8547 MUTEX_EXIT(&peer->peer_lock);
8550 if (rxi_monitor_processStats) {
8551 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8552 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8553 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8556 MUTEX_EXIT(&rx_rpc_stats);
8560 * Increment the times and count for a particular rpc function.
8562 * Traditionally this call was invoked from rxgen stubs. Modern stubs
8563 * call rx_RecordCallStatistics instead, so the public version of this
8564 * function is left purely for legacy callers.
8567 * The peer who invoked the rpc
8569 * @param rxInterface
8570 * A unique number that identifies the rpc interface
8572 * @param currentFunc
8573 * The index of the function being invoked
8576 * The total number of functions in this interface
8579 * The amount of time this function waited for a thread
8582 * The amount of time this function invocation took to execute
8585 * The number bytes sent by this invocation
8588 * The number bytes received by this invocation
8591 * If true, this invocation was made to a server
8595 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8596 afs_uint32 currentFunc, afs_uint32 totalFunc,
8597 struct clock *queueTime, struct clock *execTime,
8598 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8604 sent64 = ((afs_uint64)bytesSent->high << 32) + bytesSent->low;
8605 rcvd64 = ((afs_uint64)bytesRcvd->high << 32) + bytesRcvd->low;
8607 rxi_IncrementTimeAndCount(peer, rxInterface, currentFunc, totalFunc,
8608 queueTime, execTime, sent64, rcvd64,
8615 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8619 * IN callerVersion - the rpc stat version of the caller.
8621 * IN count - the number of entries to marshall.
8623 * IN stats - pointer to stats to be marshalled.
8625 * OUT ptr - Where to store the marshalled data.
8632 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8633 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8639 * We only support the first version
8641 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8642 *(ptr++) = stats->remote_peer;
8643 *(ptr++) = stats->remote_port;
8644 *(ptr++) = stats->remote_is_server;
8645 *(ptr++) = stats->interfaceId;
8646 *(ptr++) = stats->func_total;
8647 *(ptr++) = stats->func_index;
8648 *(ptr++) = stats->invocations >> 32;
8649 *(ptr++) = stats->invocations & MAX_AFS_UINT32;
8650 *(ptr++) = stats->bytes_sent >> 32;
8651 *(ptr++) = stats->bytes_sent & MAX_AFS_UINT32;
8652 *(ptr++) = stats->bytes_rcvd >> 32;
8653 *(ptr++) = stats->bytes_rcvd & MAX_AFS_UINT32;
8654 *(ptr++) = stats->queue_time_sum.sec;
8655 *(ptr++) = stats->queue_time_sum.usec;
8656 *(ptr++) = stats->queue_time_sum_sqr.sec;
8657 *(ptr++) = stats->queue_time_sum_sqr.usec;
8658 *(ptr++) = stats->queue_time_min.sec;
8659 *(ptr++) = stats->queue_time_min.usec;
8660 *(ptr++) = stats->queue_time_max.sec;
8661 *(ptr++) = stats->queue_time_max.usec;
8662 *(ptr++) = stats->execution_time_sum.sec;
8663 *(ptr++) = stats->execution_time_sum.usec;
8664 *(ptr++) = stats->execution_time_sum_sqr.sec;
8665 *(ptr++) = stats->execution_time_sum_sqr.usec;
8666 *(ptr++) = stats->execution_time_min.sec;
8667 *(ptr++) = stats->execution_time_min.usec;
8668 *(ptr++) = stats->execution_time_max.sec;
8669 *(ptr++) = stats->execution_time_max.usec;
8675 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8680 * IN callerVersion - the rpc stat version of the caller
8682 * OUT myVersion - the rpc stat version of this function
8684 * OUT clock_sec - local time seconds
8686 * OUT clock_usec - local time microseconds
8688 * OUT allocSize - the number of bytes allocated to contain stats
8690 * OUT statCount - the number stats retrieved from this process.
8692 * OUT stats - the actual stats retrieved from this process.
8696 * Returns void. If successful, stats will != NULL.
8700 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8701 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8702 size_t * allocSize, afs_uint32 * statCount,
8703 afs_uint32 ** stats)
8713 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8716 * Check to see if stats are enabled
8719 MUTEX_ENTER(&rx_rpc_stats);
8720 if (!rxi_monitor_processStats) {
8721 MUTEX_EXIT(&rx_rpc_stats);
8725 clock_GetTime(&now);
8726 *clock_sec = now.sec;
8727 *clock_usec = now.usec;
8730 * Allocate the space based upon the caller version
8732 * If the client is at an older version than we are,
8733 * we return the statistic data in the older data format, but
8734 * we still return our version number so the client knows we
8735 * are maintaining more data than it can retrieve.
8738 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8739 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8740 *statCount = rxi_rpc_process_stat_cnt;
8743 * This can't happen yet, but in the future version changes
8744 * can be handled by adding additional code here
8748 if (space > (size_t) 0) {
8750 ptr = *stats = rxi_Alloc(space);
8753 struct opr_queue *cursor;
8755 for (opr_queue_Scan(&processStats, cursor)) {
8756 struct rx_interface_stat *rpc_stat =
8757 opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8759 * Copy the data based upon the caller version
8761 rx_MarshallProcessRPCStats(callerVersion,
8762 rpc_stat->stats[0].func_total,
8763 rpc_stat->stats, &ptr);
8769 MUTEX_EXIT(&rx_rpc_stats);
8774 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8778 * IN callerVersion - the rpc stat version of the caller
8780 * OUT myVersion - the rpc stat version of this function
8782 * OUT clock_sec - local time seconds
8784 * OUT clock_usec - local time microseconds
8786 * OUT allocSize - the number of bytes allocated to contain stats
8788 * OUT statCount - the number of stats retrieved from the individual
8791 * OUT stats - the actual stats retrieved from the individual peer structures.
8795 * Returns void. If successful, stats will != NULL.
8799 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8800 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8801 size_t * allocSize, afs_uint32 * statCount,
8802 afs_uint32 ** stats)
8812 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8815 * Check to see if stats are enabled
8818 MUTEX_ENTER(&rx_rpc_stats);
8819 if (!rxi_monitor_peerStats) {
8820 MUTEX_EXIT(&rx_rpc_stats);
8824 clock_GetTime(&now);
8825 *clock_sec = now.sec;
8826 *clock_usec = now.usec;
8829 * Allocate the space based upon the caller version
8831 * If the client is at an older version than we are,
8832 * we return the statistic data in the older data format, but
8833 * we still return our version number so the client knows we
8834 * are maintaining more data than it can retrieve.
8837 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8838 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8839 *statCount = rxi_rpc_peer_stat_cnt;
8842 * This can't happen yet, but in the future version changes
8843 * can be handled by adding additional code here
8847 if (space > (size_t) 0) {
8849 ptr = *stats = rxi_Alloc(space);
8852 struct opr_queue *cursor;
8854 for (opr_queue_Scan(&peerStats, cursor)) {
8855 struct rx_interface_stat *rpc_stat
8856 = opr_queue_Entry(cursor, struct rx_interface_stat,
8860 * Copy the data based upon the caller version
8862 rx_MarshallProcessRPCStats(callerVersion,
8863 rpc_stat->stats[0].func_total,
8864 rpc_stat->stats, &ptr);
8870 MUTEX_EXIT(&rx_rpc_stats);
8875 * rx_FreeRPCStats - free memory allocated by
8876 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8880 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8881 * rx_RetrievePeerRPCStats
8883 * IN allocSize - the number of bytes in stats.
8891 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8893 rxi_Free(stats, allocSize);
8897 * rx_queryProcessRPCStats - see if process rpc stat collection is
8898 * currently enabled.
8904 * Returns 0 if stats are not enabled != 0 otherwise
8908 rx_queryProcessRPCStats(void)
8911 MUTEX_ENTER(&rx_rpc_stats);
8912 rc = rxi_monitor_processStats;
8913 MUTEX_EXIT(&rx_rpc_stats);
8918 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8924 * Returns 0 if stats are not enabled != 0 otherwise
8928 rx_queryPeerRPCStats(void)
8931 MUTEX_ENTER(&rx_rpc_stats);
8932 rc = rxi_monitor_peerStats;
8933 MUTEX_EXIT(&rx_rpc_stats);
8938 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8948 rx_enableProcessRPCStats(void)
8950 MUTEX_ENTER(&rx_rpc_stats);
8951 rx_enable_stats = 1;
8952 rxi_monitor_processStats = 1;
8953 MUTEX_EXIT(&rx_rpc_stats);
8957 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8967 rx_enablePeerRPCStats(void)
8969 MUTEX_ENTER(&rx_rpc_stats);
8970 rx_enable_stats = 1;
8971 rxi_monitor_peerStats = 1;
8972 MUTEX_EXIT(&rx_rpc_stats);
8976 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8986 rx_disableProcessRPCStats(void)
8988 struct opr_queue *cursor, *store;
8991 MUTEX_ENTER(&rx_rpc_stats);
8994 * Turn off process statistics and if peer stats is also off, turn
8998 rxi_monitor_processStats = 0;
8999 if (rxi_monitor_peerStats == 0) {
9000 rx_enable_stats = 0;
9003 for (opr_queue_ScanSafe(&processStats, cursor, store)) {
9004 unsigned int num_funcs = 0;
9005 struct rx_interface_stat *rpc_stat
9006 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
9008 opr_queue_Remove(&rpc_stat->entry);
9010 num_funcs = rpc_stat->stats[0].func_total;
9012 sizeof(rx_interface_stat_t) +
9013 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
9015 rxi_Free(rpc_stat, space);
9016 rxi_rpc_process_stat_cnt -= num_funcs;
9018 MUTEX_EXIT(&rx_rpc_stats);
9022 * rx_disablePeerRPCStats - stop rpc stat collection for peers
9032 rx_disablePeerRPCStats(void)
9034 struct rx_peer **peer_ptr, **peer_end;
9038 * Turn off peer statistics and if process stats is also off, turn
9042 rxi_monitor_peerStats = 0;
9043 if (rxi_monitor_processStats == 0) {
9044 rx_enable_stats = 0;
9047 for (peer_ptr = &rx_peerHashTable[0], peer_end =
9048 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
9050 struct rx_peer *peer, *next, *prev;
9052 MUTEX_ENTER(&rx_peerHashTable_lock);
9053 MUTEX_ENTER(&rx_rpc_stats);
9054 for (prev = peer = *peer_ptr; peer; peer = next) {
9056 code = MUTEX_TRYENTER(&peer->peer_lock);
9059 struct opr_queue *cursor, *store;
9061 if (prev == *peer_ptr) {
9072 MUTEX_EXIT(&rx_peerHashTable_lock);
9074 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
9075 unsigned int num_funcs = 0;
9076 struct rx_interface_stat *rpc_stat
9077 = opr_queue_Entry(cursor, struct rx_interface_stat,
9080 opr_queue_Remove(&rpc_stat->entry);
9081 opr_queue_Remove(&rpc_stat->entryPeers);
9082 num_funcs = rpc_stat->stats[0].func_total;
9084 sizeof(rx_interface_stat_t) +
9085 rpc_stat->stats[0].func_total *
9086 sizeof(rx_function_entry_v1_t);
9088 rxi_Free(rpc_stat, space);
9089 rxi_rpc_peer_stat_cnt -= num_funcs;
9091 MUTEX_EXIT(&peer->peer_lock);
9093 MUTEX_ENTER(&rx_peerHashTable_lock);
9103 MUTEX_EXIT(&rx_rpc_stats);
9104 MUTEX_EXIT(&rx_peerHashTable_lock);
9109 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
9114 * IN clearFlag - flag indicating which stats to clear
9122 rx_clearProcessRPCStats(afs_uint32 clearFlag)
9124 struct opr_queue *cursor;
9126 MUTEX_ENTER(&rx_rpc_stats);
9128 for (opr_queue_Scan(&processStats, cursor)) {
9129 unsigned int num_funcs = 0, i;
9130 struct rx_interface_stat *rpc_stat
9131 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
9133 num_funcs = rpc_stat->stats[0].func_total;
9134 for (i = 0; i < num_funcs; i++) {
9135 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9136 rpc_stat->stats[i].invocations = 0;
9138 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9139 rpc_stat->stats[i].bytes_sent = 0;
9141 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9142 rpc_stat->stats[i].bytes_rcvd = 0;
9144 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9145 rpc_stat->stats[i].queue_time_sum.sec = 0;
9146 rpc_stat->stats[i].queue_time_sum.usec = 0;
9148 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9149 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9150 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9152 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9153 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9154 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9156 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9157 rpc_stat->stats[i].queue_time_max.sec = 0;
9158 rpc_stat->stats[i].queue_time_max.usec = 0;
9160 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9161 rpc_stat->stats[i].execution_time_sum.sec = 0;
9162 rpc_stat->stats[i].execution_time_sum.usec = 0;
9164 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9165 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9166 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9168 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9169 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9170 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9172 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9173 rpc_stat->stats[i].execution_time_max.sec = 0;
9174 rpc_stat->stats[i].execution_time_max.usec = 0;
9179 MUTEX_EXIT(&rx_rpc_stats);
9183 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
9188 * IN clearFlag - flag indicating which stats to clear
9196 rx_clearPeerRPCStats(afs_uint32 clearFlag)
9198 struct opr_queue *cursor;
9200 MUTEX_ENTER(&rx_rpc_stats);
9202 for (opr_queue_Scan(&peerStats, cursor)) {
9203 unsigned int num_funcs, i;
9204 struct rx_interface_stat *rpc_stat
9205 = opr_queue_Entry(cursor, struct rx_interface_stat, entryPeers);
9207 num_funcs = rpc_stat->stats[0].func_total;
9208 for (i = 0; i < num_funcs; i++) {
9209 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9210 rpc_stat->stats[i].invocations = 0;
9212 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9213 rpc_stat->stats[i].bytes_sent = 0;
9215 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9216 rpc_stat->stats[i].bytes_rcvd = 0;
9218 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9219 rpc_stat->stats[i].queue_time_sum.sec = 0;
9220 rpc_stat->stats[i].queue_time_sum.usec = 0;
9222 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9223 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9224 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9226 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9227 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9228 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9230 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9231 rpc_stat->stats[i].queue_time_max.sec = 0;
9232 rpc_stat->stats[i].queue_time_max.usec = 0;
9234 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9235 rpc_stat->stats[i].execution_time_sum.sec = 0;
9236 rpc_stat->stats[i].execution_time_sum.usec = 0;
9238 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9239 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9240 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9242 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9243 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9244 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9246 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9247 rpc_stat->stats[i].execution_time_max.sec = 0;
9248 rpc_stat->stats[i].execution_time_max.usec = 0;
9253 MUTEX_EXIT(&rx_rpc_stats);
9257 * rxi_rxstat_userok points to a routine that returns 1 if the caller
9258 * is authorized to enable/disable/clear RX statistics.
9260 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
9263 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
9265 rxi_rxstat_userok = proc;
9269 rx_RxStatUserOk(struct rx_call *call)
9271 if (!rxi_rxstat_userok)
9273 return rxi_rxstat_userok(call);
9278 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
9279 * function in the MSVC runtime DLL (msvcrt.dll).
9281 * Note: the system serializes calls to this function.
9284 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
9285 DWORD reason, /* reason function is being called */
9286 LPVOID reserved) /* reserved for future use */
9289 case DLL_PROCESS_ATTACH:
9290 /* library is being attached to a process */
9294 case DLL_PROCESS_DETACH:
9301 #endif /* AFS_NT40_ENV */
9304 int rx_DumpCalls(FILE *outputFile, char *cookie)
9306 #ifdef RXDEBUG_PACKET
9307 #ifdef KDUMP_RX_LOCK
9308 struct rx_call_rx_lock *c;
9315 #define RXDPRINTF sprintf
9316 #define RXDPRINTOUT output
9318 #define RXDPRINTF fprintf
9319 #define RXDPRINTOUT outputFile
9322 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9324 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9327 for (c = rx_allCallsp; c; c = c->allNextp) {
9328 u_short rqc, tqc, iovqc;
9330 MUTEX_ENTER(&c->lock);
9331 rqc = opr_queue_Count(&c->rq);
9332 tqc = opr_queue_Count(&c->tq);
9333 iovqc = opr_queue_Count(&c->app.iovq);
9335 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, "
9336 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9337 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9338 "resendEvent=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9339 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
9340 #ifdef RX_ENABLE_LOCKS
9343 #ifdef RX_REFCOUNT_CHECK
9344 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9345 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9348 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,
9349 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9350 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9351 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9352 c->resendEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9353 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
9354 #ifdef RX_ENABLE_LOCKS
9355 , (afs_uint32)c->refCount
9357 #ifdef RX_REFCOUNT_CHECK
9358 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9361 MUTEX_EXIT(&c->lock);
9364 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9367 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9369 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9371 #endif /* RXDEBUG_PACKET */